Subject: Re: "fast" syscalls?
To: None <email@example.com, firstname.lastname@example.org>
From: List Mail User <track@Plectere.com>
Date: 03/17/2006 14:10:44
>On Mar 17, 2006, at 10:30 AM, Garrett D'Amore wrote:
>> This is almost exactly what I had in mind. It is the ioctl that I
>> wanted to optimize away.
>I can think of a couple of ways to make it less expensive for the u-
> >k notification... but I'll have to stew on it for a bit first...
>My brain is doing in the direction of a generic message queue
>description / notification API... If we had a generic subsystem, it
>would be reasonable to implement a generic "fast trap" for this on
>platforms that can do it.
>Anyway, for now I would suggest "use an ioctl to get your stuff
>working", and then we can always speed it up later..
I don't know how much is publicly documented, but try looking at
the mechanisms in IRIX for u->k (the GLqueue handled by code in the page
fault path) and k->u (shmiq and qcntl). The biggest problem becomes doing
validation on the data passed each way, not merely passing the data. At
the cost of a test and branch in the page fault handler, you can implement
this easily on any machine with an MMU (at least any I know of).