projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
module: remove mod arg from module_free, rename module_memfree().
[linux-2.6-block.git] / arch / sparc / net / bpf_jit_comp.c
CommitLineData
2809a208
DM		 1#include <linux/moduleloader.h>
2#include <linux/workqueue.h>
3#include <linux/netdevice.h>
4#include <linux/filter.h>
5#include <linux/cache.h>
54e7e2df 6#include <linux/if_vlan.h>
2809a208
DM		 7
8#include <asm/cacheflush.h>
9#include <asm/ptrace.h>
10
11#include "bpf_jit.h"
12
13int bpf_jit_enable __read_mostly;
14
2809a208
DM		 15static inline bool is_simm13(unsigned int value)
16{
17 return value + 0x1000 < 0x2000;
18}
19
20static void bpf_flush_icache(void *start_, void *end_)
21{
22#ifdef CONFIG_SPARC64
23 /* Cheetah's I-cache is fully coherent. */
24 if (tlb_type == spitfire) {
25 unsigned long start = (unsigned long) start_;
26 unsigned long end = (unsigned long) end_;
27
28 start &= ~7UL;
29 end = (end + 7UL) & ~7UL;
30 while (start < end) {
31 flushi(start);
32 start += 32;
33 }
34 }
35#endif
36}
37
38#define SEEN_DATAREF 1 /* might call external helpers */
39#define SEEN_XREG 2 /* ebx is used */
40#define SEEN_MEM 4 /* use mem[] for temporary storage */
41
42#define S13(X) ((X) & 0x1fff)
43#define IMMED 0x00002000
44#define RD(X) ((X) << 25)
45#define RS1(X) ((X) << 14)
46#define RS2(X) ((X))
47#define OP(X) ((X) << 30)
48#define OP2(X) ((X) << 22)
49#define OP3(X) ((X) << 19)
50#define COND(X) ((X) << 25)
51#define F1(X) OP(X)
52#define F2(X, Y) (OP(X) | OP2(Y))
53#define F3(X, Y) (OP(X) | OP3(Y))
54
584c5e2a
DM		 55#define CONDN COND(0x0)
56#define CONDE COND(0x1)
57#define CONDLE COND(0x2)
58#define CONDL COND(0x3)
59#define CONDLEU COND(0x4)
60#define CONDCS COND(0x5)
61#define CONDNEG COND(0x6)
62#define CONDVC COND(0x7)
63#define CONDA COND(0x8)
64#define CONDNE COND(0x9)
65#define CONDG COND(0xa)
66#define CONDGE COND(0xb)
67#define CONDGU COND(0xc)
68#define CONDCC COND(0xd)
69#define CONDPOS COND(0xe)
70#define CONDVS COND(0xf)
2809a208
DM		 71
72#define CONDGEU CONDCC
73#define CONDLU CONDCS
74
75#define WDISP22(X) (((X) >> 2) & 0x3fffff)
76
77#define BA (F2(0, 2) | CONDA)
78#define BGU (F2(0, 2) | CONDGU)
79#define BLEU (F2(0, 2) | CONDLEU)
80#define BGEU (F2(0, 2) | CONDGEU)
81#define BLU (F2(0, 2) | CONDLU)
82#define BE (F2(0, 2) | CONDE)
83#define BNE (F2(0, 2) | CONDNE)
84
85#ifdef CONFIG_SPARC64
569810d1 86#define BE_PTR (F2(0, 1) | CONDE | (2 << 20))
2809a208 87#else
588f5d62 88#define BE_PTR BE
2809a208
DM		 89#endif
90
91#define SETHI(K, REG) \
92 (F2(0, 0x4) | RD(REG) | (((K) >> 10) & 0x3fffff))
93#define OR_LO(K, REG) \
94 (F3(2, 0x02) | IMMED | RS1(REG) | ((K) & 0x3ff) | RD(REG))
95
96#define ADD F3(2, 0x00)
97#define AND F3(2, 0x01)
98#define ANDCC F3(2, 0x11)
99#define OR F3(2, 0x02)
d7ce8a5f 100#define XOR F3(2, 0x03)
2809a208
DM		 101#define SUB F3(2, 0x04)
102#define SUBCC F3(2, 0x14)
103#define MUL F3(2, 0x0a) /* umul */
104#define DIV F3(2, 0x0e) /* udiv */
105#define SLL F3(2, 0x25)
106#define SRL F3(2, 0x26)
107#define JMPL F3(2, 0x38)
108#define CALL F1(1)
109#define BR F2(0, 0x01)
110#define RD_Y F3(2, 0x28)
111#define WR_Y F3(2, 0x30)
112
113#define LD32 F3(3, 0x00)
114#define LD8 F3(3, 0x01)
115#define LD16 F3(3, 0x02)
116#define LD64 F3(3, 0x0b)
117#define ST32 F3(3, 0x04)
118
119#ifdef CONFIG_SPARC64
120#define LDPTR LD64
121#define BASE_STACKFRAME 176
122#else
123#define LDPTR LD32
124#define BASE_STACKFRAME 96
125#endif
126
127#define LD32I (LD32 | IMMED)
128#define LD8I (LD8 | IMMED)
129#define LD16I (LD16 | IMMED)
130#define LD64I (LD64 | IMMED)
131#define LDPTRI (LDPTR | IMMED)
132#define ST32I (ST32 | IMMED)
133
134#define emit_nop() \
135do { \
136 *prog++ = SETHI(0, G0); \
137} while (0)
138
139#define emit_neg() \
140do { /* sub %g0, r_A, r_A */ \
141 *prog++ = SUB | RS1(G0) | RS2(r_A) | RD(r_A); \
142} while (0)
143
144#define emit_reg_move(FROM, TO) \
145do { /* or %g0, FROM, TO */ \
146 *prog++ = OR | RS1(G0) | RS2(FROM) | RD(TO); \
147} while (0)
148
149#define emit_clear(REG) \
150do { /* or %g0, %g0, REG */ \
151 *prog++ = OR | RS1(G0) | RS2(G0) | RD(REG); \
152} while (0)
153
154#define emit_set_const(K, REG) \
155do { /* sethi %hi(K), REG */ \
156 *prog++ = SETHI(K, REG); \
157 /* or REG, %lo(K), REG */ \
158 *prog++ = OR_LO(K, REG); \
159} while (0)
160
161 /* Emit
162 *
584c5e2a 163 * OP r_A, r_X, r_A
2809a208
DM		 164 */
165#define emit_alu_X(OPCODE) \
166do { \
167 seen |= SEEN_XREG; \
168 *prog++ = OPCODE | RS1(r_A) | RS2(r_X) | RD(r_A); \
169} while (0)
170
171 /* Emit either:
172 *
173 * OP r_A, K, r_A
174 *
175 * or
176 *
177 * sethi %hi(K), r_TMP
178 * or r_TMP, %lo(K), r_TMP
179 * OP r_A, r_TMP, r_A
180 *
181 * depending upon whether K fits in a signed 13-bit
182 * immediate instruction field. Emit nothing if K
183 * is zero.
184 */
185#define emit_alu_K(OPCODE, K) \
584c5e2a 186do { \
35607b02 187 if (K || OPCODE == AND || OPCODE == MUL) { \
2809a208
DM		 188 unsigned int _insn = OPCODE; \
189 _insn |= RS1(r_A) | RD(r_A); \
190 if (is_simm13(K)) { \
191 *prog++ = _insn | IMMED | S13(K); \
192 } else { \
193 emit_set_const(K, r_TMP); \
194 *prog++ = _insn | RS2(r_TMP); \
584c5e2a 195 } \
2809a208
DM		 196 } \
197} while (0)
198
199#define emit_loadimm(K, DEST) \
200do { \
201 if (is_simm13(K)) { \
202 /* or %g0, K, DEST */ \
203 *prog++ = OR | IMMED | RS1(G0) | S13(K) | RD(DEST); \
204 } else { \
205 emit_set_const(K, DEST); \
206 } \
207} while (0)
208
209#define emit_loadptr(BASE, STRUCT, FIELD, DEST) \
210do { unsigned int _off = offsetof(STRUCT, FIELD); \
211 BUILD_BUG_ON(FIELD_SIZEOF(STRUCT, FIELD) != sizeof(void *)); \
212 *prog++ = LDPTRI | RS1(BASE) | S13(_off) | RD(DEST); \
584c5e2a 213} while (0)
2809a208
DM		 214
215#define emit_load32(BASE, STRUCT, FIELD, DEST) \
216do { unsigned int _off = offsetof(STRUCT, FIELD); \
217 BUILD_BUG_ON(FIELD_SIZEOF(STRUCT, FIELD) != sizeof(u32)); \
218 *prog++ = LD32I | RS1(BASE) | S13(_off) | RD(DEST); \
584c5e2a 219} while (0)
2809a208
DM		 220
221#define emit_load16(BASE, STRUCT, FIELD, DEST) \
222do { unsigned int _off = offsetof(STRUCT, FIELD); \
223 BUILD_BUG_ON(FIELD_SIZEOF(STRUCT, FIELD) != sizeof(u16)); \
224 *prog++ = LD16I | RS1(BASE) | S13(_off) | RD(DEST); \
584c5e2a 225} while (0)
2809a208
DM		 226
227#define __emit_load8(BASE, STRUCT, FIELD, DEST) \
228do { unsigned int _off = offsetof(STRUCT, FIELD); \
229 *prog++ = LD8I | RS1(BASE) | S13(_off) | RD(DEST); \
584c5e2a 230} while (0)
2809a208
DM		 231
232#define emit_load8(BASE, STRUCT, FIELD, DEST) \
233do { BUILD_BUG_ON(FIELD_SIZEOF(STRUCT, FIELD) != sizeof(u8)); \
234 __emit_load8(BASE, STRUCT, FIELD, DEST); \
584c5e2a 235} while (0)
2809a208 236
f6f2332d
AS		 237#ifdef CONFIG_SPARC64
238#define BIAS (STACK_BIAS - 4)
239#else
240#define BIAS (-4)
241#endif
242
243#define emit_ldmem(OFF, DEST) \
244do { *prog++ = LD32I | RS1(SP) | S13(BIAS - (OFF)) | RD(DEST); \
584c5e2a 245} while (0)
2809a208 246
f6f2332d
AS		 247#define emit_stmem(OFF, SRC) \
248do { *prog++ = ST32I | RS1(SP) | S13(BIAS - (OFF)) | RD(SRC); \
584c5e2a 249} while (0)
2809a208 250
2809a208
DM		 251#ifdef CONFIG_SMP
252#ifdef CONFIG_SPARC64
253#define emit_load_cpu(REG) \
254 emit_load16(G6, struct thread_info, cpu, REG)
255#else
256#define emit_load_cpu(REG) \
257 emit_load32(G6, struct thread_info, cpu, REG)
258#endif
259#else
260#define emit_load_cpu(REG) emit_clear(REG)
261#endif
262
263#define emit_skb_loadptr(FIELD, DEST) \
264 emit_loadptr(r_SKB, struct sk_buff, FIELD, DEST)
265#define emit_skb_load32(FIELD, DEST) \
266 emit_load32(r_SKB, struct sk_buff, FIELD, DEST)
267#define emit_skb_load16(FIELD, DEST) \
268 emit_load16(r_SKB, struct sk_buff, FIELD, DEST)
269#define __emit_skb_load8(FIELD, DEST) \
270 __emit_load8(r_SKB, struct sk_buff, FIELD, DEST)
271#define emit_skb_load8(FIELD, DEST) \
272 emit_load8(r_SKB, struct sk_buff, FIELD, DEST)
273
274#define emit_jmpl(BASE, IMM_OFF, LREG) \
275 *prog++ = (JMPL | IMMED | RS1(BASE) | S13(IMM_OFF) | RD(LREG))
276
277#define emit_call(FUNC) \
278do { void *_here = image + addrs[i] - 8; \
279 unsigned int _off = (void *)(FUNC) - _here; \
280 *prog++ = CALL | (((_off) >> 2) & 0x3fffffff); \
281 emit_nop(); \
282} while (0)
283
284#define emit_branch(BR_OPC, DEST) \
285do { unsigned int _here = addrs[i] - 8; \
286 *prog++ = BR_OPC | WDISP22((DEST) - _here); \
584c5e2a 287} while (0)
2809a208
DM		 288
289#define emit_branch_off(BR_OPC, OFF) \
290do { *prog++ = BR_OPC | WDISP22(OFF); \
584c5e2a 291} while (0)
2809a208
DM		 292
293#define emit_jump(DEST) emit_branch(BA, DEST)
294
584c5e2a
DM		 295#define emit_read_y(REG) *prog++ = RD_Y | RD(REG)
296#define emit_write_y(REG) *prog++ = WR_Y | IMMED | RS1(REG) | S13(0)
2809a208
DM		 297
298#define emit_cmp(R1, R2) \
299 *prog++ = (SUBCC | RS1(R1) | RS2(R2) | RD(G0))
300
301#define emit_cmpi(R1, IMM) \
302 *prog++ = (SUBCC | IMMED | RS1(R1) | S13(IMM) | RD(G0));
303
304#define emit_btst(R1, R2) \
305 *prog++ = (ANDCC | RS1(R1) | RS2(R2) | RD(G0))
306
307#define emit_btsti(R1, IMM) \
308 *prog++ = (ANDCC | IMMED | RS1(R1) | S13(IMM) | RD(G0));
309
310#define emit_sub(R1, R2, R3) \
311 *prog++ = (SUB | RS1(R1) | RS2(R2) | RD(R3))
312
313#define emit_subi(R1, IMM, R3) \
314 *prog++ = (SUB | IMMED | RS1(R1) | S13(IMM) | RD(R3))
315
316#define emit_add(R1, R2, R3) \
317 *prog++ = (ADD | RS1(R1) | RS2(R2) | RD(R3))
318
319#define emit_addi(R1, IMM, R3) \
320 *prog++ = (ADD | IMMED | RS1(R1) | S13(IMM) | RD(R3))
321
54e7e2df
DB		 322#define emit_and(R1, R2, R3) \
323 *prog++ = (AND | RS1(R1) | RS2(R2) | RD(R3))
324
325#define emit_andi(R1, IMM, R3) \
326 *prog++ = (AND | IMMED | RS1(R1) | S13(IMM) | RD(R3))
327
2809a208
DM		 328#define emit_alloc_stack(SZ) \
329 *prog++ = (SUB | IMMED | RS1(SP) | S13(SZ) | RD(SP))
330
331#define emit_release_stack(SZ) \
332 *prog++ = (ADD | IMMED | RS1(SP) | S13(SZ) | RD(SP))
333
584c5e2a
DM		 334/* A note about branch offset calculations. The addrs[] array,
335 * indexed by BPF instruction, records the address after all the
336 * sparc instructions emitted for that BPF instruction.
337 *
338 * The most common case is to emit a branch at the end of such
339 * a code sequence. So this would be two instructions, the
340 * branch and it's delay slot.
341 *
342 * Therefore by default the branch emitters calculate the branch
343 * offset field as:
344 *
345 * destination - (addrs[i] - 8)
346 *
347 * This "addrs[i] - 8" is the address of the branch itself or
348 * what "." would be in assembler notation. The "8" part is
349 * how we take into consideration the branch and it's delay
350 * slot mentioned above.
351 *
352 * Sometimes we need to emit a branch earlier in the code
353 * sequence. And in these situations we adjust "destination"
354 * to accomodate this difference. For example, if we needed
355 * to emit a branch (and it's delay slot) right before the
356 * final instruction emitted for a BPF opcode, we'd use
357 * "destination + 4" instead of just plain "destination" above.
358 *
359 * This is why you see all of these funny emit_branch() and
360 * emit_jump() calls with adjusted offsets.
361 */
362
7ae457c1 363void bpf_jit_compile(struct bpf_prog *fp)
2809a208
DM		 364{
365 unsigned int cleanup_addr, proglen, oldproglen = 0;
366 u32 temp[8], *prog, *func, seen = 0, pass;
367 const struct sock_filter *filter = fp->insns;
368 int i, flen = fp->len, pc_ret0 = -1;
369 unsigned int *addrs;
370 void *image;
371
372 if (!bpf_jit_enable)
373 return;
374
375 addrs = kmalloc(flen * sizeof(*addrs), GFP_KERNEL);
376 if (addrs == NULL)
377 return;
378
379 /* Before first pass, make a rough estimation of addrs[]
380 * each bpf instruction is translated to less than 64 bytes
381 */
382 for (proglen = 0, i = 0; i < flen; i++) {
383 proglen += 64;
384 addrs[i] = proglen;
385 }
386 cleanup_addr = proglen; /* epilogue address */
387 image = NULL;
388 for (pass = 0; pass < 10; pass++) {
389 u8 seen_or_pass0 = (pass == 0) ? (SEEN_XREG | SEEN_DATAREF | SEEN_MEM) : seen;
390
391 /* no prologue/epilogue for trivial filters (RET something) */
392 proglen = 0;
393 prog = temp;
394
395 /* Prologue */
396 if (seen_or_pass0) {
397 if (seen_or_pass0 & SEEN_MEM) {
398 unsigned int sz = BASE_STACKFRAME;
399 sz += BPF_MEMWORDS * sizeof(u32);
400 emit_alloc_stack(sz);
401 }
402
403 /* Make sure we dont leek kernel memory. */
404 if (seen_or_pass0 & SEEN_XREG)
405 emit_clear(r_X);
406
407 /* If this filter needs to access skb data,
7b56f76e 408 * load %o4 and %o5 with:
2809a208
DM		 409 * %o4 = skb->len - skb->data_len
410 * %o5 = skb->data
411 * And also back up %o7 into r_saved_O7 so we can
412 * invoke the stubs using 'call'.
413 */
414 if (seen_or_pass0 & SEEN_DATAREF) {
415 emit_load32(r_SKB, struct sk_buff, len, r_HEADLEN);
416 emit_load32(r_SKB, struct sk_buff, data_len, r_TMP);
417 emit_sub(r_HEADLEN, r_TMP, r_HEADLEN);
418 emit_loadptr(r_SKB, struct sk_buff, data, r_SKB_DATA);
419 }
420 }
421 emit_reg_move(O7, r_saved_O7);
422
423 switch (filter[0].code) {
34805931
DB		 424 case BPF_RET | BPF_K:
425 case BPF_LD | BPF_W | BPF_LEN:
426 case BPF_LD | BPF_W | BPF_ABS:
427 case BPF_LD | BPF_H | BPF_ABS:
428 case BPF_LD | BPF_B | BPF_ABS:
2809a208
DM		 429 /* The first instruction sets the A register (or is
430 * a "RET 'constant'")
431 */
432 break;
433 default:
434 /* Make sure we dont leak kernel information to the
435 * user.
436 */
437 emit_clear(r_A); /* A = 0 */
438 }
439
440 for (i = 0; i < flen; i++) {
441 unsigned int K = filter[i].k;
442 unsigned int t_offset;
443 unsigned int f_offset;
444 u32 t_op, f_op;
34805931 445 u16 code = bpf_anc_helper(&filter[i]);
2809a208
DM		 446 int ilen;
447
34805931
DB		 448 switch (code) {
449 case BPF_ALU | BPF_ADD | BPF_X: /* A += X; */
2809a208
DM		 450 emit_alu_X(ADD);
451 break;
34805931 452 case BPF_ALU | BPF_ADD | BPF_K: /* A += K; */
2809a208
DM		 453 emit_alu_K(ADD, K);
454 break;
34805931 455 case BPF_ALU | BPF_SUB | BPF_X: /* A -= X; */
2809a208
DM		 456 emit_alu_X(SUB);
457 break;
34805931 458 case BPF_ALU | BPF_SUB | BPF_K: /* A -= K */
2809a208
DM		 459 emit_alu_K(SUB, K);
460 break;
34805931 461 case BPF_ALU | BPF_AND | BPF_X: /* A &= X */
2809a208
DM		 462 emit_alu_X(AND);
463 break;
34805931 464 case BPF_ALU | BPF_AND | BPF_K: /* A &= K */
2809a208
DM		 465 emit_alu_K(AND, K);
466 break;
34805931 467 case BPF_ALU | BPF_OR | BPF_X: /* A |= X */
2809a208
DM		 468 emit_alu_X(OR);
469 break;
34805931 470 case BPF_ALU | BPF_OR | BPF_K: /* A |= K */
2809a208
DM		 471 emit_alu_K(OR, K);
472 break;
34805931
DB		 473 case BPF_ANC | SKF_AD_ALU_XOR_X: /* A ^= X; */
474 case BPF_ALU | BPF_XOR | BPF_X:
d7ce8a5f
DM		 475 emit_alu_X(XOR);
476 break;
34805931 477 case BPF_ALU | BPF_XOR | BPF_K: /* A ^= K */
aa99521e
DB		 478 emit_alu_K(XOR, K);
479 break;
34805931 480 case BPF_ALU | BPF_LSH | BPF_X: /* A <<= X */
2809a208
DM		 481 emit_alu_X(SLL);
482 break;
34805931 483 case BPF_ALU | BPF_LSH | BPF_K: /* A <<= K */
2809a208
DM		 484 emit_alu_K(SLL, K);
485 break;
34805931 486 case BPF_ALU | BPF_RSH | BPF_X: /* A >>= X */
2809a208
DM		 487 emit_alu_X(SRL);
488 break;
34805931 489 case BPF_ALU | BPF_RSH | BPF_K: /* A >>= K */
2809a208
DM		 490 emit_alu_K(SRL, K);
491 break;
34805931 492 case BPF_ALU | BPF_MUL | BPF_X: /* A *= X; */
2809a208
DM		 493 emit_alu_X(MUL);
494 break;
34805931 495 case BPF_ALU | BPF_MUL | BPF_K: /* A *= K */
2809a208
DM		 496 emit_alu_K(MUL, K);
497 break;
34805931 498 case BPF_ALU | BPF_DIV | BPF_K: /* A /= K with K != 0*/
aee636c4
ED		 499 if (K == 1)
500 break;
501 emit_write_y(G0);
502#ifdef CONFIG_SPARC32
503 /* The Sparc v8 architecture requires
504 * three instructions between a %y
505 * register write and the first use.
506 */
507 emit_nop();
508 emit_nop();
509 emit_nop();
510#endif
511 emit_alu_K(DIV, K);
2809a208 512 break;
34805931 513 case BPF_ALU | BPF_DIV | BPF_X: /* A /= X; */
2809a208
DM		 514 emit_cmpi(r_X, 0);
515 if (pc_ret0 > 0) {
516 t_offset = addrs[pc_ret0 - 1];
517#ifdef CONFIG_SPARC32
518 emit_branch(BE, t_offset + 20);
519#else
520 emit_branch(BE, t_offset + 8);
521#endif
522 emit_nop(); /* delay slot */
523 } else {
524 emit_branch_off(BNE, 16);
525 emit_nop();
526#ifdef CONFIG_SPARC32
527 emit_jump(cleanup_addr + 20);
528#else
529 emit_jump(cleanup_addr + 8);
530#endif
531 emit_clear(r_A);
532 }
533 emit_write_y(G0);
534#ifdef CONFIG_SPARC32
584c5e2a
DM		 535 /* The Sparc v8 architecture requires
536 * three instructions between a %y
537 * register write and the first use.
538 */
2809a208
DM		 539 emit_nop();
540 emit_nop();
541 emit_nop();
542#endif
543 emit_alu_X(DIV);
544 break;
34805931 545 case BPF_ALU | BPF_NEG:
2809a208
DM		 546 emit_neg();
547 break;
34805931 548 case BPF_RET | BPF_K:
2809a208
DM		 549 if (!K) {
550 if (pc_ret0 == -1)
551 pc_ret0 = i;
552 emit_clear(r_A);
553 } else {
554 emit_loadimm(K, r_A);
555 }
556 /* Fallthrough */
34805931 557 case BPF_RET | BPF_A:
2809a208
DM		 558 if (seen_or_pass0) {
559 if (i != flen - 1) {
560 emit_jump(cleanup_addr);
561 emit_nop();
562 break;
563 }
564 if (seen_or_pass0 & SEEN_MEM) {
565 unsigned int sz = BASE_STACKFRAME;
566 sz += BPF_MEMWORDS * sizeof(u32);
567 emit_release_stack(sz);
568 }
569 }
570 /* jmpl %r_saved_O7 + 8, %g0 */
571 emit_jmpl(r_saved_O7, 8, G0);
572 emit_reg_move(r_A, O0); /* delay slot */
573 break;
34805931 574 case BPF_MISC | BPF_TAX:
2809a208
DM		 575 seen |= SEEN_XREG;
576 emit_reg_move(r_A, r_X);
577 break;
34805931 578 case BPF_MISC | BPF_TXA:
2809a208
DM		 579 seen |= SEEN_XREG;
580 emit_reg_move(r_X, r_A);
581 break;
34805931 582 case BPF_ANC | SKF_AD_CPU:
2809a208
DM		 583 emit_load_cpu(r_A);
584 break;
34805931 585 case BPF_ANC | SKF_AD_PROTOCOL:
2809a208
DM		 586 emit_skb_load16(protocol, r_A);
587 break;
34805931 588 case BPF_ANC | SKF_AD_PKTTYPE:
709f6c58
AS		 589 __emit_skb_load8(__pkt_type_offset, r_A);
590 emit_andi(r_A, PKT_TYPE_MAX, r_A);
2809a208
DM		 591 emit_alu_K(SRL, 5);
592 break;
34805931 593 case BPF_ANC | SKF_AD_IFINDEX:
2809a208
DM		 594 emit_skb_loadptr(dev, r_A);
595 emit_cmpi(r_A, 0);
569810d1 596 emit_branch(BE_PTR, cleanup_addr + 4);
2809a208
DM		 597 emit_nop();
598 emit_load32(r_A, struct net_device, ifindex, r_A);
599 break;
34805931 600 case BPF_ANC | SKF_AD_MARK:
2809a208
DM		 601 emit_skb_load32(mark, r_A);
602 break;
34805931 603 case BPF_ANC | SKF_AD_QUEUE:
2809a208
DM		 604 emit_skb_load16(queue_mapping, r_A);
605 break;
34805931 606 case BPF_ANC | SKF_AD_HATYPE:
2809a208
DM		 607 emit_skb_loadptr(dev, r_A);
608 emit_cmpi(r_A, 0);
569810d1 609 emit_branch(BE_PTR, cleanup_addr + 4);
2809a208
DM		 610 emit_nop();
611 emit_load16(r_A, struct net_device, type, r_A);
612 break;
34805931 613 case BPF_ANC | SKF_AD_RXHASH:
61b905da 614 emit_skb_load32(hash, r_A);
2809a208 615 break;
34805931
DB		 616 case BPF_ANC | SKF_AD_VLAN_TAG:
617 case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
54e7e2df 618 emit_skb_load16(vlan_tci, r_A);
f6f2332d
AS		 619 if (code != (BPF_ANC | SKF_AD_VLAN_TAG)) {
620 emit_alu_K(SRL, 12);
621 emit_andi(r_A, 1, r_A);
54e7e2df 622 } else {
f6f2332d 623 emit_loadimm(~VLAN_TAG_PRESENT, r_TMP);
54e7e2df
DB		 624 emit_and(r_A, r_TMP, r_A);
625 }
626 break;
cec08315
AS		 627 case BPF_LD | BPF_W | BPF_LEN:
628 emit_skb_load32(len, r_A);
629 break;
630 case BPF_LDX | BPF_W | BPF_LEN:
631 emit_skb_load32(len, r_X);
632 break;
34805931 633 case BPF_LD | BPF_IMM:
2809a208
DM		 634 emit_loadimm(K, r_A);
635 break;
34805931 636 case BPF_LDX | BPF_IMM:
2809a208
DM		 637 emit_loadimm(K, r_X);
638 break;
34805931 639 case BPF_LD | BPF_MEM:
f6f2332d 640 seen |= SEEN_MEM;
2809a208
DM		 641 emit_ldmem(K * 4, r_A);
642 break;
34805931 643 case BPF_LDX | BPF_MEM:
f6f2332d 644 seen |= SEEN_MEM | SEEN_XREG;
2809a208
DM		 645 emit_ldmem(K * 4, r_X);
646 break;
34805931 647 case BPF_ST:
f6f2332d 648 seen |= SEEN_MEM;
2809a208
DM		 649 emit_stmem(K * 4, r_A);
650 break;
34805931 651 case BPF_STX:
f6f2332d 652 seen |= SEEN_MEM | SEEN_XREG;
2809a208
DM		 653 emit_stmem(K * 4, r_X);
654 break;
655
656#define CHOOSE_LOAD_FUNC(K, func) \
657 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
658
34805931 659 case BPF_LD | BPF_W | BPF_ABS:
2809a208
DM		 660 func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_word);
661common_load: seen |= SEEN_DATAREF;
662 emit_loadimm(K, r_OFF);
663 emit_call(func);
664 break;
34805931 665 case BPF_LD | BPF_H | BPF_ABS:
2809a208
DM		 666 func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_half);
667 goto common_load;
34805931 668 case BPF_LD | BPF_B | BPF_ABS:
2809a208
DM		 669 func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_byte);
670 goto common_load;
34805931 671 case BPF_LDX | BPF_B | BPF_MSH:
2809a208
DM		 672 func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_byte_msh);
673 goto common_load;
34805931 674 case BPF_LD | BPF_W | BPF_IND:
2809a208
DM		 675 func = bpf_jit_load_word;
676common_load_ind: seen |= SEEN_DATAREF | SEEN_XREG;
677 if (K) {
678 if (is_simm13(K)) {
679 emit_addi(r_X, K, r_OFF);
680 } else {
681 emit_loadimm(K, r_TMP);
682 emit_add(r_X, r_TMP, r_OFF);
683 }
684 } else {
685 emit_reg_move(r_X, r_OFF);
686 }
687 emit_call(func);
688 break;
34805931 689 case BPF_LD | BPF_H | BPF_IND:
2809a208
DM		 690 func = bpf_jit_load_half;
691 goto common_load_ind;
34805931 692 case BPF_LD | BPF_B | BPF_IND:
2809a208
DM		 693 func = bpf_jit_load_byte;
694 goto common_load_ind;
34805931 695 case BPF_JMP | BPF_JA:
2809a208
DM		 696 emit_jump(addrs[i + K]);
697 emit_nop();
698 break;
699
700#define COND_SEL(CODE, TOP, FOP) \
701 case CODE: \
702 t_op = TOP; \
703 f_op = FOP; \
704 goto cond_branch
705
34805931
DB		 706 COND_SEL(BPF_JMP | BPF_JGT | BPF_K, BGU, BLEU);
707 COND_SEL(BPF_JMP | BPF_JGE | BPF_K, BGEU, BLU);
708 COND_SEL(BPF_JMP | BPF_JEQ | BPF_K, BE, BNE);
709 COND_SEL(BPF_JMP | BPF_JSET | BPF_K, BNE, BE);
710 COND_SEL(BPF_JMP | BPF_JGT | BPF_X, BGU, BLEU);
711 COND_SEL(BPF_JMP | BPF_JGE | BPF_X, BGEU, BLU);
712 COND_SEL(BPF_JMP | BPF_JEQ | BPF_X, BE, BNE);
713 COND_SEL(BPF_JMP | BPF_JSET | BPF_X, BNE, BE);
2809a208
DM		 714
715cond_branch: f_offset = addrs[i + filter[i].jf];
716 t_offset = addrs[i + filter[i].jt];
717
718 /* same targets, can avoid doing the test :) */
719 if (filter[i].jt == filter[i].jf) {
720 emit_jump(t_offset);
721 emit_nop();
722 break;
723 }
724
34805931
DB		 725 switch (code) {
726 case BPF_JMP | BPF_JGT | BPF_X:
727 case BPF_JMP | BPF_JGE | BPF_X:
728 case BPF_JMP | BPF_JEQ | BPF_X:
2809a208
DM		 729 seen |= SEEN_XREG;
730 emit_cmp(r_A, r_X);
731 break;
34805931 732 case BPF_JMP | BPF_JSET | BPF_X:
2809a208
DM		 733 seen |= SEEN_XREG;
734 emit_btst(r_A, r_X);
735 break;
34805931
DB		 736 case BPF_JMP | BPF_JEQ | BPF_K:
737 case BPF_JMP | BPF_JGT | BPF_K:
738 case BPF_JMP | BPF_JGE | BPF_K:
2809a208
DM		 739 if (is_simm13(K)) {
740 emit_cmpi(r_A, K);
741 } else {
742 emit_loadimm(K, r_TMP);
743 emit_cmp(r_A, r_TMP);
744 }
745 break;
34805931 746 case BPF_JMP | BPF_JSET | BPF_K:
2809a208
DM		 747 if (is_simm13(K)) {
748 emit_btsti(r_A, K);
749 } else {
750 emit_loadimm(K, r_TMP);
751 emit_btst(r_A, r_TMP);
752 }
753 break;
754 }
755 if (filter[i].jt != 0) {
756 if (filter[i].jf)
757 t_offset += 8;
758 emit_branch(t_op, t_offset);
759 emit_nop(); /* delay slot */
760 if (filter[i].jf) {
761 emit_jump(f_offset);
762 emit_nop();
763 }
764 break;
765 }
766 emit_branch(f_op, f_offset);
767 emit_nop(); /* delay slot */
768 break;
769
770 default:
771 /* hmm, too complex filter, give up with jit compiler */
772 goto out;
773 }
774 ilen = (void *) prog - (void *) temp;
775 if (image) {
776 if (unlikely(proglen + ilen > oldproglen)) {
777 pr_err("bpb_jit_compile fatal error\n");
778 kfree(addrs);
be1f221c 779 module_memfree(image);
2809a208
DM		 780 return;
781 }
782 memcpy(image + proglen, temp, ilen);
783 }
784 proglen += ilen;
785 addrs[i] = proglen;
786 prog = temp;
787 }
788 /* last bpf instruction is always a RET :
789 * use it to give the cleanup instruction(s) addr
790 */
791 cleanup_addr = proglen - 8; /* jmpl; mov r_A,%o0; */
792 if (seen_or_pass0 & SEEN_MEM)
793 cleanup_addr -= 4; /* add %sp, X, %sp; */
794
795 if (image) {
796 if (proglen != oldproglen)
797 pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n",
798 proglen, oldproglen);
799 break;
800 }
801 if (proglen == oldproglen) {
5199dfe5 802 image = module_alloc(proglen);
2809a208
DM		 803 if (!image)
804 goto out;
805 }
806 oldproglen = proglen;
807 }
808
809 if (bpf_jit_enable > 1)
79617801 810 bpf_jit_dump(flen, proglen, pass, image);
2809a208
DM		 811
812 if (image) {
2809a208
DM		 813 bpf_flush_icache(image, image + proglen);
814 fp->bpf_func = (void *)image;
286aad3c 815 fp->jited = true;
2809a208
DM		 816 }
817out:
818 kfree(addrs);
819 return;
820}
821
7ae457c1 822void bpf_jit_free(struct bpf_prog *fp)
2809a208 823{
f8bbbfc3 824 if (fp->jited)
be1f221c 825 module_memfree(fp->bpf_func);
60a3b225
DB		 826
827 bpf_prog_unlock_free(fp);
2809a208 828}
Linux 6.x block layer and io_uring tree(s)