Loading...
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
11struct xfs_buf;
12struct xlog;
13struct xlog_ticket;
14struct 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 */
28static 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 */
36enum 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
147typedef 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} xlog_ticket_t;
159
160/*
161 * - A log record header is 512 bytes. There is plenty of room to grow the
162 * xlog_rec_header_t into the reserved space.
163 * - ic_data follows, so a write to disk can start at the beginning of
164 * the iclog.
165 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
166 * - ic_next is the pointer to the next iclog in the ring.
167 * - ic_log is a pointer back to the global log structure.
168 * - ic_size is the full size of the log buffer, minus the cycle headers.
169 * - ic_offset is the current number of bytes written to in this iclog.
170 * - ic_refcnt is bumped when someone is writing to the log.
171 * - ic_state is the state of the iclog.
172 *
173 * Because of cacheline contention on large machines, we need to separate
174 * various resources onto different cachelines. To start with, make the
175 * structure cacheline aligned. The following fields can be contended on
176 * by independent processes:
177 *
178 * - ic_callbacks
179 * - ic_refcnt
180 * - fields protected by the global l_icloglock
181 *
182 * so we need to ensure that these fields are located in separate cachelines.
183 * We'll put all the read-only and l_icloglock fields in the first cacheline,
184 * and move everything else out to subsequent cachelines.
185 */
186typedef struct xlog_in_core {
187 wait_queue_head_t ic_force_wait;
188 wait_queue_head_t ic_write_wait;
189 struct xlog_in_core *ic_next;
190 struct xlog_in_core *ic_prev;
191 struct xlog *ic_log;
192 u32 ic_size;
193 u32 ic_offset;
194 enum xlog_iclog_state ic_state;
195 unsigned int ic_flags;
196 void *ic_datap; /* pointer to iclog data */
197 struct list_head ic_callbacks;
198
199 /* reference counts need their own cacheline */
200 atomic_t ic_refcnt ____cacheline_aligned_in_smp;
201 xlog_in_core_2_t *ic_data;
202#define ic_header ic_data->hic_header
203#ifdef DEBUG
204 bool ic_fail_crc : 1;
205#endif
206 struct semaphore ic_sema;
207 struct work_struct ic_end_io_work;
208 struct bio ic_bio;
209 struct bio_vec ic_bvec[];
210} xlog_in_core_t;
211
212/*
213 * The CIL context is used to aggregate per-transaction details as well be
214 * passed to the iclog for checkpoint post-commit processing. After being
215 * passed to the iclog, another context needs to be allocated for tracking the
216 * next set of transactions to be aggregated into a checkpoint.
217 */
218struct xfs_cil;
219
220struct xfs_cil_ctx {
221 struct xfs_cil *cil;
222 xfs_csn_t sequence; /* chkpt sequence # */
223 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */
224 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */
225 struct xlog_in_core *commit_iclog;
226 struct xlog_ticket *ticket; /* chkpt ticket */
227 atomic_t space_used; /* aggregate size of regions */
228 struct xfs_busy_extents busy_extents;
229 struct list_head log_items; /* log items in chkpt */
230 struct list_head lv_chain; /* logvecs being pushed */
231 struct list_head iclog_entry;
232 struct list_head committing; /* ctx committing list */
233 struct work_struct push_work;
234 atomic_t order_id;
235
236 /*
237 * CPUs that could have added items to the percpu CIL data. Access is
238 * coordinated with xc_ctx_lock.
239 */
240 struct cpumask cil_pcpmask;
241};
242
243/*
244 * Per-cpu CIL tracking items
245 */
246struct xlog_cil_pcp {
247 int32_t space_used;
248 uint32_t space_reserved;
249 struct list_head busy_extents;
250 struct list_head log_items;
251};
252
253/*
254 * Committed Item List structure
255 *
256 * This structure is used to track log items that have been committed but not
257 * yet written into the log. It is used only when the delayed logging mount
258 * option is enabled.
259 *
260 * This structure tracks the list of committing checkpoint contexts so
261 * we can avoid the problem of having to hold out new transactions during a
262 * flush until we have a the commit record LSN of the checkpoint. We can
263 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
264 * sequence match and extract the commit LSN directly from there. If the
265 * checkpoint is still in the process of committing, we can block waiting for
266 * the commit LSN to be determined as well. This should make synchronous
267 * operations almost as efficient as the old logging methods.
268 */
269struct xfs_cil {
270 struct xlog *xc_log;
271 unsigned long xc_flags;
272 atomic_t xc_iclog_hdrs;
273 struct workqueue_struct *xc_push_wq;
274
275 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp;
276 struct xfs_cil_ctx *xc_ctx;
277
278 spinlock_t xc_push_lock ____cacheline_aligned_in_smp;
279 xfs_csn_t xc_push_seq;
280 bool xc_push_commit_stable;
281 struct list_head xc_committing;
282 wait_queue_head_t xc_commit_wait;
283 wait_queue_head_t xc_start_wait;
284 xfs_csn_t xc_current_sequence;
285 wait_queue_head_t xc_push_wait; /* background push throttle */
286
287 void __percpu *xc_pcp; /* percpu CIL structures */
288} ____cacheline_aligned_in_smp;
289
290/* xc_flags bit values */
291#define XLOG_CIL_EMPTY 1
292#define XLOG_CIL_PCP_SPACE 2
293
294/*
295 * The amount of log space we allow the CIL to aggregate is difficult to size.
296 * Whatever we choose, we have to make sure we can get a reservation for the
297 * log space effectively, that it is large enough to capture sufficient
298 * relogging to reduce log buffer IO significantly, but it is not too large for
299 * the log or induces too much latency when writing out through the iclogs. We
300 * track both space consumed and the number of vectors in the checkpoint
301 * context, so we need to decide which to use for limiting.
302 *
303 * Every log buffer we write out during a push needs a header reserved, which
304 * is at least one sector and more for v2 logs. Hence we need a reservation of
305 * at least 512 bytes per 32k of log space just for the LR headers. That means
306 * 16KB of reservation per megabyte of delayed logging space we will consume,
307 * plus various headers. The number of headers will vary based on the num of
308 * io vectors, so limiting on a specific number of vectors is going to result
309 * in transactions of varying size. IOWs, it is more consistent to track and
310 * limit space consumed in the log rather than by the number of objects being
311 * logged in order to prevent checkpoint ticket overruns.
312 *
313 * Further, use of static reservations through the log grant mechanism is
314 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
315 * grant) and a significant deadlock potential because regranting write space
316 * can block on log pushes. Hence if we have to regrant log space during a log
317 * push, we can deadlock.
318 *
319 * However, we can avoid this by use of a dynamic "reservation stealing"
320 * technique during transaction commit whereby unused reservation space in the
321 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
322 * space needed by the checkpoint transaction. This means that we never need to
323 * specifically reserve space for the CIL checkpoint transaction, nor do we
324 * need to regrant space once the checkpoint completes. This also means the
325 * checkpoint transaction ticket is specific to the checkpoint context, rather
326 * than the CIL itself.
327 *
328 * With dynamic reservations, we can effectively make up arbitrary limits for
329 * the checkpoint size so long as they don't violate any other size rules.
330 * Recovery imposes a rule that no transaction exceed half the log, so we are
331 * limited by that. Furthermore, the log transaction reservation subsystem
332 * tries to keep 25% of the log free, so we need to keep below that limit or we
333 * risk running out of free log space to start any new transactions.
334 *
335 * In order to keep background CIL push efficient, we only need to ensure the
336 * CIL is large enough to maintain sufficient in-memory relogging to avoid
337 * repeated physical writes of frequently modified metadata. If we allow the CIL
338 * to grow to a substantial fraction of the log, then we may be pinning hundreds
339 * of megabytes of metadata in memory until the CIL flushes. This can cause
340 * issues when we are running low on memory - pinned memory cannot be reclaimed,
341 * and the CIL consumes a lot of memory. Hence we need to set an upper physical
342 * size limit for the CIL that limits the maximum amount of memory pinned by the
343 * CIL but does not limit performance by reducing relogging efficiency
344 * significantly.
345 *
346 * As such, the CIL push threshold ends up being the smaller of two thresholds:
347 * - a threshold large enough that it allows CIL to be pushed and progress to be
348 * made without excessive blocking of incoming transaction commits. This is
349 * defined to be 12.5% of the log space - half the 25% push threshold of the
350 * AIL.
351 * - small enough that it doesn't pin excessive amounts of memory but maintains
352 * close to peak relogging efficiency. This is defined to be 16x the iclog
353 * buffer window (32MB) as measurements have shown this to be roughly the
354 * point of diminishing performance increases under highly concurrent
355 * modification workloads.
356 *
357 * To prevent the CIL from overflowing upper commit size bounds, we introduce a
358 * new threshold at which we block committing transactions until the background
359 * CIL commit commences and switches to a new context. While this is not a hard
360 * limit, it forces the process committing a transaction to the CIL to block and
361 * yeild the CPU, giving the CIL push work a chance to be scheduled and start
362 * work. This prevents a process running lots of transactions from overfilling
363 * the CIL because it is not yielding the CPU. We set the blocking limit at
364 * twice the background push space threshold so we keep in line with the AIL
365 * push thresholds.
366 *
367 * Note: this is not a -hard- limit as blocking is applied after the transaction
368 * is inserted into the CIL and the push has been triggered. It is largely a
369 * throttling mechanism that allows the CIL push to be scheduled and run. A hard
370 * limit will be difficult to implement without introducing global serialisation
371 * in the CIL commit fast path, and it's not at all clear that we actually need
372 * such hard limits given the ~7 years we've run without a hard limit before
373 * finding the first situation where a checkpoint size overflow actually
374 * occurred. Hence the simple throttle, and an ASSERT check to tell us that
375 * we've overrun the max size.
376 */
377#define XLOG_CIL_SPACE_LIMIT(log) \
378 min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4)
379
380#define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \
381 (XLOG_CIL_SPACE_LIMIT(log) * 2)
382
383/*
384 * ticket grant locks, queues and accounting have their own cachlines
385 * as these are quite hot and can be operated on concurrently.
386 */
387struct xlog_grant_head {
388 spinlock_t lock ____cacheline_aligned_in_smp;
389 struct list_head waiters;
390 atomic64_t grant;
391};
392
393/*
394 * The reservation head lsn is not made up of a cycle number and block number.
395 * Instead, it uses a cycle number and byte number. Logs don't expect to
396 * overflow 31 bits worth of byte offset, so using a byte number will mean
397 * that round off problems won't occur when releasing partial reservations.
398 */
399struct xlog {
400 /* The following fields don't need locking */
401 struct xfs_mount *l_mp; /* mount point */
402 struct xfs_ail *l_ailp; /* AIL log is working with */
403 struct xfs_cil *l_cilp; /* CIL log is working with */
404 struct xfs_buftarg *l_targ; /* buftarg of log */
405 struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */
406 struct delayed_work l_work; /* background flush work */
407 long l_opstate; /* operational state */
408 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
409 struct list_head *l_buf_cancel_table;
410 struct list_head r_dfops; /* recovered log intent items */
411 int l_iclog_hsize; /* size of iclog header */
412 int l_iclog_heads; /* # of iclog header sectors */
413 uint l_sectBBsize; /* sector size in BBs (2^n) */
414 int l_iclog_size; /* size of log in bytes */
415 int l_iclog_bufs; /* number of iclog buffers */
416 xfs_daddr_t l_logBBstart; /* start block of log */
417 int l_logsize; /* size of log in bytes */
418 int l_logBBsize; /* size of log in BB chunks */
419
420 /* The following block of fields are changed while holding icloglock */
421 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
422 /* waiting for iclog flush */
423 int l_covered_state;/* state of "covering disk
424 * log entries" */
425 xlog_in_core_t *l_iclog; /* head log queue */
426 spinlock_t l_icloglock; /* grab to change iclog state */
427 int l_curr_cycle; /* Cycle number of log writes */
428 int l_prev_cycle; /* Cycle number before last
429 * block increment */
430 int l_curr_block; /* current logical log block */
431 int l_prev_block; /* previous logical log block */
432
433 /*
434 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and
435 * read without needing to hold specific locks. To avoid operations
436 * contending with other hot objects, place each of them on a separate
437 * cacheline.
438 */
439 /* lsn of last LR on disk */
440 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp;
441 /* lsn of 1st LR with unflushed * buffers */
442 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
443
444 struct xlog_grant_head l_reserve_head;
445 struct xlog_grant_head l_write_head;
446
447 struct xfs_kobj l_kobj;
448
449 /* log recovery lsn tracking (for buffer submission */
450 xfs_lsn_t l_recovery_lsn;
451
452 uint32_t l_iclog_roundoff;/* padding roundoff */
453
454 /* Users of log incompat features should take a read lock. */
455 struct rw_semaphore l_incompat_users;
456};
457
458/*
459 * Bits for operational state
460 */
461#define XLOG_ACTIVE_RECOVERY 0 /* in the middle of recovery */
462#define XLOG_RECOVERY_NEEDED 1 /* log was recovered */
463#define XLOG_IO_ERROR 2 /* log hit an I/O error, and being
464 shutdown */
465#define XLOG_TAIL_WARN 3 /* log tail verify warning issued */
466
467static inline bool
468xlog_recovery_needed(struct xlog *log)
469{
470 return test_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
471}
472
473static inline bool
474xlog_in_recovery(struct xlog *log)
475{
476 return test_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
477}
478
479static inline bool
480xlog_is_shutdown(struct xlog *log)
481{
482 return test_bit(XLOG_IO_ERROR, &log->l_opstate);
483}
484
485/*
486 * Wait until the xlog_force_shutdown() has marked the log as shut down
487 * so xlog_is_shutdown() will always return true.
488 */
489static inline void
490xlog_shutdown_wait(
491 struct xlog *log)
492{
493 wait_var_event(&log->l_opstate, xlog_is_shutdown(log));
494}
495
496/* common routines */
497extern int
498xlog_recover(
499 struct xlog *log);
500extern int
501xlog_recover_finish(
502 struct xlog *log);
503extern void
504xlog_recover_cancel(struct xlog *);
505
506extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
507 char *dp, int size);
508
509extern struct kmem_cache *xfs_log_ticket_cache;
510struct xlog_ticket *xlog_ticket_alloc(struct xlog *log, int unit_bytes,
511 int count, bool permanent);
512
513void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
514void xlog_print_trans(struct xfs_trans *);
515int xlog_write(struct xlog *log, struct xfs_cil_ctx *ctx,
516 struct list_head *lv_chain, struct xlog_ticket *tic,
517 uint32_t len);
518void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket);
519void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket);
520
521void xlog_state_switch_iclogs(struct xlog *log, struct xlog_in_core *iclog,
522 int eventual_size);
523int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog,
524 struct xlog_ticket *ticket);
525
526/*
527 * When we crack an atomic LSN, we sample it first so that the value will not
528 * change while we are cracking it into the component values. This means we
529 * will always get consistent component values to work from. This should always
530 * be used to sample and crack LSNs that are stored and updated in atomic
531 * variables.
532 */
533static inline void
534xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
535{
536 xfs_lsn_t val = atomic64_read(lsn);
537
538 *cycle = CYCLE_LSN(val);
539 *block = BLOCK_LSN(val);
540}
541
542/*
543 * Calculate and assign a value to an atomic LSN variable from component pieces.
544 */
545static inline void
546xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
547{
548 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
549}
550
551/*
552 * When we crack the grant head, we sample it first so that the value will not
553 * change while we are cracking it into the component values. This means we
554 * will always get consistent component values to work from.
555 */
556static inline void
557xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
558{
559 *cycle = val >> 32;
560 *space = val & 0xffffffff;
561}
562
563static inline void
564xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
565{
566 xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
567}
568
569static inline int64_t
570xlog_assign_grant_head_val(int cycle, int space)
571{
572 return ((int64_t)cycle << 32) | space;
573}
574
575static inline void
576xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
577{
578 atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
579}
580
581/*
582 * Committed Item List interfaces
583 */
584int xlog_cil_init(struct xlog *log);
585void xlog_cil_init_post_recovery(struct xlog *log);
586void xlog_cil_destroy(struct xlog *log);
587bool xlog_cil_empty(struct xlog *log);
588void xlog_cil_commit(struct xlog *log, struct xfs_trans *tp,
589 xfs_csn_t *commit_seq, bool regrant);
590void xlog_cil_set_ctx_write_state(struct xfs_cil_ctx *ctx,
591 struct xlog_in_core *iclog);
592
593
594/*
595 * CIL force routines
596 */
597void xlog_cil_flush(struct xlog *log);
598xfs_lsn_t xlog_cil_force_seq(struct xlog *log, xfs_csn_t sequence);
599
600static inline void
601xlog_cil_force(struct xlog *log)
602{
603 xlog_cil_force_seq(log, log->l_cilp->xc_current_sequence);
604}
605
606/*
607 * Wrapper function for waiting on a wait queue serialised against wakeups
608 * by a spinlock. This matches the semantics of all the wait queues used in the
609 * log code.
610 */
611static inline void
612xlog_wait(
613 struct wait_queue_head *wq,
614 struct spinlock *lock)
615 __releases(lock)
616{
617 DECLARE_WAITQUEUE(wait, current);
618
619 add_wait_queue_exclusive(wq, &wait);
620 __set_current_state(TASK_UNINTERRUPTIBLE);
621 spin_unlock(lock);
622 schedule();
623 remove_wait_queue(wq, &wait);
624}
625
626int xlog_wait_on_iclog(struct xlog_in_core *iclog);
627
628/*
629 * The LSN is valid so long as it is behind the current LSN. If it isn't, this
630 * means that the next log record that includes this metadata could have a
631 * smaller LSN. In turn, this means that the modification in the log would not
632 * replay.
633 */
634static inline bool
635xlog_valid_lsn(
636 struct xlog *log,
637 xfs_lsn_t lsn)
638{
639 int cur_cycle;
640 int cur_block;
641 bool valid = true;
642
643 /*
644 * First, sample the current lsn without locking to avoid added
645 * contention from metadata I/O. The current cycle and block are updated
646 * (in xlog_state_switch_iclogs()) and read here in a particular order
647 * to avoid false negatives (e.g., thinking the metadata LSN is valid
648 * when it is not).
649 *
650 * The current block is always rewound before the cycle is bumped in
651 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in
652 * a transiently forward state. Instead, we can see the LSN in a
653 * transiently behind state if we happen to race with a cycle wrap.
654 */
655 cur_cycle = READ_ONCE(log->l_curr_cycle);
656 smp_rmb();
657 cur_block = READ_ONCE(log->l_curr_block);
658
659 if ((CYCLE_LSN(lsn) > cur_cycle) ||
660 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
661 /*
662 * If the metadata LSN appears invalid, it's possible the check
663 * above raced with a wrap to the next log cycle. Grab the lock
664 * to check for sure.
665 */
666 spin_lock(&log->l_icloglock);
667 cur_cycle = log->l_curr_cycle;
668 cur_block = log->l_curr_block;
669 spin_unlock(&log->l_icloglock);
670
671 if ((CYCLE_LSN(lsn) > cur_cycle) ||
672 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
673 valid = false;
674 }
675
676 return valid;
677}
678
679/*
680 * Log vector and shadow buffers can be large, so we need to use kvmalloc() here
681 * to ensure success. Unfortunately, kvmalloc() only allows GFP_KERNEL contexts
682 * to fall back to vmalloc, so we can't actually do anything useful with gfp
683 * flags to control the kmalloc() behaviour within kvmalloc(). Hence kmalloc()
684 * will do direct reclaim and compaction in the slow path, both of which are
685 * horrendously expensive. We just want kmalloc to fail fast and fall back to
686 * vmalloc if it can't get somethign straight away from the free lists or
687 * buddy allocator. Hence we have to open code kvmalloc outselves here.
688 *
689 * This assumes that the caller uses memalloc_nofs_save task context here, so
690 * despite the use of GFP_KERNEL here, we are going to be doing GFP_NOFS
691 * allocations. This is actually the only way to make vmalloc() do GFP_NOFS
692 * allocations, so lets just all pretend this is a GFP_KERNEL context
693 * operation....
694 */
695static inline void *
696xlog_kvmalloc(
697 size_t buf_size)
698{
699 gfp_t flags = GFP_KERNEL;
700 void *p;
701
702 flags &= ~__GFP_DIRECT_RECLAIM;
703 flags |= __GFP_NOWARN | __GFP_NORETRY;
704 do {
705 p = kmalloc(buf_size, flags);
706 if (!p)
707 p = vmalloc(buf_size);
708 } while (!p);
709
710 return p;
711}
712
713#endif /* __XFS_LOG_PRIV_H__ */
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
11struct xfs_buf;
12struct xlog;
13struct xlog_ticket;
14struct 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 */
28static 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 */
36enum 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
147typedef 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} xlog_ticket_t;
159
160/*
161 * - A log record header is 512 bytes. There is plenty of room to grow the
162 * xlog_rec_header_t into the reserved space.
163 * - ic_data follows, so a write to disk can start at the beginning of
164 * the iclog.
165 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
166 * - ic_next is the pointer to the next iclog in the ring.
167 * - ic_log is a pointer back to the global log structure.
168 * - ic_size is the full size of the log buffer, minus the cycle headers.
169 * - ic_offset is the current number of bytes written to in this iclog.
170 * - ic_refcnt is bumped when someone is writing to the log.
171 * - ic_state is the state of the iclog.
172 *
173 * Because of cacheline contention on large machines, we need to separate
174 * various resources onto different cachelines. To start with, make the
175 * structure cacheline aligned. The following fields can be contended on
176 * by independent processes:
177 *
178 * - ic_callbacks
179 * - ic_refcnt
180 * - fields protected by the global l_icloglock
181 *
182 * so we need to ensure that these fields are located in separate cachelines.
183 * We'll put all the read-only and l_icloglock fields in the first cacheline,
184 * and move everything else out to subsequent cachelines.
185 */
186typedef struct xlog_in_core {
187 wait_queue_head_t ic_force_wait;
188 wait_queue_head_t ic_write_wait;
189 struct xlog_in_core *ic_next;
190 struct xlog_in_core *ic_prev;
191 struct xlog *ic_log;
192 u32 ic_size;
193 u32 ic_offset;
194 enum xlog_iclog_state ic_state;
195 unsigned int ic_flags;
196 void *ic_datap; /* pointer to iclog data */
197 struct list_head ic_callbacks;
198
199 /* reference counts need their own cacheline */
200 atomic_t ic_refcnt ____cacheline_aligned_in_smp;
201 xlog_in_core_2_t *ic_data;
202#define ic_header ic_data->hic_header
203#ifdef DEBUG
204 bool ic_fail_crc : 1;
205#endif
206 struct semaphore ic_sema;
207 struct work_struct ic_end_io_work;
208 struct bio ic_bio;
209 struct bio_vec ic_bvec[];
210} xlog_in_core_t;
211
212/*
213 * The CIL context is used to aggregate per-transaction details as well be
214 * passed to the iclog for checkpoint post-commit processing. After being
215 * passed to the iclog, another context needs to be allocated for tracking the
216 * next set of transactions to be aggregated into a checkpoint.
217 */
218struct xfs_cil;
219
220struct xfs_cil_ctx {
221 struct xfs_cil *cil;
222 xfs_csn_t sequence; /* chkpt sequence # */
223 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */
224 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */
225 struct xlog_in_core *commit_iclog;
226 struct xlog_ticket *ticket; /* chkpt ticket */
227 atomic_t space_used; /* aggregate size of regions */
228 struct xfs_busy_extents busy_extents;
229 struct list_head log_items; /* log items in chkpt */
230 struct list_head lv_chain; /* logvecs being pushed */
231 struct list_head iclog_entry;
232 struct list_head committing; /* ctx committing list */
233 struct work_struct push_work;
234 atomic_t order_id;
235
236 /*
237 * CPUs that could have added items to the percpu CIL data. Access is
238 * coordinated with xc_ctx_lock.
239 */
240 struct cpumask cil_pcpmask;
241};
242
243/*
244 * Per-cpu CIL tracking items
245 */
246struct xlog_cil_pcp {
247 int32_t space_used;
248 uint32_t space_reserved;
249 struct list_head busy_extents;
250 struct list_head log_items;
251};
252
253/*
254 * Committed Item List structure
255 *
256 * This structure is used to track log items that have been committed but not
257 * yet written into the log. It is used only when the delayed logging mount
258 * option is enabled.
259 *
260 * This structure tracks the list of committing checkpoint contexts so
261 * we can avoid the problem of having to hold out new transactions during a
262 * flush until we have a the commit record LSN of the checkpoint. We can
263 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
264 * sequence match and extract the commit LSN directly from there. If the
265 * checkpoint is still in the process of committing, we can block waiting for
266 * the commit LSN to be determined as well. This should make synchronous
267 * operations almost as efficient as the old logging methods.
268 */
269struct xfs_cil {
270 struct xlog *xc_log;
271 unsigned long xc_flags;
272 atomic_t xc_iclog_hdrs;
273 struct workqueue_struct *xc_push_wq;
274
275 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp;
276 struct xfs_cil_ctx *xc_ctx;
277
278 spinlock_t xc_push_lock ____cacheline_aligned_in_smp;
279 xfs_csn_t xc_push_seq;
280 bool xc_push_commit_stable;
281 struct list_head xc_committing;
282 wait_queue_head_t xc_commit_wait;
283 wait_queue_head_t xc_start_wait;
284 xfs_csn_t xc_current_sequence;
285 wait_queue_head_t xc_push_wait; /* background push throttle */
286
287 void __percpu *xc_pcp; /* percpu CIL structures */
288} ____cacheline_aligned_in_smp;
289
290/* xc_flags bit values */
291#define XLOG_CIL_EMPTY 1
292#define XLOG_CIL_PCP_SPACE 2
293
294/*
295 * The amount of log space we allow the CIL to aggregate is difficult to size.
296 * Whatever we choose, we have to make sure we can get a reservation for the
297 * log space effectively, that it is large enough to capture sufficient
298 * relogging to reduce log buffer IO significantly, but it is not too large for
299 * the log or induces too much latency when writing out through the iclogs. We
300 * track both space consumed and the number of vectors in the checkpoint
301 * context, so we need to decide which to use for limiting.
302 *
303 * Every log buffer we write out during a push needs a header reserved, which
304 * is at least one sector and more for v2 logs. Hence we need a reservation of
305 * at least 512 bytes per 32k of log space just for the LR headers. That means
306 * 16KB of reservation per megabyte of delayed logging space we will consume,
307 * plus various headers. The number of headers will vary based on the num of
308 * io vectors, so limiting on a specific number of vectors is going to result
309 * in transactions of varying size. IOWs, it is more consistent to track and
310 * limit space consumed in the log rather than by the number of objects being
311 * logged in order to prevent checkpoint ticket overruns.
312 *
313 * Further, use of static reservations through the log grant mechanism is
314 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
315 * grant) and a significant deadlock potential because regranting write space
316 * can block on log pushes. Hence if we have to regrant log space during a log
317 * push, we can deadlock.
318 *
319 * However, we can avoid this by use of a dynamic "reservation stealing"
320 * technique during transaction commit whereby unused reservation space in the
321 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
322 * space needed by the checkpoint transaction. This means that we never need to
323 * specifically reserve space for the CIL checkpoint transaction, nor do we
324 * need to regrant space once the checkpoint completes. This also means the
325 * checkpoint transaction ticket is specific to the checkpoint context, rather
326 * than the CIL itself.
327 *
328 * With dynamic reservations, we can effectively make up arbitrary limits for
329 * the checkpoint size so long as they don't violate any other size rules.
330 * Recovery imposes a rule that no transaction exceed half the log, so we are
331 * limited by that. Furthermore, the log transaction reservation subsystem
332 * tries to keep 25% of the log free, so we need to keep below that limit or we
333 * risk running out of free log space to start any new transactions.
334 *
335 * In order to keep background CIL push efficient, we only need to ensure the
336 * CIL is large enough to maintain sufficient in-memory relogging to avoid
337 * repeated physical writes of frequently modified metadata. If we allow the CIL
338 * to grow to a substantial fraction of the log, then we may be pinning hundreds
339 * of megabytes of metadata in memory until the CIL flushes. This can cause
340 * issues when we are running low on memory - pinned memory cannot be reclaimed,
341 * and the CIL consumes a lot of memory. Hence we need to set an upper physical
342 * size limit for the CIL that limits the maximum amount of memory pinned by the
343 * CIL but does not limit performance by reducing relogging efficiency
344 * significantly.
345 *
346 * As such, the CIL push threshold ends up being the smaller of two thresholds:
347 * - a threshold large enough that it allows CIL to be pushed and progress to be
348 * made without excessive blocking of incoming transaction commits. This is
349 * defined to be 12.5% of the log space - half the 25% push threshold of the
350 * AIL.
351 * - small enough that it doesn't pin excessive amounts of memory but maintains
352 * close to peak relogging efficiency. This is defined to be 16x the iclog
353 * buffer window (32MB) as measurements have shown this to be roughly the
354 * point of diminishing performance increases under highly concurrent
355 * modification workloads.
356 *
357 * To prevent the CIL from overflowing upper commit size bounds, we introduce a
358 * new threshold at which we block committing transactions until the background
359 * CIL commit commences and switches to a new context. While this is not a hard
360 * limit, it forces the process committing a transaction to the CIL to block and
361 * yeild the CPU, giving the CIL push work a chance to be scheduled and start
362 * work. This prevents a process running lots of transactions from overfilling
363 * the CIL because it is not yielding the CPU. We set the blocking limit at
364 * twice the background push space threshold so we keep in line with the AIL
365 * push thresholds.
366 *
367 * Note: this is not a -hard- limit as blocking is applied after the transaction
368 * is inserted into the CIL and the push has been triggered. It is largely a
369 * throttling mechanism that allows the CIL push to be scheduled and run. A hard
370 * limit will be difficult to implement without introducing global serialisation
371 * in the CIL commit fast path, and it's not at all clear that we actually need
372 * such hard limits given the ~7 years we've run without a hard limit before
373 * finding the first situation where a checkpoint size overflow actually
374 * occurred. Hence the simple throttle, and an ASSERT check to tell us that
375 * we've overrun the max size.
376 */
377#define XLOG_CIL_SPACE_LIMIT(log) \
378 min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4)
379
380#define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \
381 (XLOG_CIL_SPACE_LIMIT(log) * 2)
382
383/*
384 * ticket grant locks, queues and accounting have their own cachlines
385 * as these are quite hot and can be operated on concurrently.
386 */
387struct xlog_grant_head {
388 spinlock_t lock ____cacheline_aligned_in_smp;
389 struct list_head waiters;
390 atomic64_t grant;
391};
392
393/*
394 * The reservation head lsn is not made up of a cycle number and block number.
395 * Instead, it uses a cycle number and byte number. Logs don't expect to
396 * overflow 31 bits worth of byte offset, so using a byte number will mean
397 * that round off problems won't occur when releasing partial reservations.
398 */
399struct xlog {
400 /* The following fields don't need locking */
401 struct xfs_mount *l_mp; /* mount point */
402 struct xfs_ail *l_ailp; /* AIL log is working with */
403 struct xfs_cil *l_cilp; /* CIL log is working with */
404 struct xfs_buftarg *l_targ; /* buftarg of log */
405 struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */
406 struct delayed_work l_work; /* background flush work */
407 long l_opstate; /* operational state */
408 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
409 struct list_head *l_buf_cancel_table;
410 struct list_head r_dfops; /* recovered log intent items */
411 int l_iclog_hsize; /* size of iclog header */
412 int l_iclog_heads; /* # of iclog header sectors */
413 uint l_sectBBsize; /* sector size in BBs (2^n) */
414 int l_iclog_size; /* size of log in bytes */
415 int l_iclog_bufs; /* number of iclog buffers */
416 xfs_daddr_t l_logBBstart; /* start block of log */
417 int l_logsize; /* size of log in bytes */
418 int l_logBBsize; /* size of log in BB chunks */
419
420 /* The following block of fields are changed while holding icloglock */
421 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
422 /* waiting for iclog flush */
423 int l_covered_state;/* state of "covering disk
424 * log entries" */
425 xlog_in_core_t *l_iclog; /* head log queue */
426 spinlock_t l_icloglock; /* grab to change iclog state */
427 int l_curr_cycle; /* Cycle number of log writes */
428 int l_prev_cycle; /* Cycle number before last
429 * block increment */
430 int l_curr_block; /* current logical log block */
431 int l_prev_block; /* previous logical log block */
432
433 /*
434 * l_tail_lsn is atomic so it can be set and read without needing to
435 * hold specific locks. To avoid operations contending with other hot
436 * objects, it on a separate cacheline.
437 */
438 /* lsn of 1st LR with unflushed * buffers */
439 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
440
441 struct xlog_grant_head l_reserve_head;
442 struct xlog_grant_head l_write_head;
443 uint64_t l_tail_space;
444
445 struct xfs_kobj l_kobj;
446
447 /* log recovery lsn tracking (for buffer submission */
448 xfs_lsn_t l_recovery_lsn;
449
450 uint32_t l_iclog_roundoff;/* padding roundoff */
451};
452
453/*
454 * Bits for operational state
455 */
456#define XLOG_ACTIVE_RECOVERY 0 /* in the middle of recovery */
457#define XLOG_RECOVERY_NEEDED 1 /* log was recovered */
458#define XLOG_IO_ERROR 2 /* log hit an I/O error, and being
459 shutdown */
460#define XLOG_TAIL_WARN 3 /* log tail verify warning issued */
461#define XLOG_SHUTDOWN_STARTED 4 /* xlog_force_shutdown() exclusion */
462
463static inline bool
464xlog_recovery_needed(struct xlog *log)
465{
466 return test_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
467}
468
469static inline bool
470xlog_in_recovery(struct xlog *log)
471{
472 return test_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
473}
474
475static inline bool
476xlog_is_shutdown(struct xlog *log)
477{
478 return test_bit(XLOG_IO_ERROR, &log->l_opstate);
479}
480
481/*
482 * Wait until the xlog_force_shutdown() has marked the log as shut down
483 * so xlog_is_shutdown() will always return true.
484 */
485static inline void
486xlog_shutdown_wait(
487 struct xlog *log)
488{
489 wait_var_event(&log->l_opstate, xlog_is_shutdown(log));
490}
491
492/* common routines */
493extern int
494xlog_recover(
495 struct xlog *log);
496extern int
497xlog_recover_finish(
498 struct xlog *log);
499extern void
500xlog_recover_cancel(struct xlog *);
501
502extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
503 char *dp, int size);
504
505extern struct kmem_cache *xfs_log_ticket_cache;
506struct xlog_ticket *xlog_ticket_alloc(struct xlog *log, int unit_bytes,
507 int count, bool permanent);
508
509void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
510void xlog_print_trans(struct xfs_trans *);
511int xlog_write(struct xlog *log, struct xfs_cil_ctx *ctx,
512 struct list_head *lv_chain, struct xlog_ticket *tic,
513 uint32_t len);
514void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket);
515void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket);
516
517void xlog_state_switch_iclogs(struct xlog *log, struct xlog_in_core *iclog,
518 int eventual_size);
519int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog,
520 struct xlog_ticket *ticket);
521
522/*
523 * When we crack an atomic LSN, we sample it first so that the value will not
524 * change while we are cracking it into the component values. This means we
525 * will always get consistent component values to work from. This should always
526 * be used to sample and crack LSNs that are stored and updated in atomic
527 * variables.
528 */
529static inline void
530xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
531{
532 xfs_lsn_t val = atomic64_read(lsn);
533
534 *cycle = CYCLE_LSN(val);
535 *block = BLOCK_LSN(val);
536}
537
538/*
539 * Calculate and assign a value to an atomic LSN variable from component pieces.
540 */
541static inline void
542xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
543{
544 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
545}
546
547/*
548 * Committed Item List interfaces
549 */
550int xlog_cil_init(struct xlog *log);
551void xlog_cil_init_post_recovery(struct xlog *log);
552void xlog_cil_destroy(struct xlog *log);
553bool xlog_cil_empty(struct xlog *log);
554void xlog_cil_commit(struct xlog *log, struct xfs_trans *tp,
555 xfs_csn_t *commit_seq, bool regrant);
556void xlog_cil_set_ctx_write_state(struct xfs_cil_ctx *ctx,
557 struct xlog_in_core *iclog);
558
559
560/*
561 * CIL force routines
562 */
563void xlog_cil_flush(struct xlog *log);
564xfs_lsn_t xlog_cil_force_seq(struct xlog *log, xfs_csn_t sequence);
565
566static inline void
567xlog_cil_force(struct xlog *log)
568{
569 xlog_cil_force_seq(log, log->l_cilp->xc_current_sequence);
570}
571
572/*
573 * Wrapper function for waiting on a wait queue serialised against wakeups
574 * by a spinlock. This matches the semantics of all the wait queues used in the
575 * log code.
576 */
577static inline void
578xlog_wait(
579 struct wait_queue_head *wq,
580 struct spinlock *lock)
581 __releases(lock)
582{
583 DECLARE_WAITQUEUE(wait, current);
584
585 add_wait_queue_exclusive(wq, &wait);
586 __set_current_state(TASK_UNINTERRUPTIBLE);
587 spin_unlock(lock);
588 schedule();
589 remove_wait_queue(wq, &wait);
590}
591
592int xlog_wait_on_iclog(struct xlog_in_core *iclog)
593 __releases(iclog->ic_log->l_icloglock);
594
595/* Calculate the distance between two LSNs in bytes */
596static inline uint64_t
597xlog_lsn_sub(
598 struct xlog *log,
599 xfs_lsn_t high,
600 xfs_lsn_t low)
601{
602 uint32_t hi_cycle = CYCLE_LSN(high);
603 uint32_t hi_block = BLOCK_LSN(high);
604 uint32_t lo_cycle = CYCLE_LSN(low);
605 uint32_t lo_block = BLOCK_LSN(low);
606
607 if (hi_cycle == lo_cycle)
608 return BBTOB(hi_block - lo_block);
609 ASSERT((hi_cycle == lo_cycle + 1) || xlog_is_shutdown(log));
610 return (uint64_t)log->l_logsize - BBTOB(lo_block - hi_block);
611}
612
613void xlog_grant_return_space(struct xlog *log, xfs_lsn_t old_head,
614 xfs_lsn_t new_head);
615
616/*
617 * The LSN is valid so long as it is behind the current LSN. If it isn't, this
618 * means that the next log record that includes this metadata could have a
619 * smaller LSN. In turn, this means that the modification in the log would not
620 * replay.
621 */
622static inline bool
623xlog_valid_lsn(
624 struct xlog *log,
625 xfs_lsn_t lsn)
626{
627 int cur_cycle;
628 int cur_block;
629 bool valid = true;
630
631 /*
632 * First, sample the current lsn without locking to avoid added
633 * contention from metadata I/O. The current cycle and block are updated
634 * (in xlog_state_switch_iclogs()) and read here in a particular order
635 * to avoid false negatives (e.g., thinking the metadata LSN is valid
636 * when it is not).
637 *
638 * The current block is always rewound before the cycle is bumped in
639 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in
640 * a transiently forward state. Instead, we can see the LSN in a
641 * transiently behind state if we happen to race with a cycle wrap.
642 */
643 cur_cycle = READ_ONCE(log->l_curr_cycle);
644 smp_rmb();
645 cur_block = READ_ONCE(log->l_curr_block);
646
647 if ((CYCLE_LSN(lsn) > cur_cycle) ||
648 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
649 /*
650 * If the metadata LSN appears invalid, it's possible the check
651 * above raced with a wrap to the next log cycle. Grab the lock
652 * to check for sure.
653 */
654 spin_lock(&log->l_icloglock);
655 cur_cycle = log->l_curr_cycle;
656 cur_block = log->l_curr_block;
657 spin_unlock(&log->l_icloglock);
658
659 if ((CYCLE_LSN(lsn) > cur_cycle) ||
660 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
661 valid = false;
662 }
663
664 return valid;
665}
666
667/*
668 * Log vector and shadow buffers can be large, so we need to use kvmalloc() here
669 * to ensure success. Unfortunately, kvmalloc() only allows GFP_KERNEL contexts
670 * to fall back to vmalloc, so we can't actually do anything useful with gfp
671 * flags to control the kmalloc() behaviour within kvmalloc(). Hence kmalloc()
672 * will do direct reclaim and compaction in the slow path, both of which are
673 * horrendously expensive. We just want kmalloc to fail fast and fall back to
674 * vmalloc if it can't get something straight away from the free lists or
675 * buddy allocator. Hence we have to open code kvmalloc outselves here.
676 *
677 * This assumes that the caller uses memalloc_nofs_save task context here, so
678 * despite the use of GFP_KERNEL here, we are going to be doing GFP_NOFS
679 * allocations. This is actually the only way to make vmalloc() do GFP_NOFS
680 * allocations, so lets just all pretend this is a GFP_KERNEL context
681 * operation....
682 */
683static inline void *
684xlog_kvmalloc(
685 size_t buf_size)
686{
687 gfp_t flags = GFP_KERNEL;
688 void *p;
689
690 flags &= ~__GFP_DIRECT_RECLAIM;
691 flags |= __GFP_NOWARN | __GFP_NORETRY;
692 do {
693 p = kmalloc(buf_size, flags);
694 if (!p)
695 p = vmalloc(buf_size);
696 } while (!p);
697
698 return p;
699}
700
701#endif /* __XFS_LOG_PRIV_H__ */