?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 5.7.6
[linux/fpc-iii.git] / arch / s390 / net / bpf_jit_comp.c
blob0f37a1b635f858f10f440c741ae3322ff51be881
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 <linux/mm.h>
27 #include <linux/kernel.h>
28 #include <asm/cacheflush.h>
29 #include <asm/dis.h>
30 #include <asm/facility.h>
31 #include <asm/nospec-branch.h>
32 #include <asm/set_memory.h>
33 #include "bpf_jit.h"
34
35 struct bpf_jit {
36 u32 seen; /* Flags to remember seen eBPF instructions */
37 u32 seen_reg[16]; /* Array to remember which registers are used */
38 u32 *addrs; /* Array with relative instruction addresses */
39 u8 *prg_buf; /* Start of program */
40 int size; /* Size of program and literal pool */
41 int size_prg; /* Size of program */
42 int prg; /* Current position in program */
43 int lit32_start; /* Start of 32-bit literal pool */
44 int lit32; /* Current position in 32-bit literal pool */
45 int lit64_start; /* Start of 64-bit literal pool */
46 int lit64; /* Current position in 64-bit literal pool */
47 int base_ip; /* Base address for literal pool */
48 int exit_ip; /* Address of exit */
49 int r1_thunk_ip; /* Address of expoline thunk for 'br %r1' */
50 int r14_thunk_ip; /* Address of expoline thunk for 'br %r14' */
51 int tail_call_start; /* Tail call start offset */
52 int labels[1]; /* Labels for local jumps */
53 };
54
55 #define SEEN_MEM BIT(0) /* use mem[] for temporary storage */
56 #define SEEN_LITERAL BIT(1) /* code uses literals */
57 #define SEEN_FUNC BIT(2) /* calls C functions */
58 #define SEEN_TAIL_CALL BIT(3) /* code uses tail calls */
59 #define SEEN_STACK (SEEN_FUNC | SEEN_MEM)
60
61 /*
62 * s390 registers
63 */
64 #define REG_W0 (MAX_BPF_JIT_REG + 0) /* Work register 1 (even) */
65 #define REG_W1 (MAX_BPF_JIT_REG + 1) /* Work register 2 (odd) */
66 #define REG_L (MAX_BPF_JIT_REG + 2) /* Literal pool register */
67 #define REG_15 (MAX_BPF_JIT_REG + 3) /* 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 */
93 [BPF_REG_AX] = 12,
94 /* Work registers for s390x backend */
95 [REG_W0] = 0,
96 [REG_W1] = 1,
97 [REG_L] = 11,
98 [REG_15] = 15,
99 };
100
101 static inline u32 reg(u32 dst_reg, u32 src_reg)
102 {
103 return reg2hex[dst_reg] << 4 | reg2hex[src_reg];
104 }
105
106 static inline u32 reg_high(u32 reg)
107 {
108 return reg2hex[reg] << 4;
109 }
110
111 static inline void reg_set_seen(struct bpf_jit *jit, u32 b1)
112 {
113 u32 r1 = reg2hex[b1];
114
115 if (!jit->seen_reg[r1] && r1 >= 6 && r1 <= 15)
116 jit->seen_reg[r1] = 1;
117 }
118
119 #define REG_SET_SEEN(b1) \
120 ({ \
121 reg_set_seen(jit, b1); \
122 })
123
124 #define REG_SEEN(b1) jit->seen_reg[reg2hex[(b1)]]
125
126 /*
127 * EMIT macros for code generation
128 */
129
130 #define _EMIT2(op) \
131 ({ \
132 if (jit->prg_buf) \
133 *(u16 *) (jit->prg_buf + jit->prg) = (op); \
134 jit->prg += 2; \
135 })
136
137 #define EMIT2(op, b1, b2) \
138 ({ \
139 _EMIT2((op) | reg(b1, b2)); \
140 REG_SET_SEEN(b1); \
141 REG_SET_SEEN(b2); \
142 })
143
144 #define _EMIT4(op) \
145 ({ \
146 if (jit->prg_buf) \
147 *(u32 *) (jit->prg_buf + jit->prg) = (op); \
148 jit->prg += 4; \
149 })
150
151 #define EMIT4(op, b1, b2) \
152 ({ \
153 _EMIT4((op) | reg(b1, b2)); \
154 REG_SET_SEEN(b1); \
155 REG_SET_SEEN(b2); \
156 })
157
158 #define EMIT4_RRF(op, b1, b2, b3) \
159 ({ \
160 _EMIT4((op) | reg_high(b3) << 8 | reg(b1, b2)); \
161 REG_SET_SEEN(b1); \
162 REG_SET_SEEN(b2); \
163 REG_SET_SEEN(b3); \
164 })
165
166 #define _EMIT4_DISP(op, disp) \
167 ({ \
168 unsigned int __disp = (disp) & 0xfff; \
169 _EMIT4((op) | __disp); \
170 })
171
172 #define EMIT4_DISP(op, b1, b2, disp) \
173 ({ \
174 _EMIT4_DISP((op) | reg_high(b1) << 16 | \
175 reg_high(b2) << 8, (disp)); \
176 REG_SET_SEEN(b1); \
177 REG_SET_SEEN(b2); \
178 })
179
180 #define EMIT4_IMM(op, b1, imm) \
181 ({ \
182 unsigned int __imm = (imm) & 0xffff; \
183 _EMIT4((op) | reg_high(b1) << 16 | __imm); \
184 REG_SET_SEEN(b1); \
185 })
186
187 #define EMIT4_PCREL(op, pcrel) \
188 ({ \
189 long __pcrel = ((pcrel) >> 1) & 0xffff; \
190 _EMIT4((op) | __pcrel); \
191 })
192
193 #define EMIT4_PCREL_RIC(op, mask, target) \
194 ({ \
195 int __rel = ((target) - jit->prg) / 2; \
196 _EMIT4((op) | (mask) << 20 | (__rel & 0xffff)); \
197 })
198
199 #define _EMIT6(op1, op2) \
200 ({ \
201 if (jit->prg_buf) { \
202 *(u32 *) (jit->prg_buf + jit->prg) = (op1); \
203 *(u16 *) (jit->prg_buf + jit->prg + 4) = (op2); \
204 } \
205 jit->prg += 6; \
206 })
207
208 #define _EMIT6_DISP(op1, op2, disp) \
209 ({ \
210 unsigned int __disp = (disp) & 0xfff; \
211 _EMIT6((op1) | __disp, op2); \
212 })
213
214 #define _EMIT6_DISP_LH(op1, op2, disp) \
215 ({ \
216 u32 _disp = (u32) (disp); \
217 unsigned int __disp_h = _disp & 0xff000; \
218 unsigned int __disp_l = _disp & 0x00fff; \
219 _EMIT6((op1) | __disp_l, (op2) | __disp_h >> 4); \
220 })
221
222 #define EMIT6_DISP_LH(op1, op2, b1, b2, b3, disp) \
223 ({ \
224 _EMIT6_DISP_LH((op1) | reg(b1, b2) << 16 | \
225 reg_high(b3) << 8, op2, disp); \
226 REG_SET_SEEN(b1); \
227 REG_SET_SEEN(b2); \
228 REG_SET_SEEN(b3); \
229 })
230
231 #define EMIT6_PCREL_LABEL(op1, op2, b1, b2, label, mask) \
232 ({ \
233 int rel = (jit->labels[label] - jit->prg) >> 1; \
234 _EMIT6((op1) | reg(b1, b2) << 16 | (rel & 0xffff), \
235 (op2) | (mask) << 12); \
236 REG_SET_SEEN(b1); \
237 REG_SET_SEEN(b2); \
238 })
239
240 #define EMIT6_PCREL_IMM_LABEL(op1, op2, b1, imm, label, mask) \
241 ({ \
242 int rel = (jit->labels[label] - jit->prg) >> 1; \
243 _EMIT6((op1) | (reg_high(b1) | (mask)) << 16 | \
244 (rel & 0xffff), (op2) | ((imm) & 0xff) << 8); \
245 REG_SET_SEEN(b1); \
246 BUILD_BUG_ON(((unsigned long) (imm)) > 0xff); \
247 })
248
249 #define EMIT6_PCREL(op1, op2, b1, b2, i, off, mask) \
250 ({ \
251 /* Branch instruction needs 6 bytes */ \
252 int rel = (addrs[(i) + (off) + 1] - (addrs[(i) + 1] - 6)) / 2;\
253 _EMIT6((op1) | reg(b1, b2) << 16 | (rel & 0xffff), (op2) | (mask));\
254 REG_SET_SEEN(b1); \
255 REG_SET_SEEN(b2); \
256 })
257
258 #define EMIT6_PCREL_RILB(op, b, target) \
259 ({ \
260 unsigned int rel = (int)((target) - jit->prg) / 2; \
261 _EMIT6((op) | reg_high(b) << 16 | rel >> 16, rel & 0xffff);\
262 REG_SET_SEEN(b); \
263 })
264
265 #define EMIT6_PCREL_RIL(op, target) \
266 ({ \
267 unsigned int rel = (int)((target) - jit->prg) / 2; \
268 _EMIT6((op) | rel >> 16, rel & 0xffff); \
269 })
270
271 #define EMIT6_PCREL_RILC(op, mask, target) \
272 ({ \
273 EMIT6_PCREL_RIL((op) | (mask) << 20, (target)); \
274 })
275
276 #define _EMIT6_IMM(op, imm) \
277 ({ \
278 unsigned int __imm = (imm); \
279 _EMIT6((op) | (__imm >> 16), __imm & 0xffff); \
280 })
281
282 #define EMIT6_IMM(op, b1, imm) \
283 ({ \
284 _EMIT6_IMM((op) | reg_high(b1) << 16, imm); \
285 REG_SET_SEEN(b1); \
286 })
287
288 #define _EMIT_CONST_U32(val) \
289 ({ \
290 unsigned int ret; \
291 ret = jit->lit32; \
292 if (jit->prg_buf) \
293 *(u32 *)(jit->prg_buf + jit->lit32) = (u32)(val);\
294 jit->lit32 += 4; \
295 ret; \
296 })
297
298 #define EMIT_CONST_U32(val) \
299 ({ \
300 jit->seen |= SEEN_LITERAL; \
301 _EMIT_CONST_U32(val) - jit->base_ip; \
302 })
303
304 #define _EMIT_CONST_U64(val) \
305 ({ \
306 unsigned int ret; \
307 ret = jit->lit64; \
308 if (jit->prg_buf) \
309 *(u64 *)(jit->prg_buf + jit->lit64) = (u64)(val);\
310 jit->lit64 += 8; \
311 ret; \
312 })
313
314 #define EMIT_CONST_U64(val) \
315 ({ \
316 jit->seen |= SEEN_LITERAL; \
317 _EMIT_CONST_U64(val) - jit->base_ip; \
318 })
319
320 #define EMIT_ZERO(b1) \
321 ({ \
322 if (!fp->aux->verifier_zext) { \
323 /* llgfr %dst,%dst (zero extend to 64 bit) */ \
324 EMIT4(0xb9160000, b1, b1); \
325 REG_SET_SEEN(b1); \
326 } \
327 })
328
329 /*
330 * Return whether this is the first pass. The first pass is special, since we
331 * don't know any sizes yet, and thus must be conservative.
332 */
333 static bool is_first_pass(struct bpf_jit *jit)
334 {
335 return jit->size == 0;
336 }
337
338 /*
339 * Return whether this is the code generation pass. The code generation pass is
340 * special, since we should change as little as possible.
341 */
342 static bool is_codegen_pass(struct bpf_jit *jit)
343 {
344 return jit->prg_buf;
345 }
346
347 /*
348 * Return whether "rel" can be encoded as a short PC-relative offset
349 */
350 static bool is_valid_rel(int rel)
351 {
352 return rel >= -65536 && rel <= 65534;
353 }
354
355 /*
356 * Return whether "off" can be reached using a short PC-relative offset
357 */
358 static bool can_use_rel(struct bpf_jit *jit, int off)
359 {
360 return is_valid_rel(off - jit->prg);
361 }
362
363 /*
364 * Return whether given displacement can be encoded using
365 * Long-Displacement Facility
366 */
367 static bool is_valid_ldisp(int disp)
368 {
369 return disp >= -524288 && disp <= 524287;
370 }
371
372 /*
373 * Return whether the next 32-bit literal pool entry can be referenced using
374 * Long-Displacement Facility
375 */
376 static bool can_use_ldisp_for_lit32(struct bpf_jit *jit)
377 {
378 return is_valid_ldisp(jit->lit32 - jit->base_ip);
379 }
380
381 /*
382 * Return whether the next 64-bit literal pool entry can be referenced using
383 * Long-Displacement Facility
384 */
385 static bool can_use_ldisp_for_lit64(struct bpf_jit *jit)
386 {
387 return is_valid_ldisp(jit->lit64 - jit->base_ip);
388 }
389
390 /*
391 * Fill whole space with illegal instructions
392 */
393 static void jit_fill_hole(void *area, unsigned int size)
394 {
395 memset(area, 0, size);
396 }
397
398 /*
399 * Save registers from "rs" (register start) to "re" (register end) on stack
400 */
401 static void save_regs(struct bpf_jit *jit, u32 rs, u32 re)
402 {
403 u32 off = STK_OFF_R6 + (rs - 6) * 8;
404
405 if (rs == re)
406 /* stg %rs,off(%r15) */
407 _EMIT6(0xe300f000 | rs << 20 | off, 0x0024);
408 else
409 /* stmg %rs,%re,off(%r15) */
410 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0024, off);
411 }
412
413 /*
414 * Restore registers from "rs" (register start) to "re" (register end) on stack
415 */
416 static void restore_regs(struct bpf_jit *jit, u32 rs, u32 re, u32 stack_depth)
417 {
418 u32 off = STK_OFF_R6 + (rs - 6) * 8;
419
420 if (jit->seen & SEEN_STACK)
421 off += STK_OFF + stack_depth;
422
423 if (rs == re)
424 /* lg %rs,off(%r15) */
425 _EMIT6(0xe300f000 | rs << 20 | off, 0x0004);
426 else
427 /* lmg %rs,%re,off(%r15) */
428 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0004, off);
429 }
430
431 /*
432 * Return first seen register (from start)
433 */
434 static int get_start(struct bpf_jit *jit, int start)
435 {
436 int i;
437
438 for (i = start; i <= 15; i++) {
439 if (jit->seen_reg[i])
440 return i;
441 }
442 return 0;
443 }
444
445 /*
446 * Return last seen register (from start) (gap >= 2)
447 */
448 static int get_end(struct bpf_jit *jit, int start)
449 {
450 int i;
451
452 for (i = start; i < 15; i++) {
453 if (!jit->seen_reg[i] && !jit->seen_reg[i + 1])
454 return i - 1;
455 }
456 return jit->seen_reg[15] ? 15 : 14;
457 }
458
459 #define REGS_SAVE 1
460 #define REGS_RESTORE 0
461 /*
462 * Save and restore clobbered registers (6-15) on stack.
463 * We save/restore registers in chunks with gap >= 2 registers.
464 */
465 static void save_restore_regs(struct bpf_jit *jit, int op, u32 stack_depth)
466 {
467 const int last = 15, save_restore_size = 6;
468 int re = 6, rs;
469
470 if (is_first_pass(jit)) {
471 /*
472 * We don't know yet which registers are used. Reserve space
473 * conservatively.
474 */
475 jit->prg += (last - re + 1) * save_restore_size;
476 return;
477 }
478
479 do {
480 rs = get_start(jit, re);
481 if (!rs)
482 break;
483 re = get_end(jit, rs + 1);
484 if (op == REGS_SAVE)
485 save_regs(jit, rs, re);
486 else
487 restore_regs(jit, rs, re, stack_depth);
488 re++;
489 } while (re <= last);
490 }
491
492 /*
493 * Emit function prologue
494 *
495 * Save registers and create stack frame if necessary.
496 * See stack frame layout desription in "bpf_jit.h"!
497 */
498 static void bpf_jit_prologue(struct bpf_jit *jit, u32 stack_depth)
499 {
500 if (jit->seen & SEEN_TAIL_CALL) {
501 /* xc STK_OFF_TCCNT(4,%r15),STK_OFF_TCCNT(%r15) */
502 _EMIT6(0xd703f000 | STK_OFF_TCCNT, 0xf000 | STK_OFF_TCCNT);
503 } else {
504 /* j tail_call_start: NOP if no tail calls are used */
505 EMIT4_PCREL(0xa7f40000, 6);
506 _EMIT2(0);
507 }
508 /* Tail calls have to skip above initialization */
509 jit->tail_call_start = jit->prg;
510 /* Save registers */
511 save_restore_regs(jit, REGS_SAVE, stack_depth);
512 /* Setup literal pool */
513 if (is_first_pass(jit) || (jit->seen & SEEN_LITERAL)) {
514 if (!is_first_pass(jit) &&
515 is_valid_ldisp(jit->size - (jit->prg + 2))) {
516 /* basr %l,0 */
517 EMIT2(0x0d00, REG_L, REG_0);
518 jit->base_ip = jit->prg;
519 } else {
520 /* larl %l,lit32_start */
521 EMIT6_PCREL_RILB(0xc0000000, REG_L, jit->lit32_start);
522 jit->base_ip = jit->lit32_start;
523 }
524 }
525 /* Setup stack and backchain */
526 if (is_first_pass(jit) || (jit->seen & SEEN_STACK)) {
527 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC))
528 /* lgr %w1,%r15 (backchain) */
529 EMIT4(0xb9040000, REG_W1, REG_15);
530 /* la %bfp,STK_160_UNUSED(%r15) (BPF frame pointer) */
531 EMIT4_DISP(0x41000000, BPF_REG_FP, REG_15, STK_160_UNUSED);
532 /* aghi %r15,-STK_OFF */
533 EMIT4_IMM(0xa70b0000, REG_15, -(STK_OFF + stack_depth));
534 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC))
535 /* stg %w1,152(%r15) (backchain) */
536 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W1, REG_0,
537 REG_15, 152);
538 }
539 }
540
541 /*
542 * Function epilogue
543 */
544 static void bpf_jit_epilogue(struct bpf_jit *jit, u32 stack_depth)
545 {
546 jit->exit_ip = jit->prg;
547 /* Load exit code: lgr %r2,%b0 */
548 EMIT4(0xb9040000, REG_2, BPF_REG_0);
549 /* Restore registers */
550 save_restore_regs(jit, REGS_RESTORE, stack_depth);
551 if (__is_defined(CC_USING_EXPOLINE) && !nospec_disable) {
552 jit->r14_thunk_ip = jit->prg;
553 /* Generate __s390_indirect_jump_r14 thunk */
554 if (test_facility(35)) {
555 /* exrl %r0,.+10 */
556 EMIT6_PCREL_RIL(0xc6000000, jit->prg + 10);
557 } else {
558 /* larl %r1,.+14 */
559 EMIT6_PCREL_RILB(0xc0000000, REG_1, jit->prg + 14);
560 /* ex 0,0(%r1) */
561 EMIT4_DISP(0x44000000, REG_0, REG_1, 0);
562 }
563 /* j . */
564 EMIT4_PCREL(0xa7f40000, 0);
565 }
566 /* br %r14 */
567 _EMIT2(0x07fe);
568
569 if (__is_defined(CC_USING_EXPOLINE) && !nospec_disable &&
570 (is_first_pass(jit) || (jit->seen & SEEN_FUNC))) {
571 jit->r1_thunk_ip = jit->prg;
572 /* Generate __s390_indirect_jump_r1 thunk */
573 if (test_facility(35)) {
574 /* exrl %r0,.+10 */
575 EMIT6_PCREL_RIL(0xc6000000, jit->prg + 10);
576 /* j . */
577 EMIT4_PCREL(0xa7f40000, 0);
578 /* br %r1 */
579 _EMIT2(0x07f1);
580 } else {
581 /* ex 0,S390_lowcore.br_r1_tampoline */
582 EMIT4_DISP(0x44000000, REG_0, REG_0,
583 offsetof(struct lowcore, br_r1_trampoline));
584 /* j . */
585 EMIT4_PCREL(0xa7f40000, 0);
586 }
587 }
588 }
589
590 /*
591 * Compile one eBPF instruction into s390x code
592 *
593 * NOTE: Use noinline because for gcov (-fprofile-arcs) gcc allocates a lot of
594 * stack space for the large switch statement.
595 */
596 static noinline int bpf_jit_insn(struct bpf_jit *jit, struct bpf_prog *fp,
597 int i, bool extra_pass, u32 stack_depth)
598 {
599 struct bpf_insn *insn = &fp->insnsi[i];
600 u32 dst_reg = insn->dst_reg;
601 u32 src_reg = insn->src_reg;
602 int last, insn_count = 1;
603 u32 *addrs = jit->addrs;
604 s32 imm = insn->imm;
605 s16 off = insn->off;
606 unsigned int mask;
607
608 switch (insn->code) {
609 /*
610 * BPF_MOV
611 */
612 case BPF_ALU | BPF_MOV | BPF_X: /* dst = (u32) src */
613 /* llgfr %dst,%src */
614 EMIT4(0xb9160000, dst_reg, src_reg);
615 if (insn_is_zext(&insn[1]))
616 insn_count = 2;
617 break;
618 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */
619 /* lgr %dst,%src */
620 EMIT4(0xb9040000, dst_reg, src_reg);
621 break;
622 case BPF_ALU | BPF_MOV | BPF_K: /* dst = (u32) imm */
623 /* llilf %dst,imm */
624 EMIT6_IMM(0xc00f0000, dst_reg, imm);
625 if (insn_is_zext(&insn[1]))
626 insn_count = 2;
627 break;
628 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = imm */
629 /* lgfi %dst,imm */
630 EMIT6_IMM(0xc0010000, dst_reg, imm);
631 break;
632 /*
633 * BPF_LD 64
634 */
635 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
636 {
637 /* 16 byte instruction that uses two 'struct bpf_insn' */
638 u64 imm64;
639
640 imm64 = (u64)(u32) insn[0].imm | ((u64)(u32) insn[1].imm) << 32;
641 /* lgrl %dst,imm */
642 EMIT6_PCREL_RILB(0xc4080000, dst_reg, _EMIT_CONST_U64(imm64));
643 insn_count = 2;
644 break;
645 }
646 /*
647 * BPF_ADD
648 */
649 case BPF_ALU | BPF_ADD | BPF_X: /* dst = (u32) dst + (u32) src */
650 /* ar %dst,%src */
651 EMIT2(0x1a00, dst_reg, src_reg);
652 EMIT_ZERO(dst_reg);
653 break;
654 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst = dst + src */
655 /* agr %dst,%src */
656 EMIT4(0xb9080000, dst_reg, src_reg);
657 break;
658 case BPF_ALU | BPF_ADD | BPF_K: /* dst = (u32) dst + (u32) imm */
659 if (!imm)
660 break;
661 /* alfi %dst,imm */
662 EMIT6_IMM(0xc20b0000, dst_reg, imm);
663 EMIT_ZERO(dst_reg);
664 break;
665 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst = dst + imm */
666 if (!imm)
667 break;
668 /* agfi %dst,imm */
669 EMIT6_IMM(0xc2080000, dst_reg, imm);
670 break;
671 /*
672 * BPF_SUB
673 */
674 case BPF_ALU | BPF_SUB | BPF_X: /* dst = (u32) dst - (u32) src */
675 /* sr %dst,%src */
676 EMIT2(0x1b00, dst_reg, src_reg);
677 EMIT_ZERO(dst_reg);
678 break;
679 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst = dst - src */
680 /* sgr %dst,%src */
681 EMIT4(0xb9090000, dst_reg, src_reg);
682 break;
683 case BPF_ALU | BPF_SUB | BPF_K: /* dst = (u32) dst - (u32) imm */
684 if (!imm)
685 break;
686 /* alfi %dst,-imm */
687 EMIT6_IMM(0xc20b0000, dst_reg, -imm);
688 EMIT_ZERO(dst_reg);
689 break;
690 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst = dst - imm */
691 if (!imm)
692 break;
693 /* agfi %dst,-imm */
694 EMIT6_IMM(0xc2080000, dst_reg, -imm);
695 break;
696 /*
697 * BPF_MUL
698 */
699 case BPF_ALU | BPF_MUL | BPF_X: /* dst = (u32) dst * (u32) src */
700 /* msr %dst,%src */
701 EMIT4(0xb2520000, dst_reg, src_reg);
702 EMIT_ZERO(dst_reg);
703 break;
704 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst = dst * src */
705 /* msgr %dst,%src */
706 EMIT4(0xb90c0000, dst_reg, src_reg);
707 break;
708 case BPF_ALU | BPF_MUL | BPF_K: /* dst = (u32) dst * (u32) imm */
709 if (imm == 1)
710 break;
711 /* msfi %r5,imm */
712 EMIT6_IMM(0xc2010000, dst_reg, imm);
713 EMIT_ZERO(dst_reg);
714 break;
715 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst = dst * imm */
716 if (imm == 1)
717 break;
718 /* msgfi %dst,imm */
719 EMIT6_IMM(0xc2000000, dst_reg, imm);
720 break;
721 /*
722 * BPF_DIV / BPF_MOD
723 */
724 case BPF_ALU | BPF_DIV | BPF_X: /* dst = (u32) dst / (u32) src */
725 case BPF_ALU | BPF_MOD | BPF_X: /* dst = (u32) dst % (u32) src */
726 {
727 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
728
729 /* lhi %w0,0 */
730 EMIT4_IMM(0xa7080000, REG_W0, 0);
731 /* lr %w1,%dst */
732 EMIT2(0x1800, REG_W1, dst_reg);
733 /* dlr %w0,%src */
734 EMIT4(0xb9970000, REG_W0, src_reg);
735 /* llgfr %dst,%rc */
736 EMIT4(0xb9160000, dst_reg, rc_reg);
737 if (insn_is_zext(&insn[1]))
738 insn_count = 2;
739 break;
740 }
741 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst = dst / src */
742 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst = dst % src */
743 {
744 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
745
746 /* lghi %w0,0 */
747 EMIT4_IMM(0xa7090000, REG_W0, 0);
748 /* lgr %w1,%dst */
749 EMIT4(0xb9040000, REG_W1, dst_reg);
750 /* dlgr %w0,%dst */
751 EMIT4(0xb9870000, REG_W0, src_reg);
752 /* lgr %dst,%rc */
753 EMIT4(0xb9040000, dst_reg, rc_reg);
754 break;
755 }
756 case BPF_ALU | BPF_DIV | BPF_K: /* dst = (u32) dst / (u32) imm */
757 case BPF_ALU | BPF_MOD | BPF_K: /* dst = (u32) dst % (u32) imm */
758 {
759 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
760
761 if (imm == 1) {
762 if (BPF_OP(insn->code) == BPF_MOD)
763 /* lhgi %dst,0 */
764 EMIT4_IMM(0xa7090000, dst_reg, 0);
765 break;
766 }
767 /* lhi %w0,0 */
768 EMIT4_IMM(0xa7080000, REG_W0, 0);
769 /* lr %w1,%dst */
770 EMIT2(0x1800, REG_W1, dst_reg);
771 if (!is_first_pass(jit) && can_use_ldisp_for_lit32(jit)) {
772 /* dl %w0,<d(imm)>(%l) */
773 EMIT6_DISP_LH(0xe3000000, 0x0097, REG_W0, REG_0, REG_L,
774 EMIT_CONST_U32(imm));
775 } else {
776 /* lgfrl %dst,imm */
777 EMIT6_PCREL_RILB(0xc40c0000, dst_reg,
778 _EMIT_CONST_U32(imm));
779 jit->seen |= SEEN_LITERAL;
780 /* dlr %w0,%dst */
781 EMIT4(0xb9970000, REG_W0, dst_reg);
782 }
783 /* llgfr %dst,%rc */
784 EMIT4(0xb9160000, dst_reg, rc_reg);
785 if (insn_is_zext(&insn[1]))
786 insn_count = 2;
787 break;
788 }
789 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst = dst / imm */
790 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst = dst % imm */
791 {
792 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
793
794 if (imm == 1) {
795 if (BPF_OP(insn->code) == BPF_MOD)
796 /* lhgi %dst,0 */
797 EMIT4_IMM(0xa7090000, dst_reg, 0);
798 break;
799 }
800 /* lghi %w0,0 */
801 EMIT4_IMM(0xa7090000, REG_W0, 0);
802 /* lgr %w1,%dst */
803 EMIT4(0xb9040000, REG_W1, dst_reg);
804 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
805 /* dlg %w0,<d(imm)>(%l) */
806 EMIT6_DISP_LH(0xe3000000, 0x0087, REG_W0, REG_0, REG_L,
807 EMIT_CONST_U64(imm));
808 } else {
809 /* lgrl %dst,imm */
810 EMIT6_PCREL_RILB(0xc4080000, dst_reg,
811 _EMIT_CONST_U64(imm));
812 jit->seen |= SEEN_LITERAL;
813 /* dlgr %w0,%dst */
814 EMIT4(0xb9870000, REG_W0, dst_reg);
815 }
816 /* lgr %dst,%rc */
817 EMIT4(0xb9040000, dst_reg, rc_reg);
818 break;
819 }
820 /*
821 * BPF_AND
822 */
823 case BPF_ALU | BPF_AND | BPF_X: /* dst = (u32) dst & (u32) src */
824 /* nr %dst,%src */
825 EMIT2(0x1400, dst_reg, src_reg);
826 EMIT_ZERO(dst_reg);
827 break;
828 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
829 /* ngr %dst,%src */
830 EMIT4(0xb9800000, dst_reg, src_reg);
831 break;
832 case BPF_ALU | BPF_AND | BPF_K: /* dst = (u32) dst & (u32) imm */
833 /* nilf %dst,imm */
834 EMIT6_IMM(0xc00b0000, dst_reg, imm);
835 EMIT_ZERO(dst_reg);
836 break;
837 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
838 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
839 /* ng %dst,<d(imm)>(%l) */
840 EMIT6_DISP_LH(0xe3000000, 0x0080,
841 dst_reg, REG_0, REG_L,
842 EMIT_CONST_U64(imm));
843 } else {
844 /* lgrl %w0,imm */
845 EMIT6_PCREL_RILB(0xc4080000, REG_W0,
846 _EMIT_CONST_U64(imm));
847 jit->seen |= SEEN_LITERAL;
848 /* ngr %dst,%w0 */
849 EMIT4(0xb9800000, dst_reg, REG_W0);
850 }
851 break;
852 /*
853 * BPF_OR
854 */
855 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
856 /* or %dst,%src */
857 EMIT2(0x1600, dst_reg, src_reg);
858 EMIT_ZERO(dst_reg);
859 break;
860 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
861 /* ogr %dst,%src */
862 EMIT4(0xb9810000, dst_reg, src_reg);
863 break;
864 case BPF_ALU | BPF_OR | BPF_K: /* dst = (u32) dst | (u32) imm */
865 /* oilf %dst,imm */
866 EMIT6_IMM(0xc00d0000, dst_reg, imm);
867 EMIT_ZERO(dst_reg);
868 break;
869 case BPF_ALU64 | BPF_OR | BPF_K: /* dst = dst | imm */
870 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
871 /* og %dst,<d(imm)>(%l) */
872 EMIT6_DISP_LH(0xe3000000, 0x0081,
873 dst_reg, REG_0, REG_L,
874 EMIT_CONST_U64(imm));
875 } else {
876 /* lgrl %w0,imm */
877 EMIT6_PCREL_RILB(0xc4080000, REG_W0,
878 _EMIT_CONST_U64(imm));
879 jit->seen |= SEEN_LITERAL;
880 /* ogr %dst,%w0 */
881 EMIT4(0xb9810000, dst_reg, REG_W0);
882 }
883 break;
884 /*
885 * BPF_XOR
886 */
887 case BPF_ALU | BPF_XOR | BPF_X: /* dst = (u32) dst ^ (u32) src */
888 /* xr %dst,%src */
889 EMIT2(0x1700, dst_reg, src_reg);
890 EMIT_ZERO(dst_reg);
891 break;
892 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst = dst ^ src */
893 /* xgr %dst,%src */
894 EMIT4(0xb9820000, dst_reg, src_reg);
895 break;
896 case BPF_ALU | BPF_XOR | BPF_K: /* dst = (u32) dst ^ (u32) imm */
897 if (!imm)
898 break;
899 /* xilf %dst,imm */
900 EMIT6_IMM(0xc0070000, dst_reg, imm);
901 EMIT_ZERO(dst_reg);
902 break;
903 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst = dst ^ imm */
904 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
905 /* xg %dst,<d(imm)>(%l) */
906 EMIT6_DISP_LH(0xe3000000, 0x0082,
907 dst_reg, REG_0, REG_L,
908 EMIT_CONST_U64(imm));
909 } else {
910 /* lgrl %w0,imm */
911 EMIT6_PCREL_RILB(0xc4080000, REG_W0,
912 _EMIT_CONST_U64(imm));
913 jit->seen |= SEEN_LITERAL;
914 /* xgr %dst,%w0 */
915 EMIT4(0xb9820000, dst_reg, REG_W0);
916 }
917 break;
918 /*
919 * BPF_LSH
920 */
921 case BPF_ALU | BPF_LSH | BPF_X: /* dst = (u32) dst << (u32) src */
922 /* sll %dst,0(%src) */
923 EMIT4_DISP(0x89000000, dst_reg, src_reg, 0);
924 EMIT_ZERO(dst_reg);
925 break;
926 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst = dst << src */
927 /* sllg %dst,%dst,0(%src) */
928 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, src_reg, 0);
929 break;
930 case BPF_ALU | BPF_LSH | BPF_K: /* dst = (u32) dst << (u32) imm */
931 if (imm == 0)
932 break;
933 /* sll %dst,imm(%r0) */
934 EMIT4_DISP(0x89000000, dst_reg, REG_0, imm);
935 EMIT_ZERO(dst_reg);
936 break;
937 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst = dst << imm */
938 if (imm == 0)
939 break;
940 /* sllg %dst,%dst,imm(%r0) */
941 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, REG_0, imm);
942 break;
943 /*
944 * BPF_RSH
945 */
946 case BPF_ALU | BPF_RSH | BPF_X: /* dst = (u32) dst >> (u32) src */
947 /* srl %dst,0(%src) */
948 EMIT4_DISP(0x88000000, dst_reg, src_reg, 0);
949 EMIT_ZERO(dst_reg);
950 break;
951 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst = dst >> src */
952 /* srlg %dst,%dst,0(%src) */
953 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, src_reg, 0);
954 break;
955 case BPF_ALU | BPF_RSH | BPF_K: /* dst = (u32) dst >> (u32) imm */
956 if (imm == 0)
957 break;
958 /* srl %dst,imm(%r0) */
959 EMIT4_DISP(0x88000000, dst_reg, REG_0, imm);
960 EMIT_ZERO(dst_reg);
961 break;
962 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst = dst >> imm */
963 if (imm == 0)
964 break;
965 /* srlg %dst,%dst,imm(%r0) */
966 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, REG_0, imm);
967 break;
968 /*
969 * BPF_ARSH
970 */
971 case BPF_ALU | BPF_ARSH | BPF_X: /* ((s32) dst) >>= src */
972 /* sra %dst,%dst,0(%src) */
973 EMIT4_DISP(0x8a000000, dst_reg, src_reg, 0);
974 EMIT_ZERO(dst_reg);
975 break;
976 case BPF_ALU64 | BPF_ARSH | BPF_X: /* ((s64) dst) >>= src */
977 /* srag %dst,%dst,0(%src) */
978 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, src_reg, 0);
979 break;
980 case BPF_ALU | BPF_ARSH | BPF_K: /* ((s32) dst >> imm */
981 if (imm == 0)
982 break;
983 /* sra %dst,imm(%r0) */
984 EMIT4_DISP(0x8a000000, dst_reg, REG_0, imm);
985 EMIT_ZERO(dst_reg);
986 break;
987 case BPF_ALU64 | BPF_ARSH | BPF_K: /* ((s64) dst) >>= imm */
988 if (imm == 0)
989 break;
990 /* srag %dst,%dst,imm(%r0) */
991 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, REG_0, imm);
992 break;
993 /*
994 * BPF_NEG
995 */
996 case BPF_ALU | BPF_NEG: /* dst = (u32) -dst */
997 /* lcr %dst,%dst */
998 EMIT2(0x1300, dst_reg, dst_reg);
999 EMIT_ZERO(dst_reg);
1000 break;
1001 case BPF_ALU64 | BPF_NEG: /* dst = -dst */
1002 /* lcgr %dst,%dst */
1003 EMIT4(0xb9030000, dst_reg, dst_reg);
1004 break;
1005 /*
1006 * BPF_FROM_BE/LE
1007 */
1008 case BPF_ALU | BPF_END | BPF_FROM_BE:
1009 /* s390 is big endian, therefore only clear high order bytes */
1010 switch (imm) {
1011 case 16: /* dst = (u16) cpu_to_be16(dst) */
1012 /* llghr %dst,%dst */
1013 EMIT4(0xb9850000, dst_reg, dst_reg);
1014 if (insn_is_zext(&insn[1]))
1015 insn_count = 2;
1016 break;
1017 case 32: /* dst = (u32) cpu_to_be32(dst) */
1018 if (!fp->aux->verifier_zext)
1019 /* llgfr %dst,%dst */
1020 EMIT4(0xb9160000, dst_reg, dst_reg);
1021 break;
1022 case 64: /* dst = (u64) cpu_to_be64(dst) */
1023 break;
1024 }
1025 break;
1026 case BPF_ALU | BPF_END | BPF_FROM_LE:
1027 switch (imm) {
1028 case 16: /* dst = (u16) cpu_to_le16(dst) */
1029 /* lrvr %dst,%dst */
1030 EMIT4(0xb91f0000, dst_reg, dst_reg);
1031 /* srl %dst,16(%r0) */
1032 EMIT4_DISP(0x88000000, dst_reg, REG_0, 16);
1033 /* llghr %dst,%dst */
1034 EMIT4(0xb9850000, dst_reg, dst_reg);
1035 if (insn_is_zext(&insn[1]))
1036 insn_count = 2;
1037 break;
1038 case 32: /* dst = (u32) cpu_to_le32(dst) */
1039 /* lrvr %dst,%dst */
1040 EMIT4(0xb91f0000, dst_reg, dst_reg);
1041 if (!fp->aux->verifier_zext)
1042 /* llgfr %dst,%dst */
1043 EMIT4(0xb9160000, dst_reg, dst_reg);
1044 break;
1045 case 64: /* dst = (u64) cpu_to_le64(dst) */
1046 /* lrvgr %dst,%dst */
1047 EMIT4(0xb90f0000, dst_reg, dst_reg);
1048 break;
1049 }
1050 break;
1051 /*
1052 * BPF_ST(X)
1053 */
1054 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src_reg */
1055 /* stcy %src,off(%dst) */
1056 EMIT6_DISP_LH(0xe3000000, 0x0072, src_reg, dst_reg, REG_0, off);
1057 jit->seen |= SEEN_MEM;
1058 break;
1059 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
1060 /* sthy %src,off(%dst) */
1061 EMIT6_DISP_LH(0xe3000000, 0x0070, src_reg, dst_reg, REG_0, off);
1062 jit->seen |= SEEN_MEM;
1063 break;
1064 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
1065 /* sty %src,off(%dst) */
1066 EMIT6_DISP_LH(0xe3000000, 0x0050, src_reg, dst_reg, REG_0, off);
1067 jit->seen |= SEEN_MEM;
1068 break;
1069 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
1070 /* stg %src,off(%dst) */
1071 EMIT6_DISP_LH(0xe3000000, 0x0024, src_reg, dst_reg, REG_0, off);
1072 jit->seen |= SEEN_MEM;
1073 break;
1074 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
1075 /* lhi %w0,imm */
1076 EMIT4_IMM(0xa7080000, REG_W0, (u8) imm);
1077 /* stcy %w0,off(dst) */
1078 EMIT6_DISP_LH(0xe3000000, 0x0072, REG_W0, dst_reg, REG_0, off);
1079 jit->seen |= SEEN_MEM;
1080 break;
1081 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
1082 /* lhi %w0,imm */
1083 EMIT4_IMM(0xa7080000, REG_W0, (u16) imm);
1084 /* sthy %w0,off(dst) */
1085 EMIT6_DISP_LH(0xe3000000, 0x0070, REG_W0, dst_reg, REG_0, off);
1086 jit->seen |= SEEN_MEM;
1087 break;
1088 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
1089 /* llilf %w0,imm */
1090 EMIT6_IMM(0xc00f0000, REG_W0, (u32) imm);
1091 /* sty %w0,off(%dst) */
1092 EMIT6_DISP_LH(0xe3000000, 0x0050, REG_W0, dst_reg, REG_0, off);
1093 jit->seen |= SEEN_MEM;
1094 break;
1095 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
1096 /* lgfi %w0,imm */
1097 EMIT6_IMM(0xc0010000, REG_W0, imm);
1098 /* stg %w0,off(%dst) */
1099 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W0, dst_reg, REG_0, off);
1100 jit->seen |= SEEN_MEM;
1101 break;
1102 /*
1103 * BPF_STX XADD (atomic_add)
1104 */
1105 case BPF_STX | BPF_XADD | BPF_W: /* *(u32 *)(dst + off) += src */
1106 /* laal %w0,%src,off(%dst) */
1107 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W0, src_reg,
1108 dst_reg, off);
1109 jit->seen |= SEEN_MEM;
1110 break;
1111 case BPF_STX | BPF_XADD | BPF_DW: /* *(u64 *)(dst + off) += src */
1112 /* laalg %w0,%src,off(%dst) */
1113 EMIT6_DISP_LH(0xeb000000, 0x00ea, REG_W0, src_reg,
1114 dst_reg, off);
1115 jit->seen |= SEEN_MEM;
1116 break;
1117 /*
1118 * BPF_LDX
1119 */
1120 case BPF_LDX | BPF_MEM | BPF_B: /* dst = *(u8 *)(ul) (src + off) */
1121 /* llgc %dst,0(off,%src) */
1122 EMIT6_DISP_LH(0xe3000000, 0x0090, dst_reg, src_reg, REG_0, off);
1123 jit->seen |= SEEN_MEM;
1124 if (insn_is_zext(&insn[1]))
1125 insn_count = 2;
1126 break;
1127 case BPF_LDX | BPF_MEM | BPF_H: /* dst = *(u16 *)(ul) (src + off) */
1128 /* llgh %dst,0(off,%src) */
1129 EMIT6_DISP_LH(0xe3000000, 0x0091, dst_reg, src_reg, REG_0, off);
1130 jit->seen |= SEEN_MEM;
1131 if (insn_is_zext(&insn[1]))
1132 insn_count = 2;
1133 break;
1134 case BPF_LDX | BPF_MEM | BPF_W: /* dst = *(u32 *)(ul) (src + off) */
1135 /* llgf %dst,off(%src) */
1136 jit->seen |= SEEN_MEM;
1137 EMIT6_DISP_LH(0xe3000000, 0x0016, dst_reg, src_reg, REG_0, off);
1138 if (insn_is_zext(&insn[1]))
1139 insn_count = 2;
1140 break;
1141 case BPF_LDX | BPF_MEM | BPF_DW: /* dst = *(u64 *)(ul) (src + off) */
1142 /* lg %dst,0(off,%src) */
1143 jit->seen |= SEEN_MEM;
1144 EMIT6_DISP_LH(0xe3000000, 0x0004, dst_reg, src_reg, REG_0, off);
1145 break;
1146 /*
1147 * BPF_JMP / CALL
1148 */
1149 case BPF_JMP | BPF_CALL:
1150 {
1151 u64 func;
1152 bool func_addr_fixed;
1153 int ret;
1154
1155 ret = bpf_jit_get_func_addr(fp, insn, extra_pass,
1156 &func, &func_addr_fixed);
1157 if (ret < 0)
1158 return -1;
1159
1160 REG_SET_SEEN(BPF_REG_5);
1161 jit->seen |= SEEN_FUNC;
1162 /* lgrl %w1,func */
1163 EMIT6_PCREL_RILB(0xc4080000, REG_W1, _EMIT_CONST_U64(func));
1164 if (__is_defined(CC_USING_EXPOLINE) && !nospec_disable) {
1165 /* brasl %r14,__s390_indirect_jump_r1 */
1166 EMIT6_PCREL_RILB(0xc0050000, REG_14, jit->r1_thunk_ip);
1167 } else {
1168 /* basr %r14,%w1 */
1169 EMIT2(0x0d00, REG_14, REG_W1);
1170 }
1171 /* lgr %b0,%r2: load return value into %b0 */
1172 EMIT4(0xb9040000, BPF_REG_0, REG_2);
1173 break;
1174 }
1175 case BPF_JMP | BPF_TAIL_CALL:
1176 /*
1177 * Implicit input:
1178 * B1: pointer to ctx
1179 * B2: pointer to bpf_array
1180 * B3: index in bpf_array
1181 */
1182 jit->seen |= SEEN_TAIL_CALL;
1183
1184 /*
1185 * if (index >= array->map.max_entries)
1186 * goto out;
1187 */
1188
1189 /* llgf %w1,map.max_entries(%b2) */
1190 EMIT6_DISP_LH(0xe3000000, 0x0016, REG_W1, REG_0, BPF_REG_2,
1191 offsetof(struct bpf_array, map.max_entries));
1192 /* if ((u32)%b3 >= (u32)%w1) goto out; */
1193 if (!is_first_pass(jit) && can_use_rel(jit, jit->labels[0])) {
1194 /* clrj %b3,%w1,0xa,label0 */
1195 EMIT6_PCREL_LABEL(0xec000000, 0x0077, BPF_REG_3,
1196 REG_W1, 0, 0xa);
1197 } else {
1198 /* clr %b3,%w1 */
1199 EMIT2(0x1500, BPF_REG_3, REG_W1);
1200 /* brcl 0xa,label0 */
1201 EMIT6_PCREL_RILC(0xc0040000, 0xa, jit->labels[0]);
1202 }
1203
1204 /*
1205 * if (tail_call_cnt++ > MAX_TAIL_CALL_CNT)
1206 * goto out;
1207 */
1208
1209 if (jit->seen & SEEN_STACK)
1210 off = STK_OFF_TCCNT + STK_OFF + stack_depth;
1211 else
1212 off = STK_OFF_TCCNT;
1213 /* lhi %w0,1 */
1214 EMIT4_IMM(0xa7080000, REG_W0, 1);
1215 /* laal %w1,%w0,off(%r15) */
1216 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W1, REG_W0, REG_15, off);
1217 if (!is_first_pass(jit) && can_use_rel(jit, jit->labels[0])) {
1218 /* clij %w1,MAX_TAIL_CALL_CNT,0x2,label0 */
1219 EMIT6_PCREL_IMM_LABEL(0xec000000, 0x007f, REG_W1,
1220 MAX_TAIL_CALL_CNT, 0, 0x2);
1221 } else {
1222 /* clfi %w1,MAX_TAIL_CALL_CNT */
1223 EMIT6_IMM(0xc20f0000, REG_W1, MAX_TAIL_CALL_CNT);
1224 /* brcl 0x2,label0 */
1225 EMIT6_PCREL_RILC(0xc0040000, 0x2, jit->labels[0]);
1226 }
1227
1228 /*
1229 * prog = array->ptrs[index];
1230 * if (prog == NULL)
1231 * goto out;
1232 */
1233
1234 /* llgfr %r1,%b3: %r1 = (u32) index */
1235 EMIT4(0xb9160000, REG_1, BPF_REG_3);
1236 /* sllg %r1,%r1,3: %r1 *= 8 */
1237 EMIT6_DISP_LH(0xeb000000, 0x000d, REG_1, REG_1, REG_0, 3);
1238 /* ltg %r1,prog(%b2,%r1) */
1239 EMIT6_DISP_LH(0xe3000000, 0x0002, REG_1, BPF_REG_2,
1240 REG_1, offsetof(struct bpf_array, ptrs));
1241 if (!is_first_pass(jit) && can_use_rel(jit, jit->labels[0])) {
1242 /* brc 0x8,label0 */
1243 EMIT4_PCREL_RIC(0xa7040000, 0x8, jit->labels[0]);
1244 } else {
1245 /* brcl 0x8,label0 */
1246 EMIT6_PCREL_RILC(0xc0040000, 0x8, jit->labels[0]);
1247 }
1248
1249 /*
1250 * Restore registers before calling function
1251 */
1252 save_restore_regs(jit, REGS_RESTORE, stack_depth);
1253
1254 /*
1255 * goto *(prog->bpf_func + tail_call_start);
1256 */
1257
1258 /* lg %r1,bpf_func(%r1) */
1259 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, REG_1, REG_0,
1260 offsetof(struct bpf_prog, bpf_func));
1261 /* bc 0xf,tail_call_start(%r1) */
1262 _EMIT4(0x47f01000 + jit->tail_call_start);
1263 /* out: */
1264 jit->labels[0] = jit->prg;
1265 break;
1266 case BPF_JMP | BPF_EXIT: /* return b0 */
1267 last = (i == fp->len - 1) ? 1 : 0;
1268 if (last)
1269 break;
1270 /* j <exit> */
1271 EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg);
1272 break;
1273 /*
1274 * Branch relative (number of skipped instructions) to offset on
1275 * condition.
1276 *
1277 * Condition code to mask mapping:
1278 *
1279 * CC | Description | Mask
1280 * ------------------------------
1281 * 0 | Operands equal | 8
1282 * 1 | First operand low | 4
1283 * 2 | First operand high | 2
1284 * 3 | Unused | 1
1285 *
1286 * For s390x relative branches: ip = ip + off_bytes
1287 * For BPF relative branches: insn = insn + off_insns + 1
1288 *
1289 * For example for s390x with offset 0 we jump to the branch
1290 * instruction itself (loop) and for BPF with offset 0 we
1291 * branch to the instruction behind the branch.
1292 */
1293 case BPF_JMP | BPF_JA: /* if (true) */
1294 mask = 0xf000; /* j */
1295 goto branch_oc;
1296 case BPF_JMP | BPF_JSGT | BPF_K: /* ((s64) dst > (s64) imm) */
1297 case BPF_JMP32 | BPF_JSGT | BPF_K: /* ((s32) dst > (s32) imm) */
1298 mask = 0x2000; /* jh */
1299 goto branch_ks;
1300 case BPF_JMP | BPF_JSLT | BPF_K: /* ((s64) dst < (s64) imm) */
1301 case BPF_JMP32 | BPF_JSLT | BPF_K: /* ((s32) dst < (s32) imm) */
1302 mask = 0x4000; /* jl */
1303 goto branch_ks;
1304 case BPF_JMP | BPF_JSGE | BPF_K: /* ((s64) dst >= (s64) imm) */
1305 case BPF_JMP32 | BPF_JSGE | BPF_K: /* ((s32) dst >= (s32) imm) */
1306 mask = 0xa000; /* jhe */
1307 goto branch_ks;
1308 case BPF_JMP | BPF_JSLE | BPF_K: /* ((s64) dst <= (s64) imm) */
1309 case BPF_JMP32 | BPF_JSLE | BPF_K: /* ((s32) dst <= (s32) imm) */
1310 mask = 0xc000; /* jle */
1311 goto branch_ks;
1312 case BPF_JMP | BPF_JGT | BPF_K: /* (dst_reg > imm) */
1313 case BPF_JMP32 | BPF_JGT | BPF_K: /* ((u32) dst_reg > (u32) imm) */
1314 mask = 0x2000; /* jh */
1315 goto branch_ku;
1316 case BPF_JMP | BPF_JLT | BPF_K: /* (dst_reg < imm) */
1317 case BPF_JMP32 | BPF_JLT | BPF_K: /* ((u32) dst_reg < (u32) imm) */
1318 mask = 0x4000; /* jl */
1319 goto branch_ku;
1320 case BPF_JMP | BPF_JGE | BPF_K: /* (dst_reg >= imm) */
1321 case BPF_JMP32 | BPF_JGE | BPF_K: /* ((u32) dst_reg >= (u32) imm) */
1322 mask = 0xa000; /* jhe */
1323 goto branch_ku;
1324 case BPF_JMP | BPF_JLE | BPF_K: /* (dst_reg <= imm) */
1325 case BPF_JMP32 | BPF_JLE | BPF_K: /* ((u32) dst_reg <= (u32) imm) */
1326 mask = 0xc000; /* jle */
1327 goto branch_ku;
1328 case BPF_JMP | BPF_JNE | BPF_K: /* (dst_reg != imm) */
1329 case BPF_JMP32 | BPF_JNE | BPF_K: /* ((u32) dst_reg != (u32) imm) */
1330 mask = 0x7000; /* jne */
1331 goto branch_ku;
1332 case BPF_JMP | BPF_JEQ | BPF_K: /* (dst_reg == imm) */
1333 case BPF_JMP32 | BPF_JEQ | BPF_K: /* ((u32) dst_reg == (u32) imm) */
1334 mask = 0x8000; /* je */
1335 goto branch_ku;
1336 case BPF_JMP | BPF_JSET | BPF_K: /* (dst_reg & imm) */
1337 case BPF_JMP32 | BPF_JSET | BPF_K: /* ((u32) dst_reg & (u32) imm) */
1338 mask = 0x7000; /* jnz */
1339 if (BPF_CLASS(insn->code) == BPF_JMP32) {
1340 /* llilf %w1,imm (load zero extend imm) */
1341 EMIT6_IMM(0xc00f0000, REG_W1, imm);
1342 /* nr %w1,%dst */
1343 EMIT2(0x1400, REG_W1, dst_reg);
1344 } else {
1345 /* lgfi %w1,imm (load sign extend imm) */
1346 EMIT6_IMM(0xc0010000, REG_W1, imm);
1347 /* ngr %w1,%dst */
1348 EMIT4(0xb9800000, REG_W1, dst_reg);
1349 }
1350 goto branch_oc;
1351
1352 case BPF_JMP | BPF_JSGT | BPF_X: /* ((s64) dst > (s64) src) */
1353 case BPF_JMP32 | BPF_JSGT | BPF_X: /* ((s32) dst > (s32) src) */
1354 mask = 0x2000; /* jh */
1355 goto branch_xs;
1356 case BPF_JMP | BPF_JSLT | BPF_X: /* ((s64) dst < (s64) src) */
1357 case BPF_JMP32 | BPF_JSLT | BPF_X: /* ((s32) dst < (s32) src) */
1358 mask = 0x4000; /* jl */
1359 goto branch_xs;
1360 case BPF_JMP | BPF_JSGE | BPF_X: /* ((s64) dst >= (s64) src) */
1361 case BPF_JMP32 | BPF_JSGE | BPF_X: /* ((s32) dst >= (s32) src) */
1362 mask = 0xa000; /* jhe */
1363 goto branch_xs;
1364 case BPF_JMP | BPF_JSLE | BPF_X: /* ((s64) dst <= (s64) src) */
1365 case BPF_JMP32 | BPF_JSLE | BPF_X: /* ((s32) dst <= (s32) src) */
1366 mask = 0xc000; /* jle */
1367 goto branch_xs;
1368 case BPF_JMP | BPF_JGT | BPF_X: /* (dst > src) */
1369 case BPF_JMP32 | BPF_JGT | BPF_X: /* ((u32) dst > (u32) src) */
1370 mask = 0x2000; /* jh */
1371 goto branch_xu;
1372 case BPF_JMP | BPF_JLT | BPF_X: /* (dst < src) */
1373 case BPF_JMP32 | BPF_JLT | BPF_X: /* ((u32) dst < (u32) src) */
1374 mask = 0x4000; /* jl */
1375 goto branch_xu;
1376 case BPF_JMP | BPF_JGE | BPF_X: /* (dst >= src) */
1377 case BPF_JMP32 | BPF_JGE | BPF_X: /* ((u32) dst >= (u32) src) */
1378 mask = 0xa000; /* jhe */
1379 goto branch_xu;
1380 case BPF_JMP | BPF_JLE | BPF_X: /* (dst <= src) */
1381 case BPF_JMP32 | BPF_JLE | BPF_X: /* ((u32) dst <= (u32) src) */
1382 mask = 0xc000; /* jle */
1383 goto branch_xu;
1384 case BPF_JMP | BPF_JNE | BPF_X: /* (dst != src) */
1385 case BPF_JMP32 | BPF_JNE | BPF_X: /* ((u32) dst != (u32) src) */
1386 mask = 0x7000; /* jne */
1387 goto branch_xu;
1388 case BPF_JMP | BPF_JEQ | BPF_X: /* (dst == src) */
1389 case BPF_JMP32 | BPF_JEQ | BPF_X: /* ((u32) dst == (u32) src) */
1390 mask = 0x8000; /* je */
1391 goto branch_xu;
1392 case BPF_JMP | BPF_JSET | BPF_X: /* (dst & src) */
1393 case BPF_JMP32 | BPF_JSET | BPF_X: /* ((u32) dst & (u32) src) */
1394 {
1395 bool is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
1396
1397 mask = 0x7000; /* jnz */
1398 /* nrk or ngrk %w1,%dst,%src */
1399 EMIT4_RRF((is_jmp32 ? 0xb9f40000 : 0xb9e40000),
1400 REG_W1, dst_reg, src_reg);
1401 goto branch_oc;
1402 branch_ks:
1403 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
1404 /* cfi or cgfi %dst,imm */
1405 EMIT6_IMM(is_jmp32 ? 0xc20d0000 : 0xc20c0000,
1406 dst_reg, imm);
1407 if (!is_first_pass(jit) &&
1408 can_use_rel(jit, addrs[i + off + 1])) {
1409 /* brc mask,off */
1410 EMIT4_PCREL_RIC(0xa7040000,
1411 mask >> 12, addrs[i + off + 1]);
1412 } else {
1413 /* brcl mask,off */
1414 EMIT6_PCREL_RILC(0xc0040000,
1415 mask >> 12, addrs[i + off + 1]);
1416 }
1417 break;
1418 branch_ku:
1419 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
1420 /* clfi or clgfi %dst,imm */
1421 EMIT6_IMM(is_jmp32 ? 0xc20f0000 : 0xc20e0000,
1422 dst_reg, imm);
1423 if (!is_first_pass(jit) &&
1424 can_use_rel(jit, addrs[i + off + 1])) {
1425 /* brc mask,off */
1426 EMIT4_PCREL_RIC(0xa7040000,
1427 mask >> 12, addrs[i + off + 1]);
1428 } else {
1429 /* brcl mask,off */
1430 EMIT6_PCREL_RILC(0xc0040000,
1431 mask >> 12, addrs[i + off + 1]);
1432 }
1433 break;
1434 branch_xs:
1435 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
1436 if (!is_first_pass(jit) &&
1437 can_use_rel(jit, addrs[i + off + 1])) {
1438 /* crj or cgrj %dst,%src,mask,off */
1439 EMIT6_PCREL(0xec000000, (is_jmp32 ? 0x0076 : 0x0064),
1440 dst_reg, src_reg, i, off, mask);
1441 } else {
1442 /* cr or cgr %dst,%src */
1443 if (is_jmp32)
1444 EMIT2(0x1900, dst_reg, src_reg);
1445 else
1446 EMIT4(0xb9200000, dst_reg, src_reg);
1447 /* brcl mask,off */
1448 EMIT6_PCREL_RILC(0xc0040000,
1449 mask >> 12, addrs[i + off + 1]);
1450 }
1451 break;
1452 branch_xu:
1453 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
1454 if (!is_first_pass(jit) &&
1455 can_use_rel(jit, addrs[i + off + 1])) {
1456 /* clrj or clgrj %dst,%src,mask,off */
1457 EMIT6_PCREL(0xec000000, (is_jmp32 ? 0x0077 : 0x0065),
1458 dst_reg, src_reg, i, off, mask);
1459 } else {
1460 /* clr or clgr %dst,%src */
1461 if (is_jmp32)
1462 EMIT2(0x1500, dst_reg, src_reg);
1463 else
1464 EMIT4(0xb9210000, dst_reg, src_reg);
1465 /* brcl mask,off */
1466 EMIT6_PCREL_RILC(0xc0040000,
1467 mask >> 12, addrs[i + off + 1]);
1468 }
1469 break;
1470 branch_oc:
1471 if (!is_first_pass(jit) &&
1472 can_use_rel(jit, addrs[i + off + 1])) {
1473 /* brc mask,off */
1474 EMIT4_PCREL_RIC(0xa7040000,
1475 mask >> 12, addrs[i + off + 1]);
1476 } else {
1477 /* brcl mask,off */
1478 EMIT6_PCREL_RILC(0xc0040000,
1479 mask >> 12, addrs[i + off + 1]);
1480 }
1481 break;
1482 }
1483 default: /* too complex, give up */
1484 pr_err("Unknown opcode %02x\n", insn->code);
1485 return -1;
1486 }
1487 return insn_count;
1488 }
1489
1490 /*
1491 * Return whether new i-th instruction address does not violate any invariant
1492 */
1493 static bool bpf_is_new_addr_sane(struct bpf_jit *jit, int i)
1494 {
1495 /* On the first pass anything goes */
1496 if (is_first_pass(jit))
1497 return true;
1498
1499 /* The codegen pass must not change anything */
1500 if (is_codegen_pass(jit))
1501 return jit->addrs[i] == jit->prg;
1502
1503 /* Passes in between must not increase code size */
1504 return jit->addrs[i] >= jit->prg;
1505 }
1506
1507 /*
1508 * Update the address of i-th instruction
1509 */
1510 static int bpf_set_addr(struct bpf_jit *jit, int i)
1511 {
1512 if (!bpf_is_new_addr_sane(jit, i))
1513 return -1;
1514 jit->addrs[i] = jit->prg;
1515 return 0;
1516 }
1517
1518 /*
1519 * Compile eBPF program into s390x code
1520 */
1521 static int bpf_jit_prog(struct bpf_jit *jit, struct bpf_prog *fp,
1522 bool extra_pass, u32 stack_depth)
1523 {
1524 int i, insn_count, lit32_size, lit64_size;
1525
1526 jit->lit32 = jit->lit32_start;
1527 jit->lit64 = jit->lit64_start;
1528 jit->prg = 0;
1529
1530 bpf_jit_prologue(jit, stack_depth);
1531 if (bpf_set_addr(jit, 0) < 0)
1532 return -1;
1533 for (i = 0; i < fp->len; i += insn_count) {
1534 insn_count = bpf_jit_insn(jit, fp, i, extra_pass, stack_depth);
1535 if (insn_count < 0)
1536 return -1;
1537 /* Next instruction address */
1538 if (bpf_set_addr(jit, i + insn_count) < 0)
1539 return -1;
1540 }
1541 bpf_jit_epilogue(jit, stack_depth);
1542
1543 lit32_size = jit->lit32 - jit->lit32_start;
1544 lit64_size = jit->lit64 - jit->lit64_start;
1545 jit->lit32_start = jit->prg;
1546 if (lit32_size)
1547 jit->lit32_start = ALIGN(jit->lit32_start, 4);
1548 jit->lit64_start = jit->lit32_start + lit32_size;
1549 if (lit64_size)
1550 jit->lit64_start = ALIGN(jit->lit64_start, 8);
1551 jit->size = jit->lit64_start + lit64_size;
1552 jit->size_prg = jit->prg;
1553 return 0;
1554 }
1555
1556 bool bpf_jit_needs_zext(void)
1557 {
1558 return true;
1559 }
1560
1561 struct s390_jit_data {
1562 struct bpf_binary_header *header;
1563 struct bpf_jit ctx;
1564 int pass;
1565 };
1566
1567 /*
1568 * Compile eBPF program "fp"
1569 */
1570 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
1571 {
1572 u32 stack_depth = round_up(fp->aux->stack_depth, 8);
1573 struct bpf_prog *tmp, *orig_fp = fp;
1574 struct bpf_binary_header *header;
1575 struct s390_jit_data *jit_data;
1576 bool tmp_blinded = false;
1577 bool extra_pass = false;
1578 struct bpf_jit jit;
1579 int pass;
1580
1581 if (!fp->jit_requested)
1582 return orig_fp;
1583
1584 tmp = bpf_jit_blind_constants(fp);
1585 /*
1586 * If blinding was requested and we failed during blinding,
1587 * we must fall back to the interpreter.
1588 */
1589 if (IS_ERR(tmp))
1590 return orig_fp;
1591 if (tmp != fp) {
1592 tmp_blinded = true;
1593 fp = tmp;
1594 }
1595
1596 jit_data = fp->aux->jit_data;
1597 if (!jit_data) {
1598 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1599 if (!jit_data) {
1600 fp = orig_fp;
1601 goto out;
1602 }
1603 fp->aux->jit_data = jit_data;
1604 }
1605 if (jit_data->ctx.addrs) {
1606 jit = jit_data->ctx;
1607 header = jit_data->header;
1608 extra_pass = true;
1609 pass = jit_data->pass + 1;
1610 goto skip_init_ctx;
1611 }
1612
1613 memset(&jit, 0, sizeof(jit));
1614 jit.addrs = kvcalloc(fp->len + 1, sizeof(*jit.addrs), GFP_KERNEL);
1615 if (jit.addrs == NULL) {
1616 fp = orig_fp;
1617 goto out;
1618 }
1619 /*
1620 * Three initial passes:
1621 * - 1/2: Determine clobbered registers
1622 * - 3: Calculate program size and addrs arrray
1623 */
1624 for (pass = 1; pass <= 3; pass++) {
1625 if (bpf_jit_prog(&jit, fp, extra_pass, stack_depth)) {
1626 fp = orig_fp;
1627 goto free_addrs;
1628 }
1629 }
1630 /*
1631 * Final pass: Allocate and generate program
1632 */
1633 header = bpf_jit_binary_alloc(jit.size, &jit.prg_buf, 8, jit_fill_hole);
1634 if (!header) {
1635 fp = orig_fp;
1636 goto free_addrs;
1637 }
1638 skip_init_ctx:
1639 if (bpf_jit_prog(&jit, fp, extra_pass, stack_depth)) {
1640 bpf_jit_binary_free(header);
1641 fp = orig_fp;
1642 goto free_addrs;
1643 }
1644 if (bpf_jit_enable > 1) {
1645 bpf_jit_dump(fp->len, jit.size, pass, jit.prg_buf);
1646 print_fn_code(jit.prg_buf, jit.size_prg);
1647 }
1648 if (!fp->is_func || extra_pass) {
1649 bpf_jit_binary_lock_ro(header);
1650 } else {
1651 jit_data->header = header;
1652 jit_data->ctx = jit;
1653 jit_data->pass = pass;
1654 }
1655 fp->bpf_func = (void *) jit.prg_buf;
1656 fp->jited = 1;
1657 fp->jited_len = jit.size;
1658
1659 if (!fp->is_func || extra_pass) {
1660 bpf_prog_fill_jited_linfo(fp, jit.addrs + 1);
1661 free_addrs:
1662 kvfree(jit.addrs);
1663 kfree(jit_data);
1664 fp->aux->jit_data = NULL;
1665 }
1666 out:
1667 if (tmp_blinded)
1668 bpf_jit_prog_release_other(fp, fp == orig_fp ?
1669 tmp : orig_fp);
1670 return fp;
1671 }
Linux with added FPC-III support