1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Copyright (C) 2016 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_format.h"
9#include "xfs_log_format.h"
10#include "xfs_trans_resv.h"
11#include "xfs_bit.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"
18#include "xfs_log.h"
19#include "xfs_refcount.h"
20#include "xfs_error.h"
21#include "xfs_log_priv.h"
22#include "xfs_log_recover.h"
23#include "xfs_ag.h"
24
25struct kmem_cache *xfs_cui_cache;
26struct kmem_cache *xfs_cud_cache;
27
28static const struct xfs_item_ops xfs_cui_item_ops;
29
30static inline struct xfs_cui_log_item *CUI_ITEM(struct xfs_log_item *lip)
31{
32 return container_of(lip, struct xfs_cui_log_item, cui_item);
33}
34
35STATIC void
36xfs_cui_item_free(
37 struct xfs_cui_log_item *cuip)
38{
39 kvfree(addr: cuip->cui_item.li_lv_shadow);
40 if (cuip->cui_format.cui_nextents > XFS_CUI_MAX_FAST_EXTENTS)
41 kfree(objp: cuip);
42 else
43 kmem_cache_free(s: xfs_cui_cache, objp: cuip);
44}
45
46/*
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.
52 */
53STATIC void
54xfs_cui_release(
55 struct xfs_cui_log_item *cuip)
56{
57 ASSERT(atomic_read(&cuip->cui_refcount) > 0);
58 if (!atomic_dec_and_test(v: &cuip->cui_refcount))
59 return;
60
61 xfs_trans_ail_delete(lip: &cuip->cui_item, shutdown_type: 0);
62 xfs_cui_item_free(cuip);
63}
64
65
66STATIC void
67xfs_cui_item_size(
68 struct xfs_log_item *lip,
69 int *nvecs,
70 int *nbytes)
71{
72 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
73
74 *nvecs += 1;
75 *nbytes += xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents);
76}
77
78/*
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.
84 */
85STATIC void
86xfs_cui_item_format(
87 struct xfs_log_item *lip,
88 struct xfs_log_vec *lv)
89{
90 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
91 struct xfs_log_iovec *vecp = NULL;
92
93 ASSERT(atomic_read(&cuip->cui_next_extent) ==
94 cuip->cui_format.cui_nextents);
95
96 cuip->cui_format.cui_type = XFS_LI_CUI;
97 cuip->cui_format.cui_size = 1;
98
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));
101}
102
103/*
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.
110 */
111STATIC void
112xfs_cui_item_unpin(
113 struct xfs_log_item *lip,
114 int remove)
115{
116 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
117
118 xfs_cui_release(cuip);
119}
120
121/*
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.
125 */
126STATIC void
127xfs_cui_item_release(
128 struct xfs_log_item *lip)
129{
130 xfs_cui_release(cuip: CUI_ITEM(lip));
131}
132
133/*
134 * Allocate and initialize an cui item with the given number of extents.
135 */
136STATIC struct xfs_cui_log_item *
137xfs_cui_init(
138 struct xfs_mount *mp,
139 uint nextents)
140
141{
142 struct xfs_cui_log_item *cuip;
143
144 ASSERT(nextents > 0);
145 if (nextents > XFS_CUI_MAX_FAST_EXTENTS)
146 cuip = kzalloc(size: xfs_cui_log_item_sizeof(nr: nextents),
147 GFP_KERNEL | __GFP_NOFAIL);
148 else
149 cuip = kmem_cache_zalloc(k: xfs_cui_cache,
150 GFP_KERNEL | __GFP_NOFAIL);
151
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(v: &cuip->cui_next_extent, i: 0);
156 atomic_set(v: &cuip->cui_refcount, i: 2);
157
158 return cuip;
159}
160
161static inline struct xfs_cud_log_item *CUD_ITEM(struct xfs_log_item *lip)
162{
163 return container_of(lip, struct xfs_cud_log_item, cud_item);
164}
165
166STATIC void
167xfs_cud_item_size(
168 struct xfs_log_item *lip,
169 int *nvecs,
170 int *nbytes)
171{
172 *nvecs += 1;
173 *nbytes += sizeof(struct xfs_cud_log_format);
174}
175
176/*
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.
182 */
183STATIC void
184xfs_cud_item_format(
185 struct xfs_log_item *lip,
186 struct xfs_log_vec *lv)
187{
188 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
189 struct xfs_log_iovec *vecp = NULL;
190
191 cudp->cud_format.cud_type = XFS_LI_CUD;
192 cudp->cud_format.cud_size = 1;
193
194 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUD_FORMAT, &cudp->cud_format,
195 sizeof(struct xfs_cud_log_format));
196}
197
198/*
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
201 * CUD.
202 */
203STATIC void
204xfs_cud_item_release(
205 struct xfs_log_item *lip)
206{
207 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
208
209 xfs_cui_release(cuip: cudp->cud_cuip);
210 kvfree(addr: cudp->cud_item.li_lv_shadow);
211 kmem_cache_free(s: xfs_cud_cache, objp: cudp);
212}
213
214static struct xfs_log_item *
215xfs_cud_item_intent(
216 struct xfs_log_item *lip)
217{
218 return &CUD_ITEM(lip)->cud_cuip->cui_item;
219}
220
221static 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,
228};
229
230/* Sort refcount intents by AG. */
231static int
232xfs_refcount_update_diff_items(
233 void *priv,
234 const struct list_head *a,
235 const struct list_head *b)
236{
237 struct xfs_refcount_intent *ra;
238 struct xfs_refcount_intent *rb;
239
240 ra = container_of(a, struct xfs_refcount_intent, ri_list);
241 rb = container_of(b, struct xfs_refcount_intent, ri_list);
242
243 return ra->ri_pag->pag_agno - rb->ri_pag->pag_agno;
244}
245
246/* Set the phys extent flags for this reverse mapping. */
247static void
248xfs_trans_set_refcount_flags(
249 struct xfs_phys_extent *pmap,
250 enum xfs_refcount_intent_type type)
251{
252 pmap->pe_flags = 0;
253 switch (type) {
254 case XFS_REFCOUNT_INCREASE:
255 case XFS_REFCOUNT_DECREASE:
256 case XFS_REFCOUNT_ALLOC_COW:
257 case XFS_REFCOUNT_FREE_COW:
258 pmap->pe_flags |= type;
259 break;
260 default:
261 ASSERT(0);
262 }
263}
264
265/* Log refcount updates in the intent item. */
266STATIC void
267xfs_refcount_update_log_item(
268 struct xfs_trans *tp,
269 struct xfs_cui_log_item *cuip,
270 struct xfs_refcount_intent *ri)
271{
272 uint next_extent;
273 struct xfs_phys_extent *pmap;
274
275 /*
276 * atomic_inc_return gives us the value after the increment;
277 * we want to use it as an array index so we need to subtract 1 from
278 * it.
279 */
280 next_extent = atomic_inc_return(v: &cuip->cui_next_extent) - 1;
281 ASSERT(next_extent < cuip->cui_format.cui_nextents);
282 pmap = &cuip->cui_format.cui_extents[next_extent];
283 pmap->pe_startblock = ri->ri_startblock;
284 pmap->pe_len = ri->ri_blockcount;
285 xfs_trans_set_refcount_flags(pmap, type: ri->ri_type);
286}
287
288static struct xfs_log_item *
289xfs_refcount_update_create_intent(
290 struct xfs_trans *tp,
291 struct list_head *items,
292 unsigned int count,
293 bool sort)
294{
295 struct xfs_mount *mp = tp->t_mountp;
296 struct xfs_cui_log_item *cuip = xfs_cui_init(mp, nextents: count);
297 struct xfs_refcount_intent *ri;
298
299 ASSERT(count > 0);
300
301 if (sort)
302 list_sort(priv: mp, head: items, cmp: xfs_refcount_update_diff_items);
303 list_for_each_entry(ri, items, ri_list)
304 xfs_refcount_update_log_item(tp, cuip, ri);
305 return &cuip->cui_item;
306}
307
308/* Get an CUD so we can process all the deferred refcount updates. */
309static struct xfs_log_item *
310xfs_refcount_update_create_done(
311 struct xfs_trans *tp,
312 struct xfs_log_item *intent,
313 unsigned int count)
314{
315 struct xfs_cui_log_item *cuip = CUI_ITEM(lip: intent);
316 struct xfs_cud_log_item *cudp;
317
318 cudp = kmem_cache_zalloc(k: xfs_cud_cache, GFP_KERNEL | __GFP_NOFAIL);
319 xfs_log_item_init(tp->t_mountp, &cudp->cud_item, XFS_LI_CUD,
320 &xfs_cud_item_ops);
321 cudp->cud_cuip = cuip;
322 cudp->cud_format.cud_cui_id = cuip->cui_format.cui_id;
323
324 return &cudp->cud_item;
325}
326
327/* Take a passive ref to the AG containing the space we're refcounting. */
328void
329xfs_refcount_update_get_group(
330 struct xfs_mount *mp,
331 struct xfs_refcount_intent *ri)
332{
333 xfs_agnumber_t agno;
334
335 agno = XFS_FSB_TO_AGNO(mp, ri->ri_startblock);
336 ri->ri_pag = xfs_perag_intent_get(mp, agno);
337}
338
339/* Release a passive AG ref after finishing refcounting work. */
340static inline void
341xfs_refcount_update_put_group(
342 struct xfs_refcount_intent *ri)
343{
344 xfs_perag_intent_put(pag: ri->ri_pag);
345}
346
347/* Process a deferred refcount update. */
348STATIC int
349xfs_refcount_update_finish_item(
350 struct xfs_trans *tp,
351 struct xfs_log_item *done,
352 struct list_head *item,
353 struct xfs_btree_cur **state)
354{
355 struct xfs_refcount_intent *ri;
356 int error;
357
358 ri = container_of(item, struct xfs_refcount_intent, ri_list);
359
360 /* Did we run out of reservation? Requeue what we didn't finish. */
361 error = xfs_refcount_finish_one(tp, ri, state);
362 if (!error && ri->ri_blockcount > 0) {
363 ASSERT(ri->ri_type == XFS_REFCOUNT_INCREASE ||
364 ri->ri_type == XFS_REFCOUNT_DECREASE);
365 return -EAGAIN;
366 }
367
368 xfs_refcount_update_put_group(ri);
369 kmem_cache_free(xfs_refcount_intent_cache, ri);
370 return error;
371}
372
373/* Abort all pending CUIs. */
374STATIC void
375xfs_refcount_update_abort_intent(
376 struct xfs_log_item *intent)
377{
378 xfs_cui_release(cuip: CUI_ITEM(lip: intent));
379}
380
381/* Cancel a deferred refcount update. */
382STATIC void
383xfs_refcount_update_cancel_item(
384 struct list_head *item)
385{
386 struct xfs_refcount_intent *ri;
387
388 ri = container_of(item, struct xfs_refcount_intent, ri_list);
389
390 xfs_refcount_update_put_group(ri);
391 kmem_cache_free(xfs_refcount_intent_cache, ri);
392}
393
394/* Is this recovered CUI ok? */
395static inline bool
396xfs_cui_validate_phys(
397 struct xfs_mount *mp,
398 struct xfs_phys_extent *pmap)
399{
400 if (!xfs_has_reflink(mp))
401 return false;
402
403 if (pmap->pe_flags & ~XFS_REFCOUNT_EXTENT_FLAGS)
404 return false;
405
406 switch (pmap->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK) {
407 case XFS_REFCOUNT_INCREASE:
408 case XFS_REFCOUNT_DECREASE:
409 case XFS_REFCOUNT_ALLOC_COW:
410 case XFS_REFCOUNT_FREE_COW:
411 break;
412 default:
413 return false;
414 }
415
416 return xfs_verify_fsbext(mp, pmap->pe_startblock, pmap->pe_len);
417}
418
419static inline void
420xfs_cui_recover_work(
421 struct xfs_mount *mp,
422 struct xfs_defer_pending *dfp,
423 struct xfs_phys_extent *pmap)
424{
425 struct xfs_refcount_intent *ri;
426
427 ri = kmem_cache_alloc(xfs_refcount_intent_cache,
428 GFP_KERNEL | __GFP_NOFAIL);
429 ri->ri_type = pmap->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK;
430 ri->ri_startblock = pmap->pe_startblock;
431 ri->ri_blockcount = pmap->pe_len;
432 xfs_refcount_update_get_group(mp, ri);
433
434 xfs_defer_add_item(dfp, &ri->ri_list);
435}
436
437/*
438 * Process a refcount update intent item that was recovered from the log.
439 * We need to update the refcountbt.
440 */
441STATIC int
442xfs_refcount_recover_work(
443 struct xfs_defer_pending *dfp,
444 struct list_head *capture_list)
445{
446 struct xfs_trans_res resv;
447 struct xfs_log_item *lip = dfp->dfp_intent;
448 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
449 struct xfs_trans *tp;
450 struct xfs_mount *mp = lip->li_log->l_mp;
451 int i;
452 int error = 0;
453
454 /*
455 * First check the validity of the extents described by the
456 * CUI. If any are bad, then assume that all are bad and
457 * just toss the CUI.
458 */
459 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
460 if (!xfs_cui_validate_phys(mp,
461 pmap: &cuip->cui_format.cui_extents[i])) {
462 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
463 &cuip->cui_format,
464 sizeof(cuip->cui_format));
465 return -EFSCORRUPTED;
466 }
467
468 xfs_cui_recover_work(mp, dfp, pmap: &cuip->cui_format.cui_extents[i]);
469 }
470
471 /*
472 * Under normal operation, refcount updates are deferred, so we
473 * wouldn't be adding them directly to a transaction. All
474 * refcount updates manage reservation usage internally and
475 * dynamically by deferring work that won't fit in the
476 * transaction. Normally, any work that needs to be deferred
477 * gets attached to the same defer_ops that scheduled the
478 * refcount update. However, we're in log recovery here, so we
479 * use the passed in defer_ops and to finish up any work that
480 * doesn't fit. We need to reserve enough blocks to handle a
481 * full btree split on either end of the refcount range.
482 */
483 resv = xlog_recover_resv(&M_RES(mp)->tr_itruncate);
484 error = xfs_trans_alloc(mp, &resv, mp->m_refc_maxlevels * 2, 0,
485 XFS_TRANS_RESERVE, &tp);
486 if (error)
487 return error;
488
489 error = xlog_recover_finish_intent(tp, dfp);
490 if (error == -EFSCORRUPTED)
491 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
492 &cuip->cui_format,
493 sizeof(cuip->cui_format));
494 if (error)
495 goto abort_error;
496
497 return xfs_defer_ops_capture_and_commit(tp, capture_list);
498
499abort_error:
500 xfs_trans_cancel(tp);
501 return error;
502}
503
504/* Relog an intent item to push the log tail forward. */
505static struct xfs_log_item *
506xfs_refcount_relog_intent(
507 struct xfs_trans *tp,
508 struct xfs_log_item *intent,
509 struct xfs_log_item *done_item)
510{
511 struct xfs_cui_log_item *cuip;
512 struct xfs_phys_extent *pmap;
513 unsigned int count;
514
515 count = CUI_ITEM(lip: intent)->cui_format.cui_nextents;
516 pmap = CUI_ITEM(lip: intent)->cui_format.cui_extents;
517
518 cuip = xfs_cui_init(mp: tp->t_mountp, nextents: count);
519 memcpy(cuip->cui_format.cui_extents, pmap, count * sizeof(*pmap));
520 atomic_set(v: &cuip->cui_next_extent, i: count);
521
522 return &cuip->cui_item;
523}
524
525const struct xfs_defer_op_type xfs_refcount_update_defer_type = {
526 .name = "refcount",
527 .max_items = XFS_CUI_MAX_FAST_EXTENTS,
528 .create_intent = xfs_refcount_update_create_intent,
529 .abort_intent = xfs_refcount_update_abort_intent,
530 .create_done = xfs_refcount_update_create_done,
531 .finish_item = xfs_refcount_update_finish_item,
532 .finish_cleanup = xfs_refcount_finish_one_cleanup,
533 .cancel_item = xfs_refcount_update_cancel_item,
534 .recover_work = xfs_refcount_recover_work,
535 .relog_intent = xfs_refcount_relog_intent,
536};
537
538STATIC bool
539xfs_cui_item_match(
540 struct xfs_log_item *lip,
541 uint64_t intent_id)
542{
543 return CUI_ITEM(lip)->cui_format.cui_id == intent_id;
544}
545
546static const struct xfs_item_ops xfs_cui_item_ops = {
547 .flags = XFS_ITEM_INTENT,
548 .iop_size = xfs_cui_item_size,
549 .iop_format = xfs_cui_item_format,
550 .iop_unpin = xfs_cui_item_unpin,
551 .iop_release = xfs_cui_item_release,
552 .iop_match = xfs_cui_item_match,
553};
554
555static inline void
556xfs_cui_copy_format(
557 struct xfs_cui_log_format *dst,
558 const struct xfs_cui_log_format *src)
559{
560 unsigned int i;
561
562 memcpy(dst, src, offsetof(struct xfs_cui_log_format, cui_extents));
563
564 for (i = 0; i < src->cui_nextents; i++)
565 memcpy(&dst->cui_extents[i], &src->cui_extents[i],
566 sizeof(struct xfs_phys_extent));
567}
568
569/*
570 * This routine is called to create an in-core extent refcount update
571 * item from the cui format structure which was logged on disk.
572 * It allocates an in-core cui, copies the extents from the format
573 * structure into it, and adds the cui to the AIL with the given
574 * LSN.
575 */
576STATIC int
577xlog_recover_cui_commit_pass2(
578 struct xlog *log,
579 struct list_head *buffer_list,
580 struct xlog_recover_item *item,
581 xfs_lsn_t lsn)
582{
583 struct xfs_mount *mp = log->l_mp;
584 struct xfs_cui_log_item *cuip;
585 struct xfs_cui_log_format *cui_formatp;
586 size_t len;
587
588 cui_formatp = item->ri_buf[0].i_addr;
589
590 if (item->ri_buf[0].i_len < xfs_cui_log_format_sizeof(0)) {
591 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
592 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
593 return -EFSCORRUPTED;
594 }
595
596 len = xfs_cui_log_format_sizeof(cui_formatp->cui_nextents);
597 if (item->ri_buf[0].i_len != len) {
598 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
599 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
600 return -EFSCORRUPTED;
601 }
602
603 cuip = xfs_cui_init(mp, nextents: cui_formatp->cui_nextents);
604 xfs_cui_copy_format(dst: &cuip->cui_format, src: cui_formatp);
605 atomic_set(v: &cuip->cui_next_extent, i: cui_formatp->cui_nextents);
606
607 xlog_recover_intent_item(log, &cuip->cui_item, lsn,
608 &xfs_refcount_update_defer_type);
609 return 0;
610}
611
612const struct xlog_recover_item_ops xlog_cui_item_ops = {
613 .item_type = XFS_LI_CUI,
614 .commit_pass2 = xlog_recover_cui_commit_pass2,
615};
616
617/*
618 * This routine is called when an CUD format structure is found in a committed
619 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
620 * was still in the log. To do this it searches the AIL for the CUI with an id
621 * equal to that in the CUD format structure. If we find it we drop the CUD
622 * reference, which removes the CUI from the AIL and frees it.
623 */
624STATIC int
625xlog_recover_cud_commit_pass2(
626 struct xlog *log,
627 struct list_head *buffer_list,
628 struct xlog_recover_item *item,
629 xfs_lsn_t lsn)
630{
631 struct xfs_cud_log_format *cud_formatp;
632
633 cud_formatp = item->ri_buf[0].i_addr;
634 if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format)) {
635 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, log->l_mp,
636 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
637 return -EFSCORRUPTED;
638 }
639
640 xlog_recover_release_intent(log, XFS_LI_CUI, cud_formatp->cud_cui_id);
641 return 0;
642}
643
644const struct xlog_recover_item_ops xlog_cud_item_ops = {
645 .item_type = XFS_LI_CUD,
646 .commit_pass2 = xlog_recover_cud_commit_pass2,
647};
648

source code of linux/fs/xfs/xfs_refcount_item.c