xref: /linux/arch/mips/net/bpf_jit_comp64.c (revision 148ab873d8194cb44f29474da3b4428fa2e100d1)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Just-In-Time compiler for eBPF bytecode on MIPS.
4  * Implementation of JIT functions for 64-bit CPUs.
5  *
6  * Copyright (c) 2021 Anyfi Networks AB.
7  * Author: Johan Almbladh <johan.almbladh@gmail.com>
8  *
9  * Based on code and ideas from
10  * Copyright (c) 2017 Cavium, Inc.
11  * Copyright (c) 2017 Shubham Bansal <illusionist.neo@gmail.com>
12  * Copyright (c) 2011 Mircea Gherzan <mgherzan@gmail.com>
13  */
14 
15 #include <linux/errno.h>
16 #include <linux/filter.h>
17 #include <linux/bpf.h>
18 #include <asm/cpu-features.h>
19 #include <asm/isa-rev.h>
20 #include <asm/uasm.h>
21 
22 #include "bpf_jit_comp.h"
23 
24 /* MIPS t0-t3 are not available in the n64 ABI */
25 #undef MIPS_R_T0
26 #undef MIPS_R_T1
27 #undef MIPS_R_T2
28 #undef MIPS_R_T3
29 
30 /* Stack is 16-byte aligned in n64 ABI */
31 #define MIPS_STACK_ALIGNMENT 16
32 
33 /* Extra 64-bit eBPF registers used by JIT */
34 #define JIT_REG_TC (MAX_BPF_JIT_REG + 0)
35 #define JIT_REG_ZX (MAX_BPF_JIT_REG + 1)
36 
37 /* Number of prologue bytes to skip when doing a tail call */
38 #define JIT_TCALL_SKIP 4
39 
40 /* Callee-saved CPU registers that the JIT must preserve */
41 #define JIT_CALLEE_REGS   \
42 	(BIT(MIPS_R_S0) | \
43 	 BIT(MIPS_R_S1) | \
44 	 BIT(MIPS_R_S2) | \
45 	 BIT(MIPS_R_S3) | \
46 	 BIT(MIPS_R_S4) | \
47 	 BIT(MIPS_R_S5) | \
48 	 BIT(MIPS_R_S6) | \
49 	 BIT(MIPS_R_S7) | \
50 	 BIT(MIPS_R_GP) | \
51 	 BIT(MIPS_R_FP) | \
52 	 BIT(MIPS_R_RA))
53 
54 /* Caller-saved CPU registers available for JIT use */
55 #define JIT_CALLER_REGS	  \
56 	(BIT(MIPS_R_A5) | \
57 	 BIT(MIPS_R_A6) | \
58 	 BIT(MIPS_R_A7))
59 /*
60  * Mapping of 64-bit eBPF registers to 64-bit native MIPS registers.
61  * MIPS registers t4 - t7 may be used by the JIT as temporary registers.
62  * MIPS registers t8 - t9 are reserved for single-register common functions.
63  */
64 static const u8 bpf2mips64[] = {
65 	/* Return value from in-kernel function, and exit value from eBPF */
66 	[BPF_REG_0] = MIPS_R_V0,
67 	/* Arguments from eBPF program to in-kernel function */
68 	[BPF_REG_1] = MIPS_R_A0,
69 	[BPF_REG_2] = MIPS_R_A1,
70 	[BPF_REG_3] = MIPS_R_A2,
71 	[BPF_REG_4] = MIPS_R_A3,
72 	[BPF_REG_5] = MIPS_R_A4,
73 	/* Callee-saved registers that in-kernel function will preserve */
74 	[BPF_REG_6] = MIPS_R_S0,
75 	[BPF_REG_7] = MIPS_R_S1,
76 	[BPF_REG_8] = MIPS_R_S2,
77 	[BPF_REG_9] = MIPS_R_S3,
78 	/* Read-only frame pointer to access the eBPF stack */
79 	[BPF_REG_FP] = MIPS_R_FP,
80 	/* Temporary register for blinding constants */
81 	[BPF_REG_AX] = MIPS_R_AT,
82 	/* Tail call count register, caller-saved */
83 	[JIT_REG_TC] = MIPS_R_A5,
84 	/* Constant for register zero-extension */
85 	[JIT_REG_ZX] = MIPS_R_V1,
86 };
87 
88 /*
89  * MIPS 32-bit operations on 64-bit registers generate a sign-extended
90  * result. However, the eBPF ISA mandates zero-extension, so we rely on the
91  * verifier to add that for us (emit_zext_ver). In addition, ALU arithmetic
92  * operations, right shift and byte swap require properly sign-extended
93  * operands or the result is unpredictable. We emit explicit sign-extensions
94  * in those cases.
95  */
96 
97 /* Sign extension */
98 static void emit_sext(struct jit_context *ctx, u8 dst, u8 src)
99 {
100 	emit(ctx, sll, dst, src, 0);
101 	clobber_reg(ctx, dst);
102 }
103 
104 /* Zero extension */
105 static void emit_zext(struct jit_context *ctx, u8 dst)
106 {
107 	if (cpu_has_mips64r2 || cpu_has_mips64r6) {
108 		emit(ctx, dinsu, dst, MIPS_R_ZERO, 32, 32);
109 	} else {
110 		emit(ctx, and, dst, dst, bpf2mips64[JIT_REG_ZX]);
111 		access_reg(ctx, JIT_REG_ZX); /* We need the ZX register */
112 	}
113 	clobber_reg(ctx, dst);
114 }
115 
116 /* Zero extension, if verifier does not do it for us  */
117 static void emit_zext_ver(struct jit_context *ctx, u8 dst)
118 {
119 	if (!ctx->program->aux->verifier_zext)
120 		emit_zext(ctx, dst);
121 }
122 
123 /* dst = imm (64-bit) */
124 static void emit_mov_i64(struct jit_context *ctx, u8 dst, u64 imm64)
125 {
126 	if (imm64 >= 0xffffffffffff8000ULL || imm64 < 0x8000ULL) {
127 		emit(ctx, daddiu, dst, MIPS_R_ZERO, (s16)imm64);
128 	} else if (imm64 >= 0xffffffff80000000ULL ||
129 		   (imm64 < 0x80000000 && imm64 > 0xffff)) {
130 		emit(ctx, lui, dst, (s16)(imm64 >> 16));
131 		emit(ctx, ori, dst, dst, (u16)imm64 & 0xffff);
132 	} else {
133 		u8 acc = MIPS_R_ZERO;
134 		int shift = 0;
135 		int k;
136 
137 		for (k = 0; k < 4; k++) {
138 			u16 half = imm64 >> (48 - 16 * k);
139 
140 			if (acc == dst)
141 				shift += 16;
142 
143 			if (half) {
144 				if (shift)
145 					emit(ctx, dsll_safe, dst, dst, shift);
146 				emit(ctx, ori, dst, acc, half);
147 				acc = dst;
148 				shift = 0;
149 			}
150 		}
151 		if (shift)
152 			emit(ctx, dsll_safe, dst, dst, shift);
153 	}
154 	clobber_reg(ctx, dst);
155 }
156 
157 /* ALU immediate operation (64-bit) */
158 static void emit_alu_i64(struct jit_context *ctx, u8 dst, s32 imm, u8 op)
159 {
160 	switch (BPF_OP(op)) {
161 	/* dst = dst | imm */
162 	case BPF_OR:
163 		emit(ctx, ori, dst, dst, (u16)imm);
164 		break;
165 	/* dst = dst ^ imm */
166 	case BPF_XOR:
167 		emit(ctx, xori, dst, dst, (u16)imm);
168 		break;
169 	/* dst = -dst */
170 	case BPF_NEG:
171 		emit(ctx, dsubu, dst, MIPS_R_ZERO, dst);
172 		break;
173 	/* dst = dst << imm */
174 	case BPF_LSH:
175 		emit(ctx, dsll_safe, dst, dst, imm);
176 		break;
177 	/* dst = dst >> imm */
178 	case BPF_RSH:
179 		emit(ctx, dsrl_safe, dst, dst, imm);
180 		break;
181 	/* dst = dst >> imm (arithmetic) */
182 	case BPF_ARSH:
183 		emit(ctx, dsra_safe, dst, dst, imm);
184 		break;
185 	/* dst = dst + imm */
186 	case BPF_ADD:
187 		emit(ctx, daddiu, dst, dst, imm);
188 		break;
189 	/* dst = dst - imm */
190 	case BPF_SUB:
191 		emit(ctx, daddiu, dst, dst, -imm);
192 		break;
193 	default:
194 		/* Width-generic operations */
195 		emit_alu_i(ctx, dst, imm, op);
196 	}
197 	clobber_reg(ctx, dst);
198 }
199 
200 /* ALU register operation (64-bit) */
201 static void emit_alu_r64(struct jit_context *ctx, u8 dst, u8 src, u8 op)
202 {
203 	switch (BPF_OP(op)) {
204 	/* dst = dst << src */
205 	case BPF_LSH:
206 		emit(ctx, dsllv, dst, dst, src);
207 		break;
208 	/* dst = dst >> src */
209 	case BPF_RSH:
210 		emit(ctx, dsrlv, dst, dst, src);
211 		break;
212 	/* dst = dst >> src (arithmetic) */
213 	case BPF_ARSH:
214 		emit(ctx, dsrav, dst, dst, src);
215 		break;
216 	/* dst = dst + src */
217 	case BPF_ADD:
218 		emit(ctx, daddu, dst, dst, src);
219 		break;
220 	/* dst = dst - src */
221 	case BPF_SUB:
222 		emit(ctx, dsubu, dst, dst, src);
223 		break;
224 	/* dst = dst * src */
225 	case BPF_MUL:
226 		if (cpu_has_mips64r6) {
227 			emit(ctx, dmulu, dst, dst, src);
228 		} else {
229 			emit(ctx, dmultu, dst, src);
230 			emit(ctx, mflo, dst);
231 			/* Ensure multiplication is completed */
232 			if (IS_ENABLED(CONFIG_CPU_R4000_WORKAROUNDS))
233 				emit(ctx, mfhi, MIPS_R_ZERO);
234 		}
235 		break;
236 	/* dst = dst / src */
237 	case BPF_DIV:
238 		if (cpu_has_mips64r6) {
239 			emit(ctx, ddivu_r6, dst, dst, src);
240 		} else {
241 			emit(ctx, ddivu, dst, src);
242 			emit(ctx, mflo, dst);
243 		}
244 		break;
245 	/* dst = dst % src */
246 	case BPF_MOD:
247 		if (cpu_has_mips64r6) {
248 			emit(ctx, dmodu, dst, dst, src);
249 		} else {
250 			emit(ctx, ddivu, dst, src);
251 			emit(ctx, mfhi, dst);
252 		}
253 		break;
254 	default:
255 		/* Width-generic operations */
256 		emit_alu_r(ctx, dst, src, op);
257 	}
258 	clobber_reg(ctx, dst);
259 }
260 
261 /* Swap sub words in a register double word */
262 static void emit_swap_r64(struct jit_context *ctx, u8 dst, u8 mask, u32 bits)
263 {
264 	u8 tmp = MIPS_R_T9;
265 
266 	emit(ctx, and, tmp, dst, mask);  /* tmp = dst & mask  */
267 	emit(ctx, dsll, tmp, tmp, bits); /* tmp = tmp << bits */
268 	emit(ctx, dsrl, dst, dst, bits); /* dst = dst >> bits */
269 	emit(ctx, and, dst, dst, mask);  /* dst = dst & mask  */
270 	emit(ctx, or, dst, dst, tmp);    /* dst = dst | tmp   */
271 }
272 
273 /* Swap bytes and truncate a register double word, word or half word */
274 static void emit_bswap_r64(struct jit_context *ctx, u8 dst, u32 width)
275 {
276 	switch (width) {
277 	/* Swap bytes in a double word */
278 	case 64:
279 		if (cpu_has_mips64r2 || cpu_has_mips64r6) {
280 			emit(ctx, dsbh, dst, dst);
281 			emit(ctx, dshd, dst, dst);
282 		} else {
283 			u8 t1 = MIPS_R_T6;
284 			u8 t2 = MIPS_R_T7;
285 
286 			emit(ctx, dsll32, t2, dst, 0);  /* t2 = dst << 32    */
287 			emit(ctx, dsrl32, dst, dst, 0); /* dst = dst >> 32   */
288 			emit(ctx, or, dst, dst, t2);    /* dst = dst | t2    */
289 
290 			emit(ctx, ori, t2, MIPS_R_ZERO, 0xffff);
291 			emit(ctx, dsll32, t1, t2, 0);   /* t1 = t2 << 32     */
292 			emit(ctx, or, t1, t1, t2);      /* t1 = t1 | t2      */
293 			emit_swap_r64(ctx, dst, t1, 16);/* dst = swap16(dst) */
294 
295 			emit(ctx, lui, t2, 0xff);       /* t2 = 0x00ff0000   */
296 			emit(ctx, ori, t2, t2, 0xff);   /* t2 = t2 | 0x00ff  */
297 			emit(ctx, dsll32, t1, t2, 0);   /* t1 = t2 << 32     */
298 			emit(ctx, or, t1, t1, t2);      /* t1 = t1 | t2      */
299 			emit_swap_r64(ctx, dst, t1, 8); /* dst = swap8(dst)  */
300 		}
301 		break;
302 	/* Swap bytes in a half word */
303 	/* Swap bytes in a word */
304 	case 32:
305 	case 16:
306 		emit_sext(ctx, dst, dst);
307 		emit_bswap_r(ctx, dst, width);
308 		emit_zext(ctx, dst);
309 		break;
310 	}
311 	clobber_reg(ctx, dst);
312 }
313 
314 /* Truncate a register double word, word or half word */
315 static void emit_trunc_r64(struct jit_context *ctx, u8 dst, u32 width)
316 {
317 	switch (width) {
318 	case 64:
319 		break;
320 	/* Zero-extend a word */
321 	case 32:
322 		emit_zext(ctx, dst);
323 		break;
324 	/* Zero-extend a half word */
325 	case 16:
326 		emit(ctx, andi, dst, dst, 0xffff);
327 		break;
328 	}
329 	clobber_reg(ctx, dst);
330 }
331 
332 /* Load operation: dst = *(size*)(src + off) */
333 static void emit_ldx(struct jit_context *ctx, u8 dst, u8 src, s16 off, u8 size)
334 {
335 	switch (size) {
336 	/* Load a byte */
337 	case BPF_B:
338 		emit(ctx, lbu, dst, off, src);
339 		break;
340 	/* Load a half word */
341 	case BPF_H:
342 		emit(ctx, lhu, dst, off, src);
343 		break;
344 	/* Load a word */
345 	case BPF_W:
346 		emit(ctx, lwu, dst, off, src);
347 		break;
348 	/* Load a double word */
349 	case BPF_DW:
350 		emit(ctx, ld, dst, off, src);
351 		break;
352 	}
353 	clobber_reg(ctx, dst);
354 }
355 
356 /* Store operation: *(size *)(dst + off) = src */
357 static void emit_stx(struct jit_context *ctx, u8 dst, u8 src, s16 off, u8 size)
358 {
359 	switch (size) {
360 	/* Store a byte */
361 	case BPF_B:
362 		emit(ctx, sb, src, off, dst);
363 		break;
364 	/* Store a half word */
365 	case BPF_H:
366 		emit(ctx, sh, src, off, dst);
367 		break;
368 	/* Store a word */
369 	case BPF_W:
370 		emit(ctx, sw, src, off, dst);
371 		break;
372 	/* Store a double word */
373 	case BPF_DW:
374 		emit(ctx, sd, src, off, dst);
375 		break;
376 	}
377 }
378 
379 /* Atomic read-modify-write */
380 static void emit_atomic_r64(struct jit_context *ctx,
381 			    u8 dst, u8 src, s16 off, u8 code)
382 {
383 	u8 t1 = MIPS_R_T6;
384 	u8 t2 = MIPS_R_T7;
385 
386 	LLSC_sync(ctx);
387 	emit(ctx, lld, t1, off, dst);
388 	switch (code) {
389 	case BPF_ADD:
390 	case BPF_ADD | BPF_FETCH:
391 		emit(ctx, daddu, t2, t1, src);
392 		break;
393 	case BPF_AND:
394 	case BPF_AND | BPF_FETCH:
395 		emit(ctx, and, t2, t1, src);
396 		break;
397 	case BPF_OR:
398 	case BPF_OR | BPF_FETCH:
399 		emit(ctx, or, t2, t1, src);
400 		break;
401 	case BPF_XOR:
402 	case BPF_XOR | BPF_FETCH:
403 		emit(ctx, xor, t2, t1, src);
404 		break;
405 	case BPF_XCHG:
406 		emit(ctx, move, t2, src);
407 		break;
408 	}
409 	emit(ctx, scd, t2, off, dst);
410 	emit(ctx, LLSC_beqz, t2, -16 - LLSC_offset);
411 	emit(ctx, nop); /* Delay slot */
412 
413 	if (code & BPF_FETCH) {
414 		emit(ctx, move, src, t1);
415 		clobber_reg(ctx, src);
416 	}
417 }
418 
419 /* Atomic compare-and-exchange */
420 static void emit_cmpxchg_r64(struct jit_context *ctx, u8 dst, u8 src, s16 off)
421 {
422 	u8 r0 = bpf2mips64[BPF_REG_0];
423 	u8 t1 = MIPS_R_T6;
424 	u8 t2 = MIPS_R_T7;
425 
426 	LLSC_sync(ctx);
427 	emit(ctx, lld, t1, off, dst);
428 	emit(ctx, bne, t1, r0, 12);
429 	emit(ctx, move, t2, src);      /* Delay slot */
430 	emit(ctx, scd, t2, off, dst);
431 	emit(ctx, LLSC_beqz, t2, -20 - LLSC_offset);
432 	emit(ctx, move, r0, t1);       /* Delay slot */
433 
434 	clobber_reg(ctx, r0);
435 }
436 
437 /* Function call */
438 static int emit_call(struct jit_context *ctx, const struct bpf_insn *insn)
439 {
440 	u8 zx = bpf2mips64[JIT_REG_ZX];
441 	u8 tmp = MIPS_R_T6;
442 	bool fixed;
443 	u64 addr;
444 
445 	/* Decode the call address */
446 	if (bpf_jit_get_func_addr(ctx->program, insn, false,
447 				  &addr, &fixed) < 0)
448 		return -1;
449 	if (!fixed)
450 		return -1;
451 
452 	/* Push caller-saved registers on stack */
453 	push_regs(ctx, ctx->clobbered & JIT_CALLER_REGS, 0, 0);
454 
455 	/* Emit function call */
456 	emit_mov_i64(ctx, tmp, addr & JALR_MASK);
457 	emit(ctx, jalr, MIPS_R_RA, tmp);
458 	emit(ctx, nop); /* Delay slot */
459 
460 	/* Restore caller-saved registers */
461 	pop_regs(ctx, ctx->clobbered & JIT_CALLER_REGS, 0, 0);
462 
463 	/* Re-initialize the JIT zero-extension register if accessed */
464 	if (ctx->accessed & BIT(JIT_REG_ZX)) {
465 		emit(ctx, daddiu, zx, MIPS_R_ZERO, -1);
466 		emit(ctx, dsrl32, zx, zx, 0);
467 	}
468 
469 	clobber_reg(ctx, MIPS_R_RA);
470 	clobber_reg(ctx, MIPS_R_V0);
471 	clobber_reg(ctx, MIPS_R_V1);
472 	return 0;
473 }
474 
475 /* Function tail call */
476 static int emit_tail_call(struct jit_context *ctx)
477 {
478 	u8 ary = bpf2mips64[BPF_REG_2];
479 	u8 ind = bpf2mips64[BPF_REG_3];
480 	u8 tcc = bpf2mips64[JIT_REG_TC];
481 	u8 tmp = MIPS_R_T6;
482 	int off;
483 
484 	/*
485 	 * Tail call:
486 	 * eBPF R1 - function argument (context ptr), passed in a0-a1
487 	 * eBPF R2 - ptr to object with array of function entry points
488 	 * eBPF R3 - array index of function to be called
489 	 */
490 
491 	/* if (ind >= ary->map.max_entries) goto out */
492 	off = offsetof(struct bpf_array, map.max_entries);
493 	if (off > 0x7fff)
494 		return -1;
495 	emit(ctx, lwu, tmp, off, ary);            /* tmp = ary->map.max_entrs*/
496 	emit(ctx, sltu, tmp, ind, tmp);           /* tmp = ind < t1          */
497 	emit(ctx, beqz, tmp, get_offset(ctx, 1)); /* PC += off(1) if tmp == 0*/
498 
499 	/* if (--TCC < 0) goto out */
500 	emit(ctx, daddiu, tcc, tcc, -1);          /* tcc-- (delay slot)      */
501 	emit(ctx, bltz, tcc, get_offset(ctx, 1)); /* PC += off(1) if tcc < 0 */
502 						  /* (next insn delay slot)  */
503 	/* prog = ary->ptrs[ind] */
504 	off = offsetof(struct bpf_array, ptrs);
505 	if (off > 0x7fff)
506 		return -1;
507 	emit(ctx, dsll, tmp, ind, 3);             /* tmp = ind << 3          */
508 	emit(ctx, daddu, tmp, tmp, ary);          /* tmp += ary              */
509 	emit(ctx, ld, tmp, off, tmp);             /* tmp = *(tmp + off)      */
510 
511 	/* if (prog == 0) goto out */
512 	emit(ctx, beqz, tmp, get_offset(ctx, 1)); /* PC += off(1) if tmp == 0*/
513 	emit(ctx, nop);                           /* Delay slot              */
514 
515 	/* func = prog->bpf_func + 8 (prologue skip offset) */
516 	off = offsetof(struct bpf_prog, bpf_func);
517 	if (off > 0x7fff)
518 		return -1;
519 	emit(ctx, ld, tmp, off, tmp);                /* tmp = *(tmp + off)   */
520 	emit(ctx, daddiu, tmp, tmp, JIT_TCALL_SKIP); /* tmp += skip (4)      */
521 
522 	/* goto func */
523 	build_epilogue(ctx, tmp);
524 	access_reg(ctx, JIT_REG_TC);
525 	return 0;
526 }
527 
528 /*
529  * Stack frame layout for a JITed program (stack grows down).
530  *
531  * Higher address  : Previous stack frame      :
532  *                 +===========================+  <--- MIPS sp before call
533  *                 | Callee-saved registers,   |
534  *                 | including RA and FP       |
535  *                 +---------------------------+  <--- eBPF FP (MIPS fp)
536  *                 | Local eBPF variables      |
537  *                 | allocated by program      |
538  *                 +---------------------------+
539  *                 | Reserved for caller-saved |
540  *                 | registers                 |
541  * Lower address   +===========================+  <--- MIPS sp
542  */
543 
544 /* Build program prologue to set up the stack and registers */
545 void build_prologue(struct jit_context *ctx)
546 {
547 	u8 fp = bpf2mips64[BPF_REG_FP];
548 	u8 tc = bpf2mips64[JIT_REG_TC];
549 	u8 zx = bpf2mips64[JIT_REG_ZX];
550 	int stack, saved, locals, reserved;
551 
552 	/*
553 	 * In the unlikely event that the TCC limit is raised to more
554 	 * than 16 bits, it is clamped to the maximum value allowed for
555 	 * the generated code (0xffff). It is better fail to compile
556 	 * instead of degrading gracefully.
557 	 */
558 	BUILD_BUG_ON(MAX_TAIL_CALL_CNT > 0xffff);
559 
560 	/*
561 	 * The first instruction initializes the tail call count register.
562 	 * On a tail call, the calling function jumps into the prologue
563 	 * after this instruction.
564 	 */
565 	emit(ctx, ori, tc, MIPS_R_ZERO, MAX_TAIL_CALL_CNT);
566 
567 	/* === Entry-point for tail calls === */
568 
569 	/*
570 	 * If the eBPF frame pointer and tail call count registers were
571 	 * accessed they must be preserved. Mark them as clobbered here
572 	 * to save and restore them on the stack as needed.
573 	 */
574 	if (ctx->accessed & BIT(BPF_REG_FP))
575 		clobber_reg(ctx, fp);
576 	if (ctx->accessed & BIT(JIT_REG_TC))
577 		clobber_reg(ctx, tc);
578 	if (ctx->accessed & BIT(JIT_REG_ZX))
579 		clobber_reg(ctx, zx);
580 
581 	/* Compute the stack space needed for callee-saved registers */
582 	saved = hweight32(ctx->clobbered & JIT_CALLEE_REGS) * sizeof(u64);
583 	saved = ALIGN(saved, MIPS_STACK_ALIGNMENT);
584 
585 	/* Stack space used by eBPF program local data */
586 	locals = ALIGN(ctx->program->aux->stack_depth, MIPS_STACK_ALIGNMENT);
587 
588 	/*
589 	 * If we are emitting function calls, reserve extra stack space for
590 	 * caller-saved registers needed by the JIT. The required space is
591 	 * computed automatically during resource usage discovery (pass 1).
592 	 */
593 	reserved = ctx->stack_used;
594 
595 	/* Allocate the stack frame */
596 	stack = ALIGN(saved + locals + reserved, MIPS_STACK_ALIGNMENT);
597 	if (stack)
598 		emit(ctx, daddiu, MIPS_R_SP, MIPS_R_SP, -stack);
599 
600 	/* Store callee-saved registers on stack */
601 	push_regs(ctx, ctx->clobbered & JIT_CALLEE_REGS, 0, stack - saved);
602 
603 	/* Initialize the eBPF frame pointer if accessed */
604 	if (ctx->accessed & BIT(BPF_REG_FP))
605 		emit(ctx, daddiu, fp, MIPS_R_SP, stack - saved);
606 
607 	/* Initialize the ePF JIT zero-extension register if accessed */
608 	if (ctx->accessed & BIT(JIT_REG_ZX)) {
609 		emit(ctx, daddiu, zx, MIPS_R_ZERO, -1);
610 		emit(ctx, dsrl32, zx, zx, 0);
611 	}
612 
613 	ctx->saved_size = saved;
614 	ctx->stack_size = stack;
615 }
616 
617 /* Build the program epilogue to restore the stack and registers */
618 void build_epilogue(struct jit_context *ctx, int dest_reg)
619 {
620 	/* Restore callee-saved registers from stack */
621 	pop_regs(ctx, ctx->clobbered & JIT_CALLEE_REGS, 0,
622 		 ctx->stack_size - ctx->saved_size);
623 
624 	/* Release the stack frame */
625 	if (ctx->stack_size)
626 		emit(ctx, daddiu, MIPS_R_SP, MIPS_R_SP, ctx->stack_size);
627 
628 	/* Jump to return address and sign-extend the 32-bit return value */
629 	emit(ctx, jr, dest_reg);
630 	emit(ctx, sll, MIPS_R_V0, MIPS_R_V0, 0); /* Delay slot */
631 }
632 
633 /* Build one eBPF instruction */
634 int build_insn(const struct bpf_insn *insn, struct jit_context *ctx)
635 {
636 	u8 dst = bpf2mips64[insn->dst_reg];
637 	u8 src = bpf2mips64[insn->src_reg];
638 	u8 res = bpf2mips64[BPF_REG_0];
639 	u8 code = insn->code;
640 	s16 off = insn->off;
641 	s32 imm = insn->imm;
642 	s32 val, rel;
643 	u8 alu, jmp;
644 
645 	switch (code) {
646 	/* ALU operations */
647 	/* dst = imm */
648 	case BPF_ALU | BPF_MOV | BPF_K:
649 		emit_mov_i(ctx, dst, imm);
650 		emit_zext_ver(ctx, dst);
651 		break;
652 	/* dst = src */
653 	case BPF_ALU | BPF_MOV | BPF_X:
654 		if (imm == 1) {
655 			/* Special mov32 for zext */
656 			emit_zext(ctx, dst);
657 		} else {
658 			emit_mov_r(ctx, dst, src);
659 			emit_zext_ver(ctx, dst);
660 		}
661 		break;
662 	/* dst = -dst */
663 	case BPF_ALU | BPF_NEG:
664 		emit_sext(ctx, dst, dst);
665 		emit_alu_i(ctx, dst, 0, BPF_NEG);
666 		emit_zext_ver(ctx, dst);
667 		break;
668 	/* dst = dst & imm */
669 	/* dst = dst | imm */
670 	/* dst = dst ^ imm */
671 	/* dst = dst << imm */
672 	case BPF_ALU | BPF_OR | BPF_K:
673 	case BPF_ALU | BPF_AND | BPF_K:
674 	case BPF_ALU | BPF_XOR | BPF_K:
675 	case BPF_ALU | BPF_LSH | BPF_K:
676 		if (!valid_alu_i(BPF_OP(code), imm)) {
677 			emit_mov_i(ctx, MIPS_R_T4, imm);
678 			emit_alu_r(ctx, dst, MIPS_R_T4, BPF_OP(code));
679 		} else if (rewrite_alu_i(BPF_OP(code), imm, &alu, &val)) {
680 			emit_alu_i(ctx, dst, val, alu);
681 		}
682 		emit_zext_ver(ctx, dst);
683 		break;
684 	/* dst = dst >> imm */
685 	/* dst = dst >> imm (arithmetic) */
686 	/* dst = dst + imm */
687 	/* dst = dst - imm */
688 	/* dst = dst * imm */
689 	/* dst = dst / imm */
690 	/* dst = dst % imm */
691 	case BPF_ALU | BPF_RSH | BPF_K:
692 	case BPF_ALU | BPF_ARSH | BPF_K:
693 	case BPF_ALU | BPF_ADD | BPF_K:
694 	case BPF_ALU | BPF_SUB | BPF_K:
695 	case BPF_ALU | BPF_MUL | BPF_K:
696 	case BPF_ALU | BPF_DIV | BPF_K:
697 	case BPF_ALU | BPF_MOD | BPF_K:
698 		if (!valid_alu_i(BPF_OP(code), imm)) {
699 			emit_sext(ctx, dst, dst);
700 			emit_mov_i(ctx, MIPS_R_T4, imm);
701 			emit_alu_r(ctx, dst, MIPS_R_T4, BPF_OP(code));
702 		} else if (rewrite_alu_i(BPF_OP(code), imm, &alu, &val)) {
703 			emit_sext(ctx, dst, dst);
704 			emit_alu_i(ctx, dst, val, alu);
705 		}
706 		emit_zext_ver(ctx, dst);
707 		break;
708 	/* dst = dst & src */
709 	/* dst = dst | src */
710 	/* dst = dst ^ src */
711 	/* dst = dst << src */
712 	case BPF_ALU | BPF_AND | BPF_X:
713 	case BPF_ALU | BPF_OR | BPF_X:
714 	case BPF_ALU | BPF_XOR | BPF_X:
715 	case BPF_ALU | BPF_LSH | BPF_X:
716 		emit_alu_r(ctx, dst, src, BPF_OP(code));
717 		emit_zext_ver(ctx, dst);
718 		break;
719 	/* dst = dst >> src */
720 	/* dst = dst >> src (arithmetic) */
721 	/* dst = dst + src */
722 	/* dst = dst - src */
723 	/* dst = dst * src */
724 	/* dst = dst / src */
725 	/* dst = dst % src */
726 	case BPF_ALU | BPF_RSH | BPF_X:
727 	case BPF_ALU | BPF_ARSH | BPF_X:
728 	case BPF_ALU | BPF_ADD | BPF_X:
729 	case BPF_ALU | BPF_SUB | BPF_X:
730 	case BPF_ALU | BPF_MUL | BPF_X:
731 	case BPF_ALU | BPF_DIV | BPF_X:
732 	case BPF_ALU | BPF_MOD | BPF_X:
733 		emit_sext(ctx, dst, dst);
734 		emit_sext(ctx, MIPS_R_T4, src);
735 		emit_alu_r(ctx, dst, MIPS_R_T4, BPF_OP(code));
736 		emit_zext_ver(ctx, dst);
737 		break;
738 	/* dst = imm (64-bit) */
739 	case BPF_ALU64 | BPF_MOV | BPF_K:
740 		emit_mov_i(ctx, dst, imm);
741 		break;
742 	/* dst = src (64-bit) */
743 	case BPF_ALU64 | BPF_MOV | BPF_X:
744 		emit_mov_r(ctx, dst, src);
745 		break;
746 	/* dst = -dst (64-bit) */
747 	case BPF_ALU64 | BPF_NEG:
748 		emit_alu_i64(ctx, dst, 0, BPF_NEG);
749 		break;
750 	/* dst = dst & imm (64-bit) */
751 	/* dst = dst | imm (64-bit) */
752 	/* dst = dst ^ imm (64-bit) */
753 	/* dst = dst << imm (64-bit) */
754 	/* dst = dst >> imm (64-bit) */
755 	/* dst = dst >> imm ((64-bit, arithmetic) */
756 	/* dst = dst + imm (64-bit) */
757 	/* dst = dst - imm (64-bit) */
758 	/* dst = dst * imm (64-bit) */
759 	/* dst = dst / imm (64-bit) */
760 	/* dst = dst % imm (64-bit) */
761 	case BPF_ALU64 | BPF_AND | BPF_K:
762 	case BPF_ALU64 | BPF_OR | BPF_K:
763 	case BPF_ALU64 | BPF_XOR | BPF_K:
764 	case BPF_ALU64 | BPF_LSH | BPF_K:
765 	case BPF_ALU64 | BPF_RSH | BPF_K:
766 	case BPF_ALU64 | BPF_ARSH | BPF_K:
767 	case BPF_ALU64 | BPF_ADD | BPF_K:
768 	case BPF_ALU64 | BPF_SUB | BPF_K:
769 	case BPF_ALU64 | BPF_MUL | BPF_K:
770 	case BPF_ALU64 | BPF_DIV | BPF_K:
771 	case BPF_ALU64 | BPF_MOD | BPF_K:
772 		if (!valid_alu_i(BPF_OP(code), imm)) {
773 			emit_mov_i(ctx, MIPS_R_T4, imm);
774 			emit_alu_r64(ctx, dst, MIPS_R_T4, BPF_OP(code));
775 		} else if (rewrite_alu_i(BPF_OP(code), imm, &alu, &val)) {
776 			emit_alu_i64(ctx, dst, val, alu);
777 		}
778 		break;
779 	/* dst = dst & src (64-bit) */
780 	/* dst = dst | src (64-bit) */
781 	/* dst = dst ^ src (64-bit) */
782 	/* dst = dst << src (64-bit) */
783 	/* dst = dst >> src (64-bit) */
784 	/* dst = dst >> src (64-bit, arithmetic) */
785 	/* dst = dst + src (64-bit) */
786 	/* dst = dst - src (64-bit) */
787 	/* dst = dst * src (64-bit) */
788 	/* dst = dst / src (64-bit) */
789 	/* dst = dst % src (64-bit) */
790 	case BPF_ALU64 | BPF_AND | BPF_X:
791 	case BPF_ALU64 | BPF_OR | BPF_X:
792 	case BPF_ALU64 | BPF_XOR | BPF_X:
793 	case BPF_ALU64 | BPF_LSH | BPF_X:
794 	case BPF_ALU64 | BPF_RSH | BPF_X:
795 	case BPF_ALU64 | BPF_ARSH | BPF_X:
796 	case BPF_ALU64 | BPF_ADD | BPF_X:
797 	case BPF_ALU64 | BPF_SUB | BPF_X:
798 	case BPF_ALU64 | BPF_MUL | BPF_X:
799 	case BPF_ALU64 | BPF_DIV | BPF_X:
800 	case BPF_ALU64 | BPF_MOD | BPF_X:
801 		emit_alu_r64(ctx, dst, src, BPF_OP(code));
802 		break;
803 	/* dst = htole(dst) */
804 	/* dst = htobe(dst) */
805 	case BPF_ALU | BPF_END | BPF_FROM_LE:
806 	case BPF_ALU | BPF_END | BPF_FROM_BE:
807 		if (BPF_SRC(code) ==
808 #ifdef __BIG_ENDIAN
809 		    BPF_FROM_LE
810 #else
811 		    BPF_FROM_BE
812 #endif
813 		    )
814 			emit_bswap_r64(ctx, dst, imm);
815 		else
816 			emit_trunc_r64(ctx, dst, imm);
817 		break;
818 	/* dst = imm64 */
819 	case BPF_LD | BPF_IMM | BPF_DW:
820 		emit_mov_i64(ctx, dst, (u32)imm | ((u64)insn[1].imm << 32));
821 		return 1;
822 	/* LDX: dst = *(size *)(src + off) */
823 	case BPF_LDX | BPF_MEM | BPF_W:
824 	case BPF_LDX | BPF_MEM | BPF_H:
825 	case BPF_LDX | BPF_MEM | BPF_B:
826 	case BPF_LDX | BPF_MEM | BPF_DW:
827 		emit_ldx(ctx, dst, src, off, BPF_SIZE(code));
828 		break;
829 	/* ST: *(size *)(dst + off) = imm */
830 	case BPF_ST | BPF_MEM | BPF_W:
831 	case BPF_ST | BPF_MEM | BPF_H:
832 	case BPF_ST | BPF_MEM | BPF_B:
833 	case BPF_ST | BPF_MEM | BPF_DW:
834 		emit_mov_i(ctx, MIPS_R_T4, imm);
835 		emit_stx(ctx, dst, MIPS_R_T4, off, BPF_SIZE(code));
836 		break;
837 	/* STX: *(size *)(dst + off) = src */
838 	case BPF_STX | BPF_MEM | BPF_W:
839 	case BPF_STX | BPF_MEM | BPF_H:
840 	case BPF_STX | BPF_MEM | BPF_B:
841 	case BPF_STX | BPF_MEM | BPF_DW:
842 		emit_stx(ctx, dst, src, off, BPF_SIZE(code));
843 		break;
844 	/* Speculation barrier */
845 	case BPF_ST | BPF_NOSPEC:
846 		break;
847 	/* Atomics */
848 	case BPF_STX | BPF_ATOMIC | BPF_W:
849 	case BPF_STX | BPF_ATOMIC | BPF_DW:
850 		switch (imm) {
851 		case BPF_ADD:
852 		case BPF_ADD | BPF_FETCH:
853 		case BPF_AND:
854 		case BPF_AND | BPF_FETCH:
855 		case BPF_OR:
856 		case BPF_OR | BPF_FETCH:
857 		case BPF_XOR:
858 		case BPF_XOR | BPF_FETCH:
859 		case BPF_XCHG:
860 			if (BPF_SIZE(code) == BPF_DW) {
861 				emit_atomic_r64(ctx, dst, src, off, imm);
862 			} else if (imm & BPF_FETCH) {
863 				u8 tmp = dst;
864 
865 				if (src == dst) { /* Don't overwrite dst */
866 					emit_mov_r(ctx, MIPS_R_T4, dst);
867 					tmp = MIPS_R_T4;
868 				}
869 				emit_sext(ctx, src, src);
870 				emit_atomic_r(ctx, tmp, src, off, imm);
871 				emit_zext_ver(ctx, src);
872 			} else { /* 32-bit, no fetch */
873 				emit_sext(ctx, MIPS_R_T4, src);
874 				emit_atomic_r(ctx, dst, MIPS_R_T4, off, imm);
875 			}
876 			break;
877 		case BPF_CMPXCHG:
878 			if (BPF_SIZE(code) == BPF_DW) {
879 				emit_cmpxchg_r64(ctx, dst, src, off);
880 			} else {
881 				u8 tmp = res;
882 
883 				if (res == dst)   /* Don't overwrite dst */
884 					tmp = MIPS_R_T4;
885 				emit_sext(ctx, tmp, res);
886 				emit_sext(ctx, MIPS_R_T5, src);
887 				emit_cmpxchg_r(ctx, dst, MIPS_R_T5, tmp, off);
888 				if (res == dst)   /* Restore result */
889 					emit_mov_r(ctx, res, MIPS_R_T4);
890 				/* Result zext inserted by verifier */
891 			}
892 			break;
893 		default:
894 			goto notyet;
895 		}
896 		break;
897 	/* PC += off if dst == src */
898 	/* PC += off if dst != src */
899 	/* PC += off if dst & src */
900 	/* PC += off if dst > src */
901 	/* PC += off if dst >= src */
902 	/* PC += off if dst < src */
903 	/* PC += off if dst <= src */
904 	/* PC += off if dst > src (signed) */
905 	/* PC += off if dst >= src (signed) */
906 	/* PC += off if dst < src (signed) */
907 	/* PC += off if dst <= src (signed) */
908 	case BPF_JMP32 | BPF_JEQ | BPF_X:
909 	case BPF_JMP32 | BPF_JNE | BPF_X:
910 	case BPF_JMP32 | BPF_JSET | BPF_X:
911 	case BPF_JMP32 | BPF_JGT | BPF_X:
912 	case BPF_JMP32 | BPF_JGE | BPF_X:
913 	case BPF_JMP32 | BPF_JLT | BPF_X:
914 	case BPF_JMP32 | BPF_JLE | BPF_X:
915 	case BPF_JMP32 | BPF_JSGT | BPF_X:
916 	case BPF_JMP32 | BPF_JSGE | BPF_X:
917 	case BPF_JMP32 | BPF_JSLT | BPF_X:
918 	case BPF_JMP32 | BPF_JSLE | BPF_X:
919 		if (off == 0)
920 			break;
921 		setup_jmp_r(ctx, dst == src, BPF_OP(code), off, &jmp, &rel);
922 		emit_sext(ctx, MIPS_R_T4, dst); /* Sign-extended dst */
923 		emit_sext(ctx, MIPS_R_T5, src); /* Sign-extended src */
924 		emit_jmp_r(ctx, MIPS_R_T4, MIPS_R_T5, rel, jmp);
925 		if (finish_jmp(ctx, jmp, off) < 0)
926 			goto toofar;
927 		break;
928 	/* PC += off if dst == imm */
929 	/* PC += off if dst != imm */
930 	/* PC += off if dst & imm */
931 	/* PC += off if dst > imm */
932 	/* PC += off if dst >= imm */
933 	/* PC += off if dst < imm */
934 	/* PC += off if dst <= imm */
935 	/* PC += off if dst > imm (signed) */
936 	/* PC += off if dst >= imm (signed) */
937 	/* PC += off if dst < imm (signed) */
938 	/* PC += off if dst <= imm (signed) */
939 	case BPF_JMP32 | BPF_JEQ | BPF_K:
940 	case BPF_JMP32 | BPF_JNE | BPF_K:
941 	case BPF_JMP32 | BPF_JSET | BPF_K:
942 	case BPF_JMP32 | BPF_JGT | BPF_K:
943 	case BPF_JMP32 | BPF_JGE | BPF_K:
944 	case BPF_JMP32 | BPF_JLT | BPF_K:
945 	case BPF_JMP32 | BPF_JLE | BPF_K:
946 	case BPF_JMP32 | BPF_JSGT | BPF_K:
947 	case BPF_JMP32 | BPF_JSGE | BPF_K:
948 	case BPF_JMP32 | BPF_JSLT | BPF_K:
949 	case BPF_JMP32 | BPF_JSLE | BPF_K:
950 		if (off == 0)
951 			break;
952 		setup_jmp_i(ctx, imm, 32, BPF_OP(code), off, &jmp, &rel);
953 		emit_sext(ctx, MIPS_R_T4, dst); /* Sign-extended dst */
954 		if (valid_jmp_i(jmp, imm)) {
955 			emit_jmp_i(ctx, MIPS_R_T4, imm, rel, jmp);
956 		} else {
957 			/* Move large immediate to register, sign-extended */
958 			emit_mov_i(ctx, MIPS_R_T5, imm);
959 			emit_jmp_r(ctx, MIPS_R_T4, MIPS_R_T5, rel, jmp);
960 		}
961 		if (finish_jmp(ctx, jmp, off) < 0)
962 			goto toofar;
963 		break;
964 	/* PC += off if dst == src */
965 	/* PC += off if dst != src */
966 	/* PC += off if dst & src */
967 	/* PC += off if dst > src */
968 	/* PC += off if dst >= src */
969 	/* PC += off if dst < src */
970 	/* PC += off if dst <= src */
971 	/* PC += off if dst > src (signed) */
972 	/* PC += off if dst >= src (signed) */
973 	/* PC += off if dst < src (signed) */
974 	/* PC += off if dst <= src (signed) */
975 	case BPF_JMP | BPF_JEQ | BPF_X:
976 	case BPF_JMP | BPF_JNE | BPF_X:
977 	case BPF_JMP | BPF_JSET | BPF_X:
978 	case BPF_JMP | BPF_JGT | BPF_X:
979 	case BPF_JMP | BPF_JGE | BPF_X:
980 	case BPF_JMP | BPF_JLT | BPF_X:
981 	case BPF_JMP | BPF_JLE | BPF_X:
982 	case BPF_JMP | BPF_JSGT | BPF_X:
983 	case BPF_JMP | BPF_JSGE | BPF_X:
984 	case BPF_JMP | BPF_JSLT | BPF_X:
985 	case BPF_JMP | BPF_JSLE | BPF_X:
986 		if (off == 0)
987 			break;
988 		setup_jmp_r(ctx, dst == src, BPF_OP(code), off, &jmp, &rel);
989 		emit_jmp_r(ctx, dst, src, rel, jmp);
990 		if (finish_jmp(ctx, jmp, off) < 0)
991 			goto toofar;
992 		break;
993 	/* PC += off if dst == imm */
994 	/* PC += off if dst != imm */
995 	/* PC += off if dst & imm */
996 	/* PC += off if dst > imm */
997 	/* PC += off if dst >= imm */
998 	/* PC += off if dst < imm */
999 	/* PC += off if dst <= imm */
1000 	/* PC += off if dst > imm (signed) */
1001 	/* PC += off if dst >= imm (signed) */
1002 	/* PC += off if dst < imm (signed) */
1003 	/* PC += off if dst <= imm (signed) */
1004 	case BPF_JMP | BPF_JEQ | BPF_K:
1005 	case BPF_JMP | BPF_JNE | BPF_K:
1006 	case BPF_JMP | BPF_JSET | BPF_K:
1007 	case BPF_JMP | BPF_JGT | BPF_K:
1008 	case BPF_JMP | BPF_JGE | BPF_K:
1009 	case BPF_JMP | BPF_JLT | BPF_K:
1010 	case BPF_JMP | BPF_JLE | BPF_K:
1011 	case BPF_JMP | BPF_JSGT | BPF_K:
1012 	case BPF_JMP | BPF_JSGE | BPF_K:
1013 	case BPF_JMP | BPF_JSLT | BPF_K:
1014 	case BPF_JMP | BPF_JSLE | BPF_K:
1015 		if (off == 0)
1016 			break;
1017 		setup_jmp_i(ctx, imm, 64, BPF_OP(code), off, &jmp, &rel);
1018 		if (valid_jmp_i(jmp, imm)) {
1019 			emit_jmp_i(ctx, dst, imm, rel, jmp);
1020 		} else {
1021 			/* Move large immediate to register */
1022 			emit_mov_i(ctx, MIPS_R_T4, imm);
1023 			emit_jmp_r(ctx, dst, MIPS_R_T4, rel, jmp);
1024 		}
1025 		if (finish_jmp(ctx, jmp, off) < 0)
1026 			goto toofar;
1027 		break;
1028 	/* PC += off */
1029 	case BPF_JMP | BPF_JA:
1030 		if (off == 0)
1031 			break;
1032 		if (emit_ja(ctx, off) < 0)
1033 			goto toofar;
1034 		break;
1035 	/* Tail call */
1036 	case BPF_JMP | BPF_TAIL_CALL:
1037 		if (emit_tail_call(ctx) < 0)
1038 			goto invalid;
1039 		break;
1040 	/* Function call */
1041 	case BPF_JMP | BPF_CALL:
1042 		if (emit_call(ctx, insn) < 0)
1043 			goto invalid;
1044 		break;
1045 	/* Function return */
1046 	case BPF_JMP | BPF_EXIT:
1047 		/*
1048 		 * Optimization: when last instruction is EXIT
1049 		 * simply continue to epilogue.
1050 		 */
1051 		if (ctx->bpf_index == ctx->program->len - 1)
1052 			break;
1053 		if (emit_exit(ctx) < 0)
1054 			goto toofar;
1055 		break;
1056 
1057 	default:
1058 invalid:
1059 		pr_err_once("unknown opcode %02x\n", code);
1060 		return -EINVAL;
1061 notyet:
1062 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
1063 		return -EFAULT;
1064 toofar:
1065 		pr_info_once("*** TOO FAR: jump at %u opcode %02x ***\n",
1066 			     ctx->bpf_index, code);
1067 		return -E2BIG;
1068 	}
1069 	return 0;
1070 }
1071