Re: GUP guarantees wrt to userspace mappings

From: Andrea Arcangeli
Date: Mon May 02 2016 - 15:11:54 EST

On Mon, May 02, 2016 at 05:22:49PM +0200, Jerome Glisse wrote:
> I think this is still fine as it means that device will read only and thus
> you can migrate to different page (ie the guest is not expecting to read back
> anything writen by the device and device writting to the page would be illegal
> and a proper IOMMU would forbid it). So it is like direct-io when you write
> from anonymous memory to a file.

Agreed. write=1 is so that if there's an O_DIRECT write() and the app
is only reading, there will be no COW generated on shared anonymous

> Now that i think again about it, i don't think it exist. pmdp_collapse_flush()
> will flush the tlb and thus send an IPI but get_user_pages_fast() can't be
> preempted so the flush will have to wait for existing get_user_pages_fast() to
> complete. Or am i missunderstanding flush ? So khugepaged is safe from GUP_fast
> point of view like the comment, inside it, says.

This is exactly correct, there's no race window.

The IPI (or the quiescent point in case of the gup_fast RCU version)
are the things that flush away get_user_pages_fast with pmdp_collapse_flush().

> Well you can't not rely on special vma here. Qemu alloc anonymous memory and
> hand it over to guest, then a guest driver (ie runing in the guest not on the
> host) try to map that memory and need valid DMA address for it, this is when
> vfio (on the host kernel) starts pining memory of regular anonymous vma (on
> the host). That same memory might back some special vma with ->mmap callback
> but in the guest. Point is there is no driver on the host and no special vma.
> From host point of view this is anonymous memory, but from guest POV it is
> just memory.

It's quite important it stays regular tmpfs/anon as device memory is
managed by the device and we'd lose everything (KSM/swapping/NUMA
balancing/compaction/memory-hotunplug/CMA etc..).