| 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 | #include "xfs_extent_busy.h" /* for struct xfs_busy_extents */ |
| 10 | |
| 11 | struct xfs_buf; |
| 12 | struct xlog; |
| 13 | struct xlog_ticket; |
| 14 | struct xfs_mount; |
| 15 | |
| 16 | /* |
| 17 | * get client id from packed copy. |
| 18 | * |
| 19 | * this hack is here because the xlog_pack code copies four bytes |
| 20 | * of xlog_op_header containing the fields oh_clientid, oh_flags |
| 21 | * and oh_res2 into the packed copy. |
| 22 | * |
| 23 | * later on this four byte chunk is treated as an int and the |
| 24 | * client id is pulled out. |
| 25 | * |
| 26 | * this has endian issues, of course. |
| 27 | */ |
| 28 | static inline uint xlog_get_client_id(__be32 i) |
| 29 | { |
| 30 | return be32_to_cpu(i) >> 24; |
| 31 | } |
| 32 | |
| 33 | /* |
| 34 | * In core log state |
| 35 | */ |
| 36 | enum xlog_iclog_state { |
| 37 | XLOG_STATE_ACTIVE, /* Current IC log being written to */ |
| 38 | XLOG_STATE_WANT_SYNC, /* Want to sync this iclog; no more writes */ |
| 39 | XLOG_STATE_SYNCING, /* This IC log is syncing */ |
| 40 | XLOG_STATE_DONE_SYNC, /* Done syncing to disk */ |
| 41 | XLOG_STATE_CALLBACK, /* Callback functions now */ |
| 42 | XLOG_STATE_DIRTY, /* Dirty IC log, not ready for ACTIVE status */ |
| 43 | }; |
| 44 | |
| 45 | #define XLOG_STATE_STRINGS \ |
| 46 | { XLOG_STATE_ACTIVE, "XLOG_STATE_ACTIVE" }, \ |
| 47 | { XLOG_STATE_WANT_SYNC, "XLOG_STATE_WANT_SYNC" }, \ |
| 48 | { XLOG_STATE_SYNCING, "XLOG_STATE_SYNCING" }, \ |
| 49 | { XLOG_STATE_DONE_SYNC, "XLOG_STATE_DONE_SYNC" }, \ |
| 50 | { XLOG_STATE_CALLBACK, "XLOG_STATE_CALLBACK" }, \ |
| 51 | { XLOG_STATE_DIRTY, "XLOG_STATE_DIRTY" } |
| 52 | |
| 53 | /* |
| 54 | * In core log flags |
| 55 | */ |
| 56 | #define XLOG_ICL_NEED_FLUSH (1u << 0) /* iclog needs REQ_PREFLUSH */ |
| 57 | #define XLOG_ICL_NEED_FUA (1u << 1) /* iclog needs REQ_FUA */ |
| 58 | |
| 59 | #define XLOG_ICL_STRINGS \ |
| 60 | { XLOG_ICL_NEED_FLUSH, "XLOG_ICL_NEED_FLUSH" }, \ |
| 61 | { XLOG_ICL_NEED_FUA, "XLOG_ICL_NEED_FUA" } |
| 62 | |
| 63 | |
| 64 | /* |
| 65 | * Log ticket flags |
| 66 | */ |
| 67 | #define XLOG_TIC_PERM_RESERV (1u << 0) /* permanent reservation */ |
| 68 | |
| 69 | #define XLOG_TIC_FLAGS \ |
| 70 | { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" } |
| 71 | |
| 72 | /* |
| 73 | * Below are states for covering allocation transactions. |
| 74 | * By covering, we mean changing the h_tail_lsn in the last on-disk |
| 75 | * log write such that no allocation transactions will be re-done during |
| 76 | * recovery after a system crash. Recovery starts at the last on-disk |
| 77 | * log write. |
| 78 | * |
| 79 | * These states are used to insert dummy log entries to cover |
| 80 | * space allocation transactions which can undo non-transactional changes |
| 81 | * after a crash. Writes to a file with space |
| 82 | * already allocated do not result in any transactions. Allocations |
| 83 | * might include space beyond the EOF. So if we just push the EOF a |
| 84 | * little, the last transaction for the file could contain the wrong |
| 85 | * size. If there is no file system activity, after an allocation |
| 86 | * transaction, and the system crashes, the allocation transaction |
| 87 | * will get replayed and the file will be truncated. This could |
| 88 | * be hours/days/... after the allocation occurred. |
| 89 | * |
| 90 | * The fix for this is to do two dummy transactions when the |
| 91 | * system is idle. We need two dummy transaction because the h_tail_lsn |
| 92 | * in the log record header needs to point beyond the last possible |
| 93 | * non-dummy transaction. The first dummy changes the h_tail_lsn to |
| 94 | * the first transaction before the dummy. The second dummy causes |
| 95 | * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn. |
| 96 | * |
| 97 | * These dummy transactions get committed when everything |
| 98 | * is idle (after there has been some activity). |
| 99 | * |
| 100 | * There are 5 states used to control this. |
| 101 | * |
| 102 | * IDLE -- no logging has been done on the file system or |
| 103 | * we are done covering previous transactions. |
| 104 | * NEED -- logging has occurred and we need a dummy transaction |
| 105 | * when the log becomes idle. |
| 106 | * DONE -- we were in the NEED state and have committed a dummy |
| 107 | * transaction. |
| 108 | * NEED2 -- we detected that a dummy transaction has gone to the |
| 109 | * on disk log with no other transactions. |
| 110 | * DONE2 -- we committed a dummy transaction when in the NEED2 state. |
| 111 | * |
| 112 | * There are two places where we switch states: |
| 113 | * |
| 114 | * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2. |
| 115 | * We commit the dummy transaction and switch to DONE or DONE2, |
| 116 | * respectively. In all other states, we don't do anything. |
| 117 | * |
| 118 | * 2.) When we finish writing the on-disk log (xlog_state_clean_log). |
| 119 | * |
| 120 | * No matter what state we are in, if this isn't the dummy |
| 121 | * transaction going out, the next state is NEED. |
| 122 | * So, if we aren't in the DONE or DONE2 states, the next state |
| 123 | * is NEED. We can't be finishing a write of the dummy record |
| 124 | * unless it was committed and the state switched to DONE or DONE2. |
| 125 | * |
| 126 | * If we are in the DONE state and this was a write of the |
| 127 | * dummy transaction, we move to NEED2. |
| 128 | * |
| 129 | * If we are in the DONE2 state and this was a write of the |
| 130 | * dummy transaction, we move to IDLE. |
| 131 | * |
| 132 | * |
| 133 | * Writing only one dummy transaction can get appended to |
| 134 | * one file space allocation. When this happens, the log recovery |
| 135 | * code replays the space allocation and a file could be truncated. |
| 136 | * This is why we have the NEED2 and DONE2 states before going idle. |
| 137 | */ |
| 138 | |
| 139 | #define XLOG_STATE_COVER_IDLE 0 |
| 140 | #define XLOG_STATE_COVER_NEED 1 |
| 141 | #define XLOG_STATE_COVER_DONE 2 |
| 142 | #define XLOG_STATE_COVER_NEED2 3 |
| 143 | #define XLOG_STATE_COVER_DONE2 4 |
| 144 | |
| 145 | #define XLOG_COVER_OPS 5 |
| 146 | |
| 147 | struct xlog_ticket { |
| 148 | struct list_head t_queue; /* reserve/write queue */ |
| 149 | struct task_struct *t_task; /* task that owns this ticket */ |
| 150 | xlog_tid_t t_tid; /* transaction identifier */ |
| 151 | atomic_t t_ref; /* ticket reference count */ |
| 152 | int t_curr_res; /* current reservation */ |
| 153 | int t_unit_res; /* unit reservation */ |
| 154 | char t_ocnt; /* original unit count */ |
| 155 | char t_cnt; /* current unit count */ |
| 156 | uint8_t t_flags; /* properties of reservation */ |
| 157 | int t_iclog_hdrs; /* iclog hdrs in t_curr_res */ |
| 158 | }; |
| 159 | |
| 160 | /* |
| 161 | * In-core log structure. |
| 162 | * |
| 163 | * - ic_forcewait is used to implement synchronous forcing of the iclog to disk. |
| 164 | * - ic_next is the pointer to the next iclog in the ring. |
| 165 | * - ic_log is a pointer back to the global log structure. |
| 166 | * - ic_size is the full size of the log buffer, minus the cycle headers. |
| 167 | * - ic_offset is the current number of bytes written to in this iclog. |
| 168 | * - ic_refcnt is bumped when someone is writing to the log. |
| 169 | * - ic_state is the state of the iclog. |
| 170 | * |
| 171 | * Because of cacheline contention on large machines, we need to separate |
| 172 | * various resources onto different cachelines. To start with, make the |
| 173 | * structure cacheline aligned. The following fields can be contended on |
| 174 | * by independent processes: |
| 175 | * |
| 176 | * - ic_callbacks |
| 177 | * - ic_refcnt |
| 178 | * - fields protected by the global l_icloglock |
| 179 | * |
| 180 | * so we need to ensure that these fields are located in separate cachelines. |
| 181 | * We'll put all the read-only and l_icloglock fields in the first cacheline, |
| 182 | * and move everything else out to subsequent cachelines. |
| 183 | */ |
| 184 | struct xlog_in_core { |
| 185 | wait_queue_head_t ic_force_wait; |
| 186 | wait_queue_head_t ic_write_wait; |
| 187 | struct xlog_in_core *ic_next; |
| 188 | struct xlog_in_core *ic_prev; |
| 189 | struct xlog *ic_log; |
| 190 | u32 ic_size; |
| 191 | u32 ic_offset; |
| 192 | enum xlog_iclog_state ic_state; |
| 193 | unsigned int ic_flags; |
| 194 | void *ic_datap; /* pointer to iclog data */ |
| 195 | struct list_head ic_callbacks; |
| 196 | |
| 197 | /* reference counts need their own cacheline */ |
| 198 | atomic_t ic_refcnt ____cacheline_aligned_in_smp; |
| 199 | struct xlog_rec_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 | }; |
| 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 | atomic_t space_used; /* aggregate size of regions */ |
| 225 | struct xfs_busy_extents busy_extents; |
| 226 | struct list_head log_items; /* log items in chkpt */ |
| 227 | struct list_head lv_chain; /* logvecs being pushed */ |
| 228 | struct list_head iclog_entry; |
| 229 | struct list_head committing; /* ctx committing list */ |
| 230 | struct work_struct push_work; |
| 231 | atomic_t order_id; |
| 232 | |
| 233 | /* |
| 234 | * CPUs that could have added items to the percpu CIL data. Access is |
| 235 | * coordinated with xc_ctx_lock. |
| 236 | */ |
| 237 | struct cpumask cil_pcpmask; |
| 238 | }; |
| 239 | |
| 240 | /* |
| 241 | * Per-cpu CIL tracking items |
| 242 | */ |
| 243 | struct xlog_cil_pcp { |
| 244 | int32_t space_used; |
| 245 | uint32_t space_reserved; |
| 246 | struct list_head busy_extents; |
| 247 | struct list_head log_items; |
| 248 | }; |
| 249 | |
| 250 | /* |
| 251 | * Committed Item List structure |
| 252 | * |
| 253 | * This structure is used to track log items that have been committed but not |
| 254 | * yet written into the log. It is used only when the delayed logging mount |
| 255 | * option is enabled. |
| 256 | * |
| 257 | * This structure tracks the list of committing checkpoint contexts so |
| 258 | * we can avoid the problem of having to hold out new transactions during a |
| 259 | * flush until we have a the commit record LSN of the checkpoint. We can |
| 260 | * traverse the list of committing contexts in xlog_cil_push_lsn() to find a |
| 261 | * sequence match and extract the commit LSN directly from there. If the |
| 262 | * checkpoint is still in the process of committing, we can block waiting for |
| 263 | * the commit LSN to be determined as well. This should make synchronous |
| 264 | * operations almost as efficient as the old logging methods. |
| 265 | */ |
| 266 | struct xfs_cil { |
| 267 | struct xlog *xc_log; |
| 268 | unsigned long xc_flags; |
| 269 | atomic_t xc_iclog_hdrs; |
| 270 | struct workqueue_struct *xc_push_wq; |
| 271 | |
| 272 | struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp; |
| 273 | struct xfs_cil_ctx *xc_ctx; |
| 274 | |
| 275 | spinlock_t xc_push_lock ____cacheline_aligned_in_smp; |
| 276 | xfs_csn_t xc_push_seq; |
| 277 | bool xc_push_commit_stable; |
| 278 | struct list_head xc_committing; |
| 279 | wait_queue_head_t xc_commit_wait; |
| 280 | wait_queue_head_t xc_start_wait; |
| 281 | xfs_csn_t xc_current_sequence; |
| 282 | wait_queue_head_t xc_push_wait; /* background push throttle */ |
| 283 | |
| 284 | void __percpu *xc_pcp; /* percpu CIL structures */ |
| 285 | } ____cacheline_aligned_in_smp; |
| 286 | |
| 287 | /* xc_flags bit values */ |
| 288 | #define XLOG_CIL_EMPTY 1 |
| 289 | #define XLOG_CIL_PCP_SPACE 2 |
| 290 | |
| 291 | /* |
| 292 | * The amount of log space we allow the CIL to aggregate is difficult to size. |
| 293 | * Whatever we choose, we have to make sure we can get a reservation for the |
| 294 | * log space effectively, that it is large enough to capture sufficient |
| 295 | * relogging to reduce log buffer IO significantly, but it is not too large for |
| 296 | * the log or induces too much latency when writing out through the iclogs. We |
| 297 | * track both space consumed and the number of vectors in the checkpoint |
| 298 | * context, so we need to decide which to use for limiting. |
| 299 | * |
| 300 | * Every log buffer we write out during a push needs a header reserved, which |
| 301 | * is at least one sector and more for v2 logs. Hence we need a reservation of |
| 302 | * at least 512 bytes per 32k of log space just for the LR headers. That means |
| 303 | * 16KB of reservation per megabyte of delayed logging space we will consume, |
| 304 | * plus various headers. The number of headers will vary based on the num of |
| 305 | * io vectors, so limiting on a specific number of vectors is going to result |
| 306 | * in transactions of varying size. IOWs, it is more consistent to track and |
| 307 | * limit space consumed in the log rather than by the number of objects being |
| 308 | * logged in order to prevent checkpoint ticket overruns. |
| 309 | * |
| 310 | * Further, use of static reservations through the log grant mechanism is |
| 311 | * problematic. It introduces a lot of complexity (e.g. reserve grant vs write |
| 312 | * grant) and a significant deadlock potential because regranting write space |
| 313 | * can block on log pushes. Hence if we have to regrant log space during a log |
| 314 | * push, we can deadlock. |
| 315 | * |
| 316 | * However, we can avoid this by use of a dynamic "reservation stealing" |
| 317 | * technique during transaction commit whereby unused reservation space in the |
| 318 | * transaction ticket is transferred to the CIL ctx commit ticket to cover the |
| 319 | * space needed by the checkpoint transaction. This means that we never need to |
| 320 | * specifically reserve space for the CIL checkpoint transaction, nor do we |
| 321 | * need to regrant space once the checkpoint completes. This also means the |
| 322 | * checkpoint transaction ticket is specific to the checkpoint context, rather |
| 323 | * than the CIL itself. |
| 324 | * |
| 325 | * With dynamic reservations, we can effectively make up arbitrary limits for |
| 326 | * the checkpoint size so long as they don't violate any other size rules. |
| 327 | * Recovery imposes a rule that no transaction exceed half the log, so we are |
| 328 | * limited by that. Furthermore, the log transaction reservation subsystem |
| 329 | * tries to keep 25% of the log free, so we need to keep below that limit or we |
| 330 | * risk running out of free log space to start any new transactions. |
| 331 | * |
| 332 | * In order to keep background CIL push efficient, we only need to ensure the |
| 333 | * CIL is large enough to maintain sufficient in-memory relogging to avoid |
| 334 | * repeated physical writes of frequently modified metadata. If we allow the CIL |
| 335 | * to grow to a substantial fraction of the log, then we may be pinning hundreds |
| 336 | * of megabytes of metadata in memory until the CIL flushes. This can cause |
| 337 | * issues when we are running low on memory - pinned memory cannot be reclaimed, |
| 338 | * and the CIL consumes a lot of memory. Hence we need to set an upper physical |
| 339 | * size limit for the CIL that limits the maximum amount of memory pinned by the |
| 340 | * CIL but does not limit performance by reducing relogging efficiency |
| 341 | * significantly. |
| 342 | * |
| 343 | * As such, the CIL push threshold ends up being the smaller of two thresholds: |
| 344 | * - a threshold large enough that it allows CIL to be pushed and progress to be |
| 345 | * made without excessive blocking of incoming transaction commits. This is |
| 346 | * defined to be 12.5% of the log space - half the 25% push threshold of the |
| 347 | * AIL. |
| 348 | * - small enough that it doesn't pin excessive amounts of memory but maintains |
| 349 | * close to peak relogging efficiency. This is defined to be 16x the iclog |
| 350 | * buffer window (32MB) as measurements have shown this to be roughly the |
| 351 | * point of diminishing performance increases under highly concurrent |
| 352 | * modification workloads. |
| 353 | * |
| 354 | * To prevent the CIL from overflowing upper commit size bounds, we introduce a |
| 355 | * new threshold at which we block committing transactions until the background |
| 356 | * CIL commit commences and switches to a new context. While this is not a hard |
| 357 | * limit, it forces the process committing a transaction to the CIL to block and |
| 358 | * yeild the CPU, giving the CIL push work a chance to be scheduled and start |
| 359 | * work. This prevents a process running lots of transactions from overfilling |
| 360 | * the CIL because it is not yielding the CPU. We set the blocking limit at |
| 361 | * twice the background push space threshold so we keep in line with the AIL |
| 362 | * push thresholds. |
| 363 | * |
| 364 | * Note: this is not a -hard- limit as blocking is applied after the transaction |
| 365 | * is inserted into the CIL and the push has been triggered. It is largely a |
| 366 | * throttling mechanism that allows the CIL push to be scheduled and run. A hard |
| 367 | * limit will be difficult to implement without introducing global serialisation |
| 368 | * in the CIL commit fast path, and it's not at all clear that we actually need |
| 369 | * such hard limits given the ~7 years we've run without a hard limit before |
| 370 | * finding the first situation where a checkpoint size overflow actually |
| 371 | * occurred. Hence the simple throttle, and an ASSERT check to tell us that |
| 372 | * we've overrun the max size. |
| 373 | */ |
| 374 | #define XLOG_CIL_SPACE_LIMIT(log) \ |
| 375 | min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4) |
| 376 | |
| 377 | #define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \ |
| 378 | (XLOG_CIL_SPACE_LIMIT(log) * 2) |
| 379 | |
| 380 | /* |
| 381 | * ticket grant locks, queues and accounting have their own cachlines |
| 382 | * as these are quite hot and can be operated on concurrently. |
| 383 | */ |
| 384 | struct xlog_grant_head { |
| 385 | spinlock_t lock ____cacheline_aligned_in_smp; |
| 386 | struct list_head waiters; |
| 387 | atomic64_t grant; |
| 388 | }; |
| 389 | |
| 390 | /* |
| 391 | * The reservation head lsn is not made up of a cycle number and block number. |
| 392 | * Instead, it uses a cycle number and byte number. Logs don't expect to |
| 393 | * overflow 31 bits worth of byte offset, so using a byte number will mean |
| 394 | * that round off problems won't occur when releasing partial reservations. |
| 395 | */ |
| 396 | struct xlog { |
| 397 | /* The following fields don't need locking */ |
| 398 | struct xfs_mount *l_mp; /* mount point */ |
| 399 | struct xfs_ail *l_ailp; /* AIL log is working with */ |
| 400 | struct xfs_cil *l_cilp; /* CIL log is working with */ |
| 401 | struct xfs_buftarg *l_targ; /* buftarg of log */ |
| 402 | struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */ |
| 403 | struct delayed_work l_work; /* background flush work */ |
| 404 | long l_opstate; /* operational state */ |
| 405 | uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */ |
| 406 | struct list_head *l_buf_cancel_table; |
| 407 | struct list_head r_dfops; /* recovered log intent items */ |
| 408 | int l_iclog_hsize; /* size of iclog header */ |
| 409 | uint l_sectBBsize; /* sector size in BBs (2^n) */ |
| 410 | int l_iclog_size; /* size of log in bytes */ |
| 411 | int l_iclog_bufs; /* number of iclog buffers */ |
| 412 | xfs_daddr_t l_logBBstart; /* start block of log */ |
| 413 | int l_logsize; /* size of log in bytes */ |
| 414 | int l_logBBsize; /* size of log in BB chunks */ |
| 415 | |
| 416 | /* The following block of fields are changed while holding icloglock */ |
| 417 | wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp; |
| 418 | /* waiting for iclog flush */ |
| 419 | int l_covered_state;/* state of "covering disk |
| 420 | * log entries" */ |
| 421 | struct xlog_in_core *l_iclog; /* head log queue */ |
| 422 | spinlock_t l_icloglock; /* grab to change iclog state */ |
| 423 | int l_curr_cycle; /* Cycle number of log writes */ |
| 424 | int l_prev_cycle; /* Cycle number before last |
| 425 | * block increment */ |
| 426 | int l_curr_block; /* current logical log block */ |
| 427 | int l_prev_block; /* previous logical log block */ |
| 428 | |
| 429 | /* |
| 430 | * l_tail_lsn is atomic so it can be set and read without needing to |
| 431 | * hold specific locks. To avoid operations contending with other hot |
| 432 | * objects, it on a separate cacheline. |
| 433 | */ |
| 434 | /* lsn of 1st LR with unflushed * buffers */ |
| 435 | atomic64_t l_tail_lsn ____cacheline_aligned_in_smp; |
| 436 | |
| 437 | struct xlog_grant_head l_reserve_head; |
| 438 | struct xlog_grant_head l_write_head; |
| 439 | uint64_t l_tail_space; |
| 440 | |
| 441 | struct xfs_kobj l_kobj; |
| 442 | |
| 443 | /* log recovery lsn tracking (for buffer submission */ |
| 444 | xfs_lsn_t l_recovery_lsn; |
| 445 | |
| 446 | uint32_t l_iclog_roundoff;/* padding roundoff */ |
| 447 | }; |
| 448 | |
| 449 | /* |
| 450 | * Bits for operational state |
| 451 | */ |
| 452 | #define XLOG_ACTIVE_RECOVERY 0 /* in the middle of recovery */ |
| 453 | #define XLOG_RECOVERY_NEEDED 1 /* log was recovered */ |
| 454 | #define XLOG_IO_ERROR 2 /* log hit an I/O error, and being |
| 455 | shutdown */ |
| 456 | #define XLOG_TAIL_WARN 3 /* log tail verify warning issued */ |
| 457 | #define XLOG_SHUTDOWN_STARTED 4 /* xlog_force_shutdown() exclusion */ |
| 458 | |
| 459 | static inline bool |
| 460 | xlog_recovery_needed(struct xlog *log) |
| 461 | { |
| 462 | return test_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate); |
| 463 | } |
| 464 | |
| 465 | static inline bool |
| 466 | xlog_in_recovery(struct xlog *log) |
| 467 | { |
| 468 | return test_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate); |
| 469 | } |
| 470 | |
| 471 | static inline bool |
| 472 | xlog_is_shutdown(struct xlog *log) |
| 473 | { |
| 474 | return test_bit(XLOG_IO_ERROR, &log->l_opstate); |
| 475 | } |
| 476 | |
| 477 | /* |
| 478 | * Wait until the xlog_force_shutdown() has marked the log as shut down |
| 479 | * so xlog_is_shutdown() will always return true. |
| 480 | */ |
| 481 | static inline void |
| 482 | xlog_shutdown_wait( |
| 483 | struct xlog *log) |
| 484 | { |
| 485 | wait_var_event(&log->l_opstate, xlog_is_shutdown(log)); |
| 486 | } |
| 487 | |
| 488 | /* common routines */ |
| 489 | extern int |
| 490 | xlog_recover( |
| 491 | struct xlog *log); |
| 492 | extern int |
| 493 | xlog_recover_finish( |
| 494 | struct xlog *log); |
| 495 | extern void |
| 496 | xlog_recover_cancel(struct xlog *); |
| 497 | |
| 498 | __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead, |
| 499 | char *dp, unsigned int hdrsize, unsigned int size); |
| 500 | |
| 501 | extern struct kmem_cache *xfs_log_ticket_cache; |
| 502 | struct xlog_ticket *xlog_ticket_alloc(struct xlog *log, int unit_bytes, |
| 503 | int count, bool permanent); |
| 504 | |
| 505 | void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket); |
| 506 | void xlog_print_trans(struct xfs_trans *); |
| 507 | int xlog_write(struct xlog *log, struct xfs_cil_ctx *ctx, |
| 508 | struct list_head *lv_chain, struct xlog_ticket *tic, |
| 509 | uint32_t len); |
| 510 | void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket); |
| 511 | void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket); |
| 512 | |
| 513 | void xlog_state_switch_iclogs(struct xlog *log, struct xlog_in_core *iclog, |
| 514 | int eventual_size); |
| 515 | int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog, |
| 516 | struct xlog_ticket *ticket); |
| 517 | |
| 518 | /* |
| 519 | * When we crack an atomic LSN, we sample it first so that the value will not |
| 520 | * change while we are cracking it into the component values. This means we |
| 521 | * will always get consistent component values to work from. This should always |
| 522 | * be used to sample and crack LSNs that are stored and updated in atomic |
| 523 | * variables. |
| 524 | */ |
| 525 | static inline void |
| 526 | xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block) |
| 527 | { |
| 528 | xfs_lsn_t val = atomic64_read(lsn); |
| 529 | |
| 530 | *cycle = CYCLE_LSN(val); |
| 531 | *block = BLOCK_LSN(val); |
| 532 | } |
| 533 | |
| 534 | /* |
| 535 | * Calculate and assign a value to an atomic LSN variable from component pieces. |
| 536 | */ |
| 537 | static inline void |
| 538 | xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block) |
| 539 | { |
| 540 | atomic64_set(v: lsn, i: xlog_assign_lsn(cycle, block)); |
| 541 | } |
| 542 | |
| 543 | /* |
| 544 | * Committed Item List interfaces |
| 545 | */ |
| 546 | int xlog_cil_init(struct xlog *log); |
| 547 | void xlog_cil_init_post_recovery(struct xlog *log); |
| 548 | void xlog_cil_destroy(struct xlog *log); |
| 549 | bool xlog_cil_empty(struct xlog *log); |
| 550 | void xlog_cil_commit(struct xlog *log, struct xfs_trans *tp, |
| 551 | xfs_csn_t *commit_seq, bool regrant); |
| 552 | void xlog_cil_set_ctx_write_state(struct xfs_cil_ctx *ctx, |
| 553 | struct xlog_in_core *iclog); |
| 554 | |
| 555 | |
| 556 | /* |
| 557 | * CIL force routines |
| 558 | */ |
| 559 | void xlog_cil_flush(struct xlog *log); |
| 560 | xfs_lsn_t xlog_cil_force_seq(struct xlog *log, xfs_csn_t sequence); |
| 561 | |
| 562 | static inline void |
| 563 | xlog_cil_force(struct xlog *log) |
| 564 | { |
| 565 | xlog_cil_force_seq(log, log->l_cilp->xc_current_sequence); |
| 566 | } |
| 567 | |
| 568 | /* |
| 569 | * Wrapper function for waiting on a wait queue serialised against wakeups |
| 570 | * by a spinlock. This matches the semantics of all the wait queues used in the |
| 571 | * log code. |
| 572 | */ |
| 573 | static inline void |
| 574 | xlog_wait( |
| 575 | struct wait_queue_head *wq, |
| 576 | struct spinlock *lock) |
| 577 | __releases(lock) |
| 578 | { |
| 579 | DECLARE_WAITQUEUE(wait, current); |
| 580 | |
| 581 | add_wait_queue_exclusive(wq_head: wq, wq_entry: &wait); |
| 582 | __set_current_state(TASK_UNINTERRUPTIBLE); |
| 583 | spin_unlock(lock); |
| 584 | schedule(); |
| 585 | remove_wait_queue(wq_head: wq, wq_entry: &wait); |
| 586 | } |
| 587 | |
| 588 | int xlog_wait_on_iclog(struct xlog_in_core *iclog) |
| 589 | __releases(iclog->ic_log->l_icloglock); |
| 590 | |
| 591 | /* Calculate the distance between two LSNs in bytes */ |
| 592 | static inline uint64_t |
| 593 | xlog_lsn_sub( |
| 594 | struct xlog *log, |
| 595 | xfs_lsn_t high, |
| 596 | xfs_lsn_t low) |
| 597 | { |
| 598 | uint32_t hi_cycle = CYCLE_LSN(high); |
| 599 | uint32_t hi_block = BLOCK_LSN(high); |
| 600 | uint32_t lo_cycle = CYCLE_LSN(low); |
| 601 | uint32_t lo_block = BLOCK_LSN(low); |
| 602 | |
| 603 | if (hi_cycle == lo_cycle) |
| 604 | return BBTOB(hi_block - lo_block); |
| 605 | ASSERT((hi_cycle == lo_cycle + 1) || xlog_is_shutdown(log)); |
| 606 | return (uint64_t)log->l_logsize - BBTOB(lo_block - hi_block); |
| 607 | } |
| 608 | |
| 609 | void xlog_grant_return_space(struct xlog *log, xfs_lsn_t old_head, |
| 610 | xfs_lsn_t new_head); |
| 611 | |
| 612 | /* |
| 613 | * The LSN is valid so long as it is behind the current LSN. If it isn't, this |
| 614 | * means that the next log record that includes this metadata could have a |
| 615 | * smaller LSN. In turn, this means that the modification in the log would not |
| 616 | * replay. |
| 617 | */ |
| 618 | static inline bool |
| 619 | xlog_valid_lsn( |
| 620 | struct xlog *log, |
| 621 | xfs_lsn_t lsn) |
| 622 | { |
| 623 | int cur_cycle; |
| 624 | int cur_block; |
| 625 | bool valid = true; |
| 626 | |
| 627 | /* |
| 628 | * First, sample the current lsn without locking to avoid added |
| 629 | * contention from metadata I/O. The current cycle and block are updated |
| 630 | * (in xlog_state_switch_iclogs()) and read here in a particular order |
| 631 | * to avoid false negatives (e.g., thinking the metadata LSN is valid |
| 632 | * when it is not). |
| 633 | * |
| 634 | * The current block is always rewound before the cycle is bumped in |
| 635 | * xlog_state_switch_iclogs() to ensure the current LSN is never seen in |
| 636 | * a transiently forward state. Instead, we can see the LSN in a |
| 637 | * transiently behind state if we happen to race with a cycle wrap. |
| 638 | */ |
| 639 | cur_cycle = READ_ONCE(log->l_curr_cycle); |
| 640 | smp_rmb(); |
| 641 | cur_block = READ_ONCE(log->l_curr_block); |
| 642 | |
| 643 | if ((CYCLE_LSN(lsn) > cur_cycle) || |
| 644 | (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) { |
| 645 | /* |
| 646 | * If the metadata LSN appears invalid, it's possible the check |
| 647 | * above raced with a wrap to the next log cycle. Grab the lock |
| 648 | * to check for sure. |
| 649 | */ |
| 650 | spin_lock(lock: &log->l_icloglock); |
| 651 | cur_cycle = log->l_curr_cycle; |
| 652 | cur_block = log->l_curr_block; |
| 653 | spin_unlock(lock: &log->l_icloglock); |
| 654 | |
| 655 | if ((CYCLE_LSN(lsn) > cur_cycle) || |
| 656 | (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) |
| 657 | valid = false; |
| 658 | } |
| 659 | |
| 660 | return valid; |
| 661 | } |
| 662 | |
| 663 | /* |
| 664 | * Log vector and shadow buffers can be large, so we need to use kvmalloc() here |
| 665 | * to ensure success. Unfortunately, kvmalloc() only allows GFP_KERNEL contexts |
| 666 | * to fall back to vmalloc, so we can't actually do anything useful with gfp |
| 667 | * flags to control the kmalloc() behaviour within kvmalloc(). Hence kmalloc() |
| 668 | * will do direct reclaim and compaction in the slow path, both of which are |
| 669 | * horrendously expensive. We just want kmalloc to fail fast and fall back to |
| 670 | * vmalloc if it can't get something straight away from the free lists or |
| 671 | * buddy allocator. Hence we have to open code kvmalloc outselves here. |
| 672 | * |
| 673 | * This assumes that the caller uses memalloc_nofs_save task context here, so |
| 674 | * despite the use of GFP_KERNEL here, we are going to be doing GFP_NOFS |
| 675 | * allocations. This is actually the only way to make vmalloc() do GFP_NOFS |
| 676 | * allocations, so lets just all pretend this is a GFP_KERNEL context |
| 677 | * operation.... |
| 678 | */ |
| 679 | static inline void * |
| 680 | xlog_kvmalloc( |
| 681 | size_t buf_size) |
| 682 | { |
| 683 | gfp_t flags = GFP_KERNEL; |
| 684 | void *p; |
| 685 | |
| 686 | flags &= ~__GFP_DIRECT_RECLAIM; |
| 687 | flags |= __GFP_NOWARN | __GFP_NORETRY; |
| 688 | do { |
| 689 | p = kmalloc(buf_size, flags); |
| 690 | if (!p) |
| 691 | p = vmalloc(buf_size); |
| 692 | } while (!p); |
| 693 | |
| 694 | return p; |
| 695 | } |
| 696 | |
| 697 | /* |
| 698 | * Given a count of iovecs and space for a log item, compute the space we need |
| 699 | * in the log to store that data plus the log headers. |
| 700 | */ |
| 701 | static inline unsigned int |
| 702 | xlog_item_space( |
| 703 | unsigned int niovecs, |
| 704 | unsigned int nbytes) |
| 705 | { |
| 706 | nbytes += niovecs * (sizeof(uint64_t) + sizeof(struct xlog_op_header)); |
| 707 | return round_up(nbytes, sizeof(uint64_t)); |
| 708 | } |
| 709 | |
| 710 | /* |
| 711 | * Cycles over XLOG_CYCLE_DATA_SIZE overflow into the extended header that was |
| 712 | * added for v2 logs. Addressing for the cycles array there is off by one, |
| 713 | * because the first batch of cycles is in the original header. |
| 714 | */ |
| 715 | static inline __be32 *xlog_cycle_data(struct xlog_rec_header *rhead, unsigned i) |
| 716 | { |
| 717 | if (i >= XLOG_CYCLE_DATA_SIZE) { |
| 718 | unsigned j = i / XLOG_CYCLE_DATA_SIZE; |
| 719 | unsigned k = i % XLOG_CYCLE_DATA_SIZE; |
| 720 | |
| 721 | return &rhead->h_ext[j - 1].xh_cycle_data[k]; |
| 722 | } |
| 723 | |
| 724 | return &rhead->h_cycle_data[i]; |
| 725 | } |
| 726 | |
| 727 | #endif /* __XFS_LOG_PRIV_H__ */ |
| 728 | |