From: Jarkko Sakkinen
> Sent: 06 March 2022 05:32
> 
> For device memory (aka VM_IO | VM_PFNMAP) MAP_POPULATE does nothing. Allow
> to use that for initializing the device memory by providing a new callback
> f_ops->populate() for the purpose.
> 
> SGX patches are provided to show the callback in context.
> 
> An obvious alternative is a ioctl but it is less elegant and requires
> two syscalls (mmap + ioctl) per memory range, instead of just one
> (mmap).

Is this all about trying to stop the vm_operations_struct.fault()
function being called?

It is pretty easy to ensure the mappings are setup in the driver's
mmap() function.
Then the fault() function can just return -VM_FAULT_SIGBUS;

If it is actually device memory you just need to call vm_iomap_memory()
That quite nicely mmap()s PCIe memory space into a user process.

Mapping driver memory is slightly more difficult.
For buffers allocated with dma_alloc_coherent() you can
probably use dma_mmap_coherent().
But I have a loop calling remap_pfn_range() because the
buffer area is made of multiple 16kB kernel buffers that
need to be mapped to contiguous user pages.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)

On Sun, Mar 06, 2022 at 08:30:14AM +0000, David Laight wrote:
> From: Jarkko Sakkinen
> > Sent: 06 March 2022 05:32
> > 
> > For device memory (aka VM_IO | VM_PFNMAP) MAP_POPULATE does nothing. Allow
> > to use that for initializing the device memory by providing a new callback
> > f_ops->populate() for the purpose.
> > 
> > SGX patches are provided to show the callback in context.
> > 
> > An obvious alternative is a ioctl but it is less elegant and requires
> > two syscalls (mmap + ioctl) per memory range, instead of just one
> > (mmap).
> 
> Is this all about trying to stop the vm_operations_struct.fault()
> function being called?

In SGX protected memory is actually encrypted normal memory and CPU access
control semantics (marked as reserved, e.g. struct page's).

In SGX you need call ENCLS[EAUG] outside the protected memory to add new
pages to the protected memory. Then when CPU is executing inside this
protected memory, also known as enclaves, it commits the memory as part of
the enclave either with ENCLU[EACCEPT] or ENCLU[EACCEPTCOPY].

So the point is not prevent page faults but to prepare the memory for
pending state so that the enclave can then accept them without round-trips,
and in some cases thus improve performance (in our case in enarx.dev
platform that we are developing).

In fact, #PF handler in SGX driver in the current SGX2 patch set also does
EAUG on-demand. Optimal is to have both routes available. And said, this
can be of course also implemented as ioctl.

BR, Jarkko