Re: [PATCH RFC] x86: Add SGX_IOC_ENCLAVE_AUGMENT_PAGES

[Jarkko Sakkinen]

On Sat, Mar 05, 2022 at 04:32:12AM +0200, Jarkko Sakkinen wrote:
> On Sat, Mar 05, 2022 at 04:17:53AM +0200, Jarkko Sakkinen wrote:
> > On Fri, Mar 04, 2022 at 09:08:20AM -0800, Dave Hansen wrote:
> > > On 3/4/22 04:28, Jarkko Sakkinen wrote:
> > > > Explicit EAUG ioctl is a better choice than an implicit EAUG from a page
> > > > fault handler because it allows to have O(1) number of kernel-enclave round
> > > > trips for EAUG-EACCEPT{COPY} process, instead of O(n), as it is in the case
> > > > when a page fault handler EAUG single page at a time.
> > > 
> > > So this is basically an optimization?  It's MADV_WILLNEED or
> > > MAP_POPULATE to the cost of avoid future faults?
> > 
> > Yes.
> > 
> > So the idea would be that based on these the #PF handler would have more
> > smartness, and it would do a batch of EAUG's?
> > 
> > That could be possibly acceptable but I also had other concern.
> > 
> > I would like to see this:
> > 
> > 1. Removal of vm_run_prot_bits.
> > 2. Use RWX vm_max_prot_bits for EAUG'd pages.
> > 
> > During run-time kernel controls PTE's, and enclave has full control of the
> > EPCM (EACCEPT, EACCEPTCOPY, EMODPE). By creating artificial limitations how
> > to operate with these, it can limit various optimizations in the user space
> > code. E.g. a syscall shim can require clever co-operation between in-enclave
> > opcodes and what you do with the kernel in various situations.
> > 
> > RWX sounds provocative yes, but here it means only the limits where kernel
> > can set its PTE's and nothing else, not that page table is filled with RWX
> > pages, and enclave dictates what is in EPCM, and that's how it actually
> > should be (e.g. you can sometimes deliver mmap() without ever going out
> > of the enclave with EMODPE).
> > 
> > If MADV_WILLNEED/MAP_POPULATE approach is combined with this what I
> > discussed here, then I think we could have solution to write an efficient
> > memory management shims.
> 
> Do you already have a rough idea what needs to be done? I can take anyway a
> look but just in case you had processed this further, please tell what you
> have.

MAP_POPULATE would almost exact match to the ioctl, so I guess that could
really work. I can give this a shot.

Does MAP_POPULATE "come through" the driver's mmap callback?

BR, Jarkko