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 #include "xfs_btree.h"
25 #include "xfs_trace.h"
26
27 struct kmem_cache *xfs_cui_cache;
28 struct kmem_cache *xfs_cud_cache;
29
30 static const struct xfs_item_ops xfs_cui_item_ops;
31
32 static inline struct xfs_cui_log_item *CUI_ITEM(struct xfs_log_item *lip)
33 {
34 return container_of(lip, struct xfs_cui_log_item, cui_item);
35 }
36
37 STATIC void
38 xfs_cui_item_free(
39 struct xfs_cui_log_item *cuip)
40 {
41 kvfree(cuip->cui_item.li_lv_shadow);
42 if (cuip->cui_format.cui_nextents > XFS_CUI_MAX_FAST_EXTENTS)
43 kfree(cuip);
44 else
45 kmem_cache_free(xfs_cui_cache, cuip);
46 }
47
48 /*
49 * Freeing the CUI requires that we remove it from the AIL if it has already
50 * been placed there. However, the CUI may not yet have been placed in the AIL
51 * when called by xfs_cui_release() from CUD processing due to the ordering of
52 * committed vs unpin operations in bulk insert operations. Hence the reference
53 * count to ensure only the last caller frees the CUI.
54 */
55 STATIC void
56 xfs_cui_release(
57 struct xfs_cui_log_item *cuip)
58 {
59 ASSERT(atomic_read(&cuip->cui_refcount) > 0);
60 if (!atomic_dec_and_test(&cuip->cui_refcount))
61 return;
62
63 xfs_trans_ail_delete(&cuip->cui_item, 0);
64 xfs_cui_item_free(cuip);
65 }
66
67
68 STATIC void
69 xfs_cui_item_size(
70 struct xfs_log_item *lip,
71 int *nvecs,
72 int *nbytes)
73 {
74 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
75
76 *nvecs += 1;
77 *nbytes += xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents);
78 }
79
80 /*
81 * This is called to fill in the vector of log iovecs for the
82 * given cui log item. We use only 1 iovec, and we point that
83 * at the cui_log_format structure embedded in the cui item.
84 * It is at this point that we assert that all of the extent
85 * slots in the cui item have been filled.
86 */
87 STATIC void
88 xfs_cui_item_format(
89 struct xfs_log_item *lip,
90 struct xfs_log_vec *lv)
91 {
92 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
93 struct xfs_log_iovec *vecp = NULL;
94
95 ASSERT(atomic_read(&cuip->cui_next_extent) ==
96 cuip->cui_format.cui_nextents);
97
98 cuip->cui_format.cui_type = XFS_LI_CUI;
99 cuip->cui_format.cui_size = 1;
100
101 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUI_FORMAT, &cuip->cui_format,
102 xfs_cui_log_format_sizeof(cuip->cui_format.cui_nextents));
103 }
104
105 /*
106 * The unpin operation is the last place an CUI is manipulated in the log. It is
107 * either inserted in the AIL or aborted in the event of a log I/O error. In
108 * either case, the CUI transaction has been successfully committed to make it
109 * this far. Therefore, we expect whoever committed the CUI to either construct
110 * and commit the CUD or drop the CUD's reference in the event of error. Simply
111 * drop the log's CUI reference now that the log is done with it.
112 */
113 STATIC void
114 xfs_cui_item_unpin(
115 struct xfs_log_item *lip,
116 int remove)
117 {
118 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
119
120 xfs_cui_release(cuip);
121 }
122
123 /*
124 * The CUI has been either committed or aborted if the transaction has been
125 * cancelled. If the transaction was cancelled, an CUD isn't going to be
126 * constructed and thus we free the CUI here directly.
127 */
128 STATIC void
129 xfs_cui_item_release(
130 struct xfs_log_item *lip)
131 {
132 xfs_cui_release(CUI_ITEM(lip));
133 }
134
135 /*
136 * Allocate and initialize an cui item with the given number of extents.
137 */
138 STATIC struct xfs_cui_log_item *
139 xfs_cui_init(
140 struct xfs_mount *mp,
141 uint nextents)
142
143 {
144 struct xfs_cui_log_item *cuip;
145
146 ASSERT(nextents > 0);
147 if (nextents > XFS_CUI_MAX_FAST_EXTENTS)
148 cuip = kzalloc(xfs_cui_log_item_sizeof(nextents),
149 GFP_KERNEL | __GFP_NOFAIL);
150 else
151 cuip = kmem_cache_zalloc(xfs_cui_cache,
152 GFP_KERNEL | __GFP_NOFAIL);
153
154 xfs_log_item_init(mp, &cuip->cui_item, XFS_LI_CUI, &xfs_cui_item_ops);
155 cuip->cui_format.cui_nextents = nextents;
156 cuip->cui_format.cui_id = (uintptr_t)(void *)cuip;
157 atomic_set(&cuip->cui_next_extent, 0);
158 atomic_set(&cuip->cui_refcount, 2);
159
160 return cuip;
161 }
162
163 static inline struct xfs_cud_log_item *CUD_ITEM(struct xfs_log_item *lip)
164 {
165 return container_of(lip, struct xfs_cud_log_item, cud_item);
166 }
167
168 STATIC void
169 xfs_cud_item_size(
170 struct xfs_log_item *lip,
171 int *nvecs,
172 int *nbytes)
173 {
174 *nvecs += 1;
175 *nbytes += sizeof(struct xfs_cud_log_format);
176 }
177
178 /*
179 * This is called to fill in the vector of log iovecs for the
180 * given cud log item. We use only 1 iovec, and we point that
181 * at the cud_log_format structure embedded in the cud item.
182 * It is at this point that we assert that all of the extent
183 * slots in the cud item have been filled.
184 */
185 STATIC void
186 xfs_cud_item_format(
187 struct xfs_log_item *lip,
188 struct xfs_log_vec *lv)
189 {
190 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
191 struct xfs_log_iovec *vecp = NULL;
192
193 cudp->cud_format.cud_type = XFS_LI_CUD;
194 cudp->cud_format.cud_size = 1;
195
196 xlog_copy_iovec(lv, &vecp, XLOG_REG_TYPE_CUD_FORMAT, &cudp->cud_format,
197 sizeof(struct xfs_cud_log_format));
198 }
199
200 /*
201 * The CUD is either committed or aborted if the transaction is cancelled. If
202 * the transaction is cancelled, drop our reference to the CUI and free the
203 * CUD.
204 */
205 STATIC void
206 xfs_cud_item_release(
207 struct xfs_log_item *lip)
208 {
209 struct xfs_cud_log_item *cudp = CUD_ITEM(lip);
210
211 xfs_cui_release(cudp->cud_cuip);
212 kvfree(cudp->cud_item.li_lv_shadow);
213 kmem_cache_free(xfs_cud_cache, cudp);
214 }
215
216 static struct xfs_log_item *
217 xfs_cud_item_intent(
218 struct xfs_log_item *lip)
219 {
220 return &CUD_ITEM(lip)->cud_cuip->cui_item;
221 }
222
223 static const struct xfs_item_ops xfs_cud_item_ops = {
224 .flags = XFS_ITEM_RELEASE_WHEN_COMMITTED |
225 XFS_ITEM_INTENT_DONE,
226 .iop_size = xfs_cud_item_size,
227 .iop_format = xfs_cud_item_format,
228 .iop_release = xfs_cud_item_release,
229 .iop_intent = xfs_cud_item_intent,
230 };
231
232 static inline struct xfs_refcount_intent *ci_entry(const struct list_head *e)
233 {
234 return list_entry(e, struct xfs_refcount_intent, ri_list);
235 }
236
237 /* Sort refcount intents by AG. */
238 static int
239 xfs_refcount_update_diff_items(
240 void *priv,
241 const struct list_head *a,
242 const struct list_head *b)
243 {
244 struct xfs_refcount_intent *ra = ci_entry(a);
245 struct xfs_refcount_intent *rb = ci_entry(b);
246
247 return ra->ri_pag->pag_agno - rb->ri_pag->pag_agno;
248 }
249
250 /* Log refcount updates in the intent item. */
251 STATIC void
252 xfs_refcount_update_log_item(
253 struct xfs_trans *tp,
254 struct xfs_cui_log_item *cuip,
255 struct xfs_refcount_intent *ri)
256 {
257 uint next_extent;
258 struct xfs_phys_extent *pmap;
259
260 /*
261 * atomic_inc_return gives us the value after the increment;
262 * we want to use it as an array index so we need to subtract 1 from
263 * it.
264 */
265 next_extent = atomic_inc_return(&cuip->cui_next_extent) - 1;
266 ASSERT(next_extent < cuip->cui_format.cui_nextents);
267 pmap = &cuip->cui_format.cui_extents[next_extent];
268 pmap->pe_startblock = ri->ri_startblock;
269 pmap->pe_len = ri->ri_blockcount;
270
271 pmap->pe_flags = 0;
272 switch (ri->ri_type) {
273 case XFS_REFCOUNT_INCREASE:
274 case XFS_REFCOUNT_DECREASE:
275 case XFS_REFCOUNT_ALLOC_COW:
276 case XFS_REFCOUNT_FREE_COW:
277 pmap->pe_flags |= ri->ri_type;
278 break;
279 default:
280 ASSERT(0);
281 }
282 }
283
284 static struct xfs_log_item *
285 xfs_refcount_update_create_intent(
286 struct xfs_trans *tp,
287 struct list_head *items,
288 unsigned int count,
289 bool sort)
290 {
291 struct xfs_mount *mp = tp->t_mountp;
292 struct xfs_cui_log_item *cuip = xfs_cui_init(mp, count);
293 struct xfs_refcount_intent *ri;
294
295 ASSERT(count > 0);
296
297 if (sort)
298 list_sort(mp, items, xfs_refcount_update_diff_items);
299 list_for_each_entry(ri, items, ri_list)
300 xfs_refcount_update_log_item(tp, cuip, ri);
301 return &cuip->cui_item;
302 }
303
304 /* Get an CUD so we can process all the deferred refcount updates. */
305 static struct xfs_log_item *
306 xfs_refcount_update_create_done(
307 struct xfs_trans *tp,
308 struct xfs_log_item *intent,
309 unsigned int count)
310 {
311 struct xfs_cui_log_item *cuip = CUI_ITEM(intent);
312 struct xfs_cud_log_item *cudp;
313
314 cudp = kmem_cache_zalloc(xfs_cud_cache, GFP_KERNEL | __GFP_NOFAIL);
315 xfs_log_item_init(tp->t_mountp, &cudp->cud_item, XFS_LI_CUD,
316 &xfs_cud_item_ops);
317 cudp->cud_cuip = cuip;
318 cudp->cud_format.cud_cui_id = cuip->cui_format.cui_id;
319
320 return &cudp->cud_item;
321 }
322
323 /* Add this deferred CUI to the transaction. */
324 void
325 xfs_refcount_defer_add(
326 struct xfs_trans *tp,
327 struct xfs_refcount_intent *ri)
328 {
329 struct xfs_mount *mp = tp->t_mountp;
330
331 trace_xfs_refcount_defer(mp, ri);
332
333 ri->ri_pag = xfs_perag_intent_get(mp, ri->ri_startblock);
334 xfs_defer_add(tp, &ri->ri_list, &xfs_refcount_update_defer_type);
335 }
336
337 /* Cancel a deferred refcount update. */
338 STATIC void
339 xfs_refcount_update_cancel_item(
340 struct list_head *item)
341 {
342 struct xfs_refcount_intent *ri = ci_entry(item);
343
344 xfs_perag_intent_put(ri->ri_pag);
345 kmem_cache_free(xfs_refcount_intent_cache, ri);
346 }
347
348 /* Process a deferred refcount update. */
349 STATIC int
350 xfs_refcount_update_finish_item(
351 struct xfs_trans *tp,
352 struct xfs_log_item *done,
353 struct list_head *item,
354 struct xfs_btree_cur **state)
355 {
356 struct xfs_refcount_intent *ri = ci_entry(item);
357 int error;
358
359 /* Did we run out of reservation? Requeue what we didn't finish. */
360 error = xfs_refcount_finish_one(tp, ri, state);
361 if (!error && ri->ri_blockcount > 0) {
362 ASSERT(ri->ri_type == XFS_REFCOUNT_INCREASE ||
363 ri->ri_type == XFS_REFCOUNT_DECREASE);
364 return -EAGAIN;
365 }
366
367 xfs_refcount_update_cancel_item(item);
368 return error;
369 }
370
371 /* Clean up after calling xfs_refcount_finish_one. */
372 STATIC void
373 xfs_refcount_finish_one_cleanup(
374 struct xfs_trans *tp,
375 struct xfs_btree_cur *rcur,
376 int error)
377 {
378 struct xfs_buf *agbp;
379
380 if (rcur == NULL)
381 return;
382 agbp = rcur->bc_ag.agbp;
383 xfs_btree_del_cursor(rcur, error);
384 if (error)
385 xfs_trans_brelse(tp, agbp);
386 }
387
388 /* Abort all pending CUIs. */
389 STATIC void
390 xfs_refcount_update_abort_intent(
391 struct xfs_log_item *intent)
392 {
393 xfs_cui_release(CUI_ITEM(intent));
394 }
395
396 /* Is this recovered CUI ok? */
397 static inline bool
398 xfs_cui_validate_phys(
399 struct xfs_mount *mp,
400 struct xfs_phys_extent *pmap)
401 {
402 if (!xfs_has_reflink(mp))
403 return false;
404
405 if (pmap->pe_flags & ~XFS_REFCOUNT_EXTENT_FLAGS)
406 return false;
407
408 switch (pmap->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK) {
409 case XFS_REFCOUNT_INCREASE:
410 case XFS_REFCOUNT_DECREASE:
411 case XFS_REFCOUNT_ALLOC_COW:
412 case XFS_REFCOUNT_FREE_COW:
413 break;
414 default:
415 return false;
416 }
417
418 return xfs_verify_fsbext(mp, pmap->pe_startblock, pmap->pe_len);
419 }
420
421 static inline void
422 xfs_cui_recover_work(
423 struct xfs_mount *mp,
424 struct xfs_defer_pending *dfp,
425 struct xfs_phys_extent *pmap)
426 {
427 struct xfs_refcount_intent *ri;
428
429 ri = kmem_cache_alloc(xfs_refcount_intent_cache,
430 GFP_KERNEL | __GFP_NOFAIL);
431 ri->ri_type = pmap->pe_flags & XFS_REFCOUNT_EXTENT_TYPE_MASK;
432 ri->ri_startblock = pmap->pe_startblock;
433 ri->ri_blockcount = pmap->pe_len;
434 ri->ri_pag = xfs_perag_intent_get(mp, pmap->pe_startblock);
435
436 xfs_defer_add_item(dfp, &ri->ri_list);
437 }
438
439 /*
440 * Process a refcount update intent item that was recovered from the log.
441 * We need to update the refcountbt.
442 */
443 STATIC int
444 xfs_refcount_recover_work(
445 struct xfs_defer_pending *dfp,
446 struct list_head *capture_list)
447 {
448 struct xfs_trans_res resv;
449 struct xfs_log_item *lip = dfp->dfp_intent;
450 struct xfs_cui_log_item *cuip = CUI_ITEM(lip);
451 struct xfs_trans *tp;
452 struct xfs_mount *mp = lip->li_log->l_mp;
453 int i;
454 int error = 0;
455
456 /*
457 * First check the validity of the extents described by the
458 * CUI. If any are bad, then assume that all are bad and
459 * just toss the CUI.
460 */
461 for (i = 0; i < cuip->cui_format.cui_nextents; i++) {
462 if (!xfs_cui_validate_phys(mp,
463 &cuip->cui_format.cui_extents[i])) {
464 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
465 &cuip->cui_format,
466 sizeof(cuip->cui_format));
467 return -EFSCORRUPTED;
468 }
469
470 xfs_cui_recover_work(mp, dfp, &cuip->cui_format.cui_extents[i]);
471 }
472
473 /*
474 * Under normal operation, refcount updates are deferred, so we
475 * wouldn't be adding them directly to a transaction. All
476 * refcount updates manage reservation usage internally and
477 * dynamically by deferring work that won't fit in the
478 * transaction. Normally, any work that needs to be deferred
479 * gets attached to the same defer_ops that scheduled the
480 * refcount update. However, we're in log recovery here, so we
481 * use the passed in defer_ops and to finish up any work that
482 * doesn't fit. We need to reserve enough blocks to handle a
483 * full btree split on either end of the refcount range.
484 */
485 resv = xlog_recover_resv(&M_RES(mp)->tr_itruncate);
486 error = xfs_trans_alloc(mp, &resv, mp->m_refc_maxlevels * 2, 0,
487 XFS_TRANS_RESERVE, &tp);
488 if (error)
489 return error;
490
491 error = xlog_recover_finish_intent(tp, dfp);
492 if (error == -EFSCORRUPTED)
493 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
494 &cuip->cui_format,
495 sizeof(cuip->cui_format));
496 if (error)
497 goto abort_error;
498
499 return xfs_defer_ops_capture_and_commit(tp, capture_list);
500
501 abort_error:
502 xfs_trans_cancel(tp);
503 return error;
504 }
505
506 /* Relog an intent item to push the log tail forward. */
507 static struct xfs_log_item *
508 xfs_refcount_relog_intent(
509 struct xfs_trans *tp,
510 struct xfs_log_item *intent,
511 struct xfs_log_item *done_item)
512 {
513 struct xfs_cui_log_item *cuip;
514 struct xfs_phys_extent *pmap;
515 unsigned int count;
516
517 count = CUI_ITEM(intent)->cui_format.cui_nextents;
518 pmap = CUI_ITEM(intent)->cui_format.cui_extents;
519
520 cuip = xfs_cui_init(tp->t_mountp, count);
521 memcpy(cuip->cui_format.cui_extents, pmap, count * sizeof(*pmap));
522 atomic_set(&cuip->cui_next_extent, count);
523
524 return &cuip->cui_item;
525 }
526
527 const struct xfs_defer_op_type xfs_refcount_update_defer_type = {
528 .name = "refcount",
529 .max_items = XFS_CUI_MAX_FAST_EXTENTS,
530 .create_intent = xfs_refcount_update_create_intent,
531 .abort_intent = xfs_refcount_update_abort_intent,
532 .create_done = xfs_refcount_update_create_done,
533 .finish_item = xfs_refcount_update_finish_item,
534 .finish_cleanup = xfs_refcount_finish_one_cleanup,
535 .cancel_item = xfs_refcount_update_cancel_item,
536 .recover_work = xfs_refcount_recover_work,
537 .relog_intent = xfs_refcount_relog_intent,
538 };
539
540 STATIC bool
541 xfs_cui_item_match(
542 struct xfs_log_item *lip,
543 uint64_t intent_id)
544 {
545 return CUI_ITEM(lip)->cui_format.cui_id == intent_id;
546 }
547
548 static const struct xfs_item_ops xfs_cui_item_ops = {
549 .flags = XFS_ITEM_INTENT,
550 .iop_size = xfs_cui_item_size,
551 .iop_format = xfs_cui_item_format,
552 .iop_unpin = xfs_cui_item_unpin,
553 .iop_release = xfs_cui_item_release,
554 .iop_match = xfs_cui_item_match,
555 };
556
557 static inline void
558 xfs_cui_copy_format(
559 struct xfs_cui_log_format *dst,
560 const struct xfs_cui_log_format *src)
561 {
562 unsigned int i;
563
564 memcpy(dst, src, offsetof(struct xfs_cui_log_format, cui_extents));
565
566 for (i = 0; i < src->cui_nextents; i++)
567 memcpy(&dst->cui_extents[i], &src->cui_extents[i],
568 sizeof(struct xfs_phys_extent));
569 }
570
571 /*
572 * This routine is called to create an in-core extent refcount update
573 * item from the cui format structure which was logged on disk.
574 * It allocates an in-core cui, copies the extents from the format
575 * structure into it, and adds the cui to the AIL with the given
576 * LSN.
577 */
578 STATIC int
579 xlog_recover_cui_commit_pass2(
580 struct xlog *log,
581 struct list_head *buffer_list,
582 struct xlog_recover_item *item,
583 xfs_lsn_t lsn)
584 {
585 struct xfs_mount *mp = log->l_mp;
586 struct xfs_cui_log_item *cuip;
587 struct xfs_cui_log_format *cui_formatp;
588 size_t len;
589
590 cui_formatp = item->ri_buf[0].i_addr;
591
592 if (item->ri_buf[0].i_len < xfs_cui_log_format_sizeof(0)) {
593 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
594 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
595 return -EFSCORRUPTED;
596 }
597
598 len = xfs_cui_log_format_sizeof(cui_formatp->cui_nextents);
599 if (item->ri_buf[0].i_len != len) {
600 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
601 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
602 return -EFSCORRUPTED;
603 }
604
605 cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
606 xfs_cui_copy_format(&cuip->cui_format, cui_formatp);
607 atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
608
609 xlog_recover_intent_item(log, &cuip->cui_item, lsn,
610 &xfs_refcount_update_defer_type);
611 return 0;
612 }
613
614 const struct xlog_recover_item_ops xlog_cui_item_ops = {
615 .item_type = XFS_LI_CUI,
616 .commit_pass2 = xlog_recover_cui_commit_pass2,
617 };
618
619 /*
620 * This routine is called when an CUD format structure is found in a committed
621 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
622 * was still in the log. To do this it searches the AIL for the CUI with an id
623 * equal to that in the CUD format structure. If we find it we drop the CUD
624 * reference, which removes the CUI from the AIL and frees it.
625 */
626 STATIC int
627 xlog_recover_cud_commit_pass2(
628 struct xlog *log,
629 struct list_head *buffer_list,
630 struct xlog_recover_item *item,
631 xfs_lsn_t lsn)
632 {
633 struct xfs_cud_log_format *cud_formatp;
634
635 cud_formatp = item->ri_buf[0].i_addr;
636 if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format)) {
637 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, log->l_mp,
638 item->ri_buf[0].i_addr, item->ri_buf[0].i_len);
639 return -EFSCORRUPTED;
640 }
641
642 xlog_recover_release_intent(log, XFS_LI_CUI, cud_formatp->cud_cui_id);
643 return 0;
644 }
645
646 const struct xlog_recover_item_ops xlog_cud_item_ops = {
647 .item_type = XFS_LI_CUD,
648 .commit_pass2 = xlog_recover_cud_commit_pass2,
649 };
650
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.