[PATCH RFC 0/3] MAP_POPULATE for device memory

[Jarkko Sakkinen]

On Mon, Mar 07, 2022 at 10:11:19PM +0000, David Laight wrote:
> From: Christoph Hellwig
> > Sent: 07 March 2022 15:57
> > 
> > On Mon, Mar 07, 2022 at 03:29:35PM +0200, Jarkko Sakkinen wrote:
> > > So what would you suggest to sort out the issue? I'm happy to go with
> > > ioctl if nothing else is acceptable.
> > 
> > PLenty of drivers treat all mmaps as if MAP_POPULATE was specified,
> > typically by using (io_)remap_pfn_range.  If there any reason to only
> > optionally have the pre-fault semantics for sgx?  If not this should
> > be really simple.  And if we have a real need for it to be optional
> > we'll just need to find a sane way to pass that information to ->mmap.
> 
> Is there any space in vma->vm_flags ?
> 
> That would be better than an extra argument or function.

It's very dense but I'll give a shot for callback route based on Dave's
comments in this thread. I.e. use it as filter inside __mm_populate() and
populate_vma_page_range().

For Enarx, which we are implementing being able to use MAP_POPULATE and get
the full range EAUG'd would be best way to optimize the performance of wasm
JIT (Enarx is a wasm run-time capable of running inside an SGX enclave, AMD
SEV-SNP VM etc.). More so than any predictor (ra_state, madvice etc.) inside
#PF handler, which have been suggested in this thread.

After some research on how we implement user space, I'd rather keep the #PF
handler working on a single page (EAUG a single page) and have either ioctl
or MAP_POPULATE to do the batch fill.

We can still "not trust the user space" i.e. the populate does not have to
guarantee to do the full length since the #PF handler will then fill the
holes. This was one concern in this thread but it is not hard to address.

BR, Jarkko