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 #include <asm/security_features.h>
19 
20 #include "bpf_jit64.h"
21 
22 static inline bool bpf_has_stack_frame(struct codegen_context *ctx)
23 {
24 	/*
25 	 * We only need a stack frame if:
26 	 * - we call other functions (kernel helpers), or
27 	 * - the bpf program uses its stack area
28 	 * The latter condition is deduced from the usage of BPF_REG_FP
29 	 */
30 	return ctx->seen & SEEN_FUNC || bpf_is_seen_register(ctx, b2p[BPF_REG_FP]);
31 }
32 
33 /*
34  * When not setting up our own stackframe, the redzone usage is:
35  *
36  *		[	prev sp		] <-------------
37  *		[	  ...       	] 		|
38  * sp (r1) --->	[    stack pointer	] --------------
39  *		[   nv gpr save area	] 5*8
40  *		[    tail_call_cnt	] 8
41  *		[    local_tmp_var	] 16
42  *		[   unused red zone	] 208 bytes protected
43  */
44 static int bpf_jit_stack_local(struct codegen_context *ctx)
45 {
46 	if (bpf_has_stack_frame(ctx))
47 		return STACK_FRAME_MIN_SIZE + ctx->stack_size;
48 	else
49 		return -(BPF_PPC_STACK_SAVE + 24);
50 }
51 
52 static int bpf_jit_stack_tailcallcnt(struct codegen_context *ctx)
53 {
54 	return bpf_jit_stack_local(ctx) + 16;
55 }
56 
57 static int bpf_jit_stack_offsetof(struct codegen_context *ctx, int reg)
58 {
59 	if (reg >= BPF_PPC_NVR_MIN && reg < 32)
60 		return (bpf_has_stack_frame(ctx) ?
61 			(BPF_PPC_STACKFRAME + ctx->stack_size) : 0)
62 				- (8 * (32 - reg));
63 
64 	pr_err("BPF JIT is asking about unknown registers");
65 	BUG();
66 }
67 
68 void bpf_jit_realloc_regs(struct codegen_context *ctx)
69 {
70 }
71 
72 void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx)
73 {
74 	int i;
75 
76 	/*
77 	 * Initialize tail_call_cnt if we do tail calls.
78 	 * Otherwise, put in NOPs so that it can be skipped when we are
79 	 * invoked through a tail call.
80 	 */
81 	if (ctx->seen & SEEN_TAILCALL) {
82 		EMIT(PPC_RAW_LI(b2p[TMP_REG_1], 0));
83 		/* this goes in the redzone */
84 		PPC_BPF_STL(b2p[TMP_REG_1], 1, -(BPF_PPC_STACK_SAVE + 8));
85 	} else {
86 		EMIT(PPC_RAW_NOP());
87 		EMIT(PPC_RAW_NOP());
88 	}
89 
90 #define BPF_TAILCALL_PROLOGUE_SIZE	8
91 
92 	if (bpf_has_stack_frame(ctx)) {
93 		/*
94 		 * We need a stack frame, but we don't necessarily need to
95 		 * save/restore LR unless we call other functions
96 		 */
97 		if (ctx->seen & SEEN_FUNC) {
98 			EMIT(PPC_RAW_MFLR(_R0));
99 			PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
100 		}
101 
102 		PPC_BPF_STLU(1, 1, -(BPF_PPC_STACKFRAME + ctx->stack_size));
103 	}
104 
105 	/*
106 	 * Back up non-volatile regs -- BPF registers 6-10
107 	 * If we haven't created our own stack frame, we save these
108 	 * in the protected zone below the previous stack frame
109 	 */
110 	for (i = BPF_REG_6; i <= BPF_REG_10; i++)
111 		if (bpf_is_seen_register(ctx, b2p[i]))
112 			PPC_BPF_STL(b2p[i], 1, bpf_jit_stack_offsetof(ctx, b2p[i]));
113 
114 	/* Setup frame pointer to point to the bpf stack area */
115 	if (bpf_is_seen_register(ctx, b2p[BPF_REG_FP]))
116 		EMIT(PPC_RAW_ADDI(b2p[BPF_REG_FP], 1,
117 				STACK_FRAME_MIN_SIZE + ctx->stack_size));
118 }
119 
120 static void bpf_jit_emit_common_epilogue(u32 *image, struct codegen_context *ctx)
121 {
122 	int i;
123 
124 	/* Restore NVRs */
125 	for (i = BPF_REG_6; i <= BPF_REG_10; i++)
126 		if (bpf_is_seen_register(ctx, b2p[i]))
127 			PPC_BPF_LL(b2p[i], 1, bpf_jit_stack_offsetof(ctx, b2p[i]));
128 
129 	/* Tear down our stack frame */
130 	if (bpf_has_stack_frame(ctx)) {
131 		EMIT(PPC_RAW_ADDI(1, 1, BPF_PPC_STACKFRAME + ctx->stack_size));
132 		if (ctx->seen & SEEN_FUNC) {
133 			PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
134 			EMIT(PPC_RAW_MTLR(0));
135 		}
136 	}
137 }
138 
139 void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
140 {
141 	bpf_jit_emit_common_epilogue(image, ctx);
142 
143 	/* Move result to r3 */
144 	EMIT(PPC_RAW_MR(3, b2p[BPF_REG_0]));
145 
146 	EMIT(PPC_RAW_BLR());
147 }
148 
149 static void bpf_jit_emit_func_call_hlp(u32 *image, struct codegen_context *ctx,
150 				       u64 func)
151 {
152 #ifdef PPC64_ELF_ABI_v1
153 	/* func points to the function descriptor */
154 	PPC_LI64(b2p[TMP_REG_2], func);
155 	/* Load actual entry point from function descriptor */
156 	PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_2], 0);
157 	/* ... and move it to CTR */
158 	EMIT(PPC_RAW_MTCTR(b2p[TMP_REG_1]));
159 	/*
160 	 * Load TOC from function descriptor at offset 8.
161 	 * We can clobber r2 since we get called through a
162 	 * function pointer (so caller will save/restore r2)
163 	 * and since we don't use a TOC ourself.
164 	 */
165 	PPC_BPF_LL(2, b2p[TMP_REG_2], 8);
166 #else
167 	/* We can clobber r12 */
168 	PPC_FUNC_ADDR(12, func);
169 	EMIT(PPC_RAW_MTCTR(12));
170 #endif
171 	EMIT(PPC_RAW_BCTRL());
172 }
173 
174 void bpf_jit_emit_func_call_rel(u32 *image, struct codegen_context *ctx, u64 func)
175 {
176 	unsigned int i, ctx_idx = ctx->idx;
177 
178 	/* Load function address into r12 */
179 	PPC_LI64(12, func);
180 
181 	/* For bpf-to-bpf function calls, the callee's address is unknown
182 	 * until the last extra pass. As seen above, we use PPC_LI64() to
183 	 * load the callee's address, but this may optimize the number of
184 	 * instructions required based on the nature of the address.
185 	 *
186 	 * Since we don't want the number of instructions emitted to change,
187 	 * we pad the optimized PPC_LI64() call with NOPs to guarantee that
188 	 * we always have a five-instruction sequence, which is the maximum
189 	 * that PPC_LI64() can emit.
190 	 */
191 	for (i = ctx->idx - ctx_idx; i < 5; i++)
192 		EMIT(PPC_RAW_NOP());
193 
194 #ifdef PPC64_ELF_ABI_v1
195 	/*
196 	 * Load TOC from function descriptor at offset 8.
197 	 * We can clobber r2 since we get called through a
198 	 * function pointer (so caller will save/restore r2)
199 	 * and since we don't use a TOC ourself.
200 	 */
201 	PPC_BPF_LL(2, 12, 8);
202 	/* Load actual entry point from function descriptor */
203 	PPC_BPF_LL(12, 12, 0);
204 #endif
205 
206 	EMIT(PPC_RAW_MTCTR(12));
207 	EMIT(PPC_RAW_BCTRL());
208 }
209 
210 static int bpf_jit_emit_tail_call(u32 *image, struct codegen_context *ctx, u32 out)
211 {
212 	/*
213 	 * By now, the eBPF program has already setup parameters in r3, r4 and r5
214 	 * r3/BPF_REG_1 - pointer to ctx -- passed as is to the next bpf program
215 	 * r4/BPF_REG_2 - pointer to bpf_array
216 	 * r5/BPF_REG_3 - index in bpf_array
217 	 */
218 	int b2p_bpf_array = b2p[BPF_REG_2];
219 	int b2p_index = b2p[BPF_REG_3];
220 
221 	/*
222 	 * if (index >= array->map.max_entries)
223 	 *   goto out;
224 	 */
225 	EMIT(PPC_RAW_LWZ(b2p[TMP_REG_1], b2p_bpf_array, offsetof(struct bpf_array, map.max_entries)));
226 	EMIT(PPC_RAW_RLWINM(b2p_index, b2p_index, 0, 0, 31));
227 	EMIT(PPC_RAW_CMPLW(b2p_index, b2p[TMP_REG_1]));
228 	PPC_BCC(COND_GE, out);
229 
230 	/*
231 	 * if (tail_call_cnt >= MAX_TAIL_CALL_CNT)
232 	 *   goto out;
233 	 */
234 	PPC_BPF_LL(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
235 	EMIT(PPC_RAW_CMPLWI(b2p[TMP_REG_1], MAX_TAIL_CALL_CNT));
236 	PPC_BCC(COND_GE, out);
237 
238 	/*
239 	 * tail_call_cnt++;
240 	 */
241 	EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], 1));
242 	PPC_BPF_STL(b2p[TMP_REG_1], 1, bpf_jit_stack_tailcallcnt(ctx));
243 
244 	/* prog = array->ptrs[index]; */
245 	EMIT(PPC_RAW_MULI(b2p[TMP_REG_1], b2p_index, 8));
246 	EMIT(PPC_RAW_ADD(b2p[TMP_REG_1], b2p[TMP_REG_1], b2p_bpf_array));
247 	PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_1], offsetof(struct bpf_array, ptrs));
248 
249 	/*
250 	 * if (prog == NULL)
251 	 *   goto out;
252 	 */
253 	EMIT(PPC_RAW_CMPLDI(b2p[TMP_REG_1], 0));
254 	PPC_BCC(COND_EQ, out);
255 
256 	/* goto *(prog->bpf_func + prologue_size); */
257 	PPC_BPF_LL(b2p[TMP_REG_1], b2p[TMP_REG_1], offsetof(struct bpf_prog, bpf_func));
258 #ifdef PPC64_ELF_ABI_v1
259 	/* skip past the function descriptor */
260 	EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1],
261 			FUNCTION_DESCR_SIZE + BPF_TAILCALL_PROLOGUE_SIZE));
262 #else
263 	EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], b2p[TMP_REG_1], BPF_TAILCALL_PROLOGUE_SIZE));
264 #endif
265 	EMIT(PPC_RAW_MTCTR(b2p[TMP_REG_1]));
266 
267 	/* tear down stack, restore NVRs, ... */
268 	bpf_jit_emit_common_epilogue(image, ctx);
269 
270 	EMIT(PPC_RAW_BCTR());
271 
272 	/* out: */
273 	return 0;
274 }
275 
276 /*
277  * We spill into the redzone always, even if the bpf program has its own stackframe.
278  * Offsets hardcoded based on BPF_PPC_STACK_SAVE -- see bpf_jit_stack_local()
279  */
280 void bpf_stf_barrier(void);
281 
282 asm (
283 "		.global bpf_stf_barrier		;"
284 "	bpf_stf_barrier:			;"
285 "		std	21,-64(1)		;"
286 "		std	22,-56(1)		;"
287 "		sync				;"
288 "		ld	21,-64(1)		;"
289 "		ld	22,-56(1)		;"
290 "		ori	31,31,0			;"
291 "		.rept 14			;"
292 "		b	1f			;"
293 "	1:					;"
294 "		.endr				;"
295 "		blr				;"
296 );
297 
298 /* Assemble the body code between the prologue & epilogue */
299 int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *ctx,
300 		       u32 *addrs, int pass)
301 {
302 	enum stf_barrier_type stf_barrier = stf_barrier_type_get();
303 	const struct bpf_insn *insn = fp->insnsi;
304 	int flen = fp->len;
305 	int i, ret;
306 
307 	/* Start of epilogue code - will only be valid 2nd pass onwards */
308 	u32 exit_addr = addrs[flen];
309 
310 	for (i = 0; i < flen; i++) {
311 		u32 code = insn[i].code;
312 		u32 dst_reg = b2p[insn[i].dst_reg];
313 		u32 src_reg = b2p[insn[i].src_reg];
314 		u32 size = BPF_SIZE(code);
315 		s16 off = insn[i].off;
316 		s32 imm = insn[i].imm;
317 		bool func_addr_fixed;
318 		u64 func_addr;
319 		u64 imm64;
320 		u32 true_cond;
321 		u32 tmp_idx;
322 
323 		/*
324 		 * addrs[] maps a BPF bytecode address into a real offset from
325 		 * the start of the body code.
326 		 */
327 		addrs[i] = ctx->idx * 4;
328 
329 		/*
330 		 * As an optimization, we note down which non-volatile registers
331 		 * are used so that we can only save/restore those in our
332 		 * prologue and epilogue. We do this here regardless of whether
333 		 * the actual BPF instruction uses src/dst registers or not
334 		 * (for instance, BPF_CALL does not use them). The expectation
335 		 * is that those instructions will have src_reg/dst_reg set to
336 		 * 0. Even otherwise, we just lose some prologue/epilogue
337 		 * optimization but everything else should work without
338 		 * any issues.
339 		 */
340 		if (dst_reg >= BPF_PPC_NVR_MIN && dst_reg < 32)
341 			bpf_set_seen_register(ctx, dst_reg);
342 		if (src_reg >= BPF_PPC_NVR_MIN && src_reg < 32)
343 			bpf_set_seen_register(ctx, src_reg);
344 
345 		switch (code) {
346 		/*
347 		 * Arithmetic operations: ADD/SUB/MUL/DIV/MOD/NEG
348 		 */
349 		case BPF_ALU | BPF_ADD | BPF_X: /* (u32) dst += (u32) src */
350 		case BPF_ALU64 | BPF_ADD | BPF_X: /* dst += src */
351 			EMIT(PPC_RAW_ADD(dst_reg, dst_reg, src_reg));
352 			goto bpf_alu32_trunc;
353 		case BPF_ALU | BPF_SUB | BPF_X: /* (u32) dst -= (u32) src */
354 		case BPF_ALU64 | BPF_SUB | BPF_X: /* dst -= src */
355 			EMIT(PPC_RAW_SUB(dst_reg, dst_reg, src_reg));
356 			goto bpf_alu32_trunc;
357 		case BPF_ALU | BPF_ADD | BPF_K: /* (u32) dst += (u32) imm */
358 		case BPF_ALU64 | BPF_ADD | BPF_K: /* dst += imm */
359 			if (!imm) {
360 				goto bpf_alu32_trunc;
361 			} else if (imm >= -32768 && imm < 32768) {
362 				EMIT(PPC_RAW_ADDI(dst_reg, dst_reg, IMM_L(imm)));
363 			} else {
364 				PPC_LI32(b2p[TMP_REG_1], imm);
365 				EMIT(PPC_RAW_ADD(dst_reg, dst_reg, b2p[TMP_REG_1]));
366 			}
367 			goto bpf_alu32_trunc;
368 		case BPF_ALU | BPF_SUB | BPF_K: /* (u32) dst -= (u32) imm */
369 		case BPF_ALU64 | BPF_SUB | BPF_K: /* dst -= imm */
370 			if (!imm) {
371 				goto bpf_alu32_trunc;
372 			} else if (imm > -32768 && imm <= 32768) {
373 				EMIT(PPC_RAW_ADDI(dst_reg, dst_reg, IMM_L(-imm)));
374 			} else {
375 				PPC_LI32(b2p[TMP_REG_1], imm);
376 				EMIT(PPC_RAW_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]));
377 			}
378 			goto bpf_alu32_trunc;
379 		case BPF_ALU | BPF_MUL | BPF_X: /* (u32) dst *= (u32) src */
380 		case BPF_ALU64 | BPF_MUL | BPF_X: /* dst *= src */
381 			if (BPF_CLASS(code) == BPF_ALU)
382 				EMIT(PPC_RAW_MULW(dst_reg, dst_reg, src_reg));
383 			else
384 				EMIT(PPC_RAW_MULD(dst_reg, dst_reg, src_reg));
385 			goto bpf_alu32_trunc;
386 		case BPF_ALU | BPF_MUL | BPF_K: /* (u32) dst *= (u32) imm */
387 		case BPF_ALU64 | BPF_MUL | BPF_K: /* dst *= imm */
388 			if (imm >= -32768 && imm < 32768)
389 				EMIT(PPC_RAW_MULI(dst_reg, dst_reg, IMM_L(imm)));
390 			else {
391 				PPC_LI32(b2p[TMP_REG_1], imm);
392 				if (BPF_CLASS(code) == BPF_ALU)
393 					EMIT(PPC_RAW_MULW(dst_reg, dst_reg,
394 							b2p[TMP_REG_1]));
395 				else
396 					EMIT(PPC_RAW_MULD(dst_reg, dst_reg,
397 							b2p[TMP_REG_1]));
398 			}
399 			goto bpf_alu32_trunc;
400 		case BPF_ALU | BPF_DIV | BPF_X: /* (u32) dst /= (u32) src */
401 		case BPF_ALU | BPF_MOD | BPF_X: /* (u32) dst %= (u32) src */
402 			if (BPF_OP(code) == BPF_MOD) {
403 				EMIT(PPC_RAW_DIVWU(b2p[TMP_REG_1], dst_reg, src_reg));
404 				EMIT(PPC_RAW_MULW(b2p[TMP_REG_1], src_reg,
405 						b2p[TMP_REG_1]));
406 				EMIT(PPC_RAW_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]));
407 			} else
408 				EMIT(PPC_RAW_DIVWU(dst_reg, dst_reg, src_reg));
409 			goto bpf_alu32_trunc;
410 		case BPF_ALU64 | BPF_DIV | BPF_X: /* dst /= src */
411 		case BPF_ALU64 | BPF_MOD | BPF_X: /* dst %= src */
412 			if (BPF_OP(code) == BPF_MOD) {
413 				EMIT(PPC_RAW_DIVDU(b2p[TMP_REG_1], dst_reg, src_reg));
414 				EMIT(PPC_RAW_MULD(b2p[TMP_REG_1], src_reg,
415 						b2p[TMP_REG_1]));
416 				EMIT(PPC_RAW_SUB(dst_reg, dst_reg, b2p[TMP_REG_1]));
417 			} else
418 				EMIT(PPC_RAW_DIVDU(dst_reg, dst_reg, src_reg));
419 			break;
420 		case BPF_ALU | BPF_MOD | BPF_K: /* (u32) dst %= (u32) imm */
421 		case BPF_ALU | BPF_DIV | BPF_K: /* (u32) dst /= (u32) imm */
422 		case BPF_ALU64 | BPF_MOD | BPF_K: /* dst %= imm */
423 		case BPF_ALU64 | BPF_DIV | BPF_K: /* dst /= imm */
424 			if (imm == 0)
425 				return -EINVAL;
426 			if (imm == 1) {
427 				if (BPF_OP(code) == BPF_DIV) {
428 					goto bpf_alu32_trunc;
429 				} else {
430 					EMIT(PPC_RAW_LI(dst_reg, 0));
431 					break;
432 				}
433 			}
434 
435 			PPC_LI32(b2p[TMP_REG_1], imm);
436 			switch (BPF_CLASS(code)) {
437 			case BPF_ALU:
438 				if (BPF_OP(code) == BPF_MOD) {
439 					EMIT(PPC_RAW_DIVWU(b2p[TMP_REG_2],
440 							dst_reg,
441 							b2p[TMP_REG_1]));
442 					EMIT(PPC_RAW_MULW(b2p[TMP_REG_1],
443 							b2p[TMP_REG_1],
444 							b2p[TMP_REG_2]));
445 					EMIT(PPC_RAW_SUB(dst_reg, dst_reg,
446 							b2p[TMP_REG_1]));
447 				} else
448 					EMIT(PPC_RAW_DIVWU(dst_reg, dst_reg,
449 							b2p[TMP_REG_1]));
450 				break;
451 			case BPF_ALU64:
452 				if (BPF_OP(code) == BPF_MOD) {
453 					EMIT(PPC_RAW_DIVDU(b2p[TMP_REG_2],
454 							dst_reg,
455 							b2p[TMP_REG_1]));
456 					EMIT(PPC_RAW_MULD(b2p[TMP_REG_1],
457 							b2p[TMP_REG_1],
458 							b2p[TMP_REG_2]));
459 					EMIT(PPC_RAW_SUB(dst_reg, dst_reg,
460 							b2p[TMP_REG_1]));
461 				} else
462 					EMIT(PPC_RAW_DIVDU(dst_reg, dst_reg,
463 							b2p[TMP_REG_1]));
464 				break;
465 			}
466 			goto bpf_alu32_trunc;
467 		case BPF_ALU | BPF_NEG: /* (u32) dst = -dst */
468 		case BPF_ALU64 | BPF_NEG: /* dst = -dst */
469 			EMIT(PPC_RAW_NEG(dst_reg, dst_reg));
470 			goto bpf_alu32_trunc;
471 
472 		/*
473 		 * Logical operations: AND/OR/XOR/[A]LSH/[A]RSH
474 		 */
475 		case BPF_ALU | BPF_AND | BPF_X: /* (u32) dst = dst & src */
476 		case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
477 			EMIT(PPC_RAW_AND(dst_reg, dst_reg, src_reg));
478 			goto bpf_alu32_trunc;
479 		case BPF_ALU | BPF_AND | BPF_K: /* (u32) dst = dst & imm */
480 		case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
481 			if (!IMM_H(imm))
482 				EMIT(PPC_RAW_ANDI(dst_reg, dst_reg, IMM_L(imm)));
483 			else {
484 				/* Sign-extended */
485 				PPC_LI32(b2p[TMP_REG_1], imm);
486 				EMIT(PPC_RAW_AND(dst_reg, dst_reg, b2p[TMP_REG_1]));
487 			}
488 			goto bpf_alu32_trunc;
489 		case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
490 		case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
491 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, src_reg));
492 			goto bpf_alu32_trunc;
493 		case BPF_ALU | BPF_OR | BPF_K:/* dst = (u32) dst | (u32) imm */
494 		case BPF_ALU64 | BPF_OR | BPF_K:/* dst = dst | imm */
495 			if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
496 				/* Sign-extended */
497 				PPC_LI32(b2p[TMP_REG_1], imm);
498 				EMIT(PPC_RAW_OR(dst_reg, dst_reg, b2p[TMP_REG_1]));
499 			} else {
500 				if (IMM_L(imm))
501 					EMIT(PPC_RAW_ORI(dst_reg, dst_reg, IMM_L(imm)));
502 				if (IMM_H(imm))
503 					EMIT(PPC_RAW_ORIS(dst_reg, dst_reg, IMM_H(imm)));
504 			}
505 			goto bpf_alu32_trunc;
506 		case BPF_ALU | BPF_XOR | BPF_X: /* (u32) dst ^= src */
507 		case BPF_ALU64 | BPF_XOR | BPF_X: /* dst ^= src */
508 			EMIT(PPC_RAW_XOR(dst_reg, dst_reg, src_reg));
509 			goto bpf_alu32_trunc;
510 		case BPF_ALU | BPF_XOR | BPF_K: /* (u32) dst ^= (u32) imm */
511 		case BPF_ALU64 | BPF_XOR | BPF_K: /* dst ^= imm */
512 			if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) {
513 				/* Sign-extended */
514 				PPC_LI32(b2p[TMP_REG_1], imm);
515 				EMIT(PPC_RAW_XOR(dst_reg, dst_reg, b2p[TMP_REG_1]));
516 			} else {
517 				if (IMM_L(imm))
518 					EMIT(PPC_RAW_XORI(dst_reg, dst_reg, IMM_L(imm)));
519 				if (IMM_H(imm))
520 					EMIT(PPC_RAW_XORIS(dst_reg, dst_reg, IMM_H(imm)));
521 			}
522 			goto bpf_alu32_trunc;
523 		case BPF_ALU | BPF_LSH | BPF_X: /* (u32) dst <<= (u32) src */
524 			/* slw clears top 32 bits */
525 			EMIT(PPC_RAW_SLW(dst_reg, dst_reg, src_reg));
526 			/* skip zero extension move, but set address map. */
527 			if (insn_is_zext(&insn[i + 1]))
528 				addrs[++i] = ctx->idx * 4;
529 			break;
530 		case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */
531 			EMIT(PPC_RAW_SLD(dst_reg, dst_reg, src_reg));
532 			break;
533 		case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<== (u32) imm */
534 			/* with imm 0, we still need to clear top 32 bits */
535 			EMIT(PPC_RAW_SLWI(dst_reg, dst_reg, imm));
536 			if (insn_is_zext(&insn[i + 1]))
537 				addrs[++i] = ctx->idx * 4;
538 			break;
539 		case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<== imm */
540 			if (imm != 0)
541 				EMIT(PPC_RAW_SLDI(dst_reg, dst_reg, imm));
542 			break;
543 		case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */
544 			EMIT(PPC_RAW_SRW(dst_reg, dst_reg, src_reg));
545 			if (insn_is_zext(&insn[i + 1]))
546 				addrs[++i] = ctx->idx * 4;
547 			break;
548 		case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */
549 			EMIT(PPC_RAW_SRD(dst_reg, dst_reg, src_reg));
550 			break;
551 		case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */
552 			EMIT(PPC_RAW_SRWI(dst_reg, dst_reg, imm));
553 			if (insn_is_zext(&insn[i + 1]))
554 				addrs[++i] = ctx->idx * 4;
555 			break;
556 		case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */
557 			if (imm != 0)
558 				EMIT(PPC_RAW_SRDI(dst_reg, dst_reg, imm));
559 			break;
560 		case BPF_ALU | BPF_ARSH | BPF_X: /* (s32) dst >>= src */
561 			EMIT(PPC_RAW_SRAW(dst_reg, dst_reg, src_reg));
562 			goto bpf_alu32_trunc;
563 		case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */
564 			EMIT(PPC_RAW_SRAD(dst_reg, dst_reg, src_reg));
565 			break;
566 		case BPF_ALU | BPF_ARSH | BPF_K: /* (s32) dst >>= imm */
567 			EMIT(PPC_RAW_SRAWI(dst_reg, dst_reg, imm));
568 			goto bpf_alu32_trunc;
569 		case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */
570 			if (imm != 0)
571 				EMIT(PPC_RAW_SRADI(dst_reg, dst_reg, imm));
572 			break;
573 
574 		/*
575 		 * MOV
576 		 */
577 		case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
578 		case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
579 			if (imm == 1) {
580 				/* special mov32 for zext */
581 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 0, 31));
582 				break;
583 			}
584 			EMIT(PPC_RAW_MR(dst_reg, src_reg));
585 			goto bpf_alu32_trunc;
586 		case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */
587 		case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */
588 			PPC_LI32(dst_reg, imm);
589 			if (imm < 0)
590 				goto bpf_alu32_trunc;
591 			else if (insn_is_zext(&insn[i + 1]))
592 				addrs[++i] = ctx->idx * 4;
593 			break;
594 
595 bpf_alu32_trunc:
596 		/* Truncate to 32-bits */
597 		if (BPF_CLASS(code) == BPF_ALU && !fp->aux->verifier_zext)
598 			EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 0, 31));
599 		break;
600 
601 		/*
602 		 * BPF_FROM_BE/LE
603 		 */
604 		case BPF_ALU | BPF_END | BPF_FROM_LE:
605 		case BPF_ALU | BPF_END | BPF_FROM_BE:
606 #ifdef __BIG_ENDIAN__
607 			if (BPF_SRC(code) == BPF_FROM_BE)
608 				goto emit_clear;
609 #else /* !__BIG_ENDIAN__ */
610 			if (BPF_SRC(code) == BPF_FROM_LE)
611 				goto emit_clear;
612 #endif
613 			switch (imm) {
614 			case 16:
615 				/* Rotate 8 bits left & mask with 0x0000ff00 */
616 				EMIT(PPC_RAW_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 16, 23));
617 				/* Rotate 8 bits right & insert LSB to reg */
618 				EMIT(PPC_RAW_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 24, 31));
619 				/* Move result back to dst_reg */
620 				EMIT(PPC_RAW_MR(dst_reg, b2p[TMP_REG_1]));
621 				break;
622 			case 32:
623 				/*
624 				 * Rotate word left by 8 bits:
625 				 * 2 bytes are already in their final position
626 				 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4)
627 				 */
628 				EMIT(PPC_RAW_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 0, 31));
629 				/* Rotate 24 bits and insert byte 1 */
630 				EMIT(PPC_RAW_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 0, 7));
631 				/* Rotate 24 bits and insert byte 3 */
632 				EMIT(PPC_RAW_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 16, 23));
633 				EMIT(PPC_RAW_MR(dst_reg, b2p[TMP_REG_1]));
634 				break;
635 			case 64:
636 				/*
637 				 * Way easier and faster(?) to store the value
638 				 * into stack and then use ldbrx
639 				 *
640 				 * ctx->seen will be reliable in pass2, but
641 				 * the instructions generated will remain the
642 				 * same across all passes
643 				 */
644 				PPC_BPF_STL(dst_reg, 1, bpf_jit_stack_local(ctx));
645 				EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], 1, bpf_jit_stack_local(ctx)));
646 				EMIT(PPC_RAW_LDBRX(dst_reg, 0, b2p[TMP_REG_1]));
647 				break;
648 			}
649 			break;
650 
651 emit_clear:
652 			switch (imm) {
653 			case 16:
654 				/* zero-extend 16 bits into 64 bits */
655 				EMIT(PPC_RAW_RLDICL(dst_reg, dst_reg, 0, 48));
656 				if (insn_is_zext(&insn[i + 1]))
657 					addrs[++i] = ctx->idx * 4;
658 				break;
659 			case 32:
660 				if (!fp->aux->verifier_zext)
661 					/* zero-extend 32 bits into 64 bits */
662 					EMIT(PPC_RAW_RLDICL(dst_reg, dst_reg, 0, 32));
663 				break;
664 			case 64:
665 				/* nop */
666 				break;
667 			}
668 			break;
669 
670 		/*
671 		 * BPF_ST NOSPEC (speculation barrier)
672 		 */
673 		case BPF_ST | BPF_NOSPEC:
674 			if (!security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) ||
675 					!security_ftr_enabled(SEC_FTR_STF_BARRIER))
676 				break;
677 
678 			switch (stf_barrier) {
679 			case STF_BARRIER_EIEIO:
680 				EMIT(PPC_RAW_EIEIO() | 0x02000000);
681 				break;
682 			case STF_BARRIER_SYNC_ORI:
683 				EMIT(PPC_RAW_SYNC());
684 				EMIT(PPC_RAW_LD(b2p[TMP_REG_1], _R13, 0));
685 				EMIT(PPC_RAW_ORI(_R31, _R31, 0));
686 				break;
687 			case STF_BARRIER_FALLBACK:
688 				EMIT(PPC_RAW_MFLR(b2p[TMP_REG_1]));
689 				PPC_LI64(12, dereference_kernel_function_descriptor(bpf_stf_barrier));
690 				EMIT(PPC_RAW_MTCTR(12));
691 				EMIT(PPC_RAW_BCTRL());
692 				EMIT(PPC_RAW_MTLR(b2p[TMP_REG_1]));
693 				break;
694 			case STF_BARRIER_NONE:
695 				break;
696 			}
697 			break;
698 
699 		/*
700 		 * BPF_ST(X)
701 		 */
702 		case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */
703 		case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
704 			if (BPF_CLASS(code) == BPF_ST) {
705 				EMIT(PPC_RAW_LI(b2p[TMP_REG_1], imm));
706 				src_reg = b2p[TMP_REG_1];
707 			}
708 			EMIT(PPC_RAW_STB(src_reg, dst_reg, off));
709 			break;
710 		case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
711 		case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
712 			if (BPF_CLASS(code) == BPF_ST) {
713 				EMIT(PPC_RAW_LI(b2p[TMP_REG_1], imm));
714 				src_reg = b2p[TMP_REG_1];
715 			}
716 			EMIT(PPC_RAW_STH(src_reg, dst_reg, off));
717 			break;
718 		case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
719 		case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
720 			if (BPF_CLASS(code) == BPF_ST) {
721 				PPC_LI32(b2p[TMP_REG_1], imm);
722 				src_reg = b2p[TMP_REG_1];
723 			}
724 			EMIT(PPC_RAW_STW(src_reg, dst_reg, off));
725 			break;
726 		case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
727 		case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
728 			if (BPF_CLASS(code) == BPF_ST) {
729 				PPC_LI32(b2p[TMP_REG_1], imm);
730 				src_reg = b2p[TMP_REG_1];
731 			}
732 			PPC_BPF_STL(src_reg, dst_reg, off);
733 			break;
734 
735 		/*
736 		 * BPF_STX ATOMIC (atomic ops)
737 		 */
738 		case BPF_STX | BPF_ATOMIC | BPF_W:
739 			if (imm != BPF_ADD) {
740 				pr_err_ratelimited(
741 					"eBPF filter atomic op code %02x (@%d) unsupported\n",
742 					code, i);
743 				return -ENOTSUPP;
744 			}
745 
746 			/* *(u32 *)(dst + off) += src */
747 
748 			/* Get EA into TMP_REG_1 */
749 			EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], dst_reg, off));
750 			tmp_idx = ctx->idx * 4;
751 			/* load value from memory into TMP_REG_2 */
752 			EMIT(PPC_RAW_LWARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0));
753 			/* add value from src_reg into this */
754 			EMIT(PPC_RAW_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg));
755 			/* store result back */
756 			EMIT(PPC_RAW_STWCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]));
757 			/* we're done if this succeeded */
758 			PPC_BCC_SHORT(COND_NE, tmp_idx);
759 			break;
760 		case BPF_STX | BPF_ATOMIC | BPF_DW:
761 			if (imm != BPF_ADD) {
762 				pr_err_ratelimited(
763 					"eBPF filter atomic op code %02x (@%d) unsupported\n",
764 					code, i);
765 				return -ENOTSUPP;
766 			}
767 			/* *(u64 *)(dst + off) += src */
768 
769 			EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], dst_reg, off));
770 			tmp_idx = ctx->idx * 4;
771 			EMIT(PPC_RAW_LDARX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1], 0));
772 			EMIT(PPC_RAW_ADD(b2p[TMP_REG_2], b2p[TMP_REG_2], src_reg));
773 			EMIT(PPC_RAW_STDCX(b2p[TMP_REG_2], 0, b2p[TMP_REG_1]));
774 			PPC_BCC_SHORT(COND_NE, tmp_idx);
775 			break;
776 
777 		/*
778 		 * BPF_LDX
779 		 */
780 		/* dst = *(u8 *)(ul) (src + off) */
781 		case BPF_LDX | BPF_MEM | BPF_B:
782 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
783 		/* dst = *(u16 *)(ul) (src + off) */
784 		case BPF_LDX | BPF_MEM | BPF_H:
785 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
786 		/* dst = *(u32 *)(ul) (src + off) */
787 		case BPF_LDX | BPF_MEM | BPF_W:
788 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
789 		/* dst = *(u64 *)(ul) (src + off) */
790 		case BPF_LDX | BPF_MEM | BPF_DW:
791 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
792 			/*
793 			 * As PTR_TO_BTF_ID that uses BPF_PROBE_MEM mode could either be a valid
794 			 * kernel pointer or NULL but not a userspace address, execute BPF_PROBE_MEM
795 			 * load only if addr is kernel address (see is_kernel_addr()), otherwise
796 			 * set dst_reg=0 and move on.
797 			 */
798 			if (BPF_MODE(code) == BPF_PROBE_MEM) {
799 				EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], src_reg, off));
800 				if (IS_ENABLED(CONFIG_PPC_BOOK3E_64))
801 					PPC_LI64(b2p[TMP_REG_2], 0x8000000000000000ul);
802 				else /* BOOK3S_64 */
803 					PPC_LI64(b2p[TMP_REG_2], PAGE_OFFSET);
804 				EMIT(PPC_RAW_CMPLD(b2p[TMP_REG_1], b2p[TMP_REG_2]));
805 				PPC_BCC(COND_GT, (ctx->idx + 4) * 4);
806 				EMIT(PPC_RAW_LI(dst_reg, 0));
807 				/*
808 				 * Check if 'off' is word aligned because PPC_BPF_LL()
809 				 * (BPF_DW case) generates two instructions if 'off' is not
810 				 * word-aligned and one instruction otherwise.
811 				 */
812 				if (BPF_SIZE(code) == BPF_DW && (off & 3))
813 					PPC_JMP((ctx->idx + 3) * 4);
814 				else
815 					PPC_JMP((ctx->idx + 2) * 4);
816 			}
817 
818 			switch (size) {
819 			case BPF_B:
820 				EMIT(PPC_RAW_LBZ(dst_reg, src_reg, off));
821 				break;
822 			case BPF_H:
823 				EMIT(PPC_RAW_LHZ(dst_reg, src_reg, off));
824 				break;
825 			case BPF_W:
826 				EMIT(PPC_RAW_LWZ(dst_reg, src_reg, off));
827 				break;
828 			case BPF_DW:
829 				PPC_BPF_LL(dst_reg, src_reg, off);
830 				break;
831 			}
832 
833 			if (size != BPF_DW && insn_is_zext(&insn[i + 1]))
834 				addrs[++i] = ctx->idx * 4;
835 
836 			if (BPF_MODE(code) == BPF_PROBE_MEM) {
837 				ret = bpf_add_extable_entry(fp, image, pass, ctx, ctx->idx - 1,
838 							    4, dst_reg);
839 				if (ret)
840 					return ret;
841 			}
842 			break;
843 
844 		/*
845 		 * Doubleword load
846 		 * 16 byte instruction that uses two 'struct bpf_insn'
847 		 */
848 		case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
849 			imm64 = ((u64)(u32) insn[i].imm) |
850 				    (((u64)(u32) insn[i+1].imm) << 32);
851 			/* Adjust for two bpf instructions */
852 			addrs[++i] = ctx->idx * 4;
853 			PPC_LI64(dst_reg, imm64);
854 			break;
855 
856 		/*
857 		 * Return/Exit
858 		 */
859 		case BPF_JMP | BPF_EXIT:
860 			/*
861 			 * If this isn't the very last instruction, branch to
862 			 * the epilogue. If we _are_ the last instruction,
863 			 * we'll just fall through to the epilogue.
864 			 */
865 			if (i != flen - 1)
866 				PPC_JMP(exit_addr);
867 			/* else fall through to the epilogue */
868 			break;
869 
870 		/*
871 		 * Call kernel helper or bpf function
872 		 */
873 		case BPF_JMP | BPF_CALL:
874 			ctx->seen |= SEEN_FUNC;
875 
876 			ret = bpf_jit_get_func_addr(fp, &insn[i], false,
877 						    &func_addr, &func_addr_fixed);
878 			if (ret < 0)
879 				return ret;
880 
881 			if (func_addr_fixed)
882 				bpf_jit_emit_func_call_hlp(image, ctx, func_addr);
883 			else
884 				bpf_jit_emit_func_call_rel(image, ctx, func_addr);
885 			/* move return value from r3 to BPF_REG_0 */
886 			EMIT(PPC_RAW_MR(b2p[BPF_REG_0], 3));
887 			break;
888 
889 		/*
890 		 * Jumps and branches
891 		 */
892 		case BPF_JMP | BPF_JA:
893 			PPC_JMP(addrs[i + 1 + off]);
894 			break;
895 
896 		case BPF_JMP | BPF_JGT | BPF_K:
897 		case BPF_JMP | BPF_JGT | BPF_X:
898 		case BPF_JMP | BPF_JSGT | BPF_K:
899 		case BPF_JMP | BPF_JSGT | BPF_X:
900 		case BPF_JMP32 | BPF_JGT | BPF_K:
901 		case BPF_JMP32 | BPF_JGT | BPF_X:
902 		case BPF_JMP32 | BPF_JSGT | BPF_K:
903 		case BPF_JMP32 | BPF_JSGT | BPF_X:
904 			true_cond = COND_GT;
905 			goto cond_branch;
906 		case BPF_JMP | BPF_JLT | BPF_K:
907 		case BPF_JMP | BPF_JLT | BPF_X:
908 		case BPF_JMP | BPF_JSLT | BPF_K:
909 		case BPF_JMP | BPF_JSLT | BPF_X:
910 		case BPF_JMP32 | BPF_JLT | BPF_K:
911 		case BPF_JMP32 | BPF_JLT | BPF_X:
912 		case BPF_JMP32 | BPF_JSLT | BPF_K:
913 		case BPF_JMP32 | BPF_JSLT | BPF_X:
914 			true_cond = COND_LT;
915 			goto cond_branch;
916 		case BPF_JMP | BPF_JGE | BPF_K:
917 		case BPF_JMP | BPF_JGE | BPF_X:
918 		case BPF_JMP | BPF_JSGE | BPF_K:
919 		case BPF_JMP | BPF_JSGE | BPF_X:
920 		case BPF_JMP32 | BPF_JGE | BPF_K:
921 		case BPF_JMP32 | BPF_JGE | BPF_X:
922 		case BPF_JMP32 | BPF_JSGE | BPF_K:
923 		case BPF_JMP32 | BPF_JSGE | BPF_X:
924 			true_cond = COND_GE;
925 			goto cond_branch;
926 		case BPF_JMP | BPF_JLE | BPF_K:
927 		case BPF_JMP | BPF_JLE | BPF_X:
928 		case BPF_JMP | BPF_JSLE | BPF_K:
929 		case BPF_JMP | BPF_JSLE | BPF_X:
930 		case BPF_JMP32 | BPF_JLE | BPF_K:
931 		case BPF_JMP32 | BPF_JLE | BPF_X:
932 		case BPF_JMP32 | BPF_JSLE | BPF_K:
933 		case BPF_JMP32 | BPF_JSLE | BPF_X:
934 			true_cond = COND_LE;
935 			goto cond_branch;
936 		case BPF_JMP | BPF_JEQ | BPF_K:
937 		case BPF_JMP | BPF_JEQ | BPF_X:
938 		case BPF_JMP32 | BPF_JEQ | BPF_K:
939 		case BPF_JMP32 | BPF_JEQ | BPF_X:
940 			true_cond = COND_EQ;
941 			goto cond_branch;
942 		case BPF_JMP | BPF_JNE | BPF_K:
943 		case BPF_JMP | BPF_JNE | BPF_X:
944 		case BPF_JMP32 | BPF_JNE | BPF_K:
945 		case BPF_JMP32 | BPF_JNE | BPF_X:
946 			true_cond = COND_NE;
947 			goto cond_branch;
948 		case BPF_JMP | BPF_JSET | BPF_K:
949 		case BPF_JMP | BPF_JSET | BPF_X:
950 		case BPF_JMP32 | BPF_JSET | BPF_K:
951 		case BPF_JMP32 | BPF_JSET | BPF_X:
952 			true_cond = COND_NE;
953 			/* Fall through */
954 
955 cond_branch:
956 			switch (code) {
957 			case BPF_JMP | BPF_JGT | BPF_X:
958 			case BPF_JMP | BPF_JLT | BPF_X:
959 			case BPF_JMP | BPF_JGE | BPF_X:
960 			case BPF_JMP | BPF_JLE | BPF_X:
961 			case BPF_JMP | BPF_JEQ | BPF_X:
962 			case BPF_JMP | BPF_JNE | BPF_X:
963 			case BPF_JMP32 | BPF_JGT | BPF_X:
964 			case BPF_JMP32 | BPF_JLT | BPF_X:
965 			case BPF_JMP32 | BPF_JGE | BPF_X:
966 			case BPF_JMP32 | BPF_JLE | BPF_X:
967 			case BPF_JMP32 | BPF_JEQ | BPF_X:
968 			case BPF_JMP32 | BPF_JNE | BPF_X:
969 				/* unsigned comparison */
970 				if (BPF_CLASS(code) == BPF_JMP32)
971 					EMIT(PPC_RAW_CMPLW(dst_reg, src_reg));
972 				else
973 					EMIT(PPC_RAW_CMPLD(dst_reg, src_reg));
974 				break;
975 			case BPF_JMP | BPF_JSGT | BPF_X:
976 			case BPF_JMP | BPF_JSLT | BPF_X:
977 			case BPF_JMP | BPF_JSGE | BPF_X:
978 			case BPF_JMP | BPF_JSLE | BPF_X:
979 			case BPF_JMP32 | BPF_JSGT | BPF_X:
980 			case BPF_JMP32 | BPF_JSLT | BPF_X:
981 			case BPF_JMP32 | BPF_JSGE | BPF_X:
982 			case BPF_JMP32 | BPF_JSLE | BPF_X:
983 				/* signed comparison */
984 				if (BPF_CLASS(code) == BPF_JMP32)
985 					EMIT(PPC_RAW_CMPW(dst_reg, src_reg));
986 				else
987 					EMIT(PPC_RAW_CMPD(dst_reg, src_reg));
988 				break;
989 			case BPF_JMP | BPF_JSET | BPF_X:
990 			case BPF_JMP32 | BPF_JSET | BPF_X:
991 				if (BPF_CLASS(code) == BPF_JMP) {
992 					EMIT(PPC_RAW_AND_DOT(b2p[TMP_REG_1], dst_reg,
993 						    src_reg));
994 				} else {
995 					int tmp_reg = b2p[TMP_REG_1];
996 
997 					EMIT(PPC_RAW_AND(tmp_reg, dst_reg, src_reg));
998 					EMIT(PPC_RAW_RLWINM_DOT(tmp_reg, tmp_reg, 0, 0,
999 						       31));
1000 				}
1001 				break;
1002 			case BPF_JMP | BPF_JNE | BPF_K:
1003 			case BPF_JMP | BPF_JEQ | BPF_K:
1004 			case BPF_JMP | BPF_JGT | BPF_K:
1005 			case BPF_JMP | BPF_JLT | BPF_K:
1006 			case BPF_JMP | BPF_JGE | BPF_K:
1007 			case BPF_JMP | BPF_JLE | BPF_K:
1008 			case BPF_JMP32 | BPF_JNE | BPF_K:
1009 			case BPF_JMP32 | BPF_JEQ | BPF_K:
1010 			case BPF_JMP32 | BPF_JGT | BPF_K:
1011 			case BPF_JMP32 | BPF_JLT | BPF_K:
1012 			case BPF_JMP32 | BPF_JGE | BPF_K:
1013 			case BPF_JMP32 | BPF_JLE | BPF_K:
1014 			{
1015 				bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
1016 
1017 				/*
1018 				 * Need sign-extended load, so only positive
1019 				 * values can be used as imm in cmpldi
1020 				 */
1021 				if (imm >= 0 && imm < 32768) {
1022 					if (is_jmp32)
1023 						EMIT(PPC_RAW_CMPLWI(dst_reg, imm));
1024 					else
1025 						EMIT(PPC_RAW_CMPLDI(dst_reg, imm));
1026 				} else {
1027 					/* sign-extending load */
1028 					PPC_LI32(b2p[TMP_REG_1], imm);
1029 					/* ... but unsigned comparison */
1030 					if (is_jmp32)
1031 						EMIT(PPC_RAW_CMPLW(dst_reg,
1032 							  b2p[TMP_REG_1]));
1033 					else
1034 						EMIT(PPC_RAW_CMPLD(dst_reg,
1035 							  b2p[TMP_REG_1]));
1036 				}
1037 				break;
1038 			}
1039 			case BPF_JMP | BPF_JSGT | BPF_K:
1040 			case BPF_JMP | BPF_JSLT | BPF_K:
1041 			case BPF_JMP | BPF_JSGE | BPF_K:
1042 			case BPF_JMP | BPF_JSLE | BPF_K:
1043 			case BPF_JMP32 | BPF_JSGT | BPF_K:
1044 			case BPF_JMP32 | BPF_JSLT | BPF_K:
1045 			case BPF_JMP32 | BPF_JSGE | BPF_K:
1046 			case BPF_JMP32 | BPF_JSLE | BPF_K:
1047 			{
1048 				bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32;
1049 
1050 				/*
1051 				 * signed comparison, so any 16-bit value
1052 				 * can be used in cmpdi
1053 				 */
1054 				if (imm >= -32768 && imm < 32768) {
1055 					if (is_jmp32)
1056 						EMIT(PPC_RAW_CMPWI(dst_reg, imm));
1057 					else
1058 						EMIT(PPC_RAW_CMPDI(dst_reg, imm));
1059 				} else {
1060 					PPC_LI32(b2p[TMP_REG_1], imm);
1061 					if (is_jmp32)
1062 						EMIT(PPC_RAW_CMPW(dst_reg,
1063 							 b2p[TMP_REG_1]));
1064 					else
1065 						EMIT(PPC_RAW_CMPD(dst_reg,
1066 							 b2p[TMP_REG_1]));
1067 				}
1068 				break;
1069 			}
1070 			case BPF_JMP | BPF_JSET | BPF_K:
1071 			case BPF_JMP32 | BPF_JSET | BPF_K:
1072 				/* andi does not sign-extend the immediate */
1073 				if (imm >= 0 && imm < 32768)
1074 					/* PPC_ANDI is _only/always_ dot-form */
1075 					EMIT(PPC_RAW_ANDI(b2p[TMP_REG_1], dst_reg, imm));
1076 				else {
1077 					int tmp_reg = b2p[TMP_REG_1];
1078 
1079 					PPC_LI32(tmp_reg, imm);
1080 					if (BPF_CLASS(code) == BPF_JMP) {
1081 						EMIT(PPC_RAW_AND_DOT(tmp_reg, dst_reg,
1082 							    tmp_reg));
1083 					} else {
1084 						EMIT(PPC_RAW_AND(tmp_reg, dst_reg,
1085 							tmp_reg));
1086 						EMIT(PPC_RAW_RLWINM_DOT(tmp_reg, tmp_reg,
1087 							       0, 0, 31));
1088 					}
1089 				}
1090 				break;
1091 			}
1092 			PPC_BCC(true_cond, addrs[i + 1 + off]);
1093 			break;
1094 
1095 		/*
1096 		 * Tail call
1097 		 */
1098 		case BPF_JMP | BPF_TAIL_CALL:
1099 			ctx->seen |= SEEN_TAILCALL;
1100 			ret = bpf_jit_emit_tail_call(image, ctx, addrs[i + 1]);
1101 			if (ret < 0)
1102 				return ret;
1103 			break;
1104 
1105 		default:
1106 			/*
1107 			 * The filter contains something cruel & unusual.
1108 			 * We don't handle it, but also there shouldn't be
1109 			 * anything missing from our list.
1110 			 */
1111 			pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n",
1112 					code, i);
1113 			return -ENOTSUPP;
1114 		}
1115 	}
1116 
1117 	/* Set end-of-body-code address for exit. */
1118 	addrs[i] = ctx->idx * 4;
1119 
1120 	return 0;
1121 }
1122