1 /* 2 * Shared Memory Communications over RDMA (SMC-R) and RoCE 3 * 4 * Manage send buffer. 5 * Producer: 6 * Copy user space data into send buffer, if send buffer space available. 7 * Consumer: 8 * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available. 9 * 10 * Copyright IBM Corp. 2016 11 * 12 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com> 13 */ 14 15 #include <linux/net.h> 16 #include <linux/rcupdate.h> 17 #include <linux/workqueue.h> 18 #include <linux/sched/signal.h> 19 20 #include <net/sock.h> 21 22 #include "smc.h" 23 #include "smc_wr.h" 24 #include "smc_cdc.h" 25 #include "smc_tx.h" 26 27 /***************************** sndbuf producer *******************************/ 28 29 /* callback implementation for sk.sk_write_space() 30 * to wakeup sndbuf producers that blocked with smc_tx_wait_memory(). 31 * called under sk_socket lock. 32 */ 33 static void smc_tx_write_space(struct sock *sk) 34 { 35 struct socket *sock = sk->sk_socket; 36 struct smc_sock *smc = smc_sk(sk); 37 struct socket_wq *wq; 38 39 /* similar to sk_stream_write_space */ 40 if (atomic_read(&smc->conn.sndbuf_space) && sock) { 41 clear_bit(SOCK_NOSPACE, &sock->flags); 42 rcu_read_lock(); 43 wq = rcu_dereference(sk->sk_wq); 44 if (skwq_has_sleeper(wq)) 45 wake_up_interruptible_poll(&wq->wait, 46 POLLOUT | POLLWRNORM | 47 POLLWRBAND); 48 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN)) 49 sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT); 50 rcu_read_unlock(); 51 } 52 } 53 54 /* Wakeup sndbuf producers that blocked with smc_tx_wait_memory(). 55 * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space(). 56 */ 57 void smc_tx_sndbuf_nonfull(struct smc_sock *smc) 58 { 59 if (smc->sk.sk_socket && 60 test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags)) 61 smc->sk.sk_write_space(&smc->sk); 62 } 63 64 /* blocks sndbuf producer until at least one byte of free space available */ 65 static int smc_tx_wait_memory(struct smc_sock *smc, int flags) 66 { 67 DEFINE_WAIT_FUNC(wait, woken_wake_function); 68 struct smc_connection *conn = &smc->conn; 69 struct sock *sk = &smc->sk; 70 bool noblock; 71 long timeo; 72 int rc = 0; 73 74 /* similar to sk_stream_wait_memory */ 75 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 76 noblock = timeo ? false : true; 77 add_wait_queue(sk_sleep(sk), &wait); 78 while (1) { 79 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk); 80 if (sk->sk_err || 81 (sk->sk_shutdown & SEND_SHUTDOWN) || 82 conn->local_tx_ctrl.conn_state_flags.peer_done_writing) { 83 rc = -EPIPE; 84 break; 85 } 86 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 87 rc = -ECONNRESET; 88 break; 89 } 90 if (!timeo) { 91 if (noblock) 92 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 93 rc = -EAGAIN; 94 break; 95 } 96 if (signal_pending(current)) { 97 rc = sock_intr_errno(timeo); 98 break; 99 } 100 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 101 if (atomic_read(&conn->sndbuf_space)) 102 break; /* at least 1 byte of free space available */ 103 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 104 sk->sk_write_pending++; 105 sk_wait_event(sk, &timeo, 106 sk->sk_err || 107 (sk->sk_shutdown & SEND_SHUTDOWN) || 108 smc_cdc_rxed_any_close_or_senddone(conn) || 109 atomic_read(&conn->sndbuf_space), 110 &wait); 111 sk->sk_write_pending--; 112 } 113 remove_wait_queue(sk_sleep(sk), &wait); 114 return rc; 115 } 116 117 /* sndbuf producer: main API called by socket layer. 118 * called under sock lock. 119 */ 120 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len) 121 { 122 size_t copylen, send_done = 0, send_remaining = len; 123 size_t chunk_len, chunk_off, chunk_len_sum; 124 struct smc_connection *conn = &smc->conn; 125 union smc_host_cursor prep; 126 struct sock *sk = &smc->sk; 127 char *sndbuf_base; 128 int tx_cnt_prep; 129 int writespace; 130 int rc, chunk; 131 132 /* This should be in poll */ 133 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 134 135 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) { 136 rc = -EPIPE; 137 goto out_err; 138 } 139 140 while (msg_data_left(msg)) { 141 if (sk->sk_state == SMC_INIT) 142 return -ENOTCONN; 143 if (smc->sk.sk_shutdown & SEND_SHUTDOWN || 144 (smc->sk.sk_err == ECONNABORTED) || 145 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort) 146 return -EPIPE; 147 if (smc_cdc_rxed_any_close(conn)) 148 return send_done ?: -ECONNRESET; 149 150 if (!atomic_read(&conn->sndbuf_space)) { 151 rc = smc_tx_wait_memory(smc, msg->msg_flags); 152 if (rc) { 153 if (send_done) 154 return send_done; 155 goto out_err; 156 } 157 continue; 158 } 159 160 /* initialize variables for 1st iteration of subsequent loop */ 161 /* could be just 1 byte, even after smc_tx_wait_memory above */ 162 writespace = atomic_read(&conn->sndbuf_space); 163 /* not more than what user space asked for */ 164 copylen = min_t(size_t, send_remaining, writespace); 165 /* determine start of sndbuf */ 166 sndbuf_base = conn->sndbuf_desc->cpu_addr; 167 smc_curs_write(&prep, 168 smc_curs_read(&conn->tx_curs_prep, conn), 169 conn); 170 tx_cnt_prep = prep.count; 171 /* determine chunks where to write into sndbuf */ 172 /* either unwrapped case, or 1st chunk of wrapped case */ 173 chunk_len = min_t(size_t, 174 copylen, conn->sndbuf_size - tx_cnt_prep); 175 chunk_len_sum = chunk_len; 176 chunk_off = tx_cnt_prep; 177 for (chunk = 0; chunk < 2; chunk++) { 178 rc = memcpy_from_msg(sndbuf_base + chunk_off, 179 msg, chunk_len); 180 if (rc) { 181 if (send_done) 182 return send_done; 183 goto out_err; 184 } 185 send_done += chunk_len; 186 send_remaining -= chunk_len; 187 188 if (chunk_len_sum == copylen) 189 break; /* either on 1st or 2nd iteration */ 190 /* prepare next (== 2nd) iteration */ 191 chunk_len = copylen - chunk_len; /* remainder */ 192 chunk_len_sum += chunk_len; 193 chunk_off = 0; /* modulo offset in send ring buffer */ 194 } 195 /* update cursors */ 196 smc_curs_add(conn->sndbuf_size, &prep, copylen); 197 smc_curs_write(&conn->tx_curs_prep, 198 smc_curs_read(&prep, conn), 199 conn); 200 /* increased in send tasklet smc_cdc_tx_handler() */ 201 smp_mb__before_atomic(); 202 atomic_sub(copylen, &conn->sndbuf_space); 203 /* guarantee 0 <= sndbuf_space <= sndbuf_size */ 204 smp_mb__after_atomic(); 205 /* since we just produced more new data into sndbuf, 206 * trigger sndbuf consumer: RDMA write into peer RMBE and CDC 207 */ 208 smc_tx_sndbuf_nonempty(conn); 209 } /* while (msg_data_left(msg)) */ 210 211 return send_done; 212 213 out_err: 214 rc = sk_stream_error(sk, msg->msg_flags, rc); 215 /* make sure we wake any epoll edge trigger waiter */ 216 if (unlikely(rc == -EAGAIN)) 217 sk->sk_write_space(sk); 218 return rc; 219 } 220 221 /***************************** sndbuf consumer *******************************/ 222 223 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */ 224 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset, 225 int num_sges, struct ib_sge sges[]) 226 { 227 struct smc_link_group *lgr = conn->lgr; 228 struct ib_send_wr *failed_wr = NULL; 229 struct ib_rdma_wr rdma_wr; 230 struct smc_link *link; 231 int rc; 232 233 memset(&rdma_wr, 0, sizeof(rdma_wr)); 234 link = &lgr->lnk[SMC_SINGLE_LINK]; 235 rdma_wr.wr.wr_id = smc_wr_tx_get_next_wr_id(link); 236 rdma_wr.wr.sg_list = sges; 237 rdma_wr.wr.num_sge = num_sges; 238 rdma_wr.wr.opcode = IB_WR_RDMA_WRITE; 239 rdma_wr.remote_addr = 240 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr + 241 /* RMBE within RMB */ 242 ((conn->peer_conn_idx - 1) * conn->peer_rmbe_size) + 243 /* offset within RMBE */ 244 peer_rmbe_offset; 245 rdma_wr.rkey = lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey; 246 rc = ib_post_send(link->roce_qp, &rdma_wr.wr, &failed_wr); 247 if (rc) 248 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 249 return rc; 250 } 251 252 /* sndbuf consumer */ 253 static inline void smc_tx_advance_cursors(struct smc_connection *conn, 254 union smc_host_cursor *prod, 255 union smc_host_cursor *sent, 256 size_t len) 257 { 258 smc_curs_add(conn->peer_rmbe_size, prod, len); 259 /* increased in recv tasklet smc_cdc_msg_rcv() */ 260 smp_mb__before_atomic(); 261 /* data in flight reduces usable snd_wnd */ 262 atomic_sub(len, &conn->peer_rmbe_space); 263 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 264 smp_mb__after_atomic(); 265 smc_curs_add(conn->sndbuf_size, sent, len); 266 } 267 268 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit; 269 * usable snd_wnd as max transmit 270 */ 271 static int smc_tx_rdma_writes(struct smc_connection *conn) 272 { 273 size_t src_off, src_len, dst_off, dst_len; /* current chunk values */ 274 size_t len, dst_len_sum, src_len_sum, dstchunk, srcchunk; 275 union smc_host_cursor sent, prep, prod, cons; 276 struct ib_sge sges[SMC_IB_MAX_SEND_SGE]; 277 struct smc_link_group *lgr = conn->lgr; 278 int to_send, rmbespace; 279 struct smc_link *link; 280 int num_sges; 281 int rc; 282 283 /* source: sndbuf */ 284 smc_curs_write(&sent, smc_curs_read(&conn->tx_curs_sent, conn), conn); 285 smc_curs_write(&prep, smc_curs_read(&conn->tx_curs_prep, conn), conn); 286 /* cf. wmem_alloc - (snd_max - snd_una) */ 287 to_send = smc_curs_diff(conn->sndbuf_size, &sent, &prep); 288 if (to_send <= 0) 289 return 0; 290 291 /* destination: RMBE */ 292 /* cf. snd_wnd */ 293 rmbespace = atomic_read(&conn->peer_rmbe_space); 294 if (rmbespace <= 0) 295 return 0; 296 smc_curs_write(&prod, 297 smc_curs_read(&conn->local_tx_ctrl.prod, conn), 298 conn); 299 smc_curs_write(&cons, 300 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 301 conn); 302 303 /* if usable snd_wnd closes ask peer to advertise once it opens again */ 304 conn->local_tx_ctrl.prod_flags.write_blocked = (to_send >= rmbespace); 305 /* cf. usable snd_wnd */ 306 len = min(to_send, rmbespace); 307 308 /* initialize variables for first iteration of subsequent nested loop */ 309 link = &lgr->lnk[SMC_SINGLE_LINK]; 310 dst_off = prod.count; 311 if (prod.wrap == cons.wrap) { 312 /* the filled destination area is unwrapped, 313 * hence the available free destination space is wrapped 314 * and we need 2 destination chunks of sum len; start with 1st 315 * which is limited by what's available in sndbuf 316 */ 317 dst_len = min_t(size_t, 318 conn->peer_rmbe_size - prod.count, len); 319 } else { 320 /* the filled destination area is wrapped, 321 * hence the available free destination space is unwrapped 322 * and we need a single destination chunk of entire len 323 */ 324 dst_len = len; 325 } 326 dst_len_sum = dst_len; 327 src_off = sent.count; 328 /* dst_len determines the maximum src_len */ 329 if (sent.count + dst_len <= conn->sndbuf_size) { 330 /* unwrapped src case: single chunk of entire dst_len */ 331 src_len = dst_len; 332 } else { 333 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */ 334 src_len = conn->sndbuf_size - sent.count; 335 } 336 src_len_sum = src_len; 337 for (dstchunk = 0; dstchunk < 2; dstchunk++) { 338 num_sges = 0; 339 for (srcchunk = 0; srcchunk < 2; srcchunk++) { 340 sges[srcchunk].addr = 341 conn->sndbuf_desc->dma_addr[SMC_SINGLE_LINK] + 342 src_off; 343 sges[srcchunk].length = src_len; 344 sges[srcchunk].lkey = link->roce_pd->local_dma_lkey; 345 num_sges++; 346 src_off += src_len; 347 if (src_off >= conn->sndbuf_size) 348 src_off -= conn->sndbuf_size; 349 /* modulo in send ring */ 350 if (src_len_sum == dst_len) 351 break; /* either on 1st or 2nd iteration */ 352 /* prepare next (== 2nd) iteration */ 353 src_len = dst_len - src_len; /* remainder */ 354 src_len_sum += src_len; 355 } 356 rc = smc_tx_rdma_write(conn, dst_off, num_sges, sges); 357 if (rc) 358 return rc; 359 if (dst_len_sum == len) 360 break; /* either on 1st or 2nd iteration */ 361 /* prepare next (== 2nd) iteration */ 362 dst_off = 0; /* modulo offset in RMBE ring buffer */ 363 dst_len = len - dst_len; /* remainder */ 364 dst_len_sum += dst_len; 365 src_len = min_t(int, 366 dst_len, conn->sndbuf_size - sent.count); 367 src_len_sum = src_len; 368 } 369 370 smc_tx_advance_cursors(conn, &prod, &sent, len); 371 /* update connection's cursors with advanced local cursors */ 372 smc_curs_write(&conn->local_tx_ctrl.prod, 373 smc_curs_read(&prod, conn), 374 conn); 375 /* dst: peer RMBE */ 376 smc_curs_write(&conn->tx_curs_sent, 377 smc_curs_read(&sent, conn), 378 conn); 379 /* src: local sndbuf */ 380 381 return 0; 382 } 383 384 /* Wakeup sndbuf consumers from any context (IRQ or process) 385 * since there is more data to transmit; usable snd_wnd as max transmit 386 */ 387 int smc_tx_sndbuf_nonempty(struct smc_connection *conn) 388 { 389 struct smc_cdc_tx_pend *pend; 390 struct smc_wr_buf *wr_buf; 391 int rc; 392 393 spin_lock_bh(&conn->send_lock); 394 rc = smc_cdc_get_free_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], &wr_buf, 395 &pend); 396 if (rc < 0) { 397 if (rc == -EBUSY) { 398 struct smc_sock *smc = 399 container_of(conn, struct smc_sock, conn); 400 401 if (smc->sk.sk_err == ECONNABORTED) { 402 rc = sock_error(&smc->sk); 403 goto out_unlock; 404 } 405 rc = 0; 406 schedule_work(&conn->tx_work); 407 } 408 goto out_unlock; 409 } 410 411 rc = smc_tx_rdma_writes(conn); 412 if (rc) { 413 smc_wr_tx_put_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], 414 (struct smc_wr_tx_pend_priv *)pend); 415 goto out_unlock; 416 } 417 418 rc = smc_cdc_msg_send(conn, wr_buf, pend); 419 420 out_unlock: 421 spin_unlock_bh(&conn->send_lock); 422 return rc; 423 } 424 425 /* Wakeup sndbuf consumers from process context 426 * since there is more data to transmit 427 */ 428 static void smc_tx_work(struct work_struct *work) 429 { 430 struct smc_connection *conn = container_of(work, 431 struct smc_connection, 432 tx_work); 433 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 434 int rc; 435 436 lock_sock(&smc->sk); 437 rc = smc_tx_sndbuf_nonempty(conn); 438 if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked && 439 !atomic_read(&conn->bytes_to_rcv)) 440 conn->local_rx_ctrl.prod_flags.write_blocked = 0; 441 release_sock(&smc->sk); 442 } 443 444 void smc_tx_consumer_update(struct smc_connection *conn) 445 { 446 union smc_host_cursor cfed, cons; 447 struct smc_cdc_tx_pend *pend; 448 struct smc_wr_buf *wr_buf; 449 int to_confirm, rc; 450 451 smc_curs_write(&cons, 452 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 453 conn); 454 smc_curs_write(&cfed, 455 smc_curs_read(&conn->rx_curs_confirmed, conn), 456 conn); 457 to_confirm = smc_curs_diff(conn->rmbe_size, &cfed, &cons); 458 459 if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 460 ((to_confirm > conn->rmbe_update_limit) && 461 ((to_confirm > (conn->rmbe_size / 2)) || 462 conn->local_rx_ctrl.prod_flags.write_blocked))) { 463 rc = smc_cdc_get_free_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], 464 &wr_buf, &pend); 465 if (!rc) 466 rc = smc_cdc_msg_send(conn, wr_buf, pend); 467 if (rc < 0) { 468 schedule_work(&conn->tx_work); 469 return; 470 } 471 smc_curs_write(&conn->rx_curs_confirmed, 472 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 473 conn); 474 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 475 } 476 if (conn->local_rx_ctrl.prod_flags.write_blocked && 477 !atomic_read(&conn->bytes_to_rcv)) 478 conn->local_rx_ctrl.prod_flags.write_blocked = 0; 479 } 480 481 /***************************** send initialize *******************************/ 482 483 /* Initialize send properties on connection establishment. NB: not __init! */ 484 void smc_tx_init(struct smc_sock *smc) 485 { 486 smc->sk.sk_write_space = smc_tx_write_space; 487 INIT_WORK(&smc->conn.tx_work, smc_tx_work); 488 spin_lock_init(&smc->conn.send_lock); 489 } 490