Re: [RFC][PATCH v2 4/5] x86/uaccess: Implement unsafe_try_cmpxchg_user()

On Thu, Jan 27, 2022 at 02:17:20AM +0000, Sean Christopherson wrote:

On Thu, Jan 20, 2022, Peter Zijlstra wrote:

quoted

Do try_cmpxchg() loops on userspace addresses.

Cc: Sean Christopherson <seanjc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 arch/x86/include/asm/uaccess.h |   67 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 67 insertions(+)

--- a/arch/x86/include/asm/uaccess.h
+++ b/arch/x86/include/asm/uaccess.h

@@ -342,6 +342,24 @@ do {									\
 		     : [umem] "m" (__m(addr))				\
 		     : : label)
 
+#define __try_cmpxchg_user_asm(itype, ltype, _ptr, _pold, _new, label)	({ \
+	bool success;							\
+	__typeof__(_ptr) _old = (__typeof__(_ptr))(_pold);		\
+	__typeof__(*(_ptr)) __old = *_old;				\
+	__typeof__(*(_ptr)) __new = (_new);				\
+	asm_volatile_goto("\n"						\
+		     "1: " LOCK_PREFIX "cmpxchg"itype" %[new], %[ptr]\n"\
+		     _ASM_EXTABLE_UA(1b, %l[label])			\
+		     : CC_OUT(z) (success),				\
+		       [ptr] "+m" (*_ptr),				\
+		       [old] "+a" (__old)				\
+		     : [new] ltype (__new)				\
+		     : "memory", "cc"					\

IIUC, the "cc" clobber is unnecessary as CONFIG_CC_HAS_ASM_GOTO_OUTPUT=y implies
__GCC_ASM_FLAG_OUTPUTS__=y, i.e. CC_OUT() will resolve to "=@cc".

Yeah, even without that GCC always assumes 'cc' is clobbered due to
hysterical raisins.

quoted

+		     : label);						\
+	if (unlikely(!success))						\
+		*_old = __old;						\
+	likely(success);					})
+
 #else // !CONFIG_CC_HAS_ASM_GOTO_OUTPUT

...

quoted

+extern void __try_cmpxchg_user_wrong_size(void);
+
+#define unsafe_try_cmpxchg_user(_ptr, _oldp, _nval, _label) ({		\
+	__typeof__(*(_ptr)) __ret;					\

This should probably be a bool, the return from the lower level helpers is a bool
that's true if the exchange succeed.  Declaring the type of the target implies
that they return the raw result, which is confusing.

Fair enough.

quoted

+	switch (sizeof(__ret)) {					\
+	case 1:	__ret = __try_cmpxchg_user_asm("b", "q",		\
+					       (_ptr), (_oldp),		\
+					       (_nval), _label);	\
+		break;							\
+	case 2:	__ret = __try_cmpxchg_user_asm("w", "r",		\
+					       (_ptr), (_oldp),		\
+					       (_nval), _label);	\
+		break;							\
+	case 4:	__ret = __try_cmpxchg_user_asm("l", "r",		\
+					       (_ptr), (_oldp),		\
+					       (_nval), _label);	\
+		break;							\
+	case 8:	__ret = __try_cmpxchg_user_asm("q", "r",		\
+					       (_ptr), (_oldp),		\
+					       (_nval), _label);	\

Doh, I should have specified that KVM needs 8-byte CMPXCHG on 32-bit kernels due
to using it to atomically update guest PAE PTEs and LTR descriptors (yay).

:-) I'm so trying to de-feature 32bit.

Also, KVM's use case isn't a tight loop, how gross would it be to add a slightly
less unsafe version that does __uaccess_begin_nospec()?  KVM pre-checks the address
way ahead of time, so the access_ok() check can be omitted.  Alternatively, KVM
could add its own macro, but that seems a little silly.  E.g. somethign like this,
though I don't think this is correct (something is getting inverted somewhere and
the assembly output is a nightmare):

/* "Returns" 0 on success, 1 on failure, -EFAULT if the access faults. */
#define ___try_cmpxchg_user(_ptr, _oldp, _nval, _label)	({		\
	int ____ret = -EFAULT;						\
	__uaccess_begin_nospec();					\
	____ret = !unsafe_try_cmpxchg_user(_ptr, _oldp, _nval, _label);	\
_label:									\
	__uaccess_end();						\
	____ret;							\
						})

Works for me I suppose, but we really ought to keep usage of that in
arch code.

Lastly, assuming I get my crap working, mind if I post a variant (Cc'd to stable@) in
the context of KVM series?  

Not at all.