?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
printf: Remove unused 'bprintf'
[drm/drm-misc.git] / arch / s390 / net / bpf_jit_comp.c
blob9d440a0b729eb71337f6824e9f8e49970b2ef63e
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 * - 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 <linux/mm.h>
26 #include <linux/kernel.h>
27 #include <asm/cacheflush.h>
28 #include <asm/extable.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 <asm/text-patching.h>
34 #include <asm/unwind.h>
35 #include "bpf_jit.h"
36
37 struct bpf_jit {
38 u32 seen; /* Flags to remember seen eBPF instructions */
39 u16 seen_regs; /* Mask to remember which registers are used */
40 u32 *addrs; /* Array with relative instruction addresses */
41 u8 *prg_buf; /* Start of program */
42 int size; /* Size of program and literal pool */
43 int size_prg; /* Size of program */
44 int prg; /* Current position in program */
45 int lit32_start; /* Start of 32-bit literal pool */
46 int lit32; /* Current position in 32-bit literal pool */
47 int lit64_start; /* Start of 64-bit literal pool */
48 int lit64; /* Current position in 64-bit literal pool */
49 int base_ip; /* Base address for literal pool */
50 int exit_ip; /* Address of exit */
51 int r1_thunk_ip; /* Address of expoline thunk for 'br %r1' */
52 int r14_thunk_ip; /* Address of expoline thunk for 'br %r14' */
53 int tail_call_start; /* Tail call start offset */
54 int excnt; /* Number of exception table entries */
55 int prologue_plt_ret; /* Return address for prologue hotpatch PLT */
56 int prologue_plt; /* Start of prologue hotpatch PLT */
57 int kern_arena; /* Pool offset of kernel arena address */
58 u64 user_arena; /* User arena address */
59 };
60
61 #define SEEN_MEM BIT(0) /* use mem[] for temporary storage */
62 #define SEEN_LITERAL BIT(1) /* code uses literals */
63 #define SEEN_FUNC BIT(2) /* calls C functions */
64 #define SEEN_STACK (SEEN_FUNC | SEEN_MEM)
65
66 #define NVREGS 0xffc0 /* %r6-%r15 */
67
68 /*
69 * s390 registers
70 */
71 #define REG_W0 (MAX_BPF_JIT_REG + 0) /* Work register 1 (even) */
72 #define REG_W1 (MAX_BPF_JIT_REG + 1) /* Work register 2 (odd) */
73 #define REG_L (MAX_BPF_JIT_REG + 2) /* Literal pool register */
74 #define REG_15 (MAX_BPF_JIT_REG + 3) /* Register 15 */
75 #define REG_0 REG_W0 /* Register 0 */
76 #define REG_1 REG_W1 /* Register 1 */
77 #define REG_2 BPF_REG_1 /* Register 2 */
78 #define REG_3 BPF_REG_2 /* Register 3 */
79 #define REG_4 BPF_REG_3 /* Register 4 */
80 #define REG_7 BPF_REG_6 /* Register 7 */
81 #define REG_8 BPF_REG_7 /* Register 8 */
82 #define REG_14 BPF_REG_0 /* Register 14 */
83
84 /*
85 * Mapping of BPF registers to s390 registers
86 */
87 static const int reg2hex[] = {
88 /* Return code */
89 [BPF_REG_0] = 14,
90 /* Function parameters */
91 [BPF_REG_1] = 2,
92 [BPF_REG_2] = 3,
93 [BPF_REG_3] = 4,
94 [BPF_REG_4] = 5,
95 [BPF_REG_5] = 6,
96 /* Call saved registers */
97 [BPF_REG_6] = 7,
98 [BPF_REG_7] = 8,
99 [BPF_REG_8] = 9,
100 [BPF_REG_9] = 10,
101 /* BPF stack pointer */
102 [BPF_REG_FP] = 13,
103 /* Register for blinding */
104 [BPF_REG_AX] = 12,
105 /* Work registers for s390x backend */
106 [REG_W0] = 0,
107 [REG_W1] = 1,
108 [REG_L] = 11,
109 [REG_15] = 15,
110 };
111
112 static inline u32 reg(u32 dst_reg, u32 src_reg)
113 {
114 return reg2hex[dst_reg] << 4 | reg2hex[src_reg];
115 }
116
117 static inline u32 reg_high(u32 reg)
118 {
119 return reg2hex[reg] << 4;
120 }
121
122 static inline void reg_set_seen(struct bpf_jit *jit, u32 b1)
123 {
124 u32 r1 = reg2hex[b1];
125
126 if (r1 >= 6 && r1 <= 15)
127 jit->seen_regs |= (1 << r1);
128 }
129
130 #define REG_SET_SEEN(b1) \
131 ({ \
132 reg_set_seen(jit, b1); \
133 })
134
135 /*
136 * EMIT macros for code generation
137 */
138
139 #define _EMIT2(op) \
140 ({ \
141 if (jit->prg_buf) \
142 *(u16 *) (jit->prg_buf + jit->prg) = (op); \
143 jit->prg += 2; \
144 })
145
146 #define EMIT2(op, b1, b2) \
147 ({ \
148 _EMIT2((op) | reg(b1, b2)); \
149 REG_SET_SEEN(b1); \
150 REG_SET_SEEN(b2); \
151 })
152
153 #define _EMIT4(op) \
154 ({ \
155 if (jit->prg_buf) \
156 *(u32 *) (jit->prg_buf + jit->prg) = (op); \
157 jit->prg += 4; \
158 })
159
160 #define EMIT4(op, b1, b2) \
161 ({ \
162 _EMIT4((op) | reg(b1, b2)); \
163 REG_SET_SEEN(b1); \
164 REG_SET_SEEN(b2); \
165 })
166
167 #define EMIT4_RRF(op, b1, b2, b3) \
168 ({ \
169 _EMIT4((op) | reg_high(b3) << 8 | reg(b1, b2)); \
170 REG_SET_SEEN(b1); \
171 REG_SET_SEEN(b2); \
172 REG_SET_SEEN(b3); \
173 })
174
175 #define _EMIT4_DISP(op, disp) \
176 ({ \
177 unsigned int __disp = (disp) & 0xfff; \
178 _EMIT4((op) | __disp); \
179 })
180
181 #define EMIT4_DISP(op, b1, b2, disp) \
182 ({ \
183 _EMIT4_DISP((op) | reg_high(b1) << 16 | \
184 reg_high(b2) << 8, (disp)); \
185 REG_SET_SEEN(b1); \
186 REG_SET_SEEN(b2); \
187 })
188
189 #define EMIT4_IMM(op, b1, imm) \
190 ({ \
191 unsigned int __imm = (imm) & 0xffff; \
192 _EMIT4((op) | reg_high(b1) << 16 | __imm); \
193 REG_SET_SEEN(b1); \
194 })
195
196 #define EMIT4_PCREL(op, pcrel) \
197 ({ \
198 long __pcrel = ((pcrel) >> 1) & 0xffff; \
199 _EMIT4((op) | __pcrel); \
200 })
201
202 #define EMIT4_PCREL_RIC(op, mask, target) \
203 ({ \
204 int __rel = ((target) - jit->prg) / 2; \
205 _EMIT4((op) | (mask) << 20 | (__rel & 0xffff)); \
206 })
207
208 #define _EMIT6(op1, op2) \
209 ({ \
210 if (jit->prg_buf) { \
211 *(u32 *) (jit->prg_buf + jit->prg) = (op1); \
212 *(u16 *) (jit->prg_buf + jit->prg + 4) = (op2); \
213 } \
214 jit->prg += 6; \
215 })
216
217 #define _EMIT6_DISP(op1, op2, disp) \
218 ({ \
219 unsigned int __disp = (disp) & 0xfff; \
220 _EMIT6((op1) | __disp, op2); \
221 })
222
223 #define _EMIT6_DISP_LH(op1, op2, disp) \
224 ({ \
225 u32 _disp = (u32) (disp); \
226 unsigned int __disp_h = _disp & 0xff000; \
227 unsigned int __disp_l = _disp & 0x00fff; \
228 _EMIT6((op1) | __disp_l, (op2) | __disp_h >> 4); \
229 })
230
231 #define EMIT6_DISP_LH(op1, op2, b1, b2, b3, disp) \
232 ({ \
233 _EMIT6_DISP_LH((op1) | reg(b1, b2) << 16 | \
234 reg_high(b3) << 8, op2, disp); \
235 REG_SET_SEEN(b1); \
236 REG_SET_SEEN(b2); \
237 REG_SET_SEEN(b3); \
238 })
239
240 #define EMIT6_PCREL_RIEB(op1, op2, b1, b2, mask, target) \
241 ({ \
242 unsigned int rel = (int)((target) - jit->prg) / 2; \
243 _EMIT6((op1) | reg(b1, b2) << 16 | (rel & 0xffff), \
244 (op2) | (mask) << 12); \
245 REG_SET_SEEN(b1); \
246 REG_SET_SEEN(b2); \
247 })
248
249 #define EMIT6_PCREL_RIEC(op1, op2, b1, imm, mask, target) \
250 ({ \
251 unsigned int rel = (int)((target) - jit->prg) / 2; \
252 _EMIT6((op1) | (reg_high(b1) | (mask)) << 16 | \
253 (rel & 0xffff), (op2) | ((imm) & 0xff) << 8); \
254 REG_SET_SEEN(b1); \
255 BUILD_BUG_ON(((unsigned long) (imm)) > 0xff); \
256 })
257
258 #define EMIT6_PCREL(op1, op2, b1, b2, i, off, mask) \
259 ({ \
260 int rel = (addrs[(i) + (off) + 1] - jit->prg) / 2; \
261 _EMIT6((op1) | reg(b1, b2) << 16 | (rel & 0xffff), (op2) | (mask));\
262 REG_SET_SEEN(b1); \
263 REG_SET_SEEN(b2); \
264 })
265
266 #define EMIT6_PCREL_RILB(op, b, target) \
267 ({ \
268 unsigned int rel = (int)((target) - jit->prg) / 2; \
269 _EMIT6((op) | reg_high(b) << 16 | rel >> 16, rel & 0xffff);\
270 REG_SET_SEEN(b); \
271 })
272
273 #define EMIT6_PCREL_RIL(op, target) \
274 ({ \
275 unsigned int rel = (int)((target) - jit->prg) / 2; \
276 _EMIT6((op) | rel >> 16, rel & 0xffff); \
277 })
278
279 #define EMIT6_PCREL_RILC(op, mask, target) \
280 ({ \
281 EMIT6_PCREL_RIL((op) | (mask) << 20, (target)); \
282 })
283
284 #define _EMIT6_IMM(op, imm) \
285 ({ \
286 unsigned int __imm = (imm); \
287 _EMIT6((op) | (__imm >> 16), __imm & 0xffff); \
288 })
289
290 #define EMIT6_IMM(op, b1, imm) \
291 ({ \
292 _EMIT6_IMM((op) | reg_high(b1) << 16, imm); \
293 REG_SET_SEEN(b1); \
294 })
295
296 #define _EMIT_CONST_U32(val) \
297 ({ \
298 unsigned int ret; \
299 ret = jit->lit32; \
300 if (jit->prg_buf) \
301 *(u32 *)(jit->prg_buf + jit->lit32) = (u32)(val);\
302 jit->lit32 += 4; \
303 ret; \
304 })
305
306 #define EMIT_CONST_U32(val) \
307 ({ \
308 jit->seen |= SEEN_LITERAL; \
309 _EMIT_CONST_U32(val) - jit->base_ip; \
310 })
311
312 #define _EMIT_CONST_U64(val) \
313 ({ \
314 unsigned int ret; \
315 ret = jit->lit64; \
316 if (jit->prg_buf) \
317 *(u64 *)(jit->prg_buf + jit->lit64) = (u64)(val);\
318 jit->lit64 += 8; \
319 ret; \
320 })
321
322 #define EMIT_CONST_U64(val) \
323 ({ \
324 jit->seen |= SEEN_LITERAL; \
325 _EMIT_CONST_U64(val) - jit->base_ip; \
326 })
327
328 #define EMIT_ZERO(b1) \
329 ({ \
330 if (!fp->aux->verifier_zext) { \
331 /* llgfr %dst,%dst (zero extend to 64 bit) */ \
332 EMIT4(0xb9160000, b1, b1); \
333 REG_SET_SEEN(b1); \
334 } \
335 })
336
337 /*
338 * Return whether this is the first pass. The first pass is special, since we
339 * don't know any sizes yet, and thus must be conservative.
340 */
341 static bool is_first_pass(struct bpf_jit *jit)
342 {
343 return jit->size == 0;
344 }
345
346 /*
347 * Return whether this is the code generation pass. The code generation pass is
348 * special, since we should change as little as possible.
349 */
350 static bool is_codegen_pass(struct bpf_jit *jit)
351 {
352 return jit->prg_buf;
353 }
354
355 /*
356 * Return whether "rel" can be encoded as a short PC-relative offset
357 */
358 static bool is_valid_rel(int rel)
359 {
360 return rel >= -65536 && rel <= 65534;
361 }
362
363 /*
364 * Return whether "off" can be reached using a short PC-relative offset
365 */
366 static bool can_use_rel(struct bpf_jit *jit, int off)
367 {
368 return is_valid_rel(off - jit->prg);
369 }
370
371 /*
372 * Return whether given displacement can be encoded using
373 * Long-Displacement Facility
374 */
375 static bool is_valid_ldisp(int disp)
376 {
377 return disp >= -524288 && disp <= 524287;
378 }
379
380 /*
381 * Return whether the next 32-bit literal pool entry can be referenced using
382 * Long-Displacement Facility
383 */
384 static bool can_use_ldisp_for_lit32(struct bpf_jit *jit)
385 {
386 return is_valid_ldisp(jit->lit32 - jit->base_ip);
387 }
388
389 /*
390 * Return whether the next 64-bit literal pool entry can be referenced using
391 * Long-Displacement Facility
392 */
393 static bool can_use_ldisp_for_lit64(struct bpf_jit *jit)
394 {
395 return is_valid_ldisp(jit->lit64 - jit->base_ip);
396 }
397
398 /*
399 * Fill whole space with illegal instructions
400 */
401 static void jit_fill_hole(void *area, unsigned int size)
402 {
403 memset(area, 0, size);
404 }
405
406 /*
407 * Save registers from "rs" (register start) to "re" (register end) on stack
408 */
409 static void save_regs(struct bpf_jit *jit, u32 rs, u32 re)
410 {
411 u32 off = STK_OFF_R6 + (rs - 6) * 8;
412
413 if (rs == re)
414 /* stg %rs,off(%r15) */
415 _EMIT6(0xe300f000 | rs << 20 | off, 0x0024);
416 else
417 /* stmg %rs,%re,off(%r15) */
418 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0024, off);
419 }
420
421 /*
422 * Restore registers from "rs" (register start) to "re" (register end) on stack
423 */
424 static void restore_regs(struct bpf_jit *jit, u32 rs, u32 re, u32 stack_depth)
425 {
426 u32 off = STK_OFF_R6 + (rs - 6) * 8;
427
428 if (jit->seen & SEEN_STACK)
429 off += STK_OFF + stack_depth;
430
431 if (rs == re)
432 /* lg %rs,off(%r15) */
433 _EMIT6(0xe300f000 | rs << 20 | off, 0x0004);
434 else
435 /* lmg %rs,%re,off(%r15) */
436 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0004, off);
437 }
438
439 /*
440 * Return first seen register (from start)
441 */
442 static int get_start(u16 seen_regs, int start)
443 {
444 int i;
445
446 for (i = start; i <= 15; i++) {
447 if (seen_regs & (1 << i))
448 return i;
449 }
450 return 0;
451 }
452
453 /*
454 * Return last seen register (from start) (gap >= 2)
455 */
456 static int get_end(u16 seen_regs, int start)
457 {
458 int i;
459
460 for (i = start; i < 15; i++) {
461 if (!(seen_regs & (3 << i)))
462 return i - 1;
463 }
464 return (seen_regs & (1 << 15)) ? 15 : 14;
465 }
466
467 #define REGS_SAVE 1
468 #define REGS_RESTORE 0
469 /*
470 * Save and restore clobbered registers (6-15) on stack.
471 * We save/restore registers in chunks with gap >= 2 registers.
472 */
473 static void save_restore_regs(struct bpf_jit *jit, int op, u32 stack_depth,
474 u16 extra_regs)
475 {
476 u16 seen_regs = jit->seen_regs | extra_regs;
477 const int last = 15, save_restore_size = 6;
478 int re = 6, rs;
479
480 if (is_first_pass(jit)) {
481 /*
482 * We don't know yet which registers are used. Reserve space
483 * conservatively.
484 */
485 jit->prg += (last - re + 1) * save_restore_size;
486 return;
487 }
488
489 do {
490 rs = get_start(seen_regs, re);
491 if (!rs)
492 break;
493 re = get_end(seen_regs, rs + 1);
494 if (op == REGS_SAVE)
495 save_regs(jit, rs, re);
496 else
497 restore_regs(jit, rs, re, stack_depth);
498 re++;
499 } while (re <= last);
500 }
501
502 static void bpf_skip(struct bpf_jit *jit, int size)
503 {
504 if (size >= 6 && !is_valid_rel(size)) {
505 /* brcl 0xf,size */
506 EMIT6_PCREL_RIL(0xc0f4000000, size);
507 size -= 6;
508 } else if (size >= 4 && is_valid_rel(size)) {
509 /* brc 0xf,size */
510 EMIT4_PCREL(0xa7f40000, size);
511 size -= 4;
512 }
513 while (size >= 2) {
514 /* bcr 0,%0 */
515 _EMIT2(0x0700);
516 size -= 2;
517 }
518 }
519
520 /*
521 * PLT for hotpatchable calls. The calling convention is the same as for the
522 * ftrace hotpatch trampolines: %r0 is return address, %r1 is clobbered.
523 */
524 struct bpf_plt {
525 char code[16];
526 void *ret;
527 void *target;
528 } __packed;
529 extern const struct bpf_plt bpf_plt;
530 asm(
531 ".pushsection .rodata\n"
532 " .balign 8\n"
533 "bpf_plt:\n"
534 " lgrl %r0,bpf_plt_ret\n"
535 " lgrl %r1,bpf_plt_target\n"
536 " br %r1\n"
537 " .balign 8\n"
538 "bpf_plt_ret: .quad 0\n"
539 "bpf_plt_target: .quad 0\n"
540 " .popsection\n"
541 );
542
543 static void bpf_jit_plt(struct bpf_plt *plt, void *ret, void *target)
544 {
545 memcpy(plt, &bpf_plt, sizeof(*plt));
546 plt->ret = ret;
547 plt->target = target;
548 }
549
550 /*
551 * Emit function prologue
552 *
553 * Save registers and create stack frame if necessary.
554 * See stack frame layout description in "bpf_jit.h"!
555 */
556 static void bpf_jit_prologue(struct bpf_jit *jit, struct bpf_prog *fp,
557 u32 stack_depth)
558 {
559 /* No-op for hotpatching */
560 /* brcl 0,prologue_plt */
561 EMIT6_PCREL_RILC(0xc0040000, 0, jit->prologue_plt);
562 jit->prologue_plt_ret = jit->prg;
563
564 if (!bpf_is_subprog(fp)) {
565 /* Initialize the tail call counter in the main program. */
566 /* xc STK_OFF_TCCNT(4,%r15),STK_OFF_TCCNT(%r15) */
567 _EMIT6(0xd703f000 | STK_OFF_TCCNT, 0xf000 | STK_OFF_TCCNT);
568 } else {
569 /*
570 * Skip the tail call counter initialization in subprograms.
571 * Insert nops in order to have tail_call_start at a
572 * predictable offset.
573 */
574 bpf_skip(jit, 6);
575 }
576 /* Tail calls have to skip above initialization */
577 jit->tail_call_start = jit->prg;
578 if (fp->aux->exception_cb) {
579 /*
580 * Switch stack, the new address is in the 2nd parameter.
581 *
582 * Arrange the restoration of %r6-%r15 in the epilogue.
583 * Do not restore them now, the prog does not need them.
584 */
585 /* lgr %r15,%r3 */
586 EMIT4(0xb9040000, REG_15, REG_3);
587 jit->seen_regs |= NVREGS;
588 } else {
589 /* Save registers */
590 save_restore_regs(jit, REGS_SAVE, stack_depth,
591 fp->aux->exception_boundary ? NVREGS : 0);
592 }
593 /* Setup literal pool */
594 if (is_first_pass(jit) || (jit->seen & SEEN_LITERAL)) {
595 if (!is_first_pass(jit) &&
596 is_valid_ldisp(jit->size - (jit->prg + 2))) {
597 /* basr %l,0 */
598 EMIT2(0x0d00, REG_L, REG_0);
599 jit->base_ip = jit->prg;
600 } else {
601 /* larl %l,lit32_start */
602 EMIT6_PCREL_RILB(0xc0000000, REG_L, jit->lit32_start);
603 jit->base_ip = jit->lit32_start;
604 }
605 }
606 /* Setup stack and backchain */
607 if (is_first_pass(jit) || (jit->seen & SEEN_STACK)) {
608 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC))
609 /* lgr %w1,%r15 (backchain) */
610 EMIT4(0xb9040000, REG_W1, REG_15);
611 /* la %bfp,STK_160_UNUSED(%r15) (BPF frame pointer) */
612 EMIT4_DISP(0x41000000, BPF_REG_FP, REG_15, STK_160_UNUSED);
613 /* aghi %r15,-STK_OFF */
614 EMIT4_IMM(0xa70b0000, REG_15, -(STK_OFF + stack_depth));
615 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC))
616 /* stg %w1,152(%r15) (backchain) */
617 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W1, REG_0,
618 REG_15, 152);
619 }
620 }
621
622 /*
623 * Emit an expoline for a jump that follows
624 */
625 static void emit_expoline(struct bpf_jit *jit)
626 {
627 /* exrl %r0,.+10 */
628 EMIT6_PCREL_RIL(0xc6000000, jit->prg + 10);
629 /* j . */
630 EMIT4_PCREL(0xa7f40000, 0);
631 }
632
633 /*
634 * Emit __s390_indirect_jump_r1 thunk if necessary
635 */
636 static void emit_r1_thunk(struct bpf_jit *jit)
637 {
638 if (nospec_uses_trampoline()) {
639 jit->r1_thunk_ip = jit->prg;
640 emit_expoline(jit);
641 /* br %r1 */
642 _EMIT2(0x07f1);
643 }
644 }
645
646 /*
647 * Call r1 either directly or via __s390_indirect_jump_r1 thunk
648 */
649 static void call_r1(struct bpf_jit *jit)
650 {
651 if (nospec_uses_trampoline())
652 /* brasl %r14,__s390_indirect_jump_r1 */
653 EMIT6_PCREL_RILB(0xc0050000, REG_14, jit->r1_thunk_ip);
654 else
655 /* basr %r14,%r1 */
656 EMIT2(0x0d00, REG_14, REG_1);
657 }
658
659 /*
660 * Function epilogue
661 */
662 static void bpf_jit_epilogue(struct bpf_jit *jit, u32 stack_depth)
663 {
664 jit->exit_ip = jit->prg;
665 /* Load exit code: lgr %r2,%b0 */
666 EMIT4(0xb9040000, REG_2, BPF_REG_0);
667 /* Restore registers */
668 save_restore_regs(jit, REGS_RESTORE, stack_depth, 0);
669 if (nospec_uses_trampoline()) {
670 jit->r14_thunk_ip = jit->prg;
671 /* Generate __s390_indirect_jump_r14 thunk */
672 emit_expoline(jit);
673 }
674 /* br %r14 */
675 _EMIT2(0x07fe);
676
677 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC))
678 emit_r1_thunk(jit);
679
680 jit->prg = ALIGN(jit->prg, 8);
681 jit->prologue_plt = jit->prg;
682 if (jit->prg_buf)
683 bpf_jit_plt((struct bpf_plt *)(jit->prg_buf + jit->prg),
684 jit->prg_buf + jit->prologue_plt_ret, NULL);
685 jit->prg += sizeof(struct bpf_plt);
686 }
687
688 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
689 {
690 regs->psw.addr = extable_fixup(x);
691 if (x->data != -1)
692 regs->gprs[x->data] = 0;
693 return true;
694 }
695
696 /*
697 * A single BPF probe instruction
698 */
699 struct bpf_jit_probe {
700 int prg; /* JITed instruction offset */
701 int nop_prg; /* JITed nop offset */
702 int reg; /* Register to clear on exception */
703 int arena_reg; /* Register to use for arena addressing */
704 };
705
706 static void bpf_jit_probe_init(struct bpf_jit_probe *probe)
707 {
708 probe->prg = -1;
709 probe->nop_prg = -1;
710 probe->reg = -1;
711 probe->arena_reg = REG_0;
712 }
713
714 /*
715 * Handlers of certain exceptions leave psw.addr pointing to the instruction
716 * directly after the failing one. Therefore, create two exception table
717 * entries and also add a nop in case two probing instructions come directly
718 * after each other.
719 */
720 static void bpf_jit_probe_emit_nop(struct bpf_jit *jit,
721 struct bpf_jit_probe *probe)
722 {
723 if (probe->prg == -1 || probe->nop_prg != -1)
724 /* The probe is not armed or nop is already emitted. */
725 return;
726
727 probe->nop_prg = jit->prg;
728 /* bcr 0,%0 */
729 _EMIT2(0x0700);
730 }
731
732 static void bpf_jit_probe_load_pre(struct bpf_jit *jit, struct bpf_insn *insn,
733 struct bpf_jit_probe *probe)
734 {
735 if (BPF_MODE(insn->code) != BPF_PROBE_MEM &&
736 BPF_MODE(insn->code) != BPF_PROBE_MEMSX &&
737 BPF_MODE(insn->code) != BPF_PROBE_MEM32)
738 return;
739
740 if (BPF_MODE(insn->code) == BPF_PROBE_MEM32) {
741 /* lgrl %r1,kern_arena */
742 EMIT6_PCREL_RILB(0xc4080000, REG_W1, jit->kern_arena);
743 probe->arena_reg = REG_W1;
744 }
745 probe->prg = jit->prg;
746 probe->reg = reg2hex[insn->dst_reg];
747 }
748
749 static void bpf_jit_probe_store_pre(struct bpf_jit *jit, struct bpf_insn *insn,
750 struct bpf_jit_probe *probe)
751 {
752 if (BPF_MODE(insn->code) != BPF_PROBE_MEM32)
753 return;
754
755 /* lgrl %r1,kern_arena */
756 EMIT6_PCREL_RILB(0xc4080000, REG_W1, jit->kern_arena);
757 probe->arena_reg = REG_W1;
758 probe->prg = jit->prg;
759 }
760
761 static void bpf_jit_probe_atomic_pre(struct bpf_jit *jit,
762 struct bpf_insn *insn,
763 struct bpf_jit_probe *probe)
764 {
765 if (BPF_MODE(insn->code) != BPF_PROBE_ATOMIC)
766 return;
767
768 /* lgrl %r1,kern_arena */
769 EMIT6_PCREL_RILB(0xc4080000, REG_W1, jit->kern_arena);
770 /* agr %r1,%dst */
771 EMIT4(0xb9080000, REG_W1, insn->dst_reg);
772 probe->arena_reg = REG_W1;
773 probe->prg = jit->prg;
774 }
775
776 static int bpf_jit_probe_post(struct bpf_jit *jit, struct bpf_prog *fp,
777 struct bpf_jit_probe *probe)
778 {
779 struct exception_table_entry *ex;
780 int i, prg;
781 s64 delta;
782 u8 *insn;
783
784 if (probe->prg == -1)
785 /* The probe is not armed. */
786 return 0;
787 bpf_jit_probe_emit_nop(jit, probe);
788 if (!fp->aux->extable)
789 /* Do nothing during early JIT passes. */
790 return 0;
791 insn = jit->prg_buf + probe->prg;
792 if (WARN_ON_ONCE(probe->prg + insn_length(*insn) != probe->nop_prg))
793 /* JIT bug - gap between probe and nop instructions. */
794 return -1;
795 for (i = 0; i < 2; i++) {
796 if (WARN_ON_ONCE(jit->excnt >= fp->aux->num_exentries))
797 /* Verifier bug - not enough entries. */
798 return -1;
799 ex = &fp->aux->extable[jit->excnt];
800 /* Add extable entries for probe and nop instructions. */
801 prg = i == 0 ? probe->prg : probe->nop_prg;
802 delta = jit->prg_buf + prg - (u8 *)&ex->insn;
803 if (WARN_ON_ONCE(delta < INT_MIN || delta > INT_MAX))
804 /* JIT bug - code and extable must be close. */
805 return -1;
806 ex->insn = delta;
807 /*
808 * Land on the current instruction. Note that the extable
809 * infrastructure ignores the fixup field; it is handled by
810 * ex_handler_bpf().
811 */
812 delta = jit->prg_buf + jit->prg - (u8 *)&ex->fixup;
813 if (WARN_ON_ONCE(delta < INT_MIN || delta > INT_MAX))
814 /* JIT bug - landing pad and extable must be close. */
815 return -1;
816 ex->fixup = delta;
817 ex->type = EX_TYPE_BPF;
818 ex->data = probe->reg;
819 jit->excnt++;
820 }
821 return 0;
822 }
823
824 /*
825 * Sign-extend the register if necessary
826 */
827 static int sign_extend(struct bpf_jit *jit, int r, u8 size, u8 flags)
828 {
829 if (!(flags & BTF_FMODEL_SIGNED_ARG))
830 return 0;
831
832 switch (size) {
833 case 1:
834 /* lgbr %r,%r */
835 EMIT4(0xb9060000, r, r);
836 return 0;
837 case 2:
838 /* lghr %r,%r */
839 EMIT4(0xb9070000, r, r);
840 return 0;
841 case 4:
842 /* lgfr %r,%r */
843 EMIT4(0xb9140000, r, r);
844 return 0;
845 case 8:
846 return 0;
847 default:
848 return -1;
849 }
850 }
851
852 /*
853 * Compile one eBPF instruction into s390x code
854 *
855 * NOTE: Use noinline because for gcov (-fprofile-arcs) gcc allocates a lot of
856 * stack space for the large switch statement.
857 */
858 static noinline int bpf_jit_insn(struct bpf_jit *jit, struct bpf_prog *fp,
859 int i, bool extra_pass, u32 stack_depth)
860 {
861 struct bpf_insn *insn = &fp->insnsi[i];
862 s32 branch_oc_off = insn->off;
863 u32 dst_reg = insn->dst_reg;
864 u32 src_reg = insn->src_reg;
865 struct bpf_jit_probe probe;
866 int last, insn_count = 1;
867 u32 *addrs = jit->addrs;
868 s32 imm = insn->imm;
869 s16 off = insn->off;
870 unsigned int mask;
871 int err;
872
873 bpf_jit_probe_init(&probe);
874
875 switch (insn->code) {
876 /*
877 * BPF_MOV
878 */
879 case BPF_ALU | BPF_MOV | BPF_X:
880 switch (insn->off) {
881 case 0: /* DST = (u32) SRC */
882 /* llgfr %dst,%src */
883 EMIT4(0xb9160000, dst_reg, src_reg);
884 if (insn_is_zext(&insn[1]))
885 insn_count = 2;
886 break;
887 case 8: /* DST = (u32)(s8) SRC */
888 /* lbr %dst,%src */
889 EMIT4(0xb9260000, dst_reg, src_reg);
890 /* llgfr %dst,%dst */
891 EMIT4(0xb9160000, dst_reg, dst_reg);
892 break;
893 case 16: /* DST = (u32)(s16) SRC */
894 /* lhr %dst,%src */
895 EMIT4(0xb9270000, dst_reg, src_reg);
896 /* llgfr %dst,%dst */
897 EMIT4(0xb9160000, dst_reg, dst_reg);
898 break;
899 }
900 break;
901 case BPF_ALU64 | BPF_MOV | BPF_X:
902 if (insn_is_cast_user(insn)) {
903 int patch_brc;
904
905 /* ltgr %dst,%src */
906 EMIT4(0xb9020000, dst_reg, src_reg);
907 /* brc 8,0f */
908 patch_brc = jit->prg;
909 EMIT4_PCREL_RIC(0xa7040000, 8, 0);
910 /* iihf %dst,user_arena>>32 */
911 EMIT6_IMM(0xc0080000, dst_reg, jit->user_arena >> 32);
912 /* 0: */
913 if (jit->prg_buf)
914 *(u16 *)(jit->prg_buf + patch_brc + 2) =
915 (jit->prg - patch_brc) >> 1;
916 break;
917 }
918 switch (insn->off) {
919 case 0: /* DST = SRC */
920 /* lgr %dst,%src */
921 EMIT4(0xb9040000, dst_reg, src_reg);
922 break;
923 case 8: /* DST = (s8) SRC */
924 /* lgbr %dst,%src */
925 EMIT4(0xb9060000, dst_reg, src_reg);
926 break;
927 case 16: /* DST = (s16) SRC */
928 /* lghr %dst,%src */
929 EMIT4(0xb9070000, dst_reg, src_reg);
930 break;
931 case 32: /* DST = (s32) SRC */
932 /* lgfr %dst,%src */
933 EMIT4(0xb9140000, dst_reg, src_reg);
934 break;
935 }
936 break;
937 case BPF_ALU | BPF_MOV | BPF_K: /* dst = (u32) imm */
938 /* llilf %dst,imm */
939 EMIT6_IMM(0xc00f0000, dst_reg, imm);
940 if (insn_is_zext(&insn[1]))
941 insn_count = 2;
942 break;
943 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = imm */
944 /* lgfi %dst,imm */
945 EMIT6_IMM(0xc0010000, dst_reg, imm);
946 break;
947 /*
948 * BPF_LD 64
949 */
950 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */
951 {
952 /* 16 byte instruction that uses two 'struct bpf_insn' */
953 u64 imm64;
954
955 imm64 = (u64)(u32) insn[0].imm | ((u64)(u32) insn[1].imm) << 32;
956 /* lgrl %dst,imm */
957 EMIT6_PCREL_RILB(0xc4080000, dst_reg, _EMIT_CONST_U64(imm64));
958 insn_count = 2;
959 break;
960 }
961 /*
962 * BPF_ADD
963 */
964 case BPF_ALU | BPF_ADD | BPF_X: /* dst = (u32) dst + (u32) src */
965 /* ar %dst,%src */
966 EMIT2(0x1a00, dst_reg, src_reg);
967 EMIT_ZERO(dst_reg);
968 break;
969 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst = dst + src */
970 /* agr %dst,%src */
971 EMIT4(0xb9080000, dst_reg, src_reg);
972 break;
973 case BPF_ALU | BPF_ADD | BPF_K: /* dst = (u32) dst + (u32) imm */
974 if (imm != 0) {
975 /* alfi %dst,imm */
976 EMIT6_IMM(0xc20b0000, dst_reg, imm);
977 }
978 EMIT_ZERO(dst_reg);
979 break;
980 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst = dst + imm */
981 if (!imm)
982 break;
983 /* agfi %dst,imm */
984 EMIT6_IMM(0xc2080000, dst_reg, imm);
985 break;
986 /*
987 * BPF_SUB
988 */
989 case BPF_ALU | BPF_SUB | BPF_X: /* dst = (u32) dst - (u32) src */
990 /* sr %dst,%src */
991 EMIT2(0x1b00, dst_reg, src_reg);
992 EMIT_ZERO(dst_reg);
993 break;
994 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst = dst - src */
995 /* sgr %dst,%src */
996 EMIT4(0xb9090000, dst_reg, src_reg);
997 break;
998 case BPF_ALU | BPF_SUB | BPF_K: /* dst = (u32) dst - (u32) imm */
999 if (imm != 0) {
1000 /* alfi %dst,-imm */
1001 EMIT6_IMM(0xc20b0000, dst_reg, -imm);
1002 }
1003 EMIT_ZERO(dst_reg);
1004 break;
1005 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst = dst - imm */
1006 if (!imm)
1007 break;
1008 if (imm == -0x80000000) {
1009 /* algfi %dst,0x80000000 */
1010 EMIT6_IMM(0xc20a0000, dst_reg, 0x80000000);
1011 } else {
1012 /* agfi %dst,-imm */
1013 EMIT6_IMM(0xc2080000, dst_reg, -imm);
1014 }
1015 break;
1016 /*
1017 * BPF_MUL
1018 */
1019 case BPF_ALU | BPF_MUL | BPF_X: /* dst = (u32) dst * (u32) src */
1020 /* msr %dst,%src */
1021 EMIT4(0xb2520000, dst_reg, src_reg);
1022 EMIT_ZERO(dst_reg);
1023 break;
1024 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst = dst * src */
1025 /* msgr %dst,%src */
1026 EMIT4(0xb90c0000, dst_reg, src_reg);
1027 break;
1028 case BPF_ALU | BPF_MUL | BPF_K: /* dst = (u32) dst * (u32) imm */
1029 if (imm != 1) {
1030 /* msfi %r5,imm */
1031 EMIT6_IMM(0xc2010000, dst_reg, imm);
1032 }
1033 EMIT_ZERO(dst_reg);
1034 break;
1035 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst = dst * imm */
1036 if (imm == 1)
1037 break;
1038 /* msgfi %dst,imm */
1039 EMIT6_IMM(0xc2000000, dst_reg, imm);
1040 break;
1041 /*
1042 * BPF_DIV / BPF_MOD
1043 */
1044 case BPF_ALU | BPF_DIV | BPF_X:
1045 case BPF_ALU | BPF_MOD | BPF_X:
1046 {
1047 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
1048
1049 switch (off) {
1050 case 0: /* dst = (u32) dst {/,%} (u32) src */
1051 /* xr %w0,%w0 */
1052 EMIT2(0x1700, REG_W0, REG_W0);
1053 /* lr %w1,%dst */
1054 EMIT2(0x1800, REG_W1, dst_reg);
1055 /* dlr %w0,%src */
1056 EMIT4(0xb9970000, REG_W0, src_reg);
1057 break;
1058 case 1: /* dst = (u32) ((s32) dst {/,%} (s32) src) */
1059 /* lgfr %r1,%dst */
1060 EMIT4(0xb9140000, REG_W1, dst_reg);
1061 /* dsgfr %r0,%src */
1062 EMIT4(0xb91d0000, REG_W0, src_reg);
1063 break;
1064 }
1065 /* llgfr %dst,%rc */
1066 EMIT4(0xb9160000, dst_reg, rc_reg);
1067 if (insn_is_zext(&insn[1]))
1068 insn_count = 2;
1069 break;
1070 }
1071 case BPF_ALU64 | BPF_DIV | BPF_X:
1072 case BPF_ALU64 | BPF_MOD | BPF_X:
1073 {
1074 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
1075
1076 switch (off) {
1077 case 0: /* dst = dst {/,%} src */
1078 /* lghi %w0,0 */
1079 EMIT4_IMM(0xa7090000, REG_W0, 0);
1080 /* lgr %w1,%dst */
1081 EMIT4(0xb9040000, REG_W1, dst_reg);
1082 /* dlgr %w0,%src */
1083 EMIT4(0xb9870000, REG_W0, src_reg);
1084 break;
1085 case 1: /* dst = (s64) dst {/,%} (s64) src */
1086 /* lgr %w1,%dst */
1087 EMIT4(0xb9040000, REG_W1, dst_reg);
1088 /* dsgr %w0,%src */
1089 EMIT4(0xb90d0000, REG_W0, src_reg);
1090 break;
1091 }
1092 /* lgr %dst,%rc */
1093 EMIT4(0xb9040000, dst_reg, rc_reg);
1094 break;
1095 }
1096 case BPF_ALU | BPF_DIV | BPF_K:
1097 case BPF_ALU | BPF_MOD | BPF_K:
1098 {
1099 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
1100
1101 if (imm == 1) {
1102 if (BPF_OP(insn->code) == BPF_MOD)
1103 /* lghi %dst,0 */
1104 EMIT4_IMM(0xa7090000, dst_reg, 0);
1105 else
1106 EMIT_ZERO(dst_reg);
1107 break;
1108 }
1109 if (!is_first_pass(jit) && can_use_ldisp_for_lit32(jit)) {
1110 switch (off) {
1111 case 0: /* dst = (u32) dst {/,%} (u32) imm */
1112 /* xr %w0,%w0 */
1113 EMIT2(0x1700, REG_W0, REG_W0);
1114 /* lr %w1,%dst */
1115 EMIT2(0x1800, REG_W1, dst_reg);
1116 /* dl %w0,<d(imm)>(%l) */
1117 EMIT6_DISP_LH(0xe3000000, 0x0097, REG_W0, REG_0,
1118 REG_L, EMIT_CONST_U32(imm));
1119 break;
1120 case 1: /* dst = (s32) dst {/,%} (s32) imm */
1121 /* lgfr %r1,%dst */
1122 EMIT4(0xb9140000, REG_W1, dst_reg);
1123 /* dsgf %r0,<d(imm)>(%l) */
1124 EMIT6_DISP_LH(0xe3000000, 0x001d, REG_W0, REG_0,
1125 REG_L, EMIT_CONST_U32(imm));
1126 break;
1127 }
1128 } else {
1129 switch (off) {
1130 case 0: /* dst = (u32) dst {/,%} (u32) imm */
1131 /* xr %w0,%w0 */
1132 EMIT2(0x1700, REG_W0, REG_W0);
1133 /* lr %w1,%dst */
1134 EMIT2(0x1800, REG_W1, dst_reg);
1135 /* lrl %dst,imm */
1136 EMIT6_PCREL_RILB(0xc40d0000, dst_reg,
1137 _EMIT_CONST_U32(imm));
1138 jit->seen |= SEEN_LITERAL;
1139 /* dlr %w0,%dst */
1140 EMIT4(0xb9970000, REG_W0, dst_reg);
1141 break;
1142 case 1: /* dst = (s32) dst {/,%} (s32) imm */
1143 /* lgfr %w1,%dst */
1144 EMIT4(0xb9140000, REG_W1, dst_reg);
1145 /* lgfrl %dst,imm */
1146 EMIT6_PCREL_RILB(0xc40c0000, dst_reg,
1147 _EMIT_CONST_U32(imm));
1148 jit->seen |= SEEN_LITERAL;
1149 /* dsgr %w0,%dst */
1150 EMIT4(0xb90d0000, REG_W0, dst_reg);
1151 break;
1152 }
1153 }
1154 /* llgfr %dst,%rc */
1155 EMIT4(0xb9160000, dst_reg, rc_reg);
1156 if (insn_is_zext(&insn[1]))
1157 insn_count = 2;
1158 break;
1159 }
1160 case BPF_ALU64 | BPF_DIV | BPF_K:
1161 case BPF_ALU64 | BPF_MOD | BPF_K:
1162 {
1163 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0;
1164
1165 if (imm == 1) {
1166 if (BPF_OP(insn->code) == BPF_MOD)
1167 /* lhgi %dst,0 */
1168 EMIT4_IMM(0xa7090000, dst_reg, 0);
1169 break;
1170 }
1171 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
1172 switch (off) {
1173 case 0: /* dst = dst {/,%} imm */
1174 /* lghi %w0,0 */
1175 EMIT4_IMM(0xa7090000, REG_W0, 0);
1176 /* lgr %w1,%dst */
1177 EMIT4(0xb9040000, REG_W1, dst_reg);
1178 /* dlg %w0,<d(imm)>(%l) */
1179 EMIT6_DISP_LH(0xe3000000, 0x0087, REG_W0, REG_0,
1180 REG_L, EMIT_CONST_U64(imm));
1181 break;
1182 case 1: /* dst = (s64) dst {/,%} (s64) imm */
1183 /* lgr %w1,%dst */
1184 EMIT4(0xb9040000, REG_W1, dst_reg);
1185 /* dsg %w0,<d(imm)>(%l) */
1186 EMIT6_DISP_LH(0xe3000000, 0x000d, REG_W0, REG_0,
1187 REG_L, EMIT_CONST_U64(imm));
1188 break;
1189 }
1190 } else {
1191 switch (off) {
1192 case 0: /* dst = dst {/,%} imm */
1193 /* lghi %w0,0 */
1194 EMIT4_IMM(0xa7090000, REG_W0, 0);
1195 /* lgr %w1,%dst */
1196 EMIT4(0xb9040000, REG_W1, dst_reg);
1197 /* lgrl %dst,imm */
1198 EMIT6_PCREL_RILB(0xc4080000, dst_reg,
1199 _EMIT_CONST_U64(imm));
1200 jit->seen |= SEEN_LITERAL;
1201 /* dlgr %w0,%dst */
1202 EMIT4(0xb9870000, REG_W0, dst_reg);
1203 break;
1204 case 1: /* dst = (s64) dst {/,%} (s64) imm */
1205 /* lgr %w1,%dst */
1206 EMIT4(0xb9040000, REG_W1, dst_reg);
1207 /* lgrl %dst,imm */
1208 EMIT6_PCREL_RILB(0xc4080000, dst_reg,
1209 _EMIT_CONST_U64(imm));
1210 jit->seen |= SEEN_LITERAL;
1211 /* dsgr %w0,%dst */
1212 EMIT4(0xb90d0000, REG_W0, dst_reg);
1213 break;
1214 }
1215 }
1216 /* lgr %dst,%rc */
1217 EMIT4(0xb9040000, dst_reg, rc_reg);
1218 break;
1219 }
1220 /*
1221 * BPF_AND
1222 */
1223 case BPF_ALU | BPF_AND | BPF_X: /* dst = (u32) dst & (u32) src */
1224 /* nr %dst,%src */
1225 EMIT2(0x1400, dst_reg, src_reg);
1226 EMIT_ZERO(dst_reg);
1227 break;
1228 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */
1229 /* ngr %dst,%src */
1230 EMIT4(0xb9800000, dst_reg, src_reg);
1231 break;
1232 case BPF_ALU | BPF_AND | BPF_K: /* dst = (u32) dst & (u32) imm */
1233 /* nilf %dst,imm */
1234 EMIT6_IMM(0xc00b0000, dst_reg, imm);
1235 EMIT_ZERO(dst_reg);
1236 break;
1237 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */
1238 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
1239 /* ng %dst,<d(imm)>(%l) */
1240 EMIT6_DISP_LH(0xe3000000, 0x0080,
1241 dst_reg, REG_0, REG_L,
1242 EMIT_CONST_U64(imm));
1243 } else {
1244 /* lgrl %w0,imm */
1245 EMIT6_PCREL_RILB(0xc4080000, REG_W0,
1246 _EMIT_CONST_U64(imm));
1247 jit->seen |= SEEN_LITERAL;
1248 /* ngr %dst,%w0 */
1249 EMIT4(0xb9800000, dst_reg, REG_W0);
1250 }
1251 break;
1252 /*
1253 * BPF_OR
1254 */
1255 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */
1256 /* or %dst,%src */
1257 EMIT2(0x1600, dst_reg, src_reg);
1258 EMIT_ZERO(dst_reg);
1259 break;
1260 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */
1261 /* ogr %dst,%src */
1262 EMIT4(0xb9810000, dst_reg, src_reg);
1263 break;
1264 case BPF_ALU | BPF_OR | BPF_K: /* dst = (u32) dst | (u32) imm */
1265 /* oilf %dst,imm */
1266 EMIT6_IMM(0xc00d0000, dst_reg, imm);
1267 EMIT_ZERO(dst_reg);
1268 break;
1269 case BPF_ALU64 | BPF_OR | BPF_K: /* dst = dst | imm */
1270 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
1271 /* og %dst,<d(imm)>(%l) */
1272 EMIT6_DISP_LH(0xe3000000, 0x0081,
1273 dst_reg, REG_0, REG_L,
1274 EMIT_CONST_U64(imm));
1275 } else {
1276 /* lgrl %w0,imm */
1277 EMIT6_PCREL_RILB(0xc4080000, REG_W0,
1278 _EMIT_CONST_U64(imm));
1279 jit->seen |= SEEN_LITERAL;
1280 /* ogr %dst,%w0 */
1281 EMIT4(0xb9810000, dst_reg, REG_W0);
1282 }
1283 break;
1284 /*
1285 * BPF_XOR
1286 */
1287 case BPF_ALU | BPF_XOR | BPF_X: /* dst = (u32) dst ^ (u32) src */
1288 /* xr %dst,%src */
1289 EMIT2(0x1700, dst_reg, src_reg);
1290 EMIT_ZERO(dst_reg);
1291 break;
1292 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst = dst ^ src */
1293 /* xgr %dst,%src */
1294 EMIT4(0xb9820000, dst_reg, src_reg);
1295 break;
1296 case BPF_ALU | BPF_XOR | BPF_K: /* dst = (u32) dst ^ (u32) imm */
1297 if (imm != 0) {
1298 /* xilf %dst,imm */
1299 EMIT6_IMM(0xc0070000, dst_reg, imm);
1300 }
1301 EMIT_ZERO(dst_reg);
1302 break;
1303 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst = dst ^ imm */
1304 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) {
1305 /* xg %dst,<d(imm)>(%l) */
1306 EMIT6_DISP_LH(0xe3000000, 0x0082,
1307 dst_reg, REG_0, REG_L,
1308 EMIT_CONST_U64(imm));
1309 } else {
1310 /* lgrl %w0,imm */
1311 EMIT6_PCREL_RILB(0xc4080000, REG_W0,
1312 _EMIT_CONST_U64(imm));
1313 jit->seen |= SEEN_LITERAL;
1314 /* xgr %dst,%w0 */
1315 EMIT4(0xb9820000, dst_reg, REG_W0);
1316 }
1317 break;
1318 /*
1319 * BPF_LSH
1320 */
1321 case BPF_ALU | BPF_LSH | BPF_X: /* dst = (u32) dst << (u32) src */
1322 /* sll %dst,0(%src) */
1323 EMIT4_DISP(0x89000000, dst_reg, src_reg, 0);
1324 EMIT_ZERO(dst_reg);
1325 break;
1326 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst = dst << src */
1327 /* sllg %dst,%dst,0(%src) */
1328 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, src_reg, 0);
1329 break;
1330 case BPF_ALU | BPF_LSH | BPF_K: /* dst = (u32) dst << (u32) imm */
1331 if (imm != 0) {
1332 /* sll %dst,imm(%r0) */
1333 EMIT4_DISP(0x89000000, dst_reg, REG_0, imm);
1334 }
1335 EMIT_ZERO(dst_reg);
1336 break;
1337 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst = dst << imm */
1338 if (imm == 0)
1339 break;
1340 /* sllg %dst,%dst,imm(%r0) */
1341 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, REG_0, imm);
1342 break;
1343 /*
1344 * BPF_RSH
1345 */
1346 case BPF_ALU | BPF_RSH | BPF_X: /* dst = (u32) dst >> (u32) src */
1347 /* srl %dst,0(%src) */
1348 EMIT4_DISP(0x88000000, dst_reg, src_reg, 0);
1349 EMIT_ZERO(dst_reg);
1350 break;
1351 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst = dst >> src */
1352 /* srlg %dst,%dst,0(%src) */
1353 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, src_reg, 0);
1354 break;
1355 case BPF_ALU | BPF_RSH | BPF_K: /* dst = (u32) dst >> (u32) imm */
1356 if (imm != 0) {
1357 /* srl %dst,imm(%r0) */
1358 EMIT4_DISP(0x88000000, dst_reg, REG_0, imm);
1359 }
1360 EMIT_ZERO(dst_reg);
1361 break;
1362 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst = dst >> imm */
1363 if (imm == 0)
1364 break;
1365 /* srlg %dst,%dst,imm(%r0) */
1366 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, REG_0, imm);
1367 break;
1368 /*
1369 * BPF_ARSH
1370 */
1371 case BPF_ALU | BPF_ARSH | BPF_X: /* ((s32) dst) >>= src */
1372 /* sra %dst,%dst,0(%src) */
1373 EMIT4_DISP(0x8a000000, dst_reg, src_reg, 0);
1374 EMIT_ZERO(dst_reg);
1375 break;
1376 case BPF_ALU64 | BPF_ARSH | BPF_X: /* ((s64) dst) >>= src */
1377 /* srag %dst,%dst,0(%src) */
1378 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, src_reg, 0);
1379 break;
1380 case BPF_ALU | BPF_ARSH | BPF_K: /* ((s32) dst >> imm */
1381 if (imm != 0) {
1382 /* sra %dst,imm(%r0) */
1383 EMIT4_DISP(0x8a000000, dst_reg, REG_0, imm);
1384 }
1385 EMIT_ZERO(dst_reg);
1386 break;
1387 case BPF_ALU64 | BPF_ARSH | BPF_K: /* ((s64) dst) >>= imm */
1388 if (imm == 0)
1389 break;
1390 /* srag %dst,%dst,imm(%r0) */
1391 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, REG_0, imm);
1392 break;
1393 /*
1394 * BPF_NEG
1395 */
1396 case BPF_ALU | BPF_NEG: /* dst = (u32) -dst */
1397 /* lcr %dst,%dst */
1398 EMIT2(0x1300, dst_reg, dst_reg);
1399 EMIT_ZERO(dst_reg);
1400 break;
1401 case BPF_ALU64 | BPF_NEG: /* dst = -dst */
1402 /* lcgr %dst,%dst */
1403 EMIT4(0xb9030000, dst_reg, dst_reg);
1404 break;
1405 /*
1406 * BPF_FROM_BE/LE
1407 */
1408 case BPF_ALU | BPF_END | BPF_FROM_BE:
1409 /* s390 is big endian, therefore only clear high order bytes */
1410 switch (imm) {
1411 case 16: /* dst = (u16) cpu_to_be16(dst) */
1412 /* llghr %dst,%dst */
1413 EMIT4(0xb9850000, dst_reg, dst_reg);
1414 if (insn_is_zext(&insn[1]))
1415 insn_count = 2;
1416 break;
1417 case 32: /* dst = (u32) cpu_to_be32(dst) */
1418 if (!fp->aux->verifier_zext)
1419 /* llgfr %dst,%dst */
1420 EMIT4(0xb9160000, dst_reg, dst_reg);
1421 break;
1422 case 64: /* dst = (u64) cpu_to_be64(dst) */
1423 break;
1424 }
1425 break;
1426 case BPF_ALU | BPF_END | BPF_FROM_LE:
1427 case BPF_ALU64 | BPF_END | BPF_FROM_LE:
1428 switch (imm) {
1429 case 16: /* dst = (u16) cpu_to_le16(dst) */
1430 /* lrvr %dst,%dst */
1431 EMIT4(0xb91f0000, dst_reg, dst_reg);
1432 /* srl %dst,16(%r0) */
1433 EMIT4_DISP(0x88000000, dst_reg, REG_0, 16);
1434 /* llghr %dst,%dst */
1435 EMIT4(0xb9850000, dst_reg, dst_reg);
1436 if (insn_is_zext(&insn[1]))
1437 insn_count = 2;
1438 break;
1439 case 32: /* dst = (u32) cpu_to_le32(dst) */
1440 /* lrvr %dst,%dst */
1441 EMIT4(0xb91f0000, dst_reg, dst_reg);
1442 if (!fp->aux->verifier_zext)
1443 /* llgfr %dst,%dst */
1444 EMIT4(0xb9160000, dst_reg, dst_reg);
1445 break;
1446 case 64: /* dst = (u64) cpu_to_le64(dst) */
1447 /* lrvgr %dst,%dst */
1448 EMIT4(0xb90f0000, dst_reg, dst_reg);
1449 break;
1450 }
1451 break;
1452 /*
1453 * BPF_NOSPEC (speculation barrier)
1454 */
1455 case BPF_ST | BPF_NOSPEC:
1456 break;
1457 /*
1458 * BPF_ST(X)
1459 */
1460 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src_reg */
1461 case BPF_STX | BPF_PROBE_MEM32 | BPF_B:
1462 bpf_jit_probe_store_pre(jit, insn, &probe);
1463 /* stcy %src,off(%dst,%arena) */
1464 EMIT6_DISP_LH(0xe3000000, 0x0072, src_reg, dst_reg,
1465 probe.arena_reg, off);
1466 err = bpf_jit_probe_post(jit, fp, &probe);
1467 if (err < 0)
1468 return err;
1469 jit->seen |= SEEN_MEM;
1470 break;
1471 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */
1472 case BPF_STX | BPF_PROBE_MEM32 | BPF_H:
1473 bpf_jit_probe_store_pre(jit, insn, &probe);
1474 /* sthy %src,off(%dst,%arena) */
1475 EMIT6_DISP_LH(0xe3000000, 0x0070, src_reg, dst_reg,
1476 probe.arena_reg, off);
1477 err = bpf_jit_probe_post(jit, fp, &probe);
1478 if (err < 0)
1479 return err;
1480 jit->seen |= SEEN_MEM;
1481 break;
1482 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */
1483 case BPF_STX | BPF_PROBE_MEM32 | BPF_W:
1484 bpf_jit_probe_store_pre(jit, insn, &probe);
1485 /* sty %src,off(%dst,%arena) */
1486 EMIT6_DISP_LH(0xe3000000, 0x0050, src_reg, dst_reg,
1487 probe.arena_reg, off);
1488 err = bpf_jit_probe_post(jit, fp, &probe);
1489 if (err < 0)
1490 return err;
1491 jit->seen |= SEEN_MEM;
1492 break;
1493 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */
1494 case BPF_STX | BPF_PROBE_MEM32 | BPF_DW:
1495 bpf_jit_probe_store_pre(jit, insn, &probe);
1496 /* stg %src,off(%dst,%arena) */
1497 EMIT6_DISP_LH(0xe3000000, 0x0024, src_reg, dst_reg,
1498 probe.arena_reg, off);
1499 err = bpf_jit_probe_post(jit, fp, &probe);
1500 if (err < 0)
1501 return err;
1502 jit->seen |= SEEN_MEM;
1503 break;
1504 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */
1505 case BPF_ST | BPF_PROBE_MEM32 | BPF_B:
1506 /* lhi %w0,imm */
1507 EMIT4_IMM(0xa7080000, REG_W0, (u8) imm);
1508 bpf_jit_probe_store_pre(jit, insn, &probe);
1509 /* stcy %w0,off(%dst,%arena) */
1510 EMIT6_DISP_LH(0xe3000000, 0x0072, REG_W0, dst_reg,
1511 probe.arena_reg, off);
1512 err = bpf_jit_probe_post(jit, fp, &probe);
1513 if (err < 0)
1514 return err;
1515 jit->seen |= SEEN_MEM;
1516 break;
1517 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */
1518 case BPF_ST | BPF_PROBE_MEM32 | BPF_H:
1519 /* lhi %w0,imm */
1520 EMIT4_IMM(0xa7080000, REG_W0, (u16) imm);
1521 bpf_jit_probe_store_pre(jit, insn, &probe);
1522 /* sthy %w0,off(%dst,%arena) */
1523 EMIT6_DISP_LH(0xe3000000, 0x0070, REG_W0, dst_reg,
1524 probe.arena_reg, off);
1525 err = bpf_jit_probe_post(jit, fp, &probe);
1526 if (err < 0)
1527 return err;
1528 jit->seen |= SEEN_MEM;
1529 break;
1530 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */
1531 case BPF_ST | BPF_PROBE_MEM32 | BPF_W:
1532 /* llilf %w0,imm */
1533 EMIT6_IMM(0xc00f0000, REG_W0, (u32) imm);
1534 bpf_jit_probe_store_pre(jit, insn, &probe);
1535 /* sty %w0,off(%dst,%arena) */
1536 EMIT6_DISP_LH(0xe3000000, 0x0050, REG_W0, dst_reg,
1537 probe.arena_reg, off);
1538 err = bpf_jit_probe_post(jit, fp, &probe);
1539 if (err < 0)
1540 return err;
1541 jit->seen |= SEEN_MEM;
1542 break;
1543 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */
1544 case BPF_ST | BPF_PROBE_MEM32 | BPF_DW:
1545 /* lgfi %w0,imm */
1546 EMIT6_IMM(0xc0010000, REG_W0, imm);
1547 bpf_jit_probe_store_pre(jit, insn, &probe);
1548 /* stg %w0,off(%dst,%arena) */
1549 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W0, dst_reg,
1550 probe.arena_reg, off);
1551 err = bpf_jit_probe_post(jit, fp, &probe);
1552 if (err < 0)
1553 return err;
1554 jit->seen |= SEEN_MEM;
1555 break;
1556 /*
1557 * BPF_ATOMIC
1558 */
1559 case BPF_STX | BPF_ATOMIC | BPF_DW:
1560 case BPF_STX | BPF_ATOMIC | BPF_W:
1561 case BPF_STX | BPF_PROBE_ATOMIC | BPF_DW:
1562 case BPF_STX | BPF_PROBE_ATOMIC | BPF_W:
1563 {
1564 bool is32 = BPF_SIZE(insn->code) == BPF_W;
1565
1566 /*
1567 * Unlike loads and stores, atomics have only a base register,
1568 * but no index register. For the non-arena case, simply use
1569 * %dst as a base. For the arena case, use the work register
1570 * %r1: first, load the arena base into it, and then add %dst
1571 * to it.
1572 */
1573 probe.arena_reg = dst_reg;
1574
1575 switch (insn->imm) {
1576 #define EMIT_ATOMIC(op32, op64) do { \
1577 bpf_jit_probe_atomic_pre(jit, insn, &probe); \
1578 /* {op32|op64} {%w0|%src},%src,off(%arena) */ \
1579 EMIT6_DISP_LH(0xeb000000, is32 ? (op32) : (op64), \
1580 (insn->imm & BPF_FETCH) ? src_reg : REG_W0, \
1581 src_reg, probe.arena_reg, off); \
1582 err = bpf_jit_probe_post(jit, fp, &probe); \
1583 if (err < 0) \
1584 return err; \
1585 if (insn->imm & BPF_FETCH) { \
1586 /* bcr 14,0 - see atomic_fetch_{add,and,or,xor}() */ \
1587 _EMIT2(0x07e0); \
1588 if (is32) \
1589 EMIT_ZERO(src_reg); \
1590 } \
1591 } while (0)
1592 case BPF_ADD:
1593 case BPF_ADD | BPF_FETCH:
1594 /* {laal|laalg} */
1595 EMIT_ATOMIC(0x00fa, 0x00ea);
1596 break;
1597 case BPF_AND:
1598 case BPF_AND | BPF_FETCH:
1599 /* {lan|lang} */
1600 EMIT_ATOMIC(0x00f4, 0x00e4);
1601 break;
1602 case BPF_OR:
1603 case BPF_OR | BPF_FETCH:
1604 /* {lao|laog} */
1605 EMIT_ATOMIC(0x00f6, 0x00e6);
1606 break;
1607 case BPF_XOR:
1608 case BPF_XOR | BPF_FETCH:
1609 /* {lax|laxg} */
1610 EMIT_ATOMIC(0x00f7, 0x00e7);
1611 break;
1612 #undef EMIT_ATOMIC
1613 case BPF_XCHG: {
1614 struct bpf_jit_probe load_probe = probe;
1615 int loop_start;
1616
1617 bpf_jit_probe_atomic_pre(jit, insn, &load_probe);
1618 /* {ly|lg} %w0,off(%arena) */
1619 EMIT6_DISP_LH(0xe3000000,
1620 is32 ? 0x0058 : 0x0004, REG_W0, REG_0,
1621 load_probe.arena_reg, off);
1622 bpf_jit_probe_emit_nop(jit, &load_probe);
1623 /* Reuse {ly|lg}'s arena_reg for {csy|csg}. */
1624 if (load_probe.prg != -1) {
1625 probe.prg = jit->prg;
1626 probe.arena_reg = load_probe.arena_reg;
1627 }
1628 loop_start = jit->prg;
1629 /* 0: {csy|csg} %w0,%src,off(%arena) */
1630 EMIT6_DISP_LH(0xeb000000, is32 ? 0x0014 : 0x0030,
1631 REG_W0, src_reg, probe.arena_reg, off);
1632 bpf_jit_probe_emit_nop(jit, &probe);
1633 /* brc 4,0b */
1634 EMIT4_PCREL_RIC(0xa7040000, 4, loop_start);
1635 /* {llgfr|lgr} %src,%w0 */
1636 EMIT4(is32 ? 0xb9160000 : 0xb9040000, src_reg, REG_W0);
1637 /* Both probes should land here on exception. */
1638 err = bpf_jit_probe_post(jit, fp, &load_probe);
1639 if (err < 0)
1640 return err;
1641 err = bpf_jit_probe_post(jit, fp, &probe);
1642 if (err < 0)
1643 return err;
1644 if (is32 && insn_is_zext(&insn[1]))
1645 insn_count = 2;
1646 break;
1647 }
1648 case BPF_CMPXCHG:
1649 bpf_jit_probe_atomic_pre(jit, insn, &probe);
1650 /* 0: {csy|csg} %b0,%src,off(%arena) */
1651 EMIT6_DISP_LH(0xeb000000, is32 ? 0x0014 : 0x0030,
1652 BPF_REG_0, src_reg,
1653 probe.arena_reg, off);
1654 err = bpf_jit_probe_post(jit, fp, &probe);
1655 if (err < 0)
1656 return err;
1657 break;
1658 default:
1659 pr_err("Unknown atomic operation %02x\n", insn->imm);
1660 return -1;
1661 }
1662
1663 jit->seen |= SEEN_MEM;
1664 break;
1665 }
1666 /*
1667 * BPF_LDX
1668 */
1669 case BPF_LDX | BPF_MEM | BPF_B: /* dst = *(u8 *)(ul) (src + off) */
1670 case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1671 case BPF_LDX | BPF_PROBE_MEM32 | BPF_B:
1672 bpf_jit_probe_load_pre(jit, insn, &probe);
1673 /* llgc %dst,off(%src,%arena) */
1674 EMIT6_DISP_LH(0xe3000000, 0x0090, dst_reg, src_reg,
1675 probe.arena_reg, off);
1676 err = bpf_jit_probe_post(jit, fp, &probe);
1677 if (err < 0)
1678 return err;
1679 jit->seen |= SEEN_MEM;
1680 if (insn_is_zext(&insn[1]))
1681 insn_count = 2;
1682 break;
1683 case BPF_LDX | BPF_MEMSX | BPF_B: /* dst = *(s8 *)(ul) (src + off) */
1684 case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1685 bpf_jit_probe_load_pre(jit, insn, &probe);
1686 /* lgb %dst,off(%src) */
1687 EMIT6_DISP_LH(0xe3000000, 0x0077, dst_reg, src_reg, REG_0, off);
1688 err = bpf_jit_probe_post(jit, fp, &probe);
1689 if (err < 0)
1690 return err;
1691 jit->seen |= SEEN_MEM;
1692 break;
1693 case BPF_LDX | BPF_MEM | BPF_H: /* dst = *(u16 *)(ul) (src + off) */
1694 case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1695 case BPF_LDX | BPF_PROBE_MEM32 | BPF_H:
1696 bpf_jit_probe_load_pre(jit, insn, &probe);
1697 /* llgh %dst,off(%src,%arena) */
1698 EMIT6_DISP_LH(0xe3000000, 0x0091, dst_reg, src_reg,
1699 probe.arena_reg, off);
1700 err = bpf_jit_probe_post(jit, fp, &probe);
1701 if (err < 0)
1702 return err;
1703 jit->seen |= SEEN_MEM;
1704 if (insn_is_zext(&insn[1]))
1705 insn_count = 2;
1706 break;
1707 case BPF_LDX | BPF_MEMSX | BPF_H: /* dst = *(s16 *)(ul) (src + off) */
1708 case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1709 bpf_jit_probe_load_pre(jit, insn, &probe);
1710 /* lgh %dst,off(%src) */
1711 EMIT6_DISP_LH(0xe3000000, 0x0015, dst_reg, src_reg, REG_0, off);
1712 err = bpf_jit_probe_post(jit, fp, &probe);
1713 if (err < 0)
1714 return err;
1715 jit->seen |= SEEN_MEM;
1716 break;
1717 case BPF_LDX | BPF_MEM | BPF_W: /* dst = *(u32 *)(ul) (src + off) */
1718 case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1719 case BPF_LDX | BPF_PROBE_MEM32 | BPF_W:
1720 bpf_jit_probe_load_pre(jit, insn, &probe);
1721 /* llgf %dst,off(%src) */
1722 jit->seen |= SEEN_MEM;
1723 EMIT6_DISP_LH(0xe3000000, 0x0016, dst_reg, src_reg,
1724 probe.arena_reg, off);
1725 err = bpf_jit_probe_post(jit, fp, &probe);
1726 if (err < 0)
1727 return err;
1728 if (insn_is_zext(&insn[1]))
1729 insn_count = 2;
1730 break;
1731 case BPF_LDX | BPF_MEMSX | BPF_W: /* dst = *(s32 *)(ul) (src + off) */
1732 case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1733 bpf_jit_probe_load_pre(jit, insn, &probe);
1734 /* lgf %dst,off(%src) */
1735 jit->seen |= SEEN_MEM;
1736 EMIT6_DISP_LH(0xe3000000, 0x0014, dst_reg, src_reg, REG_0, off);
1737 err = bpf_jit_probe_post(jit, fp, &probe);
1738 if (err < 0)
1739 return err;
1740 break;
1741 case BPF_LDX | BPF_MEM | BPF_DW: /* dst = *(u64 *)(ul) (src + off) */
1742 case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1743 case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW:
1744 bpf_jit_probe_load_pre(jit, insn, &probe);
1745 /* lg %dst,off(%src,%arena) */
1746 jit->seen |= SEEN_MEM;
1747 EMIT6_DISP_LH(0xe3000000, 0x0004, dst_reg, src_reg,
1748 probe.arena_reg, off);
1749 err = bpf_jit_probe_post(jit, fp, &probe);
1750 if (err < 0)
1751 return err;
1752 break;
1753 /*
1754 * BPF_JMP / CALL
1755 */
1756 case BPF_JMP | BPF_CALL:
1757 {
1758 const struct btf_func_model *m;
1759 bool func_addr_fixed;
1760 int j, ret;
1761 u64 func;
1762
1763 ret = bpf_jit_get_func_addr(fp, insn, extra_pass,
1764 &func, &func_addr_fixed);
1765 if (ret < 0)
1766 return -1;
1767
1768 REG_SET_SEEN(BPF_REG_5);
1769 jit->seen |= SEEN_FUNC;
1770 /*
1771 * Copy the tail call counter to where the callee expects it.
1772 *
1773 * Note 1: The callee can increment the tail call counter, but
1774 * we do not load it back, since the x86 JIT does not do this
1775 * either.
1776 *
1777 * Note 2: We assume that the verifier does not let us call the
1778 * main program, which clears the tail call counter on entry.
1779 */
1780 /* mvc STK_OFF_TCCNT(4,%r15),N(%r15) */
1781 _EMIT6(0xd203f000 | STK_OFF_TCCNT,
1782 0xf000 | (STK_OFF_TCCNT + STK_OFF + stack_depth));
1783
1784 /* Sign-extend the kfunc arguments. */
1785 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
1786 m = bpf_jit_find_kfunc_model(fp, insn);
1787 if (!m)
1788 return -1;
1789
1790 for (j = 0; j < m->nr_args; j++) {
1791 if (sign_extend(jit, BPF_REG_1 + j,
1792 m->arg_size[j],
1793 m->arg_flags[j]))
1794 return -1;
1795 }
1796 }
1797
1798 /* lgrl %w1,func */
1799 EMIT6_PCREL_RILB(0xc4080000, REG_W1, _EMIT_CONST_U64(func));
1800 /* %r1() */
1801 call_r1(jit);
1802 /* lgr %b0,%r2: load return value into %b0 */
1803 EMIT4(0xb9040000, BPF_REG_0, REG_2);
1804 break;
1805 }
1806 case BPF_JMP | BPF_TAIL_CALL: {
1807 int patch_1_clrj, patch_2_clij, patch_3_brc;
1808
1809 /*
1810 * Implicit input:
1811 * B1: pointer to ctx
1812 * B2: pointer to bpf_array
1813 * B3: index in bpf_array
1814 *
1815 * if (index >= array->map.max_entries)
1816 * goto out;
1817 */
1818
1819 /* llgf %w1,map.max_entries(%b2) */
1820 EMIT6_DISP_LH(0xe3000000, 0x0016, REG_W1, REG_0, BPF_REG_2,
1821 offsetof(struct bpf_array, map.max_entries));
1822 /* if ((u32)%b3 >= (u32)%w1) goto out; */
1823 /* clrj %b3,%w1,0xa,out */
1824 patch_1_clrj = jit->prg;
1825 EMIT6_PCREL_RIEB(0xec000000, 0x0077, BPF_REG_3, REG_W1, 0xa,
1826 jit->prg);
1827
1828 /*
1829 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
1830 * goto out;
1831 */
1832
1833 if (jit->seen & SEEN_STACK)
1834 off = STK_OFF_TCCNT + STK_OFF + stack_depth;
1835 else
1836 off = STK_OFF_TCCNT;
1837 /* lhi %w0,1 */
1838 EMIT4_IMM(0xa7080000, REG_W0, 1);
1839 /* laal %w1,%w0,off(%r15) */
1840 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W1, REG_W0, REG_15, off);
1841 /* clij %w1,MAX_TAIL_CALL_CNT-1,0x2,out */
1842 patch_2_clij = jit->prg;
1843 EMIT6_PCREL_RIEC(0xec000000, 0x007f, REG_W1, MAX_TAIL_CALL_CNT - 1,
1844 2, jit->prg);
1845
1846 /*
1847 * prog = array->ptrs[index];
1848 * if (prog == NULL)
1849 * goto out;
1850 */
1851
1852 /* llgfr %r1,%b3: %r1 = (u32) index */
1853 EMIT4(0xb9160000, REG_1, BPF_REG_3);
1854 /* sllg %r1,%r1,3: %r1 *= 8 */
1855 EMIT6_DISP_LH(0xeb000000, 0x000d, REG_1, REG_1, REG_0, 3);
1856 /* ltg %r1,prog(%b2,%r1) */
1857 EMIT6_DISP_LH(0xe3000000, 0x0002, REG_1, BPF_REG_2,
1858 REG_1, offsetof(struct bpf_array, ptrs));
1859 /* brc 0x8,out */
1860 patch_3_brc = jit->prg;
1861 EMIT4_PCREL_RIC(0xa7040000, 8, jit->prg);
1862
1863 /*
1864 * Restore registers before calling function
1865 */
1866 save_restore_regs(jit, REGS_RESTORE, stack_depth, 0);
1867
1868 /*
1869 * goto *(prog->bpf_func + tail_call_start);
1870 */
1871
1872 /* lg %r1,bpf_func(%r1) */
1873 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, REG_1, REG_0,
1874 offsetof(struct bpf_prog, bpf_func));
1875 if (nospec_uses_trampoline()) {
1876 jit->seen |= SEEN_FUNC;
1877 /* aghi %r1,tail_call_start */
1878 EMIT4_IMM(0xa70b0000, REG_1, jit->tail_call_start);
1879 /* brcl 0xf,__s390_indirect_jump_r1 */
1880 EMIT6_PCREL_RILC(0xc0040000, 0xf, jit->r1_thunk_ip);
1881 } else {
1882 /* bc 0xf,tail_call_start(%r1) */
1883 _EMIT4(0x47f01000 + jit->tail_call_start);
1884 }
1885 /* out: */
1886 if (jit->prg_buf) {
1887 *(u16 *)(jit->prg_buf + patch_1_clrj + 2) =
1888 (jit->prg - patch_1_clrj) >> 1;
1889 *(u16 *)(jit->prg_buf + patch_2_clij + 2) =
1890 (jit->prg - patch_2_clij) >> 1;
1891 *(u16 *)(jit->prg_buf + patch_3_brc + 2) =
1892 (jit->prg - patch_3_brc) >> 1;
1893 }
1894 break;
1895 }
1896 case BPF_JMP | BPF_EXIT: /* return b0 */
1897 last = (i == fp->len - 1) ? 1 : 0;
1898 if (last)
1899 break;
1900 if (!is_first_pass(jit) && can_use_rel(jit, jit->exit_ip))
1901 /* brc 0xf, <exit> */
1902 EMIT4_PCREL_RIC(0xa7040000, 0xf, jit->exit_ip);
1903 else
1904 /* brcl 0xf, <exit> */
1905 EMIT6_PCREL_RILC(0xc0040000, 0xf, jit->exit_ip);
1906 break;
1907 /*
1908 * Branch relative (number of skipped instructions) to offset on
1909 * condition.
1910 *
1911 * Condition code to mask mapping:
1912 *
1913 * CC | Description | Mask
1914 * ------------------------------
1915 * 0 | Operands equal | 8
1916 * 1 | First operand low | 4
1917 * 2 | First operand high | 2
1918 * 3 | Unused | 1
1919 *
1920 * For s390x relative branches: ip = ip + off_bytes
1921 * For BPF relative branches: insn = insn + off_insns + 1
1922 *
1923 * For example for s390x with offset 0 we jump to the branch
1924 * instruction itself (loop) and for BPF with offset 0 we
1925 * branch to the instruction behind the branch.
1926 */
1927 case BPF_JMP32 | BPF_JA: /* if (true) */
1928 branch_oc_off = imm;
1929 fallthrough;
1930 case BPF_JMP | BPF_JA: /* if (true) */
1931 mask = 0xf000; /* j */
1932 goto branch_oc;
1933 case BPF_JMP | BPF_JSGT | BPF_K: /* ((s64) dst > (s64) imm) */
1934 case BPF_JMP32 | BPF_JSGT | BPF_K: /* ((s32) dst > (s32) imm) */
1935 mask = 0x2000; /* jh */
1936 goto branch_ks;
1937 case BPF_JMP | BPF_JSLT | BPF_K: /* ((s64) dst < (s64) imm) */
1938 case BPF_JMP32 | BPF_JSLT | BPF_K: /* ((s32) dst < (s32) imm) */
1939 mask = 0x4000; /* jl */
1940 goto branch_ks;
1941 case BPF_JMP | BPF_JSGE | BPF_K: /* ((s64) dst >= (s64) imm) */
1942 case BPF_JMP32 | BPF_JSGE | BPF_K: /* ((s32) dst >= (s32) imm) */
1943 mask = 0xa000; /* jhe */
1944 goto branch_ks;
1945 case BPF_JMP | BPF_JSLE | BPF_K: /* ((s64) dst <= (s64) imm) */
1946 case BPF_JMP32 | BPF_JSLE | BPF_K: /* ((s32) dst <= (s32) imm) */
1947 mask = 0xc000; /* jle */
1948 goto branch_ks;
1949 case BPF_JMP | BPF_JGT | BPF_K: /* (dst_reg > imm) */
1950 case BPF_JMP32 | BPF_JGT | BPF_K: /* ((u32) dst_reg > (u32) imm) */
1951 mask = 0x2000; /* jh */
1952 goto branch_ku;
1953 case BPF_JMP | BPF_JLT | BPF_K: /* (dst_reg < imm) */
1954 case BPF_JMP32 | BPF_JLT | BPF_K: /* ((u32) dst_reg < (u32) imm) */
1955 mask = 0x4000; /* jl */
1956 goto branch_ku;
1957 case BPF_JMP | BPF_JGE | BPF_K: /* (dst_reg >= imm) */
1958 case BPF_JMP32 | BPF_JGE | BPF_K: /* ((u32) dst_reg >= (u32) imm) */
1959 mask = 0xa000; /* jhe */
1960 goto branch_ku;
1961 case BPF_JMP | BPF_JLE | BPF_K: /* (dst_reg <= imm) */
1962 case BPF_JMP32 | BPF_JLE | BPF_K: /* ((u32) dst_reg <= (u32) imm) */
1963 mask = 0xc000; /* jle */
1964 goto branch_ku;
1965 case BPF_JMP | BPF_JNE | BPF_K: /* (dst_reg != imm) */
1966 case BPF_JMP32 | BPF_JNE | BPF_K: /* ((u32) dst_reg != (u32) imm) */
1967 mask = 0x7000; /* jne */
1968 goto branch_ku;
1969 case BPF_JMP | BPF_JEQ | BPF_K: /* (dst_reg == imm) */
1970 case BPF_JMP32 | BPF_JEQ | BPF_K: /* ((u32) dst_reg == (u32) imm) */
1971 mask = 0x8000; /* je */
1972 goto branch_ku;
1973 case BPF_JMP | BPF_JSET | BPF_K: /* (dst_reg & imm) */
1974 case BPF_JMP32 | BPF_JSET | BPF_K: /* ((u32) dst_reg & (u32) imm) */
1975 mask = 0x7000; /* jnz */
1976 if (BPF_CLASS(insn->code) == BPF_JMP32) {
1977 /* llilf %w1,imm (load zero extend imm) */
1978 EMIT6_IMM(0xc00f0000, REG_W1, imm);
1979 /* nr %w1,%dst */
1980 EMIT2(0x1400, REG_W1, dst_reg);
1981 } else {
1982 /* lgfi %w1,imm (load sign extend imm) */
1983 EMIT6_IMM(0xc0010000, REG_W1, imm);
1984 /* ngr %w1,%dst */
1985 EMIT4(0xb9800000, REG_W1, dst_reg);
1986 }
1987 goto branch_oc;
1988
1989 case BPF_JMP | BPF_JSGT | BPF_X: /* ((s64) dst > (s64) src) */
1990 case BPF_JMP32 | BPF_JSGT | BPF_X: /* ((s32) dst > (s32) src) */
1991 mask = 0x2000; /* jh */
1992 goto branch_xs;
1993 case BPF_JMP | BPF_JSLT | BPF_X: /* ((s64) dst < (s64) src) */
1994 case BPF_JMP32 | BPF_JSLT | BPF_X: /* ((s32) dst < (s32) src) */
1995 mask = 0x4000; /* jl */
1996 goto branch_xs;
1997 case BPF_JMP | BPF_JSGE | BPF_X: /* ((s64) dst >= (s64) src) */
1998 case BPF_JMP32 | BPF_JSGE | BPF_X: /* ((s32) dst >= (s32) src) */
1999 mask = 0xa000; /* jhe */
2000 goto branch_xs;
2001 case BPF_JMP | BPF_JSLE | BPF_X: /* ((s64) dst <= (s64) src) */
2002 case BPF_JMP32 | BPF_JSLE | BPF_X: /* ((s32) dst <= (s32) src) */
2003 mask = 0xc000; /* jle */
2004 goto branch_xs;
2005 case BPF_JMP | BPF_JGT | BPF_X: /* (dst > src) */
2006 case BPF_JMP32 | BPF_JGT | BPF_X: /* ((u32) dst > (u32) src) */
2007 mask = 0x2000; /* jh */
2008 goto branch_xu;
2009 case BPF_JMP | BPF_JLT | BPF_X: /* (dst < src) */
2010 case BPF_JMP32 | BPF_JLT | BPF_X: /* ((u32) dst < (u32) src) */
2011 mask = 0x4000; /* jl */
2012 goto branch_xu;
2013 case BPF_JMP | BPF_JGE | BPF_X: /* (dst >= src) */
2014 case BPF_JMP32 | BPF_JGE | BPF_X: /* ((u32) dst >= (u32) src) */
2015 mask = 0xa000; /* jhe */
2016 goto branch_xu;
2017 case BPF_JMP | BPF_JLE | BPF_X: /* (dst <= src) */
2018 case BPF_JMP32 | BPF_JLE | BPF_X: /* ((u32) dst <= (u32) src) */
2019 mask = 0xc000; /* jle */
2020 goto branch_xu;
2021 case BPF_JMP | BPF_JNE | BPF_X: /* (dst != src) */
2022 case BPF_JMP32 | BPF_JNE | BPF_X: /* ((u32) dst != (u32) src) */
2023 mask = 0x7000; /* jne */
2024 goto branch_xu;
2025 case BPF_JMP | BPF_JEQ | BPF_X: /* (dst == src) */
2026 case BPF_JMP32 | BPF_JEQ | BPF_X: /* ((u32) dst == (u32) src) */
2027 mask = 0x8000; /* je */
2028 goto branch_xu;
2029 case BPF_JMP | BPF_JSET | BPF_X: /* (dst & src) */
2030 case BPF_JMP32 | BPF_JSET | BPF_X: /* ((u32) dst & (u32) src) */
2031 {
2032 bool is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
2033
2034 mask = 0x7000; /* jnz */
2035 /* nrk or ngrk %w1,%dst,%src */
2036 EMIT4_RRF((is_jmp32 ? 0xb9f40000 : 0xb9e40000),
2037 REG_W1, dst_reg, src_reg);
2038 goto branch_oc;
2039 branch_ks:
2040 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
2041 /* cfi or cgfi %dst,imm */
2042 EMIT6_IMM(is_jmp32 ? 0xc20d0000 : 0xc20c0000,
2043 dst_reg, imm);
2044 if (!is_first_pass(jit) &&
2045 can_use_rel(jit, addrs[i + off + 1])) {
2046 /* brc mask,off */
2047 EMIT4_PCREL_RIC(0xa7040000,
2048 mask >> 12, addrs[i + off + 1]);
2049 } else {
2050 /* brcl mask,off */
2051 EMIT6_PCREL_RILC(0xc0040000,
2052 mask >> 12, addrs[i + off + 1]);
2053 }
2054 break;
2055 branch_ku:
2056 /* lgfi %w1,imm (load sign extend imm) */
2057 src_reg = REG_1;
2058 EMIT6_IMM(0xc0010000, src_reg, imm);
2059 goto branch_xu;
2060 branch_xs:
2061 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
2062 if (!is_first_pass(jit) &&
2063 can_use_rel(jit, addrs[i + off + 1])) {
2064 /* crj or cgrj %dst,%src,mask,off */
2065 EMIT6_PCREL(0xec000000, (is_jmp32 ? 0x0076 : 0x0064),
2066 dst_reg, src_reg, i, off, mask);
2067 } else {
2068 /* cr or cgr %dst,%src */
2069 if (is_jmp32)
2070 EMIT2(0x1900, dst_reg, src_reg);
2071 else
2072 EMIT4(0xb9200000, dst_reg, src_reg);
2073 /* brcl mask,off */
2074 EMIT6_PCREL_RILC(0xc0040000,
2075 mask >> 12, addrs[i + off + 1]);
2076 }
2077 break;
2078 branch_xu:
2079 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
2080 if (!is_first_pass(jit) &&
2081 can_use_rel(jit, addrs[i + off + 1])) {
2082 /* clrj or clgrj %dst,%src,mask,off */
2083 EMIT6_PCREL(0xec000000, (is_jmp32 ? 0x0077 : 0x0065),
2084 dst_reg, src_reg, i, off, mask);
2085 } else {
2086 /* clr or clgr %dst,%src */
2087 if (is_jmp32)
2088 EMIT2(0x1500, dst_reg, src_reg);
2089 else
2090 EMIT4(0xb9210000, dst_reg, src_reg);
2091 /* brcl mask,off */
2092 EMIT6_PCREL_RILC(0xc0040000,
2093 mask >> 12, addrs[i + off + 1]);
2094 }
2095 break;
2096 branch_oc:
2097 if (!is_first_pass(jit) &&
2098 can_use_rel(jit, addrs[i + branch_oc_off + 1])) {
2099 /* brc mask,off */
2100 EMIT4_PCREL_RIC(0xa7040000,
2101 mask >> 12,
2102 addrs[i + branch_oc_off + 1]);
2103 } else {
2104 /* brcl mask,off */
2105 EMIT6_PCREL_RILC(0xc0040000,
2106 mask >> 12,
2107 addrs[i + branch_oc_off + 1]);
2108 }
2109 break;
2110 }
2111 default: /* too complex, give up */
2112 pr_err("Unknown opcode %02x\n", insn->code);
2113 return -1;
2114 }
2115
2116 return insn_count;
2117 }
2118
2119 /*
2120 * Return whether new i-th instruction address does not violate any invariant
2121 */
2122 static bool bpf_is_new_addr_sane(struct bpf_jit *jit, int i)
2123 {
2124 /* On the first pass anything goes */
2125 if (is_first_pass(jit))
2126 return true;
2127
2128 /* The codegen pass must not change anything */
2129 if (is_codegen_pass(jit))
2130 return jit->addrs[i] == jit->prg;
2131
2132 /* Passes in between must not increase code size */
2133 return jit->addrs[i] >= jit->prg;
2134 }
2135
2136 /*
2137 * Update the address of i-th instruction
2138 */
2139 static int bpf_set_addr(struct bpf_jit *jit, int i)
2140 {
2141 int delta;
2142
2143 if (is_codegen_pass(jit)) {
2144 delta = jit->prg - jit->addrs[i];
2145 if (delta < 0)
2146 bpf_skip(jit, -delta);
2147 }
2148 if (WARN_ON_ONCE(!bpf_is_new_addr_sane(jit, i)))
2149 return -1;
2150 jit->addrs[i] = jit->prg;
2151 return 0;
2152 }
2153
2154 /*
2155 * Compile eBPF program into s390x code
2156 */
2157 static int bpf_jit_prog(struct bpf_jit *jit, struct bpf_prog *fp,
2158 bool extra_pass, u32 stack_depth)
2159 {
2160 int i, insn_count, lit32_size, lit64_size;
2161 u64 kern_arena;
2162
2163 jit->lit32 = jit->lit32_start;
2164 jit->lit64 = jit->lit64_start;
2165 jit->prg = 0;
2166 jit->excnt = 0;
2167
2168 kern_arena = bpf_arena_get_kern_vm_start(fp->aux->arena);
2169 if (kern_arena)
2170 jit->kern_arena = _EMIT_CONST_U64(kern_arena);
2171 jit->user_arena = bpf_arena_get_user_vm_start(fp->aux->arena);
2172
2173 bpf_jit_prologue(jit, fp, stack_depth);
2174 if (bpf_set_addr(jit, 0) < 0)
2175 return -1;
2176 for (i = 0; i < fp->len; i += insn_count) {
2177 insn_count = bpf_jit_insn(jit, fp, i, extra_pass, stack_depth);
2178 if (insn_count < 0)
2179 return -1;
2180 /* Next instruction address */
2181 if (bpf_set_addr(jit, i + insn_count) < 0)
2182 return -1;
2183 }
2184 bpf_jit_epilogue(jit, stack_depth);
2185
2186 lit32_size = jit->lit32 - jit->lit32_start;
2187 lit64_size = jit->lit64 - jit->lit64_start;
2188 jit->lit32_start = jit->prg;
2189 if (lit32_size)
2190 jit->lit32_start = ALIGN(jit->lit32_start, 4);
2191 jit->lit64_start = jit->lit32_start + lit32_size;
2192 if (lit64_size)
2193 jit->lit64_start = ALIGN(jit->lit64_start, 8);
2194 jit->size = jit->lit64_start + lit64_size;
2195 jit->size_prg = jit->prg;
2196
2197 if (WARN_ON_ONCE(fp->aux->extable &&
2198 jit->excnt != fp->aux->num_exentries))
2199 /* Verifier bug - too many entries. */
2200 return -1;
2201
2202 return 0;
2203 }
2204
2205 bool bpf_jit_needs_zext(void)
2206 {
2207 return true;
2208 }
2209
2210 struct s390_jit_data {
2211 struct bpf_binary_header *header;
2212 struct bpf_jit ctx;
2213 int pass;
2214 };
2215
2216 static struct bpf_binary_header *bpf_jit_alloc(struct bpf_jit *jit,
2217 struct bpf_prog *fp)
2218 {
2219 struct bpf_binary_header *header;
2220 struct bpf_insn *insn;
2221 u32 extable_size;
2222 u32 code_size;
2223 int i;
2224
2225 for (i = 0; i < fp->len; i++) {
2226 insn = &fp->insnsi[i];
2227
2228 if (BPF_CLASS(insn->code) == BPF_STX &&
2229 BPF_MODE(insn->code) == BPF_PROBE_ATOMIC &&
2230 (BPF_SIZE(insn->code) == BPF_DW ||
2231 BPF_SIZE(insn->code) == BPF_W) &&
2232 insn->imm == BPF_XCHG)
2233 /*
2234 * bpf_jit_insn() emits a load and a compare-and-swap,
2235 * both of which need to be probed.
2236 */
2237 fp->aux->num_exentries += 1;
2238 }
2239 /* We need two entries per insn. */
2240 fp->aux->num_exentries *= 2;
2241
2242 code_size = roundup(jit->size,
2243 __alignof__(struct exception_table_entry));
2244 extable_size = fp->aux->num_exentries *
2245 sizeof(struct exception_table_entry);
2246 header = bpf_jit_binary_alloc(code_size + extable_size, &jit->prg_buf,
2247 8, jit_fill_hole);
2248 if (!header)
2249 return NULL;
2250 fp->aux->extable = (struct exception_table_entry *)
2251 (jit->prg_buf + code_size);
2252 return header;
2253 }
2254
2255 /*
2256 * Compile eBPF program "fp"
2257 */
2258 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
2259 {
2260 u32 stack_depth = round_up(fp->aux->stack_depth, 8);
2261 struct bpf_prog *tmp, *orig_fp = fp;
2262 struct bpf_binary_header *header;
2263 struct s390_jit_data *jit_data;
2264 bool tmp_blinded = false;
2265 bool extra_pass = false;
2266 struct bpf_jit jit;
2267 int pass;
2268
2269 if (!fp->jit_requested)
2270 return orig_fp;
2271
2272 tmp = bpf_jit_blind_constants(fp);
2273 /*
2274 * If blinding was requested and we failed during blinding,
2275 * we must fall back to the interpreter.
2276 */
2277 if (IS_ERR(tmp))
2278 return orig_fp;
2279 if (tmp != fp) {
2280 tmp_blinded = true;
2281 fp = tmp;
2282 }
2283
2284 jit_data = fp->aux->jit_data;
2285 if (!jit_data) {
2286 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2287 if (!jit_data) {
2288 fp = orig_fp;
2289 goto out;
2290 }
2291 fp->aux->jit_data = jit_data;
2292 }
2293 if (jit_data->ctx.addrs) {
2294 jit = jit_data->ctx;
2295 header = jit_data->header;
2296 extra_pass = true;
2297 pass = jit_data->pass + 1;
2298 goto skip_init_ctx;
2299 }
2300
2301 memset(&jit, 0, sizeof(jit));
2302 jit.addrs = kvcalloc(fp->len + 1, sizeof(*jit.addrs), GFP_KERNEL);
2303 if (jit.addrs == NULL) {
2304 fp = orig_fp;
2305 goto free_addrs;
2306 }
2307 /*
2308 * Three initial passes:
2309 * - 1/2: Determine clobbered registers
2310 * - 3: Calculate program size and addrs array
2311 */
2312 for (pass = 1; pass <= 3; pass++) {
2313 if (bpf_jit_prog(&jit, fp, extra_pass, stack_depth)) {
2314 fp = orig_fp;
2315 goto free_addrs;
2316 }
2317 }
2318 /*
2319 * Final pass: Allocate and generate program
2320 */
2321 header = bpf_jit_alloc(&jit, fp);
2322 if (!header) {
2323 fp = orig_fp;
2324 goto free_addrs;
2325 }
2326 skip_init_ctx:
2327 if (bpf_jit_prog(&jit, fp, extra_pass, stack_depth)) {
2328 bpf_jit_binary_free(header);
2329 fp = orig_fp;
2330 goto free_addrs;
2331 }
2332 if (bpf_jit_enable > 1) {
2333 bpf_jit_dump(fp->len, jit.size, pass, jit.prg_buf);
2334 print_fn_code(jit.prg_buf, jit.size_prg);
2335 }
2336 if (!fp->is_func || extra_pass) {
2337 if (bpf_jit_binary_lock_ro(header)) {
2338 bpf_jit_binary_free(header);
2339 fp = orig_fp;
2340 goto free_addrs;
2341 }
2342 } else {
2343 jit_data->header = header;
2344 jit_data->ctx = jit;
2345 jit_data->pass = pass;
2346 }
2347 fp->bpf_func = (void *) jit.prg_buf;
2348 fp->jited = 1;
2349 fp->jited_len = jit.size;
2350
2351 if (!fp->is_func || extra_pass) {
2352 bpf_prog_fill_jited_linfo(fp, jit.addrs + 1);
2353 free_addrs:
2354 kvfree(jit.addrs);
2355 kfree(jit_data);
2356 fp->aux->jit_data = NULL;
2357 }
2358 out:
2359 if (tmp_blinded)
2360 bpf_jit_prog_release_other(fp, fp == orig_fp ?
2361 tmp : orig_fp);
2362 return fp;
2363 }
2364
2365 bool bpf_jit_supports_kfunc_call(void)
2366 {
2367 return true;
2368 }
2369
2370 bool bpf_jit_supports_far_kfunc_call(void)
2371 {
2372 return true;
2373 }
2374
2375 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2376 void *old_addr, void *new_addr)
2377 {
2378 struct bpf_plt expected_plt, current_plt, new_plt, *plt;
2379 struct {
2380 u16 opc;
2381 s32 disp;
2382 } __packed insn;
2383 char *ret;
2384 int err;
2385
2386 /* Verify the branch to be patched. */
2387 err = copy_from_kernel_nofault(&insn, ip, sizeof(insn));
2388 if (err < 0)
2389 return err;
2390 if (insn.opc != (0xc004 | (old_addr ? 0xf0 : 0)))
2391 return -EINVAL;
2392
2393 if (t == BPF_MOD_JUMP &&
2394 insn.disp == ((char *)new_addr - (char *)ip) >> 1) {
2395 /*
2396 * The branch already points to the destination,
2397 * there is no PLT.
2398 */
2399 } else {
2400 /* Verify the PLT. */
2401 plt = ip + (insn.disp << 1);
2402 err = copy_from_kernel_nofault(&current_plt, plt,
2403 sizeof(current_plt));
2404 if (err < 0)
2405 return err;
2406 ret = (char *)ip + 6;
2407 bpf_jit_plt(&expected_plt, ret, old_addr);
2408 if (memcmp(&current_plt, &expected_plt, sizeof(current_plt)))
2409 return -EINVAL;
2410 /* Adjust the call address. */
2411 bpf_jit_plt(&new_plt, ret, new_addr);
2412 s390_kernel_write(&plt->target, &new_plt.target,
2413 sizeof(void *));
2414 }
2415
2416 /* Adjust the mask of the branch. */
2417 insn.opc = 0xc004 | (new_addr ? 0xf0 : 0);
2418 s390_kernel_write((char *)ip + 1, (char *)&insn.opc + 1, 1);
2419
2420 /* Make the new code visible to the other CPUs. */
2421 text_poke_sync_lock();
2422
2423 return 0;
2424 }
2425
2426 struct bpf_tramp_jit {
2427 struct bpf_jit common;
2428 int orig_stack_args_off;/* Offset of arguments placed on stack by the
2429 * func_addr's original caller
2430 */
2431 int stack_size; /* Trampoline stack size */
2432 int backchain_off; /* Offset of backchain */
2433 int stack_args_off; /* Offset of stack arguments for calling
2434 * func_addr, has to be at the top
2435 */
2436 int reg_args_off; /* Offset of register arguments for calling
2437 * func_addr
2438 */
2439 int ip_off; /* For bpf_get_func_ip(), has to be at
2440 * (ctx - 16)
2441 */
2442 int arg_cnt_off; /* For bpf_get_func_arg_cnt(), has to be at
2443 * (ctx - 8)
2444 */
2445 int bpf_args_off; /* Offset of BPF_PROG context, which consists
2446 * of BPF arguments followed by return value
2447 */
2448 int retval_off; /* Offset of return value (see above) */
2449 int r7_r8_off; /* Offset of saved %r7 and %r8, which are used
2450 * for __bpf_prog_enter() return value and
2451 * func_addr respectively
2452 */
2453 int run_ctx_off; /* Offset of struct bpf_tramp_run_ctx */
2454 int tccnt_off; /* Offset of saved tailcall counter */
2455 int r14_off; /* Offset of saved %r14, has to be at the
2456 * bottom */
2457 int do_fexit; /* do_fexit: label */
2458 };
2459
2460 static void load_imm64(struct bpf_jit *jit, int dst_reg, u64 val)
2461 {
2462 /* llihf %dst_reg,val_hi */
2463 EMIT6_IMM(0xc00e0000, dst_reg, (val >> 32));
2464 /* oilf %rdst_reg,val_lo */
2465 EMIT6_IMM(0xc00d0000, dst_reg, val);
2466 }
2467
2468 static int invoke_bpf_prog(struct bpf_tramp_jit *tjit,
2469 const struct btf_func_model *m,
2470 struct bpf_tramp_link *tlink, bool save_ret)
2471 {
2472 struct bpf_jit *jit = &tjit->common;
2473 int cookie_off = tjit->run_ctx_off +
2474 offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
2475 struct bpf_prog *p = tlink->link.prog;
2476 int patch;
2477
2478 /*
2479 * run_ctx.cookie = tlink->cookie;
2480 */
2481
2482 /* %r0 = tlink->cookie */
2483 load_imm64(jit, REG_W0, tlink->cookie);
2484 /* stg %r0,cookie_off(%r15) */
2485 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W0, REG_0, REG_15, cookie_off);
2486
2487 /*
2488 * if ((start = __bpf_prog_enter(p, &run_ctx)) == 0)
2489 * goto skip;
2490 */
2491
2492 /* %r1 = __bpf_prog_enter */
2493 load_imm64(jit, REG_1, (u64)bpf_trampoline_enter(p));
2494 /* %r2 = p */
2495 load_imm64(jit, REG_2, (u64)p);
2496 /* la %r3,run_ctx_off(%r15) */
2497 EMIT4_DISP(0x41000000, REG_3, REG_15, tjit->run_ctx_off);
2498 /* %r1() */
2499 call_r1(jit);
2500 /* ltgr %r7,%r2 */
2501 EMIT4(0xb9020000, REG_7, REG_2);
2502 /* brcl 8,skip */
2503 patch = jit->prg;
2504 EMIT6_PCREL_RILC(0xc0040000, 8, 0);
2505
2506 /*
2507 * retval = bpf_func(args, p->insnsi);
2508 */
2509
2510 /* %r1 = p->bpf_func */
2511 load_imm64(jit, REG_1, (u64)p->bpf_func);
2512 /* la %r2,bpf_args_off(%r15) */
2513 EMIT4_DISP(0x41000000, REG_2, REG_15, tjit->bpf_args_off);
2514 /* %r3 = p->insnsi */
2515 if (!p->jited)
2516 load_imm64(jit, REG_3, (u64)p->insnsi);
2517 /* %r1() */
2518 call_r1(jit);
2519 /* stg %r2,retval_off(%r15) */
2520 if (save_ret) {
2521 if (sign_extend(jit, REG_2, m->ret_size, m->ret_flags))
2522 return -1;
2523 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_2, REG_0, REG_15,
2524 tjit->retval_off);
2525 }
2526
2527 /* skip: */
2528 if (jit->prg_buf)
2529 *(u32 *)&jit->prg_buf[patch + 2] = (jit->prg - patch) >> 1;
2530
2531 /*
2532 * __bpf_prog_exit(p, start, &run_ctx);
2533 */
2534
2535 /* %r1 = __bpf_prog_exit */
2536 load_imm64(jit, REG_1, (u64)bpf_trampoline_exit(p));
2537 /* %r2 = p */
2538 load_imm64(jit, REG_2, (u64)p);
2539 /* lgr %r3,%r7 */
2540 EMIT4(0xb9040000, REG_3, REG_7);
2541 /* la %r4,run_ctx_off(%r15) */
2542 EMIT4_DISP(0x41000000, REG_4, REG_15, tjit->run_ctx_off);
2543 /* %r1() */
2544 call_r1(jit);
2545
2546 return 0;
2547 }
2548
2549 static int alloc_stack(struct bpf_tramp_jit *tjit, size_t size)
2550 {
2551 int stack_offset = tjit->stack_size;
2552
2553 tjit->stack_size += size;
2554 return stack_offset;
2555 }
2556
2557 /* ABI uses %r2 - %r6 for parameter passing. */
2558 #define MAX_NR_REG_ARGS 5
2559
2560 /* The "L" field of the "mvc" instruction is 8 bits. */
2561 #define MAX_MVC_SIZE 256
2562 #define MAX_NR_STACK_ARGS (MAX_MVC_SIZE / sizeof(u64))
2563
2564 /* -mfentry generates a 6-byte nop on s390x. */
2565 #define S390X_PATCH_SIZE 6
2566
2567 static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im,
2568 struct bpf_tramp_jit *tjit,
2569 const struct btf_func_model *m,
2570 u32 flags,
2571 struct bpf_tramp_links *tlinks,
2572 void *func_addr)
2573 {
2574 struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2575 struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2576 struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2577 int nr_bpf_args, nr_reg_args, nr_stack_args;
2578 struct bpf_jit *jit = &tjit->common;
2579 int arg, bpf_arg_off;
2580 int i, j;
2581
2582 /* Support as many stack arguments as "mvc" instruction can handle. */
2583 nr_reg_args = min_t(int, m->nr_args, MAX_NR_REG_ARGS);
2584 nr_stack_args = m->nr_args - nr_reg_args;
2585 if (nr_stack_args > MAX_NR_STACK_ARGS)
2586 return -ENOTSUPP;
2587
2588 /* Return to %r14, since func_addr and %r0 are not available. */
2589 if ((!func_addr && !(flags & BPF_TRAMP_F_ORIG_STACK)) ||
2590 (flags & BPF_TRAMP_F_INDIRECT))
2591 flags |= BPF_TRAMP_F_SKIP_FRAME;
2592
2593 /*
2594 * Compute how many arguments we need to pass to BPF programs.
2595 * BPF ABI mirrors that of x86_64: arguments that are 16 bytes or
2596 * smaller are packed into 1 or 2 registers; larger arguments are
2597 * passed via pointers.
2598 * In s390x ABI, arguments that are 8 bytes or smaller are packed into
2599 * a register; larger arguments are passed via pointers.
2600 * We need to deal with this difference.
2601 */
2602 nr_bpf_args = 0;
2603 for (i = 0; i < m->nr_args; i++) {
2604 if (m->arg_size[i] <= 8)
2605 nr_bpf_args += 1;
2606 else if (m->arg_size[i] <= 16)
2607 nr_bpf_args += 2;
2608 else
2609 return -ENOTSUPP;
2610 }
2611
2612 /*
2613 * Calculate the stack layout.
2614 */
2615
2616 /*
2617 * Allocate STACK_FRAME_OVERHEAD bytes for the callees. As the s390x
2618 * ABI requires, put our backchain at the end of the allocated memory.
2619 */
2620 tjit->stack_size = STACK_FRAME_OVERHEAD;
2621 tjit->backchain_off = tjit->stack_size - sizeof(u64);
2622 tjit->stack_args_off = alloc_stack(tjit, nr_stack_args * sizeof(u64));
2623 tjit->reg_args_off = alloc_stack(tjit, nr_reg_args * sizeof(u64));
2624 tjit->ip_off = alloc_stack(tjit, sizeof(u64));
2625 tjit->arg_cnt_off = alloc_stack(tjit, sizeof(u64));
2626 tjit->bpf_args_off = alloc_stack(tjit, nr_bpf_args * sizeof(u64));
2627 tjit->retval_off = alloc_stack(tjit, sizeof(u64));
2628 tjit->r7_r8_off = alloc_stack(tjit, 2 * sizeof(u64));
2629 tjit->run_ctx_off = alloc_stack(tjit,
2630 sizeof(struct bpf_tramp_run_ctx));
2631 tjit->tccnt_off = alloc_stack(tjit, sizeof(u64));
2632 tjit->r14_off = alloc_stack(tjit, sizeof(u64) * 2);
2633 /*
2634 * In accordance with the s390x ABI, the caller has allocated
2635 * STACK_FRAME_OVERHEAD bytes for us. 8 of them contain the caller's
2636 * backchain, and the rest we can use.
2637 */
2638 tjit->stack_size -= STACK_FRAME_OVERHEAD - sizeof(u64);
2639 tjit->orig_stack_args_off = tjit->stack_size + STACK_FRAME_OVERHEAD;
2640
2641 /* lgr %r1,%r15 */
2642 EMIT4(0xb9040000, REG_1, REG_15);
2643 /* aghi %r15,-stack_size */
2644 EMIT4_IMM(0xa70b0000, REG_15, -tjit->stack_size);
2645 /* stg %r1,backchain_off(%r15) */
2646 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_1, REG_0, REG_15,
2647 tjit->backchain_off);
2648 /* mvc tccnt_off(4,%r15),stack_size+STK_OFF_TCCNT(%r15) */
2649 _EMIT6(0xd203f000 | tjit->tccnt_off,
2650 0xf000 | (tjit->stack_size + STK_OFF_TCCNT));
2651 /* stmg %r2,%rN,fwd_reg_args_off(%r15) */
2652 if (nr_reg_args)
2653 EMIT6_DISP_LH(0xeb000000, 0x0024, REG_2,
2654 REG_2 + (nr_reg_args - 1), REG_15,
2655 tjit->reg_args_off);
2656 for (i = 0, j = 0; i < m->nr_args; i++) {
2657 if (i < MAX_NR_REG_ARGS)
2658 arg = REG_2 + i;
2659 else
2660 arg = tjit->orig_stack_args_off +
2661 (i - MAX_NR_REG_ARGS) * sizeof(u64);
2662 bpf_arg_off = tjit->bpf_args_off + j * sizeof(u64);
2663 if (m->arg_size[i] <= 8) {
2664 if (i < MAX_NR_REG_ARGS)
2665 /* stg %arg,bpf_arg_off(%r15) */
2666 EMIT6_DISP_LH(0xe3000000, 0x0024, arg,
2667 REG_0, REG_15, bpf_arg_off);
2668 else
2669 /* mvc bpf_arg_off(8,%r15),arg(%r15) */
2670 _EMIT6(0xd207f000 | bpf_arg_off,
2671 0xf000 | arg);
2672 j += 1;
2673 } else {
2674 if (i < MAX_NR_REG_ARGS) {
2675 /* mvc bpf_arg_off(16,%r15),0(%arg) */
2676 _EMIT6(0xd20ff000 | bpf_arg_off,
2677 reg2hex[arg] << 12);
2678 } else {
2679 /* lg %r1,arg(%r15) */
2680 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, REG_0,
2681 REG_15, arg);
2682 /* mvc bpf_arg_off(16,%r15),0(%r1) */
2683 _EMIT6(0xd20ff000 | bpf_arg_off, 0x1000);
2684 }
2685 j += 2;
2686 }
2687 }
2688 /* stmg %r7,%r8,r7_r8_off(%r15) */
2689 EMIT6_DISP_LH(0xeb000000, 0x0024, REG_7, REG_8, REG_15,
2690 tjit->r7_r8_off);
2691 /* stg %r14,r14_off(%r15) */
2692 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_14, REG_0, REG_15, tjit->r14_off);
2693
2694 if (flags & BPF_TRAMP_F_ORIG_STACK) {
2695 /*
2696 * The ftrace trampoline puts the return address (which is the
2697 * address of the original function + S390X_PATCH_SIZE) into
2698 * %r0; see ftrace_shared_hotpatch_trampoline_br and
2699 * ftrace_init_nop() for details.
2700 */
2701
2702 /* lgr %r8,%r0 */
2703 EMIT4(0xb9040000, REG_8, REG_0);
2704 } else {
2705 /* %r8 = func_addr + S390X_PATCH_SIZE */
2706 load_imm64(jit, REG_8, (u64)func_addr + S390X_PATCH_SIZE);
2707 }
2708
2709 /*
2710 * ip = func_addr;
2711 * arg_cnt = m->nr_args;
2712 */
2713
2714 if (flags & BPF_TRAMP_F_IP_ARG) {
2715 /* %r0 = func_addr */
2716 load_imm64(jit, REG_0, (u64)func_addr);
2717 /* stg %r0,ip_off(%r15) */
2718 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_0, REG_0, REG_15,
2719 tjit->ip_off);
2720 }
2721 /* lghi %r0,nr_bpf_args */
2722 EMIT4_IMM(0xa7090000, REG_0, nr_bpf_args);
2723 /* stg %r0,arg_cnt_off(%r15) */
2724 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_0, REG_0, REG_15,
2725 tjit->arg_cnt_off);
2726
2727 if (flags & BPF_TRAMP_F_CALL_ORIG) {
2728 /*
2729 * __bpf_tramp_enter(im);
2730 */
2731
2732 /* %r1 = __bpf_tramp_enter */
2733 load_imm64(jit, REG_1, (u64)__bpf_tramp_enter);
2734 /* %r2 = im */
2735 load_imm64(jit, REG_2, (u64)im);
2736 /* %r1() */
2737 call_r1(jit);
2738 }
2739
2740 for (i = 0; i < fentry->nr_links; i++)
2741 if (invoke_bpf_prog(tjit, m, fentry->links[i],
2742 flags & BPF_TRAMP_F_RET_FENTRY_RET))
2743 return -EINVAL;
2744
2745 if (fmod_ret->nr_links) {
2746 /*
2747 * retval = 0;
2748 */
2749
2750 /* xc retval_off(8,%r15),retval_off(%r15) */
2751 _EMIT6(0xd707f000 | tjit->retval_off,
2752 0xf000 | tjit->retval_off);
2753
2754 for (i = 0; i < fmod_ret->nr_links; i++) {
2755 if (invoke_bpf_prog(tjit, m, fmod_ret->links[i], true))
2756 return -EINVAL;
2757
2758 /*
2759 * if (retval)
2760 * goto do_fexit;
2761 */
2762
2763 /* ltg %r0,retval_off(%r15) */
2764 EMIT6_DISP_LH(0xe3000000, 0x0002, REG_0, REG_0, REG_15,
2765 tjit->retval_off);
2766 /* brcl 7,do_fexit */
2767 EMIT6_PCREL_RILC(0xc0040000, 7, tjit->do_fexit);
2768 }
2769 }
2770
2771 if (flags & BPF_TRAMP_F_CALL_ORIG) {
2772 /*
2773 * retval = func_addr(args);
2774 */
2775
2776 /* lmg %r2,%rN,reg_args_off(%r15) */
2777 if (nr_reg_args)
2778 EMIT6_DISP_LH(0xeb000000, 0x0004, REG_2,
2779 REG_2 + (nr_reg_args - 1), REG_15,
2780 tjit->reg_args_off);
2781 /* mvc stack_args_off(N,%r15),orig_stack_args_off(%r15) */
2782 if (nr_stack_args)
2783 _EMIT6(0xd200f000 |
2784 (nr_stack_args * sizeof(u64) - 1) << 16 |
2785 tjit->stack_args_off,
2786 0xf000 | tjit->orig_stack_args_off);
2787 /* mvc STK_OFF_TCCNT(4,%r15),tccnt_off(%r15) */
2788 _EMIT6(0xd203f000 | STK_OFF_TCCNT, 0xf000 | tjit->tccnt_off);
2789 /* lgr %r1,%r8 */
2790 EMIT4(0xb9040000, REG_1, REG_8);
2791 /* %r1() */
2792 call_r1(jit);
2793 /* stg %r2,retval_off(%r15) */
2794 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_2, REG_0, REG_15,
2795 tjit->retval_off);
2796
2797 im->ip_after_call = jit->prg_buf + jit->prg;
2798
2799 /*
2800 * The following nop will be patched by bpf_tramp_image_put().
2801 */
2802
2803 /* brcl 0,im->ip_epilogue */
2804 EMIT6_PCREL_RILC(0xc0040000, 0, (u64)im->ip_epilogue);
2805 }
2806
2807 /* do_fexit: */
2808 tjit->do_fexit = jit->prg;
2809 for (i = 0; i < fexit->nr_links; i++)
2810 if (invoke_bpf_prog(tjit, m, fexit->links[i], false))
2811 return -EINVAL;
2812
2813 if (flags & BPF_TRAMP_F_CALL_ORIG) {
2814 im->ip_epilogue = jit->prg_buf + jit->prg;
2815
2816 /*
2817 * __bpf_tramp_exit(im);
2818 */
2819
2820 /* %r1 = __bpf_tramp_exit */
2821 load_imm64(jit, REG_1, (u64)__bpf_tramp_exit);
2822 /* %r2 = im */
2823 load_imm64(jit, REG_2, (u64)im);
2824 /* %r1() */
2825 call_r1(jit);
2826 }
2827
2828 /* lmg %r2,%rN,reg_args_off(%r15) */
2829 if ((flags & BPF_TRAMP_F_RESTORE_REGS) && nr_reg_args)
2830 EMIT6_DISP_LH(0xeb000000, 0x0004, REG_2,
2831 REG_2 + (nr_reg_args - 1), REG_15,
2832 tjit->reg_args_off);
2833 /* lgr %r1,%r8 */
2834 if (!(flags & BPF_TRAMP_F_SKIP_FRAME))
2835 EMIT4(0xb9040000, REG_1, REG_8);
2836 /* lmg %r7,%r8,r7_r8_off(%r15) */
2837 EMIT6_DISP_LH(0xeb000000, 0x0004, REG_7, REG_8, REG_15,
2838 tjit->r7_r8_off);
2839 /* lg %r14,r14_off(%r15) */
2840 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_14, REG_0, REG_15, tjit->r14_off);
2841 /* lg %r2,retval_off(%r15) */
2842 if (flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET))
2843 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_2, REG_0, REG_15,
2844 tjit->retval_off);
2845 /* mvc stack_size+STK_OFF_TCCNT(4,%r15),tccnt_off(%r15) */
2846 _EMIT6(0xd203f000 | (tjit->stack_size + STK_OFF_TCCNT),
2847 0xf000 | tjit->tccnt_off);
2848 /* aghi %r15,stack_size */
2849 EMIT4_IMM(0xa70b0000, REG_15, tjit->stack_size);
2850 /* Emit an expoline for the following indirect jump. */
2851 if (nospec_uses_trampoline())
2852 emit_expoline(jit);
2853 if (flags & BPF_TRAMP_F_SKIP_FRAME)
2854 /* br %r14 */
2855 _EMIT2(0x07fe);
2856 else
2857 /* br %r1 */
2858 _EMIT2(0x07f1);
2859
2860 emit_r1_thunk(jit);
2861
2862 return 0;
2863 }
2864
2865 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
2866 struct bpf_tramp_links *tlinks, void *orig_call)
2867 {
2868 struct bpf_tramp_image im;
2869 struct bpf_tramp_jit tjit;
2870 int ret;
2871
2872 memset(&tjit, 0, sizeof(tjit));
2873
2874 ret = __arch_prepare_bpf_trampoline(&im, &tjit, m, flags,
2875 tlinks, orig_call);
2876
2877 return ret < 0 ? ret : tjit.common.prg;
2878 }
2879
2880 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image,
2881 void *image_end, const struct btf_func_model *m,
2882 u32 flags, struct bpf_tramp_links *tlinks,
2883 void *func_addr)
2884 {
2885 struct bpf_tramp_jit tjit;
2886 int ret;
2887
2888 /* Compute offsets, check whether the code fits. */
2889 memset(&tjit, 0, sizeof(tjit));
2890 ret = __arch_prepare_bpf_trampoline(im, &tjit, m, flags,
2891 tlinks, func_addr);
2892
2893 if (ret < 0)
2894 return ret;
2895 if (tjit.common.prg > (char *)image_end - (char *)image)
2896 /*
2897 * Use the same error code as for exceeding
2898 * BPF_MAX_TRAMP_LINKS.
2899 */
2900 return -E2BIG;
2901
2902 tjit.common.prg = 0;
2903 tjit.common.prg_buf = image;
2904 ret = __arch_prepare_bpf_trampoline(im, &tjit, m, flags,
2905 tlinks, func_addr);
2906
2907 return ret < 0 ? ret : tjit.common.prg;
2908 }
2909
2910 bool bpf_jit_supports_subprog_tailcalls(void)
2911 {
2912 return true;
2913 }
2914
2915 bool bpf_jit_supports_arena(void)
2916 {
2917 return true;
2918 }
2919
2920 bool bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
2921 {
2922 /*
2923 * Currently the verifier uses this function only to check which
2924 * atomic stores to arena are supported, and they all are.
2925 */
2926 return true;
2927 }
2928
2929 bool bpf_jit_supports_exceptions(void)
2930 {
2931 /*
2932 * Exceptions require unwinding support, which is always available,
2933 * because the kernel is always built with backchain.
2934 */
2935 return true;
2936 }
2937
2938 void arch_bpf_stack_walk(bool (*consume_fn)(void *, u64, u64, u64),
2939 void *cookie)
2940 {
2941 unsigned long addr, prev_addr = 0;
2942 struct unwind_state state;
2943
2944 unwind_for_each_frame(&state, NULL, NULL, 0) {
2945 addr = unwind_get_return_address(&state);
2946 if (!addr)
2947 break;
2948 /*
2949 * addr is a return address and state.sp is the value of %r15
2950 * at this address. exception_cb needs %r15 at entry to the
2951 * function containing addr, so take the next state.sp.
2952 *
2953 * There is no bp, and the exception_cb prog does not need one
2954 * to perform a quasi-longjmp. The common code requires a
2955 * non-zero bp, so pass sp there as well.
2956 */
2957 if (prev_addr && !consume_fn(cookie, prev_addr, state.sp,
2958 state.sp))
2959 break;
2960 prev_addr = addr;
2961 }
2962 }
Kernel DRM miscellaneous fixes and cross-tree changes