ARM: add a kuser_cmpxchg64 user space helper

Some user space applications are designed around the ability to perform
atomic operations on 64 bit values.  Since this is natively possible
only with ARMv6k and above, let's provide a new kuser helper to perform
the operation with kernel supervision on pre ARMv6k hardware.
Signed-off-by: NNicolas Pitre <nicolas.pitre@linaro.org>
Tested-by: NDave Martin <dave.martin@linaro.org>

Documentation/arm/kernel_user_helpers.txt

@@ -201,3 +201,67 @@ typedef void (__kuser_dmb_t)(void);
 Notes:
 
   - Valid only if __kuser_helper_version >= 3 (from kernel version 2.6.15).
+
+kuser_cmpxchg64
+---------------
+
+Location:	0xffff0f60
+
+Reference prototype:
+
+  int __kuser_cmpxchg64(const int64_t *oldval,
+                        const int64_t *newval,
+                        volatile int64_t *ptr);
+
+Input:
+
+  r0 = pointer to oldval
+  r1 = pointer to newval
+  r2 = pointer to target value
+  lr = return address
+
+Output:
+
+  r0 = success code (zero or non-zero)
+  C flag = set if r0 == 0, clear if r0 != 0
+
+Clobbered registers:
+
+  r3, lr, flags
+
+Definition:
+
+  Atomically store the 64-bit value pointed by *newval in *ptr only if *ptr
+  is equal to the 64-bit value pointed by *oldval.  Return zero if *ptr was
+  changed or non-zero if no exchange happened.
+
+  The C flag is also set if *ptr was changed to allow for assembly
+  optimization in the calling code.
+
+Usage example:
+
+typedef int (__kuser_cmpxchg64_t)(const int64_t *oldval,
+                                  const int64_t *newval,
+                                  volatile int64_t *ptr);
+#define __kuser_cmpxchg64 (*(__kuser_cmpxchg64_t *)0xffff0f60)
+
+int64_t atomic_add64(volatile int64_t *ptr, int64_t val)
+{
+	int64_t old, new;
+
+	do {
+		old = *ptr;
+		new = old + val;
+	} while(__kuser_cmpxchg64(&old, &new, ptr));
+
+	return new;
+}
+
+Notes:
+
+  - This routine already includes memory barriers as needed.
+
+  - Due to the length of this sequence, this spans 2 conventional kuser
+    "slots", therefore 0xffff0f80 is not used as a valid entry point.
+
+  - Valid only if __kuser_helper_version >= 5 (from kernel version 3.1).

arch/arm/kernel/entry-armv.S

@@ -383,7 +383,7 @@ ENDPROC(__pabt_svc)
 	.endm
 
 	.macro	kuser_cmpxchg_check
-#if __LINUX_ARM_ARCH__ < 6 && !defined(CONFIG_NEEDS_SYSCALL_FOR_CMPXCHG)
+#if !defined(CONFIG_CPU_32v6K) && !defined(CONFIG_NEEDS_SYSCALL_FOR_CMPXCHG)
 #ifndef CONFIG_MMU
 #warning "NPTL on non MMU needs fixing"
 #else
@@ -392,7 +392,7 @@ ENDPROC(__pabt_svc)
 	@ perform a quick test inline since it should be false
 	@ 99.9999% of the time.  The rest is done out of line.
 	cmp	r2, #TASK_SIZE
-	blhs	kuser_cmpxchg_fixup
+	blhs	kuser_cmpxchg64_fixup
 #endif
 #endif
 	.endm
@@ -775,6 +775,99 @@ ENDPROC(__switch_to)
 	.globl	__kuser_helper_start
 __kuser_helper_start:
 
+/*
+ * Due to the length of some sequences, __kuser_cmpxchg64 spans 2 regular
+ * kuser "slots", therefore 0xffff0f80 is not used as a valid entry point.
+ */
+
+__kuser_cmpxchg64:				@ 0xffff0f60
+
+#if defined(CONFIG_NEEDS_SYSCALL_FOR_CMPXCHG)
+
+	/*
+	 * Poor you.  No fast solution possible...
+	 * The kernel itself must perform the operation.
+	 * A special ghost syscall is used for that (see traps.c).
+	 */
+	stmfd	sp!, {r7, lr}
+	ldr	r7, 1f			@ it's 20 bits
+	swi	__ARM_NR_cmpxchg64
+	ldmfd	sp!, {r7, pc}
+1:	.word	__ARM_NR_cmpxchg64
+
+#elif defined(CONFIG_CPU_32v6K)
+
+	stmfd	sp!, {r4, r5, r6, r7}
+	ldrd	r4, r5, [r0]			@ load old val
+	ldrd	r6, r7, [r1]			@ load new val
+	smp_dmb	arm
+1:	ldrexd	r0, r1, [r2]			@ load current val
+	eors	r3, r0, r4			@ compare with oldval (1)
+	eoreqs	r3, r1, r5			@ compare with oldval (2)
+	strexdeq r3, r6, r7, [r2]		@ store newval if eq
+	teqeq	r3, #1				@ success?
+	beq	1b				@ if no then retry
+	smp_dmb	arm
+	rsbs	r0, r3, #0			@ set returned val and C flag
+	ldmfd	sp!, {r4, r5, r6, r7}
+	bx	lr
+
+#elif !defined(CONFIG_SMP)
+
+#ifdef CONFIG_MMU
+
+	/*
+	 * The only thing that can break atomicity in this cmpxchg64
+	 * implementation is either an IRQ or a data abort exception
+	 * causing another process/thread to be scheduled in the middle of
+	 * the critical sequence.  The same strategy as for cmpxchg is used.
+	 */
+	stmfd	sp!, {r4, r5, r6, lr}
+	ldmia	r0, {r4, r5}			@ load old val
+	ldmia	r1, {r6, lr}			@ load new val
+1:	ldmia	r2, {r0, r1}			@ load current val
+	eors	r3, r0, r4			@ compare with oldval (1)
+	eoreqs	r3, r1, r5			@ compare with oldval (2)
+2:	stmeqia	r2, {r6, lr}			@ store newval if eq
+	rsbs	r0, r3, #0			@ set return val and C flag
+	ldmfd	sp!, {r4, r5, r6, pc}
+
+	.text
+kuser_cmpxchg64_fixup:
+	@ Called from kuser_cmpxchg_fixup.
+	@ r2 = address of interrupted insn (must be preserved).
+	@ sp = saved regs. r7 and r8 are clobbered.
+	@ 1b = first critical insn, 2b = last critical insn.
+	@ If r2 >= 1b and r2 <= 2b then saved pc_usr is set to 1b.
+	mov	r7, #0xffff0fff
+	sub	r7, r7, #(0xffff0fff - (0xffff0f60 + (1b - __kuser_cmpxchg64)))
+	subs	r8, r2, r7
+	rsbcss	r8, r8, #(2b - 1b)
+	strcs	r7, [sp, #S_PC]
+#if __LINUX_ARM_ARCH__ < 6
+	bcc	kuser_cmpxchg32_fixup
+#endif
+	mov	pc, lr
+	.previous
+
+#else
+#warning "NPTL on non MMU needs fixing"
+	mov	r0, #-1
+	adds	r0, r0, #0
+	usr_ret	lr
+#endif
+
+#else
+#error "incoherent kernel configuration"
+#endif
+
+	/* pad to next slot */
+	.rept	(16 - (. - __kuser_cmpxchg64)/4)
+	.word	0
+	.endr
+
+	.align	5
+
 __kuser_memory_barrier:				@ 0xffff0fa0
 	smp_dmb	arm
 	usr_ret	lr
@@ -816,7 +909,7 @@ __kuser_cmpxchg:				@ 0xffff0fc0
 	usr_ret	lr
 
 	.text
-kuser_cmpxchg_fixup:
+kuser_cmpxchg32_fixup:
 	@ Called from kuser_cmpxchg_check macro.
 	@ r2 = address of interrupted insn (must be preserved).
 	@ sp = saved regs. r7 and r8 are clobbered.