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 9 /* dummy _ops. The verifier will operate on target program's ops. */ 10 const struct bpf_verifier_ops bpf_extension_verifier_ops = { 11 }; 12 const struct bpf_prog_ops bpf_extension_prog_ops = { 13 }; 14 15 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */ 16 #define TRAMPOLINE_HASH_BITS 10 17 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS) 18 19 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE]; 20 static struct latch_tree_root image_tree __cacheline_aligned; 21 22 /* serializes access to trampoline_table and image_tree */ 23 static DEFINE_MUTEX(trampoline_mutex); 24 25 static void *bpf_jit_alloc_exec_page(void) 26 { 27 void *image; 28 29 image = bpf_jit_alloc_exec(PAGE_SIZE); 30 if (!image) 31 return NULL; 32 33 set_vm_flush_reset_perms(image); 34 /* Keep image as writeable. The alternative is to keep flipping ro/rw 35 * everytime new program is attached or detached. 36 */ 37 set_memory_x((long)image, 1); 38 return image; 39 } 40 41 static __always_inline bool image_tree_less(struct latch_tree_node *a, 42 struct latch_tree_node *b) 43 { 44 struct bpf_image *ia = container_of(a, struct bpf_image, tnode); 45 struct bpf_image *ib = container_of(b, struct bpf_image, tnode); 46 47 return ia < ib; 48 } 49 50 static __always_inline int image_tree_comp(void *addr, struct latch_tree_node *n) 51 { 52 void *image = container_of(n, struct bpf_image, tnode); 53 54 if (addr < image) 55 return -1; 56 if (addr >= image + PAGE_SIZE) 57 return 1; 58 59 return 0; 60 } 61 62 static const struct latch_tree_ops image_tree_ops = { 63 .less = image_tree_less, 64 .comp = image_tree_comp, 65 }; 66 67 static void *__bpf_image_alloc(bool lock) 68 { 69 struct bpf_image *image; 70 71 image = bpf_jit_alloc_exec_page(); 72 if (!image) 73 return NULL; 74 75 if (lock) 76 mutex_lock(&trampoline_mutex); 77 latch_tree_insert(&image->tnode, &image_tree, &image_tree_ops); 78 if (lock) 79 mutex_unlock(&trampoline_mutex); 80 return image->data; 81 } 82 83 void *bpf_image_alloc(void) 84 { 85 return __bpf_image_alloc(true); 86 } 87 88 bool is_bpf_image_address(unsigned long addr) 89 { 90 bool ret; 91 92 rcu_read_lock(); 93 ret = latch_tree_find((void *) addr, &image_tree, &image_tree_ops) != NULL; 94 rcu_read_unlock(); 95 96 return ret; 97 } 98 99 struct bpf_trampoline *bpf_trampoline_lookup(u64 key) 100 { 101 struct bpf_trampoline *tr; 102 struct hlist_head *head; 103 void *image; 104 int i; 105 106 mutex_lock(&trampoline_mutex); 107 head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)]; 108 hlist_for_each_entry(tr, head, hlist) { 109 if (tr->key == key) { 110 refcount_inc(&tr->refcnt); 111 goto out; 112 } 113 } 114 tr = kzalloc(sizeof(*tr), GFP_KERNEL); 115 if (!tr) 116 goto out; 117 118 /* is_root was checked earlier. No need for bpf_jit_charge_modmem() */ 119 image = __bpf_image_alloc(false); 120 if (!image) { 121 kfree(tr); 122 tr = NULL; 123 goto out; 124 } 125 126 tr->key = key; 127 INIT_HLIST_NODE(&tr->hlist); 128 hlist_add_head(&tr->hlist, head); 129 refcount_set(&tr->refcnt, 1); 130 mutex_init(&tr->mutex); 131 for (i = 0; i < BPF_TRAMP_MAX; i++) 132 INIT_HLIST_HEAD(&tr->progs_hlist[i]); 133 tr->image = image; 134 out: 135 mutex_unlock(&trampoline_mutex); 136 return tr; 137 } 138 139 static int is_ftrace_location(void *ip) 140 { 141 long addr; 142 143 addr = ftrace_location((long)ip); 144 if (!addr) 145 return 0; 146 if (WARN_ON_ONCE(addr != (long)ip)) 147 return -EFAULT; 148 return 1; 149 } 150 151 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr) 152 { 153 void *ip = tr->func.addr; 154 int ret; 155 156 if (tr->func.ftrace_managed) 157 ret = unregister_ftrace_direct((long)ip, (long)old_addr); 158 else 159 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL); 160 return ret; 161 } 162 163 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr) 164 { 165 void *ip = tr->func.addr; 166 int ret; 167 168 if (tr->func.ftrace_managed) 169 ret = modify_ftrace_direct((long)ip, (long)old_addr, (long)new_addr); 170 else 171 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr); 172 return ret; 173 } 174 175 /* first time registering */ 176 static int register_fentry(struct bpf_trampoline *tr, void *new_addr) 177 { 178 void *ip = tr->func.addr; 179 int ret; 180 181 ret = is_ftrace_location(ip); 182 if (ret < 0) 183 return ret; 184 tr->func.ftrace_managed = ret; 185 186 if (tr->func.ftrace_managed) 187 ret = register_ftrace_direct((long)ip, (long)new_addr); 188 else 189 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr); 190 return ret; 191 } 192 193 static struct bpf_tramp_progs * 194 bpf_trampoline_get_progs(const struct bpf_trampoline *tr, int *total) 195 { 196 const struct bpf_prog_aux *aux; 197 struct bpf_tramp_progs *tprogs; 198 struct bpf_prog **progs; 199 int kind; 200 201 *total = 0; 202 tprogs = kcalloc(BPF_TRAMP_MAX, sizeof(*tprogs), GFP_KERNEL); 203 if (!tprogs) 204 return ERR_PTR(-ENOMEM); 205 206 for (kind = 0; kind < BPF_TRAMP_MAX; kind++) { 207 tprogs[kind].nr_progs = tr->progs_cnt[kind]; 208 *total += tr->progs_cnt[kind]; 209 progs = tprogs[kind].progs; 210 211 hlist_for_each_entry(aux, &tr->progs_hlist[kind], tramp_hlist) 212 *progs++ = aux->prog; 213 } 214 return tprogs; 215 } 216 217 static int bpf_trampoline_update(struct bpf_trampoline *tr) 218 { 219 void *old_image = tr->image + ((tr->selector + 1) & 1) * BPF_IMAGE_SIZE/2; 220 void *new_image = tr->image + (tr->selector & 1) * BPF_IMAGE_SIZE/2; 221 struct bpf_tramp_progs *tprogs; 222 u32 flags = BPF_TRAMP_F_RESTORE_REGS; 223 int err, total; 224 225 tprogs = bpf_trampoline_get_progs(tr, &total); 226 if (IS_ERR(tprogs)) 227 return PTR_ERR(tprogs); 228 229 if (total == 0) { 230 err = unregister_fentry(tr, old_image); 231 tr->selector = 0; 232 goto out; 233 } 234 235 if (tprogs[BPF_TRAMP_FEXIT].nr_progs || 236 tprogs[BPF_TRAMP_MODIFY_RETURN].nr_progs) 237 flags = BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME; 238 239 /* Though the second half of trampoline page is unused a task could be 240 * preempted in the middle of the first half of trampoline and two 241 * updates to trampoline would change the code from underneath the 242 * preempted task. Hence wait for tasks to voluntarily schedule or go 243 * to userspace. 244 */ 245 246 synchronize_rcu_tasks(); 247 248 err = arch_prepare_bpf_trampoline(new_image, new_image + BPF_IMAGE_SIZE / 2, 249 &tr->func.model, flags, tprogs, 250 tr->func.addr); 251 if (err < 0) 252 goto out; 253 254 if (tr->selector) 255 /* progs already running at this address */ 256 err = modify_fentry(tr, old_image, new_image); 257 else 258 /* first time registering */ 259 err = register_fentry(tr, new_image); 260 if (err) 261 goto out; 262 tr->selector++; 263 out: 264 kfree(tprogs); 265 return err; 266 } 267 268 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(enum bpf_attach_type t) 269 { 270 switch (t) { 271 case BPF_TRACE_FENTRY: 272 return BPF_TRAMP_FENTRY; 273 case BPF_MODIFY_RETURN: 274 return BPF_TRAMP_MODIFY_RETURN; 275 case BPF_TRACE_FEXIT: 276 return BPF_TRAMP_FEXIT; 277 default: 278 return BPF_TRAMP_REPLACE; 279 } 280 } 281 282 int bpf_trampoline_link_prog(struct bpf_prog *prog) 283 { 284 enum bpf_tramp_prog_type kind; 285 struct bpf_trampoline *tr; 286 int err = 0; 287 int cnt; 288 289 tr = prog->aux->trampoline; 290 kind = bpf_attach_type_to_tramp(prog->expected_attach_type); 291 mutex_lock(&tr->mutex); 292 if (tr->extension_prog) { 293 /* cannot attach fentry/fexit if extension prog is attached. 294 * cannot overwrite extension prog either. 295 */ 296 err = -EBUSY; 297 goto out; 298 } 299 cnt = tr->progs_cnt[BPF_TRAMP_FENTRY] + tr->progs_cnt[BPF_TRAMP_FEXIT]; 300 if (kind == BPF_TRAMP_REPLACE) { 301 /* Cannot attach extension if fentry/fexit are in use. */ 302 if (cnt) { 303 err = -EBUSY; 304 goto out; 305 } 306 tr->extension_prog = prog; 307 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL, 308 prog->bpf_func); 309 goto out; 310 } 311 if (cnt >= BPF_MAX_TRAMP_PROGS) { 312 err = -E2BIG; 313 goto out; 314 } 315 if (!hlist_unhashed(&prog->aux->tramp_hlist)) { 316 /* prog already linked */ 317 err = -EBUSY; 318 goto out; 319 } 320 hlist_add_head(&prog->aux->tramp_hlist, &tr->progs_hlist[kind]); 321 tr->progs_cnt[kind]++; 322 err = bpf_trampoline_update(prog->aux->trampoline); 323 if (err) { 324 hlist_del(&prog->aux->tramp_hlist); 325 tr->progs_cnt[kind]--; 326 } 327 out: 328 mutex_unlock(&tr->mutex); 329 return err; 330 } 331 332 /* bpf_trampoline_unlink_prog() should never fail. */ 333 int bpf_trampoline_unlink_prog(struct bpf_prog *prog) 334 { 335 enum bpf_tramp_prog_type kind; 336 struct bpf_trampoline *tr; 337 int err; 338 339 tr = prog->aux->trampoline; 340 kind = bpf_attach_type_to_tramp(prog->expected_attach_type); 341 mutex_lock(&tr->mutex); 342 if (kind == BPF_TRAMP_REPLACE) { 343 WARN_ON_ONCE(!tr->extension_prog); 344 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, 345 tr->extension_prog->bpf_func, NULL); 346 tr->extension_prog = NULL; 347 goto out; 348 } 349 hlist_del(&prog->aux->tramp_hlist); 350 tr->progs_cnt[kind]--; 351 err = bpf_trampoline_update(prog->aux->trampoline); 352 out: 353 mutex_unlock(&tr->mutex); 354 return err; 355 } 356 357 void bpf_trampoline_put(struct bpf_trampoline *tr) 358 { 359 struct bpf_image *image; 360 361 if (!tr) 362 return; 363 mutex_lock(&trampoline_mutex); 364 if (!refcount_dec_and_test(&tr->refcnt)) 365 goto out; 366 WARN_ON_ONCE(mutex_is_locked(&tr->mutex)); 367 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FENTRY]))) 368 goto out; 369 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FEXIT]))) 370 goto out; 371 image = container_of(tr->image, struct bpf_image, data); 372 latch_tree_erase(&image->tnode, &image_tree, &image_tree_ops); 373 /* wait for tasks to get out of trampoline before freeing it */ 374 synchronize_rcu_tasks(); 375 bpf_jit_free_exec(image); 376 hlist_del(&tr->hlist); 377 kfree(tr); 378 out: 379 mutex_unlock(&trampoline_mutex); 380 } 381 382 /* The logic is similar to BPF_PROG_RUN, but with an explicit 383 * rcu_read_lock() and migrate_disable() which are required 384 * for the trampoline. The macro is split into 385 * call _bpf_prog_enter 386 * call prog->bpf_func 387 * call __bpf_prog_exit 388 */ 389 u64 notrace __bpf_prog_enter(void) 390 { 391 u64 start = 0; 392 393 rcu_read_lock(); 394 migrate_disable(); 395 if (static_branch_unlikely(&bpf_stats_enabled_key)) 396 start = sched_clock(); 397 return start; 398 } 399 400 void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start) 401 { 402 struct bpf_prog_stats *stats; 403 404 if (static_branch_unlikely(&bpf_stats_enabled_key) && 405 /* static_key could be enabled in __bpf_prog_enter 406 * and disabled in __bpf_prog_exit. 407 * And vice versa. 408 * Hence check that 'start' is not zero. 409 */ 410 start) { 411 stats = this_cpu_ptr(prog->aux->stats); 412 u64_stats_update_begin(&stats->syncp); 413 stats->cnt++; 414 stats->nsecs += sched_clock() - start; 415 u64_stats_update_end(&stats->syncp); 416 } 417 migrate_enable(); 418 rcu_read_unlock(); 419 } 420 421 int __weak 422 arch_prepare_bpf_trampoline(void *image, void *image_end, 423 const struct btf_func_model *m, u32 flags, 424 struct bpf_tramp_progs *tprogs, 425 void *orig_call) 426 { 427 return -ENOTSUPP; 428 } 429 430 static int __init init_trampolines(void) 431 { 432 int i; 433 434 for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++) 435 INIT_HLIST_HEAD(&trampoline_table[i]); 436 return 0; 437 } 438 late_initcall(init_trampolines); 439