fs/ceph/dir.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/spinlock.h>
   5 #include <linux/fs_struct.h>
   6 #include <linux/namei.h>
   7 #include <linux/slab.h>
   8 #include <linux/sched.h>
   9 #include <linux/xattr.h>
  10
  11 #include "super.h"
  12 #include "mds_client.h"
  13
  14 /*
  15  * Directory operations: readdir, lookup, create, link, unlink,
  16  * rename, etc.
  17  */
  18
  19 /*
  20  * Ceph MDS operations are specified in terms of a base ino and
  21  * relative path.  Thus, the client can specify an operation on a
  22  * specific inode (e.g., a getattr due to fstat(2)), or as a path
  23  * relative to, say, the root directory.
  24  *
  25  * Normally, we limit ourselves to strict inode ops (no path component)
  26  * or dentry operations (a single path component relative to an ino).  The
  27  * exception to this is open_root_dentry(), which will open the mount
  28  * point by name.
  29  */
  30
  31 const struct dentry_operations ceph_dentry_ops;
  32
  33 /*
  34  * Initialize ceph dentry state.
  35  */
  36 static int ceph_d_init(struct dentry *dentry)
  37 {
  38         struct ceph_dentry_info *di;
  39
  40         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  41         if (!di)
  42                 return -ENOMEM;          /* oh well */
  43
  44         di->dentry = dentry;
  45         di->lease_session = NULL;
  46         di->time = jiffies;
  47         dentry->d_fsdata = di;
  48         ceph_dentry_lru_add(dentry);
  49         return 0;
  50 }
  51
  52 /*
  53  * for f_pos for readdir:
  54  * - hash order:
  55  *      (0xff << 52) | ((24 bits hash) << 28) |
  56  *      (the nth entry has hash collision);
  57  * - frag+name order;
  58  *      ((frag value) << 28) | (the nth entry in frag);
  59  */
  60 #define OFFSET_BITS     28
  61 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  62 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  63 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  64 {
  65         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  66         if (hash_order)
  67                 fpos |= HASH_ORDER;
  68         return fpos;
  69 }
  70
  71 static bool is_hash_order(loff_t p)
  72 {
  73         return (p & HASH_ORDER) == HASH_ORDER;
  74 }
  75
  76 static unsigned fpos_frag(loff_t p)
  77 {
  78         return p >> OFFSET_BITS;
  79 }
  80
  81 static unsigned fpos_hash(loff_t p)
  82 {
  83         return ceph_frag_value(fpos_frag(p));
  84 }
  85
  86 static unsigned fpos_off(loff_t p)
  87 {
  88         return p & OFFSET_MASK;
  89 }
  90
  91 static int fpos_cmp(loff_t l, loff_t r)
  92 {
  93         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
  94         if (v)
  95                 return v;
  96         return (int)(fpos_off(l) - fpos_off(r));
  97 }
  98
  99 /*
 100  * make note of the last dentry we read, so we can
 101  * continue at the same lexicographical point,
 102  * regardless of what dir changes take place on the
 103  * server.
 104  */
 105 static int note_last_dentry(struct ceph_file_info *fi, const char *name,
 106                             int len, unsigned next_offset)
 107 {
 108         char *buf = kmalloc(len+1, GFP_KERNEL);
 109         if (!buf)
 110                 return -ENOMEM;
 111         kfree(fi->last_name);
 112         fi->last_name = buf;
 113         memcpy(fi->last_name, name, len);
 114         fi->last_name[len] = 0;
 115         fi->next_offset = next_offset;
 116         dout("note_last_dentry '%s'\n", fi->last_name);
 117         return 0;
 118 }
 119
 120
 121 static struct dentry *
 122 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 123                         struct ceph_readdir_cache_control *cache_ctl)
 124 {
 125         struct inode *dir = d_inode(parent);
 126         struct dentry *dentry;
 127         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 128         loff_t ptr_pos = idx * sizeof(struct dentry *);
 129         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 130
 131         if (ptr_pos >= i_size_read(dir))
 132                 return NULL;
 133
 134         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 135                 ceph_readdir_cache_release(cache_ctl);
 136                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 137                 if (!cache_ctl->page) {
 138                         dout(" page %lu not found\n", ptr_pgoff);
 139                         return ERR_PTR(-EAGAIN);
 140                 }
 141                 /* reading/filling the cache are serialized by
 142                    i_mutex, no need to use page lock */
 143                 unlock_page(cache_ctl->page);
 144                 cache_ctl->dentries = kmap(cache_ctl->page);
 145         }
 146
 147         cache_ctl->index = idx & idx_mask;
 148
 149         rcu_read_lock();
 150         spin_lock(&parent->d_lock);
 151         /* check i_size again here, because empty directory can be
 152          * marked as complete while not holding the i_mutex. */
 153         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 154                 dentry = cache_ctl->dentries[cache_ctl->index];
 155         else
 156                 dentry = NULL;
 157         spin_unlock(&parent->d_lock);
 158         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 159                 dentry = NULL;
 160         rcu_read_unlock();
 161         return dentry ? : ERR_PTR(-EAGAIN);
 162 }
 163
 164 /*
 165  * When possible, we try to satisfy a readdir by peeking at the
 166  * dcache.  We make this work by carefully ordering dentries on
 167  * d_child when we initially get results back from the MDS, and
 168  * falling back to a "normal" sync readdir if any dentries in the dir
 169  * are dropped.
 170  *
 171  * Complete dir indicates that we have all dentries in the dir.  It is
 172  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 173  * the MDS if/when the directory is modified).
 174  */
 175 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 176                             u32 shared_gen)
 177 {
 178         struct ceph_file_info *fi = file->private_data;
 179         struct dentry *parent = file->f_path.dentry;
 180         struct inode *dir = d_inode(parent);
 181         struct dentry *dentry, *last = NULL;
 182         struct ceph_dentry_info *di;
 183         struct ceph_readdir_cache_control cache_ctl = {};
 184         u64 idx = 0;
 185         int err = 0;
 186
 187         dout("__dcache_readdir %p v%u at %llx\n", dir, shared_gen, ctx->pos);
 188
 189         /* search start position */
 190         if (ctx->pos > 2) {
 191                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 192                 while (count > 0) {
 193                         u64 step = count >> 1;
 194                         dentry = __dcache_find_get_entry(parent, idx + step,
 195                                                          &cache_ctl);
 196                         if (!dentry) {
 197                                 /* use linar search */
 198                                 idx = 0;
 199                                 break;
 200                         }
 201                         if (IS_ERR(dentry)) {
 202                                 err = PTR_ERR(dentry);
 203                                 goto out;
 204                         }
 205                         di = ceph_dentry(dentry);
 206                         spin_lock(&dentry->d_lock);
 207                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 208                                 idx += step + 1;
 209                                 count -= step + 1;
 210                         } else {
 211                                 count = step;
 212                         }
 213                         spin_unlock(&dentry->d_lock);
 214                         dput(dentry);
 215                 }
 216
 217                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 218         }
 219
 220
 221         for (;;) {
 222                 bool emit_dentry = false;
 223                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 224                 if (!dentry) {
 225                         fi->flags |= CEPH_F_ATEND;
 226                         err = 0;
 227                         break;
 228                 }
 229                 if (IS_ERR(dentry)) {
 230                         err = PTR_ERR(dentry);
 231                         goto out;
 232                 }
 233
 234                 di = ceph_dentry(dentry);
 235                 spin_lock(&dentry->d_lock);
 236                 if (di->lease_shared_gen == shared_gen &&
 237                     d_really_is_positive(dentry) &&
 238                     fpos_cmp(ctx->pos, di->offset) <= 0) {
 239                         emit_dentry = true;
 240                 }
 241                 spin_unlock(&dentry->d_lock);
 242
 243                 if (emit_dentry) {
 244                         dout(" %llx dentry %p %pd %p\n", di->offset,
 245                              dentry, dentry, d_inode(dentry));
 246                         ctx->pos = di->offset;
 247                         if (!dir_emit(ctx, dentry->d_name.name,
 248                                       dentry->d_name.len,
 249                                       ceph_translate_ino(dentry->d_sb,
 250                                                          d_inode(dentry)->i_ino),
 251                                       d_inode(dentry)->i_mode >> 12)) {
 252                                 dput(dentry);
 253                                 err = 0;
 254                                 break;
 255                         }
 256                         ctx->pos++;
 257
 258                         if (last)
 259                                 dput(last);
 260                         last = dentry;
 261                 } else {
 262                         dput(dentry);
 263                 }
 264         }
 265 out:
 266         ceph_readdir_cache_release(&cache_ctl);
 267         if (last) {
 268                 int ret;
 269                 di = ceph_dentry(last);
 270                 ret = note_last_dentry(fi, last->d_name.name, last->d_name.len,
 271                                        fpos_off(di->offset) + 1);
 272                 if (ret < 0)
 273                         err = ret;
 274                 dput(last);
 275                 /* last_name no longer match cache index */
 276                 if (fi->readdir_cache_idx >= 0) {
 277                         fi->readdir_cache_idx = -1;
 278                         fi->dir_release_count = 0;
 279                 }
 280         }
 281         return err;
 282 }
 283
 284 static bool need_send_readdir(struct ceph_file_info *fi, loff_t pos)
 285 {
 286         if (!fi->last_readdir)
 287                 return true;
 288         if (is_hash_order(pos))
 289                 return !ceph_frag_contains_value(fi->frag, fpos_hash(pos));
 290         else
 291                 return fi->frag != fpos_frag(pos);
 292 }
 293
 294 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 295 {
 296         struct ceph_file_info *fi = file->private_data;
 297         struct inode *inode = file_inode(file);
 298         struct ceph_inode_info *ci = ceph_inode(inode);
 299         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 300         struct ceph_mds_client *mdsc = fsc->mdsc;
 301         int i;
 302         int err;
 303         unsigned frag = -1;
 304         struct ceph_mds_reply_info_parsed *rinfo;
 305
 306         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 307         if (fi->flags & CEPH_F_ATEND)
 308                 return 0;
 309
 310         /* always start with . and .. */
 311         if (ctx->pos == 0) {
 312                 dout("readdir off 0 -> '.'\n");
 313                 if (!dir_emit(ctx, ".", 1,
 314                             ceph_translate_ino(inode->i_sb, inode->i_ino),
 315                             inode->i_mode >> 12))
 316                         return 0;
 317                 ctx->pos = 1;
 318         }
 319         if (ctx->pos == 1) {
 320                 ino_t ino = parent_ino(file->f_path.dentry);
 321                 dout("readdir off 1 -> '..'\n");
 322                 if (!dir_emit(ctx, "..", 2,
 323                             ceph_translate_ino(inode->i_sb, ino),
 324                             inode->i_mode >> 12))
 325                         return 0;
 326                 ctx->pos = 2;
 327         }
 328
 329         /* can we use the dcache? */
 330         spin_lock(&ci->i_ceph_lock);
 331         if (ceph_test_mount_opt(fsc, DCACHE) &&
 332             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 333             ceph_snap(inode) != CEPH_SNAPDIR &&
 334             __ceph_dir_is_complete_ordered(ci) &&
 335             __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
 336                 u32 shared_gen = ci->i_shared_gen;
 337                 spin_unlock(&ci->i_ceph_lock);
 338                 err = __dcache_readdir(file, ctx, shared_gen);
 339                 if (err != -EAGAIN)
 340                         return err;
 341         } else {
 342                 spin_unlock(&ci->i_ceph_lock);
 343         }
 344
 345         /* proceed with a normal readdir */
 346 more:
 347         /* do we have the correct frag content buffered? */
 348         if (need_send_readdir(fi, ctx->pos)) {
 349                 struct ceph_mds_request *req;
 350                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 351                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 352
 353                 /* discard old result, if any */
 354                 if (fi->last_readdir) {
 355                         ceph_mdsc_put_request(fi->last_readdir);
 356                         fi->last_readdir = NULL;
 357                 }
 358
 359                 if (is_hash_order(ctx->pos)) {
 360                         /* fragtree isn't always accurate. choose frag
 361                          * based on previous reply when possible. */
 362                         if (frag == (unsigned)-1)
 363                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 364                                                         NULL, NULL);
 365                 } else {
 366                         frag = fpos_frag(ctx->pos);
 367                 }
 368
 369                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 370                      ceph_vinop(inode), frag, fi->last_name);
 371                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 372                 if (IS_ERR(req))
 373                         return PTR_ERR(req);
 374                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 375                 if (err) {
 376                         ceph_mdsc_put_request(req);
 377                         return err;
 378                 }
 379                 /* hints to request -> mds selection code */
 380                 req->r_direct_mode = USE_AUTH_MDS;
 381                 if (op == CEPH_MDS_OP_READDIR) {
 382                         req->r_direct_hash = ceph_frag_value(frag);
 383                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 384                 }
 385                 if (fi->last_name) {
 386                         req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL);
 387                         if (!req->r_path2) {
 388                                 ceph_mdsc_put_request(req);
 389                                 return -ENOMEM;
 390                         }
 391                 } else if (is_hash_order(ctx->pos)) {
 392                         req->r_args.readdir.offset_hash =
 393                                 cpu_to_le32(fpos_hash(ctx->pos));
 394                 }
 395
 396                 req->r_dir_release_cnt = fi->dir_release_count;
 397                 req->r_dir_ordered_cnt = fi->dir_ordered_count;
 398                 req->r_readdir_cache_idx = fi->readdir_cache_idx;
 399                 req->r_readdir_offset = fi->next_offset;
 400                 req->r_args.readdir.frag = cpu_to_le32(frag);
 401                 req->r_args.readdir.flags =
 402                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 403
 404                 req->r_inode = inode;
 405                 ihold(inode);
 406                 req->r_dentry = dget(file->f_path.dentry);
 407                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 408                 if (err < 0) {
 409                         ceph_mdsc_put_request(req);
 410                         return err;
 411                 }
 412                 dout("readdir got and parsed readdir result=%d on "
 413                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 414                      err, frag,
 415                      (int)req->r_reply_info.dir_end,
 416                      (int)req->r_reply_info.dir_complete,
 417                      (int)req->r_reply_info.hash_order);
 418
 419                 rinfo = &req->r_reply_info;
 420                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 421                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 422                         if (!rinfo->hash_order) {
 423                                 fi->next_offset = req->r_readdir_offset;
 424                                 /* adjust ctx->pos to beginning of frag */
 425                                 ctx->pos = ceph_make_fpos(frag,
 426                                                           fi->next_offset,
 427                                                           false);
 428                         }
 429                 }
 430
 431                 fi->frag = frag;
 432                 fi->last_readdir = req;
 433
 434                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 435                         fi->readdir_cache_idx = req->r_readdir_cache_idx;
 436                         if (fi->readdir_cache_idx < 0) {
 437                                 /* preclude from marking dir ordered */
 438                                 fi->dir_ordered_count = 0;
 439                         } else if (ceph_frag_is_leftmost(frag) &&
 440                                    fi->next_offset == 2) {
 441                                 /* note dir version at start of readdir so
 442                                  * we can tell if any dentries get dropped */
 443                                 fi->dir_release_count = req->r_dir_release_cnt;
 444                                 fi->dir_ordered_count = req->r_dir_ordered_cnt;
 445                         }
 446                 } else {
 447                         dout("readdir !did_prepopulate");
 448                         /* disable readdir cache */
 449                         fi->readdir_cache_idx = -1;
 450                         /* preclude from marking dir complete */
 451                         fi->dir_release_count = 0;
 452                 }
 453
 454                 /* note next offset and last dentry name */
 455                 if (rinfo->dir_nr > 0) {
 456                         struct ceph_mds_reply_dir_entry *rde =
 457                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 458                         unsigned next_offset = req->r_reply_info.dir_end ?
 459                                         2 : (fpos_off(rde->offset) + 1);
 460                         err = note_last_dentry(fi, rde->name, rde->name_len,
 461                                                next_offset);
 462                         if (err)
 463                                 return err;
 464                 } else if (req->r_reply_info.dir_end) {
 465                         fi->next_offset = 2;
 466                         /* keep last name */
 467                 }
 468         }
 469
 470         rinfo = &fi->last_readdir->r_reply_info;
 471         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 472              fi->frag, rinfo->dir_nr, ctx->pos,
 473              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 474
 475         i = 0;
 476         /* search start position */
 477         if (rinfo->dir_nr > 0) {
 478                 int step, nr = rinfo->dir_nr;
 479                 while (nr > 0) {
 480                         step = nr >> 1;
 481                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 482                                 i +=  step + 1;
 483                                 nr -= step + 1;
 484                         } else {
 485                                 nr = step;
 486                         }
 487                 }
 488         }
 489         for (; i < rinfo->dir_nr; i++) {
 490                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 491                 struct ceph_vino vino;
 492                 ino_t ino;
 493                 u32 ftype;
 494
 495                 BUG_ON(rde->offset < ctx->pos);
 496
 497                 ctx->pos = rde->offset;
 498                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 499                      i, rinfo->dir_nr, ctx->pos,
 500                      rde->name_len, rde->name, &rde->inode.in);
 501
 502                 BUG_ON(!rde->inode.in);
 503                 ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
 504                 vino.ino = le64_to_cpu(rde->inode.in->ino);
 505                 vino.snap = le64_to_cpu(rde->inode.in->snapid);
 506                 ino = ceph_vino_to_ino(vino);
 507
 508                 if (!dir_emit(ctx, rde->name, rde->name_len,
 509                               ceph_translate_ino(inode->i_sb, ino), ftype)) {
 510                         dout("filldir stopping us...\n");
 511                         return 0;
 512                 }
 513                 ctx->pos++;
 514         }
 515
 516         ceph_mdsc_put_request(fi->last_readdir);
 517         fi->last_readdir = NULL;
 518
 519         if (fi->next_offset > 2) {
 520                 frag = fi->frag;
 521                 goto more;
 522         }
 523
 524         /* more frags? */
 525         if (!ceph_frag_is_rightmost(fi->frag)) {
 526                 frag = ceph_frag_next(fi->frag);
 527                 if (is_hash_order(ctx->pos)) {
 528                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 529                                                         fi->next_offset, true);
 530                         if (new_pos > ctx->pos)
 531                                 ctx->pos = new_pos;
 532                         /* keep last_name */
 533                 } else {
 534                         ctx->pos = ceph_make_fpos(frag, fi->next_offset, false);
 535                         kfree(fi->last_name);
 536                         fi->last_name = NULL;
 537                 }
 538                 dout("readdir next frag is %x\n", frag);
 539                 goto more;
 540         }
 541         fi->flags |= CEPH_F_ATEND;
 542
 543         /*
 544          * if dir_release_count still matches the dir, no dentries
 545          * were released during the whole readdir, and we should have
 546          * the complete dir contents in our cache.
 547          */
 548         if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) {
 549                 spin_lock(&ci->i_ceph_lock);
 550                 if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) {
 551                         dout(" marking %p complete and ordered\n", inode);
 552                         /* use i_size to track number of entries in
 553                          * readdir cache */
 554                         BUG_ON(fi->readdir_cache_idx < 0);
 555                         i_size_write(inode, fi->readdir_cache_idx *
 556                                      sizeof(struct dentry*));
 557                 } else {
 558                         dout(" marking %p complete\n", inode);
 559                 }
 560                 __ceph_dir_set_complete(ci, fi->dir_release_count,
 561                                         fi->dir_ordered_count);
 562                 spin_unlock(&ci->i_ceph_lock);
 563         }
 564
 565         dout("readdir %p file %p done.\n", inode, file);
 566         return 0;
 567 }
 568
 569 static void reset_readdir(struct ceph_file_info *fi)
 570 {
 571         if (fi->last_readdir) {
 572                 ceph_mdsc_put_request(fi->last_readdir);
 573                 fi->last_readdir = NULL;
 574         }
 575         kfree(fi->last_name);
 576         fi->last_name = NULL;
 577         fi->dir_release_count = 0;
 578         fi->readdir_cache_idx = -1;
 579         fi->next_offset = 2;  /* compensate for . and .. */
 580         fi->flags &= ~CEPH_F_ATEND;
 581 }
 582
 583 /*
 584  * discard buffered readdir content on seekdir(0), or seek to new frag,
 585  * or seek prior to current chunk
 586  */
 587 static bool need_reset_readdir(struct ceph_file_info *fi, loff_t new_pos)
 588 {
 589         struct ceph_mds_reply_info_parsed *rinfo;
 590         loff_t chunk_offset;
 591         if (new_pos == 0)
 592                 return true;
 593         if (is_hash_order(new_pos)) {
 594                 /* no need to reset last_name for a forward seek when
 595                  * dentries are sotred in hash order */
 596         } else if (fi->frag != fpos_frag(new_pos)) {
 597                 return true;
 598         }
 599         rinfo = fi->last_readdir ? &fi->last_readdir->r_reply_info : NULL;
 600         if (!rinfo || !rinfo->dir_nr)
 601                 return true;
 602         chunk_offset = rinfo->dir_entries[0].offset;
 603         return new_pos < chunk_offset ||
 604                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 605 }
 606
 607 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 608 {
 609         struct ceph_file_info *fi = file->private_data;
 610         struct inode *inode = file->f_mapping->host;
 611         loff_t retval;
 612
 613         inode_lock(inode);
 614         retval = -EINVAL;
 615         switch (whence) {
 616         case SEEK_CUR:
 617                 offset += file->f_pos;
 618         case SEEK_SET:
 619                 break;
 620         case SEEK_END:
 621                 retval = -EOPNOTSUPP;
 622         default:
 623                 goto out;
 624         }
 625
 626         if (offset >= 0) {
 627                 if (need_reset_readdir(fi, offset)) {
 628                         dout("dir_llseek dropping %p content\n", file);
 629                         reset_readdir(fi);
 630                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 631                         /* for hash offset, we don't know if a forward seek
 632                          * is within same frag */
 633                         fi->dir_release_count = 0;
 634                         fi->readdir_cache_idx = -1;
 635                 }
 636
 637                 if (offset != file->f_pos) {
 638                         file->f_pos = offset;
 639                         file->f_version = 0;
 640                         fi->flags &= ~CEPH_F_ATEND;
 641                 }
 642                 retval = offset;
 643         }
 644 out:
 645         inode_unlock(inode);
 646         return retval;
 647 }
 648
 649 /*
 650  * Handle lookups for the hidden .snap directory.
 651  */
 652 int ceph_handle_snapdir(struct ceph_mds_request *req,
 653                         struct dentry *dentry, int err)
 654 {
 655         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 656         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 657
 658         /* .snap dir? */
 659         if (err == -ENOENT &&
 660             ceph_snap(parent) == CEPH_NOSNAP &&
 661             strcmp(dentry->d_name.name,
 662                    fsc->mount_options->snapdir_name) == 0) {
 663                 struct inode *inode = ceph_get_snapdir(parent);
 664                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 665                      dentry, dentry, inode);
 666                 BUG_ON(!d_unhashed(dentry));
 667                 d_add(dentry, inode);
 668                 err = 0;
 669         }
 670         return err;
 671 }
 672
 673 /*
 674  * Figure out final result of a lookup/open request.
 675  *
 676  * Mainly, make sure we return the final req->r_dentry (if it already
 677  * existed) in place of the original VFS-provided dentry when they
 678  * differ.
 679  *
 680  * Gracefully handle the case where the MDS replies with -ENOENT and
 681  * no trace (which it may do, at its discretion, e.g., if it doesn't
 682  * care to issue a lease on the negative dentry).
 683  */
 684 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 685                                   struct dentry *dentry, int err)
 686 {
 687         if (err == -ENOENT) {
 688                 /* no trace? */
 689                 err = 0;
 690                 if (!req->r_reply_info.head->is_dentry) {
 691                         dout("ENOENT and no trace, dentry %p inode %p\n",
 692                              dentry, d_inode(dentry));
 693                         if (d_really_is_positive(dentry)) {
 694                                 d_drop(dentry);
 695                                 err = -ENOENT;
 696                         } else {
 697                                 d_add(dentry, NULL);
 698                         }
 699                 }
 700         }
 701         if (err)
 702                 dentry = ERR_PTR(err);
 703         else if (dentry != req->r_dentry)
 704                 dentry = dget(req->r_dentry);   /* we got spliced */
 705         else
 706                 dentry = NULL;
 707         return dentry;
 708 }
 709
 710 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 711 {
 712         return ceph_ino(inode) == CEPH_INO_ROOT &&
 713                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 714 }
 715
 716 /*
 717  * Look up a single dir entry.  If there is a lookup intent, inform
 718  * the MDS so that it gets our 'caps wanted' value in a single op.
 719  */
 720 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 721                                   unsigned int flags)
 722 {
 723         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 724         struct ceph_mds_client *mdsc = fsc->mdsc;
 725         struct ceph_mds_request *req;
 726         int op;
 727         int mask;
 728         int err;
 729
 730         dout("lookup %p dentry %p '%pd'\n",
 731              dir, dentry, dentry);
 732
 733         if (dentry->d_name.len > NAME_MAX)
 734                 return ERR_PTR(-ENAMETOOLONG);
 735
 736         /* can we conclude ENOENT locally? */
 737         if (d_really_is_negative(dentry)) {
 738                 struct ceph_inode_info *ci = ceph_inode(dir);
 739                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 740
 741                 spin_lock(&ci->i_ceph_lock);
 742                 dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
 743                 if (strncmp(dentry->d_name.name,
 744                             fsc->mount_options->snapdir_name,
 745                             dentry->d_name.len) &&
 746                     !is_root_ceph_dentry(dir, dentry) &&
 747                     ceph_test_mount_opt(fsc, DCACHE) &&
 748                     __ceph_dir_is_complete(ci) &&
 749                     (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
 750                         spin_unlock(&ci->i_ceph_lock);
 751                         dout(" dir %p complete, -ENOENT\n", dir);
 752                         d_add(dentry, NULL);
 753                         di->lease_shared_gen = ci->i_shared_gen;
 754                         return NULL;
 755                 }
 756                 spin_unlock(&ci->i_ceph_lock);
 757         }
 758
 759         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 760                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 761         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 762         if (IS_ERR(req))
 763                 return ERR_CAST(req);
 764         req->r_dentry = dget(dentry);
 765         req->r_num_caps = 2;
 766
 767         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 768         if (ceph_security_xattr_wanted(dir))
 769                 mask |= CEPH_CAP_XATTR_SHARED;
 770         req->r_args.getattr.mask = cpu_to_le32(mask);
 771
 772         req->r_parent = dir;
 773         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 774         err = ceph_mdsc_do_request(mdsc, NULL, req);
 775         err = ceph_handle_snapdir(req, dentry, err);
 776         dentry = ceph_finish_lookup(req, dentry, err);
 777         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 778         dout("lookup result=%p\n", dentry);
 779         return dentry;
 780 }
 781
 782 /*
 783  * If we do a create but get no trace back from the MDS, follow up with
 784  * a lookup (the VFS expects us to link up the provided dentry).
 785  */
 786 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 787 {
 788         struct dentry *result = ceph_lookup(dir, dentry, 0);
 789
 790         if (result && !IS_ERR(result)) {
 791                 /*
 792                  * We created the item, then did a lookup, and found
 793                  * it was already linked to another inode we already
 794                  * had in our cache (and thus got spliced). To not
 795                  * confuse VFS (especially when inode is a directory),
 796                  * we don't link our dentry to that inode, return an
 797                  * error instead.
 798                  *
 799                  * This event should be rare and it happens only when
 800                  * we talk to old MDS. Recent MDS does not send traceless
 801                  * reply for request that creates new inode.
 802                  */
 803                 d_drop(result);
 804                 return -ESTALE;
 805         }
 806         return PTR_ERR(result);
 807 }
 808
 809 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 810                       umode_t mode, dev_t rdev)
 811 {
 812         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 813         struct ceph_mds_client *mdsc = fsc->mdsc;
 814         struct ceph_mds_request *req;
 815         struct ceph_acls_info acls = {};
 816         int err;
 817
 818         if (ceph_snap(dir) != CEPH_NOSNAP)
 819                 return -EROFS;
 820
 821         err = ceph_pre_init_acls(dir, &mode, &acls);
 822         if (err < 0)
 823                 return err;
 824
 825         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 826              dir, dentry, mode, rdev);
 827         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 828         if (IS_ERR(req)) {
 829                 err = PTR_ERR(req);
 830                 goto out;
 831         }
 832         req->r_dentry = dget(dentry);
 833         req->r_num_caps = 2;
 834         req->r_parent = dir;
 835         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 836         req->r_args.mknod.mode = cpu_to_le32(mode);
 837         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 838         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 839         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 840         if (acls.pagelist) {
 841                 req->r_pagelist = acls.pagelist;
 842                 acls.pagelist = NULL;
 843         }
 844         err = ceph_mdsc_do_request(mdsc, dir, req);
 845         if (!err && !req->r_reply_info.head->is_dentry)
 846                 err = ceph_handle_notrace_create(dir, dentry);
 847         ceph_mdsc_put_request(req);
 848 out:
 849         if (!err)
 850                 ceph_init_inode_acls(d_inode(dentry), &acls);
 851         else
 852                 d_drop(dentry);
 853         ceph_release_acls_info(&acls);
 854         return err;
 855 }
 856
 857 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 858                        bool excl)
 859 {
 860         return ceph_mknod(dir, dentry, mode, 0);
 861 }
 862
 863 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 864                             const char *dest)
 865 {
 866         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 867         struct ceph_mds_client *mdsc = fsc->mdsc;
 868         struct ceph_mds_request *req;
 869         int err;
 870
 871         if (ceph_snap(dir) != CEPH_NOSNAP)
 872                 return -EROFS;
 873
 874         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 875         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 876         if (IS_ERR(req)) {
 877                 err = PTR_ERR(req);
 878                 goto out;
 879         }
 880         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 881         if (!req->r_path2) {
 882                 err = -ENOMEM;
 883                 ceph_mdsc_put_request(req);
 884                 goto out;
 885         }
 886         req->r_parent = dir;
 887         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 888         req->r_dentry = dget(dentry);
 889         req->r_num_caps = 2;
 890         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 891         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 892         err = ceph_mdsc_do_request(mdsc, dir, req);
 893         if (!err && !req->r_reply_info.head->is_dentry)
 894                 err = ceph_handle_notrace_create(dir, dentry);
 895         ceph_mdsc_put_request(req);
 896 out:
 897         if (err)
 898                 d_drop(dentry);
 899         return err;
 900 }
 901
 902 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 903 {
 904         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 905         struct ceph_mds_client *mdsc = fsc->mdsc;
 906         struct ceph_mds_request *req;
 907         struct ceph_acls_info acls = {};
 908         int err = -EROFS;
 909         int op;
 910
 911         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 912                 /* mkdir .snap/foo is a MKSNAP */
 913                 op = CEPH_MDS_OP_MKSNAP;
 914                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 915                      dentry, dentry);
 916         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 917                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 918                 op = CEPH_MDS_OP_MKDIR;
 919         } else {
 920                 goto out;
 921         }
 922
 923         mode |= S_IFDIR;
 924         err = ceph_pre_init_acls(dir, &mode, &acls);
 925         if (err < 0)
 926                 goto out;
 927
 928         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 929         if (IS_ERR(req)) {
 930                 err = PTR_ERR(req);
 931                 goto out;
 932         }
 933
 934         req->r_dentry = dget(dentry);
 935         req->r_num_caps = 2;
 936         req->r_parent = dir;
 937         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 938         req->r_args.mkdir.mode = cpu_to_le32(mode);
 939         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 940         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 941         if (acls.pagelist) {
 942                 req->r_pagelist = acls.pagelist;
 943                 acls.pagelist = NULL;
 944         }
 945         err = ceph_mdsc_do_request(mdsc, dir, req);
 946         if (!err &&
 947             !req->r_reply_info.head->is_target &&
 948             !req->r_reply_info.head->is_dentry)
 949                 err = ceph_handle_notrace_create(dir, dentry);
 950         ceph_mdsc_put_request(req);
 951 out:
 952         if (!err)
 953                 ceph_init_inode_acls(d_inode(dentry), &acls);
 954         else
 955                 d_drop(dentry);
 956         ceph_release_acls_info(&acls);
 957         return err;
 958 }
 959
 960 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
 961                      struct dentry *dentry)
 962 {
 963         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 964         struct ceph_mds_client *mdsc = fsc->mdsc;
 965         struct ceph_mds_request *req;
 966         int err;
 967
 968         if (ceph_snap(dir) != CEPH_NOSNAP)
 969                 return -EROFS;
 970
 971         dout("link in dir %p old_dentry %p dentry %p\n", dir,
 972              old_dentry, dentry);
 973         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
 974         if (IS_ERR(req)) {
 975                 d_drop(dentry);
 976                 return PTR_ERR(req);
 977         }
 978         req->r_dentry = dget(dentry);
 979         req->r_num_caps = 2;
 980         req->r_old_dentry = dget(old_dentry);
 981         req->r_parent = dir;
 982         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 983         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 984         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 985         /* release LINK_SHARED on source inode (mds will lock it) */
 986         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
 987         err = ceph_mdsc_do_request(mdsc, dir, req);
 988         if (err) {
 989                 d_drop(dentry);
 990         } else if (!req->r_reply_info.head->is_dentry) {
 991                 ihold(d_inode(old_dentry));
 992                 d_instantiate(dentry, d_inode(old_dentry));
 993         }
 994         ceph_mdsc_put_request(req);
 995         return err;
 996 }
 997
 998 /*
 999  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
1000  * looks like the link count will hit 0, drop any other caps (other
1001  * than PIN) we don't specifically want (due to the file still being
1002  * open).
1003  */
1004 static int drop_caps_for_unlink(struct inode *inode)
1005 {
1006         struct ceph_inode_info *ci = ceph_inode(inode);
1007         int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1008
1009         spin_lock(&ci->i_ceph_lock);
1010         if (inode->i_nlink == 1) {
1011                 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
1012                 ci->i_ceph_flags |= CEPH_I_NODELAY;
1013         }
1014         spin_unlock(&ci->i_ceph_lock);
1015         return drop;
1016 }
1017
1018 /*
1019  * rmdir and unlink are differ only by the metadata op code
1020  */
1021 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1022 {
1023         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1024         struct ceph_mds_client *mdsc = fsc->mdsc;
1025         struct inode *inode = d_inode(dentry);
1026         struct ceph_mds_request *req;
1027         int err = -EROFS;
1028         int op;
1029
1030         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1031                 /* rmdir .snap/foo is RMSNAP */
1032                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1033                 op = CEPH_MDS_OP_RMSNAP;
1034         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1035                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1036                      dir, dentry, inode);
1037                 op = d_is_dir(dentry) ?
1038                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1039         } else
1040                 goto out;
1041         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1042         if (IS_ERR(req)) {
1043                 err = PTR_ERR(req);
1044                 goto out;
1045         }
1046         req->r_dentry = dget(dentry);
1047         req->r_num_caps = 2;
1048         req->r_parent = dir;
1049         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1050         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1051         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1052         req->r_inode_drop = drop_caps_for_unlink(inode);
1053         err = ceph_mdsc_do_request(mdsc, dir, req);
1054         if (!err && !req->r_reply_info.head->is_dentry)
1055                 d_delete(dentry);
1056         ceph_mdsc_put_request(req);
1057 out:
1058         return err;
1059 }
1060
1061 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1062                        struct inode *new_dir, struct dentry *new_dentry,
1063                        unsigned int flags)
1064 {
1065         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1066         struct ceph_mds_client *mdsc = fsc->mdsc;
1067         struct ceph_mds_request *req;
1068         int op = CEPH_MDS_OP_RENAME;
1069         int err;
1070
1071         if (flags)
1072                 return -EINVAL;
1073
1074         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1075                 return -EXDEV;
1076         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1077                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1078                         op = CEPH_MDS_OP_RENAMESNAP;
1079                 else
1080                         return -EROFS;
1081         }
1082         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1083              old_dir, old_dentry, new_dir, new_dentry);
1084         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1085         if (IS_ERR(req))
1086                 return PTR_ERR(req);
1087         ihold(old_dir);
1088         req->r_dentry = dget(new_dentry);
1089         req->r_num_caps = 2;
1090         req->r_old_dentry = dget(old_dentry);
1091         req->r_old_dentry_dir = old_dir;
1092         req->r_parent = new_dir;
1093         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1094         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1095         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1096         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1097         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1098         /* release LINK_RDCACHE on source inode (mds will lock it) */
1099         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
1100         if (d_really_is_positive(new_dentry))
1101                 req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry));
1102         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1103         if (!err && !req->r_reply_info.head->is_dentry) {
1104                 /*
1105                  * Normally d_move() is done by fill_trace (called by
1106                  * do_request, above).  If there is no trace, we need
1107                  * to do it here.
1108                  */
1109
1110                 /* d_move screws up sibling dentries' offsets */
1111                 ceph_dir_clear_complete(old_dir);
1112                 ceph_dir_clear_complete(new_dir);
1113
1114                 d_move(old_dentry, new_dentry);
1115
1116                 /* ensure target dentry is invalidated, despite
1117                    rehashing bug in vfs_rename_dir */
1118                 ceph_invalidate_dentry_lease(new_dentry);
1119         }
1120         ceph_mdsc_put_request(req);
1121         return err;
1122 }
1123
1124 /*
1125  * Ensure a dentry lease will no longer revalidate.
1126  */
1127 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1128 {
1129         spin_lock(&dentry->d_lock);
1130         ceph_dentry(dentry)->time = jiffies;
1131         ceph_dentry(dentry)->lease_shared_gen = 0;
1132         spin_unlock(&dentry->d_lock);
1133 }
1134
1135 /*
1136  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1137  * renew if the least is more than half up.
1138  */
1139 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags,
1140                                  struct inode *dir)
1141 {
1142         struct ceph_dentry_info *di;
1143         struct ceph_mds_session *s;
1144         int valid = 0;
1145         u32 gen;
1146         unsigned long ttl;
1147         struct ceph_mds_session *session = NULL;
1148         u32 seq = 0;
1149
1150         spin_lock(&dentry->d_lock);
1151         di = ceph_dentry(dentry);
1152         if (di && di->lease_session) {
1153                 s = di->lease_session;
1154                 spin_lock(&s->s_gen_ttl_lock);
1155                 gen = s->s_cap_gen;
1156                 ttl = s->s_cap_ttl;
1157                 spin_unlock(&s->s_gen_ttl_lock);
1158
1159                 if (di->lease_gen == gen &&
1160                     time_before(jiffies, di->time) &&
1161                     time_before(jiffies, ttl)) {
1162                         valid = 1;
1163                         if (di->lease_renew_after &&
1164                             time_after(jiffies, di->lease_renew_after)) {
1165                                 /*
1166                                  * We should renew. If we're in RCU walk mode
1167                                  * though, we can't do that so just return
1168                                  * -ECHILD.
1169                                  */
1170                                 if (flags & LOOKUP_RCU) {
1171                                         valid = -ECHILD;
1172                                 } else {
1173                                         session = ceph_get_mds_session(s);
1174                                         seq = di->lease_seq;
1175                                         di->lease_renew_after = 0;
1176                                         di->lease_renew_from = jiffies;
1177                                 }
1178                         }
1179                 }
1180         }
1181         spin_unlock(&dentry->d_lock);
1182
1183         if (session) {
1184                 ceph_mdsc_lease_send_msg(session, dir, dentry,
1185                                          CEPH_MDS_LEASE_RENEW, seq);
1186                 ceph_put_mds_session(session);
1187         }
1188         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1189         return valid;
1190 }
1191
1192 /*
1193  * Check if directory-wide content lease/cap is valid.
1194  */
1195 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1196 {
1197         struct ceph_inode_info *ci = ceph_inode(dir);
1198         struct ceph_dentry_info *di = ceph_dentry(dentry);
1199         int valid = 0;
1200
1201         spin_lock(&ci->i_ceph_lock);
1202         if (ci->i_shared_gen == di->lease_shared_gen)
1203                 valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1204         spin_unlock(&ci->i_ceph_lock);
1205         dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1206              dir, (unsigned)ci->i_shared_gen, dentry,
1207              (unsigned)di->lease_shared_gen, valid);
1208         return valid;
1209 }
1210
1211 /*
1212  * Check if cached dentry can be trusted.
1213  */
1214 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1215 {
1216         int valid = 0;
1217         struct dentry *parent;
1218         struct inode *dir;
1219
1220         if (flags & LOOKUP_RCU) {
1221                 parent = READ_ONCE(dentry->d_parent);
1222                 dir = d_inode_rcu(parent);
1223                 if (!dir)
1224                         return -ECHILD;
1225         } else {
1226                 parent = dget_parent(dentry);
1227                 dir = d_inode(parent);
1228         }
1229
1230         dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1231              dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1232
1233         /* always trust cached snapped dentries, snapdir dentry */
1234         if (ceph_snap(dir) != CEPH_NOSNAP) {
1235                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1236                      dentry, d_inode(dentry));
1237                 valid = 1;
1238         } else if (d_really_is_positive(dentry) &&
1239                    ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1240                 valid = 1;
1241         } else {
1242                 valid = dentry_lease_is_valid(dentry, flags, dir);
1243                 if (valid == -ECHILD)
1244                         return valid;
1245                 if (valid || dir_lease_is_valid(dir, dentry)) {
1246                         if (d_really_is_positive(dentry))
1247                                 valid = ceph_is_any_caps(d_inode(dentry));
1248                         else
1249                                 valid = 1;
1250                 }
1251         }
1252
1253         if (!valid) {
1254                 struct ceph_mds_client *mdsc =
1255                         ceph_sb_to_client(dir->i_sb)->mdsc;
1256                 struct ceph_mds_request *req;
1257                 int op, err;
1258                 u32 mask;
1259
1260                 if (flags & LOOKUP_RCU)
1261                         return -ECHILD;
1262
1263                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1264                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1265                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1266                 if (!IS_ERR(req)) {
1267                         req->r_dentry = dget(dentry);
1268                         req->r_num_caps = 2;
1269                         req->r_parent = dir;
1270
1271                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1272                         if (ceph_security_xattr_wanted(dir))
1273                                 mask |= CEPH_CAP_XATTR_SHARED;
1274                         req->r_args.getattr.mask = cpu_to_le32(mask);
1275
1276                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1277                         switch (err) {
1278                         case 0:
1279                                 if (d_really_is_positive(dentry) &&
1280                                     d_inode(dentry) == req->r_target_inode)
1281                                         valid = 1;
1282                                 break;
1283                         case -ENOENT:
1284                                 if (d_really_is_negative(dentry))
1285                                         valid = 1;
1286                                 /* Fallthrough */
1287                         default:
1288                                 break;
1289                         }
1290                         ceph_mdsc_put_request(req);
1291                         dout("d_revalidate %p lookup result=%d\n",
1292                              dentry, err);
1293                 }
1294         }
1295
1296         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1297         if (valid) {
1298                 ceph_dentry_lru_touch(dentry);
1299         } else {
1300                 ceph_dir_clear_complete(dir);
1301         }
1302
1303         if (!(flags & LOOKUP_RCU))
1304                 dput(parent);
1305         return valid;
1306 }
1307
1308 /*
1309  * Release our ceph_dentry_info.
1310  */
1311 static void ceph_d_release(struct dentry *dentry)
1312 {
1313         struct ceph_dentry_info *di = ceph_dentry(dentry);
1314
1315         dout("d_release %p\n", dentry);
1316         ceph_dentry_lru_del(dentry);
1317
1318         spin_lock(&dentry->d_lock);
1319         dentry->d_fsdata = NULL;
1320         spin_unlock(&dentry->d_lock);
1321
1322         if (di->lease_session)
1323                 ceph_put_mds_session(di->lease_session);
1324         kmem_cache_free(ceph_dentry_cachep, di);
1325 }
1326
1327 /*
1328  * When the VFS prunes a dentry from the cache, we need to clear the
1329  * complete flag on the parent directory.
1330  *
1331  * Called under dentry->d_lock.
1332  */
1333 static void ceph_d_prune(struct dentry *dentry)
1334 {
1335         dout("ceph_d_prune %p\n", dentry);
1336
1337         /* do we have a valid parent? */
1338         if (IS_ROOT(dentry))
1339                 return;
1340
1341         /* if we are not hashed, we don't affect dir's completeness */
1342         if (d_unhashed(dentry))
1343                 return;
1344
1345         if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_SNAPDIR)
1346                 return;
1347
1348         /*
1349          * we hold d_lock, so d_parent is stable, and d_fsdata is never
1350          * cleared until d_release
1351          */
1352         ceph_dir_clear_complete(d_inode(dentry->d_parent));
1353 }
1354
1355 /*
1356  * read() on a dir.  This weird interface hack only works if mounted
1357  * with '-o dirstat'.
1358  */
1359 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1360                              loff_t *ppos)
1361 {
1362         struct ceph_file_info *cf = file->private_data;
1363         struct inode *inode = file_inode(file);
1364         struct ceph_inode_info *ci = ceph_inode(inode);
1365         int left;
1366         const int bufsize = 1024;
1367
1368         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1369                 return -EISDIR;
1370
1371         if (!cf->dir_info) {
1372                 cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1373                 if (!cf->dir_info)
1374                         return -ENOMEM;
1375                 cf->dir_info_len =
1376                         snprintf(cf->dir_info, bufsize,
1377                                 "entries:   %20lld\n"
1378                                 " files:    %20lld\n"
1379                                 " subdirs:  %20lld\n"
1380                                 "rentries:  %20lld\n"
1381                                 " rfiles:   %20lld\n"
1382                                 " rsubdirs: %20lld\n"
1383                                 "rbytes:    %20lld\n"
1384                                 "rctime:    %10ld.%09ld\n",
1385                                 ci->i_files + ci->i_subdirs,
1386                                 ci->i_files,
1387                                 ci->i_subdirs,
1388                                 ci->i_rfiles + ci->i_rsubdirs,
1389                                 ci->i_rfiles,
1390                                 ci->i_rsubdirs,
1391                                 ci->i_rbytes,
1392                                 (long)ci->i_rctime.tv_sec,
1393                                 (long)ci->i_rctime.tv_nsec);
1394         }
1395
1396         if (*ppos >= cf->dir_info_len)
1397                 return 0;
1398         size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1399         left = copy_to_user(buf, cf->dir_info + *ppos, size);
1400         if (left == size)
1401                 return -EFAULT;
1402         *ppos += (size - left);
1403         return size - left;
1404 }
1405
1406 /*
1407  * We maintain a private dentry LRU.
1408  *
1409  * FIXME: this needs to be changed to a per-mds lru to be useful.
1410  */
1411 void ceph_dentry_lru_add(struct dentry *dn)
1412 {
1413         struct ceph_dentry_info *di = ceph_dentry(dn);
1414         struct ceph_mds_client *mdsc;
1415
1416         dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1417         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1418         spin_lock(&mdsc->dentry_lru_lock);
1419         list_add_tail(&di->lru, &mdsc->dentry_lru);
1420         mdsc->num_dentry++;
1421         spin_unlock(&mdsc->dentry_lru_lock);
1422 }
1423
1424 void ceph_dentry_lru_touch(struct dentry *dn)
1425 {
1426         struct ceph_dentry_info *di = ceph_dentry(dn);
1427         struct ceph_mds_client *mdsc;
1428
1429         dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1430              di->offset);
1431         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1432         spin_lock(&mdsc->dentry_lru_lock);
1433         list_move_tail(&di->lru, &mdsc->dentry_lru);
1434         spin_unlock(&mdsc->dentry_lru_lock);
1435 }
1436
1437 void ceph_dentry_lru_del(struct dentry *dn)
1438 {
1439         struct ceph_dentry_info *di = ceph_dentry(dn);
1440         struct ceph_mds_client *mdsc;
1441
1442         dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1443         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1444         spin_lock(&mdsc->dentry_lru_lock);
1445         list_del_init(&di->lru);
1446         mdsc->num_dentry--;
1447         spin_unlock(&mdsc->dentry_lru_lock);
1448 }
1449
1450 /*
1451  * Return name hash for a given dentry.  This is dependent on
1452  * the parent directory's hash function.
1453  */
1454 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1455 {
1456         struct ceph_inode_info *dci = ceph_inode(dir);
1457
1458         switch (dci->i_dir_layout.dl_dir_hash) {
1459         case 0: /* for backward compat */
1460         case CEPH_STR_HASH_LINUX:
1461                 return dn->d_name.hash;
1462
1463         default:
1464                 return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1465                                      dn->d_name.name, dn->d_name.len);
1466         }
1467 }
1468
1469 const struct file_operations ceph_dir_fops = {
1470         .read = ceph_read_dir,
1471         .iterate = ceph_readdir,
1472         .llseek = ceph_dir_llseek,
1473         .open = ceph_open,
1474         .release = ceph_release,
1475         .unlocked_ioctl = ceph_ioctl,
1476         .fsync = ceph_fsync,
1477 };
1478
1479 const struct file_operations ceph_snapdir_fops = {
1480         .iterate = ceph_readdir,
1481         .llseek = ceph_dir_llseek,
1482         .open = ceph_open,
1483         .release = ceph_release,
1484 };
1485
1486 const struct inode_operations ceph_dir_iops = {
1487         .lookup = ceph_lookup,
1488         .permission = ceph_permission,
1489         .getattr = ceph_getattr,
1490         .setattr = ceph_setattr,
1491         .listxattr = ceph_listxattr,
1492         .get_acl = ceph_get_acl,
1493         .set_acl = ceph_set_acl,
1494         .mknod = ceph_mknod,
1495         .symlink = ceph_symlink,
1496         .mkdir = ceph_mkdir,
1497         .link = ceph_link,
1498         .unlink = ceph_unlink,
1499         .rmdir = ceph_unlink,
1500         .rename = ceph_rename,
1501         .create = ceph_create,
1502         .atomic_open = ceph_atomic_open,
1503 };
1504
1505 const struct inode_operations ceph_snapdir_iops = {
1506         .lookup = ceph_lookup,
1507         .permission = ceph_permission,
1508         .getattr = ceph_getattr,
1509         .mkdir = ceph_mkdir,
1510         .rmdir = ceph_unlink,
1511         .rename = ceph_rename,
1512 };
1513
1514 const struct dentry_operations ceph_dentry_ops = {
1515         .d_revalidate = ceph_d_revalidate,
1516         .d_release = ceph_d_release,
1517         .d_prune = ceph_d_prune,
1518         .d_init = ceph_d_init,
1519 };