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>
19 /* There are endianness assumptions herein. */
20 #error "Little-endian PPC not supported in BPF compiler"
23 int bpf_jit_enable __read_mostly;
26 static inline void bpf_flush_icache(void *start, void *end)
29 flush_icache_range((unsigned long)start, (unsigned long)end);
32 static void bpf_jit_build_prologue(struct sk_filter *fp, u32 *image,
33 struct codegen_context *ctx)
36 const struct sock_filter *filter = fp->insns;
38 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
40 if (ctx->seen & SEEN_DATAREF) {
41 /* If we call any helpers (for loads), save LR */
42 EMIT(PPC_INST_MFLR | __PPC_RT(0));
45 /* Back up non-volatile regs. */
46 PPC_STD(r_D, 1, -(8*(32-r_D)));
47 PPC_STD(r_HL, 1, -(8*(32-r_HL)));
49 if (ctx->seen & SEEN_MEM) {
51 * Conditionally save regs r15-r31 as some will be used
54 for (i = r_M; i < (r_M+16); i++) {
55 if (ctx->seen & (1 << (i-r_M)))
56 PPC_STD(i, 1, -(8*(32-i)));
59 EMIT(PPC_INST_STDU | __PPC_RS(1) | __PPC_RA(1) |
60 (-BPF_PPC_STACKFRAME & 0xfffc));
63 if (ctx->seen & SEEN_DATAREF) {
65 * If this filter needs to access skb data,
66 * prepare r_D and r_HL:
67 * r_HL = skb->len - skb->data_len
70 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
72 PPC_LWZ_OFFS(r_HL, r_skb, offsetof(struct sk_buff, len));
73 PPC_SUB(r_HL, r_HL, r_scratch1);
74 PPC_LD_OFFS(r_D, r_skb, offsetof(struct sk_buff, data));
77 if (ctx->seen & SEEN_XREG) {
79 * TODO: Could also detect whether first instr. sets X and
80 * avoid this (as below, with A).
85 switch (filter[0].code) {
88 case BPF_S_ANC_PROTOCOL:
89 case BPF_S_ANC_IFINDEX:
91 case BPF_S_ANC_RXHASH:
97 /* first instruction sets A register (or is RET 'constant') */
100 /* make sure we dont leak kernel information to user */
105 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
109 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
110 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
111 if (ctx->seen & SEEN_DATAREF) {
114 PPC_LD(r_D, 1, -(8*(32-r_D)));
115 PPC_LD(r_HL, 1, -(8*(32-r_HL)));
117 if (ctx->seen & SEEN_MEM) {
118 /* Restore any saved non-vol registers */
119 for (i = r_M; i < (r_M+16); i++) {
120 if (ctx->seen & (1 << (i-r_M)))
121 PPC_LD(i, 1, -(8*(32-i)));
125 /* The RETs have left a return value in R3. */
130 /* Assemble the body code between the prologue & epilogue. */
131 static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
132 struct codegen_context *ctx,
135 const struct sock_filter *filter = fp->insns;
138 unsigned int true_cond;
141 /* Start of epilogue code */
142 unsigned int exit_addr = addrs[flen];
144 for (i = 0; i < flen; i++) {
145 unsigned int K = filter[i].k;
148 * addrs[] maps a BPF bytecode address into a real offset from
149 * the start of the body code.
151 addrs[i] = ctx->idx * 4;
153 switch (filter[i].code) {
155 case BPF_S_ALU_ADD_X: /* A += X; */
156 ctx->seen |= SEEN_XREG;
157 PPC_ADD(r_A, r_A, r_X);
159 case BPF_S_ALU_ADD_K: /* A += K; */
162 PPC_ADDI(r_A, r_A, IMM_L(K));
164 PPC_ADDIS(r_A, r_A, IMM_HA(K));
166 case BPF_S_ALU_SUB_X: /* A -= X; */
167 ctx->seen |= SEEN_XREG;
168 PPC_SUB(r_A, r_A, r_X);
170 case BPF_S_ALU_SUB_K: /* A -= K */
173 PPC_ADDI(r_A, r_A, IMM_L(-K));
175 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
177 case BPF_S_ALU_MUL_X: /* A *= X; */
178 ctx->seen |= SEEN_XREG;
179 PPC_MUL(r_A, r_A, r_X);
181 case BPF_S_ALU_MUL_K: /* A *= K */
183 PPC_MULI(r_A, r_A, K);
185 PPC_LI32(r_scratch1, K);
186 PPC_MUL(r_A, r_A, r_scratch1);
189 case BPF_S_ALU_DIV_X: /* A /= X; */
190 ctx->seen |= SEEN_XREG;
192 if (ctx->pc_ret0 != -1) {
193 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
196 * Exit, returning 0; first pass hits here
197 * (longer worst-case code size).
199 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
203 PPC_DIVWU(r_A, r_A, r_X);
205 case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */
206 PPC_LI32(r_scratch1, K);
207 /* Top 32 bits of 64bit result -> A */
208 PPC_MULHWU(r_A, r_A, r_scratch1);
210 case BPF_S_ALU_AND_X:
211 ctx->seen |= SEEN_XREG;
212 PPC_AND(r_A, r_A, r_X);
214 case BPF_S_ALU_AND_K:
216 PPC_ANDI(r_A, r_A, K);
218 PPC_LI32(r_scratch1, K);
219 PPC_AND(r_A, r_A, r_scratch1);
223 ctx->seen |= SEEN_XREG;
224 PPC_OR(r_A, r_A, r_X);
228 PPC_ORI(r_A, r_A, IMM_L(K));
230 PPC_ORIS(r_A, r_A, IMM_H(K));
232 case BPF_S_ALU_LSH_X: /* A <<= X; */
233 ctx->seen |= SEEN_XREG;
234 PPC_SLW(r_A, r_A, r_X);
236 case BPF_S_ALU_LSH_K:
240 PPC_SLWI(r_A, r_A, K);
242 case BPF_S_ALU_RSH_X: /* A >>= X; */
243 ctx->seen |= SEEN_XREG;
244 PPC_SRW(r_A, r_A, r_X);
246 case BPF_S_ALU_RSH_K: /* A >>= K; */
250 PPC_SRWI(r_A, r_A, K);
258 if (ctx->pc_ret0 == -1)
262 * If this isn't the very last instruction, branch to
263 * the epilogue if we've stuff to clean up. Otherwise,
264 * if there's nothing to tidy, just return. If we /are/
265 * the last instruction, we're about to fall through to
266 * the epilogue to return.
270 * Note: 'seen' is properly valid only on pass
271 * #2. Both parts of this conditional are the
272 * same instruction size though, meaning the
273 * first pass will still correctly determine the
274 * code size/addresses.
291 case BPF_S_MISC_TAX: /* X = A */
294 case BPF_S_MISC_TXA: /* A = X */
295 ctx->seen |= SEEN_XREG;
299 /*** Constant loads/M[] access ***/
300 case BPF_S_LD_IMM: /* A = K */
303 case BPF_S_LDX_IMM: /* X = K */
306 case BPF_S_LD_MEM: /* A = mem[K] */
307 PPC_MR(r_A, r_M + (K & 0xf));
308 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
310 case BPF_S_LDX_MEM: /* X = mem[K] */
311 PPC_MR(r_X, r_M + (K & 0xf));
312 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
314 case BPF_S_ST: /* mem[K] = A */
315 PPC_MR(r_M + (K & 0xf), r_A);
316 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
318 case BPF_S_STX: /* mem[K] = X */
319 PPC_MR(r_M + (K & 0xf), r_X);
320 ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf));
322 case BPF_S_LD_W_LEN: /* A = skb->len; */
323 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
324 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len));
326 case BPF_S_LDX_W_LEN: /* X = skb->len; */
327 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
330 /*** Ancillary info loads ***/
332 /* None of the BPF_S_ANC* codes appear to be passed by
333 * sk_chk_filter(). The interpreter and the x86 BPF
334 * compiler implement them so we do too -- they may be
337 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
338 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
340 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
342 /* ntohs is a NOP with BE loads. */
344 case BPF_S_ANC_IFINDEX:
345 PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
347 PPC_CMPDI(r_scratch1, 0);
348 if (ctx->pc_ret0 != -1) {
349 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
351 /* Exit, returning 0; first pass hits here. */
352 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
356 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
358 PPC_LWZ_OFFS(r_A, r_scratch1,
359 offsetof(struct net_device, ifindex));
362 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
363 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
366 case BPF_S_ANC_RXHASH:
367 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
368 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
371 case BPF_S_ANC_QUEUE:
372 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
373 queue_mapping) != 2);
374 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
381 * raw_smp_processor_id() = local_paca->paca_index
383 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct,
385 PPC_LHZ_OFFS(r_A, 13,
386 offsetof(struct paca_struct, paca_index));
392 /*** Absolute loads from packet header/data ***/
403 * Load from [K]. Reference with the (negative)
404 * SKF_NET_OFF/SKF_LL_OFF offsets is unsupported.
406 ctx->seen |= SEEN_DATAREF;
409 PPC_LI64(r_scratch1, func);
410 PPC_MTLR(r_scratch1);
414 * Helper returns 'lt' condition on error, and an
415 * appropriate return value in r3
417 PPC_BCC(COND_LT, exit_addr);
420 /*** Indirect loads from packet header/data ***/
423 goto common_load_ind;
426 goto common_load_ind;
431 * Load from [X + K]. Negative offsets are tested for
432 * in the helper functions, and result in a 'ret 0'.
434 ctx->seen |= SEEN_DATAREF | SEEN_XREG;
435 PPC_LI64(r_scratch1, func);
436 PPC_MTLR(r_scratch1);
437 PPC_ADDI(r_addr, r_X, IMM_L(K));
439 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
441 /* If error, cr0.LT set */
442 PPC_BCC(COND_LT, exit_addr);
445 case BPF_S_LDX_B_MSH:
447 * x86 version drops packet (RET 0) when K<0, whereas
448 * interpreter does allow K<0 (__load_pointer, special
449 * ancillary data). common_load returns ENOTSUPP if K<0,
450 * so we fall back to interpreter & filter works.
452 func = sk_load_byte_msh;
456 /*** Jump and branches ***/
459 PPC_JMP(addrs[i + 1 + K]);
462 case BPF_S_JMP_JGT_K:
463 case BPF_S_JMP_JGT_X:
466 case BPF_S_JMP_JGE_K:
467 case BPF_S_JMP_JGE_X:
470 case BPF_S_JMP_JEQ_K:
471 case BPF_S_JMP_JEQ_X:
474 case BPF_S_JMP_JSET_K:
475 case BPF_S_JMP_JSET_X:
479 /* same targets, can avoid doing the test :) */
480 if (filter[i].jt == filter[i].jf) {
481 if (filter[i].jt > 0)
482 PPC_JMP(addrs[i + 1 + filter[i].jt]);
486 switch (filter[i].code) {
487 case BPF_S_JMP_JGT_X:
488 case BPF_S_JMP_JGE_X:
489 case BPF_S_JMP_JEQ_X:
490 ctx->seen |= SEEN_XREG;
493 case BPF_S_JMP_JSET_X:
494 ctx->seen |= SEEN_XREG;
495 PPC_AND_DOT(r_scratch1, r_A, r_X);
497 case BPF_S_JMP_JEQ_K:
498 case BPF_S_JMP_JGT_K:
499 case BPF_S_JMP_JGE_K:
503 PPC_LI32(r_scratch1, K);
504 PPC_CMPLW(r_A, r_scratch1);
507 case BPF_S_JMP_JSET_K:
509 /* PPC_ANDI is /only/ dot-form */
510 PPC_ANDI(r_scratch1, r_A, K);
512 PPC_LI32(r_scratch1, K);
513 PPC_AND_DOT(r_scratch1, r_A,
518 /* Sometimes branches are constructed "backward", with
519 * the false path being the branch and true path being
520 * a fallthrough to the next instruction.
522 if (filter[i].jt == 0)
523 /* Swap the sense of the branch */
524 PPC_BCC(true_cond ^ COND_CMP_TRUE,
525 addrs[i + 1 + filter[i].jf]);
527 PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]);
528 if (filter[i].jf != 0)
529 PPC_JMP(addrs[i + 1 + filter[i].jf]);
533 /* The filter contains something cruel & unusual.
534 * We don't handle it, but also there shouldn't be
535 * anything missing from our list.
537 if (printk_ratelimit())
538 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
544 /* Set end-of-body-code address for exit. */
545 addrs[i] = ctx->idx * 4;
550 void bpf_jit_compile(struct sk_filter *fp)
552 unsigned int proglen;
553 unsigned int alloclen;
557 struct codegen_context cgctx;
564 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
569 * There are multiple assembly passes as the generated code will change
570 * size as it settles down, figuring out the max branch offsets/exit
573 * The range of standard conditional branches is +/- 32Kbytes. Since
574 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
575 * finish with 8 bytes/instruction. Not feasible, so long jumps are
576 * used, distinct from short branches.
580 * For now, both branch types assemble to 2 words (short branches padded
581 * with a NOP); this is less efficient, but assembly will always complete
582 * after exactly 3 passes:
584 * First pass: No code buffer; Program is "faux-generated" -- no code
585 * emitted but maximum size of output determined (and addrs[] filled
586 * in). Also, we note whether we use M[], whether we use skb data, etc.
587 * All generation choices assumed to be 'worst-case', e.g. branches all
588 * far (2 instructions), return path code reduction not available, etc.
590 * Second pass: Code buffer allocated with size determined previously.
591 * Prologue generated to support features we have seen used. Exit paths
592 * determined and addrs[] is filled in again, as code may be slightly
593 * smaller as a result.
595 * Third pass: Code generated 'for real', and branch destinations
596 * determined from now-accurate addrs[] map.
600 * If we optimise this, near branches will be shorter. On the
601 * first assembly pass, we should err on the side of caution and
602 * generate the biggest code. On subsequent passes, branches will be
603 * generated short or long and code size will reduce. With smaller
604 * code, more branches may fall into the short category, and code will
607 * Finally, if we see one pass generate code the same size as the
608 * previous pass we have converged and should now generate code for
609 * real. Allocating at the end will also save the memory that would
610 * otherwise be wasted by the (small) current code shrinkage.
611 * Preferably, we should do a small number of passes (e.g. 5) and if we
612 * haven't converged by then, get impatient and force code to generate
613 * as-is, even if the odd branch would be left long. The chances of a
614 * long jump are tiny with all but the most enormous of BPF filter
615 * inputs, so we should usually converge on the third pass.
621 /* Scouting faux-generate pass 0 */
622 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
623 /* We hit something illegal or unsupported. */
627 * Pretend to build prologue, given the features we've seen. This will
628 * update ctgtx.idx as it pretends to output instructions, then we can
629 * calculate total size from idx.
631 bpf_jit_build_prologue(fp, 0, &cgctx);
632 bpf_jit_build_epilogue(0, &cgctx);
634 proglen = cgctx.idx * 4;
635 alloclen = proglen + FUNCTION_DESCR_SIZE;
636 image = module_alloc(max_t(unsigned int, alloclen,
637 sizeof(struct work_struct)));
641 code_base = image + (FUNCTION_DESCR_SIZE/4);
643 /* Code generation passes 1-2 */
644 for (pass = 1; pass < 3; pass++) {
645 /* Now build the prologue, body code & epilogue for real. */
647 bpf_jit_build_prologue(fp, code_base, &cgctx);
648 bpf_jit_build_body(fp, code_base, &cgctx, addrs);
649 bpf_jit_build_epilogue(code_base, &cgctx);
651 if (bpf_jit_enable > 1)
652 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
653 proglen - (cgctx.idx * 4), cgctx.seen);
656 if (bpf_jit_enable > 1)
657 pr_info("flen=%d proglen=%u pass=%d image=%p\n",
658 flen, proglen, pass, image);
661 if (bpf_jit_enable > 1)
662 print_hex_dump(KERN_ERR, "JIT code: ",
667 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;
671 fp->bpf_func = (void *)image;
678 static void jit_free_defer(struct work_struct *arg)
680 module_free(NULL, arg);
683 /* run from softirq, we must use a work_struct to call
684 * module_free() from process context
686 void bpf_jit_free(struct sk_filter *fp)
688 if (fp->bpf_func != sk_run_filter) {
689 struct work_struct *work = (struct work_struct *)fp->bpf_func;
691 INIT_WORK(work, jit_free_defer);