1 /* bpf_jit_comp.c: BPF JIT compiler for PPC64
3 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
5 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com)
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; version 2
12 #include <linux/moduleloader.h>
13 #include <asm/cacheflush.h>
14 #include <linux/netdevice.h>
15 #include <linux/filter.h>
16 #include <linux/if_vlan.h>
21 /* There are endianness assumptions herein. */
22 #error "Little-endian PPC not supported in BPF compiler"
25 int bpf_jit_enable __read_mostly
;
28 static inline void bpf_flush_icache(void *start
, void *end
)
31 flush_icache_range((unsigned long)start
, (unsigned long)end
);
34 static void bpf_jit_build_prologue(struct sk_filter
*fp
, u32
*image
,
35 struct codegen_context
*ctx
)
38 const struct sock_filter
*filter
= fp
->insns
;
40 if (ctx
->seen
& (SEEN_MEM
| SEEN_DATAREF
)) {
42 if (ctx
->seen
& SEEN_DATAREF
) {
43 /* If we call any helpers (for loads), save LR */
44 EMIT(PPC_INST_MFLR
| __PPC_RT(R0
));
47 /* Back up non-volatile regs. */
48 PPC_STD(r_D
, 1, -(8*(32-r_D
)));
49 PPC_STD(r_HL
, 1, -(8*(32-r_HL
)));
51 if (ctx
->seen
& SEEN_MEM
) {
53 * Conditionally save regs r15-r31 as some will be used
56 for (i
= r_M
; i
< (r_M
+16); i
++) {
57 if (ctx
->seen
& (1 << (i
-r_M
)))
58 PPC_STD(i
, 1, -(8*(32-i
)));
61 EMIT(PPC_INST_STDU
| __PPC_RS(R1
) | __PPC_RA(R1
) |
62 (-BPF_PPC_STACKFRAME
& 0xfffc));
65 if (ctx
->seen
& SEEN_DATAREF
) {
67 * If this filter needs to access skb data,
68 * prepare r_D and r_HL:
69 * r_HL = skb->len - skb->data_len
72 PPC_LWZ_OFFS(r_scratch1
, r_skb
, offsetof(struct sk_buff
,
74 PPC_LWZ_OFFS(r_HL
, r_skb
, offsetof(struct sk_buff
, len
));
75 PPC_SUB(r_HL
, r_HL
, r_scratch1
);
76 PPC_LD_OFFS(r_D
, r_skb
, offsetof(struct sk_buff
, data
));
79 if (ctx
->seen
& SEEN_XREG
) {
81 * TODO: Could also detect whether first instr. sets X and
82 * avoid this (as below, with A).
87 switch (filter
[0].code
) {
90 case BPF_S_ANC_PROTOCOL
:
91 case BPF_S_ANC_IFINDEX
:
93 case BPF_S_ANC_RXHASH
:
94 case BPF_S_ANC_VLAN_TAG
:
95 case BPF_S_ANC_VLAN_TAG_PRESENT
:
101 /* first instruction sets A register (or is RET 'constant') */
104 /* make sure we dont leak kernel information to user */
109 static void bpf_jit_build_epilogue(u32
*image
, struct codegen_context
*ctx
)
113 if (ctx
->seen
& (SEEN_MEM
| SEEN_DATAREF
)) {
114 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME
);
115 if (ctx
->seen
& SEEN_DATAREF
) {
118 PPC_LD(r_D
, 1, -(8*(32-r_D
)));
119 PPC_LD(r_HL
, 1, -(8*(32-r_HL
)));
121 if (ctx
->seen
& SEEN_MEM
) {
122 /* Restore any saved non-vol registers */
123 for (i
= r_M
; i
< (r_M
+16); i
++) {
124 if (ctx
->seen
& (1 << (i
-r_M
)))
125 PPC_LD(i
, 1, -(8*(32-i
)));
129 /* The RETs have left a return value in R3. */
134 #define CHOOSE_LOAD_FUNC(K, func) \
135 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
137 /* Assemble the body code between the prologue & epilogue. */
138 static int bpf_jit_build_body(struct sk_filter
*fp
, u32
*image
,
139 struct codegen_context
*ctx
,
142 const struct sock_filter
*filter
= fp
->insns
;
145 unsigned int true_cond
;
148 /* Start of epilogue code */
149 unsigned int exit_addr
= addrs
[flen
];
151 for (i
= 0; i
< flen
; i
++) {
152 unsigned int K
= filter
[i
].k
;
155 * addrs[] maps a BPF bytecode address into a real offset from
156 * the start of the body code.
158 addrs
[i
] = ctx
->idx
* 4;
160 switch (filter
[i
].code
) {
162 case BPF_S_ALU_ADD_X
: /* A += X; */
163 ctx
->seen
|= SEEN_XREG
;
164 PPC_ADD(r_A
, r_A
, r_X
);
166 case BPF_S_ALU_ADD_K
: /* A += K; */
169 PPC_ADDI(r_A
, r_A
, IMM_L(K
));
171 PPC_ADDIS(r_A
, r_A
, IMM_HA(K
));
173 case BPF_S_ALU_SUB_X
: /* A -= X; */
174 ctx
->seen
|= SEEN_XREG
;
175 PPC_SUB(r_A
, r_A
, r_X
);
177 case BPF_S_ALU_SUB_K
: /* A -= K */
180 PPC_ADDI(r_A
, r_A
, IMM_L(-K
));
182 PPC_ADDIS(r_A
, r_A
, IMM_HA(-K
));
184 case BPF_S_ALU_MUL_X
: /* A *= X; */
185 ctx
->seen
|= SEEN_XREG
;
186 PPC_MUL(r_A
, r_A
, r_X
);
188 case BPF_S_ALU_MUL_K
: /* A *= K */
190 PPC_MULI(r_A
, r_A
, K
);
192 PPC_LI32(r_scratch1
, K
);
193 PPC_MUL(r_A
, r_A
, r_scratch1
);
196 case BPF_S_ALU_DIV_X
: /* A /= X; */
197 ctx
->seen
|= SEEN_XREG
;
199 if (ctx
->pc_ret0
!= -1) {
200 PPC_BCC(COND_EQ
, addrs
[ctx
->pc_ret0
]);
203 * Exit, returning 0; first pass hits here
204 * (longer worst-case code size).
206 PPC_BCC_SHORT(COND_NE
, (ctx
->idx
*4)+12);
210 PPC_DIVWU(r_A
, r_A
, r_X
);
212 case BPF_S_ALU_DIV_K
: /* A = reciprocal_divide(A, K); */
213 PPC_LI32(r_scratch1
, K
);
214 /* Top 32 bits of 64bit result -> A */
215 PPC_MULHWU(r_A
, r_A
, r_scratch1
);
217 case BPF_S_ALU_AND_X
:
218 ctx
->seen
|= SEEN_XREG
;
219 PPC_AND(r_A
, r_A
, r_X
);
221 case BPF_S_ALU_AND_K
:
223 PPC_ANDI(r_A
, r_A
, K
);
225 PPC_LI32(r_scratch1
, K
);
226 PPC_AND(r_A
, r_A
, r_scratch1
);
230 ctx
->seen
|= SEEN_XREG
;
231 PPC_OR(r_A
, r_A
, r_X
);
235 PPC_ORI(r_A
, r_A
, IMM_L(K
));
237 PPC_ORIS(r_A
, r_A
, IMM_H(K
));
239 case BPF_S_ANC_ALU_XOR_X
:
240 case BPF_S_ALU_XOR_X
: /* A ^= X */
241 ctx
->seen
|= SEEN_XREG
;
242 PPC_XOR(r_A
, r_A
, r_X
);
244 case BPF_S_ALU_XOR_K
: /* A ^= K */
246 PPC_XORI(r_A
, r_A
, IMM_L(K
));
248 PPC_XORIS(r_A
, r_A
, IMM_H(K
));
250 case BPF_S_ALU_LSH_X
: /* A <<= X; */
251 ctx
->seen
|= SEEN_XREG
;
252 PPC_SLW(r_A
, r_A
, r_X
);
254 case BPF_S_ALU_LSH_K
:
258 PPC_SLWI(r_A
, r_A
, K
);
260 case BPF_S_ALU_RSH_X
: /* A >>= X; */
261 ctx
->seen
|= SEEN_XREG
;
262 PPC_SRW(r_A
, r_A
, r_X
);
264 case BPF_S_ALU_RSH_K
: /* A >>= K; */
268 PPC_SRWI(r_A
, r_A
, K
);
276 if (ctx
->pc_ret0
== -1)
280 * If this isn't the very last instruction, branch to
281 * the epilogue if we've stuff to clean up. Otherwise,
282 * if there's nothing to tidy, just return. If we /are/
283 * the last instruction, we're about to fall through to
284 * the epilogue to return.
288 * Note: 'seen' is properly valid only on pass
289 * #2. Both parts of this conditional are the
290 * same instruction size though, meaning the
291 * first pass will still correctly determine the
292 * code size/addresses.
309 case BPF_S_MISC_TAX
: /* X = A */
312 case BPF_S_MISC_TXA
: /* A = X */
313 ctx
->seen
|= SEEN_XREG
;
317 /*** Constant loads/M[] access ***/
318 case BPF_S_LD_IMM
: /* A = K */
321 case BPF_S_LDX_IMM
: /* X = K */
324 case BPF_S_LD_MEM
: /* A = mem[K] */
325 PPC_MR(r_A
, r_M
+ (K
& 0xf));
326 ctx
->seen
|= SEEN_MEM
| (1<<(K
& 0xf));
328 case BPF_S_LDX_MEM
: /* X = mem[K] */
329 PPC_MR(r_X
, r_M
+ (K
& 0xf));
330 ctx
->seen
|= SEEN_MEM
| (1<<(K
& 0xf));
332 case BPF_S_ST
: /* mem[K] = A */
333 PPC_MR(r_M
+ (K
& 0xf), r_A
);
334 ctx
->seen
|= SEEN_MEM
| (1<<(K
& 0xf));
336 case BPF_S_STX
: /* mem[K] = X */
337 PPC_MR(r_M
+ (K
& 0xf), r_X
);
338 ctx
->seen
|= SEEN_XREG
| SEEN_MEM
| (1<<(K
& 0xf));
340 case BPF_S_LD_W_LEN
: /* A = skb->len; */
341 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
, len
) != 4);
342 PPC_LWZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
, len
));
344 case BPF_S_LDX_W_LEN
: /* X = skb->len; */
345 PPC_LWZ_OFFS(r_X
, r_skb
, offsetof(struct sk_buff
, len
));
348 /*** Ancillary info loads ***/
350 /* None of the BPF_S_ANC* codes appear to be passed by
351 * sk_chk_filter(). The interpreter and the x86 BPF
352 * compiler implement them so we do too -- they may be
355 case BPF_S_ANC_PROTOCOL
: /* A = ntohs(skb->protocol); */
356 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
,
358 PPC_LHZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
360 /* ntohs is a NOP with BE loads. */
362 case BPF_S_ANC_IFINDEX
:
363 PPC_LD_OFFS(r_scratch1
, r_skb
, offsetof(struct sk_buff
,
365 PPC_CMPDI(r_scratch1
, 0);
366 if (ctx
->pc_ret0
!= -1) {
367 PPC_BCC(COND_EQ
, addrs
[ctx
->pc_ret0
]);
369 /* Exit, returning 0; first pass hits here. */
370 PPC_BCC_SHORT(COND_NE
, (ctx
->idx
*4)+12);
374 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device
,
376 PPC_LWZ_OFFS(r_A
, r_scratch1
,
377 offsetof(struct net_device
, ifindex
));
380 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
, mark
) != 4);
381 PPC_LWZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
384 case BPF_S_ANC_RXHASH
:
385 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
, rxhash
) != 4);
386 PPC_LWZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
389 case BPF_S_ANC_VLAN_TAG
:
390 case BPF_S_ANC_VLAN_TAG_PRESENT
:
391 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
, vlan_tci
) != 2);
392 PPC_LHZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
394 if (filter
[i
].code
== BPF_S_ANC_VLAN_TAG
)
395 PPC_ANDI(r_A
, r_A
, VLAN_VID_MASK
);
397 PPC_ANDI(r_A
, r_A
, VLAN_TAG_PRESENT
);
399 case BPF_S_ANC_QUEUE
:
400 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
,
401 queue_mapping
) != 2);
402 PPC_LHZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
409 * raw_smp_processor_id() = local_paca->paca_index
411 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct
,
413 PPC_LHZ_OFFS(r_A
, 13,
414 offsetof(struct paca_struct
, paca_index
));
420 /*** Absolute loads from packet header/data ***/
422 func
= CHOOSE_LOAD_FUNC(K
, sk_load_word
);
425 func
= CHOOSE_LOAD_FUNC(K
, sk_load_half
);
428 func
= CHOOSE_LOAD_FUNC(K
, sk_load_byte
);
431 ctx
->seen
|= SEEN_DATAREF
;
432 PPC_LI64(r_scratch1
, func
);
433 PPC_MTLR(r_scratch1
);
437 * Helper returns 'lt' condition on error, and an
438 * appropriate return value in r3
440 PPC_BCC(COND_LT
, exit_addr
);
443 /*** Indirect loads from packet header/data ***/
446 goto common_load_ind
;
449 goto common_load_ind
;
454 * Load from [X + K]. Negative offsets are tested for
455 * in the helper functions.
457 ctx
->seen
|= SEEN_DATAREF
| SEEN_XREG
;
458 PPC_LI64(r_scratch1
, func
);
459 PPC_MTLR(r_scratch1
);
460 PPC_ADDI(r_addr
, r_X
, IMM_L(K
));
462 PPC_ADDIS(r_addr
, r_addr
, IMM_HA(K
));
464 /* If error, cr0.LT set */
465 PPC_BCC(COND_LT
, exit_addr
);
468 case BPF_S_LDX_B_MSH
:
469 func
= CHOOSE_LOAD_FUNC(K
, sk_load_byte_msh
);
473 /*** Jump and branches ***/
476 PPC_JMP(addrs
[i
+ 1 + K
]);
479 case BPF_S_JMP_JGT_K
:
480 case BPF_S_JMP_JGT_X
:
483 case BPF_S_JMP_JGE_K
:
484 case BPF_S_JMP_JGE_X
:
487 case BPF_S_JMP_JEQ_K
:
488 case BPF_S_JMP_JEQ_X
:
491 case BPF_S_JMP_JSET_K
:
492 case BPF_S_JMP_JSET_X
:
496 /* same targets, can avoid doing the test :) */
497 if (filter
[i
].jt
== filter
[i
].jf
) {
498 if (filter
[i
].jt
> 0)
499 PPC_JMP(addrs
[i
+ 1 + filter
[i
].jt
]);
503 switch (filter
[i
].code
) {
504 case BPF_S_JMP_JGT_X
:
505 case BPF_S_JMP_JGE_X
:
506 case BPF_S_JMP_JEQ_X
:
507 ctx
->seen
|= SEEN_XREG
;
510 case BPF_S_JMP_JSET_X
:
511 ctx
->seen
|= SEEN_XREG
;
512 PPC_AND_DOT(r_scratch1
, r_A
, r_X
);
514 case BPF_S_JMP_JEQ_K
:
515 case BPF_S_JMP_JGT_K
:
516 case BPF_S_JMP_JGE_K
:
520 PPC_LI32(r_scratch1
, K
);
521 PPC_CMPLW(r_A
, r_scratch1
);
524 case BPF_S_JMP_JSET_K
:
526 /* PPC_ANDI is /only/ dot-form */
527 PPC_ANDI(r_scratch1
, r_A
, K
);
529 PPC_LI32(r_scratch1
, K
);
530 PPC_AND_DOT(r_scratch1
, r_A
,
535 /* Sometimes branches are constructed "backward", with
536 * the false path being the branch and true path being
537 * a fallthrough to the next instruction.
539 if (filter
[i
].jt
== 0)
540 /* Swap the sense of the branch */
541 PPC_BCC(true_cond
^ COND_CMP_TRUE
,
542 addrs
[i
+ 1 + filter
[i
].jf
]);
544 PPC_BCC(true_cond
, addrs
[i
+ 1 + filter
[i
].jt
]);
545 if (filter
[i
].jf
!= 0)
546 PPC_JMP(addrs
[i
+ 1 + filter
[i
].jf
]);
550 /* The filter contains something cruel & unusual.
551 * We don't handle it, but also there shouldn't be
552 * anything missing from our list.
554 if (printk_ratelimit())
555 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
561 /* Set end-of-body-code address for exit. */
562 addrs
[i
] = ctx
->idx
* 4;
567 void bpf_jit_compile(struct sk_filter
*fp
)
569 unsigned int proglen
;
570 unsigned int alloclen
;
574 struct codegen_context cgctx
;
581 addrs
= kzalloc((flen
+1) * sizeof(*addrs
), GFP_KERNEL
);
586 * There are multiple assembly passes as the generated code will change
587 * size as it settles down, figuring out the max branch offsets/exit
590 * The range of standard conditional branches is +/- 32Kbytes. Since
591 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
592 * finish with 8 bytes/instruction. Not feasible, so long jumps are
593 * used, distinct from short branches.
597 * For now, both branch types assemble to 2 words (short branches padded
598 * with a NOP); this is less efficient, but assembly will always complete
599 * after exactly 3 passes:
601 * First pass: No code buffer; Program is "faux-generated" -- no code
602 * emitted but maximum size of output determined (and addrs[] filled
603 * in). Also, we note whether we use M[], whether we use skb data, etc.
604 * All generation choices assumed to be 'worst-case', e.g. branches all
605 * far (2 instructions), return path code reduction not available, etc.
607 * Second pass: Code buffer allocated with size determined previously.
608 * Prologue generated to support features we have seen used. Exit paths
609 * determined and addrs[] is filled in again, as code may be slightly
610 * smaller as a result.
612 * Third pass: Code generated 'for real', and branch destinations
613 * determined from now-accurate addrs[] map.
617 * If we optimise this, near branches will be shorter. On the
618 * first assembly pass, we should err on the side of caution and
619 * generate the biggest code. On subsequent passes, branches will be
620 * generated short or long and code size will reduce. With smaller
621 * code, more branches may fall into the short category, and code will
624 * Finally, if we see one pass generate code the same size as the
625 * previous pass we have converged and should now generate code for
626 * real. Allocating at the end will also save the memory that would
627 * otherwise be wasted by the (small) current code shrinkage.
628 * Preferably, we should do a small number of passes (e.g. 5) and if we
629 * haven't converged by then, get impatient and force code to generate
630 * as-is, even if the odd branch would be left long. The chances of a
631 * long jump are tiny with all but the most enormous of BPF filter
632 * inputs, so we should usually converge on the third pass.
638 /* Scouting faux-generate pass 0 */
639 if (bpf_jit_build_body(fp
, 0, &cgctx
, addrs
))
640 /* We hit something illegal or unsupported. */
644 * Pretend to build prologue, given the features we've seen. This will
645 * update ctgtx.idx as it pretends to output instructions, then we can
646 * calculate total size from idx.
648 bpf_jit_build_prologue(fp
, 0, &cgctx
);
649 bpf_jit_build_epilogue(0, &cgctx
);
651 proglen
= cgctx
.idx
* 4;
652 alloclen
= proglen
+ FUNCTION_DESCR_SIZE
;
653 image
= module_alloc(max_t(unsigned int, alloclen
,
654 sizeof(struct work_struct
)));
658 code_base
= image
+ (FUNCTION_DESCR_SIZE
/4);
660 /* Code generation passes 1-2 */
661 for (pass
= 1; pass
< 3; pass
++) {
662 /* Now build the prologue, body code & epilogue for real. */
664 bpf_jit_build_prologue(fp
, code_base
, &cgctx
);
665 bpf_jit_build_body(fp
, code_base
, &cgctx
, addrs
);
666 bpf_jit_build_epilogue(code_base
, &cgctx
);
668 if (bpf_jit_enable
> 1)
669 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass
,
670 proglen
- (cgctx
.idx
* 4), cgctx
.seen
);
673 if (bpf_jit_enable
> 1)
674 pr_info("flen=%d proglen=%u pass=%d image=%p\n",
675 flen
, proglen
, pass
, image
);
678 if (bpf_jit_enable
> 1)
679 print_hex_dump(KERN_ERR
, "JIT code: ",
684 bpf_flush_icache(code_base
, code_base
+ (proglen
/4));
685 /* Function descriptor nastiness: Address + TOC */
686 ((u64
*)image
)[0] = (u64
)code_base
;
687 ((u64
*)image
)[1] = local_paca
->kernel_toc
;
688 fp
->bpf_func
= (void *)image
;
695 static void jit_free_defer(struct work_struct
*arg
)
697 module_free(NULL
, arg
);
700 /* run from softirq, we must use a work_struct to call
701 * module_free() from process context
703 void bpf_jit_free(struct sk_filter
*fp
)
705 if (fp
->bpf_func
!= sk_run_filter
) {
706 struct work_struct
*work
= (struct work_struct
*)fp
->bpf_func
;
708 INIT_WORK(work
, jit_free_defer
);