xref: /linux/arch/mips/net/bpf_jit_comp.c (revision ea8a163e02d6925773129e2dd86e419e491b791d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Just-In-Time compiler for eBPF bytecode on MIPS.
4  * Implementation of JIT functions common to 32-bit and 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 /*
16  * Code overview
17  * =============
18  *
19  * - bpf_jit_comp.h
20  *   Common definitions and utilities.
21  *
22  * - bpf_jit_comp.c
23  *   Implementation of JIT top-level logic and exported JIT API functions.
24  *   Implementation of internal operations shared by 32-bit and 64-bit code.
25  *   JMP and ALU JIT control code, register control code, shared ALU and
26  *   JMP/JMP32 JIT operations.
27  *
28  * - bpf_jit_comp32.c
29  *   Implementation of functions to JIT prologue, epilogue and a single eBPF
30  *   instruction for 32-bit MIPS CPUs. The functions use shared operations
31  *   where possible, and implement the rest for 32-bit MIPS such as ALU64
32  *   operations.
33  *
34  * - bpf_jit_comp64.c
35  *   Ditto, for 64-bit MIPS CPUs.
36  *
37  * Zero and sign extension
38  * ========================
39  * 32-bit MIPS instructions on 64-bit MIPS registers use sign extension,
40  * but the eBPF instruction set mandates zero extension. We let the verifier
41  * insert explicit zero-extensions after 32-bit ALU operations, both for
42  * 32-bit and 64-bit MIPS JITs. Conditional JMP32 operations on 64-bit MIPs
43  * are JITed with sign extensions inserted when so expected.
44  *
45  * ALU operations
46  * ==============
47  * ALU operations on 32/64-bit MIPS and ALU64 operations on 64-bit MIPS are
48  * JITed in the following steps. ALU64 operations on 32-bit MIPS are more
49  * complicated and therefore only processed by special implementations in
50  * step (3).
51  *
52  * 1) valid_alu_i:
53  *    Determine if an immediate operation can be emitted as such, or if
54  *    we must fall back to the register version.
55  *
56  * 2) rewrite_alu_i:
57  *    Convert BPF operation and immediate value to a canonical form for
58  *    JITing. In some degenerate cases this form may be a no-op.
59  *
60  * 3) emit_alu_{i,i64,r,64}:
61  *    Emit instructions for an ALU or ALU64 immediate or register operation.
62  *
63  * JMP operations
64  * ==============
65  * JMP and JMP32 operations require an JIT instruction offset table for
66  * translating the jump offset. This table is computed by dry-running the
67  * JIT without actually emitting anything. However, the computed PC-relative
68  * offset may overflow the 18-bit offset field width of the native MIPS
69  * branch instruction. In such cases, the long jump is converted into the
70  * following sequence.
71  *
72  *    <branch> !<cond> +2    Inverted PC-relative branch
73  *    nop                    Delay slot
74  *    j <offset>             Unconditional absolute long jump
75  *    nop                    Delay slot
76  *
77  * Since this converted sequence alters the offset table, all offsets must
78  * be re-calculated. This may in turn trigger new branch conversions, so
79  * the process is repeated until no further changes are made. Normally it
80  * completes in 1-2 iterations. If JIT_MAX_ITERATIONS should reached, we
81  * fall back to converting every remaining jump operation. The branch
82  * conversion is independent of how the JMP or JMP32 condition is JITed.
83  *
84  * JMP32 and JMP operations are JITed as follows.
85  *
86  * 1) setup_jmp_{i,r}:
87  *    Convert jump conditional and offset into a form that can be JITed.
88  *    This form may be a no-op, a canonical form, or an inverted PC-relative
89  *    jump if branch conversion is necessary.
90  *
91  * 2) valid_jmp_i:
92  *    Determine if an immediate operations can be emitted as such, or if
93  *    we must fall back to the register version. Applies to JMP32 for 32-bit
94  *    MIPS, and both JMP and JMP32 for 64-bit MIPS.
95  *
96  * 3) emit_jmp_{i,i64,r,r64}:
97  *    Emit instructions for an JMP or JMP32 immediate or register operation.
98  *
99  * 4) finish_jmp_{i,r}:
100  *    Emit any instructions needed to finish the jump. This includes a nop
101  *    for the delay slot if a branch was emitted, and a long absolute jump
102  *    if the branch was converted.
103  */
104 
105 #include <linux/limits.h>
106 #include <linux/bitops.h>
107 #include <linux/errno.h>
108 #include <linux/filter.h>
109 #include <linux/bpf.h>
110 #include <linux/slab.h>
111 #include <asm/bitops.h>
112 #include <asm/cacheflush.h>
113 #include <asm/cpu-features.h>
114 #include <asm/isa-rev.h>
115 #include <asm/uasm.h>
116 
117 #include "bpf_jit_comp.h"
118 
119 /* Convenience macros for descriptor access */
120 #define CONVERTED(desc)	((desc) & JIT_DESC_CONVERT)
121 #define INDEX(desc)	((desc) & ~JIT_DESC_CONVERT)
122 
123 /*
124  * Push registers on the stack, starting at a given depth from the stack
125  * pointer and increasing. The next depth to be written is returned.
126  */
127 int push_regs(struct jit_context *ctx, u32 mask, u32 excl, int depth)
128 {
129 	int reg;
130 
131 	for (reg = 0; reg < BITS_PER_BYTE * sizeof(mask); reg++)
132 		if (mask & BIT(reg)) {
133 			if ((excl & BIT(reg)) == 0) {
134 				if (sizeof(long) == 4)
135 					emit(ctx, sw, reg, depth, MIPS_R_SP);
136 				else /* sizeof(long) == 8 */
137 					emit(ctx, sd, reg, depth, MIPS_R_SP);
138 			}
139 			depth += sizeof(long);
140 		}
141 
142 	ctx->stack_used = max((int)ctx->stack_used, depth);
143 	return depth;
144 }
145 
146 /*
147  * Pop registers from the stack, starting at a given depth from the stack
148  * pointer and increasing. The next depth to be read is returned.
149  */
150 int pop_regs(struct jit_context *ctx, u32 mask, u32 excl, int depth)
151 {
152 	int reg;
153 
154 	for (reg = 0; reg < BITS_PER_BYTE * sizeof(mask); reg++)
155 		if (mask & BIT(reg)) {
156 			if ((excl & BIT(reg)) == 0) {
157 				if (sizeof(long) == 4)
158 					emit(ctx, lw, reg, depth, MIPS_R_SP);
159 				else /* sizeof(long) == 8 */
160 					emit(ctx, ld, reg, depth, MIPS_R_SP);
161 			}
162 			depth += sizeof(long);
163 		}
164 
165 	return depth;
166 }
167 
168 /* Compute the 28-bit jump target address from a BPF program location */
169 int get_target(struct jit_context *ctx, u32 loc)
170 {
171 	u32 index = INDEX(ctx->descriptors[loc]);
172 	unsigned long pc = (unsigned long)&ctx->target[ctx->jit_index];
173 	unsigned long addr = (unsigned long)&ctx->target[index];
174 
175 	if (!ctx->target)
176 		return 0;
177 
178 	if ((addr ^ pc) & ~MIPS_JMP_MASK)
179 		return -1;
180 
181 	return addr & MIPS_JMP_MASK;
182 }
183 
184 /* Compute the PC-relative offset to relative BPF program offset */
185 int get_offset(const struct jit_context *ctx, int off)
186 {
187 	return (INDEX(ctx->descriptors[ctx->bpf_index + off]) -
188 		ctx->jit_index - 1) * sizeof(u32);
189 }
190 
191 /* dst = imm (register width) */
192 void emit_mov_i(struct jit_context *ctx, u8 dst, s32 imm)
193 {
194 	if (imm >= -0x8000 && imm <= 0x7fff) {
195 		emit(ctx, addiu, dst, MIPS_R_ZERO, imm);
196 	} else {
197 		emit(ctx, lui, dst, (s16)((u32)imm >> 16));
198 		emit(ctx, ori, dst, dst, (u16)(imm & 0xffff));
199 	}
200 	clobber_reg(ctx, dst);
201 }
202 
203 /* dst = src (register width) */
204 void emit_mov_r(struct jit_context *ctx, u8 dst, u8 src)
205 {
206 	emit(ctx, ori, dst, src, 0);
207 	clobber_reg(ctx, dst);
208 }
209 
210 /* Validate ALU immediate range */
211 bool valid_alu_i(u8 op, s32 imm)
212 {
213 	switch (BPF_OP(op)) {
214 	case BPF_NEG:
215 	case BPF_LSH:
216 	case BPF_RSH:
217 	case BPF_ARSH:
218 		/* All legal eBPF values are valid */
219 		return true;
220 	case BPF_ADD:
221 		/* imm must be 16 bits */
222 		return imm >= -0x8000 && imm <= 0x7fff;
223 	case BPF_SUB:
224 		/* -imm must be 16 bits */
225 		return imm >= -0x7fff && imm <= 0x8000;
226 	case BPF_AND:
227 	case BPF_OR:
228 	case BPF_XOR:
229 		/* imm must be 16 bits unsigned */
230 		return imm >= 0 && imm <= 0xffff;
231 	case BPF_MUL:
232 		/* imm must be zero or a positive power of two */
233 		return imm == 0 || (imm > 0 && is_power_of_2(imm));
234 	case BPF_DIV:
235 	case BPF_MOD:
236 		/* imm must be an 17-bit power of two */
237 		return (u32)imm <= 0x10000 && is_power_of_2((u32)imm);
238 	}
239 	return false;
240 }
241 
242 /* Rewrite ALU immediate operation */
243 bool rewrite_alu_i(u8 op, s32 imm, u8 *alu, s32 *val)
244 {
245 	bool act = true;
246 
247 	switch (BPF_OP(op)) {
248 	case BPF_LSH:
249 	case BPF_RSH:
250 	case BPF_ARSH:
251 	case BPF_ADD:
252 	case BPF_SUB:
253 	case BPF_OR:
254 	case BPF_XOR:
255 		/* imm == 0 is a no-op */
256 		act = imm != 0;
257 		break;
258 	case BPF_MUL:
259 		if (imm == 1) {
260 			/* dst * 1 is a no-op */
261 			act = false;
262 		} else if (imm == 0) {
263 			/* dst * 0 is dst & 0 */
264 			op = BPF_AND;
265 		} else {
266 			/* dst * (1 << n) is dst << n */
267 			op = BPF_LSH;
268 			imm = ilog2(abs(imm));
269 		}
270 		break;
271 	case BPF_DIV:
272 		if (imm == 1) {
273 			/* dst / 1 is a no-op */
274 			act = false;
275 		} else {
276 			/* dst / (1 << n) is dst >> n */
277 			op = BPF_RSH;
278 			imm = ilog2(imm);
279 		}
280 		break;
281 	case BPF_MOD:
282 		/* dst % (1 << n) is dst & ((1 << n) - 1) */
283 		op = BPF_AND;
284 		imm--;
285 		break;
286 	}
287 
288 	*alu = op;
289 	*val = imm;
290 	return act;
291 }
292 
293 /* ALU immediate operation (32-bit) */
294 void emit_alu_i(struct jit_context *ctx, u8 dst, s32 imm, u8 op)
295 {
296 	switch (BPF_OP(op)) {
297 	/* dst = -dst */
298 	case BPF_NEG:
299 		emit(ctx, subu, dst, MIPS_R_ZERO, dst);
300 		break;
301 	/* dst = dst & imm */
302 	case BPF_AND:
303 		emit(ctx, andi, dst, dst, (u16)imm);
304 		break;
305 	/* dst = dst | imm */
306 	case BPF_OR:
307 		emit(ctx, ori, dst, dst, (u16)imm);
308 		break;
309 	/* dst = dst ^ imm */
310 	case BPF_XOR:
311 		emit(ctx, xori, dst, dst, (u16)imm);
312 		break;
313 	/* dst = dst << imm */
314 	case BPF_LSH:
315 		emit(ctx, sll, dst, dst, imm);
316 		break;
317 	/* dst = dst >> imm */
318 	case BPF_RSH:
319 		emit(ctx, srl, dst, dst, imm);
320 		break;
321 	/* dst = dst >> imm (arithmetic) */
322 	case BPF_ARSH:
323 		emit(ctx, sra, dst, dst, imm);
324 		break;
325 	/* dst = dst + imm */
326 	case BPF_ADD:
327 		emit(ctx, addiu, dst, dst, imm);
328 		break;
329 	/* dst = dst - imm */
330 	case BPF_SUB:
331 		emit(ctx, addiu, dst, dst, -imm);
332 		break;
333 	}
334 	clobber_reg(ctx, dst);
335 }
336 
337 /* ALU register operation (32-bit) */
338 void emit_alu_r(struct jit_context *ctx, u8 dst, u8 src, u8 op)
339 {
340 	switch (BPF_OP(op)) {
341 	/* dst = dst & src */
342 	case BPF_AND:
343 		emit(ctx, and, dst, dst, src);
344 		break;
345 	/* dst = dst | src */
346 	case BPF_OR:
347 		emit(ctx, or, dst, dst, src);
348 		break;
349 	/* dst = dst ^ src */
350 	case BPF_XOR:
351 		emit(ctx, xor, dst, dst, src);
352 		break;
353 	/* dst = dst << src */
354 	case BPF_LSH:
355 		emit(ctx, sllv, dst, dst, src);
356 		break;
357 	/* dst = dst >> src */
358 	case BPF_RSH:
359 		emit(ctx, srlv, dst, dst, src);
360 		break;
361 	/* dst = dst >> src (arithmetic) */
362 	case BPF_ARSH:
363 		emit(ctx, srav, dst, dst, src);
364 		break;
365 	/* dst = dst + src */
366 	case BPF_ADD:
367 		emit(ctx, addu, dst, dst, src);
368 		break;
369 	/* dst = dst - src */
370 	case BPF_SUB:
371 		emit(ctx, subu, dst, dst, src);
372 		break;
373 	/* dst = dst * src */
374 	case BPF_MUL:
375 		if (cpu_has_mips32r1 || cpu_has_mips32r6) {
376 			emit(ctx, mul, dst, dst, src);
377 		} else {
378 			emit(ctx, multu, dst, src);
379 			emit(ctx, mflo, dst);
380 		}
381 		break;
382 	/* dst = dst / src */
383 	case BPF_DIV:
384 		if (cpu_has_mips32r6) {
385 			emit(ctx, divu_r6, dst, dst, src);
386 		} else {
387 			emit(ctx, divu, dst, src);
388 			emit(ctx, mflo, dst);
389 		}
390 		break;
391 	/* dst = dst % src */
392 	case BPF_MOD:
393 		if (cpu_has_mips32r6) {
394 			emit(ctx, modu, dst, dst, src);
395 		} else {
396 			emit(ctx, divu, dst, src);
397 			emit(ctx, mfhi, dst);
398 		}
399 		break;
400 	}
401 	clobber_reg(ctx, dst);
402 }
403 
404 /* Atomic read-modify-write (32-bit) */
405 void emit_atomic_r(struct jit_context *ctx, u8 dst, u8 src, s16 off, u8 code)
406 {
407 	LLSC_sync(ctx);
408 	emit(ctx, ll, MIPS_R_T9, off, dst);
409 	switch (code) {
410 	case BPF_ADD:
411 	case BPF_ADD | BPF_FETCH:
412 		emit(ctx, addu, MIPS_R_T8, MIPS_R_T9, src);
413 		break;
414 	case BPF_AND:
415 	case BPF_AND | BPF_FETCH:
416 		emit(ctx, and, MIPS_R_T8, MIPS_R_T9, src);
417 		break;
418 	case BPF_OR:
419 	case BPF_OR | BPF_FETCH:
420 		emit(ctx, or, MIPS_R_T8, MIPS_R_T9, src);
421 		break;
422 	case BPF_XOR:
423 	case BPF_XOR | BPF_FETCH:
424 		emit(ctx, xor, MIPS_R_T8, MIPS_R_T9, src);
425 		break;
426 	case BPF_XCHG:
427 		emit(ctx, move, MIPS_R_T8, src);
428 		break;
429 	}
430 	emit(ctx, sc, MIPS_R_T8, off, dst);
431 	emit(ctx, LLSC_beqz, MIPS_R_T8, -16 - LLSC_offset);
432 	emit(ctx, nop); /* Delay slot */
433 
434 	if (code & BPF_FETCH) {
435 		emit(ctx, move, src, MIPS_R_T9);
436 		clobber_reg(ctx, src);
437 	}
438 }
439 
440 /* Atomic compare-and-exchange (32-bit) */
441 void emit_cmpxchg_r(struct jit_context *ctx, u8 dst, u8 src, u8 res, s16 off)
442 {
443 	LLSC_sync(ctx);
444 	emit(ctx, ll, MIPS_R_T9, off, dst);
445 	emit(ctx, bne, MIPS_R_T9, res, 12);
446 	emit(ctx, move, MIPS_R_T8, src);     /* Delay slot */
447 	emit(ctx, sc, MIPS_R_T8, off, dst);
448 	emit(ctx, LLSC_beqz, MIPS_R_T8, -20 - LLSC_offset);
449 	emit(ctx, move, res, MIPS_R_T9);     /* Delay slot */
450 	clobber_reg(ctx, res);
451 }
452 
453 /* Swap bytes and truncate a register word or half word */
454 void emit_bswap_r(struct jit_context *ctx, u8 dst, u32 width)
455 {
456 	u8 tmp = MIPS_R_T8;
457 	u8 msk = MIPS_R_T9;
458 
459 	switch (width) {
460 	/* Swap bytes in a word */
461 	case 32:
462 		if (cpu_has_mips32r2 || cpu_has_mips32r6) {
463 			emit(ctx, wsbh, dst, dst);
464 			emit(ctx, rotr, dst, dst, 16);
465 		} else {
466 			emit(ctx, sll, tmp, dst, 16);    /* tmp  = dst << 16 */
467 			emit(ctx, srl, dst, dst, 16);    /* dst = dst >> 16  */
468 			emit(ctx, or, dst, dst, tmp);    /* dst = dst | tmp  */
469 
470 			emit(ctx, lui, msk, 0xff);       /* msk = 0x00ff0000 */
471 			emit(ctx, ori, msk, msk, 0xff);  /* msk = msk | 0xff */
472 
473 			emit(ctx, and, tmp, dst, msk);   /* tmp = dst & msk  */
474 			emit(ctx, sll, tmp, tmp, 8);     /* tmp = tmp << 8   */
475 			emit(ctx, srl, dst, dst, 8);     /* dst = dst >> 8   */
476 			emit(ctx, and, dst, dst, msk);   /* dst = dst & msk  */
477 			emit(ctx, or, dst, dst, tmp);    /* reg = dst | tmp  */
478 		}
479 		break;
480 	/* Swap bytes in a half word */
481 	case 16:
482 		if (cpu_has_mips32r2 || cpu_has_mips32r6) {
483 			emit(ctx, wsbh, dst, dst);
484 			emit(ctx, andi, dst, dst, 0xffff);
485 		} else {
486 			emit(ctx, andi, tmp, dst, 0xff00); /* t = d & 0xff00 */
487 			emit(ctx, srl, tmp, tmp, 8);       /* t = t >> 8     */
488 			emit(ctx, andi, dst, dst, 0x00ff); /* d = d & 0x00ff */
489 			emit(ctx, sll, dst, dst, 8);       /* d = d << 8     */
490 			emit(ctx, or,  dst, dst, tmp);     /* d = d | t      */
491 		}
492 		break;
493 	}
494 	clobber_reg(ctx, dst);
495 }
496 
497 /* Validate jump immediate range */
498 bool valid_jmp_i(u8 op, s32 imm)
499 {
500 	switch (op) {
501 	case JIT_JNOP:
502 		/* Immediate value not used */
503 		return true;
504 	case BPF_JEQ:
505 	case BPF_JNE:
506 		/* No immediate operation */
507 		return false;
508 	case BPF_JSET:
509 	case JIT_JNSET:
510 		/* imm must be 16 bits unsigned */
511 		return imm >= 0 && imm <= 0xffff;
512 	case BPF_JGE:
513 	case BPF_JLT:
514 	case BPF_JSGE:
515 	case BPF_JSLT:
516 		/* imm must be 16 bits */
517 		return imm >= -0x8000 && imm <= 0x7fff;
518 	case BPF_JGT:
519 	case BPF_JLE:
520 	case BPF_JSGT:
521 	case BPF_JSLE:
522 		/* imm + 1 must be 16 bits */
523 		return imm >= -0x8001 && imm <= 0x7ffe;
524 	}
525 	return false;
526 }
527 
528 /* Invert a conditional jump operation */
529 static u8 invert_jmp(u8 op)
530 {
531 	switch (op) {
532 	case BPF_JA: return JIT_JNOP;
533 	case BPF_JEQ: return BPF_JNE;
534 	case BPF_JNE: return BPF_JEQ;
535 	case BPF_JSET: return JIT_JNSET;
536 	case BPF_JGT: return BPF_JLE;
537 	case BPF_JGE: return BPF_JLT;
538 	case BPF_JLT: return BPF_JGE;
539 	case BPF_JLE: return BPF_JGT;
540 	case BPF_JSGT: return BPF_JSLE;
541 	case BPF_JSGE: return BPF_JSLT;
542 	case BPF_JSLT: return BPF_JSGE;
543 	case BPF_JSLE: return BPF_JSGT;
544 	}
545 	return 0;
546 }
547 
548 /* Prepare a PC-relative jump operation */
549 static void setup_jmp(struct jit_context *ctx, u8 bpf_op,
550 		      s16 bpf_off, u8 *jit_op, s32 *jit_off)
551 {
552 	u32 *descp = &ctx->descriptors[ctx->bpf_index];
553 	int op = bpf_op;
554 	int offset = 0;
555 
556 	/* Do not compute offsets on the first pass */
557 	if (INDEX(*descp) == 0)
558 		goto done;
559 
560 	/* Skip jumps never taken */
561 	if (bpf_op == JIT_JNOP)
562 		goto done;
563 
564 	/* Convert jumps always taken */
565 	if (bpf_op == BPF_JA)
566 		*descp |= JIT_DESC_CONVERT;
567 
568 	/*
569 	 * Current ctx->jit_index points to the start of the branch preamble.
570 	 * Since the preamble differs among different branch conditionals,
571 	 * the current index cannot be used to compute the branch offset.
572 	 * Instead, we use the offset table value for the next instruction,
573 	 * which gives the index immediately after the branch delay slot.
574 	 */
575 	if (!CONVERTED(*descp)) {
576 		int target = ctx->bpf_index + bpf_off + 1;
577 		int origin = ctx->bpf_index + 1;
578 
579 		offset = (INDEX(ctx->descriptors[target]) -
580 			  INDEX(ctx->descriptors[origin]) + 1) * sizeof(u32);
581 	}
582 
583 	/*
584 	 * The PC-relative branch offset field on MIPS is 18 bits signed,
585 	 * so if the computed offset is larger than this we generate a an
586 	 * absolute jump that we skip with an inverted conditional branch.
587 	 */
588 	if (CONVERTED(*descp) || offset < -0x20000 || offset > 0x1ffff) {
589 		offset = 3 * sizeof(u32);
590 		op = invert_jmp(bpf_op);
591 		ctx->changes += !CONVERTED(*descp);
592 		*descp |= JIT_DESC_CONVERT;
593 	}
594 
595 done:
596 	*jit_off = offset;
597 	*jit_op = op;
598 }
599 
600 /* Prepare a PC-relative jump operation with immediate conditional */
601 void setup_jmp_i(struct jit_context *ctx, s32 imm, u8 width,
602 		 u8 bpf_op, s16 bpf_off, u8 *jit_op, s32 *jit_off)
603 {
604 	bool always = false;
605 	bool never = false;
606 
607 	switch (bpf_op) {
608 	case BPF_JEQ:
609 	case BPF_JNE:
610 		break;
611 	case BPF_JSET:
612 	case BPF_JLT:
613 		never = imm == 0;
614 		break;
615 	case BPF_JGE:
616 		always = imm == 0;
617 		break;
618 	case BPF_JGT:
619 		never = (u32)imm == U32_MAX;
620 		break;
621 	case BPF_JLE:
622 		always = (u32)imm == U32_MAX;
623 		break;
624 	case BPF_JSGT:
625 		never = imm == S32_MAX && width == 32;
626 		break;
627 	case BPF_JSGE:
628 		always = imm == S32_MIN && width == 32;
629 		break;
630 	case BPF_JSLT:
631 		never = imm == S32_MIN && width == 32;
632 		break;
633 	case BPF_JSLE:
634 		always = imm == S32_MAX && width == 32;
635 		break;
636 	}
637 
638 	if (never)
639 		bpf_op = JIT_JNOP;
640 	if (always)
641 		bpf_op = BPF_JA;
642 	setup_jmp(ctx, bpf_op, bpf_off, jit_op, jit_off);
643 }
644 
645 /* Prepare a PC-relative jump operation with register conditional */
646 void setup_jmp_r(struct jit_context *ctx, bool same_reg,
647 		 u8 bpf_op, s16 bpf_off, u8 *jit_op, s32 *jit_off)
648 {
649 	switch (bpf_op) {
650 	case BPF_JSET:
651 		break;
652 	case BPF_JEQ:
653 	case BPF_JGE:
654 	case BPF_JLE:
655 	case BPF_JSGE:
656 	case BPF_JSLE:
657 		if (same_reg)
658 			bpf_op = BPF_JA;
659 		break;
660 	case BPF_JNE:
661 	case BPF_JLT:
662 	case BPF_JGT:
663 	case BPF_JSGT:
664 	case BPF_JSLT:
665 		if (same_reg)
666 			bpf_op = JIT_JNOP;
667 		break;
668 	}
669 	setup_jmp(ctx, bpf_op, bpf_off, jit_op, jit_off);
670 }
671 
672 /* Finish a PC-relative jump operation */
673 int finish_jmp(struct jit_context *ctx, u8 jit_op, s16 bpf_off)
674 {
675 	/* Emit conditional branch delay slot */
676 	if (jit_op != JIT_JNOP)
677 		emit(ctx, nop);
678 	/*
679 	 * Emit an absolute long jump with delay slot,
680 	 * if the PC-relative branch was converted.
681 	 */
682 	if (CONVERTED(ctx->descriptors[ctx->bpf_index])) {
683 		int target = get_target(ctx, ctx->bpf_index + bpf_off + 1);
684 
685 		if (target < 0)
686 			return -1;
687 		emit(ctx, j, target);
688 		emit(ctx, nop);
689 	}
690 	return 0;
691 }
692 
693 /* Jump immediate (32-bit) */
694 void emit_jmp_i(struct jit_context *ctx, u8 dst, s32 imm, s32 off, u8 op)
695 {
696 	switch (op) {
697 	/* No-op, used internally for branch optimization */
698 	case JIT_JNOP:
699 		break;
700 	/* PC += off if dst & imm */
701 	case BPF_JSET:
702 		emit(ctx, andi, MIPS_R_T9, dst, (u16)imm);
703 		emit(ctx, bnez, MIPS_R_T9, off);
704 		break;
705 	/* PC += off if (dst & imm) == 0 (not in BPF, used for long jumps) */
706 	case JIT_JNSET:
707 		emit(ctx, andi, MIPS_R_T9, dst, (u16)imm);
708 		emit(ctx, beqz, MIPS_R_T9, off);
709 		break;
710 	/* PC += off if dst > imm */
711 	case BPF_JGT:
712 		emit(ctx, sltiu, MIPS_R_T9, dst, imm + 1);
713 		emit(ctx, beqz, MIPS_R_T9, off);
714 		break;
715 	/* PC += off if dst >= imm */
716 	case BPF_JGE:
717 		emit(ctx, sltiu, MIPS_R_T9, dst, imm);
718 		emit(ctx, beqz, MIPS_R_T9, off);
719 		break;
720 	/* PC += off if dst < imm */
721 	case BPF_JLT:
722 		emit(ctx, sltiu, MIPS_R_T9, dst, imm);
723 		emit(ctx, bnez, MIPS_R_T9, off);
724 		break;
725 	/* PC += off if dst <= imm */
726 	case BPF_JLE:
727 		emit(ctx, sltiu, MIPS_R_T9, dst, imm + 1);
728 		emit(ctx, bnez, MIPS_R_T9, off);
729 		break;
730 	/* PC += off if dst > imm (signed) */
731 	case BPF_JSGT:
732 		emit(ctx, slti, MIPS_R_T9, dst, imm + 1);
733 		emit(ctx, beqz, MIPS_R_T9, off);
734 		break;
735 	/* PC += off if dst >= imm (signed) */
736 	case BPF_JSGE:
737 		emit(ctx, slti, MIPS_R_T9, dst, imm);
738 		emit(ctx, beqz, MIPS_R_T9, off);
739 		break;
740 	/* PC += off if dst < imm (signed) */
741 	case BPF_JSLT:
742 		emit(ctx, slti, MIPS_R_T9, dst, imm);
743 		emit(ctx, bnez, MIPS_R_T9, off);
744 		break;
745 	/* PC += off if dst <= imm (signed) */
746 	case BPF_JSLE:
747 		emit(ctx, slti, MIPS_R_T9, dst, imm + 1);
748 		emit(ctx, bnez, MIPS_R_T9, off);
749 		break;
750 	}
751 }
752 
753 /* Jump register (32-bit) */
754 void emit_jmp_r(struct jit_context *ctx, u8 dst, u8 src, s32 off, u8 op)
755 {
756 	switch (op) {
757 	/* No-op, used internally for branch optimization */
758 	case JIT_JNOP:
759 		break;
760 	/* PC += off if dst == src */
761 	case BPF_JEQ:
762 		emit(ctx, beq, dst, src, off);
763 		break;
764 	/* PC += off if dst != src */
765 	case BPF_JNE:
766 		emit(ctx, bne, dst, src, off);
767 		break;
768 	/* PC += off if dst & src */
769 	case BPF_JSET:
770 		emit(ctx, and, MIPS_R_T9, dst, src);
771 		emit(ctx, bnez, MIPS_R_T9, off);
772 		break;
773 	/* PC += off if (dst & imm) == 0 (not in BPF, used for long jumps) */
774 	case JIT_JNSET:
775 		emit(ctx, and, MIPS_R_T9, dst, src);
776 		emit(ctx, beqz, MIPS_R_T9, off);
777 		break;
778 	/* PC += off if dst > src */
779 	case BPF_JGT:
780 		emit(ctx, sltu, MIPS_R_T9, src, dst);
781 		emit(ctx, bnez, MIPS_R_T9, off);
782 		break;
783 	/* PC += off if dst >= src */
784 	case BPF_JGE:
785 		emit(ctx, sltu, MIPS_R_T9, dst, src);
786 		emit(ctx, beqz, MIPS_R_T9, off);
787 		break;
788 	/* PC += off if dst < src */
789 	case BPF_JLT:
790 		emit(ctx, sltu, MIPS_R_T9, dst, src);
791 		emit(ctx, bnez, MIPS_R_T9, off);
792 		break;
793 	/* PC += off if dst <= src */
794 	case BPF_JLE:
795 		emit(ctx, sltu, MIPS_R_T9, src, dst);
796 		emit(ctx, beqz, MIPS_R_T9, off);
797 		break;
798 	/* PC += off if dst > src (signed) */
799 	case BPF_JSGT:
800 		emit(ctx, slt, MIPS_R_T9, src, dst);
801 		emit(ctx, bnez, MIPS_R_T9, off);
802 		break;
803 	/* PC += off if dst >= src (signed) */
804 	case BPF_JSGE:
805 		emit(ctx, slt, MIPS_R_T9, dst, src);
806 		emit(ctx, beqz, MIPS_R_T9, off);
807 		break;
808 	/* PC += off if dst < src (signed) */
809 	case BPF_JSLT:
810 		emit(ctx, slt, MIPS_R_T9, dst, src);
811 		emit(ctx, bnez, MIPS_R_T9, off);
812 		break;
813 	/* PC += off if dst <= src (signed) */
814 	case BPF_JSLE:
815 		emit(ctx, slt, MIPS_R_T9, src, dst);
816 		emit(ctx, beqz, MIPS_R_T9, off);
817 		break;
818 	}
819 }
820 
821 /* Jump always */
822 int emit_ja(struct jit_context *ctx, s16 off)
823 {
824 	int target = get_target(ctx, ctx->bpf_index + off + 1);
825 
826 	if (target < 0)
827 		return -1;
828 	emit(ctx, j, target);
829 	emit(ctx, nop);
830 	return 0;
831 }
832 
833 /* Jump to epilogue */
834 int emit_exit(struct jit_context *ctx)
835 {
836 	int target = get_target(ctx, ctx->program->len);
837 
838 	if (target < 0)
839 		return -1;
840 	emit(ctx, j, target);
841 	emit(ctx, nop);
842 	return 0;
843 }
844 
845 /* Build the program body from eBPF bytecode */
846 static int build_body(struct jit_context *ctx)
847 {
848 	const struct bpf_prog *prog = ctx->program;
849 	unsigned int i;
850 
851 	ctx->stack_used = 0;
852 	for (i = 0; i < prog->len; i++) {
853 		const struct bpf_insn *insn = &prog->insnsi[i];
854 		u32 *descp = &ctx->descriptors[i];
855 		int ret;
856 
857 		access_reg(ctx, insn->src_reg);
858 		access_reg(ctx, insn->dst_reg);
859 
860 		ctx->bpf_index = i;
861 		if (ctx->target == NULL) {
862 			ctx->changes += INDEX(*descp) != ctx->jit_index;
863 			*descp &= JIT_DESC_CONVERT;
864 			*descp |= ctx->jit_index;
865 		}
866 
867 		ret = build_insn(insn, ctx);
868 		if (ret < 0)
869 			return ret;
870 
871 		if (ret > 0) {
872 			i++;
873 			if (ctx->target == NULL)
874 				descp[1] = ctx->jit_index;
875 		}
876 	}
877 
878 	/* Store the end offset, where the epilogue begins */
879 	ctx->descriptors[prog->len] = ctx->jit_index;
880 	return 0;
881 }
882 
883 /* Set the branch conversion flag on all instructions */
884 static void set_convert_flag(struct jit_context *ctx, bool enable)
885 {
886 	const struct bpf_prog *prog = ctx->program;
887 	u32 flag = enable ? JIT_DESC_CONVERT : 0;
888 	unsigned int i;
889 
890 	for (i = 0; i <= prog->len; i++)
891 		ctx->descriptors[i] = INDEX(ctx->descriptors[i]) | flag;
892 }
893 
894 static void jit_fill_hole(void *area, unsigned int size)
895 {
896 	u32 *p;
897 
898 	/* We are guaranteed to have aligned memory. */
899 	for (p = area; size >= sizeof(u32); size -= sizeof(u32))
900 		uasm_i_break(&p, BRK_BUG); /* Increments p */
901 }
902 
903 bool bpf_jit_needs_zext(void)
904 {
905 	return true;
906 }
907 
908 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
909 {
910 	struct bpf_prog *tmp, *orig_prog = prog;
911 	struct bpf_binary_header *header = NULL;
912 	struct jit_context ctx;
913 	bool tmp_blinded = false;
914 	unsigned int tmp_idx;
915 	unsigned int image_size;
916 	u8 *image_ptr;
917 	int tries;
918 
919 	/*
920 	 * If BPF JIT was not enabled then we must fall back to
921 	 * the interpreter.
922 	 */
923 	if (!prog->jit_requested)
924 		return orig_prog;
925 	/*
926 	 * If constant blinding was enabled and we failed during blinding
927 	 * then we must fall back to the interpreter. Otherwise, we save
928 	 * the new JITed code.
929 	 */
930 	tmp = bpf_jit_blind_constants(prog);
931 	if (IS_ERR(tmp))
932 		return orig_prog;
933 	if (tmp != prog) {
934 		tmp_blinded = true;
935 		prog = tmp;
936 	}
937 
938 	memset(&ctx, 0, sizeof(ctx));
939 	ctx.program = prog;
940 
941 	/*
942 	 * Not able to allocate memory for descriptors[], then
943 	 * we must fall back to the interpreter
944 	 */
945 	ctx.descriptors = kcalloc(prog->len + 1, sizeof(*ctx.descriptors),
946 				  GFP_KERNEL);
947 	if (ctx.descriptors == NULL)
948 		goto out_err;
949 
950 	/* First pass discovers used resources */
951 	if (build_body(&ctx) < 0)
952 		goto out_err;
953 	/*
954 	 * Second pass computes instruction offsets.
955 	 * If any PC-relative branches are out of range, a sequence of
956 	 * a PC-relative branch + a jump is generated, and we have to
957 	 * try again from the beginning to generate the new offsets.
958 	 * This is done until no additional conversions are necessary.
959 	 * The last two iterations are done with all branches being
960 	 * converted, to guarantee offset table convergence within a
961 	 * fixed number of iterations.
962 	 */
963 	ctx.jit_index = 0;
964 	build_prologue(&ctx);
965 	tmp_idx = ctx.jit_index;
966 
967 	tries = JIT_MAX_ITERATIONS;
968 	do {
969 		ctx.jit_index = tmp_idx;
970 		ctx.changes = 0;
971 		if (tries == 2)
972 			set_convert_flag(&ctx, true);
973 		if (build_body(&ctx) < 0)
974 			goto out_err;
975 	} while (ctx.changes > 0 && --tries > 0);
976 
977 	if (WARN_ONCE(ctx.changes > 0, "JIT offsets failed to converge"))
978 		goto out_err;
979 
980 	build_epilogue(&ctx, MIPS_R_RA);
981 
982 	/* Now we know the size of the structure to make */
983 	image_size = sizeof(u32) * ctx.jit_index;
984 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
985 				      sizeof(u32), jit_fill_hole);
986 	/*
987 	 * Not able to allocate memory for the structure then
988 	 * we must fall back to the interpretation
989 	 */
990 	if (header == NULL)
991 		goto out_err;
992 
993 	/* Actual pass to generate final JIT code */
994 	ctx.target = (u32 *)image_ptr;
995 	ctx.jit_index = 0;
996 
997 	/*
998 	 * If building the JITed code fails somehow,
999 	 * we fall back to the interpretation.
1000 	 */
1001 	build_prologue(&ctx);
1002 	if (build_body(&ctx) < 0)
1003 		goto out_err;
1004 	build_epilogue(&ctx, MIPS_R_RA);
1005 
1006 	/* Populate line info meta data */
1007 	set_convert_flag(&ctx, false);
1008 	bpf_prog_fill_jited_linfo(prog, &ctx.descriptors[1]);
1009 
1010 	/* Set as read-only exec and flush instruction cache */
1011 	bpf_jit_binary_lock_ro(header);
1012 	flush_icache_range((unsigned long)header,
1013 			   (unsigned long)&ctx.target[ctx.jit_index]);
1014 
1015 	if (bpf_jit_enable > 1)
1016 		bpf_jit_dump(prog->len, image_size, 2, ctx.target);
1017 
1018 	prog->bpf_func = (void *)ctx.target;
1019 	prog->jited = 1;
1020 	prog->jited_len = image_size;
1021 
1022 out:
1023 	if (tmp_blinded)
1024 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1025 					   tmp : orig_prog);
1026 	kfree(ctx.descriptors);
1027 	return prog;
1028 
1029 out_err:
1030 	prog = orig_prog;
1031 	if (header)
1032 		bpf_jit_binary_free(header);
1033 	goto out;
1034 }
1035