Loading...
1/*
2 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#ifndef __XFS_LOG_PRIV_H__
19#define __XFS_LOG_PRIV_H__
20
21struct xfs_buf;
22struct log;
23struct xlog_ticket;
24struct xfs_mount;
25
26/*
27 * Macros, structures, prototypes for internal log manager use.
28 */
29
30#define XLOG_MIN_ICLOGS 2
31#define XLOG_MAX_ICLOGS 8
32#define XLOG_HEADER_MAGIC_NUM 0xFEEDbabe /* Invalid cycle number */
33#define XLOG_VERSION_1 1
34#define XLOG_VERSION_2 2 /* Large IClogs, Log sunit */
35#define XLOG_VERSION_OKBITS (XLOG_VERSION_1 | XLOG_VERSION_2)
36#define XLOG_MIN_RECORD_BSIZE (16*1024) /* eventually 32k */
37#define XLOG_BIG_RECORD_BSIZE (32*1024) /* 32k buffers */
38#define XLOG_MAX_RECORD_BSIZE (256*1024)
39#define XLOG_HEADER_CYCLE_SIZE (32*1024) /* cycle data in header */
40#define XLOG_MIN_RECORD_BSHIFT 14 /* 16384 == 1 << 14 */
41#define XLOG_BIG_RECORD_BSHIFT 15 /* 32k == 1 << 15 */
42#define XLOG_MAX_RECORD_BSHIFT 18 /* 256k == 1 << 18 */
43#define XLOG_BTOLSUNIT(log, b) (((b)+(log)->l_mp->m_sb.sb_logsunit-1) / \
44 (log)->l_mp->m_sb.sb_logsunit)
45#define XLOG_LSUNITTOB(log, su) ((su) * (log)->l_mp->m_sb.sb_logsunit)
46
47#define XLOG_HEADER_SIZE 512
48
49#define XLOG_REC_SHIFT(log) \
50 BTOBB(1 << (xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? \
51 XLOG_MAX_RECORD_BSHIFT : XLOG_BIG_RECORD_BSHIFT))
52#define XLOG_TOTAL_REC_SHIFT(log) \
53 BTOBB(XLOG_MAX_ICLOGS << (xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? \
54 XLOG_MAX_RECORD_BSHIFT : XLOG_BIG_RECORD_BSHIFT))
55
56static inline xfs_lsn_t xlog_assign_lsn(uint cycle, uint block)
57{
58 return ((xfs_lsn_t)cycle << 32) | block;
59}
60
61static inline uint xlog_get_cycle(char *ptr)
62{
63 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
64 return be32_to_cpu(*((__be32 *)ptr + 1));
65 else
66 return be32_to_cpu(*(__be32 *)ptr);
67}
68
69#define BLK_AVG(blk1, blk2) ((blk1+blk2) >> 1)
70
71#ifdef __KERNEL__
72
73/*
74 * get client id from packed copy.
75 *
76 * this hack is here because the xlog_pack code copies four bytes
77 * of xlog_op_header containing the fields oh_clientid, oh_flags
78 * and oh_res2 into the packed copy.
79 *
80 * later on this four byte chunk is treated as an int and the
81 * client id is pulled out.
82 *
83 * this has endian issues, of course.
84 */
85static inline uint xlog_get_client_id(__be32 i)
86{
87 return be32_to_cpu(i) >> 24;
88}
89
90/*
91 * In core log state
92 */
93#define XLOG_STATE_ACTIVE 0x0001 /* Current IC log being written to */
94#define XLOG_STATE_WANT_SYNC 0x0002 /* Want to sync this iclog; no more writes */
95#define XLOG_STATE_SYNCING 0x0004 /* This IC log is syncing */
96#define XLOG_STATE_DONE_SYNC 0x0008 /* Done syncing to disk */
97#define XLOG_STATE_DO_CALLBACK \
98 0x0010 /* Process callback functions */
99#define XLOG_STATE_CALLBACK 0x0020 /* Callback functions now */
100#define XLOG_STATE_DIRTY 0x0040 /* Dirty IC log, not ready for ACTIVE status*/
101#define XLOG_STATE_IOERROR 0x0080 /* IO error happened in sync'ing log */
102#define XLOG_STATE_ALL 0x7FFF /* All possible valid flags */
103#define XLOG_STATE_NOTUSED 0x8000 /* This IC log not being used */
104#endif /* __KERNEL__ */
105
106/*
107 * Flags to log operation header
108 *
109 * The first write of a new transaction will be preceded with a start
110 * record, XLOG_START_TRANS. Once a transaction is committed, a commit
111 * record is written, XLOG_COMMIT_TRANS. If a single region can not fit into
112 * the remainder of the current active in-core log, it is split up into
113 * multiple regions. Each partial region will be marked with a
114 * XLOG_CONTINUE_TRANS until the last one, which gets marked with XLOG_END_TRANS.
115 *
116 */
117#define XLOG_START_TRANS 0x01 /* Start a new transaction */
118#define XLOG_COMMIT_TRANS 0x02 /* Commit this transaction */
119#define XLOG_CONTINUE_TRANS 0x04 /* Cont this trans into new region */
120#define XLOG_WAS_CONT_TRANS 0x08 /* Cont this trans into new region */
121#define XLOG_END_TRANS 0x10 /* End a continued transaction */
122#define XLOG_UNMOUNT_TRANS 0x20 /* Unmount a filesystem transaction */
123
124#ifdef __KERNEL__
125/*
126 * Flags to log ticket
127 */
128#define XLOG_TIC_INITED 0x1 /* has been initialized */
129#define XLOG_TIC_PERM_RESERV 0x2 /* permanent reservation */
130
131#define XLOG_TIC_FLAGS \
132 { XLOG_TIC_INITED, "XLOG_TIC_INITED" }, \
133 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }
134
135#endif /* __KERNEL__ */
136
137#define XLOG_UNMOUNT_TYPE 0x556e /* Un for Unmount */
138
139/*
140 * Flags for log structure
141 */
142#define XLOG_CHKSUM_MISMATCH 0x1 /* used only during recovery */
143#define XLOG_ACTIVE_RECOVERY 0x2 /* in the middle of recovery */
144#define XLOG_RECOVERY_NEEDED 0x4 /* log was recovered */
145#define XLOG_IO_ERROR 0x8 /* log hit an I/O error, and being
146 shutdown */
147#define XLOG_TAIL_WARN 0x10 /* log tail verify warning issued */
148
149typedef __uint32_t xlog_tid_t;
150
151#ifdef __KERNEL__
152/*
153 * Below are states for covering allocation transactions.
154 * By covering, we mean changing the h_tail_lsn in the last on-disk
155 * log write such that no allocation transactions will be re-done during
156 * recovery after a system crash. Recovery starts at the last on-disk
157 * log write.
158 *
159 * These states are used to insert dummy log entries to cover
160 * space allocation transactions which can undo non-transactional changes
161 * after a crash. Writes to a file with space
162 * already allocated do not result in any transactions. Allocations
163 * might include space beyond the EOF. So if we just push the EOF a
164 * little, the last transaction for the file could contain the wrong
165 * size. If there is no file system activity, after an allocation
166 * transaction, and the system crashes, the allocation transaction
167 * will get replayed and the file will be truncated. This could
168 * be hours/days/... after the allocation occurred.
169 *
170 * The fix for this is to do two dummy transactions when the
171 * system is idle. We need two dummy transaction because the h_tail_lsn
172 * in the log record header needs to point beyond the last possible
173 * non-dummy transaction. The first dummy changes the h_tail_lsn to
174 * the first transaction before the dummy. The second dummy causes
175 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn.
176 *
177 * These dummy transactions get committed when everything
178 * is idle (after there has been some activity).
179 *
180 * There are 5 states used to control this.
181 *
182 * IDLE -- no logging has been done on the file system or
183 * we are done covering previous transactions.
184 * NEED -- logging has occurred and we need a dummy transaction
185 * when the log becomes idle.
186 * DONE -- we were in the NEED state and have committed a dummy
187 * transaction.
188 * NEED2 -- we detected that a dummy transaction has gone to the
189 * on disk log with no other transactions.
190 * DONE2 -- we committed a dummy transaction when in the NEED2 state.
191 *
192 * There are two places where we switch states:
193 *
194 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
195 * We commit the dummy transaction and switch to DONE or DONE2,
196 * respectively. In all other states, we don't do anything.
197 *
198 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
199 *
200 * No matter what state we are in, if this isn't the dummy
201 * transaction going out, the next state is NEED.
202 * So, if we aren't in the DONE or DONE2 states, the next state
203 * is NEED. We can't be finishing a write of the dummy record
204 * unless it was committed and the state switched to DONE or DONE2.
205 *
206 * If we are in the DONE state and this was a write of the
207 * dummy transaction, we move to NEED2.
208 *
209 * If we are in the DONE2 state and this was a write of the
210 * dummy transaction, we move to IDLE.
211 *
212 *
213 * Writing only one dummy transaction can get appended to
214 * one file space allocation. When this happens, the log recovery
215 * code replays the space allocation and a file could be truncated.
216 * This is why we have the NEED2 and DONE2 states before going idle.
217 */
218
219#define XLOG_STATE_COVER_IDLE 0
220#define XLOG_STATE_COVER_NEED 1
221#define XLOG_STATE_COVER_DONE 2
222#define XLOG_STATE_COVER_NEED2 3
223#define XLOG_STATE_COVER_DONE2 4
224
225#define XLOG_COVER_OPS 5
226
227
228/* Ticket reservation region accounting */
229#define XLOG_TIC_LEN_MAX 15
230
231/*
232 * Reservation region
233 * As would be stored in xfs_log_iovec but without the i_addr which
234 * we don't care about.
235 */
236typedef struct xlog_res {
237 uint r_len; /* region length :4 */
238 uint r_type; /* region's transaction type :4 */
239} xlog_res_t;
240
241typedef struct xlog_ticket {
242 wait_queue_head_t t_wait; /* ticket wait queue */
243 struct list_head t_queue; /* reserve/write queue */
244 xlog_tid_t t_tid; /* transaction identifier : 4 */
245 atomic_t t_ref; /* ticket reference count : 4 */
246 int t_curr_res; /* current reservation in bytes : 4 */
247 int t_unit_res; /* unit reservation in bytes : 4 */
248 char t_ocnt; /* original count : 1 */
249 char t_cnt; /* current count : 1 */
250 char t_clientid; /* who does this belong to; : 1 */
251 char t_flags; /* properties of reservation : 1 */
252 uint t_trans_type; /* transaction type : 4 */
253
254 /* reservation array fields */
255 uint t_res_num; /* num in array : 4 */
256 uint t_res_num_ophdrs; /* num op hdrs : 4 */
257 uint t_res_arr_sum; /* array sum : 4 */
258 uint t_res_o_flow; /* sum overflow : 4 */
259 xlog_res_t t_res_arr[XLOG_TIC_LEN_MAX]; /* array of res : 8 * 15 */
260} xlog_ticket_t;
261
262#endif
263
264
265typedef struct xlog_op_header {
266 __be32 oh_tid; /* transaction id of operation : 4 b */
267 __be32 oh_len; /* bytes in data region : 4 b */
268 __u8 oh_clientid; /* who sent me this : 1 b */
269 __u8 oh_flags; /* : 1 b */
270 __u16 oh_res2; /* 32 bit align : 2 b */
271} xlog_op_header_t;
272
273
274/* valid values for h_fmt */
275#define XLOG_FMT_UNKNOWN 0
276#define XLOG_FMT_LINUX_LE 1
277#define XLOG_FMT_LINUX_BE 2
278#define XLOG_FMT_IRIX_BE 3
279
280/* our fmt */
281#ifdef XFS_NATIVE_HOST
282#define XLOG_FMT XLOG_FMT_LINUX_BE
283#else
284#define XLOG_FMT XLOG_FMT_LINUX_LE
285#endif
286
287typedef struct xlog_rec_header {
288 __be32 h_magicno; /* log record (LR) identifier : 4 */
289 __be32 h_cycle; /* write cycle of log : 4 */
290 __be32 h_version; /* LR version : 4 */
291 __be32 h_len; /* len in bytes; should be 64-bit aligned: 4 */
292 __be64 h_lsn; /* lsn of this LR : 8 */
293 __be64 h_tail_lsn; /* lsn of 1st LR w/ buffers not committed: 8 */
294 __be32 h_chksum; /* may not be used; non-zero if used : 4 */
295 __be32 h_prev_block; /* block number to previous LR : 4 */
296 __be32 h_num_logops; /* number of log operations in this LR : 4 */
297 __be32 h_cycle_data[XLOG_HEADER_CYCLE_SIZE / BBSIZE];
298 /* new fields */
299 __be32 h_fmt; /* format of log record : 4 */
300 uuid_t h_fs_uuid; /* uuid of FS : 16 */
301 __be32 h_size; /* iclog size : 4 */
302} xlog_rec_header_t;
303
304typedef struct xlog_rec_ext_header {
305 __be32 xh_cycle; /* write cycle of log : 4 */
306 __be32 xh_cycle_data[XLOG_HEADER_CYCLE_SIZE / BBSIZE]; /* : 256 */
307} xlog_rec_ext_header_t;
308
309#ifdef __KERNEL__
310
311/*
312 * Quite misnamed, because this union lays out the actual on-disk log buffer.
313 */
314typedef union xlog_in_core2 {
315 xlog_rec_header_t hic_header;
316 xlog_rec_ext_header_t hic_xheader;
317 char hic_sector[XLOG_HEADER_SIZE];
318} xlog_in_core_2_t;
319
320/*
321 * - A log record header is 512 bytes. There is plenty of room to grow the
322 * xlog_rec_header_t into the reserved space.
323 * - ic_data follows, so a write to disk can start at the beginning of
324 * the iclog.
325 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
326 * - ic_next is the pointer to the next iclog in the ring.
327 * - ic_bp is a pointer to the buffer used to write this incore log to disk.
328 * - ic_log is a pointer back to the global log structure.
329 * - ic_callback is a linked list of callback function/argument pairs to be
330 * called after an iclog finishes writing.
331 * - ic_size is the full size of the header plus data.
332 * - ic_offset is the current number of bytes written to in this iclog.
333 * - ic_refcnt is bumped when someone is writing to the log.
334 * - ic_state is the state of the iclog.
335 *
336 * Because of cacheline contention on large machines, we need to separate
337 * various resources onto different cachelines. To start with, make the
338 * structure cacheline aligned. The following fields can be contended on
339 * by independent processes:
340 *
341 * - ic_callback_*
342 * - ic_refcnt
343 * - fields protected by the global l_icloglock
344 *
345 * so we need to ensure that these fields are located in separate cachelines.
346 * We'll put all the read-only and l_icloglock fields in the first cacheline,
347 * and move everything else out to subsequent cachelines.
348 */
349typedef struct xlog_in_core {
350 wait_queue_head_t ic_force_wait;
351 wait_queue_head_t ic_write_wait;
352 struct xlog_in_core *ic_next;
353 struct xlog_in_core *ic_prev;
354 struct xfs_buf *ic_bp;
355 struct log *ic_log;
356 int ic_size;
357 int ic_offset;
358 int ic_bwritecnt;
359 unsigned short ic_state;
360 char *ic_datap; /* pointer to iclog data */
361
362 /* Callback structures need their own cacheline */
363 spinlock_t ic_callback_lock ____cacheline_aligned_in_smp;
364 xfs_log_callback_t *ic_callback;
365 xfs_log_callback_t **ic_callback_tail;
366
367 /* reference counts need their own cacheline */
368 atomic_t ic_refcnt ____cacheline_aligned_in_smp;
369 xlog_in_core_2_t *ic_data;
370#define ic_header ic_data->hic_header
371} xlog_in_core_t;
372
373/*
374 * The CIL context is used to aggregate per-transaction details as well be
375 * passed to the iclog for checkpoint post-commit processing. After being
376 * passed to the iclog, another context needs to be allocated for tracking the
377 * next set of transactions to be aggregated into a checkpoint.
378 */
379struct xfs_cil;
380
381struct xfs_cil_ctx {
382 struct xfs_cil *cil;
383 xfs_lsn_t sequence; /* chkpt sequence # */
384 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */
385 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */
386 struct xlog_ticket *ticket; /* chkpt ticket */
387 int nvecs; /* number of regions */
388 int space_used; /* aggregate size of regions */
389 struct list_head busy_extents; /* busy extents in chkpt */
390 struct xfs_log_vec *lv_chain; /* logvecs being pushed */
391 xfs_log_callback_t log_cb; /* completion callback hook. */
392 struct list_head committing; /* ctx committing list */
393};
394
395/*
396 * Committed Item List structure
397 *
398 * This structure is used to track log items that have been committed but not
399 * yet written into the log. It is used only when the delayed logging mount
400 * option is enabled.
401 *
402 * This structure tracks the list of committing checkpoint contexts so
403 * we can avoid the problem of having to hold out new transactions during a
404 * flush until we have a the commit record LSN of the checkpoint. We can
405 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
406 * sequence match and extract the commit LSN directly from there. If the
407 * checkpoint is still in the process of committing, we can block waiting for
408 * the commit LSN to be determined as well. This should make synchronous
409 * operations almost as efficient as the old logging methods.
410 */
411struct xfs_cil {
412 struct log *xc_log;
413 struct list_head xc_cil;
414 spinlock_t xc_cil_lock;
415 struct xfs_cil_ctx *xc_ctx;
416 struct rw_semaphore xc_ctx_lock;
417 struct list_head xc_committing;
418 wait_queue_head_t xc_commit_wait;
419 xfs_lsn_t xc_current_sequence;
420};
421
422/*
423 * The amount of log space we allow the CIL to aggregate is difficult to size.
424 * Whatever we choose, we have to make sure we can get a reservation for the
425 * log space effectively, that it is large enough to capture sufficient
426 * relogging to reduce log buffer IO significantly, but it is not too large for
427 * the log or induces too much latency when writing out through the iclogs. We
428 * track both space consumed and the number of vectors in the checkpoint
429 * context, so we need to decide which to use for limiting.
430 *
431 * Every log buffer we write out during a push needs a header reserved, which
432 * is at least one sector and more for v2 logs. Hence we need a reservation of
433 * at least 512 bytes per 32k of log space just for the LR headers. That means
434 * 16KB of reservation per megabyte of delayed logging space we will consume,
435 * plus various headers. The number of headers will vary based on the num of
436 * io vectors, so limiting on a specific number of vectors is going to result
437 * in transactions of varying size. IOWs, it is more consistent to track and
438 * limit space consumed in the log rather than by the number of objects being
439 * logged in order to prevent checkpoint ticket overruns.
440 *
441 * Further, use of static reservations through the log grant mechanism is
442 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
443 * grant) and a significant deadlock potential because regranting write space
444 * can block on log pushes. Hence if we have to regrant log space during a log
445 * push, we can deadlock.
446 *
447 * However, we can avoid this by use of a dynamic "reservation stealing"
448 * technique during transaction commit whereby unused reservation space in the
449 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
450 * space needed by the checkpoint transaction. This means that we never need to
451 * specifically reserve space for the CIL checkpoint transaction, nor do we
452 * need to regrant space once the checkpoint completes. This also means the
453 * checkpoint transaction ticket is specific to the checkpoint context, rather
454 * than the CIL itself.
455 *
456 * With dynamic reservations, we can effectively make up arbitrary limits for
457 * the checkpoint size so long as they don't violate any other size rules.
458 * Recovery imposes a rule that no transaction exceed half the log, so we are
459 * limited by that. Furthermore, the log transaction reservation subsystem
460 * tries to keep 25% of the log free, so we need to keep below that limit or we
461 * risk running out of free log space to start any new transactions.
462 *
463 * In order to keep background CIL push efficient, we will set a lower
464 * threshold at which background pushing is attempted without blocking current
465 * transaction commits. A separate, higher bound defines when CIL pushes are
466 * enforced to ensure we stay within our maximum checkpoint size bounds.
467 * threshold, yet give us plenty of space for aggregation on large logs.
468 */
469#define XLOG_CIL_SPACE_LIMIT(log) (log->l_logsize >> 3)
470#define XLOG_CIL_HARD_SPACE_LIMIT(log) (3 * (log->l_logsize >> 4))
471
472/*
473 * The reservation head lsn is not made up of a cycle number and block number.
474 * Instead, it uses a cycle number and byte number. Logs don't expect to
475 * overflow 31 bits worth of byte offset, so using a byte number will mean
476 * that round off problems won't occur when releasing partial reservations.
477 */
478typedef struct log {
479 /* The following fields don't need locking */
480 struct xfs_mount *l_mp; /* mount point */
481 struct xfs_ail *l_ailp; /* AIL log is working with */
482 struct xfs_cil *l_cilp; /* CIL log is working with */
483 struct xfs_buf *l_xbuf; /* extra buffer for log
484 * wrapping */
485 struct xfs_buftarg *l_targ; /* buftarg of log */
486 uint l_flags;
487 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
488 struct list_head *l_buf_cancel_table;
489 int l_iclog_hsize; /* size of iclog header */
490 int l_iclog_heads; /* # of iclog header sectors */
491 uint l_sectBBsize; /* sector size in BBs (2^n) */
492 int l_iclog_size; /* size of log in bytes */
493 int l_iclog_size_log; /* log power size of log */
494 int l_iclog_bufs; /* number of iclog buffers */
495 xfs_daddr_t l_logBBstart; /* start block of log */
496 int l_logsize; /* size of log in bytes */
497 int l_logBBsize; /* size of log in BB chunks */
498
499 /* The following block of fields are changed while holding icloglock */
500 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
501 /* waiting for iclog flush */
502 int l_covered_state;/* state of "covering disk
503 * log entries" */
504 xlog_in_core_t *l_iclog; /* head log queue */
505 spinlock_t l_icloglock; /* grab to change iclog state */
506 int l_curr_cycle; /* Cycle number of log writes */
507 int l_prev_cycle; /* Cycle number before last
508 * block increment */
509 int l_curr_block; /* current logical log block */
510 int l_prev_block; /* previous logical log block */
511
512 /*
513 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and
514 * read without needing to hold specific locks. To avoid operations
515 * contending with other hot objects, place each of them on a separate
516 * cacheline.
517 */
518 /* lsn of last LR on disk */
519 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp;
520 /* lsn of 1st LR with unflushed * buffers */
521 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
522
523 /*
524 * ticket grant locks, queues and accounting have their own cachlines
525 * as these are quite hot and can be operated on concurrently.
526 */
527 spinlock_t l_grant_reserve_lock ____cacheline_aligned_in_smp;
528 struct list_head l_reserveq;
529 atomic64_t l_grant_reserve_head;
530
531 spinlock_t l_grant_write_lock ____cacheline_aligned_in_smp;
532 struct list_head l_writeq;
533 atomic64_t l_grant_write_head;
534
535 /* The following field are used for debugging; need to hold icloglock */
536#ifdef DEBUG
537 char *l_iclog_bak[XLOG_MAX_ICLOGS];
538#endif
539
540} xlog_t;
541
542#define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
543 ((log)->l_buf_cancel_table + ((__uint64_t)blkno % XLOG_BC_TABLE_SIZE))
544
545#define XLOG_FORCED_SHUTDOWN(log) ((log)->l_flags & XLOG_IO_ERROR)
546
547/* common routines */
548extern xfs_lsn_t xlog_assign_tail_lsn(struct xfs_mount *mp);
549extern int xlog_recover(xlog_t *log);
550extern int xlog_recover_finish(xlog_t *log);
551extern void xlog_pack_data(xlog_t *log, xlog_in_core_t *iclog, int);
552
553extern kmem_zone_t *xfs_log_ticket_zone;
554struct xlog_ticket *xlog_ticket_alloc(struct log *log, int unit_bytes,
555 int count, char client, uint xflags,
556 int alloc_flags);
557
558
559static inline void
560xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes)
561{
562 *ptr += bytes;
563 *len -= bytes;
564 *off += bytes;
565}
566
567void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
568int xlog_write(struct log *log, struct xfs_log_vec *log_vector,
569 struct xlog_ticket *tic, xfs_lsn_t *start_lsn,
570 xlog_in_core_t **commit_iclog, uint flags);
571
572/*
573 * When we crack an atomic LSN, we sample it first so that the value will not
574 * change while we are cracking it into the component values. This means we
575 * will always get consistent component values to work from. This should always
576 * be used to sample and crack LSNs that are stored and updated in atomic
577 * variables.
578 */
579static inline void
580xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
581{
582 xfs_lsn_t val = atomic64_read(lsn);
583
584 *cycle = CYCLE_LSN(val);
585 *block = BLOCK_LSN(val);
586}
587
588/*
589 * Calculate and assign a value to an atomic LSN variable from component pieces.
590 */
591static inline void
592xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
593{
594 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
595}
596
597/*
598 * When we crack the grant head, we sample it first so that the value will not
599 * change while we are cracking it into the component values. This means we
600 * will always get consistent component values to work from.
601 */
602static inline void
603xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
604{
605 *cycle = val >> 32;
606 *space = val & 0xffffffff;
607}
608
609static inline void
610xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
611{
612 xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
613}
614
615static inline int64_t
616xlog_assign_grant_head_val(int cycle, int space)
617{
618 return ((int64_t)cycle << 32) | space;
619}
620
621static inline void
622xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
623{
624 atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
625}
626
627/*
628 * Committed Item List interfaces
629 */
630int xlog_cil_init(struct log *log);
631void xlog_cil_init_post_recovery(struct log *log);
632void xlog_cil_destroy(struct log *log);
633
634/*
635 * CIL force routines
636 */
637xfs_lsn_t xlog_cil_force_lsn(struct log *log, xfs_lsn_t sequence);
638
639static inline void
640xlog_cil_force(struct log *log)
641{
642 xlog_cil_force_lsn(log, log->l_cilp->xc_current_sequence);
643}
644
645/*
646 * Unmount record type is used as a pseudo transaction type for the ticket.
647 * It's value must be outside the range of XFS_TRANS_* values.
648 */
649#define XLOG_UNMOUNT_REC_TYPE (-1U)
650
651/*
652 * Wrapper function for waiting on a wait queue serialised against wakeups
653 * by a spinlock. This matches the semantics of all the wait queues used in the
654 * log code.
655 */
656static inline void xlog_wait(wait_queue_head_t *wq, spinlock_t *lock)
657{
658 DECLARE_WAITQUEUE(wait, current);
659
660 add_wait_queue_exclusive(wq, &wait);
661 __set_current_state(TASK_UNINTERRUPTIBLE);
662 spin_unlock(lock);
663 schedule();
664 remove_wait_queue(wq, &wait);
665}
666#endif /* __KERNEL__ */
667
668#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
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/*
54 * Log ticket flags
55 */
56#define XLOG_TIC_PERM_RESERV 0x1 /* permanent reservation */
57
58#define XLOG_TIC_FLAGS \
59 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }
60
61/*
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
66 * log write.
67 *
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.
78 *
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.
85 *
86 * These dummy transactions get committed when everything
87 * is idle (after there has been some activity).
88 *
89 * There are 5 states used to control this.
90 *
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
96 * transaction.
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.
100 *
101 * There are two places where we switch states:
102 *
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.
106 *
107 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
108 *
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.
114 *
115 * If we are in the DONE state and this was a write of the
116 * dummy transaction, we move to NEED2.
117 *
118 * If we are in the DONE2 state and this was a write of the
119 * dummy transaction, we move to IDLE.
120 *
121 *
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.
126 */
127
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
133
134#define XLOG_COVER_OPS 5
135
136/* Ticket reservation region accounting */
137#define XLOG_TIC_LEN_MAX 15
138
139/*
140 * Reservation region
141 * As would be stored in xfs_log_iovec but without the i_addr which
142 * we don't care about.
143 */
144typedef struct xlog_res {
145 uint r_len; /* region length :4 */
146 uint r_type; /* region's transaction type :4 */
147} xlog_res_t;
148
149typedef 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 */
160
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 */
167} xlog_ticket_t;
168
169/*
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
173 * the iclog.
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.
181 *
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:
186 *
187 * - ic_callbacks
188 * - ic_refcnt
189 * - fields protected by the global l_icloglock
190 *
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.
194 */
195typedef 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;
200 struct xlog *ic_log;
201 u32 ic_size;
202 u32 ic_offset;
203 enum xlog_iclog_state ic_state;
204 char *ic_datap; /* pointer to iclog data */
205
206 /* Callback structures need their own cacheline */
207 spinlock_t ic_callback_lock ____cacheline_aligned_in_smp;
208 struct list_head ic_callbacks;
209
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
214#ifdef DEBUG
215 bool ic_fail_crc : 1;
216#endif
217 struct semaphore ic_sema;
218 struct work_struct ic_end_io_work;
219 struct bio ic_bio;
220 struct bio_vec ic_bvec[];
221} xlog_in_core_t;
222
223/*
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.
228 */
229struct xfs_cil;
230
231struct xfs_cil_ctx {
232 struct xfs_cil *cil;
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;
244};
245
246/*
247 * Committed Item List structure
248 *
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
251 * option is enabled.
252 *
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.
261 */
262struct xfs_cil {
263 struct xlog *xc_log;
264 struct list_head xc_cil;
265 spinlock_t xc_cil_lock;
266
267 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp;
268 struct xfs_cil_ctx *xc_ctx;
269
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;
278
279/*
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.
287 *
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.
297 *
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.
303 *
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.
312 *
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.
319 *
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
329 * significantly.
330 *
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
335 * AIL.
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.
341 *
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
350 * push thresholds.
351 *
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.
361 */
362#define XLOG_CIL_SPACE_LIMIT(log) \
363 min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4)
364
365#define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \
366 (XLOG_CIL_SPACE_LIMIT(log) * 2)
367
368/*
369 * ticket grant locks, queues and accounting have their own cachlines
370 * as these are quite hot and can be operated on concurrently.
371 */
372struct xlog_grant_head {
373 spinlock_t lock ____cacheline_aligned_in_smp;
374 struct list_head waiters;
375 atomic64_t grant;
376};
377
378/*
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.
383 */
384struct xlog {
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 */
392 uint l_flags;
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 */
403
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
408 * log entries" */
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
413 * block increment */
414 int l_curr_block; /* current logical log block */
415 int l_prev_block; /* previous logical log block */
416
417 /*
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
421 * cacheline.
422 */
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;
427
428 struct xlog_grant_head l_reserve_head;
429 struct xlog_grant_head l_write_head;
430
431 struct xfs_kobj l_kobj;
432
433 /* The following field are used for debugging; need to hold icloglock */
434#ifdef DEBUG
435 void *l_iclog_bak[XLOG_MAX_ICLOGS];
436#endif
437 /* log recovery lsn tracking (for buffer submission */
438 xfs_lsn_t l_recovery_lsn;
439};
440
441#define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
442 ((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE))
443
444#define XLOG_FORCED_SHUTDOWN(log) \
445 (unlikely((log)->l_flags & XLOG_IO_ERROR))
446
447/* common routines */
448extern int
449xlog_recover(
450 struct xlog *log);
451extern int
452xlog_recover_finish(
453 struct xlog *log);
454extern void
455xlog_recover_cancel(struct xlog *);
456
457extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
458 char *dp, int size);
459
460extern kmem_zone_t *xfs_log_ticket_zone;
461struct xlog_ticket *
462xlog_ticket_alloc(
463 struct xlog *log,
464 int unit_bytes,
465 int count,
466 char client,
467 bool permanent);
468
469static inline void
470xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes)
471{
472 *ptr += bytes;
473 *len -= bytes;
474 *off += bytes;
475}
476
477void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
478void xlog_print_trans(struct xfs_trans *);
479int 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);
483int xlog_commit_record(struct xlog *log, struct xlog_ticket *ticket,
484 struct xlog_in_core **iclog, xfs_lsn_t *lsn);
485void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket);
486void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket);
487
488/*
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
493 * variables.
494 */
495static inline void
496xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
497{
498 xfs_lsn_t val = atomic64_read(lsn);
499
500 *cycle = CYCLE_LSN(val);
501 *block = BLOCK_LSN(val);
502}
503
504/*
505 * Calculate and assign a value to an atomic LSN variable from component pieces.
506 */
507static inline void
508xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
509{
510 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
511}
512
513/*
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.
517 */
518static inline void
519xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
520{
521 *cycle = val >> 32;
522 *space = val & 0xffffffff;
523}
524
525static inline void
526xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
527{
528 xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
529}
530
531static inline int64_t
532xlog_assign_grant_head_val(int cycle, int space)
533{
534 return ((int64_t)cycle << 32) | space;
535}
536
537static inline void
538xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
539{
540 atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
541}
542
543/*
544 * Committed Item List interfaces
545 */
546int xlog_cil_init(struct xlog *log);
547void xlog_cil_init_post_recovery(struct xlog *log);
548void xlog_cil_destroy(struct xlog *log);
549bool xlog_cil_empty(struct xlog *log);
550
551/*
552 * CIL force routines
553 */
554xfs_lsn_t
555xlog_cil_force_lsn(
556 struct xlog *log,
557 xfs_lsn_t sequence);
558
559static inline void
560xlog_cil_force(struct xlog *log)
561{
562 xlog_cil_force_lsn(log, log->l_cilp->xc_current_sequence);
563}
564
565/*
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
568 * log code.
569 */
570static inline void
571xlog_wait(
572 struct wait_queue_head *wq,
573 struct spinlock *lock)
574 __releases(lock)
575{
576 DECLARE_WAITQUEUE(wait, current);
577
578 add_wait_queue_exclusive(wq, &wait);
579 __set_current_state(TASK_UNINTERRUPTIBLE);
580 spin_unlock(lock);
581 schedule();
582 remove_wait_queue(wq, &wait);
583}
584
585/*
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
589 * replay.
590 */
591static inline bool
592xlog_valid_lsn(
593 struct xlog *log,
594 xfs_lsn_t lsn)
595{
596 int cur_cycle;
597 int cur_block;
598 bool valid = true;
599
600 /*
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
605 * when it is not).
606 *
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.
611 */
612 cur_cycle = READ_ONCE(log->l_curr_cycle);
613 smp_rmb();
614 cur_block = READ_ONCE(log->l_curr_block);
615
616 if ((CYCLE_LSN(lsn) > cur_cycle) ||
617 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
618 /*
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
621 * to check for sure.
622 */
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);
627
628 if ((CYCLE_LSN(lsn) > cur_cycle) ||
629 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
630 valid = false;
631 }
632
633 return valid;
634}
635
636#endif /* __XFS_LOG_PRIV_H__ */