On 18 October 2016 at 14:04, Mike Travis travis@sgi.com wrote:
On 10/18/2016 5:49 AM, Ard Biesheuvel wrote:
On 18 October 2016 at 13:46, Mike Travis travis@sgi.com wrote:
Have you tested this change on a system which has physical addressing > 40 bits to insure that addresses > 40 bits are not simply truncated to an incorrect (mod 40) address?
The device should never see addresses > 40 bits, given that the DMA mask is set to 40 bits not 64.
I agree that the device should never see that. But how is that restriction imposed? If a user presents an buffer address > 40 bits how would that be converted to useful buffer at <= 40 bits?
I believe I tried this (at the time the kernel was at version 2.6) and it did not work. Since audio I/O was such a low (relative) bandwidth, it was not worth the effort to rework the audio code to allow buffers > 32 bits but <= 40 bits. It may have changed since then?
This is very platform dependent. Note that the 40 bit DMA mask describes the memory view from the PCI side, which may not be mapped 1:1 with system memory (although it usually is on x86)
In general, a platform will wire up the device to a struct dma_map_ops {} instance that describes how DMA is performed for this particular device on this platform. Whether it resorts to bounce buffering or simply maps the buffer into the device accessible window using an IOMMU is thus also platform dependent. But in general, all these drivers ensure that the device address that results from a mapping adheres to the DMA mask that has been set at probe time.