On Sun, Jun 03, 2001 at 09:00:24AM -0700, David Brownell wrote:
> > From: "Georg Acher" <[EMAIL PROTECTED]>
> > On Sat, Jun 02, 2001 at 01:22:00PM +0200, Pavel Machek wrote:
> > > 
> > > With Acher's uhci, even ifconfig up of usb-to-usb networking device
> > > [plusb handled by generic usb-to-usb driver; see -ac series].
> > > does 50% slowdown. When fsbr is being used, systems slows down by 350%
> > 
> > Hm, the 50% make me curious... have to look what's happening...
> 
> For PL-2301/2302 devices, "ifconfig up" is mostly just posting a bulk read.
> True with both "usbnet" and its (now obsolete) predecessor drivers "plusb"
> and (for different devices) "net1080".
> 
> Laplink-style cables can often support another mode (poll via USB "interrupts",
> and then issue reads only when data is available) but not every device can work
> that way (like, I seem to recall, an iPaq PDA).  And that'd increase the latency
> per packet by a couple milliseconds, even when it's possible.

I thought that it is a dangling bulk read, but I have to check why it is
still consuming PCI bandwidth, even without transmitting data... 

> > > (running more than 4 times slower than normally. Ouch).
> > 
> > Blame Intel. Either low latency or low PCI usage, you can choose...
> 
> This problem is UHCI-specific, not USB-generic, yes?
> Doesn't happen with OHCI?

UHCI runs its descriptor chain only once a millisecond, which is bad, if you
use breadth access mode to provide fair bandwith. Then each transfer can
only get one TD (64byte) in one millisecond through. To avoid that, the
bandwith reclamation is used, a nice word for a design/concept flaw, it
simply loops the descriptor chain. That causes a high PCI load, even with no
transfers in-progress, the descriptors are always read via PCI.
Unfortunately, there's no delay-TD or such...

-- 
         Georg Acher, [EMAIL PROTECTED]         
         http://www.in.tum.de/~acher/
          "Oh no, not again !" The bowl of petunias          

_______________________________________________
[EMAIL PROTECTED]
To unsubscribe, use the last form field at:
http://lists.sourceforge.net/lists/listinfo/linux-usb-devel

Reply via email to