1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
6 #ifndef __XFS_LOG_PRIV_H__
7 #define __XFS_LOG_PRIV_H__
15 * Flags for log structure
17 #define XLOG_ACTIVE_RECOVERY 0x2 /* in the middle of recovery */
18 #define XLOG_RECOVERY_NEEDED 0x4 /* log was recovered */
19 #define XLOG_IO_ERROR 0x8 /* log hit an I/O error, and being
21 #define XLOG_TAIL_WARN 0x10 /* log tail verify warning issued */
24 * get client id from packed copy.
26 * this hack is here because the xlog_pack code copies four bytes
27 * of xlog_op_header containing the fields oh_clientid, oh_flags
28 * and oh_res2 into the packed copy.
30 * later on this four byte chunk is treated as an int and the
31 * client id is pulled out.
33 * this has endian issues, of course.
35 static inline uint
xlog_get_client_id(__be32 i
)
37 return be32_to_cpu(i
) >> 24;
43 enum xlog_iclog_state
{
44 XLOG_STATE_ACTIVE
, /* Current IC log being written to */
45 XLOG_STATE_WANT_SYNC
, /* Want to sync this iclog; no more writes */
46 XLOG_STATE_SYNCING
, /* This IC log is syncing */
47 XLOG_STATE_DONE_SYNC
, /* Done syncing to disk */
48 XLOG_STATE_CALLBACK
, /* Callback functions now */
49 XLOG_STATE_DIRTY
, /* Dirty IC log, not ready for ACTIVE status */
50 XLOG_STATE_IOERROR
, /* IO error happened in sync'ing log */
56 #define XLOG_TIC_PERM_RESERV 0x1 /* permanent reservation */
58 #define XLOG_TIC_FLAGS \
59 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }
62 * Below are states for covering allocation transactions.
63 * By covering, we mean changing the h_tail_lsn in the last on-disk
64 * log write such that no allocation transactions will be re-done during
65 * recovery after a system crash. Recovery starts at the last on-disk
68 * These states are used to insert dummy log entries to cover
69 * space allocation transactions which can undo non-transactional changes
70 * after a crash. Writes to a file with space
71 * already allocated do not result in any transactions. Allocations
72 * might include space beyond the EOF. So if we just push the EOF a
73 * little, the last transaction for the file could contain the wrong
74 * size. If there is no file system activity, after an allocation
75 * transaction, and the system crashes, the allocation transaction
76 * will get replayed and the file will be truncated. This could
77 * be hours/days/... after the allocation occurred.
79 * The fix for this is to do two dummy transactions when the
80 * system is idle. We need two dummy transaction because the h_tail_lsn
81 * in the log record header needs to point beyond the last possible
82 * non-dummy transaction. The first dummy changes the h_tail_lsn to
83 * the first transaction before the dummy. The second dummy causes
84 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn.
86 * These dummy transactions get committed when everything
87 * is idle (after there has been some activity).
89 * There are 5 states used to control this.
91 * IDLE -- no logging has been done on the file system or
92 * we are done covering previous transactions.
93 * NEED -- logging has occurred and we need a dummy transaction
94 * when the log becomes idle.
95 * DONE -- we were in the NEED state and have committed a dummy
97 * NEED2 -- we detected that a dummy transaction has gone to the
98 * on disk log with no other transactions.
99 * DONE2 -- we committed a dummy transaction when in the NEED2 state.
101 * There are two places where we switch states:
103 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
104 * We commit the dummy transaction and switch to DONE or DONE2,
105 * respectively. In all other states, we don't do anything.
107 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
109 * No matter what state we are in, if this isn't the dummy
110 * transaction going out, the next state is NEED.
111 * So, if we aren't in the DONE or DONE2 states, the next state
112 * is NEED. We can't be finishing a write of the dummy record
113 * unless it was committed and the state switched to DONE or DONE2.
115 * If we are in the DONE state and this was a write of the
116 * dummy transaction, we move to NEED2.
118 * If we are in the DONE2 state and this was a write of the
119 * dummy transaction, we move to IDLE.
122 * Writing only one dummy transaction can get appended to
123 * one file space allocation. When this happens, the log recovery
124 * code replays the space allocation and a file could be truncated.
125 * This is why we have the NEED2 and DONE2 states before going idle.
128 #define XLOG_STATE_COVER_IDLE 0
129 #define XLOG_STATE_COVER_NEED 1
130 #define XLOG_STATE_COVER_DONE 2
131 #define XLOG_STATE_COVER_NEED2 3
132 #define XLOG_STATE_COVER_DONE2 4
134 #define XLOG_COVER_OPS 5
136 /* Ticket reservation region accounting */
137 #define XLOG_TIC_LEN_MAX 15
141 * As would be stored in xfs_log_iovec but without the i_addr which
142 * we don't care about.
144 typedef struct xlog_res
{
145 uint r_len
; /* region length :4 */
146 uint r_type
; /* region's transaction type :4 */
149 typedef struct xlog_ticket
{
150 struct list_head t_queue
; /* reserve/write queue */
151 struct task_struct
*t_task
; /* task that owns this ticket */
152 xlog_tid_t t_tid
; /* transaction identifier : 4 */
153 atomic_t t_ref
; /* ticket reference count : 4 */
154 int t_curr_res
; /* current reservation in bytes : 4 */
155 int t_unit_res
; /* unit reservation in bytes : 4 */
156 char t_ocnt
; /* original count : 1 */
157 char t_cnt
; /* current count : 1 */
158 char t_clientid
; /* who does this belong to; : 1 */
159 char t_flags
; /* properties of reservation : 1 */
161 /* reservation array fields */
162 uint t_res_num
; /* num in array : 4 */
163 uint t_res_num_ophdrs
; /* num op hdrs : 4 */
164 uint t_res_arr_sum
; /* array sum : 4 */
165 uint t_res_o_flow
; /* sum overflow : 4 */
166 xlog_res_t t_res_arr
[XLOG_TIC_LEN_MAX
]; /* array of res : 8 * 15 */
170 * - A log record header is 512 bytes. There is plenty of room to grow the
171 * xlog_rec_header_t into the reserved space.
172 * - ic_data follows, so a write to disk can start at the beginning of
174 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
175 * - ic_next is the pointer to the next iclog in the ring.
176 * - ic_log is a pointer back to the global log structure.
177 * - ic_size is the full size of the log buffer, minus the cycle headers.
178 * - ic_offset is the current number of bytes written to in this iclog.
179 * - ic_refcnt is bumped when someone is writing to the log.
180 * - ic_state is the state of the iclog.
182 * Because of cacheline contention on large machines, we need to separate
183 * various resources onto different cachelines. To start with, make the
184 * structure cacheline aligned. The following fields can be contended on
185 * by independent processes:
189 * - fields protected by the global l_icloglock
191 * so we need to ensure that these fields are located in separate cachelines.
192 * We'll put all the read-only and l_icloglock fields in the first cacheline,
193 * and move everything else out to subsequent cachelines.
195 typedef struct xlog_in_core
{
196 wait_queue_head_t ic_force_wait
;
197 wait_queue_head_t ic_write_wait
;
198 struct xlog_in_core
*ic_next
;
199 struct xlog_in_core
*ic_prev
;
203 enum xlog_iclog_state ic_state
;
204 char *ic_datap
; /* pointer to iclog data */
206 /* Callback structures need their own cacheline */
207 spinlock_t ic_callback_lock ____cacheline_aligned_in_smp
;
208 struct list_head ic_callbacks
;
210 /* reference counts need their own cacheline */
211 atomic_t ic_refcnt ____cacheline_aligned_in_smp
;
212 xlog_in_core_2_t
*ic_data
;
213 #define ic_header ic_data->hic_header
215 bool ic_fail_crc
: 1;
217 struct semaphore ic_sema
;
218 struct work_struct ic_end_io_work
;
220 struct bio_vec ic_bvec
[];
224 * The CIL context is used to aggregate per-transaction details as well be
225 * passed to the iclog for checkpoint post-commit processing. After being
226 * passed to the iclog, another context needs to be allocated for tracking the
227 * next set of transactions to be aggregated into a checkpoint.
233 xfs_lsn_t sequence
; /* chkpt sequence # */
234 xfs_lsn_t start_lsn
; /* first LSN of chkpt commit */
235 xfs_lsn_t commit_lsn
; /* chkpt commit record lsn */
236 struct xlog_ticket
*ticket
; /* chkpt ticket */
237 int nvecs
; /* number of regions */
238 int space_used
; /* aggregate size of regions */
239 struct list_head busy_extents
; /* busy extents in chkpt */
240 struct xfs_log_vec
*lv_chain
; /* logvecs being pushed */
241 struct list_head iclog_entry
;
242 struct list_head committing
; /* ctx committing list */
243 struct work_struct discard_endio_work
;
247 * Committed Item List structure
249 * This structure is used to track log items that have been committed but not
250 * yet written into the log. It is used only when the delayed logging mount
253 * This structure tracks the list of committing checkpoint contexts so
254 * we can avoid the problem of having to hold out new transactions during a
255 * flush until we have a the commit record LSN of the checkpoint. We can
256 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
257 * sequence match and extract the commit LSN directly from there. If the
258 * checkpoint is still in the process of committing, we can block waiting for
259 * the commit LSN to be determined as well. This should make synchronous
260 * operations almost as efficient as the old logging methods.
264 struct list_head xc_cil
;
265 spinlock_t xc_cil_lock
;
267 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp
;
268 struct xfs_cil_ctx
*xc_ctx
;
270 spinlock_t xc_push_lock ____cacheline_aligned_in_smp
;
271 xfs_lsn_t xc_push_seq
;
272 struct list_head xc_committing
;
273 wait_queue_head_t xc_commit_wait
;
274 xfs_lsn_t xc_current_sequence
;
275 struct work_struct xc_push_work
;
276 wait_queue_head_t xc_push_wait
; /* background push throttle */
277 } ____cacheline_aligned_in_smp
;
280 * The amount of log space we allow the CIL to aggregate is difficult to size.
281 * Whatever we choose, we have to make sure we can get a reservation for the
282 * log space effectively, that it is large enough to capture sufficient
283 * relogging to reduce log buffer IO significantly, but it is not too large for
284 * the log or induces too much latency when writing out through the iclogs. We
285 * track both space consumed and the number of vectors in the checkpoint
286 * context, so we need to decide which to use for limiting.
288 * Every log buffer we write out during a push needs a header reserved, which
289 * is at least one sector and more for v2 logs. Hence we need a reservation of
290 * at least 512 bytes per 32k of log space just for the LR headers. That means
291 * 16KB of reservation per megabyte of delayed logging space we will consume,
292 * plus various headers. The number of headers will vary based on the num of
293 * io vectors, so limiting on a specific number of vectors is going to result
294 * in transactions of varying size. IOWs, it is more consistent to track and
295 * limit space consumed in the log rather than by the number of objects being
296 * logged in order to prevent checkpoint ticket overruns.
298 * Further, use of static reservations through the log grant mechanism is
299 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
300 * grant) and a significant deadlock potential because regranting write space
301 * can block on log pushes. Hence if we have to regrant log space during a log
302 * push, we can deadlock.
304 * However, we can avoid this by use of a dynamic "reservation stealing"
305 * technique during transaction commit whereby unused reservation space in the
306 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
307 * space needed by the checkpoint transaction. This means that we never need to
308 * specifically reserve space for the CIL checkpoint transaction, nor do we
309 * need to regrant space once the checkpoint completes. This also means the
310 * checkpoint transaction ticket is specific to the checkpoint context, rather
311 * than the CIL itself.
313 * With dynamic reservations, we can effectively make up arbitrary limits for
314 * the checkpoint size so long as they don't violate any other size rules.
315 * Recovery imposes a rule that no transaction exceed half the log, so we are
316 * limited by that. Furthermore, the log transaction reservation subsystem
317 * tries to keep 25% of the log free, so we need to keep below that limit or we
318 * risk running out of free log space to start any new transactions.
320 * In order to keep background CIL push efficient, we only need to ensure the
321 * CIL is large enough to maintain sufficient in-memory relogging to avoid
322 * repeated physical writes of frequently modified metadata. If we allow the CIL
323 * to grow to a substantial fraction of the log, then we may be pinning hundreds
324 * of megabytes of metadata in memory until the CIL flushes. This can cause
325 * issues when we are running low on memory - pinned memory cannot be reclaimed,
326 * and the CIL consumes a lot of memory. Hence we need to set an upper physical
327 * size limit for the CIL that limits the maximum amount of memory pinned by the
328 * CIL but does not limit performance by reducing relogging efficiency
331 * As such, the CIL push threshold ends up being the smaller of two thresholds:
332 * - a threshold large enough that it allows CIL to be pushed and progress to be
333 * made without excessive blocking of incoming transaction commits. This is
334 * defined to be 12.5% of the log space - half the 25% push threshold of the
336 * - small enough that it doesn't pin excessive amounts of memory but maintains
337 * close to peak relogging efficiency. This is defined to be 16x the iclog
338 * buffer window (32MB) as measurements have shown this to be roughly the
339 * point of diminishing performance increases under highly concurrent
340 * modification workloads.
342 * To prevent the CIL from overflowing upper commit size bounds, we introduce a
343 * new threshold at which we block committing transactions until the background
344 * CIL commit commences and switches to a new context. While this is not a hard
345 * limit, it forces the process committing a transaction to the CIL to block and
346 * yeild the CPU, giving the CIL push work a chance to be scheduled and start
347 * work. This prevents a process running lots of transactions from overfilling
348 * the CIL because it is not yielding the CPU. We set the blocking limit at
349 * twice the background push space threshold so we keep in line with the AIL
352 * Note: this is not a -hard- limit as blocking is applied after the transaction
353 * is inserted into the CIL and the push has been triggered. It is largely a
354 * throttling mechanism that allows the CIL push to be scheduled and run. A hard
355 * limit will be difficult to implement without introducing global serialisation
356 * in the CIL commit fast path, and it's not at all clear that we actually need
357 * such hard limits given the ~7 years we've run without a hard limit before
358 * finding the first situation where a checkpoint size overflow actually
359 * occurred. Hence the simple throttle, and an ASSERT check to tell us that
360 * we've overrun the max size.
362 #define XLOG_CIL_SPACE_LIMIT(log) \
363 min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4)
365 #define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \
366 (XLOG_CIL_SPACE_LIMIT(log) * 2)
369 * ticket grant locks, queues and accounting have their own cachlines
370 * as these are quite hot and can be operated on concurrently.
372 struct xlog_grant_head
{
373 spinlock_t lock ____cacheline_aligned_in_smp
;
374 struct list_head waiters
;
379 * The reservation head lsn is not made up of a cycle number and block number.
380 * Instead, it uses a cycle number and byte number. Logs don't expect to
381 * overflow 31 bits worth of byte offset, so using a byte number will mean
382 * that round off problems won't occur when releasing partial reservations.
385 /* The following fields don't need locking */
386 struct xfs_mount
*l_mp
; /* mount point */
387 struct xfs_ail
*l_ailp
; /* AIL log is working with */
388 struct xfs_cil
*l_cilp
; /* CIL log is working with */
389 struct xfs_buftarg
*l_targ
; /* buftarg of log */
390 struct workqueue_struct
*l_ioend_workqueue
; /* for I/O completions */
391 struct delayed_work l_work
; /* background flush work */
393 uint l_quotaoffs_flag
; /* XFS_DQ_*, for QUOTAOFFs */
394 struct list_head
*l_buf_cancel_table
;
395 int l_iclog_hsize
; /* size of iclog header */
396 int l_iclog_heads
; /* # of iclog header sectors */
397 uint l_sectBBsize
; /* sector size in BBs (2^n) */
398 int l_iclog_size
; /* size of log in bytes */
399 int l_iclog_bufs
; /* number of iclog buffers */
400 xfs_daddr_t l_logBBstart
; /* start block of log */
401 int l_logsize
; /* size of log in bytes */
402 int l_logBBsize
; /* size of log in BB chunks */
404 /* The following block of fields are changed while holding icloglock */
405 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp
;
406 /* waiting for iclog flush */
407 int l_covered_state
;/* state of "covering disk
409 xlog_in_core_t
*l_iclog
; /* head log queue */
410 spinlock_t l_icloglock
; /* grab to change iclog state */
411 int l_curr_cycle
; /* Cycle number of log writes */
412 int l_prev_cycle
; /* Cycle number before last
414 int l_curr_block
; /* current logical log block */
415 int l_prev_block
; /* previous logical log block */
418 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and
419 * read without needing to hold specific locks. To avoid operations
420 * contending with other hot objects, place each of them on a separate
423 /* lsn of last LR on disk */
424 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp
;
425 /* lsn of 1st LR with unflushed * buffers */
426 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp
;
428 struct xlog_grant_head l_reserve_head
;
429 struct xlog_grant_head l_write_head
;
431 struct xfs_kobj l_kobj
;
433 /* The following field are used for debugging; need to hold icloglock */
435 void *l_iclog_bak
[XLOG_MAX_ICLOGS
];
437 /* log recovery lsn tracking (for buffer submission */
438 xfs_lsn_t l_recovery_lsn
;
441 #define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
442 ((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE))
444 #define XLOG_FORCED_SHUTDOWN(log) \
445 (unlikely((log)->l_flags & XLOG_IO_ERROR))
447 /* common routines */
455 xlog_recover_cancel(struct xlog
*);
457 extern __le32
xlog_cksum(struct xlog
*log
, struct xlog_rec_header
*rhead
,
460 extern kmem_zone_t
*xfs_log_ticket_zone
;
470 xlog_write_adv_cnt(void **ptr
, int *len
, int *off
, size_t bytes
)
477 void xlog_print_tic_res(struct xfs_mount
*mp
, struct xlog_ticket
*ticket
);
478 void xlog_print_trans(struct xfs_trans
*);
479 int xlog_write(struct xlog
*log
, struct xfs_log_vec
*log_vector
,
480 struct xlog_ticket
*tic
, xfs_lsn_t
*start_lsn
,
481 struct xlog_in_core
**commit_iclog
, uint flags
,
482 bool need_start_rec
);
483 int xlog_commit_record(struct xlog
*log
, struct xlog_ticket
*ticket
,
484 struct xlog_in_core
**iclog
, xfs_lsn_t
*lsn
);
485 void xfs_log_ticket_ungrant(struct xlog
*log
, struct xlog_ticket
*ticket
);
486 void xfs_log_ticket_regrant(struct xlog
*log
, struct xlog_ticket
*ticket
);
489 * When we crack an atomic LSN, we sample it first so that the value will not
490 * change while we are cracking it into the component values. This means we
491 * will always get consistent component values to work from. This should always
492 * be used to sample and crack LSNs that are stored and updated in atomic
496 xlog_crack_atomic_lsn(atomic64_t
*lsn
, uint
*cycle
, uint
*block
)
498 xfs_lsn_t val
= atomic64_read(lsn
);
500 *cycle
= CYCLE_LSN(val
);
501 *block
= BLOCK_LSN(val
);
505 * Calculate and assign a value to an atomic LSN variable from component pieces.
508 xlog_assign_atomic_lsn(atomic64_t
*lsn
, uint cycle
, uint block
)
510 atomic64_set(lsn
, xlog_assign_lsn(cycle
, block
));
514 * When we crack the grant head, we sample it first so that the value will not
515 * change while we are cracking it into the component values. This means we
516 * will always get consistent component values to work from.
519 xlog_crack_grant_head_val(int64_t val
, int *cycle
, int *space
)
522 *space
= val
& 0xffffffff;
526 xlog_crack_grant_head(atomic64_t
*head
, int *cycle
, int *space
)
528 xlog_crack_grant_head_val(atomic64_read(head
), cycle
, space
);
531 static inline int64_t
532 xlog_assign_grant_head_val(int cycle
, int space
)
534 return ((int64_t)cycle
<< 32) | space
;
538 xlog_assign_grant_head(atomic64_t
*head
, int cycle
, int space
)
540 atomic64_set(head
, xlog_assign_grant_head_val(cycle
, space
));
544 * Committed Item List interfaces
546 int xlog_cil_init(struct xlog
*log
);
547 void xlog_cil_init_post_recovery(struct xlog
*log
);
548 void xlog_cil_destroy(struct xlog
*log
);
549 bool xlog_cil_empty(struct xlog
*log
);
560 xlog_cil_force(struct xlog
*log
)
562 xlog_cil_force_lsn(log
, log
->l_cilp
->xc_current_sequence
);
566 * Wrapper function for waiting on a wait queue serialised against wakeups
567 * by a spinlock. This matches the semantics of all the wait queues used in the
572 struct wait_queue_head
*wq
,
573 struct spinlock
*lock
)
576 DECLARE_WAITQUEUE(wait
, current
);
578 add_wait_queue_exclusive(wq
, &wait
);
579 __set_current_state(TASK_UNINTERRUPTIBLE
);
582 remove_wait_queue(wq
, &wait
);
586 * The LSN is valid so long as it is behind the current LSN. If it isn't, this
587 * means that the next log record that includes this metadata could have a
588 * smaller LSN. In turn, this means that the modification in the log would not
601 * First, sample the current lsn without locking to avoid added
602 * contention from metadata I/O. The current cycle and block are updated
603 * (in xlog_state_switch_iclogs()) and read here in a particular order
604 * to avoid false negatives (e.g., thinking the metadata LSN is valid
607 * The current block is always rewound before the cycle is bumped in
608 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in
609 * a transiently forward state. Instead, we can see the LSN in a
610 * transiently behind state if we happen to race with a cycle wrap.
612 cur_cycle
= READ_ONCE(log
->l_curr_cycle
);
614 cur_block
= READ_ONCE(log
->l_curr_block
);
616 if ((CYCLE_LSN(lsn
) > cur_cycle
) ||
617 (CYCLE_LSN(lsn
) == cur_cycle
&& BLOCK_LSN(lsn
) > cur_block
)) {
619 * If the metadata LSN appears invalid, it's possible the check
620 * above raced with a wrap to the next log cycle. Grab the lock
623 spin_lock(&log
->l_icloglock
);
624 cur_cycle
= log
->l_curr_cycle
;
625 cur_block
= log
->l_curr_block
;
626 spin_unlock(&log
->l_icloglock
);
628 if ((CYCLE_LSN(lsn
) > cur_cycle
) ||
629 (CYCLE_LSN(lsn
) == cur_cycle
&& BLOCK_LSN(lsn
) > cur_block
))
636 #endif /* __XFS_LOG_PRIV_H__ */