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 for PPC64
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 *
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
10 * of the License.
11 */
12#include <linux/moduleloader.h>
13#include <asm/cacheflush.h>
14#include <linux/netdevice.h>
15#include <linux/filter.h>
16#include "bpf_jit.h"
17
18#ifndef __BIG_ENDIAN
19/* There are endianness assumptions herein. */
20#error "Little-endian PPC not supported in BPF compiler"
21#endif
22
23int bpf_jit_enable __read_mostly;
24
25
26static inline void bpf_flush_icache(void *start, void *end)
27{
28 smp_wmb();
29 flush_icache_range((unsigned long)start, (unsigned long)end);
30}
31
32static void bpf_jit_build_prologue(struct sk_filter *fp, u32 *image,
33 struct codegen_context *ctx)
34{
35 int i;
36 const struct sock_filter *filter = fp->insns;
37
38 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
39 /* Make stackframe */
40 if (ctx->seen & SEEN_DATAREF) {
41 /* If we call any helpers (for loads), save LR */
42 EMIT(PPC_INST_MFLR | __PPC_RT(0));
43 PPC_STD(0, 1, 16);
44
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)));
48 }
49 if (ctx->seen & SEEN_MEM) {
50 /*
51 * Conditionally save regs r15-r31 as some will be used
52 * for M[] data.
53 */
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)));
57 }
58 }
59 EMIT(PPC_INST_STDU | __PPC_RS(1) | __PPC_RA(1) |
60 (-BPF_PPC_STACKFRAME & 0xfffc));
61 }
62
63 if (ctx->seen & SEEN_DATAREF) {
64 /*
65 * If this filter needs to access skb data,
66 * prepare r_D and r_HL:
67 * r_HL = skb->len - skb->data_len
68 * r_D = skb->data
69 */
70 PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
71 data_len));
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));
75 }
76
77 if (ctx->seen & SEEN_XREG) {
78 /*
79 * TODO: Could also detect whether first instr. sets X and
80 * avoid this (as below, with A).
81 */
82 PPC_LI(r_X, 0);
83 }
84
85 switch (filter[0].code) {
86 case BPF_S_RET_K:
87 case BPF_S_LD_W_LEN:
88 case BPF_S_ANC_PROTOCOL:
89 case BPF_S_ANC_IFINDEX:
90 case BPF_S_ANC_MARK:
91 case BPF_S_ANC_RXHASH:
92 case BPF_S_ANC_CPU:
93 case BPF_S_ANC_QUEUE:
94 case BPF_S_LD_W_ABS:
95 case BPF_S_LD_H_ABS:
96 case BPF_S_LD_B_ABS:
97 /* first instruction sets A register (or is RET 'constant') */
98 break;
99 default:
100 /* make sure we dont leak kernel information to user */
101 PPC_LI(r_A, 0);
102 }
103}
104
105static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
106{
107 int i;
108
109 if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
110 PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
111 if (ctx->seen & SEEN_DATAREF) {
112 PPC_LD(0, 1, 16);
113 PPC_MTLR(0);
114 PPC_LD(r_D, 1, -(8*(32-r_D)));
115 PPC_LD(r_HL, 1, -(8*(32-r_HL)));
116 }
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)));
122 }
123 }
124 }
125 /* The RETs have left a return value in R3. */
126
127 PPC_BLR();
128}
129
130#define CHOOSE_LOAD_FUNC(K, func) \
131 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
132
133/* Assemble the body code between the prologue & epilogue. */
134static int bpf_jit_build_body(struct sk_filter *fp, u32 *image,
135 struct codegen_context *ctx,
136 unsigned int *addrs)
137{
138 const struct sock_filter *filter = fp->insns;
139 int flen = fp->len;
140 u8 *func;
141 unsigned int true_cond;
142 int i;
143
144 /* Start of epilogue code */
145 unsigned int exit_addr = addrs[flen];
146
147 for (i = 0; i < flen; i++) {
148 unsigned int K = filter[i].k;
149
150 /*
151 * addrs[] maps a BPF bytecode address into a real offset from
152 * the start of the body code.
153 */
154 addrs[i] = ctx->idx * 4;
155
156 switch (filter[i].code) {
157 /*** ALU ops ***/
158 case BPF_S_ALU_ADD_X: /* A += X; */
159 ctx->seen |= SEEN_XREG;
160 PPC_ADD(r_A, r_A, r_X);
161 break;
162 case BPF_S_ALU_ADD_K: /* A += K; */
163 if (!K)
164 break;
165 PPC_ADDI(r_A, r_A, IMM_L(K));
166 if (K >= 32768)
167 PPC_ADDIS(r_A, r_A, IMM_HA(K));
168 break;
169 case BPF_S_ALU_SUB_X: /* A -= X; */
170 ctx->seen |= SEEN_XREG;
171 PPC_SUB(r_A, r_A, r_X);
172 break;
173 case BPF_S_ALU_SUB_K: /* A -= K */
174 if (!K)
175 break;
176 PPC_ADDI(r_A, r_A, IMM_L(-K));
177 if (K >= 32768)
178 PPC_ADDIS(r_A, r_A, IMM_HA(-K));
179 break;
180 case BPF_S_ALU_MUL_X: /* A *= X; */
181 ctx->seen |= SEEN_XREG;
182 PPC_MUL(r_A, r_A, r_X);
183 break;
184 case BPF_S_ALU_MUL_K: /* A *= K */
185 if (K < 32768)
186 PPC_MULI(r_A, r_A, K);
187 else {
188 PPC_LI32(r_scratch1, K);
189 PPC_MUL(r_A, r_A, r_scratch1);
190 }
191 break;
192 case BPF_S_ALU_DIV_X: /* A /= X; */
193 ctx->seen |= SEEN_XREG;
194 PPC_CMPWI(r_X, 0);
195 if (ctx->pc_ret0 != -1) {
196 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
197 } else {
198 /*
199 * Exit, returning 0; first pass hits here
200 * (longer worst-case code size).
201 */
202 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
203 PPC_LI(r_ret, 0);
204 PPC_JMP(exit_addr);
205 }
206 PPC_DIVWU(r_A, r_A, r_X);
207 break;
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);
212 break;
213 case BPF_S_ALU_AND_X:
214 ctx->seen |= SEEN_XREG;
215 PPC_AND(r_A, r_A, r_X);
216 break;
217 case BPF_S_ALU_AND_K:
218 if (!IMM_H(K))
219 PPC_ANDI(r_A, r_A, K);
220 else {
221 PPC_LI32(r_scratch1, K);
222 PPC_AND(r_A, r_A, r_scratch1);
223 }
224 break;
225 case BPF_S_ALU_OR_X:
226 ctx->seen |= SEEN_XREG;
227 PPC_OR(r_A, r_A, r_X);
228 break;
229 case BPF_S_ALU_OR_K:
230 if (IMM_L(K))
231 PPC_ORI(r_A, r_A, IMM_L(K));
232 if (K >= 65536)
233 PPC_ORIS(r_A, r_A, IMM_H(K));
234 break;
235 case BPF_S_ALU_LSH_X: /* A <<= X; */
236 ctx->seen |= SEEN_XREG;
237 PPC_SLW(r_A, r_A, r_X);
238 break;
239 case BPF_S_ALU_LSH_K:
240 if (K == 0)
241 break;
242 else
243 PPC_SLWI(r_A, r_A, K);
244 break;
245 case BPF_S_ALU_RSH_X: /* A >>= X; */
246 ctx->seen |= SEEN_XREG;
247 PPC_SRW(r_A, r_A, r_X);
248 break;
249 case BPF_S_ALU_RSH_K: /* A >>= K; */
250 if (K == 0)
251 break;
252 else
253 PPC_SRWI(r_A, r_A, K);
254 break;
255 case BPF_S_ALU_NEG:
256 PPC_NEG(r_A, r_A);
257 break;
258 case BPF_S_RET_K:
259 PPC_LI32(r_ret, K);
260 if (!K) {
261 if (ctx->pc_ret0 == -1)
262 ctx->pc_ret0 = i;
263 }
264 /*
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.
270 */
271 if (i != flen - 1) {
272 /*
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.
278 */
279 if (ctx->seen)
280 PPC_JMP(exit_addr);
281 else
282 PPC_BLR();
283 }
284 break;
285 case BPF_S_RET_A:
286 PPC_MR(r_ret, r_A);
287 if (i != flen - 1) {
288 if (ctx->seen)
289 PPC_JMP(exit_addr);
290 else
291 PPC_BLR();
292 }
293 break;
294 case BPF_S_MISC_TAX: /* X = A */
295 PPC_MR(r_X, r_A);
296 break;
297 case BPF_S_MISC_TXA: /* A = X */
298 ctx->seen |= SEEN_XREG;
299 PPC_MR(r_A, r_X);
300 break;
301
302 /*** Constant loads/M[] access ***/
303 case BPF_S_LD_IMM: /* A = K */
304 PPC_LI32(r_A, K);
305 break;
306 case BPF_S_LDX_IMM: /* X = K */
307 PPC_LI32(r_X, K);
308 break;
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));
312 break;
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));
316 break;
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));
320 break;
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));
324 break;
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));
328 break;
329 case BPF_S_LDX_W_LEN: /* X = skb->len; */
330 PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
331 break;
332
333 /*** Ancillary info loads ***/
334
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
338 * planted in future.
339 */
340 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
341 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
342 protocol) != 2);
343 PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
344 protocol));
345 /* ntohs is a NOP with BE loads. */
346 break;
347 case BPF_S_ANC_IFINDEX:
348 PPC_LD_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
349 dev));
350 PPC_CMPDI(r_scratch1, 0);
351 if (ctx->pc_ret0 != -1) {
352 PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
353 } else {
354 /* Exit, returning 0; first pass hits here. */
355 PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
356 PPC_LI(r_ret, 0);
357 PPC_JMP(exit_addr);
358 }
359 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
360 ifindex) != 4);
361 PPC_LWZ_OFFS(r_A, r_scratch1,
362 offsetof(struct net_device, ifindex));
363 break;
364 case BPF_S_ANC_MARK:
365 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
366 PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
367 mark));
368 break;
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,
372 rxhash));
373 break;
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,
378 queue_mapping));
379 break;
380 case BPF_S_ANC_CPU:
381#ifdef CONFIG_SMP
382 /*
383 * PACA ptr is r13:
384 * raw_smp_processor_id() = local_paca->paca_index
385 */
386 BUILD_BUG_ON(FIELD_SIZEOF(struct paca_struct,
387 paca_index) != 2);
388 PPC_LHZ_OFFS(r_A, 13,
389 offsetof(struct paca_struct, paca_index));
390#else
391 PPC_LI(r_A, 0);
392#endif
393 break;
394
395 /*** Absolute loads from packet header/data ***/
396 case BPF_S_LD_W_ABS:
397 func = CHOOSE_LOAD_FUNC(K, sk_load_word);
398 goto common_load;
399 case BPF_S_LD_H_ABS:
400 func = CHOOSE_LOAD_FUNC(K, sk_load_half);
401 goto common_load;
402 case BPF_S_LD_B_ABS:
403 func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
404 common_load:
405 /* Load from [K]. */
406 ctx->seen |= SEEN_DATAREF;
407 PPC_LI64(r_scratch1, func);
408 PPC_MTLR(r_scratch1);
409 PPC_LI32(r_addr, K);
410 PPC_BLRL();
411 /*
412 * Helper returns 'lt' condition on error, and an
413 * appropriate return value in r3
414 */
415 PPC_BCC(COND_LT, exit_addr);
416 break;
417
418 /*** Indirect loads from packet header/data ***/
419 case BPF_S_LD_W_IND:
420 func = sk_load_word;
421 goto common_load_ind;
422 case BPF_S_LD_H_IND:
423 func = sk_load_half;
424 goto common_load_ind;
425 case BPF_S_LD_B_IND:
426 func = sk_load_byte;
427 common_load_ind:
428 /*
429 * Load from [X + K]. Negative offsets are tested for
430 * in the helper functions.
431 */
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));
436 if (K >= 32768)
437 PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
438 PPC_BLRL();
439 /* If error, cr0.LT set */
440 PPC_BCC(COND_LT, exit_addr);
441 break;
442
443 case BPF_S_LDX_B_MSH:
444 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
445 goto common_load;
446 break;
447
448 /*** Jump and branches ***/
449 case BPF_S_JMP_JA:
450 if (K != 0)
451 PPC_JMP(addrs[i + 1 + K]);
452 break;
453
454 case BPF_S_JMP_JGT_K:
455 case BPF_S_JMP_JGT_X:
456 true_cond = COND_GT;
457 goto cond_branch;
458 case BPF_S_JMP_JGE_K:
459 case BPF_S_JMP_JGE_X:
460 true_cond = COND_GE;
461 goto cond_branch;
462 case BPF_S_JMP_JEQ_K:
463 case BPF_S_JMP_JEQ_X:
464 true_cond = COND_EQ;
465 goto cond_branch;
466 case BPF_S_JMP_JSET_K:
467 case BPF_S_JMP_JSET_X:
468 true_cond = COND_NE;
469 /* Fall through */
470 cond_branch:
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]);
475 break;
476 }
477
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;
483 PPC_CMPLW(r_A, r_X);
484 break;
485 case BPF_S_JMP_JSET_X:
486 ctx->seen |= SEEN_XREG;
487 PPC_AND_DOT(r_scratch1, r_A, r_X);
488 break;
489 case BPF_S_JMP_JEQ_K:
490 case BPF_S_JMP_JGT_K:
491 case BPF_S_JMP_JGE_K:
492 if (K < 32768)
493 PPC_CMPLWI(r_A, K);
494 else {
495 PPC_LI32(r_scratch1, K);
496 PPC_CMPLW(r_A, r_scratch1);
497 }
498 break;
499 case BPF_S_JMP_JSET_K:
500 if (K < 32768)
501 /* PPC_ANDI is /only/ dot-form */
502 PPC_ANDI(r_scratch1, r_A, K);
503 else {
504 PPC_LI32(r_scratch1, K);
505 PPC_AND_DOT(r_scratch1, r_A,
506 r_scratch1);
507 }
508 break;
509 }
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.
513 */
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]);
518 else {
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]);
522 }
523 break;
524 default:
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.
528 */
529 if (printk_ratelimit())
530 pr_err("BPF filter opcode %04x (@%d) unsupported\n",
531 filter[i].code, i);
532 return -ENOTSUPP;
533 }
534
535 }
536 /* Set end-of-body-code address for exit. */
537 addrs[i] = ctx->idx * 4;
538
539 return 0;
540}
541
542void bpf_jit_compile(struct sk_filter *fp)
543{
544 unsigned int proglen;
545 unsigned int alloclen;
546 u32 *image = NULL;
547 u32 *code_base;
548 unsigned int *addrs;
549 struct codegen_context cgctx;
550 int pass;
551 int flen = fp->len;
552
553 if (!bpf_jit_enable)
554 return;
555
556 addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
557 if (addrs == NULL)
558 return;
559
560 /*
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
563 * paths required.
564 *
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.
569 *
570 * Current:
571 *
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:
575 *
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.
581 *
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.
586 *
587 * Third pass: Code generated 'for real', and branch destinations
588 * determined from now-accurate addrs[] map.
589 *
590 * Ideal:
591 *
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
597 * reduce more.
598 *
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.
608 */
609
610 cgctx.idx = 0;
611 cgctx.seen = 0;
612 cgctx.pc_ret0 = -1;
613 /* Scouting faux-generate pass 0 */
614 if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
615 /* We hit something illegal or unsupported. */
616 goto out;
617
618 /*
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.
622 */
623 bpf_jit_build_prologue(fp, 0, &cgctx);
624 bpf_jit_build_epilogue(0, &cgctx);
625
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)));
630 if (!image)
631 goto out;
632
633 code_base = image + (FUNCTION_DESCR_SIZE/4);
634
635 /* Code generation passes 1-2 */
636 for (pass = 1; pass < 3; pass++) {
637 /* Now build the prologue, body code & epilogue for real. */
638 cgctx.idx = 0;
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);
642
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);
646 }
647
648 if (bpf_jit_enable > 1)
649 pr_info("flen=%d proglen=%u pass=%d image=%p\n",
650 flen, proglen, pass, image);
651
652 if (image) {
653 if (bpf_jit_enable > 1)
654 print_hex_dump(KERN_ERR, "JIT code: ",
655 DUMP_PREFIX_ADDRESS,
656 16, 1, code_base,
657 proglen, false);
658
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;
664 }
665out:
666 kfree(addrs);
667 return;
668}
669
670static void jit_free_defer(struct work_struct *arg)
671{
672 module_free(NULL, arg);
673}
674
675/* run from softirq, we must use a work_struct to call
676 * module_free() from process context
677 */
678void bpf_jit_free(struct sk_filter *fp)
679{
680 if (fp->bpf_func != sk_run_filter) {
681 struct work_struct *work = (struct work_struct *)fp->bpf_func;
682
683 INIT_WORK(work, jit_free_defer);
684 schedule_work(work);
685 }
686}