dmaengine: imx-sdma: (trivial) fix a typo
[linux-2.6/btrfs-unstable.git] / arch / powerpc / net / bpf_jit_comp64.c
blob6073b78516f69348b7e80ed3680224c67a026e2c
1 /*
2 * bpf_jit_comp64.c: eBPF JIT compiler
4 * Copyright 2016 Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
5 * IBM Corporation
7 * Based on the powerpc classic BPF JIT compiler by Matt Evans
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.
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>
21 #include "bpf_jit64.h"
23 int bpf_jit_enable __read_mostly;
25 static void bpf_jit_fill_ill_insns(void *area, unsigned int size)
27 int *p = area;
29 /* Fill whole space with trap instructions */
30 while (p < (int *)((char *)area + size))
31 *p++ = BREAKPOINT_INSTRUCTION;
34 static inline void bpf_flush_icache(void *start, void *end)
36 smp_wmb();
37 flush_icache_range((unsigned long)start, (unsigned long)end);
40 static inline bool bpf_is_seen_register(struct codegen_context *ctx, int i)
42 return (ctx->seen & (1 << (31 - b2p[i])));
45 static inline void bpf_set_seen_register(struct codegen_context *ctx, int i)
47 ctx->seen |= (1 << (31 - b2p[i]));
50 static inline bool bpf_has_stack_frame(struct codegen_context *ctx)
53 * We only need a stack frame if:
54 * - we call other functions (kernel helpers), or
55 * - the bpf program uses its stack area
56 * The latter condition is deduced from the usage of BPF_REG_FP
58 return ctx->seen & SEEN_FUNC || bpf_is_seen_register(ctx, BPF_REG_FP);
61 static void bpf_jit_emit_skb_loads(u32 *image, struct codegen_context *ctx)
64 * Load skb->len and skb->data_len
65 * r3 points to skb
67 PPC_LWZ(b2p[SKB_HLEN_REG], 3, offsetof(struct sk_buff, len));
68 PPC_LWZ(b2p[TMP_REG_1], 3, offsetof(struct sk_buff, data_len));
69 /* header_len = len - data_len */
70 PPC_SUB(b2p[SKB_HLEN_REG], b2p[SKB_HLEN_REG], b2p[TMP_REG_1]);
72 /* skb->data pointer */
73 PPC_BPF_LL(b2p[SKB_DATA_REG], 3, offsetof(struct sk_buff, data));
76 static void bpf_jit_emit_func_call(u32 *image, struct codegen_context *ctx, u64 func)
78 #ifdef PPC64_ELF_ABI_v1
79 /* func points to the function descriptor */
80 PPC_LI64(b2p[TMP_REG_2], func);
81 /* Load actual entry point from function descriptor */
82 PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_2], 0);
83 /* ... and move it to LR */
84 PPC_MTLR(b2p[TMP_REG_1]);
86 * Load TOC from function descriptor at offset 8.
87 * We can clobber r2 since we get called through a
88 * function pointer (so caller will save/restore r2)
89 * and since we don't use a TOC ourself.
91 PPC_BPF_LL(2, b2p[TMP_REG_2], 8);
92 #else
93 /* We can clobber r12 */
94 PPC_FUNC_ADDR(12, func);
95 PPC_MTLR(12);
96 #endif
97 PPC_BLRL();
100 static void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx)
102 int i;
103 bool new_stack_frame = bpf_has_stack_frame(ctx);
105 if (new_stack_frame) {
107 * We need a stack frame, but we don't necessarily need to
108 * save/restore LR unless we call other functions
110 if (ctx->seen & SEEN_FUNC) {
111 EMIT(PPC_INST_MFLR | __PPC_RT(R0));
112 PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
115 PPC_BPF_STLU(1, 1, -BPF_PPC_STACKFRAME);
119 * Back up non-volatile regs -- BPF registers 6-10
120 * If we haven't created our own stack frame, we save these
121 * in the protected zone below the previous stack frame
123 for (i = BPF_REG_6; i <= BPF_REG_10; i++)
124 if (bpf_is_seen_register(ctx, i))
125 PPC_BPF_STL(b2p[i], 1,
126 (new_stack_frame ? BPF_PPC_STACKFRAME : 0) -
127 (8 * (32 - b2p[i])));
130 * Save additional non-volatile regs if we cache skb
131 * Also, setup skb data
133 if (ctx->seen & SEEN_SKB) {
134 PPC_BPF_STL(b2p[SKB_HLEN_REG], 1,
135 BPF_PPC_STACKFRAME - (8 * (32 - b2p[SKB_HLEN_REG])));
136 PPC_BPF_STL(b2p[SKB_DATA_REG], 1,
137 BPF_PPC_STACKFRAME - (8 * (32 - b2p[SKB_DATA_REG])));
138 bpf_jit_emit_skb_loads(image, ctx);
141 /* Setup frame pointer to point to the bpf stack area */
142 if (bpf_is_seen_register(ctx, BPF_REG_FP))
143 PPC_ADDI(b2p[BPF_REG_FP], 1,
144 BPF_PPC_STACKFRAME - BPF_PPC_STACK_SAVE);
147 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
149 int i;
150 bool new_stack_frame = bpf_has_stack_frame(ctx);
152 /* Move result to r3 */
153 PPC_MR(3, b2p[BPF_REG_0]);
155 /* Restore NVRs */
156 for (i = BPF_REG_6; i <= BPF_REG_10; i++)
157 if (bpf_is_seen_register(ctx, i))
158 PPC_BPF_LL(b2p[i], 1,
159 (new_stack_frame ? BPF_PPC_STACKFRAME : 0) -
160 (8 * (32 - b2p[i])));
162 /* Restore non-volatile registers used for skb cache */
163 if (ctx->seen & SEEN_SKB) {
164 PPC_BPF_LL(b2p[SKB_HLEN_REG], 1,
165 BPF_PPC_STACKFRAME - (8 * (32 - b2p[SKB_HLEN_REG])));
166 PPC_BPF_LL(b2p[SKB_DATA_REG], 1,
167 BPF_PPC_STACKFRAME - (8 * (32 - b2p[SKB_DATA_REG])));
170 /* Tear down our stack frame */
171 if (new_stack_frame) {
172 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
173 if (ctx->seen & SEEN_FUNC) {
174 PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
175 PPC_MTLR(0);
179 PPC_BLR();
182 /* Assemble the body code between the prologue & epilogue */
183 static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
184 struct codegen_context *ctx,
185 u32 *addrs)
187 const struct bpf_insn *insn = fp->insnsi;
188 int flen = fp->len;
189 int i;
191 /* Start of epilogue code - will only be valid 2nd pass onwards */
192 u32 exit_addr = addrs[flen];
194 for (i = 0; i < flen; i++) {
195 u32 code = insn[i].code;
196 u32 dst_reg = b2p[insn[i].dst_reg];
197 u32 src_reg = b2p[insn[i].src_reg];
198 s16 off = insn[i].off;
199 s32 imm = insn[i].imm;
200 u64 imm64;
201 u8 *func;
202 u32 true_cond;
203 int stack_local_off;
206 * addrs[] maps a BPF bytecode address into a real offset from
207 * the start of the body code.
209 addrs[i] = ctx->idx * 4;
212 * As an optimization, we note down which non-volatile registers
213 * are used so that we can only save/restore those in our
214 * prologue and epilogue. We do this here regardless of whether
215 * the actual BPF instruction uses src/dst registers or not
216 * (for instance, BPF_CALL does not use them). The expectation
217 * is that those instructions will have src_reg/dst_reg set to
218 * 0. Even otherwise, we just lose some prologue/epilogue
219 * optimization but everything else should work without
220 * any issues.
222 if (dst_reg >= 24 && dst_reg <= 31)
223 bpf_set_seen_register(ctx, insn[i].dst_reg);
224 if (src_reg >= 24 && src_reg <= 31)
225 bpf_set_seen_register(ctx, insn[i].src_reg);
227 switch (code) {
229 * Arithmetic operations: ADD/SUB/MUL/DIV/MOD/NEG
231 case BPF_ALU | BPF_ADD | BPF_X: /* (u32) dst += (u32) src */
232 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst += src */
233 PPC_ADD(dst_reg, dst_reg, src_reg);
234 goto bpf_alu32_trunc;
235 case BPF_ALU | BPF_SUB | BPF_X: /* (u32) dst -= (u32) src */
236 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst -= src */
237 PPC_SUB(dst_reg, dst_reg, src_reg);
238 goto bpf_alu32_trunc;
239 case BPF_ALU | BPF_ADD | BPF_K: /* (u32) dst += (u32) imm */
240 case BPF_ALU | BPF_SUB | BPF_K: /* (u32) dst -= (u32) imm */
241 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst += imm */
242 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst -= imm */
243 if (BPF_OP(code) == BPF_SUB)
244 imm = -imm;
245 if (imm) {
246 if (imm >= -32768 && imm < 32768)
247 PPC_ADDI(dst_reg, dst_reg, IMM_L(imm));
248 else {
249 PPC_LI32(b2p[TMP_REG_1], imm);
250 PPC_ADD(dst_reg, dst_reg, b2p[TMP_REG_1]);
253 goto bpf_alu32_trunc;
254 case BPF_ALU | BPF_MUL | BPF_X: /* (u32) dst *= (u32) src */
255 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst *= src */
256 if (BPF_CLASS(code) == BPF_ALU)
257 PPC_MULW(dst_reg, dst_reg, src_reg);
258 else
259 PPC_MULD(dst_reg, dst_reg, src_reg);
260 goto bpf_alu32_trunc;
261 case BPF_ALU | BPF_MUL | BPF_K: /* (u32) dst *= (u32) imm */
262 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst *= imm */
263 if (imm >= -32768 && imm < 32768)
264 PPC_MULI(dst_reg, dst_reg, IMM_L(imm));
265 else {
266 PPC_LI32(b2p[TMP_REG_1], imm);
267 if (BPF_CLASS(code) == BPF_ALU)
268 PPC_MULW(dst_reg, dst_reg,
269 b2p[TMP_REG_1]);
270 else
271 PPC_MULD(dst_reg, dst_reg,
272 b2p[TMP_REG_1]);
274 goto bpf_alu32_trunc;
275 case BPF_ALU | BPF_DIV | BPF_X: /* (u32) dst /= (u32) src */
276 case BPF_ALU | BPF_MOD | BPF_X: /* (u32) dst %= (u32) src */
277 PPC_CMPWI(src_reg, 0);
278 PPC_BCC_SHORT(COND_NE, (ctx->idx * 4) + 12);
279 PPC_LI(b2p[BPF_REG_0], 0);
280 PPC_JMP(exit_addr);
281 if (BPF_OP(code) == BPF_MOD) {
282 PPC_DIVWU(b2p[TMP_REG_1], dst_reg, src_reg);
283 PPC_MULW(b2p[TMP_REG_1], src_reg,
284 b2p[TMP_REG_1]);
285 PPC_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]);
286 } else
287 PPC_DIVWU(dst_reg, dst_reg, src_reg);
288 goto bpf_alu32_trunc;
289 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst /= src */
290 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst %= src */
291 PPC_CMPDI(src_reg, 0);
292 PPC_BCC_SHORT(COND_NE, (ctx->idx * 4) + 12);
293 PPC_LI(b2p[BPF_REG_0], 0);
294 PPC_JMP(exit_addr);
295 if (BPF_OP(code) == BPF_MOD) {
296 PPC_DIVD(b2p[TMP_REG_1], dst_reg, src_reg);
297 PPC_MULD(b2p[TMP_REG_1], src_reg,
298 b2p[TMP_REG_1]);
299 PPC_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]);
300 } else
301 PPC_DIVD(dst_reg, dst_reg, src_reg);
302 break;
303 case BPF_ALU | BPF_MOD | BPF_K: /* (u32) dst %= (u32) imm */
304 case BPF_ALU | BPF_DIV | BPF_K: /* (u32) dst /= (u32) imm */
305 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst %= imm */
306 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst /= imm */
307 if (imm == 0)
308 return -EINVAL;
309 else if (imm == 1)
310 goto bpf_alu32_trunc;
312 PPC_LI32(b2p[TMP_REG_1], imm);
313 switch (BPF_CLASS(code)) {
314 case BPF_ALU:
315 if (BPF_OP(code) == BPF_MOD) {
316 PPC_DIVWU(b2p[TMP_REG_2], dst_reg,
317 b2p[TMP_REG_1]);
318 PPC_MULW(b2p[TMP_REG_1],
319 b2p[TMP_REG_1],
320 b2p[TMP_REG_2]);
321 PPC_SUB(dst_reg, dst_reg,
322 b2p[TMP_REG_1]);
323 } else
324 PPC_DIVWU(dst_reg, dst_reg,
325 b2p[TMP_REG_1]);
326 break;
327 case BPF_ALU64:
328 if (BPF_OP(code) == BPF_MOD) {
329 PPC_DIVD(b2p[TMP_REG_2], dst_reg,
330 b2p[TMP_REG_1]);
331 PPC_MULD(b2p[TMP_REG_1],
332 b2p[TMP_REG_1],
333 b2p[TMP_REG_2]);
334 PPC_SUB(dst_reg, dst_reg,
335 b2p[TMP_REG_1]);
336 } else
337 PPC_DIVD(dst_reg, dst_reg,
338 b2p[TMP_REG_1]);
339 break;
341 goto bpf_alu32_trunc;
342 case BPF_ALU | BPF_NEG: /* (u32) dst = -dst */
343 case BPF_ALU64 | BPF_NEG: /* dst = -dst */
344 PPC_NEG(dst_reg, dst_reg);
345 goto bpf_alu32_trunc;
348 * Logical operations: AND/OR/XOR/[A]LSH/[A]RSH
350 case BPF_ALU | BPF_AND | BPF_X: /* (u32) dst = dst & src */
351 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
352 PPC_AND(dst_reg, dst_reg, src_reg);
353 goto bpf_alu32_trunc;
354 case BPF_ALU | BPF_AND | BPF_K: /* (u32) dst = dst & imm */
355 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
356 if (!IMM_H(imm))
357 PPC_ANDI(dst_reg, dst_reg, IMM_L(imm));
358 else {
359 /* Sign-extended */
360 PPC_LI32(b2p[TMP_REG_1], imm);
361 PPC_AND(dst_reg, dst_reg, b2p[TMP_REG_1]);
363 goto bpf_alu32_trunc;
364 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
365 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
366 PPC_OR(dst_reg, dst_reg, src_reg);
367 goto bpf_alu32_trunc;
368 case BPF_ALU | BPF_OR | BPF_K:/* dst = (u32) dst | (u32) imm */
369 case BPF_ALU64 | BPF_OR | BPF_K:/* dst = dst | imm */
370 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
371 /* Sign-extended */
372 PPC_LI32(b2p[TMP_REG_1], imm);
373 PPC_OR(dst_reg, dst_reg, b2p[TMP_REG_1]);
374 } else {
375 if (IMM_L(imm))
376 PPC_ORI(dst_reg, dst_reg, IMM_L(imm));
377 if (IMM_H(imm))
378 PPC_ORIS(dst_reg, dst_reg, IMM_H(imm));
380 goto bpf_alu32_trunc;
381 case BPF_ALU | BPF_XOR | BPF_X: /* (u32) dst ^= src */
382 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst ^= src */
383 PPC_XOR(dst_reg, dst_reg, src_reg);
384 goto bpf_alu32_trunc;
385 case BPF_ALU | BPF_XOR | BPF_K: /* (u32) dst ^= (u32) imm */
386 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst ^= imm */
387 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
388 /* Sign-extended */
389 PPC_LI32(b2p[TMP_REG_1], imm);
390 PPC_XOR(dst_reg, dst_reg, b2p[TMP_REG_1]);
391 } else {
392 if (IMM_L(imm))
393 PPC_XORI(dst_reg, dst_reg, IMM_L(imm));
394 if (IMM_H(imm))
395 PPC_XORIS(dst_reg, dst_reg, IMM_H(imm));
397 goto bpf_alu32_trunc;
398 case BPF_ALU | BPF_LSH | BPF_X: /* (u32) dst <<= (u32) src */
399 /* slw clears top 32 bits */
400 PPC_SLW(dst_reg, dst_reg, src_reg);
401 break;
402 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */
403 PPC_SLD(dst_reg, dst_reg, src_reg);
404 break;
405 case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<== (u32) imm */
406 /* with imm 0, we still need to clear top 32 bits */
407 PPC_SLWI(dst_reg, dst_reg, imm);
408 break;
409 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<== imm */
410 if (imm != 0)
411 PPC_SLDI(dst_reg, dst_reg, imm);
412 break;
413 case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */
414 PPC_SRW(dst_reg, dst_reg, src_reg);
415 break;
416 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */
417 PPC_SRD(dst_reg, dst_reg, src_reg);
418 break;
419 case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */
420 PPC_SRWI(dst_reg, dst_reg, imm);
421 break;
422 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */
423 if (imm != 0)
424 PPC_SRDI(dst_reg, dst_reg, imm);
425 break;
426 case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */
427 PPC_SRAD(dst_reg, dst_reg, src_reg);
428 break;
429 case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */
430 if (imm != 0)
431 PPC_SRADI(dst_reg, dst_reg, imm);
432 break;
435 * MOV
437 case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
438 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
439 PPC_MR(dst_reg, src_reg);
440 goto bpf_alu32_trunc;
441 case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */
442 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */
443 PPC_LI32(dst_reg, imm);
444 if (imm < 0)
445 goto bpf_alu32_trunc;
446 break;
448 bpf_alu32_trunc:
449 /* Truncate to 32-bits */
450 if (BPF_CLASS(code) == BPF_ALU)
451 PPC_RLWINM(dst_reg, dst_reg, 0, 0, 31);
452 break;
455 * BPF_FROM_BE/LE
457 case BPF_ALU | BPF_END | BPF_FROM_LE:
458 case BPF_ALU | BPF_END | BPF_FROM_BE:
459 #ifdef __BIG_ENDIAN__
460 if (BPF_SRC(code) == BPF_FROM_BE)
461 goto emit_clear;
462 #else /* !__BIG_ENDIAN__ */
463 if (BPF_SRC(code) == BPF_FROM_LE)
464 goto emit_clear;
465 #endif
466 switch (imm) {
467 case 16:
468 /* Rotate 8 bits left & mask with 0x0000ff00 */
469 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 16, 23);
470 /* Rotate 8 bits right & insert LSB to reg */
471 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 24, 31);
472 /* Move result back to dst_reg */
473 PPC_MR(dst_reg, b2p[TMP_REG_1]);
474 break;
475 case 32:
477 * Rotate word left by 8 bits:
478 * 2 bytes are already in their final position
479 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4)
481 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 0, 31);
482 /* Rotate 24 bits and insert byte 1 */
483 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 0, 7);
484 /* Rotate 24 bits and insert byte 3 */
485 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 16, 23);
486 PPC_MR(dst_reg, b2p[TMP_REG_1]);
487 break;
488 case 64:
490 * Way easier and faster(?) to store the value
491 * into stack and then use ldbrx
493 * First, determine where in stack we can store
494 * this:
495 * - if we have allotted a stack frame, then we
496 * will utilize the area set aside by
497 * BPF_PPC_STACK_LOCALS
498 * - else, we use the area beneath the NV GPR
499 * save area
501 * ctx->seen will be reliable in pass2, but
502 * the instructions generated will remain the
503 * same across all passes
505 if (bpf_has_stack_frame(ctx))
506 stack_local_off = STACK_FRAME_MIN_SIZE;
507 else
508 stack_local_off = -(BPF_PPC_STACK_SAVE + 8);
510 PPC_STD(dst_reg, 1, stack_local_off);
511 PPC_ADDI(b2p[TMP_REG_1], 1, stack_local_off);
512 PPC_LDBRX(dst_reg, 0, b2p[TMP_REG_1]);
513 break;
515 break;
517 emit_clear:
518 switch (imm) {
519 case 16:
520 /* zero-extend 16 bits into 64 bits */
521 PPC_RLDICL(dst_reg, dst_reg, 0, 48);
522 break;
523 case 32:
524 /* zero-extend 32 bits into 64 bits */
525 PPC_RLDICL(dst_reg, dst_reg, 0, 32);
526 break;
527 case 64:
528 /* nop */
529 break;
531 break;
534 * BPF_ST(X)
536 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */
537 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
538 if (BPF_CLASS(code) == BPF_ST) {
539 PPC_LI(b2p[TMP_REG_1], imm);
540 src_reg = b2p[TMP_REG_1];
542 PPC_STB(src_reg, dst_reg, off);
543 break;
544 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
545 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
546 if (BPF_CLASS(code) == BPF_ST) {
547 PPC_LI(b2p[TMP_REG_1], imm);
548 src_reg = b2p[TMP_REG_1];
550 PPC_STH(src_reg, dst_reg, off);
551 break;
552 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
553 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
554 if (BPF_CLASS(code) == BPF_ST) {
555 PPC_LI32(b2p[TMP_REG_1], imm);
556 src_reg = b2p[TMP_REG_1];
558 PPC_STW(src_reg, dst_reg, off);
559 break;
560 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
561 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
562 if (BPF_CLASS(code) == BPF_ST) {
563 PPC_LI32(b2p[TMP_REG_1], imm);
564 src_reg = b2p[TMP_REG_1];
566 PPC_STD(src_reg, dst_reg, off);
567 break;
570 * BPF_STX XADD (atomic_add)
572 /* *(u32 *)(dst + off) += src */
573 case BPF_STX | BPF_XADD | BPF_W:
574 /* Get EA into TMP_REG_1 */
575 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
576 /* error if EA is not word-aligned */
577 PPC_ANDI(b2p[TMP_REG_2], b2p[TMP_REG_1], 0x03);
578 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + 12);
579 PPC_LI(b2p[BPF_REG_0], 0);
580 PPC_JMP(exit_addr);
581 /* load value from memory into TMP_REG_2 */
582 PPC_BPF_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
583 /* add value from src_reg into this */
584 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
585 /* store result back */
586 PPC_BPF_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
587 /* we're done if this succeeded */
588 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + (7*4));
589 /* otherwise, let's try once more */
590 PPC_BPF_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
591 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
592 PPC_BPF_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
593 /* exit if the store was not successful */
594 PPC_LI(b2p[BPF_REG_0], 0);
595 PPC_BCC(COND_NE, exit_addr);
596 break;
597 /* *(u64 *)(dst + off) += src */
598 case BPF_STX | BPF_XADD | BPF_DW:
599 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
600 /* error if EA is not doubleword-aligned */
601 PPC_ANDI(b2p[TMP_REG_2], b2p[TMP_REG_1], 0x07);
602 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + (3*4));
603 PPC_LI(b2p[BPF_REG_0], 0);
604 PPC_JMP(exit_addr);
605 PPC_BPF_LDARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
606 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
607 PPC_BPF_STDCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
608 PPC_BCC_SHORT(COND_EQ, (ctx->idx * 4) + (7*4));
609 PPC_BPF_LDARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
610 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
611 PPC_BPF_STDCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
612 PPC_LI(b2p[BPF_REG_0], 0);
613 PPC_BCC(COND_NE, exit_addr);
614 break;
617 * BPF_LDX
619 /* dst = *(u8 *)(ul) (src + off) */
620 case BPF_LDX | BPF_MEM | BPF_B:
621 PPC_LBZ(dst_reg, src_reg, off);
622 break;
623 /* dst = *(u16 *)(ul) (src + off) */
624 case BPF_LDX | BPF_MEM | BPF_H:
625 PPC_LHZ(dst_reg, src_reg, off);
626 break;
627 /* dst = *(u32 *)(ul) (src + off) */
628 case BPF_LDX | BPF_MEM | BPF_W:
629 PPC_LWZ(dst_reg, src_reg, off);
630 break;
631 /* dst = *(u64 *)(ul) (src + off) */
632 case BPF_LDX | BPF_MEM | BPF_DW:
633 PPC_LD(dst_reg, src_reg, off);
634 break;
637 * Doubleword load
638 * 16 byte instruction that uses two 'struct bpf_insn'
640 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
641 imm64 = ((u64)(u32) insn[i].imm) |
642 (((u64)(u32) insn[i+1].imm) << 32);
643 /* Adjust for two bpf instructions */
644 addrs[++i] = ctx->idx * 4;
645 PPC_LI64(dst_reg, imm64);
646 break;
649 * Return/Exit
651 case BPF_JMP | BPF_EXIT:
653 * If this isn't the very last instruction, branch to
654 * the epilogue. If we _are_ the last instruction,
655 * we'll just fall through to the epilogue.
657 if (i != flen - 1)
658 PPC_JMP(exit_addr);
659 /* else fall through to the epilogue */
660 break;
663 * Call kernel helper
665 case BPF_JMP | BPF_CALL:
666 ctx->seen |= SEEN_FUNC;
667 func = (u8 *) __bpf_call_base + imm;
669 /* Save skb pointer if we need to re-cache skb data */
670 if (bpf_helper_changes_skb_data(func))
671 PPC_BPF_STL(3, 1, STACK_FRAME_MIN_SIZE);
673 bpf_jit_emit_func_call(image, ctx, (u64)func);
675 /* move return value from r3 to BPF_REG_0 */
676 PPC_MR(b2p[BPF_REG_0], 3);
678 /* refresh skb cache */
679 if (bpf_helper_changes_skb_data(func)) {
680 /* reload skb pointer to r3 */
681 PPC_BPF_LL(3, 1, STACK_FRAME_MIN_SIZE);
682 bpf_jit_emit_skb_loads(image, ctx);
684 break;
687 * Jumps and branches
689 case BPF_JMP | BPF_JA:
690 PPC_JMP(addrs[i + 1 + off]);
691 break;
693 case BPF_JMP | BPF_JGT | BPF_K:
694 case BPF_JMP | BPF_JGT | BPF_X:
695 case BPF_JMP | BPF_JSGT | BPF_K:
696 case BPF_JMP | BPF_JSGT | BPF_X:
697 true_cond = COND_GT;
698 goto cond_branch;
699 case BPF_JMP | BPF_JGE | BPF_K:
700 case BPF_JMP | BPF_JGE | BPF_X:
701 case BPF_JMP | BPF_JSGE | BPF_K:
702 case BPF_JMP | BPF_JSGE | BPF_X:
703 true_cond = COND_GE;
704 goto cond_branch;
705 case BPF_JMP | BPF_JEQ | BPF_K:
706 case BPF_JMP | BPF_JEQ | BPF_X:
707 true_cond = COND_EQ;
708 goto cond_branch;
709 case BPF_JMP | BPF_JNE | BPF_K:
710 case BPF_JMP | BPF_JNE | BPF_X:
711 true_cond = COND_NE;
712 goto cond_branch;
713 case BPF_JMP | BPF_JSET | BPF_K:
714 case BPF_JMP | BPF_JSET | BPF_X:
715 true_cond = COND_NE;
716 /* Fall through */
718 cond_branch:
719 switch (code) {
720 case BPF_JMP | BPF_JGT | BPF_X:
721 case BPF_JMP | BPF_JGE | BPF_X:
722 case BPF_JMP | BPF_JEQ | BPF_X:
723 case BPF_JMP | BPF_JNE | BPF_X:
724 /* unsigned comparison */
725 PPC_CMPLD(dst_reg, src_reg);
726 break;
727 case BPF_JMP | BPF_JSGT | BPF_X:
728 case BPF_JMP | BPF_JSGE | BPF_X:
729 /* signed comparison */
730 PPC_CMPD(dst_reg, src_reg);
731 break;
732 case BPF_JMP | BPF_JSET | BPF_X:
733 PPC_AND_DOT(b2p[TMP_REG_1], dst_reg, src_reg);
734 break;
735 case BPF_JMP | BPF_JNE | BPF_K:
736 case BPF_JMP | BPF_JEQ | BPF_K:
737 case BPF_JMP | BPF_JGT | BPF_K:
738 case BPF_JMP | BPF_JGE | BPF_K:
740 * Need sign-extended load, so only positive
741 * values can be used as imm in cmpldi
743 if (imm >= 0 && imm < 32768)
744 PPC_CMPLDI(dst_reg, imm);
745 else {
746 /* sign-extending load */
747 PPC_LI32(b2p[TMP_REG_1], imm);
748 /* ... but unsigned comparison */
749 PPC_CMPLD(dst_reg, b2p[TMP_REG_1]);
751 break;
752 case BPF_JMP | BPF_JSGT | BPF_K:
753 case BPF_JMP | BPF_JSGE | BPF_K:
755 * signed comparison, so any 16-bit value
756 * can be used in cmpdi
758 if (imm >= -32768 && imm < 32768)
759 PPC_CMPDI(dst_reg, imm);
760 else {
761 PPC_LI32(b2p[TMP_REG_1], imm);
762 PPC_CMPD(dst_reg, b2p[TMP_REG_1]);
764 break;
765 case BPF_JMP | BPF_JSET | BPF_K:
766 /* andi does not sign-extend the immediate */
767 if (imm >= 0 && imm < 32768)
768 /* PPC_ANDI is _only/always_ dot-form */
769 PPC_ANDI(b2p[TMP_REG_1], dst_reg, imm);
770 else {
771 PPC_LI32(b2p[TMP_REG_1], imm);
772 PPC_AND_DOT(b2p[TMP_REG_1], dst_reg,
773 b2p[TMP_REG_1]);
775 break;
777 PPC_BCC(true_cond, addrs[i + 1 + off]);
778 break;
781 * Loads from packet header/data
782 * Assume 32-bit input value in imm and X (src_reg)
785 /* Absolute loads */
786 case BPF_LD | BPF_W | BPF_ABS:
787 func = (u8 *)CHOOSE_LOAD_FUNC(imm, sk_load_word);
788 goto common_load_abs;
789 case BPF_LD | BPF_H | BPF_ABS:
790 func = (u8 *)CHOOSE_LOAD_FUNC(imm, sk_load_half);
791 goto common_load_abs;
792 case BPF_LD | BPF_B | BPF_ABS:
793 func = (u8 *)CHOOSE_LOAD_FUNC(imm, sk_load_byte);
794 common_load_abs:
796 * Load from [imm]
797 * Load into r4, which can just be passed onto
798 * skb load helpers as the second parameter
800 PPC_LI32(4, imm);
801 goto common_load;
803 /* Indirect loads */
804 case BPF_LD | BPF_W | BPF_IND:
805 func = (u8 *)sk_load_word;
806 goto common_load_ind;
807 case BPF_LD | BPF_H | BPF_IND:
808 func = (u8 *)sk_load_half;
809 goto common_load_ind;
810 case BPF_LD | BPF_B | BPF_IND:
811 func = (u8 *)sk_load_byte;
812 common_load_ind:
814 * Load from [src_reg + imm]
815 * Treat src_reg as a 32-bit value
817 PPC_EXTSW(4, src_reg);
818 if (imm) {
819 if (imm >= -32768 && imm < 32768)
820 PPC_ADDI(4, 4, IMM_L(imm));
821 else {
822 PPC_LI32(b2p[TMP_REG_1], imm);
823 PPC_ADD(4, 4, b2p[TMP_REG_1]);
827 common_load:
828 ctx->seen |= SEEN_SKB;
829 ctx->seen |= SEEN_FUNC;
830 bpf_jit_emit_func_call(image, ctx, (u64)func);
833 * Helper returns 'lt' condition on error, and an
834 * appropriate return value in BPF_REG_0
836 PPC_BCC(COND_LT, exit_addr);
837 break;
840 * TODO: Tail call
842 case BPF_JMP | BPF_CALL | BPF_X:
844 default:
846 * The filter contains something cruel & unusual.
847 * We don't handle it, but also there shouldn't be
848 * anything missing from our list.
850 pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n",
851 code, i);
852 return -ENOTSUPP;
856 /* Set end-of-body-code address for exit. */
857 addrs[i] = ctx->idx * 4;
859 return 0;
862 void bpf_jit_compile(struct bpf_prog *fp) { }
864 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
866 u32 proglen;
867 u32 alloclen;
868 u8 *image = NULL;
869 u32 *code_base;
870 u32 *addrs;
871 struct codegen_context cgctx;
872 int pass;
873 int flen;
874 struct bpf_binary_header *bpf_hdr;
876 if (!bpf_jit_enable)
877 return fp;
879 flen = fp->len;
880 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
881 if (addrs == NULL)
882 return fp;
884 cgctx.idx = 0;
885 cgctx.seen = 0;
886 /* Scouting faux-generate pass 0 */
887 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
888 /* We hit something illegal or unsupported. */
889 goto out;
892 * Pretend to build prologue, given the features we've seen. This will
893 * update ctgtx.idx as it pretends to output instructions, then we can
894 * calculate total size from idx.
896 bpf_jit_build_prologue(0, &cgctx);
897 bpf_jit_build_epilogue(0, &cgctx);
899 proglen = cgctx.idx * 4;
900 alloclen = proglen + FUNCTION_DESCR_SIZE;
902 bpf_hdr = bpf_jit_binary_alloc(alloclen, &image, 4,
903 bpf_jit_fill_ill_insns);
904 if (!bpf_hdr)
905 goto out;
907 code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
909 /* Code generation passes 1-2 */
910 for (pass = 1; pass < 3; pass++) {
911 /* Now build the prologue, body code & epilogue for real. */
912 cgctx.idx = 0;
913 bpf_jit_build_prologue(code_base, &cgctx);
914 bpf_jit_build_body(fp, code_base, &cgctx, addrs);
915 bpf_jit_build_epilogue(code_base, &cgctx);
917 if (bpf_jit_enable > 1)
918 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
919 proglen - (cgctx.idx * 4), cgctx.seen);
922 if (bpf_jit_enable > 1)
924 * Note that we output the base address of the code_base
925 * rather than image, since opcodes are in code_base.
927 bpf_jit_dump(flen, proglen, pass, code_base);
929 if (image) {
930 bpf_flush_icache(bpf_hdr, image + alloclen);
931 #ifdef PPC64_ELF_ABI_v1
932 /* Function descriptor nastiness: Address + TOC */
933 ((u64 *)image)[0] = (u64)code_base;
934 ((u64 *)image)[1] = local_paca->kernel_toc;
935 #endif
936 fp->bpf_func = (void *)image;
937 fp->jited = 1;
940 out:
941 kfree(addrs);
942 return fp;
945 void bpf_jit_free(struct bpf_prog *fp)
947 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
948 struct bpf_binary_header *bpf_hdr = (void *)addr;
950 if (fp->jited)
951 bpf_jit_binary_free(bpf_hdr);
953 bpf_prog_unlock_free(fp);