On Tue, Aug 20, 2024 at 03:59:21PM +0100, Catalin Marinas wrote:

On Mon, Aug 19, 2024 at 05:33:24PM +0100, Mark Brown wrote:

quoted

On Mon, Aug 19, 2024 at 10:10:36AM +0100, Catalin Marinas wrote:

quoted

quoted

At a quick look, do_mmap() seems to always set VM_MAYEXEC but discard it
for non-executable file mmap. Last time I looked (when doing MTE) there
wasn't a way for the arch code to clear specific VM_* flags, only to
validate them. But I think we should just clear VM_MAYEXEC and also
return an error for VM_EXEC in the core do_mmap() if VM_SHADOW_STACK. It
would cover the other architectures doing shadow stacks.

quoted

Yes, I think adding something generic would make sense here.  That feels
like a cleanup which could be split out?

It can be done separately. It doesn't look like x86 has such checks.
Adding it generically would be a slight ABI tightening but I doubt it
matters, no sane software would use an executable shadow stack.

OK.

quoted

quoted

Is there any arch restriction with setting BTI and GCS? It doesn't make
sense but curious if it matters. We block the exec permission anyway
(unless the BTI pages moved to PIE as well, I don't remember).

quoted

As you say BTI should be meaningless for a non-executable page like GCS,
I'm not aware of any way in which it matters.  BTI is separate to PIE.

My thoughts were whether we can get rid of this hunk entirely by
handling it in the core code. We'd allow BTI if one wants such useless
combination but clear VM_MAYEXEC in the core code (and ignore VM_SHARED
since you can't set it anyway).

I have to admit that the BTI because I was shoving _EXEC in there rather
than because it specifically needed to be blocked.  So change the check
for VM_SHARED to a VM_WARN_ON(), and leave the _EXEC check for now
pending the above core change?