1 #include "ceph_debug.h" 2 3 #include <linux/fs.h> 4 #include <linux/kernel.h> 5 #include <linux/sched.h> 6 #include <linux/vmalloc.h> 7 #include <linux/wait.h> 8 9 #include "super.h" 10 #include "decode.h" 11 #include "messenger.h" 12 13 /* 14 * Capability management 15 * 16 * The Ceph metadata servers control client access to inode metadata 17 * and file data by issuing capabilities, granting clients permission 18 * to read and/or write both inode field and file data to OSDs 19 * (storage nodes). Each capability consists of a set of bits 20 * indicating which operations are allowed. 21 * 22 * If the client holds a *_SHARED cap, the client has a coherent value 23 * that can be safely read from the cached inode. 24 * 25 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the 26 * client is allowed to change inode attributes (e.g., file size, 27 * mtime), note its dirty state in the ceph_cap, and asynchronously 28 * flush that metadata change to the MDS. 29 * 30 * In the event of a conflicting operation (perhaps by another 31 * client), the MDS will revoke the conflicting client capabilities. 32 * 33 * In order for a client to cache an inode, it must hold a capability 34 * with at least one MDS server. When inodes are released, release 35 * notifications are batched and periodically sent en masse to the MDS 36 * cluster to release server state. 37 */ 38 39 40 /* 41 * Generate readable cap strings for debugging output. 42 */ 43 #define MAX_CAP_STR 20 44 static char cap_str[MAX_CAP_STR][40]; 45 static DEFINE_SPINLOCK(cap_str_lock); 46 static int last_cap_str; 47 48 static char *gcap_string(char *s, int c) 49 { 50 if (c & CEPH_CAP_GSHARED) 51 *s++ = 's'; 52 if (c & CEPH_CAP_GEXCL) 53 *s++ = 'x'; 54 if (c & CEPH_CAP_GCACHE) 55 *s++ = 'c'; 56 if (c & CEPH_CAP_GRD) 57 *s++ = 'r'; 58 if (c & CEPH_CAP_GWR) 59 *s++ = 'w'; 60 if (c & CEPH_CAP_GBUFFER) 61 *s++ = 'b'; 62 if (c & CEPH_CAP_GLAZYIO) 63 *s++ = 'l'; 64 return s; 65 } 66 67 const char *ceph_cap_string(int caps) 68 { 69 int i; 70 char *s; 71 int c; 72 73 spin_lock(&cap_str_lock); 74 i = last_cap_str++; 75 if (last_cap_str == MAX_CAP_STR) 76 last_cap_str = 0; 77 spin_unlock(&cap_str_lock); 78 79 s = cap_str[i]; 80 81 if (caps & CEPH_CAP_PIN) 82 *s++ = 'p'; 83 84 c = (caps >> CEPH_CAP_SAUTH) & 3; 85 if (c) { 86 *s++ = 'A'; 87 s = gcap_string(s, c); 88 } 89 90 c = (caps >> CEPH_CAP_SLINK) & 3; 91 if (c) { 92 *s++ = 'L'; 93 s = gcap_string(s, c); 94 } 95 96 c = (caps >> CEPH_CAP_SXATTR) & 3; 97 if (c) { 98 *s++ = 'X'; 99 s = gcap_string(s, c); 100 } 101 102 c = caps >> CEPH_CAP_SFILE; 103 if (c) { 104 *s++ = 'F'; 105 s = gcap_string(s, c); 106 } 107 108 if (s == cap_str[i]) 109 *s++ = '-'; 110 *s = 0; 111 return cap_str[i]; 112 } 113 114 /* 115 * Cap reservations 116 * 117 * Maintain a global pool of preallocated struct ceph_caps, referenced 118 * by struct ceph_caps_reservations. This ensures that we preallocate 119 * memory needed to successfully process an MDS response. (If an MDS 120 * sends us cap information and we fail to process it, we will have 121 * problems due to the client and MDS being out of sync.) 122 * 123 * Reservations are 'owned' by a ceph_cap_reservation context. 124 */ 125 static spinlock_t caps_list_lock; 126 static struct list_head caps_list; /* unused (reserved or unreserved) */ 127 static int caps_total_count; /* total caps allocated */ 128 static int caps_use_count; /* in use */ 129 static int caps_reserve_count; /* unused, reserved */ 130 static int caps_avail_count; /* unused, unreserved */ 131 132 void __init ceph_caps_init(void) 133 { 134 INIT_LIST_HEAD(&caps_list); 135 spin_lock_init(&caps_list_lock); 136 } 137 138 void ceph_caps_finalize(void) 139 { 140 struct ceph_cap *cap; 141 142 spin_lock(&caps_list_lock); 143 while (!list_empty(&caps_list)) { 144 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item); 145 list_del(&cap->caps_item); 146 kmem_cache_free(ceph_cap_cachep, cap); 147 } 148 caps_total_count = 0; 149 caps_avail_count = 0; 150 caps_use_count = 0; 151 caps_reserve_count = 0; 152 spin_unlock(&caps_list_lock); 153 } 154 155 int ceph_reserve_caps(struct ceph_cap_reservation *ctx, int need) 156 { 157 int i; 158 struct ceph_cap *cap; 159 int have; 160 int alloc = 0; 161 LIST_HEAD(newcaps); 162 int ret = 0; 163 164 dout("reserve caps ctx=%p need=%d\n", ctx, need); 165 166 /* first reserve any caps that are already allocated */ 167 spin_lock(&caps_list_lock); 168 if (caps_avail_count >= need) 169 have = need; 170 else 171 have = caps_avail_count; 172 caps_avail_count -= have; 173 caps_reserve_count += have; 174 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 175 caps_avail_count); 176 spin_unlock(&caps_list_lock); 177 178 for (i = have; i < need; i++) { 179 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 180 if (!cap) { 181 ret = -ENOMEM; 182 goto out_alloc_count; 183 } 184 list_add(&cap->caps_item, &newcaps); 185 alloc++; 186 } 187 BUG_ON(have + alloc != need); 188 189 spin_lock(&caps_list_lock); 190 caps_total_count += alloc; 191 caps_reserve_count += alloc; 192 list_splice(&newcaps, &caps_list); 193 194 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 195 caps_avail_count); 196 spin_unlock(&caps_list_lock); 197 198 ctx->count = need; 199 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n", 200 ctx, caps_total_count, caps_use_count, caps_reserve_count, 201 caps_avail_count); 202 return 0; 203 204 out_alloc_count: 205 /* we didn't manage to reserve as much as we needed */ 206 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n", 207 ctx, need, have); 208 return ret; 209 } 210 211 int ceph_unreserve_caps(struct ceph_cap_reservation *ctx) 212 { 213 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count); 214 if (ctx->count) { 215 spin_lock(&caps_list_lock); 216 BUG_ON(caps_reserve_count < ctx->count); 217 caps_reserve_count -= ctx->count; 218 caps_avail_count += ctx->count; 219 ctx->count = 0; 220 dout("unreserve caps %d = %d used + %d resv + %d avail\n", 221 caps_total_count, caps_use_count, caps_reserve_count, 222 caps_avail_count); 223 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 224 caps_avail_count); 225 spin_unlock(&caps_list_lock); 226 } 227 return 0; 228 } 229 230 static struct ceph_cap *get_cap(struct ceph_cap_reservation *ctx) 231 { 232 struct ceph_cap *cap = NULL; 233 234 /* temporary, until we do something about cap import/export */ 235 if (!ctx) 236 return kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 237 238 spin_lock(&caps_list_lock); 239 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n", 240 ctx, ctx->count, caps_total_count, caps_use_count, 241 caps_reserve_count, caps_avail_count); 242 BUG_ON(!ctx->count); 243 BUG_ON(ctx->count > caps_reserve_count); 244 BUG_ON(list_empty(&caps_list)); 245 246 ctx->count--; 247 caps_reserve_count--; 248 caps_use_count++; 249 250 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item); 251 list_del(&cap->caps_item); 252 253 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 254 caps_avail_count); 255 spin_unlock(&caps_list_lock); 256 return cap; 257 } 258 259 static void put_cap(struct ceph_cap *cap, 260 struct ceph_cap_reservation *ctx) 261 { 262 spin_lock(&caps_list_lock); 263 dout("put_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n", 264 ctx, ctx ? ctx->count : 0, caps_total_count, caps_use_count, 265 caps_reserve_count, caps_avail_count); 266 caps_use_count--; 267 /* 268 * Keep some preallocated caps around, at least enough to do a 269 * readdir (which needs to preallocate lots of them), to avoid 270 * lots of free/alloc churn. 271 */ 272 if (caps_avail_count >= caps_reserve_count + 273 ceph_client(cap->ci->vfs_inode.i_sb)->mount_args.max_readdir) { 274 caps_total_count--; 275 kmem_cache_free(ceph_cap_cachep, cap); 276 } else { 277 if (ctx) { 278 ctx->count++; 279 caps_reserve_count++; 280 } else { 281 caps_avail_count++; 282 } 283 list_add(&cap->caps_item, &caps_list); 284 } 285 286 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 287 caps_avail_count); 288 spin_unlock(&caps_list_lock); 289 } 290 291 void ceph_reservation_status(struct ceph_client *client, 292 int *total, int *avail, int *used, int *reserved) 293 { 294 if (total) 295 *total = caps_total_count; 296 if (avail) 297 *avail = caps_avail_count; 298 if (used) 299 *used = caps_use_count; 300 if (reserved) 301 *reserved = caps_reserve_count; 302 } 303 304 /* 305 * Find ceph_cap for given mds, if any. 306 * 307 * Called with i_lock held. 308 */ 309 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds) 310 { 311 struct ceph_cap *cap; 312 struct rb_node *n = ci->i_caps.rb_node; 313 314 while (n) { 315 cap = rb_entry(n, struct ceph_cap, ci_node); 316 if (mds < cap->mds) 317 n = n->rb_left; 318 else if (mds > cap->mds) 319 n = n->rb_right; 320 else 321 return cap; 322 } 323 return NULL; 324 } 325 326 /* 327 * Return id of any MDS with a cap, preferably FILE_WR|WRBUFFER|EXCL, else 328 * -1. 329 */ 330 static int __ceph_get_cap_mds(struct ceph_inode_info *ci, u32 *mseq) 331 { 332 struct ceph_cap *cap; 333 int mds = -1; 334 struct rb_node *p; 335 336 /* prefer mds with WR|WRBUFFER|EXCL caps */ 337 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 338 cap = rb_entry(p, struct ceph_cap, ci_node); 339 mds = cap->mds; 340 if (mseq) 341 *mseq = cap->mseq; 342 if (cap->issued & (CEPH_CAP_FILE_WR | 343 CEPH_CAP_FILE_BUFFER | 344 CEPH_CAP_FILE_EXCL)) 345 break; 346 } 347 return mds; 348 } 349 350 int ceph_get_cap_mds(struct inode *inode) 351 { 352 int mds; 353 spin_lock(&inode->i_lock); 354 mds = __ceph_get_cap_mds(ceph_inode(inode), NULL); 355 spin_unlock(&inode->i_lock); 356 return mds; 357 } 358 359 /* 360 * Called under i_lock. 361 */ 362 static void __insert_cap_node(struct ceph_inode_info *ci, 363 struct ceph_cap *new) 364 { 365 struct rb_node **p = &ci->i_caps.rb_node; 366 struct rb_node *parent = NULL; 367 struct ceph_cap *cap = NULL; 368 369 while (*p) { 370 parent = *p; 371 cap = rb_entry(parent, struct ceph_cap, ci_node); 372 if (new->mds < cap->mds) 373 p = &(*p)->rb_left; 374 else if (new->mds > cap->mds) 375 p = &(*p)->rb_right; 376 else 377 BUG(); 378 } 379 380 rb_link_node(&new->ci_node, parent, p); 381 rb_insert_color(&new->ci_node, &ci->i_caps); 382 } 383 384 /* 385 * (re)set cap hold timeouts, which control the delayed release 386 * of unused caps back to the MDS. Should be called on cap use. 387 */ 388 static void __cap_set_timeouts(struct ceph_mds_client *mdsc, 389 struct ceph_inode_info *ci) 390 { 391 struct ceph_mount_args *ma = &mdsc->client->mount_args; 392 393 ci->i_hold_caps_min = round_jiffies(jiffies + 394 ma->caps_wanted_delay_min * HZ); 395 ci->i_hold_caps_max = round_jiffies(jiffies + 396 ma->caps_wanted_delay_max * HZ); 397 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode, 398 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies); 399 } 400 401 /* 402 * (Re)queue cap at the end of the delayed cap release list. 403 * 404 * If I_FLUSH is set, leave the inode at the front of the list. 405 * 406 * Caller holds i_lock 407 * -> we take mdsc->cap_delay_lock 408 */ 409 static void __cap_delay_requeue(struct ceph_mds_client *mdsc, 410 struct ceph_inode_info *ci) 411 { 412 __cap_set_timeouts(mdsc, ci); 413 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode, 414 ci->i_ceph_flags, ci->i_hold_caps_max); 415 if (!mdsc->stopping) { 416 spin_lock(&mdsc->cap_delay_lock); 417 if (!list_empty(&ci->i_cap_delay_list)) { 418 if (ci->i_ceph_flags & CEPH_I_FLUSH) 419 goto no_change; 420 list_del_init(&ci->i_cap_delay_list); 421 } 422 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 423 no_change: 424 spin_unlock(&mdsc->cap_delay_lock); 425 } 426 } 427 428 /* 429 * Queue an inode for immediate writeback. Mark inode with I_FLUSH, 430 * indicating we should send a cap message to flush dirty metadata 431 * asap, and move to the front of the delayed cap list. 432 */ 433 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc, 434 struct ceph_inode_info *ci) 435 { 436 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode); 437 spin_lock(&mdsc->cap_delay_lock); 438 ci->i_ceph_flags |= CEPH_I_FLUSH; 439 if (!list_empty(&ci->i_cap_delay_list)) 440 list_del_init(&ci->i_cap_delay_list); 441 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 442 spin_unlock(&mdsc->cap_delay_lock); 443 } 444 445 /* 446 * Cancel delayed work on cap. 447 * 448 * Caller must hold i_lock. 449 */ 450 static void __cap_delay_cancel(struct ceph_mds_client *mdsc, 451 struct ceph_inode_info *ci) 452 { 453 dout("__cap_delay_cancel %p\n", &ci->vfs_inode); 454 if (list_empty(&ci->i_cap_delay_list)) 455 return; 456 spin_lock(&mdsc->cap_delay_lock); 457 list_del_init(&ci->i_cap_delay_list); 458 spin_unlock(&mdsc->cap_delay_lock); 459 } 460 461 /* 462 * Common issue checks for add_cap, handle_cap_grant. 463 */ 464 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap, 465 unsigned issued) 466 { 467 unsigned had = __ceph_caps_issued(ci, NULL); 468 469 /* 470 * Each time we receive FILE_CACHE anew, we increment 471 * i_rdcache_gen. 472 */ 473 if ((issued & CEPH_CAP_FILE_CACHE) && 474 (had & CEPH_CAP_FILE_CACHE) == 0) 475 ci->i_rdcache_gen++; 476 477 /* 478 * if we are newly issued FILE_SHARED, clear I_COMPLETE; we 479 * don't know what happened to this directory while we didn't 480 * have the cap. 481 */ 482 if ((issued & CEPH_CAP_FILE_SHARED) && 483 (had & CEPH_CAP_FILE_SHARED) == 0) { 484 ci->i_shared_gen++; 485 if (S_ISDIR(ci->vfs_inode.i_mode)) { 486 dout(" marking %p NOT complete\n", &ci->vfs_inode); 487 ci->i_ceph_flags &= ~CEPH_I_COMPLETE; 488 } 489 } 490 } 491 492 /* 493 * Add a capability under the given MDS session. 494 * 495 * Caller should hold session snap_rwsem (read) and s_mutex. 496 * 497 * @fmode is the open file mode, if we are opening a file, otherwise 498 * it is < 0. (This is so we can atomically add the cap and add an 499 * open file reference to it.) 500 */ 501 int ceph_add_cap(struct inode *inode, 502 struct ceph_mds_session *session, u64 cap_id, 503 int fmode, unsigned issued, unsigned wanted, 504 unsigned seq, unsigned mseq, u64 realmino, int flags, 505 struct ceph_cap_reservation *caps_reservation) 506 { 507 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 508 struct ceph_inode_info *ci = ceph_inode(inode); 509 struct ceph_cap *new_cap = NULL; 510 struct ceph_cap *cap; 511 int mds = session->s_mds; 512 int actual_wanted; 513 514 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode, 515 session->s_mds, cap_id, ceph_cap_string(issued), seq); 516 517 /* 518 * If we are opening the file, include file mode wanted bits 519 * in wanted. 520 */ 521 if (fmode >= 0) 522 wanted |= ceph_caps_for_mode(fmode); 523 524 retry: 525 spin_lock(&inode->i_lock); 526 cap = __get_cap_for_mds(ci, mds); 527 if (!cap) { 528 if (new_cap) { 529 cap = new_cap; 530 new_cap = NULL; 531 } else { 532 spin_unlock(&inode->i_lock); 533 new_cap = get_cap(caps_reservation); 534 if (new_cap == NULL) 535 return -ENOMEM; 536 goto retry; 537 } 538 539 cap->issued = 0; 540 cap->implemented = 0; 541 cap->mds = mds; 542 cap->mds_wanted = 0; 543 544 cap->ci = ci; 545 __insert_cap_node(ci, cap); 546 547 /* clear out old exporting info? (i.e. on cap import) */ 548 if (ci->i_cap_exporting_mds == mds) { 549 ci->i_cap_exporting_issued = 0; 550 ci->i_cap_exporting_mseq = 0; 551 ci->i_cap_exporting_mds = -1; 552 } 553 554 /* add to session cap list */ 555 cap->session = session; 556 spin_lock(&session->s_cap_lock); 557 list_add_tail(&cap->session_caps, &session->s_caps); 558 session->s_nr_caps++; 559 spin_unlock(&session->s_cap_lock); 560 } 561 562 if (!ci->i_snap_realm) { 563 /* 564 * add this inode to the appropriate snap realm 565 */ 566 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc, 567 realmino); 568 if (realm) { 569 ceph_get_snap_realm(mdsc, realm); 570 spin_lock(&realm->inodes_with_caps_lock); 571 ci->i_snap_realm = realm; 572 list_add(&ci->i_snap_realm_item, 573 &realm->inodes_with_caps); 574 spin_unlock(&realm->inodes_with_caps_lock); 575 } else { 576 pr_err("ceph_add_cap: couldn't find snap realm %llx\n", 577 realmino); 578 } 579 } 580 581 __check_cap_issue(ci, cap, issued); 582 583 /* 584 * If we are issued caps we don't want, or the mds' wanted 585 * value appears to be off, queue a check so we'll release 586 * later and/or update the mds wanted value. 587 */ 588 actual_wanted = __ceph_caps_wanted(ci); 589 if ((wanted & ~actual_wanted) || 590 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) { 591 dout(" issued %s, mds wanted %s, actual %s, queueing\n", 592 ceph_cap_string(issued), ceph_cap_string(wanted), 593 ceph_cap_string(actual_wanted)); 594 __cap_delay_requeue(mdsc, ci); 595 } 596 597 if (flags & CEPH_CAP_FLAG_AUTH) 598 ci->i_auth_cap = cap; 599 else if (ci->i_auth_cap == cap) 600 ci->i_auth_cap = NULL; 601 602 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n", 603 inode, ceph_vinop(inode), cap, ceph_cap_string(issued), 604 ceph_cap_string(issued|cap->issued), seq, mds); 605 cap->cap_id = cap_id; 606 cap->issued = issued; 607 cap->implemented |= issued; 608 cap->mds_wanted |= wanted; 609 cap->seq = seq; 610 cap->issue_seq = seq; 611 cap->mseq = mseq; 612 cap->gen = session->s_cap_gen; 613 614 if (fmode >= 0) 615 __ceph_get_fmode(ci, fmode); 616 spin_unlock(&inode->i_lock); 617 wake_up(&ci->i_cap_wq); 618 return 0; 619 } 620 621 /* 622 * Return true if cap has not timed out and belongs to the current 623 * generation of the MDS session (i.e. has not gone 'stale' due to 624 * us losing touch with the mds). 625 */ 626 static int __cap_is_valid(struct ceph_cap *cap) 627 { 628 unsigned long ttl; 629 u32 gen; 630 631 spin_lock(&cap->session->s_cap_lock); 632 gen = cap->session->s_cap_gen; 633 ttl = cap->session->s_cap_ttl; 634 spin_unlock(&cap->session->s_cap_lock); 635 636 if (cap->gen < gen || time_after_eq(jiffies, ttl)) { 637 dout("__cap_is_valid %p cap %p issued %s " 638 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode, 639 cap, ceph_cap_string(cap->issued), cap->gen, gen); 640 return 0; 641 } 642 643 return 1; 644 } 645 646 /* 647 * Return set of valid cap bits issued to us. Note that caps time 648 * out, and may be invalidated in bulk if the client session times out 649 * and session->s_cap_gen is bumped. 650 */ 651 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented) 652 { 653 int have = ci->i_snap_caps; 654 struct ceph_cap *cap; 655 struct rb_node *p; 656 657 if (implemented) 658 *implemented = 0; 659 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 660 cap = rb_entry(p, struct ceph_cap, ci_node); 661 if (!__cap_is_valid(cap)) 662 continue; 663 dout("__ceph_caps_issued %p cap %p issued %s\n", 664 &ci->vfs_inode, cap, ceph_cap_string(cap->issued)); 665 have |= cap->issued; 666 if (implemented) 667 *implemented |= cap->implemented; 668 } 669 return have; 670 } 671 672 /* 673 * Get cap bits issued by caps other than @ocap 674 */ 675 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap) 676 { 677 int have = ci->i_snap_caps; 678 struct ceph_cap *cap; 679 struct rb_node *p; 680 681 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 682 cap = rb_entry(p, struct ceph_cap, ci_node); 683 if (cap == ocap) 684 continue; 685 if (!__cap_is_valid(cap)) 686 continue; 687 have |= cap->issued; 688 } 689 return have; 690 } 691 692 /* 693 * Move a cap to the end of the LRU (oldest caps at list head, newest 694 * at list tail). 695 */ 696 static void __touch_cap(struct ceph_cap *cap) 697 { 698 struct ceph_mds_session *s = cap->session; 699 700 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap, 701 s->s_mds); 702 spin_lock(&s->s_cap_lock); 703 list_move_tail(&cap->session_caps, &s->s_caps); 704 spin_unlock(&s->s_cap_lock); 705 } 706 707 /* 708 * Check if we hold the given mask. If so, move the cap(s) to the 709 * front of their respective LRUs. (This is the preferred way for 710 * callers to check for caps they want.) 711 */ 712 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch) 713 { 714 struct ceph_cap *cap; 715 struct rb_node *p; 716 int have = ci->i_snap_caps; 717 718 if ((have & mask) == mask) { 719 dout("__ceph_caps_issued_mask %p snap issued %s" 720 " (mask %s)\n", &ci->vfs_inode, 721 ceph_cap_string(have), 722 ceph_cap_string(mask)); 723 return 1; 724 } 725 726 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 727 cap = rb_entry(p, struct ceph_cap, ci_node); 728 if (!__cap_is_valid(cap)) 729 continue; 730 if ((cap->issued & mask) == mask) { 731 dout("__ceph_caps_issued_mask %p cap %p issued %s" 732 " (mask %s)\n", &ci->vfs_inode, cap, 733 ceph_cap_string(cap->issued), 734 ceph_cap_string(mask)); 735 if (touch) 736 __touch_cap(cap); 737 return 1; 738 } 739 740 /* does a combination of caps satisfy mask? */ 741 have |= cap->issued; 742 if ((have & mask) == mask) { 743 dout("__ceph_caps_issued_mask %p combo issued %s" 744 " (mask %s)\n", &ci->vfs_inode, 745 ceph_cap_string(cap->issued), 746 ceph_cap_string(mask)); 747 if (touch) { 748 struct rb_node *q; 749 750 /* touch this + preceeding caps */ 751 __touch_cap(cap); 752 for (q = rb_first(&ci->i_caps); q != p; 753 q = rb_next(q)) { 754 cap = rb_entry(q, struct ceph_cap, 755 ci_node); 756 if (!__cap_is_valid(cap)) 757 continue; 758 __touch_cap(cap); 759 } 760 } 761 return 1; 762 } 763 } 764 765 return 0; 766 } 767 768 /* 769 * Return true if mask caps are currently being revoked by an MDS. 770 */ 771 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask) 772 { 773 struct inode *inode = &ci->vfs_inode; 774 struct ceph_cap *cap; 775 struct rb_node *p; 776 int ret = 0; 777 778 spin_lock(&inode->i_lock); 779 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 780 cap = rb_entry(p, struct ceph_cap, ci_node); 781 if (__cap_is_valid(cap) && 782 (cap->implemented & ~cap->issued & mask)) { 783 ret = 1; 784 break; 785 } 786 } 787 spin_unlock(&inode->i_lock); 788 dout("ceph_caps_revoking %p %s = %d\n", inode, 789 ceph_cap_string(mask), ret); 790 return ret; 791 } 792 793 int __ceph_caps_used(struct ceph_inode_info *ci) 794 { 795 int used = 0; 796 if (ci->i_pin_ref) 797 used |= CEPH_CAP_PIN; 798 if (ci->i_rd_ref) 799 used |= CEPH_CAP_FILE_RD; 800 if (ci->i_rdcache_ref || ci->i_rdcache_gen) 801 used |= CEPH_CAP_FILE_CACHE; 802 if (ci->i_wr_ref) 803 used |= CEPH_CAP_FILE_WR; 804 if (ci->i_wrbuffer_ref) 805 used |= CEPH_CAP_FILE_BUFFER; 806 return used; 807 } 808 809 /* 810 * wanted, by virtue of open file modes 811 */ 812 int __ceph_caps_file_wanted(struct ceph_inode_info *ci) 813 { 814 int want = 0; 815 int mode; 816 for (mode = 0; mode < 4; mode++) 817 if (ci->i_nr_by_mode[mode]) 818 want |= ceph_caps_for_mode(mode); 819 return want; 820 } 821 822 /* 823 * Return caps we have registered with the MDS(s) as 'wanted'. 824 */ 825 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci) 826 { 827 struct ceph_cap *cap; 828 struct rb_node *p; 829 int mds_wanted = 0; 830 831 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 832 cap = rb_entry(p, struct ceph_cap, ci_node); 833 if (!__cap_is_valid(cap)) 834 continue; 835 mds_wanted |= cap->mds_wanted; 836 } 837 return mds_wanted; 838 } 839 840 /* 841 * called under i_lock 842 */ 843 static int __ceph_is_any_caps(struct ceph_inode_info *ci) 844 { 845 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0; 846 } 847 848 /* 849 * caller should hold i_lock, and session s_mutex. 850 * returns true if this is the last cap. if so, caller should iput. 851 */ 852 void __ceph_remove_cap(struct ceph_cap *cap, 853 struct ceph_cap_reservation *ctx) 854 { 855 struct ceph_mds_session *session = cap->session; 856 struct ceph_inode_info *ci = cap->ci; 857 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc; 858 859 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode); 860 861 /* remove from session list */ 862 spin_lock(&session->s_cap_lock); 863 list_del_init(&cap->session_caps); 864 session->s_nr_caps--; 865 spin_unlock(&session->s_cap_lock); 866 867 /* remove from inode list */ 868 rb_erase(&cap->ci_node, &ci->i_caps); 869 cap->session = NULL; 870 if (ci->i_auth_cap == cap) 871 ci->i_auth_cap = NULL; 872 873 put_cap(cap, ctx); 874 875 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) { 876 struct ceph_snap_realm *realm = ci->i_snap_realm; 877 spin_lock(&realm->inodes_with_caps_lock); 878 list_del_init(&ci->i_snap_realm_item); 879 ci->i_snap_realm_counter++; 880 ci->i_snap_realm = NULL; 881 spin_unlock(&realm->inodes_with_caps_lock); 882 ceph_put_snap_realm(mdsc, realm); 883 } 884 if (!__ceph_is_any_real_caps(ci)) 885 __cap_delay_cancel(mdsc, ci); 886 } 887 888 /* 889 * Build and send a cap message to the given MDS. 890 * 891 * Caller should be holding s_mutex. 892 */ 893 static int send_cap_msg(struct ceph_mds_session *session, 894 u64 ino, u64 cid, int op, 895 int caps, int wanted, int dirty, 896 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq, 897 u64 size, u64 max_size, 898 struct timespec *mtime, struct timespec *atime, 899 u64 time_warp_seq, 900 uid_t uid, gid_t gid, mode_t mode, 901 u64 xattr_version, 902 struct ceph_buffer *xattrs_buf, 903 u64 follows) 904 { 905 struct ceph_mds_caps *fc; 906 struct ceph_msg *msg; 907 908 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s" 909 " seq %u/%u mseq %u follows %lld size %llu/%llu" 910 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op), 911 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted), 912 ceph_cap_string(dirty), 913 seq, issue_seq, mseq, follows, size, max_size, 914 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0); 915 916 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL); 917 if (IS_ERR(msg)) 918 return PTR_ERR(msg); 919 920 fc = msg->front.iov_base; 921 922 memset(fc, 0, sizeof(*fc)); 923 924 fc->cap_id = cpu_to_le64(cid); 925 fc->op = cpu_to_le32(op); 926 fc->seq = cpu_to_le32(seq); 927 fc->client_tid = cpu_to_le64(flush_tid); 928 fc->issue_seq = cpu_to_le32(issue_seq); 929 fc->migrate_seq = cpu_to_le32(mseq); 930 fc->caps = cpu_to_le32(caps); 931 fc->wanted = cpu_to_le32(wanted); 932 fc->dirty = cpu_to_le32(dirty); 933 fc->ino = cpu_to_le64(ino); 934 fc->snap_follows = cpu_to_le64(follows); 935 936 fc->size = cpu_to_le64(size); 937 fc->max_size = cpu_to_le64(max_size); 938 if (mtime) 939 ceph_encode_timespec(&fc->mtime, mtime); 940 if (atime) 941 ceph_encode_timespec(&fc->atime, atime); 942 fc->time_warp_seq = cpu_to_le32(time_warp_seq); 943 944 fc->uid = cpu_to_le32(uid); 945 fc->gid = cpu_to_le32(gid); 946 fc->mode = cpu_to_le32(mode); 947 948 fc->xattr_version = cpu_to_le64(xattr_version); 949 if (xattrs_buf) { 950 msg->middle = ceph_buffer_get(xattrs_buf); 951 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len); 952 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len); 953 } 954 955 ceph_con_send(&session->s_con, msg); 956 return 0; 957 } 958 959 /* 960 * Queue cap releases when an inode is dropped from our 961 * cache. 962 */ 963 void ceph_queue_caps_release(struct inode *inode) 964 { 965 struct ceph_inode_info *ci = ceph_inode(inode); 966 struct rb_node *p; 967 968 spin_lock(&inode->i_lock); 969 p = rb_first(&ci->i_caps); 970 while (p) { 971 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node); 972 struct ceph_mds_session *session = cap->session; 973 struct ceph_msg *msg; 974 struct ceph_mds_cap_release *head; 975 struct ceph_mds_cap_item *item; 976 977 spin_lock(&session->s_cap_lock); 978 BUG_ON(!session->s_num_cap_releases); 979 msg = list_first_entry(&session->s_cap_releases, 980 struct ceph_msg, list_head); 981 982 dout(" adding %p release to mds%d msg %p (%d left)\n", 983 inode, session->s_mds, msg, session->s_num_cap_releases); 984 985 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE); 986 head = msg->front.iov_base; 987 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1); 988 item = msg->front.iov_base + msg->front.iov_len; 989 item->ino = cpu_to_le64(ceph_ino(inode)); 990 item->cap_id = cpu_to_le64(cap->cap_id); 991 item->migrate_seq = cpu_to_le32(cap->mseq); 992 item->seq = cpu_to_le32(cap->issue_seq); 993 994 session->s_num_cap_releases--; 995 996 msg->front.iov_len += sizeof(*item); 997 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) { 998 dout(" release msg %p full\n", msg); 999 list_move_tail(&msg->list_head, 1000 &session->s_cap_releases_done); 1001 } else { 1002 dout(" release msg %p at %d/%d (%d)\n", msg, 1003 (int)le32_to_cpu(head->num), 1004 (int)CEPH_CAPS_PER_RELEASE, 1005 (int)msg->front.iov_len); 1006 } 1007 spin_unlock(&session->s_cap_lock); 1008 p = rb_next(p); 1009 __ceph_remove_cap(cap, NULL); 1010 1011 } 1012 spin_unlock(&inode->i_lock); 1013 } 1014 1015 /* 1016 * Send a cap msg on the given inode. Update our caps state, then 1017 * drop i_lock and send the message. 1018 * 1019 * Make note of max_size reported/requested from mds, revoked caps 1020 * that have now been implemented. 1021 * 1022 * Make half-hearted attempt ot to invalidate page cache if we are 1023 * dropping RDCACHE. Note that this will leave behind locked pages 1024 * that we'll then need to deal with elsewhere. 1025 * 1026 * Return non-zero if delayed release, or we experienced an error 1027 * such that the caller should requeue + retry later. 1028 * 1029 * called with i_lock, then drops it. 1030 * caller should hold snap_rwsem (read), s_mutex. 1031 */ 1032 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap, 1033 int op, int used, int want, int retain, int flushing, 1034 unsigned *pflush_tid) 1035 __releases(cap->ci->vfs_inode->i_lock) 1036 { 1037 struct ceph_inode_info *ci = cap->ci; 1038 struct inode *inode = &ci->vfs_inode; 1039 u64 cap_id = cap->cap_id; 1040 int held = cap->issued | cap->implemented; 1041 int revoking = cap->implemented & ~cap->issued; 1042 int dropping = cap->issued & ~retain; 1043 int keep; 1044 u64 seq, issue_seq, mseq, time_warp_seq, follows; 1045 u64 size, max_size; 1046 struct timespec mtime, atime; 1047 int wake = 0; 1048 mode_t mode; 1049 uid_t uid; 1050 gid_t gid; 1051 struct ceph_mds_session *session; 1052 u64 xattr_version = 0; 1053 int delayed = 0; 1054 u64 flush_tid = 0; 1055 int i; 1056 int ret; 1057 1058 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n", 1059 inode, cap, cap->session, 1060 ceph_cap_string(held), ceph_cap_string(held & retain), 1061 ceph_cap_string(revoking)); 1062 BUG_ON((retain & CEPH_CAP_PIN) == 0); 1063 1064 session = cap->session; 1065 1066 /* don't release wanted unless we've waited a bit. */ 1067 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1068 time_before(jiffies, ci->i_hold_caps_min)) { 1069 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n", 1070 ceph_cap_string(cap->issued), 1071 ceph_cap_string(cap->issued & retain), 1072 ceph_cap_string(cap->mds_wanted), 1073 ceph_cap_string(want)); 1074 want |= cap->mds_wanted; 1075 retain |= cap->issued; 1076 delayed = 1; 1077 } 1078 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH); 1079 1080 cap->issued &= retain; /* drop bits we don't want */ 1081 if (cap->implemented & ~cap->issued) { 1082 /* 1083 * Wake up any waiters on wanted -> needed transition. 1084 * This is due to the weird transition from buffered 1085 * to sync IO... we need to flush dirty pages _before_ 1086 * allowing sync writes to avoid reordering. 1087 */ 1088 wake = 1; 1089 } 1090 cap->implemented &= cap->issued | used; 1091 cap->mds_wanted = want; 1092 1093 if (flushing) { 1094 /* 1095 * assign a tid for flush operations so we can avoid 1096 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark 1097 * clean type races. track latest tid for every bit 1098 * so we can handle flush AxFw, flush Fw, and have the 1099 * first ack clean Ax. 1100 */ 1101 flush_tid = ++ci->i_cap_flush_last_tid; 1102 if (pflush_tid) 1103 *pflush_tid = flush_tid; 1104 dout(" cap_flush_tid %d\n", (int)flush_tid); 1105 for (i = 0; i < CEPH_CAP_BITS; i++) 1106 if (flushing & (1 << i)) 1107 ci->i_cap_flush_tid[i] = flush_tid; 1108 } 1109 1110 keep = cap->implemented; 1111 seq = cap->seq; 1112 issue_seq = cap->issue_seq; 1113 mseq = cap->mseq; 1114 size = inode->i_size; 1115 ci->i_reported_size = size; 1116 max_size = ci->i_wanted_max_size; 1117 ci->i_requested_max_size = max_size; 1118 mtime = inode->i_mtime; 1119 atime = inode->i_atime; 1120 time_warp_seq = ci->i_time_warp_seq; 1121 follows = ci->i_snap_realm->cached_context->seq; 1122 uid = inode->i_uid; 1123 gid = inode->i_gid; 1124 mode = inode->i_mode; 1125 1126 if (dropping & CEPH_CAP_XATTR_EXCL) { 1127 __ceph_build_xattrs_blob(ci); 1128 xattr_version = ci->i_xattrs.version + 1; 1129 } 1130 1131 spin_unlock(&inode->i_lock); 1132 1133 if (dropping & CEPH_CAP_FILE_CACHE) { 1134 /* invalidate what we can */ 1135 dout("invalidating pages on %p\n", inode); 1136 invalidate_mapping_pages(&inode->i_data, 0, -1); 1137 } 1138 1139 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id, 1140 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq, 1141 size, max_size, &mtime, &atime, time_warp_seq, 1142 uid, gid, mode, 1143 xattr_version, 1144 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL, 1145 follows); 1146 if (ret < 0) { 1147 dout("error sending cap msg, must requeue %p\n", inode); 1148 delayed = 1; 1149 } 1150 1151 if (wake) 1152 wake_up(&ci->i_cap_wq); 1153 1154 return delayed; 1155 } 1156 1157 /* 1158 * When a snapshot is taken, clients accumulate dirty metadata on 1159 * inodes with capabilities in ceph_cap_snaps to describe the file 1160 * state at the time the snapshot was taken. This must be flushed 1161 * asynchronously back to the MDS once sync writes complete and dirty 1162 * data is written out. 1163 * 1164 * Called under i_lock. Takes s_mutex as needed. 1165 */ 1166 void __ceph_flush_snaps(struct ceph_inode_info *ci, 1167 struct ceph_mds_session **psession) 1168 { 1169 struct inode *inode = &ci->vfs_inode; 1170 int mds; 1171 struct ceph_cap_snap *capsnap; 1172 u32 mseq; 1173 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 1174 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold 1175 session->s_mutex */ 1176 u64 next_follows = 0; /* keep track of how far we've gotten through the 1177 i_cap_snaps list, and skip these entries next time 1178 around to avoid an infinite loop */ 1179 1180 if (psession) 1181 session = *psession; 1182 1183 dout("__flush_snaps %p\n", inode); 1184 retry: 1185 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 1186 /* avoid an infiniute loop after retry */ 1187 if (capsnap->follows < next_follows) 1188 continue; 1189 /* 1190 * we need to wait for sync writes to complete and for dirty 1191 * pages to be written out. 1192 */ 1193 if (capsnap->dirty_pages || capsnap->writing) 1194 continue; 1195 1196 /* pick mds, take s_mutex */ 1197 mds = __ceph_get_cap_mds(ci, &mseq); 1198 if (session && session->s_mds != mds) { 1199 dout("oops, wrong session %p mutex\n", session); 1200 mutex_unlock(&session->s_mutex); 1201 ceph_put_mds_session(session); 1202 session = NULL; 1203 } 1204 if (!session) { 1205 spin_unlock(&inode->i_lock); 1206 mutex_lock(&mdsc->mutex); 1207 session = __ceph_lookup_mds_session(mdsc, mds); 1208 mutex_unlock(&mdsc->mutex); 1209 if (session) { 1210 dout("inverting session/ino locks on %p\n", 1211 session); 1212 mutex_lock(&session->s_mutex); 1213 } 1214 /* 1215 * if session == NULL, we raced against a cap 1216 * deletion. retry, and we'll get a better 1217 * @mds value next time. 1218 */ 1219 spin_lock(&inode->i_lock); 1220 goto retry; 1221 } 1222 1223 capsnap->flush_tid = ++ci->i_cap_flush_last_tid; 1224 atomic_inc(&capsnap->nref); 1225 if (!list_empty(&capsnap->flushing_item)) 1226 list_del_init(&capsnap->flushing_item); 1227 list_add_tail(&capsnap->flushing_item, 1228 &session->s_cap_snaps_flushing); 1229 spin_unlock(&inode->i_lock); 1230 1231 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n", 1232 inode, capsnap, next_follows, capsnap->size); 1233 send_cap_msg(session, ceph_vino(inode).ino, 0, 1234 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0, 1235 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq, 1236 capsnap->size, 0, 1237 &capsnap->mtime, &capsnap->atime, 1238 capsnap->time_warp_seq, 1239 capsnap->uid, capsnap->gid, capsnap->mode, 1240 0, NULL, 1241 capsnap->follows); 1242 1243 next_follows = capsnap->follows + 1; 1244 ceph_put_cap_snap(capsnap); 1245 1246 spin_lock(&inode->i_lock); 1247 goto retry; 1248 } 1249 1250 /* we flushed them all; remove this inode from the queue */ 1251 spin_lock(&mdsc->snap_flush_lock); 1252 list_del_init(&ci->i_snap_flush_item); 1253 spin_unlock(&mdsc->snap_flush_lock); 1254 1255 if (psession) 1256 *psession = session; 1257 else if (session) { 1258 mutex_unlock(&session->s_mutex); 1259 ceph_put_mds_session(session); 1260 } 1261 } 1262 1263 static void ceph_flush_snaps(struct ceph_inode_info *ci) 1264 { 1265 struct inode *inode = &ci->vfs_inode; 1266 1267 spin_lock(&inode->i_lock); 1268 __ceph_flush_snaps(ci, NULL); 1269 spin_unlock(&inode->i_lock); 1270 } 1271 1272 /* 1273 * Add dirty inode to the flushing list. Assigned a seq number so we 1274 * can wait for caps to flush without starving. 1275 * 1276 * Called under i_lock. 1277 */ 1278 static int __mark_caps_flushing(struct inode *inode, 1279 struct ceph_mds_session *session) 1280 { 1281 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1282 struct ceph_inode_info *ci = ceph_inode(inode); 1283 int flushing; 1284 1285 BUG_ON(ci->i_dirty_caps == 0); 1286 BUG_ON(list_empty(&ci->i_dirty_item)); 1287 1288 flushing = ci->i_dirty_caps; 1289 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", 1290 ceph_cap_string(flushing), 1291 ceph_cap_string(ci->i_flushing_caps), 1292 ceph_cap_string(ci->i_flushing_caps | flushing)); 1293 ci->i_flushing_caps |= flushing; 1294 ci->i_dirty_caps = 0; 1295 1296 spin_lock(&mdsc->cap_dirty_lock); 1297 if (list_empty(&ci->i_flushing_item)) { 1298 list_del_init(&ci->i_dirty_item); 1299 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1300 mdsc->num_cap_flushing++; 1301 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq; 1302 dout(" inode %p now flushing seq %lld\n", &ci->vfs_inode, 1303 ci->i_cap_flush_seq); 1304 } 1305 spin_unlock(&mdsc->cap_dirty_lock); 1306 1307 return flushing; 1308 } 1309 1310 /* 1311 * Swiss army knife function to examine currently used and wanted 1312 * versus held caps. Release, flush, ack revoked caps to mds as 1313 * appropriate. 1314 * 1315 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay 1316 * cap release further. 1317 * CHECK_CAPS_AUTHONLY - we should only check the auth cap 1318 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without 1319 * further delay. 1320 */ 1321 void ceph_check_caps(struct ceph_inode_info *ci, int flags, 1322 struct ceph_mds_session *session) 1323 { 1324 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode); 1325 struct ceph_mds_client *mdsc = &client->mdsc; 1326 struct inode *inode = &ci->vfs_inode; 1327 struct ceph_cap *cap; 1328 int file_wanted, used; 1329 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ 1330 int drop_session_lock = session ? 0 : 1; 1331 int want, retain, revoking, flushing = 0; 1332 int mds = -1; /* keep track of how far we've gone through i_caps list 1333 to avoid an infinite loop on retry */ 1334 struct rb_node *p; 1335 int tried_invalidate = 0; 1336 int delayed = 0, sent = 0, force_requeue = 0, num; 1337 int is_delayed = flags & CHECK_CAPS_NODELAY; 1338 1339 /* if we are unmounting, flush any unused caps immediately. */ 1340 if (mdsc->stopping) 1341 is_delayed = 1; 1342 1343 spin_lock(&inode->i_lock); 1344 1345 if (ci->i_ceph_flags & CEPH_I_FLUSH) 1346 flags |= CHECK_CAPS_FLUSH; 1347 1348 /* flush snaps first time around only */ 1349 if (!list_empty(&ci->i_cap_snaps)) 1350 __ceph_flush_snaps(ci, &session); 1351 goto retry_locked; 1352 retry: 1353 spin_lock(&inode->i_lock); 1354 retry_locked: 1355 file_wanted = __ceph_caps_file_wanted(ci); 1356 used = __ceph_caps_used(ci); 1357 want = file_wanted | used; 1358 1359 retain = want | CEPH_CAP_PIN; 1360 if (!mdsc->stopping && inode->i_nlink > 0) { 1361 if (want) { 1362 retain |= CEPH_CAP_ANY; /* be greedy */ 1363 } else { 1364 retain |= CEPH_CAP_ANY_SHARED; 1365 /* 1366 * keep RD only if we didn't have the file open RW, 1367 * because then the mds would revoke it anyway to 1368 * journal max_size=0. 1369 */ 1370 if (ci->i_max_size == 0) 1371 retain |= CEPH_CAP_ANY_RD; 1372 } 1373 } 1374 1375 dout("check_caps %p file_want %s used %s dirty %s flushing %s" 1376 " issued %s retain %s %s%s%s\n", inode, 1377 ceph_cap_string(file_wanted), 1378 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps), 1379 ceph_cap_string(ci->i_flushing_caps), 1380 ceph_cap_string(__ceph_caps_issued(ci, NULL)), 1381 ceph_cap_string(retain), 1382 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "", 1383 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "", 1384 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : ""); 1385 1386 /* 1387 * If we no longer need to hold onto old our caps, and we may 1388 * have cached pages, but don't want them, then try to invalidate. 1389 * If we fail, it's because pages are locked.... try again later. 1390 */ 1391 if ((!is_delayed || mdsc->stopping) && 1392 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */ 1393 ci->i_rdcache_gen && /* may have cached pages */ 1394 file_wanted == 0 && /* no open files */ 1395 !ci->i_truncate_pending && 1396 !tried_invalidate) { 1397 u32 invalidating_gen = ci->i_rdcache_gen; 1398 int ret; 1399 1400 dout("check_caps trying to invalidate on %p\n", inode); 1401 spin_unlock(&inode->i_lock); 1402 ret = invalidate_inode_pages2(&inode->i_data); 1403 spin_lock(&inode->i_lock); 1404 if (ret == 0 && invalidating_gen == ci->i_rdcache_gen) { 1405 /* success. */ 1406 ci->i_rdcache_gen = 0; 1407 ci->i_rdcache_revoking = 0; 1408 } else { 1409 dout("check_caps failed to invalidate pages\n"); 1410 /* we failed to invalidate pages. check these 1411 caps again later. */ 1412 force_requeue = 1; 1413 __cap_set_timeouts(mdsc, ci); 1414 } 1415 tried_invalidate = 1; 1416 goto retry_locked; 1417 } 1418 1419 num = 0; 1420 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 1421 cap = rb_entry(p, struct ceph_cap, ci_node); 1422 num++; 1423 1424 /* avoid looping forever */ 1425 if (mds >= cap->mds || 1426 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap)) 1427 continue; 1428 1429 /* NOTE: no side-effects allowed, until we take s_mutex */ 1430 1431 revoking = cap->implemented & ~cap->issued; 1432 if (revoking) 1433 dout("mds%d revoking %s\n", cap->mds, 1434 ceph_cap_string(revoking)); 1435 1436 if (cap == ci->i_auth_cap && 1437 (cap->issued & CEPH_CAP_FILE_WR)) { 1438 /* request larger max_size from MDS? */ 1439 if (ci->i_wanted_max_size > ci->i_max_size && 1440 ci->i_wanted_max_size > ci->i_requested_max_size) { 1441 dout("requesting new max_size\n"); 1442 goto ack; 1443 } 1444 1445 /* approaching file_max? */ 1446 if ((inode->i_size << 1) >= ci->i_max_size && 1447 (ci->i_reported_size << 1) < ci->i_max_size) { 1448 dout("i_size approaching max_size\n"); 1449 goto ack; 1450 } 1451 } 1452 /* flush anything dirty? */ 1453 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) && 1454 ci->i_dirty_caps) { 1455 dout("flushing dirty caps\n"); 1456 goto ack; 1457 } 1458 1459 /* completed revocation? going down and there are no caps? */ 1460 if (revoking && (revoking & used) == 0) { 1461 dout("completed revocation of %s\n", 1462 ceph_cap_string(cap->implemented & ~cap->issued)); 1463 goto ack; 1464 } 1465 1466 /* want more caps from mds? */ 1467 if (want & ~(cap->mds_wanted | cap->issued)) 1468 goto ack; 1469 1470 /* things we might delay */ 1471 if ((cap->issued & ~retain) == 0 && 1472 cap->mds_wanted == want) 1473 continue; /* nope, all good */ 1474 1475 if (is_delayed) 1476 goto ack; 1477 1478 /* delay? */ 1479 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1480 time_before(jiffies, ci->i_hold_caps_max)) { 1481 dout(" delaying issued %s -> %s, wanted %s -> %s\n", 1482 ceph_cap_string(cap->issued), 1483 ceph_cap_string(cap->issued & retain), 1484 ceph_cap_string(cap->mds_wanted), 1485 ceph_cap_string(want)); 1486 delayed++; 1487 continue; 1488 } 1489 1490 ack: 1491 if (session && session != cap->session) { 1492 dout("oops, wrong session %p mutex\n", session); 1493 mutex_unlock(&session->s_mutex); 1494 session = NULL; 1495 } 1496 if (!session) { 1497 session = cap->session; 1498 if (mutex_trylock(&session->s_mutex) == 0) { 1499 dout("inverting session/ino locks on %p\n", 1500 session); 1501 spin_unlock(&inode->i_lock); 1502 if (took_snap_rwsem) { 1503 up_read(&mdsc->snap_rwsem); 1504 took_snap_rwsem = 0; 1505 } 1506 mutex_lock(&session->s_mutex); 1507 goto retry; 1508 } 1509 } 1510 /* take snap_rwsem after session mutex */ 1511 if (!took_snap_rwsem) { 1512 if (down_read_trylock(&mdsc->snap_rwsem) == 0) { 1513 dout("inverting snap/in locks on %p\n", 1514 inode); 1515 spin_unlock(&inode->i_lock); 1516 down_read(&mdsc->snap_rwsem); 1517 took_snap_rwsem = 1; 1518 goto retry; 1519 } 1520 took_snap_rwsem = 1; 1521 } 1522 1523 if (cap == ci->i_auth_cap && ci->i_dirty_caps) 1524 flushing = __mark_caps_flushing(inode, session); 1525 1526 mds = cap->mds; /* remember mds, so we don't repeat */ 1527 sent++; 1528 1529 /* __send_cap drops i_lock */ 1530 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want, 1531 retain, flushing, NULL); 1532 goto retry; /* retake i_lock and restart our cap scan. */ 1533 } 1534 1535 /* 1536 * Reschedule delayed caps release if we delayed anything, 1537 * otherwise cancel. 1538 */ 1539 if (delayed && is_delayed) 1540 force_requeue = 1; /* __send_cap delayed release; requeue */ 1541 if (!delayed && !is_delayed) 1542 __cap_delay_cancel(mdsc, ci); 1543 else if (!is_delayed || force_requeue) 1544 __cap_delay_requeue(mdsc, ci); 1545 1546 spin_unlock(&inode->i_lock); 1547 1548 if (session && drop_session_lock) 1549 mutex_unlock(&session->s_mutex); 1550 if (took_snap_rwsem) 1551 up_read(&mdsc->snap_rwsem); 1552 } 1553 1554 /* 1555 * Mark caps dirty. If inode is newly dirty, add to the global dirty 1556 * list. 1557 */ 1558 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask) 1559 { 1560 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc; 1561 struct inode *inode = &ci->vfs_inode; 1562 int was = __ceph_caps_dirty(ci); 1563 int dirty = 0; 1564 1565 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode, 1566 ceph_cap_string(mask), ceph_cap_string(ci->i_dirty_caps), 1567 ceph_cap_string(ci->i_dirty_caps | mask)); 1568 ci->i_dirty_caps |= mask; 1569 if (!was) { 1570 dout(" inode %p now dirty\n", &ci->vfs_inode); 1571 spin_lock(&mdsc->cap_dirty_lock); 1572 list_add(&ci->i_dirty_item, &mdsc->cap_dirty); 1573 spin_unlock(&mdsc->cap_dirty_lock); 1574 igrab(inode); 1575 dirty |= I_DIRTY_SYNC; 1576 } 1577 if ((was & CEPH_CAP_FILE_BUFFER) && 1578 (mask & CEPH_CAP_FILE_BUFFER)) 1579 dirty |= I_DIRTY_DATASYNC; 1580 if (dirty) 1581 __mark_inode_dirty(inode, dirty); 1582 __cap_delay_requeue(mdsc, ci); 1583 return was; 1584 } 1585 1586 /* 1587 * Try to flush dirty caps back to the auth mds. 1588 */ 1589 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session, 1590 unsigned *flush_tid) 1591 { 1592 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1593 struct ceph_inode_info *ci = ceph_inode(inode); 1594 int unlock_session = session ? 0 : 1; 1595 int flushing = 0; 1596 1597 retry: 1598 spin_lock(&inode->i_lock); 1599 if (ci->i_dirty_caps && ci->i_auth_cap) { 1600 struct ceph_cap *cap = ci->i_auth_cap; 1601 int used = __ceph_caps_used(ci); 1602 int want = __ceph_caps_wanted(ci); 1603 int delayed; 1604 1605 if (!session) { 1606 spin_unlock(&inode->i_lock); 1607 session = cap->session; 1608 mutex_lock(&session->s_mutex); 1609 goto retry; 1610 } 1611 BUG_ON(session != cap->session); 1612 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) 1613 goto out; 1614 1615 flushing = __mark_caps_flushing(inode, session); 1616 1617 /* __send_cap drops i_lock */ 1618 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want, 1619 cap->issued | cap->implemented, flushing, 1620 flush_tid); 1621 if (!delayed) 1622 goto out_unlocked; 1623 1624 spin_lock(&inode->i_lock); 1625 __cap_delay_requeue(mdsc, ci); 1626 } 1627 out: 1628 spin_unlock(&inode->i_lock); 1629 out_unlocked: 1630 if (session && unlock_session) 1631 mutex_unlock(&session->s_mutex); 1632 return flushing; 1633 } 1634 1635 /* 1636 * Return true if we've flushed caps through the given flush_tid. 1637 */ 1638 static int caps_are_flushed(struct inode *inode, unsigned tid) 1639 { 1640 struct ceph_inode_info *ci = ceph_inode(inode); 1641 int dirty, i, ret = 1; 1642 1643 spin_lock(&inode->i_lock); 1644 dirty = __ceph_caps_dirty(ci); 1645 for (i = 0; i < CEPH_CAP_BITS; i++) 1646 if ((ci->i_flushing_caps & (1 << i)) && 1647 ci->i_cap_flush_tid[i] <= tid) { 1648 /* still flushing this bit */ 1649 ret = 0; 1650 break; 1651 } 1652 spin_unlock(&inode->i_lock); 1653 return ret; 1654 } 1655 1656 /* 1657 * Wait on any unsafe replies for the given inode. First wait on the 1658 * newest request, and make that the upper bound. Then, if there are 1659 * more requests, keep waiting on the oldest as long as it is still older 1660 * than the original request. 1661 */ 1662 static void sync_write_wait(struct inode *inode) 1663 { 1664 struct ceph_inode_info *ci = ceph_inode(inode); 1665 struct list_head *head = &ci->i_unsafe_writes; 1666 struct ceph_osd_request *req; 1667 u64 last_tid; 1668 1669 spin_lock(&ci->i_unsafe_lock); 1670 if (list_empty(head)) 1671 goto out; 1672 1673 /* set upper bound as _last_ entry in chain */ 1674 req = list_entry(head->prev, struct ceph_osd_request, 1675 r_unsafe_item); 1676 last_tid = req->r_tid; 1677 1678 do { 1679 ceph_osdc_get_request(req); 1680 spin_unlock(&ci->i_unsafe_lock); 1681 dout("sync_write_wait on tid %llu (until %llu)\n", 1682 req->r_tid, last_tid); 1683 wait_for_completion(&req->r_safe_completion); 1684 spin_lock(&ci->i_unsafe_lock); 1685 ceph_osdc_put_request(req); 1686 1687 /* 1688 * from here on look at first entry in chain, since we 1689 * only want to wait for anything older than last_tid 1690 */ 1691 if (list_empty(head)) 1692 break; 1693 req = list_entry(head->next, struct ceph_osd_request, 1694 r_unsafe_item); 1695 } while (req->r_tid < last_tid); 1696 out: 1697 spin_unlock(&ci->i_unsafe_lock); 1698 } 1699 1700 int ceph_fsync(struct file *file, struct dentry *dentry, int datasync) 1701 { 1702 struct inode *inode = dentry->d_inode; 1703 struct ceph_inode_info *ci = ceph_inode(inode); 1704 unsigned flush_tid; 1705 int ret; 1706 int dirty; 1707 1708 dout("fsync %p%s\n", inode, datasync ? " datasync" : ""); 1709 sync_write_wait(inode); 1710 1711 ret = filemap_write_and_wait(inode->i_mapping); 1712 if (ret < 0) 1713 return ret; 1714 1715 dirty = try_flush_caps(inode, NULL, &flush_tid); 1716 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); 1717 1718 /* 1719 * only wait on non-file metadata writeback (the mds 1720 * can recover size and mtime, so we don't need to 1721 * wait for that) 1722 */ 1723 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { 1724 dout("fsync waiting for flush_tid %u\n", flush_tid); 1725 ret = wait_event_interruptible(ci->i_cap_wq, 1726 caps_are_flushed(inode, flush_tid)); 1727 } 1728 1729 dout("fsync %p%s done\n", inode, datasync ? " datasync" : ""); 1730 return ret; 1731 } 1732 1733 /* 1734 * Flush any dirty caps back to the mds. If we aren't asked to wait, 1735 * queue inode for flush but don't do so immediately, because we can 1736 * get by with fewer MDS messages if we wait for data writeback to 1737 * complete first. 1738 */ 1739 int ceph_write_inode(struct inode *inode, int wait) 1740 { 1741 struct ceph_inode_info *ci = ceph_inode(inode); 1742 unsigned flush_tid; 1743 int err = 0; 1744 int dirty; 1745 1746 dout("write_inode %p wait=%d\n", inode, wait); 1747 if (wait) { 1748 dirty = try_flush_caps(inode, NULL, &flush_tid); 1749 if (dirty) 1750 err = wait_event_interruptible(ci->i_cap_wq, 1751 caps_are_flushed(inode, flush_tid)); 1752 } else { 1753 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1754 1755 spin_lock(&inode->i_lock); 1756 if (__ceph_caps_dirty(ci)) 1757 __cap_delay_requeue_front(mdsc, ci); 1758 spin_unlock(&inode->i_lock); 1759 } 1760 return err; 1761 } 1762 1763 /* 1764 * After a recovering MDS goes active, we need to resend any caps 1765 * we were flushing. 1766 * 1767 * Caller holds session->s_mutex. 1768 */ 1769 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc, 1770 struct ceph_mds_session *session) 1771 { 1772 struct ceph_cap_snap *capsnap; 1773 1774 dout("kick_flushing_capsnaps mds%d\n", session->s_mds); 1775 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing, 1776 flushing_item) { 1777 struct ceph_inode_info *ci = capsnap->ci; 1778 struct inode *inode = &ci->vfs_inode; 1779 struct ceph_cap *cap; 1780 1781 spin_lock(&inode->i_lock); 1782 cap = ci->i_auth_cap; 1783 if (cap && cap->session == session) { 1784 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode, 1785 cap, capsnap); 1786 __ceph_flush_snaps(ci, &session); 1787 } else { 1788 pr_err("%p auth cap %p not mds%d ???\n", inode, 1789 cap, session->s_mds); 1790 spin_unlock(&inode->i_lock); 1791 } 1792 } 1793 } 1794 1795 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 1796 struct ceph_mds_session *session) 1797 { 1798 struct ceph_inode_info *ci; 1799 1800 kick_flushing_capsnaps(mdsc, session); 1801 1802 dout("kick_flushing_caps mds%d\n", session->s_mds); 1803 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 1804 struct inode *inode = &ci->vfs_inode; 1805 struct ceph_cap *cap; 1806 int delayed = 0; 1807 1808 spin_lock(&inode->i_lock); 1809 cap = ci->i_auth_cap; 1810 if (cap && cap->session == session) { 1811 dout("kick_flushing_caps %p cap %p %s\n", inode, 1812 cap, ceph_cap_string(ci->i_flushing_caps)); 1813 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, 1814 __ceph_caps_used(ci), 1815 __ceph_caps_wanted(ci), 1816 cap->issued | cap->implemented, 1817 ci->i_flushing_caps, NULL); 1818 if (delayed) { 1819 spin_lock(&inode->i_lock); 1820 __cap_delay_requeue(mdsc, ci); 1821 spin_unlock(&inode->i_lock); 1822 } 1823 } else { 1824 pr_err("%p auth cap %p not mds%d ???\n", inode, 1825 cap, session->s_mds); 1826 spin_unlock(&inode->i_lock); 1827 } 1828 } 1829 } 1830 1831 1832 /* 1833 * Take references to capabilities we hold, so that we don't release 1834 * them to the MDS prematurely. 1835 * 1836 * Protected by i_lock. 1837 */ 1838 static void __take_cap_refs(struct ceph_inode_info *ci, int got) 1839 { 1840 if (got & CEPH_CAP_PIN) 1841 ci->i_pin_ref++; 1842 if (got & CEPH_CAP_FILE_RD) 1843 ci->i_rd_ref++; 1844 if (got & CEPH_CAP_FILE_CACHE) 1845 ci->i_rdcache_ref++; 1846 if (got & CEPH_CAP_FILE_WR) 1847 ci->i_wr_ref++; 1848 if (got & CEPH_CAP_FILE_BUFFER) { 1849 if (ci->i_wrbuffer_ref == 0) 1850 igrab(&ci->vfs_inode); 1851 ci->i_wrbuffer_ref++; 1852 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n", 1853 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref); 1854 } 1855 } 1856 1857 /* 1858 * Try to grab cap references. Specify those refs we @want, and the 1859 * minimal set we @need. Also include the larger offset we are writing 1860 * to (when applicable), and check against max_size here as well. 1861 * Note that caller is responsible for ensuring max_size increases are 1862 * requested from the MDS. 1863 */ 1864 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want, 1865 int *got, loff_t endoff, int *check_max, int *err) 1866 { 1867 struct inode *inode = &ci->vfs_inode; 1868 int ret = 0; 1869 int have, implemented; 1870 1871 dout("get_cap_refs %p need %s want %s\n", inode, 1872 ceph_cap_string(need), ceph_cap_string(want)); 1873 spin_lock(&inode->i_lock); 1874 1875 /* make sure we _have_ some caps! */ 1876 if (!__ceph_is_any_caps(ci)) { 1877 dout("get_cap_refs %p no real caps\n", inode); 1878 *err = -EBADF; 1879 ret = 1; 1880 goto out; 1881 } 1882 1883 if (need & CEPH_CAP_FILE_WR) { 1884 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) { 1885 dout("get_cap_refs %p endoff %llu > maxsize %llu\n", 1886 inode, endoff, ci->i_max_size); 1887 if (endoff > ci->i_wanted_max_size) { 1888 *check_max = 1; 1889 ret = 1; 1890 } 1891 goto out; 1892 } 1893 /* 1894 * If a sync write is in progress, we must wait, so that we 1895 * can get a final snapshot value for size+mtime. 1896 */ 1897 if (__ceph_have_pending_cap_snap(ci)) { 1898 dout("get_cap_refs %p cap_snap_pending\n", inode); 1899 goto out; 1900 } 1901 } 1902 have = __ceph_caps_issued(ci, &implemented); 1903 1904 /* 1905 * disallow writes while a truncate is pending 1906 */ 1907 if (ci->i_truncate_pending) 1908 have &= ~CEPH_CAP_FILE_WR; 1909 1910 if ((have & need) == need) { 1911 /* 1912 * Look at (implemented & ~have & not) so that we keep waiting 1913 * on transition from wanted -> needed caps. This is needed 1914 * for WRBUFFER|WR -> WR to avoid a new WR sync write from 1915 * going before a prior buffered writeback happens. 1916 */ 1917 int not = want & ~(have & need); 1918 int revoking = implemented & ~have; 1919 dout("get_cap_refs %p have %s but not %s (revoking %s)\n", 1920 inode, ceph_cap_string(have), ceph_cap_string(not), 1921 ceph_cap_string(revoking)); 1922 if ((revoking & not) == 0) { 1923 *got = need | (have & want); 1924 __take_cap_refs(ci, *got); 1925 ret = 1; 1926 } 1927 } else { 1928 dout("get_cap_refs %p have %s needed %s\n", inode, 1929 ceph_cap_string(have), ceph_cap_string(need)); 1930 } 1931 out: 1932 spin_unlock(&inode->i_lock); 1933 dout("get_cap_refs %p ret %d got %s\n", inode, 1934 ret, ceph_cap_string(*got)); 1935 return ret; 1936 } 1937 1938 /* 1939 * Check the offset we are writing up to against our current 1940 * max_size. If necessary, tell the MDS we want to write to 1941 * a larger offset. 1942 */ 1943 static void check_max_size(struct inode *inode, loff_t endoff) 1944 { 1945 struct ceph_inode_info *ci = ceph_inode(inode); 1946 int check = 0; 1947 1948 /* do we need to explicitly request a larger max_size? */ 1949 spin_lock(&inode->i_lock); 1950 if ((endoff >= ci->i_max_size || 1951 endoff > (inode->i_size << 1)) && 1952 endoff > ci->i_wanted_max_size) { 1953 dout("write %p at large endoff %llu, req max_size\n", 1954 inode, endoff); 1955 ci->i_wanted_max_size = endoff; 1956 check = 1; 1957 } 1958 spin_unlock(&inode->i_lock); 1959 if (check) 1960 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 1961 } 1962 1963 /* 1964 * Wait for caps, and take cap references. If we can't get a WR cap 1965 * due to a small max_size, make sure we check_max_size (and possibly 1966 * ask the mds) so we don't get hung up indefinitely. 1967 */ 1968 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got, 1969 loff_t endoff) 1970 { 1971 int check_max, ret, err; 1972 1973 retry: 1974 if (endoff > 0) 1975 check_max_size(&ci->vfs_inode, endoff); 1976 check_max = 0; 1977 err = 0; 1978 ret = wait_event_interruptible(ci->i_cap_wq, 1979 try_get_cap_refs(ci, need, want, 1980 got, endoff, 1981 &check_max, &err)); 1982 if (err) 1983 ret = err; 1984 if (check_max) 1985 goto retry; 1986 return ret; 1987 } 1988 1989 /* 1990 * Take cap refs. Caller must already know we hold at least one ref 1991 * on the caps in question or we don't know this is safe. 1992 */ 1993 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) 1994 { 1995 spin_lock(&ci->vfs_inode.i_lock); 1996 __take_cap_refs(ci, caps); 1997 spin_unlock(&ci->vfs_inode.i_lock); 1998 } 1999 2000 /* 2001 * Release cap refs. 2002 * 2003 * If we released the last ref on any given cap, call ceph_check_caps 2004 * to release (or schedule a release). 2005 * 2006 * If we are releasing a WR cap (from a sync write), finalize any affected 2007 * cap_snap, and wake up any waiters. 2008 */ 2009 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had) 2010 { 2011 struct inode *inode = &ci->vfs_inode; 2012 int last = 0, put = 0, flushsnaps = 0, wake = 0; 2013 struct ceph_cap_snap *capsnap; 2014 2015 spin_lock(&inode->i_lock); 2016 if (had & CEPH_CAP_PIN) 2017 --ci->i_pin_ref; 2018 if (had & CEPH_CAP_FILE_RD) 2019 if (--ci->i_rd_ref == 0) 2020 last++; 2021 if (had & CEPH_CAP_FILE_CACHE) 2022 if (--ci->i_rdcache_ref == 0) 2023 last++; 2024 if (had & CEPH_CAP_FILE_BUFFER) { 2025 if (--ci->i_wrbuffer_ref == 0) { 2026 last++; 2027 put++; 2028 } 2029 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n", 2030 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref); 2031 } 2032 if (had & CEPH_CAP_FILE_WR) 2033 if (--ci->i_wr_ref == 0) { 2034 last++; 2035 if (!list_empty(&ci->i_cap_snaps)) { 2036 capsnap = list_first_entry(&ci->i_cap_snaps, 2037 struct ceph_cap_snap, 2038 ci_item); 2039 if (capsnap->writing) { 2040 capsnap->writing = 0; 2041 flushsnaps = 2042 __ceph_finish_cap_snap(ci, 2043 capsnap); 2044 wake = 1; 2045 } 2046 } 2047 } 2048 spin_unlock(&inode->i_lock); 2049 2050 dout("put_cap_refs %p had %s %s\n", inode, ceph_cap_string(had), 2051 last ? "last" : ""); 2052 2053 if (last && !flushsnaps) 2054 ceph_check_caps(ci, 0, NULL); 2055 else if (flushsnaps) 2056 ceph_flush_snaps(ci); 2057 if (wake) 2058 wake_up(&ci->i_cap_wq); 2059 if (put) 2060 iput(inode); 2061 } 2062 2063 /* 2064 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap 2065 * context. Adjust per-snap dirty page accounting as appropriate. 2066 * Once all dirty data for a cap_snap is flushed, flush snapped file 2067 * metadata back to the MDS. If we dropped the last ref, call 2068 * ceph_check_caps. 2069 */ 2070 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr, 2071 struct ceph_snap_context *snapc) 2072 { 2073 struct inode *inode = &ci->vfs_inode; 2074 int last = 0; 2075 int last_snap = 0; 2076 int found = 0; 2077 struct ceph_cap_snap *capsnap = NULL; 2078 2079 spin_lock(&inode->i_lock); 2080 ci->i_wrbuffer_ref -= nr; 2081 last = !ci->i_wrbuffer_ref; 2082 2083 if (ci->i_head_snapc == snapc) { 2084 ci->i_wrbuffer_ref_head -= nr; 2085 if (!ci->i_wrbuffer_ref_head) { 2086 ceph_put_snap_context(ci->i_head_snapc); 2087 ci->i_head_snapc = NULL; 2088 } 2089 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n", 2090 inode, 2091 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr, 2092 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 2093 last ? " LAST" : ""); 2094 } else { 2095 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2096 if (capsnap->context == snapc) { 2097 found = 1; 2098 capsnap->dirty_pages -= nr; 2099 last_snap = !capsnap->dirty_pages; 2100 break; 2101 } 2102 } 2103 BUG_ON(!found); 2104 dout("put_wrbuffer_cap_refs on %p cap_snap %p " 2105 " snap %lld %d/%d -> %d/%d %s%s\n", 2106 inode, capsnap, capsnap->context->seq, 2107 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, 2108 ci->i_wrbuffer_ref, capsnap->dirty_pages, 2109 last ? " (wrbuffer last)" : "", 2110 last_snap ? " (capsnap last)" : ""); 2111 } 2112 2113 spin_unlock(&inode->i_lock); 2114 2115 if (last) { 2116 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2117 iput(inode); 2118 } else if (last_snap) { 2119 ceph_flush_snaps(ci); 2120 wake_up(&ci->i_cap_wq); 2121 } 2122 } 2123 2124 /* 2125 * Handle a cap GRANT message from the MDS. (Note that a GRANT may 2126 * actually be a revocation if it specifies a smaller cap set.) 2127 * 2128 * caller holds s_mutex. 2129 * return value: 2130 * 0 - ok 2131 * 1 - check_caps on auth cap only (writeback) 2132 * 2 - check_caps (ack revoke) 2133 */ 2134 static int handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant, 2135 struct ceph_mds_session *session, 2136 struct ceph_cap *cap, 2137 struct ceph_buffer *xattr_buf) 2138 __releases(inode->i_lock) 2139 2140 { 2141 struct ceph_inode_info *ci = ceph_inode(inode); 2142 int mds = session->s_mds; 2143 int seq = le32_to_cpu(grant->seq); 2144 int newcaps = le32_to_cpu(grant->caps); 2145 int issued, implemented, used, wanted, dirty; 2146 u64 size = le64_to_cpu(grant->size); 2147 u64 max_size = le64_to_cpu(grant->max_size); 2148 struct timespec mtime, atime, ctime; 2149 int reply = 0; 2150 int wake = 0; 2151 int writeback = 0; 2152 int revoked_rdcache = 0; 2153 int invalidate_async = 0; 2154 int tried_invalidate = 0; 2155 int ret; 2156 2157 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n", 2158 inode, cap, mds, seq, ceph_cap_string(newcaps)); 2159 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size, 2160 inode->i_size); 2161 2162 /* 2163 * If CACHE is being revoked, and we have no dirty buffers, 2164 * try to invalidate (once). (If there are dirty buffers, we 2165 * will invalidate _after_ writeback.) 2166 */ 2167 restart: 2168 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && 2169 !ci->i_wrbuffer_ref && !tried_invalidate) { 2170 dout("CACHE invalidation\n"); 2171 spin_unlock(&inode->i_lock); 2172 tried_invalidate = 1; 2173 2174 ret = invalidate_inode_pages2(&inode->i_data); 2175 spin_lock(&inode->i_lock); 2176 if (ret < 0) { 2177 /* there were locked pages.. invalidate later 2178 in a separate thread. */ 2179 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) { 2180 invalidate_async = 1; 2181 ci->i_rdcache_revoking = ci->i_rdcache_gen; 2182 } 2183 } else { 2184 /* we successfully invalidated those pages */ 2185 revoked_rdcache = 1; 2186 ci->i_rdcache_gen = 0; 2187 ci->i_rdcache_revoking = 0; 2188 } 2189 goto restart; 2190 } 2191 2192 /* side effects now are allowed */ 2193 2194 issued = __ceph_caps_issued(ci, &implemented); 2195 issued |= implemented | __ceph_caps_dirty(ci); 2196 2197 cap->gen = session->s_cap_gen; 2198 2199 __check_cap_issue(ci, cap, newcaps); 2200 2201 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) { 2202 inode->i_mode = le32_to_cpu(grant->mode); 2203 inode->i_uid = le32_to_cpu(grant->uid); 2204 inode->i_gid = le32_to_cpu(grant->gid); 2205 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode, 2206 inode->i_uid, inode->i_gid); 2207 } 2208 2209 if ((issued & CEPH_CAP_LINK_EXCL) == 0) 2210 inode->i_nlink = le32_to_cpu(grant->nlink); 2211 2212 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) { 2213 int len = le32_to_cpu(grant->xattr_len); 2214 u64 version = le64_to_cpu(grant->xattr_version); 2215 2216 if (version > ci->i_xattrs.version) { 2217 dout(" got new xattrs v%llu on %p len %d\n", 2218 version, inode, len); 2219 if (ci->i_xattrs.blob) 2220 ceph_buffer_put(ci->i_xattrs.blob); 2221 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf); 2222 ci->i_xattrs.version = version; 2223 } 2224 } 2225 2226 /* size/ctime/mtime/atime? */ 2227 ceph_fill_file_size(inode, issued, 2228 le32_to_cpu(grant->truncate_seq), 2229 le64_to_cpu(grant->truncate_size), size); 2230 ceph_decode_timespec(&mtime, &grant->mtime); 2231 ceph_decode_timespec(&atime, &grant->atime); 2232 ceph_decode_timespec(&ctime, &grant->ctime); 2233 ceph_fill_file_time(inode, issued, 2234 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime, 2235 &atime); 2236 2237 /* max size increase? */ 2238 if (max_size != ci->i_max_size) { 2239 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size); 2240 ci->i_max_size = max_size; 2241 if (max_size >= ci->i_wanted_max_size) { 2242 ci->i_wanted_max_size = 0; /* reset */ 2243 ci->i_requested_max_size = 0; 2244 } 2245 wake = 1; 2246 } 2247 2248 /* check cap bits */ 2249 wanted = __ceph_caps_wanted(ci); 2250 used = __ceph_caps_used(ci); 2251 dirty = __ceph_caps_dirty(ci); 2252 dout(" my wanted = %s, used = %s, dirty %s\n", 2253 ceph_cap_string(wanted), 2254 ceph_cap_string(used), 2255 ceph_cap_string(dirty)); 2256 if (wanted != le32_to_cpu(grant->wanted)) { 2257 dout("mds wanted %s -> %s\n", 2258 ceph_cap_string(le32_to_cpu(grant->wanted)), 2259 ceph_cap_string(wanted)); 2260 grant->wanted = cpu_to_le32(wanted); 2261 } 2262 2263 cap->seq = seq; 2264 2265 /* file layout may have changed */ 2266 ci->i_layout = grant->layout; 2267 2268 /* revocation, grant, or no-op? */ 2269 if (cap->issued & ~newcaps) { 2270 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued), 2271 ceph_cap_string(newcaps)); 2272 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER) 2273 writeback = 1; /* will delay ack */ 2274 else if (dirty & ~newcaps) 2275 reply = 1; /* initiate writeback in check_caps */ 2276 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 || 2277 revoked_rdcache) 2278 reply = 2; /* send revoke ack in check_caps */ 2279 cap->issued = newcaps; 2280 } else if (cap->issued == newcaps) { 2281 dout("caps unchanged: %s -> %s\n", 2282 ceph_cap_string(cap->issued), ceph_cap_string(newcaps)); 2283 } else { 2284 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued), 2285 ceph_cap_string(newcaps)); 2286 cap->issued = newcaps; 2287 cap->implemented |= newcaps; /* add bits only, to 2288 * avoid stepping on a 2289 * pending revocation */ 2290 wake = 1; 2291 } 2292 2293 spin_unlock(&inode->i_lock); 2294 if (writeback) { 2295 /* 2296 * queue inode for writeback: we can't actually call 2297 * filemap_write_and_wait, etc. from message handler 2298 * context. 2299 */ 2300 dout("queueing %p for writeback\n", inode); 2301 if (ceph_queue_writeback(inode)) 2302 igrab(inode); 2303 } 2304 if (invalidate_async) { 2305 dout("queueing %p for page invalidation\n", inode); 2306 if (ceph_queue_page_invalidation(inode)) 2307 igrab(inode); 2308 } 2309 if (wake) 2310 wake_up(&ci->i_cap_wq); 2311 return reply; 2312 } 2313 2314 /* 2315 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the 2316 * MDS has been safely committed. 2317 */ 2318 static void handle_cap_flush_ack(struct inode *inode, 2319 struct ceph_mds_caps *m, 2320 struct ceph_mds_session *session, 2321 struct ceph_cap *cap) 2322 __releases(inode->i_lock) 2323 { 2324 struct ceph_inode_info *ci = ceph_inode(inode); 2325 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 2326 unsigned seq = le32_to_cpu(m->seq); 2327 int dirty = le32_to_cpu(m->dirty); 2328 int cleaned = 0; 2329 u64 flush_tid = le64_to_cpu(m->client_tid); 2330 int old_dirty = 0, new_dirty = 0; 2331 int i; 2332 2333 for (i = 0; i < CEPH_CAP_BITS; i++) 2334 if ((dirty & (1 << i)) && 2335 flush_tid == ci->i_cap_flush_tid[i]) 2336 cleaned |= 1 << i; 2337 2338 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," 2339 " flushing %s -> %s\n", 2340 inode, session->s_mds, seq, ceph_cap_string(dirty), 2341 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), 2342 ceph_cap_string(ci->i_flushing_caps & ~cleaned)); 2343 2344 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned)) 2345 goto out; 2346 2347 old_dirty = ci->i_dirty_caps | ci->i_flushing_caps; 2348 ci->i_flushing_caps &= ~cleaned; 2349 new_dirty = ci->i_dirty_caps | ci->i_flushing_caps; 2350 2351 spin_lock(&mdsc->cap_dirty_lock); 2352 if (ci->i_flushing_caps == 0) { 2353 list_del_init(&ci->i_flushing_item); 2354 if (!list_empty(&session->s_cap_flushing)) 2355 dout(" mds%d still flushing cap on %p\n", 2356 session->s_mds, 2357 &list_entry(session->s_cap_flushing.next, 2358 struct ceph_inode_info, 2359 i_flushing_item)->vfs_inode); 2360 mdsc->num_cap_flushing--; 2361 wake_up(&mdsc->cap_flushing_wq); 2362 dout(" inode %p now !flushing\n", inode); 2363 } 2364 if (old_dirty && !new_dirty) { 2365 dout(" inode %p now clean\n", inode); 2366 list_del_init(&ci->i_dirty_item); 2367 } 2368 spin_unlock(&mdsc->cap_dirty_lock); 2369 wake_up(&ci->i_cap_wq); 2370 2371 out: 2372 spin_unlock(&inode->i_lock); 2373 if (old_dirty && !new_dirty) 2374 iput(inode); 2375 } 2376 2377 /* 2378 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can 2379 * throw away our cap_snap. 2380 * 2381 * Caller hold s_mutex. 2382 */ 2383 static void handle_cap_flushsnap_ack(struct inode *inode, 2384 struct ceph_mds_caps *m, 2385 struct ceph_mds_session *session) 2386 { 2387 struct ceph_inode_info *ci = ceph_inode(inode); 2388 u64 follows = le64_to_cpu(m->snap_follows); 2389 u64 flush_tid = le64_to_cpu(m->client_tid); 2390 struct ceph_cap_snap *capsnap; 2391 int drop = 0; 2392 2393 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n", 2394 inode, ci, session->s_mds, follows); 2395 2396 spin_lock(&inode->i_lock); 2397 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2398 if (capsnap->follows == follows) { 2399 if (capsnap->flush_tid != flush_tid) { 2400 dout(" cap_snap %p follows %lld tid %lld !=" 2401 " %lld\n", capsnap, follows, 2402 flush_tid, capsnap->flush_tid); 2403 break; 2404 } 2405 WARN_ON(capsnap->dirty_pages || capsnap->writing); 2406 dout(" removing cap_snap %p follows %lld\n", 2407 capsnap, follows); 2408 ceph_put_snap_context(capsnap->context); 2409 list_del(&capsnap->ci_item); 2410 list_del(&capsnap->flushing_item); 2411 ceph_put_cap_snap(capsnap); 2412 drop = 1; 2413 break; 2414 } else { 2415 dout(" skipping cap_snap %p follows %lld\n", 2416 capsnap, capsnap->follows); 2417 } 2418 } 2419 spin_unlock(&inode->i_lock); 2420 if (drop) 2421 iput(inode); 2422 } 2423 2424 /* 2425 * Handle TRUNC from MDS, indicating file truncation. 2426 * 2427 * caller hold s_mutex. 2428 */ 2429 static void handle_cap_trunc(struct inode *inode, 2430 struct ceph_mds_caps *trunc, 2431 struct ceph_mds_session *session) 2432 __releases(inode->i_lock) 2433 { 2434 struct ceph_inode_info *ci = ceph_inode(inode); 2435 int mds = session->s_mds; 2436 int seq = le32_to_cpu(trunc->seq); 2437 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq); 2438 u64 truncate_size = le64_to_cpu(trunc->truncate_size); 2439 u64 size = le64_to_cpu(trunc->size); 2440 int implemented = 0; 2441 int dirty = __ceph_caps_dirty(ci); 2442 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented); 2443 int queue_trunc = 0; 2444 2445 issued |= implemented | dirty; 2446 2447 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n", 2448 inode, mds, seq, truncate_size, truncate_seq); 2449 queue_trunc = ceph_fill_file_size(inode, issued, 2450 truncate_seq, truncate_size, size); 2451 spin_unlock(&inode->i_lock); 2452 2453 if (queue_trunc) 2454 if (queue_work(ceph_client(inode->i_sb)->trunc_wq, 2455 &ci->i_vmtruncate_work)) 2456 igrab(inode); 2457 } 2458 2459 /* 2460 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a 2461 * different one. If we are the most recent migration we've seen (as 2462 * indicated by mseq), make note of the migrating cap bits for the 2463 * duration (until we see the corresponding IMPORT). 2464 * 2465 * caller holds s_mutex 2466 */ 2467 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex, 2468 struct ceph_mds_session *session) 2469 { 2470 struct ceph_inode_info *ci = ceph_inode(inode); 2471 int mds = session->s_mds; 2472 unsigned mseq = le32_to_cpu(ex->migrate_seq); 2473 struct ceph_cap *cap = NULL, *t; 2474 struct rb_node *p; 2475 int remember = 1; 2476 2477 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n", 2478 inode, ci, mds, mseq); 2479 2480 spin_lock(&inode->i_lock); 2481 2482 /* make sure we haven't seen a higher mseq */ 2483 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 2484 t = rb_entry(p, struct ceph_cap, ci_node); 2485 if (ceph_seq_cmp(t->mseq, mseq) > 0) { 2486 dout(" higher mseq on cap from mds%d\n", 2487 t->session->s_mds); 2488 remember = 0; 2489 } 2490 if (t->session->s_mds == mds) 2491 cap = t; 2492 } 2493 2494 if (cap) { 2495 if (remember) { 2496 /* make note */ 2497 ci->i_cap_exporting_mds = mds; 2498 ci->i_cap_exporting_mseq = mseq; 2499 ci->i_cap_exporting_issued = cap->issued; 2500 } 2501 __ceph_remove_cap(cap, NULL); 2502 } else { 2503 WARN_ON(!cap); 2504 } 2505 2506 spin_unlock(&inode->i_lock); 2507 } 2508 2509 /* 2510 * Handle cap IMPORT. If there are temp bits from an older EXPORT, 2511 * clean them up. 2512 * 2513 * caller holds s_mutex. 2514 */ 2515 static void handle_cap_import(struct ceph_mds_client *mdsc, 2516 struct inode *inode, struct ceph_mds_caps *im, 2517 struct ceph_mds_session *session, 2518 void *snaptrace, int snaptrace_len) 2519 { 2520 struct ceph_inode_info *ci = ceph_inode(inode); 2521 int mds = session->s_mds; 2522 unsigned issued = le32_to_cpu(im->caps); 2523 unsigned wanted = le32_to_cpu(im->wanted); 2524 unsigned seq = le32_to_cpu(im->seq); 2525 unsigned mseq = le32_to_cpu(im->migrate_seq); 2526 u64 realmino = le64_to_cpu(im->realm); 2527 u64 cap_id = le64_to_cpu(im->cap_id); 2528 2529 if (ci->i_cap_exporting_mds >= 0 && 2530 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) { 2531 dout("handle_cap_import inode %p ci %p mds%d mseq %d" 2532 " - cleared exporting from mds%d\n", 2533 inode, ci, mds, mseq, 2534 ci->i_cap_exporting_mds); 2535 ci->i_cap_exporting_issued = 0; 2536 ci->i_cap_exporting_mseq = 0; 2537 ci->i_cap_exporting_mds = -1; 2538 } else { 2539 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n", 2540 inode, ci, mds, mseq); 2541 } 2542 2543 down_write(&mdsc->snap_rwsem); 2544 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len, 2545 false); 2546 downgrade_write(&mdsc->snap_rwsem); 2547 ceph_add_cap(inode, session, cap_id, -1, 2548 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH, 2549 NULL /* no caps context */); 2550 try_flush_caps(inode, session, NULL); 2551 up_read(&mdsc->snap_rwsem); 2552 } 2553 2554 /* 2555 * Handle a caps message from the MDS. 2556 * 2557 * Identify the appropriate session, inode, and call the right handler 2558 * based on the cap op. 2559 */ 2560 void ceph_handle_caps(struct ceph_mds_session *session, 2561 struct ceph_msg *msg) 2562 { 2563 struct ceph_mds_client *mdsc = session->s_mdsc; 2564 struct super_block *sb = mdsc->client->sb; 2565 struct inode *inode; 2566 struct ceph_cap *cap; 2567 struct ceph_mds_caps *h; 2568 int mds = le64_to_cpu(msg->hdr.src.name.num); 2569 int op; 2570 u32 seq; 2571 struct ceph_vino vino; 2572 u64 cap_id; 2573 u64 size, max_size; 2574 int check_caps = 0; 2575 int r; 2576 2577 dout("handle_caps from mds%d\n", mds); 2578 2579 /* decode */ 2580 if (msg->front.iov_len < sizeof(*h)) 2581 goto bad; 2582 h = msg->front.iov_base; 2583 op = le32_to_cpu(h->op); 2584 vino.ino = le64_to_cpu(h->ino); 2585 vino.snap = CEPH_NOSNAP; 2586 cap_id = le64_to_cpu(h->cap_id); 2587 seq = le32_to_cpu(h->seq); 2588 size = le64_to_cpu(h->size); 2589 max_size = le64_to_cpu(h->max_size); 2590 2591 mutex_lock(&session->s_mutex); 2592 session->s_seq++; 2593 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, 2594 (unsigned)seq); 2595 2596 /* lookup ino */ 2597 inode = ceph_find_inode(sb, vino); 2598 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino, 2599 vino.snap, inode); 2600 if (!inode) { 2601 dout(" i don't have ino %llx\n", vino.ino); 2602 goto done; 2603 } 2604 2605 /* these will work even if we don't have a cap yet */ 2606 switch (op) { 2607 case CEPH_CAP_OP_FLUSHSNAP_ACK: 2608 handle_cap_flushsnap_ack(inode, h, session); 2609 goto done; 2610 2611 case CEPH_CAP_OP_EXPORT: 2612 handle_cap_export(inode, h, session); 2613 goto done; 2614 2615 case CEPH_CAP_OP_IMPORT: 2616 handle_cap_import(mdsc, inode, h, session, 2617 msg->middle, 2618 le32_to_cpu(h->snap_trace_len)); 2619 check_caps = 1; /* we may have sent a RELEASE to the old auth */ 2620 goto done; 2621 } 2622 2623 /* the rest require a cap */ 2624 spin_lock(&inode->i_lock); 2625 cap = __get_cap_for_mds(ceph_inode(inode), mds); 2626 if (!cap) { 2627 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n", 2628 inode, ceph_ino(inode), ceph_snap(inode), mds); 2629 spin_unlock(&inode->i_lock); 2630 goto done; 2631 } 2632 2633 /* note that each of these drops i_lock for us */ 2634 switch (op) { 2635 case CEPH_CAP_OP_REVOKE: 2636 case CEPH_CAP_OP_GRANT: 2637 r = handle_cap_grant(inode, h, session, cap, msg->middle); 2638 if (r == 1) 2639 ceph_check_caps(ceph_inode(inode), 2640 CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY, 2641 session); 2642 else if (r == 2) 2643 ceph_check_caps(ceph_inode(inode), 2644 CHECK_CAPS_NODELAY, 2645 session); 2646 break; 2647 2648 case CEPH_CAP_OP_FLUSH_ACK: 2649 handle_cap_flush_ack(inode, h, session, cap); 2650 break; 2651 2652 case CEPH_CAP_OP_TRUNC: 2653 handle_cap_trunc(inode, h, session); 2654 break; 2655 2656 default: 2657 spin_unlock(&inode->i_lock); 2658 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op, 2659 ceph_cap_op_name(op)); 2660 } 2661 2662 done: 2663 mutex_unlock(&session->s_mutex); 2664 2665 if (check_caps) 2666 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY, NULL); 2667 if (inode) 2668 iput(inode); 2669 return; 2670 2671 bad: 2672 pr_err("ceph_handle_caps: corrupt message\n"); 2673 return; 2674 } 2675 2676 /* 2677 * Delayed work handler to process end of delayed cap release LRU list. 2678 */ 2679 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc, int flushdirty) 2680 { 2681 struct ceph_inode_info *ci; 2682 int flags = CHECK_CAPS_NODELAY; 2683 2684 if (flushdirty) 2685 flags |= CHECK_CAPS_FLUSH; 2686 2687 dout("check_delayed_caps\n"); 2688 while (1) { 2689 spin_lock(&mdsc->cap_delay_lock); 2690 if (list_empty(&mdsc->cap_delay_list)) 2691 break; 2692 ci = list_first_entry(&mdsc->cap_delay_list, 2693 struct ceph_inode_info, 2694 i_cap_delay_list); 2695 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 && 2696 time_before(jiffies, ci->i_hold_caps_max)) 2697 break; 2698 list_del_init(&ci->i_cap_delay_list); 2699 spin_unlock(&mdsc->cap_delay_lock); 2700 dout("check_delayed_caps on %p\n", &ci->vfs_inode); 2701 ceph_check_caps(ci, flags, NULL); 2702 } 2703 spin_unlock(&mdsc->cap_delay_lock); 2704 } 2705 2706 /* 2707 * Drop open file reference. If we were the last open file, 2708 * we may need to release capabilities to the MDS (or schedule 2709 * their delayed release). 2710 */ 2711 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode) 2712 { 2713 struct inode *inode = &ci->vfs_inode; 2714 int last = 0; 2715 2716 spin_lock(&inode->i_lock); 2717 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode, 2718 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1); 2719 BUG_ON(ci->i_nr_by_mode[fmode] == 0); 2720 if (--ci->i_nr_by_mode[fmode] == 0) 2721 last++; 2722 spin_unlock(&inode->i_lock); 2723 2724 if (last && ci->i_vino.snap == CEPH_NOSNAP) 2725 ceph_check_caps(ci, 0, NULL); 2726 } 2727 2728 /* 2729 * Helpers for embedding cap and dentry lease releases into mds 2730 * requests. 2731 * 2732 * @force is used by dentry_release (below) to force inclusion of a 2733 * record for the directory inode, even when there aren't any caps to 2734 * drop. 2735 */ 2736 int ceph_encode_inode_release(void **p, struct inode *inode, 2737 int mds, int drop, int unless, int force) 2738 { 2739 struct ceph_inode_info *ci = ceph_inode(inode); 2740 struct ceph_cap *cap; 2741 struct ceph_mds_request_release *rel = *p; 2742 int ret = 0; 2743 2744 dout("encode_inode_release %p mds%d drop %s unless %s\n", inode, 2745 mds, ceph_cap_string(drop), ceph_cap_string(unless)); 2746 2747 spin_lock(&inode->i_lock); 2748 cap = __get_cap_for_mds(ci, mds); 2749 if (cap && __cap_is_valid(cap)) { 2750 if (force || 2751 ((cap->issued & drop) && 2752 (cap->issued & unless) == 0)) { 2753 if ((cap->issued & drop) && 2754 (cap->issued & unless) == 0) { 2755 dout("encode_inode_release %p cap %p %s -> " 2756 "%s\n", inode, cap, 2757 ceph_cap_string(cap->issued), 2758 ceph_cap_string(cap->issued & ~drop)); 2759 cap->issued &= ~drop; 2760 cap->implemented &= ~drop; 2761 if (ci->i_ceph_flags & CEPH_I_NODELAY) { 2762 int wanted = __ceph_caps_wanted(ci); 2763 dout(" wanted %s -> %s (act %s)\n", 2764 ceph_cap_string(cap->mds_wanted), 2765 ceph_cap_string(cap->mds_wanted & 2766 ~wanted), 2767 ceph_cap_string(wanted)); 2768 cap->mds_wanted &= wanted; 2769 } 2770 } else { 2771 dout("encode_inode_release %p cap %p %s" 2772 " (force)\n", inode, cap, 2773 ceph_cap_string(cap->issued)); 2774 } 2775 2776 rel->ino = cpu_to_le64(ceph_ino(inode)); 2777 rel->cap_id = cpu_to_le64(cap->cap_id); 2778 rel->seq = cpu_to_le32(cap->seq); 2779 rel->issue_seq = cpu_to_le32(cap->issue_seq), 2780 rel->mseq = cpu_to_le32(cap->mseq); 2781 rel->caps = cpu_to_le32(cap->issued); 2782 rel->wanted = cpu_to_le32(cap->mds_wanted); 2783 rel->dname_len = 0; 2784 rel->dname_seq = 0; 2785 *p += sizeof(*rel); 2786 ret = 1; 2787 } else { 2788 dout("encode_inode_release %p cap %p %s\n", 2789 inode, cap, ceph_cap_string(cap->issued)); 2790 } 2791 } 2792 spin_unlock(&inode->i_lock); 2793 return ret; 2794 } 2795 2796 int ceph_encode_dentry_release(void **p, struct dentry *dentry, 2797 int mds, int drop, int unless) 2798 { 2799 struct inode *dir = dentry->d_parent->d_inode; 2800 struct ceph_mds_request_release *rel = *p; 2801 struct ceph_dentry_info *di = ceph_dentry(dentry); 2802 int force = 0; 2803 int ret; 2804 2805 /* 2806 * force an record for the directory caps if we have a dentry lease. 2807 * this is racy (can't take i_lock and d_lock together), but it 2808 * doesn't have to be perfect; the mds will revoke anything we don't 2809 * release. 2810 */ 2811 spin_lock(&dentry->d_lock); 2812 if (di->lease_session && di->lease_session->s_mds == mds) 2813 force = 1; 2814 spin_unlock(&dentry->d_lock); 2815 2816 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force); 2817 2818 spin_lock(&dentry->d_lock); 2819 if (ret && di->lease_session && di->lease_session->s_mds == mds) { 2820 dout("encode_dentry_release %p mds%d seq %d\n", 2821 dentry, mds, (int)di->lease_seq); 2822 rel->dname_len = cpu_to_le32(dentry->d_name.len); 2823 memcpy(*p, dentry->d_name.name, dentry->d_name.len); 2824 *p += dentry->d_name.len; 2825 rel->dname_seq = cpu_to_le32(di->lease_seq); 2826 } 2827 spin_unlock(&dentry->d_lock); 2828 return ret; 2829 } 2830