kernel/bpf/bpf_task_storage.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2020 Facebook
   4  * Copyright 2020 Google LLC.
   5  */
   6
   7 #include <linux/pid.h>
   8 #include <linux/sched.h>
   9 #include <linux/rculist.h>
  10 #include <linux/list.h>
  11 #include <linux/hash.h>
  12 #include <linux/types.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/bpf.h>
  15 #include <linux/bpf_local_storage.h>
  16 #include <linux/filter.h>
  17 #include <uapi/linux/btf.h>
  18 #include <linux/btf_ids.h>
  19 #include <linux/rcupdate_trace.h>
  20
  21 DEFINE_BPF_STORAGE_CACHE(task_cache);
  22
  23 static DEFINE_PER_CPU(int, bpf_task_storage_busy);
  24
  25 static void bpf_task_storage_lock(void)
  26 {
  27         migrate_disable();
  28         this_cpu_inc(bpf_task_storage_busy);
  29 }
  30
  31 static void bpf_task_storage_unlock(void)
  32 {
  33         this_cpu_dec(bpf_task_storage_busy);
  34         migrate_enable();
  35 }
  36
  37 static bool bpf_task_storage_trylock(void)
  38 {
  39         migrate_disable();
  40         if (unlikely(this_cpu_inc_return(bpf_task_storage_busy) != 1)) {
  41                 this_cpu_dec(bpf_task_storage_busy);
  42                 migrate_enable();
  43                 return false;
  44         }
  45         return true;
  46 }
  47
  48 static struct bpf_local_storage __rcu **task_storage_ptr(void *owner)
  49 {
  50         struct task_struct *task = owner;
  51
  52         return &task->bpf_storage;
  53 }
  54
  55 static struct bpf_local_storage_data *
  56 task_storage_lookup(struct task_struct *task, struct bpf_map *map,
  57                     bool cacheit_lockit)
  58 {
  59         struct bpf_local_storage *task_storage;
  60         struct bpf_local_storage_map *smap;
  61
  62         task_storage =
  63                 rcu_dereference_check(task->bpf_storage, bpf_rcu_lock_held());
  64         if (!task_storage)
  65                 return NULL;
  66
  67         smap = (struct bpf_local_storage_map *)map;
  68         return bpf_local_storage_lookup(task_storage, smap, cacheit_lockit);
  69 }
  70
  71 void bpf_task_storage_free(struct task_struct *task)
  72 {
  73         struct bpf_local_storage *local_storage;
  74
  75         rcu_read_lock();
  76
  77         local_storage = rcu_dereference(task->bpf_storage);
  78         if (!local_storage) {
  79                 rcu_read_unlock();
  80                 return;
  81         }
  82
  83         bpf_task_storage_lock();
  84         bpf_local_storage_destroy(local_storage);
  85         bpf_task_storage_unlock();
  86         rcu_read_unlock();
  87 }
  88
  89 static void *bpf_pid_task_storage_lookup_elem(struct bpf_map *map, void *key)
  90 {
  91         struct bpf_local_storage_data *sdata;
  92         struct task_struct *task;
  93         unsigned int f_flags;
  94         struct pid *pid;
  95         int fd, err;
  96
  97         fd = *(int *)key;
  98         pid = pidfd_get_pid(fd, &f_flags);
  99         if (IS_ERR(pid))
 100                 return ERR_CAST(pid);
 101
 102         /* We should be in an RCU read side critical section, it should be safe
 103          * to call pid_task.
 104          */
 105         WARN_ON_ONCE(!rcu_read_lock_held());
 106         task = pid_task(pid, PIDTYPE_PID);
 107         if (!task) {
 108                 err = -ENOENT;
 109                 goto out;
 110         }
 111
 112         bpf_task_storage_lock();
 113         sdata = task_storage_lookup(task, map, true);
 114         bpf_task_storage_unlock();
 115         put_pid(pid);
 116         return sdata ? sdata->data : NULL;
 117 out:
 118         put_pid(pid);
 119         return ERR_PTR(err);
 120 }
 121
 122 static long bpf_pid_task_storage_update_elem(struct bpf_map *map, void *key,
 123                                              void *value, u64 map_flags)
 124 {
 125         struct bpf_local_storage_data *sdata;
 126         struct task_struct *task;
 127         unsigned int f_flags;
 128         struct pid *pid;
 129         int fd, err;
 130
 131         if ((map_flags & BPF_F_LOCK) && btf_record_has_field(map->record, BPF_UPTR))
 132                 return -EOPNOTSUPP;
 133
 134         fd = *(int *)key;
 135         pid = pidfd_get_pid(fd, &f_flags);
 136         if (IS_ERR(pid))
 137                 return PTR_ERR(pid);
 138
 139         /* We should be in an RCU read side critical section, it should be safe
 140          * to call pid_task.
 141          */
 142         WARN_ON_ONCE(!rcu_read_lock_held());
 143         task = pid_task(pid, PIDTYPE_PID);
 144         if (!task) {
 145                 err = -ENOENT;
 146                 goto out;
 147         }
 148
 149         bpf_task_storage_lock();
 150         sdata = bpf_local_storage_update(
 151                 task, (struct bpf_local_storage_map *)map, value, map_flags,
 152                 true, GFP_ATOMIC);
 153         bpf_task_storage_unlock();
 154
 155         err = PTR_ERR_OR_ZERO(sdata);
 156 out:
 157         put_pid(pid);
 158         return err;
 159 }
 160
 161 static int task_storage_delete(struct task_struct *task, struct bpf_map *map,
 162                                bool nobusy)
 163 {
 164         struct bpf_local_storage_data *sdata;
 165
 166         sdata = task_storage_lookup(task, map, false);
 167         if (!sdata)
 168                 return -ENOENT;
 169
 170         if (!nobusy)
 171                 return -EBUSY;
 172
 173         bpf_selem_unlink(SELEM(sdata), false);
 174
 175         return 0;
 176 }
 177
 178 static long bpf_pid_task_storage_delete_elem(struct bpf_map *map, void *key)
 179 {
 180         struct task_struct *task;
 181         unsigned int f_flags;
 182         struct pid *pid;
 183         int fd, err;
 184
 185         fd = *(int *)key;
 186         pid = pidfd_get_pid(fd, &f_flags);
 187         if (IS_ERR(pid))
 188                 return PTR_ERR(pid);
 189
 190         /* We should be in an RCU read side critical section, it should be safe
 191          * to call pid_task.
 192          */
 193         WARN_ON_ONCE(!rcu_read_lock_held());
 194         task = pid_task(pid, PIDTYPE_PID);
 195         if (!task) {
 196                 err = -ENOENT;
 197                 goto out;
 198         }
 199
 200         bpf_task_storage_lock();
 201         err = task_storage_delete(task, map, true);
 202         bpf_task_storage_unlock();
 203 out:
 204         put_pid(pid);
 205         return err;
 206 }
 207
 208 /* Called by bpf_task_storage_get*() helpers */
 209 static void *__bpf_task_storage_get(struct bpf_map *map,
 210                                     struct task_struct *task, void *value,
 211                                     u64 flags, gfp_t gfp_flags, bool nobusy)
 212 {
 213         struct bpf_local_storage_data *sdata;
 214
 215         sdata = task_storage_lookup(task, map, nobusy);
 216         if (sdata)
 217                 return sdata->data;
 218
 219         /* only allocate new storage, when the task is refcounted */
 220         if (refcount_read(&task->usage) &&
 221             (flags & BPF_LOCAL_STORAGE_GET_F_CREATE) && nobusy) {
 222                 sdata = bpf_local_storage_update(
 223                         task, (struct bpf_local_storage_map *)map, value,
 224                         BPF_NOEXIST, false, gfp_flags);
 225                 return IS_ERR(sdata) ? NULL : sdata->data;
 226         }
 227
 228         return NULL;
 229 }
 230
 231 /* *gfp_flags* is a hidden argument provided by the verifier */
 232 BPF_CALL_5(bpf_task_storage_get_recur, struct bpf_map *, map, struct task_struct *,
 233            task, void *, value, u64, flags, gfp_t, gfp_flags)
 234 {
 235         bool nobusy;
 236         void *data;
 237
 238         WARN_ON_ONCE(!bpf_rcu_lock_held());
 239         if (flags & ~BPF_LOCAL_STORAGE_GET_F_CREATE || !task)
 240                 return (unsigned long)NULL;
 241
 242         nobusy = bpf_task_storage_trylock();
 243         data = __bpf_task_storage_get(map, task, value, flags,
 244                                       gfp_flags, nobusy);
 245         if (nobusy)
 246                 bpf_task_storage_unlock();
 247         return (unsigned long)data;
 248 }
 249
 250 /* *gfp_flags* is a hidden argument provided by the verifier */
 251 BPF_CALL_5(bpf_task_storage_get, struct bpf_map *, map, struct task_struct *,
 252            task, void *, value, u64, flags, gfp_t, gfp_flags)
 253 {
 254         void *data;
 255
 256         WARN_ON_ONCE(!bpf_rcu_lock_held());
 257         if (flags & ~BPF_LOCAL_STORAGE_GET_F_CREATE || !task)
 258                 return (unsigned long)NULL;
 259
 260         bpf_task_storage_lock();
 261         data = __bpf_task_storage_get(map, task, value, flags,
 262                                       gfp_flags, true);
 263         bpf_task_storage_unlock();
 264         return (unsigned long)data;
 265 }
 266
 267 BPF_CALL_2(bpf_task_storage_delete_recur, struct bpf_map *, map, struct task_struct *,
 268            task)
 269 {
 270         bool nobusy;
 271         int ret;
 272
 273         WARN_ON_ONCE(!bpf_rcu_lock_held());
 274         if (!task)
 275                 return -EINVAL;
 276
 277         nobusy = bpf_task_storage_trylock();
 278         /* This helper must only be called from places where the lifetime of the task
 279          * is guaranteed. Either by being refcounted or by being protected
 280          * by an RCU read-side critical section.
 281          */
 282         ret = task_storage_delete(task, map, nobusy);
 283         if (nobusy)
 284                 bpf_task_storage_unlock();
 285         return ret;
 286 }
 287
 288 BPF_CALL_2(bpf_task_storage_delete, struct bpf_map *, map, struct task_struct *,
 289            task)
 290 {
 291         int ret;
 292
 293         WARN_ON_ONCE(!bpf_rcu_lock_held());
 294         if (!task)
 295                 return -EINVAL;
 296
 297         bpf_task_storage_lock();
 298         /* This helper must only be called from places where the lifetime of the task
 299          * is guaranteed. Either by being refcounted or by being protected
 300          * by an RCU read-side critical section.
 301          */
 302         ret = task_storage_delete(task, map, true);
 303         bpf_task_storage_unlock();
 304         return ret;
 305 }
 306
 307 static int notsupp_get_next_key(struct bpf_map *map, void *key, void *next_key)
 308 {
 309         return -ENOTSUPP;
 310 }
 311
 312 static struct bpf_map *task_storage_map_alloc(union bpf_attr *attr)
 313 {
 314         return bpf_local_storage_map_alloc(attr, &task_cache, true);
 315 }
 316
 317 static void task_storage_map_free(struct bpf_map *map)
 318 {
 319         bpf_local_storage_map_free(map, &task_cache, &bpf_task_storage_busy);
 320 }
 321
 322 BTF_ID_LIST_GLOBAL_SINGLE(bpf_local_storage_map_btf_id, struct, bpf_local_storage_map)
 323 const struct bpf_map_ops task_storage_map_ops = {
 324         .map_meta_equal = bpf_map_meta_equal,
 325         .map_alloc_check = bpf_local_storage_map_alloc_check,
 326         .map_alloc = task_storage_map_alloc,
 327         .map_free = task_storage_map_free,
 328         .map_get_next_key = notsupp_get_next_key,
 329         .map_lookup_elem = bpf_pid_task_storage_lookup_elem,
 330         .map_update_elem = bpf_pid_task_storage_update_elem,
 331         .map_delete_elem = bpf_pid_task_storage_delete_elem,
 332         .map_check_btf = bpf_local_storage_map_check_btf,
 333         .map_mem_usage = bpf_local_storage_map_mem_usage,
 334         .map_btf_id = &bpf_local_storage_map_btf_id[0],
 335         .map_owner_storage_ptr = task_storage_ptr,
 336 };
 337
 338 const struct bpf_func_proto bpf_task_storage_get_recur_proto = {
 339         .func = bpf_task_storage_get_recur,
 340         .gpl_only = false,
 341         .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
 342         .arg1_type = ARG_CONST_MAP_PTR,
 343         .arg2_type = ARG_PTR_TO_BTF_ID_OR_NULL,
 344         .arg2_btf_id = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 345         .arg3_type = ARG_PTR_TO_MAP_VALUE_OR_NULL,
 346         .arg4_type = ARG_ANYTHING,
 347 };
 348
 349 const struct bpf_func_proto bpf_task_storage_get_proto = {
 350         .func = bpf_task_storage_get,
 351         .gpl_only = false,
 352         .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
 353         .arg1_type = ARG_CONST_MAP_PTR,
 354         .arg2_type = ARG_PTR_TO_BTF_ID_OR_NULL,
 355         .arg2_btf_id = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 356         .arg3_type = ARG_PTR_TO_MAP_VALUE_OR_NULL,
 357         .arg4_type = ARG_ANYTHING,
 358 };
 359
 360 const struct bpf_func_proto bpf_task_storage_delete_recur_proto = {
 361         .func = bpf_task_storage_delete_recur,
 362         .gpl_only = false,
 363         .ret_type = RET_INTEGER,
 364         .arg1_type = ARG_CONST_MAP_PTR,
 365         .arg2_type = ARG_PTR_TO_BTF_ID_OR_NULL,
 366         .arg2_btf_id = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 367 };
 368
 369 const struct bpf_func_proto bpf_task_storage_delete_proto = {
 370         .func = bpf_task_storage_delete,
 371         .gpl_only = false,
 372         .ret_type = RET_INTEGER,
 373         .arg1_type = ARG_CONST_MAP_PTR,
 374         .arg2_type = ARG_PTR_TO_BTF_ID_OR_NULL,
 375         .arg2_btf_id = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 376 };