1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2019 Facebook */
3 #include <linux/hash.h>
5 #include <linux/filter.h>
6 #include <linux/ftrace.h>
7 #include <linux/rbtree_latch.h>
8 #include <linux/perf_event.h>
10 #include <linux/rcupdate_trace.h>
11 #include <linux/rcupdate_wait.h>
13 /* dummy _ops. The verifier will operate on target program's ops. */
14 const struct bpf_verifier_ops bpf_extension_verifier_ops = {
16 const struct bpf_prog_ops bpf_extension_prog_ops = {
19 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */
20 #define TRAMPOLINE_HASH_BITS 10
21 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS)
23 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE];
25 /* serializes access to trampoline_table */
26 static DEFINE_MUTEX(trampoline_mutex);
28 void *bpf_jit_alloc_exec_page(void)
32 image = bpf_jit_alloc_exec(PAGE_SIZE);
36 set_vm_flush_reset_perms(image);
37 /* Keep image as writeable. The alternative is to keep flipping ro/rw
38 * everytime new program is attached or detached.
40 set_memory_x((long)image, 1);
44 void bpf_image_ksym_add(void *data, struct bpf_ksym *ksym)
46 ksym->start = (unsigned long) data;
47 ksym->end = ksym->start + PAGE_SIZE;
49 perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
50 PAGE_SIZE, false, ksym->name);
53 void bpf_image_ksym_del(struct bpf_ksym *ksym)
56 perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
57 PAGE_SIZE, true, ksym->name);
60 static void bpf_trampoline_ksym_add(struct bpf_trampoline *tr)
62 struct bpf_ksym *ksym = &tr->ksym;
64 snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu", tr->key);
65 bpf_image_ksym_add(tr->image, ksym);
68 static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
70 struct bpf_trampoline *tr;
71 struct hlist_head *head;
75 mutex_lock(&trampoline_mutex);
76 head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
77 hlist_for_each_entry(tr, head, hlist) {
79 refcount_inc(&tr->refcnt);
83 tr = kzalloc(sizeof(*tr), GFP_KERNEL);
87 /* is_root was checked earlier. No need for bpf_jit_charge_modmem() */
88 image = bpf_jit_alloc_exec_page();
96 INIT_HLIST_NODE(&tr->hlist);
97 hlist_add_head(&tr->hlist, head);
98 refcount_set(&tr->refcnt, 1);
99 mutex_init(&tr->mutex);
100 for (i = 0; i < BPF_TRAMP_MAX; i++)
101 INIT_HLIST_HEAD(&tr->progs_hlist[i]);
103 INIT_LIST_HEAD_RCU(&tr->ksym.lnode);
104 bpf_trampoline_ksym_add(tr);
106 mutex_unlock(&trampoline_mutex);
110 static int is_ftrace_location(void *ip)
114 addr = ftrace_location((long)ip);
117 if (WARN_ON_ONCE(addr != (long)ip))
122 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
124 void *ip = tr->func.addr;
127 if (tr->func.ftrace_managed)
128 ret = unregister_ftrace_direct((long)ip, (long)old_addr);
130 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
134 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr)
136 void *ip = tr->func.addr;
139 if (tr->func.ftrace_managed)
140 ret = modify_ftrace_direct((long)ip, (long)old_addr, (long)new_addr);
142 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr);
146 /* first time registering */
147 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
149 void *ip = tr->func.addr;
152 ret = is_ftrace_location(ip);
155 tr->func.ftrace_managed = ret;
157 if (tr->func.ftrace_managed)
158 ret = register_ftrace_direct((long)ip, (long)new_addr);
160 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
164 static struct bpf_tramp_progs *
165 bpf_trampoline_get_progs(const struct bpf_trampoline *tr, int *total)
167 const struct bpf_prog_aux *aux;
168 struct bpf_tramp_progs *tprogs;
169 struct bpf_prog **progs;
173 tprogs = kcalloc(BPF_TRAMP_MAX, sizeof(*tprogs), GFP_KERNEL);
175 return ERR_PTR(-ENOMEM);
177 for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
178 tprogs[kind].nr_progs = tr->progs_cnt[kind];
179 *total += tr->progs_cnt[kind];
180 progs = tprogs[kind].progs;
182 hlist_for_each_entry(aux, &tr->progs_hlist[kind], tramp_hlist)
183 *progs++ = aux->prog;
188 static int bpf_trampoline_update(struct bpf_trampoline *tr)
190 void *old_image = tr->image + ((tr->selector + 1) & 1) * PAGE_SIZE/2;
191 void *new_image = tr->image + (tr->selector & 1) * PAGE_SIZE/2;
192 struct bpf_tramp_progs *tprogs;
193 u32 flags = BPF_TRAMP_F_RESTORE_REGS;
196 tprogs = bpf_trampoline_get_progs(tr, &total);
198 return PTR_ERR(tprogs);
201 err = unregister_fentry(tr, old_image);
206 if (tprogs[BPF_TRAMP_FEXIT].nr_progs ||
207 tprogs[BPF_TRAMP_MODIFY_RETURN].nr_progs)
208 flags = BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
210 /* Though the second half of trampoline page is unused a task could be
211 * preempted in the middle of the first half of trampoline and two
212 * updates to trampoline would change the code from underneath the
213 * preempted task. Hence wait for tasks to voluntarily schedule or go
215 * The same trampoline can hold both sleepable and non-sleepable progs.
216 * synchronize_rcu_tasks_trace() is needed to make sure all sleepable
217 * programs finish executing.
218 * Wait for these two grace periods together.
220 synchronize_rcu_mult(call_rcu_tasks, call_rcu_tasks_trace);
222 err = arch_prepare_bpf_trampoline(new_image, new_image + PAGE_SIZE / 2,
223 &tr->func.model, flags, tprogs,
229 /* progs already running at this address */
230 err = modify_fentry(tr, old_image, new_image);
232 /* first time registering */
233 err = register_fentry(tr, new_image);
242 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
244 switch (prog->expected_attach_type) {
245 case BPF_TRACE_FENTRY:
246 return BPF_TRAMP_FENTRY;
247 case BPF_MODIFY_RETURN:
248 return BPF_TRAMP_MODIFY_RETURN;
249 case BPF_TRACE_FEXIT:
250 return BPF_TRAMP_FEXIT;
252 if (!prog->aux->attach_func_proto->type)
253 /* The function returns void, we cannot modify its
256 return BPF_TRAMP_FEXIT;
258 return BPF_TRAMP_MODIFY_RETURN;
260 return BPF_TRAMP_REPLACE;
264 int bpf_trampoline_link_prog(struct bpf_prog *prog, struct bpf_trampoline *tr)
266 enum bpf_tramp_prog_type kind;
270 kind = bpf_attach_type_to_tramp(prog);
271 mutex_lock(&tr->mutex);
272 if (tr->extension_prog) {
273 /* cannot attach fentry/fexit if extension prog is attached.
274 * cannot overwrite extension prog either.
279 cnt = tr->progs_cnt[BPF_TRAMP_FENTRY] + tr->progs_cnt[BPF_TRAMP_FEXIT];
280 if (kind == BPF_TRAMP_REPLACE) {
281 /* Cannot attach extension if fentry/fexit are in use. */
286 tr->extension_prog = prog;
287 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
291 if (cnt >= BPF_MAX_TRAMP_PROGS) {
295 if (!hlist_unhashed(&prog->aux->tramp_hlist)) {
296 /* prog already linked */
300 hlist_add_head(&prog->aux->tramp_hlist, &tr->progs_hlist[kind]);
301 tr->progs_cnt[kind]++;
302 err = bpf_trampoline_update(tr);
304 hlist_del(&prog->aux->tramp_hlist);
305 tr->progs_cnt[kind]--;
308 mutex_unlock(&tr->mutex);
312 /* bpf_trampoline_unlink_prog() should never fail. */
313 int bpf_trampoline_unlink_prog(struct bpf_prog *prog, struct bpf_trampoline *tr)
315 enum bpf_tramp_prog_type kind;
318 kind = bpf_attach_type_to_tramp(prog);
319 mutex_lock(&tr->mutex);
320 if (kind == BPF_TRAMP_REPLACE) {
321 WARN_ON_ONCE(!tr->extension_prog);
322 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
323 tr->extension_prog->bpf_func, NULL);
324 tr->extension_prog = NULL;
327 hlist_del(&prog->aux->tramp_hlist);
328 tr->progs_cnt[kind]--;
329 err = bpf_trampoline_update(tr);
331 mutex_unlock(&tr->mutex);
335 struct bpf_trampoline *bpf_trampoline_get(u64 key,
336 struct bpf_attach_target_info *tgt_info)
338 struct bpf_trampoline *tr;
340 tr = bpf_trampoline_lookup(key);
344 mutex_lock(&tr->mutex);
348 memcpy(&tr->func.model, &tgt_info->fmodel, sizeof(tgt_info->fmodel));
349 tr->func.addr = (void *)tgt_info->tgt_addr;
351 mutex_unlock(&tr->mutex);
355 void bpf_trampoline_put(struct bpf_trampoline *tr)
359 mutex_lock(&trampoline_mutex);
360 if (!refcount_dec_and_test(&tr->refcnt))
362 WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
363 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FENTRY])))
365 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FEXIT])))
367 bpf_image_ksym_del(&tr->ksym);
368 /* This code will be executed when all bpf progs (both sleepable and
369 * non-sleepable) went through
370 * bpf_prog_put()->call_rcu[_tasks_trace]()->bpf_prog_free_deferred().
371 * Hence no need for another synchronize_rcu_tasks_trace() here,
372 * but synchronize_rcu_tasks() is still needed, since trampoline
373 * may not have had any sleepable programs and we need to wait
374 * for tasks to get out of trampoline code before freeing it.
376 synchronize_rcu_tasks();
377 bpf_jit_free_exec(tr->image);
378 hlist_del(&tr->hlist);
381 mutex_unlock(&trampoline_mutex);
384 #define NO_START_TIME 1
385 static u64 notrace bpf_prog_start_time(void)
387 u64 start = NO_START_TIME;
389 if (static_branch_unlikely(&bpf_stats_enabled_key)) {
390 start = sched_clock();
391 if (unlikely(!start))
392 start = NO_START_TIME;
397 static void notrace inc_misses_counter(struct bpf_prog *prog)
399 struct bpf_prog_stats *stats;
401 stats = this_cpu_ptr(prog->stats);
402 u64_stats_update_begin(&stats->syncp);
404 u64_stats_update_end(&stats->syncp);
407 /* The logic is similar to BPF_PROG_RUN, but with an explicit
408 * rcu_read_lock() and migrate_disable() which are required
409 * for the trampoline. The macro is split into
410 * call __bpf_prog_enter
411 * call prog->bpf_func
412 * call __bpf_prog_exit
414 * __bpf_prog_enter returns:
415 * 0 - skip execution of the bpf prog
416 * 1 - execute bpf prog
417 * [2..MAX_U64] - excute bpf prog and record execution time.
418 * This is start time.
420 u64 notrace __bpf_prog_enter(struct bpf_prog *prog)
425 if (unlikely(__this_cpu_inc_return(*(prog->active)) != 1)) {
426 inc_misses_counter(prog);
429 return bpf_prog_start_time();
432 static void notrace update_prog_stats(struct bpf_prog *prog,
435 struct bpf_prog_stats *stats;
437 if (static_branch_unlikely(&bpf_stats_enabled_key) &&
438 /* static_key could be enabled in __bpf_prog_enter*
439 * and disabled in __bpf_prog_exit*.
441 * Hence check that 'start' is valid.
443 start > NO_START_TIME) {
444 stats = this_cpu_ptr(prog->stats);
445 u64_stats_update_begin(&stats->syncp);
447 stats->nsecs += sched_clock() - start;
448 u64_stats_update_end(&stats->syncp);
452 void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start)
455 update_prog_stats(prog, start);
456 __this_cpu_dec(*(prog->active));
461 u64 notrace __bpf_prog_enter_sleepable(struct bpf_prog *prog)
463 rcu_read_lock_trace();
466 if (unlikely(__this_cpu_inc_return(*(prog->active)) != 1)) {
467 inc_misses_counter(prog);
470 return bpf_prog_start_time();
473 void notrace __bpf_prog_exit_sleepable(struct bpf_prog *prog, u64 start)
475 update_prog_stats(prog, start);
476 __this_cpu_dec(*(prog->active));
478 rcu_read_unlock_trace();
482 arch_prepare_bpf_trampoline(void *image, void *image_end,
483 const struct btf_func_model *m, u32 flags,
484 struct bpf_tramp_progs *tprogs,
490 static int __init init_trampolines(void)
494 for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
495 INIT_HLIST_HEAD(&trampoline_table[i]);
498 late_initcall(init_trampolines);