NetBSD/sbin/dump/rcache.c
bouyer 262ce6cb8d Default the read block size (-k default value) to kern.maxphys (usually
64k these days). This gives a noticable performance boost on large filesystems.
2015-08-24 17:34:03 +00:00

475 lines
12 KiB
C

/* $NetBSD: rcache.c,v 1.25 2015/08/24 17:34:03 bouyer Exp $ */
/*-
* Copyright (c) 1999 The NetBSD Foundation, Inc.
* All rights reserved.
*
* This code is derived from software contributed to The NetBSD Foundation
* by Martin J. Laubach <mjl@emsi.priv.at> and
* Manuel Bouyer <Manuel.Bouyer@lip6.fr>.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
* ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
* TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#include <sys/cdefs.h>
#ifndef lint
__RCSID("$NetBSD: rcache.c,v 1.25 2015/08/24 17:34:03 bouyer Exp $");
#endif /* not lint */
#include <sys/types.h>
#include <sys/uio.h>
#include <sys/mman.h>
#include <sys/param.h>
#include <sys/sysctl.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <errno.h>
#include <string.h>
#include "dump.h"
/*-----------------------------------------------------------------------*/
#define MAXCACHEBUFS 512 /* max 512 buffers */
#define MAXMEMPART 6 /* max 15% of the user mem */
/*-----------------------------------------------------------------------*/
union cdesc {
volatile size_t cd_count;
struct {
volatile daddr_t blkstart;
volatile daddr_t blkend; /* start + nblksread */
volatile daddr_t blocksRead;
volatile size_t time;
#ifdef DIAGNOSTICS
volatile pid_t owner;
#endif
} desc;
#define cd_blkstart desc.blkstart
#define cd_blkend desc.blkend
#define cd_blocksRead desc.blocksRead
#define cd_time desc.time
#define cd_owner desc.owner
};
static int findlru(void);
static void *shareBuffer = NULL;
static union cdesc *cheader;
static union cdesc *cdesc;
static char *cdata;
static int cachebufs;
static int nblksread;
#ifdef STATS
static int nreads;
static int nphysread;
static int64_t readsize;
static int64_t physreadsize;
#endif
#define CSIZE (nblksread << dev_bshift) /* cache buf size */
#define CDATA(desc) (cdata + ((desc) - cdesc) * CSIZE)
void
initcache(int cachesize, int readblksize)
{
size_t len;
size_t sharedSize;
if (readblksize == -1) { /* use kern.maxphys */
int kern_maxphys;
int mib[2] = { CTL_KERN, KERN_MAXPHYS };
len = sizeof(kern_maxphys);
if (sysctl(mib, 2, &kern_maxphys, &len, NULL, 0) < 0) {
msg("sysctl(kern.maxphys) failed: %s\n",
strerror(errno));
return;
}
readblksize = kern_maxphys;
}
/* Convert read block size in terms of filesystem block size */
nblksread = howmany(readblksize, ufsib->ufs_bsize);
/* Then, convert it in terms of device block size */
nblksread <<= ufsib->ufs_bshift - dev_bshift;
if (cachesize == -1) { /* Compute from memory available */
uint64_t usermem, cachetmp;
int mib[2] = { CTL_HW, HW_USERMEM64 };
len = sizeof(usermem);
if (sysctl(mib, 2, &usermem, &len, NULL, 0) < 0) {
msg("sysctl(hw.usermem) failed: %s\n",
strerror(errno));
return;
}
cachetmp = (usermem / MAXMEMPART) / CSIZE;
/* for those with TB of RAM */
cachebufs = (cachetmp > INT_MAX) ? INT_MAX : cachetmp;
} else { /* User specified */
cachebufs = cachesize;
}
if (cachebufs) { /* Don't allocate if zero --> no caching */
if (cachebufs > MAXCACHEBUFS)
cachebufs = MAXCACHEBUFS;
sharedSize = sizeof(union cdesc) +
sizeof(union cdesc) * cachebufs +
cachebufs * CSIZE;
#ifdef STATS
fprintf(stderr, "Using %d buffers (%d bytes)\n", cachebufs,
sharedSize);
#endif
shareBuffer = mmap(NULL, sharedSize, PROT_READ | PROT_WRITE,
MAP_ANON | MAP_SHARED, -1, 0);
if (shareBuffer == MAP_FAILED) {
msg("can't mmap shared memory for buffer: %s\n",
strerror(errno));
return;
}
cheader = shareBuffer;
cdesc = (union cdesc *) (((char *) shareBuffer) +
sizeof(union cdesc));
cdata = ((char *) shareBuffer) + sizeof(union cdesc) +
sizeof(union cdesc) * cachebufs;
memset(shareBuffer, '\0', sharedSize);
}
}
/*
* Find the cache buffer descriptor that shows the minimal access time
*/
static int
findlru(void)
{
int i;
size_t minTime = cdesc[0].cd_time;
int minIdx = 0;
for (i = 0; i < cachebufs; i++) {
if (cdesc[i].cd_time < minTime) {
minIdx = i;
minTime = cdesc[i].cd_time;
}
}
return minIdx;
}
/*
* Read data directly from disk, with smart error handling.
* Try to recover from hard errors by reading in sector sized pieces.
* Error recovery is attempted at most BREADEMAX times before seeking
* consent from the operator to continue.
*/
static int breaderrors = 0;
#define BREADEMAX 32
void
rawread(daddr_t blkno, char *buf, int size)
{
int cnt, i;
#ifdef STATS
nphysread++;
physreadsize += size;
#endif
loop:
if (lseek(diskfd, ((off_t) blkno << dev_bshift), SEEK_SET) == -1) {
msg("rawread: lseek fails\n");
goto err;
}
if ((cnt = read(diskfd, buf, size)) == size)
return;
if (blkno + (size >> dev_bshift) > ufsib->ufs_dsize) {
/*
* Trying to read the final fragment.
*
* NB - dump only works in TP_BSIZE blocks, hence
* rounds `dev_bsize' fragments up to TP_BSIZE pieces.
* It should be smarter about not actually trying to
* read more than it can get, but for the time being
* we punt and scale back the read only when it gets
* us into trouble. (mkm 9/25/83)
*/
size -= dev_bsize;
goto loop;
}
if (cnt == -1)
msg("read error from %s: %s: [block %lld]: count=%d\n",
disk, strerror(errno), (long long)blkno, size);
else
msg("short read error from %s: [block %lld]: "
"count=%d, got=%d\n",
disk, (long long)blkno, size, cnt);
err:
if (++breaderrors > BREADEMAX) {
msg("More than %d block read errors from %s\n",
BREADEMAX, disk);
broadcast("DUMP IS AILING!\n");
msg("This is an unrecoverable error.\n");
if (!query("Do you want to attempt to continue?")) {
dumpabort(0);
/*NOTREACHED*/
} else
breaderrors = 0;
}
/*
* Zero buffer, then try to read each sector of buffer separately.
*/
memset(buf, 0, size);
for (i = 0; i < size; i += dev_bsize, buf += dev_bsize, blkno++) {
if (lseek(diskfd, ((off_t)blkno << dev_bshift),
SEEK_SET) == -1) {
msg("rawread: lseek2 fails: %s!\n",
strerror(errno));
continue;
}
if ((cnt = read(diskfd, buf, (int)dev_bsize)) == dev_bsize)
continue;
if (cnt == -1) {
msg("read error from %s: %s: [sector %lld]: "
"count=%ld\n", disk, strerror(errno),
(long long)blkno, dev_bsize);
continue;
}
msg("short read error from %s: [sector %lld]: "
"count=%ld, got=%d\n",
disk, (long long)blkno, dev_bsize, cnt);
}
}
void
bread(daddr_t blkno, char *buf, int size)
{
int osize = size, idx;
daddr_t oblkno = blkno;
char *obuf = buf;
daddr_t numBlocks = howmany(size, dev_bsize);
#ifdef STATS
nreads++;
readsize += size;
#endif
if (!shareBuffer) {
rawread(blkno, buf, size);
return;
}
if (flock(diskfd, LOCK_EX)) {
msg("flock(LOCK_EX) failed: %s\n",
strerror(errno));
rawread(blkno, buf, size);
return;
}
retry:
idx = 0;
while (size > 0) {
int i;
for (i = 0; i < cachebufs; i++) {
union cdesc *curr = &cdesc[(i + idx) % cachebufs];
#ifdef DIAGNOSTICS
if (curr->cd_owner) {
fprintf(stderr, "Owner is set (%d, me=%d), can"
"not happen.\n", curr->cd_owner, getpid());
}
#endif
if (curr->cd_blkend == 0)
continue;
/*
* If we find a bit of the read in the buffers,
* now compute how many blocks we can copy,
* copy them out, adjust blkno, buf and size,
* and restart
*/
if (curr->cd_blkstart <= blkno &&
blkno < curr->cd_blkend) {
/* Number of data blocks to be copied */
int toCopy = MIN(size,
(curr->cd_blkend - blkno) << dev_bshift);
#ifdef DIAGNOSTICS
if (toCopy <= 0 || toCopy > CSIZE) {
fprintf(stderr, "toCopy %d !\n",
toCopy);
dumpabort(0);
}
if (CDATA(curr) +
((blkno - curr->cd_blkstart) <<
dev_bshift) < CDATA(curr) ||
CDATA(curr) +
((blkno - curr->cd_blkstart) <<
dev_bshift) > CDATA(curr) + CSIZE) {
fprintf(stderr, "%p < %p !!!\n",
CDATA(curr) + ((blkno -
curr->cd_blkstart) << dev_bshift),
CDATA(curr));
fprintf(stderr,
"cdesc[i].cd_blkstart %lld "
"blkno %lld dev_bsize %ld\n",
(long long)curr->cd_blkstart,
(long long)blkno,
dev_bsize);
dumpabort(0);
}
#endif
memcpy(buf, CDATA(curr) +
((blkno - curr->cd_blkstart) <<
dev_bshift),
toCopy);
buf += toCopy;
size -= toCopy;
blkno += howmany(toCopy, dev_bsize);
numBlocks -= howmany(toCopy, dev_bsize);
curr->cd_time = cheader->cd_count++;
/*
* If all data of a cache block have been
* read, chances are good no more reads
* will occur, so expire the cache immediately
*/
curr->cd_blocksRead +=
howmany(toCopy, dev_bsize);
if (curr->cd_blocksRead >= nblksread)
curr->cd_time = 0;
goto retry;
}
}
/* No more to do? */
if (size == 0)
break;
/*
* This does actually not happen if fs blocks are not greater
* than nblksread.
*/
if (numBlocks > nblksread || blkno >= ufsib->ufs_dsize) {
rawread(oblkno, obuf, osize);
break;
} else {
ssize_t rsize;
daddr_t blockBlkNo;
blockBlkNo = (blkno / nblksread) * nblksread;
idx = findlru();
rsize = MIN(nblksread,
ufsib->ufs_dsize - blockBlkNo) << dev_bshift;
#ifdef DIAGNOSTICS
if (cdesc[idx].cd_owner)
fprintf(stderr, "Owner is set (%d, me=%d), can"
"not happen(2).\n", cdesc[idx].cd_owner,
getpid());
cdesc[idx].cd_owner = getpid();
#endif
cdesc[idx].cd_time = cheader->cd_count++;
cdesc[idx].cd_blkstart = blockBlkNo;
cdesc[idx].cd_blkend = 0;
cdesc[idx].cd_blocksRead = 0;
if (lseek(diskfd, ((off_t) blockBlkNo << dev_bshift),
SEEK_SET) == -1) {
msg("readBlocks: lseek fails: %s\n",
strerror(errno));
rsize = -1;
} else {
rsize = read(diskfd,
CDATA(&cdesc[idx]), rsize);
if (rsize < 0) {
msg("readBlocks: read fails: %s\n",
strerror(errno));
}
}
/* On errors, panic, punt, try to read without
* cache and let raw read routine do the rest.
*/
if (rsize <= 0) {
rawread(oblkno, obuf, osize);
#ifdef DIAGNOSTICS
if (cdesc[idx].cd_owner != getpid())
fprintf(stderr, "Owner changed from "
"%d to %d, can't happen\n",
getpid(), cdesc[idx].cd_owner);
cdesc[idx].cd_owner = 0;
#endif
break;
}
/* On short read, just note the fact and go on */
cdesc[idx].cd_blkend = blockBlkNo + rsize / dev_bsize;
#ifdef STATS
nphysread++;
physreadsize += rsize;
#endif
#ifdef DIAGNOSTICS
if (cdesc[idx].cd_owner != getpid())
fprintf(stderr, "Owner changed from "
"%d to %d, can't happen\n",
getpid(), cdesc[idx].cd_owner);
cdesc[idx].cd_owner = 0;
#endif
/*
* We swapped some of data in, let the loop fetch
* them from cache
*/
}
}
if (flock(diskfd, LOCK_UN))
msg("flock(LOCK_UN) failed: %s\n",
strerror(errno));
}
void
printcachestats(void)
{
#ifdef STATS
fprintf(stderr, "Pid %d: %d reads (%u bytes) "
"%d physical reads (%u bytes) %d%% hits, %d%% overhead\n",
getpid(), nreads, (u_int) readsize, nphysread,
(u_int) physreadsize, (nreads - nphysread) * 100 / nreads,
(int) (((physreadsize - readsize) * 100) / readsize));
#endif
}