1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Shared Memory Communications over RDMA (SMC-R) and RoCE 4 * 5 * Manage RMBE 6 * copy new RMBE data into user space 7 * 8 * Copyright IBM Corp. 2016 9 * 10 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com> 11 */ 12 13 #include <linux/net.h> 14 #include <linux/rcupdate.h> 15 #include <linux/sched/signal.h> 16 17 #include <net/sock.h> 18 19 #include "smc.h" 20 #include "smc_core.h" 21 #include "smc_cdc.h" 22 #include "smc_tx.h" /* smc_tx_consumer_update() */ 23 #include "smc_rx.h" 24 25 /* callback implementation to wakeup consumers blocked with smc_rx_wait(). 26 * indirectly called by smc_cdc_msg_recv_action(). 27 */ 28 static void smc_rx_wake_up(struct sock *sk) 29 { 30 struct socket_wq *wq; 31 32 /* derived from sock_def_readable() */ 33 /* called already in smc_listen_work() */ 34 rcu_read_lock(); 35 wq = rcu_dereference(sk->sk_wq); 36 if (skwq_has_sleeper(wq)) 37 wake_up_interruptible_sync_poll(&wq->wait, EPOLLIN | EPOLLPRI | 38 EPOLLRDNORM | EPOLLRDBAND); 39 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 40 if ((sk->sk_shutdown == SHUTDOWN_MASK) || 41 (sk->sk_state == SMC_CLOSED)) 42 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 43 rcu_read_unlock(); 44 } 45 46 /* Update consumer cursor 47 * @conn connection to update 48 * @cons consumer cursor 49 * @len number of Bytes consumed 50 * Returns: 51 * 1 if we should end our receive, 0 otherwise 52 */ 53 static int smc_rx_update_consumer(struct smc_sock *smc, 54 union smc_host_cursor cons, size_t len) 55 { 56 struct smc_connection *conn = &smc->conn; 57 struct sock *sk = &smc->sk; 58 bool force = false; 59 int diff, rc = 0; 60 61 smc_curs_add(conn->rmb_desc->len, &cons, len); 62 63 /* did we process urgent data? */ 64 if (conn->urg_state == SMC_URG_VALID || conn->urg_rx_skip_pend) { 65 diff = smc_curs_comp(conn->rmb_desc->len, &cons, 66 &conn->urg_curs); 67 if (sock_flag(sk, SOCK_URGINLINE)) { 68 if (diff == 0) { 69 force = true; 70 rc = 1; 71 conn->urg_state = SMC_URG_READ; 72 } 73 } else { 74 if (diff == 1) { 75 /* skip urgent byte */ 76 force = true; 77 smc_curs_add(conn->rmb_desc->len, &cons, 1); 78 conn->urg_rx_skip_pend = false; 79 } else if (diff < -1) 80 /* we read past urgent byte */ 81 conn->urg_state = SMC_URG_READ; 82 } 83 } 84 85 smc_curs_copy(&conn->local_tx_ctrl.cons, &cons, conn); 86 87 /* send consumer cursor update if required */ 88 /* similar to advertising new TCP rcv_wnd if required */ 89 smc_tx_consumer_update(conn, force); 90 91 return rc; 92 } 93 94 static void smc_rx_update_cons(struct smc_sock *smc, size_t len) 95 { 96 struct smc_connection *conn = &smc->conn; 97 union smc_host_cursor cons; 98 99 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn); 100 smc_rx_update_consumer(smc, cons, len); 101 } 102 103 struct smc_spd_priv { 104 struct smc_sock *smc; 105 size_t len; 106 }; 107 108 static void smc_rx_pipe_buf_release(struct pipe_inode_info *pipe, 109 struct pipe_buffer *buf) 110 { 111 struct smc_spd_priv *priv = (struct smc_spd_priv *)buf->private; 112 struct smc_sock *smc = priv->smc; 113 struct smc_connection *conn; 114 struct sock *sk = &smc->sk; 115 116 if (sk->sk_state == SMC_CLOSED || 117 sk->sk_state == SMC_PEERFINCLOSEWAIT || 118 sk->sk_state == SMC_APPFINCLOSEWAIT) 119 goto out; 120 conn = &smc->conn; 121 lock_sock(sk); 122 smc_rx_update_cons(smc, priv->len); 123 release_sock(sk); 124 if (atomic_sub_and_test(priv->len, &conn->splice_pending)) 125 smc_rx_wake_up(sk); 126 out: 127 kfree(priv); 128 put_page(buf->page); 129 sock_put(sk); 130 } 131 132 static int smc_rx_pipe_buf_nosteal(struct pipe_inode_info *pipe, 133 struct pipe_buffer *buf) 134 { 135 return 1; 136 } 137 138 static const struct pipe_buf_operations smc_pipe_ops = { 139 .can_merge = 0, 140 .confirm = generic_pipe_buf_confirm, 141 .release = smc_rx_pipe_buf_release, 142 .steal = smc_rx_pipe_buf_nosteal, 143 .get = generic_pipe_buf_get 144 }; 145 146 static void smc_rx_spd_release(struct splice_pipe_desc *spd, 147 unsigned int i) 148 { 149 put_page(spd->pages[i]); 150 } 151 152 static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len, 153 struct smc_sock *smc) 154 { 155 struct splice_pipe_desc spd; 156 struct partial_page partial; 157 struct smc_spd_priv *priv; 158 int bytes; 159 160 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 161 if (!priv) 162 return -ENOMEM; 163 priv->len = len; 164 priv->smc = smc; 165 partial.offset = src - (char *)smc->conn.rmb_desc->cpu_addr; 166 partial.len = len; 167 partial.private = (unsigned long)priv; 168 169 spd.nr_pages_max = 1; 170 spd.nr_pages = 1; 171 spd.pages = &smc->conn.rmb_desc->pages; 172 spd.partial = &partial; 173 spd.ops = &smc_pipe_ops; 174 spd.spd_release = smc_rx_spd_release; 175 176 bytes = splice_to_pipe(pipe, &spd); 177 if (bytes > 0) { 178 sock_hold(&smc->sk); 179 get_page(smc->conn.rmb_desc->pages); 180 atomic_add(bytes, &smc->conn.splice_pending); 181 } 182 183 return bytes; 184 } 185 186 static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn) 187 { 188 return atomic_read(&conn->bytes_to_rcv) && 189 !atomic_read(&conn->splice_pending); 190 } 191 192 /* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted 193 * @smc smc socket 194 * @timeo pointer to max seconds to wait, pointer to value 0 for no timeout 195 * @fcrit add'l criterion to evaluate as function pointer 196 * Returns: 197 * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown. 198 * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted). 199 */ 200 int smc_rx_wait(struct smc_sock *smc, long *timeo, 201 int (*fcrit)(struct smc_connection *conn)) 202 { 203 DEFINE_WAIT_FUNC(wait, woken_wake_function); 204 struct smc_connection *conn = &smc->conn; 205 struct sock *sk = &smc->sk; 206 int rc; 207 208 if (fcrit(conn)) 209 return 1; 210 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 211 add_wait_queue(sk_sleep(sk), &wait); 212 rc = sk_wait_event(sk, timeo, 213 sk->sk_err || 214 sk->sk_shutdown & RCV_SHUTDOWN || 215 fcrit(conn) || 216 smc_cdc_rxed_any_close_or_senddone(conn), 217 &wait); 218 remove_wait_queue(sk_sleep(sk), &wait); 219 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 220 return rc; 221 } 222 223 static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len, 224 int flags) 225 { 226 struct smc_connection *conn = &smc->conn; 227 union smc_host_cursor cons; 228 struct sock *sk = &smc->sk; 229 int rc = 0; 230 231 if (sock_flag(sk, SOCK_URGINLINE) || 232 !(conn->urg_state == SMC_URG_VALID) || 233 conn->urg_state == SMC_URG_READ) 234 return -EINVAL; 235 236 if (conn->urg_state == SMC_URG_VALID) { 237 if (!(flags & MSG_PEEK)) 238 smc->conn.urg_state = SMC_URG_READ; 239 msg->msg_flags |= MSG_OOB; 240 if (len > 0) { 241 if (!(flags & MSG_TRUNC)) 242 rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1); 243 len = 1; 244 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn); 245 if (smc_curs_diff(conn->rmb_desc->len, &cons, 246 &conn->urg_curs) > 1) 247 conn->urg_rx_skip_pend = true; 248 /* Urgent Byte was already accounted for, but trigger 249 * skipping the urgent byte in non-inline case 250 */ 251 if (!(flags & MSG_PEEK)) 252 smc_rx_update_consumer(smc, cons, 0); 253 } else { 254 msg->msg_flags |= MSG_TRUNC; 255 } 256 257 return rc ? -EFAULT : len; 258 } 259 260 if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN) 261 return 0; 262 263 return -EAGAIN; 264 } 265 266 /* smc_rx_recvmsg - receive data from RMBE 267 * @msg: copy data to receive buffer 268 * @pipe: copy data to pipe if set - indicates splice() call 269 * 270 * rcvbuf consumer: main API called by socket layer. 271 * Called under sk lock. 272 */ 273 int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg, 274 struct pipe_inode_info *pipe, size_t len, int flags) 275 { 276 size_t copylen, read_done = 0, read_remaining = len; 277 size_t chunk_len, chunk_off, chunk_len_sum; 278 struct smc_connection *conn = &smc->conn; 279 int (*func)(struct smc_connection *conn); 280 union smc_host_cursor cons; 281 int readable, chunk; 282 char *rcvbuf_base; 283 struct sock *sk; 284 int splbytes; 285 long timeo; 286 int target; /* Read at least these many bytes */ 287 int rc; 288 289 if (unlikely(flags & MSG_ERRQUEUE)) 290 return -EINVAL; /* future work for sk.sk_family == AF_SMC */ 291 292 sk = &smc->sk; 293 if (sk->sk_state == SMC_LISTEN) 294 return -ENOTCONN; 295 if (flags & MSG_OOB) 296 return smc_rx_recv_urg(smc, msg, len, flags); 297 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 298 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 299 300 /* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */ 301 rcvbuf_base = conn->rx_off + conn->rmb_desc->cpu_addr; 302 303 do { /* while (read_remaining) */ 304 if (read_done >= target || (pipe && read_done)) 305 break; 306 307 if (atomic_read(&conn->bytes_to_rcv)) 308 goto copy; 309 else if (conn->urg_state == SMC_URG_VALID) 310 /* we received a single urgent Byte - skip */ 311 smc_rx_update_cons(smc, 0); 312 313 if (sk->sk_shutdown & RCV_SHUTDOWN || 314 smc_cdc_rxed_any_close_or_senddone(conn) || 315 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort) 316 break; 317 318 if (read_done) { 319 if (sk->sk_err || 320 sk->sk_state == SMC_CLOSED || 321 !timeo || 322 signal_pending(current)) 323 break; 324 } else { 325 if (sk->sk_err) { 326 read_done = sock_error(sk); 327 break; 328 } 329 if (sk->sk_state == SMC_CLOSED) { 330 if (!sock_flag(sk, SOCK_DONE)) { 331 /* This occurs when user tries to read 332 * from never connected socket. 333 */ 334 read_done = -ENOTCONN; 335 break; 336 } 337 break; 338 } 339 if (signal_pending(current)) { 340 read_done = sock_intr_errno(timeo); 341 break; 342 } 343 if (!timeo) 344 return -EAGAIN; 345 } 346 347 if (!smc_rx_data_available(conn)) { 348 smc_rx_wait(smc, &timeo, smc_rx_data_available); 349 continue; 350 } 351 352 copy: 353 /* initialize variables for 1st iteration of subsequent loop */ 354 /* could be just 1 byte, even after waiting on data above */ 355 readable = atomic_read(&conn->bytes_to_rcv); 356 splbytes = atomic_read(&conn->splice_pending); 357 if (!readable || (msg && splbytes)) { 358 if (splbytes) 359 func = smc_rx_data_available_and_no_splice_pend; 360 else 361 func = smc_rx_data_available; 362 smc_rx_wait(smc, &timeo, func); 363 continue; 364 } 365 366 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn); 367 /* subsequent splice() calls pick up where previous left */ 368 if (splbytes) 369 smc_curs_add(conn->rmb_desc->len, &cons, splbytes); 370 if (conn->urg_state == SMC_URG_VALID && 371 sock_flag(&smc->sk, SOCK_URGINLINE) && 372 readable > 1) 373 readable--; /* always stop at urgent Byte */ 374 /* not more than what user space asked for */ 375 copylen = min_t(size_t, read_remaining, readable); 376 /* determine chunks where to read from rcvbuf */ 377 /* either unwrapped case, or 1st chunk of wrapped case */ 378 chunk_len = min_t(size_t, copylen, conn->rmb_desc->len - 379 cons.count); 380 chunk_len_sum = chunk_len; 381 chunk_off = cons.count; 382 smc_rmb_sync_sg_for_cpu(conn); 383 for (chunk = 0; chunk < 2; chunk++) { 384 if (!(flags & MSG_TRUNC)) { 385 if (msg) { 386 rc = memcpy_to_msg(msg, rcvbuf_base + 387 chunk_off, 388 chunk_len); 389 } else { 390 rc = smc_rx_splice(pipe, rcvbuf_base + 391 chunk_off, chunk_len, 392 smc); 393 } 394 if (rc < 0) { 395 if (!read_done) 396 read_done = -EFAULT; 397 smc_rmb_sync_sg_for_device(conn); 398 goto out; 399 } 400 } 401 read_remaining -= chunk_len; 402 read_done += chunk_len; 403 404 if (chunk_len_sum == copylen) 405 break; /* either on 1st or 2nd iteration */ 406 /* prepare next (== 2nd) iteration */ 407 chunk_len = copylen - chunk_len; /* remainder */ 408 chunk_len_sum += chunk_len; 409 chunk_off = 0; /* modulo offset in recv ring buffer */ 410 } 411 smc_rmb_sync_sg_for_device(conn); 412 413 /* update cursors */ 414 if (!(flags & MSG_PEEK)) { 415 /* increased in recv tasklet smc_cdc_msg_rcv() */ 416 smp_mb__before_atomic(); 417 atomic_sub(copylen, &conn->bytes_to_rcv); 418 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 419 smp_mb__after_atomic(); 420 if (msg && smc_rx_update_consumer(smc, cons, copylen)) 421 goto out; 422 } 423 } while (read_remaining); 424 out: 425 return read_done; 426 } 427 428 /* Initialize receive properties on connection establishment. NB: not __init! */ 429 void smc_rx_init(struct smc_sock *smc) 430 { 431 smc->sk.sk_data_ready = smc_rx_wake_up; 432 atomic_set(&smc->conn.splice_pending, 0); 433 smc->conn.urg_state = SMC_URG_READ; 434 } 435