1 // SPDX-License-Identifier: GPL-2.0-only
3 * bpf_jit_comp64.c: eBPF JIT compiler
5 * Copyright 2016 Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
8 * Based on the powerpc classic BPF JIT compiler by Matt Evans
10 #include <linux/moduleloader.h>
11 #include <asm/cacheflush.h>
12 #include <asm/asm-compat.h>
13 #include <linux/netdevice.h>
14 #include <linux/filter.h>
15 #include <linux/if_vlan.h>
16 #include <asm/kprobes.h>
17 #include <linux/bpf.h>
19 #include "bpf_jit64.h"
21 static void bpf_jit_fill_ill_insns(void *area, unsigned int size)
23 memset32(area, BREAKPOINT_INSTRUCTION, size/4);
26 static inline void bpf_flush_icache(void *start, void *end)
29 flush_icache_range((unsigned long)start, (unsigned long)end);
32 static inline bool bpf_is_seen_register(struct codegen_context *ctx, int i)
34 return (ctx->seen & (1 << (31 - b2p[i])));
37 static inline void bpf_set_seen_register(struct codegen_context *ctx, int i)
39 ctx->seen |= (1 << (31 - b2p[i]));
42 static inline bool bpf_has_stack_frame(struct codegen_context *ctx)
45 * We only need a stack frame if:
46 * - we call other functions (kernel helpers), or
47 * - the bpf program uses its stack area
48 * The latter condition is deduced from the usage of BPF_REG_FP
50 return ctx->seen & SEEN_FUNC || bpf_is_seen_register(ctx, BPF_REG_FP);
54 * When not setting up our own stackframe, the redzone usage is:
56 * [ prev sp ] <-------------
58 * sp (r1) ---> [ stack pointer ] --------------
59 * [ nv gpr save area ] 6*8
62 * [ unused red zone ] 208 bytes protected
64 static int bpf_jit_stack_local(struct codegen_context *ctx)
66 if (bpf_has_stack_frame(ctx))
67 return STACK_FRAME_MIN_SIZE + ctx->stack_size;
69 return -(BPF_PPC_STACK_SAVE + 16);
72 static int bpf_jit_stack_tailcallcnt(struct codegen_context *ctx)
74 return bpf_jit_stack_local(ctx) + 8;
77 static int bpf_jit_stack_offsetof(struct codegen_context *ctx, int reg)
79 if (reg >= BPF_PPC_NVR_MIN && reg < 32)
80 return (bpf_has_stack_frame(ctx) ?
81 (BPF_PPC_STACKFRAME + ctx->stack_size) : 0)
84 pr_err("BPF JIT is asking about unknown registers");
88 static void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx)
93 * Initialize tail_call_cnt if we do tail calls.
94 * Otherwise, put in NOPs so that it can be skipped when we are
95 * invoked through a tail call.
97 if (ctx->seen & SEEN_TAILCALL) {
98 PPC_LI(b2p[TMP_REG_1], 0);
99 /* this goes in the redzone */
100 PPC_BPF_STL(b2p[TMP_REG_1], 1, -(BPF_PPC_STACK_SAVE + 8));
106 #define BPF_TAILCALL_PROLOGUE_SIZE 8
108 if (bpf_has_stack_frame(ctx)) {
110 * We need a stack frame, but we don't necessarily need to
111 * save/restore LR unless we call other functions
113 if (ctx->seen & SEEN_FUNC) {
114 EMIT(PPC_INST_MFLR | __PPC_RT(R0));
115 PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
118 PPC_BPF_STLU(1, 1, -(BPF_PPC_STACKFRAME + ctx->stack_size));
122 * Back up non-volatile regs -- BPF registers 6-10
123 * If we haven't created our own stack frame, we save these
124 * in the protected zone below the previous stack frame
126 for (i = BPF_REG_6; i <= BPF_REG_10; i++)
127 if (bpf_is_seen_register(ctx, i))
128 PPC_BPF_STL(b2p[i], 1, bpf_jit_stack_offsetof(ctx, b2p[i]));
130 /* Setup frame pointer to point to the bpf stack area */
131 if (bpf_is_seen_register(ctx, BPF_REG_FP))
132 PPC_ADDI(b2p[BPF_REG_FP], 1,
133 STACK_FRAME_MIN_SIZE + ctx->stack_size);
136 static void bpf_jit_emit_common_epilogue(u32 *image, struct codegen_context *ctx)
141 for (i = BPF_REG_6; i <= BPF_REG_10; i++)
142 if (bpf_is_seen_register(ctx, i))
143 PPC_BPF_LL(b2p[i], 1, bpf_jit_stack_offsetof(ctx, b2p[i]));
145 /* Tear down our stack frame */
146 if (bpf_has_stack_frame(ctx)) {
147 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME + ctx->stack_size);
148 if (ctx->seen & SEEN_FUNC) {
149 PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
155 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
157 bpf_jit_emit_common_epilogue(image, ctx);
159 /* Move result to r3 */
160 PPC_MR(3, b2p[BPF_REG_0]);
165 static void bpf_jit_emit_func_call_hlp(u32 *image, struct codegen_context *ctx,
168 #ifdef PPC64_ELF_ABI_v1
169 /* func points to the function descriptor */
170 PPC_LI64(b2p[TMP_REG_2], func);
171 /* Load actual entry point from function descriptor */
172 PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_2], 0);
173 /* ... and move it to LR */
174 PPC_MTLR(b2p[TMP_REG_1]);
176 * Load TOC from function descriptor at offset 8.
177 * We can clobber r2 since we get called through a
178 * function pointer (so caller will save/restore r2)
179 * and since we don't use a TOC ourself.
181 PPC_BPF_LL(2, b2p[TMP_REG_2], 8);
183 /* We can clobber r12 */
184 PPC_FUNC_ADDR(12, func);
190 static void bpf_jit_emit_func_call_rel(u32 *image, struct codegen_context *ctx,
193 unsigned int i, ctx_idx = ctx->idx;
195 /* Load function address into r12 */
198 /* For bpf-to-bpf function calls, the callee's address is unknown
199 * until the last extra pass. As seen above, we use PPC_LI64() to
200 * load the callee's address, but this may optimize the number of
201 * instructions required based on the nature of the address.
203 * Since we don't want the number of instructions emitted to change,
204 * we pad the optimized PPC_LI64() call with NOPs to guarantee that
205 * we always have a five-instruction sequence, which is the maximum
206 * that PPC_LI64() can emit.
208 for (i = ctx->idx - ctx_idx; i < 5; i++)
211 #ifdef PPC64_ELF_ABI_v1
213 * Load TOC from function descriptor at offset 8.
214 * We can clobber r2 since we get called through a
215 * function pointer (so caller will save/restore r2)
216 * and since we don't use a TOC ourself.
218 PPC_BPF_LL(2, 12, 8);
219 /* Load actual entry point from function descriptor */
220 PPC_BPF_LL(12, 12, 0);
227 static void bpf_jit_emit_tail_call(u32 *image, struct codegen_context *ctx, u32 out)
230 * By now, the eBPF program has already setup parameters in r3, r4 and r5
231 * r3/BPF_REG_1 - pointer to ctx -- passed as is to the next bpf program
232 * r4/BPF_REG_2 - pointer to bpf_array
233 * r5/BPF_REG_3 - index in bpf_array
235 int b2p_bpf_array = b2p[BPF_REG_2];
236 int b2p_index = b2p[BPF_REG_3];
239 * if (index >= array->map.max_entries)
242 PPC_LWZ(b2p[TMP_REG_1], b2p_bpf_array, offsetof(struct bpf_array, map.max_entries));
243 PPC_RLWINM(b2p_index, b2p_index, 0, 0, 31);
244 PPC_CMPLW(b2p_index, b2p[TMP_REG_1]);
245 PPC_BCC(COND_GE, out);
248 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
251 PPC_BPF_LL(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
252 PPC_CMPLWI(b2p[TMP_REG_1], MAX_TAIL_CALL_CNT);
253 PPC_BCC(COND_GT, out);
258 PPC_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], 1);
259 PPC_BPF_STL(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
261 /* prog = array->ptrs[index]; */
262 PPC_MULI(b2p[TMP_REG_1], b2p_index, 8);
263 PPC_ADD(b2p[TMP_REG_1], b2p[TMP_REG_1], b2p_bpf_array);
264 PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_1], offsetof(struct bpf_array, ptrs));
270 PPC_CMPLDI(b2p[TMP_REG_1], 0);
271 PPC_BCC(COND_EQ, out);
273 /* goto *(prog->bpf_func + prologue_size); */
274 PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_1], offsetof(struct bpf_prog, bpf_func));
275 #ifdef PPC64_ELF_ABI_v1
276 /* skip past the function descriptor */
277 PPC_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1],
278 FUNCTION_DESCR_SIZE + BPF_TAILCALL_PROLOGUE_SIZE);
280 PPC_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], BPF_TAILCALL_PROLOGUE_SIZE);
282 PPC_MTCTR(b2p[TMP_REG_1]);
284 /* tear down stack, restore NVRs, ... */
285 bpf_jit_emit_common_epilogue(image, ctx);
291 /* Assemble the body code between the prologue & epilogue */
292 static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
293 struct codegen_context *ctx,
294 u32 *addrs, bool extra_pass)
296 const struct bpf_insn *insn = fp->insnsi;
300 /* Start of epilogue code - will only be valid 2nd pass onwards */
301 u32 exit_addr = addrs[flen];
303 for (i = 0; i < flen; i++) {
304 u32 code = insn[i].code;
305 u32 dst_reg = b2p[insn[i].dst_reg];
306 u32 src_reg = b2p[insn[i].src_reg];
307 s16 off = insn[i].off;
308 s32 imm = insn[i].imm;
309 bool func_addr_fixed;
316 * addrs[] maps a BPF bytecode address into a real offset from
317 * the start of the body code.
319 addrs[i] = ctx->idx * 4;
322 * As an optimization, we note down which non-volatile registers
323 * are used so that we can only save/restore those in our
324 * prologue and epilogue. We do this here regardless of whether
325 * the actual BPF instruction uses src/dst registers or not
326 * (for instance, BPF_CALL does not use them). The expectation
327 * is that those instructions will have src_reg/dst_reg set to
328 * 0. Even otherwise, we just lose some prologue/epilogue
329 * optimization but everything else should work without
332 if (dst_reg >= BPF_PPC_NVR_MIN && dst_reg < 32)
333 bpf_set_seen_register(ctx, insn[i].dst_reg);
334 if (src_reg >= BPF_PPC_NVR_MIN && src_reg < 32)
335 bpf_set_seen_register(ctx, insn[i].src_reg);
339 * Arithmetic operations: ADD/SUB/MUL/DIV/MOD/NEG
341 case BPF_ALU | BPF_ADD | BPF_X: /* (u32) dst += (u32) src */
342 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst += src */
343 PPC_ADD(dst_reg, dst_reg, src_reg);
344 goto bpf_alu32_trunc;
345 case BPF_ALU | BPF_SUB | BPF_X: /* (u32) dst -= (u32) src */
346 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst -= src */
347 PPC_SUB(dst_reg, dst_reg, src_reg);
348 goto bpf_alu32_trunc;
349 case BPF_ALU | BPF_ADD | BPF_K: /* (u32) dst += (u32) imm */
350 case BPF_ALU | BPF_SUB | BPF_K: /* (u32) dst -= (u32) imm */
351 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst += imm */
352 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst -= imm */
353 if (BPF_OP(code) == BPF_SUB)
356 if (imm >= -32768 && imm < 32768)
357 PPC_ADDI(dst_reg, dst_reg, IMM_L(imm));
359 PPC_LI32(b2p[TMP_REG_1], imm);
360 PPC_ADD(dst_reg, dst_reg, b2p[TMP_REG_1]);
363 goto bpf_alu32_trunc;
364 case BPF_ALU | BPF_MUL | BPF_X: /* (u32) dst *= (u32) src */
365 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst *= src */
366 if (BPF_CLASS(code) == BPF_ALU)
367 PPC_MULW(dst_reg, dst_reg, src_reg);
369 PPC_MULD(dst_reg, dst_reg, src_reg);
370 goto bpf_alu32_trunc;
371 case BPF_ALU | BPF_MUL | BPF_K: /* (u32) dst *= (u32) imm */
372 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst *= imm */
373 if (imm >= -32768 && imm < 32768)
374 PPC_MULI(dst_reg, dst_reg, IMM_L(imm));
376 PPC_LI32(b2p[TMP_REG_1], imm);
377 if (BPF_CLASS(code) == BPF_ALU)
378 PPC_MULW(dst_reg, dst_reg,
381 PPC_MULD(dst_reg, dst_reg,
384 goto bpf_alu32_trunc;
385 case BPF_ALU | BPF_DIV | BPF_X: /* (u32) dst /= (u32) src */
386 case BPF_ALU | BPF_MOD | BPF_X: /* (u32) dst %= (u32) src */
387 if (BPF_OP(code) == BPF_MOD) {
388 PPC_DIVWU(b2p[TMP_REG_1], dst_reg, src_reg);
389 PPC_MULW(b2p[TMP_REG_1], src_reg,
391 PPC_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]);
393 PPC_DIVWU(dst_reg, dst_reg, src_reg);
394 goto bpf_alu32_trunc;
395 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst /= src */
396 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst %= src */
397 if (BPF_OP(code) == BPF_MOD) {
398 PPC_DIVDU(b2p[TMP_REG_1], dst_reg, src_reg);
399 PPC_MULD(b2p[TMP_REG_1], src_reg,
401 PPC_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]);
403 PPC_DIVDU(dst_reg, dst_reg, src_reg);
405 case BPF_ALU | BPF_MOD | BPF_K: /* (u32) dst %= (u32) imm */
406 case BPF_ALU | BPF_DIV | BPF_K: /* (u32) dst /= (u32) imm */
407 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst %= imm */
408 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst /= imm */
412 goto bpf_alu32_trunc;
414 PPC_LI32(b2p[TMP_REG_1], imm);
415 switch (BPF_CLASS(code)) {
417 if (BPF_OP(code) == BPF_MOD) {
418 PPC_DIVWU(b2p[TMP_REG_2], dst_reg,
420 PPC_MULW(b2p[TMP_REG_1],
423 PPC_SUB(dst_reg, dst_reg,
426 PPC_DIVWU(dst_reg, dst_reg,
430 if (BPF_OP(code) == BPF_MOD) {
431 PPC_DIVDU(b2p[TMP_REG_2], dst_reg,
433 PPC_MULD(b2p[TMP_REG_1],
436 PPC_SUB(dst_reg, dst_reg,
439 PPC_DIVDU(dst_reg, dst_reg,
443 goto bpf_alu32_trunc;
444 case BPF_ALU | BPF_NEG: /* (u32) dst = -dst */
445 case BPF_ALU64 | BPF_NEG: /* dst = -dst */
446 PPC_NEG(dst_reg, dst_reg);
447 goto bpf_alu32_trunc;
450 * Logical operations: AND/OR/XOR/[A]LSH/[A]RSH
452 case BPF_ALU | BPF_AND | BPF_X: /* (u32) dst = dst & src */
453 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
454 PPC_AND(dst_reg, dst_reg, src_reg);
455 goto bpf_alu32_trunc;
456 case BPF_ALU | BPF_AND | BPF_K: /* (u32) dst = dst & imm */
457 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
459 PPC_ANDI(dst_reg, dst_reg, IMM_L(imm));
462 PPC_LI32(b2p[TMP_REG_1], imm);
463 PPC_AND(dst_reg, dst_reg, b2p[TMP_REG_1]);
465 goto bpf_alu32_trunc;
466 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
467 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
468 PPC_OR(dst_reg, dst_reg, src_reg);
469 goto bpf_alu32_trunc;
470 case BPF_ALU | BPF_OR | BPF_K:/* dst = (u32) dst | (u32) imm */
471 case BPF_ALU64 | BPF_OR | BPF_K:/* dst = dst | imm */
472 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
474 PPC_LI32(b2p[TMP_REG_1], imm);
475 PPC_OR(dst_reg, dst_reg, b2p[TMP_REG_1]);
478 PPC_ORI(dst_reg, dst_reg, IMM_L(imm));
480 PPC_ORIS(dst_reg, dst_reg, IMM_H(imm));
482 goto bpf_alu32_trunc;
483 case BPF_ALU | BPF_XOR | BPF_X: /* (u32) dst ^= src */
484 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst ^= src */
485 PPC_XOR(dst_reg, dst_reg, src_reg);
486 goto bpf_alu32_trunc;
487 case BPF_ALU | BPF_XOR | BPF_K: /* (u32) dst ^= (u32) imm */
488 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst ^= imm */
489 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
491 PPC_LI32(b2p[TMP_REG_1], imm);
492 PPC_XOR(dst_reg, dst_reg, b2p[TMP_REG_1]);
495 PPC_XORI(dst_reg, dst_reg, IMM_L(imm));
497 PPC_XORIS(dst_reg, dst_reg, IMM_H(imm));
499 goto bpf_alu32_trunc;
500 case BPF_ALU | BPF_LSH | BPF_X: /* (u32) dst <<= (u32) src */
501 /* slw clears top 32 bits */
502 PPC_SLW(dst_reg, dst_reg, src_reg);
504 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */
505 PPC_SLD(dst_reg, dst_reg, src_reg);
507 case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<== (u32) imm */
508 /* with imm 0, we still need to clear top 32 bits */
509 PPC_SLWI(dst_reg, dst_reg, imm);
511 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<== imm */
513 PPC_SLDI(dst_reg, dst_reg, imm);
515 case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */
516 PPC_SRW(dst_reg, dst_reg, src_reg);
518 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */
519 PPC_SRD(dst_reg, dst_reg, src_reg);
521 case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */
522 PPC_SRWI(dst_reg, dst_reg, imm);
524 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */
526 PPC_SRDI(dst_reg, dst_reg, imm);
528 case BPF_ALU | BPF_ARSH | BPF_X: /* (s32) dst >>= src */
529 PPC_SRAW(dst_reg, dst_reg, src_reg);
530 goto bpf_alu32_trunc;
531 case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */
532 PPC_SRAD(dst_reg, dst_reg, src_reg);
534 case BPF_ALU | BPF_ARSH | BPF_K: /* (s32) dst >>= imm */
535 PPC_SRAWI(dst_reg, dst_reg, imm);
536 goto bpf_alu32_trunc;
537 case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */
539 PPC_SRADI(dst_reg, dst_reg, imm);
545 case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
546 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
547 PPC_MR(dst_reg, src_reg);
548 goto bpf_alu32_trunc;
549 case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */
550 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */
551 PPC_LI32(dst_reg, imm);
553 goto bpf_alu32_trunc;
557 /* Truncate to 32-bits */
558 if (BPF_CLASS(code) == BPF_ALU)
559 PPC_RLWINM(dst_reg, dst_reg, 0, 0, 31);
565 case BPF_ALU | BPF_END | BPF_FROM_LE:
566 case BPF_ALU | BPF_END | BPF_FROM_BE:
567 #ifdef __BIG_ENDIAN__
568 if (BPF_SRC(code) == BPF_FROM_BE)
570 #else /* !__BIG_ENDIAN__ */
571 if (BPF_SRC(code) == BPF_FROM_LE)
576 /* Rotate 8 bits left & mask with 0x0000ff00 */
577 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 16, 23);
578 /* Rotate 8 bits right & insert LSB to reg */
579 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 24, 31);
580 /* Move result back to dst_reg */
581 PPC_MR(dst_reg, b2p[TMP_REG_1]);
585 * Rotate word left by 8 bits:
586 * 2 bytes are already in their final position
587 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4)
589 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 0, 31);
590 /* Rotate 24 bits and insert byte 1 */
591 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 0, 7);
592 /* Rotate 24 bits and insert byte 3 */
593 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 16, 23);
594 PPC_MR(dst_reg, b2p[TMP_REG_1]);
598 * Way easier and faster(?) to store the value
599 * into stack and then use ldbrx
601 * ctx->seen will be reliable in pass2, but
602 * the instructions generated will remain the
603 * same across all passes
605 PPC_BPF_STL(dst_reg, 1, bpf_jit_stack_local(ctx));
606 PPC_ADDI(b2p[TMP_REG_1], 1, bpf_jit_stack_local(ctx));
607 PPC_LDBRX(dst_reg, 0, b2p[TMP_REG_1]);
615 /* zero-extend 16 bits into 64 bits */
616 PPC_RLDICL(dst_reg, dst_reg, 0, 48);
619 /* zero-extend 32 bits into 64 bits */
620 PPC_RLDICL(dst_reg, dst_reg, 0, 32);
631 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */
632 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
633 if (BPF_CLASS(code) == BPF_ST) {
634 PPC_LI(b2p[TMP_REG_1], imm);
635 src_reg = b2p[TMP_REG_1];
637 PPC_STB(src_reg, dst_reg, off);
639 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
640 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
641 if (BPF_CLASS(code) == BPF_ST) {
642 PPC_LI(b2p[TMP_REG_1], imm);
643 src_reg = b2p[TMP_REG_1];
645 PPC_STH(src_reg, dst_reg, off);
647 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
648 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
649 if (BPF_CLASS(code) == BPF_ST) {
650 PPC_LI32(b2p[TMP_REG_1], imm);
651 src_reg = b2p[TMP_REG_1];
653 PPC_STW(src_reg, dst_reg, off);
655 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
656 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
657 if (BPF_CLASS(code) == BPF_ST) {
658 PPC_LI32(b2p[TMP_REG_1], imm);
659 src_reg = b2p[TMP_REG_1];
661 PPC_BPF_STL(src_reg, dst_reg, off);
665 * BPF_STX XADD (atomic_add)
667 /* *(u32 *)(dst + off) += src */
668 case BPF_STX | BPF_XADD | BPF_W:
669 /* Get EA into TMP_REG_1 */
670 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
671 tmp_idx = ctx->idx * 4;
672 /* load value from memory into TMP_REG_2 */
673 PPC_BPF_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
674 /* add value from src_reg into this */
675 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
676 /* store result back */
677 PPC_BPF_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
678 /* we're done if this succeeded */
679 PPC_BCC_SHORT(COND_NE, tmp_idx);
681 /* *(u64 *)(dst + off) += src */
682 case BPF_STX | BPF_XADD | BPF_DW:
683 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
684 tmp_idx = ctx->idx * 4;
685 PPC_BPF_LDARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
686 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
687 PPC_BPF_STDCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
688 PPC_BCC_SHORT(COND_NE, tmp_idx);
694 /* dst = *(u8 *)(ul) (src + off) */
695 case BPF_LDX | BPF_MEM | BPF_B:
696 PPC_LBZ(dst_reg, src_reg, off);
698 /* dst = *(u16 *)(ul) (src + off) */
699 case BPF_LDX | BPF_MEM | BPF_H:
700 PPC_LHZ(dst_reg, src_reg, off);
702 /* dst = *(u32 *)(ul) (src + off) */
703 case BPF_LDX | BPF_MEM | BPF_W:
704 PPC_LWZ(dst_reg, src_reg, off);
706 /* dst = *(u64 *)(ul) (src + off) */
707 case BPF_LDX | BPF_MEM | BPF_DW:
708 PPC_BPF_LL(dst_reg, src_reg, off);
713 * 16 byte instruction that uses two 'struct bpf_insn'
715 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
716 imm64 = ((u64)(u32) insn[i].imm) |
717 (((u64)(u32) insn[i+1].imm) << 32);
718 /* Adjust for two bpf instructions */
719 addrs[++i] = ctx->idx * 4;
720 PPC_LI64(dst_reg, imm64);
726 case BPF_JMP | BPF_EXIT:
728 * If this isn't the very last instruction, branch to
729 * the epilogue. If we _are_ the last instruction,
730 * we'll just fall through to the epilogue.
734 /* else fall through to the epilogue */
738 * Call kernel helper or bpf function
740 case BPF_JMP | BPF_CALL:
741 ctx->seen |= SEEN_FUNC;
743 ret = bpf_jit_get_func_addr(fp, &insn[i], extra_pass,
744 &func_addr, &func_addr_fixed);
749 bpf_jit_emit_func_call_hlp(image, ctx, func_addr);
751 bpf_jit_emit_func_call_rel(image, ctx, func_addr);
752 /* move return value from r3 to BPF_REG_0 */
753 PPC_MR(b2p[BPF_REG_0], 3);
759 case BPF_JMP | BPF_JA:
760 PPC_JMP(addrs[i + 1 + off]);
763 case BPF_JMP | BPF_JGT | BPF_K:
764 case BPF_JMP | BPF_JGT | BPF_X:
765 case BPF_JMP | BPF_JSGT | BPF_K:
766 case BPF_JMP | BPF_JSGT | BPF_X:
767 case BPF_JMP32 | BPF_JGT | BPF_K:
768 case BPF_JMP32 | BPF_JGT | BPF_X:
769 case BPF_JMP32 | BPF_JSGT | BPF_K:
770 case BPF_JMP32 | BPF_JSGT | BPF_X:
773 case BPF_JMP | BPF_JLT | BPF_K:
774 case BPF_JMP | BPF_JLT | BPF_X:
775 case BPF_JMP | BPF_JSLT | BPF_K:
776 case BPF_JMP | BPF_JSLT | BPF_X:
777 case BPF_JMP32 | BPF_JLT | BPF_K:
778 case BPF_JMP32 | BPF_JLT | BPF_X:
779 case BPF_JMP32 | BPF_JSLT | BPF_K:
780 case BPF_JMP32 | BPF_JSLT | BPF_X:
783 case BPF_JMP | BPF_JGE | BPF_K:
784 case BPF_JMP | BPF_JGE | BPF_X:
785 case BPF_JMP | BPF_JSGE | BPF_K:
786 case BPF_JMP | BPF_JSGE | BPF_X:
787 case BPF_JMP32 | BPF_JGE | BPF_K:
788 case BPF_JMP32 | BPF_JGE | BPF_X:
789 case BPF_JMP32 | BPF_JSGE | BPF_K:
790 case BPF_JMP32 | BPF_JSGE | BPF_X:
793 case BPF_JMP | BPF_JLE | BPF_K:
794 case BPF_JMP | BPF_JLE | BPF_X:
795 case BPF_JMP | BPF_JSLE | BPF_K:
796 case BPF_JMP | BPF_JSLE | BPF_X:
797 case BPF_JMP32 | BPF_JLE | BPF_K:
798 case BPF_JMP32 | BPF_JLE | BPF_X:
799 case BPF_JMP32 | BPF_JSLE | BPF_K:
800 case BPF_JMP32 | BPF_JSLE | BPF_X:
803 case BPF_JMP | BPF_JEQ | BPF_K:
804 case BPF_JMP | BPF_JEQ | BPF_X:
805 case BPF_JMP32 | BPF_JEQ | BPF_K:
806 case BPF_JMP32 | BPF_JEQ | BPF_X:
809 case BPF_JMP | BPF_JNE | BPF_K:
810 case BPF_JMP | BPF_JNE | BPF_X:
811 case BPF_JMP32 | BPF_JNE | BPF_K:
812 case BPF_JMP32 | BPF_JNE | BPF_X:
815 case BPF_JMP | BPF_JSET | BPF_K:
816 case BPF_JMP | BPF_JSET | BPF_X:
817 case BPF_JMP32 | BPF_JSET | BPF_K:
818 case BPF_JMP32 | BPF_JSET | BPF_X:
824 case BPF_JMP | BPF_JGT | BPF_X:
825 case BPF_JMP | BPF_JLT | BPF_X:
826 case BPF_JMP | BPF_JGE | BPF_X:
827 case BPF_JMP | BPF_JLE | BPF_X:
828 case BPF_JMP | BPF_JEQ | BPF_X:
829 case BPF_JMP | BPF_JNE | BPF_X:
830 case BPF_JMP32 | BPF_JGT | BPF_X:
831 case BPF_JMP32 | BPF_JLT | BPF_X:
832 case BPF_JMP32 | BPF_JGE | BPF_X:
833 case BPF_JMP32 | BPF_JLE | BPF_X:
834 case BPF_JMP32 | BPF_JEQ | BPF_X:
835 case BPF_JMP32 | BPF_JNE | BPF_X:
836 /* unsigned comparison */
837 if (BPF_CLASS(code) == BPF_JMP32)
838 PPC_CMPLW(dst_reg, src_reg);
840 PPC_CMPLD(dst_reg, src_reg);
842 case BPF_JMP | BPF_JSGT | BPF_X:
843 case BPF_JMP | BPF_JSLT | BPF_X:
844 case BPF_JMP | BPF_JSGE | BPF_X:
845 case BPF_JMP | BPF_JSLE | BPF_X:
846 case BPF_JMP32 | BPF_JSGT | BPF_X:
847 case BPF_JMP32 | BPF_JSLT | BPF_X:
848 case BPF_JMP32 | BPF_JSGE | BPF_X:
849 case BPF_JMP32 | BPF_JSLE | BPF_X:
850 /* signed comparison */
851 if (BPF_CLASS(code) == BPF_JMP32)
852 PPC_CMPW(dst_reg, src_reg);
854 PPC_CMPD(dst_reg, src_reg);
856 case BPF_JMP | BPF_JSET | BPF_X:
857 case BPF_JMP32 | BPF_JSET | BPF_X:
858 if (BPF_CLASS(code) == BPF_JMP) {
859 PPC_AND_DOT(b2p[TMP_REG_1], dst_reg,
862 int tmp_reg = b2p[TMP_REG_1];
864 PPC_AND(tmp_reg, dst_reg, src_reg);
865 PPC_RLWINM_DOT(tmp_reg, tmp_reg, 0, 0,
869 case BPF_JMP | BPF_JNE | BPF_K:
870 case BPF_JMP | BPF_JEQ | BPF_K:
871 case BPF_JMP | BPF_JGT | BPF_K:
872 case BPF_JMP | BPF_JLT | BPF_K:
873 case BPF_JMP | BPF_JGE | BPF_K:
874 case BPF_JMP | BPF_JLE | BPF_K:
875 case BPF_JMP32 | BPF_JNE | BPF_K:
876 case BPF_JMP32 | BPF_JEQ | BPF_K:
877 case BPF_JMP32 | BPF_JGT | BPF_K:
878 case BPF_JMP32 | BPF_JLT | BPF_K:
879 case BPF_JMP32 | BPF_JGE | BPF_K:
880 case BPF_JMP32 | BPF_JLE | BPF_K:
882 bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
885 * Need sign-extended load, so only positive
886 * values can be used as imm in cmpldi
888 if (imm >= 0 && imm < 32768) {
890 PPC_CMPLWI(dst_reg, imm);
892 PPC_CMPLDI(dst_reg, imm);
894 /* sign-extending load */
895 PPC_LI32(b2p[TMP_REG_1], imm);
896 /* ... but unsigned comparison */
906 case BPF_JMP | BPF_JSGT | BPF_K:
907 case BPF_JMP | BPF_JSLT | BPF_K:
908 case BPF_JMP | BPF_JSGE | BPF_K:
909 case BPF_JMP | BPF_JSLE | BPF_K:
910 case BPF_JMP32 | BPF_JSGT | BPF_K:
911 case BPF_JMP32 | BPF_JSLT | BPF_K:
912 case BPF_JMP32 | BPF_JSGE | BPF_K:
913 case BPF_JMP32 | BPF_JSLE | BPF_K:
915 bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
918 * signed comparison, so any 16-bit value
919 * can be used in cmpdi
921 if (imm >= -32768 && imm < 32768) {
923 PPC_CMPWI(dst_reg, imm);
925 PPC_CMPDI(dst_reg, imm);
927 PPC_LI32(b2p[TMP_REG_1], imm);
937 case BPF_JMP | BPF_JSET | BPF_K:
938 case BPF_JMP32 | BPF_JSET | BPF_K:
939 /* andi does not sign-extend the immediate */
940 if (imm >= 0 && imm < 32768)
941 /* PPC_ANDI is _only/always_ dot-form */
942 PPC_ANDI(b2p[TMP_REG_1], dst_reg, imm);
944 int tmp_reg = b2p[TMP_REG_1];
946 PPC_LI32(tmp_reg, imm);
947 if (BPF_CLASS(code) == BPF_JMP) {
948 PPC_AND_DOT(tmp_reg, dst_reg,
951 PPC_AND(tmp_reg, dst_reg,
953 PPC_RLWINM_DOT(tmp_reg, tmp_reg,
959 PPC_BCC(true_cond, addrs[i + 1 + off]);
965 case BPF_JMP | BPF_TAIL_CALL:
966 ctx->seen |= SEEN_TAILCALL;
967 bpf_jit_emit_tail_call(image, ctx, addrs[i + 1]);
972 * The filter contains something cruel & unusual.
973 * We don't handle it, but also there shouldn't be
974 * anything missing from our list.
976 pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n",
982 /* Set end-of-body-code address for exit. */
983 addrs[i] = ctx->idx * 4;
988 /* Fix the branch target addresses for subprog calls */
989 static int bpf_jit_fixup_subprog_calls(struct bpf_prog *fp, u32 *image,
990 struct codegen_context *ctx, u32 *addrs)
992 const struct bpf_insn *insn = fp->insnsi;
993 bool func_addr_fixed;
998 for (i = 0; i < fp->len; i++) {
1000 * During the extra pass, only the branch target addresses for
1001 * the subprog calls need to be fixed. All other instructions
1002 * can left untouched.
1004 * The JITed image length does not change because we already
1005 * ensure that the JITed instruction sequence for these calls
1006 * are of fixed length by padding them with NOPs.
1008 if (insn[i].code == (BPF_JMP | BPF_CALL) &&
1009 insn[i].src_reg == BPF_PSEUDO_CALL) {
1010 ret = bpf_jit_get_func_addr(fp, &insn[i], true,
1017 * Save ctx->idx as this would currently point to the
1018 * end of the JITed image and set it to the offset of
1019 * the instruction sequence corresponding to the
1020 * subprog call temporarily.
1023 ctx->idx = addrs[i] / 4;
1024 bpf_jit_emit_func_call_rel(image, ctx, func_addr);
1027 * Restore ctx->idx here. This is safe as the length
1028 * of the JITed sequence remains unchanged.
1037 struct powerpc64_jit_data {
1038 struct bpf_binary_header *header;
1042 struct codegen_context ctx;
1045 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
1052 struct powerpc64_jit_data *jit_data;
1053 struct codegen_context cgctx;
1056 struct bpf_binary_header *bpf_hdr;
1057 struct bpf_prog *org_fp = fp;
1058 struct bpf_prog *tmp_fp;
1059 bool bpf_blinded = false;
1060 bool extra_pass = false;
1062 if (!fp->jit_requested)
1065 tmp_fp = bpf_jit_blind_constants(org_fp);
1069 if (tmp_fp != org_fp) {
1074 jit_data = fp->aux->jit_data;
1076 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1081 fp->aux->jit_data = jit_data;
1085 addrs = jit_data->addrs;
1087 cgctx = jit_data->ctx;
1088 image = jit_data->image;
1089 bpf_hdr = jit_data->header;
1090 proglen = jit_data->proglen;
1091 alloclen = proglen + FUNCTION_DESCR_SIZE;
1096 addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
1097 if (addrs == NULL) {
1102 memset(&cgctx, 0, sizeof(struct codegen_context));
1104 /* Make sure that the stack is quadword aligned. */
1105 cgctx.stack_size = round_up(fp->aux->stack_depth, 16);
1107 /* Scouting faux-generate pass 0 */
1108 if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
1109 /* We hit something illegal or unsupported. */
1115 * Pretend to build prologue, given the features we've seen. This will
1116 * update ctgtx.idx as it pretends to output instructions, then we can
1117 * calculate total size from idx.
1119 bpf_jit_build_prologue(0, &cgctx);
1120 bpf_jit_build_epilogue(0, &cgctx);
1122 proglen = cgctx.idx * 4;
1123 alloclen = proglen + FUNCTION_DESCR_SIZE;
1125 bpf_hdr = bpf_jit_binary_alloc(alloclen, &image, 4,
1126 bpf_jit_fill_ill_insns);
1133 code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
1137 * Do not touch the prologue and epilogue as they will remain
1138 * unchanged. Only fix the branch target address for subprog
1139 * calls in the body.
1141 * This does not change the offsets and lengths of the subprog
1142 * call instruction sequences and hence, the size of the JITed
1145 bpf_jit_fixup_subprog_calls(fp, code_base, &cgctx, addrs);
1147 /* There is no need to perform the usual passes. */
1148 goto skip_codegen_passes;
1151 /* Code generation passes 1-2 */
1152 for (pass = 1; pass < 3; pass++) {
1153 /* Now build the prologue, body code & epilogue for real. */
1155 bpf_jit_build_prologue(code_base, &cgctx);
1156 bpf_jit_build_body(fp, code_base, &cgctx, addrs, extra_pass);
1157 bpf_jit_build_epilogue(code_base, &cgctx);
1159 if (bpf_jit_enable > 1)
1160 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
1161 proglen - (cgctx.idx * 4), cgctx.seen);
1164 skip_codegen_passes:
1165 if (bpf_jit_enable > 1)
1167 * Note that we output the base address of the code_base
1168 * rather than image, since opcodes are in code_base.
1170 bpf_jit_dump(flen, proglen, pass, code_base);
1172 #ifdef PPC64_ELF_ABI_v1
1173 /* Function descriptor nastiness: Address + TOC */
1174 ((u64 *)image)[0] = (u64)code_base;
1175 ((u64 *)image)[1] = local_paca->kernel_toc;
1178 fp->bpf_func = (void *)image;
1180 fp->jited_len = alloclen;
1182 bpf_flush_icache(bpf_hdr, (u8 *)bpf_hdr + (bpf_hdr->pages * PAGE_SIZE));
1183 if (!fp->is_func || extra_pass) {
1184 bpf_prog_fill_jited_linfo(fp, addrs);
1188 fp->aux->jit_data = NULL;
1190 jit_data->addrs = addrs;
1191 jit_data->ctx = cgctx;
1192 jit_data->proglen = proglen;
1193 jit_data->image = image;
1194 jit_data->header = bpf_hdr;
1199 bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
1204 /* Overriding bpf_jit_free() as we don't set images read-only. */
1205 void bpf_jit_free(struct bpf_prog *fp)
1207 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
1208 struct bpf_binary_header *bpf_hdr = (void *)addr;
1211 bpf_jit_binary_free(bpf_hdr);
1213 bpf_prog_unlock_free(fp);