1 /* bpf_jit_comp.c: BPF JIT compiler for PPC64
3 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
5 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com)
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; version 2
12 #include <linux/moduleloader.h>
13 #include <asm/cacheflush.h>
14 #include <linux/netdevice.h>
15 #include <linux/filter.h>
16 #include <linux/if_vlan.h>
20 int bpf_jit_enable __read_mostly;
22 static inline void bpf_flush_icache(void *start, void *end)
25 flush_icache_range((unsigned long)start, (unsigned long)end);
28 static void bpf_jit_build_prologue(struct sk_filter *fp, u32 *image,
29 struct codegen_context *ctx)
32 const struct sock_filter *filter = fp->insns;
34 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
36 if (ctx->seen & SEEN_DATAREF) {
37 /* If we call any helpers (for loads), save LR */
38 EMIT(PPC_INST_MFLR | __PPC_RT(R0));
41 /* Back up non-volatile regs. */
42 PPC_STD(r_D, 1, -(8*(32-r_D)));
43 PPC_STD(r_HL, 1, -(8*(32-r_HL)));
45 if (ctx->seen & SEEN_MEM) {
47 * Conditionally save regs r15-r31 as some will be used
50 for (i = r_M; i < (r_M+16); i++) {
51 if (ctx->seen & (1 << (i-r_M)))
52 PPC_STD(i, 1, -(8*(32-i)));
55 EMIT(PPC_INST_STDU | __PPC_RS(R1) | __PPC_RA(R1) |
56 (-BPF_PPC_STACKFRAME & 0xfffc));
59 if (ctx->seen & SEEN_DATAREF) {
61 * If this filter needs to access skb data,
62 * prepare r_D and r_HL:
63 * r_HL = skb->len - skb->data_len
66 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
68 PPC_LWZ_OFFS(r_HL, r_skb, offsetof(struct sk_buff, len));
69 PPC_SUB(r_HL, r_HL, r_scratch1);
70 PPC_LD_OFFS(r_D, r_skb, offsetof(struct sk_buff, data));
73 if (ctx->seen & SEEN_XREG) {
75 * TODO: Could also detect whether first instr. sets X and
76 * avoid this (as below, with A).
81 switch (filter[0].code) {
84 case BPF_S_ANC_PROTOCOL:
85 case BPF_S_ANC_IFINDEX:
87 case BPF_S_ANC_RXHASH:
88 case BPF_S_ANC_VLAN_TAG:
89 case BPF_S_ANC_VLAN_TAG_PRESENT:
95 /* first instruction sets A register (or is RET 'constant') */
98 /* make sure we dont leak kernel information to user */
103 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
107 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
108 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
109 if (ctx->seen & SEEN_DATAREF) {
112 PPC_LD(r_D, 1, -(8*(32-r_D)));
113 PPC_LD(r_HL, 1, -(8*(32-r_HL)));
115 if (ctx->seen & SEEN_MEM) {
116 /* Restore any saved non-vol registers */
117 for (i = r_M; i < (r_M+16); i++) {
118 if (ctx->seen & (1 << (i-r_M)))
119 PPC_LD(i, 1, -(8*(32-i)));
123 /* The RETs have left a return value in R3. */
128 #define CHOOSE_LOAD_FUNC(K, func) \
129 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
131 /* Assemble the body code between the prologue & epilogue. */
132 static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
133 struct codegen_context *ctx,
136 const struct sock_filter *filter = fp->insns;
139 unsigned int true_cond;
142 /* Start of epilogue code */
143 unsigned int exit_addr = addrs[flen];
145 for (i = 0; i < flen; i++) {
146 unsigned int K = filter[i].k;
149 * addrs[] maps a BPF bytecode address into a real offset from
150 * the start of the body code.
152 addrs[i] = ctx->idx * 4;
154 switch (filter[i].code) {
156 case BPF_S_ALU_ADD_X: /* A += X; */
157 ctx->seen |= SEEN_XREG;
158 PPC_ADD(r_A, r_A, r_X);
160 case BPF_S_ALU_ADD_K: /* A += K; */
163 PPC_ADDI(r_A, r_A, IMM_L(K));
165 PPC_ADDIS(r_A, r_A, IMM_HA(K));
167 case BPF_S_ALU_SUB_X: /* A -= X; */
168 ctx->seen |= SEEN_XREG;
169 PPC_SUB(r_A, r_A, r_X);
171 case BPF_S_ALU_SUB_K: /* A -= K */
174 PPC_ADDI(r_A, r_A, IMM_L(-K));
176 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
178 case BPF_S_ALU_MUL_X: /* A *= X; */
179 ctx->seen |= SEEN_XREG;
180 PPC_MUL(r_A, r_A, r_X);
182 case BPF_S_ALU_MUL_K: /* A *= K */
184 PPC_MULI(r_A, r_A, K);
186 PPC_LI32(r_scratch1, K);
187 PPC_MUL(r_A, r_A, r_scratch1);
190 case BPF_S_ALU_MOD_X: /* A %= X; */
191 ctx->seen |= SEEN_XREG;
193 if (ctx->pc_ret0 != -1) {
194 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
196 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
200 PPC_DIVWU(r_scratch1, r_A, r_X);
201 PPC_MUL(r_scratch1, r_X, r_scratch1);
202 PPC_SUB(r_A, r_A, r_scratch1);
204 case BPF_S_ALU_MOD_K: /* A %= K; */
205 PPC_LI32(r_scratch2, K);
206 PPC_DIVWU(r_scratch1, r_A, r_scratch2);
207 PPC_MUL(r_scratch1, r_scratch2, r_scratch1);
208 PPC_SUB(r_A, r_A, r_scratch1);
210 case BPF_S_ALU_DIV_X: /* A /= X; */
211 ctx->seen |= SEEN_XREG;
213 if (ctx->pc_ret0 != -1) {
214 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
217 * Exit, returning 0; first pass hits here
218 * (longer worst-case code size).
220 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
224 PPC_DIVWU(r_A, r_A, r_X);
226 case BPF_S_ALU_DIV_K: /* A /= K */
229 PPC_LI32(r_scratch1, K);
230 PPC_DIVWU(r_A, r_A, r_scratch1);
232 case BPF_S_ALU_AND_X:
233 ctx->seen |= SEEN_XREG;
234 PPC_AND(r_A, r_A, r_X);
236 case BPF_S_ALU_AND_K:
238 PPC_ANDI(r_A, r_A, K);
240 PPC_LI32(r_scratch1, K);
241 PPC_AND(r_A, r_A, r_scratch1);
245 ctx->seen |= SEEN_XREG;
246 PPC_OR(r_A, r_A, r_X);
250 PPC_ORI(r_A, r_A, IMM_L(K));
252 PPC_ORIS(r_A, r_A, IMM_H(K));
254 case BPF_S_ANC_ALU_XOR_X:
255 case BPF_S_ALU_XOR_X: /* A ^= X */
256 ctx->seen |= SEEN_XREG;
257 PPC_XOR(r_A, r_A, r_X);
259 case BPF_S_ALU_XOR_K: /* A ^= K */
261 PPC_XORI(r_A, r_A, IMM_L(K));
263 PPC_XORIS(r_A, r_A, IMM_H(K));
265 case BPF_S_ALU_LSH_X: /* A <<= X; */
266 ctx->seen |= SEEN_XREG;
267 PPC_SLW(r_A, r_A, r_X);
269 case BPF_S_ALU_LSH_K:
273 PPC_SLWI(r_A, r_A, K);
275 case BPF_S_ALU_RSH_X: /* A >>= X; */
276 ctx->seen |= SEEN_XREG;
277 PPC_SRW(r_A, r_A, r_X);
279 case BPF_S_ALU_RSH_K: /* A >>= K; */
283 PPC_SRWI(r_A, r_A, K);
291 if (ctx->pc_ret0 == -1)
295 * If this isn't the very last instruction, branch to
296 * the epilogue if we've stuff to clean up. Otherwise,
297 * if there's nothing to tidy, just return. If we /are/
298 * the last instruction, we're about to fall through to
299 * the epilogue to return.
303 * Note: 'seen' is properly valid only on pass
304 * #2. Both parts of this conditional are the
305 * same instruction size though, meaning the
306 * first pass will still correctly determine the
307 * code size/addresses.
324 case BPF_S_MISC_TAX: /* X = A */
327 case BPF_S_MISC_TXA: /* A = X */
328 ctx->seen |= SEEN_XREG;
332 /*** Constant loads/M[] access ***/
333 case BPF_S_LD_IMM: /* A = K */
336 case BPF_S_LDX_IMM: /* X = K */
339 case BPF_S_LD_MEM: /* A = mem[K] */
340 PPC_MR(r_A, r_M + (K & 0xf));
341 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
343 case BPF_S_LDX_MEM: /* X = mem[K] */
344 PPC_MR(r_X, r_M + (K & 0xf));
345 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
347 case BPF_S_ST: /* mem[K] = A */
348 PPC_MR(r_M + (K & 0xf), r_A);
349 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
351 case BPF_S_STX: /* mem[K] = X */
352 PPC_MR(r_M + (K & 0xf), r_X);
353 ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf));
355 case BPF_S_LD_W_LEN: /* A = skb->len; */
356 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
357 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len));
359 case BPF_S_LDX_W_LEN: /* X = skb->len; */
360 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
363 /*** Ancillary info loads ***/
364 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
365 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
367 PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff,
370 case BPF_S_ANC_IFINDEX:
371 PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
373 PPC_CMPDI(r_scratch1, 0);
374 if (ctx->pc_ret0 != -1) {
375 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
377 /* Exit, returning 0; first pass hits here. */
378 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
382 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
384 PPC_LWZ_OFFS(r_A, r_scratch1,
385 offsetof(struct net_device, ifindex));
388 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
389 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
392 case BPF_S_ANC_RXHASH:
393 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
394 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
397 case BPF_S_ANC_VLAN_TAG:
398 case BPF_S_ANC_VLAN_TAG_PRESENT:
399 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
400 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
402 if (filter[i].code == BPF_S_ANC_VLAN_TAG)
403 PPC_ANDI(r_A, r_A, VLAN_VID_MASK);
405 PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT);
407 case BPF_S_ANC_QUEUE:
408 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
409 queue_mapping) != 2);
410 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
417 * raw_smp_processor_id() = local_paca->paca_index
419 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct,
421 PPC_LHZ_OFFS(r_A, 13,
422 offsetof(struct paca_struct, paca_index));
428 /*** Absolute loads from packet header/data ***/
430 func = CHOOSE_LOAD_FUNC(K, sk_load_word);
433 func = CHOOSE_LOAD_FUNC(K, sk_load_half);
436 func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
439 ctx->seen |= SEEN_DATAREF;
440 PPC_LI64(r_scratch1, func);
441 PPC_MTLR(r_scratch1);
445 * Helper returns 'lt' condition on error, and an
446 * appropriate return value in r3
448 PPC_BCC(COND_LT, exit_addr);
451 /*** Indirect loads from packet header/data ***/
454 goto common_load_ind;
457 goto common_load_ind;
462 * Load from [X + K]. Negative offsets are tested for
463 * in the helper functions.
465 ctx->seen |= SEEN_DATAREF | SEEN_XREG;
466 PPC_LI64(r_scratch1, func);
467 PPC_MTLR(r_scratch1);
468 PPC_ADDI(r_addr, r_X, IMM_L(K));
470 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
472 /* If error, cr0.LT set */
473 PPC_BCC(COND_LT, exit_addr);
476 case BPF_S_LDX_B_MSH:
477 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
481 /*** Jump and branches ***/
484 PPC_JMP(addrs[i + 1 + K]);
487 case BPF_S_JMP_JGT_K:
488 case BPF_S_JMP_JGT_X:
491 case BPF_S_JMP_JGE_K:
492 case BPF_S_JMP_JGE_X:
495 case BPF_S_JMP_JEQ_K:
496 case BPF_S_JMP_JEQ_X:
499 case BPF_S_JMP_JSET_K:
500 case BPF_S_JMP_JSET_X:
504 /* same targets, can avoid doing the test :) */
505 if (filter[i].jt == filter[i].jf) {
506 if (filter[i].jt > 0)
507 PPC_JMP(addrs[i + 1 + filter[i].jt]);
511 switch (filter[i].code) {
512 case BPF_S_JMP_JGT_X:
513 case BPF_S_JMP_JGE_X:
514 case BPF_S_JMP_JEQ_X:
515 ctx->seen |= SEEN_XREG;
518 case BPF_S_JMP_JSET_X:
519 ctx->seen |= SEEN_XREG;
520 PPC_AND_DOT(r_scratch1, r_A, r_X);
522 case BPF_S_JMP_JEQ_K:
523 case BPF_S_JMP_JGT_K:
524 case BPF_S_JMP_JGE_K:
528 PPC_LI32(r_scratch1, K);
529 PPC_CMPLW(r_A, r_scratch1);
532 case BPF_S_JMP_JSET_K:
534 /* PPC_ANDI is /only/ dot-form */
535 PPC_ANDI(r_scratch1, r_A, K);
537 PPC_LI32(r_scratch1, K);
538 PPC_AND_DOT(r_scratch1, r_A,
543 /* Sometimes branches are constructed "backward", with
544 * the false path being the branch and true path being
545 * a fallthrough to the next instruction.
547 if (filter[i].jt == 0)
548 /* Swap the sense of the branch */
549 PPC_BCC(true_cond ^ COND_CMP_TRUE,
550 addrs[i + 1 + filter[i].jf]);
552 PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]);
553 if (filter[i].jf != 0)
554 PPC_JMP(addrs[i + 1 + filter[i].jf]);
558 /* The filter contains something cruel & unusual.
559 * We don't handle it, but also there shouldn't be
560 * anything missing from our list.
562 if (printk_ratelimit())
563 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
569 /* Set end-of-body-code address for exit. */
570 addrs[i] = ctx->idx * 4;
575 void bpf_jit_compile(struct sk_filter *fp)
577 unsigned int proglen;
578 unsigned int alloclen;
582 struct codegen_context cgctx;
589 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
594 * There are multiple assembly passes as the generated code will change
595 * size as it settles down, figuring out the max branch offsets/exit
598 * The range of standard conditional branches is +/- 32Kbytes. Since
599 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
600 * finish with 8 bytes/instruction. Not feasible, so long jumps are
601 * used, distinct from short branches.
605 * For now, both branch types assemble to 2 words (short branches padded
606 * with a NOP); this is less efficient, but assembly will always complete
607 * after exactly 3 passes:
609 * First pass: No code buffer; Program is "faux-generated" -- no code
610 * emitted but maximum size of output determined (and addrs[] filled
611 * in). Also, we note whether we use M[], whether we use skb data, etc.
612 * All generation choices assumed to be 'worst-case', e.g. branches all
613 * far (2 instructions), return path code reduction not available, etc.
615 * Second pass: Code buffer allocated with size determined previously.
616 * Prologue generated to support features we have seen used. Exit paths
617 * determined and addrs[] is filled in again, as code may be slightly
618 * smaller as a result.
620 * Third pass: Code generated 'for real', and branch destinations
621 * determined from now-accurate addrs[] map.
625 * If we optimise this, near branches will be shorter. On the
626 * first assembly pass, we should err on the side of caution and
627 * generate the biggest code. On subsequent passes, branches will be
628 * generated short or long and code size will reduce. With smaller
629 * code, more branches may fall into the short category, and code will
632 * Finally, if we see one pass generate code the same size as the
633 * previous pass we have converged and should now generate code for
634 * real. Allocating at the end will also save the memory that would
635 * otherwise be wasted by the (small) current code shrinkage.
636 * Preferably, we should do a small number of passes (e.g. 5) and if we
637 * haven't converged by then, get impatient and force code to generate
638 * as-is, even if the odd branch would be left long. The chances of a
639 * long jump are tiny with all but the most enormous of BPF filter
640 * inputs, so we should usually converge on the third pass.
646 /* Scouting faux-generate pass 0 */
647 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
648 /* We hit something illegal or unsupported. */
652 * Pretend to build prologue, given the features we've seen. This will
653 * update ctgtx.idx as it pretends to output instructions, then we can
654 * calculate total size from idx.
656 bpf_jit_build_prologue(fp, 0, &cgctx);
657 bpf_jit_build_epilogue(0, &cgctx);
659 proglen = cgctx.idx * 4;
660 alloclen = proglen + FUNCTION_DESCR_SIZE;
661 image = module_alloc(alloclen);
665 code_base = image + (FUNCTION_DESCR_SIZE/4);
667 /* Code generation passes 1-2 */
668 for (pass = 1; pass < 3; pass++) {
669 /* Now build the prologue, body code & epilogue for real. */
671 bpf_jit_build_prologue(fp, code_base, &cgctx);
672 bpf_jit_build_body(fp, code_base, &cgctx, addrs);
673 bpf_jit_build_epilogue(code_base, &cgctx);
675 if (bpf_jit_enable > 1)
676 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
677 proglen - (cgctx.idx * 4), cgctx.seen);
680 if (bpf_jit_enable > 1)
681 /* Note that we output the base address of the code_base
682 * rather than image, since opcodes are in code_base.
684 bpf_jit_dump(flen, proglen, pass, code_base);
687 bpf_flush_icache(code_base, code_base + (proglen/4));
688 /* Function descriptor nastiness: Address + TOC */
689 ((u64 *)image)[0] = (u64)code_base;
690 ((u64 *)image)[1] = local_paca->kernel_toc;
691 fp->bpf_func = (void *)image;
699 void bpf_jit_free(struct sk_filter *fp)
702 module_free(NULL, fp->bpf_func);