Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/moduleloader.h>
3#include <linux/workqueue.h>
4#include <linux/netdevice.h>
5#include <linux/filter.h>
6#include <linux/bpf.h>
7#include <linux/cache.h>
8#include <linux/if_vlan.h>
9
10#include <asm/cacheflush.h>
11#include <asm/ptrace.h>
12
13#include "bpf_jit_64.h"
14
15static inline bool is_simm13(unsigned int value)
16{
17 return value + 0x1000 < 0x2000;
18}
19
20static inline bool is_simm10(unsigned int value)
21{
22 return value + 0x200 < 0x400;
23}
24
25static inline bool is_simm5(unsigned int value)
26{
27 return value + 0x10 < 0x20;
28}
29
30static inline bool is_sethi(unsigned int value)
31{
32 return (value & ~0x3fffff) == 0;
33}
34
35static void bpf_flush_icache(void *start_, void *end_)
36{
37 /* Cheetah's I-cache is fully coherent. */
38 if (tlb_type == spitfire) {
39 unsigned long start = (unsigned long) start_;
40 unsigned long end = (unsigned long) end_;
41
42 start &= ~7UL;
43 end = (end + 7UL) & ~7UL;
44 while (start < end) {
45 flushi(start);
46 start += 32;
47 }
48 }
49}
50
51#define SEEN_DATAREF 1 /* might call external helpers */
52#define SEEN_XREG 2 /* ebx is used */
53#define SEEN_MEM 4 /* use mem[] for temporary storage */
54
55#define S13(X) ((X) & 0x1fff)
56#define S5(X) ((X) & 0x1f)
57#define IMMED 0x00002000
58#define RD(X) ((X) << 25)
59#define RS1(X) ((X) << 14)
60#define RS2(X) ((X))
61#define OP(X) ((X) << 30)
62#define OP2(X) ((X) << 22)
63#define OP3(X) ((X) << 19)
64#define COND(X) (((X) & 0xf) << 25)
65#define CBCOND(X) (((X) & 0x1f) << 25)
66#define F1(X) OP(X)
67#define F2(X, Y) (OP(X) | OP2(Y))
68#define F3(X, Y) (OP(X) | OP3(Y))
69#define ASI(X) (((X) & 0xff) << 5)
70
71#define CONDN COND(0x0)
72#define CONDE COND(0x1)
73#define CONDLE COND(0x2)
74#define CONDL COND(0x3)
75#define CONDLEU COND(0x4)
76#define CONDCS COND(0x5)
77#define CONDNEG COND(0x6)
78#define CONDVC COND(0x7)
79#define CONDA COND(0x8)
80#define CONDNE COND(0x9)
81#define CONDG COND(0xa)
82#define CONDGE COND(0xb)
83#define CONDGU COND(0xc)
84#define CONDCC COND(0xd)
85#define CONDPOS COND(0xe)
86#define CONDVS COND(0xf)
87
88#define CONDGEU CONDCC
89#define CONDLU CONDCS
90
91#define WDISP22(X) (((X) >> 2) & 0x3fffff)
92#define WDISP19(X) (((X) >> 2) & 0x7ffff)
93
94/* The 10-bit branch displacement for CBCOND is split into two fields */
95static u32 WDISP10(u32 off)
96{
97 u32 ret = ((off >> 2) & 0xff) << 5;
98
99 ret |= ((off >> (2 + 8)) & 0x03) << 19;
100
101 return ret;
102}
103
104#define CBCONDE CBCOND(0x09)
105#define CBCONDLE CBCOND(0x0a)
106#define CBCONDL CBCOND(0x0b)
107#define CBCONDLEU CBCOND(0x0c)
108#define CBCONDCS CBCOND(0x0d)
109#define CBCONDN CBCOND(0x0e)
110#define CBCONDVS CBCOND(0x0f)
111#define CBCONDNE CBCOND(0x19)
112#define CBCONDG CBCOND(0x1a)
113#define CBCONDGE CBCOND(0x1b)
114#define CBCONDGU CBCOND(0x1c)
115#define CBCONDCC CBCOND(0x1d)
116#define CBCONDPOS CBCOND(0x1e)
117#define CBCONDVC CBCOND(0x1f)
118
119#define CBCONDGEU CBCONDCC
120#define CBCONDLU CBCONDCS
121
122#define ANNUL (1 << 29)
123#define XCC (1 << 21)
124
125#define BRANCH (F2(0, 1) | XCC)
126#define CBCOND_OP (F2(0, 3) | XCC)
127
128#define BA (BRANCH | CONDA)
129#define BG (BRANCH | CONDG)
130#define BL (BRANCH | CONDL)
131#define BLE (BRANCH | CONDLE)
132#define BGU (BRANCH | CONDGU)
133#define BLEU (BRANCH | CONDLEU)
134#define BGE (BRANCH | CONDGE)
135#define BGEU (BRANCH | CONDGEU)
136#define BLU (BRANCH | CONDLU)
137#define BE (BRANCH | CONDE)
138#define BNE (BRANCH | CONDNE)
139
140#define SETHI(K, REG) \
141 (F2(0, 0x4) | RD(REG) | (((K) >> 10) & 0x3fffff))
142#define OR_LO(K, REG) \
143 (F3(2, 0x02) | IMMED | RS1(REG) | ((K) & 0x3ff) | RD(REG))
144
145#define ADD F3(2, 0x00)
146#define AND F3(2, 0x01)
147#define ANDCC F3(2, 0x11)
148#define OR F3(2, 0x02)
149#define XOR F3(2, 0x03)
150#define SUB F3(2, 0x04)
151#define SUBCC F3(2, 0x14)
152#define MUL F3(2, 0x0a)
153#define MULX F3(2, 0x09)
154#define UDIVX F3(2, 0x0d)
155#define DIV F3(2, 0x0e)
156#define SLL F3(2, 0x25)
157#define SLLX (F3(2, 0x25)|(1<<12))
158#define SRA F3(2, 0x27)
159#define SRAX (F3(2, 0x27)|(1<<12))
160#define SRL F3(2, 0x26)
161#define SRLX (F3(2, 0x26)|(1<<12))
162#define JMPL F3(2, 0x38)
163#define SAVE F3(2, 0x3c)
164#define RESTORE F3(2, 0x3d)
165#define CALL F1(1)
166#define BR F2(0, 0x01)
167#define RD_Y F3(2, 0x28)
168#define WR_Y F3(2, 0x30)
169
170#define LD32 F3(3, 0x00)
171#define LD8 F3(3, 0x01)
172#define LD16 F3(3, 0x02)
173#define LD64 F3(3, 0x0b)
174#define LD64A F3(3, 0x1b)
175#define ST8 F3(3, 0x05)
176#define ST16 F3(3, 0x06)
177#define ST32 F3(3, 0x04)
178#define ST64 F3(3, 0x0e)
179
180#define CAS F3(3, 0x3c)
181#define CASX F3(3, 0x3e)
182
183#define LDPTR LD64
184#define BASE_STACKFRAME 176
185
186#define LD32I (LD32 | IMMED)
187#define LD8I (LD8 | IMMED)
188#define LD16I (LD16 | IMMED)
189#define LD64I (LD64 | IMMED)
190#define LDPTRI (LDPTR | IMMED)
191#define ST32I (ST32 | IMMED)
192
193struct jit_ctx {
194 struct bpf_prog *prog;
195 unsigned int *offset;
196 int idx;
197 int epilogue_offset;
198 bool tmp_1_used;
199 bool tmp_2_used;
200 bool tmp_3_used;
201 bool saw_ld_abs_ind;
202 bool saw_frame_pointer;
203 bool saw_call;
204 bool saw_tail_call;
205 u32 *image;
206};
207
208#define TMP_REG_1 (MAX_BPF_JIT_REG + 0)
209#define TMP_REG_2 (MAX_BPF_JIT_REG + 1)
210#define SKB_HLEN_REG (MAX_BPF_JIT_REG + 2)
211#define SKB_DATA_REG (MAX_BPF_JIT_REG + 3)
212#define TMP_REG_3 (MAX_BPF_JIT_REG + 4)
213
214/* Map BPF registers to SPARC registers */
215static const int bpf2sparc[] = {
216 /* return value from in-kernel function, and exit value from eBPF */
217 [BPF_REG_0] = O5,
218
219 /* arguments from eBPF program to in-kernel function */
220 [BPF_REG_1] = O0,
221 [BPF_REG_2] = O1,
222 [BPF_REG_3] = O2,
223 [BPF_REG_4] = O3,
224 [BPF_REG_5] = O4,
225
226 /* callee saved registers that in-kernel function will preserve */
227 [BPF_REG_6] = L0,
228 [BPF_REG_7] = L1,
229 [BPF_REG_8] = L2,
230 [BPF_REG_9] = L3,
231
232 /* read-only frame pointer to access stack */
233 [BPF_REG_FP] = L6,
234
235 [BPF_REG_AX] = G7,
236
237 /* temporary register for internal BPF JIT */
238 [TMP_REG_1] = G1,
239 [TMP_REG_2] = G2,
240 [TMP_REG_3] = G3,
241
242 [SKB_HLEN_REG] = L4,
243 [SKB_DATA_REG] = L5,
244};
245
246static void emit(const u32 insn, struct jit_ctx *ctx)
247{
248 if (ctx->image != NULL)
249 ctx->image[ctx->idx] = insn;
250
251 ctx->idx++;
252}
253
254static void emit_call(u32 *func, struct jit_ctx *ctx)
255{
256 if (ctx->image != NULL) {
257 void *here = &ctx->image[ctx->idx];
258 unsigned int off;
259
260 off = (void *)func - here;
261 ctx->image[ctx->idx] = CALL | ((off >> 2) & 0x3fffffff);
262 }
263 ctx->idx++;
264}
265
266static void emit_nop(struct jit_ctx *ctx)
267{
268 emit(SETHI(0, G0), ctx);
269}
270
271static void emit_reg_move(u32 from, u32 to, struct jit_ctx *ctx)
272{
273 emit(OR | RS1(G0) | RS2(from) | RD(to), ctx);
274}
275
276/* Emit 32-bit constant, zero extended. */
277static void emit_set_const(s32 K, u32 reg, struct jit_ctx *ctx)
278{
279 emit(SETHI(K, reg), ctx);
280 emit(OR_LO(K, reg), ctx);
281}
282
283/* Emit 32-bit constant, sign extended. */
284static void emit_set_const_sext(s32 K, u32 reg, struct jit_ctx *ctx)
285{
286 if (K >= 0) {
287 emit(SETHI(K, reg), ctx);
288 emit(OR_LO(K, reg), ctx);
289 } else {
290 u32 hbits = ~(u32) K;
291 u32 lbits = -0x400 | (u32) K;
292
293 emit(SETHI(hbits, reg), ctx);
294 emit(XOR | IMMED | RS1(reg) | S13(lbits) | RD(reg), ctx);
295 }
296}
297
298static void emit_alu(u32 opcode, u32 src, u32 dst, struct jit_ctx *ctx)
299{
300 emit(opcode | RS1(dst) | RS2(src) | RD(dst), ctx);
301}
302
303static void emit_alu3(u32 opcode, u32 a, u32 b, u32 c, struct jit_ctx *ctx)
304{
305 emit(opcode | RS1(a) | RS2(b) | RD(c), ctx);
306}
307
308static void emit_alu_K(unsigned int opcode, unsigned int dst, unsigned int imm,
309 struct jit_ctx *ctx)
310{
311 bool small_immed = is_simm13(imm);
312 unsigned int insn = opcode;
313
314 insn |= RS1(dst) | RD(dst);
315 if (small_immed) {
316 emit(insn | IMMED | S13(imm), ctx);
317 } else {
318 unsigned int tmp = bpf2sparc[TMP_REG_1];
319
320 ctx->tmp_1_used = true;
321
322 emit_set_const_sext(imm, tmp, ctx);
323 emit(insn | RS2(tmp), ctx);
324 }
325}
326
327static void emit_alu3_K(unsigned int opcode, unsigned int src, unsigned int imm,
328 unsigned int dst, struct jit_ctx *ctx)
329{
330 bool small_immed = is_simm13(imm);
331 unsigned int insn = opcode;
332
333 insn |= RS1(src) | RD(dst);
334 if (small_immed) {
335 emit(insn | IMMED | S13(imm), ctx);
336 } else {
337 unsigned int tmp = bpf2sparc[TMP_REG_1];
338
339 ctx->tmp_1_used = true;
340
341 emit_set_const_sext(imm, tmp, ctx);
342 emit(insn | RS2(tmp), ctx);
343 }
344}
345
346static void emit_loadimm32(s32 K, unsigned int dest, struct jit_ctx *ctx)
347{
348 if (K >= 0 && is_simm13(K)) {
349 /* or %g0, K, DEST */
350 emit(OR | IMMED | RS1(G0) | S13(K) | RD(dest), ctx);
351 } else {
352 emit_set_const(K, dest, ctx);
353 }
354}
355
356static void emit_loadimm(s32 K, unsigned int dest, struct jit_ctx *ctx)
357{
358 if (is_simm13(K)) {
359 /* or %g0, K, DEST */
360 emit(OR | IMMED | RS1(G0) | S13(K) | RD(dest), ctx);
361 } else {
362 emit_set_const(K, dest, ctx);
363 }
364}
365
366static void emit_loadimm_sext(s32 K, unsigned int dest, struct jit_ctx *ctx)
367{
368 if (is_simm13(K)) {
369 /* or %g0, K, DEST */
370 emit(OR | IMMED | RS1(G0) | S13(K) | RD(dest), ctx);
371 } else {
372 emit_set_const_sext(K, dest, ctx);
373 }
374}
375
376static void analyze_64bit_constant(u32 high_bits, u32 low_bits,
377 int *hbsp, int *lbsp, int *abbasp)
378{
379 int lowest_bit_set, highest_bit_set, all_bits_between_are_set;
380 int i;
381
382 lowest_bit_set = highest_bit_set = -1;
383 i = 0;
384 do {
385 if ((lowest_bit_set == -1) && ((low_bits >> i) & 1))
386 lowest_bit_set = i;
387 if ((highest_bit_set == -1) && ((high_bits >> (32 - i - 1)) & 1))
388 highest_bit_set = (64 - i - 1);
389 } while (++i < 32 && (highest_bit_set == -1 ||
390 lowest_bit_set == -1));
391 if (i == 32) {
392 i = 0;
393 do {
394 if (lowest_bit_set == -1 && ((high_bits >> i) & 1))
395 lowest_bit_set = i + 32;
396 if (highest_bit_set == -1 &&
397 ((low_bits >> (32 - i - 1)) & 1))
398 highest_bit_set = 32 - i - 1;
399 } while (++i < 32 && (highest_bit_set == -1 ||
400 lowest_bit_set == -1));
401 }
402
403 all_bits_between_are_set = 1;
404 for (i = lowest_bit_set; i <= highest_bit_set; i++) {
405 if (i < 32) {
406 if ((low_bits & (1 << i)) != 0)
407 continue;
408 } else {
409 if ((high_bits & (1 << (i - 32))) != 0)
410 continue;
411 }
412 all_bits_between_are_set = 0;
413 break;
414 }
415 *hbsp = highest_bit_set;
416 *lbsp = lowest_bit_set;
417 *abbasp = all_bits_between_are_set;
418}
419
420static unsigned long create_simple_focus_bits(unsigned long high_bits,
421 unsigned long low_bits,
422 int lowest_bit_set, int shift)
423{
424 long hi, lo;
425
426 if (lowest_bit_set < 32) {
427 lo = (low_bits >> lowest_bit_set) << shift;
428 hi = ((high_bits << (32 - lowest_bit_set)) << shift);
429 } else {
430 lo = 0;
431 hi = ((high_bits >> (lowest_bit_set - 32)) << shift);
432 }
433 return hi | lo;
434}
435
436static bool const64_is_2insns(unsigned long high_bits,
437 unsigned long low_bits)
438{
439 int highest_bit_set, lowest_bit_set, all_bits_between_are_set;
440
441 if (high_bits == 0 || high_bits == 0xffffffff)
442 return true;
443
444 analyze_64bit_constant(high_bits, low_bits,
445 &highest_bit_set, &lowest_bit_set,
446 &all_bits_between_are_set);
447
448 if ((highest_bit_set == 63 || lowest_bit_set == 0) &&
449 all_bits_between_are_set != 0)
450 return true;
451
452 if (highest_bit_set - lowest_bit_set < 21)
453 return true;
454
455 return false;
456}
457
458static void sparc_emit_set_const64_quick2(unsigned long high_bits,
459 unsigned long low_imm,
460 unsigned int dest,
461 int shift_count, struct jit_ctx *ctx)
462{
463 emit_loadimm32(high_bits, dest, ctx);
464
465 /* Now shift it up into place. */
466 emit_alu_K(SLLX, dest, shift_count, ctx);
467
468 /* If there is a low immediate part piece, finish up by
469 * putting that in as well.
470 */
471 if (low_imm != 0)
472 emit(OR | IMMED | RS1(dest) | S13(low_imm) | RD(dest), ctx);
473}
474
475static void emit_loadimm64(u64 K, unsigned int dest, struct jit_ctx *ctx)
476{
477 int all_bits_between_are_set, lowest_bit_set, highest_bit_set;
478 unsigned int tmp = bpf2sparc[TMP_REG_1];
479 u32 low_bits = (K & 0xffffffff);
480 u32 high_bits = (K >> 32);
481
482 /* These two tests also take care of all of the one
483 * instruction cases.
484 */
485 if (high_bits == 0xffffffff && (low_bits & 0x80000000))
486 return emit_loadimm_sext(K, dest, ctx);
487 if (high_bits == 0x00000000)
488 return emit_loadimm32(K, dest, ctx);
489
490 analyze_64bit_constant(high_bits, low_bits, &highest_bit_set,
491 &lowest_bit_set, &all_bits_between_are_set);
492
493 /* 1) mov -1, %reg
494 * sllx %reg, shift, %reg
495 * 2) mov -1, %reg
496 * srlx %reg, shift, %reg
497 * 3) mov some_small_const, %reg
498 * sllx %reg, shift, %reg
499 */
500 if (((highest_bit_set == 63 || lowest_bit_set == 0) &&
501 all_bits_between_are_set != 0) ||
502 ((highest_bit_set - lowest_bit_set) < 12)) {
503 int shift = lowest_bit_set;
504 long the_const = -1;
505
506 if ((highest_bit_set != 63 && lowest_bit_set != 0) ||
507 all_bits_between_are_set == 0) {
508 the_const =
509 create_simple_focus_bits(high_bits, low_bits,
510 lowest_bit_set, 0);
511 } else if (lowest_bit_set == 0)
512 shift = -(63 - highest_bit_set);
513
514 emit(OR | IMMED | RS1(G0) | S13(the_const) | RD(dest), ctx);
515 if (shift > 0)
516 emit_alu_K(SLLX, dest, shift, ctx);
517 else if (shift < 0)
518 emit_alu_K(SRLX, dest, -shift, ctx);
519
520 return;
521 }
522
523 /* Now a range of 22 or less bits set somewhere.
524 * 1) sethi %hi(focus_bits), %reg
525 * sllx %reg, shift, %reg
526 * 2) sethi %hi(focus_bits), %reg
527 * srlx %reg, shift, %reg
528 */
529 if ((highest_bit_set - lowest_bit_set) < 21) {
530 unsigned long focus_bits =
531 create_simple_focus_bits(high_bits, low_bits,
532 lowest_bit_set, 10);
533
534 emit(SETHI(focus_bits, dest), ctx);
535
536 /* If lowest_bit_set == 10 then a sethi alone could
537 * have done it.
538 */
539 if (lowest_bit_set < 10)
540 emit_alu_K(SRLX, dest, 10 - lowest_bit_set, ctx);
541 else if (lowest_bit_set > 10)
542 emit_alu_K(SLLX, dest, lowest_bit_set - 10, ctx);
543 return;
544 }
545
546 /* Ok, now 3 instruction sequences. */
547 if (low_bits == 0) {
548 emit_loadimm32(high_bits, dest, ctx);
549 emit_alu_K(SLLX, dest, 32, ctx);
550 return;
551 }
552
553 /* We may be able to do something quick
554 * when the constant is negated, so try that.
555 */
556 if (const64_is_2insns((~high_bits) & 0xffffffff,
557 (~low_bits) & 0xfffffc00)) {
558 /* NOTE: The trailing bits get XOR'd so we need the
559 * non-negated bits, not the negated ones.
560 */
561 unsigned long trailing_bits = low_bits & 0x3ff;
562
563 if ((((~high_bits) & 0xffffffff) == 0 &&
564 ((~low_bits) & 0x80000000) == 0) ||
565 (((~high_bits) & 0xffffffff) == 0xffffffff &&
566 ((~low_bits) & 0x80000000) != 0)) {
567 unsigned long fast_int = (~low_bits & 0xffffffff);
568
569 if ((is_sethi(fast_int) &&
570 (~high_bits & 0xffffffff) == 0)) {
571 emit(SETHI(fast_int, dest), ctx);
572 } else if (is_simm13(fast_int)) {
573 emit(OR | IMMED | RS1(G0) | S13(fast_int) | RD(dest), ctx);
574 } else {
575 emit_loadimm64(fast_int, dest, ctx);
576 }
577 } else {
578 u64 n = ((~low_bits) & 0xfffffc00) |
579 (((unsigned long)((~high_bits) & 0xffffffff))<<32);
580 emit_loadimm64(n, dest, ctx);
581 }
582
583 low_bits = -0x400 | trailing_bits;
584
585 emit(XOR | IMMED | RS1(dest) | S13(low_bits) | RD(dest), ctx);
586 return;
587 }
588
589 /* 1) sethi %hi(xxx), %reg
590 * or %reg, %lo(xxx), %reg
591 * sllx %reg, yyy, %reg
592 */
593 if ((highest_bit_set - lowest_bit_set) < 32) {
594 unsigned long focus_bits =
595 create_simple_focus_bits(high_bits, low_bits,
596 lowest_bit_set, 0);
597
598 /* So what we know is that the set bits straddle the
599 * middle of the 64-bit word.
600 */
601 sparc_emit_set_const64_quick2(focus_bits, 0, dest,
602 lowest_bit_set, ctx);
603 return;
604 }
605
606 /* 1) sethi %hi(high_bits), %reg
607 * or %reg, %lo(high_bits), %reg
608 * sllx %reg, 32, %reg
609 * or %reg, low_bits, %reg
610 */
611 if (is_simm13(low_bits) && ((int)low_bits > 0)) {
612 sparc_emit_set_const64_quick2(high_bits, low_bits,
613 dest, 32, ctx);
614 return;
615 }
616
617 /* Oh well, we tried... Do a full 64-bit decomposition. */
618 ctx->tmp_1_used = true;
619
620 emit_loadimm32(high_bits, tmp, ctx);
621 emit_loadimm32(low_bits, dest, ctx);
622 emit_alu_K(SLLX, tmp, 32, ctx);
623 emit(OR | RS1(dest) | RS2(tmp) | RD(dest), ctx);
624}
625
626static void emit_branch(unsigned int br_opc, unsigned int from_idx, unsigned int to_idx,
627 struct jit_ctx *ctx)
628{
629 unsigned int off = to_idx - from_idx;
630
631 if (br_opc & XCC)
632 emit(br_opc | WDISP19(off << 2), ctx);
633 else
634 emit(br_opc | WDISP22(off << 2), ctx);
635}
636
637static void emit_cbcond(unsigned int cb_opc, unsigned int from_idx, unsigned int to_idx,
638 const u8 dst, const u8 src, struct jit_ctx *ctx)
639{
640 unsigned int off = to_idx - from_idx;
641
642 emit(cb_opc | WDISP10(off << 2) | RS1(dst) | RS2(src), ctx);
643}
644
645static void emit_cbcondi(unsigned int cb_opc, unsigned int from_idx, unsigned int to_idx,
646 const u8 dst, s32 imm, struct jit_ctx *ctx)
647{
648 unsigned int off = to_idx - from_idx;
649
650 emit(cb_opc | IMMED | WDISP10(off << 2) | RS1(dst) | S5(imm), ctx);
651}
652
653#define emit_read_y(REG, CTX) emit(RD_Y | RD(REG), CTX)
654#define emit_write_y(REG, CTX) emit(WR_Y | IMMED | RS1(REG) | S13(0), CTX)
655
656#define emit_cmp(R1, R2, CTX) \
657 emit(SUBCC | RS1(R1) | RS2(R2) | RD(G0), CTX)
658
659#define emit_cmpi(R1, IMM, CTX) \
660 emit(SUBCC | IMMED | RS1(R1) | S13(IMM) | RD(G0), CTX)
661
662#define emit_btst(R1, R2, CTX) \
663 emit(ANDCC | RS1(R1) | RS2(R2) | RD(G0), CTX)
664
665#define emit_btsti(R1, IMM, CTX) \
666 emit(ANDCC | IMMED | RS1(R1) | S13(IMM) | RD(G0), CTX)
667
668static int emit_compare_and_branch(const u8 code, const u8 dst, u8 src,
669 const s32 imm, bool is_imm, int branch_dst,
670 struct jit_ctx *ctx)
671{
672 bool use_cbcond = (sparc64_elf_hwcap & AV_SPARC_CBCOND) != 0;
673 const u8 tmp = bpf2sparc[TMP_REG_1];
674
675 branch_dst = ctx->offset[branch_dst];
676
677 if (!is_simm10(branch_dst - ctx->idx) ||
678 BPF_OP(code) == BPF_JSET)
679 use_cbcond = false;
680
681 if (is_imm) {
682 bool fits = true;
683
684 if (use_cbcond) {
685 if (!is_simm5(imm))
686 fits = false;
687 } else if (!is_simm13(imm)) {
688 fits = false;
689 }
690 if (!fits) {
691 ctx->tmp_1_used = true;
692 emit_loadimm_sext(imm, tmp, ctx);
693 src = tmp;
694 is_imm = false;
695 }
696 }
697
698 if (!use_cbcond) {
699 u32 br_opcode;
700
701 if (BPF_OP(code) == BPF_JSET) {
702 if (is_imm)
703 emit_btsti(dst, imm, ctx);
704 else
705 emit_btst(dst, src, ctx);
706 } else {
707 if (is_imm)
708 emit_cmpi(dst, imm, ctx);
709 else
710 emit_cmp(dst, src, ctx);
711 }
712 switch (BPF_OP(code)) {
713 case BPF_JEQ:
714 br_opcode = BE;
715 break;
716 case BPF_JGT:
717 br_opcode = BGU;
718 break;
719 case BPF_JLT:
720 br_opcode = BLU;
721 break;
722 case BPF_JGE:
723 br_opcode = BGEU;
724 break;
725 case BPF_JLE:
726 br_opcode = BLEU;
727 break;
728 case BPF_JSET:
729 case BPF_JNE:
730 br_opcode = BNE;
731 break;
732 case BPF_JSGT:
733 br_opcode = BG;
734 break;
735 case BPF_JSLT:
736 br_opcode = BL;
737 break;
738 case BPF_JSGE:
739 br_opcode = BGE;
740 break;
741 case BPF_JSLE:
742 br_opcode = BLE;
743 break;
744 default:
745 /* Make sure we dont leak kernel information to the
746 * user.
747 */
748 return -EFAULT;
749 }
750 emit_branch(br_opcode, ctx->idx, branch_dst, ctx);
751 emit_nop(ctx);
752 } else {
753 u32 cbcond_opcode;
754
755 switch (BPF_OP(code)) {
756 case BPF_JEQ:
757 cbcond_opcode = CBCONDE;
758 break;
759 case BPF_JGT:
760 cbcond_opcode = CBCONDGU;
761 break;
762 case BPF_JLT:
763 cbcond_opcode = CBCONDLU;
764 break;
765 case BPF_JGE:
766 cbcond_opcode = CBCONDGEU;
767 break;
768 case BPF_JLE:
769 cbcond_opcode = CBCONDLEU;
770 break;
771 case BPF_JNE:
772 cbcond_opcode = CBCONDNE;
773 break;
774 case BPF_JSGT:
775 cbcond_opcode = CBCONDG;
776 break;
777 case BPF_JSLT:
778 cbcond_opcode = CBCONDL;
779 break;
780 case BPF_JSGE:
781 cbcond_opcode = CBCONDGE;
782 break;
783 case BPF_JSLE:
784 cbcond_opcode = CBCONDLE;
785 break;
786 default:
787 /* Make sure we dont leak kernel information to the
788 * user.
789 */
790 return -EFAULT;
791 }
792 cbcond_opcode |= CBCOND_OP;
793 if (is_imm)
794 emit_cbcondi(cbcond_opcode, ctx->idx, branch_dst,
795 dst, imm, ctx);
796 else
797 emit_cbcond(cbcond_opcode, ctx->idx, branch_dst,
798 dst, src, ctx);
799 }
800 return 0;
801}
802
803static void load_skb_regs(struct jit_ctx *ctx, u8 r_skb)
804{
805 const u8 r_headlen = bpf2sparc[SKB_HLEN_REG];
806 const u8 r_data = bpf2sparc[SKB_DATA_REG];
807 const u8 r_tmp = bpf2sparc[TMP_REG_1];
808 unsigned int off;
809
810 off = offsetof(struct sk_buff, len);
811 emit(LD32I | RS1(r_skb) | S13(off) | RD(r_headlen), ctx);
812
813 off = offsetof(struct sk_buff, data_len);
814 emit(LD32I | RS1(r_skb) | S13(off) | RD(r_tmp), ctx);
815
816 emit(SUB | RS1(r_headlen) | RS2(r_tmp) | RD(r_headlen), ctx);
817
818 off = offsetof(struct sk_buff, data);
819 emit(LDPTRI | RS1(r_skb) | S13(off) | RD(r_data), ctx);
820}
821
822/* Just skip the save instruction and the ctx register move. */
823#define BPF_TAILCALL_PROLOGUE_SKIP 16
824#define BPF_TAILCALL_CNT_SP_OFF (STACK_BIAS + 128)
825
826static void build_prologue(struct jit_ctx *ctx)
827{
828 s32 stack_needed = BASE_STACKFRAME;
829
830 if (ctx->saw_frame_pointer || ctx->saw_tail_call) {
831 struct bpf_prog *prog = ctx->prog;
832 u32 stack_depth;
833
834 stack_depth = prog->aux->stack_depth;
835 stack_needed += round_up(stack_depth, 16);
836 }
837
838 if (ctx->saw_tail_call)
839 stack_needed += 8;
840
841 /* save %sp, -176, %sp */
842 emit(SAVE | IMMED | RS1(SP) | S13(-stack_needed) | RD(SP), ctx);
843
844 /* tail_call_cnt = 0 */
845 if (ctx->saw_tail_call) {
846 u32 off = BPF_TAILCALL_CNT_SP_OFF;
847
848 emit(ST32 | IMMED | RS1(SP) | S13(off) | RD(G0), ctx);
849 } else {
850 emit_nop(ctx);
851 }
852 if (ctx->saw_frame_pointer) {
853 const u8 vfp = bpf2sparc[BPF_REG_FP];
854
855 emit(ADD | IMMED | RS1(FP) | S13(STACK_BIAS) | RD(vfp), ctx);
856 }
857
858 emit_reg_move(I0, O0, ctx);
859 /* If you add anything here, adjust BPF_TAILCALL_PROLOGUE_SKIP above. */
860
861 if (ctx->saw_ld_abs_ind)
862 load_skb_regs(ctx, bpf2sparc[BPF_REG_1]);
863}
864
865static void build_epilogue(struct jit_ctx *ctx)
866{
867 ctx->epilogue_offset = ctx->idx;
868
869 /* ret (jmpl %i7 + 8, %g0) */
870 emit(JMPL | IMMED | RS1(I7) | S13(8) | RD(G0), ctx);
871
872 /* restore %i5, %g0, %o0 */
873 emit(RESTORE | RS1(bpf2sparc[BPF_REG_0]) | RS2(G0) | RD(O0), ctx);
874}
875
876static void emit_tail_call(struct jit_ctx *ctx)
877{
878 const u8 bpf_array = bpf2sparc[BPF_REG_2];
879 const u8 bpf_index = bpf2sparc[BPF_REG_3];
880 const u8 tmp = bpf2sparc[TMP_REG_1];
881 u32 off;
882
883 ctx->saw_tail_call = true;
884
885 off = offsetof(struct bpf_array, map.max_entries);
886 emit(LD32 | IMMED | RS1(bpf_array) | S13(off) | RD(tmp), ctx);
887 emit_cmp(bpf_index, tmp, ctx);
888#define OFFSET1 17
889 emit_branch(BGEU, ctx->idx, ctx->idx + OFFSET1, ctx);
890 emit_nop(ctx);
891
892 off = BPF_TAILCALL_CNT_SP_OFF;
893 emit(LD32 | IMMED | RS1(SP) | S13(off) | RD(tmp), ctx);
894 emit_cmpi(tmp, MAX_TAIL_CALL_CNT, ctx);
895#define OFFSET2 13
896 emit_branch(BGU, ctx->idx, ctx->idx + OFFSET2, ctx);
897 emit_nop(ctx);
898
899 emit_alu_K(ADD, tmp, 1, ctx);
900 off = BPF_TAILCALL_CNT_SP_OFF;
901 emit(ST32 | IMMED | RS1(SP) | S13(off) | RD(tmp), ctx);
902
903 emit_alu3_K(SLL, bpf_index, 3, tmp, ctx);
904 emit_alu(ADD, bpf_array, tmp, ctx);
905 off = offsetof(struct bpf_array, ptrs);
906 emit(LD64 | IMMED | RS1(tmp) | S13(off) | RD(tmp), ctx);
907
908 emit_cmpi(tmp, 0, ctx);
909#define OFFSET3 5
910 emit_branch(BE, ctx->idx, ctx->idx + OFFSET3, ctx);
911 emit_nop(ctx);
912
913 off = offsetof(struct bpf_prog, bpf_func);
914 emit(LD64 | IMMED | RS1(tmp) | S13(off) | RD(tmp), ctx);
915
916 off = BPF_TAILCALL_PROLOGUE_SKIP;
917 emit(JMPL | IMMED | RS1(tmp) | S13(off) | RD(G0), ctx);
918 emit_nop(ctx);
919}
920
921static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
922{
923 const u8 code = insn->code;
924 const u8 dst = bpf2sparc[insn->dst_reg];
925 const u8 src = bpf2sparc[insn->src_reg];
926 const int i = insn - ctx->prog->insnsi;
927 const s16 off = insn->off;
928 const s32 imm = insn->imm;
929 u32 *func;
930
931 if (insn->src_reg == BPF_REG_FP)
932 ctx->saw_frame_pointer = true;
933
934 switch (code) {
935 /* dst = src */
936 case BPF_ALU | BPF_MOV | BPF_X:
937 emit_alu3_K(SRL, src, 0, dst, ctx);
938 break;
939 case BPF_ALU64 | BPF_MOV | BPF_X:
940 emit_reg_move(src, dst, ctx);
941 break;
942 /* dst = dst OP src */
943 case BPF_ALU | BPF_ADD | BPF_X:
944 case BPF_ALU64 | BPF_ADD | BPF_X:
945 emit_alu(ADD, src, dst, ctx);
946 goto do_alu32_trunc;
947 case BPF_ALU | BPF_SUB | BPF_X:
948 case BPF_ALU64 | BPF_SUB | BPF_X:
949 emit_alu(SUB, src, dst, ctx);
950 goto do_alu32_trunc;
951 case BPF_ALU | BPF_AND | BPF_X:
952 case BPF_ALU64 | BPF_AND | BPF_X:
953 emit_alu(AND, src, dst, ctx);
954 goto do_alu32_trunc;
955 case BPF_ALU | BPF_OR | BPF_X:
956 case BPF_ALU64 | BPF_OR | BPF_X:
957 emit_alu(OR, src, dst, ctx);
958 goto do_alu32_trunc;
959 case BPF_ALU | BPF_XOR | BPF_X:
960 case BPF_ALU64 | BPF_XOR | BPF_X:
961 emit_alu(XOR, src, dst, ctx);
962 goto do_alu32_trunc;
963 case BPF_ALU | BPF_MUL | BPF_X:
964 emit_alu(MUL, src, dst, ctx);
965 goto do_alu32_trunc;
966 case BPF_ALU64 | BPF_MUL | BPF_X:
967 emit_alu(MULX, src, dst, ctx);
968 break;
969 case BPF_ALU | BPF_DIV | BPF_X:
970 emit_write_y(G0, ctx);
971 emit_alu(DIV, src, dst, ctx);
972 break;
973 case BPF_ALU64 | BPF_DIV | BPF_X:
974 emit_alu(UDIVX, src, dst, ctx);
975 break;
976 case BPF_ALU | BPF_MOD | BPF_X: {
977 const u8 tmp = bpf2sparc[TMP_REG_1];
978
979 ctx->tmp_1_used = true;
980
981 emit_write_y(G0, ctx);
982 emit_alu3(DIV, dst, src, tmp, ctx);
983 emit_alu3(MULX, tmp, src, tmp, ctx);
984 emit_alu3(SUB, dst, tmp, dst, ctx);
985 goto do_alu32_trunc;
986 }
987 case BPF_ALU64 | BPF_MOD | BPF_X: {
988 const u8 tmp = bpf2sparc[TMP_REG_1];
989
990 ctx->tmp_1_used = true;
991
992 emit_alu3(UDIVX, dst, src, tmp, ctx);
993 emit_alu3(MULX, tmp, src, tmp, ctx);
994 emit_alu3(SUB, dst, tmp, dst, ctx);
995 break;
996 }
997 case BPF_ALU | BPF_LSH | BPF_X:
998 emit_alu(SLL, src, dst, ctx);
999 goto do_alu32_trunc;
1000 case BPF_ALU64 | BPF_LSH | BPF_X:
1001 emit_alu(SLLX, src, dst, ctx);
1002 break;
1003 case BPF_ALU | BPF_RSH | BPF_X:
1004 emit_alu(SRL, src, dst, ctx);
1005 break;
1006 case BPF_ALU64 | BPF_RSH | BPF_X:
1007 emit_alu(SRLX, src, dst, ctx);
1008 break;
1009 case BPF_ALU | BPF_ARSH | BPF_X:
1010 emit_alu(SRA, src, dst, ctx);
1011 goto do_alu32_trunc;
1012 case BPF_ALU64 | BPF_ARSH | BPF_X:
1013 emit_alu(SRAX, src, dst, ctx);
1014 break;
1015
1016 /* dst = -dst */
1017 case BPF_ALU | BPF_NEG:
1018 case BPF_ALU64 | BPF_NEG:
1019 emit(SUB | RS1(0) | RS2(dst) | RD(dst), ctx);
1020 goto do_alu32_trunc;
1021
1022 case BPF_ALU | BPF_END | BPF_FROM_BE:
1023 switch (imm) {
1024 case 16:
1025 emit_alu_K(SLL, dst, 16, ctx);
1026 emit_alu_K(SRL, dst, 16, ctx);
1027 break;
1028 case 32:
1029 emit_alu_K(SRL, dst, 0, ctx);
1030 break;
1031 case 64:
1032 /* nop */
1033 break;
1034
1035 }
1036 break;
1037
1038 /* dst = BSWAP##imm(dst) */
1039 case BPF_ALU | BPF_END | BPF_FROM_LE: {
1040 const u8 tmp = bpf2sparc[TMP_REG_1];
1041 const u8 tmp2 = bpf2sparc[TMP_REG_2];
1042
1043 ctx->tmp_1_used = true;
1044 switch (imm) {
1045 case 16:
1046 emit_alu3_K(AND, dst, 0xff, tmp, ctx);
1047 emit_alu3_K(SRL, dst, 8, dst, ctx);
1048 emit_alu3_K(AND, dst, 0xff, dst, ctx);
1049 emit_alu3_K(SLL, tmp, 8, tmp, ctx);
1050 emit_alu(OR, tmp, dst, ctx);
1051 break;
1052
1053 case 32:
1054 ctx->tmp_2_used = true;
1055 emit_alu3_K(SRL, dst, 24, tmp, ctx); /* tmp = dst >> 24 */
1056 emit_alu3_K(SRL, dst, 16, tmp2, ctx); /* tmp2 = dst >> 16 */
1057 emit_alu3_K(AND, tmp2, 0xff, tmp2, ctx);/* tmp2 = tmp2 & 0xff */
1058 emit_alu3_K(SLL, tmp2, 8, tmp2, ctx); /* tmp2 = tmp2 << 8 */
1059 emit_alu(OR, tmp2, tmp, ctx); /* tmp = tmp | tmp2 */
1060 emit_alu3_K(SRL, dst, 8, tmp2, ctx); /* tmp2 = dst >> 8 */
1061 emit_alu3_K(AND, tmp2, 0xff, tmp2, ctx);/* tmp2 = tmp2 & 0xff */
1062 emit_alu3_K(SLL, tmp2, 16, tmp2, ctx); /* tmp2 = tmp2 << 16 */
1063 emit_alu(OR, tmp2, tmp, ctx); /* tmp = tmp | tmp2 */
1064 emit_alu3_K(AND, dst, 0xff, dst, ctx); /* dst = dst & 0xff */
1065 emit_alu3_K(SLL, dst, 24, dst, ctx); /* dst = dst << 24 */
1066 emit_alu(OR, tmp, dst, ctx); /* dst = dst | tmp */
1067 break;
1068
1069 case 64:
1070 emit_alu3_K(ADD, SP, STACK_BIAS + 128, tmp, ctx);
1071 emit(ST64 | RS1(tmp) | RS2(G0) | RD(dst), ctx);
1072 emit(LD64A | ASI(ASI_PL) | RS1(tmp) | RS2(G0) | RD(dst), ctx);
1073 break;
1074 }
1075 break;
1076 }
1077 /* dst = imm */
1078 case BPF_ALU | BPF_MOV | BPF_K:
1079 emit_loadimm32(imm, dst, ctx);
1080 break;
1081 case BPF_ALU64 | BPF_MOV | BPF_K:
1082 emit_loadimm_sext(imm, dst, ctx);
1083 break;
1084 /* dst = dst OP imm */
1085 case BPF_ALU | BPF_ADD | BPF_K:
1086 case BPF_ALU64 | BPF_ADD | BPF_K:
1087 emit_alu_K(ADD, dst, imm, ctx);
1088 goto do_alu32_trunc;
1089 case BPF_ALU | BPF_SUB | BPF_K:
1090 case BPF_ALU64 | BPF_SUB | BPF_K:
1091 emit_alu_K(SUB, dst, imm, ctx);
1092 goto do_alu32_trunc;
1093 case BPF_ALU | BPF_AND | BPF_K:
1094 case BPF_ALU64 | BPF_AND | BPF_K:
1095 emit_alu_K(AND, dst, imm, ctx);
1096 goto do_alu32_trunc;
1097 case BPF_ALU | BPF_OR | BPF_K:
1098 case BPF_ALU64 | BPF_OR | BPF_K:
1099 emit_alu_K(OR, dst, imm, ctx);
1100 goto do_alu32_trunc;
1101 case BPF_ALU | BPF_XOR | BPF_K:
1102 case BPF_ALU64 | BPF_XOR | BPF_K:
1103 emit_alu_K(XOR, dst, imm, ctx);
1104 goto do_alu32_trunc;
1105 case BPF_ALU | BPF_MUL | BPF_K:
1106 emit_alu_K(MUL, dst, imm, ctx);
1107 goto do_alu32_trunc;
1108 case BPF_ALU64 | BPF_MUL | BPF_K:
1109 emit_alu_K(MULX, dst, imm, ctx);
1110 break;
1111 case BPF_ALU | BPF_DIV | BPF_K:
1112 if (imm == 0)
1113 return -EINVAL;
1114
1115 emit_write_y(G0, ctx);
1116 emit_alu_K(DIV, dst, imm, ctx);
1117 goto do_alu32_trunc;
1118 case BPF_ALU64 | BPF_DIV | BPF_K:
1119 if (imm == 0)
1120 return -EINVAL;
1121
1122 emit_alu_K(UDIVX, dst, imm, ctx);
1123 break;
1124 case BPF_ALU64 | BPF_MOD | BPF_K:
1125 case BPF_ALU | BPF_MOD | BPF_K: {
1126 const u8 tmp = bpf2sparc[TMP_REG_2];
1127 unsigned int div;
1128
1129 if (imm == 0)
1130 return -EINVAL;
1131
1132 div = (BPF_CLASS(code) == BPF_ALU64) ? UDIVX : DIV;
1133
1134 ctx->tmp_2_used = true;
1135
1136 if (BPF_CLASS(code) != BPF_ALU64)
1137 emit_write_y(G0, ctx);
1138 if (is_simm13(imm)) {
1139 emit(div | IMMED | RS1(dst) | S13(imm) | RD(tmp), ctx);
1140 emit(MULX | IMMED | RS1(tmp) | S13(imm) | RD(tmp), ctx);
1141 emit(SUB | RS1(dst) | RS2(tmp) | RD(dst), ctx);
1142 } else {
1143 const u8 tmp1 = bpf2sparc[TMP_REG_1];
1144
1145 ctx->tmp_1_used = true;
1146
1147 emit_set_const_sext(imm, tmp1, ctx);
1148 emit(div | RS1(dst) | RS2(tmp1) | RD(tmp), ctx);
1149 emit(MULX | RS1(tmp) | RS2(tmp1) | RD(tmp), ctx);
1150 emit(SUB | RS1(dst) | RS2(tmp) | RD(dst), ctx);
1151 }
1152 goto do_alu32_trunc;
1153 }
1154 case BPF_ALU | BPF_LSH | BPF_K:
1155 emit_alu_K(SLL, dst, imm, ctx);
1156 goto do_alu32_trunc;
1157 case BPF_ALU64 | BPF_LSH | BPF_K:
1158 emit_alu_K(SLLX, dst, imm, ctx);
1159 break;
1160 case BPF_ALU | BPF_RSH | BPF_K:
1161 emit_alu_K(SRL, dst, imm, ctx);
1162 break;
1163 case BPF_ALU64 | BPF_RSH | BPF_K:
1164 emit_alu_K(SRLX, dst, imm, ctx);
1165 break;
1166 case BPF_ALU | BPF_ARSH | BPF_K:
1167 emit_alu_K(SRA, dst, imm, ctx);
1168 goto do_alu32_trunc;
1169 case BPF_ALU64 | BPF_ARSH | BPF_K:
1170 emit_alu_K(SRAX, dst, imm, ctx);
1171 break;
1172
1173 do_alu32_trunc:
1174 if (BPF_CLASS(code) == BPF_ALU)
1175 emit_alu_K(SRL, dst, 0, ctx);
1176 break;
1177
1178 /* JUMP off */
1179 case BPF_JMP | BPF_JA:
1180 emit_branch(BA, ctx->idx, ctx->offset[i + off], ctx);
1181 emit_nop(ctx);
1182 break;
1183 /* IF (dst COND src) JUMP off */
1184 case BPF_JMP | BPF_JEQ | BPF_X:
1185 case BPF_JMP | BPF_JGT | BPF_X:
1186 case BPF_JMP | BPF_JLT | BPF_X:
1187 case BPF_JMP | BPF_JGE | BPF_X:
1188 case BPF_JMP | BPF_JLE | BPF_X:
1189 case BPF_JMP | BPF_JNE | BPF_X:
1190 case BPF_JMP | BPF_JSGT | BPF_X:
1191 case BPF_JMP | BPF_JSLT | BPF_X:
1192 case BPF_JMP | BPF_JSGE | BPF_X:
1193 case BPF_JMP | BPF_JSLE | BPF_X:
1194 case BPF_JMP | BPF_JSET | BPF_X: {
1195 int err;
1196
1197 err = emit_compare_and_branch(code, dst, src, 0, false, i + off, ctx);
1198 if (err)
1199 return err;
1200 break;
1201 }
1202 /* IF (dst COND imm) JUMP off */
1203 case BPF_JMP | BPF_JEQ | BPF_K:
1204 case BPF_JMP | BPF_JGT | BPF_K:
1205 case BPF_JMP | BPF_JLT | BPF_K:
1206 case BPF_JMP | BPF_JGE | BPF_K:
1207 case BPF_JMP | BPF_JLE | BPF_K:
1208 case BPF_JMP | BPF_JNE | BPF_K:
1209 case BPF_JMP | BPF_JSGT | BPF_K:
1210 case BPF_JMP | BPF_JSLT | BPF_K:
1211 case BPF_JMP | BPF_JSGE | BPF_K:
1212 case BPF_JMP | BPF_JSLE | BPF_K:
1213 case BPF_JMP | BPF_JSET | BPF_K: {
1214 int err;
1215
1216 err = emit_compare_and_branch(code, dst, 0, imm, true, i + off, ctx);
1217 if (err)
1218 return err;
1219 break;
1220 }
1221
1222 /* function call */
1223 case BPF_JMP | BPF_CALL:
1224 {
1225 u8 *func = ((u8 *)__bpf_call_base) + imm;
1226
1227 ctx->saw_call = true;
1228 if (ctx->saw_ld_abs_ind && bpf_helper_changes_pkt_data(func))
1229 emit_reg_move(bpf2sparc[BPF_REG_1], L7, ctx);
1230
1231 emit_call((u32 *)func, ctx);
1232 emit_nop(ctx);
1233
1234 emit_reg_move(O0, bpf2sparc[BPF_REG_0], ctx);
1235
1236 if (ctx->saw_ld_abs_ind && bpf_helper_changes_pkt_data(func))
1237 load_skb_regs(ctx, L7);
1238 break;
1239 }
1240
1241 /* tail call */
1242 case BPF_JMP | BPF_TAIL_CALL:
1243 emit_tail_call(ctx);
1244 break;
1245
1246 /* function return */
1247 case BPF_JMP | BPF_EXIT:
1248 /* Optimization: when last instruction is EXIT,
1249 simply fallthrough to epilogue. */
1250 if (i == ctx->prog->len - 1)
1251 break;
1252 emit_branch(BA, ctx->idx, ctx->epilogue_offset, ctx);
1253 emit_nop(ctx);
1254 break;
1255
1256 /* dst = imm64 */
1257 case BPF_LD | BPF_IMM | BPF_DW:
1258 {
1259 const struct bpf_insn insn1 = insn[1];
1260 u64 imm64;
1261
1262 imm64 = (u64)insn1.imm << 32 | (u32)imm;
1263 emit_loadimm64(imm64, dst, ctx);
1264
1265 return 1;
1266 }
1267
1268 /* LDX: dst = *(size *)(src + off) */
1269 case BPF_LDX | BPF_MEM | BPF_W:
1270 case BPF_LDX | BPF_MEM | BPF_H:
1271 case BPF_LDX | BPF_MEM | BPF_B:
1272 case BPF_LDX | BPF_MEM | BPF_DW: {
1273 const u8 tmp = bpf2sparc[TMP_REG_1];
1274 u32 opcode = 0, rs2;
1275
1276 ctx->tmp_1_used = true;
1277 switch (BPF_SIZE(code)) {
1278 case BPF_W:
1279 opcode = LD32;
1280 break;
1281 case BPF_H:
1282 opcode = LD16;
1283 break;
1284 case BPF_B:
1285 opcode = LD8;
1286 break;
1287 case BPF_DW:
1288 opcode = LD64;
1289 break;
1290 }
1291
1292 if (is_simm13(off)) {
1293 opcode |= IMMED;
1294 rs2 = S13(off);
1295 } else {
1296 emit_loadimm(off, tmp, ctx);
1297 rs2 = RS2(tmp);
1298 }
1299 emit(opcode | RS1(src) | rs2 | RD(dst), ctx);
1300 break;
1301 }
1302 /* ST: *(size *)(dst + off) = imm */
1303 case BPF_ST | BPF_MEM | BPF_W:
1304 case BPF_ST | BPF_MEM | BPF_H:
1305 case BPF_ST | BPF_MEM | BPF_B:
1306 case BPF_ST | BPF_MEM | BPF_DW: {
1307 const u8 tmp = bpf2sparc[TMP_REG_1];
1308 const u8 tmp2 = bpf2sparc[TMP_REG_2];
1309 u32 opcode = 0, rs2;
1310
1311 ctx->tmp_2_used = true;
1312 emit_loadimm(imm, tmp2, ctx);
1313
1314 switch (BPF_SIZE(code)) {
1315 case BPF_W:
1316 opcode = ST32;
1317 break;
1318 case BPF_H:
1319 opcode = ST16;
1320 break;
1321 case BPF_B:
1322 opcode = ST8;
1323 break;
1324 case BPF_DW:
1325 opcode = ST64;
1326 break;
1327 }
1328
1329 if (is_simm13(off)) {
1330 opcode |= IMMED;
1331 rs2 = S13(off);
1332 } else {
1333 ctx->tmp_1_used = true;
1334 emit_loadimm(off, tmp, ctx);
1335 rs2 = RS2(tmp);
1336 }
1337 emit(opcode | RS1(dst) | rs2 | RD(tmp2), ctx);
1338 break;
1339 }
1340
1341 /* STX: *(size *)(dst + off) = src */
1342 case BPF_STX | BPF_MEM | BPF_W:
1343 case BPF_STX | BPF_MEM | BPF_H:
1344 case BPF_STX | BPF_MEM | BPF_B:
1345 case BPF_STX | BPF_MEM | BPF_DW: {
1346 const u8 tmp = bpf2sparc[TMP_REG_1];
1347 u32 opcode = 0, rs2;
1348
1349 switch (BPF_SIZE(code)) {
1350 case BPF_W:
1351 opcode = ST32;
1352 break;
1353 case BPF_H:
1354 opcode = ST16;
1355 break;
1356 case BPF_B:
1357 opcode = ST8;
1358 break;
1359 case BPF_DW:
1360 opcode = ST64;
1361 break;
1362 }
1363 if (is_simm13(off)) {
1364 opcode |= IMMED;
1365 rs2 = S13(off);
1366 } else {
1367 ctx->tmp_1_used = true;
1368 emit_loadimm(off, tmp, ctx);
1369 rs2 = RS2(tmp);
1370 }
1371 emit(opcode | RS1(dst) | rs2 | RD(src), ctx);
1372 break;
1373 }
1374
1375 /* STX XADD: lock *(u32 *)(dst + off) += src */
1376 case BPF_STX | BPF_XADD | BPF_W: {
1377 const u8 tmp = bpf2sparc[TMP_REG_1];
1378 const u8 tmp2 = bpf2sparc[TMP_REG_2];
1379 const u8 tmp3 = bpf2sparc[TMP_REG_3];
1380
1381 ctx->tmp_1_used = true;
1382 ctx->tmp_2_used = true;
1383 ctx->tmp_3_used = true;
1384 emit_loadimm(off, tmp, ctx);
1385 emit_alu3(ADD, dst, tmp, tmp, ctx);
1386
1387 emit(LD32 | RS1(tmp) | RS2(G0) | RD(tmp2), ctx);
1388 emit_alu3(ADD, tmp2, src, tmp3, ctx);
1389 emit(CAS | ASI(ASI_P) | RS1(tmp) | RS2(tmp2) | RD(tmp3), ctx);
1390 emit_cmp(tmp2, tmp3, ctx);
1391 emit_branch(BNE, 4, 0, ctx);
1392 emit_nop(ctx);
1393 break;
1394 }
1395 /* STX XADD: lock *(u64 *)(dst + off) += src */
1396 case BPF_STX | BPF_XADD | BPF_DW: {
1397 const u8 tmp = bpf2sparc[TMP_REG_1];
1398 const u8 tmp2 = bpf2sparc[TMP_REG_2];
1399 const u8 tmp3 = bpf2sparc[TMP_REG_3];
1400
1401 ctx->tmp_1_used = true;
1402 ctx->tmp_2_used = true;
1403 ctx->tmp_3_used = true;
1404 emit_loadimm(off, tmp, ctx);
1405 emit_alu3(ADD, dst, tmp, tmp, ctx);
1406
1407 emit(LD64 | RS1(tmp) | RS2(G0) | RD(tmp2), ctx);
1408 emit_alu3(ADD, tmp2, src, tmp3, ctx);
1409 emit(CASX | ASI(ASI_P) | RS1(tmp) | RS2(tmp2) | RD(tmp3), ctx);
1410 emit_cmp(tmp2, tmp3, ctx);
1411 emit_branch(BNE, 4, 0, ctx);
1412 emit_nop(ctx);
1413 break;
1414 }
1415#define CHOOSE_LOAD_FUNC(K, func) \
1416 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
1417
1418 /* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + imm)) */
1419 case BPF_LD | BPF_ABS | BPF_W:
1420 func = CHOOSE_LOAD_FUNC(imm, bpf_jit_load_word);
1421 goto common_load;
1422 case BPF_LD | BPF_ABS | BPF_H:
1423 func = CHOOSE_LOAD_FUNC(imm, bpf_jit_load_half);
1424 goto common_load;
1425 case BPF_LD | BPF_ABS | BPF_B:
1426 func = CHOOSE_LOAD_FUNC(imm, bpf_jit_load_byte);
1427 goto common_load;
1428 /* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + src + imm)) */
1429 case BPF_LD | BPF_IND | BPF_W:
1430 func = bpf_jit_load_word;
1431 goto common_load;
1432 case BPF_LD | BPF_IND | BPF_H:
1433 func = bpf_jit_load_half;
1434 goto common_load;
1435
1436 case BPF_LD | BPF_IND | BPF_B:
1437 func = bpf_jit_load_byte;
1438 common_load:
1439 ctx->saw_ld_abs_ind = true;
1440
1441 emit_reg_move(bpf2sparc[BPF_REG_6], O0, ctx);
1442 emit_loadimm(imm, O1, ctx);
1443
1444 if (BPF_MODE(code) == BPF_IND)
1445 emit_alu(ADD, src, O1, ctx);
1446
1447 emit_call(func, ctx);
1448 emit_alu_K(SRA, O1, 0, ctx);
1449
1450 emit_reg_move(O0, bpf2sparc[BPF_REG_0], ctx);
1451 break;
1452
1453 default:
1454 pr_err_once("unknown opcode %02x\n", code);
1455 return -EINVAL;
1456 }
1457
1458 return 0;
1459}
1460
1461static int build_body(struct jit_ctx *ctx)
1462{
1463 const struct bpf_prog *prog = ctx->prog;
1464 int i;
1465
1466 for (i = 0; i < prog->len; i++) {
1467 const struct bpf_insn *insn = &prog->insnsi[i];
1468 int ret;
1469
1470 ret = build_insn(insn, ctx);
1471
1472 if (ret > 0) {
1473 i++;
1474 ctx->offset[i] = ctx->idx;
1475 continue;
1476 }
1477 ctx->offset[i] = ctx->idx;
1478 if (ret)
1479 return ret;
1480 }
1481 return 0;
1482}
1483
1484static void jit_fill_hole(void *area, unsigned int size)
1485{
1486 u32 *ptr;
1487 /* We are guaranteed to have aligned memory. */
1488 for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
1489 *ptr++ = 0x91d02005; /* ta 5 */
1490}
1491
1492struct sparc64_jit_data {
1493 struct bpf_binary_header *header;
1494 u8 *image;
1495 struct jit_ctx ctx;
1496};
1497
1498struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1499{
1500 struct bpf_prog *tmp, *orig_prog = prog;
1501 struct sparc64_jit_data *jit_data;
1502 struct bpf_binary_header *header;
1503 bool tmp_blinded = false;
1504 bool extra_pass = false;
1505 struct jit_ctx ctx;
1506 u32 image_size;
1507 u8 *image_ptr;
1508 int pass;
1509
1510 if (!prog->jit_requested)
1511 return orig_prog;
1512
1513 tmp = bpf_jit_blind_constants(prog);
1514 /* If blinding was requested and we failed during blinding,
1515 * we must fall back to the interpreter.
1516 */
1517 if (IS_ERR(tmp))
1518 return orig_prog;
1519 if (tmp != prog) {
1520 tmp_blinded = true;
1521 prog = tmp;
1522 }
1523
1524 jit_data = prog->aux->jit_data;
1525 if (!jit_data) {
1526 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1527 if (!jit_data) {
1528 prog = orig_prog;
1529 goto out;
1530 }
1531 prog->aux->jit_data = jit_data;
1532 }
1533 if (jit_data->ctx.offset) {
1534 ctx = jit_data->ctx;
1535 image_ptr = jit_data->image;
1536 header = jit_data->header;
1537 extra_pass = true;
1538 image_size = sizeof(u32) * ctx.idx;
1539 goto skip_init_ctx;
1540 }
1541
1542 memset(&ctx, 0, sizeof(ctx));
1543 ctx.prog = prog;
1544
1545 ctx.offset = kcalloc(prog->len, sizeof(unsigned int), GFP_KERNEL);
1546 if (ctx.offset == NULL) {
1547 prog = orig_prog;
1548 goto out_off;
1549 }
1550
1551 /* Fake pass to detect features used, and get an accurate assessment
1552 * of what the final image size will be.
1553 */
1554 if (build_body(&ctx)) {
1555 prog = orig_prog;
1556 goto out_off;
1557 }
1558 build_prologue(&ctx);
1559 build_epilogue(&ctx);
1560
1561 /* Now we know the actual image size. */
1562 image_size = sizeof(u32) * ctx.idx;
1563 header = bpf_jit_binary_alloc(image_size, &image_ptr,
1564 sizeof(u32), jit_fill_hole);
1565 if (header == NULL) {
1566 prog = orig_prog;
1567 goto out_off;
1568 }
1569
1570 ctx.image = (u32 *)image_ptr;
1571skip_init_ctx:
1572 for (pass = 1; pass < 3; pass++) {
1573 ctx.idx = 0;
1574
1575 build_prologue(&ctx);
1576
1577 if (build_body(&ctx)) {
1578 bpf_jit_binary_free(header);
1579 prog = orig_prog;
1580 goto out_off;
1581 }
1582
1583 build_epilogue(&ctx);
1584
1585 if (bpf_jit_enable > 1)
1586 pr_info("Pass %d: shrink = %d, seen = [%c%c%c%c%c%c%c]\n", pass,
1587 image_size - (ctx.idx * 4),
1588 ctx.tmp_1_used ? '1' : ' ',
1589 ctx.tmp_2_used ? '2' : ' ',
1590 ctx.tmp_3_used ? '3' : ' ',
1591 ctx.saw_ld_abs_ind ? 'L' : ' ',
1592 ctx.saw_frame_pointer ? 'F' : ' ',
1593 ctx.saw_call ? 'C' : ' ',
1594 ctx.saw_tail_call ? 'T' : ' ');
1595 }
1596
1597 if (bpf_jit_enable > 1)
1598 bpf_jit_dump(prog->len, image_size, pass, ctx.image);
1599
1600 bpf_flush_icache(header, (u8 *)header + (header->pages * PAGE_SIZE));
1601
1602 if (!prog->is_func || extra_pass) {
1603 bpf_jit_binary_lock_ro(header);
1604 } else {
1605 jit_data->ctx = ctx;
1606 jit_data->image = image_ptr;
1607 jit_data->header = header;
1608 }
1609
1610 prog->bpf_func = (void *)ctx.image;
1611 prog->jited = 1;
1612 prog->jited_len = image_size;
1613
1614 if (!prog->is_func || extra_pass) {
1615out_off:
1616 kfree(ctx.offset);
1617 kfree(jit_data);
1618 prog->aux->jit_data = NULL;
1619 }
1620out:
1621 if (tmp_blinded)
1622 bpf_jit_prog_release_other(prog, prog == orig_prog ?
1623 tmp : orig_prog);
1624 return prog;
1625}