?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.11-rc5
[linux/fpc-iii.git] / arch / powerpc / net / bpf_jit_comp64.c
blobaee2bb817ac68c4a8dfa7a28b521ac73c0f77b3f
1 /*
2 * bpf_jit_comp64.c: eBPF JIT compiler
3 *
4 * Copyright 2016 Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
5 * IBM Corporation
6 *
7 * Based on the powerpc classic BPF JIT compiler by Matt Evans
8 *
9 * This program is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU General Public License
11 * as published by the Free Software Foundation; version 2
12 * of the License.
13 */
14 #include <linux/moduleloader.h>
15 #include <asm/cacheflush.h>
16 #include <linux/netdevice.h>
17 #include <linux/filter.h>
18 #include <linux/if_vlan.h>
19 #include <asm/kprobes.h>
20 #include <linux/bpf.h>
21
22 #include "bpf_jit64.h"
23
24 int bpf_jit_enable __read_mostly;
25
26 static void bpf_jit_fill_ill_insns(void *area, unsigned int size)
27 {
28 int *p = area;
29
30 /* Fill whole space with trap instructions */
31 while (p < (int *)((char *)area + size))
32 *p++ = BREAKPOINT_INSTRUCTION;
33 }
34
35 static inline void bpf_flush_icache(void *start, void *end)
36 {
37 smp_wmb();
38 flush_icache_range((unsigned long)start, (unsigned long)end);
39 }
40
41 static inline bool bpf_is_seen_register(struct codegen_context *ctx, int i)
42 {
43 return (ctx->seen & (1 << (31 - b2p[i])));
44 }
45
46 static inline void bpf_set_seen_register(struct codegen_context *ctx, int i)
47 {
48 ctx->seen |= (1 << (31 - b2p[i]));
49 }
50
51 static inline bool bpf_has_stack_frame(struct codegen_context *ctx)
52 {
53 /*
54 * We only need a stack frame if:
55 * - we call other functions (kernel helpers), or
56 * - the bpf program uses its stack area
57 * The latter condition is deduced from the usage of BPF_REG_FP
58 */
59 return ctx->seen & SEEN_FUNC || bpf_is_seen_register(ctx, BPF_REG_FP);
60 }
61
62 /*
63 * When not setting up our own stackframe, the redzone usage is:
64 *
65 * [ prev sp ] <-------------
66 * [ ... ] |
67 * sp (r1) ---> [ stack pointer ] --------------
68 * [ nv gpr save area ] 8*8
69 * [ tail_call_cnt ] 8
70 * [ local_tmp_var ] 8
71 * [ unused red zone ] 208 bytes protected
72 */
73 static int bpf_jit_stack_local(struct codegen_context *ctx)
74 {
75 if (bpf_has_stack_frame(ctx))
76 return STACK_FRAME_MIN_SIZE + MAX_BPF_STACK;
77 else
78 return -(BPF_PPC_STACK_SAVE + 16);
79 }
80
81 static int bpf_jit_stack_tailcallcnt(struct codegen_context *ctx)
82 {
83 return bpf_jit_stack_local(ctx) + 8;
84 }
85
86 static int bpf_jit_stack_offsetof(struct codegen_context *ctx, int reg)
87 {
88 if (reg >= BPF_PPC_NVR_MIN && reg < 32)
89 return (bpf_has_stack_frame(ctx) ? BPF_PPC_STACKFRAME : 0)
90 - (8 * (32 - reg));
91
92 pr_err("BPF JIT is asking about unknown registers");
93 BUG();
94 }
95
96 static void bpf_jit_emit_skb_loads(u32 *image, struct codegen_context *ctx)
97 {
98 /*
99 * Load skb->len and skb->data_len
100 * r3 points to skb
101 */
102 PPC_LWZ(b2p[SKB_HLEN_REG], 3, offsetof(struct sk_buff, len));
103 PPC_LWZ(b2p[TMP_REG_1], 3, offsetof(struct sk_buff, data_len));
104 /* header_len = len - data_len */
105 PPC_SUB(b2p[SKB_HLEN_REG], b2p[SKB_HLEN_REG], b2p[TMP_REG_1]);
106
107 /* skb->data pointer */
108 PPC_BPF_LL(b2p[SKB_DATA_REG], 3, offsetof(struct sk_buff, data));
109 }
110
111 static void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx)
112 {
113 int i;
114
115 /*
116 * Initialize tail_call_cnt if we do tail calls.
117 * Otherwise, put in NOPs so that it can be skipped when we are
118 * invoked through a tail call.
119 */
120 if (ctx->seen & SEEN_TAILCALL) {
121 PPC_LI(b2p[TMP_REG_1], 0);
122 /* this goes in the redzone */
123 PPC_BPF_STL(b2p[TMP_REG_1], 1, -(BPF_PPC_STACK_SAVE + 8));
124 } else {
125 PPC_NOP();
126 PPC_NOP();
127 }
128
129 #define BPF_TAILCALL_PROLOGUE_SIZE 8
130
131 if (bpf_has_stack_frame(ctx)) {
132 /*
133 * We need a stack frame, but we don't necessarily need to
134 * save/restore LR unless we call other functions
135 */
136 if (ctx->seen & SEEN_FUNC) {
137 EMIT(PPC_INST_MFLR | __PPC_RT(R0));
138 PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
139 }
140
141 PPC_BPF_STLU(1, 1, -BPF_PPC_STACKFRAME);
142 }
143
144 /*
145 * Back up non-volatile regs -- BPF registers 6-10
146 * If we haven't created our own stack frame, we save these
147 * in the protected zone below the previous stack frame
148 */
149 for (i = BPF_REG_6; i <= BPF_REG_10; i++)
150 if (bpf_is_seen_register(ctx, i))
151 PPC_BPF_STL(b2p[i], 1, bpf_jit_stack_offsetof(ctx, b2p[i]));
152
153 /*
154 * Save additional non-volatile regs if we cache skb
155 * Also, setup skb data
156 */
157 if (ctx->seen & SEEN_SKB) {
158 PPC_BPF_STL(b2p[SKB_HLEN_REG], 1,
159 bpf_jit_stack_offsetof(ctx, b2p[SKB_HLEN_REG]));
160 PPC_BPF_STL(b2p[SKB_DATA_REG], 1,
161 bpf_jit_stack_offsetof(ctx, b2p[SKB_DATA_REG]));
162 bpf_jit_emit_skb_loads(image, ctx);
163 }
164
165 /* Setup frame pointer to point to the bpf stack area */
166 if (bpf_is_seen_register(ctx, BPF_REG_FP))
167 PPC_ADDI(b2p[BPF_REG_FP], 1,
168 STACK_FRAME_MIN_SIZE + MAX_BPF_STACK);
169 }
170
171 static void bpf_jit_emit_common_epilogue(u32 *image, struct codegen_context *ctx)
172 {
173 int i;
174
175 /* Restore NVRs */
176 for (i = BPF_REG_6; i <= BPF_REG_10; i++)
177 if (bpf_is_seen_register(ctx, i))
178 PPC_BPF_LL(b2p[i], 1, bpf_jit_stack_offsetof(ctx, b2p[i]));
179
180 /* Restore non-volatile registers used for skb cache */
181 if (ctx->seen & SEEN_SKB) {
182 PPC_BPF_LL(b2p[SKB_HLEN_REG], 1,
183 bpf_jit_stack_offsetof(ctx, b2p[SKB_HLEN_REG]));
184 PPC_BPF_LL(b2p[SKB_DATA_REG], 1,
185 bpf_jit_stack_offsetof(ctx, b2p[SKB_DATA_REG]));
186 }
187
188 /* Tear down our stack frame */
189 if (bpf_has_stack_frame(ctx)) {
190 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
191 if (ctx->seen & SEEN_FUNC) {
192 PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
193 PPC_MTLR(0);
194 }
195 }
196 }
197
198 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
199 {
200 bpf_jit_emit_common_epilogue(image, ctx);
201
202 /* Move result to r3 */
203 PPC_MR(3, b2p[BPF_REG_0]);
204
205 PPC_BLR();
206 }
207
208 static void bpf_jit_emit_func_call(u32 *image, struct codegen_context *ctx, u64 func)
209 {
210 #ifdef PPC64_ELF_ABI_v1
211 /* func points to the function descriptor */
212 PPC_LI64(b2p[TMP_REG_2], func);
213 /* Load actual entry point from function descriptor */
214 PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_2], 0);
215 /* ... and move it to LR */
216 PPC_MTLR(b2p[TMP_REG_1]);
217 /*
218 * Load TOC from function descriptor at offset 8.
219 * We can clobber r2 since we get called through a
220 * function pointer (so caller will save/restore r2)
221 * and since we don't use a TOC ourself.
222 */
223 PPC_BPF_LL(2, b2p[TMP_REG_2], 8);
224 #else
225 /* We can clobber r12 */
226 PPC_FUNC_ADDR(12, func);
227 PPC_MTLR(12);
228 #endif
229 PPC_BLRL();
230 }
231
232 static void bpf_jit_emit_tail_call(u32 *image, struct codegen_context *ctx, u32 out)
233 {
234 /*
235 * By now, the eBPF program has already setup parameters in r3, r4 and r5
236 * r3/BPF_REG_1 - pointer to ctx -- passed as is to the next bpf program
237 * r4/BPF_REG_2 - pointer to bpf_array
238 * r5/BPF_REG_3 - index in bpf_array
239 */
240 int b2p_bpf_array = b2p[BPF_REG_2];
241 int b2p_index = b2p[BPF_REG_3];
242
243 /*
244 * if (index >= array->map.max_entries)
245 * goto out;
246 */
247 PPC_LWZ(b2p[TMP_REG_1], b2p_bpf_array, offsetof(struct bpf_array, map.max_entries));
248 PPC_CMPLW(b2p_index, b2p[TMP_REG_1]);
249 PPC_BCC(COND_GE, out);
250
251 /*
252 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
253 * goto out;
254 */
255 PPC_LD(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
256 PPC_CMPLWI(b2p[TMP_REG_1], MAX_TAIL_CALL_CNT);
257 PPC_BCC(COND_GT, out);
258
259 /*
260 * tail_call_cnt++;
261 */
262 PPC_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], 1);
263 PPC_BPF_STL(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
264
265 /* prog = array->ptrs[index]; */
266 PPC_MULI(b2p[TMP_REG_1], b2p_index, 8);
267 PPC_ADD(b2p[TMP_REG_1], b2p[TMP_REG_1], b2p_bpf_array);
268 PPC_LD(b2p[TMP_REG_1], b2p[TMP_REG_1], offsetof(struct bpf_array, ptrs));
269
270 /*
271 * if (prog == NULL)
272 * goto out;
273 */
274 PPC_CMPLDI(b2p[TMP_REG_1], 0);
275 PPC_BCC(COND_EQ, out);
276
277 /* goto *(prog->bpf_func + prologue_size); */
278 PPC_LD(b2p[TMP_REG_1], b2p[TMP_REG_1], offsetof(struct bpf_prog, bpf_func));
279 #ifdef PPC64_ELF_ABI_v1
280 /* skip past the function descriptor */
281 PPC_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1],
282 FUNCTION_DESCR_SIZE + BPF_TAILCALL_PROLOGUE_SIZE);
283 #else
284 PPC_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], BPF_TAILCALL_PROLOGUE_SIZE);
285 #endif
286 PPC_MTCTR(b2p[TMP_REG_1]);
287
288 /* tear down stack, restore NVRs, ... */
289 bpf_jit_emit_common_epilogue(image, ctx);
290
291 PPC_BCTR();
292 /* out: */
293 }
294
295 /* Assemble the body code between the prologue & epilogue */
296 static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
297 struct codegen_context *ctx,
298 u32 *addrs)
299 {
300 const struct bpf_insn *insn = fp->insnsi;
301 int flen = fp->len;
302 int i;
303
304 /* Start of epilogue code - will only be valid 2nd pass onwards */
305 u32 exit_addr = addrs[flen];
306
307 for (i = 0; i < flen; i++) {
308 u32 code = insn[i].code;
309 u32 dst_reg = b2p[insn[i].dst_reg];
310 u32 src_reg = b2p[insn[i].src_reg];
311 s16 off = insn[i].off;
312 s32 imm = insn[i].imm;
313 u64 imm64;
314 u8 *func;
315 u32 true_cond;
316
317 /*
318 * addrs[] maps a BPF bytecode address into a real offset from
319 * the start of the body code.
320 */
321 addrs[i] = ctx->idx * 4;
322
323 /*
324 * As an optimization, we note down which non-volatile registers
325 * are used so that we can only save/restore those in our
326 * prologue and epilogue. We do this here regardless of whether
327 * the actual BPF instruction uses src/dst registers or not
328 * (for instance, BPF_CALL does not use them). The expectation
329 * is that those instructions will have src_reg/dst_reg set to
330 * 0. Even otherwise, we just lose some prologue/epilogue
331 * optimization but everything else should work without
332 * any issues.
333 */
334 if (dst_reg >= BPF_PPC_NVR_MIN && dst_reg < 32)
335 bpf_set_seen_register(ctx, insn[i].dst_reg);
336 if (src_reg >= BPF_PPC_NVR_MIN && src_reg < 32)
337 bpf_set_seen_register(ctx, insn[i].src_reg);
338
339 switch (code) {
340 /*
341 * Arithmetic operations: ADD/SUB/MUL/DIV/MOD/NEG
342 */
343 case BPF_ALU | BPF_ADD | BPF_X: /* (u32) dst += (u32) src */
344 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst += src */
345 PPC_ADD(dst_reg, dst_reg, src_reg);
346 goto bpf_alu32_trunc;
347 case BPF_ALU | BPF_SUB | BPF_X: /* (u32) dst -= (u32) src */
348 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst -= src */
349 PPC_SUB(dst_reg, dst_reg, src_reg);
350 goto bpf_alu32_trunc;
351 case BPF_ALU | BPF_ADD | BPF_K: /* (u32) dst += (u32) imm */
352 case BPF_ALU | BPF_SUB | BPF_K: /* (u32) dst -= (u32) imm */
353 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst += imm */
354 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst -= imm */
355 if (BPF_OP(code) == BPF_SUB)
356 imm = -imm;
357 if (imm) {
358 if (imm >= -32768 && imm < 32768)
359 PPC_ADDI(dst_reg, dst_reg, IMM_L(imm));
360 else {
361 PPC_LI32(b2p[TMP_REG_1], imm);
362 PPC_ADD(dst_reg, dst_reg, b2p[TMP_REG_1]);
363 }
364 }
365 goto bpf_alu32_trunc;
366 case BPF_ALU | BPF_MUL | BPF_X: /* (u32) dst *= (u32) src */
367 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst *= src */
368 if (BPF_CLASS(code) == BPF_ALU)
369 PPC_MULW(dst_reg, dst_reg, src_reg);
370 else
371 PPC_MULD(dst_reg, dst_reg, src_reg);
372 goto bpf_alu32_trunc;
373 case BPF_ALU | BPF_MUL | BPF_K: /* (u32) dst *= (u32) imm */
374 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst *= imm */
375 if (imm >= -32768 && imm < 32768)
376 PPC_MULI(dst_reg, dst_reg, IMM_L(imm));
377 else {
378 PPC_LI32(b2p[TMP_REG_1], imm);
379 if (BPF_CLASS(code) == BPF_ALU)
380 PPC_MULW(dst_reg, dst_reg,
381 b2p[TMP_REG_1]);
382 else
383 PPC_MULD(dst_reg, dst_reg,
384 b2p[TMP_REG_1]);
385 }
386 goto bpf_alu32_trunc;
387 case BPF_ALU | BPF_DIV | BPF_X: /* (u32) dst /= (u32) src */
388 case BPF_ALU | BPF_MOD | BPF_X: /* (u32) dst %= (u32) src */
389 PPC_CMPWI(src_reg, 0);
390 PPC_BCC_SHORT(COND_NE, (ctx->idx * 4) + 12);
391 PPC_LI(b2p[BPF_REG_0], 0);
392 PPC_JMP(exit_addr);
393 if (BPF_OP(code) == BPF_MOD) {
394 PPC_DIVWU(b2p[TMP_REG_1], dst_reg, src_reg);
395 PPC_MULW(b2p[TMP_REG_1], src_reg,
396 b2p[TMP_REG_1]);
397 PPC_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]);
398 } else
399 PPC_DIVWU(dst_reg, dst_reg, src_reg);
400 goto bpf_alu32_trunc;
401 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst /= src */
402 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst %= src */
403 PPC_CMPDI(src_reg, 0);
404 PPC_BCC_SHORT(COND_NE, (ctx->idx * 4) + 12);
405 PPC_LI(b2p[BPF_REG_0], 0);
406 PPC_JMP(exit_addr);
407 if (BPF_OP(code) == BPF_MOD) {
408 PPC_DIVD(b2p[TMP_REG_1], dst_reg, src_reg);
409 PPC_MULD(b2p[TMP_REG_1], src_reg,
410 b2p[TMP_REG_1]);
411 PPC_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]);
412 } else
413 PPC_DIVD(dst_reg, dst_reg, src_reg);
414 break;
415 case BPF_ALU | BPF_MOD | BPF_K: /* (u32) dst %= (u32) imm */
416 case BPF_ALU | BPF_DIV | BPF_K: /* (u32) dst /= (u32) imm */
417 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst %= imm */
418 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst /= imm */
419 if (imm == 0)
420 return -EINVAL;
421 else if (imm == 1)
422 goto bpf_alu32_trunc;
423
424 PPC_LI32(b2p[TMP_REG_1], imm);
425 switch (BPF_CLASS(code)) {
426 case BPF_ALU:
427 if (BPF_OP(code) == BPF_MOD) {
428 PPC_DIVWU(b2p[TMP_REG_2], dst_reg,
429 b2p[TMP_REG_1]);
430 PPC_MULW(b2p[TMP_REG_1],
431 b2p[TMP_REG_1],
432 b2p[TMP_REG_2]);
433 PPC_SUB(dst_reg, dst_reg,
434 b2p[TMP_REG_1]);
435 } else
436 PPC_DIVWU(dst_reg, dst_reg,
437 b2p[TMP_REG_1]);
438 break;
439 case BPF_ALU64:
440 if (BPF_OP(code) == BPF_MOD) {
441 PPC_DIVD(b2p[TMP_REG_2], dst_reg,
442 b2p[TMP_REG_1]);
443 PPC_MULD(b2p[TMP_REG_1],
444 b2p[TMP_REG_1],
445 b2p[TMP_REG_2]);
446 PPC_SUB(dst_reg, dst_reg,
447 b2p[TMP_REG_1]);
448 } else
449 PPC_DIVD(dst_reg, dst_reg,
450 b2p[TMP_REG_1]);
451 break;
452 }
453 goto bpf_alu32_trunc;
454 case BPF_ALU | BPF_NEG: /* (u32) dst = -dst */
455 case BPF_ALU64 | BPF_NEG: /* dst = -dst */
456 PPC_NEG(dst_reg, dst_reg);
457 goto bpf_alu32_trunc;
458
459 /*
460 * Logical operations: AND/OR/XOR/[A]LSH/[A]RSH
461 */
462 case BPF_ALU | BPF_AND | BPF_X: /* (u32) dst = dst & src */
463 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
464 PPC_AND(dst_reg, dst_reg, src_reg);
465 goto bpf_alu32_trunc;
466 case BPF_ALU | BPF_AND | BPF_K: /* (u32) dst = dst & imm */
467 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
468 if (!IMM_H(imm))
469 PPC_ANDI(dst_reg, dst_reg, IMM_L(imm));
470 else {
471 /* Sign-extended */
472 PPC_LI32(b2p[TMP_REG_1], imm);
473 PPC_AND(dst_reg, dst_reg, b2p[TMP_REG_1]);
474 }
475 goto bpf_alu32_trunc;
476 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
477 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
478 PPC_OR(dst_reg, dst_reg, src_reg);
479 goto bpf_alu32_trunc;
480 case BPF_ALU | BPF_OR | BPF_K:/* dst = (u32) dst | (u32) imm */
481 case BPF_ALU64 | BPF_OR | BPF_K:/* dst = dst | imm */
482 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
483 /* Sign-extended */
484 PPC_LI32(b2p[TMP_REG_1], imm);
485 PPC_OR(dst_reg, dst_reg, b2p[TMP_REG_1]);
486 } else {
487 if (IMM_L(imm))
488 PPC_ORI(dst_reg, dst_reg, IMM_L(imm));
489 if (IMM_H(imm))
490 PPC_ORIS(dst_reg, dst_reg, IMM_H(imm));
491 }
492 goto bpf_alu32_trunc;
493 case BPF_ALU | BPF_XOR | BPF_X: /* (u32) dst ^= src */
494 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst ^= src */
495 PPC_XOR(dst_reg, dst_reg, src_reg);
496 goto bpf_alu32_trunc;
497 case BPF_ALU | BPF_XOR | BPF_K: /* (u32) dst ^= (u32) imm */
498 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst ^= imm */
499 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
500 /* Sign-extended */
501 PPC_LI32(b2p[TMP_REG_1], imm);
502 PPC_XOR(dst_reg, dst_reg, b2p[TMP_REG_1]);
503 } else {
504 if (IMM_L(imm))
505 PPC_XORI(dst_reg, dst_reg, IMM_L(imm));
506 if (IMM_H(imm))
507 PPC_XORIS(dst_reg, dst_reg, IMM_H(imm));
508 }
509 goto bpf_alu32_trunc;
510 case BPF_ALU | BPF_LSH | BPF_X: /* (u32) dst <<= (u32) src */
511 /* slw clears top 32 bits */
512 PPC_SLW(dst_reg, dst_reg, src_reg);
513 break;
514 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */
515 PPC_SLD(dst_reg, dst_reg, src_reg);
516 break;
517 case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<== (u32) imm */
518 /* with imm 0, we still need to clear top 32 bits */
519 PPC_SLWI(dst_reg, dst_reg, imm);
520 break;
521 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<== imm */
522 if (imm != 0)
523 PPC_SLDI(dst_reg, dst_reg, imm);
524 break;
525 case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */
526 PPC_SRW(dst_reg, dst_reg, src_reg);
527 break;
528 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */
529 PPC_SRD(dst_reg, dst_reg, src_reg);
530 break;
531 case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */
532 PPC_SRWI(dst_reg, dst_reg, imm);
533 break;
534 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */
535 if (imm != 0)
536 PPC_SRDI(dst_reg, dst_reg, imm);
537 break;
538 case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */
539 PPC_SRAD(dst_reg, dst_reg, src_reg);
540 break;
541 case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */
542 if (imm != 0)
543 PPC_SRADI(dst_reg, dst_reg, imm);
544 break;
545
546 /*
547 * MOV
548 */
549 case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
550 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
551 PPC_MR(dst_reg, src_reg);
552 goto bpf_alu32_trunc;
553 case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */
554 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */
555 PPC_LI32(dst_reg, imm);
556 if (imm < 0)
557 goto bpf_alu32_trunc;
558 break;
559
560 bpf_alu32_trunc:
561 /* Truncate to 32-bits */
562 if (BPF_CLASS(code) == BPF_ALU)
563 PPC_RLWINM(dst_reg, dst_reg, 0, 0, 31);
564 break;
565
566 /*
567 * BPF_FROM_BE/LE
568 */
569 case BPF_ALU | BPF_END | BPF_FROM_LE:
570 case BPF_ALU | BPF_END | BPF_FROM_BE:
571 #ifdef __BIG_ENDIAN__
572 if (BPF_SRC(code) == BPF_FROM_BE)
573 goto emit_clear;
574 #else /* !__BIG_ENDIAN__ */
575 if (BPF_SRC(code) == BPF_FROM_LE)
576 goto emit_clear;
577 #endif
578 switch (imm) {
579 case 16:
580 /* Rotate 8 bits left & mask with 0x0000ff00 */
581 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 16, 23);
582 /* Rotate 8 bits right & insert LSB to reg */
583 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 24, 31);
584 /* Move result back to dst_reg */
585 PPC_MR(dst_reg, b2p[TMP_REG_1]);
586 break;
587 case 32:
588 /*
589 * Rotate word left by 8 bits:
590 * 2 bytes are already in their final position
591 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4)
592 */
593 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 0, 31);
594 /* Rotate 24 bits and insert byte 1 */
595 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 0, 7);
596 /* Rotate 24 bits and insert byte 3 */
597 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 16, 23);
598 PPC_MR(dst_reg, b2p[TMP_REG_1]);
599 break;
600 case 64:
601 /*
602 * Way easier and faster(?) to store the value
603 * into stack and then use ldbrx
604 *
605 * ctx->seen will be reliable in pass2, but
606 * the instructions generated will remain the
607 * same across all passes
608 */
609 PPC_STD(dst_reg, 1, bpf_jit_stack_local(ctx));
610 PPC_ADDI(b2p[TMP_REG_1], 1, bpf_jit_stack_local(ctx));
611 PPC_LDBRX(dst_reg, 0, b2p[TMP_REG_1]);
612 break;
613 }
614 break;
615
616 emit_clear:
617 switch (imm) {
618 case 16:
619 /* zero-extend 16 bits into 64 bits */
620 PPC_RLDICL(dst_reg, dst_reg, 0, 48);
621 break;
622 case 32:
623 /* zero-extend 32 bits into 64 bits */
624 PPC_RLDICL(dst_reg, dst_reg, 0, 32);
625 break;
626 case 64:
627 /* nop */
628 break;
629 }
630 break;
631
632 /*
633 * BPF_ST(X)
634 */
635 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */
636 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
637 if (BPF_CLASS(code) == BPF_ST) {
638 PPC_LI(b2p[TMP_REG_1], imm);
639 src_reg = b2p[TMP_REG_1];
640 }
641 PPC_STB(src_reg, dst_reg, off);
642 break;
643 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
644 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
645 if (BPF_CLASS(code) == BPF_ST) {
646 PPC_LI(b2p[TMP_REG_1], imm);
647 src_reg = b2p[TMP_REG_1];
648 }
649 PPC_STH(src_reg, dst_reg, off);
650 break;
651 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
652 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
653 if (BPF_CLASS(code) == BPF_ST) {
654 PPC_LI32(b2p[TMP_REG_1], imm);
655 src_reg = b2p[TMP_REG_1];
656 }
657 PPC_STW(src_reg, dst_reg, off);
658 break;
659 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
660 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
661 if (BPF_CLASS(code) == BPF_ST) {
662 PPC_LI32(b2p[TMP_REG_1], imm);
663 src_reg = b2p[TMP_REG_1];
664 }
665 PPC_STD(src_reg, dst_reg, off);
666 break;
667
668 /*
669 * BPF_STX XADD (atomic_add)
670 */
671 /* *(u32 *)(dst + off) += src */
672 case BPF_STX | BPF_XADD | BPF_W:
673 /* Get EA into TMP_REG_1 */
674 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
675 /* error if EA is not word-aligned */
676 PPC_ANDI(b2p[TMP_REG_2], b2p[TMP_REG_1], 0x03);
677 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + 12);
678 PPC_LI(b2p[BPF_REG_0], 0);
679 PPC_JMP(exit_addr);
680 /* load value from memory into TMP_REG_2 */
681 PPC_BPF_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
682 /* add value from src_reg into this */
683 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
684 /* store result back */
685 PPC_BPF_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
686 /* we're done if this succeeded */
687 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + (7*4));
688 /* otherwise, let's try once more */
689 PPC_BPF_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
690 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
691 PPC_BPF_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
692 /* exit if the store was not successful */
693 PPC_LI(b2p[BPF_REG_0], 0);
694 PPC_BCC(COND_NE, exit_addr);
695 break;
696 /* *(u64 *)(dst + off) += src */
697 case BPF_STX | BPF_XADD | BPF_DW:
698 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
699 /* error if EA is not doubleword-aligned */
700 PPC_ANDI(b2p[TMP_REG_2], b2p[TMP_REG_1], 0x07);
701 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + (3*4));
702 PPC_LI(b2p[BPF_REG_0], 0);
703 PPC_JMP(exit_addr);
704 PPC_BPF_LDARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
705 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
706 PPC_BPF_STDCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
707 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + (7*4));
708 PPC_BPF_LDARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
709 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
710 PPC_BPF_STDCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
711 PPC_LI(b2p[BPF_REG_0], 0);
712 PPC_BCC(COND_NE, exit_addr);
713 break;
714
715 /*
716 * BPF_LDX
717 */
718 /* dst = *(u8 *)(ul) (src + off) */
719 case BPF_LDX | BPF_MEM | BPF_B:
720 PPC_LBZ(dst_reg, src_reg, off);
721 break;
722 /* dst = *(u16 *)(ul) (src + off) */
723 case BPF_LDX | BPF_MEM | BPF_H:
724 PPC_LHZ(dst_reg, src_reg, off);
725 break;
726 /* dst = *(u32 *)(ul) (src + off) */
727 case BPF_LDX | BPF_MEM | BPF_W:
728 PPC_LWZ(dst_reg, src_reg, off);
729 break;
730 /* dst = *(u64 *)(ul) (src + off) */
731 case BPF_LDX | BPF_MEM | BPF_DW:
732 PPC_LD(dst_reg, src_reg, off);
733 break;
734
735 /*
736 * Doubleword load
737 * 16 byte instruction that uses two 'struct bpf_insn'
738 */
739 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
740 imm64 = ((u64)(u32) insn[i].imm) |
741 (((u64)(u32) insn[i+1].imm) << 32);
742 /* Adjust for two bpf instructions */
743 addrs[++i] = ctx->idx * 4;
744 PPC_LI64(dst_reg, imm64);
745 break;
746
747 /*
748 * Return/Exit
749 */
750 case BPF_JMP | BPF_EXIT:
751 /*
752 * If this isn't the very last instruction, branch to
753 * the epilogue. If we _are_ the last instruction,
754 * we'll just fall through to the epilogue.
755 */
756 if (i != flen - 1)
757 PPC_JMP(exit_addr);
758 /* else fall through to the epilogue */
759 break;
760
761 /*
762 * Call kernel helper
763 */
764 case BPF_JMP | BPF_CALL:
765 ctx->seen |= SEEN_FUNC;
766 func = (u8 *) __bpf_call_base + imm;
767
768 /* Save skb pointer if we need to re-cache skb data */
769 if (bpf_helper_changes_pkt_data(func))
770 PPC_BPF_STL(3, 1, bpf_jit_stack_local(ctx));
771
772 bpf_jit_emit_func_call(image, ctx, (u64)func);
773
774 /* move return value from r3 to BPF_REG_0 */
775 PPC_MR(b2p[BPF_REG_0], 3);
776
777 /* refresh skb cache */
778 if (bpf_helper_changes_pkt_data(func)) {
779 /* reload skb pointer to r3 */
780 PPC_BPF_LL(3, 1, bpf_jit_stack_local(ctx));
781 bpf_jit_emit_skb_loads(image, ctx);
782 }
783 break;
784
785 /*
786 * Jumps and branches
787 */
788 case BPF_JMP | BPF_JA:
789 PPC_JMP(addrs[i + 1 + off]);
790 break;
791
792 case BPF_JMP | BPF_JGT | BPF_K:
793 case BPF_JMP | BPF_JGT | BPF_X:
794 case BPF_JMP | BPF_JSGT | BPF_K:
795 case BPF_JMP | BPF_JSGT | BPF_X:
796 true_cond = COND_GT;
797 goto cond_branch;
798 case BPF_JMP | BPF_JGE | BPF_K:
799 case BPF_JMP | BPF_JGE | BPF_X:
800 case BPF_JMP | BPF_JSGE | BPF_K:
801 case BPF_JMP | BPF_JSGE | BPF_X:
802 true_cond = COND_GE;
803 goto cond_branch;
804 case BPF_JMP | BPF_JEQ | BPF_K:
805 case BPF_JMP | BPF_JEQ | BPF_X:
806 true_cond = COND_EQ;
807 goto cond_branch;
808 case BPF_JMP | BPF_JNE | BPF_K:
809 case BPF_JMP | BPF_JNE | BPF_X:
810 true_cond = COND_NE;
811 goto cond_branch;
812 case BPF_JMP | BPF_JSET | BPF_K:
813 case BPF_JMP | BPF_JSET | BPF_X:
814 true_cond = COND_NE;
815 /* Fall through */
816
817 cond_branch:
818 switch (code) {
819 case BPF_JMP | BPF_JGT | BPF_X:
820 case BPF_JMP | BPF_JGE | BPF_X:
821 case BPF_JMP | BPF_JEQ | BPF_X:
822 case BPF_JMP | BPF_JNE | BPF_X:
823 /* unsigned comparison */
824 PPC_CMPLD(dst_reg, src_reg);
825 break;
826 case BPF_JMP | BPF_JSGT | BPF_X:
827 case BPF_JMP | BPF_JSGE | BPF_X:
828 /* signed comparison */
829 PPC_CMPD(dst_reg, src_reg);
830 break;
831 case BPF_JMP | BPF_JSET | BPF_X:
832 PPC_AND_DOT(b2p[TMP_REG_1], dst_reg, src_reg);
833 break;
834 case BPF_JMP | BPF_JNE | BPF_K:
835 case BPF_JMP | BPF_JEQ | BPF_K:
836 case BPF_JMP | BPF_JGT | BPF_K:
837 case BPF_JMP | BPF_JGE | BPF_K:
838 /*
839 * Need sign-extended load, so only positive
840 * values can be used as imm in cmpldi
841 */
842 if (imm >= 0 && imm < 32768)
843 PPC_CMPLDI(dst_reg, imm);
844 else {
845 /* sign-extending load */
846 PPC_LI32(b2p[TMP_REG_1], imm);
847 /* ... but unsigned comparison */
848 PPC_CMPLD(dst_reg, b2p[TMP_REG_1]);
849 }
850 break;
851 case BPF_JMP | BPF_JSGT | BPF_K:
852 case BPF_JMP | BPF_JSGE | BPF_K:
853 /*
854 * signed comparison, so any 16-bit value
855 * can be used in cmpdi
856 */
857 if (imm >= -32768 && imm < 32768)
858 PPC_CMPDI(dst_reg, imm);
859 else {
860 PPC_LI32(b2p[TMP_REG_1], imm);
861 PPC_CMPD(dst_reg, b2p[TMP_REG_1]);
862 }
863 break;
864 case BPF_JMP | BPF_JSET | BPF_K:
865 /* andi does not sign-extend the immediate */
866 if (imm >= 0 && imm < 32768)
867 /* PPC_ANDI is _only/always_ dot-form */
868 PPC_ANDI(b2p[TMP_REG_1], dst_reg, imm);
869 else {
870 PPC_LI32(b2p[TMP_REG_1], imm);
871 PPC_AND_DOT(b2p[TMP_REG_1], dst_reg,
872 b2p[TMP_REG_1]);
873 }
874 break;
875 }
876 PPC_BCC(true_cond, addrs[i + 1 + off]);
877 break;
878
879 /*
880 * Loads from packet header/data
881 * Assume 32-bit input value in imm and X (src_reg)
882 */
883
884 /* Absolute loads */
885 case BPF_LD | BPF_W | BPF_ABS:
886 func = (u8 *)CHOOSE_LOAD_FUNC(imm, sk_load_word);
887 goto common_load_abs;
888 case BPF_LD | BPF_H | BPF_ABS:
889 func = (u8 *)CHOOSE_LOAD_FUNC(imm, sk_load_half);
890 goto common_load_abs;
891 case BPF_LD | BPF_B | BPF_ABS:
892 func = (u8 *)CHOOSE_LOAD_FUNC(imm, sk_load_byte);
893 common_load_abs:
894 /*
895 * Load from [imm]
896 * Load into r4, which can just be passed onto
897 * skb load helpers as the second parameter
898 */
899 PPC_LI32(4, imm);
900 goto common_load;
901
902 /* Indirect loads */
903 case BPF_LD | BPF_W | BPF_IND:
904 func = (u8 *)sk_load_word;
905 goto common_load_ind;
906 case BPF_LD | BPF_H | BPF_IND:
907 func = (u8 *)sk_load_half;
908 goto common_load_ind;
909 case BPF_LD | BPF_B | BPF_IND:
910 func = (u8 *)sk_load_byte;
911 common_load_ind:
912 /*
913 * Load from [src_reg + imm]
914 * Treat src_reg as a 32-bit value
915 */
916 PPC_EXTSW(4, src_reg);
917 if (imm) {
918 if (imm >= -32768 && imm < 32768)
919 PPC_ADDI(4, 4, IMM_L(imm));
920 else {
921 PPC_LI32(b2p[TMP_REG_1], imm);
922 PPC_ADD(4, 4, b2p[TMP_REG_1]);
923 }
924 }
925
926 common_load:
927 ctx->seen |= SEEN_SKB;
928 ctx->seen |= SEEN_FUNC;
929 bpf_jit_emit_func_call(image, ctx, (u64)func);
930
931 /*
932 * Helper returns 'lt' condition on error, and an
933 * appropriate return value in BPF_REG_0
934 */
935 PPC_BCC(COND_LT, exit_addr);
936 break;
937
938 /*
939 * Tail call
940 */
941 case BPF_JMP | BPF_CALL | BPF_X:
942 ctx->seen |= SEEN_TAILCALL;
943 bpf_jit_emit_tail_call(image, ctx, addrs[i + 1]);
944 break;
945
946 default:
947 /*
948 * The filter contains something cruel & unusual.
949 * We don't handle it, but also there shouldn't be
950 * anything missing from our list.
951 */
952 pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n",
953 code, i);
954 return -ENOTSUPP;
955 }
956 }
957
958 /* Set end-of-body-code address for exit. */
959 addrs[i] = ctx->idx * 4;
960
961 return 0;
962 }
963
964 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
965 {
966 u32 proglen;
967 u32 alloclen;
968 u8 *image = NULL;
969 u32 *code_base;
970 u32 *addrs;
971 struct codegen_context cgctx;
972 int pass;
973 int flen;
974 struct bpf_binary_header *bpf_hdr;
975 struct bpf_prog *org_fp = fp;
976 struct bpf_prog *tmp_fp;
977 bool bpf_blinded = false;
978
979 if (!bpf_jit_enable)
980 return org_fp;
981
982 tmp_fp = bpf_jit_blind_constants(org_fp);
983 if (IS_ERR(tmp_fp))
984 return org_fp;
985
986 if (tmp_fp != org_fp) {
987 bpf_blinded = true;
988 fp = tmp_fp;
989 }
990
991 flen = fp->len;
992 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
993 if (addrs == NULL) {
994 fp = org_fp;
995 goto out;
996 }
997
998 memset(&cgctx, 0, sizeof(struct codegen_context));
999
1000 /* Scouting faux-generate pass 0 */
1001 if (bpf_jit_build_body(fp, 0, &cgctx, addrs)) {
1002 /* We hit something illegal or unsupported. */
1003 fp = org_fp;
1004 goto out;
1005 }
1006
1007 /*
1008 * Pretend to build prologue, given the features we've seen. This will
1009 * update ctgtx.idx as it pretends to output instructions, then we can
1010 * calculate total size from idx.
1011 */
1012 bpf_jit_build_prologue(0, &cgctx);
1013 bpf_jit_build_epilogue(0, &cgctx);
1014
1015 proglen = cgctx.idx * 4;
1016 alloclen = proglen + FUNCTION_DESCR_SIZE;
1017
1018 bpf_hdr = bpf_jit_binary_alloc(alloclen, &image, 4,
1019 bpf_jit_fill_ill_insns);
1020 if (!bpf_hdr) {
1021 fp = org_fp;
1022 goto out;
1023 }
1024
1025 code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
1026
1027 /* Code generation passes 1-2 */
1028 for (pass = 1; pass < 3; pass++) {
1029 /* Now build the prologue, body code & epilogue for real. */
1030 cgctx.idx = 0;
1031 bpf_jit_build_prologue(code_base, &cgctx);
1032 bpf_jit_build_body(fp, code_base, &cgctx, addrs);
1033 bpf_jit_build_epilogue(code_base, &cgctx);
1034
1035 if (bpf_jit_enable > 1)
1036 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
1037 proglen - (cgctx.idx * 4), cgctx.seen);
1038 }
1039
1040 if (bpf_jit_enable > 1)
1041 /*
1042 * Note that we output the base address of the code_base
1043 * rather than image, since opcodes are in code_base.
1044 */
1045 bpf_jit_dump(flen, proglen, pass, code_base);
1046
1047 #ifdef PPC64_ELF_ABI_v1
1048 /* Function descriptor nastiness: Address + TOC */
1049 ((u64 *)image)[0] = (u64)code_base;
1050 ((u64 *)image)[1] = local_paca->kernel_toc;
1051 #endif
1052
1053 fp->bpf_func = (void *)image;
1054 fp->jited = 1;
1055
1056 bpf_flush_icache(bpf_hdr, (u8 *)bpf_hdr + (bpf_hdr->pages * PAGE_SIZE));
1057
1058 out:
1059 kfree(addrs);
1060
1061 if (bpf_blinded)
1062 bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
1063
1064 return fp;
1065 }
1066
1067 /* Overriding bpf_jit_free() as we don't set images read-only. */
1068 void bpf_jit_free(struct bpf_prog *fp)
1069 {
1070 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
1071 struct bpf_binary_header *bpf_hdr = (void *)addr;
1072
1073 if (fp->jited)
1074 bpf_jit_binary_free(bpf_hdr);
1075
1076 bpf_prog_unlock_free(fp);
1077 }
Linux with added FPC-III support