1604326b4SDaniel Borkmann // SPDX-License-Identifier: GPL-2.0 2604326b4SDaniel Borkmann /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */ 3604326b4SDaniel Borkmann 4604326b4SDaniel Borkmann #include <linux/skmsg.h> 5604326b4SDaniel Borkmann #include <linux/filter.h> 6604326b4SDaniel Borkmann #include <linux/bpf.h> 7604326b4SDaniel Borkmann #include <linux/init.h> 8604326b4SDaniel Borkmann #include <linux/wait.h> 9604326b4SDaniel Borkmann 10604326b4SDaniel Borkmann #include <net/inet_common.h> 110608c69cSJohn Fastabend #include <net/tls.h> 12604326b4SDaniel Borkmann 13604326b4SDaniel Borkmann int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock, 1402c558b2SJohn Fastabend struct msghdr *msg, int len, int flags) 15604326b4SDaniel Borkmann { 16604326b4SDaniel Borkmann struct iov_iter *iter = &msg->msg_iter; 1702c558b2SJohn Fastabend int peek = flags & MSG_PEEK; 18604326b4SDaniel Borkmann int i, ret, copied = 0; 19604326b4SDaniel Borkmann struct sk_msg *msg_rx; 20604326b4SDaniel Borkmann 21604326b4SDaniel Borkmann msg_rx = list_first_entry_or_null(&psock->ingress_msg, 22604326b4SDaniel Borkmann struct sk_msg, list); 2302c558b2SJohn Fastabend 2402c558b2SJohn Fastabend while (copied != len) { 2502c558b2SJohn Fastabend struct scatterlist *sge; 2602c558b2SJohn Fastabend 27604326b4SDaniel Borkmann if (unlikely(!msg_rx)) 28604326b4SDaniel Borkmann break; 29604326b4SDaniel Borkmann 30604326b4SDaniel Borkmann i = msg_rx->sg.start; 31604326b4SDaniel Borkmann do { 32604326b4SDaniel Borkmann struct page *page; 33604326b4SDaniel Borkmann int copy; 34604326b4SDaniel Borkmann 35604326b4SDaniel Borkmann sge = sk_msg_elem(msg_rx, i); 36604326b4SDaniel Borkmann copy = sge->length; 37604326b4SDaniel Borkmann page = sg_page(sge); 38604326b4SDaniel Borkmann if (copied + copy > len) 39604326b4SDaniel Borkmann copy = len - copied; 40604326b4SDaniel Borkmann ret = copy_page_to_iter(page, sge->offset, copy, iter); 41604326b4SDaniel Borkmann if (ret != copy) { 42604326b4SDaniel Borkmann msg_rx->sg.start = i; 43604326b4SDaniel Borkmann return -EFAULT; 44604326b4SDaniel Borkmann } 45604326b4SDaniel Borkmann 46604326b4SDaniel Borkmann copied += copy; 4702c558b2SJohn Fastabend if (likely(!peek)) { 48604326b4SDaniel Borkmann sge->offset += copy; 49604326b4SDaniel Borkmann sge->length -= copy; 50604326b4SDaniel Borkmann sk_mem_uncharge(sk, copy); 513f4c3127SJohn Fastabend msg_rx->sg.size -= copy; 5202c558b2SJohn Fastabend 53604326b4SDaniel Borkmann if (!sge->length) { 5402c558b2SJohn Fastabend sk_msg_iter_var_next(i); 55604326b4SDaniel Borkmann if (!msg_rx->skb) 56604326b4SDaniel Borkmann put_page(page); 57604326b4SDaniel Borkmann } 5802c558b2SJohn Fastabend } else { 5902c558b2SJohn Fastabend sk_msg_iter_var_next(i); 6002c558b2SJohn Fastabend } 61604326b4SDaniel Borkmann 62604326b4SDaniel Borkmann if (copied == len) 63604326b4SDaniel Borkmann break; 64604326b4SDaniel Borkmann } while (i != msg_rx->sg.end); 65604326b4SDaniel Borkmann 6602c558b2SJohn Fastabend if (unlikely(peek)) { 6702c558b2SJohn Fastabend msg_rx = list_next_entry(msg_rx, list); 6802c558b2SJohn Fastabend continue; 6902c558b2SJohn Fastabend } 7002c558b2SJohn Fastabend 71604326b4SDaniel Borkmann msg_rx->sg.start = i; 72604326b4SDaniel Borkmann if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) { 73604326b4SDaniel Borkmann list_del(&msg_rx->list); 74604326b4SDaniel Borkmann if (msg_rx->skb) 75604326b4SDaniel Borkmann consume_skb(msg_rx->skb); 76604326b4SDaniel Borkmann kfree(msg_rx); 77604326b4SDaniel Borkmann } 7802c558b2SJohn Fastabend msg_rx = list_first_entry_or_null(&psock->ingress_msg, 7902c558b2SJohn Fastabend struct sk_msg, list); 80604326b4SDaniel Borkmann } 81604326b4SDaniel Borkmann 82604326b4SDaniel Borkmann return copied; 83604326b4SDaniel Borkmann } 84604326b4SDaniel Borkmann EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg); 85604326b4SDaniel Borkmann 86604326b4SDaniel Borkmann static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock, 87604326b4SDaniel Borkmann struct sk_msg *msg, u32 apply_bytes, int flags) 88604326b4SDaniel Borkmann { 89604326b4SDaniel Borkmann bool apply = apply_bytes; 90604326b4SDaniel Borkmann struct scatterlist *sge; 91604326b4SDaniel Borkmann u32 size, copied = 0; 92604326b4SDaniel Borkmann struct sk_msg *tmp; 93604326b4SDaniel Borkmann int i, ret = 0; 94604326b4SDaniel Borkmann 95604326b4SDaniel Borkmann tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL); 96604326b4SDaniel Borkmann if (unlikely(!tmp)) 97604326b4SDaniel Borkmann return -ENOMEM; 98604326b4SDaniel Borkmann 99604326b4SDaniel Borkmann lock_sock(sk); 100604326b4SDaniel Borkmann tmp->sg.start = msg->sg.start; 101604326b4SDaniel Borkmann i = msg->sg.start; 102604326b4SDaniel Borkmann do { 103604326b4SDaniel Borkmann sge = sk_msg_elem(msg, i); 104604326b4SDaniel Borkmann size = (apply && apply_bytes < sge->length) ? 105604326b4SDaniel Borkmann apply_bytes : sge->length; 106604326b4SDaniel Borkmann if (!sk_wmem_schedule(sk, size)) { 107604326b4SDaniel Borkmann if (!copied) 108604326b4SDaniel Borkmann ret = -ENOMEM; 109604326b4SDaniel Borkmann break; 110604326b4SDaniel Borkmann } 111604326b4SDaniel Borkmann 112604326b4SDaniel Borkmann sk_mem_charge(sk, size); 113604326b4SDaniel Borkmann sk_msg_xfer(tmp, msg, i, size); 114604326b4SDaniel Borkmann copied += size; 115604326b4SDaniel Borkmann if (sge->length) 116604326b4SDaniel Borkmann get_page(sk_msg_page(tmp, i)); 117604326b4SDaniel Borkmann sk_msg_iter_var_next(i); 118604326b4SDaniel Borkmann tmp->sg.end = i; 119604326b4SDaniel Borkmann if (apply) { 120604326b4SDaniel Borkmann apply_bytes -= size; 121604326b4SDaniel Borkmann if (!apply_bytes) 122604326b4SDaniel Borkmann break; 123604326b4SDaniel Borkmann } 124604326b4SDaniel Borkmann } while (i != msg->sg.end); 125604326b4SDaniel Borkmann 126604326b4SDaniel Borkmann if (!ret) { 127604326b4SDaniel Borkmann msg->sg.start = i; 128604326b4SDaniel Borkmann msg->sg.size -= apply_bytes; 129604326b4SDaniel Borkmann sk_psock_queue_msg(psock, tmp); 130552de910SJohn Fastabend sk_psock_data_ready(sk, psock); 131604326b4SDaniel Borkmann } else { 132604326b4SDaniel Borkmann sk_msg_free(sk, tmp); 133604326b4SDaniel Borkmann kfree(tmp); 134604326b4SDaniel Borkmann } 135604326b4SDaniel Borkmann 136604326b4SDaniel Borkmann release_sock(sk); 137604326b4SDaniel Borkmann return ret; 138604326b4SDaniel Borkmann } 139604326b4SDaniel Borkmann 140604326b4SDaniel Borkmann static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes, 141604326b4SDaniel Borkmann int flags, bool uncharge) 142604326b4SDaniel Borkmann { 143604326b4SDaniel Borkmann bool apply = apply_bytes; 144604326b4SDaniel Borkmann struct scatterlist *sge; 145604326b4SDaniel Borkmann struct page *page; 146604326b4SDaniel Borkmann int size, ret = 0; 147604326b4SDaniel Borkmann u32 off; 148604326b4SDaniel Borkmann 149604326b4SDaniel Borkmann while (1) { 1500608c69cSJohn Fastabend bool has_tx_ulp; 1510608c69cSJohn Fastabend 152604326b4SDaniel Borkmann sge = sk_msg_elem(msg, msg->sg.start); 153604326b4SDaniel Borkmann size = (apply && apply_bytes < sge->length) ? 154604326b4SDaniel Borkmann apply_bytes : sge->length; 155604326b4SDaniel Borkmann off = sge->offset; 156604326b4SDaniel Borkmann page = sg_page(sge); 157604326b4SDaniel Borkmann 158604326b4SDaniel Borkmann tcp_rate_check_app_limited(sk); 159604326b4SDaniel Borkmann retry: 1600608c69cSJohn Fastabend has_tx_ulp = tls_sw_has_ctx_tx(sk); 1610608c69cSJohn Fastabend if (has_tx_ulp) { 1620608c69cSJohn Fastabend flags |= MSG_SENDPAGE_NOPOLICY; 1630608c69cSJohn Fastabend ret = kernel_sendpage_locked(sk, 1640608c69cSJohn Fastabend page, off, size, flags); 1650608c69cSJohn Fastabend } else { 166604326b4SDaniel Borkmann ret = do_tcp_sendpages(sk, page, off, size, flags); 1670608c69cSJohn Fastabend } 1680608c69cSJohn Fastabend 169604326b4SDaniel Borkmann if (ret <= 0) 170604326b4SDaniel Borkmann return ret; 171604326b4SDaniel Borkmann if (apply) 172604326b4SDaniel Borkmann apply_bytes -= ret; 173604326b4SDaniel Borkmann msg->sg.size -= ret; 174604326b4SDaniel Borkmann sge->offset += ret; 175604326b4SDaniel Borkmann sge->length -= ret; 176604326b4SDaniel Borkmann if (uncharge) 177604326b4SDaniel Borkmann sk_mem_uncharge(sk, ret); 178604326b4SDaniel Borkmann if (ret != size) { 179604326b4SDaniel Borkmann size -= ret; 180604326b4SDaniel Borkmann off += ret; 181604326b4SDaniel Borkmann goto retry; 182604326b4SDaniel Borkmann } 183604326b4SDaniel Borkmann if (!sge->length) { 184604326b4SDaniel Borkmann put_page(page); 185604326b4SDaniel Borkmann sk_msg_iter_next(msg, start); 186604326b4SDaniel Borkmann sg_init_table(sge, 1); 187604326b4SDaniel Borkmann if (msg->sg.start == msg->sg.end) 188604326b4SDaniel Borkmann break; 189604326b4SDaniel Borkmann } 190604326b4SDaniel Borkmann if (apply && !apply_bytes) 191604326b4SDaniel Borkmann break; 192604326b4SDaniel Borkmann } 193604326b4SDaniel Borkmann 194604326b4SDaniel Borkmann return 0; 195604326b4SDaniel Borkmann } 196604326b4SDaniel Borkmann 197604326b4SDaniel Borkmann static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg, 198604326b4SDaniel Borkmann u32 apply_bytes, int flags, bool uncharge) 199604326b4SDaniel Borkmann { 200604326b4SDaniel Borkmann int ret; 201604326b4SDaniel Borkmann 202604326b4SDaniel Borkmann lock_sock(sk); 203604326b4SDaniel Borkmann ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge); 204604326b4SDaniel Borkmann release_sock(sk); 205604326b4SDaniel Borkmann return ret; 206604326b4SDaniel Borkmann } 207604326b4SDaniel Borkmann 208604326b4SDaniel Borkmann int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg, 209604326b4SDaniel Borkmann u32 bytes, int flags) 210604326b4SDaniel Borkmann { 211604326b4SDaniel Borkmann bool ingress = sk_msg_to_ingress(msg); 212604326b4SDaniel Borkmann struct sk_psock *psock = sk_psock_get(sk); 213604326b4SDaniel Borkmann int ret; 214604326b4SDaniel Borkmann 215604326b4SDaniel Borkmann if (unlikely(!psock)) { 216604326b4SDaniel Borkmann sk_msg_free(sk, msg); 217604326b4SDaniel Borkmann return 0; 218604326b4SDaniel Borkmann } 219604326b4SDaniel Borkmann ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) : 220604326b4SDaniel Borkmann tcp_bpf_push_locked(sk, msg, bytes, flags, false); 221604326b4SDaniel Borkmann sk_psock_put(sk, psock); 222604326b4SDaniel Borkmann return ret; 223604326b4SDaniel Borkmann } 224604326b4SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir); 225604326b4SDaniel Borkmann 226a2652798SYueHaibing #ifdef CONFIG_BPF_STREAM_PARSER 227a2652798SYueHaibing static bool tcp_bpf_stream_read(const struct sock *sk) 228a2652798SYueHaibing { 229a2652798SYueHaibing struct sk_psock *psock; 230a2652798SYueHaibing bool empty = true; 231a2652798SYueHaibing 232a2652798SYueHaibing rcu_read_lock(); 233a2652798SYueHaibing psock = sk_psock(sk); 234a2652798SYueHaibing if (likely(psock)) 235a2652798SYueHaibing empty = list_empty(&psock->ingress_msg); 236a2652798SYueHaibing rcu_read_unlock(); 237a2652798SYueHaibing return !empty; 238a2652798SYueHaibing } 239a2652798SYueHaibing 240c0fd336eSYueHaibing static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock, 241c0fd336eSYueHaibing int flags, long timeo, int *err) 242c0fd336eSYueHaibing { 243c0fd336eSYueHaibing DEFINE_WAIT_FUNC(wait, woken_wake_function); 244c0fd336eSYueHaibing int ret = 0; 245c0fd336eSYueHaibing 246c0fd336eSYueHaibing if (!timeo) 247c0fd336eSYueHaibing return ret; 248c0fd336eSYueHaibing 249c0fd336eSYueHaibing add_wait_queue(sk_sleep(sk), &wait); 250c0fd336eSYueHaibing sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 251c0fd336eSYueHaibing ret = sk_wait_event(sk, &timeo, 252c0fd336eSYueHaibing !list_empty(&psock->ingress_msg) || 253c0fd336eSYueHaibing !skb_queue_empty(&sk->sk_receive_queue), &wait); 254c0fd336eSYueHaibing sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 255c0fd336eSYueHaibing remove_wait_queue(sk_sleep(sk), &wait); 256c0fd336eSYueHaibing return ret; 257c0fd336eSYueHaibing } 258c0fd336eSYueHaibing 259c0fd336eSYueHaibing static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 260c0fd336eSYueHaibing int nonblock, int flags, int *addr_len) 261c0fd336eSYueHaibing { 262c0fd336eSYueHaibing struct sk_psock *psock; 263c0fd336eSYueHaibing int copied, ret; 264c0fd336eSYueHaibing 265*18f02ad1SXiyu Yang if (unlikely(flags & MSG_ERRQUEUE)) 266*18f02ad1SXiyu Yang return inet_recv_error(sk, msg, len, addr_len); 267*18f02ad1SXiyu Yang 268c0fd336eSYueHaibing psock = sk_psock_get(sk); 269c0fd336eSYueHaibing if (unlikely(!psock)) 270c0fd336eSYueHaibing return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 271c0fd336eSYueHaibing if (!skb_queue_empty(&sk->sk_receive_queue) && 272*18f02ad1SXiyu Yang sk_psock_queue_empty(psock)) { 273*18f02ad1SXiyu Yang sk_psock_put(sk, psock); 274c0fd336eSYueHaibing return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 275*18f02ad1SXiyu Yang } 276c0fd336eSYueHaibing lock_sock(sk); 277c0fd336eSYueHaibing msg_bytes_ready: 278c0fd336eSYueHaibing copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags); 279c0fd336eSYueHaibing if (!copied) { 280c0fd336eSYueHaibing int data, err = 0; 281c0fd336eSYueHaibing long timeo; 282c0fd336eSYueHaibing 283c0fd336eSYueHaibing timeo = sock_rcvtimeo(sk, nonblock); 284c0fd336eSYueHaibing data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err); 285c0fd336eSYueHaibing if (data) { 286c0fd336eSYueHaibing if (!sk_psock_queue_empty(psock)) 287c0fd336eSYueHaibing goto msg_bytes_ready; 288c0fd336eSYueHaibing release_sock(sk); 289c0fd336eSYueHaibing sk_psock_put(sk, psock); 290c0fd336eSYueHaibing return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 291c0fd336eSYueHaibing } 292c0fd336eSYueHaibing if (err) { 293c0fd336eSYueHaibing ret = err; 294c0fd336eSYueHaibing goto out; 295c0fd336eSYueHaibing } 296c0fd336eSYueHaibing copied = -EAGAIN; 297c0fd336eSYueHaibing } 298c0fd336eSYueHaibing ret = copied; 299c0fd336eSYueHaibing out: 300c0fd336eSYueHaibing release_sock(sk); 301c0fd336eSYueHaibing sk_psock_put(sk, psock); 302c0fd336eSYueHaibing return ret; 303c0fd336eSYueHaibing } 304c0fd336eSYueHaibing 305604326b4SDaniel Borkmann static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock, 306604326b4SDaniel Borkmann struct sk_msg *msg, int *copied, int flags) 307604326b4SDaniel Borkmann { 308031097d9SJakub Kicinski bool cork = false, enospc = sk_msg_full(msg); 309604326b4SDaniel Borkmann struct sock *sk_redir; 3107246d8edSJohn Fastabend u32 tosend, delta = 0; 311604326b4SDaniel Borkmann int ret; 312604326b4SDaniel Borkmann 313604326b4SDaniel Borkmann more_data: 3147246d8edSJohn Fastabend if (psock->eval == __SK_NONE) { 3157246d8edSJohn Fastabend /* Track delta in msg size to add/subtract it on SK_DROP from 3167246d8edSJohn Fastabend * returned to user copied size. This ensures user doesn't 3177246d8edSJohn Fastabend * get a positive return code with msg_cut_data and SK_DROP 3187246d8edSJohn Fastabend * verdict. 3197246d8edSJohn Fastabend */ 3207246d8edSJohn Fastabend delta = msg->sg.size; 321604326b4SDaniel Borkmann psock->eval = sk_psock_msg_verdict(sk, psock, msg); 3227246d8edSJohn Fastabend delta -= msg->sg.size; 3237246d8edSJohn Fastabend } 324604326b4SDaniel Borkmann 325604326b4SDaniel Borkmann if (msg->cork_bytes && 326604326b4SDaniel Borkmann msg->cork_bytes > msg->sg.size && !enospc) { 327604326b4SDaniel Borkmann psock->cork_bytes = msg->cork_bytes - msg->sg.size; 328604326b4SDaniel Borkmann if (!psock->cork) { 329604326b4SDaniel Borkmann psock->cork = kzalloc(sizeof(*psock->cork), 330604326b4SDaniel Borkmann GFP_ATOMIC | __GFP_NOWARN); 331604326b4SDaniel Borkmann if (!psock->cork) 332604326b4SDaniel Borkmann return -ENOMEM; 333604326b4SDaniel Borkmann } 334604326b4SDaniel Borkmann memcpy(psock->cork, msg, sizeof(*msg)); 335604326b4SDaniel Borkmann return 0; 336604326b4SDaniel Borkmann } 337604326b4SDaniel Borkmann 338604326b4SDaniel Borkmann tosend = msg->sg.size; 339604326b4SDaniel Borkmann if (psock->apply_bytes && psock->apply_bytes < tosend) 340604326b4SDaniel Borkmann tosend = psock->apply_bytes; 341604326b4SDaniel Borkmann 342604326b4SDaniel Borkmann switch (psock->eval) { 343604326b4SDaniel Borkmann case __SK_PASS: 344604326b4SDaniel Borkmann ret = tcp_bpf_push(sk, msg, tosend, flags, true); 345604326b4SDaniel Borkmann if (unlikely(ret)) { 346604326b4SDaniel Borkmann *copied -= sk_msg_free(sk, msg); 347604326b4SDaniel Borkmann break; 348604326b4SDaniel Borkmann } 349604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 350604326b4SDaniel Borkmann break; 351604326b4SDaniel Borkmann case __SK_REDIRECT: 352604326b4SDaniel Borkmann sk_redir = psock->sk_redir; 353604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 354604326b4SDaniel Borkmann if (psock->cork) { 355604326b4SDaniel Borkmann cork = true; 356604326b4SDaniel Borkmann psock->cork = NULL; 357604326b4SDaniel Borkmann } 358604326b4SDaniel Borkmann sk_msg_return(sk, msg, tosend); 359604326b4SDaniel Borkmann release_sock(sk); 360604326b4SDaniel Borkmann ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags); 361604326b4SDaniel Borkmann lock_sock(sk); 362604326b4SDaniel Borkmann if (unlikely(ret < 0)) { 363604326b4SDaniel Borkmann int free = sk_msg_free_nocharge(sk, msg); 364604326b4SDaniel Borkmann 365604326b4SDaniel Borkmann if (!cork) 366604326b4SDaniel Borkmann *copied -= free; 367604326b4SDaniel Borkmann } 368604326b4SDaniel Borkmann if (cork) { 369604326b4SDaniel Borkmann sk_msg_free(sk, msg); 370604326b4SDaniel Borkmann kfree(msg); 371604326b4SDaniel Borkmann msg = NULL; 372604326b4SDaniel Borkmann ret = 0; 373604326b4SDaniel Borkmann } 374604326b4SDaniel Borkmann break; 375604326b4SDaniel Borkmann case __SK_DROP: 376604326b4SDaniel Borkmann default: 377604326b4SDaniel Borkmann sk_msg_free_partial(sk, msg, tosend); 378604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 3797246d8edSJohn Fastabend *copied -= (tosend + delta); 380604326b4SDaniel Borkmann return -EACCES; 381604326b4SDaniel Borkmann } 382604326b4SDaniel Borkmann 383604326b4SDaniel Borkmann if (likely(!ret)) { 384604326b4SDaniel Borkmann if (!psock->apply_bytes) { 385604326b4SDaniel Borkmann psock->eval = __SK_NONE; 386604326b4SDaniel Borkmann if (psock->sk_redir) { 387604326b4SDaniel Borkmann sock_put(psock->sk_redir); 388604326b4SDaniel Borkmann psock->sk_redir = NULL; 389604326b4SDaniel Borkmann } 390604326b4SDaniel Borkmann } 391604326b4SDaniel Borkmann if (msg && 392604326b4SDaniel Borkmann msg->sg.data[msg->sg.start].page_link && 393604326b4SDaniel Borkmann msg->sg.data[msg->sg.start].length) 394604326b4SDaniel Borkmann goto more_data; 395604326b4SDaniel Borkmann } 396604326b4SDaniel Borkmann return ret; 397604326b4SDaniel Borkmann } 398604326b4SDaniel Borkmann 399604326b4SDaniel Borkmann static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 400604326b4SDaniel Borkmann { 401604326b4SDaniel Borkmann struct sk_msg tmp, *msg_tx = NULL; 402604326b4SDaniel Borkmann int copied = 0, err = 0; 403604326b4SDaniel Borkmann struct sk_psock *psock; 404604326b4SDaniel Borkmann long timeo; 40541477662SJakub Kicinski int flags; 40641477662SJakub Kicinski 40741477662SJakub Kicinski /* Don't let internal do_tcp_sendpages() flags through */ 40841477662SJakub Kicinski flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED); 40941477662SJakub Kicinski flags |= MSG_NO_SHARED_FRAGS; 410604326b4SDaniel Borkmann 411604326b4SDaniel Borkmann psock = sk_psock_get(sk); 412604326b4SDaniel Borkmann if (unlikely(!psock)) 413604326b4SDaniel Borkmann return tcp_sendmsg(sk, msg, size); 414604326b4SDaniel Borkmann 415604326b4SDaniel Borkmann lock_sock(sk); 416604326b4SDaniel Borkmann timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 417604326b4SDaniel Borkmann while (msg_data_left(msg)) { 418604326b4SDaniel Borkmann bool enospc = false; 419604326b4SDaniel Borkmann u32 copy, osize; 420604326b4SDaniel Borkmann 421604326b4SDaniel Borkmann if (sk->sk_err) { 422604326b4SDaniel Borkmann err = -sk->sk_err; 423604326b4SDaniel Borkmann goto out_err; 424604326b4SDaniel Borkmann } 425604326b4SDaniel Borkmann 426604326b4SDaniel Borkmann copy = msg_data_left(msg); 427604326b4SDaniel Borkmann if (!sk_stream_memory_free(sk)) 428604326b4SDaniel Borkmann goto wait_for_sndbuf; 429604326b4SDaniel Borkmann if (psock->cork) { 430604326b4SDaniel Borkmann msg_tx = psock->cork; 431604326b4SDaniel Borkmann } else { 432604326b4SDaniel Borkmann msg_tx = &tmp; 433604326b4SDaniel Borkmann sk_msg_init(msg_tx); 434604326b4SDaniel Borkmann } 435604326b4SDaniel Borkmann 436604326b4SDaniel Borkmann osize = msg_tx->sg.size; 437604326b4SDaniel Borkmann err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1); 438604326b4SDaniel Borkmann if (err) { 439604326b4SDaniel Borkmann if (err != -ENOSPC) 440604326b4SDaniel Borkmann goto wait_for_memory; 441604326b4SDaniel Borkmann enospc = true; 442604326b4SDaniel Borkmann copy = msg_tx->sg.size - osize; 443604326b4SDaniel Borkmann } 444604326b4SDaniel Borkmann 445604326b4SDaniel Borkmann err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx, 446604326b4SDaniel Borkmann copy); 447604326b4SDaniel Borkmann if (err < 0) { 448604326b4SDaniel Borkmann sk_msg_trim(sk, msg_tx, osize); 449604326b4SDaniel Borkmann goto out_err; 450604326b4SDaniel Borkmann } 451604326b4SDaniel Borkmann 452604326b4SDaniel Borkmann copied += copy; 453604326b4SDaniel Borkmann if (psock->cork_bytes) { 454604326b4SDaniel Borkmann if (size > psock->cork_bytes) 455604326b4SDaniel Borkmann psock->cork_bytes = 0; 456604326b4SDaniel Borkmann else 457604326b4SDaniel Borkmann psock->cork_bytes -= size; 458604326b4SDaniel Borkmann if (psock->cork_bytes && !enospc) 459604326b4SDaniel Borkmann goto out_err; 460604326b4SDaniel Borkmann /* All cork bytes are accounted, rerun the prog. */ 461604326b4SDaniel Borkmann psock->eval = __SK_NONE; 462604326b4SDaniel Borkmann psock->cork_bytes = 0; 463604326b4SDaniel Borkmann } 464604326b4SDaniel Borkmann 465604326b4SDaniel Borkmann err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags); 466604326b4SDaniel Borkmann if (unlikely(err < 0)) 467604326b4SDaniel Borkmann goto out_err; 468604326b4SDaniel Borkmann continue; 469604326b4SDaniel Borkmann wait_for_sndbuf: 470604326b4SDaniel Borkmann set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 471604326b4SDaniel Borkmann wait_for_memory: 472604326b4SDaniel Borkmann err = sk_stream_wait_memory(sk, &timeo); 473604326b4SDaniel Borkmann if (err) { 474604326b4SDaniel Borkmann if (msg_tx && msg_tx != psock->cork) 475604326b4SDaniel Borkmann sk_msg_free(sk, msg_tx); 476604326b4SDaniel Borkmann goto out_err; 477604326b4SDaniel Borkmann } 478604326b4SDaniel Borkmann } 479604326b4SDaniel Borkmann out_err: 480604326b4SDaniel Borkmann if (err < 0) 481604326b4SDaniel Borkmann err = sk_stream_error(sk, msg->msg_flags, err); 482604326b4SDaniel Borkmann release_sock(sk); 483604326b4SDaniel Borkmann sk_psock_put(sk, psock); 484604326b4SDaniel Borkmann return copied ? copied : err; 485604326b4SDaniel Borkmann } 486604326b4SDaniel Borkmann 487604326b4SDaniel Borkmann static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset, 488604326b4SDaniel Borkmann size_t size, int flags) 489604326b4SDaniel Borkmann { 490604326b4SDaniel Borkmann struct sk_msg tmp, *msg = NULL; 491604326b4SDaniel Borkmann int err = 0, copied = 0; 492604326b4SDaniel Borkmann struct sk_psock *psock; 493604326b4SDaniel Borkmann bool enospc = false; 494604326b4SDaniel Borkmann 495604326b4SDaniel Borkmann psock = sk_psock_get(sk); 496604326b4SDaniel Borkmann if (unlikely(!psock)) 497604326b4SDaniel Borkmann return tcp_sendpage(sk, page, offset, size, flags); 498604326b4SDaniel Borkmann 499604326b4SDaniel Borkmann lock_sock(sk); 500604326b4SDaniel Borkmann if (psock->cork) { 501604326b4SDaniel Borkmann msg = psock->cork; 502604326b4SDaniel Borkmann } else { 503604326b4SDaniel Borkmann msg = &tmp; 504604326b4SDaniel Borkmann sk_msg_init(msg); 505604326b4SDaniel Borkmann } 506604326b4SDaniel Borkmann 507604326b4SDaniel Borkmann /* Catch case where ring is full and sendpage is stalled. */ 508604326b4SDaniel Borkmann if (unlikely(sk_msg_full(msg))) 509604326b4SDaniel Borkmann goto out_err; 510604326b4SDaniel Borkmann 511604326b4SDaniel Borkmann sk_msg_page_add(msg, page, size, offset); 512604326b4SDaniel Borkmann sk_mem_charge(sk, size); 513604326b4SDaniel Borkmann copied = size; 514604326b4SDaniel Borkmann if (sk_msg_full(msg)) 515604326b4SDaniel Borkmann enospc = true; 516604326b4SDaniel Borkmann if (psock->cork_bytes) { 517604326b4SDaniel Borkmann if (size > psock->cork_bytes) 518604326b4SDaniel Borkmann psock->cork_bytes = 0; 519604326b4SDaniel Borkmann else 520604326b4SDaniel Borkmann psock->cork_bytes -= size; 521604326b4SDaniel Borkmann if (psock->cork_bytes && !enospc) 522604326b4SDaniel Borkmann goto out_err; 523604326b4SDaniel Borkmann /* All cork bytes are accounted, rerun the prog. */ 524604326b4SDaniel Borkmann psock->eval = __SK_NONE; 525604326b4SDaniel Borkmann psock->cork_bytes = 0; 526604326b4SDaniel Borkmann } 527604326b4SDaniel Borkmann 528604326b4SDaniel Borkmann err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags); 529604326b4SDaniel Borkmann out_err: 530604326b4SDaniel Borkmann release_sock(sk); 531604326b4SDaniel Borkmann sk_psock_put(sk, psock); 532604326b4SDaniel Borkmann return copied ? copied : err; 533604326b4SDaniel Borkmann } 534604326b4SDaniel Borkmann 535604326b4SDaniel Borkmann enum { 536604326b4SDaniel Borkmann TCP_BPF_IPV4, 537604326b4SDaniel Borkmann TCP_BPF_IPV6, 538604326b4SDaniel Borkmann TCP_BPF_NUM_PROTS, 539604326b4SDaniel Borkmann }; 540604326b4SDaniel Borkmann 541604326b4SDaniel Borkmann enum { 542604326b4SDaniel Borkmann TCP_BPF_BASE, 543604326b4SDaniel Borkmann TCP_BPF_TX, 544604326b4SDaniel Borkmann TCP_BPF_NUM_CFGS, 545604326b4SDaniel Borkmann }; 546604326b4SDaniel Borkmann 547604326b4SDaniel Borkmann static struct proto *tcpv6_prot_saved __read_mostly; 548604326b4SDaniel Borkmann static DEFINE_SPINLOCK(tcpv6_prot_lock); 549604326b4SDaniel Borkmann static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS]; 550604326b4SDaniel Borkmann 551604326b4SDaniel Borkmann static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS], 552604326b4SDaniel Borkmann struct proto *base) 553604326b4SDaniel Borkmann { 554604326b4SDaniel Borkmann prot[TCP_BPF_BASE] = *base; 555f747632bSLorenz Bauer prot[TCP_BPF_BASE].unhash = sock_map_unhash; 556f747632bSLorenz Bauer prot[TCP_BPF_BASE].close = sock_map_close; 557604326b4SDaniel Borkmann prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg; 558604326b4SDaniel Borkmann prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read; 559604326b4SDaniel Borkmann 560604326b4SDaniel Borkmann prot[TCP_BPF_TX] = prot[TCP_BPF_BASE]; 561604326b4SDaniel Borkmann prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg; 562604326b4SDaniel Borkmann prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage; 563604326b4SDaniel Borkmann } 564604326b4SDaniel Borkmann 565604326b4SDaniel Borkmann static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops) 566604326b4SDaniel Borkmann { 567604326b4SDaniel Borkmann if (sk->sk_family == AF_INET6 && 568604326b4SDaniel Borkmann unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) { 569604326b4SDaniel Borkmann spin_lock_bh(&tcpv6_prot_lock); 570604326b4SDaniel Borkmann if (likely(ops != tcpv6_prot_saved)) { 571604326b4SDaniel Borkmann tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops); 572604326b4SDaniel Borkmann smp_store_release(&tcpv6_prot_saved, ops); 573604326b4SDaniel Borkmann } 574604326b4SDaniel Borkmann spin_unlock_bh(&tcpv6_prot_lock); 575604326b4SDaniel Borkmann } 576604326b4SDaniel Borkmann } 577604326b4SDaniel Borkmann 578604326b4SDaniel Borkmann static int __init tcp_bpf_v4_build_proto(void) 579604326b4SDaniel Borkmann { 580604326b4SDaniel Borkmann tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot); 581604326b4SDaniel Borkmann return 0; 582604326b4SDaniel Borkmann } 583604326b4SDaniel Borkmann core_initcall(tcp_bpf_v4_build_proto); 584604326b4SDaniel Borkmann 585604326b4SDaniel Borkmann static int tcp_bpf_assert_proto_ops(struct proto *ops) 586604326b4SDaniel Borkmann { 587604326b4SDaniel Borkmann /* In order to avoid retpoline, we make assumptions when we call 588604326b4SDaniel Borkmann * into ops if e.g. a psock is not present. Make sure they are 589604326b4SDaniel Borkmann * indeed valid assumptions. 590604326b4SDaniel Borkmann */ 591604326b4SDaniel Borkmann return ops->recvmsg == tcp_recvmsg && 592604326b4SDaniel Borkmann ops->sendmsg == tcp_sendmsg && 593604326b4SDaniel Borkmann ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP; 594604326b4SDaniel Borkmann } 595604326b4SDaniel Borkmann 596f747632bSLorenz Bauer struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock) 597604326b4SDaniel Borkmann { 598d19da360SLorenz Bauer int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 599d19da360SLorenz Bauer int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 600604326b4SDaniel Borkmann 601d19da360SLorenz Bauer if (!psock->sk_proto) { 602d19da360SLorenz Bauer struct proto *ops = READ_ONCE(sk->sk_prot); 603604326b4SDaniel Borkmann 604d19da360SLorenz Bauer if (tcp_bpf_assert_proto_ops(ops)) 605d19da360SLorenz Bauer return ERR_PTR(-EINVAL); 606d19da360SLorenz Bauer 607d19da360SLorenz Bauer tcp_bpf_check_v6_needs_rebuild(sk, ops); 608d19da360SLorenz Bauer } 609d19da360SLorenz Bauer 610d19da360SLorenz Bauer return &tcp_bpf_prots[family][config]; 611604326b4SDaniel Borkmann } 612604326b4SDaniel Borkmann 613e8025155SJakub Sitnicki /* If a child got cloned from a listening socket that had tcp_bpf 614e8025155SJakub Sitnicki * protocol callbacks installed, we need to restore the callbacks to 615e8025155SJakub Sitnicki * the default ones because the child does not inherit the psock state 616e8025155SJakub Sitnicki * that tcp_bpf callbacks expect. 617e8025155SJakub Sitnicki */ 618e8025155SJakub Sitnicki void tcp_bpf_clone(const struct sock *sk, struct sock *newsk) 619e8025155SJakub Sitnicki { 620e8025155SJakub Sitnicki int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 621e8025155SJakub Sitnicki struct proto *prot = newsk->sk_prot; 622e8025155SJakub Sitnicki 623e8025155SJakub Sitnicki if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE]) 624e8025155SJakub Sitnicki newsk->sk_prot = sk->sk_prot_creator; 625e8025155SJakub Sitnicki } 626f747632bSLorenz Bauer #endif /* CONFIG_BPF_STREAM_PARSER */ 627