Re: [PATCH v5 03/10] powerpc/signal64: Move non-inline functions out of setup_sigcontext()

On Sun Feb 7, 2021 at 10:44 PM CST, Daniel Axtens wrote:

Hi Chris,

These two paragraphs are a little confusing and they seem slightly
repetitive. But I get the general idea. Two specific comments below:

Umm... yeah only one of those was supposed to be sent. I will reword
this for the next spin and address the comment below about how it is
not entirely clear that the inline functions are being moved out.

quoted

There are non-inline functions which get called in setup_sigcontext() to
save register state to the thread struct. Move these functions into a
separate prepare_setup_sigcontext() function so that
setup_sigcontext() can be refactored later into an "unsafe" version
which assumes an open uaccess window. Non-inline functions should be
avoided when uaccess is open.

Why do we want to avoid non-inline functions? We came up with:

- we want KUAP protection for as much of the kernel as possible: each
extra bit of code run with the window open is another piece of attack
surface.
   
- non-inline functions default to traceable, which means we could end
up ftracing while uaccess is enabled. That's a pretty big hole in the
defences that KUAP provides.

I think we've also had problems with the window being opened or closed
unexpectedly by various bits of code? So the less code runs in uaccess
context the less likely that is to occur.

That is my understanding as well.

quoted

The majority of setup_sigcontext() can be refactored to execute in an
"unsafe" context (uaccess window is opened) except for some non-inline
functions. Move these out into a separate prepare_setup_sigcontext()
function which must be called first and before opening up a uaccess
window. A follow-up commit converts setup_sigcontext() to be "unsafe".

This was a bit confusing until we realise that you're moving the _calls_
to the non-inline functions out, not the non-inline functions
themselves.

quoted

Signed-off-by: Christopher M. Riedl <redacted>
---
 arch/powerpc/kernel/signal_64.c | 32 +++++++++++++++++++++-----------
 1 file changed, 21 insertions(+), 11 deletions(-)

diff --git a/arch/powerpc/kernel/signal_64.c b/arch/powerpc/kernel/signal_64.c
index f9e4a1ac440f..b211a8ea4f6e 100644
--- a/arch/powerpc/kernel/signal_64.c
+++ b/arch/powerpc/kernel/signal_64.c

@@ -79,6 +79,24 @@ static elf_vrreg_t __user *sigcontext_vmx_regs(struct sigcontext __user *sc)
 }
 #endif
 
+static void prepare_setup_sigcontext(struct task_struct *tsk, int ctx_has_vsx_region)

ctx_has_vsx_region should probably be a bool? Although setup_sigcontext
also has it as an int so I guess that's arguable, and maybe it's better
to stick with this for constency.

I've been told not to introduce unrelated changes in my patches before
so chose to keep this as an int for consistency.

quoted

+{
+#ifdef CONFIG_ALTIVEC
+	/* save altivec registers */
+	if (tsk->thread.used_vr)
+		flush_altivec_to_thread(tsk);
+	if (cpu_has_feature(CPU_FTR_ALTIVEC))
+		tsk->thread.vrsave = mfspr(SPRN_VRSAVE);
+#endif /* CONFIG_ALTIVEC */
+
+	flush_fp_to_thread(tsk);
+
+#ifdef CONFIG_VSX
+	if (tsk->thread.used_vsr && ctx_has_vsx_region)
+		flush_vsx_to_thread(tsk);
+#endif /* CONFIG_VSX */

Alternatively, given that this is the only use of ctx_has_vsx_region,
mpe suggested that perhaps we could drop it entirely and always
flush_vsx if used_vsr. The function is only ever called with either
`current` or wth ctx_has_vsx_region set to 1, so in either case I think
that's safe? I'm not sure if it would have performance implications.

I think that could work as long as we can guarantee that the context
passed to swapcontext will always be sufficiently sized if used_vsr,
which I think *has* to be the case?

Should we move this and the altivec ifdef to IS_ENABLED(CONFIG_VSX) etc?
I'm not sure if that runs into any problems with things like 'used_vsr'
only being defined if CONFIG_VSX is set, but I thought I'd ask.

That's why I didn't use IS_ENABLED(CONFIG_...) here - all of these
field (used_vr, vrsave, used_vsr) declarations are guarded by #ifdefs :/

quoted

+}
+
 /*
  * Set up the sigcontext for the signal frame.
  */

@@ -97,7 +115,6 @@ static long setup_sigcontext(struct sigcontext __user *sc,
 	 */
 #ifdef CONFIG_ALTIVEC
 	elf_vrreg_t __user *v_regs = sigcontext_vmx_regs(sc);
-	unsigned long vrsave;
 #endif
 	struct pt_regs *regs = tsk->thread.regs;
 	unsigned long msr = regs->msr;

@@ -112,7 +129,6 @@ static long setup_sigcontext(struct sigcontext __user *sc,
 
 	/* save altivec registers */
 	if (tsk->thread.used_vr) {
-		flush_altivec_to_thread(tsk);
 		/* Copy 33 vec registers (vr0..31 and vscr) to the stack */
 		err |= __copy_to_user(v_regs, &tsk->thread.vr_state,
 				      33 * sizeof(vector128));

@@ -124,17 +140,10 @@ static long setup_sigcontext(struct sigcontext __user *sc,
 	/* We always copy to/from vrsave, it's 0 if we don't have or don't
 	 * use altivec.
 	 */
-	vrsave = 0;
-	if (cpu_has_feature(CPU_FTR_ALTIVEC)) {
-		vrsave = mfspr(SPRN_VRSAVE);
-		tsk->thread.vrsave = vrsave;
-	}
-
-	err |= __put_user(vrsave, (u32 __user *)&v_regs[33]);
+	err |= __put_user(tsk->thread.vrsave, (u32 __user *)&v_regs[33]);

Previously, if !cpu_has_feature(ALTIVEC), v_regs[33] had vrsave stored,
which was set to 0 explicitly. Now we store thread.vrsave instead of the
local vrsave. That should be safe - it is initalised to 0 elsewhere.

So you don't have to do anything here, this is just letting you know
that we checked it and thought about it.

Thanks! I thought about adding a comment/note here as I had to convince
myself that thread.vrsave is indeed initialized to 0 before making this
change as well. I will mention it in the word-smithed commit message for
posterity.

quoted

 #else /* CONFIG_ALTIVEC */
 	err |= __put_user(0, &sc->v_regs);
 #endif /* CONFIG_ALTIVEC */
-	flush_fp_to_thread(tsk);
 	/* copy fpr regs and fpscr */
 	err |= copy_fpr_to_user(&sc->fp_regs, tsk);
 

@@ -150,7 +159,6 @@ static long setup_sigcontext(struct sigcontext __user *sc,
 	 * VMX data.
 	 */
 	if (tsk->thread.used_vsr && ctx_has_vsx_region) {
-		flush_vsx_to_thread(tsk);
 		v_regs += ELF_NVRREG;
 		err |= copy_vsx_to_user(v_regs, tsk);
 		/* set MSR_VSX in the MSR value in the frame to

@@ -655,6 +663,7 @@ SYSCALL_DEFINE3(swapcontext, struct ucontext __user *, old_ctx,
 		ctx_has_vsx_region = 1;
 
 	if (old_ctx != NULL) {
+		prepare_setup_sigcontext(current, ctx_has_vsx_region);
 		if (!access_ok(old_ctx, ctx_size)
 		    || setup_sigcontext(&old_ctx->uc_mcontext, current, 0, NULL, 0,
 					ctx_has_vsx_region)

I had a think about whether there was a problem with bubbling
prepare_setup_sigcontext over the access_ok() test, but given that
prepare_setup_sigcontext(current ...) doesn't access any of old_ctx, I'm
satisfied that it's OK - no changes needed.

Not sure I understand what you mean by 'bubbling over'?

quoted

@@ -842,6 +851,7 @@ int handle_rt_signal64(struct ksignal *ksig, sigset_t *set,
 #endif
 	{
 		err |= __put_user(0, &frame->uc.uc_link);
+		prepare_setup_sigcontext(tsk, 1);

Why do we call with ctx_has_vsx_region = 1 here? It's not immediately
clear to me why this is correct, but mpe and Mikey seem pretty convinced
that it is.

I think it's because we always have a "complete" sigcontext w/ the VSX
save area here, unlike in swapcontext where we have to check. Also, the
following unsafe_setup_sigcontext() is called with ctx_has_vsx_region=1
so assumes that the VSX data was copied by prepare_setup_sigcontext().

quoted

 		err |= setup_sigcontext(&frame->uc.uc_mcontext, tsk, ksig->sig,
 					NULL, (unsigned long)ksig->ka.sa.sa_handler,
 					1);

Finally, it's a bit hard to figure out where to put this, but we spent
some time making sure that the various things you moved into the
prepare_setup_sigcontext() function were called under the same
circumstances as they were before, and there were no concerns there.

Thanks for reviewing and double checking my work :)

Kind regards,
Daniel