1 /* bpf_jit_comp.c: BPF JIT compiler
3 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
5 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com)
6 * Ported to ppc32 by Denis Kirjanov <kda@linux-powerpc.org>
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License
10 * as published by the Free Software Foundation; version 2
13 #include <linux/moduleloader.h>
14 #include <asm/cacheflush.h>
15 #include <linux/netdevice.h>
16 #include <linux/filter.h>
17 #include <linux/if_vlan.h>
19 #include "bpf_jit32.h"
21 int bpf_jit_enable __read_mostly;
23 static inline void bpf_flush_icache(void *start, void *end)
26 flush_icache_range((unsigned long)start, (unsigned long)end);
29 static void bpf_jit_build_prologue(struct bpf_prog *fp, u32 *image,
30 struct codegen_context *ctx)
33 const struct sock_filter *filter = fp->insns;
35 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
37 if (ctx->seen & SEEN_DATAREF) {
38 /* If we call any helpers (for loads), save LR */
39 EMIT(PPC_INST_MFLR | __PPC_RT(R0));
40 PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
42 /* Back up non-volatile regs. */
43 PPC_BPF_STL(r_D, 1, -(REG_SZ*(32-r_D)));
44 PPC_BPF_STL(r_HL, 1, -(REG_SZ*(32-r_HL)));
46 if (ctx->seen & SEEN_MEM) {
48 * Conditionally save regs r15-r31 as some will be used
51 for (i = r_M; i < (r_M+16); i++) {
52 if (ctx->seen & (1 << (i-r_M)))
53 PPC_BPF_STL(i, 1, -(REG_SZ*(32-i)));
56 PPC_BPF_STLU(1, 1, -BPF_PPC_STACKFRAME);
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_LL_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 /* make sure we dont leak kernel information to user */
82 if (bpf_needs_clear_a(&filter[0]))
86 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
90 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
91 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
92 if (ctx->seen & SEEN_DATAREF) {
93 PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
95 PPC_BPF_LL(r_D, 1, -(REG_SZ*(32-r_D)));
96 PPC_BPF_LL(r_HL, 1, -(REG_SZ*(32-r_HL)));
98 if (ctx->seen & SEEN_MEM) {
99 /* Restore any saved non-vol registers */
100 for (i = r_M; i < (r_M+16); i++) {
101 if (ctx->seen & (1 << (i-r_M)))
102 PPC_BPF_LL(i, 1, -(REG_SZ*(32-i)));
106 /* The RETs have left a return value in R3. */
111 #define CHOOSE_LOAD_FUNC(K, func) \
112 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
114 /* Assemble the body code between the prologue & epilogue. */
115 static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
116 struct codegen_context *ctx,
119 const struct sock_filter *filter = fp->insns;
122 unsigned int true_cond;
125 /* Start of epilogue code */
126 unsigned int exit_addr = addrs[flen];
128 for (i = 0; i < flen; i++) {
129 unsigned int K = filter[i].k;
130 u16 code = bpf_anc_helper(&filter[i]);
133 * addrs[] maps a BPF bytecode address into a real offset from
134 * the start of the body code.
136 addrs[i] = ctx->idx * 4;
140 case BPF_ALU | BPF_ADD | BPF_X: /* A += X; */
141 ctx->seen |= SEEN_XREG;
142 PPC_ADD(r_A, r_A, r_X);
144 case BPF_ALU | BPF_ADD | BPF_K: /* A += K; */
147 PPC_ADDI(r_A, r_A, IMM_L(K));
149 PPC_ADDIS(r_A, r_A, IMM_HA(K));
151 case BPF_ALU | BPF_SUB | BPF_X: /* A -= X; */
152 ctx->seen |= SEEN_XREG;
153 PPC_SUB(r_A, r_A, r_X);
155 case BPF_ALU | BPF_SUB | BPF_K: /* A -= K */
158 PPC_ADDI(r_A, r_A, IMM_L(-K));
160 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
162 case BPF_ALU | BPF_MUL | BPF_X: /* A *= X; */
163 ctx->seen |= SEEN_XREG;
164 PPC_MULW(r_A, r_A, r_X);
166 case BPF_ALU | BPF_MUL | BPF_K: /* A *= K */
168 PPC_MULI(r_A, r_A, K);
170 PPC_LI32(r_scratch1, K);
171 PPC_MULW(r_A, r_A, r_scratch1);
174 case BPF_ALU | BPF_MOD | BPF_X: /* A %= X; */
175 case BPF_ALU | BPF_DIV | BPF_X: /* A /= X; */
176 ctx->seen |= SEEN_XREG;
178 if (ctx->pc_ret0 != -1) {
179 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
181 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
185 if (code == (BPF_ALU | BPF_MOD | BPF_X)) {
186 PPC_DIVWU(r_scratch1, r_A, r_X);
187 PPC_MULW(r_scratch1, r_X, r_scratch1);
188 PPC_SUB(r_A, r_A, r_scratch1);
190 PPC_DIVWU(r_A, r_A, r_X);
193 case BPF_ALU | BPF_MOD | BPF_K: /* A %= K; */
194 PPC_LI32(r_scratch2, K);
195 PPC_DIVWU(r_scratch1, r_A, r_scratch2);
196 PPC_MULW(r_scratch1, r_scratch2, r_scratch1);
197 PPC_SUB(r_A, r_A, r_scratch1);
199 case BPF_ALU | BPF_DIV | BPF_K: /* A /= K */
202 PPC_LI32(r_scratch1, K);
203 PPC_DIVWU(r_A, r_A, r_scratch1);
205 case BPF_ALU | BPF_AND | BPF_X:
206 ctx->seen |= SEEN_XREG;
207 PPC_AND(r_A, r_A, r_X);
209 case BPF_ALU | BPF_AND | BPF_K:
211 PPC_ANDI(r_A, r_A, K);
213 PPC_LI32(r_scratch1, K);
214 PPC_AND(r_A, r_A, r_scratch1);
217 case BPF_ALU | BPF_OR | BPF_X:
218 ctx->seen |= SEEN_XREG;
219 PPC_OR(r_A, r_A, r_X);
221 case BPF_ALU | BPF_OR | BPF_K:
223 PPC_ORI(r_A, r_A, IMM_L(K));
225 PPC_ORIS(r_A, r_A, IMM_H(K));
227 case BPF_ANC | SKF_AD_ALU_XOR_X:
228 case BPF_ALU | BPF_XOR | BPF_X: /* A ^= X */
229 ctx->seen |= SEEN_XREG;
230 PPC_XOR(r_A, r_A, r_X);
232 case BPF_ALU | BPF_XOR | BPF_K: /* A ^= K */
234 PPC_XORI(r_A, r_A, IMM_L(K));
236 PPC_XORIS(r_A, r_A, IMM_H(K));
238 case BPF_ALU | BPF_LSH | BPF_X: /* A <<= X; */
239 ctx->seen |= SEEN_XREG;
240 PPC_SLW(r_A, r_A, r_X);
242 case BPF_ALU | BPF_LSH | BPF_K:
246 PPC_SLWI(r_A, r_A, K);
248 case BPF_ALU | BPF_RSH | BPF_X: /* A >>= X; */
249 ctx->seen |= SEEN_XREG;
250 PPC_SRW(r_A, r_A, r_X);
252 case BPF_ALU | BPF_RSH | BPF_K: /* A >>= K; */
256 PPC_SRWI(r_A, r_A, K);
258 case BPF_ALU | BPF_NEG:
261 case BPF_RET | BPF_K:
264 if (ctx->pc_ret0 == -1)
268 * If this isn't the very last instruction, branch to
269 * the epilogue if we've stuff to clean up. Otherwise,
270 * if there's nothing to tidy, just return. If we /are/
271 * the last instruction, we're about to fall through to
272 * the epilogue to return.
276 * Note: 'seen' is properly valid only on pass
277 * #2. Both parts of this conditional are the
278 * same instruction size though, meaning the
279 * first pass will still correctly determine the
280 * code size/addresses.
288 case BPF_RET | BPF_A:
297 case BPF_MISC | BPF_TAX: /* X = A */
300 case BPF_MISC | BPF_TXA: /* A = X */
301 ctx->seen |= SEEN_XREG;
305 /*** Constant loads/M[] access ***/
306 case BPF_LD | BPF_IMM: /* A = K */
309 case BPF_LDX | BPF_IMM: /* X = K */
312 case BPF_LD | BPF_MEM: /* A = mem[K] */
313 PPC_MR(r_A, r_M + (K & 0xf));
314 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
316 case BPF_LDX | BPF_MEM: /* X = mem[K] */
317 PPC_MR(r_X, r_M + (K & 0xf));
318 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
320 case BPF_ST: /* mem[K] = A */
321 PPC_MR(r_M + (K & 0xf), r_A);
322 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
324 case BPF_STX: /* mem[K] = X */
325 PPC_MR(r_M + (K & 0xf), r_X);
326 ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf));
328 case BPF_LD | BPF_W | BPF_LEN: /* A = skb->len; */
329 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
330 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len));
332 case BPF_LDX | BPF_W | BPF_LEN: /* X = skb->len; */
333 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
336 /*** Ancillary info loads ***/
337 case BPF_ANC | SKF_AD_PROTOCOL: /* A = ntohs(skb->protocol); */
338 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
340 PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff,
343 case BPF_ANC | SKF_AD_IFINDEX:
344 case BPF_ANC | SKF_AD_HATYPE:
345 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
347 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
349 PPC_LL_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
351 PPC_CMPDI(r_scratch1, 0);
352 if (ctx->pc_ret0 != -1) {
353 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
355 /* Exit, returning 0; first pass hits here. */
356 PPC_BCC_SHORT(COND_NE, ctx->idx * 4 + 12);
360 if (code == (BPF_ANC | SKF_AD_IFINDEX)) {
361 PPC_LWZ_OFFS(r_A, r_scratch1,
362 offsetof(struct net_device, ifindex));
364 PPC_LHZ_OFFS(r_A, r_scratch1,
365 offsetof(struct net_device, type));
369 case BPF_ANC | SKF_AD_MARK:
370 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
371 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
374 case BPF_ANC | SKF_AD_RXHASH:
375 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
376 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
379 case BPF_ANC | SKF_AD_VLAN_TAG:
380 case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
381 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
382 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
384 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
386 if (code == (BPF_ANC | SKF_AD_VLAN_TAG)) {
387 PPC_ANDI(r_A, r_A, ~VLAN_TAG_PRESENT);
389 PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT);
390 PPC_SRWI(r_A, r_A, 12);
393 case BPF_ANC | SKF_AD_QUEUE:
394 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
395 queue_mapping) != 2);
396 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
399 case BPF_ANC | SKF_AD_PKTTYPE:
400 PPC_LBZ_OFFS(r_A, r_skb, PKT_TYPE_OFFSET());
401 PPC_ANDI(r_A, r_A, PKT_TYPE_MAX);
402 PPC_SRWI(r_A, r_A, 5);
404 case BPF_ANC | SKF_AD_CPU:
405 PPC_BPF_LOAD_CPU(r_A);
407 /*** Absolute loads from packet header/data ***/
408 case BPF_LD | BPF_W | BPF_ABS:
409 func = CHOOSE_LOAD_FUNC(K, sk_load_word);
411 case BPF_LD | BPF_H | BPF_ABS:
412 func = CHOOSE_LOAD_FUNC(K, sk_load_half);
414 case BPF_LD | BPF_B | BPF_ABS:
415 func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
418 ctx->seen |= SEEN_DATAREF;
419 PPC_FUNC_ADDR(r_scratch1, func);
420 PPC_MTLR(r_scratch1);
424 * Helper returns 'lt' condition on error, and an
425 * appropriate return value in r3
427 PPC_BCC(COND_LT, exit_addr);
430 /*** Indirect loads from packet header/data ***/
431 case BPF_LD | BPF_W | BPF_IND:
433 goto common_load_ind;
434 case BPF_LD | BPF_H | BPF_IND:
436 goto common_load_ind;
437 case BPF_LD | BPF_B | BPF_IND:
441 * Load from [X + K]. Negative offsets are tested for
442 * in the helper functions.
444 ctx->seen |= SEEN_DATAREF | SEEN_XREG;
445 PPC_FUNC_ADDR(r_scratch1, func);
446 PPC_MTLR(r_scratch1);
447 PPC_ADDI(r_addr, r_X, IMM_L(K));
449 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
451 /* If error, cr0.LT set */
452 PPC_BCC(COND_LT, exit_addr);
455 case BPF_LDX | BPF_B | BPF_MSH:
456 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
460 /*** Jump and branches ***/
461 case BPF_JMP | BPF_JA:
463 PPC_JMP(addrs[i + 1 + K]);
466 case BPF_JMP | BPF_JGT | BPF_K:
467 case BPF_JMP | BPF_JGT | BPF_X:
470 case BPF_JMP | BPF_JGE | BPF_K:
471 case BPF_JMP | BPF_JGE | BPF_X:
474 case BPF_JMP | BPF_JEQ | BPF_K:
475 case BPF_JMP | BPF_JEQ | BPF_X:
478 case BPF_JMP | BPF_JSET | BPF_K:
479 case BPF_JMP | BPF_JSET | BPF_X:
483 /* same targets, can avoid doing the test :) */
484 if (filter[i].jt == filter[i].jf) {
485 if (filter[i].jt > 0)
486 PPC_JMP(addrs[i + 1 + filter[i].jt]);
491 case BPF_JMP | BPF_JGT | BPF_X:
492 case BPF_JMP | BPF_JGE | BPF_X:
493 case BPF_JMP | BPF_JEQ | BPF_X:
494 ctx->seen |= SEEN_XREG;
497 case BPF_JMP | BPF_JSET | BPF_X:
498 ctx->seen |= SEEN_XREG;
499 PPC_AND_DOT(r_scratch1, r_A, r_X);
501 case BPF_JMP | BPF_JEQ | BPF_K:
502 case BPF_JMP | BPF_JGT | BPF_K:
503 case BPF_JMP | BPF_JGE | BPF_K:
507 PPC_LI32(r_scratch1, K);
508 PPC_CMPLW(r_A, r_scratch1);
511 case BPF_JMP | BPF_JSET | BPF_K:
513 /* PPC_ANDI is /only/ dot-form */
514 PPC_ANDI(r_scratch1, r_A, K);
516 PPC_LI32(r_scratch1, K);
517 PPC_AND_DOT(r_scratch1, r_A,
522 /* Sometimes branches are constructed "backward", with
523 * the false path being the branch and true path being
524 * a fallthrough to the next instruction.
526 if (filter[i].jt == 0)
527 /* Swap the sense of the branch */
528 PPC_BCC(true_cond ^ COND_CMP_TRUE,
529 addrs[i + 1 + filter[i].jf]);
531 PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]);
532 if (filter[i].jf != 0)
533 PPC_JMP(addrs[i + 1 + filter[i].jf]);
537 /* The filter contains something cruel & unusual.
538 * We don't handle it, but also there shouldn't be
539 * anything missing from our list.
541 if (printk_ratelimit())
542 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
548 /* Set end-of-body-code address for exit. */
549 addrs[i] = ctx->idx * 4;
554 void bpf_jit_compile(struct bpf_prog *fp)
556 unsigned int proglen;
557 unsigned int alloclen;
561 struct codegen_context cgctx;
568 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
573 * There are multiple assembly passes as the generated code will change
574 * size as it settles down, figuring out the max branch offsets/exit
577 * The range of standard conditional branches is +/- 32Kbytes. Since
578 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
579 * finish with 8 bytes/instruction. Not feasible, so long jumps are
580 * used, distinct from short branches.
584 * For now, both branch types assemble to 2 words (short branches padded
585 * with a NOP); this is less efficient, but assembly will always complete
586 * after exactly 3 passes:
588 * First pass: No code buffer; Program is "faux-generated" -- no code
589 * emitted but maximum size of output determined (and addrs[] filled
590 * in). Also, we note whether we use M[], whether we use skb data, etc.
591 * All generation choices assumed to be 'worst-case', e.g. branches all
592 * far (2 instructions), return path code reduction not available, etc.
594 * Second pass: Code buffer allocated with size determined previously.
595 * Prologue generated to support features we have seen used. Exit paths
596 * determined and addrs[] is filled in again, as code may be slightly
597 * smaller as a result.
599 * Third pass: Code generated 'for real', and branch destinations
600 * determined from now-accurate addrs[] map.
604 * If we optimise this, near branches will be shorter. On the
605 * first assembly pass, we should err on the side of caution and
606 * generate the biggest code. On subsequent passes, branches will be
607 * generated short or long and code size will reduce. With smaller
608 * code, more branches may fall into the short category, and code will
611 * Finally, if we see one pass generate code the same size as the
612 * previous pass we have converged and should now generate code for
613 * real. Allocating at the end will also save the memory that would
614 * otherwise be wasted by the (small) current code shrinkage.
615 * Preferably, we should do a small number of passes (e.g. 5) and if we
616 * haven't converged by then, get impatient and force code to generate
617 * as-is, even if the odd branch would be left long. The chances of a
618 * long jump are tiny with all but the most enormous of BPF filter
619 * inputs, so we should usually converge on the third pass.
625 /* Scouting faux-generate pass 0 */
626 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
627 /* We hit something illegal or unsupported. */
631 * Pretend to build prologue, given the features we've seen. This will
632 * update ctgtx.idx as it pretends to output instructions, then we can
633 * calculate total size from idx.
635 bpf_jit_build_prologue(fp, 0, &cgctx);
636 bpf_jit_build_epilogue(0, &cgctx);
638 proglen = cgctx.idx * 4;
639 alloclen = proglen + FUNCTION_DESCR_SIZE;
640 image = module_alloc(alloclen);
644 code_base = image + (FUNCTION_DESCR_SIZE/4);
646 /* Code generation passes 1-2 */
647 for (pass = 1; pass < 3; pass++) {
648 /* Now build the prologue, body code & epilogue for real. */
650 bpf_jit_build_prologue(fp, code_base, &cgctx);
651 bpf_jit_build_body(fp, code_base, &cgctx, addrs);
652 bpf_jit_build_epilogue(code_base, &cgctx);
654 if (bpf_jit_enable > 1)
655 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
656 proglen - (cgctx.idx * 4), cgctx.seen);
659 if (bpf_jit_enable > 1)
660 /* Note that we output the base address of the code_base
661 * rather than image, since opcodes are in code_base.
663 bpf_jit_dump(flen, proglen, pass, code_base);
666 bpf_flush_icache(code_base, code_base + (proglen/4));
668 /* Function descriptor nastiness: Address + TOC */
669 ((u64 *)image)[0] = (u64)code_base;
670 ((u64 *)image)[1] = local_paca->kernel_toc;
672 fp->bpf_func = (void *)image;
680 void bpf_jit_free(struct bpf_prog *fp)
683 module_memfree(fp->bpf_func);
685 bpf_prog_unlock_free(fp);