Re: [PATCH V2 16/32] x86/sgx: Support restricting of enclave page permissions

[Jarkko Sakkinen <jarkko@kernel.org>]

On Thu, Mar 17, 2022 at 11:50:41PM +0200, Jarkko Sakkinen wrote:
> On Thu, Mar 17, 2022 at 09:28:45AM -0500, Haitao Huang wrote:
> > Hi
> > 
> > On Thu, 17 Mar 2022 02:11:28 -0500, Jarkko Sakkinen <jarkko@kernel.org>
> > wrote:
> > 
> > > On Thu, Mar 17, 2022 at 09:01:07AM +0200, Jarkko Sakkinen wrote:
> > > > On Mon, Mar 14, 2022 at 10:39:36AM -0500, Haitao Huang wrote:
> > > > > Hi Jarkko
> > > > >
> > > > > On Sun, 13 Mar 2022 21:58:51 -0500, Jarkko Sakkinen
> > > > <jarkko@kernel.org>
> > > > > wrote:
> > > > >
> > > > > > On Mon, Mar 14, 2022 at 04:50:56AM +0200, Jarkko Sakkinen wrote:
> > > > > > > On Mon, Mar 14, 2022 at 04:49:37AM +0200, Jarkko Sakkinen wrote:
> > > > > > > > On Fri, Mar 11, 2022 at 09:53:29AM -0800, Reinette Chatre wrote:
> > > > > > > >
> > > > > > > > > I saw Haitao's note that EMODPE requires "Read access
> > > > permitted
> > > > > > > by enclave".
> > > > > > > > > This motivates that EMODPR->PROT_NONE should not be allowed
> > > > > > > since it would
> > > > > > > > > not be possible to relax permissions (run EMODPE) after that.
> > > > > > > Even so, I
> > > > > > > > > also found in the SDM that EACCEPT has the note "Read access
> > > > > > > permitted
> > > > > > > > > by enclave". That seems to indicate that EMODPR->PROT_NONE is
> > > > > > > not practical
> > > > > > > > > from that perspective either since the enclave will not be
> > > > able to
> > > > > > > > > EACCEPT the change. Does that match your understanding?
> > > > > > > >
> > > > > > > > Yes, PROT_NONE should not be allowed.
> > > > > > > >
> > > > > > > > This is however the real problem.
> > > > > > > >
> > > > > > > > The current kernel patch set has inconsistent API and EMODPR
> > > > ioctl is
> > > > > > > > simply unacceptable. It  also requires more concurrency
> > > > management
> > > > > > > from
> > > > > > > > user space run-time, which would be heck a lot easier to do
> > > > in the
> > > > > > > kernel.
> > > > > > > >
> > > > > > > > If you really want EMODPR as ioctl, then for consistencys sake,
> > > > > > > then EAUG
> > > > > > > > should be too. Like this when things go opposite directions,
> > > > this
> > > > > > > patch set
> > > > > > > > plain and simply will not work out.
> > > > > > > >
> > > > > > > > I would pick EAUG's strategy from these two as it requires half
> > > > > > > the back
> > > > > > > > calls to host from an enclave. I.e. please combine
> > > > mprotect() and
> > > > > > > EMODPR,
> > > > > > > > either in the #PF handler or as part of mprotect(), which ever
> > > > > > > suits you
> > > > > > > > best.
> > > > > > > >
> > > > > > > > I'll try demonstrate this with two examples.
> > > > > > > >
> > > > > > > > mmap() could go something like this() (simplified):
> > > > > > > > 1. Execution #UD's to SYSCALL.
> > > > > > > > 2. Host calls enclave's mmap() handler with mmap() parameters.
> > > > > > > > 3. Enclave up-calls host's mmap().
> > > > > > > > 4. Loops the range with EACCEPTCOPY.
> > > > > > > >
> > > > > > > > mprotect() has to be done like this:
> > > > > > > > 1. Execution #UD's to SYSCALL.
> > > > > > > > 2. Host calls enclave's mprotect() handler.
> > > > > > > > 3. Enclave up-calls host's mprotect().
> > > > > > > > 4. Enclave up-calls host's ioctl() to
> > > > SGX_IOC_ENCLAVE_PERMISSIONS.
> > > > >
> > > > > I assume up-calls here are ocalls as we call them in our
> > > > implementation,
> > > > > which are the calls enclave make to untrusted side via EEXIT.
> > > > >ar
> > > > > If so, can your implementation combine this two up-calls into one,
> > > > then host
> > > > > side just do ioctl() and mprotect to kernel? If so, would that
> > > > address your
> > > > > concern about extra up-calls?
> > > > >
> > > > >
> > > > > > > > 3. Loops the range with EACCEPT.
> > > > > > >   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > > > > >   5. Loops the range with EACCEPT + EMODPE.
> > > > > > >
> > > > > > > > This is just terrible IMHO. I hope these examples bring some
> > > > insight.
> > > > > >
> > > > > > E.g. in Enarx we have to add a special up-call (so called
> > > > enarxcall in
> > > > > > intermediate that we call sallyport, which provides shared buffer to
> > > > > > communicate with the enclave) just for reseting the range with
> > > > PROT_READ.
> > > > > > Feel very redundant, adds ugly cruft and is completely opposite
> > > > strategy
> > > > > > to
> > > > > > what you've chosen to do with EAUG, which is I think correct
> > > > choice as
> > > > > > far
> > > > > > as API is concerned.
> > > > >
> > > > > The problem with EMODPR on #PF is that kernel needs to know what
> > > > permissions
> > > > > requested from enclave at the time of #PF. So enclave has to make
> > > > at least
> > > > > one call to kernel (again via ocall in our case, I assume up-call
> > > > in your
> > > > > case) to make the change.
> > > > 
> > > > The #PF handler should do unconditionally EMODPR with PROT_READ.
> > > 
> > > Or mprotect(), as long as secinfo contains PROT_READ. I don't care about
> > > this detail hugely anymore because it does not affect uapi.
> > > 
> > > Using EMODPR as a permission control mechanism is a ridiculous idea, and
> > > I cannot commit to maintain a broken uapi.
> > > 
> > 
> > Jarkko, how would automatically forcing PROT_READ on #PF work for this
> > sequence?
> > 
> > 1) EAUG a page (has to be RW)
> > 2) EACCEPT(RW)
> > 3) enclave copies some data to page
> > 4) enclave wants to change permission to R
> > 
> > If you are proposing mprotect, then as I indicated earlier, please address
> > concerns raised by Reinette:
> > https://lore.kernel.org/linux-sgx/e1c04077-0165-c5ec-53be-7fd732965e80@intel.com/
> 
> For EAUG you can choose between #PF handler and having it as part of
> mmap() with the same uapi.
> 
> For EMODPR clearly #PF handler would be tricky but nothing prevents
> resetting the permissions as part of mprotect() flow, which is trivial.
> 
> One good reason to have a fixed EMODPR is that e.g. emulating properly
> mprotect() is almost undoable if you don't do it otherwise. Specifically

s/don't//g

> the scenario where your address range spans through multiple adjacent
> VMAs. It's even without EMODPR complex enough scenario that you really
> don't want to ask yourself for more trouble than use EMODPR in a super
> conservative manner.
> 
> Having EMODPR fully exposed will only make more difficult API to do with
> extra round-trips. If you want to use ring-0 instructions fully exposed,
> please don't use a kernel. There's a bunch of hardware features in Intel
> CPUs for which Linux does not provide 1:1 all wide open interfaces.
> 
> BR, Jarkko