Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
xfs_log_cil.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it would be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11  * GNU General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public License
14  * along with this program; if not, write the Free Software Foundation,
15  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
16  */
17 
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_log.h"
22 #include "xfs_trans.h"
23 #include "xfs_trans_priv.h"
24 #include "xfs_log_priv.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_alloc.h"
30 #include "xfs_extent_busy.h"
31 #include "xfs_discard.h"
32 
33 /*
34  * Allocate a new ticket. Failing to get a new ticket makes it really hard to
35  * recover, so we don't allow failure here. Also, we allocate in a context that
36  * we don't want to be issuing transactions from, so we need to tell the
37  * allocation code this as well.
38  *
39  * We don't reserve any space for the ticket - we are going to steal whatever
40  * space we require from transactions as they commit. To ensure we reserve all
41  * the space required, we need to set the current reservation of the ticket to
42  * zero so that we know to steal the initial transaction overhead from the
43  * first transaction commit.
44  */
45 static struct xlog_ticket *
46 xlog_cil_ticket_alloc(
47  struct xlog *log)
48 {
49  struct xlog_ticket *tic;
50 
51  tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
53  tic->t_trans_type = XFS_TRANS_CHECKPOINT;
54 
55  /*
56  * set the current reservation to zero so we know to steal the basic
57  * transaction overhead reservation from the first transaction commit.
58  */
59  tic->t_curr_res = 0;
60  return tic;
61 }
62 
63 /*
64  * After the first stage of log recovery is done, we know where the head and
65  * tail of the log are. We need this log initialisation done before we can
66  * initialise the first CIL checkpoint context.
67  *
68  * Here we allocate a log ticket to track space usage during a CIL push. This
69  * ticket is passed to xlog_write() directly so that we don't slowly leak log
70  * space by failing to account for space used by log headers and additional
71  * region headers for split regions.
72  */
73 void
75  struct xlog *log)
76 {
77  log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
78  log->l_cilp->xc_ctx->sequence = 1;
79  log->l_cilp->xc_ctx->commit_lsn = xlog_assign_lsn(log->l_curr_cycle,
80  log->l_curr_block);
81 }
82 
83 /*
84  * Format log item into a flat buffers
85  *
86  * For delayed logging, we need to hold a formatted buffer containing all the
87  * changes on the log item. This enables us to relog the item in memory and
88  * write it out asynchronously without needing to relock the object that was
89  * modified at the time it gets written into the iclog.
90  *
91  * This function builds a vector for the changes in each log item in the
92  * transaction. It then works out the length of the buffer needed for each log
93  * item, allocates them and formats the vector for the item into the buffer.
94  * The buffer is then attached to the log item are then inserted into the
95  * Committed Item List for tracking until the next checkpoint is written out.
96  *
97  * We don't set up region headers during this process; we simply copy the
98  * regions into the flat buffer. We can do this because we still have to do a
99  * formatting step to write the regions into the iclog buffer. Writing the
100  * ophdrs during the iclog write means that we can support splitting large
101  * regions across iclog boundares without needing a change in the format of the
102  * item/region encapsulation.
103  *
104  * Hence what we need to do now is change the rewrite the vector array to point
105  * to the copied region inside the buffer we just allocated. This allows us to
106  * format the regions into the iclog as though they are being formatted
107  * directly out of the objects themselves.
108  */
109 static struct xfs_log_vec *
110 xlog_cil_prepare_log_vecs(
111  struct xfs_trans *tp)
112 {
113  struct xfs_log_item_desc *lidp;
114  struct xfs_log_vec *lv = NULL;
115  struct xfs_log_vec *ret_lv = NULL;
116 
117 
118  /* Bail out if we didn't find a log item. */
119  if (list_empty(&tp->t_items)) {
120  ASSERT(0);
121  return NULL;
122  }
123 
124  list_for_each_entry(lidp, &tp->t_items, lid_trans) {
125  struct xfs_log_vec *new_lv;
126  void *ptr;
127  int index;
128  int len = 0;
129  uint niovecs;
130 
131  /* Skip items which aren't dirty in this transaction. */
132  if (!(lidp->lid_flags & XFS_LID_DIRTY))
133  continue;
134 
135  /* Skip items that do not have any vectors for writing */
136  niovecs = IOP_SIZE(lidp->lid_item);
137  if (!niovecs)
138  continue;
139 
140  new_lv = kmem_zalloc(sizeof(*new_lv) +
141  niovecs * sizeof(struct xfs_log_iovec),
142  KM_SLEEP);
143 
144  /* The allocated iovec region lies beyond the log vector. */
145  new_lv->lv_iovecp = (struct xfs_log_iovec *)&new_lv[1];
146  new_lv->lv_niovecs = niovecs;
147  new_lv->lv_item = lidp->lid_item;
148 
149  /* build the vector array and calculate it's length */
150  IOP_FORMAT(new_lv->lv_item, new_lv->lv_iovecp);
151  for (index = 0; index < new_lv->lv_niovecs; index++)
152  len += new_lv->lv_iovecp[index].i_len;
153 
154  new_lv->lv_buf_len = len;
155  new_lv->lv_buf = kmem_alloc(new_lv->lv_buf_len,
156  KM_SLEEP|KM_NOFS);
157  ptr = new_lv->lv_buf;
158 
159  for (index = 0; index < new_lv->lv_niovecs; index++) {
160  struct xfs_log_iovec *vec = &new_lv->lv_iovecp[index];
161 
162  memcpy(ptr, vec->i_addr, vec->i_len);
163  vec->i_addr = ptr;
164  ptr += vec->i_len;
165  }
166  ASSERT(ptr == new_lv->lv_buf + new_lv->lv_buf_len);
167 
168  if (!ret_lv)
169  ret_lv = new_lv;
170  else
171  lv->lv_next = new_lv;
172  lv = new_lv;
173  }
174 
175  return ret_lv;
176 }
177 
178 /*
179  * Prepare the log item for insertion into the CIL. Calculate the difference in
180  * log space and vectors it will consume, and if it is a new item pin it as
181  * well.
182  */
183 STATIC void
185  struct xlog *log,
186  struct xfs_log_vec *lv,
187  int *len,
188  int *diff_iovecs)
189 {
190  struct xfs_log_vec *old = lv->lv_item->li_lv;
191 
192  if (old) {
193  /* existing lv on log item, space used is a delta */
194  ASSERT(!list_empty(&lv->lv_item->li_cil));
195  ASSERT(old->lv_buf && old->lv_buf_len && old->lv_niovecs);
196 
197  *len += lv->lv_buf_len - old->lv_buf_len;
198  *diff_iovecs += lv->lv_niovecs - old->lv_niovecs;
199  kmem_free(old->lv_buf);
200  kmem_free(old);
201  } else {
202  /* new lv, must pin the log item */
203  ASSERT(!lv->lv_item->li_lv);
204  ASSERT(list_empty(&lv->lv_item->li_cil));
205 
206  *len += lv->lv_buf_len;
207  *diff_iovecs += lv->lv_niovecs;
208  IOP_PIN(lv->lv_item);
209 
210  }
211 
212  /* attach new log vector to log item */
213  lv->lv_item->li_lv = lv;
214 
215  /*
216  * If this is the first time the item is being committed to the
217  * CIL, store the sequence number on the log item so we can
218  * tell in future commits whether this is the first checkpoint
219  * the item is being committed into.
220  */
221  if (!lv->lv_item->li_seq)
222  lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
223 }
224 
225 /*
226  * Insert the log items into the CIL and calculate the difference in space
227  * consumed by the item. Add the space to the checkpoint ticket and calculate
228  * if the change requires additional log metadata. If it does, take that space
229  * as well. Remove the amount of space we added to the checkpoint ticket from
230  * the current transaction ticket so that the accounting works out correctly.
231  */
232 static void
233 xlog_cil_insert_items(
234  struct xlog *log,
235  struct xfs_log_vec *log_vector,
236  struct xlog_ticket *ticket)
237 {
238  struct xfs_cil *cil = log->l_cilp;
239  struct xfs_cil_ctx *ctx = cil->xc_ctx;
240  struct xfs_log_vec *lv;
241  int len = 0;
242  int diff_iovecs = 0;
243  int iclog_space;
244 
245  ASSERT(log_vector);
246 
247  /*
248  * Do all the accounting aggregation and switching of log vectors
249  * around in a separate loop to the insertion of items into the CIL.
250  * Then we can do a separate loop to update the CIL within a single
251  * lock/unlock pair. This reduces the number of round trips on the CIL
252  * lock from O(nr_logvectors) to O(1) and greatly reduces the overall
253  * hold time for the transaction commit.
254  *
255  * If this is the first time the item is being placed into the CIL in
256  * this context, pin it so it can't be written to disk until the CIL is
257  * flushed to the iclog and the iclog written to disk.
258  *
259  * We can do this safely because the context can't checkpoint until we
260  * are done so it doesn't matter exactly how we update the CIL.
261  */
262  for (lv = log_vector; lv; lv = lv->lv_next)
263  xfs_cil_prepare_item(log, lv, &len, &diff_iovecs);
264 
265  /* account for space used by new iovec headers */
266  len += diff_iovecs * sizeof(xlog_op_header_t);
267 
268  spin_lock(&cil->xc_cil_lock);
269 
270  /* move the items to the tail of the CIL */
271  for (lv = log_vector; lv; lv = lv->lv_next)
272  list_move_tail(&lv->lv_item->li_cil, &cil->xc_cil);
273 
274  ctx->nvecs += diff_iovecs;
275 
276  /*
277  * Now transfer enough transaction reservation to the context ticket
278  * for the checkpoint. The context ticket is special - the unit
279  * reservation has to grow as well as the current reservation as we
280  * steal from tickets so we can correctly determine the space used
281  * during the transaction commit.
282  */
283  if (ctx->ticket->t_curr_res == 0) {
284  /* first commit in checkpoint, steal the header reservation */
285  ASSERT(ticket->t_curr_res >= ctx->ticket->t_unit_res + len);
286  ctx->ticket->t_curr_res = ctx->ticket->t_unit_res;
287  ticket->t_curr_res -= ctx->ticket->t_unit_res;
288  }
289 
290  /* do we need space for more log record headers? */
291  iclog_space = log->l_iclog_size - log->l_iclog_hsize;
292  if (len > 0 && (ctx->space_used / iclog_space !=
293  (ctx->space_used + len) / iclog_space)) {
294  int hdrs;
295 
296  hdrs = (len + iclog_space - 1) / iclog_space;
297  /* need to take into account split region headers, too */
298  hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
299  ctx->ticket->t_unit_res += hdrs;
300  ctx->ticket->t_curr_res += hdrs;
301  ticket->t_curr_res -= hdrs;
302  ASSERT(ticket->t_curr_res >= len);
303  }
304  ticket->t_curr_res -= len;
305  ctx->space_used += len;
306 
307  spin_unlock(&cil->xc_cil_lock);
308 }
309 
310 static void
311 xlog_cil_free_logvec(
312  struct xfs_log_vec *log_vector)
313 {
314  struct xfs_log_vec *lv;
315 
316  for (lv = log_vector; lv; ) {
317  struct xfs_log_vec *next = lv->lv_next;
318  kmem_free(lv->lv_buf);
319  kmem_free(lv);
320  lv = next;
321  }
322 }
323 
324 /*
325  * Mark all items committed and clear busy extents. We free the log vector
326  * chains in a separate pass so that we unpin the log items as quickly as
327  * possible.
328  */
329 static void
330 xlog_cil_committed(
331  void *args,
332  int abort)
333 {
334  struct xfs_cil_ctx *ctx = args;
335  struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
336 
337  xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
338  ctx->start_lsn, abort);
339 
340  xfs_extent_busy_sort(&ctx->busy_extents);
341  xfs_extent_busy_clear(mp, &ctx->busy_extents,
342  (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
343 
344  spin_lock(&ctx->cil->xc_cil_lock);
345  list_del(&ctx->committing);
346  spin_unlock(&ctx->cil->xc_cil_lock);
347 
348  xlog_cil_free_logvec(ctx->lv_chain);
349 
350  if (!list_empty(&ctx->busy_extents)) {
351  ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
352 
353  xfs_discard_extents(mp, &ctx->busy_extents);
354  xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
355  }
356 
357  kmem_free(ctx);
358 }
359 
360 /*
361  * Push the Committed Item List to the log. If @push_seq flag is zero, then it
362  * is a background flush and so we can chose to ignore it. Otherwise, if the
363  * current sequence is the same as @push_seq we need to do a flush. If
364  * @push_seq is less than the current sequence, then it has already been
365  * flushed and we don't need to do anything - the caller will wait for it to
366  * complete if necessary.
367  *
368  * @push_seq is a value rather than a flag because that allows us to do an
369  * unlocked check of the sequence number for a match. Hence we can allows log
370  * forces to run racily and not issue pushes for the same sequence twice. If we
371  * get a race between multiple pushes for the same sequence they will block on
372  * the first one and then abort, hence avoiding needless pushes.
373  */
374 STATIC int
376  struct xlog *log)
377 {
378  struct xfs_cil *cil = log->l_cilp;
379  struct xfs_log_vec *lv;
380  struct xfs_cil_ctx *ctx;
381  struct xfs_cil_ctx *new_ctx;
382  struct xlog_in_core *commit_iclog;
383  struct xlog_ticket *tic;
384  int num_lv;
385  int num_iovecs;
386  int len;
387  int error = 0;
388  struct xfs_trans_header thdr;
389  struct xfs_log_iovec lhdr;
390  struct xfs_log_vec lvhdr = { NULL };
391  xfs_lsn_t commit_lsn;
392  xfs_lsn_t push_seq;
393 
394  if (!cil)
395  return 0;
396 
397  new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
398  new_ctx->ticket = xlog_cil_ticket_alloc(log);
399 
400  down_write(&cil->xc_ctx_lock);
401  ctx = cil->xc_ctx;
402 
403  spin_lock(&cil->xc_cil_lock);
404  push_seq = cil->xc_push_seq;
405  ASSERT(push_seq <= ctx->sequence);
406 
407  /*
408  * Check if we've anything to push. If there is nothing, then we don't
409  * move on to a new sequence number and so we have to be able to push
410  * this sequence again later.
411  */
412  if (list_empty(&cil->xc_cil)) {
413  cil->xc_push_seq = 0;
414  spin_unlock(&cil->xc_cil_lock);
415  goto out_skip;
416  }
417  spin_unlock(&cil->xc_cil_lock);
418 
419 
420  /* check for a previously pushed seqeunce */
421  if (push_seq < cil->xc_ctx->sequence)
422  goto out_skip;
423 
424  /*
425  * pull all the log vectors off the items in the CIL, and
426  * remove the items from the CIL. We don't need the CIL lock
427  * here because it's only needed on the transaction commit
428  * side which is currently locked out by the flush lock.
429  */
430  lv = NULL;
431  num_lv = 0;
432  num_iovecs = 0;
433  len = 0;
434  while (!list_empty(&cil->xc_cil)) {
435  struct xfs_log_item *item;
436  int i;
437 
438  item = list_first_entry(&cil->xc_cil,
439  struct xfs_log_item, li_cil);
440  list_del_init(&item->li_cil);
441  if (!ctx->lv_chain)
442  ctx->lv_chain = item->li_lv;
443  else
444  lv->lv_next = item->li_lv;
445  lv = item->li_lv;
446  item->li_lv = NULL;
447 
448  num_lv++;
449  num_iovecs += lv->lv_niovecs;
450  for (i = 0; i < lv->lv_niovecs; i++)
451  len += lv->lv_iovecp[i].i_len;
452  }
453 
454  /*
455  * initialise the new context and attach it to the CIL. Then attach
456  * the current context to the CIL committing lsit so it can be found
457  * during log forces to extract the commit lsn of the sequence that
458  * needs to be forced.
459  */
460  INIT_LIST_HEAD(&new_ctx->committing);
461  INIT_LIST_HEAD(&new_ctx->busy_extents);
462  new_ctx->sequence = ctx->sequence + 1;
463  new_ctx->cil = cil;
464  cil->xc_ctx = new_ctx;
465 
466  /*
467  * mirror the new sequence into the cil structure so that we can do
468  * unlocked checks against the current sequence in log forces without
469  * risking deferencing a freed context pointer.
470  */
471  cil->xc_current_sequence = new_ctx->sequence;
472 
473  /*
474  * The switch is now done, so we can drop the context lock and move out
475  * of a shared context. We can't just go straight to the commit record,
476  * though - we need to synchronise with previous and future commits so
477  * that the commit records are correctly ordered in the log to ensure
478  * that we process items during log IO completion in the correct order.
479  *
480  * For example, if we get an EFI in one checkpoint and the EFD in the
481  * next (e.g. due to log forces), we do not want the checkpoint with
482  * the EFD to be committed before the checkpoint with the EFI. Hence
483  * we must strictly order the commit records of the checkpoints so
484  * that: a) the checkpoint callbacks are attached to the iclogs in the
485  * correct order; and b) the checkpoints are replayed in correct order
486  * in log recovery.
487  *
488  * Hence we need to add this context to the committing context list so
489  * that higher sequences will wait for us to write out a commit record
490  * before they do.
491  */
492  spin_lock(&cil->xc_cil_lock);
493  list_add(&ctx->committing, &cil->xc_committing);
494  spin_unlock(&cil->xc_cil_lock);
495  up_write(&cil->xc_ctx_lock);
496 
497  /*
498  * Build a checkpoint transaction header and write it to the log to
499  * begin the transaction. We need to account for the space used by the
500  * transaction header here as it is not accounted for in xlog_write().
501  *
502  * The LSN we need to pass to the log items on transaction commit is
503  * the LSN reported by the first log vector write. If we use the commit
504  * record lsn then we can move the tail beyond the grant write head.
505  */
506  tic = ctx->ticket;
509  thdr.th_tid = tic->t_tid;
510  thdr.th_num_items = num_iovecs;
511  lhdr.i_addr = &thdr;
512  lhdr.i_len = sizeof(xfs_trans_header_t);
514  tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
515 
516  lvhdr.lv_niovecs = 1;
517  lvhdr.lv_iovecp = &lhdr;
518  lvhdr.lv_next = ctx->lv_chain;
519 
520  error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
521  if (error)
522  goto out_abort_free_ticket;
523 
524  /*
525  * now that we've written the checkpoint into the log, strictly
526  * order the commit records so replay will get them in the right order.
527  */
528 restart:
529  spin_lock(&cil->xc_cil_lock);
530  list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
531  /*
532  * Higher sequences will wait for this one so skip them.
533  * Don't wait for own own sequence, either.
534  */
535  if (new_ctx->sequence >= ctx->sequence)
536  continue;
537  if (!new_ctx->commit_lsn) {
538  /*
539  * It is still being pushed! Wait for the push to
540  * complete, then start again from the beginning.
541  */
542  xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock);
543  goto restart;
544  }
545  }
546  spin_unlock(&cil->xc_cil_lock);
547 
548  /* xfs_log_done always frees the ticket on error. */
549  commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, 0);
550  if (commit_lsn == -1)
551  goto out_abort;
552 
553  /* attach all the transactions w/ busy extents to iclog */
554  ctx->log_cb.cb_func = xlog_cil_committed;
555  ctx->log_cb.cb_arg = ctx;
556  error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
557  if (error)
558  goto out_abort;
559 
560  /*
561  * now the checkpoint commit is complete and we've attached the
562  * callbacks to the iclog we can assign the commit LSN to the context
563  * and wake up anyone who is waiting for the commit to complete.
564  */
565  spin_lock(&cil->xc_cil_lock);
566  ctx->commit_lsn = commit_lsn;
567  wake_up_all(&cil->xc_commit_wait);
568  spin_unlock(&cil->xc_cil_lock);
569 
570  /* release the hounds! */
571  return xfs_log_release_iclog(log->l_mp, commit_iclog);
572 
573 out_skip:
574  up_write(&cil->xc_ctx_lock);
575  xfs_log_ticket_put(new_ctx->ticket);
576  kmem_free(new_ctx);
577  return 0;
578 
579 out_abort_free_ticket:
580  xfs_log_ticket_put(tic);
581 out_abort:
582  xlog_cil_committed(ctx, XFS_LI_ABORTED);
583  return XFS_ERROR(EIO);
584 }
585 
586 static void
587 xlog_cil_push_work(
588  struct work_struct *work)
589 {
590  struct xfs_cil *cil = container_of(work, struct xfs_cil,
591  xc_push_work);
592  xlog_cil_push(cil->xc_log);
593 }
594 
595 /*
596  * We need to push CIL every so often so we don't cache more than we can fit in
597  * the log. The limit really is that a checkpoint can't be more than half the
598  * log (the current checkpoint is not allowed to overwrite the previous
599  * checkpoint), but commit latency and memory usage limit this to a smaller
600  * size.
601  */
602 static void
603 xlog_cil_push_background(
604  struct xlog *log)
605 {
606  struct xfs_cil *cil = log->l_cilp;
607 
608  /*
609  * The cil won't be empty because we are called while holding the
610  * context lock so whatever we added to the CIL will still be there
611  */
612  ASSERT(!list_empty(&cil->xc_cil));
613 
614  /*
615  * don't do a background push if we haven't used up all the
616  * space available yet.
617  */
618  if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
619  return;
620 
621  spin_lock(&cil->xc_cil_lock);
622  if (cil->xc_push_seq < cil->xc_current_sequence) {
623  cil->xc_push_seq = cil->xc_current_sequence;
624  queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
625  }
626  spin_unlock(&cil->xc_cil_lock);
627 
628 }
629 
630 static void
631 xlog_cil_push_foreground(
632  struct xlog *log,
633  xfs_lsn_t push_seq)
634 {
635  struct xfs_cil *cil = log->l_cilp;
636 
637  if (!cil)
638  return;
639 
640  ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
641 
642  /* start on any pending background push to minimise wait time on it */
643  flush_work(&cil->xc_push_work);
644 
645  /*
646  * If the CIL is empty or we've already pushed the sequence then
647  * there's no work we need to do.
648  */
649  spin_lock(&cil->xc_cil_lock);
650  if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
651  spin_unlock(&cil->xc_cil_lock);
652  return;
653  }
654 
655  cil->xc_push_seq = push_seq;
656  spin_unlock(&cil->xc_cil_lock);
657 
658  /* do the push now */
659  xlog_cil_push(log);
660 }
661 
662 /*
663  * Commit a transaction with the given vector to the Committed Item List.
664  *
665  * To do this, we need to format the item, pin it in memory if required and
666  * account for the space used by the transaction. Once we have done that we
667  * need to release the unused reservation for the transaction, attach the
668  * transaction to the checkpoint context so we carry the busy extents through
669  * to checkpoint completion, and then unlock all the items in the transaction.
670  *
671  * For more specific information about the order of operations in
672  * xfs_log_commit_cil() please refer to the comments in
673  * xfs_trans_commit_iclog().
674  *
675  * Called with the context lock already held in read mode to lock out
676  * background commit, returns without it held once background commits are
677  * allowed again.
678  */
679 int
681  struct xfs_mount *mp,
682  struct xfs_trans *tp,
683  xfs_lsn_t *commit_lsn,
684  int flags)
685 {
686  struct xlog *log = mp->m_log;
687  int log_flags = 0;
688  struct xfs_log_vec *log_vector;
689 
690  if (flags & XFS_TRANS_RELEASE_LOG_RES)
691  log_flags = XFS_LOG_REL_PERM_RESERV;
692 
693  /*
694  * Do all the hard work of formatting items (including memory
695  * allocation) outside the CIL context lock. This prevents stalling CIL
696  * pushes when we are low on memory and a transaction commit spends a
697  * lot of time in memory reclaim.
698  */
699  log_vector = xlog_cil_prepare_log_vecs(tp);
700  if (!log_vector)
701  return ENOMEM;
702 
703  /* lock out background commit */
704  down_read(&log->l_cilp->xc_ctx_lock);
705  if (commit_lsn)
706  *commit_lsn = log->l_cilp->xc_ctx->sequence;
707 
708  xlog_cil_insert_items(log, log_vector, tp->t_ticket);
709 
710  /* check we didn't blow the reservation */
711  if (tp->t_ticket->t_curr_res < 0)
712  xlog_print_tic_res(log->l_mp, tp->t_ticket);
713 
714  /* attach the transaction to the CIL if it has any busy extents */
715  if (!list_empty(&tp->t_busy)) {
716  spin_lock(&log->l_cilp->xc_cil_lock);
717  list_splice_init(&tp->t_busy,
718  &log->l_cilp->xc_ctx->busy_extents);
719  spin_unlock(&log->l_cilp->xc_cil_lock);
720  }
721 
722  tp->t_commit_lsn = *commit_lsn;
723  xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
725 
726  /*
727  * Once all the items of the transaction have been copied to the CIL,
728  * the items can be unlocked and freed.
729  *
730  * This needs to be done before we drop the CIL context lock because we
731  * have to update state in the log items and unlock them before they go
732  * to disk. If we don't, then the CIL checkpoint can race with us and
733  * we can run checkpoint completion before we've updated and unlocked
734  * the log items. This affects (at least) processing of stale buffers,
735  * inodes and EFIs.
736  */
737  xfs_trans_free_items(tp, *commit_lsn, 0);
738 
739  xlog_cil_push_background(log);
740 
741  up_read(&log->l_cilp->xc_ctx_lock);
742  return 0;
743 }
744 
745 /*
746  * Conditionally push the CIL based on the sequence passed in.
747  *
748  * We only need to push if we haven't already pushed the sequence
749  * number given. Hence the only time we will trigger a push here is
750  * if the push sequence is the same as the current context.
751  *
752  * We return the current commit lsn to allow the callers to determine if a
753  * iclog flush is necessary following this call.
754  */
755 xfs_lsn_t
757  struct xlog *log,
759 {
760  struct xfs_cil *cil = log->l_cilp;
761  struct xfs_cil_ctx *ctx;
762  xfs_lsn_t commit_lsn = NULLCOMMITLSN;
763 
764  ASSERT(sequence <= cil->xc_current_sequence);
765 
766  /*
767  * check to see if we need to force out the current context.
768  * xlog_cil_push() handles racing pushes for the same sequence,
769  * so no need to deal with it here.
770  */
771  xlog_cil_push_foreground(log, sequence);
772 
773  /*
774  * See if we can find a previous sequence still committing.
775  * We need to wait for all previous sequence commits to complete
776  * before allowing the force of push_seq to go ahead. Hence block
777  * on commits for those as well.
778  */
779 restart:
780  spin_lock(&cil->xc_cil_lock);
781  list_for_each_entry(ctx, &cil->xc_committing, committing) {
782  if (ctx->sequence > sequence)
783  continue;
784  if (!ctx->commit_lsn) {
785  /*
786  * It is still being pushed! Wait for the push to
787  * complete, then start again from the beginning.
788  */
789  xlog_wait(&cil->xc_commit_wait, &cil->xc_cil_lock);
790  goto restart;
791  }
792  if (ctx->sequence != sequence)
793  continue;
794  /* found it! */
795  commit_lsn = ctx->commit_lsn;
796  }
797  spin_unlock(&cil->xc_cil_lock);
798  return commit_lsn;
799 }
800 
801 /*
802  * Check if the current log item was first committed in this sequence.
803  * We can't rely on just the log item being in the CIL, we have to check
804  * the recorded commit sequence number.
805  *
806  * Note: for this to be used in a non-racy manner, it has to be called with
807  * CIL flushing locked out. As a result, it should only be used during the
808  * transaction commit process when deciding what to format into the item.
809  */
810 bool
812  struct xfs_log_item *lip)
813 {
814  struct xfs_cil_ctx *ctx;
815 
816  if (list_empty(&lip->li_cil))
817  return false;
818 
819  ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
820 
821  /*
822  * li_seq is written on the first commit of a log item to record the
823  * first checkpoint it is written to. Hence if it is different to the
824  * current sequence, we're in a new checkpoint.
825  */
826  if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
827  return false;
828  return true;
829 }
830 
831 /*
832  * Perform initial CIL structure initialisation.
833  */
834 int
836  struct xlog *log)
837 {
838  struct xfs_cil *cil;
839  struct xfs_cil_ctx *ctx;
840 
841  cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
842  if (!cil)
843  return ENOMEM;
844 
845  ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
846  if (!ctx) {
847  kmem_free(cil);
848  return ENOMEM;
849  }
850 
851  INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
852  INIT_LIST_HEAD(&cil->xc_cil);
853  INIT_LIST_HEAD(&cil->xc_committing);
854  spin_lock_init(&cil->xc_cil_lock);
855  init_rwsem(&cil->xc_ctx_lock);
856  init_waitqueue_head(&cil->xc_commit_wait);
857 
858  INIT_LIST_HEAD(&ctx->committing);
859  INIT_LIST_HEAD(&ctx->busy_extents);
860  ctx->sequence = 1;
861  ctx->cil = cil;
862  cil->xc_ctx = ctx;
863  cil->xc_current_sequence = ctx->sequence;
864 
865  cil->xc_log = log;
866  log->l_cilp = cil;
867  return 0;
868 }
869 
870 void
872  struct xlog *log)
873 {
874  if (log->l_cilp->xc_ctx) {
875  if (log->l_cilp->xc_ctx->ticket)
876  xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
877  kmem_free(log->l_cilp->xc_ctx);
878  }
879 
880  ASSERT(list_empty(&log->l_cilp->xc_cil));
881  kmem_free(log->l_cilp);
882 }
883