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 static int caps_min_count; /* keep at least this many (unreserved) */ 132 133 void __init ceph_caps_init(void) 134 { 135 INIT_LIST_HEAD(&caps_list); 136 spin_lock_init(&caps_list_lock); 137 } 138 139 void ceph_caps_finalize(void) 140 { 141 struct ceph_cap *cap; 142 143 spin_lock(&caps_list_lock); 144 while (!list_empty(&caps_list)) { 145 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item); 146 list_del(&cap->caps_item); 147 kmem_cache_free(ceph_cap_cachep, cap); 148 } 149 caps_total_count = 0; 150 caps_avail_count = 0; 151 caps_use_count = 0; 152 caps_reserve_count = 0; 153 caps_min_count = 0; 154 spin_unlock(&caps_list_lock); 155 } 156 157 void ceph_adjust_min_caps(int delta) 158 { 159 spin_lock(&caps_list_lock); 160 caps_min_count += delta; 161 BUG_ON(caps_min_count < 0); 162 spin_unlock(&caps_list_lock); 163 } 164 165 int ceph_reserve_caps(struct ceph_cap_reservation *ctx, int need) 166 { 167 int i; 168 struct ceph_cap *cap; 169 int have; 170 int alloc = 0; 171 LIST_HEAD(newcaps); 172 int ret = 0; 173 174 dout("reserve caps ctx=%p need=%d\n", ctx, need); 175 176 /* first reserve any caps that are already allocated */ 177 spin_lock(&caps_list_lock); 178 if (caps_avail_count >= need) 179 have = need; 180 else 181 have = caps_avail_count; 182 caps_avail_count -= have; 183 caps_reserve_count += have; 184 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 185 caps_avail_count); 186 spin_unlock(&caps_list_lock); 187 188 for (i = have; i < need; i++) { 189 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 190 if (!cap) { 191 ret = -ENOMEM; 192 goto out_alloc_count; 193 } 194 list_add(&cap->caps_item, &newcaps); 195 alloc++; 196 } 197 BUG_ON(have + alloc != need); 198 199 spin_lock(&caps_list_lock); 200 caps_total_count += alloc; 201 caps_reserve_count += alloc; 202 list_splice(&newcaps, &caps_list); 203 204 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 205 caps_avail_count); 206 spin_unlock(&caps_list_lock); 207 208 ctx->count = need; 209 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n", 210 ctx, caps_total_count, caps_use_count, caps_reserve_count, 211 caps_avail_count); 212 return 0; 213 214 out_alloc_count: 215 /* we didn't manage to reserve as much as we needed */ 216 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n", 217 ctx, need, have); 218 return ret; 219 } 220 221 int ceph_unreserve_caps(struct ceph_cap_reservation *ctx) 222 { 223 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count); 224 if (ctx->count) { 225 spin_lock(&caps_list_lock); 226 BUG_ON(caps_reserve_count < ctx->count); 227 caps_reserve_count -= ctx->count; 228 caps_avail_count += ctx->count; 229 ctx->count = 0; 230 dout("unreserve caps %d = %d used + %d resv + %d avail\n", 231 caps_total_count, caps_use_count, caps_reserve_count, 232 caps_avail_count); 233 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 234 caps_avail_count); 235 spin_unlock(&caps_list_lock); 236 } 237 return 0; 238 } 239 240 static struct ceph_cap *get_cap(struct ceph_cap_reservation *ctx) 241 { 242 struct ceph_cap *cap = NULL; 243 244 /* temporary, until we do something about cap import/export */ 245 if (!ctx) 246 return kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 247 248 spin_lock(&caps_list_lock); 249 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n", 250 ctx, ctx->count, caps_total_count, caps_use_count, 251 caps_reserve_count, caps_avail_count); 252 BUG_ON(!ctx->count); 253 BUG_ON(ctx->count > caps_reserve_count); 254 BUG_ON(list_empty(&caps_list)); 255 256 ctx->count--; 257 caps_reserve_count--; 258 caps_use_count++; 259 260 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item); 261 list_del(&cap->caps_item); 262 263 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 264 caps_avail_count); 265 spin_unlock(&caps_list_lock); 266 return cap; 267 } 268 269 void ceph_put_cap(struct ceph_cap *cap) 270 { 271 spin_lock(&caps_list_lock); 272 dout("put_cap %p %d = %d used + %d resv + %d avail\n", 273 cap, caps_total_count, caps_use_count, 274 caps_reserve_count, caps_avail_count); 275 caps_use_count--; 276 /* 277 * Keep some preallocated caps around (ceph_min_count), to 278 * avoid lots of free/alloc churn. 279 */ 280 if (caps_avail_count >= caps_reserve_count + caps_min_count) { 281 caps_total_count--; 282 kmem_cache_free(ceph_cap_cachep, cap); 283 } else { 284 caps_avail_count++; 285 list_add(&cap->caps_item, &caps_list); 286 } 287 288 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count + 289 caps_avail_count); 290 spin_unlock(&caps_list_lock); 291 } 292 293 void ceph_reservation_status(struct ceph_client *client, 294 int *total, int *avail, int *used, int *reserved, 295 int *min) 296 { 297 if (total) 298 *total = caps_total_count; 299 if (avail) 300 *avail = caps_avail_count; 301 if (used) 302 *used = caps_use_count; 303 if (reserved) 304 *reserved = caps_reserve_count; 305 if (min) 306 *min = caps_min_count; 307 } 308 309 /* 310 * Find ceph_cap for given mds, if any. 311 * 312 * Called with i_lock held. 313 */ 314 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds) 315 { 316 struct ceph_cap *cap; 317 struct rb_node *n = ci->i_caps.rb_node; 318 319 while (n) { 320 cap = rb_entry(n, struct ceph_cap, ci_node); 321 if (mds < cap->mds) 322 n = n->rb_left; 323 else if (mds > cap->mds) 324 n = n->rb_right; 325 else 326 return cap; 327 } 328 return NULL; 329 } 330 331 /* 332 * Return id of any MDS with a cap, preferably FILE_WR|WRBUFFER|EXCL, else 333 * -1. 334 */ 335 static int __ceph_get_cap_mds(struct ceph_inode_info *ci, u32 *mseq) 336 { 337 struct ceph_cap *cap; 338 int mds = -1; 339 struct rb_node *p; 340 341 /* prefer mds with WR|WRBUFFER|EXCL caps */ 342 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 343 cap = rb_entry(p, struct ceph_cap, ci_node); 344 mds = cap->mds; 345 if (mseq) 346 *mseq = cap->mseq; 347 if (cap->issued & (CEPH_CAP_FILE_WR | 348 CEPH_CAP_FILE_BUFFER | 349 CEPH_CAP_FILE_EXCL)) 350 break; 351 } 352 return mds; 353 } 354 355 int ceph_get_cap_mds(struct inode *inode) 356 { 357 int mds; 358 spin_lock(&inode->i_lock); 359 mds = __ceph_get_cap_mds(ceph_inode(inode), NULL); 360 spin_unlock(&inode->i_lock); 361 return mds; 362 } 363 364 /* 365 * Called under i_lock. 366 */ 367 static void __insert_cap_node(struct ceph_inode_info *ci, 368 struct ceph_cap *new) 369 { 370 struct rb_node **p = &ci->i_caps.rb_node; 371 struct rb_node *parent = NULL; 372 struct ceph_cap *cap = NULL; 373 374 while (*p) { 375 parent = *p; 376 cap = rb_entry(parent, struct ceph_cap, ci_node); 377 if (new->mds < cap->mds) 378 p = &(*p)->rb_left; 379 else if (new->mds > cap->mds) 380 p = &(*p)->rb_right; 381 else 382 BUG(); 383 } 384 385 rb_link_node(&new->ci_node, parent, p); 386 rb_insert_color(&new->ci_node, &ci->i_caps); 387 } 388 389 /* 390 * (re)set cap hold timeouts, which control the delayed release 391 * of unused caps back to the MDS. Should be called on cap use. 392 */ 393 static void __cap_set_timeouts(struct ceph_mds_client *mdsc, 394 struct ceph_inode_info *ci) 395 { 396 struct ceph_mount_args *ma = mdsc->client->mount_args; 397 398 ci->i_hold_caps_min = round_jiffies(jiffies + 399 ma->caps_wanted_delay_min * HZ); 400 ci->i_hold_caps_max = round_jiffies(jiffies + 401 ma->caps_wanted_delay_max * HZ); 402 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode, 403 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies); 404 } 405 406 /* 407 * (Re)queue cap at the end of the delayed cap release list. 408 * 409 * If I_FLUSH is set, leave the inode at the front of the list. 410 * 411 * Caller holds i_lock 412 * -> we take mdsc->cap_delay_lock 413 */ 414 static void __cap_delay_requeue(struct ceph_mds_client *mdsc, 415 struct ceph_inode_info *ci) 416 { 417 __cap_set_timeouts(mdsc, ci); 418 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode, 419 ci->i_ceph_flags, ci->i_hold_caps_max); 420 if (!mdsc->stopping) { 421 spin_lock(&mdsc->cap_delay_lock); 422 if (!list_empty(&ci->i_cap_delay_list)) { 423 if (ci->i_ceph_flags & CEPH_I_FLUSH) 424 goto no_change; 425 list_del_init(&ci->i_cap_delay_list); 426 } 427 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 428 no_change: 429 spin_unlock(&mdsc->cap_delay_lock); 430 } 431 } 432 433 /* 434 * Queue an inode for immediate writeback. Mark inode with I_FLUSH, 435 * indicating we should send a cap message to flush dirty metadata 436 * asap, and move to the front of the delayed cap list. 437 */ 438 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc, 439 struct ceph_inode_info *ci) 440 { 441 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode); 442 spin_lock(&mdsc->cap_delay_lock); 443 ci->i_ceph_flags |= CEPH_I_FLUSH; 444 if (!list_empty(&ci->i_cap_delay_list)) 445 list_del_init(&ci->i_cap_delay_list); 446 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 447 spin_unlock(&mdsc->cap_delay_lock); 448 } 449 450 /* 451 * Cancel delayed work on cap. 452 * 453 * Caller must hold i_lock. 454 */ 455 static void __cap_delay_cancel(struct ceph_mds_client *mdsc, 456 struct ceph_inode_info *ci) 457 { 458 dout("__cap_delay_cancel %p\n", &ci->vfs_inode); 459 if (list_empty(&ci->i_cap_delay_list)) 460 return; 461 spin_lock(&mdsc->cap_delay_lock); 462 list_del_init(&ci->i_cap_delay_list); 463 spin_unlock(&mdsc->cap_delay_lock); 464 } 465 466 /* 467 * Common issue checks for add_cap, handle_cap_grant. 468 */ 469 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap, 470 unsigned issued) 471 { 472 unsigned had = __ceph_caps_issued(ci, NULL); 473 474 /* 475 * Each time we receive FILE_CACHE anew, we increment 476 * i_rdcache_gen. 477 */ 478 if ((issued & CEPH_CAP_FILE_CACHE) && 479 (had & CEPH_CAP_FILE_CACHE) == 0) 480 ci->i_rdcache_gen++; 481 482 /* 483 * if we are newly issued FILE_SHARED, clear I_COMPLETE; we 484 * don't know what happened to this directory while we didn't 485 * have the cap. 486 */ 487 if ((issued & CEPH_CAP_FILE_SHARED) && 488 (had & CEPH_CAP_FILE_SHARED) == 0) { 489 ci->i_shared_gen++; 490 if (S_ISDIR(ci->vfs_inode.i_mode)) { 491 dout(" marking %p NOT complete\n", &ci->vfs_inode); 492 ci->i_ceph_flags &= ~CEPH_I_COMPLETE; 493 } 494 } 495 } 496 497 /* 498 * Add a capability under the given MDS session. 499 * 500 * Caller should hold session snap_rwsem (read) and s_mutex. 501 * 502 * @fmode is the open file mode, if we are opening a file, otherwise 503 * it is < 0. (This is so we can atomically add the cap and add an 504 * open file reference to it.) 505 */ 506 int ceph_add_cap(struct inode *inode, 507 struct ceph_mds_session *session, u64 cap_id, 508 int fmode, unsigned issued, unsigned wanted, 509 unsigned seq, unsigned mseq, u64 realmino, int flags, 510 struct ceph_cap_reservation *caps_reservation) 511 { 512 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 513 struct ceph_inode_info *ci = ceph_inode(inode); 514 struct ceph_cap *new_cap = NULL; 515 struct ceph_cap *cap; 516 int mds = session->s_mds; 517 int actual_wanted; 518 519 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode, 520 session->s_mds, cap_id, ceph_cap_string(issued), seq); 521 522 /* 523 * If we are opening the file, include file mode wanted bits 524 * in wanted. 525 */ 526 if (fmode >= 0) 527 wanted |= ceph_caps_for_mode(fmode); 528 529 retry: 530 spin_lock(&inode->i_lock); 531 cap = __get_cap_for_mds(ci, mds); 532 if (!cap) { 533 if (new_cap) { 534 cap = new_cap; 535 new_cap = NULL; 536 } else { 537 spin_unlock(&inode->i_lock); 538 new_cap = get_cap(caps_reservation); 539 if (new_cap == NULL) 540 return -ENOMEM; 541 goto retry; 542 } 543 544 cap->issued = 0; 545 cap->implemented = 0; 546 cap->mds = mds; 547 cap->mds_wanted = 0; 548 549 cap->ci = ci; 550 __insert_cap_node(ci, cap); 551 552 /* clear out old exporting info? (i.e. on cap import) */ 553 if (ci->i_cap_exporting_mds == mds) { 554 ci->i_cap_exporting_issued = 0; 555 ci->i_cap_exporting_mseq = 0; 556 ci->i_cap_exporting_mds = -1; 557 } 558 559 /* add to session cap list */ 560 cap->session = session; 561 spin_lock(&session->s_cap_lock); 562 list_add_tail(&cap->session_caps, &session->s_caps); 563 session->s_nr_caps++; 564 spin_unlock(&session->s_cap_lock); 565 } 566 567 if (!ci->i_snap_realm) { 568 /* 569 * add this inode to the appropriate snap realm 570 */ 571 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc, 572 realmino); 573 if (realm) { 574 ceph_get_snap_realm(mdsc, realm); 575 spin_lock(&realm->inodes_with_caps_lock); 576 ci->i_snap_realm = realm; 577 list_add(&ci->i_snap_realm_item, 578 &realm->inodes_with_caps); 579 spin_unlock(&realm->inodes_with_caps_lock); 580 } else { 581 pr_err("ceph_add_cap: couldn't find snap realm %llx\n", 582 realmino); 583 } 584 } 585 586 __check_cap_issue(ci, cap, issued); 587 588 /* 589 * If we are issued caps we don't want, or the mds' wanted 590 * value appears to be off, queue a check so we'll release 591 * later and/or update the mds wanted value. 592 */ 593 actual_wanted = __ceph_caps_wanted(ci); 594 if ((wanted & ~actual_wanted) || 595 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) { 596 dout(" issued %s, mds wanted %s, actual %s, queueing\n", 597 ceph_cap_string(issued), ceph_cap_string(wanted), 598 ceph_cap_string(actual_wanted)); 599 __cap_delay_requeue(mdsc, ci); 600 } 601 602 if (flags & CEPH_CAP_FLAG_AUTH) 603 ci->i_auth_cap = cap; 604 else if (ci->i_auth_cap == cap) 605 ci->i_auth_cap = NULL; 606 607 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n", 608 inode, ceph_vinop(inode), cap, ceph_cap_string(issued), 609 ceph_cap_string(issued|cap->issued), seq, mds); 610 cap->cap_id = cap_id; 611 cap->issued = issued; 612 cap->implemented |= issued; 613 cap->mds_wanted |= wanted; 614 cap->seq = seq; 615 cap->issue_seq = seq; 616 cap->mseq = mseq; 617 cap->cap_gen = session->s_cap_gen; 618 619 if (fmode >= 0) 620 __ceph_get_fmode(ci, fmode); 621 spin_unlock(&inode->i_lock); 622 wake_up(&ci->i_cap_wq); 623 return 0; 624 } 625 626 /* 627 * Return true if cap has not timed out and belongs to the current 628 * generation of the MDS session (i.e. has not gone 'stale' due to 629 * us losing touch with the mds). 630 */ 631 static int __cap_is_valid(struct ceph_cap *cap) 632 { 633 unsigned long ttl; 634 u32 gen; 635 636 spin_lock(&cap->session->s_cap_lock); 637 gen = cap->session->s_cap_gen; 638 ttl = cap->session->s_cap_ttl; 639 spin_unlock(&cap->session->s_cap_lock); 640 641 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) { 642 dout("__cap_is_valid %p cap %p issued %s " 643 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode, 644 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen); 645 return 0; 646 } 647 648 return 1; 649 } 650 651 /* 652 * Return set of valid cap bits issued to us. Note that caps time 653 * out, and may be invalidated in bulk if the client session times out 654 * and session->s_cap_gen is bumped. 655 */ 656 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented) 657 { 658 int have = ci->i_snap_caps | ci->i_cap_exporting_issued; 659 struct ceph_cap *cap; 660 struct rb_node *p; 661 662 if (implemented) 663 *implemented = 0; 664 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 665 cap = rb_entry(p, struct ceph_cap, ci_node); 666 if (!__cap_is_valid(cap)) 667 continue; 668 dout("__ceph_caps_issued %p cap %p issued %s\n", 669 &ci->vfs_inode, cap, ceph_cap_string(cap->issued)); 670 have |= cap->issued; 671 if (implemented) 672 *implemented |= cap->implemented; 673 } 674 return have; 675 } 676 677 /* 678 * Get cap bits issued by caps other than @ocap 679 */ 680 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap) 681 { 682 int have = ci->i_snap_caps; 683 struct ceph_cap *cap; 684 struct rb_node *p; 685 686 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 687 cap = rb_entry(p, struct ceph_cap, ci_node); 688 if (cap == ocap) 689 continue; 690 if (!__cap_is_valid(cap)) 691 continue; 692 have |= cap->issued; 693 } 694 return have; 695 } 696 697 /* 698 * Move a cap to the end of the LRU (oldest caps at list head, newest 699 * at list tail). 700 */ 701 static void __touch_cap(struct ceph_cap *cap) 702 { 703 struct ceph_mds_session *s = cap->session; 704 705 spin_lock(&s->s_cap_lock); 706 if (s->s_cap_iterator == NULL) { 707 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap, 708 s->s_mds); 709 list_move_tail(&cap->session_caps, &s->s_caps); 710 } else { 711 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n", 712 &cap->ci->vfs_inode, cap, s->s_mds); 713 } 714 spin_unlock(&s->s_cap_lock); 715 } 716 717 /* 718 * Check if we hold the given mask. If so, move the cap(s) to the 719 * front of their respective LRUs. (This is the preferred way for 720 * callers to check for caps they want.) 721 */ 722 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch) 723 { 724 struct ceph_cap *cap; 725 struct rb_node *p; 726 int have = ci->i_snap_caps; 727 728 if ((have & mask) == mask) { 729 dout("__ceph_caps_issued_mask %p snap issued %s" 730 " (mask %s)\n", &ci->vfs_inode, 731 ceph_cap_string(have), 732 ceph_cap_string(mask)); 733 return 1; 734 } 735 736 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 737 cap = rb_entry(p, struct ceph_cap, ci_node); 738 if (!__cap_is_valid(cap)) 739 continue; 740 if ((cap->issued & mask) == mask) { 741 dout("__ceph_caps_issued_mask %p cap %p issued %s" 742 " (mask %s)\n", &ci->vfs_inode, cap, 743 ceph_cap_string(cap->issued), 744 ceph_cap_string(mask)); 745 if (touch) 746 __touch_cap(cap); 747 return 1; 748 } 749 750 /* does a combination of caps satisfy mask? */ 751 have |= cap->issued; 752 if ((have & mask) == mask) { 753 dout("__ceph_caps_issued_mask %p combo issued %s" 754 " (mask %s)\n", &ci->vfs_inode, 755 ceph_cap_string(cap->issued), 756 ceph_cap_string(mask)); 757 if (touch) { 758 struct rb_node *q; 759 760 /* touch this + preceeding caps */ 761 __touch_cap(cap); 762 for (q = rb_first(&ci->i_caps); q != p; 763 q = rb_next(q)) { 764 cap = rb_entry(q, struct ceph_cap, 765 ci_node); 766 if (!__cap_is_valid(cap)) 767 continue; 768 __touch_cap(cap); 769 } 770 } 771 return 1; 772 } 773 } 774 775 return 0; 776 } 777 778 /* 779 * Return true if mask caps are currently being revoked by an MDS. 780 */ 781 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask) 782 { 783 struct inode *inode = &ci->vfs_inode; 784 struct ceph_cap *cap; 785 struct rb_node *p; 786 int ret = 0; 787 788 spin_lock(&inode->i_lock); 789 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 790 cap = rb_entry(p, struct ceph_cap, ci_node); 791 if (__cap_is_valid(cap) && 792 (cap->implemented & ~cap->issued & mask)) { 793 ret = 1; 794 break; 795 } 796 } 797 spin_unlock(&inode->i_lock); 798 dout("ceph_caps_revoking %p %s = %d\n", inode, 799 ceph_cap_string(mask), ret); 800 return ret; 801 } 802 803 int __ceph_caps_used(struct ceph_inode_info *ci) 804 { 805 int used = 0; 806 if (ci->i_pin_ref) 807 used |= CEPH_CAP_PIN; 808 if (ci->i_rd_ref) 809 used |= CEPH_CAP_FILE_RD; 810 if (ci->i_rdcache_ref || ci->i_rdcache_gen) 811 used |= CEPH_CAP_FILE_CACHE; 812 if (ci->i_wr_ref) 813 used |= CEPH_CAP_FILE_WR; 814 if (ci->i_wrbuffer_ref) 815 used |= CEPH_CAP_FILE_BUFFER; 816 return used; 817 } 818 819 /* 820 * wanted, by virtue of open file modes 821 */ 822 int __ceph_caps_file_wanted(struct ceph_inode_info *ci) 823 { 824 int want = 0; 825 int mode; 826 for (mode = 0; mode < 4; mode++) 827 if (ci->i_nr_by_mode[mode]) 828 want |= ceph_caps_for_mode(mode); 829 return want; 830 } 831 832 /* 833 * Return caps we have registered with the MDS(s) as 'wanted'. 834 */ 835 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci) 836 { 837 struct ceph_cap *cap; 838 struct rb_node *p; 839 int mds_wanted = 0; 840 841 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 842 cap = rb_entry(p, struct ceph_cap, ci_node); 843 if (!__cap_is_valid(cap)) 844 continue; 845 mds_wanted |= cap->mds_wanted; 846 } 847 return mds_wanted; 848 } 849 850 /* 851 * called under i_lock 852 */ 853 static int __ceph_is_any_caps(struct ceph_inode_info *ci) 854 { 855 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0; 856 } 857 858 /* 859 * caller should hold i_lock. 860 * caller will not hold session s_mutex if called from destroy_inode. 861 */ 862 void __ceph_remove_cap(struct ceph_cap *cap) 863 { 864 struct ceph_mds_session *session = cap->session; 865 struct ceph_inode_info *ci = cap->ci; 866 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc; 867 868 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode); 869 870 /* remove from inode list */ 871 rb_erase(&cap->ci_node, &ci->i_caps); 872 cap->ci = NULL; 873 if (ci->i_auth_cap == cap) 874 ci->i_auth_cap = NULL; 875 876 /* remove from session list */ 877 spin_lock(&session->s_cap_lock); 878 if (session->s_cap_iterator == cap) { 879 /* not yet, we are iterating over this very cap */ 880 dout("__ceph_remove_cap delaying %p removal from session %p\n", 881 cap, cap->session); 882 } else { 883 list_del_init(&cap->session_caps); 884 session->s_nr_caps--; 885 cap->session = NULL; 886 } 887 spin_unlock(&session->s_cap_lock); 888 889 if (cap->session == NULL) 890 ceph_put_cap(cap); 891 892 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) { 893 struct ceph_snap_realm *realm = ci->i_snap_realm; 894 spin_lock(&realm->inodes_with_caps_lock); 895 list_del_init(&ci->i_snap_realm_item); 896 ci->i_snap_realm_counter++; 897 ci->i_snap_realm = NULL; 898 spin_unlock(&realm->inodes_with_caps_lock); 899 ceph_put_snap_realm(mdsc, realm); 900 } 901 if (!__ceph_is_any_real_caps(ci)) 902 __cap_delay_cancel(mdsc, ci); 903 } 904 905 /* 906 * Build and send a cap message to the given MDS. 907 * 908 * Caller should be holding s_mutex. 909 */ 910 static int send_cap_msg(struct ceph_mds_session *session, 911 u64 ino, u64 cid, int op, 912 int caps, int wanted, int dirty, 913 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq, 914 u64 size, u64 max_size, 915 struct timespec *mtime, struct timespec *atime, 916 u64 time_warp_seq, 917 uid_t uid, gid_t gid, mode_t mode, 918 u64 xattr_version, 919 struct ceph_buffer *xattrs_buf, 920 u64 follows) 921 { 922 struct ceph_mds_caps *fc; 923 struct ceph_msg *msg; 924 925 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s" 926 " seq %u/%u mseq %u follows %lld size %llu/%llu" 927 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op), 928 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted), 929 ceph_cap_string(dirty), 930 seq, issue_seq, mseq, follows, size, max_size, 931 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0); 932 933 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL); 934 if (IS_ERR(msg)) 935 return PTR_ERR(msg); 936 937 msg->hdr.tid = cpu_to_le64(flush_tid); 938 939 fc = msg->front.iov_base; 940 memset(fc, 0, sizeof(*fc)); 941 942 fc->cap_id = cpu_to_le64(cid); 943 fc->op = cpu_to_le32(op); 944 fc->seq = cpu_to_le32(seq); 945 fc->issue_seq = cpu_to_le32(issue_seq); 946 fc->migrate_seq = cpu_to_le32(mseq); 947 fc->caps = cpu_to_le32(caps); 948 fc->wanted = cpu_to_le32(wanted); 949 fc->dirty = cpu_to_le32(dirty); 950 fc->ino = cpu_to_le64(ino); 951 fc->snap_follows = cpu_to_le64(follows); 952 953 fc->size = cpu_to_le64(size); 954 fc->max_size = cpu_to_le64(max_size); 955 if (mtime) 956 ceph_encode_timespec(&fc->mtime, mtime); 957 if (atime) 958 ceph_encode_timespec(&fc->atime, atime); 959 fc->time_warp_seq = cpu_to_le32(time_warp_seq); 960 961 fc->uid = cpu_to_le32(uid); 962 fc->gid = cpu_to_le32(gid); 963 fc->mode = cpu_to_le32(mode); 964 965 fc->xattr_version = cpu_to_le64(xattr_version); 966 if (xattrs_buf) { 967 msg->middle = ceph_buffer_get(xattrs_buf); 968 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len); 969 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len); 970 } 971 972 ceph_con_send(&session->s_con, msg); 973 return 0; 974 } 975 976 /* 977 * Queue cap releases when an inode is dropped from our cache. Since 978 * inode is about to be destroyed, there is no need for i_lock. 979 */ 980 void ceph_queue_caps_release(struct inode *inode) 981 { 982 struct ceph_inode_info *ci = ceph_inode(inode); 983 struct rb_node *p; 984 985 p = rb_first(&ci->i_caps); 986 while (p) { 987 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node); 988 struct ceph_mds_session *session = cap->session; 989 struct ceph_msg *msg; 990 struct ceph_mds_cap_release *head; 991 struct ceph_mds_cap_item *item; 992 993 spin_lock(&session->s_cap_lock); 994 BUG_ON(!session->s_num_cap_releases); 995 msg = list_first_entry(&session->s_cap_releases, 996 struct ceph_msg, list_head); 997 998 dout(" adding %p release to mds%d msg %p (%d left)\n", 999 inode, session->s_mds, msg, session->s_num_cap_releases); 1000 1001 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE); 1002 head = msg->front.iov_base; 1003 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1); 1004 item = msg->front.iov_base + msg->front.iov_len; 1005 item->ino = cpu_to_le64(ceph_ino(inode)); 1006 item->cap_id = cpu_to_le64(cap->cap_id); 1007 item->migrate_seq = cpu_to_le32(cap->mseq); 1008 item->seq = cpu_to_le32(cap->issue_seq); 1009 1010 session->s_num_cap_releases--; 1011 1012 msg->front.iov_len += sizeof(*item); 1013 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) { 1014 dout(" release msg %p full\n", msg); 1015 list_move_tail(&msg->list_head, 1016 &session->s_cap_releases_done); 1017 } else { 1018 dout(" release msg %p at %d/%d (%d)\n", msg, 1019 (int)le32_to_cpu(head->num), 1020 (int)CEPH_CAPS_PER_RELEASE, 1021 (int)msg->front.iov_len); 1022 } 1023 spin_unlock(&session->s_cap_lock); 1024 p = rb_next(p); 1025 __ceph_remove_cap(cap); 1026 } 1027 } 1028 1029 /* 1030 * Send a cap msg on the given inode. Update our caps state, then 1031 * drop i_lock and send the message. 1032 * 1033 * Make note of max_size reported/requested from mds, revoked caps 1034 * that have now been implemented. 1035 * 1036 * Make half-hearted attempt ot to invalidate page cache if we are 1037 * dropping RDCACHE. Note that this will leave behind locked pages 1038 * that we'll then need to deal with elsewhere. 1039 * 1040 * Return non-zero if delayed release, or we experienced an error 1041 * such that the caller should requeue + retry later. 1042 * 1043 * called with i_lock, then drops it. 1044 * caller should hold snap_rwsem (read), s_mutex. 1045 */ 1046 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap, 1047 int op, int used, int want, int retain, int flushing, 1048 unsigned *pflush_tid) 1049 __releases(cap->ci->vfs_inode->i_lock) 1050 { 1051 struct ceph_inode_info *ci = cap->ci; 1052 struct inode *inode = &ci->vfs_inode; 1053 u64 cap_id = cap->cap_id; 1054 int held, revoking, dropping, keep; 1055 u64 seq, issue_seq, mseq, time_warp_seq, follows; 1056 u64 size, max_size; 1057 struct timespec mtime, atime; 1058 int wake = 0; 1059 mode_t mode; 1060 uid_t uid; 1061 gid_t gid; 1062 struct ceph_mds_session *session; 1063 u64 xattr_version = 0; 1064 int delayed = 0; 1065 u64 flush_tid = 0; 1066 int i; 1067 int ret; 1068 1069 held = cap->issued | cap->implemented; 1070 revoking = cap->implemented & ~cap->issued; 1071 retain &= ~revoking; 1072 dropping = cap->issued & ~retain; 1073 1074 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n", 1075 inode, cap, cap->session, 1076 ceph_cap_string(held), ceph_cap_string(held & retain), 1077 ceph_cap_string(revoking)); 1078 BUG_ON((retain & CEPH_CAP_PIN) == 0); 1079 1080 session = cap->session; 1081 1082 /* don't release wanted unless we've waited a bit. */ 1083 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1084 time_before(jiffies, ci->i_hold_caps_min)) { 1085 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n", 1086 ceph_cap_string(cap->issued), 1087 ceph_cap_string(cap->issued & retain), 1088 ceph_cap_string(cap->mds_wanted), 1089 ceph_cap_string(want)); 1090 want |= cap->mds_wanted; 1091 retain |= cap->issued; 1092 delayed = 1; 1093 } 1094 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH); 1095 1096 cap->issued &= retain; /* drop bits we don't want */ 1097 if (cap->implemented & ~cap->issued) { 1098 /* 1099 * Wake up any waiters on wanted -> needed transition. 1100 * This is due to the weird transition from buffered 1101 * to sync IO... we need to flush dirty pages _before_ 1102 * allowing sync writes to avoid reordering. 1103 */ 1104 wake = 1; 1105 } 1106 cap->implemented &= cap->issued | used; 1107 cap->mds_wanted = want; 1108 1109 if (flushing) { 1110 /* 1111 * assign a tid for flush operations so we can avoid 1112 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark 1113 * clean type races. track latest tid for every bit 1114 * so we can handle flush AxFw, flush Fw, and have the 1115 * first ack clean Ax. 1116 */ 1117 flush_tid = ++ci->i_cap_flush_last_tid; 1118 if (pflush_tid) 1119 *pflush_tid = flush_tid; 1120 dout(" cap_flush_tid %d\n", (int)flush_tid); 1121 for (i = 0; i < CEPH_CAP_BITS; i++) 1122 if (flushing & (1 << i)) 1123 ci->i_cap_flush_tid[i] = flush_tid; 1124 } 1125 1126 keep = cap->implemented; 1127 seq = cap->seq; 1128 issue_seq = cap->issue_seq; 1129 mseq = cap->mseq; 1130 size = inode->i_size; 1131 ci->i_reported_size = size; 1132 max_size = ci->i_wanted_max_size; 1133 ci->i_requested_max_size = max_size; 1134 mtime = inode->i_mtime; 1135 atime = inode->i_atime; 1136 time_warp_seq = ci->i_time_warp_seq; 1137 follows = ci->i_snap_realm->cached_context->seq; 1138 uid = inode->i_uid; 1139 gid = inode->i_gid; 1140 mode = inode->i_mode; 1141 1142 if (dropping & CEPH_CAP_XATTR_EXCL) { 1143 __ceph_build_xattrs_blob(ci); 1144 xattr_version = ci->i_xattrs.version + 1; 1145 } 1146 1147 spin_unlock(&inode->i_lock); 1148 1149 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id, 1150 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq, 1151 size, max_size, &mtime, &atime, time_warp_seq, 1152 uid, gid, mode, 1153 xattr_version, 1154 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL, 1155 follows); 1156 if (ret < 0) { 1157 dout("error sending cap msg, must requeue %p\n", inode); 1158 delayed = 1; 1159 } 1160 1161 if (wake) 1162 wake_up(&ci->i_cap_wq); 1163 1164 return delayed; 1165 } 1166 1167 /* 1168 * When a snapshot is taken, clients accumulate dirty metadata on 1169 * inodes with capabilities in ceph_cap_snaps to describe the file 1170 * state at the time the snapshot was taken. This must be flushed 1171 * asynchronously back to the MDS once sync writes complete and dirty 1172 * data is written out. 1173 * 1174 * Called under i_lock. Takes s_mutex as needed. 1175 */ 1176 void __ceph_flush_snaps(struct ceph_inode_info *ci, 1177 struct ceph_mds_session **psession) 1178 { 1179 struct inode *inode = &ci->vfs_inode; 1180 int mds; 1181 struct ceph_cap_snap *capsnap; 1182 u32 mseq; 1183 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 1184 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold 1185 session->s_mutex */ 1186 u64 next_follows = 0; /* keep track of how far we've gotten through the 1187 i_cap_snaps list, and skip these entries next time 1188 around to avoid an infinite loop */ 1189 1190 if (psession) 1191 session = *psession; 1192 1193 dout("__flush_snaps %p\n", inode); 1194 retry: 1195 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 1196 /* avoid an infiniute loop after retry */ 1197 if (capsnap->follows < next_follows) 1198 continue; 1199 /* 1200 * we need to wait for sync writes to complete and for dirty 1201 * pages to be written out. 1202 */ 1203 if (capsnap->dirty_pages || capsnap->writing) 1204 continue; 1205 1206 /* pick mds, take s_mutex */ 1207 mds = __ceph_get_cap_mds(ci, &mseq); 1208 if (session && session->s_mds != mds) { 1209 dout("oops, wrong session %p mutex\n", session); 1210 mutex_unlock(&session->s_mutex); 1211 ceph_put_mds_session(session); 1212 session = NULL; 1213 } 1214 if (!session) { 1215 spin_unlock(&inode->i_lock); 1216 mutex_lock(&mdsc->mutex); 1217 session = __ceph_lookup_mds_session(mdsc, mds); 1218 mutex_unlock(&mdsc->mutex); 1219 if (session) { 1220 dout("inverting session/ino locks on %p\n", 1221 session); 1222 mutex_lock(&session->s_mutex); 1223 } 1224 /* 1225 * if session == NULL, we raced against a cap 1226 * deletion. retry, and we'll get a better 1227 * @mds value next time. 1228 */ 1229 spin_lock(&inode->i_lock); 1230 goto retry; 1231 } 1232 1233 capsnap->flush_tid = ++ci->i_cap_flush_last_tid; 1234 atomic_inc(&capsnap->nref); 1235 if (!list_empty(&capsnap->flushing_item)) 1236 list_del_init(&capsnap->flushing_item); 1237 list_add_tail(&capsnap->flushing_item, 1238 &session->s_cap_snaps_flushing); 1239 spin_unlock(&inode->i_lock); 1240 1241 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n", 1242 inode, capsnap, next_follows, capsnap->size); 1243 send_cap_msg(session, ceph_vino(inode).ino, 0, 1244 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0, 1245 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq, 1246 capsnap->size, 0, 1247 &capsnap->mtime, &capsnap->atime, 1248 capsnap->time_warp_seq, 1249 capsnap->uid, capsnap->gid, capsnap->mode, 1250 0, NULL, 1251 capsnap->follows); 1252 1253 next_follows = capsnap->follows + 1; 1254 ceph_put_cap_snap(capsnap); 1255 1256 spin_lock(&inode->i_lock); 1257 goto retry; 1258 } 1259 1260 /* we flushed them all; remove this inode from the queue */ 1261 spin_lock(&mdsc->snap_flush_lock); 1262 list_del_init(&ci->i_snap_flush_item); 1263 spin_unlock(&mdsc->snap_flush_lock); 1264 1265 if (psession) 1266 *psession = session; 1267 else if (session) { 1268 mutex_unlock(&session->s_mutex); 1269 ceph_put_mds_session(session); 1270 } 1271 } 1272 1273 static void ceph_flush_snaps(struct ceph_inode_info *ci) 1274 { 1275 struct inode *inode = &ci->vfs_inode; 1276 1277 spin_lock(&inode->i_lock); 1278 __ceph_flush_snaps(ci, NULL); 1279 spin_unlock(&inode->i_lock); 1280 } 1281 1282 /* 1283 * Mark caps dirty. If inode is newly dirty, add to the global dirty 1284 * list. 1285 */ 1286 void __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask) 1287 { 1288 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc; 1289 struct inode *inode = &ci->vfs_inode; 1290 int was = ci->i_dirty_caps; 1291 int dirty = 0; 1292 1293 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode, 1294 ceph_cap_string(mask), ceph_cap_string(was), 1295 ceph_cap_string(was | mask)); 1296 ci->i_dirty_caps |= mask; 1297 if (was == 0) { 1298 dout(" inode %p now dirty\n", &ci->vfs_inode); 1299 BUG_ON(!list_empty(&ci->i_dirty_item)); 1300 spin_lock(&mdsc->cap_dirty_lock); 1301 list_add(&ci->i_dirty_item, &mdsc->cap_dirty); 1302 spin_unlock(&mdsc->cap_dirty_lock); 1303 if (ci->i_flushing_caps == 0) { 1304 igrab(inode); 1305 dirty |= I_DIRTY_SYNC; 1306 } 1307 } 1308 BUG_ON(list_empty(&ci->i_dirty_item)); 1309 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) && 1310 (mask & CEPH_CAP_FILE_BUFFER)) 1311 dirty |= I_DIRTY_DATASYNC; 1312 if (dirty) 1313 __mark_inode_dirty(inode, dirty); 1314 __cap_delay_requeue(mdsc, ci); 1315 } 1316 1317 /* 1318 * Add dirty inode to the flushing list. Assigned a seq number so we 1319 * can wait for caps to flush without starving. 1320 * 1321 * Called under i_lock. 1322 */ 1323 static int __mark_caps_flushing(struct inode *inode, 1324 struct ceph_mds_session *session) 1325 { 1326 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1327 struct ceph_inode_info *ci = ceph_inode(inode); 1328 int flushing; 1329 1330 BUG_ON(ci->i_dirty_caps == 0); 1331 BUG_ON(list_empty(&ci->i_dirty_item)); 1332 1333 flushing = ci->i_dirty_caps; 1334 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", 1335 ceph_cap_string(flushing), 1336 ceph_cap_string(ci->i_flushing_caps), 1337 ceph_cap_string(ci->i_flushing_caps | flushing)); 1338 ci->i_flushing_caps |= flushing; 1339 ci->i_dirty_caps = 0; 1340 dout(" inode %p now !dirty\n", inode); 1341 1342 spin_lock(&mdsc->cap_dirty_lock); 1343 list_del_init(&ci->i_dirty_item); 1344 1345 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq; 1346 if (list_empty(&ci->i_flushing_item)) { 1347 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1348 mdsc->num_cap_flushing++; 1349 dout(" inode %p now flushing seq %lld\n", inode, 1350 ci->i_cap_flush_seq); 1351 } else { 1352 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1353 dout(" inode %p now flushing (more) seq %lld\n", inode, 1354 ci->i_cap_flush_seq); 1355 } 1356 spin_unlock(&mdsc->cap_dirty_lock); 1357 1358 return flushing; 1359 } 1360 1361 /* 1362 * try to invalidate mapping pages without blocking. 1363 */ 1364 static int mapping_is_empty(struct address_space *mapping) 1365 { 1366 struct page *page = find_get_page(mapping, 0); 1367 1368 if (!page) 1369 return 1; 1370 1371 put_page(page); 1372 return 0; 1373 } 1374 1375 static int try_nonblocking_invalidate(struct inode *inode) 1376 { 1377 struct ceph_inode_info *ci = ceph_inode(inode); 1378 u32 invalidating_gen = ci->i_rdcache_gen; 1379 1380 spin_unlock(&inode->i_lock); 1381 invalidate_mapping_pages(&inode->i_data, 0, -1); 1382 spin_lock(&inode->i_lock); 1383 1384 if (mapping_is_empty(&inode->i_data) && 1385 invalidating_gen == ci->i_rdcache_gen) { 1386 /* success. */ 1387 dout("try_nonblocking_invalidate %p success\n", inode); 1388 ci->i_rdcache_gen = 0; 1389 ci->i_rdcache_revoking = 0; 1390 return 0; 1391 } 1392 dout("try_nonblocking_invalidate %p failed\n", inode); 1393 return -1; 1394 } 1395 1396 /* 1397 * Swiss army knife function to examine currently used and wanted 1398 * versus held caps. Release, flush, ack revoked caps to mds as 1399 * appropriate. 1400 * 1401 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay 1402 * cap release further. 1403 * CHECK_CAPS_AUTHONLY - we should only check the auth cap 1404 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without 1405 * further delay. 1406 */ 1407 void ceph_check_caps(struct ceph_inode_info *ci, int flags, 1408 struct ceph_mds_session *session) 1409 { 1410 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode); 1411 struct ceph_mds_client *mdsc = &client->mdsc; 1412 struct inode *inode = &ci->vfs_inode; 1413 struct ceph_cap *cap; 1414 int file_wanted, used; 1415 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ 1416 int drop_session_lock = session ? 0 : 1; 1417 int issued, implemented, want, retain, revoking, flushing = 0; 1418 int mds = -1; /* keep track of how far we've gone through i_caps list 1419 to avoid an infinite loop on retry */ 1420 struct rb_node *p; 1421 int tried_invalidate = 0; 1422 int delayed = 0, sent = 0, force_requeue = 0, num; 1423 int queue_invalidate = 0; 1424 int is_delayed = flags & CHECK_CAPS_NODELAY; 1425 1426 /* if we are unmounting, flush any unused caps immediately. */ 1427 if (mdsc->stopping) 1428 is_delayed = 1; 1429 1430 spin_lock(&inode->i_lock); 1431 1432 if (ci->i_ceph_flags & CEPH_I_FLUSH) 1433 flags |= CHECK_CAPS_FLUSH; 1434 1435 /* flush snaps first time around only */ 1436 if (!list_empty(&ci->i_cap_snaps)) 1437 __ceph_flush_snaps(ci, &session); 1438 goto retry_locked; 1439 retry: 1440 spin_lock(&inode->i_lock); 1441 retry_locked: 1442 file_wanted = __ceph_caps_file_wanted(ci); 1443 used = __ceph_caps_used(ci); 1444 want = file_wanted | used; 1445 issued = __ceph_caps_issued(ci, &implemented); 1446 revoking = implemented & ~issued; 1447 1448 retain = want | CEPH_CAP_PIN; 1449 if (!mdsc->stopping && inode->i_nlink > 0) { 1450 if (want) { 1451 retain |= CEPH_CAP_ANY; /* be greedy */ 1452 } else { 1453 retain |= CEPH_CAP_ANY_SHARED; 1454 /* 1455 * keep RD only if we didn't have the file open RW, 1456 * because then the mds would revoke it anyway to 1457 * journal max_size=0. 1458 */ 1459 if (ci->i_max_size == 0) 1460 retain |= CEPH_CAP_ANY_RD; 1461 } 1462 } 1463 1464 dout("check_caps %p file_want %s used %s dirty %s flushing %s" 1465 " issued %s revoking %s retain %s %s%s%s\n", inode, 1466 ceph_cap_string(file_wanted), 1467 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps), 1468 ceph_cap_string(ci->i_flushing_caps), 1469 ceph_cap_string(issued), ceph_cap_string(revoking), 1470 ceph_cap_string(retain), 1471 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "", 1472 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "", 1473 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : ""); 1474 1475 /* 1476 * If we no longer need to hold onto old our caps, and we may 1477 * have cached pages, but don't want them, then try to invalidate. 1478 * If we fail, it's because pages are locked.... try again later. 1479 */ 1480 if ((!is_delayed || mdsc->stopping) && 1481 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */ 1482 ci->i_rdcache_gen && /* may have cached pages */ 1483 (file_wanted == 0 || /* no open files */ 1484 (revoking & CEPH_CAP_FILE_CACHE)) && /* or revoking cache */ 1485 !tried_invalidate) { 1486 dout("check_caps trying to invalidate on %p\n", inode); 1487 if (try_nonblocking_invalidate(inode) < 0) { 1488 if (revoking & CEPH_CAP_FILE_CACHE) { 1489 dout("check_caps queuing invalidate\n"); 1490 queue_invalidate = 1; 1491 ci->i_rdcache_revoking = ci->i_rdcache_gen; 1492 } else { 1493 dout("check_caps failed to invalidate pages\n"); 1494 /* we failed to invalidate pages. check these 1495 caps again later. */ 1496 force_requeue = 1; 1497 __cap_set_timeouts(mdsc, ci); 1498 } 1499 } 1500 tried_invalidate = 1; 1501 goto retry_locked; 1502 } 1503 1504 num = 0; 1505 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 1506 cap = rb_entry(p, struct ceph_cap, ci_node); 1507 num++; 1508 1509 /* avoid looping forever */ 1510 if (mds >= cap->mds || 1511 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap)) 1512 continue; 1513 1514 /* NOTE: no side-effects allowed, until we take s_mutex */ 1515 1516 revoking = cap->implemented & ~cap->issued; 1517 if (revoking) 1518 dout(" mds%d revoking %s\n", cap->mds, 1519 ceph_cap_string(revoking)); 1520 1521 if (cap == ci->i_auth_cap && 1522 (cap->issued & CEPH_CAP_FILE_WR)) { 1523 /* request larger max_size from MDS? */ 1524 if (ci->i_wanted_max_size > ci->i_max_size && 1525 ci->i_wanted_max_size > ci->i_requested_max_size) { 1526 dout("requesting new max_size\n"); 1527 goto ack; 1528 } 1529 1530 /* approaching file_max? */ 1531 if ((inode->i_size << 1) >= ci->i_max_size && 1532 (ci->i_reported_size << 1) < ci->i_max_size) { 1533 dout("i_size approaching max_size\n"); 1534 goto ack; 1535 } 1536 } 1537 /* flush anything dirty? */ 1538 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) && 1539 ci->i_dirty_caps) { 1540 dout("flushing dirty caps\n"); 1541 goto ack; 1542 } 1543 1544 /* completed revocation? going down and there are no caps? */ 1545 if (revoking && (revoking & used) == 0) { 1546 dout("completed revocation of %s\n", 1547 ceph_cap_string(cap->implemented & ~cap->issued)); 1548 goto ack; 1549 } 1550 1551 /* want more caps from mds? */ 1552 if (want & ~(cap->mds_wanted | cap->issued)) 1553 goto ack; 1554 1555 /* things we might delay */ 1556 if ((cap->issued & ~retain) == 0 && 1557 cap->mds_wanted == want) 1558 continue; /* nope, all good */ 1559 1560 if (is_delayed) 1561 goto ack; 1562 1563 /* delay? */ 1564 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1565 time_before(jiffies, ci->i_hold_caps_max)) { 1566 dout(" delaying issued %s -> %s, wanted %s -> %s\n", 1567 ceph_cap_string(cap->issued), 1568 ceph_cap_string(cap->issued & retain), 1569 ceph_cap_string(cap->mds_wanted), 1570 ceph_cap_string(want)); 1571 delayed++; 1572 continue; 1573 } 1574 1575 ack: 1576 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { 1577 dout(" skipping %p I_NOFLUSH set\n", inode); 1578 continue; 1579 } 1580 1581 if (session && session != cap->session) { 1582 dout("oops, wrong session %p mutex\n", session); 1583 mutex_unlock(&session->s_mutex); 1584 session = NULL; 1585 } 1586 if (!session) { 1587 session = cap->session; 1588 if (mutex_trylock(&session->s_mutex) == 0) { 1589 dout("inverting session/ino locks on %p\n", 1590 session); 1591 spin_unlock(&inode->i_lock); 1592 if (took_snap_rwsem) { 1593 up_read(&mdsc->snap_rwsem); 1594 took_snap_rwsem = 0; 1595 } 1596 mutex_lock(&session->s_mutex); 1597 goto retry; 1598 } 1599 } 1600 /* take snap_rwsem after session mutex */ 1601 if (!took_snap_rwsem) { 1602 if (down_read_trylock(&mdsc->snap_rwsem) == 0) { 1603 dout("inverting snap/in locks on %p\n", 1604 inode); 1605 spin_unlock(&inode->i_lock); 1606 down_read(&mdsc->snap_rwsem); 1607 took_snap_rwsem = 1; 1608 goto retry; 1609 } 1610 took_snap_rwsem = 1; 1611 } 1612 1613 if (cap == ci->i_auth_cap && ci->i_dirty_caps) 1614 flushing = __mark_caps_flushing(inode, session); 1615 1616 mds = cap->mds; /* remember mds, so we don't repeat */ 1617 sent++; 1618 1619 /* __send_cap drops i_lock */ 1620 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want, 1621 retain, flushing, NULL); 1622 goto retry; /* retake i_lock and restart our cap scan. */ 1623 } 1624 1625 /* 1626 * Reschedule delayed caps release if we delayed anything, 1627 * otherwise cancel. 1628 */ 1629 if (delayed && is_delayed) 1630 force_requeue = 1; /* __send_cap delayed release; requeue */ 1631 if (!delayed && !is_delayed) 1632 __cap_delay_cancel(mdsc, ci); 1633 else if (!is_delayed || force_requeue) 1634 __cap_delay_requeue(mdsc, ci); 1635 1636 spin_unlock(&inode->i_lock); 1637 1638 if (queue_invalidate) 1639 ceph_queue_invalidate(inode); 1640 1641 if (session && drop_session_lock) 1642 mutex_unlock(&session->s_mutex); 1643 if (took_snap_rwsem) 1644 up_read(&mdsc->snap_rwsem); 1645 } 1646 1647 /* 1648 * Try to flush dirty caps back to the auth mds. 1649 */ 1650 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session, 1651 unsigned *flush_tid) 1652 { 1653 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1654 struct ceph_inode_info *ci = ceph_inode(inode); 1655 int unlock_session = session ? 0 : 1; 1656 int flushing = 0; 1657 1658 retry: 1659 spin_lock(&inode->i_lock); 1660 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { 1661 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode); 1662 goto out; 1663 } 1664 if (ci->i_dirty_caps && ci->i_auth_cap) { 1665 struct ceph_cap *cap = ci->i_auth_cap; 1666 int used = __ceph_caps_used(ci); 1667 int want = __ceph_caps_wanted(ci); 1668 int delayed; 1669 1670 if (!session) { 1671 spin_unlock(&inode->i_lock); 1672 session = cap->session; 1673 mutex_lock(&session->s_mutex); 1674 goto retry; 1675 } 1676 BUG_ON(session != cap->session); 1677 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) 1678 goto out; 1679 1680 flushing = __mark_caps_flushing(inode, session); 1681 1682 /* __send_cap drops i_lock */ 1683 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want, 1684 cap->issued | cap->implemented, flushing, 1685 flush_tid); 1686 if (!delayed) 1687 goto out_unlocked; 1688 1689 spin_lock(&inode->i_lock); 1690 __cap_delay_requeue(mdsc, ci); 1691 } 1692 out: 1693 spin_unlock(&inode->i_lock); 1694 out_unlocked: 1695 if (session && unlock_session) 1696 mutex_unlock(&session->s_mutex); 1697 return flushing; 1698 } 1699 1700 /* 1701 * Return true if we've flushed caps through the given flush_tid. 1702 */ 1703 static int caps_are_flushed(struct inode *inode, unsigned tid) 1704 { 1705 struct ceph_inode_info *ci = ceph_inode(inode); 1706 int dirty, i, ret = 1; 1707 1708 spin_lock(&inode->i_lock); 1709 dirty = __ceph_caps_dirty(ci); 1710 for (i = 0; i < CEPH_CAP_BITS; i++) 1711 if ((ci->i_flushing_caps & (1 << i)) && 1712 ci->i_cap_flush_tid[i] <= tid) { 1713 /* still flushing this bit */ 1714 ret = 0; 1715 break; 1716 } 1717 spin_unlock(&inode->i_lock); 1718 return ret; 1719 } 1720 1721 /* 1722 * Wait on any unsafe replies for the given inode. First wait on the 1723 * newest request, and make that the upper bound. Then, if there are 1724 * more requests, keep waiting on the oldest as long as it is still older 1725 * than the original request. 1726 */ 1727 static void sync_write_wait(struct inode *inode) 1728 { 1729 struct ceph_inode_info *ci = ceph_inode(inode); 1730 struct list_head *head = &ci->i_unsafe_writes; 1731 struct ceph_osd_request *req; 1732 u64 last_tid; 1733 1734 spin_lock(&ci->i_unsafe_lock); 1735 if (list_empty(head)) 1736 goto out; 1737 1738 /* set upper bound as _last_ entry in chain */ 1739 req = list_entry(head->prev, struct ceph_osd_request, 1740 r_unsafe_item); 1741 last_tid = req->r_tid; 1742 1743 do { 1744 ceph_osdc_get_request(req); 1745 spin_unlock(&ci->i_unsafe_lock); 1746 dout("sync_write_wait on tid %llu (until %llu)\n", 1747 req->r_tid, last_tid); 1748 wait_for_completion(&req->r_safe_completion); 1749 spin_lock(&ci->i_unsafe_lock); 1750 ceph_osdc_put_request(req); 1751 1752 /* 1753 * from here on look at first entry in chain, since we 1754 * only want to wait for anything older than last_tid 1755 */ 1756 if (list_empty(head)) 1757 break; 1758 req = list_entry(head->next, struct ceph_osd_request, 1759 r_unsafe_item); 1760 } while (req->r_tid < last_tid); 1761 out: 1762 spin_unlock(&ci->i_unsafe_lock); 1763 } 1764 1765 int ceph_fsync(struct file *file, struct dentry *dentry, int datasync) 1766 { 1767 struct inode *inode = dentry->d_inode; 1768 struct ceph_inode_info *ci = ceph_inode(inode); 1769 unsigned flush_tid; 1770 int ret; 1771 int dirty; 1772 1773 dout("fsync %p%s\n", inode, datasync ? " datasync" : ""); 1774 sync_write_wait(inode); 1775 1776 ret = filemap_write_and_wait(inode->i_mapping); 1777 if (ret < 0) 1778 return ret; 1779 1780 dirty = try_flush_caps(inode, NULL, &flush_tid); 1781 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); 1782 1783 /* 1784 * only wait on non-file metadata writeback (the mds 1785 * can recover size and mtime, so we don't need to 1786 * wait for that) 1787 */ 1788 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { 1789 dout("fsync waiting for flush_tid %u\n", flush_tid); 1790 ret = wait_event_interruptible(ci->i_cap_wq, 1791 caps_are_flushed(inode, flush_tid)); 1792 } 1793 1794 dout("fsync %p%s done\n", inode, datasync ? " datasync" : ""); 1795 return ret; 1796 } 1797 1798 /* 1799 * Flush any dirty caps back to the mds. If we aren't asked to wait, 1800 * queue inode for flush but don't do so immediately, because we can 1801 * get by with fewer MDS messages if we wait for data writeback to 1802 * complete first. 1803 */ 1804 int ceph_write_inode(struct inode *inode, int wait) 1805 { 1806 struct ceph_inode_info *ci = ceph_inode(inode); 1807 unsigned flush_tid; 1808 int err = 0; 1809 int dirty; 1810 1811 dout("write_inode %p wait=%d\n", inode, wait); 1812 if (wait) { 1813 dirty = try_flush_caps(inode, NULL, &flush_tid); 1814 if (dirty) 1815 err = wait_event_interruptible(ci->i_cap_wq, 1816 caps_are_flushed(inode, flush_tid)); 1817 } else { 1818 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1819 1820 spin_lock(&inode->i_lock); 1821 if (__ceph_caps_dirty(ci)) 1822 __cap_delay_requeue_front(mdsc, ci); 1823 spin_unlock(&inode->i_lock); 1824 } 1825 return err; 1826 } 1827 1828 /* 1829 * After a recovering MDS goes active, we need to resend any caps 1830 * we were flushing. 1831 * 1832 * Caller holds session->s_mutex. 1833 */ 1834 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc, 1835 struct ceph_mds_session *session) 1836 { 1837 struct ceph_cap_snap *capsnap; 1838 1839 dout("kick_flushing_capsnaps mds%d\n", session->s_mds); 1840 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing, 1841 flushing_item) { 1842 struct ceph_inode_info *ci = capsnap->ci; 1843 struct inode *inode = &ci->vfs_inode; 1844 struct ceph_cap *cap; 1845 1846 spin_lock(&inode->i_lock); 1847 cap = ci->i_auth_cap; 1848 if (cap && cap->session == session) { 1849 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode, 1850 cap, capsnap); 1851 __ceph_flush_snaps(ci, &session); 1852 } else { 1853 pr_err("%p auth cap %p not mds%d ???\n", inode, 1854 cap, session->s_mds); 1855 spin_unlock(&inode->i_lock); 1856 } 1857 } 1858 } 1859 1860 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 1861 struct ceph_mds_session *session) 1862 { 1863 struct ceph_inode_info *ci; 1864 1865 kick_flushing_capsnaps(mdsc, session); 1866 1867 dout("kick_flushing_caps mds%d\n", session->s_mds); 1868 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 1869 struct inode *inode = &ci->vfs_inode; 1870 struct ceph_cap *cap; 1871 int delayed = 0; 1872 1873 spin_lock(&inode->i_lock); 1874 cap = ci->i_auth_cap; 1875 if (cap && cap->session == session) { 1876 dout("kick_flushing_caps %p cap %p %s\n", inode, 1877 cap, ceph_cap_string(ci->i_flushing_caps)); 1878 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, 1879 __ceph_caps_used(ci), 1880 __ceph_caps_wanted(ci), 1881 cap->issued | cap->implemented, 1882 ci->i_flushing_caps, NULL); 1883 if (delayed) { 1884 spin_lock(&inode->i_lock); 1885 __cap_delay_requeue(mdsc, ci); 1886 spin_unlock(&inode->i_lock); 1887 } 1888 } else { 1889 pr_err("%p auth cap %p not mds%d ???\n", inode, 1890 cap, session->s_mds); 1891 spin_unlock(&inode->i_lock); 1892 } 1893 } 1894 } 1895 1896 1897 /* 1898 * Take references to capabilities we hold, so that we don't release 1899 * them to the MDS prematurely. 1900 * 1901 * Protected by i_lock. 1902 */ 1903 static void __take_cap_refs(struct ceph_inode_info *ci, int got) 1904 { 1905 if (got & CEPH_CAP_PIN) 1906 ci->i_pin_ref++; 1907 if (got & CEPH_CAP_FILE_RD) 1908 ci->i_rd_ref++; 1909 if (got & CEPH_CAP_FILE_CACHE) 1910 ci->i_rdcache_ref++; 1911 if (got & CEPH_CAP_FILE_WR) 1912 ci->i_wr_ref++; 1913 if (got & CEPH_CAP_FILE_BUFFER) { 1914 if (ci->i_wrbuffer_ref == 0) 1915 igrab(&ci->vfs_inode); 1916 ci->i_wrbuffer_ref++; 1917 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n", 1918 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref); 1919 } 1920 } 1921 1922 /* 1923 * Try to grab cap references. Specify those refs we @want, and the 1924 * minimal set we @need. Also include the larger offset we are writing 1925 * to (when applicable), and check against max_size here as well. 1926 * Note that caller is responsible for ensuring max_size increases are 1927 * requested from the MDS. 1928 */ 1929 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want, 1930 int *got, loff_t endoff, int *check_max, int *err) 1931 { 1932 struct inode *inode = &ci->vfs_inode; 1933 int ret = 0; 1934 int have, implemented; 1935 int file_wanted; 1936 1937 dout("get_cap_refs %p need %s want %s\n", inode, 1938 ceph_cap_string(need), ceph_cap_string(want)); 1939 spin_lock(&inode->i_lock); 1940 1941 /* make sure file is actually open */ 1942 file_wanted = __ceph_caps_file_wanted(ci); 1943 if ((file_wanted & need) == 0) { 1944 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n", 1945 ceph_cap_string(need), ceph_cap_string(file_wanted)); 1946 *err = -EBADF; 1947 ret = 1; 1948 goto out; 1949 } 1950 1951 if (need & CEPH_CAP_FILE_WR) { 1952 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) { 1953 dout("get_cap_refs %p endoff %llu > maxsize %llu\n", 1954 inode, endoff, ci->i_max_size); 1955 if (endoff > ci->i_wanted_max_size) { 1956 *check_max = 1; 1957 ret = 1; 1958 } 1959 goto out; 1960 } 1961 /* 1962 * If a sync write is in progress, we must wait, so that we 1963 * can get a final snapshot value for size+mtime. 1964 */ 1965 if (__ceph_have_pending_cap_snap(ci)) { 1966 dout("get_cap_refs %p cap_snap_pending\n", inode); 1967 goto out; 1968 } 1969 } 1970 have = __ceph_caps_issued(ci, &implemented); 1971 1972 /* 1973 * disallow writes while a truncate is pending 1974 */ 1975 if (ci->i_truncate_pending) 1976 have &= ~CEPH_CAP_FILE_WR; 1977 1978 if ((have & need) == need) { 1979 /* 1980 * Look at (implemented & ~have & not) so that we keep waiting 1981 * on transition from wanted -> needed caps. This is needed 1982 * for WRBUFFER|WR -> WR to avoid a new WR sync write from 1983 * going before a prior buffered writeback happens. 1984 */ 1985 int not = want & ~(have & need); 1986 int revoking = implemented & ~have; 1987 dout("get_cap_refs %p have %s but not %s (revoking %s)\n", 1988 inode, ceph_cap_string(have), ceph_cap_string(not), 1989 ceph_cap_string(revoking)); 1990 if ((revoking & not) == 0) { 1991 *got = need | (have & want); 1992 __take_cap_refs(ci, *got); 1993 ret = 1; 1994 } 1995 } else { 1996 dout("get_cap_refs %p have %s needed %s\n", inode, 1997 ceph_cap_string(have), ceph_cap_string(need)); 1998 } 1999 out: 2000 spin_unlock(&inode->i_lock); 2001 dout("get_cap_refs %p ret %d got %s\n", inode, 2002 ret, ceph_cap_string(*got)); 2003 return ret; 2004 } 2005 2006 /* 2007 * Check the offset we are writing up to against our current 2008 * max_size. If necessary, tell the MDS we want to write to 2009 * a larger offset. 2010 */ 2011 static void check_max_size(struct inode *inode, loff_t endoff) 2012 { 2013 struct ceph_inode_info *ci = ceph_inode(inode); 2014 int check = 0; 2015 2016 /* do we need to explicitly request a larger max_size? */ 2017 spin_lock(&inode->i_lock); 2018 if ((endoff >= ci->i_max_size || 2019 endoff > (inode->i_size << 1)) && 2020 endoff > ci->i_wanted_max_size) { 2021 dout("write %p at large endoff %llu, req max_size\n", 2022 inode, endoff); 2023 ci->i_wanted_max_size = endoff; 2024 check = 1; 2025 } 2026 spin_unlock(&inode->i_lock); 2027 if (check) 2028 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2029 } 2030 2031 /* 2032 * Wait for caps, and take cap references. If we can't get a WR cap 2033 * due to a small max_size, make sure we check_max_size (and possibly 2034 * ask the mds) so we don't get hung up indefinitely. 2035 */ 2036 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got, 2037 loff_t endoff) 2038 { 2039 int check_max, ret, err; 2040 2041 retry: 2042 if (endoff > 0) 2043 check_max_size(&ci->vfs_inode, endoff); 2044 check_max = 0; 2045 err = 0; 2046 ret = wait_event_interruptible(ci->i_cap_wq, 2047 try_get_cap_refs(ci, need, want, 2048 got, endoff, 2049 &check_max, &err)); 2050 if (err) 2051 ret = err; 2052 if (check_max) 2053 goto retry; 2054 return ret; 2055 } 2056 2057 /* 2058 * Take cap refs. Caller must already know we hold at least one ref 2059 * on the caps in question or we don't know this is safe. 2060 */ 2061 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) 2062 { 2063 spin_lock(&ci->vfs_inode.i_lock); 2064 __take_cap_refs(ci, caps); 2065 spin_unlock(&ci->vfs_inode.i_lock); 2066 } 2067 2068 /* 2069 * Release cap refs. 2070 * 2071 * If we released the last ref on any given cap, call ceph_check_caps 2072 * to release (or schedule a release). 2073 * 2074 * If we are releasing a WR cap (from a sync write), finalize any affected 2075 * cap_snap, and wake up any waiters. 2076 */ 2077 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had) 2078 { 2079 struct inode *inode = &ci->vfs_inode; 2080 int last = 0, put = 0, flushsnaps = 0, wake = 0; 2081 struct ceph_cap_snap *capsnap; 2082 2083 spin_lock(&inode->i_lock); 2084 if (had & CEPH_CAP_PIN) 2085 --ci->i_pin_ref; 2086 if (had & CEPH_CAP_FILE_RD) 2087 if (--ci->i_rd_ref == 0) 2088 last++; 2089 if (had & CEPH_CAP_FILE_CACHE) 2090 if (--ci->i_rdcache_ref == 0) 2091 last++; 2092 if (had & CEPH_CAP_FILE_BUFFER) { 2093 if (--ci->i_wrbuffer_ref == 0) { 2094 last++; 2095 put++; 2096 } 2097 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n", 2098 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref); 2099 } 2100 if (had & CEPH_CAP_FILE_WR) 2101 if (--ci->i_wr_ref == 0) { 2102 last++; 2103 if (!list_empty(&ci->i_cap_snaps)) { 2104 capsnap = list_first_entry(&ci->i_cap_snaps, 2105 struct ceph_cap_snap, 2106 ci_item); 2107 if (capsnap->writing) { 2108 capsnap->writing = 0; 2109 flushsnaps = 2110 __ceph_finish_cap_snap(ci, 2111 capsnap); 2112 wake = 1; 2113 } 2114 } 2115 } 2116 spin_unlock(&inode->i_lock); 2117 2118 dout("put_cap_refs %p had %s %s\n", inode, ceph_cap_string(had), 2119 last ? "last" : ""); 2120 2121 if (last && !flushsnaps) 2122 ceph_check_caps(ci, 0, NULL); 2123 else if (flushsnaps) 2124 ceph_flush_snaps(ci); 2125 if (wake) 2126 wake_up(&ci->i_cap_wq); 2127 if (put) 2128 iput(inode); 2129 } 2130 2131 /* 2132 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap 2133 * context. Adjust per-snap dirty page accounting as appropriate. 2134 * Once all dirty data for a cap_snap is flushed, flush snapped file 2135 * metadata back to the MDS. If we dropped the last ref, call 2136 * ceph_check_caps. 2137 */ 2138 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr, 2139 struct ceph_snap_context *snapc) 2140 { 2141 struct inode *inode = &ci->vfs_inode; 2142 int last = 0; 2143 int last_snap = 0; 2144 int found = 0; 2145 struct ceph_cap_snap *capsnap = NULL; 2146 2147 spin_lock(&inode->i_lock); 2148 ci->i_wrbuffer_ref -= nr; 2149 last = !ci->i_wrbuffer_ref; 2150 2151 if (ci->i_head_snapc == snapc) { 2152 ci->i_wrbuffer_ref_head -= nr; 2153 if (!ci->i_wrbuffer_ref_head) { 2154 ceph_put_snap_context(ci->i_head_snapc); 2155 ci->i_head_snapc = NULL; 2156 } 2157 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n", 2158 inode, 2159 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr, 2160 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 2161 last ? " LAST" : ""); 2162 } else { 2163 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2164 if (capsnap->context == snapc) { 2165 found = 1; 2166 capsnap->dirty_pages -= nr; 2167 last_snap = !capsnap->dirty_pages; 2168 break; 2169 } 2170 } 2171 BUG_ON(!found); 2172 dout("put_wrbuffer_cap_refs on %p cap_snap %p " 2173 " snap %lld %d/%d -> %d/%d %s%s\n", 2174 inode, capsnap, capsnap->context->seq, 2175 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, 2176 ci->i_wrbuffer_ref, capsnap->dirty_pages, 2177 last ? " (wrbuffer last)" : "", 2178 last_snap ? " (capsnap last)" : ""); 2179 } 2180 2181 spin_unlock(&inode->i_lock); 2182 2183 if (last) { 2184 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2185 iput(inode); 2186 } else if (last_snap) { 2187 ceph_flush_snaps(ci); 2188 wake_up(&ci->i_cap_wq); 2189 } 2190 } 2191 2192 /* 2193 * Handle a cap GRANT message from the MDS. (Note that a GRANT may 2194 * actually be a revocation if it specifies a smaller cap set.) 2195 * 2196 * caller holds s_mutex. 2197 * return value: 2198 * 0 - ok 2199 * 1 - check_caps on auth cap only (writeback) 2200 * 2 - check_caps (ack revoke) 2201 */ 2202 static int handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant, 2203 struct ceph_mds_session *session, 2204 struct ceph_cap *cap, 2205 struct ceph_buffer *xattr_buf) 2206 __releases(inode->i_lock) 2207 2208 { 2209 struct ceph_inode_info *ci = ceph_inode(inode); 2210 int mds = session->s_mds; 2211 int seq = le32_to_cpu(grant->seq); 2212 int newcaps = le32_to_cpu(grant->caps); 2213 int issued, implemented, used, wanted, dirty; 2214 u64 size = le64_to_cpu(grant->size); 2215 u64 max_size = le64_to_cpu(grant->max_size); 2216 struct timespec mtime, atime, ctime; 2217 int reply = 0; 2218 int wake = 0; 2219 int writeback = 0; 2220 int revoked_rdcache = 0; 2221 int queue_invalidate = 0; 2222 2223 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n", 2224 inode, cap, mds, seq, ceph_cap_string(newcaps)); 2225 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size, 2226 inode->i_size); 2227 2228 /* 2229 * If CACHE is being revoked, and we have no dirty buffers, 2230 * try to invalidate (once). (If there are dirty buffers, we 2231 * will invalidate _after_ writeback.) 2232 */ 2233 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && 2234 !ci->i_wrbuffer_ref) { 2235 if (try_nonblocking_invalidate(inode) == 0) { 2236 revoked_rdcache = 1; 2237 } else { 2238 /* there were locked pages.. invalidate later 2239 in a separate thread. */ 2240 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) { 2241 queue_invalidate = 1; 2242 ci->i_rdcache_revoking = ci->i_rdcache_gen; 2243 } 2244 } 2245 } 2246 2247 /* side effects now are allowed */ 2248 2249 issued = __ceph_caps_issued(ci, &implemented); 2250 issued |= implemented | __ceph_caps_dirty(ci); 2251 2252 cap->cap_gen = session->s_cap_gen; 2253 2254 __check_cap_issue(ci, cap, newcaps); 2255 2256 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) { 2257 inode->i_mode = le32_to_cpu(grant->mode); 2258 inode->i_uid = le32_to_cpu(grant->uid); 2259 inode->i_gid = le32_to_cpu(grant->gid); 2260 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode, 2261 inode->i_uid, inode->i_gid); 2262 } 2263 2264 if ((issued & CEPH_CAP_LINK_EXCL) == 0) 2265 inode->i_nlink = le32_to_cpu(grant->nlink); 2266 2267 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) { 2268 int len = le32_to_cpu(grant->xattr_len); 2269 u64 version = le64_to_cpu(grant->xattr_version); 2270 2271 if (version > ci->i_xattrs.version) { 2272 dout(" got new xattrs v%llu on %p len %d\n", 2273 version, inode, len); 2274 if (ci->i_xattrs.blob) 2275 ceph_buffer_put(ci->i_xattrs.blob); 2276 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf); 2277 ci->i_xattrs.version = version; 2278 } 2279 } 2280 2281 /* size/ctime/mtime/atime? */ 2282 ceph_fill_file_size(inode, issued, 2283 le32_to_cpu(grant->truncate_seq), 2284 le64_to_cpu(grant->truncate_size), size); 2285 ceph_decode_timespec(&mtime, &grant->mtime); 2286 ceph_decode_timespec(&atime, &grant->atime); 2287 ceph_decode_timespec(&ctime, &grant->ctime); 2288 ceph_fill_file_time(inode, issued, 2289 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime, 2290 &atime); 2291 2292 /* max size increase? */ 2293 if (max_size != ci->i_max_size) { 2294 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size); 2295 ci->i_max_size = max_size; 2296 if (max_size >= ci->i_wanted_max_size) { 2297 ci->i_wanted_max_size = 0; /* reset */ 2298 ci->i_requested_max_size = 0; 2299 } 2300 wake = 1; 2301 } 2302 2303 /* check cap bits */ 2304 wanted = __ceph_caps_wanted(ci); 2305 used = __ceph_caps_used(ci); 2306 dirty = __ceph_caps_dirty(ci); 2307 dout(" my wanted = %s, used = %s, dirty %s\n", 2308 ceph_cap_string(wanted), 2309 ceph_cap_string(used), 2310 ceph_cap_string(dirty)); 2311 if (wanted != le32_to_cpu(grant->wanted)) { 2312 dout("mds wanted %s -> %s\n", 2313 ceph_cap_string(le32_to_cpu(grant->wanted)), 2314 ceph_cap_string(wanted)); 2315 grant->wanted = cpu_to_le32(wanted); 2316 } 2317 2318 cap->seq = seq; 2319 2320 /* file layout may have changed */ 2321 ci->i_layout = grant->layout; 2322 2323 /* revocation, grant, or no-op? */ 2324 if (cap->issued & ~newcaps) { 2325 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued), 2326 ceph_cap_string(newcaps)); 2327 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER) 2328 writeback = 1; /* will delay ack */ 2329 else if (dirty & ~newcaps) 2330 reply = 1; /* initiate writeback in check_caps */ 2331 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 || 2332 revoked_rdcache) 2333 reply = 2; /* send revoke ack in check_caps */ 2334 cap->issued = newcaps; 2335 } else if (cap->issued == newcaps) { 2336 dout("caps unchanged: %s -> %s\n", 2337 ceph_cap_string(cap->issued), ceph_cap_string(newcaps)); 2338 } else { 2339 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued), 2340 ceph_cap_string(newcaps)); 2341 cap->issued = newcaps; 2342 cap->implemented |= newcaps; /* add bits only, to 2343 * avoid stepping on a 2344 * pending revocation */ 2345 wake = 1; 2346 } 2347 2348 spin_unlock(&inode->i_lock); 2349 if (writeback) 2350 /* 2351 * queue inode for writeback: we can't actually call 2352 * filemap_write_and_wait, etc. from message handler 2353 * context. 2354 */ 2355 ceph_queue_writeback(inode); 2356 if (queue_invalidate) 2357 ceph_queue_invalidate(inode); 2358 if (wake) 2359 wake_up(&ci->i_cap_wq); 2360 return reply; 2361 } 2362 2363 /* 2364 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the 2365 * MDS has been safely committed. 2366 */ 2367 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid, 2368 struct ceph_mds_caps *m, 2369 struct ceph_mds_session *session, 2370 struct ceph_cap *cap) 2371 __releases(inode->i_lock) 2372 { 2373 struct ceph_inode_info *ci = ceph_inode(inode); 2374 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 2375 unsigned seq = le32_to_cpu(m->seq); 2376 int dirty = le32_to_cpu(m->dirty); 2377 int cleaned = 0; 2378 int drop = 0; 2379 int i; 2380 2381 for (i = 0; i < CEPH_CAP_BITS; i++) 2382 if ((dirty & (1 << i)) && 2383 flush_tid == ci->i_cap_flush_tid[i]) 2384 cleaned |= 1 << i; 2385 2386 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," 2387 " flushing %s -> %s\n", 2388 inode, session->s_mds, seq, ceph_cap_string(dirty), 2389 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), 2390 ceph_cap_string(ci->i_flushing_caps & ~cleaned)); 2391 2392 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned)) 2393 goto out; 2394 2395 ci->i_flushing_caps &= ~cleaned; 2396 2397 spin_lock(&mdsc->cap_dirty_lock); 2398 if (ci->i_flushing_caps == 0) { 2399 list_del_init(&ci->i_flushing_item); 2400 if (!list_empty(&session->s_cap_flushing)) 2401 dout(" mds%d still flushing cap on %p\n", 2402 session->s_mds, 2403 &list_entry(session->s_cap_flushing.next, 2404 struct ceph_inode_info, 2405 i_flushing_item)->vfs_inode); 2406 mdsc->num_cap_flushing--; 2407 wake_up(&mdsc->cap_flushing_wq); 2408 dout(" inode %p now !flushing\n", inode); 2409 2410 if (ci->i_dirty_caps == 0) { 2411 dout(" inode %p now clean\n", inode); 2412 BUG_ON(!list_empty(&ci->i_dirty_item)); 2413 drop = 1; 2414 } else { 2415 BUG_ON(list_empty(&ci->i_dirty_item)); 2416 } 2417 } 2418 spin_unlock(&mdsc->cap_dirty_lock); 2419 wake_up(&ci->i_cap_wq); 2420 2421 out: 2422 spin_unlock(&inode->i_lock); 2423 if (drop) 2424 iput(inode); 2425 } 2426 2427 /* 2428 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can 2429 * throw away our cap_snap. 2430 * 2431 * Caller hold s_mutex. 2432 */ 2433 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid, 2434 struct ceph_mds_caps *m, 2435 struct ceph_mds_session *session) 2436 { 2437 struct ceph_inode_info *ci = ceph_inode(inode); 2438 u64 follows = le64_to_cpu(m->snap_follows); 2439 struct ceph_cap_snap *capsnap; 2440 int drop = 0; 2441 2442 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n", 2443 inode, ci, session->s_mds, follows); 2444 2445 spin_lock(&inode->i_lock); 2446 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2447 if (capsnap->follows == follows) { 2448 if (capsnap->flush_tid != flush_tid) { 2449 dout(" cap_snap %p follows %lld tid %lld !=" 2450 " %lld\n", capsnap, follows, 2451 flush_tid, capsnap->flush_tid); 2452 break; 2453 } 2454 WARN_ON(capsnap->dirty_pages || capsnap->writing); 2455 dout(" removing cap_snap %p follows %lld\n", 2456 capsnap, follows); 2457 ceph_put_snap_context(capsnap->context); 2458 list_del(&capsnap->ci_item); 2459 list_del(&capsnap->flushing_item); 2460 ceph_put_cap_snap(capsnap); 2461 drop = 1; 2462 break; 2463 } else { 2464 dout(" skipping cap_snap %p follows %lld\n", 2465 capsnap, capsnap->follows); 2466 } 2467 } 2468 spin_unlock(&inode->i_lock); 2469 if (drop) 2470 iput(inode); 2471 } 2472 2473 /* 2474 * Handle TRUNC from MDS, indicating file truncation. 2475 * 2476 * caller hold s_mutex. 2477 */ 2478 static void handle_cap_trunc(struct inode *inode, 2479 struct ceph_mds_caps *trunc, 2480 struct ceph_mds_session *session) 2481 __releases(inode->i_lock) 2482 { 2483 struct ceph_inode_info *ci = ceph_inode(inode); 2484 int mds = session->s_mds; 2485 int seq = le32_to_cpu(trunc->seq); 2486 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq); 2487 u64 truncate_size = le64_to_cpu(trunc->truncate_size); 2488 u64 size = le64_to_cpu(trunc->size); 2489 int implemented = 0; 2490 int dirty = __ceph_caps_dirty(ci); 2491 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented); 2492 int queue_trunc = 0; 2493 2494 issued |= implemented | dirty; 2495 2496 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n", 2497 inode, mds, seq, truncate_size, truncate_seq); 2498 queue_trunc = ceph_fill_file_size(inode, issued, 2499 truncate_seq, truncate_size, size); 2500 spin_unlock(&inode->i_lock); 2501 2502 if (queue_trunc) 2503 ceph_queue_vmtruncate(inode); 2504 } 2505 2506 /* 2507 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a 2508 * different one. If we are the most recent migration we've seen (as 2509 * indicated by mseq), make note of the migrating cap bits for the 2510 * duration (until we see the corresponding IMPORT). 2511 * 2512 * caller holds s_mutex 2513 */ 2514 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex, 2515 struct ceph_mds_session *session) 2516 { 2517 struct ceph_inode_info *ci = ceph_inode(inode); 2518 int mds = session->s_mds; 2519 unsigned mseq = le32_to_cpu(ex->migrate_seq); 2520 struct ceph_cap *cap = NULL, *t; 2521 struct rb_node *p; 2522 int remember = 1; 2523 2524 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n", 2525 inode, ci, mds, mseq); 2526 2527 spin_lock(&inode->i_lock); 2528 2529 /* make sure we haven't seen a higher mseq */ 2530 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 2531 t = rb_entry(p, struct ceph_cap, ci_node); 2532 if (ceph_seq_cmp(t->mseq, mseq) > 0) { 2533 dout(" higher mseq on cap from mds%d\n", 2534 t->session->s_mds); 2535 remember = 0; 2536 } 2537 if (t->session->s_mds == mds) 2538 cap = t; 2539 } 2540 2541 if (cap) { 2542 if (remember) { 2543 /* make note */ 2544 ci->i_cap_exporting_mds = mds; 2545 ci->i_cap_exporting_mseq = mseq; 2546 ci->i_cap_exporting_issued = cap->issued; 2547 } 2548 __ceph_remove_cap(cap); 2549 } else { 2550 WARN_ON(!cap); 2551 } 2552 2553 spin_unlock(&inode->i_lock); 2554 } 2555 2556 /* 2557 * Handle cap IMPORT. If there are temp bits from an older EXPORT, 2558 * clean them up. 2559 * 2560 * caller holds s_mutex. 2561 */ 2562 static void handle_cap_import(struct ceph_mds_client *mdsc, 2563 struct inode *inode, struct ceph_mds_caps *im, 2564 struct ceph_mds_session *session, 2565 void *snaptrace, int snaptrace_len) 2566 { 2567 struct ceph_inode_info *ci = ceph_inode(inode); 2568 int mds = session->s_mds; 2569 unsigned issued = le32_to_cpu(im->caps); 2570 unsigned wanted = le32_to_cpu(im->wanted); 2571 unsigned seq = le32_to_cpu(im->seq); 2572 unsigned mseq = le32_to_cpu(im->migrate_seq); 2573 u64 realmino = le64_to_cpu(im->realm); 2574 u64 cap_id = le64_to_cpu(im->cap_id); 2575 2576 if (ci->i_cap_exporting_mds >= 0 && 2577 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) { 2578 dout("handle_cap_import inode %p ci %p mds%d mseq %d" 2579 " - cleared exporting from mds%d\n", 2580 inode, ci, mds, mseq, 2581 ci->i_cap_exporting_mds); 2582 ci->i_cap_exporting_issued = 0; 2583 ci->i_cap_exporting_mseq = 0; 2584 ci->i_cap_exporting_mds = -1; 2585 } else { 2586 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n", 2587 inode, ci, mds, mseq); 2588 } 2589 2590 down_write(&mdsc->snap_rwsem); 2591 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len, 2592 false); 2593 downgrade_write(&mdsc->snap_rwsem); 2594 ceph_add_cap(inode, session, cap_id, -1, 2595 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH, 2596 NULL /* no caps context */); 2597 try_flush_caps(inode, session, NULL); 2598 up_read(&mdsc->snap_rwsem); 2599 } 2600 2601 /* 2602 * Handle a caps message from the MDS. 2603 * 2604 * Identify the appropriate session, inode, and call the right handler 2605 * based on the cap op. 2606 */ 2607 void ceph_handle_caps(struct ceph_mds_session *session, 2608 struct ceph_msg *msg) 2609 { 2610 struct ceph_mds_client *mdsc = session->s_mdsc; 2611 struct super_block *sb = mdsc->client->sb; 2612 struct inode *inode; 2613 struct ceph_cap *cap; 2614 struct ceph_mds_caps *h; 2615 int mds = session->s_mds; 2616 int op; 2617 u32 seq; 2618 struct ceph_vino vino; 2619 u64 cap_id; 2620 u64 size, max_size; 2621 u64 tid; 2622 int check_caps = 0; 2623 void *snaptrace; 2624 int r; 2625 2626 dout("handle_caps from mds%d\n", mds); 2627 2628 /* decode */ 2629 tid = le64_to_cpu(msg->hdr.tid); 2630 if (msg->front.iov_len < sizeof(*h)) 2631 goto bad; 2632 h = msg->front.iov_base; 2633 snaptrace = h + 1; 2634 op = le32_to_cpu(h->op); 2635 vino.ino = le64_to_cpu(h->ino); 2636 vino.snap = CEPH_NOSNAP; 2637 cap_id = le64_to_cpu(h->cap_id); 2638 seq = le32_to_cpu(h->seq); 2639 size = le64_to_cpu(h->size); 2640 max_size = le64_to_cpu(h->max_size); 2641 2642 mutex_lock(&session->s_mutex); 2643 session->s_seq++; 2644 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, 2645 (unsigned)seq); 2646 2647 /* lookup ino */ 2648 inode = ceph_find_inode(sb, vino); 2649 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino, 2650 vino.snap, inode); 2651 if (!inode) { 2652 dout(" i don't have ino %llx\n", vino.ino); 2653 goto done; 2654 } 2655 2656 /* these will work even if we don't have a cap yet */ 2657 switch (op) { 2658 case CEPH_CAP_OP_FLUSHSNAP_ACK: 2659 handle_cap_flushsnap_ack(inode, tid, h, session); 2660 goto done; 2661 2662 case CEPH_CAP_OP_EXPORT: 2663 handle_cap_export(inode, h, session); 2664 goto done; 2665 2666 case CEPH_CAP_OP_IMPORT: 2667 handle_cap_import(mdsc, inode, h, session, 2668 snaptrace, le32_to_cpu(h->snap_trace_len)); 2669 check_caps = 1; /* we may have sent a RELEASE to the old auth */ 2670 goto done; 2671 } 2672 2673 /* the rest require a cap */ 2674 spin_lock(&inode->i_lock); 2675 cap = __get_cap_for_mds(ceph_inode(inode), mds); 2676 if (!cap) { 2677 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n", 2678 inode, ceph_ino(inode), ceph_snap(inode), mds); 2679 spin_unlock(&inode->i_lock); 2680 goto done; 2681 } 2682 2683 /* note that each of these drops i_lock for us */ 2684 switch (op) { 2685 case CEPH_CAP_OP_REVOKE: 2686 case CEPH_CAP_OP_GRANT: 2687 r = handle_cap_grant(inode, h, session, cap, msg->middle); 2688 if (r == 1) 2689 ceph_check_caps(ceph_inode(inode), 2690 CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY, 2691 session); 2692 else if (r == 2) 2693 ceph_check_caps(ceph_inode(inode), 2694 CHECK_CAPS_NODELAY, 2695 session); 2696 break; 2697 2698 case CEPH_CAP_OP_FLUSH_ACK: 2699 handle_cap_flush_ack(inode, tid, h, session, cap); 2700 break; 2701 2702 case CEPH_CAP_OP_TRUNC: 2703 handle_cap_trunc(inode, h, session); 2704 break; 2705 2706 default: 2707 spin_unlock(&inode->i_lock); 2708 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op, 2709 ceph_cap_op_name(op)); 2710 } 2711 2712 done: 2713 mutex_unlock(&session->s_mutex); 2714 2715 if (check_caps) 2716 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY, NULL); 2717 if (inode) 2718 iput(inode); 2719 return; 2720 2721 bad: 2722 pr_err("ceph_handle_caps: corrupt message\n"); 2723 ceph_msg_dump(msg); 2724 return; 2725 } 2726 2727 /* 2728 * Delayed work handler to process end of delayed cap release LRU list. 2729 */ 2730 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc) 2731 { 2732 struct ceph_inode_info *ci; 2733 int flags = CHECK_CAPS_NODELAY; 2734 2735 dout("check_delayed_caps\n"); 2736 while (1) { 2737 spin_lock(&mdsc->cap_delay_lock); 2738 if (list_empty(&mdsc->cap_delay_list)) 2739 break; 2740 ci = list_first_entry(&mdsc->cap_delay_list, 2741 struct ceph_inode_info, 2742 i_cap_delay_list); 2743 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 && 2744 time_before(jiffies, ci->i_hold_caps_max)) 2745 break; 2746 list_del_init(&ci->i_cap_delay_list); 2747 spin_unlock(&mdsc->cap_delay_lock); 2748 dout("check_delayed_caps on %p\n", &ci->vfs_inode); 2749 ceph_check_caps(ci, flags, NULL); 2750 } 2751 spin_unlock(&mdsc->cap_delay_lock); 2752 } 2753 2754 /* 2755 * Flush all dirty caps to the mds 2756 */ 2757 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc) 2758 { 2759 struct ceph_inode_info *ci, *nci = NULL; 2760 struct inode *inode, *ninode = NULL; 2761 struct list_head *p, *n; 2762 2763 dout("flush_dirty_caps\n"); 2764 spin_lock(&mdsc->cap_dirty_lock); 2765 list_for_each_safe(p, n, &mdsc->cap_dirty) { 2766 if (nci) { 2767 ci = nci; 2768 inode = ninode; 2769 ci->i_ceph_flags &= ~CEPH_I_NOFLUSH; 2770 dout("flush_dirty_caps inode %p (was next inode)\n", 2771 inode); 2772 } else { 2773 ci = list_entry(p, struct ceph_inode_info, 2774 i_dirty_item); 2775 inode = igrab(&ci->vfs_inode); 2776 BUG_ON(!inode); 2777 dout("flush_dirty_caps inode %p\n", inode); 2778 } 2779 if (n != &mdsc->cap_dirty) { 2780 nci = list_entry(n, struct ceph_inode_info, 2781 i_dirty_item); 2782 ninode = igrab(&nci->vfs_inode); 2783 BUG_ON(!ninode); 2784 nci->i_ceph_flags |= CEPH_I_NOFLUSH; 2785 dout("flush_dirty_caps next inode %p, noflush\n", 2786 ninode); 2787 } else { 2788 nci = NULL; 2789 ninode = NULL; 2790 } 2791 spin_unlock(&mdsc->cap_dirty_lock); 2792 if (inode) { 2793 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, 2794 NULL); 2795 iput(inode); 2796 } 2797 spin_lock(&mdsc->cap_dirty_lock); 2798 } 2799 spin_unlock(&mdsc->cap_dirty_lock); 2800 } 2801 2802 /* 2803 * Drop open file reference. If we were the last open file, 2804 * we may need to release capabilities to the MDS (or schedule 2805 * their delayed release). 2806 */ 2807 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode) 2808 { 2809 struct inode *inode = &ci->vfs_inode; 2810 int last = 0; 2811 2812 spin_lock(&inode->i_lock); 2813 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode, 2814 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1); 2815 BUG_ON(ci->i_nr_by_mode[fmode] == 0); 2816 if (--ci->i_nr_by_mode[fmode] == 0) 2817 last++; 2818 spin_unlock(&inode->i_lock); 2819 2820 if (last && ci->i_vino.snap == CEPH_NOSNAP) 2821 ceph_check_caps(ci, 0, NULL); 2822 } 2823 2824 /* 2825 * Helpers for embedding cap and dentry lease releases into mds 2826 * requests. 2827 * 2828 * @force is used by dentry_release (below) to force inclusion of a 2829 * record for the directory inode, even when there aren't any caps to 2830 * drop. 2831 */ 2832 int ceph_encode_inode_release(void **p, struct inode *inode, 2833 int mds, int drop, int unless, int force) 2834 { 2835 struct ceph_inode_info *ci = ceph_inode(inode); 2836 struct ceph_cap *cap; 2837 struct ceph_mds_request_release *rel = *p; 2838 int ret = 0; 2839 2840 dout("encode_inode_release %p mds%d drop %s unless %s\n", inode, 2841 mds, ceph_cap_string(drop), ceph_cap_string(unless)); 2842 2843 spin_lock(&inode->i_lock); 2844 cap = __get_cap_for_mds(ci, mds); 2845 if (cap && __cap_is_valid(cap)) { 2846 if (force || 2847 ((cap->issued & drop) && 2848 (cap->issued & unless) == 0)) { 2849 if ((cap->issued & drop) && 2850 (cap->issued & unless) == 0) { 2851 dout("encode_inode_release %p cap %p %s -> " 2852 "%s\n", inode, cap, 2853 ceph_cap_string(cap->issued), 2854 ceph_cap_string(cap->issued & ~drop)); 2855 cap->issued &= ~drop; 2856 cap->implemented &= ~drop; 2857 if (ci->i_ceph_flags & CEPH_I_NODELAY) { 2858 int wanted = __ceph_caps_wanted(ci); 2859 dout(" wanted %s -> %s (act %s)\n", 2860 ceph_cap_string(cap->mds_wanted), 2861 ceph_cap_string(cap->mds_wanted & 2862 ~wanted), 2863 ceph_cap_string(wanted)); 2864 cap->mds_wanted &= wanted; 2865 } 2866 } else { 2867 dout("encode_inode_release %p cap %p %s" 2868 " (force)\n", inode, cap, 2869 ceph_cap_string(cap->issued)); 2870 } 2871 2872 rel->ino = cpu_to_le64(ceph_ino(inode)); 2873 rel->cap_id = cpu_to_le64(cap->cap_id); 2874 rel->seq = cpu_to_le32(cap->seq); 2875 rel->issue_seq = cpu_to_le32(cap->issue_seq), 2876 rel->mseq = cpu_to_le32(cap->mseq); 2877 rel->caps = cpu_to_le32(cap->issued); 2878 rel->wanted = cpu_to_le32(cap->mds_wanted); 2879 rel->dname_len = 0; 2880 rel->dname_seq = 0; 2881 *p += sizeof(*rel); 2882 ret = 1; 2883 } else { 2884 dout("encode_inode_release %p cap %p %s\n", 2885 inode, cap, ceph_cap_string(cap->issued)); 2886 } 2887 } 2888 spin_unlock(&inode->i_lock); 2889 return ret; 2890 } 2891 2892 int ceph_encode_dentry_release(void **p, struct dentry *dentry, 2893 int mds, int drop, int unless) 2894 { 2895 struct inode *dir = dentry->d_parent->d_inode; 2896 struct ceph_mds_request_release *rel = *p; 2897 struct ceph_dentry_info *di = ceph_dentry(dentry); 2898 int force = 0; 2899 int ret; 2900 2901 /* 2902 * force an record for the directory caps if we have a dentry lease. 2903 * this is racy (can't take i_lock and d_lock together), but it 2904 * doesn't have to be perfect; the mds will revoke anything we don't 2905 * release. 2906 */ 2907 spin_lock(&dentry->d_lock); 2908 if (di->lease_session && di->lease_session->s_mds == mds) 2909 force = 1; 2910 spin_unlock(&dentry->d_lock); 2911 2912 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force); 2913 2914 spin_lock(&dentry->d_lock); 2915 if (ret && di->lease_session && di->lease_session->s_mds == mds) { 2916 dout("encode_dentry_release %p mds%d seq %d\n", 2917 dentry, mds, (int)di->lease_seq); 2918 rel->dname_len = cpu_to_le32(dentry->d_name.len); 2919 memcpy(*p, dentry->d_name.name, dentry->d_name.len); 2920 *p += dentry->d_name.len; 2921 rel->dname_seq = cpu_to_le32(di->lease_seq); 2922 } 2923 spin_unlock(&dentry->d_lock); 2924 return ret; 2925 } 2926