Greg KH wrote:
> On Sat, Jul 20, 2002 at 06:15:46PM -0700, David Brownell wrote:
> 
>>Part 1:
>>
>>  - HCDs remove special (a) "can't queue" and (b) "only one packet"
>>    limitations for interrupt transfers (*)
> 
> 
> Is this _really_ needed?  Do any drivers do anything special right now
> to work around this?  If this is a limitation of UHCI hardware, I don't
> think it's worth the extra effort and code to queue these in the host
> controller, when no drivers want this right now.

Yes.  For "only one packet", I think many folk using INTERRRUPT-OUT work
around it.  And not that I've spent time with HID, but I understand that
multi-packet reports are possible there too ... both directions.

This is the guts of why INTR-OUT transfers are painful today.  Writing
a 73 byte report takes N fullsize and one short packet.  Using the one
packet model, that's simple enough with queuing; or using bigger writes,
the queueing might not be needed.  And so on; arbitrary restrictions
just making things complicated.

Someone once reported they weren't getting 1/msec interrupt transfers
in some setup (ISTR), and the generic fix for that is ensuring that the
hardware can have a queue to munch on while the host CPU is tied up.

It's not a UHCI limitation as near as I can tell.  The spec even talks
about interrupt TDs living in a QH, and the "breadth first" mode seems
designed for this particular usage.


>>Part 2:
>>
>>  - HCDs remove (c) automagic resubmit
>>  - USB device drivers updated to not rely on it (including usbfs)
> 
> 
> This is something I see as needed, but does "Part 1" really have to
> happen before this is done?

I don't remember such a dependency, though it might be smoother for
some driver updates to sequence it one way or another.  If nothing
else it's how I'm thinking of staging this stuff.


>>I suspect that the best solution here is not to completely remove
>>this capability, but just define a new transfer flag to explicitly
>>disable the automagic resubmit.
> 
> 
> No.  Let's rip it out.  Make the driver do the resubmit, just like they
> have to do for bulk urbs.  No special cases or flags needed.

Nicer for sure, but unmodified drivers will have silent failures.
I suppose that'll show ... somebody something!  :)


>>Note that this doesn't address the SL-811 and CRIS host controller
>>drivers, both of which are in the 2.5 tree.  I've cc'd the author
>>of the SL-811 code; there's no email address on the CRIS code.
>>Or any of the various host controller drivers that aren't living in
>>any of the standard Linux kernel trees.
> 
> 
> If it isn't in the kernel tree, don't worry about it :)

OK.

- Dave




-------------------------------------------------------
This sf.net email is sponsored by:ThinkGeek
Welcome to geek heaven.
http://thinkgeek.com/sf
_______________________________________________
[EMAIL PROTECTED]
To unsubscribe, use the last form field at:
https://lists.sourceforge.net/lists/listinfo/linux-usb-devel

Reply via email to