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 POLLOUT | POLLWRNORM | 50 POLLWRBAND); 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 (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 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_or_senddone(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 return rc; 254 } 255 256 /* sndbuf consumer */ 257 static inline void smc_tx_advance_cursors(struct smc_connection *conn, 258 union smc_host_cursor *prod, 259 union smc_host_cursor *sent, 260 size_t len) 261 { 262 smc_curs_add(conn->peer_rmbe_size, prod, len); 263 /* increased in recv tasklet smc_cdc_msg_rcv() */ 264 smp_mb__before_atomic(); 265 /* data in flight reduces usable snd_wnd */ 266 atomic_sub(len, &conn->peer_rmbe_space); 267 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 268 smp_mb__after_atomic(); 269 smc_curs_add(conn->sndbuf_size, sent, len); 270 } 271 272 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit; 273 * usable snd_wnd as max transmit 274 */ 275 static int smc_tx_rdma_writes(struct smc_connection *conn) 276 { 277 size_t src_off, src_len, dst_off, dst_len; /* current chunk values */ 278 size_t len, dst_len_sum, src_len_sum, dstchunk, srcchunk; 279 union smc_host_cursor sent, prep, prod, cons; 280 struct ib_sge sges[SMC_IB_MAX_SEND_SGE]; 281 struct smc_link_group *lgr = conn->lgr; 282 int to_send, rmbespace; 283 struct smc_link *link; 284 dma_addr_t dma_addr; 285 int num_sges; 286 int rc; 287 288 /* source: sndbuf */ 289 smc_curs_write(&sent, smc_curs_read(&conn->tx_curs_sent, conn), conn); 290 smc_curs_write(&prep, smc_curs_read(&conn->tx_curs_prep, conn), conn); 291 /* cf. wmem_alloc - (snd_max - snd_una) */ 292 to_send = smc_curs_diff(conn->sndbuf_size, &sent, &prep); 293 if (to_send <= 0) 294 return 0; 295 296 /* destination: RMBE */ 297 /* cf. snd_wnd */ 298 rmbespace = atomic_read(&conn->peer_rmbe_space); 299 if (rmbespace <= 0) 300 return 0; 301 smc_curs_write(&prod, 302 smc_curs_read(&conn->local_tx_ctrl.prod, conn), 303 conn); 304 smc_curs_write(&cons, 305 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 306 conn); 307 308 /* if usable snd_wnd closes ask peer to advertise once it opens again */ 309 conn->local_tx_ctrl.prod_flags.write_blocked = (to_send >= rmbespace); 310 /* cf. usable snd_wnd */ 311 len = min(to_send, rmbespace); 312 313 /* initialize variables for first iteration of subsequent nested loop */ 314 link = &lgr->lnk[SMC_SINGLE_LINK]; 315 dst_off = prod.count; 316 if (prod.wrap == cons.wrap) { 317 /* the filled destination area is unwrapped, 318 * hence the available free destination space is wrapped 319 * and we need 2 destination chunks of sum len; start with 1st 320 * which is limited by what's available in sndbuf 321 */ 322 dst_len = min_t(size_t, 323 conn->peer_rmbe_size - prod.count, len); 324 } else { 325 /* the filled destination area is wrapped, 326 * hence the available free destination space is unwrapped 327 * and we need a single destination chunk of entire len 328 */ 329 dst_len = len; 330 } 331 dst_len_sum = dst_len; 332 src_off = sent.count; 333 /* dst_len determines the maximum src_len */ 334 if (sent.count + dst_len <= conn->sndbuf_size) { 335 /* unwrapped src case: single chunk of entire dst_len */ 336 src_len = dst_len; 337 } else { 338 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */ 339 src_len = conn->sndbuf_size - sent.count; 340 } 341 src_len_sum = src_len; 342 dma_addr = sg_dma_address(conn->sndbuf_desc->sgt[SMC_SINGLE_LINK].sgl); 343 for (dstchunk = 0; dstchunk < 2; dstchunk++) { 344 num_sges = 0; 345 for (srcchunk = 0; srcchunk < 2; srcchunk++) { 346 sges[srcchunk].addr = dma_addr + src_off; 347 sges[srcchunk].length = src_len; 348 sges[srcchunk].lkey = link->roce_pd->local_dma_lkey; 349 num_sges++; 350 src_off += src_len; 351 if (src_off >= conn->sndbuf_size) 352 src_off -= conn->sndbuf_size; 353 /* modulo in send ring */ 354 if (src_len_sum == dst_len) 355 break; /* either on 1st or 2nd iteration */ 356 /* prepare next (== 2nd) iteration */ 357 src_len = dst_len - src_len; /* remainder */ 358 src_len_sum += src_len; 359 } 360 rc = smc_tx_rdma_write(conn, dst_off, num_sges, sges); 361 if (rc) 362 return rc; 363 if (dst_len_sum == len) 364 break; /* either on 1st or 2nd iteration */ 365 /* prepare next (== 2nd) iteration */ 366 dst_off = 0; /* modulo offset in RMBE ring buffer */ 367 dst_len = len - dst_len; /* remainder */ 368 dst_len_sum += dst_len; 369 src_len = min_t(int, 370 dst_len, conn->sndbuf_size - sent.count); 371 src_len_sum = src_len; 372 } 373 374 smc_tx_advance_cursors(conn, &prod, &sent, len); 375 /* update connection's cursors with advanced local cursors */ 376 smc_curs_write(&conn->local_tx_ctrl.prod, 377 smc_curs_read(&prod, conn), 378 conn); 379 /* dst: peer RMBE */ 380 smc_curs_write(&conn->tx_curs_sent, 381 smc_curs_read(&sent, conn), 382 conn); 383 /* src: local sndbuf */ 384 385 return 0; 386 } 387 388 /* Wakeup sndbuf consumers from any context (IRQ or process) 389 * since there is more data to transmit; usable snd_wnd as max transmit 390 */ 391 int smc_tx_sndbuf_nonempty(struct smc_connection *conn) 392 { 393 struct smc_cdc_tx_pend *pend; 394 struct smc_wr_buf *wr_buf; 395 int rc; 396 397 spin_lock_bh(&conn->send_lock); 398 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend); 399 if (rc < 0) { 400 if (rc == -EBUSY) { 401 struct smc_sock *smc = 402 container_of(conn, struct smc_sock, conn); 403 404 if (smc->sk.sk_err == ECONNABORTED) { 405 rc = sock_error(&smc->sk); 406 goto out_unlock; 407 } 408 rc = 0; 409 schedule_delayed_work(&conn->tx_work, 410 SMC_TX_WORK_DELAY); 411 } 412 goto out_unlock; 413 } 414 415 rc = smc_tx_rdma_writes(conn); 416 if (rc) { 417 smc_wr_tx_put_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], 418 (struct smc_wr_tx_pend_priv *)pend); 419 goto out_unlock; 420 } 421 422 rc = smc_cdc_msg_send(conn, wr_buf, pend); 423 424 out_unlock: 425 spin_unlock_bh(&conn->send_lock); 426 return rc; 427 } 428 429 /* Wakeup sndbuf consumers from process context 430 * since there is more data to transmit 431 */ 432 static void smc_tx_work(struct work_struct *work) 433 { 434 struct smc_connection *conn = container_of(to_delayed_work(work), 435 struct smc_connection, 436 tx_work); 437 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 438 int rc; 439 440 lock_sock(&smc->sk); 441 rc = smc_tx_sndbuf_nonempty(conn); 442 if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked && 443 !atomic_read(&conn->bytes_to_rcv)) 444 conn->local_rx_ctrl.prod_flags.write_blocked = 0; 445 release_sock(&smc->sk); 446 } 447 448 void smc_tx_consumer_update(struct smc_connection *conn) 449 { 450 union smc_host_cursor cfed, cons; 451 int to_confirm; 452 453 smc_curs_write(&cons, 454 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 455 conn); 456 smc_curs_write(&cfed, 457 smc_curs_read(&conn->rx_curs_confirmed, conn), 458 conn); 459 to_confirm = smc_curs_diff(conn->rmbe_size, &cfed, &cons); 460 461 if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 462 ((to_confirm > conn->rmbe_update_limit) && 463 ((to_confirm > (conn->rmbe_size / 2)) || 464 conn->local_rx_ctrl.prod_flags.write_blocked))) { 465 if (smc_cdc_get_slot_and_msg_send(conn) < 0) { 466 schedule_delayed_work(&conn->tx_work, 467 SMC_TX_WORK_DELAY); 468 return; 469 } 470 smc_curs_write(&conn->rx_curs_confirmed, 471 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 472 conn); 473 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 474 } 475 if (conn->local_rx_ctrl.prod_flags.write_blocked && 476 !atomic_read(&conn->bytes_to_rcv)) 477 conn->local_rx_ctrl.prod_flags.write_blocked = 0; 478 } 479 480 /***************************** send initialize *******************************/ 481 482 /* Initialize send properties on connection establishment. NB: not __init! */ 483 void smc_tx_init(struct smc_sock *smc) 484 { 485 smc->sk.sk_write_space = smc_tx_write_space; 486 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work); 487 spin_lock_init(&smc->conn.send_lock); 488 } 489