blob: 4dad756962d02e1b29c703b5b54ccee07d91fabf [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Nathan Scott7b718762005-11-02 14:58:39 +11002 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
Nathan Scott7b718762005-11-02 14:58:39 +11005 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
Linus Torvalds1da177e2005-04-16 15:20:36 -07007 * published by the Free Software Foundation.
8 *
Nathan Scott7b718762005-11-02 14:58:39 +11009 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070013 *
Nathan Scott7b718762005-11-02 14:58:39 +110014 * 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
Linus Torvalds1da177e2005-04-16 15:20:36 -070017 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include "xfs.h"
Nathan Scotta844f452005-11-02 14:38:42 +110019#include "xfs_fs.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include "xfs_types.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include "xfs_log.h"
22#include "xfs_trans.h"
Nathan Scotta844f452005-11-02 14:38:42 +110023#include "xfs_sb.h"
24#include "xfs_ag.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include "xfs_mount.h"
26#include "xfs_error.h"
27#include "xfs_log_priv.h"
28#include "xfs_buf_item.h"
Nathan Scotta844f452005-11-02 14:38:42 +110029#include "xfs_bmap_btree.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include "xfs_alloc_btree.h"
Nathan Scotta844f452005-11-02 14:38:42 +110031#include "xfs_ialloc_btree.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include "xfs_log_recover.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070033#include "xfs_trans_priv.h"
Nathan Scotta844f452005-11-02 14:38:42 +110034#include "xfs_dinode.h"
35#include "xfs_inode.h"
Christoph Hellwig0b1b2132009-12-14 23:14:59 +000036#include "xfs_trace.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
David Chinnereb01c9c2008-04-10 12:18:46 +100038kmem_zone_t *xfs_log_ticket_zone;
Linus Torvalds1da177e2005-04-16 15:20:36 -070039
Linus Torvalds1da177e2005-04-16 15:20:36 -070040/* Local miscellaneous function prototypes */
Mark Tinguelyad223e62012-06-14 09:22:15 -050041STATIC int
42xlog_commit_record(
43 struct xlog *log,
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
47
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -050048STATIC struct xlog *
49xlog_alloc_log(
50 struct xfs_mount *mp,
51 struct xfs_buftarg *log_target,
52 xfs_daddr_t blk_offset,
53 int num_bblks);
Mark Tinguelyad223e62012-06-14 09:22:15 -050054STATIC int
55xlog_space_left(
56 struct xlog *log,
57 atomic64_t *head);
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -050058STATIC int
59xlog_sync(
60 struct xlog *log,
61 struct xlog_in_core *iclog);
62STATIC void
63xlog_dealloc_log(
64 struct xlog *log);
Linus Torvalds1da177e2005-04-16 15:20:36 -070065
66/* local state machine functions */
67STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -050068STATIC void
69xlog_state_do_callback(
70 struct xlog *log,
71 int aborted,
72 struct xlog_in_core *iclog);
73STATIC int
74xlog_state_get_iclog_space(
75 struct xlog *log,
76 int len,
77 struct xlog_in_core **iclog,
78 struct xlog_ticket *ticket,
79 int *continued_write,
80 int *logoffsetp);
81STATIC int
82xlog_state_release_iclog(
83 struct xlog *log,
84 struct xlog_in_core *iclog);
85STATIC void
86xlog_state_switch_iclogs(
87 struct xlog *log,
88 struct xlog_in_core *iclog,
89 int eventual_size);
90STATIC void
91xlog_state_want_sync(
92 struct xlog *log,
93 struct xlog_in_core *iclog);
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
Mark Tinguelyad223e62012-06-14 09:22:15 -050095STATIC void
96xlog_grant_push_ail(
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -050097 struct xlog *log,
98 int need_bytes);
99STATIC void
100xlog_regrant_reserve_log_space(
101 struct xlog *log,
102 struct xlog_ticket *ticket);
103STATIC void
104xlog_ungrant_log_space(
105 struct xlog *log,
106 struct xlog_ticket *ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
Nathan Scottcfcbbbd2005-11-02 15:12:04 +1100108#if defined(DEBUG)
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -0500109STATIC void
110xlog_verify_dest_ptr(
111 struct xlog *log,
112 char *ptr);
Mark Tinguelyad223e62012-06-14 09:22:15 -0500113STATIC void
114xlog_verify_grant_tail(
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -0500115 struct xlog *log);
116STATIC void
117xlog_verify_iclog(
118 struct xlog *log,
119 struct xlog_in_core *iclog,
120 int count,
121 boolean_t syncing);
122STATIC void
123xlog_verify_tail_lsn(
124 struct xlog *log,
125 struct xlog_in_core *iclog,
126 xfs_lsn_t tail_lsn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127#else
128#define xlog_verify_dest_ptr(a,b)
Dave Chinner3f336c62010-12-21 12:02:52 +1100129#define xlog_verify_grant_tail(a)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130#define xlog_verify_iclog(a,b,c,d)
131#define xlog_verify_tail_lsn(a,b,c)
132#endif
133
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -0500134STATIC int
135xlog_iclogs_empty(
136 struct xlog *log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137
Christoph Hellwigdd954c62006-01-11 15:34:50 +1100138static void
Dave Chinner663e4962010-12-21 12:06:05 +1100139xlog_grant_sub_space(
Mark Tinguelyad223e62012-06-14 09:22:15 -0500140 struct xlog *log,
141 atomic64_t *head,
142 int bytes)
Christoph Hellwigdd954c62006-01-11 15:34:50 +1100143{
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100144 int64_t head_val = atomic64_read(head);
145 int64_t new, old;
Dave Chinnera69ed032010-12-21 12:08:20 +1100146
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100147 do {
148 int cycle, space;
Dave Chinnera69ed032010-12-21 12:08:20 +1100149
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100150 xlog_crack_grant_head_val(head_val, &cycle, &space);
Dave Chinnera69ed032010-12-21 12:08:20 +1100151
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100152 space -= bytes;
153 if (space < 0) {
154 space += log->l_logsize;
155 cycle--;
156 }
157
158 old = head_val;
159 new = xlog_assign_grant_head_val(cycle, space);
160 head_val = atomic64_cmpxchg(head, old, new);
161 } while (head_val != old);
Christoph Hellwigdd954c62006-01-11 15:34:50 +1100162}
163
164static void
Dave Chinner663e4962010-12-21 12:06:05 +1100165xlog_grant_add_space(
Mark Tinguelyad223e62012-06-14 09:22:15 -0500166 struct xlog *log,
167 atomic64_t *head,
168 int bytes)
Christoph Hellwigdd954c62006-01-11 15:34:50 +1100169{
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100170 int64_t head_val = atomic64_read(head);
171 int64_t new, old;
Dave Chinnera69ed032010-12-21 12:08:20 +1100172
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100173 do {
174 int tmp;
175 int cycle, space;
Dave Chinnera69ed032010-12-21 12:08:20 +1100176
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100177 xlog_crack_grant_head_val(head_val, &cycle, &space);
Dave Chinnera69ed032010-12-21 12:08:20 +1100178
Dave Chinnerd0eb2f32010-12-21 12:29:14 +1100179 tmp = log->l_logsize - space;
180 if (tmp > bytes)
181 space += bytes;
182 else {
183 space = bytes - tmp;
184 cycle++;
185 }
186
187 old = head_val;
188 new = xlog_assign_grant_head_val(cycle, space);
189 head_val = atomic64_cmpxchg(head, old, new);
190 } while (head_val != old);
Christoph Hellwigdd954c62006-01-11 15:34:50 +1100191}
Dave Chinnera69ed032010-12-21 12:08:20 +1100192
Christoph Hellwigc303c5b2012-02-20 02:31:26 +0000193STATIC void
194xlog_grant_head_init(
195 struct xlog_grant_head *head)
196{
197 xlog_assign_grant_head(&head->grant, 1, 0);
198 INIT_LIST_HEAD(&head->waiters);
199 spin_lock_init(&head->lock);
200}
201
Christoph Hellwiga79bf2d2012-02-20 02:31:27 +0000202STATIC void
203xlog_grant_head_wake_all(
204 struct xlog_grant_head *head)
205{
206 struct xlog_ticket *tic;
207
208 spin_lock(&head->lock);
209 list_for_each_entry(tic, &head->waiters, t_queue)
210 wake_up_process(tic->t_task);
211 spin_unlock(&head->lock);
212}
213
Christoph Hellwige179840d2012-02-20 02:31:29 +0000214static inline int
215xlog_ticket_reservation(
Mark Tinguelyad223e62012-06-14 09:22:15 -0500216 struct xlog *log,
Christoph Hellwige179840d2012-02-20 02:31:29 +0000217 struct xlog_grant_head *head,
218 struct xlog_ticket *tic)
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000219{
Christoph Hellwige179840d2012-02-20 02:31:29 +0000220 if (head == &log->l_write_head) {
221 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
222 return tic->t_unit_res;
223 } else {
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000224 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
Christoph Hellwige179840d2012-02-20 02:31:29 +0000225 return tic->t_unit_res * tic->t_cnt;
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000226 else
Christoph Hellwige179840d2012-02-20 02:31:29 +0000227 return tic->t_unit_res;
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000228 }
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000229}
230
231STATIC bool
Christoph Hellwige179840d2012-02-20 02:31:29 +0000232xlog_grant_head_wake(
Mark Tinguelyad223e62012-06-14 09:22:15 -0500233 struct xlog *log,
Christoph Hellwige179840d2012-02-20 02:31:29 +0000234 struct xlog_grant_head *head,
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000235 int *free_bytes)
236{
237 struct xlog_ticket *tic;
238 int need_bytes;
239
Christoph Hellwige179840d2012-02-20 02:31:29 +0000240 list_for_each_entry(tic, &head->waiters, t_queue) {
241 need_bytes = xlog_ticket_reservation(log, head, tic);
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000242 if (*free_bytes < need_bytes)
243 return false;
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000244
Christoph Hellwige179840d2012-02-20 02:31:29 +0000245 *free_bytes -= need_bytes;
246 trace_xfs_log_grant_wake_up(log, tic);
Christoph Hellwig14a72352012-02-20 02:31:24 +0000247 wake_up_process(tic->t_task);
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000248 }
249
250 return true;
251}
252
253STATIC int
Christoph Hellwig23ee3df2012-02-20 02:31:28 +0000254xlog_grant_head_wait(
Mark Tinguelyad223e62012-06-14 09:22:15 -0500255 struct xlog *log,
Christoph Hellwig23ee3df2012-02-20 02:31:28 +0000256 struct xlog_grant_head *head,
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000257 struct xlog_ticket *tic,
258 int need_bytes)
259{
Christoph Hellwig23ee3df2012-02-20 02:31:28 +0000260 list_add_tail(&tic->t_queue, &head->waiters);
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000261
262 do {
263 if (XLOG_FORCED_SHUTDOWN(log))
264 goto shutdown;
265 xlog_grant_push_ail(log, need_bytes);
266
Christoph Hellwig14a72352012-02-20 02:31:24 +0000267 __set_current_state(TASK_UNINTERRUPTIBLE);
Christoph Hellwig23ee3df2012-02-20 02:31:28 +0000268 spin_unlock(&head->lock);
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000269
Christoph Hellwig14a72352012-02-20 02:31:24 +0000270 XFS_STATS_INC(xs_sleep_logspace);
271
272 trace_xfs_log_grant_sleep(log, tic);
273 schedule();
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000274 trace_xfs_log_grant_wake(log, tic);
275
Christoph Hellwig23ee3df2012-02-20 02:31:28 +0000276 spin_lock(&head->lock);
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000277 if (XLOG_FORCED_SHUTDOWN(log))
278 goto shutdown;
Christoph Hellwig23ee3df2012-02-20 02:31:28 +0000279 } while (xlog_space_left(log, &head->grant) < need_bytes);
Christoph Hellwig9f9c19e2011-11-28 08:17:36 +0000280
281 list_del_init(&tic->t_queue);
282 return 0;
283shutdown:
284 list_del_init(&tic->t_queue);
285 return XFS_ERROR(EIO);
286}
287
Christoph Hellwig42ceedb2012-02-20 02:31:30 +0000288/*
289 * Atomically get the log space required for a log ticket.
290 *
291 * Once a ticket gets put onto head->waiters, it will only return after the
292 * needed reservation is satisfied.
293 *
294 * This function is structured so that it has a lock free fast path. This is
295 * necessary because every new transaction reservation will come through this
296 * path. Hence any lock will be globally hot if we take it unconditionally on
297 * every pass.
298 *
299 * As tickets are only ever moved on and off head->waiters under head->lock, we
300 * only need to take that lock if we are going to add the ticket to the queue
301 * and sleep. We can avoid taking the lock if the ticket was never added to
302 * head->waiters because the t_queue list head will be empty and we hold the
303 * only reference to it so it can safely be checked unlocked.
304 */
305STATIC int
306xlog_grant_head_check(
Mark Tinguelyad223e62012-06-14 09:22:15 -0500307 struct xlog *log,
Christoph Hellwig42ceedb2012-02-20 02:31:30 +0000308 struct xlog_grant_head *head,
309 struct xlog_ticket *tic,
310 int *need_bytes)
311{
312 int free_bytes;
313 int error = 0;
314
315 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
316
317 /*
318 * If there are other waiters on the queue then give them a chance at
319 * logspace before us. Wake up the first waiters, if we do not wake
320 * up all the waiters then go to sleep waiting for more free space,
321 * otherwise try to get some space for this transaction.
322 */
323 *need_bytes = xlog_ticket_reservation(log, head, tic);
324 free_bytes = xlog_space_left(log, &head->grant);
325 if (!list_empty_careful(&head->waiters)) {
326 spin_lock(&head->lock);
327 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
328 free_bytes < *need_bytes) {
329 error = xlog_grant_head_wait(log, head, tic,
330 *need_bytes);
331 }
332 spin_unlock(&head->lock);
333 } else if (free_bytes < *need_bytes) {
334 spin_lock(&head->lock);
335 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
336 spin_unlock(&head->lock);
337 }
338
339 return error;
340}
341
Christoph Hellwig0adba532007-08-30 17:21:46 +1000342static void
343xlog_tic_reset_res(xlog_ticket_t *tic)
344{
345 tic->t_res_num = 0;
346 tic->t_res_arr_sum = 0;
347 tic->t_res_num_ophdrs = 0;
348}
349
350static void
351xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
352{
353 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
354 /* add to overflow and start again */
355 tic->t_res_o_flow += tic->t_res_arr_sum;
356 tic->t_res_num = 0;
357 tic->t_res_arr_sum = 0;
358 }
359
360 tic->t_res_arr[tic->t_res_num].r_len = len;
361 tic->t_res_arr[tic->t_res_num].r_type = type;
362 tic->t_res_arr_sum += len;
363 tic->t_res_num++;
364}
Christoph Hellwigdd954c62006-01-11 15:34:50 +1100365
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366/*
Christoph Hellwig9006fb92012-02-20 02:31:31 +0000367 * Replenish the byte reservation required by moving the grant write head.
368 */
369int
370xfs_log_regrant(
371 struct xfs_mount *mp,
372 struct xlog_ticket *tic)
373{
Mark Tinguelyad223e62012-06-14 09:22:15 -0500374 struct xlog *log = mp->m_log;
Christoph Hellwig9006fb92012-02-20 02:31:31 +0000375 int need_bytes;
376 int error = 0;
377
378 if (XLOG_FORCED_SHUTDOWN(log))
379 return XFS_ERROR(EIO);
380
381 XFS_STATS_INC(xs_try_logspace);
382
383 /*
384 * This is a new transaction on the ticket, so we need to change the
385 * transaction ID so that the next transaction has a different TID in
386 * the log. Just add one to the existing tid so that we can see chains
387 * of rolling transactions in the log easily.
388 */
389 tic->t_tid++;
390
391 xlog_grant_push_ail(log, tic->t_unit_res);
392
393 tic->t_curr_res = tic->t_unit_res;
394 xlog_tic_reset_res(tic);
395
396 if (tic->t_cnt > 0)
397 return 0;
398
399 trace_xfs_log_regrant(log, tic);
400
401 error = xlog_grant_head_check(log, &log->l_write_head, tic,
402 &need_bytes);
403 if (error)
404 goto out_error;
405
406 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
407 trace_xfs_log_regrant_exit(log, tic);
408 xlog_verify_grant_tail(log);
409 return 0;
410
411out_error:
412 /*
413 * If we are failing, make sure the ticket doesn't have any current
414 * reservations. We don't want to add this back when the ticket/
415 * transaction gets cancelled.
416 */
417 tic->t_curr_res = 0;
418 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
419 return error;
420}
421
422/*
423 * Reserve log space and return a ticket corresponding the reservation.
424 *
425 * Each reservation is going to reserve extra space for a log record header.
426 * When writes happen to the on-disk log, we don't subtract the length of the
427 * log record header from any reservation. By wasting space in each
428 * reservation, we prevent over allocation problems.
429 */
430int
431xfs_log_reserve(
432 struct xfs_mount *mp,
433 int unit_bytes,
434 int cnt,
435 struct xlog_ticket **ticp,
436 __uint8_t client,
437 bool permanent,
438 uint t_type)
439{
Mark Tinguelyad223e62012-06-14 09:22:15 -0500440 struct xlog *log = mp->m_log;
Christoph Hellwig9006fb92012-02-20 02:31:31 +0000441 struct xlog_ticket *tic;
442 int need_bytes;
443 int error = 0;
444
445 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
446
447 if (XLOG_FORCED_SHUTDOWN(log))
448 return XFS_ERROR(EIO);
449
450 XFS_STATS_INC(xs_try_logspace);
451
452 ASSERT(*ticp == NULL);
453 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
454 KM_SLEEP | KM_MAYFAIL);
455 if (!tic)
456 return XFS_ERROR(ENOMEM);
457
458 tic->t_trans_type = t_type;
459 *ticp = tic;
460
461 xlog_grant_push_ail(log, tic->t_unit_res * tic->t_cnt);
462
463 trace_xfs_log_reserve(log, tic);
464
465 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
466 &need_bytes);
467 if (error)
468 goto out_error;
469
470 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
471 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
472 trace_xfs_log_reserve_exit(log, tic);
473 xlog_verify_grant_tail(log);
474 return 0;
475
476out_error:
477 /*
478 * If we are failing, make sure the ticket doesn't have any current
479 * reservations. We don't want to add this back when the ticket/
480 * transaction gets cancelled.
481 */
482 tic->t_curr_res = 0;
483 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
484 return error;
485}
486
487
488/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 * NOTES:
490 *
491 * 1. currblock field gets updated at startup and after in-core logs
492 * marked as with WANT_SYNC.
493 */
494
495/*
496 * This routine is called when a user of a log manager ticket is done with
497 * the reservation. If the ticket was ever used, then a commit record for
498 * the associated transaction is written out as a log operation header with
499 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
500 * a given ticket. If the ticket was one with a permanent reservation, then
501 * a few operations are done differently. Permanent reservation tickets by
502 * default don't release the reservation. They just commit the current
503 * transaction with the belief that the reservation is still needed. A flag
504 * must be passed in before permanent reservations are actually released.
505 * When these type of tickets are not released, they need to be set into
506 * the inited state again. By doing this, a start record will be written
507 * out when the next write occurs.
508 */
509xfs_lsn_t
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000510xfs_log_done(
511 struct xfs_mount *mp,
512 struct xlog_ticket *ticket,
513 struct xlog_in_core **iclog,
514 uint flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515{
Mark Tinguelyad223e62012-06-14 09:22:15 -0500516 struct xlog *log = mp->m_log;
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000517 xfs_lsn_t lsn = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 if (XLOG_FORCED_SHUTDOWN(log) ||
520 /*
521 * If nothing was ever written, don't write out commit record.
522 * If we get an error, just continue and give back the log ticket.
523 */
524 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
Dave Chinner55b66332010-03-23 11:43:17 +1100525 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 lsn = (xfs_lsn_t) -1;
527 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
528 flags |= XFS_LOG_REL_PERM_RESERV;
529 }
530 }
531
532
533 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
534 (flags & XFS_LOG_REL_PERM_RESERV)) {
Christoph Hellwig0b1b2132009-12-14 23:14:59 +0000535 trace_xfs_log_done_nonperm(log, ticket);
536
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 /*
Nathan Scottc41564b2006-03-29 08:55:14 +1000538 * Release ticket if not permanent reservation or a specific
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 * request has been made to release a permanent reservation.
540 */
541 xlog_ungrant_log_space(log, ticket);
Dave Chinnercc09c0d2008-11-17 17:37:10 +1100542 xfs_log_ticket_put(ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 } else {
Christoph Hellwig0b1b2132009-12-14 23:14:59 +0000544 trace_xfs_log_done_perm(log, ticket);
545
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 xlog_regrant_reserve_log_space(log, ticket);
Lachlan McIlroyc6a7b0f2008-08-13 16:52:50 +1000547 /* If this ticket was a permanent reservation and we aren't
548 * trying to release it, reset the inited flags; so next time
549 * we write, a start record will be written out.
550 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 ticket->t_flags |= XLOG_TIC_INITED;
Lachlan McIlroyc6a7b0f2008-08-13 16:52:50 +1000552 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553
554 return lsn;
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000555}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557/*
558 * Attaches a new iclog I/O completion callback routine during
559 * transaction commit. If the log is in error state, a non-zero
560 * return code is handed back and the caller is responsible for
561 * executing the callback at an appropriate time.
562 */
563int
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000564xfs_log_notify(
565 struct xfs_mount *mp,
566 struct xlog_in_core *iclog,
567 xfs_log_callback_t *cb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568{
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000569 int abortflg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570
David Chinner114d23a2008-04-10 12:18:39 +1000571 spin_lock(&iclog->ic_callback_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
573 if (!abortflg) {
574 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
575 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
576 cb->cb_next = NULL;
577 *(iclog->ic_callback_tail) = cb;
578 iclog->ic_callback_tail = &(cb->cb_next);
579 }
David Chinner114d23a2008-04-10 12:18:39 +1000580 spin_unlock(&iclog->ic_callback_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 return abortflg;
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000582}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583
584int
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000585xfs_log_release_iclog(
586 struct xfs_mount *mp,
587 struct xlog_in_core *iclog)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588{
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000589 if (xlog_state_release_iclog(mp->m_log, iclog)) {
Nathan Scott7d04a332006-06-09 14:58:38 +1000590 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
Jesper Juhl014c2542006-01-15 02:37:08 +0100591 return EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 }
593
594 return 0;
595}
596
597/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 * Mount a log filesystem
599 *
600 * mp - ubiquitous xfs mount point structure
601 * log_target - buftarg of on-disk log device
602 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
603 * num_bblocks - Number of BBSIZE blocks in on-disk log
604 *
605 * Return error or zero.
606 */
607int
David Chinner249a8c12008-02-05 12:13:32 +1100608xfs_log_mount(
609 xfs_mount_t *mp,
610 xfs_buftarg_t *log_target,
611 xfs_daddr_t blk_offset,
612 int num_bblks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613{
David Chinner249a8c12008-02-05 12:13:32 +1100614 int error;
615
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
Dave Chinnera0fa2b62011-03-07 10:01:35 +1100617 xfs_notice(mp, "Mounting Filesystem");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 else {
Dave Chinnera0fa2b62011-03-07 10:01:35 +1100619 xfs_notice(mp,
620"Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
Christoph Hellwigbd186aa2007-08-30 17:21:12 +1000621 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 }
623
624 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
Dave Chinnera6cb7672009-04-06 18:39:27 +0200625 if (IS_ERR(mp->m_log)) {
626 error = -PTR_ERR(mp->m_log);
Dave Chinner644c3562008-11-10 16:50:24 +1100627 goto out;
628 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 /*
David Chinner249a8c12008-02-05 12:13:32 +1100631 * Initialize the AIL now we have a log.
632 */
David Chinner249a8c12008-02-05 12:13:32 +1100633 error = xfs_trans_ail_init(mp);
634 if (error) {
Dave Chinnera0fa2b62011-03-07 10:01:35 +1100635 xfs_warn(mp, "AIL initialisation failed: error %d", error);
Christoph Hellwig26430752009-02-12 19:55:48 +0100636 goto out_free_log;
David Chinner249a8c12008-02-05 12:13:32 +1100637 }
David Chinnera9c21c12008-10-30 17:39:35 +1100638 mp->m_log->l_ailp = mp->m_ail;
David Chinner249a8c12008-02-05 12:13:32 +1100639
640 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 * skip log recovery on a norecovery mount. pretend it all
642 * just worked.
643 */
644 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
David Chinner249a8c12008-02-05 12:13:32 +1100645 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
647 if (readonly)
Christoph Hellwigbd186aa2007-08-30 17:21:12 +1000648 mp->m_flags &= ~XFS_MOUNT_RDONLY;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649
Eric Sandeen65be6052006-01-11 15:34:19 +1100650 error = xlog_recover(mp->m_log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
652 if (readonly)
Christoph Hellwigbd186aa2007-08-30 17:21:12 +1000653 mp->m_flags |= XFS_MOUNT_RDONLY;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 if (error) {
Dave Chinnera0fa2b62011-03-07 10:01:35 +1100655 xfs_warn(mp, "log mount/recovery failed: error %d",
656 error);
Christoph Hellwig26430752009-02-12 19:55:48 +0100657 goto out_destroy_ail;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 }
659 }
660
661 /* Normal transactions can now occur */
662 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
663
Dave Chinner71e330b2010-05-21 14:37:18 +1000664 /*
665 * Now the log has been fully initialised and we know were our
666 * space grant counters are, we can initialise the permanent ticket
667 * needed for delayed logging to work.
668 */
669 xlog_cil_init_post_recovery(mp->m_log);
670
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 return 0;
Christoph Hellwig26430752009-02-12 19:55:48 +0100672
673out_destroy_ail:
674 xfs_trans_ail_destroy(mp);
675out_free_log:
676 xlog_dealloc_log(mp->m_log);
Dave Chinner644c3562008-11-10 16:50:24 +1100677out:
David Chinner249a8c12008-02-05 12:13:32 +1100678 return error;
Christoph Hellwig26430752009-02-12 19:55:48 +0100679}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
681/*
682 * Finish the recovery of the file system. This is separate from
683 * the xfs_log_mount() call, because it depends on the code in
684 * xfs_mountfs() to read in the root and real-time bitmap inodes
685 * between calling xfs_log_mount() and here.
686 *
687 * mp - ubiquitous xfs mount point structure
688 */
689int
Christoph Hellwig42490232008-08-13 16:49:32 +1000690xfs_log_mount_finish(xfs_mount_t *mp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
692 int error;
693
694 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
Christoph Hellwig42490232008-08-13 16:49:32 +1000695 error = xlog_recover_finish(mp->m_log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 else {
697 error = 0;
Christoph Hellwigbd186aa2007-08-30 17:21:12 +1000698 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 }
700
701 return error;
702}
703
704/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 * Final log writes as part of unmount.
706 *
707 * Mark the filesystem clean as unmount happens. Note that during relocation
708 * this routine needs to be executed as part of source-bag while the
709 * deallocation must not be done until source-end.
710 */
711
712/*
713 * Unmount record used to have a string "Unmount filesystem--" in the
714 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
715 * We just write the magic number now since that particular field isn't
716 * currently architecture converted and "nUmount" is a bit foo.
717 * As far as I know, there weren't any dependencies on the old behaviour.
718 */
719
720int
721xfs_log_unmount_write(xfs_mount_t *mp)
722{
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -0500723 struct xlog *log = mp->m_log;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 xlog_in_core_t *iclog;
725#ifdef DEBUG
726 xlog_in_core_t *first_iclog;
727#endif
Christoph Hellwig35a8a722010-02-15 23:34:54 +0000728 xlog_ticket_t *tic = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 xfs_lsn_t lsn;
730 int error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 /*
733 * Don't write out unmount record on read-only mounts.
734 * Or, if we are doing a forced umount (typically because of IO errors).
735 */
Christoph Hellwigbd186aa2007-08-30 17:21:12 +1000736 if (mp->m_flags & XFS_MOUNT_RDONLY)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 return 0;
738
Christoph Hellwiga14a3482010-01-19 09:56:46 +0000739 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
David Chinnerb911ca02008-04-10 12:24:30 +1000740 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741
742#ifdef DEBUG
743 first_iclog = iclog = log->l_iclog;
744 do {
745 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
746 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
747 ASSERT(iclog->ic_offset == 0);
748 }
749 iclog = iclog->ic_next;
750 } while (iclog != first_iclog);
751#endif
752 if (! (XLOG_FORCED_SHUTDOWN(log))) {
Tim Shimmin955e47a2006-09-28 11:04:16 +1000753 error = xfs_log_reserve(mp, 600, 1, &tic,
754 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 if (!error) {
Dave Chinner55b66332010-03-23 11:43:17 +1100756 /* the data section must be 32 bit size aligned */
757 struct {
758 __uint16_t magic;
759 __uint16_t pad1;
760 __uint32_t pad2; /* may as well make it 64 bits */
761 } magic = {
762 .magic = XLOG_UNMOUNT_TYPE,
763 };
764 struct xfs_log_iovec reg = {
Christoph Hellwig4e0d5f92010-06-23 18:11:15 +1000765 .i_addr = &magic,
Dave Chinner55b66332010-03-23 11:43:17 +1100766 .i_len = sizeof(magic),
767 .i_type = XLOG_REG_TYPE_UNMOUNT,
768 };
769 struct xfs_log_vec vec = {
770 .lv_niovecs = 1,
771 .lv_iovecp = &reg,
772 };
773
Dave Chinner39486592012-03-22 05:15:11 +0000774 /* remove inited flag, and account for space used */
Dave Chinner55b66332010-03-23 11:43:17 +1100775 tic->t_flags = 0;
Dave Chinner39486592012-03-22 05:15:11 +0000776 tic->t_curr_res -= sizeof(magic);
Dave Chinner55b66332010-03-23 11:43:17 +1100777 error = xlog_write(log, &vec, tic, &lsn,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 NULL, XLOG_UNMOUNT_TRANS);
779 /*
780 * At this point, we're umounting anyway,
781 * so there's no point in transitioning log state
782 * to IOERROR. Just continue...
783 */
784 }
785
Dave Chinnera0fa2b62011-03-07 10:01:35 +1100786 if (error)
787 xfs_alert(mp, "%s: unmount record failed", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
789
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000790 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 iclog = log->l_iclog;
David Chinner155cc6b2008-03-06 13:44:14 +1100792 atomic_inc(&iclog->ic_refcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 xlog_state_want_sync(log, iclog);
Christoph Hellwig39e2def2008-12-03 12:20:28 +0100794 spin_unlock(&log->l_icloglock);
David Chinner1bb7d6b2008-04-10 12:24:38 +1000795 error = xlog_state_release_iclog(log, iclog);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000797 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
799 iclog->ic_state == XLOG_STATE_DIRTY)) {
800 if (!XLOG_FORCED_SHUTDOWN(log)) {
Dave Chinnereb40a872010-12-21 12:09:01 +1100801 xlog_wait(&iclog->ic_force_wait,
802 &log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 } else {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000804 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 }
806 } else {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000807 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 }
Tim Shimmin955e47a2006-09-28 11:04:16 +1000809 if (tic) {
Christoph Hellwig0b1b2132009-12-14 23:14:59 +0000810 trace_xfs_log_umount_write(log, tic);
Tim Shimmin955e47a2006-09-28 11:04:16 +1000811 xlog_ungrant_log_space(log, tic);
Dave Chinnercc09c0d2008-11-17 17:37:10 +1100812 xfs_log_ticket_put(tic);
Tim Shimmin955e47a2006-09-28 11:04:16 +1000813 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 } else {
815 /*
816 * We're already in forced_shutdown mode, couldn't
817 * even attempt to write out the unmount transaction.
818 *
819 * Go through the motions of sync'ing and releasing
820 * the iclog, even though no I/O will actually happen,
Nathan Scottc41564b2006-03-29 08:55:14 +1000821 * we need to wait for other log I/Os that may already
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 * be in progress. Do this as a separate section of
823 * code so we'll know if we ever get stuck here that
824 * we're in this odd situation of trying to unmount
825 * a file system that went into forced_shutdown as
826 * the result of an unmount..
827 */
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000828 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 iclog = log->l_iclog;
David Chinner155cc6b2008-03-06 13:44:14 +1100830 atomic_inc(&iclog->ic_refcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
832 xlog_state_want_sync(log, iclog);
Christoph Hellwig39e2def2008-12-03 12:20:28 +0100833 spin_unlock(&log->l_icloglock);
David Chinner1bb7d6b2008-04-10 12:24:38 +1000834 error = xlog_state_release_iclog(log, iclog);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000836 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837
838 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
839 || iclog->ic_state == XLOG_STATE_DIRTY
840 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
841
Dave Chinnereb40a872010-12-21 12:09:01 +1100842 xlog_wait(&iclog->ic_force_wait,
843 &log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 } else {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000845 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 }
847 }
848
David Chinner1bb7d6b2008-04-10 12:24:38 +1000849 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850} /* xfs_log_unmount_write */
851
852/*
853 * Deallocate log structures for unmount/relocation.
David Chinner249a8c12008-02-05 12:13:32 +1100854 *
855 * We need to stop the aild from running before we destroy
856 * and deallocate the log as the aild references the log.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 */
858void
Christoph Hellwig21b699c2009-03-16 08:19:29 +0100859xfs_log_unmount(xfs_mount_t *mp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860{
Ben Myers11159a02012-05-25 15:45:36 -0500861 cancel_delayed_work_sync(&mp->m_sync_work);
David Chinner249a8c12008-02-05 12:13:32 +1100862 xfs_trans_ail_destroy(mp);
Nathan Scottc41564b2006-03-29 08:55:14 +1000863 xlog_dealloc_log(mp->m_log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864}
865
Dave Chinner43f5efc2010-03-23 10:10:00 +1100866void
867xfs_log_item_init(
868 struct xfs_mount *mp,
869 struct xfs_log_item *item,
870 int type,
Christoph Hellwig272e42b2011-10-28 09:54:24 +0000871 const struct xfs_item_ops *ops)
Dave Chinner43f5efc2010-03-23 10:10:00 +1100872{
873 item->li_mountp = mp;
874 item->li_ailp = mp->m_ail;
875 item->li_type = type;
876 item->li_ops = ops;
Dave Chinner71e330b2010-05-21 14:37:18 +1000877 item->li_lv = NULL;
878
879 INIT_LIST_HEAD(&item->li_ail);
880 INIT_LIST_HEAD(&item->li_cil);
Dave Chinner43f5efc2010-03-23 10:10:00 +1100881}
882
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000883/*
884 * Wake up processes waiting for log space after we have moved the log tail.
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000885 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886void
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000887xfs_log_space_wake(
Christoph Hellwigcfb7cdc2012-02-20 02:31:23 +0000888 struct xfs_mount *mp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
Mark Tinguelyad223e62012-06-14 09:22:15 -0500890 struct xlog *log = mp->m_log;
Christoph Hellwigcfb7cdc2012-02-20 02:31:23 +0000891 int free_bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 if (XLOG_FORCED_SHUTDOWN(log))
894 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895
Christoph Hellwig28496962012-02-20 02:31:25 +0000896 if (!list_empty_careful(&log->l_write_head.waiters)) {
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000897 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
898
Christoph Hellwig28496962012-02-20 02:31:25 +0000899 spin_lock(&log->l_write_head.lock);
900 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
Christoph Hellwige179840d2012-02-20 02:31:29 +0000901 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
Christoph Hellwig28496962012-02-20 02:31:25 +0000902 spin_unlock(&log->l_write_head.lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 }
Dave Chinner10547942010-12-21 12:02:25 +1100904
Christoph Hellwig28496962012-02-20 02:31:25 +0000905 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000906 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
907
Christoph Hellwig28496962012-02-20 02:31:25 +0000908 spin_lock(&log->l_reserve_head.lock);
909 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
Christoph Hellwige179840d2012-02-20 02:31:29 +0000910 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
Christoph Hellwig28496962012-02-20 02:31:25 +0000911 spin_unlock(&log->l_reserve_head.lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 }
Dave Chinner3f16b982010-12-21 12:29:01 +1100913}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914
915/*
916 * Determine if we have a transaction that has gone to disk
Dave Chinnerb6f8dd42010-04-13 15:06:44 +1000917 * that needs to be covered. To begin the transition to the idle state
918 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
919 * If we are then in a state where covering is needed, the caller is informed
920 * that dummy transactions are required to move the log into the idle state.
921 *
922 * Because this is called as part of the sync process, we should also indicate
923 * that dummy transactions should be issued in anything but the covered or
924 * idle states. This ensures that the log tail is accurately reflected in
925 * the log at the end of the sync, hence if a crash occurrs avoids replay
926 * of transactions where the metadata is already on disk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 */
928int
929xfs_log_need_covered(xfs_mount_t *mp)
930{
David Chinner27d8d5f2008-10-30 17:38:39 +1100931 int needed = 0;
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -0500932 struct xlog *log = mp->m_log;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933
David Chinner92821e22007-05-24 15:26:31 +1000934 if (!xfs_fs_writable(mp))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 return 0;
936
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000937 spin_lock(&log->l_icloglock);
Dave Chinnerb6f8dd42010-04-13 15:06:44 +1000938 switch (log->l_covered_state) {
939 case XLOG_STATE_COVER_DONE:
940 case XLOG_STATE_COVER_DONE2:
941 case XLOG_STATE_COVER_IDLE:
942 break;
943 case XLOG_STATE_COVER_NEED:
944 case XLOG_STATE_COVER_NEED2:
Dave Chinnerfd074842011-04-08 12:45:07 +1000945 if (!xfs_ail_min_lsn(log->l_ailp) &&
Dave Chinnerb6f8dd42010-04-13 15:06:44 +1000946 xlog_iclogs_empty(log)) {
947 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
948 log->l_covered_state = XLOG_STATE_COVER_DONE;
949 else
950 log->l_covered_state = XLOG_STATE_COVER_DONE2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 }
Dave Chinnerb6f8dd42010-04-13 15:06:44 +1000952 /* FALLTHRU */
953 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 needed = 1;
Dave Chinnerb6f8dd42010-04-13 15:06:44 +1000955 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 }
Eric Sandeenb22cd72c2007-10-11 17:37:10 +1000957 spin_unlock(&log->l_icloglock);
Jesper Juhl014c2542006-01-15 02:37:08 +0100958 return needed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959}
960
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000961/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 * We may be holding the log iclog lock upon entering this routine.
963 */
964xfs_lsn_t
Christoph Hellwig1c304622012-04-23 15:58:33 +1000965xlog_assign_tail_lsn_locked(
Dave Chinner1c3cb9e2010-12-21 12:28:39 +1100966 struct xfs_mount *mp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967{
Mark Tinguelyad223e62012-06-14 09:22:15 -0500968 struct xlog *log = mp->m_log;
Christoph Hellwig1c304622012-04-23 15:58:33 +1000969 struct xfs_log_item *lip;
970 xfs_lsn_t tail_lsn;
971
972 assert_spin_locked(&mp->m_ail->xa_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000974 /*
975 * To make sure we always have a valid LSN for the log tail we keep
976 * track of the last LSN which was committed in log->l_last_sync_lsn,
Christoph Hellwig1c304622012-04-23 15:58:33 +1000977 * and use that when the AIL was empty.
Christoph Hellwig09a423a2012-02-20 02:31:20 +0000978 */
Christoph Hellwig1c304622012-04-23 15:58:33 +1000979 lip = xfs_ail_min(mp->m_ail);
980 if (lip)
981 tail_lsn = lip->li_lsn;
982 else
Dave Chinner84f3c682010-12-03 22:11:29 +1100983 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
Dave Chinner1c3cb9e2010-12-21 12:28:39 +1100984 atomic64_set(&log->l_tail_lsn, tail_lsn);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 return tail_lsn;
Dave Chinner1c3cb9e2010-12-21 12:28:39 +1100986}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987
Christoph Hellwig1c304622012-04-23 15:58:33 +1000988xfs_lsn_t
989xlog_assign_tail_lsn(
990 struct xfs_mount *mp)
991{
992 xfs_lsn_t tail_lsn;
993
994 spin_lock(&mp->m_ail->xa_lock);
995 tail_lsn = xlog_assign_tail_lsn_locked(mp);
996 spin_unlock(&mp->m_ail->xa_lock);
997
998 return tail_lsn;
999}
1000
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001/*
1002 * Return the space in the log between the tail and the head. The head
1003 * is passed in the cycle/bytes formal parms. In the special case where
1004 * the reserve head has wrapped passed the tail, this calculation is no
1005 * longer valid. In this case, just return 0 which means there is no space
1006 * in the log. This works for all places where this function is called
1007 * with the reserve head. Of course, if the write head were to ever
1008 * wrap the tail, we should blow up. Rather than catch this case here,
1009 * we depend on other ASSERTions in other parts of the code. XXXmiken
1010 *
1011 * This code also handles the case where the reservation head is behind
1012 * the tail. The details of this case are described below, but the end
1013 * result is that we return the size of the log as the amount of space left.
1014 */
David Chinnera8272ce2007-11-23 16:28:09 +11001015STATIC int
Dave Chinnera69ed032010-12-21 12:08:20 +11001016xlog_space_left(
Mark Tinguelyad223e62012-06-14 09:22:15 -05001017 struct xlog *log,
Dave Chinnerc8a09ff2010-12-04 00:02:40 +11001018 atomic64_t *head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019{
Dave Chinnera69ed032010-12-21 12:08:20 +11001020 int free_bytes;
1021 int tail_bytes;
1022 int tail_cycle;
1023 int head_cycle;
1024 int head_bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025
Dave Chinnera69ed032010-12-21 12:08:20 +11001026 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
Dave Chinner1c3cb9e2010-12-21 12:28:39 +11001027 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1028 tail_bytes = BBTOB(tail_bytes);
Dave Chinnera69ed032010-12-21 12:08:20 +11001029 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1030 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1031 else if (tail_cycle + 1 < head_cycle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032 return 0;
Dave Chinnera69ed032010-12-21 12:08:20 +11001033 else if (tail_cycle < head_cycle) {
1034 ASSERT(tail_cycle == (head_cycle - 1));
1035 free_bytes = tail_bytes - head_bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 } else {
1037 /*
1038 * The reservation head is behind the tail.
1039 * In this case we just want to return the size of the
1040 * log as the amount of space left.
1041 */
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001042 xfs_alert(log->l_mp,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 "xlog_space_left: head behind tail\n"
1044 " tail_cycle = %d, tail_bytes = %d\n"
1045 " GH cycle = %d, GH bytes = %d",
Dave Chinnera69ed032010-12-21 12:08:20 +11001046 tail_cycle, tail_bytes, head_cycle, head_bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047 ASSERT(0);
1048 free_bytes = log->l_logsize;
1049 }
1050 return free_bytes;
Dave Chinnera69ed032010-12-21 12:08:20 +11001051}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052
1053
1054/*
1055 * Log function which is called when an io completes.
1056 *
1057 * The log manager needs its own routine, in order to control what
1058 * happens with the buffer after the write completes.
1059 */
1060void
1061xlog_iodone(xfs_buf_t *bp)
1062{
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001063 struct xlog_in_core *iclog = bp->b_fspriv;
1064 struct xlog *l = iclog->ic_log;
1065 int aborted = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066
1067 /*
1068 * Race to shutdown the filesystem if we see an error.
1069 */
Chandra Seetharaman5a52c2a582011-07-22 23:39:51 +00001070 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
Christoph Hellwig901796a2011-10-10 16:52:49 +00001072 xfs_buf_ioerror_alert(bp, __func__);
Christoph Hellwigc867cb62011-10-10 16:52:46 +00001073 xfs_buf_stale(bp);
Nathan Scott7d04a332006-06-09 14:58:38 +10001074 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 /*
1076 * This flag will be propagated to the trans-committed
1077 * callback routines to let them know that the log-commit
1078 * didn't succeed.
1079 */
1080 aborted = XFS_LI_ABORTED;
1081 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1082 aborted = XFS_LI_ABORTED;
1083 }
David Chinner3db296f2007-05-14 18:24:16 +10001084
1085 /* log I/O is always issued ASYNC */
1086 ASSERT(XFS_BUF_ISASYNC(bp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 xlog_state_done_syncing(iclog, aborted);
David Chinner3db296f2007-05-14 18:24:16 +10001088 /*
1089 * do not reference the buffer (bp) here as we could race
1090 * with it being freed after writing the unmount record to the
1091 * log.
1092 */
1093
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094} /* xlog_iodone */
1095
1096/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 * Return size of each in-core log record buffer.
1098 *
Malcolm Parsons9da096f2009-03-29 09:55:42 +02001099 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 *
1101 * If the filesystem blocksize is too large, we may need to choose a
1102 * larger size since the directory code currently logs entire blocks.
1103 */
1104
1105STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001106xlog_get_iclog_buffer_size(
1107 struct xfs_mount *mp,
1108 struct xlog *log)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109{
1110 int size;
1111 int xhdrs;
1112
Eric Sandeen1cb51252007-08-16 16:24:43 +10001113 if (mp->m_logbufs <= 0)
1114 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1115 else
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001116 log->l_iclog_bufs = mp->m_logbufs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117
1118 /*
1119 * Buffer size passed in from mount system call.
1120 */
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001121 if (mp->m_logbsize > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 size = log->l_iclog_size = mp->m_logbsize;
1123 log->l_iclog_size_log = 0;
1124 while (size != 1) {
1125 log->l_iclog_size_log++;
1126 size >>= 1;
1127 }
1128
Eric Sandeen62118702008-03-06 13:44:28 +11001129 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
Malcolm Parsons9da096f2009-03-29 09:55:42 +02001130 /* # headers = size / 32k
1131 * one header holds cycles from 32k of data
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 */
1133
1134 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1135 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1136 xhdrs++;
1137 log->l_iclog_hsize = xhdrs << BBSHIFT;
1138 log->l_iclog_heads = xhdrs;
1139 } else {
1140 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1141 log->l_iclog_hsize = BBSIZE;
1142 log->l_iclog_heads = 1;
1143 }
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001144 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 }
1146
Malcolm Parsons9da096f2009-03-29 09:55:42 +02001147 /* All machines use 32kB buffers by default. */
Eric Sandeen1cb51252007-08-16 16:24:43 +10001148 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1149 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
1151 /* the default log size is 16k or 32k which is one header sector */
1152 log->l_iclog_hsize = BBSIZE;
1153 log->l_iclog_heads = 1;
1154
Christoph Hellwig7153f8b2009-02-09 08:36:46 +01001155done:
1156 /* are we being asked to make the sizes selected above visible? */
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001157 if (mp->m_logbufs == 0)
1158 mp->m_logbufs = log->l_iclog_bufs;
1159 if (mp->m_logbsize == 0)
1160 mp->m_logbsize = log->l_iclog_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161} /* xlog_get_iclog_buffer_size */
1162
1163
1164/*
1165 * This routine initializes some of the log structure for a given mount point.
1166 * Its primary purpose is to fill in enough, so recovery can occur. However,
1167 * some other stuff may be filled in too.
1168 */
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001169STATIC struct xlog *
1170xlog_alloc_log(
1171 struct xfs_mount *mp,
1172 struct xfs_buftarg *log_target,
1173 xfs_daddr_t blk_offset,
1174 int num_bblks)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175{
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001176 struct xlog *log;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 xlog_rec_header_t *head;
1178 xlog_in_core_t **iclogp;
1179 xlog_in_core_t *iclog, *prev_iclog=NULL;
1180 xfs_buf_t *bp;
1181 int i;
Dave Chinnera6cb7672009-04-06 18:39:27 +02001182 int error = ENOMEM;
Alex Elder69ce58f2010-04-20 17:09:59 +10001183 uint log2_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001185 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
Dave Chinnera6cb7672009-04-06 18:39:27 +02001186 if (!log) {
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001187 xfs_warn(mp, "Log allocation failed: No memory!");
Dave Chinnera6cb7672009-04-06 18:39:27 +02001188 goto out;
1189 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190
1191 log->l_mp = mp;
1192 log->l_targ = log_target;
1193 log->l_logsize = BBTOB(num_bblks);
1194 log->l_logBBstart = blk_offset;
1195 log->l_logBBsize = num_bblks;
1196 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1197 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1198
1199 log->l_prev_block = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
Dave Chinner1c3cb9e2010-12-21 12:28:39 +11001201 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1202 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
Christoph Hellwigc303c5b2012-02-20 02:31:26 +00001204
1205 xlog_grant_head_init(&log->l_reserve_head);
1206 xlog_grant_head_init(&log->l_write_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207
Dave Chinnera6cb7672009-04-06 18:39:27 +02001208 error = EFSCORRUPTED;
Eric Sandeen62118702008-03-06 13:44:28 +11001209 if (xfs_sb_version_hassector(&mp->m_sb)) {
Alex Elder69ce58f2010-04-20 17:09:59 +10001210 log2_size = mp->m_sb.sb_logsectlog;
1211 if (log2_size < BBSHIFT) {
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001212 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1213 log2_size, BBSHIFT);
Alex Elder69ce58f2010-04-20 17:09:59 +10001214 goto out_free_log;
1215 }
1216
1217 log2_size -= BBSHIFT;
1218 if (log2_size > mp->m_sectbb_log) {
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001219 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1220 log2_size, mp->m_sectbb_log);
Dave Chinnera6cb7672009-04-06 18:39:27 +02001221 goto out_free_log;
1222 }
1223
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 /* for larger sector sizes, must have v2 or external log */
Alex Elder69ce58f2010-04-20 17:09:59 +10001225 if (log2_size && log->l_logBBstart > 0 &&
1226 !xfs_sb_version_haslogv2(&mp->m_sb)) {
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001227 xfs_warn(mp,
1228 "log sector size (0x%x) invalid for configuration.",
1229 log2_size);
Dave Chinnera6cb7672009-04-06 18:39:27 +02001230 goto out_free_log;
1231 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 }
Alex Elder69ce58f2010-04-20 17:09:59 +10001233 log->l_sectBBsize = 1 << log2_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234
1235 xlog_get_iclog_buffer_size(mp, log);
1236
Dave Chinnera6cb7672009-04-06 18:39:27 +02001237 error = ENOMEM;
Dave Chinnere70b73f2012-04-23 15:58:49 +10001238 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
Dave Chinner644c3562008-11-10 16:50:24 +11001239 if (!bp)
1240 goto out_free_log;
Christoph Hellwigcb669ca2011-07-13 13:43:49 +02001241 bp->b_iodone = xlog_iodone;
Christoph Hellwig0c842ad2011-07-08 14:36:19 +02001242 ASSERT(xfs_buf_islocked(bp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 log->l_xbuf = bp;
1244
Eric Sandeen007c61c2007-10-11 17:43:56 +10001245 spin_lock_init(&log->l_icloglock);
Dave Chinnereb40a872010-12-21 12:09:01 +11001246 init_waitqueue_head(&log->l_flush_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 iclogp = &log->l_iclog;
1249 /*
1250 * The amount of memory to allocate for the iclog structure is
1251 * rather funky due to the way the structure is defined. It is
1252 * done this way so that we can use different sizes for machines
1253 * with different amounts of memory. See the definition of
1254 * xlog_in_core_t in xfs_log_priv.h for details.
1255 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 ASSERT(log->l_iclog_size >= 4096);
1257 for (i=0; i < log->l_iclog_bufs; i++) {
Dave Chinner644c3562008-11-10 16:50:24 +11001258 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1259 if (!*iclogp)
1260 goto out_free_iclog;
1261
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 iclog = *iclogp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 iclog->ic_prev = prev_iclog;
1264 prev_iclog = iclog;
Christoph Hellwig1fa40b02007-05-14 18:23:50 +10001265
Dave Chinner686865f2010-09-24 20:07:47 +10001266 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
Dave Chinnere70b73f2012-04-23 15:58:49 +10001267 BTOBB(log->l_iclog_size), 0);
Dave Chinner644c3562008-11-10 16:50:24 +11001268 if (!bp)
1269 goto out_free_iclog;
Christoph Hellwigc8da0fa2011-07-08 14:36:25 +02001270
Christoph Hellwigcb669ca2011-07-13 13:43:49 +02001271 bp->b_iodone = xlog_iodone;
Christoph Hellwig1fa40b02007-05-14 18:23:50 +10001272 iclog->ic_bp = bp;
Christoph Hellwigb28708d2008-11-28 14:23:38 +11001273 iclog->ic_data = bp->b_addr;
David Chinner4679b2d2008-04-10 12:18:54 +10001274#ifdef DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
David Chinner4679b2d2008-04-10 12:18:54 +10001276#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 head = &iclog->ic_header;
1278 memset(head, 0, sizeof(xlog_rec_header_t));
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001279 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1280 head->h_version = cpu_to_be32(
Eric Sandeen62118702008-03-06 13:44:28 +11001281 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001282 head->h_size = cpu_to_be32(log->l_iclog_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 /* new fields */
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001284 head->h_fmt = cpu_to_be32(XLOG_FMT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1286
Dave Chinner4e94b712012-04-23 15:58:51 +10001287 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 iclog->ic_state = XLOG_STATE_ACTIVE;
1289 iclog->ic_log = log;
David Chinner114d23a2008-04-10 12:18:39 +10001290 atomic_set(&iclog->ic_refcnt, 0);
1291 spin_lock_init(&iclog->ic_callback_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 iclog->ic_callback_tail = &(iclog->ic_callback);
Christoph Hellwigb28708d2008-11-28 14:23:38 +11001293 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294
Christoph Hellwig0c842ad2011-07-08 14:36:19 +02001295 ASSERT(xfs_buf_islocked(iclog->ic_bp));
Dave Chinnereb40a872010-12-21 12:09:01 +11001296 init_waitqueue_head(&iclog->ic_force_wait);
1297 init_waitqueue_head(&iclog->ic_write_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298
1299 iclogp = &iclog->ic_next;
1300 }
1301 *iclogp = log->l_iclog; /* complete ring */
1302 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1303
Dave Chinner71e330b2010-05-21 14:37:18 +10001304 error = xlog_cil_init(log);
1305 if (error)
1306 goto out_free_iclog;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 return log;
Dave Chinner644c3562008-11-10 16:50:24 +11001308
1309out_free_iclog:
1310 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1311 prev_iclog = iclog->ic_next;
Dave Chinnereb40a872010-12-21 12:09:01 +11001312 if (iclog->ic_bp)
Dave Chinner644c3562008-11-10 16:50:24 +11001313 xfs_buf_free(iclog->ic_bp);
Dave Chinner644c3562008-11-10 16:50:24 +11001314 kmem_free(iclog);
1315 }
1316 spinlock_destroy(&log->l_icloglock);
Dave Chinner644c3562008-11-10 16:50:24 +11001317 xfs_buf_free(log->l_xbuf);
1318out_free_log:
1319 kmem_free(log);
Dave Chinnera6cb7672009-04-06 18:39:27 +02001320out:
1321 return ERR_PTR(-error);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322} /* xlog_alloc_log */
1323
1324
1325/*
1326 * Write out the commit record of a transaction associated with the given
1327 * ticket. Return the lsn of the commit record.
1328 */
1329STATIC int
Dave Chinner55b66332010-03-23 11:43:17 +11001330xlog_commit_record(
Mark Tinguelyad223e62012-06-14 09:22:15 -05001331 struct xlog *log,
Dave Chinner55b66332010-03-23 11:43:17 +11001332 struct xlog_ticket *ticket,
1333 struct xlog_in_core **iclog,
1334 xfs_lsn_t *commitlsnp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335{
Dave Chinner55b66332010-03-23 11:43:17 +11001336 struct xfs_mount *mp = log->l_mp;
1337 int error;
1338 struct xfs_log_iovec reg = {
1339 .i_addr = NULL,
1340 .i_len = 0,
1341 .i_type = XLOG_REG_TYPE_COMMIT,
1342 };
1343 struct xfs_log_vec vec = {
1344 .lv_niovecs = 1,
1345 .lv_iovecp = &reg,
1346 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
1348 ASSERT_ALWAYS(iclog);
Dave Chinner55b66332010-03-23 11:43:17 +11001349 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1350 XLOG_COMMIT_TRANS);
1351 if (error)
Nathan Scott7d04a332006-06-09 14:58:38 +10001352 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
Jesper Juhl014c2542006-01-15 02:37:08 +01001353 return error;
Dave Chinner55b66332010-03-23 11:43:17 +11001354}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355
1356/*
1357 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1358 * log space. This code pushes on the lsn which would supposedly free up
1359 * the 25% which we want to leave free. We may need to adopt a policy which
1360 * pushes on an lsn which is further along in the log once we reach the high
1361 * water mark. In this manner, we would be creating a low water mark.
1362 */
David Chinnera8272ce2007-11-23 16:28:09 +11001363STATIC void
Dave Chinner2ced19c2010-12-21 12:09:20 +11001364xlog_grant_push_ail(
Mark Tinguelyad223e62012-06-14 09:22:15 -05001365 struct xlog *log,
Dave Chinner2ced19c2010-12-21 12:09:20 +11001366 int need_bytes)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367{
Dave Chinner2ced19c2010-12-21 12:09:20 +11001368 xfs_lsn_t threshold_lsn = 0;
Dave Chinner84f3c682010-12-03 22:11:29 +11001369 xfs_lsn_t last_sync_lsn;
Dave Chinner2ced19c2010-12-21 12:09:20 +11001370 int free_blocks;
1371 int free_bytes;
1372 int threshold_block;
1373 int threshold_cycle;
1374 int free_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
Dave Chinner2ced19c2010-12-21 12:09:20 +11001376 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377
Christoph Hellwig28496962012-02-20 02:31:25 +00001378 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
Dave Chinner2ced19c2010-12-21 12:09:20 +11001379 free_blocks = BTOBBT(free_bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380
Dave Chinner2ced19c2010-12-21 12:09:20 +11001381 /*
1382 * Set the threshold for the minimum number of free blocks in the
1383 * log to the maximum of what the caller needs, one quarter of the
1384 * log, and 256 blocks.
1385 */
1386 free_threshold = BTOBB(need_bytes);
1387 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1388 free_threshold = MAX(free_threshold, 256);
1389 if (free_blocks >= free_threshold)
1390 return;
1391
Dave Chinner1c3cb9e2010-12-21 12:28:39 +11001392 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1393 &threshold_block);
1394 threshold_block += free_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 if (threshold_block >= log->l_logBBsize) {
Dave Chinner2ced19c2010-12-21 12:09:20 +11001396 threshold_block -= log->l_logBBsize;
1397 threshold_cycle += 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 }
Dave Chinner2ced19c2010-12-21 12:09:20 +11001399 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1400 threshold_block);
1401 /*
1402 * Don't pass in an lsn greater than the lsn of the last
Dave Chinner84f3c682010-12-03 22:11:29 +11001403 * log record known to be on disk. Use a snapshot of the last sync lsn
1404 * so that it doesn't change between the compare and the set.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 */
Dave Chinner84f3c682010-12-03 22:11:29 +11001406 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1407 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1408 threshold_lsn = last_sync_lsn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409
Dave Chinner2ced19c2010-12-21 12:09:20 +11001410 /*
1411 * Get the transaction layer to kick the dirty buffers out to
1412 * disk asynchronously. No point in trying to do this if
1413 * the filesystem is shutting down.
1414 */
1415 if (!XLOG_FORCED_SHUTDOWN(log))
Dave Chinnerfd074842011-04-08 12:45:07 +10001416 xfs_ail_push(log->l_ailp, threshold_lsn);
Dave Chinner2ced19c2010-12-21 12:09:20 +11001417}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
Christoph Hellwig873ff552010-01-13 22:17:57 +00001419/*
1420 * The bdstrat callback function for log bufs. This gives us a central
1421 * place to trap bufs in case we get hit by a log I/O error and need to
1422 * shutdown. Actually, in practice, even when we didn't get a log error,
1423 * we transition the iclogs to IOERROR state *after* flushing all existing
1424 * iclogs to disk. This is because we don't want anymore new transactions to be
1425 * started or completed afterwards.
1426 */
1427STATIC int
1428xlog_bdstrat(
1429 struct xfs_buf *bp)
1430{
Christoph Hellwigadadbee2011-07-13 13:43:49 +02001431 struct xlog_in_core *iclog = bp->b_fspriv;
Christoph Hellwig873ff552010-01-13 22:17:57 +00001432
Christoph Hellwig873ff552010-01-13 22:17:57 +00001433 if (iclog->ic_state & XLOG_STATE_IOERROR) {
Chandra Seetharaman5a52c2a582011-07-22 23:39:51 +00001434 xfs_buf_ioerror(bp, EIO);
Christoph Hellwigc867cb62011-10-10 16:52:46 +00001435 xfs_buf_stale(bp);
Christoph Hellwig1a1a3e92010-10-06 18:41:18 +00001436 xfs_buf_ioend(bp, 0);
Christoph Hellwig873ff552010-01-13 22:17:57 +00001437 /*
1438 * It would seem logical to return EIO here, but we rely on
1439 * the log state machine to propagate I/O errors instead of
1440 * doing it here.
1441 */
1442 return 0;
1443 }
1444
Christoph Hellwig873ff552010-01-13 22:17:57 +00001445 xfs_buf_iorequest(bp);
1446 return 0;
1447}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448
1449/*
1450 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1451 * fashion. Previously, we should have moved the current iclog
1452 * ptr in the log to point to the next available iclog. This allows further
1453 * write to continue while this code syncs out an iclog ready to go.
1454 * Before an in-core log can be written out, the data section must be scanned
1455 * to save away the 1st word of each BBSIZE block into the header. We replace
1456 * it with the current cycle count. Each BBSIZE block is tagged with the
1457 * cycle count because there in an implicit assumption that drives will
1458 * guarantee that entire 512 byte blocks get written at once. In other words,
1459 * we can't have part of a 512 byte block written and part not written. By
1460 * tagging each block, we will know which blocks are valid when recovering
1461 * after an unclean shutdown.
1462 *
1463 * This routine is single threaded on the iclog. No other thread can be in
1464 * this routine with the same iclog. Changing contents of iclog can there-
1465 * fore be done without grabbing the state machine lock. Updating the global
1466 * log will require grabbing the lock though.
1467 *
1468 * The entire log manager uses a logical block numbering scheme. Only
1469 * log_sync (and then only bwrite()) know about the fact that the log may
1470 * not start with block zero on a given device. The log block start offset
1471 * is added immediately before calling bwrite().
1472 */
1473
David Chinnera8272ce2007-11-23 16:28:09 +11001474STATIC int
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001475xlog_sync(
1476 struct xlog *log,
1477 struct xlog_in_core *iclog)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478{
1479 xfs_caddr_t dptr; /* pointer to byte sized element */
1480 xfs_buf_t *bp;
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001481 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 uint count; /* byte count of bwrite */
1483 uint count_init; /* initial count before roundup */
1484 int roundoff; /* roundoff to BB or stripe */
1485 int split = 0; /* split write into two regions */
1486 int error;
Eric Sandeen62118702008-03-06 13:44:28 +11001487 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488
1489 XFS_STATS_INC(xs_log_writes);
David Chinner155cc6b2008-03-06 13:44:14 +11001490 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491
1492 /* Add for LR header */
1493 count_init = log->l_iclog_hsize + iclog->ic_offset;
1494
1495 /* Round out the log write size */
1496 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1497 /* we have a v2 stripe unit to use */
1498 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1499 } else {
1500 count = BBTOB(BTOBB(count_init));
1501 }
1502 roundoff = count - count_init;
1503 ASSERT(roundoff >= 0);
1504 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1505 roundoff < log->l_mp->m_sb.sb_logsunit)
1506 ||
1507 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1508 roundoff < BBTOB(1)));
1509
1510 /* move grant heads by roundoff in sync */
Christoph Hellwig28496962012-02-20 02:31:25 +00001511 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1512 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513
1514 /* put cycle number in every block */
1515 xlog_pack_data(log, iclog, roundoff);
1516
1517 /* real byte length */
1518 if (v2) {
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001519 iclog->ic_header.h_len =
1520 cpu_to_be32(iclog->ic_offset + roundoff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 } else {
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001522 iclog->ic_header.h_len =
1523 cpu_to_be32(iclog->ic_offset);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 }
1525
Nathan Scottf5faad72006-07-28 17:04:44 +10001526 bp = iclog->ic_bp;
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001527 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528
1529 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1530
1531 /* Do we need to split this write into 2 parts? */
1532 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1533 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1534 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1535 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1536 } else {
1537 iclog->ic_bwritecnt = 1;
1538 }
Dave Chinneraa0e8832012-04-23 15:58:52 +10001539 bp->b_io_length = BTOBB(count);
Christoph Hellwigadadbee2011-07-13 13:43:49 +02001540 bp->b_fspriv = iclog;
Nathan Scottf5faad72006-07-28 17:04:44 +10001541 XFS_BUF_ZEROFLAGS(bp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 XFS_BUF_ASYNC(bp);
Christoph Hellwig1d5ae5d2011-07-08 14:36:32 +02001543 bp->b_flags |= XBF_SYNCIO;
Christoph Hellwig651701d2010-06-28 10:34:34 -04001544
Christoph Hellwiga27a2632011-06-16 12:02:23 +00001545 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
Christoph Hellwige163cbd2011-07-08 14:36:36 +02001546 bp->b_flags |= XBF_FUA;
1547
Christoph Hellwiga27a2632011-06-16 12:02:23 +00001548 /*
Christoph Hellwige163cbd2011-07-08 14:36:36 +02001549 * Flush the data device before flushing the log to make
1550 * sure all meta data written back from the AIL actually made
1551 * it to disk before stamping the new log tail LSN into the
1552 * log buffer. For an external log we need to issue the
1553 * flush explicitly, and unfortunately synchronously here;
1554 * for an internal log we can simply use the block layer
1555 * state machine for preflushes.
Christoph Hellwiga27a2632011-06-16 12:02:23 +00001556 */
1557 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1558 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
Christoph Hellwige163cbd2011-07-08 14:36:36 +02001559 else
1560 bp->b_flags |= XBF_FLUSH;
Christoph Hellwiga27a2632011-06-16 12:02:23 +00001561 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562
1563 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1564 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1565
1566 xlog_verify_iclog(log, iclog, count, B_TRUE);
1567
1568 /* account for log which doesn't start at block #0 */
1569 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1570 /*
1571 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1572 * is shutting down.
1573 */
1574 XFS_BUF_WRITE(bp);
1575
Christoph Hellwig901796a2011-10-10 16:52:49 +00001576 error = xlog_bdstrat(bp);
1577 if (error) {
1578 xfs_buf_ioerror_alert(bp, "xlog_sync");
Jesper Juhl014c2542006-01-15 02:37:08 +01001579 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 }
1581 if (split) {
Nathan Scottf5faad72006-07-28 17:04:44 +10001582 bp = iclog->ic_log->l_xbuf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
Chandra Seetharaman02fe03d2011-07-22 23:40:22 +00001584 xfs_buf_associate_memory(bp,
1585 (char *)&iclog->ic_header + count, split);
Christoph Hellwigadadbee2011-07-13 13:43:49 +02001586 bp->b_fspriv = iclog;
Nathan Scottf5faad72006-07-28 17:04:44 +10001587 XFS_BUF_ZEROFLAGS(bp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 XFS_BUF_ASYNC(bp);
Christoph Hellwig1d5ae5d2011-07-08 14:36:32 +02001589 bp->b_flags |= XBF_SYNCIO;
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11001590 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
Christoph Hellwige163cbd2011-07-08 14:36:36 +02001591 bp->b_flags |= XBF_FUA;
Chandra Seetharaman62926042011-07-22 23:40:15 +00001592 dptr = bp->b_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 /*
1594 * Bump the cycle numbers at the start of each block
1595 * since this part of the buffer is at the start of
1596 * a new cycle. Watch out for the header magic number
1597 * case, though.
1598 */
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001599 for (i = 0; i < split; i += BBSIZE) {
Marcin Slusarz413d57c2008-02-13 15:03:29 -08001600 be32_add_cpu((__be32 *)dptr, 1);
Christoph Hellwigb53e6752007-10-12 10:59:34 +10001601 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
Marcin Slusarz413d57c2008-02-13 15:03:29 -08001602 be32_add_cpu((__be32 *)dptr, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 dptr += BBSIZE;
1604 }
1605
1606 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1607 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1608
Nathan Scottc41564b2006-03-29 08:55:14 +10001609 /* account for internal log which doesn't start at block #0 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1611 XFS_BUF_WRITE(bp);
Christoph Hellwig901796a2011-10-10 16:52:49 +00001612 error = xlog_bdstrat(bp);
1613 if (error) {
1614 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
Jesper Juhl014c2542006-01-15 02:37:08 +01001615 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 }
1617 }
Jesper Juhl014c2542006-01-15 02:37:08 +01001618 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619} /* xlog_sync */
1620
1621
1622/*
Nathan Scottc41564b2006-03-29 08:55:14 +10001623 * Deallocate a log structure
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 */
David Chinnera8272ce2007-11-23 16:28:09 +11001625STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001626xlog_dealloc_log(
1627 struct xlog *log)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628{
1629 xlog_in_core_t *iclog, *next_iclog;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 int i;
1631
Dave Chinner71e330b2010-05-21 14:37:18 +10001632 xlog_cil_destroy(log);
1633
Dave Chinner44396472011-04-21 09:34:27 +00001634 /*
1635 * always need to ensure that the extra buffer does not point to memory
1636 * owned by another log buffer before we free it.
1637 */
Dave Chinnere70b73f2012-04-23 15:58:49 +10001638 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
Dave Chinner44396472011-04-21 09:34:27 +00001639 xfs_buf_free(log->l_xbuf);
1640
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 iclog = log->l_iclog;
1642 for (i=0; i<log->l_iclog_bufs; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 xfs_buf_free(iclog->ic_bp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 next_iclog = iclog->ic_next;
Denys Vlasenkof0e2d932008-05-19 16:31:57 +10001645 kmem_free(iclog);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 iclog = next_iclog;
1647 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 spinlock_destroy(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 log->l_mp->m_log = NULL;
Denys Vlasenkof0e2d932008-05-19 16:31:57 +10001651 kmem_free(log);
Nathan Scottc41564b2006-03-29 08:55:14 +10001652} /* xlog_dealloc_log */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653
1654/*
1655 * Update counters atomically now that memcpy is done.
1656 */
1657/* ARGSUSED */
1658static inline void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05001659xlog_state_finish_copy(
1660 struct xlog *log,
1661 struct xlog_in_core *iclog,
1662 int record_cnt,
1663 int copy_bytes)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664{
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10001665 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666
Marcin Slusarz413d57c2008-02-13 15:03:29 -08001667 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 iclog->ic_offset += copy_bytes;
1669
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10001670 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671} /* xlog_state_finish_copy */
1672
1673
1674
1675
1676/*
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001677 * print out info relating to regions written which consume
1678 * the reservation
1679 */
Dave Chinner71e330b2010-05-21 14:37:18 +10001680void
1681xlog_print_tic_res(
1682 struct xfs_mount *mp,
1683 struct xlog_ticket *ticket)
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001684{
1685 uint i;
1686 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1687
1688 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1689 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1690 "bformat",
1691 "bchunk",
1692 "efi_format",
1693 "efd_format",
1694 "iformat",
1695 "icore",
1696 "iext",
1697 "ibroot",
1698 "ilocal",
1699 "iattr_ext",
1700 "iattr_broot",
1701 "iattr_local",
1702 "qformat",
1703 "dquot",
1704 "quotaoff",
1705 "LR header",
1706 "unmount",
1707 "commit",
1708 "trans header"
1709 };
1710 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1711 "SETATTR_NOT_SIZE",
1712 "SETATTR_SIZE",
1713 "INACTIVE",
1714 "CREATE",
1715 "CREATE_TRUNC",
1716 "TRUNCATE_FILE",
1717 "REMOVE",
1718 "LINK",
1719 "RENAME",
1720 "MKDIR",
1721 "RMDIR",
1722 "SYMLINK",
1723 "SET_DMATTRS",
1724 "GROWFS",
1725 "STRAT_WRITE",
1726 "DIOSTRAT",
1727 "WRITE_SYNC",
1728 "WRITEID",
1729 "ADDAFORK",
1730 "ATTRINVAL",
1731 "ATRUNCATE",
1732 "ATTR_SET",
1733 "ATTR_RM",
1734 "ATTR_FLAG",
1735 "CLEAR_AGI_BUCKET",
1736 "QM_SBCHANGE",
1737 "DUMMY1",
1738 "DUMMY2",
1739 "QM_QUOTAOFF",
1740 "QM_DQALLOC",
1741 "QM_SETQLIM",
1742 "QM_DQCLUSTER",
1743 "QM_QINOCREATE",
1744 "QM_QUOTAOFF_END",
1745 "SB_UNIT",
1746 "FSYNC_TS",
1747 "GROWFSRT_ALLOC",
1748 "GROWFSRT_ZERO",
1749 "GROWFSRT_FREE",
1750 "SWAPEXT"
1751 };
1752
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001753 xfs_warn(mp,
Christoph Hellwig93b8a582011-12-06 21:58:07 +00001754 "xlog_write: reservation summary:\n"
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001755 " trans type = %s (%u)\n"
1756 " unit res = %d bytes\n"
1757 " current res = %d bytes\n"
1758 " total reg = %u bytes (o/flow = %u bytes)\n"
1759 " ophdrs = %u (ophdr space = %u bytes)\n"
1760 " ophdr + reg = %u bytes\n"
1761 " num regions = %u\n",
1762 ((ticket->t_trans_type <= 0 ||
1763 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1764 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1765 ticket->t_trans_type,
1766 ticket->t_unit_res,
1767 ticket->t_curr_res,
1768 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1769 ticket->t_res_num_ophdrs, ophdr_spc,
1770 ticket->t_res_arr_sum +
1771 ticket->t_res_o_flow + ophdr_spc,
1772 ticket->t_res_num);
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001773
1774 for (i = 0; i < ticket->t_res_num; i++) {
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001775 uint r_type = ticket->t_res_arr[i].r_type;
1776 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001777 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1778 "bad-rtype" : res_type_str[r_type-1]),
1779 ticket->t_res_arr[i].r_len);
1780 }
Dave Chinner169a7b02010-05-07 11:05:31 +10001781
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001782 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
Christoph Hellwig93b8a582011-12-06 21:58:07 +00001783 "xlog_write: reservation ran out. Need to up reservation");
Dave Chinner169a7b02010-05-07 11:05:31 +10001784 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001785}
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001786
1787/*
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001788 * Calculate the potential space needed by the log vector. Each region gets
1789 * its own xlog_op_header_t and may need to be double word aligned.
1790 */
1791static int
1792xlog_write_calc_vec_length(
1793 struct xlog_ticket *ticket,
Dave Chinner55b66332010-03-23 11:43:17 +11001794 struct xfs_log_vec *log_vector)
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001795{
Dave Chinner55b66332010-03-23 11:43:17 +11001796 struct xfs_log_vec *lv;
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001797 int headers = 0;
1798 int len = 0;
1799 int i;
1800
1801 /* acct for start rec of xact */
1802 if (ticket->t_flags & XLOG_TIC_INITED)
1803 headers++;
1804
Dave Chinner55b66332010-03-23 11:43:17 +11001805 for (lv = log_vector; lv; lv = lv->lv_next) {
1806 headers += lv->lv_niovecs;
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001807
Dave Chinner55b66332010-03-23 11:43:17 +11001808 for (i = 0; i < lv->lv_niovecs; i++) {
1809 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1810
1811 len += vecp->i_len;
1812 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1813 }
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001814 }
1815
1816 ticket->t_res_num_ophdrs += headers;
1817 len += headers * sizeof(struct xlog_op_header);
1818
1819 return len;
1820}
1821
1822/*
1823 * If first write for transaction, insert start record We can't be trying to
1824 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1825 */
1826static int
1827xlog_write_start_rec(
Christoph Hellwige6b1f272010-03-23 11:47:38 +11001828 struct xlog_op_header *ophdr,
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001829 struct xlog_ticket *ticket)
1830{
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001831 if (!(ticket->t_flags & XLOG_TIC_INITED))
1832 return 0;
1833
1834 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1835 ophdr->oh_clientid = ticket->t_clientid;
1836 ophdr->oh_len = 0;
1837 ophdr->oh_flags = XLOG_START_TRANS;
1838 ophdr->oh_res2 = 0;
1839
1840 ticket->t_flags &= ~XLOG_TIC_INITED;
1841
1842 return sizeof(struct xlog_op_header);
1843}
1844
1845static xlog_op_header_t *
1846xlog_write_setup_ophdr(
Mark Tinguelyad223e62012-06-14 09:22:15 -05001847 struct xlog *log,
Christoph Hellwige6b1f272010-03-23 11:47:38 +11001848 struct xlog_op_header *ophdr,
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001849 struct xlog_ticket *ticket,
1850 uint flags)
1851{
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001852 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1853 ophdr->oh_clientid = ticket->t_clientid;
1854 ophdr->oh_res2 = 0;
1855
1856 /* are we copying a commit or unmount record? */
1857 ophdr->oh_flags = flags;
1858
1859 /*
1860 * We've seen logs corrupted with bad transaction client ids. This
1861 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1862 * and shut down the filesystem.
1863 */
1864 switch (ophdr->oh_clientid) {
1865 case XFS_TRANSACTION:
1866 case XFS_VOLUME:
1867 case XFS_LOG:
1868 break;
1869 default:
Dave Chinnera0fa2b62011-03-07 10:01:35 +11001870 xfs_warn(log->l_mp,
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001871 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1872 ophdr->oh_clientid, ticket);
1873 return NULL;
1874 }
1875
1876 return ophdr;
1877}
1878
1879/*
1880 * Set up the parameters of the region copy into the log. This has
1881 * to handle region write split across multiple log buffers - this
1882 * state is kept external to this function so that this code can
1883 * can be written in an obvious, self documenting manner.
1884 */
1885static int
1886xlog_write_setup_copy(
1887 struct xlog_ticket *ticket,
1888 struct xlog_op_header *ophdr,
1889 int space_available,
1890 int space_required,
1891 int *copy_off,
1892 int *copy_len,
1893 int *last_was_partial_copy,
1894 int *bytes_consumed)
1895{
1896 int still_to_copy;
1897
1898 still_to_copy = space_required - *bytes_consumed;
1899 *copy_off = *bytes_consumed;
1900
1901 if (still_to_copy <= space_available) {
1902 /* write of region completes here */
1903 *copy_len = still_to_copy;
1904 ophdr->oh_len = cpu_to_be32(*copy_len);
1905 if (*last_was_partial_copy)
1906 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1907 *last_was_partial_copy = 0;
1908 *bytes_consumed = 0;
1909 return 0;
1910 }
1911
1912 /* partial write of region, needs extra log op header reservation */
1913 *copy_len = space_available;
1914 ophdr->oh_len = cpu_to_be32(*copy_len);
1915 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1916 if (*last_was_partial_copy)
1917 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1918 *bytes_consumed += *copy_len;
1919 (*last_was_partial_copy)++;
1920
1921 /* account for new log op header */
1922 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1923 ticket->t_res_num_ophdrs++;
1924
1925 return sizeof(struct xlog_op_header);
1926}
1927
1928static int
1929xlog_write_copy_finish(
Mark Tinguelyad223e62012-06-14 09:22:15 -05001930 struct xlog *log,
Dave Chinnerb5203cd2010-03-23 11:29:44 +11001931 struct xlog_in_core *iclog,
1932 uint flags,
1933 int *record_cnt,
1934 int *data_cnt,
1935 int *partial_copy,
1936 int *partial_copy_len,
1937 int log_offset,
1938 struct xlog_in_core **commit_iclog)
1939{
1940 if (*partial_copy) {
1941 /*
1942 * This iclog has already been marked WANT_SYNC by
1943 * xlog_state_get_iclog_space.
1944 */
1945 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1946 *record_cnt = 0;
1947 *data_cnt = 0;
1948 return xlog_state_release_iclog(log, iclog);
1949 }
1950
1951 *partial_copy = 0;
1952 *partial_copy_len = 0;
1953
1954 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1955 /* no more space in this iclog - push it. */
1956 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1957 *record_cnt = 0;
1958 *data_cnt = 0;
1959
1960 spin_lock(&log->l_icloglock);
1961 xlog_state_want_sync(log, iclog);
1962 spin_unlock(&log->l_icloglock);
1963
1964 if (!commit_iclog)
1965 return xlog_state_release_iclog(log, iclog);
1966 ASSERT(flags & XLOG_COMMIT_TRANS);
1967 *commit_iclog = iclog;
1968 }
1969
1970 return 0;
1971}
1972
1973/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974 * Write some region out to in-core log
1975 *
1976 * This will be called when writing externally provided regions or when
1977 * writing out a commit record for a given transaction.
1978 *
1979 * General algorithm:
1980 * 1. Find total length of this write. This may include adding to the
1981 * lengths passed in.
1982 * 2. Check whether we violate the tickets reservation.
1983 * 3. While writing to this iclog
1984 * A. Reserve as much space in this iclog as can get
1985 * B. If this is first write, save away start lsn
1986 * C. While writing this region:
1987 * 1. If first write of transaction, write start record
1988 * 2. Write log operation header (header per region)
1989 * 3. Find out if we can fit entire region into this iclog
1990 * 4. Potentially, verify destination memcpy ptr
1991 * 5. Memcpy (partial) region
1992 * 6. If partial copy, release iclog; otherwise, continue
1993 * copying more regions into current iclog
1994 * 4. Mark want sync bit (in simulation mode)
1995 * 5. Release iclog for potential flush to on-disk log.
1996 *
1997 * ERRORS:
1998 * 1. Panic if reservation is overrun. This should never happen since
1999 * reservation amounts are generated internal to the filesystem.
2000 * NOTES:
2001 * 1. Tickets are single threaded data structures.
2002 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2003 * syncing routine. When a single log_write region needs to span
2004 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2005 * on all log operation writes which don't contain the end of the
2006 * region. The XLOG_END_TRANS bit is used for the in-core log
2007 * operation which contains the end of the continued log_write region.
2008 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2009 * we don't really know exactly how much space will be used. As a result,
2010 * we don't update ic_offset until the end when we know exactly how many
2011 * bytes have been written out.
2012 */
Dave Chinner71e330b2010-05-21 14:37:18 +10002013int
Christoph Hellwig35a8a722010-02-15 23:34:54 +00002014xlog_write(
Mark Tinguelyad223e62012-06-14 09:22:15 -05002015 struct xlog *log,
Dave Chinner55b66332010-03-23 11:43:17 +11002016 struct xfs_log_vec *log_vector,
Christoph Hellwig35a8a722010-02-15 23:34:54 +00002017 struct xlog_ticket *ticket,
2018 xfs_lsn_t *start_lsn,
2019 struct xlog_in_core **commit_iclog,
2020 uint flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021{
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002022 struct xlog_in_core *iclog = NULL;
Dave Chinner55b66332010-03-23 11:43:17 +11002023 struct xfs_log_iovec *vecp;
2024 struct xfs_log_vec *lv;
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002025 int len;
2026 int index;
2027 int partial_copy = 0;
2028 int partial_copy_len = 0;
2029 int contwr = 0;
2030 int record_cnt = 0;
2031 int data_cnt = 0;
2032 int error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002034 *start_lsn = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035
Dave Chinner55b66332010-03-23 11:43:17 +11002036 len = xlog_write_calc_vec_length(ticket, log_vector);
Dave Chinner71e330b2010-05-21 14:37:18 +10002037
Christoph Hellwig93b8a582011-12-06 21:58:07 +00002038 /*
2039 * Region headers and bytes are already accounted for.
2040 * We only need to take into account start records and
2041 * split regions in this function.
2042 */
2043 if (ticket->t_flags & XLOG_TIC_INITED)
2044 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2045
2046 /*
2047 * Commit record headers need to be accounted for. These
2048 * come in as separate writes so are easy to detect.
2049 */
2050 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2051 ticket->t_curr_res -= sizeof(xlog_op_header_t);
Dave Chinner71e330b2010-05-21 14:37:18 +10002052
2053 if (ticket->t_curr_res < 0)
Dave Chinner55b66332010-03-23 11:43:17 +11002054 xlog_print_tic_res(log->l_mp, ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055
Dave Chinner55b66332010-03-23 11:43:17 +11002056 index = 0;
2057 lv = log_vector;
2058 vecp = lv->lv_iovecp;
2059 while (lv && index < lv->lv_niovecs) {
Christoph Hellwige6b1f272010-03-23 11:47:38 +11002060 void *ptr;
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002061 int log_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002063 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2064 &contwr, &log_offset);
2065 if (error)
2066 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002068 ASSERT(log_offset <= iclog->ic_size - 1);
Christoph Hellwige6b1f272010-03-23 11:47:38 +11002069 ptr = iclog->ic_datap + log_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002071 /* start_lsn is the first lsn written to. That's all we need. */
2072 if (!*start_lsn)
2073 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002075 /*
2076 * This loop writes out as many regions as can fit in the amount
2077 * of space which was allocated by xlog_state_get_iclog_space().
2078 */
Dave Chinner55b66332010-03-23 11:43:17 +11002079 while (lv && index < lv->lv_niovecs) {
2080 struct xfs_log_iovec *reg = &vecp[index];
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002081 struct xlog_op_header *ophdr;
2082 int start_rec_copy;
2083 int copy_len;
2084 int copy_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085
Dave Chinner55b66332010-03-23 11:43:17 +11002086 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
Christoph Hellwige6b1f272010-03-23 11:47:38 +11002087 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002089 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2090 if (start_rec_copy) {
2091 record_cnt++;
Christoph Hellwige6b1f272010-03-23 11:47:38 +11002092 xlog_write_adv_cnt(&ptr, &len, &log_offset,
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002093 start_rec_copy);
2094 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002096 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2097 if (!ophdr)
2098 return XFS_ERROR(EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099
Christoph Hellwige6b1f272010-03-23 11:47:38 +11002100 xlog_write_adv_cnt(&ptr, &len, &log_offset,
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002101 sizeof(struct xlog_op_header));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002103 len += xlog_write_setup_copy(ticket, ophdr,
2104 iclog->ic_size-log_offset,
Dave Chinner55b66332010-03-23 11:43:17 +11002105 reg->i_len,
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002106 &copy_off, &copy_len,
2107 &partial_copy,
2108 &partial_copy_len);
2109 xlog_verify_dest_ptr(log, ptr);
Dave Chinnerb5203cd2010-03-23 11:29:44 +11002110
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002111 /* copy region */
2112 ASSERT(copy_len >= 0);
Christoph Hellwige6b1f272010-03-23 11:47:38 +11002113 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2114 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
Dave Chinnerb5203cd2010-03-23 11:29:44 +11002115
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002116 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2117 record_cnt++;
2118 data_cnt += contwr ? copy_len : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002120 error = xlog_write_copy_finish(log, iclog, flags,
2121 &record_cnt, &data_cnt,
2122 &partial_copy,
2123 &partial_copy_len,
2124 log_offset,
2125 commit_iclog);
2126 if (error)
2127 return error;
2128
2129 /*
2130 * if we had a partial copy, we need to get more iclog
2131 * space but we don't want to increment the region
2132 * index because there is still more is this region to
2133 * write.
2134 *
2135 * If we completed writing this region, and we flushed
2136 * the iclog (indicated by resetting of the record
2137 * count), then we also need to get more log space. If
2138 * this was the last record, though, we are done and
2139 * can just return.
2140 */
2141 if (partial_copy)
2142 break;
2143
Dave Chinner55b66332010-03-23 11:43:17 +11002144 if (++index == lv->lv_niovecs) {
2145 lv = lv->lv_next;
2146 index = 0;
2147 if (lv)
2148 vecp = lv->lv_iovecp;
2149 }
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002150 if (record_cnt == 0) {
Dave Chinner55b66332010-03-23 11:43:17 +11002151 if (!lv)
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002152 return 0;
2153 break;
2154 }
2155 }
2156 }
2157
2158 ASSERT(len == 0);
2159
2160 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2161 if (!commit_iclog)
2162 return xlog_state_release_iclog(log, iclog);
2163
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 ASSERT(flags & XLOG_COMMIT_TRANS);
2165 *commit_iclog = iclog;
2166 return 0;
Christoph Hellwig99428ad2010-03-23 11:35:45 +11002167}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168
2169
2170/*****************************************************************************
2171 *
2172 * State Machine functions
2173 *
2174 *****************************************************************************
2175 */
2176
2177/* Clean iclogs starting from the head. This ordering must be
2178 * maintained, so an iclog doesn't become ACTIVE beyond one that
2179 * is SYNCING. This is also required to maintain the notion that we use
David Chinner12017fa2008-08-13 16:34:31 +10002180 * a ordered wait queue to hold off would be writers to the log when every
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 * iclog is trying to sync to disk.
2182 *
2183 * State Change: DIRTY -> ACTIVE
2184 */
Christoph Hellwigba0f32d2005-06-21 15:36:52 +10002185STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002186xlog_state_clean_log(
2187 struct xlog *log)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188{
2189 xlog_in_core_t *iclog;
2190 int changed = 0;
2191
2192 iclog = log->l_iclog;
2193 do {
2194 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2195 iclog->ic_state = XLOG_STATE_ACTIVE;
2196 iclog->ic_offset = 0;
David Chinner114d23a2008-04-10 12:18:39 +10002197 ASSERT(iclog->ic_callback == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 /*
2199 * If the number of ops in this iclog indicate it just
2200 * contains the dummy transaction, we can
2201 * change state into IDLE (the second time around).
2202 * Otherwise we should change the state into
2203 * NEED a dummy.
2204 * We don't need to cover the dummy.
2205 */
2206 if (!changed &&
Christoph Hellwigb53e6752007-10-12 10:59:34 +10002207 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2208 XLOG_COVER_OPS)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 changed = 1;
2210 } else {
2211 /*
2212 * We have two dirty iclogs so start over
2213 * This could also be num of ops indicates
2214 * this is not the dummy going out.
2215 */
2216 changed = 2;
2217 }
2218 iclog->ic_header.h_num_logops = 0;
2219 memset(iclog->ic_header.h_cycle_data, 0,
2220 sizeof(iclog->ic_header.h_cycle_data));
2221 iclog->ic_header.h_lsn = 0;
2222 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2223 /* do nothing */;
2224 else
2225 break; /* stop cleaning */
2226 iclog = iclog->ic_next;
2227 } while (iclog != log->l_iclog);
2228
2229 /* log is locked when we are called */
2230 /*
2231 * Change state for the dummy log recording.
2232 * We usually go to NEED. But we go to NEED2 if the changed indicates
2233 * we are done writing the dummy record.
2234 * If we are done with the second dummy recored (DONE2), then
2235 * we go to IDLE.
2236 */
2237 if (changed) {
2238 switch (log->l_covered_state) {
2239 case XLOG_STATE_COVER_IDLE:
2240 case XLOG_STATE_COVER_NEED:
2241 case XLOG_STATE_COVER_NEED2:
2242 log->l_covered_state = XLOG_STATE_COVER_NEED;
2243 break;
2244
2245 case XLOG_STATE_COVER_DONE:
2246 if (changed == 1)
2247 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2248 else
2249 log->l_covered_state = XLOG_STATE_COVER_NEED;
2250 break;
2251
2252 case XLOG_STATE_COVER_DONE2:
2253 if (changed == 1)
2254 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2255 else
2256 log->l_covered_state = XLOG_STATE_COVER_NEED;
2257 break;
2258
2259 default:
2260 ASSERT(0);
2261 }
2262 }
2263} /* xlog_state_clean_log */
2264
2265STATIC xfs_lsn_t
2266xlog_get_lowest_lsn(
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002267 struct xlog *log)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268{
2269 xlog_in_core_t *lsn_log;
2270 xfs_lsn_t lowest_lsn, lsn;
2271
2272 lsn_log = log->l_iclog;
2273 lowest_lsn = 0;
2274 do {
2275 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
Christoph Hellwigb53e6752007-10-12 10:59:34 +10002276 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 if ((lsn && !lowest_lsn) ||
2278 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2279 lowest_lsn = lsn;
2280 }
2281 }
2282 lsn_log = lsn_log->ic_next;
2283 } while (lsn_log != log->l_iclog);
Jesper Juhl014c2542006-01-15 02:37:08 +01002284 return lowest_lsn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285}
2286
2287
2288STATIC void
2289xlog_state_do_callback(
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002290 struct xlog *log,
2291 int aborted,
2292 struct xlog_in_core *ciclog)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293{
2294 xlog_in_core_t *iclog;
2295 xlog_in_core_t *first_iclog; /* used to know when we've
2296 * processed all iclogs once */
2297 xfs_log_callback_t *cb, *cb_next;
2298 int flushcnt = 0;
2299 xfs_lsn_t lowest_lsn;
2300 int ioerrors; /* counter: iclogs with errors */
2301 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2302 int funcdidcallbacks; /* flag: function did callbacks */
2303 int repeats; /* for issuing console warnings if
2304 * looping too many times */
Matthew Wilcoxd748c622008-05-19 16:34:27 +10002305 int wake = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002307 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 first_iclog = iclog = log->l_iclog;
2309 ioerrors = 0;
2310 funcdidcallbacks = 0;
2311 repeats = 0;
2312
2313 do {
2314 /*
2315 * Scan all iclogs starting with the one pointed to by the
2316 * log. Reset this starting point each time the log is
2317 * unlocked (during callbacks).
2318 *
2319 * Keep looping through iclogs until one full pass is made
2320 * without running any callbacks.
2321 */
2322 first_iclog = log->l_iclog;
2323 iclog = log->l_iclog;
2324 loopdidcallbacks = 0;
2325 repeats++;
2326
2327 do {
2328
2329 /* skip all iclogs in the ACTIVE & DIRTY states */
2330 if (iclog->ic_state &
2331 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2332 iclog = iclog->ic_next;
2333 continue;
2334 }
2335
2336 /*
2337 * Between marking a filesystem SHUTDOWN and stopping
2338 * the log, we do flush all iclogs to disk (if there
2339 * wasn't a log I/O error). So, we do want things to
2340 * go smoothly in case of just a SHUTDOWN w/o a
2341 * LOG_IO_ERROR.
2342 */
2343 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2344 /*
2345 * Can only perform callbacks in order. Since
2346 * this iclog is not in the DONE_SYNC/
2347 * DO_CALLBACK state, we skip the rest and
2348 * just try to clean up. If we set our iclog
2349 * to DO_CALLBACK, we will not process it when
2350 * we retry since a previous iclog is in the
2351 * CALLBACK and the state cannot change since
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002352 * we are holding the l_icloglock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 */
2354 if (!(iclog->ic_state &
2355 (XLOG_STATE_DONE_SYNC |
2356 XLOG_STATE_DO_CALLBACK))) {
2357 if (ciclog && (ciclog->ic_state ==
2358 XLOG_STATE_DONE_SYNC)) {
2359 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2360 }
2361 break;
2362 }
2363 /*
2364 * We now have an iclog that is in either the
2365 * DO_CALLBACK or DONE_SYNC states. The other
2366 * states (WANT_SYNC, SYNCING, or CALLBACK were
2367 * caught by the above if and are going to
2368 * clean (i.e. we aren't doing their callbacks)
2369 * see the above if.
2370 */
2371
2372 /*
2373 * We will do one more check here to see if we
2374 * have chased our tail around.
2375 */
2376
2377 lowest_lsn = xlog_get_lowest_lsn(log);
Christoph Hellwigb53e6752007-10-12 10:59:34 +10002378 if (lowest_lsn &&
2379 XFS_LSN_CMP(lowest_lsn,
Dave Chinner84f3c682010-12-03 22:11:29 +11002380 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 iclog = iclog->ic_next;
2382 continue; /* Leave this iclog for
2383 * another thread */
2384 }
2385
2386 iclog->ic_state = XLOG_STATE_CALLBACK;
2387
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
Dave Chinner84f3c682010-12-03 22:11:29 +11002389 /*
Dave Chinner7e9620f2012-10-08 21:56:12 +11002390 * Completion of a iclog IO does not imply that
2391 * a transaction has completed, as transactions
2392 * can be large enough to span many iclogs. We
2393 * cannot change the tail of the log half way
2394 * through a transaction as this may be the only
2395 * transaction in the log and moving th etail to
2396 * point to the middle of it will prevent
2397 * recovery from finding the start of the
2398 * transaction. Hence we should only update the
2399 * last_sync_lsn if this iclog contains
2400 * transaction completion callbacks on it.
2401 *
2402 * We have to do this before we drop the
Dave Chinner84f3c682010-12-03 22:11:29 +11002403 * icloglock to ensure we are the only one that
2404 * can update it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 */
Dave Chinner84f3c682010-12-03 22:11:29 +11002406 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2407 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
Dave Chinner7e9620f2012-10-08 21:56:12 +11002408 if (iclog->ic_callback)
2409 atomic64_set(&log->l_last_sync_lsn,
2410 be64_to_cpu(iclog->ic_header.h_lsn));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411
Dave Chinner84f3c682010-12-03 22:11:29 +11002412 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 ioerrors++;
Dave Chinner84f3c682010-12-03 22:11:29 +11002414
2415 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416
David Chinner114d23a2008-04-10 12:18:39 +10002417 /*
2418 * Keep processing entries in the callback list until
2419 * we come around and it is empty. We need to
2420 * atomically see that the list is empty and change the
2421 * state to DIRTY so that we don't miss any more
2422 * callbacks being added.
2423 */
2424 spin_lock(&iclog->ic_callback_lock);
2425 cb = iclog->ic_callback;
Christoph Hellwig4b809162007-08-16 15:37:36 +10002426 while (cb) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 iclog->ic_callback_tail = &(iclog->ic_callback);
2428 iclog->ic_callback = NULL;
David Chinner114d23a2008-04-10 12:18:39 +10002429 spin_unlock(&iclog->ic_callback_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430
2431 /* perform callbacks in the order given */
Christoph Hellwig4b809162007-08-16 15:37:36 +10002432 for (; cb; cb = cb_next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433 cb_next = cb->cb_next;
2434 cb->cb_func(cb->cb_arg, aborted);
2435 }
David Chinner114d23a2008-04-10 12:18:39 +10002436 spin_lock(&iclog->ic_callback_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 cb = iclog->ic_callback;
2438 }
2439
2440 loopdidcallbacks++;
2441 funcdidcallbacks++;
2442
David Chinner114d23a2008-04-10 12:18:39 +10002443 spin_lock(&log->l_icloglock);
Christoph Hellwig4b809162007-08-16 15:37:36 +10002444 ASSERT(iclog->ic_callback == NULL);
David Chinner114d23a2008-04-10 12:18:39 +10002445 spin_unlock(&iclog->ic_callback_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2447 iclog->ic_state = XLOG_STATE_DIRTY;
2448
2449 /*
2450 * Transition from DIRTY to ACTIVE if applicable.
2451 * NOP if STATE_IOERROR.
2452 */
2453 xlog_state_clean_log(log);
2454
2455 /* wake up threads waiting in xfs_log_force() */
Dave Chinnereb40a872010-12-21 12:09:01 +11002456 wake_up_all(&iclog->ic_force_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
2458 iclog = iclog->ic_next;
2459 } while (first_iclog != iclog);
Nathan Scotta3c6685e2006-09-28 11:02:14 +10002460
2461 if (repeats > 5000) {
2462 flushcnt += repeats;
2463 repeats = 0;
Dave Chinnera0fa2b62011-03-07 10:01:35 +11002464 xfs_warn(log->l_mp,
Nathan Scotta3c6685e2006-09-28 11:02:14 +10002465 "%s: possible infinite loop (%d iterations)",
Harvey Harrison34a622b2008-04-10 12:19:21 +10002466 __func__, flushcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 }
2468 } while (!ioerrors && loopdidcallbacks);
2469
2470 /*
2471 * make one last gasp attempt to see if iclogs are being left in
2472 * limbo..
2473 */
2474#ifdef DEBUG
2475 if (funcdidcallbacks) {
2476 first_iclog = iclog = log->l_iclog;
2477 do {
2478 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2479 /*
2480 * Terminate the loop if iclogs are found in states
2481 * which will cause other threads to clean up iclogs.
2482 *
2483 * SYNCING - i/o completion will go through logs
2484 * DONE_SYNC - interrupt thread should be waiting for
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002485 * l_icloglock
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 * IOERROR - give up hope all ye who enter here
2487 */
2488 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2489 iclog->ic_state == XLOG_STATE_SYNCING ||
2490 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2491 iclog->ic_state == XLOG_STATE_IOERROR )
2492 break;
2493 iclog = iclog->ic_next;
2494 } while (first_iclog != iclog);
2495 }
2496#endif
2497
Matthew Wilcoxd748c622008-05-19 16:34:27 +10002498 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2499 wake = 1;
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002500 spin_unlock(&log->l_icloglock);
Matthew Wilcoxd748c622008-05-19 16:34:27 +10002501
2502 if (wake)
Dave Chinnereb40a872010-12-21 12:09:01 +11002503 wake_up_all(&log->l_flush_wait);
Matthew Wilcoxd748c622008-05-19 16:34:27 +10002504}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505
2506
2507/*
2508 * Finish transitioning this iclog to the dirty state.
2509 *
2510 * Make sure that we completely execute this routine only when this is
2511 * the last call to the iclog. There is a good chance that iclog flushes,
2512 * when we reach the end of the physical log, get turned into 2 separate
2513 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2514 * routine. By using the reference count bwritecnt, we guarantee that only
2515 * the second completion goes through.
2516 *
2517 * Callbacks could take time, so they are done outside the scope of the
David Chinner12017fa2008-08-13 16:34:31 +10002518 * global state machine log lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 */
David Chinnera8272ce2007-11-23 16:28:09 +11002520STATIC void
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521xlog_state_done_syncing(
2522 xlog_in_core_t *iclog,
2523 int aborted)
2524{
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002525 struct xlog *log = iclog->ic_log;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002527 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528
2529 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2530 iclog->ic_state == XLOG_STATE_IOERROR);
David Chinner155cc6b2008-03-06 13:44:14 +11002531 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2533
2534
2535 /*
2536 * If we got an error, either on the first buffer, or in the case of
2537 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2538 * and none should ever be attempted to be written to disk
2539 * again.
2540 */
2541 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2542 if (--iclog->ic_bwritecnt == 1) {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002543 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 return;
2545 }
2546 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2547 }
2548
2549 /*
2550 * Someone could be sleeping prior to writing out the next
2551 * iclog buffer, we wake them all, one will get to do the
2552 * I/O, the others get to wait for the result.
2553 */
Dave Chinnereb40a872010-12-21 12:09:01 +11002554 wake_up_all(&iclog->ic_write_wait);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002555 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2557} /* xlog_state_done_syncing */
2558
2559
2560/*
2561 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
David Chinner12017fa2008-08-13 16:34:31 +10002562 * sleep. We wait on the flush queue on the head iclog as that should be
2563 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2564 * we will wait here and all new writes will sleep until a sync completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 *
2566 * The in-core logs are used in a circular fashion. They are not used
2567 * out-of-order even when an iclog past the head is free.
2568 *
2569 * return:
2570 * * log_offset where xlog_write() can start writing into the in-core
2571 * log's data space.
2572 * * in-core log pointer to which xlog_write() should write.
2573 * * boolean indicating this is a continued write to an in-core log.
2574 * If this is the last write, then the in-core log's offset field
2575 * needs to be incremented, depending on the amount of data which
2576 * is copied.
2577 */
David Chinnera8272ce2007-11-23 16:28:09 +11002578STATIC int
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002579xlog_state_get_iclog_space(
2580 struct xlog *log,
2581 int len,
2582 struct xlog_in_core **iclogp,
2583 struct xlog_ticket *ticket,
2584 int *continued_write,
2585 int *logoffsetp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587 int log_offset;
2588 xlog_rec_header_t *head;
2589 xlog_in_core_t *iclog;
2590 int error;
2591
2592restart:
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002593 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 if (XLOG_FORCED_SHUTDOWN(log)) {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002595 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 return XFS_ERROR(EIO);
2597 }
2598
2599 iclog = log->l_iclog;
Matthew Wilcoxd748c622008-05-19 16:34:27 +10002600 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 XFS_STATS_INC(xs_log_noiclogs);
Matthew Wilcoxd748c622008-05-19 16:34:27 +10002602
2603 /* Wait for log writes to have flushed */
Dave Chinnereb40a872010-12-21 12:09:01 +11002604 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 goto restart;
2606 }
Matthew Wilcoxd748c622008-05-19 16:34:27 +10002607
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 head = &iclog->ic_header;
2609
David Chinner155cc6b2008-03-06 13:44:14 +11002610 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 log_offset = iclog->ic_offset;
2612
2613 /* On the 1st write to an iclog, figure out lsn. This works
2614 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2615 * committing to. If the offset is set, that's how many blocks
2616 * must be written.
2617 */
2618 if (log_offset == 0) {
2619 ticket->t_curr_res -= log->l_iclog_hsize;
Christoph Hellwig0adba532007-08-30 17:21:46 +10002620 xlog_tic_add_region(ticket,
Tim Shimmin7e9c6392005-09-02 16:42:05 +10002621 log->l_iclog_hsize,
2622 XLOG_REG_TYPE_LRHEADER);
Christoph Hellwigb53e6752007-10-12 10:59:34 +10002623 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2624 head->h_lsn = cpu_to_be64(
Christoph Hellwig03bea6f2007-10-12 10:58:05 +10002625 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 ASSERT(log->l_curr_block >= 0);
2627 }
2628
2629 /* If there is enough room to write everything, then do it. Otherwise,
2630 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2631 * bit is on, so this will get flushed out. Don't update ic_offset
2632 * until you know exactly how many bytes get copied. Therefore, wait
2633 * until later to update ic_offset.
2634 *
2635 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2636 * can fit into remaining data section.
2637 */
2638 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2639 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2640
Dave Chinner49641f12008-07-11 17:43:55 +10002641 /*
2642 * If I'm the only one writing to this iclog, sync it to disk.
2643 * We need to do an atomic compare and decrement here to avoid
2644 * racing with concurrent atomic_dec_and_lock() calls in
2645 * xlog_state_release_iclog() when there is more than one
2646 * reference to the iclog.
2647 */
2648 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2649 /* we are the only one */
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002650 spin_unlock(&log->l_icloglock);
Dave Chinner49641f12008-07-11 17:43:55 +10002651 error = xlog_state_release_iclog(log, iclog);
2652 if (error)
Jesper Juhl014c2542006-01-15 02:37:08 +01002653 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 } else {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002655 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 }
2657 goto restart;
2658 }
2659
2660 /* Do we have enough room to write the full amount in the remainder
2661 * of this iclog? Or must we continue a write on the next iclog and
2662 * mark this iclog as completely taken? In the case where we switch
2663 * iclogs (to mark it taken), this particular iclog will release/sync
2664 * to disk in xlog_write().
2665 */
2666 if (len <= iclog->ic_size - iclog->ic_offset) {
2667 *continued_write = 0;
2668 iclog->ic_offset += len;
2669 } else {
2670 *continued_write = 1;
2671 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2672 }
2673 *iclogp = iclog;
2674
2675 ASSERT(iclog->ic_offset <= iclog->ic_size);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002676 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677
2678 *logoffsetp = log_offset;
2679 return 0;
2680} /* xlog_state_get_iclog_space */
2681
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682/* The first cnt-1 times through here we don't need to
2683 * move the grant write head because the permanent
2684 * reservation has reserved cnt times the unit amount.
2685 * Release part of current permanent unit reservation and
2686 * reset current reservation to be one units worth. Also
2687 * move grant reservation head forward.
2688 */
2689STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002690xlog_regrant_reserve_log_space(
2691 struct xlog *log,
2692 struct xlog_ticket *ticket)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693{
Christoph Hellwig0b1b2132009-12-14 23:14:59 +00002694 trace_xfs_log_regrant_reserve_enter(log, ticket);
2695
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696 if (ticket->t_cnt > 0)
2697 ticket->t_cnt--;
2698
Christoph Hellwig28496962012-02-20 02:31:25 +00002699 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
Dave Chinnera69ed032010-12-21 12:08:20 +11002700 ticket->t_curr_res);
Christoph Hellwig28496962012-02-20 02:31:25 +00002701 xlog_grant_sub_space(log, &log->l_write_head.grant,
Dave Chinnera69ed032010-12-21 12:08:20 +11002702 ticket->t_curr_res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 ticket->t_curr_res = ticket->t_unit_res;
Christoph Hellwig0adba532007-08-30 17:21:46 +10002704 xlog_tic_reset_res(ticket);
Christoph Hellwig0b1b2132009-12-14 23:14:59 +00002705
2706 trace_xfs_log_regrant_reserve_sub(log, ticket);
2707
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 /* just return if we still have some of the pre-reserved space */
Dave Chinnerd0eb2f32010-12-21 12:29:14 +11002709 if (ticket->t_cnt > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711
Christoph Hellwig28496962012-02-20 02:31:25 +00002712 xlog_grant_add_space(log, &log->l_reserve_head.grant,
Dave Chinnera69ed032010-12-21 12:08:20 +11002713 ticket->t_unit_res);
Christoph Hellwig0b1b2132009-12-14 23:14:59 +00002714
2715 trace_xfs_log_regrant_reserve_exit(log, ticket);
2716
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 ticket->t_curr_res = ticket->t_unit_res;
Christoph Hellwig0adba532007-08-30 17:21:46 +10002718 xlog_tic_reset_res(ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719} /* xlog_regrant_reserve_log_space */
2720
2721
2722/*
2723 * Give back the space left from a reservation.
2724 *
2725 * All the information we need to make a correct determination of space left
2726 * is present. For non-permanent reservations, things are quite easy. The
2727 * count should have been decremented to zero. We only need to deal with the
2728 * space remaining in the current reservation part of the ticket. If the
2729 * ticket contains a permanent reservation, there may be left over space which
2730 * needs to be released. A count of N means that N-1 refills of the current
2731 * reservation can be done before we need to ask for more space. The first
2732 * one goes to fill up the first current reservation. Once we run out of
2733 * space, the count will stay at zero and the only space remaining will be
2734 * in the current reservation field.
2735 */
2736STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002737xlog_ungrant_log_space(
2738 struct xlog *log,
2739 struct xlog_ticket *ticket)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740{
Dave Chinner663e4962010-12-21 12:06:05 +11002741 int bytes;
2742
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 if (ticket->t_cnt > 0)
2744 ticket->t_cnt--;
2745
Christoph Hellwig0b1b2132009-12-14 23:14:59 +00002746 trace_xfs_log_ungrant_enter(log, ticket);
Christoph Hellwig0b1b2132009-12-14 23:14:59 +00002747 trace_xfs_log_ungrant_sub(log, ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748
Dave Chinner663e4962010-12-21 12:06:05 +11002749 /*
2750 * If this is a permanent reservation ticket, we may be able to free
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 * up more space based on the remaining count.
2752 */
Dave Chinner663e4962010-12-21 12:06:05 +11002753 bytes = ticket->t_curr_res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 if (ticket->t_cnt > 0) {
2755 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
Dave Chinner663e4962010-12-21 12:06:05 +11002756 bytes += ticket->t_unit_res*ticket->t_cnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 }
2758
Christoph Hellwig28496962012-02-20 02:31:25 +00002759 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2760 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
Dave Chinner663e4962010-12-21 12:06:05 +11002761
Christoph Hellwig0b1b2132009-12-14 23:14:59 +00002762 trace_xfs_log_ungrant_exit(log, ticket);
2763
Christoph Hellwigcfb7cdc2012-02-20 02:31:23 +00002764 xfs_log_space_wake(log->l_mp);
Christoph Hellwig09a423a2012-02-20 02:31:20 +00002765}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
2767/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 * Flush iclog to disk if this is the last reference to the given iclog and
2769 * the WANT_SYNC bit is set.
2770 *
2771 * When this function is entered, the iclog is not necessarily in the
2772 * WANT_SYNC state. It may be sitting around waiting to get filled.
2773 *
2774 *
2775 */
David Chinnera8272ce2007-11-23 16:28:09 +11002776STATIC int
David Chinnerb5893342008-03-06 13:44:06 +11002777xlog_state_release_iclog(
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002778 struct xlog *log,
2779 struct xlog_in_core *iclog)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 int sync = 0; /* do we sync? */
2782
David Chinner155cc6b2008-03-06 13:44:14 +11002783 if (iclog->ic_state & XLOG_STATE_IOERROR)
2784 return XFS_ERROR(EIO);
2785
2786 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2787 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2788 return 0;
2789
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 if (iclog->ic_state & XLOG_STATE_IOERROR) {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002791 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 return XFS_ERROR(EIO);
2793 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2795 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2796
David Chinner155cc6b2008-03-06 13:44:14 +11002797 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
David Chinnerb5893342008-03-06 13:44:06 +11002798 /* update tail before writing to iclog */
Dave Chinner1c3cb9e2010-12-21 12:28:39 +11002799 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 sync++;
2801 iclog->ic_state = XLOG_STATE_SYNCING;
Dave Chinner1c3cb9e2010-12-21 12:28:39 +11002802 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2803 xlog_verify_tail_lsn(log, iclog, tail_lsn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 /* cycle incremented when incrementing curr_block */
2805 }
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002806 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807
2808 /*
2809 * We let the log lock go, so it's possible that we hit a log I/O
Nathan Scottc41564b2006-03-29 08:55:14 +10002810 * error or some other SHUTDOWN condition that marks the iclog
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2812 * this iclog has consistent data, so we ignore IOERROR
2813 * flags after this point.
2814 */
David Chinnerb5893342008-03-06 13:44:06 +11002815 if (sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 return xlog_sync(log, iclog);
Jesper Juhl014c2542006-01-15 02:37:08 +01002817 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818} /* xlog_state_release_iclog */
2819
2820
2821/*
2822 * This routine will mark the current iclog in the ring as WANT_SYNC
2823 * and move the current iclog pointer to the next iclog in the ring.
2824 * When this routine is called from xlog_state_get_iclog_space(), the
2825 * exact size of the iclog has not yet been determined. All we know is
2826 * that every data block. We have run out of space in this log record.
2827 */
2828STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05002829xlog_state_switch_iclogs(
2830 struct xlog *log,
2831 struct xlog_in_core *iclog,
2832 int eventual_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833{
2834 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2835 if (!eventual_size)
2836 eventual_size = iclog->ic_offset;
2837 iclog->ic_state = XLOG_STATE_WANT_SYNC;
Christoph Hellwigb53e6752007-10-12 10:59:34 +10002838 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 log->l_prev_block = log->l_curr_block;
2840 log->l_prev_cycle = log->l_curr_cycle;
2841
2842 /* roll log?: ic_offset changed later */
2843 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2844
2845 /* Round up to next log-sunit */
Eric Sandeen62118702008-03-06 13:44:28 +11002846 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 log->l_mp->m_sb.sb_logsunit > 1) {
2848 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2849 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2850 }
2851
2852 if (log->l_curr_block >= log->l_logBBsize) {
2853 log->l_curr_cycle++;
2854 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2855 log->l_curr_cycle++;
2856 log->l_curr_block -= log->l_logBBsize;
2857 ASSERT(log->l_curr_block >= 0);
2858 }
2859 ASSERT(iclog == log->l_iclog);
2860 log->l_iclog = iclog->ic_next;
2861} /* xlog_state_switch_iclogs */
2862
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863/*
2864 * Write out all data in the in-core log as of this exact moment in time.
2865 *
2866 * Data may be written to the in-core log during this call. However,
2867 * we don't guarantee this data will be written out. A change from past
2868 * implementation means this routine will *not* write out zero length LRs.
2869 *
2870 * Basically, we try and perform an intelligent scan of the in-core logs.
2871 * If we determine there is no flushable data, we just return. There is no
2872 * flushable data if:
2873 *
2874 * 1. the current iclog is active and has no data; the previous iclog
2875 * is in the active or dirty state.
2876 * 2. the current iclog is drity, and the previous iclog is in the
2877 * active or dirty state.
2878 *
David Chinner12017fa2008-08-13 16:34:31 +10002879 * We may sleep if:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 *
2881 * 1. the current iclog is not in the active nor dirty state.
2882 * 2. the current iclog dirty, and the previous iclog is not in the
2883 * active nor dirty state.
2884 * 3. the current iclog is active, and there is another thread writing
2885 * to this particular iclog.
2886 * 4. a) the current iclog is active and has no other writers
2887 * b) when we return from flushing out this iclog, it is still
2888 * not in the active nor dirty state.
2889 */
Christoph Hellwiga14a3482010-01-19 09:56:46 +00002890int
2891_xfs_log_force(
2892 struct xfs_mount *mp,
2893 uint flags,
2894 int *log_flushed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895{
Mark Tinguelyad223e62012-06-14 09:22:15 -05002896 struct xlog *log = mp->m_log;
Christoph Hellwiga14a3482010-01-19 09:56:46 +00002897 struct xlog_in_core *iclog;
2898 xfs_lsn_t lsn;
2899
2900 XFS_STATS_INC(xs_log_force);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901
Christoph Hellwig93b8a582011-12-06 21:58:07 +00002902 xlog_cil_force(log);
Dave Chinner71e330b2010-05-21 14:37:18 +10002903
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002904 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905
2906 iclog = log->l_iclog;
2907 if (iclog->ic_state & XLOG_STATE_IOERROR) {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002908 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 return XFS_ERROR(EIO);
2910 }
2911
2912 /* If the head iclog is not active nor dirty, we just attach
2913 * ourselves to the head and go to sleep.
2914 */
2915 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2916 iclog->ic_state == XLOG_STATE_DIRTY) {
2917 /*
2918 * If the head is dirty or (active and empty), then
2919 * we need to look at the previous iclog. If the previous
2920 * iclog is active or dirty we are done. There is nothing
2921 * to sync out. Otherwise, we attach ourselves to the
2922 * previous iclog and go to sleep.
2923 */
2924 if (iclog->ic_state == XLOG_STATE_DIRTY ||
David Chinner155cc6b2008-03-06 13:44:14 +11002925 (atomic_read(&iclog->ic_refcnt) == 0
2926 && iclog->ic_offset == 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927 iclog = iclog->ic_prev;
2928 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2929 iclog->ic_state == XLOG_STATE_DIRTY)
2930 goto no_sleep;
2931 else
2932 goto maybe_sleep;
2933 } else {
David Chinner155cc6b2008-03-06 13:44:14 +11002934 if (atomic_read(&iclog->ic_refcnt) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 /* We are the only one with access to this
2936 * iclog. Flush it out now. There should
2937 * be a roundoff of zero to show that someone
2938 * has already taken care of the roundoff from
2939 * the previous sync.
2940 */
David Chinner155cc6b2008-03-06 13:44:14 +11002941 atomic_inc(&iclog->ic_refcnt);
Christoph Hellwigb53e6752007-10-12 10:59:34 +10002942 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 xlog_state_switch_iclogs(log, iclog, 0);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002944 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945
2946 if (xlog_state_release_iclog(log, iclog))
2947 return XFS_ERROR(EIO);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00002948
2949 if (log_flushed)
2950 *log_flushed = 1;
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002951 spin_lock(&log->l_icloglock);
Christoph Hellwigb53e6752007-10-12 10:59:34 +10002952 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 iclog->ic_state != XLOG_STATE_DIRTY)
2954 goto maybe_sleep;
2955 else
2956 goto no_sleep;
2957 } else {
2958 /* Someone else is writing to this iclog.
2959 * Use its call to flush out the data. However,
2960 * the other thread may not force out this LR,
2961 * so we mark it WANT_SYNC.
2962 */
2963 xlog_state_switch_iclogs(log, iclog, 0);
2964 goto maybe_sleep;
2965 }
2966 }
2967 }
2968
2969 /* By the time we come around again, the iclog could've been filled
2970 * which would give it another lsn. If we have a new lsn, just
2971 * return because the relevant data has been flushed.
2972 */
2973maybe_sleep:
2974 if (flags & XFS_LOG_SYNC) {
2975 /*
2976 * We must check if we're shutting down here, before
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002977 * we wait, while we're holding the l_icloglock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 * Then we check again after waking up, in case our
2979 * sleep was disturbed by a bad news.
2980 */
2981 if (iclog->ic_state & XLOG_STATE_IOERROR) {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002982 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983 return XFS_ERROR(EIO);
2984 }
2985 XFS_STATS_INC(xs_log_force_sleep);
Dave Chinnereb40a872010-12-21 12:09:01 +11002986 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 /*
2988 * No need to grab the log lock here since we're
2989 * only deciding whether or not to return EIO
2990 * and the memory read should be atomic.
2991 */
2992 if (iclog->ic_state & XLOG_STATE_IOERROR)
2993 return XFS_ERROR(EIO);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00002994 if (log_flushed)
2995 *log_flushed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 } else {
2997
2998no_sleep:
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10002999 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000 }
3001 return 0;
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003002}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003
3004/*
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003005 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3006 * about errors or whether the log was flushed or not. This is the normal
3007 * interface to use when trying to unpin items or move the log forward.
3008 */
3009void
3010xfs_log_force(
3011 xfs_mount_t *mp,
3012 uint flags)
3013{
3014 int error;
3015
Dave Chinner14c26c62012-04-24 16:33:31 +10003016 trace_xfs_log_force(mp, 0);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003017 error = _xfs_log_force(mp, flags, NULL);
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003018 if (error)
3019 xfs_warn(mp, "%s: error %d returned.", __func__, error);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003020}
3021
3022/*
3023 * Force the in-core log to disk for a specific LSN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 *
3025 * Find in-core log with lsn.
3026 * If it is in the DIRTY state, just return.
3027 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3028 * state and go to sleep or return.
3029 * If it is in any other state, go to sleep or return.
3030 *
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003031 * Synchronous forces are implemented with a signal variable. All callers
3032 * to force a given lsn to disk will wait on a the sv attached to the
3033 * specific in-core log. When given in-core log finally completes its
3034 * write to disk, that thread will wake up all threads waiting on the
3035 * sv.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 */
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003037int
3038_xfs_log_force_lsn(
3039 struct xfs_mount *mp,
3040 xfs_lsn_t lsn,
3041 uint flags,
3042 int *log_flushed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043{
Mark Tinguelyad223e62012-06-14 09:22:15 -05003044 struct xlog *log = mp->m_log;
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003045 struct xlog_in_core *iclog;
3046 int already_slept = 0;
3047
3048 ASSERT(lsn != 0);
3049
3050 XFS_STATS_INC(xs_log_force);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051
Christoph Hellwig93b8a582011-12-06 21:58:07 +00003052 lsn = xlog_cil_force_lsn(log, lsn);
3053 if (lsn == NULLCOMMITLSN)
3054 return 0;
Dave Chinner71e330b2010-05-21 14:37:18 +10003055
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056try_again:
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003057 spin_lock(&log->l_icloglock);
3058 iclog = log->l_iclog;
3059 if (iclog->ic_state & XLOG_STATE_IOERROR) {
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003060 spin_unlock(&log->l_icloglock);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003061 return XFS_ERROR(EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 }
3063
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003064 do {
3065 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3066 iclog = iclog->ic_next;
3067 continue;
3068 }
3069
3070 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3071 spin_unlock(&log->l_icloglock);
3072 return 0;
3073 }
3074
3075 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3076 /*
3077 * We sleep here if we haven't already slept (e.g.
3078 * this is the first time we've looked at the correct
3079 * iclog buf) and the buffer before us is going to
3080 * be sync'ed. The reason for this is that if we
3081 * are doing sync transactions here, by waiting for
3082 * the previous I/O to complete, we can allow a few
3083 * more transactions into this iclog before we close
3084 * it down.
3085 *
3086 * Otherwise, we mark the buffer WANT_SYNC, and bump
3087 * up the refcnt so we can release the log (which
3088 * drops the ref count). The state switch keeps new
3089 * transaction commits from using this buffer. When
3090 * the current commits finish writing into the buffer,
3091 * the refcount will drop to zero and the buffer will
3092 * go out then.
3093 */
3094 if (!already_slept &&
3095 (iclog->ic_prev->ic_state &
3096 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3097 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3098
3099 XFS_STATS_INC(xs_log_force_sleep);
3100
Dave Chinnereb40a872010-12-21 12:09:01 +11003101 xlog_wait(&iclog->ic_prev->ic_write_wait,
3102 &log->l_icloglock);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003103 if (log_flushed)
3104 *log_flushed = 1;
3105 already_slept = 1;
3106 goto try_again;
3107 }
David Chinner155cc6b2008-03-06 13:44:14 +11003108 atomic_inc(&iclog->ic_refcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 xlog_state_switch_iclogs(log, iclog, 0);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003110 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111 if (xlog_state_release_iclog(log, iclog))
3112 return XFS_ERROR(EIO);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003113 if (log_flushed)
3114 *log_flushed = 1;
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003115 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003118 if ((flags & XFS_LOG_SYNC) && /* sleep */
3119 !(iclog->ic_state &
3120 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3121 /*
3122 * Don't wait on completion if we know that we've
3123 * gotten a log write error.
3124 */
3125 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3126 spin_unlock(&log->l_icloglock);
3127 return XFS_ERROR(EIO);
3128 }
3129 XFS_STATS_INC(xs_log_force_sleep);
Dave Chinnereb40a872010-12-21 12:09:01 +11003130 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003131 /*
3132 * No need to grab the log lock here since we're
3133 * only deciding whether or not to return EIO
3134 * and the memory read should be atomic.
3135 */
3136 if (iclog->ic_state & XLOG_STATE_IOERROR)
3137 return XFS_ERROR(EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003139 if (log_flushed)
3140 *log_flushed = 1;
3141 } else { /* just return */
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003142 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 }
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003144
3145 return 0;
3146 } while (iclog != log->l_iclog);
3147
3148 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149 return 0;
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003150}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003151
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003152/*
3153 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3154 * about errors or whether the log was flushed or not. This is the normal
3155 * interface to use when trying to unpin items or move the log forward.
3156 */
3157void
3158xfs_log_force_lsn(
3159 xfs_mount_t *mp,
3160 xfs_lsn_t lsn,
3161 uint flags)
3162{
3163 int error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164
Dave Chinner14c26c62012-04-24 16:33:31 +10003165 trace_xfs_log_force(mp, lsn);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003166 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003167 if (error)
3168 xfs_warn(mp, "%s: error %d returned.", __func__, error);
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003169}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170
3171/*
3172 * Called when we want to mark the current iclog as being ready to sync to
3173 * disk.
3174 */
David Chinnera8272ce2007-11-23 16:28:09 +11003175STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05003176xlog_state_want_sync(
3177 struct xlog *log,
3178 struct xlog_in_core *iclog)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179{
Christoph Hellwiga8914f32009-08-10 11:32:44 -03003180 assert_spin_locked(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003181
3182 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3183 xlog_state_switch_iclogs(log, iclog, 0);
3184 } else {
3185 ASSERT(iclog->ic_state &
3186 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3187 }
Christoph Hellwig39e2def2008-12-03 12:20:28 +01003188}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189
3190
3191/*****************************************************************************
3192 *
3193 * TICKET functions
3194 *
3195 *****************************************************************************
3196 */
3197
3198/*
Malcolm Parsons9da096f2009-03-29 09:55:42 +02003199 * Free a used ticket when its refcount falls to zero.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200 */
Dave Chinnercc09c0d2008-11-17 17:37:10 +11003201void
3202xfs_log_ticket_put(
3203 xlog_ticket_t *ticket)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204{
Dave Chinnercc09c0d2008-11-17 17:37:10 +11003205 ASSERT(atomic_read(&ticket->t_ref) > 0);
Dave Chinnereb40a872010-12-21 12:09:01 +11003206 if (atomic_dec_and_test(&ticket->t_ref))
Dave Chinnercc09c0d2008-11-17 17:37:10 +11003207 kmem_zone_free(xfs_log_ticket_zone, ticket);
Dave Chinnercc09c0d2008-11-17 17:37:10 +11003208}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209
Dave Chinnercc09c0d2008-11-17 17:37:10 +11003210xlog_ticket_t *
3211xfs_log_ticket_get(
3212 xlog_ticket_t *ticket)
3213{
3214 ASSERT(atomic_read(&ticket->t_ref) > 0);
3215 atomic_inc(&ticket->t_ref);
3216 return ticket;
3217}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218
3219/*
David Chinnereb01c9c2008-04-10 12:18:46 +10003220 * Allocate and initialise a new log ticket.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221 */
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05003222struct xlog_ticket *
Dave Chinner9b9fc2b72010-03-23 11:21:11 +11003223xlog_ticket_alloc(
Mark Tinguelyad223e62012-06-14 09:22:15 -05003224 struct xlog *log,
Dave Chinner9b9fc2b72010-03-23 11:21:11 +11003225 int unit_bytes,
3226 int cnt,
3227 char client,
Christoph Hellwig9006fb92012-02-20 02:31:31 +00003228 bool permanent,
Al Viro77ba7872012-04-02 06:24:04 -04003229 xfs_km_flags_t alloc_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230{
Dave Chinner9b9fc2b72010-03-23 11:21:11 +11003231 struct xlog_ticket *tic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232 uint num_headers;
Dave Chinner9b9fc2b72010-03-23 11:21:11 +11003233 int iclog_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234
Dave Chinner3383ca52010-05-07 11:04:17 +10003235 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
David Chinnereb01c9c2008-04-10 12:18:46 +10003236 if (!tic)
3237 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238
3239 /*
3240 * Permanent reservations have up to 'cnt'-1 active log operations
3241 * in the log. A unit in this case is the amount of space for one
3242 * of these log operations. Normal reservations have a cnt of 1
3243 * and their unit amount is the total amount of space required.
3244 *
3245 * The following lines of code account for non-transaction data
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003246 * which occupy space in the on-disk log.
3247 *
3248 * Normal form of a transaction is:
3249 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3250 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3251 *
3252 * We need to account for all the leadup data and trailer data
3253 * around the transaction data.
3254 * And then we need to account for the worst case in terms of using
3255 * more space.
3256 * The worst case will happen if:
3257 * - the placement of the transaction happens to be such that the
3258 * roundoff is at its maximum
3259 * - the transaction data is synced before the commit record is synced
3260 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3261 * Therefore the commit record is in its own Log Record.
3262 * This can happen as the commit record is called with its
3263 * own region to xlog_write().
3264 * This then means that in the worst case, roundoff can happen for
3265 * the commit-rec as well.
3266 * The commit-rec is smaller than padding in this scenario and so it is
3267 * not added separately.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 */
3269
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003270 /* for trans header */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 unit_bytes += sizeof(xlog_op_header_t);
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003272 unit_bytes += sizeof(xfs_trans_header_t);
3273
3274 /* for start-rec */
3275 unit_bytes += sizeof(xlog_op_header_t);
3276
Dave Chinner9b9fc2b72010-03-23 11:21:11 +11003277 /*
3278 * for LR headers - the space for data in an iclog is the size minus
3279 * the space used for the headers. If we use the iclog size, then we
3280 * undercalculate the number of headers required.
3281 *
3282 * Furthermore - the addition of op headers for split-recs might
3283 * increase the space required enough to require more log and op
3284 * headers, so take that into account too.
3285 *
3286 * IMPORTANT: This reservation makes the assumption that if this
3287 * transaction is the first in an iclog and hence has the LR headers
3288 * accounted to it, then the remaining space in the iclog is
3289 * exclusively for this transaction. i.e. if the transaction is larger
3290 * than the iclog, it will be the only thing in that iclog.
3291 * Fundamentally, this means we must pass the entire log vector to
3292 * xlog_write to guarantee this.
3293 */
3294 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3295 num_headers = howmany(unit_bytes, iclog_space);
3296
3297 /* for split-recs - ophdrs added when data split over LRs */
3298 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3299
3300 /* add extra header reservations if we overrun */
3301 while (!num_headers ||
3302 howmany(unit_bytes, iclog_space) > num_headers) {
3303 unit_bytes += sizeof(xlog_op_header_t);
3304 num_headers++;
3305 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 unit_bytes += log->l_iclog_hsize * num_headers;
3307
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003308 /* for commit-rec LR header - note: padding will subsume the ophdr */
3309 unit_bytes += log->l_iclog_hsize;
3310
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003311 /* for roundoff padding for transaction data and one for commit record */
Eric Sandeen62118702008-03-06 13:44:28 +11003312 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003313 log->l_mp->m_sb.sb_logsunit > 1) {
3314 /* log su roundoff */
3315 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3316 } else {
3317 /* BB roundoff */
3318 unit_bytes += 2*BBSIZE;
3319 }
3320
Dave Chinnercc09c0d2008-11-17 17:37:10 +11003321 atomic_set(&tic->t_ref, 1);
Christoph Hellwig14a72352012-02-20 02:31:24 +00003322 tic->t_task = current;
Dave Chinner10547942010-12-21 12:02:25 +11003323 INIT_LIST_HEAD(&tic->t_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324 tic->t_unit_res = unit_bytes;
3325 tic->t_curr_res = unit_bytes;
3326 tic->t_cnt = cnt;
3327 tic->t_ocnt = cnt;
Dave Chinnerf9837102010-04-14 15:47:55 +10003328 tic->t_tid = random32();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 tic->t_clientid = client;
3330 tic->t_flags = XLOG_TIC_INITED;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10003331 tic->t_trans_type = 0;
Christoph Hellwig9006fb92012-02-20 02:31:31 +00003332 if (permanent)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333 tic->t_flags |= XLOG_TIC_PERM_RESERV;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334
Christoph Hellwig0adba532007-08-30 17:21:46 +10003335 xlog_tic_reset_res(tic);
Tim Shimmin7e9c6392005-09-02 16:42:05 +10003336
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 return tic;
Dave Chinnercc09c0d2008-11-17 17:37:10 +11003338}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339
3340
3341/******************************************************************************
3342 *
3343 * Log debug routines
3344 *
3345 ******************************************************************************
3346 */
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11003347#if defined(DEBUG)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348/*
3349 * Make sure that the destination ptr is within the valid data region of
3350 * one of the iclogs. This uses backup pointers stored in a different
3351 * part of the log in case we trash the log structure.
3352 */
3353void
Christoph Hellwige6b1f272010-03-23 11:47:38 +11003354xlog_verify_dest_ptr(
Mark Tinguelyad223e62012-06-14 09:22:15 -05003355 struct xlog *log,
Christoph Hellwige6b1f272010-03-23 11:47:38 +11003356 char *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357{
3358 int i;
3359 int good_ptr = 0;
3360
Christoph Hellwige6b1f272010-03-23 11:47:38 +11003361 for (i = 0; i < log->l_iclog_bufs; i++) {
3362 if (ptr >= log->l_iclog_bak[i] &&
3363 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364 good_ptr++;
3365 }
Christoph Hellwige6b1f272010-03-23 11:47:38 +11003366
3367 if (!good_ptr)
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003368 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
Christoph Hellwige6b1f272010-03-23 11:47:38 +11003369}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
Dave Chinnerda8a1a42011-04-08 12:45:07 +10003371/*
3372 * Check to make sure the grant write head didn't just over lap the tail. If
3373 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3374 * the cycles differ by exactly one and check the byte count.
3375 *
3376 * This check is run unlocked, so can give false positives. Rather than assert
3377 * on failures, use a warn-once flag and a panic tag to allow the admin to
3378 * determine if they want to panic the machine when such an error occurs. For
3379 * debug kernels this will have the same effect as using an assert but, unlinke
3380 * an assert, it can be turned off at runtime.
3381 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382STATIC void
Dave Chinner3f336c62010-12-21 12:02:52 +11003383xlog_verify_grant_tail(
Mark Tinguelyad223e62012-06-14 09:22:15 -05003384 struct xlog *log)
Dave Chinner3f336c62010-12-21 12:02:52 +11003385{
Dave Chinner1c3cb9e2010-12-21 12:28:39 +11003386 int tail_cycle, tail_blocks;
Dave Chinnera69ed032010-12-21 12:08:20 +11003387 int cycle, space;
Dave Chinner3f336c62010-12-21 12:02:52 +11003388
Christoph Hellwig28496962012-02-20 02:31:25 +00003389 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
Dave Chinner1c3cb9e2010-12-21 12:28:39 +11003390 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3391 if (tail_cycle != cycle) {
Dave Chinnerda8a1a42011-04-08 12:45:07 +10003392 if (cycle - 1 != tail_cycle &&
3393 !(log->l_flags & XLOG_TAIL_WARN)) {
3394 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3395 "%s: cycle - 1 != tail_cycle", __func__);
3396 log->l_flags |= XLOG_TAIL_WARN;
3397 }
3398
3399 if (space > BBTOB(tail_blocks) &&
3400 !(log->l_flags & XLOG_TAIL_WARN)) {
3401 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3402 "%s: space > BBTOB(tail_blocks)", __func__);
3403 log->l_flags |= XLOG_TAIL_WARN;
3404 }
Dave Chinner3f336c62010-12-21 12:02:52 +11003405 }
3406}
3407
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408/* check if it will fit */
3409STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05003410xlog_verify_tail_lsn(
3411 struct xlog *log,
3412 struct xlog_in_core *iclog,
3413 xfs_lsn_t tail_lsn)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414{
3415 int blocks;
3416
3417 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3418 blocks =
3419 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3420 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003421 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 } else {
3423 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3424
3425 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003426 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427
3428 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3429 if (blocks < BTOBB(iclog->ic_offset) + 1)
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003430 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 }
3432} /* xlog_verify_tail_lsn */
3433
3434/*
3435 * Perform a number of checks on the iclog before writing to disk.
3436 *
3437 * 1. Make sure the iclogs are still circular
3438 * 2. Make sure we have a good magic number
3439 * 3. Make sure we don't have magic numbers in the data
3440 * 4. Check fields of each log operation header for:
3441 * A. Valid client identifier
3442 * B. tid ptr value falls in valid ptr space (user space code)
3443 * C. Length in log record header is correct according to the
3444 * individual operation headers within record.
3445 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3446 * log, check the preceding blocks of the physical log to make sure all
3447 * the cycle numbers agree with the current cycle number.
3448 */
3449STATIC void
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05003450xlog_verify_iclog(
3451 struct xlog *log,
3452 struct xlog_in_core *iclog,
3453 int count,
3454 boolean_t syncing)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455{
3456 xlog_op_header_t *ophead;
3457 xlog_in_core_t *icptr;
3458 xlog_in_core_2_t *xhdr;
3459 xfs_caddr_t ptr;
3460 xfs_caddr_t base_ptr;
3461 __psint_t field_offset;
3462 __uint8_t clientid;
3463 int len, i, j, k, op_len;
3464 int idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
3466 /* check validity of iclog pointers */
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003467 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468 icptr = log->l_iclog;
3469 for (i=0; i < log->l_iclog_bufs; i++) {
Christoph Hellwig4b809162007-08-16 15:37:36 +10003470 if (icptr == NULL)
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003471 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 icptr = icptr->ic_next;
3473 }
3474 if (icptr != log->l_iclog)
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003475 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003476 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477
3478 /* check log magic numbers */
Christoph Hellwig69ef9212011-07-08 14:36:05 +02003479 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003480 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481
Christoph Hellwigb53e6752007-10-12 10:59:34 +10003482 ptr = (xfs_caddr_t) &iclog->ic_header;
3483 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 ptr += BBSIZE) {
Christoph Hellwig69ef9212011-07-08 14:36:05 +02003485 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003486 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3487 __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 }
3489
3490 /* check fields */
Christoph Hellwigb53e6752007-10-12 10:59:34 +10003491 len = be32_to_cpu(iclog->ic_header.h_num_logops);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492 ptr = iclog->ic_datap;
3493 base_ptr = ptr;
3494 ophead = (xlog_op_header_t *)ptr;
Christoph Hellwigb28708d2008-11-28 14:23:38 +11003495 xhdr = iclog->ic_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 for (i = 0; i < len; i++) {
3497 ophead = (xlog_op_header_t *)ptr;
3498
3499 /* clientid is only 1 byte */
3500 field_offset = (__psint_t)
3501 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3502 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3503 clientid = ophead->oh_clientid;
3504 } else {
3505 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3506 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3507 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3508 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
Christoph Hellwig03bea6f2007-10-12 10:58:05 +10003509 clientid = xlog_get_client_id(
3510 xhdr[j].hic_xheader.xh_cycle_data[k]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511 } else {
Christoph Hellwig03bea6f2007-10-12 10:58:05 +10003512 clientid = xlog_get_client_id(
3513 iclog->ic_header.h_cycle_data[idx]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 }
3515 }
3516 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
Dave Chinnera0fa2b62011-03-07 10:01:35 +11003517 xfs_warn(log->l_mp,
3518 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3519 __func__, clientid, ophead,
3520 (unsigned long)field_offset);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521
3522 /* check length */
3523 field_offset = (__psint_t)
3524 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3525 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
Christoph Hellwig67fcb7b2007-10-12 10:58:59 +10003526 op_len = be32_to_cpu(ophead->oh_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 } else {
3528 idx = BTOBBT((__psint_t)&ophead->oh_len -
3529 (__psint_t)iclog->ic_datap);
3530 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3531 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3532 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
Christoph Hellwigb53e6752007-10-12 10:59:34 +10003533 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 } else {
Christoph Hellwigb53e6752007-10-12 10:59:34 +10003535 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536 }
3537 }
3538 ptr += sizeof(xlog_op_header_t) + op_len;
3539 }
3540} /* xlog_verify_iclog */
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11003541#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542
3543/*
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003544 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 */
3546STATIC int
3547xlog_state_ioerror(
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05003548 struct xlog *log)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549{
3550 xlog_in_core_t *iclog, *ic;
3551
3552 iclog = log->l_iclog;
3553 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3554 /*
3555 * Mark all the incore logs IOERROR.
3556 * From now on, no log flushes will result.
3557 */
3558 ic = iclog;
3559 do {
3560 ic->ic_state = XLOG_STATE_IOERROR;
3561 ic = ic->ic_next;
3562 } while (ic != iclog);
Jesper Juhl014c2542006-01-15 02:37:08 +01003563 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 }
3565 /*
3566 * Return non-zero, if state transition has already happened.
3567 */
Jesper Juhl014c2542006-01-15 02:37:08 +01003568 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569}
3570
3571/*
3572 * This is called from xfs_force_shutdown, when we're forcibly
3573 * shutting down the filesystem, typically because of an IO error.
3574 * Our main objectives here are to make sure that:
3575 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3576 * parties to find out, 'atomically'.
3577 * b. those who're sleeping on log reservations, pinned objects and
3578 * other resources get woken up, and be told the bad news.
3579 * c. nothing new gets queued up after (a) and (b) are done.
3580 * d. if !logerror, flush the iclogs to disk, then seal them off
3581 * for business.
Dave Chinner9da1ab12010-05-17 15:51:59 +10003582 *
3583 * Note: for delayed logging the !logerror case needs to flush the regions
3584 * held in memory out to the iclogs before flushing them to disk. This needs
3585 * to be done before the log is marked as shutdown, otherwise the flush to the
3586 * iclogs will fail.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 */
3588int
3589xfs_log_force_umount(
3590 struct xfs_mount *mp,
3591 int logerror)
3592{
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05003593 struct xlog *log;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595
3596 log = mp->m_log;
3597
3598 /*
3599 * If this happens during log recovery, don't worry about
3600 * locking; the log isn't open for business yet.
3601 */
3602 if (!log ||
3603 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3604 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
Christoph Hellwigbac8dca2008-11-28 14:23:31 +11003605 if (mp->m_sb_bp)
3606 XFS_BUF_DONE(mp->m_sb_bp);
Jesper Juhl014c2542006-01-15 02:37:08 +01003607 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 }
3609
3610 /*
3611 * Somebody could've already done the hard work for us.
3612 * No need to get locks for this.
3613 */
3614 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3615 ASSERT(XLOG_FORCED_SHUTDOWN(log));
Jesper Juhl014c2542006-01-15 02:37:08 +01003616 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 }
3618 retval = 0;
Dave Chinner9da1ab12010-05-17 15:51:59 +10003619
3620 /*
3621 * Flush the in memory commit item list before marking the log as
3622 * being shut down. We need to do it in this order to ensure all the
3623 * completed transactions are flushed to disk with the xfs_log_force()
3624 * call below.
3625 */
Christoph Hellwig93b8a582011-12-06 21:58:07 +00003626 if (!logerror)
Dave Chinnera44f13e2010-08-24 11:40:03 +10003627 xlog_cil_force(log);
Dave Chinner9da1ab12010-05-17 15:51:59 +10003628
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629 /*
Dave Chinner3f16b982010-12-21 12:29:01 +11003630 * mark the filesystem and the as in a shutdown state and wake
3631 * everybody up to tell them the bad news.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 */
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003633 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
Christoph Hellwigbac8dca2008-11-28 14:23:31 +11003635 if (mp->m_sb_bp)
3636 XFS_BUF_DONE(mp->m_sb_bp);
3637
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638 /*
3639 * This flag is sort of redundant because of the mount flag, but
3640 * it's good to maintain the separation between the log and the rest
3641 * of XFS.
3642 */
3643 log->l_flags |= XLOG_IO_ERROR;
3644
3645 /*
3646 * If we hit a log error, we want to mark all the iclogs IOERROR
3647 * while we're still holding the loglock.
3648 */
3649 if (logerror)
3650 retval = xlog_state_ioerror(log);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003651 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652
3653 /*
Dave Chinner10547942010-12-21 12:02:25 +11003654 * We don't want anybody waiting for log reservations after this. That
3655 * means we have to wake up everybody queued up on reserveq as well as
3656 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3657 * we don't enqueue anything once the SHUTDOWN flag is set, and this
Dave Chinner3f16b982010-12-21 12:29:01 +11003658 * action is protected by the grant locks.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659 */
Christoph Hellwiga79bf2d2012-02-20 02:31:27 +00003660 xlog_grant_head_wake_all(&log->l_reserve_head);
3661 xlog_grant_head_wake_all(&log->l_write_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003663 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 ASSERT(!logerror);
3665 /*
3666 * Force the incore logs to disk before shutting the
3667 * log down completely.
3668 */
Christoph Hellwiga14a3482010-01-19 09:56:46 +00003669 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3670
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003671 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672 retval = xlog_state_ioerror(log);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003673 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674 }
3675 /*
3676 * Wake up everybody waiting on xfs_log_force.
3677 * Callback all log item committed functions as if the
3678 * log writes were completed.
3679 */
3680 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3681
3682#ifdef XFSERRORDEBUG
3683 {
3684 xlog_in_core_t *iclog;
3685
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003686 spin_lock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687 iclog = log->l_iclog;
3688 do {
3689 ASSERT(iclog->ic_callback == 0);
3690 iclog = iclog->ic_next;
3691 } while (iclog != log->l_iclog);
Eric Sandeenb22cd72c2007-10-11 17:37:10 +10003692 spin_unlock(&log->l_icloglock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693 }
3694#endif
3695 /* return non-zero if log IOERROR transition had already happened */
Jesper Juhl014c2542006-01-15 02:37:08 +01003696 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697}
3698
Christoph Hellwigba0f32d2005-06-21 15:36:52 +10003699STATIC int
Mark Tinguely9a8d2fd2012-06-14 09:22:16 -05003700xlog_iclogs_empty(
3701 struct xlog *log)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702{
3703 xlog_in_core_t *iclog;
3704
3705 iclog = log->l_iclog;
3706 do {
3707 /* endianness does not matter here, zero is zero in
3708 * any language.
3709 */
3710 if (iclog->ic_header.h_num_logops)
Jesper Juhl014c2542006-01-15 02:37:08 +01003711 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712 iclog = iclog->ic_next;
3713 } while (iclog != log->l_iclog);
Jesper Juhl014c2542006-01-15 02:37:08 +01003714 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715}