1 /* 2 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved. 3 */ 4 5 /* 6 * This file contains code imported from the OFED rds source file ib_send.c 7 * Oracle elects to have and use the contents of ib_send.c under and governed 8 * by the OpenIB.org BSD license (see below for full license text). However, 9 * the following notice accompanied the original version of this file: 10 */ 11 12 /* 13 * Copyright (c) 2006 Oracle. All rights reserved. 14 * 15 * This software is available to you under a choice of one of two 16 * licenses. You may choose to be licensed under the terms of the GNU 17 * General Public License (GPL) Version 2, available from the file 18 * COPYING in the main directory of this source tree, or the 19 * OpenIB.org BSD license below: 20 * 21 * Redistribution and use in source and binary forms, with or 22 * without modification, are permitted provided that the following 23 * conditions are met: 24 * 25 * - Redistributions of source code must retain the above 26 * copyright notice, this list of conditions and the following 27 * disclaimer. 28 * 29 * - Redistributions in binary form must reproduce the above 30 * copyright notice, this list of conditions and the following 31 * disclaimer in the documentation and/or other materials 32 * provided with the distribution. 33 * 34 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 35 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 36 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 37 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 38 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 39 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 40 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 41 * SOFTWARE. 42 * 43 */ 44 #include <sys/rds.h> 45 46 #include <sys/ib/clients/rdsv3/rdsv3.h> 47 #include <sys/ib/clients/rdsv3/rdma.h> 48 #include <sys/ib/clients/rdsv3/ib.h> 49 #include <sys/ib/clients/rdsv3/rdsv3_debug.h> 50 51 static void 52 rdsv3_ib_send_rdma_complete(struct rdsv3_message *rm, 53 int wc_status) 54 { 55 int notify_status; 56 57 RDSV3_DPRINTF4("rdsv3_ib_send_rdma_complete", "rm: %p, wc_status: %d", 58 rm, wc_status); 59 60 switch (wc_status) { 61 case IBT_WC_WR_FLUSHED_ERR: 62 return; 63 64 case IBT_WC_SUCCESS: 65 notify_status = RDS_RDMA_SUCCESS; 66 break; 67 68 case IBT_WC_REMOTE_ACCESS_ERR: 69 notify_status = RDS_RDMA_REMOTE_ERROR; 70 break; 71 72 default: 73 notify_status = RDS_RDMA_OTHER_ERROR; 74 break; 75 } 76 rdsv3_rdma_send_complete(rm, notify_status); 77 78 RDSV3_DPRINTF4("rdsv3_ib_send_rdma_complete", "rm: %p, wc_status: %d", 79 rm, wc_status); 80 } 81 82 static void rdsv3_ib_dma_unmap_sg_rdma(struct ib_device *dev, 83 uint_t num, struct rdsv3_rdma_sg scat[]); 84 85 void 86 rdsv3_ib_send_unmap_rdma(struct rdsv3_ib_connection *ic, 87 struct rdsv3_rdma_op *op) 88 { 89 RDSV3_DPRINTF4("rdsv3_ib_send_unmap_rdma", "ic: %p, op: %p", ic, op); 90 if (op->r_mapped) { 91 op->r_mapped = 0; 92 if (ic->i_cm_id) { 93 rdsv3_ib_dma_unmap_sg_rdma(ic->i_cm_id->device, 94 op->r_nents, op->r_rdma_sg); 95 } else { 96 rdsv3_ib_dma_unmap_sg_rdma((struct ib_device *)NULL, 97 op->r_nents, op->r_rdma_sg); 98 } 99 } 100 } 101 102 static void 103 rdsv3_ib_send_unmap_rm(struct rdsv3_ib_connection *ic, 104 struct rdsv3_ib_send_work *send, 105 int wc_status) 106 { 107 struct rdsv3_message *rm = send->s_rm; 108 109 RDSV3_DPRINTF4("rdsv3_ib_send_unmap_rm", "ic %p send %p rm %p\n", 110 ic, send, rm); 111 112 mutex_enter(&rm->m_rs_lock); 113 if (rm->m_count) { 114 rdsv3_ib_dma_unmap_sg(ic->i_cm_id->device, 115 rm->m_sg, rm->m_count); 116 rm->m_count = 0; 117 } 118 mutex_exit(&rm->m_rs_lock); 119 120 if (rm->m_rdma_op != NULL) { 121 rdsv3_ib_send_unmap_rdma(ic, rm->m_rdma_op); 122 123 /* 124 * If the user asked for a completion notification on this 125 * message, we can implement three different semantics: 126 * 1. Notify when we received the ACK on the RDS message 127 * that was queued with the RDMA. This provides reliable 128 * notification of RDMA status at the expense of a one-way 129 * packet delay. 130 * 2. Notify when the IB stack gives us the completion 131 * event for the RDMA operation. 132 * 3. Notify when the IB stack gives us the completion 133 * event for the accompanying RDS messages. 134 * Here, we implement approach #3. To implement approach #2, 135 * call rdsv3_rdma_send_complete from the cq_handler. 136 * To implement #1, 137 * don't call rdsv3_rdma_send_complete at all, and fall back to 138 * the notify 139 * handling in the ACK processing code. 140 * 141 * Note: There's no need to explicitly sync any RDMA buffers 142 * using 143 * ib_dma_sync_sg_for_cpu - the completion for the RDMA 144 * operation itself unmapped the RDMA buffers, which takes care 145 * of synching. 146 */ 147 rdsv3_ib_send_rdma_complete(rm, wc_status); 148 149 if (rm->m_rdma_op->r_write) 150 rdsv3_stats_add(s_send_rdma_bytes, 151 rm->m_rdma_op->r_bytes); 152 else 153 rdsv3_stats_add(s_recv_rdma_bytes, 154 rm->m_rdma_op->r_bytes); 155 } 156 157 /* 158 * If anyone waited for this message to get flushed out, wake 159 * them up now 160 */ 161 rdsv3_message_unmapped(rm); 162 163 rdsv3_message_put(rm); 164 send->s_rm = NULL; 165 } 166 167 void 168 rdsv3_ib_send_init_ring(struct rdsv3_ib_connection *ic) 169 { 170 struct rdsv3_ib_send_work *send; 171 uint32_t i; 172 173 RDSV3_DPRINTF4("rdsv3_ib_send_init_ring", "ic: %p", ic); 174 175 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { 176 send->s_rm = NULL; 177 send->s_op = NULL; 178 } 179 } 180 181 void 182 rdsv3_ib_send_clear_ring(struct rdsv3_ib_connection *ic) 183 { 184 struct rdsv3_ib_send_work *send; 185 uint32_t i; 186 187 RDSV3_DPRINTF4("rdsv3_ib_send_clear_ring", "ic: %p", ic); 188 189 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { 190 if (send->s_opcode == 0xdd) 191 continue; 192 if (send->s_rm) 193 rdsv3_ib_send_unmap_rm(ic, send, IBT_WC_WR_FLUSHED_ERR); 194 if (send->s_op) 195 rdsv3_ib_send_unmap_rdma(ic, send->s_op); 196 } 197 198 RDSV3_DPRINTF4("rdsv3_ib_send_clear_ring", "Return: ic: %p", ic); 199 } 200 201 /* 202 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc 203 * operations performed in the send path. As the sender allocs and potentially 204 * unallocs the next free entry in the ring it doesn't alter which is 205 * the next to be freed, which is what this is concerned with. 206 */ 207 void 208 rdsv3_ib_send_cqe_handler(struct rdsv3_ib_connection *ic, ibt_wc_t *wc) 209 { 210 struct rdsv3_connection *conn = ic->conn; 211 struct rdsv3_ib_send_work *send; 212 uint32_t completed, polled; 213 uint32_t oldest; 214 uint32_t i = 0; 215 int ret; 216 217 RDSV3_DPRINTF4("rdsv3_ib_send_cqe_handler", 218 "wc wc_id 0x%llx status %u byte_len %u imm_data %u\n", 219 (unsigned long long)wc->wc_id, wc->wc_status, 220 wc->wc_bytes_xfer, ntohl(wc->wc_immed_data)); 221 222 rdsv3_ib_stats_inc(s_ib_tx_cq_event); 223 224 if (wc->wc_id == RDSV3_IB_ACK_WR_ID) { 225 if (ic->i_ack_queued + HZ/2 < jiffies) 226 rdsv3_ib_stats_inc(s_ib_tx_stalled); 227 rdsv3_ib_ack_send_complete(ic); 228 return; 229 } 230 231 oldest = rdsv3_ib_ring_oldest(&ic->i_send_ring); 232 233 completed = rdsv3_ib_ring_completed(&ic->i_send_ring, 234 (wc->wc_id & ~RDSV3_IB_SEND_OP), oldest); 235 236 for (i = 0; i < completed; i++) { 237 send = &ic->i_sends[oldest]; 238 239 /* 240 * In the error case, wc->opcode sometimes contains 241 * garbage 242 */ 243 switch (send->s_opcode) { 244 case IBT_WRC_SEND: 245 if (send->s_rm) 246 rdsv3_ib_send_unmap_rm(ic, send, 247 wc->wc_status); 248 break; 249 case IBT_WRC_RDMAW: 250 case IBT_WRC_RDMAR: 251 /* 252 * Nothing to be done - the SG list will 253 * be unmapped 254 * when the SEND completes. 255 */ 256 break; 257 default: 258 #ifndef __lock_lint 259 RDSV3_DPRINTF2("rdsv3_ib_send_cq_comp_handler", 260 "RDS/IB: %s: unexpected opcode " 261 "0x%x in WR!", 262 __func__, send->s_opcode); 263 #endif 264 break; 265 } 266 267 send->s_opcode = 0xdd; 268 if (send->s_queued + HZ/2 < jiffies) 269 rdsv3_ib_stats_inc(s_ib_tx_stalled); 270 271 /* 272 * If a RDMA operation produced an error, signal 273 * this right 274 * away. If we don't, the subsequent SEND that goes 275 * with this 276 * RDMA will be canceled with ERR_WFLUSH, and the 277 * application 278 * never learn that the RDMA failed. 279 */ 280 if (wc->wc_status == 281 IBT_WC_REMOTE_ACCESS_ERR && send->s_op) { 282 struct rdsv3_message *rm; 283 284 rm = rdsv3_send_get_message(conn, send->s_op); 285 if (rm) { 286 if (rm->m_rdma_op != NULL) 287 rdsv3_ib_send_unmap_rdma(ic, 288 rm->m_rdma_op); 289 rdsv3_ib_send_rdma_complete(rm, 290 wc->wc_status); 291 rdsv3_message_put(rm); 292 } 293 } 294 295 oldest = (oldest + 1) % ic->i_send_ring.w_nr; 296 } 297 298 rdsv3_ib_ring_free(&ic->i_send_ring, completed); 299 300 clear_bit(RDSV3_LL_SEND_FULL, &conn->c_flags); 301 302 /* We expect errors as the qp is drained during shutdown */ 303 if (wc->wc_status != IBT_WC_SUCCESS && rdsv3_conn_up(conn)) { 304 RDSV3_DPRINTF2("rdsv3_ib_send_cqe_handler", 305 "send completion on %u.%u.%u.%u " 306 "had status %u, disconnecting and reconnecting\n", 307 NIPQUAD(conn->c_faddr), wc->wc_status); 308 rdsv3_conn_drop(conn); 309 } 310 311 RDSV3_DPRINTF4("rdsv3_ib_send_cqe_handler", "Return: conn: %p", ic); 312 } 313 314 /* 315 * This is the main function for allocating credits when sending 316 * messages. 317 * 318 * Conceptually, we have two counters: 319 * - send credits: this tells us how many WRs we're allowed 320 * to submit without overruning the reciever's queue. For 321 * each SEND WR we post, we decrement this by one. 322 * 323 * - posted credits: this tells us how many WRs we recently 324 * posted to the receive queue. This value is transferred 325 * to the peer as a "credit update" in a RDS header field. 326 * Every time we transmit credits to the peer, we subtract 327 * the amount of transferred credits from this counter. 328 * 329 * It is essential that we avoid situations where both sides have 330 * exhausted their send credits, and are unable to send new credits 331 * to the peer. We achieve this by requiring that we send at least 332 * one credit update to the peer before exhausting our credits. 333 * When new credits arrive, we subtract one credit that is withheld 334 * until we've posted new buffers and are ready to transmit these 335 * credits (see rdsv3_ib_send_add_credits below). 336 * 337 * The RDS send code is essentially single-threaded; rdsv3_send_xmit 338 * grabs c_send_lock to ensure exclusive access to the send ring. 339 * However, the ACK sending code is independent and can race with 340 * message SENDs. 341 * 342 * In the send path, we need to update the counters for send credits 343 * and the counter of posted buffers atomically - when we use the 344 * last available credit, we cannot allow another thread to race us 345 * and grab the posted credits counter. Hence, we have to use a 346 * spinlock to protect the credit counter, or use atomics. 347 * 348 * Spinlocks shared between the send and the receive path are bad, 349 * because they create unnecessary delays. An early implementation 350 * using a spinlock showed a 5% degradation in throughput at some 351 * loads. 352 * 353 * This implementation avoids spinlocks completely, putting both 354 * counters into a single atomic, and updating that atomic using 355 * atomic_add (in the receive path, when receiving fresh credits), 356 * and using atomic_cmpxchg when updating the two counters. 357 */ 358 int 359 rdsv3_ib_send_grab_credits(struct rdsv3_ib_connection *ic, 360 uint32_t wanted, uint32_t *adv_credits, int need_posted) 361 { 362 unsigned int avail, posted, got = 0, advertise; 363 long oldval, newval; 364 365 RDSV3_DPRINTF4("rdsv3_ib_send_grab_credits", "ic: %p, %d %d %d", 366 ic, wanted, *adv_credits, need_posted); 367 368 *adv_credits = 0; 369 if (!ic->i_flowctl) 370 return (wanted); 371 372 try_again: 373 advertise = 0; 374 oldval = newval = atomic_get(&ic->i_credits); 375 posted = IB_GET_POST_CREDITS(oldval); 376 avail = IB_GET_SEND_CREDITS(oldval); 377 378 RDSV3_DPRINTF5("rdsv3_ib_send_grab_credits", 379 "wanted (%u): credits=%u posted=%u\n", wanted, avail, posted); 380 381 /* The last credit must be used to send a credit update. */ 382 if (avail && !posted) 383 avail--; 384 385 if (avail < wanted) { 386 struct rdsv3_connection *conn = ic->i_cm_id->context; 387 388 /* Oops, there aren't that many credits left! */ 389 set_bit(RDSV3_LL_SEND_FULL, &conn->c_flags); 390 got = avail; 391 } else { 392 /* Sometimes you get what you want, lalala. */ 393 got = wanted; 394 } 395 newval -= IB_SET_SEND_CREDITS(got); 396 397 /* 398 * If need_posted is non-zero, then the caller wants 399 * the posted regardless of whether any send credits are 400 * available. 401 */ 402 if (posted && (got || need_posted)) { 403 advertise = min(posted, RDSV3_MAX_ADV_CREDIT); 404 newval -= IB_SET_POST_CREDITS(advertise); 405 } 406 407 /* Finally bill everything */ 408 if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval) 409 goto try_again; 410 411 *adv_credits = advertise; 412 413 RDSV3_DPRINTF4("rdsv3_ib_send_grab_credits", "ic: %p, %d %d %d", 414 ic, got, *adv_credits, need_posted); 415 416 return (got); 417 } 418 419 void 420 rdsv3_ib_send_add_credits(struct rdsv3_connection *conn, unsigned int credits) 421 { 422 struct rdsv3_ib_connection *ic = conn->c_transport_data; 423 424 if (credits == 0) 425 return; 426 427 RDSV3_DPRINTF5("rdsv3_ib_send_add_credits", 428 "credits (%u): current=%u%s\n", 429 credits, 430 IB_GET_SEND_CREDITS(atomic_get(&ic->i_credits)), 431 test_bit(RDSV3_LL_SEND_FULL, &conn->c_flags) ? 432 ", ll_send_full" : ""); 433 434 atomic_add_32(&ic->i_credits, IB_SET_SEND_CREDITS(credits)); 435 if (test_and_clear_bit(RDSV3_LL_SEND_FULL, &conn->c_flags)) 436 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0); 437 438 ASSERT(!(IB_GET_SEND_CREDITS(credits) >= 16384)); 439 440 rdsv3_ib_stats_inc(s_ib_rx_credit_updates); 441 442 RDSV3_DPRINTF4("rdsv3_ib_send_add_credits", 443 "Return: conn: %p, credits: %d", 444 conn, credits); 445 } 446 447 void 448 rdsv3_ib_advertise_credits(struct rdsv3_connection *conn, unsigned int posted) 449 { 450 struct rdsv3_ib_connection *ic = conn->c_transport_data; 451 452 RDSV3_DPRINTF4("rdsv3_ib_advertise_credits", "conn: %p, posted: %d", 453 conn, posted); 454 455 if (posted == 0) 456 return; 457 458 atomic_add_32(&ic->i_credits, IB_SET_POST_CREDITS(posted)); 459 460 /* 461 * Decide whether to send an update to the peer now. 462 * If we would send a credit update for every single buffer we 463 * post, we would end up with an ACK storm (ACK arrives, 464 * consumes buffer, we refill the ring, send ACK to remote 465 * advertising the newly posted buffer... ad inf) 466 * 467 * Performance pretty much depends on how often we send 468 * credit updates - too frequent updates mean lots of ACKs. 469 * Too infrequent updates, and the peer will run out of 470 * credits and has to throttle. 471 * For the time being, 16 seems to be a good compromise. 472 */ 473 if (IB_GET_POST_CREDITS(atomic_get(&ic->i_credits)) >= 16) 474 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 475 } 476 477 static inline void 478 rdsv3_ib_xmit_populate_wr(struct rdsv3_ib_connection *ic, 479 ibt_send_wr_t *wr, unsigned int pos, 480 struct rdsv3_scatterlist *scat, unsigned int off, unsigned int length, 481 int send_flags) 482 { 483 ibt_wr_ds_t *sge; 484 485 RDSV3_DPRINTF4("rdsv3_ib_xmit_populate_wr", 486 "ic: %p, wr: %p scat: %p %d %d %d %d", 487 ic, wr, scat, pos, off, length, send_flags); 488 489 wr->wr_id = pos | RDSV3_IB_SEND_OP; 490 wr->wr_trans = IBT_RC_SRV; 491 wr->wr_flags = send_flags; 492 wr->wr_opcode = IBT_WRC_SEND; 493 494 if (length != 0) { 495 int ix, len, assigned; 496 ibt_wr_ds_t *sgl; 497 498 ASSERT(length <= scat->length - off); 499 500 sgl = scat->sgl; 501 if (off != 0) { 502 /* find the right sgl to begin with */ 503 while (sgl->ds_len <= off) { 504 off -= sgl->ds_len; 505 sgl++; 506 } 507 } 508 509 ix = 1; /* first data sgl is at 1 */ 510 assigned = 0; 511 len = length; 512 do { 513 sge = &wr->wr_sgl[ix++]; 514 sge->ds_va = sgl->ds_va + off; 515 assigned = min(len, sgl->ds_len - off); 516 sge->ds_len = assigned; 517 sge->ds_key = sgl->ds_key; 518 len -= assigned; 519 if (len != 0) { 520 sgl++; 521 off = 0; 522 } 523 } while (len > 0); 524 525 wr->wr_nds = ix; 526 } else { 527 /* 528 * We're sending a packet with no payload. There is only 529 * one SGE 530 */ 531 wr->wr_nds = 1; 532 } 533 534 sge = &wr->wr_sgl[0]; 535 sge->ds_va = ic->i_send_hdrs_dma + (pos * sizeof (struct rdsv3_header)); 536 sge->ds_len = sizeof (struct rdsv3_header); 537 sge->ds_key = ic->i_mr->lkey; 538 539 RDSV3_DPRINTF4("rdsv3_ib_xmit_populate_wr", 540 "Return: ic: %p, wr: %p scat: %p", ic, wr, scat); 541 } 542 543 /* 544 * This can be called multiple times for a given message. The first time 545 * we see a message we map its scatterlist into the IB device so that 546 * we can provide that mapped address to the IB scatter gather entries 547 * in the IB work requests. We translate the scatterlist into a series 548 * of work requests that fragment the message. These work requests complete 549 * in order so we pass ownership of the message to the completion handler 550 * once we send the final fragment. 551 * 552 * The RDS core uses the c_send_lock to only enter this function once 553 * per connection. This makes sure that the tx ring alloc/unalloc pairs 554 * don't get out of sync and confuse the ring. 555 */ 556 int 557 rdsv3_ib_xmit(struct rdsv3_connection *conn, struct rdsv3_message *rm, 558 unsigned int hdr_off, unsigned int sg, unsigned int off) 559 { 560 struct rdsv3_ib_connection *ic = conn->c_transport_data; 561 struct ib_device *dev = ic->i_cm_id->device; 562 struct rdsv3_ib_send_work *send = NULL; 563 struct rdsv3_ib_send_work *first; 564 struct rdsv3_ib_send_work *prev; 565 ibt_send_wr_t *wr; 566 struct rdsv3_scatterlist *scat; 567 uint32_t pos; 568 uint32_t i; 569 uint32_t work_alloc; 570 uint32_t credit_alloc; 571 uint32_t posted; 572 uint32_t adv_credits = 0; 573 int send_flags = 0; 574 int sent; 575 int ret; 576 int flow_controlled = 0; 577 578 RDSV3_DPRINTF4("rdsv3_ib_xmit", "conn: %p, rm: %p", conn, rm); 579 580 ASSERT(!(off % RDSV3_FRAG_SIZE)); 581 ASSERT(!(hdr_off != 0 && hdr_off != sizeof (struct rdsv3_header))); 582 583 /* Do not send cong updates to IB loopback */ 584 if (conn->c_loopback && 585 rm->m_inc.i_hdr.h_flags & RDSV3_FLAG_CONG_BITMAP) { 586 rdsv3_cong_map_updated(conn->c_fcong, ~(uint64_t)0); 587 return (sizeof (struct rdsv3_header) + RDSV3_CONG_MAP_BYTES); 588 } 589 590 #ifndef __lock_lint 591 /* FIXME we may overallocate here */ 592 if (ntohl(rm->m_inc.i_hdr.h_len) == 0) 593 i = 1; 594 else 595 i = ceil(ntohl(rm->m_inc.i_hdr.h_len), RDSV3_FRAG_SIZE); 596 #endif 597 598 work_alloc = rdsv3_ib_ring_alloc(&ic->i_send_ring, i, &pos); 599 if (work_alloc != i) { 600 rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 601 set_bit(RDSV3_LL_SEND_FULL, &conn->c_flags); 602 rdsv3_ib_stats_inc(s_ib_tx_ring_full); 603 ret = -ENOMEM; 604 goto out; 605 } 606 607 credit_alloc = work_alloc; 608 if (ic->i_flowctl) { 609 credit_alloc = rdsv3_ib_send_grab_credits(ic, work_alloc, 610 &posted, 0); 611 adv_credits += posted; 612 if (credit_alloc < work_alloc) { 613 rdsv3_ib_ring_unalloc(&ic->i_send_ring, 614 work_alloc - credit_alloc); 615 work_alloc = credit_alloc; 616 flow_controlled++; 617 } 618 if (work_alloc == 0) { 619 rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 620 rdsv3_ib_stats_inc(s_ib_tx_throttle); 621 ret = -ENOMEM; 622 goto out; 623 } 624 } 625 626 /* map the message the first time we see it */ 627 if (ic->i_rm == NULL) { 628 /* 629 * printk(KERN_NOTICE 630 * "rdsv3_ib_xmit prep msg dport=%u flags=0x%x len=%d\n", 631 * be16_to_cpu(rm->m_inc.i_hdr.h_dport), 632 * rm->m_inc.i_hdr.h_flags, 633 * be32_to_cpu(rm->m_inc.i_hdr.h_len)); 634 */ 635 if (rm->m_nents) { 636 rm->m_count = rdsv3_ib_dma_map_sg(dev, 637 rm->m_sg, rm->m_nents); 638 RDSV3_DPRINTF5("rdsv3_ib_xmit", 639 "ic %p mapping rm %p: %d\n", ic, rm, rm->m_count); 640 if (rm->m_count == 0) { 641 rdsv3_ib_stats_inc(s_ib_tx_sg_mapping_failure); 642 rdsv3_ib_ring_unalloc(&ic->i_send_ring, 643 work_alloc); 644 ret = -ENOMEM; /* XXX ? */ 645 RDSV3_DPRINTF2("rdsv3_ib_xmit", 646 "fail: ic %p mapping rm %p: %d\n", 647 ic, rm, rm->m_count); 648 goto out; 649 } 650 } else { 651 rm->m_count = 0; 652 } 653 654 ic->i_unsignaled_wrs = rdsv3_ib_sysctl_max_unsig_wrs; 655 ic->i_unsignaled_bytes = rdsv3_ib_sysctl_max_unsig_bytes; 656 rdsv3_message_addref(rm); 657 ic->i_rm = rm; 658 659 /* Finalize the header */ 660 if (test_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags)) 661 rm->m_inc.i_hdr.h_flags |= RDSV3_FLAG_ACK_REQUIRED; 662 if (test_bit(RDSV3_MSG_RETRANSMITTED, &rm->m_flags)) 663 rm->m_inc.i_hdr.h_flags |= RDSV3_FLAG_RETRANSMITTED; 664 665 /* 666 * If it has a RDMA op, tell the peer we did it. This is 667 * used by the peer to release use-once RDMA MRs. 668 */ 669 if (rm->m_rdma_op) { 670 struct rdsv3_ext_header_rdma ext_hdr; 671 672 ext_hdr.h_rdma_rkey = htonl(rm->m_rdma_op->r_key); 673 (void) rdsv3_message_add_extension(&rm->m_inc.i_hdr, 674 RDSV3_EXTHDR_RDMA, &ext_hdr, 675 sizeof (ext_hdr)); 676 } 677 if (rm->m_rdma_cookie) { 678 (void) rdsv3_message_add_rdma_dest_extension( 679 &rm->m_inc.i_hdr, 680 rdsv3_rdma_cookie_key(rm->m_rdma_cookie), 681 rdsv3_rdma_cookie_offset(rm->m_rdma_cookie)); 682 } 683 684 /* 685 * Note - rdsv3_ib_piggyb_ack clears the ACK_REQUIRED bit, so 686 * we should not do this unless we have a chance of at least 687 * sticking the header into the send ring. Which is why we 688 * should call rdsv3_ib_ring_alloc first. 689 */ 690 rm->m_inc.i_hdr.h_ack = htonll(rdsv3_ib_piggyb_ack(ic)); 691 rdsv3_message_make_checksum(&rm->m_inc.i_hdr); 692 693 /* 694 * Update adv_credits since we reset the ACK_REQUIRED bit. 695 */ 696 (void) rdsv3_ib_send_grab_credits(ic, 0, &posted, 1); 697 adv_credits += posted; 698 ASSERT(adv_credits <= 255); 699 } 700 701 send = &ic->i_sends[pos]; 702 first = send; 703 prev = NULL; 704 scat = &rm->m_sg[sg]; 705 sent = 0; 706 i = 0; 707 708 /* 709 * Sometimes you want to put a fence between an RDMA 710 * READ and the following SEND. 711 * We could either do this all the time 712 * or when requested by the user. Right now, we let 713 * the application choose. 714 */ 715 if (rm->m_rdma_op && rm->m_rdma_op->r_fence) 716 send_flags = IBT_WR_SEND_FENCE; 717 718 /* 719 * We could be copying the header into the unused tail of the page. 720 * That would need to be changed in the future when those pages might 721 * be mapped userspace pages or page cache pages. So instead we always 722 * use a second sge and our long-lived ring of mapped headers. We send 723 * the header after the data so that the data payload can be aligned on 724 * the receiver. 725 */ 726 727 /* handle a 0-len message */ 728 if (ntohl(rm->m_inc.i_hdr.h_len) == 0) { 729 wr = &ic->i_send_wrs[0]; 730 rdsv3_ib_xmit_populate_wr(ic, wr, pos, NULL, 0, 0, send_flags); 731 send->s_queued = jiffies; 732 send->s_op = NULL; 733 send->s_opcode = wr->wr_opcode; 734 goto add_header; 735 } 736 737 /* if there's data reference it with a chain of work reqs */ 738 for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) { 739 unsigned int len; 740 741 send = &ic->i_sends[pos]; 742 743 wr = &ic->i_send_wrs[i]; 744 len = min(RDSV3_FRAG_SIZE, 745 rdsv3_ib_sg_dma_len(dev, scat) - off); 746 rdsv3_ib_xmit_populate_wr(ic, wr, pos, scat, off, len, 747 send_flags); 748 send->s_queued = jiffies; 749 send->s_op = NULL; 750 send->s_opcode = wr->wr_opcode; 751 752 /* 753 * We want to delay signaling completions just enough to get 754 * the batching benefits but not so much that we create dead 755 * time 756 * on the wire. 757 */ 758 if (ic->i_unsignaled_wrs-- == 0) { 759 ic->i_unsignaled_wrs = rdsv3_ib_sysctl_max_unsig_wrs; 760 wr->wr_flags |= 761 IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT; 762 } 763 764 ic->i_unsignaled_bytes -= len; 765 if (ic->i_unsignaled_bytes <= 0) { 766 ic->i_unsignaled_bytes = 767 rdsv3_ib_sysctl_max_unsig_bytes; 768 wr->wr_flags |= 769 IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT; 770 } 771 772 /* 773 * Always signal the last one if we're stopping due to flow 774 * control. 775 */ 776 if (flow_controlled && i == (work_alloc-1)) { 777 wr->wr_flags |= 778 IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT; 779 } 780 781 RDSV3_DPRINTF5("rdsv3_ib_xmit", "send %p wr %p num_sge %u \n", 782 send, wr, wr->wr_nds); 783 784 sent += len; 785 off += len; 786 if (off == rdsv3_ib_sg_dma_len(dev, scat)) { 787 scat++; 788 off = 0; 789 } 790 791 add_header: 792 /* 793 * Tack on the header after the data. The header SGE 794 * should already 795 * have been set up to point to the right header buffer. 796 */ 797 (void) memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, 798 sizeof (struct rdsv3_header)); 799 800 if (0) { 801 struct rdsv3_header *hdr = &ic->i_send_hdrs[pos]; 802 803 RDSV3_DPRINTF2("rdsv3_ib_xmit", 804 "send WR dport=%u flags=0x%x len=%d", 805 ntohs(hdr->h_dport), 806 hdr->h_flags, 807 ntohl(hdr->h_len)); 808 } 809 if (adv_credits) { 810 struct rdsv3_header *hdr = &ic->i_send_hdrs[pos]; 811 812 /* add credit and redo the header checksum */ 813 hdr->h_credit = adv_credits; 814 rdsv3_message_make_checksum(hdr); 815 adv_credits = 0; 816 rdsv3_ib_stats_inc(s_ib_tx_credit_updates); 817 } 818 819 prev = send; 820 821 pos = (pos + 1) % ic->i_send_ring.w_nr; 822 } 823 824 /* 825 * Account the RDS header in the number of bytes we sent, but just once. 826 * The caller has no concept of fragmentation. 827 */ 828 if (hdr_off == 0) 829 sent += sizeof (struct rdsv3_header); 830 831 /* if we finished the message then send completion owns it */ 832 if (scat == &rm->m_sg[rm->m_count]) { 833 prev->s_rm = ic->i_rm; 834 wr->wr_flags |= IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT; 835 ic->i_rm = NULL; 836 } 837 838 if (i < work_alloc) { 839 rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i); 840 work_alloc = i; 841 } 842 if (ic->i_flowctl && i < credit_alloc) 843 rdsv3_ib_send_add_credits(conn, credit_alloc - i); 844 845 /* XXX need to worry about failed_wr and partial sends. */ 846 ret = ibt_post_send(ib_get_ibt_channel_hdl(ic->i_cm_id), 847 ic->i_send_wrs, i, &posted); 848 if (posted != i) { 849 RDSV3_DPRINTF2("rdsv3_ib_xmit", 850 "ic %p first %p nwr: %d ret %d:%d", 851 ic, first, i, ret, posted); 852 } 853 if (ret) { 854 RDSV3_DPRINTF2("rdsv3_ib_xmit", 855 "RDS/IB: ib_post_send to %u.%u.%u.%u " 856 "returned %d\n", NIPQUAD(conn->c_faddr), ret); 857 rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 858 if (prev->s_rm) { 859 ic->i_rm = prev->s_rm; 860 prev->s_rm = NULL; 861 } 862 RDSV3_DPRINTF2("rdsv3_ib_xmit", "ibt_post_send failed\n"); 863 rdsv3_conn_drop(ic->conn); 864 ret = -EAGAIN; 865 goto out; 866 } 867 868 ret = sent; 869 870 RDSV3_DPRINTF4("rdsv3_ib_xmit", "Return: conn: %p, rm: %p", conn, rm); 871 out: 872 ASSERT(!adv_credits); 873 return (ret); 874 } 875 876 static void 877 rdsv3_ib_dma_unmap_sg_rdma(struct ib_device *dev, uint_t num, 878 struct rdsv3_rdma_sg scat[]) 879 { 880 ibt_hca_hdl_t hca_hdl; 881 int i; 882 int num_sgl; 883 884 RDSV3_DPRINTF4("rdsv3_ib_dma_unmap_sg", "rdma_sg: %p", scat); 885 886 if (dev) { 887 hca_hdl = ib_get_ibt_hca_hdl(dev); 888 } else { 889 hca_hdl = scat[0].hca_hdl; 890 RDSV3_DPRINTF2("rdsv3_ib_dma_unmap_sg_rdma", 891 "NULL dev use cached hca_hdl %p", hca_hdl); 892 } 893 894 if (hca_hdl == NULL) 895 return; 896 scat[0].hca_hdl = NULL; 897 898 for (i = 0; i < num; i++) { 899 if (scat[i].mihdl != NULL) { 900 num_sgl = (scat[i].iovec.bytes / PAGESIZE) + 2; 901 kmem_free(scat[i].swr.wr_sgl, 902 (num_sgl * sizeof (ibt_wr_ds_t))); 903 scat[i].swr.wr_sgl = NULL; 904 (void) ibt_unmap_mem_iov(hca_hdl, scat[i].mihdl); 905 scat[i].mihdl = NULL; 906 } else 907 break; 908 } 909 } 910 911 /* ARGSUSED */ 912 uint_t 913 rdsv3_ib_dma_map_sg_rdma(struct ib_device *dev, struct rdsv3_rdma_sg scat[], 914 uint_t num, struct rdsv3_scatterlist **scatl) 915 { 916 ibt_hca_hdl_t hca_hdl; 917 ibt_iov_attr_t iov_attr; 918 struct buf *bp; 919 uint_t i, j, k; 920 uint_t count; 921 struct rdsv3_scatterlist *sg; 922 int ret; 923 924 RDSV3_DPRINTF4("rdsv3_ib_dma_map_sg_rdma", "scat: %p, num: %d", 925 scat, num); 926 927 hca_hdl = ib_get_ibt_hca_hdl(dev); 928 scat[0].hca_hdl = hca_hdl; 929 bzero(&iov_attr, sizeof (ibt_iov_attr_t)); 930 iov_attr.iov_flags = IBT_IOV_BUF; 931 iov_attr.iov_lso_hdr_sz = 0; 932 933 for (i = 0, count = 0; i < num; i++) { 934 /* transpose umem_cookie to buf structure */ 935 bp = ddi_umem_iosetup(scat[i].umem_cookie, 936 scat[i].iovec.addr & PAGEOFFSET, scat[i].iovec.bytes, 937 B_WRITE, 0, 0, NULL, DDI_UMEM_SLEEP); 938 if (bp == NULL) { 939 /* free resources and return error */ 940 goto out; 941 } 942 /* setup ibt_map_mem_iov() attributes */ 943 iov_attr.iov_buf = bp; 944 iov_attr.iov_wr_nds = (scat[i].iovec.bytes / PAGESIZE) + 2; 945 scat[i].swr.wr_sgl = 946 kmem_zalloc(iov_attr.iov_wr_nds * sizeof (ibt_wr_ds_t), 947 KM_SLEEP); 948 949 ret = ibt_map_mem_iov(hca_hdl, &iov_attr, 950 (ibt_all_wr_t *)&scat[i].swr, &scat[i].mihdl); 951 freerbuf(bp); 952 if (ret != IBT_SUCCESS) { 953 RDSV3_DPRINTF2("rdsv3_ib_dma_map_sg_rdma", 954 "ibt_map_mem_iov returned: %d", ret); 955 /* free resources and return error */ 956 kmem_free(scat[i].swr.wr_sgl, 957 iov_attr.iov_wr_nds * sizeof (ibt_wr_ds_t)); 958 goto out; 959 } 960 count += scat[i].swr.wr_nds; 961 962 #ifdef DEBUG 963 for (j = 0; j < scat[i].swr.wr_nds; j++) { 964 RDSV3_DPRINTF5("rdsv3_ib_dma_map_sg_rdma", 965 "sgl[%d] va %llx len %x", j, 966 scat[i].swr.wr_sgl[j].ds_va, 967 scat[i].swr.wr_sgl[j].ds_len); 968 } 969 #endif 970 RDSV3_DPRINTF4("rdsv3_ib_dma_map_sg_rdma", 971 "iovec.bytes: 0x%x scat[%d]swr.wr_nds: %d", 972 scat[i].iovec.bytes, i, scat[i].swr.wr_nds); 973 } 974 975 count = ((count - 1) / RDSV3_IB_MAX_SGE) + 1; 976 RDSV3_DPRINTF4("rdsv3_ib_dma_map_sg_rdma", "Ret: num: %d", count); 977 return (count); 978 979 out: 980 rdsv3_ib_dma_unmap_sg_rdma(dev, num, scat); 981 return (0); 982 } 983 984 int 985 rdsv3_ib_xmit_rdma(struct rdsv3_connection *conn, struct rdsv3_rdma_op *op) 986 { 987 struct rdsv3_ib_connection *ic = conn->c_transport_data; 988 struct rdsv3_ib_send_work *send = NULL; 989 struct rdsv3_rdma_sg *scat; 990 uint64_t remote_addr; 991 uint32_t pos; 992 uint32_t work_alloc; 993 uint32_t i, j, k, idx; 994 uint32_t left, count; 995 uint32_t posted; 996 int sent; 997 ibt_status_t status; 998 ibt_send_wr_t *wr; 999 ibt_wr_ds_t *sge; 1000 1001 RDSV3_DPRINTF4("rdsv3_ib_xmit_rdma", "rdsv3_ib_conn: %p", ic); 1002 1003 /* map the message the first time we see it */ 1004 if (!op->r_mapped) { 1005 op->r_count = rdsv3_ib_dma_map_sg_rdma(ic->i_cm_id->device, 1006 op->r_rdma_sg, op->r_nents, &op->r_sg); 1007 RDSV3_DPRINTF5("rdsv3_ib_xmit_rdma", "ic %p mapping op %p: %d", 1008 ic, op, op->r_count); 1009 if (op->r_count == 0) { 1010 rdsv3_ib_stats_inc(s_ib_tx_sg_mapping_failure); 1011 RDSV3_DPRINTF2("rdsv3_ib_xmit_rdma", 1012 "fail: ic %p mapping op %p: %d", 1013 ic, op, op->r_count); 1014 return (-ENOMEM); /* XXX ? */ 1015 } 1016 op->r_mapped = 1; 1017 } 1018 1019 /* 1020 * Instead of knowing how to return a partial rdma read/write 1021 * we insist that there 1022 * be enough work requests to send the entire message. 1023 */ 1024 work_alloc = rdsv3_ib_ring_alloc(&ic->i_send_ring, op->r_count, &pos); 1025 if (work_alloc != op->r_count) { 1026 rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 1027 rdsv3_ib_stats_inc(s_ib_tx_ring_full); 1028 return (-ENOMEM); 1029 } 1030 1031 RDSV3_DPRINTF4("rdsv3_ib_xmit_rdma", "pos %u cnt %u", pos, op->r_count); 1032 /* 1033 * take the scatter list and transpose into a list of 1034 * send wr's each with a scatter list of RDSV3_IB_MAX_SGE 1035 */ 1036 scat = &op->r_rdma_sg[0]; 1037 sent = 0; 1038 remote_addr = op->r_remote_addr; 1039 1040 for (i = 0, k = 0; i < op->r_nents; i++) { 1041 left = scat[i].swr.wr_nds; 1042 for (idx = 0; left > 0; k++) { 1043 send = &ic->i_sends[pos]; 1044 send->s_queued = jiffies; 1045 send->s_opcode = op->r_write ? IBT_WRC_RDMAW : 1046 IBT_WRC_RDMAR; 1047 send->s_op = op; 1048 1049 wr = &ic->i_send_wrs[k]; 1050 wr->wr_flags = 0; 1051 wr->wr_id = pos | RDSV3_IB_SEND_OP; 1052 wr->wr_trans = IBT_RC_SRV; 1053 wr->wr_opcode = op->r_write ? IBT_WRC_RDMAW : 1054 IBT_WRC_RDMAR; 1055 wr->wr.rc.rcwr.rdma.rdma_raddr = remote_addr; 1056 wr->wr.rc.rcwr.rdma.rdma_rkey = op->r_key; 1057 1058 if (left > RDSV3_IB_MAX_SGE) { 1059 count = RDSV3_IB_MAX_SGE; 1060 left -= RDSV3_IB_MAX_SGE; 1061 } else { 1062 count = left; 1063 left = 0; 1064 } 1065 wr->wr_nds = count; 1066 1067 for (j = 0; j < count; j++) { 1068 sge = &wr->wr_sgl[j]; 1069 *sge = scat[i].swr.wr_sgl[idx]; 1070 remote_addr += scat[i].swr.wr_sgl[idx].ds_len; 1071 sent += scat[i].swr.wr_sgl[idx].ds_len; 1072 idx++; 1073 RDSV3_DPRINTF5("xmit_rdma", 1074 "send_wrs[%d]sgl[%d] va %llx len %x", 1075 k, j, sge->ds_va, sge->ds_len); 1076 } 1077 RDSV3_DPRINTF5("rdsv3_ib_xmit_rdma", 1078 "wr[%d] %p key: %x code: %d tlen: %d", 1079 k, wr, wr->wr.rc.rcwr.rdma.rdma_rkey, 1080 wr->wr_opcode, sent); 1081 1082 /* 1083 * We want to delay signaling completions just enough 1084 * to get the batching benefits but not so much that 1085 * we create dead time on the wire. 1086 */ 1087 if (ic->i_unsignaled_wrs-- == 0) { 1088 ic->i_unsignaled_wrs = 1089 rdsv3_ib_sysctl_max_unsig_wrs; 1090 wr->wr_flags = IBT_WR_SEND_SIGNAL; 1091 } 1092 1093 pos = (pos + 1) % ic->i_send_ring.w_nr; 1094 } 1095 } 1096 1097 status = ibt_post_send(ib_get_ibt_channel_hdl(ic->i_cm_id), 1098 ic->i_send_wrs, k, &posted); 1099 if (status != IBT_SUCCESS) { 1100 RDSV3_DPRINTF2("rdsv3_ib_xmit_rdma", 1101 "RDS/IB: rdma ib_post_send to %u.%u.%u.%u " 1102 "returned %d", NIPQUAD(conn->c_faddr), status); 1103 rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 1104 } 1105 RDSV3_DPRINTF4("rdsv3_ib_xmit_rdma", "Ret: %p", ic); 1106 return (status); 1107 } 1108 1109 void 1110 rdsv3_ib_xmit_complete(struct rdsv3_connection *conn) 1111 { 1112 struct rdsv3_ib_connection *ic = conn->c_transport_data; 1113 1114 RDSV3_DPRINTF4("rdsv3_ib_xmit_complete", "conn: %p", conn); 1115 1116 /* 1117 * We may have a pending ACK or window update we were unable 1118 * to send previously (due to flow control). Try again. 1119 */ 1120 rdsv3_ib_attempt_ack(ic); 1121 } 1122