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> 11604326b4SDaniel Borkmann 12604326b4SDaniel Borkmann static bool tcp_bpf_stream_read(const struct sock *sk) 13604326b4SDaniel Borkmann { 14604326b4SDaniel Borkmann struct sk_psock *psock; 15604326b4SDaniel Borkmann bool empty = true; 16604326b4SDaniel Borkmann 17604326b4SDaniel Borkmann rcu_read_lock(); 18604326b4SDaniel Borkmann psock = sk_psock(sk); 19604326b4SDaniel Borkmann if (likely(psock)) 20604326b4SDaniel Borkmann empty = list_empty(&psock->ingress_msg); 21604326b4SDaniel Borkmann rcu_read_unlock(); 22604326b4SDaniel Borkmann return !empty; 23604326b4SDaniel Borkmann } 24604326b4SDaniel Borkmann 25604326b4SDaniel Borkmann static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock, 26604326b4SDaniel Borkmann int flags, long timeo, int *err) 27604326b4SDaniel Borkmann { 28604326b4SDaniel Borkmann DEFINE_WAIT_FUNC(wait, woken_wake_function); 29604326b4SDaniel Borkmann int ret; 30604326b4SDaniel Borkmann 31604326b4SDaniel Borkmann add_wait_queue(sk_sleep(sk), &wait); 32604326b4SDaniel Borkmann sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 33604326b4SDaniel Borkmann ret = sk_wait_event(sk, &timeo, 34604326b4SDaniel Borkmann !list_empty(&psock->ingress_msg) || 35604326b4SDaniel Borkmann !skb_queue_empty(&sk->sk_receive_queue), &wait); 36604326b4SDaniel Borkmann sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 37604326b4SDaniel Borkmann remove_wait_queue(sk_sleep(sk), &wait); 38604326b4SDaniel Borkmann return ret; 39604326b4SDaniel Borkmann } 40604326b4SDaniel Borkmann 41604326b4SDaniel Borkmann int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock, 4202c558b2SJohn Fastabend struct msghdr *msg, int len, int flags) 43604326b4SDaniel Borkmann { 44604326b4SDaniel Borkmann struct iov_iter *iter = &msg->msg_iter; 4502c558b2SJohn Fastabend int peek = flags & MSG_PEEK; 46604326b4SDaniel Borkmann int i, ret, copied = 0; 47604326b4SDaniel Borkmann struct sk_msg *msg_rx; 48604326b4SDaniel Borkmann 49604326b4SDaniel Borkmann msg_rx = list_first_entry_or_null(&psock->ingress_msg, 50604326b4SDaniel Borkmann struct sk_msg, list); 5102c558b2SJohn Fastabend 5202c558b2SJohn Fastabend while (copied != len) { 5302c558b2SJohn Fastabend struct scatterlist *sge; 5402c558b2SJohn Fastabend 55604326b4SDaniel Borkmann if (unlikely(!msg_rx)) 56604326b4SDaniel Borkmann break; 57604326b4SDaniel Borkmann 58604326b4SDaniel Borkmann i = msg_rx->sg.start; 59604326b4SDaniel Borkmann do { 60604326b4SDaniel Borkmann struct page *page; 61604326b4SDaniel Borkmann int copy; 62604326b4SDaniel Borkmann 63604326b4SDaniel Borkmann sge = sk_msg_elem(msg_rx, i); 64604326b4SDaniel Borkmann copy = sge->length; 65604326b4SDaniel Borkmann page = sg_page(sge); 66604326b4SDaniel Borkmann if (copied + copy > len) 67604326b4SDaniel Borkmann copy = len - copied; 68604326b4SDaniel Borkmann ret = copy_page_to_iter(page, sge->offset, copy, iter); 69604326b4SDaniel Borkmann if (ret != copy) { 70604326b4SDaniel Borkmann msg_rx->sg.start = i; 71604326b4SDaniel Borkmann return -EFAULT; 72604326b4SDaniel Borkmann } 73604326b4SDaniel Borkmann 74604326b4SDaniel Borkmann copied += copy; 7502c558b2SJohn Fastabend if (likely(!peek)) { 76604326b4SDaniel Borkmann sge->offset += copy; 77604326b4SDaniel Borkmann sge->length -= copy; 78604326b4SDaniel Borkmann sk_mem_uncharge(sk, copy); 793f4c3127SJohn Fastabend msg_rx->sg.size -= copy; 8002c558b2SJohn Fastabend 81604326b4SDaniel Borkmann if (!sge->length) { 8202c558b2SJohn Fastabend sk_msg_iter_var_next(i); 83604326b4SDaniel Borkmann if (!msg_rx->skb) 84604326b4SDaniel Borkmann put_page(page); 85604326b4SDaniel Borkmann } 8602c558b2SJohn Fastabend } else { 8702c558b2SJohn Fastabend sk_msg_iter_var_next(i); 8802c558b2SJohn Fastabend } 89604326b4SDaniel Borkmann 90604326b4SDaniel Borkmann if (copied == len) 91604326b4SDaniel Borkmann break; 92604326b4SDaniel Borkmann } while (i != msg_rx->sg.end); 93604326b4SDaniel Borkmann 9402c558b2SJohn Fastabend if (unlikely(peek)) { 9502c558b2SJohn Fastabend msg_rx = list_next_entry(msg_rx, list); 9602c558b2SJohn Fastabend continue; 9702c558b2SJohn Fastabend } 9802c558b2SJohn Fastabend 99604326b4SDaniel Borkmann msg_rx->sg.start = i; 100604326b4SDaniel Borkmann if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) { 101604326b4SDaniel Borkmann list_del(&msg_rx->list); 102604326b4SDaniel Borkmann if (msg_rx->skb) 103604326b4SDaniel Borkmann consume_skb(msg_rx->skb); 104604326b4SDaniel Borkmann kfree(msg_rx); 105604326b4SDaniel Borkmann } 10602c558b2SJohn Fastabend msg_rx = list_first_entry_or_null(&psock->ingress_msg, 10702c558b2SJohn Fastabend struct sk_msg, list); 108604326b4SDaniel Borkmann } 109604326b4SDaniel Borkmann 110604326b4SDaniel Borkmann return copied; 111604326b4SDaniel Borkmann } 112604326b4SDaniel Borkmann EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg); 113604326b4SDaniel Borkmann 114604326b4SDaniel Borkmann int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 115604326b4SDaniel Borkmann int nonblock, int flags, int *addr_len) 116604326b4SDaniel Borkmann { 117604326b4SDaniel Borkmann struct sk_psock *psock; 118604326b4SDaniel Borkmann int copied, ret; 119604326b4SDaniel Borkmann 120604326b4SDaniel Borkmann if (unlikely(flags & MSG_ERRQUEUE)) 121604326b4SDaniel Borkmann return inet_recv_error(sk, msg, len, addr_len); 122604326b4SDaniel Borkmann if (!skb_queue_empty(&sk->sk_receive_queue)) 123604326b4SDaniel Borkmann return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 124604326b4SDaniel Borkmann 125604326b4SDaniel Borkmann psock = sk_psock_get(sk); 126604326b4SDaniel Borkmann if (unlikely(!psock)) 127604326b4SDaniel Borkmann return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 128604326b4SDaniel Borkmann lock_sock(sk); 129604326b4SDaniel Borkmann msg_bytes_ready: 13002c558b2SJohn Fastabend copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags); 131604326b4SDaniel Borkmann if (!copied) { 132604326b4SDaniel Borkmann int data, err = 0; 133604326b4SDaniel Borkmann long timeo; 134604326b4SDaniel Borkmann 135604326b4SDaniel Borkmann timeo = sock_rcvtimeo(sk, nonblock); 136604326b4SDaniel Borkmann data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err); 137604326b4SDaniel Borkmann if (data) { 138604326b4SDaniel Borkmann if (skb_queue_empty(&sk->sk_receive_queue)) 139604326b4SDaniel Borkmann goto msg_bytes_ready; 140604326b4SDaniel Borkmann release_sock(sk); 141604326b4SDaniel Borkmann sk_psock_put(sk, psock); 142604326b4SDaniel Borkmann return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 143604326b4SDaniel Borkmann } 144604326b4SDaniel Borkmann if (err) { 145604326b4SDaniel Borkmann ret = err; 146604326b4SDaniel Borkmann goto out; 147604326b4SDaniel Borkmann } 14827b31e68SJohn Fastabend copied = -EAGAIN; 149604326b4SDaniel Borkmann } 150604326b4SDaniel Borkmann ret = copied; 151604326b4SDaniel Borkmann out: 152604326b4SDaniel Borkmann release_sock(sk); 153604326b4SDaniel Borkmann sk_psock_put(sk, psock); 154604326b4SDaniel Borkmann return ret; 155604326b4SDaniel Borkmann } 156604326b4SDaniel Borkmann 157604326b4SDaniel Borkmann static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock, 158604326b4SDaniel Borkmann struct sk_msg *msg, u32 apply_bytes, int flags) 159604326b4SDaniel Borkmann { 160604326b4SDaniel Borkmann bool apply = apply_bytes; 161604326b4SDaniel Borkmann struct scatterlist *sge; 162604326b4SDaniel Borkmann u32 size, copied = 0; 163604326b4SDaniel Borkmann struct sk_msg *tmp; 164604326b4SDaniel Borkmann int i, ret = 0; 165604326b4SDaniel Borkmann 166604326b4SDaniel Borkmann tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL); 167604326b4SDaniel Borkmann if (unlikely(!tmp)) 168604326b4SDaniel Borkmann return -ENOMEM; 169604326b4SDaniel Borkmann 170604326b4SDaniel Borkmann lock_sock(sk); 171604326b4SDaniel Borkmann tmp->sg.start = msg->sg.start; 172604326b4SDaniel Borkmann i = msg->sg.start; 173604326b4SDaniel Borkmann do { 174604326b4SDaniel Borkmann sge = sk_msg_elem(msg, i); 175604326b4SDaniel Borkmann size = (apply && apply_bytes < sge->length) ? 176604326b4SDaniel Borkmann apply_bytes : sge->length; 177604326b4SDaniel Borkmann if (!sk_wmem_schedule(sk, size)) { 178604326b4SDaniel Borkmann if (!copied) 179604326b4SDaniel Borkmann ret = -ENOMEM; 180604326b4SDaniel Borkmann break; 181604326b4SDaniel Borkmann } 182604326b4SDaniel Borkmann 183604326b4SDaniel Borkmann sk_mem_charge(sk, size); 184604326b4SDaniel Borkmann sk_msg_xfer(tmp, msg, i, size); 185604326b4SDaniel Borkmann copied += size; 186604326b4SDaniel Borkmann if (sge->length) 187604326b4SDaniel Borkmann get_page(sk_msg_page(tmp, i)); 188604326b4SDaniel Borkmann sk_msg_iter_var_next(i); 189604326b4SDaniel Borkmann tmp->sg.end = i; 190604326b4SDaniel Borkmann if (apply) { 191604326b4SDaniel Borkmann apply_bytes -= size; 192604326b4SDaniel Borkmann if (!apply_bytes) 193604326b4SDaniel Borkmann break; 194604326b4SDaniel Borkmann } 195604326b4SDaniel Borkmann } while (i != msg->sg.end); 196604326b4SDaniel Borkmann 197604326b4SDaniel Borkmann if (!ret) { 198604326b4SDaniel Borkmann msg->sg.start = i; 199604326b4SDaniel Borkmann msg->sg.size -= apply_bytes; 200604326b4SDaniel Borkmann sk_psock_queue_msg(psock, tmp); 201*552de910SJohn Fastabend sk_psock_data_ready(sk, psock); 202604326b4SDaniel Borkmann } else { 203604326b4SDaniel Borkmann sk_msg_free(sk, tmp); 204604326b4SDaniel Borkmann kfree(tmp); 205604326b4SDaniel Borkmann } 206604326b4SDaniel Borkmann 207604326b4SDaniel Borkmann release_sock(sk); 208604326b4SDaniel Borkmann return ret; 209604326b4SDaniel Borkmann } 210604326b4SDaniel Borkmann 211604326b4SDaniel Borkmann static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes, 212604326b4SDaniel Borkmann int flags, bool uncharge) 213604326b4SDaniel Borkmann { 214604326b4SDaniel Borkmann bool apply = apply_bytes; 215604326b4SDaniel Borkmann struct scatterlist *sge; 216604326b4SDaniel Borkmann struct page *page; 217604326b4SDaniel Borkmann int size, ret = 0; 218604326b4SDaniel Borkmann u32 off; 219604326b4SDaniel Borkmann 220604326b4SDaniel Borkmann while (1) { 221604326b4SDaniel Borkmann sge = sk_msg_elem(msg, msg->sg.start); 222604326b4SDaniel Borkmann size = (apply && apply_bytes < sge->length) ? 223604326b4SDaniel Borkmann apply_bytes : sge->length; 224604326b4SDaniel Borkmann off = sge->offset; 225604326b4SDaniel Borkmann page = sg_page(sge); 226604326b4SDaniel Borkmann 227604326b4SDaniel Borkmann tcp_rate_check_app_limited(sk); 228604326b4SDaniel Borkmann retry: 229604326b4SDaniel Borkmann ret = do_tcp_sendpages(sk, page, off, size, flags); 230604326b4SDaniel Borkmann if (ret <= 0) 231604326b4SDaniel Borkmann return ret; 232604326b4SDaniel Borkmann if (apply) 233604326b4SDaniel Borkmann apply_bytes -= ret; 234604326b4SDaniel Borkmann msg->sg.size -= ret; 235604326b4SDaniel Borkmann sge->offset += ret; 236604326b4SDaniel Borkmann sge->length -= ret; 237604326b4SDaniel Borkmann if (uncharge) 238604326b4SDaniel Borkmann sk_mem_uncharge(sk, ret); 239604326b4SDaniel Borkmann if (ret != size) { 240604326b4SDaniel Borkmann size -= ret; 241604326b4SDaniel Borkmann off += ret; 242604326b4SDaniel Borkmann goto retry; 243604326b4SDaniel Borkmann } 244604326b4SDaniel Borkmann if (!sge->length) { 245604326b4SDaniel Borkmann put_page(page); 246604326b4SDaniel Borkmann sk_msg_iter_next(msg, start); 247604326b4SDaniel Borkmann sg_init_table(sge, 1); 248604326b4SDaniel Borkmann if (msg->sg.start == msg->sg.end) 249604326b4SDaniel Borkmann break; 250604326b4SDaniel Borkmann } 251604326b4SDaniel Borkmann if (apply && !apply_bytes) 252604326b4SDaniel Borkmann break; 253604326b4SDaniel Borkmann } 254604326b4SDaniel Borkmann 255604326b4SDaniel Borkmann return 0; 256604326b4SDaniel Borkmann } 257604326b4SDaniel Borkmann 258604326b4SDaniel Borkmann static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg, 259604326b4SDaniel Borkmann u32 apply_bytes, int flags, bool uncharge) 260604326b4SDaniel Borkmann { 261604326b4SDaniel Borkmann int ret; 262604326b4SDaniel Borkmann 263604326b4SDaniel Borkmann lock_sock(sk); 264604326b4SDaniel Borkmann ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge); 265604326b4SDaniel Borkmann release_sock(sk); 266604326b4SDaniel Borkmann return ret; 267604326b4SDaniel Borkmann } 268604326b4SDaniel Borkmann 269604326b4SDaniel Borkmann int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg, 270604326b4SDaniel Borkmann u32 bytes, int flags) 271604326b4SDaniel Borkmann { 272604326b4SDaniel Borkmann bool ingress = sk_msg_to_ingress(msg); 273604326b4SDaniel Borkmann struct sk_psock *psock = sk_psock_get(sk); 274604326b4SDaniel Borkmann int ret; 275604326b4SDaniel Borkmann 276604326b4SDaniel Borkmann if (unlikely(!psock)) { 277604326b4SDaniel Borkmann sk_msg_free(sk, msg); 278604326b4SDaniel Borkmann return 0; 279604326b4SDaniel Borkmann } 280604326b4SDaniel Borkmann ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) : 281604326b4SDaniel Borkmann tcp_bpf_push_locked(sk, msg, bytes, flags, false); 282604326b4SDaniel Borkmann sk_psock_put(sk, psock); 283604326b4SDaniel Borkmann return ret; 284604326b4SDaniel Borkmann } 285604326b4SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir); 286604326b4SDaniel Borkmann 287604326b4SDaniel Borkmann static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock, 288604326b4SDaniel Borkmann struct sk_msg *msg, int *copied, int flags) 289604326b4SDaniel Borkmann { 290604326b4SDaniel Borkmann bool cork = false, enospc = msg->sg.start == msg->sg.end; 291604326b4SDaniel Borkmann struct sock *sk_redir; 2927246d8edSJohn Fastabend u32 tosend, delta = 0; 293604326b4SDaniel Borkmann int ret; 294604326b4SDaniel Borkmann 295604326b4SDaniel Borkmann more_data: 2967246d8edSJohn Fastabend if (psock->eval == __SK_NONE) { 2977246d8edSJohn Fastabend /* Track delta in msg size to add/subtract it on SK_DROP from 2987246d8edSJohn Fastabend * returned to user copied size. This ensures user doesn't 2997246d8edSJohn Fastabend * get a positive return code with msg_cut_data and SK_DROP 3007246d8edSJohn Fastabend * verdict. 3017246d8edSJohn Fastabend */ 3027246d8edSJohn Fastabend delta = msg->sg.size; 303604326b4SDaniel Borkmann psock->eval = sk_psock_msg_verdict(sk, psock, msg); 3047246d8edSJohn Fastabend if (msg->sg.size < delta) 3057246d8edSJohn Fastabend delta -= msg->sg.size; 3067246d8edSJohn Fastabend else 3077246d8edSJohn Fastabend delta = 0; 3087246d8edSJohn Fastabend } 309604326b4SDaniel Borkmann 310604326b4SDaniel Borkmann if (msg->cork_bytes && 311604326b4SDaniel Borkmann msg->cork_bytes > msg->sg.size && !enospc) { 312604326b4SDaniel Borkmann psock->cork_bytes = msg->cork_bytes - msg->sg.size; 313604326b4SDaniel Borkmann if (!psock->cork) { 314604326b4SDaniel Borkmann psock->cork = kzalloc(sizeof(*psock->cork), 315604326b4SDaniel Borkmann GFP_ATOMIC | __GFP_NOWARN); 316604326b4SDaniel Borkmann if (!psock->cork) 317604326b4SDaniel Borkmann return -ENOMEM; 318604326b4SDaniel Borkmann } 319604326b4SDaniel Borkmann memcpy(psock->cork, msg, sizeof(*msg)); 320604326b4SDaniel Borkmann return 0; 321604326b4SDaniel Borkmann } 322604326b4SDaniel Borkmann 323604326b4SDaniel Borkmann tosend = msg->sg.size; 324604326b4SDaniel Borkmann if (psock->apply_bytes && psock->apply_bytes < tosend) 325604326b4SDaniel Borkmann tosend = psock->apply_bytes; 326604326b4SDaniel Borkmann 327604326b4SDaniel Borkmann switch (psock->eval) { 328604326b4SDaniel Borkmann case __SK_PASS: 329604326b4SDaniel Borkmann ret = tcp_bpf_push(sk, msg, tosend, flags, true); 330604326b4SDaniel Borkmann if (unlikely(ret)) { 331604326b4SDaniel Borkmann *copied -= sk_msg_free(sk, msg); 332604326b4SDaniel Borkmann break; 333604326b4SDaniel Borkmann } 334604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 335604326b4SDaniel Borkmann break; 336604326b4SDaniel Borkmann case __SK_REDIRECT: 337604326b4SDaniel Borkmann sk_redir = psock->sk_redir; 338604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 339604326b4SDaniel Borkmann if (psock->cork) { 340604326b4SDaniel Borkmann cork = true; 341604326b4SDaniel Borkmann psock->cork = NULL; 342604326b4SDaniel Borkmann } 343604326b4SDaniel Borkmann sk_msg_return(sk, msg, tosend); 344604326b4SDaniel Borkmann release_sock(sk); 345604326b4SDaniel Borkmann ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags); 346604326b4SDaniel Borkmann lock_sock(sk); 347604326b4SDaniel Borkmann if (unlikely(ret < 0)) { 348604326b4SDaniel Borkmann int free = sk_msg_free_nocharge(sk, msg); 349604326b4SDaniel Borkmann 350604326b4SDaniel Borkmann if (!cork) 351604326b4SDaniel Borkmann *copied -= free; 352604326b4SDaniel Borkmann } 353604326b4SDaniel Borkmann if (cork) { 354604326b4SDaniel Borkmann sk_msg_free(sk, msg); 355604326b4SDaniel Borkmann kfree(msg); 356604326b4SDaniel Borkmann msg = NULL; 357604326b4SDaniel Borkmann ret = 0; 358604326b4SDaniel Borkmann } 359604326b4SDaniel Borkmann break; 360604326b4SDaniel Borkmann case __SK_DROP: 361604326b4SDaniel Borkmann default: 362604326b4SDaniel Borkmann sk_msg_free_partial(sk, msg, tosend); 363604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 3647246d8edSJohn Fastabend *copied -= (tosend + delta); 365604326b4SDaniel Borkmann return -EACCES; 366604326b4SDaniel Borkmann } 367604326b4SDaniel Borkmann 368604326b4SDaniel Borkmann if (likely(!ret)) { 369604326b4SDaniel Borkmann if (!psock->apply_bytes) { 370604326b4SDaniel Borkmann psock->eval = __SK_NONE; 371604326b4SDaniel Borkmann if (psock->sk_redir) { 372604326b4SDaniel Borkmann sock_put(psock->sk_redir); 373604326b4SDaniel Borkmann psock->sk_redir = NULL; 374604326b4SDaniel Borkmann } 375604326b4SDaniel Borkmann } 376604326b4SDaniel Borkmann if (msg && 377604326b4SDaniel Borkmann msg->sg.data[msg->sg.start].page_link && 378604326b4SDaniel Borkmann msg->sg.data[msg->sg.start].length) 379604326b4SDaniel Borkmann goto more_data; 380604326b4SDaniel Borkmann } 381604326b4SDaniel Borkmann return ret; 382604326b4SDaniel Borkmann } 383604326b4SDaniel Borkmann 384604326b4SDaniel Borkmann static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 385604326b4SDaniel Borkmann { 386604326b4SDaniel Borkmann struct sk_msg tmp, *msg_tx = NULL; 387604326b4SDaniel Borkmann int flags = msg->msg_flags | MSG_NO_SHARED_FRAGS; 388604326b4SDaniel Borkmann int copied = 0, err = 0; 389604326b4SDaniel Borkmann struct sk_psock *psock; 390604326b4SDaniel Borkmann long timeo; 391604326b4SDaniel Borkmann 392604326b4SDaniel Borkmann psock = sk_psock_get(sk); 393604326b4SDaniel Borkmann if (unlikely(!psock)) 394604326b4SDaniel Borkmann return tcp_sendmsg(sk, msg, size); 395604326b4SDaniel Borkmann 396604326b4SDaniel Borkmann lock_sock(sk); 397604326b4SDaniel Borkmann timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 398604326b4SDaniel Borkmann while (msg_data_left(msg)) { 399604326b4SDaniel Borkmann bool enospc = false; 400604326b4SDaniel Borkmann u32 copy, osize; 401604326b4SDaniel Borkmann 402604326b4SDaniel Borkmann if (sk->sk_err) { 403604326b4SDaniel Borkmann err = -sk->sk_err; 404604326b4SDaniel Borkmann goto out_err; 405604326b4SDaniel Borkmann } 406604326b4SDaniel Borkmann 407604326b4SDaniel Borkmann copy = msg_data_left(msg); 408604326b4SDaniel Borkmann if (!sk_stream_memory_free(sk)) 409604326b4SDaniel Borkmann goto wait_for_sndbuf; 410604326b4SDaniel Borkmann if (psock->cork) { 411604326b4SDaniel Borkmann msg_tx = psock->cork; 412604326b4SDaniel Borkmann } else { 413604326b4SDaniel Borkmann msg_tx = &tmp; 414604326b4SDaniel Borkmann sk_msg_init(msg_tx); 415604326b4SDaniel Borkmann } 416604326b4SDaniel Borkmann 417604326b4SDaniel Borkmann osize = msg_tx->sg.size; 418604326b4SDaniel Borkmann err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1); 419604326b4SDaniel Borkmann if (err) { 420604326b4SDaniel Borkmann if (err != -ENOSPC) 421604326b4SDaniel Borkmann goto wait_for_memory; 422604326b4SDaniel Borkmann enospc = true; 423604326b4SDaniel Borkmann copy = msg_tx->sg.size - osize; 424604326b4SDaniel Borkmann } 425604326b4SDaniel Borkmann 426604326b4SDaniel Borkmann err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx, 427604326b4SDaniel Borkmann copy); 428604326b4SDaniel Borkmann if (err < 0) { 429604326b4SDaniel Borkmann sk_msg_trim(sk, msg_tx, osize); 430604326b4SDaniel Borkmann goto out_err; 431604326b4SDaniel Borkmann } 432604326b4SDaniel Borkmann 433604326b4SDaniel Borkmann copied += copy; 434604326b4SDaniel Borkmann if (psock->cork_bytes) { 435604326b4SDaniel Borkmann if (size > psock->cork_bytes) 436604326b4SDaniel Borkmann psock->cork_bytes = 0; 437604326b4SDaniel Borkmann else 438604326b4SDaniel Borkmann psock->cork_bytes -= size; 439604326b4SDaniel Borkmann if (psock->cork_bytes && !enospc) 440604326b4SDaniel Borkmann goto out_err; 441604326b4SDaniel Borkmann /* All cork bytes are accounted, rerun the prog. */ 442604326b4SDaniel Borkmann psock->eval = __SK_NONE; 443604326b4SDaniel Borkmann psock->cork_bytes = 0; 444604326b4SDaniel Borkmann } 445604326b4SDaniel Borkmann 446604326b4SDaniel Borkmann err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags); 447604326b4SDaniel Borkmann if (unlikely(err < 0)) 448604326b4SDaniel Borkmann goto out_err; 449604326b4SDaniel Borkmann continue; 450604326b4SDaniel Borkmann wait_for_sndbuf: 451604326b4SDaniel Borkmann set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 452604326b4SDaniel Borkmann wait_for_memory: 453604326b4SDaniel Borkmann err = sk_stream_wait_memory(sk, &timeo); 454604326b4SDaniel Borkmann if (err) { 455604326b4SDaniel Borkmann if (msg_tx && msg_tx != psock->cork) 456604326b4SDaniel Borkmann sk_msg_free(sk, msg_tx); 457604326b4SDaniel Borkmann goto out_err; 458604326b4SDaniel Borkmann } 459604326b4SDaniel Borkmann } 460604326b4SDaniel Borkmann out_err: 461604326b4SDaniel Borkmann if (err < 0) 462604326b4SDaniel Borkmann err = sk_stream_error(sk, msg->msg_flags, err); 463604326b4SDaniel Borkmann release_sock(sk); 464604326b4SDaniel Borkmann sk_psock_put(sk, psock); 465604326b4SDaniel Borkmann return copied ? copied : err; 466604326b4SDaniel Borkmann } 467604326b4SDaniel Borkmann 468604326b4SDaniel Borkmann static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset, 469604326b4SDaniel Borkmann size_t size, int flags) 470604326b4SDaniel Borkmann { 471604326b4SDaniel Borkmann struct sk_msg tmp, *msg = NULL; 472604326b4SDaniel Borkmann int err = 0, copied = 0; 473604326b4SDaniel Borkmann struct sk_psock *psock; 474604326b4SDaniel Borkmann bool enospc = false; 475604326b4SDaniel Borkmann 476604326b4SDaniel Borkmann psock = sk_psock_get(sk); 477604326b4SDaniel Borkmann if (unlikely(!psock)) 478604326b4SDaniel Borkmann return tcp_sendpage(sk, page, offset, size, flags); 479604326b4SDaniel Borkmann 480604326b4SDaniel Borkmann lock_sock(sk); 481604326b4SDaniel Borkmann if (psock->cork) { 482604326b4SDaniel Borkmann msg = psock->cork; 483604326b4SDaniel Borkmann } else { 484604326b4SDaniel Borkmann msg = &tmp; 485604326b4SDaniel Borkmann sk_msg_init(msg); 486604326b4SDaniel Borkmann } 487604326b4SDaniel Borkmann 488604326b4SDaniel Borkmann /* Catch case where ring is full and sendpage is stalled. */ 489604326b4SDaniel Borkmann if (unlikely(sk_msg_full(msg))) 490604326b4SDaniel Borkmann goto out_err; 491604326b4SDaniel Borkmann 492604326b4SDaniel Borkmann sk_msg_page_add(msg, page, size, offset); 493604326b4SDaniel Borkmann sk_mem_charge(sk, size); 494604326b4SDaniel Borkmann copied = size; 495604326b4SDaniel Borkmann if (sk_msg_full(msg)) 496604326b4SDaniel Borkmann enospc = true; 497604326b4SDaniel Borkmann if (psock->cork_bytes) { 498604326b4SDaniel Borkmann if (size > psock->cork_bytes) 499604326b4SDaniel Borkmann psock->cork_bytes = 0; 500604326b4SDaniel Borkmann else 501604326b4SDaniel Borkmann psock->cork_bytes -= size; 502604326b4SDaniel Borkmann if (psock->cork_bytes && !enospc) 503604326b4SDaniel Borkmann goto out_err; 504604326b4SDaniel Borkmann /* All cork bytes are accounted, rerun the prog. */ 505604326b4SDaniel Borkmann psock->eval = __SK_NONE; 506604326b4SDaniel Borkmann psock->cork_bytes = 0; 507604326b4SDaniel Borkmann } 508604326b4SDaniel Borkmann 509604326b4SDaniel Borkmann err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags); 510604326b4SDaniel Borkmann out_err: 511604326b4SDaniel Borkmann release_sock(sk); 512604326b4SDaniel Borkmann sk_psock_put(sk, psock); 513604326b4SDaniel Borkmann return copied ? copied : err; 514604326b4SDaniel Borkmann } 515604326b4SDaniel Borkmann 516604326b4SDaniel Borkmann static void tcp_bpf_remove(struct sock *sk, struct sk_psock *psock) 517604326b4SDaniel Borkmann { 518604326b4SDaniel Borkmann struct sk_psock_link *link; 519604326b4SDaniel Borkmann 520604326b4SDaniel Borkmann sk_psock_cork_free(psock); 521604326b4SDaniel Borkmann __sk_psock_purge_ingress_msg(psock); 522604326b4SDaniel Borkmann while ((link = sk_psock_link_pop(psock))) { 523604326b4SDaniel Borkmann sk_psock_unlink(sk, link); 524604326b4SDaniel Borkmann sk_psock_free_link(link); 525604326b4SDaniel Borkmann } 526604326b4SDaniel Borkmann } 527604326b4SDaniel Borkmann 528604326b4SDaniel Borkmann static void tcp_bpf_unhash(struct sock *sk) 529604326b4SDaniel Borkmann { 530604326b4SDaniel Borkmann void (*saved_unhash)(struct sock *sk); 531604326b4SDaniel Borkmann struct sk_psock *psock; 532604326b4SDaniel Borkmann 533604326b4SDaniel Borkmann rcu_read_lock(); 534604326b4SDaniel Borkmann psock = sk_psock(sk); 535604326b4SDaniel Borkmann if (unlikely(!psock)) { 536604326b4SDaniel Borkmann rcu_read_unlock(); 537604326b4SDaniel Borkmann if (sk->sk_prot->unhash) 538604326b4SDaniel Borkmann sk->sk_prot->unhash(sk); 539604326b4SDaniel Borkmann return; 540604326b4SDaniel Borkmann } 541604326b4SDaniel Borkmann 542604326b4SDaniel Borkmann saved_unhash = psock->saved_unhash; 543604326b4SDaniel Borkmann tcp_bpf_remove(sk, psock); 544604326b4SDaniel Borkmann rcu_read_unlock(); 545604326b4SDaniel Borkmann saved_unhash(sk); 546604326b4SDaniel Borkmann } 547604326b4SDaniel Borkmann 548604326b4SDaniel Borkmann static void tcp_bpf_close(struct sock *sk, long timeout) 549604326b4SDaniel Borkmann { 550604326b4SDaniel Borkmann void (*saved_close)(struct sock *sk, long timeout); 551604326b4SDaniel Borkmann struct sk_psock *psock; 552604326b4SDaniel Borkmann 553604326b4SDaniel Borkmann lock_sock(sk); 554604326b4SDaniel Borkmann rcu_read_lock(); 555604326b4SDaniel Borkmann psock = sk_psock(sk); 556604326b4SDaniel Borkmann if (unlikely(!psock)) { 557604326b4SDaniel Borkmann rcu_read_unlock(); 558604326b4SDaniel Borkmann release_sock(sk); 559604326b4SDaniel Borkmann return sk->sk_prot->close(sk, timeout); 560604326b4SDaniel Borkmann } 561604326b4SDaniel Borkmann 562604326b4SDaniel Borkmann saved_close = psock->saved_close; 563604326b4SDaniel Borkmann tcp_bpf_remove(sk, psock); 564604326b4SDaniel Borkmann rcu_read_unlock(); 565604326b4SDaniel Borkmann release_sock(sk); 566604326b4SDaniel Borkmann saved_close(sk, timeout); 567604326b4SDaniel Borkmann } 568604326b4SDaniel Borkmann 569604326b4SDaniel Borkmann enum { 570604326b4SDaniel Borkmann TCP_BPF_IPV4, 571604326b4SDaniel Borkmann TCP_BPF_IPV6, 572604326b4SDaniel Borkmann TCP_BPF_NUM_PROTS, 573604326b4SDaniel Borkmann }; 574604326b4SDaniel Borkmann 575604326b4SDaniel Borkmann enum { 576604326b4SDaniel Borkmann TCP_BPF_BASE, 577604326b4SDaniel Borkmann TCP_BPF_TX, 578604326b4SDaniel Borkmann TCP_BPF_NUM_CFGS, 579604326b4SDaniel Borkmann }; 580604326b4SDaniel Borkmann 581604326b4SDaniel Borkmann static struct proto *tcpv6_prot_saved __read_mostly; 582604326b4SDaniel Borkmann static DEFINE_SPINLOCK(tcpv6_prot_lock); 583604326b4SDaniel Borkmann static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS]; 584604326b4SDaniel Borkmann 585604326b4SDaniel Borkmann static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS], 586604326b4SDaniel Borkmann struct proto *base) 587604326b4SDaniel Borkmann { 588604326b4SDaniel Borkmann prot[TCP_BPF_BASE] = *base; 589604326b4SDaniel Borkmann prot[TCP_BPF_BASE].unhash = tcp_bpf_unhash; 590604326b4SDaniel Borkmann prot[TCP_BPF_BASE].close = tcp_bpf_close; 591604326b4SDaniel Borkmann prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg; 592604326b4SDaniel Borkmann prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read; 593604326b4SDaniel Borkmann 594604326b4SDaniel Borkmann prot[TCP_BPF_TX] = prot[TCP_BPF_BASE]; 595604326b4SDaniel Borkmann prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg; 596604326b4SDaniel Borkmann prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage; 597604326b4SDaniel Borkmann } 598604326b4SDaniel Borkmann 599604326b4SDaniel Borkmann static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops) 600604326b4SDaniel Borkmann { 601604326b4SDaniel Borkmann if (sk->sk_family == AF_INET6 && 602604326b4SDaniel Borkmann unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) { 603604326b4SDaniel Borkmann spin_lock_bh(&tcpv6_prot_lock); 604604326b4SDaniel Borkmann if (likely(ops != tcpv6_prot_saved)) { 605604326b4SDaniel Borkmann tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops); 606604326b4SDaniel Borkmann smp_store_release(&tcpv6_prot_saved, ops); 607604326b4SDaniel Borkmann } 608604326b4SDaniel Borkmann spin_unlock_bh(&tcpv6_prot_lock); 609604326b4SDaniel Borkmann } 610604326b4SDaniel Borkmann } 611604326b4SDaniel Borkmann 612604326b4SDaniel Borkmann static int __init tcp_bpf_v4_build_proto(void) 613604326b4SDaniel Borkmann { 614604326b4SDaniel Borkmann tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot); 615604326b4SDaniel Borkmann return 0; 616604326b4SDaniel Borkmann } 617604326b4SDaniel Borkmann core_initcall(tcp_bpf_v4_build_proto); 618604326b4SDaniel Borkmann 619604326b4SDaniel Borkmann static void tcp_bpf_update_sk_prot(struct sock *sk, struct sk_psock *psock) 620604326b4SDaniel Borkmann { 621604326b4SDaniel Borkmann int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 622604326b4SDaniel Borkmann int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 623604326b4SDaniel Borkmann 624604326b4SDaniel Borkmann sk_psock_update_proto(sk, psock, &tcp_bpf_prots[family][config]); 625604326b4SDaniel Borkmann } 626604326b4SDaniel Borkmann 627604326b4SDaniel Borkmann static void tcp_bpf_reinit_sk_prot(struct sock *sk, struct sk_psock *psock) 628604326b4SDaniel Borkmann { 629604326b4SDaniel Borkmann int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 630604326b4SDaniel Borkmann int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 631604326b4SDaniel Borkmann 632604326b4SDaniel Borkmann /* Reinit occurs when program types change e.g. TCP_BPF_TX is removed 633604326b4SDaniel Borkmann * or added requiring sk_prot hook updates. We keep original saved 634604326b4SDaniel Borkmann * hooks in this case. 635604326b4SDaniel Borkmann */ 636604326b4SDaniel Borkmann sk->sk_prot = &tcp_bpf_prots[family][config]; 637604326b4SDaniel Borkmann } 638604326b4SDaniel Borkmann 639604326b4SDaniel Borkmann static int tcp_bpf_assert_proto_ops(struct proto *ops) 640604326b4SDaniel Borkmann { 641604326b4SDaniel Borkmann /* In order to avoid retpoline, we make assumptions when we call 642604326b4SDaniel Borkmann * into ops if e.g. a psock is not present. Make sure they are 643604326b4SDaniel Borkmann * indeed valid assumptions. 644604326b4SDaniel Borkmann */ 645604326b4SDaniel Borkmann return ops->recvmsg == tcp_recvmsg && 646604326b4SDaniel Borkmann ops->sendmsg == tcp_sendmsg && 647604326b4SDaniel Borkmann ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP; 648604326b4SDaniel Borkmann } 649604326b4SDaniel Borkmann 650604326b4SDaniel Borkmann void tcp_bpf_reinit(struct sock *sk) 651604326b4SDaniel Borkmann { 652604326b4SDaniel Borkmann struct sk_psock *psock; 653604326b4SDaniel Borkmann 654604326b4SDaniel Borkmann sock_owned_by_me(sk); 655604326b4SDaniel Borkmann 656604326b4SDaniel Borkmann rcu_read_lock(); 657604326b4SDaniel Borkmann psock = sk_psock(sk); 658604326b4SDaniel Borkmann tcp_bpf_reinit_sk_prot(sk, psock); 659604326b4SDaniel Borkmann rcu_read_unlock(); 660604326b4SDaniel Borkmann } 661604326b4SDaniel Borkmann 662604326b4SDaniel Borkmann int tcp_bpf_init(struct sock *sk) 663604326b4SDaniel Borkmann { 664604326b4SDaniel Borkmann struct proto *ops = READ_ONCE(sk->sk_prot); 665604326b4SDaniel Borkmann struct sk_psock *psock; 666604326b4SDaniel Borkmann 667604326b4SDaniel Borkmann sock_owned_by_me(sk); 668604326b4SDaniel Borkmann 669604326b4SDaniel Borkmann rcu_read_lock(); 670604326b4SDaniel Borkmann psock = sk_psock(sk); 671604326b4SDaniel Borkmann if (unlikely(!psock || psock->sk_proto || 672604326b4SDaniel Borkmann tcp_bpf_assert_proto_ops(ops))) { 673604326b4SDaniel Borkmann rcu_read_unlock(); 674604326b4SDaniel Borkmann return -EINVAL; 675604326b4SDaniel Borkmann } 676604326b4SDaniel Borkmann tcp_bpf_check_v6_needs_rebuild(sk, ops); 677604326b4SDaniel Borkmann tcp_bpf_update_sk_prot(sk, psock); 678604326b4SDaniel Borkmann rcu_read_unlock(); 679604326b4SDaniel Borkmann return 0; 680604326b4SDaniel Borkmann } 681