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 #include <linux/rcupdate_trace.h>
11 #include <linux/rcupdate_wait.h>
12 #include <linux/static_call.h>
13 #include <linux/bpf_verifier.h>
14 #include <linux/bpf_lsm.h>
15 #include <linux/delay.h>
16
17 /* dummy _ops. The verifier will operate on target program's ops. */
18 const struct bpf_verifier_ops bpf_extension_verifier_ops = {
19 };
20 const struct bpf_prog_ops bpf_extension_prog_ops = {
21 };
22
23 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */
24 #define TRAMPOLINE_HASH_BITS 10
25 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS)
26
27 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE];
28
29 /* serializes access to trampoline_table */
30 static DEFINE_MUTEX(trampoline_mutex);
31
32 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
33 static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mutex);
34
35 static int bpf_tramp_ftrace_ops_func(struct ftrace_ops *ops, enum ftrace_ops_cmd cmd)
36 {
37 struct bpf_trampoline *tr = ops->private;
38 int ret = 0;
39
40 if (cmd == FTRACE_OPS_CMD_ENABLE_SHARE_IPMODIFY_SELF) {
41 /* This is called inside register_ftrace_direct_multi(), so
42 * tr->mutex is already locked.
43 */
44 lockdep_assert_held_once(&tr->mutex);
45
46 /* Instead of updating the trampoline here, we propagate
47 * -EAGAIN to register_ftrace_direct(). Then we can
48 * retry register_ftrace_direct() after updating the
49 * trampoline.
50 */
51 if ((tr->flags & BPF_TRAMP_F_CALL_ORIG) &&
52 !(tr->flags & BPF_TRAMP_F_ORIG_STACK)) {
53 if (WARN_ON_ONCE(tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY))
54 return -EBUSY;
55
56 tr->flags |= BPF_TRAMP_F_SHARE_IPMODIFY;
57 return -EAGAIN;
58 }
59
60 return 0;
61 }
62
63 /* The normal locking order is
64 * tr->mutex => direct_mutex (ftrace.c) => ftrace_lock (ftrace.c)
65 *
66 * The following two commands are called from
67 *
68 * prepare_direct_functions_for_ipmodify
69 * cleanup_direct_functions_after_ipmodify
70 *
71 * In both cases, direct_mutex is already locked. Use
72 * mutex_trylock(&tr->mutex) to avoid deadlock in race condition
73 * (something else is making changes to this same trampoline).
74 */
75 if (!mutex_trylock(&tr->mutex)) {
76 /* sleep 1 ms to make sure whatever holding tr->mutex makes
77 * some progress.
78 */
79 msleep(1);
80 return -EAGAIN;
81 }
82
83 switch (cmd) {
84 case FTRACE_OPS_CMD_ENABLE_SHARE_IPMODIFY_PEER:
85 tr->flags |= BPF_TRAMP_F_SHARE_IPMODIFY;
86
87 if ((tr->flags & BPF_TRAMP_F_CALL_ORIG) &&
88 !(tr->flags & BPF_TRAMP_F_ORIG_STACK))
89 ret = bpf_trampoline_update(tr, false /* lock_direct_mutex */);
90 break;
91 case FTRACE_OPS_CMD_DISABLE_SHARE_IPMODIFY_PEER:
92 tr->flags &= ~BPF_TRAMP_F_SHARE_IPMODIFY;
93
94 if (tr->flags & BPF_TRAMP_F_ORIG_STACK)
95 ret = bpf_trampoline_update(tr, false /* lock_direct_mutex */);
96 break;
97 default:
98 ret = -EINVAL;
99 break;
100 }
101
102 mutex_unlock(&tr->mutex);
103 return ret;
104 }
105 #endif
106
107 bool bpf_prog_has_trampoline(const struct bpf_prog *prog)
108 {
109 enum bpf_attach_type eatype = prog->expected_attach_type;
110 enum bpf_prog_type ptype = prog->type;
111
112 return (ptype == BPF_PROG_TYPE_TRACING &&
113 (eatype == BPF_TRACE_FENTRY || eatype == BPF_TRACE_FEXIT ||
114 eatype == BPF_MODIFY_RETURN)) ||
115 (ptype == BPF_PROG_TYPE_LSM && eatype == BPF_LSM_MAC);
116 }
117
118 void bpf_image_ksym_add(void *data, unsigned int size, struct bpf_ksym *ksym)
119 {
120 ksym->start = (unsigned long) data;
121 ksym->end = ksym->start + size;
122 bpf_ksym_add(ksym);
123 perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
124 PAGE_SIZE, false, ksym->name);
125 }
126
127 void bpf_image_ksym_del(struct bpf_ksym *ksym)
128 {
129 bpf_ksym_del(ksym);
130 perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
131 PAGE_SIZE, true, ksym->name);
132 }
133
134 static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
135 {
136 struct bpf_trampoline *tr;
137 struct hlist_head *head;
138 int i;
139
140 mutex_lock(&trampoline_mutex);
141 head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
142 hlist_for_each_entry(tr, head, hlist) {
143 if (tr->key == key) {
144 refcount_inc(&tr->refcnt);
145 goto out;
146 }
147 }
148 tr = kzalloc(sizeof(*tr), GFP_KERNEL);
149 if (!tr)
150 goto out;
151 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
152 tr->fops = kzalloc(sizeof(struct ftrace_ops), GFP_KERNEL);
153 if (!tr->fops) {
154 kfree(tr);
155 tr = NULL;
156 goto out;
157 }
158 tr->fops->private = tr;
159 tr->fops->ops_func = bpf_tramp_ftrace_ops_func;
160 #endif
161
162 tr->key = key;
163 INIT_HLIST_NODE(&tr->hlist);
164 hlist_add_head(&tr->hlist, head);
165 refcount_set(&tr->refcnt, 1);
166 mutex_init(&tr->mutex);
167 for (i = 0; i < BPF_TRAMP_MAX; i++)
168 INIT_HLIST_HEAD(&tr->progs_hlist[i]);
169 out:
170 mutex_unlock(&trampoline_mutex);
171 return tr;
172 }
173
174 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
175 {
176 void *ip = tr->func.addr;
177 int ret;
178
179 if (tr->func.ftrace_managed)
180 ret = unregister_ftrace_direct(tr->fops, (long)old_addr, false);
181 else
182 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
183
184 return ret;
185 }
186
187 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr,
188 bool lock_direct_mutex)
189 {
190 void *ip = tr->func.addr;
191 int ret;
192
193 if (tr->func.ftrace_managed) {
194 if (lock_direct_mutex)
195 ret = modify_ftrace_direct(tr->fops, (long)new_addr);
196 else
197 ret = modify_ftrace_direct_nolock(tr->fops, (long)new_addr);
198 } else {
199 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr);
200 }
201 return ret;
202 }
203
204 /* first time registering */
205 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
206 {
207 void *ip = tr->func.addr;
208 unsigned long faddr;
209 int ret;
210
211 faddr = ftrace_location((unsigned long)ip);
212 if (faddr) {
213 if (!tr->fops)
214 return -ENOTSUPP;
215 tr->func.ftrace_managed = true;
216 }
217
218 if (tr->func.ftrace_managed) {
219 ftrace_set_filter_ip(tr->fops, (unsigned long)ip, 0, 1);
220 ret = register_ftrace_direct(tr->fops, (long)new_addr);
221 } else {
222 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
223 }
224
225 return ret;
226 }
227
228 static struct bpf_tramp_links *
229 bpf_trampoline_get_progs(const struct bpf_trampoline *tr, int *total, bool *ip_arg)
230 {
231 struct bpf_tramp_link *link;
232 struct bpf_tramp_links *tlinks;
233 struct bpf_tramp_link **links;
234 int kind;
235
236 *total = 0;
237 tlinks = kcalloc(BPF_TRAMP_MAX, sizeof(*tlinks), GFP_KERNEL);
238 if (!tlinks)
239 return ERR_PTR(-ENOMEM);
240
241 for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
242 tlinks[kind].nr_links = tr->progs_cnt[kind];
243 *total += tr->progs_cnt[kind];
244 links = tlinks[kind].links;
245
246 hlist_for_each_entry(link, &tr->progs_hlist[kind], tramp_hlist) {
247 *ip_arg |= link->link.prog->call_get_func_ip;
248 *links++ = link;
249 }
250 }
251 return tlinks;
252 }
253
254 static void bpf_tramp_image_free(struct bpf_tramp_image *im)
255 {
256 bpf_image_ksym_del(&im->ksym);
257 arch_free_bpf_trampoline(im->image, im->size);
258 bpf_jit_uncharge_modmem(im->size);
259 percpu_ref_exit(&im->pcref);
260 kfree_rcu(im, rcu);
261 }
262
263 static void __bpf_tramp_image_put_deferred(struct work_struct *work)
264 {
265 struct bpf_tramp_image *im;
266
267 im = container_of(work, struct bpf_tramp_image, work);
268 bpf_tramp_image_free(im);
269 }
270
271 /* callback, fexit step 3 or fentry step 2 */
272 static void __bpf_tramp_image_put_rcu(struct rcu_head *rcu)
273 {
274 struct bpf_tramp_image *im;
275
276 im = container_of(rcu, struct bpf_tramp_image, rcu);
277 INIT_WORK(&im->work, __bpf_tramp_image_put_deferred);
278 schedule_work(&im->work);
279 }
280
281 /* callback, fexit step 2. Called after percpu_ref_kill confirms. */
282 static void __bpf_tramp_image_release(struct percpu_ref *pcref)
283 {
284 struct bpf_tramp_image *im;
285
286 im = container_of(pcref, struct bpf_tramp_image, pcref);
287 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
288 }
289
290 /* callback, fexit or fentry step 1 */
291 static void __bpf_tramp_image_put_rcu_tasks(struct rcu_head *rcu)
292 {
293 struct bpf_tramp_image *im;
294
295 im = container_of(rcu, struct bpf_tramp_image, rcu);
296 if (im->ip_after_call)
297 /* the case of fmod_ret/fexit trampoline and CONFIG_PREEMPTION=y */
298 percpu_ref_kill(&im->pcref);
299 else
300 /* the case of fentry trampoline */
301 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
302 }
303
304 static void bpf_tramp_image_put(struct bpf_tramp_image *im)
305 {
306 /* The trampoline image that calls original function is using:
307 * rcu_read_lock_trace to protect sleepable bpf progs
308 * rcu_read_lock to protect normal bpf progs
309 * percpu_ref to protect trampoline itself
310 * rcu tasks to protect trampoline asm not covered by percpu_ref
311 * (which are few asm insns before __bpf_tramp_enter and
312 * after __bpf_tramp_exit)
313 *
314 * The trampoline is unreachable before bpf_tramp_image_put().
315 *
316 * First, patch the trampoline to avoid calling into fexit progs.
317 * The progs will be freed even if the original function is still
318 * executing or sleeping.
319 * In case of CONFIG_PREEMPT=y use call_rcu_tasks() to wait on
320 * first few asm instructions to execute and call into
321 * __bpf_tramp_enter->percpu_ref_get.
322 * Then use percpu_ref_kill to wait for the trampoline and the original
323 * function to finish.
324 * Then use call_rcu_tasks() to make sure few asm insns in
325 * the trampoline epilogue are done as well.
326 *
327 * In !PREEMPT case the task that got interrupted in the first asm
328 * insns won't go through an RCU quiescent state which the
329 * percpu_ref_kill will be waiting for. Hence the first
330 * call_rcu_tasks() is not necessary.
331 */
332 if (im->ip_after_call) {
333 int err = bpf_arch_text_poke(im->ip_after_call, BPF_MOD_JUMP,
334 NULL, im->ip_epilogue);
335 WARN_ON(err);
336 if (IS_ENABLED(CONFIG_TASKS_RCU))
337 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
338 else
339 percpu_ref_kill(&im->pcref);
340 return;
341 }
342
343 /* The trampoline without fexit and fmod_ret progs doesn't call original
344 * function and doesn't use percpu_ref.
345 * Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
346 * Then use call_rcu_tasks() to wait for the rest of trampoline asm
347 * and normal progs.
348 */
349 call_rcu_tasks_trace(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
350 }
351
352 static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, int size)
353 {
354 struct bpf_tramp_image *im;
355 struct bpf_ksym *ksym;
356 void *image;
357 int err = -ENOMEM;
358
359 im = kzalloc(sizeof(*im), GFP_KERNEL);
360 if (!im)
361 goto out;
362
363 err = bpf_jit_charge_modmem(size);
364 if (err)
365 goto out_free_im;
366 im->size = size;
367
368 err = -ENOMEM;
369 im->image = image = arch_alloc_bpf_trampoline(size);
370 if (!image)
371 goto out_uncharge;
372
373 err = percpu_ref_init(&im->pcref, __bpf_tramp_image_release, 0, GFP_KERNEL);
374 if (err)
375 goto out_free_image;
376
377 ksym = &im->ksym;
378 INIT_LIST_HEAD_RCU(&ksym->lnode);
379 snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu", key);
380 bpf_image_ksym_add(image, size, ksym);
381 return im;
382
383 out_free_image:
384 arch_free_bpf_trampoline(im->image, im->size);
385 out_uncharge:
386 bpf_jit_uncharge_modmem(size);
387 out_free_im:
388 kfree(im);
389 out:
390 return ERR_PTR(err);
391 }
392
393 static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mutex)
394 {
395 struct bpf_tramp_image *im;
396 struct bpf_tramp_links *tlinks;
397 u32 orig_flags = tr->flags;
398 bool ip_arg = false;
399 int err, total, size;
400
401 tlinks = bpf_trampoline_get_progs(tr, &total, &ip_arg);
402 if (IS_ERR(tlinks))
403 return PTR_ERR(tlinks);
404
405 if (total == 0) {
406 err = unregister_fentry(tr, tr->cur_image->image);
407 bpf_tramp_image_put(tr->cur_image);
408 tr->cur_image = NULL;
409 goto out;
410 }
411
412 /* clear all bits except SHARE_IPMODIFY and TAIL_CALL_CTX */
413 tr->flags &= (BPF_TRAMP_F_SHARE_IPMODIFY | BPF_TRAMP_F_TAIL_CALL_CTX);
414
415 if (tlinks[BPF_TRAMP_FEXIT].nr_links ||
416 tlinks[BPF_TRAMP_MODIFY_RETURN].nr_links) {
417 /* NOTE: BPF_TRAMP_F_RESTORE_REGS and BPF_TRAMP_F_SKIP_FRAME
418 * should not be set together.
419 */
420 tr->flags |= BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
421 } else {
422 tr->flags |= BPF_TRAMP_F_RESTORE_REGS;
423 }
424
425 if (ip_arg)
426 tr->flags |= BPF_TRAMP_F_IP_ARG;
427
428 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
429 again:
430 if ((tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY) &&
431 (tr->flags & BPF_TRAMP_F_CALL_ORIG))
432 tr->flags |= BPF_TRAMP_F_ORIG_STACK;
433 #endif
434
435 size = arch_bpf_trampoline_size(&tr->func.model, tr->flags,
436 tlinks, tr->func.addr);
437 if (size < 0) {
438 err = size;
439 goto out;
440 }
441
442 if (size > PAGE_SIZE) {
443 err = -E2BIG;
444 goto out;
445 }
446
447 im = bpf_tramp_image_alloc(tr->key, size);
448 if (IS_ERR(im)) {
449 err = PTR_ERR(im);
450 goto out;
451 }
452
453 err = arch_prepare_bpf_trampoline(im, im->image, im->image + size,
454 &tr->func.model, tr->flags, tlinks,
455 tr->func.addr);
456 if (err < 0)
457 goto out_free;
458
459 err = arch_protect_bpf_trampoline(im->image, im->size);
460 if (err)
461 goto out_free;
462
463 WARN_ON(tr->cur_image && total == 0);
464 if (tr->cur_image)
465 /* progs already running at this address */
466 err = modify_fentry(tr, tr->cur_image->image, im->image, lock_direct_mutex);
467 else
468 /* first time registering */
469 err = register_fentry(tr, im->image);
470
471 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
472 if (err == -EAGAIN) {
473 /* -EAGAIN from bpf_tramp_ftrace_ops_func. Now
474 * BPF_TRAMP_F_SHARE_IPMODIFY is set, we can generate the
475 * trampoline again, and retry register.
476 */
477 /* reset fops->func and fops->trampoline for re-register */
478 tr->fops->func = NULL;
479 tr->fops->trampoline = 0;
480
481 /* free im memory and reallocate later */
482 bpf_tramp_image_free(im);
483 goto again;
484 }
485 #endif
486 if (err)
487 goto out_free;
488
489 if (tr->cur_image)
490 bpf_tramp_image_put(tr->cur_image);
491 tr->cur_image = im;
492 out:
493 /* If any error happens, restore previous flags */
494 if (err)
495 tr->flags = orig_flags;
496 kfree(tlinks);
497 return err;
498
499 out_free:
500 bpf_tramp_image_free(im);
501 goto out;
502 }
503
504 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
505 {
506 switch (prog->expected_attach_type) {
507 case BPF_TRACE_FENTRY:
508 return BPF_TRAMP_FENTRY;
509 case BPF_MODIFY_RETURN:
510 return BPF_TRAMP_MODIFY_RETURN;
511 case BPF_TRACE_FEXIT:
512 return BPF_TRAMP_FEXIT;
513 case BPF_LSM_MAC:
514 if (!prog->aux->attach_func_proto->type)
515 /* The function returns void, we cannot modify its
516 * return value.
517 */
518 return BPF_TRAMP_FEXIT;
519 else
520 return BPF_TRAMP_MODIFY_RETURN;
521 default:
522 return BPF_TRAMP_REPLACE;
523 }
524 }
525
526 static int __bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
527 {
528 enum bpf_tramp_prog_type kind;
529 struct bpf_tramp_link *link_exiting;
530 int err = 0;
531 int cnt = 0, i;
532
533 kind = bpf_attach_type_to_tramp(link->link.prog);
534 if (tr->extension_prog)
535 /* cannot attach fentry/fexit if extension prog is attached.
536 * cannot overwrite extension prog either.
537 */
538 return -EBUSY;
539
540 for (i = 0; i < BPF_TRAMP_MAX; i++)
541 cnt += tr->progs_cnt[i];
542
543 if (kind == BPF_TRAMP_REPLACE) {
544 /* Cannot attach extension if fentry/fexit are in use. */
545 if (cnt)
546 return -EBUSY;
547 tr->extension_prog = link->link.prog;
548 return bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
549 link->link.prog->bpf_func);
550 }
551 if (cnt >= BPF_MAX_TRAMP_LINKS)
552 return -E2BIG;
553 if (!hlist_unhashed(&link->tramp_hlist))
554 /* prog already linked */
555 return -EBUSY;
556 hlist_for_each_entry(link_exiting, &tr->progs_hlist[kind], tramp_hlist) {
557 if (link_exiting->link.prog != link->link.prog)
558 continue;
559 /* prog already linked */
560 return -EBUSY;
561 }
562
563 hlist_add_head(&link->tramp_hlist, &tr->progs_hlist[kind]);
564 tr->progs_cnt[kind]++;
565 err = bpf_trampoline_update(tr, true /* lock_direct_mutex */);
566 if (err) {
567 hlist_del_init(&link->tramp_hlist);
568 tr->progs_cnt[kind]--;
569 }
570 return err;
571 }
572
573 int bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
574 {
575 int err;
576
577 mutex_lock(&tr->mutex);
578 err = __bpf_trampoline_link_prog(link, tr);
579 mutex_unlock(&tr->mutex);
580 return err;
581 }
582
583 static int __bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
584 {
585 enum bpf_tramp_prog_type kind;
586 int err;
587
588 kind = bpf_attach_type_to_tramp(link->link.prog);
589 if (kind == BPF_TRAMP_REPLACE) {
590 WARN_ON_ONCE(!tr->extension_prog);
591 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
592 tr->extension_prog->bpf_func, NULL);
593 tr->extension_prog = NULL;
594 return err;
595 }
596 hlist_del_init(&link->tramp_hlist);
597 tr->progs_cnt[kind]--;
598 return bpf_trampoline_update(tr, true /* lock_direct_mutex */);
599 }
600
601 /* bpf_trampoline_unlink_prog() should never fail. */
602 int bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
603 {
604 int err;
605
606 mutex_lock(&tr->mutex);
607 err = __bpf_trampoline_unlink_prog(link, tr);
608 mutex_unlock(&tr->mutex);
609 return err;
610 }
611
612 #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_LSM)
613 static void bpf_shim_tramp_link_release(struct bpf_link *link)
614 {
615 struct bpf_shim_tramp_link *shim_link =
616 container_of(link, struct bpf_shim_tramp_link, link.link);
617
618 /* paired with 'shim_link->trampoline = tr' in bpf_trampoline_link_cgroup_shim */
619 if (!shim_link->trampoline)
620 return;
621
622 WARN_ON_ONCE(bpf_trampoline_unlink_prog(&shim_link->link, shim_link->trampoline));
623 bpf_trampoline_put(shim_link->trampoline);
624 }
625
626 static void bpf_shim_tramp_link_dealloc(struct bpf_link *link)
627 {
628 struct bpf_shim_tramp_link *shim_link =
629 container_of(link, struct bpf_shim_tramp_link, link.link);
630
631 kfree(shim_link);
632 }
633
634 static const struct bpf_link_ops bpf_shim_tramp_link_lops = {
635 .release = bpf_shim_tramp_link_release,
636 .dealloc = bpf_shim_tramp_link_dealloc,
637 };
638
639 static struct bpf_shim_tramp_link *cgroup_shim_alloc(const struct bpf_prog *prog,
640 bpf_func_t bpf_func,
641 int cgroup_atype)
642 {
643 struct bpf_shim_tramp_link *shim_link = NULL;
644 struct bpf_prog *p;
645
646 shim_link = kzalloc(sizeof(*shim_link), GFP_USER);
647 if (!shim_link)
648 return NULL;
649
650 p = bpf_prog_alloc(1, 0);
651 if (!p) {
652 kfree(shim_link);
653 return NULL;
654 }
655
656 p->jited = false;
657 p->bpf_func = bpf_func;
658
659 p->aux->cgroup_atype = cgroup_atype;
660 p->aux->attach_func_proto = prog->aux->attach_func_proto;
661 p->aux->attach_btf_id = prog->aux->attach_btf_id;
662 p->aux->attach_btf = prog->aux->attach_btf;
663 btf_get(p->aux->attach_btf);
664 p->type = BPF_PROG_TYPE_LSM;
665 p->expected_attach_type = BPF_LSM_MAC;
666 bpf_prog_inc(p);
667 bpf_link_init(&shim_link->link.link, BPF_LINK_TYPE_UNSPEC,
668 &bpf_shim_tramp_link_lops, p);
669 bpf_cgroup_atype_get(p->aux->attach_btf_id, cgroup_atype);
670
671 return shim_link;
672 }
673
674 static struct bpf_shim_tramp_link *cgroup_shim_find(struct bpf_trampoline *tr,
675 bpf_func_t bpf_func)
676 {
677 struct bpf_tramp_link *link;
678 int kind;
679
680 for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
681 hlist_for_each_entry(link, &tr->progs_hlist[kind], tramp_hlist) {
682 struct bpf_prog *p = link->link.prog;
683
684 if (p->bpf_func == bpf_func)
685 return container_of(link, struct bpf_shim_tramp_link, link);
686 }
687 }
688
689 return NULL;
690 }
691
692 int bpf_trampoline_link_cgroup_shim(struct bpf_prog *prog,
693 int cgroup_atype)
694 {
695 struct bpf_shim_tramp_link *shim_link = NULL;
696 struct bpf_attach_target_info tgt_info = {};
697 struct bpf_trampoline *tr;
698 bpf_func_t bpf_func;
699 u64 key;
700 int err;
701
702 err = bpf_check_attach_target(NULL, prog, NULL,
703 prog->aux->attach_btf_id,
704 &tgt_info);
705 if (err)
706 return err;
707
708 key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
709 prog->aux->attach_btf_id);
710
711 bpf_lsm_find_cgroup_shim(prog, &bpf_func);
712 tr = bpf_trampoline_get(key, &tgt_info);
713 if (!tr)
714 return -ENOMEM;
715
716 mutex_lock(&tr->mutex);
717
718 shim_link = cgroup_shim_find(tr, bpf_func);
719 if (shim_link) {
720 /* Reusing existing shim attached by the other program. */
721 bpf_link_inc(&shim_link->link.link);
722
723 mutex_unlock(&tr->mutex);
724 bpf_trampoline_put(tr); /* bpf_trampoline_get above */
725 return 0;
726 }
727
728 /* Allocate and install new shim. */
729
730 shim_link = cgroup_shim_alloc(prog, bpf_func, cgroup_atype);
731 if (!shim_link) {
732 err = -ENOMEM;
733 goto err;
734 }
735
736 err = __bpf_trampoline_link_prog(&shim_link->link, tr);
737 if (err)
738 goto err;
739
740 shim_link->trampoline = tr;
741 /* note, we're still holding tr refcnt from above */
742
743 mutex_unlock(&tr->mutex);
744
745 return 0;
746 err:
747 mutex_unlock(&tr->mutex);
748
749 if (shim_link)
750 bpf_link_put(&shim_link->link.link);
751
752 /* have to release tr while _not_ holding its mutex */
753 bpf_trampoline_put(tr); /* bpf_trampoline_get above */
754
755 return err;
756 }
757
758 void bpf_trampoline_unlink_cgroup_shim(struct bpf_prog *prog)
759 {
760 struct bpf_shim_tramp_link *shim_link = NULL;
761 struct bpf_trampoline *tr;
762 bpf_func_t bpf_func;
763 u64 key;
764
765 key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
766 prog->aux->attach_btf_id);
767
768 bpf_lsm_find_cgroup_shim(prog, &bpf_func);
769 tr = bpf_trampoline_lookup(key);
770 if (WARN_ON_ONCE(!tr))
771 return;
772
773 mutex_lock(&tr->mutex);
774 shim_link = cgroup_shim_find(tr, bpf_func);
775 mutex_unlock(&tr->mutex);
776
777 if (shim_link)
778 bpf_link_put(&shim_link->link.link);
779
780 bpf_trampoline_put(tr); /* bpf_trampoline_lookup above */
781 }
782 #endif
783
784 struct bpf_trampoline *bpf_trampoline_get(u64 key,
785 struct bpf_attach_target_info *tgt_info)
786 {
787 struct bpf_trampoline *tr;
788
789 tr = bpf_trampoline_lookup(key);
790 if (!tr)
791 return NULL;
792
793 mutex_lock(&tr->mutex);
794 if (tr->func.addr)
795 goto out;
796
797 memcpy(&tr->func.model, &tgt_info->fmodel, sizeof(tgt_info->fmodel));
798 tr->func.addr = (void *)tgt_info->tgt_addr;
799 out:
800 mutex_unlock(&tr->mutex);
801 return tr;
802 }
803
804 void bpf_trampoline_put(struct bpf_trampoline *tr)
805 {
806 int i;
807
808 if (!tr)
809 return;
810 mutex_lock(&trampoline_mutex);
811 if (!refcount_dec_and_test(&tr->refcnt))
812 goto out;
813 WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
814
815 for (i = 0; i < BPF_TRAMP_MAX; i++)
816 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[i])))
817 goto out;
818
819 /* This code will be executed even when the last bpf_tramp_image
820 * is alive. All progs are detached from the trampoline and the
821 * trampoline image is patched with jmp into epilogue to skip
822 * fexit progs. The fentry-only trampoline will be freed via
823 * multiple rcu callbacks.
824 */
825 hlist_del(&tr->hlist);
826 if (tr->fops) {
827 ftrace_free_filter(tr->fops);
828 kfree(tr->fops);
829 }
830 kfree(tr);
831 out:
832 mutex_unlock(&trampoline_mutex);
833 }
834
835 #define NO_START_TIME 1
836 static __always_inline u64 notrace bpf_prog_start_time(void)
837 {
838 u64 start = NO_START_TIME;
839
840 if (static_branch_unlikely(&bpf_stats_enabled_key)) {
841 start = sched_clock();
842 if (unlikely(!start))
843 start = NO_START_TIME;
844 }
845 return start;
846 }
847
848 /* The logic is similar to bpf_prog_run(), but with an explicit
849 * rcu_read_lock() and migrate_disable() which are required
850 * for the trampoline. The macro is split into
851 * call __bpf_prog_enter
852 * call prog->bpf_func
853 * call __bpf_prog_exit
854 *
855 * __bpf_prog_enter returns:
856 * 0 - skip execution of the bpf prog
857 * 1 - execute bpf prog
858 * [2..MAX_U64] - execute bpf prog and record execution time.
859 * This is start time.
860 */
861 static u64 notrace __bpf_prog_enter_recur(struct bpf_prog *prog, struct bpf_tramp_run_ctx *run_ctx)
862 __acquires(RCU)
863 {
864 rcu_read_lock();
865 migrate_disable();
866
867 run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
868
869 if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
870 bpf_prog_inc_misses_counter(prog);
871 return 0;
872 }
873 return bpf_prog_start_time();
874 }
875
876 static void notrace update_prog_stats(struct bpf_prog *prog,
877 u64 start)
878 {
879 struct bpf_prog_stats *stats;
880
881 if (static_branch_unlikely(&bpf_stats_enabled_key) &&
882 /* static_key could be enabled in __bpf_prog_enter*
883 * and disabled in __bpf_prog_exit*.
884 * And vice versa.
885 * Hence check that 'start' is valid.
886 */
887 start > NO_START_TIME) {
888 u64 duration = sched_clock() - start;
889 unsigned long flags;
890
891 stats = this_cpu_ptr(prog->stats);
892 flags = u64_stats_update_begin_irqsave(&stats->syncp);
893 u64_stats_inc(&stats->cnt);
894 u64_stats_add(&stats->nsecs, duration);
895 u64_stats_update_end_irqrestore(&stats->syncp, flags);
896 }
897 }
898
899 static void notrace __bpf_prog_exit_recur(struct bpf_prog *prog, u64 start,
900 struct bpf_tramp_run_ctx *run_ctx)
901 __releases(RCU)
902 {
903 bpf_reset_run_ctx(run_ctx->saved_run_ctx);
904
905 update_prog_stats(prog, start);
906 this_cpu_dec(*(prog->active));
907 migrate_enable();
908 rcu_read_unlock();
909 }
910
911 static u64 notrace __bpf_prog_enter_lsm_cgroup(struct bpf_prog *prog,
912 struct bpf_tramp_run_ctx *run_ctx)
913 __acquires(RCU)
914 {
915 /* Runtime stats are exported via actual BPF_LSM_CGROUP
916 * programs, not the shims.
917 */
918 rcu_read_lock();
919 migrate_disable();
920
921 run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
922
923 return NO_START_TIME;
924 }
925
926 static void notrace __bpf_prog_exit_lsm_cgroup(struct bpf_prog *prog, u64 start,
927 struct bpf_tramp_run_ctx *run_ctx)
928 __releases(RCU)
929 {
930 bpf_reset_run_ctx(run_ctx->saved_run_ctx);
931
932 migrate_enable();
933 rcu_read_unlock();
934 }
935
936 u64 notrace __bpf_prog_enter_sleepable_recur(struct bpf_prog *prog,
937 struct bpf_tramp_run_ctx *run_ctx)
938 {
939 rcu_read_lock_trace();
940 migrate_disable();
941 might_fault();
942
943 run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
944
945 if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
946 bpf_prog_inc_misses_counter(prog);
947 return 0;
948 }
949 return bpf_prog_start_time();
950 }
951
952 void notrace __bpf_prog_exit_sleepable_recur(struct bpf_prog *prog, u64 start,
953 struct bpf_tramp_run_ctx *run_ctx)
954 {
955 bpf_reset_run_ctx(run_ctx->saved_run_ctx);
956
957 update_prog_stats(prog, start);
958 this_cpu_dec(*(prog->active));
959 migrate_enable();
960 rcu_read_unlock_trace();
961 }
962
963 static u64 notrace __bpf_prog_enter_sleepable(struct bpf_prog *prog,
964 struct bpf_tramp_run_ctx *run_ctx)
965 {
966 rcu_read_lock_trace();
967 migrate_disable();
968 might_fault();
969
970 run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
971
972 return bpf_prog_start_time();
973 }
974
975 static void notrace __bpf_prog_exit_sleepable(struct bpf_prog *prog, u64 start,
976 struct bpf_tramp_run_ctx *run_ctx)
977 {
978 bpf_reset_run_ctx(run_ctx->saved_run_ctx);
979
980 update_prog_stats(prog, start);
981 migrate_enable();
982 rcu_read_unlock_trace();
983 }
984
985 static u64 notrace __bpf_prog_enter(struct bpf_prog *prog,
986 struct bpf_tramp_run_ctx *run_ctx)
987 __acquires(RCU)
988 {
989 rcu_read_lock();
990 migrate_disable();
991
992 run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
993
994 return bpf_prog_start_time();
995 }
996
997 static void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start,
998 struct bpf_tramp_run_ctx *run_ctx)
999 __releases(RCU)
1000 {
1001 bpf_reset_run_ctx(run_ctx->saved_run_ctx);
1002
1003 update_prog_stats(prog, start);
1004 migrate_enable();
1005 rcu_read_unlock();
1006 }
1007
1008 void notrace __bpf_tramp_enter(struct bpf_tramp_image *tr)
1009 {
1010 percpu_ref_get(&tr->pcref);
1011 }
1012
1013 void notrace __bpf_tramp_exit(struct bpf_tramp_image *tr)
1014 {
1015 percpu_ref_put(&tr->pcref);
1016 }
1017
1018 bpf_trampoline_enter_t bpf_trampoline_enter(const struct bpf_prog *prog)
1019 {
1020 bool sleepable = prog->sleepable;
1021
1022 if (bpf_prog_check_recur(prog))
1023 return sleepable ? __bpf_prog_enter_sleepable_recur :
1024 __bpf_prog_enter_recur;
1025
1026 if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1027 prog->expected_attach_type == BPF_LSM_CGROUP)
1028 return __bpf_prog_enter_lsm_cgroup;
1029
1030 return sleepable ? __bpf_prog_enter_sleepable : __bpf_prog_enter;
1031 }
1032
1033 bpf_trampoline_exit_t bpf_trampoline_exit(const struct bpf_prog *prog)
1034 {
1035 bool sleepable = prog->sleepable;
1036
1037 if (bpf_prog_check_recur(prog))
1038 return sleepable ? __bpf_prog_exit_sleepable_recur :
1039 __bpf_prog_exit_recur;
1040
1041 if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1042 prog->expected_attach_type == BPF_LSM_CGROUP)
1043 return __bpf_prog_exit_lsm_cgroup;
1044
1045 return sleepable ? __bpf_prog_exit_sleepable : __bpf_prog_exit;
1046 }
1047
1048 int __weak
1049 arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
1050 const struct btf_func_model *m, u32 flags,
1051 struct bpf_tramp_links *tlinks,
1052 void *func_addr)
1053 {
1054 return -ENOTSUPP;
1055 }
1056
1057 void * __weak arch_alloc_bpf_trampoline(unsigned int size)
1058 {
1059 void *image;
1060
1061 if (WARN_ON_ONCE(size > PAGE_SIZE))
1062 return NULL;
1063 image = bpf_jit_alloc_exec(PAGE_SIZE);
1064 if (image)
1065 set_vm_flush_reset_perms(image);
1066 return image;
1067 }
1068
1069 void __weak arch_free_bpf_trampoline(void *image, unsigned int size)
1070 {
1071 WARN_ON_ONCE(size > PAGE_SIZE);
1072 /* bpf_jit_free_exec doesn't need "size", but
1073 * bpf_prog_pack_free() needs it.
1074 */
1075 bpf_jit_free_exec(image);
1076 }
1077
1078 int __weak arch_protect_bpf_trampoline(void *image, unsigned int size)
1079 {
1080 WARN_ON_ONCE(size > PAGE_SIZE);
1081 return set_memory_rox((long)image, 1);
1082 }
1083
1084 int __weak arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
1085 struct bpf_tramp_links *tlinks, void *func_addr)
1086 {
1087 return -ENOTSUPP;
1088 }
1089
1090 static int __init init_trampolines(void)
1091 {
1092 int i;
1093
1094 for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
1095 INIT_HLIST_HEAD(&trampoline_table[i]);
1096 return 0;
1097 }
1098 late_initcall(init_trampolines);
1099
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.