xref: /linux/kernel/bpf/trampoline.c (revision a85dd3a5170c8812cd835ea968ccadf0ebf1648e)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2019 Facebook */
3 #include <linux/hash.h>
4 #include <linux/bpf.h>
5 #include <linux/filter.h>
6 #include <linux/ftrace.h>
7 
8 /* dummy _ops. The verifier will operate on target program's ops. */
9 const struct bpf_verifier_ops bpf_extension_verifier_ops = {
10 };
11 const struct bpf_prog_ops bpf_extension_prog_ops = {
12 };
13 
14 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */
15 #define TRAMPOLINE_HASH_BITS 10
16 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS)
17 
18 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE];
19 
20 /* serializes access to trampoline_table */
21 static DEFINE_MUTEX(trampoline_mutex);
22 
23 void *bpf_jit_alloc_exec_page(void)
24 {
25 	void *image;
26 
27 	image = bpf_jit_alloc_exec(PAGE_SIZE);
28 	if (!image)
29 		return NULL;
30 
31 	set_vm_flush_reset_perms(image);
32 	/* Keep image as writeable. The alternative is to keep flipping ro/rw
33 	 * everytime new program is attached or detached.
34 	 */
35 	set_memory_x((long)image, 1);
36 	return image;
37 }
38 
39 struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
40 {
41 	struct bpf_trampoline *tr;
42 	struct hlist_head *head;
43 	void *image;
44 	int i;
45 
46 	mutex_lock(&trampoline_mutex);
47 	head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
48 	hlist_for_each_entry(tr, head, hlist) {
49 		if (tr->key == key) {
50 			refcount_inc(&tr->refcnt);
51 			goto out;
52 		}
53 	}
54 	tr = kzalloc(sizeof(*tr), GFP_KERNEL);
55 	if (!tr)
56 		goto out;
57 
58 	/* is_root was checked earlier. No need for bpf_jit_charge_modmem() */
59 	image = bpf_jit_alloc_exec_page();
60 	if (!image) {
61 		kfree(tr);
62 		tr = NULL;
63 		goto out;
64 	}
65 
66 	tr->key = key;
67 	INIT_HLIST_NODE(&tr->hlist);
68 	hlist_add_head(&tr->hlist, head);
69 	refcount_set(&tr->refcnt, 1);
70 	mutex_init(&tr->mutex);
71 	for (i = 0; i < BPF_TRAMP_MAX; i++)
72 		INIT_HLIST_HEAD(&tr->progs_hlist[i]);
73 	tr->image = image;
74 out:
75 	mutex_unlock(&trampoline_mutex);
76 	return tr;
77 }
78 
79 static int is_ftrace_location(void *ip)
80 {
81 	long addr;
82 
83 	addr = ftrace_location((long)ip);
84 	if (!addr)
85 		return 0;
86 	if (WARN_ON_ONCE(addr != (long)ip))
87 		return -EFAULT;
88 	return 1;
89 }
90 
91 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
92 {
93 	void *ip = tr->func.addr;
94 	int ret;
95 
96 	if (tr->func.ftrace_managed)
97 		ret = unregister_ftrace_direct((long)ip, (long)old_addr);
98 	else
99 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
100 	return ret;
101 }
102 
103 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr)
104 {
105 	void *ip = tr->func.addr;
106 	int ret;
107 
108 	if (tr->func.ftrace_managed)
109 		ret = modify_ftrace_direct((long)ip, (long)old_addr, (long)new_addr);
110 	else
111 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr);
112 	return ret;
113 }
114 
115 /* first time registering */
116 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
117 {
118 	void *ip = tr->func.addr;
119 	int ret;
120 
121 	ret = is_ftrace_location(ip);
122 	if (ret < 0)
123 		return ret;
124 	tr->func.ftrace_managed = ret;
125 
126 	if (tr->func.ftrace_managed)
127 		ret = register_ftrace_direct((long)ip, (long)new_addr);
128 	else
129 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
130 	return ret;
131 }
132 
133 /* Each call __bpf_prog_enter + call bpf_func + call __bpf_prog_exit is ~50
134  * bytes on x86.  Pick a number to fit into PAGE_SIZE / 2
135  */
136 #define BPF_MAX_TRAMP_PROGS 40
137 
138 static int bpf_trampoline_update(struct bpf_trampoline *tr)
139 {
140 	void *old_image = tr->image + ((tr->selector + 1) & 1) * PAGE_SIZE/2;
141 	void *new_image = tr->image + (tr->selector & 1) * PAGE_SIZE/2;
142 	struct bpf_prog *progs_to_run[BPF_MAX_TRAMP_PROGS];
143 	int fentry_cnt = tr->progs_cnt[BPF_TRAMP_FENTRY];
144 	int fexit_cnt = tr->progs_cnt[BPF_TRAMP_FEXIT];
145 	struct bpf_prog **progs, **fentry, **fexit;
146 	u32 flags = BPF_TRAMP_F_RESTORE_REGS;
147 	struct bpf_prog_aux *aux;
148 	int err;
149 
150 	if (fentry_cnt + fexit_cnt == 0) {
151 		err = unregister_fentry(tr, old_image);
152 		tr->selector = 0;
153 		goto out;
154 	}
155 
156 	/* populate fentry progs */
157 	fentry = progs = progs_to_run;
158 	hlist_for_each_entry(aux, &tr->progs_hlist[BPF_TRAMP_FENTRY], tramp_hlist)
159 		*progs++ = aux->prog;
160 
161 	/* populate fexit progs */
162 	fexit = progs;
163 	hlist_for_each_entry(aux, &tr->progs_hlist[BPF_TRAMP_FEXIT], tramp_hlist)
164 		*progs++ = aux->prog;
165 
166 	if (fexit_cnt)
167 		flags = BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
168 
169 	/* Though the second half of trampoline page is unused a task could be
170 	 * preempted in the middle of the first half of trampoline and two
171 	 * updates to trampoline would change the code from underneath the
172 	 * preempted task. Hence wait for tasks to voluntarily schedule or go
173 	 * to userspace.
174 	 */
175 	synchronize_rcu_tasks();
176 
177 	err = arch_prepare_bpf_trampoline(new_image, new_image + PAGE_SIZE / 2,
178 					  &tr->func.model, flags,
179 					  fentry, fentry_cnt,
180 					  fexit, fexit_cnt,
181 					  tr->func.addr);
182 	if (err < 0)
183 		goto out;
184 
185 	if (tr->selector)
186 		/* progs already running at this address */
187 		err = modify_fentry(tr, old_image, new_image);
188 	else
189 		/* first time registering */
190 		err = register_fentry(tr, new_image);
191 	if (err)
192 		goto out;
193 	tr->selector++;
194 out:
195 	return err;
196 }
197 
198 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(enum bpf_attach_type t)
199 {
200 	switch (t) {
201 	case BPF_TRACE_FENTRY:
202 		return BPF_TRAMP_FENTRY;
203 	case BPF_TRACE_FEXIT:
204 		return BPF_TRAMP_FEXIT;
205 	default:
206 		return BPF_TRAMP_REPLACE;
207 	}
208 }
209 
210 int bpf_trampoline_link_prog(struct bpf_prog *prog)
211 {
212 	enum bpf_tramp_prog_type kind;
213 	struct bpf_trampoline *tr;
214 	int err = 0;
215 	int cnt;
216 
217 	tr = prog->aux->trampoline;
218 	kind = bpf_attach_type_to_tramp(prog->expected_attach_type);
219 	mutex_lock(&tr->mutex);
220 	if (tr->extension_prog) {
221 		/* cannot attach fentry/fexit if extension prog is attached.
222 		 * cannot overwrite extension prog either.
223 		 */
224 		err = -EBUSY;
225 		goto out;
226 	}
227 	cnt = tr->progs_cnt[BPF_TRAMP_FENTRY] + tr->progs_cnt[BPF_TRAMP_FEXIT];
228 	if (kind == BPF_TRAMP_REPLACE) {
229 		/* Cannot attach extension if fentry/fexit are in use. */
230 		if (cnt) {
231 			err = -EBUSY;
232 			goto out;
233 		}
234 		tr->extension_prog = prog;
235 		err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
236 					 prog->bpf_func);
237 		goto out;
238 	}
239 	if (cnt >= BPF_MAX_TRAMP_PROGS) {
240 		err = -E2BIG;
241 		goto out;
242 	}
243 	if (!hlist_unhashed(&prog->aux->tramp_hlist)) {
244 		/* prog already linked */
245 		err = -EBUSY;
246 		goto out;
247 	}
248 	hlist_add_head(&prog->aux->tramp_hlist, &tr->progs_hlist[kind]);
249 	tr->progs_cnt[kind]++;
250 	err = bpf_trampoline_update(prog->aux->trampoline);
251 	if (err) {
252 		hlist_del(&prog->aux->tramp_hlist);
253 		tr->progs_cnt[kind]--;
254 	}
255 out:
256 	mutex_unlock(&tr->mutex);
257 	return err;
258 }
259 
260 /* bpf_trampoline_unlink_prog() should never fail. */
261 int bpf_trampoline_unlink_prog(struct bpf_prog *prog)
262 {
263 	enum bpf_tramp_prog_type kind;
264 	struct bpf_trampoline *tr;
265 	int err;
266 
267 	tr = prog->aux->trampoline;
268 	kind = bpf_attach_type_to_tramp(prog->expected_attach_type);
269 	mutex_lock(&tr->mutex);
270 	if (kind == BPF_TRAMP_REPLACE) {
271 		WARN_ON_ONCE(!tr->extension_prog);
272 		err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
273 					 tr->extension_prog->bpf_func, NULL);
274 		tr->extension_prog = NULL;
275 		goto out;
276 	}
277 	hlist_del(&prog->aux->tramp_hlist);
278 	tr->progs_cnt[kind]--;
279 	err = bpf_trampoline_update(prog->aux->trampoline);
280 out:
281 	mutex_unlock(&tr->mutex);
282 	return err;
283 }
284 
285 void bpf_trampoline_put(struct bpf_trampoline *tr)
286 {
287 	if (!tr)
288 		return;
289 	mutex_lock(&trampoline_mutex);
290 	if (!refcount_dec_and_test(&tr->refcnt))
291 		goto out;
292 	WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
293 	if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FENTRY])))
294 		goto out;
295 	if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FEXIT])))
296 		goto out;
297 	/* wait for tasks to get out of trampoline before freeing it */
298 	synchronize_rcu_tasks();
299 	bpf_jit_free_exec(tr->image);
300 	hlist_del(&tr->hlist);
301 	kfree(tr);
302 out:
303 	mutex_unlock(&trampoline_mutex);
304 }
305 
306 /* The logic is similar to BPF_PROG_RUN, but with explicit rcu and preempt that
307  * are needed for trampoline. The macro is split into
308  * call _bpf_prog_enter
309  * call prog->bpf_func
310  * call __bpf_prog_exit
311  */
312 u64 notrace __bpf_prog_enter(void)
313 {
314 	u64 start = 0;
315 
316 	rcu_read_lock();
317 	preempt_disable();
318 	if (static_branch_unlikely(&bpf_stats_enabled_key))
319 		start = sched_clock();
320 	return start;
321 }
322 
323 void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start)
324 {
325 	struct bpf_prog_stats *stats;
326 
327 	if (static_branch_unlikely(&bpf_stats_enabled_key) &&
328 	    /* static_key could be enabled in __bpf_prog_enter
329 	     * and disabled in __bpf_prog_exit.
330 	     * And vice versa.
331 	     * Hence check that 'start' is not zero.
332 	     */
333 	    start) {
334 		stats = this_cpu_ptr(prog->aux->stats);
335 		u64_stats_update_begin(&stats->syncp);
336 		stats->cnt++;
337 		stats->nsecs += sched_clock() - start;
338 		u64_stats_update_end(&stats->syncp);
339 	}
340 	preempt_enable();
341 	rcu_read_unlock();
342 }
343 
344 int __weak
345 arch_prepare_bpf_trampoline(void *image, void *image_end,
346 			    const struct btf_func_model *m, u32 flags,
347 			    struct bpf_prog **fentry_progs, int fentry_cnt,
348 			    struct bpf_prog **fexit_progs, int fexit_cnt,
349 			    void *orig_call)
350 {
351 	return -ENOTSUPP;
352 }
353 
354 static int __init init_trampolines(void)
355 {
356 	int i;
357 
358 	for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
359 		INIT_HLIST_HEAD(&trampoline_table[i]);
360 	return 0;
361 }
362 late_initcall(init_trampolines);
363