fs/ceph/snap.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/sort.h>
   4 #include <linux/slab.h>
   5
   6 #include "super.h"
   7 #include "mds_client.h"
   8
   9 #include <linux/ceph/decode.h>
  10
  11 /*
  12  * Snapshots in ceph are driven in large part by cooperation from the
  13  * client.  In contrast to local file systems or file servers that
  14  * implement snapshots at a single point in the system, ceph's
  15  * distributed access to storage requires clients to help decide
  16  * whether a write logically occurs before or after a recently created
  17  * snapshot.
  18  *
  19  * This provides a perfect instantanous client-wide snapshot.  Between
  20  * clients, however, snapshots may appear to be applied at slightly
  21  * different points in time, depending on delays in delivering the
  22  * snapshot notification.
  23  *
  24  * Snapshots are _not_ file system-wide.  Instead, each snapshot
  25  * applies to the subdirectory nested beneath some directory.  This
  26  * effectively divides the hierarchy into multiple "realms," where all
  27  * of the files contained by each realm share the same set of
  28  * snapshots.  An individual realm's snap set contains snapshots
  29  * explicitly created on that realm, as well as any snaps in its
  30  * parent's snap set _after_ the point at which the parent became it's
  31  * parent (due to, say, a rename).  Similarly, snaps from prior parents
  32  * during the time intervals during which they were the parent are included.
  33  *
  34  * The client is spared most of this detail, fortunately... it must only
  35  * maintains a hierarchy of realms reflecting the current parent/child
  36  * realm relationship, and for each realm has an explicit list of snaps
  37  * inherited from prior parents.
  38  *
  39  * A snap_realm struct is maintained for realms containing every inode
  40  * with an open cap in the system.  (The needed snap realm information is
  41  * provided by the MDS whenever a cap is issued, i.e., on open.)  A 'seq'
  42  * version number is used to ensure that as realm parameters change (new
  43  * snapshot, new parent, etc.) the client's realm hierarchy is updated.
  44  *
  45  * The realm hierarchy drives the generation of a 'snap context' for each
  46  * realm, which simply lists the resulting set of snaps for the realm.  This
  47  * is attached to any writes sent to OSDs.
  48  */
  49 /*
  50  * Unfortunately error handling is a bit mixed here.  If we get a snap
  51  * update, but don't have enough memory to update our realm hierarchy,
  52  * it's not clear what we can do about it (besides complaining to the
  53  * console).
  54  */
  55
  56
  57 /*
  58  * increase ref count for the realm
  59  *
  60  * caller must hold snap_rwsem for write.
  61  */
  62 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
  63                          struct ceph_snap_realm *realm)
  64 {
  65         dout("get_realm %p %d -> %d\n", realm,
  66              atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
  67         /*
  68          * since we _only_ increment realm refs or empty the empty
  69          * list with snap_rwsem held, adjusting the empty list here is
  70          * safe.  we do need to protect against concurrent empty list
  71          * additions, however.
  72          */
  73         if (atomic_read(&realm->nref) == 0) {
  74                 spin_lock(&mdsc->snap_empty_lock);
  75                 list_del_init(&realm->empty_item);
  76                 spin_unlock(&mdsc->snap_empty_lock);
  77         }
  78
  79         atomic_inc(&realm->nref);
  80 }
  81
  82 static void __insert_snap_realm(struct rb_root *root,
  83                                 struct ceph_snap_realm *new)
  84 {
  85         struct rb_node **p = &root->rb_node;
  86         struct rb_node *parent = NULL;
  87         struct ceph_snap_realm *r = NULL;
  88
  89         while (*p) {
  90                 parent = *p;
  91                 r = rb_entry(parent, struct ceph_snap_realm, node);
  92                 if (new->ino < r->ino)
  93                         p = &(*p)->rb_left;
  94                 else if (new->ino > r->ino)
  95                         p = &(*p)->rb_right;
  96                 else
  97                         BUG();
  98         }
  99
 100         rb_link_node(&new->node, parent, p);
 101         rb_insert_color(&new->node, root);
 102 }
 103
 104 /*
 105  * create and get the realm rooted at @ino and bump its ref count.
 106  *
 107  * caller must hold snap_rwsem for write.
 108  */
 109 static struct ceph_snap_realm *ceph_create_snap_realm(
 110         struct ceph_mds_client *mdsc,
 111         u64 ino)
 112 {
 113         struct ceph_snap_realm *realm;
 114
 115         realm = kzalloc(sizeof(*realm), GFP_NOFS);
 116         if (!realm)
 117                 return ERR_PTR(-ENOMEM);
 118
 119         atomic_set(&realm->nref, 0);    /* tree does not take a ref */
 120         realm->ino = ino;
 121         INIT_LIST_HEAD(&realm->children);
 122         INIT_LIST_HEAD(&realm->child_item);
 123         INIT_LIST_HEAD(&realm->empty_item);
 124         INIT_LIST_HEAD(&realm->dirty_item);
 125         INIT_LIST_HEAD(&realm->inodes_with_caps);
 126         spin_lock_init(&realm->inodes_with_caps_lock);
 127         __insert_snap_realm(&mdsc->snap_realms, realm);
 128         dout("create_snap_realm %llx %p\n", realm->ino, realm);
 129         return realm;
 130 }
 131
 132 /*
 133  * lookup the realm rooted at @ino.
 134  *
 135  * caller must hold snap_rwsem for write.
 136  */
 137 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
 138                                                u64 ino)
 139 {
 140         struct rb_node *n = mdsc->snap_realms.rb_node;
 141         struct ceph_snap_realm *r;
 142
 143         while (n) {
 144                 r = rb_entry(n, struct ceph_snap_realm, node);
 145                 if (ino < r->ino)
 146                         n = n->rb_left;
 147                 else if (ino > r->ino)
 148                         n = n->rb_right;
 149                 else {
 150                         dout("lookup_snap_realm %llx %p\n", r->ino, r);
 151                         return r;
 152                 }
 153         }
 154         return NULL;
 155 }
 156
 157 static void __put_snap_realm(struct ceph_mds_client *mdsc,
 158                              struct ceph_snap_realm *realm);
 159
 160 /*
 161  * called with snap_rwsem (write)
 162  */
 163 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
 164                                  struct ceph_snap_realm *realm)
 165 {
 166         dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
 167
 168         rb_erase(&realm->node, &mdsc->snap_realms);
 169
 170         if (realm->parent) {
 171                 list_del_init(&realm->child_item);
 172                 __put_snap_realm(mdsc, realm->parent);
 173         }
 174
 175         kfree(realm->prior_parent_snaps);
 176         kfree(realm->snaps);
 177         ceph_put_snap_context(realm->cached_context);
 178         kfree(realm);
 179 }
 180
 181 /*
 182  * caller holds snap_rwsem (write)
 183  */
 184 static void __put_snap_realm(struct ceph_mds_client *mdsc,
 185                              struct ceph_snap_realm *realm)
 186 {
 187         dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
 188              atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
 189         if (atomic_dec_and_test(&realm->nref))
 190                 __destroy_snap_realm(mdsc, realm);
 191 }
 192
 193 /*
 194  * caller needn't hold any locks
 195  */
 196 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
 197                          struct ceph_snap_realm *realm)
 198 {
 199         dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
 200              atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
 201         if (!atomic_dec_and_test(&realm->nref))
 202                 return;
 203
 204         if (down_write_trylock(&mdsc->snap_rwsem)) {
 205                 __destroy_snap_realm(mdsc, realm);
 206                 up_write(&mdsc->snap_rwsem);
 207         } else {
 208                 spin_lock(&mdsc->snap_empty_lock);
 209                 list_add(&realm->empty_item, &mdsc->snap_empty);
 210                 spin_unlock(&mdsc->snap_empty_lock);
 211         }
 212 }
 213
 214 /*
 215  * Clean up any realms whose ref counts have dropped to zero.  Note
 216  * that this does not include realms who were created but not yet
 217  * used.
 218  *
 219  * Called under snap_rwsem (write)
 220  */
 221 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
 222 {
 223         struct ceph_snap_realm *realm;
 224
 225         spin_lock(&mdsc->snap_empty_lock);
 226         while (!list_empty(&mdsc->snap_empty)) {
 227                 realm = list_first_entry(&mdsc->snap_empty,
 228                                    struct ceph_snap_realm, empty_item);
 229                 list_del(&realm->empty_item);
 230                 spin_unlock(&mdsc->snap_empty_lock);
 231                 __destroy_snap_realm(mdsc, realm);
 232                 spin_lock(&mdsc->snap_empty_lock);
 233         }
 234         spin_unlock(&mdsc->snap_empty_lock);
 235 }
 236
 237 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
 238 {
 239         down_write(&mdsc->snap_rwsem);
 240         __cleanup_empty_realms(mdsc);
 241         up_write(&mdsc->snap_rwsem);
 242 }
 243
 244 /*
 245  * adjust the parent realm of a given @realm.  adjust child list, and parent
 246  * pointers, and ref counts appropriately.
 247  *
 248  * return true if parent was changed, 0 if unchanged, <0 on error.
 249  *
 250  * caller must hold snap_rwsem for write.
 251  */
 252 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
 253                                     struct ceph_snap_realm *realm,
 254                                     u64 parentino)
 255 {
 256         struct ceph_snap_realm *parent;
 257
 258         if (realm->parent_ino == parentino)
 259                 return 0;
 260
 261         parent = ceph_lookup_snap_realm(mdsc, parentino);
 262         if (!parent) {
 263                 parent = ceph_create_snap_realm(mdsc, parentino);
 264                 if (IS_ERR(parent))
 265                         return PTR_ERR(parent);
 266         }
 267         dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
 268              realm->ino, realm, realm->parent_ino, realm->parent,
 269              parentino, parent);
 270         if (realm->parent) {
 271                 list_del_init(&realm->child_item);
 272                 ceph_put_snap_realm(mdsc, realm->parent);
 273         }
 274         realm->parent_ino = parentino;
 275         realm->parent = parent;
 276         ceph_get_snap_realm(mdsc, parent);
 277         list_add(&realm->child_item, &parent->children);
 278         return 1;
 279 }
 280
 281
 282 static int cmpu64_rev(const void *a, const void *b)
 283 {
 284         if (*(u64 *)a < *(u64 *)b)
 285                 return 1;
 286         if (*(u64 *)a > *(u64 *)b)
 287                 return -1;
 288         return 0;
 289 }
 290
 291 /*
 292  * build the snap context for a given realm.
 293  */
 294 static int build_snap_context(struct ceph_snap_realm *realm)
 295 {
 296         struct ceph_snap_realm *parent = realm->parent;
 297         struct ceph_snap_context *snapc;
 298         int err = 0;
 299         int i;
 300         int num = realm->num_prior_parent_snaps + realm->num_snaps;
 301
 302         /*
 303          * build parent context, if it hasn't been built.
 304          * conservatively estimate that all parent snaps might be
 305          * included by us.
 306          */
 307         if (parent) {
 308                 if (!parent->cached_context) {
 309                         err = build_snap_context(parent);
 310                         if (err)
 311                                 goto fail;
 312                 }
 313                 num += parent->cached_context->num_snaps;
 314         }
 315
 316         /* do i actually need to update?  not if my context seq
 317            matches realm seq, and my parents' does to.  (this works
 318            because we rebuild_snap_realms() works _downward_ in
 319            hierarchy after each update.) */
 320         if (realm->cached_context &&
 321             realm->cached_context->seq == realm->seq &&
 322             (!parent ||
 323              realm->cached_context->seq >= parent->cached_context->seq)) {
 324                 dout("build_snap_context %llx %p: %p seq %lld (%d snaps)"
 325                      " (unchanged)\n",
 326                      realm->ino, realm, realm->cached_context,
 327                      realm->cached_context->seq,
 328                      realm->cached_context->num_snaps);
 329                 return 0;
 330         }
 331
 332         /* alloc new snap context */
 333         err = -ENOMEM;
 334         if (num > (ULONG_MAX - sizeof(*snapc)) / sizeof(u64))
 335                 goto fail;
 336         snapc = kzalloc(sizeof(*snapc) + num*sizeof(u64), GFP_NOFS);
 337         if (!snapc)
 338                 goto fail;
 339         atomic_set(&snapc->nref, 1);
 340
 341         /* build (reverse sorted) snap vector */
 342         num = 0;
 343         snapc->seq = realm->seq;
 344         if (parent) {
 345                 /* include any of parent's snaps occurring _after_ my
 346                    parent became my parent */
 347                 for (i = 0; i < parent->cached_context->num_snaps; i++)
 348                         if (parent->cached_context->snaps[i] >=
 349                             realm->parent_since)
 350                                 snapc->snaps[num++] =
 351                                         parent->cached_context->snaps[i];
 352                 if (parent->cached_context->seq > snapc->seq)
 353                         snapc->seq = parent->cached_context->seq;
 354         }
 355         memcpy(snapc->snaps + num, realm->snaps,
 356                sizeof(u64)*realm->num_snaps);
 357         num += realm->num_snaps;
 358         memcpy(snapc->snaps + num, realm->prior_parent_snaps,
 359                sizeof(u64)*realm->num_prior_parent_snaps);
 360         num += realm->num_prior_parent_snaps;
 361
 362         sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
 363         snapc->num_snaps = num;
 364         dout("build_snap_context %llx %p: %p seq %lld (%d snaps)\n",
 365              realm->ino, realm, snapc, snapc->seq, snapc->num_snaps);
 366
 367         if (realm->cached_context)
 368                 ceph_put_snap_context(realm->cached_context);
 369         realm->cached_context = snapc;
 370         return 0;
 371
 372 fail:
 373         /*
 374          * if we fail, clear old (incorrect) cached_context... hopefully
 375          * we'll have better luck building it later
 376          */
 377         if (realm->cached_context) {
 378                 ceph_put_snap_context(realm->cached_context);
 379                 realm->cached_context = NULL;
 380         }
 381         pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
 382                realm, err);
 383         return err;
 384 }
 385
 386 /*
 387  * rebuild snap context for the given realm and all of its children.
 388  */
 389 static void rebuild_snap_realms(struct ceph_snap_realm *realm)
 390 {
 391         struct ceph_snap_realm *child;
 392
 393         dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
 394         build_snap_context(realm);
 395
 396         list_for_each_entry(child, &realm->children, child_item)
 397                 rebuild_snap_realms(child);
 398 }
 399
 400
 401 /*
 402  * helper to allocate and decode an array of snapids.  free prior
 403  * instance, if any.
 404  */
 405 static int dup_array(u64 **dst, __le64 *src, int num)
 406 {
 407         int i;
 408
 409         kfree(*dst);
 410         if (num) {
 411                 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
 412                 if (!*dst)
 413                         return -ENOMEM;
 414                 for (i = 0; i < num; i++)
 415                         (*dst)[i] = get_unaligned_le64(src + i);
 416         } else {
 417                 *dst = NULL;
 418         }
 419         return 0;
 420 }
 421
 422
 423 /*
 424  * When a snapshot is applied, the size/mtime inode metadata is queued
 425  * in a ceph_cap_snap (one for each snapshot) until writeback
 426  * completes and the metadata can be flushed back to the MDS.
 427  *
 428  * However, if a (sync) write is currently in-progress when we apply
 429  * the snapshot, we have to wait until the write succeeds or fails
 430  * (and a final size/mtime is known).  In this case the
 431  * cap_snap->writing = 1, and is said to be "pending."  When the write
 432  * finishes, we __ceph_finish_cap_snap().
 433  *
 434  * Caller must hold snap_rwsem for read (i.e., the realm topology won't
 435  * change).
 436  */
 437 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
 438 {
 439         struct inode *inode = &ci->vfs_inode;
 440         struct ceph_cap_snap *capsnap;
 441         int used, dirty;
 442
 443         capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
 444         if (!capsnap) {
 445                 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
 446                 return;
 447         }
 448
 449         spin_lock(&ci->i_ceph_lock);
 450         used = __ceph_caps_used(ci);
 451         dirty = __ceph_caps_dirty(ci);
 452
 453         /*
 454          * If there is a write in progress, treat that as a dirty Fw,
 455          * even though it hasn't completed yet; by the time we finish
 456          * up this capsnap it will be.
 457          */
 458         if (used & CEPH_CAP_FILE_WR)
 459                 dirty |= CEPH_CAP_FILE_WR;
 460
 461         if (__ceph_have_pending_cap_snap(ci)) {
 462                 /* there is no point in queuing multiple "pending" cap_snaps,
 463                    as no new writes are allowed to start when pending, so any
 464                    writes in progress now were started before the previous
 465                    cap_snap.  lucky us. */
 466                 dout("queue_cap_snap %p already pending\n", inode);
 467                 kfree(capsnap);
 468         } else if (dirty & (CEPH_CAP_AUTH_EXCL|CEPH_CAP_XATTR_EXCL|
 469                             CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_WR)) {
 470                 struct ceph_snap_context *snapc = ci->i_head_snapc;
 471
 472                 /*
 473                  * if we are a sync write, we may need to go to the snaprealm
 474                  * to get the current snapc.
 475                  */
 476                 if (!snapc)
 477                         snapc = ci->i_snap_realm->cached_context;
 478
 479                 dout("queue_cap_snap %p cap_snap %p queuing under %p %s\n",
 480                      inode, capsnap, snapc, ceph_cap_string(dirty));
 481                 ihold(inode);
 482
 483                 atomic_set(&capsnap->nref, 1);
 484                 capsnap->ci = ci;
 485                 INIT_LIST_HEAD(&capsnap->ci_item);
 486                 INIT_LIST_HEAD(&capsnap->flushing_item);
 487
 488                 capsnap->follows = snapc->seq;
 489                 capsnap->issued = __ceph_caps_issued(ci, NULL);
 490                 capsnap->dirty = dirty;
 491
 492                 capsnap->mode = inode->i_mode;
 493                 capsnap->uid = inode->i_uid;
 494                 capsnap->gid = inode->i_gid;
 495
 496                 if (dirty & CEPH_CAP_XATTR_EXCL) {
 497                         __ceph_build_xattrs_blob(ci);
 498                         capsnap->xattr_blob =
 499                                 ceph_buffer_get(ci->i_xattrs.blob);
 500                         capsnap->xattr_version = ci->i_xattrs.version;
 501                 } else {
 502                         capsnap->xattr_blob = NULL;
 503                         capsnap->xattr_version = 0;
 504                 }
 505
 506                 /* dirty page count moved from _head to this cap_snap;
 507                    all subsequent writes page dirties occur _after_ this
 508                    snapshot. */
 509                 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
 510                 ci->i_wrbuffer_ref_head = 0;
 511                 capsnap->context = snapc;
 512                 ci->i_head_snapc =
 513                         ceph_get_snap_context(ci->i_snap_realm->cached_context);
 514                 dout(" new snapc is %p\n", ci->i_head_snapc);
 515                 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
 516
 517                 if (used & CEPH_CAP_FILE_WR) {
 518                         dout("queue_cap_snap %p cap_snap %p snapc %p"
 519                              " seq %llu used WR, now pending\n", inode,
 520                              capsnap, snapc, snapc->seq);
 521                         capsnap->writing = 1;
 522                 } else {
 523                         /* note mtime, size NOW. */
 524                         __ceph_finish_cap_snap(ci, capsnap);
 525                 }
 526         } else {
 527                 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
 528                 kfree(capsnap);
 529         }
 530
 531         spin_unlock(&ci->i_ceph_lock);
 532 }
 533
 534 /*
 535  * Finalize the size, mtime for a cap_snap.. that is, settle on final values
 536  * to be used for the snapshot, to be flushed back to the mds.
 537  *
 538  * If capsnap can now be flushed, add to snap_flush list, and return 1.
 539  *
 540  * Caller must hold i_ceph_lock.
 541  */
 542 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
 543                             struct ceph_cap_snap *capsnap)
 544 {
 545         struct inode *inode = &ci->vfs_inode;
 546         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
 547
 548         BUG_ON(capsnap->writing);
 549         capsnap->size = inode->i_size;
 550         capsnap->mtime = inode->i_mtime;
 551         capsnap->atime = inode->i_atime;
 552         capsnap->ctime = inode->i_ctime;
 553         capsnap->time_warp_seq = ci->i_time_warp_seq;
 554         if (capsnap->dirty_pages) {
 555                 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
 556                      "still has %d dirty pages\n", inode, capsnap,
 557                      capsnap->context, capsnap->context->seq,
 558                      ceph_cap_string(capsnap->dirty), capsnap->size,
 559                      capsnap->dirty_pages);
 560                 return 0;
 561         }
 562         dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
 563              inode, capsnap, capsnap->context,
 564              capsnap->context->seq, ceph_cap_string(capsnap->dirty),
 565              capsnap->size);
 566
 567         spin_lock(&mdsc->snap_flush_lock);
 568         list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
 569         spin_unlock(&mdsc->snap_flush_lock);
 570         return 1;  /* caller may want to ceph_flush_snaps */
 571 }
 572
 573 /*
 574  * Queue cap_snaps for snap writeback for this realm and its children.
 575  * Called under snap_rwsem, so realm topology won't change.
 576  */
 577 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
 578 {
 579         struct ceph_inode_info *ci;
 580         struct inode *lastinode = NULL;
 581         struct ceph_snap_realm *child;
 582
 583         dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
 584
 585         spin_lock(&realm->inodes_with_caps_lock);
 586         list_for_each_entry(ci, &realm->inodes_with_caps,
 587                             i_snap_realm_item) {
 588                 struct inode *inode = igrab(&ci->vfs_inode);
 589                 if (!inode)
 590                         continue;
 591                 spin_unlock(&realm->inodes_with_caps_lock);
 592                 if (lastinode)
 593                         iput(lastinode);
 594                 lastinode = inode;
 595                 ceph_queue_cap_snap(ci);
 596                 spin_lock(&realm->inodes_with_caps_lock);
 597         }
 598         spin_unlock(&realm->inodes_with_caps_lock);
 599         if (lastinode)
 600                 iput(lastinode);
 601
 602         list_for_each_entry(child, &realm->children, child_item) {
 603                 dout("queue_realm_cap_snaps %p %llx queue child %p %llx\n",
 604                      realm, realm->ino, child, child->ino);
 605                 list_del_init(&child->dirty_item);
 606                 list_add(&child->dirty_item, &realm->dirty_item);
 607         }
 608
 609         list_del_init(&realm->dirty_item);
 610         dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
 611 }
 612
 613 /*
 614  * Parse and apply a snapblob "snap trace" from the MDS.  This specifies
 615  * the snap realm parameters from a given realm and all of its ancestors,
 616  * up to the root.
 617  *
 618  * Caller must hold snap_rwsem for write.
 619  */
 620 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
 621                            void *p, void *e, bool deletion)
 622 {
 623         struct ceph_mds_snap_realm *ri;    /* encoded */
 624         __le64 *snaps;                     /* encoded */
 625         __le64 *prior_parent_snaps;        /* encoded */
 626         struct ceph_snap_realm *realm;
 627         int invalidate = 0;
 628         int err = -ENOMEM;
 629         LIST_HEAD(dirty_realms);
 630
 631         dout("update_snap_trace deletion=%d\n", deletion);
 632 more:
 633         ceph_decode_need(&p, e, sizeof(*ri), bad);
 634         ri = p;
 635         p += sizeof(*ri);
 636         ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
 637                             le32_to_cpu(ri->num_prior_parent_snaps)), bad);
 638         snaps = p;
 639         p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
 640         prior_parent_snaps = p;
 641         p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
 642
 643         realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
 644         if (!realm) {
 645                 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
 646                 if (IS_ERR(realm)) {
 647                         err = PTR_ERR(realm);
 648                         goto fail;
 649                 }
 650         }
 651
 652         /* ensure the parent is correct */
 653         err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
 654         if (err < 0)
 655                 goto fail;
 656         invalidate += err;
 657
 658         if (le64_to_cpu(ri->seq) > realm->seq) {
 659                 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
 660                      realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
 661                 /* update realm parameters, snap lists */
 662                 realm->seq = le64_to_cpu(ri->seq);
 663                 realm->created = le64_to_cpu(ri->created);
 664                 realm->parent_since = le64_to_cpu(ri->parent_since);
 665
 666                 realm->num_snaps = le32_to_cpu(ri->num_snaps);
 667                 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
 668                 if (err < 0)
 669                         goto fail;
 670
 671                 realm->num_prior_parent_snaps =
 672                         le32_to_cpu(ri->num_prior_parent_snaps);
 673                 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
 674                                 realm->num_prior_parent_snaps);
 675                 if (err < 0)
 676                         goto fail;
 677
 678                 /* queue realm for cap_snap creation */
 679                 list_add(&realm->dirty_item, &dirty_realms);
 680
 681                 invalidate = 1;
 682         } else if (!realm->cached_context) {
 683                 dout("update_snap_trace %llx %p seq %lld new\n",
 684                      realm->ino, realm, realm->seq);
 685                 invalidate = 1;
 686         } else {
 687                 dout("update_snap_trace %llx %p seq %lld unchanged\n",
 688                      realm->ino, realm, realm->seq);
 689         }
 690
 691         dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
 692              realm, invalidate, p, e);
 693
 694         if (p < e)
 695                 goto more;
 696
 697         /* invalidate when we reach the _end_ (root) of the trace */
 698         if (invalidate)
 699                 rebuild_snap_realms(realm);
 700
 701         /*
 702          * queue cap snaps _after_ we've built the new snap contexts,
 703          * so that i_head_snapc can be set appropriately.
 704          */
 705         while (!list_empty(&dirty_realms)) {
 706                 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
 707                                          dirty_item);
 708                 queue_realm_cap_snaps(realm);
 709         }
 710
 711         __cleanup_empty_realms(mdsc);
 712         return 0;
 713
 714 bad:
 715         err = -EINVAL;
 716 fail:
 717         pr_err("update_snap_trace error %d\n", err);
 718         return err;
 719 }
 720
 721
 722 /*
 723  * Send any cap_snaps that are queued for flush.  Try to carry
 724  * s_mutex across multiple snap flushes to avoid locking overhead.
 725  *
 726  * Caller holds no locks.
 727  */
 728 static void flush_snaps(struct ceph_mds_client *mdsc)
 729 {
 730         struct ceph_inode_info *ci;
 731         struct inode *inode;
 732         struct ceph_mds_session *session = NULL;
 733
 734         dout("flush_snaps\n");
 735         spin_lock(&mdsc->snap_flush_lock);
 736         while (!list_empty(&mdsc->snap_flush_list)) {
 737                 ci = list_first_entry(&mdsc->snap_flush_list,
 738                                 struct ceph_inode_info, i_snap_flush_item);
 739                 inode = &ci->vfs_inode;
 740                 ihold(inode);
 741                 spin_unlock(&mdsc->snap_flush_lock);
 742                 spin_lock(&ci->i_ceph_lock);
 743                 __ceph_flush_snaps(ci, &session, 0);
 744                 spin_unlock(&ci->i_ceph_lock);
 745                 iput(inode);
 746                 spin_lock(&mdsc->snap_flush_lock);
 747         }
 748         spin_unlock(&mdsc->snap_flush_lock);
 749
 750         if (session) {
 751                 mutex_unlock(&session->s_mutex);
 752                 ceph_put_mds_session(session);
 753         }
 754         dout("flush_snaps done\n");
 755 }
 756
 757
 758 /*
 759  * Handle a snap notification from the MDS.
 760  *
 761  * This can take two basic forms: the simplest is just a snap creation
 762  * or deletion notification on an existing realm.  This should update the
 763  * realm and its children.
 764  *
 765  * The more difficult case is realm creation, due to snap creation at a
 766  * new point in the file hierarchy, or due to a rename that moves a file or
 767  * directory into another realm.
 768  */
 769 void ceph_handle_snap(struct ceph_mds_client *mdsc,
 770                       struct ceph_mds_session *session,
 771                       struct ceph_msg *msg)
 772 {
 773         struct super_block *sb = mdsc->fsc->sb;
 774         int mds = session->s_mds;
 775         u64 split;
 776         int op;
 777         int trace_len;
 778         struct ceph_snap_realm *realm = NULL;
 779         void *p = msg->front.iov_base;
 780         void *e = p + msg->front.iov_len;
 781         struct ceph_mds_snap_head *h;
 782         int num_split_inos, num_split_realms;
 783         __le64 *split_inos = NULL, *split_realms = NULL;
 784         int i;
 785         int locked_rwsem = 0;
 786
 787         /* decode */
 788         if (msg->front.iov_len < sizeof(*h))
 789                 goto bad;
 790         h = p;
 791         op = le32_to_cpu(h->op);
 792         split = le64_to_cpu(h->split);   /* non-zero if we are splitting an
 793                                           * existing realm */
 794         num_split_inos = le32_to_cpu(h->num_split_inos);
 795         num_split_realms = le32_to_cpu(h->num_split_realms);
 796         trace_len = le32_to_cpu(h->trace_len);
 797         p += sizeof(*h);
 798
 799         dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
 800              ceph_snap_op_name(op), split, trace_len);
 801
 802         mutex_lock(&session->s_mutex);
 803         session->s_seq++;
 804         mutex_unlock(&session->s_mutex);
 805
 806         down_write(&mdsc->snap_rwsem);
 807         locked_rwsem = 1;
 808
 809         if (op == CEPH_SNAP_OP_SPLIT) {
 810                 struct ceph_mds_snap_realm *ri;
 811
 812                 /*
 813                  * A "split" breaks part of an existing realm off into
 814                  * a new realm.  The MDS provides a list of inodes
 815                  * (with caps) and child realms that belong to the new
 816                  * child.
 817                  */
 818                 split_inos = p;
 819                 p += sizeof(u64) * num_split_inos;
 820                 split_realms = p;
 821                 p += sizeof(u64) * num_split_realms;
 822                 ceph_decode_need(&p, e, sizeof(*ri), bad);
 823                 /* we will peek at realm info here, but will _not_
 824                  * advance p, as the realm update will occur below in
 825                  * ceph_update_snap_trace. */
 826                 ri = p;
 827
 828                 realm = ceph_lookup_snap_realm(mdsc, split);
 829                 if (!realm) {
 830                         realm = ceph_create_snap_realm(mdsc, split);
 831                         if (IS_ERR(realm))
 832                                 goto out;
 833                 }
 834                 ceph_get_snap_realm(mdsc, realm);
 835
 836                 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
 837                 for (i = 0; i < num_split_inos; i++) {
 838                         struct ceph_vino vino = {
 839                                 .ino = le64_to_cpu(split_inos[i]),
 840                                 .snap = CEPH_NOSNAP,
 841                         };
 842                         struct inode *inode = ceph_find_inode(sb, vino);
 843                         struct ceph_inode_info *ci;
 844                         struct ceph_snap_realm *oldrealm;
 845
 846                         if (!inode)
 847                                 continue;
 848                         ci = ceph_inode(inode);
 849
 850                         spin_lock(&ci->i_ceph_lock);
 851                         if (!ci->i_snap_realm)
 852                                 goto skip_inode;
 853                         /*
 854                          * If this inode belongs to a realm that was
 855                          * created after our new realm, we experienced
 856                          * a race (due to another split notifications
 857                          * arriving from a different MDS).  So skip
 858                          * this inode.
 859                          */
 860                         if (ci->i_snap_realm->created >
 861                             le64_to_cpu(ri->created)) {
 862                                 dout(" leaving %p in newer realm %llx %p\n",
 863                                      inode, ci->i_snap_realm->ino,
 864                                      ci->i_snap_realm);
 865                                 goto skip_inode;
 866                         }
 867                         dout(" will move %p to split realm %llx %p\n",
 868                              inode, realm->ino, realm);
 869                         /*
 870                          * Move the inode to the new realm
 871                          */
 872                         spin_lock(&realm->inodes_with_caps_lock);
 873                         list_del_init(&ci->i_snap_realm_item);
 874                         list_add(&ci->i_snap_realm_item,
 875                                  &realm->inodes_with_caps);
 876                         oldrealm = ci->i_snap_realm;
 877                         ci->i_snap_realm = realm;
 878                         spin_unlock(&realm->inodes_with_caps_lock);
 879                         spin_unlock(&ci->i_ceph_lock);
 880
 881                         ceph_get_snap_realm(mdsc, realm);
 882                         ceph_put_snap_realm(mdsc, oldrealm);
 883
 884                         iput(inode);
 885                         continue;
 886
 887 skip_inode:
 888                         spin_unlock(&ci->i_ceph_lock);
 889                         iput(inode);
 890                 }
 891
 892                 /* we may have taken some of the old realm's children. */
 893                 for (i = 0; i < num_split_realms; i++) {
 894                         struct ceph_snap_realm *child =
 895                                 ceph_lookup_snap_realm(mdsc,
 896                                            le64_to_cpu(split_realms[i]));
 897                         if (!child)
 898                                 continue;
 899                         adjust_snap_realm_parent(mdsc, child, realm->ino);
 900                 }
 901         }
 902
 903         /*
 904          * update using the provided snap trace. if we are deleting a
 905          * snap, we can avoid queueing cap_snaps.
 906          */
 907         ceph_update_snap_trace(mdsc, p, e,
 908                                op == CEPH_SNAP_OP_DESTROY);
 909
 910         if (op == CEPH_SNAP_OP_SPLIT)
 911                 /* we took a reference when we created the realm, above */
 912                 ceph_put_snap_realm(mdsc, realm);
 913
 914         __cleanup_empty_realms(mdsc);
 915
 916         up_write(&mdsc->snap_rwsem);
 917
 918         flush_snaps(mdsc);
 919         return;
 920
 921 bad:
 922         pr_err("corrupt snap message from mds%d\n", mds);
 923         ceph_msg_dump(msg);
 924 out:
 925         if (locked_rwsem)
 926                 up_write(&mdsc->snap_rwsem);
 927         return;
 928 }
 929
 930
 931