2005-04-17 00:20:36 +02:00
|
|
|
/*
|
2005-11-02 04:58:39 +01:00
|
|
|
* Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
|
|
|
|
* All Rights Reserved.
|
2005-04-17 00:20:36 +02:00
|
|
|
*
|
2005-11-02 04:58:39 +01:00
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public License as
|
2005-04-17 00:20:36 +02:00
|
|
|
* published by the Free Software Foundation.
|
|
|
|
*
|
2005-11-02 04:58:39 +01:00
|
|
|
* This program is distributed in the hope that it would be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
* GNU General Public License for more details.
|
2005-04-17 00:20:36 +02:00
|
|
|
*
|
2005-11-02 04:58:39 +01:00
|
|
|
* You should have received a copy of the GNU General Public License
|
|
|
|
* along with this program; if not, write the Free Software Foundation,
|
|
|
|
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
|
2005-04-17 00:20:36 +02:00
|
|
|
*/
|
|
|
|
#include "xfs.h"
|
2005-11-02 04:38:42 +01:00
|
|
|
#include "xfs_fs.h"
|
2005-04-17 00:20:36 +02:00
|
|
|
#include "xfs_types.h"
|
2005-11-02 04:38:42 +01:00
|
|
|
#include "xfs_bit.h"
|
2005-04-17 00:20:36 +02:00
|
|
|
#include "xfs_log.h"
|
2005-11-02 04:38:42 +01:00
|
|
|
#include "xfs_inum.h"
|
2005-04-17 00:20:36 +02:00
|
|
|
#include "xfs_trans.h"
|
|
|
|
#include "xfs_sb.h"
|
|
|
|
#include "xfs_ag.h"
|
|
|
|
#include "xfs_mount.h"
|
|
|
|
#include "xfs_bmap_btree.h"
|
2005-11-02 04:38:42 +01:00
|
|
|
#include "xfs_alloc_btree.h"
|
2005-04-17 00:20:36 +02:00
|
|
|
#include "xfs_ialloc_btree.h"
|
2005-11-02 04:38:42 +01:00
|
|
|
#include "xfs_dinode.h"
|
|
|
|
#include "xfs_inode.h"
|
2005-04-17 00:20:36 +02:00
|
|
|
#include "xfs_btree.h"
|
|
|
|
#include "xfs_alloc.h"
|
|
|
|
#include "xfs_error.h"
|
2009-12-15 00:14:59 +01:00
|
|
|
#include "xfs_trace.h"
|
2005-04-17 00:20:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
#define XFS_ABSDIFF(a,b) (((a) <= (b)) ? ((b) - (a)) : ((a) - (b)))
|
|
|
|
|
|
|
|
#define XFSA_FIXUP_BNO_OK 1
|
|
|
|
#define XFSA_FIXUP_CNT_OK 2
|
|
|
|
|
|
|
|
STATIC int xfs_alloc_ag_vextent_exact(xfs_alloc_arg_t *);
|
|
|
|
STATIC int xfs_alloc_ag_vextent_near(xfs_alloc_arg_t *);
|
|
|
|
STATIC int xfs_alloc_ag_vextent_size(xfs_alloc_arg_t *);
|
|
|
|
STATIC int xfs_alloc_ag_vextent_small(xfs_alloc_arg_t *,
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_btree_cur_t *, xfs_agblock_t *, xfs_extlen_t *, int *);
|
|
|
|
STATIC void xfs_alloc_busy_trim(struct xfs_alloc_arg *,
|
|
|
|
xfs_agblock_t, xfs_extlen_t, xfs_agblock_t *, xfs_extlen_t *);
|
2005-04-17 00:20:36 +02:00
|
|
|
|
2008-10-30 06:56:09 +01:00
|
|
|
/*
|
|
|
|
* Lookup the record equal to [bno, len] in the btree given by cur.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_lookup_eq(
|
|
|
|
struct xfs_btree_cur *cur, /* btree cursor */
|
|
|
|
xfs_agblock_t bno, /* starting block of extent */
|
|
|
|
xfs_extlen_t len, /* length of extent */
|
|
|
|
int *stat) /* success/failure */
|
|
|
|
{
|
|
|
|
cur->bc_rec.a.ar_startblock = bno;
|
|
|
|
cur->bc_rec.a.ar_blockcount = len;
|
|
|
|
return xfs_btree_lookup(cur, XFS_LOOKUP_EQ, stat);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Lookup the first record greater than or equal to [bno, len]
|
|
|
|
* in the btree given by cur.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_lookup_ge(
|
|
|
|
struct xfs_btree_cur *cur, /* btree cursor */
|
|
|
|
xfs_agblock_t bno, /* starting block of extent */
|
|
|
|
xfs_extlen_t len, /* length of extent */
|
|
|
|
int *stat) /* success/failure */
|
|
|
|
{
|
|
|
|
cur->bc_rec.a.ar_startblock = bno;
|
|
|
|
cur->bc_rec.a.ar_blockcount = len;
|
|
|
|
return xfs_btree_lookup(cur, XFS_LOOKUP_GE, stat);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Lookup the first record less than or equal to [bno, len]
|
|
|
|
* in the btree given by cur.
|
|
|
|
*/
|
2011-01-07 14:02:04 +01:00
|
|
|
int /* error */
|
2008-10-30 06:56:09 +01:00
|
|
|
xfs_alloc_lookup_le(
|
|
|
|
struct xfs_btree_cur *cur, /* btree cursor */
|
|
|
|
xfs_agblock_t bno, /* starting block of extent */
|
|
|
|
xfs_extlen_t len, /* length of extent */
|
|
|
|
int *stat) /* success/failure */
|
|
|
|
{
|
|
|
|
cur->bc_rec.a.ar_startblock = bno;
|
|
|
|
cur->bc_rec.a.ar_blockcount = len;
|
|
|
|
return xfs_btree_lookup(cur, XFS_LOOKUP_LE, stat);
|
|
|
|
}
|
|
|
|
|
2008-10-30 06:56:32 +01:00
|
|
|
/*
|
|
|
|
* Update the record referred to by cur to the value given
|
|
|
|
* by [bno, len].
|
|
|
|
* This either works (return 0) or gets an EFSCORRUPTED error.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_update(
|
|
|
|
struct xfs_btree_cur *cur, /* btree cursor */
|
|
|
|
xfs_agblock_t bno, /* starting block of extent */
|
|
|
|
xfs_extlen_t len) /* length of extent */
|
|
|
|
{
|
|
|
|
union xfs_btree_rec rec;
|
|
|
|
|
|
|
|
rec.alloc.ar_startblock = cpu_to_be32(bno);
|
|
|
|
rec.alloc.ar_blockcount = cpu_to_be32(len);
|
|
|
|
return xfs_btree_update(cur, &rec);
|
|
|
|
}
|
2008-10-30 06:56:09 +01:00
|
|
|
|
2008-10-30 06:58:11 +01:00
|
|
|
/*
|
|
|
|
* Get the data from the pointed-to record.
|
|
|
|
*/
|
2011-01-07 14:02:04 +01:00
|
|
|
int /* error */
|
2008-10-30 06:58:11 +01:00
|
|
|
xfs_alloc_get_rec(
|
|
|
|
struct xfs_btree_cur *cur, /* btree cursor */
|
|
|
|
xfs_agblock_t *bno, /* output: starting block of extent */
|
|
|
|
xfs_extlen_t *len, /* output: length of extent */
|
|
|
|
int *stat) /* output: success/failure */
|
|
|
|
{
|
|
|
|
union xfs_btree_rec *rec;
|
|
|
|
int error;
|
|
|
|
|
|
|
|
error = xfs_btree_get_rec(cur, &rec, stat);
|
|
|
|
if (!error && *stat == 1) {
|
|
|
|
*bno = be32_to_cpu(rec->alloc.ar_startblock);
|
|
|
|
*len = be32_to_cpu(rec->alloc.ar_blockcount);
|
|
|
|
}
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Compute aligned version of the found extent.
|
|
|
|
* Takes alignment and min length into account.
|
|
|
|
*/
|
2008-04-10 04:21:32 +02:00
|
|
|
STATIC void
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_alloc_compute_aligned(
|
2011-03-04 13:59:54 +01:00
|
|
|
xfs_alloc_arg_t *args, /* allocation argument structure */
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_agblock_t foundbno, /* starting block in found extent */
|
|
|
|
xfs_extlen_t foundlen, /* length in found extent */
|
|
|
|
xfs_agblock_t *resbno, /* result block number */
|
|
|
|
xfs_extlen_t *reslen) /* result length */
|
|
|
|
{
|
|
|
|
xfs_agblock_t bno;
|
|
|
|
xfs_extlen_t len;
|
|
|
|
|
2011-04-24 21:06:15 +02:00
|
|
|
/* Trim busy sections out of found extent */
|
|
|
|
xfs_alloc_busy_trim(args, foundbno, foundlen, &bno, &len);
|
|
|
|
|
|
|
|
if (args->alignment > 1 && len >= args->minlen) {
|
|
|
|
xfs_agblock_t aligned_bno = roundup(bno, args->alignment);
|
|
|
|
xfs_extlen_t diff = aligned_bno - bno;
|
|
|
|
|
|
|
|
*resbno = aligned_bno;
|
|
|
|
*reslen = diff >= len ? 0 : len - diff;
|
2005-04-17 00:20:36 +02:00
|
|
|
} else {
|
2011-04-24 21:06:15 +02:00
|
|
|
*resbno = bno;
|
|
|
|
*reslen = len;
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compute best start block and diff for "near" allocations.
|
|
|
|
* freelen >= wantlen already checked by caller.
|
|
|
|
*/
|
|
|
|
STATIC xfs_extlen_t /* difference value (absolute) */
|
|
|
|
xfs_alloc_compute_diff(
|
|
|
|
xfs_agblock_t wantbno, /* target starting block */
|
|
|
|
xfs_extlen_t wantlen, /* target length */
|
|
|
|
xfs_extlen_t alignment, /* target alignment */
|
|
|
|
xfs_agblock_t freebno, /* freespace's starting block */
|
|
|
|
xfs_extlen_t freelen, /* freespace's length */
|
|
|
|
xfs_agblock_t *newbnop) /* result: best start block from free */
|
|
|
|
{
|
|
|
|
xfs_agblock_t freeend; /* end of freespace extent */
|
|
|
|
xfs_agblock_t newbno1; /* return block number */
|
|
|
|
xfs_agblock_t newbno2; /* other new block number */
|
|
|
|
xfs_extlen_t newlen1=0; /* length with newbno1 */
|
|
|
|
xfs_extlen_t newlen2=0; /* length with newbno2 */
|
|
|
|
xfs_agblock_t wantend; /* end of target extent */
|
|
|
|
|
|
|
|
ASSERT(freelen >= wantlen);
|
|
|
|
freeend = freebno + freelen;
|
|
|
|
wantend = wantbno + wantlen;
|
|
|
|
if (freebno >= wantbno) {
|
|
|
|
if ((newbno1 = roundup(freebno, alignment)) >= freeend)
|
|
|
|
newbno1 = NULLAGBLOCK;
|
|
|
|
} else if (freeend >= wantend && alignment > 1) {
|
|
|
|
newbno1 = roundup(wantbno, alignment);
|
|
|
|
newbno2 = newbno1 - alignment;
|
|
|
|
if (newbno1 >= freeend)
|
|
|
|
newbno1 = NULLAGBLOCK;
|
|
|
|
else
|
|
|
|
newlen1 = XFS_EXTLEN_MIN(wantlen, freeend - newbno1);
|
|
|
|
if (newbno2 < freebno)
|
|
|
|
newbno2 = NULLAGBLOCK;
|
|
|
|
else
|
|
|
|
newlen2 = XFS_EXTLEN_MIN(wantlen, freeend - newbno2);
|
|
|
|
if (newbno1 != NULLAGBLOCK && newbno2 != NULLAGBLOCK) {
|
|
|
|
if (newlen1 < newlen2 ||
|
|
|
|
(newlen1 == newlen2 &&
|
|
|
|
XFS_ABSDIFF(newbno1, wantbno) >
|
|
|
|
XFS_ABSDIFF(newbno2, wantbno)))
|
|
|
|
newbno1 = newbno2;
|
|
|
|
} else if (newbno2 != NULLAGBLOCK)
|
|
|
|
newbno1 = newbno2;
|
|
|
|
} else if (freeend >= wantend) {
|
|
|
|
newbno1 = wantbno;
|
|
|
|
} else if (alignment > 1) {
|
|
|
|
newbno1 = roundup(freeend - wantlen, alignment);
|
|
|
|
if (newbno1 > freeend - wantlen &&
|
|
|
|
newbno1 - alignment >= freebno)
|
|
|
|
newbno1 -= alignment;
|
|
|
|
else if (newbno1 >= freeend)
|
|
|
|
newbno1 = NULLAGBLOCK;
|
|
|
|
} else
|
|
|
|
newbno1 = freeend - wantlen;
|
|
|
|
*newbnop = newbno1;
|
|
|
|
return newbno1 == NULLAGBLOCK ? 0 : XFS_ABSDIFF(newbno1, wantbno);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Fix up the length, based on mod and prod.
|
|
|
|
* len should be k * prod + mod for some k.
|
|
|
|
* If len is too small it is returned unchanged.
|
|
|
|
* If len hits maxlen it is left alone.
|
|
|
|
*/
|
|
|
|
STATIC void
|
|
|
|
xfs_alloc_fix_len(
|
|
|
|
xfs_alloc_arg_t *args) /* allocation argument structure */
|
|
|
|
{
|
|
|
|
xfs_extlen_t k;
|
|
|
|
xfs_extlen_t rlen;
|
|
|
|
|
|
|
|
ASSERT(args->mod < args->prod);
|
|
|
|
rlen = args->len;
|
|
|
|
ASSERT(rlen >= args->minlen);
|
|
|
|
ASSERT(rlen <= args->maxlen);
|
|
|
|
if (args->prod <= 1 || rlen < args->mod || rlen == args->maxlen ||
|
|
|
|
(args->mod == 0 && rlen < args->prod))
|
|
|
|
return;
|
|
|
|
k = rlen % args->prod;
|
|
|
|
if (k == args->mod)
|
|
|
|
return;
|
|
|
|
if (k > args->mod) {
|
|
|
|
if ((int)(rlen = rlen - k - args->mod) < (int)args->minlen)
|
|
|
|
return;
|
|
|
|
} else {
|
|
|
|
if ((int)(rlen = rlen - args->prod - (args->mod - k)) <
|
|
|
|
(int)args->minlen)
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
ASSERT(rlen >= args->minlen);
|
|
|
|
ASSERT(rlen <= args->maxlen);
|
|
|
|
args->len = rlen;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Fix up length if there is too little space left in the a.g.
|
|
|
|
* Return 1 if ok, 0 if too little, should give up.
|
|
|
|
*/
|
|
|
|
STATIC int
|
|
|
|
xfs_alloc_fix_minleft(
|
|
|
|
xfs_alloc_arg_t *args) /* allocation argument structure */
|
|
|
|
{
|
|
|
|
xfs_agf_t *agf; /* a.g. freelist header */
|
|
|
|
int diff; /* free space difference */
|
|
|
|
|
|
|
|
if (args->minleft == 0)
|
|
|
|
return 1;
|
|
|
|
agf = XFS_BUF_TO_AGF(args->agbp);
|
2005-11-02 05:11:25 +01:00
|
|
|
diff = be32_to_cpu(agf->agf_freeblks)
|
2005-04-17 00:20:36 +02:00
|
|
|
- args->len - args->minleft;
|
|
|
|
if (diff >= 0)
|
|
|
|
return 1;
|
|
|
|
args->len += diff; /* shrink the allocated space */
|
|
|
|
if (args->len >= args->minlen)
|
|
|
|
return 1;
|
|
|
|
args->agbno = NULLAGBLOCK;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Update the two btrees, logically removing from freespace the extent
|
|
|
|
* starting at rbno, rlen blocks. The extent is contained within the
|
|
|
|
* actual (current) free extent fbno for flen blocks.
|
|
|
|
* Flags are passed in indicating whether the cursors are set to the
|
|
|
|
* relevant records.
|
|
|
|
*/
|
|
|
|
STATIC int /* error code */
|
|
|
|
xfs_alloc_fixup_trees(
|
|
|
|
xfs_btree_cur_t *cnt_cur, /* cursor for by-size btree */
|
|
|
|
xfs_btree_cur_t *bno_cur, /* cursor for by-block btree */
|
|
|
|
xfs_agblock_t fbno, /* starting block of free extent */
|
|
|
|
xfs_extlen_t flen, /* length of free extent */
|
|
|
|
xfs_agblock_t rbno, /* starting block of returned extent */
|
|
|
|
xfs_extlen_t rlen, /* length of returned extent */
|
|
|
|
int flags) /* flags, XFSA_FIXUP_... */
|
|
|
|
{
|
|
|
|
int error; /* error code */
|
|
|
|
int i; /* operation results */
|
|
|
|
xfs_agblock_t nfbno1; /* first new free startblock */
|
|
|
|
xfs_agblock_t nfbno2; /* second new free startblock */
|
|
|
|
xfs_extlen_t nflen1=0; /* first new free length */
|
|
|
|
xfs_extlen_t nflen2=0; /* second new free length */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Look up the record in the by-size tree if necessary.
|
|
|
|
*/
|
|
|
|
if (flags & XFSA_FIXUP_CNT_OK) {
|
|
|
|
#ifdef DEBUG
|
|
|
|
if ((error = xfs_alloc_get_rec(cnt_cur, &nfbno1, &nflen1, &i)))
|
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(
|
|
|
|
i == 1 && nfbno1 == fbno && nflen1 == flen);
|
|
|
|
#endif
|
|
|
|
} else {
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, fbno, flen, &i)))
|
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 1);
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Look up the record in the by-block tree if necessary.
|
|
|
|
*/
|
|
|
|
if (flags & XFSA_FIXUP_BNO_OK) {
|
|
|
|
#ifdef DEBUG
|
|
|
|
if ((error = xfs_alloc_get_rec(bno_cur, &nfbno1, &nflen1, &i)))
|
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(
|
|
|
|
i == 1 && nfbno1 == fbno && nflen1 == flen);
|
|
|
|
#endif
|
|
|
|
} else {
|
|
|
|
if ((error = xfs_alloc_lookup_eq(bno_cur, fbno, flen, &i)))
|
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 1);
|
|
|
|
}
|
2008-10-30 07:14:34 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
#ifdef DEBUG
|
2008-10-30 07:14:34 +01:00
|
|
|
if (bno_cur->bc_nlevels == 1 && cnt_cur->bc_nlevels == 1) {
|
|
|
|
struct xfs_btree_block *bnoblock;
|
|
|
|
struct xfs_btree_block *cntblock;
|
|
|
|
|
|
|
|
bnoblock = XFS_BUF_TO_BLOCK(bno_cur->bc_bufs[0]);
|
|
|
|
cntblock = XFS_BUF_TO_BLOCK(cnt_cur->bc_bufs[0]);
|
2005-04-17 00:20:36 +02:00
|
|
|
|
2008-10-30 07:14:34 +01:00
|
|
|
XFS_WANT_CORRUPTED_RETURN(
|
|
|
|
bnoblock->bb_numrecs == cntblock->bb_numrecs);
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
#endif
|
2008-10-30 07:14:34 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Deal with all four cases: the allocated record is contained
|
|
|
|
* within the freespace record, so we can have new freespace
|
|
|
|
* at either (or both) end, or no freespace remaining.
|
|
|
|
*/
|
|
|
|
if (rbno == fbno && rlen == flen)
|
|
|
|
nfbno1 = nfbno2 = NULLAGBLOCK;
|
|
|
|
else if (rbno == fbno) {
|
|
|
|
nfbno1 = rbno + rlen;
|
|
|
|
nflen1 = flen - rlen;
|
|
|
|
nfbno2 = NULLAGBLOCK;
|
|
|
|
} else if (rbno + rlen == fbno + flen) {
|
|
|
|
nfbno1 = fbno;
|
|
|
|
nflen1 = flen - rlen;
|
|
|
|
nfbno2 = NULLAGBLOCK;
|
|
|
|
} else {
|
|
|
|
nfbno1 = fbno;
|
|
|
|
nflen1 = rbno - fbno;
|
|
|
|
nfbno2 = rbno + rlen;
|
|
|
|
nflen2 = (fbno + flen) - nfbno2;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Delete the entry from the by-size btree.
|
|
|
|
*/
|
2008-10-30 06:58:01 +01:00
|
|
|
if ((error = xfs_btree_delete(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 1);
|
|
|
|
/*
|
|
|
|
* Add new by-size btree entry(s).
|
|
|
|
*/
|
|
|
|
if (nfbno1 != NULLAGBLOCK) {
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, nfbno1, nflen1, &i)))
|
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 0);
|
2008-10-30 06:57:40 +01:00
|
|
|
if ((error = xfs_btree_insert(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 1);
|
|
|
|
}
|
|
|
|
if (nfbno2 != NULLAGBLOCK) {
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, nfbno2, nflen2, &i)))
|
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 0);
|
2008-10-30 06:57:40 +01:00
|
|
|
if ((error = xfs_btree_insert(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 1);
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Fix up the by-block btree entry(s).
|
|
|
|
*/
|
|
|
|
if (nfbno1 == NULLAGBLOCK) {
|
|
|
|
/*
|
|
|
|
* No remaining freespace, just delete the by-block tree entry.
|
|
|
|
*/
|
2008-10-30 06:58:01 +01:00
|
|
|
if ((error = xfs_btree_delete(bno_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 1);
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* Update the by-block entry to start later|be shorter.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_update(bno_cur, nfbno1, nflen1)))
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
if (nfbno2 != NULLAGBLOCK) {
|
|
|
|
/*
|
|
|
|
* 2 resulting free entries, need to add one.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_eq(bno_cur, nfbno2, nflen2, &i)))
|
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 0);
|
2008-10-30 06:57:40 +01:00
|
|
|
if ((error = xfs_btree_insert(bno_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
XFS_WANT_CORRUPTED_RETURN(i == 1);
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read in the allocation group free block array.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_read_agfl(
|
|
|
|
xfs_mount_t *mp, /* mount point structure */
|
|
|
|
xfs_trans_t *tp, /* transaction pointer */
|
|
|
|
xfs_agnumber_t agno, /* allocation group number */
|
|
|
|
xfs_buf_t **bpp) /* buffer for the ag free block array */
|
|
|
|
{
|
|
|
|
xfs_buf_t *bp; /* return value */
|
|
|
|
int error;
|
|
|
|
|
|
|
|
ASSERT(agno != NULLAGNUMBER);
|
|
|
|
error = xfs_trans_read_buf(
|
|
|
|
mp, tp, mp->m_ddev_targp,
|
|
|
|
XFS_AG_DADDR(mp, agno, XFS_AGFL_DADDR(mp)),
|
|
|
|
XFS_FSS_TO_BB(mp, 1), 0, &bp);
|
|
|
|
if (error)
|
|
|
|
return error;
|
|
|
|
ASSERT(bp);
|
|
|
|
ASSERT(!XFS_BUF_GETERROR(bp));
|
|
|
|
XFS_BUF_SET_VTYPE_REF(bp, B_FS_AGFL, XFS_AGFL_REF);
|
|
|
|
*bpp = bp;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2011-03-04 13:59:55 +01:00
|
|
|
STATIC int
|
|
|
|
xfs_alloc_update_counters(
|
|
|
|
struct xfs_trans *tp,
|
|
|
|
struct xfs_perag *pag,
|
|
|
|
struct xfs_buf *agbp,
|
|
|
|
long len)
|
|
|
|
{
|
|
|
|
struct xfs_agf *agf = XFS_BUF_TO_AGF(agbp);
|
|
|
|
|
|
|
|
pag->pagf_freeblks += len;
|
|
|
|
be32_add_cpu(&agf->agf_freeblks, len);
|
|
|
|
|
|
|
|
xfs_trans_agblocks_delta(tp, len);
|
|
|
|
if (unlikely(be32_to_cpu(agf->agf_freeblks) >
|
|
|
|
be32_to_cpu(agf->agf_length)))
|
|
|
|
return EFSCORRUPTED;
|
|
|
|
|
|
|
|
xfs_alloc_log_agf(tp, agbp, XFS_AGF_FREEBLKS);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Allocation group level functions.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allocate a variable extent in the allocation group agno.
|
|
|
|
* Type and bno are used to determine where in the allocation group the
|
|
|
|
* extent will start.
|
|
|
|
* Extent's length (returned in *len) will be between minlen and maxlen,
|
|
|
|
* and of the form k * prod + mod unless there's nothing that large.
|
|
|
|
* Return the starting a.g. block, or NULLAGBLOCK if we can't do it.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_ag_vextent(
|
|
|
|
xfs_alloc_arg_t *args) /* argument structure for allocation */
|
|
|
|
{
|
|
|
|
int error=0;
|
|
|
|
|
|
|
|
ASSERT(args->minlen > 0);
|
|
|
|
ASSERT(args->maxlen > 0);
|
|
|
|
ASSERT(args->minlen <= args->maxlen);
|
|
|
|
ASSERT(args->mod < args->prod);
|
|
|
|
ASSERT(args->alignment > 0);
|
|
|
|
/*
|
|
|
|
* Branch to correct routine based on the type.
|
|
|
|
*/
|
|
|
|
args->wasfromfl = 0;
|
|
|
|
switch (args->type) {
|
|
|
|
case XFS_ALLOCTYPE_THIS_AG:
|
|
|
|
error = xfs_alloc_ag_vextent_size(args);
|
|
|
|
break;
|
|
|
|
case XFS_ALLOCTYPE_NEAR_BNO:
|
|
|
|
error = xfs_alloc_ag_vextent_near(args);
|
|
|
|
break;
|
|
|
|
case XFS_ALLOCTYPE_THIS_BNO:
|
|
|
|
error = xfs_alloc_ag_vextent_exact(args);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
ASSERT(0);
|
|
|
|
/* NOTREACHED */
|
|
|
|
}
|
2011-03-04 13:59:55 +01:00
|
|
|
|
|
|
|
if (error || args->agbno == NULLAGBLOCK)
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
2011-03-04 13:59:55 +01:00
|
|
|
|
|
|
|
ASSERT(args->len >= args->minlen);
|
|
|
|
ASSERT(args->len <= args->maxlen);
|
|
|
|
ASSERT(!args->wasfromfl || !args->isfl);
|
|
|
|
ASSERT(args->agbno % args->alignment == 0);
|
|
|
|
|
|
|
|
if (!args->wasfromfl) {
|
|
|
|
error = xfs_alloc_update_counters(args->tp, args->pag,
|
|
|
|
args->agbp,
|
|
|
|
-((long)(args->len)));
|
|
|
|
if (error)
|
|
|
|
return error;
|
|
|
|
|
2011-04-24 21:06:15 +02:00
|
|
|
ASSERT(!xfs_alloc_busy_search(args->mp, args->agno,
|
|
|
|
args->agbno, args->len));
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2011-03-04 13:59:55 +01:00
|
|
|
|
|
|
|
if (!args->isfl) {
|
|
|
|
xfs_trans_mod_sb(args->tp, args->wasdel ?
|
|
|
|
XFS_TRANS_SB_RES_FDBLOCKS :
|
|
|
|
XFS_TRANS_SB_FDBLOCKS,
|
|
|
|
-((long)(args->len)));
|
|
|
|
}
|
|
|
|
|
|
|
|
XFS_STATS_INC(xs_allocx);
|
|
|
|
XFS_STATS_ADD(xs_allocb, args->len);
|
|
|
|
return error;
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allocate a variable extent at exactly agno/bno.
|
|
|
|
* Extent's length (returned in *len) will be between minlen and maxlen,
|
|
|
|
* and of the form k * prod + mod unless there's nothing that large.
|
|
|
|
* Return the starting a.g. block (bno), or NULLAGBLOCK if we can't do it.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_ag_vextent_exact(
|
|
|
|
xfs_alloc_arg_t *args) /* allocation argument structure */
|
|
|
|
{
|
|
|
|
xfs_btree_cur_t *bno_cur;/* by block-number btree cursor */
|
|
|
|
xfs_btree_cur_t *cnt_cur;/* by count btree cursor */
|
|
|
|
int error;
|
|
|
|
xfs_agblock_t fbno; /* start block of found extent */
|
|
|
|
xfs_extlen_t flen; /* length of found extent */
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_agblock_t tbno; /* start block of trimmed extent */
|
|
|
|
xfs_extlen_t tlen; /* length of trimmed extent */
|
|
|
|
xfs_agblock_t tend; /* end block of trimmed extent */
|
|
|
|
xfs_agblock_t end; /* end of allocated extent */
|
2005-04-17 00:20:36 +02:00
|
|
|
int i; /* success/failure of operation */
|
|
|
|
xfs_extlen_t rlen; /* length of returned extent */
|
|
|
|
|
|
|
|
ASSERT(args->alignment == 1);
|
2010-12-10 16:03:57 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Allocate/initialize a cursor for the by-number freespace btree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
bno_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
|
2010-12-10 16:03:57 +01:00
|
|
|
args->agno, XFS_BTNUM_BNO);
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Lookup bno and minlen in the btree (minlen is irrelevant, really).
|
|
|
|
* Look for the closest free block <= bno, it must contain bno
|
|
|
|
* if any free block does.
|
|
|
|
*/
|
2010-12-10 16:03:57 +01:00
|
|
|
error = xfs_alloc_lookup_le(bno_cur, args->agbno, args->minlen, &i);
|
|
|
|
if (error)
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
2010-12-10 16:03:57 +01:00
|
|
|
if (!i)
|
|
|
|
goto not_found;
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Grab the freespace record.
|
|
|
|
*/
|
2010-12-10 16:03:57 +01:00
|
|
|
error = xfs_alloc_get_rec(bno_cur, &fbno, &flen, &i);
|
|
|
|
if (error)
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
ASSERT(fbno <= args->agbno);
|
2010-12-10 16:03:57 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
2011-04-24 21:06:15 +02:00
|
|
|
* Check for overlapping busy extents.
|
2005-04-17 00:20:36 +02:00
|
|
|
*/
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_alloc_busy_trim(args, fbno, flen, &tbno, &tlen);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Give up if the start of the extent is busy, or the freespace isn't
|
|
|
|
* long enough for the minimum request.
|
|
|
|
*/
|
|
|
|
if (tbno > args->agbno)
|
|
|
|
goto not_found;
|
|
|
|
if (tlen < args->minlen)
|
|
|
|
goto not_found;
|
|
|
|
tend = tbno + tlen;
|
|
|
|
if (tend < args->agbno + args->minlen)
|
2010-12-10 16:03:57 +01:00
|
|
|
goto not_found;
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* End of extent will be smaller of the freespace end and the
|
|
|
|
* maximal requested end.
|
2010-12-10 16:03:57 +01:00
|
|
|
*
|
2005-04-17 00:20:36 +02:00
|
|
|
* Fix the length according to mod and prod if given.
|
|
|
|
*/
|
2011-04-24 21:06:15 +02:00
|
|
|
end = XFS_AGBLOCK_MIN(tend, args->agbno + args->maxlen);
|
2005-04-17 00:20:36 +02:00
|
|
|
args->len = end - args->agbno;
|
|
|
|
xfs_alloc_fix_len(args);
|
2010-12-10 16:03:57 +01:00
|
|
|
if (!xfs_alloc_fix_minleft(args))
|
|
|
|
goto not_found;
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
rlen = args->len;
|
2011-04-24 21:06:15 +02:00
|
|
|
ASSERT(args->agbno + rlen <= tend);
|
2005-04-17 00:20:36 +02:00
|
|
|
end = args->agbno + rlen;
|
2010-12-10 16:03:57 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* We are allocating agbno for rlen [agbno .. end]
|
|
|
|
* Allocate/initialize a cursor for the by-size btree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
cnt_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
|
|
|
|
args->agno, XFS_BTNUM_CNT);
|
2005-04-17 00:20:36 +02:00
|
|
|
ASSERT(args->agbno + args->len <=
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length));
|
2010-12-10 16:03:57 +01:00
|
|
|
error = xfs_alloc_fixup_trees(cnt_cur, bno_cur, fbno, flen, args->agbno,
|
|
|
|
args->len, XFSA_FIXUP_BNO_OK);
|
|
|
|
if (error) {
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
|
|
|
|
goto error0;
|
|
|
|
}
|
2010-12-10 16:03:57 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
2009-12-15 00:14:59 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
args->wasfromfl = 0;
|
2010-12-10 16:03:57 +01:00
|
|
|
trace_xfs_alloc_exact_done(args);
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
not_found:
|
|
|
|
/* Didn't find it, return null. */
|
|
|
|
xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
|
|
|
|
args->agbno = NULLAGBLOCK;
|
|
|
|
trace_xfs_alloc_exact_notfound(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
error0:
|
|
|
|
xfs_btree_del_cursor(bno_cur, XFS_BTREE_ERROR);
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_exact_error(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
2010-12-10 16:04:11 +01:00
|
|
|
/*
|
|
|
|
* Search the btree in a given direction via the search cursor and compare
|
|
|
|
* the records found against the good extent we've already found.
|
|
|
|
*/
|
|
|
|
STATIC int
|
|
|
|
xfs_alloc_find_best_extent(
|
|
|
|
struct xfs_alloc_arg *args, /* allocation argument structure */
|
|
|
|
struct xfs_btree_cur **gcur, /* good cursor */
|
|
|
|
struct xfs_btree_cur **scur, /* searching cursor */
|
|
|
|
xfs_agblock_t gdiff, /* difference for search comparison */
|
|
|
|
xfs_agblock_t *sbno, /* extent found by search */
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_extlen_t *slen, /* extent length */
|
|
|
|
xfs_agblock_t *sbnoa, /* aligned extent found by search */
|
|
|
|
xfs_extlen_t *slena, /* aligned extent length */
|
2010-12-10 16:04:11 +01:00
|
|
|
int dir) /* 0 = search right, 1 = search left */
|
|
|
|
{
|
|
|
|
xfs_agblock_t new;
|
|
|
|
xfs_agblock_t sdiff;
|
|
|
|
int error;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
/* The good extent is perfect, no need to search. */
|
|
|
|
if (!gdiff)
|
|
|
|
goto out_use_good;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Look until we find a better one, run out of space or run off the end.
|
|
|
|
*/
|
|
|
|
do {
|
|
|
|
error = xfs_alloc_get_rec(*scur, sbno, slen, &i);
|
|
|
|
if (error)
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_alloc_compute_aligned(args, *sbno, *slen, sbnoa, slena);
|
2010-12-10 16:04:11 +01:00
|
|
|
|
|
|
|
/*
|
|
|
|
* The good extent is closer than this one.
|
|
|
|
*/
|
|
|
|
if (!dir) {
|
2011-04-24 21:06:15 +02:00
|
|
|
if (*sbnoa >= args->agbno + gdiff)
|
2010-12-10 16:04:11 +01:00
|
|
|
goto out_use_good;
|
|
|
|
} else {
|
2011-04-24 21:06:15 +02:00
|
|
|
if (*sbnoa <= args->agbno - gdiff)
|
2010-12-10 16:04:11 +01:00
|
|
|
goto out_use_good;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Same distance, compare length and pick the best.
|
|
|
|
*/
|
|
|
|
if (*slena >= args->minlen) {
|
|
|
|
args->len = XFS_EXTLEN_MIN(*slena, args->maxlen);
|
|
|
|
xfs_alloc_fix_len(args);
|
|
|
|
|
|
|
|
sdiff = xfs_alloc_compute_diff(args->agbno, args->len,
|
2011-04-24 21:06:15 +02:00
|
|
|
args->alignment, *sbnoa,
|
|
|
|
*slena, &new);
|
2010-12-10 16:04:11 +01:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Choose closer size and invalidate other cursor.
|
|
|
|
*/
|
|
|
|
if (sdiff < gdiff)
|
|
|
|
goto out_use_search;
|
|
|
|
goto out_use_good;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!dir)
|
|
|
|
error = xfs_btree_increment(*scur, 0, &i);
|
|
|
|
else
|
|
|
|
error = xfs_btree_decrement(*scur, 0, &i);
|
|
|
|
if (error)
|
|
|
|
goto error0;
|
|
|
|
} while (i);
|
|
|
|
|
|
|
|
out_use_good:
|
|
|
|
xfs_btree_del_cursor(*scur, XFS_BTREE_NOERROR);
|
|
|
|
*scur = NULL;
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
out_use_search:
|
|
|
|
xfs_btree_del_cursor(*gcur, XFS_BTREE_NOERROR);
|
|
|
|
*gcur = NULL;
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
error0:
|
|
|
|
/* caller invalidates cursors */
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Allocate a variable extent near bno in the allocation group agno.
|
|
|
|
* Extent's length (returned in len) will be between minlen and maxlen,
|
|
|
|
* and of the form k * prod + mod unless there's nothing that large.
|
|
|
|
* Return the starting a.g. block, or NULLAGBLOCK if we can't do it.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_ag_vextent_near(
|
|
|
|
xfs_alloc_arg_t *args) /* allocation argument structure */
|
|
|
|
{
|
|
|
|
xfs_btree_cur_t *bno_cur_gt; /* cursor for bno btree, right side */
|
|
|
|
xfs_btree_cur_t *bno_cur_lt; /* cursor for bno btree, left side */
|
|
|
|
xfs_btree_cur_t *cnt_cur; /* cursor for count btree */
|
|
|
|
xfs_agblock_t gtbno; /* start bno of right side entry */
|
|
|
|
xfs_agblock_t gtbnoa; /* aligned ... */
|
|
|
|
xfs_extlen_t gtdiff; /* difference to right side entry */
|
|
|
|
xfs_extlen_t gtlen; /* length of right side entry */
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_extlen_t gtlena; /* aligned ... */
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_agblock_t gtnew; /* useful start bno of right side */
|
|
|
|
int error; /* error code */
|
|
|
|
int i; /* result code, temporary */
|
|
|
|
int j; /* result code, temporary */
|
|
|
|
xfs_agblock_t ltbno; /* start bno of left side entry */
|
|
|
|
xfs_agblock_t ltbnoa; /* aligned ... */
|
|
|
|
xfs_extlen_t ltdiff; /* difference to left side entry */
|
|
|
|
xfs_extlen_t ltlen; /* length of left side entry */
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_extlen_t ltlena; /* aligned ... */
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_agblock_t ltnew; /* useful start bno of left side */
|
|
|
|
xfs_extlen_t rlen; /* length of returned extent */
|
2011-04-24 21:06:15 +02:00
|
|
|
int forced = 0;
|
2005-04-17 00:20:36 +02:00
|
|
|
#if defined(DEBUG) && defined(__KERNEL__)
|
|
|
|
/*
|
|
|
|
* Randomly don't execute the first algorithm.
|
|
|
|
*/
|
|
|
|
int dofirst; /* set to do first algorithm */
|
|
|
|
|
2007-05-08 05:49:03 +02:00
|
|
|
dofirst = random32() & 1;
|
2005-04-17 00:20:36 +02:00
|
|
|
#endif
|
2011-04-24 21:06:15 +02:00
|
|
|
|
|
|
|
restart:
|
|
|
|
bno_cur_lt = NULL;
|
|
|
|
bno_cur_gt = NULL;
|
|
|
|
ltlen = 0;
|
|
|
|
gtlena = 0;
|
|
|
|
ltlena = 0;
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Get a cursor for the by-size btree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
cnt_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
|
|
|
|
args->agno, XFS_BTNUM_CNT);
|
2011-04-24 21:06:15 +02:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* See if there are any free extents as big as maxlen.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_ge(cnt_cur, 0, args->maxlen, &i)))
|
|
|
|
goto error0;
|
|
|
|
/*
|
|
|
|
* If none, then pick up the last entry in the tree unless the
|
|
|
|
* tree is empty.
|
|
|
|
*/
|
|
|
|
if (!i) {
|
|
|
|
if ((error = xfs_alloc_ag_vextent_small(args, cnt_cur, <bno,
|
|
|
|
<len, &i)))
|
|
|
|
goto error0;
|
|
|
|
if (i == 0 || ltlen == 0) {
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
2011-04-24 21:06:15 +02:00
|
|
|
trace_xfs_alloc_near_noentry(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
ASSERT(i == 1);
|
|
|
|
}
|
|
|
|
args->wasfromfl = 0;
|
2011-04-24 21:06:15 +02:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* First algorithm.
|
|
|
|
* If the requested extent is large wrt the freespaces available
|
|
|
|
* in this a.g., then the cursor will be pointing to a btree entry
|
|
|
|
* near the right edge of the tree. If it's in the last btree leaf
|
|
|
|
* block, then we just examine all the entries in that block
|
|
|
|
* that are big enough, and pick the best one.
|
|
|
|
* This is written as a while loop so we can break out of it,
|
|
|
|
* but we never loop back to the top.
|
|
|
|
*/
|
|
|
|
while (xfs_btree_islastblock(cnt_cur, 0)) {
|
|
|
|
xfs_extlen_t bdiff;
|
|
|
|
int besti=0;
|
|
|
|
xfs_extlen_t blen=0;
|
|
|
|
xfs_agblock_t bnew=0;
|
|
|
|
|
|
|
|
#if defined(DEBUG) && defined(__KERNEL__)
|
|
|
|
if (!dofirst)
|
|
|
|
break;
|
|
|
|
#endif
|
|
|
|
/*
|
|
|
|
* Start from the entry that lookup found, sequence through
|
|
|
|
* all larger free blocks. If we're actually pointing at a
|
|
|
|
* record smaller than maxlen, go to the start of this block,
|
|
|
|
* and skip all those smaller than minlen.
|
|
|
|
*/
|
|
|
|
if (ltlen || args->alignment > 1) {
|
|
|
|
cnt_cur->bc_ptrs[0] = 1;
|
|
|
|
do {
|
|
|
|
if ((error = xfs_alloc_get_rec(cnt_cur, <bno,
|
|
|
|
<len, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
if (ltlen >= args->minlen)
|
|
|
|
break;
|
2008-10-30 06:55:45 +01:00
|
|
|
if ((error = xfs_btree_increment(cnt_cur, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
} while (i);
|
|
|
|
ASSERT(ltlen >= args->minlen);
|
|
|
|
if (!i)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
i = cnt_cur->bc_ptrs[0];
|
|
|
|
for (j = 1, blen = 0, bdiff = 0;
|
|
|
|
!error && j && (blen < args->maxlen || bdiff > 0);
|
2008-10-30 06:55:45 +01:00
|
|
|
error = xfs_btree_increment(cnt_cur, 0, &j)) {
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* For each entry, decide if it's better than
|
|
|
|
* the previous best entry.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_get_rec(cnt_cur, <bno, <len, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2011-03-04 13:59:54 +01:00
|
|
|
xfs_alloc_compute_aligned(args, ltbno, ltlen,
|
|
|
|
<bnoa, <lena);
|
2008-04-17 08:49:49 +02:00
|
|
|
if (ltlena < args->minlen)
|
2005-04-17 00:20:36 +02:00
|
|
|
continue;
|
|
|
|
args->len = XFS_EXTLEN_MIN(ltlena, args->maxlen);
|
|
|
|
xfs_alloc_fix_len(args);
|
|
|
|
ASSERT(args->len >= args->minlen);
|
|
|
|
if (args->len < blen)
|
|
|
|
continue;
|
|
|
|
ltdiff = xfs_alloc_compute_diff(args->agbno, args->len,
|
2011-04-24 21:06:15 +02:00
|
|
|
args->alignment, ltbnoa, ltlena, <new);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (ltnew != NULLAGBLOCK &&
|
|
|
|
(args->len > blen || ltdiff < bdiff)) {
|
|
|
|
bdiff = ltdiff;
|
|
|
|
bnew = ltnew;
|
|
|
|
blen = args->len;
|
|
|
|
besti = cnt_cur->bc_ptrs[0];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* It didn't work. We COULD be in a case where
|
|
|
|
* there's a good record somewhere, so try again.
|
|
|
|
*/
|
|
|
|
if (blen == 0)
|
|
|
|
break;
|
|
|
|
/*
|
|
|
|
* Point at the best entry, and retrieve it again.
|
|
|
|
*/
|
|
|
|
cnt_cur->bc_ptrs[0] = besti;
|
|
|
|
if ((error = xfs_alloc_get_rec(cnt_cur, <bno, <len, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2010-07-20 09:54:45 +02:00
|
|
|
ASSERT(ltbno + ltlen <= be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length));
|
2005-04-17 00:20:36 +02:00
|
|
|
args->len = blen;
|
|
|
|
if (!xfs_alloc_fix_minleft(args)) {
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_near_nominleft(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
blen = args->len;
|
|
|
|
/*
|
|
|
|
* We are allocating starting at bnew for blen blocks.
|
|
|
|
*/
|
|
|
|
args->agbno = bnew;
|
|
|
|
ASSERT(bnew >= ltbno);
|
2010-07-20 09:54:45 +02:00
|
|
|
ASSERT(bnew + blen <= ltbno + ltlen);
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Set up a cursor for the by-bno tree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
bno_cur_lt = xfs_allocbt_init_cursor(args->mp, args->tp,
|
|
|
|
args->agbp, args->agno, XFS_BTNUM_BNO);
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Fix up the btree entries.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_fixup_trees(cnt_cur, bno_cur_lt, ltbno,
|
|
|
|
ltlen, bnew, blen, XFSA_FIXUP_CNT_OK)))
|
|
|
|
goto error0;
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
|
|
|
xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_NOERROR);
|
2009-12-15 00:14:59 +01:00
|
|
|
|
|
|
|
trace_xfs_alloc_near_first(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Second algorithm.
|
|
|
|
* Search in the by-bno tree to the left and to the right
|
|
|
|
* simultaneously, until in each case we find a space big enough,
|
|
|
|
* or run into the edge of the tree. When we run into the edge,
|
|
|
|
* we deallocate that cursor.
|
|
|
|
* If both searches succeed, we compare the two spaces and pick
|
|
|
|
* the better one.
|
|
|
|
* With alignment, it's possible for both to fail; the upper
|
|
|
|
* level algorithm that picks allocation groups for allocations
|
|
|
|
* is not supposed to do this.
|
|
|
|
*/
|
|
|
|
/*
|
|
|
|
* Allocate and initialize the cursor for the leftward search.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
bno_cur_lt = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
|
|
|
|
args->agno, XFS_BTNUM_BNO);
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Lookup <= bno to find the leftward search's starting point.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_le(bno_cur_lt, args->agbno, args->maxlen, &i)))
|
|
|
|
goto error0;
|
|
|
|
if (!i) {
|
|
|
|
/*
|
|
|
|
* Didn't find anything; use this cursor for the rightward
|
|
|
|
* search.
|
|
|
|
*/
|
|
|
|
bno_cur_gt = bno_cur_lt;
|
|
|
|
bno_cur_lt = NULL;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Found something. Duplicate the cursor for the rightward search.
|
|
|
|
*/
|
|
|
|
else if ((error = xfs_btree_dup_cursor(bno_cur_lt, &bno_cur_gt)))
|
|
|
|
goto error0;
|
|
|
|
/*
|
|
|
|
* Increment the cursor, so we will point at the entry just right
|
|
|
|
* of the leftward entry if any, or to the leftmost entry.
|
|
|
|
*/
|
2008-10-30 06:55:45 +01:00
|
|
|
if ((error = xfs_btree_increment(bno_cur_gt, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (!i) {
|
|
|
|
/*
|
|
|
|
* It failed, there are no rightward entries.
|
|
|
|
*/
|
|
|
|
xfs_btree_del_cursor(bno_cur_gt, XFS_BTREE_NOERROR);
|
|
|
|
bno_cur_gt = NULL;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Loop going left with the leftward cursor, right with the
|
|
|
|
* rightward cursor, until either both directions give up or
|
|
|
|
* we find an entry at least as big as minlen.
|
|
|
|
*/
|
|
|
|
do {
|
|
|
|
if (bno_cur_lt) {
|
|
|
|
if ((error = xfs_alloc_get_rec(bno_cur_lt, <bno, <len, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2011-03-04 13:59:54 +01:00
|
|
|
xfs_alloc_compute_aligned(args, ltbno, ltlen,
|
|
|
|
<bnoa, <lena);
|
2008-04-10 04:21:32 +02:00
|
|
|
if (ltlena >= args->minlen)
|
2005-04-17 00:20:36 +02:00
|
|
|
break;
|
2008-10-30 06:55:58 +01:00
|
|
|
if ((error = xfs_btree_decrement(bno_cur_lt, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (!i) {
|
|
|
|
xfs_btree_del_cursor(bno_cur_lt,
|
|
|
|
XFS_BTREE_NOERROR);
|
|
|
|
bno_cur_lt = NULL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (bno_cur_gt) {
|
|
|
|
if ((error = xfs_alloc_get_rec(bno_cur_gt, >bno, >len, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2011-03-04 13:59:54 +01:00
|
|
|
xfs_alloc_compute_aligned(args, gtbno, gtlen,
|
|
|
|
>bnoa, >lena);
|
2008-04-10 04:21:32 +02:00
|
|
|
if (gtlena >= args->minlen)
|
2005-04-17 00:20:36 +02:00
|
|
|
break;
|
2008-10-30 06:55:45 +01:00
|
|
|
if ((error = xfs_btree_increment(bno_cur_gt, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (!i) {
|
|
|
|
xfs_btree_del_cursor(bno_cur_gt,
|
|
|
|
XFS_BTREE_NOERROR);
|
|
|
|
bno_cur_gt = NULL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} while (bno_cur_lt || bno_cur_gt);
|
2010-12-10 16:04:11 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Got both cursors still active, need to find better entry.
|
|
|
|
*/
|
|
|
|
if (bno_cur_lt && bno_cur_gt) {
|
|
|
|
if (ltlena >= args->minlen) {
|
|
|
|
/*
|
2010-12-10 16:04:11 +01:00
|
|
|
* Left side is good, look for a right side entry.
|
2005-04-17 00:20:36 +02:00
|
|
|
*/
|
|
|
|
args->len = XFS_EXTLEN_MIN(ltlena, args->maxlen);
|
|
|
|
xfs_alloc_fix_len(args);
|
2010-12-10 16:04:11 +01:00
|
|
|
ltdiff = xfs_alloc_compute_diff(args->agbno, args->len,
|
2011-04-24 21:06:15 +02:00
|
|
|
args->alignment, ltbnoa, ltlena, <new);
|
2010-12-10 16:04:11 +01:00
|
|
|
|
|
|
|
error = xfs_alloc_find_best_extent(args,
|
|
|
|
&bno_cur_lt, &bno_cur_gt,
|
2011-04-24 21:06:15 +02:00
|
|
|
ltdiff, >bno, >len,
|
|
|
|
>bnoa, >lena,
|
2010-12-10 16:04:11 +01:00
|
|
|
0 /* search right */);
|
|
|
|
} else {
|
|
|
|
ASSERT(gtlena >= args->minlen);
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
2010-12-10 16:04:11 +01:00
|
|
|
* Right side is good, look for a left side entry.
|
2005-04-17 00:20:36 +02:00
|
|
|
*/
|
|
|
|
args->len = XFS_EXTLEN_MIN(gtlena, args->maxlen);
|
|
|
|
xfs_alloc_fix_len(args);
|
2010-12-10 16:04:11 +01:00
|
|
|
gtdiff = xfs_alloc_compute_diff(args->agbno, args->len,
|
2011-04-24 21:06:15 +02:00
|
|
|
args->alignment, gtbnoa, gtlena, >new);
|
2010-12-10 16:04:11 +01:00
|
|
|
|
|
|
|
error = xfs_alloc_find_best_extent(args,
|
|
|
|
&bno_cur_gt, &bno_cur_lt,
|
2011-04-24 21:06:15 +02:00
|
|
|
gtdiff, <bno, <len,
|
|
|
|
<bnoa, <lena,
|
2010-12-10 16:04:11 +01:00
|
|
|
1 /* search left */);
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2010-12-10 16:04:11 +01:00
|
|
|
|
|
|
|
if (error)
|
|
|
|
goto error0;
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2010-12-10 16:04:11 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* If we couldn't get anything, give up.
|
|
|
|
*/
|
|
|
|
if (bno_cur_lt == NULL && bno_cur_gt == NULL) {
|
2011-04-24 21:06:15 +02:00
|
|
|
if (!forced++) {
|
|
|
|
trace_xfs_alloc_near_busy(args);
|
|
|
|
xfs_log_force(args->mp, XFS_LOG_SYNC);
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_size_neither(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
args->agbno = NULLAGBLOCK;
|
|
|
|
return 0;
|
|
|
|
}
|
2010-12-10 16:04:11 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* At this point we have selected a freespace entry, either to the
|
|
|
|
* left or to the right. If it's on the right, copy all the
|
|
|
|
* useful variables to the "left" set so we only have one
|
|
|
|
* copy of this code.
|
|
|
|
*/
|
|
|
|
if (bno_cur_gt) {
|
|
|
|
bno_cur_lt = bno_cur_gt;
|
|
|
|
bno_cur_gt = NULL;
|
|
|
|
ltbno = gtbno;
|
|
|
|
ltbnoa = gtbnoa;
|
|
|
|
ltlen = gtlen;
|
|
|
|
ltlena = gtlena;
|
|
|
|
j = 1;
|
|
|
|
} else
|
|
|
|
j = 0;
|
2010-12-10 16:04:11 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Fix up the length and compute the useful address.
|
|
|
|
*/
|
|
|
|
args->len = XFS_EXTLEN_MIN(ltlena, args->maxlen);
|
|
|
|
xfs_alloc_fix_len(args);
|
|
|
|
if (!xfs_alloc_fix_minleft(args)) {
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_near_nominleft(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_NOERROR);
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
rlen = args->len;
|
2011-04-24 21:06:15 +02:00
|
|
|
(void)xfs_alloc_compute_diff(args->agbno, rlen, args->alignment,
|
|
|
|
ltbnoa, ltlena, <new);
|
2005-04-17 00:20:36 +02:00
|
|
|
ASSERT(ltnew >= ltbno);
|
2011-04-24 21:06:15 +02:00
|
|
|
ASSERT(ltnew + rlen <= ltbnoa + ltlena);
|
2005-11-02 05:11:25 +01:00
|
|
|
ASSERT(ltnew + rlen <= be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length));
|
2005-04-17 00:20:36 +02:00
|
|
|
args->agbno = ltnew;
|
2011-04-24 21:06:15 +02:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
if ((error = xfs_alloc_fixup_trees(cnt_cur, bno_cur_lt, ltbno, ltlen,
|
|
|
|
ltnew, rlen, XFSA_FIXUP_BNO_OK)))
|
|
|
|
goto error0;
|
2009-12-15 00:14:59 +01:00
|
|
|
|
|
|
|
if (j)
|
|
|
|
trace_xfs_alloc_near_greater(args);
|
|
|
|
else
|
|
|
|
trace_xfs_alloc_near_lesser(args);
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
|
|
|
xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_NOERROR);
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
error0:
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_near_error(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (cnt_cur != NULL)
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
|
|
|
|
if (bno_cur_lt != NULL)
|
|
|
|
xfs_btree_del_cursor(bno_cur_lt, XFS_BTREE_ERROR);
|
|
|
|
if (bno_cur_gt != NULL)
|
|
|
|
xfs_btree_del_cursor(bno_cur_gt, XFS_BTREE_ERROR);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allocate a variable extent anywhere in the allocation group agno.
|
|
|
|
* Extent's length (returned in len) will be between minlen and maxlen,
|
|
|
|
* and of the form k * prod + mod unless there's nothing that large.
|
|
|
|
* Return the starting a.g. block, or NULLAGBLOCK if we can't do it.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_ag_vextent_size(
|
|
|
|
xfs_alloc_arg_t *args) /* allocation argument structure */
|
|
|
|
{
|
|
|
|
xfs_btree_cur_t *bno_cur; /* cursor for bno btree */
|
|
|
|
xfs_btree_cur_t *cnt_cur; /* cursor for cnt btree */
|
|
|
|
int error; /* error result */
|
|
|
|
xfs_agblock_t fbno; /* start of found freespace */
|
|
|
|
xfs_extlen_t flen; /* length of found freespace */
|
|
|
|
int i; /* temp status variable */
|
|
|
|
xfs_agblock_t rbno; /* returned block number */
|
|
|
|
xfs_extlen_t rlen; /* length of returned extent */
|
2011-04-24 21:06:15 +02:00
|
|
|
int forced = 0;
|
2005-04-17 00:20:36 +02:00
|
|
|
|
2011-04-24 21:06:15 +02:00
|
|
|
restart:
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Allocate and initialize a cursor for the by-size btree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
cnt_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
|
|
|
|
args->agno, XFS_BTNUM_CNT);
|
2005-04-17 00:20:36 +02:00
|
|
|
bno_cur = NULL;
|
2011-04-24 21:06:15 +02:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Look for an entry >= maxlen+alignment-1 blocks.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_ge(cnt_cur, 0,
|
|
|
|
args->maxlen + args->alignment - 1, &i)))
|
|
|
|
goto error0;
|
2011-04-24 21:06:15 +02:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
2011-04-24 21:06:15 +02:00
|
|
|
* If none or we have busy extents that we cannot allocate from, then
|
|
|
|
* we have to settle for a smaller extent. In the case that there are
|
|
|
|
* no large extents, this will return the last entry in the tree unless
|
|
|
|
* the tree is empty. In the case that there are only busy large
|
|
|
|
* extents, this will return the largest small extent unless there
|
|
|
|
* are no smaller extents available.
|
2005-04-17 00:20:36 +02:00
|
|
|
*/
|
2011-04-24 21:06:15 +02:00
|
|
|
if (!i || forced > 1) {
|
|
|
|
error = xfs_alloc_ag_vextent_small(args, cnt_cur,
|
|
|
|
&fbno, &flen, &i);
|
|
|
|
if (error)
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (i == 0 || flen == 0) {
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_size_noentry(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
ASSERT(i == 1);
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_alloc_compute_aligned(args, fbno, flen, &rbno, &rlen);
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* Search for a non-busy extent that is large enough.
|
|
|
|
* If we are at low space, don't check, or if we fall of
|
|
|
|
* the end of the btree, turn off the busy check and
|
|
|
|
* restart.
|
|
|
|
*/
|
|
|
|
for (;;) {
|
|
|
|
error = xfs_alloc_get_rec(cnt_cur, &fbno, &flen, &i);
|
|
|
|
if (error)
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
|
|
|
|
xfs_alloc_compute_aligned(args, fbno, flen,
|
|
|
|
&rbno, &rlen);
|
|
|
|
|
|
|
|
if (rlen >= args->maxlen)
|
|
|
|
break;
|
|
|
|
|
|
|
|
error = xfs_btree_increment(cnt_cur, 0, &i);
|
|
|
|
if (error)
|
|
|
|
goto error0;
|
|
|
|
if (i == 0) {
|
|
|
|
/*
|
|
|
|
* Our only valid extents must have been busy.
|
|
|
|
* Make it unbusy by forcing the log out and
|
|
|
|
* retrying. If we've been here before, forcing
|
|
|
|
* the log isn't making the extents available,
|
|
|
|
* which means they have probably been freed in
|
|
|
|
* this transaction. In that case, we have to
|
|
|
|
* give up on them and we'll attempt a minlen
|
|
|
|
* allocation the next time around.
|
|
|
|
*/
|
|
|
|
xfs_btree_del_cursor(cnt_cur,
|
|
|
|
XFS_BTREE_NOERROR);
|
|
|
|
trace_xfs_alloc_size_busy(args);
|
|
|
|
if (!forced++)
|
|
|
|
xfs_log_force(args->mp, XFS_LOG_SYNC);
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2011-04-24 21:06:15 +02:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* In the first case above, we got the last entry in the
|
|
|
|
* by-size btree. Now we check to see if the space hits maxlen
|
|
|
|
* once aligned; if not, we search left for something better.
|
|
|
|
* This can't happen in the second case above.
|
|
|
|
*/
|
|
|
|
rlen = XFS_EXTLEN_MIN(args->maxlen, rlen);
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(rlen == 0 ||
|
|
|
|
(rlen <= flen && rbno + rlen <= fbno + flen), error0);
|
|
|
|
if (rlen < args->maxlen) {
|
|
|
|
xfs_agblock_t bestfbno;
|
|
|
|
xfs_extlen_t bestflen;
|
|
|
|
xfs_agblock_t bestrbno;
|
|
|
|
xfs_extlen_t bestrlen;
|
|
|
|
|
|
|
|
bestrlen = rlen;
|
|
|
|
bestrbno = rbno;
|
|
|
|
bestflen = flen;
|
|
|
|
bestfbno = fbno;
|
|
|
|
for (;;) {
|
2008-10-30 06:55:58 +01:00
|
|
|
if ((error = xfs_btree_decrement(cnt_cur, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (i == 0)
|
|
|
|
break;
|
|
|
|
if ((error = xfs_alloc_get_rec(cnt_cur, &fbno, &flen,
|
|
|
|
&i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
if (flen < bestrlen)
|
|
|
|
break;
|
2011-03-04 13:59:54 +01:00
|
|
|
xfs_alloc_compute_aligned(args, fbno, flen,
|
|
|
|
&rbno, &rlen);
|
2005-04-17 00:20:36 +02:00
|
|
|
rlen = XFS_EXTLEN_MIN(args->maxlen, rlen);
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(rlen == 0 ||
|
|
|
|
(rlen <= flen && rbno + rlen <= fbno + flen),
|
|
|
|
error0);
|
|
|
|
if (rlen > bestrlen) {
|
|
|
|
bestrlen = rlen;
|
|
|
|
bestrbno = rbno;
|
|
|
|
bestflen = flen;
|
|
|
|
bestfbno = fbno;
|
|
|
|
if (rlen == args->maxlen)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, bestfbno, bestflen,
|
|
|
|
&i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
rlen = bestrlen;
|
|
|
|
rbno = bestrbno;
|
|
|
|
flen = bestflen;
|
|
|
|
fbno = bestfbno;
|
|
|
|
}
|
|
|
|
args->wasfromfl = 0;
|
|
|
|
/*
|
|
|
|
* Fix up the length.
|
|
|
|
*/
|
|
|
|
args->len = rlen;
|
2011-04-24 21:06:15 +02:00
|
|
|
if (rlen < args->minlen) {
|
|
|
|
if (!forced++) {
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
|
|
|
trace_xfs_alloc_size_busy(args);
|
|
|
|
xfs_log_force(args->mp, XFS_LOG_SYNC);
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
goto out_nominleft;
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2011-04-24 21:06:15 +02:00
|
|
|
xfs_alloc_fix_len(args);
|
|
|
|
|
|
|
|
if (!xfs_alloc_fix_minleft(args))
|
|
|
|
goto out_nominleft;
|
2005-04-17 00:20:36 +02:00
|
|
|
rlen = args->len;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(rlen <= flen, error0);
|
|
|
|
/*
|
|
|
|
* Allocate and initialize a cursor for the by-block tree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
bno_cur = xfs_allocbt_init_cursor(args->mp, args->tp, args->agbp,
|
|
|
|
args->agno, XFS_BTNUM_BNO);
|
2005-04-17 00:20:36 +02:00
|
|
|
if ((error = xfs_alloc_fixup_trees(cnt_cur, bno_cur, fbno, flen,
|
|
|
|
rbno, rlen, XFSA_FIXUP_CNT_OK)))
|
|
|
|
goto error0;
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
|
|
|
xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
|
|
|
|
cnt_cur = bno_cur = NULL;
|
|
|
|
args->len = rlen;
|
|
|
|
args->agbno = rbno;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(
|
|
|
|
args->agbno + args->len <=
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length),
|
2005-04-17 00:20:36 +02:00
|
|
|
error0);
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_size_done(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
error0:
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_size_error(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (cnt_cur)
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
|
|
|
|
if (bno_cur)
|
|
|
|
xfs_btree_del_cursor(bno_cur, XFS_BTREE_ERROR);
|
|
|
|
return error;
|
2011-04-24 21:06:15 +02:00
|
|
|
|
|
|
|
out_nominleft:
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
|
|
|
trace_xfs_alloc_size_nominleft(args);
|
|
|
|
args->agbno = NULLAGBLOCK;
|
|
|
|
return 0;
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Deal with the case where only small freespaces remain.
|
|
|
|
* Either return the contents of the last freespace record,
|
|
|
|
* or allocate space from the freelist if there is nothing in the tree.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_ag_vextent_small(
|
|
|
|
xfs_alloc_arg_t *args, /* allocation argument structure */
|
|
|
|
xfs_btree_cur_t *ccur, /* by-size cursor */
|
|
|
|
xfs_agblock_t *fbnop, /* result block number */
|
|
|
|
xfs_extlen_t *flenp, /* result length */
|
|
|
|
int *stat) /* status: 0-freelist, 1-normal/none */
|
|
|
|
{
|
|
|
|
int error;
|
|
|
|
xfs_agblock_t fbno;
|
|
|
|
xfs_extlen_t flen;
|
|
|
|
int i;
|
|
|
|
|
2008-10-30 06:55:58 +01:00
|
|
|
if ((error = xfs_btree_decrement(ccur, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (i) {
|
|
|
|
if ((error = xfs_alloc_get_rec(ccur, &fbno, &flen, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Nothing in the btree, try the freelist. Make sure
|
|
|
|
* to respect minleft even when pulling from the
|
|
|
|
* freelist.
|
|
|
|
*/
|
|
|
|
else if (args->minlen == 1 && args->alignment == 1 && !args->isfl &&
|
2005-11-02 05:11:25 +01:00
|
|
|
(be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_flcount)
|
|
|
|
> args->minleft)) {
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
error = xfs_alloc_get_freelist(args->tp, args->agbp, &fbno, 0);
|
|
|
|
if (error)
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (fbno != NULLAGBLOCK) {
|
2011-04-24 21:06:16 +02:00
|
|
|
xfs_alloc_busy_reuse(args->mp, args->agno, fbno, 1,
|
|
|
|
args->userdata);
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
if (args->userdata) {
|
|
|
|
xfs_buf_t *bp;
|
|
|
|
|
|
|
|
bp = xfs_btree_get_bufs(args->mp, args->tp,
|
|
|
|
args->agno, fbno, 0);
|
|
|
|
xfs_trans_binval(args->tp, bp);
|
|
|
|
}
|
|
|
|
args->len = 1;
|
|
|
|
args->agbno = fbno;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(
|
|
|
|
args->agbno + args->len <=
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length),
|
2005-04-17 00:20:36 +02:00
|
|
|
error0);
|
|
|
|
args->wasfromfl = 1;
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_small_freelist(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
*stat = 0;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Nothing in the freelist.
|
|
|
|
*/
|
|
|
|
else
|
|
|
|
flen = 0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Can't allocate from the freelist for some reason.
|
|
|
|
*/
|
2006-09-28 03:03:44 +02:00
|
|
|
else {
|
|
|
|
fbno = NULLAGBLOCK;
|
2005-04-17 00:20:36 +02:00
|
|
|
flen = 0;
|
2006-09-28 03:03:44 +02:00
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Can't do the allocation, give up.
|
|
|
|
*/
|
|
|
|
if (flen < args->minlen) {
|
|
|
|
args->agbno = NULLAGBLOCK;
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_small_notenough(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
flen = 0;
|
|
|
|
}
|
|
|
|
*fbnop = fbno;
|
|
|
|
*flenp = flen;
|
|
|
|
*stat = 1;
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_small_done(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
error0:
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_small_error(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Free the extent starting at agno/bno for length.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_free_ag_extent(
|
|
|
|
xfs_trans_t *tp, /* transaction pointer */
|
|
|
|
xfs_buf_t *agbp, /* buffer for a.g. freelist header */
|
|
|
|
xfs_agnumber_t agno, /* allocation group number */
|
|
|
|
xfs_agblock_t bno, /* starting block number */
|
|
|
|
xfs_extlen_t len, /* length of extent */
|
|
|
|
int isfl) /* set if is freelist blocks - no sb acctg */
|
|
|
|
{
|
|
|
|
xfs_btree_cur_t *bno_cur; /* cursor for by-block btree */
|
|
|
|
xfs_btree_cur_t *cnt_cur; /* cursor for by-size btree */
|
|
|
|
int error; /* error return value */
|
|
|
|
xfs_agblock_t gtbno; /* start of right neighbor block */
|
|
|
|
xfs_extlen_t gtlen; /* length of right neighbor block */
|
|
|
|
int haveleft; /* have a left neighbor block */
|
|
|
|
int haveright; /* have a right neighbor block */
|
|
|
|
int i; /* temp, result code */
|
|
|
|
xfs_agblock_t ltbno; /* start of left neighbor block */
|
|
|
|
xfs_extlen_t ltlen; /* length of left neighbor block */
|
|
|
|
xfs_mount_t *mp; /* mount point struct for filesystem */
|
|
|
|
xfs_agblock_t nbno; /* new starting block of freespace */
|
|
|
|
xfs_extlen_t nlen; /* new length of freespace */
|
2011-03-04 13:59:55 +01:00
|
|
|
xfs_perag_t *pag; /* per allocation group data */
|
2005-04-17 00:20:36 +02:00
|
|
|
|
|
|
|
mp = tp->t_mountp;
|
|
|
|
/*
|
|
|
|
* Allocate and initialize a cursor for the by-block btree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
bno_cur = xfs_allocbt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_BNO);
|
2005-04-17 00:20:36 +02:00
|
|
|
cnt_cur = NULL;
|
|
|
|
/*
|
|
|
|
* Look for a neighboring block on the left (lower block numbers)
|
|
|
|
* that is contiguous with this space.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_le(bno_cur, bno, len, &haveleft)))
|
|
|
|
goto error0;
|
|
|
|
if (haveleft) {
|
|
|
|
/*
|
|
|
|
* There is a block to our left.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_get_rec(bno_cur, <bno, <len, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
/*
|
|
|
|
* It's not contiguous, though.
|
|
|
|
*/
|
|
|
|
if (ltbno + ltlen < bno)
|
|
|
|
haveleft = 0;
|
|
|
|
else {
|
|
|
|
/*
|
|
|
|
* If this failure happens the request to free this
|
|
|
|
* space was invalid, it's (partly) already free.
|
|
|
|
* Very bad.
|
|
|
|
*/
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(ltbno + ltlen <= bno, error0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Look for a neighboring block on the right (higher block numbers)
|
|
|
|
* that is contiguous with this space.
|
|
|
|
*/
|
2008-10-30 06:55:45 +01:00
|
|
|
if ((error = xfs_btree_increment(bno_cur, 0, &haveright)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
if (haveright) {
|
|
|
|
/*
|
|
|
|
* There is a block to our right.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_get_rec(bno_cur, >bno, >len, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
/*
|
|
|
|
* It's not contiguous, though.
|
|
|
|
*/
|
|
|
|
if (bno + len < gtbno)
|
|
|
|
haveright = 0;
|
|
|
|
else {
|
|
|
|
/*
|
|
|
|
* If this failure happens the request to free this
|
|
|
|
* space was invalid, it's (partly) already free.
|
|
|
|
* Very bad.
|
|
|
|
*/
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(gtbno >= bno + len, error0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Now allocate and initialize a cursor for the by-size tree.
|
|
|
|
*/
|
2008-10-30 06:53:59 +01:00
|
|
|
cnt_cur = xfs_allocbt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_CNT);
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Have both left and right contiguous neighbors.
|
|
|
|
* Merge all three into a single free block.
|
|
|
|
*/
|
|
|
|
if (haveleft && haveright) {
|
|
|
|
/*
|
|
|
|
* Delete the old by-size entry on the left.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, ltbno, ltlen, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2008-10-30 06:58:01 +01:00
|
|
|
if ((error = xfs_btree_delete(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
/*
|
|
|
|
* Delete the old by-size entry on the right.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, gtbno, gtlen, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2008-10-30 06:58:01 +01:00
|
|
|
if ((error = xfs_btree_delete(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
/*
|
|
|
|
* Delete the old by-block entry for the right block.
|
|
|
|
*/
|
2008-10-30 06:58:01 +01:00
|
|
|
if ((error = xfs_btree_delete(bno_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
/*
|
|
|
|
* Move the by-block cursor back to the left neighbor.
|
|
|
|
*/
|
2008-10-30 06:55:58 +01:00
|
|
|
if ((error = xfs_btree_decrement(bno_cur, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
#ifdef DEBUG
|
|
|
|
/*
|
|
|
|
* Check that this is the right record: delete didn't
|
|
|
|
* mangle the cursor.
|
|
|
|
*/
|
|
|
|
{
|
|
|
|
xfs_agblock_t xxbno;
|
|
|
|
xfs_extlen_t xxlen;
|
|
|
|
|
|
|
|
if ((error = xfs_alloc_get_rec(bno_cur, &xxbno, &xxlen,
|
|
|
|
&i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(
|
|
|
|
i == 1 && xxbno == ltbno && xxlen == ltlen,
|
|
|
|
error0);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
/*
|
|
|
|
* Update remaining by-block entry to the new, joined block.
|
|
|
|
*/
|
|
|
|
nbno = ltbno;
|
|
|
|
nlen = len + ltlen + gtlen;
|
|
|
|
if ((error = xfs_alloc_update(bno_cur, nbno, nlen)))
|
|
|
|
goto error0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Have only a left contiguous neighbor.
|
|
|
|
* Merge it together with the new freespace.
|
|
|
|
*/
|
|
|
|
else if (haveleft) {
|
|
|
|
/*
|
|
|
|
* Delete the old by-size entry on the left.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, ltbno, ltlen, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2008-10-30 06:58:01 +01:00
|
|
|
if ((error = xfs_btree_delete(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
/*
|
|
|
|
* Back up the by-block cursor to the left neighbor, and
|
|
|
|
* update its length.
|
|
|
|
*/
|
2008-10-30 06:55:58 +01:00
|
|
|
if ((error = xfs_btree_decrement(bno_cur, 0, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
nbno = ltbno;
|
|
|
|
nlen = len + ltlen;
|
|
|
|
if ((error = xfs_alloc_update(bno_cur, nbno, nlen)))
|
|
|
|
goto error0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Have only a right contiguous neighbor.
|
|
|
|
* Merge it together with the new freespace.
|
|
|
|
*/
|
|
|
|
else if (haveright) {
|
|
|
|
/*
|
|
|
|
* Delete the old by-size entry on the right.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, gtbno, gtlen, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
2008-10-30 06:58:01 +01:00
|
|
|
if ((error = xfs_btree_delete(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
/*
|
|
|
|
* Update the starting block and length of the right
|
|
|
|
* neighbor in the by-block tree.
|
|
|
|
*/
|
|
|
|
nbno = bno;
|
|
|
|
nlen = len + gtlen;
|
|
|
|
if ((error = xfs_alloc_update(bno_cur, nbno, nlen)))
|
|
|
|
goto error0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* No contiguous neighbors.
|
|
|
|
* Insert the new freespace into the by-block tree.
|
|
|
|
*/
|
|
|
|
else {
|
|
|
|
nbno = bno;
|
|
|
|
nlen = len;
|
2008-10-30 06:57:40 +01:00
|
|
|
if ((error = xfs_btree_insert(bno_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
}
|
|
|
|
xfs_btree_del_cursor(bno_cur, XFS_BTREE_NOERROR);
|
|
|
|
bno_cur = NULL;
|
|
|
|
/*
|
|
|
|
* In all cases we need to insert the new freespace in the by-size tree.
|
|
|
|
*/
|
|
|
|
if ((error = xfs_alloc_lookup_eq(cnt_cur, nbno, nlen, &i)))
|
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 0, error0);
|
2008-10-30 06:57:40 +01:00
|
|
|
if ((error = xfs_btree_insert(cnt_cur, &i)))
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
XFS_WANT_CORRUPTED_GOTO(i == 1, error0);
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_NOERROR);
|
|
|
|
cnt_cur = NULL;
|
2011-03-04 13:59:55 +01:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Update the freespace totals in the ag and superblock.
|
|
|
|
*/
|
2011-03-04 13:59:55 +01:00
|
|
|
pag = xfs_perag_get(mp, agno);
|
|
|
|
error = xfs_alloc_update_counters(tp, pag, agbp, len);
|
|
|
|
xfs_perag_put(pag);
|
|
|
|
if (error)
|
|
|
|
goto error0;
|
|
|
|
|
|
|
|
if (!isfl)
|
|
|
|
xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (long)len);
|
|
|
|
XFS_STATS_INC(xs_freex);
|
|
|
|
XFS_STATS_ADD(xs_freeb, len);
|
2009-12-15 00:14:59 +01:00
|
|
|
|
|
|
|
trace_xfs_free_extent(mp, agno, bno, len, isfl, haveleft, haveright);
|
2005-04-17 00:20:36 +02:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
error0:
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_free_extent(mp, agno, bno, len, isfl, -1, -1);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (bno_cur)
|
|
|
|
xfs_btree_del_cursor(bno_cur, XFS_BTREE_ERROR);
|
|
|
|
if (cnt_cur)
|
|
|
|
xfs_btree_del_cursor(cnt_cur, XFS_BTREE_ERROR);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Visible (exported) allocation/free functions.
|
|
|
|
* Some of these are used just by xfs_alloc_btree.c and this file.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compute and fill in value of m_ag_maxlevels.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
xfs_alloc_compute_maxlevels(
|
|
|
|
xfs_mount_t *mp) /* file system mount structure */
|
|
|
|
{
|
|
|
|
int level;
|
|
|
|
uint maxblocks;
|
|
|
|
uint maxleafents;
|
|
|
|
int minleafrecs;
|
|
|
|
int minnoderecs;
|
|
|
|
|
|
|
|
maxleafents = (mp->m_sb.sb_agblocks + 1) / 2;
|
|
|
|
minleafrecs = mp->m_alloc_mnr[0];
|
|
|
|
minnoderecs = mp->m_alloc_mnr[1];
|
|
|
|
maxblocks = (maxleafents + minleafrecs - 1) / minleafrecs;
|
|
|
|
for (level = 1; maxblocks > 1; level++)
|
|
|
|
maxblocks = (maxblocks + minnoderecs - 1) / minnoderecs;
|
|
|
|
mp->m_ag_maxlevels = level;
|
|
|
|
}
|
|
|
|
|
2009-03-16 08:29:46 +01:00
|
|
|
/*
|
|
|
|
* Find the length of the longest extent in an AG.
|
|
|
|
*/
|
|
|
|
xfs_extlen_t
|
|
|
|
xfs_alloc_longest_free_extent(
|
|
|
|
struct xfs_mount *mp,
|
|
|
|
struct xfs_perag *pag)
|
|
|
|
{
|
|
|
|
xfs_extlen_t need, delta = 0;
|
|
|
|
|
|
|
|
need = XFS_MIN_FREELIST_PAG(pag, mp);
|
|
|
|
if (need > pag->pagf_flcount)
|
|
|
|
delta = need - pag->pagf_flcount;
|
|
|
|
|
|
|
|
if (pag->pagf_longest > delta)
|
|
|
|
return pag->pagf_longest - delta;
|
|
|
|
return pag->pagf_flcount > 0 || pag->pagf_longest > 0;
|
|
|
|
}
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Decide whether to use this allocation group for this allocation.
|
|
|
|
* If so, fix up the btree freelist's size.
|
|
|
|
*/
|
|
|
|
STATIC int /* error */
|
|
|
|
xfs_alloc_fix_freelist(
|
|
|
|
xfs_alloc_arg_t *args, /* allocation argument structure */
|
|
|
|
int flags) /* XFS_ALLOC_FLAG_... */
|
|
|
|
{
|
|
|
|
xfs_buf_t *agbp; /* agf buffer pointer */
|
|
|
|
xfs_agf_t *agf; /* a.g. freespace structure pointer */
|
|
|
|
xfs_buf_t *agflbp;/* agfl buffer pointer */
|
|
|
|
xfs_agblock_t bno; /* freelist block */
|
|
|
|
xfs_extlen_t delta; /* new blocks needed in freelist */
|
|
|
|
int error; /* error result code */
|
|
|
|
xfs_extlen_t longest;/* longest extent in allocation group */
|
|
|
|
xfs_mount_t *mp; /* file system mount point structure */
|
|
|
|
xfs_extlen_t need; /* total blocks needed in freelist */
|
|
|
|
xfs_perag_t *pag; /* per-ag information structure */
|
|
|
|
xfs_alloc_arg_t targs; /* local allocation arguments */
|
|
|
|
xfs_trans_t *tp; /* transaction pointer */
|
|
|
|
|
|
|
|
mp = args->mp;
|
|
|
|
|
|
|
|
pag = args->pag;
|
|
|
|
tp = args->tp;
|
|
|
|
if (!pag->pagf_init) {
|
|
|
|
if ((error = xfs_alloc_read_agf(mp, tp, args->agno, flags,
|
|
|
|
&agbp)))
|
|
|
|
return error;
|
|
|
|
if (!pag->pagf_init) {
|
2006-08-10 06:40:41 +02:00
|
|
|
ASSERT(flags & XFS_ALLOC_FLAG_TRYLOCK);
|
|
|
|
ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
|
2005-04-17 00:20:36 +02:00
|
|
|
args->agbp = NULL;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
} else
|
|
|
|
agbp = NULL;
|
|
|
|
|
2006-08-10 06:40:41 +02:00
|
|
|
/*
|
|
|
|
* If this is a metadata preferred pag and we are user data
|
2005-04-17 00:20:36 +02:00
|
|
|
* then try somewhere else if we are not being asked to
|
|
|
|
* try harder at this point
|
|
|
|
*/
|
2006-08-10 06:40:41 +02:00
|
|
|
if (pag->pagf_metadata && args->userdata &&
|
|
|
|
(flags & XFS_ALLOC_FLAG_TRYLOCK)) {
|
|
|
|
ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
|
2005-04-17 00:20:36 +02:00
|
|
|
args->agbp = NULL;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2006-08-10 06:40:41 +02:00
|
|
|
if (!(flags & XFS_ALLOC_FLAG_FREEING)) {
|
|
|
|
/*
|
|
|
|
* If it looks like there isn't a long enough extent, or enough
|
|
|
|
* total blocks, reject it.
|
|
|
|
*/
|
2009-03-16 08:29:46 +01:00
|
|
|
need = XFS_MIN_FREELIST_PAG(pag, mp);
|
|
|
|
longest = xfs_alloc_longest_free_extent(mp, pag);
|
2006-08-10 06:40:41 +02:00
|
|
|
if ((args->minlen + args->alignment + args->minalignslop - 1) >
|
|
|
|
longest ||
|
|
|
|
((int)(pag->pagf_freeblks + pag->pagf_flcount -
|
|
|
|
need - args->total) < (int)args->minleft)) {
|
|
|
|
if (agbp)
|
|
|
|
xfs_trans_brelse(tp, agbp);
|
|
|
|
args->agbp = NULL;
|
|
|
|
return 0;
|
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2006-08-10 06:40:41 +02:00
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Get the a.g. freespace buffer.
|
|
|
|
* Can fail if we're not blocking on locks, and it's held.
|
|
|
|
*/
|
|
|
|
if (agbp == NULL) {
|
|
|
|
if ((error = xfs_alloc_read_agf(mp, tp, args->agno, flags,
|
|
|
|
&agbp)))
|
|
|
|
return error;
|
|
|
|
if (agbp == NULL) {
|
2006-08-10 06:40:41 +02:00
|
|
|
ASSERT(flags & XFS_ALLOC_FLAG_TRYLOCK);
|
|
|
|
ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
|
2005-04-17 00:20:36 +02:00
|
|
|
args->agbp = NULL;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Figure out how many blocks we should have in the freelist.
|
|
|
|
*/
|
|
|
|
agf = XFS_BUF_TO_AGF(agbp);
|
|
|
|
need = XFS_MIN_FREELIST(agf, mp);
|
|
|
|
/*
|
|
|
|
* If there isn't enough total or single-extent, reject it.
|
|
|
|
*/
|
2006-08-10 06:40:41 +02:00
|
|
|
if (!(flags & XFS_ALLOC_FLAG_FREEING)) {
|
|
|
|
delta = need > be32_to_cpu(agf->agf_flcount) ?
|
|
|
|
(need - be32_to_cpu(agf->agf_flcount)) : 0;
|
|
|
|
longest = be32_to_cpu(agf->agf_longest);
|
|
|
|
longest = (longest > delta) ? (longest - delta) :
|
|
|
|
(be32_to_cpu(agf->agf_flcount) > 0 || longest > 0);
|
|
|
|
if ((args->minlen + args->alignment + args->minalignslop - 1) >
|
|
|
|
longest ||
|
|
|
|
((int)(be32_to_cpu(agf->agf_freeblks) +
|
|
|
|
be32_to_cpu(agf->agf_flcount) - need - args->total) <
|
|
|
|
(int)args->minleft)) {
|
|
|
|
xfs_trans_brelse(tp, agbp);
|
|
|
|
args->agbp = NULL;
|
|
|
|
return 0;
|
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Make the freelist shorter if it's too long.
|
|
|
|
*/
|
2005-11-02 05:11:25 +01:00
|
|
|
while (be32_to_cpu(agf->agf_flcount) > need) {
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_buf_t *bp;
|
|
|
|
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
error = xfs_alloc_get_freelist(tp, agbp, &bno, 0);
|
|
|
|
if (error)
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
if ((error = xfs_free_ag_extent(tp, agbp, args->agno, bno, 1, 1)))
|
|
|
|
return error;
|
|
|
|
bp = xfs_btree_get_bufs(mp, tp, args->agno, bno, 0);
|
|
|
|
xfs_trans_binval(tp, bp);
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Initialize the args structure.
|
|
|
|
*/
|
|
|
|
targs.tp = tp;
|
|
|
|
targs.mp = mp;
|
|
|
|
targs.agbp = agbp;
|
|
|
|
targs.agno = args->agno;
|
|
|
|
targs.mod = targs.minleft = targs.wasdel = targs.userdata =
|
|
|
|
targs.minalignslop = 0;
|
|
|
|
targs.alignment = targs.minlen = targs.prod = targs.isfl = 1;
|
|
|
|
targs.type = XFS_ALLOCTYPE_THIS_AG;
|
|
|
|
targs.pag = pag;
|
|
|
|
if ((error = xfs_alloc_read_agfl(mp, tp, targs.agno, &agflbp)))
|
|
|
|
return error;
|
|
|
|
/*
|
|
|
|
* Make the freelist longer if it's too short.
|
|
|
|
*/
|
2005-11-02 05:11:25 +01:00
|
|
|
while (be32_to_cpu(agf->agf_flcount) < need) {
|
2005-04-17 00:20:36 +02:00
|
|
|
targs.agbno = 0;
|
2005-11-02 05:11:25 +01:00
|
|
|
targs.maxlen = need - be32_to_cpu(agf->agf_flcount);
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Allocate as many blocks as possible at once.
|
|
|
|
*/
|
2006-05-08 11:51:58 +02:00
|
|
|
if ((error = xfs_alloc_ag_vextent(&targs))) {
|
|
|
|
xfs_trans_brelse(tp, agflbp);
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
2006-05-08 11:51:58 +02:00
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Stop if we run out. Won't happen if callers are obeying
|
|
|
|
* the restrictions correctly. Can happen for free calls
|
|
|
|
* on a completely full ag.
|
|
|
|
*/
|
2006-06-09 06:55:18 +02:00
|
|
|
if (targs.agbno == NULLAGBLOCK) {
|
2006-08-10 06:40:41 +02:00
|
|
|
if (flags & XFS_ALLOC_FLAG_FREEING)
|
|
|
|
break;
|
|
|
|
xfs_trans_brelse(tp, agflbp);
|
|
|
|
args->agbp = NULL;
|
|
|
|
return 0;
|
2006-06-09 06:55:18 +02:00
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Put each allocated block on the list.
|
|
|
|
*/
|
|
|
|
for (bno = targs.agbno; bno < targs.agbno + targs.len; bno++) {
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
error = xfs_alloc_put_freelist(tp, agbp,
|
|
|
|
agflbp, bno, 0);
|
|
|
|
if (error)
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
}
|
|
|
|
}
|
2006-05-08 11:51:58 +02:00
|
|
|
xfs_trans_brelse(tp, agflbp);
|
2005-04-17 00:20:36 +02:00
|
|
|
args->agbp = agbp;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Get a block from the freelist.
|
|
|
|
* Returns with the buffer for the block gotten.
|
|
|
|
*/
|
|
|
|
int /* error */
|
|
|
|
xfs_alloc_get_freelist(
|
|
|
|
xfs_trans_t *tp, /* transaction pointer */
|
|
|
|
xfs_buf_t *agbp, /* buffer containing the agf structure */
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
xfs_agblock_t *bnop, /* block address retrieved from freelist */
|
|
|
|
int btreeblk) /* destination is a AGF btree */
|
2005-04-17 00:20:36 +02:00
|
|
|
{
|
|
|
|
xfs_agf_t *agf; /* a.g. freespace structure */
|
|
|
|
xfs_agfl_t *agfl; /* a.g. freelist structure */
|
|
|
|
xfs_buf_t *agflbp;/* buffer for a.g. freelist structure */
|
|
|
|
xfs_agblock_t bno; /* block number returned */
|
|
|
|
int error;
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
int logflags;
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_mount_t *mp; /* mount structure */
|
|
|
|
xfs_perag_t *pag; /* per allocation group data */
|
|
|
|
|
|
|
|
agf = XFS_BUF_TO_AGF(agbp);
|
|
|
|
/*
|
|
|
|
* Freelist is empty, give up.
|
|
|
|
*/
|
|
|
|
if (!agf->agf_flcount) {
|
|
|
|
*bnop = NULLAGBLOCK;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Read the array of free blocks.
|
|
|
|
*/
|
|
|
|
mp = tp->t_mountp;
|
|
|
|
if ((error = xfs_alloc_read_agfl(mp, tp,
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(agf->agf_seqno), &agflbp)))
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
agfl = XFS_BUF_TO_AGFL(agflbp);
|
|
|
|
/*
|
|
|
|
* Get the block number and update the data structures.
|
|
|
|
*/
|
2006-09-28 02:56:51 +02:00
|
|
|
bno = be32_to_cpu(agfl->agfl_bno[be32_to_cpu(agf->agf_flfirst)]);
|
2008-02-14 00:03:29 +01:00
|
|
|
be32_add_cpu(&agf->agf_flfirst, 1);
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_trans_brelse(tp, agflbp);
|
2005-11-02 05:11:25 +01:00
|
|
|
if (be32_to_cpu(agf->agf_flfirst) == XFS_AGFL_SIZE(mp))
|
2005-04-17 00:20:36 +02:00
|
|
|
agf->agf_flfirst = 0;
|
2010-01-11 12:47:41 +01:00
|
|
|
|
|
|
|
pag = xfs_perag_get(mp, be32_to_cpu(agf->agf_seqno));
|
2008-02-14 00:03:29 +01:00
|
|
|
be32_add_cpu(&agf->agf_flcount, -1);
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_trans_agflist_delta(tp, -1);
|
|
|
|
pag->pagf_flcount--;
|
2010-01-11 12:47:41 +01:00
|
|
|
xfs_perag_put(pag);
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
|
|
|
|
logflags = XFS_AGF_FLFIRST | XFS_AGF_FLCOUNT;
|
|
|
|
if (btreeblk) {
|
2008-02-14 00:03:29 +01:00
|
|
|
be32_add_cpu(&agf->agf_btreeblks, 1);
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
pag->pagf_btreeblks++;
|
|
|
|
logflags |= XFS_AGF_BTREEBLKS;
|
|
|
|
}
|
|
|
|
|
|
|
|
xfs_alloc_log_agf(tp, agbp, logflags);
|
2005-04-17 00:20:36 +02:00
|
|
|
*bnop = bno;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Log the given fields from the agf structure.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
xfs_alloc_log_agf(
|
|
|
|
xfs_trans_t *tp, /* transaction pointer */
|
|
|
|
xfs_buf_t *bp, /* buffer for a.g. freelist header */
|
|
|
|
int fields) /* mask of fields to be logged (XFS_AGF_...) */
|
|
|
|
{
|
|
|
|
int first; /* first byte offset */
|
|
|
|
int last; /* last byte offset */
|
|
|
|
static const short offsets[] = {
|
|
|
|
offsetof(xfs_agf_t, agf_magicnum),
|
|
|
|
offsetof(xfs_agf_t, agf_versionnum),
|
|
|
|
offsetof(xfs_agf_t, agf_seqno),
|
|
|
|
offsetof(xfs_agf_t, agf_length),
|
|
|
|
offsetof(xfs_agf_t, agf_roots[0]),
|
|
|
|
offsetof(xfs_agf_t, agf_levels[0]),
|
|
|
|
offsetof(xfs_agf_t, agf_flfirst),
|
|
|
|
offsetof(xfs_agf_t, agf_fllast),
|
|
|
|
offsetof(xfs_agf_t, agf_flcount),
|
|
|
|
offsetof(xfs_agf_t, agf_freeblks),
|
|
|
|
offsetof(xfs_agf_t, agf_longest),
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
offsetof(xfs_agf_t, agf_btreeblks),
|
2005-04-17 00:20:36 +02:00
|
|
|
sizeof(xfs_agf_t)
|
|
|
|
};
|
|
|
|
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_agf(tp->t_mountp, XFS_BUF_TO_AGF(bp), fields, _RET_IP_);
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_btree_offsets(fields, offsets, XFS_AGF_NUM_BITS, &first, &last);
|
|
|
|
xfs_trans_log_buf(tp, bp, (uint)first, (uint)last);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Interface for inode allocation to force the pag data to be initialized.
|
|
|
|
*/
|
|
|
|
int /* error */
|
|
|
|
xfs_alloc_pagf_init(
|
|
|
|
xfs_mount_t *mp, /* file system mount structure */
|
|
|
|
xfs_trans_t *tp, /* transaction pointer */
|
|
|
|
xfs_agnumber_t agno, /* allocation group number */
|
|
|
|
int flags) /* XFS_ALLOC_FLAGS_... */
|
|
|
|
{
|
|
|
|
xfs_buf_t *bp;
|
|
|
|
int error;
|
|
|
|
|
|
|
|
if ((error = xfs_alloc_read_agf(mp, tp, agno, flags, &bp)))
|
|
|
|
return error;
|
|
|
|
if (bp)
|
|
|
|
xfs_trans_brelse(tp, bp);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Put the block on the freelist for the allocation group.
|
|
|
|
*/
|
|
|
|
int /* error */
|
|
|
|
xfs_alloc_put_freelist(
|
|
|
|
xfs_trans_t *tp, /* transaction pointer */
|
|
|
|
xfs_buf_t *agbp, /* buffer for a.g. freelist header */
|
|
|
|
xfs_buf_t *agflbp,/* buffer for a.g. free block array */
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
xfs_agblock_t bno, /* block being freed */
|
|
|
|
int btreeblk) /* block came from a AGF btree */
|
2005-04-17 00:20:36 +02:00
|
|
|
{
|
|
|
|
xfs_agf_t *agf; /* a.g. freespace structure */
|
|
|
|
xfs_agfl_t *agfl; /* a.g. free block array */
|
2006-09-28 02:56:51 +02:00
|
|
|
__be32 *blockp;/* pointer to array entry */
|
2005-04-17 00:20:36 +02:00
|
|
|
int error;
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
int logflags;
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_mount_t *mp; /* mount structure */
|
|
|
|
xfs_perag_t *pag; /* per allocation group data */
|
|
|
|
|
|
|
|
agf = XFS_BUF_TO_AGF(agbp);
|
|
|
|
mp = tp->t_mountp;
|
|
|
|
|
|
|
|
if (!agflbp && (error = xfs_alloc_read_agfl(mp, tp,
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(agf->agf_seqno), &agflbp)))
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
agfl = XFS_BUF_TO_AGFL(agflbp);
|
2008-02-14 00:03:29 +01:00
|
|
|
be32_add_cpu(&agf->agf_fllast, 1);
|
2005-11-02 05:11:25 +01:00
|
|
|
if (be32_to_cpu(agf->agf_fllast) == XFS_AGFL_SIZE(mp))
|
2005-04-17 00:20:36 +02:00
|
|
|
agf->agf_fllast = 0;
|
2010-01-11 12:47:41 +01:00
|
|
|
|
|
|
|
pag = xfs_perag_get(mp, be32_to_cpu(agf->agf_seqno));
|
2008-02-14 00:03:29 +01:00
|
|
|
be32_add_cpu(&agf->agf_flcount, 1);
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_trans_agflist_delta(tp, 1);
|
|
|
|
pag->pagf_flcount++;
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
|
|
|
|
logflags = XFS_AGF_FLLAST | XFS_AGF_FLCOUNT;
|
|
|
|
if (btreeblk) {
|
2008-02-14 00:03:29 +01:00
|
|
|
be32_add_cpu(&agf->agf_btreeblks, -1);
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
pag->pagf_btreeblks--;
|
|
|
|
logflags |= XFS_AGF_BTREEBLKS;
|
|
|
|
}
|
2010-01-11 12:47:41 +01:00
|
|
|
xfs_perag_put(pag);
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
|
|
|
|
xfs_alloc_log_agf(tp, agbp, logflags);
|
|
|
|
|
2005-11-02 05:11:25 +01:00
|
|
|
ASSERT(be32_to_cpu(agf->agf_flcount) <= XFS_AGFL_SIZE(mp));
|
|
|
|
blockp = &agfl->agfl_bno[be32_to_cpu(agf->agf_fllast)];
|
2006-09-28 02:56:51 +02:00
|
|
|
*blockp = cpu_to_be32(bno);
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
xfs_alloc_log_agf(tp, agbp, logflags);
|
2005-04-17 00:20:36 +02:00
|
|
|
xfs_trans_log_buf(tp, agflbp,
|
|
|
|
(int)((xfs_caddr_t)blockp - (xfs_caddr_t)agfl),
|
|
|
|
(int)((xfs_caddr_t)blockp - (xfs_caddr_t)agfl +
|
|
|
|
sizeof(xfs_agblock_t) - 1));
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read in the allocation group header (free/alloc section).
|
|
|
|
*/
|
|
|
|
int /* error */
|
2008-11-28 04:23:38 +01:00
|
|
|
xfs_read_agf(
|
|
|
|
struct xfs_mount *mp, /* mount point structure */
|
|
|
|
struct xfs_trans *tp, /* transaction pointer */
|
|
|
|
xfs_agnumber_t agno, /* allocation group number */
|
|
|
|
int flags, /* XFS_BUF_ */
|
|
|
|
struct xfs_buf **bpp) /* buffer for the ag freelist header */
|
2005-04-17 00:20:36 +02:00
|
|
|
{
|
2008-11-28 04:23:38 +01:00
|
|
|
struct xfs_agf *agf; /* ag freelist header */
|
2005-04-17 00:20:36 +02:00
|
|
|
int agf_ok; /* set if agf is consistent */
|
|
|
|
int error;
|
|
|
|
|
|
|
|
ASSERT(agno != NULLAGNUMBER);
|
|
|
|
error = xfs_trans_read_buf(
|
|
|
|
mp, tp, mp->m_ddev_targp,
|
|
|
|
XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp)),
|
2008-11-28 04:23:38 +01:00
|
|
|
XFS_FSS_TO_BB(mp, 1), flags, bpp);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (error)
|
|
|
|
return error;
|
2008-11-28 04:23:38 +01:00
|
|
|
if (!*bpp)
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
2008-11-28 04:23:38 +01:00
|
|
|
|
|
|
|
ASSERT(!XFS_BUF_GETERROR(*bpp));
|
|
|
|
agf = XFS_BUF_TO_AGF(*bpp);
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Validate the magic number of the agf block.
|
|
|
|
*/
|
|
|
|
agf_ok =
|
2011-07-08 14:36:05 +02:00
|
|
|
agf->agf_magicnum == cpu_to_be32(XFS_AGF_MAGIC) &&
|
2005-11-02 05:11:25 +01:00
|
|
|
XFS_AGF_GOOD_VERSION(be32_to_cpu(agf->agf_versionnum)) &&
|
|
|
|
be32_to_cpu(agf->agf_freeblks) <= be32_to_cpu(agf->agf_length) &&
|
|
|
|
be32_to_cpu(agf->agf_flfirst) < XFS_AGFL_SIZE(mp) &&
|
|
|
|
be32_to_cpu(agf->agf_fllast) < XFS_AGFL_SIZE(mp) &&
|
2008-11-28 04:23:38 +01:00
|
|
|
be32_to_cpu(agf->agf_flcount) <= XFS_AGFL_SIZE(mp) &&
|
|
|
|
be32_to_cpu(agf->agf_seqno) == agno;
|
2008-10-30 07:05:49 +01:00
|
|
|
if (xfs_sb_version_haslazysbcount(&mp->m_sb))
|
|
|
|
agf_ok = agf_ok && be32_to_cpu(agf->agf_btreeblks) <=
|
|
|
|
be32_to_cpu(agf->agf_length);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (unlikely(XFS_TEST_ERROR(!agf_ok, mp, XFS_ERRTAG_ALLOC_READ_AGF,
|
|
|
|
XFS_RANDOM_ALLOC_READ_AGF))) {
|
|
|
|
XFS_CORRUPTION_ERROR("xfs_alloc_read_agf",
|
|
|
|
XFS_ERRLEVEL_LOW, mp, agf);
|
2008-11-28 04:23:38 +01:00
|
|
|
xfs_trans_brelse(tp, *bpp);
|
2005-04-17 00:20:36 +02:00
|
|
|
return XFS_ERROR(EFSCORRUPTED);
|
|
|
|
}
|
2008-11-28 04:23:38 +01:00
|
|
|
XFS_BUF_SET_VTYPE_REF(*bpp, B_FS_AGF, XFS_AGF_REF);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read in the allocation group header (free/alloc section).
|
|
|
|
*/
|
|
|
|
int /* error */
|
|
|
|
xfs_alloc_read_agf(
|
|
|
|
struct xfs_mount *mp, /* mount point structure */
|
|
|
|
struct xfs_trans *tp, /* transaction pointer */
|
|
|
|
xfs_agnumber_t agno, /* allocation group number */
|
|
|
|
int flags, /* XFS_ALLOC_FLAG_... */
|
|
|
|
struct xfs_buf **bpp) /* buffer for the ag freelist header */
|
|
|
|
{
|
|
|
|
struct xfs_agf *agf; /* ag freelist header */
|
|
|
|
struct xfs_perag *pag; /* per allocation group data */
|
|
|
|
int error;
|
|
|
|
|
|
|
|
ASSERT(agno != NULLAGNUMBER);
|
|
|
|
|
|
|
|
error = xfs_read_agf(mp, tp, agno,
|
2010-01-19 10:56:44 +01:00
|
|
|
(flags & XFS_ALLOC_FLAG_TRYLOCK) ? XBF_TRYLOCK : 0,
|
2008-11-28 04:23:38 +01:00
|
|
|
bpp);
|
|
|
|
if (error)
|
|
|
|
return error;
|
|
|
|
if (!*bpp)
|
|
|
|
return 0;
|
|
|
|
ASSERT(!XFS_BUF_GETERROR(*bpp));
|
|
|
|
|
|
|
|
agf = XFS_BUF_TO_AGF(*bpp);
|
2010-01-11 12:47:41 +01:00
|
|
|
pag = xfs_perag_get(mp, agno);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (!pag->pagf_init) {
|
2005-11-02 05:11:25 +01:00
|
|
|
pag->pagf_freeblks = be32_to_cpu(agf->agf_freeblks);
|
[XFS] Lazy Superblock Counters
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
2007-05-24 07:26:31 +02:00
|
|
|
pag->pagf_btreeblks = be32_to_cpu(agf->agf_btreeblks);
|
2005-11-02 05:11:25 +01:00
|
|
|
pag->pagf_flcount = be32_to_cpu(agf->agf_flcount);
|
|
|
|
pag->pagf_longest = be32_to_cpu(agf->agf_longest);
|
2005-04-17 00:20:36 +02:00
|
|
|
pag->pagf_levels[XFS_BTNUM_BNOi] =
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(agf->agf_levels[XFS_BTNUM_BNOi]);
|
2005-04-17 00:20:36 +02:00
|
|
|
pag->pagf_levels[XFS_BTNUM_CNTi] =
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(agf->agf_levels[XFS_BTNUM_CNTi]);
|
2007-10-11 09:43:56 +02:00
|
|
|
spin_lock_init(&pag->pagb_lock);
|
2010-01-11 12:47:49 +01:00
|
|
|
pag->pagb_count = 0;
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
pag->pagb_tree = RB_ROOT;
|
2005-04-17 00:20:36 +02:00
|
|
|
pag->pagf_init = 1;
|
|
|
|
}
|
|
|
|
#ifdef DEBUG
|
|
|
|
else if (!XFS_FORCED_SHUTDOWN(mp)) {
|
2005-11-02 05:11:25 +01:00
|
|
|
ASSERT(pag->pagf_freeblks == be32_to_cpu(agf->agf_freeblks));
|
2008-10-30 07:05:49 +01:00
|
|
|
ASSERT(pag->pagf_btreeblks == be32_to_cpu(agf->agf_btreeblks));
|
2005-11-02 05:11:25 +01:00
|
|
|
ASSERT(pag->pagf_flcount == be32_to_cpu(agf->agf_flcount));
|
|
|
|
ASSERT(pag->pagf_longest == be32_to_cpu(agf->agf_longest));
|
2005-04-17 00:20:36 +02:00
|
|
|
ASSERT(pag->pagf_levels[XFS_BTNUM_BNOi] ==
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(agf->agf_levels[XFS_BTNUM_BNOi]));
|
2005-04-17 00:20:36 +02:00
|
|
|
ASSERT(pag->pagf_levels[XFS_BTNUM_CNTi] ==
|
2005-11-02 05:11:25 +01:00
|
|
|
be32_to_cpu(agf->agf_levels[XFS_BTNUM_CNTi]));
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
#endif
|
2010-01-11 12:47:41 +01:00
|
|
|
xfs_perag_put(pag);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allocate an extent (variable-size).
|
|
|
|
* Depending on the allocation type, we either look in a single allocation
|
|
|
|
* group or loop over the allocation groups to find the result.
|
|
|
|
*/
|
|
|
|
int /* error */
|
|
|
|
xfs_alloc_vextent(
|
|
|
|
xfs_alloc_arg_t *args) /* allocation argument structure */
|
|
|
|
{
|
|
|
|
xfs_agblock_t agsize; /* allocation group size */
|
|
|
|
int error;
|
|
|
|
int flags; /* XFS_ALLOC_FLAG_... locking flags */
|
|
|
|
xfs_extlen_t minleft;/* minimum left value, temp copy */
|
|
|
|
xfs_mount_t *mp; /* mount structure pointer */
|
|
|
|
xfs_agnumber_t sagno; /* starting allocation group number */
|
|
|
|
xfs_alloctype_t type; /* input allocation type */
|
|
|
|
int bump_rotor = 0;
|
|
|
|
int no_min = 0;
|
|
|
|
xfs_agnumber_t rotorstep = xfs_rotorstep; /* inode32 agf stepper */
|
|
|
|
|
|
|
|
mp = args->mp;
|
|
|
|
type = args->otype = args->type;
|
|
|
|
args->agbno = NULLAGBLOCK;
|
|
|
|
/*
|
|
|
|
* Just fix this up, for the case where the last a.g. is shorter
|
|
|
|
* (or there's only one a.g.) and the caller couldn't easily figure
|
|
|
|
* that out (xfs_bmap_alloc).
|
|
|
|
*/
|
|
|
|
agsize = mp->m_sb.sb_agblocks;
|
|
|
|
if (args->maxlen > agsize)
|
|
|
|
args->maxlen = agsize;
|
|
|
|
if (args->alignment == 0)
|
|
|
|
args->alignment = 1;
|
|
|
|
ASSERT(XFS_FSB_TO_AGNO(mp, args->fsbno) < mp->m_sb.sb_agcount);
|
|
|
|
ASSERT(XFS_FSB_TO_AGBNO(mp, args->fsbno) < agsize);
|
|
|
|
ASSERT(args->minlen <= args->maxlen);
|
|
|
|
ASSERT(args->minlen <= agsize);
|
|
|
|
ASSERT(args->mod < args->prod);
|
|
|
|
if (XFS_FSB_TO_AGNO(mp, args->fsbno) >= mp->m_sb.sb_agcount ||
|
|
|
|
XFS_FSB_TO_AGBNO(mp, args->fsbno) >= agsize ||
|
|
|
|
args->minlen > args->maxlen || args->minlen > agsize ||
|
|
|
|
args->mod >= args->prod) {
|
|
|
|
args->fsbno = NULLFSBLOCK;
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_vextent_badargs(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
minleft = args->minleft;
|
|
|
|
|
|
|
|
switch (type) {
|
|
|
|
case XFS_ALLOCTYPE_THIS_AG:
|
|
|
|
case XFS_ALLOCTYPE_NEAR_BNO:
|
|
|
|
case XFS_ALLOCTYPE_THIS_BNO:
|
|
|
|
/*
|
|
|
|
* These three force us into a single a.g.
|
|
|
|
*/
|
|
|
|
args->agno = XFS_FSB_TO_AGNO(mp, args->fsbno);
|
2010-01-11 12:47:41 +01:00
|
|
|
args->pag = xfs_perag_get(mp, args->agno);
|
2005-04-17 00:20:36 +02:00
|
|
|
args->minleft = 0;
|
|
|
|
error = xfs_alloc_fix_freelist(args, 0);
|
|
|
|
args->minleft = minleft;
|
|
|
|
if (error) {
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_vextent_nofix(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
}
|
|
|
|
if (!args->agbp) {
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_vextent_noagbp(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
args->agbno = XFS_FSB_TO_AGBNO(mp, args->fsbno);
|
|
|
|
if ((error = xfs_alloc_ag_vextent(args)))
|
|
|
|
goto error0;
|
|
|
|
break;
|
|
|
|
case XFS_ALLOCTYPE_START_BNO:
|
|
|
|
/*
|
|
|
|
* Try near allocation first, then anywhere-in-ag after
|
|
|
|
* the first a.g. fails.
|
|
|
|
*/
|
|
|
|
if ((args->userdata == XFS_ALLOC_INITIAL_USER_DATA) &&
|
|
|
|
(mp->m_flags & XFS_MOUNT_32BITINODES)) {
|
|
|
|
args->fsbno = XFS_AGB_TO_FSB(mp,
|
|
|
|
((mp->m_agfrotor / rotorstep) %
|
|
|
|
mp->m_sb.sb_agcount), 0);
|
|
|
|
bump_rotor = 1;
|
|
|
|
}
|
|
|
|
args->agbno = XFS_FSB_TO_AGBNO(mp, args->fsbno);
|
|
|
|
args->type = XFS_ALLOCTYPE_NEAR_BNO;
|
|
|
|
/* FALLTHROUGH */
|
|
|
|
case XFS_ALLOCTYPE_ANY_AG:
|
|
|
|
case XFS_ALLOCTYPE_START_AG:
|
|
|
|
case XFS_ALLOCTYPE_FIRST_AG:
|
|
|
|
/*
|
|
|
|
* Rotate through the allocation groups looking for a winner.
|
|
|
|
*/
|
|
|
|
if (type == XFS_ALLOCTYPE_ANY_AG) {
|
|
|
|
/*
|
|
|
|
* Start with the last place we left off.
|
|
|
|
*/
|
|
|
|
args->agno = sagno = (mp->m_agfrotor / rotorstep) %
|
|
|
|
mp->m_sb.sb_agcount;
|
|
|
|
args->type = XFS_ALLOCTYPE_THIS_AG;
|
|
|
|
flags = XFS_ALLOC_FLAG_TRYLOCK;
|
|
|
|
} else if (type == XFS_ALLOCTYPE_FIRST_AG) {
|
|
|
|
/*
|
|
|
|
* Start with allocation group given by bno.
|
|
|
|
*/
|
|
|
|
args->agno = XFS_FSB_TO_AGNO(mp, args->fsbno);
|
|
|
|
args->type = XFS_ALLOCTYPE_THIS_AG;
|
|
|
|
sagno = 0;
|
|
|
|
flags = 0;
|
|
|
|
} else {
|
|
|
|
if (type == XFS_ALLOCTYPE_START_AG)
|
|
|
|
args->type = XFS_ALLOCTYPE_THIS_AG;
|
|
|
|
/*
|
|
|
|
* Start with the given allocation group.
|
|
|
|
*/
|
|
|
|
args->agno = sagno = XFS_FSB_TO_AGNO(mp, args->fsbno);
|
|
|
|
flags = XFS_ALLOC_FLAG_TRYLOCK;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Loop over allocation groups twice; first time with
|
|
|
|
* trylock set, second time without.
|
|
|
|
*/
|
|
|
|
for (;;) {
|
2010-01-11 12:47:41 +01:00
|
|
|
args->pag = xfs_perag_get(mp, args->agno);
|
2005-04-17 00:20:36 +02:00
|
|
|
if (no_min) args->minleft = 0;
|
|
|
|
error = xfs_alloc_fix_freelist(args, flags);
|
|
|
|
args->minleft = minleft;
|
|
|
|
if (error) {
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_vextent_nofix(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* If we get a buffer back then the allocation will fly.
|
|
|
|
*/
|
|
|
|
if (args->agbp) {
|
|
|
|
if ((error = xfs_alloc_ag_vextent(args)))
|
|
|
|
goto error0;
|
|
|
|
break;
|
|
|
|
}
|
2009-12-15 00:14:59 +01:00
|
|
|
|
|
|
|
trace_xfs_alloc_vextent_loopfailed(args);
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Didn't work, figure out the next iteration.
|
|
|
|
*/
|
|
|
|
if (args->agno == sagno &&
|
|
|
|
type == XFS_ALLOCTYPE_START_BNO)
|
|
|
|
args->type = XFS_ALLOCTYPE_THIS_AG;
|
2006-06-09 06:55:18 +02:00
|
|
|
/*
|
|
|
|
* For the first allocation, we can try any AG to get
|
|
|
|
* space. However, if we already have allocated a
|
|
|
|
* block, we don't want to try AGs whose number is below
|
|
|
|
* sagno. Otherwise, we may end up with out-of-order
|
|
|
|
* locking of AGF, which might cause deadlock.
|
|
|
|
*/
|
|
|
|
if (++(args->agno) == mp->m_sb.sb_agcount) {
|
|
|
|
if (args->firstblock != NULLFSBLOCK)
|
|
|
|
args->agno = sagno;
|
|
|
|
else
|
|
|
|
args->agno = 0;
|
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
/*
|
|
|
|
* Reached the starting a.g., must either be done
|
|
|
|
* or switch to non-trylock mode.
|
|
|
|
*/
|
|
|
|
if (args->agno == sagno) {
|
|
|
|
if (no_min == 1) {
|
|
|
|
args->agbno = NULLAGBLOCK;
|
2009-12-15 00:14:59 +01:00
|
|
|
trace_xfs_alloc_vextent_allfailed(args);
|
2005-04-17 00:20:36 +02:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (flags == 0) {
|
|
|
|
no_min = 1;
|
|
|
|
} else {
|
|
|
|
flags = 0;
|
|
|
|
if (type == XFS_ALLOCTYPE_START_BNO) {
|
|
|
|
args->agbno = XFS_FSB_TO_AGBNO(mp,
|
|
|
|
args->fsbno);
|
|
|
|
args->type = XFS_ALLOCTYPE_NEAR_BNO;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2010-01-11 12:47:41 +01:00
|
|
|
xfs_perag_put(args->pag);
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
if (bump_rotor || (type == XFS_ALLOCTYPE_ANY_AG)) {
|
|
|
|
if (args->agno == sagno)
|
|
|
|
mp->m_agfrotor = (mp->m_agfrotor + 1) %
|
|
|
|
(mp->m_sb.sb_agcount * rotorstep);
|
|
|
|
else
|
|
|
|
mp->m_agfrotor = (args->agno * rotorstep + 1) %
|
|
|
|
(mp->m_sb.sb_agcount * rotorstep);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
ASSERT(0);
|
|
|
|
/* NOTREACHED */
|
|
|
|
}
|
|
|
|
if (args->agbno == NULLAGBLOCK)
|
|
|
|
args->fsbno = NULLFSBLOCK;
|
|
|
|
else {
|
|
|
|
args->fsbno = XFS_AGB_TO_FSB(mp, args->agno, args->agbno);
|
|
|
|
#ifdef DEBUG
|
|
|
|
ASSERT(args->len >= args->minlen);
|
|
|
|
ASSERT(args->len <= args->maxlen);
|
|
|
|
ASSERT(args->agbno % args->alignment == 0);
|
|
|
|
XFS_AG_CHECK_DADDR(mp, XFS_FSB_TO_DADDR(mp, args->fsbno),
|
|
|
|
args->len);
|
|
|
|
#endif
|
|
|
|
}
|
2010-01-11 12:47:41 +01:00
|
|
|
xfs_perag_put(args->pag);
|
2005-04-17 00:20:36 +02:00
|
|
|
return 0;
|
|
|
|
error0:
|
2010-01-11 12:47:41 +01:00
|
|
|
xfs_perag_put(args->pag);
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Free an extent.
|
|
|
|
* Just break up the extent address and hand off to xfs_free_ag_extent
|
|
|
|
* after fixing up the freelist.
|
|
|
|
*/
|
|
|
|
int /* error */
|
|
|
|
xfs_free_extent(
|
|
|
|
xfs_trans_t *tp, /* transaction pointer */
|
|
|
|
xfs_fsblock_t bno, /* starting block number of extent */
|
|
|
|
xfs_extlen_t len) /* length of extent */
|
|
|
|
{
|
2006-08-10 06:40:41 +02:00
|
|
|
xfs_alloc_arg_t args;
|
2005-04-17 00:20:36 +02:00
|
|
|
int error;
|
|
|
|
|
|
|
|
ASSERT(len != 0);
|
2006-08-10 06:40:41 +02:00
|
|
|
memset(&args, 0, sizeof(xfs_alloc_arg_t));
|
2005-04-17 00:20:36 +02:00
|
|
|
args.tp = tp;
|
|
|
|
args.mp = tp->t_mountp;
|
2011-04-08 04:45:07 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
* validate that the block number is legal - the enables us to detect
|
|
|
|
* and handle a silent filesystem corruption rather than crashing.
|
|
|
|
*/
|
2005-04-17 00:20:36 +02:00
|
|
|
args.agno = XFS_FSB_TO_AGNO(args.mp, bno);
|
2011-04-08 04:45:07 +02:00
|
|
|
if (args.agno >= args.mp->m_sb.sb_agcount)
|
|
|
|
return EFSCORRUPTED;
|
|
|
|
|
2005-04-17 00:20:36 +02:00
|
|
|
args.agbno = XFS_FSB_TO_AGBNO(args.mp, bno);
|
2011-04-08 04:45:07 +02:00
|
|
|
if (args.agbno >= args.mp->m_sb.sb_agblocks)
|
|
|
|
return EFSCORRUPTED;
|
|
|
|
|
2010-01-11 12:47:41 +01:00
|
|
|
args.pag = xfs_perag_get(args.mp, args.agno);
|
2011-04-08 04:45:07 +02:00
|
|
|
ASSERT(args.pag);
|
|
|
|
|
|
|
|
error = xfs_alloc_fix_freelist(&args, XFS_ALLOC_FLAG_FREEING);
|
|
|
|
if (error)
|
2005-04-17 00:20:36 +02:00
|
|
|
goto error0;
|
2011-04-08 04:45:07 +02:00
|
|
|
|
|
|
|
/* validate the extent size is legal now we have the agf locked */
|
|
|
|
if (args.agbno + len >
|
|
|
|
be32_to_cpu(XFS_BUF_TO_AGF(args.agbp)->agf_length)) {
|
|
|
|
error = EFSCORRUPTED;
|
|
|
|
goto error0;
|
|
|
|
}
|
|
|
|
|
2006-08-10 06:40:41 +02:00
|
|
|
error = xfs_free_ag_extent(tp, args.agbp, args.agno, args.agbno, len, 0);
|
2011-04-24 21:06:14 +02:00
|
|
|
if (!error)
|
2011-05-04 20:55:15 +02:00
|
|
|
xfs_alloc_busy_insert(tp, args.agno, args.agbno, len, 0);
|
2005-04-17 00:20:36 +02:00
|
|
|
error0:
|
2010-01-11 12:47:41 +01:00
|
|
|
xfs_perag_put(args.pag);
|
2005-04-17 00:20:36 +02:00
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
xfs_alloc_busy_insert(
|
|
|
|
struct xfs_trans *tp,
|
|
|
|
xfs_agnumber_t agno,
|
|
|
|
xfs_agblock_t bno,
|
2011-05-04 20:55:15 +02:00
|
|
|
xfs_extlen_t len,
|
|
|
|
unsigned int flags)
|
2005-04-17 00:20:36 +02:00
|
|
|
{
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
struct xfs_busy_extent *new;
|
|
|
|
struct xfs_busy_extent *busyp;
|
2010-01-11 12:47:41 +01:00
|
|
|
struct xfs_perag *pag;
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
struct rb_node **rbp;
|
2011-04-24 21:06:16 +02:00
|
|
|
struct rb_node *parent = NULL;
|
2005-04-17 00:20:36 +02:00
|
|
|
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
new = kmem_zalloc(sizeof(struct xfs_busy_extent), KM_MAYFAIL);
|
|
|
|
if (!new) {
|
|
|
|
/*
|
|
|
|
* No Memory! Since it is now not possible to track the free
|
|
|
|
* block, make this a synchronous transaction to insure that
|
|
|
|
* the block is not reused before this transaction commits.
|
|
|
|
*/
|
2011-04-24 21:06:16 +02:00
|
|
|
trace_xfs_alloc_busy_enomem(tp->t_mountp, agno, bno, len);
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
xfs_trans_set_sync(tp);
|
|
|
|
return;
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
new->agno = agno;
|
|
|
|
new->bno = bno;
|
|
|
|
new->length = len;
|
|
|
|
INIT_LIST_HEAD(&new->list);
|
2011-05-04 20:55:15 +02:00
|
|
|
new->flags = flags;
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
|
|
|
|
/* trace before insert to be able to see failed inserts */
|
2011-04-24 21:06:16 +02:00
|
|
|
trace_xfs_alloc_busy(tp->t_mountp, agno, bno, len);
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
|
|
|
|
pag = xfs_perag_get(tp->t_mountp, new->agno);
|
|
|
|
spin_lock(&pag->pagb_lock);
|
|
|
|
rbp = &pag->pagb_tree.rb_node;
|
2011-04-24 21:06:16 +02:00
|
|
|
while (*rbp) {
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
parent = *rbp;
|
|
|
|
busyp = rb_entry(parent, struct xfs_busy_extent, rb_node);
|
|
|
|
|
|
|
|
if (new->bno < busyp->bno) {
|
|
|
|
rbp = &(*rbp)->rb_left;
|
2011-04-24 21:06:16 +02:00
|
|
|
ASSERT(new->bno + new->length <= busyp->bno);
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
} else if (new->bno > busyp->bno) {
|
|
|
|
rbp = &(*rbp)->rb_right;
|
2011-04-24 21:06:16 +02:00
|
|
|
ASSERT(bno >= busyp->bno + busyp->length);
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
} else {
|
2011-04-24 21:06:16 +02:00
|
|
|
ASSERT(0);
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
}
|
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
rb_link_node(&new->rb_node, parent, rbp);
|
|
|
|
rb_insert_color(&new->rb_node, &pag->pagb_tree);
|
|
|
|
|
|
|
|
list_add(&new->list, &tp->t_busy);
|
2010-01-11 12:47:41 +01:00
|
|
|
spin_unlock(&pag->pagb_lock);
|
|
|
|
xfs_perag_put(pag);
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
/*
|
|
|
|
* Search for a busy extent within the range of the extent we are about to
|
|
|
|
* allocate. You need to be holding the busy extent tree lock when calling
|
|
|
|
* xfs_alloc_busy_search(). This function returns 0 for no overlapping busy
|
|
|
|
* extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
|
|
|
|
* match. This is done so that a non-zero return indicates an overlap that
|
|
|
|
* will require a synchronous transaction, but it can still be
|
|
|
|
* used to distinguish between a partial or exact match.
|
|
|
|
*/
|
2011-01-07 14:02:04 +01:00
|
|
|
int
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
xfs_alloc_busy_search(
|
|
|
|
struct xfs_mount *mp,
|
|
|
|
xfs_agnumber_t agno,
|
|
|
|
xfs_agblock_t bno,
|
|
|
|
xfs_extlen_t len)
|
2005-04-17 00:20:36 +02:00
|
|
|
{
|
2010-01-11 12:47:41 +01:00
|
|
|
struct xfs_perag *pag;
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
struct rb_node *rbp;
|
|
|
|
struct xfs_busy_extent *busyp;
|
|
|
|
int match = 0;
|
2005-04-17 00:20:36 +02:00
|
|
|
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
pag = xfs_perag_get(mp, agno);
|
2010-01-11 12:47:41 +01:00
|
|
|
spin_lock(&pag->pagb_lock);
|
2009-12-15 00:14:59 +01:00
|
|
|
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
rbp = pag->pagb_tree.rb_node;
|
|
|
|
|
|
|
|
/* find closest start bno overlap */
|
|
|
|
while (rbp) {
|
|
|
|
busyp = rb_entry(rbp, struct xfs_busy_extent, rb_node);
|
|
|
|
if (bno < busyp->bno) {
|
|
|
|
/* may overlap, but exact start block is lower */
|
|
|
|
if (bno + len > busyp->bno)
|
|
|
|
match = -1;
|
|
|
|
rbp = rbp->rb_left;
|
|
|
|
} else if (bno > busyp->bno) {
|
|
|
|
/* may overlap, but exact start block is higher */
|
|
|
|
if (bno < busyp->bno + busyp->length)
|
|
|
|
match = -1;
|
|
|
|
rbp = rbp->rb_right;
|
|
|
|
} else {
|
|
|
|
/* bno matches busyp, length determines exact match */
|
|
|
|
match = (busyp->length == len) ? 1 : -1;
|
|
|
|
break;
|
|
|
|
}
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2010-01-11 12:47:41 +01:00
|
|
|
spin_unlock(&pag->pagb_lock);
|
|
|
|
xfs_perag_put(pag);
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
return match;
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
|
|
|
|
2011-04-24 21:06:16 +02:00
|
|
|
/*
|
|
|
|
* The found free extent [fbno, fend] overlaps part or all of the given busy
|
|
|
|
* extent. If the overlap covers the beginning, the end, or all of the busy
|
|
|
|
* extent, the overlapping portion can be made unbusy and used for the
|
|
|
|
* allocation. We can't split a busy extent because we can't modify a
|
|
|
|
* transaction/CIL context busy list, but we can update an entries block
|
|
|
|
* number or length.
|
|
|
|
*
|
|
|
|
* Returns true if the extent can safely be reused, or false if the search
|
|
|
|
* needs to be restarted.
|
|
|
|
*/
|
|
|
|
STATIC bool
|
|
|
|
xfs_alloc_busy_update_extent(
|
|
|
|
struct xfs_mount *mp,
|
|
|
|
struct xfs_perag *pag,
|
|
|
|
struct xfs_busy_extent *busyp,
|
|
|
|
xfs_agblock_t fbno,
|
|
|
|
xfs_extlen_t flen,
|
|
|
|
bool userdata)
|
|
|
|
{
|
|
|
|
xfs_agblock_t fend = fbno + flen;
|
|
|
|
xfs_agblock_t bbno = busyp->bno;
|
|
|
|
xfs_agblock_t bend = bbno + busyp->length;
|
|
|
|
|
2011-05-20 15:45:32 +02:00
|
|
|
/*
|
|
|
|
* This extent is currently being discarded. Give the thread
|
|
|
|
* performing the discard a chance to mark the extent unbusy
|
|
|
|
* and retry.
|
|
|
|
*/
|
|
|
|
if (busyp->flags & XFS_ALLOC_BUSY_DISCARDED) {
|
|
|
|
spin_unlock(&pag->pagb_lock);
|
|
|
|
delay(1);
|
|
|
|
spin_lock(&pag->pagb_lock);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2011-04-24 21:06:16 +02:00
|
|
|
/*
|
|
|
|
* If there is a busy extent overlapping a user allocation, we have
|
|
|
|
* no choice but to force the log and retry the search.
|
|
|
|
*
|
|
|
|
* Fortunately this does not happen during normal operation, but
|
|
|
|
* only if the filesystem is very low on space and has to dip into
|
|
|
|
* the AGFL for normal allocations.
|
|
|
|
*/
|
|
|
|
if (userdata)
|
|
|
|
goto out_force_log;
|
|
|
|
|
|
|
|
if (bbno < fbno && bend > fend) {
|
|
|
|
/*
|
|
|
|
* Case 1:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +---------+
|
|
|
|
* fbno fend
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We would have to split the busy extent to be able to track
|
|
|
|
* it correct, which we cannot do because we would have to
|
|
|
|
* modify the list of busy extents attached to the transaction
|
|
|
|
* or CIL context, which is immutable.
|
|
|
|
*
|
|
|
|
* Force out the log to clear the busy extent and retry the
|
|
|
|
* search.
|
|
|
|
*/
|
|
|
|
goto out_force_log;
|
|
|
|
} else if (bbno >= fbno && bend <= fend) {
|
|
|
|
/*
|
|
|
|
* Case 2:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +-----------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 3:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +--------------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 4:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +--------------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 5:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +-----------------------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The busy extent is fully covered by the extent we are
|
|
|
|
* allocating, and can simply be removed from the rbtree.
|
|
|
|
* However we cannot remove it from the immutable list
|
|
|
|
* tracking busy extents in the transaction or CIL context,
|
|
|
|
* so set the length to zero to mark it invalid.
|
|
|
|
*
|
|
|
|
* We also need to restart the busy extent search from the
|
|
|
|
* tree root, because erasing the node can rearrange the
|
|
|
|
* tree topology.
|
|
|
|
*/
|
|
|
|
rb_erase(&busyp->rb_node, &pag->pagb_tree);
|
|
|
|
busyp->length = 0;
|
|
|
|
return false;
|
|
|
|
} else if (fend < bend) {
|
|
|
|
/*
|
|
|
|
* Case 6:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +---------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 7:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
busyp->bno = fend;
|
|
|
|
} else if (bbno < fbno) {
|
|
|
|
/*
|
|
|
|
* Case 8:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +-------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 9:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +----------------------+
|
|
|
|
* fbno fend
|
|
|
|
*/
|
|
|
|
busyp->length = fbno - busyp->bno;
|
|
|
|
} else {
|
|
|
|
ASSERT(0);
|
|
|
|
}
|
|
|
|
|
|
|
|
trace_xfs_alloc_busy_reuse(mp, pag->pag_agno, fbno, flen);
|
|
|
|
return true;
|
|
|
|
|
|
|
|
out_force_log:
|
|
|
|
spin_unlock(&pag->pagb_lock);
|
|
|
|
xfs_log_force(mp, XFS_LOG_SYNC);
|
|
|
|
trace_xfs_alloc_busy_force(mp, pag->pag_agno, fbno, flen);
|
|
|
|
spin_lock(&pag->pagb_lock);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* For a given extent [fbno, flen], make sure we can reuse it safely.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
xfs_alloc_busy_reuse(
|
|
|
|
struct xfs_mount *mp,
|
|
|
|
xfs_agnumber_t agno,
|
|
|
|
xfs_agblock_t fbno,
|
|
|
|
xfs_extlen_t flen,
|
|
|
|
bool userdata)
|
|
|
|
{
|
|
|
|
struct xfs_perag *pag;
|
|
|
|
struct rb_node *rbp;
|
|
|
|
|
|
|
|
ASSERT(flen > 0);
|
|
|
|
|
|
|
|
pag = xfs_perag_get(mp, agno);
|
|
|
|
spin_lock(&pag->pagb_lock);
|
|
|
|
restart:
|
|
|
|
rbp = pag->pagb_tree.rb_node;
|
|
|
|
while (rbp) {
|
|
|
|
struct xfs_busy_extent *busyp =
|
|
|
|
rb_entry(rbp, struct xfs_busy_extent, rb_node);
|
|
|
|
xfs_agblock_t bbno = busyp->bno;
|
|
|
|
xfs_agblock_t bend = bbno + busyp->length;
|
|
|
|
|
|
|
|
if (fbno + flen <= bbno) {
|
|
|
|
rbp = rbp->rb_left;
|
|
|
|
continue;
|
|
|
|
} else if (fbno >= bend) {
|
|
|
|
rbp = rbp->rb_right;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!xfs_alloc_busy_update_extent(mp, pag, busyp, fbno, flen,
|
|
|
|
userdata))
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
spin_unlock(&pag->pagb_lock);
|
|
|
|
xfs_perag_put(pag);
|
|
|
|
}
|
|
|
|
|
2011-04-24 21:06:15 +02:00
|
|
|
/*
|
|
|
|
* For a given extent [fbno, flen], search the busy extent list to find a
|
|
|
|
* subset of the extent that is not busy. If *rlen is smaller than
|
|
|
|
* args->minlen no suitable extent could be found, and the higher level
|
|
|
|
* code needs to force out the log and retry the allocation.
|
|
|
|
*/
|
|
|
|
STATIC void
|
|
|
|
xfs_alloc_busy_trim(
|
|
|
|
struct xfs_alloc_arg *args,
|
|
|
|
xfs_agblock_t bno,
|
|
|
|
xfs_extlen_t len,
|
|
|
|
xfs_agblock_t *rbno,
|
|
|
|
xfs_extlen_t *rlen)
|
|
|
|
{
|
2011-04-24 21:06:16 +02:00
|
|
|
xfs_agblock_t fbno;
|
|
|
|
xfs_extlen_t flen;
|
2011-04-24 21:06:15 +02:00
|
|
|
struct rb_node *rbp;
|
|
|
|
|
2011-04-24 21:06:16 +02:00
|
|
|
ASSERT(len > 0);
|
2011-04-24 21:06:15 +02:00
|
|
|
|
|
|
|
spin_lock(&args->pag->pagb_lock);
|
2011-04-24 21:06:16 +02:00
|
|
|
restart:
|
|
|
|
fbno = bno;
|
|
|
|
flen = len;
|
2011-04-24 21:06:15 +02:00
|
|
|
rbp = args->pag->pagb_tree.rb_node;
|
|
|
|
while (rbp && flen >= args->minlen) {
|
|
|
|
struct xfs_busy_extent *busyp =
|
|
|
|
rb_entry(rbp, struct xfs_busy_extent, rb_node);
|
|
|
|
xfs_agblock_t fend = fbno + flen;
|
|
|
|
xfs_agblock_t bbno = busyp->bno;
|
|
|
|
xfs_agblock_t bend = bbno + busyp->length;
|
|
|
|
|
|
|
|
if (fend <= bbno) {
|
|
|
|
rbp = rbp->rb_left;
|
|
|
|
continue;
|
|
|
|
} else if (fbno >= bend) {
|
|
|
|
rbp = rbp->rb_right;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2011-04-24 21:06:16 +02:00
|
|
|
/*
|
|
|
|
* If this is a metadata allocation, try to reuse the busy
|
|
|
|
* extent instead of trimming the allocation.
|
|
|
|
*/
|
2011-05-20 15:45:32 +02:00
|
|
|
if (!args->userdata &&
|
|
|
|
!(busyp->flags & XFS_ALLOC_BUSY_DISCARDED)) {
|
2011-04-24 21:06:16 +02:00
|
|
|
if (!xfs_alloc_busy_update_extent(args->mp, args->pag,
|
|
|
|
busyp, fbno, flen,
|
|
|
|
false))
|
|
|
|
goto restart;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2011-04-24 21:06:15 +02:00
|
|
|
if (bbno <= fbno) {
|
|
|
|
/* start overlap */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Case 1:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +---------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 2:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +-------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 3:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +-------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 4:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +-----------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* No unbusy region in extent, return failure.
|
|
|
|
*/
|
|
|
|
if (fend <= bend)
|
|
|
|
goto fail;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Case 5:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +----------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 6:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +--------------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Needs to be trimmed to:
|
|
|
|
* +-------+
|
|
|
|
* fbno fend
|
|
|
|
*/
|
|
|
|
fbno = bend;
|
|
|
|
} else if (bend >= fend) {
|
|
|
|
/* end overlap */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Case 7:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Case 8:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +--------------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Needs to be trimmed to:
|
|
|
|
* +-------+
|
|
|
|
* fbno fend
|
|
|
|
*/
|
|
|
|
fend = bbno;
|
|
|
|
} else {
|
|
|
|
/* middle overlap */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Case 9:
|
|
|
|
* bbno bend
|
|
|
|
* +BBBBBBBBBBBBBBBBB+
|
|
|
|
* +-----------------------------------+
|
|
|
|
* fbno fend
|
|
|
|
*
|
|
|
|
* Can be trimmed to:
|
|
|
|
* +-------+ OR +-------+
|
|
|
|
* fbno fend fbno fend
|
|
|
|
*
|
|
|
|
* Backward allocation leads to significant
|
|
|
|
* fragmentation of directories, which degrades
|
|
|
|
* directory performance, therefore we always want to
|
|
|
|
* choose the option that produces forward allocation
|
|
|
|
* patterns.
|
|
|
|
* Preferring the lower bno extent will make the next
|
|
|
|
* request use "fend" as the start of the next
|
|
|
|
* allocation; if the segment is no longer busy at
|
|
|
|
* that point, we'll get a contiguous allocation, but
|
|
|
|
* even if it is still busy, we will get a forward
|
|
|
|
* allocation.
|
|
|
|
* We try to avoid choosing the segment at "bend",
|
|
|
|
* because that can lead to the next allocation
|
|
|
|
* taking the segment at "fbno", which would be a
|
|
|
|
* backward allocation. We only use the segment at
|
|
|
|
* "fbno" if it is much larger than the current
|
|
|
|
* requested size, because in that case there's a
|
|
|
|
* good chance subsequent allocations will be
|
|
|
|
* contiguous.
|
|
|
|
*/
|
|
|
|
if (bbno - fbno >= args->maxlen) {
|
|
|
|
/* left candidate fits perfect */
|
|
|
|
fend = bbno;
|
|
|
|
} else if (fend - bend >= args->maxlen * 4) {
|
|
|
|
/* right candidate has enough free space */
|
|
|
|
fbno = bend;
|
|
|
|
} else if (bbno - fbno >= args->minlen) {
|
|
|
|
/* left candidate fits minimum requirement */
|
|
|
|
fend = bbno;
|
|
|
|
} else {
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
flen = fend - fbno;
|
|
|
|
}
|
|
|
|
spin_unlock(&args->pag->pagb_lock);
|
|
|
|
|
|
|
|
if (fbno != bno || flen != len) {
|
|
|
|
trace_xfs_alloc_busy_trim(args->mp, args->agno, bno, len,
|
|
|
|
fbno, flen);
|
|
|
|
}
|
|
|
|
*rbno = fbno;
|
|
|
|
*rlen = flen;
|
|
|
|
return;
|
|
|
|
fail:
|
|
|
|
/*
|
|
|
|
* Return a zero extent length as failure indications. All callers
|
|
|
|
* re-check if the trimmed extent satisfies the minlen requirement.
|
|
|
|
*/
|
|
|
|
spin_unlock(&args->pag->pagb_lock);
|
|
|
|
trace_xfs_alloc_busy_trim(args->mp, args->agno, bno, len, fbno, 0);
|
|
|
|
*rbno = fbno;
|
|
|
|
*rlen = 0;
|
|
|
|
}
|
|
|
|
|
2011-04-24 21:06:17 +02:00
|
|
|
static void
|
|
|
|
xfs_alloc_busy_clear_one(
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
struct xfs_mount *mp,
|
2011-04-24 21:06:17 +02:00
|
|
|
struct xfs_perag *pag,
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
struct xfs_busy_extent *busyp)
|
2005-04-17 00:20:36 +02:00
|
|
|
{
|
2011-04-24 21:06:16 +02:00
|
|
|
if (busyp->length) {
|
|
|
|
trace_xfs_alloc_busy_clear(mp, busyp->agno, busyp->bno,
|
|
|
|
busyp->length);
|
|
|
|
rb_erase(&busyp->rb_node, &pag->pagb_tree);
|
|
|
|
}
|
2009-12-15 00:14:59 +01:00
|
|
|
|
2011-04-24 21:06:17 +02:00
|
|
|
list_del_init(&busyp->list);
|
xfs: Improve scalability of busy extent tracking
When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.
Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.
To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.
So, use a per-AG rbtree indexed by block number to track busy
extents. This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.
However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.
Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.
By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.
The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.
Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.
Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
2010-05-21 04:07:08 +02:00
|
|
|
kmem_free(busyp);
|
2005-04-17 00:20:36 +02:00
|
|
|
}
|
2011-04-24 21:06:17 +02:00
|
|
|
|
2011-05-20 15:45:32 +02:00
|
|
|
/*
|
|
|
|
* Remove all extents on the passed in list from the busy extents tree.
|
|
|
|
* If do_discard is set skip extents that need to be discarded, and mark
|
|
|
|
* these as undergoing a discard operation instead.
|
|
|
|
*/
|
2011-04-24 21:06:17 +02:00
|
|
|
void
|
|
|
|
xfs_alloc_busy_clear(
|
|
|
|
struct xfs_mount *mp,
|
2011-05-20 15:45:32 +02:00
|
|
|
struct list_head *list,
|
|
|
|
bool do_discard)
|
2011-04-24 21:06:17 +02:00
|
|
|
{
|
|
|
|
struct xfs_busy_extent *busyp, *n;
|
|
|
|
struct xfs_perag *pag = NULL;
|
|
|
|
xfs_agnumber_t agno = NULLAGNUMBER;
|
|
|
|
|
|
|
|
list_for_each_entry_safe(busyp, n, list, list) {
|
|
|
|
if (busyp->agno != agno) {
|
|
|
|
if (pag) {
|
|
|
|
spin_unlock(&pag->pagb_lock);
|
|
|
|
xfs_perag_put(pag);
|
|
|
|
}
|
|
|
|
pag = xfs_perag_get(mp, busyp->agno);
|
|
|
|
spin_lock(&pag->pagb_lock);
|
|
|
|
agno = busyp->agno;
|
|
|
|
}
|
|
|
|
|
2011-05-04 20:55:15 +02:00
|
|
|
if (do_discard && busyp->length &&
|
|
|
|
!(busyp->flags & XFS_ALLOC_BUSY_SKIP_DISCARD))
|
2011-05-20 15:45:32 +02:00
|
|
|
busyp->flags = XFS_ALLOC_BUSY_DISCARDED;
|
|
|
|
else
|
|
|
|
xfs_alloc_busy_clear_one(mp, pag, busyp);
|
2011-04-24 21:06:17 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (pag) {
|
|
|
|
spin_unlock(&pag->pagb_lock);
|
|
|
|
xfs_perag_put(pag);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Callback for list_sort to sort busy extents by the AG they reside in.
|
|
|
|
*/
|
|
|
|
int
|
|
|
|
xfs_busy_extent_ag_cmp(
|
|
|
|
void *priv,
|
|
|
|
struct list_head *a,
|
|
|
|
struct list_head *b)
|
|
|
|
{
|
|
|
|
return container_of(a, struct xfs_busy_extent, list)->agno -
|
|
|
|
container_of(b, struct xfs_busy_extent, list)->agno;
|
|
|
|
}
|