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_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)
 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", key);
 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                 goto out;
 433         }
 434
 435         im = bpf_tramp_image_alloc(tr->key);
 436         if (IS_ERR(im)) {
 437                 err = PTR_ERR(im);
 438                 goto out;
 439         }
 440
 441         /* clear all bits except SHARE_IPMODIFY */
 442         tr->flags &= BPF_TRAMP_F_SHARE_IPMODIFY;
 443
 444         if (tlinks[BPF_TRAMP_FEXIT].nr_links ||
 445             tlinks[BPF_TRAMP_MODIFY_RETURN].nr_links) {
 446                 /* NOTE: BPF_TRAMP_F_RESTORE_REGS and BPF_TRAMP_F_SKIP_FRAME
 447                  * should not be set together.
 448                  */
 449                 tr->flags |= BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
 450         } else {
 451                 tr->flags |= BPF_TRAMP_F_RESTORE_REGS;
 452         }
 453
 454         if (ip_arg)
 455                 tr->flags |= BPF_TRAMP_F_IP_ARG;
 456
 457 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 458 again:
 459         if ((tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY) &&
 460             (tr->flags & BPF_TRAMP_F_CALL_ORIG))
 461                 tr->flags |= BPF_TRAMP_F_ORIG_STACK;
 462 #endif
 463
 464         err = arch_prepare_bpf_trampoline(im, im->image, im->image + PAGE_SIZE,
 465                                           &tr->func.model, tr->flags, tlinks,
 466                                           tr->func.addr);
 467         if (err < 0)
 468                 goto out;
 469
 470         set_memory_ro((long)im->image, 1);
 471         set_memory_x((long)im->image, 1);
 472
 473         WARN_ON(tr->cur_image && total == 0);
 474         if (tr->cur_image)
 475                 /* progs already running at this address */
 476                 err = modify_fentry(tr, tr->cur_image->image, im->image, lock_direct_mutex);
 477         else
 478                 /* first time registering */
 479                 err = register_fentry(tr, im->image);
 480
 481 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 482         if (err == -EAGAIN) {
 483                 /* -EAGAIN from bpf_tramp_ftrace_ops_func. Now
 484                  * BPF_TRAMP_F_SHARE_IPMODIFY is set, we can generate the
 485                  * trampoline again, and retry register.
 486                  */
 487                 /* reset fops->func and fops->trampoline for re-register */
 488                 tr->fops->func = NULL;
 489                 tr->fops->trampoline = 0;
 490
 491                 /* reset im->image memory attr for arch_prepare_bpf_trampoline */
 492                 set_memory_nx((long)im->image, 1);
 493                 set_memory_rw((long)im->image, 1);
 494                 goto again;
 495         }
 496 #endif
 497         if (err)
 498                 goto out;
 499
 500         if (tr->cur_image)
 501                 bpf_tramp_image_put(tr->cur_image);
 502         tr->cur_image = im;
 503 out:
 504         /* If any error happens, restore previous flags */
 505         if (err)
 506                 tr->flags = orig_flags;
 507         kfree(tlinks);
 508         return err;
 509 }
 510
 511 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
 512 {
 513         switch (prog->expected_attach_type) {
 514         case BPF_TRACE_FENTRY:
 515                 return BPF_TRAMP_FENTRY;
 516         case BPF_MODIFY_RETURN:
 517                 return BPF_TRAMP_MODIFY_RETURN;
 518         case BPF_TRACE_FEXIT:
 519                 return BPF_TRAMP_FEXIT;
 520         case BPF_LSM_MAC:
 521                 if (!prog->aux->attach_func_proto->type)
 522                         /* The function returns void, we cannot modify its
 523                          * return value.
 524                          */
 525                         return BPF_TRAMP_FEXIT;
 526                 else
 527                         return BPF_TRAMP_MODIFY_RETURN;
 528         default:
 529                 return BPF_TRAMP_REPLACE;
 530         }
 531 }
 532
 533 static int __bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 534 {
 535         enum bpf_tramp_prog_type kind;
 536         struct bpf_tramp_link *link_exiting;
 537         int err = 0;
 538         int cnt = 0, i;
 539
 540         kind = bpf_attach_type_to_tramp(link->link.prog);
 541         if (tr->extension_prog)
 542                 /* cannot attach fentry/fexit if extension prog is attached.
 543                  * cannot overwrite extension prog either.
 544                  */
 545                 return -EBUSY;
 546
 547         for (i = 0; i < BPF_TRAMP_MAX; i++)
 548                 cnt += tr->progs_cnt[i];
 549
 550         if (kind == BPF_TRAMP_REPLACE) {
 551                 /* Cannot attach extension if fentry/fexit are in use. */
 552                 if (cnt)
 553                         return -EBUSY;
 554                 tr->extension_prog = link->link.prog;
 555                 return bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
 556                                           link->link.prog->bpf_func);
 557         }
 558         if (cnt >= BPF_MAX_TRAMP_LINKS)
 559                 return -E2BIG;
 560         if (!hlist_unhashed(&link->tramp_hlist))
 561                 /* prog already linked */
 562                 return -EBUSY;
 563         hlist_for_each_entry(link_exiting, &tr->progs_hlist[kind], tramp_hlist) {
 564                 if (link_exiting->link.prog != link->link.prog)
 565                         continue;
 566                 /* prog already linked */
 567                 return -EBUSY;
 568         }
 569
 570         hlist_add_head(&link->tramp_hlist, &tr->progs_hlist[kind]);
 571         tr->progs_cnt[kind]++;
 572         err = bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 573         if (err) {
 574                 hlist_del_init(&link->tramp_hlist);
 575                 tr->progs_cnt[kind]--;
 576         }
 577         return err;
 578 }
 579
 580 int bpf_trampoline_link_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 581 {
 582         int err;
 583
 584         mutex_lock(&tr->mutex);
 585         err = __bpf_trampoline_link_prog(link, tr);
 586         mutex_unlock(&tr->mutex);
 587         return err;
 588 }
 589
 590 static int __bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 591 {
 592         enum bpf_tramp_prog_type kind;
 593         int err;
 594
 595         kind = bpf_attach_type_to_tramp(link->link.prog);
 596         if (kind == BPF_TRAMP_REPLACE) {
 597                 WARN_ON_ONCE(!tr->extension_prog);
 598                 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
 599                                          tr->extension_prog->bpf_func, NULL);
 600                 tr->extension_prog = NULL;
 601                 return err;
 602         }
 603         hlist_del_init(&link->tramp_hlist);
 604         tr->progs_cnt[kind]--;
 605         return bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 606 }
 607
 608 /* bpf_trampoline_unlink_prog() should never fail. */
 609 int bpf_trampoline_unlink_prog(struct bpf_tramp_link *link, struct bpf_trampoline *tr)
 610 {
 611         int err;
 612
 613         mutex_lock(&tr->mutex);
 614         err = __bpf_trampoline_unlink_prog(link, tr);
 615         mutex_unlock(&tr->mutex);
 616         return err;
 617 }
 618
 619 #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_LSM)
 620 static void bpf_shim_tramp_link_release(struct bpf_link *link)
 621 {
 622         struct bpf_shim_tramp_link *shim_link =
 623                 container_of(link, struct bpf_shim_tramp_link, link.link);
 624
 625         /* paired with 'shim_link->trampoline = tr' in bpf_trampoline_link_cgroup_shim */
 626         if (!shim_link->trampoline)
 627                 return;
 628
 629         WARN_ON_ONCE(bpf_trampoline_unlink_prog(&shim_link->link, shim_link->trampoline));
 630         bpf_trampoline_put(shim_link->trampoline);
 631 }
 632
 633 static void bpf_shim_tramp_link_dealloc(struct bpf_link *link)
 634 {
 635         struct bpf_shim_tramp_link *shim_link =
 636                 container_of(link, struct bpf_shim_tramp_link, link.link);
 637
 638         kfree(shim_link);
 639 }
 640
 641 static const struct bpf_link_ops bpf_shim_tramp_link_lops = {
 642         .release = bpf_shim_tramp_link_release,
 643         .dealloc = bpf_shim_tramp_link_dealloc,
 644 };
 645
 646 static struct bpf_shim_tramp_link *cgroup_shim_alloc(const struct bpf_prog *prog,
 647                                                      bpf_func_t bpf_func,
 648                                                      int cgroup_atype)
 649 {
 650         struct bpf_shim_tramp_link *shim_link = NULL;
 651         struct bpf_prog *p;
 652
 653         shim_link = kzalloc(sizeof(*shim_link), GFP_USER);
 654         if (!shim_link)
 655                 return NULL;
 656
 657         p = bpf_prog_alloc(1, 0);
 658         if (!p) {
 659                 kfree(shim_link);
 660                 return NULL;
 661         }
 662
 663         p->jited = false;
 664         p->bpf_func = bpf_func;
 665
 666         p->aux->cgroup_atype = cgroup_atype;
 667         p->aux->attach_func_proto = prog->aux->attach_func_proto;
 668         p->aux->attach_btf_id = prog->aux->attach_btf_id;
 669         p->aux->attach_btf = prog->aux->attach_btf;
 670         btf_get(p->aux->attach_btf);
 671         p->type = BPF_PROG_TYPE_LSM;
 672         p->expected_attach_type = BPF_LSM_MAC;
 673         bpf_prog_inc(p);
 674         bpf_link_init(&shim_link->link.link, BPF_LINK_TYPE_UNSPEC,
 675                       &bpf_shim_tramp_link_lops, p);
 676         bpf_cgroup_atype_get(p->aux->attach_btf_id, cgroup_atype);
 677
 678         return shim_link;
 679 }
 680
 681 static struct bpf_shim_tramp_link *cgroup_shim_find(struct bpf_trampoline *tr,
 682                                                     bpf_func_t bpf_func)
 683 {
 684         struct bpf_tramp_link *link;
 685         int kind;
 686
 687         for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
 688                 hlist_for_each_entry(link, &tr->progs_hlist[kind], tramp_hlist) {
 689                         struct bpf_prog *p = link->link.prog;
 690
 691                         if (p->bpf_func == bpf_func)
 692                                 return container_of(link, struct bpf_shim_tramp_link, link);
 693                 }
 694         }
 695
 696         return NULL;
 697 }
 698
 699 int bpf_trampoline_link_cgroup_shim(struct bpf_prog *prog,
 700                                     int cgroup_atype)
 701 {
 702         struct bpf_shim_tramp_link *shim_link = NULL;
 703         struct bpf_attach_target_info tgt_info = {};
 704         struct bpf_trampoline *tr;
 705         bpf_func_t bpf_func;
 706         u64 key;
 707         int err;
 708
 709         err = bpf_check_attach_target(NULL, prog, NULL,
 710                                       prog->aux->attach_btf_id,
 711                                       &tgt_info);
 712         if (err)
 713                 return err;
 714
 715         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 716                                          prog->aux->attach_btf_id);
 717
 718         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 719         tr = bpf_trampoline_get(key, &tgt_info);
 720         if (!tr)
 721                 return  -ENOMEM;
 722
 723         mutex_lock(&tr->mutex);
 724
 725         shim_link = cgroup_shim_find(tr, bpf_func);
 726         if (shim_link) {
 727                 /* Reusing existing shim attached by the other program. */
 728                 bpf_link_inc(&shim_link->link.link);
 729
 730                 mutex_unlock(&tr->mutex);
 731                 bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 732                 return 0;
 733         }
 734
 735         /* Allocate and install new shim. */
 736
 737         shim_link = cgroup_shim_alloc(prog, bpf_func, cgroup_atype);
 738         if (!shim_link) {
 739                 err = -ENOMEM;
 740                 goto err;
 741         }
 742
 743         err = __bpf_trampoline_link_prog(&shim_link->link, tr);
 744         if (err)
 745                 goto err;
 746
 747         shim_link->trampoline = tr;
 748         /* note, we're still holding tr refcnt from above */
 749
 750         mutex_unlock(&tr->mutex);
 751
 752         return 0;
 753 err:
 754         mutex_unlock(&tr->mutex);
 755
 756         if (shim_link)
 757                 bpf_link_put(&shim_link->link.link);
 758
 759         /* have to release tr while _not_ holding its mutex */
 760         bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 761
 762         return err;
 763 }
 764
 765 void bpf_trampoline_unlink_cgroup_shim(struct bpf_prog *prog)
 766 {
 767         struct bpf_shim_tramp_link *shim_link = NULL;
 768         struct bpf_trampoline *tr;
 769         bpf_func_t bpf_func;
 770         u64 key;
 771
 772         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 773                                          prog->aux->attach_btf_id);
 774
 775         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 776         tr = bpf_trampoline_lookup(key);
 777         if (WARN_ON_ONCE(!tr))
 778                 return;
 779
 780         mutex_lock(&tr->mutex);
 781         shim_link = cgroup_shim_find(tr, bpf_func);
 782         mutex_unlock(&tr->mutex);
 783
 784         if (shim_link)
 785                 bpf_link_put(&shim_link->link.link);
 786
 787         bpf_trampoline_put(tr); /* bpf_trampoline_lookup above */
 788 }
 789 #endif
 790
 791 struct bpf_trampoline *bpf_trampoline_get(u64 key,
 792                                           struct bpf_attach_target_info *tgt_info)
 793 {
 794         struct bpf_trampoline *tr;
 795
 796         tr = bpf_trampoline_lookup(key);
 797         if (!tr)
 798                 return NULL;
 799
 800         mutex_lock(&tr->mutex);
 801         if (tr->func.addr)
 802                 goto out;
 803
 804         memcpy(&tr->func.model, &tgt_info->fmodel, sizeof(tgt_info->fmodel));
 805         tr->func.addr = (void *)tgt_info->tgt_addr;
 806 out:
 807         mutex_unlock(&tr->mutex);
 808         return tr;
 809 }
 810
 811 void bpf_trampoline_put(struct bpf_trampoline *tr)
 812 {
 813         int i;
 814
 815         if (!tr)
 816                 return;
 817         mutex_lock(&trampoline_mutex);
 818         if (!refcount_dec_and_test(&tr->refcnt))
 819                 goto out;
 820         WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
 821
 822         for (i = 0; i < BPF_TRAMP_MAX; i++)
 823                 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[i])))
 824                         goto out;
 825
 826         /* This code will be executed even when the last bpf_tramp_image
 827          * is alive. All progs are detached from the trampoline and the
 828          * trampoline image is patched with jmp into epilogue to skip
 829          * fexit progs. The fentry-only trampoline will be freed via
 830          * multiple rcu callbacks.
 831          */
 832         hlist_del(&tr->hlist);
 833         if (tr->fops) {
 834                 ftrace_free_filter(tr->fops);
 835                 kfree(tr->fops);
 836         }
 837         kfree(tr);
 838 out:
 839         mutex_unlock(&trampoline_mutex);
 840 }
 841
 842 #define NO_START_TIME 1
 843 static __always_inline u64 notrace bpf_prog_start_time(void)
 844 {
 845         u64 start = NO_START_TIME;
 846
 847         if (static_branch_unlikely(&bpf_stats_enabled_key)) {
 848                 start = sched_clock();
 849                 if (unlikely(!start))
 850                         start = NO_START_TIME;
 851         }
 852         return start;
 853 }
 854
 855 /* The logic is similar to bpf_prog_run(), but with an explicit
 856  * rcu_read_lock() and migrate_disable() which are required
 857  * for the trampoline. The macro is split into
 858  * call __bpf_prog_enter
 859  * call prog->bpf_func
 860  * call __bpf_prog_exit
 861  *
 862  * __bpf_prog_enter returns:
 863  * 0 - skip execution of the bpf prog
 864  * 1 - execute bpf prog
 865  * [2..MAX_U64] - execute bpf prog and record execution time.
 866  *     This is start time.
 867  */
 868 u64 notrace __bpf_prog_enter(struct bpf_prog *prog, struct bpf_tramp_run_ctx *run_ctx)
 869         __acquires(RCU)
 870 {
 871         rcu_read_lock();
 872         migrate_disable();
 873
 874         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 875
 876         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 877                 bpf_prog_inc_misses_counter(prog);
 878                 return 0;
 879         }
 880         return bpf_prog_start_time();
 881 }
 882
 883 static void notrace update_prog_stats(struct bpf_prog *prog,
 884                                       u64 start)
 885 {
 886         struct bpf_prog_stats *stats;
 887
 888         if (static_branch_unlikely(&bpf_stats_enabled_key) &&
 889             /* static_key could be enabled in __bpf_prog_enter*
 890              * and disabled in __bpf_prog_exit*.
 891              * And vice versa.
 892              * Hence check that 'start' is valid.
 893              */
 894             start > NO_START_TIME) {
 895                 unsigned long flags;
 896
 897                 stats = this_cpu_ptr(prog->stats);
 898                 flags = u64_stats_update_begin_irqsave(&stats->syncp);
 899                 u64_stats_inc(&stats->cnt);
 900                 u64_stats_add(&stats->nsecs, sched_clock() - start);
 901                 u64_stats_update_end_irqrestore(&stats->syncp, flags);
 902         }
 903 }
 904
 905 void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start, struct bpf_tramp_run_ctx *run_ctx)
 906         __releases(RCU)
 907 {
 908         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 909
 910         update_prog_stats(prog, start);
 911         this_cpu_dec(*(prog->active));
 912         migrate_enable();
 913         rcu_read_unlock();
 914 }
 915
 916 u64 notrace __bpf_prog_enter_lsm_cgroup(struct bpf_prog *prog,
 917                                         struct bpf_tramp_run_ctx *run_ctx)
 918         __acquires(RCU)
 919 {
 920         /* Runtime stats are exported via actual BPF_LSM_CGROUP
 921          * programs, not the shims.
 922          */
 923         rcu_read_lock();
 924         migrate_disable();
 925
 926         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 927
 928         return NO_START_TIME;
 929 }
 930
 931 void notrace __bpf_prog_exit_lsm_cgroup(struct bpf_prog *prog, u64 start,
 932                                         struct bpf_tramp_run_ctx *run_ctx)
 933         __releases(RCU)
 934 {
 935         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 936
 937         migrate_enable();
 938         rcu_read_unlock();
 939 }
 940
 941 u64 notrace __bpf_prog_enter_sleepable(struct bpf_prog *prog, struct bpf_tramp_run_ctx *run_ctx)
 942 {
 943         rcu_read_lock_trace();
 944         migrate_disable();
 945         might_fault();
 946
 947         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 948                 bpf_prog_inc_misses_counter(prog);
 949                 return 0;
 950         }
 951
 952         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 953
 954         return bpf_prog_start_time();
 955 }
 956
 957 void notrace __bpf_prog_exit_sleepable(struct bpf_prog *prog, u64 start,
 958                                        struct bpf_tramp_run_ctx *run_ctx)
 959 {
 960         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 961
 962         update_prog_stats(prog, start);
 963         this_cpu_dec(*(prog->active));
 964         migrate_enable();
 965         rcu_read_unlock_trace();
 966 }
 967
 968 u64 notrace __bpf_prog_enter_struct_ops(struct bpf_prog *prog,
 969                                         struct bpf_tramp_run_ctx *run_ctx)
 970         __acquires(RCU)
 971 {
 972         rcu_read_lock();
 973         migrate_disable();
 974
 975         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 976
 977         return bpf_prog_start_time();
 978 }
 979
 980 void notrace __bpf_prog_exit_struct_ops(struct bpf_prog *prog, u64 start,
 981                                         struct bpf_tramp_run_ctx *run_ctx)
 982         __releases(RCU)
 983 {
 984         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 985
 986         update_prog_stats(prog, start);
 987         migrate_enable();
 988         rcu_read_unlock();
 989 }
 990
 991 void notrace __bpf_tramp_enter(struct bpf_tramp_image *tr)
 992 {
 993         percpu_ref_get(&tr->pcref);
 994 }
 995
 996 void notrace __bpf_tramp_exit(struct bpf_tramp_image *tr)
 997 {
 998         percpu_ref_put(&tr->pcref);
 999 }
1000
1001 int __weak
1002 arch_prepare_bpf_trampoline(struct bpf_tramp_image *tr, void *image, void *image_end,
1003                             const struct btf_func_model *m, u32 flags,
1004                             struct bpf_tramp_links *tlinks,
1005                             void *orig_call)
1006 {
1007         return -ENOTSUPP;
1008 }
1009
1010 static int __init init_trampolines(void)
1011 {
1012         int i;
1013
1014         for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
1015                 INIT_HLIST_HEAD(&trampoline_table[i]);
1016         return 0;
1017 }
1018 late_initcall(init_trampolines);