/fs/xfs/xfs_log.c
C | 3753 lines | 2261 code | 440 blank | 1052 comment | 420 complexity | 0d6ab3726121eb7735de7103e0722919 MD5 | raw file
Possible License(s): LGPL-2.0, AGPL-1.0, GPL-2.0
1/*
2 * Copyright (c) 2000-2005 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#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_bit.h"
22#include "xfs_log.h"
23#include "xfs_inum.h"
24#include "xfs_trans.h"
25#include "xfs_sb.h"
26#include "xfs_ag.h"
27#include "xfs_mount.h"
28#include "xfs_error.h"
29#include "xfs_log_priv.h"
30#include "xfs_buf_item.h"
31#include "xfs_bmap_btree.h"
32#include "xfs_alloc_btree.h"
33#include "xfs_ialloc_btree.h"
34#include "xfs_log_recover.h"
35#include "xfs_trans_priv.h"
36#include "xfs_dinode.h"
37#include "xfs_inode.h"
38#include "xfs_rw.h"
39#include "xfs_trace.h"
40
41kmem_zone_t *xfs_log_ticket_zone;
42
43/* Local miscellaneous function prototypes */
44STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
45 xlog_in_core_t **, xfs_lsn_t *);
46STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
47 xfs_buftarg_t *log_target,
48 xfs_daddr_t blk_offset,
49 int num_bblks);
50STATIC int xlog_space_left(struct log *log, atomic64_t *head);
51STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
52STATIC void xlog_dealloc_log(xlog_t *log);
53
54/* local state machine functions */
55STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
56STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
57STATIC int xlog_state_get_iclog_space(xlog_t *log,
58 int len,
59 xlog_in_core_t **iclog,
60 xlog_ticket_t *ticket,
61 int *continued_write,
62 int *logoffsetp);
63STATIC int xlog_state_release_iclog(xlog_t *log,
64 xlog_in_core_t *iclog);
65STATIC void xlog_state_switch_iclogs(xlog_t *log,
66 xlog_in_core_t *iclog,
67 int eventual_size);
68STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
69
70/* local functions to manipulate grant head */
71STATIC int xlog_grant_log_space(xlog_t *log,
72 xlog_ticket_t *xtic);
73STATIC void xlog_grant_push_ail(struct log *log,
74 int need_bytes);
75STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
76 xlog_ticket_t *ticket);
77STATIC int xlog_regrant_write_log_space(xlog_t *log,
78 xlog_ticket_t *ticket);
79STATIC void xlog_ungrant_log_space(xlog_t *log,
80 xlog_ticket_t *ticket);
81
82#if defined(DEBUG)
83STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
84STATIC void xlog_verify_grant_tail(struct log *log);
85STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
86 int count, boolean_t syncing);
87STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
88 xfs_lsn_t tail_lsn);
89#else
90#define xlog_verify_dest_ptr(a,b)
91#define xlog_verify_grant_tail(a)
92#define xlog_verify_iclog(a,b,c,d)
93#define xlog_verify_tail_lsn(a,b,c)
94#endif
95
96STATIC int xlog_iclogs_empty(xlog_t *log);
97
98static void
99xlog_grant_sub_space(
100 struct log *log,
101 atomic64_t *head,
102 int bytes)
103{
104 int64_t head_val = atomic64_read(head);
105 int64_t new, old;
106
107 do {
108 int cycle, space;
109
110 xlog_crack_grant_head_val(head_val, &cycle, &space);
111
112 space -= bytes;
113 if (space < 0) {
114 space += log->l_logsize;
115 cycle--;
116 }
117
118 old = head_val;
119 new = xlog_assign_grant_head_val(cycle, space);
120 head_val = atomic64_cmpxchg(head, old, new);
121 } while (head_val != old);
122}
123
124static void
125xlog_grant_add_space(
126 struct log *log,
127 atomic64_t *head,
128 int bytes)
129{
130 int64_t head_val = atomic64_read(head);
131 int64_t new, old;
132
133 do {
134 int tmp;
135 int cycle, space;
136
137 xlog_crack_grant_head_val(head_val, &cycle, &space);
138
139 tmp = log->l_logsize - space;
140 if (tmp > bytes)
141 space += bytes;
142 else {
143 space = bytes - tmp;
144 cycle++;
145 }
146
147 old = head_val;
148 new = xlog_assign_grant_head_val(cycle, space);
149 head_val = atomic64_cmpxchg(head, old, new);
150 } while (head_val != old);
151}
152
153static void
154xlog_tic_reset_res(xlog_ticket_t *tic)
155{
156 tic->t_res_num = 0;
157 tic->t_res_arr_sum = 0;
158 tic->t_res_num_ophdrs = 0;
159}
160
161static void
162xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
163{
164 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
165 /* add to overflow and start again */
166 tic->t_res_o_flow += tic->t_res_arr_sum;
167 tic->t_res_num = 0;
168 tic->t_res_arr_sum = 0;
169 }
170
171 tic->t_res_arr[tic->t_res_num].r_len = len;
172 tic->t_res_arr[tic->t_res_num].r_type = type;
173 tic->t_res_arr_sum += len;
174 tic->t_res_num++;
175}
176
177/*
178 * NOTES:
179 *
180 * 1. currblock field gets updated at startup and after in-core logs
181 * marked as with WANT_SYNC.
182 */
183
184/*
185 * This routine is called when a user of a log manager ticket is done with
186 * the reservation. If the ticket was ever used, then a commit record for
187 * the associated transaction is written out as a log operation header with
188 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
189 * a given ticket. If the ticket was one with a permanent reservation, then
190 * a few operations are done differently. Permanent reservation tickets by
191 * default don't release the reservation. They just commit the current
192 * transaction with the belief that the reservation is still needed. A flag
193 * must be passed in before permanent reservations are actually released.
194 * When these type of tickets are not released, they need to be set into
195 * the inited state again. By doing this, a start record will be written
196 * out when the next write occurs.
197 */
198xfs_lsn_t
199xfs_log_done(
200 struct xfs_mount *mp,
201 struct xlog_ticket *ticket,
202 struct xlog_in_core **iclog,
203 uint flags)
204{
205 struct log *log = mp->m_log;
206 xfs_lsn_t lsn = 0;
207
208 if (XLOG_FORCED_SHUTDOWN(log) ||
209 /*
210 * If nothing was ever written, don't write out commit record.
211 * If we get an error, just continue and give back the log ticket.
212 */
213 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
214 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
215 lsn = (xfs_lsn_t) -1;
216 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
217 flags |= XFS_LOG_REL_PERM_RESERV;
218 }
219 }
220
221
222 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
223 (flags & XFS_LOG_REL_PERM_RESERV)) {
224 trace_xfs_log_done_nonperm(log, ticket);
225
226 /*
227 * Release ticket if not permanent reservation or a specific
228 * request has been made to release a permanent reservation.
229 */
230 xlog_ungrant_log_space(log, ticket);
231 xfs_log_ticket_put(ticket);
232 } else {
233 trace_xfs_log_done_perm(log, ticket);
234
235 xlog_regrant_reserve_log_space(log, ticket);
236 /* If this ticket was a permanent reservation and we aren't
237 * trying to release it, reset the inited flags; so next time
238 * we write, a start record will be written out.
239 */
240 ticket->t_flags |= XLOG_TIC_INITED;
241 }
242
243 return lsn;
244}
245
246/*
247 * Attaches a new iclog I/O completion callback routine during
248 * transaction commit. If the log is in error state, a non-zero
249 * return code is handed back and the caller is responsible for
250 * executing the callback at an appropriate time.
251 */
252int
253xfs_log_notify(
254 struct xfs_mount *mp,
255 struct xlog_in_core *iclog,
256 xfs_log_callback_t *cb)
257{
258 int abortflg;
259
260 spin_lock(&iclog->ic_callback_lock);
261 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
262 if (!abortflg) {
263 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
264 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
265 cb->cb_next = NULL;
266 *(iclog->ic_callback_tail) = cb;
267 iclog->ic_callback_tail = &(cb->cb_next);
268 }
269 spin_unlock(&iclog->ic_callback_lock);
270 return abortflg;
271}
272
273int
274xfs_log_release_iclog(
275 struct xfs_mount *mp,
276 struct xlog_in_core *iclog)
277{
278 if (xlog_state_release_iclog(mp->m_log, iclog)) {
279 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
280 return EIO;
281 }
282
283 return 0;
284}
285
286/*
287 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
288 * to the reservation.
289 * 2. Potentially, push buffers at tail of log to disk.
290 *
291 * Each reservation is going to reserve extra space for a log record header.
292 * When writes happen to the on-disk log, we don't subtract the length of the
293 * log record header from any reservation. By wasting space in each
294 * reservation, we prevent over allocation problems.
295 */
296int
297xfs_log_reserve(
298 struct xfs_mount *mp,
299 int unit_bytes,
300 int cnt,
301 struct xlog_ticket **ticket,
302 __uint8_t client,
303 uint flags,
304 uint t_type)
305{
306 struct log *log = mp->m_log;
307 struct xlog_ticket *internal_ticket;
308 int retval = 0;
309
310 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
311
312 if (XLOG_FORCED_SHUTDOWN(log))
313 return XFS_ERROR(EIO);
314
315 XFS_STATS_INC(xs_try_logspace);
316
317
318 if (*ticket != NULL) {
319 ASSERT(flags & XFS_LOG_PERM_RESERV);
320 internal_ticket = *ticket;
321
322 /*
323 * this is a new transaction on the ticket, so we need to
324 * change the transaction ID so that the next transaction has a
325 * different TID in the log. Just add one to the existing tid
326 * so that we can see chains of rolling transactions in the log
327 * easily.
328 */
329 internal_ticket->t_tid++;
330
331 trace_xfs_log_reserve(log, internal_ticket);
332
333 xlog_grant_push_ail(log, internal_ticket->t_unit_res);
334 retval = xlog_regrant_write_log_space(log, internal_ticket);
335 } else {
336 /* may sleep if need to allocate more tickets */
337 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
338 client, flags,
339 KM_SLEEP|KM_MAYFAIL);
340 if (!internal_ticket)
341 return XFS_ERROR(ENOMEM);
342 internal_ticket->t_trans_type = t_type;
343 *ticket = internal_ticket;
344
345 trace_xfs_log_reserve(log, internal_ticket);
346
347 xlog_grant_push_ail(log,
348 (internal_ticket->t_unit_res *
349 internal_ticket->t_cnt));
350 retval = xlog_grant_log_space(log, internal_ticket);
351 }
352
353 return retval;
354} /* xfs_log_reserve */
355
356
357/*
358 * Mount a log filesystem
359 *
360 * mp - ubiquitous xfs mount point structure
361 * log_target - buftarg of on-disk log device
362 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
363 * num_bblocks - Number of BBSIZE blocks in on-disk log
364 *
365 * Return error or zero.
366 */
367int
368xfs_log_mount(
369 xfs_mount_t *mp,
370 xfs_buftarg_t *log_target,
371 xfs_daddr_t blk_offset,
372 int num_bblks)
373{
374 int error;
375
376 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
377 xfs_notice(mp, "Mounting Filesystem");
378 else {
379 xfs_notice(mp,
380"Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
381 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
382 }
383
384 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
385 if (IS_ERR(mp->m_log)) {
386 error = -PTR_ERR(mp->m_log);
387 goto out;
388 }
389
390 /*
391 * Initialize the AIL now we have a log.
392 */
393 error = xfs_trans_ail_init(mp);
394 if (error) {
395 xfs_warn(mp, "AIL initialisation failed: error %d", error);
396 goto out_free_log;
397 }
398 mp->m_log->l_ailp = mp->m_ail;
399
400 /*
401 * skip log recovery on a norecovery mount. pretend it all
402 * just worked.
403 */
404 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
405 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
406
407 if (readonly)
408 mp->m_flags &= ~XFS_MOUNT_RDONLY;
409
410 error = xlog_recover(mp->m_log);
411
412 if (readonly)
413 mp->m_flags |= XFS_MOUNT_RDONLY;
414 if (error) {
415 xfs_warn(mp, "log mount/recovery failed: error %d",
416 error);
417 goto out_destroy_ail;
418 }
419 }
420
421 /* Normal transactions can now occur */
422 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
423
424 /*
425 * Now the log has been fully initialised and we know were our
426 * space grant counters are, we can initialise the permanent ticket
427 * needed for delayed logging to work.
428 */
429 xlog_cil_init_post_recovery(mp->m_log);
430
431 return 0;
432
433out_destroy_ail:
434 xfs_trans_ail_destroy(mp);
435out_free_log:
436 xlog_dealloc_log(mp->m_log);
437out:
438 return error;
439}
440
441/*
442 * Finish the recovery of the file system. This is separate from
443 * the xfs_log_mount() call, because it depends on the code in
444 * xfs_mountfs() to read in the root and real-time bitmap inodes
445 * between calling xfs_log_mount() and here.
446 *
447 * mp - ubiquitous xfs mount point structure
448 */
449int
450xfs_log_mount_finish(xfs_mount_t *mp)
451{
452 int error;
453
454 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
455 error = xlog_recover_finish(mp->m_log);
456 else {
457 error = 0;
458 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
459 }
460
461 return error;
462}
463
464/*
465 * Final log writes as part of unmount.
466 *
467 * Mark the filesystem clean as unmount happens. Note that during relocation
468 * this routine needs to be executed as part of source-bag while the
469 * deallocation must not be done until source-end.
470 */
471
472/*
473 * Unmount record used to have a string "Unmount filesystem--" in the
474 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
475 * We just write the magic number now since that particular field isn't
476 * currently architecture converted and "nUmount" is a bit foo.
477 * As far as I know, there weren't any dependencies on the old behaviour.
478 */
479
480int
481xfs_log_unmount_write(xfs_mount_t *mp)
482{
483 xlog_t *log = mp->m_log;
484 xlog_in_core_t *iclog;
485#ifdef DEBUG
486 xlog_in_core_t *first_iclog;
487#endif
488 xlog_ticket_t *tic = NULL;
489 xfs_lsn_t lsn;
490 int error;
491
492 /*
493 * Don't write out unmount record on read-only mounts.
494 * Or, if we are doing a forced umount (typically because of IO errors).
495 */
496 if (mp->m_flags & XFS_MOUNT_RDONLY)
497 return 0;
498
499 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
500 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
501
502#ifdef DEBUG
503 first_iclog = iclog = log->l_iclog;
504 do {
505 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
506 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
507 ASSERT(iclog->ic_offset == 0);
508 }
509 iclog = iclog->ic_next;
510 } while (iclog != first_iclog);
511#endif
512 if (! (XLOG_FORCED_SHUTDOWN(log))) {
513 error = xfs_log_reserve(mp, 600, 1, &tic,
514 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
515 if (!error) {
516 /* the data section must be 32 bit size aligned */
517 struct {
518 __uint16_t magic;
519 __uint16_t pad1;
520 __uint32_t pad2; /* may as well make it 64 bits */
521 } magic = {
522 .magic = XLOG_UNMOUNT_TYPE,
523 };
524 struct xfs_log_iovec reg = {
525 .i_addr = &magic,
526 .i_len = sizeof(magic),
527 .i_type = XLOG_REG_TYPE_UNMOUNT,
528 };
529 struct xfs_log_vec vec = {
530 .lv_niovecs = 1,
531 .lv_iovecp = ®,
532 };
533
534 /* remove inited flag */
535 tic->t_flags = 0;
536 error = xlog_write(log, &vec, tic, &lsn,
537 NULL, XLOG_UNMOUNT_TRANS);
538 /*
539 * At this point, we're umounting anyway,
540 * so there's no point in transitioning log state
541 * to IOERROR. Just continue...
542 */
543 }
544
545 if (error)
546 xfs_alert(mp, "%s: unmount record failed", __func__);
547
548
549 spin_lock(&log->l_icloglock);
550 iclog = log->l_iclog;
551 atomic_inc(&iclog->ic_refcnt);
552 xlog_state_want_sync(log, iclog);
553 spin_unlock(&log->l_icloglock);
554 error = xlog_state_release_iclog(log, iclog);
555
556 spin_lock(&log->l_icloglock);
557 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
558 iclog->ic_state == XLOG_STATE_DIRTY)) {
559 if (!XLOG_FORCED_SHUTDOWN(log)) {
560 xlog_wait(&iclog->ic_force_wait,
561 &log->l_icloglock);
562 } else {
563 spin_unlock(&log->l_icloglock);
564 }
565 } else {
566 spin_unlock(&log->l_icloglock);
567 }
568 if (tic) {
569 trace_xfs_log_umount_write(log, tic);
570 xlog_ungrant_log_space(log, tic);
571 xfs_log_ticket_put(tic);
572 }
573 } else {
574 /*
575 * We're already in forced_shutdown mode, couldn't
576 * even attempt to write out the unmount transaction.
577 *
578 * Go through the motions of sync'ing and releasing
579 * the iclog, even though no I/O will actually happen,
580 * we need to wait for other log I/Os that may already
581 * be in progress. Do this as a separate section of
582 * code so we'll know if we ever get stuck here that
583 * we're in this odd situation of trying to unmount
584 * a file system that went into forced_shutdown as
585 * the result of an unmount..
586 */
587 spin_lock(&log->l_icloglock);
588 iclog = log->l_iclog;
589 atomic_inc(&iclog->ic_refcnt);
590
591 xlog_state_want_sync(log, iclog);
592 spin_unlock(&log->l_icloglock);
593 error = xlog_state_release_iclog(log, iclog);
594
595 spin_lock(&log->l_icloglock);
596
597 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
598 || iclog->ic_state == XLOG_STATE_DIRTY
599 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
600
601 xlog_wait(&iclog->ic_force_wait,
602 &log->l_icloglock);
603 } else {
604 spin_unlock(&log->l_icloglock);
605 }
606 }
607
608 return error;
609} /* xfs_log_unmount_write */
610
611/*
612 * Deallocate log structures for unmount/relocation.
613 *
614 * We need to stop the aild from running before we destroy
615 * and deallocate the log as the aild references the log.
616 */
617void
618xfs_log_unmount(xfs_mount_t *mp)
619{
620 xfs_trans_ail_destroy(mp);
621 xlog_dealloc_log(mp->m_log);
622}
623
624void
625xfs_log_item_init(
626 struct xfs_mount *mp,
627 struct xfs_log_item *item,
628 int type,
629 struct xfs_item_ops *ops)
630{
631 item->li_mountp = mp;
632 item->li_ailp = mp->m_ail;
633 item->li_type = type;
634 item->li_ops = ops;
635 item->li_lv = NULL;
636
637 INIT_LIST_HEAD(&item->li_ail);
638 INIT_LIST_HEAD(&item->li_cil);
639}
640
641/*
642 * Write region vectors to log. The write happens using the space reservation
643 * of the ticket (tic). It is not a requirement that all writes for a given
644 * transaction occur with one call to xfs_log_write(). However, it is important
645 * to note that the transaction reservation code makes an assumption about the
646 * number of log headers a transaction requires that may be violated if you
647 * don't pass all the transaction vectors in one call....
648 */
649int
650xfs_log_write(
651 struct xfs_mount *mp,
652 struct xfs_log_iovec reg[],
653 int nentries,
654 struct xlog_ticket *tic,
655 xfs_lsn_t *start_lsn)
656{
657 struct log *log = mp->m_log;
658 int error;
659 struct xfs_log_vec vec = {
660 .lv_niovecs = nentries,
661 .lv_iovecp = reg,
662 };
663
664 if (XLOG_FORCED_SHUTDOWN(log))
665 return XFS_ERROR(EIO);
666
667 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
668 if (error)
669 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
670 return error;
671}
672
673void
674xfs_log_move_tail(xfs_mount_t *mp,
675 xfs_lsn_t tail_lsn)
676{
677 xlog_ticket_t *tic;
678 xlog_t *log = mp->m_log;
679 int need_bytes, free_bytes;
680
681 if (XLOG_FORCED_SHUTDOWN(log))
682 return;
683
684 if (tail_lsn == 0)
685 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
686
687 /* tail_lsn == 1 implies that we weren't passed a valid value. */
688 if (tail_lsn != 1)
689 atomic64_set(&log->l_tail_lsn, tail_lsn);
690
691 if (!list_empty_careful(&log->l_writeq)) {
692#ifdef DEBUG
693 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
694 panic("Recovery problem");
695#endif
696 spin_lock(&log->l_grant_write_lock);
697 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
698 list_for_each_entry(tic, &log->l_writeq, t_queue) {
699 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
700
701 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
702 break;
703 tail_lsn = 0;
704 free_bytes -= tic->t_unit_res;
705 trace_xfs_log_regrant_write_wake_up(log, tic);
706 wake_up(&tic->t_wait);
707 }
708 spin_unlock(&log->l_grant_write_lock);
709 }
710
711 if (!list_empty_careful(&log->l_reserveq)) {
712#ifdef DEBUG
713 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
714 panic("Recovery problem");
715#endif
716 spin_lock(&log->l_grant_reserve_lock);
717 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
718 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
719 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
720 need_bytes = tic->t_unit_res*tic->t_cnt;
721 else
722 need_bytes = tic->t_unit_res;
723 if (free_bytes < need_bytes && tail_lsn != 1)
724 break;
725 tail_lsn = 0;
726 free_bytes -= need_bytes;
727 trace_xfs_log_grant_wake_up(log, tic);
728 wake_up(&tic->t_wait);
729 }
730 spin_unlock(&log->l_grant_reserve_lock);
731 }
732}
733
734/*
735 * Determine if we have a transaction that has gone to disk
736 * that needs to be covered. To begin the transition to the idle state
737 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
738 * If we are then in a state where covering is needed, the caller is informed
739 * that dummy transactions are required to move the log into the idle state.
740 *
741 * Because this is called as part of the sync process, we should also indicate
742 * that dummy transactions should be issued in anything but the covered or
743 * idle states. This ensures that the log tail is accurately reflected in
744 * the log at the end of the sync, hence if a crash occurrs avoids replay
745 * of transactions where the metadata is already on disk.
746 */
747int
748xfs_log_need_covered(xfs_mount_t *mp)
749{
750 int needed = 0;
751 xlog_t *log = mp->m_log;
752
753 if (!xfs_fs_writable(mp))
754 return 0;
755
756 spin_lock(&log->l_icloglock);
757 switch (log->l_covered_state) {
758 case XLOG_STATE_COVER_DONE:
759 case XLOG_STATE_COVER_DONE2:
760 case XLOG_STATE_COVER_IDLE:
761 break;
762 case XLOG_STATE_COVER_NEED:
763 case XLOG_STATE_COVER_NEED2:
764 if (!xfs_ail_min_lsn(log->l_ailp) &&
765 xlog_iclogs_empty(log)) {
766 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
767 log->l_covered_state = XLOG_STATE_COVER_DONE;
768 else
769 log->l_covered_state = XLOG_STATE_COVER_DONE2;
770 }
771 /* FALLTHRU */
772 default:
773 needed = 1;
774 break;
775 }
776 spin_unlock(&log->l_icloglock);
777 return needed;
778}
779
780/******************************************************************************
781 *
782 * local routines
783 *
784 ******************************************************************************
785 */
786
787/* xfs_trans_tail_ail returns 0 when there is nothing in the list.
788 * The log manager must keep track of the last LR which was committed
789 * to disk. The lsn of this LR will become the new tail_lsn whenever
790 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
791 * the situation where stuff could be written into the log but nothing
792 * was ever in the AIL when asked. Eventually, we panic since the
793 * tail hits the head.
794 *
795 * We may be holding the log iclog lock upon entering this routine.
796 */
797xfs_lsn_t
798xlog_assign_tail_lsn(
799 struct xfs_mount *mp)
800{
801 xfs_lsn_t tail_lsn;
802 struct log *log = mp->m_log;
803
804 tail_lsn = xfs_ail_min_lsn(mp->m_ail);
805 if (!tail_lsn)
806 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
807
808 atomic64_set(&log->l_tail_lsn, tail_lsn);
809 return tail_lsn;
810}
811
812/*
813 * Return the space in the log between the tail and the head. The head
814 * is passed in the cycle/bytes formal parms. In the special case where
815 * the reserve head has wrapped passed the tail, this calculation is no
816 * longer valid. In this case, just return 0 which means there is no space
817 * in the log. This works for all places where this function is called
818 * with the reserve head. Of course, if the write head were to ever
819 * wrap the tail, we should blow up. Rather than catch this case here,
820 * we depend on other ASSERTions in other parts of the code. XXXmiken
821 *
822 * This code also handles the case where the reservation head is behind
823 * the tail. The details of this case are described below, but the end
824 * result is that we return the size of the log as the amount of space left.
825 */
826STATIC int
827xlog_space_left(
828 struct log *log,
829 atomic64_t *head)
830{
831 int free_bytes;
832 int tail_bytes;
833 int tail_cycle;
834 int head_cycle;
835 int head_bytes;
836
837 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
838 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
839 tail_bytes = BBTOB(tail_bytes);
840 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
841 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
842 else if (tail_cycle + 1 < head_cycle)
843 return 0;
844 else if (tail_cycle < head_cycle) {
845 ASSERT(tail_cycle == (head_cycle - 1));
846 free_bytes = tail_bytes - head_bytes;
847 } else {
848 /*
849 * The reservation head is behind the tail.
850 * In this case we just want to return the size of the
851 * log as the amount of space left.
852 */
853 xfs_alert(log->l_mp,
854 "xlog_space_left: head behind tail\n"
855 " tail_cycle = %d, tail_bytes = %d\n"
856 " GH cycle = %d, GH bytes = %d",
857 tail_cycle, tail_bytes, head_cycle, head_bytes);
858 ASSERT(0);
859 free_bytes = log->l_logsize;
860 }
861 return free_bytes;
862}
863
864
865/*
866 * Log function which is called when an io completes.
867 *
868 * The log manager needs its own routine, in order to control what
869 * happens with the buffer after the write completes.
870 */
871void
872xlog_iodone(xfs_buf_t *bp)
873{
874 xlog_in_core_t *iclog = bp->b_fspriv;
875 xlog_t *l = iclog->ic_log;
876 int aborted = 0;
877
878 /*
879 * Race to shutdown the filesystem if we see an error.
880 */
881 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
882 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
883 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
884 XFS_BUF_STALE(bp);
885 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
886 /*
887 * This flag will be propagated to the trans-committed
888 * callback routines to let them know that the log-commit
889 * didn't succeed.
890 */
891 aborted = XFS_LI_ABORTED;
892 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
893 aborted = XFS_LI_ABORTED;
894 }
895
896 /* log I/O is always issued ASYNC */
897 ASSERT(XFS_BUF_ISASYNC(bp));
898 xlog_state_done_syncing(iclog, aborted);
899 /*
900 * do not reference the buffer (bp) here as we could race
901 * with it being freed after writing the unmount record to the
902 * log.
903 */
904
905} /* xlog_iodone */
906
907/*
908 * Return size of each in-core log record buffer.
909 *
910 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
911 *
912 * If the filesystem blocksize is too large, we may need to choose a
913 * larger size since the directory code currently logs entire blocks.
914 */
915
916STATIC void
917xlog_get_iclog_buffer_size(xfs_mount_t *mp,
918 xlog_t *log)
919{
920 int size;
921 int xhdrs;
922
923 if (mp->m_logbufs <= 0)
924 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
925 else
926 log->l_iclog_bufs = mp->m_logbufs;
927
928 /*
929 * Buffer size passed in from mount system call.
930 */
931 if (mp->m_logbsize > 0) {
932 size = log->l_iclog_size = mp->m_logbsize;
933 log->l_iclog_size_log = 0;
934 while (size != 1) {
935 log->l_iclog_size_log++;
936 size >>= 1;
937 }
938
939 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
940 /* # headers = size / 32k
941 * one header holds cycles from 32k of data
942 */
943
944 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
945 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
946 xhdrs++;
947 log->l_iclog_hsize = xhdrs << BBSHIFT;
948 log->l_iclog_heads = xhdrs;
949 } else {
950 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
951 log->l_iclog_hsize = BBSIZE;
952 log->l_iclog_heads = 1;
953 }
954 goto done;
955 }
956
957 /* All machines use 32kB buffers by default. */
958 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
959 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
960
961 /* the default log size is 16k or 32k which is one header sector */
962 log->l_iclog_hsize = BBSIZE;
963 log->l_iclog_heads = 1;
964
965done:
966 /* are we being asked to make the sizes selected above visible? */
967 if (mp->m_logbufs == 0)
968 mp->m_logbufs = log->l_iclog_bufs;
969 if (mp->m_logbsize == 0)
970 mp->m_logbsize = log->l_iclog_size;
971} /* xlog_get_iclog_buffer_size */
972
973
974/*
975 * This routine initializes some of the log structure for a given mount point.
976 * Its primary purpose is to fill in enough, so recovery can occur. However,
977 * some other stuff may be filled in too.
978 */
979STATIC xlog_t *
980xlog_alloc_log(xfs_mount_t *mp,
981 xfs_buftarg_t *log_target,
982 xfs_daddr_t blk_offset,
983 int num_bblks)
984{
985 xlog_t *log;
986 xlog_rec_header_t *head;
987 xlog_in_core_t **iclogp;
988 xlog_in_core_t *iclog, *prev_iclog=NULL;
989 xfs_buf_t *bp;
990 int i;
991 int error = ENOMEM;
992 uint log2_size = 0;
993
994 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
995 if (!log) {
996 xfs_warn(mp, "Log allocation failed: No memory!");
997 goto out;
998 }
999
1000 log->l_mp = mp;
1001 log->l_targ = log_target;
1002 log->l_logsize = BBTOB(num_bblks);
1003 log->l_logBBstart = blk_offset;
1004 log->l_logBBsize = num_bblks;
1005 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1006 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1007
1008 log->l_prev_block = -1;
1009 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1010 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1011 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1012 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1013 xlog_assign_grant_head(&log->l_grant_reserve_head, 1, 0);
1014 xlog_assign_grant_head(&log->l_grant_write_head, 1, 0);
1015 INIT_LIST_HEAD(&log->l_reserveq);
1016 INIT_LIST_HEAD(&log->l_writeq);
1017 spin_lock_init(&log->l_grant_reserve_lock);
1018 spin_lock_init(&log->l_grant_write_lock);
1019
1020 error = EFSCORRUPTED;
1021 if (xfs_sb_version_hassector(&mp->m_sb)) {
1022 log2_size = mp->m_sb.sb_logsectlog;
1023 if (log2_size < BBSHIFT) {
1024 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1025 log2_size, BBSHIFT);
1026 goto out_free_log;
1027 }
1028
1029 log2_size -= BBSHIFT;
1030 if (log2_size > mp->m_sectbb_log) {
1031 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1032 log2_size, mp->m_sectbb_log);
1033 goto out_free_log;
1034 }
1035
1036 /* for larger sector sizes, must have v2 or external log */
1037 if (log2_size && log->l_logBBstart > 0 &&
1038 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1039 xfs_warn(mp,
1040 "log sector size (0x%x) invalid for configuration.",
1041 log2_size);
1042 goto out_free_log;
1043 }
1044 }
1045 log->l_sectBBsize = 1 << log2_size;
1046
1047 xlog_get_iclog_buffer_size(mp, log);
1048
1049 error = ENOMEM;
1050 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1051 if (!bp)
1052 goto out_free_log;
1053 bp->b_iodone = xlog_iodone;
1054 ASSERT(xfs_buf_islocked(bp));
1055 log->l_xbuf = bp;
1056
1057 spin_lock_init(&log->l_icloglock);
1058 init_waitqueue_head(&log->l_flush_wait);
1059
1060 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1061 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1062
1063 iclogp = &log->l_iclog;
1064 /*
1065 * The amount of memory to allocate for the iclog structure is
1066 * rather funky due to the way the structure is defined. It is
1067 * done this way so that we can use different sizes for machines
1068 * with different amounts of memory. See the definition of
1069 * xlog_in_core_t in xfs_log_priv.h for details.
1070 */
1071 ASSERT(log->l_iclog_size >= 4096);
1072 for (i=0; i < log->l_iclog_bufs; i++) {
1073 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1074 if (!*iclogp)
1075 goto out_free_iclog;
1076
1077 iclog = *iclogp;
1078 iclog->ic_prev = prev_iclog;
1079 prev_iclog = iclog;
1080
1081 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1082 log->l_iclog_size, 0);
1083 if (!bp)
1084 goto out_free_iclog;
1085
1086 bp->b_iodone = xlog_iodone;
1087 iclog->ic_bp = bp;
1088 iclog->ic_data = bp->b_addr;
1089#ifdef DEBUG
1090 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1091#endif
1092 head = &iclog->ic_header;
1093 memset(head, 0, sizeof(xlog_rec_header_t));
1094 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1095 head->h_version = cpu_to_be32(
1096 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1097 head->h_size = cpu_to_be32(log->l_iclog_size);
1098 /* new fields */
1099 head->h_fmt = cpu_to_be32(XLOG_FMT);
1100 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1101
1102 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1103 iclog->ic_state = XLOG_STATE_ACTIVE;
1104 iclog->ic_log = log;
1105 atomic_set(&iclog->ic_refcnt, 0);
1106 spin_lock_init(&iclog->ic_callback_lock);
1107 iclog->ic_callback_tail = &(iclog->ic_callback);
1108 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1109
1110 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1111 init_waitqueue_head(&iclog->ic_force_wait);
1112 init_waitqueue_head(&iclog->ic_write_wait);
1113
1114 iclogp = &iclog->ic_next;
1115 }
1116 *iclogp = log->l_iclog; /* complete ring */
1117 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1118
1119 error = xlog_cil_init(log);
1120 if (error)
1121 goto out_free_iclog;
1122 return log;
1123
1124out_free_iclog:
1125 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1126 prev_iclog = iclog->ic_next;
1127 if (iclog->ic_bp)
1128 xfs_buf_free(iclog->ic_bp);
1129 kmem_free(iclog);
1130 }
1131 spinlock_destroy(&log->l_icloglock);
1132 xfs_buf_free(log->l_xbuf);
1133out_free_log:
1134 kmem_free(log);
1135out:
1136 return ERR_PTR(-error);
1137} /* xlog_alloc_log */
1138
1139
1140/*
1141 * Write out the commit record of a transaction associated with the given
1142 * ticket. Return the lsn of the commit record.
1143 */
1144STATIC int
1145xlog_commit_record(
1146 struct log *log,
1147 struct xlog_ticket *ticket,
1148 struct xlog_in_core **iclog,
1149 xfs_lsn_t *commitlsnp)
1150{
1151 struct xfs_mount *mp = log->l_mp;
1152 int error;
1153 struct xfs_log_iovec reg = {
1154 .i_addr = NULL,
1155 .i_len = 0,
1156 .i_type = XLOG_REG_TYPE_COMMIT,
1157 };
1158 struct xfs_log_vec vec = {
1159 .lv_niovecs = 1,
1160 .lv_iovecp = ®,
1161 };
1162
1163 ASSERT_ALWAYS(iclog);
1164 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1165 XLOG_COMMIT_TRANS);
1166 if (error)
1167 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1168 return error;
1169}
1170
1171/*
1172 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1173 * log space. This code pushes on the lsn which would supposedly free up
1174 * the 25% which we want to leave free. We may need to adopt a policy which
1175 * pushes on an lsn which is further along in the log once we reach the high
1176 * water mark. In this manner, we would be creating a low water mark.
1177 */
1178STATIC void
1179xlog_grant_push_ail(
1180 struct log *log,
1181 int need_bytes)
1182{
1183 xfs_lsn_t threshold_lsn = 0;
1184 xfs_lsn_t last_sync_lsn;
1185 int free_blocks;
1186 int free_bytes;
1187 int threshold_block;
1188 int threshold_cycle;
1189 int free_threshold;
1190
1191 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1192
1193 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
1194 free_blocks = BTOBBT(free_bytes);
1195
1196 /*
1197 * Set the threshold for the minimum number of free blocks in the
1198 * log to the maximum of what the caller needs, one quarter of the
1199 * log, and 256 blocks.
1200 */
1201 free_threshold = BTOBB(need_bytes);
1202 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1203 free_threshold = MAX(free_threshold, 256);
1204 if (free_blocks >= free_threshold)
1205 return;
1206
1207 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1208 &threshold_block);
1209 threshold_block += free_threshold;
1210 if (threshold_block >= log->l_logBBsize) {
1211 threshold_block -= log->l_logBBsize;
1212 threshold_cycle += 1;
1213 }
1214 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1215 threshold_block);
1216 /*
1217 * Don't pass in an lsn greater than the lsn of the last
1218 * log record known to be on disk. Use a snapshot of the last sync lsn
1219 * so that it doesn't change between the compare and the set.
1220 */
1221 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1222 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1223 threshold_lsn = last_sync_lsn;
1224
1225 /*
1226 * Get the transaction layer to kick the dirty buffers out to
1227 * disk asynchronously. No point in trying to do this if
1228 * the filesystem is shutting down.
1229 */
1230 if (!XLOG_FORCED_SHUTDOWN(log))
1231 xfs_ail_push(log->l_ailp, threshold_lsn);
1232}
1233
1234/*
1235 * The bdstrat callback function for log bufs. This gives us a central
1236 * place to trap bufs in case we get hit by a log I/O error and need to
1237 * shutdown. Actually, in practice, even when we didn't get a log error,
1238 * we transition the iclogs to IOERROR state *after* flushing all existing
1239 * iclogs to disk. This is because we don't want anymore new transactions to be
1240 * started or completed afterwards.
1241 */
1242STATIC int
1243xlog_bdstrat(
1244 struct xfs_buf *bp)
1245{
1246 struct xlog_in_core *iclog = bp->b_fspriv;
1247
1248 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1249 xfs_buf_ioerror(bp, EIO);
1250 XFS_BUF_STALE(bp);
1251 xfs_buf_ioend(bp, 0);
1252 /*
1253 * It would seem logical to return EIO here, but we rely on
1254 * the log state machine to propagate I/O errors instead of
1255 * doing it here.
1256 */
1257 return 0;
1258 }
1259
1260 xfs_buf_iorequest(bp);
1261 return 0;
1262}
1263
1264/*
1265 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1266 * fashion. Previously, we should have moved the current iclog
1267 * ptr in the log to point to the next available iclog. This allows further
1268 * write to continue while this code syncs out an iclog ready to go.
1269 * Before an in-core log can be written out, the data section must be scanned
1270 * to save away the 1st word of each BBSIZE block into the header. We replace
1271 * it with the current cycle count. Each BBSIZE block is tagged with the
1272 * cycle count because there in an implicit assumption that drives will
1273 * guarantee that entire 512 byte blocks get written at once. In other words,
1274 * we can't have part of a 512 byte block written and part not written. By
1275 * tagging each block, we will know which blocks are valid when recovering
1276 * after an unclean shutdown.
1277 *
1278 * This routine is single threaded on the iclog. No other thread can be in
1279 * this routine with the same iclog. Changing contents of iclog can there-
1280 * fore be done without grabbing the state machine lock. Updating the global
1281 * log will require grabbing the lock though.
1282 *
1283 * The entire log manager uses a logical block numbering scheme. Only
1284 * log_sync (and then only bwrite()) know about the fact that the log may
1285 * not start with block zero on a given device. The log block start offset
1286 * is added immediately before calling bwrite().
1287 */
1288
1289STATIC int
1290xlog_sync(xlog_t *log,
1291 xlog_in_core_t *iclog)
1292{
1293 xfs_caddr_t dptr; /* pointer to byte sized element */
1294 xfs_buf_t *bp;
1295 int i;
1296 uint count; /* byte count of bwrite */
1297 uint count_init; /* initial count before roundup */
1298 int roundoff; /* roundoff to BB or stripe */
1299 int split = 0; /* split write into two regions */
1300 int error;
1301 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1302
1303 XFS_STATS_INC(xs_log_writes);
1304 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1305
1306 /* Add for LR header */
1307 count_init = log->l_iclog_hsize + iclog->ic_offset;
1308
1309 /* Round out the log write size */
1310 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1311 /* we have a v2 stripe unit to use */
1312 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1313 } else {
1314 count = BBTOB(BTOBB(count_init));
1315 }
1316 roundoff = count - count_init;
1317 ASSERT(roundoff >= 0);
1318 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1319 roundoff < log->l_mp->m_sb.sb_logsunit)
1320 ||
1321 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1322 roundoff < BBTOB(1)));
1323
1324 /* move grant heads by roundoff in sync */
1325 xlog_grant_add_space(log, &log->l_grant_reserve_head, roundoff);
1326 xlog_grant_add_space(log, &log->l_grant_write_head, roundoff);
1327
1328 /* put cycle number in every block */
1329 xlog_pack_data(log, iclog, roundoff);
1330
1331 /* real byte length */
1332 if (v2) {
1333 iclog->ic_header.h_len =
1334 cpu_to_be32(iclog->ic_offset + roundoff);
1335 } else {
1336 iclog->ic_header.h_len =
1337 cpu_to_be32(iclog->ic_offset);
1338 }
1339
1340 bp = iclog->ic_bp;
1341 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1342
1343 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1344
1345 /* Do we need to split this write into 2 parts? */
1346 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1347 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1348 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1349 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1350 } else {
1351 iclog->ic_bwritecnt = 1;
1352 }
1353 XFS_BUF_SET_COUNT(bp, count);
1354 bp->b_fspriv = iclog;
1355 XFS_BUF_ZEROFLAGS(bp);
1356 XFS_BUF_ASYNC(bp);
1357 bp->b_flags |= XBF_SYNCIO;
1358
1359 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1360 bp->b_flags |= XBF_FUA;
1361
1362 /*
1363 * Flush the data device before flushing the log to make
1364 * sure all meta data written back from the AIL actually made
1365 * it to disk before stamping the new log tail LSN into the
1366 * log buffer. For an external log we need to issue the
1367 * flush explicitly, and unfortunately synchronously here;
1368 * for an internal log we can simply use the block layer
1369 * state machine for preflushes.
1370 */
1371 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1372 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1373 else
1374 bp->b_flags |= XBF_FLUSH;
1375 }
1376
1377 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1378 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1379
1380 xlog_verify_iclog(log, iclog, count, B_TRUE);
1381
1382 /* account for log which doesn't start at block #0 */
1383 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1384 /*
1385 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1386 * is shutting down.
1387 */
1388 XFS_BUF_WRITE(bp);
1389
1390 if ((error = xlog_bdstrat(bp))) {
1391 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1392 XFS_BUF_ADDR(bp));
1393 return error;
1394 }
1395 if (split) {
1396 bp = iclog->ic_log->l_xbuf;
1397 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1398 xfs_buf_associate_memory(bp,
1399 (char *)&iclog->ic_header + count, split);
1400 bp->b_fspriv = iclog;
1401 XFS_BUF_ZEROFLAGS(bp);
1402 XFS_BUF_ASYNC(bp);
1403 bp->b_flags |= XBF_SYNCIO;
1404 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1405 bp->b_flags |= XBF_FUA;
1406 dptr = bp->b_addr;
1407 /*
1408 * Bump the cycle numbers at the start of each block
1409 * since this part of the buffer is at the start of
1410 * a new cycle. Watch out for the header magic number
1411 * case, though.
1412 */
1413 for (i = 0; i < split; i += BBSIZE) {
1414 be32_add_cpu((__be32 *)dptr, 1);
1415 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1416 be32_add_cpu((__be32 *)dptr, 1);
1417 dptr += BBSIZE;
1418 }
1419
1420 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1421 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1422
1423 /* account for internal log which doesn't start at block #0 */
1424 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1425 XFS_BUF_WRITE(bp);
1426 if ((error = xlog_bdstrat(bp))) {
1427 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1428 bp, XFS_BUF_ADDR(bp));
1429 return error;
1430 }
1431 }
1432 return 0;
1433} /* xlog_sync */
1434
1435
1436/*
1437 * Deallocate a log structure
1438 */
1439STATIC void
1440xlog_dealloc_log(xlog_t *log)
1441{
1442 xlog_in_core_t *iclog, *next_iclog;
1443 int i;
1444
1445 xlog_cil_destroy(log);
1446
1447 /*
1448 * always need to ensure that the extra buffer does not point to memory
1449 * owned by another log buffer before we free it.
1450 */
1451 xfs_buf_set_empty(log->l_xbuf, log->l_iclog_size);
1452 xfs_buf_free(log->l_xbuf);
1453
1454 iclog = log->l_iclog;
1455 for (i=0; i<log->l_iclog_bufs; i++) {
1456 xfs_buf_free(iclog->ic_bp);
1457 next_iclog = iclog->ic_next;
1458 kmem_free(iclog);
1459 iclog = next_iclog;
1460 }
1461 spinlock_destroy(&log->l_icloglock);
1462
1463 log->l_mp->m_log = NULL;
1464 kmem_free(log);
1465} /* xlog_dealloc_log */
1466
1467/*
1468 * Update counters atomically now that memcpy is done.
1469 */
1470/* ARGSUSED */
1471static inline void
1472xlog_state_finish_copy(xlog_t *log,
1473 xlog_in_core_t *iclog,
1474 int record_cnt,
1475 int copy_bytes)
1476{
1477 spin_lock(&log->l_icloglock);
1478
1479 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1480 iclog->ic_offset += copy_bytes;
1481
1482 spin_unlock(&log->l_icloglock);
1483} /* xlog_state_finish_copy */
1484
1485
1486
1487
1488/*
1489 * print out info relating to regions written which consume
1490 * the reservation
1491 */
1492void
1493xlog_print_tic_res(
1494 struct xfs_mount *mp,
1495 struct xlog_ticket *ticket)
1496{
1497 uint i;
1498 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1499
1500 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1501 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1502 "bformat",
1503 "bchunk",
1504 "efi_format",
1505 "efd_format",
1506 "iformat",
1507 "icore",
1508 "iext",
1509 "ibroot",
1510 "ilocal",
1511 "iattr_ext",
1512 "iattr_broot",
1513 "iattr_local",
1514 "qformat",
1515 "dquot",
1516 "quotaoff",
1517 "LR header",
1518 "unmount",
1519 "commit",
1520 "trans header"
1521 };
1522 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1523 "SETATTR_NOT_SIZE",
1524 "SETATTR_SIZE",
1525 "INACTIVE",
1526 "CREATE",
1527 "CREATE_TRUNC",
1528 "TRUNCATE_FILE",
1529 "REMOVE",
1530 "LINK",
1531 "RENAME",
1532 "MKDIR",
1533 "RMDIR",
1534 "SYMLINK",
1535 "SET_DMATTRS",
1536 "GROWFS",
1537 "STRAT_WRITE",
1538 "DIOSTRAT",
1539 "WRITE_SYNC",
1540 "WRITEID",
1541 "ADDAFORK",
1542 "ATTRINVAL",
1543 "ATRUNCATE",
1544 "ATTR_SET",
1545 "ATTR_RM",
1546 "ATTR_FLAG",
1547 "CLEAR_AGI_BUCKET",
1548 "QM_SBCHANGE",
1549 "DUMMY1",
1550 "DUMMY2",
1551 "QM_QUOTAOFF",
1552 "QM_DQALLOC",
1553 "QM_SETQLIM",
1554 "QM_DQCLUSTER",
1555 "QM_QINOCREATE",
1556 "QM_QUOTAOFF_END",
1557 "SB_UNIT",
1558 "FSYNC_TS",
1559 "GROWFSRT_ALLOC",
1560 "GROWFSRT_ZERO",
1561 "GROWFSRT_FREE",
1562 "SWAPEXT"
1563 };
1564
1565 xfs_warn(mp,
1566 "xfs_log_write: reservation summary:\n"
1567 " trans type = %s (%u)\n"
1568 " unit res = %d bytes\n"
1569 " current res = %d bytes\n"
1570 " total reg = %u bytes (o/flow = %u bytes)\n"
1571 " ophdrs = %u (ophdr space = %u bytes)\n"
1572 " ophdr + reg = %u bytes\n"
1573 " num regions = %u\n",
1574 ((ticket->t_trans_type <= 0 ||
1575 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1576 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1577 ticket->t_trans_type,
1578 ticket->t_unit_res,
1579 ticket->t_curr_res,
1580 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1581 ticket->t_res_num_ophdrs, ophdr_spc,
1582 ticket->t_res_arr_sum +
1583 ticket->t_res_o_flow + ophdr_spc,
1584 ticket->t_res_num);
1585
1586 for (i = 0; i < ticket->t_res_num; i++) {
1587 uint r_type = ticket->t_res_arr[i].r_type;
1588 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1589 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1590 "bad-rtype" : res_type_str[r_type-1]),
1591 ticket->t_res_arr[i].r_len);
1592 }
1593
1594 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1595 "xfs_log_write: reservation ran out. Need to up reservation");
1596 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1597}
1598
1599/*
1600 * Calculate the potential space needed by the log vector. Each region gets
1601 * its own xlog_op_header_t and may need to be double word aligned.
1602 */
1603static int
1604xlog_write_calc_vec_length(
1605 struct xlog_ticket *ticket,
1606 struct xfs_log_vec *log_vector)
1607{
1608 struct xfs_log_vec *lv;
1609 int headers = 0;
1610 int len = 0;
1611 int i;
1612
1613 /* acct for start rec of xact */
1614 if (ticket->t_flags & XLOG_TIC_INITED)
1615 headers++;
1616
1617 for (lv = log_vector; lv; lv = lv->lv_next) {
1618 headers += lv->lv_niovecs;
1619
1620 for (i = 0; i < lv->lv_niovecs; i++) {
1621 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1622
1623 len += vecp->i_len;
1624 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1625 }
1626 }
1627
1628 ticket->t_res_num_ophdrs += headers;
1629 len += headers * sizeof(struct xlog_op_header);
1630
1631 return len;
1632}
1633
1634/*
1635 * If first write for transaction, insert start record We can't be trying to
1636 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1637 */
1638static int
1639xlog_write_start_rec(
1640 struct xlog_op_header *ophdr,
1641 struct xlog_ticket *ticket)
1642{
1643 if (!(ticket->t_flags & XLOG_TIC_INITED))
1644 return 0;
1645
1646 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1647 ophdr->oh_clientid = ticket->t_clientid;
1648 ophdr->oh_len = 0;
1649 ophdr->oh_flags = XLOG_START_TRANS;
1650 ophdr->oh_res2 = 0;
1651
1652 ticket->t_flags &= ~XLOG_TIC_INITED;
1653
1654 return sizeof(struct xlog_op_header);
1655}
1656
1657static xlog_op_header_t *
1658xlog_write_setup_ophdr(
1659 struct log *log,
1660 struct xlog_op_header *ophdr,
1661 struct xlog_ticket *ticket,
1662 uint flags)
1663{
1664 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1665 ophdr->oh_clientid = ticket->t_clientid;
1666 ophdr->oh_res2 = 0;
1667
1668 /* are we copying a commit or unmount record? */
1669 ophdr->oh_flags = flags;
1670
1671 /*
1672 * We've seen logs corrupted with bad transaction client ids. This
1673 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1674 * and shut down the filesystem.
1675 */
1676 switch (ophdr->oh_clientid) {
1677 case XFS_TRANSACTION:
1678 case XFS_VOLUME:
1679 case XFS_LOG:
1680 break;
1681 default:
1682 xfs_warn(log->l_mp,
1683 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1684 ophdr->oh_clientid, ticket);
1685 return NULL;
1686 }
1687
1688 return ophdr;
1689}
1690
1691/*
1692 * Set up the parameters of the region copy into the log. This has
1693 * to handle region write split across multiple log buffers - this
1694 * state is kept external to this function so that this code can
1695 * can be written in an obvious, self documenting manner.
1696 */
1697static int
1698xlog_write_setup_copy(
1699 struct xlog_ticket *ticket,
1700 struct xlog_op_header *ophdr,
1701 int space_available,
1702 int space_required,
1703 int *copy_off,
1704 int *copy_len,
1705 int *last_was_partial_copy,
1706 int *bytes_consumed)
1707{
1708 int still_to_copy;
1709
1710 still_to_copy = space_required - *bytes_consumed;
1711 *copy_off = *bytes_consumed;
1712
1713 if (still_to_copy <= space_available) {
1714 /* write of region completes here */
1715 *copy_len = still_to_copy;
1716 ophdr->oh_len = cpu_to_be32(*copy_len);
1717 if (*last_was_partial_copy)
1718 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1719 *last_was_partial_copy = 0;
1720 *bytes_consumed = 0;
1721 return 0;
1722 }
1723
1724 /* partial write of region, needs extra log op header reservation */
1725 *copy_len = space_available;
1726 ophdr->oh_len = cpu_to_be32(*copy_len);
1727 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1728 if (*last_was_partial_copy)
1729 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1730 *bytes_consumed += *copy_len;
1731 (*last_was_partial_copy)++;
1732
1733 /* account for new log op header */
1734 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1735 ticket->t_res_num_ophdrs++;
1736
1737 return sizeof(struct xlog_op_header);
1738}
1739
1740static int
1741xlog_write_copy_finish(
1742 struct log *log,
1743 struct xlog_in_core *iclog,
1744 uint flags,
1745 int *record_cnt,
1746 int *data_cnt,
1747 int *partial_copy,
1748 int *partial_copy_len,
1749 int log_offset,
1750 struct xlog_in_core **commit_iclog)
1751{
1752 if (*partial_copy) {
1753 /*
1754 * This iclog has already been marked WANT_SYNC by
1755 * xlog_state_get_iclog_space.
1756 */
1757 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1758 *record_cnt = 0;
1759 *data_cnt = 0;
1760 return xlog_state_release_iclog(log, iclog);
1761 }
1762
1763 *partial_copy = 0;
1764 *partial_copy_len = 0;
1765
1766 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1767 /* no more space in this iclog - push it. */
1768 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1769 *record_cnt = 0;
1770 *data_cnt = 0;
1771
1772 spin_lock(&log->l_icloglock);
1773 xlog_state_want_sync(log, iclog);
1774 spin_unlock(&log->l_icloglock);
1775
1776 if (!commit_iclog)
1777 return xlog_state_release_iclog(log, iclog);
1778 ASSERT(flags & XLOG_COMMIT_TRANS);
1779 *commit_iclog = iclog;
1780 }
1781
1782 return 0;
1783}
1784
1785/*
1786 * Write some region out to in-core log
1787 *
1788 * This will be called when writing externally provided regions or when
1789 * writing out a commit record for a given transaction.
1790 *
1791 * General algorithm:
1792 * 1. Find total length of this write. This may include adding to the
1793 * lengths passed in.
1794 * 2. Check whether we violate the tickets reservation.
1795 * 3. While writing to this iclog
1796 * A. Reserve as much space in this iclog as can get
1797 * B. If this is first write, save away start lsn
1798 * C. While writing this region:
1799 * 1. If first write of transaction, write start record
1800 * 2. Write log operation header (header per region)
1801 * 3. Find out if we can fit entire region into this iclog
1802 * 4. Potentially, verify destination memcpy ptr
1803 * 5. Memcpy (partial) region
1804 * 6. If partial copy, release iclog; otherwise, continue
1805 * copying more regions into current iclog
1806 * 4. Mark want sync bit (in simulation mode)
1807 * 5. Release iclog for potential flush to on-disk log.
1808 *
1809 * ERRORS:
1810 * 1. Panic if reservation is overrun. This should never happen since
1811 * reservation amounts are generated internal to the filesystem.
1812 * NOTES:
1813 * 1. Tickets are single threaded data structures.
1814 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1815 * syncing routine. When a single log_write region needs to span
1816 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1817 * on all log operation writes which don't contain the end of the
1818 * region. The XLOG_END_TRANS bit is used for the in-core log
1819 * operation which contains the end of the continued log_write region.
1820 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1821 * we don't really know exactly how much space will be used. As a result,
1822 * we don't update ic_offset until the end when we know exactly how many
1823 * bytes have been written out.
1824 */
1825int
1826xlog_write(
1827 struct log *log,
1828 struct xfs_log_vec *log_vector,
1829 struct xlog_ticket *ticket,
1830 xfs_lsn_t *start_lsn,
1831 struct xlog_in_core **commit_iclog,
1832 uint flags)
1833{
1834 struct xlog_in_core *iclog = NULL;
1835 struct xfs_log_iovec *vecp;
1836 struct xfs_log_vec *lv;
1837 int len;
1838 int index;
1839 int partial_copy = 0;
1840 int partial_copy_len = 0;
1841 int contwr = 0;
1842 int record_cnt = 0;
1843 int data_cnt = 0;
1844 int error;
1845
1846 *start_lsn = 0;
1847
1848 len = xlog_write_calc_vec_length(ticket, log_vector);
1849 if (log->l_cilp) {
1850 /*
1851 * Region headers and bytes are already accounted for.
1852 * We only need to take into account start records and
1853 * split regions in this function.
1854 */
1855 if (ticket->t_flags & XLOG_TIC_INITED)
1856 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1857
1858 /*
1859 * Commit record headers need to be accounted for. These
1860 * come in as separate writes so are easy to detect.
1861 */
1862 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1863 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1864 } else
1865 ticket->t_curr_res -= len;
1866
1867 if (ticket->t_curr_res < 0)
1868 xlog_print_tic_res(log->l_mp, ticket);
1869
1870 index = 0;
1871 lv = log_vector;
1872 vecp = lv->lv_iovecp;
1873 while (lv && index < lv->lv_niovecs) {
1874 void *ptr;
1875 int log_offset;
1876
1877 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1878 &contwr, &log_offset);
1879 if (error)
1880 return error;
1881
1882 ASSERT(log_offset <= iclog->ic_size - 1);
1883 ptr = iclog->ic_datap + log_offset;
1884
1885 /* start_lsn is the first lsn written to. That's all we need. */
1886 if (!*start_lsn)
1887 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1888
1889 /*
1890 * This loop writes out as many regions as can fit in the amount
1891 * of space which was allocated by xlog_state_get_iclog_space().
1892 */
1893 while (lv && index < lv->lv_niovecs) {
1894 struct xfs_log_iovec *reg = &vecp[index];
1895 struct xlog_op_header *ophdr;
1896 int start_rec_copy;
1897 int copy_len;
1898 int copy_off;
1899
1900 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1901 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1902
1903 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1904 if (start_rec_copy) {
1905 record_cnt++;
1906 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1907 start_rec_copy);
1908 }
1909
1910 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1911 if (!ophdr)
1912 return XFS_ERROR(EIO);
1913
1914 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1915 sizeof(struct xlog_op_header));
1916
1917 len += xlog_write_setup_copy(ticket, ophdr,
1918 iclog->ic_size-log_offset,
1919 reg->i_len,
1920 ©_off, ©_len,
1921 &partial_copy,
1922 &partial_copy_len);
1923 xlog_verify_dest_ptr(log, ptr);
1924
1925 /* copy region */
1926 ASSERT(copy_len >= 0);
1927 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1928 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1929
1930 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1931 record_cnt++;
1932 data_cnt += contwr ? copy_len : 0;
1933
1934 error = xlog_write_copy_finish(log, iclog, flags,
1935 &record_cnt, &data_cnt,
1936 &partial_copy,
1937 &partial_copy_len,
1938 log_offset,
1939 commit_iclog);
1940 if (error)
1941 return error;
1942
1943 /*
1944 * if we had a partial copy, we need to get more iclog
1945 * space but we don't want to increment the region
1946 * index because there is still more is this region to
1947 * write.
1948 *
1949 * If we completed writing this region, and we flushed
1950 * the iclog (indicated by resetting of the record
1951 * count), then we also need to get more log space. If
1952 * this was the last record, though, we are done and
1953 * can just return.
1954 */
1955 if (partial_copy)
1956 break;
1957
1958 if (++index == lv->lv_niovecs) {
1959 lv = lv->lv_next;
1960 index = 0;
1961 if (lv)
1962 vecp = lv->lv_iovecp;
1963 }
1964 if (record_cnt == 0) {
1965 if (!lv)
1966 return 0;
1967 break;
1968 }
1969 }
1970 }
1971
1972 ASSERT(len == 0);
1973
1974 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1975 if (!commit_iclog)
1976 return xlog_state_release_iclog(log, iclog);
1977
1978 ASSERT(flags & XLOG_COMMIT_TRANS);
1979 *commit_iclog = iclog;
1980 return 0;
1981}
1982
1983
1984/*****************************************************************************
1985 *
1986 * State Machine functions
1987 *
1988 *****************************************************************************
1989 */
1990
1991/* Clean iclogs starting from the head. This ordering must be
1992 * maintained, so an iclog doesn't become ACTIVE beyond one that
1993 * is SYNCING. This is also required to maintain the notion that we use
1994 * a ordered wait queue to hold off would be writers to the log when every
1995 * iclog is trying to sync to disk.
1996 *
1997 * State Change: DIRTY -> ACTIVE
1998 */
1999STATIC void
2000xlog_state_clean_log(xlog_t *log)
2001{
2002 xlog_in_core_t *iclog;
2003 int changed = 0;
2004
2005 iclog = log->l_iclog;
2006 do {
2007 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2008 iclog->ic_state = XLOG_STATE_ACTIVE;
2009 iclog->ic_offset = 0;
2010 ASSERT(iclog->ic_callback == NULL);
2011 /*
2012 * If the number of ops in this iclog indicate it just
2013 * contains the dummy transaction, we can
2014 * change state into IDLE (the second time around).
2015 * Otherwise we should change the state into
2016 * NEED a dummy.
2017 * We don't need to cover the dummy.
2018 */
2019 if (!changed &&
2020 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2021 XLOG_COVER_OPS)) {
2022 changed = 1;
2023 } else {
2024 /*
2025 * We have two dirty iclogs so start over
2026 * This could also be num of ops indicates
2027 * this is not the dummy going out.
2028 */
2029 changed = 2;
2030 }
2031 iclog->ic_header.h_num_logops = 0;
2032 memset(iclog->ic_header.h_cycle_data, 0,
2033 sizeof(iclog->ic_header.h_cycle_data));
2034 iclog->ic_header.h_lsn = 0;
2035 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2036 /* do nothing */;
2037 else
2038 break; /* stop cleaning */
2039 iclog = iclog->ic_next;
2040 } while (iclog != log->l_iclog);
2041
2042 /* log is locked when we are called */
2043 /*
2044 * Change state for the dummy log recording.
2045 * We usually go to NEED. But we go to NEED2 if the changed indicates
2046 * we are done writing the dummy record.
2047 * If we are done with the second dummy recored (DONE2), then
2048 * we go to IDLE.
2049 */
2050 if (changed) {
2051 switch (log->l_covered_state) {
2052 case XLOG_STATE_COVER_IDLE:
2053 case XLOG_STATE_COVER_NEED:
2054 case XLOG_STATE_COVER_NEED2:
2055 log->l_covered_state = XLOG_STATE_COVER_NEED;
2056 break;
2057
2058 case XLOG_STATE_COVER_DONE:
2059 if (changed == 1)
2060 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2061 else
2062 log->l_covered_state = XLOG_STATE_COVER_NEED;
2063 break;
2064
2065 case XLOG_STATE_COVER_DONE2:
2066 if (changed == 1)
2067 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2068 else
2069 log->l_covered_state = XLOG_STATE_COVER_NEED;
2070 break;
2071
2072 default:
2073 ASSERT(0);
2074 }
2075 }
2076} /* xlog_state_clean_log */
2077
2078STATIC xfs_lsn_t
2079xlog_get_lowest_lsn(
2080 xlog_t *log)
2081{
2082 xlog_in_core_t *lsn_log;
2083 xfs_lsn_t lowest_lsn, lsn;
2084
2085 lsn_log = log->l_iclog;
2086 lowest_lsn = 0;
2087 do {
2088 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2089 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2090 if ((lsn && !lowest_lsn) ||
2091 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2092 lowest_lsn = lsn;
2093 }
2094 }
2095 lsn_log = lsn_log->ic_next;
2096 } while (lsn_log != log->l_iclog);
2097 return lowest_lsn;
2098}
2099
2100
2101STATIC void
2102xlog_state_do_callback(
2103 xlog_t *log,
2104 int aborted,
2105 xlog_in_core_t *ciclog)
2106{
2107 xlog_in_core_t *iclog;
2108 xlog_in_core_t *first_iclog; /* used to know when we've
2109 * processed all iclogs once */
2110 xfs_log_callback_t *cb, *cb_next;
2111 int flushcnt = 0;
2112 xfs_lsn_t lowest_lsn;
2113 int ioerrors; /* counter: iclogs with errors */
2114 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2115 int funcdidcallbacks; /* flag: function did callbacks */
2116 int repeats; /* for issuing console warnings if
2117 * looping too many times */
2118 int wake = 0;
2119
2120 spin_lock(&log->l_icloglock);
2121 first_iclog = iclog = log->l_iclog;
2122 ioerrors = 0;
2123 funcdidcallbacks = 0;
2124 repeats = 0;
2125
2126 do {
2127 /*
2128 * Scan all iclogs starting with the one pointed to by the
2129 * log. Reset this starting point each time the log is
2130 * unlocked (during callbacks).
2131 *
2132 * Keep looping through iclogs until one full pass is made
2133 * without running any callbacks.
2134 */
2135 first_iclog = log->l_iclog;
2136 iclog = log->l_iclog;
2137 loopdidcallbacks = 0;
2138 repeats++;
2139
2140 do {
2141
2142 /* skip all iclogs in the ACTIVE & DIRTY states */
2143 if (iclog->ic_state &
2144 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2145 iclog = iclog->ic_next;
2146 continue;
2147 }
2148
2149 /*
2150 * Between marking a filesystem SHUTDOWN and stopping
2151 * the log, we do flush all iclogs to disk (if there
2152 * wasn't a log I/O error). So, we do want things to
2153 * go smoothly in case of just a SHUTDOWN w/o a
2154 * LOG_IO_ERROR.
2155 */
2156 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2157 /*
2158 * Can only perform callbacks in order. Since
2159 * this iclog is not in the DONE_SYNC/
2160 * DO_CALLBACK state, we skip the rest and
2161 * just try to clean up. If we set our iclog
2162 * to DO_CALLBACK, we will not process it when
2163 * we retry since a previous iclog is in the
2164 * CALLBACK and the state cannot change since
2165 * we are holding the l_icloglock.
2166 */
2167 if (!(iclog->ic_state &
2168 (XLOG_STATE_DONE_SYNC |
2169 XLOG_STATE_DO_CALLBACK))) {
2170 if (ciclog && (ciclog->ic_state ==
2171 XLOG_STATE_DONE_SYNC)) {
2172 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2173 }
2174 break;
2175 }
2176 /*
2177 * We now have an iclog that is in either the
2178 * DO_CALLBACK or DONE_SYNC states. The other
2179 * states (WANT_SYNC, SYNCING, or CALLBACK were
2180 * caught by the above if and are going to
2181 * clean (i.e. we aren't doing their callbacks)
2182 * see the above if.
2183 */
2184
2185 /*
2186 * We will do one more check here to see if we
2187 * have chased our tail around.
2188 */
2189
2190 lowest_lsn = xlog_get_lowest_lsn(log);
2191 if (lowest_lsn &&
2192 XFS_LSN_CMP(lowest_lsn,
2193 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2194 iclog = iclog->ic_next;
2195 continue; /* Leave this iclog for
2196 * another thread */
2197 }
2198
2199 iclog->ic_state = XLOG_STATE_CALLBACK;
2200
2201
2202 /*
2203 * update the last_sync_lsn before we drop the
2204 * icloglock to ensure we are the only one that
2205 * can update it.
2206 */
2207 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2208 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2209 atomic64_set(&log->l_last_sync_lsn,
2210 be64_to_cpu(iclog->ic_header.h_lsn));
2211
2212 } else
2213 ioerrors++;
2214
2215 spin_unlock(&log->l_icloglock);
2216
2217 /*
2218 * Keep processing entries in the callback list until
2219 * we come around and it is empty. We need to
2220 * atomically see that the list is empty and change the
2221 * state to DIRTY so that we don't miss any more
2222 * callbacks being added.
2223 */
2224 spin_lock(&iclog->ic_callback_lock);
2225 cb = iclog->ic_callback;
2226 while (cb) {
2227 iclog->ic_callback_tail = &(iclog->ic_callback);
2228 iclog->ic_callback = NULL;
2229 spin_unlock(&iclog->ic_callback_lock);
2230
2231 /* perform callbacks in the order given */
2232 for (; cb; cb = cb_next) {
2233 cb_next = cb->cb_next;
2234 cb->cb_func(cb->cb_arg, aborted);
2235 }
2236 spin_lock(&iclog->ic_callback_lock);
2237 cb = iclog->ic_callback;
2238 }
2239
2240 loopdidcallbacks++;
2241 funcdidcallbacks++;
2242
2243 spin_lock(&log->l_icloglock);
2244 ASSERT(iclog->ic_callback == NULL);
2245 spin_unlock(&iclog->ic_callback_lock);
2246 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2247 iclog->ic_state = XLOG_STATE_DIRTY;
2248
2249 /*
2250 * Transition from DIRTY to ACTIVE if applicable.
2251 * NOP if STATE_IOERROR.
2252 */
2253 xlog_state_clean_log(log);
2254
2255 /* wake up threads waiting in xfs_log_force() */
2256 wake_up_all(&iclog->ic_force_wait);
2257
2258 iclog = iclog->ic_next;
2259 } while (first_iclog != iclog);
2260
2261 if (repeats > 5000) {
2262 flushcnt += repeats;
2263 repeats = 0;
2264 xfs_warn(log->l_mp,
2265 "%s: possible infinite loop (%d iterations)",
2266 __func__, flushcnt);
2267 }
2268 } while (!ioerrors && loopdidcallbacks);
2269
2270 /*
2271 * make one last gasp attempt to see if iclogs are being left in
2272 * limbo..
2273 */
2274#ifdef DEBUG
2275 if (funcdidcallbacks) {
2276 first_iclog = iclog = log->l_iclog;
2277 do {
2278 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2279 /*
2280 * Terminate the loop if iclogs are found in states
2281 * which will cause other threads to clean up iclogs.
2282 *
2283 * SYNCING - i/o completion will go through logs
2284 * DONE_SYNC - interrupt thread should be waiting for
2285 * l_icloglock
2286 * IOERROR - give up hope all ye who enter here
2287 */
2288 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2289 iclog->ic_state == XLOG_STATE_SYNCING ||
2290 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2291 iclog->ic_state == XLOG_STATE_IOERROR )
2292 break;
2293 iclog = iclog->ic_next;
2294 } while (first_iclog != iclog);
2295 }
2296#endif
2297
2298 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2299 wake = 1;
2300 spin_unlock(&log->l_icloglock);
2301
2302 if (wake)
2303 wake_up_all(&log->l_flush_wait);
2304}
2305
2306
2307/*
2308 * Finish transitioning this iclog to the dirty state.
2309 *
2310 * Make sure that we completely execute this routine only when this is
2311 * the last call to the iclog. There is a good chance that iclog flushes,
2312 * when we reach the end of the physical log, get turned into 2 separate
2313 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2314 * routine. By using the reference count bwritecnt, we guarantee that only
2315 * the second completion goes through.
2316 *
2317 * Callbacks could take time, so they are done outside the scope of the
2318 * global state machine log lock.
2319 */
2320STATIC void
2321xlog_state_done_syncing(
2322 xlog_in_core_t *iclog,
2323 int aborted)
2324{
2325 xlog_t *log = iclog->ic_log;
2326
2327 spin_lock(&log->l_icloglock);
2328
2329 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2330 iclog->ic_state == XLOG_STATE_IOERROR);
2331 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2332 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2333
2334
2335 /*
2336 * If we got an error, either on the first buffer, or in the case of
2337 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2338 * and none should ever be attempted to be written to disk
2339 * again.
2340 */
2341 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2342 if (--iclog->ic_bwritecnt == 1) {
2343 spin_unlock(&log->l_icloglock);
2344 return;
2345 }
2346 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2347 }
2348
2349 /*
2350 * Someone could be sleeping prior to writing out the next
2351 * iclog buffer, we wake them all, one will get to do the
2352 * I/O, the others get to wait for the result.
2353 */
2354 wake_up_all(&iclog->ic_write_wait);
2355 spin_unlock(&log->l_icloglock);
2356 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2357} /* xlog_state_done_syncing */
2358
2359
2360/*
2361 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2362 * sleep. We wait on the flush queue on the head iclog as that should be
2363 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2364 * we will wait here and all new writes will sleep until a sync completes.
2365 *
2366 * The in-core logs are used in a circular fashion. They are not used
2367 * out-of-order even when an iclog past the head is free.
2368 *
2369 * return:
2370 * * log_offset where xlog_write() can start writing into the in-core
2371 * log's data space.
2372 * * in-core log pointer to which xlog_write() should write.
2373 * * boolean indicating this is a continued write to an in-core log.
2374 * If this is the last write, then the in-core log's offset field
2375 * needs to be incremented, depending on the amount of data which
2376 * is copied.
2377 */
2378STATIC int
2379xlog_state_get_iclog_space(xlog_t *log,
2380 int len,
2381 xlog_in_core_t **iclogp,
2382 xlog_ticket_t *ticket,
2383 int *continued_write,
2384 int *logoffsetp)
2385{
2386 int log_offset;
2387 xlog_rec_header_t *head;
2388 xlog_in_core_t *iclog;
2389 int error;
2390
2391restart:
2392 spin_lock(&log->l_icloglock);
2393 if (XLOG_FORCED_SHUTDOWN(log)) {
2394 spin_unlock(&log->l_icloglock);
2395 return XFS_ERROR(EIO);
2396 }
2397
2398 iclog = log->l_iclog;
2399 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2400 XFS_STATS_INC(xs_log_noiclogs);
2401
2402 /* Wait for log writes to have flushed */
2403 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2404 goto restart;
2405 }
2406
2407 head = &iclog->ic_header;
2408
2409 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2410 log_offset = iclog->ic_offset;
2411
2412 /* On the 1st write to an iclog, figure out lsn. This works
2413 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2414 * committing to. If the offset is set, that's how many blocks
2415 * must be written.
2416 */
2417 if (log_offset == 0) {
2418 ticket->t_curr_res -= log->l_iclog_hsize;
2419 xlog_tic_add_region(ticket,
2420 log->l_iclog_hsize,
2421 XLOG_REG_TYPE_LRHEADER);
2422 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2423 head->h_lsn = cpu_to_be64(
2424 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2425 ASSERT(log->l_curr_block >= 0);
2426 }
2427
2428 /* If there is enough room to write everything, then do it. Otherwise,
2429 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2430 * bit is on, so this will get flushed out. Don't update ic_offset
2431 * until you know exactly how many bytes get copied. Therefore, wait
2432 * until later to update ic_offset.
2433 *
2434 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2435 * can fit into remaining data section.
2436 */
2437 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2438 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2439
2440 /*
2441 * If I'm the only one writing to this iclog, sync it to disk.
2442 * We need to do an atomic compare and decrement here to avoid
2443 * racing with concurrent atomic_dec_and_lock() calls in
2444 * xlog_state_release_iclog() when there is more than one
2445 * reference to the iclog.
2446 */
2447 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2448 /* we are the only one */
2449 spin_unlock(&log->l_icloglock);
2450 error = xlog_state_release_iclog(log, iclog);
2451 if (error)
2452 return error;
2453 } else {
2454 spin_unlock(&log->l_icloglock);
2455 }
2456 goto restart;
2457 }
2458
2459 /* Do we have enough room to write the full amount in the remainder
2460 * of this iclog? Or must we continue a write on the next iclog and
2461 * mark this iclog as completely taken? In the case where we switch
2462 * iclogs (to mark it taken), this particular iclog will release/sync
2463 * to disk in xlog_write().
2464 */
2465 if (len <= iclog->ic_size - iclog->ic_offset) {
2466 *continued_write = 0;
2467 iclog->ic_offset += len;
2468 } else {
2469 *continued_write = 1;
2470 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2471 }
2472 *iclogp = iclog;
2473
2474 ASSERT(iclog->ic_offset <= iclog->ic_size);
2475 spin_unlock(&log->l_icloglock);
2476
2477 *logoffsetp = log_offset;
2478 return 0;
2479} /* xlog_state_get_iclog_space */
2480
2481/*
2482 * Atomically get the log space required for a log ticket.
2483 *
2484 * Once a ticket gets put onto the reserveq, it will only return after
2485 * the needed reservation is satisfied.
2486 *
2487 * This function is structured so that it has a lock free fast path. This is
2488 * necessary because every new transaction reservation will come through this
2489 * path. Hence any lock will be globally hot if we take it unconditionally on
2490 * every pass.
2491 *
2492 * As tickets are only ever moved on and off the reserveq under the
2493 * l_grant_reserve_lock, we only need to take that lock if we are going
2494 * to add the ticket to the queue and sleep. We can avoid taking the lock if the
2495 * ticket was never added to the reserveq because the t_queue list head will be
2496 * empty and we hold the only reference to it so it can safely be checked
2497 * unlocked.
2498 */
2499STATIC int
2500xlog_grant_log_space(xlog_t *log,
2501 xlog_ticket_t *tic)
2502{
2503 int free_bytes;
2504 int need_bytes;
2505
2506#ifdef DEBUG
2507 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2508 panic("grant Recovery problem");
2509#endif
2510
2511 trace_xfs_log_grant_enter(log, tic);
2512
2513 need_bytes = tic->t_unit_res;
2514 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2515 need_bytes *= tic->t_ocnt;
2516
2517 /* something is already sleeping; insert new transaction at end */
2518 if (!list_empty_careful(&log->l_reserveq)) {
2519 spin_lock(&log->l_grant_reserve_lock);
2520 /* recheck the queue now we are locked */
2521 if (list_empty(&log->l_reserveq)) {
2522 spin_unlock(&log->l_grant_reserve_lock);
2523 goto redo;
2524 }
2525 list_add_tail(&tic->t_queue, &log->l_reserveq);
2526
2527 trace_xfs_log_grant_sleep1(log, tic);
2528
2529 /*
2530 * Gotta check this before going to sleep, while we're
2531 * holding the grant lock.
2532 */
2533 if (XLOG_FORCED_SHUTDOWN(log))
2534 goto error_return;
2535
2536 XFS_STATS_INC(xs_sleep_logspace);
2537 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2538
2539 /*
2540 * If we got an error, and the filesystem is shutting down,
2541 * we'll catch it down below. So just continue...
2542 */
2543 trace_xfs_log_grant_wake1(log, tic);
2544 }
2545
2546redo:
2547 if (XLOG_FORCED_SHUTDOWN(log))
2548 goto error_return_unlocked;
2549
2550 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
2551 if (free_bytes < need_bytes) {
2552 spin_lock(&log->l_grant_reserve_lock);
2553 if (list_empty(&tic->t_queue))
2554 list_add_tail(&tic->t_queue, &log->l_reserveq);
2555
2556 trace_xfs_log_grant_sleep2(log, tic);
2557
2558 if (XLOG_FORCED_SHUTDOWN(log))
2559 goto error_return;
2560
2561 xlog_grant_push_ail(log, need_bytes);
2562
2563 XFS_STATS_INC(xs_sleep_logspace);
2564 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2565
2566 trace_xfs_log_grant_wake2(log, tic);
2567 goto redo;
2568 }
2569
2570 if (!list_empty(&tic->t_queue)) {
2571 spin_lock(&log->l_grant_reserve_lock);
2572 list_del_init(&tic->t_queue);
2573 spin_unlock(&log->l_grant_reserve_lock);
2574 }
2575
2576 /* we've got enough space */
2577 xlog_grant_add_space(log, &log->l_grant_reserve_head, need_bytes);
2578 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2579 trace_xfs_log_grant_exit(log, tic);
2580 xlog_verify_grant_tail(log);
2581 return 0;
2582
2583error_return_unlocked:
2584 spin_lock(&log->l_grant_reserve_lock);
2585error_return:
2586 list_del_init(&tic->t_queue);
2587 spin_unlock(&log->l_grant_reserve_lock);
2588 trace_xfs_log_grant_error(log, tic);
2589
2590 /*
2591 * If we are failing, make sure the ticket doesn't have any
2592 * current reservations. We don't want to add this back when
2593 * the ticket/transaction gets cancelled.
2594 */
2595 tic->t_curr_res = 0;
2596 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2597 return XFS_ERROR(EIO);
2598} /* xlog_grant_log_space */
2599
2600
2601/*
2602 * Replenish the byte reservation required by moving the grant write head.
2603 *
2604 * Similar to xlog_grant_log_space, the function is structured to have a lock
2605 * free fast path.
2606 */
2607STATIC int
2608xlog_regrant_write_log_space(xlog_t *log,
2609 xlog_ticket_t *tic)
2610{
2611 int free_bytes, need_bytes;
2612
2613 tic->t_curr_res = tic->t_unit_res;
2614 xlog_tic_reset_res(tic);
2615
2616 if (tic->t_cnt > 0)
2617 return 0;
2618
2619#ifdef DEBUG
2620 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2621 panic("regrant Recovery problem");
2622#endif
2623
2624 trace_xfs_log_regrant_write_enter(log, tic);
2625 if (XLOG_FORCED_SHUTDOWN(log))
2626 goto error_return_unlocked;
2627
2628 /* If there are other waiters on the queue then give them a
2629 * chance at logspace before us. Wake up the first waiters,
2630 * if we do not wake up all the waiters then go to sleep waiting
2631 * for more free space, otherwise try to get some space for
2632 * this transaction.
2633 */
2634 need_bytes = tic->t_unit_res;
2635 if (!list_empty_careful(&log->l_writeq)) {
2636 struct xlog_ticket *ntic;
2637
2638 spin_lock(&log->l_grant_write_lock);
2639 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2640 list_for_each_entry(ntic, &log->l_writeq, t_queue) {
2641 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2642
2643 if (free_bytes < ntic->t_unit_res)
2644 break;
2645 free_bytes -= ntic->t_unit_res;
2646 wake_up(&ntic->t_wait);
2647 }
2648
2649 if (ntic != list_first_entry(&log->l_writeq,
2650 struct xlog_ticket, t_queue)) {
2651 if (list_empty(&tic->t_queue))
2652 list_add_tail(&tic->t_queue, &log->l_writeq);
2653 trace_xfs_log_regrant_write_sleep1(log, tic);
2654
2655 xlog_grant_push_ail(log, need_bytes);
2656
2657 XFS_STATS_INC(xs_sleep_logspace);
2658 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2659 trace_xfs_log_regrant_write_wake1(log, tic);
2660 } else
2661 spin_unlock(&log->l_grant_write_lock);
2662 }
2663
2664redo:
2665 if (XLOG_FORCED_SHUTDOWN(log))
2666 goto error_return_unlocked;
2667
2668 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2669 if (free_bytes < need_bytes) {
2670 spin_lock(&log->l_grant_write_lock);
2671 if (list_empty(&tic->t_queue))
2672 list_add_tail(&tic->t_queue, &log->l_writeq);
2673
2674 if (XLOG_FORCED_SHUTDOWN(log))
2675 goto error_return;
2676
2677 xlog_grant_push_ail(log, need_bytes);
2678
2679 XFS_STATS_INC(xs_sleep_logspace);
2680 trace_xfs_log_regrant_write_sleep2(log, tic);
2681 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2682
2683 trace_xfs_log_regrant_write_wake2(log, tic);
2684 goto redo;
2685 }
2686
2687 if (!list_empty(&tic->t_queue)) {
2688 spin_lock(&log->l_grant_write_lock);
2689 list_del_init(&tic->t_queue);
2690 spin_unlock(&log->l_grant_write_lock);
2691 }
2692
2693 /* we've got enough space */
2694 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2695 trace_xfs_log_regrant_write_exit(log, tic);
2696 xlog_verify_grant_tail(log);
2697 return 0;
2698
2699
2700 error_return_unlocked:
2701 spin_lock(&log->l_grant_write_lock);
2702 error_return:
2703 list_del_init(&tic->t_queue);
2704 spin_unlock(&log->l_grant_write_lock);
2705 trace_xfs_log_regrant_write_error(log, tic);
2706
2707 /*
2708 * If we are failing, make sure the ticket doesn't have any
2709 * current reservations. We don't want to add this back when
2710 * the ticket/transaction gets cancelled.
2711 */
2712 tic->t_curr_res = 0;
2713 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2714 return XFS_ERROR(EIO);
2715} /* xlog_regrant_write_log_space */
2716
2717
2718/* The first cnt-1 times through here we don't need to
2719 * move the grant write head because the permanent
2720 * reservation has reserved cnt times the unit amount.
2721 * Release part of current permanent unit reservation and
2722 * reset current reservation to be one units worth. Also
2723 * move grant reservation head forward.
2724 */
2725STATIC void
2726xlog_regrant_reserve_log_space(xlog_t *log,
2727 xlog_ticket_t *ticket)
2728{
2729 trace_xfs_log_regrant_reserve_enter(log, ticket);
2730
2731 if (ticket->t_cnt > 0)
2732 ticket->t_cnt--;
2733
2734 xlog_grant_sub_space(log, &log->l_grant_reserve_head,
2735 ticket->t_curr_res);
2736 xlog_grant_sub_space(log, &log->l_grant_write_head,
2737 ticket->t_curr_res);
2738 ticket->t_curr_res = ticket->t_unit_res;
2739 xlog_tic_reset_res(ticket);
2740
2741 trace_xfs_log_regrant_reserve_sub(log, ticket);
2742
2743 /* just return if we still have some of the pre-reserved space */
2744 if (ticket->t_cnt > 0)
2745 return;
2746
2747 xlog_grant_add_space(log, &log->l_grant_reserve_head,
2748 ticket->t_unit_res);
2749
2750 trace_xfs_log_regrant_reserve_exit(log, ticket);
2751
2752 ticket->t_curr_res = ticket->t_unit_res;
2753 xlog_tic_reset_res(ticket);
2754} /* xlog_regrant_reserve_log_space */
2755
2756
2757/*
2758 * Give back the space left from a reservation.
2759 *
2760 * All the information we need to make a correct determination of space left
2761 * is present. For non-permanent reservations, things are quite easy. The
2762 * count should have been decremented to zero. We only need to deal with the
2763 * space remaining in the current reservation part of the ticket. If the
2764 * ticket contains a permanent reservation, there may be left over space which
2765 * needs to be released. A count of N means that N-1 refills of the current
2766 * reservation can be done before we need to ask for more space. The first
2767 * one goes to fill up the first current reservation. Once we run out of
2768 * space, the count will stay at zero and the only space remaining will be
2769 * in the current reservation field.
2770 */
2771STATIC void
2772xlog_ungrant_log_space(xlog_t *log,
2773 xlog_ticket_t *ticket)
2774{
2775 int bytes;
2776
2777 if (ticket->t_cnt > 0)
2778 ticket->t_cnt--;
2779
2780 trace_xfs_log_ungrant_enter(log, ticket);
2781 trace_xfs_log_ungrant_sub(log, ticket);
2782
2783 /*
2784 * If this is a permanent reservation ticket, we may be able to free
2785 * up more space based on the remaining count.
2786 */
2787 bytes = ticket->t_curr_res;
2788 if (ticket->t_cnt > 0) {
2789 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2790 bytes += ticket->t_unit_res*ticket->t_cnt;
2791 }
2792
2793 xlog_grant_sub_space(log, &log->l_grant_reserve_head, bytes);
2794 xlog_grant_sub_space(log, &log->l_grant_write_head, bytes);
2795
2796 trace_xfs_log_ungrant_exit(log, ticket);
2797
2798 xfs_log_move_tail(log->l_mp, 1);
2799} /* xlog_ungrant_log_space */
2800
2801
2802/*
2803 * Flush iclog to disk if this is the last reference to the given iclog and
2804 * the WANT_SYNC bit is set.
2805 *
2806 * When this function is entered, the iclog is not necessarily in the
2807 * WANT_SYNC state. It may be sitting around waiting to get filled.
2808 *
2809 *
2810 */
2811STATIC int
2812xlog_state_release_iclog(
2813 xlog_t *log,
2814 xlog_in_core_t *iclog)
2815{
2816 int sync = 0; /* do we sync? */
2817
2818 if (iclog->ic_state & XLOG_STATE_IOERROR)
2819 return XFS_ERROR(EIO);
2820
2821 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2822 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2823 return 0;
2824
2825 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2826 spin_unlock(&log->l_icloglock);
2827 return XFS_ERROR(EIO);
2828 }
2829 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2830 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2831
2832 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2833 /* update tail before writing to iclog */
2834 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2835 sync++;
2836 iclog->ic_state = XLOG_STATE_SYNCING;
2837 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2838 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2839 /* cycle incremented when incrementing curr_block */
2840 }
2841 spin_unlock(&log->l_icloglock);
2842
2843 /*
2844 * We let the log lock go, so it's possible that we hit a log I/O
2845 * error or some other SHUTDOWN condition that marks the iclog
2846 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2847 * this iclog has consistent data, so we ignore IOERROR
2848 * flags after this point.
2849 */
2850 if (sync)
2851 return xlog_sync(log, iclog);
2852 return 0;
2853} /* xlog_state_release_iclog */
2854
2855
2856/*
2857 * This routine will mark the current iclog in the ring as WANT_SYNC
2858 * and move the current iclog pointer to the next iclog in the ring.
2859 * When this routine is called from xlog_state_get_iclog_space(), the
2860 * exact size of the iclog has not yet been determined. All we know is
2861 * that every data block. We have run out of space in this log record.
2862 */
2863STATIC void
2864xlog_state_switch_iclogs(xlog_t *log,
2865 xlog_in_core_t *iclog,
2866 int eventual_size)
2867{
2868 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2869 if (!eventual_size)
2870 eventual_size = iclog->ic_offset;
2871 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2872 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2873 log->l_prev_block = log->l_curr_block;
2874 log->l_prev_cycle = log->l_curr_cycle;
2875
2876 /* roll log?: ic_offset changed later */
2877 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2878
2879 /* Round up to next log-sunit */
2880 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2881 log->l_mp->m_sb.sb_logsunit > 1) {
2882 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2883 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2884 }
2885
2886 if (log->l_curr_block >= log->l_logBBsize) {
2887 log->l_curr_cycle++;
2888 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2889 log->l_curr_cycle++;
2890 log->l_curr_block -= log->l_logBBsize;
2891 ASSERT(log->l_curr_block >= 0);
2892 }
2893 ASSERT(iclog == log->l_iclog);
2894 log->l_iclog = iclog->ic_next;
2895} /* xlog_state_switch_iclogs */
2896
2897/*
2898 * Write out all data in the in-core log as of this exact moment in time.
2899 *
2900 * Data may be written to the in-core log during this call. However,
2901 * we don't guarantee this data will be written out. A change from past
2902 * implementation means this routine will *not* write out zero length LRs.
2903 *
2904 * Basically, we try and perform an intelligent scan of the in-core logs.
2905 * If we determine there is no flushable data, we just return. There is no
2906 * flushable data if:
2907 *
2908 * 1. the current iclog is active and has no data; the previous iclog
2909 * is in the active or dirty state.
2910 * 2. the current iclog is drity, and the previous iclog is in the
2911 * active or dirty state.
2912 *
2913 * We may sleep if:
2914 *
2915 * 1. the current iclog is not in the active nor dirty state.
2916 * 2. the current iclog dirty, and the previous iclog is not in the
2917 * active nor dirty state.
2918 * 3. the current iclog is active, and there is another thread writing
2919 * to this particular iclog.
2920 * 4. a) the current iclog is active and has no other writers
2921 * b) when we return from flushing out this iclog, it is still
2922 * not in the active nor dirty state.
2923 */
2924int
2925_xfs_log_force(
2926 struct xfs_mount *mp,
2927 uint flags,
2928 int *log_flushed)
2929{
2930 struct log *log = mp->m_log;
2931 struct xlog_in_core *iclog;
2932 xfs_lsn_t lsn;
2933
2934 XFS_STATS_INC(xs_log_force);
2935
2936 if (log->l_cilp)
2937 xlog_cil_force(log);
2938
2939 spin_lock(&log->l_icloglock);
2940
2941 iclog = log->l_iclog;
2942 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2943 spin_unlock(&log->l_icloglock);
2944 return XFS_ERROR(EIO);
2945 }
2946
2947 /* If the head iclog is not active nor dirty, we just attach
2948 * ourselves to the head and go to sleep.
2949 */
2950 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2951 iclog->ic_state == XLOG_STATE_DIRTY) {
2952 /*
2953 * If the head is dirty or (active and empty), then
2954 * we need to look at the previous iclog. If the previous
2955 * iclog is active or dirty we are done. There is nothing
2956 * to sync out. Otherwise, we attach ourselves to the
2957 * previous iclog and go to sleep.
2958 */
2959 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2960 (atomic_read(&iclog->ic_refcnt) == 0
2961 && iclog->ic_offset == 0)) {
2962 iclog = iclog->ic_prev;
2963 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2964 iclog->ic_state == XLOG_STATE_DIRTY)
2965 goto no_sleep;
2966 else
2967 goto maybe_sleep;
2968 } else {
2969 if (atomic_read(&iclog->ic_refcnt) == 0) {
2970 /* We are the only one with access to this
2971 * iclog. Flush it out now. There should
2972 * be a roundoff of zero to show that someone
2973 * has already taken care of the roundoff from
2974 * the previous sync.
2975 */
2976 atomic_inc(&iclog->ic_refcnt);
2977 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2978 xlog_state_switch_iclogs(log, iclog, 0);
2979 spin_unlock(&log->l_icloglock);
2980
2981 if (xlog_state_release_iclog(log, iclog))
2982 return XFS_ERROR(EIO);
2983
2984 if (log_flushed)
2985 *log_flushed = 1;
2986 spin_lock(&log->l_icloglock);
2987 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2988 iclog->ic_state != XLOG_STATE_DIRTY)
2989 goto maybe_sleep;
2990 else
2991 goto no_sleep;
2992 } else {
2993 /* Someone else is writing to this iclog.
2994 * Use its call to flush out the data. However,
2995 * the other thread may not force out this LR,
2996 * so we mark it WANT_SYNC.
2997 */
2998 xlog_state_switch_iclogs(log, iclog, 0);
2999 goto maybe_sleep;
3000 }
3001 }
3002 }
3003
3004 /* By the time we come around again, the iclog could've been filled
3005 * which would give it another lsn. If we have a new lsn, just
3006 * return because the relevant data has been flushed.
3007 */
3008maybe_sleep:
3009 if (flags & XFS_LOG_SYNC) {
3010 /*
3011 * We must check if we're shutting down here, before
3012 * we wait, while we're holding the l_icloglock.
3013 * Then we check again after waking up, in case our
3014 * sleep was disturbed by a bad news.
3015 */
3016 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3017 spin_unlock(&log->l_icloglock);
3018 return XFS_ERROR(EIO);
3019 }
3020 XFS_STATS_INC(xs_log_force_sleep);
3021 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3022 /*
3023 * No need to grab the log lock here since we're
3024 * only deciding whether or not to return EIO
3025 * and the memory read should be atomic.
3026 */
3027 if (iclog->ic_state & XLOG_STATE_IOERROR)
3028 return XFS_ERROR(EIO);
3029 if (log_flushed)
3030 *log_flushed = 1;
3031 } else {
3032
3033no_sleep:
3034 spin_unlock(&log->l_icloglock);
3035 }
3036 return 0;
3037}
3038
3039/*
3040 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3041 * about errors or whether the log was flushed or not. This is the normal
3042 * interface to use when trying to unpin items or move the log forward.
3043 */
3044void
3045xfs_log_force(
3046 xfs_mount_t *mp,
3047 uint flags)
3048{
3049 int error;
3050
3051 error = _xfs_log_force(mp, flags, NULL);
3052 if (error)
3053 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3054}
3055
3056/*
3057 * Force the in-core log to disk for a specific LSN.
3058 *
3059 * Find in-core log with lsn.
3060 * If it is in the DIRTY state, just return.
3061 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3062 * state and go to sleep or return.
3063 * If it is in any other state, go to sleep or return.
3064 *
3065 * Synchronous forces are implemented with a signal variable. All callers
3066 * to force a given lsn to disk will wait on a the sv attached to the
3067 * specific in-core log. When given in-core log finally completes its
3068 * write to disk, that thread will wake up all threads waiting on the
3069 * sv.
3070 */
3071int
3072_xfs_log_force_lsn(
3073 struct xfs_mount *mp,
3074 xfs_lsn_t lsn,
3075 uint flags,
3076 int *log_flushed)
3077{
3078 struct log *log = mp->m_log;
3079 struct xlog_in_core *iclog;
3080 int already_slept = 0;
3081
3082 ASSERT(lsn != 0);
3083
3084 XFS_STATS_INC(xs_log_force);
3085
3086 if (log->l_cilp) {
3087 lsn = xlog_cil_force_lsn(log, lsn);
3088 if (lsn == NULLCOMMITLSN)
3089 return 0;
3090 }
3091
3092try_again:
3093 spin_lock(&log->l_icloglock);
3094 iclog = log->l_iclog;
3095 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3096 spin_unlock(&log->l_icloglock);
3097 return XFS_ERROR(EIO);
3098 }
3099
3100 do {
3101 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3102 iclog = iclog->ic_next;
3103 continue;
3104 }
3105
3106 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3107 spin_unlock(&log->l_icloglock);
3108 return 0;
3109 }
3110
3111 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3112 /*
3113 * We sleep here if we haven't already slept (e.g.
3114 * this is the first time we've looked at the correct
3115 * iclog buf) and the buffer before us is going to
3116 * be sync'ed. The reason for this is that if we
3117 * are doing sync transactions here, by waiting for
3118 * the previous I/O to complete, we can allow a few
3119 * more transactions into this iclog before we close
3120 * it down.
3121 *
3122 * Otherwise, we mark the buffer WANT_SYNC, and bump
3123 * up the refcnt so we can release the log (which
3124 * drops the ref count). The state switch keeps new
3125 * transaction commits from using this buffer. When
3126 * the current commits finish writing into the buffer,
3127 * the refcount will drop to zero and the buffer will
3128 * go out then.
3129 */
3130 if (!already_slept &&
3131 (iclog->ic_prev->ic_state &
3132 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3133 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3134
3135 XFS_STATS_INC(xs_log_force_sleep);
3136
3137 xlog_wait(&iclog->ic_prev->ic_write_wait,
3138 &log->l_icloglock);
3139 if (log_flushed)
3140 *log_flushed = 1;
3141 already_slept = 1;
3142 goto try_again;
3143 }
3144 atomic_inc(&iclog->ic_refcnt);
3145 xlog_state_switch_iclogs(log, iclog, 0);
3146 spin_unlock(&log->l_icloglock);
3147 if (xlog_state_release_iclog(log, iclog))
3148 return XFS_ERROR(EIO);
3149 if (log_flushed)
3150 *log_flushed = 1;
3151 spin_lock(&log->l_icloglock);
3152 }
3153
3154 if ((flags & XFS_LOG_SYNC) && /* sleep */
3155 !(iclog->ic_state &
3156 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3157 /*
3158 * Don't wait on completion if we know that we've
3159 * gotten a log write error.
3160 */
3161 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3162 spin_unlock(&log->l_icloglock);
3163 return XFS_ERROR(EIO);
3164 }
3165 XFS_STATS_INC(xs_log_force_sleep);
3166 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3167 /*
3168 * No need to grab the log lock here since we're
3169 * only deciding whether or not to return EIO
3170 * and the memory read should be atomic.
3171 */
3172 if (iclog->ic_state & XLOG_STATE_IOERROR)
3173 return XFS_ERROR(EIO);
3174
3175 if (log_flushed)
3176 *log_flushed = 1;
3177 } else { /* just return */
3178 spin_unlock(&log->l_icloglock);
3179 }
3180
3181 return 0;
3182 } while (iclog != log->l_iclog);
3183
3184 spin_unlock(&log->l_icloglock);
3185 return 0;
3186}
3187
3188/*
3189 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3190 * about errors or whether the log was flushed or not. This is the normal
3191 * interface to use when trying to unpin items or move the log forward.
3192 */
3193void
3194xfs_log_force_lsn(
3195 xfs_mount_t *mp,
3196 xfs_lsn_t lsn,
3197 uint flags)
3198{
3199 int error;
3200
3201 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3202 if (error)
3203 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3204}
3205
3206/*
3207 * Called when we want to mark the current iclog as being ready to sync to
3208 * disk.
3209 */
3210STATIC void
3211xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3212{
3213 assert_spin_locked(&log->l_icloglock);
3214
3215 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3216 xlog_state_switch_iclogs(log, iclog, 0);
3217 } else {
3218 ASSERT(iclog->ic_state &
3219 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3220 }
3221}
3222
3223
3224/*****************************************************************************
3225 *
3226 * TICKET functions
3227 *
3228 *****************************************************************************
3229 */
3230
3231/*
3232 * Free a used ticket when its refcount falls to zero.
3233 */
3234void
3235xfs_log_ticket_put(
3236 xlog_ticket_t *ticket)
3237{
3238 ASSERT(atomic_read(&ticket->t_ref) > 0);
3239 if (atomic_dec_and_test(&ticket->t_ref))
3240 kmem_zone_free(xfs_log_ticket_zone, ticket);
3241}
3242
3243xlog_ticket_t *
3244xfs_log_ticket_get(
3245 xlog_ticket_t *ticket)
3246{
3247 ASSERT(atomic_read(&ticket->t_ref) > 0);
3248 atomic_inc(&ticket->t_ref);
3249 return ticket;
3250}
3251
3252/*
3253 * Allocate and initialise a new log ticket.
3254 */
3255xlog_ticket_t *
3256xlog_ticket_alloc(
3257 struct log *log,
3258 int unit_bytes,
3259 int cnt,
3260 char client,
3261 uint xflags,
3262 int alloc_flags)
3263{
3264 struct xlog_ticket *tic;
3265 uint num_headers;
3266 int iclog_space;
3267
3268 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3269 if (!tic)
3270 return NULL;
3271
3272 /*
3273 * Permanent reservations have up to 'cnt'-1 active log operations
3274 * in the log. A unit in this case is the amount of space for one
3275 * of these log operations. Normal reservations have a cnt of 1
3276 * and their unit amount is the total amount of space required.
3277 *
3278 * The following lines of code account for non-transaction data
3279 * which occupy space in the on-disk log.
3280 *
3281 * Normal form of a transaction is:
3282 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3283 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3284 *
3285 * We need to account for all the leadup data and trailer data
3286 * around the transaction data.
3287 * And then we need to account for the worst case in terms of using
3288 * more space.
3289 * The worst case will happen if:
3290 * - the placement of the transaction happens to be such that the
3291 * roundoff is at its maximum
3292 * - the transaction data is synced before the commit record is synced
3293 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3294 * Therefore the commit record is in its own Log Record.
3295 * This can happen as the commit record is called with its
3296 * own region to xlog_write().
3297 * This then means that in the worst case, roundoff can happen for
3298 * the commit-rec as well.
3299 * The commit-rec is smaller than padding in this scenario and so it is
3300 * not added separately.
3301 */
3302
3303 /* for trans header */
3304 unit_bytes += sizeof(xlog_op_header_t);
3305 unit_bytes += sizeof(xfs_trans_header_t);
3306
3307 /* for start-rec */
3308 unit_bytes += sizeof(xlog_op_header_t);
3309
3310 /*
3311 * for LR headers - the space for data in an iclog is the size minus
3312 * the space used for the headers. If we use the iclog size, then we
3313 * undercalculate the number of headers required.
3314 *
3315 * Furthermore - the addition of op headers for split-recs might
3316 * increase the space required enough to require more log and op
3317 * headers, so take that into account too.
3318 *
3319 * IMPORTANT: This reservation makes the assumption that if this
3320 * transaction is the first in an iclog and hence has the LR headers
3321 * accounted to it, then the remaining space in the iclog is
3322 * exclusively for this transaction. i.e. if the transaction is larger
3323 * than the iclog, it will be the only thing in that iclog.
3324 * Fundamentally, this means we must pass the entire log vector to
3325 * xlog_write to guarantee this.
3326 */
3327 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3328 num_headers = howmany(unit_bytes, iclog_space);
3329
3330 /* for split-recs - ophdrs added when data split over LRs */
3331 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3332
3333 /* add extra header reservations if we overrun */
3334 while (!num_headers ||
3335 howmany(unit_bytes, iclog_space) > num_headers) {
3336 unit_bytes += sizeof(xlog_op_header_t);
3337 num_headers++;
3338 }
3339 unit_bytes += log->l_iclog_hsize * num_headers;
3340
3341 /* for commit-rec LR header - note: padding will subsume the ophdr */
3342 unit_bytes += log->l_iclog_hsize;
3343
3344 /* for roundoff padding for transaction data and one for commit record */
3345 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3346 log->l_mp->m_sb.sb_logsunit > 1) {
3347 /* log su roundoff */
3348 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3349 } else {
3350 /* BB roundoff */
3351 unit_bytes += 2*BBSIZE;
3352 }
3353
3354 atomic_set(&tic->t_ref, 1);
3355 INIT_LIST_HEAD(&tic->t_queue);
3356 tic->t_unit_res = unit_bytes;
3357 tic->t_curr_res = unit_bytes;
3358 tic->t_cnt = cnt;
3359 tic->t_ocnt = cnt;
3360 tic->t_tid = random32();
3361 tic->t_clientid = client;
3362 tic->t_flags = XLOG_TIC_INITED;
3363 tic->t_trans_type = 0;
3364 if (xflags & XFS_LOG_PERM_RESERV)
3365 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3366 init_waitqueue_head(&tic->t_wait);
3367
3368 xlog_tic_reset_res(tic);
3369
3370 return tic;
3371}
3372
3373
3374/******************************************************************************
3375 *
3376 * Log debug routines
3377 *
3378 ******************************************************************************
3379 */
3380#if defined(DEBUG)
3381/*
3382 * Make sure that the destination ptr is within the valid data region of
3383 * one of the iclogs. This uses backup pointers stored in a different
3384 * part of the log in case we trash the log structure.
3385 */
3386void
3387xlog_verify_dest_ptr(
3388 struct log *log,
3389 char *ptr)
3390{
3391 int i;
3392 int good_ptr = 0;
3393
3394 for (i = 0; i < log->l_iclog_bufs; i++) {
3395 if (ptr >= log->l_iclog_bak[i] &&
3396 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3397 good_ptr++;
3398 }
3399
3400 if (!good_ptr)
3401 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3402}
3403
3404/*
3405 * Check to make sure the grant write head didn't just over lap the tail. If
3406 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3407 * the cycles differ by exactly one and check the byte count.
3408 *
3409 * This check is run unlocked, so can give false positives. Rather than assert
3410 * on failures, use a warn-once flag and a panic tag to allow the admin to
3411 * determine if they want to panic the machine when such an error occurs. For
3412 * debug kernels this will have the same effect as using an assert but, unlinke
3413 * an assert, it can be turned off at runtime.
3414 */
3415STATIC void
3416xlog_verify_grant_tail(
3417 struct log *log)
3418{
3419 int tail_cycle, tail_blocks;
3420 int cycle, space;
3421
3422 xlog_crack_grant_head(&log->l_grant_write_head, &cycle, &space);
3423 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3424 if (tail_cycle != cycle) {
3425 if (cycle - 1 != tail_cycle &&
3426 !(log->l_flags & XLOG_TAIL_WARN)) {
3427 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3428 "%s: cycle - 1 != tail_cycle", __func__);
3429 log->l_flags |= XLOG_TAIL_WARN;
3430 }
3431
3432 if (space > BBTOB(tail_blocks) &&
3433 !(log->l_flags & XLOG_TAIL_WARN)) {
3434 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3435 "%s: space > BBTOB(tail_blocks)", __func__);
3436 log->l_flags |= XLOG_TAIL_WARN;
3437 }
3438 }
3439}
3440
3441/* check if it will fit */
3442STATIC void
3443xlog_verify_tail_lsn(xlog_t *log,
3444 xlog_in_core_t *iclog,
3445 xfs_lsn_t tail_lsn)
3446{
3447 int blocks;
3448
3449 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3450 blocks =
3451 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3452 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3453 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3454 } else {
3455 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3456
3457 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3458 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3459
3460 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3461 if (blocks < BTOBB(iclog->ic_offset) + 1)
3462 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3463 }
3464} /* xlog_verify_tail_lsn */
3465
3466/*
3467 * Perform a number of checks on the iclog before writing to disk.
3468 *
3469 * 1. Make sure the iclogs are still circular
3470 * 2. Make sure we have a good magic number
3471 * 3. Make sure we don't have magic numbers in the data
3472 * 4. Check fields of each log operation header for:
3473 * A. Valid client identifier
3474 * B. tid ptr value falls in valid ptr space (user space code)
3475 * C. Length in log record header is correct according to the
3476 * individual operation headers within record.
3477 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3478 * log, check the preceding blocks of the physical log to make sure all
3479 * the cycle numbers agree with the current cycle number.
3480 */
3481STATIC void
3482xlog_verify_iclog(xlog_t *log,
3483 xlog_in_core_t *iclog,
3484 int count,
3485 boolean_t syncing)
3486{
3487 xlog_op_header_t *ophead;
3488 xlog_in_core_t *icptr;
3489 xlog_in_core_2_t *xhdr;
3490 xfs_caddr_t ptr;
3491 xfs_caddr_t base_ptr;
3492 __psint_t field_offset;
3493 __uint8_t clientid;
3494 int len, i, j, k, op_len;
3495 int idx;
3496
3497 /* check validity of iclog pointers */
3498 spin_lock(&log->l_icloglock);
3499 icptr = log->l_iclog;
3500 for (i=0; i < log->l_iclog_bufs; i++) {
3501 if (icptr == NULL)
3502 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3503 icptr = icptr->ic_next;
3504 }
3505 if (icptr != log->l_iclog)
3506 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3507 spin_unlock(&log->l_icloglock);
3508
3509 /* check log magic numbers */
3510 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3511 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3512
3513 ptr = (xfs_caddr_t) &iclog->ic_header;
3514 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3515 ptr += BBSIZE) {
3516 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3517 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3518 __func__);
3519 }
3520
3521 /* check fields */
3522 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3523 ptr = iclog->ic_datap;
3524 base_ptr = ptr;
3525 ophead = (xlog_op_header_t *)ptr;
3526 xhdr = iclog->ic_data;
3527 for (i = 0; i < len; i++) {
3528 ophead = (xlog_op_header_t *)ptr;
3529
3530 /* clientid is only 1 byte */
3531 field_offset = (__psint_t)
3532 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3533 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3534 clientid = ophead->oh_clientid;
3535 } else {
3536 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3537 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3538 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3539 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3540 clientid = xlog_get_client_id(
3541 xhdr[j].hic_xheader.xh_cycle_data[k]);
3542 } else {
3543 clientid = xlog_get_client_id(
3544 iclog->ic_header.h_cycle_data[idx]);
3545 }
3546 }
3547 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3548 xfs_warn(log->l_mp,
3549 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3550 __func__, clientid, ophead,
3551 (unsigned long)field_offset);
3552
3553 /* check length */
3554 field_offset = (__psint_t)
3555 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3556 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3557 op_len = be32_to_cpu(ophead->oh_len);
3558 } else {
3559 idx = BTOBBT((__psint_t)&ophead->oh_len -
3560 (__psint_t)iclog->ic_datap);
3561 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3562 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3563 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3564 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3565 } else {
3566 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3567 }
3568 }
3569 ptr += sizeof(xlog_op_header_t) + op_len;
3570 }
3571} /* xlog_verify_iclog */
3572#endif
3573
3574/*
3575 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3576 */
3577STATIC int
3578xlog_state_ioerror(
3579 xlog_t *log)
3580{
3581 xlog_in_core_t *iclog, *ic;
3582
3583 iclog = log->l_iclog;
3584 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3585 /*
3586 * Mark all the incore logs IOERROR.
3587 * From now on, no log flushes will result.
3588 */
3589 ic = iclog;
3590 do {
3591 ic->ic_state = XLOG_STATE_IOERROR;
3592 ic = ic->ic_next;
3593 } while (ic != iclog);
3594 return 0;
3595 }
3596 /*
3597 * Return non-zero, if state transition has already happened.
3598 */
3599 return 1;
3600}
3601
3602/*
3603 * This is called from xfs_force_shutdown, when we're forcibly
3604 * shutting down the filesystem, typically because of an IO error.
3605 * Our main objectives here are to make sure that:
3606 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3607 * parties to find out, 'atomically'.
3608 * b. those who're sleeping on log reservations, pinned objects and
3609 * other resources get woken up, and be told the bad news.
3610 * c. nothing new gets queued up after (a) and (b) are done.
3611 * d. if !logerror, flush the iclogs to disk, then seal them off
3612 * for business.
3613 *
3614 * Note: for delayed logging the !logerror case needs to flush the regions
3615 * held in memory out to the iclogs before flushing them to disk. This needs
3616 * to be done before the log is marked as shutdown, otherwise the flush to the
3617 * iclogs will fail.
3618 */
3619int
3620xfs_log_force_umount(
3621 struct xfs_mount *mp,
3622 int logerror)
3623{
3624 xlog_ticket_t *tic;
3625 xlog_t *log;
3626 int retval;
3627
3628 log = mp->m_log;
3629
3630 /*
3631 * If this happens during log recovery, don't worry about
3632 * locking; the log isn't open for business yet.
3633 */
3634 if (!log ||
3635 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3636 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3637 if (mp->m_sb_bp)
3638 XFS_BUF_DONE(mp->m_sb_bp);
3639 return 0;
3640 }
3641
3642 /*
3643 * Somebody could've already done the hard work for us.
3644 * No need to get locks for this.
3645 */
3646 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3647 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3648 return 1;
3649 }
3650 retval = 0;
3651
3652 /*
3653 * Flush the in memory commit item list before marking the log as
3654 * being shut down. We need to do it in this order to ensure all the
3655 * completed transactions are flushed to disk with the xfs_log_force()
3656 * call below.
3657 */
3658 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3659 xlog_cil_force(log);
3660
3661 /*
3662 * mark the filesystem and the as in a shutdown state and wake
3663 * everybody up to tell them the bad news.
3664 */
3665 spin_lock(&log->l_icloglock);
3666 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3667 if (mp->m_sb_bp)
3668 XFS_BUF_DONE(mp->m_sb_bp);
3669
3670 /*
3671 * This flag is sort of redundant because of the mount flag, but
3672 * it's good to maintain the separation between the log and the rest
3673 * of XFS.
3674 */
3675 log->l_flags |= XLOG_IO_ERROR;
3676
3677 /*
3678 * If we hit a log error, we want to mark all the iclogs IOERROR
3679 * while we're still holding the loglock.
3680 */
3681 if (logerror)
3682 retval = xlog_state_ioerror(log);
3683 spin_unlock(&log->l_icloglock);
3684
3685 /*
3686 * We don't want anybody waiting for log reservations after this. That
3687 * means we have to wake up everybody queued up on reserveq as well as
3688 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3689 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3690 * action is protected by the grant locks.
3691 */
3692 spin_lock(&log->l_grant_reserve_lock);
3693 list_for_each_entry(tic, &log->l_reserveq, t_queue)
3694 wake_up(&tic->t_wait);
3695 spin_unlock(&log->l_grant_reserve_lock);
3696
3697 spin_lock(&log->l_grant_write_lock);
3698 list_for_each_entry(tic, &log->l_writeq, t_queue)
3699 wake_up(&tic->t_wait);
3700 spin_unlock(&log->l_grant_write_lock);
3701
3702 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3703 ASSERT(!logerror);
3704 /*
3705 * Force the incore logs to disk before shutting the
3706 * log down completely.
3707 */
3708 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3709
3710 spin_lock(&log->l_icloglock);
3711 retval = xlog_state_ioerror(log);
3712 spin_unlock(&log->l_icloglock);
3713 }
3714 /*
3715 * Wake up everybody waiting on xfs_log_force.
3716 * Callback all log item committed functions as if the
3717 * log writes were completed.
3718 */
3719 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3720
3721#ifdef XFSERRORDEBUG
3722 {
3723 xlog_in_core_t *iclog;
3724
3725 spin_lock(&log->l_icloglock);
3726 iclog = log->l_iclog;
3727 do {
3728 ASSERT(iclog->ic_callback == 0);
3729 iclog = iclog->ic_next;
3730 } while (iclog != log->l_iclog);
3731 spin_unlock(&log->l_icloglock);
3732 }
3733#endif
3734 /* return non-zero if log IOERROR transition had already happened */
3735 return retval;
3736}
3737
3738STATIC int
3739xlog_iclogs_empty(xlog_t *log)
3740{
3741 xlog_in_core_t *iclog;
3742
3743 iclog = log->l_iclog;
3744 do {
3745 /* endianness does not matter here, zero is zero in
3746 * any language.
3747 */
3748 if (iclog->ic_header.h_num_logops)
3749 return 0;
3750 iclog = iclog->ic_next;
3751 } while (iclog != log->l_iclog);
3752 return 1;
3753}