1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
6 #ifndef __XFS_LOG_PRIV_H__
7 #define __XFS_LOG_PRIV_H__
15 * Flags for log structure
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
21 #define XLOG_TAIL_WARN 0x10 /* log tail verify warning issued */
24 * get client id from packed copy.
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.
30 * later on this four byte chunk is treated as an int and the
31 * client id is pulled out.
33 * this has endian issues, of course.
35 static inline uint xlog_get_client_id(__be32 i)
37 return be32_to_cpu(i) >> 24;
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 */
53 #define XLOG_STATE_STRINGS \
54 { XLOG_STATE_ACTIVE, "XLOG_STATE_ACTIVE" }, \
55 { XLOG_STATE_WANT_SYNC, "XLOG_STATE_WANT_SYNC" }, \
56 { XLOG_STATE_SYNCING, "XLOG_STATE_SYNCING" }, \
57 { XLOG_STATE_DONE_SYNC, "XLOG_STATE_DONE_SYNC" }, \
58 { XLOG_STATE_CALLBACK, "XLOG_STATE_CALLBACK" }, \
59 { XLOG_STATE_DIRTY, "XLOG_STATE_DIRTY" }, \
60 { XLOG_STATE_IOERROR, "XLOG_STATE_IOERROR" }
65 #define XLOG_ICL_NEED_FLUSH (1 << 0) /* iclog needs REQ_PREFLUSH */
66 #define XLOG_ICL_NEED_FUA (1 << 1) /* iclog needs REQ_FUA */
68 #define XLOG_ICL_STRINGS \
69 { XLOG_ICL_NEED_FLUSH, "XLOG_ICL_NEED_FLUSH" }, \
70 { XLOG_ICL_NEED_FUA, "XLOG_ICL_NEED_FUA" }
76 #define XLOG_TIC_PERM_RESERV 0x1 /* permanent reservation */
78 #define XLOG_TIC_FLAGS \
79 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }
82 * Below are states for covering allocation transactions.
83 * By covering, we mean changing the h_tail_lsn in the last on-disk
84 * log write such that no allocation transactions will be re-done during
85 * recovery after a system crash. Recovery starts at the last on-disk
88 * These states are used to insert dummy log entries to cover
89 * space allocation transactions which can undo non-transactional changes
90 * after a crash. Writes to a file with space
91 * already allocated do not result in any transactions. Allocations
92 * might include space beyond the EOF. So if we just push the EOF a
93 * little, the last transaction for the file could contain the wrong
94 * size. If there is no file system activity, after an allocation
95 * transaction, and the system crashes, the allocation transaction
96 * will get replayed and the file will be truncated. This could
97 * be hours/days/... after the allocation occurred.
99 * The fix for this is to do two dummy transactions when the
100 * system is idle. We need two dummy transaction because the h_tail_lsn
101 * in the log record header needs to point beyond the last possible
102 * non-dummy transaction. The first dummy changes the h_tail_lsn to
103 * the first transaction before the dummy. The second dummy causes
104 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn.
106 * These dummy transactions get committed when everything
107 * is idle (after there has been some activity).
109 * There are 5 states used to control this.
111 * IDLE -- no logging has been done on the file system or
112 * we are done covering previous transactions.
113 * NEED -- logging has occurred and we need a dummy transaction
114 * when the log becomes idle.
115 * DONE -- we were in the NEED state and have committed a dummy
117 * NEED2 -- we detected that a dummy transaction has gone to the
118 * on disk log with no other transactions.
119 * DONE2 -- we committed a dummy transaction when in the NEED2 state.
121 * There are two places where we switch states:
123 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
124 * We commit the dummy transaction and switch to DONE or DONE2,
125 * respectively. In all other states, we don't do anything.
127 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
129 * No matter what state we are in, if this isn't the dummy
130 * transaction going out, the next state is NEED.
131 * So, if we aren't in the DONE or DONE2 states, the next state
132 * is NEED. We can't be finishing a write of the dummy record
133 * unless it was committed and the state switched to DONE or DONE2.
135 * If we are in the DONE state and this was a write of the
136 * dummy transaction, we move to NEED2.
138 * If we are in the DONE2 state and this was a write of the
139 * dummy transaction, we move to IDLE.
142 * Writing only one dummy transaction can get appended to
143 * one file space allocation. When this happens, the log recovery
144 * code replays the space allocation and a file could be truncated.
145 * This is why we have the NEED2 and DONE2 states before going idle.
148 #define XLOG_STATE_COVER_IDLE 0
149 #define XLOG_STATE_COVER_NEED 1
150 #define XLOG_STATE_COVER_DONE 2
151 #define XLOG_STATE_COVER_NEED2 3
152 #define XLOG_STATE_COVER_DONE2 4
154 #define XLOG_COVER_OPS 5
156 /* Ticket reservation region accounting */
157 #define XLOG_TIC_LEN_MAX 15
161 * As would be stored in xfs_log_iovec but without the i_addr which
162 * we don't care about.
164 typedef struct xlog_res {
165 uint r_len; /* region length :4 */
166 uint r_type; /* region's transaction type :4 */
169 typedef struct xlog_ticket {
170 struct list_head t_queue; /* reserve/write queue */
171 struct task_struct *t_task; /* task that owns this ticket */
172 xlog_tid_t t_tid; /* transaction identifier : 4 */
173 atomic_t t_ref; /* ticket reference count : 4 */
174 int t_curr_res; /* current reservation in bytes : 4 */
175 int t_unit_res; /* unit reservation in bytes : 4 */
176 char t_ocnt; /* original count : 1 */
177 char t_cnt; /* current count : 1 */
178 char t_clientid; /* who does this belong to; : 1 */
179 char t_flags; /* properties of reservation : 1 */
181 /* reservation array fields */
182 uint t_res_num; /* num in array : 4 */
183 uint t_res_num_ophdrs; /* num op hdrs : 4 */
184 uint t_res_arr_sum; /* array sum : 4 */
185 uint t_res_o_flow; /* sum overflow : 4 */
186 xlog_res_t t_res_arr[XLOG_TIC_LEN_MAX]; /* array of res : 8 * 15 */
190 * - A log record header is 512 bytes. There is plenty of room to grow the
191 * xlog_rec_header_t into the reserved space.
192 * - ic_data follows, so a write to disk can start at the beginning of
194 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
195 * - ic_next is the pointer to the next iclog in the ring.
196 * - ic_log is a pointer back to the global log structure.
197 * - ic_size is the full size of the log buffer, minus the cycle headers.
198 * - ic_offset is the current number of bytes written to in this iclog.
199 * - ic_refcnt is bumped when someone is writing to the log.
200 * - ic_state is the state of the iclog.
202 * Because of cacheline contention on large machines, we need to separate
203 * various resources onto different cachelines. To start with, make the
204 * structure cacheline aligned. The following fields can be contended on
205 * by independent processes:
209 * - fields protected by the global l_icloglock
211 * so we need to ensure that these fields are located in separate cachelines.
212 * We'll put all the read-only and l_icloglock fields in the first cacheline,
213 * and move everything else out to subsequent cachelines.
215 typedef struct xlog_in_core {
216 wait_queue_head_t ic_force_wait;
217 wait_queue_head_t ic_write_wait;
218 struct xlog_in_core *ic_next;
219 struct xlog_in_core *ic_prev;
223 enum xlog_iclog_state ic_state;
224 unsigned int ic_flags;
225 char *ic_datap; /* pointer to iclog data */
226 struct list_head ic_callbacks;
228 /* reference counts need their own cacheline */
229 atomic_t ic_refcnt ____cacheline_aligned_in_smp;
230 xlog_in_core_2_t *ic_data;
231 #define ic_header ic_data->hic_header
233 bool ic_fail_crc : 1;
235 struct semaphore ic_sema;
236 struct work_struct ic_end_io_work;
238 struct bio_vec ic_bvec[];
242 * The CIL context is used to aggregate per-transaction details as well be
243 * passed to the iclog for checkpoint post-commit processing. After being
244 * passed to the iclog, another context needs to be allocated for tracking the
245 * next set of transactions to be aggregated into a checkpoint.
251 xfs_csn_t sequence; /* chkpt sequence # */
252 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */
253 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */
254 struct xlog_ticket *ticket; /* chkpt ticket */
255 int nvecs; /* number of regions */
256 int space_used; /* aggregate size of regions */
257 struct list_head busy_extents; /* busy extents in chkpt */
258 struct xfs_log_vec *lv_chain; /* logvecs being pushed */
259 struct list_head iclog_entry;
260 struct list_head committing; /* ctx committing list */
261 struct work_struct discard_endio_work;
265 * Committed Item List structure
267 * This structure is used to track log items that have been committed but not
268 * yet written into the log. It is used only when the delayed logging mount
271 * This structure tracks the list of committing checkpoint contexts so
272 * we can avoid the problem of having to hold out new transactions during a
273 * flush until we have a the commit record LSN of the checkpoint. We can
274 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
275 * sequence match and extract the commit LSN directly from there. If the
276 * checkpoint is still in the process of committing, we can block waiting for
277 * the commit LSN to be determined as well. This should make synchronous
278 * operations almost as efficient as the old logging methods.
282 struct list_head xc_cil;
283 spinlock_t xc_cil_lock;
285 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp;
286 struct xfs_cil_ctx *xc_ctx;
288 spinlock_t xc_push_lock ____cacheline_aligned_in_smp;
289 xfs_csn_t xc_push_seq;
290 struct list_head xc_committing;
291 wait_queue_head_t xc_commit_wait;
292 xfs_csn_t xc_current_sequence;
293 struct work_struct xc_push_work;
294 wait_queue_head_t xc_push_wait; /* background push throttle */
295 } ____cacheline_aligned_in_smp;
298 * The amount of log space we allow the CIL to aggregate is difficult to size.
299 * Whatever we choose, we have to make sure we can get a reservation for the
300 * log space effectively, that it is large enough to capture sufficient
301 * relogging to reduce log buffer IO significantly, but it is not too large for
302 * the log or induces too much latency when writing out through the iclogs. We
303 * track both space consumed and the number of vectors in the checkpoint
304 * context, so we need to decide which to use for limiting.
306 * Every log buffer we write out during a push needs a header reserved, which
307 * is at least one sector and more for v2 logs. Hence we need a reservation of
308 * at least 512 bytes per 32k of log space just for the LR headers. That means
309 * 16KB of reservation per megabyte of delayed logging space we will consume,
310 * plus various headers. The number of headers will vary based on the num of
311 * io vectors, so limiting on a specific number of vectors is going to result
312 * in transactions of varying size. IOWs, it is more consistent to track and
313 * limit space consumed in the log rather than by the number of objects being
314 * logged in order to prevent checkpoint ticket overruns.
316 * Further, use of static reservations through the log grant mechanism is
317 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
318 * grant) and a significant deadlock potential because regranting write space
319 * can block on log pushes. Hence if we have to regrant log space during a log
320 * push, we can deadlock.
322 * However, we can avoid this by use of a dynamic "reservation stealing"
323 * technique during transaction commit whereby unused reservation space in the
324 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
325 * space needed by the checkpoint transaction. This means that we never need to
326 * specifically reserve space for the CIL checkpoint transaction, nor do we
327 * need to regrant space once the checkpoint completes. This also means the
328 * checkpoint transaction ticket is specific to the checkpoint context, rather
329 * than the CIL itself.
331 * With dynamic reservations, we can effectively make up arbitrary limits for
332 * the checkpoint size so long as they don't violate any other size rules.
333 * Recovery imposes a rule that no transaction exceed half the log, so we are
334 * limited by that. Furthermore, the log transaction reservation subsystem
335 * tries to keep 25% of the log free, so we need to keep below that limit or we
336 * risk running out of free log space to start any new transactions.
338 * In order to keep background CIL push efficient, we only need to ensure the
339 * CIL is large enough to maintain sufficient in-memory relogging to avoid
340 * repeated physical writes of frequently modified metadata. If we allow the CIL
341 * to grow to a substantial fraction of the log, then we may be pinning hundreds
342 * of megabytes of metadata in memory until the CIL flushes. This can cause
343 * issues when we are running low on memory - pinned memory cannot be reclaimed,
344 * and the CIL consumes a lot of memory. Hence we need to set an upper physical
345 * size limit for the CIL that limits the maximum amount of memory pinned by the
346 * CIL but does not limit performance by reducing relogging efficiency
349 * As such, the CIL push threshold ends up being the smaller of two thresholds:
350 * - a threshold large enough that it allows CIL to be pushed and progress to be
351 * made without excessive blocking of incoming transaction commits. This is
352 * defined to be 12.5% of the log space - half the 25% push threshold of the
354 * - small enough that it doesn't pin excessive amounts of memory but maintains
355 * close to peak relogging efficiency. This is defined to be 16x the iclog
356 * buffer window (32MB) as measurements have shown this to be roughly the
357 * point of diminishing performance increases under highly concurrent
358 * modification workloads.
360 * To prevent the CIL from overflowing upper commit size bounds, we introduce a
361 * new threshold at which we block committing transactions until the background
362 * CIL commit commences and switches to a new context. While this is not a hard
363 * limit, it forces the process committing a transaction to the CIL to block and
364 * yeild the CPU, giving the CIL push work a chance to be scheduled and start
365 * work. This prevents a process running lots of transactions from overfilling
366 * the CIL because it is not yielding the CPU. We set the blocking limit at
367 * twice the background push space threshold so we keep in line with the AIL
370 * Note: this is not a -hard- limit as blocking is applied after the transaction
371 * is inserted into the CIL and the push has been triggered. It is largely a
372 * throttling mechanism that allows the CIL push to be scheduled and run. A hard
373 * limit will be difficult to implement without introducing global serialisation
374 * in the CIL commit fast path, and it's not at all clear that we actually need
375 * such hard limits given the ~7 years we've run without a hard limit before
376 * finding the first situation where a checkpoint size overflow actually
377 * occurred. Hence the simple throttle, and an ASSERT check to tell us that
378 * we've overrun the max size.
380 #define XLOG_CIL_SPACE_LIMIT(log) \
381 min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4)
383 #define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \
384 (XLOG_CIL_SPACE_LIMIT(log) * 2)
387 * ticket grant locks, queues and accounting have their own cachlines
388 * as these are quite hot and can be operated on concurrently.
390 struct xlog_grant_head {
391 spinlock_t lock ____cacheline_aligned_in_smp;
392 struct list_head waiters;
397 * The reservation head lsn is not made up of a cycle number and block number.
398 * Instead, it uses a cycle number and byte number. Logs don't expect to
399 * overflow 31 bits worth of byte offset, so using a byte number will mean
400 * that round off problems won't occur when releasing partial reservations.
403 /* The following fields don't need locking */
404 struct xfs_mount *l_mp; /* mount point */
405 struct xfs_ail *l_ailp; /* AIL log is working with */
406 struct xfs_cil *l_cilp; /* CIL log is working with */
407 struct xfs_buftarg *l_targ; /* buftarg of log */
408 struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */
409 struct delayed_work l_work; /* background flush work */
411 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
412 struct list_head *l_buf_cancel_table;
413 int l_iclog_hsize; /* size of iclog header */
414 int l_iclog_heads; /* # of iclog header sectors */
415 uint l_sectBBsize; /* sector size in BBs (2^n) */
416 int l_iclog_size; /* size of log in bytes */
417 int l_iclog_bufs; /* number of iclog buffers */
418 xfs_daddr_t l_logBBstart; /* start block of log */
419 int l_logsize; /* size of log in bytes */
420 int l_logBBsize; /* size of log in BB chunks */
422 /* The following block of fields are changed while holding icloglock */
423 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
424 /* waiting for iclog flush */
425 int l_covered_state;/* state of "covering disk
427 xlog_in_core_t *l_iclog; /* head log queue */
428 spinlock_t l_icloglock; /* grab to change iclog state */
429 int l_curr_cycle; /* Cycle number of log writes */
430 int l_prev_cycle; /* Cycle number before last
432 int l_curr_block; /* current logical log block */
433 int l_prev_block; /* previous logical log block */
436 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and
437 * read without needing to hold specific locks. To avoid operations
438 * contending with other hot objects, place each of them on a separate
441 /* lsn of last LR on disk */
442 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp;
443 /* lsn of 1st LR with unflushed * buffers */
444 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
446 struct xlog_grant_head l_reserve_head;
447 struct xlog_grant_head l_write_head;
449 struct xfs_kobj l_kobj;
451 /* The following field are used for debugging; need to hold icloglock */
453 void *l_iclog_bak[XLOG_MAX_ICLOGS];
455 /* log recovery lsn tracking (for buffer submission */
456 xfs_lsn_t l_recovery_lsn;
458 uint32_t l_iclog_roundoff;/* padding roundoff */
461 #define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
462 ((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE))
464 #define XLOG_FORCED_SHUTDOWN(log) \
465 (unlikely((log)->l_flags & XLOG_IO_ERROR))
467 /* common routines */
475 xlog_recover_cancel(struct xlog *);
477 extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
480 extern kmem_zone_t *xfs_log_ticket_zone;
490 xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes)
497 void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
498 void xlog_print_trans(struct xfs_trans *);
499 int xlog_write(struct xlog *log, struct xfs_log_vec *log_vector,
500 struct xlog_ticket *tic, xfs_lsn_t *start_lsn,
501 struct xlog_in_core **commit_iclog, uint optype);
502 int xlog_commit_record(struct xlog *log, struct xlog_ticket *ticket,
503 struct xlog_in_core **iclog, xfs_lsn_t *lsn);
504 void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket);
505 void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket);
507 int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog,
508 xfs_lsn_t log_tail_lsn);
511 * When we crack an atomic LSN, we sample it first so that the value will not
512 * change while we are cracking it into the component values. This means we
513 * will always get consistent component values to work from. This should always
514 * be used to sample and crack LSNs that are stored and updated in atomic
518 xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
520 xfs_lsn_t val = atomic64_read(lsn);
522 *cycle = CYCLE_LSN(val);
523 *block = BLOCK_LSN(val);
527 * Calculate and assign a value to an atomic LSN variable from component pieces.
530 xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
532 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
536 * When we crack the grant head, we sample it first so that the value will not
537 * change while we are cracking it into the component values. This means we
538 * will always get consistent component values to work from.
541 xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
544 *space = val & 0xffffffff;
548 xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
550 xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
553 static inline int64_t
554 xlog_assign_grant_head_val(int cycle, int space)
556 return ((int64_t)cycle << 32) | space;
560 xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
562 atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
566 * Committed Item List interfaces
568 int xlog_cil_init(struct xlog *log);
569 void xlog_cil_init_post_recovery(struct xlog *log);
570 void xlog_cil_destroy(struct xlog *log);
571 bool xlog_cil_empty(struct xlog *log);
572 void xlog_cil_commit(struct xlog *log, struct xfs_trans *tp,
573 xfs_csn_t *commit_seq, bool regrant);
578 xfs_lsn_t xlog_cil_force_seq(struct xlog *log, xfs_csn_t sequence);
581 xlog_cil_force(struct xlog *log)
583 xlog_cil_force_seq(log, log->l_cilp->xc_current_sequence);
587 * Wrapper function for waiting on a wait queue serialised against wakeups
588 * by a spinlock. This matches the semantics of all the wait queues used in the
593 struct wait_queue_head *wq,
594 struct spinlock *lock)
597 DECLARE_WAITQUEUE(wait, current);
599 add_wait_queue_exclusive(wq, &wait);
600 __set_current_state(TASK_UNINTERRUPTIBLE);
603 remove_wait_queue(wq, &wait);
606 int xlog_wait_on_iclog(struct xlog_in_core *iclog);
609 * The LSN is valid so long as it is behind the current LSN. If it isn't, this
610 * means that the next log record that includes this metadata could have a
611 * smaller LSN. In turn, this means that the modification in the log would not
624 * First, sample the current lsn without locking to avoid added
625 * contention from metadata I/O. The current cycle and block are updated
626 * (in xlog_state_switch_iclogs()) and read here in a particular order
627 * to avoid false negatives (e.g., thinking the metadata LSN is valid
630 * The current block is always rewound before the cycle is bumped in
631 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in
632 * a transiently forward state. Instead, we can see the LSN in a
633 * transiently behind state if we happen to race with a cycle wrap.
635 cur_cycle = READ_ONCE(log->l_curr_cycle);
637 cur_block = READ_ONCE(log->l_curr_block);
639 if ((CYCLE_LSN(lsn) > cur_cycle) ||
640 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
642 * If the metadata LSN appears invalid, it's possible the check
643 * above raced with a wrap to the next log cycle. Grab the lock
646 spin_lock(&log->l_icloglock);
647 cur_cycle = log->l_curr_cycle;
648 cur_block = log->l_curr_block;
649 spin_unlock(&log->l_icloglock);
651 if ((CYCLE_LSN(lsn) > cur_cycle) ||
652 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
659 #endif /* __XFS_LOG_PRIV_H__ */