On 10/11/16 15:59, Joerg Roedel wrote: > On Tue, Nov 08, 2016 at 11:37:23AM +0000, Robin Murphy wrote: >> TBH I've been pondering ripping the size stuff out of dma-iommu, as it >> all stems from me originally failing to understand what dma_32bit_pfn is >> actually for. > > The point of dma_32bit_pfn is to allocate dma-address below 4G by > default. This is a performance optimization so that even devices capable > of 64bit DMA are using SAC by default instead of DAC. > > Since it is the goal to share a dma-iommu implemenation between > architectures, I would rather prefer not to rip this stuff out.
Oh, I didn't mean rip it out entirely, just get rid of the bogus assumption that it's the "size" of the domain, especially when given a >32-bit DMA mask, since that defeats the very optimisation I do now understand (although it might still be OK for platform devices where SAC/DAC doesn't apply, to avoid the rb_last() overhead every time). >From the patch I've started, "rip it out" turns out to actually be mostly "rewrite the comments" anyway - I'll post something soon. Robin.
