> This won't probably work with IOMMU. E.g. on sparc64 once we kill GFP_DMA
> limiting to first 4GB physical. That's why buses using IOMMU will need some
> kind of acquire_dma/acquire_dma_sg/release_dma/release_dma_sg, so that IOMMU
> page tables can be dynamically filled up. And if you want to go between two
> different cards, then you'll have to use some special function which will
> fill two IO page tables if necessary (e.g. SBUS and PCI one). I don't know
> if SBUS<->SBUS DMA actually works, there are some bits for it but have never
> tried, but PCI<->other PCI should definitely work.
What about doing it this way: (imagine we need to access memory on bus1
from bus2)
(1) Convert bus1 address to physical address via bus1-specific function.
(2) Call bus2->acquire_dma on bus2 (a bus2-specific function).
This sets up the IOMMU.
(3) Do the transfer.
(4) Call bus2->release_dma.
BTW what should happen if the IOMMU address space is full? I guess
we cannot block in acquire_dma() since it can be run from a BH.
> But you need to take care of the bus endianity in case a system has more
> buses with different endians. E.g. say you have SBUS and PCI, SBUS is big
> endian and PCI little endian.
> So you either stick them into bus_ops (I'd recommend that),
Accessing readl() et al. via bus_ops is awfully slow.
> plus have some
> possibly inlineable functions for each different bus e.g. for cases where a
> driver is single bus only and will always be (many are).
> So you'd have
> read_32(bus_t bus, ptr) which would do bus->bus_ops->read_32(ptr) and then
> pci_read_32(ptr) which would be the pci version of the same. All these would
> work in its bus native order.
The inline functions are much better, but what if a single driver is able
to handle the same chip connected via different bus interfaces (for example
an ethernet chip available in both ISA and PCI cards)?
Have a nice fortnight
-- Martin `MJ' Mares <mj@ucw.cz> http://atrey.karlin.mff.cuni.cz/~mj/ Faculty of Math and Physics, Charles University, Prague, Czech Rep., Earth "If a train station is where the train stops, what is a work station?"- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/