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
  10 /* dummy _ops. The verifier will operate on target program's ops. */
  11 const struct bpf_verifier_ops bpf_extension_verifier_ops = {
  12 };
  13 const struct bpf_prog_ops bpf_extension_prog_ops = {
  14 };
  15
  16 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */
  17 #define TRAMPOLINE_HASH_BITS 10
  18 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS)
  19
  20 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE];
  21
  22 /* serializes access to trampoline_table */
  23 static DEFINE_MUTEX(trampoline_mutex);
  24
  25 void *bpf_jit_alloc_exec_page(void)
  26 {
  27         void *image;
  28
  29         image = bpf_jit_alloc_exec(PAGE_SIZE);
  30         if (!image)
  31                 return NULL;
  32
  33         set_vm_flush_reset_perms(image);
  34         /* Keep image as writeable. The alternative is to keep flipping ro/rw
  35          * everytime new program is attached or detached.
  36          */
  37         set_memory_x((long)image, 1);
  38         return image;
  39 }
  40
  41 void bpf_image_ksym_add(void *data, struct bpf_ksym *ksym)
  42 {
  43         ksym->start = (unsigned long) data;
  44         ksym->end = ksym->start + PAGE_SIZE;
  45         bpf_ksym_add(ksym);
  46         perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
  47                            PAGE_SIZE, false, ksym->name);
  48 }
  49
  50 void bpf_image_ksym_del(struct bpf_ksym *ksym)
  51 {
  52         bpf_ksym_del(ksym);
  53         perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start,
  54                            PAGE_SIZE, true, ksym->name);
  55 }
  56
  57 static void bpf_trampoline_ksym_add(struct bpf_trampoline *tr)
  58 {
  59         struct bpf_ksym *ksym = &tr->ksym;
  60
  61         snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu", tr->key);
  62         bpf_image_ksym_add(tr->image, ksym);
  63 }
  64
  65 struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
  66 {
  67         struct bpf_trampoline *tr;
  68         struct hlist_head *head;
  69         void *image;
  70         int i;
  71
  72         mutex_lock(&trampoline_mutex);
  73         head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)];
  74         hlist_for_each_entry(tr, head, hlist) {
  75                 if (tr->key == key) {
  76                         refcount_inc(&tr->refcnt);
  77                         goto out;
  78                 }
  79         }
  80         tr = kzalloc(sizeof(*tr), GFP_KERNEL);
  81         if (!tr)
  82                 goto out;
  83
  84         /* is_root was checked earlier. No need for bpf_jit_charge_modmem() */
  85         image = bpf_jit_alloc_exec_page();
  86         if (!image) {
  87                 kfree(tr);
  88                 tr = NULL;
  89                 goto out;
  90         }
  91
  92         tr->key = key;
  93         INIT_HLIST_NODE(&tr->hlist);
  94         hlist_add_head(&tr->hlist, head);
  95         refcount_set(&tr->refcnt, 1);
  96         mutex_init(&tr->mutex);
  97         for (i = 0; i < BPF_TRAMP_MAX; i++)
  98                 INIT_HLIST_HEAD(&tr->progs_hlist[i]);
  99         tr->image = image;
 100         INIT_LIST_HEAD_RCU(&tr->ksym.lnode);
 101         bpf_trampoline_ksym_add(tr);
 102 out:
 103         mutex_unlock(&trampoline_mutex);
 104         return tr;
 105 }
 106
 107 static int is_ftrace_location(void *ip)
 108 {
 109         long addr;
 110
 111         addr = ftrace_location((long)ip);
 112         if (!addr)
 113                 return 0;
 114         if (WARN_ON_ONCE(addr != (long)ip))
 115                 return -EFAULT;
 116         return 1;
 117 }
 118
 119 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr)
 120 {
 121         void *ip = tr->func.addr;
 122         int ret;
 123
 124         if (tr->func.ftrace_managed)
 125                 ret = unregister_ftrace_direct((long)ip, (long)old_addr);
 126         else
 127                 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL);
 128         return ret;
 129 }
 130
 131 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr)
 132 {
 133         void *ip = tr->func.addr;
 134         int ret;
 135
 136         if (tr->func.ftrace_managed)
 137                 ret = modify_ftrace_direct((long)ip, (long)old_addr, (long)new_addr);
 138         else
 139                 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr);
 140         return ret;
 141 }
 142
 143 /* first time registering */
 144 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
 145 {
 146         void *ip = tr->func.addr;
 147         int ret;
 148
 149         ret = is_ftrace_location(ip);
 150         if (ret < 0)
 151                 return ret;
 152         tr->func.ftrace_managed = ret;
 153
 154         if (tr->func.ftrace_managed)
 155                 ret = register_ftrace_direct((long)ip, (long)new_addr);
 156         else
 157                 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr);
 158         return ret;
 159 }
 160
 161 static struct bpf_tramp_progs *
 162 bpf_trampoline_get_progs(const struct bpf_trampoline *tr, int *total)
 163 {
 164         const struct bpf_prog_aux *aux;
 165         struct bpf_tramp_progs *tprogs;
 166         struct bpf_prog **progs;
 167         int kind;
 168
 169         *total = 0;
 170         tprogs = kcalloc(BPF_TRAMP_MAX, sizeof(*tprogs), GFP_KERNEL);
 171         if (!tprogs)
 172                 return ERR_PTR(-ENOMEM);
 173
 174         for (kind = 0; kind < BPF_TRAMP_MAX; kind++) {
 175                 tprogs[kind].nr_progs = tr->progs_cnt[kind];
 176                 *total += tr->progs_cnt[kind];
 177                 progs = tprogs[kind].progs;
 178
 179                 hlist_for_each_entry(aux, &tr->progs_hlist[kind], tramp_hlist)
 180                         *progs++ = aux->prog;
 181         }
 182         return tprogs;
 183 }
 184
 185 static int bpf_trampoline_update(struct bpf_trampoline *tr)
 186 {
 187         void *old_image = tr->image + ((tr->selector + 1) & 1) * PAGE_SIZE/2;
 188         void *new_image = tr->image + (tr->selector & 1) * PAGE_SIZE/2;
 189         struct bpf_tramp_progs *tprogs;
 190         u32 flags = BPF_TRAMP_F_RESTORE_REGS;
 191         int err, total;
 192
 193         tprogs = bpf_trampoline_get_progs(tr, &total);
 194         if (IS_ERR(tprogs))
 195                 return PTR_ERR(tprogs);
 196
 197         if (total == 0) {
 198                 err = unregister_fentry(tr, old_image);
 199                 tr->selector = 0;
 200                 goto out;
 201         }
 202
 203         if (tprogs[BPF_TRAMP_FEXIT].nr_progs ||
 204             tprogs[BPF_TRAMP_MODIFY_RETURN].nr_progs)
 205                 flags = BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME;
 206
 207         /* Though the second half of trampoline page is unused a task could be
 208          * preempted in the middle of the first half of trampoline and two
 209          * updates to trampoline would change the code from underneath the
 210          * preempted task. Hence wait for tasks to voluntarily schedule or go
 211          * to userspace.
 212          */
 213
 214         synchronize_rcu_tasks();
 215
 216         err = arch_prepare_bpf_trampoline(new_image, new_image + PAGE_SIZE / 2,
 217                                           &tr->func.model, flags, tprogs,
 218                                           tr->func.addr);
 219         if (err < 0)
 220                 goto out;
 221
 222         if (tr->selector)
 223                 /* progs already running at this address */
 224                 err = modify_fentry(tr, old_image, new_image);
 225         else
 226                 /* first time registering */
 227                 err = register_fentry(tr, new_image);
 228         if (err)
 229                 goto out;
 230         tr->selector++;
 231 out:
 232         kfree(tprogs);
 233         return err;
 234 }
 235
 236 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(enum bpf_attach_type t)
 237 {
 238         switch (t) {
 239         case BPF_TRACE_FENTRY:
 240                 return BPF_TRAMP_FENTRY;
 241         case BPF_MODIFY_RETURN:
 242                 return BPF_TRAMP_MODIFY_RETURN;
 243         case BPF_TRACE_FEXIT:
 244                 return BPF_TRAMP_FEXIT;
 245         default:
 246                 return BPF_TRAMP_REPLACE;
 247         }
 248 }
 249
 250 int bpf_trampoline_link_prog(struct bpf_prog *prog)
 251 {
 252         enum bpf_tramp_prog_type kind;
 253         struct bpf_trampoline *tr;
 254         int err = 0;
 255         int cnt;
 256
 257         tr = prog->aux->trampoline;
 258         kind = bpf_attach_type_to_tramp(prog->expected_attach_type);
 259         mutex_lock(&tr->mutex);
 260         if (tr->extension_prog) {
 261                 /* cannot attach fentry/fexit if extension prog is attached.
 262                  * cannot overwrite extension prog either.
 263                  */
 264                 err = -EBUSY;
 265                 goto out;
 266         }
 267         cnt = tr->progs_cnt[BPF_TRAMP_FENTRY] + tr->progs_cnt[BPF_TRAMP_FEXIT];
 268         if (kind == BPF_TRAMP_REPLACE) {
 269                 /* Cannot attach extension if fentry/fexit are in use. */
 270                 if (cnt) {
 271                         err = -EBUSY;
 272                         goto out;
 273                 }
 274                 tr->extension_prog = prog;
 275                 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL,
 276                                          prog->bpf_func);
 277                 goto out;
 278         }
 279         if (cnt >= BPF_MAX_TRAMP_PROGS) {
 280                 err = -E2BIG;
 281                 goto out;
 282         }
 283         if (!hlist_unhashed(&prog->aux->tramp_hlist)) {
 284                 /* prog already linked */
 285                 err = -EBUSY;
 286                 goto out;
 287         }
 288         hlist_add_head(&prog->aux->tramp_hlist, &tr->progs_hlist[kind]);
 289         tr->progs_cnt[kind]++;
 290         err = bpf_trampoline_update(prog->aux->trampoline);
 291         if (err) {
 292                 hlist_del(&prog->aux->tramp_hlist);
 293                 tr->progs_cnt[kind]--;
 294         }
 295 out:
 296         mutex_unlock(&tr->mutex);
 297         return err;
 298 }
 299
 300 /* bpf_trampoline_unlink_prog() should never fail. */
 301 int bpf_trampoline_unlink_prog(struct bpf_prog *prog)
 302 {
 303         enum bpf_tramp_prog_type kind;
 304         struct bpf_trampoline *tr;
 305         int err;
 306
 307         tr = prog->aux->trampoline;
 308         kind = bpf_attach_type_to_tramp(prog->expected_attach_type);
 309         mutex_lock(&tr->mutex);
 310         if (kind == BPF_TRAMP_REPLACE) {
 311                 WARN_ON_ONCE(!tr->extension_prog);
 312                 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP,
 313                                          tr->extension_prog->bpf_func, NULL);
 314                 tr->extension_prog = NULL;
 315                 goto out;
 316         }
 317         hlist_del(&prog->aux->tramp_hlist);
 318         tr->progs_cnt[kind]--;
 319         err = bpf_trampoline_update(prog->aux->trampoline);
 320 out:
 321         mutex_unlock(&tr->mutex);
 322         return err;
 323 }
 324
 325 void bpf_trampoline_put(struct bpf_trampoline *tr)
 326 {
 327         if (!tr)
 328                 return;
 329         mutex_lock(&trampoline_mutex);
 330         if (!refcount_dec_and_test(&tr->refcnt))
 331                 goto out;
 332         WARN_ON_ONCE(mutex_is_locked(&tr->mutex));
 333         if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FENTRY])))
 334                 goto out;
 335         if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FEXIT])))
 336                 goto out;
 337         bpf_image_ksym_del(&tr->ksym);
 338         /* wait for tasks to get out of trampoline before freeing it */
 339         synchronize_rcu_tasks();
 340         bpf_jit_free_exec(tr->image);
 341         hlist_del(&tr->hlist);
 342         kfree(tr);
 343 out:
 344         mutex_unlock(&trampoline_mutex);
 345 }
 346
 347 /* The logic is similar to BPF_PROG_RUN, but with an explicit
 348  * rcu_read_lock() and migrate_disable() which are required
 349  * for the trampoline. The macro is split into
 350  * call _bpf_prog_enter
 351  * call prog->bpf_func
 352  * call __bpf_prog_exit
 353  */
 354 u64 notrace __bpf_prog_enter(void)
 355 {
 356         u64 start = 0;
 357
 358         rcu_read_lock();
 359         migrate_disable();
 360         if (static_branch_unlikely(&bpf_stats_enabled_key))
 361                 start = sched_clock();
 362         return start;
 363 }
 364
 365 void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start)
 366 {
 367         struct bpf_prog_stats *stats;
 368
 369         if (static_branch_unlikely(&bpf_stats_enabled_key) &&
 370             /* static_key could be enabled in __bpf_prog_enter
 371              * and disabled in __bpf_prog_exit.
 372              * And vice versa.
 373              * Hence check that 'start' is not zero.
 374              */
 375             start) {
 376                 stats = this_cpu_ptr(prog->aux->stats);
 377                 u64_stats_update_begin(&stats->syncp);
 378                 stats->cnt++;
 379                 stats->nsecs += sched_clock() - start;
 380                 u64_stats_update_end(&stats->syncp);
 381         }
 382         migrate_enable();
 383         rcu_read_unlock();
 384 }
 385
 386 int __weak
 387 arch_prepare_bpf_trampoline(void *image, void *image_end,
 388                             const struct btf_func_model *m, u32 flags,
 389                             struct bpf_tramp_progs *tprogs,
 390                             void *orig_call)
 391 {
 392         return -ENOTSUPP;
 393 }
 394
 395 static int __init init_trampolines(void)
 396 {
 397         int i;
 398
 399         for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++)
 400                 INIT_HLIST_HEAD(&trampoline_table[i]);
 401         return 0;
 402 }
 403 late_initcall(init_trampolines);