1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc. 4 * All Rights Reserved. 5 */ 6 #ifndef __XFS_LOG_PRIV_H__ 7 #define __XFS_LOG_PRIV_H__ 8 9 struct xfs_buf; 10 struct xlog; 11 struct xlog_ticket; 12 struct xfs_mount; 13 14 /* 15 * get client id from packed copy. 16 * 17 * this hack is here because the xlog_pack code copies four bytes 18 * of xlog_op_header containing the fields oh_clientid, oh_flags 19 * and oh_res2 into the packed copy. 20 * 21 * later on this four byte chunk is treated as an int and the 22 * client id is pulled out. 23 * 24 * this has endian issues, of course. 25 */ 26 static inline uint xlog_get_client_id(__be32 i) 27 { 28 return be32_to_cpu(i) >> 24; 29 } 30 31 /* 32 * In core log state 33 */ 34 enum xlog_iclog_state { 35 XLOG_STATE_ACTIVE, /* Current IC log being written to */ 36 XLOG_STATE_WANT_SYNC, /* Want to sync this iclog; no more writes */ 37 XLOG_STATE_SYNCING, /* This IC log is syncing */ 38 XLOG_STATE_DONE_SYNC, /* Done syncing to disk */ 39 XLOG_STATE_CALLBACK, /* Callback functions now */ 40 XLOG_STATE_DIRTY, /* Dirty IC log, not ready for ACTIVE status */ 41 }; 42 43 #define XLOG_STATE_STRINGS \ 44 { XLOG_STATE_ACTIVE, "XLOG_STATE_ACTIVE" }, \ 45 { XLOG_STATE_WANT_SYNC, "XLOG_STATE_WANT_SYNC" }, \ 46 { XLOG_STATE_SYNCING, "XLOG_STATE_SYNCING" }, \ 47 { XLOG_STATE_DONE_SYNC, "XLOG_STATE_DONE_SYNC" }, \ 48 { XLOG_STATE_CALLBACK, "XLOG_STATE_CALLBACK" }, \ 49 { XLOG_STATE_DIRTY, "XLOG_STATE_DIRTY" } 50 51 /* 52 * In core log flags 53 */ 54 #define XLOG_ICL_NEED_FLUSH (1u << 0) /* iclog needs REQ_PREFLUSH */ 55 #define XLOG_ICL_NEED_FUA (1u << 1) /* iclog needs REQ_FUA */ 56 57 #define XLOG_ICL_STRINGS \ 58 { XLOG_ICL_NEED_FLUSH, "XLOG_ICL_NEED_FLUSH" }, \ 59 { XLOG_ICL_NEED_FUA, "XLOG_ICL_NEED_FUA" } 60 61 62 /* 63 * Log ticket flags 64 */ 65 #define XLOG_TIC_PERM_RESERV (1u << 0) /* permanent reservation */ 66 67 #define XLOG_TIC_FLAGS \ 68 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" } 69 70 /* 71 * Below are states for covering allocation transactions. 72 * By covering, we mean changing the h_tail_lsn in the last on-disk 73 * log write such that no allocation transactions will be re-done during 74 * recovery after a system crash. Recovery starts at the last on-disk 75 * log write. 76 * 77 * These states are used to insert dummy log entries to cover 78 * space allocation transactions which can undo non-transactional changes 79 * after a crash. Writes to a file with space 80 * already allocated do not result in any transactions. Allocations 81 * might include space beyond the EOF. So if we just push the EOF a 82 * little, the last transaction for the file could contain the wrong 83 * size. If there is no file system activity, after an allocation 84 * transaction, and the system crashes, the allocation transaction 85 * will get replayed and the file will be truncated. This could 86 * be hours/days/... after the allocation occurred. 87 * 88 * The fix for this is to do two dummy transactions when the 89 * system is idle. We need two dummy transaction because the h_tail_lsn 90 * in the log record header needs to point beyond the last possible 91 * non-dummy transaction. The first dummy changes the h_tail_lsn to 92 * the first transaction before the dummy. The second dummy causes 93 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn. 94 * 95 * These dummy transactions get committed when everything 96 * is idle (after there has been some activity). 97 * 98 * There are 5 states used to control this. 99 * 100 * IDLE -- no logging has been done on the file system or 101 * we are done covering previous transactions. 102 * NEED -- logging has occurred and we need a dummy transaction 103 * when the log becomes idle. 104 * DONE -- we were in the NEED state and have committed a dummy 105 * transaction. 106 * NEED2 -- we detected that a dummy transaction has gone to the 107 * on disk log with no other transactions. 108 * DONE2 -- we committed a dummy transaction when in the NEED2 state. 109 * 110 * There are two places where we switch states: 111 * 112 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2. 113 * We commit the dummy transaction and switch to DONE or DONE2, 114 * respectively. In all other states, we don't do anything. 115 * 116 * 2.) When we finish writing the on-disk log (xlog_state_clean_log). 117 * 118 * No matter what state we are in, if this isn't the dummy 119 * transaction going out, the next state is NEED. 120 * So, if we aren't in the DONE or DONE2 states, the next state 121 * is NEED. We can't be finishing a write of the dummy record 122 * unless it was committed and the state switched to DONE or DONE2. 123 * 124 * If we are in the DONE state and this was a write of the 125 * dummy transaction, we move to NEED2. 126 * 127 * If we are in the DONE2 state and this was a write of the 128 * dummy transaction, we move to IDLE. 129 * 130 * 131 * Writing only one dummy transaction can get appended to 132 * one file space allocation. When this happens, the log recovery 133 * code replays the space allocation and a file could be truncated. 134 * This is why we have the NEED2 and DONE2 states before going idle. 135 */ 136 137 #define XLOG_STATE_COVER_IDLE 0 138 #define XLOG_STATE_COVER_NEED 1 139 #define XLOG_STATE_COVER_DONE 2 140 #define XLOG_STATE_COVER_NEED2 3 141 #define XLOG_STATE_COVER_DONE2 4 142 143 #define XLOG_COVER_OPS 5 144 145 typedef struct xlog_ticket { 146 struct list_head t_queue; /* reserve/write queue */ 147 struct task_struct *t_task; /* task that owns this ticket */ 148 xlog_tid_t t_tid; /* transaction identifier : 4 */ 149 atomic_t t_ref; /* ticket reference count : 4 */ 150 int t_curr_res; /* current reservation in bytes : 4 */ 151 int t_unit_res; /* unit reservation in bytes : 4 */ 152 char t_ocnt; /* original count : 1 */ 153 char t_cnt; /* current count : 1 */ 154 uint8_t t_flags; /* properties of reservation : 1 */ 155 } xlog_ticket_t; 156 157 /* 158 * - A log record header is 512 bytes. There is plenty of room to grow the 159 * xlog_rec_header_t into the reserved space. 160 * - ic_data follows, so a write to disk can start at the beginning of 161 * the iclog. 162 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk. 163 * - ic_next is the pointer to the next iclog in the ring. 164 * - ic_log is a pointer back to the global log structure. 165 * - ic_size is the full size of the log buffer, minus the cycle headers. 166 * - ic_offset is the current number of bytes written to in this iclog. 167 * - ic_refcnt is bumped when someone is writing to the log. 168 * - ic_state is the state of the iclog. 169 * 170 * Because of cacheline contention on large machines, we need to separate 171 * various resources onto different cachelines. To start with, make the 172 * structure cacheline aligned. The following fields can be contended on 173 * by independent processes: 174 * 175 * - ic_callbacks 176 * - ic_refcnt 177 * - fields protected by the global l_icloglock 178 * 179 * so we need to ensure that these fields are located in separate cachelines. 180 * We'll put all the read-only and l_icloglock fields in the first cacheline, 181 * and move everything else out to subsequent cachelines. 182 */ 183 typedef struct xlog_in_core { 184 wait_queue_head_t ic_force_wait; 185 wait_queue_head_t ic_write_wait; 186 struct xlog_in_core *ic_next; 187 struct xlog_in_core *ic_prev; 188 struct xlog *ic_log; 189 u32 ic_size; 190 u32 ic_offset; 191 enum xlog_iclog_state ic_state; 192 unsigned int ic_flags; 193 void *ic_datap; /* pointer to iclog data */ 194 struct list_head ic_callbacks; 195 196 /* reference counts need their own cacheline */ 197 atomic_t ic_refcnt ____cacheline_aligned_in_smp; 198 xlog_in_core_2_t *ic_data; 199 #define ic_header ic_data->hic_header 200 #ifdef DEBUG 201 bool ic_fail_crc : 1; 202 #endif 203 struct semaphore ic_sema; 204 struct work_struct ic_end_io_work; 205 struct bio ic_bio; 206 struct bio_vec ic_bvec[]; 207 } xlog_in_core_t; 208 209 /* 210 * The CIL context is used to aggregate per-transaction details as well be 211 * passed to the iclog for checkpoint post-commit processing. After being 212 * passed to the iclog, another context needs to be allocated for tracking the 213 * next set of transactions to be aggregated into a checkpoint. 214 */ 215 struct xfs_cil; 216 217 struct xfs_cil_ctx { 218 struct xfs_cil *cil; 219 xfs_csn_t sequence; /* chkpt sequence # */ 220 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */ 221 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */ 222 struct xlog_in_core *commit_iclog; 223 struct xlog_ticket *ticket; /* chkpt ticket */ 224 int space_used; /* aggregate size of regions */ 225 struct list_head busy_extents; /* busy extents in chkpt */ 226 struct xfs_log_vec *lv_chain; /* logvecs being pushed */ 227 struct list_head iclog_entry; 228 struct list_head committing; /* ctx committing list */ 229 struct work_struct discard_endio_work; 230 struct work_struct push_work; 231 }; 232 233 /* 234 * Committed Item List structure 235 * 236 * This structure is used to track log items that have been committed but not 237 * yet written into the log. It is used only when the delayed logging mount 238 * option is enabled. 239 * 240 * This structure tracks the list of committing checkpoint contexts so 241 * we can avoid the problem of having to hold out new transactions during a 242 * flush until we have a the commit record LSN of the checkpoint. We can 243 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a 244 * sequence match and extract the commit LSN directly from there. If the 245 * checkpoint is still in the process of committing, we can block waiting for 246 * the commit LSN to be determined as well. This should make synchronous 247 * operations almost as efficient as the old logging methods. 248 */ 249 struct xfs_cil { 250 struct xlog *xc_log; 251 struct list_head xc_cil; 252 spinlock_t xc_cil_lock; 253 struct workqueue_struct *xc_push_wq; 254 255 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp; 256 struct xfs_cil_ctx *xc_ctx; 257 258 spinlock_t xc_push_lock ____cacheline_aligned_in_smp; 259 xfs_csn_t xc_push_seq; 260 bool xc_push_commit_stable; 261 struct list_head xc_committing; 262 wait_queue_head_t xc_commit_wait; 263 wait_queue_head_t xc_start_wait; 264 xfs_csn_t xc_current_sequence; 265 wait_queue_head_t xc_push_wait; /* background push throttle */ 266 } ____cacheline_aligned_in_smp; 267 268 /* 269 * The amount of log space we allow the CIL to aggregate is difficult to size. 270 * Whatever we choose, we have to make sure we can get a reservation for the 271 * log space effectively, that it is large enough to capture sufficient 272 * relogging to reduce log buffer IO significantly, but it is not too large for 273 * the log or induces too much latency when writing out through the iclogs. We 274 * track both space consumed and the number of vectors in the checkpoint 275 * context, so we need to decide which to use for limiting. 276 * 277 * Every log buffer we write out during a push needs a header reserved, which 278 * is at least one sector and more for v2 logs. Hence we need a reservation of 279 * at least 512 bytes per 32k of log space just for the LR headers. That means 280 * 16KB of reservation per megabyte of delayed logging space we will consume, 281 * plus various headers. The number of headers will vary based on the num of 282 * io vectors, so limiting on a specific number of vectors is going to result 283 * in transactions of varying size. IOWs, it is more consistent to track and 284 * limit space consumed in the log rather than by the number of objects being 285 * logged in order to prevent checkpoint ticket overruns. 286 * 287 * Further, use of static reservations through the log grant mechanism is 288 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write 289 * grant) and a significant deadlock potential because regranting write space 290 * can block on log pushes. Hence if we have to regrant log space during a log 291 * push, we can deadlock. 292 * 293 * However, we can avoid this by use of a dynamic "reservation stealing" 294 * technique during transaction commit whereby unused reservation space in the 295 * transaction ticket is transferred to the CIL ctx commit ticket to cover the 296 * space needed by the checkpoint transaction. This means that we never need to 297 * specifically reserve space for the CIL checkpoint transaction, nor do we 298 * need to regrant space once the checkpoint completes. This also means the 299 * checkpoint transaction ticket is specific to the checkpoint context, rather 300 * than the CIL itself. 301 * 302 * With dynamic reservations, we can effectively make up arbitrary limits for 303 * the checkpoint size so long as they don't violate any other size rules. 304 * Recovery imposes a rule that no transaction exceed half the log, so we are 305 * limited by that. Furthermore, the log transaction reservation subsystem 306 * tries to keep 25% of the log free, so we need to keep below that limit or we 307 * risk running out of free log space to start any new transactions. 308 * 309 * In order to keep background CIL push efficient, we only need to ensure the 310 * CIL is large enough to maintain sufficient in-memory relogging to avoid 311 * repeated physical writes of frequently modified metadata. If we allow the CIL 312 * to grow to a substantial fraction of the log, then we may be pinning hundreds 313 * of megabytes of metadata in memory until the CIL flushes. This can cause 314 * issues when we are running low on memory - pinned memory cannot be reclaimed, 315 * and the CIL consumes a lot of memory. Hence we need to set an upper physical 316 * size limit for the CIL that limits the maximum amount of memory pinned by the 317 * CIL but does not limit performance by reducing relogging efficiency 318 * significantly. 319 * 320 * As such, the CIL push threshold ends up being the smaller of two thresholds: 321 * - a threshold large enough that it allows CIL to be pushed and progress to be 322 * made without excessive blocking of incoming transaction commits. This is 323 * defined to be 12.5% of the log space - half the 25% push threshold of the 324 * AIL. 325 * - small enough that it doesn't pin excessive amounts of memory but maintains 326 * close to peak relogging efficiency. This is defined to be 16x the iclog 327 * buffer window (32MB) as measurements have shown this to be roughly the 328 * point of diminishing performance increases under highly concurrent 329 * modification workloads. 330 * 331 * To prevent the CIL from overflowing upper commit size bounds, we introduce a 332 * new threshold at which we block committing transactions until the background 333 * CIL commit commences and switches to a new context. While this is not a hard 334 * limit, it forces the process committing a transaction to the CIL to block and 335 * yeild the CPU, giving the CIL push work a chance to be scheduled and start 336 * work. This prevents a process running lots of transactions from overfilling 337 * the CIL because it is not yielding the CPU. We set the blocking limit at 338 * twice the background push space threshold so we keep in line with the AIL 339 * push thresholds. 340 * 341 * Note: this is not a -hard- limit as blocking is applied after the transaction 342 * is inserted into the CIL and the push has been triggered. It is largely a 343 * throttling mechanism that allows the CIL push to be scheduled and run. A hard 344 * limit will be difficult to implement without introducing global serialisation 345 * in the CIL commit fast path, and it's not at all clear that we actually need 346 * such hard limits given the ~7 years we've run without a hard limit before 347 * finding the first situation where a checkpoint size overflow actually 348 * occurred. Hence the simple throttle, and an ASSERT check to tell us that 349 * we've overrun the max size. 350 */ 351 #define XLOG_CIL_SPACE_LIMIT(log) \ 352 min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4) 353 354 #define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \ 355 (XLOG_CIL_SPACE_LIMIT(log) * 2) 356 357 /* 358 * ticket grant locks, queues and accounting have their own cachlines 359 * as these are quite hot and can be operated on concurrently. 360 */ 361 struct xlog_grant_head { 362 spinlock_t lock ____cacheline_aligned_in_smp; 363 struct list_head waiters; 364 atomic64_t grant; 365 }; 366 367 /* 368 * The reservation head lsn is not made up of a cycle number and block number. 369 * Instead, it uses a cycle number and byte number. Logs don't expect to 370 * overflow 31 bits worth of byte offset, so using a byte number will mean 371 * that round off problems won't occur when releasing partial reservations. 372 */ 373 struct xlog { 374 /* The following fields don't need locking */ 375 struct xfs_mount *l_mp; /* mount point */ 376 struct xfs_ail *l_ailp; /* AIL log is working with */ 377 struct xfs_cil *l_cilp; /* CIL log is working with */ 378 struct xfs_buftarg *l_targ; /* buftarg of log */ 379 struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */ 380 struct delayed_work l_work; /* background flush work */ 381 long l_opstate; /* operational state */ 382 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */ 383 struct list_head *l_buf_cancel_table; 384 int l_iclog_hsize; /* size of iclog header */ 385 int l_iclog_heads; /* # of iclog header sectors */ 386 uint l_sectBBsize; /* sector size in BBs (2^n) */ 387 int l_iclog_size; /* size of log in bytes */ 388 int l_iclog_bufs; /* number of iclog buffers */ 389 xfs_daddr_t l_logBBstart; /* start block of log */ 390 int l_logsize; /* size of log in bytes */ 391 int l_logBBsize; /* size of log in BB chunks */ 392 393 /* The following block of fields are changed while holding icloglock */ 394 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp; 395 /* waiting for iclog flush */ 396 int l_covered_state;/* state of "covering disk 397 * log entries" */ 398 xlog_in_core_t *l_iclog; /* head log queue */ 399 spinlock_t l_icloglock; /* grab to change iclog state */ 400 int l_curr_cycle; /* Cycle number of log writes */ 401 int l_prev_cycle; /* Cycle number before last 402 * block increment */ 403 int l_curr_block; /* current logical log block */ 404 int l_prev_block; /* previous logical log block */ 405 406 /* 407 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and 408 * read without needing to hold specific locks. To avoid operations 409 * contending with other hot objects, place each of them on a separate 410 * cacheline. 411 */ 412 /* lsn of last LR on disk */ 413 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp; 414 /* lsn of 1st LR with unflushed * buffers */ 415 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp; 416 417 struct xlog_grant_head l_reserve_head; 418 struct xlog_grant_head l_write_head; 419 420 struct xfs_kobj l_kobj; 421 422 /* log recovery lsn tracking (for buffer submission */ 423 xfs_lsn_t l_recovery_lsn; 424 425 uint32_t l_iclog_roundoff;/* padding roundoff */ 426 427 /* Users of log incompat features should take a read lock. */ 428 struct rw_semaphore l_incompat_users; 429 }; 430 431 /* 432 * Bits for operational state 433 */ 434 #define XLOG_ACTIVE_RECOVERY 0 /* in the middle of recovery */ 435 #define XLOG_RECOVERY_NEEDED 1 /* log was recovered */ 436 #define XLOG_IO_ERROR 2 /* log hit an I/O error, and being 437 shutdown */ 438 #define XLOG_TAIL_WARN 3 /* log tail verify warning issued */ 439 440 static inline bool 441 xlog_recovery_needed(struct xlog *log) 442 { 443 return test_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate); 444 } 445 446 static inline bool 447 xlog_in_recovery(struct xlog *log) 448 { 449 return test_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate); 450 } 451 452 static inline bool 453 xlog_is_shutdown(struct xlog *log) 454 { 455 return test_bit(XLOG_IO_ERROR, &log->l_opstate); 456 } 457 458 /* 459 * Wait until the xlog_force_shutdown() has marked the log as shut down 460 * so xlog_is_shutdown() will always return true. 461 */ 462 static inline void 463 xlog_shutdown_wait( 464 struct xlog *log) 465 { 466 wait_var_event(&log->l_opstate, xlog_is_shutdown(log)); 467 } 468 469 /* common routines */ 470 extern int 471 xlog_recover( 472 struct xlog *log); 473 extern int 474 xlog_recover_finish( 475 struct xlog *log); 476 extern void 477 xlog_recover_cancel(struct xlog *); 478 479 extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead, 480 char *dp, int size); 481 482 extern struct kmem_cache *xfs_log_ticket_cache; 483 struct xlog_ticket *xlog_ticket_alloc(struct xlog *log, int unit_bytes, 484 int count, bool permanent); 485 486 void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket); 487 void xlog_print_trans(struct xfs_trans *); 488 int xlog_write(struct xlog *log, struct xfs_cil_ctx *ctx, 489 struct xfs_log_vec *log_vector, struct xlog_ticket *tic, 490 uint32_t len); 491 void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket); 492 void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket); 493 494 void xlog_state_switch_iclogs(struct xlog *log, struct xlog_in_core *iclog, 495 int eventual_size); 496 int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog); 497 498 /* 499 * When we crack an atomic LSN, we sample it first so that the value will not 500 * change while we are cracking it into the component values. This means we 501 * will always get consistent component values to work from. This should always 502 * be used to sample and crack LSNs that are stored and updated in atomic 503 * variables. 504 */ 505 static inline void 506 xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block) 507 { 508 xfs_lsn_t val = atomic64_read(lsn); 509 510 *cycle = CYCLE_LSN(val); 511 *block = BLOCK_LSN(val); 512 } 513 514 /* 515 * Calculate and assign a value to an atomic LSN variable from component pieces. 516 */ 517 static inline void 518 xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block) 519 { 520 atomic64_set(lsn, xlog_assign_lsn(cycle, block)); 521 } 522 523 /* 524 * When we crack the grant head, we sample it first so that the value will not 525 * change while we are cracking it into the component values. This means we 526 * will always get consistent component values to work from. 527 */ 528 static inline void 529 xlog_crack_grant_head_val(int64_t val, int *cycle, int *space) 530 { 531 *cycle = val >> 32; 532 *space = val & 0xffffffff; 533 } 534 535 static inline void 536 xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space) 537 { 538 xlog_crack_grant_head_val(atomic64_read(head), cycle, space); 539 } 540 541 static inline int64_t 542 xlog_assign_grant_head_val(int cycle, int space) 543 { 544 return ((int64_t)cycle << 32) | space; 545 } 546 547 static inline void 548 xlog_assign_grant_head(atomic64_t *head, int cycle, int space) 549 { 550 atomic64_set(head, xlog_assign_grant_head_val(cycle, space)); 551 } 552 553 /* 554 * Committed Item List interfaces 555 */ 556 int xlog_cil_init(struct xlog *log); 557 void xlog_cil_init_post_recovery(struct xlog *log); 558 void xlog_cil_destroy(struct xlog *log); 559 bool xlog_cil_empty(struct xlog *log); 560 void xlog_cil_commit(struct xlog *log, struct xfs_trans *tp, 561 xfs_csn_t *commit_seq, bool regrant); 562 void xlog_cil_set_ctx_write_state(struct xfs_cil_ctx *ctx, 563 struct xlog_in_core *iclog); 564 565 566 /* 567 * CIL force routines 568 */ 569 void xlog_cil_flush(struct xlog *log); 570 xfs_lsn_t xlog_cil_force_seq(struct xlog *log, xfs_csn_t sequence); 571 572 static inline void 573 xlog_cil_force(struct xlog *log) 574 { 575 xlog_cil_force_seq(log, log->l_cilp->xc_current_sequence); 576 } 577 578 /* 579 * Wrapper function for waiting on a wait queue serialised against wakeups 580 * by a spinlock. This matches the semantics of all the wait queues used in the 581 * log code. 582 */ 583 static inline void 584 xlog_wait( 585 struct wait_queue_head *wq, 586 struct spinlock *lock) 587 __releases(lock) 588 { 589 DECLARE_WAITQUEUE(wait, current); 590 591 add_wait_queue_exclusive(wq, &wait); 592 __set_current_state(TASK_UNINTERRUPTIBLE); 593 spin_unlock(lock); 594 schedule(); 595 remove_wait_queue(wq, &wait); 596 } 597 598 int xlog_wait_on_iclog(struct xlog_in_core *iclog); 599 600 /* 601 * The LSN is valid so long as it is behind the current LSN. If it isn't, this 602 * means that the next log record that includes this metadata could have a 603 * smaller LSN. In turn, this means that the modification in the log would not 604 * replay. 605 */ 606 static inline bool 607 xlog_valid_lsn( 608 struct xlog *log, 609 xfs_lsn_t lsn) 610 { 611 int cur_cycle; 612 int cur_block; 613 bool valid = true; 614 615 /* 616 * First, sample the current lsn without locking to avoid added 617 * contention from metadata I/O. The current cycle and block are updated 618 * (in xlog_state_switch_iclogs()) and read here in a particular order 619 * to avoid false negatives (e.g., thinking the metadata LSN is valid 620 * when it is not). 621 * 622 * The current block is always rewound before the cycle is bumped in 623 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in 624 * a transiently forward state. Instead, we can see the LSN in a 625 * transiently behind state if we happen to race with a cycle wrap. 626 */ 627 cur_cycle = READ_ONCE(log->l_curr_cycle); 628 smp_rmb(); 629 cur_block = READ_ONCE(log->l_curr_block); 630 631 if ((CYCLE_LSN(lsn) > cur_cycle) || 632 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) { 633 /* 634 * If the metadata LSN appears invalid, it's possible the check 635 * above raced with a wrap to the next log cycle. Grab the lock 636 * to check for sure. 637 */ 638 spin_lock(&log->l_icloglock); 639 cur_cycle = log->l_curr_cycle; 640 cur_block = log->l_curr_block; 641 spin_unlock(&log->l_icloglock); 642 643 if ((CYCLE_LSN(lsn) > cur_cycle) || 644 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) 645 valid = false; 646 } 647 648 return valid; 649 } 650 651 /* 652 * Log vector and shadow buffers can be large, so we need to use kvmalloc() here 653 * to ensure success. Unfortunately, kvmalloc() only allows GFP_KERNEL contexts 654 * to fall back to vmalloc, so we can't actually do anything useful with gfp 655 * flags to control the kmalloc() behaviour within kvmalloc(). Hence kmalloc() 656 * will do direct reclaim and compaction in the slow path, both of which are 657 * horrendously expensive. We just want kmalloc to fail fast and fall back to 658 * vmalloc if it can't get somethign straight away from the free lists or 659 * buddy allocator. Hence we have to open code kvmalloc outselves here. 660 * 661 * This assumes that the caller uses memalloc_nofs_save task context here, so 662 * despite the use of GFP_KERNEL here, we are going to be doing GFP_NOFS 663 * allocations. This is actually the only way to make vmalloc() do GFP_NOFS 664 * allocations, so lets just all pretend this is a GFP_KERNEL context 665 * operation.... 666 */ 667 static inline void * 668 xlog_kvmalloc( 669 size_t buf_size) 670 { 671 gfp_t flags = GFP_KERNEL; 672 void *p; 673 674 flags &= ~__GFP_DIRECT_RECLAIM; 675 flags |= __GFP_NOWARN | __GFP_NORETRY; 676 do { 677 p = kmalloc(buf_size, flags); 678 if (!p) 679 p = vmalloc(buf_size); 680 } while (!p); 681 682 return p; 683 } 684 685 #endif /* __XFS_LOG_PRIV_H__ */ 686