1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * eBPF JIT compiler for PPC32
4  *
5  * Copyright 2020 Christophe Leroy <christophe.leroy@csgroup.eu>
6  *		  CS GROUP France
7  *
8  * Based on PPC64 eBPF JIT compiler by Naveen N. Rao
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_jit.h"
20 
21 /*
22  * Stack layout:
23  *
24  *		[	prev sp		] <-------------
25  *		[   nv gpr save area	] 16 * 4	|
26  * fp (r31) -->	[   ebpf stack space	] upto 512	|
27  *		[     frame header	] 16		|
28  * sp (r1) --->	[    stack pointer	] --------------
29  */
30 
31 /* for gpr non volatile registers r17 to r31 (14) + tail call */
32 #define BPF_PPC_STACK_SAVE	(15 * 4 + 4)
33 /* stack frame, ensure this is quadword aligned */
34 #define BPF_PPC_STACKFRAME(ctx)	(STACK_FRAME_MIN_SIZE + BPF_PPC_STACK_SAVE + (ctx)->stack_size)
35 
36 #define PPC_EX32(r, i)		EMIT(PPC_RAW_LI((r), (i) < 0 ? -1 : 0))
37 
38 /* PPC NVR range -- update this if we ever use NVRs below r17 */
39 #define BPF_PPC_NVR_MIN		_R17
40 #define BPF_PPC_TC		_R16
41 
42 /* BPF register usage */
43 #define TMP_REG			(MAX_BPF_JIT_REG + 0)
44 
45 /* BPF to ppc register mappings */
46 void bpf_jit_init_reg_mapping(struct codegen_context *ctx)
47 {
48 	/* function return value */
49 	ctx->b2p[BPF_REG_0] = _R12;
50 	/* function arguments */
51 	ctx->b2p[BPF_REG_1] = _R4;
52 	ctx->b2p[BPF_REG_2] = _R6;
53 	ctx->b2p[BPF_REG_3] = _R8;
54 	ctx->b2p[BPF_REG_4] = _R10;
55 	ctx->b2p[BPF_REG_5] = _R22;
56 	/* non volatile registers */
57 	ctx->b2p[BPF_REG_6] = _R24;
58 	ctx->b2p[BPF_REG_7] = _R26;
59 	ctx->b2p[BPF_REG_8] = _R28;
60 	ctx->b2p[BPF_REG_9] = _R30;
61 	/* frame pointer aka BPF_REG_10 */
62 	ctx->b2p[BPF_REG_FP] = _R18;
63 	/* eBPF jit internal registers */
64 	ctx->b2p[BPF_REG_AX] = _R20;
65 	ctx->b2p[TMP_REG] = _R31;		/* 32 bits */
66 }
67 
68 static int bpf_jit_stack_offsetof(struct codegen_context *ctx, int reg)
69 {
70 	if ((reg >= BPF_PPC_NVR_MIN && reg < 32) || reg == BPF_PPC_TC)
71 		return BPF_PPC_STACKFRAME(ctx) - 4 * (32 - reg);
72 
73 	WARN(true, "BPF JIT is asking about unknown registers, will crash the stack");
74 	/* Use the hole we have left for alignment */
75 	return BPF_PPC_STACKFRAME(ctx) - 4;
76 }
77 
78 #define SEEN_VREG_MASK		0x1ff80000 /* Volatile registers r3-r12 */
79 #define SEEN_NVREG_FULL_MASK	0x0003ffff /* Non volatile registers r14-r31 */
80 #define SEEN_NVREG_TEMP_MASK	0x00001e01 /* BPF_REG_5, BPF_REG_AX, TMP_REG */
81 
82 void bpf_jit_realloc_regs(struct codegen_context *ctx)
83 {
84 	unsigned int nvreg_mask;
85 
86 	if (ctx->seen & SEEN_FUNC)
87 		nvreg_mask = SEEN_NVREG_TEMP_MASK;
88 	else
89 		nvreg_mask = SEEN_NVREG_FULL_MASK;
90 
91 	while (ctx->seen & nvreg_mask &&
92 	      (ctx->seen & SEEN_VREG_MASK) != SEEN_VREG_MASK) {
93 		int old = 32 - fls(ctx->seen & (nvreg_mask & 0xaaaaaaab));
94 		int new = 32 - fls(~ctx->seen & (SEEN_VREG_MASK & 0xaaaaaaaa));
95 		int i;
96 
97 		for (i = BPF_REG_0; i <= TMP_REG; i++) {
98 			if (ctx->b2p[i] != old)
99 				continue;
100 			ctx->b2p[i] = new;
101 			bpf_set_seen_register(ctx, new);
102 			bpf_clear_seen_register(ctx, old);
103 			if (i != TMP_REG) {
104 				bpf_set_seen_register(ctx, new - 1);
105 				bpf_clear_seen_register(ctx, old - 1);
106 			}
107 			break;
108 		}
109 	}
110 }
111 
112 void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx)
113 {
114 	int i;
115 
116 	/* First arg comes in as a 32 bits pointer. */
117 	EMIT(PPC_RAW_MR(bpf_to_ppc(BPF_REG_1), _R3));
118 	EMIT(PPC_RAW_LI(bpf_to_ppc(BPF_REG_1) - 1, 0));
119 	EMIT(PPC_RAW_STWU(_R1, _R1, -BPF_PPC_STACKFRAME(ctx)));
120 
121 	/*
122 	 * Initialize tail_call_cnt in stack frame if we do tail calls.
123 	 * Otherwise, put in NOPs so that it can be skipped when we are
124 	 * invoked through a tail call.
125 	 */
126 	if (ctx->seen & SEEN_TAILCALL)
127 		EMIT(PPC_RAW_STW(bpf_to_ppc(BPF_REG_1) - 1, _R1,
128 				 bpf_jit_stack_offsetof(ctx, BPF_PPC_TC)));
129 	else
130 		EMIT(PPC_RAW_NOP());
131 
132 #define BPF_TAILCALL_PROLOGUE_SIZE	16
133 
134 	/*
135 	 * We need a stack frame, but we don't necessarily need to
136 	 * save/restore LR unless we call other functions
137 	 */
138 	if (ctx->seen & SEEN_FUNC)
139 		EMIT(PPC_RAW_MFLR(_R0));
140 
141 	/*
142 	 * Back up non-volatile regs -- registers r18-r31
143 	 */
144 	for (i = BPF_PPC_NVR_MIN; i <= 31; i++)
145 		if (bpf_is_seen_register(ctx, i))
146 			EMIT(PPC_RAW_STW(i, _R1, bpf_jit_stack_offsetof(ctx, i)));
147 
148 	/* If needed retrieve arguments 9 and 10, ie 5th 64 bits arg.*/
149 	if (bpf_is_seen_register(ctx, bpf_to_ppc(BPF_REG_5))) {
150 		EMIT(PPC_RAW_LWZ(bpf_to_ppc(BPF_REG_5) - 1, _R1, BPF_PPC_STACKFRAME(ctx)) + 8);
151 		EMIT(PPC_RAW_LWZ(bpf_to_ppc(BPF_REG_5), _R1, BPF_PPC_STACKFRAME(ctx)) + 12);
152 	}
153 
154 	/* Setup frame pointer to point to the bpf stack area */
155 	if (bpf_is_seen_register(ctx, bpf_to_ppc(BPF_REG_FP))) {
156 		EMIT(PPC_RAW_LI(bpf_to_ppc(BPF_REG_FP) - 1, 0));
157 		EMIT(PPC_RAW_ADDI(bpf_to_ppc(BPF_REG_FP), _R1,
158 				  STACK_FRAME_MIN_SIZE + ctx->stack_size));
159 	}
160 
161 	if (ctx->seen & SEEN_FUNC)
162 		EMIT(PPC_RAW_STW(_R0, _R1, BPF_PPC_STACKFRAME(ctx) + PPC_LR_STKOFF));
163 }
164 
165 static void bpf_jit_emit_common_epilogue(u32 *image, struct codegen_context *ctx)
166 {
167 	int i;
168 
169 	/* Restore NVRs */
170 	for (i = BPF_PPC_NVR_MIN; i <= 31; i++)
171 		if (bpf_is_seen_register(ctx, i))
172 			EMIT(PPC_RAW_LWZ(i, _R1, bpf_jit_stack_offsetof(ctx, i)));
173 }
174 
175 void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
176 {
177 	EMIT(PPC_RAW_MR(_R3, bpf_to_ppc(BPF_REG_0)));
178 
179 	bpf_jit_emit_common_epilogue(image, ctx);
180 
181 	/* Tear down our stack frame */
182 
183 	if (ctx->seen & SEEN_FUNC)
184 		EMIT(PPC_RAW_LWZ(_R0, _R1, BPF_PPC_STACKFRAME(ctx) + PPC_LR_STKOFF));
185 
186 	EMIT(PPC_RAW_ADDI(_R1, _R1, BPF_PPC_STACKFRAME(ctx)));
187 
188 	if (ctx->seen & SEEN_FUNC)
189 		EMIT(PPC_RAW_MTLR(_R0));
190 
191 	EMIT(PPC_RAW_BLR());
192 }
193 
194 int bpf_jit_emit_func_call_rel(u32 *image, struct codegen_context *ctx, u64 func)
195 {
196 	s32 rel = (s32)func - (s32)(image + ctx->idx);
197 
198 	if (image && rel < 0x2000000 && rel >= -0x2000000) {
199 		PPC_BL(func);
200 		EMIT(PPC_RAW_NOP());
201 		EMIT(PPC_RAW_NOP());
202 		EMIT(PPC_RAW_NOP());
203 	} else {
204 		/* Load function address into r0 */
205 		EMIT(PPC_RAW_LIS(_R0, IMM_H(func)));
206 		EMIT(PPC_RAW_ORI(_R0, _R0, IMM_L(func)));
207 		EMIT(PPC_RAW_MTCTR(_R0));
208 		EMIT(PPC_RAW_BCTRL());
209 	}
210 
211 	return 0;
212 }
213 
214 static int bpf_jit_emit_tail_call(u32 *image, struct codegen_context *ctx, u32 out)
215 {
216 	/*
217 	 * By now, the eBPF program has already setup parameters in r3-r6
218 	 * r3-r4/BPF_REG_1 - pointer to ctx -- passed as is to the next bpf program
219 	 * r5-r6/BPF_REG_2 - pointer to bpf_array
220 	 * r7-r8/BPF_REG_3 - index in bpf_array
221 	 */
222 	int b2p_bpf_array = bpf_to_ppc(BPF_REG_2);
223 	int b2p_index = bpf_to_ppc(BPF_REG_3);
224 
225 	/*
226 	 * if (index >= array->map.max_entries)
227 	 *   goto out;
228 	 */
229 	EMIT(PPC_RAW_LWZ(_R0, b2p_bpf_array, offsetof(struct bpf_array, map.max_entries)));
230 	EMIT(PPC_RAW_CMPLW(b2p_index, _R0));
231 	EMIT(PPC_RAW_LWZ(_R0, _R1, bpf_jit_stack_offsetof(ctx, BPF_PPC_TC)));
232 	PPC_BCC_SHORT(COND_GE, out);
233 
234 	/*
235 	 * if (tail_call_cnt >= MAX_TAIL_CALL_CNT)
236 	 *   goto out;
237 	 */
238 	EMIT(PPC_RAW_CMPLWI(_R0, MAX_TAIL_CALL_CNT));
239 	/* tail_call_cnt++; */
240 	EMIT(PPC_RAW_ADDIC(_R0, _R0, 1));
241 	PPC_BCC_SHORT(COND_GE, out);
242 
243 	/* prog = array->ptrs[index]; */
244 	EMIT(PPC_RAW_RLWINM(_R3, b2p_index, 2, 0, 29));
245 	EMIT(PPC_RAW_ADD(_R3, _R3, b2p_bpf_array));
246 	EMIT(PPC_RAW_LWZ(_R3, _R3, offsetof(struct bpf_array, ptrs)));
247 	EMIT(PPC_RAW_STW(_R0, _R1, bpf_jit_stack_offsetof(ctx, BPF_PPC_TC)));
248 
249 	/*
250 	 * if (prog == NULL)
251 	 *   goto out;
252 	 */
253 	EMIT(PPC_RAW_CMPLWI(_R3, 0));
254 	PPC_BCC_SHORT(COND_EQ, out);
255 
256 	/* goto *(prog->bpf_func + prologue_size); */
257 	EMIT(PPC_RAW_LWZ(_R3, _R3, offsetof(struct bpf_prog, bpf_func)));
258 
259 	if (ctx->seen & SEEN_FUNC)
260 		EMIT(PPC_RAW_LWZ(_R0, _R1, BPF_PPC_STACKFRAME(ctx) + PPC_LR_STKOFF));
261 
262 	EMIT(PPC_RAW_ADDIC(_R3, _R3, BPF_TAILCALL_PROLOGUE_SIZE));
263 
264 	if (ctx->seen & SEEN_FUNC)
265 		EMIT(PPC_RAW_MTLR(_R0));
266 
267 	EMIT(PPC_RAW_MTCTR(_R3));
268 
269 	EMIT(PPC_RAW_MR(_R3, bpf_to_ppc(BPF_REG_1)));
270 
271 	/* tear restore NVRs, ... */
272 	bpf_jit_emit_common_epilogue(image, ctx);
273 
274 	EMIT(PPC_RAW_BCTR());
275 
276 	/* out: */
277 	return 0;
278 }
279 
280 /* Assemble the body code between the prologue & epilogue */
281 int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *ctx,
282 		       u32 *addrs, int pass)
283 {
284 	const struct bpf_insn *insn = fp->insnsi;
285 	int flen = fp->len;
286 	int i, ret;
287 
288 	/* Start of epilogue code - will only be valid 2nd pass onwards */
289 	u32 exit_addr = addrs[flen];
290 
291 	for (i = 0; i < flen; i++) {
292 		u32 code = insn[i].code;
293 		u32 dst_reg = bpf_to_ppc(insn[i].dst_reg);
294 		u32 dst_reg_h = dst_reg - 1;
295 		u32 src_reg = bpf_to_ppc(insn[i].src_reg);
296 		u32 src_reg_h = src_reg - 1;
297 		u32 tmp_reg = bpf_to_ppc(TMP_REG);
298 		u32 size = BPF_SIZE(code);
299 		s16 off = insn[i].off;
300 		s32 imm = insn[i].imm;
301 		bool func_addr_fixed;
302 		u64 func_addr;
303 		u32 true_cond;
304 		u32 tmp_idx;
305 		int j;
306 
307 		/*
308 		 * addrs[] maps a BPF bytecode address into a real offset from
309 		 * the start of the body code.
310 		 */
311 		addrs[i] = ctx->idx * 4;
312 
313 		/*
314 		 * As an optimization, we note down which registers
315 		 * are used so that we can only save/restore those in our
316 		 * prologue and epilogue. We do this here regardless of whether
317 		 * the actual BPF instruction uses src/dst registers or not
318 		 * (for instance, BPF_CALL does not use them). The expectation
319 		 * is that those instructions will have src_reg/dst_reg set to
320 		 * 0. Even otherwise, we just lose some prologue/epilogue
321 		 * optimization but everything else should work without
322 		 * any issues.
323 		 */
324 		if (dst_reg >= 3 && dst_reg < 32) {
325 			bpf_set_seen_register(ctx, dst_reg);
326 			bpf_set_seen_register(ctx, dst_reg_h);
327 		}
328 
329 		if (src_reg >= 3 && src_reg < 32) {
330 			bpf_set_seen_register(ctx, src_reg);
331 			bpf_set_seen_register(ctx, src_reg_h);
332 		}
333 
334 		switch (code) {
335 		/*
336 		 * Arithmetic operations: ADD/SUB/MUL/DIV/MOD/NEG
337 		 */
338 		case BPF_ALU | BPF_ADD | BPF_X: /* (u32) dst += (u32) src */
339 			EMIT(PPC_RAW_ADD(dst_reg, dst_reg, src_reg));
340 			break;
341 		case BPF_ALU64 | BPF_ADD | BPF_X: /* dst += src */
342 			EMIT(PPC_RAW_ADDC(dst_reg, dst_reg, src_reg));
343 			EMIT(PPC_RAW_ADDE(dst_reg_h, dst_reg_h, src_reg_h));
344 			break;
345 		case BPF_ALU | BPF_SUB | BPF_X: /* (u32) dst -= (u32) src */
346 			EMIT(PPC_RAW_SUB(dst_reg, dst_reg, src_reg));
347 			break;
348 		case BPF_ALU64 | BPF_SUB | BPF_X: /* dst -= src */
349 			EMIT(PPC_RAW_SUBFC(dst_reg, src_reg, dst_reg));
350 			EMIT(PPC_RAW_SUBFE(dst_reg_h, src_reg_h, dst_reg_h));
351 			break;
352 		case BPF_ALU | BPF_SUB | BPF_K: /* (u32) dst -= (u32) imm */
353 			imm = -imm;
354 			fallthrough;
355 		case BPF_ALU | BPF_ADD | BPF_K: /* (u32) dst += (u32) imm */
356 			if (IMM_HA(imm) & 0xffff)
357 				EMIT(PPC_RAW_ADDIS(dst_reg, dst_reg, IMM_HA(imm)));
358 			if (IMM_L(imm))
359 				EMIT(PPC_RAW_ADDI(dst_reg, dst_reg, IMM_L(imm)));
360 			break;
361 		case BPF_ALU64 | BPF_SUB | BPF_K: /* dst -= imm */
362 			imm = -imm;
363 			fallthrough;
364 		case BPF_ALU64 | BPF_ADD | BPF_K: /* dst += imm */
365 			if (!imm)
366 				break;
367 
368 			if (imm >= -32768 && imm < 32768) {
369 				EMIT(PPC_RAW_ADDIC(dst_reg, dst_reg, imm));
370 			} else {
371 				PPC_LI32(_R0, imm);
372 				EMIT(PPC_RAW_ADDC(dst_reg, dst_reg, _R0));
373 			}
374 			if (imm >= 0 || (BPF_OP(code) == BPF_SUB && imm == 0x80000000))
375 				EMIT(PPC_RAW_ADDZE(dst_reg_h, dst_reg_h));
376 			else
377 				EMIT(PPC_RAW_ADDME(dst_reg_h, dst_reg_h));
378 			break;
379 		case BPF_ALU64 | BPF_MUL | BPF_X: /* dst *= src */
380 			bpf_set_seen_register(ctx, tmp_reg);
381 			EMIT(PPC_RAW_MULW(_R0, dst_reg, src_reg_h));
382 			EMIT(PPC_RAW_MULW(dst_reg_h, dst_reg_h, src_reg));
383 			EMIT(PPC_RAW_MULHWU(tmp_reg, dst_reg, src_reg));
384 			EMIT(PPC_RAW_MULW(dst_reg, dst_reg, src_reg));
385 			EMIT(PPC_RAW_ADD(dst_reg_h, dst_reg_h, _R0));
386 			EMIT(PPC_RAW_ADD(dst_reg_h, dst_reg_h, tmp_reg));
387 			break;
388 		case BPF_ALU | BPF_MUL | BPF_X: /* (u32) dst *= (u32) src */
389 			EMIT(PPC_RAW_MULW(dst_reg, dst_reg, src_reg));
390 			break;
391 		case BPF_ALU | BPF_MUL | BPF_K: /* (u32) dst *= (u32) imm */
392 			if (imm >= -32768 && imm < 32768) {
393 				EMIT(PPC_RAW_MULI(dst_reg, dst_reg, imm));
394 			} else {
395 				PPC_LI32(_R0, imm);
396 				EMIT(PPC_RAW_MULW(dst_reg, dst_reg, _R0));
397 			}
398 			break;
399 		case BPF_ALU64 | BPF_MUL | BPF_K: /* dst *= imm */
400 			if (!imm) {
401 				PPC_LI32(dst_reg, 0);
402 				PPC_LI32(dst_reg_h, 0);
403 				break;
404 			}
405 			if (imm == 1)
406 				break;
407 			if (imm == -1) {
408 				EMIT(PPC_RAW_SUBFIC(dst_reg, dst_reg, 0));
409 				EMIT(PPC_RAW_SUBFZE(dst_reg_h, dst_reg_h));
410 				break;
411 			}
412 			bpf_set_seen_register(ctx, tmp_reg);
413 			PPC_LI32(tmp_reg, imm);
414 			EMIT(PPC_RAW_MULW(dst_reg_h, dst_reg_h, tmp_reg));
415 			if (imm < 0)
416 				EMIT(PPC_RAW_SUB(dst_reg_h, dst_reg_h, dst_reg));
417 			EMIT(PPC_RAW_MULHWU(_R0, dst_reg, tmp_reg));
418 			EMIT(PPC_RAW_MULW(dst_reg, dst_reg, tmp_reg));
419 			EMIT(PPC_RAW_ADD(dst_reg_h, dst_reg_h, _R0));
420 			break;
421 		case BPF_ALU | BPF_DIV | BPF_X: /* (u32) dst /= (u32) src */
422 			EMIT(PPC_RAW_DIVWU(dst_reg, dst_reg, src_reg));
423 			break;
424 		case BPF_ALU | BPF_MOD | BPF_X: /* (u32) dst %= (u32) src */
425 			EMIT(PPC_RAW_DIVWU(_R0, dst_reg, src_reg));
426 			EMIT(PPC_RAW_MULW(_R0, src_reg, _R0));
427 			EMIT(PPC_RAW_SUB(dst_reg, dst_reg, _R0));
428 			break;
429 		case BPF_ALU64 | BPF_DIV | BPF_X: /* dst /= src */
430 			return -EOPNOTSUPP;
431 		case BPF_ALU64 | BPF_MOD | BPF_X: /* dst %= src */
432 			return -EOPNOTSUPP;
433 		case BPF_ALU | BPF_DIV | BPF_K: /* (u32) dst /= (u32) imm */
434 			if (!imm)
435 				return -EINVAL;
436 			if (imm == 1)
437 				break;
438 
439 			PPC_LI32(_R0, imm);
440 			EMIT(PPC_RAW_DIVWU(dst_reg, dst_reg, _R0));
441 			break;
442 		case BPF_ALU | BPF_MOD | BPF_K: /* (u32) dst %= (u32) imm */
443 			if (!imm)
444 				return -EINVAL;
445 
446 			if (!is_power_of_2((u32)imm)) {
447 				bpf_set_seen_register(ctx, tmp_reg);
448 				PPC_LI32(tmp_reg, imm);
449 				EMIT(PPC_RAW_DIVWU(_R0, dst_reg, tmp_reg));
450 				EMIT(PPC_RAW_MULW(_R0, tmp_reg, _R0));
451 				EMIT(PPC_RAW_SUB(dst_reg, dst_reg, _R0));
452 				break;
453 			}
454 			if (imm == 1)
455 				EMIT(PPC_RAW_LI(dst_reg, 0));
456 			else
457 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 32 - ilog2((u32)imm), 31));
458 
459 			break;
460 		case BPF_ALU64 | BPF_MOD | BPF_K: /* dst %= imm */
461 			if (!imm)
462 				return -EINVAL;
463 			if (imm < 0)
464 				imm = -imm;
465 			if (!is_power_of_2(imm))
466 				return -EOPNOTSUPP;
467 			if (imm == 1)
468 				EMIT(PPC_RAW_LI(dst_reg, 0));
469 			else
470 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 32 - ilog2(imm), 31));
471 			EMIT(PPC_RAW_LI(dst_reg_h, 0));
472 			break;
473 		case BPF_ALU64 | BPF_DIV | BPF_K: /* dst /= imm */
474 			if (!imm)
475 				return -EINVAL;
476 			if (!is_power_of_2(abs(imm)))
477 				return -EOPNOTSUPP;
478 
479 			if (imm < 0) {
480 				EMIT(PPC_RAW_SUBFIC(dst_reg, dst_reg, 0));
481 				EMIT(PPC_RAW_SUBFZE(dst_reg_h, dst_reg_h));
482 				imm = -imm;
483 			}
484 			if (imm == 1)
485 				break;
486 			imm = ilog2(imm);
487 			EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 32 - imm, imm, 31));
488 			EMIT(PPC_RAW_RLWIMI(dst_reg, dst_reg_h, 32 - imm, 0, imm - 1));
489 			EMIT(PPC_RAW_SRAWI(dst_reg_h, dst_reg_h, imm));
490 			break;
491 		case BPF_ALU | BPF_NEG: /* (u32) dst = -dst */
492 			EMIT(PPC_RAW_NEG(dst_reg, dst_reg));
493 			break;
494 		case BPF_ALU64 | BPF_NEG: /* dst = -dst */
495 			EMIT(PPC_RAW_SUBFIC(dst_reg, dst_reg, 0));
496 			EMIT(PPC_RAW_SUBFZE(dst_reg_h, dst_reg_h));
497 			break;
498 
499 		/*
500 		 * Logical operations: AND/OR/XOR/[A]LSH/[A]RSH
501 		 */
502 		case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
503 			EMIT(PPC_RAW_AND(dst_reg, dst_reg, src_reg));
504 			EMIT(PPC_RAW_AND(dst_reg_h, dst_reg_h, src_reg_h));
505 			break;
506 		case BPF_ALU | BPF_AND | BPF_X: /* (u32) dst = dst & src */
507 			EMIT(PPC_RAW_AND(dst_reg, dst_reg, src_reg));
508 			break;
509 		case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
510 			if (imm >= 0)
511 				EMIT(PPC_RAW_LI(dst_reg_h, 0));
512 			fallthrough;
513 		case BPF_ALU | BPF_AND | BPF_K: /* (u32) dst = dst & imm */
514 			if (!IMM_H(imm)) {
515 				EMIT(PPC_RAW_ANDI(dst_reg, dst_reg, IMM_L(imm)));
516 			} else if (!IMM_L(imm)) {
517 				EMIT(PPC_RAW_ANDIS(dst_reg, dst_reg, IMM_H(imm)));
518 			} else if (imm == (((1 << fls(imm)) - 1) ^ ((1 << (ffs(i) - 1)) - 1))) {
519 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0,
520 						    32 - fls(imm), 32 - ffs(imm)));
521 			} else {
522 				PPC_LI32(_R0, imm);
523 				EMIT(PPC_RAW_AND(dst_reg, dst_reg, _R0));
524 			}
525 			break;
526 		case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
527 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, src_reg));
528 			EMIT(PPC_RAW_OR(dst_reg_h, dst_reg_h, src_reg_h));
529 			break;
530 		case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
531 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, src_reg));
532 			break;
533 		case BPF_ALU64 | BPF_OR | BPF_K:/* dst = dst | imm */
534 			/* Sign-extended */
535 			if (imm < 0)
536 				EMIT(PPC_RAW_LI(dst_reg_h, -1));
537 			fallthrough;
538 		case BPF_ALU | BPF_OR | BPF_K:/* dst = (u32) dst | (u32) imm */
539 			if (IMM_L(imm))
540 				EMIT(PPC_RAW_ORI(dst_reg, dst_reg, IMM_L(imm)));
541 			if (IMM_H(imm))
542 				EMIT(PPC_RAW_ORIS(dst_reg, dst_reg, IMM_H(imm)));
543 			break;
544 		case BPF_ALU64 | BPF_XOR | BPF_X: /* dst ^= src */
545 			if (dst_reg == src_reg) {
546 				EMIT(PPC_RAW_LI(dst_reg, 0));
547 				EMIT(PPC_RAW_LI(dst_reg_h, 0));
548 			} else {
549 				EMIT(PPC_RAW_XOR(dst_reg, dst_reg, src_reg));
550 				EMIT(PPC_RAW_XOR(dst_reg_h, dst_reg_h, src_reg_h));
551 			}
552 			break;
553 		case BPF_ALU | BPF_XOR | BPF_X: /* (u32) dst ^= src */
554 			if (dst_reg == src_reg)
555 				EMIT(PPC_RAW_LI(dst_reg, 0));
556 			else
557 				EMIT(PPC_RAW_XOR(dst_reg, dst_reg, src_reg));
558 			break;
559 		case BPF_ALU64 | BPF_XOR | BPF_K: /* dst ^= imm */
560 			if (imm < 0)
561 				EMIT(PPC_RAW_NOR(dst_reg_h, dst_reg_h, dst_reg_h));
562 			fallthrough;
563 		case BPF_ALU | BPF_XOR | BPF_K: /* (u32) dst ^= (u32) imm */
564 			if (IMM_L(imm))
565 				EMIT(PPC_RAW_XORI(dst_reg, dst_reg, IMM_L(imm)));
566 			if (IMM_H(imm))
567 				EMIT(PPC_RAW_XORIS(dst_reg, dst_reg, IMM_H(imm)));
568 			break;
569 		case BPF_ALU | BPF_LSH | BPF_X: /* (u32) dst <<= (u32) src */
570 			EMIT(PPC_RAW_SLW(dst_reg, dst_reg, src_reg));
571 			break;
572 		case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */
573 			bpf_set_seen_register(ctx, tmp_reg);
574 			EMIT(PPC_RAW_SUBFIC(_R0, src_reg, 32));
575 			EMIT(PPC_RAW_SLW(dst_reg_h, dst_reg_h, src_reg));
576 			EMIT(PPC_RAW_ADDI(tmp_reg, src_reg, 32));
577 			EMIT(PPC_RAW_SRW(_R0, dst_reg, _R0));
578 			EMIT(PPC_RAW_SLW(tmp_reg, dst_reg, tmp_reg));
579 			EMIT(PPC_RAW_OR(dst_reg_h, dst_reg_h, _R0));
580 			EMIT(PPC_RAW_SLW(dst_reg, dst_reg, src_reg));
581 			EMIT(PPC_RAW_OR(dst_reg_h, dst_reg_h, tmp_reg));
582 			break;
583 		case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<= (u32) imm */
584 			if (!imm)
585 				break;
586 			EMIT(PPC_RAW_SLWI(dst_reg, dst_reg, imm));
587 			break;
588 		case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<= imm */
589 			if (imm < 0)
590 				return -EINVAL;
591 			if (!imm)
592 				break;
593 			if (imm < 32) {
594 				EMIT(PPC_RAW_RLWINM(dst_reg_h, dst_reg_h, imm, 0, 31 - imm));
595 				EMIT(PPC_RAW_RLWIMI(dst_reg_h, dst_reg, imm, 32 - imm, 31));
596 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, imm, 0, 31 - imm));
597 				break;
598 			}
599 			if (imm < 64)
600 				EMIT(PPC_RAW_RLWINM(dst_reg_h, dst_reg, imm, 0, 31 - imm));
601 			else
602 				EMIT(PPC_RAW_LI(dst_reg_h, 0));
603 			EMIT(PPC_RAW_LI(dst_reg, 0));
604 			break;
605 		case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */
606 			EMIT(PPC_RAW_SRW(dst_reg, dst_reg, src_reg));
607 			break;
608 		case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */
609 			bpf_set_seen_register(ctx, tmp_reg);
610 			EMIT(PPC_RAW_SUBFIC(_R0, src_reg, 32));
611 			EMIT(PPC_RAW_SRW(dst_reg, dst_reg, src_reg));
612 			EMIT(PPC_RAW_ADDI(tmp_reg, src_reg, 32));
613 			EMIT(PPC_RAW_SLW(_R0, dst_reg_h, _R0));
614 			EMIT(PPC_RAW_SRW(tmp_reg, dst_reg_h, tmp_reg));
615 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, _R0));
616 			EMIT(PPC_RAW_SRW(dst_reg_h, dst_reg_h, src_reg));
617 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, tmp_reg));
618 			break;
619 		case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */
620 			if (!imm)
621 				break;
622 			EMIT(PPC_RAW_SRWI(dst_reg, dst_reg, imm));
623 			break;
624 		case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */
625 			if (imm < 0)
626 				return -EINVAL;
627 			if (!imm)
628 				break;
629 			if (imm < 32) {
630 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 32 - imm, imm, 31));
631 				EMIT(PPC_RAW_RLWIMI(dst_reg, dst_reg_h, 32 - imm, 0, imm - 1));
632 				EMIT(PPC_RAW_RLWINM(dst_reg_h, dst_reg_h, 32 - imm, imm, 31));
633 				break;
634 			}
635 			if (imm < 64)
636 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg_h, 64 - imm, imm - 32, 31));
637 			else
638 				EMIT(PPC_RAW_LI(dst_reg, 0));
639 			EMIT(PPC_RAW_LI(dst_reg_h, 0));
640 			break;
641 		case BPF_ALU | BPF_ARSH | BPF_X: /* (s32) dst >>= src */
642 			EMIT(PPC_RAW_SRAW(dst_reg, dst_reg, src_reg));
643 			break;
644 		case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */
645 			bpf_set_seen_register(ctx, tmp_reg);
646 			EMIT(PPC_RAW_SUBFIC(_R0, src_reg, 32));
647 			EMIT(PPC_RAW_SRW(dst_reg, dst_reg, src_reg));
648 			EMIT(PPC_RAW_SLW(_R0, dst_reg_h, _R0));
649 			EMIT(PPC_RAW_ADDI(tmp_reg, src_reg, 32));
650 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, _R0));
651 			EMIT(PPC_RAW_RLWINM(_R0, tmp_reg, 0, 26, 26));
652 			EMIT(PPC_RAW_SRAW(tmp_reg, dst_reg_h, tmp_reg));
653 			EMIT(PPC_RAW_SRAW(dst_reg_h, dst_reg_h, src_reg));
654 			EMIT(PPC_RAW_SLW(tmp_reg, tmp_reg, _R0));
655 			EMIT(PPC_RAW_OR(dst_reg, dst_reg, tmp_reg));
656 			break;
657 		case BPF_ALU | BPF_ARSH | BPF_K: /* (s32) dst >>= imm */
658 			if (!imm)
659 				break;
660 			EMIT(PPC_RAW_SRAWI(dst_reg, dst_reg, imm));
661 			break;
662 		case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */
663 			if (imm < 0)
664 				return -EINVAL;
665 			if (!imm)
666 				break;
667 			if (imm < 32) {
668 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 32 - imm, imm, 31));
669 				EMIT(PPC_RAW_RLWIMI(dst_reg, dst_reg_h, 32 - imm, 0, imm - 1));
670 				EMIT(PPC_RAW_SRAWI(dst_reg_h, dst_reg_h, imm));
671 				break;
672 			}
673 			if (imm < 64)
674 				EMIT(PPC_RAW_SRAWI(dst_reg, dst_reg_h, imm - 32));
675 			else
676 				EMIT(PPC_RAW_SRAWI(dst_reg, dst_reg_h, 31));
677 			EMIT(PPC_RAW_SRAWI(dst_reg_h, dst_reg_h, 31));
678 			break;
679 
680 		/*
681 		 * MOV
682 		 */
683 		case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
684 			if (dst_reg == src_reg)
685 				break;
686 			EMIT(PPC_RAW_MR(dst_reg, src_reg));
687 			EMIT(PPC_RAW_MR(dst_reg_h, src_reg_h));
688 			break;
689 		case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */
690 			/* special mov32 for zext */
691 			if (imm == 1)
692 				EMIT(PPC_RAW_LI(dst_reg_h, 0));
693 			else if (dst_reg != src_reg)
694 				EMIT(PPC_RAW_MR(dst_reg, src_reg));
695 			break;
696 		case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */
697 			PPC_LI32(dst_reg, imm);
698 			PPC_EX32(dst_reg_h, imm);
699 			break;
700 		case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */
701 			PPC_LI32(dst_reg, imm);
702 			break;
703 
704 		/*
705 		 * BPF_FROM_BE/LE
706 		 */
707 		case BPF_ALU | BPF_END | BPF_FROM_LE:
708 			switch (imm) {
709 			case 16:
710 				/* Copy 16 bits to upper part */
711 				EMIT(PPC_RAW_RLWIMI(dst_reg, dst_reg, 16, 0, 15));
712 				/* Rotate 8 bits right & mask */
713 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 24, 16, 31));
714 				break;
715 			case 32:
716 				/*
717 				 * Rotate word left by 8 bits:
718 				 * 2 bytes are already in their final position
719 				 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4)
720 				 */
721 				EMIT(PPC_RAW_RLWINM(_R0, dst_reg, 8, 0, 31));
722 				/* Rotate 24 bits and insert byte 1 */
723 				EMIT(PPC_RAW_RLWIMI(_R0, dst_reg, 24, 0, 7));
724 				/* Rotate 24 bits and insert byte 3 */
725 				EMIT(PPC_RAW_RLWIMI(_R0, dst_reg, 24, 16, 23));
726 				EMIT(PPC_RAW_MR(dst_reg, _R0));
727 				break;
728 			case 64:
729 				bpf_set_seen_register(ctx, tmp_reg);
730 				EMIT(PPC_RAW_RLWINM(tmp_reg, dst_reg, 8, 0, 31));
731 				EMIT(PPC_RAW_RLWINM(_R0, dst_reg_h, 8, 0, 31));
732 				/* Rotate 24 bits and insert byte 1 */
733 				EMIT(PPC_RAW_RLWIMI(tmp_reg, dst_reg, 24, 0, 7));
734 				EMIT(PPC_RAW_RLWIMI(_R0, dst_reg_h, 24, 0, 7));
735 				/* Rotate 24 bits and insert byte 3 */
736 				EMIT(PPC_RAW_RLWIMI(tmp_reg, dst_reg, 24, 16, 23));
737 				EMIT(PPC_RAW_RLWIMI(_R0, dst_reg_h, 24, 16, 23));
738 				EMIT(PPC_RAW_MR(dst_reg, _R0));
739 				EMIT(PPC_RAW_MR(dst_reg_h, tmp_reg));
740 				break;
741 			}
742 			break;
743 		case BPF_ALU | BPF_END | BPF_FROM_BE:
744 			switch (imm) {
745 			case 16:
746 				/* zero-extend 16 bits into 32 bits */
747 				EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 16, 31));
748 				break;
749 			case 32:
750 			case 64:
751 				/* nop */
752 				break;
753 			}
754 			break;
755 
756 		/*
757 		 * BPF_ST NOSPEC (speculation barrier)
758 		 */
759 		case BPF_ST | BPF_NOSPEC:
760 			break;
761 
762 		/*
763 		 * BPF_ST(X)
764 		 */
765 		case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */
766 			EMIT(PPC_RAW_STB(src_reg, dst_reg, off));
767 			break;
768 		case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
769 			PPC_LI32(_R0, imm);
770 			EMIT(PPC_RAW_STB(_R0, dst_reg, off));
771 			break;
772 		case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
773 			EMIT(PPC_RAW_STH(src_reg, dst_reg, off));
774 			break;
775 		case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
776 			PPC_LI32(_R0, imm);
777 			EMIT(PPC_RAW_STH(_R0, dst_reg, off));
778 			break;
779 		case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
780 			EMIT(PPC_RAW_STW(src_reg, dst_reg, off));
781 			break;
782 		case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
783 			PPC_LI32(_R0, imm);
784 			EMIT(PPC_RAW_STW(_R0, dst_reg, off));
785 			break;
786 		case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
787 			EMIT(PPC_RAW_STW(src_reg_h, dst_reg, off));
788 			EMIT(PPC_RAW_STW(src_reg, dst_reg, off + 4));
789 			break;
790 		case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
791 			PPC_LI32(_R0, imm);
792 			EMIT(PPC_RAW_STW(_R0, dst_reg, off + 4));
793 			PPC_EX32(_R0, imm);
794 			EMIT(PPC_RAW_STW(_R0, dst_reg, off));
795 			break;
796 
797 		/*
798 		 * BPF_STX ATOMIC (atomic ops)
799 		 */
800 		case BPF_STX | BPF_ATOMIC | BPF_W:
801 			if (imm != BPF_ADD) {
802 				pr_err_ratelimited("eBPF filter atomic op code %02x (@%d) unsupported\n",
803 						   code, i);
804 				return -ENOTSUPP;
805 			}
806 
807 			/* *(u32 *)(dst + off) += src */
808 
809 			bpf_set_seen_register(ctx, tmp_reg);
810 			/* Get offset into TMP_REG */
811 			EMIT(PPC_RAW_LI(tmp_reg, off));
812 			/* load value from memory into r0 */
813 			EMIT(PPC_RAW_LWARX(_R0, tmp_reg, dst_reg, 0));
814 			/* add value from src_reg into this */
815 			EMIT(PPC_RAW_ADD(_R0, _R0, src_reg));
816 			/* store result back */
817 			EMIT(PPC_RAW_STWCX(_R0, tmp_reg, dst_reg));
818 			/* we're done if this succeeded */
819 			PPC_BCC_SHORT(COND_NE, (ctx->idx - 3) * 4);
820 			break;
821 
822 		case BPF_STX | BPF_ATOMIC | BPF_DW: /* *(u64 *)(dst + off) += src */
823 			return -EOPNOTSUPP;
824 
825 		/*
826 		 * BPF_LDX
827 		 */
828 		case BPF_LDX | BPF_MEM | BPF_B: /* dst = *(u8 *)(ul) (src + off) */
829 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
830 		case BPF_LDX | BPF_MEM | BPF_H: /* dst = *(u16 *)(ul) (src + off) */
831 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
832 		case BPF_LDX | BPF_MEM | BPF_W: /* dst = *(u32 *)(ul) (src + off) */
833 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
834 		case BPF_LDX | BPF_MEM | BPF_DW: /* dst = *(u64 *)(ul) (src + off) */
835 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
836 			/*
837 			 * As PTR_TO_BTF_ID that uses BPF_PROBE_MEM mode could either be a valid
838 			 * kernel pointer or NULL but not a userspace address, execute BPF_PROBE_MEM
839 			 * load only if addr is kernel address (see is_kernel_addr()), otherwise
840 			 * set dst_reg=0 and move on.
841 			 */
842 			if (BPF_MODE(code) == BPF_PROBE_MEM) {
843 				PPC_LI32(_R0, TASK_SIZE - off);
844 				EMIT(PPC_RAW_CMPLW(src_reg, _R0));
845 				PPC_BCC_SHORT(COND_GT, (ctx->idx + 4) * 4);
846 				EMIT(PPC_RAW_LI(dst_reg, 0));
847 				/*
848 				 * For BPF_DW case, "li reg_h,0" would be needed when
849 				 * !fp->aux->verifier_zext. Emit NOP otherwise.
850 				 *
851 				 * Note that "li reg_h,0" is emitted for BPF_B/H/W case,
852 				 * if necessary. So, jump there insted of emitting an
853 				 * additional "li reg_h,0" instruction.
854 				 */
855 				if (size == BPF_DW && !fp->aux->verifier_zext)
856 					EMIT(PPC_RAW_LI(dst_reg_h, 0));
857 				else
858 					EMIT(PPC_RAW_NOP());
859 				/*
860 				 * Need to jump two instructions instead of one for BPF_DW case
861 				 * as there are two load instructions for dst_reg_h & dst_reg
862 				 * respectively.
863 				 */
864 				if (size == BPF_DW)
865 					PPC_JMP((ctx->idx + 3) * 4);
866 				else
867 					PPC_JMP((ctx->idx + 2) * 4);
868 			}
869 
870 			switch (size) {
871 			case BPF_B:
872 				EMIT(PPC_RAW_LBZ(dst_reg, src_reg, off));
873 				break;
874 			case BPF_H:
875 				EMIT(PPC_RAW_LHZ(dst_reg, src_reg, off));
876 				break;
877 			case BPF_W:
878 				EMIT(PPC_RAW_LWZ(dst_reg, src_reg, off));
879 				break;
880 			case BPF_DW:
881 				EMIT(PPC_RAW_LWZ(dst_reg_h, src_reg, off));
882 				EMIT(PPC_RAW_LWZ(dst_reg, src_reg, off + 4));
883 				break;
884 			}
885 
886 			if (size != BPF_DW && !fp->aux->verifier_zext)
887 				EMIT(PPC_RAW_LI(dst_reg_h, 0));
888 
889 			if (BPF_MODE(code) == BPF_PROBE_MEM) {
890 				int insn_idx = ctx->idx - 1;
891 				int jmp_off = 4;
892 
893 				/*
894 				 * In case of BPF_DW, two lwz instructions are emitted, one
895 				 * for higher 32-bit and another for lower 32-bit. So, set
896 				 * ex->insn to the first of the two and jump over both
897 				 * instructions in fixup.
898 				 *
899 				 * Similarly, with !verifier_zext, two instructions are
900 				 * emitted for BPF_B/H/W case. So, set ex->insn to the
901 				 * instruction that could fault and skip over both
902 				 * instructions.
903 				 */
904 				if (size == BPF_DW || !fp->aux->verifier_zext) {
905 					insn_idx -= 1;
906 					jmp_off += 4;
907 				}
908 
909 				ret = bpf_add_extable_entry(fp, image, pass, ctx, insn_idx,
910 							    jmp_off, dst_reg);
911 				if (ret)
912 					return ret;
913 			}
914 			break;
915 
916 		/*
917 		 * Doubleword load
918 		 * 16 byte instruction that uses two 'struct bpf_insn'
919 		 */
920 		case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
921 			tmp_idx = ctx->idx;
922 			PPC_LI32(dst_reg_h, (u32)insn[i + 1].imm);
923 			PPC_LI32(dst_reg, (u32)insn[i].imm);
924 			/* padding to allow full 4 instructions for later patching */
925 			for (j = ctx->idx - tmp_idx; j < 4; j++)
926 				EMIT(PPC_RAW_NOP());
927 			/* Adjust for two bpf instructions */
928 			addrs[++i] = ctx->idx * 4;
929 			break;
930 
931 		/*
932 		 * Return/Exit
933 		 */
934 		case BPF_JMP | BPF_EXIT:
935 			/*
936 			 * If this isn't the very last instruction, branch to
937 			 * the epilogue. If we _are_ the last instruction,
938 			 * we'll just fall through to the epilogue.
939 			 */
940 			if (i != flen - 1) {
941 				ret = bpf_jit_emit_exit_insn(image, ctx, _R0, exit_addr);
942 				if (ret)
943 					return ret;
944 			}
945 			/* else fall through to the epilogue */
946 			break;
947 
948 		/*
949 		 * Call kernel helper or bpf function
950 		 */
951 		case BPF_JMP | BPF_CALL:
952 			ctx->seen |= SEEN_FUNC;
953 
954 			ret = bpf_jit_get_func_addr(fp, &insn[i], false,
955 						    &func_addr, &func_addr_fixed);
956 			if (ret < 0)
957 				return ret;
958 
959 			if (bpf_is_seen_register(ctx, bpf_to_ppc(BPF_REG_5))) {
960 				EMIT(PPC_RAW_STW(bpf_to_ppc(BPF_REG_5) - 1, _R1, 8));
961 				EMIT(PPC_RAW_STW(bpf_to_ppc(BPF_REG_5), _R1, 12));
962 			}
963 
964 			ret = bpf_jit_emit_func_call_rel(image, ctx, func_addr);
965 			if (ret)
966 				return ret;
967 
968 			EMIT(PPC_RAW_MR(bpf_to_ppc(BPF_REG_0) - 1, _R3));
969 			EMIT(PPC_RAW_MR(bpf_to_ppc(BPF_REG_0), _R4));
970 			break;
971 
972 		/*
973 		 * Jumps and branches
974 		 */
975 		case BPF_JMP | BPF_JA:
976 			PPC_JMP(addrs[i + 1 + off]);
977 			break;
978 
979 		case BPF_JMP | BPF_JGT | BPF_K:
980 		case BPF_JMP | BPF_JGT | BPF_X:
981 		case BPF_JMP | BPF_JSGT | BPF_K:
982 		case BPF_JMP | BPF_JSGT | BPF_X:
983 		case BPF_JMP32 | BPF_JGT | BPF_K:
984 		case BPF_JMP32 | BPF_JGT | BPF_X:
985 		case BPF_JMP32 | BPF_JSGT | BPF_K:
986 		case BPF_JMP32 | BPF_JSGT | BPF_X:
987 			true_cond = COND_GT;
988 			goto cond_branch;
989 		case BPF_JMP | BPF_JLT | BPF_K:
990 		case BPF_JMP | BPF_JLT | BPF_X:
991 		case BPF_JMP | BPF_JSLT | BPF_K:
992 		case BPF_JMP | BPF_JSLT | BPF_X:
993 		case BPF_JMP32 | BPF_JLT | BPF_K:
994 		case BPF_JMP32 | BPF_JLT | BPF_X:
995 		case BPF_JMP32 | BPF_JSLT | BPF_K:
996 		case BPF_JMP32 | BPF_JSLT | BPF_X:
997 			true_cond = COND_LT;
998 			goto cond_branch;
999 		case BPF_JMP | BPF_JGE | BPF_K:
1000 		case BPF_JMP | BPF_JGE | BPF_X:
1001 		case BPF_JMP | BPF_JSGE | BPF_K:
1002 		case BPF_JMP | BPF_JSGE | BPF_X:
1003 		case BPF_JMP32 | BPF_JGE | BPF_K:
1004 		case BPF_JMP32 | BPF_JGE | BPF_X:
1005 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1006 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1007 			true_cond = COND_GE;
1008 			goto cond_branch;
1009 		case BPF_JMP | BPF_JLE | BPF_K:
1010 		case BPF_JMP | BPF_JLE | BPF_X:
1011 		case BPF_JMP | BPF_JSLE | BPF_K:
1012 		case BPF_JMP | BPF_JSLE | BPF_X:
1013 		case BPF_JMP32 | BPF_JLE | BPF_K:
1014 		case BPF_JMP32 | BPF_JLE | BPF_X:
1015 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1016 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1017 			true_cond = COND_LE;
1018 			goto cond_branch;
1019 		case BPF_JMP | BPF_JEQ | BPF_K:
1020 		case BPF_JMP | BPF_JEQ | BPF_X:
1021 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1022 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1023 			true_cond = COND_EQ;
1024 			goto cond_branch;
1025 		case BPF_JMP | BPF_JNE | BPF_K:
1026 		case BPF_JMP | BPF_JNE | BPF_X:
1027 		case BPF_JMP32 | BPF_JNE | BPF_K:
1028 		case BPF_JMP32 | BPF_JNE | BPF_X:
1029 			true_cond = COND_NE;
1030 			goto cond_branch;
1031 		case BPF_JMP | BPF_JSET | BPF_K:
1032 		case BPF_JMP | BPF_JSET | BPF_X:
1033 		case BPF_JMP32 | BPF_JSET | BPF_K:
1034 		case BPF_JMP32 | BPF_JSET | BPF_X:
1035 			true_cond = COND_NE;
1036 			/* fallthrough; */
1037 
1038 cond_branch:
1039 			switch (code) {
1040 			case BPF_JMP | BPF_JGT | BPF_X:
1041 			case BPF_JMP | BPF_JLT | BPF_X:
1042 			case BPF_JMP | BPF_JGE | BPF_X:
1043 			case BPF_JMP | BPF_JLE | BPF_X:
1044 			case BPF_JMP | BPF_JEQ | BPF_X:
1045 			case BPF_JMP | BPF_JNE | BPF_X:
1046 				/* unsigned comparison */
1047 				EMIT(PPC_RAW_CMPLW(dst_reg_h, src_reg_h));
1048 				PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1049 				EMIT(PPC_RAW_CMPLW(dst_reg, src_reg));
1050 				break;
1051 			case BPF_JMP32 | BPF_JGT | BPF_X:
1052 			case BPF_JMP32 | BPF_JLT | BPF_X:
1053 			case BPF_JMP32 | BPF_JGE | BPF_X:
1054 			case BPF_JMP32 | BPF_JLE | BPF_X:
1055 			case BPF_JMP32 | BPF_JEQ | BPF_X:
1056 			case BPF_JMP32 | BPF_JNE | BPF_X:
1057 				/* unsigned comparison */
1058 				EMIT(PPC_RAW_CMPLW(dst_reg, src_reg));
1059 				break;
1060 			case BPF_JMP | BPF_JSGT | BPF_X:
1061 			case BPF_JMP | BPF_JSLT | BPF_X:
1062 			case BPF_JMP | BPF_JSGE | BPF_X:
1063 			case BPF_JMP | BPF_JSLE | BPF_X:
1064 				/* signed comparison */
1065 				EMIT(PPC_RAW_CMPW(dst_reg_h, src_reg_h));
1066 				PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1067 				EMIT(PPC_RAW_CMPLW(dst_reg, src_reg));
1068 				break;
1069 			case BPF_JMP32 | BPF_JSGT | BPF_X:
1070 			case BPF_JMP32 | BPF_JSLT | BPF_X:
1071 			case BPF_JMP32 | BPF_JSGE | BPF_X:
1072 			case BPF_JMP32 | BPF_JSLE | BPF_X:
1073 				/* signed comparison */
1074 				EMIT(PPC_RAW_CMPW(dst_reg, src_reg));
1075 				break;
1076 			case BPF_JMP | BPF_JSET | BPF_X:
1077 				EMIT(PPC_RAW_AND_DOT(_R0, dst_reg_h, src_reg_h));
1078 				PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1079 				EMIT(PPC_RAW_AND_DOT(_R0, dst_reg, src_reg));
1080 				break;
1081 			case BPF_JMP32 | BPF_JSET | BPF_X: {
1082 				EMIT(PPC_RAW_AND_DOT(_R0, dst_reg, src_reg));
1083 				break;
1084 			case BPF_JMP | BPF_JNE | BPF_K:
1085 			case BPF_JMP | BPF_JEQ | BPF_K:
1086 			case BPF_JMP | BPF_JGT | BPF_K:
1087 			case BPF_JMP | BPF_JLT | BPF_K:
1088 			case BPF_JMP | BPF_JGE | BPF_K:
1089 			case BPF_JMP | BPF_JLE | BPF_K:
1090 				/*
1091 				 * Need sign-extended load, so only positive
1092 				 * values can be used as imm in cmplwi
1093 				 */
1094 				if (imm >= 0 && imm < 32768) {
1095 					EMIT(PPC_RAW_CMPLWI(dst_reg_h, 0));
1096 					PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1097 					EMIT(PPC_RAW_CMPLWI(dst_reg, imm));
1098 				} else {
1099 					/* sign-extending load ... but unsigned comparison */
1100 					PPC_EX32(_R0, imm);
1101 					EMIT(PPC_RAW_CMPLW(dst_reg_h, _R0));
1102 					PPC_LI32(_R0, imm);
1103 					PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1104 					EMIT(PPC_RAW_CMPLW(dst_reg, _R0));
1105 				}
1106 				break;
1107 			case BPF_JMP32 | BPF_JNE | BPF_K:
1108 			case BPF_JMP32 | BPF_JEQ | BPF_K:
1109 			case BPF_JMP32 | BPF_JGT | BPF_K:
1110 			case BPF_JMP32 | BPF_JLT | BPF_K:
1111 			case BPF_JMP32 | BPF_JGE | BPF_K:
1112 			case BPF_JMP32 | BPF_JLE | BPF_K:
1113 				if (imm >= 0 && imm < 65536) {
1114 					EMIT(PPC_RAW_CMPLWI(dst_reg, imm));
1115 				} else {
1116 					PPC_LI32(_R0, imm);
1117 					EMIT(PPC_RAW_CMPLW(dst_reg, _R0));
1118 				}
1119 				break;
1120 			}
1121 			case BPF_JMP | BPF_JSGT | BPF_K:
1122 			case BPF_JMP | BPF_JSLT | BPF_K:
1123 			case BPF_JMP | BPF_JSGE | BPF_K:
1124 			case BPF_JMP | BPF_JSLE | BPF_K:
1125 				if (imm >= 0 && imm < 65536) {
1126 					EMIT(PPC_RAW_CMPWI(dst_reg_h, imm < 0 ? -1 : 0));
1127 					PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1128 					EMIT(PPC_RAW_CMPLWI(dst_reg, imm));
1129 				} else {
1130 					/* sign-extending load */
1131 					EMIT(PPC_RAW_CMPWI(dst_reg_h, imm < 0 ? -1 : 0));
1132 					PPC_LI32(_R0, imm);
1133 					PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1134 					EMIT(PPC_RAW_CMPLW(dst_reg, _R0));
1135 				}
1136 				break;
1137 			case BPF_JMP32 | BPF_JSGT | BPF_K:
1138 			case BPF_JMP32 | BPF_JSLT | BPF_K:
1139 			case BPF_JMP32 | BPF_JSGE | BPF_K:
1140 			case BPF_JMP32 | BPF_JSLE | BPF_K:
1141 				/*
1142 				 * signed comparison, so any 16-bit value
1143 				 * can be used in cmpwi
1144 				 */
1145 				if (imm >= -32768 && imm < 32768) {
1146 					EMIT(PPC_RAW_CMPWI(dst_reg, imm));
1147 				} else {
1148 					/* sign-extending load */
1149 					PPC_LI32(_R0, imm);
1150 					EMIT(PPC_RAW_CMPW(dst_reg, _R0));
1151 				}
1152 				break;
1153 			case BPF_JMP | BPF_JSET | BPF_K:
1154 				/* andi does not sign-extend the immediate */
1155 				if (imm >= 0 && imm < 32768) {
1156 					/* PPC_ANDI is _only/always_ dot-form */
1157 					EMIT(PPC_RAW_ANDI(_R0, dst_reg, imm));
1158 				} else {
1159 					PPC_LI32(_R0, imm);
1160 					if (imm < 0) {
1161 						EMIT(PPC_RAW_CMPWI(dst_reg_h, 0));
1162 						PPC_BCC_SHORT(COND_NE, (ctx->idx + 2) * 4);
1163 					}
1164 					EMIT(PPC_RAW_AND_DOT(_R0, dst_reg, _R0));
1165 				}
1166 				break;
1167 			case BPF_JMP32 | BPF_JSET | BPF_K:
1168 				/* andi does not sign-extend the immediate */
1169 				if (imm >= 0 && imm < 32768) {
1170 					/* PPC_ANDI is _only/always_ dot-form */
1171 					EMIT(PPC_RAW_ANDI(_R0, dst_reg, imm));
1172 				} else {
1173 					PPC_LI32(_R0, imm);
1174 					EMIT(PPC_RAW_AND_DOT(_R0, dst_reg, _R0));
1175 				}
1176 				break;
1177 			}
1178 			PPC_BCC(true_cond, addrs[i + 1 + off]);
1179 			break;
1180 
1181 		/*
1182 		 * Tail call
1183 		 */
1184 		case BPF_JMP | BPF_TAIL_CALL:
1185 			ctx->seen |= SEEN_TAILCALL;
1186 			ret = bpf_jit_emit_tail_call(image, ctx, addrs[i + 1]);
1187 			if (ret < 0)
1188 				return ret;
1189 			break;
1190 
1191 		default:
1192 			/*
1193 			 * The filter contains something cruel & unusual.
1194 			 * We don't handle it, but also there shouldn't be
1195 			 * anything missing from our list.
1196 			 */
1197 			pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n", code, i);
1198 			return -EOPNOTSUPP;
1199 		}
1200 		if (BPF_CLASS(code) == BPF_ALU && !fp->aux->verifier_zext &&
1201 		    !insn_is_zext(&insn[i + 1]) && !(BPF_OP(code) == BPF_END && imm == 64))
1202 			EMIT(PPC_RAW_LI(dst_reg_h, 0));
1203 	}
1204 
1205 	/* Set end-of-body-code address for exit. */
1206 	addrs[i] = ctx->idx * 4;
1207 
1208 	return 0;
1209 }
1210