Re: [9fans] book chapters

[William Ahern <william@25thandClement.com>]

On Fri, Jun 27, 2003 at 06:08:13PM -0400, Joel Salomon wrote:
> On Fri, 27 Jun 2003, pac wrote:
<snip>
> * Unix Assumes a Static File System
> I'm out of my depths here - anyone able to answer this?

Linux has this solution *wrong*. the DNOTIFY mechanism is accessed
thru fcntl(), and DNOTIFY uses signals. ugh!

I don't know if BSD _copied_ Linux. The free BSD's have kqueue(),
which is an interface that smoothes over file descriptors, signals,
process status and async-io context pointers. It includes a mechanism
to be notified on all file events. DNOTIFY on Linux only notifies
on directories (tho it will trigger a directory event when children
are modified).

However, Linux might get epoll(), which is supposed to be similar to
Solaris' /dev/poll, but a tad more general. It looks similar to kqueue(),
but not nearly as robust. (for better or worse)

I'm working on a shell utility, `watch', which outputs file event
notifications.

The code is ugly, because I ran into a lot of caveats figuring out
the semantics of kqueue vnode events.

	[http://www.25thandclement.com/~william/projects/watch.html]

I guess one tie-in to Plan 9 is that none of the BSD's (free ones, at least)
keep track of path names in their fs layer. Linux does, and in fact on Linux
/dev/pid/[fd] is analagous to fd2path(), except Linux tracks rename()s. Ted
Unangst, an OpenBSD developer, wrote me to say that he might hack OpenBSD to
keep some path info. I talked to him about it because if I'm monitoring a
large tree for events (say in a CMS), kqueue() forces me to open every last
file, while in Linux I need only open every directory; the issue being
descriptor table limits.

- Bill