1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2022-2023 Oracle. All Rights Reserved.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_trans_resv.h"
11 #include "xfs_mount.h"
12 #include "xfs_btree.h"
13 #include "xfs_btree_staging.h"
14 #include "xfs_log_format.h"
15 #include "xfs_trans.h"
17 #include "xfs_inode.h"
18 #include "xfs_alloc.h"
21 #include "xfs_defer.h"
22 #include "scrub/scrub.h"
23 #include "scrub/common.h"
24 #include "scrub/trace.h"
25 #include "scrub/repair.h"
26 #include "scrub/newbt.h"
29 * Estimate proper slack values for a btree that's being reloaded.
31 * Under most circumstances, we'll take whatever default loading value the
32 * btree bulk loading code calculates for us. However, there are some
33 * exceptions to this rule:
35 * (0) If someone turned one of the debug knobs.
36 * (1) If this is a per-AG btree and the AG has less than 10% space free.
37 * (2) If this is an inode btree and the FS has less than 10% space free.
39 * In either case, format the new btree blocks almost completely full to
40 * minimize space usage.
43 xrep_newbt_estimate_slack(
44 struct xrep_newbt *xnr)
46 struct xfs_scrub *sc = xnr->sc;
47 struct xfs_btree_bload *bload = &xnr->bload;
52 * The xfs_globals values are set to -1 (i.e. take the bload defaults)
53 * unless someone has set them otherwise, so we just pull the values
56 bload->leaf_slack = xfs_globals.bload_leaf_slack;
57 bload->node_slack = xfs_globals.bload_node_slack;
59 if (sc->ops->type == ST_PERAG) {
60 free = sc->sa.pag->pagf_freeblks;
61 sz = xfs_ag_block_count(sc->mp, sc->sa.pag->pag_agno);
63 free = percpu_counter_sum(&sc->mp->m_fdblocks);
64 sz = sc->mp->m_sb.sb_dblocks;
67 /* No further changes if there's more than 10% free space left. */
68 if (free >= div_u64(sz, 10))
72 * We're low on space; load the btrees as tightly as possible. Leave
73 * a couple of open slots in each btree block so that we don't end up
74 * splitting the btrees like crazy after a mount.
76 if (bload->leaf_slack < 0)
77 bload->leaf_slack = 2;
78 if (bload->node_slack < 0)
79 bload->node_slack = 2;
82 /* Initialize accounting resources for staging a new AG btree. */
85 struct xrep_newbt *xnr,
87 const struct xfs_owner_info *oinfo,
88 xfs_fsblock_t alloc_hint,
89 enum xfs_ag_resv_type resv)
91 memset(xnr, 0, sizeof(struct xrep_newbt));
93 xnr->oinfo = *oinfo; /* structure copy */
94 xnr->alloc_hint = alloc_hint;
96 INIT_LIST_HEAD(&xnr->resv_list);
97 xnr->bload.max_dirty = XFS_B_TO_FSBT(sc->mp, 256U << 10); /* 256K */
98 xrep_newbt_estimate_slack(xnr);
101 /* Initialize accounting resources for staging a new inode fork btree. */
103 xrep_newbt_init_inode(
104 struct xrep_newbt *xnr,
105 struct xfs_scrub *sc,
107 const struct xfs_owner_info *oinfo)
109 struct xfs_ifork *ifp;
111 ifp = kmem_cache_zalloc(xfs_ifork_cache, XCHK_GFP_FLAGS);
115 xrep_newbt_init_ag(xnr, sc, oinfo,
116 XFS_INO_TO_FSB(sc->mp, sc->ip->i_ino),
118 xnr->ifake.if_fork = ifp;
119 xnr->ifake.if_fork_size = xfs_inode_fork_size(sc->ip, whichfork);
124 * Initialize accounting resources for staging a new btree. Callers are
125 * expected to add their own reservations (and clean them up) manually.
128 xrep_newbt_init_bare(
129 struct xrep_newbt *xnr,
130 struct xfs_scrub *sc)
132 xrep_newbt_init_ag(xnr, sc, &XFS_RMAP_OINFO_ANY_OWNER, NULLFSBLOCK,
137 * Designate specific blocks to be used to build our new btree. @pag must be
138 * a passive reference.
141 xrep_newbt_add_blocks(
142 struct xrep_newbt *xnr,
143 struct xfs_perag *pag,
144 const struct xfs_alloc_arg *args)
146 struct xfs_mount *mp = xnr->sc->mp;
147 struct xrep_newbt_resv *resv;
150 resv = kmalloc(sizeof(struct xrep_newbt_resv), XCHK_GFP_FLAGS);
154 INIT_LIST_HEAD(&resv->list);
155 resv->agbno = XFS_FSB_TO_AGBNO(mp, args->fsbno);
156 resv->len = args->len;
158 resv->pag = xfs_perag_hold(pag);
161 ASSERT(xnr->oinfo.oi_offset == 0);
163 error = xfs_alloc_schedule_autoreap(args, true, &resv->autoreap);
168 list_add_tail(&resv->list, &xnr->resv_list);
171 xfs_perag_put(resv->pag);
177 * Add an extent to the new btree reservation pool. Callers are required to
178 * reap this reservation manually if the repair is cancelled. @pag must be a
182 xrep_newbt_add_extent(
183 struct xrep_newbt *xnr,
184 struct xfs_perag *pag,
188 struct xfs_mount *mp = xnr->sc->mp;
189 struct xfs_alloc_arg args = {
190 .tp = NULL, /* no autoreap */
192 .fsbno = XFS_AGB_TO_FSB(mp, pag->pag_agno, agbno),
197 return xrep_newbt_add_blocks(xnr, pag, &args);
200 /* Don't let our allocation hint take us beyond this AG */
202 xrep_newbt_validate_ag_alloc_hint(
203 struct xrep_newbt *xnr)
205 struct xfs_scrub *sc = xnr->sc;
206 xfs_agnumber_t agno = XFS_FSB_TO_AGNO(sc->mp, xnr->alloc_hint);
208 if (agno == sc->sa.pag->pag_agno &&
209 xfs_verify_fsbno(sc->mp, xnr->alloc_hint))
212 xnr->alloc_hint = XFS_AGB_TO_FSB(sc->mp, sc->sa.pag->pag_agno,
213 XFS_AGFL_BLOCK(sc->mp) + 1);
216 /* Allocate disk space for a new per-AG btree. */
218 xrep_newbt_alloc_ag_blocks(
219 struct xrep_newbt *xnr,
222 struct xfs_scrub *sc = xnr->sc;
223 struct xfs_mount *mp = sc->mp;
226 ASSERT(sc->sa.pag != NULL);
228 while (nr_blocks > 0) {
229 struct xfs_alloc_arg args = {
240 xrep_newbt_validate_ag_alloc_hint(xnr);
242 if (xnr->alloc_vextent)
243 error = xnr->alloc_vextent(sc, &args, xnr->alloc_hint);
245 error = xfs_alloc_vextent_near_bno(&args,
249 if (args.fsbno == NULLFSBLOCK)
252 agno = XFS_FSB_TO_AGNO(mp, args.fsbno);
254 trace_xrep_newbt_alloc_ag_blocks(mp, agno,
255 XFS_FSB_TO_AGBNO(mp, args.fsbno), args.len,
256 xnr->oinfo.oi_owner);
258 if (agno != sc->sa.pag->pag_agno) {
259 ASSERT(agno == sc->sa.pag->pag_agno);
260 return -EFSCORRUPTED;
263 error = xrep_newbt_add_blocks(xnr, sc->sa.pag, &args);
267 nr_blocks -= args.len;
268 xnr->alloc_hint = args.fsbno + args.len;
270 error = xrep_defer_finish(sc);
278 /* Don't let our allocation hint take us beyond EOFS */
280 xrep_newbt_validate_file_alloc_hint(
281 struct xrep_newbt *xnr)
283 struct xfs_scrub *sc = xnr->sc;
285 if (xfs_verify_fsbno(sc->mp, xnr->alloc_hint))
288 xnr->alloc_hint = XFS_AGB_TO_FSB(sc->mp, 0, XFS_AGFL_BLOCK(sc->mp) + 1);
291 /* Allocate disk space for our new file-based btree. */
293 xrep_newbt_alloc_file_blocks(
294 struct xrep_newbt *xnr,
297 struct xfs_scrub *sc = xnr->sc;
298 struct xfs_mount *mp = sc->mp;
301 while (nr_blocks > 0) {
302 struct xfs_alloc_arg args = {
311 struct xfs_perag *pag;
314 xrep_newbt_validate_file_alloc_hint(xnr);
316 if (xnr->alloc_vextent)
317 error = xnr->alloc_vextent(sc, &args, xnr->alloc_hint);
319 error = xfs_alloc_vextent_start_ag(&args,
323 if (args.fsbno == NULLFSBLOCK)
326 agno = XFS_FSB_TO_AGNO(mp, args.fsbno);
328 trace_xrep_newbt_alloc_file_blocks(mp, agno,
329 XFS_FSB_TO_AGBNO(mp, args.fsbno), args.len,
330 xnr->oinfo.oi_owner);
332 pag = xfs_perag_get(mp, agno);
335 return -EFSCORRUPTED;
338 error = xrep_newbt_add_blocks(xnr, pag, &args);
343 nr_blocks -= args.len;
344 xnr->alloc_hint = args.fsbno + args.len;
346 error = xrep_defer_finish(sc);
354 /* Allocate disk space for our new btree. */
356 xrep_newbt_alloc_blocks(
357 struct xrep_newbt *xnr,
361 return xrep_newbt_alloc_file_blocks(xnr, nr_blocks);
362 return xrep_newbt_alloc_ag_blocks(xnr, nr_blocks);
366 * Free the unused part of a space extent that was reserved for a new ondisk
367 * structure. Returns the number of EFIs logged or a negative errno.
370 xrep_newbt_free_extent(
371 struct xrep_newbt *xnr,
372 struct xrep_newbt_resv *resv,
373 bool btree_committed)
375 struct xfs_scrub *sc = xnr->sc;
376 xfs_agblock_t free_agbno = resv->agbno;
377 xfs_extlen_t free_aglen = resv->len;
381 if (!btree_committed || resv->used == 0) {
383 * If we're not committing a new btree or we didn't use the
384 * space reservation, let the existing EFI free the entire
387 trace_xrep_newbt_free_blocks(sc->mp, resv->pag->pag_agno,
388 free_agbno, free_aglen, xnr->oinfo.oi_owner);
389 xfs_alloc_commit_autoreap(sc->tp, &resv->autoreap);
394 * We used space and committed the btree. Cancel the autoreap, remove
395 * the written blocks from the reservation, and possibly log a new EFI
396 * to free any unused reservation space.
398 xfs_alloc_cancel_autoreap(sc->tp, &resv->autoreap);
399 free_agbno += resv->used;
400 free_aglen -= resv->used;
405 trace_xrep_newbt_free_blocks(sc->mp, resv->pag->pag_agno, free_agbno,
406 free_aglen, xnr->oinfo.oi_owner);
408 ASSERT(xnr->resv != XFS_AG_RESV_AGFL);
409 ASSERT(xnr->resv != XFS_AG_RESV_IGNORE);
412 * Use EFIs to free the reservations. This reduces the chance
413 * that we leak blocks if the system goes down.
415 fsbno = XFS_AGB_TO_FSB(sc->mp, resv->pag->pag_agno, free_agbno);
416 error = xfs_free_extent_later(sc->tp, fsbno, free_aglen, &xnr->oinfo,
424 /* Free all the accounting info and disk space we reserved for a new btree. */
427 struct xrep_newbt *xnr,
428 bool btree_committed)
430 struct xfs_scrub *sc = xnr->sc;
431 struct xrep_newbt_resv *resv, *n;
432 unsigned int freed = 0;
436 * If the filesystem already went down, we can't free the blocks. Skip
437 * ahead to freeing the incore metadata because we can't fix anything.
439 if (xfs_is_shutdown(sc->mp))
442 list_for_each_entry_safe(resv, n, &xnr->resv_list, list) {
445 ret = xrep_newbt_free_extent(xnr, resv, btree_committed);
446 list_del(&resv->list);
447 xfs_perag_put(resv->pag);
455 if (freed >= XREP_MAX_ITRUNCATE_EFIS) {
456 error = xrep_defer_finish(sc);
464 error = xrep_defer_finish(sc);
468 * If we still have reservations attached to @newbt, cleanup must have
469 * failed and the filesystem is about to go down. Clean up the incore
470 * reservations and try to commit to freeing the space we used.
472 list_for_each_entry_safe(resv, n, &xnr->resv_list, list) {
473 xfs_alloc_commit_autoreap(sc->tp, &resv->autoreap);
474 list_del(&resv->list);
475 xfs_perag_put(resv->pag);
480 kmem_cache_free(xfs_ifork_cache, xnr->ifake.if_fork);
481 xnr->ifake.if_fork = NULL;
488 * Free all the accounting info and unused disk space allocations after
489 * committing a new btree.
493 struct xrep_newbt *xnr)
495 return xrep_newbt_free(xnr, true);
499 * Free all the accounting info and all of the disk space we reserved for a new
500 * btree that we're not going to commit. We want to try to roll things back
501 * cleanly for things like ENOSPC midway through allocation.
505 struct xrep_newbt *xnr)
507 xrep_newbt_free(xnr, false);
510 /* Feed one of the reserved btree blocks to the bulk loader. */
512 xrep_newbt_claim_block(
513 struct xfs_btree_cur *cur,
514 struct xrep_newbt *xnr,
515 union xfs_btree_ptr *ptr)
517 struct xrep_newbt_resv *resv;
518 struct xfs_mount *mp = cur->bc_mp;
522 * The first item in the list should always have a free block unless
523 * we're completely out.
525 resv = list_first_entry(&xnr->resv_list, struct xrep_newbt_resv, list);
526 if (resv->used == resv->len)
530 * Peel off a block from the start of the reservation. We allocate
531 * blocks in order to place blocks on disk in increasing record or key
532 * order. The block reservations tend to end up on the list in
533 * decreasing order, which hopefully results in leaf blocks ending up
536 agbno = resv->agbno + resv->used;
539 /* If we used all the blocks in this reservation, move it to the end. */
540 if (resv->used == resv->len)
541 list_move_tail(&resv->list, &xnr->resv_list);
543 trace_xrep_newbt_claim_block(mp, resv->pag->pag_agno, agbno, 1,
544 xnr->oinfo.oi_owner);
546 if (cur->bc_ops->ptr_len == XFS_BTREE_LONG_PTR_LEN)
547 ptr->l = cpu_to_be64(XFS_AGB_TO_FSB(mp, resv->pag->pag_agno,
550 ptr->s = cpu_to_be32(agbno);
552 /* Relog all the EFIs. */
553 return xrep_defer_finish(xnr->sc);
556 /* How many reserved blocks are unused? */
558 xrep_newbt_unused_blocks(
559 struct xrep_newbt *xnr)
561 struct xrep_newbt_resv *resv;
562 unsigned int unused = 0;
564 list_for_each_entry(resv, &xnr->resv_list, list)
565 unused += resv->len - resv->used;