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
9#include "xfs_extent_busy.h" /* for struct xfs_busy_extents */
10
11struct xfs_buf;
12struct xlog;
13struct xlog_ticket;
14struct xfs_mount;
15
16/*
17 * get client id from packed copy.
18 *
19 * this hack is here because the xlog_pack code copies four bytes
20 * of xlog_op_header containing the fields oh_clientid, oh_flags
21 * and oh_res2 into the packed copy.
22 *
23 * later on this four byte chunk is treated as an int and the
24 * client id is pulled out.
25 *
26 * this has endian issues, of course.
27 */
28static inline uint xlog_get_client_id(__be32 i)
29{
30 return be32_to_cpu(i) >> 24;
31}
32
33/*
34 * In core log state
35 */
36enum xlog_iclog_state {
37 XLOG_STATE_ACTIVE, /* Current IC log being written to */
38 XLOG_STATE_WANT_SYNC, /* Want to sync this iclog; no more writes */
39 XLOG_STATE_SYNCING, /* This IC log is syncing */
40 XLOG_STATE_DONE_SYNC, /* Done syncing to disk */
41 XLOG_STATE_CALLBACK, /* Callback functions now */
42 XLOG_STATE_DIRTY, /* Dirty IC log, not ready for ACTIVE status */
43};
44
45#define XLOG_STATE_STRINGS \
46 { XLOG_STATE_ACTIVE, "XLOG_STATE_ACTIVE" }, \
47 { XLOG_STATE_WANT_SYNC, "XLOG_STATE_WANT_SYNC" }, \
48 { XLOG_STATE_SYNCING, "XLOG_STATE_SYNCING" }, \
49 { XLOG_STATE_DONE_SYNC, "XLOG_STATE_DONE_SYNC" }, \
50 { XLOG_STATE_CALLBACK, "XLOG_STATE_CALLBACK" }, \
51 { XLOG_STATE_DIRTY, "XLOG_STATE_DIRTY" }
52
53/*
54 * In core log flags
55 */
56#define XLOG_ICL_NEED_FLUSH (1u << 0) /* iclog needs REQ_PREFLUSH */
57#define XLOG_ICL_NEED_FUA (1u << 1) /* iclog needs REQ_FUA */
58
59#define XLOG_ICL_STRINGS \
60 { XLOG_ICL_NEED_FLUSH, "XLOG_ICL_NEED_FLUSH" }, \
61 { XLOG_ICL_NEED_FUA, "XLOG_ICL_NEED_FUA" }
62
63
64/*
65 * Log ticket flags
66 */
67#define XLOG_TIC_PERM_RESERV (1u << 0) /* permanent reservation */
68
69#define XLOG_TIC_FLAGS \
70 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }
71
72/*
73 * Below are states for covering allocation transactions.
74 * By covering, we mean changing the h_tail_lsn in the last on-disk
75 * log write such that no allocation transactions will be re-done during
76 * recovery after a system crash. Recovery starts at the last on-disk
77 * log write.
78 *
79 * These states are used to insert dummy log entries to cover
80 * space allocation transactions which can undo non-transactional changes
81 * after a crash. Writes to a file with space
82 * already allocated do not result in any transactions. Allocations
83 * might include space beyond the EOF. So if we just push the EOF a
84 * little, the last transaction for the file could contain the wrong
85 * size. If there is no file system activity, after an allocation
86 * transaction, and the system crashes, the allocation transaction
87 * will get replayed and the file will be truncated. This could
88 * be hours/days/... after the allocation occurred.
89 *
90 * The fix for this is to do two dummy transactions when the
91 * system is idle. We need two dummy transaction because the h_tail_lsn
92 * in the log record header needs to point beyond the last possible
93 * non-dummy transaction. The first dummy changes the h_tail_lsn to
94 * the first transaction before the dummy. The second dummy causes
95 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn.
96 *
97 * These dummy transactions get committed when everything
98 * is idle (after there has been some activity).
99 *
100 * There are 5 states used to control this.
101 *
102 * IDLE -- no logging has been done on the file system or
103 * we are done covering previous transactions.
104 * NEED -- logging has occurred and we need a dummy transaction
105 * when the log becomes idle.
106 * DONE -- we were in the NEED state and have committed a dummy
107 * transaction.
108 * NEED2 -- we detected that a dummy transaction has gone to the
109 * on disk log with no other transactions.
110 * DONE2 -- we committed a dummy transaction when in the NEED2 state.
111 *
112 * There are two places where we switch states:
113 *
114 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
115 * We commit the dummy transaction and switch to DONE or DONE2,
116 * respectively. In all other states, we don't do anything.
117 *
118 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
119 *
120 * No matter what state we are in, if this isn't the dummy
121 * transaction going out, the next state is NEED.
122 * So, if we aren't in the DONE or DONE2 states, the next state
123 * is NEED. We can't be finishing a write of the dummy record
124 * unless it was committed and the state switched to DONE or DONE2.
125 *
126 * If we are in the DONE state and this was a write of the
127 * dummy transaction, we move to NEED2.
128 *
129 * If we are in the DONE2 state and this was a write of the
130 * dummy transaction, we move to IDLE.
131 *
132 *
133 * Writing only one dummy transaction can get appended to
134 * one file space allocation. When this happens, the log recovery
135 * code replays the space allocation and a file could be truncated.
136 * This is why we have the NEED2 and DONE2 states before going idle.
137 */
138
139#define XLOG_STATE_COVER_IDLE 0
140#define XLOG_STATE_COVER_NEED 1
141#define XLOG_STATE_COVER_DONE 2
142#define XLOG_STATE_COVER_NEED2 3
143#define XLOG_STATE_COVER_DONE2 4
144
145#define XLOG_COVER_OPS 5
146
147typedef struct xlog_ticket {
148 struct list_head t_queue; /* reserve/write queue */
149 struct task_struct *t_task; /* task that owns this ticket */
150 xlog_tid_t t_tid; /* transaction identifier */
151 atomic_t t_ref; /* ticket reference count */
152 int t_curr_res; /* current reservation */
153 int t_unit_res; /* unit reservation */
154 char t_ocnt; /* original unit count */
155 char t_cnt; /* current unit count */
156 uint8_t t_flags; /* properties of reservation */
157 int t_iclog_hdrs; /* iclog hdrs in t_curr_res */
158} xlog_ticket_t;
159
160/*
161 * - A log record header is 512 bytes. There is plenty of room to grow the
162 * xlog_rec_header_t into the reserved space.
163 * - ic_data follows, so a write to disk can start at the beginning of
164 * the iclog.
165 * - ic_forcewait is used to implement synchronous forcing of the iclog to disk.
166 * - ic_next is the pointer to the next iclog in the ring.
167 * - ic_log is a pointer back to the global log structure.
168 * - ic_size is the full size of the log buffer, minus the cycle headers.
169 * - ic_offset is the current number of bytes written to in this iclog.
170 * - ic_refcnt is bumped when someone is writing to the log.
171 * - ic_state is the state of the iclog.
172 *
173 * Because of cacheline contention on large machines, we need to separate
174 * various resources onto different cachelines. To start with, make the
175 * structure cacheline aligned. The following fields can be contended on
176 * by independent processes:
177 *
178 * - ic_callbacks
179 * - ic_refcnt
180 * - fields protected by the global l_icloglock
181 *
182 * so we need to ensure that these fields are located in separate cachelines.
183 * We'll put all the read-only and l_icloglock fields in the first cacheline,
184 * and move everything else out to subsequent cachelines.
185 */
186typedef struct xlog_in_core {
187 wait_queue_head_t ic_force_wait;
188 wait_queue_head_t ic_write_wait;
189 struct xlog_in_core *ic_next;
190 struct xlog_in_core *ic_prev;
191 struct xlog *ic_log;
192 u32 ic_size;
193 u32 ic_offset;
194 enum xlog_iclog_state ic_state;
195 unsigned int ic_flags;
196 void *ic_datap; /* pointer to iclog data */
197 struct list_head ic_callbacks;
198
199 /* reference counts need their own cacheline */
200 atomic_t ic_refcnt ____cacheline_aligned_in_smp;
201 xlog_in_core_2_t *ic_data;
202#define ic_header ic_data->hic_header
203#ifdef DEBUG
204 bool ic_fail_crc : 1;
205#endif
206 struct semaphore ic_sema;
207 struct work_struct ic_end_io_work;
208 struct bio ic_bio;
209 struct bio_vec ic_bvec[];
210} xlog_in_core_t;
211
212/*
213 * The CIL context is used to aggregate per-transaction details as well be
214 * passed to the iclog for checkpoint post-commit processing. After being
215 * passed to the iclog, another context needs to be allocated for tracking the
216 * next set of transactions to be aggregated into a checkpoint.
217 */
218struct xfs_cil;
219
220struct xfs_cil_ctx {
221 struct xfs_cil *cil;
222 xfs_csn_t sequence; /* chkpt sequence # */
223 xfs_lsn_t start_lsn; /* first LSN of chkpt commit */
224 xfs_lsn_t commit_lsn; /* chkpt commit record lsn */
225 struct xlog_in_core *commit_iclog;
226 struct xlog_ticket *ticket; /* chkpt ticket */
227 atomic_t space_used; /* aggregate size of regions */
228 struct xfs_busy_extents busy_extents;
229 struct list_head log_items; /* log items in chkpt */
230 struct list_head lv_chain; /* logvecs being pushed */
231 struct list_head iclog_entry;
232 struct list_head committing; /* ctx committing list */
233 struct work_struct push_work;
234 atomic_t order_id;
235
236 /*
237 * CPUs that could have added items to the percpu CIL data. Access is
238 * coordinated with xc_ctx_lock.
239 */
240 struct cpumask cil_pcpmask;
241};
242
243/*
244 * Per-cpu CIL tracking items
245 */
246struct xlog_cil_pcp {
247 int32_t space_used;
248 uint32_t space_reserved;
249 struct list_head busy_extents;
250 struct list_head log_items;
251};
252
253/*
254 * Committed Item List structure
255 *
256 * This structure is used to track log items that have been committed but not
257 * yet written into the log. It is used only when the delayed logging mount
258 * option is enabled.
259 *
260 * This structure tracks the list of committing checkpoint contexts so
261 * we can avoid the problem of having to hold out new transactions during a
262 * flush until we have a the commit record LSN of the checkpoint. We can
263 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
264 * sequence match and extract the commit LSN directly from there. If the
265 * checkpoint is still in the process of committing, we can block waiting for
266 * the commit LSN to be determined as well. This should make synchronous
267 * operations almost as efficient as the old logging methods.
268 */
269struct xfs_cil {
270 struct xlog *xc_log;
271 unsigned long xc_flags;
272 atomic_t xc_iclog_hdrs;
273 struct workqueue_struct *xc_push_wq;
274
275 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp;
276 struct xfs_cil_ctx *xc_ctx;
277
278 spinlock_t xc_push_lock ____cacheline_aligned_in_smp;
279 xfs_csn_t xc_push_seq;
280 bool xc_push_commit_stable;
281 struct list_head xc_committing;
282 wait_queue_head_t xc_commit_wait;
283 wait_queue_head_t xc_start_wait;
284 xfs_csn_t xc_current_sequence;
285 wait_queue_head_t xc_push_wait; /* background push throttle */
286
287 void __percpu *xc_pcp; /* percpu CIL structures */
288} ____cacheline_aligned_in_smp;
289
290/* xc_flags bit values */
291#define XLOG_CIL_EMPTY 1
292#define XLOG_CIL_PCP_SPACE 2
293
294/*
295 * The amount of log space we allow the CIL to aggregate is difficult to size.
296 * Whatever we choose, we have to make sure we can get a reservation for the
297 * log space effectively, that it is large enough to capture sufficient
298 * relogging to reduce log buffer IO significantly, but it is not too large for
299 * the log or induces too much latency when writing out through the iclogs. We
300 * track both space consumed and the number of vectors in the checkpoint
301 * context, so we need to decide which to use for limiting.
302 *
303 * Every log buffer we write out during a push needs a header reserved, which
304 * is at least one sector and more for v2 logs. Hence we need a reservation of
305 * at least 512 bytes per 32k of log space just for the LR headers. That means
306 * 16KB of reservation per megabyte of delayed logging space we will consume,
307 * plus various headers. The number of headers will vary based on the num of
308 * io vectors, so limiting on a specific number of vectors is going to result
309 * in transactions of varying size. IOWs, it is more consistent to track and
310 * limit space consumed in the log rather than by the number of objects being
311 * logged in order to prevent checkpoint ticket overruns.
312 *
313 * Further, use of static reservations through the log grant mechanism is
314 * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
315 * grant) and a significant deadlock potential because regranting write space
316 * can block on log pushes. Hence if we have to regrant log space during a log
317 * push, we can deadlock.
318 *
319 * However, we can avoid this by use of a dynamic "reservation stealing"
320 * technique during transaction commit whereby unused reservation space in the
321 * transaction ticket is transferred to the CIL ctx commit ticket to cover the
322 * space needed by the checkpoint transaction. This means that we never need to
323 * specifically reserve space for the CIL checkpoint transaction, nor do we
324 * need to regrant space once the checkpoint completes. This also means the
325 * checkpoint transaction ticket is specific to the checkpoint context, rather
326 * than the CIL itself.
327 *
328 * With dynamic reservations, we can effectively make up arbitrary limits for
329 * the checkpoint size so long as they don't violate any other size rules.
330 * Recovery imposes a rule that no transaction exceed half the log, so we are
331 * limited by that. Furthermore, the log transaction reservation subsystem
332 * tries to keep 25% of the log free, so we need to keep below that limit or we
333 * risk running out of free log space to start any new transactions.
334 *
335 * In order to keep background CIL push efficient, we only need to ensure the
336 * CIL is large enough to maintain sufficient in-memory relogging to avoid
337 * repeated physical writes of frequently modified metadata. If we allow the CIL
338 * to grow to a substantial fraction of the log, then we may be pinning hundreds
339 * of megabytes of metadata in memory until the CIL flushes. This can cause
340 * issues when we are running low on memory - pinned memory cannot be reclaimed,
341 * and the CIL consumes a lot of memory. Hence we need to set an upper physical
342 * size limit for the CIL that limits the maximum amount of memory pinned by the
343 * CIL but does not limit performance by reducing relogging efficiency
344 * significantly.
345 *
346 * As such, the CIL push threshold ends up being the smaller of two thresholds:
347 * - a threshold large enough that it allows CIL to be pushed and progress to be
348 * made without excessive blocking of incoming transaction commits. This is
349 * defined to be 12.5% of the log space - half the 25% push threshold of the
350 * AIL.
351 * - small enough that it doesn't pin excessive amounts of memory but maintains
352 * close to peak relogging efficiency. This is defined to be 16x the iclog
353 * buffer window (32MB) as measurements have shown this to be roughly the
354 * point of diminishing performance increases under highly concurrent
355 * modification workloads.
356 *
357 * To prevent the CIL from overflowing upper commit size bounds, we introduce a
358 * new threshold at which we block committing transactions until the background
359 * CIL commit commences and switches to a new context. While this is not a hard
360 * limit, it forces the process committing a transaction to the CIL to block and
361 * yeild the CPU, giving the CIL push work a chance to be scheduled and start
362 * work. This prevents a process running lots of transactions from overfilling
363 * the CIL because it is not yielding the CPU. We set the blocking limit at
364 * twice the background push space threshold so we keep in line with the AIL
365 * push thresholds.
366 *
367 * Note: this is not a -hard- limit as blocking is applied after the transaction
368 * is inserted into the CIL and the push has been triggered. It is largely a
369 * throttling mechanism that allows the CIL push to be scheduled and run. A hard
370 * limit will be difficult to implement without introducing global serialisation
371 * in the CIL commit fast path, and it's not at all clear that we actually need
372 * such hard limits given the ~7 years we've run without a hard limit before
373 * finding the first situation where a checkpoint size overflow actually
374 * occurred. Hence the simple throttle, and an ASSERT check to tell us that
375 * we've overrun the max size.
376 */
377#define XLOG_CIL_SPACE_LIMIT(log) \
378 min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4)
379
380#define XLOG_CIL_BLOCKING_SPACE_LIMIT(log) \
381 (XLOG_CIL_SPACE_LIMIT(log) * 2)
382
383/*
384 * ticket grant locks, queues and accounting have their own cachlines
385 * as these are quite hot and can be operated on concurrently.
386 */
387struct xlog_grant_head {
388 spinlock_t lock ____cacheline_aligned_in_smp;
389 struct list_head waiters;
390 atomic64_t grant;
391};
392
393/*
394 * The reservation head lsn is not made up of a cycle number and block number.
395 * Instead, it uses a cycle number and byte number. Logs don't expect to
396 * overflow 31 bits worth of byte offset, so using a byte number will mean
397 * that round off problems won't occur when releasing partial reservations.
398 */
399struct xlog {
400 /* The following fields don't need locking */
401 struct xfs_mount *l_mp; /* mount point */
402 struct xfs_ail *l_ailp; /* AIL log is working with */
403 struct xfs_cil *l_cilp; /* CIL log is working with */
404 struct xfs_buftarg *l_targ; /* buftarg of log */
405 struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */
406 struct delayed_work l_work; /* background flush work */
407 long l_opstate; /* operational state */
408 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
409 struct list_head *l_buf_cancel_table;
410 struct list_head r_dfops; /* recovered log intent items */
411 int l_iclog_hsize; /* size of iclog header */
412 int l_iclog_heads; /* # of iclog header sectors */
413 uint l_sectBBsize; /* sector size in BBs (2^n) */
414 int l_iclog_size; /* size of log in bytes */
415 int l_iclog_bufs; /* number of iclog buffers */
416 xfs_daddr_t l_logBBstart; /* start block of log */
417 int l_logsize; /* size of log in bytes */
418 int l_logBBsize; /* size of log in BB chunks */
419
420 /* The following block of fields are changed while holding icloglock */
421 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
422 /* waiting for iclog flush */
423 int l_covered_state;/* state of "covering disk
424 * log entries" */
425 xlog_in_core_t *l_iclog; /* head log queue */
426 spinlock_t l_icloglock; /* grab to change iclog state */
427 int l_curr_cycle; /* Cycle number of log writes */
428 int l_prev_cycle; /* Cycle number before last
429 * block increment */
430 int l_curr_block; /* current logical log block */
431 int l_prev_block; /* previous logical log block */
432
433 /*
434 * l_tail_lsn is atomic so it can be set and read without needing to
435 * hold specific locks. To avoid operations contending with other hot
436 * objects, it on a separate cacheline.
437 */
438 /* lsn of 1st LR with unflushed * buffers */
439 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
440
441 struct xlog_grant_head l_reserve_head;
442 struct xlog_grant_head l_write_head;
443 uint64_t l_tail_space;
444
445 struct xfs_kobj l_kobj;
446
447 /* log recovery lsn tracking (for buffer submission */
448 xfs_lsn_t l_recovery_lsn;
449
450 uint32_t l_iclog_roundoff;/* padding roundoff */
451};
452
453/*
454 * Bits for operational state
455 */
456#define XLOG_ACTIVE_RECOVERY 0 /* in the middle of recovery */
457#define XLOG_RECOVERY_NEEDED 1 /* log was recovered */
458#define XLOG_IO_ERROR 2 /* log hit an I/O error, and being
459 shutdown */
460#define XLOG_TAIL_WARN 3 /* log tail verify warning issued */
461#define XLOG_SHUTDOWN_STARTED 4 /* xlog_force_shutdown() exclusion */
462
463static inline bool
464xlog_recovery_needed(struct xlog *log)
465{
466 return test_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
467}
468
469static inline bool
470xlog_in_recovery(struct xlog *log)
471{
472 return test_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
473}
474
475static inline bool
476xlog_is_shutdown(struct xlog *log)
477{
478 return test_bit(XLOG_IO_ERROR, &log->l_opstate);
479}
480
481/*
482 * Wait until the xlog_force_shutdown() has marked the log as shut down
483 * so xlog_is_shutdown() will always return true.
484 */
485static inline void
486xlog_shutdown_wait(
487 struct xlog *log)
488{
489 wait_var_event(&log->l_opstate, xlog_is_shutdown(log));
490}
491
492/* common routines */
493extern int
494xlog_recover(
495 struct xlog *log);
496extern int
497xlog_recover_finish(
498 struct xlog *log);
499extern void
500xlog_recover_cancel(struct xlog *);
501
502extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
503 char *dp, int size);
504
505extern struct kmem_cache *xfs_log_ticket_cache;
506struct xlog_ticket *xlog_ticket_alloc(struct xlog *log, int unit_bytes,
507 int count, bool permanent);
508
509void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
510void xlog_print_trans(struct xfs_trans *);
511int xlog_write(struct xlog *log, struct xfs_cil_ctx *ctx,
512 struct list_head *lv_chain, struct xlog_ticket *tic,
513 uint32_t len);
514void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket);
515void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket);
516
517void xlog_state_switch_iclogs(struct xlog *log, struct xlog_in_core *iclog,
518 int eventual_size);
519int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog,
520 struct xlog_ticket *ticket);
521
522/*
523 * When we crack an atomic LSN, we sample it first so that the value will not
524 * change while we are cracking it into the component values. This means we
525 * will always get consistent component values to work from. This should always
526 * be used to sample and crack LSNs that are stored and updated in atomic
527 * variables.
528 */
529static inline void
530xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
531{
532 xfs_lsn_t val = atomic64_read(lsn);
533
534 *cycle = CYCLE_LSN(val);
535 *block = BLOCK_LSN(val);
536}
537
538/*
539 * Calculate and assign a value to an atomic LSN variable from component pieces.
540 */
541static inline void
542xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
543{
544 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
545}
546
547/*
548 * Committed Item List interfaces
549 */
550int xlog_cil_init(struct xlog *log);
551void xlog_cil_init_post_recovery(struct xlog *log);
552void xlog_cil_destroy(struct xlog *log);
553bool xlog_cil_empty(struct xlog *log);
554void xlog_cil_commit(struct xlog *log, struct xfs_trans *tp,
555 xfs_csn_t *commit_seq, bool regrant);
556void xlog_cil_set_ctx_write_state(struct xfs_cil_ctx *ctx,
557 struct xlog_in_core *iclog);
558
559
560/*
561 * CIL force routines
562 */
563void xlog_cil_flush(struct xlog *log);
564xfs_lsn_t xlog_cil_force_seq(struct xlog *log, xfs_csn_t sequence);
565
566static inline void
567xlog_cil_force(struct xlog *log)
568{
569 xlog_cil_force_seq(log, log->l_cilp->xc_current_sequence);
570}
571
572/*
573 * Wrapper function for waiting on a wait queue serialised against wakeups
574 * by a spinlock. This matches the semantics of all the wait queues used in the
575 * log code.
576 */
577static inline void
578xlog_wait(
579 struct wait_queue_head *wq,
580 struct spinlock *lock)
581 __releases(lock)
582{
583 DECLARE_WAITQUEUE(wait, current);
584
585 add_wait_queue_exclusive(wq, &wait);
586 __set_current_state(TASK_UNINTERRUPTIBLE);
587 spin_unlock(lock);
588 schedule();
589 remove_wait_queue(wq, &wait);
590}
591
592int xlog_wait_on_iclog(struct xlog_in_core *iclog)
593 __releases(iclog->ic_log->l_icloglock);
594
595/* Calculate the distance between two LSNs in bytes */
596static inline uint64_t
597xlog_lsn_sub(
598 struct xlog *log,
599 xfs_lsn_t high,
600 xfs_lsn_t low)
601{
602 uint32_t hi_cycle = CYCLE_LSN(high);
603 uint32_t hi_block = BLOCK_LSN(high);
604 uint32_t lo_cycle = CYCLE_LSN(low);
605 uint32_t lo_block = BLOCK_LSN(low);
606
607 if (hi_cycle == lo_cycle)
608 return BBTOB(hi_block - lo_block);
609 ASSERT((hi_cycle == lo_cycle + 1) || xlog_is_shutdown(log));
610 return (uint64_t)log->l_logsize - BBTOB(lo_block - hi_block);
611}
612
613void xlog_grant_return_space(struct xlog *log, xfs_lsn_t old_head,
614 xfs_lsn_t new_head);
615
616/*
617 * The LSN is valid so long as it is behind the current LSN. If it isn't, this
618 * means that the next log record that includes this metadata could have a
619 * smaller LSN. In turn, this means that the modification in the log would not
620 * replay.
621 */
622static inline bool
623xlog_valid_lsn(
624 struct xlog *log,
625 xfs_lsn_t lsn)
626{
627 int cur_cycle;
628 int cur_block;
629 bool valid = true;
630
631 /*
632 * First, sample the current lsn without locking to avoid added
633 * contention from metadata I/O. The current cycle and block are updated
634 * (in xlog_state_switch_iclogs()) and read here in a particular order
635 * to avoid false negatives (e.g., thinking the metadata LSN is valid
636 * when it is not).
637 *
638 * The current block is always rewound before the cycle is bumped in
639 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in
640 * a transiently forward state. Instead, we can see the LSN in a
641 * transiently behind state if we happen to race with a cycle wrap.
642 */
643 cur_cycle = READ_ONCE(log->l_curr_cycle);
644 smp_rmb();
645 cur_block = READ_ONCE(log->l_curr_block);
646
647 if ((CYCLE_LSN(lsn) > cur_cycle) ||
648 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
649 /*
650 * If the metadata LSN appears invalid, it's possible the check
651 * above raced with a wrap to the next log cycle. Grab the lock
652 * to check for sure.
653 */
654 spin_lock(&log->l_icloglock);
655 cur_cycle = log->l_curr_cycle;
656 cur_block = log->l_curr_block;
657 spin_unlock(&log->l_icloglock);
658
659 if ((CYCLE_LSN(lsn) > cur_cycle) ||
660 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
661 valid = false;
662 }
663
664 return valid;
665}
666
667/*
668 * Log vector and shadow buffers can be large, so we need to use kvmalloc() here
669 * to ensure success. Unfortunately, kvmalloc() only allows GFP_KERNEL contexts
670 * to fall back to vmalloc, so we can't actually do anything useful with gfp
671 * flags to control the kmalloc() behaviour within kvmalloc(). Hence kmalloc()
672 * will do direct reclaim and compaction in the slow path, both of which are
673 * horrendously expensive. We just want kmalloc to fail fast and fall back to
674 * vmalloc if it can't get something straight away from the free lists or
675 * buddy allocator. Hence we have to open code kvmalloc outselves here.
676 *
677 * This assumes that the caller uses memalloc_nofs_save task context here, so
678 * despite the use of GFP_KERNEL here, we are going to be doing GFP_NOFS
679 * allocations. This is actually the only way to make vmalloc() do GFP_NOFS
680 * allocations, so lets just all pretend this is a GFP_KERNEL context
681 * operation....
682 */
683static inline void *
684xlog_kvmalloc(
685 size_t buf_size)
686{
687 gfp_t flags = GFP_KERNEL;
688 void *p;
689
690 flags &= ~__GFP_DIRECT_RECLAIM;
691 flags |= __GFP_NOWARN | __GFP_NORETRY;
692 do {
693 p = kmalloc(buf_size, flags);
694 if (!p)
695 p = vmalloc(buf_size);
696 } while (!p);
697
698 return p;
699}
700
701#endif /* __XFS_LOG_PRIV_H__ */