xref: /openbmc/linux/arch/powerpc/net/bpf_jit_comp.c (revision f21453b0)
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 <asm/kprobes.h>
17 #include <linux/bpf.h>
18 
19 #include "bpf_jit.h"
20 
21 static void bpf_jit_fill_ill_insns(void *area, unsigned int size)
22 {
23 	memset32(area, BREAKPOINT_INSTRUCTION, size / 4);
24 }
25 
26 /* Fix the branch target addresses for subprog calls */
27 static int bpf_jit_fixup_subprog_calls(struct bpf_prog *fp, u32 *image,
28 				       struct codegen_context *ctx, u32 *addrs)
29 {
30 	const struct bpf_insn *insn = fp->insnsi;
31 	bool func_addr_fixed;
32 	u64 func_addr;
33 	u32 tmp_idx;
34 	int i, ret;
35 
36 	for (i = 0; i < fp->len; i++) {
37 		/*
38 		 * During the extra pass, only the branch target addresses for
39 		 * the subprog calls need to be fixed. All other instructions
40 		 * can left untouched.
41 		 *
42 		 * The JITed image length does not change because we already
43 		 * ensure that the JITed instruction sequence for these calls
44 		 * are of fixed length by padding them with NOPs.
45 		 */
46 		if (insn[i].code == (BPF_JMP | BPF_CALL) &&
47 		    insn[i].src_reg == BPF_PSEUDO_CALL) {
48 			ret = bpf_jit_get_func_addr(fp, &insn[i], true,
49 						    &func_addr,
50 						    &func_addr_fixed);
51 			if (ret < 0)
52 				return ret;
53 
54 			/*
55 			 * Save ctx->idx as this would currently point to the
56 			 * end of the JITed image and set it to the offset of
57 			 * the instruction sequence corresponding to the
58 			 * subprog call temporarily.
59 			 */
60 			tmp_idx = ctx->idx;
61 			ctx->idx = addrs[i] / 4;
62 			bpf_jit_emit_func_call_rel(image, ctx, func_addr);
63 
64 			/*
65 			 * Restore ctx->idx here. This is safe as the length
66 			 * of the JITed sequence remains unchanged.
67 			 */
68 			ctx->idx = tmp_idx;
69 		}
70 	}
71 
72 	return 0;
73 }
74 
75 struct powerpc64_jit_data {
76 	struct bpf_binary_header *header;
77 	u32 *addrs;
78 	u8 *image;
79 	u32 proglen;
80 	struct codegen_context ctx;
81 };
82 
83 bool bpf_jit_needs_zext(void)
84 {
85 	return true;
86 }
87 
88 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
89 {
90 	u32 proglen;
91 	u32 alloclen;
92 	u8 *image = NULL;
93 	u32 *code_base;
94 	u32 *addrs;
95 	struct powerpc64_jit_data *jit_data;
96 	struct codegen_context cgctx;
97 	int pass;
98 	int flen;
99 	struct bpf_binary_header *bpf_hdr;
100 	struct bpf_prog *org_fp = fp;
101 	struct bpf_prog *tmp_fp;
102 	bool bpf_blinded = false;
103 	bool extra_pass = false;
104 
105 	if (!fp->jit_requested)
106 		return org_fp;
107 
108 	tmp_fp = bpf_jit_blind_constants(org_fp);
109 	if (IS_ERR(tmp_fp))
110 		return org_fp;
111 
112 	if (tmp_fp != org_fp) {
113 		bpf_blinded = true;
114 		fp = tmp_fp;
115 	}
116 
117 	jit_data = fp->aux->jit_data;
118 	if (!jit_data) {
119 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
120 		if (!jit_data) {
121 			fp = org_fp;
122 			goto out;
123 		}
124 		fp->aux->jit_data = jit_data;
125 	}
126 
127 	flen = fp->len;
128 	addrs = jit_data->addrs;
129 	if (addrs) {
130 		cgctx = jit_data->ctx;
131 		image = jit_data->image;
132 		bpf_hdr = jit_data->header;
133 		proglen = jit_data->proglen;
134 		alloclen = proglen + FUNCTION_DESCR_SIZE;
135 		extra_pass = true;
136 		goto skip_init_ctx;
137 	}
138 
139 	addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
140 	if (addrs == NULL) {
141 		fp = org_fp;
142 		goto out_addrs;
143 	}
144 
145 	memset(&cgctx, 0, sizeof(struct codegen_context));
146 	memcpy(cgctx.b2p, b2p, sizeof(cgctx.b2p));
147 
148 	/* Make sure that the stack is quadword aligned. */
149 	cgctx.stack_size = round_up(fp->aux->stack_depth, 16);
150 
151 	/* Scouting faux-generate pass 0 */
152 	if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
153 		/* We hit something illegal or unsupported. */
154 		fp = org_fp;
155 		goto out_addrs;
156 	}
157 
158 	/*
159 	 * If we have seen a tail call, we need a second pass.
160 	 * This is because bpf_jit_emit_common_epilogue() is called
161 	 * from bpf_jit_emit_tail_call() with a not yet stable ctx->seen.
162 	 */
163 	if (cgctx.seen & SEEN_TAILCALL) {
164 		cgctx.idx = 0;
165 		if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
166 			fp = org_fp;
167 			goto out_addrs;
168 		}
169 	}
170 
171 	bpf_jit_realloc_regs(&cgctx);
172 	/*
173 	 * Pretend to build prologue, given the features we've seen.  This will
174 	 * update ctgtx.idx as it pretends to output instructions, then we can
175 	 * calculate total size from idx.
176 	 */
177 	bpf_jit_build_prologue(0, &cgctx);
178 	bpf_jit_build_epilogue(0, &cgctx);
179 
180 	proglen = cgctx.idx * 4;
181 	alloclen = proglen + FUNCTION_DESCR_SIZE;
182 
183 	bpf_hdr = bpf_jit_binary_alloc(alloclen, &image, 4, bpf_jit_fill_ill_insns);
184 	if (!bpf_hdr) {
185 		fp = org_fp;
186 		goto out_addrs;
187 	}
188 
189 skip_init_ctx:
190 	code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
191 
192 	if (extra_pass) {
193 		/*
194 		 * Do not touch the prologue and epilogue as they will remain
195 		 * unchanged. Only fix the branch target address for subprog
196 		 * calls in the body.
197 		 *
198 		 * This does not change the offsets and lengths of the subprog
199 		 * call instruction sequences and hence, the size of the JITed
200 		 * image as well.
201 		 */
202 		bpf_jit_fixup_subprog_calls(fp, code_base, &cgctx, addrs);
203 
204 		/* There is no need to perform the usual passes. */
205 		goto skip_codegen_passes;
206 	}
207 
208 	/* Code generation passes 1-2 */
209 	for (pass = 1; pass < 3; pass++) {
210 		/* Now build the prologue, body code & epilogue for real. */
211 		cgctx.idx = 0;
212 		bpf_jit_build_prologue(code_base, &cgctx);
213 		bpf_jit_build_body(fp, code_base, &cgctx, addrs, extra_pass);
214 		bpf_jit_build_epilogue(code_base, &cgctx);
215 
216 		if (bpf_jit_enable > 1)
217 			pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
218 				proglen - (cgctx.idx * 4), cgctx.seen);
219 	}
220 
221 skip_codegen_passes:
222 	if (bpf_jit_enable > 1)
223 		/*
224 		 * Note that we output the base address of the code_base
225 		 * rather than image, since opcodes are in code_base.
226 		 */
227 		bpf_jit_dump(flen, proglen, pass, code_base);
228 
229 #ifdef PPC64_ELF_ABI_v1
230 	/* Function descriptor nastiness: Address + TOC */
231 	((u64 *)image)[0] = (u64)code_base;
232 	((u64 *)image)[1] = local_paca->kernel_toc;
233 #endif
234 
235 	fp->bpf_func = (void *)image;
236 	fp->jited = 1;
237 	fp->jited_len = alloclen;
238 
239 	bpf_flush_icache(bpf_hdr, (u8 *)bpf_hdr + (bpf_hdr->pages * PAGE_SIZE));
240 	bpf_jit_binary_lock_ro(bpf_hdr);
241 	if (!fp->is_func || extra_pass) {
242 		bpf_prog_fill_jited_linfo(fp, addrs);
243 out_addrs:
244 		kfree(addrs);
245 		kfree(jit_data);
246 		fp->aux->jit_data = NULL;
247 	} else {
248 		jit_data->addrs = addrs;
249 		jit_data->ctx = cgctx;
250 		jit_data->proglen = proglen;
251 		jit_data->image = image;
252 		jit_data->header = bpf_hdr;
253 	}
254 
255 out:
256 	if (bpf_blinded)
257 		bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
258 
259 	return fp;
260 }
261