fs/ceph/inode.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/module.h>
   4 #include <linux/fs.h>
   5 #include <linux/smp_lock.h>
   6 #include <linux/slab.h>
   7 #include <linux/string.h>
   8 #include <linux/uaccess.h>
   9 #include <linux/kernel.h>
  10 #include <linux/namei.h>
  11 #include <linux/writeback.h>
  12 #include <linux/vmalloc.h>
  13
  14 #include "super.h"
  15 #include "decode.h"
  16
  17 /*
  18  * Ceph inode operations
  19  *
  20  * Implement basic inode helpers (get, alloc) and inode ops (getattr,
  21  * setattr, etc.), xattr helpers, and helpers for assimilating
  22  * metadata returned by the MDS into our cache.
  23  *
  24  * Also define helpers for doing asynchronous writeback, invalidation,
  25  * and truncation for the benefit of those who can't afford to block
  26  * (typically because they are in the message handler path).
  27  */
  28
  29 static const struct inode_operations ceph_symlink_iops;
  30
  31 static void ceph_inode_invalidate_pages(struct work_struct *work);
  32
  33 /*
  34  * find or create an inode, given the ceph ino number
  35  */
  36 struct inode *ceph_get_inode(struct super_block *sb, struct ceph_vino vino)
  37 {
  38         struct inode *inode;
  39         ino_t t = ceph_vino_to_ino(vino);
  40
  41         inode = iget5_locked(sb, t, ceph_ino_compare, ceph_set_ino_cb, &vino);
  42         if (inode == NULL)
  43                 return ERR_PTR(-ENOMEM);
  44         if (inode->i_state & I_NEW) {
  45                 dout("get_inode created new inode %p %llx.%llx ino %llx\n",
  46                      inode, ceph_vinop(inode), (u64)inode->i_ino);
  47                 unlock_new_inode(inode);
  48         }
  49
  50         dout("get_inode on %lu=%llx.%llx got %p\n", inode->i_ino, vino.ino,
  51              vino.snap, inode);
  52         return inode;
  53 }
  54
  55 /*
  56  * get/constuct snapdir inode for a given directory
  57  */
  58 struct inode *ceph_get_snapdir(struct inode *parent)
  59 {
  60         struct ceph_vino vino = {
  61                 .ino = ceph_ino(parent),
  62                 .snap = CEPH_SNAPDIR,
  63         };
  64         struct inode *inode = ceph_get_inode(parent->i_sb, vino);
  65         struct ceph_inode_info *ci = ceph_inode(inode);
  66
  67         BUG_ON(!S_ISDIR(parent->i_mode));
  68         if (IS_ERR(inode))
  69                 return ERR_PTR(PTR_ERR(inode));
  70         inode->i_mode = parent->i_mode;
  71         inode->i_uid = parent->i_uid;
  72         inode->i_gid = parent->i_gid;
  73         inode->i_op = &ceph_dir_iops;
  74         inode->i_fop = &ceph_dir_fops;
  75         ci->i_snap_caps = CEPH_CAP_PIN; /* so we can open */
  76         ci->i_rbytes = 0;
  77         return inode;
  78 }
  79
  80 const struct inode_operations ceph_file_iops = {
  81         .permission = ceph_permission,
  82         .setattr = ceph_setattr,
  83         .getattr = ceph_getattr,
  84         .setxattr = ceph_setxattr,
  85         .getxattr = ceph_getxattr,
  86         .listxattr = ceph_listxattr,
  87         .removexattr = ceph_removexattr,
  88 };
  89
  90
  91 /*
  92  * We use a 'frag tree' to keep track of the MDS's directory fragments
  93  * for a given inode (usually there is just a single fragment).  We
  94  * need to know when a child frag is delegated to a new MDS, or when
  95  * it is flagged as replicated, so we can direct our requests
  96  * accordingly.
  97  */
  98
  99 /*
 100  * find/create a frag in the tree
 101  */
 102 static struct ceph_inode_frag *__get_or_create_frag(struct ceph_inode_info *ci,
 103                                                     u32 f)
 104 {
 105         struct rb_node **p;
 106         struct rb_node *parent = NULL;
 107         struct ceph_inode_frag *frag;
 108         int c;
 109
 110         p = &ci->i_fragtree.rb_node;
 111         while (*p) {
 112                 parent = *p;
 113                 frag = rb_entry(parent, struct ceph_inode_frag, node);
 114                 c = ceph_frag_compare(f, frag->frag);
 115                 if (c < 0)
 116                         p = &(*p)->rb_left;
 117                 else if (c > 0)
 118                         p = &(*p)->rb_right;
 119                 else
 120                         return frag;
 121         }
 122
 123         frag = kmalloc(sizeof(*frag), GFP_NOFS);
 124         if (!frag) {
 125                 pr_err("__get_or_create_frag ENOMEM on %p %llx.%llx "
 126                        "frag %x\n", &ci->vfs_inode,
 127                        ceph_vinop(&ci->vfs_inode), f);
 128                 return ERR_PTR(-ENOMEM);
 129         }
 130         frag->frag = f;
 131         frag->split_by = 0;
 132         frag->mds = -1;
 133         frag->ndist = 0;
 134
 135         rb_link_node(&frag->node, parent, p);
 136         rb_insert_color(&frag->node, &ci->i_fragtree);
 137
 138         dout("get_or_create_frag added %llx.%llx frag %x\n",
 139              ceph_vinop(&ci->vfs_inode), f);
 140         return frag;
 141 }
 142
 143 /*
 144  * find a specific frag @f
 145  */
 146 struct ceph_inode_frag *__ceph_find_frag(struct ceph_inode_info *ci, u32 f)
 147 {
 148         struct rb_node *n = ci->i_fragtree.rb_node;
 149
 150         while (n) {
 151                 struct ceph_inode_frag *frag =
 152                         rb_entry(n, struct ceph_inode_frag, node);
 153                 int c = ceph_frag_compare(f, frag->frag);
 154                 if (c < 0)
 155                         n = n->rb_left;
 156                 else if (c > 0)
 157                         n = n->rb_right;
 158                 else
 159                         return frag;
 160         }
 161         return NULL;
 162 }
 163
 164 /*
 165  * Choose frag containing the given value @v.  If @pfrag is
 166  * specified, copy the frag delegation info to the caller if
 167  * it is present.
 168  */
 169 u32 ceph_choose_frag(struct ceph_inode_info *ci, u32 v,
 170                      struct ceph_inode_frag *pfrag,
 171                      int *found)
 172 {
 173         u32 t = ceph_frag_make(0, 0);
 174         struct ceph_inode_frag *frag;
 175         unsigned nway, i;
 176         u32 n;
 177
 178         if (found)
 179                 *found = 0;
 180
 181         mutex_lock(&ci->i_fragtree_mutex);
 182         while (1) {
 183                 WARN_ON(!ceph_frag_contains_value(t, v));
 184                 frag = __ceph_find_frag(ci, t);
 185                 if (!frag)
 186                         break; /* t is a leaf */
 187                 if (frag->split_by == 0) {
 188                         if (pfrag)
 189                                 memcpy(pfrag, frag, sizeof(*pfrag));
 190                         if (found)
 191                                 *found = 1;
 192                         break;
 193                 }
 194
 195                 /* choose child */
 196                 nway = 1 << frag->split_by;
 197                 dout("choose_frag(%x) %x splits by %d (%d ways)\n", v, t,
 198                      frag->split_by, nway);
 199                 for (i = 0; i < nway; i++) {
 200                         n = ceph_frag_make_child(t, frag->split_by, i);
 201                         if (ceph_frag_contains_value(n, v)) {
 202                                 t = n;
 203                                 break;
 204                         }
 205                 }
 206                 BUG_ON(i == nway);
 207         }
 208         dout("choose_frag(%x) = %x\n", v, t);
 209
 210         mutex_unlock(&ci->i_fragtree_mutex);
 211         return t;
 212 }
 213
 214 /*
 215  * Process dirfrag (delegation) info from the mds.  Include leaf
 216  * fragment in tree ONLY if ndist > 0.  Otherwise, only
 217  * branches/splits are included in i_fragtree)
 218  */
 219 static int ceph_fill_dirfrag(struct inode *inode,
 220                              struct ceph_mds_reply_dirfrag *dirinfo)
 221 {
 222         struct ceph_inode_info *ci = ceph_inode(inode);
 223         struct ceph_inode_frag *frag;
 224         u32 id = le32_to_cpu(dirinfo->frag);
 225         int mds = le32_to_cpu(dirinfo->auth);
 226         int ndist = le32_to_cpu(dirinfo->ndist);
 227         int i;
 228         int err = 0;
 229
 230         mutex_lock(&ci->i_fragtree_mutex);
 231         if (ndist == 0) {
 232                 /* no delegation info needed. */
 233                 frag = __ceph_find_frag(ci, id);
 234                 if (!frag)
 235                         goto out;
 236                 if (frag->split_by == 0) {
 237                         /* tree leaf, remove */
 238                         dout("fill_dirfrag removed %llx.%llx frag %x"
 239                              " (no ref)\n", ceph_vinop(inode), id);
 240                         rb_erase(&frag->node, &ci->i_fragtree);
 241                         kfree(frag);
 242                 } else {
 243                         /* tree branch, keep and clear */
 244                         dout("fill_dirfrag cleared %llx.%llx frag %x"
 245                              " referral\n", ceph_vinop(inode), id);
 246                         frag->mds = -1;
 247                         frag->ndist = 0;
 248                 }
 249                 goto out;
 250         }
 251
 252
 253         /* find/add this frag to store mds delegation info */
 254         frag = __get_or_create_frag(ci, id);
 255         if (IS_ERR(frag)) {
 256                 /* this is not the end of the world; we can continue
 257                    with bad/inaccurate delegation info */
 258                 pr_err("fill_dirfrag ENOMEM on mds ref %llx.%llx fg %x\n",
 259                        ceph_vinop(inode), le32_to_cpu(dirinfo->frag));
 260                 err = -ENOMEM;
 261                 goto out;
 262         }
 263
 264         frag->mds = mds;
 265         frag->ndist = min_t(u32, ndist, CEPH_MAX_DIRFRAG_REP);
 266         for (i = 0; i < frag->ndist; i++)
 267                 frag->dist[i] = le32_to_cpu(dirinfo->dist[i]);
 268         dout("fill_dirfrag %llx.%llx frag %x ndist=%d\n",
 269              ceph_vinop(inode), frag->frag, frag->ndist);
 270
 271 out:
 272         mutex_unlock(&ci->i_fragtree_mutex);
 273         return err;
 274 }
 275
 276
 277 /*
 278  * initialize a newly allocated inode.
 279  */
 280 struct inode *ceph_alloc_inode(struct super_block *sb)
 281 {
 282         struct ceph_inode_info *ci;
 283         int i;
 284
 285         ci = kmem_cache_alloc(ceph_inode_cachep, GFP_NOFS);
 286         if (!ci)
 287                 return NULL;
 288
 289         dout("alloc_inode %p\n", &ci->vfs_inode);
 290
 291         ci->i_version = 0;
 292         ci->i_time_warp_seq = 0;
 293         ci->i_ceph_flags = 0;
 294         ci->i_release_count = 0;
 295         ci->i_symlink = NULL;
 296
 297         ci->i_fragtree = RB_ROOT;
 298         mutex_init(&ci->i_fragtree_mutex);
 299
 300         ci->i_xattrs.blob = NULL;
 301         ci->i_xattrs.prealloc_blob = NULL;
 302         ci->i_xattrs.dirty = false;
 303         ci->i_xattrs.index = RB_ROOT;
 304         ci->i_xattrs.count = 0;
 305         ci->i_xattrs.names_size = 0;
 306         ci->i_xattrs.vals_size = 0;
 307         ci->i_xattrs.version = 0;
 308         ci->i_xattrs.index_version = 0;
 309
 310         ci->i_caps = RB_ROOT;
 311         ci->i_auth_cap = NULL;
 312         ci->i_dirty_caps = 0;
 313         ci->i_flushing_caps = 0;
 314         INIT_LIST_HEAD(&ci->i_dirty_item);
 315         INIT_LIST_HEAD(&ci->i_flushing_item);
 316         ci->i_cap_flush_seq = 0;
 317         ci->i_cap_flush_last_tid = 0;
 318         memset(&ci->i_cap_flush_tid, 0, sizeof(ci->i_cap_flush_tid));
 319         init_waitqueue_head(&ci->i_cap_wq);
 320         ci->i_hold_caps_min = 0;
 321         ci->i_hold_caps_max = 0;
 322         INIT_LIST_HEAD(&ci->i_cap_delay_list);
 323         ci->i_cap_exporting_mds = 0;
 324         ci->i_cap_exporting_mseq = 0;
 325         ci->i_cap_exporting_issued = 0;
 326         INIT_LIST_HEAD(&ci->i_cap_snaps);
 327         ci->i_head_snapc = NULL;
 328         ci->i_snap_caps = 0;
 329
 330         for (i = 0; i < CEPH_FILE_MODE_NUM; i++)
 331                 ci->i_nr_by_mode[i] = 0;
 332
 333         ci->i_truncate_seq = 0;
 334         ci->i_truncate_size = 0;
 335         ci->i_truncate_pending = 0;
 336
 337         ci->i_max_size = 0;
 338         ci->i_reported_size = 0;
 339         ci->i_wanted_max_size = 0;
 340         ci->i_requested_max_size = 0;
 341
 342         ci->i_pin_ref = 0;
 343         ci->i_rd_ref = 0;
 344         ci->i_rdcache_ref = 0;
 345         ci->i_wr_ref = 0;
 346         ci->i_wrbuffer_ref = 0;
 347         ci->i_wrbuffer_ref_head = 0;
 348         ci->i_shared_gen = 0;
 349         ci->i_rdcache_gen = 0;
 350         ci->i_rdcache_revoking = 0;
 351
 352         INIT_LIST_HEAD(&ci->i_unsafe_writes);
 353         INIT_LIST_HEAD(&ci->i_unsafe_dirops);
 354         spin_lock_init(&ci->i_unsafe_lock);
 355
 356         ci->i_snap_realm = NULL;
 357         INIT_LIST_HEAD(&ci->i_snap_realm_item);
 358         INIT_LIST_HEAD(&ci->i_snap_flush_item);
 359
 360         INIT_WORK(&ci->i_wb_work, ceph_inode_writeback);
 361         INIT_WORK(&ci->i_pg_inv_work, ceph_inode_invalidate_pages);
 362
 363         INIT_WORK(&ci->i_vmtruncate_work, ceph_vmtruncate_work);
 364
 365         return &ci->vfs_inode;
 366 }
 367
 368 void ceph_destroy_inode(struct inode *inode)
 369 {
 370         struct ceph_inode_info *ci = ceph_inode(inode);
 371         struct ceph_inode_frag *frag;
 372         struct rb_node *n;
 373
 374         dout("destroy_inode %p ino %llx.%llx\n", inode, ceph_vinop(inode));
 375
 376         ceph_queue_caps_release(inode);
 377
 378         kfree(ci->i_symlink);
 379         while ((n = rb_first(&ci->i_fragtree)) != NULL) {
 380                 frag = rb_entry(n, struct ceph_inode_frag, node);
 381                 rb_erase(n, &ci->i_fragtree);
 382                 kfree(frag);
 383         }
 384
 385         __ceph_destroy_xattrs(ci);
 386         if (ci->i_xattrs.blob)
 387                 ceph_buffer_put(ci->i_xattrs.blob);
 388         if (ci->i_xattrs.prealloc_blob)
 389                 ceph_buffer_put(ci->i_xattrs.prealloc_blob);
 390
 391         kmem_cache_free(ceph_inode_cachep, ci);
 392 }
 393
 394
 395 /*
 396  * Helpers to fill in size, ctime, mtime, and atime.  We have to be
 397  * careful because either the client or MDS may have more up to date
 398  * info, depending on which capabilities are held, and whether
 399  * time_warp_seq or truncate_seq have increased.  (Ordinarily, mtime
 400  * and size are monotonically increasing, except when utimes() or
 401  * truncate() increments the corresponding _seq values.)
 402  */
 403 int ceph_fill_file_size(struct inode *inode, int issued,
 404                         u32 truncate_seq, u64 truncate_size, u64 size)
 405 {
 406         struct ceph_inode_info *ci = ceph_inode(inode);
 407         int queue_trunc = 0;
 408
 409         if (ceph_seq_cmp(truncate_seq, ci->i_truncate_seq) > 0 ||
 410             (truncate_seq == ci->i_truncate_seq && size > inode->i_size)) {
 411                 dout("size %lld -> %llu\n", inode->i_size, size);
 412                 inode->i_size = size;
 413                 inode->i_blocks = (size + (1<<9) - 1) >> 9;
 414                 ci->i_reported_size = size;
 415                 if (truncate_seq != ci->i_truncate_seq) {
 416                         dout("truncate_seq %u -> %u\n",
 417                              ci->i_truncate_seq, truncate_seq);
 418                         ci->i_truncate_seq = truncate_seq;
 419                         if (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_RD|
 420                                       CEPH_CAP_FILE_WR|CEPH_CAP_FILE_BUFFER|
 421                                       CEPH_CAP_FILE_EXCL)) {
 422                                 ci->i_truncate_pending++;
 423                                 queue_trunc = 1;
 424                         }
 425                 }
 426         }
 427         if (ceph_seq_cmp(truncate_seq, ci->i_truncate_seq) >= 0 &&
 428             ci->i_truncate_size != truncate_size) {
 429                 dout("truncate_size %lld -> %llu\n", ci->i_truncate_size,
 430                      truncate_size);
 431                 ci->i_truncate_size = truncate_size;
 432         }
 433         return queue_trunc;
 434 }
 435
 436 void ceph_fill_file_time(struct inode *inode, int issued,
 437                          u64 time_warp_seq, struct timespec *ctime,
 438                          struct timespec *mtime, struct timespec *atime)
 439 {
 440         struct ceph_inode_info *ci = ceph_inode(inode);
 441         int warn = 0;
 442
 443         if (issued & (CEPH_CAP_FILE_EXCL|
 444                       CEPH_CAP_FILE_WR|
 445                       CEPH_CAP_FILE_BUFFER)) {
 446                 if (timespec_compare(ctime, &inode->i_ctime) > 0) {
 447                         dout("ctime %ld.%09ld -> %ld.%09ld inc w/ cap\n",
 448                              inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
 449                              ctime->tv_sec, ctime->tv_nsec);
 450                         inode->i_ctime = *ctime;
 451                 }
 452                 if (ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) > 0) {
 453                         /* the MDS did a utimes() */
 454                         dout("mtime %ld.%09ld -> %ld.%09ld "
 455                              "tw %d -> %d\n",
 456                              inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
 457                              mtime->tv_sec, mtime->tv_nsec,
 458                              ci->i_time_warp_seq, (int)time_warp_seq);
 459
 460                         inode->i_mtime = *mtime;
 461                         inode->i_atime = *atime;
 462                         ci->i_time_warp_seq = time_warp_seq;
 463                 } else if (time_warp_seq == ci->i_time_warp_seq) {
 464                         /* nobody did utimes(); take the max */
 465                         if (timespec_compare(mtime, &inode->i_mtime) > 0) {
 466                                 dout("mtime %ld.%09ld -> %ld.%09ld inc\n",
 467                                      inode->i_mtime.tv_sec,
 468                                      inode->i_mtime.tv_nsec,
 469                                      mtime->tv_sec, mtime->tv_nsec);
 470                                 inode->i_mtime = *mtime;
 471                         }
 472                         if (timespec_compare(atime, &inode->i_atime) > 0) {
 473                                 dout("atime %ld.%09ld -> %ld.%09ld inc\n",
 474                                      inode->i_atime.tv_sec,
 475                                      inode->i_atime.tv_nsec,
 476                                      atime->tv_sec, atime->tv_nsec);
 477                                 inode->i_atime = *atime;
 478                         }
 479                 } else if (issued & CEPH_CAP_FILE_EXCL) {
 480                         /* we did a utimes(); ignore mds values */
 481                 } else {
 482                         warn = 1;
 483                 }
 484         } else {
 485                 /* we have no write caps; whatever the MDS says is true */
 486                 if (ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) >= 0) {
 487                         inode->i_ctime = *ctime;
 488                         inode->i_mtime = *mtime;
 489                         inode->i_atime = *atime;
 490                         ci->i_time_warp_seq = time_warp_seq;
 491                 } else {
 492                         warn = 1;
 493                 }
 494         }
 495         if (warn) /* time_warp_seq shouldn't go backwards */
 496                 dout("%p mds time_warp_seq %llu < %u\n",
 497                      inode, time_warp_seq, ci->i_time_warp_seq);
 498 }
 499
 500 /*
 501  * Populate an inode based on info from mds.  May be called on new or
 502  * existing inodes.
 503  */
 504 static int fill_inode(struct inode *inode,
 505                       struct ceph_mds_reply_info_in *iinfo,
 506                       struct ceph_mds_reply_dirfrag *dirinfo,
 507                       struct ceph_mds_session *session,
 508                       unsigned long ttl_from, int cap_fmode,
 509                       struct ceph_cap_reservation *caps_reservation)
 510 {
 511         struct ceph_mds_reply_inode *info = iinfo->in;
 512         struct ceph_inode_info *ci = ceph_inode(inode);
 513         int i;
 514         int issued, implemented;
 515         struct timespec mtime, atime, ctime;
 516         u32 nsplits;
 517         struct ceph_buffer *xattr_blob = NULL;
 518         int err = 0;
 519         int queue_trunc = 0;
 520
 521         dout("fill_inode %p ino %llx.%llx v %llu had %llu\n",
 522              inode, ceph_vinop(inode), le64_to_cpu(info->version),
 523              ci->i_version);
 524
 525         /*
 526          * prealloc xattr data, if it looks like we'll need it.  only
 527          * if len > 4 (meaning there are actually xattrs; the first 4
 528          * bytes are the xattr count).
 529          */
 530         if (iinfo->xattr_len > 4) {
 531                 xattr_blob = ceph_buffer_new(iinfo->xattr_len, GFP_NOFS);
 532                 if (!xattr_blob)
 533                         pr_err("fill_inode ENOMEM xattr blob %d bytes\n",
 534                                iinfo->xattr_len);
 535         }
 536
 537         spin_lock(&inode->i_lock);
 538
 539         /*
 540          * provided version will be odd if inode value is projected,
 541          * even if stable.  skip the update if we have a newer info
 542          * (e.g., due to inode info racing form multiple MDSs), or if
 543          * we are getting projected (unstable) inode info.
 544          */
 545         if (le64_to_cpu(info->version) > 0 &&
 546             (ci->i_version & ~1) > le64_to_cpu(info->version))
 547                 goto no_change;
 548
 549         issued = __ceph_caps_issued(ci, &implemented);
 550         issued |= implemented | __ceph_caps_dirty(ci);
 551
 552         /* update inode */
 553         ci->i_version = le64_to_cpu(info->version);
 554         inode->i_version++;
 555         inode->i_rdev = le32_to_cpu(info->rdev);
 556
 557         if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
 558                 inode->i_mode = le32_to_cpu(info->mode);
 559                 inode->i_uid = le32_to_cpu(info->uid);
 560                 inode->i_gid = le32_to_cpu(info->gid);
 561                 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
 562                      inode->i_uid, inode->i_gid);
 563         }
 564
 565         if ((issued & CEPH_CAP_LINK_EXCL) == 0)
 566                 inode->i_nlink = le32_to_cpu(info->nlink);
 567
 568         /* be careful with mtime, atime, size */
 569         ceph_decode_timespec(&atime, &info->atime);
 570         ceph_decode_timespec(&mtime, &info->mtime);
 571         ceph_decode_timespec(&ctime, &info->ctime);
 572         queue_trunc = ceph_fill_file_size(inode, issued,
 573                                           le32_to_cpu(info->truncate_seq),
 574                                           le64_to_cpu(info->truncate_size),
 575                                           le64_to_cpu(info->size));
 576         ceph_fill_file_time(inode, issued,
 577                             le32_to_cpu(info->time_warp_seq),
 578                             &ctime, &mtime, &atime);
 579
 580         ci->i_max_size = le64_to_cpu(info->max_size);
 581         ci->i_layout = info->layout;
 582         inode->i_blkbits = fls(le32_to_cpu(info->layout.fl_stripe_unit)) - 1;
 583
 584         /* xattrs */
 585         /* note that if i_xattrs.len <= 4, i_xattrs.data will still be NULL. */
 586         if ((issued & CEPH_CAP_XATTR_EXCL) == 0 &&
 587             le64_to_cpu(info->xattr_version) > ci->i_xattrs.version) {
 588                 if (ci->i_xattrs.blob)
 589                         ceph_buffer_put(ci->i_xattrs.blob);
 590                 ci->i_xattrs.blob = xattr_blob;
 591                 if (xattr_blob)
 592                         memcpy(ci->i_xattrs.blob->vec.iov_base,
 593                                iinfo->xattr_data, iinfo->xattr_len);
 594                 ci->i_xattrs.version = le64_to_cpu(info->xattr_version);
 595         }
 596
 597         inode->i_mapping->a_ops = &ceph_aops;
 598         inode->i_mapping->backing_dev_info =
 599                 &ceph_client(inode->i_sb)->backing_dev_info;
 600
 601         switch (inode->i_mode & S_IFMT) {
 602         case S_IFIFO:
 603         case S_IFBLK:
 604         case S_IFCHR:
 605         case S_IFSOCK:
 606                 init_special_inode(inode, inode->i_mode, inode->i_rdev);
 607                 inode->i_op = &ceph_file_iops;
 608                 break;
 609         case S_IFREG:
 610                 inode->i_op = &ceph_file_iops;
 611                 inode->i_fop = &ceph_file_fops;
 612                 break;
 613         case S_IFLNK:
 614                 inode->i_op = &ceph_symlink_iops;
 615                 if (!ci->i_symlink) {
 616                         int symlen = iinfo->symlink_len;
 617                         char *sym;
 618
 619                         BUG_ON(symlen != inode->i_size);
 620                         spin_unlock(&inode->i_lock);
 621
 622                         err = -ENOMEM;
 623                         sym = kmalloc(symlen+1, GFP_NOFS);
 624                         if (!sym)
 625                                 goto out;
 626                         memcpy(sym, iinfo->symlink, symlen);
 627                         sym[symlen] = 0;
 628
 629                         spin_lock(&inode->i_lock);
 630                         if (!ci->i_symlink)
 631                                 ci->i_symlink = sym;
 632                         else
 633                                 kfree(sym); /* lost a race */
 634                 }
 635                 break;
 636         case S_IFDIR:
 637                 inode->i_op = &ceph_dir_iops;
 638                 inode->i_fop = &ceph_dir_fops;
 639
 640                 ci->i_files = le64_to_cpu(info->files);
 641                 ci->i_subdirs = le64_to_cpu(info->subdirs);
 642                 ci->i_rbytes = le64_to_cpu(info->rbytes);
 643                 ci->i_rfiles = le64_to_cpu(info->rfiles);
 644                 ci->i_rsubdirs = le64_to_cpu(info->rsubdirs);
 645                 ceph_decode_timespec(&ci->i_rctime, &info->rctime);
 646
 647                 /* set dir completion flag? */
 648                 if (ci->i_files == 0 && ci->i_subdirs == 0 &&
 649                     ceph_snap(inode) == CEPH_NOSNAP &&
 650                     (le32_to_cpu(info->cap.caps) & CEPH_CAP_FILE_SHARED)) {
 651                         dout(" marking %p complete (empty)\n", inode);
 652                         ci->i_ceph_flags |= CEPH_I_COMPLETE;
 653                         ci->i_max_offset = 2;
 654                 }
 655
 656                 /* it may be better to set st_size in getattr instead? */
 657                 if (ceph_test_opt(ceph_client(inode->i_sb), RBYTES))
 658                         inode->i_size = ci->i_rbytes;
 659                 break;
 660         default:
 661                 pr_err("fill_inode %llx.%llx BAD mode 0%o\n",
 662                        ceph_vinop(inode), inode->i_mode);
 663         }
 664
 665 no_change:
 666         spin_unlock(&inode->i_lock);
 667
 668         /* queue truncate if we saw i_size decrease */
 669         if (queue_trunc)
 670                 if (queue_work(ceph_client(inode->i_sb)->trunc_wq,
 671                                &ci->i_vmtruncate_work))
 672                         igrab(inode);
 673
 674         /* populate frag tree */
 675         /* FIXME: move me up, if/when version reflects fragtree changes */
 676         nsplits = le32_to_cpu(info->fragtree.nsplits);
 677         mutex_lock(&ci->i_fragtree_mutex);
 678         for (i = 0; i < nsplits; i++) {
 679                 u32 id = le32_to_cpu(info->fragtree.splits[i].frag);
 680                 struct ceph_inode_frag *frag = __get_or_create_frag(ci, id);
 681
 682                 if (IS_ERR(frag))
 683                         continue;
 684                 frag->split_by = le32_to_cpu(info->fragtree.splits[i].by);
 685                 dout(" frag %x split by %d\n", frag->frag, frag->split_by);
 686         }
 687         mutex_unlock(&ci->i_fragtree_mutex);
 688
 689         /* were we issued a capability? */
 690         if (info->cap.caps) {
 691                 if (ceph_snap(inode) == CEPH_NOSNAP) {
 692                         ceph_add_cap(inode, session,
 693                                      le64_to_cpu(info->cap.cap_id),
 694                                      cap_fmode,
 695                                      le32_to_cpu(info->cap.caps),
 696                                      le32_to_cpu(info->cap.wanted),
 697                                      le32_to_cpu(info->cap.seq),
 698                                      le32_to_cpu(info->cap.mseq),
 699                                      le64_to_cpu(info->cap.realm),
 700                                      info->cap.flags,
 701                                      caps_reservation);
 702                 } else {
 703                         spin_lock(&inode->i_lock);
 704                         dout(" %p got snap_caps %s\n", inode,
 705                              ceph_cap_string(le32_to_cpu(info->cap.caps)));
 706                         ci->i_snap_caps |= le32_to_cpu(info->cap.caps);
 707                         if (cap_fmode >= 0)
 708                                 __ceph_get_fmode(ci, cap_fmode);
 709                         spin_unlock(&inode->i_lock);
 710                 }
 711         }
 712
 713         /* update delegation info? */
 714         if (dirinfo)
 715                 ceph_fill_dirfrag(inode, dirinfo);
 716
 717         err = 0;
 718
 719 out:
 720         if (xattr_blob)
 721                 ceph_buffer_put(xattr_blob);
 722         return err;
 723 }
 724
 725 /*
 726  * caller should hold session s_mutex.
 727  */
 728 static void update_dentry_lease(struct dentry *dentry,
 729                                 struct ceph_mds_reply_lease *lease,
 730                                 struct ceph_mds_session *session,
 731                                 unsigned long from_time)
 732 {
 733         struct ceph_dentry_info *di = ceph_dentry(dentry);
 734         long unsigned duration = le32_to_cpu(lease->duration_ms);
 735         long unsigned ttl = from_time + (duration * HZ) / 1000;
 736         long unsigned half_ttl = from_time + (duration * HZ / 2) / 1000;
 737         struct inode *dir;
 738
 739         /* only track leases on regular dentries */
 740         if (dentry->d_op != &ceph_dentry_ops)
 741                 return;
 742
 743         spin_lock(&dentry->d_lock);
 744         dout("update_dentry_lease %p mask %d duration %lu ms ttl %lu\n",
 745              dentry, le16_to_cpu(lease->mask), duration, ttl);
 746
 747         /* make lease_rdcache_gen match directory */
 748         dir = dentry->d_parent->d_inode;
 749         di->lease_shared_gen = ceph_inode(dir)->i_shared_gen;
 750
 751         if (lease->mask == 0)
 752                 goto out_unlock;
 753
 754         if (di->lease_gen == session->s_cap_gen &&
 755             time_before(ttl, dentry->d_time))
 756                 goto out_unlock;  /* we already have a newer lease. */
 757
 758         if (di->lease_session && di->lease_session != session)
 759                 goto out_unlock;
 760
 761         ceph_dentry_lru_touch(dentry);
 762
 763         if (!di->lease_session)
 764                 di->lease_session = ceph_get_mds_session(session);
 765         di->lease_gen = session->s_cap_gen;
 766         di->lease_seq = le32_to_cpu(lease->seq);
 767         di->lease_renew_after = half_ttl;
 768         di->lease_renew_from = 0;
 769         dentry->d_time = ttl;
 770 out_unlock:
 771         spin_unlock(&dentry->d_lock);
 772         return;
 773 }
 774
 775 /*
 776  * splice a dentry to an inode.
 777  * caller must hold directory i_mutex for this to be safe.
 778  *
 779  * we will only rehash the resulting dentry if @prehash is
 780  * true; @prehash will be set to false (for the benefit of
 781  * the caller) if we fail.
 782  */
 783 static struct dentry *splice_dentry(struct dentry *dn, struct inode *in,
 784                                     bool *prehash)
 785 {
 786         struct dentry *realdn;
 787
 788         /* dn must be unhashed */
 789         if (!d_unhashed(dn))
 790                 d_drop(dn);
 791         realdn = d_materialise_unique(dn, in);
 792         if (IS_ERR(realdn)) {
 793                 pr_err("splice_dentry error %p inode %p ino %llx.%llx\n",
 794                        dn, in, ceph_vinop(in));
 795                 if (prehash)
 796                         *prehash = false; /* don't rehash on error */
 797                 dn = realdn; /* note realdn contains the error */
 798                 goto out;
 799         } else if (realdn) {
 800                 dout("dn %p (%d) spliced with %p (%d) "
 801                      "inode %p ino %llx.%llx\n",
 802                      dn, atomic_read(&dn->d_count),
 803                      realdn, atomic_read(&realdn->d_count),
 804                      realdn->d_inode, ceph_vinop(realdn->d_inode));
 805                 dput(dn);
 806                 dn = realdn;
 807         } else {
 808                 BUG_ON(!ceph_dentry(dn));
 809
 810                 dout("dn %p attached to %p ino %llx.%llx\n",
 811                      dn, dn->d_inode, ceph_vinop(dn->d_inode));
 812         }
 813         if ((!prehash || *prehash) && d_unhashed(dn))
 814                 d_rehash(dn);
 815 out:
 816         return dn;
 817 }
 818
 819 /*
 820  * Set dentry's directory position based on the current dir's max, and
 821  * order it in d_subdirs, so that dcache_readdir behaves.
 822  */
 823 static void ceph_set_dentry_offset(struct dentry *dn)
 824 {
 825         struct dentry *dir = dn->d_parent;
 826         struct inode *inode = dn->d_parent->d_inode;
 827         struct ceph_dentry_info *di;
 828
 829         BUG_ON(!inode);
 830
 831         di = ceph_dentry(dn);
 832
 833         spin_lock(&inode->i_lock);
 834         di->offset = ceph_inode(inode)->i_max_offset++;
 835         spin_unlock(&inode->i_lock);
 836
 837         spin_lock(&dcache_lock);
 838         spin_lock(&dn->d_lock);
 839         list_move_tail(&dir->d_subdirs, &dn->d_u.d_child);
 840         dout("set_dentry_offset %p %lld (%p %p)\n", dn, di->offset,
 841              dn->d_u.d_child.prev, dn->d_u.d_child.next);
 842         spin_unlock(&dn->d_lock);
 843         spin_unlock(&dcache_lock);
 844 }
 845
 846 /*
 847  * Incorporate results into the local cache.  This is either just
 848  * one inode, or a directory, dentry, and possibly linked-to inode (e.g.,
 849  * after a lookup).
 850  *
 851  * A reply may contain
 852  *         a directory inode along with a dentry.
 853  *  and/or a target inode
 854  *
 855  * Called with snap_rwsem (read).
 856  */
 857 int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req,
 858                     struct ceph_mds_session *session)
 859 {
 860         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
 861         struct inode *in = NULL;
 862         struct ceph_mds_reply_inode *ininfo;
 863         struct ceph_vino vino;
 864         int i = 0;
 865         int err = 0;
 866
 867         dout("fill_trace %p is_dentry %d is_target %d\n", req,
 868              rinfo->head->is_dentry, rinfo->head->is_target);
 869
 870 #if 0
 871         /*
 872          * Debugging hook:
 873          *
 874          * If we resend completed ops to a recovering mds, we get no
 875          * trace.  Since that is very rare, pretend this is the case
 876          * to ensure the 'no trace' handlers in the callers behave.
 877          *
 878          * Fill in inodes unconditionally to avoid breaking cap
 879          * invariants.
 880          */
 881         if (rinfo->head->op & CEPH_MDS_OP_WRITE) {
 882                 pr_info("fill_trace faking empty trace on %lld %s\n",
 883                         req->r_tid, ceph_mds_op_name(rinfo->head->op));
 884                 if (rinfo->head->is_dentry) {
 885                         rinfo->head->is_dentry = 0;
 886                         err = fill_inode(req->r_locked_dir,
 887                                          &rinfo->diri, rinfo->dirfrag,
 888                                          session, req->r_request_started, -1);
 889                 }
 890                 if (rinfo->head->is_target) {
 891                         rinfo->head->is_target = 0;
 892                         ininfo = rinfo->targeti.in;
 893                         vino.ino = le64_to_cpu(ininfo->ino);
 894                         vino.snap = le64_to_cpu(ininfo->snapid);
 895                         in = ceph_get_inode(sb, vino);
 896                         err = fill_inode(in, &rinfo->targeti, NULL,
 897                                          session, req->r_request_started,
 898                                          req->r_fmode);
 899                         iput(in);
 900                 }
 901         }
 902 #endif
 903
 904         if (!rinfo->head->is_target && !rinfo->head->is_dentry) {
 905                 dout("fill_trace reply is empty!\n");
 906                 if (rinfo->head->result == 0 && req->r_locked_dir) {
 907                         struct ceph_inode_info *ci =
 908                                 ceph_inode(req->r_locked_dir);
 909                         dout(" clearing %p complete (empty trace)\n",
 910                              req->r_locked_dir);
 911                         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
 912                         ci->i_release_count++;
 913                 }
 914                 return 0;
 915         }
 916
 917         if (rinfo->head->is_dentry) {
 918                 struct inode *dir = req->r_locked_dir;
 919
 920                 err = fill_inode(dir, &rinfo->diri, rinfo->dirfrag,
 921                                  session, req->r_request_started, -1,
 922                                  &req->r_caps_reservation);
 923                 if (err < 0)
 924                         return err;
 925         }
 926
 927         if (rinfo->head->is_dentry && !req->r_aborted) {
 928                 /*
 929                  * lookup link rename   : null -> possibly existing inode
 930                  * mknod symlink mkdir  : null -> new inode
 931                  * unlink               : linked -> null
 932                  */
 933                 struct inode *dir = req->r_locked_dir;
 934                 struct dentry *dn = req->r_dentry;
 935                 bool have_dir_cap, have_lease;
 936
 937                 BUG_ON(!dn);
 938                 BUG_ON(!dir);
 939                 BUG_ON(dn->d_parent->d_inode != dir);
 940                 BUG_ON(ceph_ino(dir) !=
 941                        le64_to_cpu(rinfo->diri.in->ino));
 942                 BUG_ON(ceph_snap(dir) !=
 943                        le64_to_cpu(rinfo->diri.in->snapid));
 944
 945                 /* do we have a lease on the whole dir? */
 946                 have_dir_cap =
 947                         (le32_to_cpu(rinfo->diri.in->cap.caps) &
 948                          CEPH_CAP_FILE_SHARED);
 949
 950                 /* do we have a dn lease? */
 951                 have_lease = have_dir_cap ||
 952                         (le16_to_cpu(rinfo->dlease->mask) &
 953                          CEPH_LOCK_DN);
 954
 955                 if (!have_lease)
 956                         dout("fill_trace  no dentry lease or dir cap\n");
 957
 958                 /* rename? */
 959                 if (req->r_old_dentry && req->r_op == CEPH_MDS_OP_RENAME) {
 960                         dout(" src %p '%.*s' dst %p '%.*s'\n",
 961                              req->r_old_dentry,
 962                              req->r_old_dentry->d_name.len,
 963                              req->r_old_dentry->d_name.name,
 964                              dn, dn->d_name.len, dn->d_name.name);
 965                         dout("fill_trace doing d_move %p -> %p\n",
 966                              req->r_old_dentry, dn);
 967                         d_move(req->r_old_dentry, dn);
 968                         dout(" src %p '%.*s' dst %p '%.*s'\n",
 969                              req->r_old_dentry,
 970                              req->r_old_dentry->d_name.len,
 971                              req->r_old_dentry->d_name.name,
 972                              dn, dn->d_name.len, dn->d_name.name);
 973                         /* ensure target dentry is invalidated, despite
 974                            rehashing bug in vfs_rename_dir */
 975                         dn->d_time = jiffies;
 976                         ceph_dentry(dn)->lease_shared_gen = 0;
 977                         /* take overwritten dentry's readdir offset */
 978                         ceph_dentry(req->r_old_dentry)->offset =
 979                                 ceph_dentry(dn)->offset;
 980                         dn = req->r_old_dentry;  /* use old_dentry */
 981                         in = dn->d_inode;
 982                 }
 983
 984                 /* null dentry? */
 985                 if (!rinfo->head->is_target) {
 986                         dout("fill_trace null dentry\n");
 987                         if (dn->d_inode) {
 988                                 dout("d_delete %p\n", dn);
 989                                 d_delete(dn);
 990                         } else {
 991                                 dout("d_instantiate %p NULL\n", dn);
 992                                 d_instantiate(dn, NULL);
 993                                 if (have_lease && d_unhashed(dn))
 994                                         d_rehash(dn);
 995                                 update_dentry_lease(dn, rinfo->dlease,
 996                                                     session,
 997                                                     req->r_request_started);
 998                         }
 999                         goto done;
1000                 }
1001
1002                 /* attach proper inode */
1003                 ininfo = rinfo->targeti.in;
1004                 vino.ino = le64_to_cpu(ininfo->ino);
1005                 vino.snap = le64_to_cpu(ininfo->snapid);
1006                 if (!dn->d_inode) {
1007                         in = ceph_get_inode(sb, vino);
1008                         if (IS_ERR(in)) {
1009                                 pr_err("fill_trace bad get_inode "
1010                                        "%llx.%llx\n", vino.ino, vino.snap);
1011                                 err = PTR_ERR(in);
1012                                 d_delete(dn);
1013                                 goto done;
1014                         }
1015                         dn = splice_dentry(dn, in, &have_lease);
1016                         if (IS_ERR(dn)) {
1017                                 err = PTR_ERR(dn);
1018                                 goto done;
1019                         }
1020                         req->r_dentry = dn;  /* may have spliced */
1021                         ceph_set_dentry_offset(dn);
1022                         igrab(in);
1023                 } else if (ceph_ino(in) == vino.ino &&
1024                            ceph_snap(in) == vino.snap) {
1025                         igrab(in);
1026                 } else {
1027                         dout(" %p links to %p %llx.%llx, not %llx.%llx\n",
1028                              dn, in, ceph_ino(in), ceph_snap(in),
1029                              vino.ino, vino.snap);
1030                         have_lease = false;
1031                         in = NULL;
1032                 }
1033
1034                 if (have_lease)
1035                         update_dentry_lease(dn, rinfo->dlease, session,
1036                                             req->r_request_started);
1037                 dout(" final dn %p\n", dn);
1038                 i++;
1039         } else if (req->r_op == CEPH_MDS_OP_LOOKUPSNAP ||
1040                    req->r_op == CEPH_MDS_OP_MKSNAP) {
1041                 struct dentry *dn = req->r_dentry;
1042
1043                 /* fill out a snapdir LOOKUPSNAP dentry */
1044                 BUG_ON(!dn);
1045                 BUG_ON(!req->r_locked_dir);
1046                 BUG_ON(ceph_snap(req->r_locked_dir) != CEPH_SNAPDIR);
1047                 ininfo = rinfo->targeti.in;
1048                 vino.ino = le64_to_cpu(ininfo->ino);
1049                 vino.snap = le64_to_cpu(ininfo->snapid);
1050                 in = ceph_get_inode(sb, vino);
1051                 if (IS_ERR(in)) {
1052                         pr_err("fill_inode get_inode badness %llx.%llx\n",
1053                                vino.ino, vino.snap);
1054                         err = PTR_ERR(in);
1055                         d_delete(dn);
1056                         goto done;
1057                 }
1058                 dout(" linking snapped dir %p to dn %p\n", in, dn);
1059                 dn = splice_dentry(dn, in, NULL);
1060                 if (IS_ERR(dn)) {
1061                         err = PTR_ERR(dn);
1062                         goto done;
1063                 }
1064                 ceph_set_dentry_offset(dn);
1065                 req->r_dentry = dn;  /* may have spliced */
1066                 igrab(in);
1067                 rinfo->head->is_dentry = 1;  /* fool notrace handlers */
1068         }
1069
1070         if (rinfo->head->is_target) {
1071                 vino.ino = le64_to_cpu(rinfo->targeti.in->ino);
1072                 vino.snap = le64_to_cpu(rinfo->targeti.in->snapid);
1073
1074                 if (in == NULL || ceph_ino(in) != vino.ino ||
1075                     ceph_snap(in) != vino.snap) {
1076                         in = ceph_get_inode(sb, vino);
1077                         if (IS_ERR(in)) {
1078                                 err = PTR_ERR(in);
1079                                 goto done;
1080                         }
1081                 }
1082                 req->r_target_inode = in;
1083
1084                 err = fill_inode(in,
1085                                  &rinfo->targeti, NULL,
1086                                  session, req->r_request_started,
1087                                  (le32_to_cpu(rinfo->head->result) == 0) ?
1088                                  req->r_fmode : -1,
1089                                  &req->r_caps_reservation);
1090                 if (err < 0) {
1091                         pr_err("fill_inode badness %p %llx.%llx\n",
1092                                in, ceph_vinop(in));
1093                         goto done;
1094                 }
1095         }
1096
1097 done:
1098         dout("fill_trace done err=%d\n", err);
1099         return err;
1100 }
1101
1102 /*
1103  * Prepopulate our cache with readdir results, leases, etc.
1104  */
1105 int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1106                              struct ceph_mds_session *session)
1107 {
1108         struct dentry *parent = req->r_dentry;
1109         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1110         struct qstr dname;
1111         struct dentry *dn;
1112         struct inode *in;
1113         int err = 0, i;
1114         struct inode *snapdir = NULL;
1115         struct ceph_mds_request_head *rhead = req->r_request->front.iov_base;
1116         u64 frag = le32_to_cpu(rhead->args.readdir.frag);
1117         struct ceph_dentry_info *di;
1118
1119         if (le32_to_cpu(rinfo->head->op) == CEPH_MDS_OP_LSSNAP) {
1120                 snapdir = ceph_get_snapdir(parent->d_inode);
1121                 parent = d_find_alias(snapdir);
1122                 dout("readdir_prepopulate %d items under SNAPDIR dn %p\n",
1123                      rinfo->dir_nr, parent);
1124         } else {
1125                 dout("readdir_prepopulate %d items under dn %p\n",
1126                      rinfo->dir_nr, parent);
1127                 if (rinfo->dir_dir)
1128                         ceph_fill_dirfrag(parent->d_inode, rinfo->dir_dir);
1129         }
1130
1131         for (i = 0; i < rinfo->dir_nr; i++) {
1132                 struct ceph_vino vino;
1133
1134                 dname.name = rinfo->dir_dname[i];
1135                 dname.len = rinfo->dir_dname_len[i];
1136                 dname.hash = full_name_hash(dname.name, dname.len);
1137
1138                 vino.ino = le64_to_cpu(rinfo->dir_in[i].in->ino);
1139                 vino.snap = le64_to_cpu(rinfo->dir_in[i].in->snapid);
1140
1141 retry_lookup:
1142                 dn = d_lookup(parent, &dname);
1143                 dout("d_lookup on parent=%p name=%.*s got %p\n",
1144                      parent, dname.len, dname.name, dn);
1145
1146                 if (!dn) {
1147                         dn = d_alloc(parent, &dname);
1148                         dout("d_alloc %p '%.*s' = %p\n", parent,
1149                              dname.len, dname.name, dn);
1150                         if (dn == NULL) {
1151                                 dout("d_alloc badness\n");
1152                                 err = -ENOMEM;
1153                                 goto out;
1154                         }
1155                         err = ceph_init_dentry(dn);
1156                         if (err < 0)
1157                                 goto out;
1158                 } else if (dn->d_inode &&
1159                            (ceph_ino(dn->d_inode) != vino.ino ||
1160                             ceph_snap(dn->d_inode) != vino.snap)) {
1161                         dout(" dn %p points to wrong inode %p\n",
1162                              dn, dn->d_inode);
1163                         d_delete(dn);
1164                         dput(dn);
1165                         goto retry_lookup;
1166                 } else {
1167                         /* reorder parent's d_subdirs */
1168                         spin_lock(&dcache_lock);
1169                         spin_lock(&dn->d_lock);
1170                         list_move(&dn->d_u.d_child, &parent->d_subdirs);
1171                         spin_unlock(&dn->d_lock);
1172                         spin_unlock(&dcache_lock);
1173                 }
1174
1175                 di = dn->d_fsdata;
1176                 di->offset = ceph_make_fpos(frag, i + req->r_readdir_offset);
1177
1178                 /* inode */
1179                 if (dn->d_inode) {
1180                         in = dn->d_inode;
1181                 } else {
1182                         in = ceph_get_inode(parent->d_sb, vino);
1183                         if (in == NULL) {
1184                                 dout("new_inode badness\n");
1185                                 d_delete(dn);
1186                                 dput(dn);
1187                                 err = -ENOMEM;
1188                                 goto out;
1189                         }
1190                         dn = splice_dentry(dn, in, NULL);
1191                 }
1192
1193                 if (fill_inode(in, &rinfo->dir_in[i], NULL, session,
1194                                req->r_request_started, -1,
1195                                &req->r_caps_reservation) < 0) {
1196                         pr_err("fill_inode badness on %p\n", in);
1197                         dput(dn);
1198                         continue;
1199                 }
1200                 update_dentry_lease(dn, rinfo->dir_dlease[i],
1201                                     req->r_session, req->r_request_started);
1202                 dput(dn);
1203         }
1204         req->r_did_prepopulate = true;
1205
1206 out:
1207         if (snapdir) {
1208                 iput(snapdir);
1209                 dput(parent);
1210         }
1211         dout("readdir_prepopulate done\n");
1212         return err;
1213 }
1214
1215 int ceph_inode_set_size(struct inode *inode, loff_t size)
1216 {
1217         struct ceph_inode_info *ci = ceph_inode(inode);
1218         int ret = 0;
1219
1220         spin_lock(&inode->i_lock);
1221         dout("set_size %p %llu -> %llu\n", inode, inode->i_size, size);
1222         inode->i_size = size;
1223         inode->i_blocks = (size + (1 << 9) - 1) >> 9;
1224
1225         /* tell the MDS if we are approaching max_size */
1226         if ((size << 1) >= ci->i_max_size &&
1227             (ci->i_reported_size << 1) < ci->i_max_size)
1228                 ret = 1;
1229
1230         spin_unlock(&inode->i_lock);
1231         return ret;
1232 }
1233
1234 /*
1235  * Write back inode data in a worker thread.  (This can't be done
1236  * in the message handler context.)
1237  */
1238 void ceph_inode_writeback(struct work_struct *work)
1239 {
1240         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1241                                                   i_wb_work);
1242         struct inode *inode = &ci->vfs_inode;
1243
1244         dout("writeback %p\n", inode);
1245         filemap_fdatawrite(&inode->i_data);
1246         iput(inode);
1247 }
1248
1249 /*
1250  * Invalidate inode pages in a worker thread.  (This can't be done
1251  * in the message handler context.)
1252  */
1253 static void ceph_inode_invalidate_pages(struct work_struct *work)
1254 {
1255         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1256                                                   i_pg_inv_work);
1257         struct inode *inode = &ci->vfs_inode;
1258         u32 orig_gen;
1259         int check = 0;
1260
1261         spin_lock(&inode->i_lock);
1262         dout("invalidate_pages %p gen %d revoking %d\n", inode,
1263              ci->i_rdcache_gen, ci->i_rdcache_revoking);
1264         if (ci->i_rdcache_gen == 0 ||
1265             ci->i_rdcache_revoking != ci->i_rdcache_gen) {
1266                 BUG_ON(ci->i_rdcache_revoking > ci->i_rdcache_gen);
1267                 /* nevermind! */
1268                 ci->i_rdcache_revoking = 0;
1269                 spin_unlock(&inode->i_lock);
1270                 goto out;
1271         }
1272         orig_gen = ci->i_rdcache_gen;
1273         spin_unlock(&inode->i_lock);
1274
1275         truncate_inode_pages(&inode->i_data, 0);
1276
1277         spin_lock(&inode->i_lock);
1278         if (orig_gen == ci->i_rdcache_gen) {
1279                 dout("invalidate_pages %p gen %d successful\n", inode,
1280                      ci->i_rdcache_gen);
1281                 ci->i_rdcache_gen = 0;
1282                 ci->i_rdcache_revoking = 0;
1283                 check = 1;
1284         } else {
1285                 dout("invalidate_pages %p gen %d raced, gen now %d\n",
1286                      inode, orig_gen, ci->i_rdcache_gen);
1287         }
1288         spin_unlock(&inode->i_lock);
1289
1290         if (check)
1291                 ceph_check_caps(ci, 0, NULL);
1292 out:
1293         iput(inode);
1294 }
1295
1296
1297 /*
1298  * called by trunc_wq; take i_mutex ourselves
1299  *
1300  * We also truncate in a separate thread as well.
1301  */
1302 void ceph_vmtruncate_work(struct work_struct *work)
1303 {
1304         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1305                                                   i_vmtruncate_work);
1306         struct inode *inode = &ci->vfs_inode;
1307
1308         dout("vmtruncate_work %p\n", inode);
1309         mutex_lock(&inode->i_mutex);
1310         __ceph_do_pending_vmtruncate(inode);
1311         mutex_unlock(&inode->i_mutex);
1312         iput(inode);
1313 }
1314
1315 /*
1316  * called with i_mutex held.
1317  *
1318  * Make sure any pending truncation is applied before doing anything
1319  * that may depend on it.
1320  */
1321 void __ceph_do_pending_vmtruncate(struct inode *inode)
1322 {
1323         struct ceph_inode_info *ci = ceph_inode(inode);
1324         u64 to;
1325         int wrbuffer_refs, wake = 0;
1326
1327 retry:
1328         spin_lock(&inode->i_lock);
1329         if (ci->i_truncate_pending == 0) {
1330                 dout("__do_pending_vmtruncate %p none pending\n", inode);
1331                 spin_unlock(&inode->i_lock);
1332                 return;
1333         }
1334
1335         /*
1336          * make sure any dirty snapped pages are flushed before we
1337          * possibly truncate them.. so write AND block!
1338          */
1339         if (ci->i_wrbuffer_ref_head < ci->i_wrbuffer_ref) {
1340                 dout("__do_pending_vmtruncate %p flushing snaps first\n",
1341                      inode);
1342                 spin_unlock(&inode->i_lock);
1343                 filemap_write_and_wait_range(&inode->i_data, 0,
1344                                              inode->i_sb->s_maxbytes);
1345                 goto retry;
1346         }
1347
1348         to = ci->i_truncate_size;
1349         wrbuffer_refs = ci->i_wrbuffer_ref;
1350         dout("__do_pending_vmtruncate %p (%d) to %lld\n", inode,
1351              ci->i_truncate_pending, to);
1352         spin_unlock(&inode->i_lock);
1353
1354         truncate_inode_pages(inode->i_mapping, to);
1355
1356         spin_lock(&inode->i_lock);
1357         ci->i_truncate_pending--;
1358         if (ci->i_truncate_pending == 0)
1359                 wake = 1;
1360         spin_unlock(&inode->i_lock);
1361
1362         if (wrbuffer_refs == 0)
1363                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1364         if (wake)
1365                 wake_up(&ci->i_cap_wq);
1366 }
1367
1368
1369 /*
1370  * symlinks
1371  */
1372 static void *ceph_sym_follow_link(struct dentry *dentry, struct nameidata *nd)
1373 {
1374         struct ceph_inode_info *ci = ceph_inode(dentry->d_inode);
1375         nd_set_link(nd, ci->i_symlink);
1376         return NULL;
1377 }
1378
1379 static const struct inode_operations ceph_symlink_iops = {
1380         .readlink = generic_readlink,
1381         .follow_link = ceph_sym_follow_link,
1382 };
1383
1384 /*
1385  * setattr
1386  */
1387 int ceph_setattr(struct dentry *dentry, struct iattr *attr)
1388 {
1389         struct inode *inode = dentry->d_inode;
1390         struct ceph_inode_info *ci = ceph_inode(inode);
1391         struct inode *parent_inode = dentry->d_parent->d_inode;
1392         const unsigned int ia_valid = attr->ia_valid;
1393         struct ceph_mds_request *req;
1394         struct ceph_mds_client *mdsc = &ceph_client(dentry->d_sb)->mdsc;
1395         int issued;
1396         int release = 0, dirtied = 0;
1397         int mask = 0;
1398         int err = 0;
1399         int queue_trunc = 0;
1400
1401         if (ceph_snap(inode) != CEPH_NOSNAP)
1402                 return -EROFS;
1403
1404         __ceph_do_pending_vmtruncate(inode);
1405
1406         err = inode_change_ok(inode, attr);
1407         if (err != 0)
1408                 return err;
1409
1410         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETATTR,
1411                                        USE_AUTH_MDS);
1412         if (IS_ERR(req))
1413                 return PTR_ERR(req);
1414
1415         spin_lock(&inode->i_lock);
1416         issued = __ceph_caps_issued(ci, NULL);
1417         dout("setattr %p issued %s\n", inode, ceph_cap_string(issued));
1418
1419         if (ia_valid & ATTR_UID) {
1420                 dout("setattr %p uid %d -> %d\n", inode,
1421                      inode->i_uid, attr->ia_uid);
1422                 if (issued & CEPH_CAP_AUTH_EXCL) {
1423                         inode->i_uid = attr->ia_uid;
1424                         dirtied |= CEPH_CAP_AUTH_EXCL;
1425                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1426                            attr->ia_uid != inode->i_uid) {
1427                         req->r_args.setattr.uid = cpu_to_le32(attr->ia_uid);
1428                         mask |= CEPH_SETATTR_UID;
1429                         release |= CEPH_CAP_AUTH_SHARED;
1430                 }
1431         }
1432         if (ia_valid & ATTR_GID) {
1433                 dout("setattr %p gid %d -> %d\n", inode,
1434                      inode->i_gid, attr->ia_gid);
1435                 if (issued & CEPH_CAP_AUTH_EXCL) {
1436                         inode->i_gid = attr->ia_gid;
1437                         dirtied |= CEPH_CAP_AUTH_EXCL;
1438                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1439                            attr->ia_gid != inode->i_gid) {
1440                         req->r_args.setattr.gid = cpu_to_le32(attr->ia_gid);
1441                         mask |= CEPH_SETATTR_GID;
1442                         release |= CEPH_CAP_AUTH_SHARED;
1443                 }
1444         }
1445         if (ia_valid & ATTR_MODE) {
1446                 dout("setattr %p mode 0%o -> 0%o\n", inode, inode->i_mode,
1447                      attr->ia_mode);
1448                 if (issued & CEPH_CAP_AUTH_EXCL) {
1449                         inode->i_mode = attr->ia_mode;
1450                         dirtied |= CEPH_CAP_AUTH_EXCL;
1451                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1452                            attr->ia_mode != inode->i_mode) {
1453                         req->r_args.setattr.mode = cpu_to_le32(attr->ia_mode);
1454                         mask |= CEPH_SETATTR_MODE;
1455                         release |= CEPH_CAP_AUTH_SHARED;
1456                 }
1457         }
1458
1459         if (ia_valid & ATTR_ATIME) {
1460                 dout("setattr %p atime %ld.%ld -> %ld.%ld\n", inode,
1461                      inode->i_atime.tv_sec, inode->i_atime.tv_nsec,
1462                      attr->ia_atime.tv_sec, attr->ia_atime.tv_nsec);
1463                 if (issued & CEPH_CAP_FILE_EXCL) {
1464                         ci->i_time_warp_seq++;
1465                         inode->i_atime = attr->ia_atime;
1466                         dirtied |= CEPH_CAP_FILE_EXCL;
1467                 } else if ((issued & CEPH_CAP_FILE_WR) &&
1468                            timespec_compare(&inode->i_atime,
1469                                             &attr->ia_atime) < 0) {
1470                         inode->i_atime = attr->ia_atime;
1471                         dirtied |= CEPH_CAP_FILE_WR;
1472                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1473                            !timespec_equal(&inode->i_atime, &attr->ia_atime)) {
1474                         ceph_encode_timespec(&req->r_args.setattr.atime,
1475                                              &attr->ia_atime);
1476                         mask |= CEPH_SETATTR_ATIME;
1477                         release |= CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_RD |
1478                                 CEPH_CAP_FILE_WR;
1479                 }
1480         }
1481         if (ia_valid & ATTR_MTIME) {
1482                 dout("setattr %p mtime %ld.%ld -> %ld.%ld\n", inode,
1483                      inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
1484                      attr->ia_mtime.tv_sec, attr->ia_mtime.tv_nsec);
1485                 if (issued & CEPH_CAP_FILE_EXCL) {
1486                         ci->i_time_warp_seq++;
1487                         inode->i_mtime = attr->ia_mtime;
1488                         dirtied |= CEPH_CAP_FILE_EXCL;
1489                 } else if ((issued & CEPH_CAP_FILE_WR) &&
1490                            timespec_compare(&inode->i_mtime,
1491                                             &attr->ia_mtime) < 0) {
1492                         inode->i_mtime = attr->ia_mtime;
1493                         dirtied |= CEPH_CAP_FILE_WR;
1494                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1495                            !timespec_equal(&inode->i_mtime, &attr->ia_mtime)) {
1496                         ceph_encode_timespec(&req->r_args.setattr.mtime,
1497                                              &attr->ia_mtime);
1498                         mask |= CEPH_SETATTR_MTIME;
1499                         release |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
1500                                 CEPH_CAP_FILE_WR;
1501                 }
1502         }
1503         if (ia_valid & ATTR_SIZE) {
1504                 dout("setattr %p size %lld -> %lld\n", inode,
1505                      inode->i_size, attr->ia_size);
1506                 if (attr->ia_size > inode->i_sb->s_maxbytes) {
1507                         err = -EINVAL;
1508                         goto out;
1509                 }
1510                 if ((issued & CEPH_CAP_FILE_EXCL) &&
1511                     attr->ia_size > inode->i_size) {
1512                         inode->i_size = attr->ia_size;
1513                         if (attr->ia_size < inode->i_size) {
1514                                 ci->i_truncate_size = attr->ia_size;
1515                                 ci->i_truncate_pending++;
1516                                 queue_trunc = 1;
1517                         }
1518                         inode->i_blocks =
1519                                 (attr->ia_size + (1 << 9) - 1) >> 9;
1520                         inode->i_ctime = attr->ia_ctime;
1521                         ci->i_reported_size = attr->ia_size;
1522                         dirtied |= CEPH_CAP_FILE_EXCL;
1523                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1524                            attr->ia_size != inode->i_size) {
1525                         req->r_args.setattr.size = cpu_to_le64(attr->ia_size);
1526                         req->r_args.setattr.old_size =
1527                                 cpu_to_le64(inode->i_size);
1528                         mask |= CEPH_SETATTR_SIZE;
1529                         release |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
1530                                 CEPH_CAP_FILE_WR;
1531                 }
1532         }
1533
1534         /* these do nothing */
1535         if (ia_valid & ATTR_CTIME) {
1536                 bool only = (ia_valid & (ATTR_SIZE|ATTR_MTIME|ATTR_ATIME|
1537                                          ATTR_MODE|ATTR_UID|ATTR_GID)) == 0;
1538                 dout("setattr %p ctime %ld.%ld -> %ld.%ld (%s)\n", inode,
1539                      inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
1540                      attr->ia_ctime.tv_sec, attr->ia_ctime.tv_nsec,
1541                      only ? "ctime only" : "ignored");
1542                 inode->i_ctime = attr->ia_ctime;
1543                 if (only) {
1544                         /*
1545                          * if kernel wants to dirty ctime but nothing else,
1546                          * we need to choose a cap to dirty under, or do
1547                          * a almost-no-op setattr
1548                          */
1549                         if (issued & CEPH_CAP_AUTH_EXCL)
1550                                 dirtied |= CEPH_CAP_AUTH_EXCL;
1551                         else if (issued & CEPH_CAP_FILE_EXCL)
1552                                 dirtied |= CEPH_CAP_FILE_EXCL;
1553                         else if (issued & CEPH_CAP_XATTR_EXCL)
1554                                 dirtied |= CEPH_CAP_XATTR_EXCL;
1555                         else
1556                                 mask |= CEPH_SETATTR_CTIME;
1557                 }
1558         }
1559         if (ia_valid & ATTR_FILE)
1560                 dout("setattr %p ATTR_FILE ... hrm!\n", inode);
1561
1562         if (dirtied) {
1563                 __ceph_mark_dirty_caps(ci, dirtied);
1564                 inode->i_ctime = CURRENT_TIME;
1565         }
1566
1567         release &= issued;
1568         spin_unlock(&inode->i_lock);
1569
1570         if (queue_trunc)
1571                 __ceph_do_pending_vmtruncate(inode);
1572
1573         if (mask) {
1574                 req->r_inode = igrab(inode);
1575                 req->r_inode_drop = release;
1576                 req->r_args.setattr.mask = cpu_to_le32(mask);
1577                 req->r_num_caps = 1;
1578                 err = ceph_mdsc_do_request(mdsc, parent_inode, req);
1579         }
1580         dout("setattr %p result=%d (%s locally, %d remote)\n", inode, err,
1581              ceph_cap_string(dirtied), mask);
1582
1583         ceph_mdsc_put_request(req);
1584         __ceph_do_pending_vmtruncate(inode);
1585         return err;
1586 out:
1587         spin_unlock(&inode->i_lock);
1588         ceph_mdsc_put_request(req);
1589         return err;
1590 }
1591
1592 /*
1593  * Verify that we have a lease on the given mask.  If not,
1594  * do a getattr against an mds.
1595  */
1596 int ceph_do_getattr(struct inode *inode, int mask)
1597 {
1598         struct ceph_client *client = ceph_sb_to_client(inode->i_sb);
1599         struct ceph_mds_client *mdsc = &client->mdsc;
1600         struct ceph_mds_request *req;
1601         int err;
1602
1603         if (ceph_snap(inode) == CEPH_SNAPDIR) {
1604                 dout("do_getattr inode %p SNAPDIR\n", inode);
1605                 return 0;
1606         }
1607
1608         dout("do_getattr inode %p mask %s\n", inode, ceph_cap_string(mask));
1609         if (ceph_caps_issued_mask(ceph_inode(inode), mask, 1))
1610                 return 0;
1611
1612         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, USE_ANY_MDS);
1613         if (IS_ERR(req))
1614                 return PTR_ERR(req);
1615         req->r_inode = igrab(inode);
1616         req->r_num_caps = 1;
1617         req->r_args.getattr.mask = cpu_to_le32(mask);
1618         err = ceph_mdsc_do_request(mdsc, NULL, req);
1619         ceph_mdsc_put_request(req);
1620         dout("do_getattr result=%d\n", err);
1621         return err;
1622 }
1623
1624
1625 /*
1626  * Check inode permissions.  We verify we have a valid value for
1627  * the AUTH cap, then call the generic handler.
1628  */
1629 int ceph_permission(struct inode *inode, int mask)
1630 {
1631         int err = ceph_do_getattr(inode, CEPH_CAP_AUTH_SHARED);
1632
1633         if (!err)
1634                 err = generic_permission(inode, mask, NULL);
1635         return err;
1636 }
1637
1638 /*
1639  * Get all attributes.  Hopefully somedata we'll have a statlite()
1640  * and can limit the fields we require to be accurate.
1641  */
1642 int ceph_getattr(struct vfsmount *mnt, struct dentry *dentry,
1643                  struct kstat *stat)
1644 {
1645         struct inode *inode = dentry->d_inode;
1646         struct ceph_inode_info *ci = ceph_inode(inode);
1647         int err;
1648
1649         err = ceph_do_getattr(inode, CEPH_STAT_CAP_INODE_ALL);
1650         if (!err) {
1651                 generic_fillattr(inode, stat);
1652                 stat->ino = inode->i_ino;
1653                 if (ceph_snap(inode) != CEPH_NOSNAP)
1654                         stat->dev = ceph_snap(inode);
1655                 else
1656                         stat->dev = 0;
1657                 if (S_ISDIR(inode->i_mode)) {
1658                         stat->size = ci->i_rbytes;
1659                         stat->blocks = 0;
1660                         stat->blksize = 65536;
1661                 }
1662         }
1663         return err;
1664 }