aboutsummaryrefslogtreecommitdiffstats
path: root/arch/s390/net
diff options
context:
space:
mode:
authorMartin Schwidefsky <schwidefsky@de.ibm.com>2012-07-31 10:23:59 -0400
committerMartin Schwidefsky <schwidefsky@de.ibm.com>2012-09-26 09:44:49 -0400
commitc10302efe569bfd646b4c22df29577a4595b4580 (patch)
treefbfba0e5a1cab17411f6d2084477abbdd9e9e547 /arch/s390/net
parent6f0f9b6b3fcfe5e156f20d4f804f0d505c750b3c (diff)
s390/bpf,jit: BPF Just In Time compiler for s390
The s390 implementation of the JIT compiler for packet filter speedup. Reviewed-by: Heiko Carstens <heiko.carstens@de.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Diffstat (limited to 'arch/s390/net')
-rw-r--r--arch/s390/net/Makefile4
-rw-r--r--arch/s390/net/bpf_jit.S130
-rw-r--r--arch/s390/net/bpf_jit_comp.c738
3 files changed, 872 insertions, 0 deletions
diff --git a/arch/s390/net/Makefile b/arch/s390/net/Makefile
new file mode 100644
index 000000000000..90568c33ddb0
--- /dev/null
+++ b/arch/s390/net/Makefile
@@ -0,0 +1,4 @@
1#
2# Arch-specific network modules
3#
4obj-$(CONFIG_BPF_JIT) += bpf_jit.o bpf_jit_comp.o
diff --git a/arch/s390/net/bpf_jit.S b/arch/s390/net/bpf_jit.S
new file mode 100644
index 000000000000..7e45d13816c1
--- /dev/null
+++ b/arch/s390/net/bpf_jit.S
@@ -0,0 +1,130 @@
1/*
2 * BPF Jit compiler for s390, help functions.
3 *
4 * Copyright IBM Corp. 2012
5 *
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
7 */
8#include <linux/linkage.h>
9
10/*
11 * Calling convention:
12 * registers %r2, %r6-%r8, %r10-%r11, %r13, %r15 are call saved
13 * %r2: skb pointer
14 * %r3: offset parameter
15 * %r5: BPF A accumulator
16 * %r8: return address
17 * %r9: save register for skb pointer
18 * %r10: skb->data
19 * %r11: skb->len - skb->data_len (headlen)
20 * %r12: BPF X accumulator
21 *
22 * skb_copy_bits takes 4 parameters:
23 * %r2 = skb pointer
24 * %r3 = offset into skb data
25 * %r4 = length to copy
26 * %r5 = pointer to temp buffer
27 */
28#define SKBDATA %r8
29
30 /* A = *(u32 *) (skb->data+K+X) */
31ENTRY(sk_load_word_ind)
32 ar %r3,%r12 # offset += X
33 bmr %r8 # < 0 -> return with cc
34
35 /* A = *(u32 *) (skb->data+K) */
36ENTRY(sk_load_word)
37 llgfr %r1,%r3 # extend offset
38 ahi %r3,4 # offset + 4
39 clr %r11,%r3 # hlen <= offset + 4 ?
40 jl sk_load_word_slow
41 l %r5,0(%r1,%r10) # get word from skb
42 xr %r1,%r1 # set cc to zero
43 br %r8
44
45sk_load_word_slow:
46 lgr %r9,%r2 # save %r2
47 lhi %r4,4 # 4 bytes
48 la %r5,160(%r15) # pointer to temp buffer
49 brasl %r14,skb_copy_bits # get data from skb
50 l %r5,160(%r15) # load result from temp buffer
51 ltgr %r2,%r2 # set cc to (%r2 != 0)
52 lgr %r2,%r9 # restore %r2
53 br %r8
54
55 /* A = *(u16 *) (skb->data+K+X) */
56ENTRY(sk_load_half_ind)
57 ar %r3,%r12 # offset += X
58 bmr %r8 # < 0 -> return with cc
59
60 /* A = *(u16 *) (skb->data+K) */
61ENTRY(sk_load_half)
62 llgfr %r1,%r3 # extend offset
63 ahi %r3,2 # offset + 2
64 clr %r11,%r3 # hlen <= offset + 2 ?
65 jl sk_load_half_slow
66 llgh %r5,0(%r1,%r10) # get half from skb
67 xr %r1,%r1 # set cc to zero
68 br %r8
69
70sk_load_half_slow:
71 lgr %r9,%r2 # save %r2
72 lhi %r4,2 # 2 bytes
73 la %r5,162(%r15) # pointer to temp buffer
74 brasl %r14,skb_copy_bits # get data from skb
75 xc 160(2,%r15),160(%r15)
76 l %r5,160(%r15) # load result from temp buffer
77 ltgr %r2,%r2 # set cc to (%r2 != 0)
78 lgr %r2,%r9 # restore %r2
79 br %r8
80
81 /* A = *(u8 *) (skb->data+K+X) */
82ENTRY(sk_load_byte_ind)
83 ar %r3,%r12 # offset += X
84 bmr %r8 # < 0 -> return with cc
85
86 /* A = *(u8 *) (skb->data+K) */
87ENTRY(sk_load_byte)
88 llgfr %r1,%r3 # extend offset
89 clr %r11,%r3 # hlen < offset ?
90 jle sk_load_byte_slow
91 lhi %r5,0
92 ic %r5,0(%r1,%r10) # get byte from skb
93 xr %r1,%r1 # set cc to zero
94 br %r8
95
96sk_load_byte_slow:
97 lgr %r9,%r2 # save %r2
98 lhi %r4,1 # 1 bytes
99 la %r5,163(%r15) # pointer to temp buffer
100 brasl %r14,skb_copy_bits # get data from skb
101 xc 160(3,%r15),160(%r15)
102 l %r5,160(%r15) # load result from temp buffer
103 ltgr %r2,%r2 # set cc to (%r2 != 0)
104 lgr %r2,%r9 # restore %r2
105 br %r8
106
107 /* A = (*(u8 *)(skb->data+K) & 0xf) << 2 */
108ENTRY(sk_load_byte_msh)
109 llgfr %r1,%r3 # extend offset
110 clr %r11,%r3 # hlen < offset ?
111 jle sk_load_byte_slow
112 lhi %r12,0
113 ic %r12,0(%r1,%r10) # get byte from skb
114 nill %r12,0x0f
115 sll %r12,2
116 xr %r1,%r1 # set cc to zero
117 br %r8
118
119sk_load_byte_msh_slow:
120 lgr %r9,%r2 # save %r2
121 lhi %r4,2 # 2 bytes
122 la %r5,162(%r15) # pointer to temp buffer
123 brasl %r14,skb_copy_bits # get data from skb
124 xc 160(3,%r15),160(%r15)
125 l %r12,160(%r15) # load result from temp buffer
126 nill %r12,0x0f
127 sll %r12,2
128 ltgr %r2,%r2 # set cc to (%r2 != 0)
129 lgr %r2,%r9 # restore %r2
130 br %r8
diff --git a/arch/s390/net/bpf_jit_comp.c b/arch/s390/net/bpf_jit_comp.c
new file mode 100644
index 000000000000..b07aa3dc5eed
--- /dev/null
+++ b/arch/s390/net/bpf_jit_comp.c
@@ -0,0 +1,738 @@
1/*
2 * BPF Jit compiler for s390.
3 *
4 * Copyright IBM Corp. 2012
5 *
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
7 */
8#include <linux/moduleloader.h>
9#include <linux/netdevice.h>
10#include <linux/filter.h>
11#include <asm/cacheflush.h>
12#include <asm/processor.h>
13
14/*
15 * Conventions:
16 * %r2 = skb pointer
17 * %r3 = offset parameter
18 * %r4 = scratch register / length parameter
19 * %r5 = BPF A accumulator
20 * %r8 = return address
21 * %r9 = save register for skb pointer
22 * %r10 = skb->data
23 * %r11 = skb->len - skb->data_len (headlen)
24 * %r12 = BPF X accumulator
25 * %r13 = literal pool pointer
26 * 0(%r15) - 63(%r15) scratch memory array with BPF_MEMWORDS
27 */
28int bpf_jit_enable __read_mostly;
29
30/*
31 * assembly code in arch/x86/net/bpf_jit.S
32 */
33extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[];
34extern u8 sk_load_word_ind[], sk_load_half_ind[], sk_load_byte_ind[];
35
36struct bpf_jit {
37 unsigned int seen;
38 u8 *start;
39 u8 *prg;
40 u8 *mid;
41 u8 *lit;
42 u8 *end;
43 u8 *base_ip;
44 u8 *ret0_ip;
45 u8 *exit_ip;
46 unsigned int off_load_word;
47 unsigned int off_load_half;
48 unsigned int off_load_byte;
49 unsigned int off_load_bmsh;
50 unsigned int off_load_iword;
51 unsigned int off_load_ihalf;
52 unsigned int off_load_ibyte;
53};
54
55#define BPF_SIZE_MAX 4096 /* Max size for program */
56
57#define SEEN_DATAREF 1 /* might call external helpers */
58#define SEEN_XREG 2 /* ebx is used */
59#define SEEN_MEM 4 /* use mem[] for temporary storage */
60#define SEEN_RET0 8 /* pc_ret0 points to a valid return 0 */
61#define SEEN_LITERAL 16 /* code uses literals */
62#define SEEN_LOAD_WORD 32 /* code uses sk_load_word */
63#define SEEN_LOAD_HALF 64 /* code uses sk_load_half */
64#define SEEN_LOAD_BYTE 128 /* code uses sk_load_byte */
65#define SEEN_LOAD_BMSH 256 /* code uses sk_load_byte_msh */
66#define SEEN_LOAD_IWORD 512 /* code uses sk_load_word_ind */
67#define SEEN_LOAD_IHALF 1024 /* code uses sk_load_half_ind */
68#define SEEN_LOAD_IBYTE 2048 /* code uses sk_load_byte_ind */
69
70#define EMIT2(op) \
71({ \
72 if (jit->prg + 2 <= jit->mid) \
73 *(u16 *) jit->prg = op; \
74 jit->prg += 2; \
75})
76
77#define EMIT4(op) \
78({ \
79 if (jit->prg + 4 <= jit->mid) \
80 *(u32 *) jit->prg = op; \
81 jit->prg += 4; \
82})
83
84#define EMIT4_DISP(op, disp) \
85({ \
86 unsigned int __disp = (disp) & 0xfff; \
87 EMIT4(op | __disp); \
88})
89
90#define EMIT4_IMM(op, imm) \
91({ \
92 unsigned int __imm = (imm) & 0xffff; \
93 EMIT4(op | __imm); \
94})
95
96#define EMIT4_PCREL(op, pcrel) \
97({ \
98 long __pcrel = ((pcrel) >> 1) & 0xffff; \
99 EMIT4(op | __pcrel); \
100})
101
102#define EMIT6(op1, op2) \
103({ \
104 if (jit->prg + 6 <= jit->mid) { \
105 *(u32 *) jit->prg = op1; \
106 *(u16 *) (jit->prg + 4) = op2; \
107 } \
108 jit->prg += 6; \
109})
110
111#define EMIT6_DISP(op1, op2, disp) \
112({ \
113 unsigned int __disp = (disp) & 0xfff; \
114 EMIT6(op1 | __disp, op2); \
115})
116
117#define EMIT_CONST(val) \
118({ \
119 unsigned int ret; \
120 ret = (unsigned int) (jit->lit - jit->base_ip); \
121 jit->seen |= SEEN_LITERAL; \
122 if (jit->lit + 4 <= jit->end) \
123 *(u32 *) jit->lit = val; \
124 jit->lit += 4; \
125 ret; \
126})
127
128#define EMIT_FN_CONST(bit, fn) \
129({ \
130 unsigned int ret; \
131 ret = (unsigned int) (jit->lit - jit->base_ip); \
132 if (jit->seen & bit) { \
133 jit->seen |= SEEN_LITERAL; \
134 if (jit->lit + 8 <= jit->end) \
135 *(void **) jit->lit = fn; \
136 jit->lit += 8; \
137 } \
138 ret; \
139})
140
141static void bpf_jit_prologue(struct bpf_jit *jit)
142{
143 /* Save registers and create stack frame if necessary */
144 if (jit->seen & SEEN_DATAREF) {
145 /* stmg %r8,%r15,88(%r15) */
146 EMIT6(0xeb8ff058, 0x0024);
147 /* lgr %r14,%r15 */
148 EMIT4(0xb90400ef);
149 /* ahi %r15,<offset> */
150 EMIT4_IMM(0xa7fa0000, (jit->seen & SEEN_MEM) ? -112 : -80);
151 /* stg %r14,152(%r15) */
152 EMIT6(0xe3e0f098, 0x0024);
153 } else if ((jit->seen & SEEN_XREG) && (jit->seen & SEEN_LITERAL))
154 /* stmg %r12,%r13,120(%r15) */
155 EMIT6(0xebcdf078, 0x0024);
156 else if (jit->seen & SEEN_XREG)
157 /* stg %r12,120(%r15) */
158 EMIT6(0xe3c0f078, 0x0024);
159 else if (jit->seen & SEEN_LITERAL)
160 /* stg %r13,128(%r15) */
161 EMIT6(0xe3d0f080, 0x0024);
162
163 /* Setup literal pool */
164 if (jit->seen & SEEN_LITERAL) {
165 /* basr %r13,0 */
166 EMIT2(0x0dd0);
167 jit->base_ip = jit->prg;
168 }
169 jit->off_load_word = EMIT_FN_CONST(SEEN_LOAD_WORD, sk_load_word);
170 jit->off_load_half = EMIT_FN_CONST(SEEN_LOAD_HALF, sk_load_half);
171 jit->off_load_byte = EMIT_FN_CONST(SEEN_LOAD_BYTE, sk_load_byte);
172 jit->off_load_bmsh = EMIT_FN_CONST(SEEN_LOAD_BMSH, sk_load_byte_msh);
173 jit->off_load_iword = EMIT_FN_CONST(SEEN_LOAD_IWORD, sk_load_word_ind);
174 jit->off_load_ihalf = EMIT_FN_CONST(SEEN_LOAD_IHALF, sk_load_half_ind);
175 jit->off_load_ibyte = EMIT_FN_CONST(SEEN_LOAD_IBYTE, sk_load_byte_ind);
176
177 /* Filter needs to access skb data */
178 if (jit->seen & SEEN_DATAREF) {
179 /* l %r11,<len>(%r2) */
180 EMIT4_DISP(0x58b02000, offsetof(struct sk_buff, len));
181 /* s %r11,<data_len>(%r2) */
182 EMIT4_DISP(0x5bb02000, offsetof(struct sk_buff, data_len));
183 /* lg %r10,<data>(%r2) */
184 EMIT6_DISP(0xe3a02000, 0x0004,
185 offsetof(struct sk_buff, data));
186 }
187}
188
189static void bpf_jit_epilogue(struct bpf_jit *jit)
190{
191 /* Return 0 */
192 if (jit->seen & SEEN_RET0) {
193 jit->ret0_ip = jit->prg;
194 /* lghi %r2,0 */
195 EMIT4(0xa7290000);
196 }
197 jit->exit_ip = jit->prg;
198 /* Restore registers */
199 if (jit->seen & SEEN_DATAREF)
200 /* lmg %r8,%r15,<offset>(%r15) */
201 EMIT6_DISP(0xeb8ff000, 0x0004,
202 (jit->seen & SEEN_MEM) ? 200 : 168);
203 else if ((jit->seen & SEEN_XREG) && (jit->seen & SEEN_LITERAL))
204 /* lmg %r12,%r13,120(%r15) */
205 EMIT6(0xebcdf078, 0x0004);
206 else if (jit->seen & SEEN_XREG)
207 /* lg %r12,120(%r15) */
208 EMIT6(0xe3c0f078, 0x0004);
209 else if (jit->seen & SEEN_LITERAL)
210 /* lg %r13,128(%r15) */
211 EMIT6(0xe3d0f080, 0x0004);
212 /* br %r14 */
213 EMIT2(0x07fe);
214}
215
216/*
217 * make sure we dont leak kernel information to user
218 */
219static void bpf_jit_noleaks(struct bpf_jit *jit, struct sock_filter *filter)
220{
221 /* Clear temporary memory if (seen & SEEN_MEM) */
222 if (jit->seen & SEEN_MEM)
223 /* xc 0(64,%r15),0(%r15) */
224 EMIT6(0xd73ff000, 0xf000);
225 /* Clear X if (seen & SEEN_XREG) */
226 if (jit->seen & SEEN_XREG)
227 /* lhi %r12,0 */
228 EMIT4(0xa7c80000);
229 /* Clear A if the first register does not set it. */
230 switch (filter[0].code) {
231 case BPF_S_LD_W_ABS:
232 case BPF_S_LD_H_ABS:
233 case BPF_S_LD_B_ABS:
234 case BPF_S_LD_W_LEN:
235 case BPF_S_LD_W_IND:
236 case BPF_S_LD_H_IND:
237 case BPF_S_LD_B_IND:
238 case BPF_S_LDX_B_MSH:
239 case BPF_S_LD_IMM:
240 case BPF_S_LD_MEM:
241 case BPF_S_MISC_TXA:
242 case BPF_S_ANC_PROTOCOL:
243 case BPF_S_ANC_PKTTYPE:
244 case BPF_S_ANC_IFINDEX:
245 case BPF_S_ANC_MARK:
246 case BPF_S_ANC_QUEUE:
247 case BPF_S_ANC_HATYPE:
248 case BPF_S_ANC_RXHASH:
249 case BPF_S_ANC_CPU:
250 case BPF_S_RET_K:
251 /* first instruction sets A register */
252 break;
253 default: /* A = 0 */
254 /* lhi %r5,0 */
255 EMIT4(0xa7580000);
256 }
257}
258
259static int bpf_jit_insn(struct bpf_jit *jit, struct sock_filter *filter,
260 unsigned int *addrs, int i, int last)
261{
262 unsigned int K;
263 int offset;
264 unsigned int mask;
265
266 K = filter->k;
267 switch (filter->code) {
268 case BPF_S_ALU_ADD_X: /* A += X */
269 jit->seen |= SEEN_XREG;
270 /* ar %r5,%r12 */
271 EMIT2(0x1a5c);
272 break;
273 case BPF_S_ALU_ADD_K: /* A += K */
274 if (!K)
275 break;
276 if (K <= 16383)
277 /* ahi %r5,<K> */
278 EMIT4_IMM(0xa75a0000, K);
279 else
280 /* a %r5,<d(K)>(%r13) */
281 EMIT4_DISP(0x5a50d000, EMIT_CONST(K));
282 break;
283 case BPF_S_ALU_SUB_X: /* A -= X */
284 jit->seen |= SEEN_XREG;
285 /* sr %r5,%r12 */
286 EMIT2(0x1b5c);
287 break;
288 case BPF_S_ALU_SUB_K: /* A -= K */
289 if (!K)
290 break;
291 if (K <= 16384)
292 /* ahi %r5,-K */
293 EMIT4_IMM(0xa75a0000, -K);
294 else
295 /* s %r5,<d(K)>(%r13) */
296 EMIT4_DISP(0x5b50d000, EMIT_CONST(K));
297 break;
298 case BPF_S_ALU_MUL_X: /* A *= X */
299 jit->seen |= SEEN_XREG;
300 /* msr %r5,%r12 */
301 EMIT4(0xb252005c);
302 break;
303 case BPF_S_ALU_MUL_K: /* A *= K */
304 if (K <= 16383)
305 /* mhi %r5,K */
306 EMIT4_IMM(0xa75c0000, K);
307 else
308 /* ms %r5,<d(K)>(%r13) */
309 EMIT4_DISP(0x7150d000, EMIT_CONST(K));
310 break;
311 case BPF_S_ALU_DIV_X: /* A /= X */
312 jit->seen |= SEEN_XREG | SEEN_RET0;
313 /* ltr %r12,%r12 */
314 EMIT2(0x12cc);
315 /* jz <ret0> */
316 EMIT4_PCREL(0xa7840000, (jit->ret0_ip - jit->prg));
317 /* lhi %r4,0 */
318 EMIT4(0xa7480000);
319 /* dr %r4,%r12 */
320 EMIT2(0x1d4c);
321 break;
322 case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K) */
323 /* m %r4,<d(K)>(%r13) */
324 EMIT4_DISP(0x5c40d000, EMIT_CONST(K));
325 /* lr %r5,%r4 */
326 EMIT2(0x1854);
327 break;
328 case BPF_S_ALU_AND_X: /* A &= X */
329 jit->seen |= SEEN_XREG;
330 /* nr %r5,%r12 */
331 EMIT2(0x145c);
332 break;
333 case BPF_S_ALU_AND_K: /* A &= K */
334 /* n %r5,<d(K)>(%r13) */
335 EMIT4_DISP(0x5450d000, EMIT_CONST(K));
336 break;
337 case BPF_S_ALU_OR_X: /* A |= X */
338 jit->seen |= SEEN_XREG;
339 /* or %r5,%r12 */
340 EMIT2(0x165c);
341 break;
342 case BPF_S_ALU_OR_K: /* A |= K */
343 /* o %r5,<d(K)>(%r13) */
344 EMIT4_DISP(0x5650d000, EMIT_CONST(K));
345 break;
346 case BPF_S_ALU_LSH_X: /* A <<= X; */
347 jit->seen |= SEEN_XREG;
348 /* sll %r5,0(%r12) */
349 EMIT4(0x8950c000);
350 break;
351 case BPF_S_ALU_LSH_K: /* A <<= K */
352 if (K == 0)
353 break;
354 /* sll %r5,K */
355 EMIT4_DISP(0x89500000, K);
356 break;
357 case BPF_S_ALU_RSH_X: /* A >>= X; */
358 jit->seen |= SEEN_XREG;
359 /* srl %r5,0(%r12) */
360 EMIT4(0x8850c000);
361 break;
362 case BPF_S_ALU_RSH_K: /* A >>= K; */
363 if (K == 0)
364 break;
365 /* srl %r5,K */
366 EMIT4_DISP(0x88500000, K);
367 break;
368 case BPF_S_ALU_NEG: /* A = -A */
369 /* lnr %r5,%r5 */
370 EMIT2(0x1155);
371 break;
372 case BPF_S_JMP_JA: /* ip += K */
373 offset = addrs[i + K] + jit->start - jit->prg;
374 EMIT4_PCREL(0xa7f40000, offset);
375 break;
376 case BPF_S_JMP_JGT_K: /* ip += (A > K) ? jt : jf */
377 mask = 0x200000; /* jh */
378 goto kbranch;
379 case BPF_S_JMP_JGE_K: /* ip += (A >= K) ? jt : jf */
380 mask = 0xa00000; /* jhe */
381 goto kbranch;
382 case BPF_S_JMP_JEQ_K: /* ip += (A == K) ? jt : jf */
383 mask = 0x800000; /* je */
384kbranch: /* Emit compare if the branch targets are different */
385 if (filter->jt != filter->jf) {
386 if (K <= 16383)
387 /* chi %r5,<K> */
388 EMIT4_IMM(0xa75e0000, K);
389 else
390 /* c %r5,<d(K)>(%r13) */
391 EMIT4_DISP(0x5950d000, EMIT_CONST(K));
392 }
393branch: if (filter->jt == filter->jf) {
394 if (filter->jt == 0)
395 break;
396 /* j <jt> */
397 offset = addrs[i + filter->jt] + jit->start - jit->prg;
398 EMIT4_PCREL(0xa7f40000, offset);
399 break;
400 }
401 if (filter->jt != 0) {
402 /* brc <mask>,<jt> */
403 offset = addrs[i + filter->jt] + jit->start - jit->prg;
404 EMIT4_PCREL(0xa7040000 | mask, offset);
405 }
406 if (filter->jf != 0) {
407 /* brc <mask^15>,<jf> */
408 offset = addrs[i + filter->jf] + jit->start - jit->prg;
409 EMIT4_PCREL(0xa7040000 | (mask ^ 0xf00000), offset);
410 }
411 break;
412 case BPF_S_JMP_JSET_K: /* ip += (A & K) ? jt : jf */
413 mask = 0x700000; /* jnz */
414 /* Emit test if the branch targets are different */
415 if (filter->jt != filter->jf) {
416 if (K > 65535) {
417 /* lr %r4,%r5 */
418 EMIT2(0x1845);
419 /* n %r4,<d(K)>(%r13) */
420 EMIT4_DISP(0x5440d000, EMIT_CONST(K));
421 } else
422 /* tmll %r5,K */
423 EMIT4_IMM(0xa7510000, K);
424 }
425 goto branch;
426 case BPF_S_JMP_JGT_X: /* ip += (A > X) ? jt : jf */
427 mask = 0x200000; /* jh */
428 goto xbranch;
429 case BPF_S_JMP_JGE_X: /* ip += (A >= X) ? jt : jf */
430 mask = 0xa00000; /* jhe */
431 goto xbranch;
432 case BPF_S_JMP_JEQ_X: /* ip += (A == X) ? jt : jf */
433 mask = 0x800000; /* je */
434xbranch: /* Emit compare if the branch targets are different */
435 if (filter->jt != filter->jf) {
436 jit->seen |= SEEN_XREG;
437 /* cr %r5,%r12 */
438 EMIT2(0x195c);
439 }
440 goto branch;
441 case BPF_S_JMP_JSET_X: /* ip += (A & X) ? jt : jf */
442 mask = 0x700000; /* jnz */
443 /* Emit test if the branch targets are different */
444 if (filter->jt != filter->jf) {
445 jit->seen |= SEEN_XREG;
446 /* lr %r4,%r5 */
447 EMIT2(0x1845);
448 /* nr %r4,%r12 */
449 EMIT2(0x144c);
450 }
451 goto branch;
452 case BPF_S_LD_W_ABS: /* A = *(u32 *) (skb->data+K) */
453 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_WORD;
454 offset = jit->off_load_word;
455 goto load_abs;
456 case BPF_S_LD_H_ABS: /* A = *(u16 *) (skb->data+K) */
457 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_HALF;
458 offset = jit->off_load_half;
459 goto load_abs;
460 case BPF_S_LD_B_ABS: /* A = *(u8 *) (skb->data+K) */
461 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_BYTE;
462 offset = jit->off_load_byte;
463load_abs: if ((int) K < 0)
464 goto out;
465call_fn: /* lg %r1,<d(function)>(%r13) */
466 EMIT6_DISP(0xe310d000, 0x0004, offset);
467 /* l %r3,<d(K)>(%r13) */
468 EMIT4_DISP(0x5830d000, EMIT_CONST(K));
469 /* basr %r8,%r1 */
470 EMIT2(0x0d81);
471 /* jnz <ret0> */
472 EMIT4_PCREL(0xa7740000, (jit->ret0_ip - jit->prg));
473 break;
474 case BPF_S_LD_W_IND: /* A = *(u32 *) (skb->data+K+X) */
475 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IWORD;
476 offset = jit->off_load_iword;
477 goto call_fn;
478 case BPF_S_LD_H_IND: /* A = *(u16 *) (skb->data+K+X) */
479 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IHALF;
480 offset = jit->off_load_ihalf;
481 goto call_fn;
482 case BPF_S_LD_B_IND: /* A = *(u8 *) (skb->data+K+X) */
483 jit->seen |= SEEN_DATAREF | SEEN_RET0 | SEEN_LOAD_IBYTE;
484 offset = jit->off_load_ibyte;
485 goto call_fn;
486 case BPF_S_LDX_B_MSH:
487 /* X = (*(u8 *)(skb->data+K) & 0xf) << 2 */
488 jit->seen |= SEEN_RET0;
489 if ((int) K < 0) {
490 /* j <ret0> */
491 EMIT4_PCREL(0xa7f40000, (jit->ret0_ip - jit->prg));
492 break;
493 }
494 jit->seen |= SEEN_DATAREF | SEEN_LOAD_BMSH;
495 offset = jit->off_load_bmsh;
496 goto call_fn;
497 case BPF_S_LD_W_LEN: /* A = skb->len; */
498 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
499 /* l %r5,<d(len)>(%r2) */
500 EMIT4_DISP(0x58502000, offsetof(struct sk_buff, len));
501 break;
502 case BPF_S_LDX_W_LEN: /* X = skb->len; */
503 jit->seen |= SEEN_XREG;
504 /* l %r12,<d(len)>(%r2) */
505 EMIT4_DISP(0x58c02000, offsetof(struct sk_buff, len));
506 break;
507 case BPF_S_LD_IMM: /* A = K */
508 if (K <= 16383)
509 /* lhi %r5,K */
510 EMIT4_IMM(0xa7580000, K);
511 else
512 /* l %r5,<d(K)>(%r13) */
513 EMIT4_DISP(0x5850d000, EMIT_CONST(K));
514 break;
515 case BPF_S_LDX_IMM: /* X = K */
516 jit->seen |= SEEN_XREG;
517 if (K <= 16383)
518 /* lhi %r12,<K> */
519 EMIT4_IMM(0xa7c80000, K);
520 else
521 /* l %r12,<d(K)>(%r13) */
522 EMIT4_DISP(0x58c0d000, EMIT_CONST(K));
523 break;
524 case BPF_S_LD_MEM: /* A = mem[K] */
525 jit->seen |= SEEN_MEM;
526 /* l %r5,<K>(%r15) */
527 EMIT4_DISP(0x5850f000,
528 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
529 break;
530 case BPF_S_LDX_MEM: /* X = mem[K] */
531 jit->seen |= SEEN_XREG | SEEN_MEM;
532 /* l %r12,<K>(%r15) */
533 EMIT4_DISP(0x58c0f000,
534 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
535 break;
536 case BPF_S_MISC_TAX: /* X = A */
537 jit->seen |= SEEN_XREG;
538 /* lr %r12,%r5 */
539 EMIT2(0x18c5);
540 break;
541 case BPF_S_MISC_TXA: /* A = X */
542 jit->seen |= SEEN_XREG;
543 /* lr %r5,%r12 */
544 EMIT2(0x185c);
545 break;
546 case BPF_S_RET_K:
547 if (K == 0) {
548 jit->seen |= SEEN_RET0;
549 if (last)
550 break;
551 /* j <ret0> */
552 EMIT4_PCREL(0xa7f40000, jit->ret0_ip - jit->prg);
553 } else {
554 if (K <= 16383)
555 /* lghi %r2,K */
556 EMIT4_IMM(0xa7290000, K);
557 else
558 /* llgf %r2,<K>(%r13) */
559 EMIT6_DISP(0xe320d000, 0x0016, EMIT_CONST(K));
560 /* j <exit> */
561 if (last && !(jit->seen & SEEN_RET0))
562 break;
563 EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg);
564 }
565 break;
566 case BPF_S_RET_A:
567 /* llgfr %r2,%r5 */
568 EMIT4(0xb9160025);
569 /* j <exit> */
570 EMIT4_PCREL(0xa7f40000, jit->exit_ip - jit->prg);
571 break;
572 case BPF_S_ST: /* mem[K] = A */
573 jit->seen |= SEEN_MEM;
574 /* st %r5,<K>(%r15) */
575 EMIT4_DISP(0x5050f000,
576 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
577 break;
578 case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */
579 jit->seen |= SEEN_XREG | SEEN_MEM;
580 /* st %r12,<K>(%r15) */
581 EMIT4_DISP(0x50c0f000,
582 (jit->seen & SEEN_DATAREF) ? 160 + K*4 : K*4);
583 break;
584 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
585 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
586 /* lhi %r5,0 */
587 EMIT4(0xa7580000);
588 /* icm %r5,3,<d(protocol)>(%r2) */
589 EMIT4_DISP(0xbf532000, offsetof(struct sk_buff, protocol));
590 break;
591 case BPF_S_ANC_IFINDEX: /* if (!skb->dev) return 0;
592 * A = skb->dev->ifindex */
593 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
594 jit->seen |= SEEN_RET0;
595 /* lg %r1,<d(dev)>(%r2) */
596 EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev));
597 /* ltgr %r1,%r1 */
598 EMIT4(0xb9020011);
599 /* jz <ret0> */
600 EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg);
601 /* l %r5,<d(ifindex)>(%r1) */
602 EMIT4_DISP(0x58501000, offsetof(struct net_device, ifindex));
603 break;
604 case BPF_S_ANC_MARK: /* A = skb->mark */
605 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
606 /* l %r5,<d(mark)>(%r2) */
607 EMIT4_DISP(0x58502000, offsetof(struct sk_buff, mark));
608 break;
609 case BPF_S_ANC_QUEUE: /* A = skb->queue_mapping */
610 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
611 /* lhi %r5,0 */
612 EMIT4(0xa7580000);
613 /* icm %r5,3,<d(queue_mapping)>(%r2) */
614 EMIT4_DISP(0xbf532000, offsetof(struct sk_buff, queue_mapping));
615 break;
616 case BPF_S_ANC_HATYPE: /* if (!skb->dev) return 0;
617 * A = skb->dev->type */
618 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
619 jit->seen |= SEEN_RET0;
620 /* lg %r1,<d(dev)>(%r2) */
621 EMIT6_DISP(0xe3102000, 0x0004, offsetof(struct sk_buff, dev));
622 /* ltgr %r1,%r1 */
623 EMIT4(0xb9020011);
624 /* jz <ret0> */
625 EMIT4_PCREL(0xa7840000, jit->ret0_ip - jit->prg);
626 /* lhi %r5,0 */
627 EMIT4(0xa7580000);
628 /* icm %r5,3,<d(type)>(%r1) */
629 EMIT4_DISP(0xbf531000, offsetof(struct net_device, type));
630 break;
631 case BPF_S_ANC_RXHASH: /* A = skb->rxhash */
632 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
633 /* l %r5,<d(rxhash)>(%r2) */
634 EMIT4_DISP(0x58502000, offsetof(struct sk_buff, rxhash));
635 break;
636 case BPF_S_ANC_CPU: /* A = smp_processor_id() */
637#ifdef CONFIG_SMP
638 /* l %r5,<d(cpu_nr)> */
639 EMIT4_DISP(0x58500000, offsetof(struct _lowcore, cpu_nr));
640#else
641 /* lhi %r5,0 */
642 EMIT4(0xa7580000);
643#endif
644 break;
645 default: /* too complex, give up */
646 goto out;
647 }
648 addrs[i] = jit->prg - jit->start;
649 return 0;
650out:
651 return -1;
652}
653
654void bpf_jit_compile(struct sk_filter *fp)
655{
656 unsigned long size, prg_len, lit_len;
657 struct bpf_jit jit, cjit;
658 unsigned int *addrs;
659 int pass, i;
660
661 if (!bpf_jit_enable)
662 return;
663 addrs = kmalloc(fp->len * sizeof(*addrs), GFP_KERNEL);
664 if (addrs == NULL)
665 return;
666 memset(addrs, 0, fp->len * sizeof(*addrs));
667 memset(&jit, 0, sizeof(cjit));
668 memset(&cjit, 0, sizeof(cjit));
669
670 for (pass = 0; pass < 10; pass++) {
671 jit.prg = jit.start;
672 jit.lit = jit.mid;
673
674 bpf_jit_prologue(&jit);
675 bpf_jit_noleaks(&jit, fp->insns);
676 for (i = 0; i < fp->len; i++) {
677 if (bpf_jit_insn(&jit, fp->insns + i, addrs, i,
678 i == fp->len - 1))
679 goto out;
680 }
681 bpf_jit_epilogue(&jit);
682 if (jit.start) {
683 WARN_ON(jit.prg > cjit.prg || jit.lit > cjit.lit);
684 if (memcmp(&jit, &cjit, sizeof(jit)) == 0)
685 break;
686 } else if (jit.prg == cjit.prg && jit.lit == cjit.lit) {
687 prg_len = jit.prg - jit.start;
688 lit_len = jit.lit - jit.mid;
689 size = max_t(unsigned long, prg_len + lit_len,
690 sizeof(struct work_struct));
691 if (size >= BPF_SIZE_MAX)
692 goto out;
693 jit.start = module_alloc(size);
694 if (!jit.start)
695 goto out;
696 jit.prg = jit.mid = jit.start + prg_len;
697 jit.lit = jit.end = jit.start + prg_len + lit_len;
698 jit.base_ip += (unsigned long) jit.start;
699 jit.exit_ip += (unsigned long) jit.start;
700 jit.ret0_ip += (unsigned long) jit.start;
701 }
702 cjit = jit;
703 }
704 if (bpf_jit_enable > 1) {
705 pr_err("flen=%d proglen=%lu pass=%d image=%p\n",
706 fp->len, jit.end - jit.start, pass, jit.start);
707 if (jit.start) {
708 printk(KERN_ERR "JIT code:\n");
709 print_fn_code(jit.start, jit.mid - jit.start);
710 print_hex_dump(KERN_ERR, "JIT literals:\n",
711 DUMP_PREFIX_ADDRESS, 16, 1,
712 jit.mid, jit.end - jit.mid, false);
713 }
714 }
715 if (jit.start)
716 fp->bpf_func = (void *) jit.start;
717out:
718 kfree(addrs);
719}
720
721static void jit_free_defer(struct work_struct *arg)
722{
723 module_free(NULL, arg);
724}
725
726/* run from softirq, we must use a work_struct to call
727 * module_free() from process context
728 */
729void bpf_jit_free(struct sk_filter *fp)
730{
731 struct work_struct *work;
732
733 if (fp->bpf_func == sk_run_filter)
734 return;
735 work = (struct work_struct *)fp->bpf_func;
736 INIT_WORK(work, jit_free_defer);
737 schedule_work(work);
738}
generated by cgit v1.2.2 (git 2.25.0) at 2025-01-20 15:18:48 -0500