On Mon, Mar 07, 2022 at 03:39:01PM -0800, Reinette Chatre wrote: > Hi Jarkko, > > On 3/7/2022 9:22 AM, Jarkko Sakkinen wrote: > > On Mon, Mar 07, 2022 at 09:13:48AM -0800, Reinette Chatre wrote: > >> Hi Jarkko, > >> > >> On 3/7/2022 8:09 AM, Jarkko Sakkinen wrote: > >>> On Mon, Mar 07, 2022 at 06:02:03PM +0200, Jarkko Sakkinen wrote: > >>>> On Mon, Mar 07, 2022 at 05:35:04PM +0200, Jarkko Sakkinen wrote: > >>>>> vm_max_permissions was created to control the pre-initialization content > >>>>> that contributes to MRSIGNATURE. It was never meant to be as a limit to > >>>>> dynamically added pages. > >>>>> > >>>>> E.g. static content could be used as a hook for LSM's to decide whether > >>>>> certain signature is qualified for EINIT. Dynamic content has nothing to > >>>>> do with that. The current mechanisms only add to the complexity on how > >>>>> to control PTE and EPCM permissions, and do not add anything else than > >>>>> obfuscity to security side of things. > >> > >> Linux has mechanisms to enforce what can be executed. For example, with SELinux > >> a process can be required to have PROCESS__EXECHEAP or PROCESS__EXECSTACK > >> before it can be allowed to execute writable memory. > >> > >> A few SGX runtimes enables unmodified executables to be run within SGX enclaves. > >> > >> Does a change like this not enable executables prevented by existing > >> security mechanisms to circumvent such restrictions by running within > >> a SGX enclave? > > > > It does not open any extra exposure as the existing policies apply for > > the enclave content created before initialization. > > > > And I'm not sure what kind of circumvention scenario we are talking > > about. > > > >>>>> Thus add PROT_EXEC to the permissions assigned by the #PF handler. > >>>>> > >>>>> Signed-off-by: Jarkko Sakkinen <jarkko@xxxxxxxxxx> > >>>>> --- > >>>>> arch/x86/kernel/cpu/sgx/encl.c | 9 ++++----- > >>>>> 1 file changed, 4 insertions(+), 5 deletions(-) > >>>>> > >>>>> diff --git a/arch/x86/kernel/cpu/sgx/encl.c b/arch/x86/kernel/cpu/sgx/encl.c > >>>>> index 79e39bd99c09..0256918b2c2f 100644 > >>>>> --- a/arch/x86/kernel/cpu/sgx/encl.c > >>>>> +++ b/arch/x86/kernel/cpu/sgx/encl.c > >>>>> @@ -160,12 +160,11 @@ static vm_fault_t sgx_encl_eaug_page(struct vm_area_struct *vma, > >>>>> encl_page->encl = encl; > >>>>> > >>>>> /* > >>>>> - * Adding a regular page that is architecturally allowed to only > >>>>> - * be created with RW permissions. > >>>>> - * TBD: Interface with user space policy to support max permissions > >>>>> - * of RWX. > >>>>> + * Dynamic pages do not contribute to MRSIGNATURE, i.e. they are > >>>>> + * controlled only by PTE and EPCM permissions. Thus, the no limit > >>>>> + * is set here. > >>>>> */ > >>>>> - prot = PROT_READ | PROT_WRITE; > >>>>> + prot = PROT_READ | PROT_WRITE | PROT_EXEC; > >>>>> encl_page->vm_run_prot_bits = calc_vm_prot_bits(prot, 0); > >>>>> encl_page->vm_max_prot_bits = encl_page->vm_run_prot_bits; > >>>>> > >>>>> -- > >>>>> 2.35.1 > >>>>> > >>>> > >>>> This is really a show stopper. I think here's a logical mistake on for what > >>>> purpose vm_max_prot_bits are used for. They are meant for the static and > >>>> also signed content of the enclave. > >>>> > >>>> These changes in the patch set that are related to vm_max_prot_bits only > >>>> messes up what already exists, and make incredibly hard to implement > >>>> anything decent on top of SGX2 features. > >>> > >>> I.e. once signed content has passed EINIT ioctl, and whatever checks > >>> there are now or in future (e.g. LSM hooks), the system has accepted > >>> the enclave behaviour, and it includes also the use of EACCEPT opcode. > >>> > >>> It's the exec or no-exec decision point. The thing that these patches > >>> do is making an obfuscated mess of all this. When EINIT has passed, > >>> it has been decided that the enclave can do its workload. Let's not > >>> throw stick in front of it, and make everyones life misserable. > >> > >> A common use for these dynamically added pages is to increase the heap > >> and stack. Always allowing PTEs of RWX on these pages irrespective > >> whether it will be used for heap, stack, or relocatable code does > >> not match with how the kernel manages protections. > >> > >> As I said before I am not comfortable with such a change and cannot > >> sign off on this. I would defer to the maintainers to choose the > >> direction. > >> > >> Reinette > > > > My choice is to not use this existing mechanism for dynamically created > > pages because otherwise the implementation overally is just crippled. > > > > Something unusable is for sure as secure as you can get. > > > > ok, I will proceed with your guidance here and include your snippet from > this patch into the next version. Thank you. > One question, regarding "MRSIGNATURE" - did you perhaps mean "MRENCLAVE"? It could either both define "static root of trust". Does not really all that much which one you use in the comment. > Reinette BR, Jarkko
