MIPS: cmpxchg: Implement 1 byte & 2 byte cmpxchg()

Implement support for 1 & 2 byte cmpxchg() using read-modify-write atop
a 4 byte cmpxchg(). This allows us to support these atomic operations
despite the MIPS ISA only providing 4 & 8 byte atomic operations.

This is required in order to support queued rwlocks (qrwlock) in a later
patch, since these make use of a 1 byte cmpxchg() in their slow path.

Signed-off-by: Paul Burton <paul.burton@imgtec.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/16355/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

arch/mips/include/asm/cmpxchg.h

@@ -142,10 +142,17 @@ static inline unsigned long __xchg(unsigned long x, volatile void * ptr, int siz
 	__ret;								\
 })
 
+extern unsigned long __cmpxchg_small(volatile void *ptr, unsigned long old,
+				     unsigned long new, unsigned int size);
+
 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
 				      unsigned long new, unsigned int size)
 {
 	switch (size) {
+	case 1:
+	case 2:
+		return __cmpxchg_small(ptr, old, new, size);
+
 	case 4:
 		return __cmpxchg_asm("ll", "sc", (volatile u32 *)ptr, old, new);
 

arch/mips/kernel/cmpxchg.c

@@ -50,3 +50,60 @@ unsigned long __xchg_small(volatile void *ptr, unsigned long val, unsigned int s
 
 	return (load32 & mask) >> shift;
 }
+
+unsigned long __cmpxchg_small(volatile void *ptr, unsigned long old,
+			      unsigned long new, unsigned int size)
+{
+	u32 mask, old32, new32, load32;
+	volatile u32 *ptr32;
+	unsigned int shift;
+	u8 load;
+
+	/* Check that ptr is naturally aligned */
+	WARN_ON((unsigned long)ptr & (size - 1));
+
+	/* Mask inputs to the correct size. */
+	mask = GENMASK((size * BITS_PER_BYTE) - 1, 0);
+	old &= mask;
+	new &= mask;
+
+	/*
+	 * Calculate a shift & mask that correspond to the value we wish to
+	 * compare & exchange within the naturally aligned 4 byte integer
+	 * that includes it.
+	 */
+	shift = (unsigned long)ptr & 0x3;
+	if (IS_ENABLED(CONFIG_CPU_BIG_ENDIAN))
+		shift ^= sizeof(u32) - size;
+	shift *= BITS_PER_BYTE;
+	mask <<= shift;
+
+	/*
+	 * Calculate a pointer to the naturally aligned 4 byte integer that
+	 * includes our byte of interest, and load its value.
+	 */
+	ptr32 = (volatile u32 *)((unsigned long)ptr & ~0x3);
+	load32 = *ptr32;
+
+	while (true) {
+		/*
+		 * Ensure the byte we want to exchange matches the expected
+		 * old value, and if not then bail.
+		 */
+		load = (load32 & mask) >> shift;
+		if (load != old)
+			return load;
+
+		/*
+		 * Calculate the old & new values of the naturally aligned
+		 * 4 byte integer that include the byte we want to exchange.
+		 * Attempt to exchange the old value for the new value, and
+		 * return if we succeed.
+		 */
+		old32 = (load32 & ~mask) | (old << shift);
+		new32 = (load32 & ~mask) | (new << shift);
+		load32 = cmpxchg(ptr32, old32, new32);
+		if (load32 == old32)
+			return old;
+	}
+}