kasan: speed up mte_set_mem_tag_range

Use DC GVA / DC GZVA to speed up KASan memory tagging in HW tags mode.

The first cacheline is always tagged using STG/STZG even if the address is
cacheline-aligned, as benchmarks show it is faster than a conditional
branch.

Signed-off-by: Evgenii Stepanov <[e-mail]>
Co-developed-by: Peter Collingbourne <[e-mail]>
Signed-off-by: Peter Collingbourne <[e-mail]>
Reviewed-by: Catalin Marinas <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Will Deacon <[e-mail]>

arch/arm64/include/asm/mte-kasan.h

···
 	return mte_get_ptr_tag(addr);
 }
 
+static inline u64 __stg_post(u64 p)
+{
+	asm volatile(__MTE_PREAMBLE "stg %0, [%0], #16"
+		     : "+r"(p)
+		     :
+		     : "memory");
+	return p;
+}
+
+static inline u64 __stzg_post(u64 p)
+{
+	asm volatile(__MTE_PREAMBLE "stzg %0, [%0], #16"
+		     : "+r"(p)
+		     :
+		     : "memory");
+	return p;
+}
+
+static inline void __dc_gva(u64 p)
+{
+	asm volatile(__MTE_PREAMBLE "dc gva, %0" : : "r"(p) : "memory");
+}
+
+static inline void __dc_gzva(u64 p)
+{
+	asm volatile(__MTE_PREAMBLE "dc gzva, %0" : : "r"(p) : "memory");
+}
+
 /*
  * Assign allocation tags for a region of memory based on the pointer tag.
  * Note: The address must be non-NULL and MTE_GRANULE_SIZE aligned and
- * size must be non-zero and MTE_GRANULE_SIZE aligned.
+ * size must be MTE_GRANULE_SIZE aligned.
  */
-static inline void mte_set_mem_tag_range(void *addr, size_t size,
-						u8 tag, bool init)
+static inline void mte_set_mem_tag_range(void *addr, size_t size, u8 tag,
+					 bool init)
 {
-	u64 curr, end;
+	u64 curr, mask, dczid_bs, end1, end2, end3;
 
-	if (!size)
-		return;
+	/* Read DC G(Z)VA block size from the system register. */
+	dczid_bs = 4ul << (read_cpuid(DCZID_EL0) & 0xf);
 
 	curr = (u64)__tag_set(addr, tag);
-	end = curr + size;
+	mask = dczid_bs - 1;
+	/* STG/STZG up to the end of the first block. */
+	end1 = curr | mask;
+	end3 = curr + size;
+	/* DC GVA / GZVA in [end1, end2) */
+	end2 = end3 & ~mask;
 
 	/*
-	 * 'asm volatile' is required to prevent the compiler to move
-	 * the statement outside of the loop.
+	 * The following code uses STG on the first DC GVA block even if the
+	 * start address is aligned - it appears to be faster than an alignment
+	 * check + conditional branch. Also, if the range size is at least 2 DC
+	 * GVA blocks, the first two loops can use post-condition to save one
+	 * branch each.
 	 */
-	if (init) {
-		do {
-			asm volatile(__MTE_PREAMBLE "stzg %0, [%0]"
-				     :
-				     : "r" (curr)
-				     : "memory");
-			curr += MTE_GRANULE_SIZE;
-		} while (curr != end);
-	} else {
-		do {
-			asm volatile(__MTE_PREAMBLE "stg %0, [%0]"
-				     :
-				     : "r" (curr)
-				     : "memory");
-			curr += MTE_GRANULE_SIZE;
-		} while (curr != end);
-	}
+#define SET_MEMTAG_RANGE(stg_post, dc_gva)		\
+	do {						\
+		if (size >= 2 * dczid_bs) {		\
+			do {				\
+				curr = stg_post(curr);	\
+			} while (curr < end1);		\
+							\
+			do {				\
+				dc_gva(curr);		\
+				curr += dczid_bs;	\
+			} while (curr < end2);		\
+		}					\
+							\
+		while (curr < end3)			\
+			curr = stg_post(curr);		\
+	} while (0)
+
+	if (init)
+		SET_MEMTAG_RANGE(__stzg_post, __dc_gzva);
+	else
+		SET_MEMTAG_RANGE(__stg_post, __dc_gva);
+#undef SET_MEMTAG_RANGE
 }
 
 void mte_enable_kernel_sync(void);