xref: /openbmc/linux/arch/powerpc/net/bpf_jit_comp.c (revision 12eb4683)
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 
20 int bpf_jit_enable __read_mostly;
21 
22 static 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 
28 static 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 
103 static 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. */
132 static 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 = reciprocal_divide(A, K); */
227 			PPC_LI32(r_scratch1, K);
228 			/* Top 32 bits of 64bit result -> A */
229 			PPC_MULHWU(r_A, r_A, r_scratch1);
230 			break;
231 		case BPF_S_ALU_AND_X:
232 			ctx->seen |= SEEN_XREG;
233 			PPC_AND(r_A, r_A, r_X);
234 			break;
235 		case BPF_S_ALU_AND_K:
236 			if (!IMM_H(K))
237 				PPC_ANDI(r_A, r_A, K);
238 			else {
239 				PPC_LI32(r_scratch1, K);
240 				PPC_AND(r_A, r_A, r_scratch1);
241 			}
242 			break;
243 		case BPF_S_ALU_OR_X:
244 			ctx->seen |= SEEN_XREG;
245 			PPC_OR(r_A, r_A, r_X);
246 			break;
247 		case BPF_S_ALU_OR_K:
248 			if (IMM_L(K))
249 				PPC_ORI(r_A, r_A, IMM_L(K));
250 			if (K >= 65536)
251 				PPC_ORIS(r_A, r_A, IMM_H(K));
252 			break;
253 		case BPF_S_ANC_ALU_XOR_X:
254 		case BPF_S_ALU_XOR_X: /* A ^= X */
255 			ctx->seen |= SEEN_XREG;
256 			PPC_XOR(r_A, r_A, r_X);
257 			break;
258 		case BPF_S_ALU_XOR_K: /* A ^= K */
259 			if (IMM_L(K))
260 				PPC_XORI(r_A, r_A, IMM_L(K));
261 			if (K >= 65536)
262 				PPC_XORIS(r_A, r_A, IMM_H(K));
263 			break;
264 		case BPF_S_ALU_LSH_X: /* A <<= X; */
265 			ctx->seen |= SEEN_XREG;
266 			PPC_SLW(r_A, r_A, r_X);
267 			break;
268 		case BPF_S_ALU_LSH_K:
269 			if (K == 0)
270 				break;
271 			else
272 				PPC_SLWI(r_A, r_A, K);
273 			break;
274 		case BPF_S_ALU_RSH_X: /* A >>= X; */
275 			ctx->seen |= SEEN_XREG;
276 			PPC_SRW(r_A, r_A, r_X);
277 			break;
278 		case BPF_S_ALU_RSH_K: /* A >>= K; */
279 			if (K == 0)
280 				break;
281 			else
282 				PPC_SRWI(r_A, r_A, K);
283 			break;
284 		case BPF_S_ALU_NEG:
285 			PPC_NEG(r_A, r_A);
286 			break;
287 		case BPF_S_RET_K:
288 			PPC_LI32(r_ret, K);
289 			if (!K) {
290 				if (ctx->pc_ret0 == -1)
291 					ctx->pc_ret0 = i;
292 			}
293 			/*
294 			 * If this isn't the very last instruction, branch to
295 			 * the epilogue if we've stuff to clean up.  Otherwise,
296 			 * if there's nothing to tidy, just return.  If we /are/
297 			 * the last instruction, we're about to fall through to
298 			 * the epilogue to return.
299 			 */
300 			if (i != flen - 1) {
301 				/*
302 				 * Note: 'seen' is properly valid only on pass
303 				 * #2.	Both parts of this conditional are the
304 				 * same instruction size though, meaning the
305 				 * first pass will still correctly determine the
306 				 * code size/addresses.
307 				 */
308 				if (ctx->seen)
309 					PPC_JMP(exit_addr);
310 				else
311 					PPC_BLR();
312 			}
313 			break;
314 		case BPF_S_RET_A:
315 			PPC_MR(r_ret, r_A);
316 			if (i != flen - 1) {
317 				if (ctx->seen)
318 					PPC_JMP(exit_addr);
319 				else
320 					PPC_BLR();
321 			}
322 			break;
323 		case BPF_S_MISC_TAX: /* X = A */
324 			PPC_MR(r_X, r_A);
325 			break;
326 		case BPF_S_MISC_TXA: /* A = X */
327 			ctx->seen |= SEEN_XREG;
328 			PPC_MR(r_A, r_X);
329 			break;
330 
331 			/*** Constant loads/M[] access ***/
332 		case BPF_S_LD_IMM: /* A = K */
333 			PPC_LI32(r_A, K);
334 			break;
335 		case BPF_S_LDX_IMM: /* X = K */
336 			PPC_LI32(r_X, K);
337 			break;
338 		case BPF_S_LD_MEM: /* A = mem[K] */
339 			PPC_MR(r_A, r_M + (K & 0xf));
340 			ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
341 			break;
342 		case BPF_S_LDX_MEM: /* X = mem[K] */
343 			PPC_MR(r_X, r_M + (K & 0xf));
344 			ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
345 			break;
346 		case BPF_S_ST: /* mem[K] = A */
347 			PPC_MR(r_M + (K & 0xf), r_A);
348 			ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
349 			break;
350 		case BPF_S_STX: /* mem[K] = X */
351 			PPC_MR(r_M + (K & 0xf), r_X);
352 			ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf));
353 			break;
354 		case BPF_S_LD_W_LEN: /*	A = skb->len; */
355 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
356 			PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len));
357 			break;
358 		case BPF_S_LDX_W_LEN: /* X = skb->len; */
359 			PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
360 			break;
361 
362 			/*** Ancillary info loads ***/
363 		case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
364 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
365 						  protocol) != 2);
366 			PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff,
367 							    protocol));
368 			break;
369 		case BPF_S_ANC_IFINDEX:
370 			PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
371 								dev));
372 			PPC_CMPDI(r_scratch1, 0);
373 			if (ctx->pc_ret0 != -1) {
374 				PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
375 			} else {
376 				/* Exit, returning 0; first pass hits here. */
377 				PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
378 				PPC_LI(r_ret, 0);
379 				PPC_JMP(exit_addr);
380 			}
381 			BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
382 						  ifindex) != 4);
383 			PPC_LWZ_OFFS(r_A, r_scratch1,
384 				     offsetof(struct net_device, ifindex));
385 			break;
386 		case BPF_S_ANC_MARK:
387 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
388 			PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
389 							  mark));
390 			break;
391 		case BPF_S_ANC_RXHASH:
392 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
393 			PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
394 							  rxhash));
395 			break;
396 		case BPF_S_ANC_VLAN_TAG:
397 		case BPF_S_ANC_VLAN_TAG_PRESENT:
398 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
399 			PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
400 							  vlan_tci));
401 			if (filter[i].code == BPF_S_ANC_VLAN_TAG)
402 				PPC_ANDI(r_A, r_A, VLAN_VID_MASK);
403 			else
404 				PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT);
405 			break;
406 		case BPF_S_ANC_QUEUE:
407 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
408 						  queue_mapping) != 2);
409 			PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
410 							  queue_mapping));
411 			break;
412 		case BPF_S_ANC_CPU:
413 #ifdef CONFIG_SMP
414 			/*
415 			 * PACA ptr is r13:
416 			 * raw_smp_processor_id() = local_paca->paca_index
417 			 */
418 			BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct,
419 						  paca_index) != 2);
420 			PPC_LHZ_OFFS(r_A, 13,
421 				     offsetof(struct paca_struct, paca_index));
422 #else
423 			PPC_LI(r_A, 0);
424 #endif
425 			break;
426 
427 			/*** Absolute loads from packet header/data ***/
428 		case BPF_S_LD_W_ABS:
429 			func = CHOOSE_LOAD_FUNC(K, sk_load_word);
430 			goto common_load;
431 		case BPF_S_LD_H_ABS:
432 			func = CHOOSE_LOAD_FUNC(K, sk_load_half);
433 			goto common_load;
434 		case BPF_S_LD_B_ABS:
435 			func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
436 		common_load:
437 			/* Load from [K]. */
438 			ctx->seen |= SEEN_DATAREF;
439 			PPC_LI64(r_scratch1, func);
440 			PPC_MTLR(r_scratch1);
441 			PPC_LI32(r_addr, K);
442 			PPC_BLRL();
443 			/*
444 			 * Helper returns 'lt' condition on error, and an
445 			 * appropriate return value in r3
446 			 */
447 			PPC_BCC(COND_LT, exit_addr);
448 			break;
449 
450 			/*** Indirect loads from packet header/data ***/
451 		case BPF_S_LD_W_IND:
452 			func = sk_load_word;
453 			goto common_load_ind;
454 		case BPF_S_LD_H_IND:
455 			func = sk_load_half;
456 			goto common_load_ind;
457 		case BPF_S_LD_B_IND:
458 			func = sk_load_byte;
459 		common_load_ind:
460 			/*
461 			 * Load from [X + K].  Negative offsets are tested for
462 			 * in the helper functions.
463 			 */
464 			ctx->seen |= SEEN_DATAREF | SEEN_XREG;
465 			PPC_LI64(r_scratch1, func);
466 			PPC_MTLR(r_scratch1);
467 			PPC_ADDI(r_addr, r_X, IMM_L(K));
468 			if (K >= 32768)
469 				PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
470 			PPC_BLRL();
471 			/* If error, cr0.LT set */
472 			PPC_BCC(COND_LT, exit_addr);
473 			break;
474 
475 		case BPF_S_LDX_B_MSH:
476 			func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
477 			goto common_load;
478 			break;
479 
480 			/*** Jump and branches ***/
481 		case BPF_S_JMP_JA:
482 			if (K != 0)
483 				PPC_JMP(addrs[i + 1 + K]);
484 			break;
485 
486 		case BPF_S_JMP_JGT_K:
487 		case BPF_S_JMP_JGT_X:
488 			true_cond = COND_GT;
489 			goto cond_branch;
490 		case BPF_S_JMP_JGE_K:
491 		case BPF_S_JMP_JGE_X:
492 			true_cond = COND_GE;
493 			goto cond_branch;
494 		case BPF_S_JMP_JEQ_K:
495 		case BPF_S_JMP_JEQ_X:
496 			true_cond = COND_EQ;
497 			goto cond_branch;
498 		case BPF_S_JMP_JSET_K:
499 		case BPF_S_JMP_JSET_X:
500 			true_cond = COND_NE;
501 			/* Fall through */
502 		cond_branch:
503 			/* same targets, can avoid doing the test :) */
504 			if (filter[i].jt == filter[i].jf) {
505 				if (filter[i].jt > 0)
506 					PPC_JMP(addrs[i + 1 + filter[i].jt]);
507 				break;
508 			}
509 
510 			switch (filter[i].code) {
511 			case BPF_S_JMP_JGT_X:
512 			case BPF_S_JMP_JGE_X:
513 			case BPF_S_JMP_JEQ_X:
514 				ctx->seen |= SEEN_XREG;
515 				PPC_CMPLW(r_A, r_X);
516 				break;
517 			case BPF_S_JMP_JSET_X:
518 				ctx->seen |= SEEN_XREG;
519 				PPC_AND_DOT(r_scratch1, r_A, r_X);
520 				break;
521 			case BPF_S_JMP_JEQ_K:
522 			case BPF_S_JMP_JGT_K:
523 			case BPF_S_JMP_JGE_K:
524 				if (K < 32768)
525 					PPC_CMPLWI(r_A, K);
526 				else {
527 					PPC_LI32(r_scratch1, K);
528 					PPC_CMPLW(r_A, r_scratch1);
529 				}
530 				break;
531 			case BPF_S_JMP_JSET_K:
532 				if (K < 32768)
533 					/* PPC_ANDI is /only/ dot-form */
534 					PPC_ANDI(r_scratch1, r_A, K);
535 				else {
536 					PPC_LI32(r_scratch1, K);
537 					PPC_AND_DOT(r_scratch1, r_A,
538 						    r_scratch1);
539 				}
540 				break;
541 			}
542 			/* Sometimes branches are constructed "backward", with
543 			 * the false path being the branch and true path being
544 			 * a fallthrough to the next instruction.
545 			 */
546 			if (filter[i].jt == 0)
547 				/* Swap the sense of the branch */
548 				PPC_BCC(true_cond ^ COND_CMP_TRUE,
549 					addrs[i + 1 + filter[i].jf]);
550 			else {
551 				PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]);
552 				if (filter[i].jf != 0)
553 					PPC_JMP(addrs[i + 1 + filter[i].jf]);
554 			}
555 			break;
556 		default:
557 			/* The filter contains something cruel & unusual.
558 			 * We don't handle it, but also there shouldn't be
559 			 * anything missing from our list.
560 			 */
561 			if (printk_ratelimit())
562 				pr_err("BPF filter opcode %04x (@%d) unsupported\n",
563 				       filter[i].code, i);
564 			return -ENOTSUPP;
565 		}
566 
567 	}
568 	/* Set end-of-body-code address for exit. */
569 	addrs[i] = ctx->idx * 4;
570 
571 	return 0;
572 }
573 
574 void bpf_jit_compile(struct sk_filter *fp)
575 {
576 	unsigned int proglen;
577 	unsigned int alloclen;
578 	u32 *image = NULL;
579 	u32 *code_base;
580 	unsigned int *addrs;
581 	struct codegen_context cgctx;
582 	int pass;
583 	int flen = fp->len;
584 
585 	if (!bpf_jit_enable)
586 		return;
587 
588 	addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
589 	if (addrs == NULL)
590 		return;
591 
592 	/*
593 	 * There are multiple assembly passes as the generated code will change
594 	 * size as it settles down, figuring out the max branch offsets/exit
595 	 * paths required.
596 	 *
597 	 * The range of standard conditional branches is +/- 32Kbytes.	Since
598 	 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
599 	 * finish with 8 bytes/instruction.  Not feasible, so long jumps are
600 	 * used, distinct from short branches.
601 	 *
602 	 * Current:
603 	 *
604 	 * For now, both branch types assemble to 2 words (short branches padded
605 	 * with a NOP); this is less efficient, but assembly will always complete
606 	 * after exactly 3 passes:
607 	 *
608 	 * First pass: No code buffer; Program is "faux-generated" -- no code
609 	 * emitted but maximum size of output determined (and addrs[] filled
610 	 * in).	 Also, we note whether we use M[], whether we use skb data, etc.
611 	 * All generation choices assumed to be 'worst-case', e.g. branches all
612 	 * far (2 instructions), return path code reduction not available, etc.
613 	 *
614 	 * Second pass: Code buffer allocated with size determined previously.
615 	 * Prologue generated to support features we have seen used.  Exit paths
616 	 * determined and addrs[] is filled in again, as code may be slightly
617 	 * smaller as a result.
618 	 *
619 	 * Third pass: Code generated 'for real', and branch destinations
620 	 * determined from now-accurate addrs[] map.
621 	 *
622 	 * Ideal:
623 	 *
624 	 * If we optimise this, near branches will be shorter.	On the
625 	 * first assembly pass, we should err on the side of caution and
626 	 * generate the biggest code.  On subsequent passes, branches will be
627 	 * generated short or long and code size will reduce.  With smaller
628 	 * code, more branches may fall into the short category, and code will
629 	 * reduce more.
630 	 *
631 	 * Finally, if we see one pass generate code the same size as the
632 	 * previous pass we have converged and should now generate code for
633 	 * real.  Allocating at the end will also save the memory that would
634 	 * otherwise be wasted by the (small) current code shrinkage.
635 	 * Preferably, we should do a small number of passes (e.g. 5) and if we
636 	 * haven't converged by then, get impatient and force code to generate
637 	 * as-is, even if the odd branch would be left long.  The chances of a
638 	 * long jump are tiny with all but the most enormous of BPF filter
639 	 * inputs, so we should usually converge on the third pass.
640 	 */
641 
642 	cgctx.idx = 0;
643 	cgctx.seen = 0;
644 	cgctx.pc_ret0 = -1;
645 	/* Scouting faux-generate pass 0 */
646 	if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
647 		/* We hit something illegal or unsupported. */
648 		goto out;
649 
650 	/*
651 	 * Pretend to build prologue, given the features we've seen.  This will
652 	 * update ctgtx.idx as it pretends to output instructions, then we can
653 	 * calculate total size from idx.
654 	 */
655 	bpf_jit_build_prologue(fp, 0, &cgctx);
656 	bpf_jit_build_epilogue(0, &cgctx);
657 
658 	proglen = cgctx.idx * 4;
659 	alloclen = proglen + FUNCTION_DESCR_SIZE;
660 	image = module_alloc(alloclen);
661 	if (!image)
662 		goto out;
663 
664 	code_base = image + (FUNCTION_DESCR_SIZE/4);
665 
666 	/* Code generation passes 1-2 */
667 	for (pass = 1; pass < 3; pass++) {
668 		/* Now build the prologue, body code & epilogue for real. */
669 		cgctx.idx = 0;
670 		bpf_jit_build_prologue(fp, code_base, &cgctx);
671 		bpf_jit_build_body(fp, code_base, &cgctx, addrs);
672 		bpf_jit_build_epilogue(code_base, &cgctx);
673 
674 		if (bpf_jit_enable > 1)
675 			pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
676 				proglen - (cgctx.idx * 4), cgctx.seen);
677 	}
678 
679 	if (bpf_jit_enable > 1)
680 		/* Note that we output the base address of the code_base
681 		 * rather than image, since opcodes are in code_base.
682 		 */
683 		bpf_jit_dump(flen, proglen, pass, code_base);
684 
685 	if (image) {
686 		bpf_flush_icache(code_base, code_base + (proglen/4));
687 		/* Function descriptor nastiness: Address + TOC */
688 		((u64 *)image)[0] = (u64)code_base;
689 		((u64 *)image)[1] = local_paca->kernel_toc;
690 		fp->bpf_func = (void *)image;
691 	}
692 out:
693 	kfree(addrs);
694 	return;
695 }
696 
697 void bpf_jit_free(struct sk_filter *fp)
698 {
699 	if (fp->bpf_func != sk_run_filter)
700 		module_free(NULL, fp->bpf_func);
701 	kfree(fp);
702 }
703