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>
6#include <linux/if_vlan.h>
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
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
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)
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
86#define BE_PTR		(F2(0, 1) | CONDE | (2 << 20))
87#else
88#define BE_PTR		BE
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)
100#define XOR		F3(2, 0x03)
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	 *
163	 *	OP	r_A, r_X, r_A
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)					\
186do {								\
187	if (K || OPCODE == AND || OPCODE == MUL) {		\
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);		\
195		}						\
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);		\
213} while (0)
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);		\
219} while (0)
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);		\
225} while (0)
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);		\
230} while (0)
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);			\
235} while (0)
236
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);	\
245} while (0)
246
247#define emit_stmem(OFF, SRC)						\
248do {	*prog++ = ST32I | RS1(SP) | S13(BIAS - (OFF)) | RD(SRC);	\
249} while (0)
250
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);	\
287} while (0)
288
289#define emit_branch_off(BR_OPC, OFF)			\
290do {	*prog++ = BR_OPC | WDISP22(OFF);		\
291} while (0)
292
293#define emit_jump(DEST)		emit_branch(BA, DEST)
294
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)
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
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
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
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
363void bpf_jit_compile(struct bpf_prog *fp)
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,
408			 * load %o4 and %o5 with:
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		/* Make sure we dont leak kernel information to the user. */
424		if (bpf_needs_clear_a(&filter[0]))
425			emit_clear(r_A); /* A = 0 */
426
427		for (i = 0; i < flen; i++) {
428			unsigned int K = filter[i].k;
429			unsigned int t_offset;
430			unsigned int f_offset;
431			u32 t_op, f_op;
432			u16 code = bpf_anc_helper(&filter[i]);
433			int ilen;
434
435			switch (code) {
436			case BPF_ALU | BPF_ADD | BPF_X:	/* A += X; */
437				emit_alu_X(ADD);
438				break;
439			case BPF_ALU | BPF_ADD | BPF_K:	/* A += K; */
440				emit_alu_K(ADD, K);
441				break;
442			case BPF_ALU | BPF_SUB | BPF_X:	/* A -= X; */
443				emit_alu_X(SUB);
444				break;
445			case BPF_ALU | BPF_SUB | BPF_K:	/* A -= K */
446				emit_alu_K(SUB, K);
447				break;
448			case BPF_ALU | BPF_AND | BPF_X:	/* A &= X */
449				emit_alu_X(AND);
450				break;
451			case BPF_ALU | BPF_AND | BPF_K:	/* A &= K */
452				emit_alu_K(AND, K);
453				break;
454			case BPF_ALU | BPF_OR | BPF_X:	/* A |= X */
455				emit_alu_X(OR);
456				break;
457			case BPF_ALU | BPF_OR | BPF_K:	/* A |= K */
458				emit_alu_K(OR, K);
459				break;
460			case BPF_ANC | SKF_AD_ALU_XOR_X: /* A ^= X; */
461			case BPF_ALU | BPF_XOR | BPF_X:
462				emit_alu_X(XOR);
463				break;
464			case BPF_ALU | BPF_XOR | BPF_K:	/* A ^= K */
465				emit_alu_K(XOR, K);
466				break;
467			case BPF_ALU | BPF_LSH | BPF_X:	/* A <<= X */
468				emit_alu_X(SLL);
469				break;
470			case BPF_ALU | BPF_LSH | BPF_K:	/* A <<= K */
471				emit_alu_K(SLL, K);
472				break;
473			case BPF_ALU | BPF_RSH | BPF_X:	/* A >>= X */
474				emit_alu_X(SRL);
475				break;
476			case BPF_ALU | BPF_RSH | BPF_K:	/* A >>= K */
477				emit_alu_K(SRL, K);
478				break;
479			case BPF_ALU | BPF_MUL | BPF_X:	/* A *= X; */
480				emit_alu_X(MUL);
481				break;
482			case BPF_ALU | BPF_MUL | BPF_K:	/* A *= K */
483				emit_alu_K(MUL, K);
484				break;
485			case BPF_ALU | BPF_DIV | BPF_K:	/* A /= K with K != 0*/
486				if (K == 1)
487					break;
488				emit_write_y(G0);
489#ifdef CONFIG_SPARC32
490				/* The Sparc v8 architecture requires
491				 * three instructions between a %y
492				 * register write and the first use.
493				 */
494				emit_nop();
495				emit_nop();
496				emit_nop();
497#endif
498				emit_alu_K(DIV, K);
499				break;
500			case BPF_ALU | BPF_DIV | BPF_X:	/* A /= X; */
501				emit_cmpi(r_X, 0);
502				if (pc_ret0 > 0) {
503					t_offset = addrs[pc_ret0 - 1];
504#ifdef CONFIG_SPARC32
505					emit_branch(BE, t_offset + 20);
506#else
507					emit_branch(BE, t_offset + 8);
508#endif
509					emit_nop(); /* delay slot */
510				} else {
511					emit_branch_off(BNE, 16);
512					emit_nop();
513#ifdef CONFIG_SPARC32
514					emit_jump(cleanup_addr + 20);
515#else
516					emit_jump(cleanup_addr + 8);
517#endif
518					emit_clear(r_A);
519				}
520				emit_write_y(G0);
521#ifdef CONFIG_SPARC32
522				/* The Sparc v8 architecture requires
523				 * three instructions between a %y
524				 * register write and the first use.
525				 */
526				emit_nop();
527				emit_nop();
528				emit_nop();
529#endif
530				emit_alu_X(DIV);
531				break;
532			case BPF_ALU | BPF_NEG:
533				emit_neg();
534				break;
535			case BPF_RET | BPF_K:
536				if (!K) {
537					if (pc_ret0 == -1)
538						pc_ret0 = i;
539					emit_clear(r_A);
540				} else {
541					emit_loadimm(K, r_A);
542				}
543				/* Fallthrough */
544			case BPF_RET | BPF_A:
545				if (seen_or_pass0) {
546					if (i != flen - 1) {
547						emit_jump(cleanup_addr);
548						emit_nop();
549						break;
550					}
551					if (seen_or_pass0 & SEEN_MEM) {
552						unsigned int sz = BASE_STACKFRAME;
553						sz += BPF_MEMWORDS * sizeof(u32);
554						emit_release_stack(sz);
555					}
556				}
557				/* jmpl %r_saved_O7 + 8, %g0 */
558				emit_jmpl(r_saved_O7, 8, G0);
559				emit_reg_move(r_A, O0); /* delay slot */
560				break;
561			case BPF_MISC | BPF_TAX:
562				seen |= SEEN_XREG;
563				emit_reg_move(r_A, r_X);
564				break;
565			case BPF_MISC | BPF_TXA:
566				seen |= SEEN_XREG;
567				emit_reg_move(r_X, r_A);
568				break;
569			case BPF_ANC | SKF_AD_CPU:
570				emit_load_cpu(r_A);
571				break;
572			case BPF_ANC | SKF_AD_PROTOCOL:
573				emit_skb_load16(protocol, r_A);
574				break;
575			case BPF_ANC | SKF_AD_PKTTYPE:
576				__emit_skb_load8(__pkt_type_offset, r_A);
577				emit_andi(r_A, PKT_TYPE_MAX, r_A);
578				emit_alu_K(SRL, 5);
579				break;
580			case BPF_ANC | SKF_AD_IFINDEX:
581				emit_skb_loadptr(dev, r_A);
582				emit_cmpi(r_A, 0);
583				emit_branch(BE_PTR, cleanup_addr + 4);
584				emit_nop();
585				emit_load32(r_A, struct net_device, ifindex, r_A);
586				break;
587			case BPF_ANC | SKF_AD_MARK:
588				emit_skb_load32(mark, r_A);
589				break;
590			case BPF_ANC | SKF_AD_QUEUE:
591				emit_skb_load16(queue_mapping, r_A);
592				break;
593			case BPF_ANC | SKF_AD_HATYPE:
594				emit_skb_loadptr(dev, r_A);
595				emit_cmpi(r_A, 0);
596				emit_branch(BE_PTR, cleanup_addr + 4);
597				emit_nop();
598				emit_load16(r_A, struct net_device, type, r_A);
599				break;
600			case BPF_ANC | SKF_AD_RXHASH:
601				emit_skb_load32(hash, r_A);
602				break;
603			case BPF_ANC | SKF_AD_VLAN_TAG:
604			case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
605				emit_skb_load16(vlan_tci, r_A);
606				if (code != (BPF_ANC | SKF_AD_VLAN_TAG)) {
607					emit_alu_K(SRL, 12);
608					emit_andi(r_A, 1, r_A);
609				} else {
610					emit_loadimm(~VLAN_TAG_PRESENT, r_TMP);
611					emit_and(r_A, r_TMP, r_A);
612				}
613				break;
614			case BPF_LD | BPF_W | BPF_LEN:
615				emit_skb_load32(len, r_A);
616				break;
617			case BPF_LDX | BPF_W | BPF_LEN:
618				emit_skb_load32(len, r_X);
619				break;
620			case BPF_LD | BPF_IMM:
621				emit_loadimm(K, r_A);
622				break;
623			case BPF_LDX | BPF_IMM:
624				emit_loadimm(K, r_X);
625				break;
626			case BPF_LD | BPF_MEM:
627				seen |= SEEN_MEM;
628				emit_ldmem(K * 4, r_A);
629				break;
630			case BPF_LDX | BPF_MEM:
631				seen |= SEEN_MEM | SEEN_XREG;
632				emit_ldmem(K * 4, r_X);
633				break;
634			case BPF_ST:
635				seen |= SEEN_MEM;
636				emit_stmem(K * 4, r_A);
637				break;
638			case BPF_STX:
639				seen |= SEEN_MEM | SEEN_XREG;
640				emit_stmem(K * 4, r_X);
641				break;
642
643#define CHOOSE_LOAD_FUNC(K, func) \
644	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
645
646			case BPF_LD | BPF_W | BPF_ABS:
647				func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_word);
648common_load:			seen |= SEEN_DATAREF;
649				emit_loadimm(K, r_OFF);
650				emit_call(func);
651				break;
652			case BPF_LD | BPF_H | BPF_ABS:
653				func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_half);
654				goto common_load;
655			case BPF_LD | BPF_B | BPF_ABS:
656				func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_byte);
657				goto common_load;
658			case BPF_LDX | BPF_B | BPF_MSH:
659				func = CHOOSE_LOAD_FUNC(K, bpf_jit_load_byte_msh);
660				goto common_load;
661			case BPF_LD | BPF_W | BPF_IND:
662				func = bpf_jit_load_word;
663common_load_ind:		seen |= SEEN_DATAREF | SEEN_XREG;
664				if (K) {
665					if (is_simm13(K)) {
666						emit_addi(r_X, K, r_OFF);
667					} else {
668						emit_loadimm(K, r_TMP);
669						emit_add(r_X, r_TMP, r_OFF);
670					}
671				} else {
672					emit_reg_move(r_X, r_OFF);
673				}
674				emit_call(func);
675				break;
676			case BPF_LD | BPF_H | BPF_IND:
677				func = bpf_jit_load_half;
678				goto common_load_ind;
679			case BPF_LD | BPF_B | BPF_IND:
680				func = bpf_jit_load_byte;
681				goto common_load_ind;
682			case BPF_JMP | BPF_JA:
683				emit_jump(addrs[i + K]);
684				emit_nop();
685				break;
686
687#define COND_SEL(CODE, TOP, FOP)	\
688	case CODE:			\
689		t_op = TOP;		\
690		f_op = FOP;		\
691		goto cond_branch
692
693			COND_SEL(BPF_JMP | BPF_JGT | BPF_K, BGU, BLEU);
694			COND_SEL(BPF_JMP | BPF_JGE | BPF_K, BGEU, BLU);
695			COND_SEL(BPF_JMP | BPF_JEQ | BPF_K, BE, BNE);
696			COND_SEL(BPF_JMP | BPF_JSET | BPF_K, BNE, BE);
697			COND_SEL(BPF_JMP | BPF_JGT | BPF_X, BGU, BLEU);
698			COND_SEL(BPF_JMP | BPF_JGE | BPF_X, BGEU, BLU);
699			COND_SEL(BPF_JMP | BPF_JEQ | BPF_X, BE, BNE);
700			COND_SEL(BPF_JMP | BPF_JSET | BPF_X, BNE, BE);
701
702cond_branch:			f_offset = addrs[i + filter[i].jf];
703				t_offset = addrs[i + filter[i].jt];
704
705				/* same targets, can avoid doing the test :) */
706				if (filter[i].jt == filter[i].jf) {
707					emit_jump(t_offset);
708					emit_nop();
709					break;
710				}
711
712				switch (code) {
713				case BPF_JMP | BPF_JGT | BPF_X:
714				case BPF_JMP | BPF_JGE | BPF_X:
715				case BPF_JMP | BPF_JEQ | BPF_X:
716					seen |= SEEN_XREG;
717					emit_cmp(r_A, r_X);
718					break;
719				case BPF_JMP | BPF_JSET | BPF_X:
720					seen |= SEEN_XREG;
721					emit_btst(r_A, r_X);
722					break;
723				case BPF_JMP | BPF_JEQ | BPF_K:
724				case BPF_JMP | BPF_JGT | BPF_K:
725				case BPF_JMP | BPF_JGE | BPF_K:
726					if (is_simm13(K)) {
727						emit_cmpi(r_A, K);
728					} else {
729						emit_loadimm(K, r_TMP);
730						emit_cmp(r_A, r_TMP);
731					}
732					break;
733				case BPF_JMP | BPF_JSET | BPF_K:
734					if (is_simm13(K)) {
735						emit_btsti(r_A, K);
736					} else {
737						emit_loadimm(K, r_TMP);
738						emit_btst(r_A, r_TMP);
739					}
740					break;
741				}
742				if (filter[i].jt != 0) {
743					if (filter[i].jf)
744						t_offset += 8;
745					emit_branch(t_op, t_offset);
746					emit_nop(); /* delay slot */
747					if (filter[i].jf) {
748						emit_jump(f_offset);
749						emit_nop();
750					}
751					break;
752				}
753				emit_branch(f_op, f_offset);
754				emit_nop(); /* delay slot */
755				break;
756
757			default:
758				/* hmm, too complex filter, give up with jit compiler */
759				goto out;
760			}
761			ilen = (void *) prog - (void *) temp;
762			if (image) {
763				if (unlikely(proglen + ilen > oldproglen)) {
764					pr_err("bpb_jit_compile fatal error\n");
765					kfree(addrs);
766					module_memfree(image);
767					return;
768				}
769				memcpy(image + proglen, temp, ilen);
770			}
771			proglen += ilen;
772			addrs[i] = proglen;
773			prog = temp;
774		}
775		/* last bpf instruction is always a RET :
776		 * use it to give the cleanup instruction(s) addr
777		 */
778		cleanup_addr = proglen - 8; /* jmpl; mov r_A,%o0; */
779		if (seen_or_pass0 & SEEN_MEM)
780			cleanup_addr -= 4; /* add %sp, X, %sp; */
781
782		if (image) {
783			if (proglen != oldproglen)
784				pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n",
785				       proglen, oldproglen);
786			break;
787		}
788		if (proglen == oldproglen) {
789			image = module_alloc(proglen);
790			if (!image)
791				goto out;
792		}
793		oldproglen = proglen;
794	}
795
796	if (bpf_jit_enable > 1)
797		bpf_jit_dump(flen, proglen, pass, image);
798
799	if (image) {
800		bpf_flush_icache(image, image + proglen);
801		fp->bpf_func = (void *)image;
802		fp->jited = true;
803	}
804out:
805	kfree(addrs);
806	return;
807}
808
809void bpf_jit_free(struct bpf_prog *fp)
810{
811	if (fp->jited)
812		module_memfree(fp->bpf_func);
813
814	bpf_prog_unlock_free(fp);
815}
816