Re: [PATCH v16 18/22] platform/x86: Intel SGX driver

[Jarkko Sakkinen]

On Thu, Nov 15, 2018 at 10:04:06PM +0200, Jarkko Sakkinen wrote:
> On Thu, Nov 15, 2018 at 10:00:02PM +0200, Jarkko Sakkinen wrote:
> > On Thu, Nov 08, 2018 at 04:46:03PM +0200, Jarkko Sakkinen wrote:
> > > On Wed, Nov 07, 2018 at 10:00:57AM -0800, Sean Christopherson wrote:
> > > > What do we gain by a single buffer vs. separate buffers?  The ioctl()
> > > > would be slightly smaller but it seems like the actual code would be
> > > > more complex.
> > > 
> > > I'm fine with either. It was just a suggestion.
> > > 
> > > > The enclave build process also utilizes the backing as temp storage
> > > > to avoid having to alloc kernel memory when queueing pages to be added
> > > > by the worker thread (which reminds me that I wanted to document why a
> > > > worker thread is used).  Keeping this behavior would effectively make
> > > > providing backing mandatory.
> > > 
> > > Would it be a problem just allocate those pages with alloc_page() and
> > > free them in the worker thread?
> > > 
> > > > Are there any potential complications with ENCLS consuming userspace
> > > > pointers?  We'd have to wrap them with user_access_{begin,end}() and
> > > > probably tweak the fixup, but I assume having the fixup handler means
> > > > we're generally ok?
> > > 
> > > Last time I did it I used get_user_pages() for pinning. I'm not sure
> > > why I should do anything but just re-use that.
> > 
> > What about VA page swapping? Not saying that it'd have to be done right
> > now but we need to answer whether it is enclave local or a global asset.
> > If it is local it would also require an argument.
> > 
> > I will most likely won't fix this for v17 because this detail needs
> > careful consideration.
> 
> I wonder if you can map shmem file to process address space so that you
> get it accounted for the process? That would be optimal for us. This way
> this won't become an API issue.
> 
> Yeah, as I started to implement this I realized these issues with the
> API side that will arise. Even doing vm_mmap() in the kernel code would
> be better than taking addresses through the ioctl. That is another
> option.

This is how strongly think. I can go with VMAs for swapping *but* if
that is the route I would recommend using vm_mmap() instead of taking
pointer from user space. This way the way swapping is done can be always
changed. It will be a huge lock-in to do it otherwise.

/Jarkko