Date: Sun, 25 Nov 2001 12:45:35 +0100
From: Tarjei Huse <[EMAIL PROTECTED]>
This is from lwn.net, latest. How large will the effect be for
cyrus-imapd? I guess this is a good reason to advize people to use
reiser or ext3 :) Tarjei
Honestly, I thought the +S bit already did fully synchronous writes.
ext2's performance with +S has always been fairly horrible compared to
a standard FFS system, so now I'm even more puzzled.
Anyway, this is a horrible idea. If they must do this, they should
make a new mount option for the new semantics, not modifying an old
one.
Larry
How synchronous should Linux be? One day, as Andrew Morton was
wandering around the filesystem code, he noticed a seeming
inconsistency. While Linux, like most operating systems that are
concerned with performance, buffers filesystem writes in the
kernels, it does provide a couple of ways for the user to request
synchronous behavior:
The "sync" option, used with the mount command, will request
synchronous writes for every file in the filesystem.
The chattr +S command can request synchronous updates for a single
file or directory. What Andrew noticed is that, while most
operations are indeed implemented synchronously, the data actually
written to files is not. So he posted a patch making writes
synchronous as well. It turns out, though, that there are two
types of opinion regarding synchronous writes of file data. Linux
has never, in the past, had that behavior. The claim in the mount
man page ("All I/O to the file system should be done
synchronously.") is simply incorrect. Fully synchronous behavior is
very expensive, leading to horrible performance, and is, according
to some, rarely needed. It is better, according to this camp, to
expect applications to use the fsync() system call to explicitly
force synchronous behavior when it's needed in a specific
situation. Rather than implement synchronous file writes, these
folks (as typified by Stephen Tweedie) propose instead to limit
implicit synchronous behavior to directories.
On the other side, Jeff Garzik argued that implementing synchronous
file writes is the correct thing to do. Users sometimes need that
behavior, and it is impractical to hack up applications and scripts
to call fsync() explicitly. A separate dirsync mount option could
provided to request synchronous semantics for directories only.
Amusingly, Andrew appears to have been won over by the first point
of view, but his patch (implementing the second) found its way into
2.4.15-pre5. Unless something changes, fully synchronous behavior
will be the way of things in
2.4.15.
