nfp: bpf: support logic indirect shifts (BPF_[L|R]SH | BPF_X)

For indirect shifts, shift amount is not specified as constant, NFP needs
to get the shift amount through the low 5 bits of source A operand in
PREV_ALU, therefore extra instructions are needed compared with shifts by
constants.

Because NFP is 32-bit, so we are using register pair for 64-bit shifts and
therefore would need different instruction sequences depending on whether
shift amount is less than 32 or not.

NFP branch-on-bit-test instruction emitter is added by this patch and is
used for efficient runtime check on shift amount. We'd think the shift
amount is less than 32 if bit 5 is clear and greater or equal than 32
otherwise. Shift amount is greater than or equal to 64 will result in
undefined behavior.

This patch also use range info to avoid generating unnecessary runtime code
if we are certain shift amount is less than 32 or not.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

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

@@ -211,6 +211,60 @@ emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer)
 	emit_br_relo(nfp_prog, mask, addr, defer, RELO_BR_REL);
 }
 
+static void
+__emit_br_bit(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 addr, u8 defer,
+	      bool set, bool src_lmextn)
+{
+	u16 addr_lo, addr_hi;
+	u64 insn;
+
+	addr_lo = addr & (OP_BR_BIT_ADDR_LO >> __bf_shf(OP_BR_BIT_ADDR_LO));
+	addr_hi = addr != addr_lo;
+
+	insn = OP_BR_BIT_BASE |
+		FIELD_PREP(OP_BR_BIT_A_SRC, areg) |
+		FIELD_PREP(OP_BR_BIT_B_SRC, breg) |
+		FIELD_PREP(OP_BR_BIT_BV, set) |
+		FIELD_PREP(OP_BR_BIT_DEFBR, defer) |
+		FIELD_PREP(OP_BR_BIT_ADDR_LO, addr_lo) |
+		FIELD_PREP(OP_BR_BIT_ADDR_HI, addr_hi) |
+		FIELD_PREP(OP_BR_BIT_SRC_LMEXTN, src_lmextn);
+
+	nfp_prog_push(nfp_prog, insn);
+}
+
+static void
+emit_br_bit_relo(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr,
+		 u8 defer, bool set, enum nfp_relo_type relo)
+{
+	struct nfp_insn_re_regs reg;
+	int err;
+
+	/* NOTE: The bit to test is specified as an rotation amount, such that
+	 *	 the bit to test will be placed on the MSB of the result when
+	 *	 doing a rotate right. For bit X, we need right rotate X + 1.
+	 */
+	bit += 1;
+
+	err = swreg_to_restricted(reg_none(), src, reg_imm(bit), &reg, false);
+	if (err) {
+		nfp_prog->error = err;
+		return;
+	}
+
+	__emit_br_bit(nfp_prog, reg.areg, reg.breg, addr, defer, set,
+		      reg.src_lmextn);
+
+	nfp_prog->prog[nfp_prog->prog_len - 1] |=
+		FIELD_PREP(OP_RELO_TYPE, relo);
+}
+
+static void
+emit_br_bset(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr, u8 defer)
+{
+	emit_br_bit_relo(nfp_prog, src, bit, addr, defer, true, RELO_BR_REL);
+}
+
 static void
 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi,
 	     enum immed_width width, bool invert,
@@ -309,6 +363,19 @@ emit_shf(struct nfp_prog *nfp_prog, swreg dst,
 		   reg.dst_lmextn, reg.src_lmextn);
 }
 
+static void
+emit_shf_indir(struct nfp_prog *nfp_prog, swreg dst,
+	       swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc)
+{
+	if (sc == SHF_SC_R_ROT) {
+		pr_err("indirect shift is not allowed on rotation\n");
+		nfp_prog->error = -EFAULT;
+		return;
+	}
+
+	emit_shf(nfp_prog, dst, lreg, op, rreg, sc, 0);
+}
+
 static void
 __emit_alu(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab,
 	   u16 areg, enum alu_op op, u16 breg, bool swap, bool wr_both,
@@ -1629,56 +1696,226 @@ static int neg_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	return 0;
 }
 
-static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	const struct bpf_insn *insn = &meta->insn;
-	u8 dst = insn->dst_reg * 2;
-
-	if (insn->imm < 32) {
-		emit_shf(nfp_prog, reg_both(dst + 1),
-			 reg_a(dst + 1), SHF_OP_NONE, reg_b(dst),
-			 SHF_SC_R_DSHF, 32 - insn->imm);
-		emit_shf(nfp_prog, reg_both(dst),
-			 reg_none(), SHF_OP_NONE, reg_b(dst),
-			 SHF_SC_L_SHF, insn->imm);
-	} else if (insn->imm == 32) {
+/* Pseudo code:
+ *   if shift_amt >= 32
+ *     dst_high = dst_low << shift_amt[4:0]
+ *     dst_low = 0;
+ *   else
+ *     dst_high = (dst_high, dst_low) >> (32 - shift_amt)
+ *     dst_low = dst_low << shift_amt
+ *
+ * The indirect shift will use the same logic at runtime.
+ */
+static int __shl_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt)
+{
+	if (shift_amt < 32) {
+		emit_shf(nfp_prog, reg_both(dst + 1), reg_a(dst + 1),
+			 SHF_OP_NONE, reg_b(dst), SHF_SC_R_DSHF,
+			 32 - shift_amt);
+		emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
+			 reg_b(dst), SHF_SC_L_SHF, shift_amt);
+	} else if (shift_amt == 32) {
 		wrp_reg_mov(nfp_prog, dst + 1, dst);
 		wrp_immed(nfp_prog, reg_both(dst), 0);
-	} else if (insn->imm > 32) {
-		emit_shf(nfp_prog, reg_both(dst + 1),
-			 reg_none(), SHF_OP_NONE, reg_b(dst),
-			 SHF_SC_L_SHF, insn->imm - 32);
+	} else if (shift_amt > 32) {
+		emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
+			 reg_b(dst), SHF_SC_L_SHF, shift_amt - 32);
 		wrp_immed(nfp_prog, reg_both(dst), 0);
 	}
 
 	return 0;
 }
 
-static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	const struct bpf_insn *insn = &meta->insn;
 	u8 dst = insn->dst_reg * 2;
 
-	if (insn->imm < 32) {
-		emit_shf(nfp_prog, reg_both(dst),
-			 reg_a(dst + 1), SHF_OP_NONE, reg_b(dst),
-			 SHF_SC_R_DSHF, insn->imm);
-		emit_shf(nfp_prog, reg_both(dst + 1),
-			 reg_none(), SHF_OP_NONE, reg_b(dst + 1),
-			 SHF_SC_R_SHF, insn->imm);
-	} else if (insn->imm == 32) {
+	return __shl_imm64(nfp_prog, dst, insn->imm);
+}
+
+static void shl_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	emit_alu(nfp_prog, imm_both(nfp_prog), reg_imm(32), ALU_OP_SUB,
+		 reg_b(src));
+	emit_alu(nfp_prog, reg_none(), imm_a(nfp_prog), ALU_OP_OR, reg_imm(0));
+	emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_a(dst + 1), SHF_OP_NONE,
+		       reg_b(dst), SHF_SC_R_DSHF);
+}
+
+/* NOTE: for indirect left shift, HIGH part should be calculated first. */
+static void shl_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
+	emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
+		       reg_b(dst), SHF_SC_L_SHF);
+}
+
+static void shl_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	shl_reg64_lt32_high(nfp_prog, dst, src);
+	shl_reg64_lt32_low(nfp_prog, dst, src);
+}
+
+static void shl_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
+	emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
+		       reg_b(dst), SHF_SC_L_SHF);
+	wrp_immed(nfp_prog, reg_both(dst), 0);
+}
+
+static int shl_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	const struct bpf_insn *insn = &meta->insn;
+	u64 umin, umax;
+	u8 dst, src;
+
+	dst = insn->dst_reg * 2;
+	umin = meta->umin;
+	umax = meta->umax;
+	if (umin == umax)
+		return __shl_imm64(nfp_prog, dst, umin);
+
+	src = insn->src_reg * 2;
+	if (umax < 32) {
+		shl_reg64_lt32(nfp_prog, dst, src);
+	} else if (umin >= 32) {
+		shl_reg64_ge32(nfp_prog, dst, src);
+	} else {
+		/* Generate different instruction sequences depending on runtime
+		 * value of shift amount.
+		 */
+		u16 label_ge32, label_end;
+
+		label_ge32 = nfp_prog_current_offset(nfp_prog) + 7;
+		emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0);
+
+		shl_reg64_lt32_high(nfp_prog, dst, src);
+		label_end = nfp_prog_current_offset(nfp_prog) + 6;
+		emit_br(nfp_prog, BR_UNC, label_end, 2);
+		/* shl_reg64_lt32_low packed in delay slot. */
+		shl_reg64_lt32_low(nfp_prog, dst, src);
+
+		if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32))
+			return -EINVAL;
+		shl_reg64_ge32(nfp_prog, dst, src);
+
+		if (!nfp_prog_confirm_current_offset(nfp_prog, label_end))
+			return -EINVAL;
+	}
+
+	return 0;
+}
+
+/* Pseudo code:
+ *   if shift_amt >= 32
+ *     dst_high = 0;
+ *     dst_low = dst_high >> shift_amt[4:0]
+ *   else
+ *     dst_high = dst_high >> shift_amt
+ *     dst_low = (dst_high, dst_low) >> shift_amt
+ *
+ * The indirect shift will use the same logic at runtime.
+ */
+static int __shr_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt)
+{
+	if (shift_amt < 32) {
+		emit_shf(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE,
+			 reg_b(dst), SHF_SC_R_DSHF, shift_amt);
+		emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
+			 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt);
+	} else if (shift_amt == 32) {
 		wrp_reg_mov(nfp_prog, dst, dst + 1);
 		wrp_immed(nfp_prog, reg_both(dst + 1), 0);
-	} else if (insn->imm > 32) {
-		emit_shf(nfp_prog, reg_both(dst),
-			 reg_none(), SHF_OP_NONE, reg_b(dst + 1),
-			 SHF_SC_R_SHF, insn->imm - 32);
+	} else if (shift_amt > 32) {
+		emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
+			 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt - 32);
 		wrp_immed(nfp_prog, reg_both(dst + 1), 0);
 	}
 
 	return 0;
 }
 
+static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	const struct bpf_insn *insn = &meta->insn;
+	u8 dst = insn->dst_reg * 2;
+
+	return __shr_imm64(nfp_prog, dst, insn->imm);
+}
+
+/* NOTE: for indirect right shift, LOW part should be calculated first. */
+static void shr_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
+	emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
+		       reg_b(dst + 1), SHF_SC_R_SHF);
+}
+
+static void shr_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
+	emit_shf_indir(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE,
+		       reg_b(dst), SHF_SC_R_DSHF);
+}
+
+static void shr_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	shr_reg64_lt32_low(nfp_prog, dst, src);
+	shr_reg64_lt32_high(nfp_prog, dst, src);
+}
+
+static void shr_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
+{
+	emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
+	emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
+		       reg_b(dst + 1), SHF_SC_R_SHF);
+	wrp_immed(nfp_prog, reg_both(dst + 1), 0);
+}
+
+static int shr_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	const struct bpf_insn *insn = &meta->insn;
+	u64 umin, umax;
+	u8 dst, src;
+
+	dst = insn->dst_reg * 2;
+	umin = meta->umin;
+	umax = meta->umax;
+	if (umin == umax)
+		return __shr_imm64(nfp_prog, dst, umin);
+
+	src = insn->src_reg * 2;
+	if (umax < 32) {
+		shr_reg64_lt32(nfp_prog, dst, src);
+	} else if (umin >= 32) {
+		shr_reg64_ge32(nfp_prog, dst, src);
+	} else {
+		/* Generate different instruction sequences depending on runtime
+		 * value of shift amount.
+		 */
+		u16 label_ge32, label_end;
+
+		label_ge32 = nfp_prog_current_offset(nfp_prog) + 6;
+		emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0);
+		shr_reg64_lt32_low(nfp_prog, dst, src);
+		label_end = nfp_prog_current_offset(nfp_prog) + 6;
+		emit_br(nfp_prog, BR_UNC, label_end, 2);
+		/* shr_reg64_lt32_high packed in delay slot. */
+		shr_reg64_lt32_high(nfp_prog, dst, src);
+
+		if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32))
+			return -EINVAL;
+		shr_reg64_ge32(nfp_prog, dst, src);
+
+		if (!nfp_prog_confirm_current_offset(nfp_prog, label_end))
+			return -EINVAL;
+	}
+
+	return 0;
+}
+
 static int mov_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	const struct bpf_insn *insn = &meta->insn;
@@ -2501,7 +2738,9 @@ static const instr_cb_t instr_cb[256] = {
 	[BPF_ALU64 | BPF_SUB | BPF_X] =	sub_reg64,
 	[BPF_ALU64 | BPF_SUB | BPF_K] =	sub_imm64,
 	[BPF_ALU64 | BPF_NEG] =		neg_reg64,
+	[BPF_ALU64 | BPF_LSH | BPF_X] =	shl_reg64,
 	[BPF_ALU64 | BPF_LSH | BPF_K] =	shl_imm64,
+	[BPF_ALU64 | BPF_RSH | BPF_X] =	shr_reg64,
 	[BPF_ALU64 | BPF_RSH | BPF_K] =	shr_imm64,
 	[BPF_ALU | BPF_MOV | BPF_X] =	mov_reg,
 	[BPF_ALU | BPF_MOV | BPF_K] =	mov_imm,

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

@@ -263,6 +263,8 @@ struct nfp_bpf_reg_state {
  * @func_id: function id for call instructions
  * @arg1: arg1 for call instructions
  * @arg2: arg2 for call instructions
+ * @umin: copy of core verifier umin_value.
+ * @umax: copy of core verifier umax_value.
  * @off: index of first generated machine instruction (in nfp_prog.prog)
  * @n: eBPF instruction number
  * @flags: eBPF instruction extra optimization flags
@@ -298,6 +300,13 @@ struct nfp_insn_meta {
 			struct bpf_reg_state arg1;
 			struct nfp_bpf_reg_state arg2;
 		};
+		/* We are interested in range info for some operands,
+		 * for example, the shift amount.
+		 */
+		struct {
+			u64 umin;
+			u64 umax;
+		};
 	};
 	unsigned int off;
 	unsigned short n;
@@ -375,6 +384,25 @@ static inline bool is_mbpf_xadd(const struct nfp_insn_meta *meta)
 	return (meta->insn.code & ~BPF_SIZE_MASK) == (BPF_STX | BPF_XADD);
 }
 
+static inline bool is_mbpf_indir_shift(const struct nfp_insn_meta *meta)
+{
+	u8 code = meta->insn.code;
+	bool is_alu, is_shift;
+	u8 opclass, opcode;
+
+	opclass = BPF_CLASS(code);
+	is_alu = opclass == BPF_ALU64 || opclass == BPF_ALU;
+	if (!is_alu)
+		return false;
+
+	opcode = BPF_OP(code);
+	is_shift = opcode == BPF_LSH || opcode == BPF_RSH || opcode == BPF_ARSH;
+	if (!is_shift)
+		return false;
+
+	return BPF_SRC(code) == BPF_X;
+}
+
 /**
  * struct nfp_prog - nfp BPF program
  * @bpf: backpointer to the bpf app priv structure

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

@@ -190,6 +190,8 @@ nfp_prog_prepare(struct nfp_prog *nfp_prog, const struct bpf_insn *prog,
 
 		meta->insn = prog[i];
 		meta->n = i;
+		if (is_mbpf_indir_shift(meta))
+			meta->umin = U64_MAX;
 
 		list_add_tail(&meta->l, &nfp_prog->insns);
 	}

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

@@ -551,6 +551,14 @@ nfp_verify_insn(struct bpf_verifier_env *env, int insn_idx, int prev_insn_idx)
 	if (is_mbpf_xadd(meta))
 		return nfp_bpf_check_xadd(nfp_prog, meta, env);
 
+	if (is_mbpf_indir_shift(meta)) {
+		const struct bpf_reg_state *sreg =
+			cur_regs(env) + meta->insn.src_reg;
+
+		meta->umin = min(meta->umin, sreg->umin_value);
+		meta->umax = max(meta->umax, sreg->umax_value);
+	}
+
 	return 0;
 }
 

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

@@ -72,8 +72,21 @@
 #define OP_BR_ADDR_LO		0x007ffc00000ULL
 #define OP_BR_ADDR_HI		0x10000000000ULL
 
-#define nfp_is_br(_insn)				\
-	(((_insn) & OP_BR_BASE_MASK) == OP_BR_BASE)
+#define OP_BR_BIT_BASE		0x0d000000000ULL
+#define OP_BR_BIT_BASE_MASK	0x0f800080300ULL
+#define OP_BR_BIT_A_SRC		0x000000000ffULL
+#define OP_BR_BIT_B_SRC		0x0000003fc00ULL
+#define OP_BR_BIT_BV		0x00000040000ULL
+#define OP_BR_BIT_SRC_LMEXTN	0x40000000000ULL
+#define OP_BR_BIT_DEFBR		OP_BR_DEFBR
+#define OP_BR_BIT_ADDR_LO	OP_BR_ADDR_LO
+#define OP_BR_BIT_ADDR_HI	OP_BR_ADDR_HI
+
+static inline bool nfp_is_br(u64 insn)
+{
+	return (insn & OP_BR_BASE_MASK) == OP_BR_BASE ||
+	       (insn & OP_BR_BIT_BASE_MASK) == OP_BR_BIT_BASE;
+}
 
 enum br_mask {
 	BR_BEQ = 0x00,