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"
14 #include "xfs_mount.h"
15 #include "xfs_defer.h"
16 #include "xfs_trans.h"
17 #include "xfs_buf_item.h"
18 #include "xfs_inode.h"
19 #include "xfs_inode_item.h"
20 #include "xfs_trace.h"
23 * Deferred Operations in XFS
25 * Due to the way locking rules work in XFS, certain transactions (block
26 * mapping and unmapping, typically) have permanent reservations so that
27 * we can roll the transaction to adhere to AG locking order rules and
28 * to unlock buffers between metadata updates. Prior to rmap/reflink,
29 * the mapping code had a mechanism to perform these deferrals for
30 * extents that were going to be freed; this code makes that facility
33 * When adding the reverse mapping and reflink features, it became
34 * necessary to perform complex remapping multi-transactions to comply
35 * with AG locking order rules, and to be able to spread a single
36 * refcount update operation (an operation on an n-block extent can
37 * update as many as n records!) among multiple transactions. XFS can
38 * roll a transaction to facilitate this, but using this facility
39 * requires us to log "intent" items in case log recovery needs to
40 * redo the operation, and to log "done" items to indicate that redo
43 * Deferred work is tracked in xfs_defer_pending items. Each pending
44 * item tracks one type of deferred work. Incoming work items (which
45 * have not yet had an intent logged) are attached to a pending item
46 * on the dop_intake list, where they wait for the caller to finish
47 * the deferred operations.
49 * Finishing a set of deferred operations is an involved process. To
50 * start, we define "rolling a deferred-op transaction" as follows:
52 * > For each xfs_defer_pending item on the dop_intake list,
53 * - Sort the work items in AG order. XFS locking
54 * order rules require us to lock buffers in AG order.
55 * - Create a log intent item for that type.
56 * - Attach it to the pending item.
57 * - Move the pending item from the dop_intake list to the
59 * > Roll the transaction.
61 * NOTE: To avoid exceeding the transaction reservation, we limit the
62 * number of items that we attach to a given xfs_defer_pending.
64 * The actual finishing process looks like this:
66 * > For each xfs_defer_pending in the dop_pending list,
67 * - Roll the deferred-op transaction as above.
68 * - Create a log done item for that type, and attach it to the
70 * - For each work item attached to the log intent item,
71 * * Perform the described action.
72 * * Attach the work item to the log done item.
73 * * If the result of doing the work was -EAGAIN, ->finish work
74 * wants a new transaction. See the "Requesting a Fresh
75 * Transaction while Finishing Deferred Work" section below for
78 * The key here is that we must log an intent item for all pending
79 * work items every time we roll the transaction, and that we must log
80 * a done item as soon as the work is completed. With this mechanism
81 * we can perform complex remapping operations, chaining intent items
84 * Requesting a Fresh Transaction while Finishing Deferred Work
86 * If ->finish_item decides that it needs a fresh transaction to
87 * finish the work, it must ask its caller (xfs_defer_finish) for a
88 * continuation. The most likely cause of this circumstance are the
89 * refcount adjust functions deciding that they've logged enough items
90 * to be at risk of exceeding the transaction reservation.
92 * To get a fresh transaction, we want to log the existing log done
93 * item to prevent the log intent item from replaying, immediately log
94 * a new log intent item with the unfinished work items, roll the
95 * transaction, and re-call ->finish_item wherever it left off. The
96 * log done item and the new log intent item must be in the same
97 * transaction or atomicity cannot be guaranteed; defer_finish ensures
100 * This requires some coordination between ->finish_item and
101 * defer_finish. Upon deciding to request a new transaction,
102 * ->finish_item should update the current work item to reflect the
103 * unfinished work. Next, it should reset the log done item's list
104 * count to the number of items finished, and return -EAGAIN.
105 * defer_finish sees the -EAGAIN, logs the new log intent item
106 * with the remaining work items, and leaves the xfs_defer_pending
107 * item at the head of the dop_work queue. Then it rolls the
108 * transaction and picks up processing where it left off. It is
109 * required that ->finish_item must be careful to leave enough
110 * transaction reservation to fit the new log intent item.
112 * This is an example of remapping the extent (E, E+B) into file X at
113 * offset A and dealing with the extent (C, C+B) already being mapped
115 * +-------------------------------------------------+
116 * | Unmap file X startblock C offset A length B | t0
117 * | Intent to reduce refcount for extent (C, B) |
118 * | Intent to remove rmap (X, C, A, B) |
119 * | Intent to free extent (D, 1) (bmbt block) |
120 * | Intent to map (X, A, B) at startblock E |
121 * +-------------------------------------------------+
122 * | Map file X startblock E offset A length B | t1
123 * | Done mapping (X, E, A, B) |
124 * | Intent to increase refcount for extent (E, B) |
125 * | Intent to add rmap (X, E, A, B) |
126 * +-------------------------------------------------+
127 * | Reduce refcount for extent (C, B) | t2
128 * | Done reducing refcount for extent (C, 9) |
129 * | Intent to reduce refcount for extent (C+9, B-9) |
130 * | (ran out of space after 9 refcount updates) |
131 * +-------------------------------------------------+
132 * | Reduce refcount for extent (C+9, B+9) | t3
133 * | Done reducing refcount for extent (C+9, B-9) |
134 * | Increase refcount for extent (E, B) |
135 * | Done increasing refcount for extent (E, B) |
136 * | Intent to free extent (C, B) |
137 * | Intent to free extent (F, 1) (refcountbt block) |
138 * | Intent to remove rmap (F, 1, REFC) |
139 * +-------------------------------------------------+
140 * | Remove rmap (X, C, A, B) | t4
141 * | Done removing rmap (X, C, A, B) |
142 * | Add rmap (X, E, A, B) |
143 * | Done adding rmap (X, E, A, B) |
144 * | Remove rmap (F, 1, REFC) |
145 * | Done removing rmap (F, 1, REFC) |
146 * +-------------------------------------------------+
147 * | Free extent (C, B) | t5
148 * | Done freeing extent (C, B) |
149 * | Free extent (D, 1) |
150 * | Done freeing extent (D, 1) |
151 * | Free extent (F, 1) |
152 * | Done freeing extent (F, 1) |
153 * +-------------------------------------------------+
155 * If we should crash before t2 commits, log recovery replays
156 * the following intent items:
158 * - Intent to reduce refcount for extent (C, B)
159 * - Intent to remove rmap (X, C, A, B)
160 * - Intent to free extent (D, 1) (bmbt block)
161 * - Intent to increase refcount for extent (E, B)
162 * - Intent to add rmap (X, E, A, B)
164 * In the process of recovering, it should also generate and take care
165 * of these intent items:
167 * - Intent to free extent (C, B)
168 * - Intent to free extent (F, 1) (refcountbt block)
169 * - Intent to remove rmap (F, 1, REFC)
171 * Note that the continuation requested between t2 and t3 is likely to
175 static const struct xfs_defer_op_type *defer_op_types[XFS_DEFER_OPS_TYPE_MAX];
178 * For each pending item in the intake list, log its intent item and the
179 * associated extents, then add the entire intake list to the end of
183 xfs_defer_create_intents(
184 struct xfs_trans *tp)
186 struct list_head *li;
187 struct xfs_defer_pending *dfp;
189 list_for_each_entry(dfp, &tp->t_dfops, dfp_list) {
190 dfp->dfp_intent = dfp->dfp_type->create_intent(tp,
192 trace_xfs_defer_create_intent(tp->t_mountp, dfp);
193 list_sort(tp->t_mountp, &dfp->dfp_work,
194 dfp->dfp_type->diff_items);
195 list_for_each(li, &dfp->dfp_work)
196 dfp->dfp_type->log_item(tp, dfp->dfp_intent, li);
200 /* Abort all the intents that were committed. */
202 xfs_defer_trans_abort(
203 struct xfs_trans *tp,
204 struct list_head *dop_pending)
206 struct xfs_defer_pending *dfp;
208 trace_xfs_defer_trans_abort(tp, _RET_IP_);
210 /* Abort intent items that don't have a done item. */
211 list_for_each_entry(dfp, dop_pending, dfp_list) {
212 trace_xfs_defer_pending_abort(tp->t_mountp, dfp);
213 if (dfp->dfp_intent && !dfp->dfp_done) {
214 dfp->dfp_type->abort_intent(dfp->dfp_intent);
215 dfp->dfp_intent = NULL;
220 /* Roll a transaction so we can do some deferred op processing. */
222 xfs_defer_trans_roll(
223 struct xfs_trans **tpp)
225 struct xfs_trans *tp = *tpp;
226 struct xfs_buf_log_item *bli;
227 struct xfs_inode_log_item *ili;
228 struct xfs_log_item *lip;
229 struct xfs_buf *bplist[XFS_DEFER_OPS_NR_BUFS];
230 struct xfs_inode *iplist[XFS_DEFER_OPS_NR_INODES];
231 int bpcount = 0, ipcount = 0;
235 list_for_each_entry(lip, &tp->t_items, li_trans) {
236 switch (lip->li_type) {
238 bli = container_of(lip, struct xfs_buf_log_item,
240 if (bli->bli_flags & XFS_BLI_HOLD) {
241 if (bpcount >= XFS_DEFER_OPS_NR_BUFS) {
243 return -EFSCORRUPTED;
245 xfs_trans_dirty_buf(tp, bli->bli_buf);
246 bplist[bpcount++] = bli->bli_buf;
250 ili = container_of(lip, struct xfs_inode_log_item,
252 if (ili->ili_lock_flags == 0) {
253 if (ipcount >= XFS_DEFER_OPS_NR_INODES) {
255 return -EFSCORRUPTED;
257 xfs_trans_log_inode(tp, ili->ili_inode,
259 iplist[ipcount++] = ili->ili_inode;
267 trace_xfs_defer_trans_roll(tp, _RET_IP_);
270 * Roll the transaction. Rolling always given a new transaction (even
271 * if committing the old one fails!) to hand back to the caller, so we
272 * join the held resources to the new transaction so that we always
273 * return with the held resources joined to @tpp, no matter what
276 error = xfs_trans_roll(tpp);
279 /* Rejoin the joined inodes. */
280 for (i = 0; i < ipcount; i++)
281 xfs_trans_ijoin(tp, iplist[i], 0);
283 /* Rejoin the buffers and dirty them so the log moves forward. */
284 for (i = 0; i < bpcount; i++) {
285 xfs_trans_bjoin(tp, bplist[i]);
286 xfs_trans_bhold(tp, bplist[i]);
290 trace_xfs_defer_trans_roll_error(tp, error);
295 * Reset an already used dfops after finish.
299 struct xfs_trans *tp)
301 ASSERT(list_empty(&tp->t_dfops));
304 * Low mode state transfers across transaction rolls to mirror dfops
305 * lifetime. Clear it now that dfops is reset.
307 tp->t_flags &= ~XFS_TRANS_LOWMODE;
311 * Free up any items left in the list.
314 xfs_defer_cancel_list(
315 struct xfs_mount *mp,
316 struct list_head *dop_list)
318 struct xfs_defer_pending *dfp;
319 struct xfs_defer_pending *pli;
320 struct list_head *pwi;
324 * Free the pending items. Caller should already have arranged
325 * for the intent items to be released.
327 list_for_each_entry_safe(dfp, pli, dop_list, dfp_list) {
328 trace_xfs_defer_cancel_list(mp, dfp);
329 list_del(&dfp->dfp_list);
330 list_for_each_safe(pwi, n, &dfp->dfp_work) {
333 dfp->dfp_type->cancel_item(pwi);
335 ASSERT(dfp->dfp_count == 0);
341 * Finish all the pending work. This involves logging intent items for
342 * any work items that wandered in since the last transaction roll (if
343 * one has even happened), rolling the transaction, and finishing the
344 * work items in the first item on the logged-and-pending list.
346 * If an inode is provided, relog it to the new transaction.
349 xfs_defer_finish_noroll(
350 struct xfs_trans **tp)
352 struct xfs_defer_pending *dfp;
353 struct list_head *li;
357 void (*cleanup_fn)(struct xfs_trans *, void *, int);
358 LIST_HEAD(dop_pending);
360 ASSERT((*tp)->t_flags & XFS_TRANS_PERM_LOG_RES);
362 trace_xfs_defer_finish(*tp, _RET_IP_);
364 /* Until we run out of pending work to finish... */
365 while (!list_empty(&dop_pending) || !list_empty(&(*tp)->t_dfops)) {
366 /* log intents and pull in intake items */
367 xfs_defer_create_intents(*tp);
368 list_splice_tail_init(&(*tp)->t_dfops, &dop_pending);
371 * Roll the transaction.
373 error = xfs_defer_trans_roll(tp);
377 /* Log an intent-done item for the first pending item. */
378 dfp = list_first_entry(&dop_pending, struct xfs_defer_pending,
380 trace_xfs_defer_pending_finish((*tp)->t_mountp, dfp);
381 dfp->dfp_done = dfp->dfp_type->create_done(*tp, dfp->dfp_intent,
383 cleanup_fn = dfp->dfp_type->finish_cleanup;
385 /* Finish the work items. */
387 list_for_each_safe(li, n, &dfp->dfp_work) {
390 error = dfp->dfp_type->finish_item(*tp, li,
391 dfp->dfp_done, &state);
392 if (error == -EAGAIN) {
394 * Caller wants a fresh transaction;
395 * put the work item back on the list
398 list_add(li, &dfp->dfp_work);
403 * Clean up after ourselves and jump out.
404 * xfs_defer_cancel will take care of freeing
405 * all these lists and stuff.
408 cleanup_fn(*tp, state, error);
412 if (error == -EAGAIN) {
414 * Caller wants a fresh transaction, so log a
415 * new log intent item to replace the old one
416 * and roll the transaction. See "Requesting
417 * a Fresh Transaction while Finishing
418 * Deferred Work" above.
420 dfp->dfp_intent = dfp->dfp_type->create_intent(*tp,
422 dfp->dfp_done = NULL;
423 list_for_each(li, &dfp->dfp_work)
424 dfp->dfp_type->log_item(*tp, dfp->dfp_intent,
427 /* Done with the dfp, free it. */
428 list_del(&dfp->dfp_list);
433 cleanup_fn(*tp, state, error);
438 xfs_defer_trans_abort(*tp, &dop_pending);
439 xfs_force_shutdown((*tp)->t_mountp, SHUTDOWN_CORRUPT_INCORE);
440 trace_xfs_defer_finish_error(*tp, error);
441 xfs_defer_cancel_list((*tp)->t_mountp, &dop_pending);
442 xfs_defer_cancel(*tp);
446 trace_xfs_defer_finish_done(*tp, _RET_IP_);
452 struct xfs_trans **tp)
457 * Finish and roll the transaction once more to avoid returning to the
458 * caller with a dirty transaction.
460 error = xfs_defer_finish_noroll(tp);
463 if ((*tp)->t_flags & XFS_TRANS_DIRTY) {
464 error = xfs_defer_trans_roll(tp);
466 xfs_force_shutdown((*tp)->t_mountp,
467 SHUTDOWN_CORRUPT_INCORE);
471 xfs_defer_reset(*tp);
477 struct xfs_trans *tp)
479 struct xfs_mount *mp = tp->t_mountp;
481 trace_xfs_defer_cancel(tp, _RET_IP_);
482 xfs_defer_cancel_list(mp, &tp->t_dfops);
485 /* Add an item for later deferred processing. */
488 struct xfs_trans *tp,
489 enum xfs_defer_ops_type type,
490 struct list_head *li)
492 struct xfs_defer_pending *dfp = NULL;
494 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
497 * Add the item to a pending item at the end of the intake list.
498 * If the last pending item has the same type, reuse it. Else,
499 * create a new pending item at the end of the intake list.
501 if (!list_empty(&tp->t_dfops)) {
502 dfp = list_last_entry(&tp->t_dfops,
503 struct xfs_defer_pending, dfp_list);
504 if (dfp->dfp_type->type != type ||
505 (dfp->dfp_type->max_items &&
506 dfp->dfp_count >= dfp->dfp_type->max_items))
510 dfp = kmem_alloc(sizeof(struct xfs_defer_pending),
512 dfp->dfp_type = defer_op_types[type];
513 dfp->dfp_intent = NULL;
514 dfp->dfp_done = NULL;
516 INIT_LIST_HEAD(&dfp->dfp_work);
517 list_add_tail(&dfp->dfp_list, &tp->t_dfops);
520 list_add_tail(li, &dfp->dfp_work);
524 /* Initialize a deferred operation list. */
526 xfs_defer_init_op_type(
527 const struct xfs_defer_op_type *type)
529 defer_op_types[type->type] = type;
533 * Move deferred ops from one transaction to another and reset the source to
534 * initial state. This is primarily used to carry state forward across
535 * transaction rolls with pending dfops.
539 struct xfs_trans *dtp,
540 struct xfs_trans *stp)
542 list_splice_init(&stp->t_dfops, &dtp->t_dfops);
545 * Low free space mode was historically controlled by a dfops field.
546 * This meant that low mode state potentially carried across multiple
547 * transaction rolls. Transfer low mode on a dfops move to preserve
550 dtp->t_flags |= (stp->t_flags & XFS_TRANS_LOWMODE);
552 xfs_defer_reset(stp);