UFS

1. We should not caclucate the superblock summary, since it can at most lag 30 seconds’ change, and doing the caclucate is painful delay. What we should do is to transfer the responsibility to background fsck
2. newfs(8) should warn about the snapshot is actually unable to create, in case fssize is greater than maxfilesize.

The patch for 1) is available at http://people.freebsd.org/~delphij/ffs/patch-sbsummary. I will work on 2) shortly, as it’s quite simpler than 1).

It turns out that we have to optimize our FS before benchmarking. I have read some articles and maillist archives so I got some impression:

1. Give correct parameters to newfs(1)
   avgfilesize, filesperdir, will control DIRPERF hashing.

2. Turn on noatime mounts.

3. etc.

Just make a note.

While reading the snapshot code in FFS Soft Updates’s FreeBSD implementation, I got a question about the following code:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

		/*
		 * If there is a fragment, clear it here.
		 */
		blkno = 0;
		loc = howmany(xp->i_size, fs->fs_bsize) - 1;
		if (loc < NDADDR) {
			__len = fragroundup(fs, blkoff(fs, xp->i\_size));__
			if (len < fs->fs_bsize) {
				ffs_blkfree(copy_fs, vp, DIP(xp, i_db[loc]),
				    len, xp->i_number);
				blkno = DIP(xp, i_db[loc]);
				DIP_SET(xp, i_db[loc], 0);
			}
		}

主要是空格的问题。我打算接下来的一段时间琢磨琢磨SoftUpdates，如果这个patch能commit的话，delphijfork的diff一下就能减少46K :-)

The current FreeBSD Soft Updates poses to a performance peanlity when there’re plenty of i-nodes and this is very common case in badly-designed systems, however, I think this should be solved.

I have just tried to umount /usr on my test box in order to get all outstanding writes to be sync’ed to disk (as Kirk McKusick said, the umount system call will do true sync of a file system).