ipath_qp.c revision 60460dfd425e3d301831d8dd6398620d67bbda2b
1/* 2 * Copyright (c) 2006 QLogic, Inc. All rights reserved. 3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34#include <linux/err.h> 35#include <linux/vmalloc.h> 36 37#include "ipath_verbs.h" 38#include "ips_common.h" 39 40#define BITS_PER_PAGE (PAGE_SIZE*BITS_PER_BYTE) 41#define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1) 42#define mk_qpn(qpt, map, off) (((map) - (qpt)->map) * BITS_PER_PAGE + \ 43 (off)) 44#define find_next_offset(map, off) find_next_zero_bit((map)->page, \ 45 BITS_PER_PAGE, off) 46 47#define TRANS_INVALID 0 48#define TRANS_ANY2RST 1 49#define TRANS_RST2INIT 2 50#define TRANS_INIT2INIT 3 51#define TRANS_INIT2RTR 4 52#define TRANS_RTR2RTS 5 53#define TRANS_RTS2RTS 6 54#define TRANS_SQERR2RTS 7 55#define TRANS_ANY2ERR 8 56#define TRANS_RTS2SQD 9 /* XXX Wait for expected ACKs & signal event */ 57#define TRANS_SQD2SQD 10 /* error if not drained & parameter change */ 58#define TRANS_SQD2RTS 11 /* error if not drained */ 59 60/* 61 * Convert the AETH credit code into the number of credits. 62 */ 63static u32 credit_table[31] = { 64 0, /* 0 */ 65 1, /* 1 */ 66 2, /* 2 */ 67 3, /* 3 */ 68 4, /* 4 */ 69 6, /* 5 */ 70 8, /* 6 */ 71 12, /* 7 */ 72 16, /* 8 */ 73 24, /* 9 */ 74 32, /* A */ 75 48, /* B */ 76 64, /* C */ 77 96, /* D */ 78 128, /* E */ 79 192, /* F */ 80 256, /* 10 */ 81 384, /* 11 */ 82 512, /* 12 */ 83 768, /* 13 */ 84 1024, /* 14 */ 85 1536, /* 15 */ 86 2048, /* 16 */ 87 3072, /* 17 */ 88 4096, /* 18 */ 89 6144, /* 19 */ 90 8192, /* 1A */ 91 12288, /* 1B */ 92 16384, /* 1C */ 93 24576, /* 1D */ 94 32768 /* 1E */ 95}; 96 97static u32 alloc_qpn(struct ipath_qp_table *qpt) 98{ 99 u32 i, offset, max_scan, qpn; 100 struct qpn_map *map; 101 u32 ret; 102 103 qpn = qpt->last + 1; 104 if (qpn >= QPN_MAX) 105 qpn = 2; 106 offset = qpn & BITS_PER_PAGE_MASK; 107 map = &qpt->map[qpn / BITS_PER_PAGE]; 108 max_scan = qpt->nmaps - !offset; 109 for (i = 0;;) { 110 if (unlikely(!map->page)) { 111 unsigned long page = get_zeroed_page(GFP_KERNEL); 112 unsigned long flags; 113 114 /* 115 * Free the page if someone raced with us 116 * installing it: 117 */ 118 spin_lock_irqsave(&qpt->lock, flags); 119 if (map->page) 120 free_page(page); 121 else 122 map->page = (void *)page; 123 spin_unlock_irqrestore(&qpt->lock, flags); 124 if (unlikely(!map->page)) 125 break; 126 } 127 if (likely(atomic_read(&map->n_free))) { 128 do { 129 if (!test_and_set_bit(offset, map->page)) { 130 atomic_dec(&map->n_free); 131 qpt->last = qpn; 132 ret = qpn; 133 goto bail; 134 } 135 offset = find_next_offset(map, offset); 136 qpn = mk_qpn(qpt, map, offset); 137 /* 138 * This test differs from alloc_pidmap(). 139 * If find_next_offset() does find a zero 140 * bit, we don't need to check for QPN 141 * wrapping around past our starting QPN. 142 * We just need to be sure we don't loop 143 * forever. 144 */ 145 } while (offset < BITS_PER_PAGE && qpn < QPN_MAX); 146 } 147 /* 148 * In order to keep the number of pages allocated to a 149 * minimum, we scan the all existing pages before increasing 150 * the size of the bitmap table. 151 */ 152 if (++i > max_scan) { 153 if (qpt->nmaps == QPNMAP_ENTRIES) 154 break; 155 map = &qpt->map[qpt->nmaps++]; 156 offset = 0; 157 } else if (map < &qpt->map[qpt->nmaps]) { 158 ++map; 159 offset = 0; 160 } else { 161 map = &qpt->map[0]; 162 offset = 2; 163 } 164 qpn = mk_qpn(qpt, map, offset); 165 } 166 167 ret = 0; 168 169bail: 170 return ret; 171} 172 173static void free_qpn(struct ipath_qp_table *qpt, u32 qpn) 174{ 175 struct qpn_map *map; 176 177 map = qpt->map + qpn / BITS_PER_PAGE; 178 if (map->page) 179 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page); 180 atomic_inc(&map->n_free); 181} 182 183/** 184 * ipath_alloc_qpn - allocate a QP number 185 * @qpt: the QP table 186 * @qp: the QP 187 * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special) 188 * 189 * Allocate the next available QPN and put the QP into the hash table. 190 * The hash table holds a reference to the QP. 191 */ 192static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp, 193 enum ib_qp_type type) 194{ 195 unsigned long flags; 196 u32 qpn; 197 int ret; 198 199 if (type == IB_QPT_SMI) 200 qpn = 0; 201 else if (type == IB_QPT_GSI) 202 qpn = 1; 203 else { 204 /* Allocate the next available QPN */ 205 qpn = alloc_qpn(qpt); 206 if (qpn == 0) { 207 ret = -ENOMEM; 208 goto bail; 209 } 210 } 211 qp->ibqp.qp_num = qpn; 212 213 /* Add the QP to the hash table. */ 214 spin_lock_irqsave(&qpt->lock, flags); 215 216 qpn %= qpt->max; 217 qp->next = qpt->table[qpn]; 218 qpt->table[qpn] = qp; 219 atomic_inc(&qp->refcount); 220 221 spin_unlock_irqrestore(&qpt->lock, flags); 222 ret = 0; 223 224bail: 225 return ret; 226} 227 228/** 229 * ipath_free_qp - remove a QP from the QP table 230 * @qpt: the QP table 231 * @qp: the QP to remove 232 * 233 * Remove the QP from the table so it can't be found asynchronously by 234 * the receive interrupt routine. 235 */ 236static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp) 237{ 238 struct ipath_qp *q, **qpp; 239 unsigned long flags; 240 int fnd = 0; 241 242 spin_lock_irqsave(&qpt->lock, flags); 243 244 /* Remove QP from the hash table. */ 245 qpp = &qpt->table[qp->ibqp.qp_num % qpt->max]; 246 for (; (q = *qpp) != NULL; qpp = &q->next) { 247 if (q == qp) { 248 *qpp = qp->next; 249 qp->next = NULL; 250 atomic_dec(&qp->refcount); 251 fnd = 1; 252 break; 253 } 254 } 255 256 spin_unlock_irqrestore(&qpt->lock, flags); 257 258 if (!fnd) 259 return; 260 261 /* If QPN is not reserved, mark QPN free in the bitmap. */ 262 if (qp->ibqp.qp_num > 1) 263 free_qpn(qpt, qp->ibqp.qp_num); 264 265 wait_event(qp->wait, !atomic_read(&qp->refcount)); 266} 267 268/** 269 * ipath_free_all_qps - remove all QPs from the table 270 * @qpt: the QP table to empty 271 */ 272void ipath_free_all_qps(struct ipath_qp_table *qpt) 273{ 274 unsigned long flags; 275 struct ipath_qp *qp, *nqp; 276 u32 n; 277 278 for (n = 0; n < qpt->max; n++) { 279 spin_lock_irqsave(&qpt->lock, flags); 280 qp = qpt->table[n]; 281 qpt->table[n] = NULL; 282 spin_unlock_irqrestore(&qpt->lock, flags); 283 284 while (qp) { 285 nqp = qp->next; 286 if (qp->ibqp.qp_num > 1) 287 free_qpn(qpt, qp->ibqp.qp_num); 288 if (!atomic_dec_and_test(&qp->refcount) || 289 !ipath_destroy_qp(&qp->ibqp)) 290 _VERBS_INFO("QP memory leak!\n"); 291 qp = nqp; 292 } 293 } 294 295 for (n = 0; n < ARRAY_SIZE(qpt->map); n++) { 296 if (qpt->map[n].page) 297 free_page((unsigned long)qpt->map[n].page); 298 } 299} 300 301/** 302 * ipath_lookup_qpn - return the QP with the given QPN 303 * @qpt: the QP table 304 * @qpn: the QP number to look up 305 * 306 * The caller is responsible for decrementing the QP reference count 307 * when done. 308 */ 309struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn) 310{ 311 unsigned long flags; 312 struct ipath_qp *qp; 313 314 spin_lock_irqsave(&qpt->lock, flags); 315 316 for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) { 317 if (qp->ibqp.qp_num == qpn) { 318 atomic_inc(&qp->refcount); 319 break; 320 } 321 } 322 323 spin_unlock_irqrestore(&qpt->lock, flags); 324 return qp; 325} 326 327/** 328 * ipath_reset_qp - initialize the QP state to the reset state 329 * @qp: the QP to reset 330 */ 331static void ipath_reset_qp(struct ipath_qp *qp) 332{ 333 qp->remote_qpn = 0; 334 qp->qkey = 0; 335 qp->qp_access_flags = 0; 336 qp->s_hdrwords = 0; 337 qp->s_psn = 0; 338 qp->r_psn = 0; 339 atomic_set(&qp->msn, 0); 340 if (qp->ibqp.qp_type == IB_QPT_RC) { 341 qp->s_state = IB_OPCODE_RC_SEND_LAST; 342 qp->r_state = IB_OPCODE_RC_SEND_LAST; 343 } else { 344 qp->s_state = IB_OPCODE_UC_SEND_LAST; 345 qp->r_state = IB_OPCODE_UC_SEND_LAST; 346 } 347 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 348 qp->s_nak_state = 0; 349 qp->s_rnr_timeout = 0; 350 qp->s_head = 0; 351 qp->s_tail = 0; 352 qp->s_cur = 0; 353 qp->s_last = 0; 354 qp->s_ssn = 1; 355 qp->s_lsn = 0; 356 qp->r_rq.head = 0; 357 qp->r_rq.tail = 0; 358 qp->r_reuse_sge = 0; 359} 360 361/** 362 * ipath_error_qp - put a QP into an error state 363 * @qp: the QP to put into an error state 364 * 365 * Flushes both send and receive work queues. 366 * QP r_rq.lock and s_lock should be held. 367 */ 368 369static void ipath_error_qp(struct ipath_qp *qp) 370{ 371 struct ipath_ibdev *dev = to_idev(qp->ibqp.device); 372 struct ib_wc wc; 373 374 _VERBS_INFO("QP%d/%d in error state\n", 375 qp->ibqp.qp_num, qp->remote_qpn); 376 377 spin_lock(&dev->pending_lock); 378 /* XXX What if its already removed by the timeout code? */ 379 if (!list_empty(&qp->timerwait)) 380 list_del_init(&qp->timerwait); 381 if (!list_empty(&qp->piowait)) 382 list_del_init(&qp->piowait); 383 spin_unlock(&dev->pending_lock); 384 385 wc.status = IB_WC_WR_FLUSH_ERR; 386 wc.vendor_err = 0; 387 wc.byte_len = 0; 388 wc.imm_data = 0; 389 wc.qp_num = qp->ibqp.qp_num; 390 wc.src_qp = 0; 391 wc.wc_flags = 0; 392 wc.pkey_index = 0; 393 wc.slid = 0; 394 wc.sl = 0; 395 wc.dlid_path_bits = 0; 396 wc.port_num = 0; 397 398 while (qp->s_last != qp->s_head) { 399 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last); 400 401 wc.wr_id = wqe->wr.wr_id; 402 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode]; 403 if (++qp->s_last >= qp->s_size) 404 qp->s_last = 0; 405 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1); 406 } 407 qp->s_cur = qp->s_tail = qp->s_head; 408 qp->s_hdrwords = 0; 409 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 410 411 wc.opcode = IB_WC_RECV; 412 while (qp->r_rq.tail != qp->r_rq.head) { 413 wc.wr_id = get_rwqe_ptr(&qp->r_rq, qp->r_rq.tail)->wr_id; 414 if (++qp->r_rq.tail >= qp->r_rq.size) 415 qp->r_rq.tail = 0; 416 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1); 417 } 418} 419 420/** 421 * ipath_modify_qp - modify the attributes of a queue pair 422 * @ibqp: the queue pair who's attributes we're modifying 423 * @attr: the new attributes 424 * @attr_mask: the mask of attributes to modify 425 * 426 * Returns 0 on success, otherwise returns an errno. 427 */ 428int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 429 int attr_mask) 430{ 431 struct ipath_ibdev *dev = to_idev(ibqp->device); 432 struct ipath_qp *qp = to_iqp(ibqp); 433 enum ib_qp_state cur_state, new_state; 434 unsigned long flags; 435 int ret; 436 437 spin_lock_irqsave(&qp->r_rq.lock, flags); 438 spin_lock(&qp->s_lock); 439 440 cur_state = attr_mask & IB_QP_CUR_STATE ? 441 attr->cur_qp_state : qp->state; 442 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 443 444 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 445 attr_mask)) 446 goto inval; 447 448 if (attr_mask & IB_QP_AV) 449 if (attr->ah_attr.dlid == 0 || 450 attr->ah_attr.dlid >= IPS_MULTICAST_LID_BASE) 451 goto inval; 452 453 if (attr_mask & IB_QP_PKEY_INDEX) 454 if (attr->pkey_index >= ipath_layer_get_npkeys(dev->dd)) 455 goto inval; 456 457 if (attr_mask & IB_QP_MIN_RNR_TIMER) 458 if (attr->min_rnr_timer > 31) 459 goto inval; 460 461 switch (new_state) { 462 case IB_QPS_RESET: 463 ipath_reset_qp(qp); 464 break; 465 466 case IB_QPS_ERR: 467 ipath_error_qp(qp); 468 break; 469 470 default: 471 break; 472 473 } 474 475 if (attr_mask & IB_QP_PKEY_INDEX) 476 qp->s_pkey_index = attr->pkey_index; 477 478 if (attr_mask & IB_QP_DEST_QPN) 479 qp->remote_qpn = attr->dest_qp_num; 480 481 if (attr_mask & IB_QP_SQ_PSN) { 482 qp->s_next_psn = attr->sq_psn; 483 qp->s_last_psn = qp->s_next_psn - 1; 484 } 485 486 if (attr_mask & IB_QP_RQ_PSN) 487 qp->r_psn = attr->rq_psn; 488 489 if (attr_mask & IB_QP_ACCESS_FLAGS) 490 qp->qp_access_flags = attr->qp_access_flags; 491 492 if (attr_mask & IB_QP_AV) 493 qp->remote_ah_attr = attr->ah_attr; 494 495 if (attr_mask & IB_QP_PATH_MTU) 496 qp->path_mtu = attr->path_mtu; 497 498 if (attr_mask & IB_QP_RETRY_CNT) 499 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt; 500 501 if (attr_mask & IB_QP_RNR_RETRY) { 502 qp->s_rnr_retry = attr->rnr_retry; 503 if (qp->s_rnr_retry > 7) 504 qp->s_rnr_retry = 7; 505 qp->s_rnr_retry_cnt = qp->s_rnr_retry; 506 } 507 508 if (attr_mask & IB_QP_MIN_RNR_TIMER) 509 qp->s_min_rnr_timer = attr->min_rnr_timer; 510 511 if (attr_mask & IB_QP_QKEY) 512 qp->qkey = attr->qkey; 513 514 qp->state = new_state; 515 spin_unlock(&qp->s_lock); 516 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 517 518 /* 519 * If QP1 changed to the RTS state, try to move to the link to INIT 520 * even if it was ACTIVE so the SM will reinitialize the SMA's 521 * state. 522 */ 523 if (qp->ibqp.qp_num == 1 && new_state == IB_QPS_RTS) { 524 struct ipath_ibdev *dev = to_idev(ibqp->device); 525 526 ipath_layer_set_linkstate(dev->dd, IPATH_IB_LINKDOWN); 527 } 528 ret = 0; 529 goto bail; 530 531inval: 532 spin_unlock(&qp->s_lock); 533 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 534 ret = -EINVAL; 535 536bail: 537 return ret; 538} 539 540int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 541 int attr_mask, struct ib_qp_init_attr *init_attr) 542{ 543 struct ipath_qp *qp = to_iqp(ibqp); 544 545 attr->qp_state = qp->state; 546 attr->cur_qp_state = attr->qp_state; 547 attr->path_mtu = qp->path_mtu; 548 attr->path_mig_state = 0; 549 attr->qkey = qp->qkey; 550 attr->rq_psn = qp->r_psn; 551 attr->sq_psn = qp->s_next_psn; 552 attr->dest_qp_num = qp->remote_qpn; 553 attr->qp_access_flags = qp->qp_access_flags; 554 attr->cap.max_send_wr = qp->s_size - 1; 555 attr->cap.max_recv_wr = qp->r_rq.size - 1; 556 attr->cap.max_send_sge = qp->s_max_sge; 557 attr->cap.max_recv_sge = qp->r_rq.max_sge; 558 attr->cap.max_inline_data = 0; 559 attr->ah_attr = qp->remote_ah_attr; 560 memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr)); 561 attr->pkey_index = qp->s_pkey_index; 562 attr->alt_pkey_index = 0; 563 attr->en_sqd_async_notify = 0; 564 attr->sq_draining = 0; 565 attr->max_rd_atomic = 1; 566 attr->max_dest_rd_atomic = 1; 567 attr->min_rnr_timer = qp->s_min_rnr_timer; 568 attr->port_num = 1; 569 attr->timeout = 0; 570 attr->retry_cnt = qp->s_retry_cnt; 571 attr->rnr_retry = qp->s_rnr_retry; 572 attr->alt_port_num = 0; 573 attr->alt_timeout = 0; 574 575 init_attr->event_handler = qp->ibqp.event_handler; 576 init_attr->qp_context = qp->ibqp.qp_context; 577 init_attr->send_cq = qp->ibqp.send_cq; 578 init_attr->recv_cq = qp->ibqp.recv_cq; 579 init_attr->srq = qp->ibqp.srq; 580 init_attr->cap = attr->cap; 581 init_attr->sq_sig_type = 582 (qp->s_flags & (1 << IPATH_S_SIGNAL_REQ_WR)) 583 ? IB_SIGNAL_REQ_WR : 0; 584 init_attr->qp_type = qp->ibqp.qp_type; 585 init_attr->port_num = 1; 586 return 0; 587} 588 589/** 590 * ipath_compute_aeth - compute the AETH (syndrome + MSN) 591 * @qp: the queue pair to compute the AETH for 592 * 593 * Returns the AETH. 594 * 595 * The QP s_lock should be held. 596 */ 597__be32 ipath_compute_aeth(struct ipath_qp *qp) 598{ 599 u32 aeth = atomic_read(&qp->msn) & IPS_MSN_MASK; 600 601 if (qp->s_nak_state) { 602 aeth |= qp->s_nak_state << IPS_AETH_CREDIT_SHIFT; 603 } else if (qp->ibqp.srq) { 604 /* 605 * Shared receive queues don't generate credits. 606 * Set the credit field to the invalid value. 607 */ 608 aeth |= IPS_AETH_CREDIT_INVAL << IPS_AETH_CREDIT_SHIFT; 609 } else { 610 u32 min, max, x; 611 u32 credits; 612 613 /* 614 * Compute the number of credits available (RWQEs). 615 * XXX Not holding the r_rq.lock here so there is a small 616 * chance that the pair of reads are not atomic. 617 */ 618 credits = qp->r_rq.head - qp->r_rq.tail; 619 if ((int)credits < 0) 620 credits += qp->r_rq.size; 621 /* 622 * Binary search the credit table to find the code to 623 * use. 624 */ 625 min = 0; 626 max = 31; 627 for (;;) { 628 x = (min + max) / 2; 629 if (credit_table[x] == credits) 630 break; 631 if (credit_table[x] > credits) 632 max = x; 633 else if (min == x) 634 break; 635 else 636 min = x; 637 } 638 aeth |= x << IPS_AETH_CREDIT_SHIFT; 639 } 640 return cpu_to_be32(aeth); 641} 642 643/** 644 * ipath_create_qp - create a queue pair for a device 645 * @ibpd: the protection domain who's device we create the queue pair for 646 * @init_attr: the attributes of the queue pair 647 * @udata: unused by InfiniPath 648 * 649 * Returns the queue pair on success, otherwise returns an errno. 650 * 651 * Called by the ib_create_qp() core verbs function. 652 */ 653struct ib_qp *ipath_create_qp(struct ib_pd *ibpd, 654 struct ib_qp_init_attr *init_attr, 655 struct ib_udata *udata) 656{ 657 struct ipath_qp *qp; 658 int err; 659 struct ipath_swqe *swq = NULL; 660 struct ipath_ibdev *dev; 661 size_t sz; 662 struct ib_qp *ret; 663 664 if (init_attr->cap.max_send_sge > 255 || 665 init_attr->cap.max_recv_sge > 255) { 666 ret = ERR_PTR(-ENOMEM); 667 goto bail; 668 } 669 670 if (init_attr->cap.max_send_sge + 671 init_attr->cap.max_recv_sge + 672 init_attr->cap.max_send_wr + 673 init_attr->cap.max_recv_wr == 0) { 674 ret = ERR_PTR(-EINVAL); 675 goto bail; 676 } 677 678 switch (init_attr->qp_type) { 679 case IB_QPT_UC: 680 case IB_QPT_RC: 681 sz = sizeof(struct ipath_sge) * 682 init_attr->cap.max_send_sge + 683 sizeof(struct ipath_swqe); 684 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz); 685 if (swq == NULL) { 686 ret = ERR_PTR(-ENOMEM); 687 goto bail; 688 } 689 /* FALLTHROUGH */ 690 case IB_QPT_UD: 691 case IB_QPT_SMI: 692 case IB_QPT_GSI: 693 qp = kmalloc(sizeof(*qp), GFP_KERNEL); 694 if (!qp) { 695 vfree(swq); 696 ret = ERR_PTR(-ENOMEM); 697 goto bail; 698 } 699 qp->r_rq.size = init_attr->cap.max_recv_wr + 1; 700 sz = sizeof(struct ipath_sge) * 701 init_attr->cap.max_recv_sge + 702 sizeof(struct ipath_rwqe); 703 qp->r_rq.wq = vmalloc(qp->r_rq.size * sz); 704 if (!qp->r_rq.wq) { 705 kfree(qp); 706 vfree(swq); 707 ret = ERR_PTR(-ENOMEM); 708 goto bail; 709 } 710 711 /* 712 * ib_create_qp() will initialize qp->ibqp 713 * except for qp->ibqp.qp_num. 714 */ 715 spin_lock_init(&qp->s_lock); 716 spin_lock_init(&qp->r_rq.lock); 717 atomic_set(&qp->refcount, 0); 718 init_waitqueue_head(&qp->wait); 719 tasklet_init(&qp->s_task, ipath_do_ruc_send, 720 (unsigned long)qp); 721 INIT_LIST_HEAD(&qp->piowait); 722 INIT_LIST_HEAD(&qp->timerwait); 723 qp->state = IB_QPS_RESET; 724 qp->s_wq = swq; 725 qp->s_size = init_attr->cap.max_send_wr + 1; 726 qp->s_max_sge = init_attr->cap.max_send_sge; 727 qp->r_rq.max_sge = init_attr->cap.max_recv_sge; 728 qp->s_flags = init_attr->sq_sig_type == IB_SIGNAL_REQ_WR ? 729 1 << IPATH_S_SIGNAL_REQ_WR : 0; 730 dev = to_idev(ibpd->device); 731 err = ipath_alloc_qpn(&dev->qp_table, qp, 732 init_attr->qp_type); 733 if (err) { 734 vfree(swq); 735 vfree(qp->r_rq.wq); 736 kfree(qp); 737 ret = ERR_PTR(err); 738 goto bail; 739 } 740 ipath_reset_qp(qp); 741 742 /* Tell the core driver that the kernel SMA is present. */ 743 if (init_attr->qp_type == IB_QPT_SMI) 744 ipath_layer_set_verbs_flags(dev->dd, 745 IPATH_VERBS_KERNEL_SMA); 746 break; 747 748 default: 749 /* Don't support raw QPs */ 750 ret = ERR_PTR(-ENOSYS); 751 goto bail; 752 } 753 754 init_attr->cap.max_inline_data = 0; 755 756 ret = &qp->ibqp; 757 758bail: 759 return ret; 760} 761 762/** 763 * ipath_destroy_qp - destroy a queue pair 764 * @ibqp: the queue pair to destroy 765 * 766 * Returns 0 on success. 767 * 768 * Note that this can be called while the QP is actively sending or 769 * receiving! 770 */ 771int ipath_destroy_qp(struct ib_qp *ibqp) 772{ 773 struct ipath_qp *qp = to_iqp(ibqp); 774 struct ipath_ibdev *dev = to_idev(ibqp->device); 775 unsigned long flags; 776 777 /* Tell the core driver that the kernel SMA is gone. */ 778 if (qp->ibqp.qp_type == IB_QPT_SMI) 779 ipath_layer_set_verbs_flags(dev->dd, 0); 780 781 spin_lock_irqsave(&qp->r_rq.lock, flags); 782 spin_lock(&qp->s_lock); 783 qp->state = IB_QPS_ERR; 784 spin_unlock(&qp->s_lock); 785 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 786 787 /* Stop the sending tasklet. */ 788 tasklet_kill(&qp->s_task); 789 790 /* Make sure the QP isn't on the timeout list. */ 791 spin_lock_irqsave(&dev->pending_lock, flags); 792 if (!list_empty(&qp->timerwait)) 793 list_del_init(&qp->timerwait); 794 if (!list_empty(&qp->piowait)) 795 list_del_init(&qp->piowait); 796 spin_unlock_irqrestore(&dev->pending_lock, flags); 797 798 /* 799 * Make sure that the QP is not in the QPN table so receive 800 * interrupts will discard packets for this QP. XXX Also remove QP 801 * from multicast table. 802 */ 803 if (atomic_read(&qp->refcount) != 0) 804 ipath_free_qp(&dev->qp_table, qp); 805 806 vfree(qp->s_wq); 807 vfree(qp->r_rq.wq); 808 kfree(qp); 809 return 0; 810} 811 812/** 813 * ipath_init_qp_table - initialize the QP table for a device 814 * @idev: the device who's QP table we're initializing 815 * @size: the size of the QP table 816 * 817 * Returns 0 on success, otherwise returns an errno. 818 */ 819int ipath_init_qp_table(struct ipath_ibdev *idev, int size) 820{ 821 int i; 822 int ret; 823 824 idev->qp_table.last = 1; /* QPN 0 and 1 are special. */ 825 idev->qp_table.max = size; 826 idev->qp_table.nmaps = 1; 827 idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table), 828 GFP_KERNEL); 829 if (idev->qp_table.table == NULL) { 830 ret = -ENOMEM; 831 goto bail; 832 } 833 834 for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) { 835 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE); 836 idev->qp_table.map[i].page = NULL; 837 } 838 839 ret = 0; 840 841bail: 842 return ret; 843} 844 845/** 846 * ipath_sqerror_qp - put a QP's send queue into an error state 847 * @qp: QP who's send queue will be put into an error state 848 * @wc: the WC responsible for putting the QP in this state 849 * 850 * Flushes the send work queue. 851 * The QP s_lock should be held. 852 */ 853 854void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc) 855{ 856 struct ipath_ibdev *dev = to_idev(qp->ibqp.device); 857 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last); 858 859 _VERBS_INFO("Send queue error on QP%d/%d: err: %d\n", 860 qp->ibqp.qp_num, qp->remote_qpn, wc->status); 861 862 spin_lock(&dev->pending_lock); 863 /* XXX What if its already removed by the timeout code? */ 864 if (!list_empty(&qp->timerwait)) 865 list_del_init(&qp->timerwait); 866 if (!list_empty(&qp->piowait)) 867 list_del_init(&qp->piowait); 868 spin_unlock(&dev->pending_lock); 869 870 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1); 871 if (++qp->s_last >= qp->s_size) 872 qp->s_last = 0; 873 874 wc->status = IB_WC_WR_FLUSH_ERR; 875 876 while (qp->s_last != qp->s_head) { 877 wc->wr_id = wqe->wr.wr_id; 878 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode]; 879 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1); 880 if (++qp->s_last >= qp->s_size) 881 qp->s_last = 0; 882 wqe = get_swqe_ptr(qp, qp->s_last); 883 } 884 qp->s_cur = qp->s_tail = qp->s_head; 885 qp->state = IB_QPS_SQE; 886} 887 888/** 889 * ipath_get_credit - flush the send work queue of a QP 890 * @qp: the qp who's send work queue to flush 891 * @aeth: the Acknowledge Extended Transport Header 892 * 893 * The QP s_lock should be held. 894 */ 895void ipath_get_credit(struct ipath_qp *qp, u32 aeth) 896{ 897 u32 credit = (aeth >> IPS_AETH_CREDIT_SHIFT) & IPS_AETH_CREDIT_MASK; 898 899 /* 900 * If the credit is invalid, we can send 901 * as many packets as we like. Otherwise, we have to 902 * honor the credit field. 903 */ 904 if (credit == IPS_AETH_CREDIT_INVAL) 905 qp->s_lsn = (u32) -1; 906 else if (qp->s_lsn != (u32) -1) { 907 /* Compute new LSN (i.e., MSN + credit) */ 908 credit = (aeth + credit_table[credit]) & IPS_MSN_MASK; 909 if (ipath_cmp24(credit, qp->s_lsn) > 0) 910 qp->s_lsn = credit; 911 } 912 913 /* Restart sending if it was blocked due to lack of credits. */ 914 if (qp->s_cur != qp->s_head && 915 (qp->s_lsn == (u32) -1 || 916 ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn, 917 qp->s_lsn + 1) <= 0)) 918 tasklet_hi_schedule(&qp->s_task); 919} 920