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