Re: [PATCH] x86/sgx: Add SGX_PAGE_REPEAT flag for SGX_IOC_ENCLAVE_ADD_PAGES
From: Jarkko Sakkinen
Date: Mon Jun 14 2021 - 16:48:31 EST
On Thu, Jun 10, 2021 at 07:11:53AM -0700, Dave Hansen wrote:
> On 6/10/21 2:01 AM, Jarkko Sakkinen wrote:
> > On Thu, Jun 10, 2021 at 10:21:17AM +0300, Jarkko Sakkinen wrote:
> >> For uninitialized data, there's a need to add the same page multiple times,
> >> e.g. a zero page, instead of traversing the source memory forward. With the
> >> current API, this requires to call SGX_IOC_ENCLAVE_ADD_PAGES multiple
> >> times, once per page, which is not very efficient.
> >> Add a new SGX_PAGE_REPEAT flag to resolve the issue. When this flag is set
> >> to the 'flags' field of struct sgx_enclave_pages, the ioctl will apply the
> >> page at 'src' multiple times, instead of moving forward in the address
> >> space.
> >> Signed-off-by: Jarkko Sakkinen <jarkko@xxxxxxxxxx>
> > After sending this, I started to think that maybe it would actually better
> > to just add SGX_PAGE_ZERO flag, i.e. add zero pages and ignore src. That's
> > the main use case right now, and saves the user space from extra trouble of
> > having to do such page by hand.
> > That neither does prevent adding SGX_PAGE_REPEAT later on. I just see no
> > point of that generic functionality right now. It only makes simple use
> > case more complex.
> Is the main argument behind this new ABI that it increases efficiency?
> Let's say I want to add 1MB of 0'd pages to an enclave. Won't this do it?
> zeros = mmap(NULL, size, PROT_READ, MAP_ANONYMOUS|MAP_PRIVATE,
> -1, 0);
> ioctl(SGX_IOC_ENCLAVE_ADD_PAGES, zeros, size);
Later, with LSM's, it's better to mmap /dev/null.
> Sure, you'll pay the cost of faulting in the zero page size/PAGE_SIZE
> times. But, that's pretty minuscule. This zeros buffer can also be
> reused without faulting again. It can be as big or small as you want.
> Heck, it could even be 2MB in size and use the transparent huge page.
> I agree that there's definitely some optimization work to do. But, I'm
> a bit hesitant to turn to new ABI to do it.
I realized that it's good to create heap memory like in your example
because then any possible SGX_IOC_ENCLAVE_ADD_PAGES invocation map to LSM
rules, i.e. every call has a well-defined source.