ipath_qp.c revision 6b66b2da1e821181a001c00b04a807724ad803cd
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 "ipath_kernel.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/* 48 * Convert the AETH credit code into the number of credits. 49 */ 50static u32 credit_table[31] = { 51 0, /* 0 */ 52 1, /* 1 */ 53 2, /* 2 */ 54 3, /* 3 */ 55 4, /* 4 */ 56 6, /* 5 */ 57 8, /* 6 */ 58 12, /* 7 */ 59 16, /* 8 */ 60 24, /* 9 */ 61 32, /* A */ 62 48, /* B */ 63 64, /* C */ 64 96, /* D */ 65 128, /* E */ 66 192, /* F */ 67 256, /* 10 */ 68 384, /* 11 */ 69 512, /* 12 */ 70 768, /* 13 */ 71 1024, /* 14 */ 72 1536, /* 15 */ 73 2048, /* 16 */ 74 3072, /* 17 */ 75 4096, /* 18 */ 76 6144, /* 19 */ 77 8192, /* 1A */ 78 12288, /* 1B */ 79 16384, /* 1C */ 80 24576, /* 1D */ 81 32768 /* 1E */ 82}; 83 84 85static void get_map_page(struct ipath_qp_table *qpt, struct qpn_map *map) 86{ 87 unsigned long page = get_zeroed_page(GFP_KERNEL); 88 unsigned long flags; 89 90 /* 91 * Free the page if someone raced with us installing it. 92 */ 93 94 spin_lock_irqsave(&qpt->lock, flags); 95 if (map->page) 96 free_page(page); 97 else 98 map->page = (void *)page; 99 spin_unlock_irqrestore(&qpt->lock, flags); 100} 101 102 103static int alloc_qpn(struct ipath_qp_table *qpt, enum ib_qp_type type) 104{ 105 u32 i, offset, max_scan, qpn; 106 struct qpn_map *map; 107 u32 ret = -1; 108 109 if (type == IB_QPT_SMI) 110 ret = 0; 111 else if (type == IB_QPT_GSI) 112 ret = 1; 113 114 if (ret != -1) { 115 map = &qpt->map[0]; 116 if (unlikely(!map->page)) { 117 get_map_page(qpt, map); 118 if (unlikely(!map->page)) { 119 ret = -ENOMEM; 120 goto bail; 121 } 122 } 123 if (!test_and_set_bit(ret, map->page)) 124 atomic_dec(&map->n_free); 125 else 126 ret = -EBUSY; 127 goto bail; 128 } 129 130 qpn = qpt->last + 1; 131 if (qpn >= QPN_MAX) 132 qpn = 2; 133 offset = qpn & BITS_PER_PAGE_MASK; 134 map = &qpt->map[qpn / BITS_PER_PAGE]; 135 max_scan = qpt->nmaps - !offset; 136 for (i = 0;;) { 137 if (unlikely(!map->page)) { 138 get_map_page(qpt, map); 139 if (unlikely(!map->page)) 140 break; 141 } 142 if (likely(atomic_read(&map->n_free))) { 143 do { 144 if (!test_and_set_bit(offset, map->page)) { 145 atomic_dec(&map->n_free); 146 qpt->last = qpn; 147 ret = qpn; 148 goto bail; 149 } 150 offset = find_next_offset(map, offset); 151 qpn = mk_qpn(qpt, map, offset); 152 /* 153 * This test differs from alloc_pidmap(). 154 * If find_next_offset() does find a zero 155 * bit, we don't need to check for QPN 156 * wrapping around past our starting QPN. 157 * We just need to be sure we don't loop 158 * forever. 159 */ 160 } while (offset < BITS_PER_PAGE && qpn < QPN_MAX); 161 } 162 /* 163 * In order to keep the number of pages allocated to a 164 * minimum, we scan the all existing pages before increasing 165 * the size of the bitmap table. 166 */ 167 if (++i > max_scan) { 168 if (qpt->nmaps == QPNMAP_ENTRIES) 169 break; 170 map = &qpt->map[qpt->nmaps++]; 171 offset = 0; 172 } else if (map < &qpt->map[qpt->nmaps]) { 173 ++map; 174 offset = 0; 175 } else { 176 map = &qpt->map[0]; 177 offset = 2; 178 } 179 qpn = mk_qpn(qpt, map, offset); 180 } 181 182 ret = -ENOMEM; 183 184bail: 185 return ret; 186} 187 188static void free_qpn(struct ipath_qp_table *qpt, u32 qpn) 189{ 190 struct qpn_map *map; 191 192 map = qpt->map + qpn / BITS_PER_PAGE; 193 if (map->page) 194 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page); 195 atomic_inc(&map->n_free); 196} 197 198/** 199 * ipath_alloc_qpn - allocate a QP number 200 * @qpt: the QP table 201 * @qp: the QP 202 * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special) 203 * 204 * Allocate the next available QPN and put the QP into the hash table. 205 * The hash table holds a reference to the QP. 206 */ 207static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp, 208 enum ib_qp_type type) 209{ 210 unsigned long flags; 211 int ret; 212 213 ret = alloc_qpn(qpt, type); 214 if (ret < 0) 215 goto bail; 216 qp->ibqp.qp_num = ret; 217 218 /* Add the QP to the hash table. */ 219 spin_lock_irqsave(&qpt->lock, flags); 220 221 ret %= qpt->max; 222 qp->next = qpt->table[ret]; 223 qpt->table[ret] = qp; 224 atomic_inc(&qp->refcount); 225 226 spin_unlock_irqrestore(&qpt->lock, flags); 227 ret = 0; 228 229bail: 230 return ret; 231} 232 233/** 234 * ipath_free_qp - remove a QP from the QP table 235 * @qpt: the QP table 236 * @qp: the QP to remove 237 * 238 * Remove the QP from the table so it can't be found asynchronously by 239 * the receive interrupt routine. 240 */ 241static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp) 242{ 243 struct ipath_qp *q, **qpp; 244 unsigned long flags; 245 int fnd = 0; 246 247 spin_lock_irqsave(&qpt->lock, flags); 248 249 /* Remove QP from the hash table. */ 250 qpp = &qpt->table[qp->ibqp.qp_num % qpt->max]; 251 for (; (q = *qpp) != NULL; qpp = &q->next) { 252 if (q == qp) { 253 *qpp = qp->next; 254 qp->next = NULL; 255 atomic_dec(&qp->refcount); 256 fnd = 1; 257 break; 258 } 259 } 260 261 spin_unlock_irqrestore(&qpt->lock, flags); 262 263 if (!fnd) 264 return; 265 266 free_qpn(qpt, qp->ibqp.qp_num); 267 268 wait_event(qp->wait, !atomic_read(&qp->refcount)); 269} 270 271/** 272 * ipath_free_all_qps - remove all QPs from the table 273 * @qpt: the QP table to empty 274 */ 275void ipath_free_all_qps(struct ipath_qp_table *qpt) 276{ 277 unsigned long flags; 278 struct ipath_qp *qp, *nqp; 279 u32 n; 280 281 for (n = 0; n < qpt->max; n++) { 282 spin_lock_irqsave(&qpt->lock, flags); 283 qp = qpt->table[n]; 284 qpt->table[n] = NULL; 285 spin_unlock_irqrestore(&qpt->lock, flags); 286 287 while (qp) { 288 nqp = qp->next; 289 free_qpn(qpt, qp->ibqp.qp_num); 290 if (!atomic_dec_and_test(&qp->refcount) || 291 !ipath_destroy_qp(&qp->ibqp)) 292 ipath_dbg("QP memory leak!\n"); 293 qp = nqp; 294 } 295 } 296 297 for (n = 0; n < ARRAY_SIZE(qpt->map); n++) { 298 if (qpt->map[n].page) 299 free_page((unsigned long)qpt->map[n].page); 300 } 301} 302 303/** 304 * ipath_lookup_qpn - return the QP with the given QPN 305 * @qpt: the QP table 306 * @qpn: the QP number to look up 307 * 308 * The caller is responsible for decrementing the QP reference count 309 * when done. 310 */ 311struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn) 312{ 313 unsigned long flags; 314 struct ipath_qp *qp; 315 316 spin_lock_irqsave(&qpt->lock, flags); 317 318 for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) { 319 if (qp->ibqp.qp_num == qpn) { 320 atomic_inc(&qp->refcount); 321 break; 322 } 323 } 324 325 spin_unlock_irqrestore(&qpt->lock, flags); 326 return qp; 327} 328 329/** 330 * ipath_reset_qp - initialize the QP state to the reset state 331 * @qp: the QP to reset 332 */ 333static void ipath_reset_qp(struct ipath_qp *qp) 334{ 335 qp->remote_qpn = 0; 336 qp->qkey = 0; 337 qp->qp_access_flags = 0; 338 qp->s_busy = 0; 339 qp->s_flags &= ~IPATH_S_SIGNAL_REQ_WR; 340 qp->s_hdrwords = 0; 341 qp->s_psn = 0; 342 qp->r_psn = 0; 343 qp->r_msn = 0; 344 if (qp->ibqp.qp_type == IB_QPT_RC) { 345 qp->s_state = IB_OPCODE_RC_SEND_LAST; 346 qp->r_state = IB_OPCODE_RC_SEND_LAST; 347 } else { 348 qp->s_state = IB_OPCODE_UC_SEND_LAST; 349 qp->r_state = IB_OPCODE_UC_SEND_LAST; 350 } 351 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 352 qp->r_nak_state = 0; 353 qp->r_wrid_valid = 0; 354 qp->s_rnr_timeout = 0; 355 qp->s_head = 0; 356 qp->s_tail = 0; 357 qp->s_cur = 0; 358 qp->s_last = 0; 359 qp->s_ssn = 1; 360 qp->s_lsn = 0; 361 qp->s_wait_credit = 0; 362 memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue)); 363 qp->r_head_ack_queue = 0; 364 qp->s_tail_ack_queue = 0; 365 qp->s_num_rd_atomic = 0; 366 if (qp->r_rq.wq) { 367 qp->r_rq.wq->head = 0; 368 qp->r_rq.wq->tail = 0; 369 } 370 qp->r_reuse_sge = 0; 371} 372 373/** 374 * ipath_error_qp - put a QP into an error state 375 * @qp: the QP to put into an error state 376 * @err: the receive completion error to signal if a RWQE is active 377 * 378 * Flushes both send and receive work queues. 379 * The QP s_lock should be held and interrupts disabled. 380 */ 381 382void ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err) 383{ 384 struct ipath_ibdev *dev = to_idev(qp->ibqp.device); 385 struct ib_wc wc; 386 387 ipath_dbg("QP%d/%d in error state\n", 388 qp->ibqp.qp_num, qp->remote_qpn); 389 390 spin_lock(&dev->pending_lock); 391 /* XXX What if its already removed by the timeout code? */ 392 if (!list_empty(&qp->timerwait)) 393 list_del_init(&qp->timerwait); 394 if (!list_empty(&qp->piowait)) 395 list_del_init(&qp->piowait); 396 spin_unlock(&dev->pending_lock); 397 398 wc.vendor_err = 0; 399 wc.byte_len = 0; 400 wc.imm_data = 0; 401 wc.qp = &qp->ibqp; 402 wc.src_qp = 0; 403 wc.wc_flags = 0; 404 wc.pkey_index = 0; 405 wc.slid = 0; 406 wc.sl = 0; 407 wc.dlid_path_bits = 0; 408 wc.port_num = 0; 409 if (qp->r_wrid_valid) { 410 qp->r_wrid_valid = 0; 411 wc.wr_id = qp->r_wr_id; 412 wc.opcode = IB_WC_RECV; 413 wc.status = err; 414 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1); 415 } 416 wc.status = IB_WC_WR_FLUSH_ERR; 417 418 while (qp->s_last != qp->s_head) { 419 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last); 420 421 wc.wr_id = wqe->wr.wr_id; 422 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode]; 423 if (++qp->s_last >= qp->s_size) 424 qp->s_last = 0; 425 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1); 426 } 427 qp->s_cur = qp->s_tail = qp->s_head; 428 qp->s_hdrwords = 0; 429 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 430 431 if (qp->r_rq.wq) { 432 struct ipath_rwq *wq; 433 u32 head; 434 u32 tail; 435 436 spin_lock(&qp->r_rq.lock); 437 438 /* sanity check pointers before trusting them */ 439 wq = qp->r_rq.wq; 440 head = wq->head; 441 if (head >= qp->r_rq.size) 442 head = 0; 443 tail = wq->tail; 444 if (tail >= qp->r_rq.size) 445 tail = 0; 446 wc.opcode = IB_WC_RECV; 447 while (tail != head) { 448 wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id; 449 if (++tail >= qp->r_rq.size) 450 tail = 0; 451 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1); 452 } 453 wq->tail = tail; 454 455 spin_unlock(&qp->r_rq.lock); 456 } 457} 458 459/** 460 * ipath_modify_qp - modify the attributes of a queue pair 461 * @ibqp: the queue pair who's attributes we're modifying 462 * @attr: the new attributes 463 * @attr_mask: the mask of attributes to modify 464 * @udata: user data for ipathverbs.so 465 * 466 * Returns 0 on success, otherwise returns an errno. 467 */ 468int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 469 int attr_mask, struct ib_udata *udata) 470{ 471 struct ipath_ibdev *dev = to_idev(ibqp->device); 472 struct ipath_qp *qp = to_iqp(ibqp); 473 enum ib_qp_state cur_state, new_state; 474 unsigned long flags; 475 int ret; 476 477 spin_lock_irqsave(&qp->s_lock, flags); 478 479 cur_state = attr_mask & IB_QP_CUR_STATE ? 480 attr->cur_qp_state : qp->state; 481 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 482 483 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 484 attr_mask)) 485 goto inval; 486 487 if (attr_mask & IB_QP_AV) { 488 if (attr->ah_attr.dlid == 0 || 489 attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE) 490 goto inval; 491 492 if ((attr->ah_attr.ah_flags & IB_AH_GRH) && 493 (attr->ah_attr.grh.sgid_index > 1)) 494 goto inval; 495 } 496 497 if (attr_mask & IB_QP_PKEY_INDEX) 498 if (attr->pkey_index >= ipath_get_npkeys(dev->dd)) 499 goto inval; 500 501 if (attr_mask & IB_QP_MIN_RNR_TIMER) 502 if (attr->min_rnr_timer > 31) 503 goto inval; 504 505 if (attr_mask & IB_QP_PORT) 506 if (attr->port_num == 0 || 507 attr->port_num > ibqp->device->phys_port_cnt) 508 goto inval; 509 510 if (attr_mask & IB_QP_PATH_MTU) 511 if (attr->path_mtu > IB_MTU_4096) 512 goto inval; 513 514 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 515 if (attr->max_dest_rd_atomic > 1) 516 goto inval; 517 518 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) 519 if (attr->max_rd_atomic > 1) 520 goto inval; 521 522 if (attr_mask & IB_QP_PATH_MIG_STATE) 523 if (attr->path_mig_state != IB_MIG_MIGRATED && 524 attr->path_mig_state != IB_MIG_REARM) 525 goto inval; 526 527 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 528 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC) 529 goto inval; 530 531 switch (new_state) { 532 case IB_QPS_RESET: 533 ipath_reset_qp(qp); 534 break; 535 536 case IB_QPS_ERR: 537 ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR); 538 break; 539 540 default: 541 break; 542 543 } 544 545 if (attr_mask & IB_QP_PKEY_INDEX) 546 qp->s_pkey_index = attr->pkey_index; 547 548 if (attr_mask & IB_QP_DEST_QPN) 549 qp->remote_qpn = attr->dest_qp_num; 550 551 if (attr_mask & IB_QP_SQ_PSN) { 552 qp->s_psn = qp->s_next_psn = attr->sq_psn; 553 qp->s_last_psn = qp->s_next_psn - 1; 554 } 555 556 if (attr_mask & IB_QP_RQ_PSN) 557 qp->r_psn = attr->rq_psn; 558 559 if (attr_mask & IB_QP_ACCESS_FLAGS) 560 qp->qp_access_flags = attr->qp_access_flags; 561 562 if (attr_mask & IB_QP_AV) 563 qp->remote_ah_attr = attr->ah_attr; 564 565 if (attr_mask & IB_QP_PATH_MTU) 566 qp->path_mtu = attr->path_mtu; 567 568 if (attr_mask & IB_QP_RETRY_CNT) 569 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt; 570 571 if (attr_mask & IB_QP_RNR_RETRY) { 572 qp->s_rnr_retry = attr->rnr_retry; 573 if (qp->s_rnr_retry > 7) 574 qp->s_rnr_retry = 7; 575 qp->s_rnr_retry_cnt = qp->s_rnr_retry; 576 } 577 578 if (attr_mask & IB_QP_MIN_RNR_TIMER) 579 qp->r_min_rnr_timer = attr->min_rnr_timer; 580 581 if (attr_mask & IB_QP_TIMEOUT) 582 qp->timeout = attr->timeout; 583 584 if (attr_mask & IB_QP_QKEY) 585 qp->qkey = attr->qkey; 586 587 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 588 qp->r_max_rd_atomic = attr->max_dest_rd_atomic; 589 590 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) 591 qp->s_max_rd_atomic = attr->max_rd_atomic; 592 593 qp->state = new_state; 594 spin_unlock_irqrestore(&qp->s_lock, flags); 595 596 ret = 0; 597 goto bail; 598 599inval: 600 spin_unlock_irqrestore(&qp->s_lock, flags); 601 ret = -EINVAL; 602 603bail: 604 return ret; 605} 606 607int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 608 int attr_mask, struct ib_qp_init_attr *init_attr) 609{ 610 struct ipath_qp *qp = to_iqp(ibqp); 611 612 attr->qp_state = qp->state; 613 attr->cur_qp_state = attr->qp_state; 614 attr->path_mtu = qp->path_mtu; 615 attr->path_mig_state = 0; 616 attr->qkey = qp->qkey; 617 attr->rq_psn = qp->r_psn; 618 attr->sq_psn = qp->s_next_psn; 619 attr->dest_qp_num = qp->remote_qpn; 620 attr->qp_access_flags = qp->qp_access_flags; 621 attr->cap.max_send_wr = qp->s_size - 1; 622 attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1; 623 attr->cap.max_send_sge = qp->s_max_sge; 624 attr->cap.max_recv_sge = qp->r_rq.max_sge; 625 attr->cap.max_inline_data = 0; 626 attr->ah_attr = qp->remote_ah_attr; 627 memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr)); 628 attr->pkey_index = qp->s_pkey_index; 629 attr->alt_pkey_index = 0; 630 attr->en_sqd_async_notify = 0; 631 attr->sq_draining = 0; 632 attr->max_rd_atomic = qp->s_max_rd_atomic; 633 attr->max_dest_rd_atomic = qp->r_max_rd_atomic; 634 attr->min_rnr_timer = qp->r_min_rnr_timer; 635 attr->port_num = 1; 636 attr->timeout = qp->timeout; 637 attr->retry_cnt = qp->s_retry_cnt; 638 attr->rnr_retry = qp->s_rnr_retry; 639 attr->alt_port_num = 0; 640 attr->alt_timeout = 0; 641 642 init_attr->event_handler = qp->ibqp.event_handler; 643 init_attr->qp_context = qp->ibqp.qp_context; 644 init_attr->send_cq = qp->ibqp.send_cq; 645 init_attr->recv_cq = qp->ibqp.recv_cq; 646 init_attr->srq = qp->ibqp.srq; 647 init_attr->cap = attr->cap; 648 if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR) 649 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR; 650 else 651 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR; 652 init_attr->qp_type = qp->ibqp.qp_type; 653 init_attr->port_num = 1; 654 return 0; 655} 656 657/** 658 * ipath_compute_aeth - compute the AETH (syndrome + MSN) 659 * @qp: the queue pair to compute the AETH for 660 * 661 * Returns the AETH. 662 */ 663__be32 ipath_compute_aeth(struct ipath_qp *qp) 664{ 665 u32 aeth = qp->r_msn & IPATH_MSN_MASK; 666 667 if (qp->ibqp.srq) { 668 /* 669 * Shared receive queues don't generate credits. 670 * Set the credit field to the invalid value. 671 */ 672 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT; 673 } else { 674 u32 min, max, x; 675 u32 credits; 676 struct ipath_rwq *wq = qp->r_rq.wq; 677 u32 head; 678 u32 tail; 679 680 /* sanity check pointers before trusting them */ 681 head = wq->head; 682 if (head >= qp->r_rq.size) 683 head = 0; 684 tail = wq->tail; 685 if (tail >= qp->r_rq.size) 686 tail = 0; 687 /* 688 * Compute the number of credits available (RWQEs). 689 * XXX Not holding the r_rq.lock here so there is a small 690 * chance that the pair of reads are not atomic. 691 */ 692 credits = head - tail; 693 if ((int)credits < 0) 694 credits += qp->r_rq.size; 695 /* 696 * Binary search the credit table to find the code to 697 * use. 698 */ 699 min = 0; 700 max = 31; 701 for (;;) { 702 x = (min + max) / 2; 703 if (credit_table[x] == credits) 704 break; 705 if (credit_table[x] > credits) 706 max = x; 707 else if (min == x) 708 break; 709 else 710 min = x; 711 } 712 aeth |= x << IPATH_AETH_CREDIT_SHIFT; 713 } 714 return cpu_to_be32(aeth); 715} 716 717/** 718 * ipath_create_qp - create a queue pair for a device 719 * @ibpd: the protection domain who's device we create the queue pair for 720 * @init_attr: the attributes of the queue pair 721 * @udata: unused by InfiniPath 722 * 723 * Returns the queue pair on success, otherwise returns an errno. 724 * 725 * Called by the ib_create_qp() core verbs function. 726 */ 727struct ib_qp *ipath_create_qp(struct ib_pd *ibpd, 728 struct ib_qp_init_attr *init_attr, 729 struct ib_udata *udata) 730{ 731 struct ipath_qp *qp; 732 int err; 733 struct ipath_swqe *swq = NULL; 734 struct ipath_ibdev *dev; 735 size_t sz; 736 struct ib_qp *ret; 737 738 if (init_attr->cap.max_send_sge > ib_ipath_max_sges || 739 init_attr->cap.max_recv_sge > ib_ipath_max_sges || 740 init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs || 741 init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) { 742 ret = ERR_PTR(-ENOMEM); 743 goto bail; 744 } 745 746 if (init_attr->cap.max_send_sge + 747 init_attr->cap.max_recv_sge + 748 init_attr->cap.max_send_wr + 749 init_attr->cap.max_recv_wr == 0) { 750 ret = ERR_PTR(-EINVAL); 751 goto bail; 752 } 753 754 switch (init_attr->qp_type) { 755 case IB_QPT_UC: 756 case IB_QPT_RC: 757 sz = sizeof(struct ipath_sge) * 758 init_attr->cap.max_send_sge + 759 sizeof(struct ipath_swqe); 760 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz); 761 if (swq == NULL) { 762 ret = ERR_PTR(-ENOMEM); 763 goto bail; 764 } 765 /* FALLTHROUGH */ 766 case IB_QPT_UD: 767 case IB_QPT_SMI: 768 case IB_QPT_GSI: 769 sz = sizeof(*qp); 770 if (init_attr->srq) { 771 struct ipath_srq *srq = to_isrq(init_attr->srq); 772 773 sz += sizeof(*qp->r_sg_list) * 774 srq->rq.max_sge; 775 } else 776 sz += sizeof(*qp->r_sg_list) * 777 init_attr->cap.max_recv_sge; 778 qp = kmalloc(sz, GFP_KERNEL); 779 if (!qp) { 780 ret = ERR_PTR(-ENOMEM); 781 goto bail_swq; 782 } 783 if (init_attr->srq) { 784 sz = 0; 785 qp->r_rq.size = 0; 786 qp->r_rq.max_sge = 0; 787 qp->r_rq.wq = NULL; 788 init_attr->cap.max_recv_wr = 0; 789 init_attr->cap.max_recv_sge = 0; 790 } else { 791 qp->r_rq.size = init_attr->cap.max_recv_wr + 1; 792 qp->r_rq.max_sge = init_attr->cap.max_recv_sge; 793 sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + 794 sizeof(struct ipath_rwqe); 795 qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) + 796 qp->r_rq.size * sz); 797 if (!qp->r_rq.wq) { 798 ret = ERR_PTR(-ENOMEM); 799 goto bail_qp; 800 } 801 } 802 803 /* 804 * ib_create_qp() will initialize qp->ibqp 805 * except for qp->ibqp.qp_num. 806 */ 807 spin_lock_init(&qp->s_lock); 808 spin_lock_init(&qp->r_rq.lock); 809 atomic_set(&qp->refcount, 0); 810 init_waitqueue_head(&qp->wait); 811 tasklet_init(&qp->s_task, ipath_do_ruc_send, 812 (unsigned long)qp); 813 INIT_LIST_HEAD(&qp->piowait); 814 INIT_LIST_HEAD(&qp->timerwait); 815 qp->state = IB_QPS_RESET; 816 qp->s_wq = swq; 817 qp->s_size = init_attr->cap.max_send_wr + 1; 818 qp->s_max_sge = init_attr->cap.max_send_sge; 819 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) 820 qp->s_flags = IPATH_S_SIGNAL_REQ_WR; 821 else 822 qp->s_flags = 0; 823 dev = to_idev(ibpd->device); 824 err = ipath_alloc_qpn(&dev->qp_table, qp, 825 init_attr->qp_type); 826 if (err) { 827 ret = ERR_PTR(err); 828 goto bail_rwq; 829 } 830 qp->ip = NULL; 831 ipath_reset_qp(qp); 832 break; 833 834 default: 835 /* Don't support raw QPs */ 836 ret = ERR_PTR(-ENOSYS); 837 goto bail; 838 } 839 840 init_attr->cap.max_inline_data = 0; 841 842 /* 843 * Return the address of the RWQ as the offset to mmap. 844 * See ipath_mmap() for details. 845 */ 846 if (udata && udata->outlen >= sizeof(__u64)) { 847 int err; 848 849 if (!qp->r_rq.wq) { 850 __u64 offset = 0; 851 852 err = ib_copy_to_udata(udata, &offset, 853 sizeof(offset)); 854 if (err) { 855 ret = ERR_PTR(err); 856 goto bail_rwq; 857 } 858 } else { 859 u32 s = sizeof(struct ipath_rwq) + 860 qp->r_rq.size * sz; 861 862 qp->ip = 863 ipath_create_mmap_info(dev, s, 864 ibpd->uobject->context, 865 qp->r_rq.wq); 866 if (!qp->ip) { 867 ret = ERR_PTR(-ENOMEM); 868 goto bail_rwq; 869 } 870 871 err = ib_copy_to_udata(udata, &(qp->ip->offset), 872 sizeof(qp->ip->offset)); 873 if (err) { 874 ret = ERR_PTR(err); 875 goto bail_ip; 876 } 877 } 878 } 879 880 spin_lock(&dev->n_qps_lock); 881 if (dev->n_qps_allocated == ib_ipath_max_qps) { 882 spin_unlock(&dev->n_qps_lock); 883 ret = ERR_PTR(-ENOMEM); 884 goto bail_ip; 885 } 886 887 dev->n_qps_allocated++; 888 spin_unlock(&dev->n_qps_lock); 889 890 if (qp->ip) { 891 spin_lock_irq(&dev->pending_lock); 892 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps); 893 spin_unlock_irq(&dev->pending_lock); 894 } 895 896 ret = &qp->ibqp; 897 goto bail; 898 899bail_ip: 900 kfree(qp->ip); 901bail_rwq: 902 vfree(qp->r_rq.wq); 903bail_qp: 904 kfree(qp); 905bail_swq: 906 vfree(swq); 907bail: 908 return ret; 909} 910 911/** 912 * ipath_destroy_qp - destroy a queue pair 913 * @ibqp: the queue pair to destroy 914 * 915 * Returns 0 on success. 916 * 917 * Note that this can be called while the QP is actively sending or 918 * receiving! 919 */ 920int ipath_destroy_qp(struct ib_qp *ibqp) 921{ 922 struct ipath_qp *qp = to_iqp(ibqp); 923 struct ipath_ibdev *dev = to_idev(ibqp->device); 924 unsigned long flags; 925 926 spin_lock_irqsave(&qp->s_lock, flags); 927 qp->state = IB_QPS_ERR; 928 spin_unlock_irqrestore(&qp->s_lock, flags); 929 spin_lock(&dev->n_qps_lock); 930 dev->n_qps_allocated--; 931 spin_unlock(&dev->n_qps_lock); 932 933 /* Stop the sending tasklet. */ 934 tasklet_kill(&qp->s_task); 935 936 /* Make sure the QP isn't on the timeout list. */ 937 spin_lock_irqsave(&dev->pending_lock, flags); 938 if (!list_empty(&qp->timerwait)) 939 list_del_init(&qp->timerwait); 940 if (!list_empty(&qp->piowait)) 941 list_del_init(&qp->piowait); 942 spin_unlock_irqrestore(&dev->pending_lock, flags); 943 944 /* 945 * Make sure that the QP is not in the QPN table so receive 946 * interrupts will discard packets for this QP. XXX Also remove QP 947 * from multicast table. 948 */ 949 if (atomic_read(&qp->refcount) != 0) 950 ipath_free_qp(&dev->qp_table, qp); 951 952 if (qp->ip) 953 kref_put(&qp->ip->ref, ipath_release_mmap_info); 954 else 955 vfree(qp->r_rq.wq); 956 vfree(qp->s_wq); 957 kfree(qp); 958 return 0; 959} 960 961/** 962 * ipath_init_qp_table - initialize the QP table for a device 963 * @idev: the device who's QP table we're initializing 964 * @size: the size of the QP table 965 * 966 * Returns 0 on success, otherwise returns an errno. 967 */ 968int ipath_init_qp_table(struct ipath_ibdev *idev, int size) 969{ 970 int i; 971 int ret; 972 973 idev->qp_table.last = 1; /* QPN 0 and 1 are special. */ 974 idev->qp_table.max = size; 975 idev->qp_table.nmaps = 1; 976 idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table), 977 GFP_KERNEL); 978 if (idev->qp_table.table == NULL) { 979 ret = -ENOMEM; 980 goto bail; 981 } 982 983 for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) { 984 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE); 985 idev->qp_table.map[i].page = NULL; 986 } 987 988 ret = 0; 989 990bail: 991 return ret; 992} 993 994/** 995 * ipath_sqerror_qp - put a QP's send queue into an error state 996 * @qp: QP who's send queue will be put into an error state 997 * @wc: the WC responsible for putting the QP in this state 998 * 999 * Flushes the send work queue. 1000 * The QP s_lock should be held and interrupts disabled. 1001 */ 1002 1003void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc) 1004{ 1005 struct ipath_ibdev *dev = to_idev(qp->ibqp.device); 1006 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last); 1007 1008 ipath_dbg("Send queue error on QP%d/%d: err: %d\n", 1009 qp->ibqp.qp_num, qp->remote_qpn, wc->status); 1010 1011 spin_lock(&dev->pending_lock); 1012 /* XXX What if its already removed by the timeout code? */ 1013 if (!list_empty(&qp->timerwait)) 1014 list_del_init(&qp->timerwait); 1015 if (!list_empty(&qp->piowait)) 1016 list_del_init(&qp->piowait); 1017 spin_unlock(&dev->pending_lock); 1018 1019 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1); 1020 if (++qp->s_last >= qp->s_size) 1021 qp->s_last = 0; 1022 1023 wc->status = IB_WC_WR_FLUSH_ERR; 1024 1025 while (qp->s_last != qp->s_head) { 1026 wqe = get_swqe_ptr(qp, qp->s_last); 1027 wc->wr_id = wqe->wr.wr_id; 1028 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode]; 1029 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1); 1030 if (++qp->s_last >= qp->s_size) 1031 qp->s_last = 0; 1032 } 1033 qp->s_cur = qp->s_tail = qp->s_head; 1034 qp->state = IB_QPS_SQE; 1035} 1036 1037/** 1038 * ipath_get_credit - flush the send work queue of a QP 1039 * @qp: the qp who's send work queue to flush 1040 * @aeth: the Acknowledge Extended Transport Header 1041 * 1042 * The QP s_lock should be held. 1043 */ 1044void ipath_get_credit(struct ipath_qp *qp, u32 aeth) 1045{ 1046 u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK; 1047 1048 /* 1049 * If the credit is invalid, we can send 1050 * as many packets as we like. Otherwise, we have to 1051 * honor the credit field. 1052 */ 1053 if (credit == IPATH_AETH_CREDIT_INVAL) 1054 qp->s_lsn = (u32) -1; 1055 else if (qp->s_lsn != (u32) -1) { 1056 /* Compute new LSN (i.e., MSN + credit) */ 1057 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK; 1058 if (ipath_cmp24(credit, qp->s_lsn) > 0) 1059 qp->s_lsn = credit; 1060 } 1061 1062 /* Restart sending if it was blocked due to lack of credits. */ 1063 if (qp->s_cur != qp->s_head && 1064 (qp->s_lsn == (u32) -1 || 1065 ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn, 1066 qp->s_lsn + 1) <= 0)) 1067 tasklet_hi_schedule(&qp->s_task); 1068} 1069