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 xlog;
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 struct list_head t_queue; /* reserve/write queue */
243 struct task_struct *t_task; /* task that owns this ticket */
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 xlog *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 xlog *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 struct work_struct xc_push_work;
421 xfs_lsn_t xc_push_seq;
422};
423
424/*
425 * The amount of log space we allow the CIL to aggregate is difficult to size.
426 * Whatever we choose, we have to make sure we can get a reservation for the
427 * log space effectively, that it is large enough to capture sufficient
428 * relogging to reduce log buffer IO significantly, but it is not too large for
429 * the log or induces too much latency when writing out through the iclogs. We
430 * track both space consumed and the number of vectors in the checkpoint
431 * context, so we need to decide which to use for limiting.
432 *
433 * Every log buffer we write out during a push needs a header reserved, which
434 * is at least one sector and more for v2 logs. Hence we need a reservation of
435 * at least 512 bytes per 32k of log space just for the LR headers. That means
436 * 16KB of reservation per megabyte of delayed logging space we will consume,
437 * plus various headers. The number of headers will vary based on the num of
438 * io vectors, so limiting on a specific number of vectors is going to result
439 * in transactions of varying size. IOWs, it is more consistent to track and
440 * limit space consumed in the log rather than by the number of objects being
441 * logged in order to prevent checkpoint ticket overruns.
442 *
443 * Further, use of static reservations through the log grant mechanism is
444 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
445 * grant) and a significant deadlock potential because regranting write space
446 * can block on log pushes. Hence if we have to regrant log space during a log
447 * push, we can deadlock.
448 *
449 * However, we can avoid this by use of a dynamic "reservation stealing"
450 * technique during transaction commit whereby unused reservation space in the
451 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
452 * space needed by the checkpoint transaction. This means that we never need to
453 * specifically reserve space for the CIL checkpoint transaction, nor do we
454 * need to regrant space once the checkpoint completes. This also means the
455 * checkpoint transaction ticket is specific to the checkpoint context, rather
456 * than the CIL itself.
457 *
458 * With dynamic reservations, we can effectively make up arbitrary limits for
459 * the checkpoint size so long as they don't violate any other size rules.
460 * Recovery imposes a rule that no transaction exceed half the log, so we are
461 * limited by that. Furthermore, the log transaction reservation subsystem
462 * tries to keep 25% of the log free, so we need to keep below that limit or we
463 * risk running out of free log space to start any new transactions.
464 *
465 * In order to keep background CIL push efficient, we will set a lower
466 * threshold at which background pushing is attempted without blocking current
467 * transaction commits. A separate, higher bound defines when CIL pushes are
468 * enforced to ensure we stay within our maximum checkpoint size bounds.
469 * threshold, yet give us plenty of space for aggregation on large logs.
470 */
471#define XLOG_CIL_SPACE_LIMIT(log) (log->l_logsize >> 3)
472#define XLOG_CIL_HARD_SPACE_LIMIT(log) (3 * (log->l_logsize >> 4))
473
474/*
475 * ticket grant locks, queues and accounting have their own cachlines
476 * as these are quite hot and can be operated on concurrently.
477 */
478struct xlog_grant_head {
479 spinlock_t lock ____cacheline_aligned_in_smp;
480 struct list_head waiters;
481 atomic64_t grant;
482};
483
484/*
485 * The reservation head lsn is not made up of a cycle number and block number.
486 * Instead, it uses a cycle number and byte number. Logs don't expect to
487 * overflow 31 bits worth of byte offset, so using a byte number will mean
488 * that round off problems won't occur when releasing partial reservations.
489 */
490typedef struct xlog {
491 /* The following fields don't need locking */
492 struct xfs_mount *l_mp; /* mount point */
493 struct xfs_ail *l_ailp; /* AIL log is working with */
494 struct xfs_cil *l_cilp; /* CIL log is working with */
495 struct xfs_buf *l_xbuf; /* extra buffer for log
496 * wrapping */
497 struct xfs_buftarg *l_targ; /* buftarg of log */
498 uint l_flags;
499 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
500 struct list_head *l_buf_cancel_table;
501 int l_iclog_hsize; /* size of iclog header */
502 int l_iclog_heads; /* # of iclog header sectors */
503 uint l_sectBBsize; /* sector size in BBs (2^n) */
504 int l_iclog_size; /* size of log in bytes */
505 int l_iclog_size_log; /* log power size of log */
506 int l_iclog_bufs; /* number of iclog buffers */
507 xfs_daddr_t l_logBBstart; /* start block of log */
508 int l_logsize; /* size of log in bytes */
509 int l_logBBsize; /* size of log in BB chunks */
510
511 /* The following block of fields are changed while holding icloglock */
512 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
513 /* waiting for iclog flush */
514 int l_covered_state;/* state of "covering disk
515 * log entries" */
516 xlog_in_core_t *l_iclog; /* head log queue */
517 spinlock_t l_icloglock; /* grab to change iclog state */
518 int l_curr_cycle; /* Cycle number of log writes */
519 int l_prev_cycle; /* Cycle number before last
520 * block increment */
521 int l_curr_block; /* current logical log block */
522 int l_prev_block; /* previous logical log block */
523
524 /*
525 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and
526 * read without needing to hold specific locks. To avoid operations
527 * contending with other hot objects, place each of them on a separate
528 * cacheline.
529 */
530 /* lsn of last LR on disk */
531 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp;
532 /* lsn of 1st LR with unflushed * buffers */
533 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
534
535 struct xlog_grant_head l_reserve_head;
536 struct xlog_grant_head l_write_head;
537
538 /* The following field are used for debugging; need to hold icloglock */
539#ifdef DEBUG
540 char *l_iclog_bak[XLOG_MAX_ICLOGS];
541#endif
542
543} xlog_t;
544
545#define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
546 ((log)->l_buf_cancel_table + ((__uint64_t)blkno % XLOG_BC_TABLE_SIZE))
547
548#define XLOG_FORCED_SHUTDOWN(log) ((log)->l_flags & XLOG_IO_ERROR)
549
550/* common routines */
551extern int xlog_recover(xlog_t *log);
552extern int xlog_recover_finish(xlog_t *log);
553extern void xlog_pack_data(xlog_t *log, xlog_in_core_t *iclog, int);
554
555extern kmem_zone_t *xfs_log_ticket_zone;
556struct xlog_ticket *
557xlog_ticket_alloc(
558 struct xlog *log,
559 int unit_bytes,
560 int count,
561 char client,
562 bool permanent,
563 xfs_km_flags_t alloc_flags);
564
565
566static inline void
567xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes)
568{
569 *ptr += bytes;
570 *len -= bytes;
571 *off += bytes;
572}
573
574void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
575int
576xlog_write(
577 struct xlog *log,
578 struct xfs_log_vec *log_vector,
579 struct xlog_ticket *tic,
580 xfs_lsn_t *start_lsn,
581 struct xlog_in_core **commit_iclog,
582 uint flags);
583
584/*
585 * When we crack an atomic LSN, we sample it first so that the value will not
586 * change while we are cracking it into the component values. This means we
587 * will always get consistent component values to work from. This should always
588 * be used to sample and crack LSNs that are stored and updated in atomic
589 * variables.
590 */
591static inline void
592xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
593{
594 xfs_lsn_t val = atomic64_read(lsn);
595
596 *cycle = CYCLE_LSN(val);
597 *block = BLOCK_LSN(val);
598}
599
600/*
601 * Calculate and assign a value to an atomic LSN variable from component pieces.
602 */
603static inline void
604xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
605{
606 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
607}
608
609/*
610 * When we crack the grant head, we sample it first so that the value will not
611 * change while we are cracking it into the component values. This means we
612 * will always get consistent component values to work from.
613 */
614static inline void
615xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
616{
617 *cycle = val >> 32;
618 *space = val & 0xffffffff;
619}
620
621static inline void
622xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
623{
624 xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
625}
626
627static inline int64_t
628xlog_assign_grant_head_val(int cycle, int space)
629{
630 return ((int64_t)cycle << 32) | space;
631}
632
633static inline void
634xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
635{
636 atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
637}
638
639/*
640 * Committed Item List interfaces
641 */
642int
643xlog_cil_init(struct xlog *log);
644void
645xlog_cil_init_post_recovery(struct xlog *log);
646void
647xlog_cil_destroy(struct xlog *log);
648
649/*
650 * CIL force routines
651 */
652xfs_lsn_t
653xlog_cil_force_lsn(
654 struct xlog *log,
655 xfs_lsn_t sequence);
656
657static inline void
658xlog_cil_force(struct xlog *log)
659{
660 xlog_cil_force_lsn(log, log->l_cilp->xc_current_sequence);
661}
662
663/*
664 * Unmount record type is used as a pseudo transaction type for the ticket.
665 * It's value must be outside the range of XFS_TRANS_* values.
666 */
667#define XLOG_UNMOUNT_REC_TYPE (-1U)
668
669/*
670 * Wrapper function for waiting on a wait queue serialised against wakeups
671 * by a spinlock. This matches the semantics of all the wait queues used in the
672 * log code.
673 */
674static inline void xlog_wait(wait_queue_head_t *wq, spinlock_t *lock)
675{
676 DECLARE_WAITQUEUE(wait, current);
677
678 add_wait_queue_exclusive(wq, &wait);
679 __set_current_state(TASK_UNINTERRUPTIBLE);
680 spin_unlock(lock);
681 schedule();
682 remove_wait_queue(wq, &wait);
683}
684#endif /* __KERNEL__ */
685
686#endif /* __XFS_LOG_PRIV_H__ */
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 xlog;
23struct xlog_ticket;
24struct xfs_mount;
25struct xfs_log_callback;
26
27/*
28 * Flags for log structure
29 */
30#define XLOG_ACTIVE_RECOVERY 0x2 /* in the middle of recovery */
31#define XLOG_RECOVERY_NEEDED 0x4 /* log was recovered */
32#define XLOG_IO_ERROR 0x8 /* log hit an I/O error, and being
33 shutdown */
34#define XLOG_TAIL_WARN 0x10 /* log tail verify warning issued */
35
36/*
37 * get client id from packed copy.
38 *
39 * this hack is here because the xlog_pack code copies four bytes
40 * of xlog_op_header containing the fields oh_clientid, oh_flags
41 * and oh_res2 into the packed copy.
42 *
43 * later on this four byte chunk is treated as an int and the
44 * client id is pulled out.
45 *
46 * this has endian issues, of course.
47 */
48static inline uint xlog_get_client_id(__be32 i)
49{
50 return be32_to_cpu(i) >> 24;
51}
52
53/*
54 * In core log state
55 */
56#define XLOG_STATE_ACTIVE 0x0001 /* Current IC log being written to */
57#define XLOG_STATE_WANT_SYNC 0x0002 /* Want to sync this iclog; no more writes */
58#define XLOG_STATE_SYNCING 0x0004 /* This IC log is syncing */
59#define XLOG_STATE_DONE_SYNC 0x0008 /* Done syncing to disk */
60#define XLOG_STATE_DO_CALLBACK \
61 0x0010 /* Process callback functions */
62#define XLOG_STATE_CALLBACK 0x0020 /* Callback functions now */
63#define XLOG_STATE_DIRTY 0x0040 /* Dirty IC log, not ready for ACTIVE status*/
64#define XLOG_STATE_IOERROR 0x0080 /* IO error happened in sync'ing log */
65#define XLOG_STATE_IOABORT 0x0100 /* force abort on I/O completion (debug) */
66#define XLOG_STATE_ALL 0x7FFF /* All possible valid flags */
67#define XLOG_STATE_NOTUSED 0x8000 /* This IC log not being used */
68
69/*
70 * Flags to log ticket
71 */
72#define XLOG_TIC_INITED 0x1 /* has been initialized */
73#define XLOG_TIC_PERM_RESERV 0x2 /* permanent reservation */
74
75#define XLOG_TIC_FLAGS \
76 { XLOG_TIC_INITED, "XLOG_TIC_INITED" }, \
77 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }
78
79/*
80 * Below are states for covering allocation transactions.
81 * By covering, we mean changing the h_tail_lsn in the last on-disk
82 * log write such that no allocation transactions will be re-done during
83 * recovery after a system crash. Recovery starts at the last on-disk
84 * log write.
85 *
86 * These states are used to insert dummy log entries to cover
87 * space allocation transactions which can undo non-transactional changes
88 * after a crash. Writes to a file with space
89 * already allocated do not result in any transactions. Allocations
90 * might include space beyond the EOF. So if we just push the EOF a
91 * little, the last transaction for the file could contain the wrong
92 * size. If there is no file system activity, after an allocation
93 * transaction, and the system crashes, the allocation transaction
94 * will get replayed and the file will be truncated. This could
95 * be hours/days/... after the allocation occurred.
96 *
97 * The fix for this is to do two dummy transactions when the
98 * system is idle. We need two dummy transaction because the h_tail_lsn
99 * in the log record header needs to point beyond the last possible
100 * non-dummy transaction. The first dummy changes the h_tail_lsn to
101 * the first transaction before the dummy. The second dummy causes
102 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn.
103 *
104 * These dummy transactions get committed when everything
105 * is idle (after there has been some activity).
106 *
107 * There are 5 states used to control this.
108 *
109 * IDLE -- no logging has been done on the file system or
110 * we are done covering previous transactions.
111 * NEED -- logging has occurred and we need a dummy transaction
112 * when the log becomes idle.
113 * DONE -- we were in the NEED state and have committed a dummy
114 * transaction.
115 * NEED2 -- we detected that a dummy transaction has gone to the
116 * on disk log with no other transactions.
117 * DONE2 -- we committed a dummy transaction when in the NEED2 state.
118 *
119 * There are two places where we switch states:
120 *
121 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
122 * We commit the dummy transaction and switch to DONE or DONE2,
123 * respectively. In all other states, we don't do anything.
124 *
125 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
126 *
127 * No matter what state we are in, if this isn't the dummy
128 * transaction going out, the next state is NEED.
129 * So, if we aren't in the DONE or DONE2 states, the next state
130 * is NEED. We can't be finishing a write of the dummy record
131 * unless it was committed and the state switched to DONE or DONE2.
132 *
133 * If we are in the DONE state and this was a write of the
134 * dummy transaction, we move to NEED2.
135 *
136 * If we are in the DONE2 state and this was a write of the
137 * dummy transaction, we move to IDLE.
138 *
139 *
140 * Writing only one dummy transaction can get appended to
141 * one file space allocation. When this happens, the log recovery
142 * code replays the space allocation and a file could be truncated.
143 * This is why we have the NEED2 and DONE2 states before going idle.
144 */
145
146#define XLOG_STATE_COVER_IDLE 0
147#define XLOG_STATE_COVER_NEED 1
148#define XLOG_STATE_COVER_DONE 2
149#define XLOG_STATE_COVER_NEED2 3
150#define XLOG_STATE_COVER_DONE2 4
151
152#define XLOG_COVER_OPS 5
153
154/* Ticket reservation region accounting */
155#define XLOG_TIC_LEN_MAX 15
156
157/*
158 * Reservation region
159 * As would be stored in xfs_log_iovec but without the i_addr which
160 * we don't care about.
161 */
162typedef struct xlog_res {
163 uint r_len; /* region length :4 */
164 uint r_type; /* region's transaction type :4 */
165} xlog_res_t;
166
167typedef struct xlog_ticket {
168 struct list_head t_queue; /* reserve/write queue */
169 struct task_struct *t_task; /* task that owns this ticket */
170 xlog_tid_t t_tid; /* transaction identifier : 4 */
171 atomic_t t_ref; /* ticket reference count : 4 */
172 int t_curr_res; /* current reservation in bytes : 4 */
173 int t_unit_res; /* unit reservation in bytes : 4 */
174 char t_ocnt; /* original count : 1 */
175 char t_cnt; /* current count : 1 */
176 char t_clientid; /* who does this belong to; : 1 */
177 char t_flags; /* properties of reservation : 1 */
178
179 /* reservation array fields */
180 uint t_res_num; /* num in array : 4 */
181 uint t_res_num_ophdrs; /* num op hdrs : 4 */
182 uint t_res_arr_sum; /* array sum : 4 */
183 uint t_res_o_flow; /* sum overflow : 4 */
184 xlog_res_t t_res_arr[XLOG_TIC_LEN_MAX]; /* array of res : 8 * 15 */
185} xlog_ticket_t;
186
187/*
188 * - A log record header is 512 bytes. There is plenty of room to grow the
189 * xlog_rec_header_t into the reserved space.
190 * - ic_data follows, so a write to disk can start at the beginning of
191 * the iclog.
192 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
193 * - ic_next is the pointer to the next iclog in the ring.
194 * - ic_bp is a pointer to the buffer used to write this incore log to disk.
195 * - ic_log is a pointer back to the global log structure.
196 * - ic_callback is a linked list of callback function/argument pairs to be
197 * called after an iclog finishes writing.
198 * - ic_size is the full size of the header plus data.
199 * - ic_offset is the current number of bytes written to in this iclog.
200 * - ic_refcnt is bumped when someone is writing to the log.
201 * - ic_state is the state of the iclog.
202 *
203 * Because of cacheline contention on large machines, we need to separate
204 * various resources onto different cachelines. To start with, make the
205 * structure cacheline aligned. The following fields can be contended on
206 * by independent processes:
207 *
208 * - ic_callback_*
209 * - ic_refcnt
210 * - fields protected by the global l_icloglock
211 *
212 * so we need to ensure that these fields are located in separate cachelines.
213 * We'll put all the read-only and l_icloglock fields in the first cacheline,
214 * and move everything else out to subsequent cachelines.
215 */
216typedef struct xlog_in_core {
217 wait_queue_head_t ic_force_wait;
218 wait_queue_head_t ic_write_wait;
219 struct xlog_in_core *ic_next;
220 struct xlog_in_core *ic_prev;
221 struct xfs_buf *ic_bp;
222 struct xlog *ic_log;
223 int ic_size;
224 int ic_offset;
225 int ic_bwritecnt;
226 unsigned short ic_state;
227 char *ic_datap; /* pointer to iclog data */
228
229 /* Callback structures need their own cacheline */
230 spinlock_t ic_callback_lock ____cacheline_aligned_in_smp;
231 struct xfs_log_callback *ic_callback;
232 struct xfs_log_callback **ic_callback_tail;
233
234 /* reference counts need their own cacheline */
235 atomic_t ic_refcnt ____cacheline_aligned_in_smp;
236 xlog_in_core_2_t *ic_data;
237#define ic_header ic_data->hic_header
238} xlog_in_core_t;
239
240/*
241 * The CIL context is used to aggregate per-transaction details as well be
242 * passed to the iclog for checkpoint post-commit processing. After being
243 * passed to the iclog, another context needs to be allocated for tracking the
244 * next set of transactions to be aggregated into a checkpoint.
245 */
246struct xfs_cil;
247
248struct xfs_cil_ctx {
249 struct xfs_cil *cil;
250 xfs_lsn_t sequence; /* chkpt sequence # */
251 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */
252 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */
253 struct xlog_ticket *ticket; /* chkpt ticket */
254 int nvecs; /* number of regions */
255 int space_used; /* aggregate size of regions */
256 struct list_head busy_extents; /* busy extents in chkpt */
257 struct xfs_log_vec *lv_chain; /* logvecs being pushed */
258 struct xfs_log_callback log_cb; /* completion callback hook. */
259 struct list_head committing; /* ctx committing list */
260 struct work_struct discard_endio_work;
261};
262
263/*
264 * Committed Item List structure
265 *
266 * This structure is used to track log items that have been committed but not
267 * yet written into the log. It is used only when the delayed logging mount
268 * option is enabled.
269 *
270 * This structure tracks the list of committing checkpoint contexts so
271 * we can avoid the problem of having to hold out new transactions during a
272 * flush until we have a the commit record LSN of the checkpoint. We can
273 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
274 * sequence match and extract the commit LSN directly from there. If the
275 * checkpoint is still in the process of committing, we can block waiting for
276 * the commit LSN to be determined as well. This should make synchronous
277 * operations almost as efficient as the old logging methods.
278 */
279struct xfs_cil {
280 struct xlog *xc_log;
281 struct list_head xc_cil;
282 spinlock_t xc_cil_lock;
283
284 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp;
285 struct xfs_cil_ctx *xc_ctx;
286
287 spinlock_t xc_push_lock ____cacheline_aligned_in_smp;
288 xfs_lsn_t xc_push_seq;
289 struct list_head xc_committing;
290 wait_queue_head_t xc_commit_wait;
291 xfs_lsn_t xc_current_sequence;
292 struct work_struct xc_push_work;
293} ____cacheline_aligned_in_smp;
294
295/*
296 * The amount of log space we allow the CIL to aggregate is difficult to size.
297 * Whatever we choose, we have to make sure we can get a reservation for the
298 * log space effectively, that it is large enough to capture sufficient
299 * relogging to reduce log buffer IO significantly, but it is not too large for
300 * the log or induces too much latency when writing out through the iclogs. We
301 * track both space consumed and the number of vectors in the checkpoint
302 * context, so we need to decide which to use for limiting.
303 *
304 * Every log buffer we write out during a push needs a header reserved, which
305 * is at least one sector and more for v2 logs. Hence we need a reservation of
306 * at least 512 bytes per 32k of log space just for the LR headers. That means
307 * 16KB of reservation per megabyte of delayed logging space we will consume,
308 * plus various headers. The number of headers will vary based on the num of
309 * io vectors, so limiting on a specific number of vectors is going to result
310 * in transactions of varying size. IOWs, it is more consistent to track and
311 * limit space consumed in the log rather than by the number of objects being
312 * logged in order to prevent checkpoint ticket overruns.
313 *
314 * Further, use of static reservations through the log grant mechanism is
315 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
316 * grant) and a significant deadlock potential because regranting write space
317 * can block on log pushes. Hence if we have to regrant log space during a log
318 * push, we can deadlock.
319 *
320 * However, we can avoid this by use of a dynamic "reservation stealing"
321 * technique during transaction commit whereby unused reservation space in the
322 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
323 * space needed by the checkpoint transaction. This means that we never need to
324 * specifically reserve space for the CIL checkpoint transaction, nor do we
325 * need to regrant space once the checkpoint completes. This also means the
326 * checkpoint transaction ticket is specific to the checkpoint context, rather
327 * than the CIL itself.
328 *
329 * With dynamic reservations, we can effectively make up arbitrary limits for
330 * the checkpoint size so long as they don't violate any other size rules.
331 * Recovery imposes a rule that no transaction exceed half the log, so we are
332 * limited by that. Furthermore, the log transaction reservation subsystem
333 * tries to keep 25% of the log free, so we need to keep below that limit or we
334 * risk running out of free log space to start any new transactions.
335 *
336 * In order to keep background CIL push efficient, we will set a lower
337 * threshold at which background pushing is attempted without blocking current
338 * transaction commits. A separate, higher bound defines when CIL pushes are
339 * enforced to ensure we stay within our maximum checkpoint size bounds.
340 * threshold, yet give us plenty of space for aggregation on large logs.
341 */
342#define XLOG_CIL_SPACE_LIMIT(log) (log->l_logsize >> 3)
343
344/*
345 * ticket grant locks, queues and accounting have their own cachlines
346 * as these are quite hot and can be operated on concurrently.
347 */
348struct xlog_grant_head {
349 spinlock_t lock ____cacheline_aligned_in_smp;
350 struct list_head waiters;
351 atomic64_t grant;
352};
353
354/*
355 * The reservation head lsn is not made up of a cycle number and block number.
356 * Instead, it uses a cycle number and byte number. Logs don't expect to
357 * overflow 31 bits worth of byte offset, so using a byte number will mean
358 * that round off problems won't occur when releasing partial reservations.
359 */
360struct xlog {
361 /* The following fields don't need locking */
362 struct xfs_mount *l_mp; /* mount point */
363 struct xfs_ail *l_ailp; /* AIL log is working with */
364 struct xfs_cil *l_cilp; /* CIL log is working with */
365 struct xfs_buf *l_xbuf; /* extra buffer for log
366 * wrapping */
367 struct xfs_buftarg *l_targ; /* buftarg of log */
368 struct delayed_work l_work; /* background flush work */
369 uint l_flags;
370 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
371 struct list_head *l_buf_cancel_table;
372 int l_iclog_hsize; /* size of iclog header */
373 int l_iclog_heads; /* # of iclog header sectors */
374 uint l_sectBBsize; /* sector size in BBs (2^n) */
375 int l_iclog_size; /* size of log in bytes */
376 int l_iclog_size_log; /* log power size of log */
377 int l_iclog_bufs; /* number of iclog buffers */
378 xfs_daddr_t l_logBBstart; /* start block of log */
379 int l_logsize; /* size of log in bytes */
380 int l_logBBsize; /* size of log in BB chunks */
381
382 /* The following block of fields are changed while holding icloglock */
383 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
384 /* waiting for iclog flush */
385 int l_covered_state;/* state of "covering disk
386 * log entries" */
387 xlog_in_core_t *l_iclog; /* head log queue */
388 spinlock_t l_icloglock; /* grab to change iclog state */
389 int l_curr_cycle; /* Cycle number of log writes */
390 int l_prev_cycle; /* Cycle number before last
391 * block increment */
392 int l_curr_block; /* current logical log block */
393 int l_prev_block; /* previous logical log block */
394
395 /*
396 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and
397 * read without needing to hold specific locks. To avoid operations
398 * contending with other hot objects, place each of them on a separate
399 * cacheline.
400 */
401 /* lsn of last LR on disk */
402 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp;
403 /* lsn of 1st LR with unflushed * buffers */
404 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
405
406 struct xlog_grant_head l_reserve_head;
407 struct xlog_grant_head l_write_head;
408
409 struct xfs_kobj l_kobj;
410
411 /* The following field are used for debugging; need to hold icloglock */
412#ifdef DEBUG
413 void *l_iclog_bak[XLOG_MAX_ICLOGS];
414 /* log record crc error injection factor */
415 uint32_t l_badcrc_factor;
416#endif
417 /* log recovery lsn tracking (for buffer submission */
418 xfs_lsn_t l_recovery_lsn;
419};
420
421#define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
422 ((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE))
423
424#define XLOG_FORCED_SHUTDOWN(log) ((log)->l_flags & XLOG_IO_ERROR)
425
426/* common routines */
427extern int
428xlog_recover(
429 struct xlog *log);
430extern int
431xlog_recover_finish(
432 struct xlog *log);
433extern int
434xlog_recover_cancel(struct xlog *);
435
436extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
437 char *dp, int size);
438
439extern kmem_zone_t *xfs_log_ticket_zone;
440struct xlog_ticket *
441xlog_ticket_alloc(
442 struct xlog *log,
443 int unit_bytes,
444 int count,
445 char client,
446 bool permanent,
447 xfs_km_flags_t alloc_flags);
448
449
450static inline void
451xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes)
452{
453 *ptr += bytes;
454 *len -= bytes;
455 *off += bytes;
456}
457
458void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
459void xlog_print_trans(struct xfs_trans *);
460int
461xlog_write(
462 struct xlog *log,
463 struct xfs_log_vec *log_vector,
464 struct xlog_ticket *tic,
465 xfs_lsn_t *start_lsn,
466 struct xlog_in_core **commit_iclog,
467 uint flags);
468
469/*
470 * When we crack an atomic LSN, we sample it first so that the value will not
471 * change while we are cracking it into the component values. This means we
472 * will always get consistent component values to work from. This should always
473 * be used to sample and crack LSNs that are stored and updated in atomic
474 * variables.
475 */
476static inline void
477xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
478{
479 xfs_lsn_t val = atomic64_read(lsn);
480
481 *cycle = CYCLE_LSN(val);
482 *block = BLOCK_LSN(val);
483}
484
485/*
486 * Calculate and assign a value to an atomic LSN variable from component pieces.
487 */
488static inline void
489xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
490{
491 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
492}
493
494/*
495 * When we crack the grant head, we sample it first so that the value will not
496 * change while we are cracking it into the component values. This means we
497 * will always get consistent component values to work from.
498 */
499static inline void
500xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
501{
502 *cycle = val >> 32;
503 *space = val & 0xffffffff;
504}
505
506static inline void
507xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
508{
509 xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
510}
511
512static inline int64_t
513xlog_assign_grant_head_val(int cycle, int space)
514{
515 return ((int64_t)cycle << 32) | space;
516}
517
518static inline void
519xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
520{
521 atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
522}
523
524/*
525 * Committed Item List interfaces
526 */
527int xlog_cil_init(struct xlog *log);
528void xlog_cil_init_post_recovery(struct xlog *log);
529void xlog_cil_destroy(struct xlog *log);
530bool xlog_cil_empty(struct xlog *log);
531
532/*
533 * CIL force routines
534 */
535xfs_lsn_t
536xlog_cil_force_lsn(
537 struct xlog *log,
538 xfs_lsn_t sequence);
539
540static inline void
541xlog_cil_force(struct xlog *log)
542{
543 xlog_cil_force_lsn(log, log->l_cilp->xc_current_sequence);
544}
545
546/*
547 * Unmount record type is used as a pseudo transaction type for the ticket.
548 * It's value must be outside the range of XFS_TRANS_* values.
549 */
550#define XLOG_UNMOUNT_REC_TYPE (-1U)
551
552/*
553 * Wrapper function for waiting on a wait queue serialised against wakeups
554 * by a spinlock. This matches the semantics of all the wait queues used in the
555 * log code.
556 */
557static inline void xlog_wait(wait_queue_head_t *wq, spinlock_t *lock)
558{
559 DECLARE_WAITQUEUE(wait, current);
560
561 add_wait_queue_exclusive(wq, &wait);
562 __set_current_state(TASK_UNINTERRUPTIBLE);
563 spin_unlock(lock);
564 schedule();
565 remove_wait_queue(wq, &wait);
566}
567
568/*
569 * The LSN is valid so long as it is behind the current LSN. If it isn't, this
570 * means that the next log record that includes this metadata could have a
571 * smaller LSN. In turn, this means that the modification in the log would not
572 * replay.
573 */
574static inline bool
575xlog_valid_lsn(
576 struct xlog *log,
577 xfs_lsn_t lsn)
578{
579 int cur_cycle;
580 int cur_block;
581 bool valid = true;
582
583 /*
584 * First, sample the current lsn without locking to avoid added
585 * contention from metadata I/O. The current cycle and block are updated
586 * (in xlog_state_switch_iclogs()) and read here in a particular order
587 * to avoid false negatives (e.g., thinking the metadata LSN is valid
588 * when it is not).
589 *
590 * The current block is always rewound before the cycle is bumped in
591 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in
592 * a transiently forward state. Instead, we can see the LSN in a
593 * transiently behind state if we happen to race with a cycle wrap.
594 */
595 cur_cycle = READ_ONCE(log->l_curr_cycle);
596 smp_rmb();
597 cur_block = READ_ONCE(log->l_curr_block);
598
599 if ((CYCLE_LSN(lsn) > cur_cycle) ||
600 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
601 /*
602 * If the metadata LSN appears invalid, it's possible the check
603 * above raced with a wrap to the next log cycle. Grab the lock
604 * to check for sure.
605 */
606 spin_lock(&log->l_icloglock);
607 cur_cycle = log->l_curr_cycle;
608 cur_block = log->l_curr_block;
609 spin_unlock(&log->l_icloglock);
610
611 if ((CYCLE_LSN(lsn) > cur_cycle) ||
612 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
613 valid = false;
614 }
615
616 return valid;
617}
618
619#endif /* __XFS_LOG_PRIV_H__ */