Add a caching option to dump. Use -C. Note that NetBSD has a caching option

called -r but it takes 512 byte blocks instead of megabytes, and I felt a
megabytes specification would be far more useful so I did not use the same
option character.

This will *greatly* improve dump performance at the cost of possibly
missing filesystem changes that occur between passes, and does a fairly
good job making up for the loss of buffered block devices.  Caching is disabled
by default to retain historical behavior.

In tests, dump performance improved by about 40% when dumping / or /usr.

Beware that dump forks and the cache may wind up being larger then you
specify, but a more complex shared memory implementation would not produce
results that are all that much better so I kept it simple for now.

MFC after:	3 days

1 files changed, 4 insertions, 3 deletions

diff --git a/sbin/dump/traverse.c b/sbin/dump/traverse.c
index ab2ee71aa2a0..ec0d56e547ff 100644
--- a/sbin/dump/traverse.c
+++ b/sbin/dump/traverse.c
@@ -739,8 +739,8 @@ bread(ufs2_daddr_t blkno, char *buf, int size)
 	int cnt, i;
 
 loop:
-	if ((cnt = pread(diskfd, buf, size, ((off_t)blkno << dev_bshift))) ==
-						size)
+	cnt = cread(diskfd, buf, size, ((off_t)blkno << dev_bshift));
+	if (cnt == size)
 		return;
 	if (blkno + (size / dev_bsize) > fsbtodb(sblock, sblock->fs_size)) {
 		/*
@@ -774,7 +774,8 @@ loop:
 			breaderrors = 0;
 	}
 	/*
-	 * Zero buffer, then try to read each sector of buffer separately.
+	 * Zero buffer, then try to read each sector of buffer separately,
+	 * and bypass the cache.
 	 */
 	memset(buf, 0, size);
 	for (i = 0; i < size; i += dev_bsize, buf += dev_bsize, blkno++) {