1 /* 2 * Copyright(c) 2016 Intel Corporation. 3 * 4 * This file is provided under a dual BSD/GPLv2 license. When using or 5 * redistributing this file, you may do so under either license. 6 * 7 * GPL LICENSE SUMMARY 8 * 9 * This program is free software; you can redistribute it and/or modify 10 * it under the terms of version 2 of the GNU General Public License as 11 * published by the Free Software Foundation. 12 * 13 * This program is distributed in the hope that it will be useful, but 14 * WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 16 * General Public License for more details. 17 * 18 * BSD LICENSE 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * - Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * - Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in 28 * the documentation and/or other materials provided with the 29 * distribution. 30 * - Neither the name of Intel Corporation nor the names of its 31 * contributors may be used to endorse or promote products derived 32 * from this software without specific prior written permission. 33 * 34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 45 * 46 */ 47 48 #include <linux/hash.h> 49 #include <linux/bitops.h> 50 #include <linux/lockdep.h> 51 #include <linux/vmalloc.h> 52 #include <linux/slab.h> 53 #include <rdma/ib_verbs.h> 54 #include "qp.h" 55 #include "vt.h" 56 #include "trace.h" 57 58 /* 59 * Note that it is OK to post send work requests in the SQE and ERR 60 * states; rvt_do_send() will process them and generate error 61 * completions as per IB 1.2 C10-96. 62 */ 63 const int ib_rvt_state_ops[IB_QPS_ERR + 1] = { 64 [IB_QPS_RESET] = 0, 65 [IB_QPS_INIT] = RVT_POST_RECV_OK, 66 [IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK, 67 [IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | 68 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK | 69 RVT_PROCESS_NEXT_SEND_OK, 70 [IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | 71 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK, 72 [IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | 73 RVT_POST_SEND_OK | RVT_FLUSH_SEND, 74 [IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV | 75 RVT_POST_SEND_OK | RVT_FLUSH_SEND, 76 }; 77 EXPORT_SYMBOL(ib_rvt_state_ops); 78 79 static void get_map_page(struct rvt_qpn_table *qpt, 80 struct rvt_qpn_map *map, 81 gfp_t gfp) 82 { 83 unsigned long page = get_zeroed_page(gfp); 84 85 /* 86 * Free the page if someone raced with us installing it. 87 */ 88 89 spin_lock(&qpt->lock); 90 if (map->page) 91 free_page(page); 92 else 93 map->page = (void *)page; 94 spin_unlock(&qpt->lock); 95 } 96 97 /** 98 * init_qpn_table - initialize the QP number table for a device 99 * @qpt: the QPN table 100 */ 101 static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt) 102 { 103 u32 offset, i; 104 struct rvt_qpn_map *map; 105 int ret = 0; 106 107 if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start)) 108 return -EINVAL; 109 110 spin_lock_init(&qpt->lock); 111 112 qpt->last = rdi->dparms.qpn_start; 113 qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift; 114 115 /* 116 * Drivers may want some QPs beyond what we need for verbs let them use 117 * our qpn table. No need for two. Lets go ahead and mark the bitmaps 118 * for those. The reserved range must be *after* the range which verbs 119 * will pick from. 120 */ 121 122 /* Figure out number of bit maps needed before reserved range */ 123 qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE; 124 125 /* This should always be zero */ 126 offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK; 127 128 /* Starting with the first reserved bit map */ 129 map = &qpt->map[qpt->nmaps]; 130 131 rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n", 132 rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end); 133 for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) { 134 if (!map->page) { 135 get_map_page(qpt, map, GFP_KERNEL); 136 if (!map->page) { 137 ret = -ENOMEM; 138 break; 139 } 140 } 141 set_bit(offset, map->page); 142 offset++; 143 if (offset == RVT_BITS_PER_PAGE) { 144 /* next page */ 145 qpt->nmaps++; 146 map++; 147 offset = 0; 148 } 149 } 150 return ret; 151 } 152 153 /** 154 * free_qpn_table - free the QP number table for a device 155 * @qpt: the QPN table 156 */ 157 static void free_qpn_table(struct rvt_qpn_table *qpt) 158 { 159 int i; 160 161 for (i = 0; i < ARRAY_SIZE(qpt->map); i++) 162 free_page((unsigned long)qpt->map[i].page); 163 } 164 165 int rvt_driver_qp_init(struct rvt_dev_info *rdi) 166 { 167 int i; 168 int ret = -ENOMEM; 169 170 if (rdi->flags & RVT_FLAG_QP_INIT_DRIVER) { 171 rvt_pr_info(rdi, "Driver is doing QP init.\n"); 172 return 0; 173 } 174 175 if (!rdi->dparms.qp_table_size) 176 return -EINVAL; 177 178 /* 179 * If driver is not doing any QP allocation then make sure it is 180 * providing the necessary QP functions. 181 */ 182 if (!rdi->driver_f.free_all_qps || 183 !rdi->driver_f.qp_priv_alloc || 184 !rdi->driver_f.qp_priv_free || 185 !rdi->driver_f.notify_qp_reset) 186 return -EINVAL; 187 188 /* allocate parent object */ 189 rdi->qp_dev = kzalloc(sizeof(*rdi->qp_dev), GFP_KERNEL); 190 if (!rdi->qp_dev) 191 return -ENOMEM; 192 193 /* allocate hash table */ 194 rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size; 195 rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size); 196 rdi->qp_dev->qp_table = 197 kmalloc(rdi->qp_dev->qp_table_size * 198 sizeof(*rdi->qp_dev->qp_table), 199 GFP_KERNEL); 200 if (!rdi->qp_dev->qp_table) 201 goto no_qp_table; 202 203 for (i = 0; i < rdi->qp_dev->qp_table_size; i++) 204 RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL); 205 206 spin_lock_init(&rdi->qp_dev->qpt_lock); 207 208 /* initialize qpn map */ 209 if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table)) 210 goto fail_table; 211 212 spin_lock_init(&rdi->n_qps_lock); 213 214 return 0; 215 216 fail_table: 217 kfree(rdi->qp_dev->qp_table); 218 free_qpn_table(&rdi->qp_dev->qpn_table); 219 220 no_qp_table: 221 kfree(rdi->qp_dev); 222 223 return ret; 224 } 225 226 /** 227 * free_all_qps - check for QPs still in use 228 * @qpt: the QP table to empty 229 * 230 * There should not be any QPs still in use. 231 * Free memory for table. 232 */ 233 static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi) 234 { 235 unsigned long flags; 236 struct rvt_qp *qp; 237 unsigned n, qp_inuse = 0; 238 spinlock_t *ql; /* work around too long line below */ 239 240 if (rdi->driver_f.free_all_qps) 241 qp_inuse = rdi->driver_f.free_all_qps(rdi); 242 243 qp_inuse += rvt_mcast_tree_empty(rdi); 244 245 if (!rdi->qp_dev) 246 return qp_inuse; 247 248 ql = &rdi->qp_dev->qpt_lock; 249 spin_lock_irqsave(ql, flags); 250 for (n = 0; n < rdi->qp_dev->qp_table_size; n++) { 251 qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n], 252 lockdep_is_held(ql)); 253 RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL); 254 255 for (; qp; qp = rcu_dereference_protected(qp->next, 256 lockdep_is_held(ql))) 257 qp_inuse++; 258 } 259 spin_unlock_irqrestore(ql, flags); 260 synchronize_rcu(); 261 return qp_inuse; 262 } 263 264 void rvt_qp_exit(struct rvt_dev_info *rdi) 265 { 266 u32 qps_inuse = rvt_free_all_qps(rdi); 267 268 if (qps_inuse) 269 rvt_pr_err(rdi, "QP memory leak! %u still in use\n", 270 qps_inuse); 271 if (!rdi->qp_dev) 272 return; 273 274 if (rdi->flags & RVT_FLAG_QP_INIT_DRIVER) 275 return; /* driver did the qp init so nothing else to do */ 276 277 kfree(rdi->qp_dev->qp_table); 278 free_qpn_table(&rdi->qp_dev->qpn_table); 279 kfree(rdi->qp_dev); 280 } 281 282 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt, 283 struct rvt_qpn_map *map, unsigned off) 284 { 285 return (map - qpt->map) * RVT_BITS_PER_PAGE + off; 286 } 287 288 /* 289 * Allocate the next available QPN or 290 * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI. 291 */ 292 static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, 293 enum ib_qp_type type, u8 port, gfp_t gfp) 294 { 295 u32 i, offset, max_scan, qpn; 296 struct rvt_qpn_map *map; 297 u32 ret; 298 299 if (rdi->driver_f.alloc_qpn) 300 return rdi->driver_f.alloc_qpn(rdi, qpt, type, port, 301 GFP_KERNEL); 302 303 if (type == IB_QPT_SMI || type == IB_QPT_GSI) { 304 unsigned n; 305 306 ret = type == IB_QPT_GSI; 307 n = 1 << (ret + 2 * (port - 1)); 308 spin_lock(&qpt->lock); 309 if (qpt->flags & n) 310 ret = -EINVAL; 311 else 312 qpt->flags |= n; 313 spin_unlock(&qpt->lock); 314 goto bail; 315 } 316 317 qpn = qpt->last + qpt->incr; 318 if (qpn >= RVT_QPN_MAX) 319 qpn = qpt->incr | ((qpt->last & 1) ^ 1); 320 /* offset carries bit 0 */ 321 offset = qpn & RVT_BITS_PER_PAGE_MASK; 322 map = &qpt->map[qpn / RVT_BITS_PER_PAGE]; 323 max_scan = qpt->nmaps - !offset; 324 for (i = 0;;) { 325 if (unlikely(!map->page)) { 326 get_map_page(qpt, map, gfp); 327 if (unlikely(!map->page)) 328 break; 329 } 330 do { 331 if (!test_and_set_bit(offset, map->page)) { 332 qpt->last = qpn; 333 ret = qpn; 334 goto bail; 335 } 336 offset += qpt->incr; 337 /* 338 * This qpn might be bogus if offset >= BITS_PER_PAGE. 339 * That is OK. It gets re-assigned below 340 */ 341 qpn = mk_qpn(qpt, map, offset); 342 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX); 343 /* 344 * In order to keep the number of pages allocated to a 345 * minimum, we scan the all existing pages before increasing 346 * the size of the bitmap table. 347 */ 348 if (++i > max_scan) { 349 if (qpt->nmaps == RVT_QPNMAP_ENTRIES) 350 break; 351 map = &qpt->map[qpt->nmaps++]; 352 /* start at incr with current bit 0 */ 353 offset = qpt->incr | (offset & 1); 354 } else if (map < &qpt->map[qpt->nmaps]) { 355 ++map; 356 /* start at incr with current bit 0 */ 357 offset = qpt->incr | (offset & 1); 358 } else { 359 map = &qpt->map[0]; 360 /* wrap to first map page, invert bit 0 */ 361 offset = qpt->incr | ((offset & 1) ^ 1); 362 } 363 /* there can be no bits at shift and below */ 364 WARN_ON(offset & (rdi->dparms.qos_shift - 1)); 365 qpn = mk_qpn(qpt, map, offset); 366 } 367 368 ret = -ENOMEM; 369 370 bail: 371 return ret; 372 } 373 374 static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn) 375 { 376 struct rvt_qpn_map *map; 377 378 map = qpt->map + qpn / RVT_BITS_PER_PAGE; 379 if (map->page) 380 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); 381 } 382 383 /** 384 * reset_qp - initialize the QP state to the reset state 385 * @qp: the QP to reset 386 * @type: the QP type 387 * r and s lock are required to be held by the caller 388 */ 389 void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, 390 enum ib_qp_type type) 391 { 392 if (qp->state != IB_QPS_RESET) { 393 qp->state = IB_QPS_RESET; 394 395 /* Let drivers flush their waitlist */ 396 rdi->driver_f.flush_qp_waiters(qp); 397 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT); 398 spin_unlock(&qp->s_lock); 399 spin_unlock_irq(&qp->r_lock); 400 401 /* Stop the send queue and the retry timer */ 402 rdi->driver_f.stop_send_queue(qp); 403 del_timer_sync(&qp->s_timer); 404 405 /* Wait for things to stop */ 406 rdi->driver_f.quiesce_qp(qp); 407 408 /* take qp out the hash and wait for it to be unused */ 409 rvt_remove_qp(rdi, qp); 410 wait_event(qp->wait, !atomic_read(&qp->refcount)); 411 412 /* grab the lock b/c it was locked at call time */ 413 spin_lock_irq(&qp->r_lock); 414 spin_lock(&qp->s_lock); 415 416 rvt_clear_mr_refs(qp, 1); 417 } 418 419 /* 420 * Let the driver do any tear down it needs to for a qp 421 * that has been reset 422 */ 423 rdi->driver_f.notify_qp_reset(qp); 424 425 qp->remote_qpn = 0; 426 qp->qkey = 0; 427 qp->qp_access_flags = 0; 428 qp->s_flags &= RVT_S_SIGNAL_REQ_WR; 429 qp->s_hdrwords = 0; 430 qp->s_wqe = NULL; 431 qp->s_draining = 0; 432 qp->s_next_psn = 0; 433 qp->s_last_psn = 0; 434 qp->s_sending_psn = 0; 435 qp->s_sending_hpsn = 0; 436 qp->s_psn = 0; 437 qp->r_psn = 0; 438 qp->r_msn = 0; 439 if (type == IB_QPT_RC) { 440 qp->s_state = IB_OPCODE_RC_SEND_LAST; 441 qp->r_state = IB_OPCODE_RC_SEND_LAST; 442 } else { 443 qp->s_state = IB_OPCODE_UC_SEND_LAST; 444 qp->r_state = IB_OPCODE_UC_SEND_LAST; 445 } 446 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 447 qp->r_nak_state = 0; 448 qp->r_aflags = 0; 449 qp->r_flags = 0; 450 qp->s_head = 0; 451 qp->s_tail = 0; 452 qp->s_cur = 0; 453 qp->s_acked = 0; 454 qp->s_last = 0; 455 qp->s_ssn = 1; 456 qp->s_lsn = 0; 457 qp->s_mig_state = IB_MIG_MIGRATED; 458 memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue)); 459 qp->r_head_ack_queue = 0; 460 qp->s_tail_ack_queue = 0; 461 qp->s_num_rd_atomic = 0; 462 if (qp->r_rq.wq) { 463 qp->r_rq.wq->head = 0; 464 qp->r_rq.wq->tail = 0; 465 } 466 qp->r_sge.num_sge = 0; 467 } 468 EXPORT_SYMBOL(rvt_reset_qp); 469 470 /** 471 * rvt_create_qp - create a queue pair for a device 472 * @ibpd: the protection domain who's device we create the queue pair for 473 * @init_attr: the attributes of the queue pair 474 * @udata: user data for libibverbs.so 475 * 476 * Queue pair creation is mostly an rvt issue. However, drivers have their own 477 * unique idea of what queue pair numbers mean. For instance there is a reserved 478 * range for PSM. 479 * 480 * Returns the queue pair on success, otherwise returns an errno. 481 * 482 * Called by the ib_create_qp() core verbs function. 483 */ 484 struct ib_qp *rvt_create_qp(struct ib_pd *ibpd, 485 struct ib_qp_init_attr *init_attr, 486 struct ib_udata *udata) 487 { 488 struct rvt_qp *qp; 489 int err; 490 struct rvt_swqe *swq = NULL; 491 size_t sz; 492 size_t sg_list_sz; 493 struct ib_qp *ret = ERR_PTR(-ENOMEM); 494 struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device); 495 void *priv = NULL; 496 gfp_t gfp; 497 498 if (!rdi) 499 return ERR_PTR(-EINVAL); 500 501 if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge || 502 init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr || 503 init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO)) 504 return ERR_PTR(-EINVAL); 505 506 /* GFP_NOIO is applicable to RC QP's only */ 507 508 if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO && 509 init_attr->qp_type != IB_QPT_RC) 510 return ERR_PTR(-EINVAL); 511 512 gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ? 513 GFP_NOIO : GFP_KERNEL; 514 515 /* Check receive queue parameters if no SRQ is specified. */ 516 if (!init_attr->srq) { 517 if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge || 518 init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr) 519 return ERR_PTR(-EINVAL); 520 521 if (init_attr->cap.max_send_sge + 522 init_attr->cap.max_send_wr + 523 init_attr->cap.max_recv_sge + 524 init_attr->cap.max_recv_wr == 0) 525 return ERR_PTR(-EINVAL); 526 } 527 528 switch (init_attr->qp_type) { 529 case IB_QPT_SMI: 530 case IB_QPT_GSI: 531 if (init_attr->port_num == 0 || 532 init_attr->port_num > ibpd->device->phys_port_cnt) 533 return ERR_PTR(-EINVAL); 534 case IB_QPT_UC: 535 case IB_QPT_RC: 536 case IB_QPT_UD: 537 sz = sizeof(struct rvt_sge) * 538 init_attr->cap.max_send_sge + 539 sizeof(struct rvt_swqe); 540 if (gfp == GFP_NOIO) 541 swq = __vmalloc( 542 (init_attr->cap.max_send_wr + 1) * sz, 543 gfp, PAGE_KERNEL); 544 else 545 swq = vmalloc( 546 (init_attr->cap.max_send_wr + 1) * sz); 547 if (!swq) 548 return ERR_PTR(-ENOMEM); 549 550 sz = sizeof(*qp); 551 sg_list_sz = 0; 552 if (init_attr->srq) { 553 struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq); 554 555 if (srq->rq.max_sge > 1) 556 sg_list_sz = sizeof(*qp->r_sg_list) * 557 (srq->rq.max_sge - 1); 558 } else if (init_attr->cap.max_recv_sge > 1) 559 sg_list_sz = sizeof(*qp->r_sg_list) * 560 (init_attr->cap.max_recv_sge - 1); 561 qp = kzalloc(sz + sg_list_sz, gfp); 562 if (!qp) 563 goto bail_swq; 564 565 RCU_INIT_POINTER(qp->next, NULL); 566 567 /* 568 * Driver needs to set up it's private QP structure and do any 569 * initialization that is needed. 570 */ 571 priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp); 572 if (!priv) 573 goto bail_qp; 574 qp->priv = priv; 575 qp->timeout_jiffies = 576 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / 577 1000UL); 578 if (init_attr->srq) { 579 sz = 0; 580 } else { 581 qp->r_rq.size = init_attr->cap.max_recv_wr + 1; 582 qp->r_rq.max_sge = init_attr->cap.max_recv_sge; 583 sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + 584 sizeof(struct rvt_rwqe); 585 if (udata) 586 qp->r_rq.wq = vmalloc_user( 587 sizeof(struct rvt_rwq) + 588 qp->r_rq.size * sz); 589 else if (gfp == GFP_NOIO) 590 qp->r_rq.wq = __vmalloc( 591 sizeof(struct rvt_rwq) + 592 qp->r_rq.size * sz, 593 gfp, PAGE_KERNEL); 594 else 595 qp->r_rq.wq = vmalloc( 596 sizeof(struct rvt_rwq) + 597 qp->r_rq.size * sz); 598 if (!qp->r_rq.wq) 599 goto bail_driver_priv; 600 } 601 602 /* 603 * ib_create_qp() will initialize qp->ibqp 604 * except for qp->ibqp.qp_num. 605 */ 606 spin_lock_init(&qp->r_lock); 607 spin_lock_init(&qp->s_lock); 608 spin_lock_init(&qp->r_rq.lock); 609 atomic_set(&qp->refcount, 0); 610 init_waitqueue_head(&qp->wait); 611 init_timer(&qp->s_timer); 612 qp->s_timer.data = (unsigned long)qp; 613 INIT_LIST_HEAD(&qp->rspwait); 614 qp->state = IB_QPS_RESET; 615 qp->s_wq = swq; 616 qp->s_size = init_attr->cap.max_send_wr + 1; 617 qp->s_max_sge = init_attr->cap.max_send_sge; 618 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) 619 qp->s_flags = RVT_S_SIGNAL_REQ_WR; 620 621 err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table, 622 init_attr->qp_type, 623 init_attr->port_num, gfp); 624 if (err < 0) { 625 ret = ERR_PTR(err); 626 goto bail_rq_wq; 627 } 628 qp->ibqp.qp_num = err; 629 qp->port_num = init_attr->port_num; 630 rvt_reset_qp(rdi, qp, init_attr->qp_type); 631 break; 632 633 default: 634 /* Don't support raw QPs */ 635 return ERR_PTR(-EINVAL); 636 } 637 638 init_attr->cap.max_inline_data = 0; 639 640 /* 641 * Return the address of the RWQ as the offset to mmap. 642 * See rvt_mmap() for details. 643 */ 644 if (udata && udata->outlen >= sizeof(__u64)) { 645 if (!qp->r_rq.wq) { 646 __u64 offset = 0; 647 648 err = ib_copy_to_udata(udata, &offset, 649 sizeof(offset)); 650 if (err) { 651 ret = ERR_PTR(err); 652 goto bail_qpn; 653 } 654 } else { 655 u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz; 656 657 qp->ip = rvt_create_mmap_info(rdi, s, 658 ibpd->uobject->context, 659 qp->r_rq.wq); 660 if (!qp->ip) { 661 ret = ERR_PTR(-ENOMEM); 662 goto bail_qpn; 663 } 664 665 err = ib_copy_to_udata(udata, &qp->ip->offset, 666 sizeof(qp->ip->offset)); 667 if (err) { 668 ret = ERR_PTR(err); 669 goto bail_ip; 670 } 671 } 672 } 673 674 spin_lock(&rdi->n_qps_lock); 675 if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) { 676 spin_unlock(&rdi->n_qps_lock); 677 ret = ERR_PTR(-ENOMEM); 678 goto bail_ip; 679 } 680 681 rdi->n_qps_allocated++; 682 spin_unlock(&rdi->n_qps_lock); 683 684 if (qp->ip) { 685 spin_lock_irq(&rdi->pending_lock); 686 list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps); 687 spin_unlock_irq(&rdi->pending_lock); 688 } 689 690 ret = &qp->ibqp; 691 692 /* 693 * We have our QP and its good, now keep track of what types of opcodes 694 * can be processed on this QP. We do this by keeping track of what the 695 * 3 high order bits of the opcode are. 696 */ 697 switch (init_attr->qp_type) { 698 case IB_QPT_SMI: 699 case IB_QPT_GSI: 700 case IB_QPT_UD: 701 qp->allowed_ops = IB_OPCODE_UD_SEND_ONLY & RVT_OPCODE_QP_MASK; 702 break; 703 case IB_QPT_RC: 704 qp->allowed_ops = IB_OPCODE_RC_SEND_ONLY & RVT_OPCODE_QP_MASK; 705 break; 706 case IB_QPT_UC: 707 qp->allowed_ops = IB_OPCODE_UC_SEND_ONLY & RVT_OPCODE_QP_MASK; 708 break; 709 default: 710 ret = ERR_PTR(-EINVAL); 711 goto bail_ip; 712 } 713 714 return ret; 715 716 bail_ip: 717 kref_put(&qp->ip->ref, rvt_release_mmap_info); 718 719 bail_qpn: 720 free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); 721 722 bail_rq_wq: 723 vfree(qp->r_rq.wq); 724 725 bail_driver_priv: 726 rdi->driver_f.qp_priv_free(rdi, qp); 727 728 bail_qp: 729 kfree(qp); 730 731 bail_swq: 732 vfree(swq); 733 734 return ret; 735 } 736 737 void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends) 738 { 739 unsigned n; 740 741 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) 742 rvt_put_ss(&qp->s_rdma_read_sge); 743 744 rvt_put_ss(&qp->r_sge); 745 746 if (clr_sends) { 747 while (qp->s_last != qp->s_head) { 748 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last); 749 unsigned i; 750 751 for (i = 0; i < wqe->wr.num_sge; i++) { 752 struct rvt_sge *sge = &wqe->sg_list[i]; 753 754 rvt_put_mr(sge->mr); 755 } 756 if (qp->ibqp.qp_type == IB_QPT_UD || 757 qp->ibqp.qp_type == IB_QPT_SMI || 758 qp->ibqp.qp_type == IB_QPT_GSI) 759 atomic_dec(&ibah_to_rvtah( 760 wqe->ud_wr.ah)->refcount); 761 if (++qp->s_last >= qp->s_size) 762 qp->s_last = 0; 763 } 764 if (qp->s_rdma_mr) { 765 rvt_put_mr(qp->s_rdma_mr); 766 qp->s_rdma_mr = NULL; 767 } 768 } 769 770 if (qp->ibqp.qp_type != IB_QPT_RC) 771 return; 772 773 for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) { 774 struct rvt_ack_entry *e = &qp->s_ack_queue[n]; 775 776 if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST && 777 e->rdma_sge.mr) { 778 rvt_put_mr(e->rdma_sge.mr); 779 e->rdma_sge.mr = NULL; 780 } 781 } 782 } 783 EXPORT_SYMBOL(rvt_clear_mr_refs); 784 785 /** 786 * rvt_error_qp - put a QP into the error state 787 * @qp: the QP to put into the error state 788 * @err: the receive completion error to signal if a RWQE is active 789 * 790 * Flushes both send and receive work queues. 791 * Returns true if last WQE event should be generated. 792 * The QP r_lock and s_lock should be held and interrupts disabled. 793 * If we are already in error state, just return. 794 */ 795 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err) 796 { 797 struct ib_wc wc; 798 int ret = 0; 799 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 800 801 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) 802 goto bail; 803 804 qp->state = IB_QPS_ERR; 805 806 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { 807 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); 808 del_timer(&qp->s_timer); 809 } 810 811 if (qp->s_flags & RVT_S_ANY_WAIT_SEND) 812 qp->s_flags &= ~RVT_S_ANY_WAIT_SEND; 813 814 rdi->driver_f.notify_error_qp(qp); 815 816 /* Schedule the sending tasklet to drain the send work queue. */ 817 if (qp->s_last != qp->s_head) 818 rdi->driver_f.schedule_send(qp); 819 820 rvt_clear_mr_refs(qp, 0); 821 822 memset(&wc, 0, sizeof(wc)); 823 wc.qp = &qp->ibqp; 824 wc.opcode = IB_WC_RECV; 825 826 if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) { 827 wc.wr_id = qp->r_wr_id; 828 wc.status = err; 829 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 830 } 831 wc.status = IB_WC_WR_FLUSH_ERR; 832 833 if (qp->r_rq.wq) { 834 struct rvt_rwq *wq; 835 u32 head; 836 u32 tail; 837 838 spin_lock(&qp->r_rq.lock); 839 840 /* sanity check pointers before trusting them */ 841 wq = qp->r_rq.wq; 842 head = wq->head; 843 if (head >= qp->r_rq.size) 844 head = 0; 845 tail = wq->tail; 846 if (tail >= qp->r_rq.size) 847 tail = 0; 848 while (tail != head) { 849 wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id; 850 if (++tail >= qp->r_rq.size) 851 tail = 0; 852 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 853 } 854 wq->tail = tail; 855 856 spin_unlock(&qp->r_rq.lock); 857 } else if (qp->ibqp.event_handler) { 858 ret = 1; 859 } 860 861 bail: 862 return ret; 863 } 864 EXPORT_SYMBOL(rvt_error_qp); 865 866 /* 867 * Put the QP into the hash table. 868 * The hash table holds a reference to the QP. 869 */ 870 static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) 871 { 872 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; 873 unsigned long flags; 874 875 atomic_inc(&qp->refcount); 876 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); 877 878 if (qp->ibqp.qp_num <= 1) { 879 rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp); 880 } else { 881 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); 882 883 qp->next = rdi->qp_dev->qp_table[n]; 884 rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp); 885 trace_rvt_qpinsert(qp, n); 886 } 887 888 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); 889 } 890 891 /* 892 * Remove the QP from the table so it can't be found asynchronously by 893 * the receive routine. 894 */ 895 void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) 896 { 897 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; 898 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); 899 unsigned long flags; 900 int removed = 1; 901 902 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); 903 904 if (rcu_dereference_protected(rvp->qp[0], 905 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { 906 RCU_INIT_POINTER(rvp->qp[0], NULL); 907 } else if (rcu_dereference_protected(rvp->qp[1], 908 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { 909 RCU_INIT_POINTER(rvp->qp[1], NULL); 910 } else { 911 struct rvt_qp *q; 912 struct rvt_qp __rcu **qpp; 913 914 removed = 0; 915 qpp = &rdi->qp_dev->qp_table[n]; 916 for (; (q = rcu_dereference_protected(*qpp, 917 lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL; 918 qpp = &q->next) { 919 if (q == qp) { 920 RCU_INIT_POINTER(*qpp, 921 rcu_dereference_protected(qp->next, 922 lockdep_is_held(&rdi->qp_dev->qpt_lock))); 923 removed = 1; 924 trace_rvt_qpremove(qp, n); 925 break; 926 } 927 } 928 } 929 930 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); 931 if (removed) { 932 synchronize_rcu(); 933 if (atomic_dec_and_test(&qp->refcount)) 934 wake_up(&qp->wait); 935 } 936 } 937 EXPORT_SYMBOL(rvt_remove_qp); 938 939 /** 940 * qib_modify_qp - modify the attributes of a queue pair 941 * @ibqp: the queue pair who's attributes we're modifying 942 * @attr: the new attributes 943 * @attr_mask: the mask of attributes to modify 944 * @udata: user data for libibverbs.so 945 * 946 * Returns 0 on success, otherwise returns an errno. 947 */ 948 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 949 int attr_mask, struct ib_udata *udata) 950 { 951 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 952 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 953 enum ib_qp_state cur_state, new_state; 954 struct ib_event ev; 955 int lastwqe = 0; 956 int mig = 0; 957 int pmtu = 0; /* for gcc warning only */ 958 enum rdma_link_layer link; 959 960 link = rdma_port_get_link_layer(ibqp->device, qp->port_num); 961 962 spin_lock_irq(&qp->r_lock); 963 spin_lock(&qp->s_lock); 964 965 cur_state = attr_mask & IB_QP_CUR_STATE ? 966 attr->cur_qp_state : qp->state; 967 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 968 969 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 970 attr_mask, link)) 971 goto inval; 972 973 if (rdi->driver_f.check_modify_qp && 974 rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata)) 975 goto inval; 976 977 if (attr_mask & IB_QP_AV) { 978 if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) 979 goto inval; 980 if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr)) 981 goto inval; 982 } 983 984 if (attr_mask & IB_QP_ALT_PATH) { 985 if (attr->alt_ah_attr.dlid >= 986 be16_to_cpu(IB_MULTICAST_LID_BASE)) 987 goto inval; 988 if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr)) 989 goto inval; 990 if (attr->alt_pkey_index >= rvt_get_npkeys(rdi)) 991 goto inval; 992 } 993 994 if (attr_mask & IB_QP_PKEY_INDEX) 995 if (attr->pkey_index >= rvt_get_npkeys(rdi)) 996 goto inval; 997 998 if (attr_mask & IB_QP_MIN_RNR_TIMER) 999 if (attr->min_rnr_timer > 31) 1000 goto inval; 1001 1002 if (attr_mask & IB_QP_PORT) 1003 if (qp->ibqp.qp_type == IB_QPT_SMI || 1004 qp->ibqp.qp_type == IB_QPT_GSI || 1005 attr->port_num == 0 || 1006 attr->port_num > ibqp->device->phys_port_cnt) 1007 goto inval; 1008 1009 if (attr_mask & IB_QP_DEST_QPN) 1010 if (attr->dest_qp_num > RVT_QPN_MASK) 1011 goto inval; 1012 1013 if (attr_mask & IB_QP_RETRY_CNT) 1014 if (attr->retry_cnt > 7) 1015 goto inval; 1016 1017 if (attr_mask & IB_QP_RNR_RETRY) 1018 if (attr->rnr_retry > 7) 1019 goto inval; 1020 1021 /* 1022 * Don't allow invalid path_mtu values. OK to set greater 1023 * than the active mtu (or even the max_cap, if we have tuned 1024 * that to a small mtu. We'll set qp->path_mtu 1025 * to the lesser of requested attribute mtu and active, 1026 * for packetizing messages. 1027 * Note that the QP port has to be set in INIT and MTU in RTR. 1028 */ 1029 if (attr_mask & IB_QP_PATH_MTU) { 1030 pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr); 1031 if (pmtu < 0) 1032 goto inval; 1033 } 1034 1035 if (attr_mask & IB_QP_PATH_MIG_STATE) { 1036 if (attr->path_mig_state == IB_MIG_REARM) { 1037 if (qp->s_mig_state == IB_MIG_ARMED) 1038 goto inval; 1039 if (new_state != IB_QPS_RTS) 1040 goto inval; 1041 } else if (attr->path_mig_state == IB_MIG_MIGRATED) { 1042 if (qp->s_mig_state == IB_MIG_REARM) 1043 goto inval; 1044 if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD) 1045 goto inval; 1046 if (qp->s_mig_state == IB_MIG_ARMED) 1047 mig = 1; 1048 } else { 1049 goto inval; 1050 } 1051 } 1052 1053 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1054 if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic) 1055 goto inval; 1056 1057 switch (new_state) { 1058 case IB_QPS_RESET: 1059 if (qp->state != IB_QPS_RESET) 1060 rvt_reset_qp(rdi, qp, ibqp->qp_type); 1061 break; 1062 1063 case IB_QPS_RTR: 1064 /* Allow event to re-trigger if QP set to RTR more than once */ 1065 qp->r_flags &= ~RVT_R_COMM_EST; 1066 qp->state = new_state; 1067 break; 1068 1069 case IB_QPS_SQD: 1070 qp->s_draining = qp->s_last != qp->s_cur; 1071 qp->state = new_state; 1072 break; 1073 1074 case IB_QPS_SQE: 1075 if (qp->ibqp.qp_type == IB_QPT_RC) 1076 goto inval; 1077 qp->state = new_state; 1078 break; 1079 1080 case IB_QPS_ERR: 1081 lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1082 break; 1083 1084 default: 1085 qp->state = new_state; 1086 break; 1087 } 1088 1089 if (attr_mask & IB_QP_PKEY_INDEX) 1090 qp->s_pkey_index = attr->pkey_index; 1091 1092 if (attr_mask & IB_QP_PORT) 1093 qp->port_num = attr->port_num; 1094 1095 if (attr_mask & IB_QP_DEST_QPN) 1096 qp->remote_qpn = attr->dest_qp_num; 1097 1098 if (attr_mask & IB_QP_SQ_PSN) { 1099 qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask; 1100 qp->s_psn = qp->s_next_psn; 1101 qp->s_sending_psn = qp->s_next_psn; 1102 qp->s_last_psn = qp->s_next_psn - 1; 1103 qp->s_sending_hpsn = qp->s_last_psn; 1104 } 1105 1106 if (attr_mask & IB_QP_RQ_PSN) 1107 qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask; 1108 1109 if (attr_mask & IB_QP_ACCESS_FLAGS) 1110 qp->qp_access_flags = attr->qp_access_flags; 1111 1112 if (attr_mask & IB_QP_AV) { 1113 qp->remote_ah_attr = attr->ah_attr; 1114 qp->s_srate = attr->ah_attr.static_rate; 1115 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); 1116 } 1117 1118 if (attr_mask & IB_QP_ALT_PATH) { 1119 qp->alt_ah_attr = attr->alt_ah_attr; 1120 qp->s_alt_pkey_index = attr->alt_pkey_index; 1121 } 1122 1123 if (attr_mask & IB_QP_PATH_MIG_STATE) { 1124 qp->s_mig_state = attr->path_mig_state; 1125 if (mig) { 1126 qp->remote_ah_attr = qp->alt_ah_attr; 1127 qp->port_num = qp->alt_ah_attr.port_num; 1128 qp->s_pkey_index = qp->s_alt_pkey_index; 1129 1130 /* 1131 * Ignored by drivers which do not support it. Not 1132 * really worth creating a call back into the driver 1133 * just to set a flag. 1134 */ 1135 qp->s_flags |= RVT_S_AHG_CLEAR; 1136 } 1137 } 1138 1139 if (attr_mask & IB_QP_PATH_MTU) { 1140 qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu); 1141 qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu); 1142 } 1143 1144 if (attr_mask & IB_QP_RETRY_CNT) { 1145 qp->s_retry_cnt = attr->retry_cnt; 1146 qp->s_retry = attr->retry_cnt; 1147 } 1148 1149 if (attr_mask & IB_QP_RNR_RETRY) { 1150 qp->s_rnr_retry_cnt = attr->rnr_retry; 1151 qp->s_rnr_retry = attr->rnr_retry; 1152 } 1153 1154 if (attr_mask & IB_QP_MIN_RNR_TIMER) 1155 qp->r_min_rnr_timer = attr->min_rnr_timer; 1156 1157 if (attr_mask & IB_QP_TIMEOUT) { 1158 qp->timeout = attr->timeout; 1159 qp->timeout_jiffies = 1160 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / 1161 1000UL); 1162 } 1163 1164 if (attr_mask & IB_QP_QKEY) 1165 qp->qkey = attr->qkey; 1166 1167 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1168 qp->r_max_rd_atomic = attr->max_dest_rd_atomic; 1169 1170 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) 1171 qp->s_max_rd_atomic = attr->max_rd_atomic; 1172 1173 if (rdi->driver_f.modify_qp) 1174 rdi->driver_f.modify_qp(qp, attr, attr_mask, udata); 1175 1176 spin_unlock(&qp->s_lock); 1177 spin_unlock_irq(&qp->r_lock); 1178 1179 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 1180 rvt_insert_qp(rdi, qp); 1181 1182 if (lastwqe) { 1183 ev.device = qp->ibqp.device; 1184 ev.element.qp = &qp->ibqp; 1185 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 1186 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1187 } 1188 if (mig) { 1189 ev.device = qp->ibqp.device; 1190 ev.element.qp = &qp->ibqp; 1191 ev.event = IB_EVENT_PATH_MIG; 1192 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1193 } 1194 return 0; 1195 1196 inval: 1197 spin_unlock(&qp->s_lock); 1198 spin_unlock_irq(&qp->r_lock); 1199 return -EINVAL; 1200 } 1201 1202 /** 1203 * rvt_destroy_qp - destroy a queue pair 1204 * @ibqp: the queue pair to destroy 1205 * 1206 * Returns 0 on success. 1207 * 1208 * Note that this can be called while the QP is actively sending or 1209 * receiving! 1210 */ 1211 int rvt_destroy_qp(struct ib_qp *ibqp) 1212 { 1213 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1214 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1215 1216 spin_lock_irq(&qp->r_lock); 1217 spin_lock(&qp->s_lock); 1218 rvt_reset_qp(rdi, qp, ibqp->qp_type); 1219 spin_unlock(&qp->s_lock); 1220 spin_unlock_irq(&qp->r_lock); 1221 1222 /* qpn is now available for use again */ 1223 rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); 1224 1225 spin_lock(&rdi->n_qps_lock); 1226 rdi->n_qps_allocated--; 1227 spin_unlock(&rdi->n_qps_lock); 1228 1229 if (qp->ip) 1230 kref_put(&qp->ip->ref, rvt_release_mmap_info); 1231 else 1232 vfree(qp->r_rq.wq); 1233 vfree(qp->s_wq); 1234 rdi->driver_f.qp_priv_free(rdi, qp); 1235 kfree(qp); 1236 return 0; 1237 } 1238 1239 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 1240 int attr_mask, struct ib_qp_init_attr *init_attr) 1241 { 1242 return -EOPNOTSUPP; 1243 } 1244 1245 /** 1246 * rvt_post_receive - post a receive on a QP 1247 * @ibqp: the QP to post the receive on 1248 * @wr: the WR to post 1249 * @bad_wr: the first bad WR is put here 1250 * 1251 * This may be called from interrupt context. 1252 */ 1253 int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 1254 struct ib_recv_wr **bad_wr) 1255 { 1256 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1257 struct rvt_rwq *wq = qp->r_rq.wq; 1258 unsigned long flags; 1259 1260 /* Check that state is OK to post receive. */ 1261 if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) { 1262 *bad_wr = wr; 1263 return -EINVAL; 1264 } 1265 1266 for (; wr; wr = wr->next) { 1267 struct rvt_rwqe *wqe; 1268 u32 next; 1269 int i; 1270 1271 if ((unsigned)wr->num_sge > qp->r_rq.max_sge) { 1272 *bad_wr = wr; 1273 return -EINVAL; 1274 } 1275 1276 spin_lock_irqsave(&qp->r_rq.lock, flags); 1277 next = wq->head + 1; 1278 if (next >= qp->r_rq.size) 1279 next = 0; 1280 if (next == wq->tail) { 1281 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 1282 *bad_wr = wr; 1283 return -ENOMEM; 1284 } 1285 1286 wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head); 1287 wqe->wr_id = wr->wr_id; 1288 wqe->num_sge = wr->num_sge; 1289 for (i = 0; i < wr->num_sge; i++) 1290 wqe->sg_list[i] = wr->sg_list[i]; 1291 /* Make sure queue entry is written before the head index. */ 1292 smp_wmb(); 1293 wq->head = next; 1294 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 1295 } 1296 return 0; 1297 } 1298 1299 /** 1300 * rvt_post_one_wr - post one RC, UC, or UD send work request 1301 * @qp: the QP to post on 1302 * @wr: the work request to send 1303 */ 1304 static int rvt_post_one_wr(struct rvt_qp *qp, struct ib_send_wr *wr) 1305 { 1306 struct rvt_swqe *wqe; 1307 u32 next; 1308 int i; 1309 int j; 1310 int acc; 1311 struct rvt_lkey_table *rkt; 1312 struct rvt_pd *pd; 1313 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 1314 1315 /* IB spec says that num_sge == 0 is OK. */ 1316 if (unlikely(wr->num_sge > qp->s_max_sge)) 1317 return -EINVAL; 1318 1319 /* 1320 * Don't allow RDMA reads or atomic operations on UC or 1321 * undefined operations. 1322 * Make sure buffer is large enough to hold the result for atomics. 1323 */ 1324 if (qp->ibqp.qp_type == IB_QPT_UC) { 1325 if ((unsigned)wr->opcode >= IB_WR_RDMA_READ) 1326 return -EINVAL; 1327 } else if (qp->ibqp.qp_type != IB_QPT_RC) { 1328 /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */ 1329 if (wr->opcode != IB_WR_SEND && 1330 wr->opcode != IB_WR_SEND_WITH_IMM) 1331 return -EINVAL; 1332 /* Check UD destination address PD */ 1333 if (qp->ibqp.pd != ud_wr(wr)->ah->pd) 1334 return -EINVAL; 1335 } else if ((unsigned)wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) { 1336 return -EINVAL; 1337 } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP && 1338 (wr->num_sge == 0 || 1339 wr->sg_list[0].length < sizeof(u64) || 1340 wr->sg_list[0].addr & (sizeof(u64) - 1))) { 1341 return -EINVAL; 1342 } else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) { 1343 return -EINVAL; 1344 } 1345 1346 next = qp->s_head + 1; 1347 if (next >= qp->s_size) 1348 next = 0; 1349 if (next == qp->s_last) 1350 return -ENOMEM; 1351 1352 rkt = &rdi->lkey_table; 1353 pd = ibpd_to_rvtpd(qp->ibqp.pd); 1354 wqe = rvt_get_swqe_ptr(qp, qp->s_head); 1355 1356 if (qp->ibqp.qp_type != IB_QPT_UC && 1357 qp->ibqp.qp_type != IB_QPT_RC) 1358 memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr)); 1359 else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM || 1360 wr->opcode == IB_WR_RDMA_WRITE || 1361 wr->opcode == IB_WR_RDMA_READ) 1362 memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr)); 1363 else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1364 wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) 1365 memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr)); 1366 else 1367 memcpy(&wqe->wr, wr, sizeof(wqe->wr)); 1368 1369 wqe->length = 0; 1370 j = 0; 1371 if (wr->num_sge) { 1372 acc = wr->opcode >= IB_WR_RDMA_READ ? 1373 IB_ACCESS_LOCAL_WRITE : 0; 1374 for (i = 0; i < wr->num_sge; i++) { 1375 u32 length = wr->sg_list[i].length; 1376 int ok; 1377 1378 if (length == 0) 1379 continue; 1380 ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j], 1381 &wr->sg_list[i], acc); 1382 if (!ok) 1383 goto bail_inval_free; 1384 wqe->length += length; 1385 j++; 1386 } 1387 wqe->wr.num_sge = j; 1388 } 1389 if (qp->ibqp.qp_type == IB_QPT_UC || 1390 qp->ibqp.qp_type == IB_QPT_RC) { 1391 if (wqe->length > 0x80000000U) 1392 goto bail_inval_free; 1393 } else { 1394 atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount); 1395 } 1396 wqe->ssn = qp->s_ssn++; 1397 qp->s_head = next; 1398 1399 return 0; 1400 1401 bail_inval_free: 1402 /* release mr holds */ 1403 while (j) { 1404 struct rvt_sge *sge = &wqe->sg_list[--j]; 1405 1406 rvt_put_mr(sge->mr); 1407 } 1408 return -EINVAL; 1409 } 1410 1411 /** 1412 * rvt_post_send - post a send on a QP 1413 * @ibqp: the QP to post the send on 1414 * @wr: the list of work requests to post 1415 * @bad_wr: the first bad WR is put here 1416 * 1417 * This may be called from interrupt context. 1418 */ 1419 int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 1420 struct ib_send_wr **bad_wr) 1421 { 1422 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1423 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1424 unsigned long flags = 0; 1425 int call_send; 1426 unsigned nreq = 0; 1427 int err = 0; 1428 1429 spin_lock_irqsave(&qp->s_lock, flags); 1430 1431 /* 1432 * Ensure QP state is such that we can send. If not bail out early, 1433 * there is no need to do this every time we post a send. 1434 */ 1435 if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) { 1436 spin_unlock_irqrestore(&qp->s_lock, flags); 1437 return -EINVAL; 1438 } 1439 1440 /* 1441 * If the send queue is empty, and we only have a single WR then just go 1442 * ahead and kick the send engine into gear. Otherwise we will always 1443 * just schedule the send to happen later. 1444 */ 1445 call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next; 1446 1447 for (; wr; wr = wr->next) { 1448 err = rvt_post_one_wr(qp, wr); 1449 if (unlikely(err)) { 1450 *bad_wr = wr; 1451 goto bail; 1452 } 1453 nreq++; 1454 } 1455 bail: 1456 if (nreq && !call_send) 1457 rdi->driver_f.schedule_send(qp); 1458 spin_unlock_irqrestore(&qp->s_lock, flags); 1459 if (nreq && call_send) 1460 rdi->driver_f.do_send(qp); 1461 return err; 1462 } 1463 1464 /** 1465 * rvt_post_srq_receive - post a receive on a shared receive queue 1466 * @ibsrq: the SRQ to post the receive on 1467 * @wr: the list of work requests to post 1468 * @bad_wr: A pointer to the first WR to cause a problem is put here 1469 * 1470 * This may be called from interrupt context. 1471 */ 1472 int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 1473 struct ib_recv_wr **bad_wr) 1474 { 1475 return -EOPNOTSUPP; 1476 } 1477 1478 void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn) 1479 { 1480 struct rvt_qpn_map *map; 1481 1482 map = qpt->map + qpn / RVT_BITS_PER_PAGE; 1483 if (map->page) 1484 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); 1485 } 1486 EXPORT_SYMBOL(rvt_free_qpn); 1487 1488 void rvt_dec_qp_cnt(struct rvt_dev_info *rdi) 1489 { 1490 spin_lock(&rdi->n_qps_lock); 1491 rdi->n_qps_allocated--; 1492 spin_unlock(&rdi->n_qps_lock); 1493 } 1494 EXPORT_SYMBOL(rvt_dec_qp_cnt); 1495