s390/bpf,jit: BPF Just In Time compiler for s390

The s390 implementation of the JIT compiler for packet filter speedup.

Reviewed-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

3 files changed, 872 insertions, 0 deletions

diff --git a/arch/s390/net/Makefile b/arch/s390/net/Makefile
new file mode 100644
index 000000000000..90568c33ddb0
--- /dev/null
+++ b/arch/s390/net/Makefile
@@ -0,0 +1,4 @@
+#
+# Arch-specific network modules
+#
+obj-$(CONFIG_BPF_JIT) += bpf_jit.o bpf_jit_comp.o
diff --git a/arch/s390/net/bpf_jit.S b/arch/s390/net/bpf_jit.S
new file mode 100644
index 000000000000..7e45d13816c1
--- /dev/null
+++ b/arch/s390/net/bpf_jit.S
@@ -0,0 +1,130 @@
+/*
+ * BPF Jit compiler for s390, help functions.
+ *
+ * Copyright IBM Corp. 2012
+ *
+ * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
+ */
+#include <linux/linkage.h>
+
+/*
+ * Calling convention:
+ * registers %r2, %r6-%r8, %r10-%r11, %r13, %r15 are call saved
+ *   %r2: skb pointer
+ *   %r3: offset parameter
+ *   %r5: BPF A accumulator
+ *   %r8: return address
+ *   %r9: save register for skb pointer
+ *   %r10: skb->data
+ *   %r11: skb->len - skb->data_len (headlen)
+ *   %r12: BPF X accumulator
+ *
+ * skb_copy_bits takes 4 parameters:
+ *   %r2 = skb pointer
+ *   %r3 = offset into skb data
+ *   %r4 = length to copy
+ *   %r5 = pointer to temp buffer
+ */
+#define SKBDATA	%r8
+
+	/* A = *(u32 *) (skb->data+K+X) */
+ENTRY(sk_load_word_ind)
+	ar	%r3,%r12		# offset += X
+	bmr	%r8			# < 0 -> return with cc
+
+	/* A = *(u32 *) (skb->data+K) */
+ENTRY(sk_load_word)
+	llgfr	%r1,%r3			# extend offset
+	ahi	%r3,4			# offset + 4
+	clr	%r11,%r3		# hlen <= offset + 4 ?
+	jl	sk_load_word_slow
+	l	%r5,0(%r1,%r10)		# get word from skb
+	xr	%r1,%r1			# set cc to zero
+	br	%r8
+
+sk_load_word_slow:
+	lgr	%r9,%r2			# save %r2
+	lhi	%r4,4			# 4 bytes
+	la	%r5,160(%r15)		# pointer to temp buffer
+	brasl	%r14,skb_copy_bits	# get data from skb
+	l	%r5,160(%r15)		# load result from temp buffer
+	ltgr	%r2,%r2			# set cc to (%r2 != 0)
+	lgr	%r2,%r9			# restore %r2
+	br	%r8
+
+	/* A = *(u16 *) (skb->data+K+X) */
+ENTRY(sk_load_half_ind)
+	ar	%r3,%r12		# offset += X
+	bmr	%r8			# < 0 -> return with cc
+
+	/* A = *(u16 *) (skb->data+K) */
+ENTRY(sk_load_half)
+	llgfr	%r1,%r3			# extend offset
+	ahi	%r3,2			# offset + 2
+	clr	%r11,%r3		# hlen <= offset + 2 ?
+	jl	sk_load_half_slow
+	llgh	%r5,0(%r1,%r10)		# get half from skb
+	xr	%r1,%r1			# set cc to zero
+	br	%r8
+
+sk_load_half_slow:
+	lgr	%r9,%r2			# save %r2
+	lhi	%r4,2			# 2 bytes
+	la	%r5,162(%r15)		# pointer to temp buffer
+	brasl	%r14,skb_copy_bits	# get data from skb
+	xc	160(2,%r15),160(%r15)
+	l	%r5,160(%r15)		# load result from temp buffer
+	ltgr	%r2,%r2			# set cc to (%r2 != 0)
+	lgr	%r2,%r9			# restore %r2
+	br	%r8
+
+	/* A = *(u8 *) (skb->data+K+X) */
+ENTRY(sk_load_byte_ind)
+	ar	%r3,%r12		# offset += X
+	bmr	%r8			# < 0 -> return with cc
+
+	/* A = *(u8 *) (skb->data+K) */
+ENTRY(sk_load_byte)
+	llgfr	%r1,%r3			# extend offset
+	clr	%r11,%r3		# hlen < offset ?
+	jle	sk_load_byte_slow
+	lhi	%r5,0
+	ic	%r5,0(%r1,%r10)		# get byte from skb
+	xr	%r1,%r1			# set cc to zero
+	br	%r8
+
+sk_load_byte_slow:
+	lgr	%r9,%r2			# save %r2
+	lhi	%r4,1			# 1 bytes
+	la	%r5,163(%r15)		# pointer to temp buffer
+	brasl	%r14,skb_copy_bits	# get data from skb
+	xc	160(3,%r15),160(%r15)
+	l	%r5,160(%r15)		# load result from temp buffer
+	ltgr	%r2,%r2			# set cc to (%r2 != 0)
+	lgr	%r2,%r9			# restore %r2
+	br	%r8
+
+	/* A = (*(u8 *)(skb->data+K) & 0xf) << 2 */
+ENTRY(sk_load_byte_msh)
+	llgfr	%r1,%r3			# extend offset
+	clr	%r11,%r3		# hlen < offset ?
+	jle	sk_load_byte_slow
+	lhi	%r12,0
+	ic	%r12,0(%r1,%r10)	# get byte from skb
+	nill	%r12,0x0f
+	sll	%r12,2
+	xr	%r1,%r1			# set cc to zero
+	br	%r8
+
+sk_load_byte_msh_slow:
+	lgr	%r9,%r2			# save %r2
+	lhi	%r4,2			# 2 bytes
+	la	%r5,162(%r15)		# pointer to temp buffer
+	brasl	%r14,skb_copy_bits	# get data from skb
+	xc	160(3,%r15),160(%r15)
+	l	%r12,160(%r15)		# load result from temp buffer
+	nill	%r12,0x0f
+	sll	%r12,2
+	ltgr	%r2,%r2			# set cc to (%r2 != 0)
+	lgr	%r2,%r9			# restore %r2
+	br	%r8
diff --git a/arch/s390/net/bpf_jit_comp.c b/arch/s390/net/bpf_jit_comp.c
new file mode 100644
index 000000000000..b07aa3dc5eed
--- /dev/null
+++ b/arch/s390/net/bpf_jit_comp.c
@@ -0,0 +1,738 @@
+/*
+ * BPF Jit compiler for s390.
+ *
+ * Copyright IBM Corp. 2012
+ *
+ * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
+ */
+#include <linux/moduleloader.h>
+#include <linux/netdevice.h>
+#include <linux/filter.h>
+#include <asm/cacheflush.h>
+#include <asm/processor.h>
+
+/*
+ * Conventions:
+ *   %r2 = skb pointer
+ *   %r3 = offset parameter
+ *   %r4 = scratch register / length parameter
+ *   %r5 = BPF A accumulator
+ *   %r8 = return address
+ *   %r9 = save register for skb pointer
+ *   %r10 = skb->data
+ *   %r11 = skb->len - skb->data_len (headlen)
+ *   %r12 = BPF X accumulator
+ *   %r13 = literal pool pointer
+ *   0(%r15) - 63(%r15) scratch memory array with BPF_MEMWORDS
+ */
+int bpf_jit_enable __read_mostly;
+
+/*
+ * assembly code in arch/x86/net/bpf_jit.S
+ */
+extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[];
+extern u8 sk_load_word_ind[], sk_load_half_ind[], sk_load_byte_ind[];
+
+struct bpf_jit {
+	unsigned int seen;
+	u8 *start;
+	u8 *prg;
+	u8 *mid;
+	u8 *lit;
+	u8 *end;
+	u8 *base_ip;
+	u8 *ret0_ip;
+	u8 *exit_ip;
+	unsigned int off_load_word;
+	unsigned int off_load_half;
+	unsigned int off_load_byte;
+	unsigned int off_load_bmsh;
+	unsigned int off_load_iword;
+	unsigned int off_load_ihalf;
+	unsigned int off_load_ibyte;
+};
+
+#define BPF_SIZE_MAX	4096	/* Max size for program */
+
+#define SEEN_DATAREF	1	/* might call external helpers */
+#define SEEN_XREG	2	/* ebx is used */
+#define SEEN_MEM	4	/* use mem[] for temporary storage */
+#define SEEN_RET0	8	/* pc_ret0 points to a valid return 0 */
+#define SEEN_LITERAL	16	/* code uses literals */
+#define SEEN_LOAD_WORD	32	/* code uses sk_load_word */
+#define SEEN_LOAD_HALF	64	/* code uses sk_load_half */
+#define SEEN_LOAD_BYTE	128	/* code uses sk_load_byte */
+#define SEEN_LOAD_BMSH	256	/* code uses sk_load_byte_msh */
+#define SEEN_LOAD_IWORD	512	/* code uses sk_load_word_ind */
+#define SEEN_LOAD_IHALF	1024	/* code uses sk_load_half_ind */
+#define SEEN_LOAD_IBYTE	2048	/* code uses sk_load_byte_ind */
+
+#define EMIT2(op)					\
+({							\
+	if (jit->prg + 2 <= jit->mid)			\
+		*(u16 *) jit->prg = op;			\
+	jit->prg += 2;					\
+})
+
+#define EMIT4(op)					\
+({							\
+	if (jit->prg + 4 <= jit->mid)			\
+		*(u32 *) jit->prg = op;			\
+	jit->prg += 4;					\
+})
+
+#define EMIT4_DISP(op, disp)				\
+({							\
+	unsigned int __disp = (disp) & 0xfff;		\
+	EMIT4(op | __disp);				\
+})
+
+#define EMIT4_IMM(op, imm)				\
+({							\
+	unsigned int __imm = (imm) & 0xffff;		\
+	EMIT4(op | __imm);				\
+})
+
+#define EMIT4_PCREL(op, pcrel)				\
+({							\
+	long __pcrel = ((pcrel) >> 1) & 0xffff;		\
+	EMIT4(op | __pcrel);				\
+})
+
+#define EMIT6(op1, op2)					\
+({							\
+	if (jit->prg + 6 <= jit->mid) {			\
+		*(u32 *) jit->prg = op1;		\
+		*(u16 *) (jit->prg + 4) = op2;		\
+	}						\
+	jit->prg += 6;					\
+})
+
+#define EMIT6_DISP(op1, op2, disp)			\
+({							\
+	unsigned int __disp = (disp) & 0xfff;		\
+	EMIT6(op1 | __disp, op2);			\
+})
+
+#define EMIT_CONST(val)					\
+({							\
+	unsigned int ret;				\
+	ret = (unsigned int) (jit->lit - jit->base_ip);	\
+	jit->seen |= SEEN_LITERAL;			\
+	if (jit->lit + 4 <= jit->end)			\
+		*(u32 *) jit->lit = val;		\
+	jit->lit += 4;					\
+	ret;						\
+})
+
+#define EMIT_FN_CONST(bit, fn)				\
+({							\
+	unsigned int ret;				\
+	ret = (unsigned int) (jit->lit - jit->base_ip);	\
+	if (jit->seen & bit) {				\
+		jit->seen |= SEEN_LITERAL;		\
+		if (jit->lit + 8 <= jit->end)		\
+			*(void **) jit->lit = fn;	\
+		jit->lit += 8;				\
+	}						\
+	ret;						\
+})
+
+static void bpf_jit_prologue(struct bpf_jit *jit)
+{
+	/* Save registers and create stack frame if necessary */
+	if (jit->seen & SEEN_DATAREF) {
+		/* stmg %r8,%r15,88(%r15) */
+		EMIT6(0xeb8ff058, 0x0024);
+		/* lgr %r14,%r15 */
+		EMIT4(0xb90400ef);
+		/* ahi %r15,<offset> */
+		EMIT4_IMM(0xa7fa0000, (jit->seen & SEEN_MEM) ? -112 : -80);
+		/* stg %r14,152(%r15) */
+		EMIT6(0xe3e0f098, 0x0024);
+	} else if ((jit->seen & SEEN_XREG) && (jit->seen & SEEN_LITERAL))
+		/* stmg %r12,%r13,120(%r15) */
+		EMIT6(0xebcdf078, 0x0024);
+	else if (jit->seen & SEEN_XREG)
+		/* stg %r12,120(%r15) */
+		EMIT6(0xe3c0f078, 0x0024);
+	else if (jit->seen & SEEN_LITERAL)
+		/* stg %r13,128(%r15) */
+		EMIT6(0xe3d0f080, 0x0024);
+
+	/* Setup literal pool */
+	if (jit->seen & SEEN_LITERAL) {
+		/* basr %r13,0 */
+		EMIT2(0x0dd0);
+		jit->base_ip = jit->prg;
+	}
+	jit->off_load_word = EMIT_FN_CONST(SEEN_LOAD_WORD, sk_load_word);
+	jit->off_load_half = EMIT_FN_CONST(SEEN_LOAD_HALF, sk_load_half);
+	jit->off_load_byte = EMIT_FN_CONST(SEEN_LOAD_BYTE, sk_load_byte);
+	jit->off_load_bmsh = EMIT_FN_CONST(SEEN_LOAD_BMSH, sk_load_byte_msh);
+	jit->off_load_iword = EMIT_FN_CONST(SEEN_LOAD_IWORD, sk_load_word_ind);
+	jit->off_load_ihalf = EMIT_FN_CONST(SEEN_LOAD_IHALF, sk_load_half_ind);
+	jit->off_load_ibyte = EMIT_FN_CONST(SEEN_LOAD_IBYTE, sk_load_byte_ind);
+
+	/* Filter needs to access skb data */
+	if (jit->seen & SEEN_DATAREF) {
+		/* l %r11,<len>(%r2) */
+		EMIT4_DISP(0x58b02000, offsetof(struct sk_buff, len));
+		/* s %r11,<data_len>(%r2) */
+		EMIT4_DISP(0x5bb02000, offsetof(struct sk_buff, data_len));
+		/* lg %r10,<data>(%r2) */
+		EMIT6_DISP(0xe3a02000, 0x0004,
+			   offsetof(struct sk_buff, data));
+	}
+}
+
+static void bpf_jit_epilogue(struct bpf_jit *jit)
+{
+	/* Return 0 */
+	if (jit->seen & SEEN_RET0) {
+		jit->ret0_ip = jit->prg;
+		/* lghi %r2,0 */
+		EMIT4(0xa7290000);
+	}
+	jit->exit_ip = jit->prg;
+	/* Restore registers */
+	if (jit->seen & SEEN_DATAREF)
+		/* lmg %r8,%r15,<offset>(%r15) */
+		EMIT6_DISP(0xeb8ff000, 0x0004,
+			   (jit->seen & SEEN_MEM) ? 200 : 168);
+	else if ((jit->seen & SEEN_XREG) && (jit->seen & SEEN_LITERAL))
+		/* lmg %r12,%r13,120(%r15) */
+		EMIT6(0xebcdf078, 0x0004);
+	else if (jit->seen & SEEN_XREG)
+		/* lg %r12,120(%r15) */
+		EMIT6(0xe3c0f078, 0x0004);
+	else if (jit->seen & SEEN_LITERAL)
+		/* lg %r13,128(%r15) */
+		EMIT6(0xe3d0f080, 0x0004);
+	/* br %r14 */
+	EMIT2(0x07fe);
+}
+
+/*
+ * make sure we dont leak kernel information to user
+ */
+static void bpf_jit_noleaks(struct bpf_jit *jit, struct sock_filter *filter)
+{
+	/* Clear temporary memory if (seen & SEEN_MEM) */
+	if (jit->seen & SEEN_MEM)
+		/* xc 0(64,%r15),0(%r15) */
+		EMIT6(0xd73ff000, 0xf000);
+	/* Clear X if (seen & SEEN_XREG) */
+	if (jit->seen & SEEN_XREG)
+		/* lhi %r12,0 */
+		EMIT4(0xa7c80000);
+	/* Clear A if the first register does not set it. */
+	switch (filter[0].code) {
+	case BPF_S_LD_W_ABS:
+	case BPF_S_LD_H_ABS:
+	case BPF_S_LD_B_ABS:
+	case BPF_S_LD_W_LEN:
+	case BPF_S_LD_W_IND:
+	case BPF_S_LD_H_IND:
+	case BPF_S_LD_B_IND:
+	case BPF_S_LDX_B_MSH:
+	case BPF_S_LD_IMM:
+	case BPF_S_LD_MEM:
+	case BPF_S_MISC_TXA:
+	case BPF_S_ANC_PROTOCOL:
+	case BPF_S_ANC_PKTTYPE:
+	case BPF_S_ANC_IFINDEX:
+	case BPF_S_ANC_MARK:
+	case BPF_S_ANC_QUEUE:
+	case BPF_S_ANC_HATYPE:
+	case BPF_S_ANC_RXHASH:
+	case BPF_S_ANC_CPU:
+	case BPF_S_RET_K:
+		/* first instruction sets A register */
+		break;
+	default: /* A = 0 */
+		/* lhi %r5,0 */
+		EMIT4(0xa7580000);
+	}
+}
+
+static int bpf_jit_insn(struct bpf_jit *jit, struct sock_filter *filter,
+			unsigned int *addrs, int i, int last)
+{
+	unsigned int K;
+	int offset;
+	unsigned int mask;
+
+	K = filter->k;
+	switch (filter->code) {
+	case BPF_S_ALU_ADD_X: /* A += X */
+		jit->seen |= SEEN_XREG;
+		/* ar %r5,%r12 */
+		EMIT2(0x1a5c);
+		break;
+	case BPF_S_ALU_ADD_K: /* A += K */
+		if (!K)
+			break;
+		if (K <= 16383)
+			/* ahi %r5,<K> */
+			EMIT4_IMM(0xa75a0000, K);
+		else
+			/* a %r5,<d(K)>(%r13) */
+			EMIT4_DISP(0x5a50d000, EMIT_CONST(K));
+		break;
+	case BPF_S_ALU_SUB_X: /* A -= X */
+		jit->seen |= SEEN_XREG;
+		/* sr %r5,%r12 */
+		EMIT2(0x1b5c);
+		break;
+	case BPF_S_ALU_SUB_K: /* A -= K */
+		if (!K)
+			break;
+		if (K <= 16384)
+			/* ahi %r5,-K */
+			EMIT4_IMM(0xa75a0000, -K);
+		else
+			/* s %r5,<d(K)>(%r13) */
+			EMIT4_DISP(0x5b50d000, EMIT_CONST(K));
+		break;
+	case BPF_S_ALU_MUL_X: /* A *= X */
+		jit->seen |= SEEN_XREG;
+		/* msr %r5,%r12 */
+		EMIT4(0xb252005c);
+		break;
+	case BPF_S_ALU_MUL_K: /* A *= K */
+		if (K <= 16383)
+			/* mhi %r5,K */
+			EMIT4_IMM(0xa75c0000, K);
+		else
+			/* ms %r5,<d(K)>(%r13) */
+			EMIT4_DISP(0x7150d000, EMIT_CONST(K));
+		break;
+	case BPF_S_ALU_DIV_X: /* A /= X */
+		jit->seen |= SEEN_XREG | SEEN_RET0;
+		/* ltr %r12,%r12 */
+		EMIT2(0x12cc);
+		/* jz <ret0> */
+		EMIT4_PCREL(0xa7840000, (jit->ret0_ip - jit->prg));
+		/* lhi %r4,0 */
+		EMIT4(0xa7480000);
+		/* dr %r4,%r12 */
+		EMIT2(0x1d4c);
+		break;
+	case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K) */
+		/* m %r4,<d(K)>(%r13) */
+		EMIT4_DISP(0x5c40d000, EMIT_CONST(K));
+		/* lr %r5,%r4 */
+		EMIT2(0x1854);
+		break;
+	case BPF_S_ALU_AND_X: /* A &= X */
+		jit->seen |= SEEN_XREG;
+		/* nr %r5,%r12 */
+		EMIT2(0x145c);
+		break;
+	case BPF_S_ALU_AND_K: /* A &= K */
+		/* n %r5,<d(K)>(%r13) */
+		EMIT4_DISP(0x5450d000, EMIT_CONST(K));
+		break;
+	case BPF_S_ALU_OR_X: /* A |= X */
+		jit->seen |= SEEN_XREG;
+		/* or %r5,%r12 */
+		EMIT2(0x165c);
+		break;
+	case BPF_S_ALU_OR_K: /* A |= K */
+		/* o %r5,<d(K)>(%r13) */
+		EMIT4_DISP(0x5650d000, EMIT_CONST(K));
+		break;
+	case BPF_S_ALU_LSH_X: /* A <<= X; */
+		jit->seen |= SEEN_XREG;
+		/* sll %r5,0(%r12) */
+		EMIT4(0x8950c000);
+		break;
+	case BPF_S_ALU_LSH_K: /* A <<= K */
+		if (K == 0)
+			break;
+		/* sll %r5,K */
+		EMIT4_DISP(0x89500000, K);
+		break;
+	case BPF_S_ALU_RSH_X: /* A >>= X; */
+		jit->seen |= SEEN_XREG;
+		/* srl %r5,0(%r12) */
+		EMIT4(0x8850c000);
+		break;
+	case BPF_S_ALU_RSH_K: /* A >>= K; */
+		if (K == 0)
+			break;
+		/* srl %r5,K */
+		EMIT4_DISP(0x88500000, K);
+		break;
+	case BPF_S_ALU_NEG: /* A = -A */
+		/* lnr %r5,%r5 */
+		EMIT2(0x1155);
+		break;
+	case BPF_S_JMP_JA: /* ip += K */
+		offset = addrs[i + K] + jit->start - jit->prg;
+		EMIT4_PCREL(0xa7f40000, offset);
+		break;
+	case BPF_S_JMP_JGT_K: /* ip += (A > K) ? jt : jf */
+		mask = 0x200000; /* jh */
+		goto kbranch;
+	case BPF_S_JMP_JGE_K: /* ip += (A >= K) ? jt : jf */
+		mask = 0xa00000; /* jhe */
+		goto kbranch;
+	case BPF_S_JMP_JEQ_K: /* ip += (A == K) ? jt : jf */
+		mask = 0x800000; /* je */
+kbranch:	/* Emit compare if the branch targets are different */
+		if (filter->jt != filter->jf) {
+			if (K <= 16383)
+				/* chi %r5,<K> */
+				EMIT4_IMM(0xa75e0000, K);
+			else
+				/* c %r5,<d(K)>(%r13) */
+				EMIT4_DISP(0x5950d000, EMIT_CONST(K));
+		}
+branch:		if (filter->jt == filter->jf) {
+			if (filter->jt == 0)
+				break;
+			/* j <jt> */
+			offset = addrs[i + filter->jt] + jit->start - jit->prg;
+			EMIT4_PCREL(0xa7f40000, offset);
+			break;
+		}
+		if (filter->jt != 0) {
+			/* brc	<mask>,<jt> */
+			offset = addrs[i + filter->jt] + jit->start - jit->prg;
+			EMIT4_PCREL(0xa7040000 | mask, offset);
+		}
+		if (filter->jf != 0) {
+			/* brc	<mask^15>,<jf> */
+			offset = addrs[i + filter->jf] + jit->start - jit->prg;
+			EMIT4_PCREL(0xa7040000 | (mask ^ 0xf00000), offset);
+		}
+		break;
+	case BPF_S_JMP_JSET_K: /* ip += (A & K) ? jt : jf */
+		mask = 0x700000; /* jnz */
+		/* Emit test if the branch targets are different */
+		if (filter->jt != filter->jf) {
+			if (K > 65535) {
+				/* lr %r4,%r5 */
+				EMIT2(0x1845);
+				/* n %r4,<d(K)>(%r13) */
+				EMIT4_DISP(0x5440d000, EMIT_CONST(K));
+			} else
+				/* tmll %r5,K */
+				EMIT4_IMM(0xa7510000, K);
+		}
+		goto branch;
+	case BPF_S_JMP_JGT_X: /* ip += (A > X) ? jt : jf */
+		mask = 0x200000; /* jh */
+		goto xbranch;
+	case BPF_S_JMP_JGE_X: /* ip += (A >= X) ? jt : jf */
+		mask = 0xa00000; /* jhe */
+		goto xbranch;
+	case BPF_S_JMP_JEQ_X: /* ip += (A == X) ? jt : jf */
+		mask = 0x800000; /* je */
+xbranch:	/* Emit compare if the branch targets are different */
+		if (filter->jt != filter->jf) {
+			jit->seen |= SEEN_XREG;
+			/* cr %r5,%r12 */
+			EMIT2(0x195c);
+		}
+		goto branch;
+	case BPF_S_JMP_JSET_X: /* ip += (A & X) ? jt : jf */
+		mask = 0x700000; /* jnz */
+		/* Emit test if the branch targets are different */
+		if (filter->jt != filter->jf) {
+			jit->seen |= SEEN_XREG;
+			/* lr %r4,%r5 */
+			EMIT2(0x1845);
+			/* nr %r4,%r12 */
+			EMIT2(0x144c);
+		}
+		goto branch;
+	case BPF_S_LD_W_ABS: /* A = *(u32 *) (skb->data+K) */
+		jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_WORD;
+		offset = jit->off_load_word;
+		goto load_abs;
+	case BPF_S_LD_H_ABS: /* A = *(u16 *) (skb->data+K) */
+		jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_HALF;
+		offset = jit->off_load_half;
+		goto load_abs;
+	case BPF_S_LD_B_ABS: /* A = *(u8 *) (skb->data+K) */
+		jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_BYTE;
+		offset = jit->off_load_byte;
+load_abs:	if ((int) K < 0)
+			goto out;
+call_fn:	/* lg %r1,<d(function)>(%r13) */
+		EMIT6_DISP(0xe310d000, 0x0004, offset);
+		/* l %r3,<d(K)>(%r13) */
+		EMIT4_DISP(0x5830d000, EMIT_CONST(K));
+		/* basr %r8,%r1 */
+		EMIT2(0x0d81);
+		/* jnz <ret0> */
+		EMIT4_PCREL(0xa7740000, (jit->ret0_ip - jit->prg));
+		break;
+	case BPF_S_LD_W_IND: /* A = *(u32 *) (skb->data+K+X) */
+		jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IWORD;
+		offset = jit->off_load_iword;
+		goto call_fn;
+	case BPF_S_LD_H_IND: /* A = *(u16 *) (skb->data+K+X) */
+		jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IHALF;
+		offset = jit->off_load_ihalf;
+		goto call_fn;
+	case BPF_S_LD_B_IND: /* A = *(u8 *) (skb->data+K+X) */
+		jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IBYTE;
+		offset = jit->off_load_ibyte;
+		goto call_fn;
+	case BPF_S_LDX_B_MSH:
+		/* X = (*(u8 *)(skb->data+K) & 0xf) << 2 */
+		jit->seen |= SEEN_RET0;
+		if ((int) K < 0) {
+			/* j <ret0> */
+			EMIT4_PCREL(0xa7f40000, (jit->ret0_ip - jit->prg));
+			break;
+		}
+		jit->seen |= SEEN_DATAREF | SEEN_LOAD_BMSH;
+		offset = jit->off_load_bmsh;
+		goto call_fn;
+	case BPF_S_LD_W_LEN: /*	A = skb->len; */
+		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
+		/* l %r5,<d(len)>(%r2) */
+		EMIT4_DISP(0x58502000, offsetof(struct sk_buff, len));
+		break;
+	case BPF_S_LDX_W_LEN: /* X = skb->len; */
+		jit->seen |= SEEN_XREG;
+		/* l %r12,<d(len)>(%r2) */
+		EMIT4_DISP(0x58c02000, offsetof(struct sk_buff, len));
+		break;
+	case BPF_S_LD_IMM: /* A = K */
+		if (K <= 16383)
+			/* lhi %r5,K */
+			EMIT4_IMM(0xa7580000, K);
+		else
+			/* l %r5,<d(K)>(%r13) */
+			EMIT4_DISP(0x5850d000, EMIT_CONST(K));
+		break;
+	case BPF_S_LDX_IMM: /* X = K */
+		jit->seen |= SEEN_XREG;
+		if (K <= 16383)
+			/* lhi %r12,<K> */
+			EMIT4_IMM(0xa7c80000, K);
+		else
+			/* l %r12,<d(K)>(%r13) */
+			EMIT4_DISP(0x58c0d000, EMIT_CONST(K));
+		break;
+	case BPF_S_LD_MEM: /* A = mem[K] */
+		jit->seen |= SEEN_MEM;
+		/* l %r5,<K>(%r15) */
+		EMIT4_DISP(0x5850f000,
+			   (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
+		break;
+	case BPF_S_LDX_MEM: /* X = mem[K] */
+		jit->seen |= SEEN_XREG | SEEN_MEM;
+		/* l %r12,<K>(%r15) */
+		EMIT4_DISP(0x58c0f000,
+			   (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
+		break;
+	case BPF_S_MISC_TAX: /* X = A */
+		jit->seen |= SEEN_XREG;
+		/* lr %r12,%r5 */
+		EMIT2(0x18c5);
+		break;
+	case BPF_S_MISC_TXA: /* A = X */
+		jit->seen |= SEEN_XREG;
+		/* lr %r5,%r12 */
+		EMIT2(0x185c);
+		break;
+	case BPF_S_RET_K:
+		if (K == 0) {
+			jit->seen |= SEEN_RET0;
+			if (last)
+				break;
+			/* j <ret0> */
+			EMIT4_PCREL(0xa7f40000, jit->ret0_ip - jit->prg);
+		} else {
+			if (K <= 16383)
+				/* lghi %r2,K */
+				EMIT4_IMM(0xa7290000, K);
+			else
+				/* llgf %r2,<K>(%r13) */
+				EMIT6_DISP(0xe320d000, 0x0016, EMIT_CONST(K));
+			/* j <exit> */
+			if (last && !(jit->seen & SEEN_RET0))
+				break;
+			EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg);
+		}
+		break;
+	case BPF_S_RET_A:
+		/* llgfr %r2,%r5 */
+		EMIT4(0xb9160025);
+		/* j <exit> */
+		EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg);
+		break;
+	case BPF_S_ST: /* mem[K] = A */
+		jit->seen |= SEEN_MEM;
+		/* st %r5,<K>(%r15) */
+		EMIT4_DISP(0x5050f000,
+			   (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
+		break;
+	case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */
+		jit->seen |= SEEN_XREG | SEEN_MEM;
+		/* st %r12,<K>(%r15) */
+		EMIT4_DISP(0x50c0f000,
+			   (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
+		break;
+	case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
+		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
+		/* lhi %r5,0 */
+		EMIT4(0xa7580000);
+		/* icm	%r5,3,<d(protocol)>(%r2) */
+		EMIT4_DISP(0xbf532000, offsetof(struct sk_buff, protocol));
+		break;
+	case BPF_S_ANC_IFINDEX:	/* if (!skb->dev) return 0;
+				 * A = skb->dev->ifindex */
+		BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
+		jit->seen |= SEEN_RET0;
+		/* lg %r1,<d(dev)>(%r2) */
+		EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev));
+		/* ltgr %r1,%r1 */
+		EMIT4(0xb9020011);
+		/* jz <ret0> */
+		EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg);
+		/* l %r5,<d(ifindex)>(%r1) */
+		EMIT4_DISP(0x58501000, offsetof(struct net_device, ifindex));
+		break;
+	case BPF_S_ANC_MARK: /* A = skb->mark */
+		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
+		/* l %r5,<d(mark)>(%r2) */
+		EMIT4_DISP(0x58502000, offsetof(struct sk_buff, mark));
+		break;
+	case BPF_S_ANC_QUEUE: /* A = skb->queue_mapping */
+		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
+		/* lhi %r5,0 */
+		EMIT4(0xa7580000);
+		/* icm	%r5,3,<d(queue_mapping)>(%r2) */
+		EMIT4_DISP(0xbf532000, offsetof(struct sk_buff, queue_mapping));
+		break;
+	case BPF_S_ANC_HATYPE:	/* if (!skb->dev) return 0;
+				 * A = skb->dev->type */
+		BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
+		jit->seen |= SEEN_RET0;
+		/* lg %r1,<d(dev)>(%r2) */
+		EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev));
+		/* ltgr %r1,%r1 */
+		EMIT4(0xb9020011);
+		/* jz <ret0> */
+		EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg);
+		/* lhi %r5,0 */
+		EMIT4(0xa7580000);
+		/* icm	%r5,3,<d(type)>(%r1) */
+		EMIT4_DISP(0xbf531000, offsetof(struct net_device, type));
+		break;
+	case BPF_S_ANC_RXHASH: /* A = skb->rxhash */
+		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
+		/* l %r5,<d(rxhash)>(%r2) */
+		EMIT4_DISP(0x58502000, offsetof(struct sk_buff, rxhash));
+		break;
+	case BPF_S_ANC_CPU: /* A = smp_processor_id() */
+#ifdef CONFIG_SMP
+		/* l %r5,<d(cpu_nr)> */
+		EMIT4_DISP(0x58500000, offsetof(struct _lowcore, cpu_nr));
+#else
+		/* lhi %r5,0 */
+		EMIT4(0xa7580000);
+#endif
+		break;
+	default: /* too complex, give up */
+		goto out;
+	}
+	addrs[i] = jit->prg - jit->start;
+	return 0;
+out:
+	return -1;
+}
+
+void bpf_jit_compile(struct sk_filter *fp)
+{
+	unsigned long size, prg_len, lit_len;
+	struct bpf_jit jit, cjit;
+	unsigned int *addrs;
+	int pass, i;
+
+	if (!bpf_jit_enable)
+		return;
+	addrs = kmalloc(fp->len * sizeof(*addrs), GFP_KERNEL);
+	if (addrs == NULL)
+		return;
+	memset(addrs, 0, fp->len * sizeof(*addrs));
+	memset(&jit, 0, sizeof(cjit));
+	memset(&cjit, 0, sizeof(cjit));
+
+	for (pass = 0; pass < 10; pass++) {
+		jit.prg = jit.start;
+		jit.lit = jit.mid;
+
+		bpf_jit_prologue(&jit);
+		bpf_jit_noleaks(&jit, fp->insns);
+		for (i = 0; i < fp->len; i++) {
+			if (bpf_jit_insn(&jit, fp->insns + i, addrs, i,
+					 i == fp->len - 1))
+				goto out;
+		}
+		bpf_jit_epilogue(&jit);
+		if (jit.start) {
+			WARN_ON(jit.prg > cjit.prg || jit.lit > cjit.lit);
+			if (memcmp(&jit, &cjit, sizeof(jit)) == 0)
+				break;
+		} else if (jit.prg == cjit.prg && jit.lit == cjit.lit) {
+			prg_len = jit.prg - jit.start;
+			lit_len = jit.lit - jit.mid;
+			size = max_t(unsigned long, prg_len + lit_len,
+				     sizeof(struct work_struct));
+			if (size >= BPF_SIZE_MAX)
+				goto out;
+			jit.start = module_alloc(size);
+			if (!jit.start)
+				goto out;
+			jit.prg = jit.mid = jit.start + prg_len;
+			jit.lit = jit.end = jit.start + prg_len + lit_len;
+			jit.base_ip += (unsigned long) jit.start;
+			jit.exit_ip += (unsigned long) jit.start;
+			jit.ret0_ip += (unsigned long) jit.start;
+		}
+		cjit = jit;
+	}
+	if (bpf_jit_enable > 1) {
+		pr_err("flen=%d proglen=%lu pass=%d image=%p\n",
+		       fp->len, jit.end - jit.start, pass, jit.start);
+		if (jit.start) {
+			printk(KERN_ERR "JIT code:\n");
+			print_fn_code(jit.start, jit.mid - jit.start);
+			print_hex_dump(KERN_ERR, "JIT literals:\n",
+				       DUMP_PREFIX_ADDRESS, 16, 1,
+				       jit.mid, jit.end - jit.mid, false);
+		}
+	}
+	if (jit.start)
+		fp->bpf_func = (void *) jit.start;
+out:
+	kfree(addrs);
+}
+
+static void jit_free_defer(struct work_struct *arg)
+{
+	module_free(NULL, arg);
+}
+
+/* run from softirq, we must use a work_struct to call
+ * module_free() from process context
+ */
+void bpf_jit_free(struct sk_filter *fp)
+{
+	struct work_struct *work;
+
+	if (fp->bpf_func == sk_run_filter)
+		return;
+	work = (struct work_struct *)fp->bpf_func;
+	INIT_WORK(work, jit_free_defer);
+	schedule_work(work);
+}