Re: [linux-usb-devel] Re: highmem and usb [was: "sr: unaligned transfer" in 2.5.2-pre1]

[David Brownell <david-b@pacbell.net>]

> > > Not that I've seen a writeup about highmem (linux/Documentation
> > > doesn't seem to have one anyway) but if I infer correctly from that
> > > DMA-mapping.txt writeup, URBs don't support it because there's no way
> > > to specify buffers as a "struct page *" or an array of "struct
> > > scatterlist".  That's the only way that document identifies to access
> > > "highmem memory".
>
> This sounds like another good reason to have URBs take scatterlists
> directly, oddly enough. :)

If it's got to be done, I'd much rather it were "page + offset", so that the
usbcore code can be simpler.  We know how to turn scatterlists into
bulk queued requests, so there's no need for anything more ... :)


> > No, you can always ask to get pages low mem bounced. Highmem is no
> > requirement, and if your device really can't support it there's no point
> > in attempting to support it.
>
> I presume there is some overhead in bouncing to lowmem?  I imagine that
> highmem support for the HCDs wouldn't be that difficult -- they are just
> PCI devices, after all.

I'm unclear on what "bouncing to lowmem" involves, but I'd rather avoid
teaching all three HCDs a second model for addressing transfer buffers.

At least until later in the 2.5 series, when we believe they'll share a lot
more common code and so that new model can be taught to just ONE
piece of code.  Fixing bugs in one place easier than in three!


> I'd rather eliminate as much overhead as possible -- I already get
> complaints from performance fanatics about the inability of usb-storage to
> get past 92% bus saturation (sustained), and the problem will only get
> worse on USB 2.0

Well then you'll  be glad to see a patch from me, soonish, that teaches
the usb-storage "transport" code to use bulk queueing.  That'll get the
bandwidth utilization up as high as it can get.  It won't address any of
these highmem issues though.

- Dave