blob: 5bb902056e61734a44cc4ccad1e223fd992065bf [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"
Nathan Scotta844f452005-11-02 14:38:42 +110021#include "xfs_bit.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070022#include "xfs_log.h"
Nathan Scotta844f452005-11-02 14:38:42 +110023#include "xfs_inum.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include "xfs_trans.h"
Nathan Scotta844f452005-11-02 14:38:42 +110025#include "xfs_sb.h"
26#include "xfs_ag.h"
Nathan Scotta844f452005-11-02 14:38:42 +110027#include "xfs_dir2.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include "xfs_dmapi.h"
29#include "xfs_mount.h"
30#include "xfs_error.h"
31#include "xfs_log_priv.h"
32#include "xfs_buf_item.h"
Nathan Scotta844f452005-11-02 14:38:42 +110033#include "xfs_bmap_btree.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include "xfs_alloc_btree.h"
Nathan Scotta844f452005-11-02 14:38:42 +110035#include "xfs_ialloc_btree.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include "xfs_log_recover.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037#include "xfs_trans_priv.h"
Nathan Scotta844f452005-11-02 14:38:42 +110038#include "xfs_dir2_sf.h"
39#include "xfs_attr_sf.h"
40#include "xfs_dinode.h"
41#include "xfs_inode.h"
42#include "xfs_rw.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44
45#define xlog_write_adv_cnt(ptr, len, off, bytes) \
46 { (ptr) += (bytes); \
47 (len) -= (bytes); \
48 (off) += (bytes);}
49
50/* Local miscellaneous function prototypes */
51STATIC int xlog_bdstrat_cb(struct xfs_buf *);
52STATIC int xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
53 xlog_in_core_t **, xfs_lsn_t *);
54STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
55 xfs_buftarg_t *log_target,
56 xfs_daddr_t blk_offset,
57 int num_bblks);
58STATIC int xlog_space_left(xlog_t *log, int cycle, int bytes);
59STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
Nathan Scottc41564b2006-03-29 08:55:14 +100060STATIC void xlog_dealloc_log(xlog_t *log);
Linus Torvalds1da177e2005-04-16 15:20:36 -070061STATIC int xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
62 int nentries, xfs_log_ticket_t tic,
63 xfs_lsn_t *start_lsn,
64 xlog_in_core_t **commit_iclog,
65 uint flags);
66
67/* local state machine functions */
68STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
69STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
70STATIC int xlog_state_get_iclog_space(xlog_t *log,
71 int len,
72 xlog_in_core_t **iclog,
73 xlog_ticket_t *ticket,
74 int *continued_write,
75 int *logoffsetp);
76STATIC void xlog_state_put_ticket(xlog_t *log,
77 xlog_ticket_t *tic);
78STATIC int xlog_state_release_iclog(xlog_t *log,
79 xlog_in_core_t *iclog);
80STATIC void xlog_state_switch_iclogs(xlog_t *log,
81 xlog_in_core_t *iclog,
82 int eventual_size);
Christoph Hellwigf538d4d2005-11-02 10:26:59 +110083STATIC int xlog_state_sync(xlog_t *log,
84 xfs_lsn_t lsn,
85 uint flags,
86 int *log_flushed);
87STATIC int xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
Linus Torvalds1da177e2005-04-16 15:20:36 -070088STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
89
90/* local functions to manipulate grant head */
91STATIC int xlog_grant_log_space(xlog_t *log,
92 xlog_ticket_t *xtic);
93STATIC void xlog_grant_push_ail(xfs_mount_t *mp,
94 int need_bytes);
95STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
96 xlog_ticket_t *ticket);
97STATIC int xlog_regrant_write_log_space(xlog_t *log,
98 xlog_ticket_t *ticket);
99STATIC void xlog_ungrant_log_space(xlog_t *log,
100 xlog_ticket_t *ticket);
101
102
103/* local ticket functions */
104STATIC void xlog_state_ticket_alloc(xlog_t *log);
105STATIC xlog_ticket_t *xlog_ticket_get(xlog_t *log,
106 int unit_bytes,
107 int count,
108 char clientid,
109 uint flags);
110STATIC void xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
111
Nathan Scottcfcbbbd2005-11-02 15:12:04 +1100112#if defined(DEBUG)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113STATIC void xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
114STATIC void xlog_verify_grant_head(xlog_t *log, int equals);
115STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
116 int count, boolean_t syncing);
117STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
118 xfs_lsn_t tail_lsn);
119#else
120#define xlog_verify_dest_ptr(a,b)
121#define xlog_verify_grant_head(a,b)
122#define xlog_verify_iclog(a,b,c,d)
123#define xlog_verify_tail_lsn(a,b,c)
124#endif
125
Christoph Hellwigba0f32d2005-06-21 15:36:52 +1000126STATIC int xlog_iclogs_empty(xlog_t *log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128#if defined(XFS_LOG_TRACE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129void
130xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
131{
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000132 unsigned long cnts;
133
134 if (!log->l_grant_trace) {
135 log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
136 if (!log->l_grant_trace)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 return;
138 }
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000139 /* ticket counts are 1 byte each */
140 cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
142 ktrace_enter(log->l_grant_trace,
143 (void *)tic,
144 (void *)log->l_reserve_headq,
145 (void *)log->l_write_headq,
146 (void *)((unsigned long)log->l_grant_reserve_cycle),
147 (void *)((unsigned long)log->l_grant_reserve_bytes),
148 (void *)((unsigned long)log->l_grant_write_cycle),
149 (void *)((unsigned long)log->l_grant_write_bytes),
150 (void *)((unsigned long)log->l_curr_cycle),
151 (void *)((unsigned long)log->l_curr_block),
152 (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
153 (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
154 (void *)string,
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000155 (void *)((unsigned long)tic->t_trans_type),
156 (void *)cnts,
157 (void *)((unsigned long)tic->t_curr_res),
158 (void *)((unsigned long)tic->t_unit_res));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159}
160
161void
162xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
163{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 if (!iclog->ic_trace)
165 iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
166 ktrace_enter(iclog->ic_trace,
167 (void *)((unsigned long)state),
Nathan Scottcfcbbbd2005-11-02 15:12:04 +1100168 (void *)((unsigned long)current_pid()),
169 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
170 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
171 (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
172 (void *)NULL, (void *)NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174#else
175#define xlog_trace_loggrant(log,tic,string)
176#define xlog_trace_iclog(iclog,state)
177#endif /* XFS_LOG_TRACE */
178
Christoph Hellwigdd954c62006-01-11 15:34:50 +1100179
180static void
181xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
182{
183 if (*qp) {
184 tic->t_next = (*qp);
185 tic->t_prev = (*qp)->t_prev;
186 (*qp)->t_prev->t_next = tic;
187 (*qp)->t_prev = tic;
188 } else {
189 tic->t_prev = tic->t_next = tic;
190 *qp = tic;
191 }
192
193 tic->t_flags |= XLOG_TIC_IN_Q;
194}
195
196static void
197xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
198{
199 if (tic == tic->t_next) {
200 *qp = NULL;
201 } else {
202 *qp = tic->t_next;
203 tic->t_next->t_prev = tic->t_prev;
204 tic->t_prev->t_next = tic->t_next;
205 }
206
207 tic->t_next = tic->t_prev = NULL;
208 tic->t_flags &= ~XLOG_TIC_IN_Q;
209}
210
211static void
212xlog_grant_sub_space(struct log *log, int bytes)
213{
214 log->l_grant_write_bytes -= bytes;
215 if (log->l_grant_write_bytes < 0) {
216 log->l_grant_write_bytes += log->l_logsize;
217 log->l_grant_write_cycle--;
218 }
219
220 log->l_grant_reserve_bytes -= bytes;
221 if ((log)->l_grant_reserve_bytes < 0) {
222 log->l_grant_reserve_bytes += log->l_logsize;
223 log->l_grant_reserve_cycle--;
224 }
225
226}
227
228static void
229xlog_grant_add_space_write(struct log *log, int bytes)
230{
231 log->l_grant_write_bytes += bytes;
232 if (log->l_grant_write_bytes > log->l_logsize) {
233 log->l_grant_write_bytes -= log->l_logsize;
234 log->l_grant_write_cycle++;
235 }
236}
237
238static void
239xlog_grant_add_space_reserve(struct log *log, int bytes)
240{
241 log->l_grant_reserve_bytes += bytes;
242 if (log->l_grant_reserve_bytes > log->l_logsize) {
243 log->l_grant_reserve_bytes -= log->l_logsize;
244 log->l_grant_reserve_cycle++;
245 }
246}
247
248static inline void
249xlog_grant_add_space(struct log *log, int bytes)
250{
251 xlog_grant_add_space_write(log, bytes);
252 xlog_grant_add_space_reserve(log, bytes);
253}
254
255
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256/*
257 * NOTES:
258 *
259 * 1. currblock field gets updated at startup and after in-core logs
260 * marked as with WANT_SYNC.
261 */
262
263/*
264 * This routine is called when a user of a log manager ticket is done with
265 * the reservation. If the ticket was ever used, then a commit record for
266 * the associated transaction is written out as a log operation header with
267 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
268 * a given ticket. If the ticket was one with a permanent reservation, then
269 * a few operations are done differently. Permanent reservation tickets by
270 * default don't release the reservation. They just commit the current
271 * transaction with the belief that the reservation is still needed. A flag
272 * must be passed in before permanent reservations are actually released.
273 * When these type of tickets are not released, they need to be set into
274 * the inited state again. By doing this, a start record will be written
275 * out when the next write occurs.
276 */
277xfs_lsn_t
278xfs_log_done(xfs_mount_t *mp,
279 xfs_log_ticket_t xtic,
280 void **iclog,
281 uint flags)
282{
283 xlog_t *log = mp->m_log;
284 xlog_ticket_t *ticket = (xfs_log_ticket_t) xtic;
285 xfs_lsn_t lsn = 0;
286
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 if (XLOG_FORCED_SHUTDOWN(log) ||
288 /*
289 * If nothing was ever written, don't write out commit record.
290 * If we get an error, just continue and give back the log ticket.
291 */
292 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
293 (xlog_commit_record(mp, ticket,
294 (xlog_in_core_t **)iclog, &lsn)))) {
295 lsn = (xfs_lsn_t) -1;
296 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
297 flags |= XFS_LOG_REL_PERM_RESERV;
298 }
299 }
300
301
302 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
303 (flags & XFS_LOG_REL_PERM_RESERV)) {
304 /*
Nathan Scottc41564b2006-03-29 08:55:14 +1000305 * Release ticket if not permanent reservation or a specific
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 * request has been made to release a permanent reservation.
307 */
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000308 xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 xlog_ungrant_log_space(log, ticket);
310 xlog_state_put_ticket(log, ticket);
311 } else {
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000312 xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 xlog_regrant_reserve_log_space(log, ticket);
314 }
315
316 /* If this ticket was a permanent reservation and we aren't
317 * trying to release it, reset the inited flags; so next time
318 * we write, a start record will be written out.
319 */
320 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
321 (flags & XFS_LOG_REL_PERM_RESERV) == 0)
322 ticket->t_flags |= XLOG_TIC_INITED;
323
324 return lsn;
325} /* xfs_log_done */
326
327
328/*
329 * Force the in-core log to disk. If flags == XFS_LOG_SYNC,
330 * the force is done synchronously.
331 *
332 * Asynchronous forces are implemented by setting the WANT_SYNC
333 * bit in the appropriate in-core log and then returning.
334 *
335 * Synchronous forces are implemented with a semaphore. All callers
336 * to force a given lsn to disk will wait on a semaphore attached to the
337 * specific in-core log. When given in-core log finally completes its
338 * write to disk, that thread will wake up all threads waiting on the
339 * semaphore.
340 */
341int
Christoph Hellwigf538d4d2005-11-02 10:26:59 +1100342_xfs_log_force(
343 xfs_mount_t *mp,
344 xfs_lsn_t lsn,
345 uint flags,
346 int *log_flushed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347{
Christoph Hellwigf538d4d2005-11-02 10:26:59 +1100348 xlog_t *log = mp->m_log;
349 int dummy;
350
351 if (!log_flushed)
352 log_flushed = &dummy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 ASSERT(flags & XFS_LOG_FORCE);
355
356 XFS_STATS_INC(xs_log_force);
357
Christoph Hellwigf538d4d2005-11-02 10:26:59 +1100358 if (log->l_flags & XLOG_IO_ERROR)
359 return XFS_ERROR(EIO);
360 if (lsn == 0)
361 return xlog_state_sync_all(log, flags, log_flushed);
362 else
363 return xlog_state_sync(log, lsn, flags, log_flushed);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364} /* xfs_log_force */
365
366/*
367 * Attaches a new iclog I/O completion callback routine during
368 * transaction commit. If the log is in error state, a non-zero
369 * return code is handed back and the caller is responsible for
370 * executing the callback at an appropriate time.
371 */
372int
373xfs_log_notify(xfs_mount_t *mp, /* mount of partition */
374 void *iclog_hndl, /* iclog to hang callback off */
375 xfs_log_callback_t *cb)
376{
377 xlog_t *log = mp->m_log;
378 xlog_in_core_t *iclog = (xlog_in_core_t *)iclog_hndl;
379 int abortflg, spl;
380
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 cb->cb_next = NULL;
382 spl = LOG_LOCK(log);
383 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
384 if (!abortflg) {
385 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
386 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
387 cb->cb_next = NULL;
388 *(iclog->ic_callback_tail) = cb;
389 iclog->ic_callback_tail = &(cb->cb_next);
390 }
391 LOG_UNLOCK(log, spl);
392 return abortflg;
393} /* xfs_log_notify */
394
395int
396xfs_log_release_iclog(xfs_mount_t *mp,
397 void *iclog_hndl)
398{
399 xlog_t *log = mp->m_log;
400 xlog_in_core_t *iclog = (xlog_in_core_t *)iclog_hndl;
401
402 if (xlog_state_release_iclog(log, iclog)) {
Nathan Scott7d04a332006-06-09 14:58:38 +1000403 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
Jesper Juhl014c2542006-01-15 02:37:08 +0100404 return EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 }
406
407 return 0;
408}
409
410/*
411 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
412 * to the reservation.
413 * 2. Potentially, push buffers at tail of log to disk.
414 *
415 * Each reservation is going to reserve extra space for a log record header.
416 * When writes happen to the on-disk log, we don't subtract the length of the
417 * log record header from any reservation. By wasting space in each
418 * reservation, we prevent over allocation problems.
419 */
420int
421xfs_log_reserve(xfs_mount_t *mp,
422 int unit_bytes,
423 int cnt,
424 xfs_log_ticket_t *ticket,
425 __uint8_t client,
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000426 uint flags,
427 uint t_type)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428{
429 xlog_t *log = mp->m_log;
430 xlog_ticket_t *internal_ticket;
Nathan Scottcfcbbbd2005-11-02 15:12:04 +1100431 int retval = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
434 ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
435
436 if (XLOG_FORCED_SHUTDOWN(log))
437 return XFS_ERROR(EIO);
438
439 XFS_STATS_INC(xs_try_logspace);
440
441 if (*ticket != NULL) {
442 ASSERT(flags & XFS_LOG_PERM_RESERV);
443 internal_ticket = (xlog_ticket_t *)*ticket;
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000444 xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
446 retval = xlog_regrant_write_log_space(log, internal_ticket);
447 } else {
448 /* may sleep if need to allocate more tickets */
449 internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
450 client, flags);
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000451 internal_ticket->t_trans_type = t_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 *ticket = internal_ticket;
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000453 xlog_trace_loggrant(log, internal_ticket,
454 (internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
455 "xfs_log_reserve: create new ticket (permanent trans)" :
456 "xfs_log_reserve: create new ticket");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 xlog_grant_push_ail(mp,
458 (internal_ticket->t_unit_res *
459 internal_ticket->t_cnt));
460 retval = xlog_grant_log_space(log, internal_ticket);
461 }
462
463 return retval;
464} /* xfs_log_reserve */
465
466
467/*
468 * Mount a log filesystem
469 *
470 * mp - ubiquitous xfs mount point structure
471 * log_target - buftarg of on-disk log device
472 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
473 * num_bblocks - Number of BBSIZE blocks in on-disk log
474 *
475 * Return error or zero.
476 */
477int
478xfs_log_mount(xfs_mount_t *mp,
479 xfs_buftarg_t *log_target,
480 xfs_daddr_t blk_offset,
481 int num_bblks)
482{
483 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
484 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
485 else {
486 cmn_err(CE_NOTE,
487 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
488 mp->m_fsname);
489 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
490 }
491
492 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
493
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 /*
495 * skip log recovery on a norecovery mount. pretend it all
496 * just worked.
497 */
498 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
Nathan Scottb83bd132006-06-09 16:48:30 +1000499 bhv_vfs_t *vfsp = XFS_MTOVFS(mp);
500 int error, readonly = (vfsp->vfs_flag & VFS_RDONLY);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501
502 if (readonly)
503 vfsp->vfs_flag &= ~VFS_RDONLY;
504
Eric Sandeen65be6052006-01-11 15:34:19 +1100505 error = xlog_recover(mp->m_log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506
507 if (readonly)
508 vfsp->vfs_flag |= VFS_RDONLY;
509 if (error) {
510 cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
Nathan Scottc41564b2006-03-29 08:55:14 +1000511 xlog_dealloc_log(mp->m_log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 return error;
513 }
514 }
515
516 /* Normal transactions can now occur */
517 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
518
519 /* End mounting message in xfs_log_mount_finish */
520 return 0;
521} /* xfs_log_mount */
522
523/*
524 * Finish the recovery of the file system. This is separate from
525 * the xfs_log_mount() call, because it depends on the code in
526 * xfs_mountfs() to read in the root and real-time bitmap inodes
527 * between calling xfs_log_mount() and here.
528 *
529 * mp - ubiquitous xfs mount point structure
530 */
531int
532xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
533{
534 int error;
535
536 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
537 error = xlog_recover_finish(mp->m_log, mfsi_flags);
538 else {
539 error = 0;
540 ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
541 }
542
543 return error;
544}
545
546/*
547 * Unmount processing for the log.
548 */
549int
550xfs_log_unmount(xfs_mount_t *mp)
551{
552 int error;
553
554 error = xfs_log_unmount_write(mp);
555 xfs_log_unmount_dealloc(mp);
Jesper Juhl014c2542006-01-15 02:37:08 +0100556 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557}
558
559/*
560 * Final log writes as part of unmount.
561 *
562 * Mark the filesystem clean as unmount happens. Note that during relocation
563 * this routine needs to be executed as part of source-bag while the
564 * deallocation must not be done until source-end.
565 */
566
567/*
568 * Unmount record used to have a string "Unmount filesystem--" in the
569 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
570 * We just write the magic number now since that particular field isn't
571 * currently architecture converted and "nUmount" is a bit foo.
572 * As far as I know, there weren't any dependencies on the old behaviour.
573 */
574
575int
576xfs_log_unmount_write(xfs_mount_t *mp)
577{
578 xlog_t *log = mp->m_log;
579 xlog_in_core_t *iclog;
580#ifdef DEBUG
581 xlog_in_core_t *first_iclog;
582#endif
583 xfs_log_iovec_t reg[1];
584 xfs_log_ticket_t tic = NULL;
585 xfs_lsn_t lsn;
586 int error;
587 SPLDECL(s);
588
589 /* the data section must be 32 bit size aligned */
590 struct {
591 __uint16_t magic;
592 __uint16_t pad1;
593 __uint32_t pad2; /* may as well make it 64 bits */
594 } magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
595
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 /*
597 * Don't write out unmount record on read-only mounts.
598 * Or, if we are doing a forced umount (typically because of IO errors).
599 */
600 if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
601 return 0;
602
603 xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
604
605#ifdef DEBUG
606 first_iclog = iclog = log->l_iclog;
607 do {
608 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
609 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
610 ASSERT(iclog->ic_offset == 0);
611 }
612 iclog = iclog->ic_next;
613 } while (iclog != first_iclog);
614#endif
615 if (! (XLOG_FORCED_SHUTDOWN(log))) {
616 reg[0].i_addr = (void*)&magic;
617 reg[0].i_len = sizeof(magic);
Tim Shimmin7e9c6392005-09-02 16:42:05 +1000618 XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
Tim Shimmin955e47a2006-09-28 11:04:16 +1000620 error = xfs_log_reserve(mp, 600, 1, &tic,
621 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 if (!error) {
623 /* remove inited flag */
624 ((xlog_ticket_t *)tic)->t_flags = 0;
625 error = xlog_write(mp, reg, 1, tic, &lsn,
626 NULL, XLOG_UNMOUNT_TRANS);
627 /*
628 * At this point, we're umounting anyway,
629 * so there's no point in transitioning log state
630 * to IOERROR. Just continue...
631 */
632 }
633
634 if (error) {
635 xfs_fs_cmn_err(CE_ALERT, mp,
636 "xfs_log_unmount: unmount record failed");
637 }
638
639
640 s = LOG_LOCK(log);
641 iclog = log->l_iclog;
642 iclog->ic_refcnt++;
643 LOG_UNLOCK(log, s);
644 xlog_state_want_sync(log, iclog);
645 (void) xlog_state_release_iclog(log, iclog);
646
647 s = LOG_LOCK(log);
648 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
649 iclog->ic_state == XLOG_STATE_DIRTY)) {
650 if (!XLOG_FORCED_SHUTDOWN(log)) {
651 sv_wait(&iclog->ic_forcesema, PMEM,
652 &log->l_icloglock, s);
653 } else {
654 LOG_UNLOCK(log, s);
655 }
656 } else {
657 LOG_UNLOCK(log, s);
658 }
Tim Shimmin955e47a2006-09-28 11:04:16 +1000659 if (tic) {
660 xlog_trace_loggrant(log, tic, "unmount rec");
661 xlog_ungrant_log_space(log, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 xlog_state_put_ticket(log, tic);
Tim Shimmin955e47a2006-09-28 11:04:16 +1000663 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 } else {
665 /*
666 * We're already in forced_shutdown mode, couldn't
667 * even attempt to write out the unmount transaction.
668 *
669 * Go through the motions of sync'ing and releasing
670 * the iclog, even though no I/O will actually happen,
Nathan Scottc41564b2006-03-29 08:55:14 +1000671 * we need to wait for other log I/Os that may already
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 * be in progress. Do this as a separate section of
673 * code so we'll know if we ever get stuck here that
674 * we're in this odd situation of trying to unmount
675 * a file system that went into forced_shutdown as
676 * the result of an unmount..
677 */
678 s = LOG_LOCK(log);
679 iclog = log->l_iclog;
680 iclog->ic_refcnt++;
681 LOG_UNLOCK(log, s);
682
683 xlog_state_want_sync(log, iclog);
684 (void) xlog_state_release_iclog(log, iclog);
685
686 s = LOG_LOCK(log);
687
688 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
689 || iclog->ic_state == XLOG_STATE_DIRTY
690 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
691
692 sv_wait(&iclog->ic_forcesema, PMEM,
693 &log->l_icloglock, s);
694 } else {
695 LOG_UNLOCK(log, s);
696 }
697 }
698
699 return 0;
700} /* xfs_log_unmount_write */
701
702/*
703 * Deallocate log structures for unmount/relocation.
704 */
705void
706xfs_log_unmount_dealloc(xfs_mount_t *mp)
707{
Nathan Scottc41564b2006-03-29 08:55:14 +1000708 xlog_dealloc_log(mp->m_log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709}
710
711/*
712 * Write region vectors to log. The write happens using the space reservation
713 * of the ticket (tic). It is not a requirement that all writes for a given
714 * transaction occur with one call to xfs_log_write().
715 */
716int
717xfs_log_write(xfs_mount_t * mp,
718 xfs_log_iovec_t reg[],
719 int nentries,
720 xfs_log_ticket_t tic,
721 xfs_lsn_t *start_lsn)
722{
723 int error;
724 xlog_t *log = mp->m_log;
725
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 if (XLOG_FORCED_SHUTDOWN(log))
727 return XFS_ERROR(EIO);
728
729 if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
Nathan Scott7d04a332006-06-09 14:58:38 +1000730 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 }
Jesper Juhl014c2542006-01-15 02:37:08 +0100732 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733} /* xfs_log_write */
734
735
736void
737xfs_log_move_tail(xfs_mount_t *mp,
738 xfs_lsn_t tail_lsn)
739{
740 xlog_ticket_t *tic;
741 xlog_t *log = mp->m_log;
742 int need_bytes, free_bytes, cycle, bytes;
743 SPLDECL(s);
744
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 if (XLOG_FORCED_SHUTDOWN(log))
746 return;
747 ASSERT(!XFS_FORCED_SHUTDOWN(mp));
748
749 if (tail_lsn == 0) {
750 /* needed since sync_lsn is 64 bits */
751 s = LOG_LOCK(log);
752 tail_lsn = log->l_last_sync_lsn;
753 LOG_UNLOCK(log, s);
754 }
755
756 s = GRANT_LOCK(log);
757
758 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
759 * tail_lsn.
760 */
761 if (tail_lsn != 1) {
762 log->l_tail_lsn = tail_lsn;
763 }
764
765 if ((tic = log->l_write_headq)) {
766#ifdef DEBUG
767 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
768 panic("Recovery problem");
769#endif
770 cycle = log->l_grant_write_cycle;
771 bytes = log->l_grant_write_bytes;
772 free_bytes = xlog_space_left(log, cycle, bytes);
773 do {
774 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
775
776 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
777 break;
778 tail_lsn = 0;
779 free_bytes -= tic->t_unit_res;
780 sv_signal(&tic->t_sema);
781 tic = tic->t_next;
782 } while (tic != log->l_write_headq);
783 }
784 if ((tic = log->l_reserve_headq)) {
785#ifdef DEBUG
786 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
787 panic("Recovery problem");
788#endif
789 cycle = log->l_grant_reserve_cycle;
790 bytes = log->l_grant_reserve_bytes;
791 free_bytes = xlog_space_left(log, cycle, bytes);
792 do {
793 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
794 need_bytes = tic->t_unit_res*tic->t_cnt;
795 else
796 need_bytes = tic->t_unit_res;
797 if (free_bytes < need_bytes && tail_lsn != 1)
798 break;
799 tail_lsn = 0;
800 free_bytes -= need_bytes;
801 sv_signal(&tic->t_sema);
802 tic = tic->t_next;
803 } while (tic != log->l_reserve_headq);
804 }
805 GRANT_UNLOCK(log, s);
806} /* xfs_log_move_tail */
807
808/*
809 * Determine if we have a transaction that has gone to disk
810 * that needs to be covered. Log activity needs to be idle (no AIL and
811 * nothing in the iclogs). And, we need to be in the right state indicating
812 * something has gone out.
813 */
814int
815xfs_log_need_covered(xfs_mount_t *mp)
816{
817 SPLDECL(s);
818 int needed = 0, gen;
819 xlog_t *log = mp->m_log;
Nathan Scottb83bd132006-06-09 16:48:30 +1000820 bhv_vfs_t *vfsp = XFS_MTOVFS(mp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821
Nathan Scott34327e12006-06-09 17:11:55 +1000822 if (vfs_test_for_freeze(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 (vfsp->vfs_flag & VFS_RDONLY))
824 return 0;
825
826 s = LOG_LOCK(log);
827 if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
828 (log->l_covered_state == XLOG_STATE_COVER_NEED2))
829 && !xfs_trans_first_ail(mp, &gen)
830 && xlog_iclogs_empty(log)) {
831 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
832 log->l_covered_state = XLOG_STATE_COVER_DONE;
833 else {
834 ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
835 log->l_covered_state = XLOG_STATE_COVER_DONE2;
836 }
837 needed = 1;
838 }
839 LOG_UNLOCK(log, s);
Jesper Juhl014c2542006-01-15 02:37:08 +0100840 return needed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841}
842
843/******************************************************************************
844 *
845 * local routines
846 *
847 ******************************************************************************
848 */
849
850/* xfs_trans_tail_ail returns 0 when there is nothing in the list.
851 * The log manager must keep track of the last LR which was committed
852 * to disk. The lsn of this LR will become the new tail_lsn whenever
853 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
854 * the situation where stuff could be written into the log but nothing
855 * was ever in the AIL when asked. Eventually, we panic since the
856 * tail hits the head.
857 *
858 * We may be holding the log iclog lock upon entering this routine.
859 */
860xfs_lsn_t
861xlog_assign_tail_lsn(xfs_mount_t *mp)
862{
863 xfs_lsn_t tail_lsn;
864 SPLDECL(s);
865 xlog_t *log = mp->m_log;
866
867 tail_lsn = xfs_trans_tail_ail(mp);
868 s = GRANT_LOCK(log);
869 if (tail_lsn != 0) {
870 log->l_tail_lsn = tail_lsn;
871 } else {
872 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
873 }
874 GRANT_UNLOCK(log, s);
875
876 return tail_lsn;
877} /* xlog_assign_tail_lsn */
878
879
880/*
881 * Return the space in the log between the tail and the head. The head
882 * is passed in the cycle/bytes formal parms. In the special case where
883 * the reserve head has wrapped passed the tail, this calculation is no
884 * longer valid. In this case, just return 0 which means there is no space
885 * in the log. This works for all places where this function is called
886 * with the reserve head. Of course, if the write head were to ever
887 * wrap the tail, we should blow up. Rather than catch this case here,
888 * we depend on other ASSERTions in other parts of the code. XXXmiken
889 *
890 * This code also handles the case where the reservation head is behind
891 * the tail. The details of this case are described below, but the end
892 * result is that we return the size of the log as the amount of space left.
893 */
894int
895xlog_space_left(xlog_t *log, int cycle, int bytes)
896{
897 int free_bytes;
898 int tail_bytes;
899 int tail_cycle;
900
901 tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
902 tail_cycle = CYCLE_LSN(log->l_tail_lsn);
903 if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
904 free_bytes = log->l_logsize - (bytes - tail_bytes);
905 } else if ((tail_cycle + 1) < cycle) {
906 return 0;
907 } else if (tail_cycle < cycle) {
908 ASSERT(tail_cycle == (cycle - 1));
909 free_bytes = tail_bytes - bytes;
910 } else {
911 /*
912 * The reservation head is behind the tail.
913 * In this case we just want to return the size of the
914 * log as the amount of space left.
915 */
916 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
917 "xlog_space_left: head behind tail\n"
918 " tail_cycle = %d, tail_bytes = %d\n"
919 " GH cycle = %d, GH bytes = %d",
920 tail_cycle, tail_bytes, cycle, bytes);
921 ASSERT(0);
922 free_bytes = log->l_logsize;
923 }
924 return free_bytes;
925} /* xlog_space_left */
926
927
928/*
929 * Log function which is called when an io completes.
930 *
931 * The log manager needs its own routine, in order to control what
932 * happens with the buffer after the write completes.
933 */
934void
935xlog_iodone(xfs_buf_t *bp)
936{
937 xlog_in_core_t *iclog;
938 xlog_t *l;
939 int aborted;
940
941 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
942 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
943 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
944 aborted = 0;
945
946 /*
947 * Some versions of cpp barf on the recursive definition of
948 * ic_log -> hic_fields.ic_log and expand ic_log twice when
949 * it is passed through two macros. Workaround broken cpp.
950 */
951 l = iclog->ic_log;
952
953 /*
954 * Race to shutdown the filesystem if we see an error.
955 */
956 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
957 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
958 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
959 XFS_BUF_STALE(bp);
Nathan Scott7d04a332006-06-09 14:58:38 +1000960 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 /*
962 * This flag will be propagated to the trans-committed
963 * callback routines to let them know that the log-commit
964 * didn't succeed.
965 */
966 aborted = XFS_LI_ABORTED;
967 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
968 aborted = XFS_LI_ABORTED;
969 }
David Chinner3db296f2007-05-14 18:24:16 +1000970
971 /* log I/O is always issued ASYNC */
972 ASSERT(XFS_BUF_ISASYNC(bp));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 xlog_state_done_syncing(iclog, aborted);
David Chinner3db296f2007-05-14 18:24:16 +1000974 /*
975 * do not reference the buffer (bp) here as we could race
976 * with it being freed after writing the unmount record to the
977 * log.
978 */
979
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980} /* xlog_iodone */
981
982/*
983 * The bdstrat callback function for log bufs. This gives us a central
984 * place to trap bufs in case we get hit by a log I/O error and need to
985 * shutdown. Actually, in practice, even when we didn't get a log error,
986 * we transition the iclogs to IOERROR state *after* flushing all existing
987 * iclogs to disk. This is because we don't want anymore new transactions to be
988 * started or completed afterwards.
989 */
990STATIC int
991xlog_bdstrat_cb(struct xfs_buf *bp)
992{
993 xlog_in_core_t *iclog;
994
995 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
996
997 if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
998 /* note for irix bstrat will need struct bdevsw passed
999 * Fix the following macro if the code ever is merged
1000 */
1001 XFS_bdstrat(bp);
1002 return 0;
1003 }
1004
1005 xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1006 XFS_BUF_ERROR(bp, EIO);
1007 XFS_BUF_STALE(bp);
1008 xfs_biodone(bp);
Jesper Juhl014c2542006-01-15 02:37:08 +01001009 return XFS_ERROR(EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
1011
1012}
1013
1014/*
1015 * Return size of each in-core log record buffer.
1016 *
1017 * Low memory machines only get 2 16KB buffers. We don't want to waste
1018 * memory here. However, all other machines get at least 2 32KB buffers.
1019 * The number is hard coded because we don't care about the minimum
1020 * memory size, just 32MB systems.
1021 *
1022 * If the filesystem blocksize is too large, we may need to choose a
1023 * larger size since the directory code currently logs entire blocks.
1024 */
1025
1026STATIC void
1027xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1028 xlog_t *log)
1029{
1030 int size;
1031 int xhdrs;
1032
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001033 if (mp->m_logbufs <= 0) {
1034 if (xfs_physmem <= btoc(128*1024*1024)) {
1035 log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1036 } else if (xfs_physmem <= btoc(400*1024*1024)) {
1037 log->l_iclog_bufs = XLOG_MED_ICLOGS;
1038 } else { /* 256K with 32K bufs */
1039 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 }
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001041 } else {
1042 log->l_iclog_bufs = mp->m_logbufs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 }
1044
1045 /*
1046 * Buffer size passed in from mount system call.
1047 */
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001048 if (mp->m_logbsize > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 size = log->l_iclog_size = mp->m_logbsize;
1050 log->l_iclog_size_log = 0;
1051 while (size != 1) {
1052 log->l_iclog_size_log++;
1053 size >>= 1;
1054 }
1055
1056 if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
1057 /* # headers = size / 32K
1058 * one header holds cycles from 32K of data
1059 */
1060
1061 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1062 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1063 xhdrs++;
1064 log->l_iclog_hsize = xhdrs << BBSHIFT;
1065 log->l_iclog_heads = xhdrs;
1066 } else {
1067 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1068 log->l_iclog_hsize = BBSIZE;
1069 log->l_iclog_heads = 1;
1070 }
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001071 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 }
1073
1074 /*
1075 * Special case machines that have less than 32MB of memory.
1076 * All machines with more memory use 32KB buffers.
1077 */
1078 if (xfs_physmem <= btoc(32*1024*1024)) {
1079 /* Don't change; min configuration */
1080 log->l_iclog_size = XLOG_RECORD_BSIZE; /* 16k */
1081 log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1082 } else {
1083 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE; /* 32k */
1084 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1085 }
1086
1087 /* the default log size is 16k or 32k which is one header sector */
1088 log->l_iclog_hsize = BBSIZE;
1089 log->l_iclog_heads = 1;
1090
1091 /*
1092 * For 16KB, we use 3 32KB buffers. For 32KB block sizes, we use
1093 * 4 32KB buffers. For 64KB block sizes, we use 8 32KB buffers.
1094 */
1095 if (mp->m_sb.sb_blocksize >= 16*1024) {
1096 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1097 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001098 if (mp->m_logbufs <= 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 switch (mp->m_sb.sb_blocksize) {
1100 case 16*1024: /* 16 KB */
1101 log->l_iclog_bufs = 3;
1102 break;
1103 case 32*1024: /* 32 KB */
1104 log->l_iclog_bufs = 4;
1105 break;
1106 case 64*1024: /* 64 KB */
1107 log->l_iclog_bufs = 8;
1108 break;
1109 default:
1110 xlog_panic("XFS: Invalid blocksize");
1111 break;
1112 }
1113 }
1114 }
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11001115
1116done: /* are we being asked to make the sizes selected above visible? */
1117 if (mp->m_logbufs == 0)
1118 mp->m_logbufs = log->l_iclog_bufs;
1119 if (mp->m_logbsize == 0)
1120 mp->m_logbsize = log->l_iclog_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121} /* xlog_get_iclog_buffer_size */
1122
1123
1124/*
1125 * This routine initializes some of the log structure for a given mount point.
1126 * Its primary purpose is to fill in enough, so recovery can occur. However,
1127 * some other stuff may be filled in too.
1128 */
1129STATIC xlog_t *
1130xlog_alloc_log(xfs_mount_t *mp,
1131 xfs_buftarg_t *log_target,
1132 xfs_daddr_t blk_offset,
1133 int num_bblks)
1134{
1135 xlog_t *log;
1136 xlog_rec_header_t *head;
1137 xlog_in_core_t **iclogp;
1138 xlog_in_core_t *iclog, *prev_iclog=NULL;
1139 xfs_buf_t *bp;
1140 int i;
1141 int iclogsize;
1142
1143 log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1144
1145 log->l_mp = mp;
1146 log->l_targ = log_target;
1147 log->l_logsize = BBTOB(num_bblks);
1148 log->l_logBBstart = blk_offset;
1149 log->l_logBBsize = num_bblks;
1150 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1151 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1152
1153 log->l_prev_block = -1;
1154 ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1155 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1156 log->l_last_sync_lsn = log->l_tail_lsn;
1157 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1158 log->l_grant_reserve_cycle = 1;
1159 log->l_grant_write_cycle = 1;
1160
1161 if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1162 log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1163 ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1164 /* for larger sector sizes, must have v2 or external log */
1165 ASSERT(log->l_sectbb_log == 0 ||
1166 log->l_logBBstart == 0 ||
1167 XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1168 ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1169 }
1170 log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1171
1172 xlog_get_iclog_buffer_size(mp, log);
1173
1174 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1175 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1176 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1177 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1178 ASSERT(XFS_BUF_ISBUSY(bp));
1179 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1180 log->l_xbuf = bp;
1181
1182 spinlock_init(&log->l_icloglock, "iclog");
1183 spinlock_init(&log->l_grant_lock, "grhead_iclog");
1184 initnsema(&log->l_flushsema, 0, "ic-flush");
1185 xlog_state_ticket_alloc(log); /* wait until after icloglock inited */
1186
1187 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1188 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1189
1190 iclogp = &log->l_iclog;
1191 /*
1192 * The amount of memory to allocate for the iclog structure is
1193 * rather funky due to the way the structure is defined. It is
1194 * done this way so that we can use different sizes for machines
1195 * with different amounts of memory. See the definition of
1196 * xlog_in_core_t in xfs_log_priv.h for details.
1197 */
1198 iclogsize = log->l_iclog_size;
1199 ASSERT(log->l_iclog_size >= 4096);
1200 for (i=0; i < log->l_iclog_bufs; i++) {
1201 *iclogp = (xlog_in_core_t *)
1202 kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1203 iclog = *iclogp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 iclog->ic_prev = prev_iclog;
1205 prev_iclog = iclog;
Christoph Hellwig1fa40b02007-05-14 18:23:50 +10001206
1207 bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1208 if (!XFS_BUF_CPSEMA(bp))
1209 ASSERT(0);
1210 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1211 XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1212 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1213 iclog->ic_bp = bp;
1214 iclog->hic_data = bp->b_addr;
1215
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1217
1218 head = &iclog->ic_header;
1219 memset(head, 0, sizeof(xlog_rec_header_t));
1220 INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1221 INT_SET(head->h_version, ARCH_CONVERT,
1222 XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1223 INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1224 /* new fields */
1225 INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1226 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1227
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228
1229 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1230 iclog->ic_state = XLOG_STATE_ACTIVE;
1231 iclog->ic_log = log;
1232 iclog->ic_callback_tail = &(iclog->ic_callback);
1233 iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1234
1235 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1236 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1237 sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1238 sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1239
1240 iclogp = &iclog->ic_next;
1241 }
1242 *iclogp = log->l_iclog; /* complete ring */
1243 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1244
1245 return log;
1246} /* xlog_alloc_log */
1247
1248
1249/*
1250 * Write out the commit record of a transaction associated with the given
1251 * ticket. Return the lsn of the commit record.
1252 */
1253STATIC int
1254xlog_commit_record(xfs_mount_t *mp,
1255 xlog_ticket_t *ticket,
1256 xlog_in_core_t **iclog,
1257 xfs_lsn_t *commitlsnp)
1258{
1259 int error;
1260 xfs_log_iovec_t reg[1];
1261
1262 reg[0].i_addr = NULL;
1263 reg[0].i_len = 0;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001264 XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265
1266 ASSERT_ALWAYS(iclog);
1267 if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1268 iclog, XLOG_COMMIT_TRANS))) {
Nathan Scott7d04a332006-06-09 14:58:38 +10001269 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 }
Jesper Juhl014c2542006-01-15 02:37:08 +01001271 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272} /* xlog_commit_record */
1273
1274
1275/*
1276 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1277 * log space. This code pushes on the lsn which would supposedly free up
1278 * the 25% which we want to leave free. We may need to adopt a policy which
1279 * pushes on an lsn which is further along in the log once we reach the high
1280 * water mark. In this manner, we would be creating a low water mark.
1281 */
1282void
1283xlog_grant_push_ail(xfs_mount_t *mp,
1284 int need_bytes)
1285{
1286 xlog_t *log = mp->m_log; /* pointer to the log */
1287 xfs_lsn_t tail_lsn; /* lsn of the log tail */
1288 xfs_lsn_t threshold_lsn = 0; /* lsn we'd like to be at */
1289 int free_blocks; /* free blocks left to write to */
1290 int free_bytes; /* free bytes left to write to */
1291 int threshold_block; /* block in lsn we'd like to be at */
1292 int threshold_cycle; /* lsn cycle we'd like to be at */
1293 int free_threshold;
1294 SPLDECL(s);
1295
1296 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1297
1298 s = GRANT_LOCK(log);
1299 free_bytes = xlog_space_left(log,
1300 log->l_grant_reserve_cycle,
1301 log->l_grant_reserve_bytes);
1302 tail_lsn = log->l_tail_lsn;
1303 free_blocks = BTOBBT(free_bytes);
1304
1305 /*
1306 * Set the threshold for the minimum number of free blocks in the
1307 * log to the maximum of what the caller needs, one quarter of the
1308 * log, and 256 blocks.
1309 */
1310 free_threshold = BTOBB(need_bytes);
1311 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1312 free_threshold = MAX(free_threshold, 256);
1313 if (free_blocks < free_threshold) {
1314 threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1315 threshold_cycle = CYCLE_LSN(tail_lsn);
1316 if (threshold_block >= log->l_logBBsize) {
1317 threshold_block -= log->l_logBBsize;
1318 threshold_cycle += 1;
1319 }
1320 ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1321 threshold_block);
1322
1323 /* Don't pass in an lsn greater than the lsn of the last
1324 * log record known to be on disk.
1325 */
1326 if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1327 threshold_lsn = log->l_last_sync_lsn;
1328 }
1329 GRANT_UNLOCK(log, s);
1330
1331 /*
1332 * Get the transaction layer to kick the dirty buffers out to
1333 * disk asynchronously. No point in trying to do this if
1334 * the filesystem is shutting down.
1335 */
1336 if (threshold_lsn &&
1337 !XLOG_FORCED_SHUTDOWN(log))
1338 xfs_trans_push_ail(mp, threshold_lsn);
1339} /* xlog_grant_push_ail */
1340
1341
1342/*
1343 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1344 * fashion. Previously, we should have moved the current iclog
1345 * ptr in the log to point to the next available iclog. This allows further
1346 * write to continue while this code syncs out an iclog ready to go.
1347 * Before an in-core log can be written out, the data section must be scanned
1348 * to save away the 1st word of each BBSIZE block into the header. We replace
1349 * it with the current cycle count. Each BBSIZE block is tagged with the
1350 * cycle count because there in an implicit assumption that drives will
1351 * guarantee that entire 512 byte blocks get written at once. In other words,
1352 * we can't have part of a 512 byte block written and part not written. By
1353 * tagging each block, we will know which blocks are valid when recovering
1354 * after an unclean shutdown.
1355 *
1356 * This routine is single threaded on the iclog. No other thread can be in
1357 * this routine with the same iclog. Changing contents of iclog can there-
1358 * fore be done without grabbing the state machine lock. Updating the global
1359 * log will require grabbing the lock though.
1360 *
1361 * The entire log manager uses a logical block numbering scheme. Only
1362 * log_sync (and then only bwrite()) know about the fact that the log may
1363 * not start with block zero on a given device. The log block start offset
1364 * is added immediately before calling bwrite().
1365 */
1366
1367int
1368xlog_sync(xlog_t *log,
1369 xlog_in_core_t *iclog)
1370{
1371 xfs_caddr_t dptr; /* pointer to byte sized element */
1372 xfs_buf_t *bp;
1373 int i, ops;
1374 uint count; /* byte count of bwrite */
1375 uint count_init; /* initial count before roundup */
1376 int roundoff; /* roundoff to BB or stripe */
1377 int split = 0; /* split write into two regions */
1378 int error;
1379 SPLDECL(s);
1380 int v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1381
1382 XFS_STATS_INC(xs_log_writes);
1383 ASSERT(iclog->ic_refcnt == 0);
1384
1385 /* Add for LR header */
1386 count_init = log->l_iclog_hsize + iclog->ic_offset;
1387
1388 /* Round out the log write size */
1389 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1390 /* we have a v2 stripe unit to use */
1391 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1392 } else {
1393 count = BBTOB(BTOBB(count_init));
1394 }
1395 roundoff = count - count_init;
1396 ASSERT(roundoff >= 0);
1397 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1398 roundoff < log->l_mp->m_sb.sb_logsunit)
1399 ||
1400 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1401 roundoff < BBTOB(1)));
1402
1403 /* move grant heads by roundoff in sync */
1404 s = GRANT_LOCK(log);
Christoph Hellwigdd954c62006-01-11 15:34:50 +11001405 xlog_grant_add_space(log, roundoff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 GRANT_UNLOCK(log, s);
1407
1408 /* put cycle number in every block */
1409 xlog_pack_data(log, iclog, roundoff);
1410
1411 /* real byte length */
1412 if (v2) {
1413 INT_SET(iclog->ic_header.h_len,
1414 ARCH_CONVERT,
1415 iclog->ic_offset + roundoff);
1416 } else {
1417 INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1418 }
1419
1420 /* put ops count in correct order */
1421 ops = iclog->ic_header.h_num_logops;
1422 INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1423
Nathan Scottf5faad72006-07-28 17:04:44 +10001424 bp = iclog->ic_bp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1426 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1427 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1428
1429 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1430
1431 /* Do we need to split this write into 2 parts? */
1432 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1433 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1434 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1435 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1436 } else {
1437 iclog->ic_bwritecnt = 1;
1438 }
1439 XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1440 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */
Nathan Scottf5faad72006-07-28 17:04:44 +10001441 XFS_BUF_ZEROFLAGS(bp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 XFS_BUF_BUSY(bp);
1443 XFS_BUF_ASYNC(bp);
1444 /*
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11001445 * Do an ordered write for the log block.
Nathan Scottf5faad72006-07-28 17:04:44 +10001446 * Its unnecessary to flush the first split block in the log wrap case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 */
Nathan Scottf5faad72006-07-28 17:04:44 +10001448 if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11001449 XFS_BUF_ORDERED(bp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450
1451 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1452 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1453
1454 xlog_verify_iclog(log, iclog, count, B_TRUE);
1455
1456 /* account for log which doesn't start at block #0 */
1457 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1458 /*
1459 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1460 * is shutting down.
1461 */
1462 XFS_BUF_WRITE(bp);
1463
1464 if ((error = XFS_bwrite(bp))) {
1465 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1466 XFS_BUF_ADDR(bp));
Jesper Juhl014c2542006-01-15 02:37:08 +01001467 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 }
1469 if (split) {
Nathan Scottf5faad72006-07-28 17:04:44 +10001470 bp = iclog->ic_log->l_xbuf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1472 (unsigned long)1);
1473 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1474 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1475 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1476 (__psint_t)count), split);
1477 XFS_BUF_SET_FSPRIVATE(bp, iclog);
Nathan Scottf5faad72006-07-28 17:04:44 +10001478 XFS_BUF_ZEROFLAGS(bp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 XFS_BUF_BUSY(bp);
1480 XFS_BUF_ASYNC(bp);
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11001481 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1482 XFS_BUF_ORDERED(bp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 dptr = XFS_BUF_PTR(bp);
1484 /*
1485 * Bump the cycle numbers at the start of each block
1486 * since this part of the buffer is at the start of
1487 * a new cycle. Watch out for the header magic number
1488 * case, though.
1489 */
1490 for (i=0; i<split; i += BBSIZE) {
1491 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1492 if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1493 INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1494 dptr += BBSIZE;
1495 }
1496
1497 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1498 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1499
Nathan Scottc41564b2006-03-29 08:55:14 +10001500 /* account for internal log which doesn't start at block #0 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1502 XFS_BUF_WRITE(bp);
1503 if ((error = XFS_bwrite(bp))) {
1504 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1505 bp, XFS_BUF_ADDR(bp));
Jesper Juhl014c2542006-01-15 02:37:08 +01001506 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 }
1508 }
Jesper Juhl014c2542006-01-15 02:37:08 +01001509 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510} /* xlog_sync */
1511
1512
1513/*
Nathan Scottc41564b2006-03-29 08:55:14 +10001514 * Deallocate a log structure
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 */
1516void
Nathan Scottc41564b2006-03-29 08:55:14 +10001517xlog_dealloc_log(xlog_t *log)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518{
1519 xlog_in_core_t *iclog, *next_iclog;
1520 xlog_ticket_t *tic, *next_tic;
1521 int i;
1522
1523
1524 iclog = log->l_iclog;
1525 for (i=0; i<log->l_iclog_bufs; i++) {
1526 sv_destroy(&iclog->ic_forcesema);
1527 sv_destroy(&iclog->ic_writesema);
1528 xfs_buf_free(iclog->ic_bp);
1529#ifdef XFS_LOG_TRACE
1530 if (iclog->ic_trace != NULL) {
1531 ktrace_free(iclog->ic_trace);
1532 }
1533#endif
1534 next_iclog = iclog->ic_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535 kmem_free(iclog, sizeof(xlog_in_core_t));
1536 iclog = next_iclog;
1537 }
1538 freesema(&log->l_flushsema);
1539 spinlock_destroy(&log->l_icloglock);
1540 spinlock_destroy(&log->l_grant_lock);
1541
1542 /* XXXsup take a look at this again. */
1543 if ((log->l_ticket_cnt != log->l_ticket_tcnt) &&
1544 !XLOG_FORCED_SHUTDOWN(log)) {
1545 xfs_fs_cmn_err(CE_WARN, log->l_mp,
Nathan Scottc41564b2006-03-29 08:55:14 +10001546 "xlog_dealloc_log: (cnt: %d, total: %d)",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 log->l_ticket_cnt, log->l_ticket_tcnt);
1548 /* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1549
1550 } else {
1551 tic = log->l_unmount_free;
1552 while (tic) {
1553 next_tic = tic->t_next;
1554 kmem_free(tic, NBPP);
1555 tic = next_tic;
1556 }
1557 }
1558 xfs_buf_free(log->l_xbuf);
1559#ifdef XFS_LOG_TRACE
1560 if (log->l_trace != NULL) {
1561 ktrace_free(log->l_trace);
1562 }
1563 if (log->l_grant_trace != NULL) {
1564 ktrace_free(log->l_grant_trace);
1565 }
1566#endif
1567 log->l_mp->m_log = NULL;
1568 kmem_free(log, sizeof(xlog_t));
Nathan Scottc41564b2006-03-29 08:55:14 +10001569} /* xlog_dealloc_log */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570
1571/*
1572 * Update counters atomically now that memcpy is done.
1573 */
1574/* ARGSUSED */
1575static inline void
1576xlog_state_finish_copy(xlog_t *log,
1577 xlog_in_core_t *iclog,
1578 int record_cnt,
1579 int copy_bytes)
1580{
1581 SPLDECL(s);
1582
1583 s = LOG_LOCK(log);
1584
1585 iclog->ic_header.h_num_logops += record_cnt;
1586 iclog->ic_offset += copy_bytes;
1587
1588 LOG_UNLOCK(log, s);
1589} /* xlog_state_finish_copy */
1590
1591
1592
1593
1594/*
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001595 * print out info relating to regions written which consume
1596 * the reservation
1597 */
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001598STATIC void
1599xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1600{
1601 uint i;
1602 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1603
1604 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1605 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1606 "bformat",
1607 "bchunk",
1608 "efi_format",
1609 "efd_format",
1610 "iformat",
1611 "icore",
1612 "iext",
1613 "ibroot",
1614 "ilocal",
1615 "iattr_ext",
1616 "iattr_broot",
1617 "iattr_local",
1618 "qformat",
1619 "dquot",
1620 "quotaoff",
1621 "LR header",
1622 "unmount",
1623 "commit",
1624 "trans header"
1625 };
1626 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1627 "SETATTR_NOT_SIZE",
1628 "SETATTR_SIZE",
1629 "INACTIVE",
1630 "CREATE",
1631 "CREATE_TRUNC",
1632 "TRUNCATE_FILE",
1633 "REMOVE",
1634 "LINK",
1635 "RENAME",
1636 "MKDIR",
1637 "RMDIR",
1638 "SYMLINK",
1639 "SET_DMATTRS",
1640 "GROWFS",
1641 "STRAT_WRITE",
1642 "DIOSTRAT",
1643 "WRITE_SYNC",
1644 "WRITEID",
1645 "ADDAFORK",
1646 "ATTRINVAL",
1647 "ATRUNCATE",
1648 "ATTR_SET",
1649 "ATTR_RM",
1650 "ATTR_FLAG",
1651 "CLEAR_AGI_BUCKET",
1652 "QM_SBCHANGE",
1653 "DUMMY1",
1654 "DUMMY2",
1655 "QM_QUOTAOFF",
1656 "QM_DQALLOC",
1657 "QM_SETQLIM",
1658 "QM_DQCLUSTER",
1659 "QM_QINOCREATE",
1660 "QM_QUOTAOFF_END",
1661 "SB_UNIT",
1662 "FSYNC_TS",
1663 "GROWFSRT_ALLOC",
1664 "GROWFSRT_ZERO",
1665 "GROWFSRT_FREE",
1666 "SWAPEXT"
1667 };
1668
1669 xfs_fs_cmn_err(CE_WARN, mp,
1670 "xfs_log_write: reservation summary:\n"
1671 " trans type = %s (%u)\n"
1672 " unit res = %d bytes\n"
1673 " current res = %d bytes\n"
1674 " total reg = %u bytes (o/flow = %u bytes)\n"
1675 " ophdrs = %u (ophdr space = %u bytes)\n"
1676 " ophdr + reg = %u bytes\n"
1677 " num regions = %u\n",
1678 ((ticket->t_trans_type <= 0 ||
1679 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1680 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1681 ticket->t_trans_type,
1682 ticket->t_unit_res,
1683 ticket->t_curr_res,
1684 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1685 ticket->t_res_num_ophdrs, ophdr_spc,
1686 ticket->t_res_arr_sum +
Tim Shimmin12598452006-01-11 21:02:47 +11001687 ticket->t_res_o_flow + ophdr_spc,
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001688 ticket->t_res_num);
1689
1690 for (i = 0; i < ticket->t_res_num; i++) {
Tim Shimmin12598452006-01-11 21:02:47 +11001691 uint r_type = ticket->t_res_arr[i].r_type;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001692 cmn_err(CE_WARN,
1693 "region[%u]: %s - %u bytes\n",
1694 i,
1695 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1696 "bad-rtype" : res_type_str[r_type-1]),
1697 ticket->t_res_arr[i].r_len);
1698 }
1699}
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001700
1701/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 * Write some region out to in-core log
1703 *
1704 * This will be called when writing externally provided regions or when
1705 * writing out a commit record for a given transaction.
1706 *
1707 * General algorithm:
1708 * 1. Find total length of this write. This may include adding to the
1709 * lengths passed in.
1710 * 2. Check whether we violate the tickets reservation.
1711 * 3. While writing to this iclog
1712 * A. Reserve as much space in this iclog as can get
1713 * B. If this is first write, save away start lsn
1714 * C. While writing this region:
1715 * 1. If first write of transaction, write start record
1716 * 2. Write log operation header (header per region)
1717 * 3. Find out if we can fit entire region into this iclog
1718 * 4. Potentially, verify destination memcpy ptr
1719 * 5. Memcpy (partial) region
1720 * 6. If partial copy, release iclog; otherwise, continue
1721 * copying more regions into current iclog
1722 * 4. Mark want sync bit (in simulation mode)
1723 * 5. Release iclog for potential flush to on-disk log.
1724 *
1725 * ERRORS:
1726 * 1. Panic if reservation is overrun. This should never happen since
1727 * reservation amounts are generated internal to the filesystem.
1728 * NOTES:
1729 * 1. Tickets are single threaded data structures.
1730 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1731 * syncing routine. When a single log_write region needs to span
1732 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1733 * on all log operation writes which don't contain the end of the
1734 * region. The XLOG_END_TRANS bit is used for the in-core log
1735 * operation which contains the end of the continued log_write region.
1736 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1737 * we don't really know exactly how much space will be used. As a result,
1738 * we don't update ic_offset until the end when we know exactly how many
1739 * bytes have been written out.
1740 */
1741int
1742xlog_write(xfs_mount_t * mp,
1743 xfs_log_iovec_t reg[],
1744 int nentries,
1745 xfs_log_ticket_t tic,
1746 xfs_lsn_t *start_lsn,
1747 xlog_in_core_t **commit_iclog,
1748 uint flags)
1749{
Nathan Scott5493a0f2006-06-28 11:17:28 +10001750 xlog_t *log = mp->m_log;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 xlog_ticket_t *ticket = (xlog_ticket_t *)tic;
Nathan Scott5493a0f2006-06-28 11:17:28 +10001752 xlog_in_core_t *iclog = NULL; /* ptr to current in-core log */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 xlog_op_header_t *logop_head; /* ptr to log operation header */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 __psint_t ptr; /* copy address into data region */
1755 int len; /* # xlog_write() bytes 2 still copy */
1756 int index; /* region index currently copying */
1757 int log_offset; /* offset (from 0) into data region */
1758 int start_rec_copy; /* # bytes to copy for start record */
1759 int partial_copy; /* did we split a region? */
1760 int partial_copy_len;/* # bytes copied if split region */
1761 int need_copy; /* # bytes need to memcpy this region */
1762 int copy_len; /* # bytes actually memcpy'ing */
1763 int copy_off; /* # bytes from entry start */
1764 int contwr; /* continued write of in-core log? */
1765 int error;
1766 int record_cnt = 0, data_cnt = 0;
1767
1768 partial_copy_len = partial_copy = 0;
1769
1770 /* Calculate potential maximum space. Each region gets its own
1771 * xlog_op_header_t and may need to be double word aligned.
1772 */
1773 len = 0;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001774 if (ticket->t_flags & XLOG_TIC_INITED) { /* acct for start rec of xact */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 len += sizeof(xlog_op_header_t);
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001776 XLOG_TIC_ADD_OPHDR(ticket);
1777 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778
1779 for (index = 0; index < nentries; index++) {
1780 len += sizeof(xlog_op_header_t); /* each region gets >= 1 */
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001781 XLOG_TIC_ADD_OPHDR(ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 len += reg[index].i_len;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001783 XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 }
1785 contwr = *start_lsn = 0;
1786
1787 if (ticket->t_curr_res < len) {
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001788 xlog_print_tic_res(mp, ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789#ifdef DEBUG
1790 xlog_panic(
1791 "xfs_log_write: reservation ran out. Need to up reservation");
1792#else
1793 /* Customer configurable panic */
1794 xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1795 "xfs_log_write: reservation ran out. Need to up reservation");
1796 /* If we did not panic, shutdown the filesystem */
Nathan Scott7d04a332006-06-09 14:58:38 +10001797 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798#endif
1799 } else
1800 ticket->t_curr_res -= len;
1801
1802 for (index = 0; index < nentries; ) {
1803 if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1804 &contwr, &log_offset)))
Jesper Juhl014c2542006-01-15 02:37:08 +01001805 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
1807 ASSERT(log_offset <= iclog->ic_size - 1);
1808 ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1809
1810 /* start_lsn is the first lsn written to. That's all we need. */
1811 if (! *start_lsn)
1812 *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1813
1814 /* This loop writes out as many regions as can fit in the amount
1815 * of space which was allocated by xlog_state_get_iclog_space().
1816 */
1817 while (index < nentries) {
1818 ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1819 ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1820 start_rec_copy = 0;
1821
1822 /* If first write for transaction, insert start record.
1823 * We can't be trying to commit if we are inited. We can't
1824 * have any "partial_copy" if we are inited.
1825 */
1826 if (ticket->t_flags & XLOG_TIC_INITED) {
1827 logop_head = (xlog_op_header_t *)ptr;
1828 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1829 logop_head->oh_clientid = ticket->t_clientid;
1830 logop_head->oh_len = 0;
1831 logop_head->oh_flags = XLOG_START_TRANS;
1832 logop_head->oh_res2 = 0;
1833 ticket->t_flags &= ~XLOG_TIC_INITED; /* clear bit */
1834 record_cnt++;
1835
1836 start_rec_copy = sizeof(xlog_op_header_t);
1837 xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1838 }
1839
1840 /* Copy log operation header directly into data section */
1841 logop_head = (xlog_op_header_t *)ptr;
1842 INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1843 logop_head->oh_clientid = ticket->t_clientid;
1844 logop_head->oh_res2 = 0;
1845
1846 /* header copied directly */
1847 xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1848
1849 /* are we copying a commit or unmount record? */
1850 logop_head->oh_flags = flags;
1851
1852 /*
1853 * We've seen logs corrupted with bad transaction client
1854 * ids. This makes sure that XFS doesn't generate them on.
1855 * Turn this into an EIO and shut down the filesystem.
1856 */
1857 switch (logop_head->oh_clientid) {
1858 case XFS_TRANSACTION:
1859 case XFS_VOLUME:
1860 case XFS_LOG:
1861 break;
1862 default:
1863 xfs_fs_cmn_err(CE_WARN, mp,
1864 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1865 logop_head->oh_clientid, tic);
1866 return XFS_ERROR(EIO);
1867 }
1868
1869 /* Partial write last time? => (partial_copy != 0)
1870 * need_copy is the amount we'd like to copy if everything could
1871 * fit in the current memcpy.
1872 */
1873 need_copy = reg[index].i_len - partial_copy_len;
1874
1875 copy_off = partial_copy_len;
1876 if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1877 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1878 if (partial_copy)
1879 logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1880 partial_copy_len = partial_copy = 0;
1881 } else { /* partial write */
1882 copy_len = iclog->ic_size - log_offset;
1883 INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1884 logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1885 if (partial_copy)
1886 logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1887 partial_copy_len += copy_len;
1888 partial_copy++;
1889 len += sizeof(xlog_op_header_t); /* from splitting of region */
1890 /* account for new log op header */
1891 ticket->t_curr_res -= sizeof(xlog_op_header_t);
Tim Shimmin7e9c6392005-09-02 16:42:05 +10001892 XLOG_TIC_ADD_OPHDR(ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 }
1894 xlog_verify_dest_ptr(log, ptr);
1895
1896 /* copy region */
1897 ASSERT(copy_len >= 0);
1898 memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1899 xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1900
1901 /* make copy_len total bytes copied, including headers */
1902 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1903 record_cnt++;
1904 data_cnt += contwr ? copy_len : 0;
1905 if (partial_copy) { /* copied partial region */
1906 /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1907 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1908 record_cnt = data_cnt = 0;
1909 if ((error = xlog_state_release_iclog(log, iclog)))
Jesper Juhl014c2542006-01-15 02:37:08 +01001910 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 break; /* don't increment index */
1912 } else { /* copied entire region */
1913 index++;
1914 partial_copy_len = partial_copy = 0;
1915
1916 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1917 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1918 record_cnt = data_cnt = 0;
1919 xlog_state_want_sync(log, iclog);
1920 if (commit_iclog) {
1921 ASSERT(flags & XLOG_COMMIT_TRANS);
1922 *commit_iclog = iclog;
1923 } else if ((error = xlog_state_release_iclog(log, iclog)))
Jesper Juhl014c2542006-01-15 02:37:08 +01001924 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 if (index == nentries)
1926 return 0; /* we are done */
1927 else
1928 break;
1929 }
1930 } /* if (partial_copy) */
1931 } /* while (index < nentries) */
1932 } /* for (index = 0; index < nentries; ) */
1933 ASSERT(len == 0);
1934
1935 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1936 if (commit_iclog) {
1937 ASSERT(flags & XLOG_COMMIT_TRANS);
1938 *commit_iclog = iclog;
1939 return 0;
1940 }
Jesper Juhl014c2542006-01-15 02:37:08 +01001941 return xlog_state_release_iclog(log, iclog);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942} /* xlog_write */
1943
1944
1945/*****************************************************************************
1946 *
1947 * State Machine functions
1948 *
1949 *****************************************************************************
1950 */
1951
1952/* Clean iclogs starting from the head. This ordering must be
1953 * maintained, so an iclog doesn't become ACTIVE beyond one that
1954 * is SYNCING. This is also required to maintain the notion that we use
1955 * a counting semaphore to hold off would be writers to the log when every
1956 * iclog is trying to sync to disk.
1957 *
1958 * State Change: DIRTY -> ACTIVE
1959 */
Christoph Hellwigba0f32d2005-06-21 15:36:52 +10001960STATIC void
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961xlog_state_clean_log(xlog_t *log)
1962{
1963 xlog_in_core_t *iclog;
1964 int changed = 0;
1965
1966 iclog = log->l_iclog;
1967 do {
1968 if (iclog->ic_state == XLOG_STATE_DIRTY) {
1969 iclog->ic_state = XLOG_STATE_ACTIVE;
1970 iclog->ic_offset = 0;
1971 iclog->ic_callback = NULL; /* don't need to free */
1972 /*
1973 * If the number of ops in this iclog indicate it just
1974 * contains the dummy transaction, we can
1975 * change state into IDLE (the second time around).
1976 * Otherwise we should change the state into
1977 * NEED a dummy.
1978 * We don't need to cover the dummy.
1979 */
1980 if (!changed &&
1981 (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
1982 changed = 1;
1983 } else {
1984 /*
1985 * We have two dirty iclogs so start over
1986 * This could also be num of ops indicates
1987 * this is not the dummy going out.
1988 */
1989 changed = 2;
1990 }
1991 iclog->ic_header.h_num_logops = 0;
1992 memset(iclog->ic_header.h_cycle_data, 0,
1993 sizeof(iclog->ic_header.h_cycle_data));
1994 iclog->ic_header.h_lsn = 0;
1995 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1996 /* do nothing */;
1997 else
1998 break; /* stop cleaning */
1999 iclog = iclog->ic_next;
2000 } while (iclog != log->l_iclog);
2001
2002 /* log is locked when we are called */
2003 /*
2004 * Change state for the dummy log recording.
2005 * We usually go to NEED. But we go to NEED2 if the changed indicates
2006 * we are done writing the dummy record.
2007 * If we are done with the second dummy recored (DONE2), then
2008 * we go to IDLE.
2009 */
2010 if (changed) {
2011 switch (log->l_covered_state) {
2012 case XLOG_STATE_COVER_IDLE:
2013 case XLOG_STATE_COVER_NEED:
2014 case XLOG_STATE_COVER_NEED2:
2015 log->l_covered_state = XLOG_STATE_COVER_NEED;
2016 break;
2017
2018 case XLOG_STATE_COVER_DONE:
2019 if (changed == 1)
2020 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2021 else
2022 log->l_covered_state = XLOG_STATE_COVER_NEED;
2023 break;
2024
2025 case XLOG_STATE_COVER_DONE2:
2026 if (changed == 1)
2027 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2028 else
2029 log->l_covered_state = XLOG_STATE_COVER_NEED;
2030 break;
2031
2032 default:
2033 ASSERT(0);
2034 }
2035 }
2036} /* xlog_state_clean_log */
2037
2038STATIC xfs_lsn_t
2039xlog_get_lowest_lsn(
2040 xlog_t *log)
2041{
2042 xlog_in_core_t *lsn_log;
2043 xfs_lsn_t lowest_lsn, lsn;
2044
2045 lsn_log = log->l_iclog;
2046 lowest_lsn = 0;
2047 do {
2048 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2049 lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
2050 if ((lsn && !lowest_lsn) ||
2051 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2052 lowest_lsn = lsn;
2053 }
2054 }
2055 lsn_log = lsn_log->ic_next;
2056 } while (lsn_log != log->l_iclog);
Jesper Juhl014c2542006-01-15 02:37:08 +01002057 return lowest_lsn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058}
2059
2060
2061STATIC void
2062xlog_state_do_callback(
2063 xlog_t *log,
2064 int aborted,
2065 xlog_in_core_t *ciclog)
2066{
2067 xlog_in_core_t *iclog;
2068 xlog_in_core_t *first_iclog; /* used to know when we've
2069 * processed all iclogs once */
2070 xfs_log_callback_t *cb, *cb_next;
2071 int flushcnt = 0;
2072 xfs_lsn_t lowest_lsn;
2073 int ioerrors; /* counter: iclogs with errors */
2074 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2075 int funcdidcallbacks; /* flag: function did callbacks */
2076 int repeats; /* for issuing console warnings if
2077 * looping too many times */
2078 SPLDECL(s);
2079
2080 s = LOG_LOCK(log);
2081 first_iclog = iclog = log->l_iclog;
2082 ioerrors = 0;
2083 funcdidcallbacks = 0;
2084 repeats = 0;
2085
2086 do {
2087 /*
2088 * Scan all iclogs starting with the one pointed to by the
2089 * log. Reset this starting point each time the log is
2090 * unlocked (during callbacks).
2091 *
2092 * Keep looping through iclogs until one full pass is made
2093 * without running any callbacks.
2094 */
2095 first_iclog = log->l_iclog;
2096 iclog = log->l_iclog;
2097 loopdidcallbacks = 0;
2098 repeats++;
2099
2100 do {
2101
2102 /* skip all iclogs in the ACTIVE & DIRTY states */
2103 if (iclog->ic_state &
2104 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2105 iclog = iclog->ic_next;
2106 continue;
2107 }
2108
2109 /*
2110 * Between marking a filesystem SHUTDOWN and stopping
2111 * the log, we do flush all iclogs to disk (if there
2112 * wasn't a log I/O error). So, we do want things to
2113 * go smoothly in case of just a SHUTDOWN w/o a
2114 * LOG_IO_ERROR.
2115 */
2116 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2117 /*
2118 * Can only perform callbacks in order. Since
2119 * this iclog is not in the DONE_SYNC/
2120 * DO_CALLBACK state, we skip the rest and
2121 * just try to clean up. If we set our iclog
2122 * to DO_CALLBACK, we will not process it when
2123 * we retry since a previous iclog is in the
2124 * CALLBACK and the state cannot change since
2125 * we are holding the LOG_LOCK.
2126 */
2127 if (!(iclog->ic_state &
2128 (XLOG_STATE_DONE_SYNC |
2129 XLOG_STATE_DO_CALLBACK))) {
2130 if (ciclog && (ciclog->ic_state ==
2131 XLOG_STATE_DONE_SYNC)) {
2132 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2133 }
2134 break;
2135 }
2136 /*
2137 * We now have an iclog that is in either the
2138 * DO_CALLBACK or DONE_SYNC states. The other
2139 * states (WANT_SYNC, SYNCING, or CALLBACK were
2140 * caught by the above if and are going to
2141 * clean (i.e. we aren't doing their callbacks)
2142 * see the above if.
2143 */
2144
2145 /*
2146 * We will do one more check here to see if we
2147 * have chased our tail around.
2148 */
2149
2150 lowest_lsn = xlog_get_lowest_lsn(log);
2151 if (lowest_lsn && (
2152 XFS_LSN_CMP(
2153 lowest_lsn,
2154 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2155 )<0)) {
2156 iclog = iclog->ic_next;
2157 continue; /* Leave this iclog for
2158 * another thread */
2159 }
2160
2161 iclog->ic_state = XLOG_STATE_CALLBACK;
2162
2163 LOG_UNLOCK(log, s);
2164
2165 /* l_last_sync_lsn field protected by
2166 * GRANT_LOCK. Don't worry about iclog's lsn.
2167 * No one else can be here except us.
2168 */
2169 s = GRANT_LOCK(log);
2170 ASSERT(XFS_LSN_CMP(
2171 log->l_last_sync_lsn,
2172 INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2173 )<=0);
2174 log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2175 GRANT_UNLOCK(log, s);
2176
2177 /*
2178 * Keep processing entries in the callback list
2179 * until we come around and it is empty. We
2180 * need to atomically see that the list is
2181 * empty and change the state to DIRTY so that
2182 * we don't miss any more callbacks being added.
2183 */
2184 s = LOG_LOCK(log);
2185 } else {
2186 ioerrors++;
2187 }
2188 cb = iclog->ic_callback;
2189
2190 while (cb != 0) {
2191 iclog->ic_callback_tail = &(iclog->ic_callback);
2192 iclog->ic_callback = NULL;
2193 LOG_UNLOCK(log, s);
2194
2195 /* perform callbacks in the order given */
2196 for (; cb != 0; cb = cb_next) {
2197 cb_next = cb->cb_next;
2198 cb->cb_func(cb->cb_arg, aborted);
2199 }
2200 s = LOG_LOCK(log);
2201 cb = iclog->ic_callback;
2202 }
2203
2204 loopdidcallbacks++;
2205 funcdidcallbacks++;
2206
2207 ASSERT(iclog->ic_callback == 0);
2208 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2209 iclog->ic_state = XLOG_STATE_DIRTY;
2210
2211 /*
2212 * Transition from DIRTY to ACTIVE if applicable.
2213 * NOP if STATE_IOERROR.
2214 */
2215 xlog_state_clean_log(log);
2216
2217 /* wake up threads waiting in xfs_log_force() */
2218 sv_broadcast(&iclog->ic_forcesema);
2219
2220 iclog = iclog->ic_next;
2221 } while (first_iclog != iclog);
Nathan Scotta3c66852006-09-28 11:02:14 +10002222
2223 if (repeats > 5000) {
2224 flushcnt += repeats;
2225 repeats = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 xfs_fs_cmn_err(CE_WARN, log->l_mp,
Nathan Scotta3c66852006-09-28 11:02:14 +10002227 "%s: possible infinite loop (%d iterations)",
2228 __FUNCTION__, flushcnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 }
2230 } while (!ioerrors && loopdidcallbacks);
2231
2232 /*
2233 * make one last gasp attempt to see if iclogs are being left in
2234 * limbo..
2235 */
2236#ifdef DEBUG
2237 if (funcdidcallbacks) {
2238 first_iclog = iclog = log->l_iclog;
2239 do {
2240 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2241 /*
2242 * Terminate the loop if iclogs are found in states
2243 * which will cause other threads to clean up iclogs.
2244 *
2245 * SYNCING - i/o completion will go through logs
2246 * DONE_SYNC - interrupt thread should be waiting for
2247 * LOG_LOCK
2248 * IOERROR - give up hope all ye who enter here
2249 */
2250 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2251 iclog->ic_state == XLOG_STATE_SYNCING ||
2252 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2253 iclog->ic_state == XLOG_STATE_IOERROR )
2254 break;
2255 iclog = iclog->ic_next;
2256 } while (first_iclog != iclog);
2257 }
2258#endif
2259
Nathan Scotta3c66852006-09-28 11:02:14 +10002260 flushcnt = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2262 flushcnt = log->l_flushcnt;
2263 log->l_flushcnt = 0;
2264 }
2265 LOG_UNLOCK(log, s);
2266 while (flushcnt--)
2267 vsema(&log->l_flushsema);
2268} /* xlog_state_do_callback */
2269
2270
2271/*
2272 * Finish transitioning this iclog to the dirty state.
2273 *
2274 * Make sure that we completely execute this routine only when this is
2275 * the last call to the iclog. There is a good chance that iclog flushes,
2276 * when we reach the end of the physical log, get turned into 2 separate
2277 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2278 * routine. By using the reference count bwritecnt, we guarantee that only
2279 * the second completion goes through.
2280 *
2281 * Callbacks could take time, so they are done outside the scope of the
2282 * global state machine log lock. Assume that the calls to cvsema won't
2283 * take a long time. At least we know it won't sleep.
2284 */
2285void
2286xlog_state_done_syncing(
2287 xlog_in_core_t *iclog,
2288 int aborted)
2289{
2290 xlog_t *log = iclog->ic_log;
2291 SPLDECL(s);
2292
2293 s = LOG_LOCK(log);
2294
2295 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2296 iclog->ic_state == XLOG_STATE_IOERROR);
2297 ASSERT(iclog->ic_refcnt == 0);
2298 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2299
2300
2301 /*
2302 * If we got an error, either on the first buffer, or in the case of
2303 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2304 * and none should ever be attempted to be written to disk
2305 * again.
2306 */
2307 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2308 if (--iclog->ic_bwritecnt == 1) {
2309 LOG_UNLOCK(log, s);
2310 return;
2311 }
2312 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2313 }
2314
2315 /*
2316 * Someone could be sleeping prior to writing out the next
2317 * iclog buffer, we wake them all, one will get to do the
2318 * I/O, the others get to wait for the result.
2319 */
2320 sv_broadcast(&iclog->ic_writesema);
2321 LOG_UNLOCK(log, s);
2322 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2323} /* xlog_state_done_syncing */
2324
2325
2326/*
2327 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2328 * sleep. The flush semaphore is set to the number of in-core buffers and
2329 * decremented around disk syncing. Therefore, if all buffers are syncing,
2330 * this semaphore will cause new writes to sleep until a sync completes.
2331 * Otherwise, this code just does p() followed by v(). This approximates
2332 * a sleep/wakeup except we can't race.
2333 *
2334 * The in-core logs are used in a circular fashion. They are not used
2335 * out-of-order even when an iclog past the head is free.
2336 *
2337 * return:
2338 * * log_offset where xlog_write() can start writing into the in-core
2339 * log's data space.
2340 * * in-core log pointer to which xlog_write() should write.
2341 * * boolean indicating this is a continued write to an in-core log.
2342 * If this is the last write, then the in-core log's offset field
2343 * needs to be incremented, depending on the amount of data which
2344 * is copied.
2345 */
2346int
2347xlog_state_get_iclog_space(xlog_t *log,
2348 int len,
2349 xlog_in_core_t **iclogp,
2350 xlog_ticket_t *ticket,
2351 int *continued_write,
2352 int *logoffsetp)
2353{
2354 SPLDECL(s);
2355 int log_offset;
2356 xlog_rec_header_t *head;
2357 xlog_in_core_t *iclog;
2358 int error;
2359
2360restart:
2361 s = LOG_LOCK(log);
2362 if (XLOG_FORCED_SHUTDOWN(log)) {
2363 LOG_UNLOCK(log, s);
2364 return XFS_ERROR(EIO);
2365 }
2366
2367 iclog = log->l_iclog;
2368 if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2369 log->l_flushcnt++;
2370 LOG_UNLOCK(log, s);
2371 xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2372 XFS_STATS_INC(xs_log_noiclogs);
2373 /* Ensure that log writes happen */
2374 psema(&log->l_flushsema, PINOD);
2375 goto restart;
2376 }
2377 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2378 head = &iclog->ic_header;
2379
2380 iclog->ic_refcnt++; /* prevents sync */
2381 log_offset = iclog->ic_offset;
2382
2383 /* On the 1st write to an iclog, figure out lsn. This works
2384 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2385 * committing to. If the offset is set, that's how many blocks
2386 * must be written.
2387 */
2388 if (log_offset == 0) {
2389 ticket->t_curr_res -= log->l_iclog_hsize;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10002390 XLOG_TIC_ADD_REGION(ticket,
2391 log->l_iclog_hsize,
2392 XLOG_REG_TYPE_LRHEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2394 ASSIGN_LSN(head->h_lsn, log);
2395 ASSERT(log->l_curr_block >= 0);
2396 }
2397
2398 /* If there is enough room to write everything, then do it. Otherwise,
2399 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2400 * bit is on, so this will get flushed out. Don't update ic_offset
2401 * until you know exactly how many bytes get copied. Therefore, wait
2402 * until later to update ic_offset.
2403 *
2404 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2405 * can fit into remaining data section.
2406 */
2407 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2408 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2409
2410 /* If I'm the only one writing to this iclog, sync it to disk */
2411 if (iclog->ic_refcnt == 1) {
2412 LOG_UNLOCK(log, s);
2413 if ((error = xlog_state_release_iclog(log, iclog)))
Jesper Juhl014c2542006-01-15 02:37:08 +01002414 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 } else {
2416 iclog->ic_refcnt--;
2417 LOG_UNLOCK(log, s);
2418 }
2419 goto restart;
2420 }
2421
2422 /* Do we have enough room to write the full amount in the remainder
2423 * of this iclog? Or must we continue a write on the next iclog and
2424 * mark this iclog as completely taken? In the case where we switch
2425 * iclogs (to mark it taken), this particular iclog will release/sync
2426 * to disk in xlog_write().
2427 */
2428 if (len <= iclog->ic_size - iclog->ic_offset) {
2429 *continued_write = 0;
2430 iclog->ic_offset += len;
2431 } else {
2432 *continued_write = 1;
2433 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2434 }
2435 *iclogp = iclog;
2436
2437 ASSERT(iclog->ic_offset <= iclog->ic_size);
2438 LOG_UNLOCK(log, s);
2439
2440 *logoffsetp = log_offset;
2441 return 0;
2442} /* xlog_state_get_iclog_space */
2443
2444/*
2445 * Atomically get the log space required for a log ticket.
2446 *
2447 * Once a ticket gets put onto the reserveq, it will only return after
2448 * the needed reservation is satisfied.
2449 */
2450STATIC int
2451xlog_grant_log_space(xlog_t *log,
2452 xlog_ticket_t *tic)
2453{
2454 int free_bytes;
2455 int need_bytes;
2456 SPLDECL(s);
2457#ifdef DEBUG
2458 xfs_lsn_t tail_lsn;
2459#endif
2460
2461
2462#ifdef DEBUG
2463 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2464 panic("grant Recovery problem");
2465#endif
2466
2467 /* Is there space or do we need to sleep? */
2468 s = GRANT_LOCK(log);
2469 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2470
2471 /* something is already sleeping; insert new transaction at end */
2472 if (log->l_reserve_headq) {
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002473 xlog_ins_ticketq(&log->l_reserve_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 xlog_trace_loggrant(log, tic,
2475 "xlog_grant_log_space: sleep 1");
2476 /*
2477 * Gotta check this before going to sleep, while we're
2478 * holding the grant lock.
2479 */
2480 if (XLOG_FORCED_SHUTDOWN(log))
2481 goto error_return;
2482
2483 XFS_STATS_INC(xs_sleep_logspace);
2484 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2485 /*
2486 * If we got an error, and the filesystem is shutting down,
2487 * we'll catch it down below. So just continue...
2488 */
2489 xlog_trace_loggrant(log, tic,
2490 "xlog_grant_log_space: wake 1");
2491 s = GRANT_LOCK(log);
2492 }
2493 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2494 need_bytes = tic->t_unit_res*tic->t_ocnt;
2495 else
2496 need_bytes = tic->t_unit_res;
2497
2498redo:
2499 if (XLOG_FORCED_SHUTDOWN(log))
2500 goto error_return;
2501
2502 free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2503 log->l_grant_reserve_bytes);
2504 if (free_bytes < need_bytes) {
2505 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002506 xlog_ins_ticketq(&log->l_reserve_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 xlog_trace_loggrant(log, tic,
2508 "xlog_grant_log_space: sleep 2");
2509 XFS_STATS_INC(xs_sleep_logspace);
2510 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2511
2512 if (XLOG_FORCED_SHUTDOWN(log)) {
2513 s = GRANT_LOCK(log);
2514 goto error_return;
2515 }
2516
2517 xlog_trace_loggrant(log, tic,
2518 "xlog_grant_log_space: wake 2");
2519 xlog_grant_push_ail(log->l_mp, need_bytes);
2520 s = GRANT_LOCK(log);
2521 goto redo;
2522 } else if (tic->t_flags & XLOG_TIC_IN_Q)
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002523 xlog_del_ticketq(&log->l_reserve_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524
2525 /* we've got enough space */
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002526 xlog_grant_add_space(log, need_bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527#ifdef DEBUG
2528 tail_lsn = log->l_tail_lsn;
2529 /*
2530 * Check to make sure the grant write head didn't just over lap the
2531 * tail. If the cycles are the same, we can't be overlapping.
2532 * Otherwise, make sure that the cycles differ by exactly one and
2533 * check the byte count.
2534 */
2535 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2536 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2537 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2538 }
2539#endif
2540 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2541 xlog_verify_grant_head(log, 1);
2542 GRANT_UNLOCK(log, s);
2543 return 0;
2544
2545 error_return:
2546 if (tic->t_flags & XLOG_TIC_IN_Q)
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002547 xlog_del_ticketq(&log->l_reserve_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2549 /*
2550 * If we are failing, make sure the ticket doesn't have any
2551 * current reservations. We don't want to add this back when
2552 * the ticket/transaction gets cancelled.
2553 */
2554 tic->t_curr_res = 0;
2555 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2556 GRANT_UNLOCK(log, s);
2557 return XFS_ERROR(EIO);
2558} /* xlog_grant_log_space */
2559
2560
2561/*
2562 * Replenish the byte reservation required by moving the grant write head.
2563 *
2564 *
2565 */
2566STATIC int
2567xlog_regrant_write_log_space(xlog_t *log,
2568 xlog_ticket_t *tic)
2569{
2570 SPLDECL(s);
2571 int free_bytes, need_bytes;
2572 xlog_ticket_t *ntic;
2573#ifdef DEBUG
2574 xfs_lsn_t tail_lsn;
2575#endif
2576
2577 tic->t_curr_res = tic->t_unit_res;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10002578 XLOG_TIC_RESET_RES(tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579
2580 if (tic->t_cnt > 0)
Jesper Juhl014c2542006-01-15 02:37:08 +01002581 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582
2583#ifdef DEBUG
2584 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2585 panic("regrant Recovery problem");
2586#endif
2587
2588 s = GRANT_LOCK(log);
2589 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2590
2591 if (XLOG_FORCED_SHUTDOWN(log))
2592 goto error_return;
2593
2594 /* If there are other waiters on the queue then give them a
2595 * chance at logspace before us. Wake up the first waiters,
2596 * if we do not wake up all the waiters then go to sleep waiting
2597 * for more free space, otherwise try to get some space for
2598 * this transaction.
2599 */
2600
2601 if ((ntic = log->l_write_headq)) {
2602 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2603 log->l_grant_write_bytes);
2604 do {
2605 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2606
2607 if (free_bytes < ntic->t_unit_res)
2608 break;
2609 free_bytes -= ntic->t_unit_res;
2610 sv_signal(&ntic->t_sema);
2611 ntic = ntic->t_next;
2612 } while (ntic != log->l_write_headq);
2613
2614 if (ntic != log->l_write_headq) {
2615 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002616 xlog_ins_ticketq(&log->l_write_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617
2618 xlog_trace_loggrant(log, tic,
2619 "xlog_regrant_write_log_space: sleep 1");
2620 XFS_STATS_INC(xs_sleep_logspace);
2621 sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2622 &log->l_grant_lock, s);
2623
2624 /* If we're shutting down, this tic is already
2625 * off the queue */
2626 if (XLOG_FORCED_SHUTDOWN(log)) {
2627 s = GRANT_LOCK(log);
2628 goto error_return;
2629 }
2630
2631 xlog_trace_loggrant(log, tic,
2632 "xlog_regrant_write_log_space: wake 1");
2633 xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2634 s = GRANT_LOCK(log);
2635 }
2636 }
2637
2638 need_bytes = tic->t_unit_res;
2639
2640redo:
2641 if (XLOG_FORCED_SHUTDOWN(log))
2642 goto error_return;
2643
2644 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2645 log->l_grant_write_bytes);
2646 if (free_bytes < need_bytes) {
2647 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002648 xlog_ins_ticketq(&log->l_write_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 XFS_STATS_INC(xs_sleep_logspace);
2650 sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2651
2652 /* If we're shutting down, this tic is already off the queue */
2653 if (XLOG_FORCED_SHUTDOWN(log)) {
2654 s = GRANT_LOCK(log);
2655 goto error_return;
2656 }
2657
2658 xlog_trace_loggrant(log, tic,
2659 "xlog_regrant_write_log_space: wake 2");
2660 xlog_grant_push_ail(log->l_mp, need_bytes);
2661 s = GRANT_LOCK(log);
2662 goto redo;
2663 } else if (tic->t_flags & XLOG_TIC_IN_Q)
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002664 xlog_del_ticketq(&log->l_write_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002666 /* we've got enough space */
2667 xlog_grant_add_space_write(log, need_bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668#ifdef DEBUG
2669 tail_lsn = log->l_tail_lsn;
2670 if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2671 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2672 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2673 }
2674#endif
2675
2676 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2677 xlog_verify_grant_head(log, 1);
2678 GRANT_UNLOCK(log, s);
Jesper Juhl014c2542006-01-15 02:37:08 +01002679 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680
2681
2682 error_return:
2683 if (tic->t_flags & XLOG_TIC_IN_Q)
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002684 xlog_del_ticketq(&log->l_reserve_headq, tic);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2686 /*
2687 * If we are failing, make sure the ticket doesn't have any
2688 * current reservations. We don't want to add this back when
2689 * the ticket/transaction gets cancelled.
2690 */
2691 tic->t_curr_res = 0;
2692 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2693 GRANT_UNLOCK(log, s);
2694 return XFS_ERROR(EIO);
2695} /* xlog_regrant_write_log_space */
2696
2697
2698/* The first cnt-1 times through here we don't need to
2699 * move the grant write head because the permanent
2700 * reservation has reserved cnt times the unit amount.
2701 * Release part of current permanent unit reservation and
2702 * reset current reservation to be one units worth. Also
2703 * move grant reservation head forward.
2704 */
2705STATIC void
2706xlog_regrant_reserve_log_space(xlog_t *log,
2707 xlog_ticket_t *ticket)
2708{
2709 SPLDECL(s);
2710
2711 xlog_trace_loggrant(log, ticket,
2712 "xlog_regrant_reserve_log_space: enter");
2713 if (ticket->t_cnt > 0)
2714 ticket->t_cnt--;
2715
2716 s = GRANT_LOCK(log);
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002717 xlog_grant_sub_space(log, ticket->t_curr_res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 ticket->t_curr_res = ticket->t_unit_res;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10002719 XLOG_TIC_RESET_RES(ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 xlog_trace_loggrant(log, ticket,
2721 "xlog_regrant_reserve_log_space: sub current res");
2722 xlog_verify_grant_head(log, 1);
2723
2724 /* just return if we still have some of the pre-reserved space */
2725 if (ticket->t_cnt > 0) {
2726 GRANT_UNLOCK(log, s);
2727 return;
2728 }
2729
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002730 xlog_grant_add_space_reserve(log, ticket->t_unit_res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 xlog_trace_loggrant(log, ticket,
2732 "xlog_regrant_reserve_log_space: exit");
2733 xlog_verify_grant_head(log, 0);
2734 GRANT_UNLOCK(log, s);
2735 ticket->t_curr_res = ticket->t_unit_res;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10002736 XLOG_TIC_RESET_RES(ticket);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737} /* xlog_regrant_reserve_log_space */
2738
2739
2740/*
2741 * Give back the space left from a reservation.
2742 *
2743 * All the information we need to make a correct determination of space left
2744 * is present. For non-permanent reservations, things are quite easy. The
2745 * count should have been decremented to zero. We only need to deal with the
2746 * space remaining in the current reservation part of the ticket. If the
2747 * ticket contains a permanent reservation, there may be left over space which
2748 * needs to be released. A count of N means that N-1 refills of the current
2749 * reservation can be done before we need to ask for more space. The first
2750 * one goes to fill up the first current reservation. Once we run out of
2751 * space, the count will stay at zero and the only space remaining will be
2752 * in the current reservation field.
2753 */
2754STATIC void
2755xlog_ungrant_log_space(xlog_t *log,
2756 xlog_ticket_t *ticket)
2757{
2758 SPLDECL(s);
2759
2760 if (ticket->t_cnt > 0)
2761 ticket->t_cnt--;
2762
2763 s = GRANT_LOCK(log);
2764 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2765
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002766 xlog_grant_sub_space(log, ticket->t_curr_res);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767
2768 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2769
2770 /* If this is a permanent reservation ticket, we may be able to free
2771 * up more space based on the remaining count.
2772 */
2773 if (ticket->t_cnt > 0) {
2774 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
Christoph Hellwigdd954c62006-01-11 15:34:50 +11002775 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 }
2777
2778 xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2779 xlog_verify_grant_head(log, 1);
2780 GRANT_UNLOCK(log, s);
2781 xfs_log_move_tail(log->l_mp, 1);
2782} /* xlog_ungrant_log_space */
2783
2784
2785/*
2786 * Atomically put back used ticket.
2787 */
2788void
2789xlog_state_put_ticket(xlog_t *log,
2790 xlog_ticket_t *tic)
2791{
2792 unsigned long s;
2793
2794 s = LOG_LOCK(log);
2795 xlog_ticket_put(log, tic);
2796 LOG_UNLOCK(log, s);
2797} /* xlog_state_put_ticket */
2798
2799/*
2800 * Flush iclog to disk if this is the last reference to the given iclog and
2801 * the WANT_SYNC bit is set.
2802 *
2803 * When this function is entered, the iclog is not necessarily in the
2804 * WANT_SYNC state. It may be sitting around waiting to get filled.
2805 *
2806 *
2807 */
2808int
2809xlog_state_release_iclog(xlog_t *log,
2810 xlog_in_core_t *iclog)
2811{
2812 SPLDECL(s);
2813 int sync = 0; /* do we sync? */
2814
2815 xlog_assign_tail_lsn(log->l_mp);
2816
2817 s = LOG_LOCK(log);
2818
2819 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2820 LOG_UNLOCK(log, s);
2821 return XFS_ERROR(EIO);
2822 }
2823
2824 ASSERT(iclog->ic_refcnt > 0);
2825 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2826 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2827
2828 if (--iclog->ic_refcnt == 0 &&
2829 iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2830 sync++;
2831 iclog->ic_state = XLOG_STATE_SYNCING;
2832 INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2833 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2834 /* cycle incremented when incrementing curr_block */
2835 }
2836
2837 LOG_UNLOCK(log, s);
2838
2839 /*
2840 * We let the log lock go, so it's possible that we hit a log I/O
Nathan Scottc41564b2006-03-29 08:55:14 +10002841 * error or some other SHUTDOWN condition that marks the iclog
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2843 * this iclog has consistent data, so we ignore IOERROR
2844 * flags after this point.
2845 */
2846 if (sync) {
2847 return xlog_sync(log, iclog);
2848 }
Jesper Juhl014c2542006-01-15 02:37:08 +01002849 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850
2851} /* xlog_state_release_iclog */
2852
2853
2854/*
2855 * This routine will mark the current iclog in the ring as WANT_SYNC
2856 * and move the current iclog pointer to the next iclog in the ring.
2857 * When this routine is called from xlog_state_get_iclog_space(), the
2858 * exact size of the iclog has not yet been determined. All we know is
2859 * that every data block. We have run out of space in this log record.
2860 */
2861STATIC void
2862xlog_state_switch_iclogs(xlog_t *log,
2863 xlog_in_core_t *iclog,
2864 int eventual_size)
2865{
2866 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2867 if (!eventual_size)
2868 eventual_size = iclog->ic_offset;
2869 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2870 INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2871 log->l_prev_block = log->l_curr_block;
2872 log->l_prev_cycle = log->l_curr_cycle;
2873
2874 /* roll log?: ic_offset changed later */
2875 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2876
2877 /* Round up to next log-sunit */
2878 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2879 log->l_mp->m_sb.sb_logsunit > 1) {
2880 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2881 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2882 }
2883
2884 if (log->l_curr_block >= log->l_logBBsize) {
2885 log->l_curr_cycle++;
2886 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2887 log->l_curr_cycle++;
2888 log->l_curr_block -= log->l_logBBsize;
2889 ASSERT(log->l_curr_block >= 0);
2890 }
2891 ASSERT(iclog == log->l_iclog);
2892 log->l_iclog = iclog->ic_next;
2893} /* xlog_state_switch_iclogs */
2894
2895
2896/*
2897 * Write out all data in the in-core log as of this exact moment in time.
2898 *
2899 * Data may be written to the in-core log during this call. However,
2900 * we don't guarantee this data will be written out. A change from past
2901 * implementation means this routine will *not* write out zero length LRs.
2902 *
2903 * Basically, we try and perform an intelligent scan of the in-core logs.
2904 * If we determine there is no flushable data, we just return. There is no
2905 * flushable data if:
2906 *
2907 * 1. the current iclog is active and has no data; the previous iclog
2908 * is in the active or dirty state.
2909 * 2. the current iclog is drity, and the previous iclog is in the
2910 * active or dirty state.
2911 *
2912 * We may sleep (call psema) if:
2913 *
2914 * 1. the current iclog is not in the active nor dirty state.
2915 * 2. the current iclog dirty, and the previous iclog is not in the
2916 * active nor dirty state.
2917 * 3. the current iclog is active, and there is another thread writing
2918 * to this particular iclog.
2919 * 4. a) the current iclog is active and has no other writers
2920 * b) when we return from flushing out this iclog, it is still
2921 * not in the active nor dirty state.
2922 */
2923STATIC int
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11002924xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925{
2926 xlog_in_core_t *iclog;
2927 xfs_lsn_t lsn;
2928 SPLDECL(s);
2929
2930 s = LOG_LOCK(log);
2931
2932 iclog = log->l_iclog;
2933 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2934 LOG_UNLOCK(log, s);
2935 return XFS_ERROR(EIO);
2936 }
2937
2938 /* If the head iclog is not active nor dirty, we just attach
2939 * ourselves to the head and go to sleep.
2940 */
2941 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2942 iclog->ic_state == XLOG_STATE_DIRTY) {
2943 /*
2944 * If the head is dirty or (active and empty), then
2945 * we need to look at the previous iclog. If the previous
2946 * iclog is active or dirty we are done. There is nothing
2947 * to sync out. Otherwise, we attach ourselves to the
2948 * previous iclog and go to sleep.
2949 */
2950 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2951 (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2952 iclog = iclog->ic_prev;
2953 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2954 iclog->ic_state == XLOG_STATE_DIRTY)
2955 goto no_sleep;
2956 else
2957 goto maybe_sleep;
2958 } else {
2959 if (iclog->ic_refcnt == 0) {
2960 /* We are the only one with access to this
2961 * iclog. Flush it out now. There should
2962 * be a roundoff of zero to show that someone
2963 * has already taken care of the roundoff from
2964 * the previous sync.
2965 */
2966 iclog->ic_refcnt++;
2967 lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2968 xlog_state_switch_iclogs(log, iclog, 0);
2969 LOG_UNLOCK(log, s);
2970
2971 if (xlog_state_release_iclog(log, iclog))
2972 return XFS_ERROR(EIO);
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11002973 *log_flushed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 s = LOG_LOCK(log);
2975 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
2976 iclog->ic_state != XLOG_STATE_DIRTY)
2977 goto maybe_sleep;
2978 else
2979 goto no_sleep;
2980 } else {
2981 /* Someone else is writing to this iclog.
2982 * Use its call to flush out the data. However,
2983 * the other thread may not force out this LR,
2984 * so we mark it WANT_SYNC.
2985 */
2986 xlog_state_switch_iclogs(log, iclog, 0);
2987 goto maybe_sleep;
2988 }
2989 }
2990 }
2991
2992 /* By the time we come around again, the iclog could've been filled
2993 * which would give it another lsn. If we have a new lsn, just
2994 * return because the relevant data has been flushed.
2995 */
2996maybe_sleep:
2997 if (flags & XFS_LOG_SYNC) {
2998 /*
2999 * We must check if we're shutting down here, before
3000 * we wait, while we're holding the LOG_LOCK.
3001 * Then we check again after waking up, in case our
3002 * sleep was disturbed by a bad news.
3003 */
3004 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3005 LOG_UNLOCK(log, s);
3006 return XFS_ERROR(EIO);
3007 }
3008 XFS_STATS_INC(xs_log_force_sleep);
3009 sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
3010 /*
3011 * No need to grab the log lock here since we're
3012 * only deciding whether or not to return EIO
3013 * and the memory read should be atomic.
3014 */
3015 if (iclog->ic_state & XLOG_STATE_IOERROR)
3016 return XFS_ERROR(EIO);
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11003017 *log_flushed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018
3019 } else {
3020
3021no_sleep:
3022 LOG_UNLOCK(log, s);
3023 }
3024 return 0;
3025} /* xlog_state_sync_all */
3026
3027
3028/*
3029 * Used by code which implements synchronous log forces.
3030 *
3031 * Find in-core log with lsn.
3032 * If it is in the DIRTY state, just return.
3033 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3034 * state and go to sleep or return.
3035 * If it is in any other state, go to sleep or return.
3036 *
3037 * If filesystem activity goes to zero, the iclog will get flushed only by
3038 * bdflush().
3039 */
3040int
3041xlog_state_sync(xlog_t *log,
3042 xfs_lsn_t lsn,
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11003043 uint flags,
3044 int *log_flushed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045{
3046 xlog_in_core_t *iclog;
3047 int already_slept = 0;
3048 SPLDECL(s);
3049
3050
3051try_again:
3052 s = LOG_LOCK(log);
3053 iclog = log->l_iclog;
3054
3055 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3056 LOG_UNLOCK(log, s);
3057 return XFS_ERROR(EIO);
3058 }
3059
3060 do {
3061 if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
3062 iclog = iclog->ic_next;
3063 continue;
3064 }
3065
3066 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3067 LOG_UNLOCK(log, s);
3068 return 0;
3069 }
3070
3071 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3072 /*
3073 * We sleep here if we haven't already slept (e.g.
3074 * this is the first time we've looked at the correct
3075 * iclog buf) and the buffer before us is going to
3076 * be sync'ed. The reason for this is that if we
3077 * are doing sync transactions here, by waiting for
3078 * the previous I/O to complete, we can allow a few
3079 * more transactions into this iclog before we close
3080 * it down.
3081 *
3082 * Otherwise, we mark the buffer WANT_SYNC, and bump
3083 * up the refcnt so we can release the log (which drops
3084 * the ref count). The state switch keeps new transaction
3085 * commits from using this buffer. When the current commits
3086 * finish writing into the buffer, the refcount will drop to
3087 * zero and the buffer will go out then.
3088 */
3089 if (!already_slept &&
3090 (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3091 XLOG_STATE_SYNCING))) {
3092 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3093 XFS_STATS_INC(xs_log_force_sleep);
3094 sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3095 &log->l_icloglock, s);
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11003096 *log_flushed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003097 already_slept = 1;
3098 goto try_again;
3099 } else {
3100 iclog->ic_refcnt++;
3101 xlog_state_switch_iclogs(log, iclog, 0);
3102 LOG_UNLOCK(log, s);
3103 if (xlog_state_release_iclog(log, iclog))
3104 return XFS_ERROR(EIO);
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11003105 *log_flushed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 s = LOG_LOCK(log);
3107 }
3108 }
3109
3110 if ((flags & XFS_LOG_SYNC) && /* sleep */
3111 !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3112
3113 /*
3114 * Don't wait on the forcesema if we know that we've
3115 * gotten a log write error.
3116 */
3117 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3118 LOG_UNLOCK(log, s);
3119 return XFS_ERROR(EIO);
3120 }
3121 XFS_STATS_INC(xs_log_force_sleep);
3122 sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3123 /*
3124 * No need to grab the log lock here since we're
3125 * only deciding whether or not to return EIO
3126 * and the memory read should be atomic.
3127 */
3128 if (iclog->ic_state & XLOG_STATE_IOERROR)
3129 return XFS_ERROR(EIO);
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11003130 *log_flushed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 } else { /* just return */
3132 LOG_UNLOCK(log, s);
3133 }
3134 return 0;
3135
3136 } while (iclog != log->l_iclog);
3137
3138 LOG_UNLOCK(log, s);
Jesper Juhl014c2542006-01-15 02:37:08 +01003139 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140} /* xlog_state_sync */
3141
3142
3143/*
3144 * Called when we want to mark the current iclog as being ready to sync to
3145 * disk.
3146 */
3147void
3148xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3149{
3150 SPLDECL(s);
3151
3152 s = LOG_LOCK(log);
3153
3154 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3155 xlog_state_switch_iclogs(log, iclog, 0);
3156 } else {
3157 ASSERT(iclog->ic_state &
3158 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3159 }
3160
3161 LOG_UNLOCK(log, s);
3162} /* xlog_state_want_sync */
3163
3164
3165
3166/*****************************************************************************
3167 *
3168 * TICKET functions
3169 *
3170 *****************************************************************************
3171 */
3172
3173/*
3174 * Algorithm doesn't take into account page size. ;-(
3175 */
3176STATIC void
3177xlog_state_ticket_alloc(xlog_t *log)
3178{
3179 xlog_ticket_t *t_list;
3180 xlog_ticket_t *next;
3181 xfs_caddr_t buf;
3182 uint i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3183 SPLDECL(s);
3184
3185 /*
3186 * The kmem_zalloc may sleep, so we shouldn't be holding the
3187 * global lock. XXXmiken: may want to use zone allocator.
3188 */
3189 buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3190
3191 s = LOG_LOCK(log);
3192
3193 /* Attach 1st ticket to Q, so we can keep track of allocated memory */
3194 t_list = (xlog_ticket_t *)buf;
3195 t_list->t_next = log->l_unmount_free;
3196 log->l_unmount_free = t_list++;
3197 log->l_ticket_cnt++;
3198 log->l_ticket_tcnt++;
3199
3200 /* Next ticket becomes first ticket attached to ticket free list */
3201 if (log->l_freelist != NULL) {
3202 ASSERT(log->l_tail != NULL);
3203 log->l_tail->t_next = t_list;
3204 } else {
3205 log->l_freelist = t_list;
3206 }
3207 log->l_ticket_cnt++;
3208 log->l_ticket_tcnt++;
3209
3210 /* Cycle through rest of alloc'ed memory, building up free Q */
3211 for ( ; i > 0; i--) {
3212 next = t_list + 1;
3213 t_list->t_next = next;
3214 t_list = next;
3215 log->l_ticket_cnt++;
3216 log->l_ticket_tcnt++;
3217 }
3218 t_list->t_next = NULL;
3219 log->l_tail = t_list;
3220 LOG_UNLOCK(log, s);
3221} /* xlog_state_ticket_alloc */
3222
3223
3224/*
3225 * Put ticket into free list
3226 *
3227 * Assumption: log lock is held around this call.
3228 */
3229STATIC void
3230xlog_ticket_put(xlog_t *log,
3231 xlog_ticket_t *ticket)
3232{
3233 sv_destroy(&ticket->t_sema);
3234
3235 /*
3236 * Don't think caching will make that much difference. It's
3237 * more important to make debug easier.
3238 */
3239#if 0
3240 /* real code will want to use LIFO for caching */
3241 ticket->t_next = log->l_freelist;
3242 log->l_freelist = ticket;
3243 /* no need to clear fields */
3244#else
3245 /* When we debug, it is easier if tickets are cycled */
3246 ticket->t_next = NULL;
3247 if (log->l_tail != 0) {
3248 log->l_tail->t_next = ticket;
3249 } else {
3250 ASSERT(log->l_freelist == 0);
3251 log->l_freelist = ticket;
3252 }
3253 log->l_tail = ticket;
3254#endif /* DEBUG */
3255 log->l_ticket_cnt++;
3256} /* xlog_ticket_put */
3257
3258
3259/*
3260 * Grab ticket off freelist or allocation some more
3261 */
3262xlog_ticket_t *
3263xlog_ticket_get(xlog_t *log,
3264 int unit_bytes,
3265 int cnt,
3266 char client,
3267 uint xflags)
3268{
3269 xlog_ticket_t *tic;
3270 uint num_headers;
3271 SPLDECL(s);
3272
3273 alloc:
3274 if (log->l_freelist == NULL)
3275 xlog_state_ticket_alloc(log); /* potentially sleep */
3276
3277 s = LOG_LOCK(log);
3278 if (log->l_freelist == NULL) {
3279 LOG_UNLOCK(log, s);
3280 goto alloc;
3281 }
3282 tic = log->l_freelist;
3283 log->l_freelist = tic->t_next;
3284 if (log->l_freelist == NULL)
3285 log->l_tail = NULL;
3286 log->l_ticket_cnt--;
3287 LOG_UNLOCK(log, s);
3288
3289 /*
3290 * Permanent reservations have up to 'cnt'-1 active log operations
3291 * in the log. A unit in this case is the amount of space for one
3292 * of these log operations. Normal reservations have a cnt of 1
3293 * and their unit amount is the total amount of space required.
3294 *
3295 * The following lines of code account for non-transaction data
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003296 * which occupy space in the on-disk log.
3297 *
3298 * Normal form of a transaction is:
3299 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3300 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3301 *
3302 * We need to account for all the leadup data and trailer data
3303 * around the transaction data.
3304 * And then we need to account for the worst case in terms of using
3305 * more space.
3306 * The worst case will happen if:
3307 * - the placement of the transaction happens to be such that the
3308 * roundoff is at its maximum
3309 * - the transaction data is synced before the commit record is synced
3310 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3311 * Therefore the commit record is in its own Log Record.
3312 * This can happen as the commit record is called with its
3313 * own region to xlog_write().
3314 * This then means that in the worst case, roundoff can happen for
3315 * the commit-rec as well.
3316 * The commit-rec is smaller than padding in this scenario and so it is
3317 * not added separately.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 */
3319
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003320 /* for trans header */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003321 unit_bytes += sizeof(xlog_op_header_t);
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003322 unit_bytes += sizeof(xfs_trans_header_t);
3323
3324 /* for start-rec */
3325 unit_bytes += sizeof(xlog_op_header_t);
3326
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 /* for LR headers */
3328 num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3329 unit_bytes += log->l_iclog_hsize * num_headers;
3330
Tim Shimmin32fb9b52005-09-02 16:41:43 +10003331 /* for commit-rec LR header - note: padding will subsume the ophdr */
3332 unit_bytes += log->l_iclog_hsize;
3333
3334 /* for split-recs - ophdrs added when data split over LRs */
3335 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3336
3337 /* for roundoff padding for transaction data and one for commit record */
3338 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3339 log->l_mp->m_sb.sb_logsunit > 1) {
3340 /* log su roundoff */
3341 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3342 } else {
3343 /* BB roundoff */
3344 unit_bytes += 2*BBSIZE;
3345 }
3346
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 tic->t_unit_res = unit_bytes;
3348 tic->t_curr_res = unit_bytes;
3349 tic->t_cnt = cnt;
3350 tic->t_ocnt = cnt;
3351 tic->t_tid = (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3352 tic->t_clientid = client;
3353 tic->t_flags = XLOG_TIC_INITED;
Tim Shimmin7e9c6392005-09-02 16:42:05 +10003354 tic->t_trans_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355 if (xflags & XFS_LOG_PERM_RESERV)
3356 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3357 sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3358
Tim Shimmin7e9c6392005-09-02 16:42:05 +10003359 XLOG_TIC_RESET_RES(tic);
3360
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 return tic;
3362} /* xlog_ticket_get */
3363
3364
3365/******************************************************************************
3366 *
3367 * Log debug routines
3368 *
3369 ******************************************************************************
3370 */
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11003371#if defined(DEBUG)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372/*
3373 * Make sure that the destination ptr is within the valid data region of
3374 * one of the iclogs. This uses backup pointers stored in a different
3375 * part of the log in case we trash the log structure.
3376 */
3377void
3378xlog_verify_dest_ptr(xlog_t *log,
3379 __psint_t ptr)
3380{
3381 int i;
3382 int good_ptr = 0;
3383
3384 for (i=0; i < log->l_iclog_bufs; i++) {
3385 if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3386 ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3387 good_ptr++;
3388 }
3389 if (! good_ptr)
3390 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3391} /* xlog_verify_dest_ptr */
3392
3393STATIC void
3394xlog_verify_grant_head(xlog_t *log, int equals)
3395{
3396 if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3397 if (equals)
3398 ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3399 else
3400 ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3401 } else {
3402 ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3403 ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3404 }
3405} /* xlog_verify_grant_head */
3406
3407/* check if it will fit */
3408STATIC void
3409xlog_verify_tail_lsn(xlog_t *log,
3410 xlog_in_core_t *iclog,
3411 xfs_lsn_t tail_lsn)
3412{
3413 int blocks;
3414
3415 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3416 blocks =
3417 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3418 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3419 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3420 } else {
3421 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3422
3423 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3424 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3425
3426 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3427 if (blocks < BTOBB(iclog->ic_offset) + 1)
3428 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3429 }
3430} /* xlog_verify_tail_lsn */
3431
3432/*
3433 * Perform a number of checks on the iclog before writing to disk.
3434 *
3435 * 1. Make sure the iclogs are still circular
3436 * 2. Make sure we have a good magic number
3437 * 3. Make sure we don't have magic numbers in the data
3438 * 4. Check fields of each log operation header for:
3439 * A. Valid client identifier
3440 * B. tid ptr value falls in valid ptr space (user space code)
3441 * C. Length in log record header is correct according to the
3442 * individual operation headers within record.
3443 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3444 * log, check the preceding blocks of the physical log to make sure all
3445 * the cycle numbers agree with the current cycle number.
3446 */
3447STATIC void
3448xlog_verify_iclog(xlog_t *log,
3449 xlog_in_core_t *iclog,
3450 int count,
3451 boolean_t syncing)
3452{
3453 xlog_op_header_t *ophead;
3454 xlog_in_core_t *icptr;
3455 xlog_in_core_2_t *xhdr;
3456 xfs_caddr_t ptr;
3457 xfs_caddr_t base_ptr;
3458 __psint_t field_offset;
3459 __uint8_t clientid;
3460 int len, i, j, k, op_len;
3461 int idx;
3462 SPLDECL(s);
3463
3464 /* check validity of iclog pointers */
3465 s = LOG_LOCK(log);
3466 icptr = log->l_iclog;
3467 for (i=0; i < log->l_iclog_bufs; i++) {
3468 if (icptr == 0)
3469 xlog_panic("xlog_verify_iclog: invalid ptr");
3470 icptr = icptr->ic_next;
3471 }
3472 if (icptr != log->l_iclog)
3473 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3474 LOG_UNLOCK(log, s);
3475
3476 /* check log magic numbers */
3477 ptr = (xfs_caddr_t) &(iclog->ic_header);
3478 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3479 xlog_panic("xlog_verify_iclog: invalid magic num");
3480
3481 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3482 ptr += BBSIZE) {
3483 if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3484 xlog_panic("xlog_verify_iclog: unexpected magic num");
3485 }
3486
3487 /* check fields */
3488 len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3489 ptr = iclog->ic_datap;
3490 base_ptr = ptr;
3491 ophead = (xlog_op_header_t *)ptr;
3492 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3493 for (i = 0; i < len; i++) {
3494 ophead = (xlog_op_header_t *)ptr;
3495
3496 /* clientid is only 1 byte */
3497 field_offset = (__psint_t)
3498 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3499 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3500 clientid = ophead->oh_clientid;
3501 } else {
3502 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3503 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3504 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3505 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3506 clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3507 } else {
3508 clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3509 }
3510 }
3511 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
Christoph Hellwigda1650a2005-11-02 10:21:35 +11003512 cmn_err(CE_WARN, "xlog_verify_iclog: "
3513 "invalid clientid %d op 0x%p offset 0x%lx",
3514 clientid, ophead, (unsigned long)field_offset);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515
3516 /* check length */
3517 field_offset = (__psint_t)
3518 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3519 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3520 op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3521 } else {
3522 idx = BTOBBT((__psint_t)&ophead->oh_len -
3523 (__psint_t)iclog->ic_datap);
3524 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3525 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3526 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3527 op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3528 } else {
3529 op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3530 }
3531 }
3532 ptr += sizeof(xlog_op_header_t) + op_len;
3533 }
3534} /* xlog_verify_iclog */
Nathan Scottcfcbbbd2005-11-02 15:12:04 +11003535#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536
3537/*
3538 * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3539 */
3540STATIC int
3541xlog_state_ioerror(
3542 xlog_t *log)
3543{
3544 xlog_in_core_t *iclog, *ic;
3545
3546 iclog = log->l_iclog;
3547 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3548 /*
3549 * Mark all the incore logs IOERROR.
3550 * From now on, no log flushes will result.
3551 */
3552 ic = iclog;
3553 do {
3554 ic->ic_state = XLOG_STATE_IOERROR;
3555 ic = ic->ic_next;
3556 } while (ic != iclog);
Jesper Juhl014c2542006-01-15 02:37:08 +01003557 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558 }
3559 /*
3560 * Return non-zero, if state transition has already happened.
3561 */
Jesper Juhl014c2542006-01-15 02:37:08 +01003562 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563}
3564
3565/*
3566 * This is called from xfs_force_shutdown, when we're forcibly
3567 * shutting down the filesystem, typically because of an IO error.
3568 * Our main objectives here are to make sure that:
3569 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3570 * parties to find out, 'atomically'.
3571 * b. those who're sleeping on log reservations, pinned objects and
3572 * other resources get woken up, and be told the bad news.
3573 * c. nothing new gets queued up after (a) and (b) are done.
3574 * d. if !logerror, flush the iclogs to disk, then seal them off
3575 * for business.
3576 */
3577int
3578xfs_log_force_umount(
3579 struct xfs_mount *mp,
3580 int logerror)
3581{
3582 xlog_ticket_t *tic;
3583 xlog_t *log;
3584 int retval;
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11003585 int dummy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 SPLDECL(s);
3587 SPLDECL(s2);
3588
3589 log = mp->m_log;
3590
3591 /*
3592 * If this happens during log recovery, don't worry about
3593 * locking; the log isn't open for business yet.
3594 */
3595 if (!log ||
3596 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3597 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3598 XFS_BUF_DONE(mp->m_sb_bp);
Jesper Juhl014c2542006-01-15 02:37:08 +01003599 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 }
3601
3602 /*
3603 * Somebody could've already done the hard work for us.
3604 * No need to get locks for this.
3605 */
3606 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3607 ASSERT(XLOG_FORCED_SHUTDOWN(log));
Jesper Juhl014c2542006-01-15 02:37:08 +01003608 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609 }
3610 retval = 0;
3611 /*
3612 * We must hold both the GRANT lock and the LOG lock,
3613 * before we mark the filesystem SHUTDOWN and wake
3614 * everybody up to tell the bad news.
3615 */
3616 s = GRANT_LOCK(log);
3617 s2 = LOG_LOCK(log);
3618 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3619 XFS_BUF_DONE(mp->m_sb_bp);
3620 /*
3621 * This flag is sort of redundant because of the mount flag, but
3622 * it's good to maintain the separation between the log and the rest
3623 * of XFS.
3624 */
3625 log->l_flags |= XLOG_IO_ERROR;
3626
3627 /*
3628 * If we hit a log error, we want to mark all the iclogs IOERROR
3629 * while we're still holding the loglock.
3630 */
3631 if (logerror)
3632 retval = xlog_state_ioerror(log);
3633 LOG_UNLOCK(log, s2);
3634
3635 /*
3636 * We don't want anybody waiting for log reservations
3637 * after this. That means we have to wake up everybody
3638 * queued up on reserve_headq as well as write_headq.
3639 * In addition, we make sure in xlog_{re}grant_log_space
3640 * that we don't enqueue anything once the SHUTDOWN flag
3641 * is set, and this action is protected by the GRANTLOCK.
3642 */
3643 if ((tic = log->l_reserve_headq)) {
3644 do {
3645 sv_signal(&tic->t_sema);
3646 tic = tic->t_next;
3647 } while (tic != log->l_reserve_headq);
3648 }
3649
3650 if ((tic = log->l_write_headq)) {
3651 do {
3652 sv_signal(&tic->t_sema);
3653 tic = tic->t_next;
3654 } while (tic != log->l_write_headq);
3655 }
3656 GRANT_UNLOCK(log, s);
3657
3658 if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3659 ASSERT(!logerror);
3660 /*
3661 * Force the incore logs to disk before shutting the
3662 * log down completely.
3663 */
Christoph Hellwigf538d4d2005-11-02 10:26:59 +11003664 xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665 s2 = LOG_LOCK(log);
3666 retval = xlog_state_ioerror(log);
3667 LOG_UNLOCK(log, s2);
3668 }
3669 /*
3670 * Wake up everybody waiting on xfs_log_force.
3671 * Callback all log item committed functions as if the
3672 * log writes were completed.
3673 */
3674 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3675
3676#ifdef XFSERRORDEBUG
3677 {
3678 xlog_in_core_t *iclog;
3679
3680 s = LOG_LOCK(log);
3681 iclog = log->l_iclog;
3682 do {
3683 ASSERT(iclog->ic_callback == 0);
3684 iclog = iclog->ic_next;
3685 } while (iclog != log->l_iclog);
3686 LOG_UNLOCK(log, s);
3687 }
3688#endif
3689 /* return non-zero if log IOERROR transition had already happened */
Jesper Juhl014c2542006-01-15 02:37:08 +01003690 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691}
3692
Christoph Hellwigba0f32d2005-06-21 15:36:52 +10003693STATIC int
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694xlog_iclogs_empty(xlog_t *log)
3695{
3696 xlog_in_core_t *iclog;
3697
3698 iclog = log->l_iclog;
3699 do {
3700 /* endianness does not matter here, zero is zero in
3701 * any language.
3702 */
3703 if (iclog->ic_header.h_num_logops)
Jesper Juhl014c2542006-01-15 02:37:08 +01003704 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 iclog = iclog->ic_next;
3706 } while (iclog != log->l_iclog);
Jesper Juhl014c2542006-01-15 02:37:08 +01003707 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708}