Loading...
1/* bpf_jit_comp.c: BPF JIT compiler for PPC64
2 *
3 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
4 *
5 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com)
6 *
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
10 * of the License.
11 */
12#include <linux/moduleloader.h>
13#include <asm/cacheflush.h>
14#include <linux/netdevice.h>
15#include <linux/filter.h>
16#include "bpf_jit.h"
17
18#ifndef __BIG_ENDIAN
19/* There are endianness assumptions herein. */
20#error "Little-endian PPC not supported in BPF compiler"
21#endif
22
23int bpf_jit_enable __read_mostly;
24
25
26static inline void bpf_flush_icache(void *start, void *end)
27{
28 smp_wmb();
29 flush_icache_range((unsigned long)start, (unsigned long)end);
30}
31
32static void bpf_jit_build_prologue(struct sk_filter *fp, u32 *image,
33 struct codegen_context *ctx)
34{
35 int i;
36 const struct sock_filter *filter = fp->insns;
37
38 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
39 /* Make stackframe */
40 if (ctx->seen & SEEN_DATAREF) {
41 /* If we call any helpers (for loads), save LR */
42 EMIT(PPC_INST_MFLR | __PPC_RT(0));
43 PPC_STD(0, 1, 16);
44
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)));
48 }
49 if (ctx->seen & SEEN_MEM) {
50 /*
51 * Conditionally save regs r15-r31 as some will be used
52 * for M[] data.
53 */
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)));
57 }
58 }
59 EMIT(PPC_INST_STDU | __PPC_RS(1) | __PPC_RA(1) |
60 (-BPF_PPC_STACKFRAME & 0xfffc));
61 }
62
63 if (ctx->seen & SEEN_DATAREF) {
64 /*
65 * If this filter needs to access skb data,
66 * prepare r_D and r_HL:
67 * r_HL = skb->len - skb->data_len
68 * r_D = skb->data
69 */
70 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
71 data_len));
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));
75 }
76
77 if (ctx->seen & SEEN_XREG) {
78 /*
79 * TODO: Could also detect whether first instr. sets X and
80 * avoid this (as below, with A).
81 */
82 PPC_LI(r_X, 0);
83 }
84
85 switch (filter[0].code) {
86 case BPF_S_RET_K:
87 case BPF_S_LD_W_LEN:
88 case BPF_S_ANC_PROTOCOL:
89 case BPF_S_ANC_IFINDEX:
90 case BPF_S_ANC_MARK:
91 case BPF_S_ANC_RXHASH:
92 case BPF_S_ANC_CPU:
93 case BPF_S_ANC_QUEUE:
94 case BPF_S_LD_W_ABS:
95 case BPF_S_LD_H_ABS:
96 case BPF_S_LD_B_ABS:
97 /* first instruction sets A register (or is RET 'constant') */
98 break;
99 default:
100 /* make sure we dont leak kernel information to user */
101 PPC_LI(r_A, 0);
102 }
103}
104
105static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
106{
107 int i;
108
109 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
110 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
111 if (ctx->seen & SEEN_DATAREF) {
112 PPC_LD(0, 1, 16);
113 PPC_MTLR(0);
114 PPC_LD(r_D, 1, -(8*(32-r_D)));
115 PPC_LD(r_HL, 1, -(8*(32-r_HL)));
116 }
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)));
122 }
123 }
124 }
125 /* The RETs have left a return value in R3. */
126
127 PPC_BLR();
128}
129
130/* Assemble the body code between the prologue & epilogue. */
131static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
132 struct codegen_context *ctx,
133 unsigned int *addrs)
134{
135 const struct sock_filter *filter = fp->insns;
136 int flen = fp->len;
137 u8 *func;
138 unsigned int true_cond;
139 int i;
140
141 /* Start of epilogue code */
142 unsigned int exit_addr = addrs[flen];
143
144 for (i = 0; i < flen; i++) {
145 unsigned int K = filter[i].k;
146
147 /*
148 * addrs[] maps a BPF bytecode address into a real offset from
149 * the start of the body code.
150 */
151 addrs[i] = ctx->idx * 4;
152
153 switch (filter[i].code) {
154 /*** ALU ops ***/
155 case BPF_S_ALU_ADD_X: /* A += X; */
156 ctx->seen |= SEEN_XREG;
157 PPC_ADD(r_A, r_A, r_X);
158 break;
159 case BPF_S_ALU_ADD_K: /* A += K; */
160 if (!K)
161 break;
162 PPC_ADDI(r_A, r_A, IMM_L(K));
163 if (K >= 32768)
164 PPC_ADDIS(r_A, r_A, IMM_HA(K));
165 break;
166 case BPF_S_ALU_SUB_X: /* A -= X; */
167 ctx->seen |= SEEN_XREG;
168 PPC_SUB(r_A, r_A, r_X);
169 break;
170 case BPF_S_ALU_SUB_K: /* A -= K */
171 if (!K)
172 break;
173 PPC_ADDI(r_A, r_A, IMM_L(-K));
174 if (K >= 32768)
175 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
176 break;
177 case BPF_S_ALU_MUL_X: /* A *= X; */
178 ctx->seen |= SEEN_XREG;
179 PPC_MUL(r_A, r_A, r_X);
180 break;
181 case BPF_S_ALU_MUL_K: /* A *= K */
182 if (K < 32768)
183 PPC_MULI(r_A, r_A, K);
184 else {
185 PPC_LI32(r_scratch1, K);
186 PPC_MUL(r_A, r_A, r_scratch1);
187 }
188 break;
189 case BPF_S_ALU_DIV_X: /* A /= X; */
190 ctx->seen |= SEEN_XREG;
191 PPC_CMPWI(r_X, 0);
192 if (ctx->pc_ret0 != -1) {
193 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
194 } else {
195 /*
196 * Exit, returning 0; first pass hits here
197 * (longer worst-case code size).
198 */
199 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
200 PPC_LI(r_ret, 0);
201 PPC_JMP(exit_addr);
202 }
203 PPC_DIVWU(r_A, r_A, r_X);
204 break;
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);
209 break;
210 case BPF_S_ALU_AND_X:
211 ctx->seen |= SEEN_XREG;
212 PPC_AND(r_A, r_A, r_X);
213 break;
214 case BPF_S_ALU_AND_K:
215 if (!IMM_H(K))
216 PPC_ANDI(r_A, r_A, K);
217 else {
218 PPC_LI32(r_scratch1, K);
219 PPC_AND(r_A, r_A, r_scratch1);
220 }
221 break;
222 case BPF_S_ALU_OR_X:
223 ctx->seen |= SEEN_XREG;
224 PPC_OR(r_A, r_A, r_X);
225 break;
226 case BPF_S_ALU_OR_K:
227 if (IMM_L(K))
228 PPC_ORI(r_A, r_A, IMM_L(K));
229 if (K >= 65536)
230 PPC_ORIS(r_A, r_A, IMM_H(K));
231 break;
232 case BPF_S_ALU_LSH_X: /* A <<= X; */
233 ctx->seen |= SEEN_XREG;
234 PPC_SLW(r_A, r_A, r_X);
235 break;
236 case BPF_S_ALU_LSH_K:
237 if (K == 0)
238 break;
239 else
240 PPC_SLWI(r_A, r_A, K);
241 break;
242 case BPF_S_ALU_RSH_X: /* A >>= X; */
243 ctx->seen |= SEEN_XREG;
244 PPC_SRW(r_A, r_A, r_X);
245 break;
246 case BPF_S_ALU_RSH_K: /* A >>= K; */
247 if (K == 0)
248 break;
249 else
250 PPC_SRWI(r_A, r_A, K);
251 break;
252 case BPF_S_ALU_NEG:
253 PPC_NEG(r_A, r_A);
254 break;
255 case BPF_S_RET_K:
256 PPC_LI32(r_ret, K);
257 if (!K) {
258 if (ctx->pc_ret0 == -1)
259 ctx->pc_ret0 = i;
260 }
261 /*
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.
267 */
268 if (i != flen - 1) {
269 /*
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.
275 */
276 if (ctx->seen)
277 PPC_JMP(exit_addr);
278 else
279 PPC_BLR();
280 }
281 break;
282 case BPF_S_RET_A:
283 PPC_MR(r_ret, r_A);
284 if (i != flen - 1) {
285 if (ctx->seen)
286 PPC_JMP(exit_addr);
287 else
288 PPC_BLR();
289 }
290 break;
291 case BPF_S_MISC_TAX: /* X = A */
292 PPC_MR(r_X, r_A);
293 break;
294 case BPF_S_MISC_TXA: /* A = X */
295 ctx->seen |= SEEN_XREG;
296 PPC_MR(r_A, r_X);
297 break;
298
299 /*** Constant loads/M[] access ***/
300 case BPF_S_LD_IMM: /* A = K */
301 PPC_LI32(r_A, K);
302 break;
303 case BPF_S_LDX_IMM: /* X = K */
304 PPC_LI32(r_X, K);
305 break;
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));
309 break;
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));
313 break;
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));
317 break;
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));
321 break;
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));
325 break;
326 case BPF_S_LDX_W_LEN: /* X = skb->len; */
327 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
328 break;
329
330 /*** Ancillary info loads ***/
331
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
335 * planted in future.
336 */
337 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
338 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
339 protocol) != 2);
340 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
341 protocol));
342 /* ntohs is a NOP with BE loads. */
343 break;
344 case BPF_S_ANC_IFINDEX:
345 PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
346 dev));
347 PPC_CMPDI(r_scratch1, 0);
348 if (ctx->pc_ret0 != -1) {
349 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
350 } else {
351 /* Exit, returning 0; first pass hits here. */
352 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
353 PPC_LI(r_ret, 0);
354 PPC_JMP(exit_addr);
355 }
356 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
357 ifindex) != 4);
358 PPC_LWZ_OFFS(r_A, r_scratch1,
359 offsetof(struct net_device, ifindex));
360 break;
361 case BPF_S_ANC_MARK:
362 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
363 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
364 mark));
365 break;
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,
369 rxhash));
370 break;
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,
375 queue_mapping));
376 break;
377 case BPF_S_ANC_CPU:
378#ifdef CONFIG_SMP
379 /*
380 * PACA ptr is r13:
381 * raw_smp_processor_id() = local_paca->paca_index
382 */
383 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct,
384 paca_index) != 2);
385 PPC_LHZ_OFFS(r_A, 13,
386 offsetof(struct paca_struct, paca_index));
387#else
388 PPC_LI(r_A, 0);
389#endif
390 break;
391
392 /*** Absolute loads from packet header/data ***/
393 case BPF_S_LD_W_ABS:
394 func = sk_load_word;
395 goto common_load;
396 case BPF_S_LD_H_ABS:
397 func = sk_load_half;
398 goto common_load;
399 case BPF_S_LD_B_ABS:
400 func = sk_load_byte;
401 common_load:
402 /*
403 * Load from [K]. Reference with the (negative)
404 * SKF_NET_OFF/SKF_LL_OFF offsets is unsupported.
405 */
406 ctx->seen |= SEEN_DATAREF;
407 if ((int)K < 0)
408 return -ENOTSUPP;
409 PPC_LI64(r_scratch1, func);
410 PPC_MTLR(r_scratch1);
411 PPC_LI32(r_addr, K);
412 PPC_BLRL();
413 /*
414 * Helper returns 'lt' condition on error, and an
415 * appropriate return value in r3
416 */
417 PPC_BCC(COND_LT, exit_addr);
418 break;
419
420 /*** Indirect loads from packet header/data ***/
421 case BPF_S_LD_W_IND:
422 func = sk_load_word;
423 goto common_load_ind;
424 case BPF_S_LD_H_IND:
425 func = sk_load_half;
426 goto common_load_ind;
427 case BPF_S_LD_B_IND:
428 func = sk_load_byte;
429 common_load_ind:
430 /*
431 * Load from [X + K]. Negative offsets are tested for
432 * in the helper functions, and result in a 'ret 0'.
433 */
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));
438 if (K >= 32768)
439 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
440 PPC_BLRL();
441 /* If error, cr0.LT set */
442 PPC_BCC(COND_LT, exit_addr);
443 break;
444
445 case BPF_S_LDX_B_MSH:
446 /*
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.
451 */
452 func = sk_load_byte_msh;
453 goto common_load;
454 break;
455
456 /*** Jump and branches ***/
457 case BPF_S_JMP_JA:
458 if (K != 0)
459 PPC_JMP(addrs[i + 1 + K]);
460 break;
461
462 case BPF_S_JMP_JGT_K:
463 case BPF_S_JMP_JGT_X:
464 true_cond = COND_GT;
465 goto cond_branch;
466 case BPF_S_JMP_JGE_K:
467 case BPF_S_JMP_JGE_X:
468 true_cond = COND_GE;
469 goto cond_branch;
470 case BPF_S_JMP_JEQ_K:
471 case BPF_S_JMP_JEQ_X:
472 true_cond = COND_EQ;
473 goto cond_branch;
474 case BPF_S_JMP_JSET_K:
475 case BPF_S_JMP_JSET_X:
476 true_cond = COND_NE;
477 /* Fall through */
478 cond_branch:
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]);
483 break;
484 }
485
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;
491 PPC_CMPLW(r_A, r_X);
492 break;
493 case BPF_S_JMP_JSET_X:
494 ctx->seen |= SEEN_XREG;
495 PPC_AND_DOT(r_scratch1, r_A, r_X);
496 break;
497 case BPF_S_JMP_JEQ_K:
498 case BPF_S_JMP_JGT_K:
499 case BPF_S_JMP_JGE_K:
500 if (K < 32768)
501 PPC_CMPLWI(r_A, K);
502 else {
503 PPC_LI32(r_scratch1, K);
504 PPC_CMPLW(r_A, r_scratch1);
505 }
506 break;
507 case BPF_S_JMP_JSET_K:
508 if (K < 32768)
509 /* PPC_ANDI is /only/ dot-form */
510 PPC_ANDI(r_scratch1, r_A, K);
511 else {
512 PPC_LI32(r_scratch1, K);
513 PPC_AND_DOT(r_scratch1, r_A,
514 r_scratch1);
515 }
516 break;
517 }
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.
521 */
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]);
526 else {
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]);
530 }
531 break;
532 default:
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.
536 */
537 if (printk_ratelimit())
538 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
539 filter[i].code, i);
540 return -ENOTSUPP;
541 }
542
543 }
544 /* Set end-of-body-code address for exit. */
545 addrs[i] = ctx->idx * 4;
546
547 return 0;
548}
549
550void bpf_jit_compile(struct sk_filter *fp)
551{
552 unsigned int proglen;
553 unsigned int alloclen;
554 u32 *image = NULL;
555 u32 *code_base;
556 unsigned int *addrs;
557 struct codegen_context cgctx;
558 int pass;
559 int flen = fp->len;
560
561 if (!bpf_jit_enable)
562 return;
563
564 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
565 if (addrs == NULL)
566 return;
567
568 /*
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
571 * paths required.
572 *
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.
577 *
578 * Current:
579 *
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:
583 *
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.
589 *
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.
594 *
595 * Third pass: Code generated 'for real', and branch destinations
596 * determined from now-accurate addrs[] map.
597 *
598 * Ideal:
599 *
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
605 * reduce more.
606 *
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.
616 */
617
618 cgctx.idx = 0;
619 cgctx.seen = 0;
620 cgctx.pc_ret0 = -1;
621 /* Scouting faux-generate pass 0 */
622 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
623 /* We hit something illegal or unsupported. */
624 goto out;
625
626 /*
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.
630 */
631 bpf_jit_build_prologue(fp, 0, &cgctx);
632 bpf_jit_build_epilogue(0, &cgctx);
633
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)));
638 if (!image)
639 goto out;
640
641 code_base = image + (FUNCTION_DESCR_SIZE/4);
642
643 /* Code generation passes 1-2 */
644 for (pass = 1; pass < 3; pass++) {
645 /* Now build the prologue, body code & epilogue for real. */
646 cgctx.idx = 0;
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);
650
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);
654 }
655
656 if (bpf_jit_enable > 1)
657 pr_info("flen=%d proglen=%u pass=%d image=%p\n",
658 flen, proglen, pass, image);
659
660 if (image) {
661 if (bpf_jit_enable > 1)
662 print_hex_dump(KERN_ERR, "JIT code: ",
663 DUMP_PREFIX_ADDRESS,
664 16, 1, code_base,
665 proglen, false);
666
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;
672 }
673out:
674 kfree(addrs);
675 return;
676}
677
678static void jit_free_defer(struct work_struct *arg)
679{
680 module_free(NULL, arg);
681}
682
683/* run from softirq, we must use a work_struct to call
684 * module_free() from process context
685 */
686void bpf_jit_free(struct sk_filter *fp)
687{
688 if (fp->bpf_func != sk_run_filter) {
689 struct work_struct *work = (struct work_struct *)fp->bpf_func;
690
691 INIT_WORK(work, jit_free_defer);
692 schedule_work(work);
693 }
694}
1/* bpf_jit_comp.c: BPF JIT compiler for PPC64
2 *
3 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
4 *
5 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com)
6 *
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
10 * of the License.
11 */
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>
17
18#include "bpf_jit.h"
19
20int bpf_jit_enable __read_mostly;
21
22static inline void bpf_flush_icache(void *start, void *end)
23{
24 smp_wmb();
25 flush_icache_range((unsigned long)start, (unsigned long)end);
26}
27
28static void bpf_jit_build_prologue(struct sk_filter *fp, u32 *image,
29 struct codegen_context *ctx)
30{
31 int i;
32 const struct sock_filter *filter = fp->insns;
33
34 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
35 /* Make stackframe */
36 if (ctx->seen & SEEN_DATAREF) {
37 /* If we call any helpers (for loads), save LR */
38 EMIT(PPC_INST_MFLR | __PPC_RT(R0));
39 PPC_STD(0, 1, 16);
40
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)));
44 }
45 if (ctx->seen & SEEN_MEM) {
46 /*
47 * Conditionally save regs r15-r31 as some will be used
48 * for M[] data.
49 */
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)));
53 }
54 }
55 EMIT(PPC_INST_STDU | __PPC_RS(R1) | __PPC_RA(R1) |
56 (-BPF_PPC_STACKFRAME & 0xfffc));
57 }
58
59 if (ctx->seen & SEEN_DATAREF) {
60 /*
61 * If this filter needs to access skb data,
62 * prepare r_D and r_HL:
63 * r_HL = skb->len - skb->data_len
64 * r_D = skb->data
65 */
66 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
67 data_len));
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));
71 }
72
73 if (ctx->seen & SEEN_XREG) {
74 /*
75 * TODO: Could also detect whether first instr. sets X and
76 * avoid this (as below, with A).
77 */
78 PPC_LI(r_X, 0);
79 }
80
81 switch (filter[0].code) {
82 case BPF_S_RET_K:
83 case BPF_S_LD_W_LEN:
84 case BPF_S_ANC_PROTOCOL:
85 case BPF_S_ANC_IFINDEX:
86 case BPF_S_ANC_MARK:
87 case BPF_S_ANC_RXHASH:
88 case BPF_S_ANC_VLAN_TAG:
89 case BPF_S_ANC_VLAN_TAG_PRESENT:
90 case BPF_S_ANC_CPU:
91 case BPF_S_ANC_QUEUE:
92 case BPF_S_LD_W_ABS:
93 case BPF_S_LD_H_ABS:
94 case BPF_S_LD_B_ABS:
95 /* first instruction sets A register (or is RET 'constant') */
96 break;
97 default:
98 /* make sure we dont leak kernel information to user */
99 PPC_LI(r_A, 0);
100 }
101}
102
103static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
104{
105 int i;
106
107 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
108 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
109 if (ctx->seen & SEEN_DATAREF) {
110 PPC_LD(0, 1, 16);
111 PPC_MTLR(0);
112 PPC_LD(r_D, 1, -(8*(32-r_D)));
113 PPC_LD(r_HL, 1, -(8*(32-r_HL)));
114 }
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)));
120 }
121 }
122 }
123 /* The RETs have left a return value in R3. */
124
125 PPC_BLR();
126}
127
128#define CHOOSE_LOAD_FUNC(K, func) \
129 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
130
131/* Assemble the body code between the prologue & epilogue. */
132static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
133 struct codegen_context *ctx,
134 unsigned int *addrs)
135{
136 const struct sock_filter *filter = fp->insns;
137 int flen = fp->len;
138 u8 *func;
139 unsigned int true_cond;
140 int i;
141
142 /* Start of epilogue code */
143 unsigned int exit_addr = addrs[flen];
144
145 for (i = 0; i < flen; i++) {
146 unsigned int K = filter[i].k;
147
148 /*
149 * addrs[] maps a BPF bytecode address into a real offset from
150 * the start of the body code.
151 */
152 addrs[i] = ctx->idx * 4;
153
154 switch (filter[i].code) {
155 /*** ALU ops ***/
156 case BPF_S_ALU_ADD_X: /* A += X; */
157 ctx->seen |= SEEN_XREG;
158 PPC_ADD(r_A, r_A, r_X);
159 break;
160 case BPF_S_ALU_ADD_K: /* A += K; */
161 if (!K)
162 break;
163 PPC_ADDI(r_A, r_A, IMM_L(K));
164 if (K >= 32768)
165 PPC_ADDIS(r_A, r_A, IMM_HA(K));
166 break;
167 case BPF_S_ALU_SUB_X: /* A -= X; */
168 ctx->seen |= SEEN_XREG;
169 PPC_SUB(r_A, r_A, r_X);
170 break;
171 case BPF_S_ALU_SUB_K: /* A -= K */
172 if (!K)
173 break;
174 PPC_ADDI(r_A, r_A, IMM_L(-K));
175 if (K >= 32768)
176 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
177 break;
178 case BPF_S_ALU_MUL_X: /* A *= X; */
179 ctx->seen |= SEEN_XREG;
180 PPC_MUL(r_A, r_A, r_X);
181 break;
182 case BPF_S_ALU_MUL_K: /* A *= K */
183 if (K < 32768)
184 PPC_MULI(r_A, r_A, K);
185 else {
186 PPC_LI32(r_scratch1, K);
187 PPC_MUL(r_A, r_A, r_scratch1);
188 }
189 break;
190 case BPF_S_ALU_MOD_X: /* A %= X; */
191 ctx->seen |= SEEN_XREG;
192 PPC_CMPWI(r_X, 0);
193 if (ctx->pc_ret0 != -1) {
194 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
195 } else {
196 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
197 PPC_LI(r_ret, 0);
198 PPC_JMP(exit_addr);
199 }
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);
203 break;
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);
209 break;
210 case BPF_S_ALU_DIV_X: /* A /= X; */
211 ctx->seen |= SEEN_XREG;
212 PPC_CMPWI(r_X, 0);
213 if (ctx->pc_ret0 != -1) {
214 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
215 } else {
216 /*
217 * Exit, returning 0; first pass hits here
218 * (longer worst-case code size).
219 */
220 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
221 PPC_LI(r_ret, 0);
222 PPC_JMP(exit_addr);
223 }
224 PPC_DIVWU(r_A, r_A, r_X);
225 break;
226 case BPF_S_ALU_DIV_K: /* A /= K */
227 if (K == 1)
228 break;
229 PPC_LI32(r_scratch1, K);
230 PPC_DIVWU(r_A, r_A, r_scratch1);
231 break;
232 case BPF_S_ALU_AND_X:
233 ctx->seen |= SEEN_XREG;
234 PPC_AND(r_A, r_A, r_X);
235 break;
236 case BPF_S_ALU_AND_K:
237 if (!IMM_H(K))
238 PPC_ANDI(r_A, r_A, K);
239 else {
240 PPC_LI32(r_scratch1, K);
241 PPC_AND(r_A, r_A, r_scratch1);
242 }
243 break;
244 case BPF_S_ALU_OR_X:
245 ctx->seen |= SEEN_XREG;
246 PPC_OR(r_A, r_A, r_X);
247 break;
248 case BPF_S_ALU_OR_K:
249 if (IMM_L(K))
250 PPC_ORI(r_A, r_A, IMM_L(K));
251 if (K >= 65536)
252 PPC_ORIS(r_A, r_A, IMM_H(K));
253 break;
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);
258 break;
259 case BPF_S_ALU_XOR_K: /* A ^= K */
260 if (IMM_L(K))
261 PPC_XORI(r_A, r_A, IMM_L(K));
262 if (K >= 65536)
263 PPC_XORIS(r_A, r_A, IMM_H(K));
264 break;
265 case BPF_S_ALU_LSH_X: /* A <<= X; */
266 ctx->seen |= SEEN_XREG;
267 PPC_SLW(r_A, r_A, r_X);
268 break;
269 case BPF_S_ALU_LSH_K:
270 if (K == 0)
271 break;
272 else
273 PPC_SLWI(r_A, r_A, K);
274 break;
275 case BPF_S_ALU_RSH_X: /* A >>= X; */
276 ctx->seen |= SEEN_XREG;
277 PPC_SRW(r_A, r_A, r_X);
278 break;
279 case BPF_S_ALU_RSH_K: /* A >>= K; */
280 if (K == 0)
281 break;
282 else
283 PPC_SRWI(r_A, r_A, K);
284 break;
285 case BPF_S_ALU_NEG:
286 PPC_NEG(r_A, r_A);
287 break;
288 case BPF_S_RET_K:
289 PPC_LI32(r_ret, K);
290 if (!K) {
291 if (ctx->pc_ret0 == -1)
292 ctx->pc_ret0 = i;
293 }
294 /*
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.
300 */
301 if (i != flen - 1) {
302 /*
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.
308 */
309 if (ctx->seen)
310 PPC_JMP(exit_addr);
311 else
312 PPC_BLR();
313 }
314 break;
315 case BPF_S_RET_A:
316 PPC_MR(r_ret, r_A);
317 if (i != flen - 1) {
318 if (ctx->seen)
319 PPC_JMP(exit_addr);
320 else
321 PPC_BLR();
322 }
323 break;
324 case BPF_S_MISC_TAX: /* X = A */
325 PPC_MR(r_X, r_A);
326 break;
327 case BPF_S_MISC_TXA: /* A = X */
328 ctx->seen |= SEEN_XREG;
329 PPC_MR(r_A, r_X);
330 break;
331
332 /*** Constant loads/M[] access ***/
333 case BPF_S_LD_IMM: /* A = K */
334 PPC_LI32(r_A, K);
335 break;
336 case BPF_S_LDX_IMM: /* X = K */
337 PPC_LI32(r_X, K);
338 break;
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));
342 break;
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));
346 break;
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));
350 break;
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));
354 break;
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));
358 break;
359 case BPF_S_LDX_W_LEN: /* X = skb->len; */
360 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
361 break;
362
363 /*** Ancillary info loads ***/
364 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
365 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
366 protocol) != 2);
367 PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff,
368 protocol));
369 break;
370 case BPF_S_ANC_IFINDEX:
371 PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
372 dev));
373 PPC_CMPDI(r_scratch1, 0);
374 if (ctx->pc_ret0 != -1) {
375 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
376 } else {
377 /* Exit, returning 0; first pass hits here. */
378 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
379 PPC_LI(r_ret, 0);
380 PPC_JMP(exit_addr);
381 }
382 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
383 ifindex) != 4);
384 PPC_LWZ_OFFS(r_A, r_scratch1,
385 offsetof(struct net_device, ifindex));
386 break;
387 case BPF_S_ANC_MARK:
388 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
389 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
390 mark));
391 break;
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,
395 hash));
396 break;
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,
401 vlan_tci));
402 if (filter[i].code == BPF_S_ANC_VLAN_TAG)
403 PPC_ANDI(r_A, r_A, VLAN_VID_MASK);
404 else
405 PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT);
406 break;
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,
411 queue_mapping));
412 break;
413 case BPF_S_ANC_CPU:
414#ifdef CONFIG_SMP
415 /*
416 * PACA ptr is r13:
417 * raw_smp_processor_id() = local_paca->paca_index
418 */
419 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct,
420 paca_index) != 2);
421 PPC_LHZ_OFFS(r_A, 13,
422 offsetof(struct paca_struct, paca_index));
423#else
424 PPC_LI(r_A, 0);
425#endif
426 break;
427
428 /*** Absolute loads from packet header/data ***/
429 case BPF_S_LD_W_ABS:
430 func = CHOOSE_LOAD_FUNC(K, sk_load_word);
431 goto common_load;
432 case BPF_S_LD_H_ABS:
433 func = CHOOSE_LOAD_FUNC(K, sk_load_half);
434 goto common_load;
435 case BPF_S_LD_B_ABS:
436 func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
437 common_load:
438 /* Load from [K]. */
439 ctx->seen |= SEEN_DATAREF;
440 PPC_LI64(r_scratch1, func);
441 PPC_MTLR(r_scratch1);
442 PPC_LI32(r_addr, K);
443 PPC_BLRL();
444 /*
445 * Helper returns 'lt' condition on error, and an
446 * appropriate return value in r3
447 */
448 PPC_BCC(COND_LT, exit_addr);
449 break;
450
451 /*** Indirect loads from packet header/data ***/
452 case BPF_S_LD_W_IND:
453 func = sk_load_word;
454 goto common_load_ind;
455 case BPF_S_LD_H_IND:
456 func = sk_load_half;
457 goto common_load_ind;
458 case BPF_S_LD_B_IND:
459 func = sk_load_byte;
460 common_load_ind:
461 /*
462 * Load from [X + K]. Negative offsets are tested for
463 * in the helper functions.
464 */
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));
469 if (K >= 32768)
470 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
471 PPC_BLRL();
472 /* If error, cr0.LT set */
473 PPC_BCC(COND_LT, exit_addr);
474 break;
475
476 case BPF_S_LDX_B_MSH:
477 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
478 goto common_load;
479 break;
480
481 /*** Jump and branches ***/
482 case BPF_S_JMP_JA:
483 if (K != 0)
484 PPC_JMP(addrs[i + 1 + K]);
485 break;
486
487 case BPF_S_JMP_JGT_K:
488 case BPF_S_JMP_JGT_X:
489 true_cond = COND_GT;
490 goto cond_branch;
491 case BPF_S_JMP_JGE_K:
492 case BPF_S_JMP_JGE_X:
493 true_cond = COND_GE;
494 goto cond_branch;
495 case BPF_S_JMP_JEQ_K:
496 case BPF_S_JMP_JEQ_X:
497 true_cond = COND_EQ;
498 goto cond_branch;
499 case BPF_S_JMP_JSET_K:
500 case BPF_S_JMP_JSET_X:
501 true_cond = COND_NE;
502 /* Fall through */
503 cond_branch:
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]);
508 break;
509 }
510
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;
516 PPC_CMPLW(r_A, r_X);
517 break;
518 case BPF_S_JMP_JSET_X:
519 ctx->seen |= SEEN_XREG;
520 PPC_AND_DOT(r_scratch1, r_A, r_X);
521 break;
522 case BPF_S_JMP_JEQ_K:
523 case BPF_S_JMP_JGT_K:
524 case BPF_S_JMP_JGE_K:
525 if (K < 32768)
526 PPC_CMPLWI(r_A, K);
527 else {
528 PPC_LI32(r_scratch1, K);
529 PPC_CMPLW(r_A, r_scratch1);
530 }
531 break;
532 case BPF_S_JMP_JSET_K:
533 if (K < 32768)
534 /* PPC_ANDI is /only/ dot-form */
535 PPC_ANDI(r_scratch1, r_A, K);
536 else {
537 PPC_LI32(r_scratch1, K);
538 PPC_AND_DOT(r_scratch1, r_A,
539 r_scratch1);
540 }
541 break;
542 }
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.
546 */
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]);
551 else {
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]);
555 }
556 break;
557 default:
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.
561 */
562 if (printk_ratelimit())
563 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
564 filter[i].code, i);
565 return -ENOTSUPP;
566 }
567
568 }
569 /* Set end-of-body-code address for exit. */
570 addrs[i] = ctx->idx * 4;
571
572 return 0;
573}
574
575void bpf_jit_compile(struct sk_filter *fp)
576{
577 unsigned int proglen;
578 unsigned int alloclen;
579 u32 *image = NULL;
580 u32 *code_base;
581 unsigned int *addrs;
582 struct codegen_context cgctx;
583 int pass;
584 int flen = fp->len;
585
586 if (!bpf_jit_enable)
587 return;
588
589 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
590 if (addrs == NULL)
591 return;
592
593 /*
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
596 * paths required.
597 *
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.
602 *
603 * Current:
604 *
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:
608 *
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.
614 *
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.
619 *
620 * Third pass: Code generated 'for real', and branch destinations
621 * determined from now-accurate addrs[] map.
622 *
623 * Ideal:
624 *
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
630 * reduce more.
631 *
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.
641 */
642
643 cgctx.idx = 0;
644 cgctx.seen = 0;
645 cgctx.pc_ret0 = -1;
646 /* Scouting faux-generate pass 0 */
647 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
648 /* We hit something illegal or unsupported. */
649 goto out;
650
651 /*
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.
655 */
656 bpf_jit_build_prologue(fp, 0, &cgctx);
657 bpf_jit_build_epilogue(0, &cgctx);
658
659 proglen = cgctx.idx * 4;
660 alloclen = proglen + FUNCTION_DESCR_SIZE;
661 image = module_alloc(alloclen);
662 if (!image)
663 goto out;
664
665 code_base = image + (FUNCTION_DESCR_SIZE/4);
666
667 /* Code generation passes 1-2 */
668 for (pass = 1; pass < 3; pass++) {
669 /* Now build the prologue, body code & epilogue for real. */
670 cgctx.idx = 0;
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);
674
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);
678 }
679
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.
683 */
684 bpf_jit_dump(flen, proglen, pass, code_base);
685
686 if (image) {
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;
692 fp->jited = 1;
693 }
694out:
695 kfree(addrs);
696 return;
697}
698
699void bpf_jit_free(struct sk_filter *fp)
700{
701 if (fp->jited)
702 module_free(NULL, fp->bpf_func);
703 kfree(fp);
704}