Message-ID: <2c38deba-47c1-2800-1215-a3a12cc06316@fortanix.com>
Date:   Thu, 8 Apr 2021 17:27:59 +0200
From:   Jethro Beekman <jethro@...tanix.com>
To:     Dave Hansen <dave.hansen@...el.com>,
        Raoul Strackx <raoul.strackx@...tanix.com>,
        Jarkko Sakkinen <jarkko@...nel.org>,
        Dave Hansen <dave.hansen@...ux.intel.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...hat.com>, Borislav Petkov <bp@...en8.de>,
        x86@...nel.org, "H. Peter Anvin" <hpa@...or.com>,
        linux-sgx@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH RESEND 0/3] x86/sgx: eextend ioctl

On 2021-04-02 22:48, Dave Hansen wrote:
> On 4/2/21 1:20 PM, Jethro Beekman wrote:
>> On 2021-04-02 21:50, Dave Hansen wrote:
>>> Again, how does this save space?
>>>
>>> Are you literally talking about the temporary cost of allocating *one* page?
>>
>> No I'm talking about the amount of disk space/network traffic needed
>> to distribute the application.
> 
> Am I just horribly confused about how executable formats work?
> 
> Executables go through an SGX loader that copies them into SGX memory
> with the kernel's help.
> 
> That executable can have *ANY* format, really.
> 
> Then, a loader needs to read that format and turn it into data that can
> be shoved into the kernel.

Sure, you can define any compression format or conversion rules between executable formats. But the native “executable format” for SGX is very clearly defined in the Intel SDM as a specific sequence of ECREATE, EADD, EEXTEND and EINIT calls. It's that sequence that's used for loading the enclave and it's that sequence that's used for code signing. So when I say save space I mean save space in the native format.

Not EEXTENDing unnecessarily also reduces enclave load time if you're looking for additional arguments.

> The simplest way to do this is to just
> mmap() the executable and point the kernel ioctl()'s at the executable's
> pages one-by-one.
> 
> The other way a loader *could* work is to copy the data to a temporary
> location and then hand the temporary location to the SGX ioctl()s.
> 
> Let's say the kernel *REQUIRED* page-aligned and page-sized ioctl()
> arguments forever.  If an executable had a 256-byte-sized chunk of data,
> all the loader would have to do is allocate a page, copy the data in
> there, and then pass that whole page into the ioctl().
> 
> In other words, the loading restrictions (page alignment) have little to
> nothing to do with the format of the thing loading the executable.
> 
> But, in no way does a kernel page-size-based ABI *REQUIRE* that an
> underlying binary format represent things only in page-sized chunks.
> Look at how many page-sized executables there are in /bin.  Yet, they
> can only be mapped into the address space in PAGE_SIZE increments.
> 
>>>>> Does any actual, real-world enclave want this functionality?  Why?
>>>
>>> I didn't see an answer on this one.
>>
>> Yes, we have enclaves that use this functionality. They already exist
>> so they can't be changed (without changing the measurement) and we'd
>> like to stop using the out of tree driver as soon as possible.
>> However, we are not able to load the enclaves.
> OK, so please give this series another shot.  Please explain why you
> *ACTUALLY* need it and what the goals are.  Please explain why you can't
> just relax the restrictions of the existing add ioctl() to take
> <PAGE_SIZE arguments.
> 
> As far as I can tell, there are only two coherent arguments for this
> functionality:
> 1. It makes the loader simpler so that it doesn't need temporary pages
> 2. It would allow old enclaves created with non-upstream-Linux SGX
>    implementations to end up with the same signatures on these
>    implementations as upstream Linux.
> 
> I find both of those pretty weak arguments.  Doing #2 just for the
> out-of-tree Linux implementation also encourages folks to establish ABI
> out of the tree and then foist it on upstream later.  That's not super cool.
> But, I guess this would be nice to the folks that have gone to the
> trouble of building SGX enclaves for all these years with no upstream
> support.

Linux doesn't exist in a vacuum. People who write SGX applications write those applications for SGX, not for Linux. For Linux to claim to support SGX version 1, it should support all SGX version 1 applications. The current implementation is not SGX version 1, it's some limited subset of it.

The cost to Linux for supporting all of SGX version 1 seems excessively low.

SGX defines lots of things and you may not see the use case for all of this immediately. No one has a usecase for creating enclaves with SECS.SSAFRAMESIZE = 1000 or TCS.NSSA = 3. Why did you not demand checks for this in the ECREATE/EADD ioctls?

> 
> I'll try to look at it with fresh eyes once this is in place.
> 

--
Jethro Beekman | Fortanix


Download attachment "smime.p7s" of type "application/pkcs7-signature" (4490 bytes)

Powered by blists - more mailing lists