Date:   Sat, 14 Mar 2020 10:10:26 -0400
From:   Nathaniel McCallum <npmccallum@...hat.com>
To:     Sean Christopherson <sean.j.christopherson@...el.com>
Cc:     Jethro Beekman <jethro@...tanix.com>,
        Jarkko Sakkinen <jarkko.sakkinen@...ux.intel.com>,
        linux-kernel@...r.kernel.org, x86@...nel.org,
        linux-sgx@...r.kernel.org, akpm@...ux-foundation.org,
        dave.hansen@...el.com, Neil Horman <nhorman@...hat.com>,
        "Huang, Haitao" <haitao.huang@...el.com>,
        andriy.shevchenko@...ux.intel.com, tglx@...utronix.de,
        "Svahn, Kai" <kai.svahn@...el.com>, bp@...en8.de,
        Josh Triplett <josh@...htriplett.org>, luto@...nel.org,
        kai.huang@...el.com, David Rientjes <rientjes@...gle.com>,
        cedric.xing@...el.com, Patrick Uiterwijk <puiterwijk@...hat.com>,
        Andy Lutomirski <luto@...capital.net>,
        Connor Kuehl <ckuehl@...hat.com>,
        Harald Hoyer <harald@...hat.com>,
        Lily Sturmann <lsturman@...hat.com>
Subject: Re: [PATCH v28 21/22] x86/vdso: Implement a vDSO for Intel SGX
 enclave call

On Fri, Mar 13, 2020 at 6:08 PM Sean Christopherson
<sean.j.christopherson@...el.com> wrote:
>
> On Fri, Mar 13, 2020 at 04:14:01PM -0400, Nathaniel McCallum wrote:
> > On Fri, Mar 13, 2020 at 2:45 PM Sean Christopherson
> > <sean.j.christopherson@...el.com> wrote:
>  > > This doesn't compromise the ability to treat __vsdo...() like ENCLU if
> > > > you need the full power. But it does make it significantly easier to
> > > > consume when you don't have special needs. So as I see it, __vdso...()
> > > > should:
> > > >
> > > > 1. preserve %rbx
> > > > 2. take leaf in %rcx
> > > > 3. gain a void* stack param which is passed to the handler
> > >
> > > Unless I'm misunderstanding the request, this already exists.  %rsp at the
> > > time of EEXIT is passed to the handler.
> >
> > Sorry, different stack parameter. I mean this:
> >
> > typedef int (*sgx_enclave_exit_handler_t)(
> >     long rdi,
> >     long rsi,
> >     long rdx,
> >     long ursp,
> >     long r8,
> >     long r9,
> >     int ret,
> >     void *tcs,
> >     struct sgx_enclave_exception *e,
> >     void *misc
> > );
> >
> > int __vdso_sgx_enter_enclave(
> >     long rdi,
> >     long rsi,
> >     long rdx,
> >     int leaf,
> >     long r8,
> >     long r9,
> >     void *tcs,
> >     struct sgx_enclave_exception *e,
> >     void *misc,
> >     sgx_enclave_exit_handler_t handler
> > );
> >
> > This is so that the caller of __vdso...() can pass context into the
> > handler.
>
> Hrm, I'm not a fan of adding a param that is only consumed by the handler,
> especially when there are multiple alternatives, e.g. fudge the param in
> assembly before calling __vdso(), have the enclave supply the context in a
> volatile register, etc...

Yes, but all of those require assembly. The whole point of this is
ergonomics without assembly. Once you can call __vdso() without
assembly, having to resort to assembly to make it useful will feel
painful. I imagine it would be pretty common to pass something like a
jmp_buf reference or a reference to a struct for collecting output
stack arguments through misc.

> > Note that I've also reordered the stack parameters so that the stack
> > order can be reused.
>
> Ah, ret<->tcs, took me a minute...
>
> Does preserving tsc->e->misc ordering matter all that much?

Not really. I was just trying to aid the reader of the assembly. If
there are more important concerns, fine.

>  __vdso() needs
> to manually copy them either way.  I ask because putting @misc at the end
> would allow implementations that don't use @handler to omit the param (if
> I've done my math correctly, which is always a big if).  That would make
> adding the handler-only param a little more palatable.

Fine by me.

> > > > 4. sub/add to %rsp rather than save/restore
> > >
> > > Can you elaborate on why you want to sub/add to %rsp instead of having the
> > > enclave unwind the stack?  Preserving %rsp across EEXIT/ERESUME seems more
> > > in line with function call semantics, which I assume is desirable?  E.g.
> > >
> > >   push param3
> > >   push param2
> > >   push param1
> > >
> > >   enclu[EEXIT]
> > >
> > >   add $0x18, %rsp
> >
> > Before enclave EEXIT, the enclave restores %rsp to the value it had
> > before EENTER was called. Then it pushes additional output arguments
> > onto the stack. The enclave calls EENCLU[EEXIT].
> >
> > We are now in __vdso...() on the way back to the caller. However, %rsp
> > has a different value than we entered the function with. This breaks
> > x86_64 ABI, obviously. The handler needs to fix this up: how does it
> > do so?
> >
> > In the current code, __vdso..() saves the value of %rsp, calls the
> > handler and then restores %rsp. The handler can fix up the stack by
> > setting the correct value to %rbx and returning without restoring it.
>
> Ah, you're referring to the patch where the handler decides to return all
> the way back to the caller of __vdso...().
>
> > But this requires internal knowledge of the __vdso...() function,
> > which could theoretically change in the future.
> >
> > If instead the __vdso...() only did add/sub, then the handler could do:
> > 1. pop return address
> > 2. pop handler stack params
> > 3. pop enclave additional output stack params
> > 4. push handler stack params
> > 5. push return address
> >
> > While this is more work, it is standard calling convention work that
> > doesn't require internal knowledge of __vdso..(). Alternatively, if we
> > don't like the extra work, we can document the %rbx hack explicitly
> > into the handler documentation and make it part of the interface. But
> > we need some explicit way for the handler to pop enclave output stack
> > params that doesn't depend on internal knowledge of the __vdso...()
> > invariants.
>
> IIUC, this is what you're suggesting?  Having to align the stack makes this
> a bit annoying, but it's not bad by any means.
>
> diff --git a/arch/x86/entry/vdso/vsgx_enter_enclave.S b/arch/x86/entry/vdso/vsgx_enter_enclave.S
> index 94a8e5f99961..05d54f79b557 100644
> --- a/arch/x86/entry/vdso/vsgx_enter_enclave.S
> +++ b/arch/x86/entry/vdso/vsgx_enter_enclave.S
> @@ -139,8 +139,9 @@ SYM_FUNC_START(__vdso_sgx_enter_enclave)
>         /* Pass the untrusted RSP (at exit) to the callback via %rcx. */
>         mov     %rsp, %rcx
>
> -       /* Save the untrusted RSP in %rbx (non-volatile register). */
> +       /* Save the untrusted RSP offset in %rbx (non-volatile register). */
>         mov     %rsp, %rbx
> +       and     $0xf, %rbx
>
>         /*
>          * Align stack per x86_64 ABI. Note, %rsp needs to be 16-byte aligned
> @@ -161,8 +162,8 @@ SYM_FUNC_START(__vdso_sgx_enter_enclave)
>         mov     0x20(%rbp), %rax
>         call    .Lretpoline
>
> -       /* Restore %rsp to its post-exit value. */
> -       mov     %rbx, %rsp
> +       /* Undo the post-exit %rsp adjustment. */
> +       lea     0x20(%rsp,%rbx), %rsp
>
>
> That's reasonable, let's the handler play more games with minimal overhead.

Yes, exactly!

> > > > That would make this a very usable and fast interface without
> > > > sacrificing any of its current power.
> > >
> >
>

Powered by blists - more mailing lists