?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.8-rc8
[linux/fpc-iii.git] / arch / s390 / net / bpf_jit_comp.c
blobbee281f3163d039e2fa203bcdaa02b8bf87d6ced
1 /*
2 * BPF Jit compiler for s390.
3 *
4 * Minimum build requirements:
5 *
6 * - HAVE_MARCH_Z196_FEATURES: laal, laalg
7 * - HAVE_MARCH_Z10_FEATURES: msfi, cgrj, clgrj
8 * - HAVE_MARCH_Z9_109_FEATURES: alfi, llilf, clfi, oilf, nilf
9 * - PACK_STACK
10 * - 64BIT
11 *
12 * Copyright IBM Corp. 2012,2015
13 *
14 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
15 * Michael Holzheu <holzheu@linux.vnet.ibm.com>
16 */
17
18 #define KMSG_COMPONENT "bpf_jit"
19 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
20
21 #include <linux/netdevice.h>
22 #include <linux/filter.h>
23 #include <linux/init.h>
24 #include <linux/bpf.h>
25 #include <asm/cacheflush.h>
26 #include <asm/dis.h>
27 #include "bpf_jit.h"
28
29 int bpf_jit_enable __read_mostly;
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_SKB_CHANGE 64 /* code changes skb data */
57 #define SEEN_REG_AX 128 /* code uses constant blinding */
58 #define SEEN_STACK (SEEN_FUNC | SEEN_MEM | SEEN_SKB)
59
60 /*
61 * s390 registers
62 */
63 #define REG_W0 (MAX_BPF_JIT_REG + 0) /* Work register 1 (even) */
64 #define REG_W1 (MAX_BPF_JIT_REG + 1) /* Work register 2 (odd) */
65 #define REG_SKB_DATA (MAX_BPF_JIT_REG + 2) /* SKB data register */
66 #define REG_L (MAX_BPF_JIT_REG + 3) /* Literal pool register */
67 #define REG_15 (MAX_BPF_JIT_REG + 4) /* Register 15 */
68 #define REG_0 REG_W0 /* Register 0 */
69 #define REG_1 REG_W1 /* Register 1 */
70 #define REG_2 BPF_REG_1 /* Register 2 */
71 #define REG_14 BPF_REG_0 /* Register 14 */
72
73 /*
74 * Mapping of BPF registers to s390 registers
75 */
76 static const int reg2hex[] = {
77 /* Return code */
78 [BPF_REG_0] = 14,
79 /* Function parameters */
80 [BPF_REG_1] = 2,
81 [BPF_REG_2] = 3,
82 [BPF_REG_3] = 4,
83 [BPF_REG_4] = 5,
84 [BPF_REG_5] = 6,
85 /* Call saved registers */
86 [BPF_REG_6] = 7,
87 [BPF_REG_7] = 8,
88 [BPF_REG_8] = 9,
89 [BPF_REG_9] = 10,
90 /* BPF stack pointer */
91 [BPF_REG_FP] = 13,
92 /* Register for blinding (shared with REG_SKB_DATA) */
93 [BPF_REG_AX] = 12,
94 /* SKB data pointer */
95 [REG_SKB_DATA] = 12,
96 /* Work registers for s390x backend */
97 [REG_W0] = 0,
98 [REG_W1] = 1,
99 [REG_L] = 11,
100 [REG_15] = 15,
101 };
102
103 static inline u32 reg(u32 dst_reg, u32 src_reg)
104 {
105 return reg2hex[dst_reg] << 4 | reg2hex[src_reg];
106 }
107
108 static inline u32 reg_high(u32 reg)
109 {
110 return reg2hex[reg] << 4;
111 }
112
113 static inline void reg_set_seen(struct bpf_jit *jit, u32 b1)
114 {
115 u32 r1 = reg2hex[b1];
116
117 if (!jit->seen_reg[r1] && r1 >= 6 && r1 <= 15)
118 jit->seen_reg[r1] = 1;
119 }
120
121 #define REG_SET_SEEN(b1) \
122 ({ \
123 reg_set_seen(jit, b1); \
124 })
125
126 #define REG_SEEN(b1) jit->seen_reg[reg2hex[(b1)]]
127
128 /*
129 * EMIT macros for code generation
130 */
131
132 #define _EMIT2(op) \
133 ({ \
134 if (jit->prg_buf) \
135 *(u16 *) (jit->prg_buf + jit->prg) = op; \
136 jit->prg += 2; \
137 })
138
139 #define EMIT2(op, b1, b2) \
140 ({ \
141 _EMIT2(op | reg(b1, b2)); \
142 REG_SET_SEEN(b1); \
143 REG_SET_SEEN(b2); \
144 })
145
146 #define _EMIT4(op) \
147 ({ \
148 if (jit->prg_buf) \
149 *(u32 *) (jit->prg_buf + jit->prg) = op; \
150 jit->prg += 4; \
151 })
152
153 #define EMIT4(op, b1, b2) \
154 ({ \
155 _EMIT4(op | reg(b1, b2)); \
156 REG_SET_SEEN(b1); \
157 REG_SET_SEEN(b2); \
158 })
159
160 #define EMIT4_RRF(op, b1, b2, b3) \
161 ({ \
162 _EMIT4(op | reg_high(b3) << 8 | reg(b1, b2)); \
163 REG_SET_SEEN(b1); \
164 REG_SET_SEEN(b2); \
165 REG_SET_SEEN(b3); \
166 })
167
168 #define _EMIT4_DISP(op, disp) \
169 ({ \
170 unsigned int __disp = (disp) & 0xfff; \
171 _EMIT4(op | __disp); \
172 })
173
174 #define EMIT4_DISP(op, b1, b2, disp) \
175 ({ \
176 _EMIT4_DISP(op | reg_high(b1) << 16 | \
177 reg_high(b2) << 8, disp); \
178 REG_SET_SEEN(b1); \
179 REG_SET_SEEN(b2); \
180 })
181
182 #define EMIT4_IMM(op, b1, imm) \
183 ({ \
184 unsigned int __imm = (imm) & 0xffff; \
185 _EMIT4(op | reg_high(b1) << 16 | __imm); \
186 REG_SET_SEEN(b1); \
187 })
188
189 #define EMIT4_PCREL(op, pcrel) \
190 ({ \
191 long __pcrel = ((pcrel) >> 1) & 0xffff; \
192 _EMIT4(op | __pcrel); \
193 })
194
195 #define _EMIT6(op1, op2) \
196 ({ \
197 if (jit->prg_buf) { \
198 *(u32 *) (jit->prg_buf + jit->prg) = op1; \
199 *(u16 *) (jit->prg_buf + jit->prg + 4) = op2; \
200 } \
201 jit->prg += 6; \
202 })
203
204 #define _EMIT6_DISP(op1, op2, disp) \
205 ({ \
206 unsigned int __disp = (disp) & 0xfff; \
207 _EMIT6(op1 | __disp, op2); \
208 })
209
210 #define _EMIT6_DISP_LH(op1, op2, disp) \
211 ({ \
212 u32 _disp = (u32) disp; \
213 unsigned int __disp_h = _disp & 0xff000; \
214 unsigned int __disp_l = _disp & 0x00fff; \
215 _EMIT6(op1 | __disp_l, op2 | __disp_h >> 4); \
216 })
217
218 #define EMIT6_DISP_LH(op1, op2, b1, b2, b3, disp) \
219 ({ \
220 _EMIT6_DISP_LH(op1 | reg(b1, b2) << 16 | \
221 reg_high(b3) << 8, op2, disp); \
222 REG_SET_SEEN(b1); \
223 REG_SET_SEEN(b2); \
224 REG_SET_SEEN(b3); \
225 })
226
227 #define EMIT6_PCREL_LABEL(op1, op2, b1, b2, label, mask) \
228 ({ \
229 int rel = (jit->labels[label] - jit->prg) >> 1; \
230 _EMIT6(op1 | reg(b1, b2) << 16 | (rel & 0xffff), \
231 op2 | mask << 12); \
232 REG_SET_SEEN(b1); \
233 REG_SET_SEEN(b2); \
234 })
235
236 #define EMIT6_PCREL_IMM_LABEL(op1, op2, b1, imm, label, mask) \
237 ({ \
238 int rel = (jit->labels[label] - jit->prg) >> 1; \
239 _EMIT6(op1 | (reg_high(b1) | mask) << 16 | \
240 (rel & 0xffff), op2 | (imm & 0xff) << 8); \
241 REG_SET_SEEN(b1); \
242 BUILD_BUG_ON(((unsigned long) imm) > 0xff); \
243 })
244
245 #define EMIT6_PCREL(op1, op2, b1, b2, i, off, mask) \
246 ({ \
247 /* Branch instruction needs 6 bytes */ \
248 int rel = (addrs[i + off + 1] - (addrs[i + 1] - 6)) / 2;\
249 _EMIT6(op1 | reg(b1, b2) << 16 | (rel & 0xffff), op2 | mask); \
250 REG_SET_SEEN(b1); \
251 REG_SET_SEEN(b2); \
252 })
253
254 #define _EMIT6_IMM(op, imm) \
255 ({ \
256 unsigned int __imm = (imm); \
257 _EMIT6(op | (__imm >> 16), __imm & 0xffff); \
258 })
259
260 #define EMIT6_IMM(op, b1, imm) \
261 ({ \
262 _EMIT6_IMM(op | reg_high(b1) << 16, imm); \
263 REG_SET_SEEN(b1); \
264 })
265
266 #define EMIT_CONST_U32(val) \
267 ({ \
268 unsigned int ret; \
269 ret = jit->lit - jit->base_ip; \
270 jit->seen |= SEEN_LITERAL; \
271 if (jit->prg_buf) \
272 *(u32 *) (jit->prg_buf + jit->lit) = (u32) val; \
273 jit->lit += 4; \
274 ret; \
275 })
276
277 #define EMIT_CONST_U64(val) \
278 ({ \
279 unsigned int ret; \
280 ret = jit->lit - jit->base_ip; \
281 jit->seen |= SEEN_LITERAL; \
282 if (jit->prg_buf) \
283 *(u64 *) (jit->prg_buf + jit->lit) = (u64) val; \
284 jit->lit += 8; \
285 ret; \
286 })
287
288 #define EMIT_ZERO(b1) \
289 ({ \
290 /* llgfr %dst,%dst (zero extend to 64 bit) */ \
291 EMIT4(0xb9160000, b1, b1); \
292 REG_SET_SEEN(b1); \
293 })
294
295 /*
296 * Fill whole space with illegal instructions
297 */
298 static void jit_fill_hole(void *area, unsigned int size)
299 {
300 memset(area, 0, size);
301 }
302
303 /*
304 * Save registers from "rs" (register start) to "re" (register end) on stack
305 */
306 static void save_regs(struct bpf_jit *jit, u32 rs, u32 re)
307 {
308 u32 off = STK_OFF_R6 + (rs - 6) * 8;
309
310 if (rs == re)
311 /* stg %rs,off(%r15) */
312 _EMIT6(0xe300f000 | rs << 20 | off, 0x0024);
313 else
314 /* stmg %rs,%re,off(%r15) */
315 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0024, off);
316 }
317
318 /*
319 * Restore registers from "rs" (register start) to "re" (register end) on stack
320 */
321 static void restore_regs(struct bpf_jit *jit, u32 rs, u32 re)
322 {
323 u32 off = STK_OFF_R6 + (rs - 6) * 8;
324
325 if (jit->seen & SEEN_STACK)
326 off += STK_OFF;
327
328 if (rs == re)
329 /* lg %rs,off(%r15) */
330 _EMIT6(0xe300f000 | rs << 20 | off, 0x0004);
331 else
332 /* lmg %rs,%re,off(%r15) */
333 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0004, off);
334 }
335
336 /*
337 * Return first seen register (from start)
338 */
339 static int get_start(struct bpf_jit *jit, int start)
340 {
341 int i;
342
343 for (i = start; i <= 15; i++) {
344 if (jit->seen_reg[i])
345 return i;
346 }
347 return 0;
348 }
349
350 /*
351 * Return last seen register (from start) (gap >= 2)
352 */
353 static int get_end(struct bpf_jit *jit, int start)
354 {
355 int i;
356
357 for (i = start; i < 15; i++) {
358 if (!jit->seen_reg[i] && !jit->seen_reg[i + 1])
359 return i - 1;
360 }
361 return jit->seen_reg[15] ? 15 : 14;
362 }
363
364 #define REGS_SAVE 1
365 #define REGS_RESTORE 0
366 /*
367 * Save and restore clobbered registers (6-15) on stack.
368 * We save/restore registers in chunks with gap >= 2 registers.
369 */
370 static void save_restore_regs(struct bpf_jit *jit, int op)
371 {
372
373 int re = 6, rs;
374
375 do {
376 rs = get_start(jit, re);
377 if (!rs)
378 break;
379 re = get_end(jit, rs + 1);
380 if (op == REGS_SAVE)
381 save_regs(jit, rs, re);
382 else
383 restore_regs(jit, rs, re);
384 re++;
385 } while (re <= 15);
386 }
387
388 /*
389 * For SKB access %b1 contains the SKB pointer. For "bpf_jit.S"
390 * we store the SKB header length on the stack and the SKB data
391 * pointer in REG_SKB_DATA if BPF_REG_AX is not used.
392 */
393 static void emit_load_skb_data_hlen(struct bpf_jit *jit)
394 {
395 /* Header length: llgf %w1,<len>(%b1) */
396 EMIT6_DISP_LH(0xe3000000, 0x0016, REG_W1, REG_0, BPF_REG_1,
397 offsetof(struct sk_buff, len));
398 /* s %w1,<data_len>(%b1) */
399 EMIT4_DISP(0x5b000000, REG_W1, BPF_REG_1,
400 offsetof(struct sk_buff, data_len));
401 /* stg %w1,ST_OFF_HLEN(%r0,%r15) */
402 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W1, REG_0, REG_15, STK_OFF_HLEN);
403 if (!(jit->seen & SEEN_REG_AX))
404 /* lg %skb_data,data_off(%b1) */
405 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_SKB_DATA, REG_0,
406 BPF_REG_1, offsetof(struct sk_buff, data));
407 }
408
409 /*
410 * Emit function prologue
411 *
412 * Save registers and create stack frame if necessary.
413 * See stack frame layout desription in "bpf_jit.h"!
414 */
415 static void bpf_jit_prologue(struct bpf_jit *jit)
416 {
417 if (jit->seen & SEEN_TAIL_CALL) {
418 /* xc STK_OFF_TCCNT(4,%r15),STK_OFF_TCCNT(%r15) */
419 _EMIT6(0xd703f000 | STK_OFF_TCCNT, 0xf000 | STK_OFF_TCCNT);
420 } else {
421 /* j tail_call_start: NOP if no tail calls are used */
422 EMIT4_PCREL(0xa7f40000, 6);
423 _EMIT2(0);
424 }
425 /* Tail calls have to skip above initialization */
426 jit->tail_call_start = jit->prg;
427 /* Save registers */
428 save_restore_regs(jit, REGS_SAVE);
429 /* Setup literal pool */
430 if (jit->seen & SEEN_LITERAL) {
431 /* basr %r13,0 */
432 EMIT2(0x0d00, REG_L, REG_0);
433 jit->base_ip = jit->prg;
434 }
435 /* Setup stack and backchain */
436 if (jit->seen & SEEN_STACK) {
437 if (jit->seen & SEEN_FUNC)
438 /* lgr %w1,%r15 (backchain) */
439 EMIT4(0xb9040000, REG_W1, REG_15);
440 /* la %bfp,STK_160_UNUSED(%r15) (BPF frame pointer) */
441 EMIT4_DISP(0x41000000, BPF_REG_FP, REG_15, STK_160_UNUSED);
442 /* aghi %r15,-STK_OFF */
443 EMIT4_IMM(0xa70b0000, REG_15, -STK_OFF);
444 if (jit->seen & SEEN_FUNC)
445 /* stg %w1,152(%r15) (backchain) */
446 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W1, REG_0,
447 REG_15, 152);
448 }
449 if (jit->seen & SEEN_SKB)
450 emit_load_skb_data_hlen(jit);
451 if (jit->seen & SEEN_SKB_CHANGE)
452 /* stg %b1,ST_OFF_SKBP(%r0,%r15) */
453 EMIT6_DISP_LH(0xe3000000, 0x0024, BPF_REG_1, REG_0, REG_15,
454 STK_OFF_SKBP);
455 }
456
457 /*
458 * Function epilogue
459 */
460 static void bpf_jit_epilogue(struct bpf_jit *jit)
461 {
462 /* Return 0 */
463 if (jit->seen & SEEN_RET0) {
464 jit->ret0_ip = jit->prg;
465 /* lghi %b0,0 */
466 EMIT4_IMM(0xa7090000, BPF_REG_0, 0);
467 }
468 jit->exit_ip = jit->prg;
469 /* Load exit code: lgr %r2,%b0 */
470 EMIT4(0xb9040000, REG_2, BPF_REG_0);
471 /* Restore registers */
472 save_restore_regs(jit, REGS_RESTORE);
473 /* br %r14 */
474 _EMIT2(0x07fe);
475 }
476
477 /*
478 * Compile one eBPF instruction into s390x code
479 *
480 * NOTE: Use noinline because for gcov (-fprofile-arcs) gcc allocates a lot of
481 * stack space for the large switch statement.
482 */
483 static noinline int bpf_jit_insn(struct bpf_jit *jit, struct bpf_prog *fp, int i)
484 {
485 struct bpf_insn *insn = &fp->insnsi[i];
486 int jmp_off, last, insn_count = 1;
487 unsigned int func_addr, mask;
488 u32 dst_reg = insn->dst_reg;
489 u32 src_reg = insn->src_reg;
490 u32 *addrs = jit->addrs;
491 s32 imm = insn->imm;
492 s16 off = insn->off;
493
494 if (dst_reg == BPF_REG_AX || src_reg == BPF_REG_AX)
495 jit->seen |= SEEN_REG_AX;
496 switch (insn->code) {
497 /*
498 * BPF_MOV
499 */
500 case BPF_ALU | BPF_MOV | BPF_X: /* dst = (u32) src */
501 /* llgfr %dst,%src */
502 EMIT4(0xb9160000, dst_reg, src_reg);
503 break;
504 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
505 /* lgr %dst,%src */
506 EMIT4(0xb9040000, dst_reg, src_reg);
507 break;
508 case BPF_ALU | BPF_MOV | BPF_K: /* dst = (u32) imm */
509 /* llilf %dst,imm */
510 EMIT6_IMM(0xc00f0000, dst_reg, imm);
511 break;
512 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = imm */
513 /* lgfi %dst,imm */
514 EMIT6_IMM(0xc0010000, dst_reg, imm);
515 break;
516 /*
517 * BPF_LD 64
518 */
519 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
520 {
521 /* 16 byte instruction that uses two 'struct bpf_insn' */
522 u64 imm64;
523
524 imm64 = (u64)(u32) insn[0].imm | ((u64)(u32) insn[1].imm) << 32;
525 /* lg %dst,<d(imm)>(%l) */
526 EMIT6_DISP_LH(0xe3000000, 0x0004, dst_reg, REG_0, REG_L,
527 EMIT_CONST_U64(imm64));
528 insn_count = 2;
529 break;
530 }
531 /*
532 * BPF_ADD
533 */
534 case BPF_ALU | BPF_ADD | BPF_X: /* dst = (u32) dst + (u32) src */
535 /* ar %dst,%src */
536 EMIT2(0x1a00, dst_reg, src_reg);
537 EMIT_ZERO(dst_reg);
538 break;
539 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst = dst + src */
540 /* agr %dst,%src */
541 EMIT4(0xb9080000, dst_reg, src_reg);
542 break;
543 case BPF_ALU | BPF_ADD | BPF_K: /* dst = (u32) dst + (u32) imm */
544 if (!imm)
545 break;
546 /* alfi %dst,imm */
547 EMIT6_IMM(0xc20b0000, dst_reg, imm);
548 EMIT_ZERO(dst_reg);
549 break;
550 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst = dst + imm */
551 if (!imm)
552 break;
553 /* agfi %dst,imm */
554 EMIT6_IMM(0xc2080000, dst_reg, imm);
555 break;
556 /*
557 * BPF_SUB
558 */
559 case BPF_ALU | BPF_SUB | BPF_X: /* dst = (u32) dst - (u32) src */
560 /* sr %dst,%src */
561 EMIT2(0x1b00, dst_reg, src_reg);
562 EMIT_ZERO(dst_reg);
563 break;
564 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst = dst - src */
565 /* sgr %dst,%src */
566 EMIT4(0xb9090000, dst_reg, src_reg);
567 break;
568 case BPF_ALU | BPF_SUB | BPF_K: /* dst = (u32) dst - (u32) imm */
569 if (!imm)
570 break;
571 /* alfi %dst,-imm */
572 EMIT6_IMM(0xc20b0000, dst_reg, -imm);
573 EMIT_ZERO(dst_reg);
574 break;
575 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst = dst - imm */
576 if (!imm)
577 break;
578 /* agfi %dst,-imm */
579 EMIT6_IMM(0xc2080000, dst_reg, -imm);
580 break;
581 /*
582 * BPF_MUL
583 */
584 case BPF_ALU | BPF_MUL | BPF_X: /* dst = (u32) dst * (u32) src */
585 /* msr %dst,%src */
586 EMIT4(0xb2520000, dst_reg, src_reg);
587 EMIT_ZERO(dst_reg);
588 break;
589 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst = dst * src */
590 /* msgr %dst,%src */
591 EMIT4(0xb90c0000, dst_reg, src_reg);
592 break;
593 case BPF_ALU | BPF_MUL | BPF_K: /* dst = (u32) dst * (u32) imm */
594 if (imm == 1)
595 break;
596 /* msfi %r5,imm */
597 EMIT6_IMM(0xc2010000, dst_reg, imm);
598 EMIT_ZERO(dst_reg);
599 break;
600 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst = dst * imm */
601 if (imm == 1)
602 break;
603 /* msgfi %dst,imm */
604 EMIT6_IMM(0xc2000000, dst_reg, imm);
605 break;
606 /*
607 * BPF_DIV / BPF_MOD
608 */
609 case BPF_ALU | BPF_DIV | BPF_X: /* dst = (u32) dst / (u32) src */
610 case BPF_ALU | BPF_MOD | BPF_X: /* dst = (u32) dst % (u32) src */
611 {
612 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
613
614 jit->seen |= SEEN_RET0;
615 /* ltr %src,%src (if src == 0 goto fail) */
616 EMIT2(0x1200, src_reg, src_reg);
617 /* jz <ret0> */
618 EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg);
619 /* lhi %w0,0 */
620 EMIT4_IMM(0xa7080000, REG_W0, 0);
621 /* lr %w1,%dst */
622 EMIT2(0x1800, REG_W1, dst_reg);
623 /* dlr %w0,%src */
624 EMIT4(0xb9970000, REG_W0, src_reg);
625 /* llgfr %dst,%rc */
626 EMIT4(0xb9160000, dst_reg, rc_reg);
627 break;
628 }
629 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst = dst / src */
630 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst = dst % src */
631 {
632 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
633
634 jit->seen |= SEEN_RET0;
635 /* ltgr %src,%src (if src == 0 goto fail) */
636 EMIT4(0xb9020000, src_reg, src_reg);
637 /* jz <ret0> */
638 EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg);
639 /* lghi %w0,0 */
640 EMIT4_IMM(0xa7090000, REG_W0, 0);
641 /* lgr %w1,%dst */
642 EMIT4(0xb9040000, REG_W1, dst_reg);
643 /* dlgr %w0,%dst */
644 EMIT4(0xb9870000, REG_W0, src_reg);
645 /* lgr %dst,%rc */
646 EMIT4(0xb9040000, dst_reg, rc_reg);
647 break;
648 }
649 case BPF_ALU | BPF_DIV | BPF_K: /* dst = (u32) dst / (u32) imm */
650 case BPF_ALU | BPF_MOD | BPF_K: /* dst = (u32) dst % (u32) imm */
651 {
652 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
653
654 if (imm == 1) {
655 if (BPF_OP(insn->code) == BPF_MOD)
656 /* lhgi %dst,0 */
657 EMIT4_IMM(0xa7090000, dst_reg, 0);
658 break;
659 }
660 /* lhi %w0,0 */
661 EMIT4_IMM(0xa7080000, REG_W0, 0);
662 /* lr %w1,%dst */
663 EMIT2(0x1800, REG_W1, dst_reg);
664 /* dl %w0,<d(imm)>(%l) */
665 EMIT6_DISP_LH(0xe3000000, 0x0097, REG_W0, REG_0, REG_L,
666 EMIT_CONST_U32(imm));
667 /* llgfr %dst,%rc */
668 EMIT4(0xb9160000, dst_reg, rc_reg);
669 break;
670 }
671 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst = dst / imm */
672 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst = dst % imm */
673 {
674 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
675
676 if (imm == 1) {
677 if (BPF_OP(insn->code) == BPF_MOD)
678 /* lhgi %dst,0 */
679 EMIT4_IMM(0xa7090000, dst_reg, 0);
680 break;
681 }
682 /* lghi %w0,0 */
683 EMIT4_IMM(0xa7090000, REG_W0, 0);
684 /* lgr %w1,%dst */
685 EMIT4(0xb9040000, REG_W1, dst_reg);
686 /* dlg %w0,<d(imm)>(%l) */
687 EMIT6_DISP_LH(0xe3000000, 0x0087, REG_W0, REG_0, REG_L,
688 EMIT_CONST_U64(imm));
689 /* lgr %dst,%rc */
690 EMIT4(0xb9040000, dst_reg, rc_reg);
691 break;
692 }
693 /*
694 * BPF_AND
695 */
696 case BPF_ALU | BPF_AND | BPF_X: /* dst = (u32) dst & (u32) src */
697 /* nr %dst,%src */
698 EMIT2(0x1400, dst_reg, src_reg);
699 EMIT_ZERO(dst_reg);
700 break;
701 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
702 /* ngr %dst,%src */
703 EMIT4(0xb9800000, dst_reg, src_reg);
704 break;
705 case BPF_ALU | BPF_AND | BPF_K: /* dst = (u32) dst & (u32) imm */
706 /* nilf %dst,imm */
707 EMIT6_IMM(0xc00b0000, dst_reg, imm);
708 EMIT_ZERO(dst_reg);
709 break;
710 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
711 /* ng %dst,<d(imm)>(%l) */
712 EMIT6_DISP_LH(0xe3000000, 0x0080, dst_reg, REG_0, REG_L,
713 EMIT_CONST_U64(imm));
714 break;
715 /*
716 * BPF_OR
717 */
718 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
719 /* or %dst,%src */
720 EMIT2(0x1600, dst_reg, src_reg);
721 EMIT_ZERO(dst_reg);
722 break;
723 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
724 /* ogr %dst,%src */
725 EMIT4(0xb9810000, dst_reg, src_reg);
726 break;
727 case BPF_ALU | BPF_OR | BPF_K: /* dst = (u32) dst | (u32) imm */
728 /* oilf %dst,imm */
729 EMIT6_IMM(0xc00d0000, dst_reg, imm);
730 EMIT_ZERO(dst_reg);
731 break;
732 case BPF_ALU64 | BPF_OR | BPF_K: /* dst = dst | imm */
733 /* og %dst,<d(imm)>(%l) */
734 EMIT6_DISP_LH(0xe3000000, 0x0081, dst_reg, REG_0, REG_L,
735 EMIT_CONST_U64(imm));
736 break;
737 /*
738 * BPF_XOR
739 */
740 case BPF_ALU | BPF_XOR | BPF_X: /* dst = (u32) dst ^ (u32) src */
741 /* xr %dst,%src */
742 EMIT2(0x1700, dst_reg, src_reg);
743 EMIT_ZERO(dst_reg);
744 break;
745 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst = dst ^ src */
746 /* xgr %dst,%src */
747 EMIT4(0xb9820000, dst_reg, src_reg);
748 break;
749 case BPF_ALU | BPF_XOR | BPF_K: /* dst = (u32) dst ^ (u32) imm */
750 if (!imm)
751 break;
752 /* xilf %dst,imm */
753 EMIT6_IMM(0xc0070000, dst_reg, imm);
754 EMIT_ZERO(dst_reg);
755 break;
756 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst = dst ^ imm */
757 /* xg %dst,<d(imm)>(%l) */
758 EMIT6_DISP_LH(0xe3000000, 0x0082, dst_reg, REG_0, REG_L,
759 EMIT_CONST_U64(imm));
760 break;
761 /*
762 * BPF_LSH
763 */
764 case BPF_ALU | BPF_LSH | BPF_X: /* dst = (u32) dst << (u32) src */
765 /* sll %dst,0(%src) */
766 EMIT4_DISP(0x89000000, dst_reg, src_reg, 0);
767 EMIT_ZERO(dst_reg);
768 break;
769 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst = dst << src */
770 /* sllg %dst,%dst,0(%src) */
771 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, src_reg, 0);
772 break;
773 case BPF_ALU | BPF_LSH | BPF_K: /* dst = (u32) dst << (u32) imm */
774 if (imm == 0)
775 break;
776 /* sll %dst,imm(%r0) */
777 EMIT4_DISP(0x89000000, dst_reg, REG_0, imm);
778 EMIT_ZERO(dst_reg);
779 break;
780 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst = dst << imm */
781 if (imm == 0)
782 break;
783 /* sllg %dst,%dst,imm(%r0) */
784 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, REG_0, imm);
785 break;
786 /*
787 * BPF_RSH
788 */
789 case BPF_ALU | BPF_RSH | BPF_X: /* dst = (u32) dst >> (u32) src */
790 /* srl %dst,0(%src) */
791 EMIT4_DISP(0x88000000, dst_reg, src_reg, 0);
792 EMIT_ZERO(dst_reg);
793 break;
794 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst = dst >> src */
795 /* srlg %dst,%dst,0(%src) */
796 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, src_reg, 0);
797 break;
798 case BPF_ALU | BPF_RSH | BPF_K: /* dst = (u32) dst >> (u32) imm */
799 if (imm == 0)
800 break;
801 /* srl %dst,imm(%r0) */
802 EMIT4_DISP(0x88000000, dst_reg, REG_0, imm);
803 EMIT_ZERO(dst_reg);
804 break;
805 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst = dst >> imm */
806 if (imm == 0)
807 break;
808 /* srlg %dst,%dst,imm(%r0) */
809 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, REG_0, imm);
810 break;
811 /*
812 * BPF_ARSH
813 */
814 case BPF_ALU64 | BPF_ARSH | BPF_X: /* ((s64) dst) >>= src */
815 /* srag %dst,%dst,0(%src) */
816 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, src_reg, 0);
817 break;
818 case BPF_ALU64 | BPF_ARSH | BPF_K: /* ((s64) dst) >>= imm */
819 if (imm == 0)
820 break;
821 /* srag %dst,%dst,imm(%r0) */
822 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, REG_0, imm);
823 break;
824 /*
825 * BPF_NEG
826 */
827 case BPF_ALU | BPF_NEG: /* dst = (u32) -dst */
828 /* lcr %dst,%dst */
829 EMIT2(0x1300, dst_reg, dst_reg);
830 EMIT_ZERO(dst_reg);
831 break;
832 case BPF_ALU64 | BPF_NEG: /* dst = -dst */
833 /* lcgr %dst,%dst */
834 EMIT4(0xb9130000, dst_reg, dst_reg);
835 break;
836 /*
837 * BPF_FROM_BE/LE
838 */
839 case BPF_ALU | BPF_END | BPF_FROM_BE:
840 /* s390 is big endian, therefore only clear high order bytes */
841 switch (imm) {
842 case 16: /* dst = (u16) cpu_to_be16(dst) */
843 /* llghr %dst,%dst */
844 EMIT4(0xb9850000, dst_reg, dst_reg);
845 break;
846 case 32: /* dst = (u32) cpu_to_be32(dst) */
847 /* llgfr %dst,%dst */
848 EMIT4(0xb9160000, dst_reg, dst_reg);
849 break;
850 case 64: /* dst = (u64) cpu_to_be64(dst) */
851 break;
852 }
853 break;
854 case BPF_ALU | BPF_END | BPF_FROM_LE:
855 switch (imm) {
856 case 16: /* dst = (u16) cpu_to_le16(dst) */
857 /* lrvr %dst,%dst */
858 EMIT4(0xb91f0000, dst_reg, dst_reg);
859 /* srl %dst,16(%r0) */
860 EMIT4_DISP(0x88000000, dst_reg, REG_0, 16);
861 /* llghr %dst,%dst */
862 EMIT4(0xb9850000, dst_reg, dst_reg);
863 break;
864 case 32: /* dst = (u32) cpu_to_le32(dst) */
865 /* lrvr %dst,%dst */
866 EMIT4(0xb91f0000, dst_reg, dst_reg);
867 /* llgfr %dst,%dst */
868 EMIT4(0xb9160000, dst_reg, dst_reg);
869 break;
870 case 64: /* dst = (u64) cpu_to_le64(dst) */
871 /* lrvgr %dst,%dst */
872 EMIT4(0xb90f0000, dst_reg, dst_reg);
873 break;
874 }
875 break;
876 /*
877 * BPF_ST(X)
878 */
879 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src_reg */
880 /* stcy %src,off(%dst) */
881 EMIT6_DISP_LH(0xe3000000, 0x0072, src_reg, dst_reg, REG_0, off);
882 jit->seen |= SEEN_MEM;
883 break;
884 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
885 /* sthy %src,off(%dst) */
886 EMIT6_DISP_LH(0xe3000000, 0x0070, src_reg, dst_reg, REG_0, off);
887 jit->seen |= SEEN_MEM;
888 break;
889 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
890 /* sty %src,off(%dst) */
891 EMIT6_DISP_LH(0xe3000000, 0x0050, src_reg, dst_reg, REG_0, off);
892 jit->seen |= SEEN_MEM;
893 break;
894 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
895 /* stg %src,off(%dst) */
896 EMIT6_DISP_LH(0xe3000000, 0x0024, src_reg, dst_reg, REG_0, off);
897 jit->seen |= SEEN_MEM;
898 break;
899 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
900 /* lhi %w0,imm */
901 EMIT4_IMM(0xa7080000, REG_W0, (u8) imm);
902 /* stcy %w0,off(dst) */
903 EMIT6_DISP_LH(0xe3000000, 0x0072, REG_W0, dst_reg, REG_0, off);
904 jit->seen |= SEEN_MEM;
905 break;
906 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
907 /* lhi %w0,imm */
908 EMIT4_IMM(0xa7080000, REG_W0, (u16) imm);
909 /* sthy %w0,off(dst) */
910 EMIT6_DISP_LH(0xe3000000, 0x0070, REG_W0, dst_reg, REG_0, off);
911 jit->seen |= SEEN_MEM;
912 break;
913 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
914 /* llilf %w0,imm */
915 EMIT6_IMM(0xc00f0000, REG_W0, (u32) imm);
916 /* sty %w0,off(%dst) */
917 EMIT6_DISP_LH(0xe3000000, 0x0050, REG_W0, dst_reg, REG_0, off);
918 jit->seen |= SEEN_MEM;
919 break;
920 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
921 /* lgfi %w0,imm */
922 EMIT6_IMM(0xc0010000, REG_W0, imm);
923 /* stg %w0,off(%dst) */
924 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W0, dst_reg, REG_0, off);
925 jit->seen |= SEEN_MEM;
926 break;
927 /*
928 * BPF_STX XADD (atomic_add)
929 */
930 case BPF_STX | BPF_XADD | BPF_W: /* *(u32 *)(dst + off) += src */
931 /* laal %w0,%src,off(%dst) */
932 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W0, src_reg,
933 dst_reg, off);
934 jit->seen |= SEEN_MEM;
935 break;
936 case BPF_STX | BPF_XADD | BPF_DW: /* *(u64 *)(dst + off) += src */
937 /* laalg %w0,%src,off(%dst) */
938 EMIT6_DISP_LH(0xeb000000, 0x00ea, REG_W0, src_reg,
939 dst_reg, off);
940 jit->seen |= SEEN_MEM;
941 break;
942 /*
943 * BPF_LDX
944 */
945 case BPF_LDX | BPF_MEM | BPF_B: /* dst = *(u8 *)(ul) (src + off) */
946 /* llgc %dst,0(off,%src) */
947 EMIT6_DISP_LH(0xe3000000, 0x0090, dst_reg, src_reg, REG_0, off);
948 jit->seen |= SEEN_MEM;
949 break;
950 case BPF_LDX | BPF_MEM | BPF_H: /* dst = *(u16 *)(ul) (src + off) */
951 /* llgh %dst,0(off,%src) */
952 EMIT6_DISP_LH(0xe3000000, 0x0091, dst_reg, src_reg, REG_0, off);
953 jit->seen |= SEEN_MEM;
954 break;
955 case BPF_LDX | BPF_MEM | BPF_W: /* dst = *(u32 *)(ul) (src + off) */
956 /* llgf %dst,off(%src) */
957 jit->seen |= SEEN_MEM;
958 EMIT6_DISP_LH(0xe3000000, 0x0016, dst_reg, src_reg, REG_0, off);
959 break;
960 case BPF_LDX | BPF_MEM | BPF_DW: /* dst = *(u64 *)(ul) (src + off) */
961 /* lg %dst,0(off,%src) */
962 jit->seen |= SEEN_MEM;
963 EMIT6_DISP_LH(0xe3000000, 0x0004, dst_reg, src_reg, REG_0, off);
964 break;
965 /*
966 * BPF_JMP / CALL
967 */
968 case BPF_JMP | BPF_CALL:
969 {
970 /*
971 * b0 = (__bpf_call_base + imm)(b1, b2, b3, b4, b5)
972 */
973 const u64 func = (u64)__bpf_call_base + imm;
974
975 REG_SET_SEEN(BPF_REG_5);
976 jit->seen |= SEEN_FUNC;
977 /* lg %w1,<d(imm)>(%l) */
978 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_W1, REG_0, REG_L,
979 EMIT_CONST_U64(func));
980 /* basr %r14,%w1 */
981 EMIT2(0x0d00, REG_14, REG_W1);
982 /* lgr %b0,%r2: load return value into %b0 */
983 EMIT4(0xb9040000, BPF_REG_0, REG_2);
984 if (bpf_helper_changes_skb_data((void *)func)) {
985 jit->seen |= SEEN_SKB_CHANGE;
986 /* lg %b1,ST_OFF_SKBP(%r15) */
987 EMIT6_DISP_LH(0xe3000000, 0x0004, BPF_REG_1, REG_0,
988 REG_15, STK_OFF_SKBP);
989 emit_load_skb_data_hlen(jit);
990 }
991 break;
992 }
993 case BPF_JMP | BPF_CALL | BPF_X:
994 /*
995 * Implicit input:
996 * B1: pointer to ctx
997 * B2: pointer to bpf_array
998 * B3: index in bpf_array
999 */
1000 jit->seen |= SEEN_TAIL_CALL;
1001
1002 /*
1003 * if (index >= array->map.max_entries)
1004 * goto out;
1005 */
1006
1007 /* llgf %w1,map.max_entries(%b2) */
1008 EMIT6_DISP_LH(0xe3000000, 0x0016, REG_W1, REG_0, BPF_REG_2,
1009 offsetof(struct bpf_array, map.max_entries));
1010 /* clgrj %b3,%w1,0xa,label0: if %b3 >= %w1 goto out */
1011 EMIT6_PCREL_LABEL(0xec000000, 0x0065, BPF_REG_3,
1012 REG_W1, 0, 0xa);
1013
1014 /*
1015 * if (tail_call_cnt++ > MAX_TAIL_CALL_CNT)
1016 * goto out;
1017 */
1018
1019 if (jit->seen & SEEN_STACK)
1020 off = STK_OFF_TCCNT + STK_OFF;
1021 else
1022 off = STK_OFF_TCCNT;
1023 /* lhi %w0,1 */
1024 EMIT4_IMM(0xa7080000, REG_W0, 1);
1025 /* laal %w1,%w0,off(%r15) */
1026 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W1, REG_W0, REG_15, off);
1027 /* clij %w1,MAX_TAIL_CALL_CNT,0x2,label0 */
1028 EMIT6_PCREL_IMM_LABEL(0xec000000, 0x007f, REG_W1,
1029 MAX_TAIL_CALL_CNT, 0, 0x2);
1030
1031 /*
1032 * prog = array->ptrs[index];
1033 * if (prog == NULL)
1034 * goto out;
1035 */
1036
1037 /* sllg %r1,%b3,3: %r1 = index * 8 */
1038 EMIT6_DISP_LH(0xeb000000, 0x000d, REG_1, BPF_REG_3, REG_0, 3);
1039 /* lg %r1,prog(%b2,%r1) */
1040 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, BPF_REG_2,
1041 REG_1, offsetof(struct bpf_array, ptrs));
1042 /* clgij %r1,0,0x8,label0 */
1043 EMIT6_PCREL_IMM_LABEL(0xec000000, 0x007d, REG_1, 0, 0, 0x8);
1044
1045 /*
1046 * Restore registers before calling function
1047 */
1048 save_restore_regs(jit, REGS_RESTORE);
1049
1050 /*
1051 * goto *(prog->bpf_func + tail_call_start);
1052 */
1053
1054 /* lg %r1,bpf_func(%r1) */
1055 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, REG_1, REG_0,
1056 offsetof(struct bpf_prog, bpf_func));
1057 /* bc 0xf,tail_call_start(%r1) */
1058 _EMIT4(0x47f01000 + jit->tail_call_start);
1059 /* out: */
1060 jit->labels[0] = jit->prg;
1061 break;
1062 case BPF_JMP | BPF_EXIT: /* return b0 */
1063 last = (i == fp->len - 1) ? 1 : 0;
1064 if (last && !(jit->seen & SEEN_RET0))
1065 break;
1066 /* j <exit> */
1067 EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg);
1068 break;
1069 /*
1070 * Branch relative (number of skipped instructions) to offset on
1071 * condition.
1072 *
1073 * Condition code to mask mapping:
1074 *
1075 * CC | Description | Mask
1076 * ------------------------------
1077 * 0 | Operands equal | 8
1078 * 1 | First operand low | 4
1079 * 2 | First operand high | 2
1080 * 3 | Unused | 1
1081 *
1082 * For s390x relative branches: ip = ip + off_bytes
1083 * For BPF relative branches: insn = insn + off_insns + 1
1084 *
1085 * For example for s390x with offset 0 we jump to the branch
1086 * instruction itself (loop) and for BPF with offset 0 we
1087 * branch to the instruction behind the branch.
1088 */
1089 case BPF_JMP | BPF_JA: /* if (true) */
1090 mask = 0xf000; /* j */
1091 goto branch_oc;
1092 case BPF_JMP | BPF_JSGT | BPF_K: /* ((s64) dst > (s64) imm) */
1093 mask = 0x2000; /* jh */
1094 goto branch_ks;
1095 case BPF_JMP | BPF_JSGE | BPF_K: /* ((s64) dst >= (s64) imm) */
1096 mask = 0xa000; /* jhe */
1097 goto branch_ks;
1098 case BPF_JMP | BPF_JGT | BPF_K: /* (dst_reg > imm) */
1099 mask = 0x2000; /* jh */
1100 goto branch_ku;
1101 case BPF_JMP | BPF_JGE | BPF_K: /* (dst_reg >= imm) */
1102 mask = 0xa000; /* jhe */
1103 goto branch_ku;
1104 case BPF_JMP | BPF_JNE | BPF_K: /* (dst_reg != imm) */
1105 mask = 0x7000; /* jne */
1106 goto branch_ku;
1107 case BPF_JMP | BPF_JEQ | BPF_K: /* (dst_reg == imm) */
1108 mask = 0x8000; /* je */
1109 goto branch_ku;
1110 case BPF_JMP | BPF_JSET | BPF_K: /* (dst_reg & imm) */
1111 mask = 0x7000; /* jnz */
1112 /* lgfi %w1,imm (load sign extend imm) */
1113 EMIT6_IMM(0xc0010000, REG_W1, imm);
1114 /* ngr %w1,%dst */
1115 EMIT4(0xb9800000, REG_W1, dst_reg);
1116 goto branch_oc;
1117
1118 case BPF_JMP | BPF_JSGT | BPF_X: /* ((s64) dst > (s64) src) */
1119 mask = 0x2000; /* jh */
1120 goto branch_xs;
1121 case BPF_JMP | BPF_JSGE | BPF_X: /* ((s64) dst >= (s64) src) */
1122 mask = 0xa000; /* jhe */
1123 goto branch_xs;
1124 case BPF_JMP | BPF_JGT | BPF_X: /* (dst > src) */
1125 mask = 0x2000; /* jh */
1126 goto branch_xu;
1127 case BPF_JMP | BPF_JGE | BPF_X: /* (dst >= src) */
1128 mask = 0xa000; /* jhe */
1129 goto branch_xu;
1130 case BPF_JMP | BPF_JNE | BPF_X: /* (dst != src) */
1131 mask = 0x7000; /* jne */
1132 goto branch_xu;
1133 case BPF_JMP | BPF_JEQ | BPF_X: /* (dst == src) */
1134 mask = 0x8000; /* je */
1135 goto branch_xu;
1136 case BPF_JMP | BPF_JSET | BPF_X: /* (dst & src) */
1137 mask = 0x7000; /* jnz */
1138 /* ngrk %w1,%dst,%src */
1139 EMIT4_RRF(0xb9e40000, REG_W1, dst_reg, src_reg);
1140 goto branch_oc;
1141 branch_ks:
1142 /* lgfi %w1,imm (load sign extend imm) */
1143 EMIT6_IMM(0xc0010000, REG_W1, imm);
1144 /* cgrj %dst,%w1,mask,off */
1145 EMIT6_PCREL(0xec000000, 0x0064, dst_reg, REG_W1, i, off, mask);
1146 break;
1147 branch_ku:
1148 /* lgfi %w1,imm (load sign extend imm) */
1149 EMIT6_IMM(0xc0010000, REG_W1, imm);
1150 /* clgrj %dst,%w1,mask,off */
1151 EMIT6_PCREL(0xec000000, 0x0065, dst_reg, REG_W1, i, off, mask);
1152 break;
1153 branch_xs:
1154 /* cgrj %dst,%src,mask,off */
1155 EMIT6_PCREL(0xec000000, 0x0064, dst_reg, src_reg, i, off, mask);
1156 break;
1157 branch_xu:
1158 /* clgrj %dst,%src,mask,off */
1159 EMIT6_PCREL(0xec000000, 0x0065, dst_reg, src_reg, i, off, mask);
1160 break;
1161 branch_oc:
1162 /* brc mask,jmp_off (branch instruction needs 4 bytes) */
1163 jmp_off = addrs[i + off + 1] - (addrs[i + 1] - 4);
1164 EMIT4_PCREL(0xa7040000 | mask << 8, jmp_off);
1165 break;
1166 /*
1167 * BPF_LD
1168 */
1169 case BPF_LD | BPF_ABS | BPF_B: /* b0 = *(u8 *) (skb->data+imm) */
1170 case BPF_LD | BPF_IND | BPF_B: /* b0 = *(u8 *) (skb->data+imm+src) */
1171 if ((BPF_MODE(insn->code) == BPF_ABS) && (imm >= 0))
1172 func_addr = __pa(sk_load_byte_pos);
1173 else
1174 func_addr = __pa(sk_load_byte);
1175 goto call_fn;
1176 case BPF_LD | BPF_ABS | BPF_H: /* b0 = *(u16 *) (skb->data+imm) */
1177 case BPF_LD | BPF_IND | BPF_H: /* b0 = *(u16 *) (skb->data+imm+src) */
1178 if ((BPF_MODE(insn->code) == BPF_ABS) && (imm >= 0))
1179 func_addr = __pa(sk_load_half_pos);
1180 else
1181 func_addr = __pa(sk_load_half);
1182 goto call_fn;
1183 case BPF_LD | BPF_ABS | BPF_W: /* b0 = *(u32 *) (skb->data+imm) */
1184 case BPF_LD | BPF_IND | BPF_W: /* b0 = *(u32 *) (skb->data+imm+src) */
1185 if ((BPF_MODE(insn->code) == BPF_ABS) && (imm >= 0))
1186 func_addr = __pa(sk_load_word_pos);
1187 else
1188 func_addr = __pa(sk_load_word);
1189 goto call_fn;
1190 call_fn:
1191 jit->seen |= SEEN_SKB | SEEN_RET0 | SEEN_FUNC;
1192 REG_SET_SEEN(REG_14); /* Return address of possible func call */
1193
1194 /*
1195 * Implicit input:
1196 * BPF_REG_6 (R7) : skb pointer
1197 * REG_SKB_DATA (R12): skb data pointer (if no BPF_REG_AX)
1198 *
1199 * Calculated input:
1200 * BPF_REG_2 (R3) : offset of byte(s) to fetch in skb
1201 * BPF_REG_5 (R6) : return address
1202 *
1203 * Output:
1204 * BPF_REG_0 (R14): data read from skb
1205 *
1206 * Scratch registers (BPF_REG_1-5)
1207 */
1208
1209 /* Call function: llilf %w1,func_addr */
1210 EMIT6_IMM(0xc00f0000, REG_W1, func_addr);
1211
1212 /* Offset: lgfi %b2,imm */
1213 EMIT6_IMM(0xc0010000, BPF_REG_2, imm);
1214 if (BPF_MODE(insn->code) == BPF_IND)
1215 /* agfr %b2,%src (%src is s32 here) */
1216 EMIT4(0xb9180000, BPF_REG_2, src_reg);
1217
1218 /* Reload REG_SKB_DATA if BPF_REG_AX is used */
1219 if (jit->seen & SEEN_REG_AX)
1220 /* lg %skb_data,data_off(%b6) */
1221 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_SKB_DATA, REG_0,
1222 BPF_REG_6, offsetof(struct sk_buff, data));
1223 /* basr %b5,%w1 (%b5 is call saved) */
1224 EMIT2(0x0d00, BPF_REG_5, REG_W1);
1225
1226 /*
1227 * Note: For fast access we jump directly after the
1228 * jnz instruction from bpf_jit.S
1229 */
1230 /* jnz <ret0> */
1231 EMIT4_PCREL(0xa7740000, jit->ret0_ip - jit->prg);
1232 break;
1233 default: /* too complex, give up */
1234 pr_err("Unknown opcode %02x\n", insn->code);
1235 return -1;
1236 }
1237 return insn_count;
1238 }
1239
1240 /*
1241 * Compile eBPF program into s390x code
1242 */
1243 static int bpf_jit_prog(struct bpf_jit *jit, struct bpf_prog *fp)
1244 {
1245 int i, insn_count;
1246
1247 jit->lit = jit->lit_start;
1248 jit->prg = 0;
1249
1250 bpf_jit_prologue(jit);
1251 for (i = 0; i < fp->len; i += insn_count) {
1252 insn_count = bpf_jit_insn(jit, fp, i);
1253 if (insn_count < 0)
1254 return -1;
1255 jit->addrs[i + 1] = jit->prg; /* Next instruction address */
1256 }
1257 bpf_jit_epilogue(jit);
1258
1259 jit->lit_start = jit->prg;
1260 jit->size = jit->lit;
1261 jit->size_prg = jit->prg;
1262 return 0;
1263 }
1264
1265 /*
1266 * Classic BPF function stub. BPF programs will be converted into
1267 * eBPF and then bpf_int_jit_compile() will be called.
1268 */
1269 void bpf_jit_compile(struct bpf_prog *fp)
1270 {
1271 }
1272
1273 /*
1274 * Compile eBPF program "fp"
1275 */
1276 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
1277 {
1278 struct bpf_prog *tmp, *orig_fp = fp;
1279 struct bpf_binary_header *header;
1280 bool tmp_blinded = false;
1281 struct bpf_jit jit;
1282 int pass;
1283
1284 if (!bpf_jit_enable)
1285 return orig_fp;
1286
1287 tmp = bpf_jit_blind_constants(fp);
1288 /*
1289 * If blinding was requested and we failed during blinding,
1290 * we must fall back to the interpreter.
1291 */
1292 if (IS_ERR(tmp))
1293 return orig_fp;
1294 if (tmp != fp) {
1295 tmp_blinded = true;
1296 fp = tmp;
1297 }
1298
1299 memset(&jit, 0, sizeof(jit));
1300 jit.addrs = kcalloc(fp->len + 1, sizeof(*jit.addrs), GFP_KERNEL);
1301 if (jit.addrs == NULL) {
1302 fp = orig_fp;
1303 goto out;
1304 }
1305 /*
1306 * Three initial passes:
1307 * - 1/2: Determine clobbered registers
1308 * - 3: Calculate program size and addrs arrray
1309 */
1310 for (pass = 1; pass <= 3; pass++) {
1311 if (bpf_jit_prog(&jit, fp)) {
1312 fp = orig_fp;
1313 goto free_addrs;
1314 }
1315 }
1316 /*
1317 * Final pass: Allocate and generate program
1318 */
1319 if (jit.size >= BPF_SIZE_MAX) {
1320 fp = orig_fp;
1321 goto free_addrs;
1322 }
1323 header = bpf_jit_binary_alloc(jit.size, &jit.prg_buf, 2, jit_fill_hole);
1324 if (!header) {
1325 fp = orig_fp;
1326 goto free_addrs;
1327 }
1328 if (bpf_jit_prog(&jit, fp)) {
1329 fp = orig_fp;
1330 goto free_addrs;
1331 }
1332 if (bpf_jit_enable > 1) {
1333 bpf_jit_dump(fp->len, jit.size, pass, jit.prg_buf);
1334 if (jit.prg_buf)
1335 print_fn_code(jit.prg_buf, jit.size_prg);
1336 }
1337 if (jit.prg_buf) {
1338 set_memory_ro((unsigned long)header, header->pages);
1339 fp->bpf_func = (void *) jit.prg_buf;
1340 fp->jited = 1;
1341 }
1342 free_addrs:
1343 kfree(jit.addrs);
1344 out:
1345 if (tmp_blinded)
1346 bpf_jit_prog_release_other(fp, fp == orig_fp ?
1347 tmp : orig_fp);
1348 return fp;
1349 }
1350
1351 /*
1352 * Free eBPF program
1353 */
1354 void bpf_jit_free(struct bpf_prog *fp)
1355 {
1356 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
1357 struct bpf_binary_header *header = (void *)addr;
1358
1359 if (!fp->jited)
1360 goto free_filter;
1361
1362 set_memory_rw(addr, header->pages);
1363 bpf_jit_binary_free(header);
1364
1365 free_filter:
1366 bpf_prog_unlock_free(fp);
1367 }
Linux with added FPC-III support