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 enum xfs_refcount_intent_type type,
256 xfs_fsblock_t startblock,
257 xfs_extlen_t blockcount,
258 xfs_fsblock_t *new_fsb,
259 xfs_extlen_t *new_len,
260 struct xfs_btree_cur **pcur)
264 error = xfs_refcount_finish_one(tp, type, startblock,
265 blockcount, new_fsb, new_len, pcur);
268 * Mark the transaction dirty, even on error. This ensures the
269 * transaction is aborted, which:
271 * 1.) releases the CUI and frees the CUD
272 * 2.) shuts down the filesystem
274 tp->t_flags |= XFS_TRANS_DIRTY | XFS_TRANS_HAS_INTENT_DONE;
275 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
280 /* Sort refcount intents by AG. */
282 xfs_refcount_update_diff_items(
284 const struct list_head *a,
285 const struct list_head *b)
287 struct xfs_mount *mp = priv;
288 struct xfs_refcount_intent *ra;
289 struct xfs_refcount_intent *rb;
291 ra = container_of(a, struct xfs_refcount_intent, ri_list);
292 rb = container_of(b, struct xfs_refcount_intent, ri_list);
293 return XFS_FSB_TO_AGNO(mp, ra->ri_startblock) -
294 XFS_FSB_TO_AGNO(mp, rb->ri_startblock);
297 /* Set the phys extent flags for this reverse mapping. */
299 xfs_trans_set_refcount_flags(
300 struct xfs_phys_extent *refc,
301 enum xfs_refcount_intent_type type)
305 case XFS_REFCOUNT_INCREASE:
306 case XFS_REFCOUNT_DECREASE:
307 case XFS_REFCOUNT_ALLOC_COW:
308 case XFS_REFCOUNT_FREE_COW:
309 refc->pe_flags |= type;
316 /* Log refcount updates in the intent item. */
318 xfs_refcount_update_log_item(
319 struct xfs_trans *tp,
320 struct xfs_cui_log_item *cuip,
321 struct xfs_refcount_intent *refc)
324 struct xfs_phys_extent *ext;
326 tp->t_flags |= XFS_TRANS_DIRTY;
327 set_bit(XFS_LI_DIRTY, &cuip->cui_item.li_flags);
330 * atomic_inc_return gives us the value after the increment;
331 * we want to use it as an array index so we need to subtract 1 from
334 next_extent = atomic_inc_return(&cuip->cui_next_extent) - 1;
335 ASSERT(next_extent < cuip->cui_format.cui_nextents);
336 ext = &cuip->cui_format.cui_extents[next_extent];
337 ext->pe_startblock = refc->ri_startblock;
338 ext->pe_len = refc->ri_blockcount;
339 xfs_trans_set_refcount_flags(ext, refc->ri_type);
342 static struct xfs_log_item *
343 xfs_refcount_update_create_intent(
344 struct xfs_trans *tp,
345 struct list_head *items,
349 struct xfs_mount *mp = tp->t_mountp;
350 struct xfs_cui_log_item *cuip = xfs_cui_init(mp, count);
351 struct xfs_refcount_intent *refc;
355 xfs_trans_add_item(tp, &cuip->cui_item);
357 list_sort(mp, items, xfs_refcount_update_diff_items);
358 list_for_each_entry(refc, items, ri_list)
359 xfs_refcount_update_log_item(tp, cuip, refc);
360 return &cuip->cui_item;
363 /* Get an CUD so we can process all the deferred refcount updates. */
364 static struct xfs_log_item *
365 xfs_refcount_update_create_done(
366 struct xfs_trans *tp,
367 struct xfs_log_item *intent,
370 return &xfs_trans_get_cud(tp, CUI_ITEM(intent))->cud_item;
373 /* Process a deferred refcount update. */
375 xfs_refcount_update_finish_item(
376 struct xfs_trans *tp,
377 struct xfs_log_item *done,
378 struct list_head *item,
379 struct xfs_btree_cur **state)
381 struct xfs_refcount_intent *refc;
382 xfs_fsblock_t new_fsb;
383 xfs_extlen_t new_aglen;
386 refc = container_of(item, struct xfs_refcount_intent, ri_list);
387 error = xfs_trans_log_finish_refcount_update(tp, CUD_ITEM(done),
388 refc->ri_type, refc->ri_startblock, refc->ri_blockcount,
389 &new_fsb, &new_aglen, state);
391 /* Did we run out of reservation? Requeue what we didn't finish. */
392 if (!error && new_aglen > 0) {
393 ASSERT(refc->ri_type == XFS_REFCOUNT_INCREASE ||
394 refc->ri_type == XFS_REFCOUNT_DECREASE);
395 refc->ri_startblock = new_fsb;
396 refc->ri_blockcount = new_aglen;
399 kmem_cache_free(xfs_refcount_intent_cache, refc);
403 /* Abort all pending CUIs. */
405 xfs_refcount_update_abort_intent(
406 struct xfs_log_item *intent)
408 xfs_cui_release(CUI_ITEM(intent));
411 /* Cancel a deferred refcount update. */
413 xfs_refcount_update_cancel_item(
414 struct list_head *item)
416 struct xfs_refcount_intent *refc;
418 refc = container_of(item, struct xfs_refcount_intent, ri_list);
419 kmem_cache_free(xfs_refcount_intent_cache, refc);
422 const struct xfs_defer_op_type xfs_refcount_update_defer_type = {
423 .max_items = XFS_CUI_MAX_FAST_EXTENTS,
424 .create_intent = xfs_refcount_update_create_intent,
425 .abort_intent = xfs_refcount_update_abort_intent,
426 .create_done = xfs_refcount_update_create_done,
427 .finish_item = xfs_refcount_update_finish_item,
428 .finish_cleanup = xfs_refcount_finish_one_cleanup,
429 .cancel_item = xfs_refcount_update_cancel_item,
432 /* Is this recovered CUI ok? */
434 xfs_cui_validate_phys(
435 struct xfs_mount *mp,
436 struct xfs_phys_extent *refc)
438 if (!xfs_has_reflink(mp))
441 if (refc->pe_flags & ~XFS_REFCOUNT_EXTENT_FLAGS)
444 switch (refc->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK) {
445 case XFS_REFCOUNT_INCREASE:
446 case XFS_REFCOUNT_DECREASE:
447 case XFS_REFCOUNT_ALLOC_COW:
448 case XFS_REFCOUNT_FREE_COW:
454 return xfs_verify_fsbext(mp, refc->pe_startblock, refc->pe_len);
458 * Process a refcount update intent item that was recovered from the log.
459 * We need to update the refcountbt.
462 xfs_cui_item_recover(
463 struct xfs_log_item *lip,
464 struct list_head *capture_list)
466 struct xfs_bmbt_irec irec;
467 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
468 struct xfs_phys_extent *refc;
469 struct xfs_cud_log_item *cudp;
470 struct xfs_trans *tp;
471 struct xfs_btree_cur *rcur = NULL;
472 struct xfs_mount *mp = lip->li_log->l_mp;
473 xfs_fsblock_t new_fsb;
474 xfs_extlen_t new_len;
475 unsigned int refc_type;
476 bool requeue_only = false;
477 enum xfs_refcount_intent_type type;
482 * First check the validity of the extents described by the
483 * CUI. If any are bad, then assume that all are bad and
486 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
487 if (!xfs_cui_validate_phys(mp,
488 &cuip->cui_format.cui_extents[i])) {
489 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
491 sizeof(cuip->cui_format));
492 return -EFSCORRUPTED;
497 * Under normal operation, refcount updates are deferred, so we
498 * wouldn't be adding them directly to a transaction. All
499 * refcount updates manage reservation usage internally and
500 * dynamically by deferring work that won't fit in the
501 * transaction. Normally, any work that needs to be deferred
502 * gets attached to the same defer_ops that scheduled the
503 * refcount update. However, we're in log recovery here, so we
504 * use the passed in defer_ops and to finish up any work that
505 * doesn't fit. We need to reserve enough blocks to handle a
506 * full btree split on either end of the refcount range.
508 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate,
509 mp->m_refc_maxlevels * 2, 0, XFS_TRANS_RESERVE, &tp);
513 cudp = xfs_trans_get_cud(tp, cuip);
515 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
516 refc = &cuip->cui_format.cui_extents[i];
517 refc_type = refc->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK;
519 case XFS_REFCOUNT_INCREASE:
520 case XFS_REFCOUNT_DECREASE:
521 case XFS_REFCOUNT_ALLOC_COW:
522 case XFS_REFCOUNT_FREE_COW:
526 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, mp);
527 error = -EFSCORRUPTED;
531 new_fsb = refc->pe_startblock;
532 new_len = refc->pe_len;
534 error = xfs_trans_log_finish_refcount_update(tp, cudp,
535 type, refc->pe_startblock, refc->pe_len,
536 &new_fsb, &new_len, &rcur);
537 if (error == -EFSCORRUPTED)
538 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
539 refc, sizeof(*refc));
543 /* Requeue what we didn't finish. */
545 irec.br_startblock = new_fsb;
546 irec.br_blockcount = new_len;
548 case XFS_REFCOUNT_INCREASE:
549 xfs_refcount_increase_extent(tp, &irec);
551 case XFS_REFCOUNT_DECREASE:
552 xfs_refcount_decrease_extent(tp, &irec);
554 case XFS_REFCOUNT_ALLOC_COW:
555 xfs_refcount_alloc_cow_extent(tp,
559 case XFS_REFCOUNT_FREE_COW:
560 xfs_refcount_free_cow_extent(tp,
571 xfs_refcount_finish_one_cleanup(tp, rcur, error);
572 return xfs_defer_ops_capture_and_commit(tp, capture_list);
575 xfs_refcount_finish_one_cleanup(tp, rcur, error);
576 xfs_trans_cancel(tp);
582 struct xfs_log_item *lip,
585 return CUI_ITEM(lip)->cui_format.cui_id == intent_id;
588 /* Relog an intent item to push the log tail forward. */
589 static struct xfs_log_item *
591 struct xfs_log_item *intent,
592 struct xfs_trans *tp)
594 struct xfs_cud_log_item *cudp;
595 struct xfs_cui_log_item *cuip;
596 struct xfs_phys_extent *extp;
599 count = CUI_ITEM(intent)->cui_format.cui_nextents;
600 extp = CUI_ITEM(intent)->cui_format.cui_extents;
602 tp->t_flags |= XFS_TRANS_DIRTY;
603 cudp = xfs_trans_get_cud(tp, CUI_ITEM(intent));
604 set_bit(XFS_LI_DIRTY, &cudp->cud_item.li_flags);
606 cuip = xfs_cui_init(tp->t_mountp, count);
607 memcpy(cuip->cui_format.cui_extents, extp, count * sizeof(*extp));
608 atomic_set(&cuip->cui_next_extent, count);
609 xfs_trans_add_item(tp, &cuip->cui_item);
610 set_bit(XFS_LI_DIRTY, &cuip->cui_item.li_flags);
611 return &cuip->cui_item;
614 static const struct xfs_item_ops xfs_cui_item_ops = {
615 .flags = XFS_ITEM_INTENT,
616 .iop_size = xfs_cui_item_size,
617 .iop_format = xfs_cui_item_format,
618 .iop_unpin = xfs_cui_item_unpin,
619 .iop_release = xfs_cui_item_release,
620 .iop_recover = xfs_cui_item_recover,
621 .iop_match = xfs_cui_item_match,
622 .iop_relog = xfs_cui_item_relog,
626 * Copy an CUI format buffer from the given buf, and into the destination
627 * CUI format structure. The CUI/CUD items were designed not to need any
628 * special alignment handling.
632 struct xfs_log_iovec *buf,
633 struct xfs_cui_log_format *dst_cui_fmt)
635 struct xfs_cui_log_format *src_cui_fmt;
638 src_cui_fmt = buf->i_addr;
639 len = xfs_cui_log_format_sizeof(src_cui_fmt->cui_nextents);
641 if (buf->i_len == len) {
642 memcpy(dst_cui_fmt, src_cui_fmt, len);
645 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, NULL);
646 return -EFSCORRUPTED;
650 * This routine is called to create an in-core extent refcount update
651 * item from the cui format structure which was logged on disk.
652 * It allocates an in-core cui, copies the extents from the format
653 * structure into it, and adds the cui to the AIL with the given
657 xlog_recover_cui_commit_pass2(
659 struct list_head *buffer_list,
660 struct xlog_recover_item *item,
664 struct xfs_mount *mp = log->l_mp;
665 struct xfs_cui_log_item *cuip;
666 struct xfs_cui_log_format *cui_formatp;
668 cui_formatp = item->ri_buf[0].i_addr;
670 cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
671 error = xfs_cui_copy_format(&item->ri_buf[0], &cuip->cui_format);
673 xfs_cui_item_free(cuip);
676 atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
678 * Insert the intent into the AIL directly and drop one reference so
679 * that finishing or canceling the work will drop the other.
681 xfs_trans_ail_insert(log->l_ailp, &cuip->cui_item, lsn);
682 xfs_cui_release(cuip);
686 const struct xlog_recover_item_ops xlog_cui_item_ops = {
687 .item_type = XFS_LI_CUI,
688 .commit_pass2 = xlog_recover_cui_commit_pass2,
692 * This routine is called when an CUD format structure is found in a committed
693 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
694 * was still in the log. To do this it searches the AIL for the CUI with an id
695 * equal to that in the CUD format structure. If we find it we drop the CUD
696 * reference, which removes the CUI from the AIL and frees it.
699 xlog_recover_cud_commit_pass2(
701 struct list_head *buffer_list,
702 struct xlog_recover_item *item,
705 struct xfs_cud_log_format *cud_formatp;
707 cud_formatp = item->ri_buf[0].i_addr;
708 if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format)) {
709 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
710 return -EFSCORRUPTED;
713 xlog_recover_release_intent(log, XFS_LI_CUI, cud_formatp->cud_cui_id);
717 const struct xlog_recover_item_ops xlog_cud_item_ops = {
718 .item_type = XFS_LI_CUD,
719 .commit_pass2 = xlog_recover_cud_commit_pass2,