FILE
dlmglue.c
File preview with syntax highlighting.
linux-6.19.14/fs/ocfs2/dlmglue.c
1// SPDX-License-Identifier: GPL-2.0-or-later 2/* 3 * dlmglue.c 4 * 5 * Code which implements an OCFS2 specific interface to our DLM. 6 * 7 * Copyright (C) 2003, 2004 Oracle. All rights reserved. 8 */ 9 10#include <linux/types.h> 11#include <linux/slab.h> 12#include <linux/highmem.h> 13#include <linux/mm.h> 14#include <linux/kthread.h> 15#include <linux/pagemap.h> 16#include <linux/debugfs.h> 17#include <linux/seq_file.h> 18#include <linux/time.h> 19#include <linux/delay.h> 20#include <linux/quotaops.h> 21#include <linux/sched/signal.h> 22#include <linux/string_choices.h> 23 24#define MLOG_MASK_PREFIX ML_DLM_GLUE 25#include <cluster/masklog.h> 26 27#include "ocfs2.h" 28#include "ocfs2_lockingver.h" 29 30#include "alloc.h" 31#include "dcache.h" 32#include "dlmglue.h" 33#include "extent_map.h" 34#include "file.h" 35#include "heartbeat.h" 36#include "inode.h" 37#include "journal.h" 38#include "stackglue.h" 39#include "slot_map.h" 40#include "super.h" 41#include "uptodate.h" 42#include "quota.h" 43#include "refcounttree.h" 44#include "acl.h" 45 46#include "buffer_head_io.h" 47 48struct ocfs2_mask_waiter { 49 struct list_head mw_item; 50 int mw_status; 51 struct completion mw_complete; 52 unsigned long mw_mask; 53 unsigned long mw_goal; 54#ifdef CONFIG_OCFS2_FS_STATS 55 ktime_t mw_lock_start; 56#endif 57}; 58 59static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres); 60static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres); 61static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres); 62static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres); 63 64/* 65 * Return value from ->downconvert_worker functions. 66 * 67 * These control the precise actions of ocfs2_unblock_lock() 68 * and ocfs2_process_blocked_lock() 69 * 70 */ 71enum ocfs2_unblock_action { 72 UNBLOCK_CONTINUE = 0, /* Continue downconvert */ 73 UNBLOCK_CONTINUE_POST = 1, /* Continue downconvert, fire 74 * ->post_unlock callback */ 75 UNBLOCK_STOP_POST = 2, /* Do not downconvert, fire 76 * ->post_unlock() callback. */ 77}; 78 79struct ocfs2_unblock_ctl { 80 int requeue; 81 enum ocfs2_unblock_action unblock_action; 82}; 83 84/* Lockdep class keys */ 85#ifdef CONFIG_DEBUG_LOCK_ALLOC 86static struct lock_class_key lockdep_keys[OCFS2_NUM_LOCK_TYPES]; 87#endif 88 89static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres, 90 int new_level); 91static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres); 92 93static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres, 94 int blocking); 95 96static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres, 97 int blocking); 98 99static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb, 100 struct ocfs2_lock_res *lockres); 101 102static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres); 103 104static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres, 105 int new_level); 106static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres, 107 int blocking); 108 109#define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres) 110 111/* This aids in debugging situations where a bad LVB might be involved. */ 112static void ocfs2_dump_meta_lvb_info(u64 level, 113 const char *function, 114 unsigned int line, 115 struct ocfs2_lock_res *lockres) 116{ 117 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 118 119 mlog(level, "LVB information for %s (called from %s:%u):\n", 120 lockres->l_name, function, line); 121 mlog(level, "version: %u, clusters: %u, generation: 0x%x\n", 122 lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters), 123 be32_to_cpu(lvb->lvb_igeneration)); 124 mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n", 125 (unsigned long long)be64_to_cpu(lvb->lvb_isize), 126 be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid), 127 be16_to_cpu(lvb->lvb_imode)); 128 mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, " 129 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink), 130 (long long)be64_to_cpu(lvb->lvb_iatime_packed), 131 (long long)be64_to_cpu(lvb->lvb_ictime_packed), 132 (long long)be64_to_cpu(lvb->lvb_imtime_packed), 133 be32_to_cpu(lvb->lvb_iattr)); 134} 135 136 137/* 138 * OCFS2 Lock Resource Operations 139 * 140 * These fine tune the behavior of the generic dlmglue locking infrastructure. 141 * 142 * The most basic of lock types can point ->l_priv to their respective 143 * struct ocfs2_super and allow the default actions to manage things. 144 * 145 * Right now, each lock type also needs to implement an init function, 146 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres() 147 * should be called when the lock is no longer needed (i.e., object 148 * destruction time). 149 */ 150struct ocfs2_lock_res_ops { 151 /* 152 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define 153 * this callback if ->l_priv is not an ocfs2_super pointer 154 */ 155 struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *); 156 157 /* 158 * Optionally called in the downconvert thread after a 159 * successful downconvert. The lockres will not be referenced 160 * after this callback is called, so it is safe to free 161 * memory, etc. 162 * 163 * The exact semantics of when this is called are controlled 164 * by ->downconvert_worker() 165 */ 166 void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *); 167 168 /* 169 * Allow a lock type to add checks to determine whether it is 170 * safe to downconvert a lock. Return 0 to re-queue the 171 * downconvert at a later time, nonzero to continue. 172 * 173 * For most locks, the default checks that there are no 174 * incompatible holders are sufficient. 175 * 176 * Called with the lockres spinlock held. 177 */ 178 int (*check_downconvert)(struct ocfs2_lock_res *, int); 179 180 /* 181 * Allows a lock type to populate the lock value block. This 182 * is called on downconvert, and when we drop a lock. 183 * 184 * Locks that want to use this should set LOCK_TYPE_USES_LVB 185 * in the flags field. 186 * 187 * Called with the lockres spinlock held. 188 */ 189 void (*set_lvb)(struct ocfs2_lock_res *); 190 191 /* 192 * Called from the downconvert thread when it is determined 193 * that a lock will be downconverted. This is called without 194 * any locks held so the function can do work that might 195 * schedule (syncing out data, etc). 196 * 197 * This should return any one of the ocfs2_unblock_action 198 * values, depending on what it wants the thread to do. 199 */ 200 int (*downconvert_worker)(struct ocfs2_lock_res *, int); 201 202 /* 203 * LOCK_TYPE_* flags which describe the specific requirements 204 * of a lock type. Descriptions of each individual flag follow. 205 */ 206 int flags; 207}; 208 209/* 210 * Some locks want to "refresh" potentially stale data when a 211 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this 212 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the 213 * individual lockres l_flags member from the ast function. It is 214 * expected that the locking wrapper will clear the 215 * OCFS2_LOCK_NEEDS_REFRESH flag when done. 216 */ 217#define LOCK_TYPE_REQUIRES_REFRESH 0x1 218 219/* 220 * Indicate that a lock type makes use of the lock value block. The 221 * ->set_lvb lock type callback must be defined. 222 */ 223#define LOCK_TYPE_USES_LVB 0x2 224 225static const struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = { 226 .get_osb = ocfs2_get_inode_osb, 227 .flags = 0, 228}; 229 230static const struct ocfs2_lock_res_ops ocfs2_inode_inode_lops = { 231 .get_osb = ocfs2_get_inode_osb, 232 .check_downconvert = ocfs2_check_meta_downconvert, 233 .set_lvb = ocfs2_set_meta_lvb, 234 .downconvert_worker = ocfs2_data_convert_worker, 235 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB, 236}; 237 238static const struct ocfs2_lock_res_ops ocfs2_super_lops = { 239 .flags = LOCK_TYPE_REQUIRES_REFRESH, 240}; 241 242static const struct ocfs2_lock_res_ops ocfs2_rename_lops = { 243 .flags = 0, 244}; 245 246static const struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops = { 247 .flags = 0, 248}; 249 250static const struct ocfs2_lock_res_ops ocfs2_trim_fs_lops = { 251 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB, 252}; 253 254static const struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops = { 255 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB, 256}; 257 258static const struct ocfs2_lock_res_ops ocfs2_dentry_lops = { 259 .get_osb = ocfs2_get_dentry_osb, 260 .post_unlock = ocfs2_dentry_post_unlock, 261 .downconvert_worker = ocfs2_dentry_convert_worker, 262 .flags = 0, 263}; 264 265static const struct ocfs2_lock_res_ops ocfs2_inode_open_lops = { 266 .get_osb = ocfs2_get_inode_osb, 267 .flags = 0, 268}; 269 270static const struct ocfs2_lock_res_ops ocfs2_flock_lops = { 271 .get_osb = ocfs2_get_file_osb, 272 .flags = 0, 273}; 274 275static const struct ocfs2_lock_res_ops ocfs2_qinfo_lops = { 276 .set_lvb = ocfs2_set_qinfo_lvb, 277 .get_osb = ocfs2_get_qinfo_osb, 278 .flags = LOCK_TYPE_REQUIRES_REFRESH | LOCK_TYPE_USES_LVB, 279}; 280 281static const struct ocfs2_lock_res_ops ocfs2_refcount_block_lops = { 282 .check_downconvert = ocfs2_check_refcount_downconvert, 283 .downconvert_worker = ocfs2_refcount_convert_worker, 284 .flags = 0, 285}; 286 287static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres) 288{ 289 return lockres->l_type == OCFS2_LOCK_TYPE_META || 290 lockres->l_type == OCFS2_LOCK_TYPE_RW || 291 lockres->l_type == OCFS2_LOCK_TYPE_OPEN; 292} 293 294static inline struct ocfs2_lock_res *ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb) 295{ 296 return container_of(lksb, struct ocfs2_lock_res, l_lksb); 297} 298 299static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres) 300{ 301 BUG_ON(!ocfs2_is_inode_lock(lockres)); 302 303 return (struct inode *) lockres->l_priv; 304} 305 306static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres) 307{ 308 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY); 309 310 return (struct ocfs2_dentry_lock *)lockres->l_priv; 311} 312 313static inline struct ocfs2_mem_dqinfo *ocfs2_lock_res_qinfo(struct ocfs2_lock_res *lockres) 314{ 315 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_QINFO); 316 317 return (struct ocfs2_mem_dqinfo *)lockres->l_priv; 318} 319 320static inline struct ocfs2_refcount_tree * 321ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res *res) 322{ 323 return container_of(res, struct ocfs2_refcount_tree, rf_lockres); 324} 325 326static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres) 327{ 328 if (lockres->l_ops->get_osb) 329 return lockres->l_ops->get_osb(lockres); 330 331 return (struct ocfs2_super *)lockres->l_priv; 332} 333 334static int ocfs2_lock_create(struct ocfs2_super *osb, 335 struct ocfs2_lock_res *lockres, 336 int level, 337 u32 dlm_flags); 338static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres, 339 int wanted); 340static void __ocfs2_cluster_unlock(struct ocfs2_super *osb, 341 struct ocfs2_lock_res *lockres, 342 int level, unsigned long caller_ip); 343static inline void ocfs2_cluster_unlock(struct ocfs2_super *osb, 344 struct ocfs2_lock_res *lockres, 345 int level) 346{ 347 __ocfs2_cluster_unlock(osb, lockres, level, _RET_IP_); 348} 349 350static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres); 351static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres); 352static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres); 353static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level); 354static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb, 355 struct ocfs2_lock_res *lockres); 356static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres, 357 int convert); 358#define ocfs2_log_dlm_error(_func, _err, _lockres) do { \ 359 if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY) \ 360 mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n", \ 361 _err, _func, _lockres->l_name); \ 362 else \ 363 mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n", \ 364 _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name, \ 365 (unsigned int)ocfs2_get_dentry_lock_ino(_lockres)); \ 366} while (0) 367static int ocfs2_downconvert_thread(void *arg); 368static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb, 369 struct ocfs2_lock_res *lockres); 370static int ocfs2_inode_lock_update(struct inode *inode, 371 struct buffer_head **bh); 372static void ocfs2_drop_osb_locks(struct ocfs2_super *osb); 373static inline int ocfs2_highest_compat_lock_level(int level); 374static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres, 375 int new_level); 376static int ocfs2_downconvert_lock(struct ocfs2_super *osb, 377 struct ocfs2_lock_res *lockres, 378 int new_level, 379 int lvb, 380 unsigned int generation); 381static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb, 382 struct ocfs2_lock_res *lockres); 383static int ocfs2_cancel_convert(struct ocfs2_super *osb, 384 struct ocfs2_lock_res *lockres); 385 386 387static void ocfs2_build_lock_name(enum ocfs2_lock_type type, 388 u64 blkno, 389 u32 generation, 390 char *name) 391{ 392 int len; 393 394 BUG_ON(type >= OCFS2_NUM_LOCK_TYPES); 395 396 len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x", 397 ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD, 398 (long long)blkno, generation); 399 400 BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1)); 401 402 mlog(0, "built lock resource with name: %s\n", name); 403} 404 405static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock); 406 407static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res, 408 struct ocfs2_dlm_debug *dlm_debug) 409{ 410 mlog(0, "Add tracking for lockres %s\n", res->l_name); 411 412 spin_lock(&ocfs2_dlm_tracking_lock); 413 list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking); 414 spin_unlock(&ocfs2_dlm_tracking_lock); 415} 416 417static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res) 418{ 419 spin_lock(&ocfs2_dlm_tracking_lock); 420 if (!list_empty(&res->l_debug_list)) 421 list_del_init(&res->l_debug_list); 422 spin_unlock(&ocfs2_dlm_tracking_lock); 423} 424 425#ifdef CONFIG_OCFS2_FS_STATS 426static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res) 427{ 428 res->l_lock_refresh = 0; 429 res->l_lock_wait = 0; 430 memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats)); 431 memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats)); 432} 433 434static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level, 435 struct ocfs2_mask_waiter *mw, int ret) 436{ 437 u32 usec; 438 ktime_t kt; 439 struct ocfs2_lock_stats *stats; 440 441 if (level == LKM_PRMODE) 442 stats = &res->l_lock_prmode; 443 else if (level == LKM_EXMODE) 444 stats = &res->l_lock_exmode; 445 else 446 return; 447 448 kt = ktime_sub(ktime_get(), mw->mw_lock_start); 449 usec = ktime_to_us(kt); 450 451 stats->ls_gets++; 452 stats->ls_total += ktime_to_ns(kt); 453 /* overflow */ 454 if (unlikely(stats->ls_gets == 0)) { 455 stats->ls_gets++; 456 stats->ls_total = ktime_to_ns(kt); 457 } 458 459 if (stats->ls_max < usec) 460 stats->ls_max = usec; 461 462 if (ret) 463 stats->ls_fail++; 464 465 stats->ls_last = ktime_to_us(ktime_get_real()); 466} 467 468static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres) 469{ 470 lockres->l_lock_refresh++; 471} 472 473static inline void ocfs2_track_lock_wait(struct ocfs2_lock_res *lockres) 474{ 475 struct ocfs2_mask_waiter *mw; 476 477 if (list_empty(&lockres->l_mask_waiters)) { 478 lockres->l_lock_wait = 0; 479 return; 480 } 481 482 mw = list_first_entry(&lockres->l_mask_waiters, 483 struct ocfs2_mask_waiter, mw_item); 484 lockres->l_lock_wait = 485 ktime_to_us(ktime_mono_to_real(mw->mw_lock_start)); 486} 487 488static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw) 489{ 490 mw->mw_lock_start = ktime_get(); 491} 492#else 493static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res) 494{ 495} 496static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, 497 int level, struct ocfs2_mask_waiter *mw, int ret) 498{ 499} 500static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres) 501{ 502} 503static inline void ocfs2_track_lock_wait(struct ocfs2_lock_res *lockres) 504{ 505} 506static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw) 507{ 508} 509#endif 510 511static void ocfs2_lock_res_init_common(struct ocfs2_super *osb, 512 struct ocfs2_lock_res *res, 513 enum ocfs2_lock_type type, 514 const struct ocfs2_lock_res_ops *ops, 515 void *priv) 516{ 517 res->l_type = type; 518 res->l_ops = ops; 519 res->l_priv = priv; 520 521 res->l_level = DLM_LOCK_IV; 522 res->l_requested = DLM_LOCK_IV; 523 res->l_blocking = DLM_LOCK_IV; 524 res->l_action = OCFS2_AST_INVALID; 525 res->l_unlock_action = OCFS2_UNLOCK_INVALID; 526 527 res->l_flags = OCFS2_LOCK_INITIALIZED; 528 529 ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug); 530 531 ocfs2_init_lock_stats(res); 532#ifdef CONFIG_DEBUG_LOCK_ALLOC 533 if (type != OCFS2_LOCK_TYPE_OPEN) 534 lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type], 535 &lockdep_keys[type], 0); 536 else 537 res->l_lockdep_map.key = NULL; 538#endif 539} 540 541void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res) 542{ 543 /* This also clears out the lock status block */ 544 memset(res, 0, sizeof(struct ocfs2_lock_res)); 545 spin_lock_init(&res->l_lock); 546 init_waitqueue_head(&res->l_event); 547 INIT_LIST_HEAD(&res->l_blocked_list); 548 INIT_LIST_HEAD(&res->l_mask_waiters); 549 INIT_LIST_HEAD(&res->l_holders); 550} 551 552void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res, 553 enum ocfs2_lock_type type, 554 unsigned int generation, 555 struct inode *inode) 556{ 557 const struct ocfs2_lock_res_ops *ops; 558 559 switch(type) { 560 case OCFS2_LOCK_TYPE_RW: 561 ops = &ocfs2_inode_rw_lops; 562 break; 563 case OCFS2_LOCK_TYPE_META: 564 ops = &ocfs2_inode_inode_lops; 565 break; 566 case OCFS2_LOCK_TYPE_OPEN: 567 ops = &ocfs2_inode_open_lops; 568 break; 569 default: 570 mlog_bug_on_msg(1, "type: %d\n", type); 571 ops = NULL; /* thanks, gcc */ 572 break; 573 } 574 575 ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno, 576 generation, res->l_name); 577 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode); 578} 579 580static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres) 581{ 582 struct inode *inode = ocfs2_lock_res_inode(lockres); 583 584 return OCFS2_SB(inode->i_sb); 585} 586 587static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres) 588{ 589 struct ocfs2_mem_dqinfo *info = lockres->l_priv; 590 591 return OCFS2_SB(info->dqi_gi.dqi_sb); 592} 593 594static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres) 595{ 596 struct ocfs2_file_private *fp = lockres->l_priv; 597 598 return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb); 599} 600 601static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres) 602{ 603 __be64 inode_blkno_be; 604 605 memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], 606 sizeof(__be64)); 607 608 return be64_to_cpu(inode_blkno_be); 609} 610 611static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres) 612{ 613 struct ocfs2_dentry_lock *dl = lockres->l_priv; 614 615 return OCFS2_SB(dl->dl_inode->i_sb); 616} 617 618void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl, 619 u64 parent, struct inode *inode) 620{ 621 int len; 622 u64 inode_blkno = OCFS2_I(inode)->ip_blkno; 623 __be64 inode_blkno_be = cpu_to_be64(inode_blkno); 624 struct ocfs2_lock_res *lockres = &dl->dl_lockres; 625 626 ocfs2_lock_res_init_once(lockres); 627 628 /* 629 * Unfortunately, the standard lock naming scheme won't work 630 * here because we have two 16 byte values to use. Instead, 631 * we'll stuff the inode number as a binary value. We still 632 * want error prints to show something without garbling the 633 * display, so drop a null byte in there before the inode 634 * number. A future version of OCFS2 will likely use all 635 * binary lock names. The stringified names have been a 636 * tremendous aid in debugging, but now that the debugfs 637 * interface exists, we can mangle things there if need be. 638 * 639 * NOTE: We also drop the standard "pad" value (the total lock 640 * name size stays the same though - the last part is all 641 * zeros due to the memset in ocfs2_lock_res_init_once() 642 */ 643 len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START, 644 "%c%016llx", 645 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY), 646 (long long)parent); 647 648 BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1)); 649 650 memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be, 651 sizeof(__be64)); 652 653 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres, 654 OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops, 655 dl); 656} 657 658static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res, 659 struct ocfs2_super *osb) 660{ 661 /* Superblock lockres doesn't come from a slab so we call init 662 * once on it manually. */ 663 ocfs2_lock_res_init_once(res); 664 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO, 665 0, res->l_name); 666 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER, 667 &ocfs2_super_lops, osb); 668} 669 670static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res, 671 struct ocfs2_super *osb) 672{ 673 /* Rename lockres doesn't come from a slab so we call init 674 * once on it manually. */ 675 ocfs2_lock_res_init_once(res); 676 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name); 677 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME, 678 &ocfs2_rename_lops, osb); 679} 680 681static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res, 682 struct ocfs2_super *osb) 683{ 684 /* nfs_sync lockres doesn't come from a slab so we call init 685 * once on it manually. */ 686 ocfs2_lock_res_init_once(res); 687 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name); 688 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC, 689 &ocfs2_nfs_sync_lops, osb); 690} 691 692static void ocfs2_nfs_sync_lock_init(struct ocfs2_super *osb) 693{ 694 ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb); 695 init_rwsem(&osb->nfs_sync_rwlock); 696} 697 698void ocfs2_trim_fs_lock_res_init(struct ocfs2_super *osb) 699{ 700 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres; 701 702 /* Only one trimfs thread are allowed to work at the same time. */ 703 mutex_lock(&osb->obs_trim_fs_mutex); 704 705 ocfs2_lock_res_init_once(lockres); 706 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_TRIM_FS, 0, 0, lockres->l_name); 707 ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_TRIM_FS, 708 &ocfs2_trim_fs_lops, osb); 709} 710 711void ocfs2_trim_fs_lock_res_uninit(struct ocfs2_super *osb) 712{ 713 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres; 714 715 ocfs2_simple_drop_lockres(osb, lockres); 716 ocfs2_lock_res_free(lockres); 717 718 mutex_unlock(&osb->obs_trim_fs_mutex); 719} 720 721static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res, 722 struct ocfs2_super *osb) 723{ 724 ocfs2_lock_res_init_once(res); 725 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name); 726 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN, 727 &ocfs2_orphan_scan_lops, osb); 728} 729 730void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres, 731 struct ocfs2_file_private *fp) 732{ 733 struct inode *inode = fp->fp_file->f_mapping->host; 734 struct ocfs2_inode_info *oi = OCFS2_I(inode); 735 736 ocfs2_lock_res_init_once(lockres); 737 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno, 738 inode->i_generation, lockres->l_name); 739 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres, 740 OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops, 741 fp); 742 lockres->l_flags |= OCFS2_LOCK_NOCACHE; 743} 744 745void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres, 746 struct ocfs2_mem_dqinfo *info) 747{ 748 ocfs2_lock_res_init_once(lockres); 749 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type, 750 0, lockres->l_name); 751 ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres, 752 OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops, 753 info); 754} 755 756void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres, 757 struct ocfs2_super *osb, u64 ref_blkno, 758 unsigned int generation) 759{ 760 ocfs2_lock_res_init_once(lockres); 761 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno, 762 generation, lockres->l_name); 763 ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT, 764 &ocfs2_refcount_block_lops, osb); 765} 766 767void ocfs2_lock_res_free(struct ocfs2_lock_res *res) 768{ 769 if (!(res->l_flags & OCFS2_LOCK_INITIALIZED)) 770 return; 771 772 ocfs2_remove_lockres_tracking(res); 773 774 mlog_bug_on_msg(!list_empty(&res->l_blocked_list), 775 "Lockres %s is on the blocked list\n", 776 res->l_name); 777 mlog_bug_on_msg(!list_empty(&res->l_mask_waiters), 778 "Lockres %s has mask waiters pending\n", 779 res->l_name); 780 mlog_bug_on_msg(spin_is_locked(&res->l_lock), 781 "Lockres %s is locked\n", 782 res->l_name); 783 mlog_bug_on_msg(res->l_ro_holders, 784 "Lockres %s has %u ro holders\n", 785 res->l_name, res->l_ro_holders); 786 mlog_bug_on_msg(res->l_ex_holders, 787 "Lockres %s has %u ex holders\n", 788 res->l_name, res->l_ex_holders); 789 790 /* Need to clear out the lock status block for the dlm */ 791 memset(&res->l_lksb, 0, sizeof(res->l_lksb)); 792 793 res->l_flags = 0UL; 794} 795 796/* 797 * Keep a list of processes who have interest in a lockres. 798 * Note: this is now only used for check recursive cluster locking. 799 */ 800static inline void ocfs2_add_holder(struct ocfs2_lock_res *lockres, 801 struct ocfs2_lock_holder *oh) 802{ 803 INIT_LIST_HEAD(&oh->oh_list); 804 oh->oh_owner_pid = get_pid(task_pid(current)); 805 806 spin_lock(&lockres->l_lock); 807 list_add_tail(&oh->oh_list, &lockres->l_holders); 808 spin_unlock(&lockres->l_lock); 809} 810 811static struct ocfs2_lock_holder * 812ocfs2_pid_holder(struct ocfs2_lock_res *lockres, 813 struct pid *pid) 814{ 815 struct ocfs2_lock_holder *oh; 816 817 spin_lock(&lockres->l_lock); 818 list_for_each_entry(oh, &lockres->l_holders, oh_list) { 819 if (oh->oh_owner_pid == pid) { 820 spin_unlock(&lockres->l_lock); 821 return oh; 822 } 823 } 824 spin_unlock(&lockres->l_lock); 825 return NULL; 826} 827 828static inline void ocfs2_remove_holder(struct ocfs2_lock_res *lockres, 829 struct ocfs2_lock_holder *oh) 830{ 831 spin_lock(&lockres->l_lock); 832 list_del(&oh->oh_list); 833 spin_unlock(&lockres->l_lock); 834 835 put_pid(oh->oh_owner_pid); 836} 837 838 839static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres, 840 int level) 841{ 842 BUG_ON(!lockres); 843 844 switch(level) { 845 case DLM_LOCK_EX: 846 lockres->l_ex_holders++; 847 break; 848 case DLM_LOCK_PR: 849 lockres->l_ro_holders++; 850 break; 851 default: 852 BUG(); 853 } 854} 855 856static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres, 857 int level) 858{ 859 BUG_ON(!lockres); 860 861 switch(level) { 862 case DLM_LOCK_EX: 863 BUG_ON(!lockres->l_ex_holders); 864 lockres->l_ex_holders--; 865 break; 866 case DLM_LOCK_PR: 867 BUG_ON(!lockres->l_ro_holders); 868 lockres->l_ro_holders--; 869 break; 870 default: 871 BUG(); 872 } 873} 874 875/* WARNING: This function lives in a world where the only three lock 876 * levels are EX, PR, and NL. It *will* have to be adjusted when more 877 * lock types are added. */ 878static inline int ocfs2_highest_compat_lock_level(int level) 879{ 880 int new_level = DLM_LOCK_EX; 881 882 if (level == DLM_LOCK_EX) 883 new_level = DLM_LOCK_NL; 884 else if (level == DLM_LOCK_PR) 885 new_level = DLM_LOCK_PR; 886 return new_level; 887} 888 889static void lockres_set_flags(struct ocfs2_lock_res *lockres, 890 unsigned long newflags) 891{ 892 struct ocfs2_mask_waiter *mw, *tmp; 893 894 assert_spin_locked(&lockres->l_lock); 895 896 lockres->l_flags = newflags; 897 898 list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) { 899 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal) 900 continue; 901 902 list_del_init(&mw->mw_item); 903 mw->mw_status = 0; 904 complete(&mw->mw_complete); 905 ocfs2_track_lock_wait(lockres); 906 } 907} 908static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or) 909{ 910 lockres_set_flags(lockres, lockres->l_flags | or); 911} 912static void lockres_clear_flags(struct ocfs2_lock_res *lockres, 913 unsigned long clear) 914{ 915 lockres_set_flags(lockres, lockres->l_flags & ~clear); 916} 917 918static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres) 919{ 920 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY)); 921 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED)); 922 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED)); 923 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL); 924 925 lockres->l_level = lockres->l_requested; 926 if (lockres->l_level <= 927 ocfs2_highest_compat_lock_level(lockres->l_blocking)) { 928 lockres->l_blocking = DLM_LOCK_NL; 929 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED); 930 } 931 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); 932} 933 934static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres) 935{ 936 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY)); 937 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED)); 938 939 /* Convert from RO to EX doesn't really need anything as our 940 * information is already up to data. Convert from NL to 941 * *anything* however should mark ourselves as needing an 942 * update */ 943 if (lockres->l_level == DLM_LOCK_NL && 944 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH) 945 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH); 946 947 lockres->l_level = lockres->l_requested; 948 949 /* 950 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing 951 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from 952 * downconverting the lock before the upconvert has fully completed. 953 * Do not prevent the dc thread from downconverting if NONBLOCK lock 954 * had already returned. 955 */ 956 if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED)) 957 lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING); 958 else 959 lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED); 960 961 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); 962} 963 964static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres) 965{ 966 BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY))); 967 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED); 968 969 if (lockres->l_requested > DLM_LOCK_NL && 970 !(lockres->l_flags & OCFS2_LOCK_LOCAL) && 971 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH) 972 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH); 973 974 lockres->l_level = lockres->l_requested; 975 lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED); 976 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); 977} 978 979static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, 980 int level) 981{ 982 int needs_downconvert = 0; 983 984 assert_spin_locked(&lockres->l_lock); 985 986 if (level > lockres->l_blocking) { 987 /* only schedule a downconvert if we haven't already scheduled 988 * one that goes low enough to satisfy the level we're 989 * blocking. this also catches the case where we get 990 * duplicate BASTs */ 991 if (ocfs2_highest_compat_lock_level(level) < 992 ocfs2_highest_compat_lock_level(lockres->l_blocking)) 993 needs_downconvert = 1; 994 995 lockres->l_blocking = level; 996 } 997 998 mlog(ML_BASTS, "lockres %s, block %d, level %d, l_block %d, dwn %d\n", 999 lockres->l_name, level, lockres->l_level, lockres->l_blocking, 1000 needs_downconvert); 1001 1002 if (needs_downconvert) 1003 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED); 1004 mlog(0, "needs_downconvert = %d\n", needs_downconvert); 1005 return needs_downconvert; 1006} 1007 1008/* 1009 * OCFS2_LOCK_PENDING and l_pending_gen. 1010 * 1011 * Why does OCFS2_LOCK_PENDING exist? To close a race between setting 1012 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock() 1013 * for more details on the race. 1014 * 1015 * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces 1016 * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock() 1017 * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear 1018 * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns, 1019 * the caller is going to try to clear PENDING again. If nothing else is 1020 * happening, __lockres_clear_pending() sees PENDING is unset and does 1021 * nothing. 1022 * 1023 * But what if another path (eg downconvert thread) has just started a 1024 * new locking action? The other path has re-set PENDING. Our path 1025 * cannot clear PENDING, because that will re-open the original race 1026 * window. 1027 * 1028 * [Example] 1029 * 1030 * ocfs2_meta_lock() 1031 * ocfs2_cluster_lock() 1032 * set BUSY 1033 * set PENDING 1034 * drop l_lock 1035 * ocfs2_dlm_lock() 1036 * ocfs2_locking_ast() ocfs2_downconvert_thread() 1037 * clear PENDING ocfs2_unblock_lock() 1038 * take_l_lock 1039 * !BUSY 1040 * ocfs2_prepare_downconvert() 1041 * set BUSY 1042 * set PENDING 1043 * drop l_lock 1044 * take l_lock 1045 * clear PENDING 1046 * drop l_lock 1047 * <window> 1048 * ocfs2_dlm_lock() 1049 * 1050 * So as you can see, we now have a window where l_lock is not held, 1051 * PENDING is not set, and ocfs2_dlm_lock() has not been called. 1052 * 1053 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING 1054 * set by ocfs2_prepare_downconvert(). That wasn't nice. 1055 * 1056 * To solve this we introduce l_pending_gen. A call to 1057 * lockres_clear_pending() will only do so when it is passed a generation 1058 * number that matches the lockres. lockres_set_pending() will return the 1059 * current generation number. When ocfs2_cluster_lock() goes to clear 1060 * PENDING, it passes the generation it got from set_pending(). In our 1061 * example above, the generation numbers will *not* match. Thus, 1062 * ocfs2_cluster_lock() will not clear the PENDING set by 1063 * ocfs2_prepare_downconvert(). 1064 */ 1065 1066/* Unlocked version for ocfs2_locking_ast() */ 1067static void __lockres_clear_pending(struct ocfs2_lock_res *lockres, 1068 unsigned int generation, 1069 struct ocfs2_super *osb) 1070{ 1071 assert_spin_locked(&lockres->l_lock); 1072 1073 /* 1074 * The ast and locking functions can race us here. The winner 1075 * will clear pending, the loser will not. 1076 */ 1077 if (!(lockres->l_flags & OCFS2_LOCK_PENDING) || 1078 (lockres->l_pending_gen != generation)) 1079 return; 1080 1081 lockres_clear_flags(lockres, OCFS2_LOCK_PENDING); 1082 lockres->l_pending_gen++; 1083 1084 /* 1085 * The downconvert thread may have skipped us because we 1086 * were PENDING. Wake it up. 1087 */ 1088 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) 1089 ocfs2_wake_downconvert_thread(osb); 1090} 1091 1092/* Locked version for callers of ocfs2_dlm_lock() */ 1093static void lockres_clear_pending(struct ocfs2_lock_res *lockres, 1094 unsigned int generation, 1095 struct ocfs2_super *osb) 1096{ 1097 unsigned long flags; 1098 1099 spin_lock_irqsave(&lockres->l_lock, flags); 1100 __lockres_clear_pending(lockres, generation, osb); 1101 spin_unlock_irqrestore(&lockres->l_lock, flags); 1102} 1103 1104static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres) 1105{ 1106 assert_spin_locked(&lockres->l_lock); 1107 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY)); 1108 1109 lockres_or_flags(lockres, OCFS2_LOCK_PENDING); 1110 1111 return lockres->l_pending_gen; 1112} 1113 1114static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level) 1115{ 1116 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb); 1117 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres); 1118 int needs_downconvert; 1119 unsigned long flags; 1120 1121 BUG_ON(level <= DLM_LOCK_NL); 1122 1123 mlog(ML_BASTS, "BAST fired for lockres %s, blocking %d, level %d, " 1124 "type %s\n", lockres->l_name, level, lockres->l_level, 1125 ocfs2_lock_type_string(lockres->l_type)); 1126 1127 /* 1128 * We can skip the bast for locks which don't enable caching - 1129 * they'll be dropped at the earliest possible time anyway. 1130 */ 1131 if (lockres->l_flags & OCFS2_LOCK_NOCACHE) 1132 return; 1133 1134 spin_lock_irqsave(&lockres->l_lock, flags); 1135 needs_downconvert = ocfs2_generic_handle_bast(lockres, level); 1136 if (needs_downconvert) 1137 ocfs2_schedule_blocked_lock(osb, lockres); 1138 spin_unlock_irqrestore(&lockres->l_lock, flags); 1139 1140 wake_up(&lockres->l_event); 1141 1142 ocfs2_wake_downconvert_thread(osb); 1143} 1144 1145static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb) 1146{ 1147 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb); 1148 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres); 1149 unsigned long flags; 1150 int status; 1151 1152 spin_lock_irqsave(&lockres->l_lock, flags); 1153 1154 status = ocfs2_dlm_lock_status(&lockres->l_lksb); 1155 1156 if (status == -EAGAIN) { 1157 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); 1158 goto out; 1159 } 1160 1161 if (status) { 1162 mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n", 1163 lockres->l_name, status); 1164 spin_unlock_irqrestore(&lockres->l_lock, flags); 1165 return; 1166 } 1167 1168 mlog(ML_BASTS, "AST fired for lockres %s, action %d, unlock %d, " 1169 "level %d => %d\n", lockres->l_name, lockres->l_action, 1170 lockres->l_unlock_action, lockres->l_level, lockres->l_requested); 1171 1172 switch(lockres->l_action) { 1173 case OCFS2_AST_ATTACH: 1174 ocfs2_generic_handle_attach_action(lockres); 1175 lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL); 1176 break; 1177 case OCFS2_AST_CONVERT: 1178 ocfs2_generic_handle_convert_action(lockres); 1179 break; 1180 case OCFS2_AST_DOWNCONVERT: 1181 ocfs2_generic_handle_downconvert_action(lockres); 1182 break; 1183 default: 1184 mlog(ML_ERROR, "lockres %s: AST fired with invalid action: %u, " 1185 "flags 0x%lx, unlock: %u\n", 1186 lockres->l_name, lockres->l_action, lockres->l_flags, 1187 lockres->l_unlock_action); 1188 BUG(); 1189 } 1190out: 1191 /* set it to something invalid so if we get called again we 1192 * can catch it. */ 1193 lockres->l_action = OCFS2_AST_INVALID; 1194 1195 /* Did we try to cancel this lock? Clear that state */ 1196 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) 1197 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID; 1198 1199 /* 1200 * We may have beaten the locking functions here. We certainly 1201 * know that dlm_lock() has been called :-) 1202 * Because we can't have two lock calls in flight at once, we 1203 * can use lockres->l_pending_gen. 1204 */ 1205 __lockres_clear_pending(lockres, lockres->l_pending_gen, osb); 1206 1207 wake_up(&lockres->l_event); 1208 spin_unlock_irqrestore(&lockres->l_lock, flags); 1209} 1210 1211static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error) 1212{ 1213 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb); 1214 unsigned long flags; 1215 1216 mlog(ML_BASTS, "UNLOCK AST fired for lockres %s, action = %d\n", 1217 lockres->l_name, lockres->l_unlock_action); 1218 1219 spin_lock_irqsave(&lockres->l_lock, flags); 1220 if (error) { 1221 mlog(ML_ERROR, "Dlm passes error %d for lock %s, " 1222 "unlock_action %d\n", error, lockres->l_name, 1223 lockres->l_unlock_action); 1224 spin_unlock_irqrestore(&lockres->l_lock, flags); 1225 return; 1226 } 1227 1228 switch(lockres->l_unlock_action) { 1229 case OCFS2_UNLOCK_CANCEL_CONVERT: 1230 mlog(0, "Cancel convert success for %s\n", lockres->l_name); 1231 lockres->l_action = OCFS2_AST_INVALID; 1232 /* Downconvert thread may have requeued this lock, we 1233 * need to wake it. */ 1234 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) 1235 ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres)); 1236 break; 1237 case OCFS2_UNLOCK_DROP_LOCK: 1238 lockres->l_level = DLM_LOCK_IV; 1239 break; 1240 default: 1241 BUG(); 1242 } 1243 1244 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); 1245 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID; 1246 wake_up(&lockres->l_event); 1247 spin_unlock_irqrestore(&lockres->l_lock, flags); 1248} 1249 1250/* 1251 * This is the filesystem locking protocol. It provides the lock handling 1252 * hooks for the underlying DLM. It has a maximum version number. 1253 * The version number allows interoperability with systems running at 1254 * the same major number and an equal or smaller minor number. 1255 * 1256 * Whenever the filesystem does new things with locks (adds or removes a 1257 * lock, orders them differently, does different things underneath a lock), 1258 * the version must be changed. The protocol is negotiated when joining 1259 * the dlm domain. A node may join the domain if its major version is 1260 * identical to all other nodes and its minor version is greater than 1261 * or equal to all other nodes. When its minor version is greater than 1262 * the other nodes, it will run at the minor version specified by the 1263 * other nodes. 1264 * 1265 * If a locking change is made that will not be compatible with older 1266 * versions, the major number must be increased and the minor version set 1267 * to zero. If a change merely adds a behavior that can be disabled when 1268 * speaking to older versions, the minor version must be increased. If a 1269 * change adds a fully backwards compatible change (eg, LVB changes that 1270 * are just ignored by older versions), the version does not need to be 1271 * updated. 1272 */ 1273static struct ocfs2_locking_protocol lproto = { 1274 .lp_max_version = { 1275 .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR, 1276 .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR, 1277 }, 1278 .lp_lock_ast = ocfs2_locking_ast, 1279 .lp_blocking_ast = ocfs2_blocking_ast, 1280 .lp_unlock_ast = ocfs2_unlock_ast, 1281}; 1282 1283void ocfs2_set_locking_protocol(void) 1284{ 1285 ocfs2_stack_glue_set_max_proto_version(&lproto.lp_max_version); 1286} 1287 1288static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres, 1289 int convert) 1290{ 1291 unsigned long flags; 1292 1293 spin_lock_irqsave(&lockres->l_lock, flags); 1294 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); 1295 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING); 1296 if (convert) 1297 lockres->l_action = OCFS2_AST_INVALID; 1298 else 1299 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID; 1300 spin_unlock_irqrestore(&lockres->l_lock, flags); 1301 1302 wake_up(&lockres->l_event); 1303} 1304 1305/* Note: If we detect another process working on the lock (i.e., 1306 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller 1307 * to do the right thing in that case. 1308 */ 1309static int ocfs2_lock_create(struct ocfs2_super *osb, 1310 struct ocfs2_lock_res *lockres, 1311 int level, 1312 u32 dlm_flags) 1313{ 1314 int ret = 0; 1315 unsigned long flags; 1316 unsigned int gen; 1317 1318 mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level, 1319 dlm_flags); 1320 1321 spin_lock_irqsave(&lockres->l_lock, flags); 1322 if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) || 1323 (lockres->l_flags & OCFS2_LOCK_BUSY)) { 1324 spin_unlock_irqrestore(&lockres->l_lock, flags); 1325 goto bail; 1326 } 1327 1328 lockres->l_action = OCFS2_AST_ATTACH; 1329 lockres->l_requested = level; 1330 lockres_or_flags(lockres, OCFS2_LOCK_BUSY); 1331 gen = lockres_set_pending(lockres); 1332 spin_unlock_irqrestore(&lockres->l_lock, flags); 1333 1334 ret = ocfs2_dlm_lock(osb->cconn, 1335 level, 1336 &lockres->l_lksb, 1337 dlm_flags, 1338 lockres->l_name, 1339 OCFS2_LOCK_ID_MAX_LEN - 1); 1340 lockres_clear_pending(lockres, gen, osb); 1341 if (ret) { 1342 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres); 1343 ocfs2_recover_from_dlm_error(lockres, 1); 1344 } 1345 1346 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name); 1347 1348bail: 1349 return ret; 1350} 1351 1352static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres, 1353 int flag) 1354{ 1355 unsigned long flags; 1356 int ret; 1357 1358 spin_lock_irqsave(&lockres->l_lock, flags); 1359 ret = lockres->l_flags & flag; 1360 spin_unlock_irqrestore(&lockres->l_lock, flags); 1361 1362 return ret; 1363} 1364 1365static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres) 1366 1367{ 1368 wait_event(lockres->l_event, 1369 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY)); 1370} 1371 1372static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres) 1373 1374{ 1375 wait_event(lockres->l_event, 1376 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING)); 1377} 1378 1379/* predict what lock level we'll be dropping down to on behalf 1380 * of another node, and return true if the currently wanted 1381 * level will be compatible with it. */ 1382static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres, 1383 int wanted) 1384{ 1385 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED)); 1386 1387 return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking); 1388} 1389 1390static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw) 1391{ 1392 INIT_LIST_HEAD(&mw->mw_item); 1393 init_completion(&mw->mw_complete); 1394 ocfs2_init_start_time(mw); 1395} 1396 1397static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw) 1398{ 1399 wait_for_completion(&mw->mw_complete); 1400 /* Re-arm the completion in case we want to wait on it again */ 1401 reinit_completion(&mw->mw_complete); 1402 return mw->mw_status; 1403} 1404 1405static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres, 1406 struct ocfs2_mask_waiter *mw, 1407 unsigned long mask, 1408 unsigned long goal) 1409{ 1410 BUG_ON(!list_empty(&mw->mw_item)); 1411 1412 assert_spin_locked(&lockres->l_lock); 1413 1414 list_add_tail(&mw->mw_item, &lockres->l_mask_waiters); 1415 mw->mw_mask = mask; 1416 mw->mw_goal = goal; 1417 ocfs2_track_lock_wait(lockres); 1418} 1419 1420/* returns 0 if the mw that was removed was already satisfied, -EBUSY 1421 * if the mask still hadn't reached its goal */ 1422static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres, 1423 struct ocfs2_mask_waiter *mw) 1424{ 1425 int ret = 0; 1426 1427 assert_spin_locked(&lockres->l_lock); 1428 if (!list_empty(&mw->mw_item)) { 1429 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal) 1430 ret = -EBUSY; 1431 1432 list_del_init(&mw->mw_item); 1433 init_completion(&mw->mw_complete); 1434 ocfs2_track_lock_wait(lockres); 1435 } 1436 1437 return ret; 1438} 1439 1440static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres, 1441 struct ocfs2_mask_waiter *mw) 1442{ 1443 unsigned long flags; 1444 int ret = 0; 1445 1446 spin_lock_irqsave(&lockres->l_lock, flags); 1447 ret = __lockres_remove_mask_waiter(lockres, mw); 1448 spin_unlock_irqrestore(&lockres->l_lock, flags); 1449 1450 return ret; 1451 1452} 1453 1454static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw, 1455 struct ocfs2_lock_res *lockres) 1456{ 1457 int ret; 1458 1459 ret = wait_for_completion_interruptible(&mw->mw_complete); 1460 if (ret) 1461 lockres_remove_mask_waiter(lockres, mw); 1462 else 1463 ret = mw->mw_status; 1464 /* Re-arm the completion in case we want to wait on it again */ 1465 reinit_completion(&mw->mw_complete); 1466 return ret; 1467} 1468 1469static int __ocfs2_cluster_lock(struct ocfs2_super *osb, 1470 struct ocfs2_lock_res *lockres, 1471 int level, 1472 u32 lkm_flags, 1473 int arg_flags, 1474 int l_subclass, 1475 unsigned long caller_ip) 1476{ 1477 struct ocfs2_mask_waiter mw; 1478 int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR); 1479 int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */ 1480 unsigned long flags; 1481 unsigned int gen; 1482 int noqueue_attempted = 0; 1483 int dlm_locked = 0; 1484 int kick_dc = 0; 1485 1486 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED)) { 1487 mlog_errno(-EINVAL); 1488 return -EINVAL; 1489 } 1490 1491 ocfs2_init_mask_waiter(&mw); 1492 1493 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) 1494 lkm_flags |= DLM_LKF_VALBLK; 1495 1496again: 1497 wait = 0; 1498 1499 spin_lock_irqsave(&lockres->l_lock, flags); 1500 1501 if (catch_signals && signal_pending(current)) { 1502 ret = -ERESTARTSYS; 1503 goto unlock; 1504 } 1505 1506 mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING, 1507 "Cluster lock called on freeing lockres %s! flags " 1508 "0x%lx\n", lockres->l_name, lockres->l_flags); 1509 1510 /* We only compare against the currently granted level 1511 * here. If the lock is blocked waiting on a downconvert, 1512 * we'll get caught below. */ 1513 if (lockres->l_flags & OCFS2_LOCK_BUSY && 1514 level > lockres->l_level) { 1515 /* is someone sitting in dlm_lock? If so, wait on 1516 * them. */ 1517 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0); 1518 wait = 1; 1519 goto unlock; 1520 } 1521 1522 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) { 1523 /* 1524 * We've upconverted. If the lock now has a level we can 1525 * work with, we take it. If, however, the lock is not at the 1526 * required level, we go thru the full cycle. One way this could 1527 * happen is if a process requesting an upconvert to PR is 1528 * closely followed by another requesting upconvert to an EX. 1529 * If the process requesting EX lands here, we want it to 1530 * continue attempting to upconvert and let the process 1531 * requesting PR take the lock. 1532 * If multiple processes request upconvert to PR, the first one 1533 * here will take the lock. The others will have to go thru the 1534 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending 1535 * downconvert request. 1536 */ 1537 if (level <= lockres->l_level) 1538 goto update_holders; 1539 } 1540 1541 if (lockres->l_flags & OCFS2_LOCK_BLOCKED && 1542 !ocfs2_may_continue_on_blocked_lock(lockres, level)) { 1543 /* is the lock is currently blocked on behalf of 1544 * another node */ 1545 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0); 1546 wait = 1; 1547 goto unlock; 1548 } 1549 1550 if (level > lockres->l_level) { 1551 if (noqueue_attempted > 0) { 1552 ret = -EAGAIN; 1553 goto unlock; 1554 } 1555 if (lkm_flags & DLM_LKF_NOQUEUE) 1556 noqueue_attempted = 1; 1557 1558 if (lockres->l_action != OCFS2_AST_INVALID) 1559 mlog(ML_ERROR, "lockres %s has action %u pending\n", 1560 lockres->l_name, lockres->l_action); 1561 1562 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) { 1563 lockres->l_action = OCFS2_AST_ATTACH; 1564 lkm_flags &= ~DLM_LKF_CONVERT; 1565 } else { 1566 lockres->l_action = OCFS2_AST_CONVERT; 1567 lkm_flags |= DLM_LKF_CONVERT; 1568 } 1569 1570 lockres->l_requested = level; 1571 lockres_or_flags(lockres, OCFS2_LOCK_BUSY); 1572 gen = lockres_set_pending(lockres); 1573 spin_unlock_irqrestore(&lockres->l_lock, flags); 1574 1575 BUG_ON(level == DLM_LOCK_IV); 1576 BUG_ON(level == DLM_LOCK_NL); 1577 1578 mlog(ML_BASTS, "lockres %s, convert from %d to %d\n", 1579 lockres->l_name, lockres->l_level, level); 1580 1581 /* call dlm_lock to upgrade lock now */ 1582 ret = ocfs2_dlm_lock(osb->cconn, 1583 level, 1584 &lockres->l_lksb, 1585 lkm_flags, 1586 lockres->l_name, 1587 OCFS2_LOCK_ID_MAX_LEN - 1); 1588 lockres_clear_pending(lockres, gen, osb); 1589 if (ret) { 1590 if (!(lkm_flags & DLM_LKF_NOQUEUE) || 1591 (ret != -EAGAIN)) { 1592 ocfs2_log_dlm_error("ocfs2_dlm_lock", 1593 ret, lockres); 1594 } 1595 ocfs2_recover_from_dlm_error(lockres, 1); 1596 goto out; 1597 } 1598 dlm_locked = 1; 1599 1600 mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n", 1601 lockres->l_name); 1602 1603 /* At this point we've gone inside the dlm and need to 1604 * complete our work regardless. */ 1605 catch_signals = 0; 1606 1607 /* wait for busy to clear and carry on */ 1608 goto again; 1609 } 1610 1611update_holders: 1612 /* Ok, if we get here then we're good to go. */ 1613 ocfs2_inc_holders(lockres, level); 1614 1615 ret = 0; 1616unlock: 1617 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING); 1618 1619 /* ocfs2_unblock_lock request on seeing OCFS2_LOCK_UPCONVERT_FINISHING */ 1620 kick_dc = (lockres->l_flags & OCFS2_LOCK_BLOCKED); 1621 1622 spin_unlock_irqrestore(&lockres->l_lock, flags); 1623 if (kick_dc) 1624 ocfs2_wake_downconvert_thread(osb); 1625out: 1626 /* 1627 * This is helping work around a lock inversion between the page lock 1628 * and dlm locks. One path holds the page lock while calling aops 1629 * which block acquiring dlm locks. The voting thread holds dlm 1630 * locks while acquiring page locks while down converting data locks. 1631 * This block is helping an aop path notice the inversion and back 1632 * off to unlock its page lock before trying the dlm lock again. 1633 */ 1634 if (wait && arg_flags & OCFS2_LOCK_NONBLOCK && 1635 mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) { 1636 wait = 0; 1637 spin_lock_irqsave(&lockres->l_lock, flags); 1638 if (__lockres_remove_mask_waiter(lockres, &mw)) { 1639 if (dlm_locked) 1640 lockres_or_flags(lockres, 1641 OCFS2_LOCK_NONBLOCK_FINISHED); 1642 spin_unlock_irqrestore(&lockres->l_lock, flags); 1643 ret = -EAGAIN; 1644 } else { 1645 spin_unlock_irqrestore(&lockres->l_lock, flags); 1646 goto again; 1647 } 1648 } 1649 if (wait) { 1650 ret = ocfs2_wait_for_mask(&mw); 1651 if (ret == 0) 1652 goto again; 1653 mlog_errno(ret); 1654 } 1655 ocfs2_update_lock_stats(lockres, level, &mw, ret); 1656 1657#ifdef CONFIG_DEBUG_LOCK_ALLOC 1658 if (!ret && lockres->l_lockdep_map.key != NULL) { 1659 if (level == DLM_LOCK_PR) 1660 rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass, 1661 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE), 1662 caller_ip); 1663 else 1664 rwsem_acquire(&lockres->l_lockdep_map, l_subclass, 1665 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE), 1666 caller_ip); 1667 } 1668#endif 1669 return ret; 1670} 1671 1672static inline int ocfs2_cluster_lock(struct ocfs2_super *osb, 1673 struct ocfs2_lock_res *lockres, 1674 int level, 1675 u32 lkm_flags, 1676 int arg_flags) 1677{ 1678 return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags, 1679 0, _RET_IP_); 1680} 1681 1682 1683static void __ocfs2_cluster_unlock(struct ocfs2_super *osb, 1684 struct ocfs2_lock_res *lockres, 1685 int level, 1686 unsigned long caller_ip) 1687{ 1688 unsigned long flags; 1689 1690 spin_lock_irqsave(&lockres->l_lock, flags); 1691 ocfs2_dec_holders(lockres, level); 1692 ocfs2_downconvert_on_unlock(osb, lockres); 1693 spin_unlock_irqrestore(&lockres->l_lock, flags); 1694#ifdef CONFIG_DEBUG_LOCK_ALLOC 1695 if (lockres->l_lockdep_map.key != NULL) 1696 rwsem_release(&lockres->l_lockdep_map, caller_ip); 1697#endif 1698} 1699 1700static int ocfs2_create_new_lock(struct ocfs2_super *osb, 1701 struct ocfs2_lock_res *lockres, 1702 int ex, 1703 int local) 1704{ 1705 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 1706 unsigned long flags; 1707 u32 lkm_flags = local ? DLM_LKF_LOCAL : 0; 1708 1709 spin_lock_irqsave(&lockres->l_lock, flags); 1710 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED); 1711 lockres_or_flags(lockres, OCFS2_LOCK_LOCAL); 1712 spin_unlock_irqrestore(&lockres->l_lock, flags); 1713 1714 return ocfs2_lock_create(osb, lockres, level, lkm_flags); 1715} 1716 1717/* Grants us an EX lock on the data and metadata resources, skipping 1718 * the normal cluster directory lookup. Use this ONLY on newly created 1719 * inodes which other nodes can't possibly see, and which haven't been 1720 * hashed in the inode hash yet. This can give us a good performance 1721 * increase as it'll skip the network broadcast normally associated 1722 * with creating a new lock resource. */ 1723int ocfs2_create_new_inode_locks(struct inode *inode) 1724{ 1725 int ret; 1726 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1727 1728 BUG_ON(!ocfs2_inode_is_new(inode)); 1729 1730 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno); 1731 1732 /* NOTE: That we don't increment any of the holder counts, nor 1733 * do we add anything to a journal handle. Since this is 1734 * supposed to be a new inode which the cluster doesn't know 1735 * about yet, there is no need to. As far as the LVB handling 1736 * is concerned, this is basically like acquiring an EX lock 1737 * on a resource which has an invalid one -- we'll set it 1738 * valid when we release the EX. */ 1739 1740 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1); 1741 if (ret) { 1742 mlog_errno(ret); 1743 goto bail; 1744 } 1745 1746 /* 1747 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they 1748 * don't use a generation in their lock names. 1749 */ 1750 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0); 1751 if (ret) { 1752 mlog_errno(ret); 1753 goto bail; 1754 } 1755 1756 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0); 1757 if (ret) 1758 mlog_errno(ret); 1759 1760bail: 1761 return ret; 1762} 1763 1764int ocfs2_rw_lock(struct inode *inode, int write) 1765{ 1766 int status, level; 1767 struct ocfs2_lock_res *lockres; 1768 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1769 1770 mlog(0, "inode %llu take %s RW lock\n", 1771 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1772 write ? "EXMODE" : "PRMODE"); 1773 1774 if (ocfs2_mount_local(osb)) 1775 return 0; 1776 1777 lockres = &OCFS2_I(inode)->ip_rw_lockres; 1778 1779 level = write ? DLM_LOCK_EX : DLM_LOCK_PR; 1780 1781 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0); 1782 if (status < 0) 1783 mlog_errno(status); 1784 1785 return status; 1786} 1787 1788int ocfs2_try_rw_lock(struct inode *inode, int write) 1789{ 1790 int status, level; 1791 struct ocfs2_lock_res *lockres; 1792 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1793 1794 mlog(0, "inode %llu try to take %s RW lock\n", 1795 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1796 write ? "EXMODE" : "PRMODE"); 1797 1798 if (ocfs2_mount_local(osb)) 1799 return 0; 1800 1801 lockres = &OCFS2_I(inode)->ip_rw_lockres; 1802 1803 level = write ? DLM_LOCK_EX : DLM_LOCK_PR; 1804 1805 status = ocfs2_cluster_lock(osb, lockres, level, DLM_LKF_NOQUEUE, 0); 1806 return status; 1807} 1808 1809void ocfs2_rw_unlock(struct inode *inode, int write) 1810{ 1811 int level = write ? DLM_LOCK_EX : DLM_LOCK_PR; 1812 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres; 1813 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1814 1815 mlog(0, "inode %llu drop %s RW lock\n", 1816 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1817 write ? "EXMODE" : "PRMODE"); 1818 1819 if (!ocfs2_mount_local(osb)) 1820 ocfs2_cluster_unlock(osb, lockres, level); 1821} 1822 1823/* 1824 * ocfs2_open_lock always get PR mode lock. 1825 */ 1826int ocfs2_open_lock(struct inode *inode) 1827{ 1828 int status = 0; 1829 struct ocfs2_lock_res *lockres; 1830 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1831 1832 mlog(0, "inode %llu take PRMODE open lock\n", 1833 (unsigned long long)OCFS2_I(inode)->ip_blkno); 1834 1835 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb)) 1836 goto out; 1837 1838 lockres = &OCFS2_I(inode)->ip_open_lockres; 1839 1840 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_PR, 0, 0); 1841 if (status < 0) 1842 mlog_errno(status); 1843 1844out: 1845 return status; 1846} 1847 1848int ocfs2_try_open_lock(struct inode *inode, int write) 1849{ 1850 int status = 0, level; 1851 struct ocfs2_lock_res *lockres; 1852 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1853 1854 mlog(0, "inode %llu try to take %s open lock\n", 1855 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1856 write ? "EXMODE" : "PRMODE"); 1857 1858 if (ocfs2_is_hard_readonly(osb)) { 1859 if (write) 1860 status = -EROFS; 1861 goto out; 1862 } 1863 1864 if (ocfs2_mount_local(osb)) 1865 goto out; 1866 1867 lockres = &OCFS2_I(inode)->ip_open_lockres; 1868 1869 level = write ? DLM_LOCK_EX : DLM_LOCK_PR; 1870 1871 /* 1872 * The file system may already holding a PRMODE/EXMODE open lock. 1873 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on 1874 * other nodes and the -EAGAIN will indicate to the caller that 1875 * this inode is still in use. 1876 */ 1877 status = ocfs2_cluster_lock(osb, lockres, level, DLM_LKF_NOQUEUE, 0); 1878 1879out: 1880 return status; 1881} 1882 1883/* 1884 * ocfs2_open_unlock unlock PR and EX mode open locks. 1885 */ 1886void ocfs2_open_unlock(struct inode *inode) 1887{ 1888 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres; 1889 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1890 1891 mlog(0, "inode %llu drop open lock\n", 1892 (unsigned long long)OCFS2_I(inode)->ip_blkno); 1893 1894 if (ocfs2_mount_local(osb)) 1895 goto out; 1896 1897 if(lockres->l_ro_holders) 1898 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_PR); 1899 if(lockres->l_ex_holders) 1900 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX); 1901 1902out: 1903 return; 1904} 1905 1906static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres, 1907 int level) 1908{ 1909 int ret; 1910 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres); 1911 unsigned long flags; 1912 struct ocfs2_mask_waiter mw; 1913 1914 ocfs2_init_mask_waiter(&mw); 1915 1916retry_cancel: 1917 spin_lock_irqsave(&lockres->l_lock, flags); 1918 if (lockres->l_flags & OCFS2_LOCK_BUSY) { 1919 ret = ocfs2_prepare_cancel_convert(osb, lockres); 1920 if (ret) { 1921 spin_unlock_irqrestore(&lockres->l_lock, flags); 1922 ret = ocfs2_cancel_convert(osb, lockres); 1923 if (ret < 0) { 1924 mlog_errno(ret); 1925 goto out; 1926 } 1927 goto retry_cancel; 1928 } 1929 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0); 1930 spin_unlock_irqrestore(&lockres->l_lock, flags); 1931 1932 ocfs2_wait_for_mask(&mw); 1933 goto retry_cancel; 1934 } 1935 1936 ret = -ERESTARTSYS; 1937 /* 1938 * We may still have gotten the lock, in which case there's no 1939 * point to restarting the syscall. 1940 */ 1941 if (lockres->l_level == level) 1942 ret = 0; 1943 1944 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret, 1945 lockres->l_flags, lockres->l_level, lockres->l_action); 1946 1947 spin_unlock_irqrestore(&lockres->l_lock, flags); 1948 1949out: 1950 return ret; 1951} 1952 1953/* 1954 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of 1955 * flock() calls. The locking approach this requires is sufficiently 1956 * different from all other cluster lock types that we implement a 1957 * separate path to the "low-level" dlm calls. In particular: 1958 * 1959 * - No optimization of lock levels is done - we take at exactly 1960 * what's been requested. 1961 * 1962 * - No lock caching is employed. We immediately downconvert to 1963 * no-lock at unlock time. This also means flock locks never go on 1964 * the blocking list). 1965 * 1966 * - Since userspace can trivially deadlock itself with flock, we make 1967 * sure to allow cancellation of a misbehaving applications flock() 1968 * request. 1969 * 1970 * - Access to any flock lockres doesn't require concurrency, so we 1971 * can simplify the code by requiring the caller to guarantee 1972 * serialization of dlmglue flock calls. 1973 */ 1974int ocfs2_file_lock(struct file *file, int ex, int trylock) 1975{ 1976 int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 1977 unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0; 1978 unsigned long flags; 1979 struct ocfs2_file_private *fp = file->private_data; 1980 struct ocfs2_lock_res *lockres = &fp->fp_flock; 1981 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb); 1982 struct ocfs2_mask_waiter mw; 1983 1984 ocfs2_init_mask_waiter(&mw); 1985 1986 if ((lockres->l_flags & OCFS2_LOCK_BUSY) || 1987 (lockres->l_level > DLM_LOCK_NL)) { 1988 mlog(ML_ERROR, 1989 "File lock \"%s\" has busy or locked state: flags: 0x%lx, " 1990 "level: %u\n", lockres->l_name, lockres->l_flags, 1991 lockres->l_level); 1992 return -EINVAL; 1993 } 1994 1995 spin_lock_irqsave(&lockres->l_lock, flags); 1996 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) { 1997 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0); 1998 spin_unlock_irqrestore(&lockres->l_lock, flags); 1999 2000 /* 2001 * Get the lock at NLMODE to start - that way we 2002 * can cancel the upconvert request if need be. 2003 */ 2004 ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0); 2005 if (ret < 0) { 2006 mlog_errno(ret); 2007 goto out; 2008 } 2009 2010 ret = ocfs2_wait_for_mask(&mw); 2011 if (ret) { 2012 mlog_errno(ret); 2013 goto out; 2014 } 2015 spin_lock_irqsave(&lockres->l_lock, flags); 2016 } 2017 2018 lockres->l_action = OCFS2_AST_CONVERT; 2019 lkm_flags |= DLM_LKF_CONVERT; 2020 lockres->l_requested = level; 2021 lockres_or_flags(lockres, OCFS2_LOCK_BUSY); 2022 2023 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0); 2024 spin_unlock_irqrestore(&lockres->l_lock, flags); 2025 2026 ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags, 2027 lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1); 2028 if (ret) { 2029 if (!trylock || (ret != -EAGAIN)) { 2030 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres); 2031 ret = -EINVAL; 2032 } 2033 2034 ocfs2_recover_from_dlm_error(lockres, 1); 2035 lockres_remove_mask_waiter(lockres, &mw); 2036 goto out; 2037 } 2038 2039 ret = ocfs2_wait_for_mask_interruptible(&mw, lockres); 2040 if (ret == -ERESTARTSYS) { 2041 /* 2042 * Userspace can cause deadlock itself with 2043 * flock(). Current behavior locally is to allow the 2044 * deadlock, but abort the system call if a signal is 2045 * received. We follow this example, otherwise a 2046 * poorly written program could sit in kernel until 2047 * reboot. 2048 * 2049 * Handling this is a bit more complicated for Ocfs2 2050 * though. We can't exit this function with an 2051 * outstanding lock request, so a cancel convert is 2052 * required. We intentionally overwrite 'ret' - if the 2053 * cancel fails and the lock was granted, it's easier 2054 * to just bubble success back up to the user. 2055 */ 2056 ret = ocfs2_flock_handle_signal(lockres, level); 2057 } else if (!ret && (level > lockres->l_level)) { 2058 /* Trylock failed asynchronously */ 2059 BUG_ON(!trylock); 2060 ret = -EAGAIN; 2061 } 2062 2063out: 2064 2065 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n", 2066 lockres->l_name, ex, trylock, ret); 2067 return ret; 2068} 2069 2070void ocfs2_file_unlock(struct file *file) 2071{ 2072 int ret; 2073 unsigned int gen; 2074 unsigned long flags; 2075 struct ocfs2_file_private *fp = file->private_data; 2076 struct ocfs2_lock_res *lockres = &fp->fp_flock; 2077 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb); 2078 struct ocfs2_mask_waiter mw; 2079 2080 ocfs2_init_mask_waiter(&mw); 2081 2082 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) 2083 return; 2084 2085 if (lockres->l_level == DLM_LOCK_NL) 2086 return; 2087 2088 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n", 2089 lockres->l_name, lockres->l_flags, lockres->l_level, 2090 lockres->l_action); 2091 2092 spin_lock_irqsave(&lockres->l_lock, flags); 2093 /* 2094 * Fake a blocking ast for the downconvert code. 2095 */ 2096 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED); 2097 lockres->l_blocking = DLM_LOCK_EX; 2098 2099 gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL); 2100 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0); 2101 spin_unlock_irqrestore(&lockres->l_lock, flags); 2102 2103 ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen); 2104 if (ret) { 2105 mlog_errno(ret); 2106 return; 2107 } 2108 2109 ret = ocfs2_wait_for_mask(&mw); 2110 if (ret) 2111 mlog_errno(ret); 2112} 2113 2114static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb, 2115 struct ocfs2_lock_res *lockres) 2116{ 2117 int kick = 0; 2118 2119 /* If we know that another node is waiting on our lock, kick 2120 * the downconvert thread * pre-emptively when we reach a release 2121 * condition. */ 2122 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) { 2123 switch(lockres->l_blocking) { 2124 case DLM_LOCK_EX: 2125 if (!lockres->l_ex_holders && !lockres->l_ro_holders) 2126 kick = 1; 2127 break; 2128 case DLM_LOCK_PR: 2129 if (!lockres->l_ex_holders) 2130 kick = 1; 2131 break; 2132 default: 2133 BUG(); 2134 } 2135 } 2136 2137 if (kick) 2138 ocfs2_wake_downconvert_thread(osb); 2139} 2140 2141#define OCFS2_SEC_BITS 34 2142#define OCFS2_SEC_SHIFT (64 - OCFS2_SEC_BITS) 2143#define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1) 2144 2145/* LVB only has room for 64 bits of time here so we pack it for 2146 * now. */ 2147static u64 ocfs2_pack_timespec(struct timespec64 *spec) 2148{ 2149 u64 res; 2150 u64 sec = clamp_t(time64_t, spec->tv_sec, 0, 0x3ffffffffull); 2151 u32 nsec = spec->tv_nsec; 2152 2153 res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK); 2154 2155 return res; 2156} 2157 2158/* Call this with the lockres locked. I am reasonably sure we don't 2159 * need ip_lock in this function as anyone who would be changing those 2160 * values is supposed to be blocked in ocfs2_inode_lock right now. */ 2161static void __ocfs2_stuff_meta_lvb(struct inode *inode) 2162{ 2163 struct ocfs2_inode_info *oi = OCFS2_I(inode); 2164 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres; 2165 struct ocfs2_meta_lvb *lvb; 2166 struct timespec64 ts; 2167 2168 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 2169 2170 /* 2171 * Invalidate the LVB of a deleted inode - this way other 2172 * nodes are forced to go to disk and discover the new inode 2173 * status. 2174 */ 2175 if (oi->ip_flags & OCFS2_INODE_DELETED) { 2176 lvb->lvb_version = 0; 2177 goto out; 2178 } 2179 2180 lvb->lvb_version = OCFS2_LVB_VERSION; 2181 lvb->lvb_isize = cpu_to_be64(i_size_read(inode)); 2182 lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters); 2183 lvb->lvb_iuid = cpu_to_be32(i_uid_read(inode)); 2184 lvb->lvb_igid = cpu_to_be32(i_gid_read(inode)); 2185 lvb->lvb_imode = cpu_to_be16(inode->i_mode); 2186 lvb->lvb_inlink = cpu_to_be16(inode->i_nlink); 2187 ts = inode_get_atime(inode); 2188 lvb->lvb_iatime_packed = cpu_to_be64(ocfs2_pack_timespec(&ts)); 2189 ts = inode_get_ctime(inode); 2190 lvb->lvb_ictime_packed = cpu_to_be64(ocfs2_pack_timespec(&ts)); 2191 ts = inode_get_mtime(inode); 2192 lvb->lvb_imtime_packed = cpu_to_be64(ocfs2_pack_timespec(&ts)); 2193 lvb->lvb_iattr = cpu_to_be32(oi->ip_attr); 2194 lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features); 2195 lvb->lvb_igeneration = cpu_to_be32(inode->i_generation); 2196 2197out: 2198 mlog_meta_lvb(0, lockres); 2199} 2200 2201static void ocfs2_unpack_timespec(struct timespec64 *spec, 2202 u64 packed_time) 2203{ 2204 spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT; 2205 spec->tv_nsec = packed_time & OCFS2_NSEC_MASK; 2206} 2207 2208static int ocfs2_refresh_inode_from_lvb(struct inode *inode) 2209{ 2210 struct ocfs2_inode_info *oi = OCFS2_I(inode); 2211 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres; 2212 struct ocfs2_meta_lvb *lvb; 2213 struct timespec64 ts; 2214 2215 mlog_meta_lvb(0, lockres); 2216 2217 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 2218 if (inode_wrong_type(inode, be16_to_cpu(lvb->lvb_imode))) 2219 return -ESTALE; 2220 2221 /* We're safe here without the lockres lock... */ 2222 spin_lock(&oi->ip_lock); 2223 oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters); 2224 i_size_write(inode, be64_to_cpu(lvb->lvb_isize)); 2225 2226 oi->ip_attr = be32_to_cpu(lvb->lvb_iattr); 2227 oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures); 2228 ocfs2_set_inode_flags(inode); 2229 2230 /* fast-symlinks are a special case */ 2231 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters) 2232 inode->i_blocks = 0; 2233 else 2234 inode->i_blocks = ocfs2_inode_sector_count(inode); 2235 2236 i_uid_write(inode, be32_to_cpu(lvb->lvb_iuid)); 2237 i_gid_write(inode, be32_to_cpu(lvb->lvb_igid)); 2238 inode->i_mode = be16_to_cpu(lvb->lvb_imode); 2239 set_nlink(inode, be16_to_cpu(lvb->lvb_inlink)); 2240 ocfs2_unpack_timespec(&ts, be64_to_cpu(lvb->lvb_iatime_packed)); 2241 inode_set_atime_to_ts(inode, ts); 2242 ocfs2_unpack_timespec(&ts, be64_to_cpu(lvb->lvb_imtime_packed)); 2243 inode_set_mtime_to_ts(inode, ts); 2244 ocfs2_unpack_timespec(&ts, be64_to_cpu(lvb->lvb_ictime_packed)); 2245 inode_set_ctime_to_ts(inode, ts); 2246 spin_unlock(&oi->ip_lock); 2247 return 0; 2248} 2249 2250static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode, 2251 struct ocfs2_lock_res *lockres) 2252{ 2253 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 2254 2255 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) 2256 && lvb->lvb_version == OCFS2_LVB_VERSION 2257 && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation) 2258 return 1; 2259 return 0; 2260} 2261 2262/* Determine whether a lock resource needs to be refreshed, and 2263 * arbitrate who gets to refresh it. 2264 * 2265 * 0 means no refresh needed. 2266 * 2267 * > 0 means you need to refresh this and you MUST call 2268 * ocfs2_complete_lock_res_refresh afterwards. */ 2269static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres) 2270{ 2271 unsigned long flags; 2272 int status = 0; 2273 2274refresh_check: 2275 spin_lock_irqsave(&lockres->l_lock, flags); 2276 if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) { 2277 spin_unlock_irqrestore(&lockres->l_lock, flags); 2278 goto bail; 2279 } 2280 2281 if (lockres->l_flags & OCFS2_LOCK_REFRESHING) { 2282 spin_unlock_irqrestore(&lockres->l_lock, flags); 2283 2284 ocfs2_wait_on_refreshing_lock(lockres); 2285 goto refresh_check; 2286 } 2287 2288 /* Ok, I'll be the one to refresh this lock. */ 2289 lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING); 2290 spin_unlock_irqrestore(&lockres->l_lock, flags); 2291 2292 status = 1; 2293bail: 2294 mlog(0, "status %d\n", status); 2295 return status; 2296} 2297 2298/* If status is non zero, I'll mark it as not being in refresh 2299 * anymroe, but i won't clear the needs refresh flag. */ 2300static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres, 2301 int status) 2302{ 2303 unsigned long flags; 2304 2305 spin_lock_irqsave(&lockres->l_lock, flags); 2306 lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING); 2307 if (!status) 2308 lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH); 2309 spin_unlock_irqrestore(&lockres->l_lock, flags); 2310 2311 wake_up(&lockres->l_event); 2312} 2313 2314/* may or may not return a bh if it went to disk. */ 2315static int ocfs2_inode_lock_update(struct inode *inode, 2316 struct buffer_head **bh) 2317{ 2318 int status = 0; 2319 struct ocfs2_inode_info *oi = OCFS2_I(inode); 2320 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres; 2321 struct ocfs2_dinode *fe; 2322 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 2323 2324 if (ocfs2_mount_local(osb)) 2325 goto bail; 2326 2327 spin_lock(&oi->ip_lock); 2328 if (oi->ip_flags & OCFS2_INODE_DELETED) { 2329 mlog(0, "Orphaned inode %llu was deleted while we " 2330 "were waiting on a lock. ip_flags = 0x%x\n", 2331 (unsigned long long)oi->ip_blkno, oi->ip_flags); 2332 spin_unlock(&oi->ip_lock); 2333 status = -ENOENT; 2334 goto bail; 2335 } 2336 spin_unlock(&oi->ip_lock); 2337 2338 if (!ocfs2_should_refresh_lock_res(lockres)) 2339 goto bail; 2340 2341 /* This will discard any caching information we might have had 2342 * for the inode metadata. */ 2343 ocfs2_metadata_cache_purge(INODE_CACHE(inode)); 2344 2345 ocfs2_extent_map_trunc(inode, 0); 2346 2347 if (ocfs2_meta_lvb_is_trustable(inode, lockres)) { 2348 mlog(0, "Trusting LVB on inode %llu\n", 2349 (unsigned long long)oi->ip_blkno); 2350 status = ocfs2_refresh_inode_from_lvb(inode); 2351 goto bail_refresh; 2352 } else { 2353 /* Boo, we have to go to disk. */ 2354 /* read bh, cast, ocfs2_refresh_inode */ 2355 status = ocfs2_read_inode_block(inode, bh); 2356 if (status < 0) { 2357 mlog_errno(status); 2358 goto bail_refresh; 2359 } 2360 fe = (struct ocfs2_dinode *) (*bh)->b_data; 2361 if (inode_wrong_type(inode, le16_to_cpu(fe->i_mode))) { 2362 status = -ESTALE; 2363 goto bail_refresh; 2364 } 2365 2366 /* This is a good chance to make sure we're not 2367 * locking an invalid object. ocfs2_read_inode_block() 2368 * already checked that the inode block is sane. 2369 * 2370 * We bug on a stale inode here because we checked 2371 * above whether it was wiped from disk. The wiping 2372 * node provides a guarantee that we receive that 2373 * message and can mark the inode before dropping any 2374 * locks associated with it. */ 2375 mlog_bug_on_msg(inode->i_generation != 2376 le32_to_cpu(fe->i_generation), 2377 "Invalid dinode %llu disk generation: %u " 2378 "inode->i_generation: %u\n", 2379 (unsigned long long)oi->ip_blkno, 2380 le32_to_cpu(fe->i_generation), 2381 inode->i_generation); 2382 mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) || 2383 !(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)), 2384 "Stale dinode %llu dtime: %llu flags: 0x%x\n", 2385 (unsigned long long)oi->ip_blkno, 2386 (unsigned long long)le64_to_cpu(fe->i_dtime), 2387 le32_to_cpu(fe->i_flags)); 2388 2389 ocfs2_refresh_inode(inode, fe); 2390 ocfs2_track_lock_refresh(lockres); 2391 } 2392 2393 status = 0; 2394bail_refresh: 2395 ocfs2_complete_lock_res_refresh(lockres, status); 2396bail: 2397 return status; 2398} 2399 2400static int ocfs2_assign_bh(struct inode *inode, 2401 struct buffer_head **ret_bh, 2402 struct buffer_head *passed_bh) 2403{ 2404 int status; 2405 2406 if (passed_bh) { 2407 /* Ok, the update went to disk for us, use the 2408 * returned bh. */ 2409 *ret_bh = passed_bh; 2410 get_bh(*ret_bh); 2411 2412 return 0; 2413 } 2414 2415 status = ocfs2_read_inode_block(inode, ret_bh); 2416 if (status < 0) 2417 mlog_errno(status); 2418 2419 return status; 2420} 2421 2422/* 2423 * returns < 0 error if the callback will never be called, otherwise 2424 * the result of the lock will be communicated via the callback. 2425 */ 2426int ocfs2_inode_lock_full_nested(struct inode *inode, 2427 struct buffer_head **ret_bh, 2428 int ex, 2429 int arg_flags, 2430 int subclass) 2431{ 2432 int status, level, acquired; 2433 u32 dlm_flags; 2434 struct ocfs2_lock_res *lockres = NULL; 2435 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 2436 struct buffer_head *local_bh = NULL; 2437 2438 mlog(0, "inode %llu, take %s META lock\n", 2439 (unsigned long long)OCFS2_I(inode)->ip_blkno, 2440 ex ? "EXMODE" : "PRMODE"); 2441 2442 status = 0; 2443 acquired = 0; 2444 /* We'll allow faking a readonly metadata lock for 2445 * rodevices. */ 2446 if (ocfs2_is_hard_readonly(osb)) { 2447 if (ex) 2448 status = -EROFS; 2449 goto getbh; 2450 } 2451 2452 if ((arg_flags & OCFS2_META_LOCK_GETBH) || 2453 ocfs2_mount_local(osb)) 2454 goto update; 2455 2456 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY)) 2457 ocfs2_wait_for_recovery(osb); 2458 2459 lockres = &OCFS2_I(inode)->ip_inode_lockres; 2460 level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 2461 dlm_flags = 0; 2462 if (arg_flags & OCFS2_META_LOCK_NOQUEUE) 2463 dlm_flags |= DLM_LKF_NOQUEUE; 2464 2465 status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags, 2466 arg_flags, subclass, _RET_IP_); 2467 if (status < 0) { 2468 if (status != -EAGAIN) 2469 mlog_errno(status); 2470 goto bail; 2471 } 2472 2473 /* Notify the error cleanup path to drop the cluster lock. */ 2474 acquired = 1; 2475 2476 /* We wait twice because a node may have died while we were in 2477 * the lower dlm layers. The second time though, we've 2478 * committed to owning this lock so we don't allow signals to 2479 * abort the operation. */ 2480 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY)) 2481 ocfs2_wait_for_recovery(osb); 2482 2483update: 2484 /* 2485 * We only see this flag if we're being called from 2486 * ocfs2_read_locked_inode(). It means we're locking an inode 2487 * which hasn't been populated yet, so clear the refresh flag 2488 * and let the caller handle it. 2489 */ 2490 if (inode_state_read_once(inode) & I_NEW) { 2491 status = 0; 2492 if (lockres) 2493 ocfs2_complete_lock_res_refresh(lockres, 0); 2494 goto bail; 2495 } 2496 2497 /* This is fun. The caller may want a bh back, or it may 2498 * not. ocfs2_inode_lock_update definitely wants one in, but 2499 * may or may not read one, depending on what's in the 2500 * LVB. The result of all of this is that we've *only* gone to 2501 * disk if we have to, so the complexity is worthwhile. */ 2502 status = ocfs2_inode_lock_update(inode, &local_bh); 2503 if (status < 0) { 2504 if (status != -ENOENT) 2505 mlog_errno(status); 2506 goto bail; 2507 } 2508getbh: 2509 if (ret_bh) { 2510 status = ocfs2_assign_bh(inode, ret_bh, local_bh); 2511 if (status < 0) { 2512 mlog_errno(status); 2513 goto bail; 2514 } 2515 } 2516 2517bail: 2518 if (status < 0) { 2519 if (ret_bh && (*ret_bh)) { 2520 brelse(*ret_bh); 2521 *ret_bh = NULL; 2522 } 2523 if (acquired) 2524 ocfs2_inode_unlock(inode, ex); 2525 } 2526 2527 brelse(local_bh); 2528 return status; 2529} 2530 2531/* 2532 * This is working around a lock inversion between tasks acquiring DLM 2533 * locks while holding a folio lock and the downconvert thread which 2534 * blocks dlm lock acquiry while acquiring folio locks. 2535 * 2536 * ** These _with_folio variants are only intended to be called from aop 2537 * methods that hold folio locks and return a very specific *positive* error 2538 * code that aop methods pass up to the VFS -- test for errors with != 0. ** 2539 * 2540 * The DLM is called such that it returns -EAGAIN if it would have 2541 * blocked waiting for the downconvert thread. In that case we unlock 2542 * our folio so the downconvert thread can make progress. Once we've 2543 * done this we have to return AOP_TRUNCATED_PAGE so the aop method 2544 * that called us can bubble that back up into the VFS who will then 2545 * immediately retry the aop call. 2546 */ 2547int ocfs2_inode_lock_with_folio(struct inode *inode, 2548 struct buffer_head **ret_bh, int ex, struct folio *folio) 2549{ 2550 int ret; 2551 2552 ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK); 2553 if (ret == -EAGAIN) { 2554 folio_unlock(folio); 2555 /* 2556 * If we can't get inode lock immediately, we should not return 2557 * directly here, since this will lead to a softlockup problem. 2558 * The method is to get a blocking lock and immediately unlock 2559 * before returning, this can avoid CPU resource waste due to 2560 * lots of retries, and benefits fairness in getting lock. 2561 */ 2562 if (ocfs2_inode_lock(inode, ret_bh, ex) == 0) 2563 ocfs2_inode_unlock(inode, ex); 2564 ret = AOP_TRUNCATED_PAGE; 2565 } 2566 2567 return ret; 2568} 2569 2570int ocfs2_inode_lock_atime(struct inode *inode, 2571 struct vfsmount *vfsmnt, 2572 int *level, int wait) 2573{ 2574 int ret; 2575 2576 if (wait) 2577 ret = ocfs2_inode_lock(inode, NULL, 0); 2578 else 2579 ret = ocfs2_try_inode_lock(inode, NULL, 0); 2580 2581 if (ret < 0) { 2582 if (ret != -EAGAIN) 2583 mlog_errno(ret); 2584 return ret; 2585 } 2586 2587 /* 2588 * If we should update atime, we will get EX lock, 2589 * otherwise we just get PR lock. 2590 */ 2591 if (ocfs2_should_update_atime(inode, vfsmnt)) { 2592 struct buffer_head *bh = NULL; 2593 2594 ocfs2_inode_unlock(inode, 0); 2595 if (wait) 2596 ret = ocfs2_inode_lock(inode, &bh, 1); 2597 else 2598 ret = ocfs2_try_inode_lock(inode, &bh, 1); 2599 2600 if (ret < 0) { 2601 if (ret != -EAGAIN) 2602 mlog_errno(ret); 2603 return ret; 2604 } 2605 *level = 1; 2606 if (ocfs2_should_update_atime(inode, vfsmnt)) 2607 ocfs2_update_inode_atime(inode, bh); 2608 brelse(bh); 2609 } else 2610 *level = 0; 2611 2612 return ret; 2613} 2614 2615void ocfs2_inode_unlock(struct inode *inode, 2616 int ex) 2617{ 2618 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 2619 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres; 2620 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 2621 2622 mlog(0, "inode %llu drop %s META lock\n", 2623 (unsigned long long)OCFS2_I(inode)->ip_blkno, 2624 ex ? "EXMODE" : "PRMODE"); 2625 2626 if (!ocfs2_is_hard_readonly(osb) && 2627 !ocfs2_mount_local(osb)) 2628 ocfs2_cluster_unlock(osb, lockres, level); 2629} 2630 2631/* 2632 * This _tracker variants are introduced to deal with the recursive cluster 2633 * locking issue. The idea is to keep track of a lock holder on the stack of 2634 * the current process. If there's a lock holder on the stack, we know the 2635 * task context is already protected by cluster locking. Currently, they're 2636 * used in some VFS entry routines. 2637 * 2638 * return < 0 on error, return == 0 if there's no lock holder on the stack 2639 * before this call, return == 1 if this call would be a recursive locking. 2640 * return == -1 if this lock attempt will cause an upgrade which is forbidden. 2641 * 2642 * When taking lock levels into account,we face some different situations. 2643 * 2644 * 1. no lock is held 2645 * In this case, just lock the inode as requested and return 0 2646 * 2647 * 2. We are holding a lock 2648 * For this situation, things diverges into several cases 2649 * 2650 * wanted holding what to do 2651 * ex ex see 2.1 below 2652 * ex pr see 2.2 below 2653 * pr ex see 2.1 below 2654 * pr pr see 2.1 below 2655 * 2656 * 2.1 lock level that is been held is compatible 2657 * with the wanted level, so no lock action will be tacken. 2658 * 2659 * 2.2 Otherwise, an upgrade is needed, but it is forbidden. 2660 * 2661 * Reason why upgrade within a process is forbidden is that 2662 * lock upgrade may cause dead lock. The following illustrates 2663 * how it happens. 2664 * 2665 * thread on node1 thread on node2 2666 * ocfs2_inode_lock_tracker(ex=0) 2667 * 2668 * <====== ocfs2_inode_lock_tracker(ex=1) 2669 * 2670 * ocfs2_inode_lock_tracker(ex=1) 2671 */ 2672int ocfs2_inode_lock_tracker(struct inode *inode, 2673 struct buffer_head **ret_bh, 2674 int ex, 2675 struct ocfs2_lock_holder *oh) 2676{ 2677 int status = 0; 2678 struct ocfs2_lock_res *lockres; 2679 struct ocfs2_lock_holder *tmp_oh; 2680 struct pid *pid = task_pid(current); 2681 2682 2683 lockres = &OCFS2_I(inode)->ip_inode_lockres; 2684 tmp_oh = ocfs2_pid_holder(lockres, pid); 2685 2686 if (!tmp_oh) { 2687 /* 2688 * This corresponds to the case 1. 2689 * We haven't got any lock before. 2690 */ 2691 status = ocfs2_inode_lock_full(inode, ret_bh, ex, 0); 2692 if (status < 0) { 2693 if (status != -ENOENT) 2694 mlog_errno(status); 2695 return status; 2696 } 2697 2698 oh->oh_ex = ex; 2699 ocfs2_add_holder(lockres, oh); 2700 return 0; 2701 } 2702 2703 if (unlikely(ex && !tmp_oh->oh_ex)) { 2704 /* 2705 * case 2.2 upgrade may cause dead lock, forbid it. 2706 */ 2707 mlog(ML_ERROR, "Recursive locking is not permitted to " 2708 "upgrade to EX level from PR level.\n"); 2709 dump_stack(); 2710 return -EINVAL; 2711 } 2712 2713 /* 2714 * case 2.1 OCFS2_META_LOCK_GETBH flag make ocfs2_inode_lock_full. 2715 * ignore the lock level and just update it. 2716 */ 2717 if (ret_bh) { 2718 status = ocfs2_inode_lock_full(inode, ret_bh, ex, 2719 OCFS2_META_LOCK_GETBH); 2720 if (status < 0) { 2721 if (status != -ENOENT) 2722 mlog_errno(status); 2723 return status; 2724 } 2725 } 2726 return 1; 2727} 2728 2729void ocfs2_inode_unlock_tracker(struct inode *inode, 2730 int ex, 2731 struct ocfs2_lock_holder *oh, 2732 int had_lock) 2733{ 2734 struct ocfs2_lock_res *lockres; 2735 2736 lockres = &OCFS2_I(inode)->ip_inode_lockres; 2737 /* had_lock means that the current process already takes the cluster 2738 * lock previously. 2739 * If had_lock is 1, we have nothing to do here. 2740 * If had_lock is 0, we will release the lock. 2741 */ 2742 if (!had_lock) { 2743 ocfs2_inode_unlock(inode, oh->oh_ex); 2744 ocfs2_remove_holder(lockres, oh); 2745 } 2746} 2747 2748int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno) 2749{ 2750 struct ocfs2_lock_res *lockres; 2751 struct ocfs2_orphan_scan_lvb *lvb; 2752 int status = 0; 2753 2754 if (ocfs2_is_hard_readonly(osb)) 2755 return -EROFS; 2756 2757 if (ocfs2_mount_local(osb)) 2758 return 0; 2759 2760 lockres = &osb->osb_orphan_scan.os_lockres; 2761 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0); 2762 if (status < 0) 2763 return status; 2764 2765 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 2766 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) && 2767 lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION) 2768 *seqno = be32_to_cpu(lvb->lvb_os_seqno); 2769 else 2770 *seqno = osb->osb_orphan_scan.os_seqno + 1; 2771 2772 return status; 2773} 2774 2775void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno) 2776{ 2777 struct ocfs2_lock_res *lockres; 2778 struct ocfs2_orphan_scan_lvb *lvb; 2779 2780 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) { 2781 lockres = &osb->osb_orphan_scan.os_lockres; 2782 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 2783 lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION; 2784 lvb->lvb_os_seqno = cpu_to_be32(seqno); 2785 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX); 2786 } 2787} 2788 2789int ocfs2_super_lock(struct ocfs2_super *osb, 2790 int ex) 2791{ 2792 int status = 0; 2793 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 2794 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres; 2795 2796 if (ocfs2_is_hard_readonly(osb)) 2797 return -EROFS; 2798 2799 if (ocfs2_mount_local(osb)) 2800 goto bail; 2801 2802 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0); 2803 if (status < 0) { 2804 mlog_errno(status); 2805 goto bail; 2806 } 2807 2808 /* The super block lock path is really in the best position to 2809 * know when resources covered by the lock need to be 2810 * refreshed, so we do it here. Of course, making sense of 2811 * everything is up to the caller :) */ 2812 status = ocfs2_should_refresh_lock_res(lockres); 2813 if (status) { 2814 status = ocfs2_refresh_slot_info(osb); 2815 2816 ocfs2_complete_lock_res_refresh(lockres, status); 2817 2818 if (status < 0) { 2819 ocfs2_cluster_unlock(osb, lockres, level); 2820 mlog_errno(status); 2821 } 2822 ocfs2_track_lock_refresh(lockres); 2823 } 2824bail: 2825 return status; 2826} 2827 2828void ocfs2_super_unlock(struct ocfs2_super *osb, 2829 int ex) 2830{ 2831 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 2832 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres; 2833 2834 if (!ocfs2_mount_local(osb)) 2835 ocfs2_cluster_unlock(osb, lockres, level); 2836} 2837 2838int ocfs2_rename_lock(struct ocfs2_super *osb) 2839{ 2840 int status; 2841 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres; 2842 2843 if (ocfs2_is_hard_readonly(osb)) 2844 return -EROFS; 2845 2846 if (ocfs2_mount_local(osb)) 2847 return 0; 2848 2849 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0); 2850 if (status < 0) 2851 mlog_errno(status); 2852 2853 return status; 2854} 2855 2856void ocfs2_rename_unlock(struct ocfs2_super *osb) 2857{ 2858 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres; 2859 2860 if (!ocfs2_mount_local(osb)) 2861 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX); 2862} 2863 2864int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex) 2865{ 2866 int status; 2867 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres; 2868 2869 if (ocfs2_is_hard_readonly(osb)) 2870 return -EROFS; 2871 2872 if (ex) 2873 down_write(&osb->nfs_sync_rwlock); 2874 else 2875 down_read(&osb->nfs_sync_rwlock); 2876 2877 if (ocfs2_mount_local(osb)) 2878 return 0; 2879 2880 status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE, 2881 0, 0); 2882 if (status < 0) { 2883 mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status); 2884 2885 if (ex) 2886 up_write(&osb->nfs_sync_rwlock); 2887 else 2888 up_read(&osb->nfs_sync_rwlock); 2889 } 2890 2891 return status; 2892} 2893 2894void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex) 2895{ 2896 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres; 2897 2898 if (!ocfs2_mount_local(osb)) 2899 ocfs2_cluster_unlock(osb, lockres, 2900 ex ? LKM_EXMODE : LKM_PRMODE); 2901 if (ex) 2902 up_write(&osb->nfs_sync_rwlock); 2903 else 2904 up_read(&osb->nfs_sync_rwlock); 2905} 2906 2907int ocfs2_trim_fs_lock(struct ocfs2_super *osb, 2908 struct ocfs2_trim_fs_info *info, int trylock) 2909{ 2910 int status; 2911 struct ocfs2_trim_fs_lvb *lvb; 2912 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres; 2913 2914 if (info) 2915 info->tf_valid = 0; 2916 2917 if (ocfs2_is_hard_readonly(osb)) 2918 return -EROFS; 2919 2920 if (ocfs2_mount_local(osb)) 2921 return 0; 2922 2923 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 2924 trylock ? DLM_LKF_NOQUEUE : 0, 0); 2925 if (status < 0) { 2926 if (status != -EAGAIN) 2927 mlog_errno(status); 2928 return status; 2929 } 2930 2931 if (info) { 2932 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 2933 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) && 2934 lvb->lvb_version == OCFS2_TRIMFS_LVB_VERSION) { 2935 info->tf_valid = 1; 2936 info->tf_success = lvb->lvb_success; 2937 info->tf_nodenum = be32_to_cpu(lvb->lvb_nodenum); 2938 info->tf_start = be64_to_cpu(lvb->lvb_start); 2939 info->tf_len = be64_to_cpu(lvb->lvb_len); 2940 info->tf_minlen = be64_to_cpu(lvb->lvb_minlen); 2941 info->tf_trimlen = be64_to_cpu(lvb->lvb_trimlen); 2942 } 2943 } 2944 2945 return status; 2946} 2947 2948void ocfs2_trim_fs_unlock(struct ocfs2_super *osb, 2949 struct ocfs2_trim_fs_info *info) 2950{ 2951 struct ocfs2_trim_fs_lvb *lvb; 2952 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres; 2953 2954 if (ocfs2_mount_local(osb)) 2955 return; 2956 2957 if (info) { 2958 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 2959 lvb->lvb_version = OCFS2_TRIMFS_LVB_VERSION; 2960 lvb->lvb_success = info->tf_success; 2961 lvb->lvb_nodenum = cpu_to_be32(info->tf_nodenum); 2962 lvb->lvb_start = cpu_to_be64(info->tf_start); 2963 lvb->lvb_len = cpu_to_be64(info->tf_len); 2964 lvb->lvb_minlen = cpu_to_be64(info->tf_minlen); 2965 lvb->lvb_trimlen = cpu_to_be64(info->tf_trimlen); 2966 } 2967 2968 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX); 2969} 2970 2971int ocfs2_dentry_lock(struct dentry *dentry, int ex) 2972{ 2973 int ret; 2974 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 2975 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 2976 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb); 2977 2978 BUG_ON(!dl); 2979 2980 if (ocfs2_is_hard_readonly(osb)) { 2981 if (ex) 2982 return -EROFS; 2983 return 0; 2984 } 2985 2986 if (ocfs2_mount_local(osb)) 2987 return 0; 2988 2989 ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0); 2990 if (ret < 0) 2991 mlog_errno(ret); 2992 2993 return ret; 2994} 2995 2996void ocfs2_dentry_unlock(struct dentry *dentry, int ex) 2997{ 2998 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 2999 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 3000 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb); 3001 3002 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) 3003 ocfs2_cluster_unlock(osb, &dl->dl_lockres, level); 3004} 3005 3006/* Reference counting of the dlm debug structure. We want this because 3007 * open references on the debug inodes can live on after a mount, so 3008 * we can't rely on the ocfs2_super to always exist. */ 3009static void ocfs2_dlm_debug_free(struct kref *kref) 3010{ 3011 struct ocfs2_dlm_debug *dlm_debug; 3012 3013 dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt); 3014 3015 kfree(dlm_debug); 3016} 3017 3018void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug) 3019{ 3020 if (dlm_debug) 3021 kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free); 3022} 3023 3024static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug) 3025{ 3026 kref_get(&debug->d_refcnt); 3027} 3028 3029struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void) 3030{ 3031 struct ocfs2_dlm_debug *dlm_debug; 3032 3033 dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL); 3034 if (!dlm_debug) { 3035 mlog_errno(-ENOMEM); 3036 goto out; 3037 } 3038 3039 kref_init(&dlm_debug->d_refcnt); 3040 INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking); 3041 dlm_debug->d_filter_secs = 0; 3042out: 3043 return dlm_debug; 3044} 3045 3046/* Access to this is arbitrated for us via seq_file->sem. */ 3047struct ocfs2_dlm_seq_priv { 3048 struct ocfs2_dlm_debug *p_dlm_debug; 3049 struct ocfs2_lock_res p_iter_res; 3050 struct ocfs2_lock_res p_tmp_res; 3051}; 3052 3053static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start, 3054 struct ocfs2_dlm_seq_priv *priv) 3055{ 3056 struct ocfs2_lock_res *iter, *ret = NULL; 3057 struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug; 3058 3059 assert_spin_locked(&ocfs2_dlm_tracking_lock); 3060 3061 list_for_each_entry(iter, &start->l_debug_list, l_debug_list) { 3062 /* discover the head of the list */ 3063 if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) { 3064 mlog(0, "End of list found, %p\n", ret); 3065 break; 3066 } 3067 3068 /* We track our "dummy" iteration lockres' by a NULL 3069 * l_ops field. */ 3070 if (iter->l_ops != NULL) { 3071 ret = iter; 3072 break; 3073 } 3074 } 3075 3076 return ret; 3077} 3078 3079static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos) 3080{ 3081 struct ocfs2_dlm_seq_priv *priv = m->private; 3082 struct ocfs2_lock_res *iter; 3083 3084 spin_lock(&ocfs2_dlm_tracking_lock); 3085 iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv); 3086 if (iter) { 3087 /* Since lockres' have the lifetime of their container 3088 * (which can be inodes, ocfs2_supers, etc) we want to 3089 * copy this out to a temporary lockres while still 3090 * under the spinlock. Obviously after this we can't 3091 * trust any pointers on the copy returned, but that's 3092 * ok as the information we want isn't typically held 3093 * in them. */ 3094 priv->p_tmp_res = *iter; 3095 iter = &priv->p_tmp_res; 3096 } 3097 spin_unlock(&ocfs2_dlm_tracking_lock); 3098 3099 return iter; 3100} 3101 3102static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v) 3103{ 3104} 3105 3106static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos) 3107{ 3108 struct ocfs2_dlm_seq_priv *priv = m->private; 3109 struct ocfs2_lock_res *iter = v; 3110 struct ocfs2_lock_res *dummy = &priv->p_iter_res; 3111 3112 (*pos)++; 3113 spin_lock(&ocfs2_dlm_tracking_lock); 3114 iter = ocfs2_dlm_next_res(iter, priv); 3115 list_del_init(&dummy->l_debug_list); 3116 if (iter) { 3117 list_add(&dummy->l_debug_list, &iter->l_debug_list); 3118 priv->p_tmp_res = *iter; 3119 iter = &priv->p_tmp_res; 3120 } 3121 spin_unlock(&ocfs2_dlm_tracking_lock); 3122 3123 return iter; 3124} 3125 3126/* 3127 * Version is used by debugfs.ocfs2 to determine the format being used 3128 * 3129 * New in version 2 3130 * - Lock stats printed 3131 * New in version 3 3132 * - Max time in lock stats is in usecs (instead of nsecs) 3133 * New in version 4 3134 * - Add last pr/ex unlock times and first lock wait time in usecs 3135 */ 3136#define OCFS2_DLM_DEBUG_STR_VERSION 4 3137static int ocfs2_dlm_seq_show(struct seq_file *m, void *v) 3138{ 3139 int i; 3140 char *lvb; 3141 struct ocfs2_lock_res *lockres = v; 3142#ifdef CONFIG_OCFS2_FS_STATS 3143 u64 now, last; 3144 struct ocfs2_dlm_debug *dlm_debug = 3145 ((struct ocfs2_dlm_seq_priv *)m->private)->p_dlm_debug; 3146#endif 3147 3148 if (!lockres) 3149 return -EINVAL; 3150 3151#ifdef CONFIG_OCFS2_FS_STATS 3152 if (!lockres->l_lock_wait && dlm_debug->d_filter_secs) { 3153 now = ktime_to_us(ktime_get_real()); 3154 last = max(lockres->l_lock_prmode.ls_last, 3155 lockres->l_lock_exmode.ls_last); 3156 /* 3157 * Use d_filter_secs field to filter lock resources dump, 3158 * the default d_filter_secs(0) value filters nothing, 3159 * otherwise, only dump the last N seconds active lock 3160 * resources. 3161 */ 3162 if (div_u64(now - last, 1000000) > dlm_debug->d_filter_secs) 3163 return 0; 3164 } 3165#endif 3166 3167 seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION); 3168 3169 if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY) 3170 seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1, 3171 lockres->l_name, 3172 (unsigned int)ocfs2_get_dentry_lock_ino(lockres)); 3173 else 3174 seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name); 3175 3176 seq_printf(m, "%d\t" 3177 "0x%lx\t" 3178 "0x%x\t" 3179 "0x%x\t" 3180 "%u\t" 3181 "%u\t" 3182 "%d\t" 3183 "%d\t", 3184 lockres->l_level, 3185 lockres->l_flags, 3186 lockres->l_action, 3187 lockres->l_unlock_action, 3188 lockres->l_ro_holders, 3189 lockres->l_ex_holders, 3190 lockres->l_requested, 3191 lockres->l_blocking); 3192 3193 /* Dump the raw LVB */ 3194 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 3195 for(i = 0; i < DLM_LVB_LEN; i++) 3196 seq_printf(m, "0x%x\t", lvb[i]); 3197 3198#ifdef CONFIG_OCFS2_FS_STATS 3199# define lock_num_prmode(_l) ((_l)->l_lock_prmode.ls_gets) 3200# define lock_num_exmode(_l) ((_l)->l_lock_exmode.ls_gets) 3201# define lock_num_prmode_failed(_l) ((_l)->l_lock_prmode.ls_fail) 3202# define lock_num_exmode_failed(_l) ((_l)->l_lock_exmode.ls_fail) 3203# define lock_total_prmode(_l) ((_l)->l_lock_prmode.ls_total) 3204# define lock_total_exmode(_l) ((_l)->l_lock_exmode.ls_total) 3205# define lock_max_prmode(_l) ((_l)->l_lock_prmode.ls_max) 3206# define lock_max_exmode(_l) ((_l)->l_lock_exmode.ls_max) 3207# define lock_refresh(_l) ((_l)->l_lock_refresh) 3208# define lock_last_prmode(_l) ((_l)->l_lock_prmode.ls_last) 3209# define lock_last_exmode(_l) ((_l)->l_lock_exmode.ls_last) 3210# define lock_wait(_l) ((_l)->l_lock_wait) 3211#else 3212# define lock_num_prmode(_l) (0) 3213# define lock_num_exmode(_l) (0) 3214# define lock_num_prmode_failed(_l) (0) 3215# define lock_num_exmode_failed(_l) (0) 3216# define lock_total_prmode(_l) (0ULL) 3217# define lock_total_exmode(_l) (0ULL) 3218# define lock_max_prmode(_l) (0) 3219# define lock_max_exmode(_l) (0) 3220# define lock_refresh(_l) (0) 3221# define lock_last_prmode(_l) (0ULL) 3222# define lock_last_exmode(_l) (0ULL) 3223# define lock_wait(_l) (0ULL) 3224#endif 3225 /* The following seq_print was added in version 2 of this output */ 3226 seq_printf(m, "%u\t" 3227 "%u\t" 3228 "%u\t" 3229 "%u\t" 3230 "%llu\t" 3231 "%llu\t" 3232 "%u\t" 3233 "%u\t" 3234 "%u\t" 3235 "%llu\t" 3236 "%llu\t" 3237 "%llu\t", 3238 lock_num_prmode(lockres), 3239 lock_num_exmode(lockres), 3240 lock_num_prmode_failed(lockres), 3241 lock_num_exmode_failed(lockres), 3242 lock_total_prmode(lockres), 3243 lock_total_exmode(lockres), 3244 lock_max_prmode(lockres), 3245 lock_max_exmode(lockres), 3246 lock_refresh(lockres), 3247 lock_last_prmode(lockres), 3248 lock_last_exmode(lockres), 3249 lock_wait(lockres)); 3250 3251 /* End the line */ 3252 seq_printf(m, "\n"); 3253 return 0; 3254} 3255 3256static const struct seq_operations ocfs2_dlm_seq_ops = { 3257 .start = ocfs2_dlm_seq_start, 3258 .stop = ocfs2_dlm_seq_stop, 3259 .next = ocfs2_dlm_seq_next, 3260 .show = ocfs2_dlm_seq_show, 3261}; 3262 3263static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file) 3264{ 3265 struct seq_file *seq = file->private_data; 3266 struct ocfs2_dlm_seq_priv *priv = seq->private; 3267 struct ocfs2_lock_res *res = &priv->p_iter_res; 3268 3269 ocfs2_remove_lockres_tracking(res); 3270 ocfs2_put_dlm_debug(priv->p_dlm_debug); 3271 return seq_release_private(inode, file); 3272} 3273 3274static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file) 3275{ 3276 struct ocfs2_dlm_seq_priv *priv; 3277 struct ocfs2_super *osb; 3278 3279 priv = __seq_open_private(file, &ocfs2_dlm_seq_ops, sizeof(*priv)); 3280 if (!priv) { 3281 mlog_errno(-ENOMEM); 3282 return -ENOMEM; 3283 } 3284 3285 osb = inode->i_private; 3286 ocfs2_get_dlm_debug(osb->osb_dlm_debug); 3287 priv->p_dlm_debug = osb->osb_dlm_debug; 3288 INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list); 3289 3290 ocfs2_add_lockres_tracking(&priv->p_iter_res, 3291 priv->p_dlm_debug); 3292 3293 return 0; 3294} 3295 3296static const struct file_operations ocfs2_dlm_debug_fops = { 3297 .open = ocfs2_dlm_debug_open, 3298 .release = ocfs2_dlm_debug_release, 3299 .read = seq_read, 3300 .llseek = seq_lseek, 3301}; 3302 3303static void ocfs2_dlm_init_debug(struct ocfs2_super *osb) 3304{ 3305 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug; 3306 3307 debugfs_create_file("locking_state", S_IFREG|S_IRUSR, 3308 osb->osb_debug_root, osb, &ocfs2_dlm_debug_fops); 3309 3310 debugfs_create_u32("locking_filter", 0600, osb->osb_debug_root, 3311 &dlm_debug->d_filter_secs); 3312 ocfs2_get_dlm_debug(dlm_debug); 3313} 3314 3315static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb) 3316{ 3317 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug; 3318 3319 if (dlm_debug) 3320 ocfs2_put_dlm_debug(dlm_debug); 3321} 3322 3323int ocfs2_dlm_init(struct ocfs2_super *osb) 3324{ 3325 int status = 0; 3326 struct ocfs2_cluster_connection *conn = NULL; 3327 3328 if (ocfs2_mount_local(osb)) { 3329 osb->node_num = 0; 3330 goto local; 3331 } 3332 3333 ocfs2_dlm_init_debug(osb); 3334 3335 /* launch downconvert thread */ 3336 osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc-%s", 3337 osb->uuid_str); 3338 if (IS_ERR(osb->dc_task)) { 3339 status = PTR_ERR(osb->dc_task); 3340 osb->dc_task = NULL; 3341 mlog_errno(status); 3342 goto bail; 3343 } 3344 3345 /* for now, uuid == domain */ 3346 status = ocfs2_cluster_connect(osb->osb_cluster_stack, 3347 osb->osb_cluster_name, 3348 strlen(osb->osb_cluster_name), 3349 osb->uuid_str, 3350 strlen(osb->uuid_str), 3351 &lproto, ocfs2_do_node_down, osb, 3352 &conn); 3353 if (status) { 3354 mlog_errno(status); 3355 goto bail; 3356 } 3357 3358 status = ocfs2_cluster_this_node(conn, &osb->node_num); 3359 if (status < 0) { 3360 mlog_errno(status); 3361 mlog(ML_ERROR, 3362 "could not find this host's node number\n"); 3363 ocfs2_cluster_disconnect(conn, 0); 3364 goto bail; 3365 } 3366 3367local: 3368 ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb); 3369 ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb); 3370 ocfs2_nfs_sync_lock_init(osb); 3371 ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb); 3372 3373 osb->cconn = conn; 3374bail: 3375 if (status < 0) { 3376 ocfs2_dlm_shutdown_debug(osb); 3377 if (osb->dc_task) 3378 kthread_stop(osb->dc_task); 3379 } 3380 3381 return status; 3382} 3383 3384void ocfs2_dlm_shutdown(struct ocfs2_super *osb, 3385 int hangup_pending) 3386{ 3387 ocfs2_drop_osb_locks(osb); 3388 3389 /* 3390 * Now that we have dropped all locks and ocfs2_dismount_volume() 3391 * has disabled recovery, the DLM won't be talking to us. It's 3392 * safe to tear things down before disconnecting the cluster. 3393 */ 3394 3395 if (osb->dc_task) { 3396 kthread_stop(osb->dc_task); 3397 osb->dc_task = NULL; 3398 } 3399 3400 ocfs2_lock_res_free(&osb->osb_super_lockres); 3401 ocfs2_lock_res_free(&osb->osb_rename_lockres); 3402 ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres); 3403 ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres); 3404 3405 if (osb->cconn) { 3406 ocfs2_cluster_disconnect(osb->cconn, hangup_pending); 3407 osb->cconn = NULL; 3408 3409 ocfs2_dlm_shutdown_debug(osb); 3410 } 3411} 3412 3413static int ocfs2_drop_lock(struct ocfs2_super *osb, 3414 struct ocfs2_lock_res *lockres) 3415{ 3416 int ret; 3417 unsigned long flags; 3418 u32 lkm_flags = 0; 3419 3420 /* We didn't get anywhere near actually using this lockres. */ 3421 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED)) 3422 goto out; 3423 3424 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) 3425 lkm_flags |= DLM_LKF_VALBLK; 3426 3427 spin_lock_irqsave(&lockres->l_lock, flags); 3428 3429 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING), 3430 "lockres %s, flags 0x%lx\n", 3431 lockres->l_name, lockres->l_flags); 3432 3433 while (lockres->l_flags & OCFS2_LOCK_BUSY) { 3434 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = " 3435 "%u, unlock_action = %u\n", 3436 lockres->l_name, lockres->l_flags, lockres->l_action, 3437 lockres->l_unlock_action); 3438 3439 spin_unlock_irqrestore(&lockres->l_lock, flags); 3440 3441 /* XXX: Today we just wait on any busy 3442 * locks... Perhaps we need to cancel converts in the 3443 * future? */ 3444 ocfs2_wait_on_busy_lock(lockres); 3445 3446 spin_lock_irqsave(&lockres->l_lock, flags); 3447 } 3448 3449 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) { 3450 if (lockres->l_flags & OCFS2_LOCK_ATTACHED && 3451 lockres->l_level == DLM_LOCK_EX && 3452 !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) 3453 lockres->l_ops->set_lvb(lockres); 3454 } 3455 3456 if (lockres->l_flags & OCFS2_LOCK_BUSY) 3457 mlog(ML_ERROR, "destroying busy lock: \"%s\"\n", 3458 lockres->l_name); 3459 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) 3460 mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name); 3461 3462 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) { 3463 spin_unlock_irqrestore(&lockres->l_lock, flags); 3464 goto out; 3465 } 3466 3467 lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED); 3468 3469 /* make sure we never get here while waiting for an ast to 3470 * fire. */ 3471 BUG_ON(lockres->l_action != OCFS2_AST_INVALID); 3472 3473 /* is this necessary? */ 3474 lockres_or_flags(lockres, OCFS2_LOCK_BUSY); 3475 lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK; 3476 spin_unlock_irqrestore(&lockres->l_lock, flags); 3477 3478 mlog(0, "lock %s\n", lockres->l_name); 3479 3480 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags); 3481 if (ret) { 3482 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres); 3483 mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags); 3484 ocfs2_dlm_dump_lksb(&lockres->l_lksb); 3485 BUG(); 3486 } 3487 mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n", 3488 lockres->l_name); 3489 3490 ocfs2_wait_on_busy_lock(lockres); 3491out: 3492 return 0; 3493} 3494 3495static void ocfs2_process_blocked_lock(struct ocfs2_super *osb, 3496 struct ocfs2_lock_res *lockres); 3497 3498/* Mark the lockres as being dropped. It will no longer be 3499 * queued if blocking, but we still may have to wait on it 3500 * being dequeued from the downconvert thread before we can consider 3501 * it safe to drop. 3502 * 3503 * You can *not* attempt to call cluster_lock on this lockres anymore. */ 3504void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb, 3505 struct ocfs2_lock_res *lockres) 3506{ 3507 int status; 3508 struct ocfs2_mask_waiter mw; 3509 unsigned long flags, flags2; 3510 3511 ocfs2_init_mask_waiter(&mw); 3512 3513 spin_lock_irqsave(&lockres->l_lock, flags); 3514 lockres->l_flags |= OCFS2_LOCK_FREEING; 3515 if (lockres->l_flags & OCFS2_LOCK_QUEUED && current == osb->dc_task) { 3516 /* 3517 * We know the downconvert is queued but not in progress 3518 * because we are the downconvert thread and processing 3519 * different lock. So we can just remove the lock from the 3520 * queue. This is not only an optimization but also a way 3521 * to avoid the following deadlock: 3522 * ocfs2_dentry_post_unlock() 3523 * ocfs2_dentry_lock_put() 3524 * ocfs2_drop_dentry_lock() 3525 * iput() 3526 * ocfs2_evict_inode() 3527 * ocfs2_clear_inode() 3528 * ocfs2_mark_lockres_freeing() 3529 * ... blocks waiting for OCFS2_LOCK_QUEUED 3530 * since we are the downconvert thread which 3531 * should clear the flag. 3532 */ 3533 spin_unlock_irqrestore(&lockres->l_lock, flags); 3534 spin_lock_irqsave(&osb->dc_task_lock, flags2); 3535 list_del_init(&lockres->l_blocked_list); 3536 osb->blocked_lock_count--; 3537 spin_unlock_irqrestore(&osb->dc_task_lock, flags2); 3538 /* 3539 * Warn if we recurse into another post_unlock call. Strictly 3540 * speaking it isn't a problem but we need to be careful if 3541 * that happens (stack overflow, deadlocks, ...) so warn if 3542 * ocfs2 grows a path for which this can happen. 3543 */ 3544 WARN_ON_ONCE(lockres->l_ops->post_unlock); 3545 /* Since the lock is freeing we don't do much in the fn below */ 3546 ocfs2_process_blocked_lock(osb, lockres); 3547 return; 3548 } 3549 while (lockres->l_flags & OCFS2_LOCK_QUEUED) { 3550 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0); 3551 spin_unlock_irqrestore(&lockres->l_lock, flags); 3552 3553 mlog(0, "Waiting on lockres %s\n", lockres->l_name); 3554 3555 status = ocfs2_wait_for_mask(&mw); 3556 if (status) 3557 mlog_errno(status); 3558 3559 spin_lock_irqsave(&lockres->l_lock, flags); 3560 } 3561 spin_unlock_irqrestore(&lockres->l_lock, flags); 3562} 3563 3564void ocfs2_simple_drop_lockres(struct ocfs2_super *osb, 3565 struct ocfs2_lock_res *lockres) 3566{ 3567 int ret; 3568 3569 ocfs2_mark_lockres_freeing(osb, lockres); 3570 ret = ocfs2_drop_lock(osb, lockres); 3571 if (ret) 3572 mlog_errno(ret); 3573} 3574 3575static void ocfs2_drop_osb_locks(struct ocfs2_super *osb) 3576{ 3577 ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres); 3578 ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres); 3579 ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres); 3580 ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres); 3581} 3582 3583int ocfs2_drop_inode_locks(struct inode *inode) 3584{ 3585 int status, err; 3586 3587 /* No need to call ocfs2_mark_lockres_freeing here - 3588 * ocfs2_clear_inode has done it for us. */ 3589 3590 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb), 3591 &OCFS2_I(inode)->ip_open_lockres); 3592 if (err < 0) 3593 mlog_errno(err); 3594 3595 status = err; 3596 3597 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb), 3598 &OCFS2_I(inode)->ip_inode_lockres); 3599 if (err < 0) 3600 mlog_errno(err); 3601 if (err < 0 && !status) 3602 status = err; 3603 3604 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb), 3605 &OCFS2_I(inode)->ip_rw_lockres); 3606 if (err < 0) 3607 mlog_errno(err); 3608 if (err < 0 && !status) 3609 status = err; 3610 3611 return status; 3612} 3613 3614static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres, 3615 int new_level) 3616{ 3617 assert_spin_locked(&lockres->l_lock); 3618 3619 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL); 3620 3621 if (lockres->l_level <= new_level) { 3622 mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, " 3623 "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, " 3624 "block %d, pgen %d\n", lockres->l_name, lockres->l_level, 3625 new_level, list_empty(&lockres->l_blocked_list), 3626 list_empty(&lockres->l_mask_waiters), lockres->l_type, 3627 lockres->l_flags, lockres->l_ro_holders, 3628 lockres->l_ex_holders, lockres->l_action, 3629 lockres->l_unlock_action, lockres->l_requested, 3630 lockres->l_blocking, lockres->l_pending_gen); 3631 BUG(); 3632 } 3633 3634 mlog(ML_BASTS, "lockres %s, level %d => %d, blocking %d\n", 3635 lockres->l_name, lockres->l_level, new_level, lockres->l_blocking); 3636 3637 lockres->l_action = OCFS2_AST_DOWNCONVERT; 3638 lockres->l_requested = new_level; 3639 lockres_or_flags(lockres, OCFS2_LOCK_BUSY); 3640 return lockres_set_pending(lockres); 3641} 3642 3643static int ocfs2_downconvert_lock(struct ocfs2_super *osb, 3644 struct ocfs2_lock_res *lockres, 3645 int new_level, 3646 int lvb, 3647 unsigned int generation) 3648{ 3649 int ret; 3650 u32 dlm_flags = DLM_LKF_CONVERT; 3651 3652 mlog(ML_BASTS, "lockres %s, level %d => %d\n", lockres->l_name, 3653 lockres->l_level, new_level); 3654 3655 /* 3656 * On DLM_LKF_VALBLK, fsdlm behaves differently with o2cb. It always 3657 * expects DLM_LKF_VALBLK being set if the LKB has LVB, so that 3658 * we can recover correctly from node failure. Otherwise, we may get 3659 * invalid LVB in LKB, but without DLM_SBF_VALNOTVALID being set. 3660 */ 3661 if (ocfs2_userspace_stack(osb) && 3662 lockres->l_ops->flags & LOCK_TYPE_USES_LVB) 3663 lvb = 1; 3664 3665 if (lvb) 3666 dlm_flags |= DLM_LKF_VALBLK; 3667 3668 ret = ocfs2_dlm_lock(osb->cconn, 3669 new_level, 3670 &lockres->l_lksb, 3671 dlm_flags, 3672 lockres->l_name, 3673 OCFS2_LOCK_ID_MAX_LEN - 1); 3674 lockres_clear_pending(lockres, generation, osb); 3675 if (ret) { 3676 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres); 3677 ocfs2_recover_from_dlm_error(lockres, 1); 3678 goto bail; 3679 } 3680 3681 ret = 0; 3682bail: 3683 return ret; 3684} 3685 3686/* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */ 3687static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb, 3688 struct ocfs2_lock_res *lockres) 3689{ 3690 assert_spin_locked(&lockres->l_lock); 3691 3692 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) { 3693 /* If we're already trying to cancel a lock conversion 3694 * then just drop the spinlock and allow the caller to 3695 * requeue this lock. */ 3696 mlog(ML_BASTS, "lockres %s, skip convert\n", lockres->l_name); 3697 return 0; 3698 } 3699 3700 /* were we in a convert when we got the bast fire? */ 3701 BUG_ON(lockres->l_action != OCFS2_AST_CONVERT && 3702 lockres->l_action != OCFS2_AST_DOWNCONVERT); 3703 /* set things up for the unlockast to know to just 3704 * clear out the ast_action and unset busy, etc. */ 3705 lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT; 3706 3707 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY), 3708 "lock %s, invalid flags: 0x%lx\n", 3709 lockres->l_name, lockres->l_flags); 3710 3711 mlog(ML_BASTS, "lockres %s\n", lockres->l_name); 3712 3713 return 1; 3714} 3715 3716static int ocfs2_cancel_convert(struct ocfs2_super *osb, 3717 struct ocfs2_lock_res *lockres) 3718{ 3719 int ret; 3720 3721 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, 3722 DLM_LKF_CANCEL); 3723 if (ret) { 3724 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres); 3725 ocfs2_recover_from_dlm_error(lockres, 0); 3726 } 3727 3728 mlog(ML_BASTS, "lockres %s\n", lockres->l_name); 3729 3730 return ret; 3731} 3732 3733static int ocfs2_unblock_lock(struct ocfs2_super *osb, 3734 struct ocfs2_lock_res *lockres, 3735 struct ocfs2_unblock_ctl *ctl) 3736{ 3737 unsigned long flags; 3738 int blocking; 3739 int new_level; 3740 int level; 3741 int ret = 0; 3742 int set_lvb = 0; 3743 unsigned int gen; 3744 3745 spin_lock_irqsave(&lockres->l_lock, flags); 3746 3747recheck: 3748 /* 3749 * Is it still blocking? If not, we have no more work to do. 3750 */ 3751 if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) { 3752 BUG_ON(lockres->l_blocking != DLM_LOCK_NL); 3753 spin_unlock_irqrestore(&lockres->l_lock, flags); 3754 ret = 0; 3755 goto leave; 3756 } 3757 3758 if (lockres->l_flags & OCFS2_LOCK_BUSY) { 3759 /* XXX 3760 * This is a *big* race. The OCFS2_LOCK_PENDING flag 3761 * exists entirely for one reason - another thread has set 3762 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock(). 3763 * 3764 * If we do ocfs2_cancel_convert() before the other thread 3765 * calls dlm_lock(), our cancel will do nothing. We will 3766 * get no ast, and we will have no way of knowing the 3767 * cancel failed. Meanwhile, the other thread will call 3768 * into dlm_lock() and wait...forever. 3769 * 3770 * Why forever? Because another node has asked for the 3771 * lock first; that's why we're here in unblock_lock(). 3772 * 3773 * The solution is OCFS2_LOCK_PENDING. When PENDING is 3774 * set, we just requeue the unblock. Only when the other 3775 * thread has called dlm_lock() and cleared PENDING will 3776 * we then cancel their request. 3777 * 3778 * All callers of dlm_lock() must set OCFS2_DLM_PENDING 3779 * at the same time they set OCFS2_DLM_BUSY. They must 3780 * clear OCFS2_DLM_PENDING after dlm_lock() returns. 3781 */ 3782 if (lockres->l_flags & OCFS2_LOCK_PENDING) { 3783 mlog(ML_BASTS, "lockres %s, ReQ: Pending\n", 3784 lockres->l_name); 3785 goto leave_requeue; 3786 } 3787 3788 ctl->requeue = 1; 3789 ret = ocfs2_prepare_cancel_convert(osb, lockres); 3790 spin_unlock_irqrestore(&lockres->l_lock, flags); 3791 if (ret) { 3792 ret = ocfs2_cancel_convert(osb, lockres); 3793 if (ret < 0) 3794 mlog_errno(ret); 3795 } 3796 goto leave; 3797 } 3798 3799 /* 3800 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is 3801 * set when the ast is received for an upconvert just before the 3802 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast 3803 * on the heels of the ast, we want to delay the downconvert just 3804 * enough to allow the up requester to do its task. Because this 3805 * lock is in the blocked queue, the lock will be downconverted 3806 * as soon as the requester is done with the lock. 3807 */ 3808 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) 3809 goto leave_requeue; 3810 3811 /* 3812 * How can we block and yet be at NL? We were trying to upconvert 3813 * from NL and got canceled. The code comes back here, and now 3814 * we notice and clear BLOCKING. 3815 */ 3816 if (lockres->l_level == DLM_LOCK_NL) { 3817 BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders); 3818 mlog(ML_BASTS, "lockres %s, Aborting dc\n", lockres->l_name); 3819 lockres->l_blocking = DLM_LOCK_NL; 3820 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED); 3821 spin_unlock_irqrestore(&lockres->l_lock, flags); 3822 goto leave; 3823 } 3824 3825 /* if we're blocking an exclusive and we have *any* holders, 3826 * then requeue. */ 3827 if ((lockres->l_blocking == DLM_LOCK_EX) 3828 && (lockres->l_ex_holders || lockres->l_ro_holders)) { 3829 mlog(ML_BASTS, "lockres %s, ReQ: EX/PR Holders %u,%u\n", 3830 lockres->l_name, lockres->l_ex_holders, 3831 lockres->l_ro_holders); 3832 goto leave_requeue; 3833 } 3834 3835 /* If it's a PR we're blocking, then only 3836 * requeue if we've got any EX holders */ 3837 if (lockres->l_blocking == DLM_LOCK_PR && 3838 lockres->l_ex_holders) { 3839 mlog(ML_BASTS, "lockres %s, ReQ: EX Holders %u\n", 3840 lockres->l_name, lockres->l_ex_holders); 3841 goto leave_requeue; 3842 } 3843 3844 /* 3845 * Can we get a lock in this state if the holder counts are 3846 * zero? The meta data unblock code used to check this. 3847 */ 3848 if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH) 3849 && (lockres->l_flags & OCFS2_LOCK_REFRESHING)) { 3850 mlog(ML_BASTS, "lockres %s, ReQ: Lock Refreshing\n", 3851 lockres->l_name); 3852 goto leave_requeue; 3853 } 3854 3855 new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking); 3856 3857 if (lockres->l_ops->check_downconvert 3858 && !lockres->l_ops->check_downconvert(lockres, new_level)) { 3859 mlog(ML_BASTS, "lockres %s, ReQ: Checkpointing\n", 3860 lockres->l_name); 3861 goto leave_requeue; 3862 } 3863 3864 /* If we get here, then we know that there are no more 3865 * incompatible holders (and anyone asking for an incompatible 3866 * lock is blocked). We can now downconvert the lock */ 3867 if (!lockres->l_ops->downconvert_worker) 3868 goto downconvert; 3869 3870 /* Some lockres types want to do a bit of work before 3871 * downconverting a lock. Allow that here. The worker function 3872 * may sleep, so we save off a copy of what we're blocking as 3873 * it may change while we're not holding the spin lock. */ 3874 blocking = lockres->l_blocking; 3875 level = lockres->l_level; 3876 spin_unlock_irqrestore(&lockres->l_lock, flags); 3877 3878 ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking); 3879 3880 if (ctl->unblock_action == UNBLOCK_STOP_POST) { 3881 mlog(ML_BASTS, "lockres %s, UNBLOCK_STOP_POST\n", 3882 lockres->l_name); 3883 goto leave; 3884 } 3885 3886 spin_lock_irqsave(&lockres->l_lock, flags); 3887 if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) { 3888 /* If this changed underneath us, then we can't drop 3889 * it just yet. */ 3890 mlog(ML_BASTS, "lockres %s, block=%d:%d, level=%d:%d, " 3891 "Recheck\n", lockres->l_name, blocking, 3892 lockres->l_blocking, level, lockres->l_level); 3893 goto recheck; 3894 } 3895 3896downconvert: 3897 ctl->requeue = 0; 3898 3899 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) { 3900 if (lockres->l_level == DLM_LOCK_EX) 3901 set_lvb = 1; 3902 3903 /* 3904 * We only set the lvb if the lock has been fully 3905 * refreshed - otherwise we risk setting stale 3906 * data. Otherwise, there's no need to actually clear 3907 * out the lvb here as it's value is still valid. 3908 */ 3909 if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) 3910 lockres->l_ops->set_lvb(lockres); 3911 } 3912 3913 gen = ocfs2_prepare_downconvert(lockres, new_level); 3914 spin_unlock_irqrestore(&lockres->l_lock, flags); 3915 ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb, 3916 gen); 3917 /* The dlm lock convert is being cancelled in background, 3918 * ocfs2_cancel_convert() is asynchronous in fs/dlm, 3919 * requeue it, try again later. 3920 */ 3921 if (ret == -EBUSY) { 3922 ctl->requeue = 1; 3923 mlog(ML_BASTS, "lockres %s, ReQ: Downconvert busy\n", 3924 lockres->l_name); 3925 ret = 0; 3926 msleep(20); 3927 } 3928 3929leave: 3930 if (ret) 3931 mlog_errno(ret); 3932 return ret; 3933 3934leave_requeue: 3935 spin_unlock_irqrestore(&lockres->l_lock, flags); 3936 ctl->requeue = 1; 3937 3938 return 0; 3939} 3940 3941static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres, 3942 int blocking) 3943{ 3944 struct inode *inode; 3945 struct address_space *mapping; 3946 struct ocfs2_inode_info *oi; 3947 3948 inode = ocfs2_lock_res_inode(lockres); 3949 mapping = inode->i_mapping; 3950 3951 if (S_ISDIR(inode->i_mode)) { 3952 oi = OCFS2_I(inode); 3953 oi->ip_dir_lock_gen++; 3954 mlog(0, "generation: %u\n", oi->ip_dir_lock_gen); 3955 goto out_forget; 3956 } 3957 3958 if (!S_ISREG(inode->i_mode)) 3959 goto out; 3960 3961 /* 3962 * We need this before the filemap_fdatawrite() so that it can 3963 * transfer the dirty bit from the PTE to the 3964 * page. Unfortunately this means that even for EX->PR 3965 * downconverts, we'll lose our mappings and have to build 3966 * them up again. 3967 */ 3968 unmap_mapping_range(mapping, 0, 0, 0); 3969 3970 if (filemap_fdatawrite(mapping)) { 3971 mlog(ML_ERROR, "Could not sync inode %llu for downconvert!", 3972 (unsigned long long)OCFS2_I(inode)->ip_blkno); 3973 } 3974 sync_mapping_buffers(mapping); 3975 if (blocking == DLM_LOCK_EX) { 3976 truncate_inode_pages(mapping, 0); 3977 } else { 3978 /* We only need to wait on the I/O if we're not also 3979 * truncating pages because truncate_inode_pages waits 3980 * for us above. We don't truncate pages if we're 3981 * blocking anything < EXMODE because we want to keep 3982 * them around in that case. */ 3983 filemap_fdatawait(mapping); 3984 } 3985 3986out_forget: 3987 forget_all_cached_acls(inode); 3988 3989out: 3990 return UNBLOCK_CONTINUE; 3991} 3992 3993static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci, 3994 struct ocfs2_lock_res *lockres, 3995 int new_level) 3996{ 3997 int checkpointed = ocfs2_ci_fully_checkpointed(ci); 3998 3999 BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR); 4000 BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed); 4001 4002 if (checkpointed) 4003 return 1; 4004 4005 ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci))); 4006 return 0; 4007} 4008 4009static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres, 4010 int new_level) 4011{ 4012 struct inode *inode = ocfs2_lock_res_inode(lockres); 4013 4014 return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level); 4015} 4016 4017static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres) 4018{ 4019 struct inode *inode = ocfs2_lock_res_inode(lockres); 4020 4021 __ocfs2_stuff_meta_lvb(inode); 4022} 4023 4024/* 4025 * Does the final reference drop on our dentry lock. Right now this 4026 * happens in the downconvert thread, but we could choose to simplify the 4027 * dlmglue API and push these off to the ocfs2_wq in the future. 4028 */ 4029static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb, 4030 struct ocfs2_lock_res *lockres) 4031{ 4032 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres); 4033 ocfs2_dentry_lock_put(osb, dl); 4034} 4035 4036/* 4037 * d_delete() matching dentries before the lock downconvert. 4038 * 4039 * At this point, any process waiting to destroy the 4040 * dentry_lock due to last ref count is stopped by the 4041 * OCFS2_LOCK_QUEUED flag. 4042 * 4043 * We have two potential problems 4044 * 4045 * 1) If we do the last reference drop on our dentry_lock (via dput) 4046 * we'll wind up in ocfs2_release_dentry_lock(), waiting on 4047 * the downconvert to finish. Instead we take an elevated 4048 * reference and push the drop until after we've completed our 4049 * unblock processing. 4050 * 4051 * 2) There might be another process with a final reference, 4052 * waiting on us to finish processing. If this is the case, we 4053 * detect it and exit out - there's no more dentries anyway. 4054 */ 4055static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres, 4056 int blocking) 4057{ 4058 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres); 4059 struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode); 4060 struct dentry *dentry; 4061 unsigned long flags; 4062 int extra_ref = 0; 4063 4064 /* 4065 * This node is blocking another node from getting a read 4066 * lock. This happens when we've renamed within a 4067 * directory. We've forced the other nodes to d_delete(), but 4068 * we never actually dropped our lock because it's still 4069 * valid. The downconvert code will retain a PR for this node, 4070 * so there's no further work to do. 4071 */ 4072 if (blocking == DLM_LOCK_PR) 4073 return UNBLOCK_CONTINUE; 4074 4075 /* 4076 * Mark this inode as potentially orphaned. The code in 4077 * ocfs2_delete_inode() will figure out whether it actually 4078 * needs to be freed or not. 4079 */ 4080 spin_lock(&oi->ip_lock); 4081 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED; 4082 spin_unlock(&oi->ip_lock); 4083 4084 /* 4085 * Yuck. We need to make sure however that the check of 4086 * OCFS2_LOCK_FREEING and the extra reference are atomic with 4087 * respect to a reference decrement or the setting of that 4088 * flag. 4089 */ 4090 spin_lock_irqsave(&lockres->l_lock, flags); 4091 spin_lock(&dentry_attach_lock); 4092 if (!(lockres->l_flags & OCFS2_LOCK_FREEING) 4093 && dl->dl_count) { 4094 dl->dl_count++; 4095 extra_ref = 1; 4096 } 4097 spin_unlock(&dentry_attach_lock); 4098 spin_unlock_irqrestore(&lockres->l_lock, flags); 4099 4100 mlog(0, "extra_ref = %d\n", extra_ref); 4101 4102 /* 4103 * We have a process waiting on us in ocfs2_dentry_iput(), 4104 * which means we can't have any more outstanding 4105 * aliases. There's no need to do any more work. 4106 */ 4107 if (!extra_ref) 4108 return UNBLOCK_CONTINUE; 4109 4110 spin_lock(&dentry_attach_lock); 4111 while (1) { 4112 dentry = ocfs2_find_local_alias(dl->dl_inode, 4113 dl->dl_parent_blkno, 1); 4114 if (!dentry) 4115 break; 4116 spin_unlock(&dentry_attach_lock); 4117 4118 if (S_ISDIR(dl->dl_inode->i_mode)) 4119 shrink_dcache_parent(dentry); 4120 4121 mlog(0, "d_delete(%pd);\n", dentry); 4122 4123 /* 4124 * The following dcache calls may do an 4125 * iput(). Normally we don't want that from the 4126 * downconverting thread, but in this case it's ok 4127 * because the requesting node already has an 4128 * exclusive lock on the inode, so it can't be queued 4129 * for a downconvert. 4130 */ 4131 d_delete(dentry); 4132 dput(dentry); 4133 4134 spin_lock(&dentry_attach_lock); 4135 } 4136 spin_unlock(&dentry_attach_lock); 4137 4138 /* 4139 * If we are the last holder of this dentry lock, there is no 4140 * reason to downconvert so skip straight to the unlock. 4141 */ 4142 if (dl->dl_count == 1) 4143 return UNBLOCK_STOP_POST; 4144 4145 return UNBLOCK_CONTINUE_POST; 4146} 4147 4148static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres, 4149 int new_level) 4150{ 4151 struct ocfs2_refcount_tree *tree = 4152 ocfs2_lock_res_refcount_tree(lockres); 4153 4154 return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level); 4155} 4156 4157static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres, 4158 int blocking) 4159{ 4160 struct ocfs2_refcount_tree *tree = 4161 ocfs2_lock_res_refcount_tree(lockres); 4162 4163 ocfs2_metadata_cache_purge(&tree->rf_ci); 4164 4165 return UNBLOCK_CONTINUE; 4166} 4167 4168static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres) 4169{ 4170 struct ocfs2_qinfo_lvb *lvb; 4171 struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres); 4172 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb, 4173 oinfo->dqi_gi.dqi_type); 4174 4175 lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 4176 lvb->lvb_version = OCFS2_QINFO_LVB_VERSION; 4177 lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace); 4178 lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace); 4179 lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms); 4180 lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks); 4181 lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk); 4182 lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry); 4183} 4184 4185void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex) 4186{ 4187 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock; 4188 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb); 4189 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 4190 4191 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) 4192 ocfs2_cluster_unlock(osb, lockres, level); 4193} 4194 4195static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo) 4196{ 4197 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb, 4198 oinfo->dqi_gi.dqi_type); 4199 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock; 4200 struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb); 4201 struct buffer_head *bh = NULL; 4202 struct ocfs2_global_disk_dqinfo *gdinfo; 4203 int status = 0; 4204 4205 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) && 4206 lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) { 4207 info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace); 4208 info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace); 4209 oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms); 4210 oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks); 4211 oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk); 4212 oinfo->dqi_gi.dqi_free_entry = 4213 be32_to_cpu(lvb->lvb_free_entry); 4214 } else { 4215 status = ocfs2_read_quota_phys_block(oinfo->dqi_gqinode, 4216 oinfo->dqi_giblk, &bh); 4217 if (status) { 4218 mlog_errno(status); 4219 goto bail; 4220 } 4221 gdinfo = (struct ocfs2_global_disk_dqinfo *) 4222 (bh->b_data + OCFS2_GLOBAL_INFO_OFF); 4223 info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace); 4224 info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace); 4225 oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms); 4226 oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks); 4227 oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk); 4228 oinfo->dqi_gi.dqi_free_entry = 4229 le32_to_cpu(gdinfo->dqi_free_entry); 4230 brelse(bh); 4231 ocfs2_track_lock_refresh(lockres); 4232 } 4233 4234bail: 4235 return status; 4236} 4237 4238/* Lock quota info, this function expects at least shared lock on the quota file 4239 * so that we can safely refresh quota info from disk. */ 4240int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex) 4241{ 4242 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock; 4243 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb); 4244 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 4245 int status = 0; 4246 4247 /* On RO devices, locking really isn't needed... */ 4248 if (ocfs2_is_hard_readonly(osb)) { 4249 if (ex) 4250 status = -EROFS; 4251 goto bail; 4252 } 4253 if (ocfs2_mount_local(osb)) 4254 goto bail; 4255 4256 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0); 4257 if (status < 0) { 4258 mlog_errno(status); 4259 goto bail; 4260 } 4261 if (!ocfs2_should_refresh_lock_res(lockres)) 4262 goto bail; 4263 /* OK, we have the lock but we need to refresh the quota info */ 4264 status = ocfs2_refresh_qinfo(oinfo); 4265 if (status) 4266 ocfs2_qinfo_unlock(oinfo, ex); 4267 ocfs2_complete_lock_res_refresh(lockres, status); 4268bail: 4269 return status; 4270} 4271 4272int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex) 4273{ 4274 int status; 4275 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 4276 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres; 4277 struct ocfs2_super *osb = lockres->l_priv; 4278 4279 4280 if (ocfs2_is_hard_readonly(osb)) 4281 return -EROFS; 4282 4283 if (ocfs2_mount_local(osb)) 4284 return 0; 4285 4286 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0); 4287 if (status < 0) 4288 mlog_errno(status); 4289 4290 return status; 4291} 4292 4293void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex) 4294{ 4295 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR; 4296 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres; 4297 struct ocfs2_super *osb = lockres->l_priv; 4298 4299 if (!ocfs2_mount_local(osb)) 4300 ocfs2_cluster_unlock(osb, lockres, level); 4301} 4302 4303static void ocfs2_process_blocked_lock(struct ocfs2_super *osb, 4304 struct ocfs2_lock_res *lockres) 4305{ 4306 int status; 4307 struct ocfs2_unblock_ctl ctl = {0, 0,}; 4308 unsigned long flags; 4309 4310 /* Our reference to the lockres in this function can be 4311 * considered valid until we remove the OCFS2_LOCK_QUEUED 4312 * flag. */ 4313 4314 BUG_ON(!lockres); 4315 BUG_ON(!lockres->l_ops); 4316 4317 mlog(ML_BASTS, "lockres %s blocked\n", lockres->l_name); 4318 4319 /* Detect whether a lock has been marked as going away while 4320 * the downconvert thread was processing other things. A lock can 4321 * still be marked with OCFS2_LOCK_FREEING after this check, 4322 * but short circuiting here will still save us some 4323 * performance. */ 4324 spin_lock_irqsave(&lockres->l_lock, flags); 4325 if (lockres->l_flags & OCFS2_LOCK_FREEING) 4326 goto unqueue; 4327 spin_unlock_irqrestore(&lockres->l_lock, flags); 4328 4329 status = ocfs2_unblock_lock(osb, lockres, &ctl); 4330 if (status < 0) 4331 mlog_errno(status); 4332 4333 spin_lock_irqsave(&lockres->l_lock, flags); 4334unqueue: 4335 if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) { 4336 lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED); 4337 } else 4338 ocfs2_schedule_blocked_lock(osb, lockres); 4339 4340 mlog(ML_BASTS, "lockres %s, requeue = %s.\n", lockres->l_name, 4341 str_yes_no(ctl.requeue)); 4342 spin_unlock_irqrestore(&lockres->l_lock, flags); 4343 4344 if (ctl.unblock_action != UNBLOCK_CONTINUE 4345 && lockres->l_ops->post_unlock) 4346 lockres->l_ops->post_unlock(osb, lockres); 4347} 4348 4349static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb, 4350 struct ocfs2_lock_res *lockres) 4351{ 4352 unsigned long flags; 4353 4354 assert_spin_locked(&lockres->l_lock); 4355 4356 if (lockres->l_flags & OCFS2_LOCK_FREEING) { 4357 /* Do not schedule a lock for downconvert when it's on 4358 * the way to destruction - any nodes wanting access 4359 * to the resource will get it soon. */ 4360 mlog(ML_BASTS, "lockres %s won't be scheduled: flags 0x%lx\n", 4361 lockres->l_name, lockres->l_flags); 4362 return; 4363 } 4364 4365 lockres_or_flags(lockres, OCFS2_LOCK_QUEUED); 4366 4367 spin_lock_irqsave(&osb->dc_task_lock, flags); 4368 if (list_empty(&lockres->l_blocked_list)) { 4369 list_add_tail(&lockres->l_blocked_list, 4370 &osb->blocked_lock_list); 4371 osb->blocked_lock_count++; 4372 } 4373 spin_unlock_irqrestore(&osb->dc_task_lock, flags); 4374} 4375 4376static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb) 4377{ 4378 unsigned long processed; 4379 unsigned long flags; 4380 struct ocfs2_lock_res *lockres; 4381 4382 spin_lock_irqsave(&osb->dc_task_lock, flags); 4383 /* grab this early so we know to try again if a state change and 4384 * wake happens part-way through our work */ 4385 osb->dc_work_sequence = osb->dc_wake_sequence; 4386 4387 processed = osb->blocked_lock_count; 4388 /* 4389 * blocked lock processing in this loop might call iput which can 4390 * remove items off osb->blocked_lock_list. Downconvert up to 4391 * 'processed' number of locks, but stop short if we had some 4392 * removed in ocfs2_mark_lockres_freeing when downconverting. 4393 */ 4394 while (processed && !list_empty(&osb->blocked_lock_list)) { 4395 lockres = list_entry(osb->blocked_lock_list.next, 4396 struct ocfs2_lock_res, l_blocked_list); 4397 list_del_init(&lockres->l_blocked_list); 4398 osb->blocked_lock_count--; 4399 spin_unlock_irqrestore(&osb->dc_task_lock, flags); 4400 4401 BUG_ON(!processed); 4402 processed--; 4403 4404 ocfs2_process_blocked_lock(osb, lockres); 4405 4406 spin_lock_irqsave(&osb->dc_task_lock, flags); 4407 } 4408 spin_unlock_irqrestore(&osb->dc_task_lock, flags); 4409} 4410 4411static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb) 4412{ 4413 int empty = 0; 4414 unsigned long flags; 4415 4416 spin_lock_irqsave(&osb->dc_task_lock, flags); 4417 if (list_empty(&osb->blocked_lock_list)) 4418 empty = 1; 4419 4420 spin_unlock_irqrestore(&osb->dc_task_lock, flags); 4421 return empty; 4422} 4423 4424static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb) 4425{ 4426 int should_wake = 0; 4427 unsigned long flags; 4428 4429 spin_lock_irqsave(&osb->dc_task_lock, flags); 4430 if (osb->dc_work_sequence != osb->dc_wake_sequence) 4431 should_wake = 1; 4432 spin_unlock_irqrestore(&osb->dc_task_lock, flags); 4433 4434 return should_wake; 4435} 4436 4437static int ocfs2_downconvert_thread(void *arg) 4438{ 4439 struct ocfs2_super *osb = arg; 4440 4441 /* only quit once we've been asked to stop and there is no more 4442 * work available */ 4443 while (!(kthread_should_stop() && 4444 ocfs2_downconvert_thread_lists_empty(osb))) { 4445 4446 wait_event_interruptible(osb->dc_event, 4447 ocfs2_downconvert_thread_should_wake(osb) || 4448 kthread_should_stop()); 4449 4450 mlog(0, "downconvert_thread: awoken\n"); 4451 4452 ocfs2_downconvert_thread_do_work(osb); 4453 } 4454 4455 osb->dc_task = NULL; 4456 return 0; 4457} 4458 4459void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb) 4460{ 4461 unsigned long flags; 4462 4463 spin_lock_irqsave(&osb->dc_task_lock, flags); 4464 /* make sure the voting thread gets a swipe at whatever changes 4465 * the caller may have made to the voting state */ 4466 osb->dc_wake_sequence++; 4467 spin_unlock_irqrestore(&osb->dc_task_lock, flags); 4468 wake_up(&osb->dc_event); 4469}
~/
Read-only file preview.