nfp: bpf: add main logics for BPF-to-BPF calls support in nfp driver

This is the main patch for the logics of BPF-to-BPF calls in the nfp
driver.

The functions called on BPF_JUMP | BPF_CALL and BPF_JUMP | BPF_EXIT were
used to call helpers and exit from the program, respectively; make them
usable for calling into, or returning from, a BPF subprogram as well.

For all calls, push the return address as well as the callee-saved
registers (R6 to R9) to the stack, and pop them upon returning from the
calls. In order to limit the overhead in terms of instruction number,
this is done through dedicated subroutines. Jumping to the callee
actually consists in jumping to the subroutine, that "returns" to the
callee: this will require some fixup for passing the address in a later
patch. Similarly, returning consists in jumping to the subroutine, which
pops registers and then return directly to the caller (but no fixup is
needed here).

Return to the caller is performed with the RTN instruction newly added
to the JIT.

For the few steps where we need to know what subprogram an instruction
belongs to, the struct nfp_insn_meta is extended with a new subprog_idx
field.

Note that checks on the available stack size, to take into account the
additional requirements associated to BPF-to-BPF calls (storing R6-R9
and return addresses), are added in a later patch.

Signed-off-by: Quentin Monnet <quentin.monnet@netronome.com>
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

@@ -266,6 +266,38 @@ 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_br_alu(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi,
+	      u8 defer, bool dst_lmextn, bool src_lmextn)
+{
+	u64 insn;
+
+	insn = OP_BR_ALU_BASE |
+		FIELD_PREP(OP_BR_ALU_A_SRC, areg) |
+		FIELD_PREP(OP_BR_ALU_B_SRC, breg) |
+		FIELD_PREP(OP_BR_ALU_DEFBR, defer) |
+		FIELD_PREP(OP_BR_ALU_IMM_HI, imm_hi) |
+		FIELD_PREP(OP_BR_ALU_SRC_LMEXTN, src_lmextn) |
+		FIELD_PREP(OP_BR_ALU_DST_LMEXTN, dst_lmextn);
+
+	nfp_prog_push(nfp_prog, insn);
+}
+
+static void emit_rtn(struct nfp_prog *nfp_prog, swreg base, u8 defer)
+{
+	struct nfp_insn_ur_regs reg;
+	int err;
+
+	err = swreg_to_unrestricted(reg_none(), base, reg_imm(0), &reg);
+	if (err) {
+		nfp_prog->error = err;
+		return;
+	}
+
+	__emit_br_alu(nfp_prog, reg.areg, reg.breg, 0, defer, reg.dst_lmextn,
+		      reg.src_lmextn);
+}
+
 static void
 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi,
 	     enum immed_width width, bool invert,
@@ -3081,7 +3113,73 @@ static int jne_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	return wrp_test_reg(nfp_prog, meta, ALU_OP_XOR, BR_BNE);
 }
 
-static int call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+static int
+bpf_to_bpf_call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	u32 ret_tgt, stack_depth;
+	swreg tmp_reg;
+
+	stack_depth = round_up(nfp_prog->stack_frame_depth, STACK_FRAME_ALIGN);
+	/* Space for saving the return address is accounted for by the callee,
+	 * so stack_depth can be zero for the main function.
+	 */
+	if (stack_depth) {
+		tmp_reg = ur_load_imm_any(nfp_prog, stack_depth,
+					  stack_imm(nfp_prog));
+		emit_alu(nfp_prog, stack_reg(nfp_prog),
+			 stack_reg(nfp_prog), ALU_OP_ADD, tmp_reg);
+		emit_csr_wr(nfp_prog, stack_reg(nfp_prog),
+			    NFP_CSR_ACT_LM_ADDR0);
+	}
+
+	/* The following steps are performed:
+	 *     1. Put the start offset of the callee into imm_b(). This will
+	 *        require a fixup step, as we do not necessarily know this
+	 *        address yet.
+	 *     2. Put the return address from the callee to the caller into
+	 *        register ret_reg().
+	 *     3. (After defer slots are consumed) Jump to the subroutine that
+	 *        pushes the registers to the stack.
+	 * The subroutine acts as a trampoline, and returns to the address in
+	 * imm_b(), i.e. jumps to the callee.
+	 *
+	 * Using ret_reg() to pass the return address to the callee is set here
+	 * as a convention. The callee can then push this address onto its
+	 * stack frame in its prologue. The advantages of passing the return
+	 * address through ret_reg(), instead of pushing it to the stack right
+	 * here, are the following:
+	 * - It looks cleaner.
+	 * - If the called function is called multiple time, we get a lower
+	 *   program size.
+	 * - We save two no-op instructions that should be added just before
+	 *   the emit_br() when stack depth is not null otherwise.
+	 * - If we ever find a register to hold the return address during whole
+	 *   execution of the callee, we will not have to push the return
+	 *   address to the stack for leaf functions.
+	 */
+	ret_tgt = nfp_prog_current_offset(nfp_prog) + 3;
+	emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2,
+		     RELO_BR_GO_CALL_PUSH_REGS);
+	wrp_immed_relo(nfp_prog, imm_b(nfp_prog), 0, RELO_IMMED_REL);
+	wrp_immed_relo(nfp_prog, ret_reg(nfp_prog), ret_tgt, RELO_IMMED_REL);
+
+	if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt))
+		return -EINVAL;
+
+	if (stack_depth) {
+		tmp_reg = ur_load_imm_any(nfp_prog, stack_depth,
+					  stack_imm(nfp_prog));
+		emit_alu(nfp_prog, stack_reg(nfp_prog),
+			 stack_reg(nfp_prog), ALU_OP_SUB, tmp_reg);
+		emit_csr_wr(nfp_prog, stack_reg(nfp_prog),
+			    NFP_CSR_ACT_LM_ADDR0);
+		wrp_nops(nfp_prog, 3);
+	}
+
+	return 0;
+}
+
+static int helper_call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	switch (meta->insn.imm) {
 	case BPF_FUNC_xdp_adjust_head:
@@ -3102,6 +3200,19 @@ static int call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	}
 }
 
+static int call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	if (is_mbpf_pseudo_call(meta))
+		return bpf_to_bpf_call(nfp_prog, meta);
+	else
+		return helper_call(nfp_prog, meta);
+}
+
+static bool nfp_is_main_function(struct nfp_insn_meta *meta)
+{
+	return meta->subprog_idx == 0;
+}
+
 static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 0, RELO_BR_GO_OUT);
@@ -3109,6 +3220,30 @@ static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	return 0;
 }
 
+static int
+nfp_subprog_epilogue(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	/* Pop R6~R9 to the stack via related subroutine.
+	 * Pop return address for BPF-to-BPF call from the stack and load it
+	 * into ret_reg() before we jump. This means that the subroutine does
+	 * not come back here, we make it jump back to the subprogram caller
+	 * directly!
+	 */
+	emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 1,
+		     RELO_BR_GO_CALL_POP_REGS);
+	wrp_mov(nfp_prog, ret_reg(nfp_prog), reg_lm(0, 0));
+
+	return 0;
+}
+
+static int jmp_exit(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	if (nfp_is_main_function(meta))
+		return goto_out(nfp_prog, meta);
+	else
+		return nfp_subprog_epilogue(nfp_prog, meta);
+}
+
 static const instr_cb_t instr_cb[256] = {
 	[BPF_ALU64 | BPF_MOV | BPF_X] =	mov_reg64,
 	[BPF_ALU64 | BPF_MOV | BPF_K] =	mov_imm64,
@@ -3197,7 +3332,7 @@ static const instr_cb_t instr_cb[256] = {
 	[BPF_JMP | BPF_JSET | BPF_X] =	jset_reg,
 	[BPF_JMP | BPF_JNE | BPF_X] =	jne_reg,
 	[BPF_JMP | BPF_CALL] =		call,
-	[BPF_JMP | BPF_EXIT] =		goto_out,
+	[BPF_JMP | BPF_EXIT] =		jmp_exit,
 };
 
 /* --- Assembler logic --- */
@@ -3258,6 +3393,27 @@ static void nfp_intro(struct nfp_prog *nfp_prog)
 		 plen_reg(nfp_prog), ALU_OP_AND, pv_len(nfp_prog));
 }
 
+static void
+nfp_subprog_prologue(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	/* Save return address into the stack. */
+	wrp_mov(nfp_prog, reg_lm(0, 0), ret_reg(nfp_prog));
+}
+
+static void
+nfp_start_subprog(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
+{
+	unsigned int depth = nfp_prog->subprog[meta->subprog_idx].stack_depth;
+
+	nfp_prog->stack_frame_depth = round_up(depth, 4);
+	nfp_subprog_prologue(nfp_prog, meta);
+}
+
+bool nfp_is_subprog_start(struct nfp_insn_meta *meta)
+{
+	return meta->flags & FLAG_INSN_IS_SUBPROG_START;
+}
+
 static void nfp_outro_tc_da(struct nfp_prog *nfp_prog)
 {
 	/* TC direct-action mode:
@@ -3348,6 +3504,56 @@ static void nfp_outro_xdp(struct nfp_prog *nfp_prog)
 	emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16);
 }
 
+static void nfp_push_callee_registers(struct nfp_prog *nfp_prog)
+{
+	u8 reg;
+
+	/* Subroutine: Save all callee saved registers (R6 ~ R9).
+	 * imm_b() holds the return address.
+	 */
+	nfp_prog->tgt_call_push_regs = nfp_prog_current_offset(nfp_prog);
+	for (reg = BPF_REG_6; reg <= BPF_REG_9; reg++) {
+		u8 adj = (reg - BPF_REG_0) * 2;
+		u8 idx = (reg - BPF_REG_6) * 2;
+
+		/* The first slot in the stack frame is used to push the return
+		 * address in bpf_to_bpf_call(), start just after.
+		 */
+		wrp_mov(nfp_prog, reg_lm(0, 1 + idx), reg_b(adj));
+
+		if (reg == BPF_REG_8)
+			/* Prepare to jump back, last 3 insns use defer slots */
+			emit_rtn(nfp_prog, imm_b(nfp_prog), 3);
+
+		wrp_mov(nfp_prog, reg_lm(0, 1 + idx + 1), reg_b(adj + 1));
+	}
+}
+
+static void nfp_pop_callee_registers(struct nfp_prog *nfp_prog)
+{
+	u8 reg;
+
+	/* Subroutine: Restore all callee saved registers (R6 ~ R9).
+	 * ret_reg() holds the return address.
+	 */
+	nfp_prog->tgt_call_pop_regs = nfp_prog_current_offset(nfp_prog);
+	for (reg = BPF_REG_6; reg <= BPF_REG_9; reg++) {
+		u8 adj = (reg - BPF_REG_0) * 2;
+		u8 idx = (reg - BPF_REG_6) * 2;
+
+		/* The first slot in the stack frame holds the return address,
+		 * start popping just after that.
+		 */
+		wrp_mov(nfp_prog, reg_both(adj), reg_lm(0, 1 + idx));
+
+		if (reg == BPF_REG_8)
+			/* Prepare to jump back, last 3 insns use defer slots */
+			emit_rtn(nfp_prog, ret_reg(nfp_prog), 3);
+
+		wrp_mov(nfp_prog, reg_both(adj + 1), reg_lm(0, 1 + idx + 1));
+	}
+}
+
 static void nfp_outro(struct nfp_prog *nfp_prog)
 {
 	switch (nfp_prog->type) {
@@ -3360,13 +3566,23 @@ static void nfp_outro(struct nfp_prog *nfp_prog)
 	default:
 		WARN_ON(1);
 	}
+
+	if (nfp_prog->subprog_cnt == 1)
+		return;
+
+	nfp_push_callee_registers(nfp_prog);
+	nfp_pop_callee_registers(nfp_prog);
 }
 
 static int nfp_translate(struct nfp_prog *nfp_prog)
 {
 	struct nfp_insn_meta *meta;
+	unsigned int depth;
 	int err;
 
+	depth = nfp_prog->subprog[0].stack_depth;
+	nfp_prog->stack_frame_depth = round_up(depth, 4);
+
 	nfp_intro(nfp_prog);
 	if (nfp_prog->error)
 		return nfp_prog->error;
@@ -3376,6 +3592,12 @@ static int nfp_translate(struct nfp_prog *nfp_prog)
 
 		meta->off = nfp_prog_current_offset(nfp_prog);
 
+		if (nfp_is_subprog_start(meta)) {
+			nfp_start_subprog(nfp_prog, meta);
+			if (nfp_prog->error)
+				return nfp_prog->error;
+		}
+
 		if (meta->skip) {
 			nfp_prog->n_translated++;
 			continue;
@@ -4069,6 +4291,7 @@ void *nfp_bpf_relo_for_vnic(struct nfp_prog *nfp_prog, struct nfp_bpf_vnic *bv)
 	for (i = 0; i < nfp_prog->prog_len; i++) {
 		enum nfp_relo_type special;
 		u32 val;
+		u16 off;
 
 		special = FIELD_GET(OP_RELO_TYPE, prog[i]);
 		switch (special) {
@@ -4085,6 +4308,14 @@ void *nfp_bpf_relo_for_vnic(struct nfp_prog *nfp_prog, struct nfp_bpf_vnic *bv)
 			br_set_offset(&prog[i],
 				      nfp_prog->tgt_abort + bv->start_off);
 			break;
+		case RELO_BR_GO_CALL_PUSH_REGS:
+			off = nfp_prog->tgt_call_push_regs + bv->start_off;
+			br_set_offset(&prog[i], off);
+			break;
+		case RELO_BR_GO_CALL_POP_REGS:
+			off = nfp_prog->tgt_call_pop_regs + bv->start_off;
+			br_set_offset(&prog[i], off);
+			break;
 		case RELO_BR_NEXT_PKT:
 			br_set_offset(&prog[i], bv->tgt_done);
 			break;

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

@@ -61,6 +61,8 @@ enum nfp_relo_type {
 	/* internal jumps to parts of the outro */
 	RELO_BR_GO_OUT,
 	RELO_BR_GO_ABORT,
+	RELO_BR_GO_CALL_PUSH_REGS,
+	RELO_BR_GO_CALL_POP_REGS,
 	/* external jumps to fixed addresses */
 	RELO_BR_NEXT_PKT,
 	RELO_BR_HELPER,
@@ -104,6 +106,7 @@ enum pkt_vec {
 #define imma_a(np)	reg_a(STATIC_REG_IMMA)
 #define imma_b(np)	reg_b(STATIC_REG_IMMA)
 #define imm_both(np)	reg_both(STATIC_REG_IMM)
+#define ret_reg(np)	imm_a(np)
 
 #define NFP_BPF_ABI_FLAGS	reg_imm(0)
 #define   NFP_BPF_ABI_FLAG_MARK	1
@@ -290,6 +293,7 @@ struct nfp_bpf_reg_state {
  * @off: index of first generated machine instruction (in nfp_prog.prog)
  * @n: eBPF instruction number
  * @flags: eBPF instruction extra optimization flags
+ * @subprog_idx: index of subprogram to which the instruction belongs
  * @skip: skip this instruction (optimized out)
  * @double_cb: callback for second part of the instruction
  * @l: link on nfp_prog->insns list
@@ -336,6 +340,7 @@ struct nfp_insn_meta {
 	unsigned int off;
 	unsigned short n;
 	unsigned short flags;
+	unsigned short subprog_idx;
 	bool skip;
 	instr_cb_t double_cb;
 
@@ -432,6 +437,16 @@ static inline bool is_mbpf_helper_call(const struct nfp_insn_meta *meta)
 		insn.src_reg != BPF_PSEUDO_CALL;
 }
 
+static inline bool is_mbpf_pseudo_call(const struct nfp_insn_meta *meta)
+{
+	struct bpf_insn insn = meta->insn;
+
+	return insn.code == (BPF_JMP | BPF_CALL) &&
+		insn.src_reg == BPF_PSEUDO_CALL;
+}
+
+#define STACK_FRAME_ALIGN 64
+
 /**
  * struct nfp_bpf_subprog_info - nfp BPF sub-program (a.k.a. function) info
  * @stack_depth:	maximum stack depth used by this sub-program
@@ -451,6 +466,8 @@ struct nfp_bpf_subprog_info {
  * @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_call_push_regs: jump target for subroutine for saving R6~R9 to stack
+ * @tgt_call_pop_regs: jump target for subroutine used for restoring R6~R9
  * @n_translated: number of successfully translated instructions (for errors)
  * @error: error code if something went wrong
  * @stack_frame_depth: max stack depth for current frame
@@ -475,6 +492,8 @@ struct nfp_prog {
 	unsigned int last_bpf_off;
 	unsigned int tgt_out;
 	unsigned int tgt_abort;
+	unsigned int tgt_call_push_regs;
+	unsigned int tgt_call_pop_regs;
 
 	unsigned int n_translated;
 	int error;
@@ -502,6 +521,7 @@ struct nfp_bpf_vnic {
 	unsigned int tgt_done;
 };
 
+bool nfp_is_subprog_start(struct nfp_insn_meta *meta);
 void nfp_bpf_jit_prepare(struct nfp_prog *nfp_prog, unsigned int cnt);
 int nfp_bpf_jit(struct nfp_prog *prog);
 bool nfp_bpf_supported_opcode(u8 code);

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

@@ -262,7 +262,6 @@ static int nfp_bpf_translate(struct nfp_net *nn, struct bpf_prog *prog)
 			prog->aux->stack_depth, stack_size);
 		return -EOPNOTSUPP;
 	}
-	nfp_prog->stack_frame_depth = round_up(prog->aux->stack_depth, 4);
 
 	max_instr = nn_readw(nn, NFP_NET_CFG_BPF_MAX_LEN);
 	nfp_prog->__prog_alloc_len = max_instr * sizeof(u64);

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

@@ -641,6 +641,27 @@ nfp_verify_insn(struct bpf_verifier_env *env, int insn_idx, int prev_insn_idx)
 	return 0;
 }
 
+static int
+nfp_assign_subprog_idx(struct bpf_verifier_env *env, struct nfp_prog *nfp_prog)
+{
+	struct nfp_insn_meta *meta;
+	int index = 0;
+
+	list_for_each_entry(meta, &nfp_prog->insns, l) {
+		if (nfp_is_subprog_start(meta))
+			index++;
+		meta->subprog_idx = index;
+	}
+
+	if (index + 1 != nfp_prog->subprog_cnt) {
+		pr_vlog(env, "BUG: number of processed BPF functions is not consistent (processed %d, expected %d)\n",
+			index + 1, nfp_prog->subprog_cnt);
+		return -EFAULT;
+	}
+
+	return 0;
+}
+
 static int nfp_bpf_finalize(struct bpf_verifier_env *env)
 {
 	struct bpf_subprog_info *info;
@@ -654,10 +675,21 @@ static int nfp_bpf_finalize(struct bpf_verifier_env *env)
 	if (!nfp_prog->subprog)
 		return -ENOMEM;
 
+	nfp_assign_subprog_idx(env, nfp_prog);
+
 	info = env->subprog_info;
-	for (i = 0; i < nfp_prog->subprog_cnt; i++)
+	for (i = 0; i < nfp_prog->subprog_cnt; i++) {
 		nfp_prog->subprog[i].stack_depth = info[i].stack_depth;
 
+		if (i == 0)
+			continue;
+
+		/* Account for size of return address. */
+		nfp_prog->subprog[i].stack_depth += REG_WIDTH;
+		/* Account for size of saved registers. */
+		nfp_prog->subprog[i].stack_depth += BPF_REG_SIZE * 4;
+	}
+
 	return 0;
 }
 

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

@@ -82,6 +82,15 @@
 #define OP_BR_BIT_ADDR_LO	OP_BR_ADDR_LO
 #define OP_BR_BIT_ADDR_HI	OP_BR_ADDR_HI
 
+#define OP_BR_ALU_BASE		0x0e800000000ULL
+#define OP_BR_ALU_BASE_MASK	0x0ff80000000ULL
+#define OP_BR_ALU_A_SRC		0x000000003ffULL
+#define OP_BR_ALU_B_SRC		0x000000ffc00ULL
+#define OP_BR_ALU_DEFBR		0x00000300000ULL
+#define OP_BR_ALU_IMM_HI	0x0007fc00000ULL
+#define OP_BR_ALU_SRC_LMEXTN	0x40000000000ULL
+#define OP_BR_ALU_DST_LMEXTN	0x80000000000ULL
+
 static inline bool nfp_is_br(u64 insn)
 {
 	return (insn & OP_BR_BASE_MASK) == OP_BR_BASE ||