Merge branch 'bpf-nfp-jmp-memcpy-improvements'

Jiong Wang says:

====================
Currently, compiler will lower memcpy function call in XDP/eBPF C program
into a sequence of eBPF load/store pairs for some scenarios.

Compiler is thinking this "inline" optimiation is beneficial as it could
avoid function call and also increase code locality.

However, Netronome NPU is not an tranditional load/store architecture that
doing a sequence of individual load/store actions are not efficient.

This patch set tries to identify the load/store sequences composed of
load/store pairs that comes from memcpy lowering, then accelerates them
through NPU's Command Push Pull (CPP) instruction.

This patch set registered an new optimization pass before doing the actual
JIT work, it traverse through eBPF IR, once found candidate sequence then
record the memory copy source, destination and length information in the
first load instruction starting the sequence and marks all remaining
instructions in the sequence into skipable status. Later, when JITing the
first load instructoin, optimal instructions will be generated using those
record information.

For this safety of this transformation:

  - jump into the middle of the sequence will cancel the optimization.

  - overlapped memory access will cancel the optimization.

  - the load destination register still contains the same value as before
    the transformation.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

drivers/net/ethernet/netronome/nfp/bpf/jit.c

 /*
- * Copyright (C) 2016 Netronome Systems, Inc.
+ * Copyright (C) 2016-2017 Netronome Systems, Inc.
  *
  * This software is dual licensed under the GNU General License Version 2,
  * June 1991 as shown in the file COPYING in the top-level directory of this
@@ -65,12 +65,6 @@
 	     next = nfp_meta_next(pos),				\
 	     next2 = nfp_meta_next(next))
 
-static bool
-nfp_meta_has_next(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return meta->l.next != &nfp_prog->insns;
-}
-
 static bool
 nfp_meta_has_prev(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
@@ -102,7 +96,7 @@ nfp_prog_offset_to_index(struct nfp_prog *nfp_prog, unsigned int offset)
 /* --- Emitters --- */
 static void
 __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op,
-	   u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, bool sync)
+	   u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, bool sync, bool indir)
 {
 	enum cmd_ctx_swap ctx;
 	u64 insn;
@@ -120,14 +114,15 @@ __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op,
 		FIELD_PREP(OP_CMD_CNT, size) |
 		FIELD_PREP(OP_CMD_SIG, sync) |
 		FIELD_PREP(OP_CMD_TGT_CMD, cmd_tgt_act[op].tgt_cmd) |
+		FIELD_PREP(OP_CMD_INDIR, indir) |
 		FIELD_PREP(OP_CMD_MODE, mode);
 
 	nfp_prog_push(nfp_prog, insn);
 }
 
 static void
-emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op,
-	 u8 mode, u8 xfer, swreg lreg, swreg rreg, u8 size, bool sync)
+emit_cmd_any(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer,
+	     swreg lreg, swreg rreg, u8 size, bool sync, bool indir)
 {
 	struct nfp_insn_re_regs reg;
 	int err;
@@ -148,7 +143,22 @@ emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op,
 		return;
 	}
 
-	__emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, sync);
+	__emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, sync,
+		   indir);
+}
+
+static void
+emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer,
+	 swreg lreg, swreg rreg, u8 size, bool sync)
+{
+	emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, sync, false);
+}
+
+static void
+emit_cmd_indir(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer,
+	       swreg lreg, swreg rreg, u8 size, bool sync)
+{
+	emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, sync, true);
 }
 
 static void
@@ -230,9 +240,11 @@ emit_immed(struct nfp_prog *nfp_prog, swreg dst, u16 imm,
 		return;
 	}
 
-	__emit_immed(nfp_prog, reg.areg, reg.breg, imm >> 8, width,
-		     invert, shift, reg.wr_both,
-		     reg.dst_lmextn, reg.src_lmextn);
+	/* Use reg.dst when destination is No-Dest. */
+	__emit_immed(nfp_prog,
+		     swreg_type(dst) == NN_REG_NONE ? reg.dst : reg.areg,
+		     reg.breg, imm >> 8, width, invert, shift,
+		     reg.wr_both, reg.dst_lmextn, reg.src_lmextn);
 }
 
 static void
@@ -510,6 +522,147 @@ static void wrp_reg_mov(struct nfp_prog *nfp_prog, u16 dst, u16 src)
 	wrp_mov(nfp_prog, reg_both(dst), reg_b(src));
 }
 
+/* wrp_reg_subpart() - load @field_len bytes from @offset of @src, write the
+ * result to @dst from low end.
+ */
+static void
+wrp_reg_subpart(struct nfp_prog *nfp_prog, swreg dst, swreg src, u8 field_len,
+		u8 offset)
+{
+	enum shf_sc sc = offset ? SHF_SC_R_SHF : SHF_SC_NONE;
+	u8 mask = (1 << field_len) - 1;
+
+	emit_ld_field_any(nfp_prog, dst, mask, src, sc, offset * 8, true);
+}
+
+/* NFP has Command Push Pull bus which supports bluk memory operations. */
+static int nfp_cpp_memcpy(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	bool descending_seq = meta->ldst_gather_len < 0;
+	s16 len = abs(meta->ldst_gather_len);
+	swreg src_base, off;
+	unsigned int i;
+	u8 xfer_num;
+
+	off = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog));
+	src_base = reg_a(meta->insn.src_reg * 2);
+	xfer_num = round_up(len, 4) / 4;
+
+	/* Setup PREV_ALU fields to override memory read length. */
+	if (len > 32)
+		wrp_immed(nfp_prog, reg_none(),
+			  CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1));
+
+	/* Memory read from source addr into transfer-in registers. */
+	emit_cmd_any(nfp_prog, CMD_TGT_READ32_SWAP, CMD_MODE_32b, 0, src_base,
+		     off, xfer_num - 1, true, len > 32);
+
+	/* Move from transfer-in to transfer-out. */
+	for (i = 0; i < xfer_num; i++)
+		wrp_mov(nfp_prog, reg_xfer(i), reg_xfer(i));
+
+	off = re_load_imm_any(nfp_prog, meta->paired_st->off, imm_b(nfp_prog));
+
+	if (len <= 8) {
+		/* Use single direct_ref write8. */
+		emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0,
+			 reg_a(meta->paired_st->dst_reg * 2), off, len - 1,
+			 true);
+	} else if (len <= 32 && IS_ALIGNED(len, 4)) {
+		/* Use single direct_ref write32. */
+		emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
+			 reg_a(meta->paired_st->dst_reg * 2), off, xfer_num - 1,
+			 true);
+	} else if (len <= 32) {
+		/* Use single indirect_ref write8. */
+		wrp_immed(nfp_prog, reg_none(),
+			  CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, len - 1));
+		emit_cmd_indir(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0,
+			       reg_a(meta->paired_st->dst_reg * 2), off,
+			       len - 1, true);
+	} else if (IS_ALIGNED(len, 4)) {
+		/* Use single indirect_ref write32. */
+		wrp_immed(nfp_prog, reg_none(),
+			  CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1));
+		emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
+			       reg_a(meta->paired_st->dst_reg * 2), off,
+			       xfer_num - 1, true);
+	} else if (len <= 40) {
+		/* Use one direct_ref write32 to write the first 32-bytes, then
+		 * another direct_ref write8 to write the remaining bytes.
+		 */
+		emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
+			 reg_a(meta->paired_st->dst_reg * 2), off, 7,
+			 true);
+
+		off = re_load_imm_any(nfp_prog, meta->paired_st->off + 32,
+				      imm_b(nfp_prog));
+		emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 8,
+			 reg_a(meta->paired_st->dst_reg * 2), off, len - 33,
+			 true);
+	} else {
+		/* Use one indirect_ref write32 to write 4-bytes aligned length,
+		 * then another direct_ref write8 to write the remaining bytes.
+		 */
+		u8 new_off;
+
+		wrp_immed(nfp_prog, reg_none(),
+			  CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 2));
+		emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
+			       reg_a(meta->paired_st->dst_reg * 2), off,
+			       xfer_num - 2, true);
+		new_off = meta->paired_st->off + (xfer_num - 1) * 4;
+		off = re_load_imm_any(nfp_prog, new_off, imm_b(nfp_prog));
+		emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b,
+			 xfer_num - 1, reg_a(meta->paired_st->dst_reg * 2), off,
+			 (len & 0x3) - 1, true);
+	}
+
+	/* TODO: The following extra load is to make sure data flow be identical
+	 *  before and after we do memory copy optimization.
+	 *
+	 *  The load destination register is not guaranteed to be dead, so we
+	 *  need to make sure it is loaded with the value the same as before
+	 *  this transformation.
+	 *
+	 *  These extra loads could be removed once we have accurate register
+	 *  usage information.
+	 */
+	if (descending_seq)
+		xfer_num = 0;
+	else if (BPF_SIZE(meta->insn.code) != BPF_DW)
+		xfer_num = xfer_num - 1;
+	else
+		xfer_num = xfer_num - 2;
+
+	switch (BPF_SIZE(meta->insn.code)) {
+	case BPF_B:
+		wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2),
+				reg_xfer(xfer_num), 1,
+				IS_ALIGNED(len, 4) ? 3 : (len & 3) - 1);
+		break;
+	case BPF_H:
+		wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2),
+				reg_xfer(xfer_num), 2, (len & 3) ^ 2);
+		break;
+	case BPF_W:
+		wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2),
+			reg_xfer(0));
+		break;
+	case BPF_DW:
+		wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2),
+			reg_xfer(xfer_num));
+		wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1),
+			reg_xfer(xfer_num + 1));
+		break;
+	}
+
+	if (BPF_SIZE(meta->insn.code) != BPF_DW)
+		wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
+
+	return 0;
+}
+
 static int
 data_ld(struct nfp_prog *nfp_prog, swreg offset, u8 dst_gpr, int size)
 {
@@ -975,9 +1128,6 @@ wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 {
 	const struct bpf_insn *insn = &meta->insn;
 
-	if (insn->off < 0) /* TODO */
-		return -EOPNOTSUPP;
-
 	wrp_test_reg_one(nfp_prog, insn->dst_reg * 2, alu_op,
 			 insn->src_reg * 2, br_mask, insn->off);
 	wrp_test_reg_one(nfp_prog, insn->dst_reg * 2 + 1, alu_op,
@@ -995,9 +1145,6 @@ wrp_cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 	u8 reg = insn->dst_reg * 2;
 	swreg tmp_reg;
 
-	if (insn->off < 0) /* TODO */
-		return -EOPNOTSUPP;
-
 	tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
 	if (!swap)
 		emit_alu(nfp_prog, reg_none(), reg_a(reg), ALU_OP_SUB, tmp_reg);
@@ -1027,9 +1174,6 @@ wrp_cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 	areg = insn->dst_reg * 2;
 	breg = insn->src_reg * 2;
 
-	if (insn->off < 0) /* TODO */
-		return -EOPNOTSUPP;
-
 	if (swap) {
 		areg ^= breg;
 		breg ^= areg;
@@ -1494,6 +1638,9 @@ static int
 mem_ldx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 	unsigned int size)
 {
+	if (meta->ldst_gather_len)
+		return nfp_cpp_memcpy(nfp_prog, meta);
+
 	if (meta->ptr.type == PTR_TO_CTX) {
 		if (nfp_prog->type == BPF_PROG_TYPE_XDP)
 			return mem_ldx_xdp(nfp_prog, meta, size);
@@ -1630,8 +1777,6 @@ static int mem_stx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 
 static int jump(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
-	if (meta->insn.off < 0) /* TODO */
-		return -EOPNOTSUPP;
 	emit_br(nfp_prog, BR_UNC, meta->insn.off, 0);
 
 	return 0;
@@ -1646,9 +1791,6 @@ static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	or1 = reg_a(insn->dst_reg * 2);
 	or2 = reg_b(insn->dst_reg * 2 + 1);
 
-	if (insn->off < 0) /* TODO */
-		return -EOPNOTSUPP;
-
 	if (imm & ~0U) {
 		tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
 		emit_alu(nfp_prog, imm_a(nfp_prog),
@@ -1695,9 +1837,6 @@ static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	u64 imm = insn->imm; /* sign extend */
 	swreg tmp_reg;
 
-	if (insn->off < 0) /* TODO */
-		return -EOPNOTSUPP;
-
 	if (!imm) {
 		meta->skip = true;
 		return 0;
@@ -1726,9 +1865,6 @@ static int jne_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	u64 imm = insn->imm; /* sign extend */
 	swreg tmp_reg;
 
-	if (insn->off < 0) /* TODO */
-		return -EOPNOTSUPP;
-
 	if (!imm) {
 		emit_alu(nfp_prog, reg_none(), reg_a(insn->dst_reg * 2),
 			 ALU_OP_OR, reg_b(insn->dst_reg * 2 + 1));
@@ -1753,9 +1889,6 @@ static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	const struct bpf_insn *insn = &meta->insn;
 
-	if (insn->off < 0) /* TODO */
-		return -EOPNOTSUPP;
-
 	emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(insn->dst_reg * 2),
 		 ALU_OP_XOR, reg_b(insn->src_reg * 2));
 	emit_alu(nfp_prog, imm_b(nfp_prog), reg_a(insn->dst_reg * 2 + 1),
@@ -1887,17 +2020,22 @@ static void br_set_offset(u64 *instr, u16 offset)
 /* --- Assembler logic --- */
 static int nfp_fixup_branches(struct nfp_prog *nfp_prog)
 {
-	struct nfp_insn_meta *meta, *next;
-	u32 off, br_idx;
-	u32 idx;
+	struct nfp_insn_meta *meta, *jmp_dst;
+	u32 idx, br_idx;
 
-	nfp_for_each_insn_walk2(nfp_prog, meta, next) {
+	list_for_each_entry(meta, &nfp_prog->insns, l) {
 		if (meta->skip)
 			continue;
 		if (BPF_CLASS(meta->insn.code) != BPF_JMP)
 			continue;
 
-		br_idx = nfp_prog_offset_to_index(nfp_prog, next->off) - 1;
+		if (list_is_last(&meta->l, &nfp_prog->insns))
+			idx = nfp_prog->last_bpf_off;
+		else
+			idx = list_next_entry(meta, l)->off - 1;
+
+		br_idx = nfp_prog_offset_to_index(nfp_prog, idx);
+
 		if (!nfp_is_br(nfp_prog->prog[br_idx])) {
 			pr_err("Fixup found block not ending in branch %d %02x %016llx!!\n",
 			       br_idx, meta->insn.code, nfp_prog->prog[br_idx]);
@@ -1907,23 +2045,14 @@ static int nfp_fixup_branches(struct nfp_prog *nfp_prog)
 		if (FIELD_GET(OP_BR_SPECIAL, nfp_prog->prog[br_idx]))
 			continue;
 
-		/* Find the target offset in assembler realm */
-		off = meta->insn.off;
-		if (!off) {
-			pr_err("Fixup found zero offset!!\n");
+		if (!meta->jmp_dst) {
+			pr_err("Non-exit jump doesn't have destination info recorded!!\n");
 			return -ELOOP;
 		}
 
-		while (off && nfp_meta_has_next(nfp_prog, next)) {
-			next = nfp_meta_next(next);
-			off--;
-		}
-		if (off) {
-			pr_err("Fixup found too large jump!! %d\n", off);
-			return -ELOOP;
-		}
+		jmp_dst = meta->jmp_dst;
 
-		if (next->skip) {
+		if (jmp_dst->skip) {
 			pr_err("Branch landing on removed instruction!!\n");
 			return -ELOOP;
 		}
@@ -1932,7 +2061,7 @@ static int nfp_fixup_branches(struct nfp_prog *nfp_prog)
 		     idx <= br_idx; idx++) {
 			if (!nfp_is_br(nfp_prog->prog[idx]))
 				continue;
-			br_set_offset(&nfp_prog->prog[idx], next->off);
+			br_set_offset(&nfp_prog->prog[idx], jmp_dst->off);
 		}
 	}
 
@@ -2105,6 +2234,8 @@ static int nfp_translate(struct nfp_prog *nfp_prog)
 		nfp_prog->n_translated++;
 	}
 
+	nfp_prog->last_bpf_off = nfp_prog_current_offset(nfp_prog) - 1;
+
 	nfp_outro(nfp_prog);
 	if (nfp_prog->error)
 		return nfp_prog->error;
@@ -2173,6 +2304,9 @@ static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog)
 		if (next.src_reg || next.dst_reg)
 			continue;
 
+		if (meta2->flags & FLAG_INSN_IS_JUMP_DST)
+			continue;
+
 		meta2->skip = true;
 	}
 }
@@ -2209,17 +2343,258 @@ static void nfp_bpf_opt_ld_shift(struct nfp_prog *nfp_prog)
 		if (next1.imm != 0x20 || next2.imm != 0x20)
 			continue;
 
+		if (meta2->flags & FLAG_INSN_IS_JUMP_DST ||
+		    meta3->flags & FLAG_INSN_IS_JUMP_DST)
+			continue;
+
 		meta2->skip = true;
 		meta3->skip = true;
 	}
 }
 
+/* load/store pair that forms memory copy sould look like the following:
+ *
+ *   ld_width R, [addr_src + offset_src]
+ *   st_width [addr_dest + offset_dest], R
+ *
+ * The destination register of load and source register of store should
+ * be the same, load and store should also perform at the same width.
+ * If either of addr_src or addr_dest is stack pointer, we don't do the
+ * CPP optimization as stack is modelled by registers on NFP.
+ */
+static bool
+curr_pair_is_memcpy(struct nfp_insn_meta *ld_meta,
+		    struct nfp_insn_meta *st_meta)
+{
+	struct bpf_insn *ld = &ld_meta->insn;
+	struct bpf_insn *st = &st_meta->insn;
+
+	if (!is_mbpf_load(ld_meta) || !is_mbpf_store(st_meta))
+		return false;
+
+	if (ld_meta->ptr.type != PTR_TO_PACKET)
+		return false;
+
+	if (st_meta->ptr.type != PTR_TO_PACKET)
+		return false;
+
+	if (BPF_SIZE(ld->code) != BPF_SIZE(st->code))
+		return false;
+
+	if (ld->dst_reg != st->src_reg)
+		return false;
+
+	/* There is jump to the store insn in this pair. */
+	if (st_meta->flags & FLAG_INSN_IS_JUMP_DST)
+		return false;
+
+	return true;
+}
+
+/* Currently, we only support chaining load/store pairs if:
+ *
+ *  - Their address base registers are the same.
+ *  - Their address offsets are in the same order.
+ *  - They operate at the same memory width.
+ *  - There is no jump into the middle of them.
+ */
+static bool
+curr_pair_chain_with_previous(struct nfp_insn_meta *ld_meta,
+			      struct nfp_insn_meta *st_meta,
+			      struct bpf_insn *prev_ld,
+			      struct bpf_insn *prev_st)
+{
+	u8 prev_size, curr_size, prev_ld_base, prev_st_base, prev_ld_dst;
+	struct bpf_insn *ld = &ld_meta->insn;
+	struct bpf_insn *st = &st_meta->insn;
+	s16 prev_ld_off, prev_st_off;
+
+	/* This pair is the start pair. */
+	if (!prev_ld)
+		return true;
+
+	prev_size = BPF_LDST_BYTES(prev_ld);
+	curr_size = BPF_LDST_BYTES(ld);
+	prev_ld_base = prev_ld->src_reg;
+	prev_st_base = prev_st->dst_reg;
+	prev_ld_dst = prev_ld->dst_reg;
+	prev_ld_off = prev_ld->off;
+	prev_st_off = prev_st->off;
+
+	if (ld->dst_reg != prev_ld_dst)
+		return false;
+
+	if (ld->src_reg != prev_ld_base || st->dst_reg != prev_st_base)
+		return false;
+
+	if (curr_size != prev_size)
+		return false;
+
+	/* There is jump to the head of this pair. */
+	if (ld_meta->flags & FLAG_INSN_IS_JUMP_DST)
+		return false;
+
+	/* Both in ascending order. */
+	if (prev_ld_off + prev_size == ld->off &&
+	    prev_st_off + prev_size == st->off)
+		return true;
+
+	/* Both in descending order. */
+	if (ld->off + curr_size == prev_ld_off &&
+	    st->off + curr_size == prev_st_off)
+		return true;
+
+	return false;
+}
+
+/* Return TRUE if cross memory access happens. Cross memory access means
+ * store area is overlapping with load area that a later load might load
+ * the value from previous store, for this case we can't treat the sequence
+ * as an memory copy.
+ */
+static bool
+cross_mem_access(struct bpf_insn *ld, struct nfp_insn_meta *head_ld_meta,
+		 struct nfp_insn_meta *head_st_meta)
+{
+	s16 head_ld_off, head_st_off, ld_off;
+
+	/* Different pointer types does not overlap. */
+	if (head_ld_meta->ptr.type != head_st_meta->ptr.type)
+		return false;
+
+	/* load and store are both PTR_TO_PACKET, check ID info.  */
+	if (head_ld_meta->ptr.id != head_st_meta->ptr.id)
+		return true;
+
+	/* Canonicalize the offsets. Turn all of them against the original
+	 * base register.
+	 */
+	head_ld_off = head_ld_meta->insn.off + head_ld_meta->ptr.off;
+	head_st_off = head_st_meta->insn.off + head_st_meta->ptr.off;
+	ld_off = ld->off + head_ld_meta->ptr.off;
+
+	/* Ascending order cross. */
+	if (ld_off > head_ld_off &&
+	    head_ld_off < head_st_off && ld_off >= head_st_off)
+		return true;
+
+	/* Descending order cross. */
+	if (ld_off < head_ld_off &&
+	    head_ld_off > head_st_off && ld_off <= head_st_off)
+		return true;
+
+	return false;
+}
+
+/* This pass try to identify the following instructoin sequences.
+ *
+ *   load R, [regA + offA]
+ *   store [regB + offB], R
+ *   load R, [regA + offA + const_imm_A]
+ *   store [regB + offB + const_imm_A], R
+ *   load R, [regA + offA + 2 * const_imm_A]
+ *   store [regB + offB + 2 * const_imm_A], R
+ *   ...
+ *
+ * Above sequence is typically generated by compiler when lowering
+ * memcpy. NFP prefer using CPP instructions to accelerate it.
+ */
+static void nfp_bpf_opt_ldst_gather(struct nfp_prog *nfp_prog)
+{
+	struct nfp_insn_meta *head_ld_meta = NULL;
+	struct nfp_insn_meta *head_st_meta = NULL;
+	struct nfp_insn_meta *meta1, *meta2;
+	struct bpf_insn *prev_ld = NULL;
+	struct bpf_insn *prev_st = NULL;
+	u8 count = 0;
+
+	nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) {
+		struct bpf_insn *ld = &meta1->insn;
+		struct bpf_insn *st = &meta2->insn;
+
+		/* Reset record status if any of the following if true:
+		 *   - The current insn pair is not load/store.
+		 *   - The load/store pair doesn't chain with previous one.
+		 *   - The chained load/store pair crossed with previous pair.
+		 *   - The chained load/store pair has a total size of memory
+		 *     copy beyond 128 bytes which is the maximum length a
+		 *     single NFP CPP command can transfer.
+		 */
+		if (!curr_pair_is_memcpy(meta1, meta2) ||
+		    !curr_pair_chain_with_previous(meta1, meta2, prev_ld,
+						   prev_st) ||
+		    (head_ld_meta && (cross_mem_access(ld, head_ld_meta,
+						       head_st_meta) ||
+				      head_ld_meta->ldst_gather_len >= 128))) {
+			if (!count)
+				continue;
+
+			if (count > 1) {
+				s16 prev_ld_off = prev_ld->off;
+				s16 prev_st_off = prev_st->off;
+				s16 head_ld_off = head_ld_meta->insn.off;
+
+				if (prev_ld_off < head_ld_off) {
+					head_ld_meta->insn.off = prev_ld_off;
+					head_st_meta->insn.off = prev_st_off;
+					head_ld_meta->ldst_gather_len =
+						-head_ld_meta->ldst_gather_len;
+				}
+
+				head_ld_meta->paired_st = &head_st_meta->insn;
+				head_st_meta->skip = true;
+			} else {
+				head_ld_meta->ldst_gather_len = 0;
+			}
+
+			/* If the chain is ended by an load/store pair then this
+			 * could serve as the new head of the the next chain.
+			 */
+			if (curr_pair_is_memcpy(meta1, meta2)) {
+				head_ld_meta = meta1;
+				head_st_meta = meta2;
+				head_ld_meta->ldst_gather_len =
+					BPF_LDST_BYTES(ld);
+				meta1 = nfp_meta_next(meta1);
+				meta2 = nfp_meta_next(meta2);
+				prev_ld = ld;
+				prev_st = st;
+				count = 1;
+			} else {
+				head_ld_meta = NULL;
+				head_st_meta = NULL;
+				prev_ld = NULL;
+				prev_st = NULL;
+				count = 0;
+			}
+
+			continue;
+		}
+
+		if (!head_ld_meta) {
+			head_ld_meta = meta1;
+			head_st_meta = meta2;
+		} else {
+			meta1->skip = true;
+			meta2->skip = true;
+		}
+
+		head_ld_meta->ldst_gather_len += BPF_LDST_BYTES(ld);
+		meta1 = nfp_meta_next(meta1);
+		meta2 = nfp_meta_next(meta2);
+		prev_ld = ld;
+		prev_st = st;
+		count++;
+	}
+}
+
 static int nfp_bpf_optimize(struct nfp_prog *nfp_prog)
 {
 	nfp_bpf_opt_reg_init(nfp_prog);
 
 	nfp_bpf_opt_ld_mask(nfp_prog);
 	nfp_bpf_opt_ld_shift(nfp_prog);
+	nfp_bpf_opt_ldst_gather(nfp_prog);
 
 	return 0;
 }

drivers/net/ethernet/netronome/nfp/bpf/main.h

 /*
- * Copyright (C) 2016 Netronome Systems, Inc.
+ * Copyright (C) 2016-2017 Netronome Systems, Inc.
  *
  * This software is dual licensed under the GNU General License Version 2,
  * June 1991 as shown in the file COPYING in the top-level directory of this
@@ -89,23 +89,37 @@ typedef int (*instr_cb_t)(struct nfp_prog *, struct nfp_insn_meta *);
 #define nfp_meta_next(meta)	list_next_entry(meta, l)
 #define nfp_meta_prev(meta)	list_prev_entry(meta, l)
 
+#define FLAG_INSN_IS_JUMP_DST	BIT(0)
+
 /**
  * struct nfp_insn_meta - BPF instruction wrapper
  * @insn: BPF instruction
  * @ptr: pointer type for memory operations
+ * @ldst_gather_len: memcpy length gathered from load/store sequence
+ * @paired_st: the paired store insn at the head of the sequence
  * @ptr_not_const: pointer is not always constant
+ * @jmp_dst: destination info for jump instructions
  * @off: index of first generated machine instruction (in nfp_prog.prog)
  * @n: eBPF instruction number
+ * @flags: eBPF instruction extra optimization flags
  * @skip: skip this instruction (optimized out)
  * @double_cb: callback for second part of the instruction
  * @l: link on nfp_prog->insns list
  */
 struct nfp_insn_meta {
 	struct bpf_insn insn;
-	struct bpf_reg_state ptr;
-	bool ptr_not_const;
+	union {
+		struct {
+			struct bpf_reg_state ptr;
+			struct bpf_insn *paired_st;
+			s16 ldst_gather_len;
+			bool ptr_not_const;
+		};
+		struct nfp_insn_meta *jmp_dst;
+	};
 	unsigned int off;
 	unsigned short n;
+	unsigned short flags;
 	bool skip;
 	instr_cb_t double_cb;
 
@@ -134,6 +148,16 @@ static inline u8 mbpf_mode(const struct nfp_insn_meta *meta)
 	return BPF_MODE(meta->insn.code);
 }
 
+static inline bool is_mbpf_load(const struct nfp_insn_meta *meta)
+{
+	return (meta->insn.code & ~BPF_SIZE_MASK) == (BPF_LDX | BPF_MEM);
+}
+
+static inline bool is_mbpf_store(const struct nfp_insn_meta *meta)
+{
+	return (meta->insn.code & ~BPF_SIZE_MASK) == (BPF_STX | BPF_MEM);
+}
+
 /**
  * struct nfp_prog - nfp BPF program
  * @prog: machine code
@@ -142,6 +166,7 @@ static inline u8 mbpf_mode(const struct nfp_insn_meta *meta)
  * @verifier_meta: temporary storage for verifier's insn meta
  * @type: BPF program type
  * @start_off: address of the first instruction in the memory
+ * @last_bpf_off: address of the last instruction translated from BPF
  * @tgt_out: jump target for normal exit
  * @tgt_abort: jump target for abort (e.g. access outside of packet buffer)
  * @tgt_done: jump target to get the next packet
@@ -160,6 +185,7 @@ struct nfp_prog {
 	enum bpf_prog_type type;
 
 	unsigned int start_off;
+	unsigned int last_bpf_off;
 	unsigned int tgt_out;
 	unsigned int tgt_abort;
 	unsigned int tgt_done;
@@ -189,4 +215,7 @@ int nfp_bpf_translate(struct nfp_app *app, struct nfp_net *nn,
 		      struct bpf_prog *prog);
 int nfp_bpf_destroy(struct nfp_app *app, struct nfp_net *nn,
 		    struct bpf_prog *prog);
+struct nfp_insn_meta *
+nfp_bpf_goto_meta(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
+		  unsigned int insn_idx, unsigned int n_insns);
 #endif

drivers/net/ethernet/netronome/nfp/bpf/offload.c

 /*
- * Copyright (C) 2016 Netronome Systems, Inc.
+ * Copyright (C) 2016-2017 Netronome Systems, Inc.
  *
  * This software is dual licensed under the GNU General License Version 2,
  * June 1991 as shown in the file COPYING in the top-level directory of this
@@ -55,11 +55,10 @@ static int
 nfp_prog_prepare(struct nfp_prog *nfp_prog, const struct bpf_insn *prog,
 		 unsigned int cnt)
 {
+	struct nfp_insn_meta *meta;
 	unsigned int i;
 
 	for (i = 0; i < cnt; i++) {
-		struct nfp_insn_meta *meta;
-
 		meta = kzalloc(sizeof(*meta), GFP_KERNEL);
 		if (!meta)
 			return -ENOMEM;
@@ -70,6 +69,24 @@ nfp_prog_prepare(struct nfp_prog *nfp_prog, const struct bpf_insn *prog,
 		list_add_tail(&meta->l, &nfp_prog->insns);
 	}
 
+	/* Another pass to record jump information. */
+	list_for_each_entry(meta, &nfp_prog->insns, l) {
+		u64 code = meta->insn.code;
+
+		if (BPF_CLASS(code) == BPF_JMP && BPF_OP(code) != BPF_EXIT &&
+		    BPF_OP(code) != BPF_CALL) {
+			struct nfp_insn_meta *dst_meta;
+			unsigned short dst_indx;
+
+			dst_indx = meta->n + 1 + meta->insn.off;
+			dst_meta = nfp_bpf_goto_meta(nfp_prog, meta, dst_indx,
+						     cnt);
+
+			meta->jmp_dst = dst_meta;
+			dst_meta->flags |= FLAG_INSN_IS_JUMP_DST;
+		}
+	}
+
 	return 0;
 }
 

drivers/net/ethernet/netronome/nfp/bpf/verifier.c

 /*
- * Copyright (C) 2016 Netronome Systems, Inc.
+ * Copyright (C) 2016-2017 Netronome Systems, Inc.
  *
  * This software is dual licensed under the GNU General License Version 2,
  * June 1991 as shown in the file COPYING in the top-level directory of this
@@ -40,7 +40,7 @@
 
 #include "main.h"
 
-static struct nfp_insn_meta *
+struct nfp_insn_meta *
 nfp_bpf_goto_meta(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 		  unsigned int insn_idx, unsigned int n_insns)
 {
@@ -180,10 +180,10 @@ nfp_verify_insn(struct bpf_verifier_env *env, int insn_idx, int prev_insn_idx)
 	if (meta->insn.code == (BPF_JMP | BPF_EXIT))
 		return nfp_bpf_check_exit(nfp_prog, env);
 
-	if ((meta->insn.code & ~BPF_SIZE_MASK) == (BPF_LDX | BPF_MEM))
+	if (is_mbpf_load(meta))
 		return nfp_bpf_check_ptr(nfp_prog, meta, env,
 					 meta->insn.src_reg);
-	if ((meta->insn.code & ~BPF_SIZE_MASK) == (BPF_STX | BPF_MEM))
+	if (is_mbpf_store(meta))
 		return nfp_bpf_check_ptr(nfp_prog, meta, env,
 					 meta->insn.dst_reg);
 

drivers/net/ethernet/netronome/nfp/nfp_asm.c

@@ -41,6 +41,7 @@
 
 const struct cmd_tgt_act cmd_tgt_act[__CMD_TGT_MAP_SIZE] = {
 	[CMD_TGT_WRITE8_SWAP] =		{ 0x02, 0x42 },
+	[CMD_TGT_WRITE32_SWAP] =	{ 0x02, 0x5f },
 	[CMD_TGT_READ8] =		{ 0x01, 0x43 },
 	[CMD_TGT_READ32] =		{ 0x00, 0x5c },
 	[CMD_TGT_READ32_LE] =		{ 0x01, 0x5c },
@@ -120,7 +121,8 @@ int swreg_to_unrestricted(swreg dst, swreg lreg, swreg rreg,
 	reg->dst = nfp_swreg_to_unreg(dst, true);
 
 	/* Decode source operands */
-	if (swreg_type(lreg) == swreg_type(rreg))
+	if (swreg_type(lreg) == swreg_type(rreg) &&
+	    swreg_type(lreg) != NN_REG_NONE)
 		return -EFAULT;
 
 	if (swreg_type(lreg) == NN_REG_GPR_B ||
@@ -200,7 +202,8 @@ int swreg_to_restricted(swreg dst, swreg lreg, swreg rreg,
 	reg->dst = nfp_swreg_to_rereg(dst, true, false, NULL);
 
 	/* Decode source operands */
-	if (swreg_type(lreg) == swreg_type(rreg))
+	if (swreg_type(lreg) == swreg_type(rreg) &&
+	    swreg_type(lreg) != NN_REG_NONE)
 		return -EFAULT;
 
 	if (swreg_type(lreg) == NN_REG_GPR_B ||

drivers/net/ethernet/netronome/nfp/nfp_asm.h

 /*
- * Copyright (C) 2016 Netronome Systems, Inc.
+ * Copyright (C) 2016-2017 Netronome Systems, Inc.
  *
  * This software is dual licensed under the GNU General License Version 2,
  * June 1991 as shown in the file COPYING in the top-level directory of this
@@ -209,6 +209,7 @@ enum alu_dst_ab {
 #define OP_CMD_CNT		0x0000e000000ULL
 #define OP_CMD_SIG		0x000f0000000ULL
 #define OP_CMD_TGT_CMD		0x07f00000000ULL
+#define OP_CMD_INDIR		0x20000000000ULL
 #define OP_CMD_MODE	       0x1c0000000000ULL
 
 struct cmd_tgt_act {
@@ -219,6 +220,7 @@ struct cmd_tgt_act {
 enum cmd_tgt_map {
 	CMD_TGT_READ8,
 	CMD_TGT_WRITE8_SWAP,
+	CMD_TGT_WRITE32_SWAP,
 	CMD_TGT_READ32,
 	CMD_TGT_READ32_LE,
 	CMD_TGT_READ32_SWAP,
@@ -240,6 +242,9 @@ enum cmd_ctx_swap {
 	CMD_CTX_NO_SWAP = 3,
 };
 
+#define CMD_OVE_LEN	BIT(7)
+#define CMD_OV_LEN	GENMASK(12, 8)
+
 #define OP_LCSR_BASE		0x0fc00000000ULL
 #define OP_LCSR_A_SRC		0x000000003ffULL
 #define OP_LCSR_B_SRC		0x000000ffc00ULL

drivers/net/ethernet/netronome/nfp/nfp_net.h

@@ -548,6 +548,8 @@ struct nfp_net_dp {
  * @max_r_vecs:		Number of allocated interrupt vectors for RX/TX
  * @max_tx_rings:       Maximum number of TX rings supported by the Firmware
  * @max_rx_rings:       Maximum number of RX rings supported by the Firmware
+ * @stride_rx:		Queue controller RX queue spacing
+ * @stride_tx:		Queue controller TX queue spacing
  * @r_vecs:             Pre-allocated array of ring vectors
  * @irq_entries:        Pre-allocated array of MSI-X entries
  * @lsc_handler:        Handler for Link State Change interrupt

drivers/net/ethernet/netronome/nfp/nfpcore/nfp_cppcore.c

@@ -372,8 +372,7 @@ nfp_cpp_area_alloc(struct nfp_cpp *cpp, u32 dest,
  * that it can be accessed directly.
  *
  * NOTE: @address and @size must be 32-bit aligned values.
- *
- * NOTE: The area must also be 'released' when the structure is freed.
+ * The area must also be 'released' when the structure is freed.
  *
  * Return: NFP CPP Area handle, or NULL
  */
@@ -536,8 +535,7 @@ void nfp_cpp_area_release_free(struct nfp_cpp_area *area)
  * Read data from indicated CPP region.
  *
  * NOTE: @offset and @length must be 32-bit aligned values.
- *
- * NOTE: Area must have been locked down with an 'acquire'.
+ * Area must have been locked down with an 'acquire'.
  *
  * Return: length of io, or -ERRNO
  */
@@ -558,8 +556,7 @@ int nfp_cpp_area_read(struct nfp_cpp_area *area,
  * Write data to indicated CPP region.
  *
  * NOTE: @offset and @length must be 32-bit aligned values.
- *
- * NOTE: Area must have been locked down with an 'acquire'.
+ * Area must have been locked down with an 'acquire'.
  *
  * Return: length of io, or -ERRNO
  */