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"
25 struct kmem_cache *xfs_cui_cache;
26 struct kmem_cache *xfs_cud_cache;
28 static const struct xfs_item_ops xfs_cui_item_ops;
30 static inline struct xfs_cui_log_item *CUI_ITEM(struct xfs_log_item *lip)
32 return container_of(lip, struct xfs_cui_log_item, cui_item);
37 struct xfs_cui_log_item *cuip)
39 kmem_free(cuip->cui_item.li_lv_shadow);
40 if (cuip->cui_format.cui_nextents > XFS_CUI_MAX_FAST_EXTENTS)
43 kmem_cache_free(xfs_cui_cache, cuip);
47 * Freeing the CUI requires that we remove it from the AIL if it has already
48 * been placed there. However, the CUI may not yet have been placed in the AIL
49 * when called by xfs_cui_release() from CUD processing due to the ordering of
50 * committed vs unpin operations in bulk insert operations. Hence the reference
51 * count to ensure only the last caller frees the CUI.
55 struct xfs_cui_log_item *cuip)
57 ASSERT(atomic_read(&cuip->cui_refcount) > 0);
58 if (!atomic_dec_and_test(&cuip->cui_refcount))
61 xfs_trans_ail_delete(&cuip->cui_item, 0);
62 xfs_cui_item_free(cuip);
68 struct xfs_log_item *lip,
72 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
75 *nbytes += xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents);
79 * This is called to fill in the vector of log iovecs for the
80 * given cui log item. We use only 1 iovec, and we point that
81 * at the cui_log_format structure embedded in the cui item.
82 * It is at this point that we assert that all of the extent
83 * slots in the cui item have been filled.
87 struct xfs_log_item *lip,
88 struct xfs_log_vec *lv)
90 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
91 struct xfs_log_iovec *vecp = NULL;
93 ASSERT(atomic_read(&cuip->cui_next_extent) ==
94 cuip->cui_format.cui_nextents);
96 cuip->cui_format.cui_type = XFS_LI_CUI;
97 cuip->cui_format.cui_size = 1;
99 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUI_FORMAT, &cuip->cui_format,
100 xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents));
104 * The unpin operation is the last place an CUI is manipulated in the log. It is
105 * either inserted in the AIL or aborted in the event of a log I/O error. In
106 * either case, the CUI transaction has been successfully committed to make it
107 * this far. Therefore, we expect whoever committed the CUI to either construct
108 * and commit the CUD or drop the CUD's reference in the event of error. Simply
109 * drop the log's CUI reference now that the log is done with it.
113 struct xfs_log_item *lip,
116 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
118 xfs_cui_release(cuip);
122 * The CUI has been either committed or aborted if the transaction has been
123 * cancelled. If the transaction was cancelled, an CUD isn't going to be
124 * constructed and thus we free the CUI here directly.
127 xfs_cui_item_release(
128 struct xfs_log_item *lip)
130 xfs_cui_release(CUI_ITEM(lip));
134 * Allocate and initialize an cui item with the given number of extents.
136 STATIC struct xfs_cui_log_item *
138 struct xfs_mount *mp,
142 struct xfs_cui_log_item *cuip;
144 ASSERT(nextents > 0);
145 if (nextents > XFS_CUI_MAX_FAST_EXTENTS)
146 cuip = kmem_zalloc(xfs_cui_log_item_sizeof(nextents),
149 cuip = kmem_cache_zalloc(xfs_cui_cache,
150 GFP_KERNEL | __GFP_NOFAIL);
152 xfs_log_item_init(mp, &cuip->cui_item, XFS_LI_CUI, &xfs_cui_item_ops);
153 cuip->cui_format.cui_nextents = nextents;
154 cuip->cui_format.cui_id = (uintptr_t)(void *)cuip;
155 atomic_set(&cuip->cui_next_extent, 0);
156 atomic_set(&cuip->cui_refcount, 2);
161 static inline struct xfs_cud_log_item *CUD_ITEM(struct xfs_log_item *lip)
163 return container_of(lip, struct xfs_cud_log_item, cud_item);
168 struct xfs_log_item *lip,
173 *nbytes += sizeof(struct xfs_cud_log_format);
177 * This is called to fill in the vector of log iovecs for the
178 * given cud log item. We use only 1 iovec, and we point that
179 * at the cud_log_format structure embedded in the cud item.
180 * It is at this point that we assert that all of the extent
181 * slots in the cud item have been filled.
185 struct xfs_log_item *lip,
186 struct xfs_log_vec *lv)
188 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
189 struct xfs_log_iovec *vecp = NULL;
191 cudp->cud_format.cud_type = XFS_LI_CUD;
192 cudp->cud_format.cud_size = 1;
194 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUD_FORMAT, &cudp->cud_format,
195 sizeof(struct xfs_cud_log_format));
199 * The CUD is either committed or aborted if the transaction is cancelled. If
200 * the transaction is cancelled, drop our reference to the CUI and free the
204 xfs_cud_item_release(
205 struct xfs_log_item *lip)
207 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
209 xfs_cui_release(cudp->cud_cuip);
210 kmem_free(cudp->cud_item.li_lv_shadow);
211 kmem_cache_free(xfs_cud_cache, cudp);
214 static struct xfs_log_item *
216 struct xfs_log_item *lip)
218 return &CUD_ITEM(lip)->cud_cuip->cui_item;
221 static const struct xfs_item_ops xfs_cud_item_ops = {
222 .flags = XFS_ITEM_RELEASE_WHEN_COMMITTED |
223 XFS_ITEM_INTENT_DONE,
224 .iop_size = xfs_cud_item_size,
225 .iop_format = xfs_cud_item_format,
226 .iop_release = xfs_cud_item_release,
227 .iop_intent = xfs_cud_item_intent,
230 static struct xfs_cud_log_item *
232 struct xfs_trans *tp,
233 struct xfs_cui_log_item *cuip)
235 struct xfs_cud_log_item *cudp;
237 cudp = kmem_cache_zalloc(xfs_cud_cache, GFP_KERNEL | __GFP_NOFAIL);
238 xfs_log_item_init(tp->t_mountp, &cudp->cud_item, XFS_LI_CUD,
240 cudp->cud_cuip = cuip;
241 cudp->cud_format.cud_cui_id = cuip->cui_format.cui_id;
243 xfs_trans_add_item(tp, &cudp->cud_item);
248 * Finish an refcount update and log it to the CUD. Note that the
249 * transaction is marked dirty regardless of whether the refcount
250 * update succeeds or fails to support the CUI/CUD lifecycle rules.
253 xfs_trans_log_finish_refcount_update(
254 struct xfs_trans *tp,
255 struct xfs_cud_log_item *cudp,
256 struct xfs_refcount_intent *ri,
257 struct xfs_btree_cur **pcur)
261 error = xfs_refcount_finish_one(tp, ri, pcur);
264 * Mark the transaction dirty, even on error. This ensures the
265 * transaction is aborted, which:
267 * 1.) releases the CUI and frees the CUD
268 * 2.) shuts down the filesystem
270 tp->t_flags |= XFS_TRANS_DIRTY | XFS_TRANS_HAS_INTENT_DONE;
271 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
276 /* Sort refcount intents by AG. */
278 xfs_refcount_update_diff_items(
280 const struct list_head *a,
281 const struct list_head *b)
283 struct xfs_refcount_intent *ra;
284 struct xfs_refcount_intent *rb;
286 ra = container_of(a, struct xfs_refcount_intent, ri_list);
287 rb = container_of(b, struct xfs_refcount_intent, ri_list);
289 return ra->ri_pag->pag_agno - rb->ri_pag->pag_agno;
292 /* Set the phys extent flags for this reverse mapping. */
294 xfs_trans_set_refcount_flags(
295 struct xfs_phys_extent *pmap,
296 enum xfs_refcount_intent_type type)
300 case XFS_REFCOUNT_INCREASE:
301 case XFS_REFCOUNT_DECREASE:
302 case XFS_REFCOUNT_ALLOC_COW:
303 case XFS_REFCOUNT_FREE_COW:
304 pmap->pe_flags |= type;
311 /* Log refcount updates in the intent item. */
313 xfs_refcount_update_log_item(
314 struct xfs_trans *tp,
315 struct xfs_cui_log_item *cuip,
316 struct xfs_refcount_intent *ri)
319 struct xfs_phys_extent *pmap;
321 tp->t_flags |= XFS_TRANS_DIRTY;
322 set_bit(XFS_LI_DIRTY, &cuip->cui_item.li_flags);
325 * atomic_inc_return gives us the value after the increment;
326 * we want to use it as an array index so we need to subtract 1 from
329 next_extent = atomic_inc_return(&cuip->cui_next_extent) - 1;
330 ASSERT(next_extent < cuip->cui_format.cui_nextents);
331 pmap = &cuip->cui_format.cui_extents[next_extent];
332 pmap->pe_startblock = ri->ri_startblock;
333 pmap->pe_len = ri->ri_blockcount;
334 xfs_trans_set_refcount_flags(pmap, ri->ri_type);
337 static struct xfs_log_item *
338 xfs_refcount_update_create_intent(
339 struct xfs_trans *tp,
340 struct list_head *items,
344 struct xfs_mount *mp = tp->t_mountp;
345 struct xfs_cui_log_item *cuip = xfs_cui_init(mp, count);
346 struct xfs_refcount_intent *ri;
350 xfs_trans_add_item(tp, &cuip->cui_item);
352 list_sort(mp, items, xfs_refcount_update_diff_items);
353 list_for_each_entry(ri, items, ri_list)
354 xfs_refcount_update_log_item(tp, cuip, ri);
355 return &cuip->cui_item;
358 /* Get an CUD so we can process all the deferred refcount updates. */
359 static struct xfs_log_item *
360 xfs_refcount_update_create_done(
361 struct xfs_trans *tp,
362 struct xfs_log_item *intent,
365 return &xfs_trans_get_cud(tp, CUI_ITEM(intent))->cud_item;
368 /* Take a passive ref to the AG containing the space we're refcounting. */
370 xfs_refcount_update_get_group(
371 struct xfs_mount *mp,
372 struct xfs_refcount_intent *ri)
376 agno = XFS_FSB_TO_AGNO(mp, ri->ri_startblock);
377 ri->ri_pag = xfs_perag_intent_get(mp, agno);
380 /* Release a passive AG ref after finishing refcounting work. */
382 xfs_refcount_update_put_group(
383 struct xfs_refcount_intent *ri)
385 xfs_perag_intent_put(ri->ri_pag);
388 /* Process a deferred refcount update. */
390 xfs_refcount_update_finish_item(
391 struct xfs_trans *tp,
392 struct xfs_log_item *done,
393 struct list_head *item,
394 struct xfs_btree_cur **state)
396 struct xfs_refcount_intent *ri;
399 ri = container_of(item, struct xfs_refcount_intent, ri_list);
400 error = xfs_trans_log_finish_refcount_update(tp, CUD_ITEM(done), ri,
403 /* Did we run out of reservation? Requeue what we didn't finish. */
404 if (!error && ri->ri_blockcount > 0) {
405 ASSERT(ri->ri_type == XFS_REFCOUNT_INCREASE ||
406 ri->ri_type == XFS_REFCOUNT_DECREASE);
410 xfs_refcount_update_put_group(ri);
411 kmem_cache_free(xfs_refcount_intent_cache, ri);
415 /* Abort all pending CUIs. */
417 xfs_refcount_update_abort_intent(
418 struct xfs_log_item *intent)
420 xfs_cui_release(CUI_ITEM(intent));
423 /* Cancel a deferred refcount update. */
425 xfs_refcount_update_cancel_item(
426 struct list_head *item)
428 struct xfs_refcount_intent *ri;
430 ri = container_of(item, struct xfs_refcount_intent, ri_list);
432 xfs_refcount_update_put_group(ri);
433 kmem_cache_free(xfs_refcount_intent_cache, ri);
436 const struct xfs_defer_op_type xfs_refcount_update_defer_type = {
437 .max_items = XFS_CUI_MAX_FAST_EXTENTS,
438 .create_intent = xfs_refcount_update_create_intent,
439 .abort_intent = xfs_refcount_update_abort_intent,
440 .create_done = xfs_refcount_update_create_done,
441 .finish_item = xfs_refcount_update_finish_item,
442 .finish_cleanup = xfs_refcount_finish_one_cleanup,
443 .cancel_item = xfs_refcount_update_cancel_item,
446 /* Is this recovered CUI ok? */
448 xfs_cui_validate_phys(
449 struct xfs_mount *mp,
450 struct xfs_phys_extent *pmap)
452 if (!xfs_has_reflink(mp))
455 if (pmap->pe_flags & ~XFS_REFCOUNT_EXTENT_FLAGS)
458 switch (pmap->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK) {
459 case XFS_REFCOUNT_INCREASE:
460 case XFS_REFCOUNT_DECREASE:
461 case XFS_REFCOUNT_ALLOC_COW:
462 case XFS_REFCOUNT_FREE_COW:
468 return xfs_verify_fsbext(mp, pmap->pe_startblock, pmap->pe_len);
472 * Process a refcount update intent item that was recovered from the log.
473 * We need to update the refcountbt.
476 xfs_cui_item_recover(
477 struct xfs_log_item *lip,
478 struct list_head *capture_list)
480 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
481 struct xfs_cud_log_item *cudp;
482 struct xfs_trans *tp;
483 struct xfs_btree_cur *rcur = NULL;
484 struct xfs_mount *mp = lip->li_log->l_mp;
485 unsigned int refc_type;
486 bool requeue_only = false;
491 * First check the validity of the extents described by the
492 * CUI. If any are bad, then assume that all are bad and
495 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
496 if (!xfs_cui_validate_phys(mp,
497 &cuip->cui_format.cui_extents[i])) {
498 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
500 sizeof(cuip->cui_format));
501 return -EFSCORRUPTED;
506 * Under normal operation, refcount updates are deferred, so we
507 * wouldn't be adding them directly to a transaction. All
508 * refcount updates manage reservation usage internally and
509 * dynamically by deferring work that won't fit in the
510 * transaction. Normally, any work that needs to be deferred
511 * gets attached to the same defer_ops that scheduled the
512 * refcount update. However, we're in log recovery here, so we
513 * use the passed in defer_ops and to finish up any work that
514 * doesn't fit. We need to reserve enough blocks to handle a
515 * full btree split on either end of the refcount range.
517 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate,
518 mp->m_refc_maxlevels * 2, 0, XFS_TRANS_RESERVE, &tp);
522 cudp = xfs_trans_get_cud(tp, cuip);
524 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
525 struct xfs_refcount_intent fake = { };
526 struct xfs_phys_extent *pmap;
528 pmap = &cuip->cui_format.cui_extents[i];
529 refc_type = pmap->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK;
531 case XFS_REFCOUNT_INCREASE:
532 case XFS_REFCOUNT_DECREASE:
533 case XFS_REFCOUNT_ALLOC_COW:
534 case XFS_REFCOUNT_FREE_COW:
535 fake.ri_type = refc_type;
538 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
540 sizeof(cuip->cui_format));
541 error = -EFSCORRUPTED;
545 fake.ri_startblock = pmap->pe_startblock;
546 fake.ri_blockcount = pmap->pe_len;
549 xfs_refcount_update_get_group(mp, &fake);
550 error = xfs_trans_log_finish_refcount_update(tp, cudp,
552 xfs_refcount_update_put_group(&fake);
554 if (error == -EFSCORRUPTED)
555 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
557 sizeof(cuip->cui_format));
561 /* Requeue what we didn't finish. */
562 if (fake.ri_blockcount > 0) {
563 struct xfs_bmbt_irec irec = {
564 .br_startblock = fake.ri_startblock,
565 .br_blockcount = fake.ri_blockcount,
568 switch (fake.ri_type) {
569 case XFS_REFCOUNT_INCREASE:
570 xfs_refcount_increase_extent(tp, &irec);
572 case XFS_REFCOUNT_DECREASE:
573 xfs_refcount_decrease_extent(tp, &irec);
575 case XFS_REFCOUNT_ALLOC_COW:
576 xfs_refcount_alloc_cow_extent(tp,
580 case XFS_REFCOUNT_FREE_COW:
581 xfs_refcount_free_cow_extent(tp,
592 xfs_refcount_finish_one_cleanup(tp, rcur, error);
593 return xfs_defer_ops_capture_and_commit(tp, capture_list);
596 xfs_refcount_finish_one_cleanup(tp, rcur, error);
597 xfs_trans_cancel(tp);
603 struct xfs_log_item *lip,
606 return CUI_ITEM(lip)->cui_format.cui_id == intent_id;
609 /* Relog an intent item to push the log tail forward. */
610 static struct xfs_log_item *
612 struct xfs_log_item *intent,
613 struct xfs_trans *tp)
615 struct xfs_cud_log_item *cudp;
616 struct xfs_cui_log_item *cuip;
617 struct xfs_phys_extent *pmap;
620 count = CUI_ITEM(intent)->cui_format.cui_nextents;
621 pmap = CUI_ITEM(intent)->cui_format.cui_extents;
623 tp->t_flags |= XFS_TRANS_DIRTY;
624 cudp = xfs_trans_get_cud(tp, CUI_ITEM(intent));
625 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
627 cuip = xfs_cui_init(tp->t_mountp, count);
628 memcpy(cuip->cui_format.cui_extents, pmap, count * sizeof(*pmap));
629 atomic_set(&cuip->cui_next_extent, count);
630 xfs_trans_add_item(tp, &cuip->cui_item);
631 set_bit(XFS_LI_DIRTY, &cuip->cui_item.li_flags);
632 return &cuip->cui_item;
635 static const struct xfs_item_ops xfs_cui_item_ops = {
636 .flags = XFS_ITEM_INTENT,
637 .iop_size = xfs_cui_item_size,
638 .iop_format = xfs_cui_item_format,
639 .iop_unpin = xfs_cui_item_unpin,
640 .iop_release = xfs_cui_item_release,
641 .iop_recover = xfs_cui_item_recover,
642 .iop_match = xfs_cui_item_match,
643 .iop_relog = xfs_cui_item_relog,
648 struct xfs_cui_log_format *dst,
649 const struct xfs_cui_log_format *src)
653 memcpy(dst, src, offsetof(struct xfs_cui_log_format, cui_extents));
655 for (i = 0; i < src->cui_nextents; i++)
656 memcpy(&dst->cui_extents[i], &src->cui_extents[i],
657 sizeof(struct xfs_phys_extent));
661 * This routine is called to create an in-core extent refcount update
662 * item from the cui format structure which was logged on disk.
663 * It allocates an in-core cui, copies the extents from the format
664 * structure into it, and adds the cui to the AIL with the given
668 xlog_recover_cui_commit_pass2(
670 struct list_head *buffer_list,
671 struct xlog_recover_item *item,
674 struct xfs_mount *mp = log->l_mp;
675 struct xfs_cui_log_item *cuip;
676 struct xfs_cui_log_format *cui_formatp;
679 cui_formatp = item->ri_buf[0].i_addr;
681 if (item->ri_buf[0].i_len < xfs_cui_log_format_sizeof(0)) {
682 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
683 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
684 return -EFSCORRUPTED;
687 len = xfs_cui_log_format_sizeof(cui_formatp->cui_nextents);
688 if (item->ri_buf[0].i_len != len) {
689 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
690 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
691 return -EFSCORRUPTED;
694 cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
695 xfs_cui_copy_format(&cuip->cui_format, cui_formatp);
696 atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
698 * Insert the intent into the AIL directly and drop one reference so
699 * that finishing or canceling the work will drop the other.
701 xfs_trans_ail_insert(log->l_ailp, &cuip->cui_item, lsn);
702 xfs_cui_release(cuip);
706 const struct xlog_recover_item_ops xlog_cui_item_ops = {
707 .item_type = XFS_LI_CUI,
708 .commit_pass2 = xlog_recover_cui_commit_pass2,
712 * This routine is called when an CUD format structure is found in a committed
713 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
714 * was still in the log. To do this it searches the AIL for the CUI with an id
715 * equal to that in the CUD format structure. If we find it we drop the CUD
716 * reference, which removes the CUI from the AIL and frees it.
719 xlog_recover_cud_commit_pass2(
721 struct list_head *buffer_list,
722 struct xlog_recover_item *item,
725 struct xfs_cud_log_format *cud_formatp;
727 cud_formatp = item->ri_buf[0].i_addr;
728 if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format)) {
729 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, log->l_mp,
730 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
731 return -EFSCORRUPTED;
734 xlog_recover_release_intent(log, XFS_LI_CUI, cud_formatp->cud_cui_id);
738 const struct xlog_recover_item_ops xlog_cud_item_ops = {
739 .item_type = XFS_LI_CUD,
740 .commit_pass2 = xlog_recover_cud_commit_pass2,