fs/ceph/dir.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/spinlock.h>
   5 #include <linux/namei.h>
   6 #include <linux/slab.h>
   7 #include <linux/sched.h>
   8 #include <linux/xattr.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12 #include "crypto.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 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di);
  34 static int __dir_lease_try_check(const struct dentry *dentry);
  35
  36 /*
  37  * Initialize ceph dentry state.
  38  */
  39 static int ceph_d_init(struct dentry *dentry)
  40 {
  41         struct ceph_dentry_info *di;
  42         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dentry->d_sb);
  43
  44         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  45         if (!di)
  46                 return -ENOMEM;          /* oh well */
  47
  48         di->dentry = dentry;
  49         di->lease_session = NULL;
  50         di->time = jiffies;
  51         dentry->d_fsdata = di;
  52         INIT_LIST_HEAD(&di->lease_list);
  53
  54         atomic64_inc(&mdsc->metric.total_dentries);
  55
  56         return 0;
  57 }
  58
  59 /*
  60  * for f_pos for readdir:
  61  * - hash order:
  62  *      (0xff << 52) | ((24 bits hash) << 28) |
  63  *      (the nth entry has hash collision);
  64  * - frag+name order;
  65  *      ((frag value) << 28) | (the nth entry in frag);
  66  */
  67 #define OFFSET_BITS     28
  68 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  69 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  70 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  71 {
  72         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  73         if (hash_order)
  74                 fpos |= HASH_ORDER;
  75         return fpos;
  76 }
  77
  78 static bool is_hash_order(loff_t p)
  79 {
  80         return (p & HASH_ORDER) == HASH_ORDER;
  81 }
  82
  83 static unsigned fpos_frag(loff_t p)
  84 {
  85         return p >> OFFSET_BITS;
  86 }
  87
  88 static unsigned fpos_hash(loff_t p)
  89 {
  90         return ceph_frag_value(fpos_frag(p));
  91 }
  92
  93 static unsigned fpos_off(loff_t p)
  94 {
  95         return p & OFFSET_MASK;
  96 }
  97
  98 static int fpos_cmp(loff_t l, loff_t r)
  99 {
 100         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
 101         if (v)
 102                 return v;
 103         return (int)(fpos_off(l) - fpos_off(r));
 104 }
 105
 106 /*
 107  * make note of the last dentry we read, so we can
 108  * continue at the same lexicographical point,
 109  * regardless of what dir changes take place on the
 110  * server.
 111  */
 112 static int note_last_dentry(struct ceph_dir_file_info *dfi, const char *name,
 113                             int len, unsigned next_offset)
 114 {
 115         char *buf = kmalloc(len+1, GFP_KERNEL);
 116         if (!buf)
 117                 return -ENOMEM;
 118         kfree(dfi->last_name);
 119         dfi->last_name = buf;
 120         memcpy(dfi->last_name, name, len);
 121         dfi->last_name[len] = 0;
 122         dfi->next_offset = next_offset;
 123         dout("note_last_dentry '%s'\n", dfi->last_name);
 124         return 0;
 125 }
 126
 127
 128 static struct dentry *
 129 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 130                         struct ceph_readdir_cache_control *cache_ctl)
 131 {
 132         struct inode *dir = d_inode(parent);
 133         struct dentry *dentry;
 134         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 135         loff_t ptr_pos = idx * sizeof(struct dentry *);
 136         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 137
 138         if (ptr_pos >= i_size_read(dir))
 139                 return NULL;
 140
 141         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 142                 ceph_readdir_cache_release(cache_ctl);
 143                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 144                 if (!cache_ctl->page) {
 145                         dout(" page %lu not found\n", ptr_pgoff);
 146                         return ERR_PTR(-EAGAIN);
 147                 }
 148                 /* reading/filling the cache are serialized by
 149                    i_rwsem, no need to use page lock */
 150                 unlock_page(cache_ctl->page);
 151                 cache_ctl->dentries = kmap(cache_ctl->page);
 152         }
 153
 154         cache_ctl->index = idx & idx_mask;
 155
 156         rcu_read_lock();
 157         spin_lock(&parent->d_lock);
 158         /* check i_size again here, because empty directory can be
 159          * marked as complete while not holding the i_rwsem. */
 160         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 161                 dentry = cache_ctl->dentries[cache_ctl->index];
 162         else
 163                 dentry = NULL;
 164         spin_unlock(&parent->d_lock);
 165         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 166                 dentry = NULL;
 167         rcu_read_unlock();
 168         return dentry ? : ERR_PTR(-EAGAIN);
 169 }
 170
 171 /*
 172  * When possible, we try to satisfy a readdir by peeking at the
 173  * dcache.  We make this work by carefully ordering dentries on
 174  * d_child when we initially get results back from the MDS, and
 175  * falling back to a "normal" sync readdir if any dentries in the dir
 176  * are dropped.
 177  *
 178  * Complete dir indicates that we have all dentries in the dir.  It is
 179  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 180  * the MDS if/when the directory is modified).
 181  */
 182 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 183                             int shared_gen)
 184 {
 185         struct ceph_dir_file_info *dfi = file->private_data;
 186         struct dentry *parent = file->f_path.dentry;
 187         struct inode *dir = d_inode(parent);
 188         struct dentry *dentry, *last = NULL;
 189         struct ceph_dentry_info *di;
 190         struct ceph_readdir_cache_control cache_ctl = {};
 191         u64 idx = 0;
 192         int err = 0;
 193
 194         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 195
 196         /* search start position */
 197         if (ctx->pos > 2) {
 198                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 199                 while (count > 0) {
 200                         u64 step = count >> 1;
 201                         dentry = __dcache_find_get_entry(parent, idx + step,
 202                                                          &cache_ctl);
 203                         if (!dentry) {
 204                                 /* use linar search */
 205                                 idx = 0;
 206                                 break;
 207                         }
 208                         if (IS_ERR(dentry)) {
 209                                 err = PTR_ERR(dentry);
 210                                 goto out;
 211                         }
 212                         di = ceph_dentry(dentry);
 213                         spin_lock(&dentry->d_lock);
 214                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 215                                 idx += step + 1;
 216                                 count -= step + 1;
 217                         } else {
 218                                 count = step;
 219                         }
 220                         spin_unlock(&dentry->d_lock);
 221                         dput(dentry);
 222                 }
 223
 224                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 225         }
 226
 227
 228         for (;;) {
 229                 bool emit_dentry = false;
 230                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 231                 if (!dentry) {
 232                         dfi->file_info.flags |= CEPH_F_ATEND;
 233                         err = 0;
 234                         break;
 235                 }
 236                 if (IS_ERR(dentry)) {
 237                         err = PTR_ERR(dentry);
 238                         goto out;
 239                 }
 240
 241                 spin_lock(&dentry->d_lock);
 242                 di = ceph_dentry(dentry);
 243                 if (d_unhashed(dentry) ||
 244                     d_really_is_negative(dentry) ||
 245                     di->lease_shared_gen != shared_gen ||
 246                     ((dentry->d_flags & DCACHE_NOKEY_NAME) &&
 247                      fscrypt_has_encryption_key(dir))) {
 248                         spin_unlock(&dentry->d_lock);
 249                         dput(dentry);
 250                         err = -EAGAIN;
 251                         goto out;
 252                 }
 253                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 254                         __ceph_dentry_dir_lease_touch(di);
 255                         emit_dentry = true;
 256                 }
 257                 spin_unlock(&dentry->d_lock);
 258
 259                 if (emit_dentry) {
 260                         dout(" %llx dentry %p %pd %p\n", di->offset,
 261                              dentry, dentry, d_inode(dentry));
 262                         ctx->pos = di->offset;
 263                         if (!dir_emit(ctx, dentry->d_name.name,
 264                                       dentry->d_name.len, ceph_present_inode(d_inode(dentry)),
 265                                       d_inode(dentry)->i_mode >> 12)) {
 266                                 dput(dentry);
 267                                 err = 0;
 268                                 break;
 269                         }
 270                         ctx->pos++;
 271
 272                         if (last)
 273                                 dput(last);
 274                         last = dentry;
 275                 } else {
 276                         dput(dentry);
 277                 }
 278         }
 279 out:
 280         ceph_readdir_cache_release(&cache_ctl);
 281         if (last) {
 282                 int ret;
 283                 di = ceph_dentry(last);
 284                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 285                                        fpos_off(di->offset) + 1);
 286                 if (ret < 0)
 287                         err = ret;
 288                 dput(last);
 289                 /* last_name no longer match cache index */
 290                 if (dfi->readdir_cache_idx >= 0) {
 291                         dfi->readdir_cache_idx = -1;
 292                         dfi->dir_release_count = 0;
 293                 }
 294         }
 295         return err;
 296 }
 297
 298 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 299 {
 300         if (!dfi->last_readdir)
 301                 return true;
 302         if (is_hash_order(pos))
 303                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 304         else
 305                 return dfi->frag != fpos_frag(pos);
 306 }
 307
 308 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 309 {
 310         struct ceph_dir_file_info *dfi = file->private_data;
 311         struct inode *inode = file_inode(file);
 312         struct ceph_inode_info *ci = ceph_inode(inode);
 313         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 314         struct ceph_mds_client *mdsc = fsc->mdsc;
 315         int i;
 316         int err;
 317         unsigned frag = -1;
 318         struct ceph_mds_reply_info_parsed *rinfo;
 319
 320         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 321         if (dfi->file_info.flags & CEPH_F_ATEND)
 322                 return 0;
 323
 324         /* always start with . and .. */
 325         if (ctx->pos == 0) {
 326                 dout("readdir off 0 -> '.'\n");
 327                 if (!dir_emit(ctx, ".", 1, ceph_present_inode(inode),
 328                             inode->i_mode >> 12))
 329                         return 0;
 330                 ctx->pos = 1;
 331         }
 332         if (ctx->pos == 1) {
 333                 u64 ino;
 334                 struct dentry *dentry = file->f_path.dentry;
 335
 336                 spin_lock(&dentry->d_lock);
 337                 ino = ceph_present_inode(dentry->d_parent->d_inode);
 338                 spin_unlock(&dentry->d_lock);
 339
 340                 dout("readdir off 1 -> '..'\n");
 341                 if (!dir_emit(ctx, "..", 2, ino, inode->i_mode >> 12))
 342                         return 0;
 343                 ctx->pos = 2;
 344         }
 345
 346         err = ceph_fscrypt_prepare_readdir(inode);
 347         if (err < 0)
 348                 return err;
 349
 350         spin_lock(&ci->i_ceph_lock);
 351         /* request Fx cap. if have Fx, we don't need to release Fs cap
 352          * for later create/unlink. */
 353         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_WR);
 354         /* can we use the dcache? */
 355         if (ceph_test_mount_opt(fsc, DCACHE) &&
 356             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 357             ceph_snap(inode) != CEPH_SNAPDIR &&
 358             __ceph_dir_is_complete_ordered(ci) &&
 359             __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 360                 int shared_gen = atomic_read(&ci->i_shared_gen);
 361
 362                 spin_unlock(&ci->i_ceph_lock);
 363                 err = __dcache_readdir(file, ctx, shared_gen);
 364                 if (err != -EAGAIN)
 365                         return err;
 366         } else {
 367                 spin_unlock(&ci->i_ceph_lock);
 368         }
 369
 370         /* proceed with a normal readdir */
 371 more:
 372         /* do we have the correct frag content buffered? */
 373         if (need_send_readdir(dfi, ctx->pos)) {
 374                 struct ceph_mds_request *req;
 375                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 376                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 377
 378                 /* discard old result, if any */
 379                 if (dfi->last_readdir) {
 380                         ceph_mdsc_put_request(dfi->last_readdir);
 381                         dfi->last_readdir = NULL;
 382                 }
 383
 384                 if (is_hash_order(ctx->pos)) {
 385                         /* fragtree isn't always accurate. choose frag
 386                          * based on previous reply when possible. */
 387                         if (frag == (unsigned)-1)
 388                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 389                                                         NULL, NULL);
 390                 } else {
 391                         frag = fpos_frag(ctx->pos);
 392                 }
 393
 394                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 395                      ceph_vinop(inode), frag, dfi->last_name);
 396                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 397                 if (IS_ERR(req))
 398                         return PTR_ERR(req);
 399
 400                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 401                 if (err) {
 402                         ceph_mdsc_put_request(req);
 403                         return err;
 404                 }
 405                 /* hints to request -> mds selection code */
 406                 req->r_direct_mode = USE_AUTH_MDS;
 407                 if (op == CEPH_MDS_OP_READDIR) {
 408                         req->r_direct_hash = ceph_frag_value(frag);
 409                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 410                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 411                 }
 412                 if (dfi->last_name) {
 413                         struct qstr d_name = { .name = dfi->last_name,
 414                                                .len = strlen(dfi->last_name) };
 415
 416                         req->r_path2 = kzalloc(NAME_MAX + 1, GFP_KERNEL);
 417                         if (!req->r_path2) {
 418                                 ceph_mdsc_put_request(req);
 419                                 return -ENOMEM;
 420                         }
 421
 422                         err = ceph_encode_encrypted_dname(inode, &d_name,
 423                                                           req->r_path2);
 424                         if (err < 0) {
 425                                 ceph_mdsc_put_request(req);
 426                                 return err;
 427                         }
 428                 } else if (is_hash_order(ctx->pos)) {
 429                         req->r_args.readdir.offset_hash =
 430                                 cpu_to_le32(fpos_hash(ctx->pos));
 431                 }
 432
 433                 req->r_dir_release_cnt = dfi->dir_release_count;
 434                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 435                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 436                 req->r_readdir_offset = dfi->next_offset;
 437                 req->r_args.readdir.frag = cpu_to_le32(frag);
 438                 req->r_args.readdir.flags =
 439                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 440
 441                 req->r_inode = inode;
 442                 ihold(inode);
 443                 req->r_dentry = dget(file->f_path.dentry);
 444                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 445                 if (err < 0) {
 446                         ceph_mdsc_put_request(req);
 447                         return err;
 448                 }
 449                 dout("readdir got and parsed readdir result=%d on "
 450                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 451                      err, frag,
 452                      (int)req->r_reply_info.dir_end,
 453                      (int)req->r_reply_info.dir_complete,
 454                      (int)req->r_reply_info.hash_order);
 455
 456                 rinfo = &req->r_reply_info;
 457                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 458                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 459                         if (!rinfo->hash_order) {
 460                                 dfi->next_offset = req->r_readdir_offset;
 461                                 /* adjust ctx->pos to beginning of frag */
 462                                 ctx->pos = ceph_make_fpos(frag,
 463                                                           dfi->next_offset,
 464                                                           false);
 465                         }
 466                 }
 467
 468                 dfi->frag = frag;
 469                 dfi->last_readdir = req;
 470
 471                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 472                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 473                         if (dfi->readdir_cache_idx < 0) {
 474                                 /* preclude from marking dir ordered */
 475                                 dfi->dir_ordered_count = 0;
 476                         } else if (ceph_frag_is_leftmost(frag) &&
 477                                    dfi->next_offset == 2) {
 478                                 /* note dir version at start of readdir so
 479                                  * we can tell if any dentries get dropped */
 480                                 dfi->dir_release_count = req->r_dir_release_cnt;
 481                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 482                         }
 483                 } else {
 484                         dout("readdir !did_prepopulate\n");
 485                         /* disable readdir cache */
 486                         dfi->readdir_cache_idx = -1;
 487                         /* preclude from marking dir complete */
 488                         dfi->dir_release_count = 0;
 489                 }
 490
 491                 /* note next offset and last dentry name */
 492                 if (rinfo->dir_nr > 0) {
 493                         struct ceph_mds_reply_dir_entry *rde =
 494                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 495                         unsigned next_offset = req->r_reply_info.dir_end ?
 496                                         2 : (fpos_off(rde->offset) + 1);
 497                         err = note_last_dentry(dfi, rde->name, rde->name_len,
 498                                                next_offset);
 499                         if (err) {
 500                                 ceph_mdsc_put_request(dfi->last_readdir);
 501                                 dfi->last_readdir = NULL;
 502                                 return err;
 503                         }
 504                 } else if (req->r_reply_info.dir_end) {
 505                         dfi->next_offset = 2;
 506                         /* keep last name */
 507                 }
 508         }
 509
 510         rinfo = &dfi->last_readdir->r_reply_info;
 511         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 512              dfi->frag, rinfo->dir_nr, ctx->pos,
 513              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 514
 515         i = 0;
 516         /* search start position */
 517         if (rinfo->dir_nr > 0) {
 518                 int step, nr = rinfo->dir_nr;
 519                 while (nr > 0) {
 520                         step = nr >> 1;
 521                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 522                                 i +=  step + 1;
 523                                 nr -= step + 1;
 524                         } else {
 525                                 nr = step;
 526                         }
 527                 }
 528         }
 529         for (; i < rinfo->dir_nr; i++) {
 530                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 531
 532                 if (rde->offset < ctx->pos) {
 533                         pr_warn("%s: rde->offset 0x%llx ctx->pos 0x%llx\n",
 534                                 __func__, rde->offset, ctx->pos);
 535                         return -EIO;
 536                 }
 537
 538                 if (WARN_ON_ONCE(!rde->inode.in))
 539                         return -EIO;
 540
 541                 ctx->pos = rde->offset;
 542                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 543                      i, rinfo->dir_nr, ctx->pos,
 544                      rde->name_len, rde->name, &rde->inode.in);
 545
 546                 if (!dir_emit(ctx, rde->name, rde->name_len,
 547                               ceph_present_ino(inode->i_sb, le64_to_cpu(rde->inode.in->ino)),
 548                               le32_to_cpu(rde->inode.in->mode) >> 12)) {
 549                         /*
 550                          * NOTE: Here no need to put the 'dfi->last_readdir',
 551                          * because when dir_emit stops us it's most likely
 552                          * doesn't have enough memory, etc. So for next readdir
 553                          * it will continue.
 554                          */
 555                         dout("filldir stopping us...\n");
 556                         return 0;
 557                 }
 558
 559                 /* Reset the lengths to their original allocated vals */
 560                 ctx->pos++;
 561         }
 562
 563         ceph_mdsc_put_request(dfi->last_readdir);
 564         dfi->last_readdir = NULL;
 565
 566         if (dfi->next_offset > 2) {
 567                 frag = dfi->frag;
 568                 goto more;
 569         }
 570
 571         /* more frags? */
 572         if (!ceph_frag_is_rightmost(dfi->frag)) {
 573                 frag = ceph_frag_next(dfi->frag);
 574                 if (is_hash_order(ctx->pos)) {
 575                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 576                                                         dfi->next_offset, true);
 577                         if (new_pos > ctx->pos)
 578                                 ctx->pos = new_pos;
 579                         /* keep last_name */
 580                 } else {
 581                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 582                                                         false);
 583                         kfree(dfi->last_name);
 584                         dfi->last_name = NULL;
 585                 }
 586                 dout("readdir next frag is %x\n", frag);
 587                 goto more;
 588         }
 589         dfi->file_info.flags |= CEPH_F_ATEND;
 590
 591         /*
 592          * if dir_release_count still matches the dir, no dentries
 593          * were released during the whole readdir, and we should have
 594          * the complete dir contents in our cache.
 595          */
 596         if (atomic64_read(&ci->i_release_count) ==
 597                                         dfi->dir_release_count) {
 598                 spin_lock(&ci->i_ceph_lock);
 599                 if (dfi->dir_ordered_count ==
 600                                 atomic64_read(&ci->i_ordered_count)) {
 601                         dout(" marking %p complete and ordered\n", inode);
 602                         /* use i_size to track number of entries in
 603                          * readdir cache */
 604                         BUG_ON(dfi->readdir_cache_idx < 0);
 605                         i_size_write(inode, dfi->readdir_cache_idx *
 606                                      sizeof(struct dentry*));
 607                 } else {
 608                         dout(" marking %p complete\n", inode);
 609                 }
 610                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 611                                         dfi->dir_ordered_count);
 612                 spin_unlock(&ci->i_ceph_lock);
 613         }
 614         dout("readdir %p file %p done.\n", inode, file);
 615         return 0;
 616 }
 617
 618 static void reset_readdir(struct ceph_dir_file_info *dfi)
 619 {
 620         if (dfi->last_readdir) {
 621                 ceph_mdsc_put_request(dfi->last_readdir);
 622                 dfi->last_readdir = NULL;
 623         }
 624         kfree(dfi->last_name);
 625         dfi->last_name = NULL;
 626         dfi->dir_release_count = 0;
 627         dfi->readdir_cache_idx = -1;
 628         dfi->next_offset = 2;  /* compensate for . and .. */
 629         dfi->file_info.flags &= ~CEPH_F_ATEND;
 630 }
 631
 632 /*
 633  * discard buffered readdir content on seekdir(0), or seek to new frag,
 634  * or seek prior to current chunk
 635  */
 636 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 637 {
 638         struct ceph_mds_reply_info_parsed *rinfo;
 639         loff_t chunk_offset;
 640         if (new_pos == 0)
 641                 return true;
 642         if (is_hash_order(new_pos)) {
 643                 /* no need to reset last_name for a forward seek when
 644                  * dentries are sotred in hash order */
 645         } else if (dfi->frag != fpos_frag(new_pos)) {
 646                 return true;
 647         }
 648         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 649         if (!rinfo || !rinfo->dir_nr)
 650                 return true;
 651         chunk_offset = rinfo->dir_entries[0].offset;
 652         return new_pos < chunk_offset ||
 653                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 654 }
 655
 656 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 657 {
 658         struct ceph_dir_file_info *dfi = file->private_data;
 659         struct inode *inode = file->f_mapping->host;
 660         loff_t retval;
 661
 662         inode_lock(inode);
 663         retval = -EINVAL;
 664         switch (whence) {
 665         case SEEK_CUR:
 666                 offset += file->f_pos;
 667                 break;
 668         case SEEK_SET:
 669                 break;
 670         case SEEK_END:
 671                 retval = -EOPNOTSUPP;
 672                 goto out;
 673         default:
 674                 goto out;
 675         }
 676
 677         if (offset >= 0) {
 678                 if (need_reset_readdir(dfi, offset)) {
 679                         dout("dir_llseek dropping %p content\n", file);
 680                         reset_readdir(dfi);
 681                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 682                         /* for hash offset, we don't know if a forward seek
 683                          * is within same frag */
 684                         dfi->dir_release_count = 0;
 685                         dfi->readdir_cache_idx = -1;
 686                 }
 687
 688                 if (offset != file->f_pos) {
 689                         file->f_pos = offset;
 690                         file->f_version = 0;
 691                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 692                 }
 693                 retval = offset;
 694         }
 695 out:
 696         inode_unlock(inode);
 697         return retval;
 698 }
 699
 700 /*
 701  * Handle lookups for the hidden .snap directory.
 702  */
 703 struct dentry *ceph_handle_snapdir(struct ceph_mds_request *req,
 704                                    struct dentry *dentry)
 705 {
 706         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 707         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_rwsem */
 708
 709         /* .snap dir? */
 710         if (ceph_snap(parent) == CEPH_NOSNAP &&
 711             strcmp(dentry->d_name.name, fsc->mount_options->snapdir_name) == 0) {
 712                 struct dentry *res;
 713                 struct inode *inode = ceph_get_snapdir(parent);
 714
 715                 res = d_splice_alias(inode, dentry);
 716                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p. Spliced dentry %p\n",
 717                      dentry, dentry, inode, res);
 718                 if (res)
 719                         dentry = res;
 720         }
 721         return dentry;
 722 }
 723
 724 /*
 725  * Figure out final result of a lookup/open request.
 726  *
 727  * Mainly, make sure we return the final req->r_dentry (if it already
 728  * existed) in place of the original VFS-provided dentry when they
 729  * differ.
 730  *
 731  * Gracefully handle the case where the MDS replies with -ENOENT and
 732  * no trace (which it may do, at its discretion, e.g., if it doesn't
 733  * care to issue a lease on the negative dentry).
 734  */
 735 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 736                                   struct dentry *dentry, int err)
 737 {
 738         if (err == -ENOENT) {
 739                 /* no trace? */
 740                 err = 0;
 741                 if (!req->r_reply_info.head->is_dentry) {
 742                         dout("ENOENT and no trace, dentry %p inode %p\n",
 743                              dentry, d_inode(dentry));
 744                         if (d_really_is_positive(dentry)) {
 745                                 d_drop(dentry);
 746                                 err = -ENOENT;
 747                         } else {
 748                                 d_add(dentry, NULL);
 749                         }
 750                 }
 751         }
 752         if (err)
 753                 dentry = ERR_PTR(err);
 754         else if (dentry != req->r_dentry)
 755                 dentry = dget(req->r_dentry);   /* we got spliced */
 756         else
 757                 dentry = NULL;
 758         return dentry;
 759 }
 760
 761 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 762 {
 763         return ceph_ino(inode) == CEPH_INO_ROOT &&
 764                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 765 }
 766
 767 /*
 768  * Look up a single dir entry.  If there is a lookup intent, inform
 769  * the MDS so that it gets our 'caps wanted' value in a single op.
 770  */
 771 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 772                                   unsigned int flags)
 773 {
 774         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 775         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 776         struct ceph_mds_request *req;
 777         int op;
 778         int mask;
 779         int err;
 780
 781         dout("lookup %p dentry %p '%pd'\n",
 782              dir, dentry, dentry);
 783
 784         if (dentry->d_name.len > NAME_MAX)
 785                 return ERR_PTR(-ENAMETOOLONG);
 786
 787         if (IS_ENCRYPTED(dir)) {
 788                 bool had_key = fscrypt_has_encryption_key(dir);
 789
 790                 err = fscrypt_prepare_lookup_partial(dir, dentry);
 791                 if (err < 0)
 792                         return ERR_PTR(err);
 793
 794                 /* mark directory as incomplete if it has been unlocked */
 795                 if (!had_key && fscrypt_has_encryption_key(dir))
 796                         ceph_dir_clear_complete(dir);
 797         }
 798
 799         /* can we conclude ENOENT locally? */
 800         if (d_really_is_negative(dentry)) {
 801                 struct ceph_inode_info *ci = ceph_inode(dir);
 802                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 803
 804                 spin_lock(&ci->i_ceph_lock);
 805                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 806                 if (strncmp(dentry->d_name.name,
 807                             fsc->mount_options->snapdir_name,
 808                             dentry->d_name.len) &&
 809                     !is_root_ceph_dentry(dir, dentry) &&
 810                     ceph_test_mount_opt(fsc, DCACHE) &&
 811                     __ceph_dir_is_complete(ci) &&
 812                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 813                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 814                         spin_unlock(&ci->i_ceph_lock);
 815                         dout(" dir %p complete, -ENOENT\n", dir);
 816                         d_add(dentry, NULL);
 817                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 818                         return NULL;
 819                 }
 820                 spin_unlock(&ci->i_ceph_lock);
 821         }
 822
 823         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 824                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 825         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 826         if (IS_ERR(req))
 827                 return ERR_CAST(req);
 828         req->r_dentry = dget(dentry);
 829         req->r_num_caps = 2;
 830
 831         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 832         if (ceph_security_xattr_wanted(dir))
 833                 mask |= CEPH_CAP_XATTR_SHARED;
 834         req->r_args.getattr.mask = cpu_to_le32(mask);
 835
 836         ihold(dir);
 837         req->r_parent = dir;
 838         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 839         err = ceph_mdsc_do_request(mdsc, NULL, req);
 840         if (err == -ENOENT) {
 841                 struct dentry *res;
 842
 843                 res = ceph_handle_snapdir(req, dentry);
 844                 if (IS_ERR(res)) {
 845                         err = PTR_ERR(res);
 846                 } else {
 847                         dentry = res;
 848                         err = 0;
 849                 }
 850         }
 851         dentry = ceph_finish_lookup(req, dentry, err);
 852         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 853         dout("lookup result=%p\n", dentry);
 854         return dentry;
 855 }
 856
 857 /*
 858  * If we do a create but get no trace back from the MDS, follow up with
 859  * a lookup (the VFS expects us to link up the provided dentry).
 860  */
 861 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 862 {
 863         struct dentry *result = ceph_lookup(dir, dentry, 0);
 864
 865         if (result && !IS_ERR(result)) {
 866                 /*
 867                  * We created the item, then did a lookup, and found
 868                  * it was already linked to another inode we already
 869                  * had in our cache (and thus got spliced). To not
 870                  * confuse VFS (especially when inode is a directory),
 871                  * we don't link our dentry to that inode, return an
 872                  * error instead.
 873                  *
 874                  * This event should be rare and it happens only when
 875                  * we talk to old MDS. Recent MDS does not send traceless
 876                  * reply for request that creates new inode.
 877                  */
 878                 d_drop(result);
 879                 return -ESTALE;
 880         }
 881         return PTR_ERR(result);
 882 }
 883
 884 static int ceph_mknod(struct mnt_idmap *idmap, struct inode *dir,
 885                       struct dentry *dentry, umode_t mode, dev_t rdev)
 886 {
 887         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 888         struct ceph_mds_request *req;
 889         struct ceph_acl_sec_ctx as_ctx = {};
 890         int err;
 891
 892         if (ceph_snap(dir) != CEPH_NOSNAP)
 893                 return -EROFS;
 894
 895         err = ceph_wait_on_conflict_unlink(dentry);
 896         if (err)
 897                 return err;
 898
 899         if (ceph_quota_is_max_files_exceeded(dir)) {
 900                 err = -EDQUOT;
 901                 goto out;
 902         }
 903
 904         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 905              dir, dentry, mode, rdev);
 906         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 907         if (IS_ERR(req)) {
 908                 err = PTR_ERR(req);
 909                 goto out;
 910         }
 911
 912         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
 913         if (IS_ERR(req->r_new_inode)) {
 914                 err = PTR_ERR(req->r_new_inode);
 915                 req->r_new_inode = NULL;
 916                 goto out_req;
 917         }
 918
 919         if (S_ISREG(mode) && IS_ENCRYPTED(dir))
 920                 set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
 921
 922         req->r_dentry = dget(dentry);
 923         req->r_num_caps = 2;
 924         req->r_parent = dir;
 925         ihold(dir);
 926         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 927         req->r_args.mknod.mode = cpu_to_le32(mode);
 928         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 929         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
 930                              CEPH_CAP_XATTR_EXCL;
 931         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 932
 933         ceph_as_ctx_to_req(req, &as_ctx);
 934
 935         err = ceph_mdsc_do_request(mdsc, dir, req);
 936         if (!err && !req->r_reply_info.head->is_dentry)
 937                 err = ceph_handle_notrace_create(dir, dentry);
 938 out_req:
 939         ceph_mdsc_put_request(req);
 940 out:
 941         if (!err)
 942                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 943         else
 944                 d_drop(dentry);
 945         ceph_release_acl_sec_ctx(&as_ctx);
 946         return err;
 947 }
 948
 949 static int ceph_create(struct mnt_idmap *idmap, struct inode *dir,
 950                        struct dentry *dentry, umode_t mode, bool excl)
 951 {
 952         return ceph_mknod(idmap, dir, dentry, mode, 0);
 953 }
 954
 955 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
 956 static int prep_encrypted_symlink_target(struct ceph_mds_request *req,
 957                                          const char *dest)
 958 {
 959         int err;
 960         int len = strlen(dest);
 961         struct fscrypt_str osd_link = FSTR_INIT(NULL, 0);
 962
 963         err = fscrypt_prepare_symlink(req->r_parent, dest, len, PATH_MAX,
 964                                       &osd_link);
 965         if (err)
 966                 goto out;
 967
 968         err = fscrypt_encrypt_symlink(req->r_new_inode, dest, len, &osd_link);
 969         if (err)
 970                 goto out;
 971
 972         req->r_path2 = kmalloc(CEPH_BASE64_CHARS(osd_link.len) + 1, GFP_KERNEL);
 973         if (!req->r_path2) {
 974                 err = -ENOMEM;
 975                 goto out;
 976         }
 977
 978         len = ceph_base64_encode(osd_link.name, osd_link.len, req->r_path2);
 979         req->r_path2[len] = '\0';
 980 out:
 981         fscrypt_fname_free_buffer(&osd_link);
 982         return err;
 983 }
 984 #else
 985 static int prep_encrypted_symlink_target(struct ceph_mds_request *req,
 986                                          const char *dest)
 987 {
 988         return -EOPNOTSUPP;
 989 }
 990 #endif
 991
 992 static int ceph_symlink(struct mnt_idmap *idmap, struct inode *dir,
 993                         struct dentry *dentry, const char *dest)
 994 {
 995         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 996         struct ceph_mds_request *req;
 997         struct ceph_acl_sec_ctx as_ctx = {};
 998         umode_t mode = S_IFLNK | 0777;
 999         int err;
1000
1001         if (ceph_snap(dir) != CEPH_NOSNAP)
1002                 return -EROFS;
1003
1004         err = ceph_wait_on_conflict_unlink(dentry);
1005         if (err)
1006                 return err;
1007
1008         if (ceph_quota_is_max_files_exceeded(dir)) {
1009                 err = -EDQUOT;
1010                 goto out;
1011         }
1012
1013         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
1014         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
1015         if (IS_ERR(req)) {
1016                 err = PTR_ERR(req);
1017                 goto out;
1018         }
1019
1020         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
1021         if (IS_ERR(req->r_new_inode)) {
1022                 err = PTR_ERR(req->r_new_inode);
1023                 req->r_new_inode = NULL;
1024                 goto out_req;
1025         }
1026
1027         req->r_parent = dir;
1028         ihold(dir);
1029
1030         if (IS_ENCRYPTED(req->r_new_inode)) {
1031                 err = prep_encrypted_symlink_target(req, dest);
1032                 if (err)
1033                         goto out_req;
1034         } else {
1035                 req->r_path2 = kstrdup(dest, GFP_KERNEL);
1036                 if (!req->r_path2) {
1037                         err = -ENOMEM;
1038                         goto out_req;
1039                 }
1040         }
1041
1042         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1043         req->r_dentry = dget(dentry);
1044         req->r_num_caps = 2;
1045         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
1046                              CEPH_CAP_XATTR_EXCL;
1047         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1048
1049         ceph_as_ctx_to_req(req, &as_ctx);
1050
1051         err = ceph_mdsc_do_request(mdsc, dir, req);
1052         if (!err && !req->r_reply_info.head->is_dentry)
1053                 err = ceph_handle_notrace_create(dir, dentry);
1054 out_req:
1055         ceph_mdsc_put_request(req);
1056 out:
1057         if (err)
1058                 d_drop(dentry);
1059         ceph_release_acl_sec_ctx(&as_ctx);
1060         return err;
1061 }
1062
1063 static int ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1064                       struct dentry *dentry, umode_t mode)
1065 {
1066         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1067         struct ceph_mds_request *req;
1068         struct ceph_acl_sec_ctx as_ctx = {};
1069         int err;
1070         int op;
1071
1072         err = ceph_wait_on_conflict_unlink(dentry);
1073         if (err)
1074                 return err;
1075
1076         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1077                 /* mkdir .snap/foo is a MKSNAP */
1078                 op = CEPH_MDS_OP_MKSNAP;
1079                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
1080                      dentry, dentry);
1081         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1082                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
1083                 op = CEPH_MDS_OP_MKDIR;
1084         } else {
1085                 err = -EROFS;
1086                 goto out;
1087         }
1088
1089         if (op == CEPH_MDS_OP_MKDIR &&
1090             ceph_quota_is_max_files_exceeded(dir)) {
1091                 err = -EDQUOT;
1092                 goto out;
1093         }
1094         if ((op == CEPH_MDS_OP_MKSNAP) && IS_ENCRYPTED(dir) &&
1095             !fscrypt_has_encryption_key(dir)) {
1096                 err = -ENOKEY;
1097                 goto out;
1098         }
1099
1100
1101         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1102         if (IS_ERR(req)) {
1103                 err = PTR_ERR(req);
1104                 goto out;
1105         }
1106
1107         mode |= S_IFDIR;
1108         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
1109         if (IS_ERR(req->r_new_inode)) {
1110                 err = PTR_ERR(req->r_new_inode);
1111                 req->r_new_inode = NULL;
1112                 goto out_req;
1113         }
1114
1115         req->r_dentry = dget(dentry);
1116         req->r_num_caps = 2;
1117         req->r_parent = dir;
1118         ihold(dir);
1119         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1120         req->r_args.mkdir.mode = cpu_to_le32(mode);
1121         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
1122                              CEPH_CAP_XATTR_EXCL;
1123         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1124
1125         ceph_as_ctx_to_req(req, &as_ctx);
1126
1127         err = ceph_mdsc_do_request(mdsc, dir, req);
1128         if (!err &&
1129             !req->r_reply_info.head->is_target &&
1130             !req->r_reply_info.head->is_dentry)
1131                 err = ceph_handle_notrace_create(dir, dentry);
1132 out_req:
1133         ceph_mdsc_put_request(req);
1134 out:
1135         if (!err)
1136                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1137         else
1138                 d_drop(dentry);
1139         ceph_release_acl_sec_ctx(&as_ctx);
1140         return err;
1141 }
1142
1143 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1144                      struct dentry *dentry)
1145 {
1146         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1147         struct ceph_mds_request *req;
1148         int err;
1149
1150         if (dentry->d_flags & DCACHE_DISCONNECTED)
1151                 return -EINVAL;
1152
1153         err = ceph_wait_on_conflict_unlink(dentry);
1154         if (err)
1155                 return err;
1156
1157         if (ceph_snap(dir) != CEPH_NOSNAP)
1158                 return -EROFS;
1159
1160         err = fscrypt_prepare_link(old_dentry, dir, dentry);
1161         if (err)
1162                 return err;
1163
1164         dout("link in dir %p %llx.%llx old_dentry %p:'%pd' dentry %p:'%pd'\n",
1165              dir, ceph_vinop(dir), old_dentry, old_dentry, dentry, dentry);
1166         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1167         if (IS_ERR(req)) {
1168                 d_drop(dentry);
1169                 return PTR_ERR(req);
1170         }
1171         req->r_dentry = dget(dentry);
1172         req->r_num_caps = 2;
1173         req->r_old_dentry = dget(old_dentry);
1174         /*
1175          * The old_dentry maybe a DCACHE_DISCONNECTED dentry, then we
1176          * will just pass the ino# to MDSs.
1177          */
1178         if (old_dentry->d_flags & DCACHE_DISCONNECTED)
1179                 req->r_ino2 = ceph_vino(d_inode(old_dentry));
1180         req->r_parent = dir;
1181         ihold(dir);
1182         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1183         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1184         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1185         /* release LINK_SHARED on source inode (mds will lock it) */
1186         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1187         err = ceph_mdsc_do_request(mdsc, dir, req);
1188         if (err) {
1189                 d_drop(dentry);
1190         } else if (!req->r_reply_info.head->is_dentry) {
1191                 ihold(d_inode(old_dentry));
1192                 d_instantiate(dentry, d_inode(old_dentry));
1193         }
1194         ceph_mdsc_put_request(req);
1195         return err;
1196 }
1197
1198 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1199                                  struct ceph_mds_request *req)
1200 {
1201         struct dentry *dentry = req->r_dentry;
1202         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1203         struct ceph_dentry_info *di = ceph_dentry(dentry);
1204         int result = req->r_err ? req->r_err :
1205                         le32_to_cpu(req->r_reply_info.head->result);
1206
1207         if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
1208                 pr_warn("%s dentry %p:%pd async unlink bit is not set\n",
1209                         __func__, dentry, dentry);
1210
1211         spin_lock(&fsc->async_unlink_conflict_lock);
1212         hash_del_rcu(&di->hnode);
1213         spin_unlock(&fsc->async_unlink_conflict_lock);
1214
1215         spin_lock(&dentry->d_lock);
1216         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1217         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT);
1218         spin_unlock(&dentry->d_lock);
1219
1220         synchronize_rcu();
1221
1222         if (result == -EJUKEBOX)
1223                 goto out;
1224
1225         /* If op failed, mark everyone involved for errors */
1226         if (result) {
1227                 int pathlen = 0;
1228                 u64 base = 0;
1229                 char *path = ceph_mdsc_build_path(dentry, &pathlen,
1230                                                   &base, 0);
1231
1232                 /* mark error on parent + clear complete */
1233                 mapping_set_error(req->r_parent->i_mapping, result);
1234                 ceph_dir_clear_complete(req->r_parent);
1235
1236                 /* drop the dentry -- we don't know its status */
1237                 if (!d_unhashed(dentry))
1238                         d_drop(dentry);
1239
1240                 /* mark inode itself for an error (since metadata is bogus) */
1241                 mapping_set_error(req->r_old_inode->i_mapping, result);
1242
1243                 pr_warn("async unlink failure path=(%llx)%s result=%d!\n",
1244                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1245                 ceph_mdsc_free_path(path, pathlen);
1246         }
1247 out:
1248         iput(req->r_old_inode);
1249         ceph_mdsc_release_dir_caps(req);
1250 }
1251
1252 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1253 {
1254         struct ceph_inode_info *ci = ceph_inode(dir);
1255         struct ceph_dentry_info *di;
1256         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1257
1258         spin_lock(&ci->i_ceph_lock);
1259         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1260                 ceph_take_cap_refs(ci, want, false);
1261                 got = want;
1262         }
1263         spin_unlock(&ci->i_ceph_lock);
1264
1265         /* If we didn't get anything, return 0 */
1266         if (!got)
1267                 return 0;
1268
1269         spin_lock(&dentry->d_lock);
1270         di = ceph_dentry(dentry);
1271         /*
1272          * - We are holding Fx, which implies Fs caps.
1273          * - Only support async unlink for primary linkage
1274          */
1275         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1276             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1277                 want = 0;
1278         spin_unlock(&dentry->d_lock);
1279
1280         /* Do we still want what we've got? */
1281         if (want == got)
1282                 return got;
1283
1284         ceph_put_cap_refs(ci, got);
1285         return 0;
1286 }
1287
1288 /*
1289  * rmdir and unlink are differ only by the metadata op code
1290  */
1291 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1292 {
1293         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1294         struct ceph_mds_client *mdsc = fsc->mdsc;
1295         struct inode *inode = d_inode(dentry);
1296         struct ceph_mds_request *req;
1297         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1298         int err = -EROFS;
1299         int op;
1300
1301         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1302                 /* rmdir .snap/foo is RMSNAP */
1303                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1304                 op = CEPH_MDS_OP_RMSNAP;
1305         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1306                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1307                      dir, dentry, inode);
1308                 op = d_is_dir(dentry) ?
1309                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1310         } else
1311                 goto out;
1312 retry:
1313         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1314         if (IS_ERR(req)) {
1315                 err = PTR_ERR(req);
1316                 goto out;
1317         }
1318         req->r_dentry = dget(dentry);
1319         req->r_num_caps = 2;
1320         req->r_parent = dir;
1321         ihold(dir);
1322         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1323         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1324         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1325
1326         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1327             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1328                 struct ceph_dentry_info *di = ceph_dentry(dentry);
1329
1330                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1331                      dentry->d_name.len, dentry->d_name.name,
1332                      ceph_cap_string(req->r_dir_caps));
1333                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1334                 req->r_callback = ceph_async_unlink_cb;
1335                 req->r_old_inode = d_inode(dentry);
1336                 ihold(req->r_old_inode);
1337
1338                 spin_lock(&dentry->d_lock);
1339                 di->flags |= CEPH_DENTRY_ASYNC_UNLINK;
1340                 spin_unlock(&dentry->d_lock);
1341
1342                 spin_lock(&fsc->async_unlink_conflict_lock);
1343                 hash_add_rcu(fsc->async_unlink_conflict, &di->hnode,
1344                              dentry->d_name.hash);
1345                 spin_unlock(&fsc->async_unlink_conflict_lock);
1346
1347                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1348                 if (!err) {
1349                         /*
1350                          * We have enough caps, so we assume that the unlink
1351                          * will succeed. Fix up the target inode and dcache.
1352                          */
1353                         drop_nlink(inode);
1354                         d_delete(dentry);
1355                 } else {
1356                         spin_lock(&fsc->async_unlink_conflict_lock);
1357                         hash_del_rcu(&di->hnode);
1358                         spin_unlock(&fsc->async_unlink_conflict_lock);
1359
1360                         spin_lock(&dentry->d_lock);
1361                         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1362                         spin_unlock(&dentry->d_lock);
1363
1364                         if (err == -EJUKEBOX) {
1365                                 try_async = false;
1366                                 ceph_mdsc_put_request(req);
1367                                 goto retry;
1368                         }
1369                 }
1370         } else {
1371                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1372                 err = ceph_mdsc_do_request(mdsc, dir, req);
1373                 if (!err && !req->r_reply_info.head->is_dentry)
1374                         d_delete(dentry);
1375         }
1376
1377         ceph_mdsc_put_request(req);
1378 out:
1379         return err;
1380 }
1381
1382 static int ceph_rename(struct mnt_idmap *idmap, struct inode *old_dir,
1383                        struct dentry *old_dentry, struct inode *new_dir,
1384                        struct dentry *new_dentry, unsigned int flags)
1385 {
1386         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1387         struct ceph_mds_request *req;
1388         int op = CEPH_MDS_OP_RENAME;
1389         int err;
1390
1391         if (flags)
1392                 return -EINVAL;
1393
1394         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1395                 return -EXDEV;
1396         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1397                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1398                         op = CEPH_MDS_OP_RENAMESNAP;
1399                 else
1400                         return -EROFS;
1401         }
1402         /* don't allow cross-quota renames */
1403         if ((old_dir != new_dir) &&
1404             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1405                 return -EXDEV;
1406
1407         err = ceph_wait_on_conflict_unlink(new_dentry);
1408         if (err)
1409                 return err;
1410
1411         err = fscrypt_prepare_rename(old_dir, old_dentry, new_dir, new_dentry,
1412                                      flags);
1413         if (err)
1414                 return err;
1415
1416         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1417              old_dir, old_dentry, new_dir, new_dentry);
1418         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1419         if (IS_ERR(req))
1420                 return PTR_ERR(req);
1421         ihold(old_dir);
1422         req->r_dentry = dget(new_dentry);
1423         req->r_num_caps = 2;
1424         req->r_old_dentry = dget(old_dentry);
1425         req->r_old_dentry_dir = old_dir;
1426         req->r_parent = new_dir;
1427         ihold(new_dir);
1428         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1429         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1430         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1431         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1432         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1433         /* release LINK_RDCACHE on source inode (mds will lock it) */
1434         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1435         if (d_really_is_positive(new_dentry)) {
1436                 req->r_inode_drop =
1437                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1438         }
1439         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1440         if (!err && !req->r_reply_info.head->is_dentry) {
1441                 /*
1442                  * Normally d_move() is done by fill_trace (called by
1443                  * do_request, above).  If there is no trace, we need
1444                  * to do it here.
1445                  */
1446                 d_move(old_dentry, new_dentry);
1447         }
1448         ceph_mdsc_put_request(req);
1449         return err;
1450 }
1451
1452 /*
1453  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1454  * Leases at front of the list will expire first. (Assume all leases have
1455  * similar duration)
1456  *
1457  * Called under dentry->d_lock.
1458  */
1459 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1460 {
1461         struct dentry *dn = di->dentry;
1462         struct ceph_mds_client *mdsc;
1463
1464         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1465
1466         di->flags |= CEPH_DENTRY_LEASE_LIST;
1467         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1468                 di->flags |= CEPH_DENTRY_REFERENCED;
1469                 return;
1470         }
1471
1472         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1473         spin_lock(&mdsc->dentry_list_lock);
1474         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1475         spin_unlock(&mdsc->dentry_list_lock);
1476 }
1477
1478 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1479                                      struct ceph_dentry_info *di)
1480 {
1481         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1482         di->lease_gen = 0;
1483         di->time = jiffies;
1484         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1485 }
1486
1487 /*
1488  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1489  * list if it's not in the list, otherwise set 'referenced' flag.
1490  *
1491  * Called under dentry->d_lock.
1492  */
1493 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1494 {
1495         struct dentry *dn = di->dentry;
1496         struct ceph_mds_client *mdsc;
1497
1498         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1499              di, dn, dn, di->offset);
1500
1501         if (!list_empty(&di->lease_list)) {
1502                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1503                         /* don't remove dentry from dentry lease list
1504                          * if its lease is valid */
1505                         if (__dentry_lease_is_valid(di))
1506                                 return;
1507                 } else {
1508                         di->flags |= CEPH_DENTRY_REFERENCED;
1509                         return;
1510                 }
1511         }
1512
1513         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1514                 di->flags |= CEPH_DENTRY_REFERENCED;
1515                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1516                 return;
1517         }
1518
1519         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1520         spin_lock(&mdsc->dentry_list_lock);
1521         __dentry_dir_lease_touch(mdsc, di),
1522         spin_unlock(&mdsc->dentry_list_lock);
1523 }
1524
1525 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1526 {
1527         struct ceph_mds_client *mdsc;
1528         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1529                 return;
1530         if (list_empty(&di->lease_list))
1531                 return;
1532
1533         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1534         spin_lock(&mdsc->dentry_list_lock);
1535         list_del_init(&di->lease_list);
1536         spin_unlock(&mdsc->dentry_list_lock);
1537 }
1538
1539 enum {
1540         KEEP    = 0,
1541         DELETE  = 1,
1542         TOUCH   = 2,
1543         STOP    = 4,
1544 };
1545
1546 struct ceph_lease_walk_control {
1547         bool dir_lease;
1548         bool expire_dir_lease;
1549         unsigned long nr_to_scan;
1550         unsigned long dir_lease_ttl;
1551 };
1552
1553 static unsigned long
1554 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1555                      struct ceph_lease_walk_control *lwc,
1556                      int (*check)(struct dentry*, void*))
1557 {
1558         struct ceph_dentry_info *di, *tmp;
1559         struct dentry *dentry, *last = NULL;
1560         struct list_head* list;
1561         LIST_HEAD(dispose);
1562         unsigned long freed = 0;
1563         int ret = 0;
1564
1565         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1566         spin_lock(&mdsc->dentry_list_lock);
1567         list_for_each_entry_safe(di, tmp, list, lease_list) {
1568                 if (!lwc->nr_to_scan)
1569                         break;
1570                 --lwc->nr_to_scan;
1571
1572                 dentry = di->dentry;
1573                 if (last == dentry)
1574                         break;
1575
1576                 if (!spin_trylock(&dentry->d_lock))
1577                         continue;
1578
1579                 if (__lockref_is_dead(&dentry->d_lockref)) {
1580                         list_del_init(&di->lease_list);
1581                         goto next;
1582                 }
1583
1584                 ret = check(dentry, lwc);
1585                 if (ret & TOUCH) {
1586                         /* move it into tail of dir lease list */
1587                         __dentry_dir_lease_touch(mdsc, di);
1588                         if (!last)
1589                                 last = dentry;
1590                 }
1591                 if (ret & DELETE) {
1592                         /* stale lease */
1593                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1594                         if (dentry->d_lockref.count > 0) {
1595                                 /* update_dentry_lease() will re-add
1596                                  * it to lease list, or
1597                                  * ceph_d_delete() will return 1 when
1598                                  * last reference is dropped */
1599                                 list_del_init(&di->lease_list);
1600                         } else {
1601                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1602                                 list_move_tail(&di->lease_list, &dispose);
1603                                 dget_dlock(dentry);
1604                         }
1605                 }
1606 next:
1607                 spin_unlock(&dentry->d_lock);
1608                 if (ret & STOP)
1609                         break;
1610         }
1611         spin_unlock(&mdsc->dentry_list_lock);
1612
1613         while (!list_empty(&dispose)) {
1614                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1615                                       lease_list);
1616                 dentry = di->dentry;
1617                 spin_lock(&dentry->d_lock);
1618
1619                 list_del_init(&di->lease_list);
1620                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1621                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1622                         spin_lock(&mdsc->dentry_list_lock);
1623                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1624                                 list_add_tail(&di->lease_list,
1625                                               &mdsc->dentry_leases);
1626                         } else {
1627                                 __dentry_dir_lease_touch(mdsc, di);
1628                         }
1629                         spin_unlock(&mdsc->dentry_list_lock);
1630                 } else {
1631                         freed++;
1632                 }
1633
1634                 spin_unlock(&dentry->d_lock);
1635                 /* ceph_d_delete() does the trick */
1636                 dput(dentry);
1637         }
1638         return freed;
1639 }
1640
1641 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1642 {
1643         struct ceph_dentry_info *di = ceph_dentry(dentry);
1644         int ret;
1645
1646         if (__dentry_lease_is_valid(di))
1647                 return STOP;
1648         ret = __dir_lease_try_check(dentry);
1649         if (ret == -EBUSY)
1650                 return KEEP;
1651         if (ret > 0)
1652                 return TOUCH;
1653         return DELETE;
1654 }
1655
1656 static int __dir_lease_check(struct dentry *dentry, void *arg)
1657 {
1658         struct ceph_lease_walk_control *lwc = arg;
1659         struct ceph_dentry_info *di = ceph_dentry(dentry);
1660
1661         int ret = __dir_lease_try_check(dentry);
1662         if (ret == -EBUSY)
1663                 return KEEP;
1664         if (ret > 0) {
1665                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1666                         return STOP;
1667                 /* Move dentry to tail of dir lease list if we don't want
1668                  * to delete it. So dentries in the list are checked in a
1669                  * round robin manner */
1670                 if (!lwc->expire_dir_lease)
1671                         return TOUCH;
1672                 if (dentry->d_lockref.count > 0 ||
1673                     (di->flags & CEPH_DENTRY_REFERENCED))
1674                         return TOUCH;
1675                 /* invalidate dir lease */
1676                 di->lease_shared_gen = 0;
1677         }
1678         return DELETE;
1679 }
1680
1681 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1682 {
1683         struct ceph_lease_walk_control lwc;
1684         unsigned long count;
1685         unsigned long freed;
1686
1687         spin_lock(&mdsc->caps_list_lock);
1688         if (mdsc->caps_use_max > 0 &&
1689             mdsc->caps_use_count > mdsc->caps_use_max)
1690                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1691         else
1692                 count = 0;
1693         spin_unlock(&mdsc->caps_list_lock);
1694
1695         lwc.dir_lease = false;
1696         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1697         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1698         if (!lwc.nr_to_scan) /* more invalid leases */
1699                 return -EAGAIN;
1700
1701         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1702                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1703
1704         lwc.dir_lease = true;
1705         lwc.expire_dir_lease = freed < count;
1706         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1707         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1708         if (!lwc.nr_to_scan) /* more to check */
1709                 return -EAGAIN;
1710
1711         return freed > 0 ? 1 : 0;
1712 }
1713
1714 /*
1715  * Ensure a dentry lease will no longer revalidate.
1716  */
1717 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1718 {
1719         struct ceph_dentry_info *di = ceph_dentry(dentry);
1720         spin_lock(&dentry->d_lock);
1721         di->time = jiffies;
1722         di->lease_shared_gen = 0;
1723         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1724         __dentry_lease_unlist(di);
1725         spin_unlock(&dentry->d_lock);
1726 }
1727
1728 /*
1729  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1730  * renew if the least is more than half up.
1731  */
1732 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1733 {
1734         struct ceph_mds_session *session;
1735
1736         if (!di->lease_gen)
1737                 return false;
1738
1739         session = di->lease_session;
1740         if (session) {
1741                 u32 gen;
1742                 unsigned long ttl;
1743
1744                 gen = atomic_read(&session->s_cap_gen);
1745                 ttl = session->s_cap_ttl;
1746
1747                 if (di->lease_gen == gen &&
1748                     time_before(jiffies, ttl) &&
1749                     time_before(jiffies, di->time))
1750                         return true;
1751         }
1752         di->lease_gen = 0;
1753         return false;
1754 }
1755
1756 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1757 {
1758         struct ceph_dentry_info *di;
1759         struct ceph_mds_session *session = NULL;
1760         u32 seq = 0;
1761         int valid = 0;
1762
1763         spin_lock(&dentry->d_lock);
1764         di = ceph_dentry(dentry);
1765         if (di && __dentry_lease_is_valid(di)) {
1766                 valid = 1;
1767
1768                 if (di->lease_renew_after &&
1769                     time_after(jiffies, di->lease_renew_after)) {
1770                         /*
1771                          * We should renew. If we're in RCU walk mode
1772                          * though, we can't do that so just return
1773                          * -ECHILD.
1774                          */
1775                         if (flags & LOOKUP_RCU) {
1776                                 valid = -ECHILD;
1777                         } else {
1778                                 session = ceph_get_mds_session(di->lease_session);
1779                                 seq = di->lease_seq;
1780                                 di->lease_renew_after = 0;
1781                                 di->lease_renew_from = jiffies;
1782                         }
1783                 }
1784         }
1785         spin_unlock(&dentry->d_lock);
1786
1787         if (session) {
1788                 ceph_mdsc_lease_send_msg(session, dentry,
1789                                          CEPH_MDS_LEASE_RENEW, seq);
1790                 ceph_put_mds_session(session);
1791         }
1792         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1793         return valid;
1794 }
1795
1796 /*
1797  * Called under dentry->d_lock.
1798  */
1799 static int __dir_lease_try_check(const struct dentry *dentry)
1800 {
1801         struct ceph_dentry_info *di = ceph_dentry(dentry);
1802         struct inode *dir;
1803         struct ceph_inode_info *ci;
1804         int valid = 0;
1805
1806         if (!di->lease_shared_gen)
1807                 return 0;
1808         if (IS_ROOT(dentry))
1809                 return 0;
1810
1811         dir = d_inode(dentry->d_parent);
1812         ci = ceph_inode(dir);
1813
1814         if (spin_trylock(&ci->i_ceph_lock)) {
1815                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1816                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1817                         valid = 1;
1818                 spin_unlock(&ci->i_ceph_lock);
1819         } else {
1820                 valid = -EBUSY;
1821         }
1822
1823         if (!valid)
1824                 di->lease_shared_gen = 0;
1825         return valid;
1826 }
1827
1828 /*
1829  * Check if directory-wide content lease/cap is valid.
1830  */
1831 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1832                               struct ceph_mds_client *mdsc)
1833 {
1834         struct ceph_inode_info *ci = ceph_inode(dir);
1835         int valid;
1836         int shared_gen;
1837
1838         spin_lock(&ci->i_ceph_lock);
1839         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1840         if (valid) {
1841                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1842                 shared_gen = atomic_read(&ci->i_shared_gen);
1843         }
1844         spin_unlock(&ci->i_ceph_lock);
1845         if (valid) {
1846                 struct ceph_dentry_info *di;
1847                 spin_lock(&dentry->d_lock);
1848                 di = ceph_dentry(dentry);
1849                 if (dir == d_inode(dentry->d_parent) &&
1850                     di && di->lease_shared_gen == shared_gen)
1851                         __ceph_dentry_dir_lease_touch(di);
1852                 else
1853                         valid = 0;
1854                 spin_unlock(&dentry->d_lock);
1855         }
1856         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1857              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1858         return valid;
1859 }
1860
1861 /*
1862  * Check if cached dentry can be trusted.
1863  */
1864 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1865 {
1866         int valid = 0;
1867         struct dentry *parent;
1868         struct inode *dir, *inode;
1869         struct ceph_mds_client *mdsc;
1870
1871         valid = fscrypt_d_revalidate(dentry, flags);
1872         if (valid <= 0)
1873                 return valid;
1874
1875         if (flags & LOOKUP_RCU) {
1876                 parent = READ_ONCE(dentry->d_parent);
1877                 dir = d_inode_rcu(parent);
1878                 if (!dir)
1879                         return -ECHILD;
1880                 inode = d_inode_rcu(dentry);
1881         } else {
1882                 parent = dget_parent(dentry);
1883                 dir = d_inode(parent);
1884                 inode = d_inode(dentry);
1885         }
1886
1887         dout("d_revalidate %p '%pd' inode %p offset 0x%llx nokey %d\n", dentry,
1888              dentry, inode, ceph_dentry(dentry)->offset,
1889              !!(dentry->d_flags & DCACHE_NOKEY_NAME));
1890
1891         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1892
1893         /* always trust cached snapped dentries, snapdir dentry */
1894         if (ceph_snap(dir) != CEPH_NOSNAP) {
1895                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1896                      dentry, inode);
1897                 valid = 1;
1898         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1899                 valid = 1;
1900         } else {
1901                 valid = dentry_lease_is_valid(dentry, flags);
1902                 if (valid == -ECHILD)
1903                         return valid;
1904                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1905                         if (inode)
1906                                 valid = ceph_is_any_caps(inode);
1907                         else
1908                                 valid = 1;
1909                 }
1910         }
1911
1912         if (!valid) {
1913                 struct ceph_mds_request *req;
1914                 int op, err;
1915                 u32 mask;
1916
1917                 if (flags & LOOKUP_RCU)
1918                         return -ECHILD;
1919
1920                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1921
1922                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1923                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1924                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1925                 if (!IS_ERR(req)) {
1926                         req->r_dentry = dget(dentry);
1927                         req->r_num_caps = 2;
1928                         req->r_parent = dir;
1929                         ihold(dir);
1930
1931                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1932                         if (ceph_security_xattr_wanted(dir))
1933                                 mask |= CEPH_CAP_XATTR_SHARED;
1934                         req->r_args.getattr.mask = cpu_to_le32(mask);
1935
1936                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1937                         switch (err) {
1938                         case 0:
1939                                 if (d_really_is_positive(dentry) &&
1940                                     d_inode(dentry) == req->r_target_inode)
1941                                         valid = 1;
1942                                 break;
1943                         case -ENOENT:
1944                                 if (d_really_is_negative(dentry))
1945                                         valid = 1;
1946                                 fallthrough;
1947                         default:
1948                                 break;
1949                         }
1950                         ceph_mdsc_put_request(req);
1951                         dout("d_revalidate %p lookup result=%d\n",
1952                              dentry, err);
1953                 }
1954         } else {
1955                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1956         }
1957
1958         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1959         if (!valid)
1960                 ceph_dir_clear_complete(dir);
1961
1962         if (!(flags & LOOKUP_RCU))
1963                 dput(parent);
1964         return valid;
1965 }
1966
1967 /*
1968  * Delete unused dentry that doesn't have valid lease
1969  *
1970  * Called under dentry->d_lock.
1971  */
1972 static int ceph_d_delete(const struct dentry *dentry)
1973 {
1974         struct ceph_dentry_info *di;
1975
1976         /* won't release caps */
1977         if (d_really_is_negative(dentry))
1978                 return 0;
1979         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1980                 return 0;
1981         /* vaild lease? */
1982         di = ceph_dentry(dentry);
1983         if (di) {
1984                 if (__dentry_lease_is_valid(di))
1985                         return 0;
1986                 if (__dir_lease_try_check(dentry))
1987                         return 0;
1988         }
1989         return 1;
1990 }
1991
1992 /*
1993  * Release our ceph_dentry_info.
1994  */
1995 static void ceph_d_release(struct dentry *dentry)
1996 {
1997         struct ceph_dentry_info *di = ceph_dentry(dentry);
1998         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1999
2000         dout("d_release %p\n", dentry);
2001
2002         atomic64_dec(&fsc->mdsc->metric.total_dentries);
2003
2004         spin_lock(&dentry->d_lock);
2005         __dentry_lease_unlist(di);
2006         dentry->d_fsdata = NULL;
2007         spin_unlock(&dentry->d_lock);
2008
2009         ceph_put_mds_session(di->lease_session);
2010         kmem_cache_free(ceph_dentry_cachep, di);
2011 }
2012
2013 /*
2014  * When the VFS prunes a dentry from the cache, we need to clear the
2015  * complete flag on the parent directory.
2016  *
2017  * Called under dentry->d_lock.
2018  */
2019 static void ceph_d_prune(struct dentry *dentry)
2020 {
2021         struct ceph_inode_info *dir_ci;
2022         struct ceph_dentry_info *di;
2023
2024         dout("ceph_d_prune %pd %p\n", dentry, dentry);
2025
2026         /* do we have a valid parent? */
2027         if (IS_ROOT(dentry))
2028                 return;
2029
2030         /* we hold d_lock, so d_parent is stable */
2031         dir_ci = ceph_inode(d_inode(dentry->d_parent));
2032         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
2033                 return;
2034
2035         /* who calls d_delete() should also disable dcache readdir */
2036         if (d_really_is_negative(dentry))
2037                 return;
2038
2039         /* d_fsdata does not get cleared until d_release */
2040         if (!d_unhashed(dentry)) {
2041                 __ceph_dir_clear_complete(dir_ci);
2042                 return;
2043         }
2044
2045         /* Disable dcache readdir just in case that someone called d_drop()
2046          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
2047          * properly (dcache readdir is still enabled) */
2048         di = ceph_dentry(dentry);
2049         if (di->offset > 0 &&
2050             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
2051                 __ceph_dir_clear_ordered(dir_ci);
2052 }
2053
2054 /*
2055  * read() on a dir.  This weird interface hack only works if mounted
2056  * with '-o dirstat'.
2057  */
2058 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
2059                              loff_t *ppos)
2060 {
2061         struct ceph_dir_file_info *dfi = file->private_data;
2062         struct inode *inode = file_inode(file);
2063         struct ceph_inode_info *ci = ceph_inode(inode);
2064         int left;
2065         const int bufsize = 1024;
2066
2067         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
2068                 return -EISDIR;
2069
2070         if (!dfi->dir_info) {
2071                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
2072                 if (!dfi->dir_info)
2073                         return -ENOMEM;
2074                 dfi->dir_info_len =
2075                         snprintf(dfi->dir_info, bufsize,
2076                                 "entries:   %20lld\n"
2077                                 " files:    %20lld\n"
2078                                 " subdirs:  %20lld\n"
2079                                 "rentries:  %20lld\n"
2080                                 " rfiles:   %20lld\n"
2081                                 " rsubdirs: %20lld\n"
2082                                 "rbytes:    %20lld\n"
2083                                 "rctime:    %10lld.%09ld\n",
2084                                 ci->i_files + ci->i_subdirs,
2085                                 ci->i_files,
2086                                 ci->i_subdirs,
2087                                 ci->i_rfiles + ci->i_rsubdirs,
2088                                 ci->i_rfiles,
2089                                 ci->i_rsubdirs,
2090                                 ci->i_rbytes,
2091                                 ci->i_rctime.tv_sec,
2092                                 ci->i_rctime.tv_nsec);
2093         }
2094
2095         if (*ppos >= dfi->dir_info_len)
2096                 return 0;
2097         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
2098         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
2099         if (left == size)
2100                 return -EFAULT;
2101         *ppos += (size - left);
2102         return size - left;
2103 }
2104
2105
2106
2107 /*
2108  * Return name hash for a given dentry.  This is dependent on
2109  * the parent directory's hash function.
2110  */
2111 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
2112 {
2113         struct ceph_inode_info *dci = ceph_inode(dir);
2114         unsigned hash;
2115
2116         switch (dci->i_dir_layout.dl_dir_hash) {
2117         case 0: /* for backward compat */
2118         case CEPH_STR_HASH_LINUX:
2119                 return dn->d_name.hash;
2120
2121         default:
2122                 spin_lock(&dn->d_lock);
2123                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
2124                                      dn->d_name.name, dn->d_name.len);
2125                 spin_unlock(&dn->d_lock);
2126                 return hash;
2127         }
2128 }
2129
2130 WRAP_DIR_ITER(ceph_readdir) // FIXME!
2131 const struct file_operations ceph_dir_fops = {
2132         .read = ceph_read_dir,
2133         .iterate_shared = shared_ceph_readdir,
2134         .llseek = ceph_dir_llseek,
2135         .open = ceph_open,
2136         .release = ceph_release,
2137         .unlocked_ioctl = ceph_ioctl,
2138         .compat_ioctl = compat_ptr_ioctl,
2139         .fsync = ceph_fsync,
2140         .lock = ceph_lock,
2141         .flock = ceph_flock,
2142 };
2143
2144 const struct file_operations ceph_snapdir_fops = {
2145         .iterate_shared = shared_ceph_readdir,
2146         .llseek = ceph_dir_llseek,
2147         .open = ceph_open,
2148         .release = ceph_release,
2149 };
2150
2151 const struct inode_operations ceph_dir_iops = {
2152         .lookup = ceph_lookup,
2153         .permission = ceph_permission,
2154         .getattr = ceph_getattr,
2155         .setattr = ceph_setattr,
2156         .listxattr = ceph_listxattr,
2157         .get_inode_acl = ceph_get_acl,
2158         .set_acl = ceph_set_acl,
2159         .mknod = ceph_mknod,
2160         .symlink = ceph_symlink,
2161         .mkdir = ceph_mkdir,
2162         .link = ceph_link,
2163         .unlink = ceph_unlink,
2164         .rmdir = ceph_unlink,
2165         .rename = ceph_rename,
2166         .create = ceph_create,
2167         .atomic_open = ceph_atomic_open,
2168 };
2169
2170 const struct inode_operations ceph_snapdir_iops = {
2171         .lookup = ceph_lookup,
2172         .permission = ceph_permission,
2173         .getattr = ceph_getattr,
2174         .mkdir = ceph_mkdir,
2175         .rmdir = ceph_unlink,
2176         .rename = ceph_rename,
2177 };
2178
2179 const struct dentry_operations ceph_dentry_ops = {
2180         .d_revalidate = ceph_d_revalidate,
2181         .d_delete = ceph_d_delete,
2182         .d_release = ceph_d_release,
2183         .d_prune = ceph_d_prune,
2184         .d_init = ceph_d_init,
2185 };