?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / arch / s390 / net / bpf_jit_comp.c
blob78a19c93b3802ff48b4c27a90d6428100fc5566f
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * BPF Jit compiler for s390.
4 *
5 * Minimum build requirements:
6 *
7 * - HAVE_MARCH_Z196_FEATURES: laal, laalg
8 * - HAVE_MARCH_Z10_FEATURES: msfi, cgrj, clgrj
9 * - HAVE_MARCH_Z9_109_FEATURES: alfi, llilf, clfi, oilf, nilf
10 * - PACK_STACK
11 * - 64BIT
12 *
13 * Copyright IBM Corp. 2012,2015
14 *
15 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
16 * Michael Holzheu <holzheu@linux.vnet.ibm.com>
17 */
18
19 #define KMSG_COMPONENT "bpf_jit"
20 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
21
22 #include <linux/netdevice.h>
23 #include <linux/filter.h>
24 #include <linux/init.h>
25 #include <linux/bpf.h>
26 #include <asm/cacheflush.h>
27 #include <asm/dis.h>
28 #include <asm/set_memory.h>
29 #include "bpf_jit.h"
30
31 struct bpf_jit {
32 u32 seen; /* Flags to remember seen eBPF instructions */
33 u32 seen_reg[16]; /* Array to remember which registers are used */
34 u32 *addrs; /* Array with relative instruction addresses */
35 u8 *prg_buf; /* Start of program */
36 int size; /* Size of program and literal pool */
37 int size_prg; /* Size of program */
38 int prg; /* Current position in program */
39 int lit_start; /* Start of literal pool */
40 int lit; /* Current position in literal pool */
41 int base_ip; /* Base address for literal pool */
42 int ret0_ip; /* Address of return 0 */
43 int exit_ip; /* Address of exit */
44 int tail_call_start; /* Tail call start offset */
45 int labels[1]; /* Labels for local jumps */
46 };
47
48 #define BPF_SIZE_MAX 0xffff /* Max size for program (16 bit branches) */
49
50 #define SEEN_SKB 1 /* skb access */
51 #define SEEN_MEM 2 /* use mem[] for temporary storage */
52 #define SEEN_RET0 4 /* ret0_ip points to a valid return 0 */
53 #define SEEN_LITERAL 8 /* code uses literals */
54 #define SEEN_FUNC 16 /* calls C functions */
55 #define SEEN_TAIL_CALL 32 /* code uses tail calls */
56 #define SEEN_REG_AX 64 /* code uses constant blinding */
57 #define SEEN_STACK (SEEN_FUNC | SEEN_MEM | SEEN_SKB)
58
59 /*
60 * s390 registers
61 */
62 #define REG_W0 (MAX_BPF_JIT_REG + 0) /* Work register 1 (even) */
63 #define REG_W1 (MAX_BPF_JIT_REG + 1) /* Work register 2 (odd) */
64 #define REG_SKB_DATA (MAX_BPF_JIT_REG + 2) /* SKB data register */
65 #define REG_L (MAX_BPF_JIT_REG + 3) /* Literal pool register */
66 #define REG_15 (MAX_BPF_JIT_REG + 4) /* Register 15 */
67 #define REG_0 REG_W0 /* Register 0 */
68 #define REG_1 REG_W1 /* Register 1 */
69 #define REG_2 BPF_REG_1 /* Register 2 */
70 #define REG_14 BPF_REG_0 /* Register 14 */
71
72 /*
73 * Mapping of BPF registers to s390 registers
74 */
75 static const int reg2hex[] = {
76 /* Return code */
77 [BPF_REG_0] = 14,
78 /* Function parameters */
79 [BPF_REG_1] = 2,
80 [BPF_REG_2] = 3,
81 [BPF_REG_3] = 4,
82 [BPF_REG_4] = 5,
83 [BPF_REG_5] = 6,
84 /* Call saved registers */
85 [BPF_REG_6] = 7,
86 [BPF_REG_7] = 8,
87 [BPF_REG_8] = 9,
88 [BPF_REG_9] = 10,
89 /* BPF stack pointer */
90 [BPF_REG_FP] = 13,
91 /* Register for blinding (shared with REG_SKB_DATA) */
92 [BPF_REG_AX] = 12,
93 /* SKB data pointer */
94 [REG_SKB_DATA] = 12,
95 /* Work registers for s390x backend */
96 [REG_W0] = 0,
97 [REG_W1] = 1,
98 [REG_L] = 11,
99 [REG_15] = 15,
100 };
101
102 static inline u32 reg(u32 dst_reg, u32 src_reg)
103 {
104 return reg2hex[dst_reg] << 4 | reg2hex[src_reg];
105 }
106
107 static inline u32 reg_high(u32 reg)
108 {
109 return reg2hex[reg] << 4;
110 }
111
112 static inline void reg_set_seen(struct bpf_jit *jit, u32 b1)
113 {
114 u32 r1 = reg2hex[b1];
115
116 if (!jit->seen_reg[r1] && r1 >= 6 && r1 <= 15)
117 jit->seen_reg[r1] = 1;
118 }
119
120 #define REG_SET_SEEN(b1) \
121 ({ \
122 reg_set_seen(jit, b1); \
123 })
124
125 #define REG_SEEN(b1) jit->seen_reg[reg2hex[(b1)]]
126
127 /*
128 * EMIT macros for code generation
129 */
130
131 #define _EMIT2(op) \
132 ({ \
133 if (jit->prg_buf) \
134 *(u16 *) (jit->prg_buf + jit->prg) = op; \
135 jit->prg += 2; \
136 })
137
138 #define EMIT2(op, b1, b2) \
139 ({ \
140 _EMIT2(op | reg(b1, b2)); \
141 REG_SET_SEEN(b1); \
142 REG_SET_SEEN(b2); \
143 })
144
145 #define _EMIT4(op) \
146 ({ \
147 if (jit->prg_buf) \
148 *(u32 *) (jit->prg_buf + jit->prg) = op; \
149 jit->prg += 4; \
150 })
151
152 #define EMIT4(op, b1, b2) \
153 ({ \
154 _EMIT4(op | reg(b1, b2)); \
155 REG_SET_SEEN(b1); \
156 REG_SET_SEEN(b2); \
157 })
158
159 #define EMIT4_RRF(op, b1, b2, b3) \
160 ({ \
161 _EMIT4(op | reg_high(b3) << 8 | reg(b1, b2)); \
162 REG_SET_SEEN(b1); \
163 REG_SET_SEEN(b2); \
164 REG_SET_SEEN(b3); \
165 })
166
167 #define _EMIT4_DISP(op, disp) \
168 ({ \
169 unsigned int __disp = (disp) & 0xfff; \
170 _EMIT4(op | __disp); \
171 })
172
173 #define EMIT4_DISP(op, b1, b2, disp) \
174 ({ \
175 _EMIT4_DISP(op | reg_high(b1) << 16 | \
176 reg_high(b2) << 8, disp); \
177 REG_SET_SEEN(b1); \
178 REG_SET_SEEN(b2); \
179 })
180
181 #define EMIT4_IMM(op, b1, imm) \
182 ({ \
183 unsigned int __imm = (imm) & 0xffff; \
184 _EMIT4(op | reg_high(b1) << 16 | __imm); \
185 REG_SET_SEEN(b1); \
186 })
187
188 #define EMIT4_PCREL(op, pcrel) \
189 ({ \
190 long __pcrel = ((pcrel) >> 1) & 0xffff; \
191 _EMIT4(op | __pcrel); \
192 })
193
194 #define _EMIT6(op1, op2) \
195 ({ \
196 if (jit->prg_buf) { \
197 *(u32 *) (jit->prg_buf + jit->prg) = op1; \
198 *(u16 *) (jit->prg_buf + jit->prg + 4) = op2; \
199 } \
200 jit->prg += 6; \
201 })
202
203 #define _EMIT6_DISP(op1, op2, disp) \
204 ({ \
205 unsigned int __disp = (disp) & 0xfff; \
206 _EMIT6(op1 | __disp, op2); \
207 })
208
209 #define _EMIT6_DISP_LH(op1, op2, disp) \
210 ({ \
211 u32 _disp = (u32) disp; \
212 unsigned int __disp_h = _disp & 0xff000; \
213 unsigned int __disp_l = _disp & 0x00fff; \
214 _EMIT6(op1 | __disp_l, op2 | __disp_h >> 4); \
215 })
216
217 #define EMIT6_DISP_LH(op1, op2, b1, b2, b3, disp) \
218 ({ \
219 _EMIT6_DISP_LH(op1 | reg(b1, b2) << 16 | \
220 reg_high(b3) << 8, op2, disp); \
221 REG_SET_SEEN(b1); \
222 REG_SET_SEEN(b2); \
223 REG_SET_SEEN(b3); \
224 })
225
226 #define EMIT6_PCREL_LABEL(op1, op2, b1, b2, label, mask) \
227 ({ \
228 int rel = (jit->labels[label] - jit->prg) >> 1; \
229 _EMIT6(op1 | reg(b1, b2) << 16 | (rel & 0xffff), \
230 op2 | mask << 12); \
231 REG_SET_SEEN(b1); \
232 REG_SET_SEEN(b2); \
233 })
234
235 #define EMIT6_PCREL_IMM_LABEL(op1, op2, b1, imm, label, mask) \
236 ({ \
237 int rel = (jit->labels[label] - jit->prg) >> 1; \
238 _EMIT6(op1 | (reg_high(b1) | mask) << 16 | \
239 (rel & 0xffff), op2 | (imm & 0xff) << 8); \
240 REG_SET_SEEN(b1); \
241 BUILD_BUG_ON(((unsigned long) imm) > 0xff); \
242 })
243
244 #define EMIT6_PCREL(op1, op2, b1, b2, i, off, mask) \
245 ({ \
246 /* Branch instruction needs 6 bytes */ \
247 int rel = (addrs[i + off + 1] - (addrs[i + 1] - 6)) / 2;\
248 _EMIT6(op1 | reg(b1, b2) << 16 | (rel & 0xffff), op2 | mask); \
249 REG_SET_SEEN(b1); \
250 REG_SET_SEEN(b2); \
251 })
252
253 #define _EMIT6_IMM(op, imm) \
254 ({ \
255 unsigned int __imm = (imm); \
256 _EMIT6(op | (__imm >> 16), __imm & 0xffff); \
257 })
258
259 #define EMIT6_IMM(op, b1, imm) \
260 ({ \
261 _EMIT6_IMM(op | reg_high(b1) << 16, imm); \
262 REG_SET_SEEN(b1); \
263 })
264
265 #define EMIT_CONST_U32(val) \
266 ({ \
267 unsigned int ret; \
268 ret = jit->lit - jit->base_ip; \
269 jit->seen |= SEEN_LITERAL; \
270 if (jit->prg_buf) \
271 *(u32 *) (jit->prg_buf + jit->lit) = (u32) val; \
272 jit->lit += 4; \
273 ret; \
274 })
275
276 #define EMIT_CONST_U64(val) \
277 ({ \
278 unsigned int ret; \
279 ret = jit->lit - jit->base_ip; \
280 jit->seen |= SEEN_LITERAL; \
281 if (jit->prg_buf) \
282 *(u64 *) (jit->prg_buf + jit->lit) = (u64) val; \
283 jit->lit += 8; \
284 ret; \
285 })
286
287 #define EMIT_ZERO(b1) \
288 ({ \
289 /* llgfr %dst,%dst (zero extend to 64 bit) */ \
290 EMIT4(0xb9160000, b1, b1); \
291 REG_SET_SEEN(b1); \
292 })
293
294 /*
295 * Fill whole space with illegal instructions
296 */
297 static void jit_fill_hole(void *area, unsigned int size)
298 {
299 memset(area, 0, size);
300 }
301
302 /*
303 * Save registers from "rs" (register start) to "re" (register end) on stack
304 */
305 static void save_regs(struct bpf_jit *jit, u32 rs, u32 re)
306 {
307 u32 off = STK_OFF_R6 + (rs - 6) * 8;
308
309 if (rs == re)
310 /* stg %rs,off(%r15) */
311 _EMIT6(0xe300f000 | rs << 20 | off, 0x0024);
312 else
313 /* stmg %rs,%re,off(%r15) */
314 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0024, off);
315 }
316
317 /*
318 * Restore registers from "rs" (register start) to "re" (register end) on stack
319 */
320 static void restore_regs(struct bpf_jit *jit, u32 rs, u32 re, u32 stack_depth)
321 {
322 u32 off = STK_OFF_R6 + (rs - 6) * 8;
323
324 if (jit->seen & SEEN_STACK)
325 off += STK_OFF + stack_depth;
326
327 if (rs == re)
328 /* lg %rs,off(%r15) */
329 _EMIT6(0xe300f000 | rs << 20 | off, 0x0004);
330 else
331 /* lmg %rs,%re,off(%r15) */
332 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0004, off);
333 }
334
335 /*
336 * Return first seen register (from start)
337 */
338 static int get_start(struct bpf_jit *jit, int start)
339 {
340 int i;
341
342 for (i = start; i <= 15; i++) {
343 if (jit->seen_reg[i])
344 return i;
345 }
346 return 0;
347 }
348
349 /*
350 * Return last seen register (from start) (gap >= 2)
351 */
352 static int get_end(struct bpf_jit *jit, int start)
353 {
354 int i;
355
356 for (i = start; i < 15; i++) {
357 if (!jit->seen_reg[i] && !jit->seen_reg[i + 1])
358 return i - 1;
359 }
360 return jit->seen_reg[15] ? 15 : 14;
361 }
362
363 #define REGS_SAVE 1
364 #define REGS_RESTORE 0
365 /*
366 * Save and restore clobbered registers (6-15) on stack.
367 * We save/restore registers in chunks with gap >= 2 registers.
368 */
369 static void save_restore_regs(struct bpf_jit *jit, int op, u32 stack_depth)
370 {
371
372 int re = 6, rs;
373
374 do {
375 rs = get_start(jit, re);
376 if (!rs)
377 break;
378 re = get_end(jit, rs + 1);
379 if (op == REGS_SAVE)
380 save_regs(jit, rs, re);
381 else
382 restore_regs(jit, rs, re, stack_depth);
383 re++;
384 } while (re <= 15);
385 }
386
387 /*
388 * For SKB access %b1 contains the SKB pointer. For "bpf_jit.S"
389 * we store the SKB header length on the stack and the SKB data
390 * pointer in REG_SKB_DATA if BPF_REG_AX is not used.
391 */
392 static void emit_load_skb_data_hlen(struct bpf_jit *jit)
393 {
394 /* Header length: llgf %w1,<len>(%b1) */
395 EMIT6_DISP_LH(0xe3000000, 0x0016, REG_W1, REG_0, BPF_REG_1,
396 offsetof(struct sk_buff, len));
397 /* s %w1,<data_len>(%b1) */
398 EMIT4_DISP(0x5b000000, REG_W1, BPF_REG_1,
399 offsetof(struct sk_buff, data_len));
400 /* stg %w1,ST_OFF_HLEN(%r0,%r15) */
401 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W1, REG_0, REG_15, STK_OFF_HLEN);
402 if (!(jit->seen & SEEN_REG_AX))
403 /* lg %skb_data,data_off(%b1) */
404 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_SKB_DATA, REG_0,
405 BPF_REG_1, offsetof(struct sk_buff, data));
406 }
407
408 /*
409 * Emit function prologue
410 *
411 * Save registers and create stack frame if necessary.
412 * See stack frame layout desription in "bpf_jit.h"!
413 */
414 static void bpf_jit_prologue(struct bpf_jit *jit, u32 stack_depth)
415 {
416 if (jit->seen & SEEN_TAIL_CALL) {
417 /* xc STK_OFF_TCCNT(4,%r15),STK_OFF_TCCNT(%r15) */
418 _EMIT6(0xd703f000 | STK_OFF_TCCNT, 0xf000 | STK_OFF_TCCNT);
419 } else {
420 /* j tail_call_start: NOP if no tail calls are used */
421 EMIT4_PCREL(0xa7f40000, 6);
422 _EMIT2(0);
423 }
424 /* Tail calls have to skip above initialization */
425 jit->tail_call_start = jit->prg;
426 /* Save registers */
427 save_restore_regs(jit, REGS_SAVE, stack_depth);
428 /* Setup literal pool */
429 if (jit->seen & SEEN_LITERAL) {
430 /* basr %r13,0 */
431 EMIT2(0x0d00, REG_L, REG_0);
432 jit->base_ip = jit->prg;
433 }
434 /* Setup stack and backchain */
435 if (jit->seen & SEEN_STACK) {
436 if (jit->seen & SEEN_FUNC)
437 /* lgr %w1,%r15 (backchain) */
438 EMIT4(0xb9040000, REG_W1, REG_15);
439 /* la %bfp,STK_160_UNUSED(%r15) (BPF frame pointer) */
440 EMIT4_DISP(0x41000000, BPF_REG_FP, REG_15, STK_160_UNUSED);
441 /* aghi %r15,-STK_OFF */
442 EMIT4_IMM(0xa70b0000, REG_15, -(STK_OFF + stack_depth));
443 if (jit->seen & SEEN_FUNC)
444 /* stg %w1,152(%r15) (backchain) */
445 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W1, REG_0,
446 REG_15, 152);
447 }
448 if (jit->seen & SEEN_SKB) {
449 emit_load_skb_data_hlen(jit);
450 /* stg %b1,ST_OFF_SKBP(%r0,%r15) */
451 EMIT6_DISP_LH(0xe3000000, 0x0024, BPF_REG_1, REG_0, REG_15,
452 STK_OFF_SKBP);
453 }
454 }
455
456 /*
457 * Function epilogue
458 */
459 static void bpf_jit_epilogue(struct bpf_jit *jit, u32 stack_depth)
460 {
461 /* Return 0 */
462 if (jit->seen & SEEN_RET0) {
463 jit->ret0_ip = jit->prg;
464 /* lghi %b0,0 */
465 EMIT4_IMM(0xa7090000, BPF_REG_0, 0);
466 }
467 jit->exit_ip = jit->prg;
468 /* Load exit code: lgr %r2,%b0 */
469 EMIT4(0xb9040000, REG_2, BPF_REG_0);
470 /* Restore registers */
471 save_restore_regs(jit, REGS_RESTORE, stack_depth);
472 /* br %r14 */
473 _EMIT2(0x07fe);
474 }
475
476 /*
477 * Compile one eBPF instruction into s390x code
478 *
479 * NOTE: Use noinline because for gcov (-fprofile-arcs) gcc allocates a lot of
480 * stack space for the large switch statement.
481 */
482 static noinline int bpf_jit_insn(struct bpf_jit *jit, struct bpf_prog *fp, int i)
483 {
484 struct bpf_insn *insn = &fp->insnsi[i];
485 int jmp_off, last, insn_count = 1;
486 unsigned int func_addr, mask;
487 u32 dst_reg = insn->dst_reg;
488 u32 src_reg = insn->src_reg;
489 u32 *addrs = jit->addrs;
490 s32 imm = insn->imm;
491 s16 off = insn->off;
492
493 if (dst_reg == BPF_REG_AX || src_reg == BPF_REG_AX)
494 jit->seen |= SEEN_REG_AX;
495 switch (insn->code) {
496 /*
497 * BPF_MOV
498 */
499 case BPF_ALU | BPF_MOV | BPF_X: /* dst = (u32) src */
500 /* llgfr %dst,%src */
501 EMIT4(0xb9160000, dst_reg, src_reg);
502 break;
503 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
504 /* lgr %dst,%src */
505 EMIT4(0xb9040000, dst_reg, src_reg);
506 break;
507 case BPF_ALU | BPF_MOV | BPF_K: /* dst = (u32) imm */
508 /* llilf %dst,imm */
509 EMIT6_IMM(0xc00f0000, dst_reg, imm);
510 break;
511 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = imm */
512 /* lgfi %dst,imm */
513 EMIT6_IMM(0xc0010000, dst_reg, imm);
514 break;
515 /*
516 * BPF_LD 64
517 */
518 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
519 {
520 /* 16 byte instruction that uses two 'struct bpf_insn' */
521 u64 imm64;
522
523 imm64 = (u64)(u32) insn[0].imm | ((u64)(u32) insn[1].imm) << 32;
524 /* lg %dst,<d(imm)>(%l) */
525 EMIT6_DISP_LH(0xe3000000, 0x0004, dst_reg, REG_0, REG_L,
526 EMIT_CONST_U64(imm64));
527 insn_count = 2;
528 break;
529 }
530 /*
531 * BPF_ADD
532 */
533 case BPF_ALU | BPF_ADD | BPF_X: /* dst = (u32) dst + (u32) src */
534 /* ar %dst,%src */
535 EMIT2(0x1a00, dst_reg, src_reg);
536 EMIT_ZERO(dst_reg);
537 break;
538 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst = dst + src */
539 /* agr %dst,%src */
540 EMIT4(0xb9080000, dst_reg, src_reg);
541 break;
542 case BPF_ALU | BPF_ADD | BPF_K: /* dst = (u32) dst + (u32) imm */
543 if (!imm)
544 break;
545 /* alfi %dst,imm */
546 EMIT6_IMM(0xc20b0000, dst_reg, imm);
547 EMIT_ZERO(dst_reg);
548 break;
549 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst = dst + imm */
550 if (!imm)
551 break;
552 /* agfi %dst,imm */
553 EMIT6_IMM(0xc2080000, dst_reg, imm);
554 break;
555 /*
556 * BPF_SUB
557 */
558 case BPF_ALU | BPF_SUB | BPF_X: /* dst = (u32) dst - (u32) src */
559 /* sr %dst,%src */
560 EMIT2(0x1b00, dst_reg, src_reg);
561 EMIT_ZERO(dst_reg);
562 break;
563 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst = dst - src */
564 /* sgr %dst,%src */
565 EMIT4(0xb9090000, dst_reg, src_reg);
566 break;
567 case BPF_ALU | BPF_SUB | BPF_K: /* dst = (u32) dst - (u32) imm */
568 if (!imm)
569 break;
570 /* alfi %dst,-imm */
571 EMIT6_IMM(0xc20b0000, dst_reg, -imm);
572 EMIT_ZERO(dst_reg);
573 break;
574 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst = dst - imm */
575 if (!imm)
576 break;
577 /* agfi %dst,-imm */
578 EMIT6_IMM(0xc2080000, dst_reg, -imm);
579 break;
580 /*
581 * BPF_MUL
582 */
583 case BPF_ALU | BPF_MUL | BPF_X: /* dst = (u32) dst * (u32) src */
584 /* msr %dst,%src */
585 EMIT4(0xb2520000, dst_reg, src_reg);
586 EMIT_ZERO(dst_reg);
587 break;
588 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst = dst * src */
589 /* msgr %dst,%src */
590 EMIT4(0xb90c0000, dst_reg, src_reg);
591 break;
592 case BPF_ALU | BPF_MUL | BPF_K: /* dst = (u32) dst * (u32) imm */
593 if (imm == 1)
594 break;
595 /* msfi %r5,imm */
596 EMIT6_IMM(0xc2010000, dst_reg, imm);
597 EMIT_ZERO(dst_reg);
598 break;
599 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst = dst * imm */
600 if (imm == 1)
601 break;
602 /* msgfi %dst,imm */
603 EMIT6_IMM(0xc2000000, dst_reg, imm);
604 break;
605 /*
606 * BPF_DIV / BPF_MOD
607 */
608 case BPF_ALU | BPF_DIV | BPF_X: /* dst = (u32) dst / (u32) src */
609 case BPF_ALU | BPF_MOD | BPF_X: /* dst = (u32) dst % (u32) src */
610 {
611 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
612
613 /* lhi %w0,0 */
614 EMIT4_IMM(0xa7080000, REG_W0, 0);
615 /* lr %w1,%dst */
616 EMIT2(0x1800, REG_W1, dst_reg);
617 /* dlr %w0,%src */
618 EMIT4(0xb9970000, REG_W0, src_reg);
619 /* llgfr %dst,%rc */
620 EMIT4(0xb9160000, dst_reg, rc_reg);
621 break;
622 }
623 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst = dst / src */
624 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst = dst % src */
625 {
626 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
627
628 /* lghi %w0,0 */
629 EMIT4_IMM(0xa7090000, REG_W0, 0);
630 /* lgr %w1,%dst */
631 EMIT4(0xb9040000, REG_W1, dst_reg);
632 /* dlgr %w0,%dst */
633 EMIT4(0xb9870000, REG_W0, src_reg);
634 /* lgr %dst,%rc */
635 EMIT4(0xb9040000, dst_reg, rc_reg);
636 break;
637 }
638 case BPF_ALU | BPF_DIV | BPF_K: /* dst = (u32) dst / (u32) imm */
639 case BPF_ALU | BPF_MOD | BPF_K: /* dst = (u32) dst % (u32) imm */
640 {
641 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
642
643 if (imm == 1) {
644 if (BPF_OP(insn->code) == BPF_MOD)
645 /* lhgi %dst,0 */
646 EMIT4_IMM(0xa7090000, dst_reg, 0);
647 break;
648 }
649 /* lhi %w0,0 */
650 EMIT4_IMM(0xa7080000, REG_W0, 0);
651 /* lr %w1,%dst */
652 EMIT2(0x1800, REG_W1, dst_reg);
653 /* dl %w0,<d(imm)>(%l) */
654 EMIT6_DISP_LH(0xe3000000, 0x0097, REG_W0, REG_0, REG_L,
655 EMIT_CONST_U32(imm));
656 /* llgfr %dst,%rc */
657 EMIT4(0xb9160000, dst_reg, rc_reg);
658 break;
659 }
660 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst = dst / imm */
661 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst = dst % imm */
662 {
663 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
664
665 if (imm == 1) {
666 if (BPF_OP(insn->code) == BPF_MOD)
667 /* lhgi %dst,0 */
668 EMIT4_IMM(0xa7090000, dst_reg, 0);
669 break;
670 }
671 /* lghi %w0,0 */
672 EMIT4_IMM(0xa7090000, REG_W0, 0);
673 /* lgr %w1,%dst */
674 EMIT4(0xb9040000, REG_W1, dst_reg);
675 /* dlg %w0,<d(imm)>(%l) */
676 EMIT6_DISP_LH(0xe3000000, 0x0087, REG_W0, REG_0, REG_L,
677 EMIT_CONST_U64(imm));
678 /* lgr %dst,%rc */
679 EMIT4(0xb9040000, dst_reg, rc_reg);
680 break;
681 }
682 /*
683 * BPF_AND
684 */
685 case BPF_ALU | BPF_AND | BPF_X: /* dst = (u32) dst & (u32) src */
686 /* nr %dst,%src */
687 EMIT2(0x1400, dst_reg, src_reg);
688 EMIT_ZERO(dst_reg);
689 break;
690 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
691 /* ngr %dst,%src */
692 EMIT4(0xb9800000, dst_reg, src_reg);
693 break;
694 case BPF_ALU | BPF_AND | BPF_K: /* dst = (u32) dst & (u32) imm */
695 /* nilf %dst,imm */
696 EMIT6_IMM(0xc00b0000, dst_reg, imm);
697 EMIT_ZERO(dst_reg);
698 break;
699 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
700 /* ng %dst,<d(imm)>(%l) */
701 EMIT6_DISP_LH(0xe3000000, 0x0080, dst_reg, REG_0, REG_L,
702 EMIT_CONST_U64(imm));
703 break;
704 /*
705 * BPF_OR
706 */
707 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
708 /* or %dst,%src */
709 EMIT2(0x1600, dst_reg, src_reg);
710 EMIT_ZERO(dst_reg);
711 break;
712 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
713 /* ogr %dst,%src */
714 EMIT4(0xb9810000, dst_reg, src_reg);
715 break;
716 case BPF_ALU | BPF_OR | BPF_K: /* dst = (u32) dst | (u32) imm */
717 /* oilf %dst,imm */
718 EMIT6_IMM(0xc00d0000, dst_reg, imm);
719 EMIT_ZERO(dst_reg);
720 break;
721 case BPF_ALU64 | BPF_OR | BPF_K: /* dst = dst | imm */
722 /* og %dst,<d(imm)>(%l) */
723 EMIT6_DISP_LH(0xe3000000, 0x0081, dst_reg, REG_0, REG_L,
724 EMIT_CONST_U64(imm));
725 break;
726 /*
727 * BPF_XOR
728 */
729 case BPF_ALU | BPF_XOR | BPF_X: /* dst = (u32) dst ^ (u32) src */
730 /* xr %dst,%src */
731 EMIT2(0x1700, dst_reg, src_reg);
732 EMIT_ZERO(dst_reg);
733 break;
734 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst = dst ^ src */
735 /* xgr %dst,%src */
736 EMIT4(0xb9820000, dst_reg, src_reg);
737 break;
738 case BPF_ALU | BPF_XOR | BPF_K: /* dst = (u32) dst ^ (u32) imm */
739 if (!imm)
740 break;
741 /* xilf %dst,imm */
742 EMIT6_IMM(0xc0070000, dst_reg, imm);
743 EMIT_ZERO(dst_reg);
744 break;
745 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst = dst ^ imm */
746 /* xg %dst,<d(imm)>(%l) */
747 EMIT6_DISP_LH(0xe3000000, 0x0082, dst_reg, REG_0, REG_L,
748 EMIT_CONST_U64(imm));
749 break;
750 /*
751 * BPF_LSH
752 */
753 case BPF_ALU | BPF_LSH | BPF_X: /* dst = (u32) dst << (u32) src */
754 /* sll %dst,0(%src) */
755 EMIT4_DISP(0x89000000, dst_reg, src_reg, 0);
756 EMIT_ZERO(dst_reg);
757 break;
758 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst = dst << src */
759 /* sllg %dst,%dst,0(%src) */
760 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, src_reg, 0);
761 break;
762 case BPF_ALU | BPF_LSH | BPF_K: /* dst = (u32) dst << (u32) imm */
763 if (imm == 0)
764 break;
765 /* sll %dst,imm(%r0) */
766 EMIT4_DISP(0x89000000, dst_reg, REG_0, imm);
767 EMIT_ZERO(dst_reg);
768 break;
769 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst = dst << imm */
770 if (imm == 0)
771 break;
772 /* sllg %dst,%dst,imm(%r0) */
773 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, REG_0, imm);
774 break;
775 /*
776 * BPF_RSH
777 */
778 case BPF_ALU | BPF_RSH | BPF_X: /* dst = (u32) dst >> (u32) src */
779 /* srl %dst,0(%src) */
780 EMIT4_DISP(0x88000000, dst_reg, src_reg, 0);
781 EMIT_ZERO(dst_reg);
782 break;
783 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst = dst >> src */
784 /* srlg %dst,%dst,0(%src) */
785 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, src_reg, 0);
786 break;
787 case BPF_ALU | BPF_RSH | BPF_K: /* dst = (u32) dst >> (u32) imm */
788 if (imm == 0)
789 break;
790 /* srl %dst,imm(%r0) */
791 EMIT4_DISP(0x88000000, dst_reg, REG_0, imm);
792 EMIT_ZERO(dst_reg);
793 break;
794 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst = dst >> imm */
795 if (imm == 0)
796 break;
797 /* srlg %dst,%dst,imm(%r0) */
798 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, REG_0, imm);
799 break;
800 /*
801 * BPF_ARSH
802 */
803 case BPF_ALU64 | BPF_ARSH | BPF_X: /* ((s64) dst) >>= src */
804 /* srag %dst,%dst,0(%src) */
805 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, src_reg, 0);
806 break;
807 case BPF_ALU64 | BPF_ARSH | BPF_K: /* ((s64) dst) >>= imm */
808 if (imm == 0)
809 break;
810 /* srag %dst,%dst,imm(%r0) */
811 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, REG_0, imm);
812 break;
813 /*
814 * BPF_NEG
815 */
816 case BPF_ALU | BPF_NEG: /* dst = (u32) -dst */
817 /* lcr %dst,%dst */
818 EMIT2(0x1300, dst_reg, dst_reg);
819 EMIT_ZERO(dst_reg);
820 break;
821 case BPF_ALU64 | BPF_NEG: /* dst = -dst */
822 /* lcgr %dst,%dst */
823 EMIT4(0xb9130000, dst_reg, dst_reg);
824 break;
825 /*
826 * BPF_FROM_BE/LE
827 */
828 case BPF_ALU | BPF_END | BPF_FROM_BE:
829 /* s390 is big endian, therefore only clear high order bytes */
830 switch (imm) {
831 case 16: /* dst = (u16) cpu_to_be16(dst) */
832 /* llghr %dst,%dst */
833 EMIT4(0xb9850000, dst_reg, dst_reg);
834 break;
835 case 32: /* dst = (u32) cpu_to_be32(dst) */
836 /* llgfr %dst,%dst */
837 EMIT4(0xb9160000, dst_reg, dst_reg);
838 break;
839 case 64: /* dst = (u64) cpu_to_be64(dst) */
840 break;
841 }
842 break;
843 case BPF_ALU | BPF_END | BPF_FROM_LE:
844 switch (imm) {
845 case 16: /* dst = (u16) cpu_to_le16(dst) */
846 /* lrvr %dst,%dst */
847 EMIT4(0xb91f0000, dst_reg, dst_reg);
848 /* srl %dst,16(%r0) */
849 EMIT4_DISP(0x88000000, dst_reg, REG_0, 16);
850 /* llghr %dst,%dst */
851 EMIT4(0xb9850000, dst_reg, dst_reg);
852 break;
853 case 32: /* dst = (u32) cpu_to_le32(dst) */
854 /* lrvr %dst,%dst */
855 EMIT4(0xb91f0000, dst_reg, dst_reg);
856 /* llgfr %dst,%dst */
857 EMIT4(0xb9160000, dst_reg, dst_reg);
858 break;
859 case 64: /* dst = (u64) cpu_to_le64(dst) */
860 /* lrvgr %dst,%dst */
861 EMIT4(0xb90f0000, dst_reg, dst_reg);
862 break;
863 }
864 break;
865 /*
866 * BPF_ST(X)
867 */
868 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src_reg */
869 /* stcy %src,off(%dst) */
870 EMIT6_DISP_LH(0xe3000000, 0x0072, src_reg, dst_reg, REG_0, off);
871 jit->seen |= SEEN_MEM;
872 break;
873 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
874 /* sthy %src,off(%dst) */
875 EMIT6_DISP_LH(0xe3000000, 0x0070, src_reg, dst_reg, REG_0, off);
876 jit->seen |= SEEN_MEM;
877 break;
878 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
879 /* sty %src,off(%dst) */
880 EMIT6_DISP_LH(0xe3000000, 0x0050, src_reg, dst_reg, REG_0, off);
881 jit->seen |= SEEN_MEM;
882 break;
883 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
884 /* stg %src,off(%dst) */
885 EMIT6_DISP_LH(0xe3000000, 0x0024, src_reg, dst_reg, REG_0, off);
886 jit->seen |= SEEN_MEM;
887 break;
888 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
889 /* lhi %w0,imm */
890 EMIT4_IMM(0xa7080000, REG_W0, (u8) imm);
891 /* stcy %w0,off(dst) */
892 EMIT6_DISP_LH(0xe3000000, 0x0072, REG_W0, dst_reg, REG_0, off);
893 jit->seen |= SEEN_MEM;
894 break;
895 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
896 /* lhi %w0,imm */
897 EMIT4_IMM(0xa7080000, REG_W0, (u16) imm);
898 /* sthy %w0,off(dst) */
899 EMIT6_DISP_LH(0xe3000000, 0x0070, REG_W0, dst_reg, REG_0, off);
900 jit->seen |= SEEN_MEM;
901 break;
902 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
903 /* llilf %w0,imm */
904 EMIT6_IMM(0xc00f0000, REG_W0, (u32) imm);
905 /* sty %w0,off(%dst) */
906 EMIT6_DISP_LH(0xe3000000, 0x0050, REG_W0, dst_reg, REG_0, off);
907 jit->seen |= SEEN_MEM;
908 break;
909 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
910 /* lgfi %w0,imm */
911 EMIT6_IMM(0xc0010000, REG_W0, imm);
912 /* stg %w0,off(%dst) */
913 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W0, dst_reg, REG_0, off);
914 jit->seen |= SEEN_MEM;
915 break;
916 /*
917 * BPF_STX XADD (atomic_add)
918 */
919 case BPF_STX | BPF_XADD | BPF_W: /* *(u32 *)(dst + off) += src */
920 /* laal %w0,%src,off(%dst) */
921 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W0, src_reg,
922 dst_reg, off);
923 jit->seen |= SEEN_MEM;
924 break;
925 case BPF_STX | BPF_XADD | BPF_DW: /* *(u64 *)(dst + off) += src */
926 /* laalg %w0,%src,off(%dst) */
927 EMIT6_DISP_LH(0xeb000000, 0x00ea, REG_W0, src_reg,
928 dst_reg, off);
929 jit->seen |= SEEN_MEM;
930 break;
931 /*
932 * BPF_LDX
933 */
934 case BPF_LDX | BPF_MEM | BPF_B: /* dst = *(u8 *)(ul) (src + off) */
935 /* llgc %dst,0(off,%src) */
936 EMIT6_DISP_LH(0xe3000000, 0x0090, dst_reg, src_reg, REG_0, off);
937 jit->seen |= SEEN_MEM;
938 break;
939 case BPF_LDX | BPF_MEM | BPF_H: /* dst = *(u16 *)(ul) (src + off) */
940 /* llgh %dst,0(off,%src) */
941 EMIT6_DISP_LH(0xe3000000, 0x0091, dst_reg, src_reg, REG_0, off);
942 jit->seen |= SEEN_MEM;
943 break;
944 case BPF_LDX | BPF_MEM | BPF_W: /* dst = *(u32 *)(ul) (src + off) */
945 /* llgf %dst,off(%src) */
946 jit->seen |= SEEN_MEM;
947 EMIT6_DISP_LH(0xe3000000, 0x0016, dst_reg, src_reg, REG_0, off);
948 break;
949 case BPF_LDX | BPF_MEM | BPF_DW: /* dst = *(u64 *)(ul) (src + off) */
950 /* lg %dst,0(off,%src) */
951 jit->seen |= SEEN_MEM;
952 EMIT6_DISP_LH(0xe3000000, 0x0004, dst_reg, src_reg, REG_0, off);
953 break;
954 /*
955 * BPF_JMP / CALL
956 */
957 case BPF_JMP | BPF_CALL:
958 {
959 /*
960 * b0 = (__bpf_call_base + imm)(b1, b2, b3, b4, b5)
961 */
962 const u64 func = (u64)__bpf_call_base + imm;
963
964 REG_SET_SEEN(BPF_REG_5);
965 jit->seen |= SEEN_FUNC;
966 /* lg %w1,<d(imm)>(%l) */
967 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_W1, REG_0, REG_L,
968 EMIT_CONST_U64(func));
969 /* basr %r14,%w1 */
970 EMIT2(0x0d00, REG_14, REG_W1);
971 /* lgr %b0,%r2: load return value into %b0 */
972 EMIT4(0xb9040000, BPF_REG_0, REG_2);
973 if ((jit->seen & SEEN_SKB) &&
974 bpf_helper_changes_pkt_data((void *)func)) {
975 /* lg %b1,ST_OFF_SKBP(%r15) */
976 EMIT6_DISP_LH(0xe3000000, 0x0004, BPF_REG_1, REG_0,
977 REG_15, STK_OFF_SKBP);
978 emit_load_skb_data_hlen(jit);
979 }
980 break;
981 }
982 case BPF_JMP | BPF_TAIL_CALL:
983 /*
984 * Implicit input:
985 * B1: pointer to ctx
986 * B2: pointer to bpf_array
987 * B3: index in bpf_array
988 */
989 jit->seen |= SEEN_TAIL_CALL;
990
991 /*
992 * if (index >= array->map.max_entries)
993 * goto out;
994 */
995
996 /* llgf %w1,map.max_entries(%b2) */
997 EMIT6_DISP_LH(0xe3000000, 0x0016, REG_W1, REG_0, BPF_REG_2,
998 offsetof(struct bpf_array, map.max_entries));
999 /* clgrj %b3,%w1,0xa,label0: if %b3 >= %w1 goto out */
1000 EMIT6_PCREL_LABEL(0xec000000, 0x0065, BPF_REG_3,
1001 REG_W1, 0, 0xa);
1002
1003 /*
1004 * if (tail_call_cnt++ > MAX_TAIL_CALL_CNT)
1005 * goto out;
1006 */
1007
1008 if (jit->seen & SEEN_STACK)
1009 off = STK_OFF_TCCNT + STK_OFF + fp->aux->stack_depth;
1010 else
1011 off = STK_OFF_TCCNT;
1012 /* lhi %w0,1 */
1013 EMIT4_IMM(0xa7080000, REG_W0, 1);
1014 /* laal %w1,%w0,off(%r15) */
1015 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W1, REG_W0, REG_15, off);
1016 /* clij %w1,MAX_TAIL_CALL_CNT,0x2,label0 */
1017 EMIT6_PCREL_IMM_LABEL(0xec000000, 0x007f, REG_W1,
1018 MAX_TAIL_CALL_CNT, 0, 0x2);
1019
1020 /*
1021 * prog = array->ptrs[index];
1022 * if (prog == NULL)
1023 * goto out;
1024 */
1025
1026 /* sllg %r1,%b3,3: %r1 = index * 8 */
1027 EMIT6_DISP_LH(0xeb000000, 0x000d, REG_1, BPF_REG_3, REG_0, 3);
1028 /* lg %r1,prog(%b2,%r1) */
1029 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, BPF_REG_2,
1030 REG_1, offsetof(struct bpf_array, ptrs));
1031 /* clgij %r1,0,0x8,label0 */
1032 EMIT6_PCREL_IMM_LABEL(0xec000000, 0x007d, REG_1, 0, 0, 0x8);
1033
1034 /*
1035 * Restore registers before calling function
1036 */
1037 save_restore_regs(jit, REGS_RESTORE, fp->aux->stack_depth);
1038
1039 /*
1040 * goto *(prog->bpf_func + tail_call_start);
1041 */
1042
1043 /* lg %r1,bpf_func(%r1) */
1044 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, REG_1, REG_0,
1045 offsetof(struct bpf_prog, bpf_func));
1046 /* bc 0xf,tail_call_start(%r1) */
1047 _EMIT4(0x47f01000 + jit->tail_call_start);
1048 /* out: */
1049 jit->labels[0] = jit->prg;
1050 break;
1051 case BPF_JMP | BPF_EXIT: /* return b0 */
1052 last = (i == fp->len - 1) ? 1 : 0;
1053 if (last && !(jit->seen & SEEN_RET0))
1054 break;
1055 /* j <exit> */
1056 EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg);
1057 break;
1058 /*
1059 * Branch relative (number of skipped instructions) to offset on
1060 * condition.
1061 *
1062 * Condition code to mask mapping:
1063 *
1064 * CC | Description | Mask
1065 * ------------------------------
1066 * 0 | Operands equal | 8
1067 * 1 | First operand low | 4
1068 * 2 | First operand high | 2
1069 * 3 | Unused | 1
1070 *
1071 * For s390x relative branches: ip = ip + off_bytes
1072 * For BPF relative branches: insn = insn + off_insns + 1
1073 *
1074 * For example for s390x with offset 0 we jump to the branch
1075 * instruction itself (loop) and for BPF with offset 0 we
1076 * branch to the instruction behind the branch.
1077 */
1078 case BPF_JMP | BPF_JA: /* if (true) */
1079 mask = 0xf000; /* j */
1080 goto branch_oc;
1081 case BPF_JMP | BPF_JSGT | BPF_K: /* ((s64) dst > (s64) imm) */
1082 mask = 0x2000; /* jh */
1083 goto branch_ks;
1084 case BPF_JMP | BPF_JSLT | BPF_K: /* ((s64) dst < (s64) imm) */
1085 mask = 0x4000; /* jl */
1086 goto branch_ks;
1087 case BPF_JMP | BPF_JSGE | BPF_K: /* ((s64) dst >= (s64) imm) */
1088 mask = 0xa000; /* jhe */
1089 goto branch_ks;
1090 case BPF_JMP | BPF_JSLE | BPF_K: /* ((s64) dst <= (s64) imm) */
1091 mask = 0xc000; /* jle */
1092 goto branch_ks;
1093 case BPF_JMP | BPF_JGT | BPF_K: /* (dst_reg > imm) */
1094 mask = 0x2000; /* jh */
1095 goto branch_ku;
1096 case BPF_JMP | BPF_JLT | BPF_K: /* (dst_reg < imm) */
1097 mask = 0x4000; /* jl */
1098 goto branch_ku;
1099 case BPF_JMP | BPF_JGE | BPF_K: /* (dst_reg >= imm) */
1100 mask = 0xa000; /* jhe */
1101 goto branch_ku;
1102 case BPF_JMP | BPF_JLE | BPF_K: /* (dst_reg <= imm) */
1103 mask = 0xc000; /* jle */
1104 goto branch_ku;
1105 case BPF_JMP | BPF_JNE | BPF_K: /* (dst_reg != imm) */
1106 mask = 0x7000; /* jne */
1107 goto branch_ku;
1108 case BPF_JMP | BPF_JEQ | BPF_K: /* (dst_reg == imm) */
1109 mask = 0x8000; /* je */
1110 goto branch_ku;
1111 case BPF_JMP | BPF_JSET | BPF_K: /* (dst_reg & imm) */
1112 mask = 0x7000; /* jnz */
1113 /* lgfi %w1,imm (load sign extend imm) */
1114 EMIT6_IMM(0xc0010000, REG_W1, imm);
1115 /* ngr %w1,%dst */
1116 EMIT4(0xb9800000, REG_W1, dst_reg);
1117 goto branch_oc;
1118
1119 case BPF_JMP | BPF_JSGT | BPF_X: /* ((s64) dst > (s64) src) */
1120 mask = 0x2000; /* jh */
1121 goto branch_xs;
1122 case BPF_JMP | BPF_JSLT | BPF_X: /* ((s64) dst < (s64) src) */
1123 mask = 0x4000; /* jl */
1124 goto branch_xs;
1125 case BPF_JMP | BPF_JSGE | BPF_X: /* ((s64) dst >= (s64) src) */
1126 mask = 0xa000; /* jhe */
1127 goto branch_xs;
1128 case BPF_JMP | BPF_JSLE | BPF_X: /* ((s64) dst <= (s64) src) */
1129 mask = 0xc000; /* jle */
1130 goto branch_xs;
1131 case BPF_JMP | BPF_JGT | BPF_X: /* (dst > src) */
1132 mask = 0x2000; /* jh */
1133 goto branch_xu;
1134 case BPF_JMP | BPF_JLT | BPF_X: /* (dst < src) */
1135 mask = 0x4000; /* jl */
1136 goto branch_xu;
1137 case BPF_JMP | BPF_JGE | BPF_X: /* (dst >= src) */
1138 mask = 0xa000; /* jhe */
1139 goto branch_xu;
1140 case BPF_JMP | BPF_JLE | BPF_X: /* (dst <= src) */
1141 mask = 0xc000; /* jle */
1142 goto branch_xu;
1143 case BPF_JMP | BPF_JNE | BPF_X: /* (dst != src) */
1144 mask = 0x7000; /* jne */
1145 goto branch_xu;
1146 case BPF_JMP | BPF_JEQ | BPF_X: /* (dst == src) */
1147 mask = 0x8000; /* je */
1148 goto branch_xu;
1149 case BPF_JMP | BPF_JSET | BPF_X: /* (dst & src) */
1150 mask = 0x7000; /* jnz */
1151 /* ngrk %w1,%dst,%src */
1152 EMIT4_RRF(0xb9e40000, REG_W1, dst_reg, src_reg);
1153 goto branch_oc;
1154 branch_ks:
1155 /* lgfi %w1,imm (load sign extend imm) */
1156 EMIT6_IMM(0xc0010000, REG_W1, imm);
1157 /* cgrj %dst,%w1,mask,off */
1158 EMIT6_PCREL(0xec000000, 0x0064, dst_reg, REG_W1, i, off, mask);
1159 break;
1160 branch_ku:
1161 /* lgfi %w1,imm (load sign extend imm) */
1162 EMIT6_IMM(0xc0010000, REG_W1, imm);
1163 /* clgrj %dst,%w1,mask,off */
1164 EMIT6_PCREL(0xec000000, 0x0065, dst_reg, REG_W1, i, off, mask);
1165 break;
1166 branch_xs:
1167 /* cgrj %dst,%src,mask,off */
1168 EMIT6_PCREL(0xec000000, 0x0064, dst_reg, src_reg, i, off, mask);
1169 break;
1170 branch_xu:
1171 /* clgrj %dst,%src,mask,off */
1172 EMIT6_PCREL(0xec000000, 0x0065, dst_reg, src_reg, i, off, mask);
1173 break;
1174 branch_oc:
1175 /* brc mask,jmp_off (branch instruction needs 4 bytes) */
1176 jmp_off = addrs[i + off + 1] - (addrs[i + 1] - 4);
1177 EMIT4_PCREL(0xa7040000 | mask << 8, jmp_off);
1178 break;
1179 /*
1180 * BPF_LD
1181 */
1182 case BPF_LD | BPF_ABS | BPF_B: /* b0 = *(u8 *) (skb->data+imm) */
1183 case BPF_LD | BPF_IND | BPF_B: /* b0 = *(u8 *) (skb->data+imm+src) */
1184 if ((BPF_MODE(insn->code) == BPF_ABS) && (imm >= 0))
1185 func_addr = __pa(sk_load_byte_pos);
1186 else
1187 func_addr = __pa(sk_load_byte);
1188 goto call_fn;
1189 case BPF_LD | BPF_ABS | BPF_H: /* b0 = *(u16 *) (skb->data+imm) */
1190 case BPF_LD | BPF_IND | BPF_H: /* b0 = *(u16 *) (skb->data+imm+src) */
1191 if ((BPF_MODE(insn->code) == BPF_ABS) && (imm >= 0))
1192 func_addr = __pa(sk_load_half_pos);
1193 else
1194 func_addr = __pa(sk_load_half);
1195 goto call_fn;
1196 case BPF_LD | BPF_ABS | BPF_W: /* b0 = *(u32 *) (skb->data+imm) */
1197 case BPF_LD | BPF_IND | BPF_W: /* b0 = *(u32 *) (skb->data+imm+src) */
1198 if ((BPF_MODE(insn->code) == BPF_ABS) && (imm >= 0))
1199 func_addr = __pa(sk_load_word_pos);
1200 else
1201 func_addr = __pa(sk_load_word);
1202 goto call_fn;
1203 call_fn:
1204 jit->seen |= SEEN_SKB | SEEN_RET0 | SEEN_FUNC;
1205 REG_SET_SEEN(REG_14); /* Return address of possible func call */
1206
1207 /*
1208 * Implicit input:
1209 * BPF_REG_6 (R7) : skb pointer
1210 * REG_SKB_DATA (R12): skb data pointer (if no BPF_REG_AX)
1211 *
1212 * Calculated input:
1213 * BPF_REG_2 (R3) : offset of byte(s) to fetch in skb
1214 * BPF_REG_5 (R6) : return address
1215 *
1216 * Output:
1217 * BPF_REG_0 (R14): data read from skb
1218 *
1219 * Scratch registers (BPF_REG_1-5)
1220 */
1221
1222 /* Call function: llilf %w1,func_addr */
1223 EMIT6_IMM(0xc00f0000, REG_W1, func_addr);
1224
1225 /* Offset: lgfi %b2,imm */
1226 EMIT6_IMM(0xc0010000, BPF_REG_2, imm);
1227 if (BPF_MODE(insn->code) == BPF_IND)
1228 /* agfr %b2,%src (%src is s32 here) */
1229 EMIT4(0xb9180000, BPF_REG_2, src_reg);
1230
1231 /* Reload REG_SKB_DATA if BPF_REG_AX is used */
1232 if (jit->seen & SEEN_REG_AX)
1233 /* lg %skb_data,data_off(%b6) */
1234 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_SKB_DATA, REG_0,
1235 BPF_REG_6, offsetof(struct sk_buff, data));
1236 /* basr %b5,%w1 (%b5 is call saved) */
1237 EMIT2(0x0d00, BPF_REG_5, REG_W1);
1238
1239 /*
1240 * Note: For fast access we jump directly after the
1241 * jnz instruction from bpf_jit.S
1242 */
1243 /* jnz <ret0> */
1244 EMIT4_PCREL(0xa7740000, jit->ret0_ip - jit->prg);
1245 break;
1246 default: /* too complex, give up */
1247 pr_err("Unknown opcode %02x\n", insn->code);
1248 return -1;
1249 }
1250 return insn_count;
1251 }
1252
1253 /*
1254 * Compile eBPF program into s390x code
1255 */
1256 static int bpf_jit_prog(struct bpf_jit *jit, struct bpf_prog *fp)
1257 {
1258 int i, insn_count;
1259
1260 jit->lit = jit->lit_start;
1261 jit->prg = 0;
1262
1263 bpf_jit_prologue(jit, fp->aux->stack_depth);
1264 for (i = 0; i < fp->len; i += insn_count) {
1265 insn_count = bpf_jit_insn(jit, fp, i);
1266 if (insn_count < 0)
1267 return -1;
1268 /* Next instruction address */
1269 jit->addrs[i + insn_count] = jit->prg;
1270 }
1271 bpf_jit_epilogue(jit, fp->aux->stack_depth);
1272
1273 jit->lit_start = jit->prg;
1274 jit->size = jit->lit;
1275 jit->size_prg = jit->prg;
1276 return 0;
1277 }
1278
1279 /*
1280 * Compile eBPF program "fp"
1281 */
1282 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
1283 {
1284 struct bpf_prog *tmp, *orig_fp = fp;
1285 struct bpf_binary_header *header;
1286 bool tmp_blinded = false;
1287 struct bpf_jit jit;
1288 int pass;
1289
1290 if (!fp->jit_requested)
1291 return orig_fp;
1292
1293 tmp = bpf_jit_blind_constants(fp);
1294 /*
1295 * If blinding was requested and we failed during blinding,
1296 * we must fall back to the interpreter.
1297 */
1298 if (IS_ERR(tmp))
1299 return orig_fp;
1300 if (tmp != fp) {
1301 tmp_blinded = true;
1302 fp = tmp;
1303 }
1304
1305 memset(&jit, 0, sizeof(jit));
1306 jit.addrs = kcalloc(fp->len + 1, sizeof(*jit.addrs), GFP_KERNEL);
1307 if (jit.addrs == NULL) {
1308 fp = orig_fp;
1309 goto out;
1310 }
1311 /*
1312 * Three initial passes:
1313 * - 1/2: Determine clobbered registers
1314 * - 3: Calculate program size and addrs arrray
1315 */
1316 for (pass = 1; pass <= 3; pass++) {
1317 if (bpf_jit_prog(&jit, fp)) {
1318 fp = orig_fp;
1319 goto free_addrs;
1320 }
1321 }
1322 /*
1323 * Final pass: Allocate and generate program
1324 */
1325 if (jit.size >= BPF_SIZE_MAX) {
1326 fp = orig_fp;
1327 goto free_addrs;
1328 }
1329 header = bpf_jit_binary_alloc(jit.size, &jit.prg_buf, 2, jit_fill_hole);
1330 if (!header) {
1331 fp = orig_fp;
1332 goto free_addrs;
1333 }
1334 if (bpf_jit_prog(&jit, fp)) {
1335 fp = orig_fp;
1336 goto free_addrs;
1337 }
1338 if (bpf_jit_enable > 1) {
1339 bpf_jit_dump(fp->len, jit.size, pass, jit.prg_buf);
1340 print_fn_code(jit.prg_buf, jit.size_prg);
1341 }
1342 bpf_jit_binary_lock_ro(header);
1343 fp->bpf_func = (void *) jit.prg_buf;
1344 fp->jited = 1;
1345 fp->jited_len = jit.size;
1346 free_addrs:
1347 kfree(jit.addrs);
1348 out:
1349 if (tmp_blinded)
1350 bpf_jit_prog_release_other(fp, fp == orig_fp ?
1351 tmp : orig_fp);
1352 return fp;
1353 }
CRIS linux kernel tree