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_format.h"
9 #include "xfs_log_format.h"
10 #include "xfs_trans_resv.h"
12 #include "xfs_shared.h"
13 #include "xfs_mount.h"
14 #include "xfs_defer.h"
15 #include "xfs_trans.h"
16 #include "xfs_trans_priv.h"
17 #include "xfs_refcount_item.h"
19 #include "xfs_refcount.h"
20 #include "xfs_error.h"
21 #include "xfs_log_priv.h"
22 #include "xfs_log_recover.h"
24 struct kmem_cache *xfs_cui_cache;
25 struct kmem_cache *xfs_cud_cache;
27 static const struct xfs_item_ops xfs_cui_item_ops;
29 static inline struct xfs_cui_log_item *CUI_ITEM(struct xfs_log_item *lip)
31 return container_of(lip, struct xfs_cui_log_item, cui_item);
36 struct xfs_cui_log_item *cuip)
38 if (cuip->cui_format.cui_nextents > XFS_CUI_MAX_FAST_EXTENTS)
41 kmem_cache_free(xfs_cui_cache, cuip);
45 * Freeing the CUI requires that we remove it from the AIL if it has already
46 * been placed there. However, the CUI may not yet have been placed in the AIL
47 * when called by xfs_cui_release() from CUD processing due to the ordering of
48 * committed vs unpin operations in bulk insert operations. Hence the reference
49 * count to ensure only the last caller frees the CUI.
53 struct xfs_cui_log_item *cuip)
55 ASSERT(atomic_read(&cuip->cui_refcount) > 0);
56 if (atomic_dec_and_test(&cuip->cui_refcount)) {
57 xfs_trans_ail_delete(&cuip->cui_item, SHUTDOWN_LOG_IO_ERROR);
58 xfs_cui_item_free(cuip);
65 struct xfs_log_item *lip,
69 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
72 *nbytes += xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents);
76 * This is called to fill in the vector of log iovecs for the
77 * given cui log item. We use only 1 iovec, and we point that
78 * at the cui_log_format structure embedded in the cui item.
79 * It is at this point that we assert that all of the extent
80 * slots in the cui item have been filled.
84 struct xfs_log_item *lip,
85 struct xfs_log_vec *lv)
87 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
88 struct xfs_log_iovec *vecp = NULL;
90 ASSERT(atomic_read(&cuip->cui_next_extent) ==
91 cuip->cui_format.cui_nextents);
93 cuip->cui_format.cui_type = XFS_LI_CUI;
94 cuip->cui_format.cui_size = 1;
96 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUI_FORMAT, &cuip->cui_format,
97 xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents));
101 * The unpin operation is the last place an CUI is manipulated in the log. It is
102 * either inserted in the AIL or aborted in the event of a log I/O error. In
103 * either case, the CUI transaction has been successfully committed to make it
104 * this far. Therefore, we expect whoever committed the CUI to either construct
105 * and commit the CUD or drop the CUD's reference in the event of error. Simply
106 * drop the log's CUI reference now that the log is done with it.
110 struct xfs_log_item *lip,
113 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
115 xfs_cui_release(cuip);
119 * The CUI has been either committed or aborted if the transaction has been
120 * cancelled. If the transaction was cancelled, an CUD isn't going to be
121 * constructed and thus we free the CUI here directly.
124 xfs_cui_item_release(
125 struct xfs_log_item *lip)
127 xfs_cui_release(CUI_ITEM(lip));
131 * Allocate and initialize an cui item with the given number of extents.
133 STATIC struct xfs_cui_log_item *
135 struct xfs_mount *mp,
139 struct xfs_cui_log_item *cuip;
141 ASSERT(nextents > 0);
142 if (nextents > XFS_CUI_MAX_FAST_EXTENTS)
143 cuip = kmem_zalloc(xfs_cui_log_item_sizeof(nextents),
146 cuip = kmem_cache_zalloc(xfs_cui_cache,
147 GFP_KERNEL | __GFP_NOFAIL);
149 xfs_log_item_init(mp, &cuip->cui_item, XFS_LI_CUI, &xfs_cui_item_ops);
150 cuip->cui_format.cui_nextents = nextents;
151 cuip->cui_format.cui_id = (uintptr_t)(void *)cuip;
152 atomic_set(&cuip->cui_next_extent, 0);
153 atomic_set(&cuip->cui_refcount, 2);
158 static inline struct xfs_cud_log_item *CUD_ITEM(struct xfs_log_item *lip)
160 return container_of(lip, struct xfs_cud_log_item, cud_item);
165 struct xfs_log_item *lip,
170 *nbytes += sizeof(struct xfs_cud_log_format);
174 * This is called to fill in the vector of log iovecs for the
175 * given cud log item. We use only 1 iovec, and we point that
176 * at the cud_log_format structure embedded in the cud item.
177 * It is at this point that we assert that all of the extent
178 * slots in the cud item have been filled.
182 struct xfs_log_item *lip,
183 struct xfs_log_vec *lv)
185 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
186 struct xfs_log_iovec *vecp = NULL;
188 cudp->cud_format.cud_type = XFS_LI_CUD;
189 cudp->cud_format.cud_size = 1;
191 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUD_FORMAT, &cudp->cud_format,
192 sizeof(struct xfs_cud_log_format));
196 * The CUD is either committed or aborted if the transaction is cancelled. If
197 * the transaction is cancelled, drop our reference to the CUI and free the
201 xfs_cud_item_release(
202 struct xfs_log_item *lip)
204 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
206 xfs_cui_release(cudp->cud_cuip);
207 kmem_cache_free(xfs_cud_cache, cudp);
210 static const struct xfs_item_ops xfs_cud_item_ops = {
211 .flags = XFS_ITEM_RELEASE_WHEN_COMMITTED,
212 .iop_size = xfs_cud_item_size,
213 .iop_format = xfs_cud_item_format,
214 .iop_release = xfs_cud_item_release,
217 static struct xfs_cud_log_item *
219 struct xfs_trans *tp,
220 struct xfs_cui_log_item *cuip)
222 struct xfs_cud_log_item *cudp;
224 cudp = kmem_cache_zalloc(xfs_cud_cache, GFP_KERNEL | __GFP_NOFAIL);
225 xfs_log_item_init(tp->t_mountp, &cudp->cud_item, XFS_LI_CUD,
227 cudp->cud_cuip = cuip;
228 cudp->cud_format.cud_cui_id = cuip->cui_format.cui_id;
230 xfs_trans_add_item(tp, &cudp->cud_item);
235 * Finish an refcount update and log it to the CUD. Note that the
236 * transaction is marked dirty regardless of whether the refcount
237 * update succeeds or fails to support the CUI/CUD lifecycle rules.
240 xfs_trans_log_finish_refcount_update(
241 struct xfs_trans *tp,
242 struct xfs_cud_log_item *cudp,
243 enum xfs_refcount_intent_type type,
244 xfs_fsblock_t startblock,
245 xfs_extlen_t blockcount,
246 xfs_fsblock_t *new_fsb,
247 xfs_extlen_t *new_len,
248 struct xfs_btree_cur **pcur)
252 error = xfs_refcount_finish_one(tp, type, startblock,
253 blockcount, new_fsb, new_len, pcur);
256 * Mark the transaction dirty, even on error. This ensures the
257 * transaction is aborted, which:
259 * 1.) releases the CUI and frees the CUD
260 * 2.) shuts down the filesystem
262 tp->t_flags |= XFS_TRANS_DIRTY;
263 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
268 /* Sort refcount intents by AG. */
270 xfs_refcount_update_diff_items(
272 const struct list_head *a,
273 const struct list_head *b)
275 struct xfs_mount *mp = priv;
276 struct xfs_refcount_intent *ra;
277 struct xfs_refcount_intent *rb;
279 ra = container_of(a, struct xfs_refcount_intent, ri_list);
280 rb = container_of(b, struct xfs_refcount_intent, ri_list);
281 return XFS_FSB_TO_AGNO(mp, ra->ri_startblock) -
282 XFS_FSB_TO_AGNO(mp, rb->ri_startblock);
285 /* Set the phys extent flags for this reverse mapping. */
287 xfs_trans_set_refcount_flags(
288 struct xfs_phys_extent *refc,
289 enum xfs_refcount_intent_type type)
293 case XFS_REFCOUNT_INCREASE:
294 case XFS_REFCOUNT_DECREASE:
295 case XFS_REFCOUNT_ALLOC_COW:
296 case XFS_REFCOUNT_FREE_COW:
297 refc->pe_flags |= type;
304 /* Log refcount updates in the intent item. */
306 xfs_refcount_update_log_item(
307 struct xfs_trans *tp,
308 struct xfs_cui_log_item *cuip,
309 struct xfs_refcount_intent *refc)
312 struct xfs_phys_extent *ext;
314 tp->t_flags |= XFS_TRANS_DIRTY;
315 set_bit(XFS_LI_DIRTY, &cuip->cui_item.li_flags);
318 * atomic_inc_return gives us the value after the increment;
319 * we want to use it as an array index so we need to subtract 1 from
322 next_extent = atomic_inc_return(&cuip->cui_next_extent) - 1;
323 ASSERT(next_extent < cuip->cui_format.cui_nextents);
324 ext = &cuip->cui_format.cui_extents[next_extent];
325 ext->pe_startblock = refc->ri_startblock;
326 ext->pe_len = refc->ri_blockcount;
327 xfs_trans_set_refcount_flags(ext, refc->ri_type);
330 static struct xfs_log_item *
331 xfs_refcount_update_create_intent(
332 struct xfs_trans *tp,
333 struct list_head *items,
337 struct xfs_mount *mp = tp->t_mountp;
338 struct xfs_cui_log_item *cuip = xfs_cui_init(mp, count);
339 struct xfs_refcount_intent *refc;
343 xfs_trans_add_item(tp, &cuip->cui_item);
345 list_sort(mp, items, xfs_refcount_update_diff_items);
346 list_for_each_entry(refc, items, ri_list)
347 xfs_refcount_update_log_item(tp, cuip, refc);
348 return &cuip->cui_item;
351 /* Get an CUD so we can process all the deferred refcount updates. */
352 static struct xfs_log_item *
353 xfs_refcount_update_create_done(
354 struct xfs_trans *tp,
355 struct xfs_log_item *intent,
358 return &xfs_trans_get_cud(tp, CUI_ITEM(intent))->cud_item;
361 /* Process a deferred refcount update. */
363 xfs_refcount_update_finish_item(
364 struct xfs_trans *tp,
365 struct xfs_log_item *done,
366 struct list_head *item,
367 struct xfs_btree_cur **state)
369 struct xfs_refcount_intent *refc;
370 xfs_fsblock_t new_fsb;
371 xfs_extlen_t new_aglen;
374 refc = container_of(item, struct xfs_refcount_intent, ri_list);
375 error = xfs_trans_log_finish_refcount_update(tp, CUD_ITEM(done),
376 refc->ri_type, refc->ri_startblock, refc->ri_blockcount,
377 &new_fsb, &new_aglen, state);
379 /* Did we run out of reservation? Requeue what we didn't finish. */
380 if (!error && new_aglen > 0) {
381 ASSERT(refc->ri_type == XFS_REFCOUNT_INCREASE ||
382 refc->ri_type == XFS_REFCOUNT_DECREASE);
383 refc->ri_startblock = new_fsb;
384 refc->ri_blockcount = new_aglen;
387 kmem_cache_free(xfs_refcount_intent_cache, refc);
391 /* Abort all pending CUIs. */
393 xfs_refcount_update_abort_intent(
394 struct xfs_log_item *intent)
396 xfs_cui_release(CUI_ITEM(intent));
399 /* Cancel a deferred refcount update. */
401 xfs_refcount_update_cancel_item(
402 struct list_head *item)
404 struct xfs_refcount_intent *refc;
406 refc = container_of(item, struct xfs_refcount_intent, ri_list);
407 kmem_cache_free(xfs_refcount_intent_cache, refc);
410 const struct xfs_defer_op_type xfs_refcount_update_defer_type = {
411 .max_items = XFS_CUI_MAX_FAST_EXTENTS,
412 .create_intent = xfs_refcount_update_create_intent,
413 .abort_intent = xfs_refcount_update_abort_intent,
414 .create_done = xfs_refcount_update_create_done,
415 .finish_item = xfs_refcount_update_finish_item,
416 .finish_cleanup = xfs_refcount_finish_one_cleanup,
417 .cancel_item = xfs_refcount_update_cancel_item,
420 /* Is this recovered CUI ok? */
422 xfs_cui_validate_phys(
423 struct xfs_mount *mp,
424 struct xfs_phys_extent *refc)
426 if (!xfs_has_reflink(mp))
429 if (refc->pe_flags & ~XFS_REFCOUNT_EXTENT_FLAGS)
432 switch (refc->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK) {
433 case XFS_REFCOUNT_INCREASE:
434 case XFS_REFCOUNT_DECREASE:
435 case XFS_REFCOUNT_ALLOC_COW:
436 case XFS_REFCOUNT_FREE_COW:
442 return xfs_verify_fsbext(mp, refc->pe_startblock, refc->pe_len);
446 * Process a refcount update intent item that was recovered from the log.
447 * We need to update the refcountbt.
450 xfs_cui_item_recover(
451 struct xfs_log_item *lip,
452 struct list_head *capture_list)
454 struct xfs_bmbt_irec irec;
455 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
456 struct xfs_phys_extent *refc;
457 struct xfs_cud_log_item *cudp;
458 struct xfs_trans *tp;
459 struct xfs_btree_cur *rcur = NULL;
460 struct xfs_mount *mp = lip->li_log->l_mp;
461 xfs_fsblock_t new_fsb;
462 xfs_extlen_t new_len;
463 unsigned int refc_type;
464 bool requeue_only = false;
465 enum xfs_refcount_intent_type type;
470 * First check the validity of the extents described by the
471 * CUI. If any are bad, then assume that all are bad and
474 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
475 if (!xfs_cui_validate_phys(mp,
476 &cuip->cui_format.cui_extents[i])) {
477 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
479 sizeof(cuip->cui_format));
480 return -EFSCORRUPTED;
485 * Under normal operation, refcount updates are deferred, so we
486 * wouldn't be adding them directly to a transaction. All
487 * refcount updates manage reservation usage internally and
488 * dynamically by deferring work that won't fit in the
489 * transaction. Normally, any work that needs to be deferred
490 * gets attached to the same defer_ops that scheduled the
491 * refcount update. However, we're in log recovery here, so we
492 * use the passed in defer_ops and to finish up any work that
493 * doesn't fit. We need to reserve enough blocks to handle a
494 * full btree split on either end of the refcount range.
496 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate,
497 mp->m_refc_maxlevels * 2, 0, XFS_TRANS_RESERVE, &tp);
501 cudp = xfs_trans_get_cud(tp, cuip);
503 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
504 refc = &cuip->cui_format.cui_extents[i];
505 refc_type = refc->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK;
507 case XFS_REFCOUNT_INCREASE:
508 case XFS_REFCOUNT_DECREASE:
509 case XFS_REFCOUNT_ALLOC_COW:
510 case XFS_REFCOUNT_FREE_COW:
514 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, mp);
515 error = -EFSCORRUPTED;
519 new_fsb = refc->pe_startblock;
520 new_len = refc->pe_len;
522 error = xfs_trans_log_finish_refcount_update(tp, cudp,
523 type, refc->pe_startblock, refc->pe_len,
524 &new_fsb, &new_len, &rcur);
525 if (error == -EFSCORRUPTED)
526 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
527 refc, sizeof(*refc));
531 /* Requeue what we didn't finish. */
533 irec.br_startblock = new_fsb;
534 irec.br_blockcount = new_len;
536 case XFS_REFCOUNT_INCREASE:
537 xfs_refcount_increase_extent(tp, &irec);
539 case XFS_REFCOUNT_DECREASE:
540 xfs_refcount_decrease_extent(tp, &irec);
542 case XFS_REFCOUNT_ALLOC_COW:
543 xfs_refcount_alloc_cow_extent(tp,
547 case XFS_REFCOUNT_FREE_COW:
548 xfs_refcount_free_cow_extent(tp,
559 xfs_refcount_finish_one_cleanup(tp, rcur, error);
560 return xfs_defer_ops_capture_and_commit(tp, capture_list);
563 xfs_refcount_finish_one_cleanup(tp, rcur, error);
564 xfs_trans_cancel(tp);
570 struct xfs_log_item *lip,
573 return CUI_ITEM(lip)->cui_format.cui_id == intent_id;
576 /* Relog an intent item to push the log tail forward. */
577 static struct xfs_log_item *
579 struct xfs_log_item *intent,
580 struct xfs_trans *tp)
582 struct xfs_cud_log_item *cudp;
583 struct xfs_cui_log_item *cuip;
584 struct xfs_phys_extent *extp;
587 count = CUI_ITEM(intent)->cui_format.cui_nextents;
588 extp = CUI_ITEM(intent)->cui_format.cui_extents;
590 tp->t_flags |= XFS_TRANS_DIRTY;
591 cudp = xfs_trans_get_cud(tp, CUI_ITEM(intent));
592 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
594 cuip = xfs_cui_init(tp->t_mountp, count);
595 memcpy(cuip->cui_format.cui_extents, extp, count * sizeof(*extp));
596 atomic_set(&cuip->cui_next_extent, count);
597 xfs_trans_add_item(tp, &cuip->cui_item);
598 set_bit(XFS_LI_DIRTY, &cuip->cui_item.li_flags);
599 return &cuip->cui_item;
602 static const struct xfs_item_ops xfs_cui_item_ops = {
603 .iop_size = xfs_cui_item_size,
604 .iop_format = xfs_cui_item_format,
605 .iop_unpin = xfs_cui_item_unpin,
606 .iop_release = xfs_cui_item_release,
607 .iop_recover = xfs_cui_item_recover,
608 .iop_match = xfs_cui_item_match,
609 .iop_relog = xfs_cui_item_relog,
613 * Copy an CUI format buffer from the given buf, and into the destination
614 * CUI format structure. The CUI/CUD items were designed not to need any
615 * special alignment handling.
619 struct xfs_log_iovec *buf,
620 struct xfs_cui_log_format *dst_cui_fmt)
622 struct xfs_cui_log_format *src_cui_fmt;
625 src_cui_fmt = buf->i_addr;
626 len = xfs_cui_log_format_sizeof(src_cui_fmt->cui_nextents);
628 if (buf->i_len == len) {
629 memcpy(dst_cui_fmt, src_cui_fmt, len);
632 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, NULL);
633 return -EFSCORRUPTED;
637 * This routine is called to create an in-core extent refcount update
638 * item from the cui format structure which was logged on disk.
639 * It allocates an in-core cui, copies the extents from the format
640 * structure into it, and adds the cui to the AIL with the given
644 xlog_recover_cui_commit_pass2(
646 struct list_head *buffer_list,
647 struct xlog_recover_item *item,
651 struct xfs_mount *mp = log->l_mp;
652 struct xfs_cui_log_item *cuip;
653 struct xfs_cui_log_format *cui_formatp;
655 cui_formatp = item->ri_buf[0].i_addr;
657 cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
658 error = xfs_cui_copy_format(&item->ri_buf[0], &cuip->cui_format);
660 xfs_cui_item_free(cuip);
663 atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
665 * Insert the intent into the AIL directly and drop one reference so
666 * that finishing or canceling the work will drop the other.
668 xfs_trans_ail_insert(log->l_ailp, &cuip->cui_item, lsn);
669 xfs_cui_release(cuip);
673 const struct xlog_recover_item_ops xlog_cui_item_ops = {
674 .item_type = XFS_LI_CUI,
675 .commit_pass2 = xlog_recover_cui_commit_pass2,
679 * This routine is called when an CUD format structure is found in a committed
680 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
681 * was still in the log. To do this it searches the AIL for the CUI with an id
682 * equal to that in the CUD format structure. If we find it we drop the CUD
683 * reference, which removes the CUI from the AIL and frees it.
686 xlog_recover_cud_commit_pass2(
688 struct list_head *buffer_list,
689 struct xlog_recover_item *item,
692 struct xfs_cud_log_format *cud_formatp;
694 cud_formatp = item->ri_buf[0].i_addr;
695 if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format)) {
696 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
697 return -EFSCORRUPTED;
700 xlog_recover_release_intent(log, XFS_LI_CUI, cud_formatp->cud_cui_id);
704 const struct xlog_recover_item_ops xlog_cud_item_ops = {
705 .item_type = XFS_LI_CUD,
706 .commit_pass2 = xlog_recover_cud_commit_pass2,