OK, so it looks like URBs need to learn about highmem.
Here's a quick proposal. It's rather first-drafty, and needs
discussion (for a week or so?).
- Add to struct urb:
struct page *page; /* [in] transfer buffer page */
unsigned offset; /* [in] offset within page */
- When transfer_buffer is null, that means page+offset
is used to indicate the start of a buffer. The length is
still specified by transfer_buffer_length.
- EITHER:
* each HCD is modified to manage page DMA
mappings if needed, in the same places they now
handle the single shot DMA mappings
- OR:
* the (new) hcd layer does this, once. In that case
it'd be advantageous to add another field to the
URB, just for HCDs to use:
dma_addr_t dma; /* [hcd] address for transfer */
A few things to note about that proposal:
- It doesn't apply to SETUP packets. Those would still
work the current way. There's not a lot of reason to
expect those eight bytes would be in highmem ...
- It doesn't address ISO transfers. Other folk know more
than I do about how they're used, and whether they'd
even need highmem. (The "transfer_buffer" isn't used
for ISO transfers...)
- If the mappings are set up/destroyed in the HCD layer,
so that HCD internals talk "dma+transfer_length" ...
That would enable a new a new non-URB bulk transfer
request that would allow a minor optimization. It could
talk directly in terms of scatterlists, calling pci_map_sg()
which can sometimes coalesce page buffers. The API
would not be _necessary_ to support scatter/gather I/O,
since it can already be done with bulk queueing.
The main issue I'd poke at is what to do with non-bulk
transfers, notably ISO.
An alternative approach, passing a scatterlist plus length
instead of page plus offset, would change HCD internals
substantially, and thus destabilize things (IMO needlessly).
I'm not sure I see anyone interested in doing that work.
- Dave
_______________________________________________
[EMAIL PROTECTED]
To unsubscribe, use the last form field at:
https://lists.sourceforge.net/lists/listinfo/linux-usb-devel
