[dpdk-dev] [RFC 03/12] acl: add support for s390x architecture

[Vivian Kong <vivkong@gmail.com>]

Add big endian support for s390x architecture.

Signed-off-by: Vivian Kong <vivkong@ca.ibm.com>
---
 app/test-acl/main.c             |  4 ++
 lib/librte_acl/Makefile         |  2 +
 lib/librte_acl/acl_bld.c        | 69 +++++++++++++++++++++++++++------
 lib/librte_acl/acl_gen.c        |  9 +++++
 lib/librte_acl/acl_run_scalar.c |  8 ++++
 lib/librte_acl/rte_acl.c        |  4 ++
 lib/librte_acl/rte_acl.h        |  1 +
 7 files changed, 85 insertions(+), 12 deletions(-)

diff --git a/app/test-acl/main.c b/app/test-acl/main.c
index b80179417..b6c5c9abd 100644
--- a/app/test-acl/main.c
+++ b/app/test-acl/main.c
@@ -81,6 +81,10 @@ static const struct acl_alg acl_alg[] = {
 		.name = "altivec",
 		.alg = RTE_ACL_CLASSIFY_ALTIVEC,
 	},
+	{
+		.name = "s390x",
+		.alg = RTE_ACL_CLASSIFY_S390X,
+	},
 };
 
 static struct {
diff --git a/lib/librte_acl/Makefile b/lib/librte_acl/Makefile
index ea5edf00a..f87693d1e 100644
--- a/lib/librte_acl/Makefile
+++ b/lib/librte_acl/Makefile
@@ -30,6 +30,8 @@ CFLAGS_acl_run_neon.o += -Wno-maybe-uninitialized
 endif
 else ifeq ($(CONFIG_RTE_ARCH_PPC_64),y)
 SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_altivec.c
+else ifeq ($(CONFIG_RTE_ARCH_S390X),y)
+SRCS-$(CONFIG_RTE_LIBRTE_ACL) +=
 else
 SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_sse.c
 endif
diff --git a/lib/librte_acl/acl_bld.c b/lib/librte_acl/acl_bld.c
index b82191f42..16bf09304 100644
--- a/lib/librte_acl/acl_bld.c
+++ b/lib/librte_acl/acl_bld.c
@@ -777,6 +777,16 @@ acl_build_reset(struct rte_acl_ctx *ctx)
 		sizeof(*ctx) - offsetof(struct rte_acl_ctx, num_categories));
 }
 
+static uint32_t get_le_byte_index(uint32_t index, int size)
+{
+#if  __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+	(void) size;
+	return index;
+#else
+	return size - 1 - index;
+#endif
+}
+
 static void
 acl_gen_range(struct acl_build_context *context,
 	const uint8_t *hi, const uint8_t *lo, int size, int level,
@@ -786,12 +796,19 @@ acl_gen_range(struct acl_build_context *context,
 	uint32_t n;
 
 	prev = root;
+
+	/* On big endian min and max point to highest byte.
+	 * Therefore iterate in opposite direction as on
+	 * little endian with helper function.
+	 */
 	for (n = size - 1; n > 0; n--) {
+		uint32_t le_idx = get_le_byte_index(n, size);
 		node = acl_alloc_node(context, level++);
-		acl_add_ptr_range(context, prev, node, lo[n], hi[n]);
+		acl_add_ptr_range(context, prev, node, lo[le_idx], hi[le_idx]);
 		prev = node;
 	}
-	acl_add_ptr_range(context, prev, end, lo[0], hi[0]);
+	const uint32_t first_idx = get_le_byte_index(0, size);
+	acl_add_ptr_range(context, prev, end, lo[first_idx], hi[first_idx]);
 }
 
 static struct rte_acl_node *
@@ -804,10 +821,16 @@ acl_gen_range_trie(struct acl_build_context *context,
 	const uint8_t *lo = min;
 	const uint8_t *hi = max;
 
+	/* On big endian min and max point to highest byte.
+	 * Therefore iterate in opposite direction as on
+	 * little endian.
+	 */
+	const int byte_index = get_le_byte_index(size-1, size);
+
 	*pend = acl_alloc_node(context, level+size);
 	root = acl_alloc_node(context, level++);
 
-	if (lo[size - 1] == hi[size - 1]) {
+	if (lo[byte_index] == hi[byte_index]) {
 		acl_gen_range(context, hi, lo, size, level, root, *pend);
 	} else {
 		uint8_t limit_lo[64];
@@ -819,27 +842,29 @@ acl_gen_range_trie(struct acl_build_context *context,
 		memset(limit_hi, UINT8_MAX, RTE_DIM(limit_hi));
 
 		for (n = size - 2; n >= 0; n--) {
-			hi_ff = (uint8_t)(hi_ff & hi[n]);
-			lo_00 = (uint8_t)(lo_00 | lo[n]);
+			const uint32_t le_idx = get_le_byte_index(n, size);
+			hi_ff = (uint8_t)(hi_ff & hi[le_idx]);
+			lo_00 = (uint8_t)(lo_00 | lo[le_idx]);
 		}
 
 		if (hi_ff != UINT8_MAX) {
-			limit_lo[size - 1] = hi[size - 1];
+			limit_lo[byte_index] = hi[byte_index];
 			acl_gen_range(context, hi, limit_lo, size, level,
 				root, *pend);
 		}
 
 		if (lo_00 != 0) {
-			limit_hi[size - 1] = lo[size - 1];
+			limit_hi[byte_index] = lo[byte_index];
 			acl_gen_range(context, limit_hi, lo, size, level,
 				root, *pend);
 		}
 
-		if (hi[size - 1] - lo[size - 1] > 1 ||
+		if (hi[byte_index] - lo[byte_index] > 1 ||
 				lo_00 == 0 ||
 				hi_ff == UINT8_MAX) {
-			limit_lo[size-1] = (uint8_t)(lo[size-1] + (lo_00 != 0));
-			limit_hi[size-1] = (uint8_t)(hi[size-1] -
+			limit_lo[byte_index] = (uint8_t)(lo[byte_index] +
+				(lo_00 != 0));
+			limit_hi[byte_index] = (uint8_t)(hi[byte_index] -
 				(hi_ff != UINT8_MAX));
 			acl_gen_range(context, limit_hi, limit_lo, size,
 				level, root, *pend);
@@ -863,13 +888,17 @@ acl_gen_mask_trie(struct acl_build_context *context,
 	root = acl_alloc_node(context, level++);
 	prev = root;
 
+	/* On big endian val and msk point to highest byte.
+	 * Therefore iterate in opposite direction as on
+	 * little endian with helper function
+	 */
 	for (n = size - 1; n >= 0; n--) {
+		uint32_t le_idx = get_le_byte_index(n, size);
 		node = acl_alloc_node(context, level++);
-		acl_gen_mask(&bits, val[n] & msk[n], msk[n]);
+		acl_gen_mask(&bits, val[le_idx] & msk[le_idx], msk[le_idx]);
 		acl_add_ptr(context, prev, node, &bits);
 		prev = node;
 	}
-
 	*pend = prev;
 	return root;
 }
@@ -927,6 +956,14 @@ build_trie(struct acl_build_context *context, struct rte_acl_build_rule *head,
 					fld->mask_range.u32,
 					rule->config->defs[n].size);
 
+				/* Fields are aligned highest to lowest bit.
+				 * Masked needs to be shifted to follow same
+				 * convention
+				 */
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+				mask = mask << 32;
+#endif
+
 				/* gen a mini-trie for this field */
 				merge = acl_gen_mask_trie(context,
 					&fld->value,
@@ -1017,6 +1054,14 @@ acl_calc_wildness(struct rte_acl_build_rule *head,
 			uint32_t bit_len = CHAR_BIT * config->defs[n].size;
 			uint64_t msk_val = RTE_LEN2MASK(bit_len,
 				typeof(msk_val));
+
+			/* Fields are aligned highest to lowest bit.
+			 * Masked needs to be shifted to follow same
+			 * convention
+			 */
+			if (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+				msk_val <<= 32;
+
 			double size = bit_len;
 			int field_index = config->defs[n].field_index;
 			const struct rte_acl_field *fld = rule->f->field +
diff --git a/lib/librte_acl/acl_gen.c b/lib/librte_acl/acl_gen.c
index 35a0140b4..b4e8df0ce 100644
--- a/lib/librte_acl/acl_gen.c
+++ b/lib/librte_acl/acl_gen.c
@@ -360,7 +360,16 @@ acl_gen_node(struct rte_acl_node *node, uint64_t *node_array,
 		array_ptr = &node_array[index->quad_index];
 		acl_add_ptrs(node, array_ptr, no_match, 0);
 		qtrp = (uint32_t *)node->transitions;
+
+		/* Swap qtrp on big endian that transitions[0]
+		 * is at least signifcant byte.
+		 */
+#if __BYTE_ORDER == __ORDER_BIG_ENDIAN__
+		node->node_index = __bswap_32(qtrp[0]);
+#else
 		node->node_index = qtrp[0];
+#endif
+
 		node->node_index <<= sizeof(index->quad_index) * CHAR_BIT;
 		node->node_index |= index->quad_index | node->node_type;
 		index->quad_index += node->fanout;
diff --git a/lib/librte_acl/acl_run_scalar.c b/lib/librte_acl/acl_run_scalar.c
index 3d61e7940..9f01ef8d8 100644
--- a/lib/librte_acl/acl_run_scalar.c
+++ b/lib/librte_acl/acl_run_scalar.c
@@ -141,6 +141,14 @@ rte_acl_classify_scalar(const struct rte_acl_ctx *ctx, const uint8_t **data,
 		input0 = GET_NEXT_4BYTES(parms, 0);
 		input1 = GET_NEXT_4BYTES(parms, 1);
 
+		/* input needs to be swapped because the rules get
+		 * swapped while building the trie.
+		 */
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+		input0 = __bswap_32(input0);
+		input1 = __bswap_32(input1);
+#endif
+
 		for (n = 0; n < 4; n++) {
 
 			transition0 = scalar_transition(flows.trans,
diff --git a/lib/librte_acl/rte_acl.c b/lib/librte_acl/rte_acl.c
index c436a9bfd..6d4d3f239 100644
--- a/lib/librte_acl/rte_acl.c
+++ b/lib/librte_acl/rte_acl.c
@@ -64,6 +64,8 @@ static const rte_acl_classify_t classify_fns[] = {
 	[RTE_ACL_CLASSIFY_AVX2] = rte_acl_classify_avx2,
 	[RTE_ACL_CLASSIFY_NEON] = rte_acl_classify_neon,
 	[RTE_ACL_CLASSIFY_ALTIVEC] = rte_acl_classify_altivec,
+	/* use scalar for s390x for now */
+	[RTE_ACL_CLASSIFY_S390X] = rte_acl_classify_scalar,
 };
 
 /* by default, use always available scalar code path. */
@@ -103,6 +105,8 @@ RTE_INIT(rte_acl_init)
 		alg =  RTE_ACL_CLASSIFY_NEON;
 #elif defined(RTE_ARCH_PPC_64)
 	alg = RTE_ACL_CLASSIFY_ALTIVEC;
+#elif defined(RTE_ARCH_S390X)
+	alg = RTE_ACL_CLASSIFY_S390X;
 #else
 #ifdef CC_AVX2_SUPPORT
 	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
diff --git a/lib/librte_acl/rte_acl.h b/lib/librte_acl/rte_acl.h
index aa22e70c6..9537196db 100644
--- a/lib/librte_acl/rte_acl.h
+++ b/lib/librte_acl/rte_acl.h
@@ -241,6 +241,7 @@ enum rte_acl_classify_alg {
 	RTE_ACL_CLASSIFY_AVX2 = 3,    /**< requires AVX2 support. */
 	RTE_ACL_CLASSIFY_NEON = 4,    /**< requires NEON support. */
 	RTE_ACL_CLASSIFY_ALTIVEC = 5,    /**< requires ALTIVEC support. */
+	RTE_ACL_CLASSIFY_S390X = 6,    /**< requires s390x z13 support. */
 	RTE_ACL_CLASSIFY_NUM          /* should always be the last one. */
 };
 
-- 
2.17.1