SMAP bypass

Summary

SMAP is a security feature on x86 CPUs, that forbids ring0 from reading/writing to ring3 pages, making it harder to exploit entire classes of vulnerabilities.

There is a vulnerability in FreeBSD 12 that allows SMAP to be bypassed by userland. There is a very high probability that it affects the PS5 but I was unable to access a PS5 firmware to confirm it.

This vuln downgrades the security properties of the OS, and is a building block for exploitation chains.

Details

With SMAP enabled, when %RFLAGS.AC is cleared the kernel will page-fault if it tries to access a page marked as "user page". When %RFLAGS.AC is set the kernel can access the user pages as if SMAP was not enabled.

In the FreeBSD kernel, a few functions exist that temporarily set %RFLAGS.AC in order to access user pages: the copyin() and copyout() functions.

These functions are used in all syscalls and are the only ways the kernel can copy data from/to userland.

These functions handle faults gracefully, that is, if userland passes an unmapped address and the kernel tries to copy data from it, the functions will simply return an error without kernel panic.

There is a bug in the fault handling of these functions. Typically copyin() is implemented as follows:

.macro	COPYIN smap erms
	/* ... */
	movq	$copy_fault,PCB_ONFAULT(%r11)
	/* ... */
	stac // set %RFLAGS.AC, to allow access to user pages
	do_the_copyin
	clac // clear %RFLAGS.AC, to forbid access to user pages
	/* ... */

copy_fault:
	movq	$0,PCB_ONFAULT(%r11)
	movl	$EFAULT,%eax
	POP_FRAME_POINTER
	ret

void
trap(struct trapframe *frame)
{
	/* ... */
			if (curpcb->pcb_onfault != NULL) {
				frame->tf_rip = (long)curpcb->pcb_onfault;
				return;
			}
	/* ... */
}

The fault handler copy_fault is registered in pcb_onfault at the beginning, then %RFLAGS.AC is set, the copy is made, and %RFLAGS.AC is cleared back.

If the copy faults for whatever reason, an exception is raised, the trap() handler sees that pcb_onfault has a pointer registered, and simply IRETs back to the pointer that was registered.

The problem is, the fault handler copy_fault does not clear %RFLAGS.AC, meaning that it remains set after copyin()/copyout() returns. The rest of the syscall will therefore execute with SMAP effectively disabled, until the kernel returns to userland where the SMAP state gets reset back to normal.

This SMAP disablement survives context switches, so a user that disables SMAP during one of his syscalls can also disable SMAP in other user/kernel threads if a rescheduling happens after/during the syscall (taking a mutex for example).

Fix

Add a clac instruction in copy_fault to clear %RFLAGS.AC:

 copy_fault:
+	clac
 	movq	$0,PCB_ONFAULT(%r11)
 	movl	$EFAULT,%eax
 	POP_FRAME_POINTER
 	ret

Impact

Userland can open lage windows where the kernel executes with SMAP disabled.

Lack of SMAP makes exploitation of common vulnerabilities easy/trivial.