Re: Using TMPFS for /tmp and /var/run?

Am 30.03.2012 21:36, schrieb Adrian Chadd:
> Let me tell you a story.
> 
> Someone decided that ext4 could have a decent speed up if it
> implemented the posix standard for not flushing files on close().
> After all, if you needed it to be guaranteed to be written to disk,
> you would call a flush routine first, before you called close().
> 
> So they did this.
> 
> Then people testing out ext4 discovered that upon crash, their
> kde/gnome profiles were corrupted.
> 
> Why? Because KDE/Gnome authors hadn't ever called flush before
> close(), and they weren't the only ones. They didn't read the
> standard, they only used the system and fixed bugs whenever their
> system behaved against their expectations. They didn't notice that the
> system was being different from the standard.
> 
> Guess what ext4 did? :)

ext4 sprouted an option (auto_da_alloc, when used with the proper data
journalling option data=ordered) to support buggy software.

Note that ext4 isn't pioneering the "fsync() required" semantics here,
there are other precedents of "0-blocks in files after crash" in Linux
file systems, such as XFS.

I'm oblivious to the current ext4 defaults WRT these semantics (and I
haven't looked at vanilla kernels for a while anyways---distros might
have changed default settings).