Commit | Line | Data |
---|---|---|
0ca87f05 ME |
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> | |
5082dfb7 DB |
16 | #include <linux/if_vlan.h> |
17 | ||
0ca87f05 ME |
18 | #include "bpf_jit.h" |
19 | ||
0ca87f05 ME |
20 | int bpf_jit_enable __read_mostly; |
21 | ||
0ca87f05 ME |
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 */ | |
c75df6f9 | 38 | EMIT(PPC_INST_MFLR | __PPC_RT(R0)); |
0ca87f05 ME |
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 | } | |
c75df6f9 | 55 | EMIT(PPC_INST_STDU | __PPC_RS(R1) | __PPC_RA(R1) | |
0ca87f05 ME |
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: | |
5082dfb7 DB |
88 | case BPF_S_ANC_VLAN_TAG: |
89 | case BPF_S_ANC_VLAN_TAG_PRESENT: | |
0ca87f05 ME |
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 | ||
05be1824 JS |
128 | #define CHOOSE_LOAD_FUNC(K, func) \ |
129 | ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset) | |
130 | ||
0ca87f05 ME |
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_DIV_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 | /* | |
197 | * Exit, returning 0; first pass hits here | |
198 | * (longer worst-case code size). | |
199 | */ | |
200 | PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12); | |
201 | PPC_LI(r_ret, 0); | |
202 | PPC_JMP(exit_addr); | |
203 | } | |
204 | PPC_DIVWU(r_A, r_A, r_X); | |
205 | break; | |
206 | case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */ | |
207 | PPC_LI32(r_scratch1, K); | |
208 | /* Top 32 bits of 64bit result -> A */ | |
209 | PPC_MULHWU(r_A, r_A, r_scratch1); | |
210 | break; | |
211 | case BPF_S_ALU_AND_X: | |
212 | ctx->seen |= SEEN_XREG; | |
213 | PPC_AND(r_A, r_A, r_X); | |
214 | break; | |
215 | case BPF_S_ALU_AND_K: | |
216 | if (!IMM_H(K)) | |
217 | PPC_ANDI(r_A, r_A, K); | |
218 | else { | |
219 | PPC_LI32(r_scratch1, K); | |
220 | PPC_AND(r_A, r_A, r_scratch1); | |
221 | } | |
222 | break; | |
223 | case BPF_S_ALU_OR_X: | |
224 | ctx->seen |= SEEN_XREG; | |
225 | PPC_OR(r_A, r_A, r_X); | |
226 | break; | |
227 | case BPF_S_ALU_OR_K: | |
228 | if (IMM_L(K)) | |
229 | PPC_ORI(r_A, r_A, IMM_L(K)); | |
230 | if (K >= 65536) | |
231 | PPC_ORIS(r_A, r_A, IMM_H(K)); | |
232 | break; | |
02871903 DB |
233 | case BPF_S_ANC_ALU_XOR_X: |
234 | case BPF_S_ALU_XOR_X: /* A ^= X */ | |
235 | ctx->seen |= SEEN_XREG; | |
236 | PPC_XOR(r_A, r_A, r_X); | |
237 | break; | |
238 | case BPF_S_ALU_XOR_K: /* A ^= K */ | |
239 | if (IMM_L(K)) | |
240 | PPC_XORI(r_A, r_A, IMM_L(K)); | |
241 | if (K >= 65536) | |
242 | PPC_XORIS(r_A, r_A, IMM_H(K)); | |
243 | break; | |
0ca87f05 ME |
244 | case BPF_S_ALU_LSH_X: /* A <<= X; */ |
245 | ctx->seen |= SEEN_XREG; | |
246 | PPC_SLW(r_A, r_A, r_X); | |
247 | break; | |
248 | case BPF_S_ALU_LSH_K: | |
249 | if (K == 0) | |
250 | break; | |
251 | else | |
252 | PPC_SLWI(r_A, r_A, K); | |
253 | break; | |
254 | case BPF_S_ALU_RSH_X: /* A >>= X; */ | |
255 | ctx->seen |= SEEN_XREG; | |
256 | PPC_SRW(r_A, r_A, r_X); | |
257 | break; | |
258 | case BPF_S_ALU_RSH_K: /* A >>= K; */ | |
259 | if (K == 0) | |
260 | break; | |
261 | else | |
262 | PPC_SRWI(r_A, r_A, K); | |
263 | break; | |
264 | case BPF_S_ALU_NEG: | |
265 | PPC_NEG(r_A, r_A); | |
266 | break; | |
267 | case BPF_S_RET_K: | |
268 | PPC_LI32(r_ret, K); | |
269 | if (!K) { | |
270 | if (ctx->pc_ret0 == -1) | |
271 | ctx->pc_ret0 = i; | |
272 | } | |
273 | /* | |
274 | * If this isn't the very last instruction, branch to | |
275 | * the epilogue if we've stuff to clean up. Otherwise, | |
276 | * if there's nothing to tidy, just return. If we /are/ | |
277 | * the last instruction, we're about to fall through to | |
278 | * the epilogue to return. | |
279 | */ | |
280 | if (i != flen - 1) { | |
281 | /* | |
282 | * Note: 'seen' is properly valid only on pass | |
283 | * #2. Both parts of this conditional are the | |
284 | * same instruction size though, meaning the | |
285 | * first pass will still correctly determine the | |
286 | * code size/addresses. | |
287 | */ | |
288 | if (ctx->seen) | |
289 | PPC_JMP(exit_addr); | |
290 | else | |
291 | PPC_BLR(); | |
292 | } | |
293 | break; | |
294 | case BPF_S_RET_A: | |
295 | PPC_MR(r_ret, r_A); | |
296 | if (i != flen - 1) { | |
297 | if (ctx->seen) | |
298 | PPC_JMP(exit_addr); | |
299 | else | |
300 | PPC_BLR(); | |
301 | } | |
302 | break; | |
303 | case BPF_S_MISC_TAX: /* X = A */ | |
304 | PPC_MR(r_X, r_A); | |
305 | break; | |
306 | case BPF_S_MISC_TXA: /* A = X */ | |
307 | ctx->seen |= SEEN_XREG; | |
308 | PPC_MR(r_A, r_X); | |
309 | break; | |
310 | ||
311 | /*** Constant loads/M[] access ***/ | |
312 | case BPF_S_LD_IMM: /* A = K */ | |
313 | PPC_LI32(r_A, K); | |
314 | break; | |
315 | case BPF_S_LDX_IMM: /* X = K */ | |
316 | PPC_LI32(r_X, K); | |
317 | break; | |
318 | case BPF_S_LD_MEM: /* A = mem[K] */ | |
319 | PPC_MR(r_A, r_M + (K & 0xf)); | |
320 | ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); | |
321 | break; | |
322 | case BPF_S_LDX_MEM: /* X = mem[K] */ | |
323 | PPC_MR(r_X, r_M + (K & 0xf)); | |
324 | ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); | |
325 | break; | |
326 | case BPF_S_ST: /* mem[K] = A */ | |
327 | PPC_MR(r_M + (K & 0xf), r_A); | |
328 | ctx->seen |= SEEN_MEM | (1<<(K & 0xf)); | |
329 | break; | |
330 | case BPF_S_STX: /* mem[K] = X */ | |
331 | PPC_MR(r_M + (K & 0xf), r_X); | |
332 | ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf)); | |
333 | break; | |
334 | case BPF_S_LD_W_LEN: /* A = skb->len; */ | |
335 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); | |
336 | PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len)); | |
337 | break; | |
338 | case BPF_S_LDX_W_LEN: /* X = skb->len; */ | |
339 | PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len)); | |
340 | break; | |
341 | ||
342 | /*** Ancillary info loads ***/ | |
0ca87f05 ME |
343 | case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */ |
344 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, | |
345 | protocol) != 2); | |
9c662cad PB |
346 | PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff, |
347 | protocol)); | |
0ca87f05 ME |
348 | break; |
349 | case BPF_S_ANC_IFINDEX: | |
350 | PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff, | |
351 | dev)); | |
352 | PPC_CMPDI(r_scratch1, 0); | |
353 | if (ctx->pc_ret0 != -1) { | |
354 | PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]); | |
355 | } else { | |
356 | /* Exit, returning 0; first pass hits here. */ | |
357 | PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12); | |
358 | PPC_LI(r_ret, 0); | |
359 | PPC_JMP(exit_addr); | |
360 | } | |
361 | BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, | |
362 | ifindex) != 4); | |
363 | PPC_LWZ_OFFS(r_A, r_scratch1, | |
364 | offsetof(struct net_device, ifindex)); | |
365 | break; | |
366 | case BPF_S_ANC_MARK: | |
367 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); | |
368 | PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, | |
369 | mark)); | |
370 | break; | |
371 | case BPF_S_ANC_RXHASH: | |
372 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4); | |
373 | PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, | |
374 | rxhash)); | |
375 | break; | |
5082dfb7 DB |
376 | case BPF_S_ANC_VLAN_TAG: |
377 | case BPF_S_ANC_VLAN_TAG_PRESENT: | |
378 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); | |
379 | PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, | |
380 | vlan_tci)); | |
381 | if (filter[i].code == BPF_S_ANC_VLAN_TAG) | |
382 | PPC_ANDI(r_A, r_A, VLAN_VID_MASK); | |
383 | else | |
384 | PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT); | |
385 | break; | |
0ca87f05 ME |
386 | case BPF_S_ANC_QUEUE: |
387 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, | |
388 | queue_mapping) != 2); | |
389 | PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, | |
390 | queue_mapping)); | |
391 | break; | |
392 | case BPF_S_ANC_CPU: | |
393 | #ifdef CONFIG_SMP | |
394 | /* | |
395 | * PACA ptr is r13: | |
396 | * raw_smp_processor_id() = local_paca->paca_index | |
397 | */ | |
398 | BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct, | |
399 | paca_index) != 2); | |
400 | PPC_LHZ_OFFS(r_A, 13, | |
401 | offsetof(struct paca_struct, paca_index)); | |
402 | #else | |
403 | PPC_LI(r_A, 0); | |
404 | #endif | |
405 | break; | |
406 | ||
407 | /*** Absolute loads from packet header/data ***/ | |
408 | case BPF_S_LD_W_ABS: | |
05be1824 | 409 | func = CHOOSE_LOAD_FUNC(K, sk_load_word); |
0ca87f05 ME |
410 | goto common_load; |
411 | case BPF_S_LD_H_ABS: | |
05be1824 | 412 | func = CHOOSE_LOAD_FUNC(K, sk_load_half); |
0ca87f05 ME |
413 | goto common_load; |
414 | case BPF_S_LD_B_ABS: | |
05be1824 | 415 | func = CHOOSE_LOAD_FUNC(K, sk_load_byte); |
0ca87f05 | 416 | common_load: |
05be1824 | 417 | /* Load from [K]. */ |
0ca87f05 | 418 | ctx->seen |= SEEN_DATAREF; |
0ca87f05 ME |
419 | PPC_LI64(r_scratch1, func); |
420 | PPC_MTLR(r_scratch1); | |
421 | PPC_LI32(r_addr, K); | |
422 | PPC_BLRL(); | |
423 | /* | |
424 | * Helper returns 'lt' condition on error, and an | |
425 | * appropriate return value in r3 | |
426 | */ | |
427 | PPC_BCC(COND_LT, exit_addr); | |
428 | break; | |
429 | ||
430 | /*** Indirect loads from packet header/data ***/ | |
431 | case BPF_S_LD_W_IND: | |
432 | func = sk_load_word; | |
433 | goto common_load_ind; | |
434 | case BPF_S_LD_H_IND: | |
435 | func = sk_load_half; | |
436 | goto common_load_ind; | |
437 | case BPF_S_LD_B_IND: | |
438 | func = sk_load_byte; | |
439 | common_load_ind: | |
440 | /* | |
441 | * Load from [X + K]. Negative offsets are tested for | |
05be1824 | 442 | * in the helper functions. |
0ca87f05 ME |
443 | */ |
444 | ctx->seen |= SEEN_DATAREF | SEEN_XREG; | |
445 | PPC_LI64(r_scratch1, func); | |
446 | PPC_MTLR(r_scratch1); | |
447 | PPC_ADDI(r_addr, r_X, IMM_L(K)); | |
448 | if (K >= 32768) | |
449 | PPC_ADDIS(r_addr, r_addr, IMM_HA(K)); | |
450 | PPC_BLRL(); | |
451 | /* If error, cr0.LT set */ | |
452 | PPC_BCC(COND_LT, exit_addr); | |
453 | break; | |
454 | ||
455 | case BPF_S_LDX_B_MSH: | |
05be1824 | 456 | func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh); |
0ca87f05 ME |
457 | goto common_load; |
458 | break; | |
459 | ||
460 | /*** Jump and branches ***/ | |
461 | case BPF_S_JMP_JA: | |
462 | if (K != 0) | |
463 | PPC_JMP(addrs[i + 1 + K]); | |
464 | break; | |
465 | ||
466 | case BPF_S_JMP_JGT_K: | |
467 | case BPF_S_JMP_JGT_X: | |
468 | true_cond = COND_GT; | |
469 | goto cond_branch; | |
470 | case BPF_S_JMP_JGE_K: | |
471 | case BPF_S_JMP_JGE_X: | |
472 | true_cond = COND_GE; | |
473 | goto cond_branch; | |
474 | case BPF_S_JMP_JEQ_K: | |
475 | case BPF_S_JMP_JEQ_X: | |
476 | true_cond = COND_EQ; | |
477 | goto cond_branch; | |
478 | case BPF_S_JMP_JSET_K: | |
479 | case BPF_S_JMP_JSET_X: | |
480 | true_cond = COND_NE; | |
481 | /* Fall through */ | |
482 | cond_branch: | |
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]); | |
487 | break; | |
488 | } | |
489 | ||
490 | switch (filter[i].code) { | |
491 | case BPF_S_JMP_JGT_X: | |
492 | case BPF_S_JMP_JGE_X: | |
493 | case BPF_S_JMP_JEQ_X: | |
494 | ctx->seen |= SEEN_XREG; | |
495 | PPC_CMPLW(r_A, r_X); | |
496 | break; | |
497 | case BPF_S_JMP_JSET_X: | |
498 | ctx->seen |= SEEN_XREG; | |
499 | PPC_AND_DOT(r_scratch1, r_A, r_X); | |
500 | break; | |
501 | case BPF_S_JMP_JEQ_K: | |
502 | case BPF_S_JMP_JGT_K: | |
503 | case BPF_S_JMP_JGE_K: | |
504 | if (K < 32768) | |
505 | PPC_CMPLWI(r_A, K); | |
506 | else { | |
507 | PPC_LI32(r_scratch1, K); | |
508 | PPC_CMPLW(r_A, r_scratch1); | |
509 | } | |
510 | break; | |
511 | case BPF_S_JMP_JSET_K: | |
512 | if (K < 32768) | |
513 | /* PPC_ANDI is /only/ dot-form */ | |
514 | PPC_ANDI(r_scratch1, r_A, K); | |
515 | else { | |
516 | PPC_LI32(r_scratch1, K); | |
517 | PPC_AND_DOT(r_scratch1, r_A, | |
518 | r_scratch1); | |
519 | } | |
520 | break; | |
521 | } | |
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. | |
525 | */ | |
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]); | |
530 | else { | |
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]); | |
534 | } | |
535 | break; | |
536 | default: | |
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. | |
540 | */ | |
541 | if (printk_ratelimit()) | |
542 | pr_err("BPF filter opcode %04x (@%d) unsupported\n", | |
543 | filter[i].code, i); | |
544 | return -ENOTSUPP; | |
545 | } | |
546 | ||
547 | } | |
548 | /* Set end-of-body-code address for exit. */ | |
549 | addrs[i] = ctx->idx * 4; | |
550 | ||
551 | return 0; | |
552 | } | |
553 | ||
554 | void bpf_jit_compile(struct sk_filter *fp) | |
555 | { | |
556 | unsigned int proglen; | |
557 | unsigned int alloclen; | |
558 | u32 *image = NULL; | |
559 | u32 *code_base; | |
560 | unsigned int *addrs; | |
561 | struct codegen_context cgctx; | |
562 | int pass; | |
563 | int flen = fp->len; | |
564 | ||
565 | if (!bpf_jit_enable) | |
566 | return; | |
567 | ||
568 | addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL); | |
569 | if (addrs == NULL) | |
570 | return; | |
571 | ||
572 | /* | |
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 | |
575 | * paths required. | |
576 | * | |
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. | |
581 | * | |
582 | * Current: | |
583 | * | |
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: | |
587 | * | |
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. | |
593 | * | |
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. | |
598 | * | |
599 | * Third pass: Code generated 'for real', and branch destinations | |
600 | * determined from now-accurate addrs[] map. | |
601 | * | |
602 | * Ideal: | |
603 | * | |
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 | |
609 | * reduce more. | |
610 | * | |
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. | |
620 | */ | |
621 | ||
622 | cgctx.idx = 0; | |
623 | cgctx.seen = 0; | |
624 | cgctx.pc_ret0 = -1; | |
625 | /* Scouting faux-generate pass 0 */ | |
626 | if (bpf_jit_build_body(fp, 0, &cgctx, addrs)) | |
627 | /* We hit something illegal or unsupported. */ | |
628 | goto out; | |
629 | ||
630 | /* | |
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. | |
634 | */ | |
635 | bpf_jit_build_prologue(fp, 0, &cgctx); | |
636 | bpf_jit_build_epilogue(0, &cgctx); | |
637 | ||
638 | proglen = cgctx.idx * 4; | |
639 | alloclen = proglen + FUNCTION_DESCR_SIZE; | |
ed900ffb | 640 | image = module_alloc(alloclen); |
0ca87f05 ME |
641 | if (!image) |
642 | goto out; | |
643 | ||
644 | code_base = image + (FUNCTION_DESCR_SIZE/4); | |
645 | ||
646 | /* Code generation passes 1-2 */ | |
647 | for (pass = 1; pass < 3; pass++) { | |
648 | /* Now build the prologue, body code & epilogue for real. */ | |
649 | cgctx.idx = 0; | |
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); | |
653 | ||
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); | |
657 | } | |
658 | ||
659 | if (bpf_jit_enable > 1) | |
79617801 DB |
660 | /* Note that we output the base address of the code_base |
661 | * rather than image, since opcodes are in code_base. | |
662 | */ | |
663 | bpf_jit_dump(flen, proglen, pass, code_base); | |
0ca87f05 ME |
664 | |
665 | if (image) { | |
0ca87f05 ME |
666 | bpf_flush_icache(code_base, code_base + (proglen/4)); |
667 | /* Function descriptor nastiness: Address + TOC */ | |
668 | ((u64 *)image)[0] = (u64)code_base; | |
669 | ((u64 *)image)[1] = local_paca->kernel_toc; | |
670 | fp->bpf_func = (void *)image; | |
671 | } | |
672 | out: | |
673 | kfree(addrs); | |
674 | return; | |
675 | } | |
676 | ||
0ca87f05 ME |
677 | void bpf_jit_free(struct sk_filter *fp) |
678 | { | |
ed900ffb DB |
679 | if (fp->bpf_func != sk_run_filter) |
680 | module_free(NULL, fp->bpf_func); | |
0ca87f05 | 681 | } |