xref: /openbmc/linux/kernel/bpf/trampoline.c (revision 37744fee)
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 #include <linux/rbtree_latch.h>
8 #include <linux/perf_event.h>
9 #include <linux/btf.h>
10 
11 /* dummy _ops. The verifier will operate on target program's ops. */
12 const struct bpf_verifier_ops bpf_extension_verifier_ops = {
13 };
14 const struct bpf_prog_ops bpf_extension_prog_ops = {
15 };
16 
17 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */
18 #define TRAMPOLINE_HASH_BITS 10
19 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS)
20 
21 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE];
22 
23 /* serializes access to trampoline_table */
24 static DEFINE_MUTEX(trampoline_mutex);
25 
26 void *bpf_jit_alloc_exec_page(void)
27 {
28 	void *image;
29 
30 	image = bpf_jit_alloc_exec(PAGE_SIZE);
31 	if (!image)
32 		return NULL;
33 
34 	set_vm_flush_reset_perms(image);
35 	/* Keep image as writeable. The alternative is to keep flipping ro/rw
36 	 * everytime new program is attached or detached.
37 	 */
38 	set_memory_x((long)image, 1);
39 	return image;
40 }
41 
42 void bpf_image_ksym_add(void *data, struct bpf_ksym *ksym)
43 {
44 	ksym->start = (unsigned long) data;
45 	ksym->end = ksym->start + PAGE_SIZE;
46 	bpf_ksym_add(ksym);
47 	perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
48 			   PAGE_SIZE, false, ksym->name);
49 }
50 
51 void bpf_image_ksym_del(struct bpf_ksym *ksym)
52 {
53 	bpf_ksym_del(ksym);
54 	perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
55 			   PAGE_SIZE, true, ksym->name);
56 }
57 
58 static void bpf_trampoline_ksym_add(struct bpf_trampoline *tr)
59 {
60 	struct bpf_ksym *ksym = &tr->ksym;
61 
62 	snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu", tr->key);
63 	bpf_image_ksym_add(tr->image, ksym);
64 }
65 
66 struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
67 {
68 	struct bpf_trampoline *tr;
69 	struct hlist_head *head;
70 	void *image;
71 	int i;
72 
73 	mutex_lock(&trampoline_mutex);
74 	head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
75 	hlist_for_each_entry(tr, head, hlist) {
76 		if (tr->key == key) {
77 			refcount_inc(&tr->refcnt);
78 			goto out;
79 		}
80 	}
81 	tr = kzalloc(sizeof(*tr), GFP_KERNEL);
82 	if (!tr)
83 		goto out;
84 
85 	/* is_root was checked earlier. No need for bpf_jit_charge_modmem() */
86 	image = bpf_jit_alloc_exec_page();
87 	if (!image) {
88 		kfree(tr);
89 		tr = NULL;
90 		goto out;
91 	}
92 
93 	tr->key = key;
94 	INIT_HLIST_NODE(&tr->hlist);
95 	hlist_add_head(&tr->hlist, head);
96 	refcount_set(&tr->refcnt, 1);
97 	mutex_init(&tr->mutex);
98 	for (i = 0; i < BPF_TRAMP_MAX; i++)
99 		INIT_HLIST_HEAD(&tr->progs_hlist[i]);
100 	tr->image = image;
101 	INIT_LIST_HEAD_RCU(&tr->ksym.lnode);
102 	bpf_trampoline_ksym_add(tr);
103 out:
104 	mutex_unlock(&trampoline_mutex);
105 	return tr;
106 }
107 
108 static int is_ftrace_location(void *ip)
109 {
110 	long addr;
111 
112 	addr = ftrace_location((long)ip);
113 	if (!addr)
114 		return 0;
115 	if (WARN_ON_ONCE(addr != (long)ip))
116 		return -EFAULT;
117 	return 1;
118 }
119 
120 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
121 {
122 	void *ip = tr->func.addr;
123 	int ret;
124 
125 	if (tr->func.ftrace_managed)
126 		ret = unregister_ftrace_direct((long)ip, (long)old_addr);
127 	else
128 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
129 	return ret;
130 }
131 
132 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr)
133 {
134 	void *ip = tr->func.addr;
135 	int ret;
136 
137 	if (tr->func.ftrace_managed)
138 		ret = modify_ftrace_direct((long)ip, (long)old_addr, (long)new_addr);
139 	else
140 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr);
141 	return ret;
142 }
143 
144 /* first time registering */
145 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
146 {
147 	void *ip = tr->func.addr;
148 	int ret;
149 
150 	ret = is_ftrace_location(ip);
151 	if (ret < 0)
152 		return ret;
153 	tr->func.ftrace_managed = ret;
154 
155 	if (tr->func.ftrace_managed)
156 		ret = register_ftrace_direct((long)ip, (long)new_addr);
157 	else
158 		ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
159 	return ret;
160 }
161 
162 static struct bpf_tramp_progs *
163 bpf_trampoline_get_progs(const struct bpf_trampoline *tr, int *total)
164 {
165 	const struct bpf_prog_aux *aux;
166 	struct bpf_tramp_progs *tprogs;
167 	struct bpf_prog **progs;
168 	int kind;
169 
170 	*total = 0;
171 	tprogs = kcalloc(BPF_TRAMP_MAX, sizeof(*tprogs), GFP_KERNEL);
172 	if (!tprogs)
173 		return ERR_PTR(-ENOMEM);
174 
175 	for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
176 		tprogs[kind].nr_progs = tr->progs_cnt[kind];
177 		*total += tr->progs_cnt[kind];
178 		progs = tprogs[kind].progs;
179 
180 		hlist_for_each_entry(aux, &tr->progs_hlist[kind], tramp_hlist)
181 			*progs++ = aux->prog;
182 	}
183 	return tprogs;
184 }
185 
186 static int bpf_trampoline_update(struct bpf_trampoline *tr)
187 {
188 	void *old_image = tr->image + ((tr->selector + 1) & 1) * PAGE_SIZE/2;
189 	void *new_image = tr->image + (tr->selector & 1) * PAGE_SIZE/2;
190 	struct bpf_tramp_progs *tprogs;
191 	u32 flags = BPF_TRAMP_F_RESTORE_REGS;
192 	int err, total;
193 
194 	tprogs = bpf_trampoline_get_progs(tr, &total);
195 	if (IS_ERR(tprogs))
196 		return PTR_ERR(tprogs);
197 
198 	if (total == 0) {
199 		err = unregister_fentry(tr, old_image);
200 		tr->selector = 0;
201 		goto out;
202 	}
203 
204 	if (tprogs[BPF_TRAMP_FEXIT].nr_progs ||
205 	    tprogs[BPF_TRAMP_MODIFY_RETURN].nr_progs)
206 		flags = BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
207 
208 	/* Though the second half of trampoline page is unused a task could be
209 	 * preempted in the middle of the first half of trampoline and two
210 	 * updates to trampoline would change the code from underneath the
211 	 * preempted task. Hence wait for tasks to voluntarily schedule or go
212 	 * to userspace.
213 	 */
214 
215 	synchronize_rcu_tasks();
216 
217 	err = arch_prepare_bpf_trampoline(new_image, new_image + PAGE_SIZE / 2,
218 					  &tr->func.model, flags, tprogs,
219 					  tr->func.addr);
220 	if (err < 0)
221 		goto out;
222 
223 	if (tr->selector)
224 		/* progs already running at this address */
225 		err = modify_fentry(tr, old_image, new_image);
226 	else
227 		/* first time registering */
228 		err = register_fentry(tr, new_image);
229 	if (err)
230 		goto out;
231 	tr->selector++;
232 out:
233 	kfree(tprogs);
234 	return err;
235 }
236 
237 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
238 {
239 	switch (prog->expected_attach_type) {
240 	case BPF_TRACE_FENTRY:
241 		return BPF_TRAMP_FENTRY;
242 	case BPF_MODIFY_RETURN:
243 		return BPF_TRAMP_MODIFY_RETURN;
244 	case BPF_TRACE_FEXIT:
245 		return BPF_TRAMP_FEXIT;
246 	case BPF_LSM_MAC:
247 		if (!prog->aux->attach_func_proto->type)
248 			/* The function returns void, we cannot modify its
249 			 * return value.
250 			 */
251 			return BPF_TRAMP_FEXIT;
252 		else
253 			return BPF_TRAMP_MODIFY_RETURN;
254 	default:
255 		return BPF_TRAMP_REPLACE;
256 	}
257 }
258 
259 int bpf_trampoline_link_prog(struct bpf_prog *prog)
260 {
261 	enum bpf_tramp_prog_type kind;
262 	struct bpf_trampoline *tr;
263 	int err = 0;
264 	int cnt;
265 
266 	tr = prog->aux->trampoline;
267 	kind = bpf_attach_type_to_tramp(prog);
268 	mutex_lock(&tr->mutex);
269 	if (tr->extension_prog) {
270 		/* cannot attach fentry/fexit if extension prog is attached.
271 		 * cannot overwrite extension prog either.
272 		 */
273 		err = -EBUSY;
274 		goto out;
275 	}
276 	cnt = tr->progs_cnt[BPF_TRAMP_FENTRY] + tr->progs_cnt[BPF_TRAMP_FEXIT];
277 	if (kind == BPF_TRAMP_REPLACE) {
278 		/* Cannot attach extension if fentry/fexit are in use. */
279 		if (cnt) {
280 			err = -EBUSY;
281 			goto out;
282 		}
283 		tr->extension_prog = prog;
284 		err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
285 					 prog->bpf_func);
286 		goto out;
287 	}
288 	if (cnt >= BPF_MAX_TRAMP_PROGS) {
289 		err = -E2BIG;
290 		goto out;
291 	}
292 	if (!hlist_unhashed(&prog->aux->tramp_hlist)) {
293 		/* prog already linked */
294 		err = -EBUSY;
295 		goto out;
296 	}
297 	hlist_add_head(&prog->aux->tramp_hlist, &tr->progs_hlist[kind]);
298 	tr->progs_cnt[kind]++;
299 	err = bpf_trampoline_update(prog->aux->trampoline);
300 	if (err) {
301 		hlist_del(&prog->aux->tramp_hlist);
302 		tr->progs_cnt[kind]--;
303 	}
304 out:
305 	mutex_unlock(&tr->mutex);
306 	return err;
307 }
308 
309 /* bpf_trampoline_unlink_prog() should never fail. */
310 int bpf_trampoline_unlink_prog(struct bpf_prog *prog)
311 {
312 	enum bpf_tramp_prog_type kind;
313 	struct bpf_trampoline *tr;
314 	int err;
315 
316 	tr = prog->aux->trampoline;
317 	kind = bpf_attach_type_to_tramp(prog);
318 	mutex_lock(&tr->mutex);
319 	if (kind == BPF_TRAMP_REPLACE) {
320 		WARN_ON_ONCE(!tr->extension_prog);
321 		err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
322 					 tr->extension_prog->bpf_func, NULL);
323 		tr->extension_prog = NULL;
324 		goto out;
325 	}
326 	hlist_del(&prog->aux->tramp_hlist);
327 	tr->progs_cnt[kind]--;
328 	err = bpf_trampoline_update(prog->aux->trampoline);
329 out:
330 	mutex_unlock(&tr->mutex);
331 	return err;
332 }
333 
334 void bpf_trampoline_put(struct bpf_trampoline *tr)
335 {
336 	if (!tr)
337 		return;
338 	mutex_lock(&trampoline_mutex);
339 	if (!refcount_dec_and_test(&tr->refcnt))
340 		goto out;
341 	WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
342 	if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FENTRY])))
343 		goto out;
344 	if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FEXIT])))
345 		goto out;
346 	bpf_image_ksym_del(&tr->ksym);
347 	/* wait for tasks to get out of trampoline before freeing it */
348 	synchronize_rcu_tasks();
349 	bpf_jit_free_exec(tr->image);
350 	hlist_del(&tr->hlist);
351 	kfree(tr);
352 out:
353 	mutex_unlock(&trampoline_mutex);
354 }
355 
356 /* The logic is similar to BPF_PROG_RUN, but with an explicit
357  * rcu_read_lock() and migrate_disable() which are required
358  * for the trampoline. The macro is split into
359  * call _bpf_prog_enter
360  * call prog->bpf_func
361  * call __bpf_prog_exit
362  */
363 u64 notrace __bpf_prog_enter(void)
364 	__acquires(RCU)
365 {
366 	u64 start = 0;
367 
368 	rcu_read_lock();
369 	migrate_disable();
370 	if (static_branch_unlikely(&bpf_stats_enabled_key))
371 		start = sched_clock();
372 	return start;
373 }
374 
375 void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start)
376 	__releases(RCU)
377 {
378 	struct bpf_prog_stats *stats;
379 
380 	if (static_branch_unlikely(&bpf_stats_enabled_key) &&
381 	    /* static_key could be enabled in __bpf_prog_enter
382 	     * and disabled in __bpf_prog_exit.
383 	     * And vice versa.
384 	     * Hence check that 'start' is not zero.
385 	     */
386 	    start) {
387 		stats = this_cpu_ptr(prog->aux->stats);
388 		u64_stats_update_begin(&stats->syncp);
389 		stats->cnt++;
390 		stats->nsecs += sched_clock() - start;
391 		u64_stats_update_end(&stats->syncp);
392 	}
393 	migrate_enable();
394 	rcu_read_unlock();
395 }
396 
397 int __weak
398 arch_prepare_bpf_trampoline(void *image, void *image_end,
399 			    const struct btf_func_model *m, u32 flags,
400 			    struct bpf_tramp_progs *tprogs,
401 			    void *orig_call)
402 {
403 	return -ENOTSUPP;
404 }
405 
406 static int __init init_trampolines(void)
407 {
408 	int i;
409 
410 	for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
411 		INIT_HLIST_HEAD(&trampoline_table[i]);
412 	return 0;
413 }
414 late_initcall(init_trampolines);
415