On Wed, May 20, 2015 at 11:01:27AM +0100, Christoffer Dall wrote: > On Tue, May 19, 2015 at 12:18:54PM +0100, Catalin Marinas wrote: > > On Tue, May 19, 2015 at 11:03:22AM +0100, Andrew Jones wrote: > > > On Mon, May 18, 2015 at 04:53:03PM +0100, Catalin Marinas wrote: > > > > I didn't have time to follow these threads in details, but just to > > > > recap my understanding, we have two main use-cases: > > > > > > > > 1. Qemu handling guest I/O to device (e.g. PCIe BARs) > > > > 2. Qemu emulating device DMA > > > > > > > > For (1), I guess Qemu uses an anonymous mmap() and then tells KVM about > > > > this memory slot. The memory attributes in this case could be Device > > > > because that's how the guest would normally map it. The > > > > file_operations.mmap trick would work in this case but this means > > > > expanding the KVM ABI beyond just an ioctl(). > > > > > > > > For (2), since Qemu is writing to the guest "RAM" (e.g. video > > > > framebuffer allocated by the guest), I still think the simplest is to > > > > tell the guest (via DT) that such device is cache coherent rather than > > > > trying to remap the Qemu mapping as non-cacheable. > > > > > > If we need a solution for (1), then I'd prefer that it work and be > > > applied to (2) as well. Anyway, I'm still not 100% sure we can count on > > > all guest types (booloaders, different OSes) to listen to us. They may > > > assume non-cacheable is typical and safe, and thus just do that always. > > > We can certainly change some of those bootloaders and OSes, but probably > > > not all of them. > > > > That's fine by me. Once you get the vma splitting and attributes > > changing done, I think you get the second one for free. > > > > Do we want to differentiate between Device and Normal Non-cacheable > > memory? Something like KVM_MEMSLOT_DEVICE? > > > > Nitpick: I'm not sure whether "uncached" is clear enough. In Linux, > > pgprot_noncached() returns Strongly Ordered memory. For Normal > > Non-cachable we used pgprot_writecombine (e.g. a video framebuffer). > > > > Maybe something like KVM_MEMSLOT_COHERENT meaning a request to KVM to > > ensure that guest and host access it coherently (which would mean > > writecombine for ARM). That's similar naming to functions like > > dma_alloc_coherent() that return cacheable or non-cacheable memory based > > on what the device supports. Anyway, I'm not to bothered with the > > naming. > > > One thing to keep in mind for (2) is that QEMU is likely to do things > like calling regular memcpy() on the memory region, so mapping it as > device memory which would fault on unaligned accesses may be a problem, > so ideally there is a memory type for the user space mapping which > allows such behavior where we at the same time can guarantee the that > the mapping is coherent with the guest mapping through the S2 > attributes.
I agree, for (2) we need Normal memory (either cacheable or non-cacheable, though as I can see it's more likely the latter as we can't guarantee the guest honouring "dma-coherent" device properties). -- Catalin
