include/linux/skmsg.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
   3
   4 #ifndef _LINUX_SKMSG_H
   5 #define _LINUX_SKMSG_H
   6
   7 #include <linux/bpf.h>
   8 #include <linux/filter.h>
   9 #include <linux/scatterlist.h>
  10 #include <linux/skbuff.h>
  11
  12 #include <net/sock.h>
  13 #include <net/tcp.h>
  14 #include <net/strparser.h>
  15
  16 #define MAX_MSG_FRAGS                   MAX_SKB_FRAGS
  17 #define NR_MSG_FRAG_IDS                 (MAX_MSG_FRAGS + 1)
  18
  19 enum __sk_action {
  20         __SK_DROP = 0,
  21         __SK_PASS,
  22         __SK_REDIRECT,
  23         __SK_NONE,
  24 };
  25
  26 struct sk_msg_sg {
  27         u32                             start;
  28         u32                             curr;
  29         u32                             end;
  30         u32                             size;
  31         u32                             copybreak;
  32         unsigned long                   copy;
  33         /* The extra two elements:
  34          * 1) used for chaining the front and sections when the list becomes
  35          *    partitioned (e.g. end < start). The crypto APIs require the
  36          *    chaining;
  37          * 2) to chain tailer SG entries after the message.
  38          */
  39         struct scatterlist              data[MAX_MSG_FRAGS + 2];
  40 };
  41 static_assert(BITS_PER_LONG >= NR_MSG_FRAG_IDS);
  42
  43 /* UAPI in filter.c depends on struct sk_msg_sg being first element. */
  44 struct sk_msg {
  45         struct sk_msg_sg                sg;
  46         void                            *data;
  47         void                            *data_end;
  48         u32                             apply_bytes;
  49         u32                             cork_bytes;
  50         u32                             flags;
  51         struct sk_buff                  *skb;
  52         struct sock                     *sk_redir;
  53         struct sock                     *sk;
  54         struct list_head                list;
  55 };
  56
  57 struct sk_psock_progs {
  58         struct bpf_prog                 *msg_parser;
  59         struct bpf_prog                 *stream_parser;
  60         struct bpf_prog                 *stream_verdict;
  61         struct bpf_prog                 *skb_verdict;
  62 };
  63
  64 enum sk_psock_state_bits {
  65         SK_PSOCK_TX_ENABLED,
  66 };
  67
  68 struct sk_psock_link {
  69         struct list_head                list;
  70         struct bpf_map                  *map;
  71         void                            *link_raw;
  72 };
  73
  74 struct sk_psock_work_state {
  75         struct sk_buff                  *skb;
  76         u32                             len;
  77         u32                             off;
  78 };
  79
  80 struct sk_psock {
  81         struct sock                     *sk;
  82         struct sock                     *sk_redir;
  83         u32                             apply_bytes;
  84         u32                             cork_bytes;
  85         u32                             eval;
  86         struct sk_msg                   *cork;
  87         struct sk_psock_progs           progs;
  88 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
  89         struct strparser                strp;
  90 #endif
  91         struct sk_buff_head             ingress_skb;
  92         struct list_head                ingress_msg;
  93         spinlock_t                      ingress_lock;
  94         unsigned long                   state;
  95         struct list_head                link;
  96         spinlock_t                      link_lock;
  97         refcount_t                      refcnt;
  98         void (*saved_unhash)(struct sock *sk);
  99         void (*saved_close)(struct sock *sk, long timeout);
 100         void (*saved_write_space)(struct sock *sk);
 101         void (*saved_data_ready)(struct sock *sk);
 102         int  (*psock_update_sk_prot)(struct sock *sk, struct sk_psock *psock,
 103                                      bool restore);
 104         struct proto                    *sk_proto;
 105         struct mutex                    work_mutex;
 106         struct sk_psock_work_state      work_state;
 107         struct work_struct              work;
 108         struct rcu_work                 rwork;
 109 };
 110
 111 int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
 112                  int elem_first_coalesce);
 113 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
 114                  u32 off, u32 len);
 115 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len);
 116 int sk_msg_free(struct sock *sk, struct sk_msg *msg);
 117 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg);
 118 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes);
 119 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
 120                                   u32 bytes);
 121
 122 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes);
 123 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes);
 124
 125 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
 126                               struct sk_msg *msg, u32 bytes);
 127 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 128                              struct sk_msg *msg, u32 bytes);
 129 int sk_msg_recvmsg(struct sock *sk, struct sk_psock *psock, struct msghdr *msg,
 130                    int len, int flags);
 131
 132 static inline void sk_msg_check_to_free(struct sk_msg *msg, u32 i, u32 bytes)
 133 {
 134         WARN_ON(i == msg->sg.end && bytes);
 135 }
 136
 137 static inline void sk_msg_apply_bytes(struct sk_psock *psock, u32 bytes)
 138 {
 139         if (psock->apply_bytes) {
 140                 if (psock->apply_bytes < bytes)
 141                         psock->apply_bytes = 0;
 142                 else
 143                         psock->apply_bytes -= bytes;
 144         }
 145 }
 146
 147 static inline u32 sk_msg_iter_dist(u32 start, u32 end)
 148 {
 149         return end >= start ? end - start : end + (NR_MSG_FRAG_IDS - start);
 150 }
 151
 152 #define sk_msg_iter_var_prev(var)                       \
 153         do {                                            \
 154                 if (var == 0)                           \
 155                         var = NR_MSG_FRAG_IDS - 1;      \
 156                 else                                    \
 157                         var--;                          \
 158         } while (0)
 159
 160 #define sk_msg_iter_var_next(var)                       \
 161         do {                                            \
 162                 var++;                                  \
 163                 if (var == NR_MSG_FRAG_IDS)             \
 164                         var = 0;                        \
 165         } while (0)
 166
 167 #define sk_msg_iter_prev(msg, which)                    \
 168         sk_msg_iter_var_prev(msg->sg.which)
 169
 170 #define sk_msg_iter_next(msg, which)                    \
 171         sk_msg_iter_var_next(msg->sg.which)
 172
 173 static inline void sk_msg_clear_meta(struct sk_msg *msg)
 174 {
 175         memset(&msg->sg, 0, offsetofend(struct sk_msg_sg, copy));
 176 }
 177
 178 static inline void sk_msg_init(struct sk_msg *msg)
 179 {
 180         BUILD_BUG_ON(ARRAY_SIZE(msg->sg.data) - 1 != NR_MSG_FRAG_IDS);
 181         memset(msg, 0, sizeof(*msg));
 182         sg_init_marker(msg->sg.data, NR_MSG_FRAG_IDS);
 183 }
 184
 185 static inline void sk_msg_xfer(struct sk_msg *dst, struct sk_msg *src,
 186                                int which, u32 size)
 187 {
 188         dst->sg.data[which] = src->sg.data[which];
 189         dst->sg.data[which].length  = size;
 190         dst->sg.size               += size;
 191         src->sg.size               -= size;
 192         src->sg.data[which].length -= size;
 193         src->sg.data[which].offset += size;
 194 }
 195
 196 static inline void sk_msg_xfer_full(struct sk_msg *dst, struct sk_msg *src)
 197 {
 198         memcpy(dst, src, sizeof(*src));
 199         sk_msg_init(src);
 200 }
 201
 202 static inline bool sk_msg_full(const struct sk_msg *msg)
 203 {
 204         return sk_msg_iter_dist(msg->sg.start, msg->sg.end) == MAX_MSG_FRAGS;
 205 }
 206
 207 static inline u32 sk_msg_elem_used(const struct sk_msg *msg)
 208 {
 209         return sk_msg_iter_dist(msg->sg.start, msg->sg.end);
 210 }
 211
 212 static inline struct scatterlist *sk_msg_elem(struct sk_msg *msg, int which)
 213 {
 214         return &msg->sg.data[which];
 215 }
 216
 217 static inline struct scatterlist sk_msg_elem_cpy(struct sk_msg *msg, int which)
 218 {
 219         return msg->sg.data[which];
 220 }
 221
 222 static inline struct page *sk_msg_page(struct sk_msg *msg, int which)
 223 {
 224         return sg_page(sk_msg_elem(msg, which));
 225 }
 226
 227 static inline bool sk_msg_to_ingress(const struct sk_msg *msg)
 228 {
 229         return msg->flags & BPF_F_INGRESS;
 230 }
 231
 232 static inline void sk_msg_compute_data_pointers(struct sk_msg *msg)
 233 {
 234         struct scatterlist *sge = sk_msg_elem(msg, msg->sg.start);
 235
 236         if (test_bit(msg->sg.start, &msg->sg.copy)) {
 237                 msg->data = NULL;
 238                 msg->data_end = NULL;
 239         } else {
 240                 msg->data = sg_virt(sge);
 241                 msg->data_end = msg->data + sge->length;
 242         }
 243 }
 244
 245 static inline void sk_msg_page_add(struct sk_msg *msg, struct page *page,
 246                                    u32 len, u32 offset)
 247 {
 248         struct scatterlist *sge;
 249
 250         get_page(page);
 251         sge = sk_msg_elem(msg, msg->sg.end);
 252         sg_set_page(sge, page, len, offset);
 253         sg_unmark_end(sge);
 254
 255         __set_bit(msg->sg.end, &msg->sg.copy);
 256         msg->sg.size += len;
 257         sk_msg_iter_next(msg, end);
 258 }
 259
 260 static inline void sk_msg_sg_copy(struct sk_msg *msg, u32 i, bool copy_state)
 261 {
 262         do {
 263                 if (copy_state)
 264                         __set_bit(i, &msg->sg.copy);
 265                 else
 266                         __clear_bit(i, &msg->sg.copy);
 267                 sk_msg_iter_var_next(i);
 268                 if (i == msg->sg.end)
 269                         break;
 270         } while (1);
 271 }
 272
 273 static inline void sk_msg_sg_copy_set(struct sk_msg *msg, u32 start)
 274 {
 275         sk_msg_sg_copy(msg, start, true);
 276 }
 277
 278 static inline void sk_msg_sg_copy_clear(struct sk_msg *msg, u32 start)
 279 {
 280         sk_msg_sg_copy(msg, start, false);
 281 }
 282
 283 static inline struct sk_psock *sk_psock(const struct sock *sk)
 284 {
 285         return rcu_dereference_sk_user_data(sk);
 286 }
 287
 288 static inline void sk_psock_queue_msg(struct sk_psock *psock,
 289                                       struct sk_msg *msg)
 290 {
 291         spin_lock_bh(&psock->ingress_lock);
 292         list_add_tail(&msg->list, &psock->ingress_msg);
 293         spin_unlock_bh(&psock->ingress_lock);
 294 }
 295
 296 static inline struct sk_msg *sk_psock_dequeue_msg(struct sk_psock *psock)
 297 {
 298         struct sk_msg *msg;
 299
 300         spin_lock_bh(&psock->ingress_lock);
 301         msg = list_first_entry_or_null(&psock->ingress_msg, struct sk_msg, list);
 302         if (msg)
 303                 list_del(&msg->list);
 304         spin_unlock_bh(&psock->ingress_lock);
 305         return msg;
 306 }
 307
 308 static inline struct sk_msg *sk_psock_peek_msg(struct sk_psock *psock)
 309 {
 310         struct sk_msg *msg;
 311
 312         spin_lock_bh(&psock->ingress_lock);
 313         msg = list_first_entry_or_null(&psock->ingress_msg, struct sk_msg, list);
 314         spin_unlock_bh(&psock->ingress_lock);
 315         return msg;
 316 }
 317
 318 static inline struct sk_msg *sk_psock_next_msg(struct sk_psock *psock,
 319                                                struct sk_msg *msg)
 320 {
 321         struct sk_msg *ret;
 322
 323         spin_lock_bh(&psock->ingress_lock);
 324         if (list_is_last(&msg->list, &psock->ingress_msg))
 325                 ret = NULL;
 326         else
 327                 ret = list_next_entry(msg, list);
 328         spin_unlock_bh(&psock->ingress_lock);
 329         return ret;
 330 }
 331
 332 static inline bool sk_psock_queue_empty(const struct sk_psock *psock)
 333 {
 334         return psock ? list_empty(&psock->ingress_msg) : true;
 335 }
 336
 337 static inline void kfree_sk_msg(struct sk_msg *msg)
 338 {
 339         if (msg->skb)
 340                 consume_skb(msg->skb);
 341         kfree(msg);
 342 }
 343
 344 static inline void sk_psock_report_error(struct sk_psock *psock, int err)
 345 {
 346         struct sock *sk = psock->sk;
 347
 348         sk->sk_err = err;
 349         sk_error_report(sk);
 350 }
 351
 352 struct sk_psock *sk_psock_init(struct sock *sk, int node);
 353 void sk_psock_stop(struct sk_psock *psock, bool wait);
 354
 355 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
 356 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock);
 357 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock);
 358 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock);
 359 #else
 360 static inline int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
 361 {
 362         return -EOPNOTSUPP;
 363 }
 364
 365 static inline void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
 366 {
 367 }
 368
 369 static inline void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
 370 {
 371 }
 372 #endif
 373
 374 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock);
 375 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock);
 376
 377 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
 378                          struct sk_msg *msg);
 379
 380 static inline struct sk_psock_link *sk_psock_init_link(void)
 381 {
 382         return kzalloc(sizeof(struct sk_psock_link),
 383                        GFP_ATOMIC | __GFP_NOWARN);
 384 }
 385
 386 static inline void sk_psock_free_link(struct sk_psock_link *link)
 387 {
 388         kfree(link);
 389 }
 390
 391 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock);
 392
 393 static inline void sk_psock_cork_free(struct sk_psock *psock)
 394 {
 395         if (psock->cork) {
 396                 sk_msg_free(psock->sk, psock->cork);
 397                 kfree(psock->cork);
 398                 psock->cork = NULL;
 399         }
 400 }
 401
 402 static inline void sk_psock_restore_proto(struct sock *sk,
 403                                           struct sk_psock *psock)
 404 {
 405         if (psock->psock_update_sk_prot)
 406                 psock->psock_update_sk_prot(sk, psock, true);
 407 }
 408
 409 static inline void sk_psock_set_state(struct sk_psock *psock,
 410                                       enum sk_psock_state_bits bit)
 411 {
 412         set_bit(bit, &psock->state);
 413 }
 414
 415 static inline void sk_psock_clear_state(struct sk_psock *psock,
 416                                         enum sk_psock_state_bits bit)
 417 {
 418         clear_bit(bit, &psock->state);
 419 }
 420
 421 static inline bool sk_psock_test_state(const struct sk_psock *psock,
 422                                        enum sk_psock_state_bits bit)
 423 {
 424         return test_bit(bit, &psock->state);
 425 }
 426
 427 static inline struct sk_psock *sk_psock_get(struct sock *sk)
 428 {
 429         struct sk_psock *psock;
 430
 431         rcu_read_lock();
 432         psock = sk_psock(sk);
 433         if (psock && !refcount_inc_not_zero(&psock->refcnt))
 434                 psock = NULL;
 435         rcu_read_unlock();
 436         return psock;
 437 }
 438
 439 void sk_psock_drop(struct sock *sk, struct sk_psock *psock);
 440
 441 static inline void sk_psock_put(struct sock *sk, struct sk_psock *psock)
 442 {
 443         if (refcount_dec_and_test(&psock->refcnt))
 444                 sk_psock_drop(sk, psock);
 445 }
 446
 447 static inline void sk_psock_data_ready(struct sock *sk, struct sk_psock *psock)
 448 {
 449         if (psock->saved_data_ready)
 450                 psock->saved_data_ready(sk);
 451         else
 452                 sk->sk_data_ready(sk);
 453 }
 454
 455 static inline void psock_set_prog(struct bpf_prog **pprog,
 456                                   struct bpf_prog *prog)
 457 {
 458         prog = xchg(pprog, prog);
 459         if (prog)
 460                 bpf_prog_put(prog);
 461 }
 462
 463 static inline int psock_replace_prog(struct bpf_prog **pprog,
 464                                      struct bpf_prog *prog,
 465                                      struct bpf_prog *old)
 466 {
 467         if (cmpxchg(pprog, old, prog) != old)
 468                 return -ENOENT;
 469
 470         if (old)
 471                 bpf_prog_put(old);
 472
 473         return 0;
 474 }
 475
 476 static inline void psock_progs_drop(struct sk_psock_progs *progs)
 477 {
 478         psock_set_prog(&progs->msg_parser, NULL);
 479         psock_set_prog(&progs->stream_parser, NULL);
 480         psock_set_prog(&progs->stream_verdict, NULL);
 481         psock_set_prog(&progs->skb_verdict, NULL);
 482 }
 483
 484 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb);
 485
 486 static inline bool sk_psock_strp_enabled(struct sk_psock *psock)
 487 {
 488         if (!psock)
 489                 return false;
 490         return !!psock->saved_data_ready;
 491 }
 492
 493 #if IS_ENABLED(CONFIG_NET_SOCK_MSG)
 494
 495 /* We only have one bit so far. */
 496 #define BPF_F_PTR_MASK ~(BPF_F_INGRESS)
 497
 498 static inline bool skb_bpf_ingress(const struct sk_buff *skb)
 499 {
 500         unsigned long sk_redir = skb->_sk_redir;
 501
 502         return sk_redir & BPF_F_INGRESS;
 503 }
 504
 505 static inline void skb_bpf_set_ingress(struct sk_buff *skb)
 506 {
 507         skb->_sk_redir |= BPF_F_INGRESS;
 508 }
 509
 510 static inline void skb_bpf_set_redir(struct sk_buff *skb, struct sock *sk_redir,
 511                                      bool ingress)
 512 {
 513         skb->_sk_redir = (unsigned long)sk_redir;
 514         if (ingress)
 515                 skb->_sk_redir |= BPF_F_INGRESS;
 516 }
 517
 518 static inline struct sock *skb_bpf_redirect_fetch(const struct sk_buff *skb)
 519 {
 520         unsigned long sk_redir = skb->_sk_redir;
 521
 522         return (struct sock *)(sk_redir & BPF_F_PTR_MASK);
 523 }
 524
 525 static inline void skb_bpf_redirect_clear(struct sk_buff *skb)
 526 {
 527         skb->_sk_redir = 0;
 528 }
 529 #endif /* CONFIG_NET_SOCK_MSG */
 530 #endif /* _LINUX_SKMSG_H */