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_fs_client *fsc,
 113                             struct ceph_dir_file_info *dfi,
 114                             const char *name,
 115                             int len, unsigned next_offset)
 116 {
 117         char *buf = kmalloc(len+1, GFP_KERNEL);
 118         if (!buf)
 119                 return -ENOMEM;
 120         kfree(dfi->last_name);
 121         dfi->last_name = buf;
 122         memcpy(dfi->last_name, name, len);
 123         dfi->last_name[len] = 0;
 124         dfi->next_offset = next_offset;
 125         doutc(fsc->client, "'%s'\n", dfi->last_name);
 126         return 0;
 127 }
 128
 129
 130 static struct dentry *
 131 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 132                         struct ceph_readdir_cache_control *cache_ctl)
 133 {
 134         struct inode *dir = d_inode(parent);
 135         struct ceph_client *cl = ceph_inode_to_client(dir);
 136         struct dentry *dentry;
 137         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 138         loff_t ptr_pos = idx * sizeof(struct dentry *);
 139         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 140
 141         if (ptr_pos >= i_size_read(dir))
 142                 return NULL;
 143
 144         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 145                 ceph_readdir_cache_release(cache_ctl);
 146                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 147                 if (!cache_ctl->page) {
 148                         doutc(cl, " page %lu not found\n", ptr_pgoff);
 149                         return ERR_PTR(-EAGAIN);
 150                 }
 151                 /* reading/filling the cache are serialized by
 152                    i_rwsem, no need to use page lock */
 153                 unlock_page(cache_ctl->page);
 154                 cache_ctl->dentries = kmap(cache_ctl->page);
 155         }
 156
 157         cache_ctl->index = idx & idx_mask;
 158
 159         rcu_read_lock();
 160         spin_lock(&parent->d_lock);
 161         /* check i_size again here, because empty directory can be
 162          * marked as complete while not holding the i_rwsem. */
 163         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 164                 dentry = cache_ctl->dentries[cache_ctl->index];
 165         else
 166                 dentry = NULL;
 167         spin_unlock(&parent->d_lock);
 168         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 169                 dentry = NULL;
 170         rcu_read_unlock();
 171         return dentry ? : ERR_PTR(-EAGAIN);
 172 }
 173
 174 /*
 175  * When possible, we try to satisfy a readdir by peeking at the
 176  * dcache.  We make this work by carefully ordering dentries on
 177  * d_child when we initially get results back from the MDS, and
 178  * falling back to a "normal" sync readdir if any dentries in the dir
 179  * are dropped.
 180  *
 181  * Complete dir indicates that we have all dentries in the dir.  It is
 182  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 183  * the MDS if/when the directory is modified).
 184  */
 185 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 186                             int shared_gen)
 187 {
 188         struct ceph_dir_file_info *dfi = file->private_data;
 189         struct dentry *parent = file->f_path.dentry;
 190         struct inode *dir = d_inode(parent);
 191         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(dir);
 192         struct ceph_client *cl = ceph_inode_to_client(dir);
 193         struct dentry *dentry, *last = NULL;
 194         struct ceph_dentry_info *di;
 195         struct ceph_readdir_cache_control cache_ctl = {};
 196         u64 idx = 0;
 197         int err = 0;
 198
 199         doutc(cl, "%p %llx.%llx v%u at %llx\n", dir, ceph_vinop(dir),
 200               (unsigned)shared_gen, ctx->pos);
 201
 202         /* search start position */
 203         if (ctx->pos > 2) {
 204                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 205                 while (count > 0) {
 206                         u64 step = count >> 1;
 207                         dentry = __dcache_find_get_entry(parent, idx + step,
 208                                                          &cache_ctl);
 209                         if (!dentry) {
 210                                 /* use linar search */
 211                                 idx = 0;
 212                                 break;
 213                         }
 214                         if (IS_ERR(dentry)) {
 215                                 err = PTR_ERR(dentry);
 216                                 goto out;
 217                         }
 218                         di = ceph_dentry(dentry);
 219                         spin_lock(&dentry->d_lock);
 220                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 221                                 idx += step + 1;
 222                                 count -= step + 1;
 223                         } else {
 224                                 count = step;
 225                         }
 226                         spin_unlock(&dentry->d_lock);
 227                         dput(dentry);
 228                 }
 229
 230                 doutc(cl, "%p %llx.%llx cache idx %llu\n", dir,
 231                       ceph_vinop(dir), idx);
 232         }
 233
 234
 235         for (;;) {
 236                 bool emit_dentry = false;
 237                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 238                 if (!dentry) {
 239                         dfi->file_info.flags |= CEPH_F_ATEND;
 240                         err = 0;
 241                         break;
 242                 }
 243                 if (IS_ERR(dentry)) {
 244                         err = PTR_ERR(dentry);
 245                         goto out;
 246                 }
 247
 248                 spin_lock(&dentry->d_lock);
 249                 di = ceph_dentry(dentry);
 250                 if (d_unhashed(dentry) ||
 251                     d_really_is_negative(dentry) ||
 252                     di->lease_shared_gen != shared_gen ||
 253                     ((dentry->d_flags & DCACHE_NOKEY_NAME) &&
 254                      fscrypt_has_encryption_key(dir))) {
 255                         spin_unlock(&dentry->d_lock);
 256                         dput(dentry);
 257                         err = -EAGAIN;
 258                         goto out;
 259                 }
 260                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 261                         __ceph_dentry_dir_lease_touch(di);
 262                         emit_dentry = true;
 263                 }
 264                 spin_unlock(&dentry->d_lock);
 265
 266                 if (emit_dentry) {
 267                         doutc(cl, " %llx dentry %p %pd %p\n", di->offset,
 268                               dentry, dentry, d_inode(dentry));
 269                         ctx->pos = di->offset;
 270                         if (!dir_emit(ctx, dentry->d_name.name,
 271                                       dentry->d_name.len, ceph_present_inode(d_inode(dentry)),
 272                                       d_inode(dentry)->i_mode >> 12)) {
 273                                 dput(dentry);
 274                                 err = 0;
 275                                 break;
 276                         }
 277                         ctx->pos++;
 278
 279                         if (last)
 280                                 dput(last);
 281                         last = dentry;
 282                 } else {
 283                         dput(dentry);
 284                 }
 285         }
 286 out:
 287         ceph_readdir_cache_release(&cache_ctl);
 288         if (last) {
 289                 int ret;
 290                 di = ceph_dentry(last);
 291                 ret = note_last_dentry(fsc, dfi, last->d_name.name,
 292                                        last->d_name.len,
 293                                        fpos_off(di->offset) + 1);
 294                 if (ret < 0)
 295                         err = ret;
 296                 dput(last);
 297                 /* last_name no longer match cache index */
 298                 if (dfi->readdir_cache_idx >= 0) {
 299                         dfi->readdir_cache_idx = -1;
 300                         dfi->dir_release_count = 0;
 301                 }
 302         }
 303         return err;
 304 }
 305
 306 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 307 {
 308         if (!dfi->last_readdir)
 309                 return true;
 310         if (is_hash_order(pos))
 311                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 312         else
 313                 return dfi->frag != fpos_frag(pos);
 314 }
 315
 316 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 317 {
 318         struct ceph_dir_file_info *dfi = file->private_data;
 319         struct inode *inode = file_inode(file);
 320         struct ceph_inode_info *ci = ceph_inode(inode);
 321         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 322         struct ceph_mds_client *mdsc = fsc->mdsc;
 323         struct ceph_client *cl = fsc->client;
 324         int i;
 325         int err;
 326         unsigned frag = -1;
 327         struct ceph_mds_reply_info_parsed *rinfo;
 328
 329         doutc(cl, "%p %llx.%llx file %p pos %llx\n", inode,
 330               ceph_vinop(inode), file, ctx->pos);
 331         if (dfi->file_info.flags & CEPH_F_ATEND)
 332                 return 0;
 333
 334         /* always start with . and .. */
 335         if (ctx->pos == 0) {
 336                 doutc(cl, "%p %llx.%llx off 0 -> '.'\n", inode,
 337                       ceph_vinop(inode));
 338                 if (!dir_emit(ctx, ".", 1, ceph_present_inode(inode),
 339                             inode->i_mode >> 12))
 340                         return 0;
 341                 ctx->pos = 1;
 342         }
 343         if (ctx->pos == 1) {
 344                 u64 ino;
 345                 struct dentry *dentry = file->f_path.dentry;
 346
 347                 spin_lock(&dentry->d_lock);
 348                 ino = ceph_present_inode(dentry->d_parent->d_inode);
 349                 spin_unlock(&dentry->d_lock);
 350
 351                 doutc(cl, "%p %llx.%llx off 1 -> '..'\n", inode,
 352                       ceph_vinop(inode));
 353                 if (!dir_emit(ctx, "..", 2, ino, inode->i_mode >> 12))
 354                         return 0;
 355                 ctx->pos = 2;
 356         }
 357
 358         err = ceph_fscrypt_prepare_readdir(inode);
 359         if (err < 0)
 360                 return err;
 361
 362         spin_lock(&ci->i_ceph_lock);
 363         /* request Fx cap. if have Fx, we don't need to release Fs cap
 364          * for later create/unlink. */
 365         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_WR);
 366         /* can we use the dcache? */
 367         if (ceph_test_mount_opt(fsc, DCACHE) &&
 368             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 369             ceph_snap(inode) != CEPH_SNAPDIR &&
 370             __ceph_dir_is_complete_ordered(ci) &&
 371             __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 372                 int shared_gen = atomic_read(&ci->i_shared_gen);
 373
 374                 spin_unlock(&ci->i_ceph_lock);
 375                 err = __dcache_readdir(file, ctx, shared_gen);
 376                 if (err != -EAGAIN)
 377                         return err;
 378         } else {
 379                 spin_unlock(&ci->i_ceph_lock);
 380         }
 381
 382         /* proceed with a normal readdir */
 383 more:
 384         /* do we have the correct frag content buffered? */
 385         if (need_send_readdir(dfi, ctx->pos)) {
 386                 struct ceph_mds_request *req;
 387                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 388                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 389
 390                 /* discard old result, if any */
 391                 if (dfi->last_readdir) {
 392                         ceph_mdsc_put_request(dfi->last_readdir);
 393                         dfi->last_readdir = NULL;
 394                 }
 395
 396                 if (is_hash_order(ctx->pos)) {
 397                         /* fragtree isn't always accurate. choose frag
 398                          * based on previous reply when possible. */
 399                         if (frag == (unsigned)-1)
 400                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 401                                                         NULL, NULL);
 402                 } else {
 403                         frag = fpos_frag(ctx->pos);
 404                 }
 405
 406                 doutc(cl, "fetching %p %llx.%llx frag %x offset '%s'\n",
 407                       inode, ceph_vinop(inode), frag, dfi->last_name);
 408                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 409                 if (IS_ERR(req))
 410                         return PTR_ERR(req);
 411
 412                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 413                 if (err) {
 414                         ceph_mdsc_put_request(req);
 415                         return err;
 416                 }
 417                 /* hints to request -> mds selection code */
 418                 req->r_direct_mode = USE_AUTH_MDS;
 419                 if (op == CEPH_MDS_OP_READDIR) {
 420                         req->r_direct_hash = ceph_frag_value(frag);
 421                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 422                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 423                 }
 424                 if (dfi->last_name) {
 425                         struct qstr d_name = { .name = dfi->last_name,
 426                                                .len = strlen(dfi->last_name) };
 427
 428                         req->r_path2 = kzalloc(NAME_MAX + 1, GFP_KERNEL);
 429                         if (!req->r_path2) {
 430                                 ceph_mdsc_put_request(req);
 431                                 return -ENOMEM;
 432                         }
 433
 434                         err = ceph_encode_encrypted_dname(inode, &d_name,
 435                                                           req->r_path2);
 436                         if (err < 0) {
 437                                 ceph_mdsc_put_request(req);
 438                                 return err;
 439                         }
 440                 } else if (is_hash_order(ctx->pos)) {
 441                         req->r_args.readdir.offset_hash =
 442                                 cpu_to_le32(fpos_hash(ctx->pos));
 443                 }
 444
 445                 req->r_dir_release_cnt = dfi->dir_release_count;
 446                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 447                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 448                 req->r_readdir_offset = dfi->next_offset;
 449                 req->r_args.readdir.frag = cpu_to_le32(frag);
 450                 req->r_args.readdir.flags =
 451                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 452
 453                 req->r_inode = inode;
 454                 ihold(inode);
 455                 req->r_dentry = dget(file->f_path.dentry);
 456                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 457                 if (err < 0) {
 458                         ceph_mdsc_put_request(req);
 459                         return err;
 460                 }
 461                 doutc(cl, "%p %llx.%llx got and parsed readdir result=%d"
 462                       "on frag %x, end=%d, complete=%d, hash_order=%d\n",
 463                       inode, ceph_vinop(inode), err, frag,
 464                       (int)req->r_reply_info.dir_end,
 465                       (int)req->r_reply_info.dir_complete,
 466                       (int)req->r_reply_info.hash_order);
 467
 468                 rinfo = &req->r_reply_info;
 469                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 470                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 471                         if (!rinfo->hash_order) {
 472                                 dfi->next_offset = req->r_readdir_offset;
 473                                 /* adjust ctx->pos to beginning of frag */
 474                                 ctx->pos = ceph_make_fpos(frag,
 475                                                           dfi->next_offset,
 476                                                           false);
 477                         }
 478                 }
 479
 480                 dfi->frag = frag;
 481                 dfi->last_readdir = req;
 482
 483                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 484                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 485                         if (dfi->readdir_cache_idx < 0) {
 486                                 /* preclude from marking dir ordered */
 487                                 dfi->dir_ordered_count = 0;
 488                         } else if (ceph_frag_is_leftmost(frag) &&
 489                                    dfi->next_offset == 2) {
 490                                 /* note dir version at start of readdir so
 491                                  * we can tell if any dentries get dropped */
 492                                 dfi->dir_release_count = req->r_dir_release_cnt;
 493                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 494                         }
 495                 } else {
 496                         doutc(cl, "%p %llx.%llx !did_prepopulate\n", inode,
 497                               ceph_vinop(inode));
 498                         /* disable readdir cache */
 499                         dfi->readdir_cache_idx = -1;
 500                         /* preclude from marking dir complete */
 501                         dfi->dir_release_count = 0;
 502                 }
 503
 504                 /* note next offset and last dentry name */
 505                 if (rinfo->dir_nr > 0) {
 506                         struct ceph_mds_reply_dir_entry *rde =
 507                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 508                         unsigned next_offset = req->r_reply_info.dir_end ?
 509                                         2 : (fpos_off(rde->offset) + 1);
 510                         err = note_last_dentry(fsc, dfi, rde->name,
 511                                                rde->name_len, next_offset);
 512                         if (err) {
 513                                 ceph_mdsc_put_request(dfi->last_readdir);
 514                                 dfi->last_readdir = NULL;
 515                                 return err;
 516                         }
 517                 } else if (req->r_reply_info.dir_end) {
 518                         dfi->next_offset = 2;
 519                         /* keep last name */
 520                 }
 521         }
 522
 523         rinfo = &dfi->last_readdir->r_reply_info;
 524         doutc(cl, "%p %llx.%llx frag %x num %d pos %llx chunk first %llx\n",
 525               inode, ceph_vinop(inode), dfi->frag, rinfo->dir_nr, ctx->pos,
 526               rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 527
 528         i = 0;
 529         /* search start position */
 530         if (rinfo->dir_nr > 0) {
 531                 int step, nr = rinfo->dir_nr;
 532                 while (nr > 0) {
 533                         step = nr >> 1;
 534                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 535                                 i +=  step + 1;
 536                                 nr -= step + 1;
 537                         } else {
 538                                 nr = step;
 539                         }
 540                 }
 541         }
 542         for (; i < rinfo->dir_nr; i++) {
 543                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 544
 545                 if (rde->offset < ctx->pos) {
 546                         pr_warn_client(cl,
 547                                 "%p %llx.%llx rde->offset 0x%llx ctx->pos 0x%llx\n",
 548                                 inode, ceph_vinop(inode), rde->offset, ctx->pos);
 549                         return -EIO;
 550                 }
 551
 552                 if (WARN_ON_ONCE(!rde->inode.in))
 553                         return -EIO;
 554
 555                 ctx->pos = rde->offset;
 556                 doutc(cl, "%p %llx.%llx (%d/%d) -> %llx '%.*s' %p\n", inode,
 557                       ceph_vinop(inode), i, rinfo->dir_nr, ctx->pos,
 558                       rde->name_len, rde->name, &rde->inode.in);
 559
 560                 if (!dir_emit(ctx, rde->name, rde->name_len,
 561                               ceph_present_ino(inode->i_sb, le64_to_cpu(rde->inode.in->ino)),
 562                               le32_to_cpu(rde->inode.in->mode) >> 12)) {
 563                         /*
 564                          * NOTE: Here no need to put the 'dfi->last_readdir',
 565                          * because when dir_emit stops us it's most likely
 566                          * doesn't have enough memory, etc. So for next readdir
 567                          * it will continue.
 568                          */
 569                         doutc(cl, "filldir stopping us...\n");
 570                         return 0;
 571                 }
 572
 573                 /* Reset the lengths to their original allocated vals */
 574                 ctx->pos++;
 575         }
 576
 577         ceph_mdsc_put_request(dfi->last_readdir);
 578         dfi->last_readdir = NULL;
 579
 580         if (dfi->next_offset > 2) {
 581                 frag = dfi->frag;
 582                 goto more;
 583         }
 584
 585         /* more frags? */
 586         if (!ceph_frag_is_rightmost(dfi->frag)) {
 587                 frag = ceph_frag_next(dfi->frag);
 588                 if (is_hash_order(ctx->pos)) {
 589                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 590                                                         dfi->next_offset, true);
 591                         if (new_pos > ctx->pos)
 592                                 ctx->pos = new_pos;
 593                         /* keep last_name */
 594                 } else {
 595                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 596                                                         false);
 597                         kfree(dfi->last_name);
 598                         dfi->last_name = NULL;
 599                 }
 600                 doutc(cl, "%p %llx.%llx next frag is %x\n", inode,
 601                       ceph_vinop(inode), frag);
 602                 goto more;
 603         }
 604         dfi->file_info.flags |= CEPH_F_ATEND;
 605
 606         /*
 607          * if dir_release_count still matches the dir, no dentries
 608          * were released during the whole readdir, and we should have
 609          * the complete dir contents in our cache.
 610          */
 611         if (atomic64_read(&ci->i_release_count) ==
 612                                         dfi->dir_release_count) {
 613                 spin_lock(&ci->i_ceph_lock);
 614                 if (dfi->dir_ordered_count ==
 615                                 atomic64_read(&ci->i_ordered_count)) {
 616                         doutc(cl, " marking %p %llx.%llx complete and ordered\n",
 617                               inode, ceph_vinop(inode));
 618                         /* use i_size to track number of entries in
 619                          * readdir cache */
 620                         BUG_ON(dfi->readdir_cache_idx < 0);
 621                         i_size_write(inode, dfi->readdir_cache_idx *
 622                                      sizeof(struct dentry*));
 623                 } else {
 624                         doutc(cl, " marking %llx.%llx complete\n",
 625                               ceph_vinop(inode));
 626                 }
 627                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 628                                         dfi->dir_ordered_count);
 629                 spin_unlock(&ci->i_ceph_lock);
 630         }
 631         doutc(cl, "%p %llx.%llx file %p done.\n", inode, ceph_vinop(inode),
 632               file);
 633         return 0;
 634 }
 635
 636 static void reset_readdir(struct ceph_dir_file_info *dfi)
 637 {
 638         if (dfi->last_readdir) {
 639                 ceph_mdsc_put_request(dfi->last_readdir);
 640                 dfi->last_readdir = NULL;
 641         }
 642         kfree(dfi->last_name);
 643         dfi->last_name = NULL;
 644         dfi->dir_release_count = 0;
 645         dfi->readdir_cache_idx = -1;
 646         dfi->next_offset = 2;  /* compensate for . and .. */
 647         dfi->file_info.flags &= ~CEPH_F_ATEND;
 648 }
 649
 650 /*
 651  * discard buffered readdir content on seekdir(0), or seek to new frag,
 652  * or seek prior to current chunk
 653  */
 654 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 655 {
 656         struct ceph_mds_reply_info_parsed *rinfo;
 657         loff_t chunk_offset;
 658         if (new_pos == 0)
 659                 return true;
 660         if (is_hash_order(new_pos)) {
 661                 /* no need to reset last_name for a forward seek when
 662                  * dentries are sotred in hash order */
 663         } else if (dfi->frag != fpos_frag(new_pos)) {
 664                 return true;
 665         }
 666         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 667         if (!rinfo || !rinfo->dir_nr)
 668                 return true;
 669         chunk_offset = rinfo->dir_entries[0].offset;
 670         return new_pos < chunk_offset ||
 671                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 672 }
 673
 674 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 675 {
 676         struct ceph_dir_file_info *dfi = file->private_data;
 677         struct inode *inode = file->f_mapping->host;
 678         struct ceph_client *cl = ceph_inode_to_client(inode);
 679         loff_t retval;
 680
 681         inode_lock(inode);
 682         retval = -EINVAL;
 683         switch (whence) {
 684         case SEEK_CUR:
 685                 offset += file->f_pos;
 686                 break;
 687         case SEEK_SET:
 688                 break;
 689         case SEEK_END:
 690                 retval = -EOPNOTSUPP;
 691                 goto out;
 692         default:
 693                 goto out;
 694         }
 695
 696         if (offset >= 0) {
 697                 if (need_reset_readdir(dfi, offset)) {
 698                         doutc(cl, "%p %llx.%llx dropping %p content\n",
 699                               inode, ceph_vinop(inode), file);
 700                         reset_readdir(dfi);
 701                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 702                         /* for hash offset, we don't know if a forward seek
 703                          * is within same frag */
 704                         dfi->dir_release_count = 0;
 705                         dfi->readdir_cache_idx = -1;
 706                 }
 707
 708                 if (offset != file->f_pos) {
 709                         file->f_pos = offset;
 710                         file->f_version = 0;
 711                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 712                 }
 713                 retval = offset;
 714         }
 715 out:
 716         inode_unlock(inode);
 717         return retval;
 718 }
 719
 720 /*
 721  * Handle lookups for the hidden .snap directory.
 722  */
 723 struct dentry *ceph_handle_snapdir(struct ceph_mds_request *req,
 724                                    struct dentry *dentry)
 725 {
 726         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
 727         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_rwsem */
 728         struct ceph_client *cl = ceph_inode_to_client(parent);
 729
 730         /* .snap dir? */
 731         if (ceph_snap(parent) == CEPH_NOSNAP &&
 732             strcmp(dentry->d_name.name, fsc->mount_options->snapdir_name) == 0) {
 733                 struct dentry *res;
 734                 struct inode *inode = ceph_get_snapdir(parent);
 735
 736                 res = d_splice_alias(inode, dentry);
 737                 doutc(cl, "ENOENT on snapdir %p '%pd', linking to "
 738                       "snapdir %p %llx.%llx. Spliced dentry %p\n",
 739                       dentry, dentry, inode, ceph_vinop(inode), res);
 740                 if (res)
 741                         dentry = res;
 742         }
 743         return dentry;
 744 }
 745
 746 /*
 747  * Figure out final result of a lookup/open request.
 748  *
 749  * Mainly, make sure we return the final req->r_dentry (if it already
 750  * existed) in place of the original VFS-provided dentry when they
 751  * differ.
 752  *
 753  * Gracefully handle the case where the MDS replies with -ENOENT and
 754  * no trace (which it may do, at its discretion, e.g., if it doesn't
 755  * care to issue a lease on the negative dentry).
 756  */
 757 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 758                                   struct dentry *dentry, int err)
 759 {
 760         struct ceph_client *cl = req->r_mdsc->fsc->client;
 761
 762         if (err == -ENOENT) {
 763                 /* no trace? */
 764                 err = 0;
 765                 if (!req->r_reply_info.head->is_dentry) {
 766                         doutc(cl,
 767                               "ENOENT and no trace, dentry %p inode %llx.%llx\n",
 768                               dentry, ceph_vinop(d_inode(dentry)));
 769                         if (d_really_is_positive(dentry)) {
 770                                 d_drop(dentry);
 771                                 err = -ENOENT;
 772                         } else {
 773                                 d_add(dentry, NULL);
 774                         }
 775                 }
 776         }
 777         if (err)
 778                 dentry = ERR_PTR(err);
 779         else if (dentry != req->r_dentry)
 780                 dentry = dget(req->r_dentry);   /* we got spliced */
 781         else
 782                 dentry = NULL;
 783         return dentry;
 784 }
 785
 786 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 787 {
 788         return ceph_ino(inode) == CEPH_INO_ROOT &&
 789                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 790 }
 791
 792 /*
 793  * Look up a single dir entry.  If there is a lookup intent, inform
 794  * the MDS so that it gets our 'caps wanted' value in a single op.
 795  */
 796 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 797                                   unsigned int flags)
 798 {
 799         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
 800         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 801         struct ceph_client *cl = fsc->client;
 802         struct ceph_mds_request *req;
 803         int op;
 804         int mask;
 805         int err;
 806
 807         doutc(cl, "%p %llx.%llx/'%pd' dentry %p\n", dir, ceph_vinop(dir),
 808               dentry, dentry);
 809
 810         if (dentry->d_name.len > NAME_MAX)
 811                 return ERR_PTR(-ENAMETOOLONG);
 812
 813         if (IS_ENCRYPTED(dir)) {
 814                 bool had_key = fscrypt_has_encryption_key(dir);
 815
 816                 err = fscrypt_prepare_lookup_partial(dir, dentry);
 817                 if (err < 0)
 818                         return ERR_PTR(err);
 819
 820                 /* mark directory as incomplete if it has been unlocked */
 821                 if (!had_key && fscrypt_has_encryption_key(dir))
 822                         ceph_dir_clear_complete(dir);
 823         }
 824
 825         /* can we conclude ENOENT locally? */
 826         if (d_really_is_negative(dentry)) {
 827                 struct ceph_inode_info *ci = ceph_inode(dir);
 828                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 829
 830                 spin_lock(&ci->i_ceph_lock);
 831                 doutc(cl, " dir %llx.%llx flags are 0x%lx\n",
 832                       ceph_vinop(dir), ci->i_ceph_flags);
 833                 if (strncmp(dentry->d_name.name,
 834                             fsc->mount_options->snapdir_name,
 835                             dentry->d_name.len) &&
 836                     !is_root_ceph_dentry(dir, dentry) &&
 837                     ceph_test_mount_opt(fsc, DCACHE) &&
 838                     __ceph_dir_is_complete(ci) &&
 839                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 840                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 841                         spin_unlock(&ci->i_ceph_lock);
 842                         doutc(cl, " dir %llx.%llx complete, -ENOENT\n",
 843                               ceph_vinop(dir));
 844                         d_add(dentry, NULL);
 845                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 846                         return NULL;
 847                 }
 848                 spin_unlock(&ci->i_ceph_lock);
 849         }
 850
 851         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 852                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 853         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 854         if (IS_ERR(req))
 855                 return ERR_CAST(req);
 856         req->r_dentry = dget(dentry);
 857         req->r_num_caps = 2;
 858
 859         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 860         if (ceph_security_xattr_wanted(dir))
 861                 mask |= CEPH_CAP_XATTR_SHARED;
 862         req->r_args.getattr.mask = cpu_to_le32(mask);
 863
 864         ihold(dir);
 865         req->r_parent = dir;
 866         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 867         err = ceph_mdsc_do_request(mdsc, NULL, req);
 868         if (err == -ENOENT) {
 869                 struct dentry *res;
 870
 871                 res = ceph_handle_snapdir(req, dentry);
 872                 if (IS_ERR(res)) {
 873                         err = PTR_ERR(res);
 874                 } else {
 875                         dentry = res;
 876                         err = 0;
 877                 }
 878         }
 879         dentry = ceph_finish_lookup(req, dentry, err);
 880         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 881         doutc(cl, "result=%p\n", dentry);
 882         return dentry;
 883 }
 884
 885 /*
 886  * If we do a create but get no trace back from the MDS, follow up with
 887  * a lookup (the VFS expects us to link up the provided dentry).
 888  */
 889 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 890 {
 891         struct dentry *result = ceph_lookup(dir, dentry, 0);
 892
 893         if (result && !IS_ERR(result)) {
 894                 /*
 895                  * We created the item, then did a lookup, and found
 896                  * it was already linked to another inode we already
 897                  * had in our cache (and thus got spliced). To not
 898                  * confuse VFS (especially when inode is a directory),
 899                  * we don't link our dentry to that inode, return an
 900                  * error instead.
 901                  *
 902                  * This event should be rare and it happens only when
 903                  * we talk to old MDS. Recent MDS does not send traceless
 904                  * reply for request that creates new inode.
 905                  */
 906                 d_drop(result);
 907                 return -ESTALE;
 908         }
 909         return PTR_ERR(result);
 910 }
 911
 912 static int ceph_mknod(struct mnt_idmap *idmap, struct inode *dir,
 913                       struct dentry *dentry, umode_t mode, dev_t rdev)
 914 {
 915         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 916         struct ceph_client *cl = mdsc->fsc->client;
 917         struct ceph_mds_request *req;
 918         struct ceph_acl_sec_ctx as_ctx = {};
 919         int err;
 920
 921         if (ceph_snap(dir) != CEPH_NOSNAP)
 922                 return -EROFS;
 923
 924         err = ceph_wait_on_conflict_unlink(dentry);
 925         if (err)
 926                 return err;
 927
 928         if (ceph_quota_is_max_files_exceeded(dir)) {
 929                 err = -EDQUOT;
 930                 goto out;
 931         }
 932
 933         doutc(cl, "%p %llx.%llx/'%pd' dentry %p mode 0%ho rdev %d\n",
 934               dir, ceph_vinop(dir), dentry, dentry, mode, rdev);
 935         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 936         if (IS_ERR(req)) {
 937                 err = PTR_ERR(req);
 938                 goto out;
 939         }
 940
 941         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
 942         if (IS_ERR(req->r_new_inode)) {
 943                 err = PTR_ERR(req->r_new_inode);
 944                 req->r_new_inode = NULL;
 945                 goto out_req;
 946         }
 947
 948         if (S_ISREG(mode) && IS_ENCRYPTED(dir))
 949                 set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
 950
 951         req->r_dentry = dget(dentry);
 952         req->r_num_caps = 2;
 953         req->r_parent = dir;
 954         ihold(dir);
 955         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 956         req->r_mnt_idmap = mnt_idmap_get(idmap);
 957         req->r_args.mknod.mode = cpu_to_le32(mode);
 958         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 959         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
 960                              CEPH_CAP_XATTR_EXCL;
 961         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 962
 963         ceph_as_ctx_to_req(req, &as_ctx);
 964
 965         err = ceph_mdsc_do_request(mdsc, dir, req);
 966         if (!err && !req->r_reply_info.head->is_dentry)
 967                 err = ceph_handle_notrace_create(dir, dentry);
 968 out_req:
 969         ceph_mdsc_put_request(req);
 970 out:
 971         if (!err)
 972                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 973         else
 974                 d_drop(dentry);
 975         ceph_release_acl_sec_ctx(&as_ctx);
 976         return err;
 977 }
 978
 979 static int ceph_create(struct mnt_idmap *idmap, struct inode *dir,
 980                        struct dentry *dentry, umode_t mode, bool excl)
 981 {
 982         return ceph_mknod(idmap, dir, dentry, mode, 0);
 983 }
 984
 985 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
 986 static int prep_encrypted_symlink_target(struct ceph_mds_request *req,
 987                                          const char *dest)
 988 {
 989         int err;
 990         int len = strlen(dest);
 991         struct fscrypt_str osd_link = FSTR_INIT(NULL, 0);
 992
 993         err = fscrypt_prepare_symlink(req->r_parent, dest, len, PATH_MAX,
 994                                       &osd_link);
 995         if (err)
 996                 goto out;
 997
 998         err = fscrypt_encrypt_symlink(req->r_new_inode, dest, len, &osd_link);
 999         if (err)
1000                 goto out;
1001
1002         req->r_path2 = kmalloc(CEPH_BASE64_CHARS(osd_link.len) + 1, GFP_KERNEL);
1003         if (!req->r_path2) {
1004                 err = -ENOMEM;
1005                 goto out;
1006         }
1007
1008         len = ceph_base64_encode(osd_link.name, osd_link.len, req->r_path2);
1009         req->r_path2[len] = '\0';
1010 out:
1011         fscrypt_fname_free_buffer(&osd_link);
1012         return err;
1013 }
1014 #else
1015 static int prep_encrypted_symlink_target(struct ceph_mds_request *req,
1016                                          const char *dest)
1017 {
1018         return -EOPNOTSUPP;
1019 }
1020 #endif
1021
1022 static int ceph_symlink(struct mnt_idmap *idmap, struct inode *dir,
1023                         struct dentry *dentry, const char *dest)
1024 {
1025         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1026         struct ceph_client *cl = mdsc->fsc->client;
1027         struct ceph_mds_request *req;
1028         struct ceph_acl_sec_ctx as_ctx = {};
1029         umode_t mode = S_IFLNK | 0777;
1030         int err;
1031
1032         if (ceph_snap(dir) != CEPH_NOSNAP)
1033                 return -EROFS;
1034
1035         err = ceph_wait_on_conflict_unlink(dentry);
1036         if (err)
1037                 return err;
1038
1039         if (ceph_quota_is_max_files_exceeded(dir)) {
1040                 err = -EDQUOT;
1041                 goto out;
1042         }
1043
1044         doutc(cl, "%p %llx.%llx/'%pd' to '%s'\n", dir, ceph_vinop(dir), dentry,
1045               dest);
1046         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
1047         if (IS_ERR(req)) {
1048                 err = PTR_ERR(req);
1049                 goto out;
1050         }
1051
1052         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
1053         if (IS_ERR(req->r_new_inode)) {
1054                 err = PTR_ERR(req->r_new_inode);
1055                 req->r_new_inode = NULL;
1056                 goto out_req;
1057         }
1058
1059         req->r_parent = dir;
1060         ihold(dir);
1061
1062         if (IS_ENCRYPTED(req->r_new_inode)) {
1063                 err = prep_encrypted_symlink_target(req, dest);
1064                 if (err)
1065                         goto out_req;
1066         } else {
1067                 req->r_path2 = kstrdup(dest, GFP_KERNEL);
1068                 if (!req->r_path2) {
1069                         err = -ENOMEM;
1070                         goto out_req;
1071                 }
1072         }
1073
1074         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1075         req->r_mnt_idmap = mnt_idmap_get(idmap);
1076         req->r_dentry = dget(dentry);
1077         req->r_num_caps = 2;
1078         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
1079                              CEPH_CAP_XATTR_EXCL;
1080         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1081
1082         ceph_as_ctx_to_req(req, &as_ctx);
1083
1084         err = ceph_mdsc_do_request(mdsc, dir, req);
1085         if (!err && !req->r_reply_info.head->is_dentry)
1086                 err = ceph_handle_notrace_create(dir, dentry);
1087 out_req:
1088         ceph_mdsc_put_request(req);
1089 out:
1090         if (err)
1091                 d_drop(dentry);
1092         ceph_release_acl_sec_ctx(&as_ctx);
1093         return err;
1094 }
1095
1096 static int ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1097                       struct dentry *dentry, umode_t mode)
1098 {
1099         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1100         struct ceph_client *cl = mdsc->fsc->client;
1101         struct ceph_mds_request *req;
1102         struct ceph_acl_sec_ctx as_ctx = {};
1103         int err;
1104         int op;
1105
1106         err = ceph_wait_on_conflict_unlink(dentry);
1107         if (err)
1108                 return err;
1109
1110         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1111                 /* mkdir .snap/foo is a MKSNAP */
1112                 op = CEPH_MDS_OP_MKSNAP;
1113                 doutc(cl, "mksnap %llx.%llx/'%pd' dentry %p\n",
1114                       ceph_vinop(dir), dentry, dentry);
1115         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1116                 doutc(cl, "mkdir %llx.%llx/'%pd' dentry %p mode 0%ho\n",
1117                       ceph_vinop(dir), dentry, dentry, mode);
1118                 op = CEPH_MDS_OP_MKDIR;
1119         } else {
1120                 err = -EROFS;
1121                 goto out;
1122         }
1123
1124         if (op == CEPH_MDS_OP_MKDIR &&
1125             ceph_quota_is_max_files_exceeded(dir)) {
1126                 err = -EDQUOT;
1127                 goto out;
1128         }
1129         if ((op == CEPH_MDS_OP_MKSNAP) && IS_ENCRYPTED(dir) &&
1130             !fscrypt_has_encryption_key(dir)) {
1131                 err = -ENOKEY;
1132                 goto out;
1133         }
1134
1135
1136         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1137         if (IS_ERR(req)) {
1138                 err = PTR_ERR(req);
1139                 goto out;
1140         }
1141
1142         mode |= S_IFDIR;
1143         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
1144         if (IS_ERR(req->r_new_inode)) {
1145                 err = PTR_ERR(req->r_new_inode);
1146                 req->r_new_inode = NULL;
1147                 goto out_req;
1148         }
1149
1150         req->r_dentry = dget(dentry);
1151         req->r_num_caps = 2;
1152         req->r_parent = dir;
1153         ihold(dir);
1154         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1155         if (op == CEPH_MDS_OP_MKDIR)
1156                 req->r_mnt_idmap = mnt_idmap_get(idmap);
1157         req->r_args.mkdir.mode = cpu_to_le32(mode);
1158         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
1159                              CEPH_CAP_XATTR_EXCL;
1160         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1161
1162         ceph_as_ctx_to_req(req, &as_ctx);
1163
1164         err = ceph_mdsc_do_request(mdsc, dir, req);
1165         if (!err &&
1166             !req->r_reply_info.head->is_target &&
1167             !req->r_reply_info.head->is_dentry)
1168                 err = ceph_handle_notrace_create(dir, dentry);
1169 out_req:
1170         ceph_mdsc_put_request(req);
1171 out:
1172         if (!err)
1173                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1174         else
1175                 d_drop(dentry);
1176         ceph_release_acl_sec_ctx(&as_ctx);
1177         return err;
1178 }
1179
1180 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1181                      struct dentry *dentry)
1182 {
1183         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1184         struct ceph_client *cl = mdsc->fsc->client;
1185         struct ceph_mds_request *req;
1186         int err;
1187
1188         if (dentry->d_flags & DCACHE_DISCONNECTED)
1189                 return -EINVAL;
1190
1191         err = ceph_wait_on_conflict_unlink(dentry);
1192         if (err)
1193                 return err;
1194
1195         if (ceph_snap(dir) != CEPH_NOSNAP)
1196                 return -EROFS;
1197
1198         err = fscrypt_prepare_link(old_dentry, dir, dentry);
1199         if (err)
1200                 return err;
1201
1202         doutc(cl, "%p %llx.%llx/'%pd' to '%pd'\n", dir, ceph_vinop(dir),
1203               old_dentry, dentry);
1204         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1205         if (IS_ERR(req)) {
1206                 d_drop(dentry);
1207                 return PTR_ERR(req);
1208         }
1209         req->r_dentry = dget(dentry);
1210         req->r_num_caps = 2;
1211         req->r_old_dentry = dget(old_dentry);
1212         /*
1213          * The old_dentry maybe a DCACHE_DISCONNECTED dentry, then we
1214          * will just pass the ino# to MDSs.
1215          */
1216         if (old_dentry->d_flags & DCACHE_DISCONNECTED)
1217                 req->r_ino2 = ceph_vino(d_inode(old_dentry));
1218         req->r_parent = dir;
1219         ihold(dir);
1220         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1221         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1222         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1223         /* release LINK_SHARED on source inode (mds will lock it) */
1224         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1225         err = ceph_mdsc_do_request(mdsc, dir, req);
1226         if (err) {
1227                 d_drop(dentry);
1228         } else if (!req->r_reply_info.head->is_dentry) {
1229                 ihold(d_inode(old_dentry));
1230                 d_instantiate(dentry, d_inode(old_dentry));
1231         }
1232         ceph_mdsc_put_request(req);
1233         return err;
1234 }
1235
1236 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1237                                  struct ceph_mds_request *req)
1238 {
1239         struct dentry *dentry = req->r_dentry;
1240         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
1241         struct ceph_client *cl = fsc->client;
1242         struct ceph_dentry_info *di = ceph_dentry(dentry);
1243         int result = req->r_err ? req->r_err :
1244                         le32_to_cpu(req->r_reply_info.head->result);
1245
1246         if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
1247                 pr_warn_client(cl,
1248                         "dentry %p:%pd async unlink bit is not set\n",
1249                         dentry, dentry);
1250
1251         spin_lock(&fsc->async_unlink_conflict_lock);
1252         hash_del_rcu(&di->hnode);
1253         spin_unlock(&fsc->async_unlink_conflict_lock);
1254
1255         spin_lock(&dentry->d_lock);
1256         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1257         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT);
1258         spin_unlock(&dentry->d_lock);
1259
1260         synchronize_rcu();
1261
1262         if (result == -EJUKEBOX)
1263                 goto out;
1264
1265         /* If op failed, mark everyone involved for errors */
1266         if (result) {
1267                 int pathlen = 0;
1268                 u64 base = 0;
1269                 char *path = ceph_mdsc_build_path(mdsc, dentry, &pathlen,
1270                                                   &base, 0);
1271
1272                 /* mark error on parent + clear complete */
1273                 mapping_set_error(req->r_parent->i_mapping, result);
1274                 ceph_dir_clear_complete(req->r_parent);
1275
1276                 /* drop the dentry -- we don't know its status */
1277                 if (!d_unhashed(dentry))
1278                         d_drop(dentry);
1279
1280                 /* mark inode itself for an error (since metadata is bogus) */
1281                 mapping_set_error(req->r_old_inode->i_mapping, result);
1282
1283                 pr_warn_client(cl, "failure path=(%llx)%s result=%d!\n",
1284                                base, IS_ERR(path) ? "<<bad>>" : path, result);
1285                 ceph_mdsc_free_path(path, pathlen);
1286         }
1287 out:
1288         iput(req->r_old_inode);
1289         ceph_mdsc_release_dir_caps(req);
1290 }
1291
1292 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1293 {
1294         struct ceph_inode_info *ci = ceph_inode(dir);
1295         struct ceph_dentry_info *di;
1296         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1297
1298         spin_lock(&ci->i_ceph_lock);
1299         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1300                 ceph_take_cap_refs(ci, want, false);
1301                 got = want;
1302         }
1303         spin_unlock(&ci->i_ceph_lock);
1304
1305         /* If we didn't get anything, return 0 */
1306         if (!got)
1307                 return 0;
1308
1309         spin_lock(&dentry->d_lock);
1310         di = ceph_dentry(dentry);
1311         /*
1312          * - We are holding Fx, which implies Fs caps.
1313          * - Only support async unlink for primary linkage
1314          */
1315         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1316             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1317                 want = 0;
1318         spin_unlock(&dentry->d_lock);
1319
1320         /* Do we still want what we've got? */
1321         if (want == got)
1322                 return got;
1323
1324         ceph_put_cap_refs(ci, got);
1325         return 0;
1326 }
1327
1328 /*
1329  * rmdir and unlink are differ only by the metadata op code
1330  */
1331 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1332 {
1333         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
1334         struct ceph_client *cl = fsc->client;
1335         struct ceph_mds_client *mdsc = fsc->mdsc;
1336         struct inode *inode = d_inode(dentry);
1337         struct ceph_mds_request *req;
1338         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1339         int err = -EROFS;
1340         int op;
1341
1342         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1343                 /* rmdir .snap/foo is RMSNAP */
1344                 doutc(cl, "rmsnap %llx.%llx/'%pd' dn\n", ceph_vinop(dir),
1345                       dentry);
1346                 op = CEPH_MDS_OP_RMSNAP;
1347         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1348                 doutc(cl, "unlink/rmdir %llx.%llx/'%pd' inode %llx.%llx\n",
1349                       ceph_vinop(dir), dentry, ceph_vinop(inode));
1350                 op = d_is_dir(dentry) ?
1351                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1352         } else
1353                 goto out;
1354 retry:
1355         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1356         if (IS_ERR(req)) {
1357                 err = PTR_ERR(req);
1358                 goto out;
1359         }
1360         req->r_dentry = dget(dentry);
1361         req->r_num_caps = 2;
1362         req->r_parent = dir;
1363         ihold(dir);
1364         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1365         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1366         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1367
1368         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1369             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1370                 struct ceph_dentry_info *di = ceph_dentry(dentry);
1371
1372                 doutc(cl, "async unlink on %llx.%llx/'%pd' caps=%s",
1373                       ceph_vinop(dir), dentry,
1374                       ceph_cap_string(req->r_dir_caps));
1375                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1376                 req->r_callback = ceph_async_unlink_cb;
1377                 req->r_old_inode = d_inode(dentry);
1378                 ihold(req->r_old_inode);
1379
1380                 spin_lock(&dentry->d_lock);
1381                 di->flags |= CEPH_DENTRY_ASYNC_UNLINK;
1382                 spin_unlock(&dentry->d_lock);
1383
1384                 spin_lock(&fsc->async_unlink_conflict_lock);
1385                 hash_add_rcu(fsc->async_unlink_conflict, &di->hnode,
1386                              dentry->d_name.hash);
1387                 spin_unlock(&fsc->async_unlink_conflict_lock);
1388
1389                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1390                 if (!err) {
1391                         /*
1392                          * We have enough caps, so we assume that the unlink
1393                          * will succeed. Fix up the target inode and dcache.
1394                          */
1395                         drop_nlink(inode);
1396                         d_delete(dentry);
1397                 } else {
1398                         spin_lock(&fsc->async_unlink_conflict_lock);
1399                         hash_del_rcu(&di->hnode);
1400                         spin_unlock(&fsc->async_unlink_conflict_lock);
1401
1402                         spin_lock(&dentry->d_lock);
1403                         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1404                         spin_unlock(&dentry->d_lock);
1405
1406                         if (err == -EJUKEBOX) {
1407                                 try_async = false;
1408                                 ceph_mdsc_put_request(req);
1409                                 goto retry;
1410                         }
1411                 }
1412         } else {
1413                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1414                 err = ceph_mdsc_do_request(mdsc, dir, req);
1415                 if (!err && !req->r_reply_info.head->is_dentry)
1416                         d_delete(dentry);
1417         }
1418
1419         ceph_mdsc_put_request(req);
1420 out:
1421         return err;
1422 }
1423
1424 static int ceph_rename(struct mnt_idmap *idmap, struct inode *old_dir,
1425                        struct dentry *old_dentry, struct inode *new_dir,
1426                        struct dentry *new_dentry, unsigned int flags)
1427 {
1428         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1429         struct ceph_client *cl = mdsc->fsc->client;
1430         struct ceph_mds_request *req;
1431         int op = CEPH_MDS_OP_RENAME;
1432         int err;
1433
1434         if (flags)
1435                 return -EINVAL;
1436
1437         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1438                 return -EXDEV;
1439         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1440                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1441                         op = CEPH_MDS_OP_RENAMESNAP;
1442                 else
1443                         return -EROFS;
1444         }
1445         /* don't allow cross-quota renames */
1446         if ((old_dir != new_dir) &&
1447             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1448                 return -EXDEV;
1449
1450         err = ceph_wait_on_conflict_unlink(new_dentry);
1451         if (err)
1452                 return err;
1453
1454         err = fscrypt_prepare_rename(old_dir, old_dentry, new_dir, new_dentry,
1455                                      flags);
1456         if (err)
1457                 return err;
1458
1459         doutc(cl, "%llx.%llx/'%pd' to %llx.%llx/'%pd'\n",
1460               ceph_vinop(old_dir), old_dentry, ceph_vinop(new_dir),
1461               new_dentry);
1462         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1463         if (IS_ERR(req))
1464                 return PTR_ERR(req);
1465         ihold(old_dir);
1466         req->r_dentry = dget(new_dentry);
1467         req->r_num_caps = 2;
1468         req->r_old_dentry = dget(old_dentry);
1469         req->r_old_dentry_dir = old_dir;
1470         req->r_parent = new_dir;
1471         ihold(new_dir);
1472         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1473         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1474         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1475         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1476         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1477         /* release LINK_RDCACHE on source inode (mds will lock it) */
1478         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1479         if (d_really_is_positive(new_dentry)) {
1480                 req->r_inode_drop =
1481                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1482         }
1483         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1484         if (!err && !req->r_reply_info.head->is_dentry) {
1485                 /*
1486                  * Normally d_move() is done by fill_trace (called by
1487                  * do_request, above).  If there is no trace, we need
1488                  * to do it here.
1489                  */
1490                 d_move(old_dentry, new_dentry);
1491         }
1492         ceph_mdsc_put_request(req);
1493         return err;
1494 }
1495
1496 /*
1497  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1498  * Leases at front of the list will expire first. (Assume all leases have
1499  * similar duration)
1500  *
1501  * Called under dentry->d_lock.
1502  */
1503 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1504 {
1505         struct dentry *dn = di->dentry;
1506         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dn->d_sb)->mdsc;
1507         struct ceph_client *cl = mdsc->fsc->client;
1508
1509         doutc(cl, "%p %p '%pd'\n", di, dn, dn);
1510
1511         di->flags |= CEPH_DENTRY_LEASE_LIST;
1512         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1513                 di->flags |= CEPH_DENTRY_REFERENCED;
1514                 return;
1515         }
1516
1517         spin_lock(&mdsc->dentry_list_lock);
1518         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1519         spin_unlock(&mdsc->dentry_list_lock);
1520 }
1521
1522 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1523                                      struct ceph_dentry_info *di)
1524 {
1525         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1526         di->lease_gen = 0;
1527         di->time = jiffies;
1528         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1529 }
1530
1531 /*
1532  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1533  * list if it's not in the list, otherwise set 'referenced' flag.
1534  *
1535  * Called under dentry->d_lock.
1536  */
1537 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1538 {
1539         struct dentry *dn = di->dentry;
1540         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dn->d_sb)->mdsc;
1541         struct ceph_client *cl = mdsc->fsc->client;
1542
1543         doutc(cl, "%p %p '%pd' (offset 0x%llx)\n", di, dn, dn, di->offset);
1544
1545         if (!list_empty(&di->lease_list)) {
1546                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1547                         /* don't remove dentry from dentry lease list
1548                          * if its lease is valid */
1549                         if (__dentry_lease_is_valid(di))
1550                                 return;
1551                 } else {
1552                         di->flags |= CEPH_DENTRY_REFERENCED;
1553                         return;
1554                 }
1555         }
1556
1557         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1558                 di->flags |= CEPH_DENTRY_REFERENCED;
1559                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1560                 return;
1561         }
1562
1563         spin_lock(&mdsc->dentry_list_lock);
1564         __dentry_dir_lease_touch(mdsc, di),
1565         spin_unlock(&mdsc->dentry_list_lock);
1566 }
1567
1568 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1569 {
1570         struct ceph_mds_client *mdsc;
1571         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1572                 return;
1573         if (list_empty(&di->lease_list))
1574                 return;
1575
1576         mdsc = ceph_sb_to_fs_client(di->dentry->d_sb)->mdsc;
1577         spin_lock(&mdsc->dentry_list_lock);
1578         list_del_init(&di->lease_list);
1579         spin_unlock(&mdsc->dentry_list_lock);
1580 }
1581
1582 enum {
1583         KEEP    = 0,
1584         DELETE  = 1,
1585         TOUCH   = 2,
1586         STOP    = 4,
1587 };
1588
1589 struct ceph_lease_walk_control {
1590         bool dir_lease;
1591         bool expire_dir_lease;
1592         unsigned long nr_to_scan;
1593         unsigned long dir_lease_ttl;
1594 };
1595
1596 static unsigned long
1597 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1598                      struct ceph_lease_walk_control *lwc,
1599                      int (*check)(struct dentry*, void*))
1600 {
1601         struct ceph_dentry_info *di, *tmp;
1602         struct dentry *dentry, *last = NULL;
1603         struct list_head* list;
1604         LIST_HEAD(dispose);
1605         unsigned long freed = 0;
1606         int ret = 0;
1607
1608         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1609         spin_lock(&mdsc->dentry_list_lock);
1610         list_for_each_entry_safe(di, tmp, list, lease_list) {
1611                 if (!lwc->nr_to_scan)
1612                         break;
1613                 --lwc->nr_to_scan;
1614
1615                 dentry = di->dentry;
1616                 if (last == dentry)
1617                         break;
1618
1619                 if (!spin_trylock(&dentry->d_lock))
1620                         continue;
1621
1622                 if (__lockref_is_dead(&dentry->d_lockref)) {
1623                         list_del_init(&di->lease_list);
1624                         goto next;
1625                 }
1626
1627                 ret = check(dentry, lwc);
1628                 if (ret & TOUCH) {
1629                         /* move it into tail of dir lease list */
1630                         __dentry_dir_lease_touch(mdsc, di);
1631                         if (!last)
1632                                 last = dentry;
1633                 }
1634                 if (ret & DELETE) {
1635                         /* stale lease */
1636                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1637                         if (dentry->d_lockref.count > 0) {
1638                                 /* update_dentry_lease() will re-add
1639                                  * it to lease list, or
1640                                  * ceph_d_delete() will return 1 when
1641                                  * last reference is dropped */
1642                                 list_del_init(&di->lease_list);
1643                         } else {
1644                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1645                                 list_move_tail(&di->lease_list, &dispose);
1646                                 dget_dlock(dentry);
1647                         }
1648                 }
1649 next:
1650                 spin_unlock(&dentry->d_lock);
1651                 if (ret & STOP)
1652                         break;
1653         }
1654         spin_unlock(&mdsc->dentry_list_lock);
1655
1656         while (!list_empty(&dispose)) {
1657                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1658                                       lease_list);
1659                 dentry = di->dentry;
1660                 spin_lock(&dentry->d_lock);
1661
1662                 list_del_init(&di->lease_list);
1663                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1664                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1665                         spin_lock(&mdsc->dentry_list_lock);
1666                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1667                                 list_add_tail(&di->lease_list,
1668                                               &mdsc->dentry_leases);
1669                         } else {
1670                                 __dentry_dir_lease_touch(mdsc, di);
1671                         }
1672                         spin_unlock(&mdsc->dentry_list_lock);
1673                 } else {
1674                         freed++;
1675                 }
1676
1677                 spin_unlock(&dentry->d_lock);
1678                 /* ceph_d_delete() does the trick */
1679                 dput(dentry);
1680         }
1681         return freed;
1682 }
1683
1684 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1685 {
1686         struct ceph_dentry_info *di = ceph_dentry(dentry);
1687         int ret;
1688
1689         if (__dentry_lease_is_valid(di))
1690                 return STOP;
1691         ret = __dir_lease_try_check(dentry);
1692         if (ret == -EBUSY)
1693                 return KEEP;
1694         if (ret > 0)
1695                 return TOUCH;
1696         return DELETE;
1697 }
1698
1699 static int __dir_lease_check(struct dentry *dentry, void *arg)
1700 {
1701         struct ceph_lease_walk_control *lwc = arg;
1702         struct ceph_dentry_info *di = ceph_dentry(dentry);
1703
1704         int ret = __dir_lease_try_check(dentry);
1705         if (ret == -EBUSY)
1706                 return KEEP;
1707         if (ret > 0) {
1708                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1709                         return STOP;
1710                 /* Move dentry to tail of dir lease list if we don't want
1711                  * to delete it. So dentries in the list are checked in a
1712                  * round robin manner */
1713                 if (!lwc->expire_dir_lease)
1714                         return TOUCH;
1715                 if (dentry->d_lockref.count > 0 ||
1716                     (di->flags & CEPH_DENTRY_REFERENCED))
1717                         return TOUCH;
1718                 /* invalidate dir lease */
1719                 di->lease_shared_gen = 0;
1720         }
1721         return DELETE;
1722 }
1723
1724 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1725 {
1726         struct ceph_lease_walk_control lwc;
1727         unsigned long count;
1728         unsigned long freed;
1729
1730         spin_lock(&mdsc->caps_list_lock);
1731         if (mdsc->caps_use_max > 0 &&
1732             mdsc->caps_use_count > mdsc->caps_use_max)
1733                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1734         else
1735                 count = 0;
1736         spin_unlock(&mdsc->caps_list_lock);
1737
1738         lwc.dir_lease = false;
1739         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1740         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1741         if (!lwc.nr_to_scan) /* more invalid leases */
1742                 return -EAGAIN;
1743
1744         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1745                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1746
1747         lwc.dir_lease = true;
1748         lwc.expire_dir_lease = freed < count;
1749         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1750         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1751         if (!lwc.nr_to_scan) /* more to check */
1752                 return -EAGAIN;
1753
1754         return freed > 0 ? 1 : 0;
1755 }
1756
1757 /*
1758  * Ensure a dentry lease will no longer revalidate.
1759  */
1760 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1761 {
1762         struct ceph_dentry_info *di = ceph_dentry(dentry);
1763         spin_lock(&dentry->d_lock);
1764         di->time = jiffies;
1765         di->lease_shared_gen = 0;
1766         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1767         __dentry_lease_unlist(di);
1768         spin_unlock(&dentry->d_lock);
1769 }
1770
1771 /*
1772  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1773  * renew if the least is more than half up.
1774  */
1775 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1776 {
1777         struct ceph_mds_session *session;
1778
1779         if (!di->lease_gen)
1780                 return false;
1781
1782         session = di->lease_session;
1783         if (session) {
1784                 u32 gen;
1785                 unsigned long ttl;
1786
1787                 gen = atomic_read(&session->s_cap_gen);
1788                 ttl = session->s_cap_ttl;
1789
1790                 if (di->lease_gen == gen &&
1791                     time_before(jiffies, ttl) &&
1792                     time_before(jiffies, di->time))
1793                         return true;
1794         }
1795         di->lease_gen = 0;
1796         return false;
1797 }
1798
1799 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1800 {
1801         struct ceph_dentry_info *di;
1802         struct ceph_mds_session *session = NULL;
1803         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dentry->d_sb)->mdsc;
1804         struct ceph_client *cl = mdsc->fsc->client;
1805         u32 seq = 0;
1806         int valid = 0;
1807
1808         spin_lock(&dentry->d_lock);
1809         di = ceph_dentry(dentry);
1810         if (di && __dentry_lease_is_valid(di)) {
1811                 valid = 1;
1812
1813                 if (di->lease_renew_after &&
1814                     time_after(jiffies, di->lease_renew_after)) {
1815                         /*
1816                          * We should renew. If we're in RCU walk mode
1817                          * though, we can't do that so just return
1818                          * -ECHILD.
1819                          */
1820                         if (flags & LOOKUP_RCU) {
1821                                 valid = -ECHILD;
1822                         } else {
1823                                 session = ceph_get_mds_session(di->lease_session);
1824                                 seq = di->lease_seq;
1825                                 di->lease_renew_after = 0;
1826                                 di->lease_renew_from = jiffies;
1827                         }
1828                 }
1829         }
1830         spin_unlock(&dentry->d_lock);
1831
1832         if (session) {
1833                 ceph_mdsc_lease_send_msg(session, dentry,
1834                                          CEPH_MDS_LEASE_RENEW, seq);
1835                 ceph_put_mds_session(session);
1836         }
1837         doutc(cl, "dentry %p = %d\n", dentry, valid);
1838         return valid;
1839 }
1840
1841 /*
1842  * Called under dentry->d_lock.
1843  */
1844 static int __dir_lease_try_check(const struct dentry *dentry)
1845 {
1846         struct ceph_dentry_info *di = ceph_dentry(dentry);
1847         struct inode *dir;
1848         struct ceph_inode_info *ci;
1849         int valid = 0;
1850
1851         if (!di->lease_shared_gen)
1852                 return 0;
1853         if (IS_ROOT(dentry))
1854                 return 0;
1855
1856         dir = d_inode(dentry->d_parent);
1857         ci = ceph_inode(dir);
1858
1859         if (spin_trylock(&ci->i_ceph_lock)) {
1860                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1861                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1862                         valid = 1;
1863                 spin_unlock(&ci->i_ceph_lock);
1864         } else {
1865                 valid = -EBUSY;
1866         }
1867
1868         if (!valid)
1869                 di->lease_shared_gen = 0;
1870         return valid;
1871 }
1872
1873 /*
1874  * Check if directory-wide content lease/cap is valid.
1875  */
1876 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1877                               struct ceph_mds_client *mdsc)
1878 {
1879         struct ceph_inode_info *ci = ceph_inode(dir);
1880         struct ceph_client *cl = mdsc->fsc->client;
1881         int valid;
1882         int shared_gen;
1883
1884         spin_lock(&ci->i_ceph_lock);
1885         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1886         if (valid) {
1887                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1888                 shared_gen = atomic_read(&ci->i_shared_gen);
1889         }
1890         spin_unlock(&ci->i_ceph_lock);
1891         if (valid) {
1892                 struct ceph_dentry_info *di;
1893                 spin_lock(&dentry->d_lock);
1894                 di = ceph_dentry(dentry);
1895                 if (dir == d_inode(dentry->d_parent) &&
1896                     di && di->lease_shared_gen == shared_gen)
1897                         __ceph_dentry_dir_lease_touch(di);
1898                 else
1899                         valid = 0;
1900                 spin_unlock(&dentry->d_lock);
1901         }
1902         doutc(cl, "dir %p %llx.%llx v%u dentry %p '%pd' = %d\n", dir,
1903               ceph_vinop(dir), (unsigned)atomic_read(&ci->i_shared_gen),
1904               dentry, dentry, valid);
1905         return valid;
1906 }
1907
1908 /*
1909  * Check if cached dentry can be trusted.
1910  */
1911 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1912 {
1913         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dentry->d_sb)->mdsc;
1914         struct ceph_client *cl = mdsc->fsc->client;
1915         int valid = 0;
1916         struct dentry *parent;
1917         struct inode *dir, *inode;
1918
1919         valid = fscrypt_d_revalidate(dentry, flags);
1920         if (valid <= 0)
1921                 return valid;
1922
1923         if (flags & LOOKUP_RCU) {
1924                 parent = READ_ONCE(dentry->d_parent);
1925                 dir = d_inode_rcu(parent);
1926                 if (!dir)
1927                         return -ECHILD;
1928                 inode = d_inode_rcu(dentry);
1929         } else {
1930                 parent = dget_parent(dentry);
1931                 dir = d_inode(parent);
1932                 inode = d_inode(dentry);
1933         }
1934
1935         doutc(cl, "%p '%pd' inode %p offset 0x%llx nokey %d\n",
1936               dentry, dentry, inode, ceph_dentry(dentry)->offset,
1937               !!(dentry->d_flags & DCACHE_NOKEY_NAME));
1938
1939         mdsc = ceph_sb_to_fs_client(dir->i_sb)->mdsc;
1940
1941         /* always trust cached snapped dentries, snapdir dentry */
1942         if (ceph_snap(dir) != CEPH_NOSNAP) {
1943                 doutc(cl, "%p '%pd' inode %p is SNAPPED\n", dentry,
1944                       dentry, inode);
1945                 valid = 1;
1946         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1947                 valid = 1;
1948         } else {
1949                 valid = dentry_lease_is_valid(dentry, flags);
1950                 if (valid == -ECHILD)
1951                         return valid;
1952                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1953                         if (inode)
1954                                 valid = ceph_is_any_caps(inode);
1955                         else
1956                                 valid = 1;
1957                 }
1958         }
1959
1960         if (!valid) {
1961                 struct ceph_mds_request *req;
1962                 int op, err;
1963                 u32 mask;
1964
1965                 if (flags & LOOKUP_RCU)
1966                         return -ECHILD;
1967
1968                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1969
1970                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1971                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1972                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1973                 if (!IS_ERR(req)) {
1974                         req->r_dentry = dget(dentry);
1975                         req->r_num_caps = 2;
1976                         req->r_parent = dir;
1977                         ihold(dir);
1978
1979                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1980                         if (ceph_security_xattr_wanted(dir))
1981                                 mask |= CEPH_CAP_XATTR_SHARED;
1982                         req->r_args.getattr.mask = cpu_to_le32(mask);
1983
1984                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1985                         switch (err) {
1986                         case 0:
1987                                 if (d_really_is_positive(dentry) &&
1988                                     d_inode(dentry) == req->r_target_inode)
1989                                         valid = 1;
1990                                 break;
1991                         case -ENOENT:
1992                                 if (d_really_is_negative(dentry))
1993                                         valid = 1;
1994                                 fallthrough;
1995                         default:
1996                                 break;
1997                         }
1998                         ceph_mdsc_put_request(req);
1999                         doutc(cl, "%p '%pd', lookup result=%d\n", dentry,
2000                               dentry, err);
2001                 }
2002         } else {
2003                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
2004         }
2005
2006         doutc(cl, "%p '%pd' %s\n", dentry, dentry, valid ? "valid" : "invalid");
2007         if (!valid)
2008                 ceph_dir_clear_complete(dir);
2009
2010         if (!(flags & LOOKUP_RCU))
2011                 dput(parent);
2012         return valid;
2013 }
2014
2015 /*
2016  * Delete unused dentry that doesn't have valid lease
2017  *
2018  * Called under dentry->d_lock.
2019  */
2020 static int ceph_d_delete(const struct dentry *dentry)
2021 {
2022         struct ceph_dentry_info *di;
2023
2024         /* won't release caps */
2025         if (d_really_is_negative(dentry))
2026                 return 0;
2027         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
2028                 return 0;
2029         /* vaild lease? */
2030         di = ceph_dentry(dentry);
2031         if (di) {
2032                 if (__dentry_lease_is_valid(di))
2033                         return 0;
2034                 if (__dir_lease_try_check(dentry))
2035                         return 0;
2036         }
2037         return 1;
2038 }
2039
2040 /*
2041  * Release our ceph_dentry_info.
2042  */
2043 static void ceph_d_release(struct dentry *dentry)
2044 {
2045         struct ceph_dentry_info *di = ceph_dentry(dentry);
2046         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
2047
2048         doutc(fsc->client, "dentry %p '%pd'\n", dentry, dentry);
2049
2050         atomic64_dec(&fsc->mdsc->metric.total_dentries);
2051
2052         spin_lock(&dentry->d_lock);
2053         __dentry_lease_unlist(di);
2054         dentry->d_fsdata = NULL;
2055         spin_unlock(&dentry->d_lock);
2056
2057         ceph_put_mds_session(di->lease_session);
2058         kmem_cache_free(ceph_dentry_cachep, di);
2059 }
2060
2061 /*
2062  * When the VFS prunes a dentry from the cache, we need to clear the
2063  * complete flag on the parent directory.
2064  *
2065  * Called under dentry->d_lock.
2066  */
2067 static void ceph_d_prune(struct dentry *dentry)
2068 {
2069         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dentry->d_sb);
2070         struct ceph_client *cl = mdsc->fsc->client;
2071         struct ceph_inode_info *dir_ci;
2072         struct ceph_dentry_info *di;
2073
2074         doutc(cl, "dentry %p '%pd'\n", dentry, dentry);
2075
2076         /* do we have a valid parent? */
2077         if (IS_ROOT(dentry))
2078                 return;
2079
2080         /* we hold d_lock, so d_parent is stable */
2081         dir_ci = ceph_inode(d_inode(dentry->d_parent));
2082         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
2083                 return;
2084
2085         /* who calls d_delete() should also disable dcache readdir */
2086         if (d_really_is_negative(dentry))
2087                 return;
2088
2089         /* d_fsdata does not get cleared until d_release */
2090         if (!d_unhashed(dentry)) {
2091                 __ceph_dir_clear_complete(dir_ci);
2092                 return;
2093         }
2094
2095         /* Disable dcache readdir just in case that someone called d_drop()
2096          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
2097          * properly (dcache readdir is still enabled) */
2098         di = ceph_dentry(dentry);
2099         if (di->offset > 0 &&
2100             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
2101                 __ceph_dir_clear_ordered(dir_ci);
2102 }
2103
2104 /*
2105  * read() on a dir.  This weird interface hack only works if mounted
2106  * with '-o dirstat'.
2107  */
2108 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
2109                              loff_t *ppos)
2110 {
2111         struct ceph_dir_file_info *dfi = file->private_data;
2112         struct inode *inode = file_inode(file);
2113         struct ceph_inode_info *ci = ceph_inode(inode);
2114         int left;
2115         const int bufsize = 1024;
2116
2117         if (!ceph_test_mount_opt(ceph_sb_to_fs_client(inode->i_sb), DIRSTAT))
2118                 return -EISDIR;
2119
2120         if (!dfi->dir_info) {
2121                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
2122                 if (!dfi->dir_info)
2123                         return -ENOMEM;
2124                 dfi->dir_info_len =
2125                         snprintf(dfi->dir_info, bufsize,
2126                                 "entries:   %20lld\n"
2127                                 " files:    %20lld\n"
2128                                 " subdirs:  %20lld\n"
2129                                 "rentries:  %20lld\n"
2130                                 " rfiles:   %20lld\n"
2131                                 " rsubdirs: %20lld\n"
2132                                 "rbytes:    %20lld\n"
2133                                 "rctime:    %10lld.%09ld\n",
2134                                 ci->i_files + ci->i_subdirs,
2135                                 ci->i_files,
2136                                 ci->i_subdirs,
2137                                 ci->i_rfiles + ci->i_rsubdirs,
2138                                 ci->i_rfiles,
2139                                 ci->i_rsubdirs,
2140                                 ci->i_rbytes,
2141                                 ci->i_rctime.tv_sec,
2142                                 ci->i_rctime.tv_nsec);
2143         }
2144
2145         if (*ppos >= dfi->dir_info_len)
2146                 return 0;
2147         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
2148         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
2149         if (left == size)
2150                 return -EFAULT;
2151         *ppos += (size - left);
2152         return size - left;
2153 }
2154
2155
2156
2157 /*
2158  * Return name hash for a given dentry.  This is dependent on
2159  * the parent directory's hash function.
2160  */
2161 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
2162 {
2163         struct ceph_inode_info *dci = ceph_inode(dir);
2164         unsigned hash;
2165
2166         switch (dci->i_dir_layout.dl_dir_hash) {
2167         case 0: /* for backward compat */
2168         case CEPH_STR_HASH_LINUX:
2169                 return dn->d_name.hash;
2170
2171         default:
2172                 spin_lock(&dn->d_lock);
2173                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
2174                                      dn->d_name.name, dn->d_name.len);
2175                 spin_unlock(&dn->d_lock);
2176                 return hash;
2177         }
2178 }
2179
2180 WRAP_DIR_ITER(ceph_readdir) // FIXME!
2181 const struct file_operations ceph_dir_fops = {
2182         .read = ceph_read_dir,
2183         .iterate_shared = shared_ceph_readdir,
2184         .llseek = ceph_dir_llseek,
2185         .open = ceph_open,
2186         .release = ceph_release,
2187         .unlocked_ioctl = ceph_ioctl,
2188         .compat_ioctl = compat_ptr_ioctl,
2189         .fsync = ceph_fsync,
2190         .lock = ceph_lock,
2191         .flock = ceph_flock,
2192 };
2193
2194 const struct file_operations ceph_snapdir_fops = {
2195         .iterate_shared = shared_ceph_readdir,
2196         .llseek = ceph_dir_llseek,
2197         .open = ceph_open,
2198         .release = ceph_release,
2199 };
2200
2201 const struct inode_operations ceph_dir_iops = {
2202         .lookup = ceph_lookup,
2203         .permission = ceph_permission,
2204         .getattr = ceph_getattr,
2205         .setattr = ceph_setattr,
2206         .listxattr = ceph_listxattr,
2207         .get_inode_acl = ceph_get_acl,
2208         .set_acl = ceph_set_acl,
2209         .mknod = ceph_mknod,
2210         .symlink = ceph_symlink,
2211         .mkdir = ceph_mkdir,
2212         .link = ceph_link,
2213         .unlink = ceph_unlink,
2214         .rmdir = ceph_unlink,
2215         .rename = ceph_rename,
2216         .create = ceph_create,
2217         .atomic_open = ceph_atomic_open,
2218 };
2219
2220 const struct inode_operations ceph_snapdir_iops = {
2221         .lookup = ceph_lookup,
2222         .permission = ceph_permission,
2223         .getattr = ceph_getattr,
2224         .mkdir = ceph_mkdir,
2225         .rmdir = ceph_unlink,
2226         .rename = ceph_rename,
2227 };
2228
2229 const struct dentry_operations ceph_dentry_ops = {
2230         .d_revalidate = ceph_d_revalidate,
2231         .d_delete = ceph_d_delete,
2232         .d_release = ceph_d_release,
2233         .d_prune = ceph_d_prune,
2234         .d_init = ceph_d_init,
2235 };