Re: [PATCH 2 of 4] Introduce i386 fibril scheduling

[Linus Torvalds]

On Sat, 3 Feb 2007, Ingo Molnar wrote:
> 
> Well, in my picture, 'only if you block' is a pure thread utilization 
> decision: bounce a piece of work to another thread if this thread cannot 
> complete it. (if the kernel is lucky enough that the user context told 
> it "it's fine to do that".)

Sure, you can do it that way too. But at that point, your argument that we 
shouldn't do it with fibrils is wrong: you'd still need basically the 
exact same setup that Zach does in his fibril stuff, and the exact same 
hook in the scheduler, testing the exact same value ("do we have a pending 
queue of work").

So at that point, you really are arguing about a rather small detail in 
the implementation, I think.

Which is fair enough. 

But I actually think the *bigger* argument and problems are elsewhere, 
namely in the interface details. Notably, I think the *real* issues end up 
how we handle synchronization, and how we handle signalling. Those are in 
many ways (I think) more important than whether we actually can schedule 
these trivial things on multiple CPU's concurrently or not.

For example, I think serialization is potentially a much more expensive 
issue. Could we, for example, allow users to serialize with these things 
*without* having to go through the expense of doing a system call? Again, 
I'm thinking of the case of no IO happening, in which case there also 
won't be any actual threading taking place, in which case it's a total 
waste of time to do a system call at all.

And trying to do that actually has implications for the interfaces (like 
possibly returning a zero cookie for the async() system call if it was 
doable totally synchronously?)

Signal handling is similar: I actually think that a "async()" system call 
should be interruptible within the context of the caller, since we would 
want to *try* to execute it synchronously. That automatically means that 
we have semantic meaning for fibrils and signal handling.

Finally, can we actually get POSIX aio semantics with this? Can we 
implement the current aio_xyzzy() system calls using this same feature? 
And most importantly - does it perform well enough that we really can do 
that?

THOSE are to me bigger questions than what happens inside the kernel, and 
whether we actually end up using another thread if we end up doing it 
non-synchronously.

					Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/