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 kmem_free(cuip->cui_item.li_lv_shadow);
39 if (cuip->cui_format.cui_nextents > XFS_CUI_MAX_FAST_EXTENTS)
42 kmem_cache_free(xfs_cui_cache, cuip);
46 * Freeing the CUI requires that we remove it from the AIL if it has already
47 * been placed there. However, the CUI may not yet have been placed in the AIL
48 * when called by xfs_cui_release() from CUD processing due to the ordering of
49 * committed vs unpin operations in bulk insert operations. Hence the reference
50 * count to ensure only the last caller frees the CUI.
54 struct xfs_cui_log_item *cuip)
56 ASSERT(atomic_read(&cuip->cui_refcount) > 0);
57 if (!atomic_dec_and_test(&cuip->cui_refcount))
60 xfs_trans_ail_delete(&cuip->cui_item, 0);
61 xfs_cui_item_free(cuip);
67 struct xfs_log_item *lip,
71 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
74 *nbytes += xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents);
78 * This is called to fill in the vector of log iovecs for the
79 * given cui log item. We use only 1 iovec, and we point that
80 * at the cui_log_format structure embedded in the cui item.
81 * It is at this point that we assert that all of the extent
82 * slots in the cui item have been filled.
86 struct xfs_log_item *lip,
87 struct xfs_log_vec *lv)
89 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
90 struct xfs_log_iovec *vecp = NULL;
92 ASSERT(atomic_read(&cuip->cui_next_extent) ==
93 cuip->cui_format.cui_nextents);
95 cuip->cui_format.cui_type = XFS_LI_CUI;
96 cuip->cui_format.cui_size = 1;
98 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUI_FORMAT, &cuip->cui_format,
99 xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents));
103 * The unpin operation is the last place an CUI is manipulated in the log. It is
104 * either inserted in the AIL or aborted in the event of a log I/O error. In
105 * either case, the CUI transaction has been successfully committed to make it
106 * this far. Therefore, we expect whoever committed the CUI to either construct
107 * and commit the CUD or drop the CUD's reference in the event of error. Simply
108 * drop the log's CUI reference now that the log is done with it.
112 struct xfs_log_item *lip,
115 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
117 xfs_cui_release(cuip);
121 * The CUI has been either committed or aborted if the transaction has been
122 * cancelled. If the transaction was cancelled, an CUD isn't going to be
123 * constructed and thus we free the CUI here directly.
126 xfs_cui_item_release(
127 struct xfs_log_item *lip)
129 xfs_cui_release(CUI_ITEM(lip));
133 * Allocate and initialize an cui item with the given number of extents.
135 STATIC struct xfs_cui_log_item *
137 struct xfs_mount *mp,
141 struct xfs_cui_log_item *cuip;
143 ASSERT(nextents > 0);
144 if (nextents > XFS_CUI_MAX_FAST_EXTENTS)
145 cuip = kmem_zalloc(xfs_cui_log_item_sizeof(nextents),
148 cuip = kmem_cache_zalloc(xfs_cui_cache,
149 GFP_KERNEL | __GFP_NOFAIL);
151 xfs_log_item_init(mp, &cuip->cui_item, XFS_LI_CUI, &xfs_cui_item_ops);
152 cuip->cui_format.cui_nextents = nextents;
153 cuip->cui_format.cui_id = (uintptr_t)(void *)cuip;
154 atomic_set(&cuip->cui_next_extent, 0);
155 atomic_set(&cuip->cui_refcount, 2);
160 static inline struct xfs_cud_log_item *CUD_ITEM(struct xfs_log_item *lip)
162 return container_of(lip, struct xfs_cud_log_item, cud_item);
167 struct xfs_log_item *lip,
172 *nbytes += sizeof(struct xfs_cud_log_format);
176 * This is called to fill in the vector of log iovecs for the
177 * given cud log item. We use only 1 iovec, and we point that
178 * at the cud_log_format structure embedded in the cud item.
179 * It is at this point that we assert that all of the extent
180 * slots in the cud item have been filled.
184 struct xfs_log_item *lip,
185 struct xfs_log_vec *lv)
187 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
188 struct xfs_log_iovec *vecp = NULL;
190 cudp->cud_format.cud_type = XFS_LI_CUD;
191 cudp->cud_format.cud_size = 1;
193 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUD_FORMAT, &cudp->cud_format,
194 sizeof(struct xfs_cud_log_format));
198 * The CUD is either committed or aborted if the transaction is cancelled. If
199 * the transaction is cancelled, drop our reference to the CUI and free the
203 xfs_cud_item_release(
204 struct xfs_log_item *lip)
206 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
208 xfs_cui_release(cudp->cud_cuip);
209 kmem_free(cudp->cud_item.li_lv_shadow);
210 kmem_cache_free(xfs_cud_cache, cudp);
213 static struct xfs_log_item *
215 struct xfs_log_item *lip)
217 return &CUD_ITEM(lip)->cud_cuip->cui_item;
220 static const struct xfs_item_ops xfs_cud_item_ops = {
221 .flags = XFS_ITEM_RELEASE_WHEN_COMMITTED |
222 XFS_ITEM_INTENT_DONE,
223 .iop_size = xfs_cud_item_size,
224 .iop_format = xfs_cud_item_format,
225 .iop_release = xfs_cud_item_release,
226 .iop_intent = xfs_cud_item_intent,
229 static struct xfs_cud_log_item *
231 struct xfs_trans *tp,
232 struct xfs_cui_log_item *cuip)
234 struct xfs_cud_log_item *cudp;
236 cudp = kmem_cache_zalloc(xfs_cud_cache, GFP_KERNEL | __GFP_NOFAIL);
237 xfs_log_item_init(tp->t_mountp, &cudp->cud_item, XFS_LI_CUD,
239 cudp->cud_cuip = cuip;
240 cudp->cud_format.cud_cui_id = cuip->cui_format.cui_id;
242 xfs_trans_add_item(tp, &cudp->cud_item);
247 * Finish an refcount update and log it to the CUD. Note that the
248 * transaction is marked dirty regardless of whether the refcount
249 * update succeeds or fails to support the CUI/CUD lifecycle rules.
252 xfs_trans_log_finish_refcount_update(
253 struct xfs_trans *tp,
254 struct xfs_cud_log_item *cudp,
255 struct xfs_refcount_intent *ri,
256 struct xfs_btree_cur **pcur)
260 error = xfs_refcount_finish_one(tp, ri, pcur);
263 * Mark the transaction dirty, even on error. This ensures the
264 * transaction is aborted, which:
266 * 1.) releases the CUI and frees the CUD
267 * 2.) shuts down the filesystem
269 tp->t_flags |= XFS_TRANS_DIRTY | XFS_TRANS_HAS_INTENT_DONE;
270 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
275 /* Sort refcount intents by AG. */
277 xfs_refcount_update_diff_items(
279 const struct list_head *a,
280 const struct list_head *b)
282 struct xfs_mount *mp = priv;
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);
288 return XFS_FSB_TO_AGNO(mp, ra->ri_startblock) -
289 XFS_FSB_TO_AGNO(mp, rb->ri_startblock);
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 /* Process a deferred refcount update. */
370 xfs_refcount_update_finish_item(
371 struct xfs_trans *tp,
372 struct xfs_log_item *done,
373 struct list_head *item,
374 struct xfs_btree_cur **state)
376 struct xfs_refcount_intent *ri;
379 ri = container_of(item, struct xfs_refcount_intent, ri_list);
380 error = xfs_trans_log_finish_refcount_update(tp, CUD_ITEM(done), ri,
383 /* Did we run out of reservation? Requeue what we didn't finish. */
384 if (!error && ri->ri_blockcount > 0) {
385 ASSERT(ri->ri_type == XFS_REFCOUNT_INCREASE ||
386 ri->ri_type == XFS_REFCOUNT_DECREASE);
389 kmem_cache_free(xfs_refcount_intent_cache, ri);
393 /* Abort all pending CUIs. */
395 xfs_refcount_update_abort_intent(
396 struct xfs_log_item *intent)
398 xfs_cui_release(CUI_ITEM(intent));
401 /* Cancel a deferred refcount update. */
403 xfs_refcount_update_cancel_item(
404 struct list_head *item)
406 struct xfs_refcount_intent *ri;
408 ri = container_of(item, struct xfs_refcount_intent, ri_list);
409 kmem_cache_free(xfs_refcount_intent_cache, ri);
412 const struct xfs_defer_op_type xfs_refcount_update_defer_type = {
413 .max_items = XFS_CUI_MAX_FAST_EXTENTS,
414 .create_intent = xfs_refcount_update_create_intent,
415 .abort_intent = xfs_refcount_update_abort_intent,
416 .create_done = xfs_refcount_update_create_done,
417 .finish_item = xfs_refcount_update_finish_item,
418 .finish_cleanup = xfs_refcount_finish_one_cleanup,
419 .cancel_item = xfs_refcount_update_cancel_item,
422 /* Is this recovered CUI ok? */
424 xfs_cui_validate_phys(
425 struct xfs_mount *mp,
426 struct xfs_phys_extent *pmap)
428 if (!xfs_has_reflink(mp))
431 if (pmap->pe_flags & ~XFS_REFCOUNT_EXTENT_FLAGS)
434 switch (pmap->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK) {
435 case XFS_REFCOUNT_INCREASE:
436 case XFS_REFCOUNT_DECREASE:
437 case XFS_REFCOUNT_ALLOC_COW:
438 case XFS_REFCOUNT_FREE_COW:
444 return xfs_verify_fsbext(mp, pmap->pe_startblock, pmap->pe_len);
448 * Process a refcount update intent item that was recovered from the log.
449 * We need to update the refcountbt.
452 xfs_cui_item_recover(
453 struct xfs_log_item *lip,
454 struct list_head *capture_list)
456 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
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 unsigned int refc_type;
462 bool requeue_only = false;
467 * First check the validity of the extents described by the
468 * CUI. If any are bad, then assume that all are bad and
471 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
472 if (!xfs_cui_validate_phys(mp,
473 &cuip->cui_format.cui_extents[i])) {
474 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
476 sizeof(cuip->cui_format));
477 return -EFSCORRUPTED;
482 * Under normal operation, refcount updates are deferred, so we
483 * wouldn't be adding them directly to a transaction. All
484 * refcount updates manage reservation usage internally and
485 * dynamically by deferring work that won't fit in the
486 * transaction. Normally, any work that needs to be deferred
487 * gets attached to the same defer_ops that scheduled the
488 * refcount update. However, we're in log recovery here, so we
489 * use the passed in defer_ops and to finish up any work that
490 * doesn't fit. We need to reserve enough blocks to handle a
491 * full btree split on either end of the refcount range.
493 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate,
494 mp->m_refc_maxlevels * 2, 0, XFS_TRANS_RESERVE, &tp);
498 cudp = xfs_trans_get_cud(tp, cuip);
500 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
501 struct xfs_refcount_intent fake = { };
502 struct xfs_phys_extent *pmap;
504 pmap = &cuip->cui_format.cui_extents[i];
505 refc_type = pmap->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:
511 fake.ri_type = refc_type;
514 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
516 sizeof(cuip->cui_format));
517 error = -EFSCORRUPTED;
521 fake.ri_startblock = pmap->pe_startblock;
522 fake.ri_blockcount = pmap->pe_len;
524 error = xfs_trans_log_finish_refcount_update(tp, cudp,
526 if (error == -EFSCORRUPTED)
527 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
529 sizeof(cuip->cui_format));
533 /* Requeue what we didn't finish. */
534 if (fake.ri_blockcount > 0) {
535 struct xfs_bmbt_irec irec = {
536 .br_startblock = fake.ri_startblock,
537 .br_blockcount = fake.ri_blockcount,
540 switch (fake.ri_type) {
541 case XFS_REFCOUNT_INCREASE:
542 xfs_refcount_increase_extent(tp, &irec);
544 case XFS_REFCOUNT_DECREASE:
545 xfs_refcount_decrease_extent(tp, &irec);
547 case XFS_REFCOUNT_ALLOC_COW:
548 xfs_refcount_alloc_cow_extent(tp,
552 case XFS_REFCOUNT_FREE_COW:
553 xfs_refcount_free_cow_extent(tp,
564 xfs_refcount_finish_one_cleanup(tp, rcur, error);
565 return xfs_defer_ops_capture_and_commit(tp, capture_list);
568 xfs_refcount_finish_one_cleanup(tp, rcur, error);
569 xfs_trans_cancel(tp);
575 struct xfs_log_item *lip,
578 return CUI_ITEM(lip)->cui_format.cui_id == intent_id;
581 /* Relog an intent item to push the log tail forward. */
582 static struct xfs_log_item *
584 struct xfs_log_item *intent,
585 struct xfs_trans *tp)
587 struct xfs_cud_log_item *cudp;
588 struct xfs_cui_log_item *cuip;
589 struct xfs_phys_extent *pmap;
592 count = CUI_ITEM(intent)->cui_format.cui_nextents;
593 pmap = CUI_ITEM(intent)->cui_format.cui_extents;
595 tp->t_flags |= XFS_TRANS_DIRTY;
596 cudp = xfs_trans_get_cud(tp, CUI_ITEM(intent));
597 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
599 cuip = xfs_cui_init(tp->t_mountp, count);
600 memcpy(cuip->cui_format.cui_extents, pmap, count * sizeof(*pmap));
601 atomic_set(&cuip->cui_next_extent, count);
602 xfs_trans_add_item(tp, &cuip->cui_item);
603 set_bit(XFS_LI_DIRTY, &cuip->cui_item.li_flags);
604 return &cuip->cui_item;
607 static const struct xfs_item_ops xfs_cui_item_ops = {
608 .flags = XFS_ITEM_INTENT,
609 .iop_size = xfs_cui_item_size,
610 .iop_format = xfs_cui_item_format,
611 .iop_unpin = xfs_cui_item_unpin,
612 .iop_release = xfs_cui_item_release,
613 .iop_recover = xfs_cui_item_recover,
614 .iop_match = xfs_cui_item_match,
615 .iop_relog = xfs_cui_item_relog,
620 struct xfs_cui_log_format *dst,
621 const struct xfs_cui_log_format *src)
625 memcpy(dst, src, offsetof(struct xfs_cui_log_format, cui_extents));
627 for (i = 0; i < src->cui_nextents; i++)
628 memcpy(&dst->cui_extents[i], &src->cui_extents[i],
629 sizeof(struct xfs_phys_extent));
633 * This routine is called to create an in-core extent refcount update
634 * item from the cui format structure which was logged on disk.
635 * It allocates an in-core cui, copies the extents from the format
636 * structure into it, and adds the cui to the AIL with the given
640 xlog_recover_cui_commit_pass2(
642 struct list_head *buffer_list,
643 struct xlog_recover_item *item,
646 struct xfs_mount *mp = log->l_mp;
647 struct xfs_cui_log_item *cuip;
648 struct xfs_cui_log_format *cui_formatp;
651 cui_formatp = item->ri_buf[0].i_addr;
653 if (item->ri_buf[0].i_len < xfs_cui_log_format_sizeof(0)) {
654 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
655 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
656 return -EFSCORRUPTED;
659 len = xfs_cui_log_format_sizeof(cui_formatp->cui_nextents);
660 if (item->ri_buf[0].i_len != len) {
661 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
662 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
663 return -EFSCORRUPTED;
666 cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
667 xfs_cui_copy_format(&cuip->cui_format, cui_formatp);
668 atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
670 * Insert the intent into the AIL directly and drop one reference so
671 * that finishing or canceling the work will drop the other.
673 xfs_trans_ail_insert(log->l_ailp, &cuip->cui_item, lsn);
674 xfs_cui_release(cuip);
678 const struct xlog_recover_item_ops xlog_cui_item_ops = {
679 .item_type = XFS_LI_CUI,
680 .commit_pass2 = xlog_recover_cui_commit_pass2,
684 * This routine is called when an CUD format structure is found in a committed
685 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
686 * was still in the log. To do this it searches the AIL for the CUI with an id
687 * equal to that in the CUD format structure. If we find it we drop the CUD
688 * reference, which removes the CUI from the AIL and frees it.
691 xlog_recover_cud_commit_pass2(
693 struct list_head *buffer_list,
694 struct xlog_recover_item *item,
697 struct xfs_cud_log_format *cud_formatp;
699 cud_formatp = item->ri_buf[0].i_addr;
700 if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format)) {
701 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, log->l_mp,
702 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
703 return -EFSCORRUPTED;
706 xlog_recover_release_intent(log, XFS_LI_CUI, cud_formatp->cud_cui_id);
710 const struct xlog_recover_item_ops xlog_cud_item_ops = {
711 .item_type = XFS_LI_CUD,
712 .commit_pass2 = xlog_recover_cud_commit_pass2,