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 != cache_ctl->page->index) {
 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_children 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 linear 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 sorted 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                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 711                 }
 712                 retval = offset;
 713         }
 714 out:
 715         inode_unlock(inode);
 716         return retval;
 717 }
 718
 719 /*
 720  * Handle lookups for the hidden .snap directory.
 721  */
 722 struct dentry *ceph_handle_snapdir(struct ceph_mds_request *req,
 723                                    struct dentry *dentry)
 724 {
 725         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
 726         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_rwsem */
 727         struct ceph_client *cl = ceph_inode_to_client(parent);
 728
 729         /* .snap dir? */
 730         if (ceph_snap(parent) == CEPH_NOSNAP &&
 731             strcmp(dentry->d_name.name, fsc->mount_options->snapdir_name) == 0) {
 732                 struct dentry *res;
 733                 struct inode *inode = ceph_get_snapdir(parent);
 734
 735                 res = d_splice_alias(inode, dentry);
 736                 doutc(cl, "ENOENT on snapdir %p '%pd', linking to "
 737                       "snapdir %p %llx.%llx. Spliced dentry %p\n",
 738                       dentry, dentry, inode, ceph_vinop(inode), res);
 739                 if (res)
 740                         dentry = res;
 741         }
 742         return dentry;
 743 }
 744
 745 /*
 746  * Figure out final result of a lookup/open request.
 747  *
 748  * Mainly, make sure we return the final req->r_dentry (if it already
 749  * existed) in place of the original VFS-provided dentry when they
 750  * differ.
 751  *
 752  * Gracefully handle the case where the MDS replies with -ENOENT and
 753  * no trace (which it may do, at its discretion, e.g., if it doesn't
 754  * care to issue a lease on the negative dentry).
 755  */
 756 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 757                                   struct dentry *dentry, int err)
 758 {
 759         struct ceph_client *cl = req->r_mdsc->fsc->client;
 760
 761         if (err == -ENOENT) {
 762                 /* no trace? */
 763                 err = 0;
 764                 if (!req->r_reply_info.head->is_dentry) {
 765                         doutc(cl,
 766                               "ENOENT and no trace, dentry %p inode %llx.%llx\n",
 767                               dentry, ceph_vinop(d_inode(dentry)));
 768                         if (d_really_is_positive(dentry)) {
 769                                 d_drop(dentry);
 770                                 err = -ENOENT;
 771                         } else {
 772                                 d_add(dentry, NULL);
 773                         }
 774                 }
 775         }
 776         if (err)
 777                 dentry = ERR_PTR(err);
 778         else if (dentry != req->r_dentry)
 779                 dentry = dget(req->r_dentry);   /* we got spliced */
 780         else
 781                 dentry = NULL;
 782         return dentry;
 783 }
 784
 785 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 786 {
 787         return ceph_ino(inode) == CEPH_INO_ROOT &&
 788                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 789 }
 790
 791 /*
 792  * Look up a single dir entry.  If there is a lookup intent, inform
 793  * the MDS so that it gets our 'caps wanted' value in a single op.
 794  */
 795 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 796                                   unsigned int flags)
 797 {
 798         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
 799         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 800         struct ceph_client *cl = fsc->client;
 801         struct ceph_mds_request *req;
 802         int op;
 803         int mask;
 804         int err;
 805
 806         doutc(cl, "%p %llx.%llx/'%pd' dentry %p\n", dir, ceph_vinop(dir),
 807               dentry, dentry);
 808
 809         if (dentry->d_name.len > NAME_MAX)
 810                 return ERR_PTR(-ENAMETOOLONG);
 811
 812         if (IS_ENCRYPTED(dir)) {
 813                 bool had_key = fscrypt_has_encryption_key(dir);
 814
 815                 err = fscrypt_prepare_lookup_partial(dir, dentry);
 816                 if (err < 0)
 817                         return ERR_PTR(err);
 818
 819                 /* mark directory as incomplete if it has been unlocked */
 820                 if (!had_key && fscrypt_has_encryption_key(dir))
 821                         ceph_dir_clear_complete(dir);
 822         }
 823
 824         /* can we conclude ENOENT locally? */
 825         if (d_really_is_negative(dentry)) {
 826                 struct ceph_inode_info *ci = ceph_inode(dir);
 827                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 828
 829                 spin_lock(&ci->i_ceph_lock);
 830                 doutc(cl, " dir %llx.%llx flags are 0x%lx\n",
 831                       ceph_vinop(dir), ci->i_ceph_flags);
 832                 if (strncmp(dentry->d_name.name,
 833                             fsc->mount_options->snapdir_name,
 834                             dentry->d_name.len) &&
 835                     !is_root_ceph_dentry(dir, dentry) &&
 836                     ceph_test_mount_opt(fsc, DCACHE) &&
 837                     __ceph_dir_is_complete(ci) &&
 838                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 839                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 840                         spin_unlock(&ci->i_ceph_lock);
 841                         doutc(cl, " dir %llx.%llx complete, -ENOENT\n",
 842                               ceph_vinop(dir));
 843                         d_add(dentry, NULL);
 844                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 845                         return NULL;
 846                 }
 847                 spin_unlock(&ci->i_ceph_lock);
 848         }
 849
 850         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 851                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 852         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 853         if (IS_ERR(req))
 854                 return ERR_CAST(req);
 855         req->r_dentry = dget(dentry);
 856         req->r_num_caps = 2;
 857
 858         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 859         if (ceph_security_xattr_wanted(dir))
 860                 mask |= CEPH_CAP_XATTR_SHARED;
 861         req->r_args.getattr.mask = cpu_to_le32(mask);
 862
 863         ihold(dir);
 864         req->r_parent = dir;
 865         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 866         err = ceph_mdsc_do_request(mdsc, NULL, req);
 867         if (err == -ENOENT) {
 868                 struct dentry *res;
 869
 870                 res = ceph_handle_snapdir(req, dentry);
 871                 if (IS_ERR(res)) {
 872                         err = PTR_ERR(res);
 873                 } else {
 874                         dentry = res;
 875                         err = 0;
 876                 }
 877         }
 878         dentry = ceph_finish_lookup(req, dentry, err);
 879         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 880         doutc(cl, "result=%p\n", dentry);
 881         return dentry;
 882 }
 883
 884 /*
 885  * If we do a create but get no trace back from the MDS, follow up with
 886  * a lookup (the VFS expects us to link up the provided dentry).
 887  */
 888 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 889 {
 890         struct dentry *result = ceph_lookup(dir, dentry, 0);
 891
 892         if (result && !IS_ERR(result)) {
 893                 /*
 894                  * We created the item, then did a lookup, and found
 895                  * it was already linked to another inode we already
 896                  * had in our cache (and thus got spliced). To not
 897                  * confuse VFS (especially when inode is a directory),
 898                  * we don't link our dentry to that inode, return an
 899                  * error instead.
 900                  *
 901                  * This event should be rare and it happens only when
 902                  * we talk to old MDS. Recent MDS does not send traceless
 903                  * reply for request that creates new inode.
 904                  */
 905                 d_drop(result);
 906                 return -ESTALE;
 907         }
 908         return PTR_ERR(result);
 909 }
 910
 911 static int ceph_mknod(struct mnt_idmap *idmap, struct inode *dir,
 912                       struct dentry *dentry, umode_t mode, dev_t rdev)
 913 {
 914         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 915         struct ceph_client *cl = mdsc->fsc->client;
 916         struct ceph_mds_request *req;
 917         struct ceph_acl_sec_ctx as_ctx = {};
 918         int err;
 919
 920         if (ceph_snap(dir) != CEPH_NOSNAP)
 921                 return -EROFS;
 922
 923         err = ceph_wait_on_conflict_unlink(dentry);
 924         if (err)
 925                 return err;
 926
 927         if (ceph_quota_is_max_files_exceeded(dir)) {
 928                 err = -EDQUOT;
 929                 goto out;
 930         }
 931
 932         doutc(cl, "%p %llx.%llx/'%pd' dentry %p mode 0%ho rdev %d\n",
 933               dir, ceph_vinop(dir), dentry, dentry, mode, rdev);
 934         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 935         if (IS_ERR(req)) {
 936                 err = PTR_ERR(req);
 937                 goto out;
 938         }
 939
 940         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
 941         if (IS_ERR(req->r_new_inode)) {
 942                 err = PTR_ERR(req->r_new_inode);
 943                 req->r_new_inode = NULL;
 944                 goto out_req;
 945         }
 946
 947         if (S_ISREG(mode) && IS_ENCRYPTED(dir))
 948                 set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
 949
 950         req->r_dentry = dget(dentry);
 951         req->r_num_caps = 2;
 952         req->r_parent = dir;
 953         ihold(dir);
 954         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 955         req->r_mnt_idmap = mnt_idmap_get(idmap);
 956         req->r_args.mknod.mode = cpu_to_le32(mode);
 957         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 958         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
 959                              CEPH_CAP_XATTR_EXCL;
 960         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 961
 962         ceph_as_ctx_to_req(req, &as_ctx);
 963
 964         err = ceph_mdsc_do_request(mdsc, dir, req);
 965         if (!err && !req->r_reply_info.head->is_dentry)
 966                 err = ceph_handle_notrace_create(dir, dentry);
 967 out_req:
 968         ceph_mdsc_put_request(req);
 969 out:
 970         if (!err)
 971                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 972         else
 973                 d_drop(dentry);
 974         ceph_release_acl_sec_ctx(&as_ctx);
 975         return err;
 976 }
 977
 978 static int ceph_create(struct mnt_idmap *idmap, struct inode *dir,
 979                        struct dentry *dentry, umode_t mode, bool excl)
 980 {
 981         return ceph_mknod(idmap, dir, dentry, mode, 0);
 982 }
 983
 984 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
 985 static int prep_encrypted_symlink_target(struct ceph_mds_request *req,
 986                                          const char *dest)
 987 {
 988         int err;
 989         int len = strlen(dest);
 990         struct fscrypt_str osd_link = FSTR_INIT(NULL, 0);
 991
 992         err = fscrypt_prepare_symlink(req->r_parent, dest, len, PATH_MAX,
 993                                       &osd_link);
 994         if (err)
 995                 goto out;
 996
 997         err = fscrypt_encrypt_symlink(req->r_new_inode, dest, len, &osd_link);
 998         if (err)
 999                 goto out;
1000
1001         req->r_path2 = kmalloc(CEPH_BASE64_CHARS(osd_link.len) + 1, GFP_KERNEL);
1002         if (!req->r_path2) {
1003                 err = -ENOMEM;
1004                 goto out;
1005         }
1006
1007         len = ceph_base64_encode(osd_link.name, osd_link.len, req->r_path2);
1008         req->r_path2[len] = '\0';
1009 out:
1010         fscrypt_fname_free_buffer(&osd_link);
1011         return err;
1012 }
1013 #else
1014 static int prep_encrypted_symlink_target(struct ceph_mds_request *req,
1015                                          const char *dest)
1016 {
1017         return -EOPNOTSUPP;
1018 }
1019 #endif
1020
1021 static int ceph_symlink(struct mnt_idmap *idmap, struct inode *dir,
1022                         struct dentry *dentry, const char *dest)
1023 {
1024         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1025         struct ceph_client *cl = mdsc->fsc->client;
1026         struct ceph_mds_request *req;
1027         struct ceph_acl_sec_ctx as_ctx = {};
1028         umode_t mode = S_IFLNK | 0777;
1029         int err;
1030
1031         if (ceph_snap(dir) != CEPH_NOSNAP)
1032                 return -EROFS;
1033
1034         err = ceph_wait_on_conflict_unlink(dentry);
1035         if (err)
1036                 return err;
1037
1038         if (ceph_quota_is_max_files_exceeded(dir)) {
1039                 err = -EDQUOT;
1040                 goto out;
1041         }
1042
1043         doutc(cl, "%p %llx.%llx/'%pd' to '%s'\n", dir, ceph_vinop(dir), dentry,
1044               dest);
1045         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
1046         if (IS_ERR(req)) {
1047                 err = PTR_ERR(req);
1048                 goto out;
1049         }
1050
1051         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
1052         if (IS_ERR(req->r_new_inode)) {
1053                 err = PTR_ERR(req->r_new_inode);
1054                 req->r_new_inode = NULL;
1055                 goto out_req;
1056         }
1057
1058         req->r_parent = dir;
1059         ihold(dir);
1060
1061         if (IS_ENCRYPTED(req->r_new_inode)) {
1062                 err = prep_encrypted_symlink_target(req, dest);
1063                 if (err)
1064                         goto out_req;
1065         } else {
1066                 req->r_path2 = kstrdup(dest, GFP_KERNEL);
1067                 if (!req->r_path2) {
1068                         err = -ENOMEM;
1069                         goto out_req;
1070                 }
1071         }
1072
1073         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1074         req->r_mnt_idmap = mnt_idmap_get(idmap);
1075         req->r_dentry = dget(dentry);
1076         req->r_num_caps = 2;
1077         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
1078                              CEPH_CAP_XATTR_EXCL;
1079         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1080
1081         ceph_as_ctx_to_req(req, &as_ctx);
1082
1083         err = ceph_mdsc_do_request(mdsc, dir, req);
1084         if (!err && !req->r_reply_info.head->is_dentry)
1085                 err = ceph_handle_notrace_create(dir, dentry);
1086 out_req:
1087         ceph_mdsc_put_request(req);
1088 out:
1089         if (err)
1090                 d_drop(dentry);
1091         ceph_release_acl_sec_ctx(&as_ctx);
1092         return err;
1093 }
1094
1095 static int ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1096                       struct dentry *dentry, umode_t mode)
1097 {
1098         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1099         struct ceph_client *cl = mdsc->fsc->client;
1100         struct ceph_mds_request *req;
1101         struct ceph_acl_sec_ctx as_ctx = {};
1102         int err;
1103         int op;
1104
1105         err = ceph_wait_on_conflict_unlink(dentry);
1106         if (err)
1107                 return err;
1108
1109         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1110                 /* mkdir .snap/foo is a MKSNAP */
1111                 op = CEPH_MDS_OP_MKSNAP;
1112                 doutc(cl, "mksnap %llx.%llx/'%pd' dentry %p\n",
1113                       ceph_vinop(dir), dentry, dentry);
1114         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1115                 doutc(cl, "mkdir %llx.%llx/'%pd' dentry %p mode 0%ho\n",
1116                       ceph_vinop(dir), dentry, dentry, mode);
1117                 op = CEPH_MDS_OP_MKDIR;
1118         } else {
1119                 err = -EROFS;
1120                 goto out;
1121         }
1122
1123         if (op == CEPH_MDS_OP_MKDIR &&
1124             ceph_quota_is_max_files_exceeded(dir)) {
1125                 err = -EDQUOT;
1126                 goto out;
1127         }
1128         if ((op == CEPH_MDS_OP_MKSNAP) && IS_ENCRYPTED(dir) &&
1129             !fscrypt_has_encryption_key(dir)) {
1130                 err = -ENOKEY;
1131                 goto out;
1132         }
1133
1134
1135         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1136         if (IS_ERR(req)) {
1137                 err = PTR_ERR(req);
1138                 goto out;
1139         }
1140
1141         mode |= S_IFDIR;
1142         req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
1143         if (IS_ERR(req->r_new_inode)) {
1144                 err = PTR_ERR(req->r_new_inode);
1145                 req->r_new_inode = NULL;
1146                 goto out_req;
1147         }
1148
1149         req->r_dentry = dget(dentry);
1150         req->r_num_caps = 2;
1151         req->r_parent = dir;
1152         ihold(dir);
1153         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1154         if (op == CEPH_MDS_OP_MKDIR)
1155                 req->r_mnt_idmap = mnt_idmap_get(idmap);
1156         req->r_args.mkdir.mode = cpu_to_le32(mode);
1157         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
1158                              CEPH_CAP_XATTR_EXCL;
1159         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1160
1161         ceph_as_ctx_to_req(req, &as_ctx);
1162
1163         err = ceph_mdsc_do_request(mdsc, dir, req);
1164         if (!err &&
1165             !req->r_reply_info.head->is_target &&
1166             !req->r_reply_info.head->is_dentry)
1167                 err = ceph_handle_notrace_create(dir, dentry);
1168 out_req:
1169         ceph_mdsc_put_request(req);
1170 out:
1171         if (!err)
1172                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1173         else
1174                 d_drop(dentry);
1175         ceph_release_acl_sec_ctx(&as_ctx);
1176         return err;
1177 }
1178
1179 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1180                      struct dentry *dentry)
1181 {
1182         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1183         struct ceph_client *cl = mdsc->fsc->client;
1184         struct ceph_mds_request *req;
1185         int err;
1186
1187         if (dentry->d_flags & DCACHE_DISCONNECTED)
1188                 return -EINVAL;
1189
1190         err = ceph_wait_on_conflict_unlink(dentry);
1191         if (err)
1192                 return err;
1193
1194         if (ceph_snap(dir) != CEPH_NOSNAP)
1195                 return -EROFS;
1196
1197         err = fscrypt_prepare_link(old_dentry, dir, dentry);
1198         if (err)
1199                 return err;
1200
1201         doutc(cl, "%p %llx.%llx/'%pd' to '%pd'\n", dir, ceph_vinop(dir),
1202               old_dentry, dentry);
1203         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1204         if (IS_ERR(req)) {
1205                 d_drop(dentry);
1206                 return PTR_ERR(req);
1207         }
1208         req->r_dentry = dget(dentry);
1209         req->r_num_caps = 2;
1210         req->r_old_dentry = dget(old_dentry);
1211         /*
1212          * The old_dentry maybe a DCACHE_DISCONNECTED dentry, then we
1213          * will just pass the ino# to MDSs.
1214          */
1215         if (old_dentry->d_flags & DCACHE_DISCONNECTED)
1216                 req->r_ino2 = ceph_vino(d_inode(old_dentry));
1217         req->r_parent = dir;
1218         ihold(dir);
1219         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1220         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1221         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1222         /* release LINK_SHARED on source inode (mds will lock it) */
1223         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1224         err = ceph_mdsc_do_request(mdsc, dir, req);
1225         if (err) {
1226                 d_drop(dentry);
1227         } else if (!req->r_reply_info.head->is_dentry) {
1228                 ihold(d_inode(old_dentry));
1229                 d_instantiate(dentry, d_inode(old_dentry));
1230         }
1231         ceph_mdsc_put_request(req);
1232         return err;
1233 }
1234
1235 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1236                                  struct ceph_mds_request *req)
1237 {
1238         struct dentry *dentry = req->r_dentry;
1239         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
1240         struct ceph_client *cl = fsc->client;
1241         struct ceph_dentry_info *di = ceph_dentry(dentry);
1242         int result = req->r_err ? req->r_err :
1243                         le32_to_cpu(req->r_reply_info.head->result);
1244
1245         if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
1246                 pr_warn_client(cl,
1247                         "dentry %p:%pd async unlink bit is not set\n",
1248                         dentry, dentry);
1249
1250         spin_lock(&fsc->async_unlink_conflict_lock);
1251         hash_del_rcu(&di->hnode);
1252         spin_unlock(&fsc->async_unlink_conflict_lock);
1253
1254         spin_lock(&dentry->d_lock);
1255         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1256         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT);
1257         spin_unlock(&dentry->d_lock);
1258
1259         synchronize_rcu();
1260
1261         if (result == -EJUKEBOX)
1262                 goto out;
1263
1264         /* If op failed, mark everyone involved for errors */
1265         if (result) {
1266                 int pathlen = 0;
1267                 u64 base = 0;
1268                 char *path = ceph_mdsc_build_path(mdsc, dentry, &pathlen,
1269                                                   &base, 0);
1270
1271                 /* mark error on parent + clear complete */
1272                 mapping_set_error(req->r_parent->i_mapping, result);
1273                 ceph_dir_clear_complete(req->r_parent);
1274
1275                 /* drop the dentry -- we don't know its status */
1276                 if (!d_unhashed(dentry))
1277                         d_drop(dentry);
1278
1279                 /* mark inode itself for an error (since metadata is bogus) */
1280                 mapping_set_error(req->r_old_inode->i_mapping, result);
1281
1282                 pr_warn_client(cl, "failure path=(%llx)%s result=%d!\n",
1283                                base, IS_ERR(path) ? "<<bad>>" : path, result);
1284                 ceph_mdsc_free_path(path, pathlen);
1285         }
1286 out:
1287         iput(req->r_old_inode);
1288         ceph_mdsc_release_dir_caps(req);
1289 }
1290
1291 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1292 {
1293         struct ceph_inode_info *ci = ceph_inode(dir);
1294         struct ceph_dentry_info *di;
1295         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1296
1297         spin_lock(&ci->i_ceph_lock);
1298         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1299                 ceph_take_cap_refs(ci, want, false);
1300                 got = want;
1301         }
1302         spin_unlock(&ci->i_ceph_lock);
1303
1304         /* If we didn't get anything, return 0 */
1305         if (!got)
1306                 return 0;
1307
1308         spin_lock(&dentry->d_lock);
1309         di = ceph_dentry(dentry);
1310         /*
1311          * - We are holding Fx, which implies Fs caps.
1312          * - Only support async unlink for primary linkage
1313          */
1314         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1315             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1316                 want = 0;
1317         spin_unlock(&dentry->d_lock);
1318
1319         /* Do we still want what we've got? */
1320         if (want == got)
1321                 return got;
1322
1323         ceph_put_cap_refs(ci, got);
1324         return 0;
1325 }
1326
1327 /*
1328  * rmdir and unlink are differ only by the metadata op code
1329  */
1330 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1331 {
1332         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
1333         struct ceph_client *cl = fsc->client;
1334         struct ceph_mds_client *mdsc = fsc->mdsc;
1335         struct inode *inode = d_inode(dentry);
1336         struct ceph_mds_request *req;
1337         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1338         struct dentry *dn;
1339         int err = -EROFS;
1340         int op;
1341         char *path;
1342         int pathlen;
1343         u64 pathbase;
1344
1345         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1346                 /* rmdir .snap/foo is RMSNAP */
1347                 doutc(cl, "rmsnap %llx.%llx/'%pd' dn\n", ceph_vinop(dir),
1348                       dentry);
1349                 op = CEPH_MDS_OP_RMSNAP;
1350         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1351                 doutc(cl, "unlink/rmdir %llx.%llx/'%pd' inode %llx.%llx\n",
1352                       ceph_vinop(dir), dentry, ceph_vinop(inode));
1353                 op = d_is_dir(dentry) ?
1354                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1355         } else
1356                 goto out;
1357
1358         dn = d_find_alias(dir);
1359         if (!dn) {
1360                 try_async = false;
1361         } else {
1362                 path = ceph_mdsc_build_path(mdsc, dn, &pathlen, &pathbase, 0);
1363                 if (IS_ERR(path)) {
1364                         try_async = false;
1365                         err = 0;
1366                 } else {
1367                         err = ceph_mds_check_access(mdsc, path, MAY_WRITE);
1368                 }
1369                 ceph_mdsc_free_path(path, pathlen);
1370                 dput(dn);
1371
1372                 /* For none EACCES cases will let the MDS do the mds auth check */
1373                 if (err == -EACCES) {
1374                         return err;
1375                 } else if (err < 0) {
1376                         try_async = false;
1377                         err = 0;
1378                 }
1379         }
1380
1381 retry:
1382         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1383         if (IS_ERR(req)) {
1384                 err = PTR_ERR(req);
1385                 goto out;
1386         }
1387         req->r_dentry = dget(dentry);
1388         req->r_num_caps = 2;
1389         req->r_parent = dir;
1390         ihold(dir);
1391         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1392         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1393         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1394
1395         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1396             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1397                 struct ceph_dentry_info *di = ceph_dentry(dentry);
1398
1399                 doutc(cl, "async unlink on %llx.%llx/'%pd' caps=%s",
1400                       ceph_vinop(dir), dentry,
1401                       ceph_cap_string(req->r_dir_caps));
1402                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1403                 req->r_callback = ceph_async_unlink_cb;
1404                 req->r_old_inode = d_inode(dentry);
1405                 ihold(req->r_old_inode);
1406
1407                 spin_lock(&dentry->d_lock);
1408                 di->flags |= CEPH_DENTRY_ASYNC_UNLINK;
1409                 spin_unlock(&dentry->d_lock);
1410
1411                 spin_lock(&fsc->async_unlink_conflict_lock);
1412                 hash_add_rcu(fsc->async_unlink_conflict, &di->hnode,
1413                              dentry->d_name.hash);
1414                 spin_unlock(&fsc->async_unlink_conflict_lock);
1415
1416                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1417                 if (!err) {
1418                         /*
1419                          * We have enough caps, so we assume that the unlink
1420                          * will succeed. Fix up the target inode and dcache.
1421                          */
1422                         drop_nlink(inode);
1423                         d_delete(dentry);
1424                 } else {
1425                         spin_lock(&fsc->async_unlink_conflict_lock);
1426                         hash_del_rcu(&di->hnode);
1427                         spin_unlock(&fsc->async_unlink_conflict_lock);
1428
1429                         spin_lock(&dentry->d_lock);
1430                         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1431                         spin_unlock(&dentry->d_lock);
1432
1433                         if (err == -EJUKEBOX) {
1434                                 try_async = false;
1435                                 ceph_mdsc_put_request(req);
1436                                 goto retry;
1437                         }
1438                 }
1439         } else {
1440                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1441                 err = ceph_mdsc_do_request(mdsc, dir, req);
1442                 if (!err && !req->r_reply_info.head->is_dentry)
1443                         d_delete(dentry);
1444         }
1445
1446         ceph_mdsc_put_request(req);
1447 out:
1448         return err;
1449 }
1450
1451 static int ceph_rename(struct mnt_idmap *idmap, struct inode *old_dir,
1452                        struct dentry *old_dentry, struct inode *new_dir,
1453                        struct dentry *new_dentry, unsigned int flags)
1454 {
1455         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1456         struct ceph_client *cl = mdsc->fsc->client;
1457         struct ceph_mds_request *req;
1458         int op = CEPH_MDS_OP_RENAME;
1459         int err;
1460
1461         if (flags)
1462                 return -EINVAL;
1463
1464         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1465                 return -EXDEV;
1466         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1467                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1468                         op = CEPH_MDS_OP_RENAMESNAP;
1469                 else
1470                         return -EROFS;
1471         }
1472         /* don't allow cross-quota renames */
1473         if ((old_dir != new_dir) &&
1474             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1475                 return -EXDEV;
1476
1477         err = ceph_wait_on_conflict_unlink(new_dentry);
1478         if (err)
1479                 return err;
1480
1481         err = fscrypt_prepare_rename(old_dir, old_dentry, new_dir, new_dentry,
1482                                      flags);
1483         if (err)
1484                 return err;
1485
1486         doutc(cl, "%llx.%llx/'%pd' to %llx.%llx/'%pd'\n",
1487               ceph_vinop(old_dir), old_dentry, ceph_vinop(new_dir),
1488               new_dentry);
1489         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1490         if (IS_ERR(req))
1491                 return PTR_ERR(req);
1492         ihold(old_dir);
1493         req->r_dentry = dget(new_dentry);
1494         req->r_num_caps = 2;
1495         req->r_old_dentry = dget(old_dentry);
1496         req->r_old_dentry_dir = old_dir;
1497         req->r_parent = new_dir;
1498         ihold(new_dir);
1499         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1500         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1501         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1502         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_XATTR_EXCL;
1503         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1504         /* release LINK_RDCACHE on source inode (mds will lock it) */
1505         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1506         if (d_really_is_positive(new_dentry)) {
1507                 req->r_inode_drop =
1508                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1509         }
1510         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1511         if (!err && !req->r_reply_info.head->is_dentry) {
1512                 /*
1513                  * Normally d_move() is done by fill_trace (called by
1514                  * do_request, above).  If there is no trace, we need
1515                  * to do it here.
1516                  */
1517                 d_move(old_dentry, new_dentry);
1518         }
1519         ceph_mdsc_put_request(req);
1520         return err;
1521 }
1522
1523 /*
1524  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1525  * Leases at front of the list will expire first. (Assume all leases have
1526  * similar duration)
1527  *
1528  * Called under dentry->d_lock.
1529  */
1530 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1531 {
1532         struct dentry *dn = di->dentry;
1533         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dn->d_sb)->mdsc;
1534         struct ceph_client *cl = mdsc->fsc->client;
1535
1536         doutc(cl, "%p %p '%pd'\n", di, dn, dn);
1537
1538         di->flags |= CEPH_DENTRY_LEASE_LIST;
1539         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1540                 di->flags |= CEPH_DENTRY_REFERENCED;
1541                 return;
1542         }
1543
1544         spin_lock(&mdsc->dentry_list_lock);
1545         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1546         spin_unlock(&mdsc->dentry_list_lock);
1547 }
1548
1549 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1550                                      struct ceph_dentry_info *di)
1551 {
1552         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1553         di->lease_gen = 0;
1554         di->time = jiffies;
1555         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1556 }
1557
1558 /*
1559  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1560  * list if it's not in the list, otherwise set 'referenced' flag.
1561  *
1562  * Called under dentry->d_lock.
1563  */
1564 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1565 {
1566         struct dentry *dn = di->dentry;
1567         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dn->d_sb)->mdsc;
1568         struct ceph_client *cl = mdsc->fsc->client;
1569
1570         doutc(cl, "%p %p '%pd' (offset 0x%llx)\n", di, dn, dn, di->offset);
1571
1572         if (!list_empty(&di->lease_list)) {
1573                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1574                         /* don't remove dentry from dentry lease list
1575                          * if its lease is valid */
1576                         if (__dentry_lease_is_valid(di))
1577                                 return;
1578                 } else {
1579                         di->flags |= CEPH_DENTRY_REFERENCED;
1580                         return;
1581                 }
1582         }
1583
1584         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1585                 di->flags |= CEPH_DENTRY_REFERENCED;
1586                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1587                 return;
1588         }
1589
1590         spin_lock(&mdsc->dentry_list_lock);
1591         __dentry_dir_lease_touch(mdsc, di);
1592         spin_unlock(&mdsc->dentry_list_lock);
1593 }
1594
1595 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1596 {
1597         struct ceph_mds_client *mdsc;
1598         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1599                 return;
1600         if (list_empty(&di->lease_list))
1601                 return;
1602
1603         mdsc = ceph_sb_to_fs_client(di->dentry->d_sb)->mdsc;
1604         spin_lock(&mdsc->dentry_list_lock);
1605         list_del_init(&di->lease_list);
1606         spin_unlock(&mdsc->dentry_list_lock);
1607 }
1608
1609 enum {
1610         KEEP    = 0,
1611         DELETE  = 1,
1612         TOUCH   = 2,
1613         STOP    = 4,
1614 };
1615
1616 struct ceph_lease_walk_control {
1617         bool dir_lease;
1618         bool expire_dir_lease;
1619         unsigned long nr_to_scan;
1620         unsigned long dir_lease_ttl;
1621 };
1622
1623 static int __dir_lease_check(const struct dentry *, struct ceph_lease_walk_control *);
1624 static int __dentry_lease_check(const struct dentry *);
1625
1626 static unsigned long
1627 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1628                      struct ceph_lease_walk_control *lwc)
1629 {
1630         struct ceph_dentry_info *di, *tmp;
1631         struct dentry *dentry, *last = NULL;
1632         struct list_head* list;
1633         LIST_HEAD(dispose);
1634         unsigned long freed = 0;
1635         int ret = 0;
1636
1637         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1638         spin_lock(&mdsc->dentry_list_lock);
1639         list_for_each_entry_safe(di, tmp, list, lease_list) {
1640                 if (!lwc->nr_to_scan)
1641                         break;
1642                 --lwc->nr_to_scan;
1643
1644                 dentry = di->dentry;
1645                 if (last == dentry)
1646                         break;
1647
1648                 if (!spin_trylock(&dentry->d_lock))
1649                         continue;
1650
1651                 if (__lockref_is_dead(&dentry->d_lockref)) {
1652                         list_del_init(&di->lease_list);
1653                         goto next;
1654                 }
1655
1656                 if (lwc->dir_lease)
1657                         ret = __dir_lease_check(dentry, lwc);
1658                 else
1659                         ret = __dentry_lease_check(dentry);
1660                 if (ret & TOUCH) {
1661                         /* move it into tail of dir lease list */
1662                         __dentry_dir_lease_touch(mdsc, di);
1663                         if (!last)
1664                                 last = dentry;
1665                 }
1666                 if (ret & DELETE) {
1667                         /* stale lease */
1668                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1669                         if (dentry->d_lockref.count > 0) {
1670                                 /* update_dentry_lease() will re-add
1671                                  * it to lease list, or
1672                                  * ceph_d_delete() will return 1 when
1673                                  * last reference is dropped */
1674                                 list_del_init(&di->lease_list);
1675                         } else {
1676                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1677                                 list_move_tail(&di->lease_list, &dispose);
1678                                 dget_dlock(dentry);
1679                         }
1680                 }
1681 next:
1682                 spin_unlock(&dentry->d_lock);
1683                 if (ret & STOP)
1684                         break;
1685         }
1686         spin_unlock(&mdsc->dentry_list_lock);
1687
1688         while (!list_empty(&dispose)) {
1689                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1690                                       lease_list);
1691                 dentry = di->dentry;
1692                 spin_lock(&dentry->d_lock);
1693
1694                 list_del_init(&di->lease_list);
1695                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1696                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1697                         spin_lock(&mdsc->dentry_list_lock);
1698                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1699                                 list_add_tail(&di->lease_list,
1700                                               &mdsc->dentry_leases);
1701                         } else {
1702                                 __dentry_dir_lease_touch(mdsc, di);
1703                         }
1704                         spin_unlock(&mdsc->dentry_list_lock);
1705                 } else {
1706                         freed++;
1707                 }
1708
1709                 spin_unlock(&dentry->d_lock);
1710                 /* ceph_d_delete() does the trick */
1711                 dput(dentry);
1712         }
1713         return freed;
1714 }
1715
1716 static int __dentry_lease_check(const struct dentry *dentry)
1717 {
1718         struct ceph_dentry_info *di = ceph_dentry(dentry);
1719         int ret;
1720
1721         if (__dentry_lease_is_valid(di))
1722                 return STOP;
1723         ret = __dir_lease_try_check(dentry);
1724         if (ret == -EBUSY)
1725                 return KEEP;
1726         if (ret > 0)
1727                 return TOUCH;
1728         return DELETE;
1729 }
1730
1731 static int __dir_lease_check(const struct dentry *dentry,
1732                              struct ceph_lease_walk_control *lwc)
1733 {
1734         struct ceph_dentry_info *di = ceph_dentry(dentry);
1735
1736         int ret = __dir_lease_try_check(dentry);
1737         if (ret == -EBUSY)
1738                 return KEEP;
1739         if (ret > 0) {
1740                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1741                         return STOP;
1742                 /* Move dentry to tail of dir lease list if we don't want
1743                  * to delete it. So dentries in the list are checked in a
1744                  * round robin manner */
1745                 if (!lwc->expire_dir_lease)
1746                         return TOUCH;
1747                 if (dentry->d_lockref.count > 0 ||
1748                     (di->flags & CEPH_DENTRY_REFERENCED))
1749                         return TOUCH;
1750                 /* invalidate dir lease */
1751                 di->lease_shared_gen = 0;
1752         }
1753         return DELETE;
1754 }
1755
1756 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1757 {
1758         struct ceph_lease_walk_control lwc;
1759         unsigned long count;
1760         unsigned long freed;
1761
1762         spin_lock(&mdsc->caps_list_lock);
1763         if (mdsc->caps_use_max > 0 &&
1764             mdsc->caps_use_count > mdsc->caps_use_max)
1765                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1766         else
1767                 count = 0;
1768         spin_unlock(&mdsc->caps_list_lock);
1769
1770         lwc.dir_lease = false;
1771         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1772         freed = __dentry_leases_walk(mdsc, &lwc);
1773         if (!lwc.nr_to_scan) /* more invalid leases */
1774                 return -EAGAIN;
1775
1776         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1777                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1778
1779         lwc.dir_lease = true;
1780         lwc.expire_dir_lease = freed < count;
1781         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1782         freed +=__dentry_leases_walk(mdsc, &lwc);
1783         if (!lwc.nr_to_scan) /* more to check */
1784                 return -EAGAIN;
1785
1786         return freed > 0 ? 1 : 0;
1787 }
1788
1789 /*
1790  * Ensure a dentry lease will no longer revalidate.
1791  */
1792 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1793 {
1794         struct ceph_dentry_info *di = ceph_dentry(dentry);
1795         spin_lock(&dentry->d_lock);
1796         di->time = jiffies;
1797         di->lease_shared_gen = 0;
1798         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1799         __dentry_lease_unlist(di);
1800         spin_unlock(&dentry->d_lock);
1801 }
1802
1803 /*
1804  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1805  * renew if the least is more than half up.
1806  */
1807 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1808 {
1809         struct ceph_mds_session *session;
1810
1811         if (!di->lease_gen)
1812                 return false;
1813
1814         session = di->lease_session;
1815         if (session) {
1816                 u32 gen;
1817                 unsigned long ttl;
1818
1819                 gen = atomic_read(&session->s_cap_gen);
1820                 ttl = session->s_cap_ttl;
1821
1822                 if (di->lease_gen == gen &&
1823                     time_before(jiffies, ttl) &&
1824                     time_before(jiffies, di->time))
1825                         return true;
1826         }
1827         di->lease_gen = 0;
1828         return false;
1829 }
1830
1831 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1832 {
1833         struct ceph_dentry_info *di;
1834         struct ceph_mds_session *session = NULL;
1835         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dentry->d_sb)->mdsc;
1836         struct ceph_client *cl = mdsc->fsc->client;
1837         u32 seq = 0;
1838         int valid = 0;
1839
1840         spin_lock(&dentry->d_lock);
1841         di = ceph_dentry(dentry);
1842         if (di && __dentry_lease_is_valid(di)) {
1843                 valid = 1;
1844
1845                 if (di->lease_renew_after &&
1846                     time_after(jiffies, di->lease_renew_after)) {
1847                         /*
1848                          * We should renew. If we're in RCU walk mode
1849                          * though, we can't do that so just return
1850                          * -ECHILD.
1851                          */
1852                         if (flags & LOOKUP_RCU) {
1853                                 valid = -ECHILD;
1854                         } else {
1855                                 session = ceph_get_mds_session(di->lease_session);
1856                                 seq = di->lease_seq;
1857                                 di->lease_renew_after = 0;
1858                                 di->lease_renew_from = jiffies;
1859                         }
1860                 }
1861         }
1862         spin_unlock(&dentry->d_lock);
1863
1864         if (session) {
1865                 ceph_mdsc_lease_send_msg(session, dentry,
1866                                          CEPH_MDS_LEASE_RENEW, seq);
1867                 ceph_put_mds_session(session);
1868         }
1869         doutc(cl, "dentry %p = %d\n", dentry, valid);
1870         return valid;
1871 }
1872
1873 /*
1874  * Called under dentry->d_lock.
1875  */
1876 static int __dir_lease_try_check(const struct dentry *dentry)
1877 {
1878         struct ceph_dentry_info *di = ceph_dentry(dentry);
1879         struct inode *dir;
1880         struct ceph_inode_info *ci;
1881         int valid = 0;
1882
1883         if (!di->lease_shared_gen)
1884                 return 0;
1885         if (IS_ROOT(dentry))
1886                 return 0;
1887
1888         dir = d_inode(dentry->d_parent);
1889         ci = ceph_inode(dir);
1890
1891         if (spin_trylock(&ci->i_ceph_lock)) {
1892                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1893                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1894                         valid = 1;
1895                 spin_unlock(&ci->i_ceph_lock);
1896         } else {
1897                 valid = -EBUSY;
1898         }
1899
1900         if (!valid)
1901                 di->lease_shared_gen = 0;
1902         return valid;
1903 }
1904
1905 /*
1906  * Check if directory-wide content lease/cap is valid.
1907  */
1908 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1909                               struct ceph_mds_client *mdsc)
1910 {
1911         struct ceph_inode_info *ci = ceph_inode(dir);
1912         struct ceph_client *cl = mdsc->fsc->client;
1913         int valid;
1914         int shared_gen;
1915
1916         spin_lock(&ci->i_ceph_lock);
1917         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1918         if (valid) {
1919                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1920                 shared_gen = atomic_read(&ci->i_shared_gen);
1921         }
1922         spin_unlock(&ci->i_ceph_lock);
1923         if (valid) {
1924                 struct ceph_dentry_info *di;
1925                 spin_lock(&dentry->d_lock);
1926                 di = ceph_dentry(dentry);
1927                 if (dir == d_inode(dentry->d_parent) &&
1928                     di && di->lease_shared_gen == shared_gen)
1929                         __ceph_dentry_dir_lease_touch(di);
1930                 else
1931                         valid = 0;
1932                 spin_unlock(&dentry->d_lock);
1933         }
1934         doutc(cl, "dir %p %llx.%llx v%u dentry %p '%pd' = %d\n", dir,
1935               ceph_vinop(dir), (unsigned)atomic_read(&ci->i_shared_gen),
1936               dentry, dentry, valid);
1937         return valid;
1938 }
1939
1940 /*
1941  * Check if cached dentry can be trusted.
1942  */
1943 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1944 {
1945         struct ceph_mds_client *mdsc = ceph_sb_to_fs_client(dentry->d_sb)->mdsc;
1946         struct ceph_client *cl = mdsc->fsc->client;
1947         int valid = 0;
1948         struct dentry *parent;
1949         struct inode *dir, *inode;
1950
1951         valid = fscrypt_d_revalidate(dentry, flags);
1952         if (valid <= 0)
1953                 return valid;
1954
1955         if (flags & LOOKUP_RCU) {
1956                 parent = READ_ONCE(dentry->d_parent);
1957                 dir = d_inode_rcu(parent);
1958                 if (!dir)
1959                         return -ECHILD;
1960                 inode = d_inode_rcu(dentry);
1961         } else {
1962                 parent = dget_parent(dentry);
1963                 dir = d_inode(parent);
1964                 inode = d_inode(dentry);
1965         }
1966
1967         doutc(cl, "%p '%pd' inode %p offset 0x%llx nokey %d\n",
1968               dentry, dentry, inode, ceph_dentry(dentry)->offset,
1969               !!(dentry->d_flags & DCACHE_NOKEY_NAME));
1970
1971         mdsc = ceph_sb_to_fs_client(dir->i_sb)->mdsc;
1972
1973         /* always trust cached snapped dentries, snapdir dentry */
1974         if (ceph_snap(dir) != CEPH_NOSNAP) {
1975                 doutc(cl, "%p '%pd' inode %p is SNAPPED\n", dentry,
1976                       dentry, inode);
1977                 valid = 1;
1978         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1979                 valid = 1;
1980         } else {
1981                 valid = dentry_lease_is_valid(dentry, flags);
1982                 if (valid == -ECHILD)
1983                         return valid;
1984                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1985                         if (inode)
1986                                 valid = ceph_is_any_caps(inode);
1987                         else
1988                                 valid = 1;
1989                 }
1990         }
1991
1992         if (!valid) {
1993                 struct ceph_mds_request *req;
1994                 int op, err;
1995                 u32 mask;
1996
1997                 if (flags & LOOKUP_RCU)
1998                         return -ECHILD;
1999
2000                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
2001
2002                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
2003                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
2004                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
2005                 if (!IS_ERR(req)) {
2006                         req->r_dentry = dget(dentry);
2007                         req->r_num_caps = 2;
2008                         req->r_parent = dir;
2009                         ihold(dir);
2010
2011                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
2012                         if (ceph_security_xattr_wanted(dir))
2013                                 mask |= CEPH_CAP_XATTR_SHARED;
2014                         req->r_args.getattr.mask = cpu_to_le32(mask);
2015
2016                         err = ceph_mdsc_do_request(mdsc, NULL, req);
2017                         switch (err) {
2018                         case 0:
2019                                 if (d_really_is_positive(dentry) &&
2020                                     d_inode(dentry) == req->r_target_inode)
2021                                         valid = 1;
2022                                 break;
2023                         case -ENOENT:
2024                                 if (d_really_is_negative(dentry))
2025                                         valid = 1;
2026                                 fallthrough;
2027                         default:
2028                                 break;
2029                         }
2030                         ceph_mdsc_put_request(req);
2031                         doutc(cl, "%p '%pd', lookup result=%d\n", dentry,
2032                               dentry, err);
2033                 }
2034         } else {
2035                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
2036         }
2037
2038         doutc(cl, "%p '%pd' %s\n", dentry, dentry, valid ? "valid" : "invalid");
2039         if (!valid)
2040                 ceph_dir_clear_complete(dir);
2041
2042         if (!(flags & LOOKUP_RCU))
2043                 dput(parent);
2044         return valid;
2045 }
2046
2047 /*
2048  * Delete unused dentry that doesn't have valid lease
2049  *
2050  * Called under dentry->d_lock.
2051  */
2052 static int ceph_d_delete(const struct dentry *dentry)
2053 {
2054         struct ceph_dentry_info *di;
2055
2056         /* won't release caps */
2057         if (d_really_is_negative(dentry))
2058                 return 0;
2059         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
2060                 return 0;
2061         /* valid lease? */
2062         di = ceph_dentry(dentry);
2063         if (di) {
2064                 if (__dentry_lease_is_valid(di))
2065                         return 0;
2066                 if (__dir_lease_try_check(dentry))
2067                         return 0;
2068         }
2069         return 1;
2070 }
2071
2072 /*
2073  * Release our ceph_dentry_info.
2074  */
2075 static void ceph_d_release(struct dentry *dentry)
2076 {
2077         struct ceph_dentry_info *di = ceph_dentry(dentry);
2078         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
2079
2080         doutc(fsc->client, "dentry %p '%pd'\n", dentry, dentry);
2081
2082         atomic64_dec(&fsc->mdsc->metric.total_dentries);
2083
2084         spin_lock(&dentry->d_lock);
2085         __dentry_lease_unlist(di);
2086         dentry->d_fsdata = NULL;
2087         spin_unlock(&dentry->d_lock);
2088
2089         ceph_put_mds_session(di->lease_session);
2090         kmem_cache_free(ceph_dentry_cachep, di);
2091 }
2092
2093 /*
2094  * When the VFS prunes a dentry from the cache, we need to clear the
2095  * complete flag on the parent directory.
2096  *
2097  * Called under dentry->d_lock.
2098  */
2099 static void ceph_d_prune(struct dentry *dentry)
2100 {
2101         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dentry->d_sb);
2102         struct ceph_client *cl = mdsc->fsc->client;
2103         struct ceph_inode_info *dir_ci;
2104         struct ceph_dentry_info *di;
2105
2106         doutc(cl, "dentry %p '%pd'\n", dentry, dentry);
2107
2108         /* do we have a valid parent? */
2109         if (IS_ROOT(dentry))
2110                 return;
2111
2112         /* we hold d_lock, so d_parent is stable */
2113         dir_ci = ceph_inode(d_inode(dentry->d_parent));
2114         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
2115                 return;
2116
2117         /* who calls d_delete() should also disable dcache readdir */
2118         if (d_really_is_negative(dentry))
2119                 return;
2120
2121         /* d_fsdata does not get cleared until d_release */
2122         if (!d_unhashed(dentry)) {
2123                 __ceph_dir_clear_complete(dir_ci);
2124                 return;
2125         }
2126
2127         /* Disable dcache readdir just in case that someone called d_drop()
2128          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
2129          * properly (dcache readdir is still enabled) */
2130         di = ceph_dentry(dentry);
2131         if (di->offset > 0 &&
2132             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
2133                 __ceph_dir_clear_ordered(dir_ci);
2134 }
2135
2136 /*
2137  * read() on a dir.  This weird interface hack only works if mounted
2138  * with '-o dirstat'.
2139  */
2140 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
2141                              loff_t *ppos)
2142 {
2143         struct ceph_dir_file_info *dfi = file->private_data;
2144         struct inode *inode = file_inode(file);
2145         struct ceph_inode_info *ci = ceph_inode(inode);
2146         int left;
2147         const int bufsize = 1024;
2148
2149         if (!ceph_test_mount_opt(ceph_sb_to_fs_client(inode->i_sb), DIRSTAT))
2150                 return -EISDIR;
2151
2152         if (!dfi->dir_info) {
2153                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
2154                 if (!dfi->dir_info)
2155                         return -ENOMEM;
2156                 dfi->dir_info_len =
2157                         snprintf(dfi->dir_info, bufsize,
2158                                 "entries:   %20lld\n"
2159                                 " files:    %20lld\n"
2160                                 " subdirs:  %20lld\n"
2161                                 "rentries:  %20lld\n"
2162                                 " rfiles:   %20lld\n"
2163                                 " rsubdirs: %20lld\n"
2164                                 "rbytes:    %20lld\n"
2165                                 "rctime:    %10lld.%09ld\n",
2166                                 ci->i_files + ci->i_subdirs,
2167                                 ci->i_files,
2168                                 ci->i_subdirs,
2169                                 ci->i_rfiles + ci->i_rsubdirs,
2170                                 ci->i_rfiles,
2171                                 ci->i_rsubdirs,
2172                                 ci->i_rbytes,
2173                                 ci->i_rctime.tv_sec,
2174                                 ci->i_rctime.tv_nsec);
2175         }
2176
2177         if (*ppos >= dfi->dir_info_len)
2178                 return 0;
2179         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
2180         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
2181         if (left == size)
2182                 return -EFAULT;
2183         *ppos += (size - left);
2184         return size - left;
2185 }
2186
2187
2188
2189 /*
2190  * Return name hash for a given dentry.  This is dependent on
2191  * the parent directory's hash function.
2192  */
2193 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
2194 {
2195         struct ceph_inode_info *dci = ceph_inode(dir);
2196         unsigned hash;
2197
2198         switch (dci->i_dir_layout.dl_dir_hash) {
2199         case 0: /* for backward compat */
2200         case CEPH_STR_HASH_LINUX:
2201                 return dn->d_name.hash;
2202
2203         default:
2204                 spin_lock(&dn->d_lock);
2205                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
2206                                      dn->d_name.name, dn->d_name.len);
2207                 spin_unlock(&dn->d_lock);
2208                 return hash;
2209         }
2210 }
2211
2212 WRAP_DIR_ITER(ceph_readdir) // FIXME!
2213 const struct file_operations ceph_dir_fops = {
2214         .read = ceph_read_dir,
2215         .iterate_shared = shared_ceph_readdir,
2216         .llseek = ceph_dir_llseek,
2217         .open = ceph_open,
2218         .release = ceph_release,
2219         .unlocked_ioctl = ceph_ioctl,
2220         .compat_ioctl = compat_ptr_ioctl,
2221         .fsync = ceph_fsync,
2222         .lock = ceph_lock,
2223         .flock = ceph_flock,
2224 };
2225
2226 const struct file_operations ceph_snapdir_fops = {
2227         .iterate_shared = shared_ceph_readdir,
2228         .llseek = ceph_dir_llseek,
2229         .open = ceph_open,
2230         .release = ceph_release,
2231 };
2232
2233 const struct inode_operations ceph_dir_iops = {
2234         .lookup = ceph_lookup,
2235         .permission = ceph_permission,
2236         .getattr = ceph_getattr,
2237         .setattr = ceph_setattr,
2238         .listxattr = ceph_listxattr,
2239         .get_inode_acl = ceph_get_acl,
2240         .set_acl = ceph_set_acl,
2241         .mknod = ceph_mknod,
2242         .symlink = ceph_symlink,
2243         .mkdir = ceph_mkdir,
2244         .link = ceph_link,
2245         .unlink = ceph_unlink,
2246         .rmdir = ceph_unlink,
2247         .rename = ceph_rename,
2248         .create = ceph_create,
2249         .atomic_open = ceph_atomic_open,
2250 };
2251
2252 const struct inode_operations ceph_snapdir_iops = {
2253         .lookup = ceph_lookup,
2254         .permission = ceph_permission,
2255         .getattr = ceph_getattr,
2256         .mkdir = ceph_mkdir,
2257         .rmdir = ceph_unlink,
2258         .rename = ceph_rename,
2259 };
2260
2261 const struct dentry_operations ceph_dentry_ops = {
2262         .d_revalidate = ceph_d_revalidate,
2263         .d_delete = ceph_d_delete,
2264         .d_release = ceph_d_release,
2265         .d_prune = ceph_d_prune,
2266         .d_init = ceph_d_init,
2267 };