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 bpf_prog *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) {
83 case BPF_LD | BPF_W | BPF_LEN:
84 case BPF_LD | BPF_W | BPF_ABS:
85 case BPF_LD | BPF_H | BPF_ABS:
86 case BPF_LD | BPF_B | BPF_ABS:
87 /* first instruction sets A register (or is RET 'constant') */
90 /* make sure we dont leak kernel information to user */
95 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
99 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
100 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
101 if (ctx->seen & SEEN_DATAREF) {
104 PPC_LD(r_D, 1, -(8*(32-r_D)));
105 PPC_LD(r_HL, 1, -(8*(32-r_HL)));
107 if (ctx->seen & SEEN_MEM) {
108 /* Restore any saved non-vol registers */
109 for (i = r_M; i < (r_M+16); i++) {
110 if (ctx->seen & (1 << (i-r_M)))
111 PPC_LD(i, 1, -(8*(32-i)));
115 /* The RETs have left a return value in R3. */
120 #define CHOOSE_LOAD_FUNC(K, func) \
121 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
123 /* Assemble the body code between the prologue & epilogue. */
124 static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
125 struct codegen_context *ctx,
128 const struct sock_filter *filter = fp->insns;
131 unsigned int true_cond;
134 /* Start of epilogue code */
135 unsigned int exit_addr = addrs[flen];
137 for (i = 0; i < flen; i++) {
138 unsigned int K = filter[i].k;
139 u16 code = bpf_anc_helper(&filter[i]);
142 * addrs[] maps a BPF bytecode address into a real offset from
143 * the start of the body code.
145 addrs[i] = ctx->idx * 4;
149 case BPF_ALU | BPF_ADD | BPF_X: /* A += X; */
150 ctx->seen |= SEEN_XREG;
151 PPC_ADD(r_A, r_A, r_X);
153 case BPF_ALU | BPF_ADD | BPF_K: /* A += K; */
156 PPC_ADDI(r_A, r_A, IMM_L(K));
158 PPC_ADDIS(r_A, r_A, IMM_HA(K));
160 case BPF_ALU | BPF_SUB | BPF_X: /* A -= X; */
161 ctx->seen |= SEEN_XREG;
162 PPC_SUB(r_A, r_A, r_X);
164 case BPF_ALU | BPF_SUB | BPF_K: /* A -= K */
167 PPC_ADDI(r_A, r_A, IMM_L(-K));
169 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
171 case BPF_ALU | BPF_MUL | BPF_X: /* A *= X; */
172 ctx->seen |= SEEN_XREG;
173 PPC_MUL(r_A, r_A, r_X);
175 case BPF_ALU | BPF_MUL | BPF_K: /* A *= K */
177 PPC_MULI(r_A, r_A, K);
179 PPC_LI32(r_scratch1, K);
180 PPC_MUL(r_A, r_A, r_scratch1);
183 case BPF_ALU | BPF_MOD | BPF_X: /* A %= X; */
184 ctx->seen |= SEEN_XREG;
186 if (ctx->pc_ret0 != -1) {
187 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
189 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
193 PPC_DIVWU(r_scratch1, r_A, r_X);
194 PPC_MUL(r_scratch1, r_X, r_scratch1);
195 PPC_SUB(r_A, r_A, r_scratch1);
197 case BPF_ALU | BPF_MOD | BPF_K: /* A %= K; */
198 PPC_LI32(r_scratch2, K);
199 PPC_DIVWU(r_scratch1, r_A, r_scratch2);
200 PPC_MUL(r_scratch1, r_scratch2, r_scratch1);
201 PPC_SUB(r_A, r_A, r_scratch1);
203 case BPF_ALU | BPF_DIV | BPF_X: /* A /= X; */
204 ctx->seen |= SEEN_XREG;
206 if (ctx->pc_ret0 != -1) {
207 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
210 * Exit, returning 0; first pass hits here
211 * (longer worst-case code size).
213 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
217 PPC_DIVWU(r_A, r_A, r_X);
219 case BPF_ALU | BPF_DIV | BPF_K: /* A /= K */
222 PPC_LI32(r_scratch1, K);
223 PPC_DIVWU(r_A, r_A, r_scratch1);
225 case BPF_ALU | BPF_AND | BPF_X:
226 ctx->seen |= SEEN_XREG;
227 PPC_AND(r_A, r_A, r_X);
229 case BPF_ALU | BPF_AND | BPF_K:
231 PPC_ANDI(r_A, r_A, K);
233 PPC_LI32(r_scratch1, K);
234 PPC_AND(r_A, r_A, r_scratch1);
237 case BPF_ALU | BPF_OR | BPF_X:
238 ctx->seen |= SEEN_XREG;
239 PPC_OR(r_A, r_A, r_X);
241 case BPF_ALU | BPF_OR | BPF_K:
243 PPC_ORI(r_A, r_A, IMM_L(K));
245 PPC_ORIS(r_A, r_A, IMM_H(K));
247 case BPF_ANC | SKF_AD_ALU_XOR_X:
248 case BPF_ALU | BPF_XOR | BPF_X: /* A ^= X */
249 ctx->seen |= SEEN_XREG;
250 PPC_XOR(r_A, r_A, r_X);
252 case BPF_ALU | BPF_XOR | BPF_K: /* A ^= K */
254 PPC_XORI(r_A, r_A, IMM_L(K));
256 PPC_XORIS(r_A, r_A, IMM_H(K));
258 case BPF_ALU | BPF_LSH | BPF_X: /* A <<= X; */
259 ctx->seen |= SEEN_XREG;
260 PPC_SLW(r_A, r_A, r_X);
262 case BPF_ALU | BPF_LSH | BPF_K:
266 PPC_SLWI(r_A, r_A, K);
268 case BPF_ALU | BPF_RSH | BPF_X: /* A >>= X; */
269 ctx->seen |= SEEN_XREG;
270 PPC_SRW(r_A, r_A, r_X);
272 case BPF_ALU | BPF_RSH | BPF_K: /* A >>= K; */
276 PPC_SRWI(r_A, r_A, K);
278 case BPF_ALU | BPF_NEG:
281 case BPF_RET | BPF_K:
284 if (ctx->pc_ret0 == -1)
288 * If this isn't the very last instruction, branch to
289 * the epilogue if we've stuff to clean up. Otherwise,
290 * if there's nothing to tidy, just return. If we /are/
291 * the last instruction, we're about to fall through to
292 * the epilogue to return.
296 * Note: 'seen' is properly valid only on pass
297 * #2. Both parts of this conditional are the
298 * same instruction size though, meaning the
299 * first pass will still correctly determine the
300 * code size/addresses.
308 case BPF_RET | BPF_A:
317 case BPF_MISC | BPF_TAX: /* X = A */
320 case BPF_MISC | BPF_TXA: /* A = X */
321 ctx->seen |= SEEN_XREG;
325 /*** Constant loads/M[] access ***/
326 case BPF_LD | BPF_IMM: /* A = K */
329 case BPF_LDX | BPF_IMM: /* X = K */
332 case BPF_LD | BPF_MEM: /* A = mem[K] */
333 PPC_MR(r_A, r_M + (K & 0xf));
334 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
336 case BPF_LDX | BPF_MEM: /* X = mem[K] */
337 PPC_MR(r_X, r_M + (K & 0xf));
338 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
340 case BPF_ST: /* mem[K] = A */
341 PPC_MR(r_M + (K & 0xf), r_A);
342 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
344 case BPF_STX: /* mem[K] = X */
345 PPC_MR(r_M + (K & 0xf), r_X);
346 ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf));
348 case BPF_LD | BPF_W | BPF_LEN: /* A = skb->len; */
349 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
350 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len));
352 case BPF_LDX | BPF_W | BPF_LEN: /* X = skb->len; */
353 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
356 /*** Ancillary info loads ***/
357 case BPF_ANC | SKF_AD_PROTOCOL: /* A = ntohs(skb->protocol); */
358 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
360 PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff,
363 case BPF_ANC | SKF_AD_IFINDEX:
364 PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
366 PPC_CMPDI(r_scratch1, 0);
367 if (ctx->pc_ret0 != -1) {
368 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
370 /* Exit, returning 0; first pass hits here. */
371 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
375 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
377 PPC_LWZ_OFFS(r_A, r_scratch1,
378 offsetof(struct net_device, ifindex));
380 case BPF_ANC | SKF_AD_MARK:
381 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
382 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
385 case BPF_ANC | SKF_AD_RXHASH:
386 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
387 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
390 case BPF_ANC | SKF_AD_VLAN_TAG:
391 case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
392 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
393 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
395 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
397 if (code == (BPF_ANC | SKF_AD_VLAN_TAG)) {
398 PPC_ANDI(r_A, r_A, ~VLAN_TAG_PRESENT);
400 PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT);
401 PPC_SRWI(r_A, r_A, 12);
404 case BPF_ANC | SKF_AD_QUEUE:
405 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
406 queue_mapping) != 2);
407 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
410 case BPF_ANC | SKF_AD_CPU:
414 * raw_smp_processor_id() = local_paca->paca_index
416 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct,
418 PPC_LHZ_OFFS(r_A, 13,
419 offsetof(struct paca_struct, paca_index));
425 /*** Absolute loads from packet header/data ***/
426 case BPF_LD | BPF_W | BPF_ABS:
427 func = CHOOSE_LOAD_FUNC(K, sk_load_word);
429 case BPF_LD | BPF_H | BPF_ABS:
430 func = CHOOSE_LOAD_FUNC(K, sk_load_half);
432 case BPF_LD | BPF_B | BPF_ABS:
433 func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
436 ctx->seen |= SEEN_DATAREF;
437 PPC_LI64(r_scratch1, func);
438 PPC_MTLR(r_scratch1);
442 * Helper returns 'lt' condition on error, and an
443 * appropriate return value in r3
445 PPC_BCC(COND_LT, exit_addr);
448 /*** Indirect loads from packet header/data ***/
449 case BPF_LD | BPF_W | BPF_IND:
451 goto common_load_ind;
452 case BPF_LD | BPF_H | BPF_IND:
454 goto common_load_ind;
455 case BPF_LD | BPF_B | BPF_IND:
459 * Load from [X + K]. Negative offsets are tested for
460 * in the helper functions.
462 ctx->seen |= SEEN_DATAREF | SEEN_XREG;
463 PPC_LI64(r_scratch1, func);
464 PPC_MTLR(r_scratch1);
465 PPC_ADDI(r_addr, r_X, IMM_L(K));
467 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
469 /* If error, cr0.LT set */
470 PPC_BCC(COND_LT, exit_addr);
473 case BPF_LDX | BPF_B | BPF_MSH:
474 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
478 /*** Jump and branches ***/
479 case BPF_JMP | BPF_JA:
481 PPC_JMP(addrs[i + 1 + K]);
484 case BPF_JMP | BPF_JGT | BPF_K:
485 case BPF_JMP | BPF_JGT | BPF_X:
488 case BPF_JMP | BPF_JGE | BPF_K:
489 case BPF_JMP | BPF_JGE | BPF_X:
492 case BPF_JMP | BPF_JEQ | BPF_K:
493 case BPF_JMP | BPF_JEQ | BPF_X:
496 case BPF_JMP | BPF_JSET | BPF_K:
497 case BPF_JMP | BPF_JSET | BPF_X:
501 /* same targets, can avoid doing the test :) */
502 if (filter[i].jt == filter[i].jf) {
503 if (filter[i].jt > 0)
504 PPC_JMP(addrs[i + 1 + filter[i].jt]);
509 case BPF_JMP | BPF_JGT | BPF_X:
510 case BPF_JMP | BPF_JGE | BPF_X:
511 case BPF_JMP | BPF_JEQ | BPF_X:
512 ctx->seen |= SEEN_XREG;
515 case BPF_JMP | BPF_JSET | BPF_X:
516 ctx->seen |= SEEN_XREG;
517 PPC_AND_DOT(r_scratch1, r_A, r_X);
519 case BPF_JMP | BPF_JEQ | BPF_K:
520 case BPF_JMP | BPF_JGT | BPF_K:
521 case BPF_JMP | BPF_JGE | BPF_K:
525 PPC_LI32(r_scratch1, K);
526 PPC_CMPLW(r_A, r_scratch1);
529 case BPF_JMP | BPF_JSET | BPF_K:
531 /* PPC_ANDI is /only/ dot-form */
532 PPC_ANDI(r_scratch1, r_A, K);
534 PPC_LI32(r_scratch1, K);
535 PPC_AND_DOT(r_scratch1, r_A,
540 /* Sometimes branches are constructed "backward", with
541 * the false path being the branch and true path being
542 * a fallthrough to the next instruction.
544 if (filter[i].jt == 0)
545 /* Swap the sense of the branch */
546 PPC_BCC(true_cond ^ COND_CMP_TRUE,
547 addrs[i + 1 + filter[i].jf]);
549 PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]);
550 if (filter[i].jf != 0)
551 PPC_JMP(addrs[i + 1 + filter[i].jf]);
555 /* The filter contains something cruel & unusual.
556 * We don't handle it, but also there shouldn't be
557 * anything missing from our list.
559 if (printk_ratelimit())
560 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
566 /* Set end-of-body-code address for exit. */
567 addrs[i] = ctx->idx * 4;
572 void bpf_jit_compile(struct bpf_prog *fp)
574 unsigned int proglen;
575 unsigned int alloclen;
579 struct codegen_context cgctx;
586 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
591 * There are multiple assembly passes as the generated code will change
592 * size as it settles down, figuring out the max branch offsets/exit
595 * The range of standard conditional branches is +/- 32Kbytes. Since
596 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
597 * finish with 8 bytes/instruction. Not feasible, so long jumps are
598 * used, distinct from short branches.
602 * For now, both branch types assemble to 2 words (short branches padded
603 * with a NOP); this is less efficient, but assembly will always complete
604 * after exactly 3 passes:
606 * First pass: No code buffer; Program is "faux-generated" -- no code
607 * emitted but maximum size of output determined (and addrs[] filled
608 * in). Also, we note whether we use M[], whether we use skb data, etc.
609 * All generation choices assumed to be 'worst-case', e.g. branches all
610 * far (2 instructions), return path code reduction not available, etc.
612 * Second pass: Code buffer allocated with size determined previously.
613 * Prologue generated to support features we have seen used. Exit paths
614 * determined and addrs[] is filled in again, as code may be slightly
615 * smaller as a result.
617 * Third pass: Code generated 'for real', and branch destinations
618 * determined from now-accurate addrs[] map.
622 * If we optimise this, near branches will be shorter. On the
623 * first assembly pass, we should err on the side of caution and
624 * generate the biggest code. On subsequent passes, branches will be
625 * generated short or long and code size will reduce. With smaller
626 * code, more branches may fall into the short category, and code will
629 * Finally, if we see one pass generate code the same size as the
630 * previous pass we have converged and should now generate code for
631 * real. Allocating at the end will also save the memory that would
632 * otherwise be wasted by the (small) current code shrinkage.
633 * Preferably, we should do a small number of passes (e.g. 5) and if we
634 * haven't converged by then, get impatient and force code to generate
635 * as-is, even if the odd branch would be left long. The chances of a
636 * long jump are tiny with all but the most enormous of BPF filter
637 * inputs, so we should usually converge on the third pass.
643 /* Scouting faux-generate pass 0 */
644 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
645 /* We hit something illegal or unsupported. */
649 * Pretend to build prologue, given the features we've seen. This will
650 * update ctgtx.idx as it pretends to output instructions, then we can
651 * calculate total size from idx.
653 bpf_jit_build_prologue(fp, 0, &cgctx);
654 bpf_jit_build_epilogue(0, &cgctx);
656 proglen = cgctx.idx * 4;
657 alloclen = proglen + FUNCTION_DESCR_SIZE;
658 image = module_alloc(alloclen);
662 code_base = image + (FUNCTION_DESCR_SIZE/4);
664 /* Code generation passes 1-2 */
665 for (pass = 1; pass < 3; pass++) {
666 /* Now build the prologue, body code & epilogue for real. */
668 bpf_jit_build_prologue(fp, code_base, &cgctx);
669 bpf_jit_build_body(fp, code_base, &cgctx, addrs);
670 bpf_jit_build_epilogue(code_base, &cgctx);
672 if (bpf_jit_enable > 1)
673 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
674 proglen - (cgctx.idx * 4), cgctx.seen);
677 if (bpf_jit_enable > 1)
678 /* Note that we output the base address of the code_base
679 * rather than image, since opcodes are in code_base.
681 bpf_jit_dump(flen, proglen, pass, code_base);
684 bpf_flush_icache(code_base, code_base + (proglen/4));
685 /* Function descriptor nastiness: Address + TOC */
686 ((u64 *)image)[0] = (u64)code_base;
687 ((u64 *)image)[1] = local_paca->kernel_toc;
688 fp->bpf_func = (void *)image;
696 void bpf_jit_free(struct bpf_prog *fp)
699 module_free(NULL, fp->bpf_func);
701 bpf_prog_unlock_free(fp);