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(). Then we can
  49                  * retry register_ftrace_direct() 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(tr->fops, (long)old_addr, false);
 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(tr->fops, (long)new_addr);
 219                 else
 220                         ret = modify_ftrace_direct_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(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_put_deferred(struct work_struct *work)
 283 {
 284         struct bpf_tramp_image *im;
 285
 286         im = container_of(work, struct bpf_tramp_image, work);
 287         bpf_image_ksym_del(&im->ksym);
 288         bpf_jit_free_exec(im->image);
 289         bpf_jit_uncharge_modmem(PAGE_SIZE);
 290         percpu_ref_exit(&im->pcref);
 291         kfree_rcu(im, rcu);
 292 }
 293
 294 /* callback, fexit step 3 or fentry step 2 */
 295 static void __bpf_tramp_image_put_rcu(struct rcu_head *rcu)
 296 {
 297         struct bpf_tramp_image *im;
 298
 299         im = container_of(rcu, struct bpf_tramp_image, rcu);
 300         INIT_WORK(&im->work, __bpf_tramp_image_put_deferred);
 301         schedule_work(&im->work);
 302 }
 303
 304 /* callback, fexit step 2. Called after percpu_ref_kill confirms. */
 305 static void __bpf_tramp_image_release(struct percpu_ref *pcref)
 306 {
 307         struct bpf_tramp_image *im;
 308
 309         im = container_of(pcref, struct bpf_tramp_image, pcref);
 310         call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
 311 }
 312
 313 /* callback, fexit or fentry step 1 */
 314 static void __bpf_tramp_image_put_rcu_tasks(struct rcu_head *rcu)
 315 {
 316         struct bpf_tramp_image *im;
 317
 318         im = container_of(rcu, struct bpf_tramp_image, rcu);
 319         if (im->ip_after_call)
 320                 /* the case of fmod_ret/fexit trampoline and CONFIG_PREEMPTION=y */
 321                 percpu_ref_kill(&im->pcref);
 322         else
 323                 /* the case of fentry trampoline */
 324                 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
 325 }
 326
 327 static void bpf_tramp_image_put(struct bpf_tramp_image *im)
 328 {
 329         /* The trampoline image that calls original function is using:
 330          * rcu_read_lock_trace to protect sleepable bpf progs
 331          * rcu_read_lock to protect normal bpf progs
 332          * percpu_ref to protect trampoline itself
 333          * rcu tasks to protect trampoline asm not covered by percpu_ref
 334          * (which are few asm insns before __bpf_tramp_enter and
 335          *  after __bpf_tramp_exit)
 336          *
 337          * The trampoline is unreachable before bpf_tramp_image_put().
 338          *
 339          * First, patch the trampoline to avoid calling into fexit progs.
 340          * The progs will be freed even if the original function is still
 341          * executing or sleeping.
 342          * In case of CONFIG_PREEMPT=y use call_rcu_tasks() to wait on
 343          * first few asm instructions to execute and call into
 344          * __bpf_tramp_enter->percpu_ref_get.
 345          * Then use percpu_ref_kill to wait for the trampoline and the original
 346          * function to finish.
 347          * Then use call_rcu_tasks() to make sure few asm insns in
 348          * the trampoline epilogue are done as well.
 349          *
 350          * In !PREEMPT case the task that got interrupted in the first asm
 351          * insns won't go through an RCU quiescent state which the
 352          * percpu_ref_kill will be waiting for. Hence the first
 353          * call_rcu_tasks() is not necessary.
 354          */
 355         if (im->ip_after_call) {
 356                 int err = bpf_arch_text_poke(im->ip_after_call, BPF_MOD_JUMP,
 357                                              NULL, im->ip_epilogue);
 358                 WARN_ON(err);
 359                 if (IS_ENABLED(CONFIG_PREEMPTION))
 360                         call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
 361                 else
 362                         percpu_ref_kill(&im->pcref);
 363                 return;
 364         }
 365
 366         /* The trampoline without fexit and fmod_ret progs doesn't call original
 367          * function and doesn't use percpu_ref.
 368          * Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
 369          * Then use call_rcu_tasks() to wait for the rest of trampoline asm
 370          * and normal progs.
 371          */
 372         call_rcu_tasks_trace(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
 373 }
 374
 375 static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, u32 idx)
 376 {
 377         struct bpf_tramp_image *im;
 378         struct bpf_ksym *ksym;
 379         void *image;
 380         int err = -ENOMEM;
 381
 382         im = kzalloc(sizeof(*im), GFP_KERNEL);
 383         if (!im)
 384                 goto out;
 385
 386         err = bpf_jit_charge_modmem(PAGE_SIZE);
 387         if (err)
 388                 goto out_free_im;
 389
 390         err = -ENOMEM;
 391         im->image = image = bpf_jit_alloc_exec(PAGE_SIZE);
 392         if (!image)
 393                 goto out_uncharge;
 394         set_vm_flush_reset_perms(image);
 395
 396         err = percpu_ref_init(&im->pcref, __bpf_tramp_image_release, 0, GFP_KERNEL);
 397         if (err)
 398                 goto out_free_image;
 399
 400         ksym = &im->ksym;
 401         INIT_LIST_HEAD_RCU(&ksym->lnode);
 402         snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu_%u", key, idx);
 403         bpf_image_ksym_add(image, ksym);
 404         return im;
 405
 406 out_free_image:
 407         bpf_jit_free_exec(im->image);
 408 out_uncharge:
 409         bpf_jit_uncharge_modmem(PAGE_SIZE);
 410 out_free_im:
 411         kfree(im);
 412 out:
 413         return ERR_PTR(err);
 414 }
 415
 416 static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mutex)
 417 {
 418         struct bpf_tramp_image *im;
 419         struct bpf_tramp_links *tlinks;
 420         u32 orig_flags = tr->flags;
 421         bool ip_arg = false;
 422         int err, total;
 423
 424         tlinks = bpf_trampoline_get_progs(tr, &total, &ip_arg);
 425         if (IS_ERR(tlinks))
 426                 return PTR_ERR(tlinks);
 427
 428         if (total == 0) {
 429                 err = unregister_fentry(tr, tr->cur_image->image);
 430                 bpf_tramp_image_put(tr->cur_image);
 431                 tr->cur_image = NULL;
 432                 tr->selector = 0;
 433                 goto out;
 434         }
 435
 436         im = bpf_tramp_image_alloc(tr->key, tr->selector);
 437         if (IS_ERR(im)) {
 438                 err = PTR_ERR(im);
 439                 goto out;
 440         }
 441
 442         /* clear all bits except SHARE_IPMODIFY */
 443         tr->flags &= BPF_TRAMP_F_SHARE_IPMODIFY;
 444
 445         if (tlinks[BPF_TRAMP_FEXIT].nr_links ||
 446             tlinks[BPF_TRAMP_MODIFY_RETURN].nr_links) {
 447                 /* NOTE: BPF_TRAMP_F_RESTORE_REGS and BPF_TRAMP_F_SKIP_FRAME
 448                  * should not be set together.
 449                  */
 450                 tr->flags |= BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
 451         } else {
 452                 tr->flags |= BPF_TRAMP_F_RESTORE_REGS;
 453         }
 454
 455         if (ip_arg)
 456                 tr->flags |= BPF_TRAMP_F_IP_ARG;
 457
 458 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 459 again:
 460         if ((tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY) &&
 461             (tr->flags & BPF_TRAMP_F_CALL_ORIG))
 462                 tr->flags |= BPF_TRAMP_F_ORIG_STACK;
 463 #endif
 464
 465         err = arch_prepare_bpf_trampoline(im, im->image, im->image + PAGE_SIZE,
 466                                           &tr->func.model, tr->flags, tlinks,
 467                                           tr->func.addr);
 468         if (err < 0)
 469                 goto out;
 470
 471         set_memory_rox((long)im->image, 1);
 472
 473         WARN_ON(tr->cur_image && tr->selector == 0);
 474         WARN_ON(!tr->cur_image && tr->selector);
 475         if (tr->cur_image)
 476                 /* progs already running at this address */
 477                 err = modify_fentry(tr, tr->cur_image->image, im->image, lock_direct_mutex);
 478         else
 479                 /* first time registering */
 480                 err = register_fentry(tr, im->image);
 481
 482 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 483         if (err == -EAGAIN) {
 484                 /* -EAGAIN from bpf_tramp_ftrace_ops_func. Now
 485                  * BPF_TRAMP_F_SHARE_IPMODIFY is set, we can generate the
 486                  * trampoline again, and retry register.
 487                  */
 488                 /* reset fops->func and fops->trampoline for re-register */
 489                 tr->fops->func = NULL;
 490                 tr->fops->trampoline = 0;
 491
 492                 /* reset im->image memory attr for arch_prepare_bpf_trampoline */
 493                 set_memory_nx((long)im->image, 1);
 494                 set_memory_rw((long)im->image, 1);
 495                 goto again;
 496         }
 497 #endif
 498         if (err)
 499                 goto out;
 500
 501         if (tr->cur_image)
 502                 bpf_tramp_image_put(tr->cur_image);
 503         tr->cur_image = im;
 504         tr->selector++;
 505 out:
 506         /* If any error happens, restore previous flags */
 507         if (err)
 508                 tr->flags = orig_flags;
 509         kfree(tlinks);
 510         return err;
 511 }
 512
 513 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
 514 {
 515         switch (prog->expected_attach_type) {
 516         case BPF_TRACE_FENTRY:
 517                 return BPF_TRAMP_FENTRY;
 518         case BPF_MODIFY_RETURN:
 519                 return BPF_TRAMP_MODIFY_RETURN;
 520         case BPF_TRACE_FEXIT:
 521                 return BPF_TRAMP_FEXIT;
 522         case BPF_LSM_MAC:
 523                 if (!prog->aux->attach_func_proto->type)
 524                         /* The function returns void, we cannot modify its
 525                          * return value.
 526                          */
 527                         return BPF_TRAMP_FEXIT;
 528                 else
 529                         return BPF_TRAMP_MODIFY_RETURN;
 530         default:
 531                 return BPF_TRAMP_REPLACE;
 532         }
 533 }
 534
 535 static int __bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 536 {
 537         enum bpf_tramp_prog_type kind;
 538         struct bpf_tramp_link *link_exiting;
 539         int err = 0;
 540         int cnt = 0, i;
 541
 542         kind = bpf_attach_type_to_tramp(link->link.prog);
 543         if (tr->extension_prog)
 544                 /* cannot attach fentry/fexit if extension prog is attached.
 545                  * cannot overwrite extension prog either.
 546                  */
 547                 return -EBUSY;
 548
 549         for (i = 0; i < BPF_TRAMP_MAX; i++)
 550                 cnt += tr->progs_cnt[i];
 551
 552         if (kind == BPF_TRAMP_REPLACE) {
 553                 /* Cannot attach extension if fentry/fexit are in use. */
 554                 if (cnt)
 555                         return -EBUSY;
 556                 tr->extension_prog = link->link.prog;
 557                 return bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
 558                                           link->link.prog->bpf_func);
 559         }
 560         if (cnt >= BPF_MAX_TRAMP_LINKS)
 561                 return -E2BIG;
 562         if (!hlist_unhashed(&link->tramp_hlist))
 563                 /* prog already linked */
 564                 return -EBUSY;
 565         hlist_for_each_entry(link_exiting, &tr->progs_hlist[kind], tramp_hlist) {
 566                 if (link_exiting->link.prog != link->link.prog)
 567                         continue;
 568                 /* prog already linked */
 569                 return -EBUSY;
 570         }
 571
 572         hlist_add_head(&link->tramp_hlist, &tr->progs_hlist[kind]);
 573         tr->progs_cnt[kind]++;
 574         err = bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 575         if (err) {
 576                 hlist_del_init(&link->tramp_hlist);
 577                 tr->progs_cnt[kind]--;
 578         }
 579         return err;
 580 }
 581
 582 int bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 583 {
 584         int err;
 585
 586         mutex_lock(&tr->mutex);
 587         err = __bpf_trampoline_link_prog(link, tr);
 588         mutex_unlock(&tr->mutex);
 589         return err;
 590 }
 591
 592 static int __bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 593 {
 594         enum bpf_tramp_prog_type kind;
 595         int err;
 596
 597         kind = bpf_attach_type_to_tramp(link->link.prog);
 598         if (kind == BPF_TRAMP_REPLACE) {
 599                 WARN_ON_ONCE(!tr->extension_prog);
 600                 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
 601                                          tr->extension_prog->bpf_func, NULL);
 602                 tr->extension_prog = NULL;
 603                 return err;
 604         }
 605         hlist_del_init(&link->tramp_hlist);
 606         tr->progs_cnt[kind]--;
 607         return bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 608 }
 609
 610 /* bpf_trampoline_unlink_prog() should never fail. */
 611 int bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 612 {
 613         int err;
 614
 615         mutex_lock(&tr->mutex);
 616         err = __bpf_trampoline_unlink_prog(link, tr);
 617         mutex_unlock(&tr->mutex);
 618         return err;
 619 }
 620
 621 #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_LSM)
 622 static void bpf_shim_tramp_link_release(struct bpf_link *link)
 623 {
 624         struct bpf_shim_tramp_link *shim_link =
 625                 container_of(link, struct bpf_shim_tramp_link, link.link);
 626
 627         /* paired with 'shim_link->trampoline = tr' in bpf_trampoline_link_cgroup_shim */
 628         if (!shim_link->trampoline)
 629                 return;
 630
 631         WARN_ON_ONCE(bpf_trampoline_unlink_prog(&shim_link->link, shim_link->trampoline));
 632         bpf_trampoline_put(shim_link->trampoline);
 633 }
 634
 635 static void bpf_shim_tramp_link_dealloc(struct bpf_link *link)
 636 {
 637         struct bpf_shim_tramp_link *shim_link =
 638                 container_of(link, struct bpf_shim_tramp_link, link.link);
 639
 640         kfree(shim_link);
 641 }
 642
 643 static const struct bpf_link_ops bpf_shim_tramp_link_lops = {
 644         .release = bpf_shim_tramp_link_release,
 645         .dealloc = bpf_shim_tramp_link_dealloc,
 646 };
 647
 648 static struct bpf_shim_tramp_link *cgroup_shim_alloc(const struct bpf_prog *prog,
 649                                                      bpf_func_t bpf_func,
 650                                                      int cgroup_atype)
 651 {
 652         struct bpf_shim_tramp_link *shim_link = NULL;
 653         struct bpf_prog *p;
 654
 655         shim_link = kzalloc(sizeof(*shim_link), GFP_USER);
 656         if (!shim_link)
 657                 return NULL;
 658
 659         p = bpf_prog_alloc(1, 0);
 660         if (!p) {
 661                 kfree(shim_link);
 662                 return NULL;
 663         }
 664
 665         p->jited = false;
 666         p->bpf_func = bpf_func;
 667
 668         p->aux->cgroup_atype = cgroup_atype;
 669         p->aux->attach_func_proto = prog->aux->attach_func_proto;
 670         p->aux->attach_btf_id = prog->aux->attach_btf_id;
 671         p->aux->attach_btf = prog->aux->attach_btf;
 672         btf_get(p->aux->attach_btf);
 673         p->type = BPF_PROG_TYPE_LSM;
 674         p->expected_attach_type = BPF_LSM_MAC;
 675         bpf_prog_inc(p);
 676         bpf_link_init(&shim_link->link.link, BPF_LINK_TYPE_UNSPEC,
 677                       &bpf_shim_tramp_link_lops, p);
 678         bpf_cgroup_atype_get(p->aux->attach_btf_id, cgroup_atype);
 679
 680         return shim_link;
 681 }
 682
 683 static struct bpf_shim_tramp_link *cgroup_shim_find(struct bpf_trampoline *tr,
 684                                                     bpf_func_t bpf_func)
 685 {
 686         struct bpf_tramp_link *link;
 687         int kind;
 688
 689         for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
 690                 hlist_for_each_entry(link, &tr->progs_hlist[kind], tramp_hlist) {
 691                         struct bpf_prog *p = link->link.prog;
 692
 693                         if (p->bpf_func == bpf_func)
 694                                 return container_of(link, struct bpf_shim_tramp_link, link);
 695                 }
 696         }
 697
 698         return NULL;
 699 }
 700
 701 int bpf_trampoline_link_cgroup_shim(struct bpf_prog *prog,
 702                                     int cgroup_atype)
 703 {
 704         struct bpf_shim_tramp_link *shim_link = NULL;
 705         struct bpf_attach_target_info tgt_info = {};
 706         struct bpf_trampoline *tr;
 707         bpf_func_t bpf_func;
 708         u64 key;
 709         int err;
 710
 711         err = bpf_check_attach_target(NULL, prog, NULL,
 712                                       prog->aux->attach_btf_id,
 713                                       &tgt_info);
 714         if (err)
 715                 return err;
 716
 717         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 718                                          prog->aux->attach_btf_id);
 719
 720         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 721         tr = bpf_trampoline_get(key, &tgt_info);
 722         if (!tr)
 723                 return  -ENOMEM;
 724
 725         mutex_lock(&tr->mutex);
 726
 727         shim_link = cgroup_shim_find(tr, bpf_func);
 728         if (shim_link) {
 729                 /* Reusing existing shim attached by the other program. */
 730                 bpf_link_inc(&shim_link->link.link);
 731
 732                 mutex_unlock(&tr->mutex);
 733                 bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 734                 return 0;
 735         }
 736
 737         /* Allocate and install new shim. */
 738
 739         shim_link = cgroup_shim_alloc(prog, bpf_func, cgroup_atype);
 740         if (!shim_link) {
 741                 err = -ENOMEM;
 742                 goto err;
 743         }
 744
 745         err = __bpf_trampoline_link_prog(&shim_link->link, tr);
 746         if (err)
 747                 goto err;
 748
 749         shim_link->trampoline = tr;
 750         /* note, we're still holding tr refcnt from above */
 751
 752         mutex_unlock(&tr->mutex);
 753
 754         return 0;
 755 err:
 756         mutex_unlock(&tr->mutex);
 757
 758         if (shim_link)
 759                 bpf_link_put(&shim_link->link.link);
 760
 761         /* have to release tr while _not_ holding its mutex */
 762         bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 763
 764         return err;
 765 }
 766
 767 void bpf_trampoline_unlink_cgroup_shim(struct bpf_prog *prog)
 768 {
 769         struct bpf_shim_tramp_link *shim_link = NULL;
 770         struct bpf_trampoline *tr;
 771         bpf_func_t bpf_func;
 772         u64 key;
 773
 774         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 775                                          prog->aux->attach_btf_id);
 776
 777         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 778         tr = bpf_trampoline_lookup(key);
 779         if (WARN_ON_ONCE(!tr))
 780                 return;
 781
 782         mutex_lock(&tr->mutex);
 783         shim_link = cgroup_shim_find(tr, bpf_func);
 784         mutex_unlock(&tr->mutex);
 785
 786         if (shim_link)
 787                 bpf_link_put(&shim_link->link.link);
 788
 789         bpf_trampoline_put(tr); /* bpf_trampoline_lookup above */
 790 }
 791 #endif
 792
 793 struct bpf_trampoline *bpf_trampoline_get(u64 key,
 794                                           struct bpf_attach_target_info *tgt_info)
 795 {
 796         struct bpf_trampoline *tr;
 797
 798         tr = bpf_trampoline_lookup(key);
 799         if (!tr)
 800                 return NULL;
 801
 802         mutex_lock(&tr->mutex);
 803         if (tr->func.addr)
 804                 goto out;
 805
 806         memcpy(&tr->func.model, &tgt_info->fmodel, sizeof(tgt_info->fmodel));
 807         tr->func.addr = (void *)tgt_info->tgt_addr;
 808 out:
 809         mutex_unlock(&tr->mutex);
 810         return tr;
 811 }
 812
 813 void bpf_trampoline_put(struct bpf_trampoline *tr)
 814 {
 815         int i;
 816
 817         if (!tr)
 818                 return;
 819         mutex_lock(&trampoline_mutex);
 820         if (!refcount_dec_and_test(&tr->refcnt))
 821                 goto out;
 822         WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
 823
 824         for (i = 0; i < BPF_TRAMP_MAX; i++)
 825                 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[i])))
 826                         goto out;
 827
 828         /* This code will be executed even when the last bpf_tramp_image
 829          * is alive. All progs are detached from the trampoline and the
 830          * trampoline image is patched with jmp into epilogue to skip
 831          * fexit progs. The fentry-only trampoline will be freed via
 832          * multiple rcu callbacks.
 833          */
 834         hlist_del(&tr->hlist);
 835         if (tr->fops) {
 836                 ftrace_free_filter(tr->fops);
 837                 kfree(tr->fops);
 838         }
 839         kfree(tr);
 840 out:
 841         mutex_unlock(&trampoline_mutex);
 842 }
 843
 844 #define NO_START_TIME 1
 845 static __always_inline u64 notrace bpf_prog_start_time(void)
 846 {
 847         u64 start = NO_START_TIME;
 848
 849         if (static_branch_unlikely(&bpf_stats_enabled_key)) {
 850                 start = sched_clock();
 851                 if (unlikely(!start))
 852                         start = NO_START_TIME;
 853         }
 854         return start;
 855 }
 856
 857 /* The logic is similar to bpf_prog_run(), but with an explicit
 858  * rcu_read_lock() and migrate_disable() which are required
 859  * for the trampoline. The macro is split into
 860  * call __bpf_prog_enter
 861  * call prog->bpf_func
 862  * call __bpf_prog_exit
 863  *
 864  * __bpf_prog_enter returns:
 865  * 0 - skip execution of the bpf prog
 866  * 1 - execute bpf prog
 867  * [2..MAX_U64] - execute bpf prog and record execution time.
 868  *     This is start time.
 869  */
 870 static u64 notrace __bpf_prog_enter_recur(struct bpf_prog *prog, struct bpf_tramp_run_ctx *run_ctx)
 871         __acquires(RCU)
 872 {
 873         rcu_read_lock();
 874         migrate_disable();
 875
 876         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 877
 878         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 879                 bpf_prog_inc_misses_counter(prog);
 880                 return 0;
 881         }
 882         return bpf_prog_start_time();
 883 }
 884
 885 static void notrace update_prog_stats(struct bpf_prog *prog,
 886                                       u64 start)
 887 {
 888         struct bpf_prog_stats *stats;
 889
 890         if (static_branch_unlikely(&bpf_stats_enabled_key) &&
 891             /* static_key could be enabled in __bpf_prog_enter*
 892              * and disabled in __bpf_prog_exit*.
 893              * And vice versa.
 894              * Hence check that 'start' is valid.
 895              */
 896             start > NO_START_TIME) {
 897                 unsigned long flags;
 898
 899                 stats = this_cpu_ptr(prog->stats);
 900                 flags = u64_stats_update_begin_irqsave(&stats->syncp);
 901                 u64_stats_inc(&stats->cnt);
 902                 u64_stats_add(&stats->nsecs, sched_clock() - start);
 903                 u64_stats_update_end_irqrestore(&stats->syncp, flags);
 904         }
 905 }
 906
 907 static void notrace __bpf_prog_exit_recur(struct bpf_prog *prog, u64 start,
 908                                           struct bpf_tramp_run_ctx *run_ctx)
 909         __releases(RCU)
 910 {
 911         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 912
 913         update_prog_stats(prog, start);
 914         this_cpu_dec(*(prog->active));
 915         migrate_enable();
 916         rcu_read_unlock();
 917 }
 918
 919 static u64 notrace __bpf_prog_enter_lsm_cgroup(struct bpf_prog *prog,
 920                                                struct bpf_tramp_run_ctx *run_ctx)
 921         __acquires(RCU)
 922 {
 923         /* Runtime stats are exported via actual BPF_LSM_CGROUP
 924          * programs, not the shims.
 925          */
 926         rcu_read_lock();
 927         migrate_disable();
 928
 929         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 930
 931         return NO_START_TIME;
 932 }
 933
 934 static void notrace __bpf_prog_exit_lsm_cgroup(struct bpf_prog *prog, u64 start,
 935                                                struct bpf_tramp_run_ctx *run_ctx)
 936         __releases(RCU)
 937 {
 938         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 939
 940         migrate_enable();
 941         rcu_read_unlock();
 942 }
 943
 944 u64 notrace __bpf_prog_enter_sleepable_recur(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         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 952                 bpf_prog_inc_misses_counter(prog);
 953                 return 0;
 954         }
 955
 956         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 957
 958         return bpf_prog_start_time();
 959 }
 960
 961 void notrace __bpf_prog_exit_sleepable_recur(struct bpf_prog *prog, u64 start,
 962                                              struct bpf_tramp_run_ctx *run_ctx)
 963 {
 964         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 965
 966         update_prog_stats(prog, start);
 967         this_cpu_dec(*(prog->active));
 968         migrate_enable();
 969         rcu_read_unlock_trace();
 970 }
 971
 972 static u64 notrace __bpf_prog_enter_sleepable(struct bpf_prog *prog,
 973                                               struct bpf_tramp_run_ctx *run_ctx)
 974 {
 975         rcu_read_lock_trace();
 976         migrate_disable();
 977         might_fault();
 978
 979         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 980
 981         return bpf_prog_start_time();
 982 }
 983
 984 static void notrace __bpf_prog_exit_sleepable(struct bpf_prog *prog, u64 start,
 985                                               struct bpf_tramp_run_ctx *run_ctx)
 986 {
 987         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 988
 989         update_prog_stats(prog, start);
 990         migrate_enable();
 991         rcu_read_unlock_trace();
 992 }
 993
 994 static u64 notrace __bpf_prog_enter(struct bpf_prog *prog,
 995                                     struct bpf_tramp_run_ctx *run_ctx)
 996         __acquires(RCU)
 997 {
 998         rcu_read_lock();
 999         migrate_disable();
1000
1001         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
1002
1003         return bpf_prog_start_time();
1004 }
1005
1006 static void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start,
1007                                     struct bpf_tramp_run_ctx *run_ctx)
1008         __releases(RCU)
1009 {
1010         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
1011
1012         update_prog_stats(prog, start);
1013         migrate_enable();
1014         rcu_read_unlock();
1015 }
1016
1017 void notrace __bpf_tramp_enter(struct bpf_tramp_image *tr)
1018 {
1019         percpu_ref_get(&tr->pcref);
1020 }
1021
1022 void notrace __bpf_tramp_exit(struct bpf_tramp_image *tr)
1023 {
1024         percpu_ref_put(&tr->pcref);
1025 }
1026
1027 bpf_trampoline_enter_t bpf_trampoline_enter(const struct bpf_prog *prog)
1028 {
1029         bool sleepable = prog->aux->sleepable;
1030
1031         if (bpf_prog_check_recur(prog))
1032                 return sleepable ? __bpf_prog_enter_sleepable_recur :
1033                         __bpf_prog_enter_recur;
1034
1035         if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1036             prog->expected_attach_type == BPF_LSM_CGROUP)
1037                 return __bpf_prog_enter_lsm_cgroup;
1038
1039         return sleepable ? __bpf_prog_enter_sleepable : __bpf_prog_enter;
1040 }
1041
1042 bpf_trampoline_exit_t bpf_trampoline_exit(const struct bpf_prog *prog)
1043 {
1044         bool sleepable = prog->aux->sleepable;
1045
1046         if (bpf_prog_check_recur(prog))
1047                 return sleepable ? __bpf_prog_exit_sleepable_recur :
1048                         __bpf_prog_exit_recur;
1049
1050         if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1051             prog->expected_attach_type == BPF_LSM_CGROUP)
1052                 return __bpf_prog_exit_lsm_cgroup;
1053
1054         return sleepable ? __bpf_prog_exit_sleepable : __bpf_prog_exit;
1055 }
1056
1057 int __weak
1058 arch_prepare_bpf_trampoline(struct bpf_tramp_image *tr, void *image, void *image_end,
1059                             const struct btf_func_model *m, u32 flags,
1060                             struct bpf_tramp_links *tlinks,
1061                             void *orig_call)
1062 {
1063         return -ENOTSUPP;
1064 }
1065
1066 static int __init init_trampolines(void)
1067 {
1068         int i;
1069
1070         for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
1071                 INIT_HLIST_HEAD(&trampoline_table[i]);
1072         return 0;
1073 }
1074 late_initcall(init_trampolines);