Re: "legacy" usb stack fixes

On Fri, 12 Sep 2008, M. Warner Losh wrote:

> In message: <48CAB37F.50002_at_samsco.org>
>            Scott Long <scottl_at_samsco.org> writes:
> : Scott Long wrote:
> : > This is close to How Things Should Be.  Each umass target having its own
> : > SIM and bus is indeed wrong, but I'm not sure if it's correct for all
> : > USB controllers and buses to be under a single SIM.  What would be the
> : > most correct is for each physical USB controller/bus instance to have
> : > its own SIM instance.  I don't know if it's better to do the attachment
> : > in ehci/ohci/uhci controller drivers or in usb bus driver; up in the
> : > controller drivers is probably more correct.  I don't like this hack of
> : > attaching stuff in a SYSINIT.
> : >
> : > Scott
> : >
> : >
> :
> : Now that I've thought some on it, I'll go one step further and say that
> : registering a single SIM for multiple controller+bus instances in a
> : SYSINIT will be highly undesirable thing to do.  Since you have to
> : register a lock with the CAM when you register the SIM, you'll wind up
> : serializing all of the USB controllers under a single lock.  Or you'll
> : probably try something dangerous and tricky with dropping the new global
> : lock and picking up an individual lock, then swizzling locks in the
> : completion and event paths, with the result being rather unsatisfying
> : and unpleasant.  So I know that you'll do what you believe is correct,
> : but please take my advice on the matter anyways.
>
> Yes.  A SIM will serialize all operations, and the most logical place
> for that is the computer <-> usb interface, which is the host
> controller.  So having one SIM per host controller would be the
> optimal placement.  Having one SIM per usb device doesn't result in
> any more real parallelism because the host controller necessarily
> serializes things because of how USB is defined...
>

Correct.  Another argument for having a SIM per controller/bus and not per 
target is that the SIM is responsible for managing all resources on a
controller.  USB is still a bus topology, and thus certain resources are
finite and shared, be they bandwidth, arbitration, or concurrency. 
Granted, USB is simple enough that it doesn't give you much control over 
these resources, but having the SIM be at the target level gives the 
system even less control and visibility.  If a future enhancement to USB
grows the ability to do useful things like more concurrency, it'll be 
essential for the SIM to have a controller-wide view of this.

Scott