1 #include <linux/ceph/ceph_debug.h> 2 3 #include <linux/fs.h> 4 #include <linux/kernel.h> 5 #include <linux/sched.h> 6 #include <linux/slab.h> 7 #include <linux/vmalloc.h> 8 #include <linux/wait.h> 9 #include <linux/writeback.h> 10 11 #include "super.h" 12 #include "mds_client.h" 13 #include "cache.h" 14 #include <linux/ceph/decode.h> 15 #include <linux/ceph/messenger.h> 16 17 /* 18 * Capability management 19 * 20 * The Ceph metadata servers control client access to inode metadata 21 * and file data by issuing capabilities, granting clients permission 22 * to read and/or write both inode field and file data to OSDs 23 * (storage nodes). Each capability consists of a set of bits 24 * indicating which operations are allowed. 25 * 26 * If the client holds a *_SHARED cap, the client has a coherent value 27 * that can be safely read from the cached inode. 28 * 29 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the 30 * client is allowed to change inode attributes (e.g., file size, 31 * mtime), note its dirty state in the ceph_cap, and asynchronously 32 * flush that metadata change to the MDS. 33 * 34 * In the event of a conflicting operation (perhaps by another 35 * client), the MDS will revoke the conflicting client capabilities. 36 * 37 * In order for a client to cache an inode, it must hold a capability 38 * with at least one MDS server. When inodes are released, release 39 * notifications are batched and periodically sent en masse to the MDS 40 * cluster to release server state. 41 */ 42 43 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc); 44 static void __kick_flushing_caps(struct ceph_mds_client *mdsc, 45 struct ceph_mds_session *session, 46 struct ceph_inode_info *ci, 47 u64 oldest_flush_tid); 48 49 /* 50 * Generate readable cap strings for debugging output. 51 */ 52 #define MAX_CAP_STR 20 53 static char cap_str[MAX_CAP_STR][40]; 54 static DEFINE_SPINLOCK(cap_str_lock); 55 static int last_cap_str; 56 57 static char *gcap_string(char *s, int c) 58 { 59 if (c & CEPH_CAP_GSHARED) 60 *s++ = 's'; 61 if (c & CEPH_CAP_GEXCL) 62 *s++ = 'x'; 63 if (c & CEPH_CAP_GCACHE) 64 *s++ = 'c'; 65 if (c & CEPH_CAP_GRD) 66 *s++ = 'r'; 67 if (c & CEPH_CAP_GWR) 68 *s++ = 'w'; 69 if (c & CEPH_CAP_GBUFFER) 70 *s++ = 'b'; 71 if (c & CEPH_CAP_GLAZYIO) 72 *s++ = 'l'; 73 return s; 74 } 75 76 const char *ceph_cap_string(int caps) 77 { 78 int i; 79 char *s; 80 int c; 81 82 spin_lock(&cap_str_lock); 83 i = last_cap_str++; 84 if (last_cap_str == MAX_CAP_STR) 85 last_cap_str = 0; 86 spin_unlock(&cap_str_lock); 87 88 s = cap_str[i]; 89 90 if (caps & CEPH_CAP_PIN) 91 *s++ = 'p'; 92 93 c = (caps >> CEPH_CAP_SAUTH) & 3; 94 if (c) { 95 *s++ = 'A'; 96 s = gcap_string(s, c); 97 } 98 99 c = (caps >> CEPH_CAP_SLINK) & 3; 100 if (c) { 101 *s++ = 'L'; 102 s = gcap_string(s, c); 103 } 104 105 c = (caps >> CEPH_CAP_SXATTR) & 3; 106 if (c) { 107 *s++ = 'X'; 108 s = gcap_string(s, c); 109 } 110 111 c = caps >> CEPH_CAP_SFILE; 112 if (c) { 113 *s++ = 'F'; 114 s = gcap_string(s, c); 115 } 116 117 if (s == cap_str[i]) 118 *s++ = '-'; 119 *s = 0; 120 return cap_str[i]; 121 } 122 123 void ceph_caps_init(struct ceph_mds_client *mdsc) 124 { 125 INIT_LIST_HEAD(&mdsc->caps_list); 126 spin_lock_init(&mdsc->caps_list_lock); 127 } 128 129 void ceph_caps_finalize(struct ceph_mds_client *mdsc) 130 { 131 struct ceph_cap *cap; 132 133 spin_lock(&mdsc->caps_list_lock); 134 while (!list_empty(&mdsc->caps_list)) { 135 cap = list_first_entry(&mdsc->caps_list, 136 struct ceph_cap, caps_item); 137 list_del(&cap->caps_item); 138 kmem_cache_free(ceph_cap_cachep, cap); 139 } 140 mdsc->caps_total_count = 0; 141 mdsc->caps_avail_count = 0; 142 mdsc->caps_use_count = 0; 143 mdsc->caps_reserve_count = 0; 144 mdsc->caps_min_count = 0; 145 spin_unlock(&mdsc->caps_list_lock); 146 } 147 148 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta) 149 { 150 spin_lock(&mdsc->caps_list_lock); 151 mdsc->caps_min_count += delta; 152 BUG_ON(mdsc->caps_min_count < 0); 153 spin_unlock(&mdsc->caps_list_lock); 154 } 155 156 void ceph_reserve_caps(struct ceph_mds_client *mdsc, 157 struct ceph_cap_reservation *ctx, int need) 158 { 159 int i; 160 struct ceph_cap *cap; 161 int have; 162 int alloc = 0; 163 LIST_HEAD(newcaps); 164 165 dout("reserve caps ctx=%p need=%d\n", ctx, need); 166 167 /* first reserve any caps that are already allocated */ 168 spin_lock(&mdsc->caps_list_lock); 169 if (mdsc->caps_avail_count >= need) 170 have = need; 171 else 172 have = mdsc->caps_avail_count; 173 mdsc->caps_avail_count -= have; 174 mdsc->caps_reserve_count += have; 175 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 176 mdsc->caps_reserve_count + 177 mdsc->caps_avail_count); 178 spin_unlock(&mdsc->caps_list_lock); 179 180 for (i = have; i < need; i++) { 181 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 182 if (!cap) 183 break; 184 list_add(&cap->caps_item, &newcaps); 185 alloc++; 186 } 187 /* we didn't manage to reserve as much as we needed */ 188 if (have + alloc != need) 189 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n", 190 ctx, need, have + alloc); 191 192 spin_lock(&mdsc->caps_list_lock); 193 mdsc->caps_total_count += alloc; 194 mdsc->caps_reserve_count += alloc; 195 list_splice(&newcaps, &mdsc->caps_list); 196 197 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 198 mdsc->caps_reserve_count + 199 mdsc->caps_avail_count); 200 spin_unlock(&mdsc->caps_list_lock); 201 202 ctx->count = need; 203 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n", 204 ctx, mdsc->caps_total_count, mdsc->caps_use_count, 205 mdsc->caps_reserve_count, mdsc->caps_avail_count); 206 } 207 208 int ceph_unreserve_caps(struct ceph_mds_client *mdsc, 209 struct ceph_cap_reservation *ctx) 210 { 211 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count); 212 if (ctx->count) { 213 spin_lock(&mdsc->caps_list_lock); 214 BUG_ON(mdsc->caps_reserve_count < ctx->count); 215 mdsc->caps_reserve_count -= ctx->count; 216 mdsc->caps_avail_count += ctx->count; 217 ctx->count = 0; 218 dout("unreserve caps %d = %d used + %d resv + %d avail\n", 219 mdsc->caps_total_count, mdsc->caps_use_count, 220 mdsc->caps_reserve_count, mdsc->caps_avail_count); 221 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 222 mdsc->caps_reserve_count + 223 mdsc->caps_avail_count); 224 spin_unlock(&mdsc->caps_list_lock); 225 } 226 return 0; 227 } 228 229 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc, 230 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 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); 237 if (cap) { 238 spin_lock(&mdsc->caps_list_lock); 239 mdsc->caps_use_count++; 240 mdsc->caps_total_count++; 241 spin_unlock(&mdsc->caps_list_lock); 242 } 243 return cap; 244 } 245 246 spin_lock(&mdsc->caps_list_lock); 247 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n", 248 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count, 249 mdsc->caps_reserve_count, mdsc->caps_avail_count); 250 BUG_ON(!ctx->count); 251 BUG_ON(ctx->count > mdsc->caps_reserve_count); 252 BUG_ON(list_empty(&mdsc->caps_list)); 253 254 ctx->count--; 255 mdsc->caps_reserve_count--; 256 mdsc->caps_use_count++; 257 258 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item); 259 list_del(&cap->caps_item); 260 261 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 262 mdsc->caps_reserve_count + mdsc->caps_avail_count); 263 spin_unlock(&mdsc->caps_list_lock); 264 return cap; 265 } 266 267 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap) 268 { 269 spin_lock(&mdsc->caps_list_lock); 270 dout("put_cap %p %d = %d used + %d resv + %d avail\n", 271 cap, mdsc->caps_total_count, mdsc->caps_use_count, 272 mdsc->caps_reserve_count, mdsc->caps_avail_count); 273 mdsc->caps_use_count--; 274 /* 275 * Keep some preallocated caps around (ceph_min_count), to 276 * avoid lots of free/alloc churn. 277 */ 278 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count + 279 mdsc->caps_min_count) { 280 mdsc->caps_total_count--; 281 kmem_cache_free(ceph_cap_cachep, cap); 282 } else { 283 mdsc->caps_avail_count++; 284 list_add(&cap->caps_item, &mdsc->caps_list); 285 } 286 287 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + 288 mdsc->caps_reserve_count + mdsc->caps_avail_count); 289 spin_unlock(&mdsc->caps_list_lock); 290 } 291 292 void ceph_reservation_status(struct ceph_fs_client *fsc, 293 int *total, int *avail, int *used, int *reserved, 294 int *min) 295 { 296 struct ceph_mds_client *mdsc = fsc->mdsc; 297 298 if (total) 299 *total = mdsc->caps_total_count; 300 if (avail) 301 *avail = mdsc->caps_avail_count; 302 if (used) 303 *used = mdsc->caps_use_count; 304 if (reserved) 305 *reserved = mdsc->caps_reserve_count; 306 if (min) 307 *min = mdsc->caps_min_count; 308 } 309 310 /* 311 * Find ceph_cap for given mds, if any. 312 * 313 * Called with i_ceph_lock held. 314 */ 315 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds) 316 { 317 struct ceph_cap *cap; 318 struct rb_node *n = ci->i_caps.rb_node; 319 320 while (n) { 321 cap = rb_entry(n, struct ceph_cap, ci_node); 322 if (mds < cap->mds) 323 n = n->rb_left; 324 else if (mds > cap->mds) 325 n = n->rb_right; 326 else 327 return cap; 328 } 329 return NULL; 330 } 331 332 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds) 333 { 334 struct ceph_cap *cap; 335 336 spin_lock(&ci->i_ceph_lock); 337 cap = __get_cap_for_mds(ci, mds); 338 spin_unlock(&ci->i_ceph_lock); 339 return cap; 340 } 341 342 /* 343 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1. 344 */ 345 static int __ceph_get_cap_mds(struct ceph_inode_info *ci) 346 { 347 struct ceph_cap *cap; 348 int mds = -1; 349 struct rb_node *p; 350 351 /* prefer mds with WR|BUFFER|EXCL caps */ 352 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 353 cap = rb_entry(p, struct ceph_cap, ci_node); 354 mds = cap->mds; 355 if (cap->issued & (CEPH_CAP_FILE_WR | 356 CEPH_CAP_FILE_BUFFER | 357 CEPH_CAP_FILE_EXCL)) 358 break; 359 } 360 return mds; 361 } 362 363 int ceph_get_cap_mds(struct inode *inode) 364 { 365 struct ceph_inode_info *ci = ceph_inode(inode); 366 int mds; 367 spin_lock(&ci->i_ceph_lock); 368 mds = __ceph_get_cap_mds(ceph_inode(inode)); 369 spin_unlock(&ci->i_ceph_lock); 370 return mds; 371 } 372 373 /* 374 * Called under i_ceph_lock. 375 */ 376 static void __insert_cap_node(struct ceph_inode_info *ci, 377 struct ceph_cap *new) 378 { 379 struct rb_node **p = &ci->i_caps.rb_node; 380 struct rb_node *parent = NULL; 381 struct ceph_cap *cap = NULL; 382 383 while (*p) { 384 parent = *p; 385 cap = rb_entry(parent, struct ceph_cap, ci_node); 386 if (new->mds < cap->mds) 387 p = &(*p)->rb_left; 388 else if (new->mds > cap->mds) 389 p = &(*p)->rb_right; 390 else 391 BUG(); 392 } 393 394 rb_link_node(&new->ci_node, parent, p); 395 rb_insert_color(&new->ci_node, &ci->i_caps); 396 } 397 398 /* 399 * (re)set cap hold timeouts, which control the delayed release 400 * of unused caps back to the MDS. Should be called on cap use. 401 */ 402 static void __cap_set_timeouts(struct ceph_mds_client *mdsc, 403 struct ceph_inode_info *ci) 404 { 405 struct ceph_mount_options *ma = mdsc->fsc->mount_options; 406 407 ci->i_hold_caps_min = round_jiffies(jiffies + 408 ma->caps_wanted_delay_min * HZ); 409 ci->i_hold_caps_max = round_jiffies(jiffies + 410 ma->caps_wanted_delay_max * HZ); 411 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode, 412 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies); 413 } 414 415 /* 416 * (Re)queue cap at the end of the delayed cap release list. 417 * 418 * If I_FLUSH is set, leave the inode at the front of the list. 419 * 420 * Caller holds i_ceph_lock 421 * -> we take mdsc->cap_delay_lock 422 */ 423 static void __cap_delay_requeue(struct ceph_mds_client *mdsc, 424 struct ceph_inode_info *ci) 425 { 426 __cap_set_timeouts(mdsc, ci); 427 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode, 428 ci->i_ceph_flags, ci->i_hold_caps_max); 429 if (!mdsc->stopping) { 430 spin_lock(&mdsc->cap_delay_lock); 431 if (!list_empty(&ci->i_cap_delay_list)) { 432 if (ci->i_ceph_flags & CEPH_I_FLUSH) 433 goto no_change; 434 list_del_init(&ci->i_cap_delay_list); 435 } 436 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 437 no_change: 438 spin_unlock(&mdsc->cap_delay_lock); 439 } 440 } 441 442 /* 443 * Queue an inode for immediate writeback. Mark inode with I_FLUSH, 444 * indicating we should send a cap message to flush dirty metadata 445 * asap, and move to the front of the delayed cap list. 446 */ 447 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc, 448 struct ceph_inode_info *ci) 449 { 450 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode); 451 spin_lock(&mdsc->cap_delay_lock); 452 ci->i_ceph_flags |= CEPH_I_FLUSH; 453 if (!list_empty(&ci->i_cap_delay_list)) 454 list_del_init(&ci->i_cap_delay_list); 455 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list); 456 spin_unlock(&mdsc->cap_delay_lock); 457 } 458 459 /* 460 * Cancel delayed work on cap. 461 * 462 * Caller must hold i_ceph_lock. 463 */ 464 static void __cap_delay_cancel(struct ceph_mds_client *mdsc, 465 struct ceph_inode_info *ci) 466 { 467 dout("__cap_delay_cancel %p\n", &ci->vfs_inode); 468 if (list_empty(&ci->i_cap_delay_list)) 469 return; 470 spin_lock(&mdsc->cap_delay_lock); 471 list_del_init(&ci->i_cap_delay_list); 472 spin_unlock(&mdsc->cap_delay_lock); 473 } 474 475 /* 476 * Common issue checks for add_cap, handle_cap_grant. 477 */ 478 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap, 479 unsigned issued) 480 { 481 unsigned had = __ceph_caps_issued(ci, NULL); 482 483 /* 484 * Each time we receive FILE_CACHE anew, we increment 485 * i_rdcache_gen. 486 */ 487 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) && 488 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) { 489 ci->i_rdcache_gen++; 490 } 491 492 /* 493 * if we are newly issued FILE_SHARED, mark dir not complete; we 494 * don't know what happened to this directory while we didn't 495 * have the cap. 496 */ 497 if ((issued & CEPH_CAP_FILE_SHARED) && 498 (had & CEPH_CAP_FILE_SHARED) == 0) { 499 ci->i_shared_gen++; 500 if (S_ISDIR(ci->vfs_inode.i_mode)) { 501 dout(" marking %p NOT complete\n", &ci->vfs_inode); 502 __ceph_dir_clear_complete(ci); 503 } 504 } 505 } 506 507 /* 508 * Add a capability under the given MDS session. 509 * 510 * Caller should hold session snap_rwsem (read) and s_mutex. 511 * 512 * @fmode is the open file mode, if we are opening a file, otherwise 513 * it is < 0. (This is so we can atomically add the cap and add an 514 * open file reference to it.) 515 */ 516 void ceph_add_cap(struct inode *inode, 517 struct ceph_mds_session *session, u64 cap_id, 518 int fmode, unsigned issued, unsigned wanted, 519 unsigned seq, unsigned mseq, u64 realmino, int flags, 520 struct ceph_cap **new_cap) 521 { 522 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 523 struct ceph_inode_info *ci = ceph_inode(inode); 524 struct ceph_cap *cap; 525 int mds = session->s_mds; 526 int actual_wanted; 527 528 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode, 529 session->s_mds, cap_id, ceph_cap_string(issued), seq); 530 531 /* 532 * If we are opening the file, include file mode wanted bits 533 * in wanted. 534 */ 535 if (fmode >= 0) 536 wanted |= ceph_caps_for_mode(fmode); 537 538 cap = __get_cap_for_mds(ci, mds); 539 if (!cap) { 540 cap = *new_cap; 541 *new_cap = NULL; 542 543 cap->issued = 0; 544 cap->implemented = 0; 545 cap->mds = mds; 546 cap->mds_wanted = 0; 547 cap->mseq = 0; 548 549 cap->ci = ci; 550 __insert_cap_node(ci, cap); 551 552 /* add to session cap list */ 553 cap->session = session; 554 spin_lock(&session->s_cap_lock); 555 list_add_tail(&cap->session_caps, &session->s_caps); 556 session->s_nr_caps++; 557 spin_unlock(&session->s_cap_lock); 558 } else { 559 /* 560 * auth mds of the inode changed. we received the cap export 561 * message, but still haven't received the cap import message. 562 * handle_cap_export() updated the new auth MDS' cap. 563 * 564 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing 565 * a message that was send before the cap import message. So 566 * don't remove caps. 567 */ 568 if (ceph_seq_cmp(seq, cap->seq) <= 0) { 569 WARN_ON(cap != ci->i_auth_cap); 570 WARN_ON(cap->cap_id != cap_id); 571 seq = cap->seq; 572 mseq = cap->mseq; 573 issued |= cap->issued; 574 flags |= CEPH_CAP_FLAG_AUTH; 575 } 576 } 577 578 if (!ci->i_snap_realm) { 579 /* 580 * add this inode to the appropriate snap realm 581 */ 582 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc, 583 realmino); 584 if (realm) { 585 spin_lock(&realm->inodes_with_caps_lock); 586 ci->i_snap_realm = realm; 587 list_add(&ci->i_snap_realm_item, 588 &realm->inodes_with_caps); 589 spin_unlock(&realm->inodes_with_caps_lock); 590 } else { 591 pr_err("ceph_add_cap: couldn't find snap realm %llx\n", 592 realmino); 593 WARN_ON(!realm); 594 } 595 } 596 597 __check_cap_issue(ci, cap, issued); 598 599 /* 600 * If we are issued caps we don't want, or the mds' wanted 601 * value appears to be off, queue a check so we'll release 602 * later and/or update the mds wanted value. 603 */ 604 actual_wanted = __ceph_caps_wanted(ci); 605 if ((wanted & ~actual_wanted) || 606 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) { 607 dout(" issued %s, mds wanted %s, actual %s, queueing\n", 608 ceph_cap_string(issued), ceph_cap_string(wanted), 609 ceph_cap_string(actual_wanted)); 610 __cap_delay_requeue(mdsc, ci); 611 } 612 613 if (flags & CEPH_CAP_FLAG_AUTH) { 614 if (ci->i_auth_cap == NULL || 615 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) { 616 ci->i_auth_cap = cap; 617 cap->mds_wanted = wanted; 618 } 619 } else { 620 WARN_ON(ci->i_auth_cap == cap); 621 } 622 623 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n", 624 inode, ceph_vinop(inode), cap, ceph_cap_string(issued), 625 ceph_cap_string(issued|cap->issued), seq, mds); 626 cap->cap_id = cap_id; 627 cap->issued = issued; 628 cap->implemented |= issued; 629 if (ceph_seq_cmp(mseq, cap->mseq) > 0) 630 cap->mds_wanted = wanted; 631 else 632 cap->mds_wanted |= wanted; 633 cap->seq = seq; 634 cap->issue_seq = seq; 635 cap->mseq = mseq; 636 cap->cap_gen = session->s_cap_gen; 637 638 if (fmode >= 0) 639 __ceph_get_fmode(ci, fmode); 640 } 641 642 /* 643 * Return true if cap has not timed out and belongs to the current 644 * generation of the MDS session (i.e. has not gone 'stale' due to 645 * us losing touch with the mds). 646 */ 647 static int __cap_is_valid(struct ceph_cap *cap) 648 { 649 unsigned long ttl; 650 u32 gen; 651 652 spin_lock(&cap->session->s_gen_ttl_lock); 653 gen = cap->session->s_cap_gen; 654 ttl = cap->session->s_cap_ttl; 655 spin_unlock(&cap->session->s_gen_ttl_lock); 656 657 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) { 658 dout("__cap_is_valid %p cap %p issued %s " 659 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode, 660 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen); 661 return 0; 662 } 663 664 return 1; 665 } 666 667 /* 668 * Return set of valid cap bits issued to us. Note that caps time 669 * out, and may be invalidated in bulk if the client session times out 670 * and session->s_cap_gen is bumped. 671 */ 672 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented) 673 { 674 int have = ci->i_snap_caps; 675 struct ceph_cap *cap; 676 struct rb_node *p; 677 678 if (implemented) 679 *implemented = 0; 680 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 681 cap = rb_entry(p, struct ceph_cap, ci_node); 682 if (!__cap_is_valid(cap)) 683 continue; 684 dout("__ceph_caps_issued %p cap %p issued %s\n", 685 &ci->vfs_inode, cap, ceph_cap_string(cap->issued)); 686 have |= cap->issued; 687 if (implemented) 688 *implemented |= cap->implemented; 689 } 690 /* 691 * exclude caps issued by non-auth MDS, but are been revoking 692 * by the auth MDS. The non-auth MDS should be revoking/exporting 693 * these caps, but the message is delayed. 694 */ 695 if (ci->i_auth_cap) { 696 cap = ci->i_auth_cap; 697 have &= ~cap->implemented | cap->issued; 698 } 699 return have; 700 } 701 702 /* 703 * Get cap bits issued by caps other than @ocap 704 */ 705 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap) 706 { 707 int have = ci->i_snap_caps; 708 struct ceph_cap *cap; 709 struct rb_node *p; 710 711 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 712 cap = rb_entry(p, struct ceph_cap, ci_node); 713 if (cap == ocap) 714 continue; 715 if (!__cap_is_valid(cap)) 716 continue; 717 have |= cap->issued; 718 } 719 return have; 720 } 721 722 /* 723 * Move a cap to the end of the LRU (oldest caps at list head, newest 724 * at list tail). 725 */ 726 static void __touch_cap(struct ceph_cap *cap) 727 { 728 struct ceph_mds_session *s = cap->session; 729 730 spin_lock(&s->s_cap_lock); 731 if (s->s_cap_iterator == NULL) { 732 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap, 733 s->s_mds); 734 list_move_tail(&cap->session_caps, &s->s_caps); 735 } else { 736 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n", 737 &cap->ci->vfs_inode, cap, s->s_mds); 738 } 739 spin_unlock(&s->s_cap_lock); 740 } 741 742 /* 743 * Check if we hold the given mask. If so, move the cap(s) to the 744 * front of their respective LRUs. (This is the preferred way for 745 * callers to check for caps they want.) 746 */ 747 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch) 748 { 749 struct ceph_cap *cap; 750 struct rb_node *p; 751 int have = ci->i_snap_caps; 752 753 if ((have & mask) == mask) { 754 dout("__ceph_caps_issued_mask %p snap issued %s" 755 " (mask %s)\n", &ci->vfs_inode, 756 ceph_cap_string(have), 757 ceph_cap_string(mask)); 758 return 1; 759 } 760 761 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 762 cap = rb_entry(p, struct ceph_cap, ci_node); 763 if (!__cap_is_valid(cap)) 764 continue; 765 if ((cap->issued & mask) == mask) { 766 dout("__ceph_caps_issued_mask %p cap %p issued %s" 767 " (mask %s)\n", &ci->vfs_inode, cap, 768 ceph_cap_string(cap->issued), 769 ceph_cap_string(mask)); 770 if (touch) 771 __touch_cap(cap); 772 return 1; 773 } 774 775 /* does a combination of caps satisfy mask? */ 776 have |= cap->issued; 777 if ((have & mask) == mask) { 778 dout("__ceph_caps_issued_mask %p combo issued %s" 779 " (mask %s)\n", &ci->vfs_inode, 780 ceph_cap_string(cap->issued), 781 ceph_cap_string(mask)); 782 if (touch) { 783 struct rb_node *q; 784 785 /* touch this + preceding caps */ 786 __touch_cap(cap); 787 for (q = rb_first(&ci->i_caps); q != p; 788 q = rb_next(q)) { 789 cap = rb_entry(q, struct ceph_cap, 790 ci_node); 791 if (!__cap_is_valid(cap)) 792 continue; 793 __touch_cap(cap); 794 } 795 } 796 return 1; 797 } 798 } 799 800 return 0; 801 } 802 803 /* 804 * Return true if mask caps are currently being revoked by an MDS. 805 */ 806 int __ceph_caps_revoking_other(struct ceph_inode_info *ci, 807 struct ceph_cap *ocap, int mask) 808 { 809 struct ceph_cap *cap; 810 struct rb_node *p; 811 812 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 813 cap = rb_entry(p, struct ceph_cap, ci_node); 814 if (cap != ocap && 815 (cap->implemented & ~cap->issued & mask)) 816 return 1; 817 } 818 return 0; 819 } 820 821 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask) 822 { 823 struct inode *inode = &ci->vfs_inode; 824 int ret; 825 826 spin_lock(&ci->i_ceph_lock); 827 ret = __ceph_caps_revoking_other(ci, NULL, mask); 828 spin_unlock(&ci->i_ceph_lock); 829 dout("ceph_caps_revoking %p %s = %d\n", inode, 830 ceph_cap_string(mask), ret); 831 return ret; 832 } 833 834 int __ceph_caps_used(struct ceph_inode_info *ci) 835 { 836 int used = 0; 837 if (ci->i_pin_ref) 838 used |= CEPH_CAP_PIN; 839 if (ci->i_rd_ref) 840 used |= CEPH_CAP_FILE_RD; 841 if (ci->i_rdcache_ref || 842 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */ 843 ci->vfs_inode.i_data.nrpages)) 844 used |= CEPH_CAP_FILE_CACHE; 845 if (ci->i_wr_ref) 846 used |= CEPH_CAP_FILE_WR; 847 if (ci->i_wb_ref || ci->i_wrbuffer_ref) 848 used |= CEPH_CAP_FILE_BUFFER; 849 return used; 850 } 851 852 /* 853 * wanted, by virtue of open file modes 854 */ 855 int __ceph_caps_file_wanted(struct ceph_inode_info *ci) 856 { 857 int i, bits = 0; 858 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { 859 if (ci->i_nr_by_mode[i]) 860 bits |= 1 << i; 861 } 862 if (bits == 0) 863 return 0; 864 return ceph_caps_for_mode(bits >> 1); 865 } 866 867 /* 868 * Return caps we have registered with the MDS(s) as 'wanted'. 869 */ 870 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci) 871 { 872 struct ceph_cap *cap; 873 struct rb_node *p; 874 int mds_wanted = 0; 875 876 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 877 cap = rb_entry(p, struct ceph_cap, ci_node); 878 if (!__cap_is_valid(cap)) 879 continue; 880 if (cap == ci->i_auth_cap) 881 mds_wanted |= cap->mds_wanted; 882 else 883 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR); 884 } 885 return mds_wanted; 886 } 887 888 /* 889 * called under i_ceph_lock 890 */ 891 static int __ceph_is_any_caps(struct ceph_inode_info *ci) 892 { 893 return !RB_EMPTY_ROOT(&ci->i_caps); 894 } 895 896 int ceph_is_any_caps(struct inode *inode) 897 { 898 struct ceph_inode_info *ci = ceph_inode(inode); 899 int ret; 900 901 spin_lock(&ci->i_ceph_lock); 902 ret = __ceph_is_any_caps(ci); 903 spin_unlock(&ci->i_ceph_lock); 904 905 return ret; 906 } 907 908 static void drop_inode_snap_realm(struct ceph_inode_info *ci) 909 { 910 struct ceph_snap_realm *realm = ci->i_snap_realm; 911 spin_lock(&realm->inodes_with_caps_lock); 912 list_del_init(&ci->i_snap_realm_item); 913 ci->i_snap_realm_counter++; 914 ci->i_snap_realm = NULL; 915 spin_unlock(&realm->inodes_with_caps_lock); 916 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc, 917 realm); 918 } 919 920 /* 921 * Remove a cap. Take steps to deal with a racing iterate_session_caps. 922 * 923 * caller should hold i_ceph_lock. 924 * caller will not hold session s_mutex if called from destroy_inode. 925 */ 926 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release) 927 { 928 struct ceph_mds_session *session = cap->session; 929 struct ceph_inode_info *ci = cap->ci; 930 struct ceph_mds_client *mdsc = 931 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; 932 int removed = 0; 933 934 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode); 935 936 /* remove from session list */ 937 spin_lock(&session->s_cap_lock); 938 if (session->s_cap_iterator == cap) { 939 /* not yet, we are iterating over this very cap */ 940 dout("__ceph_remove_cap delaying %p removal from session %p\n", 941 cap, cap->session); 942 } else { 943 list_del_init(&cap->session_caps); 944 session->s_nr_caps--; 945 cap->session = NULL; 946 removed = 1; 947 } 948 /* protect backpointer with s_cap_lock: see iterate_session_caps */ 949 cap->ci = NULL; 950 951 /* 952 * s_cap_reconnect is protected by s_cap_lock. no one changes 953 * s_cap_gen while session is in the reconnect state. 954 */ 955 if (queue_release && 956 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) { 957 cap->queue_release = 1; 958 if (removed) { 959 list_add_tail(&cap->session_caps, 960 &session->s_cap_releases); 961 session->s_num_cap_releases++; 962 removed = 0; 963 } 964 } else { 965 cap->queue_release = 0; 966 } 967 cap->cap_ino = ci->i_vino.ino; 968 969 spin_unlock(&session->s_cap_lock); 970 971 /* remove from inode list */ 972 rb_erase(&cap->ci_node, &ci->i_caps); 973 if (ci->i_auth_cap == cap) 974 ci->i_auth_cap = NULL; 975 976 if (removed) 977 ceph_put_cap(mdsc, cap); 978 979 /* when reconnect denied, we remove session caps forcibly, 980 * i_wr_ref can be non-zero. If there are ongoing write, 981 * keep i_snap_realm. 982 */ 983 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm) 984 drop_inode_snap_realm(ci); 985 986 if (!__ceph_is_any_real_caps(ci)) 987 __cap_delay_cancel(mdsc, ci); 988 } 989 990 /* 991 * Build and send a cap message to the given MDS. 992 * 993 * Caller should be holding s_mutex. 994 */ 995 static int send_cap_msg(struct ceph_mds_session *session, 996 u64 ino, u64 cid, int op, 997 int caps, int wanted, int dirty, 998 u32 seq, u64 flush_tid, u64 oldest_flush_tid, 999 u32 issue_seq, u32 mseq, u64 size, u64 max_size, 1000 struct timespec *mtime, struct timespec *atime, 1001 struct timespec *ctime, u32 time_warp_seq, 1002 kuid_t uid, kgid_t gid, umode_t mode, 1003 u64 xattr_version, 1004 struct ceph_buffer *xattrs_buf, 1005 u64 follows, bool inline_data) 1006 { 1007 struct ceph_mds_caps *fc; 1008 struct ceph_msg *msg; 1009 void *p; 1010 size_t extra_len; 1011 1012 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s" 1013 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu" 1014 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op), 1015 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted), 1016 ceph_cap_string(dirty), 1017 seq, issue_seq, flush_tid, oldest_flush_tid, 1018 mseq, follows, size, max_size, 1019 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0); 1020 1021 /* flock buffer size + inline version + inline data size + 1022 * osd_epoch_barrier + oldest_flush_tid */ 1023 extra_len = 4 + 8 + 4 + 4 + 8; 1024 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len, 1025 GFP_NOFS, false); 1026 if (!msg) 1027 return -ENOMEM; 1028 1029 msg->hdr.version = cpu_to_le16(6); 1030 msg->hdr.tid = cpu_to_le64(flush_tid); 1031 1032 fc = msg->front.iov_base; 1033 memset(fc, 0, sizeof(*fc)); 1034 1035 fc->cap_id = cpu_to_le64(cid); 1036 fc->op = cpu_to_le32(op); 1037 fc->seq = cpu_to_le32(seq); 1038 fc->issue_seq = cpu_to_le32(issue_seq); 1039 fc->migrate_seq = cpu_to_le32(mseq); 1040 fc->caps = cpu_to_le32(caps); 1041 fc->wanted = cpu_to_le32(wanted); 1042 fc->dirty = cpu_to_le32(dirty); 1043 fc->ino = cpu_to_le64(ino); 1044 fc->snap_follows = cpu_to_le64(follows); 1045 1046 fc->size = cpu_to_le64(size); 1047 fc->max_size = cpu_to_le64(max_size); 1048 if (mtime) 1049 ceph_encode_timespec(&fc->mtime, mtime); 1050 if (atime) 1051 ceph_encode_timespec(&fc->atime, atime); 1052 if (ctime) 1053 ceph_encode_timespec(&fc->ctime, ctime); 1054 fc->time_warp_seq = cpu_to_le32(time_warp_seq); 1055 1056 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid)); 1057 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid)); 1058 fc->mode = cpu_to_le32(mode); 1059 1060 p = fc + 1; 1061 /* flock buffer size */ 1062 ceph_encode_32(&p, 0); 1063 /* inline version */ 1064 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE); 1065 /* inline data size */ 1066 ceph_encode_32(&p, 0); 1067 /* osd_epoch_barrier */ 1068 ceph_encode_32(&p, 0); 1069 /* oldest_flush_tid */ 1070 ceph_encode_64(&p, oldest_flush_tid); 1071 1072 fc->xattr_version = cpu_to_le64(xattr_version); 1073 if (xattrs_buf) { 1074 msg->middle = ceph_buffer_get(xattrs_buf); 1075 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len); 1076 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len); 1077 } 1078 1079 ceph_con_send(&session->s_con, msg); 1080 return 0; 1081 } 1082 1083 /* 1084 * Queue cap releases when an inode is dropped from our cache. Since 1085 * inode is about to be destroyed, there is no need for i_ceph_lock. 1086 */ 1087 void ceph_queue_caps_release(struct inode *inode) 1088 { 1089 struct ceph_inode_info *ci = ceph_inode(inode); 1090 struct rb_node *p; 1091 1092 p = rb_first(&ci->i_caps); 1093 while (p) { 1094 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node); 1095 p = rb_next(p); 1096 __ceph_remove_cap(cap, true); 1097 } 1098 } 1099 1100 /* 1101 * Send a cap msg on the given inode. Update our caps state, then 1102 * drop i_ceph_lock and send the message. 1103 * 1104 * Make note of max_size reported/requested from mds, revoked caps 1105 * that have now been implemented. 1106 * 1107 * Make half-hearted attempt ot to invalidate page cache if we are 1108 * dropping RDCACHE. Note that this will leave behind locked pages 1109 * that we'll then need to deal with elsewhere. 1110 * 1111 * Return non-zero if delayed release, or we experienced an error 1112 * such that the caller should requeue + retry later. 1113 * 1114 * called with i_ceph_lock, then drops it. 1115 * caller should hold snap_rwsem (read), s_mutex. 1116 */ 1117 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap, 1118 int op, int used, int want, int retain, int flushing, 1119 u64 flush_tid, u64 oldest_flush_tid) 1120 __releases(cap->ci->i_ceph_lock) 1121 { 1122 struct ceph_inode_info *ci = cap->ci; 1123 struct inode *inode = &ci->vfs_inode; 1124 u64 cap_id = cap->cap_id; 1125 int held, revoking, dropping, keep; 1126 u64 follows, size, max_size; 1127 u32 seq, issue_seq, mseq, time_warp_seq; 1128 struct timespec mtime, atime, ctime; 1129 int wake = 0; 1130 umode_t mode; 1131 kuid_t uid; 1132 kgid_t gid; 1133 struct ceph_mds_session *session; 1134 u64 xattr_version = 0; 1135 struct ceph_buffer *xattr_blob = NULL; 1136 int delayed = 0; 1137 int ret; 1138 bool inline_data; 1139 1140 held = cap->issued | cap->implemented; 1141 revoking = cap->implemented & ~cap->issued; 1142 retain &= ~revoking; 1143 dropping = cap->issued & ~retain; 1144 1145 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n", 1146 inode, cap, cap->session, 1147 ceph_cap_string(held), ceph_cap_string(held & retain), 1148 ceph_cap_string(revoking)); 1149 BUG_ON((retain & CEPH_CAP_PIN) == 0); 1150 1151 session = cap->session; 1152 1153 /* don't release wanted unless we've waited a bit. */ 1154 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1155 time_before(jiffies, ci->i_hold_caps_min)) { 1156 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n", 1157 ceph_cap_string(cap->issued), 1158 ceph_cap_string(cap->issued & retain), 1159 ceph_cap_string(cap->mds_wanted), 1160 ceph_cap_string(want)); 1161 want |= cap->mds_wanted; 1162 retain |= cap->issued; 1163 delayed = 1; 1164 } 1165 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH); 1166 1167 cap->issued &= retain; /* drop bits we don't want */ 1168 if (cap->implemented & ~cap->issued) { 1169 /* 1170 * Wake up any waiters on wanted -> needed transition. 1171 * This is due to the weird transition from buffered 1172 * to sync IO... we need to flush dirty pages _before_ 1173 * allowing sync writes to avoid reordering. 1174 */ 1175 wake = 1; 1176 } 1177 cap->implemented &= cap->issued | used; 1178 cap->mds_wanted = want; 1179 1180 follows = flushing ? ci->i_head_snapc->seq : 0; 1181 1182 keep = cap->implemented; 1183 seq = cap->seq; 1184 issue_seq = cap->issue_seq; 1185 mseq = cap->mseq; 1186 size = inode->i_size; 1187 ci->i_reported_size = size; 1188 max_size = ci->i_wanted_max_size; 1189 ci->i_requested_max_size = max_size; 1190 mtime = inode->i_mtime; 1191 atime = inode->i_atime; 1192 ctime = inode->i_ctime; 1193 time_warp_seq = ci->i_time_warp_seq; 1194 uid = inode->i_uid; 1195 gid = inode->i_gid; 1196 mode = inode->i_mode; 1197 1198 if (flushing & CEPH_CAP_XATTR_EXCL) { 1199 __ceph_build_xattrs_blob(ci); 1200 xattr_blob = ci->i_xattrs.blob; 1201 xattr_version = ci->i_xattrs.version; 1202 } 1203 1204 inline_data = ci->i_inline_version != CEPH_INLINE_NONE; 1205 1206 spin_unlock(&ci->i_ceph_lock); 1207 1208 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id, 1209 op, keep, want, flushing, seq, 1210 flush_tid, oldest_flush_tid, issue_seq, mseq, 1211 size, max_size, &mtime, &atime, &ctime, time_warp_seq, 1212 uid, gid, mode, xattr_version, xattr_blob, 1213 follows, inline_data); 1214 if (ret < 0) { 1215 dout("error sending cap msg, must requeue %p\n", inode); 1216 delayed = 1; 1217 } 1218 1219 if (wake) 1220 wake_up_all(&ci->i_cap_wq); 1221 1222 return delayed; 1223 } 1224 1225 static inline int __send_flush_snap(struct inode *inode, 1226 struct ceph_mds_session *session, 1227 struct ceph_cap_snap *capsnap, 1228 u32 mseq, u64 oldest_flush_tid) 1229 { 1230 return send_cap_msg(session, ceph_vino(inode).ino, 0, 1231 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0, 1232 capsnap->dirty, 0, capsnap->cap_flush.tid, 1233 oldest_flush_tid, 0, mseq, capsnap->size, 0, 1234 &capsnap->mtime, &capsnap->atime, 1235 &capsnap->ctime, capsnap->time_warp_seq, 1236 capsnap->uid, capsnap->gid, capsnap->mode, 1237 capsnap->xattr_version, capsnap->xattr_blob, 1238 capsnap->follows, capsnap->inline_data); 1239 } 1240 1241 /* 1242 * When a snapshot is taken, clients accumulate dirty metadata on 1243 * inodes with capabilities in ceph_cap_snaps to describe the file 1244 * state at the time the snapshot was taken. This must be flushed 1245 * asynchronously back to the MDS once sync writes complete and dirty 1246 * data is written out. 1247 * 1248 * Called under i_ceph_lock. Takes s_mutex as needed. 1249 */ 1250 static void __ceph_flush_snaps(struct ceph_inode_info *ci, 1251 struct ceph_mds_session *session) 1252 __releases(ci->i_ceph_lock) 1253 __acquires(ci->i_ceph_lock) 1254 { 1255 struct inode *inode = &ci->vfs_inode; 1256 struct ceph_mds_client *mdsc = session->s_mdsc; 1257 struct ceph_cap_snap *capsnap; 1258 u64 oldest_flush_tid = 0; 1259 u64 first_tid = 1, last_tid = 0; 1260 1261 dout("__flush_snaps %p session %p\n", inode, session); 1262 1263 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 1264 /* 1265 * we need to wait for sync writes to complete and for dirty 1266 * pages to be written out. 1267 */ 1268 if (capsnap->dirty_pages || capsnap->writing) 1269 break; 1270 1271 /* should be removed by ceph_try_drop_cap_snap() */ 1272 BUG_ON(!capsnap->need_flush); 1273 1274 /* only flush each capsnap once */ 1275 if (capsnap->cap_flush.tid > 0) { 1276 dout(" already flushed %p, skipping\n", capsnap); 1277 continue; 1278 } 1279 1280 spin_lock(&mdsc->cap_dirty_lock); 1281 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid; 1282 list_add_tail(&capsnap->cap_flush.g_list, 1283 &mdsc->cap_flush_list); 1284 if (oldest_flush_tid == 0) 1285 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1286 if (list_empty(&ci->i_flushing_item)) { 1287 list_add_tail(&ci->i_flushing_item, 1288 &session->s_cap_flushing); 1289 } 1290 spin_unlock(&mdsc->cap_dirty_lock); 1291 1292 list_add_tail(&capsnap->cap_flush.i_list, 1293 &ci->i_cap_flush_list); 1294 1295 if (first_tid == 1) 1296 first_tid = capsnap->cap_flush.tid; 1297 last_tid = capsnap->cap_flush.tid; 1298 } 1299 1300 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS; 1301 1302 while (first_tid <= last_tid) { 1303 struct ceph_cap *cap = ci->i_auth_cap; 1304 struct ceph_cap_flush *cf; 1305 int ret; 1306 1307 if (!(cap && cap->session == session)) { 1308 dout("__flush_snaps %p auth cap %p not mds%d, " 1309 "stop\n", inode, cap, session->s_mds); 1310 break; 1311 } 1312 1313 ret = -ENOENT; 1314 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) { 1315 if (cf->tid >= first_tid) { 1316 ret = 0; 1317 break; 1318 } 1319 } 1320 if (ret < 0) 1321 break; 1322 1323 first_tid = cf->tid + 1; 1324 1325 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush); 1326 atomic_inc(&capsnap->nref); 1327 spin_unlock(&ci->i_ceph_lock); 1328 1329 dout("__flush_snaps %p capsnap %p tid %llu %s\n", 1330 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty)); 1331 1332 ret = __send_flush_snap(inode, session, capsnap, cap->mseq, 1333 oldest_flush_tid); 1334 if (ret < 0) { 1335 pr_err("__flush_snaps: error sending cap flushsnap, " 1336 "ino (%llx.%llx) tid %llu follows %llu\n", 1337 ceph_vinop(inode), cf->tid, capsnap->follows); 1338 } 1339 1340 ceph_put_cap_snap(capsnap); 1341 spin_lock(&ci->i_ceph_lock); 1342 } 1343 } 1344 1345 void ceph_flush_snaps(struct ceph_inode_info *ci, 1346 struct ceph_mds_session **psession) 1347 { 1348 struct inode *inode = &ci->vfs_inode; 1349 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 1350 struct ceph_mds_session *session = *psession; 1351 int mds; 1352 dout("ceph_flush_snaps %p\n", inode); 1353 retry: 1354 spin_lock(&ci->i_ceph_lock); 1355 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) { 1356 dout(" no capsnap needs flush, doing nothing\n"); 1357 goto out; 1358 } 1359 if (!ci->i_auth_cap) { 1360 dout(" no auth cap (migrating?), doing nothing\n"); 1361 goto out; 1362 } 1363 1364 mds = ci->i_auth_cap->session->s_mds; 1365 if (session && session->s_mds != mds) { 1366 dout(" oops, wrong session %p mutex\n", session); 1367 mutex_unlock(&session->s_mutex); 1368 ceph_put_mds_session(session); 1369 session = NULL; 1370 } 1371 if (!session) { 1372 spin_unlock(&ci->i_ceph_lock); 1373 mutex_lock(&mdsc->mutex); 1374 session = __ceph_lookup_mds_session(mdsc, mds); 1375 mutex_unlock(&mdsc->mutex); 1376 if (session) { 1377 dout(" inverting session/ino locks on %p\n", session); 1378 mutex_lock(&session->s_mutex); 1379 } 1380 goto retry; 1381 } 1382 1383 __ceph_flush_snaps(ci, session); 1384 out: 1385 spin_unlock(&ci->i_ceph_lock); 1386 1387 if (psession) { 1388 *psession = session; 1389 } else { 1390 mutex_unlock(&session->s_mutex); 1391 ceph_put_mds_session(session); 1392 } 1393 /* we flushed them all; remove this inode from the queue */ 1394 spin_lock(&mdsc->snap_flush_lock); 1395 list_del_init(&ci->i_snap_flush_item); 1396 spin_unlock(&mdsc->snap_flush_lock); 1397 } 1398 1399 /* 1400 * Mark caps dirty. If inode is newly dirty, return the dirty flags. 1401 * Caller is then responsible for calling __mark_inode_dirty with the 1402 * returned flags value. 1403 */ 1404 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask, 1405 struct ceph_cap_flush **pcf) 1406 { 1407 struct ceph_mds_client *mdsc = 1408 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; 1409 struct inode *inode = &ci->vfs_inode; 1410 int was = ci->i_dirty_caps; 1411 int dirty = 0; 1412 1413 if (!ci->i_auth_cap) { 1414 pr_warn("__mark_dirty_caps %p %llx mask %s, " 1415 "but no auth cap (session was closed?)\n", 1416 inode, ceph_ino(inode), ceph_cap_string(mask)); 1417 return 0; 1418 } 1419 1420 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode, 1421 ceph_cap_string(mask), ceph_cap_string(was), 1422 ceph_cap_string(was | mask)); 1423 ci->i_dirty_caps |= mask; 1424 if (was == 0) { 1425 WARN_ON_ONCE(ci->i_prealloc_cap_flush); 1426 swap(ci->i_prealloc_cap_flush, *pcf); 1427 1428 if (!ci->i_head_snapc) { 1429 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem)); 1430 ci->i_head_snapc = ceph_get_snap_context( 1431 ci->i_snap_realm->cached_context); 1432 } 1433 dout(" inode %p now dirty snapc %p auth cap %p\n", 1434 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap); 1435 BUG_ON(!list_empty(&ci->i_dirty_item)); 1436 spin_lock(&mdsc->cap_dirty_lock); 1437 list_add(&ci->i_dirty_item, &mdsc->cap_dirty); 1438 spin_unlock(&mdsc->cap_dirty_lock); 1439 if (ci->i_flushing_caps == 0) { 1440 ihold(inode); 1441 dirty |= I_DIRTY_SYNC; 1442 } 1443 } else { 1444 WARN_ON_ONCE(!ci->i_prealloc_cap_flush); 1445 } 1446 BUG_ON(list_empty(&ci->i_dirty_item)); 1447 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) && 1448 (mask & CEPH_CAP_FILE_BUFFER)) 1449 dirty |= I_DIRTY_DATASYNC; 1450 __cap_delay_requeue(mdsc, ci); 1451 return dirty; 1452 } 1453 1454 struct ceph_cap_flush *ceph_alloc_cap_flush(void) 1455 { 1456 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL); 1457 } 1458 1459 void ceph_free_cap_flush(struct ceph_cap_flush *cf) 1460 { 1461 if (cf) 1462 kmem_cache_free(ceph_cap_flush_cachep, cf); 1463 } 1464 1465 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc) 1466 { 1467 if (!list_empty(&mdsc->cap_flush_list)) { 1468 struct ceph_cap_flush *cf = 1469 list_first_entry(&mdsc->cap_flush_list, 1470 struct ceph_cap_flush, g_list); 1471 return cf->tid; 1472 } 1473 return 0; 1474 } 1475 1476 /* 1477 * Remove cap_flush from the mdsc's or inode's flushing cap list. 1478 * Return true if caller needs to wake up flush waiters. 1479 */ 1480 static bool __finish_cap_flush(struct ceph_mds_client *mdsc, 1481 struct ceph_inode_info *ci, 1482 struct ceph_cap_flush *cf) 1483 { 1484 struct ceph_cap_flush *prev; 1485 bool wake = cf->wake; 1486 if (mdsc) { 1487 /* are there older pending cap flushes? */ 1488 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) { 1489 prev = list_prev_entry(cf, g_list); 1490 prev->wake = true; 1491 wake = false; 1492 } 1493 list_del(&cf->g_list); 1494 } else if (ci) { 1495 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) { 1496 prev = list_prev_entry(cf, i_list); 1497 prev->wake = true; 1498 wake = false; 1499 } 1500 list_del(&cf->i_list); 1501 } else { 1502 BUG_ON(1); 1503 } 1504 return wake; 1505 } 1506 1507 /* 1508 * Add dirty inode to the flushing list. Assigned a seq number so we 1509 * can wait for caps to flush without starving. 1510 * 1511 * Called under i_ceph_lock. 1512 */ 1513 static int __mark_caps_flushing(struct inode *inode, 1514 struct ceph_mds_session *session, bool wake, 1515 u64 *flush_tid, u64 *oldest_flush_tid) 1516 { 1517 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 1518 struct ceph_inode_info *ci = ceph_inode(inode); 1519 struct ceph_cap_flush *cf = NULL; 1520 int flushing; 1521 1522 BUG_ON(ci->i_dirty_caps == 0); 1523 BUG_ON(list_empty(&ci->i_dirty_item)); 1524 BUG_ON(!ci->i_prealloc_cap_flush); 1525 1526 flushing = ci->i_dirty_caps; 1527 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", 1528 ceph_cap_string(flushing), 1529 ceph_cap_string(ci->i_flushing_caps), 1530 ceph_cap_string(ci->i_flushing_caps | flushing)); 1531 ci->i_flushing_caps |= flushing; 1532 ci->i_dirty_caps = 0; 1533 dout(" inode %p now !dirty\n", inode); 1534 1535 swap(cf, ci->i_prealloc_cap_flush); 1536 cf->caps = flushing; 1537 cf->wake = wake; 1538 1539 spin_lock(&mdsc->cap_dirty_lock); 1540 list_del_init(&ci->i_dirty_item); 1541 1542 cf->tid = ++mdsc->last_cap_flush_tid; 1543 list_add_tail(&cf->g_list, &mdsc->cap_flush_list); 1544 *oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1545 1546 if (list_empty(&ci->i_flushing_item)) { 1547 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1548 mdsc->num_cap_flushing++; 1549 } 1550 spin_unlock(&mdsc->cap_dirty_lock); 1551 1552 list_add_tail(&cf->i_list, &ci->i_cap_flush_list); 1553 1554 *flush_tid = cf->tid; 1555 return flushing; 1556 } 1557 1558 /* 1559 * try to invalidate mapping pages without blocking. 1560 */ 1561 static int try_nonblocking_invalidate(struct inode *inode) 1562 { 1563 struct ceph_inode_info *ci = ceph_inode(inode); 1564 u32 invalidating_gen = ci->i_rdcache_gen; 1565 1566 spin_unlock(&ci->i_ceph_lock); 1567 invalidate_mapping_pages(&inode->i_data, 0, -1); 1568 spin_lock(&ci->i_ceph_lock); 1569 1570 if (inode->i_data.nrpages == 0 && 1571 invalidating_gen == ci->i_rdcache_gen) { 1572 /* success. */ 1573 dout("try_nonblocking_invalidate %p success\n", inode); 1574 /* save any racing async invalidate some trouble */ 1575 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1; 1576 return 0; 1577 } 1578 dout("try_nonblocking_invalidate %p failed\n", inode); 1579 return -1; 1580 } 1581 1582 /* 1583 * Swiss army knife function to examine currently used and wanted 1584 * versus held caps. Release, flush, ack revoked caps to mds as 1585 * appropriate. 1586 * 1587 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay 1588 * cap release further. 1589 * CHECK_CAPS_AUTHONLY - we should only check the auth cap 1590 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without 1591 * further delay. 1592 */ 1593 void ceph_check_caps(struct ceph_inode_info *ci, int flags, 1594 struct ceph_mds_session *session) 1595 { 1596 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode); 1597 struct ceph_mds_client *mdsc = fsc->mdsc; 1598 struct inode *inode = &ci->vfs_inode; 1599 struct ceph_cap *cap; 1600 u64 flush_tid, oldest_flush_tid; 1601 int file_wanted, used, cap_used; 1602 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ 1603 int issued, implemented, want, retain, revoking, flushing = 0; 1604 int mds = -1; /* keep track of how far we've gone through i_caps list 1605 to avoid an infinite loop on retry */ 1606 struct rb_node *p; 1607 int delayed = 0, sent = 0, num; 1608 bool is_delayed = flags & CHECK_CAPS_NODELAY; 1609 bool queue_invalidate = false; 1610 bool force_requeue = false; 1611 bool tried_invalidate = false; 1612 1613 /* if we are unmounting, flush any unused caps immediately. */ 1614 if (mdsc->stopping) 1615 is_delayed = 1; 1616 1617 spin_lock(&ci->i_ceph_lock); 1618 1619 if (ci->i_ceph_flags & CEPH_I_FLUSH) 1620 flags |= CHECK_CAPS_FLUSH; 1621 1622 goto retry_locked; 1623 retry: 1624 spin_lock(&ci->i_ceph_lock); 1625 retry_locked: 1626 file_wanted = __ceph_caps_file_wanted(ci); 1627 used = __ceph_caps_used(ci); 1628 issued = __ceph_caps_issued(ci, &implemented); 1629 revoking = implemented & ~issued; 1630 1631 want = file_wanted; 1632 retain = file_wanted | used | CEPH_CAP_PIN; 1633 if (!mdsc->stopping && inode->i_nlink > 0) { 1634 if (file_wanted) { 1635 retain |= CEPH_CAP_ANY; /* be greedy */ 1636 } else if (S_ISDIR(inode->i_mode) && 1637 (issued & CEPH_CAP_FILE_SHARED) && 1638 __ceph_dir_is_complete(ci)) { 1639 /* 1640 * If a directory is complete, we want to keep 1641 * the exclusive cap. So that MDS does not end up 1642 * revoking the shared cap on every create/unlink 1643 * operation. 1644 */ 1645 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL; 1646 retain |= want; 1647 } else { 1648 1649 retain |= CEPH_CAP_ANY_SHARED; 1650 /* 1651 * keep RD only if we didn't have the file open RW, 1652 * because then the mds would revoke it anyway to 1653 * journal max_size=0. 1654 */ 1655 if (ci->i_max_size == 0) 1656 retain |= CEPH_CAP_ANY_RD; 1657 } 1658 } 1659 1660 dout("check_caps %p file_want %s used %s dirty %s flushing %s" 1661 " issued %s revoking %s retain %s %s%s%s\n", inode, 1662 ceph_cap_string(file_wanted), 1663 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps), 1664 ceph_cap_string(ci->i_flushing_caps), 1665 ceph_cap_string(issued), ceph_cap_string(revoking), 1666 ceph_cap_string(retain), 1667 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "", 1668 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "", 1669 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : ""); 1670 1671 /* 1672 * If we no longer need to hold onto old our caps, and we may 1673 * have cached pages, but don't want them, then try to invalidate. 1674 * If we fail, it's because pages are locked.... try again later. 1675 */ 1676 if ((!is_delayed || mdsc->stopping) && 1677 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */ 1678 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */ 1679 inode->i_data.nrpages && /* have cached pages */ 1680 (revoking & (CEPH_CAP_FILE_CACHE| 1681 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */ 1682 !tried_invalidate) { 1683 dout("check_caps trying to invalidate on %p\n", inode); 1684 if (try_nonblocking_invalidate(inode) < 0) { 1685 if (revoking & (CEPH_CAP_FILE_CACHE| 1686 CEPH_CAP_FILE_LAZYIO)) { 1687 dout("check_caps queuing invalidate\n"); 1688 queue_invalidate = true; 1689 ci->i_rdcache_revoking = ci->i_rdcache_gen; 1690 } else { 1691 dout("check_caps failed to invalidate pages\n"); 1692 /* we failed to invalidate pages. check these 1693 caps again later. */ 1694 force_requeue = true; 1695 __cap_set_timeouts(mdsc, ci); 1696 } 1697 } 1698 tried_invalidate = true; 1699 goto retry_locked; 1700 } 1701 1702 num = 0; 1703 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 1704 cap = rb_entry(p, struct ceph_cap, ci_node); 1705 num++; 1706 1707 /* avoid looping forever */ 1708 if (mds >= cap->mds || 1709 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap)) 1710 continue; 1711 1712 /* NOTE: no side-effects allowed, until we take s_mutex */ 1713 1714 cap_used = used; 1715 if (ci->i_auth_cap && cap != ci->i_auth_cap) 1716 cap_used &= ~ci->i_auth_cap->issued; 1717 1718 revoking = cap->implemented & ~cap->issued; 1719 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n", 1720 cap->mds, cap, ceph_cap_string(cap_used), 1721 ceph_cap_string(cap->issued), 1722 ceph_cap_string(cap->implemented), 1723 ceph_cap_string(revoking)); 1724 1725 if (cap == ci->i_auth_cap && 1726 (cap->issued & CEPH_CAP_FILE_WR)) { 1727 /* request larger max_size from MDS? */ 1728 if (ci->i_wanted_max_size > ci->i_max_size && 1729 ci->i_wanted_max_size > ci->i_requested_max_size) { 1730 dout("requesting new max_size\n"); 1731 goto ack; 1732 } 1733 1734 /* approaching file_max? */ 1735 if ((inode->i_size << 1) >= ci->i_max_size && 1736 (ci->i_reported_size << 1) < ci->i_max_size) { 1737 dout("i_size approaching max_size\n"); 1738 goto ack; 1739 } 1740 } 1741 /* flush anything dirty? */ 1742 if (cap == ci->i_auth_cap) { 1743 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) { 1744 dout("flushing dirty caps\n"); 1745 goto ack; 1746 } 1747 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) { 1748 dout("flushing snap caps\n"); 1749 goto ack; 1750 } 1751 } 1752 1753 /* completed revocation? going down and there are no caps? */ 1754 if (revoking && (revoking & cap_used) == 0) { 1755 dout("completed revocation of %s\n", 1756 ceph_cap_string(cap->implemented & ~cap->issued)); 1757 goto ack; 1758 } 1759 1760 /* want more caps from mds? */ 1761 if (want & ~(cap->mds_wanted | cap->issued)) 1762 goto ack; 1763 1764 /* things we might delay */ 1765 if ((cap->issued & ~retain) == 0 && 1766 cap->mds_wanted == want) 1767 continue; /* nope, all good */ 1768 1769 if (is_delayed) 1770 goto ack; 1771 1772 /* delay? */ 1773 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1774 time_before(jiffies, ci->i_hold_caps_max)) { 1775 dout(" delaying issued %s -> %s, wanted %s -> %s\n", 1776 ceph_cap_string(cap->issued), 1777 ceph_cap_string(cap->issued & retain), 1778 ceph_cap_string(cap->mds_wanted), 1779 ceph_cap_string(want)); 1780 delayed++; 1781 continue; 1782 } 1783 1784 ack: 1785 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { 1786 dout(" skipping %p I_NOFLUSH set\n", inode); 1787 continue; 1788 } 1789 1790 if (session && session != cap->session) { 1791 dout("oops, wrong session %p mutex\n", session); 1792 mutex_unlock(&session->s_mutex); 1793 session = NULL; 1794 } 1795 if (!session) { 1796 session = cap->session; 1797 if (mutex_trylock(&session->s_mutex) == 0) { 1798 dout("inverting session/ino locks on %p\n", 1799 session); 1800 spin_unlock(&ci->i_ceph_lock); 1801 if (took_snap_rwsem) { 1802 up_read(&mdsc->snap_rwsem); 1803 took_snap_rwsem = 0; 1804 } 1805 mutex_lock(&session->s_mutex); 1806 goto retry; 1807 } 1808 } 1809 1810 /* kick flushing and flush snaps before sending normal 1811 * cap message */ 1812 if (cap == ci->i_auth_cap && 1813 (ci->i_ceph_flags & 1814 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) { 1815 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) { 1816 spin_lock(&mdsc->cap_dirty_lock); 1817 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1818 spin_unlock(&mdsc->cap_dirty_lock); 1819 __kick_flushing_caps(mdsc, session, ci, 1820 oldest_flush_tid); 1821 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 1822 } 1823 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) 1824 __ceph_flush_snaps(ci, session); 1825 1826 goto retry_locked; 1827 } 1828 1829 /* take snap_rwsem after session mutex */ 1830 if (!took_snap_rwsem) { 1831 if (down_read_trylock(&mdsc->snap_rwsem) == 0) { 1832 dout("inverting snap/in locks on %p\n", 1833 inode); 1834 spin_unlock(&ci->i_ceph_lock); 1835 down_read(&mdsc->snap_rwsem); 1836 took_snap_rwsem = 1; 1837 goto retry; 1838 } 1839 took_snap_rwsem = 1; 1840 } 1841 1842 if (cap == ci->i_auth_cap && ci->i_dirty_caps) { 1843 flushing = __mark_caps_flushing(inode, session, false, 1844 &flush_tid, 1845 &oldest_flush_tid); 1846 } else { 1847 flushing = 0; 1848 flush_tid = 0; 1849 spin_lock(&mdsc->cap_dirty_lock); 1850 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1851 spin_unlock(&mdsc->cap_dirty_lock); 1852 } 1853 1854 mds = cap->mds; /* remember mds, so we don't repeat */ 1855 sent++; 1856 1857 /* __send_cap drops i_ceph_lock */ 1858 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used, 1859 want, retain, flushing, 1860 flush_tid, oldest_flush_tid); 1861 goto retry; /* retake i_ceph_lock and restart our cap scan. */ 1862 } 1863 1864 /* 1865 * Reschedule delayed caps release if we delayed anything, 1866 * otherwise cancel. 1867 */ 1868 if (delayed && is_delayed) 1869 force_requeue = true; /* __send_cap delayed release; requeue */ 1870 if (!delayed && !is_delayed) 1871 __cap_delay_cancel(mdsc, ci); 1872 else if (!is_delayed || force_requeue) 1873 __cap_delay_requeue(mdsc, ci); 1874 1875 spin_unlock(&ci->i_ceph_lock); 1876 1877 if (queue_invalidate) 1878 ceph_queue_invalidate(inode); 1879 1880 if (session) 1881 mutex_unlock(&session->s_mutex); 1882 if (took_snap_rwsem) 1883 up_read(&mdsc->snap_rwsem); 1884 } 1885 1886 /* 1887 * Try to flush dirty caps back to the auth mds. 1888 */ 1889 static int try_flush_caps(struct inode *inode, u64 *ptid) 1890 { 1891 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 1892 struct ceph_inode_info *ci = ceph_inode(inode); 1893 struct ceph_mds_session *session = NULL; 1894 int flushing = 0; 1895 u64 flush_tid = 0, oldest_flush_tid = 0; 1896 1897 retry: 1898 spin_lock(&ci->i_ceph_lock); 1899 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { 1900 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode); 1901 goto out; 1902 } 1903 if (ci->i_dirty_caps && ci->i_auth_cap) { 1904 struct ceph_cap *cap = ci->i_auth_cap; 1905 int used = __ceph_caps_used(ci); 1906 int want = __ceph_caps_wanted(ci); 1907 int delayed; 1908 1909 if (!session || session != cap->session) { 1910 spin_unlock(&ci->i_ceph_lock); 1911 if (session) 1912 mutex_unlock(&session->s_mutex); 1913 session = cap->session; 1914 mutex_lock(&session->s_mutex); 1915 goto retry; 1916 } 1917 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) 1918 goto out; 1919 1920 flushing = __mark_caps_flushing(inode, session, true, 1921 &flush_tid, &oldest_flush_tid); 1922 1923 /* __send_cap drops i_ceph_lock */ 1924 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want, 1925 (cap->issued | cap->implemented), 1926 flushing, flush_tid, oldest_flush_tid); 1927 1928 if (delayed) { 1929 spin_lock(&ci->i_ceph_lock); 1930 __cap_delay_requeue(mdsc, ci); 1931 spin_unlock(&ci->i_ceph_lock); 1932 } 1933 } else { 1934 if (!list_empty(&ci->i_cap_flush_list)) { 1935 struct ceph_cap_flush *cf = 1936 list_last_entry(&ci->i_cap_flush_list, 1937 struct ceph_cap_flush, i_list); 1938 cf->wake = true; 1939 flush_tid = cf->tid; 1940 } 1941 flushing = ci->i_flushing_caps; 1942 spin_unlock(&ci->i_ceph_lock); 1943 } 1944 out: 1945 if (session) 1946 mutex_unlock(&session->s_mutex); 1947 1948 *ptid = flush_tid; 1949 return flushing; 1950 } 1951 1952 /* 1953 * Return true if we've flushed caps through the given flush_tid. 1954 */ 1955 static int caps_are_flushed(struct inode *inode, u64 flush_tid) 1956 { 1957 struct ceph_inode_info *ci = ceph_inode(inode); 1958 int ret = 1; 1959 1960 spin_lock(&ci->i_ceph_lock); 1961 if (!list_empty(&ci->i_cap_flush_list)) { 1962 struct ceph_cap_flush * cf = 1963 list_first_entry(&ci->i_cap_flush_list, 1964 struct ceph_cap_flush, i_list); 1965 if (cf->tid <= flush_tid) 1966 ret = 0; 1967 } 1968 spin_unlock(&ci->i_ceph_lock); 1969 return ret; 1970 } 1971 1972 /* 1973 * wait for any unsafe requests to complete. 1974 */ 1975 static int unsafe_request_wait(struct inode *inode) 1976 { 1977 struct ceph_inode_info *ci = ceph_inode(inode); 1978 struct ceph_mds_request *req1 = NULL, *req2 = NULL; 1979 int ret, err = 0; 1980 1981 spin_lock(&ci->i_unsafe_lock); 1982 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) { 1983 req1 = list_last_entry(&ci->i_unsafe_dirops, 1984 struct ceph_mds_request, 1985 r_unsafe_dir_item); 1986 ceph_mdsc_get_request(req1); 1987 } 1988 if (!list_empty(&ci->i_unsafe_iops)) { 1989 req2 = list_last_entry(&ci->i_unsafe_iops, 1990 struct ceph_mds_request, 1991 r_unsafe_target_item); 1992 ceph_mdsc_get_request(req2); 1993 } 1994 spin_unlock(&ci->i_unsafe_lock); 1995 1996 dout("unsafe_requeset_wait %p wait on tid %llu %llu\n", 1997 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL); 1998 if (req1) { 1999 ret = !wait_for_completion_timeout(&req1->r_safe_completion, 2000 ceph_timeout_jiffies(req1->r_timeout)); 2001 if (ret) 2002 err = -EIO; 2003 ceph_mdsc_put_request(req1); 2004 } 2005 if (req2) { 2006 ret = !wait_for_completion_timeout(&req2->r_safe_completion, 2007 ceph_timeout_jiffies(req2->r_timeout)); 2008 if (ret) 2009 err = -EIO; 2010 ceph_mdsc_put_request(req2); 2011 } 2012 return err; 2013 } 2014 2015 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync) 2016 { 2017 struct inode *inode = file->f_mapping->host; 2018 struct ceph_inode_info *ci = ceph_inode(inode); 2019 u64 flush_tid; 2020 int ret; 2021 int dirty; 2022 2023 dout("fsync %p%s\n", inode, datasync ? " datasync" : ""); 2024 2025 ceph_sync_write_wait(inode); 2026 2027 ret = filemap_write_and_wait_range(inode->i_mapping, start, end); 2028 if (ret < 0) 2029 goto out; 2030 2031 if (datasync) 2032 goto out; 2033 2034 inode_lock(inode); 2035 2036 dirty = try_flush_caps(inode, &flush_tid); 2037 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); 2038 2039 ret = unsafe_request_wait(inode); 2040 2041 /* 2042 * only wait on non-file metadata writeback (the mds 2043 * can recover size and mtime, so we don't need to 2044 * wait for that) 2045 */ 2046 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { 2047 ret = wait_event_interruptible(ci->i_cap_wq, 2048 caps_are_flushed(inode, flush_tid)); 2049 } 2050 inode_unlock(inode); 2051 out: 2052 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret); 2053 return ret; 2054 } 2055 2056 /* 2057 * Flush any dirty caps back to the mds. If we aren't asked to wait, 2058 * queue inode for flush but don't do so immediately, because we can 2059 * get by with fewer MDS messages if we wait for data writeback to 2060 * complete first. 2061 */ 2062 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc) 2063 { 2064 struct ceph_inode_info *ci = ceph_inode(inode); 2065 u64 flush_tid; 2066 int err = 0; 2067 int dirty; 2068 int wait = wbc->sync_mode == WB_SYNC_ALL; 2069 2070 dout("write_inode %p wait=%d\n", inode, wait); 2071 if (wait) { 2072 dirty = try_flush_caps(inode, &flush_tid); 2073 if (dirty) 2074 err = wait_event_interruptible(ci->i_cap_wq, 2075 caps_are_flushed(inode, flush_tid)); 2076 } else { 2077 struct ceph_mds_client *mdsc = 2078 ceph_sb_to_client(inode->i_sb)->mdsc; 2079 2080 spin_lock(&ci->i_ceph_lock); 2081 if (__ceph_caps_dirty(ci)) 2082 __cap_delay_requeue_front(mdsc, ci); 2083 spin_unlock(&ci->i_ceph_lock); 2084 } 2085 return err; 2086 } 2087 2088 static void __kick_flushing_caps(struct ceph_mds_client *mdsc, 2089 struct ceph_mds_session *session, 2090 struct ceph_inode_info *ci, 2091 u64 oldest_flush_tid) 2092 __releases(ci->i_ceph_lock) 2093 __acquires(ci->i_ceph_lock) 2094 { 2095 struct inode *inode = &ci->vfs_inode; 2096 struct ceph_cap *cap; 2097 struct ceph_cap_flush *cf; 2098 int ret; 2099 u64 first_tid = 0; 2100 2101 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) { 2102 if (cf->tid < first_tid) 2103 continue; 2104 2105 cap = ci->i_auth_cap; 2106 if (!(cap && cap->session == session)) { 2107 pr_err("%p auth cap %p not mds%d ???\n", 2108 inode, cap, session->s_mds); 2109 break; 2110 } 2111 2112 first_tid = cf->tid + 1; 2113 2114 if (cf->caps) { 2115 dout("kick_flushing_caps %p cap %p tid %llu %s\n", 2116 inode, cap, cf->tid, ceph_cap_string(cf->caps)); 2117 ci->i_ceph_flags |= CEPH_I_NODELAY; 2118 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, 2119 __ceph_caps_used(ci), 2120 __ceph_caps_wanted(ci), 2121 cap->issued | cap->implemented, 2122 cf->caps, cf->tid, oldest_flush_tid); 2123 if (ret) { 2124 pr_err("kick_flushing_caps: error sending " 2125 "cap flush, ino (%llx.%llx) " 2126 "tid %llu flushing %s\n", 2127 ceph_vinop(inode), cf->tid, 2128 ceph_cap_string(cf->caps)); 2129 } 2130 } else { 2131 struct ceph_cap_snap *capsnap = 2132 container_of(cf, struct ceph_cap_snap, 2133 cap_flush); 2134 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n", 2135 inode, capsnap, cf->tid, 2136 ceph_cap_string(capsnap->dirty)); 2137 2138 atomic_inc(&capsnap->nref); 2139 spin_unlock(&ci->i_ceph_lock); 2140 2141 ret = __send_flush_snap(inode, session, capsnap, cap->mseq, 2142 oldest_flush_tid); 2143 if (ret < 0) { 2144 pr_err("kick_flushing_caps: error sending " 2145 "cap flushsnap, ino (%llx.%llx) " 2146 "tid %llu follows %llu\n", 2147 ceph_vinop(inode), cf->tid, 2148 capsnap->follows); 2149 } 2150 2151 ceph_put_cap_snap(capsnap); 2152 } 2153 2154 spin_lock(&ci->i_ceph_lock); 2155 } 2156 } 2157 2158 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc, 2159 struct ceph_mds_session *session) 2160 { 2161 struct ceph_inode_info *ci; 2162 struct ceph_cap *cap; 2163 u64 oldest_flush_tid; 2164 2165 dout("early_kick_flushing_caps mds%d\n", session->s_mds); 2166 2167 spin_lock(&mdsc->cap_dirty_lock); 2168 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2169 spin_unlock(&mdsc->cap_dirty_lock); 2170 2171 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 2172 spin_lock(&ci->i_ceph_lock); 2173 cap = ci->i_auth_cap; 2174 if (!(cap && cap->session == session)) { 2175 pr_err("%p auth cap %p not mds%d ???\n", 2176 &ci->vfs_inode, cap, session->s_mds); 2177 spin_unlock(&ci->i_ceph_lock); 2178 continue; 2179 } 2180 2181 2182 /* 2183 * if flushing caps were revoked, we re-send the cap flush 2184 * in client reconnect stage. This guarantees MDS * processes 2185 * the cap flush message before issuing the flushing caps to 2186 * other client. 2187 */ 2188 if ((cap->issued & ci->i_flushing_caps) != 2189 ci->i_flushing_caps) { 2190 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 2191 __kick_flushing_caps(mdsc, session, ci, 2192 oldest_flush_tid); 2193 } else { 2194 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH; 2195 } 2196 2197 spin_unlock(&ci->i_ceph_lock); 2198 } 2199 } 2200 2201 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 2202 struct ceph_mds_session *session) 2203 { 2204 struct ceph_inode_info *ci; 2205 struct ceph_cap *cap; 2206 u64 oldest_flush_tid; 2207 2208 dout("kick_flushing_caps mds%d\n", session->s_mds); 2209 2210 spin_lock(&mdsc->cap_dirty_lock); 2211 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2212 spin_unlock(&mdsc->cap_dirty_lock); 2213 2214 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 2215 spin_lock(&ci->i_ceph_lock); 2216 cap = ci->i_auth_cap; 2217 if (!(cap && cap->session == session)) { 2218 pr_err("%p auth cap %p not mds%d ???\n", 2219 &ci->vfs_inode, cap, session->s_mds); 2220 spin_unlock(&ci->i_ceph_lock); 2221 continue; 2222 } 2223 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) { 2224 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 2225 __kick_flushing_caps(mdsc, session, ci, 2226 oldest_flush_tid); 2227 } 2228 spin_unlock(&ci->i_ceph_lock); 2229 } 2230 } 2231 2232 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc, 2233 struct ceph_mds_session *session, 2234 struct inode *inode) 2235 __releases(ci->i_ceph_lock) 2236 { 2237 struct ceph_inode_info *ci = ceph_inode(inode); 2238 struct ceph_cap *cap; 2239 2240 cap = ci->i_auth_cap; 2241 dout("kick_flushing_inode_caps %p flushing %s\n", inode, 2242 ceph_cap_string(ci->i_flushing_caps)); 2243 2244 if (!list_empty(&ci->i_cap_flush_list)) { 2245 u64 oldest_flush_tid; 2246 spin_lock(&mdsc->cap_dirty_lock); 2247 list_move_tail(&ci->i_flushing_item, 2248 &cap->session->s_cap_flushing); 2249 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2250 spin_unlock(&mdsc->cap_dirty_lock); 2251 2252 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 2253 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid); 2254 spin_unlock(&ci->i_ceph_lock); 2255 } else { 2256 spin_unlock(&ci->i_ceph_lock); 2257 } 2258 } 2259 2260 2261 /* 2262 * Take references to capabilities we hold, so that we don't release 2263 * them to the MDS prematurely. 2264 * 2265 * Protected by i_ceph_lock. 2266 */ 2267 static void __take_cap_refs(struct ceph_inode_info *ci, int got, 2268 bool snap_rwsem_locked) 2269 { 2270 if (got & CEPH_CAP_PIN) 2271 ci->i_pin_ref++; 2272 if (got & CEPH_CAP_FILE_RD) 2273 ci->i_rd_ref++; 2274 if (got & CEPH_CAP_FILE_CACHE) 2275 ci->i_rdcache_ref++; 2276 if (got & CEPH_CAP_FILE_WR) { 2277 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) { 2278 BUG_ON(!snap_rwsem_locked); 2279 ci->i_head_snapc = ceph_get_snap_context( 2280 ci->i_snap_realm->cached_context); 2281 } 2282 ci->i_wr_ref++; 2283 } 2284 if (got & CEPH_CAP_FILE_BUFFER) { 2285 if (ci->i_wb_ref == 0) 2286 ihold(&ci->vfs_inode); 2287 ci->i_wb_ref++; 2288 dout("__take_cap_refs %p wb %d -> %d (?)\n", 2289 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref); 2290 } 2291 } 2292 2293 /* 2294 * Try to grab cap references. Specify those refs we @want, and the 2295 * minimal set we @need. Also include the larger offset we are writing 2296 * to (when applicable), and check against max_size here as well. 2297 * Note that caller is responsible for ensuring max_size increases are 2298 * requested from the MDS. 2299 */ 2300 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want, 2301 loff_t endoff, bool nonblock, int *got, int *err) 2302 { 2303 struct inode *inode = &ci->vfs_inode; 2304 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 2305 int ret = 0; 2306 int have, implemented; 2307 int file_wanted; 2308 bool snap_rwsem_locked = false; 2309 2310 dout("get_cap_refs %p need %s want %s\n", inode, 2311 ceph_cap_string(need), ceph_cap_string(want)); 2312 2313 again: 2314 spin_lock(&ci->i_ceph_lock); 2315 2316 /* make sure file is actually open */ 2317 file_wanted = __ceph_caps_file_wanted(ci); 2318 if ((file_wanted & need) != need) { 2319 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n", 2320 ceph_cap_string(need), ceph_cap_string(file_wanted)); 2321 *err = -EBADF; 2322 ret = 1; 2323 goto out_unlock; 2324 } 2325 2326 /* finish pending truncate */ 2327 while (ci->i_truncate_pending) { 2328 spin_unlock(&ci->i_ceph_lock); 2329 if (snap_rwsem_locked) { 2330 up_read(&mdsc->snap_rwsem); 2331 snap_rwsem_locked = false; 2332 } 2333 __ceph_do_pending_vmtruncate(inode); 2334 spin_lock(&ci->i_ceph_lock); 2335 } 2336 2337 have = __ceph_caps_issued(ci, &implemented); 2338 2339 if (have & need & CEPH_CAP_FILE_WR) { 2340 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) { 2341 dout("get_cap_refs %p endoff %llu > maxsize %llu\n", 2342 inode, endoff, ci->i_max_size); 2343 if (endoff > ci->i_requested_max_size) { 2344 *err = -EAGAIN; 2345 ret = 1; 2346 } 2347 goto out_unlock; 2348 } 2349 /* 2350 * If a sync write is in progress, we must wait, so that we 2351 * can get a final snapshot value for size+mtime. 2352 */ 2353 if (__ceph_have_pending_cap_snap(ci)) { 2354 dout("get_cap_refs %p cap_snap_pending\n", inode); 2355 goto out_unlock; 2356 } 2357 } 2358 2359 if ((have & need) == need) { 2360 /* 2361 * Look at (implemented & ~have & not) so that we keep waiting 2362 * on transition from wanted -> needed caps. This is needed 2363 * for WRBUFFER|WR -> WR to avoid a new WR sync write from 2364 * going before a prior buffered writeback happens. 2365 */ 2366 int not = want & ~(have & need); 2367 int revoking = implemented & ~have; 2368 dout("get_cap_refs %p have %s but not %s (revoking %s)\n", 2369 inode, ceph_cap_string(have), ceph_cap_string(not), 2370 ceph_cap_string(revoking)); 2371 if ((revoking & not) == 0) { 2372 if (!snap_rwsem_locked && 2373 !ci->i_head_snapc && 2374 (need & CEPH_CAP_FILE_WR)) { 2375 if (!down_read_trylock(&mdsc->snap_rwsem)) { 2376 /* 2377 * we can not call down_read() when 2378 * task isn't in TASK_RUNNING state 2379 */ 2380 if (nonblock) { 2381 *err = -EAGAIN; 2382 ret = 1; 2383 goto out_unlock; 2384 } 2385 2386 spin_unlock(&ci->i_ceph_lock); 2387 down_read(&mdsc->snap_rwsem); 2388 snap_rwsem_locked = true; 2389 goto again; 2390 } 2391 snap_rwsem_locked = true; 2392 } 2393 *got = need | (have & want); 2394 if ((need & CEPH_CAP_FILE_RD) && 2395 !(*got & CEPH_CAP_FILE_CACHE)) 2396 ceph_disable_fscache_readpage(ci); 2397 __take_cap_refs(ci, *got, true); 2398 ret = 1; 2399 } 2400 } else { 2401 int session_readonly = false; 2402 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) { 2403 struct ceph_mds_session *s = ci->i_auth_cap->session; 2404 spin_lock(&s->s_cap_lock); 2405 session_readonly = s->s_readonly; 2406 spin_unlock(&s->s_cap_lock); 2407 } 2408 if (session_readonly) { 2409 dout("get_cap_refs %p needed %s but mds%d readonly\n", 2410 inode, ceph_cap_string(need), ci->i_auth_cap->mds); 2411 *err = -EROFS; 2412 ret = 1; 2413 goto out_unlock; 2414 } 2415 2416 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) { 2417 int mds_wanted; 2418 if (ACCESS_ONCE(mdsc->fsc->mount_state) == 2419 CEPH_MOUNT_SHUTDOWN) { 2420 dout("get_cap_refs %p forced umount\n", inode); 2421 *err = -EIO; 2422 ret = 1; 2423 goto out_unlock; 2424 } 2425 mds_wanted = __ceph_caps_mds_wanted(ci); 2426 if ((mds_wanted & need) != need) { 2427 dout("get_cap_refs %p caps were dropped" 2428 " (session killed?)\n", inode); 2429 *err = -ESTALE; 2430 ret = 1; 2431 goto out_unlock; 2432 } 2433 if ((mds_wanted & file_wanted) == 2434 (file_wanted & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR))) 2435 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED; 2436 } 2437 2438 dout("get_cap_refs %p have %s needed %s\n", inode, 2439 ceph_cap_string(have), ceph_cap_string(need)); 2440 } 2441 out_unlock: 2442 spin_unlock(&ci->i_ceph_lock); 2443 if (snap_rwsem_locked) 2444 up_read(&mdsc->snap_rwsem); 2445 2446 dout("get_cap_refs %p ret %d got %s\n", inode, 2447 ret, ceph_cap_string(*got)); 2448 return ret; 2449 } 2450 2451 /* 2452 * Check the offset we are writing up to against our current 2453 * max_size. If necessary, tell the MDS we want to write to 2454 * a larger offset. 2455 */ 2456 static void check_max_size(struct inode *inode, loff_t endoff) 2457 { 2458 struct ceph_inode_info *ci = ceph_inode(inode); 2459 int check = 0; 2460 2461 /* do we need to explicitly request a larger max_size? */ 2462 spin_lock(&ci->i_ceph_lock); 2463 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) { 2464 dout("write %p at large endoff %llu, req max_size\n", 2465 inode, endoff); 2466 ci->i_wanted_max_size = endoff; 2467 } 2468 /* duplicate ceph_check_caps()'s logic */ 2469 if (ci->i_auth_cap && 2470 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) && 2471 ci->i_wanted_max_size > ci->i_max_size && 2472 ci->i_wanted_max_size > ci->i_requested_max_size) 2473 check = 1; 2474 spin_unlock(&ci->i_ceph_lock); 2475 if (check) 2476 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2477 } 2478 2479 /* 2480 * Wait for caps, and take cap references. If we can't get a WR cap 2481 * due to a small max_size, make sure we check_max_size (and possibly 2482 * ask the mds) so we don't get hung up indefinitely. 2483 */ 2484 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, 2485 loff_t endoff, int *got, struct page **pinned_page) 2486 { 2487 int _got, ret, err = 0; 2488 2489 ret = ceph_pool_perm_check(ci, need); 2490 if (ret < 0) 2491 return ret; 2492 2493 while (true) { 2494 if (endoff > 0) 2495 check_max_size(&ci->vfs_inode, endoff); 2496 2497 err = 0; 2498 _got = 0; 2499 ret = try_get_cap_refs(ci, need, want, endoff, 2500 false, &_got, &err); 2501 if (ret) { 2502 if (err == -EAGAIN) 2503 continue; 2504 if (err < 0) 2505 ret = err; 2506 } else { 2507 ret = wait_event_interruptible(ci->i_cap_wq, 2508 try_get_cap_refs(ci, need, want, endoff, 2509 true, &_got, &err)); 2510 if (err == -EAGAIN) 2511 continue; 2512 if (err < 0) 2513 ret = err; 2514 } 2515 if (ret < 0) { 2516 if (err == -ESTALE) { 2517 /* session was killed, try renew caps */ 2518 ret = ceph_renew_caps(&ci->vfs_inode); 2519 if (ret == 0) 2520 continue; 2521 } 2522 return ret; 2523 } 2524 2525 if (ci->i_inline_version != CEPH_INLINE_NONE && 2526 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) && 2527 i_size_read(&ci->vfs_inode) > 0) { 2528 struct page *page = 2529 find_get_page(ci->vfs_inode.i_mapping, 0); 2530 if (page) { 2531 if (PageUptodate(page)) { 2532 *pinned_page = page; 2533 break; 2534 } 2535 put_page(page); 2536 } 2537 /* 2538 * drop cap refs first because getattr while 2539 * holding * caps refs can cause deadlock. 2540 */ 2541 ceph_put_cap_refs(ci, _got); 2542 _got = 0; 2543 2544 /* 2545 * getattr request will bring inline data into 2546 * page cache 2547 */ 2548 ret = __ceph_do_getattr(&ci->vfs_inode, NULL, 2549 CEPH_STAT_CAP_INLINE_DATA, 2550 true); 2551 if (ret < 0) 2552 return ret; 2553 continue; 2554 } 2555 break; 2556 } 2557 2558 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE)) 2559 ceph_fscache_revalidate_cookie(ci); 2560 2561 *got = _got; 2562 return 0; 2563 } 2564 2565 /* 2566 * Take cap refs. Caller must already know we hold at least one ref 2567 * on the caps in question or we don't know this is safe. 2568 */ 2569 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) 2570 { 2571 spin_lock(&ci->i_ceph_lock); 2572 __take_cap_refs(ci, caps, false); 2573 spin_unlock(&ci->i_ceph_lock); 2574 } 2575 2576 2577 /* 2578 * drop cap_snap that is not associated with any snapshot. 2579 * we don't need to send FLUSHSNAP message for it. 2580 */ 2581 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci, 2582 struct ceph_cap_snap *capsnap) 2583 { 2584 if (!capsnap->need_flush && 2585 !capsnap->writing && !capsnap->dirty_pages) { 2586 dout("dropping cap_snap %p follows %llu\n", 2587 capsnap, capsnap->follows); 2588 BUG_ON(capsnap->cap_flush.tid > 0); 2589 ceph_put_snap_context(capsnap->context); 2590 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps)) 2591 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; 2592 2593 list_del(&capsnap->ci_item); 2594 ceph_put_cap_snap(capsnap); 2595 return 1; 2596 } 2597 return 0; 2598 } 2599 2600 /* 2601 * Release cap refs. 2602 * 2603 * If we released the last ref on any given cap, call ceph_check_caps 2604 * to release (or schedule a release). 2605 * 2606 * If we are releasing a WR cap (from a sync write), finalize any affected 2607 * cap_snap, and wake up any waiters. 2608 */ 2609 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had) 2610 { 2611 struct inode *inode = &ci->vfs_inode; 2612 int last = 0, put = 0, flushsnaps = 0, wake = 0; 2613 2614 spin_lock(&ci->i_ceph_lock); 2615 if (had & CEPH_CAP_PIN) 2616 --ci->i_pin_ref; 2617 if (had & CEPH_CAP_FILE_RD) 2618 if (--ci->i_rd_ref == 0) 2619 last++; 2620 if (had & CEPH_CAP_FILE_CACHE) 2621 if (--ci->i_rdcache_ref == 0) 2622 last++; 2623 if (had & CEPH_CAP_FILE_BUFFER) { 2624 if (--ci->i_wb_ref == 0) { 2625 last++; 2626 put++; 2627 } 2628 dout("put_cap_refs %p wb %d -> %d (?)\n", 2629 inode, ci->i_wb_ref+1, ci->i_wb_ref); 2630 } 2631 if (had & CEPH_CAP_FILE_WR) 2632 if (--ci->i_wr_ref == 0) { 2633 last++; 2634 if (__ceph_have_pending_cap_snap(ci)) { 2635 struct ceph_cap_snap *capsnap = 2636 list_last_entry(&ci->i_cap_snaps, 2637 struct ceph_cap_snap, 2638 ci_item); 2639 capsnap->writing = 0; 2640 if (ceph_try_drop_cap_snap(ci, capsnap)) 2641 put++; 2642 else if (__ceph_finish_cap_snap(ci, capsnap)) 2643 flushsnaps = 1; 2644 wake = 1; 2645 } 2646 if (ci->i_wrbuffer_ref_head == 0 && 2647 ci->i_dirty_caps == 0 && 2648 ci->i_flushing_caps == 0) { 2649 BUG_ON(!ci->i_head_snapc); 2650 ceph_put_snap_context(ci->i_head_snapc); 2651 ci->i_head_snapc = NULL; 2652 } 2653 /* see comment in __ceph_remove_cap() */ 2654 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) 2655 drop_inode_snap_realm(ci); 2656 } 2657 spin_unlock(&ci->i_ceph_lock); 2658 2659 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had), 2660 last ? " last" : "", put ? " put" : ""); 2661 2662 if (last && !flushsnaps) 2663 ceph_check_caps(ci, 0, NULL); 2664 else if (flushsnaps) 2665 ceph_flush_snaps(ci, NULL); 2666 if (wake) 2667 wake_up_all(&ci->i_cap_wq); 2668 while (put-- > 0) 2669 iput(inode); 2670 } 2671 2672 /* 2673 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap 2674 * context. Adjust per-snap dirty page accounting as appropriate. 2675 * Once all dirty data for a cap_snap is flushed, flush snapped file 2676 * metadata back to the MDS. If we dropped the last ref, call 2677 * ceph_check_caps. 2678 */ 2679 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr, 2680 struct ceph_snap_context *snapc) 2681 { 2682 struct inode *inode = &ci->vfs_inode; 2683 struct ceph_cap_snap *capsnap = NULL; 2684 int put = 0; 2685 bool last = false; 2686 bool found = false; 2687 bool flush_snaps = false; 2688 bool complete_capsnap = false; 2689 2690 spin_lock(&ci->i_ceph_lock); 2691 ci->i_wrbuffer_ref -= nr; 2692 if (ci->i_wrbuffer_ref == 0) { 2693 last = true; 2694 put++; 2695 } 2696 2697 if (ci->i_head_snapc == snapc) { 2698 ci->i_wrbuffer_ref_head -= nr; 2699 if (ci->i_wrbuffer_ref_head == 0 && 2700 ci->i_wr_ref == 0 && 2701 ci->i_dirty_caps == 0 && 2702 ci->i_flushing_caps == 0) { 2703 BUG_ON(!ci->i_head_snapc); 2704 ceph_put_snap_context(ci->i_head_snapc); 2705 ci->i_head_snapc = NULL; 2706 } 2707 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n", 2708 inode, 2709 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr, 2710 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 2711 last ? " LAST" : ""); 2712 } else { 2713 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2714 if (capsnap->context == snapc) { 2715 found = true; 2716 break; 2717 } 2718 } 2719 BUG_ON(!found); 2720 capsnap->dirty_pages -= nr; 2721 if (capsnap->dirty_pages == 0) { 2722 complete_capsnap = true; 2723 if (!capsnap->writing) { 2724 if (ceph_try_drop_cap_snap(ci, capsnap)) { 2725 put++; 2726 } else { 2727 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; 2728 flush_snaps = true; 2729 } 2730 } 2731 } 2732 dout("put_wrbuffer_cap_refs on %p cap_snap %p " 2733 " snap %lld %d/%d -> %d/%d %s%s\n", 2734 inode, capsnap, capsnap->context->seq, 2735 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, 2736 ci->i_wrbuffer_ref, capsnap->dirty_pages, 2737 last ? " (wrbuffer last)" : "", 2738 complete_capsnap ? " (complete capsnap)" : ""); 2739 } 2740 2741 spin_unlock(&ci->i_ceph_lock); 2742 2743 if (last) { 2744 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2745 } else if (flush_snaps) { 2746 ceph_flush_snaps(ci, NULL); 2747 } 2748 if (complete_capsnap) 2749 wake_up_all(&ci->i_cap_wq); 2750 while (put-- > 0) 2751 iput(inode); 2752 } 2753 2754 /* 2755 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP. 2756 */ 2757 static void invalidate_aliases(struct inode *inode) 2758 { 2759 struct dentry *dn, *prev = NULL; 2760 2761 dout("invalidate_aliases inode %p\n", inode); 2762 d_prune_aliases(inode); 2763 /* 2764 * For non-directory inode, d_find_alias() only returns 2765 * hashed dentry. After calling d_invalidate(), the 2766 * dentry becomes unhashed. 2767 * 2768 * For directory inode, d_find_alias() can return 2769 * unhashed dentry. But directory inode should have 2770 * one alias at most. 2771 */ 2772 while ((dn = d_find_alias(inode))) { 2773 if (dn == prev) { 2774 dput(dn); 2775 break; 2776 } 2777 d_invalidate(dn); 2778 if (prev) 2779 dput(prev); 2780 prev = dn; 2781 } 2782 if (prev) 2783 dput(prev); 2784 } 2785 2786 /* 2787 * Handle a cap GRANT message from the MDS. (Note that a GRANT may 2788 * actually be a revocation if it specifies a smaller cap set.) 2789 * 2790 * caller holds s_mutex and i_ceph_lock, we drop both. 2791 */ 2792 static void handle_cap_grant(struct ceph_mds_client *mdsc, 2793 struct inode *inode, struct ceph_mds_caps *grant, 2794 struct ceph_string **pns, u64 inline_version, 2795 void *inline_data, u32 inline_len, 2796 struct ceph_buffer *xattr_buf, 2797 struct ceph_mds_session *session, 2798 struct ceph_cap *cap, int issued) 2799 __releases(ci->i_ceph_lock) 2800 __releases(mdsc->snap_rwsem) 2801 { 2802 struct ceph_inode_info *ci = ceph_inode(inode); 2803 int mds = session->s_mds; 2804 int seq = le32_to_cpu(grant->seq); 2805 int newcaps = le32_to_cpu(grant->caps); 2806 int used, wanted, dirty; 2807 u64 size = le64_to_cpu(grant->size); 2808 u64 max_size = le64_to_cpu(grant->max_size); 2809 struct timespec mtime, atime, ctime; 2810 int check_caps = 0; 2811 bool wake = false; 2812 bool writeback = false; 2813 bool queue_trunc = false; 2814 bool queue_invalidate = false; 2815 bool deleted_inode = false; 2816 bool fill_inline = false; 2817 2818 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n", 2819 inode, cap, mds, seq, ceph_cap_string(newcaps)); 2820 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size, 2821 inode->i_size); 2822 2823 2824 /* 2825 * auth mds of the inode changed. we received the cap export message, 2826 * but still haven't received the cap import message. handle_cap_export 2827 * updated the new auth MDS' cap. 2828 * 2829 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message 2830 * that was sent before the cap import message. So don't remove caps. 2831 */ 2832 if (ceph_seq_cmp(seq, cap->seq) <= 0) { 2833 WARN_ON(cap != ci->i_auth_cap); 2834 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id)); 2835 seq = cap->seq; 2836 newcaps |= cap->issued; 2837 } 2838 2839 /* 2840 * If CACHE is being revoked, and we have no dirty buffers, 2841 * try to invalidate (once). (If there are dirty buffers, we 2842 * will invalidate _after_ writeback.) 2843 */ 2844 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */ 2845 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && 2846 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && 2847 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) { 2848 if (try_nonblocking_invalidate(inode)) { 2849 /* there were locked pages.. invalidate later 2850 in a separate thread. */ 2851 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) { 2852 queue_invalidate = true; 2853 ci->i_rdcache_revoking = ci->i_rdcache_gen; 2854 } 2855 } 2856 } 2857 2858 /* side effects now are allowed */ 2859 cap->cap_gen = session->s_cap_gen; 2860 cap->seq = seq; 2861 2862 __check_cap_issue(ci, cap, newcaps); 2863 2864 if ((newcaps & CEPH_CAP_AUTH_SHARED) && 2865 (issued & CEPH_CAP_AUTH_EXCL) == 0) { 2866 inode->i_mode = le32_to_cpu(grant->mode); 2867 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid)); 2868 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid)); 2869 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode, 2870 from_kuid(&init_user_ns, inode->i_uid), 2871 from_kgid(&init_user_ns, inode->i_gid)); 2872 } 2873 2874 if ((newcaps & CEPH_CAP_AUTH_SHARED) && 2875 (issued & CEPH_CAP_LINK_EXCL) == 0) { 2876 set_nlink(inode, le32_to_cpu(grant->nlink)); 2877 if (inode->i_nlink == 0 && 2878 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL))) 2879 deleted_inode = true; 2880 } 2881 2882 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) { 2883 int len = le32_to_cpu(grant->xattr_len); 2884 u64 version = le64_to_cpu(grant->xattr_version); 2885 2886 if (version > ci->i_xattrs.version) { 2887 dout(" got new xattrs v%llu on %p len %d\n", 2888 version, inode, len); 2889 if (ci->i_xattrs.blob) 2890 ceph_buffer_put(ci->i_xattrs.blob); 2891 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf); 2892 ci->i_xattrs.version = version; 2893 ceph_forget_all_cached_acls(inode); 2894 } 2895 } 2896 2897 if (newcaps & CEPH_CAP_ANY_RD) { 2898 /* ctime/mtime/atime? */ 2899 ceph_decode_timespec(&mtime, &grant->mtime); 2900 ceph_decode_timespec(&atime, &grant->atime); 2901 ceph_decode_timespec(&ctime, &grant->ctime); 2902 ceph_fill_file_time(inode, issued, 2903 le32_to_cpu(grant->time_warp_seq), 2904 &ctime, &mtime, &atime); 2905 } 2906 2907 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) { 2908 /* file layout may have changed */ 2909 s64 old_pool = ci->i_layout.pool_id; 2910 struct ceph_string *old_ns; 2911 2912 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout); 2913 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns, 2914 lockdep_is_held(&ci->i_ceph_lock)); 2915 rcu_assign_pointer(ci->i_layout.pool_ns, *pns); 2916 2917 if (ci->i_layout.pool_id != old_pool || *pns != old_ns) 2918 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM; 2919 2920 *pns = old_ns; 2921 2922 /* size/truncate_seq? */ 2923 queue_trunc = ceph_fill_file_size(inode, issued, 2924 le32_to_cpu(grant->truncate_seq), 2925 le64_to_cpu(grant->truncate_size), 2926 size); 2927 /* max size increase? */ 2928 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) { 2929 dout("max_size %lld -> %llu\n", 2930 ci->i_max_size, max_size); 2931 ci->i_max_size = max_size; 2932 if (max_size >= ci->i_wanted_max_size) { 2933 ci->i_wanted_max_size = 0; /* reset */ 2934 ci->i_requested_max_size = 0; 2935 } 2936 wake = true; 2937 } 2938 } 2939 2940 /* check cap bits */ 2941 wanted = __ceph_caps_wanted(ci); 2942 used = __ceph_caps_used(ci); 2943 dirty = __ceph_caps_dirty(ci); 2944 dout(" my wanted = %s, used = %s, dirty %s\n", 2945 ceph_cap_string(wanted), 2946 ceph_cap_string(used), 2947 ceph_cap_string(dirty)); 2948 if (wanted != le32_to_cpu(grant->wanted)) { 2949 dout("mds wanted %s -> %s\n", 2950 ceph_cap_string(le32_to_cpu(grant->wanted)), 2951 ceph_cap_string(wanted)); 2952 /* imported cap may not have correct mds_wanted */ 2953 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) 2954 check_caps = 1; 2955 } 2956 2957 /* revocation, grant, or no-op? */ 2958 if (cap->issued & ~newcaps) { 2959 int revoking = cap->issued & ~newcaps; 2960 2961 dout("revocation: %s -> %s (revoking %s)\n", 2962 ceph_cap_string(cap->issued), 2963 ceph_cap_string(newcaps), 2964 ceph_cap_string(revoking)); 2965 if (revoking & used & CEPH_CAP_FILE_BUFFER) 2966 writeback = true; /* initiate writeback; will delay ack */ 2967 else if (revoking == CEPH_CAP_FILE_CACHE && 2968 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && 2969 queue_invalidate) 2970 ; /* do nothing yet, invalidation will be queued */ 2971 else if (cap == ci->i_auth_cap) 2972 check_caps = 1; /* check auth cap only */ 2973 else 2974 check_caps = 2; /* check all caps */ 2975 cap->issued = newcaps; 2976 cap->implemented |= newcaps; 2977 } else if (cap->issued == newcaps) { 2978 dout("caps unchanged: %s -> %s\n", 2979 ceph_cap_string(cap->issued), ceph_cap_string(newcaps)); 2980 } else { 2981 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued), 2982 ceph_cap_string(newcaps)); 2983 /* non-auth MDS is revoking the newly grant caps ? */ 2984 if (cap == ci->i_auth_cap && 2985 __ceph_caps_revoking_other(ci, cap, newcaps)) 2986 check_caps = 2; 2987 2988 cap->issued = newcaps; 2989 cap->implemented |= newcaps; /* add bits only, to 2990 * avoid stepping on a 2991 * pending revocation */ 2992 wake = true; 2993 } 2994 BUG_ON(cap->issued & ~cap->implemented); 2995 2996 if (inline_version > 0 && inline_version >= ci->i_inline_version) { 2997 ci->i_inline_version = inline_version; 2998 if (ci->i_inline_version != CEPH_INLINE_NONE && 2999 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO))) 3000 fill_inline = true; 3001 } 3002 3003 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) { 3004 if (newcaps & ~issued) 3005 wake = true; 3006 kick_flushing_inode_caps(mdsc, session, inode); 3007 up_read(&mdsc->snap_rwsem); 3008 } else { 3009 spin_unlock(&ci->i_ceph_lock); 3010 } 3011 3012 if (fill_inline) 3013 ceph_fill_inline_data(inode, NULL, inline_data, inline_len); 3014 3015 if (queue_trunc) 3016 ceph_queue_vmtruncate(inode); 3017 3018 if (writeback) 3019 /* 3020 * queue inode for writeback: we can't actually call 3021 * filemap_write_and_wait, etc. from message handler 3022 * context. 3023 */ 3024 ceph_queue_writeback(inode); 3025 if (queue_invalidate) 3026 ceph_queue_invalidate(inode); 3027 if (deleted_inode) 3028 invalidate_aliases(inode); 3029 if (wake) 3030 wake_up_all(&ci->i_cap_wq); 3031 3032 if (check_caps == 1) 3033 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY, 3034 session); 3035 else if (check_caps == 2) 3036 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session); 3037 else 3038 mutex_unlock(&session->s_mutex); 3039 } 3040 3041 /* 3042 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the 3043 * MDS has been safely committed. 3044 */ 3045 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid, 3046 struct ceph_mds_caps *m, 3047 struct ceph_mds_session *session, 3048 struct ceph_cap *cap) 3049 __releases(ci->i_ceph_lock) 3050 { 3051 struct ceph_inode_info *ci = ceph_inode(inode); 3052 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 3053 struct ceph_cap_flush *cf, *tmp_cf; 3054 LIST_HEAD(to_remove); 3055 unsigned seq = le32_to_cpu(m->seq); 3056 int dirty = le32_to_cpu(m->dirty); 3057 int cleaned = 0; 3058 bool drop = false; 3059 bool wake_ci = 0; 3060 bool wake_mdsc = 0; 3061 3062 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) { 3063 if (cf->tid == flush_tid) 3064 cleaned = cf->caps; 3065 if (cf->caps == 0) /* capsnap */ 3066 continue; 3067 if (cf->tid <= flush_tid) { 3068 if (__finish_cap_flush(NULL, ci, cf)) 3069 wake_ci = true; 3070 list_add_tail(&cf->i_list, &to_remove); 3071 } else { 3072 cleaned &= ~cf->caps; 3073 if (!cleaned) 3074 break; 3075 } 3076 } 3077 3078 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," 3079 " flushing %s -> %s\n", 3080 inode, session->s_mds, seq, ceph_cap_string(dirty), 3081 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), 3082 ceph_cap_string(ci->i_flushing_caps & ~cleaned)); 3083 3084 if (list_empty(&to_remove) && !cleaned) 3085 goto out; 3086 3087 ci->i_flushing_caps &= ~cleaned; 3088 3089 spin_lock(&mdsc->cap_dirty_lock); 3090 3091 list_for_each_entry(cf, &to_remove, i_list) { 3092 if (__finish_cap_flush(mdsc, NULL, cf)) 3093 wake_mdsc = true; 3094 } 3095 3096 if (ci->i_flushing_caps == 0) { 3097 if (list_empty(&ci->i_cap_flush_list)) { 3098 list_del_init(&ci->i_flushing_item); 3099 if (!list_empty(&session->s_cap_flushing)) { 3100 dout(" mds%d still flushing cap on %p\n", 3101 session->s_mds, 3102 &list_first_entry(&session->s_cap_flushing, 3103 struct ceph_inode_info, 3104 i_flushing_item)->vfs_inode); 3105 } 3106 } 3107 mdsc->num_cap_flushing--; 3108 dout(" inode %p now !flushing\n", inode); 3109 3110 if (ci->i_dirty_caps == 0) { 3111 dout(" inode %p now clean\n", inode); 3112 BUG_ON(!list_empty(&ci->i_dirty_item)); 3113 drop = true; 3114 if (ci->i_wr_ref == 0 && 3115 ci->i_wrbuffer_ref_head == 0) { 3116 BUG_ON(!ci->i_head_snapc); 3117 ceph_put_snap_context(ci->i_head_snapc); 3118 ci->i_head_snapc = NULL; 3119 } 3120 } else { 3121 BUG_ON(list_empty(&ci->i_dirty_item)); 3122 } 3123 } 3124 spin_unlock(&mdsc->cap_dirty_lock); 3125 3126 out: 3127 spin_unlock(&ci->i_ceph_lock); 3128 3129 while (!list_empty(&to_remove)) { 3130 cf = list_first_entry(&to_remove, 3131 struct ceph_cap_flush, i_list); 3132 list_del(&cf->i_list); 3133 ceph_free_cap_flush(cf); 3134 } 3135 3136 if (wake_ci) 3137 wake_up_all(&ci->i_cap_wq); 3138 if (wake_mdsc) 3139 wake_up_all(&mdsc->cap_flushing_wq); 3140 if (drop) 3141 iput(inode); 3142 } 3143 3144 /* 3145 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can 3146 * throw away our cap_snap. 3147 * 3148 * Caller hold s_mutex. 3149 */ 3150 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid, 3151 struct ceph_mds_caps *m, 3152 struct ceph_mds_session *session) 3153 { 3154 struct ceph_inode_info *ci = ceph_inode(inode); 3155 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 3156 u64 follows = le64_to_cpu(m->snap_follows); 3157 struct ceph_cap_snap *capsnap; 3158 bool flushed = false; 3159 bool wake_ci = false; 3160 bool wake_mdsc = false; 3161 3162 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n", 3163 inode, ci, session->s_mds, follows); 3164 3165 spin_lock(&ci->i_ceph_lock); 3166 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 3167 if (capsnap->follows == follows) { 3168 if (capsnap->cap_flush.tid != flush_tid) { 3169 dout(" cap_snap %p follows %lld tid %lld !=" 3170 " %lld\n", capsnap, follows, 3171 flush_tid, capsnap->cap_flush.tid); 3172 break; 3173 } 3174 flushed = true; 3175 break; 3176 } else { 3177 dout(" skipping cap_snap %p follows %lld\n", 3178 capsnap, capsnap->follows); 3179 } 3180 } 3181 if (flushed) { 3182 WARN_ON(capsnap->dirty_pages || capsnap->writing); 3183 dout(" removing %p cap_snap %p follows %lld\n", 3184 inode, capsnap, follows); 3185 list_del(&capsnap->ci_item); 3186 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush)) 3187 wake_ci = true; 3188 3189 spin_lock(&mdsc->cap_dirty_lock); 3190 3191 if (list_empty(&ci->i_cap_flush_list)) 3192 list_del_init(&ci->i_flushing_item); 3193 3194 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush)) 3195 wake_mdsc = true; 3196 3197 spin_unlock(&mdsc->cap_dirty_lock); 3198 } 3199 spin_unlock(&ci->i_ceph_lock); 3200 if (flushed) { 3201 ceph_put_snap_context(capsnap->context); 3202 ceph_put_cap_snap(capsnap); 3203 if (wake_ci) 3204 wake_up_all(&ci->i_cap_wq); 3205 if (wake_mdsc) 3206 wake_up_all(&mdsc->cap_flushing_wq); 3207 iput(inode); 3208 } 3209 } 3210 3211 /* 3212 * Handle TRUNC from MDS, indicating file truncation. 3213 * 3214 * caller hold s_mutex. 3215 */ 3216 static void handle_cap_trunc(struct inode *inode, 3217 struct ceph_mds_caps *trunc, 3218 struct ceph_mds_session *session) 3219 __releases(ci->i_ceph_lock) 3220 { 3221 struct ceph_inode_info *ci = ceph_inode(inode); 3222 int mds = session->s_mds; 3223 int seq = le32_to_cpu(trunc->seq); 3224 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq); 3225 u64 truncate_size = le64_to_cpu(trunc->truncate_size); 3226 u64 size = le64_to_cpu(trunc->size); 3227 int implemented = 0; 3228 int dirty = __ceph_caps_dirty(ci); 3229 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented); 3230 int queue_trunc = 0; 3231 3232 issued |= implemented | dirty; 3233 3234 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n", 3235 inode, mds, seq, truncate_size, truncate_seq); 3236 queue_trunc = ceph_fill_file_size(inode, issued, 3237 truncate_seq, truncate_size, size); 3238 spin_unlock(&ci->i_ceph_lock); 3239 3240 if (queue_trunc) 3241 ceph_queue_vmtruncate(inode); 3242 } 3243 3244 /* 3245 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a 3246 * different one. If we are the most recent migration we've seen (as 3247 * indicated by mseq), make note of the migrating cap bits for the 3248 * duration (until we see the corresponding IMPORT). 3249 * 3250 * caller holds s_mutex 3251 */ 3252 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex, 3253 struct ceph_mds_cap_peer *ph, 3254 struct ceph_mds_session *session) 3255 { 3256 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 3257 struct ceph_mds_session *tsession = NULL; 3258 struct ceph_cap *cap, *tcap, *new_cap = NULL; 3259 struct ceph_inode_info *ci = ceph_inode(inode); 3260 u64 t_cap_id; 3261 unsigned mseq = le32_to_cpu(ex->migrate_seq); 3262 unsigned t_seq, t_mseq; 3263 int target, issued; 3264 int mds = session->s_mds; 3265 3266 if (ph) { 3267 t_cap_id = le64_to_cpu(ph->cap_id); 3268 t_seq = le32_to_cpu(ph->seq); 3269 t_mseq = le32_to_cpu(ph->mseq); 3270 target = le32_to_cpu(ph->mds); 3271 } else { 3272 t_cap_id = t_seq = t_mseq = 0; 3273 target = -1; 3274 } 3275 3276 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n", 3277 inode, ci, mds, mseq, target); 3278 retry: 3279 spin_lock(&ci->i_ceph_lock); 3280 cap = __get_cap_for_mds(ci, mds); 3281 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id)) 3282 goto out_unlock; 3283 3284 if (target < 0) { 3285 __ceph_remove_cap(cap, false); 3286 if (!ci->i_auth_cap) 3287 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED; 3288 goto out_unlock; 3289 } 3290 3291 /* 3292 * now we know we haven't received the cap import message yet 3293 * because the exported cap still exist. 3294 */ 3295 3296 issued = cap->issued; 3297 WARN_ON(issued != cap->implemented); 3298 3299 tcap = __get_cap_for_mds(ci, target); 3300 if (tcap) { 3301 /* already have caps from the target */ 3302 if (tcap->cap_id != t_cap_id || 3303 ceph_seq_cmp(tcap->seq, t_seq) < 0) { 3304 dout(" updating import cap %p mds%d\n", tcap, target); 3305 tcap->cap_id = t_cap_id; 3306 tcap->seq = t_seq - 1; 3307 tcap->issue_seq = t_seq - 1; 3308 tcap->mseq = t_mseq; 3309 tcap->issued |= issued; 3310 tcap->implemented |= issued; 3311 if (cap == ci->i_auth_cap) 3312 ci->i_auth_cap = tcap; 3313 if (!list_empty(&ci->i_cap_flush_list) && 3314 ci->i_auth_cap == tcap) { 3315 spin_lock(&mdsc->cap_dirty_lock); 3316 list_move_tail(&ci->i_flushing_item, 3317 &tcap->session->s_cap_flushing); 3318 spin_unlock(&mdsc->cap_dirty_lock); 3319 } 3320 } 3321 __ceph_remove_cap(cap, false); 3322 goto out_unlock; 3323 } else if (tsession) { 3324 /* add placeholder for the export tagert */ 3325 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0; 3326 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0, 3327 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap); 3328 3329 __ceph_remove_cap(cap, false); 3330 goto out_unlock; 3331 } 3332 3333 spin_unlock(&ci->i_ceph_lock); 3334 mutex_unlock(&session->s_mutex); 3335 3336 /* open target session */ 3337 tsession = ceph_mdsc_open_export_target_session(mdsc, target); 3338 if (!IS_ERR(tsession)) { 3339 if (mds > target) { 3340 mutex_lock(&session->s_mutex); 3341 mutex_lock_nested(&tsession->s_mutex, 3342 SINGLE_DEPTH_NESTING); 3343 } else { 3344 mutex_lock(&tsession->s_mutex); 3345 mutex_lock_nested(&session->s_mutex, 3346 SINGLE_DEPTH_NESTING); 3347 } 3348 new_cap = ceph_get_cap(mdsc, NULL); 3349 } else { 3350 WARN_ON(1); 3351 tsession = NULL; 3352 target = -1; 3353 } 3354 goto retry; 3355 3356 out_unlock: 3357 spin_unlock(&ci->i_ceph_lock); 3358 mutex_unlock(&session->s_mutex); 3359 if (tsession) { 3360 mutex_unlock(&tsession->s_mutex); 3361 ceph_put_mds_session(tsession); 3362 } 3363 if (new_cap) 3364 ceph_put_cap(mdsc, new_cap); 3365 } 3366 3367 /* 3368 * Handle cap IMPORT. 3369 * 3370 * caller holds s_mutex. acquires i_ceph_lock 3371 */ 3372 static void handle_cap_import(struct ceph_mds_client *mdsc, 3373 struct inode *inode, struct ceph_mds_caps *im, 3374 struct ceph_mds_cap_peer *ph, 3375 struct ceph_mds_session *session, 3376 struct ceph_cap **target_cap, int *old_issued) 3377 __acquires(ci->i_ceph_lock) 3378 { 3379 struct ceph_inode_info *ci = ceph_inode(inode); 3380 struct ceph_cap *cap, *ocap, *new_cap = NULL; 3381 int mds = session->s_mds; 3382 int issued; 3383 unsigned caps = le32_to_cpu(im->caps); 3384 unsigned wanted = le32_to_cpu(im->wanted); 3385 unsigned seq = le32_to_cpu(im->seq); 3386 unsigned mseq = le32_to_cpu(im->migrate_seq); 3387 u64 realmino = le64_to_cpu(im->realm); 3388 u64 cap_id = le64_to_cpu(im->cap_id); 3389 u64 p_cap_id; 3390 int peer; 3391 3392 if (ph) { 3393 p_cap_id = le64_to_cpu(ph->cap_id); 3394 peer = le32_to_cpu(ph->mds); 3395 } else { 3396 p_cap_id = 0; 3397 peer = -1; 3398 } 3399 3400 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n", 3401 inode, ci, mds, mseq, peer); 3402 3403 retry: 3404 spin_lock(&ci->i_ceph_lock); 3405 cap = __get_cap_for_mds(ci, mds); 3406 if (!cap) { 3407 if (!new_cap) { 3408 spin_unlock(&ci->i_ceph_lock); 3409 new_cap = ceph_get_cap(mdsc, NULL); 3410 goto retry; 3411 } 3412 cap = new_cap; 3413 } else { 3414 if (new_cap) { 3415 ceph_put_cap(mdsc, new_cap); 3416 new_cap = NULL; 3417 } 3418 } 3419 3420 __ceph_caps_issued(ci, &issued); 3421 issued |= __ceph_caps_dirty(ci); 3422 3423 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq, 3424 realmino, CEPH_CAP_FLAG_AUTH, &new_cap); 3425 3426 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL; 3427 if (ocap && ocap->cap_id == p_cap_id) { 3428 dout(" remove export cap %p mds%d flags %d\n", 3429 ocap, peer, ph->flags); 3430 if ((ph->flags & CEPH_CAP_FLAG_AUTH) && 3431 (ocap->seq != le32_to_cpu(ph->seq) || 3432 ocap->mseq != le32_to_cpu(ph->mseq))) { 3433 pr_err("handle_cap_import: mismatched seq/mseq: " 3434 "ino (%llx.%llx) mds%d seq %d mseq %d " 3435 "importer mds%d has peer seq %d mseq %d\n", 3436 ceph_vinop(inode), peer, ocap->seq, 3437 ocap->mseq, mds, le32_to_cpu(ph->seq), 3438 le32_to_cpu(ph->mseq)); 3439 } 3440 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE)); 3441 } 3442 3443 /* make sure we re-request max_size, if necessary */ 3444 ci->i_wanted_max_size = 0; 3445 ci->i_requested_max_size = 0; 3446 3447 *old_issued = issued; 3448 *target_cap = cap; 3449 } 3450 3451 /* 3452 * Handle a caps message from the MDS. 3453 * 3454 * Identify the appropriate session, inode, and call the right handler 3455 * based on the cap op. 3456 */ 3457 void ceph_handle_caps(struct ceph_mds_session *session, 3458 struct ceph_msg *msg) 3459 { 3460 struct ceph_mds_client *mdsc = session->s_mdsc; 3461 struct super_block *sb = mdsc->fsc->sb; 3462 struct inode *inode; 3463 struct ceph_inode_info *ci; 3464 struct ceph_cap *cap; 3465 struct ceph_mds_caps *h; 3466 struct ceph_mds_cap_peer *peer = NULL; 3467 struct ceph_snap_realm *realm = NULL; 3468 struct ceph_string *pool_ns = NULL; 3469 int mds = session->s_mds; 3470 int op, issued; 3471 u32 seq, mseq; 3472 struct ceph_vino vino; 3473 u64 tid; 3474 u64 inline_version = 0; 3475 void *inline_data = NULL; 3476 u32 inline_len = 0; 3477 void *snaptrace; 3478 size_t snaptrace_len; 3479 void *p, *end; 3480 3481 dout("handle_caps from mds%d\n", mds); 3482 3483 /* decode */ 3484 end = msg->front.iov_base + msg->front.iov_len; 3485 tid = le64_to_cpu(msg->hdr.tid); 3486 if (msg->front.iov_len < sizeof(*h)) 3487 goto bad; 3488 h = msg->front.iov_base; 3489 op = le32_to_cpu(h->op); 3490 vino.ino = le64_to_cpu(h->ino); 3491 vino.snap = CEPH_NOSNAP; 3492 seq = le32_to_cpu(h->seq); 3493 mseq = le32_to_cpu(h->migrate_seq); 3494 3495 snaptrace = h + 1; 3496 snaptrace_len = le32_to_cpu(h->snap_trace_len); 3497 p = snaptrace + snaptrace_len; 3498 3499 if (le16_to_cpu(msg->hdr.version) >= 2) { 3500 u32 flock_len; 3501 ceph_decode_32_safe(&p, end, flock_len, bad); 3502 if (p + flock_len > end) 3503 goto bad; 3504 p += flock_len; 3505 } 3506 3507 if (le16_to_cpu(msg->hdr.version) >= 3) { 3508 if (op == CEPH_CAP_OP_IMPORT) { 3509 if (p + sizeof(*peer) > end) 3510 goto bad; 3511 peer = p; 3512 p += sizeof(*peer); 3513 } else if (op == CEPH_CAP_OP_EXPORT) { 3514 /* recorded in unused fields */ 3515 peer = (void *)&h->size; 3516 } 3517 } 3518 3519 if (le16_to_cpu(msg->hdr.version) >= 4) { 3520 ceph_decode_64_safe(&p, end, inline_version, bad); 3521 ceph_decode_32_safe(&p, end, inline_len, bad); 3522 if (p + inline_len > end) 3523 goto bad; 3524 inline_data = p; 3525 p += inline_len; 3526 } 3527 3528 if (le16_to_cpu(msg->hdr.version) >= 8) { 3529 u64 flush_tid; 3530 u32 caller_uid, caller_gid; 3531 u32 osd_epoch_barrier; 3532 u32 pool_ns_len; 3533 /* version >= 5 */ 3534 ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad); 3535 /* version >= 6 */ 3536 ceph_decode_64_safe(&p, end, flush_tid, bad); 3537 /* version >= 7 */ 3538 ceph_decode_32_safe(&p, end, caller_uid, bad); 3539 ceph_decode_32_safe(&p, end, caller_gid, bad); 3540 /* version >= 8 */ 3541 ceph_decode_32_safe(&p, end, pool_ns_len, bad); 3542 if (pool_ns_len > 0) { 3543 ceph_decode_need(&p, end, pool_ns_len, bad); 3544 pool_ns = ceph_find_or_create_string(p, pool_ns_len); 3545 p += pool_ns_len; 3546 } 3547 } 3548 3549 /* lookup ino */ 3550 inode = ceph_find_inode(sb, vino); 3551 ci = ceph_inode(inode); 3552 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino, 3553 vino.snap, inode); 3554 3555 mutex_lock(&session->s_mutex); 3556 session->s_seq++; 3557 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, 3558 (unsigned)seq); 3559 3560 if (!inode) { 3561 dout(" i don't have ino %llx\n", vino.ino); 3562 3563 if (op == CEPH_CAP_OP_IMPORT) { 3564 cap = ceph_get_cap(mdsc, NULL); 3565 cap->cap_ino = vino.ino; 3566 cap->queue_release = 1; 3567 cap->cap_id = le64_to_cpu(h->cap_id); 3568 cap->mseq = mseq; 3569 cap->seq = seq; 3570 spin_lock(&session->s_cap_lock); 3571 list_add_tail(&cap->session_caps, 3572 &session->s_cap_releases); 3573 session->s_num_cap_releases++; 3574 spin_unlock(&session->s_cap_lock); 3575 } 3576 goto flush_cap_releases; 3577 } 3578 3579 /* these will work even if we don't have a cap yet */ 3580 switch (op) { 3581 case CEPH_CAP_OP_FLUSHSNAP_ACK: 3582 handle_cap_flushsnap_ack(inode, tid, h, session); 3583 goto done; 3584 3585 case CEPH_CAP_OP_EXPORT: 3586 handle_cap_export(inode, h, peer, session); 3587 goto done_unlocked; 3588 3589 case CEPH_CAP_OP_IMPORT: 3590 realm = NULL; 3591 if (snaptrace_len) { 3592 down_write(&mdsc->snap_rwsem); 3593 ceph_update_snap_trace(mdsc, snaptrace, 3594 snaptrace + snaptrace_len, 3595 false, &realm); 3596 downgrade_write(&mdsc->snap_rwsem); 3597 } else { 3598 down_read(&mdsc->snap_rwsem); 3599 } 3600 handle_cap_import(mdsc, inode, h, peer, session, 3601 &cap, &issued); 3602 handle_cap_grant(mdsc, inode, h, &pool_ns, 3603 inline_version, inline_data, inline_len, 3604 msg->middle, session, cap, issued); 3605 if (realm) 3606 ceph_put_snap_realm(mdsc, realm); 3607 goto done_unlocked; 3608 } 3609 3610 /* the rest require a cap */ 3611 spin_lock(&ci->i_ceph_lock); 3612 cap = __get_cap_for_mds(ceph_inode(inode), mds); 3613 if (!cap) { 3614 dout(" no cap on %p ino %llx.%llx from mds%d\n", 3615 inode, ceph_ino(inode), ceph_snap(inode), mds); 3616 spin_unlock(&ci->i_ceph_lock); 3617 goto flush_cap_releases; 3618 } 3619 3620 /* note that each of these drops i_ceph_lock for us */ 3621 switch (op) { 3622 case CEPH_CAP_OP_REVOKE: 3623 case CEPH_CAP_OP_GRANT: 3624 __ceph_caps_issued(ci, &issued); 3625 issued |= __ceph_caps_dirty(ci); 3626 handle_cap_grant(mdsc, inode, h, &pool_ns, 3627 inline_version, inline_data, inline_len, 3628 msg->middle, session, cap, issued); 3629 goto done_unlocked; 3630 3631 case CEPH_CAP_OP_FLUSH_ACK: 3632 handle_cap_flush_ack(inode, tid, h, session, cap); 3633 break; 3634 3635 case CEPH_CAP_OP_TRUNC: 3636 handle_cap_trunc(inode, h, session); 3637 break; 3638 3639 default: 3640 spin_unlock(&ci->i_ceph_lock); 3641 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op, 3642 ceph_cap_op_name(op)); 3643 } 3644 3645 goto done; 3646 3647 flush_cap_releases: 3648 /* 3649 * send any cap release message to try to move things 3650 * along for the mds (who clearly thinks we still have this 3651 * cap). 3652 */ 3653 ceph_send_cap_releases(mdsc, session); 3654 3655 done: 3656 mutex_unlock(&session->s_mutex); 3657 done_unlocked: 3658 iput(inode); 3659 ceph_put_string(pool_ns); 3660 return; 3661 3662 bad: 3663 pr_err("ceph_handle_caps: corrupt message\n"); 3664 ceph_msg_dump(msg); 3665 return; 3666 } 3667 3668 /* 3669 * Delayed work handler to process end of delayed cap release LRU list. 3670 */ 3671 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc) 3672 { 3673 struct ceph_inode_info *ci; 3674 int flags = CHECK_CAPS_NODELAY; 3675 3676 dout("check_delayed_caps\n"); 3677 while (1) { 3678 spin_lock(&mdsc->cap_delay_lock); 3679 if (list_empty(&mdsc->cap_delay_list)) 3680 break; 3681 ci = list_first_entry(&mdsc->cap_delay_list, 3682 struct ceph_inode_info, 3683 i_cap_delay_list); 3684 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 && 3685 time_before(jiffies, ci->i_hold_caps_max)) 3686 break; 3687 list_del_init(&ci->i_cap_delay_list); 3688 spin_unlock(&mdsc->cap_delay_lock); 3689 dout("check_delayed_caps on %p\n", &ci->vfs_inode); 3690 ceph_check_caps(ci, flags, NULL); 3691 } 3692 spin_unlock(&mdsc->cap_delay_lock); 3693 } 3694 3695 /* 3696 * Flush all dirty caps to the mds 3697 */ 3698 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc) 3699 { 3700 struct ceph_inode_info *ci; 3701 struct inode *inode; 3702 3703 dout("flush_dirty_caps\n"); 3704 spin_lock(&mdsc->cap_dirty_lock); 3705 while (!list_empty(&mdsc->cap_dirty)) { 3706 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info, 3707 i_dirty_item); 3708 inode = &ci->vfs_inode; 3709 ihold(inode); 3710 dout("flush_dirty_caps %p\n", inode); 3711 spin_unlock(&mdsc->cap_dirty_lock); 3712 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL); 3713 iput(inode); 3714 spin_lock(&mdsc->cap_dirty_lock); 3715 } 3716 spin_unlock(&mdsc->cap_dirty_lock); 3717 dout("flush_dirty_caps done\n"); 3718 } 3719 3720 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode) 3721 { 3722 int i; 3723 int bits = (fmode << 1) | 1; 3724 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { 3725 if (bits & (1 << i)) 3726 ci->i_nr_by_mode[i]++; 3727 } 3728 } 3729 3730 /* 3731 * Drop open file reference. If we were the last open file, 3732 * we may need to release capabilities to the MDS (or schedule 3733 * their delayed release). 3734 */ 3735 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode) 3736 { 3737 int i, last = 0; 3738 int bits = (fmode << 1) | 1; 3739 spin_lock(&ci->i_ceph_lock); 3740 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { 3741 if (bits & (1 << i)) { 3742 BUG_ON(ci->i_nr_by_mode[i] == 0); 3743 if (--ci->i_nr_by_mode[i] == 0) 3744 last++; 3745 } 3746 } 3747 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n", 3748 &ci->vfs_inode, fmode, 3749 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1], 3750 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]); 3751 spin_unlock(&ci->i_ceph_lock); 3752 3753 if (last && ci->i_vino.snap == CEPH_NOSNAP) 3754 ceph_check_caps(ci, 0, NULL); 3755 } 3756 3757 /* 3758 * Helpers for embedding cap and dentry lease releases into mds 3759 * requests. 3760 * 3761 * @force is used by dentry_release (below) to force inclusion of a 3762 * record for the directory inode, even when there aren't any caps to 3763 * drop. 3764 */ 3765 int ceph_encode_inode_release(void **p, struct inode *inode, 3766 int mds, int drop, int unless, int force) 3767 { 3768 struct ceph_inode_info *ci = ceph_inode(inode); 3769 struct ceph_cap *cap; 3770 struct ceph_mds_request_release *rel = *p; 3771 int used, dirty; 3772 int ret = 0; 3773 3774 spin_lock(&ci->i_ceph_lock); 3775 used = __ceph_caps_used(ci); 3776 dirty = __ceph_caps_dirty(ci); 3777 3778 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n", 3779 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop), 3780 ceph_cap_string(unless)); 3781 3782 /* only drop unused, clean caps */ 3783 drop &= ~(used | dirty); 3784 3785 cap = __get_cap_for_mds(ci, mds); 3786 if (cap && __cap_is_valid(cap)) { 3787 if (force || 3788 ((cap->issued & drop) && 3789 (cap->issued & unless) == 0)) { 3790 if ((cap->issued & drop) && 3791 (cap->issued & unless) == 0) { 3792 int wanted = __ceph_caps_wanted(ci); 3793 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0) 3794 wanted |= cap->mds_wanted; 3795 dout("encode_inode_release %p cap %p " 3796 "%s -> %s, wanted %s -> %s\n", inode, cap, 3797 ceph_cap_string(cap->issued), 3798 ceph_cap_string(cap->issued & ~drop), 3799 ceph_cap_string(cap->mds_wanted), 3800 ceph_cap_string(wanted)); 3801 3802 cap->issued &= ~drop; 3803 cap->implemented &= ~drop; 3804 cap->mds_wanted = wanted; 3805 } else { 3806 dout("encode_inode_release %p cap %p %s" 3807 " (force)\n", inode, cap, 3808 ceph_cap_string(cap->issued)); 3809 } 3810 3811 rel->ino = cpu_to_le64(ceph_ino(inode)); 3812 rel->cap_id = cpu_to_le64(cap->cap_id); 3813 rel->seq = cpu_to_le32(cap->seq); 3814 rel->issue_seq = cpu_to_le32(cap->issue_seq); 3815 rel->mseq = cpu_to_le32(cap->mseq); 3816 rel->caps = cpu_to_le32(cap->implemented); 3817 rel->wanted = cpu_to_le32(cap->mds_wanted); 3818 rel->dname_len = 0; 3819 rel->dname_seq = 0; 3820 *p += sizeof(*rel); 3821 ret = 1; 3822 } else { 3823 dout("encode_inode_release %p cap %p %s\n", 3824 inode, cap, ceph_cap_string(cap->issued)); 3825 } 3826 } 3827 spin_unlock(&ci->i_ceph_lock); 3828 return ret; 3829 } 3830 3831 int ceph_encode_dentry_release(void **p, struct dentry *dentry, 3832 int mds, int drop, int unless) 3833 { 3834 struct inode *dir = d_inode(dentry->d_parent); 3835 struct ceph_mds_request_release *rel = *p; 3836 struct ceph_dentry_info *di = ceph_dentry(dentry); 3837 int force = 0; 3838 int ret; 3839 3840 /* 3841 * force an record for the directory caps if we have a dentry lease. 3842 * this is racy (can't take i_ceph_lock and d_lock together), but it 3843 * doesn't have to be perfect; the mds will revoke anything we don't 3844 * release. 3845 */ 3846 spin_lock(&dentry->d_lock); 3847 if (di->lease_session && di->lease_session->s_mds == mds) 3848 force = 1; 3849 spin_unlock(&dentry->d_lock); 3850 3851 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force); 3852 3853 spin_lock(&dentry->d_lock); 3854 if (ret && di->lease_session && di->lease_session->s_mds == mds) { 3855 dout("encode_dentry_release %p mds%d seq %d\n", 3856 dentry, mds, (int)di->lease_seq); 3857 rel->dname_len = cpu_to_le32(dentry->d_name.len); 3858 memcpy(*p, dentry->d_name.name, dentry->d_name.len); 3859 *p += dentry->d_name.len; 3860 rel->dname_seq = cpu_to_le32(di->lease_seq); 3861 __ceph_mdsc_drop_dentry_lease(dentry); 3862 } 3863 spin_unlock(&dentry->d_lock); 3864 return ret; 3865 } 3866