kernel/bpf/trampoline.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (c) 2019 Facebook */
   3 #include <linux/hash.h>
   4 #include <linux/bpf.h>
   5 #include <linux/filter.h>
   6 #include <linux/ftrace.h>
   7 #include <linux/rbtree_latch.h>
   8 #include <linux/perf_event.h>
   9 #include <linux/btf.h>
  10 #include <linux/rcupdate_trace.h>
  11 #include <linux/rcupdate_wait.h>
  12 #include <linux/static_call.h>
  13 #include <linux/bpf_verifier.h>
  14 #include <linux/bpf_lsm.h>
  15 #include <linux/delay.h>
  16
  17 /* dummy _ops. The verifier will operate on target program's ops. */
  18 const struct bpf_verifier_ops bpf_extension_verifier_ops = {
  19 };
  20 const struct bpf_prog_ops bpf_extension_prog_ops = {
  21 };
  22
  23 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */
  24 #define TRAMPOLINE_HASH_BITS 10
  25 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS)
  26
  27 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE];
  28
  29 /* serializes access to trampoline_table */
  30 static DEFINE_MUTEX(trampoline_mutex);
  31
  32 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
  33 static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mutex);
  34
  35 static int bpf_tramp_ftrace_ops_func(struct ftrace_ops *ops, enum ftrace_ops_cmd cmd)
  36 {
  37         struct bpf_trampoline *tr = ops->private;
  38         int ret = 0;
  39
  40         if (cmd == FTRACE_OPS_CMD_ENABLE_SHARE_IPMODIFY_SELF) {
  41                 /* This is called inside register_ftrace_direct_multi(), so
  42                  * tr->mutex is already locked.
  43                  */
  44                 lockdep_assert_held_once(&tr->mutex);
  45
  46                 /* Instead of updating the trampoline here, we propagate
  47                  * -EAGAIN to register_ftrace_direct(). Then we can
  48                  * retry register_ftrace_direct() after updating the
  49                  * trampoline.
  50                  */
  51                 if ((tr->flags & BPF_TRAMP_F_CALL_ORIG) &&
  52                     !(tr->flags & BPF_TRAMP_F_ORIG_STACK)) {
  53                         if (WARN_ON_ONCE(tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY))
  54                                 return -EBUSY;
  55
  56                         tr->flags |= BPF_TRAMP_F_SHARE_IPMODIFY;
  57                         return -EAGAIN;
  58                 }
  59
  60                 return 0;
  61         }
  62
  63         /* The normal locking order is
  64          *    tr->mutex => direct_mutex (ftrace.c) => ftrace_lock (ftrace.c)
  65          *
  66          * The following two commands are called from
  67          *
  68          *   prepare_direct_functions_for_ipmodify
  69          *   cleanup_direct_functions_after_ipmodify
  70          *
  71          * In both cases, direct_mutex is already locked. Use
  72          * mutex_trylock(&tr->mutex) to avoid deadlock in race condition
  73          * (something else is making changes to this same trampoline).
  74          */
  75         if (!mutex_trylock(&tr->mutex)) {
  76                 /* sleep 1 ms to make sure whatever holding tr->mutex makes
  77                  * some progress.
  78                  */
  79                 msleep(1);
  80                 return -EAGAIN;
  81         }
  82
  83         switch (cmd) {
  84         case FTRACE_OPS_CMD_ENABLE_SHARE_IPMODIFY_PEER:
  85                 tr->flags |= BPF_TRAMP_F_SHARE_IPMODIFY;
  86
  87                 if ((tr->flags & BPF_TRAMP_F_CALL_ORIG) &&
  88                     !(tr->flags & BPF_TRAMP_F_ORIG_STACK))
  89                         ret = bpf_trampoline_update(tr, false /* lock_direct_mutex */);
  90                 break;
  91         case FTRACE_OPS_CMD_DISABLE_SHARE_IPMODIFY_PEER:
  92                 tr->flags &= ~BPF_TRAMP_F_SHARE_IPMODIFY;
  93
  94                 if (tr->flags & BPF_TRAMP_F_ORIG_STACK)
  95                         ret = bpf_trampoline_update(tr, false /* lock_direct_mutex */);
  96                 break;
  97         default:
  98                 ret = -EINVAL;
  99                 break;
 100         }
 101
 102         mutex_unlock(&tr->mutex);
 103         return ret;
 104 }
 105 #endif
 106
 107 bool bpf_prog_has_trampoline(const struct bpf_prog *prog)
 108 {
 109         enum bpf_attach_type eatype = prog->expected_attach_type;
 110         enum bpf_prog_type ptype = prog->type;
 111
 112         return (ptype == BPF_PROG_TYPE_TRACING &&
 113                 (eatype == BPF_TRACE_FENTRY || eatype == BPF_TRACE_FEXIT ||
 114                  eatype == BPF_MODIFY_RETURN)) ||
 115                 (ptype == BPF_PROG_TYPE_LSM && eatype == BPF_LSM_MAC);
 116 }
 117
 118 void bpf_image_ksym_init(void *data, unsigned int size, struct bpf_ksym *ksym)
 119 {
 120         ksym->start = (unsigned long) data;
 121         ksym->end = ksym->start + size;
 122 }
 123
 124 void bpf_image_ksym_add(struct bpf_ksym *ksym)
 125 {
 126         bpf_ksym_add(ksym);
 127         perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
 128                            PAGE_SIZE, false, ksym->name);
 129 }
 130
 131 void bpf_image_ksym_del(struct bpf_ksym *ksym)
 132 {
 133         bpf_ksym_del(ksym);
 134         perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
 135                            PAGE_SIZE, true, ksym->name);
 136 }
 137
 138 static struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
 139 {
 140         struct bpf_trampoline *tr;
 141         struct hlist_head *head;
 142         int i;
 143
 144         mutex_lock(&trampoline_mutex);
 145         head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
 146         hlist_for_each_entry(tr, head, hlist) {
 147                 if (tr->key == key) {
 148                         refcount_inc(&tr->refcnt);
 149                         goto out;
 150                 }
 151         }
 152         tr = kzalloc(sizeof(*tr), GFP_KERNEL);
 153         if (!tr)
 154                 goto out;
 155 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 156         tr->fops = kzalloc(sizeof(struct ftrace_ops), GFP_KERNEL);
 157         if (!tr->fops) {
 158                 kfree(tr);
 159                 tr = NULL;
 160                 goto out;
 161         }
 162         tr->fops->private = tr;
 163         tr->fops->ops_func = bpf_tramp_ftrace_ops_func;
 164 #endif
 165
 166         tr->key = key;
 167         INIT_HLIST_NODE(&tr->hlist);
 168         hlist_add_head(&tr->hlist, head);
 169         refcount_set(&tr->refcnt, 1);
 170         mutex_init(&tr->mutex);
 171         for (i = 0; i < BPF_TRAMP_MAX; i++)
 172                 INIT_HLIST_HEAD(&tr->progs_hlist[i]);
 173 out:
 174         mutex_unlock(&trampoline_mutex);
 175         return tr;
 176 }
 177
 178 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
 179 {
 180         void *ip = tr->func.addr;
 181         int ret;
 182
 183         if (tr->func.ftrace_managed)
 184                 ret = unregister_ftrace_direct(tr->fops, (long)old_addr, false);
 185         else
 186                 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
 187
 188         return ret;
 189 }
 190
 191 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr,
 192                          bool lock_direct_mutex)
 193 {
 194         void *ip = tr->func.addr;
 195         int ret;
 196
 197         if (tr->func.ftrace_managed) {
 198                 if (lock_direct_mutex)
 199                         ret = modify_ftrace_direct(tr->fops, (long)new_addr);
 200                 else
 201                         ret = modify_ftrace_direct_nolock(tr->fops, (long)new_addr);
 202         } else {
 203                 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr);
 204         }
 205         return ret;
 206 }
 207
 208 /* first time registering */
 209 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
 210 {
 211         void *ip = tr->func.addr;
 212         unsigned long faddr;
 213         int ret;
 214
 215         faddr = ftrace_location((unsigned long)ip);
 216         if (faddr) {
 217                 if (!tr->fops)
 218                         return -ENOTSUPP;
 219                 tr->func.ftrace_managed = true;
 220         }
 221
 222         if (tr->func.ftrace_managed) {
 223                 ftrace_set_filter_ip(tr->fops, (unsigned long)ip, 0, 1);
 224                 ret = register_ftrace_direct(tr->fops, (long)new_addr);
 225         } else {
 226                 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
 227         }
 228
 229         return ret;
 230 }
 231
 232 static struct bpf_tramp_links *
 233 bpf_trampoline_get_progs(const struct bpf_trampoline *tr, int *total, bool *ip_arg)
 234 {
 235         struct bpf_tramp_link *link;
 236         struct bpf_tramp_links *tlinks;
 237         struct bpf_tramp_link **links;
 238         int kind;
 239
 240         *total = 0;
 241         tlinks = kcalloc(BPF_TRAMP_MAX, sizeof(*tlinks), GFP_KERNEL);
 242         if (!tlinks)
 243                 return ERR_PTR(-ENOMEM);
 244
 245         for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
 246                 tlinks[kind].nr_links = tr->progs_cnt[kind];
 247                 *total += tr->progs_cnt[kind];
 248                 links = tlinks[kind].links;
 249
 250                 hlist_for_each_entry(link, &tr->progs_hlist[kind], tramp_hlist) {
 251                         *ip_arg |= link->link.prog->call_get_func_ip;
 252                         *links++ = link;
 253                 }
 254         }
 255         return tlinks;
 256 }
 257
 258 static void bpf_tramp_image_free(struct bpf_tramp_image *im)
 259 {
 260         bpf_image_ksym_del(&im->ksym);
 261         arch_free_bpf_trampoline(im->image, im->size);
 262         bpf_jit_uncharge_modmem(im->size);
 263         percpu_ref_exit(&im->pcref);
 264         kfree_rcu(im, rcu);
 265 }
 266
 267 static void __bpf_tramp_image_put_deferred(struct work_struct *work)
 268 {
 269         struct bpf_tramp_image *im;
 270
 271         im = container_of(work, struct bpf_tramp_image, work);
 272         bpf_tramp_image_free(im);
 273 }
 274
 275 /* callback, fexit step 3 or fentry step 2 */
 276 static void __bpf_tramp_image_put_rcu(struct rcu_head *rcu)
 277 {
 278         struct bpf_tramp_image *im;
 279
 280         im = container_of(rcu, struct bpf_tramp_image, rcu);
 281         INIT_WORK(&im->work, __bpf_tramp_image_put_deferred);
 282         schedule_work(&im->work);
 283 }
 284
 285 /* callback, fexit step 2. Called after percpu_ref_kill confirms. */
 286 static void __bpf_tramp_image_release(struct percpu_ref *pcref)
 287 {
 288         struct bpf_tramp_image *im;
 289
 290         im = container_of(pcref, struct bpf_tramp_image, pcref);
 291         call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
 292 }
 293
 294 /* callback, fexit or fentry step 1 */
 295 static void __bpf_tramp_image_put_rcu_tasks(struct rcu_head *rcu)
 296 {
 297         struct bpf_tramp_image *im;
 298
 299         im = container_of(rcu, struct bpf_tramp_image, rcu);
 300         if (im->ip_after_call)
 301                 /* the case of fmod_ret/fexit trampoline and CONFIG_PREEMPTION=y */
 302                 percpu_ref_kill(&im->pcref);
 303         else
 304                 /* the case of fentry trampoline */
 305                 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu);
 306 }
 307
 308 static void bpf_tramp_image_put(struct bpf_tramp_image *im)
 309 {
 310         /* The trampoline image that calls original function is using:
 311          * rcu_read_lock_trace to protect sleepable bpf progs
 312          * rcu_read_lock to protect normal bpf progs
 313          * percpu_ref to protect trampoline itself
 314          * rcu tasks to protect trampoline asm not covered by percpu_ref
 315          * (which are few asm insns before __bpf_tramp_enter and
 316          *  after __bpf_tramp_exit)
 317          *
 318          * The trampoline is unreachable before bpf_tramp_image_put().
 319          *
 320          * First, patch the trampoline to avoid calling into fexit progs.
 321          * The progs will be freed even if the original function is still
 322          * executing or sleeping.
 323          * In case of CONFIG_PREEMPT=y use call_rcu_tasks() to wait on
 324          * first few asm instructions to execute and call into
 325          * __bpf_tramp_enter->percpu_ref_get.
 326          * Then use percpu_ref_kill to wait for the trampoline and the original
 327          * function to finish.
 328          * Then use call_rcu_tasks() to make sure few asm insns in
 329          * the trampoline epilogue are done as well.
 330          *
 331          * In !PREEMPT case the task that got interrupted in the first asm
 332          * insns won't go through an RCU quiescent state which the
 333          * percpu_ref_kill will be waiting for. Hence the first
 334          * call_rcu_tasks() is not necessary.
 335          */
 336         if (im->ip_after_call) {
 337                 int err = bpf_arch_text_poke(im->ip_after_call, BPF_MOD_JUMP,
 338                                              NULL, im->ip_epilogue);
 339                 WARN_ON(err);
 340                 if (IS_ENABLED(CONFIG_TASKS_RCU))
 341                         call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
 342                 else
 343                         percpu_ref_kill(&im->pcref);
 344                 return;
 345         }
 346
 347         /* The trampoline without fexit and fmod_ret progs doesn't call original
 348          * function and doesn't use percpu_ref.
 349          * Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
 350          * Then use call_rcu_tasks() to wait for the rest of trampoline asm
 351          * and normal progs.
 352          */
 353         call_rcu_tasks_trace(&im->rcu, __bpf_tramp_image_put_rcu_tasks);
 354 }
 355
 356 static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, int size)
 357 {
 358         struct bpf_tramp_image *im;
 359         struct bpf_ksym *ksym;
 360         void *image;
 361         int err = -ENOMEM;
 362
 363         im = kzalloc(sizeof(*im), GFP_KERNEL);
 364         if (!im)
 365                 goto out;
 366
 367         err = bpf_jit_charge_modmem(size);
 368         if (err)
 369                 goto out_free_im;
 370         im->size = size;
 371
 372         err = -ENOMEM;
 373         im->image = image = arch_alloc_bpf_trampoline(size);
 374         if (!image)
 375                 goto out_uncharge;
 376
 377         err = percpu_ref_init(&im->pcref, __bpf_tramp_image_release, 0, GFP_KERNEL);
 378         if (err)
 379                 goto out_free_image;
 380
 381         ksym = &im->ksym;
 382         INIT_LIST_HEAD_RCU(&ksym->lnode);
 383         snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu", key);
 384         bpf_image_ksym_init(image, size, ksym);
 385         bpf_image_ksym_add(ksym);
 386         return im;
 387
 388 out_free_image:
 389         arch_free_bpf_trampoline(im->image, im->size);
 390 out_uncharge:
 391         bpf_jit_uncharge_modmem(size);
 392 out_free_im:
 393         kfree(im);
 394 out:
 395         return ERR_PTR(err);
 396 }
 397
 398 static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mutex)
 399 {
 400         struct bpf_tramp_image *im;
 401         struct bpf_tramp_links *tlinks;
 402         u32 orig_flags = tr->flags;
 403         bool ip_arg = false;
 404         int err, total, size;
 405
 406         tlinks = bpf_trampoline_get_progs(tr, &total, &ip_arg);
 407         if (IS_ERR(tlinks))
 408                 return PTR_ERR(tlinks);
 409
 410         if (total == 0) {
 411                 err = unregister_fentry(tr, tr->cur_image->image);
 412                 bpf_tramp_image_put(tr->cur_image);
 413                 tr->cur_image = NULL;
 414                 goto out;
 415         }
 416
 417         /* clear all bits except SHARE_IPMODIFY and TAIL_CALL_CTX */
 418         tr->flags &= (BPF_TRAMP_F_SHARE_IPMODIFY | BPF_TRAMP_F_TAIL_CALL_CTX);
 419
 420         if (tlinks[BPF_TRAMP_FEXIT].nr_links ||
 421             tlinks[BPF_TRAMP_MODIFY_RETURN].nr_links) {
 422                 /* NOTE: BPF_TRAMP_F_RESTORE_REGS and BPF_TRAMP_F_SKIP_FRAME
 423                  * should not be set together.
 424                  */
 425                 tr->flags |= BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
 426         } else {
 427                 tr->flags |= BPF_TRAMP_F_RESTORE_REGS;
 428         }
 429
 430         if (ip_arg)
 431                 tr->flags |= BPF_TRAMP_F_IP_ARG;
 432
 433 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 434 again:
 435         if ((tr->flags & BPF_TRAMP_F_SHARE_IPMODIFY) &&
 436             (tr->flags & BPF_TRAMP_F_CALL_ORIG))
 437                 tr->flags |= BPF_TRAMP_F_ORIG_STACK;
 438 #endif
 439
 440         size = arch_bpf_trampoline_size(&tr->func.model, tr->flags,
 441                                         tlinks, tr->func.addr);
 442         if (size < 0) {
 443                 err = size;
 444                 goto out;
 445         }
 446
 447         if (size > PAGE_SIZE) {
 448                 err = -E2BIG;
 449                 goto out;
 450         }
 451
 452         im = bpf_tramp_image_alloc(tr->key, size);
 453         if (IS_ERR(im)) {
 454                 err = PTR_ERR(im);
 455                 goto out;
 456         }
 457
 458         err = arch_prepare_bpf_trampoline(im, im->image, im->image + size,
 459                                           &tr->func.model, tr->flags, tlinks,
 460                                           tr->func.addr);
 461         if (err < 0)
 462                 goto out_free;
 463
 464         err = arch_protect_bpf_trampoline(im->image, im->size);
 465         if (err)
 466                 goto out_free;
 467
 468         WARN_ON(tr->cur_image && total == 0);
 469         if (tr->cur_image)
 470                 /* progs already running at this address */
 471                 err = modify_fentry(tr, tr->cur_image->image, im->image, lock_direct_mutex);
 472         else
 473                 /* first time registering */
 474                 err = register_fentry(tr, im->image);
 475
 476 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
 477         if (err == -EAGAIN) {
 478                 /* -EAGAIN from bpf_tramp_ftrace_ops_func. Now
 479                  * BPF_TRAMP_F_SHARE_IPMODIFY is set, we can generate the
 480                  * trampoline again, and retry register.
 481                  */
 482                 /* reset fops->func and fops->trampoline for re-register */
 483                 tr->fops->func = NULL;
 484                 tr->fops->trampoline = 0;
 485
 486                 /* free im memory and reallocate later */
 487                 bpf_tramp_image_free(im);
 488                 goto again;
 489         }
 490 #endif
 491         if (err)
 492                 goto out_free;
 493
 494         if (tr->cur_image)
 495                 bpf_tramp_image_put(tr->cur_image);
 496         tr->cur_image = im;
 497 out:
 498         /* If any error happens, restore previous flags */
 499         if (err)
 500                 tr->flags = orig_flags;
 501         kfree(tlinks);
 502         return err;
 503
 504 out_free:
 505         bpf_tramp_image_free(im);
 506         goto out;
 507 }
 508
 509 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog)
 510 {
 511         switch (prog->expected_attach_type) {
 512         case BPF_TRACE_FENTRY:
 513                 return BPF_TRAMP_FENTRY;
 514         case BPF_MODIFY_RETURN:
 515                 return BPF_TRAMP_MODIFY_RETURN;
 516         case BPF_TRACE_FEXIT:
 517                 return BPF_TRAMP_FEXIT;
 518         case BPF_LSM_MAC:
 519                 if (!prog->aux->attach_func_proto->type)
 520                         /* The function returns void, we cannot modify its
 521                          * return value.
 522                          */
 523                         return BPF_TRAMP_FEXIT;
 524                 else
 525                         return BPF_TRAMP_MODIFY_RETURN;
 526         default:
 527                 return BPF_TRAMP_REPLACE;
 528         }
 529 }
 530
 531 static int bpf_freplace_check_tgt_prog(struct bpf_prog *tgt_prog)
 532 {
 533         struct bpf_prog_aux *aux = tgt_prog->aux;
 534
 535         guard(mutex)(&aux->ext_mutex);
 536         if (aux->prog_array_member_cnt)
 537                 /* Program extensions can not extend target prog when the target
 538                  * prog has been updated to any prog_array map as tail callee.
 539                  * It's to prevent a potential infinite loop like:
 540                  * tgt prog entry -> tgt prog subprog -> freplace prog entry
 541                  * --tailcall-> tgt prog entry.
 542                  */
 543                 return -EBUSY;
 544
 545         aux->is_extended = true;
 546         return 0;
 547 }
 548
 549 static int __bpf_trampoline_link_prog(struct bpf_tramp_link *link,
 550                                       struct bpf_trampoline *tr,
 551                                       struct bpf_prog *tgt_prog)
 552 {
 553         enum bpf_tramp_prog_type kind;
 554         struct bpf_tramp_link *link_exiting;
 555         int err = 0;
 556         int cnt = 0, i;
 557
 558         kind = bpf_attach_type_to_tramp(link->link.prog);
 559         if (tr->extension_prog)
 560                 /* cannot attach fentry/fexit if extension prog is attached.
 561                  * cannot overwrite extension prog either.
 562                  */
 563                 return -EBUSY;
 564
 565         for (i = 0; i < BPF_TRAMP_MAX; i++)
 566                 cnt += tr->progs_cnt[i];
 567
 568         if (kind == BPF_TRAMP_REPLACE) {
 569                 /* Cannot attach extension if fentry/fexit are in use. */
 570                 if (cnt)
 571                         return -EBUSY;
 572                 err = bpf_freplace_check_tgt_prog(tgt_prog);
 573                 if (err)
 574                         return err;
 575                 tr->extension_prog = link->link.prog;
 576                 return bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
 577                                           link->link.prog->bpf_func);
 578         }
 579         if (cnt >= BPF_MAX_TRAMP_LINKS)
 580                 return -E2BIG;
 581         if (!hlist_unhashed(&link->tramp_hlist))
 582                 /* prog already linked */
 583                 return -EBUSY;
 584         hlist_for_each_entry(link_exiting, &tr->progs_hlist[kind], tramp_hlist) {
 585                 if (link_exiting->link.prog != link->link.prog)
 586                         continue;
 587                 /* prog already linked */
 588                 return -EBUSY;
 589         }
 590
 591         hlist_add_head(&link->tramp_hlist, &tr->progs_hlist[kind]);
 592         tr->progs_cnt[kind]++;
 593         err = bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 594         if (err) {
 595                 hlist_del_init(&link->tramp_hlist);
 596                 tr->progs_cnt[kind]--;
 597         }
 598         return err;
 599 }
 600
 601 int bpf_trampoline_link_prog(struct bpf_tramp_link *link,
 602                              struct bpf_trampoline *tr,
 603                              struct bpf_prog *tgt_prog)
 604 {
 605         int err;
 606
 607         mutex_lock(&tr->mutex);
 608         err = __bpf_trampoline_link_prog(link, tr, tgt_prog);
 609         mutex_unlock(&tr->mutex);
 610         return err;
 611 }
 612
 613 static int __bpf_trampoline_unlink_prog(struct bpf_tramp_link *link,
 614                                         struct bpf_trampoline *tr,
 615                                         struct bpf_prog *tgt_prog)
 616 {
 617         enum bpf_tramp_prog_type kind;
 618         int err;
 619
 620         kind = bpf_attach_type_to_tramp(link->link.prog);
 621         if (kind == BPF_TRAMP_REPLACE) {
 622                 WARN_ON_ONCE(!tr->extension_prog);
 623                 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
 624                                          tr->extension_prog->bpf_func, NULL);
 625                 tr->extension_prog = NULL;
 626                 guard(mutex)(&tgt_prog->aux->ext_mutex);
 627                 tgt_prog->aux->is_extended = false;
 628                 return err;
 629         }
 630         hlist_del_init(&link->tramp_hlist);
 631         tr->progs_cnt[kind]--;
 632         return bpf_trampoline_update(tr, true /* lock_direct_mutex */);
 633 }
 634
 635 /* bpf_trampoline_unlink_prog() should never fail. */
 636 int bpf_trampoline_unlink_prog(struct bpf_tramp_link *link,
 637                                struct bpf_trampoline *tr,
 638                                struct bpf_prog *tgt_prog)
 639 {
 640         int err;
 641
 642         mutex_lock(&tr->mutex);
 643         err = __bpf_trampoline_unlink_prog(link, tr, tgt_prog);
 644         mutex_unlock(&tr->mutex);
 645         return err;
 646 }
 647
 648 #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_LSM)
 649 static void bpf_shim_tramp_link_release(struct bpf_link *link)
 650 {
 651         struct bpf_shim_tramp_link *shim_link =
 652                 container_of(link, struct bpf_shim_tramp_link, link.link);
 653
 654         /* paired with 'shim_link->trampoline = tr' in bpf_trampoline_link_cgroup_shim */
 655         if (!shim_link->trampoline)
 656                 return;
 657
 658         WARN_ON_ONCE(bpf_trampoline_unlink_prog(&shim_link->link, shim_link->trampoline, NULL));
 659         bpf_trampoline_put(shim_link->trampoline);
 660 }
 661
 662 static void bpf_shim_tramp_link_dealloc(struct bpf_link *link)
 663 {
 664         struct bpf_shim_tramp_link *shim_link =
 665                 container_of(link, struct bpf_shim_tramp_link, link.link);
 666
 667         kfree(shim_link);
 668 }
 669
 670 static const struct bpf_link_ops bpf_shim_tramp_link_lops = {
 671         .release = bpf_shim_tramp_link_release,
 672         .dealloc = bpf_shim_tramp_link_dealloc,
 673 };
 674
 675 static struct bpf_shim_tramp_link *cgroup_shim_alloc(const struct bpf_prog *prog,
 676                                                      bpf_func_t bpf_func,
 677                                                      int cgroup_atype)
 678 {
 679         struct bpf_shim_tramp_link *shim_link = NULL;
 680         struct bpf_prog *p;
 681
 682         shim_link = kzalloc(sizeof(*shim_link), GFP_USER);
 683         if (!shim_link)
 684                 return NULL;
 685
 686         p = bpf_prog_alloc(1, 0);
 687         if (!p) {
 688                 kfree(shim_link);
 689                 return NULL;
 690         }
 691
 692         p->jited = false;
 693         p->bpf_func = bpf_func;
 694
 695         p->aux->cgroup_atype = cgroup_atype;
 696         p->aux->attach_func_proto = prog->aux->attach_func_proto;
 697         p->aux->attach_btf_id = prog->aux->attach_btf_id;
 698         p->aux->attach_btf = prog->aux->attach_btf;
 699         btf_get(p->aux->attach_btf);
 700         p->type = BPF_PROG_TYPE_LSM;
 701         p->expected_attach_type = BPF_LSM_MAC;
 702         bpf_prog_inc(p);
 703         bpf_link_init(&shim_link->link.link, BPF_LINK_TYPE_UNSPEC,
 704                       &bpf_shim_tramp_link_lops, p);
 705         bpf_cgroup_atype_get(p->aux->attach_btf_id, cgroup_atype);
 706
 707         return shim_link;
 708 }
 709
 710 static struct bpf_shim_tramp_link *cgroup_shim_find(struct bpf_trampoline *tr,
 711                                                     bpf_func_t bpf_func)
 712 {
 713         struct bpf_tramp_link *link;
 714         int kind;
 715
 716         for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
 717                 hlist_for_each_entry(link, &tr->progs_hlist[kind], tramp_hlist) {
 718                         struct bpf_prog *p = link->link.prog;
 719
 720                         if (p->bpf_func == bpf_func)
 721                                 return container_of(link, struct bpf_shim_tramp_link, link);
 722                 }
 723         }
 724
 725         return NULL;
 726 }
 727
 728 int bpf_trampoline_link_cgroup_shim(struct bpf_prog *prog,
 729                                     int cgroup_atype)
 730 {
 731         struct bpf_shim_tramp_link *shim_link = NULL;
 732         struct bpf_attach_target_info tgt_info = {};
 733         struct bpf_trampoline *tr;
 734         bpf_func_t bpf_func;
 735         u64 key;
 736         int err;
 737
 738         err = bpf_check_attach_target(NULL, prog, NULL,
 739                                       prog->aux->attach_btf_id,
 740                                       &tgt_info);
 741         if (err)
 742                 return err;
 743
 744         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 745                                          prog->aux->attach_btf_id);
 746
 747         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 748         tr = bpf_trampoline_get(key, &tgt_info);
 749         if (!tr)
 750                 return  -ENOMEM;
 751
 752         mutex_lock(&tr->mutex);
 753
 754         shim_link = cgroup_shim_find(tr, bpf_func);
 755         if (shim_link) {
 756                 /* Reusing existing shim attached by the other program. */
 757                 bpf_link_inc(&shim_link->link.link);
 758
 759                 mutex_unlock(&tr->mutex);
 760                 bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 761                 return 0;
 762         }
 763
 764         /* Allocate and install new shim. */
 765
 766         shim_link = cgroup_shim_alloc(prog, bpf_func, cgroup_atype);
 767         if (!shim_link) {
 768                 err = -ENOMEM;
 769                 goto err;
 770         }
 771
 772         err = __bpf_trampoline_link_prog(&shim_link->link, tr, NULL);
 773         if (err)
 774                 goto err;
 775
 776         shim_link->trampoline = tr;
 777         /* note, we're still holding tr refcnt from above */
 778
 779         mutex_unlock(&tr->mutex);
 780
 781         return 0;
 782 err:
 783         mutex_unlock(&tr->mutex);
 784
 785         if (shim_link)
 786                 bpf_link_put(&shim_link->link.link);
 787
 788         /* have to release tr while _not_ holding its mutex */
 789         bpf_trampoline_put(tr); /* bpf_trampoline_get above */
 790
 791         return err;
 792 }
 793
 794 void bpf_trampoline_unlink_cgroup_shim(struct bpf_prog *prog)
 795 {
 796         struct bpf_shim_tramp_link *shim_link = NULL;
 797         struct bpf_trampoline *tr;
 798         bpf_func_t bpf_func;
 799         u64 key;
 800
 801         key = bpf_trampoline_compute_key(NULL, prog->aux->attach_btf,
 802                                          prog->aux->attach_btf_id);
 803
 804         bpf_lsm_find_cgroup_shim(prog, &bpf_func);
 805         tr = bpf_trampoline_lookup(key);
 806         if (WARN_ON_ONCE(!tr))
 807                 return;
 808
 809         mutex_lock(&tr->mutex);
 810         shim_link = cgroup_shim_find(tr, bpf_func);
 811         mutex_unlock(&tr->mutex);
 812
 813         if (shim_link)
 814                 bpf_link_put(&shim_link->link.link);
 815
 816         bpf_trampoline_put(tr); /* bpf_trampoline_lookup above */
 817 }
 818 #endif
 819
 820 struct bpf_trampoline *bpf_trampoline_get(u64 key,
 821                                           struct bpf_attach_target_info *tgt_info)
 822 {
 823         struct bpf_trampoline *tr;
 824
 825         tr = bpf_trampoline_lookup(key);
 826         if (!tr)
 827                 return NULL;
 828
 829         mutex_lock(&tr->mutex);
 830         if (tr->func.addr)
 831                 goto out;
 832
 833         memcpy(&tr->func.model, &tgt_info->fmodel, sizeof(tgt_info->fmodel));
 834         tr->func.addr = (void *)tgt_info->tgt_addr;
 835 out:
 836         mutex_unlock(&tr->mutex);
 837         return tr;
 838 }
 839
 840 void bpf_trampoline_put(struct bpf_trampoline *tr)
 841 {
 842         int i;
 843
 844         if (!tr)
 845                 return;
 846         mutex_lock(&trampoline_mutex);
 847         if (!refcount_dec_and_test(&tr->refcnt))
 848                 goto out;
 849         WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
 850
 851         for (i = 0; i < BPF_TRAMP_MAX; i++)
 852                 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[i])))
 853                         goto out;
 854
 855         /* This code will be executed even when the last bpf_tramp_image
 856          * is alive. All progs are detached from the trampoline and the
 857          * trampoline image is patched with jmp into epilogue to skip
 858          * fexit progs. The fentry-only trampoline will be freed via
 859          * multiple rcu callbacks.
 860          */
 861         hlist_del(&tr->hlist);
 862         if (tr->fops) {
 863                 ftrace_free_filter(tr->fops);
 864                 kfree(tr->fops);
 865         }
 866         kfree(tr);
 867 out:
 868         mutex_unlock(&trampoline_mutex);
 869 }
 870
 871 #define NO_START_TIME 1
 872 static __always_inline u64 notrace bpf_prog_start_time(void)
 873 {
 874         u64 start = NO_START_TIME;
 875
 876         if (static_branch_unlikely(&bpf_stats_enabled_key)) {
 877                 start = sched_clock();
 878                 if (unlikely(!start))
 879                         start = NO_START_TIME;
 880         }
 881         return start;
 882 }
 883
 884 /* The logic is similar to bpf_prog_run(), but with an explicit
 885  * rcu_read_lock() and migrate_disable() which are required
 886  * for the trampoline. The macro is split into
 887  * call __bpf_prog_enter
 888  * call prog->bpf_func
 889  * call __bpf_prog_exit
 890  *
 891  * __bpf_prog_enter returns:
 892  * 0 - skip execution of the bpf prog
 893  * 1 - execute bpf prog
 894  * [2..MAX_U64] - execute bpf prog and record execution time.
 895  *     This is start time.
 896  */
 897 static u64 notrace __bpf_prog_enter_recur(struct bpf_prog *prog, struct bpf_tramp_run_ctx *run_ctx)
 898         __acquires(RCU)
 899 {
 900         rcu_read_lock();
 901         migrate_disable();
 902
 903         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 904
 905         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 906                 bpf_prog_inc_misses_counter(prog);
 907                 if (prog->aux->recursion_detected)
 908                         prog->aux->recursion_detected(prog);
 909                 return 0;
 910         }
 911         return bpf_prog_start_time();
 912 }
 913
 914 static void notrace update_prog_stats(struct bpf_prog *prog,
 915                                       u64 start)
 916 {
 917         struct bpf_prog_stats *stats;
 918
 919         if (static_branch_unlikely(&bpf_stats_enabled_key) &&
 920             /* static_key could be enabled in __bpf_prog_enter*
 921              * and disabled in __bpf_prog_exit*.
 922              * And vice versa.
 923              * Hence check that 'start' is valid.
 924              */
 925             start > NO_START_TIME) {
 926                 u64 duration = sched_clock() - start;
 927                 unsigned long flags;
 928
 929                 stats = this_cpu_ptr(prog->stats);
 930                 flags = u64_stats_update_begin_irqsave(&stats->syncp);
 931                 u64_stats_inc(&stats->cnt);
 932                 u64_stats_add(&stats->nsecs, duration);
 933                 u64_stats_update_end_irqrestore(&stats->syncp, flags);
 934         }
 935 }
 936
 937 static void notrace __bpf_prog_exit_recur(struct bpf_prog *prog, u64 start,
 938                                           struct bpf_tramp_run_ctx *run_ctx)
 939         __releases(RCU)
 940 {
 941         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 942
 943         update_prog_stats(prog, start);
 944         this_cpu_dec(*(prog->active));
 945         migrate_enable();
 946         rcu_read_unlock();
 947 }
 948
 949 static u64 notrace __bpf_prog_enter_lsm_cgroup(struct bpf_prog *prog,
 950                                                struct bpf_tramp_run_ctx *run_ctx)
 951         __acquires(RCU)
 952 {
 953         /* Runtime stats are exported via actual BPF_LSM_CGROUP
 954          * programs, not the shims.
 955          */
 956         rcu_read_lock();
 957         migrate_disable();
 958
 959         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 960
 961         return NO_START_TIME;
 962 }
 963
 964 static void notrace __bpf_prog_exit_lsm_cgroup(struct bpf_prog *prog, u64 start,
 965                                                struct bpf_tramp_run_ctx *run_ctx)
 966         __releases(RCU)
 967 {
 968         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 969
 970         migrate_enable();
 971         rcu_read_unlock();
 972 }
 973
 974 u64 notrace __bpf_prog_enter_sleepable_recur(struct bpf_prog *prog,
 975                                              struct bpf_tramp_run_ctx *run_ctx)
 976 {
 977         rcu_read_lock_trace();
 978         migrate_disable();
 979         might_fault();
 980
 981         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
 982
 983         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
 984                 bpf_prog_inc_misses_counter(prog);
 985                 if (prog->aux->recursion_detected)
 986                         prog->aux->recursion_detected(prog);
 987                 return 0;
 988         }
 989         return bpf_prog_start_time();
 990 }
 991
 992 void notrace __bpf_prog_exit_sleepable_recur(struct bpf_prog *prog, u64 start,
 993                                              struct bpf_tramp_run_ctx *run_ctx)
 994 {
 995         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
 996
 997         update_prog_stats(prog, start);
 998         this_cpu_dec(*(prog->active));
 999         migrate_enable();
1000         rcu_read_unlock_trace();
1001 }
1002
1003 static u64 notrace __bpf_prog_enter_sleepable(struct bpf_prog *prog,
1004                                               struct bpf_tramp_run_ctx *run_ctx)
1005 {
1006         rcu_read_lock_trace();
1007         migrate_disable();
1008         might_fault();
1009
1010         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
1011
1012         return bpf_prog_start_time();
1013 }
1014
1015 static void notrace __bpf_prog_exit_sleepable(struct bpf_prog *prog, u64 start,
1016                                               struct bpf_tramp_run_ctx *run_ctx)
1017 {
1018         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
1019
1020         update_prog_stats(prog, start);
1021         migrate_enable();
1022         rcu_read_unlock_trace();
1023 }
1024
1025 static u64 notrace __bpf_prog_enter(struct bpf_prog *prog,
1026                                     struct bpf_tramp_run_ctx *run_ctx)
1027         __acquires(RCU)
1028 {
1029         rcu_read_lock();
1030         migrate_disable();
1031
1032         run_ctx->saved_run_ctx = bpf_set_run_ctx(&run_ctx->run_ctx);
1033
1034         return bpf_prog_start_time();
1035 }
1036
1037 static void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start,
1038                                     struct bpf_tramp_run_ctx *run_ctx)
1039         __releases(RCU)
1040 {
1041         bpf_reset_run_ctx(run_ctx->saved_run_ctx);
1042
1043         update_prog_stats(prog, start);
1044         migrate_enable();
1045         rcu_read_unlock();
1046 }
1047
1048 void notrace __bpf_tramp_enter(struct bpf_tramp_image *tr)
1049 {
1050         percpu_ref_get(&tr->pcref);
1051 }
1052
1053 void notrace __bpf_tramp_exit(struct bpf_tramp_image *tr)
1054 {
1055         percpu_ref_put(&tr->pcref);
1056 }
1057
1058 bpf_trampoline_enter_t bpf_trampoline_enter(const struct bpf_prog *prog)
1059 {
1060         bool sleepable = prog->sleepable;
1061
1062         if (bpf_prog_check_recur(prog))
1063                 return sleepable ? __bpf_prog_enter_sleepable_recur :
1064                         __bpf_prog_enter_recur;
1065
1066         if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1067             prog->expected_attach_type == BPF_LSM_CGROUP)
1068                 return __bpf_prog_enter_lsm_cgroup;
1069
1070         return sleepable ? __bpf_prog_enter_sleepable : __bpf_prog_enter;
1071 }
1072
1073 bpf_trampoline_exit_t bpf_trampoline_exit(const struct bpf_prog *prog)
1074 {
1075         bool sleepable = prog->sleepable;
1076
1077         if (bpf_prog_check_recur(prog))
1078                 return sleepable ? __bpf_prog_exit_sleepable_recur :
1079                         __bpf_prog_exit_recur;
1080
1081         if (resolve_prog_type(prog) == BPF_PROG_TYPE_LSM &&
1082             prog->expected_attach_type == BPF_LSM_CGROUP)
1083                 return __bpf_prog_exit_lsm_cgroup;
1084
1085         return sleepable ? __bpf_prog_exit_sleepable : __bpf_prog_exit;
1086 }
1087
1088 int __weak
1089 arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
1090                             const struct btf_func_model *m, u32 flags,
1091                             struct bpf_tramp_links *tlinks,
1092                             void *func_addr)
1093 {
1094         return -ENOTSUPP;
1095 }
1096
1097 void * __weak arch_alloc_bpf_trampoline(unsigned int size)
1098 {
1099         void *image;
1100
1101         if (WARN_ON_ONCE(size > PAGE_SIZE))
1102                 return NULL;
1103         image = bpf_jit_alloc_exec(PAGE_SIZE);
1104         if (image)
1105                 set_vm_flush_reset_perms(image);
1106         return image;
1107 }
1108
1109 void __weak arch_free_bpf_trampoline(void *image, unsigned int size)
1110 {
1111         WARN_ON_ONCE(size > PAGE_SIZE);
1112         /* bpf_jit_free_exec doesn't need "size", but
1113          * bpf_prog_pack_free() needs it.
1114          */
1115         bpf_jit_free_exec(image);
1116 }
1117
1118 int __weak arch_protect_bpf_trampoline(void *image, unsigned int size)
1119 {
1120         WARN_ON_ONCE(size > PAGE_SIZE);
1121         return set_memory_rox((long)image, 1);
1122 }
1123
1124 int __weak arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
1125                                     struct bpf_tramp_links *tlinks, void *func_addr)
1126 {
1127         return -ENOTSUPP;
1128 }
1129
1130 static int __init init_trampolines(void)
1131 {
1132         int i;
1133
1134         for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
1135                 INIT_HLIST_HEAD(&trampoline_table[i]);
1136         return 0;
1137 }
1138 late_initcall(init_trampolines);