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->cap_gen = session->s_cap_gen; 613 cap->recon_gen = session->s_recon_gen; 614 615 if (fmode >= 0) 616 __ceph_get_fmode(ci, fmode); 617 spin_unlock(&inode->i_lock); 618 wake_up(&ci->i_cap_wq); 619 return 0; 620 } 621 622 /* 623 * Return true if cap has not timed out and belongs to the current 624 * generation of the MDS session (i.e. has not gone 'stale' due to 625 * us losing touch with the mds). 626 */ 627 static int __cap_is_valid(struct ceph_cap *cap) 628 { 629 unsigned long ttl; 630 u32 gen, recon_gen; 631 632 spin_lock(&cap->session->s_cap_lock); 633 gen = cap->session->s_cap_gen; 634 recon_gen = cap->session->s_recon_gen; 635 ttl = cap->session->s_cap_ttl; 636 spin_unlock(&cap->session->s_cap_lock); 637 638 if (cap->recon_gen != recon_gen) { 639 dout("__cap_is_valid %p cap %p issued %s " 640 "but DEAD (recon_gen %u vs %u)\n", &cap->ci->vfs_inode, 641 cap, ceph_cap_string(cap->issued), cap->recon_gen, 642 recon_gen); 643 return 0; 644 } 645 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) { 646 dout("__cap_is_valid %p cap %p issued %s " 647 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode, 648 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen); 649 return 0; 650 } 651 652 return 1; 653 } 654 655 /* 656 * Return set of valid cap bits issued to us. Note that caps time 657 * out, and may be invalidated in bulk if the client session times out 658 * and session->s_cap_gen is bumped. 659 */ 660 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented) 661 { 662 int have = ci->i_snap_caps; 663 struct ceph_cap *cap; 664 struct rb_node *p; 665 666 if (implemented) 667 *implemented = 0; 668 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 669 cap = rb_entry(p, struct ceph_cap, ci_node); 670 if (!__cap_is_valid(cap)) 671 continue; 672 dout("__ceph_caps_issued %p cap %p issued %s\n", 673 &ci->vfs_inode, cap, ceph_cap_string(cap->issued)); 674 have |= cap->issued; 675 if (implemented) 676 *implemented |= cap->implemented; 677 } 678 return have; 679 } 680 681 /* 682 * Get cap bits issued by caps other than @ocap 683 */ 684 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap) 685 { 686 int have = ci->i_snap_caps; 687 struct ceph_cap *cap; 688 struct rb_node *p; 689 690 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 691 cap = rb_entry(p, struct ceph_cap, ci_node); 692 if (cap == ocap) 693 continue; 694 if (!__cap_is_valid(cap)) 695 continue; 696 have |= cap->issued; 697 } 698 return have; 699 } 700 701 /* 702 * Move a cap to the end of the LRU (oldest caps at list head, newest 703 * at list tail). 704 */ 705 static void __touch_cap(struct ceph_cap *cap) 706 { 707 struct ceph_mds_session *s = cap->session; 708 709 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap, 710 s->s_mds); 711 spin_lock(&s->s_cap_lock); 712 list_move_tail(&cap->session_caps, &s->s_caps); 713 spin_unlock(&s->s_cap_lock); 714 } 715 716 /* 717 * Check if we hold the given mask. If so, move the cap(s) to the 718 * front of their respective LRUs. (This is the preferred way for 719 * callers to check for caps they want.) 720 */ 721 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch) 722 { 723 struct ceph_cap *cap; 724 struct rb_node *p; 725 int have = ci->i_snap_caps; 726 727 if ((have & mask) == mask) { 728 dout("__ceph_caps_issued_mask %p snap issued %s" 729 " (mask %s)\n", &ci->vfs_inode, 730 ceph_cap_string(have), 731 ceph_cap_string(mask)); 732 return 1; 733 } 734 735 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 736 cap = rb_entry(p, struct ceph_cap, ci_node); 737 if (!__cap_is_valid(cap)) 738 continue; 739 if ((cap->issued & mask) == mask) { 740 dout("__ceph_caps_issued_mask %p cap %p issued %s" 741 " (mask %s)\n", &ci->vfs_inode, cap, 742 ceph_cap_string(cap->issued), 743 ceph_cap_string(mask)); 744 if (touch) 745 __touch_cap(cap); 746 return 1; 747 } 748 749 /* does a combination of caps satisfy mask? */ 750 have |= cap->issued; 751 if ((have & mask) == mask) { 752 dout("__ceph_caps_issued_mask %p combo issued %s" 753 " (mask %s)\n", &ci->vfs_inode, 754 ceph_cap_string(cap->issued), 755 ceph_cap_string(mask)); 756 if (touch) { 757 struct rb_node *q; 758 759 /* touch this + preceeding caps */ 760 __touch_cap(cap); 761 for (q = rb_first(&ci->i_caps); q != p; 762 q = rb_next(q)) { 763 cap = rb_entry(q, struct ceph_cap, 764 ci_node); 765 if (!__cap_is_valid(cap)) 766 continue; 767 __touch_cap(cap); 768 } 769 } 770 return 1; 771 } 772 } 773 774 return 0; 775 } 776 777 /* 778 * Return true if mask caps are currently being revoked by an MDS. 779 */ 780 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask) 781 { 782 struct inode *inode = &ci->vfs_inode; 783 struct ceph_cap *cap; 784 struct rb_node *p; 785 int ret = 0; 786 787 spin_lock(&inode->i_lock); 788 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 789 cap = rb_entry(p, struct ceph_cap, ci_node); 790 if (__cap_is_valid(cap) && 791 (cap->implemented & ~cap->issued & mask)) { 792 ret = 1; 793 break; 794 } 795 } 796 spin_unlock(&inode->i_lock); 797 dout("ceph_caps_revoking %p %s = %d\n", inode, 798 ceph_cap_string(mask), ret); 799 return ret; 800 } 801 802 int __ceph_caps_used(struct ceph_inode_info *ci) 803 { 804 int used = 0; 805 if (ci->i_pin_ref) 806 used |= CEPH_CAP_PIN; 807 if (ci->i_rd_ref) 808 used |= CEPH_CAP_FILE_RD; 809 if (ci->i_rdcache_ref || ci->i_rdcache_gen) 810 used |= CEPH_CAP_FILE_CACHE; 811 if (ci->i_wr_ref) 812 used |= CEPH_CAP_FILE_WR; 813 if (ci->i_wrbuffer_ref) 814 used |= CEPH_CAP_FILE_BUFFER; 815 return used; 816 } 817 818 /* 819 * wanted, by virtue of open file modes 820 */ 821 int __ceph_caps_file_wanted(struct ceph_inode_info *ci) 822 { 823 int want = 0; 824 int mode; 825 for (mode = 0; mode < 4; mode++) 826 if (ci->i_nr_by_mode[mode]) 827 want |= ceph_caps_for_mode(mode); 828 return want; 829 } 830 831 /* 832 * Return caps we have registered with the MDS(s) as 'wanted'. 833 */ 834 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci) 835 { 836 struct ceph_cap *cap; 837 struct rb_node *p; 838 int mds_wanted = 0; 839 840 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 841 cap = rb_entry(p, struct ceph_cap, ci_node); 842 if (!__cap_is_valid(cap)) 843 continue; 844 mds_wanted |= cap->mds_wanted; 845 } 846 return mds_wanted; 847 } 848 849 /* 850 * called under i_lock 851 */ 852 static int __ceph_is_any_caps(struct ceph_inode_info *ci) 853 { 854 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0; 855 } 856 857 /* 858 * caller should hold i_lock, and session s_mutex. 859 * returns true if this is the last cap. if so, caller should iput. 860 */ 861 void __ceph_remove_cap(struct ceph_cap *cap, 862 struct ceph_cap_reservation *ctx) 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 session list */ 871 spin_lock(&session->s_cap_lock); 872 list_del_init(&cap->session_caps); 873 session->s_nr_caps--; 874 spin_unlock(&session->s_cap_lock); 875 876 /* remove from inode list */ 877 rb_erase(&cap->ci_node, &ci->i_caps); 878 cap->session = NULL; 879 if (ci->i_auth_cap == cap) 880 ci->i_auth_cap = NULL; 881 882 put_cap(cap, ctx); 883 884 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) { 885 struct ceph_snap_realm *realm = ci->i_snap_realm; 886 spin_lock(&realm->inodes_with_caps_lock); 887 list_del_init(&ci->i_snap_realm_item); 888 ci->i_snap_realm_counter++; 889 ci->i_snap_realm = NULL; 890 spin_unlock(&realm->inodes_with_caps_lock); 891 ceph_put_snap_realm(mdsc, realm); 892 } 893 if (!__ceph_is_any_real_caps(ci)) 894 __cap_delay_cancel(mdsc, ci); 895 } 896 897 /* 898 * Build and send a cap message to the given MDS. 899 * 900 * Caller should be holding s_mutex. 901 */ 902 static int send_cap_msg(struct ceph_mds_session *session, 903 u64 ino, u64 cid, int op, 904 int caps, int wanted, int dirty, 905 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq, 906 u64 size, u64 max_size, 907 struct timespec *mtime, struct timespec *atime, 908 u64 time_warp_seq, 909 uid_t uid, gid_t gid, mode_t mode, 910 u64 xattr_version, 911 struct ceph_buffer *xattrs_buf, 912 u64 follows) 913 { 914 struct ceph_mds_caps *fc; 915 struct ceph_msg *msg; 916 917 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s" 918 " seq %u/%u mseq %u follows %lld size %llu/%llu" 919 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op), 920 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted), 921 ceph_cap_string(dirty), 922 seq, issue_seq, mseq, follows, size, max_size, 923 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0); 924 925 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL); 926 if (IS_ERR(msg)) 927 return PTR_ERR(msg); 928 929 fc = msg->front.iov_base; 930 931 memset(fc, 0, sizeof(*fc)); 932 933 fc->cap_id = cpu_to_le64(cid); 934 fc->op = cpu_to_le32(op); 935 fc->seq = cpu_to_le32(seq); 936 fc->client_tid = cpu_to_le64(flush_tid); 937 fc->issue_seq = cpu_to_le32(issue_seq); 938 fc->migrate_seq = cpu_to_le32(mseq); 939 fc->caps = cpu_to_le32(caps); 940 fc->wanted = cpu_to_le32(wanted); 941 fc->dirty = cpu_to_le32(dirty); 942 fc->ino = cpu_to_le64(ino); 943 fc->snap_follows = cpu_to_le64(follows); 944 945 fc->size = cpu_to_le64(size); 946 fc->max_size = cpu_to_le64(max_size); 947 if (mtime) 948 ceph_encode_timespec(&fc->mtime, mtime); 949 if (atime) 950 ceph_encode_timespec(&fc->atime, atime); 951 fc->time_warp_seq = cpu_to_le32(time_warp_seq); 952 953 fc->uid = cpu_to_le32(uid); 954 fc->gid = cpu_to_le32(gid); 955 fc->mode = cpu_to_le32(mode); 956 957 fc->xattr_version = cpu_to_le64(xattr_version); 958 if (xattrs_buf) { 959 msg->middle = ceph_buffer_get(xattrs_buf); 960 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len); 961 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len); 962 } 963 964 ceph_con_send(&session->s_con, msg); 965 return 0; 966 } 967 968 /* 969 * Queue cap releases when an inode is dropped from our 970 * cache. 971 */ 972 void ceph_queue_caps_release(struct inode *inode) 973 { 974 struct ceph_inode_info *ci = ceph_inode(inode); 975 struct rb_node *p; 976 977 spin_lock(&inode->i_lock); 978 p = rb_first(&ci->i_caps); 979 while (p) { 980 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node); 981 struct ceph_mds_session *session = cap->session; 982 struct ceph_msg *msg; 983 struct ceph_mds_cap_release *head; 984 struct ceph_mds_cap_item *item; 985 986 spin_lock(&session->s_cap_lock); 987 BUG_ON(!session->s_num_cap_releases); 988 msg = list_first_entry(&session->s_cap_releases, 989 struct ceph_msg, list_head); 990 991 dout(" adding %p release to mds%d msg %p (%d left)\n", 992 inode, session->s_mds, msg, session->s_num_cap_releases); 993 994 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE); 995 head = msg->front.iov_base; 996 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1); 997 item = msg->front.iov_base + msg->front.iov_len; 998 item->ino = cpu_to_le64(ceph_ino(inode)); 999 item->cap_id = cpu_to_le64(cap->cap_id); 1000 item->migrate_seq = cpu_to_le32(cap->mseq); 1001 item->seq = cpu_to_le32(cap->issue_seq); 1002 1003 session->s_num_cap_releases--; 1004 1005 msg->front.iov_len += sizeof(*item); 1006 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) { 1007 dout(" release msg %p full\n", msg); 1008 list_move_tail(&msg->list_head, 1009 &session->s_cap_releases_done); 1010 } else { 1011 dout(" release msg %p at %d/%d (%d)\n", msg, 1012 (int)le32_to_cpu(head->num), 1013 (int)CEPH_CAPS_PER_RELEASE, 1014 (int)msg->front.iov_len); 1015 } 1016 spin_unlock(&session->s_cap_lock); 1017 p = rb_next(p); 1018 __ceph_remove_cap(cap, NULL); 1019 1020 } 1021 spin_unlock(&inode->i_lock); 1022 } 1023 1024 /* 1025 * Send a cap msg on the given inode. Update our caps state, then 1026 * drop i_lock and send the message. 1027 * 1028 * Make note of max_size reported/requested from mds, revoked caps 1029 * that have now been implemented. 1030 * 1031 * Make half-hearted attempt ot to invalidate page cache if we are 1032 * dropping RDCACHE. Note that this will leave behind locked pages 1033 * that we'll then need to deal with elsewhere. 1034 * 1035 * Return non-zero if delayed release, or we experienced an error 1036 * such that the caller should requeue + retry later. 1037 * 1038 * called with i_lock, then drops it. 1039 * caller should hold snap_rwsem (read), s_mutex. 1040 */ 1041 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap, 1042 int op, int used, int want, int retain, int flushing, 1043 unsigned *pflush_tid) 1044 __releases(cap->ci->vfs_inode->i_lock) 1045 { 1046 struct ceph_inode_info *ci = cap->ci; 1047 struct inode *inode = &ci->vfs_inode; 1048 u64 cap_id = cap->cap_id; 1049 int held = cap->issued | cap->implemented; 1050 int revoking = cap->implemented & ~cap->issued; 1051 int dropping = cap->issued & ~retain; 1052 int keep; 1053 u64 seq, issue_seq, mseq, time_warp_seq, follows; 1054 u64 size, max_size; 1055 struct timespec mtime, atime; 1056 int wake = 0; 1057 mode_t mode; 1058 uid_t uid; 1059 gid_t gid; 1060 struct ceph_mds_session *session; 1061 u64 xattr_version = 0; 1062 int delayed = 0; 1063 u64 flush_tid = 0; 1064 int i; 1065 int ret; 1066 1067 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n", 1068 inode, cap, cap->session, 1069 ceph_cap_string(held), ceph_cap_string(held & retain), 1070 ceph_cap_string(revoking)); 1071 BUG_ON((retain & CEPH_CAP_PIN) == 0); 1072 1073 session = cap->session; 1074 1075 /* don't release wanted unless we've waited a bit. */ 1076 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1077 time_before(jiffies, ci->i_hold_caps_min)) { 1078 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n", 1079 ceph_cap_string(cap->issued), 1080 ceph_cap_string(cap->issued & retain), 1081 ceph_cap_string(cap->mds_wanted), 1082 ceph_cap_string(want)); 1083 want |= cap->mds_wanted; 1084 retain |= cap->issued; 1085 delayed = 1; 1086 } 1087 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH); 1088 1089 cap->issued &= retain; /* drop bits we don't want */ 1090 if (cap->implemented & ~cap->issued) { 1091 /* 1092 * Wake up any waiters on wanted -> needed transition. 1093 * This is due to the weird transition from buffered 1094 * to sync IO... we need to flush dirty pages _before_ 1095 * allowing sync writes to avoid reordering. 1096 */ 1097 wake = 1; 1098 } 1099 cap->implemented &= cap->issued | used; 1100 cap->mds_wanted = want; 1101 1102 if (flushing) { 1103 /* 1104 * assign a tid for flush operations so we can avoid 1105 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark 1106 * clean type races. track latest tid for every bit 1107 * so we can handle flush AxFw, flush Fw, and have the 1108 * first ack clean Ax. 1109 */ 1110 flush_tid = ++ci->i_cap_flush_last_tid; 1111 if (pflush_tid) 1112 *pflush_tid = flush_tid; 1113 dout(" cap_flush_tid %d\n", (int)flush_tid); 1114 for (i = 0; i < CEPH_CAP_BITS; i++) 1115 if (flushing & (1 << i)) 1116 ci->i_cap_flush_tid[i] = flush_tid; 1117 } 1118 1119 keep = cap->implemented; 1120 seq = cap->seq; 1121 issue_seq = cap->issue_seq; 1122 mseq = cap->mseq; 1123 size = inode->i_size; 1124 ci->i_reported_size = size; 1125 max_size = ci->i_wanted_max_size; 1126 ci->i_requested_max_size = max_size; 1127 mtime = inode->i_mtime; 1128 atime = inode->i_atime; 1129 time_warp_seq = ci->i_time_warp_seq; 1130 follows = ci->i_snap_realm->cached_context->seq; 1131 uid = inode->i_uid; 1132 gid = inode->i_gid; 1133 mode = inode->i_mode; 1134 1135 if (dropping & CEPH_CAP_XATTR_EXCL) { 1136 __ceph_build_xattrs_blob(ci); 1137 xattr_version = ci->i_xattrs.version + 1; 1138 } 1139 1140 spin_unlock(&inode->i_lock); 1141 1142 if (dropping & CEPH_CAP_FILE_CACHE) { 1143 /* invalidate what we can */ 1144 dout("invalidating pages on %p\n", inode); 1145 invalidate_mapping_pages(&inode->i_data, 0, -1); 1146 } 1147 1148 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id, 1149 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq, 1150 size, max_size, &mtime, &atime, time_warp_seq, 1151 uid, gid, mode, 1152 xattr_version, 1153 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL, 1154 follows); 1155 if (ret < 0) { 1156 dout("error sending cap msg, must requeue %p\n", inode); 1157 delayed = 1; 1158 } 1159 1160 if (wake) 1161 wake_up(&ci->i_cap_wq); 1162 1163 return delayed; 1164 } 1165 1166 /* 1167 * When a snapshot is taken, clients accumulate dirty metadata on 1168 * inodes with capabilities in ceph_cap_snaps to describe the file 1169 * state at the time the snapshot was taken. This must be flushed 1170 * asynchronously back to the MDS once sync writes complete and dirty 1171 * data is written out. 1172 * 1173 * Called under i_lock. Takes s_mutex as needed. 1174 */ 1175 void __ceph_flush_snaps(struct ceph_inode_info *ci, 1176 struct ceph_mds_session **psession) 1177 { 1178 struct inode *inode = &ci->vfs_inode; 1179 int mds; 1180 struct ceph_cap_snap *capsnap; 1181 u32 mseq; 1182 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 1183 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold 1184 session->s_mutex */ 1185 u64 next_follows = 0; /* keep track of how far we've gotten through the 1186 i_cap_snaps list, and skip these entries next time 1187 around to avoid an infinite loop */ 1188 1189 if (psession) 1190 session = *psession; 1191 1192 dout("__flush_snaps %p\n", inode); 1193 retry: 1194 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 1195 /* avoid an infiniute loop after retry */ 1196 if (capsnap->follows < next_follows) 1197 continue; 1198 /* 1199 * we need to wait for sync writes to complete and for dirty 1200 * pages to be written out. 1201 */ 1202 if (capsnap->dirty_pages || capsnap->writing) 1203 continue; 1204 1205 /* pick mds, take s_mutex */ 1206 mds = __ceph_get_cap_mds(ci, &mseq); 1207 if (session && session->s_mds != mds) { 1208 dout("oops, wrong session %p mutex\n", session); 1209 mutex_unlock(&session->s_mutex); 1210 ceph_put_mds_session(session); 1211 session = NULL; 1212 } 1213 if (!session) { 1214 spin_unlock(&inode->i_lock); 1215 mutex_lock(&mdsc->mutex); 1216 session = __ceph_lookup_mds_session(mdsc, mds); 1217 mutex_unlock(&mdsc->mutex); 1218 if (session) { 1219 dout("inverting session/ino locks on %p\n", 1220 session); 1221 mutex_lock(&session->s_mutex); 1222 } 1223 /* 1224 * if session == NULL, we raced against a cap 1225 * deletion. retry, and we'll get a better 1226 * @mds value next time. 1227 */ 1228 spin_lock(&inode->i_lock); 1229 goto retry; 1230 } 1231 1232 capsnap->flush_tid = ++ci->i_cap_flush_last_tid; 1233 atomic_inc(&capsnap->nref); 1234 if (!list_empty(&capsnap->flushing_item)) 1235 list_del_init(&capsnap->flushing_item); 1236 list_add_tail(&capsnap->flushing_item, 1237 &session->s_cap_snaps_flushing); 1238 spin_unlock(&inode->i_lock); 1239 1240 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n", 1241 inode, capsnap, next_follows, capsnap->size); 1242 send_cap_msg(session, ceph_vino(inode).ino, 0, 1243 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0, 1244 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq, 1245 capsnap->size, 0, 1246 &capsnap->mtime, &capsnap->atime, 1247 capsnap->time_warp_seq, 1248 capsnap->uid, capsnap->gid, capsnap->mode, 1249 0, NULL, 1250 capsnap->follows); 1251 1252 next_follows = capsnap->follows + 1; 1253 ceph_put_cap_snap(capsnap); 1254 1255 spin_lock(&inode->i_lock); 1256 goto retry; 1257 } 1258 1259 /* we flushed them all; remove this inode from the queue */ 1260 spin_lock(&mdsc->snap_flush_lock); 1261 list_del_init(&ci->i_snap_flush_item); 1262 spin_unlock(&mdsc->snap_flush_lock); 1263 1264 if (psession) 1265 *psession = session; 1266 else if (session) { 1267 mutex_unlock(&session->s_mutex); 1268 ceph_put_mds_session(session); 1269 } 1270 } 1271 1272 static void ceph_flush_snaps(struct ceph_inode_info *ci) 1273 { 1274 struct inode *inode = &ci->vfs_inode; 1275 1276 spin_lock(&inode->i_lock); 1277 __ceph_flush_snaps(ci, NULL); 1278 spin_unlock(&inode->i_lock); 1279 } 1280 1281 /* 1282 * Mark caps dirty. If inode is newly dirty, add to the global dirty 1283 * list. 1284 */ 1285 void __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask) 1286 { 1287 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc; 1288 struct inode *inode = &ci->vfs_inode; 1289 int was = ci->i_dirty_caps; 1290 int dirty = 0; 1291 1292 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode, 1293 ceph_cap_string(mask), ceph_cap_string(was), 1294 ceph_cap_string(was | mask)); 1295 ci->i_dirty_caps |= mask; 1296 if (was == 0) { 1297 dout(" inode %p now dirty\n", &ci->vfs_inode); 1298 BUG_ON(!list_empty(&ci->i_dirty_item)); 1299 spin_lock(&mdsc->cap_dirty_lock); 1300 list_add(&ci->i_dirty_item, &mdsc->cap_dirty); 1301 spin_unlock(&mdsc->cap_dirty_lock); 1302 if (ci->i_flushing_caps == 0) { 1303 igrab(inode); 1304 dirty |= I_DIRTY_SYNC; 1305 } 1306 } 1307 BUG_ON(list_empty(&ci->i_dirty_item)); 1308 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) && 1309 (mask & CEPH_CAP_FILE_BUFFER)) 1310 dirty |= I_DIRTY_DATASYNC; 1311 if (dirty) 1312 __mark_inode_dirty(inode, dirty); 1313 __cap_delay_requeue(mdsc, ci); 1314 } 1315 1316 /* 1317 * Add dirty inode to the flushing list. Assigned a seq number so we 1318 * can wait for caps to flush without starving. 1319 * 1320 * Called under i_lock. 1321 */ 1322 static int __mark_caps_flushing(struct inode *inode, 1323 struct ceph_mds_session *session) 1324 { 1325 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1326 struct ceph_inode_info *ci = ceph_inode(inode); 1327 int flushing; 1328 1329 BUG_ON(ci->i_dirty_caps == 0); 1330 BUG_ON(list_empty(&ci->i_dirty_item)); 1331 1332 flushing = ci->i_dirty_caps; 1333 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", 1334 ceph_cap_string(flushing), 1335 ceph_cap_string(ci->i_flushing_caps), 1336 ceph_cap_string(ci->i_flushing_caps | flushing)); 1337 ci->i_flushing_caps |= flushing; 1338 ci->i_dirty_caps = 0; 1339 dout(" inode %p now !dirty\n", inode); 1340 1341 spin_lock(&mdsc->cap_dirty_lock); 1342 list_del_init(&ci->i_dirty_item); 1343 1344 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq; 1345 if (list_empty(&ci->i_flushing_item)) { 1346 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1347 mdsc->num_cap_flushing++; 1348 dout(" inode %p now flushing seq %lld\n", inode, 1349 ci->i_cap_flush_seq); 1350 } else { 1351 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1352 dout(" inode %p now flushing (more) seq %lld\n", inode, 1353 ci->i_cap_flush_seq); 1354 } 1355 spin_unlock(&mdsc->cap_dirty_lock); 1356 1357 return flushing; 1358 } 1359 1360 /* 1361 * Swiss army knife function to examine currently used and wanted 1362 * versus held caps. Release, flush, ack revoked caps to mds as 1363 * appropriate. 1364 * 1365 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay 1366 * cap release further. 1367 * CHECK_CAPS_AUTHONLY - we should only check the auth cap 1368 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without 1369 * further delay. 1370 */ 1371 void ceph_check_caps(struct ceph_inode_info *ci, int flags, 1372 struct ceph_mds_session *session) 1373 { 1374 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode); 1375 struct ceph_mds_client *mdsc = &client->mdsc; 1376 struct inode *inode = &ci->vfs_inode; 1377 struct ceph_cap *cap; 1378 int file_wanted, used; 1379 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ 1380 int drop_session_lock = session ? 0 : 1; 1381 int want, retain, revoking, flushing = 0; 1382 int mds = -1; /* keep track of how far we've gone through i_caps list 1383 to avoid an infinite loop on retry */ 1384 struct rb_node *p; 1385 int tried_invalidate = 0; 1386 int delayed = 0, sent = 0, force_requeue = 0, num; 1387 int is_delayed = flags & CHECK_CAPS_NODELAY; 1388 1389 /* if we are unmounting, flush any unused caps immediately. */ 1390 if (mdsc->stopping) 1391 is_delayed = 1; 1392 1393 spin_lock(&inode->i_lock); 1394 1395 if (ci->i_ceph_flags & CEPH_I_FLUSH) 1396 flags |= CHECK_CAPS_FLUSH; 1397 1398 /* flush snaps first time around only */ 1399 if (!list_empty(&ci->i_cap_snaps)) 1400 __ceph_flush_snaps(ci, &session); 1401 goto retry_locked; 1402 retry: 1403 spin_lock(&inode->i_lock); 1404 retry_locked: 1405 file_wanted = __ceph_caps_file_wanted(ci); 1406 used = __ceph_caps_used(ci); 1407 want = file_wanted | used; 1408 1409 retain = want | CEPH_CAP_PIN; 1410 if (!mdsc->stopping && inode->i_nlink > 0) { 1411 if (want) { 1412 retain |= CEPH_CAP_ANY; /* be greedy */ 1413 } else { 1414 retain |= CEPH_CAP_ANY_SHARED; 1415 /* 1416 * keep RD only if we didn't have the file open RW, 1417 * because then the mds would revoke it anyway to 1418 * journal max_size=0. 1419 */ 1420 if (ci->i_max_size == 0) 1421 retain |= CEPH_CAP_ANY_RD; 1422 } 1423 } 1424 1425 dout("check_caps %p file_want %s used %s dirty %s flushing %s" 1426 " issued %s retain %s %s%s%s\n", inode, 1427 ceph_cap_string(file_wanted), 1428 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps), 1429 ceph_cap_string(ci->i_flushing_caps), 1430 ceph_cap_string(__ceph_caps_issued(ci, NULL)), 1431 ceph_cap_string(retain), 1432 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "", 1433 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "", 1434 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : ""); 1435 1436 /* 1437 * If we no longer need to hold onto old our caps, and we may 1438 * have cached pages, but don't want them, then try to invalidate. 1439 * If we fail, it's because pages are locked.... try again later. 1440 */ 1441 if ((!is_delayed || mdsc->stopping) && 1442 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */ 1443 ci->i_rdcache_gen && /* may have cached pages */ 1444 file_wanted == 0 && /* no open files */ 1445 !ci->i_truncate_pending && 1446 !tried_invalidate) { 1447 u32 invalidating_gen = ci->i_rdcache_gen; 1448 int ret; 1449 1450 dout("check_caps trying to invalidate on %p\n", inode); 1451 spin_unlock(&inode->i_lock); 1452 ret = invalidate_inode_pages2(&inode->i_data); 1453 spin_lock(&inode->i_lock); 1454 if (ret == 0 && invalidating_gen == ci->i_rdcache_gen) { 1455 /* success. */ 1456 ci->i_rdcache_gen = 0; 1457 ci->i_rdcache_revoking = 0; 1458 } else { 1459 dout("check_caps failed to invalidate pages\n"); 1460 /* we failed to invalidate pages. check these 1461 caps again later. */ 1462 force_requeue = 1; 1463 __cap_set_timeouts(mdsc, ci); 1464 } 1465 tried_invalidate = 1; 1466 goto retry_locked; 1467 } 1468 1469 num = 0; 1470 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 1471 cap = rb_entry(p, struct ceph_cap, ci_node); 1472 num++; 1473 1474 /* avoid looping forever */ 1475 if (mds >= cap->mds || 1476 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap)) 1477 continue; 1478 1479 /* NOTE: no side-effects allowed, until we take s_mutex */ 1480 1481 revoking = cap->implemented & ~cap->issued; 1482 if (revoking) 1483 dout("mds%d revoking %s\n", cap->mds, 1484 ceph_cap_string(revoking)); 1485 1486 if (cap == ci->i_auth_cap && 1487 (cap->issued & CEPH_CAP_FILE_WR)) { 1488 /* request larger max_size from MDS? */ 1489 if (ci->i_wanted_max_size > ci->i_max_size && 1490 ci->i_wanted_max_size > ci->i_requested_max_size) { 1491 dout("requesting new max_size\n"); 1492 goto ack; 1493 } 1494 1495 /* approaching file_max? */ 1496 if ((inode->i_size << 1) >= ci->i_max_size && 1497 (ci->i_reported_size << 1) < ci->i_max_size) { 1498 dout("i_size approaching max_size\n"); 1499 goto ack; 1500 } 1501 } 1502 /* flush anything dirty? */ 1503 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) && 1504 ci->i_dirty_caps) { 1505 dout("flushing dirty caps\n"); 1506 goto ack; 1507 } 1508 1509 /* completed revocation? going down and there are no caps? */ 1510 if (revoking && (revoking & used) == 0) { 1511 dout("completed revocation of %s\n", 1512 ceph_cap_string(cap->implemented & ~cap->issued)); 1513 goto ack; 1514 } 1515 1516 /* want more caps from mds? */ 1517 if (want & ~(cap->mds_wanted | cap->issued)) 1518 goto ack; 1519 1520 /* things we might delay */ 1521 if ((cap->issued & ~retain) == 0 && 1522 cap->mds_wanted == want) 1523 continue; /* nope, all good */ 1524 1525 if (is_delayed) 1526 goto ack; 1527 1528 /* delay? */ 1529 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1530 time_before(jiffies, ci->i_hold_caps_max)) { 1531 dout(" delaying issued %s -> %s, wanted %s -> %s\n", 1532 ceph_cap_string(cap->issued), 1533 ceph_cap_string(cap->issued & retain), 1534 ceph_cap_string(cap->mds_wanted), 1535 ceph_cap_string(want)); 1536 delayed++; 1537 continue; 1538 } 1539 1540 ack: 1541 if (session && session != cap->session) { 1542 dout("oops, wrong session %p mutex\n", session); 1543 mutex_unlock(&session->s_mutex); 1544 session = NULL; 1545 } 1546 if (!session) { 1547 session = cap->session; 1548 if (mutex_trylock(&session->s_mutex) == 0) { 1549 dout("inverting session/ino locks on %p\n", 1550 session); 1551 spin_unlock(&inode->i_lock); 1552 if (took_snap_rwsem) { 1553 up_read(&mdsc->snap_rwsem); 1554 took_snap_rwsem = 0; 1555 } 1556 mutex_lock(&session->s_mutex); 1557 goto retry; 1558 } 1559 } 1560 /* take snap_rwsem after session mutex */ 1561 if (!took_snap_rwsem) { 1562 if (down_read_trylock(&mdsc->snap_rwsem) == 0) { 1563 dout("inverting snap/in locks on %p\n", 1564 inode); 1565 spin_unlock(&inode->i_lock); 1566 down_read(&mdsc->snap_rwsem); 1567 took_snap_rwsem = 1; 1568 goto retry; 1569 } 1570 took_snap_rwsem = 1; 1571 } 1572 1573 if (cap == ci->i_auth_cap && ci->i_dirty_caps) 1574 flushing = __mark_caps_flushing(inode, session); 1575 1576 mds = cap->mds; /* remember mds, so we don't repeat */ 1577 sent++; 1578 1579 /* __send_cap drops i_lock */ 1580 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want, 1581 retain, flushing, NULL); 1582 goto retry; /* retake i_lock and restart our cap scan. */ 1583 } 1584 1585 /* 1586 * Reschedule delayed caps release if we delayed anything, 1587 * otherwise cancel. 1588 */ 1589 if (delayed && is_delayed) 1590 force_requeue = 1; /* __send_cap delayed release; requeue */ 1591 if (!delayed && !is_delayed) 1592 __cap_delay_cancel(mdsc, ci); 1593 else if (!is_delayed || force_requeue) 1594 __cap_delay_requeue(mdsc, ci); 1595 1596 spin_unlock(&inode->i_lock); 1597 1598 if (session && drop_session_lock) 1599 mutex_unlock(&session->s_mutex); 1600 if (took_snap_rwsem) 1601 up_read(&mdsc->snap_rwsem); 1602 } 1603 1604 /* 1605 * Try to flush dirty caps back to the auth mds. 1606 */ 1607 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session, 1608 unsigned *flush_tid) 1609 { 1610 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1611 struct ceph_inode_info *ci = ceph_inode(inode); 1612 int unlock_session = session ? 0 : 1; 1613 int flushing = 0; 1614 1615 retry: 1616 spin_lock(&inode->i_lock); 1617 if (ci->i_dirty_caps && ci->i_auth_cap) { 1618 struct ceph_cap *cap = ci->i_auth_cap; 1619 int used = __ceph_caps_used(ci); 1620 int want = __ceph_caps_wanted(ci); 1621 int delayed; 1622 1623 if (!session) { 1624 spin_unlock(&inode->i_lock); 1625 session = cap->session; 1626 mutex_lock(&session->s_mutex); 1627 goto retry; 1628 } 1629 BUG_ON(session != cap->session); 1630 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) 1631 goto out; 1632 1633 flushing = __mark_caps_flushing(inode, session); 1634 1635 /* __send_cap drops i_lock */ 1636 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want, 1637 cap->issued | cap->implemented, flushing, 1638 flush_tid); 1639 if (!delayed) 1640 goto out_unlocked; 1641 1642 spin_lock(&inode->i_lock); 1643 __cap_delay_requeue(mdsc, ci); 1644 } 1645 out: 1646 spin_unlock(&inode->i_lock); 1647 out_unlocked: 1648 if (session && unlock_session) 1649 mutex_unlock(&session->s_mutex); 1650 return flushing; 1651 } 1652 1653 /* 1654 * Return true if we've flushed caps through the given flush_tid. 1655 */ 1656 static int caps_are_flushed(struct inode *inode, unsigned tid) 1657 { 1658 struct ceph_inode_info *ci = ceph_inode(inode); 1659 int dirty, i, ret = 1; 1660 1661 spin_lock(&inode->i_lock); 1662 dirty = __ceph_caps_dirty(ci); 1663 for (i = 0; i < CEPH_CAP_BITS; i++) 1664 if ((ci->i_flushing_caps & (1 << i)) && 1665 ci->i_cap_flush_tid[i] <= tid) { 1666 /* still flushing this bit */ 1667 ret = 0; 1668 break; 1669 } 1670 spin_unlock(&inode->i_lock); 1671 return ret; 1672 } 1673 1674 /* 1675 * Wait on any unsafe replies for the given inode. First wait on the 1676 * newest request, and make that the upper bound. Then, if there are 1677 * more requests, keep waiting on the oldest as long as it is still older 1678 * than the original request. 1679 */ 1680 static void sync_write_wait(struct inode *inode) 1681 { 1682 struct ceph_inode_info *ci = ceph_inode(inode); 1683 struct list_head *head = &ci->i_unsafe_writes; 1684 struct ceph_osd_request *req; 1685 u64 last_tid; 1686 1687 spin_lock(&ci->i_unsafe_lock); 1688 if (list_empty(head)) 1689 goto out; 1690 1691 /* set upper bound as _last_ entry in chain */ 1692 req = list_entry(head->prev, struct ceph_osd_request, 1693 r_unsafe_item); 1694 last_tid = req->r_tid; 1695 1696 do { 1697 ceph_osdc_get_request(req); 1698 spin_unlock(&ci->i_unsafe_lock); 1699 dout("sync_write_wait on tid %llu (until %llu)\n", 1700 req->r_tid, last_tid); 1701 wait_for_completion(&req->r_safe_completion); 1702 spin_lock(&ci->i_unsafe_lock); 1703 ceph_osdc_put_request(req); 1704 1705 /* 1706 * from here on look at first entry in chain, since we 1707 * only want to wait for anything older than last_tid 1708 */ 1709 if (list_empty(head)) 1710 break; 1711 req = list_entry(head->next, struct ceph_osd_request, 1712 r_unsafe_item); 1713 } while (req->r_tid < last_tid); 1714 out: 1715 spin_unlock(&ci->i_unsafe_lock); 1716 } 1717 1718 int ceph_fsync(struct file *file, struct dentry *dentry, int datasync) 1719 { 1720 struct inode *inode = dentry->d_inode; 1721 struct ceph_inode_info *ci = ceph_inode(inode); 1722 unsigned flush_tid; 1723 int ret; 1724 int dirty; 1725 1726 dout("fsync %p%s\n", inode, datasync ? " datasync" : ""); 1727 sync_write_wait(inode); 1728 1729 ret = filemap_write_and_wait(inode->i_mapping); 1730 if (ret < 0) 1731 return ret; 1732 1733 dirty = try_flush_caps(inode, NULL, &flush_tid); 1734 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); 1735 1736 /* 1737 * only wait on non-file metadata writeback (the mds 1738 * can recover size and mtime, so we don't need to 1739 * wait for that) 1740 */ 1741 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { 1742 dout("fsync waiting for flush_tid %u\n", flush_tid); 1743 ret = wait_event_interruptible(ci->i_cap_wq, 1744 caps_are_flushed(inode, flush_tid)); 1745 } 1746 1747 dout("fsync %p%s done\n", inode, datasync ? " datasync" : ""); 1748 return ret; 1749 } 1750 1751 /* 1752 * Flush any dirty caps back to the mds. If we aren't asked to wait, 1753 * queue inode for flush but don't do so immediately, because we can 1754 * get by with fewer MDS messages if we wait for data writeback to 1755 * complete first. 1756 */ 1757 int ceph_write_inode(struct inode *inode, int wait) 1758 { 1759 struct ceph_inode_info *ci = ceph_inode(inode); 1760 unsigned flush_tid; 1761 int err = 0; 1762 int dirty; 1763 1764 dout("write_inode %p wait=%d\n", inode, wait); 1765 if (wait) { 1766 dirty = try_flush_caps(inode, NULL, &flush_tid); 1767 if (dirty) 1768 err = wait_event_interruptible(ci->i_cap_wq, 1769 caps_are_flushed(inode, flush_tid)); 1770 } else { 1771 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 1772 1773 spin_lock(&inode->i_lock); 1774 if (__ceph_caps_dirty(ci)) 1775 __cap_delay_requeue_front(mdsc, ci); 1776 spin_unlock(&inode->i_lock); 1777 } 1778 return err; 1779 } 1780 1781 /* 1782 * After a recovering MDS goes active, we need to resend any caps 1783 * we were flushing. 1784 * 1785 * Caller holds session->s_mutex. 1786 */ 1787 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc, 1788 struct ceph_mds_session *session) 1789 { 1790 struct ceph_cap_snap *capsnap; 1791 1792 dout("kick_flushing_capsnaps mds%d\n", session->s_mds); 1793 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing, 1794 flushing_item) { 1795 struct ceph_inode_info *ci = capsnap->ci; 1796 struct inode *inode = &ci->vfs_inode; 1797 struct ceph_cap *cap; 1798 1799 spin_lock(&inode->i_lock); 1800 cap = ci->i_auth_cap; 1801 if (cap && cap->session == session) { 1802 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode, 1803 cap, capsnap); 1804 __ceph_flush_snaps(ci, &session); 1805 } else { 1806 pr_err("%p auth cap %p not mds%d ???\n", inode, 1807 cap, session->s_mds); 1808 spin_unlock(&inode->i_lock); 1809 } 1810 } 1811 } 1812 1813 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 1814 struct ceph_mds_session *session) 1815 { 1816 struct ceph_inode_info *ci; 1817 1818 kick_flushing_capsnaps(mdsc, session); 1819 1820 dout("kick_flushing_caps mds%d\n", session->s_mds); 1821 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 1822 struct inode *inode = &ci->vfs_inode; 1823 struct ceph_cap *cap; 1824 int delayed = 0; 1825 1826 spin_lock(&inode->i_lock); 1827 cap = ci->i_auth_cap; 1828 if (cap && cap->session == session) { 1829 dout("kick_flushing_caps %p cap %p %s\n", inode, 1830 cap, ceph_cap_string(ci->i_flushing_caps)); 1831 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, 1832 __ceph_caps_used(ci), 1833 __ceph_caps_wanted(ci), 1834 cap->issued | cap->implemented, 1835 ci->i_flushing_caps, NULL); 1836 if (delayed) { 1837 spin_lock(&inode->i_lock); 1838 __cap_delay_requeue(mdsc, ci); 1839 spin_unlock(&inode->i_lock); 1840 } 1841 } else { 1842 pr_err("%p auth cap %p not mds%d ???\n", inode, 1843 cap, session->s_mds); 1844 spin_unlock(&inode->i_lock); 1845 } 1846 } 1847 } 1848 1849 1850 /* 1851 * Take references to capabilities we hold, so that we don't release 1852 * them to the MDS prematurely. 1853 * 1854 * Protected by i_lock. 1855 */ 1856 static void __take_cap_refs(struct ceph_inode_info *ci, int got) 1857 { 1858 if (got & CEPH_CAP_PIN) 1859 ci->i_pin_ref++; 1860 if (got & CEPH_CAP_FILE_RD) 1861 ci->i_rd_ref++; 1862 if (got & CEPH_CAP_FILE_CACHE) 1863 ci->i_rdcache_ref++; 1864 if (got & CEPH_CAP_FILE_WR) 1865 ci->i_wr_ref++; 1866 if (got & CEPH_CAP_FILE_BUFFER) { 1867 if (ci->i_wrbuffer_ref == 0) 1868 igrab(&ci->vfs_inode); 1869 ci->i_wrbuffer_ref++; 1870 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n", 1871 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref); 1872 } 1873 } 1874 1875 /* 1876 * Try to grab cap references. Specify those refs we @want, and the 1877 * minimal set we @need. Also include the larger offset we are writing 1878 * to (when applicable), and check against max_size here as well. 1879 * Note that caller is responsible for ensuring max_size increases are 1880 * requested from the MDS. 1881 */ 1882 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want, 1883 int *got, loff_t endoff, int *check_max, int *err) 1884 { 1885 struct inode *inode = &ci->vfs_inode; 1886 int ret = 0; 1887 int have, implemented; 1888 1889 dout("get_cap_refs %p need %s want %s\n", inode, 1890 ceph_cap_string(need), ceph_cap_string(want)); 1891 spin_lock(&inode->i_lock); 1892 1893 /* make sure we _have_ some caps! */ 1894 if (!__ceph_is_any_caps(ci)) { 1895 dout("get_cap_refs %p no real caps\n", inode); 1896 *err = -EBADF; 1897 ret = 1; 1898 goto out; 1899 } 1900 1901 if (need & CEPH_CAP_FILE_WR) { 1902 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) { 1903 dout("get_cap_refs %p endoff %llu > maxsize %llu\n", 1904 inode, endoff, ci->i_max_size); 1905 if (endoff > ci->i_wanted_max_size) { 1906 *check_max = 1; 1907 ret = 1; 1908 } 1909 goto out; 1910 } 1911 /* 1912 * If a sync write is in progress, we must wait, so that we 1913 * can get a final snapshot value for size+mtime. 1914 */ 1915 if (__ceph_have_pending_cap_snap(ci)) { 1916 dout("get_cap_refs %p cap_snap_pending\n", inode); 1917 goto out; 1918 } 1919 } 1920 have = __ceph_caps_issued(ci, &implemented); 1921 1922 /* 1923 * disallow writes while a truncate is pending 1924 */ 1925 if (ci->i_truncate_pending) 1926 have &= ~CEPH_CAP_FILE_WR; 1927 1928 if ((have & need) == need) { 1929 /* 1930 * Look at (implemented & ~have & not) so that we keep waiting 1931 * on transition from wanted -> needed caps. This is needed 1932 * for WRBUFFER|WR -> WR to avoid a new WR sync write from 1933 * going before a prior buffered writeback happens. 1934 */ 1935 int not = want & ~(have & need); 1936 int revoking = implemented & ~have; 1937 dout("get_cap_refs %p have %s but not %s (revoking %s)\n", 1938 inode, ceph_cap_string(have), ceph_cap_string(not), 1939 ceph_cap_string(revoking)); 1940 if ((revoking & not) == 0) { 1941 *got = need | (have & want); 1942 __take_cap_refs(ci, *got); 1943 ret = 1; 1944 } 1945 } else { 1946 dout("get_cap_refs %p have %s needed %s\n", inode, 1947 ceph_cap_string(have), ceph_cap_string(need)); 1948 } 1949 out: 1950 spin_unlock(&inode->i_lock); 1951 dout("get_cap_refs %p ret %d got %s\n", inode, 1952 ret, ceph_cap_string(*got)); 1953 return ret; 1954 } 1955 1956 /* 1957 * Check the offset we are writing up to against our current 1958 * max_size. If necessary, tell the MDS we want to write to 1959 * a larger offset. 1960 */ 1961 static void check_max_size(struct inode *inode, loff_t endoff) 1962 { 1963 struct ceph_inode_info *ci = ceph_inode(inode); 1964 int check = 0; 1965 1966 /* do we need to explicitly request a larger max_size? */ 1967 spin_lock(&inode->i_lock); 1968 if ((endoff >= ci->i_max_size || 1969 endoff > (inode->i_size << 1)) && 1970 endoff > ci->i_wanted_max_size) { 1971 dout("write %p at large endoff %llu, req max_size\n", 1972 inode, endoff); 1973 ci->i_wanted_max_size = endoff; 1974 check = 1; 1975 } 1976 spin_unlock(&inode->i_lock); 1977 if (check) 1978 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 1979 } 1980 1981 /* 1982 * Wait for caps, and take cap references. If we can't get a WR cap 1983 * due to a small max_size, make sure we check_max_size (and possibly 1984 * ask the mds) so we don't get hung up indefinitely. 1985 */ 1986 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got, 1987 loff_t endoff) 1988 { 1989 int check_max, ret, err; 1990 1991 retry: 1992 if (endoff > 0) 1993 check_max_size(&ci->vfs_inode, endoff); 1994 check_max = 0; 1995 err = 0; 1996 ret = wait_event_interruptible(ci->i_cap_wq, 1997 try_get_cap_refs(ci, need, want, 1998 got, endoff, 1999 &check_max, &err)); 2000 if (err) 2001 ret = err; 2002 if (check_max) 2003 goto retry; 2004 return ret; 2005 } 2006 2007 /* 2008 * Take cap refs. Caller must already know we hold at least one ref 2009 * on the caps in question or we don't know this is safe. 2010 */ 2011 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) 2012 { 2013 spin_lock(&ci->vfs_inode.i_lock); 2014 __take_cap_refs(ci, caps); 2015 spin_unlock(&ci->vfs_inode.i_lock); 2016 } 2017 2018 /* 2019 * Release cap refs. 2020 * 2021 * If we released the last ref on any given cap, call ceph_check_caps 2022 * to release (or schedule a release). 2023 * 2024 * If we are releasing a WR cap (from a sync write), finalize any affected 2025 * cap_snap, and wake up any waiters. 2026 */ 2027 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had) 2028 { 2029 struct inode *inode = &ci->vfs_inode; 2030 int last = 0, put = 0, flushsnaps = 0, wake = 0; 2031 struct ceph_cap_snap *capsnap; 2032 2033 spin_lock(&inode->i_lock); 2034 if (had & CEPH_CAP_PIN) 2035 --ci->i_pin_ref; 2036 if (had & CEPH_CAP_FILE_RD) 2037 if (--ci->i_rd_ref == 0) 2038 last++; 2039 if (had & CEPH_CAP_FILE_CACHE) 2040 if (--ci->i_rdcache_ref == 0) 2041 last++; 2042 if (had & CEPH_CAP_FILE_BUFFER) { 2043 if (--ci->i_wrbuffer_ref == 0) { 2044 last++; 2045 put++; 2046 } 2047 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n", 2048 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref); 2049 } 2050 if (had & CEPH_CAP_FILE_WR) 2051 if (--ci->i_wr_ref == 0) { 2052 last++; 2053 if (!list_empty(&ci->i_cap_snaps)) { 2054 capsnap = list_first_entry(&ci->i_cap_snaps, 2055 struct ceph_cap_snap, 2056 ci_item); 2057 if (capsnap->writing) { 2058 capsnap->writing = 0; 2059 flushsnaps = 2060 __ceph_finish_cap_snap(ci, 2061 capsnap); 2062 wake = 1; 2063 } 2064 } 2065 } 2066 spin_unlock(&inode->i_lock); 2067 2068 dout("put_cap_refs %p had %s %s\n", inode, ceph_cap_string(had), 2069 last ? "last" : ""); 2070 2071 if (last && !flushsnaps) 2072 ceph_check_caps(ci, 0, NULL); 2073 else if (flushsnaps) 2074 ceph_flush_snaps(ci); 2075 if (wake) 2076 wake_up(&ci->i_cap_wq); 2077 if (put) 2078 iput(inode); 2079 } 2080 2081 /* 2082 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap 2083 * context. Adjust per-snap dirty page accounting as appropriate. 2084 * Once all dirty data for a cap_snap is flushed, flush snapped file 2085 * metadata back to the MDS. If we dropped the last ref, call 2086 * ceph_check_caps. 2087 */ 2088 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr, 2089 struct ceph_snap_context *snapc) 2090 { 2091 struct inode *inode = &ci->vfs_inode; 2092 int last = 0; 2093 int last_snap = 0; 2094 int found = 0; 2095 struct ceph_cap_snap *capsnap = NULL; 2096 2097 spin_lock(&inode->i_lock); 2098 ci->i_wrbuffer_ref -= nr; 2099 last = !ci->i_wrbuffer_ref; 2100 2101 if (ci->i_head_snapc == snapc) { 2102 ci->i_wrbuffer_ref_head -= nr; 2103 if (!ci->i_wrbuffer_ref_head) { 2104 ceph_put_snap_context(ci->i_head_snapc); 2105 ci->i_head_snapc = NULL; 2106 } 2107 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n", 2108 inode, 2109 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr, 2110 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 2111 last ? " LAST" : ""); 2112 } else { 2113 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2114 if (capsnap->context == snapc) { 2115 found = 1; 2116 capsnap->dirty_pages -= nr; 2117 last_snap = !capsnap->dirty_pages; 2118 break; 2119 } 2120 } 2121 BUG_ON(!found); 2122 dout("put_wrbuffer_cap_refs on %p cap_snap %p " 2123 " snap %lld %d/%d -> %d/%d %s%s\n", 2124 inode, capsnap, capsnap->context->seq, 2125 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, 2126 ci->i_wrbuffer_ref, capsnap->dirty_pages, 2127 last ? " (wrbuffer last)" : "", 2128 last_snap ? " (capsnap last)" : ""); 2129 } 2130 2131 spin_unlock(&inode->i_lock); 2132 2133 if (last) { 2134 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2135 iput(inode); 2136 } else if (last_snap) { 2137 ceph_flush_snaps(ci); 2138 wake_up(&ci->i_cap_wq); 2139 } 2140 } 2141 2142 /* 2143 * Handle a cap GRANT message from the MDS. (Note that a GRANT may 2144 * actually be a revocation if it specifies a smaller cap set.) 2145 * 2146 * caller holds s_mutex. 2147 * return value: 2148 * 0 - ok 2149 * 1 - check_caps on auth cap only (writeback) 2150 * 2 - check_caps (ack revoke) 2151 */ 2152 static int handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant, 2153 struct ceph_mds_session *session, 2154 struct ceph_cap *cap, 2155 struct ceph_buffer *xattr_buf) 2156 __releases(inode->i_lock) 2157 2158 { 2159 struct ceph_inode_info *ci = ceph_inode(inode); 2160 int mds = session->s_mds; 2161 int seq = le32_to_cpu(grant->seq); 2162 int newcaps = le32_to_cpu(grant->caps); 2163 int issued, implemented, used, wanted, dirty; 2164 u64 size = le64_to_cpu(grant->size); 2165 u64 max_size = le64_to_cpu(grant->max_size); 2166 struct timespec mtime, atime, ctime; 2167 int reply = 0; 2168 int wake = 0; 2169 int writeback = 0; 2170 int revoked_rdcache = 0; 2171 int invalidate_async = 0; 2172 int tried_invalidate = 0; 2173 int ret; 2174 2175 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n", 2176 inode, cap, mds, seq, ceph_cap_string(newcaps)); 2177 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size, 2178 inode->i_size); 2179 2180 /* 2181 * If CACHE is being revoked, and we have no dirty buffers, 2182 * try to invalidate (once). (If there are dirty buffers, we 2183 * will invalidate _after_ writeback.) 2184 */ 2185 restart: 2186 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && 2187 !ci->i_wrbuffer_ref && !tried_invalidate) { 2188 dout("CACHE invalidation\n"); 2189 spin_unlock(&inode->i_lock); 2190 tried_invalidate = 1; 2191 2192 ret = invalidate_inode_pages2(&inode->i_data); 2193 spin_lock(&inode->i_lock); 2194 if (ret < 0) { 2195 /* there were locked pages.. invalidate later 2196 in a separate thread. */ 2197 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) { 2198 invalidate_async = 1; 2199 ci->i_rdcache_revoking = ci->i_rdcache_gen; 2200 } 2201 } else { 2202 /* we successfully invalidated those pages */ 2203 revoked_rdcache = 1; 2204 ci->i_rdcache_gen = 0; 2205 ci->i_rdcache_revoking = 0; 2206 } 2207 goto restart; 2208 } 2209 2210 /* side effects now are allowed */ 2211 2212 issued = __ceph_caps_issued(ci, &implemented); 2213 issued |= implemented | __ceph_caps_dirty(ci); 2214 2215 cap->cap_gen = session->s_cap_gen; 2216 cap->recon_gen = session->s_recon_gen; 2217 2218 __check_cap_issue(ci, cap, newcaps); 2219 2220 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) { 2221 inode->i_mode = le32_to_cpu(grant->mode); 2222 inode->i_uid = le32_to_cpu(grant->uid); 2223 inode->i_gid = le32_to_cpu(grant->gid); 2224 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode, 2225 inode->i_uid, inode->i_gid); 2226 } 2227 2228 if ((issued & CEPH_CAP_LINK_EXCL) == 0) 2229 inode->i_nlink = le32_to_cpu(grant->nlink); 2230 2231 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) { 2232 int len = le32_to_cpu(grant->xattr_len); 2233 u64 version = le64_to_cpu(grant->xattr_version); 2234 2235 if (version > ci->i_xattrs.version) { 2236 dout(" got new xattrs v%llu on %p len %d\n", 2237 version, inode, len); 2238 if (ci->i_xattrs.blob) 2239 ceph_buffer_put(ci->i_xattrs.blob); 2240 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf); 2241 ci->i_xattrs.version = version; 2242 } 2243 } 2244 2245 /* size/ctime/mtime/atime? */ 2246 ceph_fill_file_size(inode, issued, 2247 le32_to_cpu(grant->truncate_seq), 2248 le64_to_cpu(grant->truncate_size), size); 2249 ceph_decode_timespec(&mtime, &grant->mtime); 2250 ceph_decode_timespec(&atime, &grant->atime); 2251 ceph_decode_timespec(&ctime, &grant->ctime); 2252 ceph_fill_file_time(inode, issued, 2253 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime, 2254 &atime); 2255 2256 /* max size increase? */ 2257 if (max_size != ci->i_max_size) { 2258 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size); 2259 ci->i_max_size = max_size; 2260 if (max_size >= ci->i_wanted_max_size) { 2261 ci->i_wanted_max_size = 0; /* reset */ 2262 ci->i_requested_max_size = 0; 2263 } 2264 wake = 1; 2265 } 2266 2267 /* check cap bits */ 2268 wanted = __ceph_caps_wanted(ci); 2269 used = __ceph_caps_used(ci); 2270 dirty = __ceph_caps_dirty(ci); 2271 dout(" my wanted = %s, used = %s, dirty %s\n", 2272 ceph_cap_string(wanted), 2273 ceph_cap_string(used), 2274 ceph_cap_string(dirty)); 2275 if (wanted != le32_to_cpu(grant->wanted)) { 2276 dout("mds wanted %s -> %s\n", 2277 ceph_cap_string(le32_to_cpu(grant->wanted)), 2278 ceph_cap_string(wanted)); 2279 grant->wanted = cpu_to_le32(wanted); 2280 } 2281 2282 cap->seq = seq; 2283 2284 /* file layout may have changed */ 2285 ci->i_layout = grant->layout; 2286 2287 /* revocation, grant, or no-op? */ 2288 if (cap->issued & ~newcaps) { 2289 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued), 2290 ceph_cap_string(newcaps)); 2291 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER) 2292 writeback = 1; /* will delay ack */ 2293 else if (dirty & ~newcaps) 2294 reply = 1; /* initiate writeback in check_caps */ 2295 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 || 2296 revoked_rdcache) 2297 reply = 2; /* send revoke ack in check_caps */ 2298 cap->issued = newcaps; 2299 } else if (cap->issued == newcaps) { 2300 dout("caps unchanged: %s -> %s\n", 2301 ceph_cap_string(cap->issued), ceph_cap_string(newcaps)); 2302 } else { 2303 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued), 2304 ceph_cap_string(newcaps)); 2305 cap->issued = newcaps; 2306 cap->implemented |= newcaps; /* add bits only, to 2307 * avoid stepping on a 2308 * pending revocation */ 2309 wake = 1; 2310 } 2311 2312 spin_unlock(&inode->i_lock); 2313 if (writeback) { 2314 /* 2315 * queue inode for writeback: we can't actually call 2316 * filemap_write_and_wait, etc. from message handler 2317 * context. 2318 */ 2319 dout("queueing %p for writeback\n", inode); 2320 if (ceph_queue_writeback(inode)) 2321 igrab(inode); 2322 } 2323 if (invalidate_async) { 2324 dout("queueing %p for page invalidation\n", inode); 2325 if (ceph_queue_page_invalidation(inode)) 2326 igrab(inode); 2327 } 2328 if (wake) 2329 wake_up(&ci->i_cap_wq); 2330 return reply; 2331 } 2332 2333 /* 2334 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the 2335 * MDS has been safely committed. 2336 */ 2337 static void handle_cap_flush_ack(struct inode *inode, 2338 struct ceph_mds_caps *m, 2339 struct ceph_mds_session *session, 2340 struct ceph_cap *cap) 2341 __releases(inode->i_lock) 2342 { 2343 struct ceph_inode_info *ci = ceph_inode(inode); 2344 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc; 2345 unsigned seq = le32_to_cpu(m->seq); 2346 int dirty = le32_to_cpu(m->dirty); 2347 int cleaned = 0; 2348 u64 flush_tid = le64_to_cpu(m->client_tid); 2349 int drop = 0; 2350 int i; 2351 2352 for (i = 0; i < CEPH_CAP_BITS; i++) 2353 if ((dirty & (1 << i)) && 2354 flush_tid == ci->i_cap_flush_tid[i]) 2355 cleaned |= 1 << i; 2356 2357 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," 2358 " flushing %s -> %s\n", 2359 inode, session->s_mds, seq, ceph_cap_string(dirty), 2360 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), 2361 ceph_cap_string(ci->i_flushing_caps & ~cleaned)); 2362 2363 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned)) 2364 goto out; 2365 2366 ci->i_flushing_caps &= ~cleaned; 2367 2368 spin_lock(&mdsc->cap_dirty_lock); 2369 if (ci->i_flushing_caps == 0) { 2370 list_del_init(&ci->i_flushing_item); 2371 if (!list_empty(&session->s_cap_flushing)) 2372 dout(" mds%d still flushing cap on %p\n", 2373 session->s_mds, 2374 &list_entry(session->s_cap_flushing.next, 2375 struct ceph_inode_info, 2376 i_flushing_item)->vfs_inode); 2377 mdsc->num_cap_flushing--; 2378 wake_up(&mdsc->cap_flushing_wq); 2379 dout(" inode %p now !flushing\n", inode); 2380 2381 if (ci->i_dirty_caps == 0) { 2382 dout(" inode %p now clean\n", inode); 2383 BUG_ON(!list_empty(&ci->i_dirty_item)); 2384 drop = 1; 2385 } else { 2386 BUG_ON(list_empty(&ci->i_dirty_item)); 2387 } 2388 } 2389 spin_unlock(&mdsc->cap_dirty_lock); 2390 wake_up(&ci->i_cap_wq); 2391 2392 out: 2393 spin_unlock(&inode->i_lock); 2394 if (drop) 2395 iput(inode); 2396 } 2397 2398 /* 2399 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can 2400 * throw away our cap_snap. 2401 * 2402 * Caller hold s_mutex. 2403 */ 2404 static void handle_cap_flushsnap_ack(struct inode *inode, 2405 struct ceph_mds_caps *m, 2406 struct ceph_mds_session *session) 2407 { 2408 struct ceph_inode_info *ci = ceph_inode(inode); 2409 u64 follows = le64_to_cpu(m->snap_follows); 2410 u64 flush_tid = le64_to_cpu(m->client_tid); 2411 struct ceph_cap_snap *capsnap; 2412 int drop = 0; 2413 2414 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n", 2415 inode, ci, session->s_mds, follows); 2416 2417 spin_lock(&inode->i_lock); 2418 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2419 if (capsnap->follows == follows) { 2420 if (capsnap->flush_tid != flush_tid) { 2421 dout(" cap_snap %p follows %lld tid %lld !=" 2422 " %lld\n", capsnap, follows, 2423 flush_tid, capsnap->flush_tid); 2424 break; 2425 } 2426 WARN_ON(capsnap->dirty_pages || capsnap->writing); 2427 dout(" removing cap_snap %p follows %lld\n", 2428 capsnap, follows); 2429 ceph_put_snap_context(capsnap->context); 2430 list_del(&capsnap->ci_item); 2431 list_del(&capsnap->flushing_item); 2432 ceph_put_cap_snap(capsnap); 2433 drop = 1; 2434 break; 2435 } else { 2436 dout(" skipping cap_snap %p follows %lld\n", 2437 capsnap, capsnap->follows); 2438 } 2439 } 2440 spin_unlock(&inode->i_lock); 2441 if (drop) 2442 iput(inode); 2443 } 2444 2445 /* 2446 * Handle TRUNC from MDS, indicating file truncation. 2447 * 2448 * caller hold s_mutex. 2449 */ 2450 static void handle_cap_trunc(struct inode *inode, 2451 struct ceph_mds_caps *trunc, 2452 struct ceph_mds_session *session) 2453 __releases(inode->i_lock) 2454 { 2455 struct ceph_inode_info *ci = ceph_inode(inode); 2456 int mds = session->s_mds; 2457 int seq = le32_to_cpu(trunc->seq); 2458 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq); 2459 u64 truncate_size = le64_to_cpu(trunc->truncate_size); 2460 u64 size = le64_to_cpu(trunc->size); 2461 int implemented = 0; 2462 int dirty = __ceph_caps_dirty(ci); 2463 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented); 2464 int queue_trunc = 0; 2465 2466 issued |= implemented | dirty; 2467 2468 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n", 2469 inode, mds, seq, truncate_size, truncate_seq); 2470 queue_trunc = ceph_fill_file_size(inode, issued, 2471 truncate_seq, truncate_size, size); 2472 spin_unlock(&inode->i_lock); 2473 2474 if (queue_trunc) 2475 if (queue_work(ceph_client(inode->i_sb)->trunc_wq, 2476 &ci->i_vmtruncate_work)) 2477 igrab(inode); 2478 } 2479 2480 /* 2481 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a 2482 * different one. If we are the most recent migration we've seen (as 2483 * indicated by mseq), make note of the migrating cap bits for the 2484 * duration (until we see the corresponding IMPORT). 2485 * 2486 * caller holds s_mutex 2487 */ 2488 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex, 2489 struct ceph_mds_session *session) 2490 { 2491 struct ceph_inode_info *ci = ceph_inode(inode); 2492 int mds = session->s_mds; 2493 unsigned mseq = le32_to_cpu(ex->migrate_seq); 2494 struct ceph_cap *cap = NULL, *t; 2495 struct rb_node *p; 2496 int remember = 1; 2497 2498 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n", 2499 inode, ci, mds, mseq); 2500 2501 spin_lock(&inode->i_lock); 2502 2503 /* make sure we haven't seen a higher mseq */ 2504 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 2505 t = rb_entry(p, struct ceph_cap, ci_node); 2506 if (ceph_seq_cmp(t->mseq, mseq) > 0) { 2507 dout(" higher mseq on cap from mds%d\n", 2508 t->session->s_mds); 2509 remember = 0; 2510 } 2511 if (t->session->s_mds == mds) 2512 cap = t; 2513 } 2514 2515 if (cap) { 2516 if (remember) { 2517 /* make note */ 2518 ci->i_cap_exporting_mds = mds; 2519 ci->i_cap_exporting_mseq = mseq; 2520 ci->i_cap_exporting_issued = cap->issued; 2521 } 2522 __ceph_remove_cap(cap, NULL); 2523 } else { 2524 WARN_ON(!cap); 2525 } 2526 2527 spin_unlock(&inode->i_lock); 2528 } 2529 2530 /* 2531 * Handle cap IMPORT. If there are temp bits from an older EXPORT, 2532 * clean them up. 2533 * 2534 * caller holds s_mutex. 2535 */ 2536 static void handle_cap_import(struct ceph_mds_client *mdsc, 2537 struct inode *inode, struct ceph_mds_caps *im, 2538 struct ceph_mds_session *session, 2539 void *snaptrace, int snaptrace_len) 2540 { 2541 struct ceph_inode_info *ci = ceph_inode(inode); 2542 int mds = session->s_mds; 2543 unsigned issued = le32_to_cpu(im->caps); 2544 unsigned wanted = le32_to_cpu(im->wanted); 2545 unsigned seq = le32_to_cpu(im->seq); 2546 unsigned mseq = le32_to_cpu(im->migrate_seq); 2547 u64 realmino = le64_to_cpu(im->realm); 2548 u64 cap_id = le64_to_cpu(im->cap_id); 2549 2550 if (ci->i_cap_exporting_mds >= 0 && 2551 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) { 2552 dout("handle_cap_import inode %p ci %p mds%d mseq %d" 2553 " - cleared exporting from mds%d\n", 2554 inode, ci, mds, mseq, 2555 ci->i_cap_exporting_mds); 2556 ci->i_cap_exporting_issued = 0; 2557 ci->i_cap_exporting_mseq = 0; 2558 ci->i_cap_exporting_mds = -1; 2559 } else { 2560 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n", 2561 inode, ci, mds, mseq); 2562 } 2563 2564 down_write(&mdsc->snap_rwsem); 2565 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len, 2566 false); 2567 downgrade_write(&mdsc->snap_rwsem); 2568 ceph_add_cap(inode, session, cap_id, -1, 2569 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH, 2570 NULL /* no caps context */); 2571 try_flush_caps(inode, session, NULL); 2572 up_read(&mdsc->snap_rwsem); 2573 } 2574 2575 /* 2576 * Handle a caps message from the MDS. 2577 * 2578 * Identify the appropriate session, inode, and call the right handler 2579 * based on the cap op. 2580 */ 2581 void ceph_handle_caps(struct ceph_mds_session *session, 2582 struct ceph_msg *msg) 2583 { 2584 struct ceph_mds_client *mdsc = session->s_mdsc; 2585 struct super_block *sb = mdsc->client->sb; 2586 struct inode *inode; 2587 struct ceph_cap *cap; 2588 struct ceph_mds_caps *h; 2589 int mds = le64_to_cpu(msg->hdr.src.name.num); 2590 int op; 2591 u32 seq; 2592 struct ceph_vino vino; 2593 u64 cap_id; 2594 u64 size, max_size; 2595 int check_caps = 0; 2596 int r; 2597 2598 dout("handle_caps from mds%d\n", mds); 2599 2600 /* decode */ 2601 if (msg->front.iov_len < sizeof(*h)) 2602 goto bad; 2603 h = msg->front.iov_base; 2604 op = le32_to_cpu(h->op); 2605 vino.ino = le64_to_cpu(h->ino); 2606 vino.snap = CEPH_NOSNAP; 2607 cap_id = le64_to_cpu(h->cap_id); 2608 seq = le32_to_cpu(h->seq); 2609 size = le64_to_cpu(h->size); 2610 max_size = le64_to_cpu(h->max_size); 2611 2612 mutex_lock(&session->s_mutex); 2613 session->s_seq++; 2614 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, 2615 (unsigned)seq); 2616 2617 /* lookup ino */ 2618 inode = ceph_find_inode(sb, vino); 2619 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino, 2620 vino.snap, inode); 2621 if (!inode) { 2622 dout(" i don't have ino %llx\n", vino.ino); 2623 goto done; 2624 } 2625 2626 /* these will work even if we don't have a cap yet */ 2627 switch (op) { 2628 case CEPH_CAP_OP_FLUSHSNAP_ACK: 2629 handle_cap_flushsnap_ack(inode, h, session); 2630 goto done; 2631 2632 case CEPH_CAP_OP_EXPORT: 2633 handle_cap_export(inode, h, session); 2634 goto done; 2635 2636 case CEPH_CAP_OP_IMPORT: 2637 handle_cap_import(mdsc, inode, h, session, 2638 msg->middle, 2639 le32_to_cpu(h->snap_trace_len)); 2640 check_caps = 1; /* we may have sent a RELEASE to the old auth */ 2641 goto done; 2642 } 2643 2644 /* the rest require a cap */ 2645 spin_lock(&inode->i_lock); 2646 cap = __get_cap_for_mds(ceph_inode(inode), mds); 2647 if (!cap) { 2648 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n", 2649 inode, ceph_ino(inode), ceph_snap(inode), mds); 2650 spin_unlock(&inode->i_lock); 2651 goto done; 2652 } 2653 2654 /* note that each of these drops i_lock for us */ 2655 switch (op) { 2656 case CEPH_CAP_OP_REVOKE: 2657 case CEPH_CAP_OP_GRANT: 2658 r = handle_cap_grant(inode, h, session, cap, msg->middle); 2659 if (r == 1) 2660 ceph_check_caps(ceph_inode(inode), 2661 CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY, 2662 session); 2663 else if (r == 2) 2664 ceph_check_caps(ceph_inode(inode), 2665 CHECK_CAPS_NODELAY, 2666 session); 2667 break; 2668 2669 case CEPH_CAP_OP_FLUSH_ACK: 2670 handle_cap_flush_ack(inode, h, session, cap); 2671 break; 2672 2673 case CEPH_CAP_OP_TRUNC: 2674 handle_cap_trunc(inode, h, session); 2675 break; 2676 2677 default: 2678 spin_unlock(&inode->i_lock); 2679 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op, 2680 ceph_cap_op_name(op)); 2681 } 2682 2683 done: 2684 mutex_unlock(&session->s_mutex); 2685 2686 if (check_caps) 2687 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY, NULL); 2688 if (inode) 2689 iput(inode); 2690 return; 2691 2692 bad: 2693 pr_err("ceph_handle_caps: corrupt message\n"); 2694 return; 2695 } 2696 2697 /* 2698 * Delayed work handler to process end of delayed cap release LRU list. 2699 */ 2700 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc) 2701 { 2702 struct ceph_inode_info *ci; 2703 int flags = CHECK_CAPS_NODELAY; 2704 2705 dout("check_delayed_caps\n"); 2706 while (1) { 2707 spin_lock(&mdsc->cap_delay_lock); 2708 if (list_empty(&mdsc->cap_delay_list)) 2709 break; 2710 ci = list_first_entry(&mdsc->cap_delay_list, 2711 struct ceph_inode_info, 2712 i_cap_delay_list); 2713 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 && 2714 time_before(jiffies, ci->i_hold_caps_max)) 2715 break; 2716 list_del_init(&ci->i_cap_delay_list); 2717 spin_unlock(&mdsc->cap_delay_lock); 2718 dout("check_delayed_caps on %p\n", &ci->vfs_inode); 2719 ceph_check_caps(ci, flags, NULL); 2720 } 2721 spin_unlock(&mdsc->cap_delay_lock); 2722 } 2723 2724 /* 2725 * Flush all dirty caps to the mds 2726 */ 2727 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc) 2728 { 2729 struct ceph_inode_info *ci; 2730 struct inode *inode; 2731 2732 dout("flush_dirty_caps\n"); 2733 spin_lock(&mdsc->cap_dirty_lock); 2734 while (!list_empty(&mdsc->cap_dirty)) { 2735 ci = list_first_entry(&mdsc->cap_dirty, 2736 struct ceph_inode_info, 2737 i_dirty_item); 2738 inode = igrab(&ci->vfs_inode); 2739 spin_unlock(&mdsc->cap_dirty_lock); 2740 if (inode) { 2741 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, 2742 NULL); 2743 iput(inode); 2744 } 2745 spin_lock(&mdsc->cap_dirty_lock); 2746 } 2747 spin_unlock(&mdsc->cap_dirty_lock); 2748 } 2749 2750 /* 2751 * Drop open file reference. If we were the last open file, 2752 * we may need to release capabilities to the MDS (or schedule 2753 * their delayed release). 2754 */ 2755 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode) 2756 { 2757 struct inode *inode = &ci->vfs_inode; 2758 int last = 0; 2759 2760 spin_lock(&inode->i_lock); 2761 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode, 2762 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1); 2763 BUG_ON(ci->i_nr_by_mode[fmode] == 0); 2764 if (--ci->i_nr_by_mode[fmode] == 0) 2765 last++; 2766 spin_unlock(&inode->i_lock); 2767 2768 if (last && ci->i_vino.snap == CEPH_NOSNAP) 2769 ceph_check_caps(ci, 0, NULL); 2770 } 2771 2772 /* 2773 * Helpers for embedding cap and dentry lease releases into mds 2774 * requests. 2775 * 2776 * @force is used by dentry_release (below) to force inclusion of a 2777 * record for the directory inode, even when there aren't any caps to 2778 * drop. 2779 */ 2780 int ceph_encode_inode_release(void **p, struct inode *inode, 2781 int mds, int drop, int unless, int force) 2782 { 2783 struct ceph_inode_info *ci = ceph_inode(inode); 2784 struct ceph_cap *cap; 2785 struct ceph_mds_request_release *rel = *p; 2786 int ret = 0; 2787 2788 dout("encode_inode_release %p mds%d drop %s unless %s\n", inode, 2789 mds, ceph_cap_string(drop), ceph_cap_string(unless)); 2790 2791 spin_lock(&inode->i_lock); 2792 cap = __get_cap_for_mds(ci, mds); 2793 if (cap && __cap_is_valid(cap)) { 2794 if (force || 2795 ((cap->issued & drop) && 2796 (cap->issued & unless) == 0)) { 2797 if ((cap->issued & drop) && 2798 (cap->issued & unless) == 0) { 2799 dout("encode_inode_release %p cap %p %s -> " 2800 "%s\n", inode, cap, 2801 ceph_cap_string(cap->issued), 2802 ceph_cap_string(cap->issued & ~drop)); 2803 cap->issued &= ~drop; 2804 cap->implemented &= ~drop; 2805 if (ci->i_ceph_flags & CEPH_I_NODELAY) { 2806 int wanted = __ceph_caps_wanted(ci); 2807 dout(" wanted %s -> %s (act %s)\n", 2808 ceph_cap_string(cap->mds_wanted), 2809 ceph_cap_string(cap->mds_wanted & 2810 ~wanted), 2811 ceph_cap_string(wanted)); 2812 cap->mds_wanted &= wanted; 2813 } 2814 } else { 2815 dout("encode_inode_release %p cap %p %s" 2816 " (force)\n", inode, cap, 2817 ceph_cap_string(cap->issued)); 2818 } 2819 2820 rel->ino = cpu_to_le64(ceph_ino(inode)); 2821 rel->cap_id = cpu_to_le64(cap->cap_id); 2822 rel->seq = cpu_to_le32(cap->seq); 2823 rel->issue_seq = cpu_to_le32(cap->issue_seq), 2824 rel->mseq = cpu_to_le32(cap->mseq); 2825 rel->caps = cpu_to_le32(cap->issued); 2826 rel->wanted = cpu_to_le32(cap->mds_wanted); 2827 rel->dname_len = 0; 2828 rel->dname_seq = 0; 2829 *p += sizeof(*rel); 2830 ret = 1; 2831 } else { 2832 dout("encode_inode_release %p cap %p %s\n", 2833 inode, cap, ceph_cap_string(cap->issued)); 2834 } 2835 } 2836 spin_unlock(&inode->i_lock); 2837 return ret; 2838 } 2839 2840 int ceph_encode_dentry_release(void **p, struct dentry *dentry, 2841 int mds, int drop, int unless) 2842 { 2843 struct inode *dir = dentry->d_parent->d_inode; 2844 struct ceph_mds_request_release *rel = *p; 2845 struct ceph_dentry_info *di = ceph_dentry(dentry); 2846 int force = 0; 2847 int ret; 2848 2849 /* 2850 * force an record for the directory caps if we have a dentry lease. 2851 * this is racy (can't take i_lock and d_lock together), but it 2852 * doesn't have to be perfect; the mds will revoke anything we don't 2853 * release. 2854 */ 2855 spin_lock(&dentry->d_lock); 2856 if (di->lease_session && di->lease_session->s_mds == mds) 2857 force = 1; 2858 spin_unlock(&dentry->d_lock); 2859 2860 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force); 2861 2862 spin_lock(&dentry->d_lock); 2863 if (ret && di->lease_session && di->lease_session->s_mds == mds) { 2864 dout("encode_dentry_release %p mds%d seq %d\n", 2865 dentry, mds, (int)di->lease_seq); 2866 rel->dname_len = cpu_to_le32(dentry->d_name.len); 2867 memcpy(*p, dentry->d_name.name, dentry->d_name.len); 2868 *p += dentry->d_name.len; 2869 rel->dname_seq = cpu_to_le32(di->lease_seq); 2870 } 2871 spin_unlock(&dentry->d_lock); 2872 return ret; 2873 } 2874