From c10302efe569bfd646b4c22df29577a4595b4580 Mon Sep 17 00:00:00 2001 From: Martin Schwidefsky Date: Tue, 31 Jul 2012 16:23:59 +0200 Subject: [PATCH] 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 Signed-off-by: Martin Schwidefsky --- arch/s390/Kbuild | 1 + arch/s390/Kconfig | 1 + arch/s390/include/asm/processor.h | 1 + arch/s390/kernel/dis.c | 23 + arch/s390/net/Makefile | 4 + arch/s390/net/bpf_jit.S | 130 ++++++ arch/s390/net/bpf_jit_comp.c | 738 ++++++++++++++++++++++++++++++ 7 files changed, 898 insertions(+) create mode 100644 arch/s390/net/Makefile create mode 100644 arch/s390/net/bpf_jit.S create mode 100644 arch/s390/net/bpf_jit_comp.c diff --git a/arch/s390/Kbuild b/arch/s390/Kbuild index 9858476fa0fe..cc45d25487b0 100644 --- a/arch/s390/Kbuild +++ b/arch/s390/Kbuild @@ -5,3 +5,4 @@ obj-$(CONFIG_CRYPTO_HW) += crypto/ obj-$(CONFIG_S390_HYPFS_FS) += hypfs/ obj-$(CONFIG_APPLDATA_BASE) += appldata/ obj-$(CONFIG_MATHEMU) += math-emu/ +obj-y += net/ diff --git a/arch/s390/Kconfig b/arch/s390/Kconfig index 107610e01a29..a7088dc06d2c 100644 --- a/arch/s390/Kconfig +++ b/arch/s390/Kconfig @@ -84,6 +84,7 @@ config S390 select HAVE_KERNEL_XZ select HAVE_ARCH_MUTEX_CPU_RELAX select HAVE_ARCH_JUMP_LABEL if !MARCH_G5 + select HAVE_BPF_JIT if 64BIT && PACK_STACK select ARCH_SAVE_PAGE_KEYS if HIBERNATION select ARCH_HAS_ATOMIC64_DEC_IF_POSITIVE select HAVE_MEMBLOCK diff --git a/arch/s390/include/asm/processor.h b/arch/s390/include/asm/processor.h index 11e4e3236937..d4477ba99a16 100644 --- a/arch/s390/include/asm/processor.h +++ b/arch/s390/include/asm/processor.h @@ -140,6 +140,7 @@ extern int kernel_thread(int (*fn)(void *), void * arg, unsigned long flags); extern unsigned long thread_saved_pc(struct task_struct *t); extern void show_code(struct pt_regs *regs); +extern void print_fn_code(unsigned char *code, unsigned long len); unsigned long get_wchan(struct task_struct *p); #define task_pt_regs(tsk) ((struct pt_regs *) \ diff --git a/arch/s390/kernel/dis.c b/arch/s390/kernel/dis.c index 619c5d350726..84fd7e920bfe 100644 --- a/arch/s390/kernel/dis.c +++ b/arch/s390/kernel/dis.c @@ -1601,3 +1601,26 @@ void show_code(struct pt_regs *regs) } printk("\n"); } + +void print_fn_code(unsigned char *code, unsigned long len) +{ + char buffer[64], *ptr; + int opsize, i; + + while (len) { + ptr = buffer; + opsize = insn_length(*code); + ptr += sprintf(ptr, "%p: ", code); + for (i = 0; i < opsize; i++) + ptr += sprintf(ptr, "%02x", code[i]); + *ptr++ = '\t'; + if (i < 4) + *ptr++ = '\t'; + ptr += print_insn(ptr, code, (unsigned long) code); + *ptr++ = '\n'; + *ptr++ = 0; + printk(buffer); + code += opsize; + len -= opsize; + } +} 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 + */ +#include + +/* + * 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 + */ +#include +#include +#include +#include +#include + +/* + * 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, */ + 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,(%r2) */ + EMIT4_DISP(0x58b02000, offsetof(struct sk_buff, len)); + /* s %r11,(%r2) */ + EMIT4_DISP(0x5bb02000, offsetof(struct sk_buff, data_len)); + /* lg %r10,(%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,(%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, */ + EMIT4_IMM(0xa75a0000, K); + else + /* a %r5,(%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,(%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,(%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 */ + 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,(%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,(%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,(%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, */ + EMIT4_IMM(0xa75e0000, K); + else + /* c %r5,(%r13) */ + EMIT4_DISP(0x5950d000, EMIT_CONST(K)); + } +branch: if (filter->jt == filter->jf) { + if (filter->jt == 0) + break; + /* j */ + offset = addrs[i + filter->jt] + jit->start - jit->prg; + EMIT4_PCREL(0xa7f40000, offset); + break; + } + if (filter->jt != 0) { + /* brc , */ + offset = addrs[i + filter->jt] + jit->start - jit->prg; + EMIT4_PCREL(0xa7040000 | mask, offset); + } + if (filter->jf != 0) { + /* brc , */ + 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,(%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,(%r13) */ + EMIT6_DISP(0xe310d000, 0x0004, offset); + /* l %r3,(%r13) */ + EMIT4_DISP(0x5830d000, EMIT_CONST(K)); + /* basr %r8,%r1 */ + EMIT2(0x0d81); + /* jnz */ + 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 */ + 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,(%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,(%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,(%r13) */ + EMIT4_DISP(0x5850d000, EMIT_CONST(K)); + break; + case BPF_S_LDX_IMM: /* X = K */ + jit->seen |= SEEN_XREG; + if (K <= 16383) + /* lhi %r12, */ + EMIT4_IMM(0xa7c80000, K); + else + /* l %r12,(%r13) */ + EMIT4_DISP(0x58c0d000, EMIT_CONST(K)); + break; + case BPF_S_LD_MEM: /* A = mem[K] */ + jit->seen |= SEEN_MEM; + /* l %r5,(%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,(%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 */ + EMIT4_PCREL(0xa7f40000, jit->ret0_ip - jit->prg); + } else { + if (K <= 16383) + /* lghi %r2,K */ + EMIT4_IMM(0xa7290000, K); + else + /* llgf %r2,(%r13) */ + EMIT6_DISP(0xe320d000, 0x0016, EMIT_CONST(K)); + /* j */ + 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 */ + EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg); + break; + case BPF_S_ST: /* mem[K] = A */ + jit->seen |= SEEN_MEM; + /* st %r5,(%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,(%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,(%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,(%r2) */ + EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev)); + /* ltgr %r1,%r1 */ + EMIT4(0xb9020011); + /* jz */ + EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg); + /* l %r5,(%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,(%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,(%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,(%r2) */ + EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev)); + /* ltgr %r1,%r1 */ + EMIT4(0xb9020011); + /* jz */ + EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg); + /* lhi %r5,0 */ + EMIT4(0xa7580000); + /* icm %r5,3,(%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,(%r2) */ + EMIT4_DISP(0x58502000, offsetof(struct sk_buff, rxhash)); + break; + case BPF_S_ANC_CPU: /* A = smp_processor_id() */ +#ifdef CONFIG_SMP + /* l %r5, */ + 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); +}