Re: [PATCH v21 16/28] x86/sgx: Add the Linux SGX Enclave Driver

[Sean Christopherson]

On Fri, Aug 09, 2019 at 06:02:08PM +0300, Jarkko Sakkinen wrote:
> On Thu, 2019-08-08 at 08:40 -0700, Sean Christopherson wrote:
> > On Wed, Aug 07, 2019 at 06:15:34PM +0300, Jarkko Sakkinen wrote:
> > > On Mon, Jul 29, 2019 at 11:17:57AM +0000, Ayoun, Serge wrote:
> > > > > +	/* TCS pages need to be RW in the PTEs, but can be 0 in the EPCM. */
> > > > > +	if ((secinfo.flags & SGX_SECINFO_PAGE_TYPE_MASK) ==
> > > > > SGX_SECINFO_TCS)
> > > > > +		prot |= PROT_READ | PROT_WRITE;
> > > > 
> > > > For TCS pages you add both RD and WR maximum protection bits.
> > > > For the enclave to be able to run, user mode will have to change the
> > > > "vma->vm_flags" from PROT_NONE to PROT_READ | PROT_WRITE (otherwise
> > > > eenter fails).  This is exactly what your selftest  does.
> > > 
> > > Recap where the TCS requirements came from? Why does it need
> > > RW in PTEs and can be 0 in the EPCM? The comment should explain
> > > it rather leave it as a claim IMHO.
> > 
> > Hardware ignores SECINFO.FLAGS.{R,W,X} coming from userspace and instead
> > forces RWX=0.  It does this to prevent software from directly accessing
> > the TCS.  But hardware still accesses the TCS through a virtual address,
> > e.g. to allow software to zap the page for reclaim, which means hardware
> > generates reads and writes to the TCS, i.e. the PTEs need RW permissions.
> 
> Manipulating a PTE should not require any specific permissions on the
> page that it is defining. Why RW is required in SGX context?

By PTEs I meant the TCS needs to be mapped RW in the kernel's page tables,
e.g. hardware generates read and write accesses to the TCS when entering
an enclave.

> > So, for the EADD ioctl(), it's not unreasonable for userspace to provide
> > SECINFO.FLAGS.{R,W,X} = 0 for the TCS to match what will actually get
> > jammed into the EPCM.  Allowing userspace to specify RWX=0 means the
> > kernel needs to manually add PROT_READ and PROT_WRITE to the allowed prot
> > bits so that mmap()/mprotect() work as expected.
> 
> Anyway, appreciate your throughtout explanation, thanks.
> 
> /Jarkko
>