2 * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_types.h"
24 #include "xfs_trans.h"
27 #include "xfs_mount.h"
28 #include "xfs_bmap_btree.h"
29 #include "xfs_alloc_btree.h"
30 #include "xfs_ialloc_btree.h"
31 #include "xfs_dinode.h"
32 #include "xfs_inode.h"
33 #include "xfs_btree.h"
34 #include "xfs_alloc.h"
35 #include "xfs_error.h"
36 #include "xfs_trace.h"
38 struct workqueue_struct *xfs_alloc_wq;
40 #define XFS_ABSDIFF(a,b) (((a) <= (b)) ? ((b) - (a)) : ((a) - (b)))
42 #define XFSA_FIXUP_BNO_OK 1
43 #define XFSA_FIXUP_CNT_OK 2
45 STATIC int xfs_alloc_ag_vextent_exact(xfs_alloc_arg_t *);
46 STATIC int xfs_alloc_ag_vextent_near(xfs_alloc_arg_t *);
47 STATIC int xfs_alloc_ag_vextent_size(xfs_alloc_arg_t *);
48 STATIC int xfs_alloc_ag_vextent_small(xfs_alloc_arg_t *,
49 xfs_btree_cur_t *, xfs_agblock_t *, xfs_extlen_t *, int *);
50 STATIC void xfs_alloc_busy_trim(struct xfs_alloc_arg *,
51 xfs_agblock_t, xfs_extlen_t, xfs_agblock_t *, xfs_extlen_t *);
54 * Lookup the record equal to [bno, len] in the btree given by cur.
56 STATIC int /* error */
58 struct xfs_btree_cur *cur, /* btree cursor */
59 xfs_agblock_t bno, /* starting block of extent */
60 xfs_extlen_t len, /* length of extent */
61 int *stat) /* success/failure */
63 cur->bc_rec.a.ar_startblock = bno;
64 cur->bc_rec.a.ar_blockcount = len;
65 return xfs_btree_lookup(cur, XFS_LOOKUP_EQ, stat);
69 * Lookup the first record greater than or equal to [bno, len]
70 * in the btree given by cur.
74 struct xfs_btree_cur *cur, /* btree cursor */
75 xfs_agblock_t bno, /* starting block of extent */
76 xfs_extlen_t len, /* length of extent */
77 int *stat) /* success/failure */
79 cur->bc_rec.a.ar_startblock = bno;
80 cur->bc_rec.a.ar_blockcount = len;
81 return xfs_btree_lookup(cur, XFS_LOOKUP_GE, stat);
85 * Lookup the first record less than or equal to [bno, len]
86 * in the btree given by cur.
90 struct xfs_btree_cur *cur, /* btree cursor */
91 xfs_agblock_t bno, /* starting block of extent */
92 xfs_extlen_t len, /* length of extent */
93 int *stat) /* success/failure */
95 cur->bc_rec.a.ar_startblock = bno;
96 cur->bc_rec.a.ar_blockcount = len;
97 return xfs_btree_lookup(cur, XFS_LOOKUP_LE, stat);
101 * Update the record referred to by cur to the value given
103 * This either works (return 0) or gets an EFSCORRUPTED error.
105 STATIC int /* error */
107 struct xfs_btree_cur *cur, /* btree cursor */
108 xfs_agblock_t bno, /* starting block of extent */
109 xfs_extlen_t len) /* length of extent */
111 union xfs_btree_rec rec;
113 rec.alloc.ar_startblock = cpu_to_be32(bno);
114 rec.alloc.ar_blockcount = cpu_to_be32(len);
115 return xfs_btree_update(cur, &rec);
119 * Get the data from the pointed-to record.
123 struct xfs_btree_cur *cur, /* btree cursor */
124 xfs_agblock_t *bno, /* output: starting block of extent */
125 xfs_extlen_t *len, /* output: length of extent */
126 int *stat) /* output: success/failure */
128 union xfs_btree_rec *rec;
131 error = xfs_btree_get_rec(cur, &rec, stat);
132 if (!error && *stat == 1) {
133 *bno = be32_to_cpu(rec->alloc.ar_startblock);
134 *len = be32_to_cpu(rec->alloc.ar_blockcount);
140 * Compute aligned version of the found extent.
141 * Takes alignment and min length into account.
144 xfs_alloc_compute_aligned(
145 xfs_alloc_arg_t *args, /* allocation argument structure */
146 xfs_agblock_t foundbno, /* starting block in found extent */
147 xfs_extlen_t foundlen, /* length in found extent */
148 xfs_agblock_t *resbno, /* result block number */
149 xfs_extlen_t *reslen) /* result length */
154 /* Trim busy sections out of found extent */
155 xfs_alloc_busy_trim(args, foundbno, foundlen, &bno, &len);
157 if (args->alignment > 1 && len >= args->minlen) {
158 xfs_agblock_t aligned_bno = roundup(bno, args->alignment);
159 xfs_extlen_t diff = aligned_bno - bno;
161 *resbno = aligned_bno;
162 *reslen = diff >= len ? 0 : len - diff;
170 * Compute best start block and diff for "near" allocations.
171 * freelen >= wantlen already checked by caller.
173 STATIC xfs_extlen_t /* difference value (absolute) */
174 xfs_alloc_compute_diff(
175 xfs_agblock_t wantbno, /* target starting block */
176 xfs_extlen_t wantlen, /* target length */
177 xfs_extlen_t alignment, /* target alignment */
178 xfs_agblock_t freebno, /* freespace's starting block */
179 xfs_extlen_t freelen, /* freespace's length */
180 xfs_agblock_t *newbnop) /* result: best start block from free */
182 xfs_agblock_t freeend; /* end of freespace extent */
183 xfs_agblock_t newbno1; /* return block number */
184 xfs_agblock_t newbno2; /* other new block number */
185 xfs_extlen_t newlen1=0; /* length with newbno1 */
186 xfs_extlen_t newlen2=0; /* length with newbno2 */
187 xfs_agblock_t wantend; /* end of target extent */
189 ASSERT(freelen >= wantlen);
190 freeend = freebno + freelen;
191 wantend = wantbno + wantlen;
192 if (freebno >= wantbno) {
193 if ((newbno1 = roundup(freebno, alignment)) >= freeend)
194 newbno1 = NULLAGBLOCK;
195 } else if (freeend >= wantend && alignment > 1) {
196 newbno1 = roundup(wantbno, alignment);
197 newbno2 = newbno1 - alignment;
198 if (newbno1 >= freeend)
199 newbno1 = NULLAGBLOCK;
201 newlen1 = XFS_EXTLEN_MIN(wantlen, freeend - newbno1);
202 if (newbno2 < freebno)
203 newbno2 = NULLAGBLOCK;
205 newlen2 = XFS_EXTLEN_MIN(wantlen, freeend - newbno2);
206 if (newbno1 != NULLAGBLOCK && newbno2 != NULLAGBLOCK) {
207 if (newlen1 < newlen2 ||
208 (newlen1 == newlen2 &&
209 XFS_ABSDIFF(newbno1, wantbno) >
210 XFS_ABSDIFF(newbno2, wantbno)))
212 } else if (newbno2 != NULLAGBLOCK)
214 } else if (freeend >= wantend) {
216 } else if (alignment > 1) {
217 newbno1 = roundup(freeend - wantlen, alignment);
218 if (newbno1 > freeend - wantlen &&
219 newbno1 - alignment >= freebno)
220 newbno1 -= alignment;
221 else if (newbno1 >= freeend)
222 newbno1 = NULLAGBLOCK;
224 newbno1 = freeend - wantlen;
226 return newbno1 == NULLAGBLOCK ? 0 : XFS_ABSDIFF(newbno1, wantbno);
230 * Fix up the length, based on mod and prod.
231 * len should be k * prod + mod for some k.
232 * If len is too small it is returned unchanged.
233 * If len hits maxlen it is left alone.
237 xfs_alloc_arg_t *args) /* allocation argument structure */
242 ASSERT(args->mod < args->prod);
244 ASSERT(rlen >= args->minlen);
245 ASSERT(rlen <= args->maxlen);
246 if (args->prod <= 1 || rlen < args->mod || rlen == args->maxlen ||
247 (args->mod == 0 && rlen < args->prod))
249 k = rlen % args->prod;
253 if ((int)(rlen = rlen - k - args->mod) < (int)args->minlen)
256 if ((int)(rlen = rlen - args->prod - (args->mod - k)) <
260 ASSERT(rlen >= args->minlen);
261 ASSERT(rlen <= args->maxlen);
266 * Fix up length if there is too little space left in the a.g.
267 * Return 1 if ok, 0 if too little, should give up.
270 xfs_alloc_fix_minleft(
271 xfs_alloc_arg_t *args) /* allocation argument structure */
273 xfs_agf_t *agf; /* a.g. freelist header */
274 int diff; /* free space difference */
276 if (args->minleft == 0)
278 agf = XFS_BUF_TO_AGF(args->agbp);
279 diff = be32_to_cpu(agf->agf_freeblks)
280 - args->len - args->minleft;
283 args->len += diff; /* shrink the allocated space */
284 if (args->len >= args->minlen)
286 args->agbno = NULLAGBLOCK;
291 * Update the two btrees, logically removing from freespace the extent
292 * starting at rbno, rlen blocks. The extent is contained within the
293 * actual (current) free extent fbno for flen blocks.
294 * Flags are passed in indicating whether the cursors are set to the
297 STATIC int /* error code */
298 xfs_alloc_fixup_trees(
299 xfs_btree_cur_t *cnt_cur, /* cursor for by-size btree */
300 xfs_btree_cur_t *bno_cur, /* cursor for by-block btree */
301 xfs_agblock_t fbno, /* starting block of free extent */
302 xfs_extlen_t flen, /* length of free extent */
303 xfs_agblock_t rbno, /* starting block of returned extent */
304 xfs_extlen_t rlen, /* length of returned extent */
305 int flags) /* flags, XFSA_FIXUP_... */
307 int error; /* error code */
308 int i; /* operation results */
309 xfs_agblock_t nfbno1; /* first new free startblock */
310 xfs_agblock_t nfbno2; /* second new free startblock */
311 xfs_extlen_t nflen1=0; /* first new free length */
312 xfs_extlen_t nflen2=0; /* second new free length */
315 * Look up the record in the by-size tree if necessary.
317 if (flags & XFSA_FIXUP_CNT_OK) {
319 if ((error = xfs_alloc_get_rec(cnt_cur, &nfbno1, &nflen1, &i)))
321 XFS_WANT_CORRUPTED_RETURN(
322 i == 1 && nfbno1 == fbno && nflen1 == flen);
325 if ((error = xfs_alloc_lookup_eq(cnt_cur, fbno, flen, &i)))
327 XFS_WANT_CORRUPTED_RETURN(i == 1);
330 * Look up the record in the by-block tree if necessary.
332 if (flags & XFSA_FIXUP_BNO_OK) {
334 if ((error = xfs_alloc_get_rec(bno_cur, &nfbno1, &nflen1, &i)))
336 XFS_WANT_CORRUPTED_RETURN(
337 i == 1 && nfbno1 == fbno && nflen1 == flen);
340 if ((error = xfs_alloc_lookup_eq(bno_cur, fbno, flen, &i)))
342 XFS_WANT_CORRUPTED_RETURN(i == 1);
346 if (bno_cur->bc_nlevels == 1 && cnt_cur->bc_nlevels == 1) {
347 struct xfs_btree_block *bnoblock;
348 struct xfs_btree_block *cntblock;
350 bnoblock = XFS_BUF_TO_BLOCK(bno_cur->bc_bufs[0]);
351 cntblock = XFS_BUF_TO_BLOCK(cnt_cur->bc_bufs[0]);
353 XFS_WANT_CORRUPTED_RETURN(
354 bnoblock->bb_numrecs == cntblock->bb_numrecs);
359 * Deal with all four cases: the allocated record is contained
360 * within the freespace record, so we can have new freespace
361 * at either (or both) end, or no freespace remaining.
363 if (rbno == fbno && rlen == flen)
364 nfbno1 = nfbno2 = NULLAGBLOCK;
365 else if (rbno == fbno) {
366 nfbno1 = rbno + rlen;
367 nflen1 = flen - rlen;
368 nfbno2 = NULLAGBLOCK;
369 } else if (rbno + rlen == fbno + flen) {
371 nflen1 = flen - rlen;
372 nfbno2 = NULLAGBLOCK;
375 nflen1 = rbno - fbno;
376 nfbno2 = rbno + rlen;
377 nflen2 = (fbno + flen) - nfbno2;
380 * Delete the entry from the by-size btree.
382 if ((error = xfs_btree_delete(cnt_cur, &i)))
384 XFS_WANT_CORRUPTED_RETURN(i == 1);
386 * Add new by-size btree entry(s).
388 if (nfbno1 != NULLAGBLOCK) {
389 if ((error = xfs_alloc_lookup_eq(cnt_cur, nfbno1, nflen1, &i)))
391 XFS_WANT_CORRUPTED_RETURN(i == 0);
392 if ((error = xfs_btree_insert(cnt_cur, &i)))
394 XFS_WANT_CORRUPTED_RETURN(i == 1);
396 if (nfbno2 != NULLAGBLOCK) {
397 if ((error = xfs_alloc_lookup_eq(cnt_cur, nfbno2, nflen2, &i)))
399 XFS_WANT_CORRUPTED_RETURN(i == 0);
400 if ((error = xfs_btree_insert(cnt_cur, &i)))
402 XFS_WANT_CORRUPTED_RETURN(i == 1);
405 * Fix up the by-block btree entry(s).
407 if (nfbno1 == NULLAGBLOCK) {
409 * No remaining freespace, just delete the by-block tree entry.
411 if ((error = xfs_btree_delete(bno_cur, &i)))
413 XFS_WANT_CORRUPTED_RETURN(i == 1);
416 * Update the by-block entry to start later|be shorter.
418 if ((error = xfs_alloc_update(bno_cur, nfbno1, nflen1)))
421 if (nfbno2 != NULLAGBLOCK) {
423 * 2 resulting free entries, need to add one.
425 if ((error = xfs_alloc_lookup_eq(bno_cur, nfbno2, nflen2, &i)))
427 XFS_WANT_CORRUPTED_RETURN(i == 0);
428 if ((error = xfs_btree_insert(bno_cur, &i)))
430 XFS_WANT_CORRUPTED_RETURN(i == 1);
436 * Read in the allocation group free block array.
438 STATIC int /* error */
440 xfs_mount_t *mp, /* mount point structure */
441 xfs_trans_t *tp, /* transaction pointer */
442 xfs_agnumber_t agno, /* allocation group number */
443 xfs_buf_t **bpp) /* buffer for the ag free block array */
445 xfs_buf_t *bp; /* return value */
448 ASSERT(agno != NULLAGNUMBER);
449 error = xfs_trans_read_buf(
450 mp, tp, mp->m_ddev_targp,
451 XFS_AG_DADDR(mp, agno, XFS_AGFL_DADDR(mp)),
452 XFS_FSS_TO_BB(mp, 1), 0, &bp);
455 ASSERT(!xfs_buf_geterror(bp));
456 xfs_buf_set_ref(bp, XFS_AGFL_REF);
462 xfs_alloc_update_counters(
463 struct xfs_trans *tp,
464 struct xfs_perag *pag,
465 struct xfs_buf *agbp,
468 struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
470 pag->pagf_freeblks += len;
471 be32_add_cpu(&agf->agf_freeblks, len);
473 xfs_trans_agblocks_delta(tp, len);
474 if (unlikely(be32_to_cpu(agf->agf_freeblks) >
475 be32_to_cpu(agf->agf_length)))
478 xfs_alloc_log_agf(tp, agbp, XFS_AGF_FREEBLKS);
483 * Allocation group level functions.
487 * Allocate a variable extent in the allocation group agno.
488 * Type and bno are used to determine where in the allocation group the
490 * Extent's length (returned in *len) will be between minlen and maxlen,
491 * and of the form k * prod + mod unless there's nothing that large.
492 * Return the starting a.g. block, or NULLAGBLOCK if we can't do it.
494 STATIC int /* error */
495 xfs_alloc_ag_vextent(
496 xfs_alloc_arg_t *args) /* argument structure for allocation */
500 ASSERT(args->minlen > 0);
501 ASSERT(args->maxlen > 0);
502 ASSERT(args->minlen <= args->maxlen);
503 ASSERT(args->mod < args->prod);
504 ASSERT(args->alignment > 0);
506 * Branch to correct routine based on the type.
509 switch (args->type) {
510 case XFS_ALLOCTYPE_THIS_AG:
511 error = xfs_alloc_ag_vextent_size(args);
513 case XFS_ALLOCTYPE_NEAR_BNO:
514 error = xfs_alloc_ag_vextent_near(args);
516 case XFS_ALLOCTYPE_THIS_BNO:
517 error = xfs_alloc_ag_vextent_exact(args);
524 if (error || args->agbno == NULLAGBLOCK)
527 ASSERT(args->len >= args->minlen);
528 ASSERT(args->len <= args->maxlen);
529 ASSERT(!args->wasfromfl || !args->isfl);
530 ASSERT(args->agbno % args->alignment == 0);
532 if (!args->wasfromfl) {
533 error = xfs_alloc_update_counters(args->tp, args->pag,
535 -((long)(args->len)));
539 ASSERT(!xfs_alloc_busy_search(args->mp, args->agno,
540 args->agbno, args->len));
544 xfs_trans_mod_sb(args->tp, args->wasdel ?
545 XFS_TRANS_SB_RES_FDBLOCKS :
546 XFS_TRANS_SB_FDBLOCKS,
547 -((long)(args->len)));
550 XFS_STATS_INC(xs_allocx);
551 XFS_STATS_ADD(xs_allocb, args->len);
556 * Allocate a variable extent at exactly agno/bno.
557 * Extent's length (returned in *len) will be between minlen and maxlen,
558 * and of the form k * prod + mod unless there's nothing that large.
559 * Return the starting a.g. block (bno), or NULLAGBLOCK if we can't do it.
561 STATIC int /* error */
562 xfs_alloc_ag_vextent_exact(
563 xfs_alloc_arg_t *args) /* allocation argument structure */
565 xfs_btree_cur_t *bno_cur;/* by block-number btree cursor */
566 xfs_btree_cur_t *cnt_cur;/* by count btree cursor */
568 xfs_agblock_t fbno; /* start block of found extent */
569 xfs_extlen_t flen; /* length of found extent */
570 xfs_agblock_t tbno; /* start block of trimmed extent */
571 xfs_extlen_t tlen; /* length of trimmed extent */
572 xfs_agblock_t tend; /* end block of trimmed extent */
573 int i; /* success/failure of operation */
575 ASSERT(args->alignment == 1);
578 * Allocate/initialize a cursor for the by-number freespace btree.
580 bno_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
581 args->agno, XFS_BTNUM_BNO);
584 * Lookup bno and minlen in the btree (minlen is irrelevant, really).
585 * Look for the closest free block <= bno, it must contain bno
586 * if any free block does.
588 error = xfs_alloc_lookup_le(bno_cur, args->agbno, args->minlen, &i);
595 * Grab the freespace record.
597 error = xfs_alloc_get_rec(bno_cur, &fbno, &flen, &i);
600 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
601 ASSERT(fbno <= args->agbno);
604 * Check for overlapping busy extents.
606 xfs_alloc_busy_trim(args, fbno, flen, &tbno, &tlen);
609 * Give up if the start of the extent is busy, or the freespace isn't
610 * long enough for the minimum request.
612 if (tbno > args->agbno)
614 if (tlen < args->minlen)
617 if (tend < args->agbno + args->minlen)
621 * End of extent will be smaller of the freespace end and the
622 * maximal requested end.
624 * Fix the length according to mod and prod if given.
626 args->len = XFS_AGBLOCK_MIN(tend, args->agbno + args->maxlen)
628 xfs_alloc_fix_len(args);
629 if (!xfs_alloc_fix_minleft(args))
632 ASSERT(args->agbno + args->len <= tend);
635 * We are allocating agbno for args->len
636 * Allocate/initialize a cursor for the by-size btree.
638 cnt_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
639 args->agno, XFS_BTNUM_CNT);
640 ASSERT(args->agbno + args->len <=
641 be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length));
642 error = xfs_alloc_fixup_trees(cnt_cur, bno_cur, fbno, flen, args->agbno,
643 args->len, XFSA_FIXUP_BNO_OK);
645 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
649 xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
650 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
653 trace_xfs_alloc_exact_done(args);
657 /* Didn't find it, return null. */
658 xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
659 args->agbno = NULLAGBLOCK;
660 trace_xfs_alloc_exact_notfound(args);
664 xfs_btree_del_cursor(bno_cur, XFS_BTREE_ERROR);
665 trace_xfs_alloc_exact_error(args);
670 * Search the btree in a given direction via the search cursor and compare
671 * the records found against the good extent we've already found.
674 xfs_alloc_find_best_extent(
675 struct xfs_alloc_arg *args, /* allocation argument structure */
676 struct xfs_btree_cur **gcur, /* good cursor */
677 struct xfs_btree_cur **scur, /* searching cursor */
678 xfs_agblock_t gdiff, /* difference for search comparison */
679 xfs_agblock_t *sbno, /* extent found by search */
680 xfs_extlen_t *slen, /* extent length */
681 xfs_agblock_t *sbnoa, /* aligned extent found by search */
682 xfs_extlen_t *slena, /* aligned extent length */
683 int dir) /* 0 = search right, 1 = search left */
690 /* The good extent is perfect, no need to search. */
695 * Look until we find a better one, run out of space or run off the end.
698 error = xfs_alloc_get_rec(*scur, sbno, slen, &i);
701 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
702 xfs_alloc_compute_aligned(args, *sbno, *slen, sbnoa, slena);
705 * The good extent is closer than this one.
708 if (*sbnoa >= args->agbno + gdiff)
711 if (*sbnoa <= args->agbno - gdiff)
716 * Same distance, compare length and pick the best.
718 if (*slena >= args->minlen) {
719 args->len = XFS_EXTLEN_MIN(*slena, args->maxlen);
720 xfs_alloc_fix_len(args);
722 sdiff = xfs_alloc_compute_diff(args->agbno, args->len,
723 args->alignment, *sbnoa,
727 * Choose closer size and invalidate other cursor.
735 error = xfs_btree_increment(*scur, 0, &i);
737 error = xfs_btree_decrement(*scur, 0, &i);
743 xfs_btree_del_cursor(*scur, XFS_BTREE_NOERROR);
748 xfs_btree_del_cursor(*gcur, XFS_BTREE_NOERROR);
753 /* caller invalidates cursors */
758 * Allocate a variable extent near bno in the allocation group agno.
759 * Extent's length (returned in len) will be between minlen and maxlen,
760 * and of the form k * prod + mod unless there's nothing that large.
761 * Return the starting a.g. block, or NULLAGBLOCK if we can't do it.
763 STATIC int /* error */
764 xfs_alloc_ag_vextent_near(
765 xfs_alloc_arg_t *args) /* allocation argument structure */
767 xfs_btree_cur_t *bno_cur_gt; /* cursor for bno btree, right side */
768 xfs_btree_cur_t *bno_cur_lt; /* cursor for bno btree, left side */
769 xfs_btree_cur_t *cnt_cur; /* cursor for count btree */
770 xfs_agblock_t gtbno; /* start bno of right side entry */
771 xfs_agblock_t gtbnoa; /* aligned ... */
772 xfs_extlen_t gtdiff; /* difference to right side entry */
773 xfs_extlen_t gtlen; /* length of right side entry */
774 xfs_extlen_t gtlena; /* aligned ... */
775 xfs_agblock_t gtnew; /* useful start bno of right side */
776 int error; /* error code */
777 int i; /* result code, temporary */
778 int j; /* result code, temporary */
779 xfs_agblock_t ltbno; /* start bno of left side entry */
780 xfs_agblock_t ltbnoa; /* aligned ... */
781 xfs_extlen_t ltdiff; /* difference to left side entry */
782 xfs_extlen_t ltlen; /* length of left side entry */
783 xfs_extlen_t ltlena; /* aligned ... */
784 xfs_agblock_t ltnew; /* useful start bno of left side */
785 xfs_extlen_t rlen; /* length of returned extent */
787 #if defined(DEBUG) && defined(__KERNEL__)
789 * Randomly don't execute the first algorithm.
791 int dofirst; /* set to do first algorithm */
793 dofirst = random32() & 1;
804 * Get a cursor for the by-size btree.
806 cnt_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
807 args->agno, XFS_BTNUM_CNT);
810 * See if there are any free extents as big as maxlen.
812 if ((error = xfs_alloc_lookup_ge(cnt_cur, 0, args->maxlen, &i)))
815 * If none, then pick up the last entry in the tree unless the
819 if ((error = xfs_alloc_ag_vextent_small(args, cnt_cur, <bno,
822 if (i == 0 || ltlen == 0) {
823 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
824 trace_xfs_alloc_near_noentry(args);
833 * If the requested extent is large wrt the freespaces available
834 * in this a.g., then the cursor will be pointing to a btree entry
835 * near the right edge of the tree. If it's in the last btree leaf
836 * block, then we just examine all the entries in that block
837 * that are big enough, and pick the best one.
838 * This is written as a while loop so we can break out of it,
839 * but we never loop back to the top.
841 while (xfs_btree_islastblock(cnt_cur, 0)) {
845 xfs_agblock_t bnew=0;
847 #if defined(DEBUG) && defined(__KERNEL__)
852 * Start from the entry that lookup found, sequence through
853 * all larger free blocks. If we're actually pointing at a
854 * record smaller than maxlen, go to the start of this block,
855 * and skip all those smaller than minlen.
857 if (ltlen || args->alignment > 1) {
858 cnt_cur->bc_ptrs[0] = 1;
860 if ((error = xfs_alloc_get_rec(cnt_cur, <bno,
863 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
864 if (ltlen >= args->minlen)
866 if ((error = xfs_btree_increment(cnt_cur, 0, &i)))
869 ASSERT(ltlen >= args->minlen);
873 i = cnt_cur->bc_ptrs[0];
874 for (j = 1, blen = 0, bdiff = 0;
875 !error && j && (blen < args->maxlen || bdiff > 0);
876 error = xfs_btree_increment(cnt_cur, 0, &j)) {
878 * For each entry, decide if it's better than
879 * the previous best entry.
881 if ((error = xfs_alloc_get_rec(cnt_cur, <bno, <len, &i)))
883 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
884 xfs_alloc_compute_aligned(args, ltbno, ltlen,
886 if (ltlena < args->minlen)
888 args->len = XFS_EXTLEN_MIN(ltlena, args->maxlen);
889 xfs_alloc_fix_len(args);
890 ASSERT(args->len >= args->minlen);
891 if (args->len < blen)
893 ltdiff = xfs_alloc_compute_diff(args->agbno, args->len,
894 args->alignment, ltbnoa, ltlena, <new);
895 if (ltnew != NULLAGBLOCK &&
896 (args->len > blen || ltdiff < bdiff)) {
900 besti = cnt_cur->bc_ptrs[0];
904 * It didn't work. We COULD be in a case where
905 * there's a good record somewhere, so try again.
910 * Point at the best entry, and retrieve it again.
912 cnt_cur->bc_ptrs[0] = besti;
913 if ((error = xfs_alloc_get_rec(cnt_cur, <bno, <len, &i)))
915 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
916 ASSERT(ltbno + ltlen <= be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length));
918 if (!xfs_alloc_fix_minleft(args)) {
919 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
920 trace_xfs_alloc_near_nominleft(args);
925 * We are allocating starting at bnew for blen blocks.
928 ASSERT(bnew >= ltbno);
929 ASSERT(bnew + blen <= ltbno + ltlen);
931 * Set up a cursor for the by-bno tree.
933 bno_cur_lt = xfs_allocbt_init_cursor(args->mp, args->tp,
934 args->agbp, args->agno, XFS_BTNUM_BNO);
936 * Fix up the btree entries.
938 if ((error = xfs_alloc_fixup_trees(cnt_cur, bno_cur_lt, ltbno,
939 ltlen, bnew, blen, XFSA_FIXUP_CNT_OK)))
941 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
942 xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_NOERROR);
944 trace_xfs_alloc_near_first(args);
949 * Search in the by-bno tree to the left and to the right
950 * simultaneously, until in each case we find a space big enough,
951 * or run into the edge of the tree. When we run into the edge,
952 * we deallocate that cursor.
953 * If both searches succeed, we compare the two spaces and pick
955 * With alignment, it's possible for both to fail; the upper
956 * level algorithm that picks allocation groups for allocations
957 * is not supposed to do this.
960 * Allocate and initialize the cursor for the leftward search.
962 bno_cur_lt = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
963 args->agno, XFS_BTNUM_BNO);
965 * Lookup <= bno to find the leftward search's starting point.
967 if ((error = xfs_alloc_lookup_le(bno_cur_lt, args->agbno, args->maxlen, &i)))
971 * Didn't find anything; use this cursor for the rightward
974 bno_cur_gt = bno_cur_lt;
978 * Found something. Duplicate the cursor for the rightward search.
980 else if ((error = xfs_btree_dup_cursor(bno_cur_lt, &bno_cur_gt)))
983 * Increment the cursor, so we will point at the entry just right
984 * of the leftward entry if any, or to the leftmost entry.
986 if ((error = xfs_btree_increment(bno_cur_gt, 0, &i)))
990 * It failed, there are no rightward entries.
992 xfs_btree_del_cursor(bno_cur_gt, XFS_BTREE_NOERROR);
996 * Loop going left with the leftward cursor, right with the
997 * rightward cursor, until either both directions give up or
998 * we find an entry at least as big as minlen.
1002 if ((error = xfs_alloc_get_rec(bno_cur_lt, <bno, <len, &i)))
1004 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1005 xfs_alloc_compute_aligned(args, ltbno, ltlen,
1007 if (ltlena >= args->minlen)
1009 if ((error = xfs_btree_decrement(bno_cur_lt, 0, &i)))
1012 xfs_btree_del_cursor(bno_cur_lt,
1018 if ((error = xfs_alloc_get_rec(bno_cur_gt, >bno, >len, &i)))
1020 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1021 xfs_alloc_compute_aligned(args, gtbno, gtlen,
1023 if (gtlena >= args->minlen)
1025 if ((error = xfs_btree_increment(bno_cur_gt, 0, &i)))
1028 xfs_btree_del_cursor(bno_cur_gt,
1033 } while (bno_cur_lt || bno_cur_gt);
1036 * Got both cursors still active, need to find better entry.
1038 if (bno_cur_lt && bno_cur_gt) {
1039 if (ltlena >= args->minlen) {
1041 * Left side is good, look for a right side entry.
1043 args->len = XFS_EXTLEN_MIN(ltlena, args->maxlen);
1044 xfs_alloc_fix_len(args);
1045 ltdiff = xfs_alloc_compute_diff(args->agbno, args->len,
1046 args->alignment, ltbnoa, ltlena, <new);
1048 error = xfs_alloc_find_best_extent(args,
1049 &bno_cur_lt, &bno_cur_gt,
1050 ltdiff, >bno, >len,
1052 0 /* search right */);
1054 ASSERT(gtlena >= args->minlen);
1057 * Right side is good, look for a left side entry.
1059 args->len = XFS_EXTLEN_MIN(gtlena, args->maxlen);
1060 xfs_alloc_fix_len(args);
1061 gtdiff = xfs_alloc_compute_diff(args->agbno, args->len,
1062 args->alignment, gtbnoa, gtlena, >new);
1064 error = xfs_alloc_find_best_extent(args,
1065 &bno_cur_gt, &bno_cur_lt,
1066 gtdiff, <bno, <len,
1068 1 /* search left */);
1076 * If we couldn't get anything, give up.
1078 if (bno_cur_lt == NULL && bno_cur_gt == NULL) {
1080 trace_xfs_alloc_near_busy(args);
1081 xfs_log_force(args->mp, XFS_LOG_SYNC);
1085 trace_xfs_alloc_size_neither(args);
1086 args->agbno = NULLAGBLOCK;
1091 * At this point we have selected a freespace entry, either to the
1092 * left or to the right. If it's on the right, copy all the
1093 * useful variables to the "left" set so we only have one
1094 * copy of this code.
1097 bno_cur_lt = bno_cur_gt;
1108 * Fix up the length and compute the useful address.
1110 args->len = XFS_EXTLEN_MIN(ltlena, args->maxlen);
1111 xfs_alloc_fix_len(args);
1112 if (!xfs_alloc_fix_minleft(args)) {
1113 trace_xfs_alloc_near_nominleft(args);
1114 xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_NOERROR);
1115 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
1119 (void)xfs_alloc_compute_diff(args->agbno, rlen, args->alignment,
1120 ltbnoa, ltlena, <new);
1121 ASSERT(ltnew >= ltbno);
1122 ASSERT(ltnew + rlen <= ltbnoa + ltlena);
1123 ASSERT(ltnew + rlen <= be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length));
1124 args->agbno = ltnew;
1126 if ((error = xfs_alloc_fixup_trees(cnt_cur, bno_cur_lt, ltbno, ltlen,
1127 ltnew, rlen, XFSA_FIXUP_BNO_OK)))
1131 trace_xfs_alloc_near_greater(args);
1133 trace_xfs_alloc_near_lesser(args);
1135 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
1136 xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_NOERROR);
1140 trace_xfs_alloc_near_error(args);
1141 if (cnt_cur != NULL)
1142 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
1143 if (bno_cur_lt != NULL)
1144 xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_ERROR);
1145 if (bno_cur_gt != NULL)
1146 xfs_btree_del_cursor(bno_cur_gt, XFS_BTREE_ERROR);
1151 * Allocate a variable extent anywhere in the allocation group agno.
1152 * Extent's length (returned in len) will be between minlen and maxlen,
1153 * and of the form k * prod + mod unless there's nothing that large.
1154 * Return the starting a.g. block, or NULLAGBLOCK if we can't do it.
1156 STATIC int /* error */
1157 xfs_alloc_ag_vextent_size(
1158 xfs_alloc_arg_t *args) /* allocation argument structure */
1160 xfs_btree_cur_t *bno_cur; /* cursor for bno btree */
1161 xfs_btree_cur_t *cnt_cur; /* cursor for cnt btree */
1162 int error; /* error result */
1163 xfs_agblock_t fbno; /* start of found freespace */
1164 xfs_extlen_t flen; /* length of found freespace */
1165 int i; /* temp status variable */
1166 xfs_agblock_t rbno; /* returned block number */
1167 xfs_extlen_t rlen; /* length of returned extent */
1172 * Allocate and initialize a cursor for the by-size btree.
1174 cnt_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
1175 args->agno, XFS_BTNUM_CNT);
1179 * Look for an entry >= maxlen+alignment-1 blocks.
1181 if ((error = xfs_alloc_lookup_ge(cnt_cur, 0,
1182 args->maxlen + args->alignment - 1, &i)))
1186 * If none or we have busy extents that we cannot allocate from, then
1187 * we have to settle for a smaller extent. In the case that there are
1188 * no large extents, this will return the last entry in the tree unless
1189 * the tree is empty. In the case that there are only busy large
1190 * extents, this will return the largest small extent unless there
1191 * are no smaller extents available.
1193 if (!i || forced > 1) {
1194 error = xfs_alloc_ag_vextent_small(args, cnt_cur,
1198 if (i == 0 || flen == 0) {
1199 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
1200 trace_xfs_alloc_size_noentry(args);
1204 xfs_alloc_compute_aligned(args, fbno, flen, &rbno, &rlen);
1207 * Search for a non-busy extent that is large enough.
1208 * If we are at low space, don't check, or if we fall of
1209 * the end of the btree, turn off the busy check and
1213 error = xfs_alloc_get_rec(cnt_cur, &fbno, &flen, &i);
1216 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1218 xfs_alloc_compute_aligned(args, fbno, flen,
1221 if (rlen >= args->maxlen)
1224 error = xfs_btree_increment(cnt_cur, 0, &i);
1229 * Our only valid extents must have been busy.
1230 * Make it unbusy by forcing the log out and
1231 * retrying. If we've been here before, forcing
1232 * the log isn't making the extents available,
1233 * which means they have probably been freed in
1234 * this transaction. In that case, we have to
1235 * give up on them and we'll attempt a minlen
1236 * allocation the next time around.
1238 xfs_btree_del_cursor(cnt_cur,
1240 trace_xfs_alloc_size_busy(args);
1242 xfs_log_force(args->mp, XFS_LOG_SYNC);
1249 * In the first case above, we got the last entry in the
1250 * by-size btree. Now we check to see if the space hits maxlen
1251 * once aligned; if not, we search left for something better.
1252 * This can't happen in the second case above.
1254 rlen = XFS_EXTLEN_MIN(args->maxlen, rlen);
1255 XFS_WANT_CORRUPTED_GOTO(rlen == 0 ||
1256 (rlen <= flen && rbno + rlen <= fbno + flen), error0);
1257 if (rlen < args->maxlen) {
1258 xfs_agblock_t bestfbno;
1259 xfs_extlen_t bestflen;
1260 xfs_agblock_t bestrbno;
1261 xfs_extlen_t bestrlen;
1268 if ((error = xfs_btree_decrement(cnt_cur, 0, &i)))
1272 if ((error = xfs_alloc_get_rec(cnt_cur, &fbno, &flen,
1275 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1276 if (flen < bestrlen)
1278 xfs_alloc_compute_aligned(args, fbno, flen,
1280 rlen = XFS_EXTLEN_MIN(args->maxlen, rlen);
1281 XFS_WANT_CORRUPTED_GOTO(rlen == 0 ||
1282 (rlen <= flen && rbno + rlen <= fbno + flen),
1284 if (rlen > bestrlen) {
1289 if (rlen == args->maxlen)
1293 if ((error = xfs_alloc_lookup_eq(cnt_cur, bestfbno, bestflen,
1296 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1302 args->wasfromfl = 0;
1304 * Fix up the length.
1307 if (rlen < args->minlen) {
1309 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
1310 trace_xfs_alloc_size_busy(args);
1311 xfs_log_force(args->mp, XFS_LOG_SYNC);
1316 xfs_alloc_fix_len(args);
1318 if (!xfs_alloc_fix_minleft(args))
1321 XFS_WANT_CORRUPTED_GOTO(rlen <= flen, error0);
1323 * Allocate and initialize a cursor for the by-block tree.
1325 bno_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
1326 args->agno, XFS_BTNUM_BNO);
1327 if ((error = xfs_alloc_fixup_trees(cnt_cur, bno_cur, fbno, flen,
1328 rbno, rlen, XFSA_FIXUP_CNT_OK)))
1330 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
1331 xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
1332 cnt_cur = bno_cur = NULL;
1335 XFS_WANT_CORRUPTED_GOTO(
1336 args->agbno + args->len <=
1337 be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length),
1339 trace_xfs_alloc_size_done(args);
1343 trace_xfs_alloc_size_error(args);
1345 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
1347 xfs_btree_del_cursor(bno_cur, XFS_BTREE_ERROR);
1351 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
1352 trace_xfs_alloc_size_nominleft(args);
1353 args->agbno = NULLAGBLOCK;
1358 * Deal with the case where only small freespaces remain.
1359 * Either return the contents of the last freespace record,
1360 * or allocate space from the freelist if there is nothing in the tree.
1362 STATIC int /* error */
1363 xfs_alloc_ag_vextent_small(
1364 xfs_alloc_arg_t *args, /* allocation argument structure */
1365 xfs_btree_cur_t *ccur, /* by-size cursor */
1366 xfs_agblock_t *fbnop, /* result block number */
1367 xfs_extlen_t *flenp, /* result length */
1368 int *stat) /* status: 0-freelist, 1-normal/none */
1375 if ((error = xfs_btree_decrement(ccur, 0, &i)))
1378 if ((error = xfs_alloc_get_rec(ccur, &fbno, &flen, &i)))
1380 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1383 * Nothing in the btree, try the freelist. Make sure
1384 * to respect minleft even when pulling from the
1387 else if (args->minlen == 1 && args->alignment == 1 && !args->isfl &&
1388 (be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_flcount)
1390 error = xfs_alloc_get_freelist(args->tp, args->agbp, &fbno, 0);
1393 if (fbno != NULLAGBLOCK) {
1394 xfs_alloc_busy_reuse(args->mp, args->agno, fbno, 1,
1397 if (args->userdata) {
1400 bp = xfs_btree_get_bufs(args->mp, args->tp,
1401 args->agno, fbno, 0);
1402 xfs_trans_binval(args->tp, bp);
1406 XFS_WANT_CORRUPTED_GOTO(
1407 args->agbno + args->len <=
1408 be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length),
1410 args->wasfromfl = 1;
1411 trace_xfs_alloc_small_freelist(args);
1416 * Nothing in the freelist.
1422 * Can't allocate from the freelist for some reason.
1429 * Can't do the allocation, give up.
1431 if (flen < args->minlen) {
1432 args->agbno = NULLAGBLOCK;
1433 trace_xfs_alloc_small_notenough(args);
1439 trace_xfs_alloc_small_done(args);
1443 trace_xfs_alloc_small_error(args);
1448 * Free the extent starting at agno/bno for length.
1450 STATIC int /* error */
1452 xfs_trans_t *tp, /* transaction pointer */
1453 xfs_buf_t *agbp, /* buffer for a.g. freelist header */
1454 xfs_agnumber_t agno, /* allocation group number */
1455 xfs_agblock_t bno, /* starting block number */
1456 xfs_extlen_t len, /* length of extent */
1457 int isfl) /* set if is freelist blocks - no sb acctg */
1459 xfs_btree_cur_t *bno_cur; /* cursor for by-block btree */
1460 xfs_btree_cur_t *cnt_cur; /* cursor for by-size btree */
1461 int error; /* error return value */
1462 xfs_agblock_t gtbno; /* start of right neighbor block */
1463 xfs_extlen_t gtlen; /* length of right neighbor block */
1464 int haveleft; /* have a left neighbor block */
1465 int haveright; /* have a right neighbor block */
1466 int i; /* temp, result code */
1467 xfs_agblock_t ltbno; /* start of left neighbor block */
1468 xfs_extlen_t ltlen; /* length of left neighbor block */
1469 xfs_mount_t *mp; /* mount point struct for filesystem */
1470 xfs_agblock_t nbno; /* new starting block of freespace */
1471 xfs_extlen_t nlen; /* new length of freespace */
1472 xfs_perag_t *pag; /* per allocation group data */
1476 * Allocate and initialize a cursor for the by-block btree.
1478 bno_cur = xfs_allocbt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_BNO);
1481 * Look for a neighboring block on the left (lower block numbers)
1482 * that is contiguous with this space.
1484 if ((error = xfs_alloc_lookup_le(bno_cur, bno, len, &haveleft)))
1488 * There is a block to our left.
1490 if ((error = xfs_alloc_get_rec(bno_cur, <bno, <len, &i)))
1492 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1494 * It's not contiguous, though.
1496 if (ltbno + ltlen < bno)
1500 * If this failure happens the request to free this
1501 * space was invalid, it's (partly) already free.
1504 XFS_WANT_CORRUPTED_GOTO(ltbno + ltlen <= bno, error0);
1508 * Look for a neighboring block on the right (higher block numbers)
1509 * that is contiguous with this space.
1511 if ((error = xfs_btree_increment(bno_cur, 0, &haveright)))
1515 * There is a block to our right.
1517 if ((error = xfs_alloc_get_rec(bno_cur, >bno, >len, &i)))
1519 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1521 * It's not contiguous, though.
1523 if (bno + len < gtbno)
1527 * If this failure happens the request to free this
1528 * space was invalid, it's (partly) already free.
1531 XFS_WANT_CORRUPTED_GOTO(gtbno >= bno + len, error0);
1535 * Now allocate and initialize a cursor for the by-size tree.
1537 cnt_cur = xfs_allocbt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_CNT);
1539 * Have both left and right contiguous neighbors.
1540 * Merge all three into a single free block.
1542 if (haveleft && haveright) {
1544 * Delete the old by-size entry on the left.
1546 if ((error = xfs_alloc_lookup_eq(cnt_cur, ltbno, ltlen, &i)))
1548 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1549 if ((error = xfs_btree_delete(cnt_cur, &i)))
1551 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1553 * Delete the old by-size entry on the right.
1555 if ((error = xfs_alloc_lookup_eq(cnt_cur, gtbno, gtlen, &i)))
1557 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1558 if ((error = xfs_btree_delete(cnt_cur, &i)))
1560 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1562 * Delete the old by-block entry for the right block.
1564 if ((error = xfs_btree_delete(bno_cur, &i)))
1566 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1568 * Move the by-block cursor back to the left neighbor.
1570 if ((error = xfs_btree_decrement(bno_cur, 0, &i)))
1572 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1575 * Check that this is the right record: delete didn't
1576 * mangle the cursor.
1579 xfs_agblock_t xxbno;
1582 if ((error = xfs_alloc_get_rec(bno_cur, &xxbno, &xxlen,
1585 XFS_WANT_CORRUPTED_GOTO(
1586 i == 1 && xxbno == ltbno && xxlen == ltlen,
1591 * Update remaining by-block entry to the new, joined block.
1594 nlen = len + ltlen + gtlen;
1595 if ((error = xfs_alloc_update(bno_cur, nbno, nlen)))
1599 * Have only a left contiguous neighbor.
1600 * Merge it together with the new freespace.
1602 else if (haveleft) {
1604 * Delete the old by-size entry on the left.
1606 if ((error = xfs_alloc_lookup_eq(cnt_cur, ltbno, ltlen, &i)))
1608 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1609 if ((error = xfs_btree_delete(cnt_cur, &i)))
1611 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1613 * Back up the by-block cursor to the left neighbor, and
1614 * update its length.
1616 if ((error = xfs_btree_decrement(bno_cur, 0, &i)))
1618 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1621 if ((error = xfs_alloc_update(bno_cur, nbno, nlen)))
1625 * Have only a right contiguous neighbor.
1626 * Merge it together with the new freespace.
1628 else if (haveright) {
1630 * Delete the old by-size entry on the right.
1632 if ((error = xfs_alloc_lookup_eq(cnt_cur, gtbno, gtlen, &i)))
1634 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1635 if ((error = xfs_btree_delete(cnt_cur, &i)))
1637 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1639 * Update the starting block and length of the right
1640 * neighbor in the by-block tree.
1644 if ((error = xfs_alloc_update(bno_cur, nbno, nlen)))
1648 * No contiguous neighbors.
1649 * Insert the new freespace into the by-block tree.
1654 if ((error = xfs_btree_insert(bno_cur, &i)))
1656 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1658 xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
1661 * In all cases we need to insert the new freespace in the by-size tree.
1663 if ((error = xfs_alloc_lookup_eq(cnt_cur, nbno, nlen, &i)))
1665 XFS_WANT_CORRUPTED_GOTO(i == 0, error0);
1666 if ((error = xfs_btree_insert(cnt_cur, &i)))
1668 XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
1669 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
1673 * Update the freespace totals in the ag and superblock.
1675 pag = xfs_perag_get(mp, agno);
1676 error = xfs_alloc_update_counters(tp, pag, agbp, len);
1682 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (long)len);
1683 XFS_STATS_INC(xs_freex);
1684 XFS_STATS_ADD(xs_freeb, len);
1686 trace_xfs_free_extent(mp, agno, bno, len, isfl, haveleft, haveright);
1691 trace_xfs_free_extent(mp, agno, bno, len, isfl, -1, -1);
1693 xfs_btree_del_cursor(bno_cur, XFS_BTREE_ERROR);
1695 xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
1700 * Visible (exported) allocation/free functions.
1701 * Some of these are used just by xfs_alloc_btree.c and this file.
1705 * Compute and fill in value of m_ag_maxlevels.
1708 xfs_alloc_compute_maxlevels(
1709 xfs_mount_t *mp) /* file system mount structure */
1717 maxleafents = (mp->m_sb.sb_agblocks + 1) / 2;
1718 minleafrecs = mp->m_alloc_mnr[0];
1719 minnoderecs = mp->m_alloc_mnr[1];
1720 maxblocks = (maxleafents + minleafrecs - 1) / minleafrecs;
1721 for (level = 1; maxblocks > 1; level++)
1722 maxblocks = (maxblocks + minnoderecs - 1) / minnoderecs;
1723 mp->m_ag_maxlevels = level;
1727 * Find the length of the longest extent in an AG.
1730 xfs_alloc_longest_free_extent(
1731 struct xfs_mount *mp,
1732 struct xfs_perag *pag)
1734 xfs_extlen_t need, delta = 0;
1736 need = XFS_MIN_FREELIST_PAG(pag, mp);
1737 if (need > pag->pagf_flcount)
1738 delta = need - pag->pagf_flcount;
1740 if (pag->pagf_longest > delta)
1741 return pag->pagf_longest - delta;
1742 return pag->pagf_flcount > 0 || pag->pagf_longest > 0;
1746 * Decide whether to use this allocation group for this allocation.
1747 * If so, fix up the btree freelist's size.
1749 STATIC int /* error */
1750 xfs_alloc_fix_freelist(
1751 xfs_alloc_arg_t *args, /* allocation argument structure */
1752 int flags) /* XFS_ALLOC_FLAG_... */
1754 xfs_buf_t *agbp; /* agf buffer pointer */
1755 xfs_agf_t *agf; /* a.g. freespace structure pointer */
1756 xfs_buf_t *agflbp;/* agfl buffer pointer */
1757 xfs_agblock_t bno; /* freelist block */
1758 xfs_extlen_t delta; /* new blocks needed in freelist */
1759 int error; /* error result code */
1760 xfs_extlen_t longest;/* longest extent in allocation group */
1761 xfs_mount_t *mp; /* file system mount point structure */
1762 xfs_extlen_t need; /* total blocks needed in freelist */
1763 xfs_perag_t *pag; /* per-ag information structure */
1764 xfs_alloc_arg_t targs; /* local allocation arguments */
1765 xfs_trans_t *tp; /* transaction pointer */
1771 if (!pag->pagf_init) {
1772 if ((error = xfs_alloc_read_agf(mp, tp, args->agno, flags,
1775 if (!pag->pagf_init) {
1776 ASSERT(flags & XFS_ALLOC_FLAG_TRYLOCK);
1777 ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
1785 * If this is a metadata preferred pag and we are user data
1786 * then try somewhere else if we are not being asked to
1787 * try harder at this point
1789 if (pag->pagf_metadata && args->userdata &&
1790 (flags & XFS_ALLOC_FLAG_TRYLOCK)) {
1791 ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
1796 if (!(flags & XFS_ALLOC_FLAG_FREEING)) {
1798 * If it looks like there isn't a long enough extent, or enough
1799 * total blocks, reject it.
1801 need = XFS_MIN_FREELIST_PAG(pag, mp);
1802 longest = xfs_alloc_longest_free_extent(mp, pag);
1803 if ((args->minlen + args->alignment + args->minalignslop - 1) >
1805 ((int)(pag->pagf_freeblks + pag->pagf_flcount -
1806 need - args->total) < (int)args->minleft)) {
1808 xfs_trans_brelse(tp, agbp);
1815 * Get the a.g. freespace buffer.
1816 * Can fail if we're not blocking on locks, and it's held.
1819 if ((error = xfs_alloc_read_agf(mp, tp, args->agno, flags,
1823 ASSERT(flags & XFS_ALLOC_FLAG_TRYLOCK);
1824 ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
1830 * Figure out how many blocks we should have in the freelist.
1832 agf = XFS_BUF_TO_AGF(agbp);
1833 need = XFS_MIN_FREELIST(agf, mp);
1835 * If there isn't enough total or single-extent, reject it.
1837 if (!(flags & XFS_ALLOC_FLAG_FREEING)) {
1838 delta = need > be32_to_cpu(agf->agf_flcount) ?
1839 (need - be32_to_cpu(agf->agf_flcount)) : 0;
1840 longest = be32_to_cpu(agf->agf_longest);
1841 longest = (longest > delta) ? (longest - delta) :
1842 (be32_to_cpu(agf->agf_flcount) > 0 || longest > 0);
1843 if ((args->minlen + args->alignment + args->minalignslop - 1) >
1845 ((int)(be32_to_cpu(agf->agf_freeblks) +
1846 be32_to_cpu(agf->agf_flcount) - need - args->total) <
1847 (int)args->minleft)) {
1848 xfs_trans_brelse(tp, agbp);
1854 * Make the freelist shorter if it's too long.
1856 while (be32_to_cpu(agf->agf_flcount) > need) {
1859 error = xfs_alloc_get_freelist(tp, agbp, &bno, 0);
1862 if ((error = xfs_free_ag_extent(tp, agbp, args->agno, bno, 1, 1)))
1864 bp = xfs_btree_get_bufs(mp, tp, args->agno, bno, 0);
1865 xfs_trans_binval(tp, bp);
1868 * Initialize the args structure.
1873 targs.agno = args->agno;
1874 targs.mod = targs.minleft = targs.wasdel = targs.userdata =
1875 targs.minalignslop = 0;
1876 targs.alignment = targs.minlen = targs.prod = targs.isfl = 1;
1877 targs.type = XFS_ALLOCTYPE_THIS_AG;
1879 if ((error = xfs_alloc_read_agfl(mp, tp, targs.agno, &agflbp)))
1882 * Make the freelist longer if it's too short.
1884 while (be32_to_cpu(agf->agf_flcount) < need) {
1886 targs.maxlen = need - be32_to_cpu(agf->agf_flcount);
1888 * Allocate as many blocks as possible at once.
1890 if ((error = xfs_alloc_ag_vextent(&targs))) {
1891 xfs_trans_brelse(tp, agflbp);
1895 * Stop if we run out. Won't happen if callers are obeying
1896 * the restrictions correctly. Can happen for free calls
1897 * on a completely full ag.
1899 if (targs.agbno == NULLAGBLOCK) {
1900 if (flags & XFS_ALLOC_FLAG_FREEING)
1902 xfs_trans_brelse(tp, agflbp);
1907 * Put each allocated block on the list.
1909 for (bno = targs.agbno; bno < targs.agbno + targs.len; bno++) {
1910 error = xfs_alloc_put_freelist(tp, agbp,
1916 xfs_trans_brelse(tp, agflbp);
1922 * Get a block from the freelist.
1923 * Returns with the buffer for the block gotten.
1926 xfs_alloc_get_freelist(
1927 xfs_trans_t *tp, /* transaction pointer */
1928 xfs_buf_t *agbp, /* buffer containing the agf structure */
1929 xfs_agblock_t *bnop, /* block address retrieved from freelist */
1930 int btreeblk) /* destination is a AGF btree */
1932 xfs_agf_t *agf; /* a.g. freespace structure */
1933 xfs_agfl_t *agfl; /* a.g. freelist structure */
1934 xfs_buf_t *agflbp;/* buffer for a.g. freelist structure */
1935 xfs_agblock_t bno; /* block number returned */
1938 xfs_mount_t *mp; /* mount structure */
1939 xfs_perag_t *pag; /* per allocation group data */
1941 agf = XFS_BUF_TO_AGF(agbp);
1943 * Freelist is empty, give up.
1945 if (!agf->agf_flcount) {
1946 *bnop = NULLAGBLOCK;
1950 * Read the array of free blocks.
1953 if ((error = xfs_alloc_read_agfl(mp, tp,
1954 be32_to_cpu(agf->agf_seqno), &agflbp)))
1956 agfl = XFS_BUF_TO_AGFL(agflbp);
1958 * Get the block number and update the data structures.
1960 bno = be32_to_cpu(agfl->agfl_bno[be32_to_cpu(agf->agf_flfirst)]);
1961 be32_add_cpu(&agf->agf_flfirst, 1);
1962 xfs_trans_brelse(tp, agflbp);
1963 if (be32_to_cpu(agf->agf_flfirst) == XFS_AGFL_SIZE(mp))
1964 agf->agf_flfirst = 0;
1966 pag = xfs_perag_get(mp, be32_to_cpu(agf->agf_seqno));
1967 be32_add_cpu(&agf->agf_flcount, -1);
1968 xfs_trans_agflist_delta(tp, -1);
1969 pag->pagf_flcount--;
1972 logflags = XFS_AGF_FLFIRST | XFS_AGF_FLCOUNT;
1974 be32_add_cpu(&agf->agf_btreeblks, 1);
1975 pag->pagf_btreeblks++;
1976 logflags |= XFS_AGF_BTREEBLKS;
1979 xfs_alloc_log_agf(tp, agbp, logflags);
1986 * Log the given fields from the agf structure.
1990 xfs_trans_t *tp, /* transaction pointer */
1991 xfs_buf_t *bp, /* buffer for a.g. freelist header */
1992 int fields) /* mask of fields to be logged (XFS_AGF_...) */
1994 int first; /* first byte offset */
1995 int last; /* last byte offset */
1996 static const short offsets[] = {
1997 offsetof(xfs_agf_t, agf_magicnum),
1998 offsetof(xfs_agf_t, agf_versionnum),
1999 offsetof(xfs_agf_t, agf_seqno),
2000 offsetof(xfs_agf_t, agf_length),
2001 offsetof(xfs_agf_t, agf_roots[0]),
2002 offsetof(xfs_agf_t, agf_levels[0]),
2003 offsetof(xfs_agf_t, agf_flfirst),
2004 offsetof(xfs_agf_t, agf_fllast),
2005 offsetof(xfs_agf_t, agf_flcount),
2006 offsetof(xfs_agf_t, agf_freeblks),
2007 offsetof(xfs_agf_t, agf_longest),
2008 offsetof(xfs_agf_t, agf_btreeblks),
2012 trace_xfs_agf(tp->t_mountp, XFS_BUF_TO_AGF(bp), fields, _RET_IP_);
2014 xfs_btree_offsets(fields, offsets, XFS_AGF_NUM_BITS, &first, &last);
2015 xfs_trans_log_buf(tp, bp, (uint)first, (uint)last);
2019 * Interface for inode allocation to force the pag data to be initialized.
2022 xfs_alloc_pagf_init(
2023 xfs_mount_t *mp, /* file system mount structure */
2024 xfs_trans_t *tp, /* transaction pointer */
2025 xfs_agnumber_t agno, /* allocation group number */
2026 int flags) /* XFS_ALLOC_FLAGS_... */
2031 if ((error = xfs_alloc_read_agf(mp, tp, agno, flags, &bp)))
2034 xfs_trans_brelse(tp, bp);
2039 * Put the block on the freelist for the allocation group.
2042 xfs_alloc_put_freelist(
2043 xfs_trans_t *tp, /* transaction pointer */
2044 xfs_buf_t *agbp, /* buffer for a.g. freelist header */
2045 xfs_buf_t *agflbp,/* buffer for a.g. free block array */
2046 xfs_agblock_t bno, /* block being freed */
2047 int btreeblk) /* block came from a AGF btree */
2049 xfs_agf_t *agf; /* a.g. freespace structure */
2050 xfs_agfl_t *agfl; /* a.g. free block array */
2051 __be32 *blockp;/* pointer to array entry */
2054 xfs_mount_t *mp; /* mount structure */
2055 xfs_perag_t *pag; /* per allocation group data */
2057 agf = XFS_BUF_TO_AGF(agbp);
2060 if (!agflbp && (error = xfs_alloc_read_agfl(mp, tp,
2061 be32_to_cpu(agf->agf_seqno), &agflbp)))
2063 agfl = XFS_BUF_TO_AGFL(agflbp);
2064 be32_add_cpu(&agf->agf_fllast, 1);
2065 if (be32_to_cpu(agf->agf_fllast) == XFS_AGFL_SIZE(mp))
2066 agf->agf_fllast = 0;
2068 pag = xfs_perag_get(mp, be32_to_cpu(agf->agf_seqno));
2069 be32_add_cpu(&agf->agf_flcount, 1);
2070 xfs_trans_agflist_delta(tp, 1);
2071 pag->pagf_flcount++;
2073 logflags = XFS_AGF_FLLAST | XFS_AGF_FLCOUNT;
2075 be32_add_cpu(&agf->agf_btreeblks, -1);
2076 pag->pagf_btreeblks--;
2077 logflags |= XFS_AGF_BTREEBLKS;
2081 xfs_alloc_log_agf(tp, agbp, logflags);
2083 ASSERT(be32_to_cpu(agf->agf_flcount) <= XFS_AGFL_SIZE(mp));
2084 blockp = &agfl->agfl_bno[be32_to_cpu(agf->agf_fllast)];
2085 *blockp = cpu_to_be32(bno);
2086 xfs_alloc_log_agf(tp, agbp, logflags);
2087 xfs_trans_log_buf(tp, agflbp,
2088 (int)((xfs_caddr_t)blockp - (xfs_caddr_t)agfl),
2089 (int)((xfs_caddr_t)blockp - (xfs_caddr_t)agfl +
2090 sizeof(xfs_agblock_t) - 1));
2095 * Read in the allocation group header (free/alloc section).
2099 struct xfs_mount *mp, /* mount point structure */
2100 struct xfs_trans *tp, /* transaction pointer */
2101 xfs_agnumber_t agno, /* allocation group number */
2102 int flags, /* XFS_BUF_ */
2103 struct xfs_buf **bpp) /* buffer for the ag freelist header */
2105 struct xfs_agf *agf; /* ag freelist header */
2106 int agf_ok; /* set if agf is consistent */
2109 ASSERT(agno != NULLAGNUMBER);
2110 error = xfs_trans_read_buf(
2111 mp, tp, mp->m_ddev_targp,
2112 XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp)),
2113 XFS_FSS_TO_BB(mp, 1), flags, bpp);
2119 ASSERT(!(*bpp)->b_error);
2120 agf = XFS_BUF_TO_AGF(*bpp);
2123 * Validate the magic number of the agf block.
2126 agf->agf_magicnum == cpu_to_be32(XFS_AGF_MAGIC) &&
2127 XFS_AGF_GOOD_VERSION(be32_to_cpu(agf->agf_versionnum)) &&
2128 be32_to_cpu(agf->agf_freeblks) <= be32_to_cpu(agf->agf_length) &&
2129 be32_to_cpu(agf->agf_flfirst) < XFS_AGFL_SIZE(mp) &&
2130 be32_to_cpu(agf->agf_fllast) < XFS_AGFL_SIZE(mp) &&
2131 be32_to_cpu(agf->agf_flcount) <= XFS_AGFL_SIZE(mp) &&
2132 be32_to_cpu(agf->agf_seqno) == agno;
2133 if (xfs_sb_version_haslazysbcount(&mp->m_sb))
2134 agf_ok = agf_ok && be32_to_cpu(agf->agf_btreeblks) <=
2135 be32_to_cpu(agf->agf_length);
2136 if (unlikely(XFS_TEST_ERROR(!agf_ok, mp, XFS_ERRTAG_ALLOC_READ_AGF,
2137 XFS_RANDOM_ALLOC_READ_AGF))) {
2138 XFS_CORRUPTION_ERROR("xfs_alloc_read_agf",
2139 XFS_ERRLEVEL_LOW, mp, agf);
2140 xfs_trans_brelse(tp, *bpp);
2141 return XFS_ERROR(EFSCORRUPTED);
2143 xfs_buf_set_ref(*bpp, XFS_AGF_REF);
2148 * Read in the allocation group header (free/alloc section).
2152 struct xfs_mount *mp, /* mount point structure */
2153 struct xfs_trans *tp, /* transaction pointer */
2154 xfs_agnumber_t agno, /* allocation group number */
2155 int flags, /* XFS_ALLOC_FLAG_... */
2156 struct xfs_buf **bpp) /* buffer for the ag freelist header */
2158 struct xfs_agf *agf; /* ag freelist header */
2159 struct xfs_perag *pag; /* per allocation group data */
2162 ASSERT(agno != NULLAGNUMBER);
2164 error = xfs_read_agf(mp, tp, agno,
2165 (flags & XFS_ALLOC_FLAG_TRYLOCK) ? XBF_TRYLOCK : 0,
2171 ASSERT(!(*bpp)->b_error);
2173 agf = XFS_BUF_TO_AGF(*bpp);
2174 pag = xfs_perag_get(mp, agno);
2175 if (!pag->pagf_init) {
2176 pag->pagf_freeblks = be32_to_cpu(agf->agf_freeblks);
2177 pag->pagf_btreeblks = be32_to_cpu(agf->agf_btreeblks);
2178 pag->pagf_flcount = be32_to_cpu(agf->agf_flcount);
2179 pag->pagf_longest = be32_to_cpu(agf->agf_longest);
2180 pag->pagf_levels[XFS_BTNUM_BNOi] =
2181 be32_to_cpu(agf->agf_levels[XFS_BTNUM_BNOi]);
2182 pag->pagf_levels[XFS_BTNUM_CNTi] =
2183 be32_to_cpu(agf->agf_levels[XFS_BTNUM_CNTi]);
2184 spin_lock_init(&pag->pagb_lock);
2185 pag->pagb_count = 0;
2186 pag->pagb_tree = RB_ROOT;
2190 else if (!XFS_FORCED_SHUTDOWN(mp)) {
2191 ASSERT(pag->pagf_freeblks == be32_to_cpu(agf->agf_freeblks));
2192 ASSERT(pag->pagf_btreeblks == be32_to_cpu(agf->agf_btreeblks));
2193 ASSERT(pag->pagf_flcount == be32_to_cpu(agf->agf_flcount));
2194 ASSERT(pag->pagf_longest == be32_to_cpu(agf->agf_longest));
2195 ASSERT(pag->pagf_levels[XFS_BTNUM_BNOi] ==
2196 be32_to_cpu(agf->agf_levels[XFS_BTNUM_BNOi]));
2197 ASSERT(pag->pagf_levels[XFS_BTNUM_CNTi] ==
2198 be32_to_cpu(agf->agf_levels[XFS_BTNUM_CNTi]));
2206 * Allocate an extent (variable-size).
2207 * Depending on the allocation type, we either look in a single allocation
2208 * group or loop over the allocation groups to find the result.
2211 __xfs_alloc_vextent(
2212 xfs_alloc_arg_t *args) /* allocation argument structure */
2214 xfs_agblock_t agsize; /* allocation group size */
2216 int flags; /* XFS_ALLOC_FLAG_... locking flags */
2217 xfs_extlen_t minleft;/* minimum left value, temp copy */
2218 xfs_mount_t *mp; /* mount structure pointer */
2219 xfs_agnumber_t sagno; /* starting allocation group number */
2220 xfs_alloctype_t type; /* input allocation type */
2223 xfs_agnumber_t rotorstep = xfs_rotorstep; /* inode32 agf stepper */
2226 type = args->otype = args->type;
2227 args->agbno = NULLAGBLOCK;
2229 * Just fix this up, for the case where the last a.g. is shorter
2230 * (or there's only one a.g.) and the caller couldn't easily figure
2231 * that out (xfs_bmap_alloc).
2233 agsize = mp->m_sb.sb_agblocks;
2234 if (args->maxlen > agsize)
2235 args->maxlen = agsize;
2236 if (args->alignment == 0)
2237 args->alignment = 1;
2238 ASSERT(XFS_FSB_TO_AGNO(mp, args->fsbno) < mp->m_sb.sb_agcount);
2239 ASSERT(XFS_FSB_TO_AGBNO(mp, args->fsbno) < agsize);
2240 ASSERT(args->minlen <= args->maxlen);
2241 ASSERT(args->minlen <= agsize);
2242 ASSERT(args->mod < args->prod);
2243 if (XFS_FSB_TO_AGNO(mp, args->fsbno) >= mp->m_sb.sb_agcount ||
2244 XFS_FSB_TO_AGBNO(mp, args->fsbno) >= agsize ||
2245 args->minlen > args->maxlen || args->minlen > agsize ||
2246 args->mod >= args->prod) {
2247 args->fsbno = NULLFSBLOCK;
2248 trace_xfs_alloc_vextent_badargs(args);
2251 minleft = args->minleft;
2254 case XFS_ALLOCTYPE_THIS_AG:
2255 case XFS_ALLOCTYPE_NEAR_BNO:
2256 case XFS_ALLOCTYPE_THIS_BNO:
2258 * These three force us into a single a.g.
2260 args->agno = XFS_FSB_TO_AGNO(mp, args->fsbno);
2261 args->pag = xfs_perag_get(mp, args->agno);
2263 error = xfs_alloc_fix_freelist(args, 0);
2264 args->minleft = minleft;
2266 trace_xfs_alloc_vextent_nofix(args);
2270 trace_xfs_alloc_vextent_noagbp(args);
2273 args->agbno = XFS_FSB_TO_AGBNO(mp, args->fsbno);
2274 if ((error = xfs_alloc_ag_vextent(args)))
2277 case XFS_ALLOCTYPE_START_BNO:
2279 * Try near allocation first, then anywhere-in-ag after
2280 * the first a.g. fails.
2282 if ((args->userdata == XFS_ALLOC_INITIAL_USER_DATA) &&
2283 (mp->m_flags & XFS_MOUNT_32BITINODES)) {
2284 args->fsbno = XFS_AGB_TO_FSB(mp,
2285 ((mp->m_agfrotor / rotorstep) %
2286 mp->m_sb.sb_agcount), 0);
2289 args->agbno = XFS_FSB_TO_AGBNO(mp, args->fsbno);
2290 args->type = XFS_ALLOCTYPE_NEAR_BNO;
2292 case XFS_ALLOCTYPE_ANY_AG:
2293 case XFS_ALLOCTYPE_START_AG:
2294 case XFS_ALLOCTYPE_FIRST_AG:
2296 * Rotate through the allocation groups looking for a winner.
2298 if (type == XFS_ALLOCTYPE_ANY_AG) {
2300 * Start with the last place we left off.
2302 args->agno = sagno = (mp->m_agfrotor / rotorstep) %
2303 mp->m_sb.sb_agcount;
2304 args->type = XFS_ALLOCTYPE_THIS_AG;
2305 flags = XFS_ALLOC_FLAG_TRYLOCK;
2306 } else if (type == XFS_ALLOCTYPE_FIRST_AG) {
2308 * Start with allocation group given by bno.
2310 args->agno = XFS_FSB_TO_AGNO(mp, args->fsbno);
2311 args->type = XFS_ALLOCTYPE_THIS_AG;
2315 if (type == XFS_ALLOCTYPE_START_AG)
2316 args->type = XFS_ALLOCTYPE_THIS_AG;
2318 * Start with the given allocation group.
2320 args->agno = sagno = XFS_FSB_TO_AGNO(mp, args->fsbno);
2321 flags = XFS_ALLOC_FLAG_TRYLOCK;
2324 * Loop over allocation groups twice; first time with
2325 * trylock set, second time without.
2328 args->pag = xfs_perag_get(mp, args->agno);
2329 if (no_min) args->minleft = 0;
2330 error = xfs_alloc_fix_freelist(args, flags);
2331 args->minleft = minleft;
2333 trace_xfs_alloc_vextent_nofix(args);
2337 * If we get a buffer back then the allocation will fly.
2340 if ((error = xfs_alloc_ag_vextent(args)))
2345 trace_xfs_alloc_vextent_loopfailed(args);
2348 * Didn't work, figure out the next iteration.
2350 if (args->agno == sagno &&
2351 type == XFS_ALLOCTYPE_START_BNO)
2352 args->type = XFS_ALLOCTYPE_THIS_AG;
2354 * For the first allocation, we can try any AG to get
2355 * space. However, if we already have allocated a
2356 * block, we don't want to try AGs whose number is below
2357 * sagno. Otherwise, we may end up with out-of-order
2358 * locking of AGF, which might cause deadlock.
2360 if (++(args->agno) == mp->m_sb.sb_agcount) {
2361 if (args->firstblock != NULLFSBLOCK)
2367 * Reached the starting a.g., must either be done
2368 * or switch to non-trylock mode.
2370 if (args->agno == sagno) {
2372 args->agbno = NULLAGBLOCK;
2373 trace_xfs_alloc_vextent_allfailed(args);
2380 if (type == XFS_ALLOCTYPE_START_BNO) {
2381 args->agbno = XFS_FSB_TO_AGBNO(mp,
2383 args->type = XFS_ALLOCTYPE_NEAR_BNO;
2387 xfs_perag_put(args->pag);
2389 if (bump_rotor || (type == XFS_ALLOCTYPE_ANY_AG)) {
2390 if (args->agno == sagno)
2391 mp->m_agfrotor = (mp->m_agfrotor + 1) %
2392 (mp->m_sb.sb_agcount * rotorstep);
2394 mp->m_agfrotor = (args->agno * rotorstep + 1) %
2395 (mp->m_sb.sb_agcount * rotorstep);
2402 if (args->agbno == NULLAGBLOCK)
2403 args->fsbno = NULLFSBLOCK;
2405 args->fsbno = XFS_AGB_TO_FSB(mp, args->agno, args->agbno);
2407 ASSERT(args->len >= args->minlen);
2408 ASSERT(args->len <= args->maxlen);
2409 ASSERT(args->agbno % args->alignment == 0);
2410 XFS_AG_CHECK_DADDR(mp, XFS_FSB_TO_DADDR(mp, args->fsbno),
2414 xfs_perag_put(args->pag);
2417 xfs_perag_put(args->pag);
2422 xfs_alloc_vextent_worker(
2423 struct work_struct *work)
2425 struct xfs_alloc_arg *args = container_of(work,
2426 struct xfs_alloc_arg, work);
2427 unsigned long pflags;
2429 /* we are in a transaction context here */
2430 current_set_flags_nested(&pflags, PF_FSTRANS);
2432 args->result = __xfs_alloc_vextent(args);
2433 complete(args->done);
2435 current_restore_flags_nested(&pflags, PF_FSTRANS);
2441 xfs_alloc_arg_t *args) /* allocation argument structure */
2443 DECLARE_COMPLETION_ONSTACK(done);
2446 INIT_WORK(&args->work, xfs_alloc_vextent_worker);
2447 queue_work(xfs_alloc_wq, &args->work);
2448 wait_for_completion(&done);
2449 return args->result;
2454 * Just break up the extent address and hand off to xfs_free_ag_extent
2455 * after fixing up the freelist.
2459 xfs_trans_t *tp, /* transaction pointer */
2460 xfs_fsblock_t bno, /* starting block number of extent */
2461 xfs_extlen_t len) /* length of extent */
2463 xfs_alloc_arg_t args;
2467 memset(&args, 0, sizeof(xfs_alloc_arg_t));
2469 args.mp = tp->t_mountp;
2472 * validate that the block number is legal - the enables us to detect
2473 * and handle a silent filesystem corruption rather than crashing.
2475 args.agno = XFS_FSB_TO_AGNO(args.mp, bno);
2476 if (args.agno >= args.mp->m_sb.sb_agcount)
2477 return EFSCORRUPTED;
2479 args.agbno = XFS_FSB_TO_AGBNO(args.mp, bno);
2480 if (args.agbno >= args.mp->m_sb.sb_agblocks)
2481 return EFSCORRUPTED;
2483 args.pag = xfs_perag_get(args.mp, args.agno);
2486 error = xfs_alloc_fix_freelist(&args, XFS_ALLOC_FLAG_FREEING);
2490 /* validate the extent size is legal now we have the agf locked */
2491 if (args.agbno + len >
2492 be32_to_cpu(XFS_BUF_TO_AGF(args.agbp)->agf_length)) {
2493 error = EFSCORRUPTED;
2497 error = xfs_free_ag_extent(tp, args.agbp, args.agno, args.agbno, len, 0);
2499 xfs_alloc_busy_insert(tp, args.agno, args.agbno, len, 0);
2501 xfs_perag_put(args.pag);
2506 xfs_alloc_busy_insert(
2507 struct xfs_trans *tp,
2508 xfs_agnumber_t agno,
2513 struct xfs_busy_extent *new;
2514 struct xfs_busy_extent *busyp;
2515 struct xfs_perag *pag;
2516 struct rb_node **rbp;
2517 struct rb_node *parent = NULL;
2519 new = kmem_zalloc(sizeof(struct xfs_busy_extent), KM_MAYFAIL);
2522 * No Memory! Since it is now not possible to track the free
2523 * block, make this a synchronous transaction to insure that
2524 * the block is not reused before this transaction commits.
2526 trace_xfs_alloc_busy_enomem(tp->t_mountp, agno, bno, len);
2527 xfs_trans_set_sync(tp);
2534 INIT_LIST_HEAD(&new->list);
2537 /* trace before insert to be able to see failed inserts */
2538 trace_xfs_alloc_busy(tp->t_mountp, agno, bno, len);
2540 pag = xfs_perag_get(tp->t_mountp, new->agno);
2541 spin_lock(&pag->pagb_lock);
2542 rbp = &pag->pagb_tree.rb_node;
2545 busyp = rb_entry(parent, struct xfs_busy_extent, rb_node);
2547 if (new->bno < busyp->bno) {
2548 rbp = &(*rbp)->rb_left;
2549 ASSERT(new->bno + new->length <= busyp->bno);
2550 } else if (new->bno > busyp->bno) {
2551 rbp = &(*rbp)->rb_right;
2552 ASSERT(bno >= busyp->bno + busyp->length);
2558 rb_link_node(&new->rb_node, parent, rbp);
2559 rb_insert_color(&new->rb_node, &pag->pagb_tree);
2561 list_add(&new->list, &tp->t_busy);
2562 spin_unlock(&pag->pagb_lock);
2567 * Search for a busy extent within the range of the extent we are about to
2568 * allocate. You need to be holding the busy extent tree lock when calling
2569 * xfs_alloc_busy_search(). This function returns 0 for no overlapping busy
2570 * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
2571 * match. This is done so that a non-zero return indicates an overlap that
2572 * will require a synchronous transaction, but it can still be
2573 * used to distinguish between a partial or exact match.
2576 xfs_alloc_busy_search(
2577 struct xfs_mount *mp,
2578 xfs_agnumber_t agno,
2582 struct xfs_perag *pag;
2583 struct rb_node *rbp;
2584 struct xfs_busy_extent *busyp;
2587 pag = xfs_perag_get(mp, agno);
2588 spin_lock(&pag->pagb_lock);
2590 rbp = pag->pagb_tree.rb_node;
2592 /* find closest start bno overlap */
2594 busyp = rb_entry(rbp, struct xfs_busy_extent, rb_node);
2595 if (bno < busyp->bno) {
2596 /* may overlap, but exact start block is lower */
2597 if (bno + len > busyp->bno)
2600 } else if (bno > busyp->bno) {
2601 /* may overlap, but exact start block is higher */
2602 if (bno < busyp->bno + busyp->length)
2604 rbp = rbp->rb_right;
2606 /* bno matches busyp, length determines exact match */
2607 match = (busyp->length == len) ? 1 : -1;
2611 spin_unlock(&pag->pagb_lock);
2617 * The found free extent [fbno, fend] overlaps part or all of the given busy
2618 * extent. If the overlap covers the beginning, the end, or all of the busy
2619 * extent, the overlapping portion can be made unbusy and used for the
2620 * allocation. We can't split a busy extent because we can't modify a
2621 * transaction/CIL context busy list, but we can update an entries block
2624 * Returns true if the extent can safely be reused, or false if the search
2625 * needs to be restarted.
2628 xfs_alloc_busy_update_extent(
2629 struct xfs_mount *mp,
2630 struct xfs_perag *pag,
2631 struct xfs_busy_extent *busyp,
2636 xfs_agblock_t fend = fbno + flen;
2637 xfs_agblock_t bbno = busyp->bno;
2638 xfs_agblock_t bend = bbno + busyp->length;
2641 * This extent is currently being discarded. Give the thread
2642 * performing the discard a chance to mark the extent unbusy
2645 if (busyp->flags & XFS_ALLOC_BUSY_DISCARDED) {
2646 spin_unlock(&pag->pagb_lock);
2648 spin_lock(&pag->pagb_lock);
2653 * If there is a busy extent overlapping a user allocation, we have
2654 * no choice but to force the log and retry the search.
2656 * Fortunately this does not happen during normal operation, but
2657 * only if the filesystem is very low on space and has to dip into
2658 * the AGFL for normal allocations.
2663 if (bbno < fbno && bend > fend) {
2667 * +BBBBBBBBBBBBBBBBB+
2673 * We would have to split the busy extent to be able to track
2674 * it correct, which we cannot do because we would have to
2675 * modify the list of busy extents attached to the transaction
2676 * or CIL context, which is immutable.
2678 * Force out the log to clear the busy extent and retry the
2682 } else if (bbno >= fbno && bend <= fend) {
2686 * +BBBBBBBBBBBBBBBBB+
2687 * +-----------------+
2692 * +BBBBBBBBBBBBBBBBB+
2693 * +--------------------------+
2698 * +BBBBBBBBBBBBBBBBB+
2699 * +--------------------------+
2704 * +BBBBBBBBBBBBBBBBB+
2705 * +-----------------------------------+
2711 * The busy extent is fully covered by the extent we are
2712 * allocating, and can simply be removed from the rbtree.
2713 * However we cannot remove it from the immutable list
2714 * tracking busy extents in the transaction or CIL context,
2715 * so set the length to zero to mark it invalid.
2717 * We also need to restart the busy extent search from the
2718 * tree root, because erasing the node can rearrange the
2721 rb_erase(&busyp->rb_node, &pag->pagb_tree);
2724 } else if (fend < bend) {
2728 * +BBBBBBBBBBBBBBBBB+
2734 * +BBBBBBBBBBBBBBBBB+
2735 * +------------------+
2740 } else if (bbno < fbno) {
2744 * +BBBBBBBBBBBBBBBBB+
2750 * +BBBBBBBBBBBBBBBBB+
2751 * +----------------------+
2754 busyp->length = fbno - busyp->bno;
2759 trace_xfs_alloc_busy_reuse(mp, pag->pag_agno, fbno, flen);
2763 spin_unlock(&pag->pagb_lock);
2764 xfs_log_force(mp, XFS_LOG_SYNC);
2765 trace_xfs_alloc_busy_force(mp, pag->pag_agno, fbno, flen);
2766 spin_lock(&pag->pagb_lock);
2772 * For a given extent [fbno, flen], make sure we can reuse it safely.
2775 xfs_alloc_busy_reuse(
2776 struct xfs_mount *mp,
2777 xfs_agnumber_t agno,
2782 struct xfs_perag *pag;
2783 struct rb_node *rbp;
2787 pag = xfs_perag_get(mp, agno);
2788 spin_lock(&pag->pagb_lock);
2790 rbp = pag->pagb_tree.rb_node;
2792 struct xfs_busy_extent *busyp =
2793 rb_entry(rbp, struct xfs_busy_extent, rb_node);
2794 xfs_agblock_t bbno = busyp->bno;
2795 xfs_agblock_t bend = bbno + busyp->length;
2797 if (fbno + flen <= bbno) {
2800 } else if (fbno >= bend) {
2801 rbp = rbp->rb_right;
2805 if (!xfs_alloc_busy_update_extent(mp, pag, busyp, fbno, flen,
2809 spin_unlock(&pag->pagb_lock);
2814 * For a given extent [fbno, flen], search the busy extent list to find a
2815 * subset of the extent that is not busy. If *rlen is smaller than
2816 * args->minlen no suitable extent could be found, and the higher level
2817 * code needs to force out the log and retry the allocation.
2820 xfs_alloc_busy_trim(
2821 struct xfs_alloc_arg *args,
2824 xfs_agblock_t *rbno,
2829 struct rb_node *rbp;
2833 spin_lock(&args->pag->pagb_lock);
2837 rbp = args->pag->pagb_tree.rb_node;
2838 while (rbp && flen >= args->minlen) {
2839 struct xfs_busy_extent *busyp =
2840 rb_entry(rbp, struct xfs_busy_extent, rb_node);
2841 xfs_agblock_t fend = fbno + flen;
2842 xfs_agblock_t bbno = busyp->bno;
2843 xfs_agblock_t bend = bbno + busyp->length;
2848 } else if (fbno >= bend) {
2849 rbp = rbp->rb_right;
2854 * If this is a metadata allocation, try to reuse the busy
2855 * extent instead of trimming the allocation.
2857 if (!args->userdata &&
2858 !(busyp->flags & XFS_ALLOC_BUSY_DISCARDED)) {
2859 if (!xfs_alloc_busy_update_extent(args->mp, args->pag,
2872 * +BBBBBBBBBBBBBBBBB+
2878 * +BBBBBBBBBBBBBBBBB+
2884 * +BBBBBBBBBBBBBBBBB+
2890 * +BBBBBBBBBBBBBBBBB+
2891 * +-----------------+
2894 * No unbusy region in extent, return failure.
2902 * +BBBBBBBBBBBBBBBBB+
2903 * +----------------------+
2908 * +BBBBBBBBBBBBBBBBB+
2909 * +--------------------------+
2912 * Needs to be trimmed to:
2917 } else if (bend >= fend) {
2923 * +BBBBBBBBBBBBBBBBB+
2924 * +------------------+
2929 * +BBBBBBBBBBBBBBBBB+
2930 * +--------------------------+
2933 * Needs to be trimmed to:
2939 /* middle overlap */
2944 * +BBBBBBBBBBBBBBBBB+
2945 * +-----------------------------------+
2948 * Can be trimmed to:
2949 * +-------+ OR +-------+
2950 * fbno fend fbno fend
2952 * Backward allocation leads to significant
2953 * fragmentation of directories, which degrades
2954 * directory performance, therefore we always want to
2955 * choose the option that produces forward allocation
2957 * Preferring the lower bno extent will make the next
2958 * request use "fend" as the start of the next
2959 * allocation; if the segment is no longer busy at
2960 * that point, we'll get a contiguous allocation, but
2961 * even if it is still busy, we will get a forward
2963 * We try to avoid choosing the segment at "bend",
2964 * because that can lead to the next allocation
2965 * taking the segment at "fbno", which would be a
2966 * backward allocation. We only use the segment at
2967 * "fbno" if it is much larger than the current
2968 * requested size, because in that case there's a
2969 * good chance subsequent allocations will be
2972 if (bbno - fbno >= args->maxlen) {
2973 /* left candidate fits perfect */
2975 } else if (fend - bend >= args->maxlen * 4) {
2976 /* right candidate has enough free space */
2978 } else if (bbno - fbno >= args->minlen) {
2979 /* left candidate fits minimum requirement */
2988 spin_unlock(&args->pag->pagb_lock);
2990 if (fbno != bno || flen != len) {
2991 trace_xfs_alloc_busy_trim(args->mp, args->agno, bno, len,
2999 * Return a zero extent length as failure indications. All callers
3000 * re-check if the trimmed extent satisfies the minlen requirement.
3002 spin_unlock(&args->pag->pagb_lock);
3003 trace_xfs_alloc_busy_trim(args->mp, args->agno, bno, len, fbno, 0);
3009 xfs_alloc_busy_clear_one(
3010 struct xfs_mount *mp,
3011 struct xfs_perag *pag,
3012 struct xfs_busy_extent *busyp)
3014 if (busyp->length) {
3015 trace_xfs_alloc_busy_clear(mp, busyp->agno, busyp->bno,
3017 rb_erase(&busyp->rb_node, &pag->pagb_tree);
3020 list_del_init(&busyp->list);
3025 * Remove all extents on the passed in list from the busy extents tree.
3026 * If do_discard is set skip extents that need to be discarded, and mark
3027 * these as undergoing a discard operation instead.
3030 xfs_alloc_busy_clear(
3031 struct xfs_mount *mp,
3032 struct list_head *list,
3035 struct xfs_busy_extent *busyp, *n;
3036 struct xfs_perag *pag = NULL;
3037 xfs_agnumber_t agno = NULLAGNUMBER;
3039 list_for_each_entry_safe(busyp, n, list, list) {
3040 if (busyp->agno != agno) {
3042 spin_unlock(&pag->pagb_lock);
3045 pag = xfs_perag_get(mp, busyp->agno);
3046 spin_lock(&pag->pagb_lock);
3050 if (do_discard && busyp->length &&
3051 !(busyp->flags & XFS_ALLOC_BUSY_SKIP_DISCARD))
3052 busyp->flags = XFS_ALLOC_BUSY_DISCARDED;
3054 xfs_alloc_busy_clear_one(mp, pag, busyp);
3058 spin_unlock(&pag->pagb_lock);
3064 * Callback for list_sort to sort busy extents by the AG they reside in.
3067 xfs_busy_extent_ag_cmp(
3069 struct list_head *a,
3070 struct list_head *b)
3072 return container_of(a, struct xfs_busy_extent, list)->agno -
3073 container_of(b, struct xfs_busy_extent, list)->agno;