Re: [RFC PATCH 9/9] security/selinux: Add enclave_load() implementation

From: Sean Christopherson
Date: Mon Jun 03 2019 - 11:54:07 EST


On Mon, Jun 03, 2019 at 11:01:32AM -0400, Stephen Smalley wrote:
> On 5/31/19 7:31 PM, Sean Christopherson wrote:
> >The goal of selinux_enclave_load() is to provide a facsimile of the
> >existing selinux_file_mprotect() and file_map_prot_check() policies,
> >but tailored to the unique properties of SGX.
> >
> >For example, an enclave page is technically backed by a MAP_SHARED file,
> >but the "file" is essentially shared memory that is never persisted
> >anywhere and also requires execute permissions (for some pages).
> >
> >The basic concept is to require appropriate execute permissions on the
> >source of the enclave for pages that are requesting PROT_EXEC, e.g. if
> >an enclave page is being loaded from a regular file, require
> >FILE__EXECUTE and/or FILE__EXECMOND, and if it's coming from an
> >anonymous/private mapping, require PROCESS__EXECMEM since the process
> >is essentially executing from the mapping, albeit in a roundabout way.
> >
> >Note, FILE__READ and FILE__WRITE are intentionally not required even if
> >the source page is backed by a regular file. Writes to the enclave page
> >are contained to the EPC, i.e. never hit the original file, and read
> >permissions have already been vetted (or the VMA doesn't have PROT_READ,
> >in which case loading the page into the enclave will fail).
> >
> >Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> >---
> > security/selinux/hooks.c | 85 ++++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 85 insertions(+)
> >
> >diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c
> >index 3ec702cf46ca..f436a055dda7 100644
> >--- a/security/selinux/hooks.c
> >+++ b/security/selinux/hooks.c
> >@@ -6726,6 +6726,87 @@ static void selinux_bpf_prog_free(struct bpf_prog_aux *aux)
> > }
> > #endif
> >+#ifdef CONFIG_INTEL_SGX
> >+int selinux_enclave_load(struct vm_area_struct *vma, unsigned long prot,
> >+ unsigned long *allowed_prot)
> >+{
> >+ const struct cred *cred = current_cred();
> >+ u32 sid = cred_sid(cred);
> >+ int rc;
> >+
> >+ /* SGX is supported only in 64-bit kernels. */
> >+ WARN_ON_ONCE(!default_noexec);
> >+
> >+ /*
> >+ * SGX is responsible for checking @prot vs @allowed_prot, and SELinux
> >+ * only cares about execute related permissions for enclaves.
> >+ */
> >+ if (!(*allowed_prot & PROT_EXEC))
> >+ return 0;
> >+
> >+ /*
> >+ * Loading an executable enclave page from a VMA that is not executable
> >+ * itself requires EXECUTE permissions on the source file, or if there
> >+ * is no regular source file, EXECMEM since the page is being loaded
> >+ * from a non-executable anonymous mapping.
> >+ */
> >+ if (!(vma->vm_flags & VM_EXEC)) {
> >+ if (vma->vm_file && !IS_PRIVATE(file_inode(vma->vm_file)))
> >+ rc = file_has_perm(cred, vma->vm_file, FILE__EXECUTE);
>
> We might need an EXECMOD check here as well if (vma->vm_file &&
> vma->anon_vma). The scenario would be that the host application mapped the
> file with PROT_WRITE, modified it, but haven't mapped it PROT_EXEC. Now the
> enclave loader requests PROT_EXEC without PROT_WRITE or allows it.
> FILE__EXECUTE is insufficient for this case.

Ya, agreed.

> >+ else
> >+ rc = avc_has_perm(&selinux_state,
> >+ sid, sid, SECCLASS_PROCESS,
> >+ PROCESS__EXECMEM, NULL);
>
> These calls will audit FILE__EXECUTE or PROCESS__EXECMEM denials even when
> userspace never asked for PROT_EXEC. Possibly we should use
> avc_has_perm_noaudit() and only call avc_audit() if (prot & PROT_EXEC)? And
> similarly introduce file_has_perm_noaudit() -> inode_has_perm_noaudit() ->
> avc_has_perm_noaudit() or inline here and switch to avc_has_perm_noaudit()
> throughout?

Hmm, I think we want to audit the denials even if userspace technically
hasn't requested PROT_EXEC, yet... The idea of @allowed_prot is for
userspace to declare its intent, i.e. for all intents and purposes,
userspace is asking for PROT_EXEC by declaring PROT_EXEC in @allowed_prot.

For SGX1 enclaves, where permissions are fixed at enclave build time,
declaring PROT_EXEC but not actually mapping the page as PROT_EXEC is a
userspace bug, and a fairly egregious one at that.

For SGX2 enclaves, where enclaves can convert RW->RX (among others), not
auditing would make it difficult for userspace to debug failures due to
the denial. E.g. a properly written application that declared PROT_EXEC
on a RW page with the intent of converting it to RX would eventually fail
at mprotect() due to PROT_EXEC being cleared. Without the audit, there
wouldn't be any indication that mprotect() failed with -EACCES due to LSM
restrictions.

> >+
> >+ /*
> >+ * Reject the load if the enclave *needs* the page to be
> >+ * executable, otherwise prevent it from becoming executable.
> >+ */
> >+ if (rc) {
> >+ if (prot & PROT_EXEC)
> >+ return rc;
> >+
> >+ *allowed_prot &= ~PROT_EXEC;
> >+ }
> >+ }
> >+
> >+ /*
> >+ * An enclave page that may do RW->RX or W+X requires EXECMOD (backed
> >+ * by a regular file) or EXECMEM (loaded from an anonymous mapping).
>
> At present EXECMEM is also triggered for W+X private file mappings, to allow
> denying W+X while permitting exceptions for W->X for text relocations.

Doh, missed that one. Thanks!

> >+ * Note, this hybrid EXECMOD and EXECMEM behavior is intentional and
> >+ * reflects the nature of enclaves and the EPC, e.g. EPC is effectively
> >+ * a non-persistent shared file, but each enclave is a private domain
> >+ * within that shared file, so delegate to the source of the enclave.
> >+ */
> >+ if ((*allowed_prot & PROT_EXEC) && (*allowed_prot & PROT_WRITE)) {
> >+ if (vma->vm_file && !IS_PRIVATE(file_inode(vma->vm_file)))
> >+ rc = file_has_perm(cred, vma->vm_file, FILE__EXECMOD);
> >+ else
> >+ rc = avc_has_perm(&selinux_state,
> >+ sid, sid, SECCLASS_PROCESS,
> >+ PROCESS__EXECMEM, NULL);
>
> Same issue wrt auditing here. Could also potentially skip the EXECMEM check
> this time if we performed it above (if so, then we must have passed it
> because *allowed_prot still had PROT_EXEC set).

Skipping the second EXECMEM check crossed my mind as well. I'll play with
the code to see if I can come up with clean way of avoiding multiple
EXECMEM (or EXECMOD) checks.

> >+ /*
> >+ * Clear ALLOW_EXEC instead of ALLOWED_WRITE if permissions are
> >+ * lacking and @prot has neither PROT_WRITE or PROT_EXEC. If
> >+ * userspace wanted RX they would have requested RX, and due to
> >+ * lack of permissions they can never get RW->RX, i.e. the only
> >+ * useful transition is R->RW.
> >+ */
> >+ if (rc) {
> >+ if ((prot & PROT_EXEC) && (prot & PROT_WRITE))
> >+ return rc;
> >+
> >+ if (prot & PROT_EXEC)
> >+ *allowed_prot &= ~PROT_WRITE;
> >+ else
> >+ *allowed_prot &= ~PROT_EXEC;
> >+ }
> >+ }
> >+
> >+ return 0;
> >+}
> >+#endif
> >+
> > struct lsm_blob_sizes selinux_blob_sizes __lsm_ro_after_init = {
> > .lbs_cred = sizeof(struct task_security_struct),
> > .lbs_file = sizeof(struct file_security_struct),
> >@@ -6968,6 +7049,10 @@ static struct security_hook_list selinux_hooks[] __lsm_ro_after_init = {
> > LSM_HOOK_INIT(bpf_map_free_security, selinux_bpf_map_free),
> > LSM_HOOK_INIT(bpf_prog_free_security, selinux_bpf_prog_free),
> > #endif
> >+
> >+#ifdef CONFIG_INTEL_SGX
> >+ LSM_HOOK_INIT(enclave_load, selinux_enclave_load),
> >+#endif
> > };
> > static __init int selinux_init(void)
> >
>