MFffs ffs_balloc.c 1.5.

Long ago, bread() set b_blkno to the disk block number as a side effect
of doing physical i/o (or it just retained the setting from when the
i/o was done).  The setting is lost when buffers go away and then are
reconsituted from VM.  bread() originally compensated by doing a
VOP_BMAP() to recover b_blkno, but this was no good since it sometimes
caused extra i/o or even deadlock for bread()ing metadata to do the
bmap.  This was fixed in vfs_bio.c 1.33 (1995/03/03) and ffs_balloc.c
1.5, etc., by removing the VOP_BMAP() from bread() and breadn(), and
changing all (?) places that used b_blkno to set it if necessary.

ext2fs was not imported until later in 1995 and was still depending on
the old behaviour of bread() in at least ext2_balloc().  This caused
filesystem and file corruption by clobbering direct block numbers in
inodes.

1 files changed, 2 insertions, 0 deletions

diff --git a/sys/gnu/fs/ext2fs/ext2_balloc.c b/sys/gnu/fs/ext2fs/ext2_balloc.c
index 75c17e5..559f4a6 100644
--- a/sys/gnu/fs/ext2fs/ext2_balloc.c
+++ b/sys/gnu/fs/ext2fs/ext2_balloc.c
@@ -111,6 +111,7 @@ ext2_debug("ext2_balloc called (%d, %d, %d)\n",
 				brelse(bp);
 				return (error);
 			}
+			bp->b_blkno = fsbtodb(fs, nb);
 			*bpp = bp;
 			return (0);
 		}
@@ -126,6 +127,7 @@ ext2_debug("ext2_balloc called (%d, %d, %d)\n",
 					brelse(bp);
 					return (error);
 				}
+				bp->b_blkno = fsbtodb(fs, nb);
 			} else {
 			/* Godmar thinks: this shouldn't happen w/o fragments */
 				printf("nsize %d(%d) > osize %d(%d) nb %d\n",