Linux Audio

Check our new training course

Loading...
v6.13.7
  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__ */
v5.14.15
  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
 
 
  9struct xfs_buf;
 10struct xlog;
 11struct xlog_ticket;
 12struct xfs_mount;
 13
 14/*
 15 * Flags for log structure
 16 */
 17#define XLOG_ACTIVE_RECOVERY	0x2	/* in the middle of recovery */
 18#define	XLOG_RECOVERY_NEEDED	0x4	/* log was recovered */
 19#define XLOG_IO_ERROR		0x8	/* log hit an I/O error, and being
 20					   shutdown */
 21#define XLOG_TAIL_WARN		0x10	/* log tail verify warning issued */
 22
 23/*
 24 * get client id from packed copy.
 25 *
 26 * this hack is here because the xlog_pack code copies four bytes
 27 * of xlog_op_header containing the fields oh_clientid, oh_flags
 28 * and oh_res2 into the packed copy.
 29 *
 30 * later on this four byte chunk is treated as an int and the
 31 * client id is pulled out.
 32 *
 33 * this has endian issues, of course.
 34 */
 35static inline uint xlog_get_client_id(__be32 i)
 36{
 37	return be32_to_cpu(i) >> 24;
 38}
 39
 40/*
 41 * In core log state
 42 */
 43enum xlog_iclog_state {
 44	XLOG_STATE_ACTIVE,	/* Current IC log being written to */
 45	XLOG_STATE_WANT_SYNC,	/* Want to sync this iclog; no more writes */
 46	XLOG_STATE_SYNCING,	/* This IC log is syncing */
 47	XLOG_STATE_DONE_SYNC,	/* Done syncing to disk */
 48	XLOG_STATE_CALLBACK,	/* Callback functions now */
 49	XLOG_STATE_DIRTY,	/* Dirty IC log, not ready for ACTIVE status */
 50	XLOG_STATE_IOERROR,	/* IO error happened in sync'ing log */
 51};
 52
 53#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" }
 61
 62/*
 63 * In core log flags
 64 */
 65#define XLOG_ICL_NEED_FLUSH	(1 << 0)	/* iclog needs REQ_PREFLUSH */
 66#define XLOG_ICL_NEED_FUA	(1 << 1)	/* iclog needs REQ_FUA */
 67
 68#define XLOG_ICL_STRINGS \
 69	{ XLOG_ICL_NEED_FLUSH,	"XLOG_ICL_NEED_FLUSH" }, \
 70	{ XLOG_ICL_NEED_FUA,	"XLOG_ICL_NEED_FUA" }
 71
 72
 73/*
 74 * Log ticket flags
 75 */
 76#define XLOG_TIC_PERM_RESERV	0x1	/* permanent reservation */
 77
 78#define XLOG_TIC_FLAGS \
 79	{ XLOG_TIC_PERM_RESERV,	"XLOG_TIC_PERM_RESERV" }
 80
 81/*
 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
 86 * log write.
 87 *
 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.
 98 *
 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.
105 *
106 * These dummy transactions get committed when everything
107 * is idle (after there has been some activity).
108 *
109 * There are 5 states used to control this.
110 *
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
116 *		transaction.
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.
120 *
121 * There are two places where we switch states:
122 *
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.
126 *
127 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
128 *
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.
134 *
135 *	If we are in the DONE state and this was a write of the
136 *		dummy transaction, we move to NEED2.
137 *
138 *	If we are in the DONE2 state and this was a write of the
139 *		dummy transaction, we move to IDLE.
140 *
141 *
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.
146 */
147
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
153
154#define XLOG_COVER_OPS		5
155
156/* Ticket reservation region accounting */ 
157#define XLOG_TIC_LEN_MAX	15
158
159/*
160 * Reservation region
161 * As would be stored in xfs_log_iovec but without the i_addr which
162 * we don't care about.
163 */
164typedef struct xlog_res {
165	uint	r_len;	/* region length		:4 */
166	uint	r_type;	/* region's transaction type	:4 */
167} xlog_res_t;
168
169typedef 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  */
180
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 */ 
187} xlog_ticket_t;
188
189/*
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
193 *	the iclog.
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.
201 *
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:
206 *
207 *	- ic_callbacks
208 *	- ic_refcnt
209 *	- fields protected by the global l_icloglock
210 *
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.
214 */
215typedef 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;
220	struct xlog		*ic_log;
221	u32			ic_size;
222	u32			ic_offset;
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;
227
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
232#ifdef DEBUG
233	bool			ic_fail_crc : 1;
234#endif
235	struct semaphore	ic_sema;
236	struct work_struct	ic_end_io_work;
237	struct bio		ic_bio;
238	struct bio_vec		ic_bvec[];
239} xlog_in_core_t;
240
241/*
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.
246 */
247struct xfs_cil;
248
249struct xfs_cil_ctx {
250	struct xfs_cil		*cil;
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;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
262};
263
264/*
265 * Committed Item List structure
266 *
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
269 * option is enabled.
270 *
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.
279 */
280struct xfs_cil {
281	struct xlog		*xc_log;
282	struct list_head	xc_cil;
283	spinlock_t		xc_cil_lock;
 
284
285	struct rw_semaphore	xc_ctx_lock ____cacheline_aligned_in_smp;
286	struct xfs_cil_ctx	*xc_ctx;
287
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;
296
 
 
 
 
297/*
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.
305 *
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.
315 *
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.
321 *
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.
330 *
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.
337 *
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
347 * significantly.
348 *
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
353 *   AIL.
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.
359 *
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
368 * push thresholds.
369 *
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.
379 */
380#define XLOG_CIL_SPACE_LIMIT(log)	\
381	min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4)
382
383#define XLOG_CIL_BLOCKING_SPACE_LIMIT(log)	\
384	(XLOG_CIL_SPACE_LIMIT(log) * 2)
385
386/*
387 * ticket grant locks, queues and accounting have their own cachlines
388 * as these are quite hot and can be operated on concurrently.
389 */
390struct xlog_grant_head {
391	spinlock_t		lock ____cacheline_aligned_in_smp;
392	struct list_head	waiters;
393	atomic64_t		grant;
394};
395
396/*
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.
401 */
402struct xlog {
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 */
410	uint			l_flags;
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 */
421
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
426						 * log entries" */
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
431						 * block increment */
432	int			l_curr_block;   /* current logical log block */
433	int			l_prev_block;   /* previous logical log block */
434
435	/*
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
439	 * cacheline.
440	 */
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;
445
446	struct xlog_grant_head	l_reserve_head;
447	struct xlog_grant_head	l_write_head;
 
448
449	struct xfs_kobj		l_kobj;
450
451	/* The following field are used for debugging; need to hold icloglock */
452#ifdef DEBUG
453	void			*l_iclog_bak[XLOG_MAX_ICLOGS];
454#endif
455	/* log recovery lsn tracking (for buffer submission */
456	xfs_lsn_t		l_recovery_lsn;
457
458	uint32_t		l_iclog_roundoff;/* padding roundoff */
459};
460
461#define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
462	((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE))
 
 
 
 
 
 
 
463
464#define XLOG_FORCED_SHUTDOWN(log) \
465	(unlikely((log)->l_flags & XLOG_IO_ERROR))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
466
467/* common routines */
468extern int
469xlog_recover(
470	struct xlog		*log);
471extern int
472xlog_recover_finish(
473	struct xlog		*log);
474extern void
475xlog_recover_cancel(struct xlog *);
476
477extern __le32	 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
478			    char *dp, int size);
479
480extern kmem_zone_t *xfs_log_ticket_zone;
481struct xlog_ticket *
482xlog_ticket_alloc(
483	struct xlog	*log,
484	int		unit_bytes,
485	int		count,
486	char		client,
487	bool		permanent);
488
489static inline void
490xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes)
491{
492	*ptr += bytes;
493	*len -= bytes;
494	*off += bytes;
495}
496
497void	xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
498void	xlog_print_trans(struct xfs_trans *);
499int	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);
502int	xlog_commit_record(struct xlog *log, struct xlog_ticket *ticket,
503		struct xlog_in_core **iclog, xfs_lsn_t *lsn);
504void	xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket);
505void	xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket);
506
 
 
507int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog,
508		xfs_lsn_t log_tail_lsn);
509
510/*
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
515 * variables.
516 */
517static inline void
518xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
519{
520	xfs_lsn_t val = atomic64_read(lsn);
521
522	*cycle = CYCLE_LSN(val);
523	*block = BLOCK_LSN(val);
524}
525
526/*
527 * Calculate and assign a value to an atomic LSN variable from component pieces.
528 */
529static inline void
530xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
531{
532	atomic64_set(lsn, xlog_assign_lsn(cycle, block));
533}
534
535/*
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.
539 */
540static inline void
541xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
542{
543	*cycle = val >> 32;
544	*space = val & 0xffffffff;
545}
546
547static inline void
548xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
549{
550	xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
551}
552
553static inline int64_t
554xlog_assign_grant_head_val(int cycle, int space)
555{
556	return ((int64_t)cycle << 32) | space;
557}
558
559static inline void
560xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
561{
562	atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
563}
564
565/*
566 * Committed Item List interfaces
567 */
568int	xlog_cil_init(struct xlog *log);
569void	xlog_cil_init_post_recovery(struct xlog *log);
570void	xlog_cil_destroy(struct xlog *log);
571bool	xlog_cil_empty(struct xlog *log);
572void	xlog_cil_commit(struct xlog *log, struct xfs_trans *tp,
573			xfs_csn_t *commit_seq, bool regrant);
 
 
 
574
575/*
576 * CIL force routines
577 */
 
578xfs_lsn_t xlog_cil_force_seq(struct xlog *log, xfs_csn_t sequence);
579
580static inline void
581xlog_cil_force(struct xlog *log)
582{
583	xlog_cil_force_seq(log, log->l_cilp->xc_current_sequence);
584}
585
586/*
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
589 * log code.
590 */
591static inline void
592xlog_wait(
593	struct wait_queue_head	*wq,
594	struct spinlock		*lock)
595		__releases(lock)
596{
597	DECLARE_WAITQUEUE(wait, current);
598
599	add_wait_queue_exclusive(wq, &wait);
600	__set_current_state(TASK_UNINTERRUPTIBLE);
601	spin_unlock(lock);
602	schedule();
603	remove_wait_queue(wq, &wait);
604}
605
606int xlog_wait_on_iclog(struct xlog_in_core *iclog);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
607
608/*
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
612 * replay.
613 */
614static inline bool
615xlog_valid_lsn(
616	struct xlog	*log,
617	xfs_lsn_t	lsn)
618{
619	int		cur_cycle;
620	int		cur_block;
621	bool		valid = true;
622
623	/*
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
628	 * when it is not).
629	 *
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.
634	 */
635	cur_cycle = READ_ONCE(log->l_curr_cycle);
636	smp_rmb();
637	cur_block = READ_ONCE(log->l_curr_block);
638
639	if ((CYCLE_LSN(lsn) > cur_cycle) ||
640	    (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
641		/*
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
644		 * to check for sure.
645		 */
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);
650
651		if ((CYCLE_LSN(lsn) > cur_cycle) ||
652		    (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
653			valid = false;
654	}
655
656	return valid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
657}
658
659#endif	/* __XFS_LOG_PRIV_H__ */