1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Shared Memory Communications over RDMA (SMC-R) and RoCE 4 * 5 * Manage send buffer. 6 * Producer: 7 * Copy user space data into send buffer, if send buffer space available. 8 * Consumer: 9 * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available. 10 * 11 * Copyright IBM Corp. 2016 12 * 13 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com> 14 */ 15 16 #include <linux/net.h> 17 #include <linux/rcupdate.h> 18 #include <linux/workqueue.h> 19 #include <linux/sched/signal.h> 20 21 #include <net/sock.h> 22 23 #include "smc.h" 24 #include "smc_wr.h" 25 #include "smc_cdc.h" 26 #include "smc_tx.h" 27 28 #define SMC_TX_WORK_DELAY HZ 29 30 /***************************** sndbuf producer *******************************/ 31 32 /* callback implementation for sk.sk_write_space() 33 * to wakeup sndbuf producers that blocked with smc_tx_wait_memory(). 34 * called under sk_socket lock. 35 */ 36 static void smc_tx_write_space(struct sock *sk) 37 { 38 struct socket *sock = sk->sk_socket; 39 struct smc_sock *smc = smc_sk(sk); 40 struct socket_wq *wq; 41 42 /* similar to sk_stream_write_space */ 43 if (atomic_read(&smc->conn.sndbuf_space) && sock) { 44 clear_bit(SOCK_NOSPACE, &sock->flags); 45 rcu_read_lock(); 46 wq = rcu_dereference(sk->sk_wq); 47 if (skwq_has_sleeper(wq)) 48 wake_up_interruptible_poll(&wq->wait, 49 EPOLLOUT | EPOLLWRNORM | 50 EPOLLWRBAND); 51 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN)) 52 sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT); 53 rcu_read_unlock(); 54 } 55 } 56 57 /* Wakeup sndbuf producers that blocked with smc_tx_wait_memory(). 58 * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space(). 59 */ 60 void smc_tx_sndbuf_nonfull(struct smc_sock *smc) 61 { 62 if (smc->sk.sk_socket && 63 test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags)) 64 smc->sk.sk_write_space(&smc->sk); 65 } 66 67 /* blocks sndbuf producer until at least one byte of free space available */ 68 static int smc_tx_wait_memory(struct smc_sock *smc, int flags) 69 { 70 DEFINE_WAIT_FUNC(wait, woken_wake_function); 71 struct smc_connection *conn = &smc->conn; 72 struct sock *sk = &smc->sk; 73 bool noblock; 74 long timeo; 75 int rc = 0; 76 77 /* similar to sk_stream_wait_memory */ 78 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 79 noblock = timeo ? false : true; 80 add_wait_queue(sk_sleep(sk), &wait); 81 while (1) { 82 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk); 83 if (sk->sk_err || 84 (sk->sk_shutdown & SEND_SHUTDOWN) || 85 conn->local_tx_ctrl.conn_state_flags.peer_done_writing) { 86 rc = -EPIPE; 87 break; 88 } 89 if (smc_cdc_rxed_any_close(conn)) { 90 rc = -ECONNRESET; 91 break; 92 } 93 if (!timeo) { 94 if (noblock) 95 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 96 rc = -EAGAIN; 97 break; 98 } 99 if (signal_pending(current)) { 100 rc = sock_intr_errno(timeo); 101 break; 102 } 103 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 104 if (atomic_read(&conn->sndbuf_space)) 105 break; /* at least 1 byte of free space available */ 106 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 107 sk_wait_event(sk, &timeo, 108 sk->sk_err || 109 (sk->sk_shutdown & SEND_SHUTDOWN) || 110 smc_cdc_rxed_any_close(conn) || 111 atomic_read(&conn->sndbuf_space), 112 &wait); 113 } 114 remove_wait_queue(sk_sleep(sk), &wait); 115 return rc; 116 } 117 118 /* sndbuf producer: main API called by socket layer. 119 * called under sock lock. 120 */ 121 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len) 122 { 123 size_t copylen, send_done = 0, send_remaining = len; 124 size_t chunk_len, chunk_off, chunk_len_sum; 125 struct smc_connection *conn = &smc->conn; 126 union smc_host_cursor prep; 127 struct sock *sk = &smc->sk; 128 char *sndbuf_base; 129 int tx_cnt_prep; 130 int writespace; 131 int rc, chunk; 132 133 /* This should be in poll */ 134 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk); 135 136 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) { 137 rc = -EPIPE; 138 goto out_err; 139 } 140 141 while (msg_data_left(msg)) { 142 if (sk->sk_state == SMC_INIT) 143 return -ENOTCONN; 144 if (smc->sk.sk_shutdown & SEND_SHUTDOWN || 145 (smc->sk.sk_err == ECONNABORTED) || 146 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort) 147 return -EPIPE; 148 if (smc_cdc_rxed_any_close(conn)) 149 return send_done ?: -ECONNRESET; 150 151 if (!atomic_read(&conn->sndbuf_space)) { 152 rc = smc_tx_wait_memory(smc, msg->msg_flags); 153 if (rc) { 154 if (send_done) 155 return send_done; 156 goto out_err; 157 } 158 continue; 159 } 160 161 /* initialize variables for 1st iteration of subsequent loop */ 162 /* could be just 1 byte, even after smc_tx_wait_memory above */ 163 writespace = atomic_read(&conn->sndbuf_space); 164 /* not more than what user space asked for */ 165 copylen = min_t(size_t, send_remaining, writespace); 166 /* determine start of sndbuf */ 167 sndbuf_base = conn->sndbuf_desc->cpu_addr; 168 smc_curs_write(&prep, 169 smc_curs_read(&conn->tx_curs_prep, conn), 170 conn); 171 tx_cnt_prep = prep.count; 172 /* determine chunks where to write into sndbuf */ 173 /* either unwrapped case, or 1st chunk of wrapped case */ 174 chunk_len = min_t(size_t, 175 copylen, conn->sndbuf_size - tx_cnt_prep); 176 chunk_len_sum = chunk_len; 177 chunk_off = tx_cnt_prep; 178 smc_sndbuf_sync_sg_for_cpu(conn); 179 for (chunk = 0; chunk < 2; chunk++) { 180 rc = memcpy_from_msg(sndbuf_base + chunk_off, 181 msg, chunk_len); 182 if (rc) { 183 smc_sndbuf_sync_sg_for_device(conn); 184 if (send_done) 185 return send_done; 186 goto out_err; 187 } 188 send_done += chunk_len; 189 send_remaining -= chunk_len; 190 191 if (chunk_len_sum == copylen) 192 break; /* either on 1st or 2nd iteration */ 193 /* prepare next (== 2nd) iteration */ 194 chunk_len = copylen - chunk_len; /* remainder */ 195 chunk_len_sum += chunk_len; 196 chunk_off = 0; /* modulo offset in send ring buffer */ 197 } 198 smc_sndbuf_sync_sg_for_device(conn); 199 /* update cursors */ 200 smc_curs_add(conn->sndbuf_size, &prep, copylen); 201 smc_curs_write(&conn->tx_curs_prep, 202 smc_curs_read(&prep, conn), 203 conn); 204 /* increased in send tasklet smc_cdc_tx_handler() */ 205 smp_mb__before_atomic(); 206 atomic_sub(copylen, &conn->sndbuf_space); 207 /* guarantee 0 <= sndbuf_space <= sndbuf_size */ 208 smp_mb__after_atomic(); 209 /* since we just produced more new data into sndbuf, 210 * trigger sndbuf consumer: RDMA write into peer RMBE and CDC 211 */ 212 smc_tx_sndbuf_nonempty(conn); 213 } /* while (msg_data_left(msg)) */ 214 215 return send_done; 216 217 out_err: 218 rc = sk_stream_error(sk, msg->msg_flags, rc); 219 /* make sure we wake any epoll edge trigger waiter */ 220 if (unlikely(rc == -EAGAIN)) 221 sk->sk_write_space(sk); 222 return rc; 223 } 224 225 /***************************** sndbuf consumer *******************************/ 226 227 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */ 228 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset, 229 int num_sges, struct ib_sge sges[]) 230 { 231 struct smc_link_group *lgr = conn->lgr; 232 struct ib_send_wr *failed_wr = NULL; 233 struct ib_rdma_wr rdma_wr; 234 struct smc_link *link; 235 int rc; 236 237 memset(&rdma_wr, 0, sizeof(rdma_wr)); 238 link = &lgr->lnk[SMC_SINGLE_LINK]; 239 rdma_wr.wr.wr_id = smc_wr_tx_get_next_wr_id(link); 240 rdma_wr.wr.sg_list = sges; 241 rdma_wr.wr.num_sge = num_sges; 242 rdma_wr.wr.opcode = IB_WR_RDMA_WRITE; 243 rdma_wr.remote_addr = 244 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr + 245 /* RMBE within RMB */ 246 ((conn->peer_conn_idx - 1) * conn->peer_rmbe_size) + 247 /* offset within RMBE */ 248 peer_rmbe_offset; 249 rdma_wr.rkey = lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey; 250 rc = ib_post_send(link->roce_qp, &rdma_wr.wr, &failed_wr); 251 if (rc) { 252 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 253 smc_lgr_terminate(lgr); 254 } 255 return rc; 256 } 257 258 /* sndbuf consumer */ 259 static inline void smc_tx_advance_cursors(struct smc_connection *conn, 260 union smc_host_cursor *prod, 261 union smc_host_cursor *sent, 262 size_t len) 263 { 264 smc_curs_add(conn->peer_rmbe_size, prod, len); 265 /* increased in recv tasklet smc_cdc_msg_rcv() */ 266 smp_mb__before_atomic(); 267 /* data in flight reduces usable snd_wnd */ 268 atomic_sub(len, &conn->peer_rmbe_space); 269 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 270 smp_mb__after_atomic(); 271 smc_curs_add(conn->sndbuf_size, sent, len); 272 } 273 274 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit; 275 * usable snd_wnd as max transmit 276 */ 277 static int smc_tx_rdma_writes(struct smc_connection *conn) 278 { 279 size_t src_off, src_len, dst_off, dst_len; /* current chunk values */ 280 size_t len, dst_len_sum, src_len_sum, dstchunk, srcchunk; 281 union smc_host_cursor sent, prep, prod, cons; 282 struct ib_sge sges[SMC_IB_MAX_SEND_SGE]; 283 struct smc_link_group *lgr = conn->lgr; 284 int to_send, rmbespace; 285 struct smc_link *link; 286 dma_addr_t dma_addr; 287 int num_sges; 288 int rc; 289 290 /* source: sndbuf */ 291 smc_curs_write(&sent, smc_curs_read(&conn->tx_curs_sent, conn), conn); 292 smc_curs_write(&prep, smc_curs_read(&conn->tx_curs_prep, conn), conn); 293 /* cf. wmem_alloc - (snd_max - snd_una) */ 294 to_send = smc_curs_diff(conn->sndbuf_size, &sent, &prep); 295 if (to_send <= 0) 296 return 0; 297 298 /* destination: RMBE */ 299 /* cf. snd_wnd */ 300 rmbespace = atomic_read(&conn->peer_rmbe_space); 301 if (rmbespace <= 0) 302 return 0; 303 smc_curs_write(&prod, 304 smc_curs_read(&conn->local_tx_ctrl.prod, conn), 305 conn); 306 smc_curs_write(&cons, 307 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 308 conn); 309 310 /* if usable snd_wnd closes ask peer to advertise once it opens again */ 311 conn->local_tx_ctrl.prod_flags.write_blocked = (to_send >= rmbespace); 312 /* cf. usable snd_wnd */ 313 len = min(to_send, rmbespace); 314 315 /* initialize variables for first iteration of subsequent nested loop */ 316 link = &lgr->lnk[SMC_SINGLE_LINK]; 317 dst_off = prod.count; 318 if (prod.wrap == cons.wrap) { 319 /* the filled destination area is unwrapped, 320 * hence the available free destination space is wrapped 321 * and we need 2 destination chunks of sum len; start with 1st 322 * which is limited by what's available in sndbuf 323 */ 324 dst_len = min_t(size_t, 325 conn->peer_rmbe_size - prod.count, len); 326 } else { 327 /* the filled destination area is wrapped, 328 * hence the available free destination space is unwrapped 329 * and we need a single destination chunk of entire len 330 */ 331 dst_len = len; 332 } 333 dst_len_sum = dst_len; 334 src_off = sent.count; 335 /* dst_len determines the maximum src_len */ 336 if (sent.count + dst_len <= conn->sndbuf_size) { 337 /* unwrapped src case: single chunk of entire dst_len */ 338 src_len = dst_len; 339 } else { 340 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */ 341 src_len = conn->sndbuf_size - sent.count; 342 } 343 src_len_sum = src_len; 344 dma_addr = sg_dma_address(conn->sndbuf_desc->sgt[SMC_SINGLE_LINK].sgl); 345 for (dstchunk = 0; dstchunk < 2; dstchunk++) { 346 num_sges = 0; 347 for (srcchunk = 0; srcchunk < 2; srcchunk++) { 348 sges[srcchunk].addr = dma_addr + src_off; 349 sges[srcchunk].length = src_len; 350 sges[srcchunk].lkey = link->roce_pd->local_dma_lkey; 351 num_sges++; 352 src_off += src_len; 353 if (src_off >= conn->sndbuf_size) 354 src_off -= conn->sndbuf_size; 355 /* modulo in send ring */ 356 if (src_len_sum == dst_len) 357 break; /* either on 1st or 2nd iteration */ 358 /* prepare next (== 2nd) iteration */ 359 src_len = dst_len - src_len; /* remainder */ 360 src_len_sum += src_len; 361 } 362 rc = smc_tx_rdma_write(conn, dst_off, num_sges, sges); 363 if (rc) 364 return rc; 365 if (dst_len_sum == len) 366 break; /* either on 1st or 2nd iteration */ 367 /* prepare next (== 2nd) iteration */ 368 dst_off = 0; /* modulo offset in RMBE ring buffer */ 369 dst_len = len - dst_len; /* remainder */ 370 dst_len_sum += dst_len; 371 src_len = min_t(int, 372 dst_len, conn->sndbuf_size - sent.count); 373 src_len_sum = src_len; 374 } 375 376 smc_tx_advance_cursors(conn, &prod, &sent, len); 377 /* update connection's cursors with advanced local cursors */ 378 smc_curs_write(&conn->local_tx_ctrl.prod, 379 smc_curs_read(&prod, conn), 380 conn); 381 /* dst: peer RMBE */ 382 smc_curs_write(&conn->tx_curs_sent, 383 smc_curs_read(&sent, conn), 384 conn); 385 /* src: local sndbuf */ 386 387 return 0; 388 } 389 390 /* Wakeup sndbuf consumers from any context (IRQ or process) 391 * since there is more data to transmit; usable snd_wnd as max transmit 392 */ 393 int smc_tx_sndbuf_nonempty(struct smc_connection *conn) 394 { 395 struct smc_cdc_tx_pend *pend; 396 struct smc_wr_buf *wr_buf; 397 int rc; 398 399 spin_lock_bh(&conn->send_lock); 400 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend); 401 if (rc < 0) { 402 if (rc == -EBUSY) { 403 struct smc_sock *smc = 404 container_of(conn, struct smc_sock, conn); 405 406 if (smc->sk.sk_err == ECONNABORTED) { 407 rc = sock_error(&smc->sk); 408 goto out_unlock; 409 } 410 rc = 0; 411 if (conn->alert_token_local) /* connection healthy */ 412 schedule_delayed_work(&conn->tx_work, 413 SMC_TX_WORK_DELAY); 414 } 415 goto out_unlock; 416 } 417 418 rc = smc_tx_rdma_writes(conn); 419 if (rc) { 420 smc_wr_tx_put_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], 421 (struct smc_wr_tx_pend_priv *)pend); 422 goto out_unlock; 423 } 424 425 rc = smc_cdc_msg_send(conn, wr_buf, pend); 426 427 out_unlock: 428 spin_unlock_bh(&conn->send_lock); 429 return rc; 430 } 431 432 /* Wakeup sndbuf consumers from process context 433 * since there is more data to transmit 434 */ 435 static void smc_tx_work(struct work_struct *work) 436 { 437 struct smc_connection *conn = container_of(to_delayed_work(work), 438 struct smc_connection, 439 tx_work); 440 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 441 int rc; 442 443 lock_sock(&smc->sk); 444 if (smc->sk.sk_err || 445 !conn->alert_token_local || 446 conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) 447 goto out; 448 449 rc = smc_tx_sndbuf_nonempty(conn); 450 if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked && 451 !atomic_read(&conn->bytes_to_rcv)) 452 conn->local_rx_ctrl.prod_flags.write_blocked = 0; 453 454 out: 455 release_sock(&smc->sk); 456 } 457 458 void smc_tx_consumer_update(struct smc_connection *conn) 459 { 460 union smc_host_cursor cfed, cons; 461 int to_confirm; 462 463 smc_curs_write(&cons, 464 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 465 conn); 466 smc_curs_write(&cfed, 467 smc_curs_read(&conn->rx_curs_confirmed, conn), 468 conn); 469 to_confirm = smc_curs_diff(conn->rmbe_size, &cfed, &cons); 470 471 if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 472 ((to_confirm > conn->rmbe_update_limit) && 473 ((to_confirm > (conn->rmbe_size / 2)) || 474 conn->local_rx_ctrl.prod_flags.write_blocked))) { 475 if ((smc_cdc_get_slot_and_msg_send(conn) < 0) && 476 conn->alert_token_local) { /* connection healthy */ 477 schedule_delayed_work(&conn->tx_work, 478 SMC_TX_WORK_DELAY); 479 return; 480 } 481 smc_curs_write(&conn->rx_curs_confirmed, 482 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 483 conn); 484 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 485 } 486 if (conn->local_rx_ctrl.prod_flags.write_blocked && 487 !atomic_read(&conn->bytes_to_rcv)) 488 conn->local_rx_ctrl.prod_flags.write_blocked = 0; 489 } 490 491 /***************************** send initialize *******************************/ 492 493 /* Initialize send properties on connection establishment. NB: not __init! */ 494 void smc_tx_init(struct smc_sock *smc) 495 { 496 smc->sk.sk_write_space = smc_tx_write_space; 497 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work); 498 spin_lock_init(&smc->conn.send_lock); 499 } 500