kernel/bpf/trampoline.c

   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, struct bpf_ksym *ksym)
 119 {
 120         ksym->start = (unsigned long) data;
 121         ksym->end = ksym->start + PAGE_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_put_deferred(struct work_struct *work)
 255 {
 256         struct bpf_tramp_image *im;
 257
 258         im = container_of(work, struct bpf_tramp_image, work);
 259         bpf_image_ksym_del(&im->ksym);
 260         bpf_jit_free_exec(im->image);
 261         bpf_jit_uncharge_modmem(PAGE_SIZE);
 262         percpu_ref_exit(&im->pcref);
 263         kfree_rcu(im, rcu);
 264 }
 265
 266 /* callback, fexit step 3 or fentry step 2 */
 267 static void __bpf_tramp_image_put_rcu(struct rcu_head *rcu)
 268 {
 269         struct bpf_tramp_image *im;
 270
 271         im = container_of(rcu, struct bpf_tramp_image, rcu);
 272         INIT_WORK(&im->work, __bpf_tramp_image_put_deferred);
 273         schedule_work(&im->work);
 274 }
 275
 276 /* callback, fexit step 2. Called after percpu_ref_kill confirms. */
 277 static void __bpf_tramp_image_release(struct percpu_ref *pcref)
 278 {
 279         struct bpf_tramp_image *im;
 280
 281         im = container_of(pcref, struct bpf_tramp_image, pcref);
 282         call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
 283 }
 284
 285 /* callback, fexit or fentry step 1 */
 286 static void __bpf_tramp_image_put_rcu_tasks(struct rcu_head *rcu)
 287 {
 288         struct bpf_tramp_image *im;
 289
 290         im = container_of(rcu, struct bpf_tramp_image, rcu);
 291         if (im->ip_after_call)
 292                 /* the case of fmod_ret/fexit trampoline and CONFIG_PREEMPTION=y */
 293                 percpu_ref_kill(&im->pcref);
 294         else
 295                 /* the case of fentry trampoline */
 296                 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
 297 }
 298
 299 static void bpf_tramp_image_put(struct bpf_tramp_image *im)
 300 {
 301         /* The trampoline image that calls original function is using:
 302          * rcu_read_lock_trace to protect sleepable bpf progs
 303          * rcu_read_lock to protect normal bpf progs
 304          * percpu_ref to protect trampoline itself
 305          * rcu tasks to protect trampoline asm not covered by percpu_ref
 306          * (which are few asm insns before __bpf_tramp_enter and
 307          *  after __bpf_tramp_exit)
 308          *
 309          * The trampoline is unreachable before bpf_tramp_image_put().
 310          *
 311          * First, patch the trampoline to avoid calling into fexit progs.
 312          * The progs will be freed even if the original function is still
 313          * executing or sleeping.
 314          * In case of CONFIG_PREEMPT=y use call_rcu_tasks() to wait on
 315          * first few asm instructions to execute and call into
 316          * __bpf_tramp_enter->percpu_ref_get.
 317          * Then use percpu_ref_kill to wait for the trampoline and the original
 318          * function to finish.
 319          * Then use call_rcu_tasks() to make sure few asm insns in
 320          * the trampoline epilogue are done as well.
 321          *
 322          * In !PREEMPT case the task that got interrupted in the first asm
 323          * insns won't go through an RCU quiescent state which the
 324          * percpu_ref_kill will be waiting for. Hence the first
 325          * call_rcu_tasks() is not necessary.
 326          */
 327         if (im->ip_after_call) {
 328                 int err = bpf_arch_text_poke(im->ip_after_call, BPF_MOD_JUMP,
 329                                              NULL, im->ip_epilogue);
 330                 WARN_ON(err);
 331                 if (IS_ENABLED(CONFIG_PREEMPTION))
 332                         call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
 333                 else
 334                         percpu_ref_kill(&im->pcref);
 335                 return;
 336         }
 337
 338         /* The trampoline without fexit and fmod_ret progs doesn't call original
 339          * function and doesn't use percpu_ref.
 340          * Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
 341          * Then use call_rcu_tasks() to wait for the rest of trampoline asm
 342          * and normal progs.
 343          */
 344         call_rcu_tasks_trace(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
 345 }
 346
 347 static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, u32 idx)
 348 {
 349         struct bpf_tramp_image *im;
 350         struct bpf_ksym *ksym;
 351         void *image;
 352         int err = -ENOMEM;
 353
 354         im = kzalloc(sizeof(*im), GFP_KERNEL);
 355         if (!im)
 356                 goto out;
 357
 358         err = bpf_jit_charge_modmem(PAGE_SIZE);
 359         if (err)
 360                 goto out_free_im;
 361
 362         err = -ENOMEM;
 363         im->image = image = bpf_jit_alloc_exec(PAGE_SIZE);
 364         if (!image)
 365                 goto out_uncharge;
 366         set_vm_flush_reset_perms(image);
 367
 368         err = percpu_ref_init(&im->pcref, __bpf_tramp_image_release, 0, GFP_KERNEL);
 369         if (err)
 370                 goto out_free_image;
 371
 372         ksym = &im->ksym;
 373         INIT_LIST_HEAD_RCU(&ksym->lnode);
 374         snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu_%u", key, idx);
 375         bpf_image_ksym_add(image, ksym);
 376         return im;
 377
 378 out_free_image:
 379         bpf_jit_free_exec(im->image);
 380 out_uncharge:
 381         bpf_jit_uncharge_modmem(PAGE_SIZE);
 382 out_free_im:
 383         kfree(im);
 384 out:
 385         return ERR_PTR(err);
 386 }
 387
 388 static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mutex)
 389 {
 390         struct bpf_tramp_image *im;
 391         struct bpf_tramp_links *tlinks;
 392         u32 orig_flags = tr->flags;
 393         bool ip_arg = false;
 394         int err, total;
 395
 396         tlinks = bpf_trampoline_get_progs(tr, &total, &ip_arg);
 397         if (IS_ERR(tlinks))
 398                 return PTR_ERR(tlinks);
 399
 400         if (total == 0) {
 401                 err = unregister_fentry(tr, tr->cur_image->image);
 402                 bpf_tramp_image_put(tr->cur_image);
 403                 tr->cur_image = NULL;
 404                 tr->selector = 0;
 405                 goto out;
 406         }
 407
 408         im = bpf_tramp_image_alloc(tr->key, tr->selector);
 409         if (IS_ERR(im)) {
 410                 err = PTR_ERR(im);
 411                 goto out;
 412         }
 413
 414         /* clear all bits except SHARE_IPMODIFY */
 415         tr->flags &= BPF_TRAMP_F_SHARE_IPMODIFY;
 416
 417         if (tlinks[BPF_TRAMP_FEXIT].nr_links ||
 418             tlinks[BPF_TRAMP_MODIFY_RETURN].nr_links) {
 419                 /* NOTE: BPF_TRAMP_F_RESTORE_REGS and BPF_TRAMP_F_SKIP_FRAME
 420                  * should not be set together.
 421                  */
 422                 tr->flags |= BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
 423         } else {
 424                 tr->flags |= BPF_TRAMP_F_RESTORE_REGS;
 425         }
 426
 427         if (ip_arg)
 428                 tr->flags |= BPF_TRAMP_F_IP_ARG;
 429
 430 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 431 again:
 432         if ((tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY) &&
 433             (tr->flags & BPF_TRAMP_F_CALL_ORIG))
 434                 tr->flags |= BPF_TRAMP_F_ORIG_STACK;
 435 #endif
 436
 437         err = arch_prepare_bpf_trampoline(im, im->image, im->image + PAGE_SIZE,
 438                                           &tr->func.model, tr->flags, tlinks,
 439                                           tr->func.addr);
 440         if (err < 0)
 441                 goto out;
 442
 443         set_memory_rox((long)im->image, 1);
 444
 445         WARN_ON(tr->cur_image && tr->selector == 0);
 446         WARN_ON(!tr->cur_image && tr->selector);
 447         if (tr->cur_image)
 448                 /* progs already running at this address */
 449                 err = modify_fentry(tr, tr->cur_image->image, im->image, lock_direct_mutex);
 450         else
 451                 /* first time registering */
 452                 err = register_fentry(tr, im->image);
 453
 454 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 455         if (err == -EAGAIN) {
 456                 /* -EAGAIN from bpf_tramp_ftrace_ops_func. Now
 457                  * BPF_TRAMP_F_SHARE_IPMODIFY is set, we can generate the
 458                  * trampoline again, and retry register.
 459                  */
 460                 /* reset fops->func and fops->trampoline for re-register */
 461                 tr->fops->func = NULL;
 462                 tr->fops->trampoline = 0;
 463
 464                 /* reset im->image memory attr for arch_prepare_bpf_trampoline */
 465                 set_memory_nx((long)im->image, 1);
 466                 set_memory_rw((long)im->image, 1);
 467                 goto again;
 468         }
 469 #endif
 470         if (err)
 471                 goto out;
 472
 473         if (tr->cur_image)
 474                 bpf_tramp_image_put(tr->cur_image);
 475         tr->cur_image = im;
 476         tr->selector++;
 477 out:
 478         /* If any error happens, restore previous flags */
 479         if (err)
 480                 tr->flags = orig_flags;
 481         kfree(tlinks);
 482         return err;
 483 }
 484
 485 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
 486 {
 487         switch (prog->expected_attach_type) {
 488         case BPF_TRACE_FENTRY:
 489                 return BPF_TRAMP_FENTRY;
 490         case BPF_MODIFY_RETURN:
 491                 return BPF_TRAMP_MODIFY_RETURN;
 492         case BPF_TRACE_FEXIT:
 493                 return BPF_TRAMP_FEXIT;
 494         case BPF_LSM_MAC:
 495                 if (!prog->aux->attach_func_proto->type)
 496                         /* The function returns void, we cannot modify its
 497                          * return value.
 498                          */
 499                         return BPF_TRAMP_FEXIT;
 500                 else
 501                         return BPF_TRAMP_MODIFY_RETURN;
 502         default:
 503                 return BPF_TRAMP_REPLACE;
 504         }
 505 }
 506
 507 static int __bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 508 {
 509         enum bpf_tramp_prog_type kind;
 510         struct bpf_tramp_link *link_exiting;
 511         int err = 0;
 512         int cnt = 0, i;
 513
 514         kind = bpf_attach_type_to_tramp(link->link.prog);
 515         if (tr->extension_prog)
 516                 /* cannot attach fentry/fexit if extension prog is attached.
 517                  * cannot overwrite extension prog either.
 518                  */
 519                 return -EBUSY;
 520
 521         for (i = 0; i < BPF_TRAMP_MAX; i++)
 522                 cnt += tr->progs_cnt[i];
 523
 524         if (kind == BPF_TRAMP_REPLACE) {
 525                 /* Cannot attach extension if fentry/fexit are in use. */
 526                 if (cnt)
 527                         return -EBUSY;
 528                 tr->extension_prog = link->link.prog;
 529                 return bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
 530                                           link->link.prog->bpf_func);
 531         }
 532         if (cnt >= BPF_MAX_TRAMP_LINKS)
 533                 return -E2BIG;
 534         if (!hlist_unhashed(&link->tramp_hlist))
 535                 /* prog already linked */
 536                 return -EBUSY;
 537         hlist_for_each_entry(link_exiting, &tr->progs_hlist[kind], tramp_hlist) {
 538                 if (link_exiting->link.prog != link->link.prog)
 539                         continue;
 540                 /* prog already linked */
 541                 return -EBUSY;
 542         }
 543
 544         hlist_add_head(&link->tramp_hlist, &tr->progs_hlist[kind]);
 545         tr->progs_cnt[kind]++;
 546         err = bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 547         if (err) {
 548                 hlist_del_init(&link->tramp_hlist);
 549                 tr->progs_cnt[kind]--;
 550         }
 551         return err;
 552 }
 553
 554 int bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 555 {
 556         int err;
 557
 558         mutex_lock(&tr->mutex);
 559         err = __bpf_trampoline_link_prog(link, tr);
 560         mutex_unlock(&tr->mutex);
 561         return err;
 562 }
 563
 564 static int __bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 565 {
 566         enum bpf_tramp_prog_type kind;
 567         int err;
 568
 569         kind = bpf_attach_type_to_tramp(link->link.prog);
 570         if (kind == BPF_TRAMP_REPLACE) {
 571                 WARN_ON_ONCE(!tr->extension_prog);
 572                 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
 573                                          tr->extension_prog->bpf_func, NULL);
 574                 tr->extension_prog = NULL;
 575                 return err;
 576         }
 577         hlist_del_init(&link->tramp_hlist);
 578         tr->progs_cnt[kind]--;
 579         return bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 580 }
 581
 582 /* bpf_trampoline_unlink_prog() should never fail. */
 583 int bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 584 {
 585         int err;
 586
 587         mutex_lock(&tr->mutex);
 588         err = __bpf_trampoline_unlink_prog(link, tr);
 589         mutex_unlock(&tr->mutex);
 590         return err;
 591 }
 592
 593 #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_LSM)
 594 static void bpf_shim_tramp_link_release(struct bpf_link *link)
 595 {
 596         struct bpf_shim_tramp_link *shim_link =
 597                 container_of(link, struct bpf_shim_tramp_link, link.link);
 598
 599         /* paired with 'shim_link->trampoline = tr' in bpf_trampoline_link_cgroup_shim */
 600         if (!shim_link->trampoline)
 601                 return;
 602
 603         WARN_ON_ONCE(bpf_trampoline_unlink_prog(&shim_link->link, shim_link->trampoline));
 604         bpf_trampoline_put(shim_link->trampoline);
 605 }
 606
 607 static void bpf_shim_tramp_link_dealloc(struct bpf_link *link)
 608 {
 609         struct bpf_shim_tramp_link *shim_link =
 610                 container_of(link, struct bpf_shim_tramp_link, link.link);
 611
 612         kfree(shim_link);
 613 }
 614
 615 static const struct bpf_link_ops bpf_shim_tramp_link_lops = {
 616         .release = bpf_shim_tramp_link_release,
 617         .dealloc = bpf_shim_tramp_link_dealloc,
 618 };
 619
 620 static struct bpf_shim_tramp_link *cgroup_shim_alloc(const struct bpf_prog *prog,
 621                                                      bpf_func_t bpf_func,
 622                                                      int cgroup_atype)
 623 {
 624         struct bpf_shim_tramp_link *shim_link = NULL;
 625         struct bpf_prog *p;
 626
 627         shim_link = kzalloc(sizeof(*shim_link), GFP_USER);
 628         if (!shim_link)
 629                 return NULL;
 630
 631         p = bpf_prog_alloc(1, 0);
 632         if (!p) {
 633                 kfree(shim_link);
 634                 return NULL;
 635         }
 636
 637         p->jited = false;
 638         p->bpf_func = bpf_func;
 639
 640         p->aux->cgroup_atype = cgroup_atype;
 641         p->aux->attach_func_proto = prog->aux->attach_func_proto;
 642         p->aux->attach_btf_id = prog->aux->attach_btf_id;
 643         p->aux->attach_btf = prog->aux->attach_btf;
 644         btf_get(p->aux->attach_btf);
 645         p->type = BPF_PROG_TYPE_LSM;
 646         p->expected_attach_type = BPF_LSM_MAC;
 647         bpf_prog_inc(p);
 648         bpf_link_init(&shim_link->link.link, BPF_LINK_TYPE_UNSPEC,
 649                       &bpf_shim_tramp_link_lops, p);
 650         bpf_cgroup_atype_get(p->aux->attach_btf_id, cgroup_atype);
 651
 652         return shim_link;
 653 }
 654
 655 static struct bpf_shim_tramp_link *cgroup_shim_find(struct bpf_trampoline *tr,
 656                                                     bpf_func_t bpf_func)
 657 {
 658         struct bpf_tramp_link *link;
 659         int kind;
 660
 661         for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
 662                 hlist_for_each_entry(link, &tr->progs_hlist[kind], tramp_hlist) {
 663                         struct bpf_prog *p = link->link.prog;
 664
 665                         if (p->bpf_func == bpf_func)
 666                                 return container_of(link, struct bpf_shim_tramp_link, link);
 667                 }
 668         }
 669
 670         return NULL;
 671 }
 672
 673 int bpf_trampoline_link_cgroup_shim(struct bpf_prog *prog,
 674                                     int cgroup_atype)
 675 {
 676         struct bpf_shim_tramp_link *shim_link = NULL;
 677         struct bpf_attach_target_info tgt_info = {};
 678         struct bpf_trampoline *tr;
 679         bpf_func_t bpf_func;
 680         u64 key;
 681         int err;
 682
 683         err = bpf_check_attach_target(NULL, prog, NULL,
 684                                       prog->aux->attach_btf_id,
 685                                       &tgt_info);
 686         if (err)
 687                 return err;
 688
 689         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 690                                          prog->aux->attach_btf_id);
 691
 692         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 693         tr = bpf_trampoline_get(key, &tgt_info);
 694         if (!tr)
 695                 return  -ENOMEM;
 696
 697         mutex_lock(&tr->mutex);
 698
 699         shim_link = cgroup_shim_find(tr, bpf_func);
 700         if (shim_link) {
 701                 /* Reusing existing shim attached by the other program. */
 702                 bpf_link_inc(&shim_link->link.link);
 703
 704                 mutex_unlock(&tr->mutex);
 705                 bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 706                 return 0;
 707         }
 708
 709         /* Allocate and install new shim. */
 710
 711         shim_link = cgroup_shim_alloc(prog, bpf_func, cgroup_atype);
 712         if (!shim_link) {
 713                 err = -ENOMEM;
 714                 goto err;
 715         }
 716
 717         err = __bpf_trampoline_link_prog(&shim_link->link, tr);
 718         if (err)
 719                 goto err;
 720
 721         shim_link->trampoline = tr;
 722         /* note, we're still holding tr refcnt from above */
 723
 724         mutex_unlock(&tr->mutex);
 725
 726         return 0;
 727 err:
 728         mutex_unlock(&tr->mutex);
 729
 730         if (shim_link)
 731                 bpf_link_put(&shim_link->link.link);
 732
 733         /* have to release tr while _not_ holding its mutex */
 734         bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 735
 736         return err;
 737 }
 738
 739 void bpf_trampoline_unlink_cgroup_shim(struct bpf_prog *prog)
 740 {
 741         struct bpf_shim_tramp_link *shim_link = NULL;
 742         struct bpf_trampoline *tr;
 743         bpf_func_t bpf_func;
 744         u64 key;
 745
 746         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 747                                          prog->aux->attach_btf_id);
 748
 749         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 750         tr = bpf_trampoline_lookup(key);
 751         if (WARN_ON_ONCE(!tr))
 752                 return;
 753
 754         mutex_lock(&tr->mutex);
 755         shim_link = cgroup_shim_find(tr, bpf_func);
 756         mutex_unlock(&tr->mutex);
 757
 758         if (shim_link)
 759                 bpf_link_put(&shim_link->link.link);
 760
 761         bpf_trampoline_put(tr); /* bpf_trampoline_lookup above */
 762 }
 763 #endif
 764
 765 struct bpf_trampoline *bpf_trampoline_get(u64 key,
 766                                           struct bpf_attach_target_info *tgt_info)
 767 {
 768         struct bpf_trampoline *tr;
 769
 770         tr = bpf_trampoline_lookup(key);
 771         if (!tr)
 772                 return NULL;
 773
 774         mutex_lock(&tr->mutex);
 775         if (tr->func.addr)
 776                 goto out;
 777
 778         memcpy(&tr->func.model, &tgt_info->fmodel, sizeof(tgt_info->fmodel));
 779         tr->func.addr = (void *)tgt_info->tgt_addr;
 780 out:
 781         mutex_unlock(&tr->mutex);
 782         return tr;
 783 }
 784
 785 void bpf_trampoline_put(struct bpf_trampoline *tr)
 786 {
 787         int i;
 788
 789         if (!tr)
 790                 return;
 791         mutex_lock(&trampoline_mutex);
 792         if (!refcount_dec_and_test(&tr->refcnt))
 793                 goto out;
 794         WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
 795
 796         for (i = 0; i < BPF_TRAMP_MAX; i++)
 797                 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[i])))
 798                         goto out;
 799
 800         /* This code will be executed even when the last bpf_tramp_image
 801          * is alive. All progs are detached from the trampoline and the
 802          * trampoline image is patched with jmp into epilogue to skip
 803          * fexit progs. The fentry-only trampoline will be freed via
 804          * multiple rcu callbacks.
 805          */
 806         hlist_del(&tr->hlist);
 807         if (tr->fops) {
 808                 ftrace_free_filter(tr->fops);
 809                 kfree(tr->fops);
 810         }
 811         kfree(tr);
 812 out:
 813         mutex_unlock(&trampoline_mutex);
 814 }
 815
 816 #define NO_START_TIME 1
 817 static __always_inline u64 notrace bpf_prog_start_time(void)
 818 {
 819         u64 start = NO_START_TIME;
 820
 821         if (static_branch_unlikely(&bpf_stats_enabled_key)) {
 822                 start = sched_clock();
 823                 if (unlikely(!start))
 824                         start = NO_START_TIME;
 825         }
 826         return start;
 827 }
 828
 829 /* The logic is similar to bpf_prog_run(), but with an explicit
 830  * rcu_read_lock() and migrate_disable() which are required
 831  * for the trampoline. The macro is split into
 832  * call __bpf_prog_enter
 833  * call prog->bpf_func
 834  * call __bpf_prog_exit
 835  *
 836  * __bpf_prog_enter returns:
 837  * 0 - skip execution of the bpf prog
 838  * 1 - execute bpf prog
 839  * [2..MAX_U64] - execute bpf prog and record execution time.
 840  *     This is start time.
 841  */
 842 static u64 notrace __bpf_prog_enter_recur(struct bpf_prog *prog, struct bpf_tramp_run_ctx *run_ctx)
 843         __acquires(RCU)
 844 {
 845         rcu_read_lock();
 846         migrate_disable();
 847
 848         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 849
 850         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 851                 bpf_prog_inc_misses_counter(prog);
 852                 return 0;
 853         }
 854         return bpf_prog_start_time();
 855 }
 856
 857 static void notrace update_prog_stats(struct bpf_prog *prog,
 858                                       u64 start)
 859 {
 860         struct bpf_prog_stats *stats;
 861
 862         if (static_branch_unlikely(&bpf_stats_enabled_key) &&
 863             /* static_key could be enabled in __bpf_prog_enter*
 864              * and disabled in __bpf_prog_exit*.
 865              * And vice versa.
 866              * Hence check that 'start' is valid.
 867              */
 868             start > NO_START_TIME) {
 869                 unsigned long flags;
 870
 871                 stats = this_cpu_ptr(prog->stats);
 872                 flags = u64_stats_update_begin_irqsave(&stats->syncp);
 873                 u64_stats_inc(&stats->cnt);
 874                 u64_stats_add(&stats->nsecs, sched_clock() - start);
 875                 u64_stats_update_end_irqrestore(&stats->syncp, flags);
 876         }
 877 }
 878
 879 static void notrace __bpf_prog_exit_recur(struct bpf_prog *prog, u64 start,
 880                                           struct bpf_tramp_run_ctx *run_ctx)
 881         __releases(RCU)
 882 {
 883         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 884
 885         update_prog_stats(prog, start);
 886         this_cpu_dec(*(prog->active));
 887         migrate_enable();
 888         rcu_read_unlock();
 889 }
 890
 891 static u64 notrace __bpf_prog_enter_lsm_cgroup(struct bpf_prog *prog,
 892                                                struct bpf_tramp_run_ctx *run_ctx)
 893         __acquires(RCU)
 894 {
 895         /* Runtime stats are exported via actual BPF_LSM_CGROUP
 896          * programs, not the shims.
 897          */
 898         rcu_read_lock();
 899         migrate_disable();
 900
 901         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 902
 903         return NO_START_TIME;
 904 }
 905
 906 static void notrace __bpf_prog_exit_lsm_cgroup(struct bpf_prog *prog, u64 start,
 907                                                struct bpf_tramp_run_ctx *run_ctx)
 908         __releases(RCU)
 909 {
 910         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 911
 912         migrate_enable();
 913         rcu_read_unlock();
 914 }
 915
 916 u64 notrace __bpf_prog_enter_sleepable_recur(struct bpf_prog *prog,
 917                                              struct bpf_tramp_run_ctx *run_ctx)
 918 {
 919         rcu_read_lock_trace();
 920         migrate_disable();
 921         might_fault();
 922
 923         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 924                 bpf_prog_inc_misses_counter(prog);
 925                 return 0;
 926         }
 927
 928         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 929
 930         return bpf_prog_start_time();
 931 }
 932
 933 void notrace __bpf_prog_exit_sleepable_recur(struct bpf_prog *prog, u64 start,
 934                                              struct bpf_tramp_run_ctx *run_ctx)
 935 {
 936         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 937
 938         update_prog_stats(prog, start);
 939         this_cpu_dec(*(prog->active));
 940         migrate_enable();
 941         rcu_read_unlock_trace();
 942 }
 943
 944 static u64 notrace __bpf_prog_enter_sleepable(struct bpf_prog *prog,
 945                                               struct bpf_tramp_run_ctx *run_ctx)
 946 {
 947         rcu_read_lock_trace();
 948         migrate_disable();
 949         might_fault();
 950
 951         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 952
 953         return bpf_prog_start_time();
 954 }
 955
 956 static void notrace __bpf_prog_exit_sleepable(struct bpf_prog *prog, u64 start,
 957                                               struct bpf_tramp_run_ctx *run_ctx)
 958 {
 959         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 960
 961         update_prog_stats(prog, start);
 962         migrate_enable();
 963         rcu_read_unlock_trace();
 964 }
 965
 966 static u64 notrace __bpf_prog_enter(struct bpf_prog *prog,
 967                                     struct bpf_tramp_run_ctx *run_ctx)
 968         __acquires(RCU)
 969 {
 970         rcu_read_lock();
 971         migrate_disable();
 972
 973         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 974
 975         return bpf_prog_start_time();
 976 }
 977
 978 static void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start,
 979                                     struct bpf_tramp_run_ctx *run_ctx)
 980         __releases(RCU)
 981 {
 982         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 983
 984         update_prog_stats(prog, start);
 985         migrate_enable();
 986         rcu_read_unlock();
 987 }
 988
 989 void notrace __bpf_tramp_enter(struct bpf_tramp_image *tr)
 990 {
 991         percpu_ref_get(&tr->pcref);
 992 }
 993
 994 void notrace __bpf_tramp_exit(struct bpf_tramp_image *tr)
 995 {
 996         percpu_ref_put(&tr->pcref);
 997 }
 998
 999 bpf_trampoline_enter_t bpf_trampoline_enter(const struct bpf_prog *prog)
1000 {
1001         bool sleepable = prog->aux->sleepable;
1002
1003         if (bpf_prog_check_recur(prog))
1004                 return sleepable ? __bpf_prog_enter_sleepable_recur :
1005                         __bpf_prog_enter_recur;
1006
1007         if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1008             prog->expected_attach_type == BPF_LSM_CGROUP)
1009                 return __bpf_prog_enter_lsm_cgroup;
1010
1011         return sleepable ? __bpf_prog_enter_sleepable : __bpf_prog_enter;
1012 }
1013
1014 bpf_trampoline_exit_t bpf_trampoline_exit(const struct bpf_prog *prog)
1015 {
1016         bool sleepable = prog->aux->sleepable;
1017
1018         if (bpf_prog_check_recur(prog))
1019                 return sleepable ? __bpf_prog_exit_sleepable_recur :
1020                         __bpf_prog_exit_recur;
1021
1022         if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1023             prog->expected_attach_type == BPF_LSM_CGROUP)
1024                 return __bpf_prog_exit_lsm_cgroup;
1025
1026         return sleepable ? __bpf_prog_exit_sleepable : __bpf_prog_exit;
1027 }
1028
1029 int __weak
1030 arch_prepare_bpf_trampoline(struct bpf_tramp_image *tr, void *image, void *image_end,
1031                             const struct btf_func_model *m, u32 flags,
1032                             struct bpf_tramp_links *tlinks,
1033                             void *orig_call)
1034 {
1035         return -ENOTSUPP;
1036 }
1037
1038 static int __init init_trampolines(void)
1039 {
1040         int i;
1041
1042         for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
1043                 INIT_HLIST_HEAD(&trampoline_table[i]);
1044         return 0;
1045 }
1046 late_initcall(init_trampolines);