Re: 5.99.42/i386 crash (backtrace + core available)

[Antti Kantee <pooka%cs.hut.fi@localhost>]

On Mon Jan 10 2011 at 09:34:46 +0000, Matthias Scheler wrote:
> On Mon, Jan 10, 2011 at 11:06:58AM +0200, Antti Kantee wrote:
> > dateline 2006 a.d.:
> > 
> > Threads and locks are difficult, so let's make the default userspace
> > file system model based on coprocesses with cooperative scheduling.
> > That will make things easy.
> 
> Where am I saying that threads are difficult? AmigaOS uses a thread for
> each file-system.

I didn't suggest you were saying that in 2006.  That was my thinking
back then: use one thread per file system and things will be easy.
The rest of the mail was explaining how through practical experience
I've realized that is incorrect thinking.

> > flash-forward to present day:
> > 
> > Nope.  It turns out that file systems are difficult because file systems
> > are difficult, not because locking is difficult.
> 
> How come in that case that nobody managed to fixed our rename races in
> the last 24 month?

"because file systems are difficult"

If you think that having all fs data structures accessible from a single
thread will solve all problems while keeping performance at an acceptable
level, go right ahead and prove me wrong.  I'm just saying I've already
implemented that model and in my implementation it didn't magically
solve all problems (even for ops other than the dreaded rename!).

-- 
älä karot toivorikkauttas, kyl rätei ja lumpui piisaa