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_inc_return(&realm->nref) == 1) { 74 spin_lock(&mdsc->snap_empty_lock); 75 list_del_init(&realm->empty_item); 76 spin_unlock(&mdsc->snap_empty_lock); 77 } 78 } 79 80 static void __insert_snap_realm(struct rb_root *root, 81 struct ceph_snap_realm *new) 82 { 83 struct rb_node **p = &root->rb_node; 84 struct rb_node *parent = NULL; 85 struct ceph_snap_realm *r = NULL; 86 87 while (*p) { 88 parent = *p; 89 r = rb_entry(parent, struct ceph_snap_realm, node); 90 if (new->ino < r->ino) 91 p = &(*p)->rb_left; 92 else if (new->ino > r->ino) 93 p = &(*p)->rb_right; 94 else 95 BUG(); 96 } 97 98 rb_link_node(&new->node, parent, p); 99 rb_insert_color(&new->node, root); 100 } 101 102 /* 103 * create and get the realm rooted at @ino and bump its ref count. 104 * 105 * caller must hold snap_rwsem for write. 106 */ 107 static struct ceph_snap_realm *ceph_create_snap_realm( 108 struct ceph_mds_client *mdsc, 109 u64 ino) 110 { 111 struct ceph_snap_realm *realm; 112 113 realm = kzalloc(sizeof(*realm), GFP_NOFS); 114 if (!realm) 115 return ERR_PTR(-ENOMEM); 116 117 atomic_set(&realm->nref, 1); /* for caller */ 118 realm->ino = ino; 119 INIT_LIST_HEAD(&realm->children); 120 INIT_LIST_HEAD(&realm->child_item); 121 INIT_LIST_HEAD(&realm->empty_item); 122 INIT_LIST_HEAD(&realm->dirty_item); 123 INIT_LIST_HEAD(&realm->inodes_with_caps); 124 spin_lock_init(&realm->inodes_with_caps_lock); 125 __insert_snap_realm(&mdsc->snap_realms, realm); 126 dout("create_snap_realm %llx %p\n", realm->ino, realm); 127 return realm; 128 } 129 130 /* 131 * lookup the realm rooted at @ino. 132 * 133 * caller must hold snap_rwsem for write. 134 */ 135 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc, 136 u64 ino) 137 { 138 struct rb_node *n = mdsc->snap_realms.rb_node; 139 struct ceph_snap_realm *r; 140 141 while (n) { 142 r = rb_entry(n, struct ceph_snap_realm, node); 143 if (ino < r->ino) 144 n = n->rb_left; 145 else if (ino > r->ino) 146 n = n->rb_right; 147 else { 148 dout("lookup_snap_realm %llx %p\n", r->ino, r); 149 return r; 150 } 151 } 152 return NULL; 153 } 154 155 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc, 156 u64 ino) 157 { 158 struct ceph_snap_realm *r; 159 r = __lookup_snap_realm(mdsc, ino); 160 if (r) 161 ceph_get_snap_realm(mdsc, r); 162 return r; 163 } 164 165 static void __put_snap_realm(struct ceph_mds_client *mdsc, 166 struct ceph_snap_realm *realm); 167 168 /* 169 * called with snap_rwsem (write) 170 */ 171 static void __destroy_snap_realm(struct ceph_mds_client *mdsc, 172 struct ceph_snap_realm *realm) 173 { 174 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino); 175 176 rb_erase(&realm->node, &mdsc->snap_realms); 177 178 if (realm->parent) { 179 list_del_init(&realm->child_item); 180 __put_snap_realm(mdsc, realm->parent); 181 } 182 183 kfree(realm->prior_parent_snaps); 184 kfree(realm->snaps); 185 ceph_put_snap_context(realm->cached_context); 186 kfree(realm); 187 } 188 189 /* 190 * caller holds snap_rwsem (write) 191 */ 192 static void __put_snap_realm(struct ceph_mds_client *mdsc, 193 struct ceph_snap_realm *realm) 194 { 195 dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm, 196 atomic_read(&realm->nref), atomic_read(&realm->nref)-1); 197 if (atomic_dec_and_test(&realm->nref)) 198 __destroy_snap_realm(mdsc, realm); 199 } 200 201 /* 202 * caller needn't hold any locks 203 */ 204 void ceph_put_snap_realm(struct ceph_mds_client *mdsc, 205 struct ceph_snap_realm *realm) 206 { 207 dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm, 208 atomic_read(&realm->nref), atomic_read(&realm->nref)-1); 209 if (!atomic_dec_and_test(&realm->nref)) 210 return; 211 212 if (down_write_trylock(&mdsc->snap_rwsem)) { 213 __destroy_snap_realm(mdsc, realm); 214 up_write(&mdsc->snap_rwsem); 215 } else { 216 spin_lock(&mdsc->snap_empty_lock); 217 list_add(&realm->empty_item, &mdsc->snap_empty); 218 spin_unlock(&mdsc->snap_empty_lock); 219 } 220 } 221 222 /* 223 * Clean up any realms whose ref counts have dropped to zero. Note 224 * that this does not include realms who were created but not yet 225 * used. 226 * 227 * Called under snap_rwsem (write) 228 */ 229 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc) 230 { 231 struct ceph_snap_realm *realm; 232 233 spin_lock(&mdsc->snap_empty_lock); 234 while (!list_empty(&mdsc->snap_empty)) { 235 realm = list_first_entry(&mdsc->snap_empty, 236 struct ceph_snap_realm, empty_item); 237 list_del(&realm->empty_item); 238 spin_unlock(&mdsc->snap_empty_lock); 239 __destroy_snap_realm(mdsc, realm); 240 spin_lock(&mdsc->snap_empty_lock); 241 } 242 spin_unlock(&mdsc->snap_empty_lock); 243 } 244 245 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc) 246 { 247 down_write(&mdsc->snap_rwsem); 248 __cleanup_empty_realms(mdsc); 249 up_write(&mdsc->snap_rwsem); 250 } 251 252 /* 253 * adjust the parent realm of a given @realm. adjust child list, and parent 254 * pointers, and ref counts appropriately. 255 * 256 * return true if parent was changed, 0 if unchanged, <0 on error. 257 * 258 * caller must hold snap_rwsem for write. 259 */ 260 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc, 261 struct ceph_snap_realm *realm, 262 u64 parentino) 263 { 264 struct ceph_snap_realm *parent; 265 266 if (realm->parent_ino == parentino) 267 return 0; 268 269 parent = ceph_lookup_snap_realm(mdsc, parentino); 270 if (!parent) { 271 parent = ceph_create_snap_realm(mdsc, parentino); 272 if (IS_ERR(parent)) 273 return PTR_ERR(parent); 274 } 275 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n", 276 realm->ino, realm, realm->parent_ino, realm->parent, 277 parentino, parent); 278 if (realm->parent) { 279 list_del_init(&realm->child_item); 280 ceph_put_snap_realm(mdsc, realm->parent); 281 } 282 realm->parent_ino = parentino; 283 realm->parent = parent; 284 list_add(&realm->child_item, &parent->children); 285 return 1; 286 } 287 288 289 static int cmpu64_rev(const void *a, const void *b) 290 { 291 if (*(u64 *)a < *(u64 *)b) 292 return 1; 293 if (*(u64 *)a > *(u64 *)b) 294 return -1; 295 return 0; 296 } 297 298 299 struct ceph_snap_context *ceph_empty_snapc; 300 301 /* 302 * build the snap context for a given realm. 303 */ 304 static int build_snap_context(struct ceph_snap_realm *realm) 305 { 306 struct ceph_snap_realm *parent = realm->parent; 307 struct ceph_snap_context *snapc; 308 int err = 0; 309 u32 num = realm->num_prior_parent_snaps + realm->num_snaps; 310 311 /* 312 * build parent context, if it hasn't been built. 313 * conservatively estimate that all parent snaps might be 314 * included by us. 315 */ 316 if (parent) { 317 if (!parent->cached_context) { 318 err = build_snap_context(parent); 319 if (err) 320 goto fail; 321 } 322 num += parent->cached_context->num_snaps; 323 } 324 325 /* do i actually need to update? not if my context seq 326 matches realm seq, and my parents' does to. (this works 327 because we rebuild_snap_realms() works _downward_ in 328 hierarchy after each update.) */ 329 if (realm->cached_context && 330 realm->cached_context->seq == realm->seq && 331 (!parent || 332 realm->cached_context->seq >= parent->cached_context->seq)) { 333 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)" 334 " (unchanged)\n", 335 realm->ino, realm, realm->cached_context, 336 realm->cached_context->seq, 337 (unsigned int) realm->cached_context->num_snaps); 338 return 0; 339 } 340 341 if (num == 0 && realm->seq == ceph_empty_snapc->seq) { 342 ceph_get_snap_context(ceph_empty_snapc); 343 snapc = ceph_empty_snapc; 344 goto done; 345 } 346 347 /* alloc new snap context */ 348 err = -ENOMEM; 349 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64)) 350 goto fail; 351 snapc = ceph_create_snap_context(num, GFP_NOFS); 352 if (!snapc) 353 goto fail; 354 355 /* build (reverse sorted) snap vector */ 356 num = 0; 357 snapc->seq = realm->seq; 358 if (parent) { 359 u32 i; 360 361 /* include any of parent's snaps occurring _after_ my 362 parent became my parent */ 363 for (i = 0; i < parent->cached_context->num_snaps; i++) 364 if (parent->cached_context->snaps[i] >= 365 realm->parent_since) 366 snapc->snaps[num++] = 367 parent->cached_context->snaps[i]; 368 if (parent->cached_context->seq > snapc->seq) 369 snapc->seq = parent->cached_context->seq; 370 } 371 memcpy(snapc->snaps + num, realm->snaps, 372 sizeof(u64)*realm->num_snaps); 373 num += realm->num_snaps; 374 memcpy(snapc->snaps + num, realm->prior_parent_snaps, 375 sizeof(u64)*realm->num_prior_parent_snaps); 376 num += realm->num_prior_parent_snaps; 377 378 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL); 379 snapc->num_snaps = num; 380 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n", 381 realm->ino, realm, snapc, snapc->seq, 382 (unsigned int) snapc->num_snaps); 383 384 done: 385 ceph_put_snap_context(realm->cached_context); 386 realm->cached_context = snapc; 387 return 0; 388 389 fail: 390 /* 391 * if we fail, clear old (incorrect) cached_context... hopefully 392 * we'll have better luck building it later 393 */ 394 if (realm->cached_context) { 395 ceph_put_snap_context(realm->cached_context); 396 realm->cached_context = NULL; 397 } 398 pr_err("build_snap_context %llx %p fail %d\n", realm->ino, 399 realm, err); 400 return err; 401 } 402 403 /* 404 * rebuild snap context for the given realm and all of its children. 405 */ 406 static void rebuild_snap_realms(struct ceph_snap_realm *realm) 407 { 408 struct ceph_snap_realm *child; 409 410 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm); 411 build_snap_context(realm); 412 413 list_for_each_entry(child, &realm->children, child_item) 414 rebuild_snap_realms(child); 415 } 416 417 418 /* 419 * helper to allocate and decode an array of snapids. free prior 420 * instance, if any. 421 */ 422 static int dup_array(u64 **dst, __le64 *src, u32 num) 423 { 424 u32 i; 425 426 kfree(*dst); 427 if (num) { 428 *dst = kcalloc(num, sizeof(u64), GFP_NOFS); 429 if (!*dst) 430 return -ENOMEM; 431 for (i = 0; i < num; i++) 432 (*dst)[i] = get_unaligned_le64(src + i); 433 } else { 434 *dst = NULL; 435 } 436 return 0; 437 } 438 439 static bool has_new_snaps(struct ceph_snap_context *o, 440 struct ceph_snap_context *n) 441 { 442 if (n->num_snaps == 0) 443 return false; 444 /* snaps are in descending order */ 445 return n->snaps[0] > o->seq; 446 } 447 448 /* 449 * When a snapshot is applied, the size/mtime inode metadata is queued 450 * in a ceph_cap_snap (one for each snapshot) until writeback 451 * completes and the metadata can be flushed back to the MDS. 452 * 453 * However, if a (sync) write is currently in-progress when we apply 454 * the snapshot, we have to wait until the write succeeds or fails 455 * (and a final size/mtime is known). In this case the 456 * cap_snap->writing = 1, and is said to be "pending." When the write 457 * finishes, we __ceph_finish_cap_snap(). 458 * 459 * Caller must hold snap_rwsem for read (i.e., the realm topology won't 460 * change). 461 */ 462 void ceph_queue_cap_snap(struct ceph_inode_info *ci) 463 { 464 struct inode *inode = &ci->vfs_inode; 465 struct ceph_cap_snap *capsnap; 466 struct ceph_snap_context *old_snapc, *new_snapc; 467 int used, dirty; 468 469 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS); 470 if (!capsnap) { 471 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode); 472 return; 473 } 474 475 spin_lock(&ci->i_ceph_lock); 476 used = __ceph_caps_used(ci); 477 dirty = __ceph_caps_dirty(ci); 478 479 old_snapc = ci->i_head_snapc; 480 new_snapc = ci->i_snap_realm->cached_context; 481 482 /* 483 * If there is a write in progress, treat that as a dirty Fw, 484 * even though it hasn't completed yet; by the time we finish 485 * up this capsnap it will be. 486 */ 487 if (used & CEPH_CAP_FILE_WR) 488 dirty |= CEPH_CAP_FILE_WR; 489 490 if (__ceph_have_pending_cap_snap(ci)) { 491 /* there is no point in queuing multiple "pending" cap_snaps, 492 as no new writes are allowed to start when pending, so any 493 writes in progress now were started before the previous 494 cap_snap. lucky us. */ 495 dout("queue_cap_snap %p already pending\n", inode); 496 goto update_snapc; 497 } 498 if (ci->i_wrbuffer_ref_head == 0 && 499 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) { 500 dout("queue_cap_snap %p nothing dirty|writing\n", inode); 501 goto update_snapc; 502 } 503 504 BUG_ON(!old_snapc); 505 506 /* 507 * There is no need to send FLUSHSNAP message to MDS if there is 508 * no new snapshot. But when there is dirty pages or on-going 509 * writes, we still need to create cap_snap. cap_snap is needed 510 * by the write path and page writeback path. 511 * 512 * also see ceph_try_drop_cap_snap() 513 */ 514 if (has_new_snaps(old_snapc, new_snapc)) { 515 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR)) 516 capsnap->need_flush = true; 517 } else { 518 if (!(used & CEPH_CAP_FILE_WR) && 519 ci->i_wrbuffer_ref_head == 0) { 520 dout("queue_cap_snap %p " 521 "no new_snap|dirty_page|writing\n", inode); 522 goto update_snapc; 523 } 524 } 525 526 dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n", 527 inode, capsnap, old_snapc, ceph_cap_string(dirty), 528 capsnap->need_flush ? "" : "no_flush"); 529 ihold(inode); 530 531 atomic_set(&capsnap->nref, 1); 532 capsnap->ci = ci; 533 INIT_LIST_HEAD(&capsnap->ci_item); 534 INIT_LIST_HEAD(&capsnap->flushing_item); 535 536 capsnap->follows = old_snapc->seq; 537 capsnap->issued = __ceph_caps_issued(ci, NULL); 538 capsnap->dirty = dirty; 539 540 capsnap->mode = inode->i_mode; 541 capsnap->uid = inode->i_uid; 542 capsnap->gid = inode->i_gid; 543 544 if (dirty & CEPH_CAP_XATTR_EXCL) { 545 __ceph_build_xattrs_blob(ci); 546 capsnap->xattr_blob = 547 ceph_buffer_get(ci->i_xattrs.blob); 548 capsnap->xattr_version = ci->i_xattrs.version; 549 } else { 550 capsnap->xattr_blob = NULL; 551 capsnap->xattr_version = 0; 552 } 553 554 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE; 555 556 /* dirty page count moved from _head to this cap_snap; 557 all subsequent writes page dirties occur _after_ this 558 snapshot. */ 559 capsnap->dirty_pages = ci->i_wrbuffer_ref_head; 560 ci->i_wrbuffer_ref_head = 0; 561 capsnap->context = old_snapc; 562 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps); 563 old_snapc = NULL; 564 565 if (used & CEPH_CAP_FILE_WR) { 566 dout("queue_cap_snap %p cap_snap %p snapc %p" 567 " seq %llu used WR, now pending\n", inode, 568 capsnap, old_snapc, old_snapc->seq); 569 capsnap->writing = 1; 570 } else { 571 /* note mtime, size NOW. */ 572 __ceph_finish_cap_snap(ci, capsnap); 573 } 574 capsnap = NULL; 575 576 update_snapc: 577 if (ci->i_head_snapc) { 578 ci->i_head_snapc = ceph_get_snap_context(new_snapc); 579 dout(" new snapc is %p\n", new_snapc); 580 } 581 spin_unlock(&ci->i_ceph_lock); 582 583 kfree(capsnap); 584 ceph_put_snap_context(old_snapc); 585 } 586 587 /* 588 * Finalize the size, mtime for a cap_snap.. that is, settle on final values 589 * to be used for the snapshot, to be flushed back to the mds. 590 * 591 * If capsnap can now be flushed, add to snap_flush list, and return 1. 592 * 593 * Caller must hold i_ceph_lock. 594 */ 595 int __ceph_finish_cap_snap(struct ceph_inode_info *ci, 596 struct ceph_cap_snap *capsnap) 597 { 598 struct inode *inode = &ci->vfs_inode; 599 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 600 601 BUG_ON(capsnap->writing); 602 capsnap->size = inode->i_size; 603 capsnap->mtime = inode->i_mtime; 604 capsnap->atime = inode->i_atime; 605 capsnap->ctime = inode->i_ctime; 606 capsnap->time_warp_seq = ci->i_time_warp_seq; 607 if (capsnap->dirty_pages) { 608 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu " 609 "still has %d dirty pages\n", inode, capsnap, 610 capsnap->context, capsnap->context->seq, 611 ceph_cap_string(capsnap->dirty), capsnap->size, 612 capsnap->dirty_pages); 613 return 0; 614 } 615 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n", 616 inode, capsnap, capsnap->context, 617 capsnap->context->seq, ceph_cap_string(capsnap->dirty), 618 capsnap->size); 619 620 spin_lock(&mdsc->snap_flush_lock); 621 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list); 622 spin_unlock(&mdsc->snap_flush_lock); 623 return 1; /* caller may want to ceph_flush_snaps */ 624 } 625 626 /* 627 * Queue cap_snaps for snap writeback for this realm and its children. 628 * Called under snap_rwsem, so realm topology won't change. 629 */ 630 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm) 631 { 632 struct ceph_inode_info *ci; 633 struct inode *lastinode = NULL; 634 struct ceph_snap_realm *child; 635 636 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino); 637 638 spin_lock(&realm->inodes_with_caps_lock); 639 list_for_each_entry(ci, &realm->inodes_with_caps, 640 i_snap_realm_item) { 641 struct inode *inode = igrab(&ci->vfs_inode); 642 if (!inode) 643 continue; 644 spin_unlock(&realm->inodes_with_caps_lock); 645 iput(lastinode); 646 lastinode = inode; 647 ceph_queue_cap_snap(ci); 648 spin_lock(&realm->inodes_with_caps_lock); 649 } 650 spin_unlock(&realm->inodes_with_caps_lock); 651 iput(lastinode); 652 653 list_for_each_entry(child, &realm->children, child_item) { 654 dout("queue_realm_cap_snaps %p %llx queue child %p %llx\n", 655 realm, realm->ino, child, child->ino); 656 list_del_init(&child->dirty_item); 657 list_add(&child->dirty_item, &realm->dirty_item); 658 } 659 660 list_del_init(&realm->dirty_item); 661 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino); 662 } 663 664 /* 665 * Parse and apply a snapblob "snap trace" from the MDS. This specifies 666 * the snap realm parameters from a given realm and all of its ancestors, 667 * up to the root. 668 * 669 * Caller must hold snap_rwsem for write. 670 */ 671 int ceph_update_snap_trace(struct ceph_mds_client *mdsc, 672 void *p, void *e, bool deletion, 673 struct ceph_snap_realm **realm_ret) 674 { 675 struct ceph_mds_snap_realm *ri; /* encoded */ 676 __le64 *snaps; /* encoded */ 677 __le64 *prior_parent_snaps; /* encoded */ 678 struct ceph_snap_realm *realm = NULL; 679 struct ceph_snap_realm *first_realm = NULL; 680 int invalidate = 0; 681 int err = -ENOMEM; 682 LIST_HEAD(dirty_realms); 683 684 dout("update_snap_trace deletion=%d\n", deletion); 685 more: 686 ceph_decode_need(&p, e, sizeof(*ri), bad); 687 ri = p; 688 p += sizeof(*ri); 689 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) + 690 le32_to_cpu(ri->num_prior_parent_snaps)), bad); 691 snaps = p; 692 p += sizeof(u64) * le32_to_cpu(ri->num_snaps); 693 prior_parent_snaps = p; 694 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps); 695 696 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino)); 697 if (!realm) { 698 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino)); 699 if (IS_ERR(realm)) { 700 err = PTR_ERR(realm); 701 goto fail; 702 } 703 } 704 705 /* ensure the parent is correct */ 706 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent)); 707 if (err < 0) 708 goto fail; 709 invalidate += err; 710 711 if (le64_to_cpu(ri->seq) > realm->seq) { 712 dout("update_snap_trace updating %llx %p %lld -> %lld\n", 713 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq)); 714 /* update realm parameters, snap lists */ 715 realm->seq = le64_to_cpu(ri->seq); 716 realm->created = le64_to_cpu(ri->created); 717 realm->parent_since = le64_to_cpu(ri->parent_since); 718 719 realm->num_snaps = le32_to_cpu(ri->num_snaps); 720 err = dup_array(&realm->snaps, snaps, realm->num_snaps); 721 if (err < 0) 722 goto fail; 723 724 realm->num_prior_parent_snaps = 725 le32_to_cpu(ri->num_prior_parent_snaps); 726 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps, 727 realm->num_prior_parent_snaps); 728 if (err < 0) 729 goto fail; 730 731 /* queue realm for cap_snap creation */ 732 list_add(&realm->dirty_item, &dirty_realms); 733 if (realm->seq > mdsc->last_snap_seq) 734 mdsc->last_snap_seq = realm->seq; 735 736 invalidate = 1; 737 } else if (!realm->cached_context) { 738 dout("update_snap_trace %llx %p seq %lld new\n", 739 realm->ino, realm, realm->seq); 740 invalidate = 1; 741 } else { 742 dout("update_snap_trace %llx %p seq %lld unchanged\n", 743 realm->ino, realm, realm->seq); 744 } 745 746 dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino, 747 realm, invalidate, p, e); 748 749 /* invalidate when we reach the _end_ (root) of the trace */ 750 if (invalidate && p >= e) 751 rebuild_snap_realms(realm); 752 753 if (!first_realm) 754 first_realm = realm; 755 else 756 ceph_put_snap_realm(mdsc, realm); 757 758 if (p < e) 759 goto more; 760 761 /* 762 * queue cap snaps _after_ we've built the new snap contexts, 763 * so that i_head_snapc can be set appropriately. 764 */ 765 while (!list_empty(&dirty_realms)) { 766 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm, 767 dirty_item); 768 queue_realm_cap_snaps(realm); 769 } 770 771 if (realm_ret) 772 *realm_ret = first_realm; 773 else 774 ceph_put_snap_realm(mdsc, first_realm); 775 776 __cleanup_empty_realms(mdsc); 777 return 0; 778 779 bad: 780 err = -EINVAL; 781 fail: 782 if (realm && !IS_ERR(realm)) 783 ceph_put_snap_realm(mdsc, realm); 784 if (first_realm) 785 ceph_put_snap_realm(mdsc, first_realm); 786 pr_err("update_snap_trace error %d\n", err); 787 return err; 788 } 789 790 791 /* 792 * Send any cap_snaps that are queued for flush. Try to carry 793 * s_mutex across multiple snap flushes to avoid locking overhead. 794 * 795 * Caller holds no locks. 796 */ 797 static void flush_snaps(struct ceph_mds_client *mdsc) 798 { 799 struct ceph_inode_info *ci; 800 struct inode *inode; 801 struct ceph_mds_session *session = NULL; 802 803 dout("flush_snaps\n"); 804 spin_lock(&mdsc->snap_flush_lock); 805 while (!list_empty(&mdsc->snap_flush_list)) { 806 ci = list_first_entry(&mdsc->snap_flush_list, 807 struct ceph_inode_info, i_snap_flush_item); 808 inode = &ci->vfs_inode; 809 ihold(inode); 810 spin_unlock(&mdsc->snap_flush_lock); 811 spin_lock(&ci->i_ceph_lock); 812 __ceph_flush_snaps(ci, &session, 0); 813 spin_unlock(&ci->i_ceph_lock); 814 iput(inode); 815 spin_lock(&mdsc->snap_flush_lock); 816 } 817 spin_unlock(&mdsc->snap_flush_lock); 818 819 if (session) { 820 mutex_unlock(&session->s_mutex); 821 ceph_put_mds_session(session); 822 } 823 dout("flush_snaps done\n"); 824 } 825 826 827 /* 828 * Handle a snap notification from the MDS. 829 * 830 * This can take two basic forms: the simplest is just a snap creation 831 * or deletion notification on an existing realm. This should update the 832 * realm and its children. 833 * 834 * The more difficult case is realm creation, due to snap creation at a 835 * new point in the file hierarchy, or due to a rename that moves a file or 836 * directory into another realm. 837 */ 838 void ceph_handle_snap(struct ceph_mds_client *mdsc, 839 struct ceph_mds_session *session, 840 struct ceph_msg *msg) 841 { 842 struct super_block *sb = mdsc->fsc->sb; 843 int mds = session->s_mds; 844 u64 split; 845 int op; 846 int trace_len; 847 struct ceph_snap_realm *realm = NULL; 848 void *p = msg->front.iov_base; 849 void *e = p + msg->front.iov_len; 850 struct ceph_mds_snap_head *h; 851 int num_split_inos, num_split_realms; 852 __le64 *split_inos = NULL, *split_realms = NULL; 853 int i; 854 int locked_rwsem = 0; 855 856 /* decode */ 857 if (msg->front.iov_len < sizeof(*h)) 858 goto bad; 859 h = p; 860 op = le32_to_cpu(h->op); 861 split = le64_to_cpu(h->split); /* non-zero if we are splitting an 862 * existing realm */ 863 num_split_inos = le32_to_cpu(h->num_split_inos); 864 num_split_realms = le32_to_cpu(h->num_split_realms); 865 trace_len = le32_to_cpu(h->trace_len); 866 p += sizeof(*h); 867 868 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds, 869 ceph_snap_op_name(op), split, trace_len); 870 871 mutex_lock(&session->s_mutex); 872 session->s_seq++; 873 mutex_unlock(&session->s_mutex); 874 875 down_write(&mdsc->snap_rwsem); 876 locked_rwsem = 1; 877 878 if (op == CEPH_SNAP_OP_SPLIT) { 879 struct ceph_mds_snap_realm *ri; 880 881 /* 882 * A "split" breaks part of an existing realm off into 883 * a new realm. The MDS provides a list of inodes 884 * (with caps) and child realms that belong to the new 885 * child. 886 */ 887 split_inos = p; 888 p += sizeof(u64) * num_split_inos; 889 split_realms = p; 890 p += sizeof(u64) * num_split_realms; 891 ceph_decode_need(&p, e, sizeof(*ri), bad); 892 /* we will peek at realm info here, but will _not_ 893 * advance p, as the realm update will occur below in 894 * ceph_update_snap_trace. */ 895 ri = p; 896 897 realm = ceph_lookup_snap_realm(mdsc, split); 898 if (!realm) { 899 realm = ceph_create_snap_realm(mdsc, split); 900 if (IS_ERR(realm)) 901 goto out; 902 } 903 904 dout("splitting snap_realm %llx %p\n", realm->ino, realm); 905 for (i = 0; i < num_split_inos; i++) { 906 struct ceph_vino vino = { 907 .ino = le64_to_cpu(split_inos[i]), 908 .snap = CEPH_NOSNAP, 909 }; 910 struct inode *inode = ceph_find_inode(sb, vino); 911 struct ceph_inode_info *ci; 912 struct ceph_snap_realm *oldrealm; 913 914 if (!inode) 915 continue; 916 ci = ceph_inode(inode); 917 918 spin_lock(&ci->i_ceph_lock); 919 if (!ci->i_snap_realm) 920 goto skip_inode; 921 /* 922 * If this inode belongs to a realm that was 923 * created after our new realm, we experienced 924 * a race (due to another split notifications 925 * arriving from a different MDS). So skip 926 * this inode. 927 */ 928 if (ci->i_snap_realm->created > 929 le64_to_cpu(ri->created)) { 930 dout(" leaving %p in newer realm %llx %p\n", 931 inode, ci->i_snap_realm->ino, 932 ci->i_snap_realm); 933 goto skip_inode; 934 } 935 dout(" will move %p to split realm %llx %p\n", 936 inode, realm->ino, realm); 937 /* 938 * Move the inode to the new realm 939 */ 940 spin_lock(&realm->inodes_with_caps_lock); 941 list_del_init(&ci->i_snap_realm_item); 942 list_add(&ci->i_snap_realm_item, 943 &realm->inodes_with_caps); 944 oldrealm = ci->i_snap_realm; 945 ci->i_snap_realm = realm; 946 spin_unlock(&realm->inodes_with_caps_lock); 947 spin_unlock(&ci->i_ceph_lock); 948 949 ceph_get_snap_realm(mdsc, realm); 950 ceph_put_snap_realm(mdsc, oldrealm); 951 952 iput(inode); 953 continue; 954 955 skip_inode: 956 spin_unlock(&ci->i_ceph_lock); 957 iput(inode); 958 } 959 960 /* we may have taken some of the old realm's children. */ 961 for (i = 0; i < num_split_realms; i++) { 962 struct ceph_snap_realm *child = 963 __lookup_snap_realm(mdsc, 964 le64_to_cpu(split_realms[i])); 965 if (!child) 966 continue; 967 adjust_snap_realm_parent(mdsc, child, realm->ino); 968 } 969 } 970 971 /* 972 * update using the provided snap trace. if we are deleting a 973 * snap, we can avoid queueing cap_snaps. 974 */ 975 ceph_update_snap_trace(mdsc, p, e, 976 op == CEPH_SNAP_OP_DESTROY, NULL); 977 978 if (op == CEPH_SNAP_OP_SPLIT) 979 /* we took a reference when we created the realm, above */ 980 ceph_put_snap_realm(mdsc, realm); 981 982 __cleanup_empty_realms(mdsc); 983 984 up_write(&mdsc->snap_rwsem); 985 986 flush_snaps(mdsc); 987 return; 988 989 bad: 990 pr_err("corrupt snap message from mds%d\n", mds); 991 ceph_msg_dump(msg); 992 out: 993 if (locked_rwsem) 994 up_write(&mdsc->snap_rwsem); 995 return; 996 } 997 998 int __init ceph_snap_init(void) 999 { 1000 ceph_empty_snapc = ceph_create_snap_context(0, GFP_NOFS); 1001 if (!ceph_empty_snapc) 1002 return -ENOMEM; 1003 ceph_empty_snapc->seq = 1; 1004 return 0; 1005 } 1006 1007 void ceph_snap_exit(void) 1008 { 1009 ceph_put_snap_context(ceph_empty_snapc); 1010 } 1011