On Wed, 2014-07-30 at 14:47 +0200, Peter Zijlstra wrote: > On Wed, Jul 30, 2014 at 10:30:48PM +1000, Alexey Kardashevskiy wrote: > > > > No, this is not my intention here. Here I only want to increment the counter. > > Full and hard nack on that. It should always be tied to actual pages, we > should not detach this and make it 'a number'. But this is the only way. We *cannot* go through the whole per-page locking logic every time the guest puts a translation into the IOMMU, this will completely kill guest performances for pass-through devices. Worse, for performances, because populating the iommu is a hypercall, we want to do it in "real mode" (special MMU-off environment) where we cannot rely on most normal kernel services such as normal locks, vmalloc space isn't accessible etc... So we don't have a choice. Either we let guests randomly pin arbitrary amounts of system memory, or we have a way to predictively account for the maximum that *can* be mapped/pinned in the iommu table to enable the fast path. Another problem with the mlock logic is that it doesn't refcount how many time a page has been locked, while the guest can map a given page multiple time in the iommu. Ben. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>