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>
19 /* There are endianness assumptions herein. */
20 #error "Little-endian PPC not supported in BPF compiler"
23 int bpf_jit_enable __read_mostly
;
26 static inline void bpf_flush_icache(void *start
, void *end
)
29 flush_icache_range((unsigned long)start
, (unsigned long)end
);
32 static void bpf_jit_build_prologue(struct sk_filter
*fp
, u32
*image
,
33 struct codegen_context
*ctx
)
36 const struct sock_filter
*filter
= fp
->insns
;
38 if (ctx
->seen
& (SEEN_MEM
| SEEN_DATAREF
)) {
40 if (ctx
->seen
& SEEN_DATAREF
) {
41 /* If we call any helpers (for loads), save LR */
42 EMIT(PPC_INST_MFLR
| __PPC_RT(R0
));
45 /* Back up non-volatile regs. */
46 PPC_STD(r_D
, 1, -(8*(32-r_D
)));
47 PPC_STD(r_HL
, 1, -(8*(32-r_HL
)));
49 if (ctx
->seen
& SEEN_MEM
) {
51 * Conditionally save regs r15-r31 as some will be used
54 for (i
= r_M
; i
< (r_M
+16); i
++) {
55 if (ctx
->seen
& (1 << (i
-r_M
)))
56 PPC_STD(i
, 1, -(8*(32-i
)));
59 EMIT(PPC_INST_STDU
| __PPC_RS(R1
) | __PPC_RA(R1
) |
60 (-BPF_PPC_STACKFRAME
& 0xfffc));
63 if (ctx
->seen
& SEEN_DATAREF
) {
65 * If this filter needs to access skb data,
66 * prepare r_D and r_HL:
67 * r_HL = skb->len - skb->data_len
70 PPC_LWZ_OFFS(r_scratch1
, r_skb
, offsetof(struct sk_buff
,
72 PPC_LWZ_OFFS(r_HL
, r_skb
, offsetof(struct sk_buff
, len
));
73 PPC_SUB(r_HL
, r_HL
, r_scratch1
);
74 PPC_LD_OFFS(r_D
, r_skb
, offsetof(struct sk_buff
, data
));
77 if (ctx
->seen
& SEEN_XREG
) {
79 * TODO: Could also detect whether first instr. sets X and
80 * avoid this (as below, with A).
85 switch (filter
[0].code
) {
88 case BPF_S_ANC_PROTOCOL
:
89 case BPF_S_ANC_IFINDEX
:
91 case BPF_S_ANC_RXHASH
:
97 /* first instruction sets A register (or is RET 'constant') */
100 /* make sure we dont leak kernel information to user */
105 static void bpf_jit_build_epilogue(u32
*image
, struct codegen_context
*ctx
)
109 if (ctx
->seen
& (SEEN_MEM
| SEEN_DATAREF
)) {
110 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME
);
111 if (ctx
->seen
& SEEN_DATAREF
) {
114 PPC_LD(r_D
, 1, -(8*(32-r_D
)));
115 PPC_LD(r_HL
, 1, -(8*(32-r_HL
)));
117 if (ctx
->seen
& SEEN_MEM
) {
118 /* Restore any saved non-vol registers */
119 for (i
= r_M
; i
< (r_M
+16); i
++) {
120 if (ctx
->seen
& (1 << (i
-r_M
)))
121 PPC_LD(i
, 1, -(8*(32-i
)));
125 /* The RETs have left a return value in R3. */
130 #define CHOOSE_LOAD_FUNC(K, func) \
131 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
133 /* Assemble the body code between the prologue & epilogue. */
134 static int bpf_jit_build_body(struct sk_filter
*fp
, u32
*image
,
135 struct codegen_context
*ctx
,
138 const struct sock_filter
*filter
= fp
->insns
;
141 unsigned int true_cond
;
144 /* Start of epilogue code */
145 unsigned int exit_addr
= addrs
[flen
];
147 for (i
= 0; i
< flen
; i
++) {
148 unsigned int K
= filter
[i
].k
;
151 * addrs[] maps a BPF bytecode address into a real offset from
152 * the start of the body code.
154 addrs
[i
] = ctx
->idx
* 4;
156 switch (filter
[i
].code
) {
158 case BPF_S_ALU_ADD_X
: /* A += X; */
159 ctx
->seen
|= SEEN_XREG
;
160 PPC_ADD(r_A
, r_A
, r_X
);
162 case BPF_S_ALU_ADD_K
: /* A += K; */
165 PPC_ADDI(r_A
, r_A
, IMM_L(K
));
167 PPC_ADDIS(r_A
, r_A
, IMM_HA(K
));
169 case BPF_S_ALU_SUB_X
: /* A -= X; */
170 ctx
->seen
|= SEEN_XREG
;
171 PPC_SUB(r_A
, r_A
, r_X
);
173 case BPF_S_ALU_SUB_K
: /* A -= K */
176 PPC_ADDI(r_A
, r_A
, IMM_L(-K
));
178 PPC_ADDIS(r_A
, r_A
, IMM_HA(-K
));
180 case BPF_S_ALU_MUL_X
: /* A *= X; */
181 ctx
->seen
|= SEEN_XREG
;
182 PPC_MUL(r_A
, r_A
, r_X
);
184 case BPF_S_ALU_MUL_K
: /* A *= K */
186 PPC_MULI(r_A
, r_A
, K
);
188 PPC_LI32(r_scratch1
, K
);
189 PPC_MUL(r_A
, r_A
, r_scratch1
);
192 case BPF_S_ALU_DIV_X
: /* A /= X; */
193 ctx
->seen
|= SEEN_XREG
;
195 if (ctx
->pc_ret0
!= -1) {
196 PPC_BCC(COND_EQ
, addrs
[ctx
->pc_ret0
]);
199 * Exit, returning 0; first pass hits here
200 * (longer worst-case code size).
202 PPC_BCC_SHORT(COND_NE
, (ctx
->idx
*4)+12);
206 PPC_DIVWU(r_A
, r_A
, r_X
);
208 case BPF_S_ALU_DIV_K
: /* A = reciprocal_divide(A, K); */
209 PPC_LI32(r_scratch1
, K
);
210 /* Top 32 bits of 64bit result -> A */
211 PPC_MULHWU(r_A
, r_A
, r_scratch1
);
213 case BPF_S_ALU_AND_X
:
214 ctx
->seen
|= SEEN_XREG
;
215 PPC_AND(r_A
, r_A
, r_X
);
217 case BPF_S_ALU_AND_K
:
219 PPC_ANDI(r_A
, r_A
, K
);
221 PPC_LI32(r_scratch1
, K
);
222 PPC_AND(r_A
, r_A
, r_scratch1
);
226 ctx
->seen
|= SEEN_XREG
;
227 PPC_OR(r_A
, r_A
, r_X
);
231 PPC_ORI(r_A
, r_A
, IMM_L(K
));
233 PPC_ORIS(r_A
, r_A
, IMM_H(K
));
235 case BPF_S_ALU_LSH_X
: /* A <<= X; */
236 ctx
->seen
|= SEEN_XREG
;
237 PPC_SLW(r_A
, r_A
, r_X
);
239 case BPF_S_ALU_LSH_K
:
243 PPC_SLWI(r_A
, r_A
, K
);
245 case BPF_S_ALU_RSH_X
: /* A >>= X; */
246 ctx
->seen
|= SEEN_XREG
;
247 PPC_SRW(r_A
, r_A
, r_X
);
249 case BPF_S_ALU_RSH_K
: /* A >>= K; */
253 PPC_SRWI(r_A
, r_A
, K
);
261 if (ctx
->pc_ret0
== -1)
265 * If this isn't the very last instruction, branch to
266 * the epilogue if we've stuff to clean up. Otherwise,
267 * if there's nothing to tidy, just return. If we /are/
268 * the last instruction, we're about to fall through to
269 * the epilogue to return.
273 * Note: 'seen' is properly valid only on pass
274 * #2. Both parts of this conditional are the
275 * same instruction size though, meaning the
276 * first pass will still correctly determine the
277 * code size/addresses.
294 case BPF_S_MISC_TAX
: /* X = A */
297 case BPF_S_MISC_TXA
: /* A = X */
298 ctx
->seen
|= SEEN_XREG
;
302 /*** Constant loads/M[] access ***/
303 case BPF_S_LD_IMM
: /* A = K */
306 case BPF_S_LDX_IMM
: /* X = K */
309 case BPF_S_LD_MEM
: /* A = mem[K] */
310 PPC_MR(r_A
, r_M
+ (K
& 0xf));
311 ctx
->seen
|= SEEN_MEM
| (1<<(K
& 0xf));
313 case BPF_S_LDX_MEM
: /* X = mem[K] */
314 PPC_MR(r_X
, r_M
+ (K
& 0xf));
315 ctx
->seen
|= SEEN_MEM
| (1<<(K
& 0xf));
317 case BPF_S_ST
: /* mem[K] = A */
318 PPC_MR(r_M
+ (K
& 0xf), r_A
);
319 ctx
->seen
|= SEEN_MEM
| (1<<(K
& 0xf));
321 case BPF_S_STX
: /* mem[K] = X */
322 PPC_MR(r_M
+ (K
& 0xf), r_X
);
323 ctx
->seen
|= SEEN_XREG
| SEEN_MEM
| (1<<(K
& 0xf));
325 case BPF_S_LD_W_LEN
: /* A = skb->len; */
326 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
, len
) != 4);
327 PPC_LWZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
, len
));
329 case BPF_S_LDX_W_LEN
: /* X = skb->len; */
330 PPC_LWZ_OFFS(r_X
, r_skb
, offsetof(struct sk_buff
, len
));
333 /*** Ancillary info loads ***/
335 /* None of the BPF_S_ANC* codes appear to be passed by
336 * sk_chk_filter(). The interpreter and the x86 BPF
337 * compiler implement them so we do too -- they may be
340 case BPF_S_ANC_PROTOCOL
: /* A = ntohs(skb->protocol); */
341 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
,
343 PPC_LHZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
345 /* ntohs is a NOP with BE loads. */
347 case BPF_S_ANC_IFINDEX
:
348 PPC_LD_OFFS(r_scratch1
, r_skb
, offsetof(struct sk_buff
,
350 PPC_CMPDI(r_scratch1
, 0);
351 if (ctx
->pc_ret0
!= -1) {
352 PPC_BCC(COND_EQ
, addrs
[ctx
->pc_ret0
]);
354 /* Exit, returning 0; first pass hits here. */
355 PPC_BCC_SHORT(COND_NE
, (ctx
->idx
*4)+12);
359 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device
,
361 PPC_LWZ_OFFS(r_A
, r_scratch1
,
362 offsetof(struct net_device
, ifindex
));
365 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
, mark
) != 4);
366 PPC_LWZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
369 case BPF_S_ANC_RXHASH
:
370 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
, rxhash
) != 4);
371 PPC_LWZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
374 case BPF_S_ANC_QUEUE
:
375 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff
,
376 queue_mapping
) != 2);
377 PPC_LHZ_OFFS(r_A
, r_skb
, offsetof(struct sk_buff
,
384 * raw_smp_processor_id() = local_paca->paca_index
386 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct
,
388 PPC_LHZ_OFFS(r_A
, 13,
389 offsetof(struct paca_struct
, paca_index
));
395 /*** Absolute loads from packet header/data ***/
397 func
= CHOOSE_LOAD_FUNC(K
, sk_load_word
);
400 func
= CHOOSE_LOAD_FUNC(K
, sk_load_half
);
403 func
= CHOOSE_LOAD_FUNC(K
, sk_load_byte
);
406 ctx
->seen
|= SEEN_DATAREF
;
407 PPC_LI64(r_scratch1
, func
);
408 PPC_MTLR(r_scratch1
);
412 * Helper returns 'lt' condition on error, and an
413 * appropriate return value in r3
415 PPC_BCC(COND_LT
, exit_addr
);
418 /*** Indirect loads from packet header/data ***/
421 goto common_load_ind
;
424 goto common_load_ind
;
429 * Load from [X + K]. Negative offsets are tested for
430 * in the helper functions.
432 ctx
->seen
|= SEEN_DATAREF
| SEEN_XREG
;
433 PPC_LI64(r_scratch1
, func
);
434 PPC_MTLR(r_scratch1
);
435 PPC_ADDI(r_addr
, r_X
, IMM_L(K
));
437 PPC_ADDIS(r_addr
, r_addr
, IMM_HA(K
));
439 /* If error, cr0.LT set */
440 PPC_BCC(COND_LT
, exit_addr
);
443 case BPF_S_LDX_B_MSH
:
444 func
= CHOOSE_LOAD_FUNC(K
, sk_load_byte_msh
);
448 /*** Jump and branches ***/
451 PPC_JMP(addrs
[i
+ 1 + K
]);
454 case BPF_S_JMP_JGT_K
:
455 case BPF_S_JMP_JGT_X
:
458 case BPF_S_JMP_JGE_K
:
459 case BPF_S_JMP_JGE_X
:
462 case BPF_S_JMP_JEQ_K
:
463 case BPF_S_JMP_JEQ_X
:
466 case BPF_S_JMP_JSET_K
:
467 case BPF_S_JMP_JSET_X
:
471 /* same targets, can avoid doing the test :) */
472 if (filter
[i
].jt
== filter
[i
].jf
) {
473 if (filter
[i
].jt
> 0)
474 PPC_JMP(addrs
[i
+ 1 + filter
[i
].jt
]);
478 switch (filter
[i
].code
) {
479 case BPF_S_JMP_JGT_X
:
480 case BPF_S_JMP_JGE_X
:
481 case BPF_S_JMP_JEQ_X
:
482 ctx
->seen
|= SEEN_XREG
;
485 case BPF_S_JMP_JSET_X
:
486 ctx
->seen
|= SEEN_XREG
;
487 PPC_AND_DOT(r_scratch1
, r_A
, r_X
);
489 case BPF_S_JMP_JEQ_K
:
490 case BPF_S_JMP_JGT_K
:
491 case BPF_S_JMP_JGE_K
:
495 PPC_LI32(r_scratch1
, K
);
496 PPC_CMPLW(r_A
, r_scratch1
);
499 case BPF_S_JMP_JSET_K
:
501 /* PPC_ANDI is /only/ dot-form */
502 PPC_ANDI(r_scratch1
, r_A
, K
);
504 PPC_LI32(r_scratch1
, K
);
505 PPC_AND_DOT(r_scratch1
, r_A
,
510 /* Sometimes branches are constructed "backward", with
511 * the false path being the branch and true path being
512 * a fallthrough to the next instruction.
514 if (filter
[i
].jt
== 0)
515 /* Swap the sense of the branch */
516 PPC_BCC(true_cond
^ COND_CMP_TRUE
,
517 addrs
[i
+ 1 + filter
[i
].jf
]);
519 PPC_BCC(true_cond
, addrs
[i
+ 1 + filter
[i
].jt
]);
520 if (filter
[i
].jf
!= 0)
521 PPC_JMP(addrs
[i
+ 1 + filter
[i
].jf
]);
525 /* The filter contains something cruel & unusual.
526 * We don't handle it, but also there shouldn't be
527 * anything missing from our list.
529 if (printk_ratelimit())
530 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
536 /* Set end-of-body-code address for exit. */
537 addrs
[i
] = ctx
->idx
* 4;
542 void bpf_jit_compile(struct sk_filter
*fp
)
544 unsigned int proglen
;
545 unsigned int alloclen
;
549 struct codegen_context cgctx
;
556 addrs
= kzalloc((flen
+1) * sizeof(*addrs
), GFP_KERNEL
);
561 * There are multiple assembly passes as the generated code will change
562 * size as it settles down, figuring out the max branch offsets/exit
565 * The range of standard conditional branches is +/- 32Kbytes. Since
566 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
567 * finish with 8 bytes/instruction. Not feasible, so long jumps are
568 * used, distinct from short branches.
572 * For now, both branch types assemble to 2 words (short branches padded
573 * with a NOP); this is less efficient, but assembly will always complete
574 * after exactly 3 passes:
576 * First pass: No code buffer; Program is "faux-generated" -- no code
577 * emitted but maximum size of output determined (and addrs[] filled
578 * in). Also, we note whether we use M[], whether we use skb data, etc.
579 * All generation choices assumed to be 'worst-case', e.g. branches all
580 * far (2 instructions), return path code reduction not available, etc.
582 * Second pass: Code buffer allocated with size determined previously.
583 * Prologue generated to support features we have seen used. Exit paths
584 * determined and addrs[] is filled in again, as code may be slightly
585 * smaller as a result.
587 * Third pass: Code generated 'for real', and branch destinations
588 * determined from now-accurate addrs[] map.
592 * If we optimise this, near branches will be shorter. On the
593 * first assembly pass, we should err on the side of caution and
594 * generate the biggest code. On subsequent passes, branches will be
595 * generated short or long and code size will reduce. With smaller
596 * code, more branches may fall into the short category, and code will
599 * Finally, if we see one pass generate code the same size as the
600 * previous pass we have converged and should now generate code for
601 * real. Allocating at the end will also save the memory that would
602 * otherwise be wasted by the (small) current code shrinkage.
603 * Preferably, we should do a small number of passes (e.g. 5) and if we
604 * haven't converged by then, get impatient and force code to generate
605 * as-is, even if the odd branch would be left long. The chances of a
606 * long jump are tiny with all but the most enormous of BPF filter
607 * inputs, so we should usually converge on the third pass.
613 /* Scouting faux-generate pass 0 */
614 if (bpf_jit_build_body(fp
, 0, &cgctx
, addrs
))
615 /* We hit something illegal or unsupported. */
619 * Pretend to build prologue, given the features we've seen. This will
620 * update ctgtx.idx as it pretends to output instructions, then we can
621 * calculate total size from idx.
623 bpf_jit_build_prologue(fp
, 0, &cgctx
);
624 bpf_jit_build_epilogue(0, &cgctx
);
626 proglen
= cgctx
.idx
* 4;
627 alloclen
= proglen
+ FUNCTION_DESCR_SIZE
;
628 image
= module_alloc(max_t(unsigned int, alloclen
,
629 sizeof(struct work_struct
)));
633 code_base
= image
+ (FUNCTION_DESCR_SIZE
/4);
635 /* Code generation passes 1-2 */
636 for (pass
= 1; pass
< 3; pass
++) {
637 /* Now build the prologue, body code & epilogue for real. */
639 bpf_jit_build_prologue(fp
, code_base
, &cgctx
);
640 bpf_jit_build_body(fp
, code_base
, &cgctx
, addrs
);
641 bpf_jit_build_epilogue(code_base
, &cgctx
);
643 if (bpf_jit_enable
> 1)
644 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass
,
645 proglen
- (cgctx
.idx
* 4), cgctx
.seen
);
648 if (bpf_jit_enable
> 1)
649 pr_info("flen=%d proglen=%u pass=%d image=%p\n",
650 flen
, proglen
, pass
, image
);
653 if (bpf_jit_enable
> 1)
654 print_hex_dump(KERN_ERR
, "JIT code: ",
659 bpf_flush_icache(code_base
, code_base
+ (proglen
/4));
660 /* Function descriptor nastiness: Address + TOC */
661 ((u64
*)image
)[0] = (u64
)code_base
;
662 ((u64
*)image
)[1] = local_paca
->kernel_toc
;
663 fp
->bpf_func
= (void *)image
;
670 static void jit_free_defer(struct work_struct
*arg
)
672 module_free(NULL
, arg
);
675 /* run from softirq, we must use a work_struct to call
676 * module_free() from process context
678 void bpf_jit_free(struct sk_filter
*fp
)
680 if (fp
->bpf_func
!= sk_run_filter
) {
681 struct work_struct
*work
= (struct work_struct
*)fp
->bpf_func
;
683 INIT_WORK(work
, jit_free_defer
);