1 /* 2 * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation. 3 * All rights reserved. 4 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * OpenIB.org BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include <linux/err.h> 36 #include <linux/vmalloc.h> 37 #include <linux/jhash.h> 38 39 #include "qib.h" 40 41 #define BITS_PER_PAGE (PAGE_SIZE*BITS_PER_BYTE) 42 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1) 43 44 static inline unsigned mk_qpn(struct qib_qpn_table *qpt, 45 struct qpn_map *map, unsigned off) 46 { 47 return (map - qpt->map) * BITS_PER_PAGE + off; 48 } 49 50 static inline unsigned find_next_offset(struct qib_qpn_table *qpt, 51 struct qpn_map *map, unsigned off, 52 unsigned n) 53 { 54 if (qpt->mask) { 55 off++; 56 if (((off & qpt->mask) >> 1) >= n) 57 off = (off | qpt->mask) + 2; 58 } else 59 off = find_next_zero_bit(map->page, BITS_PER_PAGE, off); 60 return off; 61 } 62 63 /* 64 * Convert the AETH credit code into the number of credits. 65 */ 66 static u32 credit_table[31] = { 67 0, /* 0 */ 68 1, /* 1 */ 69 2, /* 2 */ 70 3, /* 3 */ 71 4, /* 4 */ 72 6, /* 5 */ 73 8, /* 6 */ 74 12, /* 7 */ 75 16, /* 8 */ 76 24, /* 9 */ 77 32, /* A */ 78 48, /* B */ 79 64, /* C */ 80 96, /* D */ 81 128, /* E */ 82 192, /* F */ 83 256, /* 10 */ 84 384, /* 11 */ 85 512, /* 12 */ 86 768, /* 13 */ 87 1024, /* 14 */ 88 1536, /* 15 */ 89 2048, /* 16 */ 90 3072, /* 17 */ 91 4096, /* 18 */ 92 6144, /* 19 */ 93 8192, /* 1A */ 94 12288, /* 1B */ 95 16384, /* 1C */ 96 24576, /* 1D */ 97 32768 /* 1E */ 98 }; 99 100 static void get_map_page(struct qib_qpn_table *qpt, struct qpn_map *map) 101 { 102 unsigned long page = get_zeroed_page(GFP_KERNEL); 103 104 /* 105 * Free the page if someone raced with us installing it. 106 */ 107 108 spin_lock(&qpt->lock); 109 if (map->page) 110 free_page(page); 111 else 112 map->page = (void *)page; 113 spin_unlock(&qpt->lock); 114 } 115 116 /* 117 * Allocate the next available QPN or 118 * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI. 119 */ 120 static int alloc_qpn(struct qib_devdata *dd, struct qib_qpn_table *qpt, 121 enum ib_qp_type type, u8 port) 122 { 123 u32 i, offset, max_scan, qpn; 124 struct qpn_map *map; 125 u32 ret; 126 127 if (type == IB_QPT_SMI || type == IB_QPT_GSI) { 128 unsigned n; 129 130 ret = type == IB_QPT_GSI; 131 n = 1 << (ret + 2 * (port - 1)); 132 spin_lock(&qpt->lock); 133 if (qpt->flags & n) 134 ret = -EINVAL; 135 else 136 qpt->flags |= n; 137 spin_unlock(&qpt->lock); 138 goto bail; 139 } 140 141 qpn = qpt->last + 2; 142 if (qpn >= QPN_MAX) 143 qpn = 2; 144 if (qpt->mask && ((qpn & qpt->mask) >> 1) >= dd->n_krcv_queues) 145 qpn = (qpn | qpt->mask) + 2; 146 offset = qpn & BITS_PER_PAGE_MASK; 147 map = &qpt->map[qpn / BITS_PER_PAGE]; 148 max_scan = qpt->nmaps - !offset; 149 for (i = 0;;) { 150 if (unlikely(!map->page)) { 151 get_map_page(qpt, map); 152 if (unlikely(!map->page)) 153 break; 154 } 155 do { 156 if (!test_and_set_bit(offset, map->page)) { 157 qpt->last = qpn; 158 ret = qpn; 159 goto bail; 160 } 161 offset = find_next_offset(qpt, map, offset, 162 dd->n_krcv_queues); 163 qpn = mk_qpn(qpt, map, offset); 164 /* 165 * This test differs from alloc_pidmap(). 166 * If find_next_offset() does find a zero 167 * bit, we don't need to check for QPN 168 * wrapping around past our starting QPN. 169 * We just need to be sure we don't loop 170 * forever. 171 */ 172 } while (offset < BITS_PER_PAGE && qpn < QPN_MAX); 173 /* 174 * In order to keep the number of pages allocated to a 175 * minimum, we scan the all existing pages before increasing 176 * the size of the bitmap table. 177 */ 178 if (++i > max_scan) { 179 if (qpt->nmaps == QPNMAP_ENTRIES) 180 break; 181 map = &qpt->map[qpt->nmaps++]; 182 offset = 0; 183 } else if (map < &qpt->map[qpt->nmaps]) { 184 ++map; 185 offset = 0; 186 } else { 187 map = &qpt->map[0]; 188 offset = 2; 189 } 190 qpn = mk_qpn(qpt, map, offset); 191 } 192 193 ret = -ENOMEM; 194 195 bail: 196 return ret; 197 } 198 199 static void free_qpn(struct qib_qpn_table *qpt, u32 qpn) 200 { 201 struct qpn_map *map; 202 203 map = qpt->map + qpn / BITS_PER_PAGE; 204 if (map->page) 205 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page); 206 } 207 208 static inline unsigned qpn_hash(struct qib_ibdev *dev, u32 qpn) 209 { 210 return jhash_1word(qpn, dev->qp_rnd) & 211 (dev->qp_table_size - 1); 212 } 213 214 215 /* 216 * Put the QP into the hash table. 217 * The hash table holds a reference to the QP. 218 */ 219 static void insert_qp(struct qib_ibdev *dev, struct qib_qp *qp) 220 { 221 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 222 unsigned long flags; 223 unsigned n = qpn_hash(dev, qp->ibqp.qp_num); 224 225 spin_lock_irqsave(&dev->qpt_lock, flags); 226 atomic_inc(&qp->refcount); 227 228 if (qp->ibqp.qp_num == 0) 229 rcu_assign_pointer(ibp->qp0, qp); 230 else if (qp->ibqp.qp_num == 1) 231 rcu_assign_pointer(ibp->qp1, qp); 232 else { 233 qp->next = dev->qp_table[n]; 234 rcu_assign_pointer(dev->qp_table[n], qp); 235 } 236 237 spin_unlock_irqrestore(&dev->qpt_lock, flags); 238 synchronize_rcu(); 239 } 240 241 /* 242 * Remove the QP from the table so it can't be found asynchronously by 243 * the receive interrupt routine. 244 */ 245 static void remove_qp(struct qib_ibdev *dev, struct qib_qp *qp) 246 { 247 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 248 unsigned n = qpn_hash(dev, qp->ibqp.qp_num); 249 unsigned long flags; 250 251 spin_lock_irqsave(&dev->qpt_lock, flags); 252 253 if (ibp->qp0 == qp) { 254 atomic_dec(&qp->refcount); 255 rcu_assign_pointer(ibp->qp0, NULL); 256 } else if (ibp->qp1 == qp) { 257 atomic_dec(&qp->refcount); 258 rcu_assign_pointer(ibp->qp1, NULL); 259 } else { 260 struct qib_qp *q, **qpp; 261 262 qpp = &dev->qp_table[n]; 263 for (; (q = *qpp) != NULL; qpp = &q->next) 264 if (q == qp) { 265 atomic_dec(&qp->refcount); 266 rcu_assign_pointer(*qpp, qp->next); 267 qp->next = NULL; 268 break; 269 } 270 } 271 272 spin_unlock_irqrestore(&dev->qpt_lock, flags); 273 synchronize_rcu(); 274 } 275 276 /** 277 * qib_free_all_qps - check for QPs still in use 278 * @qpt: the QP table to empty 279 * 280 * There should not be any QPs still in use. 281 * Free memory for table. 282 */ 283 unsigned qib_free_all_qps(struct qib_devdata *dd) 284 { 285 struct qib_ibdev *dev = &dd->verbs_dev; 286 unsigned long flags; 287 struct qib_qp *qp; 288 unsigned n, qp_inuse = 0; 289 290 for (n = 0; n < dd->num_pports; n++) { 291 struct qib_ibport *ibp = &dd->pport[n].ibport_data; 292 293 if (!qib_mcast_tree_empty(ibp)) 294 qp_inuse++; 295 rcu_read_lock(); 296 if (rcu_dereference(ibp->qp0)) 297 qp_inuse++; 298 if (rcu_dereference(ibp->qp1)) 299 qp_inuse++; 300 rcu_read_unlock(); 301 } 302 303 spin_lock_irqsave(&dev->qpt_lock, flags); 304 for (n = 0; n < dev->qp_table_size; n++) { 305 qp = dev->qp_table[n]; 306 rcu_assign_pointer(dev->qp_table[n], NULL); 307 308 for (; qp; qp = qp->next) 309 qp_inuse++; 310 } 311 spin_unlock_irqrestore(&dev->qpt_lock, flags); 312 synchronize_rcu(); 313 314 return qp_inuse; 315 } 316 317 /** 318 * qib_lookup_qpn - return the QP with the given QPN 319 * @qpt: the QP table 320 * @qpn: the QP number to look up 321 * 322 * The caller is responsible for decrementing the QP reference count 323 * when done. 324 */ 325 struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn) 326 { 327 struct qib_qp *qp = NULL; 328 329 if (unlikely(qpn <= 1)) { 330 rcu_read_lock(); 331 if (qpn == 0) 332 qp = rcu_dereference(ibp->qp0); 333 else 334 qp = rcu_dereference(ibp->qp1); 335 } else { 336 struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev; 337 unsigned n = qpn_hash(dev, qpn); 338 339 rcu_read_lock(); 340 for (qp = dev->qp_table[n]; rcu_dereference(qp); qp = qp->next) 341 if (qp->ibqp.qp_num == qpn) 342 break; 343 } 344 if (qp) 345 if (unlikely(!atomic_inc_not_zero(&qp->refcount))) 346 qp = NULL; 347 348 rcu_read_unlock(); 349 return qp; 350 } 351 352 /** 353 * qib_reset_qp - initialize the QP state to the reset state 354 * @qp: the QP to reset 355 * @type: the QP type 356 */ 357 static void qib_reset_qp(struct qib_qp *qp, enum ib_qp_type type) 358 { 359 qp->remote_qpn = 0; 360 qp->qkey = 0; 361 qp->qp_access_flags = 0; 362 atomic_set(&qp->s_dma_busy, 0); 363 qp->s_flags &= QIB_S_SIGNAL_REQ_WR; 364 qp->s_hdrwords = 0; 365 qp->s_wqe = NULL; 366 qp->s_draining = 0; 367 qp->s_next_psn = 0; 368 qp->s_last_psn = 0; 369 qp->s_sending_psn = 0; 370 qp->s_sending_hpsn = 0; 371 qp->s_psn = 0; 372 qp->r_psn = 0; 373 qp->r_msn = 0; 374 if (type == IB_QPT_RC) { 375 qp->s_state = IB_OPCODE_RC_SEND_LAST; 376 qp->r_state = IB_OPCODE_RC_SEND_LAST; 377 } else { 378 qp->s_state = IB_OPCODE_UC_SEND_LAST; 379 qp->r_state = IB_OPCODE_UC_SEND_LAST; 380 } 381 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 382 qp->r_nak_state = 0; 383 qp->r_aflags = 0; 384 qp->r_flags = 0; 385 qp->s_head = 0; 386 qp->s_tail = 0; 387 qp->s_cur = 0; 388 qp->s_acked = 0; 389 qp->s_last = 0; 390 qp->s_ssn = 1; 391 qp->s_lsn = 0; 392 qp->s_mig_state = IB_MIG_MIGRATED; 393 memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue)); 394 qp->r_head_ack_queue = 0; 395 qp->s_tail_ack_queue = 0; 396 qp->s_num_rd_atomic = 0; 397 if (qp->r_rq.wq) { 398 qp->r_rq.wq->head = 0; 399 qp->r_rq.wq->tail = 0; 400 } 401 qp->r_sge.num_sge = 0; 402 } 403 404 static void clear_mr_refs(struct qib_qp *qp, int clr_sends) 405 { 406 unsigned n; 407 408 if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags)) 409 while (qp->s_rdma_read_sge.num_sge) { 410 atomic_dec(&qp->s_rdma_read_sge.sge.mr->refcount); 411 if (--qp->s_rdma_read_sge.num_sge) 412 qp->s_rdma_read_sge.sge = 413 *qp->s_rdma_read_sge.sg_list++; 414 } 415 416 while (qp->r_sge.num_sge) { 417 atomic_dec(&qp->r_sge.sge.mr->refcount); 418 if (--qp->r_sge.num_sge) 419 qp->r_sge.sge = *qp->r_sge.sg_list++; 420 } 421 422 if (clr_sends) { 423 while (qp->s_last != qp->s_head) { 424 struct qib_swqe *wqe = get_swqe_ptr(qp, qp->s_last); 425 unsigned i; 426 427 for (i = 0; i < wqe->wr.num_sge; i++) { 428 struct qib_sge *sge = &wqe->sg_list[i]; 429 430 atomic_dec(&sge->mr->refcount); 431 } 432 if (qp->ibqp.qp_type == IB_QPT_UD || 433 qp->ibqp.qp_type == IB_QPT_SMI || 434 qp->ibqp.qp_type == IB_QPT_GSI) 435 atomic_dec(&to_iah(wqe->wr.wr.ud.ah)->refcount); 436 if (++qp->s_last >= qp->s_size) 437 qp->s_last = 0; 438 } 439 if (qp->s_rdma_mr) { 440 atomic_dec(&qp->s_rdma_mr->refcount); 441 qp->s_rdma_mr = NULL; 442 } 443 } 444 445 if (qp->ibqp.qp_type != IB_QPT_RC) 446 return; 447 448 for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) { 449 struct qib_ack_entry *e = &qp->s_ack_queue[n]; 450 451 if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST && 452 e->rdma_sge.mr) { 453 atomic_dec(&e->rdma_sge.mr->refcount); 454 e->rdma_sge.mr = NULL; 455 } 456 } 457 } 458 459 /** 460 * qib_error_qp - put a QP into the error state 461 * @qp: the QP to put into the error state 462 * @err: the receive completion error to signal if a RWQE is active 463 * 464 * Flushes both send and receive work queues. 465 * Returns true if last WQE event should be generated. 466 * The QP r_lock and s_lock should be held and interrupts disabled. 467 * If we are already in error state, just return. 468 */ 469 int qib_error_qp(struct qib_qp *qp, enum ib_wc_status err) 470 { 471 struct qib_ibdev *dev = to_idev(qp->ibqp.device); 472 struct ib_wc wc; 473 int ret = 0; 474 475 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) 476 goto bail; 477 478 qp->state = IB_QPS_ERR; 479 480 if (qp->s_flags & (QIB_S_TIMER | QIB_S_WAIT_RNR)) { 481 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_WAIT_RNR); 482 del_timer(&qp->s_timer); 483 } 484 485 if (qp->s_flags & QIB_S_ANY_WAIT_SEND) 486 qp->s_flags &= ~QIB_S_ANY_WAIT_SEND; 487 488 spin_lock(&dev->pending_lock); 489 if (!list_empty(&qp->iowait) && !(qp->s_flags & QIB_S_BUSY)) { 490 qp->s_flags &= ~QIB_S_ANY_WAIT_IO; 491 list_del_init(&qp->iowait); 492 } 493 spin_unlock(&dev->pending_lock); 494 495 if (!(qp->s_flags & QIB_S_BUSY)) { 496 qp->s_hdrwords = 0; 497 if (qp->s_rdma_mr) { 498 atomic_dec(&qp->s_rdma_mr->refcount); 499 qp->s_rdma_mr = NULL; 500 } 501 if (qp->s_tx) { 502 qib_put_txreq(qp->s_tx); 503 qp->s_tx = NULL; 504 } 505 } 506 507 /* Schedule the sending tasklet to drain the send work queue. */ 508 if (qp->s_last != qp->s_head) 509 qib_schedule_send(qp); 510 511 clear_mr_refs(qp, 0); 512 513 memset(&wc, 0, sizeof(wc)); 514 wc.qp = &qp->ibqp; 515 wc.opcode = IB_WC_RECV; 516 517 if (test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags)) { 518 wc.wr_id = qp->r_wr_id; 519 wc.status = err; 520 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1); 521 } 522 wc.status = IB_WC_WR_FLUSH_ERR; 523 524 if (qp->r_rq.wq) { 525 struct qib_rwq *wq; 526 u32 head; 527 u32 tail; 528 529 spin_lock(&qp->r_rq.lock); 530 531 /* sanity check pointers before trusting them */ 532 wq = qp->r_rq.wq; 533 head = wq->head; 534 if (head >= qp->r_rq.size) 535 head = 0; 536 tail = wq->tail; 537 if (tail >= qp->r_rq.size) 538 tail = 0; 539 while (tail != head) { 540 wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id; 541 if (++tail >= qp->r_rq.size) 542 tail = 0; 543 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1); 544 } 545 wq->tail = tail; 546 547 spin_unlock(&qp->r_rq.lock); 548 } else if (qp->ibqp.event_handler) 549 ret = 1; 550 551 bail: 552 return ret; 553 } 554 555 /** 556 * qib_modify_qp - modify the attributes of a queue pair 557 * @ibqp: the queue pair who's attributes we're modifying 558 * @attr: the new attributes 559 * @attr_mask: the mask of attributes to modify 560 * @udata: user data for libibverbs.so 561 * 562 * Returns 0 on success, otherwise returns an errno. 563 */ 564 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 565 int attr_mask, struct ib_udata *udata) 566 { 567 struct qib_ibdev *dev = to_idev(ibqp->device); 568 struct qib_qp *qp = to_iqp(ibqp); 569 enum ib_qp_state cur_state, new_state; 570 struct ib_event ev; 571 int lastwqe = 0; 572 int mig = 0; 573 int ret; 574 u32 pmtu = 0; /* for gcc warning only */ 575 576 spin_lock_irq(&qp->r_lock); 577 spin_lock(&qp->s_lock); 578 579 cur_state = attr_mask & IB_QP_CUR_STATE ? 580 attr->cur_qp_state : qp->state; 581 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 582 583 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 584 attr_mask)) 585 goto inval; 586 587 if (attr_mask & IB_QP_AV) { 588 if (attr->ah_attr.dlid >= QIB_MULTICAST_LID_BASE) 589 goto inval; 590 if (qib_check_ah(qp->ibqp.device, &attr->ah_attr)) 591 goto inval; 592 } 593 594 if (attr_mask & IB_QP_ALT_PATH) { 595 if (attr->alt_ah_attr.dlid >= QIB_MULTICAST_LID_BASE) 596 goto inval; 597 if (qib_check_ah(qp->ibqp.device, &attr->alt_ah_attr)) 598 goto inval; 599 if (attr->alt_pkey_index >= qib_get_npkeys(dd_from_dev(dev))) 600 goto inval; 601 } 602 603 if (attr_mask & IB_QP_PKEY_INDEX) 604 if (attr->pkey_index >= qib_get_npkeys(dd_from_dev(dev))) 605 goto inval; 606 607 if (attr_mask & IB_QP_MIN_RNR_TIMER) 608 if (attr->min_rnr_timer > 31) 609 goto inval; 610 611 if (attr_mask & IB_QP_PORT) 612 if (qp->ibqp.qp_type == IB_QPT_SMI || 613 qp->ibqp.qp_type == IB_QPT_GSI || 614 attr->port_num == 0 || 615 attr->port_num > ibqp->device->phys_port_cnt) 616 goto inval; 617 618 if (attr_mask & IB_QP_DEST_QPN) 619 if (attr->dest_qp_num > QIB_QPN_MASK) 620 goto inval; 621 622 if (attr_mask & IB_QP_RETRY_CNT) 623 if (attr->retry_cnt > 7) 624 goto inval; 625 626 if (attr_mask & IB_QP_RNR_RETRY) 627 if (attr->rnr_retry > 7) 628 goto inval; 629 630 /* 631 * Don't allow invalid path_mtu values. OK to set greater 632 * than the active mtu (or even the max_cap, if we have tuned 633 * that to a small mtu. We'll set qp->path_mtu 634 * to the lesser of requested attribute mtu and active, 635 * for packetizing messages. 636 * Note that the QP port has to be set in INIT and MTU in RTR. 637 */ 638 if (attr_mask & IB_QP_PATH_MTU) { 639 struct qib_devdata *dd = dd_from_dev(dev); 640 int mtu, pidx = qp->port_num - 1; 641 642 mtu = ib_mtu_enum_to_int(attr->path_mtu); 643 if (mtu == -1) 644 goto inval; 645 if (mtu > dd->pport[pidx].ibmtu) { 646 switch (dd->pport[pidx].ibmtu) { 647 case 4096: 648 pmtu = IB_MTU_4096; 649 break; 650 case 2048: 651 pmtu = IB_MTU_2048; 652 break; 653 case 1024: 654 pmtu = IB_MTU_1024; 655 break; 656 case 512: 657 pmtu = IB_MTU_512; 658 break; 659 case 256: 660 pmtu = IB_MTU_256; 661 break; 662 default: 663 pmtu = IB_MTU_2048; 664 } 665 } else 666 pmtu = attr->path_mtu; 667 } 668 669 if (attr_mask & IB_QP_PATH_MIG_STATE) { 670 if (attr->path_mig_state == IB_MIG_REARM) { 671 if (qp->s_mig_state == IB_MIG_ARMED) 672 goto inval; 673 if (new_state != IB_QPS_RTS) 674 goto inval; 675 } else if (attr->path_mig_state == IB_MIG_MIGRATED) { 676 if (qp->s_mig_state == IB_MIG_REARM) 677 goto inval; 678 if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD) 679 goto inval; 680 if (qp->s_mig_state == IB_MIG_ARMED) 681 mig = 1; 682 } else 683 goto inval; 684 } 685 686 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 687 if (attr->max_dest_rd_atomic > QIB_MAX_RDMA_ATOMIC) 688 goto inval; 689 690 switch (new_state) { 691 case IB_QPS_RESET: 692 if (qp->state != IB_QPS_RESET) { 693 qp->state = IB_QPS_RESET; 694 spin_lock(&dev->pending_lock); 695 if (!list_empty(&qp->iowait)) 696 list_del_init(&qp->iowait); 697 spin_unlock(&dev->pending_lock); 698 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT); 699 spin_unlock(&qp->s_lock); 700 spin_unlock_irq(&qp->r_lock); 701 /* Stop the sending work queue and retry timer */ 702 cancel_work_sync(&qp->s_work); 703 del_timer_sync(&qp->s_timer); 704 wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy)); 705 if (qp->s_tx) { 706 qib_put_txreq(qp->s_tx); 707 qp->s_tx = NULL; 708 } 709 remove_qp(dev, qp); 710 wait_event(qp->wait, !atomic_read(&qp->refcount)); 711 spin_lock_irq(&qp->r_lock); 712 spin_lock(&qp->s_lock); 713 clear_mr_refs(qp, 1); 714 qib_reset_qp(qp, ibqp->qp_type); 715 } 716 break; 717 718 case IB_QPS_RTR: 719 /* Allow event to retrigger if QP set to RTR more than once */ 720 qp->r_flags &= ~QIB_R_COMM_EST; 721 qp->state = new_state; 722 break; 723 724 case IB_QPS_SQD: 725 qp->s_draining = qp->s_last != qp->s_cur; 726 qp->state = new_state; 727 break; 728 729 case IB_QPS_SQE: 730 if (qp->ibqp.qp_type == IB_QPT_RC) 731 goto inval; 732 qp->state = new_state; 733 break; 734 735 case IB_QPS_ERR: 736 lastwqe = qib_error_qp(qp, IB_WC_WR_FLUSH_ERR); 737 break; 738 739 default: 740 qp->state = new_state; 741 break; 742 } 743 744 if (attr_mask & IB_QP_PKEY_INDEX) 745 qp->s_pkey_index = attr->pkey_index; 746 747 if (attr_mask & IB_QP_PORT) 748 qp->port_num = attr->port_num; 749 750 if (attr_mask & IB_QP_DEST_QPN) 751 qp->remote_qpn = attr->dest_qp_num; 752 753 if (attr_mask & IB_QP_SQ_PSN) { 754 qp->s_next_psn = attr->sq_psn & QIB_PSN_MASK; 755 qp->s_psn = qp->s_next_psn; 756 qp->s_sending_psn = qp->s_next_psn; 757 qp->s_last_psn = qp->s_next_psn - 1; 758 qp->s_sending_hpsn = qp->s_last_psn; 759 } 760 761 if (attr_mask & IB_QP_RQ_PSN) 762 qp->r_psn = attr->rq_psn & QIB_PSN_MASK; 763 764 if (attr_mask & IB_QP_ACCESS_FLAGS) 765 qp->qp_access_flags = attr->qp_access_flags; 766 767 if (attr_mask & IB_QP_AV) { 768 qp->remote_ah_attr = attr->ah_attr; 769 qp->s_srate = attr->ah_attr.static_rate; 770 } 771 772 if (attr_mask & IB_QP_ALT_PATH) { 773 qp->alt_ah_attr = attr->alt_ah_attr; 774 qp->s_alt_pkey_index = attr->alt_pkey_index; 775 } 776 777 if (attr_mask & IB_QP_PATH_MIG_STATE) { 778 qp->s_mig_state = attr->path_mig_state; 779 if (mig) { 780 qp->remote_ah_attr = qp->alt_ah_attr; 781 qp->port_num = qp->alt_ah_attr.port_num; 782 qp->s_pkey_index = qp->s_alt_pkey_index; 783 } 784 } 785 786 if (attr_mask & IB_QP_PATH_MTU) { 787 qp->path_mtu = pmtu; 788 qp->pmtu = ib_mtu_enum_to_int(pmtu); 789 } 790 791 if (attr_mask & IB_QP_RETRY_CNT) { 792 qp->s_retry_cnt = attr->retry_cnt; 793 qp->s_retry = attr->retry_cnt; 794 } 795 796 if (attr_mask & IB_QP_RNR_RETRY) { 797 qp->s_rnr_retry_cnt = attr->rnr_retry; 798 qp->s_rnr_retry = attr->rnr_retry; 799 } 800 801 if (attr_mask & IB_QP_MIN_RNR_TIMER) 802 qp->r_min_rnr_timer = attr->min_rnr_timer; 803 804 if (attr_mask & IB_QP_TIMEOUT) 805 qp->timeout = attr->timeout; 806 807 if (attr_mask & IB_QP_QKEY) 808 qp->qkey = attr->qkey; 809 810 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 811 qp->r_max_rd_atomic = attr->max_dest_rd_atomic; 812 813 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) 814 qp->s_max_rd_atomic = attr->max_rd_atomic; 815 816 spin_unlock(&qp->s_lock); 817 spin_unlock_irq(&qp->r_lock); 818 819 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 820 insert_qp(dev, qp); 821 822 if (lastwqe) { 823 ev.device = qp->ibqp.device; 824 ev.element.qp = &qp->ibqp; 825 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 826 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 827 } 828 if (mig) { 829 ev.device = qp->ibqp.device; 830 ev.element.qp = &qp->ibqp; 831 ev.event = IB_EVENT_PATH_MIG; 832 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 833 } 834 ret = 0; 835 goto bail; 836 837 inval: 838 spin_unlock(&qp->s_lock); 839 spin_unlock_irq(&qp->r_lock); 840 ret = -EINVAL; 841 842 bail: 843 return ret; 844 } 845 846 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 847 int attr_mask, struct ib_qp_init_attr *init_attr) 848 { 849 struct qib_qp *qp = to_iqp(ibqp); 850 851 attr->qp_state = qp->state; 852 attr->cur_qp_state = attr->qp_state; 853 attr->path_mtu = qp->path_mtu; 854 attr->path_mig_state = qp->s_mig_state; 855 attr->qkey = qp->qkey; 856 attr->rq_psn = qp->r_psn & QIB_PSN_MASK; 857 attr->sq_psn = qp->s_next_psn & QIB_PSN_MASK; 858 attr->dest_qp_num = qp->remote_qpn; 859 attr->qp_access_flags = qp->qp_access_flags; 860 attr->cap.max_send_wr = qp->s_size - 1; 861 attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1; 862 attr->cap.max_send_sge = qp->s_max_sge; 863 attr->cap.max_recv_sge = qp->r_rq.max_sge; 864 attr->cap.max_inline_data = 0; 865 attr->ah_attr = qp->remote_ah_attr; 866 attr->alt_ah_attr = qp->alt_ah_attr; 867 attr->pkey_index = qp->s_pkey_index; 868 attr->alt_pkey_index = qp->s_alt_pkey_index; 869 attr->en_sqd_async_notify = 0; 870 attr->sq_draining = qp->s_draining; 871 attr->max_rd_atomic = qp->s_max_rd_atomic; 872 attr->max_dest_rd_atomic = qp->r_max_rd_atomic; 873 attr->min_rnr_timer = qp->r_min_rnr_timer; 874 attr->port_num = qp->port_num; 875 attr->timeout = qp->timeout; 876 attr->retry_cnt = qp->s_retry_cnt; 877 attr->rnr_retry = qp->s_rnr_retry_cnt; 878 attr->alt_port_num = qp->alt_ah_attr.port_num; 879 attr->alt_timeout = qp->alt_timeout; 880 881 init_attr->event_handler = qp->ibqp.event_handler; 882 init_attr->qp_context = qp->ibqp.qp_context; 883 init_attr->send_cq = qp->ibqp.send_cq; 884 init_attr->recv_cq = qp->ibqp.recv_cq; 885 init_attr->srq = qp->ibqp.srq; 886 init_attr->cap = attr->cap; 887 if (qp->s_flags & QIB_S_SIGNAL_REQ_WR) 888 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR; 889 else 890 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR; 891 init_attr->qp_type = qp->ibqp.qp_type; 892 init_attr->port_num = qp->port_num; 893 return 0; 894 } 895 896 /** 897 * qib_compute_aeth - compute the AETH (syndrome + MSN) 898 * @qp: the queue pair to compute the AETH for 899 * 900 * Returns the AETH. 901 */ 902 __be32 qib_compute_aeth(struct qib_qp *qp) 903 { 904 u32 aeth = qp->r_msn & QIB_MSN_MASK; 905 906 if (qp->ibqp.srq) { 907 /* 908 * Shared receive queues don't generate credits. 909 * Set the credit field to the invalid value. 910 */ 911 aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT; 912 } else { 913 u32 min, max, x; 914 u32 credits; 915 struct qib_rwq *wq = qp->r_rq.wq; 916 u32 head; 917 u32 tail; 918 919 /* sanity check pointers before trusting them */ 920 head = wq->head; 921 if (head >= qp->r_rq.size) 922 head = 0; 923 tail = wq->tail; 924 if (tail >= qp->r_rq.size) 925 tail = 0; 926 /* 927 * Compute the number of credits available (RWQEs). 928 * XXX Not holding the r_rq.lock here so there is a small 929 * chance that the pair of reads are not atomic. 930 */ 931 credits = head - tail; 932 if ((int)credits < 0) 933 credits += qp->r_rq.size; 934 /* 935 * Binary search the credit table to find the code to 936 * use. 937 */ 938 min = 0; 939 max = 31; 940 for (;;) { 941 x = (min + max) / 2; 942 if (credit_table[x] == credits) 943 break; 944 if (credit_table[x] > credits) 945 max = x; 946 else if (min == x) 947 break; 948 else 949 min = x; 950 } 951 aeth |= x << QIB_AETH_CREDIT_SHIFT; 952 } 953 return cpu_to_be32(aeth); 954 } 955 956 /** 957 * qib_create_qp - create a queue pair for a device 958 * @ibpd: the protection domain who's device we create the queue pair for 959 * @init_attr: the attributes of the queue pair 960 * @udata: user data for libibverbs.so 961 * 962 * Returns the queue pair on success, otherwise returns an errno. 963 * 964 * Called by the ib_create_qp() core verbs function. 965 */ 966 struct ib_qp *qib_create_qp(struct ib_pd *ibpd, 967 struct ib_qp_init_attr *init_attr, 968 struct ib_udata *udata) 969 { 970 struct qib_qp *qp; 971 int err; 972 struct qib_swqe *swq = NULL; 973 struct qib_ibdev *dev; 974 struct qib_devdata *dd; 975 size_t sz; 976 size_t sg_list_sz; 977 struct ib_qp *ret; 978 979 if (init_attr->cap.max_send_sge > ib_qib_max_sges || 980 init_attr->cap.max_send_wr > ib_qib_max_qp_wrs) { 981 ret = ERR_PTR(-EINVAL); 982 goto bail; 983 } 984 985 /* Check receive queue parameters if no SRQ is specified. */ 986 if (!init_attr->srq) { 987 if (init_attr->cap.max_recv_sge > ib_qib_max_sges || 988 init_attr->cap.max_recv_wr > ib_qib_max_qp_wrs) { 989 ret = ERR_PTR(-EINVAL); 990 goto bail; 991 } 992 if (init_attr->cap.max_send_sge + 993 init_attr->cap.max_send_wr + 994 init_attr->cap.max_recv_sge + 995 init_attr->cap.max_recv_wr == 0) { 996 ret = ERR_PTR(-EINVAL); 997 goto bail; 998 } 999 } 1000 1001 switch (init_attr->qp_type) { 1002 case IB_QPT_SMI: 1003 case IB_QPT_GSI: 1004 if (init_attr->port_num == 0 || 1005 init_attr->port_num > ibpd->device->phys_port_cnt) { 1006 ret = ERR_PTR(-EINVAL); 1007 goto bail; 1008 } 1009 case IB_QPT_UC: 1010 case IB_QPT_RC: 1011 case IB_QPT_UD: 1012 sz = sizeof(struct qib_sge) * 1013 init_attr->cap.max_send_sge + 1014 sizeof(struct qib_swqe); 1015 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz); 1016 if (swq == NULL) { 1017 ret = ERR_PTR(-ENOMEM); 1018 goto bail; 1019 } 1020 sz = sizeof(*qp); 1021 sg_list_sz = 0; 1022 if (init_attr->srq) { 1023 struct qib_srq *srq = to_isrq(init_attr->srq); 1024 1025 if (srq->rq.max_sge > 1) 1026 sg_list_sz = sizeof(*qp->r_sg_list) * 1027 (srq->rq.max_sge - 1); 1028 } else if (init_attr->cap.max_recv_sge > 1) 1029 sg_list_sz = sizeof(*qp->r_sg_list) * 1030 (init_attr->cap.max_recv_sge - 1); 1031 qp = kzalloc(sz + sg_list_sz, GFP_KERNEL); 1032 if (!qp) { 1033 ret = ERR_PTR(-ENOMEM); 1034 goto bail_swq; 1035 } 1036 RCU_INIT_POINTER(qp->next, NULL); 1037 if (init_attr->srq) 1038 sz = 0; 1039 else { 1040 qp->r_rq.size = init_attr->cap.max_recv_wr + 1; 1041 qp->r_rq.max_sge = init_attr->cap.max_recv_sge; 1042 sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + 1043 sizeof(struct qib_rwqe); 1044 qp->r_rq.wq = vmalloc_user(sizeof(struct qib_rwq) + 1045 qp->r_rq.size * sz); 1046 if (!qp->r_rq.wq) { 1047 ret = ERR_PTR(-ENOMEM); 1048 goto bail_qp; 1049 } 1050 } 1051 1052 /* 1053 * ib_create_qp() will initialize qp->ibqp 1054 * except for qp->ibqp.qp_num. 1055 */ 1056 spin_lock_init(&qp->r_lock); 1057 spin_lock_init(&qp->s_lock); 1058 spin_lock_init(&qp->r_rq.lock); 1059 atomic_set(&qp->refcount, 0); 1060 init_waitqueue_head(&qp->wait); 1061 init_waitqueue_head(&qp->wait_dma); 1062 init_timer(&qp->s_timer); 1063 qp->s_timer.data = (unsigned long)qp; 1064 INIT_WORK(&qp->s_work, qib_do_send); 1065 INIT_LIST_HEAD(&qp->iowait); 1066 INIT_LIST_HEAD(&qp->rspwait); 1067 qp->state = IB_QPS_RESET; 1068 qp->s_wq = swq; 1069 qp->s_size = init_attr->cap.max_send_wr + 1; 1070 qp->s_max_sge = init_attr->cap.max_send_sge; 1071 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) 1072 qp->s_flags = QIB_S_SIGNAL_REQ_WR; 1073 dev = to_idev(ibpd->device); 1074 dd = dd_from_dev(dev); 1075 err = alloc_qpn(dd, &dev->qpn_table, init_attr->qp_type, 1076 init_attr->port_num); 1077 if (err < 0) { 1078 ret = ERR_PTR(err); 1079 vfree(qp->r_rq.wq); 1080 goto bail_qp; 1081 } 1082 qp->ibqp.qp_num = err; 1083 qp->port_num = init_attr->port_num; 1084 qib_reset_qp(qp, init_attr->qp_type); 1085 break; 1086 1087 default: 1088 /* Don't support raw QPs */ 1089 ret = ERR_PTR(-ENOSYS); 1090 goto bail; 1091 } 1092 1093 init_attr->cap.max_inline_data = 0; 1094 1095 /* 1096 * Return the address of the RWQ as the offset to mmap. 1097 * See qib_mmap() for details. 1098 */ 1099 if (udata && udata->outlen >= sizeof(__u64)) { 1100 if (!qp->r_rq.wq) { 1101 __u64 offset = 0; 1102 1103 err = ib_copy_to_udata(udata, &offset, 1104 sizeof(offset)); 1105 if (err) { 1106 ret = ERR_PTR(err); 1107 goto bail_ip; 1108 } 1109 } else { 1110 u32 s = sizeof(struct qib_rwq) + qp->r_rq.size * sz; 1111 1112 qp->ip = qib_create_mmap_info(dev, s, 1113 ibpd->uobject->context, 1114 qp->r_rq.wq); 1115 if (!qp->ip) { 1116 ret = ERR_PTR(-ENOMEM); 1117 goto bail_ip; 1118 } 1119 1120 err = ib_copy_to_udata(udata, &(qp->ip->offset), 1121 sizeof(qp->ip->offset)); 1122 if (err) { 1123 ret = ERR_PTR(err); 1124 goto bail_ip; 1125 } 1126 } 1127 } 1128 1129 spin_lock(&dev->n_qps_lock); 1130 if (dev->n_qps_allocated == ib_qib_max_qps) { 1131 spin_unlock(&dev->n_qps_lock); 1132 ret = ERR_PTR(-ENOMEM); 1133 goto bail_ip; 1134 } 1135 1136 dev->n_qps_allocated++; 1137 spin_unlock(&dev->n_qps_lock); 1138 1139 if (qp->ip) { 1140 spin_lock_irq(&dev->pending_lock); 1141 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps); 1142 spin_unlock_irq(&dev->pending_lock); 1143 } 1144 1145 ret = &qp->ibqp; 1146 goto bail; 1147 1148 bail_ip: 1149 if (qp->ip) 1150 kref_put(&qp->ip->ref, qib_release_mmap_info); 1151 else 1152 vfree(qp->r_rq.wq); 1153 free_qpn(&dev->qpn_table, qp->ibqp.qp_num); 1154 bail_qp: 1155 kfree(qp); 1156 bail_swq: 1157 vfree(swq); 1158 bail: 1159 return ret; 1160 } 1161 1162 /** 1163 * qib_destroy_qp - destroy a queue pair 1164 * @ibqp: the queue pair to destroy 1165 * 1166 * Returns 0 on success. 1167 * 1168 * Note that this can be called while the QP is actively sending or 1169 * receiving! 1170 */ 1171 int qib_destroy_qp(struct ib_qp *ibqp) 1172 { 1173 struct qib_qp *qp = to_iqp(ibqp); 1174 struct qib_ibdev *dev = to_idev(ibqp->device); 1175 1176 /* Make sure HW and driver activity is stopped. */ 1177 spin_lock_irq(&qp->s_lock); 1178 if (qp->state != IB_QPS_RESET) { 1179 qp->state = IB_QPS_RESET; 1180 spin_lock(&dev->pending_lock); 1181 if (!list_empty(&qp->iowait)) 1182 list_del_init(&qp->iowait); 1183 spin_unlock(&dev->pending_lock); 1184 qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT); 1185 spin_unlock_irq(&qp->s_lock); 1186 cancel_work_sync(&qp->s_work); 1187 del_timer_sync(&qp->s_timer); 1188 wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy)); 1189 if (qp->s_tx) { 1190 qib_put_txreq(qp->s_tx); 1191 qp->s_tx = NULL; 1192 } 1193 remove_qp(dev, qp); 1194 wait_event(qp->wait, !atomic_read(&qp->refcount)); 1195 clear_mr_refs(qp, 1); 1196 } else 1197 spin_unlock_irq(&qp->s_lock); 1198 1199 /* all user's cleaned up, mark it available */ 1200 free_qpn(&dev->qpn_table, qp->ibqp.qp_num); 1201 spin_lock(&dev->n_qps_lock); 1202 dev->n_qps_allocated--; 1203 spin_unlock(&dev->n_qps_lock); 1204 1205 if (qp->ip) 1206 kref_put(&qp->ip->ref, qib_release_mmap_info); 1207 else 1208 vfree(qp->r_rq.wq); 1209 vfree(qp->s_wq); 1210 kfree(qp); 1211 return 0; 1212 } 1213 1214 /** 1215 * qib_init_qpn_table - initialize the QP number table for a device 1216 * @qpt: the QPN table 1217 */ 1218 void qib_init_qpn_table(struct qib_devdata *dd, struct qib_qpn_table *qpt) 1219 { 1220 spin_lock_init(&qpt->lock); 1221 qpt->last = 1; /* start with QPN 2 */ 1222 qpt->nmaps = 1; 1223 qpt->mask = dd->qpn_mask; 1224 } 1225 1226 /** 1227 * qib_free_qpn_table - free the QP number table for a device 1228 * @qpt: the QPN table 1229 */ 1230 void qib_free_qpn_table(struct qib_qpn_table *qpt) 1231 { 1232 int i; 1233 1234 for (i = 0; i < ARRAY_SIZE(qpt->map); i++) 1235 if (qpt->map[i].page) 1236 free_page((unsigned long) qpt->map[i].page); 1237 } 1238 1239 /** 1240 * qib_get_credit - flush the send work queue of a QP 1241 * @qp: the qp who's send work queue to flush 1242 * @aeth: the Acknowledge Extended Transport Header 1243 * 1244 * The QP s_lock should be held. 1245 */ 1246 void qib_get_credit(struct qib_qp *qp, u32 aeth) 1247 { 1248 u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK; 1249 1250 /* 1251 * If the credit is invalid, we can send 1252 * as many packets as we like. Otherwise, we have to 1253 * honor the credit field. 1254 */ 1255 if (credit == QIB_AETH_CREDIT_INVAL) { 1256 if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) { 1257 qp->s_flags |= QIB_S_UNLIMITED_CREDIT; 1258 if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) { 1259 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT; 1260 qib_schedule_send(qp); 1261 } 1262 } 1263 } else if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) { 1264 /* Compute new LSN (i.e., MSN + credit) */ 1265 credit = (aeth + credit_table[credit]) & QIB_MSN_MASK; 1266 if (qib_cmp24(credit, qp->s_lsn) > 0) { 1267 qp->s_lsn = credit; 1268 if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) { 1269 qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT; 1270 qib_schedule_send(qp); 1271 } 1272 } 1273 } 1274 } 1275