2 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
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.
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.
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
18 #ifndef __XFS_LOG_PRIV_H__
19 #define __XFS_LOG_PRIV_H__
25 struct xfs_log_callback;
28 * Flags for log structure
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
34 #define XLOG_TAIL_WARN 0x10 /* log tail verify warning issued */
37 * get client id from packed copy.
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.
43 * later on this four byte chunk is treated as an int and the
44 * client id is pulled out.
46 * this has endian issues, of course.
48 static inline uint xlog_get_client_id(__be32 i)
50 return be32_to_cpu(i) >> 24;
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_ALL 0x7FFF /* All possible valid flags */
66 #define XLOG_STATE_NOTUSED 0x8000 /* This IC log not being used */
71 #define XLOG_TIC_INITED 0x1 /* has been initialized */
72 #define XLOG_TIC_PERM_RESERV 0x2 /* permanent reservation */
74 #define XLOG_TIC_FLAGS \
75 { XLOG_TIC_INITED, "XLOG_TIC_INITED" }, \
76 { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" }
79 * Below are states for covering allocation transactions.
80 * By covering, we mean changing the h_tail_lsn in the last on-disk
81 * log write such that no allocation transactions will be re-done during
82 * recovery after a system crash. Recovery starts at the last on-disk
85 * These states are used to insert dummy log entries to cover
86 * space allocation transactions which can undo non-transactional changes
87 * after a crash. Writes to a file with space
88 * already allocated do not result in any transactions. Allocations
89 * might include space beyond the EOF. So if we just push the EOF a
90 * little, the last transaction for the file could contain the wrong
91 * size. If there is no file system activity, after an allocation
92 * transaction, and the system crashes, the allocation transaction
93 * will get replayed and the file will be truncated. This could
94 * be hours/days/... after the allocation occurred.
96 * The fix for this is to do two dummy transactions when the
97 * system is idle. We need two dummy transaction because the h_tail_lsn
98 * in the log record header needs to point beyond the last possible
99 * non-dummy transaction. The first dummy changes the h_tail_lsn to
100 * the first transaction before the dummy. The second dummy causes
101 * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn.
103 * These dummy transactions get committed when everything
104 * is idle (after there has been some activity).
106 * There are 5 states used to control this.
108 * IDLE -- no logging has been done on the file system or
109 * we are done covering previous transactions.
110 * NEED -- logging has occurred and we need a dummy transaction
111 * when the log becomes idle.
112 * DONE -- we were in the NEED state and have committed a dummy
114 * NEED2 -- we detected that a dummy transaction has gone to the
115 * on disk log with no other transactions.
116 * DONE2 -- we committed a dummy transaction when in the NEED2 state.
118 * There are two places where we switch states:
120 * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
121 * We commit the dummy transaction and switch to DONE or DONE2,
122 * respectively. In all other states, we don't do anything.
124 * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
126 * No matter what state we are in, if this isn't the dummy
127 * transaction going out, the next state is NEED.
128 * So, if we aren't in the DONE or DONE2 states, the next state
129 * is NEED. We can't be finishing a write of the dummy record
130 * unless it was committed and the state switched to DONE or DONE2.
132 * If we are in the DONE state and this was a write of the
133 * dummy transaction, we move to NEED2.
135 * If we are in the DONE2 state and this was a write of the
136 * dummy transaction, we move to IDLE.
139 * Writing only one dummy transaction can get appended to
140 * one file space allocation. When this happens, the log recovery
141 * code replays the space allocation and a file could be truncated.
142 * This is why we have the NEED2 and DONE2 states before going idle.
145 #define XLOG_STATE_COVER_IDLE 0
146 #define XLOG_STATE_COVER_NEED 1
147 #define XLOG_STATE_COVER_DONE 2
148 #define XLOG_STATE_COVER_NEED2 3
149 #define XLOG_STATE_COVER_DONE2 4
151 #define XLOG_COVER_OPS 5
153 /* Ticket reservation region accounting */
154 #define XLOG_TIC_LEN_MAX 15
158 * As would be stored in xfs_log_iovec but without the i_addr which
159 * we don't care about.
161 typedef struct xlog_res {
162 uint r_len; /* region length :4 */
163 uint r_type; /* region's transaction type :4 */
166 typedef struct xlog_ticket {
167 struct list_head t_queue; /* reserve/write queue */
168 struct task_struct *t_task; /* task that owns this ticket */
169 xlog_tid_t t_tid; /* transaction identifier : 4 */
170 atomic_t t_ref; /* ticket reference count : 4 */
171 int t_curr_res; /* current reservation in bytes : 4 */
172 int t_unit_res; /* unit reservation in bytes : 4 */
173 char t_ocnt; /* original count : 1 */
174 char t_cnt; /* current count : 1 */
175 char t_clientid; /* who does this belong to; : 1 */
176 char t_flags; /* properties of reservation : 1 */
177 uint t_trans_type; /* transaction type : 4 */
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 */
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
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.
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:
210 * - fields protected by the global l_icloglock
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.
216 typedef 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;
226 unsigned short ic_state;
227 char *ic_datap; /* pointer to iclog data */
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;
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
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.
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 */
263 * Committed Item List structure
265 * This structure is used to track log items that have been committed but not
266 * yet written into the log. It is used only when the delayed logging mount
269 * This structure tracks the list of committing checkpoint contexts so
270 * we can avoid the problem of having to hold out new transactions during a
271 * flush until we have a the commit record LSN of the checkpoint. We can
272 * traverse the list of committing contexts in xlog_cil_push_lsn() to find a
273 * sequence match and extract the commit LSN directly from there. If the
274 * checkpoint is still in the process of committing, we can block waiting for
275 * the commit LSN to be determined as well. This should make synchronous
276 * operations almost as efficient as the old logging methods.
280 struct list_head xc_cil;
281 spinlock_t xc_cil_lock;
283 struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp;
284 struct xfs_cil_ctx *xc_ctx;
286 spinlock_t xc_push_lock ____cacheline_aligned_in_smp;
287 xfs_lsn_t xc_push_seq;
288 struct list_head xc_committing;
289 wait_queue_head_t xc_commit_wait;
290 xfs_lsn_t xc_current_sequence;
291 struct work_struct xc_push_work;
292 } ____cacheline_aligned_in_smp;
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.
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.
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.
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.
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.
335 * In order to keep background CIL push efficient, we will set a lower
336 * threshold at which background pushing is attempted without blocking current
337 * transaction commits. A separate, higher bound defines when CIL pushes are
338 * enforced to ensure we stay within our maximum checkpoint size bounds.
339 * threshold, yet give us plenty of space for aggregation on large logs.
341 #define XLOG_CIL_SPACE_LIMIT(log) (log->l_logsize >> 3)
344 * ticket grant locks, queues and accounting have their own cachlines
345 * as these are quite hot and can be operated on concurrently.
347 struct xlog_grant_head {
348 spinlock_t lock ____cacheline_aligned_in_smp;
349 struct list_head waiters;
354 * The reservation head lsn is not made up of a cycle number and block number.
355 * Instead, it uses a cycle number and byte number. Logs don't expect to
356 * overflow 31 bits worth of byte offset, so using a byte number will mean
357 * that round off problems won't occur when releasing partial reservations.
360 /* The following fields don't need locking */
361 struct xfs_mount *l_mp; /* mount point */
362 struct xfs_ail *l_ailp; /* AIL log is working with */
363 struct xfs_cil *l_cilp; /* CIL log is working with */
364 struct xfs_buf *l_xbuf; /* extra buffer for log
366 struct xfs_buftarg *l_targ; /* buftarg of log */
367 struct delayed_work l_work; /* background flush work */
369 uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */
370 struct list_head *l_buf_cancel_table;
371 int l_iclog_hsize; /* size of iclog header */
372 int l_iclog_heads; /* # of iclog header sectors */
373 uint l_sectBBsize; /* sector size in BBs (2^n) */
374 int l_iclog_size; /* size of log in bytes */
375 int l_iclog_size_log; /* log power size of log */
376 int l_iclog_bufs; /* number of iclog buffers */
377 xfs_daddr_t l_logBBstart; /* start block of log */
378 int l_logsize; /* size of log in bytes */
379 int l_logBBsize; /* size of log in BB chunks */
381 /* The following block of fields are changed while holding icloglock */
382 wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp;
383 /* waiting for iclog flush */
384 int l_covered_state;/* state of "covering disk
386 xlog_in_core_t *l_iclog; /* head log queue */
387 spinlock_t l_icloglock; /* grab to change iclog state */
388 int l_curr_cycle; /* Cycle number of log writes */
389 int l_prev_cycle; /* Cycle number before last
391 int l_curr_block; /* current logical log block */
392 int l_prev_block; /* previous logical log block */
395 * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and
396 * read without needing to hold specific locks. To avoid operations
397 * contending with other hot objects, place each of them on a separate
400 /* lsn of last LR on disk */
401 atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp;
402 /* lsn of 1st LR with unflushed * buffers */
403 atomic64_t l_tail_lsn ____cacheline_aligned_in_smp;
405 struct xlog_grant_head l_reserve_head;
406 struct xlog_grant_head l_write_head;
408 struct xfs_kobj l_kobj;
410 /* The following field are used for debugging; need to hold icloglock */
412 void *l_iclog_bak[XLOG_MAX_ICLOGS];
417 #define XLOG_BUF_CANCEL_BUCKET(log, blkno) \
418 ((log)->l_buf_cancel_table + ((__uint64_t)blkno % XLOG_BC_TABLE_SIZE))
420 #define XLOG_FORCED_SHUTDOWN(log) ((log)->l_flags & XLOG_IO_ERROR)
422 /* common routines */
430 xlog_recover_cancel(struct xlog *);
432 extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead,
435 extern kmem_zone_t *xfs_log_ticket_zone;
443 xfs_km_flags_t alloc_flags);
447 xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes)
454 void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket);
458 struct xfs_log_vec *log_vector,
459 struct xlog_ticket *tic,
460 xfs_lsn_t *start_lsn,
461 struct xlog_in_core **commit_iclog,
465 * When we crack an atomic LSN, we sample it first so that the value will not
466 * change while we are cracking it into the component values. This means we
467 * will always get consistent component values to work from. This should always
468 * be used to sample and crack LSNs that are stored and updated in atomic
472 xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block)
474 xfs_lsn_t val = atomic64_read(lsn);
476 *cycle = CYCLE_LSN(val);
477 *block = BLOCK_LSN(val);
481 * Calculate and assign a value to an atomic LSN variable from component pieces.
484 xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block)
486 atomic64_set(lsn, xlog_assign_lsn(cycle, block));
490 * When we crack the grant head, we sample it first so that the value will not
491 * change while we are cracking it into the component values. This means we
492 * will always get consistent component values to work from.
495 xlog_crack_grant_head_val(int64_t val, int *cycle, int *space)
498 *space = val & 0xffffffff;
502 xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space)
504 xlog_crack_grant_head_val(atomic64_read(head), cycle, space);
507 static inline int64_t
508 xlog_assign_grant_head_val(int cycle, int space)
510 return ((int64_t)cycle << 32) | space;
514 xlog_assign_grant_head(atomic64_t *head, int cycle, int space)
516 atomic64_set(head, xlog_assign_grant_head_val(cycle, space));
520 * Committed Item List interfaces
522 int xlog_cil_init(struct xlog *log);
523 void xlog_cil_init_post_recovery(struct xlog *log);
524 void xlog_cil_destroy(struct xlog *log);
525 bool xlog_cil_empty(struct xlog *log);
536 xlog_cil_force(struct xlog *log)
538 xlog_cil_force_lsn(log, log->l_cilp->xc_current_sequence);
542 * Unmount record type is used as a pseudo transaction type for the ticket.
543 * It's value must be outside the range of XFS_TRANS_* values.
545 #define XLOG_UNMOUNT_REC_TYPE (-1U)
548 * Wrapper function for waiting on a wait queue serialised against wakeups
549 * by a spinlock. This matches the semantics of all the wait queues used in the
552 static inline void xlog_wait(wait_queue_head_t *wq, spinlock_t *lock)
554 DECLARE_WAITQUEUE(wait, current);
556 add_wait_queue_exclusive(wq, &wait);
557 __set_current_state(TASK_UNINTERRUPTIBLE);
560 remove_wait_queue(wq, &wait);
564 * The LSN is valid so long as it is behind the current LSN. If it isn't, this
565 * means that the next log record that includes this metadata could have a
566 * smaller LSN. In turn, this means that the modification in the log would not
579 * First, sample the current lsn without locking to avoid added
580 * contention from metadata I/O. The current cycle and block are updated
581 * (in xlog_state_switch_iclogs()) and read here in a particular order
582 * to avoid false negatives (e.g., thinking the metadata LSN is valid
585 * The current block is always rewound before the cycle is bumped in
586 * xlog_state_switch_iclogs() to ensure the current LSN is never seen in
587 * a transiently forward state. Instead, we can see the LSN in a
588 * transiently behind state if we happen to race with a cycle wrap.
590 cur_cycle = ACCESS_ONCE(log->l_curr_cycle);
592 cur_block = ACCESS_ONCE(log->l_curr_block);
594 if ((CYCLE_LSN(lsn) > cur_cycle) ||
595 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) {
597 * If the metadata LSN appears invalid, it's possible the check
598 * above raced with a wrap to the next log cycle. Grab the lock
601 spin_lock(&log->l_icloglock);
602 cur_cycle = log->l_curr_cycle;
603 cur_block = log->l_curr_block;
604 spin_unlock(&log->l_icloglock);
606 if ((CYCLE_LSN(lsn) > cur_cycle) ||
607 (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block))
614 #endif /* __XFS_LOG_PRIV_H__ */