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 = NULL; 1351 int mds; 1352 1353 dout("ceph_flush_snaps %p\n", inode); 1354 if (psession) 1355 session = *psession; 1356 retry: 1357 spin_lock(&ci->i_ceph_lock); 1358 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) { 1359 dout(" no capsnap needs flush, doing nothing\n"); 1360 goto out; 1361 } 1362 if (!ci->i_auth_cap) { 1363 dout(" no auth cap (migrating?), doing nothing\n"); 1364 goto out; 1365 } 1366 1367 mds = ci->i_auth_cap->session->s_mds; 1368 if (session && session->s_mds != mds) { 1369 dout(" oops, wrong session %p mutex\n", session); 1370 mutex_unlock(&session->s_mutex); 1371 ceph_put_mds_session(session); 1372 session = NULL; 1373 } 1374 if (!session) { 1375 spin_unlock(&ci->i_ceph_lock); 1376 mutex_lock(&mdsc->mutex); 1377 session = __ceph_lookup_mds_session(mdsc, mds); 1378 mutex_unlock(&mdsc->mutex); 1379 if (session) { 1380 dout(" inverting session/ino locks on %p\n", session); 1381 mutex_lock(&session->s_mutex); 1382 } 1383 goto retry; 1384 } 1385 1386 __ceph_flush_snaps(ci, session); 1387 out: 1388 spin_unlock(&ci->i_ceph_lock); 1389 1390 if (psession) { 1391 *psession = session; 1392 } else { 1393 mutex_unlock(&session->s_mutex); 1394 ceph_put_mds_session(session); 1395 } 1396 /* we flushed them all; remove this inode from the queue */ 1397 spin_lock(&mdsc->snap_flush_lock); 1398 list_del_init(&ci->i_snap_flush_item); 1399 spin_unlock(&mdsc->snap_flush_lock); 1400 } 1401 1402 /* 1403 * Mark caps dirty. If inode is newly dirty, return the dirty flags. 1404 * Caller is then responsible for calling __mark_inode_dirty with the 1405 * returned flags value. 1406 */ 1407 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask, 1408 struct ceph_cap_flush **pcf) 1409 { 1410 struct ceph_mds_client *mdsc = 1411 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; 1412 struct inode *inode = &ci->vfs_inode; 1413 int was = ci->i_dirty_caps; 1414 int dirty = 0; 1415 1416 if (!ci->i_auth_cap) { 1417 pr_warn("__mark_dirty_caps %p %llx mask %s, " 1418 "but no auth cap (session was closed?)\n", 1419 inode, ceph_ino(inode), ceph_cap_string(mask)); 1420 return 0; 1421 } 1422 1423 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode, 1424 ceph_cap_string(mask), ceph_cap_string(was), 1425 ceph_cap_string(was | mask)); 1426 ci->i_dirty_caps |= mask; 1427 if (was == 0) { 1428 WARN_ON_ONCE(ci->i_prealloc_cap_flush); 1429 swap(ci->i_prealloc_cap_flush, *pcf); 1430 1431 if (!ci->i_head_snapc) { 1432 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem)); 1433 ci->i_head_snapc = ceph_get_snap_context( 1434 ci->i_snap_realm->cached_context); 1435 } 1436 dout(" inode %p now dirty snapc %p auth cap %p\n", 1437 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap); 1438 BUG_ON(!list_empty(&ci->i_dirty_item)); 1439 spin_lock(&mdsc->cap_dirty_lock); 1440 list_add(&ci->i_dirty_item, &mdsc->cap_dirty); 1441 spin_unlock(&mdsc->cap_dirty_lock); 1442 if (ci->i_flushing_caps == 0) { 1443 ihold(inode); 1444 dirty |= I_DIRTY_SYNC; 1445 } 1446 } else { 1447 WARN_ON_ONCE(!ci->i_prealloc_cap_flush); 1448 } 1449 BUG_ON(list_empty(&ci->i_dirty_item)); 1450 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) && 1451 (mask & CEPH_CAP_FILE_BUFFER)) 1452 dirty |= I_DIRTY_DATASYNC; 1453 __cap_delay_requeue(mdsc, ci); 1454 return dirty; 1455 } 1456 1457 struct ceph_cap_flush *ceph_alloc_cap_flush(void) 1458 { 1459 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL); 1460 } 1461 1462 void ceph_free_cap_flush(struct ceph_cap_flush *cf) 1463 { 1464 if (cf) 1465 kmem_cache_free(ceph_cap_flush_cachep, cf); 1466 } 1467 1468 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc) 1469 { 1470 if (!list_empty(&mdsc->cap_flush_list)) { 1471 struct ceph_cap_flush *cf = 1472 list_first_entry(&mdsc->cap_flush_list, 1473 struct ceph_cap_flush, g_list); 1474 return cf->tid; 1475 } 1476 return 0; 1477 } 1478 1479 /* 1480 * Remove cap_flush from the mdsc's or inode's flushing cap list. 1481 * Return true if caller needs to wake up flush waiters. 1482 */ 1483 static bool __finish_cap_flush(struct ceph_mds_client *mdsc, 1484 struct ceph_inode_info *ci, 1485 struct ceph_cap_flush *cf) 1486 { 1487 struct ceph_cap_flush *prev; 1488 bool wake = cf->wake; 1489 if (mdsc) { 1490 /* are there older pending cap flushes? */ 1491 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) { 1492 prev = list_prev_entry(cf, g_list); 1493 prev->wake = true; 1494 wake = false; 1495 } 1496 list_del(&cf->g_list); 1497 } else if (ci) { 1498 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) { 1499 prev = list_prev_entry(cf, i_list); 1500 prev->wake = true; 1501 wake = false; 1502 } 1503 list_del(&cf->i_list); 1504 } else { 1505 BUG_ON(1); 1506 } 1507 return wake; 1508 } 1509 1510 /* 1511 * Add dirty inode to the flushing list. Assigned a seq number so we 1512 * can wait for caps to flush without starving. 1513 * 1514 * Called under i_ceph_lock. 1515 */ 1516 static int __mark_caps_flushing(struct inode *inode, 1517 struct ceph_mds_session *session, bool wake, 1518 u64 *flush_tid, u64 *oldest_flush_tid) 1519 { 1520 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 1521 struct ceph_inode_info *ci = ceph_inode(inode); 1522 struct ceph_cap_flush *cf = NULL; 1523 int flushing; 1524 1525 BUG_ON(ci->i_dirty_caps == 0); 1526 BUG_ON(list_empty(&ci->i_dirty_item)); 1527 BUG_ON(!ci->i_prealloc_cap_flush); 1528 1529 flushing = ci->i_dirty_caps; 1530 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", 1531 ceph_cap_string(flushing), 1532 ceph_cap_string(ci->i_flushing_caps), 1533 ceph_cap_string(ci->i_flushing_caps | flushing)); 1534 ci->i_flushing_caps |= flushing; 1535 ci->i_dirty_caps = 0; 1536 dout(" inode %p now !dirty\n", inode); 1537 1538 swap(cf, ci->i_prealloc_cap_flush); 1539 cf->caps = flushing; 1540 cf->wake = wake; 1541 1542 spin_lock(&mdsc->cap_dirty_lock); 1543 list_del_init(&ci->i_dirty_item); 1544 1545 cf->tid = ++mdsc->last_cap_flush_tid; 1546 list_add_tail(&cf->g_list, &mdsc->cap_flush_list); 1547 *oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1548 1549 if (list_empty(&ci->i_flushing_item)) { 1550 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1551 mdsc->num_cap_flushing++; 1552 } 1553 spin_unlock(&mdsc->cap_dirty_lock); 1554 1555 list_add_tail(&cf->i_list, &ci->i_cap_flush_list); 1556 1557 *flush_tid = cf->tid; 1558 return flushing; 1559 } 1560 1561 /* 1562 * try to invalidate mapping pages without blocking. 1563 */ 1564 static int try_nonblocking_invalidate(struct inode *inode) 1565 { 1566 struct ceph_inode_info *ci = ceph_inode(inode); 1567 u32 invalidating_gen = ci->i_rdcache_gen; 1568 1569 spin_unlock(&ci->i_ceph_lock); 1570 invalidate_mapping_pages(&inode->i_data, 0, -1); 1571 spin_lock(&ci->i_ceph_lock); 1572 1573 if (inode->i_data.nrpages == 0 && 1574 invalidating_gen == ci->i_rdcache_gen) { 1575 /* success. */ 1576 dout("try_nonblocking_invalidate %p success\n", inode); 1577 /* save any racing async invalidate some trouble */ 1578 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1; 1579 return 0; 1580 } 1581 dout("try_nonblocking_invalidate %p failed\n", inode); 1582 return -1; 1583 } 1584 1585 /* 1586 * Swiss army knife function to examine currently used and wanted 1587 * versus held caps. Release, flush, ack revoked caps to mds as 1588 * appropriate. 1589 * 1590 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay 1591 * cap release further. 1592 * CHECK_CAPS_AUTHONLY - we should only check the auth cap 1593 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without 1594 * further delay. 1595 */ 1596 void ceph_check_caps(struct ceph_inode_info *ci, int flags, 1597 struct ceph_mds_session *session) 1598 { 1599 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode); 1600 struct ceph_mds_client *mdsc = fsc->mdsc; 1601 struct inode *inode = &ci->vfs_inode; 1602 struct ceph_cap *cap; 1603 u64 flush_tid, oldest_flush_tid; 1604 int file_wanted, used, cap_used; 1605 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ 1606 int issued, implemented, want, retain, revoking, flushing = 0; 1607 int mds = -1; /* keep track of how far we've gone through i_caps list 1608 to avoid an infinite loop on retry */ 1609 struct rb_node *p; 1610 int delayed = 0, sent = 0, num; 1611 bool is_delayed = flags & CHECK_CAPS_NODELAY; 1612 bool queue_invalidate = false; 1613 bool force_requeue = false; 1614 bool tried_invalidate = false; 1615 1616 /* if we are unmounting, flush any unused caps immediately. */ 1617 if (mdsc->stopping) 1618 is_delayed = 1; 1619 1620 spin_lock(&ci->i_ceph_lock); 1621 1622 if (ci->i_ceph_flags & CEPH_I_FLUSH) 1623 flags |= CHECK_CAPS_FLUSH; 1624 1625 goto retry_locked; 1626 retry: 1627 spin_lock(&ci->i_ceph_lock); 1628 retry_locked: 1629 file_wanted = __ceph_caps_file_wanted(ci); 1630 used = __ceph_caps_used(ci); 1631 issued = __ceph_caps_issued(ci, &implemented); 1632 revoking = implemented & ~issued; 1633 1634 want = file_wanted; 1635 retain = file_wanted | used | CEPH_CAP_PIN; 1636 if (!mdsc->stopping && inode->i_nlink > 0) { 1637 if (file_wanted) { 1638 retain |= CEPH_CAP_ANY; /* be greedy */ 1639 } else if (S_ISDIR(inode->i_mode) && 1640 (issued & CEPH_CAP_FILE_SHARED) && 1641 __ceph_dir_is_complete(ci)) { 1642 /* 1643 * If a directory is complete, we want to keep 1644 * the exclusive cap. So that MDS does not end up 1645 * revoking the shared cap on every create/unlink 1646 * operation. 1647 */ 1648 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL; 1649 retain |= want; 1650 } else { 1651 1652 retain |= CEPH_CAP_ANY_SHARED; 1653 /* 1654 * keep RD only if we didn't have the file open RW, 1655 * because then the mds would revoke it anyway to 1656 * journal max_size=0. 1657 */ 1658 if (ci->i_max_size == 0) 1659 retain |= CEPH_CAP_ANY_RD; 1660 } 1661 } 1662 1663 dout("check_caps %p file_want %s used %s dirty %s flushing %s" 1664 " issued %s revoking %s retain %s %s%s%s\n", inode, 1665 ceph_cap_string(file_wanted), 1666 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps), 1667 ceph_cap_string(ci->i_flushing_caps), 1668 ceph_cap_string(issued), ceph_cap_string(revoking), 1669 ceph_cap_string(retain), 1670 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "", 1671 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "", 1672 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : ""); 1673 1674 /* 1675 * If we no longer need to hold onto old our caps, and we may 1676 * have cached pages, but don't want them, then try to invalidate. 1677 * If we fail, it's because pages are locked.... try again later. 1678 */ 1679 if ((!is_delayed || mdsc->stopping) && 1680 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */ 1681 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */ 1682 inode->i_data.nrpages && /* have cached pages */ 1683 (revoking & (CEPH_CAP_FILE_CACHE| 1684 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */ 1685 !tried_invalidate) { 1686 dout("check_caps trying to invalidate on %p\n", inode); 1687 if (try_nonblocking_invalidate(inode) < 0) { 1688 if (revoking & (CEPH_CAP_FILE_CACHE| 1689 CEPH_CAP_FILE_LAZYIO)) { 1690 dout("check_caps queuing invalidate\n"); 1691 queue_invalidate = true; 1692 ci->i_rdcache_revoking = ci->i_rdcache_gen; 1693 } else { 1694 dout("check_caps failed to invalidate pages\n"); 1695 /* we failed to invalidate pages. check these 1696 caps again later. */ 1697 force_requeue = true; 1698 __cap_set_timeouts(mdsc, ci); 1699 } 1700 } 1701 tried_invalidate = true; 1702 goto retry_locked; 1703 } 1704 1705 num = 0; 1706 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { 1707 cap = rb_entry(p, struct ceph_cap, ci_node); 1708 num++; 1709 1710 /* avoid looping forever */ 1711 if (mds >= cap->mds || 1712 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap)) 1713 continue; 1714 1715 /* NOTE: no side-effects allowed, until we take s_mutex */ 1716 1717 cap_used = used; 1718 if (ci->i_auth_cap && cap != ci->i_auth_cap) 1719 cap_used &= ~ci->i_auth_cap->issued; 1720 1721 revoking = cap->implemented & ~cap->issued; 1722 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n", 1723 cap->mds, cap, ceph_cap_string(cap_used), 1724 ceph_cap_string(cap->issued), 1725 ceph_cap_string(cap->implemented), 1726 ceph_cap_string(revoking)); 1727 1728 if (cap == ci->i_auth_cap && 1729 (cap->issued & CEPH_CAP_FILE_WR)) { 1730 /* request larger max_size from MDS? */ 1731 if (ci->i_wanted_max_size > ci->i_max_size && 1732 ci->i_wanted_max_size > ci->i_requested_max_size) { 1733 dout("requesting new max_size\n"); 1734 goto ack; 1735 } 1736 1737 /* approaching file_max? */ 1738 if ((inode->i_size << 1) >= ci->i_max_size && 1739 (ci->i_reported_size << 1) < ci->i_max_size) { 1740 dout("i_size approaching max_size\n"); 1741 goto ack; 1742 } 1743 } 1744 /* flush anything dirty? */ 1745 if (cap == ci->i_auth_cap) { 1746 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) { 1747 dout("flushing dirty caps\n"); 1748 goto ack; 1749 } 1750 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) { 1751 dout("flushing snap caps\n"); 1752 goto ack; 1753 } 1754 } 1755 1756 /* completed revocation? going down and there are no caps? */ 1757 if (revoking && (revoking & cap_used) == 0) { 1758 dout("completed revocation of %s\n", 1759 ceph_cap_string(cap->implemented & ~cap->issued)); 1760 goto ack; 1761 } 1762 1763 /* want more caps from mds? */ 1764 if (want & ~(cap->mds_wanted | cap->issued)) 1765 goto ack; 1766 1767 /* things we might delay */ 1768 if ((cap->issued & ~retain) == 0 && 1769 cap->mds_wanted == want) 1770 continue; /* nope, all good */ 1771 1772 if (is_delayed) 1773 goto ack; 1774 1775 /* delay? */ 1776 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && 1777 time_before(jiffies, ci->i_hold_caps_max)) { 1778 dout(" delaying issued %s -> %s, wanted %s -> %s\n", 1779 ceph_cap_string(cap->issued), 1780 ceph_cap_string(cap->issued & retain), 1781 ceph_cap_string(cap->mds_wanted), 1782 ceph_cap_string(want)); 1783 delayed++; 1784 continue; 1785 } 1786 1787 ack: 1788 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { 1789 dout(" skipping %p I_NOFLUSH set\n", inode); 1790 continue; 1791 } 1792 1793 if (session && session != cap->session) { 1794 dout("oops, wrong session %p mutex\n", session); 1795 mutex_unlock(&session->s_mutex); 1796 session = NULL; 1797 } 1798 if (!session) { 1799 session = cap->session; 1800 if (mutex_trylock(&session->s_mutex) == 0) { 1801 dout("inverting session/ino locks on %p\n", 1802 session); 1803 spin_unlock(&ci->i_ceph_lock); 1804 if (took_snap_rwsem) { 1805 up_read(&mdsc->snap_rwsem); 1806 took_snap_rwsem = 0; 1807 } 1808 mutex_lock(&session->s_mutex); 1809 goto retry; 1810 } 1811 } 1812 1813 /* kick flushing and flush snaps before sending normal 1814 * cap message */ 1815 if (cap == ci->i_auth_cap && 1816 (ci->i_ceph_flags & 1817 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) { 1818 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) { 1819 spin_lock(&mdsc->cap_dirty_lock); 1820 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1821 spin_unlock(&mdsc->cap_dirty_lock); 1822 __kick_flushing_caps(mdsc, session, ci, 1823 oldest_flush_tid); 1824 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 1825 } 1826 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) 1827 __ceph_flush_snaps(ci, session); 1828 1829 goto retry_locked; 1830 } 1831 1832 /* take snap_rwsem after session mutex */ 1833 if (!took_snap_rwsem) { 1834 if (down_read_trylock(&mdsc->snap_rwsem) == 0) { 1835 dout("inverting snap/in locks on %p\n", 1836 inode); 1837 spin_unlock(&ci->i_ceph_lock); 1838 down_read(&mdsc->snap_rwsem); 1839 took_snap_rwsem = 1; 1840 goto retry; 1841 } 1842 took_snap_rwsem = 1; 1843 } 1844 1845 if (cap == ci->i_auth_cap && ci->i_dirty_caps) { 1846 flushing = __mark_caps_flushing(inode, session, false, 1847 &flush_tid, 1848 &oldest_flush_tid); 1849 } else { 1850 flushing = 0; 1851 flush_tid = 0; 1852 spin_lock(&mdsc->cap_dirty_lock); 1853 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 1854 spin_unlock(&mdsc->cap_dirty_lock); 1855 } 1856 1857 mds = cap->mds; /* remember mds, so we don't repeat */ 1858 sent++; 1859 1860 /* __send_cap drops i_ceph_lock */ 1861 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used, 1862 want, retain, flushing, 1863 flush_tid, oldest_flush_tid); 1864 goto retry; /* retake i_ceph_lock and restart our cap scan. */ 1865 } 1866 1867 /* 1868 * Reschedule delayed caps release if we delayed anything, 1869 * otherwise cancel. 1870 */ 1871 if (delayed && is_delayed) 1872 force_requeue = true; /* __send_cap delayed release; requeue */ 1873 if (!delayed && !is_delayed) 1874 __cap_delay_cancel(mdsc, ci); 1875 else if (!is_delayed || force_requeue) 1876 __cap_delay_requeue(mdsc, ci); 1877 1878 spin_unlock(&ci->i_ceph_lock); 1879 1880 if (queue_invalidate) 1881 ceph_queue_invalidate(inode); 1882 1883 if (session) 1884 mutex_unlock(&session->s_mutex); 1885 if (took_snap_rwsem) 1886 up_read(&mdsc->snap_rwsem); 1887 } 1888 1889 /* 1890 * Try to flush dirty caps back to the auth mds. 1891 */ 1892 static int try_flush_caps(struct inode *inode, u64 *ptid) 1893 { 1894 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 1895 struct ceph_inode_info *ci = ceph_inode(inode); 1896 struct ceph_mds_session *session = NULL; 1897 int flushing = 0; 1898 u64 flush_tid = 0, oldest_flush_tid = 0; 1899 1900 retry: 1901 spin_lock(&ci->i_ceph_lock); 1902 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { 1903 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode); 1904 goto out; 1905 } 1906 if (ci->i_dirty_caps && ci->i_auth_cap) { 1907 struct ceph_cap *cap = ci->i_auth_cap; 1908 int used = __ceph_caps_used(ci); 1909 int want = __ceph_caps_wanted(ci); 1910 int delayed; 1911 1912 if (!session || session != cap->session) { 1913 spin_unlock(&ci->i_ceph_lock); 1914 if (session) 1915 mutex_unlock(&session->s_mutex); 1916 session = cap->session; 1917 mutex_lock(&session->s_mutex); 1918 goto retry; 1919 } 1920 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) 1921 goto out; 1922 1923 flushing = __mark_caps_flushing(inode, session, true, 1924 &flush_tid, &oldest_flush_tid); 1925 1926 /* __send_cap drops i_ceph_lock */ 1927 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want, 1928 (cap->issued | cap->implemented), 1929 flushing, flush_tid, oldest_flush_tid); 1930 1931 if (delayed) { 1932 spin_lock(&ci->i_ceph_lock); 1933 __cap_delay_requeue(mdsc, ci); 1934 spin_unlock(&ci->i_ceph_lock); 1935 } 1936 } else { 1937 if (!list_empty(&ci->i_cap_flush_list)) { 1938 struct ceph_cap_flush *cf = 1939 list_last_entry(&ci->i_cap_flush_list, 1940 struct ceph_cap_flush, i_list); 1941 cf->wake = true; 1942 flush_tid = cf->tid; 1943 } 1944 flushing = ci->i_flushing_caps; 1945 spin_unlock(&ci->i_ceph_lock); 1946 } 1947 out: 1948 if (session) 1949 mutex_unlock(&session->s_mutex); 1950 1951 *ptid = flush_tid; 1952 return flushing; 1953 } 1954 1955 /* 1956 * Return true if we've flushed caps through the given flush_tid. 1957 */ 1958 static int caps_are_flushed(struct inode *inode, u64 flush_tid) 1959 { 1960 struct ceph_inode_info *ci = ceph_inode(inode); 1961 int ret = 1; 1962 1963 spin_lock(&ci->i_ceph_lock); 1964 if (!list_empty(&ci->i_cap_flush_list)) { 1965 struct ceph_cap_flush * cf = 1966 list_first_entry(&ci->i_cap_flush_list, 1967 struct ceph_cap_flush, i_list); 1968 if (cf->tid <= flush_tid) 1969 ret = 0; 1970 } 1971 spin_unlock(&ci->i_ceph_lock); 1972 return ret; 1973 } 1974 1975 /* 1976 * wait for any unsafe requests to complete. 1977 */ 1978 static int unsafe_request_wait(struct inode *inode) 1979 { 1980 struct ceph_inode_info *ci = ceph_inode(inode); 1981 struct ceph_mds_request *req1 = NULL, *req2 = NULL; 1982 int ret, err = 0; 1983 1984 spin_lock(&ci->i_unsafe_lock); 1985 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) { 1986 req1 = list_last_entry(&ci->i_unsafe_dirops, 1987 struct ceph_mds_request, 1988 r_unsafe_dir_item); 1989 ceph_mdsc_get_request(req1); 1990 } 1991 if (!list_empty(&ci->i_unsafe_iops)) { 1992 req2 = list_last_entry(&ci->i_unsafe_iops, 1993 struct ceph_mds_request, 1994 r_unsafe_target_item); 1995 ceph_mdsc_get_request(req2); 1996 } 1997 spin_unlock(&ci->i_unsafe_lock); 1998 1999 dout("unsafe_requeset_wait %p wait on tid %llu %llu\n", 2000 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL); 2001 if (req1) { 2002 ret = !wait_for_completion_timeout(&req1->r_safe_completion, 2003 ceph_timeout_jiffies(req1->r_timeout)); 2004 if (ret) 2005 err = -EIO; 2006 ceph_mdsc_put_request(req1); 2007 } 2008 if (req2) { 2009 ret = !wait_for_completion_timeout(&req2->r_safe_completion, 2010 ceph_timeout_jiffies(req2->r_timeout)); 2011 if (ret) 2012 err = -EIO; 2013 ceph_mdsc_put_request(req2); 2014 } 2015 return err; 2016 } 2017 2018 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync) 2019 { 2020 struct inode *inode = file->f_mapping->host; 2021 struct ceph_inode_info *ci = ceph_inode(inode); 2022 u64 flush_tid; 2023 int ret; 2024 int dirty; 2025 2026 dout("fsync %p%s\n", inode, datasync ? " datasync" : ""); 2027 2028 ceph_sync_write_wait(inode); 2029 2030 ret = filemap_write_and_wait_range(inode->i_mapping, start, end); 2031 if (ret < 0) 2032 goto out; 2033 2034 if (datasync) 2035 goto out; 2036 2037 inode_lock(inode); 2038 2039 dirty = try_flush_caps(inode, &flush_tid); 2040 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); 2041 2042 ret = unsafe_request_wait(inode); 2043 2044 /* 2045 * only wait on non-file metadata writeback (the mds 2046 * can recover size and mtime, so we don't need to 2047 * wait for that) 2048 */ 2049 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { 2050 ret = wait_event_interruptible(ci->i_cap_wq, 2051 caps_are_flushed(inode, flush_tid)); 2052 } 2053 inode_unlock(inode); 2054 out: 2055 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret); 2056 return ret; 2057 } 2058 2059 /* 2060 * Flush any dirty caps back to the mds. If we aren't asked to wait, 2061 * queue inode for flush but don't do so immediately, because we can 2062 * get by with fewer MDS messages if we wait for data writeback to 2063 * complete first. 2064 */ 2065 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc) 2066 { 2067 struct ceph_inode_info *ci = ceph_inode(inode); 2068 u64 flush_tid; 2069 int err = 0; 2070 int dirty; 2071 int wait = wbc->sync_mode == WB_SYNC_ALL; 2072 2073 dout("write_inode %p wait=%d\n", inode, wait); 2074 if (wait) { 2075 dirty = try_flush_caps(inode, &flush_tid); 2076 if (dirty) 2077 err = wait_event_interruptible(ci->i_cap_wq, 2078 caps_are_flushed(inode, flush_tid)); 2079 } else { 2080 struct ceph_mds_client *mdsc = 2081 ceph_sb_to_client(inode->i_sb)->mdsc; 2082 2083 spin_lock(&ci->i_ceph_lock); 2084 if (__ceph_caps_dirty(ci)) 2085 __cap_delay_requeue_front(mdsc, ci); 2086 spin_unlock(&ci->i_ceph_lock); 2087 } 2088 return err; 2089 } 2090 2091 static void __kick_flushing_caps(struct ceph_mds_client *mdsc, 2092 struct ceph_mds_session *session, 2093 struct ceph_inode_info *ci, 2094 u64 oldest_flush_tid) 2095 __releases(ci->i_ceph_lock) 2096 __acquires(ci->i_ceph_lock) 2097 { 2098 struct inode *inode = &ci->vfs_inode; 2099 struct ceph_cap *cap; 2100 struct ceph_cap_flush *cf; 2101 int ret; 2102 u64 first_tid = 0; 2103 2104 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) { 2105 if (cf->tid < first_tid) 2106 continue; 2107 2108 cap = ci->i_auth_cap; 2109 if (!(cap && cap->session == session)) { 2110 pr_err("%p auth cap %p not mds%d ???\n", 2111 inode, cap, session->s_mds); 2112 break; 2113 } 2114 2115 first_tid = cf->tid + 1; 2116 2117 if (cf->caps) { 2118 dout("kick_flushing_caps %p cap %p tid %llu %s\n", 2119 inode, cap, cf->tid, ceph_cap_string(cf->caps)); 2120 ci->i_ceph_flags |= CEPH_I_NODELAY; 2121 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, 2122 __ceph_caps_used(ci), 2123 __ceph_caps_wanted(ci), 2124 cap->issued | cap->implemented, 2125 cf->caps, cf->tid, oldest_flush_tid); 2126 if (ret) { 2127 pr_err("kick_flushing_caps: error sending " 2128 "cap flush, ino (%llx.%llx) " 2129 "tid %llu flushing %s\n", 2130 ceph_vinop(inode), cf->tid, 2131 ceph_cap_string(cf->caps)); 2132 } 2133 } else { 2134 struct ceph_cap_snap *capsnap = 2135 container_of(cf, struct ceph_cap_snap, 2136 cap_flush); 2137 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n", 2138 inode, capsnap, cf->tid, 2139 ceph_cap_string(capsnap->dirty)); 2140 2141 atomic_inc(&capsnap->nref); 2142 spin_unlock(&ci->i_ceph_lock); 2143 2144 ret = __send_flush_snap(inode, session, capsnap, cap->mseq, 2145 oldest_flush_tid); 2146 if (ret < 0) { 2147 pr_err("kick_flushing_caps: error sending " 2148 "cap flushsnap, ino (%llx.%llx) " 2149 "tid %llu follows %llu\n", 2150 ceph_vinop(inode), cf->tid, 2151 capsnap->follows); 2152 } 2153 2154 ceph_put_cap_snap(capsnap); 2155 } 2156 2157 spin_lock(&ci->i_ceph_lock); 2158 } 2159 } 2160 2161 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc, 2162 struct ceph_mds_session *session) 2163 { 2164 struct ceph_inode_info *ci; 2165 struct ceph_cap *cap; 2166 u64 oldest_flush_tid; 2167 2168 dout("early_kick_flushing_caps mds%d\n", session->s_mds); 2169 2170 spin_lock(&mdsc->cap_dirty_lock); 2171 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2172 spin_unlock(&mdsc->cap_dirty_lock); 2173 2174 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 2175 spin_lock(&ci->i_ceph_lock); 2176 cap = ci->i_auth_cap; 2177 if (!(cap && cap->session == session)) { 2178 pr_err("%p auth cap %p not mds%d ???\n", 2179 &ci->vfs_inode, cap, session->s_mds); 2180 spin_unlock(&ci->i_ceph_lock); 2181 continue; 2182 } 2183 2184 2185 /* 2186 * if flushing caps were revoked, we re-send the cap flush 2187 * in client reconnect stage. This guarantees MDS * processes 2188 * the cap flush message before issuing the flushing caps to 2189 * other client. 2190 */ 2191 if ((cap->issued & ci->i_flushing_caps) != 2192 ci->i_flushing_caps) { 2193 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 2194 __kick_flushing_caps(mdsc, session, ci, 2195 oldest_flush_tid); 2196 } else { 2197 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH; 2198 } 2199 2200 spin_unlock(&ci->i_ceph_lock); 2201 } 2202 } 2203 2204 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 2205 struct ceph_mds_session *session) 2206 { 2207 struct ceph_inode_info *ci; 2208 struct ceph_cap *cap; 2209 u64 oldest_flush_tid; 2210 2211 dout("kick_flushing_caps mds%d\n", session->s_mds); 2212 2213 spin_lock(&mdsc->cap_dirty_lock); 2214 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2215 spin_unlock(&mdsc->cap_dirty_lock); 2216 2217 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 2218 spin_lock(&ci->i_ceph_lock); 2219 cap = ci->i_auth_cap; 2220 if (!(cap && cap->session == session)) { 2221 pr_err("%p auth cap %p not mds%d ???\n", 2222 &ci->vfs_inode, cap, session->s_mds); 2223 spin_unlock(&ci->i_ceph_lock); 2224 continue; 2225 } 2226 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) { 2227 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 2228 __kick_flushing_caps(mdsc, session, ci, 2229 oldest_flush_tid); 2230 } 2231 spin_unlock(&ci->i_ceph_lock); 2232 } 2233 } 2234 2235 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc, 2236 struct ceph_mds_session *session, 2237 struct inode *inode) 2238 __releases(ci->i_ceph_lock) 2239 { 2240 struct ceph_inode_info *ci = ceph_inode(inode); 2241 struct ceph_cap *cap; 2242 2243 cap = ci->i_auth_cap; 2244 dout("kick_flushing_inode_caps %p flushing %s\n", inode, 2245 ceph_cap_string(ci->i_flushing_caps)); 2246 2247 if (!list_empty(&ci->i_cap_flush_list)) { 2248 u64 oldest_flush_tid; 2249 spin_lock(&mdsc->cap_dirty_lock); 2250 list_move_tail(&ci->i_flushing_item, 2251 &cap->session->s_cap_flushing); 2252 oldest_flush_tid = __get_oldest_flush_tid(mdsc); 2253 spin_unlock(&mdsc->cap_dirty_lock); 2254 2255 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; 2256 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid); 2257 spin_unlock(&ci->i_ceph_lock); 2258 } else { 2259 spin_unlock(&ci->i_ceph_lock); 2260 } 2261 } 2262 2263 2264 /* 2265 * Take references to capabilities we hold, so that we don't release 2266 * them to the MDS prematurely. 2267 * 2268 * Protected by i_ceph_lock. 2269 */ 2270 static void __take_cap_refs(struct ceph_inode_info *ci, int got, 2271 bool snap_rwsem_locked) 2272 { 2273 if (got & CEPH_CAP_PIN) 2274 ci->i_pin_ref++; 2275 if (got & CEPH_CAP_FILE_RD) 2276 ci->i_rd_ref++; 2277 if (got & CEPH_CAP_FILE_CACHE) 2278 ci->i_rdcache_ref++; 2279 if (got & CEPH_CAP_FILE_WR) { 2280 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) { 2281 BUG_ON(!snap_rwsem_locked); 2282 ci->i_head_snapc = ceph_get_snap_context( 2283 ci->i_snap_realm->cached_context); 2284 } 2285 ci->i_wr_ref++; 2286 } 2287 if (got & CEPH_CAP_FILE_BUFFER) { 2288 if (ci->i_wb_ref == 0) 2289 ihold(&ci->vfs_inode); 2290 ci->i_wb_ref++; 2291 dout("__take_cap_refs %p wb %d -> %d (?)\n", 2292 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref); 2293 } 2294 } 2295 2296 /* 2297 * Try to grab cap references. Specify those refs we @want, and the 2298 * minimal set we @need. Also include the larger offset we are writing 2299 * to (when applicable), and check against max_size here as well. 2300 * Note that caller is responsible for ensuring max_size increases are 2301 * requested from the MDS. 2302 */ 2303 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want, 2304 loff_t endoff, bool nonblock, int *got, int *err) 2305 { 2306 struct inode *inode = &ci->vfs_inode; 2307 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 2308 int ret = 0; 2309 int have, implemented; 2310 int file_wanted; 2311 bool snap_rwsem_locked = false; 2312 2313 dout("get_cap_refs %p need %s want %s\n", inode, 2314 ceph_cap_string(need), ceph_cap_string(want)); 2315 2316 again: 2317 spin_lock(&ci->i_ceph_lock); 2318 2319 /* make sure file is actually open */ 2320 file_wanted = __ceph_caps_file_wanted(ci); 2321 if ((file_wanted & need) != need) { 2322 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n", 2323 ceph_cap_string(need), ceph_cap_string(file_wanted)); 2324 *err = -EBADF; 2325 ret = 1; 2326 goto out_unlock; 2327 } 2328 2329 /* finish pending truncate */ 2330 while (ci->i_truncate_pending) { 2331 spin_unlock(&ci->i_ceph_lock); 2332 if (snap_rwsem_locked) { 2333 up_read(&mdsc->snap_rwsem); 2334 snap_rwsem_locked = false; 2335 } 2336 __ceph_do_pending_vmtruncate(inode); 2337 spin_lock(&ci->i_ceph_lock); 2338 } 2339 2340 have = __ceph_caps_issued(ci, &implemented); 2341 2342 if (have & need & CEPH_CAP_FILE_WR) { 2343 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) { 2344 dout("get_cap_refs %p endoff %llu > maxsize %llu\n", 2345 inode, endoff, ci->i_max_size); 2346 if (endoff > ci->i_requested_max_size) { 2347 *err = -EAGAIN; 2348 ret = 1; 2349 } 2350 goto out_unlock; 2351 } 2352 /* 2353 * If a sync write is in progress, we must wait, so that we 2354 * can get a final snapshot value for size+mtime. 2355 */ 2356 if (__ceph_have_pending_cap_snap(ci)) { 2357 dout("get_cap_refs %p cap_snap_pending\n", inode); 2358 goto out_unlock; 2359 } 2360 } 2361 2362 if ((have & need) == need) { 2363 /* 2364 * Look at (implemented & ~have & not) so that we keep waiting 2365 * on transition from wanted -> needed caps. This is needed 2366 * for WRBUFFER|WR -> WR to avoid a new WR sync write from 2367 * going before a prior buffered writeback happens. 2368 */ 2369 int not = want & ~(have & need); 2370 int revoking = implemented & ~have; 2371 dout("get_cap_refs %p have %s but not %s (revoking %s)\n", 2372 inode, ceph_cap_string(have), ceph_cap_string(not), 2373 ceph_cap_string(revoking)); 2374 if ((revoking & not) == 0) { 2375 if (!snap_rwsem_locked && 2376 !ci->i_head_snapc && 2377 (need & CEPH_CAP_FILE_WR)) { 2378 if (!down_read_trylock(&mdsc->snap_rwsem)) { 2379 /* 2380 * we can not call down_read() when 2381 * task isn't in TASK_RUNNING state 2382 */ 2383 if (nonblock) { 2384 *err = -EAGAIN; 2385 ret = 1; 2386 goto out_unlock; 2387 } 2388 2389 spin_unlock(&ci->i_ceph_lock); 2390 down_read(&mdsc->snap_rwsem); 2391 snap_rwsem_locked = true; 2392 goto again; 2393 } 2394 snap_rwsem_locked = true; 2395 } 2396 *got = need | (have & want); 2397 if ((need & CEPH_CAP_FILE_RD) && 2398 !(*got & CEPH_CAP_FILE_CACHE)) 2399 ceph_disable_fscache_readpage(ci); 2400 __take_cap_refs(ci, *got, true); 2401 ret = 1; 2402 } 2403 } else { 2404 int session_readonly = false; 2405 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) { 2406 struct ceph_mds_session *s = ci->i_auth_cap->session; 2407 spin_lock(&s->s_cap_lock); 2408 session_readonly = s->s_readonly; 2409 spin_unlock(&s->s_cap_lock); 2410 } 2411 if (session_readonly) { 2412 dout("get_cap_refs %p needed %s but mds%d readonly\n", 2413 inode, ceph_cap_string(need), ci->i_auth_cap->mds); 2414 *err = -EROFS; 2415 ret = 1; 2416 goto out_unlock; 2417 } 2418 2419 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) { 2420 int mds_wanted; 2421 if (ACCESS_ONCE(mdsc->fsc->mount_state) == 2422 CEPH_MOUNT_SHUTDOWN) { 2423 dout("get_cap_refs %p forced umount\n", inode); 2424 *err = -EIO; 2425 ret = 1; 2426 goto out_unlock; 2427 } 2428 mds_wanted = __ceph_caps_mds_wanted(ci); 2429 if ((mds_wanted & need) != need) { 2430 dout("get_cap_refs %p caps were dropped" 2431 " (session killed?)\n", inode); 2432 *err = -ESTALE; 2433 ret = 1; 2434 goto out_unlock; 2435 } 2436 if ((mds_wanted & file_wanted) == 2437 (file_wanted & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR))) 2438 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED; 2439 } 2440 2441 dout("get_cap_refs %p have %s needed %s\n", inode, 2442 ceph_cap_string(have), ceph_cap_string(need)); 2443 } 2444 out_unlock: 2445 spin_unlock(&ci->i_ceph_lock); 2446 if (snap_rwsem_locked) 2447 up_read(&mdsc->snap_rwsem); 2448 2449 dout("get_cap_refs %p ret %d got %s\n", inode, 2450 ret, ceph_cap_string(*got)); 2451 return ret; 2452 } 2453 2454 /* 2455 * Check the offset we are writing up to against our current 2456 * max_size. If necessary, tell the MDS we want to write to 2457 * a larger offset. 2458 */ 2459 static void check_max_size(struct inode *inode, loff_t endoff) 2460 { 2461 struct ceph_inode_info *ci = ceph_inode(inode); 2462 int check = 0; 2463 2464 /* do we need to explicitly request a larger max_size? */ 2465 spin_lock(&ci->i_ceph_lock); 2466 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) { 2467 dout("write %p at large endoff %llu, req max_size\n", 2468 inode, endoff); 2469 ci->i_wanted_max_size = endoff; 2470 } 2471 /* duplicate ceph_check_caps()'s logic */ 2472 if (ci->i_auth_cap && 2473 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) && 2474 ci->i_wanted_max_size > ci->i_max_size && 2475 ci->i_wanted_max_size > ci->i_requested_max_size) 2476 check = 1; 2477 spin_unlock(&ci->i_ceph_lock); 2478 if (check) 2479 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2480 } 2481 2482 /* 2483 * Wait for caps, and take cap references. If we can't get a WR cap 2484 * due to a small max_size, make sure we check_max_size (and possibly 2485 * ask the mds) so we don't get hung up indefinitely. 2486 */ 2487 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, 2488 loff_t endoff, int *got, struct page **pinned_page) 2489 { 2490 int _got, ret, err = 0; 2491 2492 ret = ceph_pool_perm_check(ci, need); 2493 if (ret < 0) 2494 return ret; 2495 2496 while (true) { 2497 if (endoff > 0) 2498 check_max_size(&ci->vfs_inode, endoff); 2499 2500 err = 0; 2501 _got = 0; 2502 ret = try_get_cap_refs(ci, need, want, endoff, 2503 false, &_got, &err); 2504 if (ret) { 2505 if (err == -EAGAIN) 2506 continue; 2507 if (err < 0) 2508 ret = err; 2509 } else { 2510 ret = wait_event_interruptible(ci->i_cap_wq, 2511 try_get_cap_refs(ci, need, want, endoff, 2512 true, &_got, &err)); 2513 if (err == -EAGAIN) 2514 continue; 2515 if (err < 0) 2516 ret = err; 2517 } 2518 if (ret < 0) { 2519 if (err == -ESTALE) { 2520 /* session was killed, try renew caps */ 2521 ret = ceph_renew_caps(&ci->vfs_inode); 2522 if (ret == 0) 2523 continue; 2524 } 2525 return ret; 2526 } 2527 2528 if (ci->i_inline_version != CEPH_INLINE_NONE && 2529 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) && 2530 i_size_read(&ci->vfs_inode) > 0) { 2531 struct page *page = 2532 find_get_page(ci->vfs_inode.i_mapping, 0); 2533 if (page) { 2534 if (PageUptodate(page)) { 2535 *pinned_page = page; 2536 break; 2537 } 2538 put_page(page); 2539 } 2540 /* 2541 * drop cap refs first because getattr while 2542 * holding * caps refs can cause deadlock. 2543 */ 2544 ceph_put_cap_refs(ci, _got); 2545 _got = 0; 2546 2547 /* 2548 * getattr request will bring inline data into 2549 * page cache 2550 */ 2551 ret = __ceph_do_getattr(&ci->vfs_inode, NULL, 2552 CEPH_STAT_CAP_INLINE_DATA, 2553 true); 2554 if (ret < 0) 2555 return ret; 2556 continue; 2557 } 2558 break; 2559 } 2560 2561 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE)) 2562 ceph_fscache_revalidate_cookie(ci); 2563 2564 *got = _got; 2565 return 0; 2566 } 2567 2568 /* 2569 * Take cap refs. Caller must already know we hold at least one ref 2570 * on the caps in question or we don't know this is safe. 2571 */ 2572 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) 2573 { 2574 spin_lock(&ci->i_ceph_lock); 2575 __take_cap_refs(ci, caps, false); 2576 spin_unlock(&ci->i_ceph_lock); 2577 } 2578 2579 2580 /* 2581 * drop cap_snap that is not associated with any snapshot. 2582 * we don't need to send FLUSHSNAP message for it. 2583 */ 2584 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci, 2585 struct ceph_cap_snap *capsnap) 2586 { 2587 if (!capsnap->need_flush && 2588 !capsnap->writing && !capsnap->dirty_pages) { 2589 dout("dropping cap_snap %p follows %llu\n", 2590 capsnap, capsnap->follows); 2591 BUG_ON(capsnap->cap_flush.tid > 0); 2592 ceph_put_snap_context(capsnap->context); 2593 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps)) 2594 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; 2595 2596 list_del(&capsnap->ci_item); 2597 ceph_put_cap_snap(capsnap); 2598 return 1; 2599 } 2600 return 0; 2601 } 2602 2603 /* 2604 * Release cap refs. 2605 * 2606 * If we released the last ref on any given cap, call ceph_check_caps 2607 * to release (or schedule a release). 2608 * 2609 * If we are releasing a WR cap (from a sync write), finalize any affected 2610 * cap_snap, and wake up any waiters. 2611 */ 2612 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had) 2613 { 2614 struct inode *inode = &ci->vfs_inode; 2615 int last = 0, put = 0, flushsnaps = 0, wake = 0; 2616 2617 spin_lock(&ci->i_ceph_lock); 2618 if (had & CEPH_CAP_PIN) 2619 --ci->i_pin_ref; 2620 if (had & CEPH_CAP_FILE_RD) 2621 if (--ci->i_rd_ref == 0) 2622 last++; 2623 if (had & CEPH_CAP_FILE_CACHE) 2624 if (--ci->i_rdcache_ref == 0) 2625 last++; 2626 if (had & CEPH_CAP_FILE_BUFFER) { 2627 if (--ci->i_wb_ref == 0) { 2628 last++; 2629 put++; 2630 } 2631 dout("put_cap_refs %p wb %d -> %d (?)\n", 2632 inode, ci->i_wb_ref+1, ci->i_wb_ref); 2633 } 2634 if (had & CEPH_CAP_FILE_WR) 2635 if (--ci->i_wr_ref == 0) { 2636 last++; 2637 if (__ceph_have_pending_cap_snap(ci)) { 2638 struct ceph_cap_snap *capsnap = 2639 list_last_entry(&ci->i_cap_snaps, 2640 struct ceph_cap_snap, 2641 ci_item); 2642 capsnap->writing = 0; 2643 if (ceph_try_drop_cap_snap(ci, capsnap)) 2644 put++; 2645 else if (__ceph_finish_cap_snap(ci, capsnap)) 2646 flushsnaps = 1; 2647 wake = 1; 2648 } 2649 if (ci->i_wrbuffer_ref_head == 0 && 2650 ci->i_dirty_caps == 0 && 2651 ci->i_flushing_caps == 0) { 2652 BUG_ON(!ci->i_head_snapc); 2653 ceph_put_snap_context(ci->i_head_snapc); 2654 ci->i_head_snapc = NULL; 2655 } 2656 /* see comment in __ceph_remove_cap() */ 2657 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) 2658 drop_inode_snap_realm(ci); 2659 } 2660 spin_unlock(&ci->i_ceph_lock); 2661 2662 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had), 2663 last ? " last" : "", put ? " put" : ""); 2664 2665 if (last && !flushsnaps) 2666 ceph_check_caps(ci, 0, NULL); 2667 else if (flushsnaps) 2668 ceph_flush_snaps(ci, NULL); 2669 if (wake) 2670 wake_up_all(&ci->i_cap_wq); 2671 while (put-- > 0) 2672 iput(inode); 2673 } 2674 2675 /* 2676 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap 2677 * context. Adjust per-snap dirty page accounting as appropriate. 2678 * Once all dirty data for a cap_snap is flushed, flush snapped file 2679 * metadata back to the MDS. If we dropped the last ref, call 2680 * ceph_check_caps. 2681 */ 2682 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr, 2683 struct ceph_snap_context *snapc) 2684 { 2685 struct inode *inode = &ci->vfs_inode; 2686 struct ceph_cap_snap *capsnap = NULL; 2687 int put = 0; 2688 bool last = false; 2689 bool found = false; 2690 bool flush_snaps = false; 2691 bool complete_capsnap = false; 2692 2693 spin_lock(&ci->i_ceph_lock); 2694 ci->i_wrbuffer_ref -= nr; 2695 if (ci->i_wrbuffer_ref == 0) { 2696 last = true; 2697 put++; 2698 } 2699 2700 if (ci->i_head_snapc == snapc) { 2701 ci->i_wrbuffer_ref_head -= nr; 2702 if (ci->i_wrbuffer_ref_head == 0 && 2703 ci->i_wr_ref == 0 && 2704 ci->i_dirty_caps == 0 && 2705 ci->i_flushing_caps == 0) { 2706 BUG_ON(!ci->i_head_snapc); 2707 ceph_put_snap_context(ci->i_head_snapc); 2708 ci->i_head_snapc = NULL; 2709 } 2710 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n", 2711 inode, 2712 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr, 2713 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 2714 last ? " LAST" : ""); 2715 } else { 2716 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 2717 if (capsnap->context == snapc) { 2718 found = true; 2719 break; 2720 } 2721 } 2722 BUG_ON(!found); 2723 capsnap->dirty_pages -= nr; 2724 if (capsnap->dirty_pages == 0) { 2725 complete_capsnap = true; 2726 if (!capsnap->writing) { 2727 if (ceph_try_drop_cap_snap(ci, capsnap)) { 2728 put++; 2729 } else { 2730 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; 2731 flush_snaps = true; 2732 } 2733 } 2734 } 2735 dout("put_wrbuffer_cap_refs on %p cap_snap %p " 2736 " snap %lld %d/%d -> %d/%d %s%s\n", 2737 inode, capsnap, capsnap->context->seq, 2738 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, 2739 ci->i_wrbuffer_ref, capsnap->dirty_pages, 2740 last ? " (wrbuffer last)" : "", 2741 complete_capsnap ? " (complete capsnap)" : ""); 2742 } 2743 2744 spin_unlock(&ci->i_ceph_lock); 2745 2746 if (last) { 2747 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); 2748 } else if (flush_snaps) { 2749 ceph_flush_snaps(ci, NULL); 2750 } 2751 if (complete_capsnap) 2752 wake_up_all(&ci->i_cap_wq); 2753 while (put-- > 0) 2754 iput(inode); 2755 } 2756 2757 /* 2758 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP. 2759 */ 2760 static void invalidate_aliases(struct inode *inode) 2761 { 2762 struct dentry *dn, *prev = NULL; 2763 2764 dout("invalidate_aliases inode %p\n", inode); 2765 d_prune_aliases(inode); 2766 /* 2767 * For non-directory inode, d_find_alias() only returns 2768 * hashed dentry. After calling d_invalidate(), the 2769 * dentry becomes unhashed. 2770 * 2771 * For directory inode, d_find_alias() can return 2772 * unhashed dentry. But directory inode should have 2773 * one alias at most. 2774 */ 2775 while ((dn = d_find_alias(inode))) { 2776 if (dn == prev) { 2777 dput(dn); 2778 break; 2779 } 2780 d_invalidate(dn); 2781 if (prev) 2782 dput(prev); 2783 prev = dn; 2784 } 2785 if (prev) 2786 dput(prev); 2787 } 2788 2789 /* 2790 * Handle a cap GRANT message from the MDS. (Note that a GRANT may 2791 * actually be a revocation if it specifies a smaller cap set.) 2792 * 2793 * caller holds s_mutex and i_ceph_lock, we drop both. 2794 */ 2795 static void handle_cap_grant(struct ceph_mds_client *mdsc, 2796 struct inode *inode, struct ceph_mds_caps *grant, 2797 struct ceph_string **pns, u64 inline_version, 2798 void *inline_data, u32 inline_len, 2799 struct ceph_buffer *xattr_buf, 2800 struct ceph_mds_session *session, 2801 struct ceph_cap *cap, int issued) 2802 __releases(ci->i_ceph_lock) 2803 __releases(mdsc->snap_rwsem) 2804 { 2805 struct ceph_inode_info *ci = ceph_inode(inode); 2806 int mds = session->s_mds; 2807 int seq = le32_to_cpu(grant->seq); 2808 int newcaps = le32_to_cpu(grant->caps); 2809 int used, wanted, dirty; 2810 u64 size = le64_to_cpu(grant->size); 2811 u64 max_size = le64_to_cpu(grant->max_size); 2812 struct timespec mtime, atime, ctime; 2813 int check_caps = 0; 2814 bool wake = false; 2815 bool writeback = false; 2816 bool queue_trunc = false; 2817 bool queue_invalidate = false; 2818 bool deleted_inode = false; 2819 bool fill_inline = false; 2820 2821 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n", 2822 inode, cap, mds, seq, ceph_cap_string(newcaps)); 2823 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size, 2824 inode->i_size); 2825 2826 2827 /* 2828 * auth mds of the inode changed. we received the cap export message, 2829 * but still haven't received the cap import message. handle_cap_export 2830 * updated the new auth MDS' cap. 2831 * 2832 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message 2833 * that was sent before the cap import message. So don't remove caps. 2834 */ 2835 if (ceph_seq_cmp(seq, cap->seq) <= 0) { 2836 WARN_ON(cap != ci->i_auth_cap); 2837 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id)); 2838 seq = cap->seq; 2839 newcaps |= cap->issued; 2840 } 2841 2842 /* 2843 * If CACHE is being revoked, and we have no dirty buffers, 2844 * try to invalidate (once). (If there are dirty buffers, we 2845 * will invalidate _after_ writeback.) 2846 */ 2847 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */ 2848 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && 2849 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && 2850 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) { 2851 if (try_nonblocking_invalidate(inode)) { 2852 /* there were locked pages.. invalidate later 2853 in a separate thread. */ 2854 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) { 2855 queue_invalidate = true; 2856 ci->i_rdcache_revoking = ci->i_rdcache_gen; 2857 } 2858 } 2859 } 2860 2861 /* side effects now are allowed */ 2862 cap->cap_gen = session->s_cap_gen; 2863 cap->seq = seq; 2864 2865 __check_cap_issue(ci, cap, newcaps); 2866 2867 if ((newcaps & CEPH_CAP_AUTH_SHARED) && 2868 (issued & CEPH_CAP_AUTH_EXCL) == 0) { 2869 inode->i_mode = le32_to_cpu(grant->mode); 2870 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid)); 2871 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid)); 2872 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode, 2873 from_kuid(&init_user_ns, inode->i_uid), 2874 from_kgid(&init_user_ns, inode->i_gid)); 2875 } 2876 2877 if ((newcaps & CEPH_CAP_AUTH_SHARED) && 2878 (issued & CEPH_CAP_LINK_EXCL) == 0) { 2879 set_nlink(inode, le32_to_cpu(grant->nlink)); 2880 if (inode->i_nlink == 0 && 2881 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL))) 2882 deleted_inode = true; 2883 } 2884 2885 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) { 2886 int len = le32_to_cpu(grant->xattr_len); 2887 u64 version = le64_to_cpu(grant->xattr_version); 2888 2889 if (version > ci->i_xattrs.version) { 2890 dout(" got new xattrs v%llu on %p len %d\n", 2891 version, inode, len); 2892 if (ci->i_xattrs.blob) 2893 ceph_buffer_put(ci->i_xattrs.blob); 2894 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf); 2895 ci->i_xattrs.version = version; 2896 ceph_forget_all_cached_acls(inode); 2897 } 2898 } 2899 2900 if (newcaps & CEPH_CAP_ANY_RD) { 2901 /* ctime/mtime/atime? */ 2902 ceph_decode_timespec(&mtime, &grant->mtime); 2903 ceph_decode_timespec(&atime, &grant->atime); 2904 ceph_decode_timespec(&ctime, &grant->ctime); 2905 ceph_fill_file_time(inode, issued, 2906 le32_to_cpu(grant->time_warp_seq), 2907 &ctime, &mtime, &atime); 2908 } 2909 2910 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) { 2911 /* file layout may have changed */ 2912 s64 old_pool = ci->i_layout.pool_id; 2913 struct ceph_string *old_ns; 2914 2915 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout); 2916 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns, 2917 lockdep_is_held(&ci->i_ceph_lock)); 2918 rcu_assign_pointer(ci->i_layout.pool_ns, *pns); 2919 2920 if (ci->i_layout.pool_id != old_pool || *pns != old_ns) 2921 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM; 2922 2923 *pns = old_ns; 2924 2925 /* size/truncate_seq? */ 2926 queue_trunc = ceph_fill_file_size(inode, issued, 2927 le32_to_cpu(grant->truncate_seq), 2928 le64_to_cpu(grant->truncate_size), 2929 size); 2930 /* max size increase? */ 2931 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) { 2932 dout("max_size %lld -> %llu\n", 2933 ci->i_max_size, max_size); 2934 ci->i_max_size = max_size; 2935 if (max_size >= ci->i_wanted_max_size) { 2936 ci->i_wanted_max_size = 0; /* reset */ 2937 ci->i_requested_max_size = 0; 2938 } 2939 wake = true; 2940 } 2941 } 2942 2943 /* check cap bits */ 2944 wanted = __ceph_caps_wanted(ci); 2945 used = __ceph_caps_used(ci); 2946 dirty = __ceph_caps_dirty(ci); 2947 dout(" my wanted = %s, used = %s, dirty %s\n", 2948 ceph_cap_string(wanted), 2949 ceph_cap_string(used), 2950 ceph_cap_string(dirty)); 2951 if (wanted != le32_to_cpu(grant->wanted)) { 2952 dout("mds wanted %s -> %s\n", 2953 ceph_cap_string(le32_to_cpu(grant->wanted)), 2954 ceph_cap_string(wanted)); 2955 /* imported cap may not have correct mds_wanted */ 2956 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) 2957 check_caps = 1; 2958 } 2959 2960 /* revocation, grant, or no-op? */ 2961 if (cap->issued & ~newcaps) { 2962 int revoking = cap->issued & ~newcaps; 2963 2964 dout("revocation: %s -> %s (revoking %s)\n", 2965 ceph_cap_string(cap->issued), 2966 ceph_cap_string(newcaps), 2967 ceph_cap_string(revoking)); 2968 if (revoking & used & CEPH_CAP_FILE_BUFFER) 2969 writeback = true; /* initiate writeback; will delay ack */ 2970 else if (revoking == CEPH_CAP_FILE_CACHE && 2971 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && 2972 queue_invalidate) 2973 ; /* do nothing yet, invalidation will be queued */ 2974 else if (cap == ci->i_auth_cap) 2975 check_caps = 1; /* check auth cap only */ 2976 else 2977 check_caps = 2; /* check all caps */ 2978 cap->issued = newcaps; 2979 cap->implemented |= newcaps; 2980 } else if (cap->issued == newcaps) { 2981 dout("caps unchanged: %s -> %s\n", 2982 ceph_cap_string(cap->issued), ceph_cap_string(newcaps)); 2983 } else { 2984 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued), 2985 ceph_cap_string(newcaps)); 2986 /* non-auth MDS is revoking the newly grant caps ? */ 2987 if (cap == ci->i_auth_cap && 2988 __ceph_caps_revoking_other(ci, cap, newcaps)) 2989 check_caps = 2; 2990 2991 cap->issued = newcaps; 2992 cap->implemented |= newcaps; /* add bits only, to 2993 * avoid stepping on a 2994 * pending revocation */ 2995 wake = true; 2996 } 2997 BUG_ON(cap->issued & ~cap->implemented); 2998 2999 if (inline_version > 0 && inline_version >= ci->i_inline_version) { 3000 ci->i_inline_version = inline_version; 3001 if (ci->i_inline_version != CEPH_INLINE_NONE && 3002 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO))) 3003 fill_inline = true; 3004 } 3005 3006 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) { 3007 if (newcaps & ~issued) 3008 wake = true; 3009 kick_flushing_inode_caps(mdsc, session, inode); 3010 up_read(&mdsc->snap_rwsem); 3011 } else { 3012 spin_unlock(&ci->i_ceph_lock); 3013 } 3014 3015 if (fill_inline) 3016 ceph_fill_inline_data(inode, NULL, inline_data, inline_len); 3017 3018 if (queue_trunc) 3019 ceph_queue_vmtruncate(inode); 3020 3021 if (writeback) 3022 /* 3023 * queue inode for writeback: we can't actually call 3024 * filemap_write_and_wait, etc. from message handler 3025 * context. 3026 */ 3027 ceph_queue_writeback(inode); 3028 if (queue_invalidate) 3029 ceph_queue_invalidate(inode); 3030 if (deleted_inode) 3031 invalidate_aliases(inode); 3032 if (wake) 3033 wake_up_all(&ci->i_cap_wq); 3034 3035 if (check_caps == 1) 3036 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY, 3037 session); 3038 else if (check_caps == 2) 3039 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session); 3040 else 3041 mutex_unlock(&session->s_mutex); 3042 } 3043 3044 /* 3045 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the 3046 * MDS has been safely committed. 3047 */ 3048 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid, 3049 struct ceph_mds_caps *m, 3050 struct ceph_mds_session *session, 3051 struct ceph_cap *cap) 3052 __releases(ci->i_ceph_lock) 3053 { 3054 struct ceph_inode_info *ci = ceph_inode(inode); 3055 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 3056 struct ceph_cap_flush *cf, *tmp_cf; 3057 LIST_HEAD(to_remove); 3058 unsigned seq = le32_to_cpu(m->seq); 3059 int dirty = le32_to_cpu(m->dirty); 3060 int cleaned = 0; 3061 bool drop = false; 3062 bool wake_ci = 0; 3063 bool wake_mdsc = 0; 3064 3065 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) { 3066 if (cf->tid == flush_tid) 3067 cleaned = cf->caps; 3068 if (cf->caps == 0) /* capsnap */ 3069 continue; 3070 if (cf->tid <= flush_tid) { 3071 if (__finish_cap_flush(NULL, ci, cf)) 3072 wake_ci = true; 3073 list_add_tail(&cf->i_list, &to_remove); 3074 } else { 3075 cleaned &= ~cf->caps; 3076 if (!cleaned) 3077 break; 3078 } 3079 } 3080 3081 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," 3082 " flushing %s -> %s\n", 3083 inode, session->s_mds, seq, ceph_cap_string(dirty), 3084 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), 3085 ceph_cap_string(ci->i_flushing_caps & ~cleaned)); 3086 3087 if (list_empty(&to_remove) && !cleaned) 3088 goto out; 3089 3090 ci->i_flushing_caps &= ~cleaned; 3091 3092 spin_lock(&mdsc->cap_dirty_lock); 3093 3094 list_for_each_entry(cf, &to_remove, i_list) { 3095 if (__finish_cap_flush(mdsc, NULL, cf)) 3096 wake_mdsc = true; 3097 } 3098 3099 if (ci->i_flushing_caps == 0) { 3100 if (list_empty(&ci->i_cap_flush_list)) { 3101 list_del_init(&ci->i_flushing_item); 3102 if (!list_empty(&session->s_cap_flushing)) { 3103 dout(" mds%d still flushing cap on %p\n", 3104 session->s_mds, 3105 &list_first_entry(&session->s_cap_flushing, 3106 struct ceph_inode_info, 3107 i_flushing_item)->vfs_inode); 3108 } 3109 } 3110 mdsc->num_cap_flushing--; 3111 dout(" inode %p now !flushing\n", inode); 3112 3113 if (ci->i_dirty_caps == 0) { 3114 dout(" inode %p now clean\n", inode); 3115 BUG_ON(!list_empty(&ci->i_dirty_item)); 3116 drop = true; 3117 if (ci->i_wr_ref == 0 && 3118 ci->i_wrbuffer_ref_head == 0) { 3119 BUG_ON(!ci->i_head_snapc); 3120 ceph_put_snap_context(ci->i_head_snapc); 3121 ci->i_head_snapc = NULL; 3122 } 3123 } else { 3124 BUG_ON(list_empty(&ci->i_dirty_item)); 3125 } 3126 } 3127 spin_unlock(&mdsc->cap_dirty_lock); 3128 3129 out: 3130 spin_unlock(&ci->i_ceph_lock); 3131 3132 while (!list_empty(&to_remove)) { 3133 cf = list_first_entry(&to_remove, 3134 struct ceph_cap_flush, i_list); 3135 list_del(&cf->i_list); 3136 ceph_free_cap_flush(cf); 3137 } 3138 3139 if (wake_ci) 3140 wake_up_all(&ci->i_cap_wq); 3141 if (wake_mdsc) 3142 wake_up_all(&mdsc->cap_flushing_wq); 3143 if (drop) 3144 iput(inode); 3145 } 3146 3147 /* 3148 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can 3149 * throw away our cap_snap. 3150 * 3151 * Caller hold s_mutex. 3152 */ 3153 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid, 3154 struct ceph_mds_caps *m, 3155 struct ceph_mds_session *session) 3156 { 3157 struct ceph_inode_info *ci = ceph_inode(inode); 3158 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 3159 u64 follows = le64_to_cpu(m->snap_follows); 3160 struct ceph_cap_snap *capsnap; 3161 bool flushed = false; 3162 bool wake_ci = false; 3163 bool wake_mdsc = false; 3164 3165 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n", 3166 inode, ci, session->s_mds, follows); 3167 3168 spin_lock(&ci->i_ceph_lock); 3169 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 3170 if (capsnap->follows == follows) { 3171 if (capsnap->cap_flush.tid != flush_tid) { 3172 dout(" cap_snap %p follows %lld tid %lld !=" 3173 " %lld\n", capsnap, follows, 3174 flush_tid, capsnap->cap_flush.tid); 3175 break; 3176 } 3177 flushed = true; 3178 break; 3179 } else { 3180 dout(" skipping cap_snap %p follows %lld\n", 3181 capsnap, capsnap->follows); 3182 } 3183 } 3184 if (flushed) { 3185 WARN_ON(capsnap->dirty_pages || capsnap->writing); 3186 dout(" removing %p cap_snap %p follows %lld\n", 3187 inode, capsnap, follows); 3188 list_del(&capsnap->ci_item); 3189 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush)) 3190 wake_ci = true; 3191 3192 spin_lock(&mdsc->cap_dirty_lock); 3193 3194 if (list_empty(&ci->i_cap_flush_list)) 3195 list_del_init(&ci->i_flushing_item); 3196 3197 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush)) 3198 wake_mdsc = true; 3199 3200 spin_unlock(&mdsc->cap_dirty_lock); 3201 } 3202 spin_unlock(&ci->i_ceph_lock); 3203 if (flushed) { 3204 ceph_put_snap_context(capsnap->context); 3205 ceph_put_cap_snap(capsnap); 3206 if (wake_ci) 3207 wake_up_all(&ci->i_cap_wq); 3208 if (wake_mdsc) 3209 wake_up_all(&mdsc->cap_flushing_wq); 3210 iput(inode); 3211 } 3212 } 3213 3214 /* 3215 * Handle TRUNC from MDS, indicating file truncation. 3216 * 3217 * caller hold s_mutex. 3218 */ 3219 static void handle_cap_trunc(struct inode *inode, 3220 struct ceph_mds_caps *trunc, 3221 struct ceph_mds_session *session) 3222 __releases(ci->i_ceph_lock) 3223 { 3224 struct ceph_inode_info *ci = ceph_inode(inode); 3225 int mds = session->s_mds; 3226 int seq = le32_to_cpu(trunc->seq); 3227 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq); 3228 u64 truncate_size = le64_to_cpu(trunc->truncate_size); 3229 u64 size = le64_to_cpu(trunc->size); 3230 int implemented = 0; 3231 int dirty = __ceph_caps_dirty(ci); 3232 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented); 3233 int queue_trunc = 0; 3234 3235 issued |= implemented | dirty; 3236 3237 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n", 3238 inode, mds, seq, truncate_size, truncate_seq); 3239 queue_trunc = ceph_fill_file_size(inode, issued, 3240 truncate_seq, truncate_size, size); 3241 spin_unlock(&ci->i_ceph_lock); 3242 3243 if (queue_trunc) 3244 ceph_queue_vmtruncate(inode); 3245 } 3246 3247 /* 3248 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a 3249 * different one. If we are the most recent migration we've seen (as 3250 * indicated by mseq), make note of the migrating cap bits for the 3251 * duration (until we see the corresponding IMPORT). 3252 * 3253 * caller holds s_mutex 3254 */ 3255 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex, 3256 struct ceph_mds_cap_peer *ph, 3257 struct ceph_mds_session *session) 3258 { 3259 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 3260 struct ceph_mds_session *tsession = NULL; 3261 struct ceph_cap *cap, *tcap, *new_cap = NULL; 3262 struct ceph_inode_info *ci = ceph_inode(inode); 3263 u64 t_cap_id; 3264 unsigned mseq = le32_to_cpu(ex->migrate_seq); 3265 unsigned t_seq, t_mseq; 3266 int target, issued; 3267 int mds = session->s_mds; 3268 3269 if (ph) { 3270 t_cap_id = le64_to_cpu(ph->cap_id); 3271 t_seq = le32_to_cpu(ph->seq); 3272 t_mseq = le32_to_cpu(ph->mseq); 3273 target = le32_to_cpu(ph->mds); 3274 } else { 3275 t_cap_id = t_seq = t_mseq = 0; 3276 target = -1; 3277 } 3278 3279 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n", 3280 inode, ci, mds, mseq, target); 3281 retry: 3282 spin_lock(&ci->i_ceph_lock); 3283 cap = __get_cap_for_mds(ci, mds); 3284 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id)) 3285 goto out_unlock; 3286 3287 if (target < 0) { 3288 __ceph_remove_cap(cap, false); 3289 if (!ci->i_auth_cap) 3290 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED; 3291 goto out_unlock; 3292 } 3293 3294 /* 3295 * now we know we haven't received the cap import message yet 3296 * because the exported cap still exist. 3297 */ 3298 3299 issued = cap->issued; 3300 WARN_ON(issued != cap->implemented); 3301 3302 tcap = __get_cap_for_mds(ci, target); 3303 if (tcap) { 3304 /* already have caps from the target */ 3305 if (tcap->cap_id != t_cap_id || 3306 ceph_seq_cmp(tcap->seq, t_seq) < 0) { 3307 dout(" updating import cap %p mds%d\n", tcap, target); 3308 tcap->cap_id = t_cap_id; 3309 tcap->seq = t_seq - 1; 3310 tcap->issue_seq = t_seq - 1; 3311 tcap->mseq = t_mseq; 3312 tcap->issued |= issued; 3313 tcap->implemented |= issued; 3314 if (cap == ci->i_auth_cap) 3315 ci->i_auth_cap = tcap; 3316 if (!list_empty(&ci->i_cap_flush_list) && 3317 ci->i_auth_cap == tcap) { 3318 spin_lock(&mdsc->cap_dirty_lock); 3319 list_move_tail(&ci->i_flushing_item, 3320 &tcap->session->s_cap_flushing); 3321 spin_unlock(&mdsc->cap_dirty_lock); 3322 } 3323 } 3324 __ceph_remove_cap(cap, false); 3325 goto out_unlock; 3326 } else if (tsession) { 3327 /* add placeholder for the export tagert */ 3328 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0; 3329 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0, 3330 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap); 3331 3332 __ceph_remove_cap(cap, false); 3333 goto out_unlock; 3334 } 3335 3336 spin_unlock(&ci->i_ceph_lock); 3337 mutex_unlock(&session->s_mutex); 3338 3339 /* open target session */ 3340 tsession = ceph_mdsc_open_export_target_session(mdsc, target); 3341 if (!IS_ERR(tsession)) { 3342 if (mds > target) { 3343 mutex_lock(&session->s_mutex); 3344 mutex_lock_nested(&tsession->s_mutex, 3345 SINGLE_DEPTH_NESTING); 3346 } else { 3347 mutex_lock(&tsession->s_mutex); 3348 mutex_lock_nested(&session->s_mutex, 3349 SINGLE_DEPTH_NESTING); 3350 } 3351 new_cap = ceph_get_cap(mdsc, NULL); 3352 } else { 3353 WARN_ON(1); 3354 tsession = NULL; 3355 target = -1; 3356 } 3357 goto retry; 3358 3359 out_unlock: 3360 spin_unlock(&ci->i_ceph_lock); 3361 mutex_unlock(&session->s_mutex); 3362 if (tsession) { 3363 mutex_unlock(&tsession->s_mutex); 3364 ceph_put_mds_session(tsession); 3365 } 3366 if (new_cap) 3367 ceph_put_cap(mdsc, new_cap); 3368 } 3369 3370 /* 3371 * Handle cap IMPORT. 3372 * 3373 * caller holds s_mutex. acquires i_ceph_lock 3374 */ 3375 static void handle_cap_import(struct ceph_mds_client *mdsc, 3376 struct inode *inode, struct ceph_mds_caps *im, 3377 struct ceph_mds_cap_peer *ph, 3378 struct ceph_mds_session *session, 3379 struct ceph_cap **target_cap, int *old_issued) 3380 __acquires(ci->i_ceph_lock) 3381 { 3382 struct ceph_inode_info *ci = ceph_inode(inode); 3383 struct ceph_cap *cap, *ocap, *new_cap = NULL; 3384 int mds = session->s_mds; 3385 int issued; 3386 unsigned caps = le32_to_cpu(im->caps); 3387 unsigned wanted = le32_to_cpu(im->wanted); 3388 unsigned seq = le32_to_cpu(im->seq); 3389 unsigned mseq = le32_to_cpu(im->migrate_seq); 3390 u64 realmino = le64_to_cpu(im->realm); 3391 u64 cap_id = le64_to_cpu(im->cap_id); 3392 u64 p_cap_id; 3393 int peer; 3394 3395 if (ph) { 3396 p_cap_id = le64_to_cpu(ph->cap_id); 3397 peer = le32_to_cpu(ph->mds); 3398 } else { 3399 p_cap_id = 0; 3400 peer = -1; 3401 } 3402 3403 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n", 3404 inode, ci, mds, mseq, peer); 3405 3406 retry: 3407 spin_lock(&ci->i_ceph_lock); 3408 cap = __get_cap_for_mds(ci, mds); 3409 if (!cap) { 3410 if (!new_cap) { 3411 spin_unlock(&ci->i_ceph_lock); 3412 new_cap = ceph_get_cap(mdsc, NULL); 3413 goto retry; 3414 } 3415 cap = new_cap; 3416 } else { 3417 if (new_cap) { 3418 ceph_put_cap(mdsc, new_cap); 3419 new_cap = NULL; 3420 } 3421 } 3422 3423 __ceph_caps_issued(ci, &issued); 3424 issued |= __ceph_caps_dirty(ci); 3425 3426 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq, 3427 realmino, CEPH_CAP_FLAG_AUTH, &new_cap); 3428 3429 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL; 3430 if (ocap && ocap->cap_id == p_cap_id) { 3431 dout(" remove export cap %p mds%d flags %d\n", 3432 ocap, peer, ph->flags); 3433 if ((ph->flags & CEPH_CAP_FLAG_AUTH) && 3434 (ocap->seq != le32_to_cpu(ph->seq) || 3435 ocap->mseq != le32_to_cpu(ph->mseq))) { 3436 pr_err("handle_cap_import: mismatched seq/mseq: " 3437 "ino (%llx.%llx) mds%d seq %d mseq %d " 3438 "importer mds%d has peer seq %d mseq %d\n", 3439 ceph_vinop(inode), peer, ocap->seq, 3440 ocap->mseq, mds, le32_to_cpu(ph->seq), 3441 le32_to_cpu(ph->mseq)); 3442 } 3443 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE)); 3444 } 3445 3446 /* make sure we re-request max_size, if necessary */ 3447 ci->i_wanted_max_size = 0; 3448 ci->i_requested_max_size = 0; 3449 3450 *old_issued = issued; 3451 *target_cap = cap; 3452 } 3453 3454 /* 3455 * Handle a caps message from the MDS. 3456 * 3457 * Identify the appropriate session, inode, and call the right handler 3458 * based on the cap op. 3459 */ 3460 void ceph_handle_caps(struct ceph_mds_session *session, 3461 struct ceph_msg *msg) 3462 { 3463 struct ceph_mds_client *mdsc = session->s_mdsc; 3464 struct super_block *sb = mdsc->fsc->sb; 3465 struct inode *inode; 3466 struct ceph_inode_info *ci; 3467 struct ceph_cap *cap; 3468 struct ceph_mds_caps *h; 3469 struct ceph_mds_cap_peer *peer = NULL; 3470 struct ceph_snap_realm *realm = NULL; 3471 struct ceph_string *pool_ns = NULL; 3472 int mds = session->s_mds; 3473 int op, issued; 3474 u32 seq, mseq; 3475 struct ceph_vino vino; 3476 u64 tid; 3477 u64 inline_version = 0; 3478 void *inline_data = NULL; 3479 u32 inline_len = 0; 3480 void *snaptrace; 3481 size_t snaptrace_len; 3482 void *p, *end; 3483 3484 dout("handle_caps from mds%d\n", mds); 3485 3486 /* decode */ 3487 end = msg->front.iov_base + msg->front.iov_len; 3488 tid = le64_to_cpu(msg->hdr.tid); 3489 if (msg->front.iov_len < sizeof(*h)) 3490 goto bad; 3491 h = msg->front.iov_base; 3492 op = le32_to_cpu(h->op); 3493 vino.ino = le64_to_cpu(h->ino); 3494 vino.snap = CEPH_NOSNAP; 3495 seq = le32_to_cpu(h->seq); 3496 mseq = le32_to_cpu(h->migrate_seq); 3497 3498 snaptrace = h + 1; 3499 snaptrace_len = le32_to_cpu(h->snap_trace_len); 3500 p = snaptrace + snaptrace_len; 3501 3502 if (le16_to_cpu(msg->hdr.version) >= 2) { 3503 u32 flock_len; 3504 ceph_decode_32_safe(&p, end, flock_len, bad); 3505 if (p + flock_len > end) 3506 goto bad; 3507 p += flock_len; 3508 } 3509 3510 if (le16_to_cpu(msg->hdr.version) >= 3) { 3511 if (op == CEPH_CAP_OP_IMPORT) { 3512 if (p + sizeof(*peer) > end) 3513 goto bad; 3514 peer = p; 3515 p += sizeof(*peer); 3516 } else if (op == CEPH_CAP_OP_EXPORT) { 3517 /* recorded in unused fields */ 3518 peer = (void *)&h->size; 3519 } 3520 } 3521 3522 if (le16_to_cpu(msg->hdr.version) >= 4) { 3523 ceph_decode_64_safe(&p, end, inline_version, bad); 3524 ceph_decode_32_safe(&p, end, inline_len, bad); 3525 if (p + inline_len > end) 3526 goto bad; 3527 inline_data = p; 3528 p += inline_len; 3529 } 3530 3531 if (le16_to_cpu(msg->hdr.version) >= 8) { 3532 u64 flush_tid; 3533 u32 caller_uid, caller_gid; 3534 u32 osd_epoch_barrier; 3535 u32 pool_ns_len; 3536 /* version >= 5 */ 3537 ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad); 3538 /* version >= 6 */ 3539 ceph_decode_64_safe(&p, end, flush_tid, bad); 3540 /* version >= 7 */ 3541 ceph_decode_32_safe(&p, end, caller_uid, bad); 3542 ceph_decode_32_safe(&p, end, caller_gid, bad); 3543 /* version >= 8 */ 3544 ceph_decode_32_safe(&p, end, pool_ns_len, bad); 3545 if (pool_ns_len > 0) { 3546 ceph_decode_need(&p, end, pool_ns_len, bad); 3547 pool_ns = ceph_find_or_create_string(p, pool_ns_len); 3548 p += pool_ns_len; 3549 } 3550 } 3551 3552 /* lookup ino */ 3553 inode = ceph_find_inode(sb, vino); 3554 ci = ceph_inode(inode); 3555 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino, 3556 vino.snap, inode); 3557 3558 mutex_lock(&session->s_mutex); 3559 session->s_seq++; 3560 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, 3561 (unsigned)seq); 3562 3563 if (!inode) { 3564 dout(" i don't have ino %llx\n", vino.ino); 3565 3566 if (op == CEPH_CAP_OP_IMPORT) { 3567 cap = ceph_get_cap(mdsc, NULL); 3568 cap->cap_ino = vino.ino; 3569 cap->queue_release = 1; 3570 cap->cap_id = le64_to_cpu(h->cap_id); 3571 cap->mseq = mseq; 3572 cap->seq = seq; 3573 spin_lock(&session->s_cap_lock); 3574 list_add_tail(&cap->session_caps, 3575 &session->s_cap_releases); 3576 session->s_num_cap_releases++; 3577 spin_unlock(&session->s_cap_lock); 3578 } 3579 goto flush_cap_releases; 3580 } 3581 3582 /* these will work even if we don't have a cap yet */ 3583 switch (op) { 3584 case CEPH_CAP_OP_FLUSHSNAP_ACK: 3585 handle_cap_flushsnap_ack(inode, tid, h, session); 3586 goto done; 3587 3588 case CEPH_CAP_OP_EXPORT: 3589 handle_cap_export(inode, h, peer, session); 3590 goto done_unlocked; 3591 3592 case CEPH_CAP_OP_IMPORT: 3593 realm = NULL; 3594 if (snaptrace_len) { 3595 down_write(&mdsc->snap_rwsem); 3596 ceph_update_snap_trace(mdsc, snaptrace, 3597 snaptrace + snaptrace_len, 3598 false, &realm); 3599 downgrade_write(&mdsc->snap_rwsem); 3600 } else { 3601 down_read(&mdsc->snap_rwsem); 3602 } 3603 handle_cap_import(mdsc, inode, h, peer, session, 3604 &cap, &issued); 3605 handle_cap_grant(mdsc, inode, h, &pool_ns, 3606 inline_version, inline_data, inline_len, 3607 msg->middle, session, cap, issued); 3608 if (realm) 3609 ceph_put_snap_realm(mdsc, realm); 3610 goto done_unlocked; 3611 } 3612 3613 /* the rest require a cap */ 3614 spin_lock(&ci->i_ceph_lock); 3615 cap = __get_cap_for_mds(ceph_inode(inode), mds); 3616 if (!cap) { 3617 dout(" no cap on %p ino %llx.%llx from mds%d\n", 3618 inode, ceph_ino(inode), ceph_snap(inode), mds); 3619 spin_unlock(&ci->i_ceph_lock); 3620 goto flush_cap_releases; 3621 } 3622 3623 /* note that each of these drops i_ceph_lock for us */ 3624 switch (op) { 3625 case CEPH_CAP_OP_REVOKE: 3626 case CEPH_CAP_OP_GRANT: 3627 __ceph_caps_issued(ci, &issued); 3628 issued |= __ceph_caps_dirty(ci); 3629 handle_cap_grant(mdsc, inode, h, &pool_ns, 3630 inline_version, inline_data, inline_len, 3631 msg->middle, session, cap, issued); 3632 goto done_unlocked; 3633 3634 case CEPH_CAP_OP_FLUSH_ACK: 3635 handle_cap_flush_ack(inode, tid, h, session, cap); 3636 break; 3637 3638 case CEPH_CAP_OP_TRUNC: 3639 handle_cap_trunc(inode, h, session); 3640 break; 3641 3642 default: 3643 spin_unlock(&ci->i_ceph_lock); 3644 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op, 3645 ceph_cap_op_name(op)); 3646 } 3647 3648 goto done; 3649 3650 flush_cap_releases: 3651 /* 3652 * send any cap release message to try to move things 3653 * along for the mds (who clearly thinks we still have this 3654 * cap). 3655 */ 3656 ceph_send_cap_releases(mdsc, session); 3657 3658 done: 3659 mutex_unlock(&session->s_mutex); 3660 done_unlocked: 3661 iput(inode); 3662 ceph_put_string(pool_ns); 3663 return; 3664 3665 bad: 3666 pr_err("ceph_handle_caps: corrupt message\n"); 3667 ceph_msg_dump(msg); 3668 return; 3669 } 3670 3671 /* 3672 * Delayed work handler to process end of delayed cap release LRU list. 3673 */ 3674 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc) 3675 { 3676 struct ceph_inode_info *ci; 3677 int flags = CHECK_CAPS_NODELAY; 3678 3679 dout("check_delayed_caps\n"); 3680 while (1) { 3681 spin_lock(&mdsc->cap_delay_lock); 3682 if (list_empty(&mdsc->cap_delay_list)) 3683 break; 3684 ci = list_first_entry(&mdsc->cap_delay_list, 3685 struct ceph_inode_info, 3686 i_cap_delay_list); 3687 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 && 3688 time_before(jiffies, ci->i_hold_caps_max)) 3689 break; 3690 list_del_init(&ci->i_cap_delay_list); 3691 spin_unlock(&mdsc->cap_delay_lock); 3692 dout("check_delayed_caps on %p\n", &ci->vfs_inode); 3693 ceph_check_caps(ci, flags, NULL); 3694 } 3695 spin_unlock(&mdsc->cap_delay_lock); 3696 } 3697 3698 /* 3699 * Flush all dirty caps to the mds 3700 */ 3701 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc) 3702 { 3703 struct ceph_inode_info *ci; 3704 struct inode *inode; 3705 3706 dout("flush_dirty_caps\n"); 3707 spin_lock(&mdsc->cap_dirty_lock); 3708 while (!list_empty(&mdsc->cap_dirty)) { 3709 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info, 3710 i_dirty_item); 3711 inode = &ci->vfs_inode; 3712 ihold(inode); 3713 dout("flush_dirty_caps %p\n", inode); 3714 spin_unlock(&mdsc->cap_dirty_lock); 3715 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL); 3716 iput(inode); 3717 spin_lock(&mdsc->cap_dirty_lock); 3718 } 3719 spin_unlock(&mdsc->cap_dirty_lock); 3720 dout("flush_dirty_caps done\n"); 3721 } 3722 3723 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode) 3724 { 3725 int i; 3726 int bits = (fmode << 1) | 1; 3727 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { 3728 if (bits & (1 << i)) 3729 ci->i_nr_by_mode[i]++; 3730 } 3731 } 3732 3733 /* 3734 * Drop open file reference. If we were the last open file, 3735 * we may need to release capabilities to the MDS (or schedule 3736 * their delayed release). 3737 */ 3738 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode) 3739 { 3740 int i, last = 0; 3741 int bits = (fmode << 1) | 1; 3742 spin_lock(&ci->i_ceph_lock); 3743 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { 3744 if (bits & (1 << i)) { 3745 BUG_ON(ci->i_nr_by_mode[i] == 0); 3746 if (--ci->i_nr_by_mode[i] == 0) 3747 last++; 3748 } 3749 } 3750 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n", 3751 &ci->vfs_inode, fmode, 3752 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1], 3753 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]); 3754 spin_unlock(&ci->i_ceph_lock); 3755 3756 if (last && ci->i_vino.snap == CEPH_NOSNAP) 3757 ceph_check_caps(ci, 0, NULL); 3758 } 3759 3760 /* 3761 * Helpers for embedding cap and dentry lease releases into mds 3762 * requests. 3763 * 3764 * @force is used by dentry_release (below) to force inclusion of a 3765 * record for the directory inode, even when there aren't any caps to 3766 * drop. 3767 */ 3768 int ceph_encode_inode_release(void **p, struct inode *inode, 3769 int mds, int drop, int unless, int force) 3770 { 3771 struct ceph_inode_info *ci = ceph_inode(inode); 3772 struct ceph_cap *cap; 3773 struct ceph_mds_request_release *rel = *p; 3774 int used, dirty; 3775 int ret = 0; 3776 3777 spin_lock(&ci->i_ceph_lock); 3778 used = __ceph_caps_used(ci); 3779 dirty = __ceph_caps_dirty(ci); 3780 3781 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n", 3782 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop), 3783 ceph_cap_string(unless)); 3784 3785 /* only drop unused, clean caps */ 3786 drop &= ~(used | dirty); 3787 3788 cap = __get_cap_for_mds(ci, mds); 3789 if (cap && __cap_is_valid(cap)) { 3790 if (force || 3791 ((cap->issued & drop) && 3792 (cap->issued & unless) == 0)) { 3793 if ((cap->issued & drop) && 3794 (cap->issued & unless) == 0) { 3795 int wanted = __ceph_caps_wanted(ci); 3796 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0) 3797 wanted |= cap->mds_wanted; 3798 dout("encode_inode_release %p cap %p " 3799 "%s -> %s, wanted %s -> %s\n", inode, cap, 3800 ceph_cap_string(cap->issued), 3801 ceph_cap_string(cap->issued & ~drop), 3802 ceph_cap_string(cap->mds_wanted), 3803 ceph_cap_string(wanted)); 3804 3805 cap->issued &= ~drop; 3806 cap->implemented &= ~drop; 3807 cap->mds_wanted = wanted; 3808 } else { 3809 dout("encode_inode_release %p cap %p %s" 3810 " (force)\n", inode, cap, 3811 ceph_cap_string(cap->issued)); 3812 } 3813 3814 rel->ino = cpu_to_le64(ceph_ino(inode)); 3815 rel->cap_id = cpu_to_le64(cap->cap_id); 3816 rel->seq = cpu_to_le32(cap->seq); 3817 rel->issue_seq = cpu_to_le32(cap->issue_seq); 3818 rel->mseq = cpu_to_le32(cap->mseq); 3819 rel->caps = cpu_to_le32(cap->implemented); 3820 rel->wanted = cpu_to_le32(cap->mds_wanted); 3821 rel->dname_len = 0; 3822 rel->dname_seq = 0; 3823 *p += sizeof(*rel); 3824 ret = 1; 3825 } else { 3826 dout("encode_inode_release %p cap %p %s\n", 3827 inode, cap, ceph_cap_string(cap->issued)); 3828 } 3829 } 3830 spin_unlock(&ci->i_ceph_lock); 3831 return ret; 3832 } 3833 3834 int ceph_encode_dentry_release(void **p, struct dentry *dentry, 3835 int mds, int drop, int unless) 3836 { 3837 struct inode *dir = d_inode(dentry->d_parent); 3838 struct ceph_mds_request_release *rel = *p; 3839 struct ceph_dentry_info *di = ceph_dentry(dentry); 3840 int force = 0; 3841 int ret; 3842 3843 /* 3844 * force an record for the directory caps if we have a dentry lease. 3845 * this is racy (can't take i_ceph_lock and d_lock together), but it 3846 * doesn't have to be perfect; the mds will revoke anything we don't 3847 * release. 3848 */ 3849 spin_lock(&dentry->d_lock); 3850 if (di->lease_session && di->lease_session->s_mds == mds) 3851 force = 1; 3852 spin_unlock(&dentry->d_lock); 3853 3854 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force); 3855 3856 spin_lock(&dentry->d_lock); 3857 if (ret && di->lease_session && di->lease_session->s_mds == mds) { 3858 dout("encode_dentry_release %p mds%d seq %d\n", 3859 dentry, mds, (int)di->lease_seq); 3860 rel->dname_len = cpu_to_le32(dentry->d_name.len); 3861 memcpy(*p, dentry->d_name.name, dentry->d_name.len); 3862 *p += dentry->d_name.len; 3863 rel->dname_seq = cpu_to_le32(di->lease_seq); 3864 __ceph_mdsc_drop_dentry_lease(dentry); 3865 } 3866 spin_unlock(&dentry->d_lock); 3867 return ret; 3868 } 3869