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 * Flags for log structure 16 */ 17 #define XLOG_ACTIVE_RECOVERY 0x2 /* in the middle of recovery */ 18 #define XLOG_RECOVERY_NEEDED 0x4 /* log was recovered */ 19 #define XLOG_IO_ERROR 0x8 /* log hit an I/O error, and being 20 shutdown */ 21 #define XLOG_TAIL_WARN 0x10 /* log tail verify warning issued */ 22 23 /* 24 * get client id from packed copy. 25 * 26 * this hack is here because the xlog_pack code copies four bytes 27 * of xlog_op_header containing the fields oh_clientid, oh_flags 28 * and oh_res2 into the packed copy. 29 * 30 * later on this four byte chunk is treated as an int and the 31 * client id is pulled out. 32 * 33 * this has endian issues, of course. 34 */ 35 static inline uint xlog_get_client_id(__be32 i) 36 { 37 return be32_to_cpu(i) >> 24; 38 } 39 40 /* 41 * In core log state 42 */ 43 enum xlog_iclog_state { 44 XLOG_STATE_ACTIVE, /* Current IC log being written to */ 45 XLOG_STATE_WANT_SYNC, /* Want to sync this iclog; no more writes */ 46 XLOG_STATE_SYNCING, /* This IC log is syncing */ 47 XLOG_STATE_DONE_SYNC, /* Done syncing to disk */ 48 XLOG_STATE_CALLBACK, /* Callback functions now */ 49 XLOG_STATE_DIRTY, /* Dirty IC log, not ready for ACTIVE status */ 50 XLOG_STATE_IOERROR, /* IO error happened in sync'ing log */ 51 }; 52 53 /* 54 * Flags to log ticket 55 */ 56 #define XLOG_TIC_INITED 0x1 /* has been initialized */ 57 #define XLOG_TIC_PERM_RESERV 0x2 /* permanent reservation */ 58 59 #define XLOG_TIC_FLAGS \ 60 { XLOG_TIC_INITED, "XLOG_TIC_INITED" }, \ 61 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" } 62 63 /* 64 * Below are states for covering allocation transactions. 65 * By covering, we mean changing the h_tail_lsn in the last on-disk 66 * log write such that no allocation transactions will be re-done during 67 * recovery after a system crash. Recovery starts at the last on-disk 68 * log write. 69 * 70 * These states are used to insert dummy log entries to cover 71 * space allocation transactions which can undo non-transactional changes 72 * after a crash. Writes to a file with space 73 * already allocated do not result in any transactions. Allocations 74 * might include space beyond the EOF. So if we just push the EOF a 75 * little, the last transaction for the file could contain the wrong 76 * size. If there is no file system activity, after an allocation 77 * transaction, and the system crashes, the allocation transaction 78 * will get replayed and the file will be truncated. This could 79 * be hours/days/... after the allocation occurred. 80 * 81 * The fix for this is to do two dummy transactions when the 82 * system is idle. We need two dummy transaction because the h_tail_lsn 83 * in the log record header needs to point beyond the last possible 84 * non-dummy transaction. The first dummy changes the h_tail_lsn to 85 * the first transaction before the dummy. The second dummy causes 86 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn. 87 * 88 * These dummy transactions get committed when everything 89 * is idle (after there has been some activity). 90 * 91 * There are 5 states used to control this. 92 * 93 * IDLE -- no logging has been done on the file system or 94 * we are done covering previous transactions. 95 * NEED -- logging has occurred and we need a dummy transaction 96 * when the log becomes idle. 97 * DONE -- we were in the NEED state and have committed a dummy 98 * transaction. 99 * NEED2 -- we detected that a dummy transaction has gone to the 100 * on disk log with no other transactions. 101 * DONE2 -- we committed a dummy transaction when in the NEED2 state. 102 * 103 * There are two places where we switch states: 104 * 105 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2. 106 * We commit the dummy transaction and switch to DONE or DONE2, 107 * respectively. In all other states, we don't do anything. 108 * 109 * 2.) When we finish writing the on-disk log (xlog_state_clean_log). 110 * 111 * No matter what state we are in, if this isn't the dummy 112 * transaction going out, the next state is NEED. 113 * So, if we aren't in the DONE or DONE2 states, the next state 114 * is NEED. We can't be finishing a write of the dummy record 115 * unless it was committed and the state switched to DONE or DONE2. 116 * 117 * If we are in the DONE state and this was a write of the 118 * dummy transaction, we move to NEED2. 119 * 120 * If we are in the DONE2 state and this was a write of the 121 * dummy transaction, we move to IDLE. 122 * 123 * 124 * Writing only one dummy transaction can get appended to 125 * one file space allocation. When this happens, the log recovery 126 * code replays the space allocation and a file could be truncated. 127 * This is why we have the NEED2 and DONE2 states before going idle. 128 */ 129 130 #define XLOG_STATE_COVER_IDLE 0 131 #define XLOG_STATE_COVER_NEED 1 132 #define XLOG_STATE_COVER_DONE 2 133 #define XLOG_STATE_COVER_NEED2 3 134 #define XLOG_STATE_COVER_DONE2 4 135 136 #define XLOG_COVER_OPS 5 137 138 /* Ticket reservation region accounting */ 139 #define XLOG_TIC_LEN_MAX 15 140 141 /* 142 * Reservation region 143 * As would be stored in xfs_log_iovec but without the i_addr which 144 * we don't care about. 145 */ 146 typedef struct xlog_res { 147 uint r_len; /* region length :4 */ 148 uint r_type; /* region's transaction type :4 */ 149 } xlog_res_t; 150 151 typedef struct xlog_ticket { 152 struct list_head t_queue; /* reserve/write queue */ 153 struct task_struct *t_task; /* task that owns this ticket */ 154 xlog_tid_t t_tid; /* transaction identifier : 4 */ 155 atomic_t t_ref; /* ticket reference count : 4 */ 156 int t_curr_res; /* current reservation in bytes : 4 */ 157 int t_unit_res; /* unit reservation in bytes : 4 */ 158 char t_ocnt; /* original count : 1 */ 159 char t_cnt; /* current count : 1 */ 160 char t_clientid; /* who does this belong to; : 1 */ 161 char t_flags; /* properties of reservation : 1 */ 162 163 /* reservation array fields */ 164 uint t_res_num; /* num in array : 4 */ 165 uint t_res_num_ophdrs; /* num op hdrs : 4 */ 166 uint t_res_arr_sum; /* array sum : 4 */ 167 uint t_res_o_flow; /* sum overflow : 4 */ 168 xlog_res_t t_res_arr[XLOG_TIC_LEN_MAX]; /* array of res : 8 * 15 */ 169 } xlog_ticket_t; 170 171 /* 172 * - A log record header is 512 bytes. There is plenty of room to grow the 173 * xlog_rec_header_t into the reserved space. 174 * - ic_data follows, so a write to disk can start at the beginning of 175 * the iclog. 176 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk. 177 * - ic_next is the pointer to the next iclog in the ring. 178 * - ic_log is a pointer back to the global log structure. 179 * - ic_size is the full size of the log buffer, minus the cycle headers. 180 * - ic_offset is the current number of bytes written to in this iclog. 181 * - ic_refcnt is bumped when someone is writing to the log. 182 * - ic_state is the state of the iclog. 183 * 184 * Because of cacheline contention on large machines, we need to separate 185 * various resources onto different cachelines. To start with, make the 186 * structure cacheline aligned. The following fields can be contended on 187 * by independent processes: 188 * 189 * - ic_callbacks 190 * - ic_refcnt 191 * - fields protected by the global l_icloglock 192 * 193 * so we need to ensure that these fields are located in separate cachelines. 194 * We'll put all the read-only and l_icloglock fields in the first cacheline, 195 * and move everything else out to subsequent cachelines. 196 */ 197 typedef struct xlog_in_core { 198 wait_queue_head_t ic_force_wait; 199 wait_queue_head_t ic_write_wait; 200 struct xlog_in_core *ic_next; 201 struct xlog_in_core *ic_prev; 202 struct xlog *ic_log; 203 u32 ic_size; 204 u32 ic_offset; 205 enum xlog_iclog_state ic_state; 206 char *ic_datap; /* pointer to iclog data */ 207 208 /* Callback structures need their own cacheline */ 209 spinlock_t ic_callback_lock ____cacheline_aligned_in_smp; 210 struct list_head ic_callbacks; 211 212 /* reference counts need their own cacheline */ 213 atomic_t ic_refcnt ____cacheline_aligned_in_smp; 214 xlog_in_core_2_t *ic_data; 215 #define ic_header ic_data->hic_header 216 #ifdef DEBUG 217 bool ic_fail_crc : 1; 218 #endif 219 struct semaphore ic_sema; 220 struct work_struct ic_end_io_work; 221 struct bio ic_bio; 222 struct bio_vec ic_bvec[]; 223 } xlog_in_core_t; 224 225 /* 226 * The CIL context is used to aggregate per-transaction details as well be 227 * passed to the iclog for checkpoint post-commit processing. After being 228 * passed to the iclog, another context needs to be allocated for tracking the 229 * next set of transactions to be aggregated into a checkpoint. 230 */ 231 struct xfs_cil; 232 233 struct xfs_cil_ctx { 234 struct xfs_cil *cil; 235 xfs_lsn_t sequence; /* chkpt sequence # */ 236 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */ 237 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */ 238 struct xlog_ticket *ticket; /* chkpt ticket */ 239 int nvecs; /* number of regions */ 240 int space_used; /* aggregate size of regions */ 241 struct list_head busy_extents; /* busy extents in chkpt */ 242 struct xfs_log_vec *lv_chain; /* logvecs being pushed */ 243 struct list_head iclog_entry; 244 struct list_head committing; /* ctx committing list */ 245 struct work_struct discard_endio_work; 246 }; 247 248 /* 249 * Committed Item List structure 250 * 251 * This structure is used to track log items that have been committed but not 252 * yet written into the log. It is used only when the delayed logging mount 253 * option is enabled. 254 * 255 * This structure tracks the list of committing checkpoint contexts so 256 * we can avoid the problem of having to hold out new transactions during a 257 * flush until we have a the commit record LSN of the checkpoint. We can 258 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a 259 * sequence match and extract the commit LSN directly from there. If the 260 * checkpoint is still in the process of committing, we can block waiting for 261 * the commit LSN to be determined as well. This should make synchronous 262 * operations almost as efficient as the old logging methods. 263 */ 264 struct xfs_cil { 265 struct xlog *xc_log; 266 struct list_head xc_cil; 267 spinlock_t xc_cil_lock; 268 269 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp; 270 struct xfs_cil_ctx *xc_ctx; 271 272 spinlock_t xc_push_lock ____cacheline_aligned_in_smp; 273 xfs_lsn_t xc_push_seq; 274 struct list_head xc_committing; 275 wait_queue_head_t xc_commit_wait; 276 xfs_lsn_t xc_current_sequence; 277 struct work_struct xc_push_work; 278 } ____cacheline_aligned_in_smp; 279 280 /* 281 * The amount of log space we allow the CIL to aggregate is difficult to size. 282 * Whatever we choose, we have to make sure we can get a reservation for the 283 * log space effectively, that it is large enough to capture sufficient 284 * relogging to reduce log buffer IO significantly, but it is not too large for 285 * the log or induces too much latency when writing out through the iclogs. We 286 * track both space consumed and the number of vectors in the checkpoint 287 * context, so we need to decide which to use for limiting. 288 * 289 * Every log buffer we write out during a push needs a header reserved, which 290 * is at least one sector and more for v2 logs. Hence we need a reservation of 291 * at least 512 bytes per 32k of log space just for the LR headers. That means 292 * 16KB of reservation per megabyte of delayed logging space we will consume, 293 * plus various headers. The number of headers will vary based on the num of 294 * io vectors, so limiting on a specific number of vectors is going to result 295 * in transactions of varying size. IOWs, it is more consistent to track and 296 * limit space consumed in the log rather than by the number of objects being 297 * logged in order to prevent checkpoint ticket overruns. 298 * 299 * Further, use of static reservations through the log grant mechanism is 300 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write 301 * grant) and a significant deadlock potential because regranting write space 302 * can block on log pushes. Hence if we have to regrant log space during a log 303 * push, we can deadlock. 304 * 305 * However, we can avoid this by use of a dynamic "reservation stealing" 306 * technique during transaction commit whereby unused reservation space in the 307 * transaction ticket is transferred to the CIL ctx commit ticket to cover the 308 * space needed by the checkpoint transaction. This means that we never need to 309 * specifically reserve space for the CIL checkpoint transaction, nor do we 310 * need to regrant space once the checkpoint completes. This also means the 311 * checkpoint transaction ticket is specific to the checkpoint context, rather 312 * than the CIL itself. 313 * 314 * With dynamic reservations, we can effectively make up arbitrary limits for 315 * the checkpoint size so long as they don't violate any other size rules. 316 * Recovery imposes a rule that no transaction exceed half the log, so we are 317 * limited by that. Furthermore, the log transaction reservation subsystem 318 * tries to keep 25% of the log free, so we need to keep below that limit or we 319 * risk running out of free log space to start any new transactions. 320 * 321 * In order to keep background CIL push efficient, we will set a lower 322 * threshold at which background pushing is attempted without blocking current 323 * transaction commits. A separate, higher bound defines when CIL pushes are 324 * enforced to ensure we stay within our maximum checkpoint size bounds. 325 * threshold, yet give us plenty of space for aggregation on large logs. 326 */ 327 #define XLOG_CIL_SPACE_LIMIT(log) (log->l_logsize >> 3) 328 329 /* 330 * ticket grant locks, queues and accounting have their own cachlines 331 * as these are quite hot and can be operated on concurrently. 332 */ 333 struct xlog_grant_head { 334 spinlock_t lock ____cacheline_aligned_in_smp; 335 struct list_head waiters; 336 atomic64_t grant; 337 }; 338 339 /* 340 * The reservation head lsn is not made up of a cycle number and block number. 341 * Instead, it uses a cycle number and byte number. Logs don't expect to 342 * overflow 31 bits worth of byte offset, so using a byte number will mean 343 * that round off problems won't occur when releasing partial reservations. 344 */ 345 struct xlog { 346 /* The following fields don't need locking */ 347 struct xfs_mount *l_mp; /* mount point */ 348 struct xfs_ail *l_ailp; /* AIL log is working with */ 349 struct xfs_cil *l_cilp; /* CIL log is working with */ 350 struct xfs_buftarg *l_targ; /* buftarg of log */ 351 struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */ 352 struct delayed_work l_work; /* background flush work */ 353 uint l_flags; 354 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */ 355 struct list_head *l_buf_cancel_table; 356 int l_iclog_hsize; /* size of iclog header */ 357 int l_iclog_heads; /* # of iclog header sectors */ 358 uint l_sectBBsize; /* sector size in BBs (2^n) */ 359 int l_iclog_size; /* size of log in bytes */ 360 int l_iclog_bufs; /* number of iclog buffers */ 361 xfs_daddr_t l_logBBstart; /* start block of log */ 362 int l_logsize; /* size of log in bytes */ 363 int l_logBBsize; /* size of log in BB chunks */ 364 365 /* The following block of fields are changed while holding icloglock */ 366 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp; 367 /* waiting for iclog flush */ 368 int l_covered_state;/* state of "covering disk 369 * log entries" */ 370 xlog_in_core_t *l_iclog; /* head log queue */ 371 spinlock_t l_icloglock; /* grab to change iclog state */ 372 int l_curr_cycle; /* Cycle number of log writes */ 373 int l_prev_cycle; /* Cycle number before last 374 * block increment */ 375 int l_curr_block; /* current logical log block */ 376 int l_prev_block; /* previous logical log block */ 377 378 /* 379 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and 380 * read without needing to hold specific locks. To avoid operations 381 * contending with other hot objects, place each of them on a separate 382 * cacheline. 383 */ 384 /* lsn of last LR on disk */ 385 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp; 386 /* lsn of 1st LR with unflushed * buffers */ 387 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp; 388 389 struct xlog_grant_head l_reserve_head; 390 struct xlog_grant_head l_write_head; 391 392 struct xfs_kobj l_kobj; 393 394 /* The following field are used for debugging; need to hold icloglock */ 395 #ifdef DEBUG 396 void *l_iclog_bak[XLOG_MAX_ICLOGS]; 397 #endif 398 /* log recovery lsn tracking (for buffer submission */ 399 xfs_lsn_t l_recovery_lsn; 400 }; 401 402 #define XLOG_BUF_CANCEL_BUCKET(log, blkno) \ 403 ((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE)) 404 405 #define XLOG_FORCED_SHUTDOWN(log) \ 406 (unlikely((log)->l_flags & XLOG_IO_ERROR)) 407 408 /* common routines */ 409 extern int 410 xlog_recover( 411 struct xlog *log); 412 extern int 413 xlog_recover_finish( 414 struct xlog *log); 415 extern void 416 xlog_recover_cancel(struct xlog *); 417 418 extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead, 419 char *dp, int size); 420 421 extern kmem_zone_t *xfs_log_ticket_zone; 422 struct xlog_ticket * 423 xlog_ticket_alloc( 424 struct xlog *log, 425 int unit_bytes, 426 int count, 427 char client, 428 bool permanent, 429 xfs_km_flags_t alloc_flags); 430 431 432 static inline void 433 xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes) 434 { 435 *ptr += bytes; 436 *len -= bytes; 437 *off += bytes; 438 } 439 440 void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket); 441 void xlog_print_trans(struct xfs_trans *); 442 int 443 xlog_write( 444 struct xlog *log, 445 struct xfs_log_vec *log_vector, 446 struct xlog_ticket *tic, 447 xfs_lsn_t *start_lsn, 448 struct xlog_in_core **commit_iclog, 449 uint flags); 450 451 /* 452 * When we crack an atomic LSN, we sample it first so that the value will not 453 * change while we are cracking it into the component values. This means we 454 * will always get consistent component values to work from. This should always 455 * be used to sample and crack LSNs that are stored and updated in atomic 456 * variables. 457 */ 458 static inline void 459 xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block) 460 { 461 xfs_lsn_t val = atomic64_read(lsn); 462 463 *cycle = CYCLE_LSN(val); 464 *block = BLOCK_LSN(val); 465 } 466 467 /* 468 * Calculate and assign a value to an atomic LSN variable from component pieces. 469 */ 470 static inline void 471 xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block) 472 { 473 atomic64_set(lsn, xlog_assign_lsn(cycle, block)); 474 } 475 476 /* 477 * When we crack the grant head, we sample it first so that the value will not 478 * change while we are cracking it into the component values. This means we 479 * will always get consistent component values to work from. 480 */ 481 static inline void 482 xlog_crack_grant_head_val(int64_t val, int *cycle, int *space) 483 { 484 *cycle = val >> 32; 485 *space = val & 0xffffffff; 486 } 487 488 static inline void 489 xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space) 490 { 491 xlog_crack_grant_head_val(atomic64_read(head), cycle, space); 492 } 493 494 static inline int64_t 495 xlog_assign_grant_head_val(int cycle, int space) 496 { 497 return ((int64_t)cycle << 32) | space; 498 } 499 500 static inline void 501 xlog_assign_grant_head(atomic64_t *head, int cycle, int space) 502 { 503 atomic64_set(head, xlog_assign_grant_head_val(cycle, space)); 504 } 505 506 /* 507 * Committed Item List interfaces 508 */ 509 int xlog_cil_init(struct xlog *log); 510 void xlog_cil_init_post_recovery(struct xlog *log); 511 void xlog_cil_destroy(struct xlog *log); 512 bool xlog_cil_empty(struct xlog *log); 513 514 /* 515 * CIL force routines 516 */ 517 xfs_lsn_t 518 xlog_cil_force_lsn( 519 struct xlog *log, 520 xfs_lsn_t sequence); 521 522 static inline void 523 xlog_cil_force(struct xlog *log) 524 { 525 xlog_cil_force_lsn(log, log->l_cilp->xc_current_sequence); 526 } 527 528 /* 529 * Wrapper function for waiting on a wait queue serialised against wakeups 530 * by a spinlock. This matches the semantics of all the wait queues used in the 531 * log code. 532 */ 533 static inline void 534 xlog_wait( 535 struct wait_queue_head *wq, 536 struct spinlock *lock) 537 __releases(lock) 538 { 539 DECLARE_WAITQUEUE(wait, current); 540 541 add_wait_queue_exclusive(wq, &wait); 542 __set_current_state(TASK_UNINTERRUPTIBLE); 543 spin_unlock(lock); 544 schedule(); 545 remove_wait_queue(wq, &wait); 546 } 547 548 /* 549 * The LSN is valid so long as it is behind the current LSN. If it isn't, this 550 * means that the next log record that includes this metadata could have a 551 * smaller LSN. In turn, this means that the modification in the log would not 552 * replay. 553 */ 554 static inline bool 555 xlog_valid_lsn( 556 struct xlog *log, 557 xfs_lsn_t lsn) 558 { 559 int cur_cycle; 560 int cur_block; 561 bool valid = true; 562 563 /* 564 * First, sample the current lsn without locking to avoid added 565 * contention from metadata I/O. The current cycle and block are updated 566 * (in xlog_state_switch_iclogs()) and read here in a particular order 567 * to avoid false negatives (e.g., thinking the metadata LSN is valid 568 * when it is not). 569 * 570 * The current block is always rewound before the cycle is bumped in 571 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in 572 * a transiently forward state. Instead, we can see the LSN in a 573 * transiently behind state if we happen to race with a cycle wrap. 574 */ 575 cur_cycle = READ_ONCE(log->l_curr_cycle); 576 smp_rmb(); 577 cur_block = READ_ONCE(log->l_curr_block); 578 579 if ((CYCLE_LSN(lsn) > cur_cycle) || 580 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) { 581 /* 582 * If the metadata LSN appears invalid, it's possible the check 583 * above raced with a wrap to the next log cycle. Grab the lock 584 * to check for sure. 585 */ 586 spin_lock(&log->l_icloglock); 587 cur_cycle = log->l_curr_cycle; 588 cur_block = log->l_curr_block; 589 spin_unlock(&log->l_icloglock); 590 591 if ((CYCLE_LSN(lsn) > cur_cycle) || 592 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) 593 valid = false; 594 } 595 596 return valid; 597 } 598 599 #endif /* __XFS_LOG_PRIV_H__ */ 600