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