xref: /linux/arch/powerpc/net/bpf_jit_comp64.c (revision 26fbb4c8c7c3ee9a4c3b4de555a8587b5a19154e)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * bpf_jit_comp64.c: eBPF JIT compiler
4  *
5  * Copyright 2016 Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
6  *		  IBM Corporation
7  *
8  * Based on the powerpc classic BPF JIT compiler by Matt Evans
9  */
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>
18 
19 #include "bpf_jit64.h"
20 
21 static void bpf_jit_fill_ill_insns(void *area, unsigned int size)
22 {
23 	memset32(area, BREAKPOINT_INSTRUCTION, size/4);
24 }
25 
26 static inline void bpf_flush_icache(void *start, void *end)
27 {
28 	smp_wmb();
29 	flush_icache_range((unsigned long)start, (unsigned long)end);
30 }
31 
32 static inline bool bpf_is_seen_register(struct codegen_context *ctx, int i)
33 {
34 	return (ctx->seen & (1 << (31 - b2p[i])));
35 }
36 
37 static inline void bpf_set_seen_register(struct codegen_context *ctx, int i)
38 {
39 	ctx->seen |= (1 << (31 - b2p[i]));
40 }
41 
42 static inline bool bpf_has_stack_frame(struct codegen_context *ctx)
43 {
44 	/*
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
49 	 */
50 	return ctx->seen & SEEN_FUNC || bpf_is_seen_register(ctx, BPF_REG_FP);
51 }
52 
53 /*
54  * When not setting up our own stackframe, the redzone usage is:
55  *
56  *		[	prev sp		] <-------------
57  *		[	  ...       	] 		|
58  * sp (r1) --->	[    stack pointer	] --------------
59  *		[   nv gpr save area	] 6*8
60  *		[    tail_call_cnt	] 8
61  *		[    local_tmp_var	] 8
62  *		[   unused red zone	] 208 bytes protected
63  */
64 static int bpf_jit_stack_local(struct codegen_context *ctx)
65 {
66 	if (bpf_has_stack_frame(ctx))
67 		return STACK_FRAME_MIN_SIZE + ctx->stack_size;
68 	else
69 		return -(BPF_PPC_STACK_SAVE + 16);
70 }
71 
72 static int bpf_jit_stack_tailcallcnt(struct codegen_context *ctx)
73 {
74 	return bpf_jit_stack_local(ctx) + 8;
75 }
76 
77 static int bpf_jit_stack_offsetof(struct codegen_context *ctx, int reg)
78 {
79 	if (reg >= BPF_PPC_NVR_MIN && reg < 32)
80 		return (bpf_has_stack_frame(ctx) ?
81 			(BPF_PPC_STACKFRAME + ctx->stack_size) : 0)
82 				- (8 * (32 - reg));
83 
84 	pr_err("BPF JIT is asking about unknown registers");
85 	BUG();
86 }
87 
88 static void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx)
89 {
90 	int i;
91 
92 	/*
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.
96 	 */
97 	if (ctx->seen & SEEN_TAILCALL) {
98 		EMIT(PPC_RAW_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));
101 	} else {
102 		EMIT(PPC_RAW_NOP());
103 		EMIT(PPC_RAW_NOP());
104 	}
105 
106 #define BPF_TAILCALL_PROLOGUE_SIZE	8
107 
108 	if (bpf_has_stack_frame(ctx)) {
109 		/*
110 		 * We need a stack frame, but we don't necessarily need to
111 		 * save/restore LR unless we call other functions
112 		 */
113 		if (ctx->seen & SEEN_FUNC) {
114 			EMIT(PPC_INST_MFLR | __PPC_RT(R0));
115 			PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
116 		}
117 
118 		PPC_BPF_STLU(1, 1, -(BPF_PPC_STACKFRAME + ctx->stack_size));
119 	}
120 
121 	/*
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
125 	 */
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]));
129 
130 	/* Setup frame pointer to point to the bpf stack area */
131 	if (bpf_is_seen_register(ctx, BPF_REG_FP))
132 		EMIT(PPC_RAW_ADDI(b2p[BPF_REG_FP], 1,
133 				STACK_FRAME_MIN_SIZE + ctx->stack_size));
134 }
135 
136 static void bpf_jit_emit_common_epilogue(u32 *image, struct codegen_context *ctx)
137 {
138 	int i;
139 
140 	/* Restore NVRs */
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]));
144 
145 	/* Tear down our stack frame */
146 	if (bpf_has_stack_frame(ctx)) {
147 		EMIT(PPC_RAW_ADDI(1, 1, BPF_PPC_STACKFRAME + ctx->stack_size));
148 		if (ctx->seen & SEEN_FUNC) {
149 			PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
150 			EMIT(PPC_RAW_MTLR(0));
151 		}
152 	}
153 }
154 
155 static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
156 {
157 	bpf_jit_emit_common_epilogue(image, ctx);
158 
159 	/* Move result to r3 */
160 	EMIT(PPC_RAW_MR(3, b2p[BPF_REG_0]));
161 
162 	EMIT(PPC_RAW_BLR());
163 }
164 
165 static void bpf_jit_emit_func_call_hlp(u32 *image, struct codegen_context *ctx,
166 				       u64 func)
167 {
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 	EMIT(PPC_RAW_MTLR(b2p[TMP_REG_1]));
175 	/*
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.
180 	 */
181 	PPC_BPF_LL(2, b2p[TMP_REG_2], 8);
182 #else
183 	/* We can clobber r12 */
184 	PPC_FUNC_ADDR(12, func);
185 	EMIT(PPC_RAW_MTLR(12));
186 #endif
187 	EMIT(PPC_RAW_BLRL());
188 }
189 
190 static void bpf_jit_emit_func_call_rel(u32 *image, struct codegen_context *ctx,
191 				       u64 func)
192 {
193 	unsigned int i, ctx_idx = ctx->idx;
194 
195 	/* Load function address into r12 */
196 	PPC_LI64(12, func);
197 
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.
202 	 *
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.
207 	 */
208 	for (i = ctx->idx - ctx_idx; i < 5; i++)
209 		EMIT(PPC_RAW_NOP());
210 
211 #ifdef PPC64_ELF_ABI_v1
212 	/*
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.
217 	 */
218 	PPC_BPF_LL(2, 12, 8);
219 	/* Load actual entry point from function descriptor */
220 	PPC_BPF_LL(12, 12, 0);
221 #endif
222 
223 	EMIT(PPC_RAW_MTLR(12));
224 	EMIT(PPC_RAW_BLRL());
225 }
226 
227 static void bpf_jit_emit_tail_call(u32 *image, struct codegen_context *ctx, u32 out)
228 {
229 	/*
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
234 	 */
235 	int b2p_bpf_array = b2p[BPF_REG_2];
236 	int b2p_index = b2p[BPF_REG_3];
237 
238 	/*
239 	 * if (index >= array->map.max_entries)
240 	 *   goto out;
241 	 */
242 	EMIT(PPC_RAW_LWZ(b2p[TMP_REG_1], b2p_bpf_array, offsetof(struct bpf_array, map.max_entries)));
243 	EMIT(PPC_RAW_RLWINM(b2p_index, b2p_index, 0, 0, 31));
244 	EMIT(PPC_RAW_CMPLW(b2p_index, b2p[TMP_REG_1]));
245 	PPC_BCC(COND_GE, out);
246 
247 	/*
248 	 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
249 	 *   goto out;
250 	 */
251 	PPC_BPF_LL(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
252 	EMIT(PPC_RAW_CMPLWI(b2p[TMP_REG_1], MAX_TAIL_CALL_CNT));
253 	PPC_BCC(COND_GT, out);
254 
255 	/*
256 	 * tail_call_cnt++;
257 	 */
258 	EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], 1));
259 	PPC_BPF_STL(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
260 
261 	/* prog = array->ptrs[index]; */
262 	EMIT(PPC_RAW_MULI(b2p[TMP_REG_1], b2p_index, 8));
263 	EMIT(PPC_RAW_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));
265 
266 	/*
267 	 * if (prog == NULL)
268 	 *   goto out;
269 	 */
270 	EMIT(PPC_RAW_CMPLDI(b2p[TMP_REG_1], 0));
271 	PPC_BCC(COND_EQ, out);
272 
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 	EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1],
278 			FUNCTION_DESCR_SIZE + BPF_TAILCALL_PROLOGUE_SIZE));
279 #else
280 	EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], BPF_TAILCALL_PROLOGUE_SIZE));
281 #endif
282 	EMIT(PPC_RAW_MTCTR(b2p[TMP_REG_1]));
283 
284 	/* tear down stack, restore NVRs, ... */
285 	bpf_jit_emit_common_epilogue(image, ctx);
286 
287 	EMIT(PPC_RAW_BCTR());
288 	/* out: */
289 }
290 
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)
295 {
296 	const struct bpf_insn *insn = fp->insnsi;
297 	int flen = fp->len;
298 	int i, ret;
299 
300 	/* Start of epilogue code - will only be valid 2nd pass onwards */
301 	u32 exit_addr = addrs[flen];
302 
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;
310 		u64 func_addr;
311 		u64 imm64;
312 		u32 true_cond;
313 		u32 tmp_idx;
314 
315 		/*
316 		 * addrs[] maps a BPF bytecode address into a real offset from
317 		 * the start of the body code.
318 		 */
319 		addrs[i] = ctx->idx * 4;
320 
321 		/*
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
330 		 * any issues.
331 		 */
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);
336 
337 		switch (code) {
338 		/*
339 		 * Arithmetic operations: ADD/SUB/MUL/DIV/MOD/NEG
340 		 */
341 		case BPF_ALU | BPF_ADD | BPF_X: /* (u32) dst += (u32) src */
342 		case BPF_ALU64 | BPF_ADD | BPF_X: /* dst += src */
343 			EMIT(PPC_RAW_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 			EMIT(PPC_RAW_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)
354 				imm = -imm;
355 			if (imm) {
356 				if (imm >= -32768 && imm < 32768)
357 					EMIT(PPC_RAW_ADDI(dst_reg, dst_reg, IMM_L(imm)));
358 				else {
359 					PPC_LI32(b2p[TMP_REG_1], imm);
360 					EMIT(PPC_RAW_ADD(dst_reg, dst_reg, b2p[TMP_REG_1]));
361 				}
362 			}
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 				EMIT(PPC_RAW_MULW(dst_reg, dst_reg, src_reg));
368 			else
369 				EMIT(PPC_RAW_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 				EMIT(PPC_RAW_MULI(dst_reg, dst_reg, IMM_L(imm)));
375 			else {
376 				PPC_LI32(b2p[TMP_REG_1], imm);
377 				if (BPF_CLASS(code) == BPF_ALU)
378 					EMIT(PPC_RAW_MULW(dst_reg, dst_reg,
379 							b2p[TMP_REG_1]));
380 				else
381 					EMIT(PPC_RAW_MULD(dst_reg, dst_reg,
382 							b2p[TMP_REG_1]));
383 			}
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 				EMIT(PPC_RAW_DIVWU(b2p[TMP_REG_1], dst_reg, src_reg));
389 				EMIT(PPC_RAW_MULW(b2p[TMP_REG_1], src_reg,
390 						b2p[TMP_REG_1]));
391 				EMIT(PPC_RAW_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]));
392 			} else
393 				EMIT(PPC_RAW_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 				EMIT(PPC_RAW_DIVDU(b2p[TMP_REG_1], dst_reg, src_reg));
399 				EMIT(PPC_RAW_MULD(b2p[TMP_REG_1], src_reg,
400 						b2p[TMP_REG_1]));
401 				EMIT(PPC_RAW_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]));
402 			} else
403 				EMIT(PPC_RAW_DIVDU(dst_reg, dst_reg, src_reg));
404 			break;
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 */
409 			if (imm == 0)
410 				return -EINVAL;
411 			else if (imm == 1)
412 				goto bpf_alu32_trunc;
413 
414 			PPC_LI32(b2p[TMP_REG_1], imm);
415 			switch (BPF_CLASS(code)) {
416 			case BPF_ALU:
417 				if (BPF_OP(code) == BPF_MOD) {
418 					EMIT(PPC_RAW_DIVWU(b2p[TMP_REG_2],
419 							dst_reg,
420 							b2p[TMP_REG_1]));
421 					EMIT(PPC_RAW_MULW(b2p[TMP_REG_1],
422 							b2p[TMP_REG_1],
423 							b2p[TMP_REG_2]));
424 					EMIT(PPC_RAW_SUB(dst_reg, dst_reg,
425 							b2p[TMP_REG_1]));
426 				} else
427 					EMIT(PPC_RAW_DIVWU(dst_reg, dst_reg,
428 							b2p[TMP_REG_1]));
429 				break;
430 			case BPF_ALU64:
431 				if (BPF_OP(code) == BPF_MOD) {
432 					EMIT(PPC_RAW_DIVDU(b2p[TMP_REG_2],
433 							dst_reg,
434 							b2p[TMP_REG_1]));
435 					EMIT(PPC_RAW_MULD(b2p[TMP_REG_1],
436 							b2p[TMP_REG_1],
437 							b2p[TMP_REG_2]));
438 					EMIT(PPC_RAW_SUB(dst_reg, dst_reg,
439 							b2p[TMP_REG_1]));
440 				} else
441 					EMIT(PPC_RAW_DIVDU(dst_reg, dst_reg,
442 							b2p[TMP_REG_1]));
443 				break;
444 			}
445 			goto bpf_alu32_trunc;
446 		case BPF_ALU | BPF_NEG: /* (u32) dst = -dst */
447 		case BPF_ALU64 | BPF_NEG: /* dst = -dst */
448 			EMIT(PPC_RAW_NEG(dst_reg, dst_reg));
449 			goto bpf_alu32_trunc;
450 
451 		/*
452 		 * Logical operations: AND/OR/XOR/[A]LSH/[A]RSH
453 		 */
454 		case BPF_ALU | BPF_AND | BPF_X: /* (u32) dst = dst & src */
455 		case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
456 			EMIT(PPC_RAW_AND(dst_reg, dst_reg, src_reg));
457 			goto bpf_alu32_trunc;
458 		case BPF_ALU | BPF_AND | BPF_K: /* (u32) dst = dst & imm */
459 		case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
460 			if (!IMM_H(imm))
461 				EMIT(PPC_RAW_ANDI(dst_reg, dst_reg, IMM_L(imm)));
462 			else {
463 				/* Sign-extended */
464 				PPC_LI32(b2p[TMP_REG_1], imm);
465 				EMIT(PPC_RAW_AND(dst_reg, dst_reg, b2p[TMP_REG_1]));
466 			}
467 			goto bpf_alu32_trunc;
468 		case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
469 		case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
470 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, src_reg));
471 			goto bpf_alu32_trunc;
472 		case BPF_ALU | BPF_OR | BPF_K:/* dst = (u32) dst | (u32) imm */
473 		case BPF_ALU64 | BPF_OR | BPF_K:/* dst = dst | imm */
474 			if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
475 				/* Sign-extended */
476 				PPC_LI32(b2p[TMP_REG_1], imm);
477 				EMIT(PPC_RAW_OR(dst_reg, dst_reg, b2p[TMP_REG_1]));
478 			} else {
479 				if (IMM_L(imm))
480 					EMIT(PPC_RAW_ORI(dst_reg, dst_reg, IMM_L(imm)));
481 				if (IMM_H(imm))
482 					EMIT(PPC_RAW_ORIS(dst_reg, dst_reg, IMM_H(imm)));
483 			}
484 			goto bpf_alu32_trunc;
485 		case BPF_ALU | BPF_XOR | BPF_X: /* (u32) dst ^= src */
486 		case BPF_ALU64 | BPF_XOR | BPF_X: /* dst ^= src */
487 			EMIT(PPC_RAW_XOR(dst_reg, dst_reg, src_reg));
488 			goto bpf_alu32_trunc;
489 		case BPF_ALU | BPF_XOR | BPF_K: /* (u32) dst ^= (u32) imm */
490 		case BPF_ALU64 | BPF_XOR | BPF_K: /* dst ^= imm */
491 			if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
492 				/* Sign-extended */
493 				PPC_LI32(b2p[TMP_REG_1], imm);
494 				EMIT(PPC_RAW_XOR(dst_reg, dst_reg, b2p[TMP_REG_1]));
495 			} else {
496 				if (IMM_L(imm))
497 					EMIT(PPC_RAW_XORI(dst_reg, dst_reg, IMM_L(imm)));
498 				if (IMM_H(imm))
499 					EMIT(PPC_RAW_XORIS(dst_reg, dst_reg, IMM_H(imm)));
500 			}
501 			goto bpf_alu32_trunc;
502 		case BPF_ALU | BPF_LSH | BPF_X: /* (u32) dst <<= (u32) src */
503 			/* slw clears top 32 bits */
504 			EMIT(PPC_RAW_SLW(dst_reg, dst_reg, src_reg));
505 			/* skip zero extension move, but set address map. */
506 			if (insn_is_zext(&insn[i + 1]))
507 				addrs[++i] = ctx->idx * 4;
508 			break;
509 		case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */
510 			EMIT(PPC_RAW_SLD(dst_reg, dst_reg, src_reg));
511 			break;
512 		case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<== (u32) imm */
513 			/* with imm 0, we still need to clear top 32 bits */
514 			EMIT(PPC_RAW_SLWI(dst_reg, dst_reg, imm));
515 			if (insn_is_zext(&insn[i + 1]))
516 				addrs[++i] = ctx->idx * 4;
517 			break;
518 		case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<== imm */
519 			if (imm != 0)
520 				EMIT(PPC_RAW_SLDI(dst_reg, dst_reg, imm));
521 			break;
522 		case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */
523 			EMIT(PPC_RAW_SRW(dst_reg, dst_reg, src_reg));
524 			if (insn_is_zext(&insn[i + 1]))
525 				addrs[++i] = ctx->idx * 4;
526 			break;
527 		case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */
528 			EMIT(PPC_RAW_SRD(dst_reg, dst_reg, src_reg));
529 			break;
530 		case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */
531 			EMIT(PPC_RAW_SRWI(dst_reg, dst_reg, imm));
532 			if (insn_is_zext(&insn[i + 1]))
533 				addrs[++i] = ctx->idx * 4;
534 			break;
535 		case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */
536 			if (imm != 0)
537 				EMIT(PPC_RAW_SRDI(dst_reg, dst_reg, imm));
538 			break;
539 		case BPF_ALU | BPF_ARSH | BPF_X: /* (s32) dst >>= src */
540 			EMIT(PPC_RAW_SRAW(dst_reg, dst_reg, src_reg));
541 			goto bpf_alu32_trunc;
542 		case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */
543 			EMIT(PPC_RAW_SRAD(dst_reg, dst_reg, src_reg));
544 			break;
545 		case BPF_ALU | BPF_ARSH | BPF_K: /* (s32) dst >>= imm */
546 			EMIT(PPC_RAW_SRAWI(dst_reg, dst_reg, imm));
547 			goto bpf_alu32_trunc;
548 		case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */
549 			if (imm != 0)
550 				EMIT(PPC_RAW_SRADI(dst_reg, dst_reg, imm));
551 			break;
552 
553 		/*
554 		 * MOV
555 		 */
556 		case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
557 		case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
558 			if (imm == 1) {
559 				/* special mov32 for zext */
560 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 0, 31));
561 				break;
562 			}
563 			EMIT(PPC_RAW_MR(dst_reg, src_reg));
564 			goto bpf_alu32_trunc;
565 		case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */
566 		case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */
567 			PPC_LI32(dst_reg, imm);
568 			if (imm < 0)
569 				goto bpf_alu32_trunc;
570 			else if (insn_is_zext(&insn[i + 1]))
571 				addrs[++i] = ctx->idx * 4;
572 			break;
573 
574 bpf_alu32_trunc:
575 		/* Truncate to 32-bits */
576 		if (BPF_CLASS(code) == BPF_ALU && !fp->aux->verifier_zext)
577 			EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 0, 31));
578 		break;
579 
580 		/*
581 		 * BPF_FROM_BE/LE
582 		 */
583 		case BPF_ALU | BPF_END | BPF_FROM_LE:
584 		case BPF_ALU | BPF_END | BPF_FROM_BE:
585 #ifdef __BIG_ENDIAN__
586 			if (BPF_SRC(code) == BPF_FROM_BE)
587 				goto emit_clear;
588 #else /* !__BIG_ENDIAN__ */
589 			if (BPF_SRC(code) == BPF_FROM_LE)
590 				goto emit_clear;
591 #endif
592 			switch (imm) {
593 			case 16:
594 				/* Rotate 8 bits left & mask with 0x0000ff00 */
595 				EMIT(PPC_RAW_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 16, 23));
596 				/* Rotate 8 bits right & insert LSB to reg */
597 				EMIT(PPC_RAW_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 24, 31));
598 				/* Move result back to dst_reg */
599 				EMIT(PPC_RAW_MR(dst_reg, b2p[TMP_REG_1]));
600 				break;
601 			case 32:
602 				/*
603 				 * Rotate word left by 8 bits:
604 				 * 2 bytes are already in their final position
605 				 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4)
606 				 */
607 				EMIT(PPC_RAW_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 0, 31));
608 				/* Rotate 24 bits and insert byte 1 */
609 				EMIT(PPC_RAW_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 0, 7));
610 				/* Rotate 24 bits and insert byte 3 */
611 				EMIT(PPC_RAW_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 16, 23));
612 				EMIT(PPC_RAW_MR(dst_reg, b2p[TMP_REG_1]));
613 				break;
614 			case 64:
615 				/*
616 				 * Way easier and faster(?) to store the value
617 				 * into stack and then use ldbrx
618 				 *
619 				 * ctx->seen will be reliable in pass2, but
620 				 * the instructions generated will remain the
621 				 * same across all passes
622 				 */
623 				PPC_BPF_STL(dst_reg, 1, bpf_jit_stack_local(ctx));
624 				EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], 1, bpf_jit_stack_local(ctx)));
625 				EMIT(PPC_RAW_LDBRX(dst_reg, 0, b2p[TMP_REG_1]));
626 				break;
627 			}
628 			break;
629 
630 emit_clear:
631 			switch (imm) {
632 			case 16:
633 				/* zero-extend 16 bits into 64 bits */
634 				EMIT(PPC_RAW_RLDICL(dst_reg, dst_reg, 0, 48));
635 				if (insn_is_zext(&insn[i + 1]))
636 					addrs[++i] = ctx->idx * 4;
637 				break;
638 			case 32:
639 				if (!fp->aux->verifier_zext)
640 					/* zero-extend 32 bits into 64 bits */
641 					EMIT(PPC_RAW_RLDICL(dst_reg, dst_reg, 0, 32));
642 				break;
643 			case 64:
644 				/* nop */
645 				break;
646 			}
647 			break;
648 
649 		/*
650 		 * BPF_ST(X)
651 		 */
652 		case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */
653 		case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
654 			if (BPF_CLASS(code) == BPF_ST) {
655 				EMIT(PPC_RAW_LI(b2p[TMP_REG_1], imm));
656 				src_reg = b2p[TMP_REG_1];
657 			}
658 			EMIT(PPC_RAW_STB(src_reg, dst_reg, off));
659 			break;
660 		case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
661 		case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
662 			if (BPF_CLASS(code) == BPF_ST) {
663 				EMIT(PPC_RAW_LI(b2p[TMP_REG_1], imm));
664 				src_reg = b2p[TMP_REG_1];
665 			}
666 			EMIT(PPC_RAW_STH(src_reg, dst_reg, off));
667 			break;
668 		case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
669 		case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
670 			if (BPF_CLASS(code) == BPF_ST) {
671 				PPC_LI32(b2p[TMP_REG_1], imm);
672 				src_reg = b2p[TMP_REG_1];
673 			}
674 			EMIT(PPC_RAW_STW(src_reg, dst_reg, off));
675 			break;
676 		case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
677 		case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
678 			if (BPF_CLASS(code) == BPF_ST) {
679 				PPC_LI32(b2p[TMP_REG_1], imm);
680 				src_reg = b2p[TMP_REG_1];
681 			}
682 			PPC_BPF_STL(src_reg, dst_reg, off);
683 			break;
684 
685 		/*
686 		 * BPF_STX XADD (atomic_add)
687 		 */
688 		/* *(u32 *)(dst + off) += src */
689 		case BPF_STX | BPF_XADD | BPF_W:
690 			/* Get EA into TMP_REG_1 */
691 			EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], dst_reg, off));
692 			tmp_idx = ctx->idx * 4;
693 			/* load value from memory into TMP_REG_2 */
694 			EMIT(PPC_RAW_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0));
695 			/* add value from src_reg into this */
696 			EMIT(PPC_RAW_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg));
697 			/* store result back */
698 			EMIT(PPC_RAW_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]));
699 			/* we're done if this succeeded */
700 			PPC_BCC_SHORT(COND_NE, tmp_idx);
701 			break;
702 		/* *(u64 *)(dst + off) += src */
703 		case BPF_STX | BPF_XADD | BPF_DW:
704 			EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], dst_reg, off));
705 			tmp_idx = ctx->idx * 4;
706 			EMIT(PPC_RAW_LDARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0));
707 			EMIT(PPC_RAW_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg));
708 			EMIT(PPC_RAW_STDCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]));
709 			PPC_BCC_SHORT(COND_NE, tmp_idx);
710 			break;
711 
712 		/*
713 		 * BPF_LDX
714 		 */
715 		/* dst = *(u8 *)(ul) (src + off) */
716 		case BPF_LDX | BPF_MEM | BPF_B:
717 			EMIT(PPC_RAW_LBZ(dst_reg, src_reg, off));
718 			if (insn_is_zext(&insn[i + 1]))
719 				addrs[++i] = ctx->idx * 4;
720 			break;
721 		/* dst = *(u16 *)(ul) (src + off) */
722 		case BPF_LDX | BPF_MEM | BPF_H:
723 			EMIT(PPC_RAW_LHZ(dst_reg, src_reg, off));
724 			if (insn_is_zext(&insn[i + 1]))
725 				addrs[++i] = ctx->idx * 4;
726 			break;
727 		/* dst = *(u32 *)(ul) (src + off) */
728 		case BPF_LDX | BPF_MEM | BPF_W:
729 			EMIT(PPC_RAW_LWZ(dst_reg, src_reg, off));
730 			if (insn_is_zext(&insn[i + 1]))
731 				addrs[++i] = ctx->idx * 4;
732 			break;
733 		/* dst = *(u64 *)(ul) (src + off) */
734 		case BPF_LDX | BPF_MEM | BPF_DW:
735 			PPC_BPF_LL(dst_reg, src_reg, off);
736 			break;
737 
738 		/*
739 		 * Doubleword load
740 		 * 16 byte instruction that uses two 'struct bpf_insn'
741 		 */
742 		case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
743 			imm64 = ((u64)(u32) insn[i].imm) |
744 				    (((u64)(u32) insn[i+1].imm) << 32);
745 			/* Adjust for two bpf instructions */
746 			addrs[++i] = ctx->idx * 4;
747 			PPC_LI64(dst_reg, imm64);
748 			break;
749 
750 		/*
751 		 * Return/Exit
752 		 */
753 		case BPF_JMP | BPF_EXIT:
754 			/*
755 			 * If this isn't the very last instruction, branch to
756 			 * the epilogue. If we _are_ the last instruction,
757 			 * we'll just fall through to the epilogue.
758 			 */
759 			if (i != flen - 1)
760 				PPC_JMP(exit_addr);
761 			/* else fall through to the epilogue */
762 			break;
763 
764 		/*
765 		 * Call kernel helper or bpf function
766 		 */
767 		case BPF_JMP | BPF_CALL:
768 			ctx->seen |= SEEN_FUNC;
769 
770 			ret = bpf_jit_get_func_addr(fp, &insn[i], extra_pass,
771 						    &func_addr, &func_addr_fixed);
772 			if (ret < 0)
773 				return ret;
774 
775 			if (func_addr_fixed)
776 				bpf_jit_emit_func_call_hlp(image, ctx, func_addr);
777 			else
778 				bpf_jit_emit_func_call_rel(image, ctx, func_addr);
779 			/* move return value from r3 to BPF_REG_0 */
780 			EMIT(PPC_RAW_MR(b2p[BPF_REG_0], 3));
781 			break;
782 
783 		/*
784 		 * Jumps and branches
785 		 */
786 		case BPF_JMP | BPF_JA:
787 			PPC_JMP(addrs[i + 1 + off]);
788 			break;
789 
790 		case BPF_JMP | BPF_JGT | BPF_K:
791 		case BPF_JMP | BPF_JGT | BPF_X:
792 		case BPF_JMP | BPF_JSGT | BPF_K:
793 		case BPF_JMP | BPF_JSGT | BPF_X:
794 		case BPF_JMP32 | BPF_JGT | BPF_K:
795 		case BPF_JMP32 | BPF_JGT | BPF_X:
796 		case BPF_JMP32 | BPF_JSGT | BPF_K:
797 		case BPF_JMP32 | BPF_JSGT | BPF_X:
798 			true_cond = COND_GT;
799 			goto cond_branch;
800 		case BPF_JMP | BPF_JLT | BPF_K:
801 		case BPF_JMP | BPF_JLT | BPF_X:
802 		case BPF_JMP | BPF_JSLT | BPF_K:
803 		case BPF_JMP | BPF_JSLT | BPF_X:
804 		case BPF_JMP32 | BPF_JLT | BPF_K:
805 		case BPF_JMP32 | BPF_JLT | BPF_X:
806 		case BPF_JMP32 | BPF_JSLT | BPF_K:
807 		case BPF_JMP32 | BPF_JSLT | BPF_X:
808 			true_cond = COND_LT;
809 			goto cond_branch;
810 		case BPF_JMP | BPF_JGE | BPF_K:
811 		case BPF_JMP | BPF_JGE | BPF_X:
812 		case BPF_JMP | BPF_JSGE | BPF_K:
813 		case BPF_JMP | BPF_JSGE | BPF_X:
814 		case BPF_JMP32 | BPF_JGE | BPF_K:
815 		case BPF_JMP32 | BPF_JGE | BPF_X:
816 		case BPF_JMP32 | BPF_JSGE | BPF_K:
817 		case BPF_JMP32 | BPF_JSGE | BPF_X:
818 			true_cond = COND_GE;
819 			goto cond_branch;
820 		case BPF_JMP | BPF_JLE | BPF_K:
821 		case BPF_JMP | BPF_JLE | BPF_X:
822 		case BPF_JMP | BPF_JSLE | BPF_K:
823 		case BPF_JMP | BPF_JSLE | BPF_X:
824 		case BPF_JMP32 | BPF_JLE | BPF_K:
825 		case BPF_JMP32 | BPF_JLE | BPF_X:
826 		case BPF_JMP32 | BPF_JSLE | BPF_K:
827 		case BPF_JMP32 | BPF_JSLE | BPF_X:
828 			true_cond = COND_LE;
829 			goto cond_branch;
830 		case BPF_JMP | BPF_JEQ | BPF_K:
831 		case BPF_JMP | BPF_JEQ | BPF_X:
832 		case BPF_JMP32 | BPF_JEQ | BPF_K:
833 		case BPF_JMP32 | BPF_JEQ | BPF_X:
834 			true_cond = COND_EQ;
835 			goto cond_branch;
836 		case BPF_JMP | BPF_JNE | BPF_K:
837 		case BPF_JMP | BPF_JNE | BPF_X:
838 		case BPF_JMP32 | BPF_JNE | BPF_K:
839 		case BPF_JMP32 | BPF_JNE | BPF_X:
840 			true_cond = COND_NE;
841 			goto cond_branch;
842 		case BPF_JMP | BPF_JSET | BPF_K:
843 		case BPF_JMP | BPF_JSET | BPF_X:
844 		case BPF_JMP32 | BPF_JSET | BPF_K:
845 		case BPF_JMP32 | BPF_JSET | BPF_X:
846 			true_cond = COND_NE;
847 			/* Fall through */
848 
849 cond_branch:
850 			switch (code) {
851 			case BPF_JMP | BPF_JGT | BPF_X:
852 			case BPF_JMP | BPF_JLT | BPF_X:
853 			case BPF_JMP | BPF_JGE | BPF_X:
854 			case BPF_JMP | BPF_JLE | BPF_X:
855 			case BPF_JMP | BPF_JEQ | BPF_X:
856 			case BPF_JMP | BPF_JNE | BPF_X:
857 			case BPF_JMP32 | BPF_JGT | BPF_X:
858 			case BPF_JMP32 | BPF_JLT | BPF_X:
859 			case BPF_JMP32 | BPF_JGE | BPF_X:
860 			case BPF_JMP32 | BPF_JLE | BPF_X:
861 			case BPF_JMP32 | BPF_JEQ | BPF_X:
862 			case BPF_JMP32 | BPF_JNE | BPF_X:
863 				/* unsigned comparison */
864 				if (BPF_CLASS(code) == BPF_JMP32)
865 					EMIT(PPC_RAW_CMPLW(dst_reg, src_reg));
866 				else
867 					EMIT(PPC_RAW_CMPLD(dst_reg, src_reg));
868 				break;
869 			case BPF_JMP | BPF_JSGT | BPF_X:
870 			case BPF_JMP | BPF_JSLT | BPF_X:
871 			case BPF_JMP | BPF_JSGE | BPF_X:
872 			case BPF_JMP | BPF_JSLE | BPF_X:
873 			case BPF_JMP32 | BPF_JSGT | BPF_X:
874 			case BPF_JMP32 | BPF_JSLT | BPF_X:
875 			case BPF_JMP32 | BPF_JSGE | BPF_X:
876 			case BPF_JMP32 | BPF_JSLE | BPF_X:
877 				/* signed comparison */
878 				if (BPF_CLASS(code) == BPF_JMP32)
879 					EMIT(PPC_RAW_CMPW(dst_reg, src_reg));
880 				else
881 					EMIT(PPC_RAW_CMPD(dst_reg, src_reg));
882 				break;
883 			case BPF_JMP | BPF_JSET | BPF_X:
884 			case BPF_JMP32 | BPF_JSET | BPF_X:
885 				if (BPF_CLASS(code) == BPF_JMP) {
886 					EMIT(PPC_RAW_AND_DOT(b2p[TMP_REG_1], dst_reg,
887 						    src_reg));
888 				} else {
889 					int tmp_reg = b2p[TMP_REG_1];
890 
891 					EMIT(PPC_RAW_AND(tmp_reg, dst_reg, src_reg));
892 					EMIT(PPC_RAW_RLWINM_DOT(tmp_reg, tmp_reg, 0, 0,
893 						       31));
894 				}
895 				break;
896 			case BPF_JMP | BPF_JNE | BPF_K:
897 			case BPF_JMP | BPF_JEQ | BPF_K:
898 			case BPF_JMP | BPF_JGT | BPF_K:
899 			case BPF_JMP | BPF_JLT | BPF_K:
900 			case BPF_JMP | BPF_JGE | BPF_K:
901 			case BPF_JMP | BPF_JLE | BPF_K:
902 			case BPF_JMP32 | BPF_JNE | BPF_K:
903 			case BPF_JMP32 | BPF_JEQ | BPF_K:
904 			case BPF_JMP32 | BPF_JGT | BPF_K:
905 			case BPF_JMP32 | BPF_JLT | BPF_K:
906 			case BPF_JMP32 | BPF_JGE | BPF_K:
907 			case BPF_JMP32 | BPF_JLE | BPF_K:
908 			{
909 				bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
910 
911 				/*
912 				 * Need sign-extended load, so only positive
913 				 * values can be used as imm in cmpldi
914 				 */
915 				if (imm >= 0 && imm < 32768) {
916 					if (is_jmp32)
917 						EMIT(PPC_RAW_CMPLWI(dst_reg, imm));
918 					else
919 						EMIT(PPC_RAW_CMPLDI(dst_reg, imm));
920 				} else {
921 					/* sign-extending load */
922 					PPC_LI32(b2p[TMP_REG_1], imm);
923 					/* ... but unsigned comparison */
924 					if (is_jmp32)
925 						EMIT(PPC_RAW_CMPLW(dst_reg,
926 							  b2p[TMP_REG_1]));
927 					else
928 						EMIT(PPC_RAW_CMPLD(dst_reg,
929 							  b2p[TMP_REG_1]));
930 				}
931 				break;
932 			}
933 			case BPF_JMP | BPF_JSGT | BPF_K:
934 			case BPF_JMP | BPF_JSLT | BPF_K:
935 			case BPF_JMP | BPF_JSGE | BPF_K:
936 			case BPF_JMP | BPF_JSLE | BPF_K:
937 			case BPF_JMP32 | BPF_JSGT | BPF_K:
938 			case BPF_JMP32 | BPF_JSLT | BPF_K:
939 			case BPF_JMP32 | BPF_JSGE | BPF_K:
940 			case BPF_JMP32 | BPF_JSLE | BPF_K:
941 			{
942 				bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
943 
944 				/*
945 				 * signed comparison, so any 16-bit value
946 				 * can be used in cmpdi
947 				 */
948 				if (imm >= -32768 && imm < 32768) {
949 					if (is_jmp32)
950 						EMIT(PPC_RAW_CMPWI(dst_reg, imm));
951 					else
952 						EMIT(PPC_RAW_CMPDI(dst_reg, imm));
953 				} else {
954 					PPC_LI32(b2p[TMP_REG_1], imm);
955 					if (is_jmp32)
956 						EMIT(PPC_RAW_CMPW(dst_reg,
957 							 b2p[TMP_REG_1]));
958 					else
959 						EMIT(PPC_RAW_CMPD(dst_reg,
960 							 b2p[TMP_REG_1]));
961 				}
962 				break;
963 			}
964 			case BPF_JMP | BPF_JSET | BPF_K:
965 			case BPF_JMP32 | BPF_JSET | BPF_K:
966 				/* andi does not sign-extend the immediate */
967 				if (imm >= 0 && imm < 32768)
968 					/* PPC_ANDI is _only/always_ dot-form */
969 					EMIT(PPC_RAW_ANDI(b2p[TMP_REG_1], dst_reg, imm));
970 				else {
971 					int tmp_reg = b2p[TMP_REG_1];
972 
973 					PPC_LI32(tmp_reg, imm);
974 					if (BPF_CLASS(code) == BPF_JMP) {
975 						EMIT(PPC_RAW_AND_DOT(tmp_reg, dst_reg,
976 							    tmp_reg));
977 					} else {
978 						EMIT(PPC_RAW_AND(tmp_reg, dst_reg,
979 							tmp_reg));
980 						EMIT(PPC_RAW_RLWINM_DOT(tmp_reg, tmp_reg,
981 							       0, 0, 31));
982 					}
983 				}
984 				break;
985 			}
986 			PPC_BCC(true_cond, addrs[i + 1 + off]);
987 			break;
988 
989 		/*
990 		 * Tail call
991 		 */
992 		case BPF_JMP | BPF_TAIL_CALL:
993 			ctx->seen |= SEEN_TAILCALL;
994 			bpf_jit_emit_tail_call(image, ctx, addrs[i + 1]);
995 			break;
996 
997 		default:
998 			/*
999 			 * The filter contains something cruel & unusual.
1000 			 * We don't handle it, but also there shouldn't be
1001 			 * anything missing from our list.
1002 			 */
1003 			pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n",
1004 					code, i);
1005 			return -ENOTSUPP;
1006 		}
1007 	}
1008 
1009 	/* Set end-of-body-code address for exit. */
1010 	addrs[i] = ctx->idx * 4;
1011 
1012 	return 0;
1013 }
1014 
1015 /* Fix the branch target addresses for subprog calls */
1016 static int bpf_jit_fixup_subprog_calls(struct bpf_prog *fp, u32 *image,
1017 				       struct codegen_context *ctx, u32 *addrs)
1018 {
1019 	const struct bpf_insn *insn = fp->insnsi;
1020 	bool func_addr_fixed;
1021 	u64 func_addr;
1022 	u32 tmp_idx;
1023 	int i, ret;
1024 
1025 	for (i = 0; i < fp->len; i++) {
1026 		/*
1027 		 * During the extra pass, only the branch target addresses for
1028 		 * the subprog calls need to be fixed. All other instructions
1029 		 * can left untouched.
1030 		 *
1031 		 * The JITed image length does not change because we already
1032 		 * ensure that the JITed instruction sequence for these calls
1033 		 * are of fixed length by padding them with NOPs.
1034 		 */
1035 		if (insn[i].code == (BPF_JMP | BPF_CALL) &&
1036 		    insn[i].src_reg == BPF_PSEUDO_CALL) {
1037 			ret = bpf_jit_get_func_addr(fp, &insn[i], true,
1038 						    &func_addr,
1039 						    &func_addr_fixed);
1040 			if (ret < 0)
1041 				return ret;
1042 
1043 			/*
1044 			 * Save ctx->idx as this would currently point to the
1045 			 * end of the JITed image and set it to the offset of
1046 			 * the instruction sequence corresponding to the
1047 			 * subprog call temporarily.
1048 			 */
1049 			tmp_idx = ctx->idx;
1050 			ctx->idx = addrs[i] / 4;
1051 			bpf_jit_emit_func_call_rel(image, ctx, func_addr);
1052 
1053 			/*
1054 			 * Restore ctx->idx here. This is safe as the length
1055 			 * of the JITed sequence remains unchanged.
1056 			 */
1057 			ctx->idx = tmp_idx;
1058 		}
1059 	}
1060 
1061 	return 0;
1062 }
1063 
1064 struct powerpc64_jit_data {
1065 	struct bpf_binary_header *header;
1066 	u32 *addrs;
1067 	u8 *image;
1068 	u32 proglen;
1069 	struct codegen_context ctx;
1070 };
1071 
1072 bool bpf_jit_needs_zext(void)
1073 {
1074 	return true;
1075 }
1076 
1077 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
1078 {
1079 	u32 proglen;
1080 	u32 alloclen;
1081 	u8 *image = NULL;
1082 	u32 *code_base;
1083 	u32 *addrs;
1084 	struct powerpc64_jit_data *jit_data;
1085 	struct codegen_context cgctx;
1086 	int pass;
1087 	int flen;
1088 	struct bpf_binary_header *bpf_hdr;
1089 	struct bpf_prog *org_fp = fp;
1090 	struct bpf_prog *tmp_fp;
1091 	bool bpf_blinded = false;
1092 	bool extra_pass = false;
1093 
1094 	if (!fp->jit_requested)
1095 		return org_fp;
1096 
1097 	tmp_fp = bpf_jit_blind_constants(org_fp);
1098 	if (IS_ERR(tmp_fp))
1099 		return org_fp;
1100 
1101 	if (tmp_fp != org_fp) {
1102 		bpf_blinded = true;
1103 		fp = tmp_fp;
1104 	}
1105 
1106 	jit_data = fp->aux->jit_data;
1107 	if (!jit_data) {
1108 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1109 		if (!jit_data) {
1110 			fp = org_fp;
1111 			goto out;
1112 		}
1113 		fp->aux->jit_data = jit_data;
1114 	}
1115 
1116 	flen = fp->len;
1117 	addrs = jit_data->addrs;
1118 	if (addrs) {
1119 		cgctx = jit_data->ctx;
1120 		image = jit_data->image;
1121 		bpf_hdr = jit_data->header;
1122 		proglen = jit_data->proglen;
1123 		alloclen = proglen + FUNCTION_DESCR_SIZE;
1124 		extra_pass = true;
1125 		goto skip_init_ctx;
1126 	}
1127 
1128 	addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
1129 	if (addrs == NULL) {
1130 		fp = org_fp;
1131 		goto out_addrs;
1132 	}
1133 
1134 	memset(&cgctx, 0, sizeof(struct codegen_context));
1135 
1136 	/* Make sure that the stack is quadword aligned. */
1137 	cgctx.stack_size = round_up(fp->aux->stack_depth, 16);
1138 
1139 	/* Scouting faux-generate pass 0 */
1140 	if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
1141 		/* We hit something illegal or unsupported. */
1142 		fp = org_fp;
1143 		goto out_addrs;
1144 	}
1145 
1146 	/*
1147 	 * If we have seen a tail call, we need a second pass.
1148 	 * This is because bpf_jit_emit_common_epilogue() is called
1149 	 * from bpf_jit_emit_tail_call() with a not yet stable ctx->seen.
1150 	 */
1151 	if (cgctx.seen & SEEN_TAILCALL) {
1152 		cgctx.idx = 0;
1153 		if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
1154 			fp = org_fp;
1155 			goto out_addrs;
1156 		}
1157 	}
1158 
1159 	/*
1160 	 * Pretend to build prologue, given the features we've seen.  This will
1161 	 * update ctgtx.idx as it pretends to output instructions, then we can
1162 	 * calculate total size from idx.
1163 	 */
1164 	bpf_jit_build_prologue(0, &cgctx);
1165 	bpf_jit_build_epilogue(0, &cgctx);
1166 
1167 	proglen = cgctx.idx * 4;
1168 	alloclen = proglen + FUNCTION_DESCR_SIZE;
1169 
1170 	bpf_hdr = bpf_jit_binary_alloc(alloclen, &image, 4,
1171 			bpf_jit_fill_ill_insns);
1172 	if (!bpf_hdr) {
1173 		fp = org_fp;
1174 		goto out_addrs;
1175 	}
1176 
1177 skip_init_ctx:
1178 	code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
1179 
1180 	if (extra_pass) {
1181 		/*
1182 		 * Do not touch the prologue and epilogue as they will remain
1183 		 * unchanged. Only fix the branch target address for subprog
1184 		 * calls in the body.
1185 		 *
1186 		 * This does not change the offsets and lengths of the subprog
1187 		 * call instruction sequences and hence, the size of the JITed
1188 		 * image as well.
1189 		 */
1190 		bpf_jit_fixup_subprog_calls(fp, code_base, &cgctx, addrs);
1191 
1192 		/* There is no need to perform the usual passes. */
1193 		goto skip_codegen_passes;
1194 	}
1195 
1196 	/* Code generation passes 1-2 */
1197 	for (pass = 1; pass < 3; pass++) {
1198 		/* Now build the prologue, body code & epilogue for real. */
1199 		cgctx.idx = 0;
1200 		bpf_jit_build_prologue(code_base, &cgctx);
1201 		bpf_jit_build_body(fp, code_base, &cgctx, addrs, extra_pass);
1202 		bpf_jit_build_epilogue(code_base, &cgctx);
1203 
1204 		if (bpf_jit_enable > 1)
1205 			pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
1206 				proglen - (cgctx.idx * 4), cgctx.seen);
1207 	}
1208 
1209 skip_codegen_passes:
1210 	if (bpf_jit_enable > 1)
1211 		/*
1212 		 * Note that we output the base address of the code_base
1213 		 * rather than image, since opcodes are in code_base.
1214 		 */
1215 		bpf_jit_dump(flen, proglen, pass, code_base);
1216 
1217 #ifdef PPC64_ELF_ABI_v1
1218 	/* Function descriptor nastiness: Address + TOC */
1219 	((u64 *)image)[0] = (u64)code_base;
1220 	((u64 *)image)[1] = local_paca->kernel_toc;
1221 #endif
1222 
1223 	fp->bpf_func = (void *)image;
1224 	fp->jited = 1;
1225 	fp->jited_len = alloclen;
1226 
1227 	bpf_flush_icache(bpf_hdr, (u8 *)bpf_hdr + (bpf_hdr->pages * PAGE_SIZE));
1228 	if (!fp->is_func || extra_pass) {
1229 		bpf_prog_fill_jited_linfo(fp, addrs);
1230 out_addrs:
1231 		kfree(addrs);
1232 		kfree(jit_data);
1233 		fp->aux->jit_data = NULL;
1234 	} else {
1235 		jit_data->addrs = addrs;
1236 		jit_data->ctx = cgctx;
1237 		jit_data->proglen = proglen;
1238 		jit_data->image = image;
1239 		jit_data->header = bpf_hdr;
1240 	}
1241 
1242 out:
1243 	if (bpf_blinded)
1244 		bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
1245 
1246 	return fp;
1247 }
1248 
1249 /* Overriding bpf_jit_free() as we don't set images read-only. */
1250 void bpf_jit_free(struct bpf_prog *fp)
1251 {
1252 	unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
1253 	struct bpf_binary_header *bpf_hdr = (void *)addr;
1254 
1255 	if (fp->jited)
1256 		bpf_jit_binary_free(bpf_hdr);
1257 
1258 	bpf_prog_unlock_free(fp);
1259 }
1260