Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * eBPF JIT compiler
4 *
5 * Copyright 2016 Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
6 * IBM Corporation
7 *
8 * Based on the powerpc classic BPF JIT compiler by Matt Evans
9 */
10#include <linux/moduleloader.h>
11#include <asm/cacheflush.h>
12#include <asm/asm-compat.h>
13#include <linux/netdevice.h>
14#include <linux/filter.h>
15#include <linux/if_vlan.h>
16#include <linux/kernel.h>
17#include <linux/memory.h>
18#include <linux/bpf.h>
19
20#include <asm/kprobes.h>
21#include <asm/code-patching.h>
22
23#include "bpf_jit.h"
24
25static void bpf_jit_fill_ill_insns(void *area, unsigned int size)
26{
27 memset32(area, BREAKPOINT_INSTRUCTION, size / 4);
28}
29
30int bpf_jit_emit_exit_insn(u32 *image, struct codegen_context *ctx, int tmp_reg, long exit_addr)
31{
32 if (!exit_addr || is_offset_in_branch_range(exit_addr - (ctx->idx * 4))) {
33 PPC_JMP(exit_addr);
34 } else if (ctx->alt_exit_addr) {
35 if (WARN_ON(!is_offset_in_branch_range((long)ctx->alt_exit_addr - (ctx->idx * 4))))
36 return -1;
37 PPC_JMP(ctx->alt_exit_addr);
38 } else {
39 ctx->alt_exit_addr = ctx->idx * 4;
40 bpf_jit_build_epilogue(image, ctx);
41 }
42
43 return 0;
44}
45
46struct powerpc_jit_data {
47 /* address of rw header */
48 struct bpf_binary_header *hdr;
49 /* address of ro final header */
50 struct bpf_binary_header *fhdr;
51 u32 *addrs;
52 u8 *fimage;
53 u32 proglen;
54 struct codegen_context ctx;
55};
56
57bool bpf_jit_needs_zext(void)
58{
59 return true;
60}
61
62struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
63{
64 u32 proglen;
65 u32 alloclen;
66 u8 *image = NULL;
67 u32 *code_base;
68 u32 *addrs;
69 struct powerpc_jit_data *jit_data;
70 struct codegen_context cgctx;
71 int pass;
72 int flen;
73 struct bpf_binary_header *fhdr = NULL;
74 struct bpf_binary_header *hdr = NULL;
75 struct bpf_prog *org_fp = fp;
76 struct bpf_prog *tmp_fp;
77 bool bpf_blinded = false;
78 bool extra_pass = false;
79 u8 *fimage = NULL;
80 u32 *fcode_base;
81 u32 extable_len;
82 u32 fixup_len;
83
84 if (!fp->jit_requested)
85 return org_fp;
86
87 tmp_fp = bpf_jit_blind_constants(org_fp);
88 if (IS_ERR(tmp_fp))
89 return org_fp;
90
91 if (tmp_fp != org_fp) {
92 bpf_blinded = true;
93 fp = tmp_fp;
94 }
95
96 jit_data = fp->aux->jit_data;
97 if (!jit_data) {
98 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
99 if (!jit_data) {
100 fp = org_fp;
101 goto out;
102 }
103 fp->aux->jit_data = jit_data;
104 }
105
106 flen = fp->len;
107 addrs = jit_data->addrs;
108 if (addrs) {
109 cgctx = jit_data->ctx;
110 /*
111 * JIT compiled to a writable location (image/code_base) first.
112 * It is then moved to the readonly final location (fimage/fcode_base)
113 * using instruction patching.
114 */
115 fimage = jit_data->fimage;
116 fhdr = jit_data->fhdr;
117 proglen = jit_data->proglen;
118 hdr = jit_data->hdr;
119 image = (void *)hdr + ((void *)fimage - (void *)fhdr);
120 extra_pass = true;
121 /* During extra pass, ensure index is reset before repopulating extable entries */
122 cgctx.exentry_idx = 0;
123 goto skip_init_ctx;
124 }
125
126 addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
127 if (addrs == NULL) {
128 fp = org_fp;
129 goto out_addrs;
130 }
131
132 memset(&cgctx, 0, sizeof(struct codegen_context));
133 bpf_jit_init_reg_mapping(&cgctx);
134
135 /* Make sure that the stack is quadword aligned. */
136 cgctx.stack_size = round_up(fp->aux->stack_depth, 16);
137
138 /* Scouting faux-generate pass 0 */
139 if (bpf_jit_build_body(fp, NULL, NULL, &cgctx, addrs, 0, false)) {
140 /* We hit something illegal or unsupported. */
141 fp = org_fp;
142 goto out_addrs;
143 }
144
145 /*
146 * If we have seen a tail call, we need a second pass.
147 * This is because bpf_jit_emit_common_epilogue() is called
148 * from bpf_jit_emit_tail_call() with a not yet stable ctx->seen.
149 * We also need a second pass if we ended up with too large
150 * a program so as to ensure BPF_EXIT branches are in range.
151 */
152 if (cgctx.seen & SEEN_TAILCALL || !is_offset_in_branch_range((long)cgctx.idx * 4)) {
153 cgctx.idx = 0;
154 if (bpf_jit_build_body(fp, NULL, NULL, &cgctx, addrs, 0, false)) {
155 fp = org_fp;
156 goto out_addrs;
157 }
158 }
159
160 bpf_jit_realloc_regs(&cgctx);
161 /*
162 * Pretend to build prologue, given the features we've seen. This will
163 * update ctgtx.idx as it pretends to output instructions, then we can
164 * calculate total size from idx.
165 */
166 bpf_jit_build_prologue(NULL, &cgctx);
167 addrs[fp->len] = cgctx.idx * 4;
168 bpf_jit_build_epilogue(NULL, &cgctx);
169
170 fixup_len = fp->aux->num_exentries * BPF_FIXUP_LEN * 4;
171 extable_len = fp->aux->num_exentries * sizeof(struct exception_table_entry);
172
173 proglen = cgctx.idx * 4;
174 alloclen = proglen + FUNCTION_DESCR_SIZE + fixup_len + extable_len;
175
176 fhdr = bpf_jit_binary_pack_alloc(alloclen, &fimage, 4, &hdr, &image,
177 bpf_jit_fill_ill_insns);
178 if (!fhdr) {
179 fp = org_fp;
180 goto out_addrs;
181 }
182
183 if (extable_len)
184 fp->aux->extable = (void *)fimage + FUNCTION_DESCR_SIZE + proglen + fixup_len;
185
186skip_init_ctx:
187 code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
188 fcode_base = (u32 *)(fimage + FUNCTION_DESCR_SIZE);
189
190 /* Code generation passes 1-2 */
191 for (pass = 1; pass < 3; pass++) {
192 /* Now build the prologue, body code & epilogue for real. */
193 cgctx.idx = 0;
194 cgctx.alt_exit_addr = 0;
195 bpf_jit_build_prologue(code_base, &cgctx);
196 if (bpf_jit_build_body(fp, code_base, fcode_base, &cgctx, addrs, pass,
197 extra_pass)) {
198 bpf_arch_text_copy(&fhdr->size, &hdr->size, sizeof(hdr->size));
199 bpf_jit_binary_pack_free(fhdr, hdr);
200 fp = org_fp;
201 goto out_addrs;
202 }
203 bpf_jit_build_epilogue(code_base, &cgctx);
204
205 if (bpf_jit_enable > 1)
206 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
207 proglen - (cgctx.idx * 4), cgctx.seen);
208 }
209
210 if (bpf_jit_enable > 1)
211 /*
212 * Note that we output the base address of the code_base
213 * rather than image, since opcodes are in code_base.
214 */
215 bpf_jit_dump(flen, proglen, pass, code_base);
216
217#ifdef CONFIG_PPC64_ELF_ABI_V1
218 /* Function descriptor nastiness: Address + TOC */
219 ((u64 *)image)[0] = (u64)fcode_base;
220 ((u64 *)image)[1] = local_paca->kernel_toc;
221#endif
222
223 fp->bpf_func = (void *)fimage;
224 fp->jited = 1;
225 fp->jited_len = proglen + FUNCTION_DESCR_SIZE;
226
227 if (!fp->is_func || extra_pass) {
228 if (bpf_jit_binary_pack_finalize(fp, fhdr, hdr)) {
229 fp = org_fp;
230 goto out_addrs;
231 }
232 bpf_prog_fill_jited_linfo(fp, addrs);
233out_addrs:
234 kfree(addrs);
235 kfree(jit_data);
236 fp->aux->jit_data = NULL;
237 } else {
238 jit_data->addrs = addrs;
239 jit_data->ctx = cgctx;
240 jit_data->proglen = proglen;
241 jit_data->fimage = fimage;
242 jit_data->fhdr = fhdr;
243 jit_data->hdr = hdr;
244 }
245
246out:
247 if (bpf_blinded)
248 bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
249
250 return fp;
251}
252
253/*
254 * The caller should check for (BPF_MODE(code) == BPF_PROBE_MEM) before calling
255 * this function, as this only applies to BPF_PROBE_MEM, for now.
256 */
257int bpf_add_extable_entry(struct bpf_prog *fp, u32 *image, u32 *fimage, int pass,
258 struct codegen_context *ctx, int insn_idx, int jmp_off,
259 int dst_reg)
260{
261 off_t offset;
262 unsigned long pc;
263 struct exception_table_entry *ex, *ex_entry;
264 u32 *fixup;
265
266 /* Populate extable entries only in the last pass */
267 if (pass != 2)
268 return 0;
269
270 if (!fp->aux->extable ||
271 WARN_ON_ONCE(ctx->exentry_idx >= fp->aux->num_exentries))
272 return -EINVAL;
273
274 /*
275 * Program is first written to image before copying to the
276 * final location (fimage). Accordingly, update in the image first.
277 * As all offsets used are relative, copying as is to the
278 * final location should be alright.
279 */
280 pc = (unsigned long)&image[insn_idx];
281 ex = (void *)fp->aux->extable - (void *)fimage + (void *)image;
282
283 fixup = (void *)ex -
284 (fp->aux->num_exentries * BPF_FIXUP_LEN * 4) +
285 (ctx->exentry_idx * BPF_FIXUP_LEN * 4);
286
287 fixup[0] = PPC_RAW_LI(dst_reg, 0);
288 if (IS_ENABLED(CONFIG_PPC32))
289 fixup[1] = PPC_RAW_LI(dst_reg - 1, 0); /* clear higher 32-bit register too */
290
291 fixup[BPF_FIXUP_LEN - 1] =
292 PPC_RAW_BRANCH((long)(pc + jmp_off) - (long)&fixup[BPF_FIXUP_LEN - 1]);
293
294 ex_entry = &ex[ctx->exentry_idx];
295
296 offset = pc - (long)&ex_entry->insn;
297 if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
298 return -ERANGE;
299 ex_entry->insn = offset;
300
301 offset = (long)fixup - (long)&ex_entry->fixup;
302 if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
303 return -ERANGE;
304 ex_entry->fixup = offset;
305
306 ctx->exentry_idx++;
307 return 0;
308}
309
310void *bpf_arch_text_copy(void *dst, void *src, size_t len)
311{
312 int err;
313
314 if (WARN_ON_ONCE(core_kernel_text((unsigned long)dst)))
315 return ERR_PTR(-EINVAL);
316
317 mutex_lock(&text_mutex);
318 err = patch_instructions(dst, src, len, false);
319 mutex_unlock(&text_mutex);
320
321 return err ? ERR_PTR(err) : dst;
322}
323
324int bpf_arch_text_invalidate(void *dst, size_t len)
325{
326 u32 insn = BREAKPOINT_INSTRUCTION;
327 int ret;
328
329 if (WARN_ON_ONCE(core_kernel_text((unsigned long)dst)))
330 return -EINVAL;
331
332 mutex_lock(&text_mutex);
333 ret = patch_instructions(dst, &insn, len, true);
334 mutex_unlock(&text_mutex);
335
336 return ret;
337}
338
339void bpf_jit_free(struct bpf_prog *fp)
340{
341 if (fp->jited) {
342 struct powerpc_jit_data *jit_data = fp->aux->jit_data;
343 struct bpf_binary_header *hdr;
344
345 /*
346 * If we fail the final pass of JIT (from jit_subprogs),
347 * the program may not be finalized yet. Call finalize here
348 * before freeing it.
349 */
350 if (jit_data) {
351 bpf_jit_binary_pack_finalize(fp, jit_data->fhdr, jit_data->hdr);
352 kvfree(jit_data->addrs);
353 kfree(jit_data);
354 }
355 hdr = bpf_jit_binary_pack_hdr(fp);
356 bpf_jit_binary_pack_free(hdr, NULL);
357 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
358 }
359
360 bpf_prog_unlock_free(fp);
361}
1/* bpf_jit_comp.c: BPF JIT compiler
2 *
3 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
4 *
5 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com)
6 * Ported to ppc32 by Denis Kirjanov <kda@linux-powerpc.org>
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License
10 * as published by the Free Software Foundation; version 2
11 * of the License.
12 */
13#include <linux/moduleloader.h>
14#include <asm/cacheflush.h>
15#include <linux/netdevice.h>
16#include <linux/filter.h>
17#include <linux/if_vlan.h>
18
19#include "bpf_jit32.h"
20
21int bpf_jit_enable __read_mostly;
22
23static inline void bpf_flush_icache(void *start, void *end)
24{
25 smp_wmb();
26 flush_icache_range((unsigned long)start, (unsigned long)end);
27}
28
29static void bpf_jit_build_prologue(struct bpf_prog *fp, u32 *image,
30 struct codegen_context *ctx)
31{
32 int i;
33 const struct sock_filter *filter = fp->insns;
34
35 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
36 /* Make stackframe */
37 if (ctx->seen & SEEN_DATAREF) {
38 /* If we call any helpers (for loads), save LR */
39 EMIT(PPC_INST_MFLR | __PPC_RT(R0));
40 PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
41
42 /* Back up non-volatile regs. */
43 PPC_BPF_STL(r_D, 1, -(REG_SZ*(32-r_D)));
44 PPC_BPF_STL(r_HL, 1, -(REG_SZ*(32-r_HL)));
45 }
46 if (ctx->seen & SEEN_MEM) {
47 /*
48 * Conditionally save regs r15-r31 as some will be used
49 * for M[] data.
50 */
51 for (i = r_M; i < (r_M+16); i++) {
52 if (ctx->seen & (1 << (i-r_M)))
53 PPC_BPF_STL(i, 1, -(REG_SZ*(32-i)));
54 }
55 }
56 PPC_BPF_STLU(1, 1, -BPF_PPC_STACKFRAME);
57 }
58
59 if (ctx->seen & SEEN_DATAREF) {
60 /*
61 * If this filter needs to access skb data,
62 * prepare r_D and r_HL:
63 * r_HL = skb->len - skb->data_len
64 * r_D = skb->data
65 */
66 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
67 data_len));
68 PPC_LWZ_OFFS(r_HL, r_skb, offsetof(struct sk_buff, len));
69 PPC_SUB(r_HL, r_HL, r_scratch1);
70 PPC_LL_OFFS(r_D, r_skb, offsetof(struct sk_buff, data));
71 }
72
73 if (ctx->seen & SEEN_XREG) {
74 /*
75 * TODO: Could also detect whether first instr. sets X and
76 * avoid this (as below, with A).
77 */
78 PPC_LI(r_X, 0);
79 }
80
81 /* make sure we dont leak kernel information to user */
82 if (bpf_needs_clear_a(&filter[0]))
83 PPC_LI(r_A, 0);
84}
85
86static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
87{
88 int i;
89
90 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
91 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
92 if (ctx->seen & SEEN_DATAREF) {
93 PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
94 PPC_MTLR(0);
95 PPC_BPF_LL(r_D, 1, -(REG_SZ*(32-r_D)));
96 PPC_BPF_LL(r_HL, 1, -(REG_SZ*(32-r_HL)));
97 }
98 if (ctx->seen & SEEN_MEM) {
99 /* Restore any saved non-vol registers */
100 for (i = r_M; i < (r_M+16); i++) {
101 if (ctx->seen & (1 << (i-r_M)))
102 PPC_BPF_LL(i, 1, -(REG_SZ*(32-i)));
103 }
104 }
105 }
106 /* The RETs have left a return value in R3. */
107
108 PPC_BLR();
109}
110
111#define CHOOSE_LOAD_FUNC(K, func) \
112 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
113
114/* Assemble the body code between the prologue & epilogue. */
115static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
116 struct codegen_context *ctx,
117 unsigned int *addrs)
118{
119 const struct sock_filter *filter = fp->insns;
120 int flen = fp->len;
121 u8 *func;
122 unsigned int true_cond;
123 int i;
124
125 /* Start of epilogue code */
126 unsigned int exit_addr = addrs[flen];
127
128 for (i = 0; i < flen; i++) {
129 unsigned int K = filter[i].k;
130 u16 code = bpf_anc_helper(&filter[i]);
131
132 /*
133 * addrs[] maps a BPF bytecode address into a real offset from
134 * the start of the body code.
135 */
136 addrs[i] = ctx->idx * 4;
137
138 switch (code) {
139 /*** ALU ops ***/
140 case BPF_ALU | BPF_ADD | BPF_X: /* A += X; */
141 ctx->seen |= SEEN_XREG;
142 PPC_ADD(r_A, r_A, r_X);
143 break;
144 case BPF_ALU | BPF_ADD | BPF_K: /* A += K; */
145 if (!K)
146 break;
147 PPC_ADDI(r_A, r_A, IMM_L(K));
148 if (K >= 32768)
149 PPC_ADDIS(r_A, r_A, IMM_HA(K));
150 break;
151 case BPF_ALU | BPF_SUB | BPF_X: /* A -= X; */
152 ctx->seen |= SEEN_XREG;
153 PPC_SUB(r_A, r_A, r_X);
154 break;
155 case BPF_ALU | BPF_SUB | BPF_K: /* A -= K */
156 if (!K)
157 break;
158 PPC_ADDI(r_A, r_A, IMM_L(-K));
159 if (K >= 32768)
160 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
161 break;
162 case BPF_ALU | BPF_MUL | BPF_X: /* A *= X; */
163 ctx->seen |= SEEN_XREG;
164 PPC_MULW(r_A, r_A, r_X);
165 break;
166 case BPF_ALU | BPF_MUL | BPF_K: /* A *= K */
167 if (K < 32768)
168 PPC_MULI(r_A, r_A, K);
169 else {
170 PPC_LI32(r_scratch1, K);
171 PPC_MULW(r_A, r_A, r_scratch1);
172 }
173 break;
174 case BPF_ALU | BPF_MOD | BPF_X: /* A %= X; */
175 case BPF_ALU | BPF_DIV | BPF_X: /* A /= X; */
176 ctx->seen |= SEEN_XREG;
177 PPC_CMPWI(r_X, 0);
178 if (ctx->pc_ret0 != -1) {
179 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
180 } else {
181 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
182 PPC_LI(r_ret, 0);
183 PPC_JMP(exit_addr);
184 }
185 if (code == (BPF_ALU | BPF_MOD | BPF_X)) {
186 PPC_DIVWU(r_scratch1, r_A, r_X);
187 PPC_MULW(r_scratch1, r_X, r_scratch1);
188 PPC_SUB(r_A, r_A, r_scratch1);
189 } else {
190 PPC_DIVWU(r_A, r_A, r_X);
191 }
192 break;
193 case BPF_ALU | BPF_MOD | BPF_K: /* A %= K; */
194 PPC_LI32(r_scratch2, K);
195 PPC_DIVWU(r_scratch1, r_A, r_scratch2);
196 PPC_MULW(r_scratch1, r_scratch2, r_scratch1);
197 PPC_SUB(r_A, r_A, r_scratch1);
198 break;
199 case BPF_ALU | BPF_DIV | BPF_K: /* A /= K */
200 if (K == 1)
201 break;
202 PPC_LI32(r_scratch1, K);
203 PPC_DIVWU(r_A, r_A, r_scratch1);
204 break;
205 case BPF_ALU | BPF_AND | BPF_X:
206 ctx->seen |= SEEN_XREG;
207 PPC_AND(r_A, r_A, r_X);
208 break;
209 case BPF_ALU | BPF_AND | BPF_K:
210 if (!IMM_H(K))
211 PPC_ANDI(r_A, r_A, K);
212 else {
213 PPC_LI32(r_scratch1, K);
214 PPC_AND(r_A, r_A, r_scratch1);
215 }
216 break;
217 case BPF_ALU | BPF_OR | BPF_X:
218 ctx->seen |= SEEN_XREG;
219 PPC_OR(r_A, r_A, r_X);
220 break;
221 case BPF_ALU | BPF_OR | BPF_K:
222 if (IMM_L(K))
223 PPC_ORI(r_A, r_A, IMM_L(K));
224 if (K >= 65536)
225 PPC_ORIS(r_A, r_A, IMM_H(K));
226 break;
227 case BPF_ANC | SKF_AD_ALU_XOR_X:
228 case BPF_ALU | BPF_XOR | BPF_X: /* A ^= X */
229 ctx->seen |= SEEN_XREG;
230 PPC_XOR(r_A, r_A, r_X);
231 break;
232 case BPF_ALU | BPF_XOR | BPF_K: /* A ^= K */
233 if (IMM_L(K))
234 PPC_XORI(r_A, r_A, IMM_L(K));
235 if (K >= 65536)
236 PPC_XORIS(r_A, r_A, IMM_H(K));
237 break;
238 case BPF_ALU | BPF_LSH | BPF_X: /* A <<= X; */
239 ctx->seen |= SEEN_XREG;
240 PPC_SLW(r_A, r_A, r_X);
241 break;
242 case BPF_ALU | BPF_LSH | BPF_K:
243 if (K == 0)
244 break;
245 else
246 PPC_SLWI(r_A, r_A, K);
247 break;
248 case BPF_ALU | BPF_RSH | BPF_X: /* A >>= X; */
249 ctx->seen |= SEEN_XREG;
250 PPC_SRW(r_A, r_A, r_X);
251 break;
252 case BPF_ALU | BPF_RSH | BPF_K: /* A >>= K; */
253 if (K == 0)
254 break;
255 else
256 PPC_SRWI(r_A, r_A, K);
257 break;
258 case BPF_ALU | BPF_NEG:
259 PPC_NEG(r_A, r_A);
260 break;
261 case BPF_RET | BPF_K:
262 PPC_LI32(r_ret, K);
263 if (!K) {
264 if (ctx->pc_ret0 == -1)
265 ctx->pc_ret0 = i;
266 }
267 /*
268 * If this isn't the very last instruction, branch to
269 * the epilogue if we've stuff to clean up. Otherwise,
270 * if there's nothing to tidy, just return. If we /are/
271 * the last instruction, we're about to fall through to
272 * the epilogue to return.
273 */
274 if (i != flen - 1) {
275 /*
276 * Note: 'seen' is properly valid only on pass
277 * #2. Both parts of this conditional are the
278 * same instruction size though, meaning the
279 * first pass will still correctly determine the
280 * code size/addresses.
281 */
282 if (ctx->seen)
283 PPC_JMP(exit_addr);
284 else
285 PPC_BLR();
286 }
287 break;
288 case BPF_RET | BPF_A:
289 PPC_MR(r_ret, r_A);
290 if (i != flen - 1) {
291 if (ctx->seen)
292 PPC_JMP(exit_addr);
293 else
294 PPC_BLR();
295 }
296 break;
297 case BPF_MISC | BPF_TAX: /* X = A */
298 PPC_MR(r_X, r_A);
299 break;
300 case BPF_MISC | BPF_TXA: /* A = X */
301 ctx->seen |= SEEN_XREG;
302 PPC_MR(r_A, r_X);
303 break;
304
305 /*** Constant loads/M[] access ***/
306 case BPF_LD | BPF_IMM: /* A = K */
307 PPC_LI32(r_A, K);
308 break;
309 case BPF_LDX | BPF_IMM: /* X = K */
310 PPC_LI32(r_X, K);
311 break;
312 case BPF_LD | BPF_MEM: /* A = mem[K] */
313 PPC_MR(r_A, r_M + (K & 0xf));
314 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
315 break;
316 case BPF_LDX | BPF_MEM: /* X = mem[K] */
317 PPC_MR(r_X, r_M + (K & 0xf));
318 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
319 break;
320 case BPF_ST: /* mem[K] = A */
321 PPC_MR(r_M + (K & 0xf), r_A);
322 ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
323 break;
324 case BPF_STX: /* mem[K] = X */
325 PPC_MR(r_M + (K & 0xf), r_X);
326 ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf));
327 break;
328 case BPF_LD | BPF_W | BPF_LEN: /* A = skb->len; */
329 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
330 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len));
331 break;
332 case BPF_LDX | BPF_W | BPF_LEN: /* X = skb->len; */
333 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
334 break;
335
336 /*** Ancillary info loads ***/
337 case BPF_ANC | SKF_AD_PROTOCOL: /* A = ntohs(skb->protocol); */
338 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
339 protocol) != 2);
340 PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff,
341 protocol));
342 break;
343 case BPF_ANC | SKF_AD_IFINDEX:
344 case BPF_ANC | SKF_AD_HATYPE:
345 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
346 ifindex) != 4);
347 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
348 type) != 2);
349 PPC_LL_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
350 dev));
351 PPC_CMPDI(r_scratch1, 0);
352 if (ctx->pc_ret0 != -1) {
353 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
354 } else {
355 /* Exit, returning 0; first pass hits here. */
356 PPC_BCC_SHORT(COND_NE, ctx->idx * 4 + 12);
357 PPC_LI(r_ret, 0);
358 PPC_JMP(exit_addr);
359 }
360 if (code == (BPF_ANC | SKF_AD_IFINDEX)) {
361 PPC_LWZ_OFFS(r_A, r_scratch1,
362 offsetof(struct net_device, ifindex));
363 } else {
364 PPC_LHZ_OFFS(r_A, r_scratch1,
365 offsetof(struct net_device, type));
366 }
367
368 break;
369 case BPF_ANC | SKF_AD_MARK:
370 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
371 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
372 mark));
373 break;
374 case BPF_ANC | SKF_AD_RXHASH:
375 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
376 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
377 hash));
378 break;
379 case BPF_ANC | SKF_AD_VLAN_TAG:
380 case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
381 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
382 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
383
384 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
385 vlan_tci));
386 if (code == (BPF_ANC | SKF_AD_VLAN_TAG)) {
387 PPC_ANDI(r_A, r_A, ~VLAN_TAG_PRESENT);
388 } else {
389 PPC_ANDI(r_A, r_A, VLAN_TAG_PRESENT);
390 PPC_SRWI(r_A, r_A, 12);
391 }
392 break;
393 case BPF_ANC | SKF_AD_QUEUE:
394 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
395 queue_mapping) != 2);
396 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
397 queue_mapping));
398 break;
399 case BPF_ANC | SKF_AD_PKTTYPE:
400 PPC_LBZ_OFFS(r_A, r_skb, PKT_TYPE_OFFSET());
401 PPC_ANDI(r_A, r_A, PKT_TYPE_MAX);
402 PPC_SRWI(r_A, r_A, 5);
403 break;
404 case BPF_ANC | SKF_AD_CPU:
405 PPC_BPF_LOAD_CPU(r_A);
406 break;
407 /*** Absolute loads from packet header/data ***/
408 case BPF_LD | BPF_W | BPF_ABS:
409 func = CHOOSE_LOAD_FUNC(K, sk_load_word);
410 goto common_load;
411 case BPF_LD | BPF_H | BPF_ABS:
412 func = CHOOSE_LOAD_FUNC(K, sk_load_half);
413 goto common_load;
414 case BPF_LD | BPF_B | BPF_ABS:
415 func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
416 common_load:
417 /* Load from [K]. */
418 ctx->seen |= SEEN_DATAREF;
419 PPC_FUNC_ADDR(r_scratch1, func);
420 PPC_MTLR(r_scratch1);
421 PPC_LI32(r_addr, K);
422 PPC_BLRL();
423 /*
424 * Helper returns 'lt' condition on error, and an
425 * appropriate return value in r3
426 */
427 PPC_BCC(COND_LT, exit_addr);
428 break;
429
430 /*** Indirect loads from packet header/data ***/
431 case BPF_LD | BPF_W | BPF_IND:
432 func = sk_load_word;
433 goto common_load_ind;
434 case BPF_LD | BPF_H | BPF_IND:
435 func = sk_load_half;
436 goto common_load_ind;
437 case BPF_LD | BPF_B | BPF_IND:
438 func = sk_load_byte;
439 common_load_ind:
440 /*
441 * Load from [X + K]. Negative offsets are tested for
442 * in the helper functions.
443 */
444 ctx->seen |= SEEN_DATAREF | SEEN_XREG;
445 PPC_FUNC_ADDR(r_scratch1, func);
446 PPC_MTLR(r_scratch1);
447 PPC_ADDI(r_addr, r_X, IMM_L(K));
448 if (K >= 32768)
449 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
450 PPC_BLRL();
451 /* If error, cr0.LT set */
452 PPC_BCC(COND_LT, exit_addr);
453 break;
454
455 case BPF_LDX | BPF_B | BPF_MSH:
456 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
457 goto common_load;
458 break;
459
460 /*** Jump and branches ***/
461 case BPF_JMP | BPF_JA:
462 if (K != 0)
463 PPC_JMP(addrs[i + 1 + K]);
464 break;
465
466 case BPF_JMP | BPF_JGT | BPF_K:
467 case BPF_JMP | BPF_JGT | BPF_X:
468 true_cond = COND_GT;
469 goto cond_branch;
470 case BPF_JMP | BPF_JGE | BPF_K:
471 case BPF_JMP | BPF_JGE | BPF_X:
472 true_cond = COND_GE;
473 goto cond_branch;
474 case BPF_JMP | BPF_JEQ | BPF_K:
475 case BPF_JMP | BPF_JEQ | BPF_X:
476 true_cond = COND_EQ;
477 goto cond_branch;
478 case BPF_JMP | BPF_JSET | BPF_K:
479 case BPF_JMP | BPF_JSET | BPF_X:
480 true_cond = COND_NE;
481 /* Fall through */
482 cond_branch:
483 /* same targets, can avoid doing the test :) */
484 if (filter[i].jt == filter[i].jf) {
485 if (filter[i].jt > 0)
486 PPC_JMP(addrs[i + 1 + filter[i].jt]);
487 break;
488 }
489
490 switch (code) {
491 case BPF_JMP | BPF_JGT | BPF_X:
492 case BPF_JMP | BPF_JGE | BPF_X:
493 case BPF_JMP | BPF_JEQ | BPF_X:
494 ctx->seen |= SEEN_XREG;
495 PPC_CMPLW(r_A, r_X);
496 break;
497 case BPF_JMP | BPF_JSET | BPF_X:
498 ctx->seen |= SEEN_XREG;
499 PPC_AND_DOT(r_scratch1, r_A, r_X);
500 break;
501 case BPF_JMP | BPF_JEQ | BPF_K:
502 case BPF_JMP | BPF_JGT | BPF_K:
503 case BPF_JMP | BPF_JGE | BPF_K:
504 if (K < 32768)
505 PPC_CMPLWI(r_A, K);
506 else {
507 PPC_LI32(r_scratch1, K);
508 PPC_CMPLW(r_A, r_scratch1);
509 }
510 break;
511 case BPF_JMP | BPF_JSET | BPF_K:
512 if (K < 32768)
513 /* PPC_ANDI is /only/ dot-form */
514 PPC_ANDI(r_scratch1, r_A, K);
515 else {
516 PPC_LI32(r_scratch1, K);
517 PPC_AND_DOT(r_scratch1, r_A,
518 r_scratch1);
519 }
520 break;
521 }
522 /* Sometimes branches are constructed "backward", with
523 * the false path being the branch and true path being
524 * a fallthrough to the next instruction.
525 */
526 if (filter[i].jt == 0)
527 /* Swap the sense of the branch */
528 PPC_BCC(true_cond ^ COND_CMP_TRUE,
529 addrs[i + 1 + filter[i].jf]);
530 else {
531 PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]);
532 if (filter[i].jf != 0)
533 PPC_JMP(addrs[i + 1 + filter[i].jf]);
534 }
535 break;
536 default:
537 /* The filter contains something cruel & unusual.
538 * We don't handle it, but also there shouldn't be
539 * anything missing from our list.
540 */
541 if (printk_ratelimit())
542 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
543 filter[i].code, i);
544 return -ENOTSUPP;
545 }
546
547 }
548 /* Set end-of-body-code address for exit. */
549 addrs[i] = ctx->idx * 4;
550
551 return 0;
552}
553
554void bpf_jit_compile(struct bpf_prog *fp)
555{
556 unsigned int proglen;
557 unsigned int alloclen;
558 u32 *image = NULL;
559 u32 *code_base;
560 unsigned int *addrs;
561 struct codegen_context cgctx;
562 int pass;
563 int flen = fp->len;
564
565 if (!bpf_jit_enable)
566 return;
567
568 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
569 if (addrs == NULL)
570 return;
571
572 /*
573 * There are multiple assembly passes as the generated code will change
574 * size as it settles down, figuring out the max branch offsets/exit
575 * paths required.
576 *
577 * The range of standard conditional branches is +/- 32Kbytes. Since
578 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
579 * finish with 8 bytes/instruction. Not feasible, so long jumps are
580 * used, distinct from short branches.
581 *
582 * Current:
583 *
584 * For now, both branch types assemble to 2 words (short branches padded
585 * with a NOP); this is less efficient, but assembly will always complete
586 * after exactly 3 passes:
587 *
588 * First pass: No code buffer; Program is "faux-generated" -- no code
589 * emitted but maximum size of output determined (and addrs[] filled
590 * in). Also, we note whether we use M[], whether we use skb data, etc.
591 * All generation choices assumed to be 'worst-case', e.g. branches all
592 * far (2 instructions), return path code reduction not available, etc.
593 *
594 * Second pass: Code buffer allocated with size determined previously.
595 * Prologue generated to support features we have seen used. Exit paths
596 * determined and addrs[] is filled in again, as code may be slightly
597 * smaller as a result.
598 *
599 * Third pass: Code generated 'for real', and branch destinations
600 * determined from now-accurate addrs[] map.
601 *
602 * Ideal:
603 *
604 * If we optimise this, near branches will be shorter. On the
605 * first assembly pass, we should err on the side of caution and
606 * generate the biggest code. On subsequent passes, branches will be
607 * generated short or long and code size will reduce. With smaller
608 * code, more branches may fall into the short category, and code will
609 * reduce more.
610 *
611 * Finally, if we see one pass generate code the same size as the
612 * previous pass we have converged and should now generate code for
613 * real. Allocating at the end will also save the memory that would
614 * otherwise be wasted by the (small) current code shrinkage.
615 * Preferably, we should do a small number of passes (e.g. 5) and if we
616 * haven't converged by then, get impatient and force code to generate
617 * as-is, even if the odd branch would be left long. The chances of a
618 * long jump are tiny with all but the most enormous of BPF filter
619 * inputs, so we should usually converge on the third pass.
620 */
621
622 cgctx.idx = 0;
623 cgctx.seen = 0;
624 cgctx.pc_ret0 = -1;
625 /* Scouting faux-generate pass 0 */
626 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
627 /* We hit something illegal or unsupported. */
628 goto out;
629
630 /*
631 * Pretend to build prologue, given the features we've seen. This will
632 * update ctgtx.idx as it pretends to output instructions, then we can
633 * calculate total size from idx.
634 */
635 bpf_jit_build_prologue(fp, 0, &cgctx);
636 bpf_jit_build_epilogue(0, &cgctx);
637
638 proglen = cgctx.idx * 4;
639 alloclen = proglen + FUNCTION_DESCR_SIZE;
640 image = module_alloc(alloclen);
641 if (!image)
642 goto out;
643
644 code_base = image + (FUNCTION_DESCR_SIZE/4);
645
646 /* Code generation passes 1-2 */
647 for (pass = 1; pass < 3; pass++) {
648 /* Now build the prologue, body code & epilogue for real. */
649 cgctx.idx = 0;
650 bpf_jit_build_prologue(fp, code_base, &cgctx);
651 bpf_jit_build_body(fp, code_base, &cgctx, addrs);
652 bpf_jit_build_epilogue(code_base, &cgctx);
653
654 if (bpf_jit_enable > 1)
655 pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
656 proglen - (cgctx.idx * 4), cgctx.seen);
657 }
658
659 if (bpf_jit_enable > 1)
660 /* Note that we output the base address of the code_base
661 * rather than image, since opcodes are in code_base.
662 */
663 bpf_jit_dump(flen, proglen, pass, code_base);
664
665 if (image) {
666 bpf_flush_icache(code_base, code_base + (proglen/4));
667#ifdef CONFIG_PPC64
668 /* Function descriptor nastiness: Address + TOC */
669 ((u64 *)image)[0] = (u64)code_base;
670 ((u64 *)image)[1] = local_paca->kernel_toc;
671#endif
672 fp->bpf_func = (void *)image;
673 fp->jited = 1;
674 }
675out:
676 kfree(addrs);
677 return;
678}
679
680void bpf_jit_free(struct bpf_prog *fp)
681{
682 if (fp->jited)
683 module_memfree(fp->bpf_func);
684
685 bpf_prog_unlock_free(fp);
686}