1 // SPDX-License-Identifier: GPL-2.0-only
3 * bpf_jit_comp64.c: eBPF JIT compiler
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);
503 /* skip zero extension move, but set address map. */
504 if (insn_is_zext(&insn[i + 1]))
505 addrs[++i] = ctx->idx * 4;
507 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */
508 PPC_SLD(dst_reg, dst_reg, src_reg);
510 case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<== (u32) imm */
511 /* with imm 0, we still need to clear top 32 bits */
512 PPC_SLWI(dst_reg, dst_reg, imm);
513 if (insn_is_zext(&insn[i + 1]))
514 addrs[++i] = ctx->idx * 4;
516 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<== imm */
518 PPC_SLDI(dst_reg, dst_reg, imm);
520 case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */
521 PPC_SRW(dst_reg, dst_reg, src_reg);
522 if (insn_is_zext(&insn[i + 1]))
523 addrs[++i] = ctx->idx * 4;
525 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */
526 PPC_SRD(dst_reg, dst_reg, src_reg);
528 case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */
529 PPC_SRWI(dst_reg, dst_reg, imm);
530 if (insn_is_zext(&insn[i + 1]))
531 addrs[++i] = ctx->idx * 4;
533 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */
535 PPC_SRDI(dst_reg, dst_reg, imm);
537 case BPF_ALU | BPF_ARSH | BPF_X: /* (s32) dst >>= src */
538 PPC_SRAW(dst_reg, dst_reg, src_reg);
539 goto bpf_alu32_trunc;
540 case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */
541 PPC_SRAD(dst_reg, dst_reg, src_reg);
543 case BPF_ALU | BPF_ARSH | BPF_K: /* (s32) dst >>= imm */
544 PPC_SRAWI(dst_reg, dst_reg, imm);
545 goto bpf_alu32_trunc;
546 case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */
548 PPC_SRADI(dst_reg, dst_reg, imm);
554 case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
555 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
557 /* special mov32 for zext */
558 PPC_RLWINM(dst_reg, dst_reg, 0, 0, 31);
561 PPC_MR(dst_reg, src_reg);
562 goto bpf_alu32_trunc;
563 case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */
564 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */
565 PPC_LI32(dst_reg, imm);
567 goto bpf_alu32_trunc;
568 else if (insn_is_zext(&insn[i + 1]))
569 addrs[++i] = ctx->idx * 4;
573 /* Truncate to 32-bits */
574 if (BPF_CLASS(code) == BPF_ALU && !fp->aux->verifier_zext)
575 PPC_RLWINM(dst_reg, dst_reg, 0, 0, 31);
581 case BPF_ALU | BPF_END | BPF_FROM_LE:
582 case BPF_ALU | BPF_END | BPF_FROM_BE:
583 #ifdef __BIG_ENDIAN__
584 if (BPF_SRC(code) == BPF_FROM_BE)
586 #else /* !__BIG_ENDIAN__ */
587 if (BPF_SRC(code) == BPF_FROM_LE)
592 /* Rotate 8 bits left & mask with 0x0000ff00 */
593 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 16, 23);
594 /* Rotate 8 bits right & insert LSB to reg */
595 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 24, 31);
596 /* Move result back to dst_reg */
597 PPC_MR(dst_reg, b2p[TMP_REG_1]);
601 * Rotate word left by 8 bits:
602 * 2 bytes are already in their final position
603 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4)
605 PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 0, 31);
606 /* Rotate 24 bits and insert byte 1 */
607 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 0, 7);
608 /* Rotate 24 bits and insert byte 3 */
609 PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 16, 23);
610 PPC_MR(dst_reg, b2p[TMP_REG_1]);
614 * Way easier and faster(?) to store the value
615 * into stack and then use ldbrx
617 * ctx->seen will be reliable in pass2, but
618 * the instructions generated will remain the
619 * same across all passes
621 PPC_BPF_STL(dst_reg, 1, bpf_jit_stack_local(ctx));
622 PPC_ADDI(b2p[TMP_REG_1], 1, bpf_jit_stack_local(ctx));
623 PPC_LDBRX(dst_reg, 0, b2p[TMP_REG_1]);
631 /* zero-extend 16 bits into 64 bits */
632 PPC_RLDICL(dst_reg, dst_reg, 0, 48);
633 if (insn_is_zext(&insn[i + 1]))
634 addrs[++i] = ctx->idx * 4;
637 if (!fp->aux->verifier_zext)
638 /* zero-extend 32 bits into 64 bits */
639 PPC_RLDICL(dst_reg, dst_reg, 0, 32);
650 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */
651 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
652 if (BPF_CLASS(code) == BPF_ST) {
653 PPC_LI(b2p[TMP_REG_1], imm);
654 src_reg = b2p[TMP_REG_1];
656 PPC_STB(src_reg, dst_reg, off);
658 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
659 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
660 if (BPF_CLASS(code) == BPF_ST) {
661 PPC_LI(b2p[TMP_REG_1], imm);
662 src_reg = b2p[TMP_REG_1];
664 PPC_STH(src_reg, dst_reg, off);
666 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
667 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
668 if (BPF_CLASS(code) == BPF_ST) {
669 PPC_LI32(b2p[TMP_REG_1], imm);
670 src_reg = b2p[TMP_REG_1];
672 PPC_STW(src_reg, dst_reg, off);
674 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
675 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
676 if (BPF_CLASS(code) == BPF_ST) {
677 PPC_LI32(b2p[TMP_REG_1], imm);
678 src_reg = b2p[TMP_REG_1];
680 PPC_BPF_STL(src_reg, dst_reg, off);
684 * BPF_STX XADD (atomic_add)
686 /* *(u32 *)(dst + off) += src */
687 case BPF_STX | BPF_XADD | BPF_W:
688 /* Get EA into TMP_REG_1 */
689 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
690 tmp_idx = ctx->idx * 4;
691 /* load value from memory into TMP_REG_2 */
692 PPC_BPF_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0);
693 /* add value from src_reg into this */
694 PPC_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg);
695 /* store result back */
696 PPC_BPF_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]);
697 /* we're done if this succeeded */
698 PPC_BCC_SHORT(COND_NE, tmp_idx);
700 /* *(u64 *)(dst + off) += src */
701 case BPF_STX | BPF_XADD | BPF_DW:
702 PPC_ADDI(b2p[TMP_REG_1], dst_reg, off);
703 tmp_idx = ctx->idx * 4;
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_NE, tmp_idx);
713 /* dst = *(u8 *)(ul) (src + off) */
714 case BPF_LDX | BPF_MEM | BPF_B:
715 PPC_LBZ(dst_reg, src_reg, off);
716 if (insn_is_zext(&insn[i + 1]))
717 addrs[++i] = ctx->idx * 4;
719 /* dst = *(u16 *)(ul) (src + off) */
720 case BPF_LDX | BPF_MEM | BPF_H:
721 PPC_LHZ(dst_reg, src_reg, off);
722 if (insn_is_zext(&insn[i + 1]))
723 addrs[++i] = ctx->idx * 4;
725 /* dst = *(u32 *)(ul) (src + off) */
726 case BPF_LDX | BPF_MEM | BPF_W:
727 PPC_LWZ(dst_reg, src_reg, off);
728 if (insn_is_zext(&insn[i + 1]))
729 addrs[++i] = ctx->idx * 4;
731 /* dst = *(u64 *)(ul) (src + off) */
732 case BPF_LDX | BPF_MEM | BPF_DW:
733 PPC_BPF_LL(dst_reg, src_reg, off);
738 * 16 byte instruction that uses two 'struct bpf_insn'
740 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
741 imm64 = ((u64)(u32) insn[i].imm) |
742 (((u64)(u32) insn[i+1].imm) << 32);
743 /* Adjust for two bpf instructions */
744 addrs[++i] = ctx->idx * 4;
745 PPC_LI64(dst_reg, imm64);
751 case BPF_JMP | BPF_EXIT:
753 * If this isn't the very last instruction, branch to
754 * the epilogue. If we _are_ the last instruction,
755 * we'll just fall through to the epilogue.
759 /* else fall through to the epilogue */
763 * Call kernel helper or bpf function
765 case BPF_JMP | BPF_CALL:
766 ctx->seen |= SEEN_FUNC;
768 ret = bpf_jit_get_func_addr(fp, &insn[i], extra_pass,
769 &func_addr, &func_addr_fixed);
774 bpf_jit_emit_func_call_hlp(image, ctx, func_addr);
776 bpf_jit_emit_func_call_rel(image, ctx, func_addr);
777 /* move return value from r3 to BPF_REG_0 */
778 PPC_MR(b2p[BPF_REG_0], 3);
784 case BPF_JMP | BPF_JA:
785 PPC_JMP(addrs[i + 1 + off]);
788 case BPF_JMP | BPF_JGT | BPF_K:
789 case BPF_JMP | BPF_JGT | BPF_X:
790 case BPF_JMP | BPF_JSGT | BPF_K:
791 case BPF_JMP | BPF_JSGT | BPF_X:
792 case BPF_JMP32 | BPF_JGT | BPF_K:
793 case BPF_JMP32 | BPF_JGT | BPF_X:
794 case BPF_JMP32 | BPF_JSGT | BPF_K:
795 case BPF_JMP32 | BPF_JSGT | BPF_X:
798 case BPF_JMP | BPF_JLT | BPF_K:
799 case BPF_JMP | BPF_JLT | BPF_X:
800 case BPF_JMP | BPF_JSLT | BPF_K:
801 case BPF_JMP | BPF_JSLT | BPF_X:
802 case BPF_JMP32 | BPF_JLT | BPF_K:
803 case BPF_JMP32 | BPF_JLT | BPF_X:
804 case BPF_JMP32 | BPF_JSLT | BPF_K:
805 case BPF_JMP32 | BPF_JSLT | BPF_X:
808 case BPF_JMP | BPF_JGE | BPF_K:
809 case BPF_JMP | BPF_JGE | BPF_X:
810 case BPF_JMP | BPF_JSGE | BPF_K:
811 case BPF_JMP | BPF_JSGE | BPF_X:
812 case BPF_JMP32 | BPF_JGE | BPF_K:
813 case BPF_JMP32 | BPF_JGE | BPF_X:
814 case BPF_JMP32 | BPF_JSGE | BPF_K:
815 case BPF_JMP32 | BPF_JSGE | BPF_X:
818 case BPF_JMP | BPF_JLE | BPF_K:
819 case BPF_JMP | BPF_JLE | BPF_X:
820 case BPF_JMP | BPF_JSLE | BPF_K:
821 case BPF_JMP | BPF_JSLE | BPF_X:
822 case BPF_JMP32 | BPF_JLE | BPF_K:
823 case BPF_JMP32 | BPF_JLE | BPF_X:
824 case BPF_JMP32 | BPF_JSLE | BPF_K:
825 case BPF_JMP32 | BPF_JSLE | BPF_X:
828 case BPF_JMP | BPF_JEQ | BPF_K:
829 case BPF_JMP | BPF_JEQ | BPF_X:
830 case BPF_JMP32 | BPF_JEQ | BPF_K:
831 case BPF_JMP32 | BPF_JEQ | BPF_X:
834 case BPF_JMP | BPF_JNE | BPF_K:
835 case BPF_JMP | BPF_JNE | BPF_X:
836 case BPF_JMP32 | BPF_JNE | BPF_K:
837 case BPF_JMP32 | BPF_JNE | BPF_X:
840 case BPF_JMP | BPF_JSET | BPF_K:
841 case BPF_JMP | BPF_JSET | BPF_X:
842 case BPF_JMP32 | BPF_JSET | BPF_K:
843 case BPF_JMP32 | BPF_JSET | BPF_X:
849 case BPF_JMP | BPF_JGT | BPF_X:
850 case BPF_JMP | BPF_JLT | BPF_X:
851 case BPF_JMP | BPF_JGE | BPF_X:
852 case BPF_JMP | BPF_JLE | BPF_X:
853 case BPF_JMP | BPF_JEQ | BPF_X:
854 case BPF_JMP | BPF_JNE | BPF_X:
855 case BPF_JMP32 | BPF_JGT | BPF_X:
856 case BPF_JMP32 | BPF_JLT | BPF_X:
857 case BPF_JMP32 | BPF_JGE | BPF_X:
858 case BPF_JMP32 | BPF_JLE | BPF_X:
859 case BPF_JMP32 | BPF_JEQ | BPF_X:
860 case BPF_JMP32 | BPF_JNE | BPF_X:
861 /* unsigned comparison */
862 if (BPF_CLASS(code) == BPF_JMP32)
863 PPC_CMPLW(dst_reg, src_reg);
865 PPC_CMPLD(dst_reg, src_reg);
867 case BPF_JMP | BPF_JSGT | BPF_X:
868 case BPF_JMP | BPF_JSLT | BPF_X:
869 case BPF_JMP | BPF_JSGE | BPF_X:
870 case BPF_JMP | BPF_JSLE | BPF_X:
871 case BPF_JMP32 | BPF_JSGT | BPF_X:
872 case BPF_JMP32 | BPF_JSLT | BPF_X:
873 case BPF_JMP32 | BPF_JSGE | BPF_X:
874 case BPF_JMP32 | BPF_JSLE | BPF_X:
875 /* signed comparison */
876 if (BPF_CLASS(code) == BPF_JMP32)
877 PPC_CMPW(dst_reg, src_reg);
879 PPC_CMPD(dst_reg, src_reg);
881 case BPF_JMP | BPF_JSET | BPF_X:
882 case BPF_JMP32 | BPF_JSET | BPF_X:
883 if (BPF_CLASS(code) == BPF_JMP) {
884 PPC_AND_DOT(b2p[TMP_REG_1], dst_reg,
887 int tmp_reg = b2p[TMP_REG_1];
889 PPC_AND(tmp_reg, dst_reg, src_reg);
890 PPC_RLWINM_DOT(tmp_reg, tmp_reg, 0, 0,
894 case BPF_JMP | BPF_JNE | BPF_K:
895 case BPF_JMP | BPF_JEQ | BPF_K:
896 case BPF_JMP | BPF_JGT | BPF_K:
897 case BPF_JMP | BPF_JLT | BPF_K:
898 case BPF_JMP | BPF_JGE | BPF_K:
899 case BPF_JMP | BPF_JLE | BPF_K:
900 case BPF_JMP32 | BPF_JNE | BPF_K:
901 case BPF_JMP32 | BPF_JEQ | BPF_K:
902 case BPF_JMP32 | BPF_JGT | BPF_K:
903 case BPF_JMP32 | BPF_JLT | BPF_K:
904 case BPF_JMP32 | BPF_JGE | BPF_K:
905 case BPF_JMP32 | BPF_JLE | BPF_K:
907 bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
910 * Need sign-extended load, so only positive
911 * values can be used as imm in cmpldi
913 if (imm >= 0 && imm < 32768) {
915 PPC_CMPLWI(dst_reg, imm);
917 PPC_CMPLDI(dst_reg, imm);
919 /* sign-extending load */
920 PPC_LI32(b2p[TMP_REG_1], imm);
921 /* ... but unsigned comparison */
931 case BPF_JMP | BPF_JSGT | BPF_K:
932 case BPF_JMP | BPF_JSLT | BPF_K:
933 case BPF_JMP | BPF_JSGE | BPF_K:
934 case BPF_JMP | BPF_JSLE | BPF_K:
935 case BPF_JMP32 | BPF_JSGT | BPF_K:
936 case BPF_JMP32 | BPF_JSLT | BPF_K:
937 case BPF_JMP32 | BPF_JSGE | BPF_K:
938 case BPF_JMP32 | BPF_JSLE | BPF_K:
940 bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
943 * signed comparison, so any 16-bit value
944 * can be used in cmpdi
946 if (imm >= -32768 && imm < 32768) {
948 PPC_CMPWI(dst_reg, imm);
950 PPC_CMPDI(dst_reg, imm);
952 PPC_LI32(b2p[TMP_REG_1], imm);
962 case BPF_JMP | BPF_JSET | BPF_K:
963 case BPF_JMP32 | BPF_JSET | BPF_K:
964 /* andi does not sign-extend the immediate */
965 if (imm >= 0 && imm < 32768)
966 /* PPC_ANDI is _only/always_ dot-form */
967 PPC_ANDI(b2p[TMP_REG_1], dst_reg, imm);
969 int tmp_reg = b2p[TMP_REG_1];
971 PPC_LI32(tmp_reg, imm);
972 if (BPF_CLASS(code) == BPF_JMP) {
973 PPC_AND_DOT(tmp_reg, dst_reg,
976 PPC_AND(tmp_reg, dst_reg,
978 PPC_RLWINM_DOT(tmp_reg, tmp_reg,
984 PPC_BCC(true_cond, addrs[i + 1 + off]);
990 case BPF_JMP | BPF_TAIL_CALL:
991 ctx->seen |= SEEN_TAILCALL;
992 bpf_jit_emit_tail_call(image, ctx, addrs[i + 1]);
997 * The filter contains something cruel & unusual.
998 * We don't handle it, but also there shouldn't be
999 * anything missing from our list.
1001 pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n",
1007 /* Set end-of-body-code address for exit. */
1008 addrs[i] = ctx->idx * 4;
1013 /* Fix the branch target addresses for subprog calls */
1014 static int bpf_jit_fixup_subprog_calls(struct bpf_prog *fp, u32 *image,
1015 struct codegen_context *ctx, u32 *addrs)
1017 const struct bpf_insn *insn = fp->insnsi;
1018 bool func_addr_fixed;
1023 for (i = 0; i < fp->len; i++) {
1025 * During the extra pass, only the branch target addresses for
1026 * the subprog calls need to be fixed. All other instructions
1027 * can left untouched.
1029 * The JITed image length does not change because we already
1030 * ensure that the JITed instruction sequence for these calls
1031 * are of fixed length by padding them with NOPs.
1033 if (insn[i].code == (BPF_JMP | BPF_CALL) &&
1034 insn[i].src_reg == BPF_PSEUDO_CALL) {
1035 ret = bpf_jit_get_func_addr(fp, &insn[i], true,
1042 * Save ctx->idx as this would currently point to the
1043 * end of the JITed image and set it to the offset of
1044 * the instruction sequence corresponding to the
1045 * subprog call temporarily.
1048 ctx->idx = addrs[i] / 4;
1049 bpf_jit_emit_func_call_rel(image, ctx, func_addr);
1052 * Restore ctx->idx here. This is safe as the length
1053 * of the JITed sequence remains unchanged.
1062 struct powerpc64_jit_data {
1063 struct bpf_binary_header *header;
1067 struct codegen_context ctx;
1070 bool bpf_jit_needs_zext(void)
1075 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
1082 struct powerpc64_jit_data *jit_data;
1083 struct codegen_context cgctx;
1086 struct bpf_binary_header *bpf_hdr;
1087 struct bpf_prog *org_fp = fp;
1088 struct bpf_prog *tmp_fp;
1089 bool bpf_blinded = false;
1090 bool extra_pass = false;
1092 if (!fp->jit_requested)
1095 tmp_fp = bpf_jit_blind_constants(org_fp);
1099 if (tmp_fp != org_fp) {
1104 jit_data = fp->aux->jit_data;
1106 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1111 fp->aux->jit_data = jit_data;
1115 addrs = jit_data->addrs;
1117 cgctx = jit_data->ctx;
1118 image = jit_data->image;
1119 bpf_hdr = jit_data->header;
1120 proglen = jit_data->proglen;
1121 alloclen = proglen + FUNCTION_DESCR_SIZE;
1126 addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
1127 if (addrs == NULL) {
1132 memset(&cgctx, 0, sizeof(struct codegen_context));
1134 /* Make sure that the stack is quadword aligned. */
1135 cgctx.stack_size = round_up(fp->aux->stack_depth, 16);
1137 /* Scouting faux-generate pass 0 */
1138 if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
1139 /* We hit something illegal or unsupported. */
1145 * Pretend to build prologue, given the features we've seen. This will
1146 * update ctgtx.idx as it pretends to output instructions, then we can
1147 * calculate total size from idx.
1149 bpf_jit_build_prologue(0, &cgctx);
1150 bpf_jit_build_epilogue(0, &cgctx);
1152 proglen = cgctx.idx * 4;
1153 alloclen = proglen + FUNCTION_DESCR_SIZE;
1155 bpf_hdr = bpf_jit_binary_alloc(alloclen, &image, 4,
1156 bpf_jit_fill_ill_insns);
1163 code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
1167 * Do not touch the prologue and epilogue as they will remain
1168 * unchanged. Only fix the branch target address for subprog
1169 * calls in the body.
1171 * This does not change the offsets and lengths of the subprog
1172 * call instruction sequences and hence, the size of the JITed
1175 bpf_jit_fixup_subprog_calls(fp, code_base, &cgctx, addrs);
1177 /* There is no need to perform the usual passes. */
1178 goto skip_codegen_passes;
1181 /* Code generation passes 1-2 */
1182 for (pass = 1; pass < 3; pass++) {
1183 /* Now build the prologue, body code & epilogue for real. */
1185 bpf_jit_build_prologue(code_base, &cgctx);
1186 bpf_jit_build_body(fp, code_base, &cgctx, addrs, extra_pass);
1187 bpf_jit_build_epilogue(code_base, &cgctx);
1189 if (bpf_jit_enable > 1)
1190 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
1191 proglen - (cgctx.idx * 4), cgctx.seen);
1194 skip_codegen_passes:
1195 if (bpf_jit_enable > 1)
1197 * Note that we output the base address of the code_base
1198 * rather than image, since opcodes are in code_base.
1200 bpf_jit_dump(flen, proglen, pass, code_base);
1202 #ifdef PPC64_ELF_ABI_v1
1203 /* Function descriptor nastiness: Address + TOC */
1204 ((u64 *)image)[0] = (u64)code_base;
1205 ((u64 *)image)[1] = local_paca->kernel_toc;
1208 fp->bpf_func = (void *)image;
1210 fp->jited_len = alloclen;
1212 bpf_flush_icache(bpf_hdr, (u8 *)bpf_hdr + (bpf_hdr->pages * PAGE_SIZE));
1213 if (!fp->is_func || extra_pass) {
1214 bpf_prog_fill_jited_linfo(fp, addrs);
1218 fp->aux->jit_data = NULL;
1220 jit_data->addrs = addrs;
1221 jit_data->ctx = cgctx;
1222 jit_data->proglen = proglen;
1223 jit_data->image = image;
1224 jit_data->header = bpf_hdr;
1229 bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
1234 /* Overriding bpf_jit_free() as we don't set images read-only. */
1235 void bpf_jit_free(struct bpf_prog *fp)
1237 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
1238 struct bpf_binary_header *bpf_hdr = (void *)addr;
1241 bpf_jit_binary_free(bpf_hdr);
1243 bpf_prog_unlock_free(fp);