1 // SPDX-License-Identifier: GPL-2.0+
3 * Copyright (C) 2016 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_defer.h"
14 #include "xfs_trans.h"
15 #include "xfs_buf_item.h"
16 #include "xfs_inode.h"
17 #include "xfs_inode_item.h"
18 #include "xfs_trace.h"
21 * Deferred Operations in XFS
23 * Due to the way locking rules work in XFS, certain transactions (block
24 * mapping and unmapping, typically) have permanent reservations so that
25 * we can roll the transaction to adhere to AG locking order rules and
26 * to unlock buffers between metadata updates. Prior to rmap/reflink,
27 * the mapping code had a mechanism to perform these deferrals for
28 * extents that were going to be freed; this code makes that facility
31 * When adding the reverse mapping and reflink features, it became
32 * necessary to perform complex remapping multi-transactions to comply
33 * with AG locking order rules, and to be able to spread a single
34 * refcount update operation (an operation on an n-block extent can
35 * update as many as n records!) among multiple transactions. XFS can
36 * roll a transaction to facilitate this, but using this facility
37 * requires us to log "intent" items in case log recovery needs to
38 * redo the operation, and to log "done" items to indicate that redo
41 * Deferred work is tracked in xfs_defer_pending items. Each pending
42 * item tracks one type of deferred work. Incoming work items (which
43 * have not yet had an intent logged) are attached to a pending item
44 * on the dop_intake list, where they wait for the caller to finish
45 * the deferred operations.
47 * Finishing a set of deferred operations is an involved process. To
48 * start, we define "rolling a deferred-op transaction" as follows:
50 * > For each xfs_defer_pending item on the dop_intake list,
51 * - Sort the work items in AG order. XFS locking
52 * order rules require us to lock buffers in AG order.
53 * - Create a log intent item for that type.
54 * - Attach it to the pending item.
55 * - Move the pending item from the dop_intake list to the
57 * > Roll the transaction.
59 * NOTE: To avoid exceeding the transaction reservation, we limit the
60 * number of items that we attach to a given xfs_defer_pending.
62 * The actual finishing process looks like this:
64 * > For each xfs_defer_pending in the dop_pending list,
65 * - Roll the deferred-op transaction as above.
66 * - Create a log done item for that type, and attach it to the
68 * - For each work item attached to the log intent item,
69 * * Perform the described action.
70 * * Attach the work item to the log done item.
71 * * If the result of doing the work was -EAGAIN, ->finish work
72 * wants a new transaction. See the "Requesting a Fresh
73 * Transaction while Finishing Deferred Work" section below for
76 * The key here is that we must log an intent item for all pending
77 * work items every time we roll the transaction, and that we must log
78 * a done item as soon as the work is completed. With this mechanism
79 * we can perform complex remapping operations, chaining intent items
82 * Requesting a Fresh Transaction while Finishing Deferred Work
84 * If ->finish_item decides that it needs a fresh transaction to
85 * finish the work, it must ask its caller (xfs_defer_finish) for a
86 * continuation. The most likely cause of this circumstance are the
87 * refcount adjust functions deciding that they've logged enough items
88 * to be at risk of exceeding the transaction reservation.
90 * To get a fresh transaction, we want to log the existing log done
91 * item to prevent the log intent item from replaying, immediately log
92 * a new log intent item with the unfinished work items, roll the
93 * transaction, and re-call ->finish_item wherever it left off. The
94 * log done item and the new log intent item must be in the same
95 * transaction or atomicity cannot be guaranteed; defer_finish ensures
98 * This requires some coordination between ->finish_item and
99 * defer_finish. Upon deciding to request a new transaction,
100 * ->finish_item should update the current work item to reflect the
101 * unfinished work. Next, it should reset the log done item's list
102 * count to the number of items finished, and return -EAGAIN.
103 * defer_finish sees the -EAGAIN, logs the new log intent item
104 * with the remaining work items, and leaves the xfs_defer_pending
105 * item at the head of the dop_work queue. Then it rolls the
106 * transaction and picks up processing where it left off. It is
107 * required that ->finish_item must be careful to leave enough
108 * transaction reservation to fit the new log intent item.
110 * This is an example of remapping the extent (E, E+B) into file X at
111 * offset A and dealing with the extent (C, C+B) already being mapped
113 * +-------------------------------------------------+
114 * | Unmap file X startblock C offset A length B | t0
115 * | Intent to reduce refcount for extent (C, B) |
116 * | Intent to remove rmap (X, C, A, B) |
117 * | Intent to free extent (D, 1) (bmbt block) |
118 * | Intent to map (X, A, B) at startblock E |
119 * +-------------------------------------------------+
120 * | Map file X startblock E offset A length B | t1
121 * | Done mapping (X, E, A, B) |
122 * | Intent to increase refcount for extent (E, B) |
123 * | Intent to add rmap (X, E, A, B) |
124 * +-------------------------------------------------+
125 * | Reduce refcount for extent (C, B) | t2
126 * | Done reducing refcount for extent (C, 9) |
127 * | Intent to reduce refcount for extent (C+9, B-9) |
128 * | (ran out of space after 9 refcount updates) |
129 * +-------------------------------------------------+
130 * | Reduce refcount for extent (C+9, B+9) | t3
131 * | Done reducing refcount for extent (C+9, B-9) |
132 * | Increase refcount for extent (E, B) |
133 * | Done increasing refcount for extent (E, B) |
134 * | Intent to free extent (C, B) |
135 * | Intent to free extent (F, 1) (refcountbt block) |
136 * | Intent to remove rmap (F, 1, REFC) |
137 * +-------------------------------------------------+
138 * | Remove rmap (X, C, A, B) | t4
139 * | Done removing rmap (X, C, A, B) |
140 * | Add rmap (X, E, A, B) |
141 * | Done adding rmap (X, E, A, B) |
142 * | Remove rmap (F, 1, REFC) |
143 * | Done removing rmap (F, 1, REFC) |
144 * +-------------------------------------------------+
145 * | Free extent (C, B) | t5
146 * | Done freeing extent (C, B) |
147 * | Free extent (D, 1) |
148 * | Done freeing extent (D, 1) |
149 * | Free extent (F, 1) |
150 * | Done freeing extent (F, 1) |
151 * +-------------------------------------------------+
153 * If we should crash before t2 commits, log recovery replays
154 * the following intent items:
156 * - Intent to reduce refcount for extent (C, B)
157 * - Intent to remove rmap (X, C, A, B)
158 * - Intent to free extent (D, 1) (bmbt block)
159 * - Intent to increase refcount for extent (E, B)
160 * - Intent to add rmap (X, E, A, B)
162 * In the process of recovering, it should also generate and take care
163 * of these intent items:
165 * - Intent to free extent (C, B)
166 * - Intent to free extent (F, 1) (refcountbt block)
167 * - Intent to remove rmap (F, 1, REFC)
169 * Note that the continuation requested between t2 and t3 is likely to
173 static const struct xfs_defer_op_type *defer_op_types[] = {
174 [XFS_DEFER_OPS_TYPE_BMAP] = &xfs_bmap_update_defer_type,
175 [XFS_DEFER_OPS_TYPE_REFCOUNT] = &xfs_refcount_update_defer_type,
176 [XFS_DEFER_OPS_TYPE_RMAP] = &xfs_rmap_update_defer_type,
177 [XFS_DEFER_OPS_TYPE_FREE] = &xfs_extent_free_defer_type,
178 [XFS_DEFER_OPS_TYPE_AGFL_FREE] = &xfs_agfl_free_defer_type,
182 xfs_defer_create_intent(
183 struct xfs_trans *tp,
184 struct xfs_defer_pending *dfp,
187 const struct xfs_defer_op_type *ops = defer_op_types[dfp->dfp_type];
189 dfp->dfp_intent = ops->create_intent(tp, &dfp->dfp_work,
190 dfp->dfp_count, sort);
194 * For each pending item in the intake list, log its intent item and the
195 * associated extents, then add the entire intake list to the end of
199 xfs_defer_create_intents(
200 struct xfs_trans *tp)
202 struct xfs_defer_pending *dfp;
204 list_for_each_entry(dfp, &tp->t_dfops, dfp_list) {
205 trace_xfs_defer_create_intent(tp->t_mountp, dfp);
206 xfs_defer_create_intent(tp, dfp, true);
210 /* Abort all the intents that were committed. */
212 xfs_defer_trans_abort(
213 struct xfs_trans *tp,
214 struct list_head *dop_pending)
216 struct xfs_defer_pending *dfp;
217 const struct xfs_defer_op_type *ops;
219 trace_xfs_defer_trans_abort(tp, _RET_IP_);
221 /* Abort intent items that don't have a done item. */
222 list_for_each_entry(dfp, dop_pending, dfp_list) {
223 ops = defer_op_types[dfp->dfp_type];
224 trace_xfs_defer_pending_abort(tp->t_mountp, dfp);
225 if (dfp->dfp_intent && !dfp->dfp_done) {
226 ops->abort_intent(dfp->dfp_intent);
227 dfp->dfp_intent = NULL;
232 /* Roll a transaction so we can do some deferred op processing. */
234 xfs_defer_trans_roll(
235 struct xfs_trans **tpp)
237 struct xfs_trans *tp = *tpp;
238 struct xfs_buf_log_item *bli;
239 struct xfs_inode_log_item *ili;
240 struct xfs_log_item *lip;
241 struct xfs_buf *bplist[XFS_DEFER_OPS_NR_BUFS];
242 struct xfs_inode *iplist[XFS_DEFER_OPS_NR_INODES];
243 unsigned int ordered = 0; /* bitmap */
244 int bpcount = 0, ipcount = 0;
248 BUILD_BUG_ON(NBBY * sizeof(ordered) < XFS_DEFER_OPS_NR_BUFS);
250 list_for_each_entry(lip, &tp->t_items, li_trans) {
251 switch (lip->li_type) {
253 bli = container_of(lip, struct xfs_buf_log_item,
255 if (bli->bli_flags & XFS_BLI_HOLD) {
256 if (bpcount >= XFS_DEFER_OPS_NR_BUFS) {
258 return -EFSCORRUPTED;
260 if (bli->bli_flags & XFS_BLI_ORDERED)
261 ordered |= (1U << bpcount);
263 xfs_trans_dirty_buf(tp, bli->bli_buf);
264 bplist[bpcount++] = bli->bli_buf;
268 ili = container_of(lip, struct xfs_inode_log_item,
270 if (ili->ili_lock_flags == 0) {
271 if (ipcount >= XFS_DEFER_OPS_NR_INODES) {
273 return -EFSCORRUPTED;
275 xfs_trans_log_inode(tp, ili->ili_inode,
277 iplist[ipcount++] = ili->ili_inode;
285 trace_xfs_defer_trans_roll(tp, _RET_IP_);
288 * Roll the transaction. Rolling always given a new transaction (even
289 * if committing the old one fails!) to hand back to the caller, so we
290 * join the held resources to the new transaction so that we always
291 * return with the held resources joined to @tpp, no matter what
294 error = xfs_trans_roll(tpp);
297 /* Rejoin the joined inodes. */
298 for (i = 0; i < ipcount; i++)
299 xfs_trans_ijoin(tp, iplist[i], 0);
301 /* Rejoin the buffers and dirty them so the log moves forward. */
302 for (i = 0; i < bpcount; i++) {
303 xfs_trans_bjoin(tp, bplist[i]);
304 if (ordered & (1U << i))
305 xfs_trans_ordered_buf(tp, bplist[i]);
306 xfs_trans_bhold(tp, bplist[i]);
310 trace_xfs_defer_trans_roll_error(tp, error);
315 * Reset an already used dfops after finish.
319 struct xfs_trans *tp)
321 ASSERT(list_empty(&tp->t_dfops));
324 * Low mode state transfers across transaction rolls to mirror dfops
325 * lifetime. Clear it now that dfops is reset.
327 tp->t_flags &= ~XFS_TRANS_LOWMODE;
331 * Free up any items left in the list.
334 xfs_defer_cancel_list(
335 struct xfs_mount *mp,
336 struct list_head *dop_list)
338 struct xfs_defer_pending *dfp;
339 struct xfs_defer_pending *pli;
340 struct list_head *pwi;
342 const struct xfs_defer_op_type *ops;
345 * Free the pending items. Caller should already have arranged
346 * for the intent items to be released.
348 list_for_each_entry_safe(dfp, pli, dop_list, dfp_list) {
349 ops = defer_op_types[dfp->dfp_type];
350 trace_xfs_defer_cancel_list(mp, dfp);
351 list_del(&dfp->dfp_list);
352 list_for_each_safe(pwi, n, &dfp->dfp_work) {
355 ops->cancel_item(pwi);
357 ASSERT(dfp->dfp_count == 0);
363 * Log an intent-done item for the first pending intent, and finish the work
367 xfs_defer_finish_one(
368 struct xfs_trans *tp,
369 struct xfs_defer_pending *dfp)
371 const struct xfs_defer_op_type *ops = defer_op_types[dfp->dfp_type];
372 struct xfs_btree_cur *state = NULL;
373 struct list_head *li, *n;
376 trace_xfs_defer_pending_finish(tp->t_mountp, dfp);
378 dfp->dfp_done = ops->create_done(tp, dfp->dfp_intent, dfp->dfp_count);
379 list_for_each_safe(li, n, &dfp->dfp_work) {
382 error = ops->finish_item(tp, dfp->dfp_done, li, &state);
383 if (error == -EAGAIN) {
385 * Caller wants a fresh transaction; put the work item
386 * back on the list and log a new log intent item to
387 * replace the old one. See "Requesting a Fresh
388 * Transaction while Finishing Deferred Work" above.
390 list_add(li, &dfp->dfp_work);
392 dfp->dfp_done = NULL;
393 xfs_defer_create_intent(tp, dfp, false);
400 /* Done with the dfp, free it. */
401 list_del(&dfp->dfp_list);
404 if (ops->finish_cleanup)
405 ops->finish_cleanup(tp, state, error);
410 * Finish all the pending work. This involves logging intent items for
411 * any work items that wandered in since the last transaction roll (if
412 * one has even happened), rolling the transaction, and finishing the
413 * work items in the first item on the logged-and-pending list.
415 * If an inode is provided, relog it to the new transaction.
418 xfs_defer_finish_noroll(
419 struct xfs_trans **tp)
421 struct xfs_defer_pending *dfp;
423 LIST_HEAD(dop_pending);
425 ASSERT((*tp)->t_flags & XFS_TRANS_PERM_LOG_RES);
427 trace_xfs_defer_finish(*tp, _RET_IP_);
429 /* Until we run out of pending work to finish... */
430 while (!list_empty(&dop_pending) || !list_empty(&(*tp)->t_dfops)) {
431 xfs_defer_create_intents(*tp);
432 list_splice_tail_init(&(*tp)->t_dfops, &dop_pending);
434 error = xfs_defer_trans_roll(tp);
438 dfp = list_first_entry(&dop_pending, struct xfs_defer_pending,
440 error = xfs_defer_finish_one(*tp, dfp);
441 if (error && error != -EAGAIN)
445 trace_xfs_defer_finish_done(*tp, _RET_IP_);
449 xfs_defer_trans_abort(*tp, &dop_pending);
450 xfs_force_shutdown((*tp)->t_mountp, SHUTDOWN_CORRUPT_INCORE);
451 trace_xfs_defer_finish_error(*tp, error);
452 xfs_defer_cancel_list((*tp)->t_mountp, &dop_pending);
453 xfs_defer_cancel(*tp);
459 struct xfs_trans **tp)
464 * Finish and roll the transaction once more to avoid returning to the
465 * caller with a dirty transaction.
467 error = xfs_defer_finish_noroll(tp);
470 if ((*tp)->t_flags & XFS_TRANS_DIRTY) {
471 error = xfs_defer_trans_roll(tp);
473 xfs_force_shutdown((*tp)->t_mountp,
474 SHUTDOWN_CORRUPT_INCORE);
478 xfs_defer_reset(*tp);
484 struct xfs_trans *tp)
486 struct xfs_mount *mp = tp->t_mountp;
488 trace_xfs_defer_cancel(tp, _RET_IP_);
489 xfs_defer_cancel_list(mp, &tp->t_dfops);
492 /* Add an item for later deferred processing. */
495 struct xfs_trans *tp,
496 enum xfs_defer_ops_type type,
497 struct list_head *li)
499 struct xfs_defer_pending *dfp = NULL;
500 const struct xfs_defer_op_type *ops;
502 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
503 BUILD_BUG_ON(ARRAY_SIZE(defer_op_types) != XFS_DEFER_OPS_TYPE_MAX);
506 * Add the item to a pending item at the end of the intake list.
507 * If the last pending item has the same type, reuse it. Else,
508 * create a new pending item at the end of the intake list.
510 if (!list_empty(&tp->t_dfops)) {
511 dfp = list_last_entry(&tp->t_dfops,
512 struct xfs_defer_pending, dfp_list);
513 ops = defer_op_types[dfp->dfp_type];
514 if (dfp->dfp_type != type ||
515 (ops->max_items && dfp->dfp_count >= ops->max_items))
519 dfp = kmem_alloc(sizeof(struct xfs_defer_pending),
521 dfp->dfp_type = type;
522 dfp->dfp_intent = NULL;
523 dfp->dfp_done = NULL;
525 INIT_LIST_HEAD(&dfp->dfp_work);
526 list_add_tail(&dfp->dfp_list, &tp->t_dfops);
529 list_add_tail(li, &dfp->dfp_work);
534 * Move deferred ops from one transaction to another and reset the source to
535 * initial state. This is primarily used to carry state forward across
536 * transaction rolls with pending dfops.
540 struct xfs_trans *dtp,
541 struct xfs_trans *stp)
543 list_splice_init(&stp->t_dfops, &dtp->t_dfops);
546 * Low free space mode was historically controlled by a dfops field.
547 * This meant that low mode state potentially carried across multiple
548 * transaction rolls. Transfer low mode on a dfops move to preserve
551 dtp->t_flags |= (stp->t_flags & XFS_TRANS_LOWMODE);
553 xfs_defer_reset(stp);