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 /** 166 * rvt_driver_qp_init - Init driver qp resources 167 * @rdi: rvt dev strucutre 168 * 169 * Return: 0 on success 170 */ 171 int rvt_driver_qp_init(struct rvt_dev_info *rdi) 172 { 173 int i; 174 int ret = -ENOMEM; 175 176 if (!rdi->dparms.qp_table_size) 177 return -EINVAL; 178 179 /* 180 * If driver is not doing any QP allocation then make sure it is 181 * providing the necessary QP functions. 182 */ 183 if (!rdi->driver_f.free_all_qps || 184 !rdi->driver_f.qp_priv_alloc || 185 !rdi->driver_f.qp_priv_free || 186 !rdi->driver_f.notify_qp_reset) 187 return -EINVAL; 188 189 /* allocate parent object */ 190 rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL, 191 rdi->dparms.node); 192 if (!rdi->qp_dev) 193 return -ENOMEM; 194 195 /* allocate hash table */ 196 rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size; 197 rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size); 198 rdi->qp_dev->qp_table = 199 kmalloc_node(rdi->qp_dev->qp_table_size * 200 sizeof(*rdi->qp_dev->qp_table), 201 GFP_KERNEL, rdi->dparms.node); 202 if (!rdi->qp_dev->qp_table) 203 goto no_qp_table; 204 205 for (i = 0; i < rdi->qp_dev->qp_table_size; i++) 206 RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL); 207 208 spin_lock_init(&rdi->qp_dev->qpt_lock); 209 210 /* initialize qpn map */ 211 if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table)) 212 goto fail_table; 213 214 spin_lock_init(&rdi->n_qps_lock); 215 216 return 0; 217 218 fail_table: 219 kfree(rdi->qp_dev->qp_table); 220 free_qpn_table(&rdi->qp_dev->qpn_table); 221 222 no_qp_table: 223 kfree(rdi->qp_dev); 224 225 return ret; 226 } 227 228 /** 229 * free_all_qps - check for QPs still in use 230 * @qpt: the QP table to empty 231 * 232 * There should not be any QPs still in use. 233 * Free memory for table. 234 */ 235 static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi) 236 { 237 unsigned long flags; 238 struct rvt_qp *qp; 239 unsigned n, qp_inuse = 0; 240 spinlock_t *ql; /* work around too long line below */ 241 242 if (rdi->driver_f.free_all_qps) 243 qp_inuse = rdi->driver_f.free_all_qps(rdi); 244 245 qp_inuse += rvt_mcast_tree_empty(rdi); 246 247 if (!rdi->qp_dev) 248 return qp_inuse; 249 250 ql = &rdi->qp_dev->qpt_lock; 251 spin_lock_irqsave(ql, flags); 252 for (n = 0; n < rdi->qp_dev->qp_table_size; n++) { 253 qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n], 254 lockdep_is_held(ql)); 255 RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL); 256 257 for (; qp; qp = rcu_dereference_protected(qp->next, 258 lockdep_is_held(ql))) 259 qp_inuse++; 260 } 261 spin_unlock_irqrestore(ql, flags); 262 synchronize_rcu(); 263 return qp_inuse; 264 } 265 266 /** 267 * rvt_qp_exit - clean up qps on device exit 268 * @rdi: rvt dev structure 269 * 270 * Check for qp leaks and free resources. 271 */ 272 void rvt_qp_exit(struct rvt_dev_info *rdi) 273 { 274 u32 qps_inuse = rvt_free_all_qps(rdi); 275 276 if (qps_inuse) 277 rvt_pr_err(rdi, "QP memory leak! %u still in use\n", 278 qps_inuse); 279 if (!rdi->qp_dev) 280 return; 281 282 kfree(rdi->qp_dev->qp_table); 283 free_qpn_table(&rdi->qp_dev->qpn_table); 284 kfree(rdi->qp_dev); 285 } 286 287 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt, 288 struct rvt_qpn_map *map, unsigned off) 289 { 290 return (map - qpt->map) * RVT_BITS_PER_PAGE + off; 291 } 292 293 /** 294 * alloc_qpn - Allocate the next available qpn or zero/one for QP type 295 * IB_QPT_SMI/IB_QPT_GSI 296 *@rdi: rvt device info structure 297 *@qpt: queue pair number table pointer 298 *@port_num: IB port number, 1 based, comes from core 299 * 300 * Return: The queue pair number 301 */ 302 static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, 303 enum ib_qp_type type, u8 port_num, gfp_t gfp) 304 { 305 u32 i, offset, max_scan, qpn; 306 struct rvt_qpn_map *map; 307 u32 ret; 308 309 if (rdi->driver_f.alloc_qpn) 310 return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num, gfp); 311 312 if (type == IB_QPT_SMI || type == IB_QPT_GSI) { 313 unsigned n; 314 315 ret = type == IB_QPT_GSI; 316 n = 1 << (ret + 2 * (port_num - 1)); 317 spin_lock(&qpt->lock); 318 if (qpt->flags & n) 319 ret = -EINVAL; 320 else 321 qpt->flags |= n; 322 spin_unlock(&qpt->lock); 323 goto bail; 324 } 325 326 qpn = qpt->last + qpt->incr; 327 if (qpn >= RVT_QPN_MAX) 328 qpn = qpt->incr | ((qpt->last & 1) ^ 1); 329 /* offset carries bit 0 */ 330 offset = qpn & RVT_BITS_PER_PAGE_MASK; 331 map = &qpt->map[qpn / RVT_BITS_PER_PAGE]; 332 max_scan = qpt->nmaps - !offset; 333 for (i = 0;;) { 334 if (unlikely(!map->page)) { 335 get_map_page(qpt, map, gfp); 336 if (unlikely(!map->page)) 337 break; 338 } 339 do { 340 if (!test_and_set_bit(offset, map->page)) { 341 qpt->last = qpn; 342 ret = qpn; 343 goto bail; 344 } 345 offset += qpt->incr; 346 /* 347 * This qpn might be bogus if offset >= BITS_PER_PAGE. 348 * That is OK. It gets re-assigned below 349 */ 350 qpn = mk_qpn(qpt, map, offset); 351 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX); 352 /* 353 * In order to keep the number of pages allocated to a 354 * minimum, we scan the all existing pages before increasing 355 * the size of the bitmap table. 356 */ 357 if (++i > max_scan) { 358 if (qpt->nmaps == RVT_QPNMAP_ENTRIES) 359 break; 360 map = &qpt->map[qpt->nmaps++]; 361 /* start at incr with current bit 0 */ 362 offset = qpt->incr | (offset & 1); 363 } else if (map < &qpt->map[qpt->nmaps]) { 364 ++map; 365 /* start at incr with current bit 0 */ 366 offset = qpt->incr | (offset & 1); 367 } else { 368 map = &qpt->map[0]; 369 /* wrap to first map page, invert bit 0 */ 370 offset = qpt->incr | ((offset & 1) ^ 1); 371 } 372 /* there can be no bits at shift and below */ 373 WARN_ON(offset & (rdi->dparms.qos_shift - 1)); 374 qpn = mk_qpn(qpt, map, offset); 375 } 376 377 ret = -ENOMEM; 378 379 bail: 380 return ret; 381 } 382 383 static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn) 384 { 385 struct rvt_qpn_map *map; 386 387 map = qpt->map + qpn / RVT_BITS_PER_PAGE; 388 if (map->page) 389 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); 390 } 391 392 /** 393 * rvt_clear_mr_refs - Drop help mr refs 394 * @qp: rvt qp data structure 395 * @clr_sends: If shoudl clear send side or not 396 */ 397 static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends) 398 { 399 unsigned n; 400 401 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) 402 rvt_put_ss(&qp->s_rdma_read_sge); 403 404 rvt_put_ss(&qp->r_sge); 405 406 if (clr_sends) { 407 while (qp->s_last != qp->s_head) { 408 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last); 409 unsigned i; 410 411 for (i = 0; i < wqe->wr.num_sge; i++) { 412 struct rvt_sge *sge = &wqe->sg_list[i]; 413 414 rvt_put_mr(sge->mr); 415 } 416 if (qp->ibqp.qp_type == IB_QPT_UD || 417 qp->ibqp.qp_type == IB_QPT_SMI || 418 qp->ibqp.qp_type == IB_QPT_GSI) 419 atomic_dec(&ibah_to_rvtah( 420 wqe->ud_wr.ah)->refcount); 421 if (++qp->s_last >= qp->s_size) 422 qp->s_last = 0; 423 smp_wmb(); /* see qp_set_savail */ 424 } 425 if (qp->s_rdma_mr) { 426 rvt_put_mr(qp->s_rdma_mr); 427 qp->s_rdma_mr = NULL; 428 } 429 } 430 431 if (qp->ibqp.qp_type != IB_QPT_RC) 432 return; 433 434 for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) { 435 struct rvt_ack_entry *e = &qp->s_ack_queue[n]; 436 437 if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST && 438 e->rdma_sge.mr) { 439 rvt_put_mr(e->rdma_sge.mr); 440 e->rdma_sge.mr = NULL; 441 } 442 } 443 } 444 445 /** 446 * rvt_remove_qp - remove qp form table 447 * @rdi: rvt dev struct 448 * @qp: qp to remove 449 * 450 * Remove the QP from the table so it can't be found asynchronously by 451 * the receive routine. 452 */ 453 static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) 454 { 455 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; 456 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); 457 unsigned long flags; 458 int removed = 1; 459 460 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); 461 462 if (rcu_dereference_protected(rvp->qp[0], 463 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { 464 RCU_INIT_POINTER(rvp->qp[0], NULL); 465 } else if (rcu_dereference_protected(rvp->qp[1], 466 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { 467 RCU_INIT_POINTER(rvp->qp[1], NULL); 468 } else { 469 struct rvt_qp *q; 470 struct rvt_qp __rcu **qpp; 471 472 removed = 0; 473 qpp = &rdi->qp_dev->qp_table[n]; 474 for (; (q = rcu_dereference_protected(*qpp, 475 lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL; 476 qpp = &q->next) { 477 if (q == qp) { 478 RCU_INIT_POINTER(*qpp, 479 rcu_dereference_protected(qp->next, 480 lockdep_is_held(&rdi->qp_dev->qpt_lock))); 481 removed = 1; 482 trace_rvt_qpremove(qp, n); 483 break; 484 } 485 } 486 } 487 488 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); 489 if (removed) { 490 synchronize_rcu(); 491 if (atomic_dec_and_test(&qp->refcount)) 492 wake_up(&qp->wait); 493 } 494 } 495 496 /** 497 * reset_qp - initialize the QP state to the reset state 498 * @qp: the QP to reset 499 * @type: the QP type 500 * r and s lock are required to be held by the caller 501 */ 502 static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, 503 enum ib_qp_type type) 504 { 505 if (qp->state != IB_QPS_RESET) { 506 qp->state = IB_QPS_RESET; 507 508 /* Let drivers flush their waitlist */ 509 rdi->driver_f.flush_qp_waiters(qp); 510 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT); 511 spin_unlock(&qp->s_lock); 512 spin_unlock(&qp->s_hlock); 513 spin_unlock_irq(&qp->r_lock); 514 515 /* Stop the send queue and the retry timer */ 516 rdi->driver_f.stop_send_queue(qp); 517 518 /* Wait for things to stop */ 519 rdi->driver_f.quiesce_qp(qp); 520 521 /* take qp out the hash and wait for it to be unused */ 522 rvt_remove_qp(rdi, qp); 523 wait_event(qp->wait, !atomic_read(&qp->refcount)); 524 525 /* grab the lock b/c it was locked at call time */ 526 spin_lock_irq(&qp->r_lock); 527 spin_lock(&qp->s_hlock); 528 spin_lock(&qp->s_lock); 529 530 rvt_clear_mr_refs(qp, 1); 531 } 532 533 /* 534 * Let the driver do any tear down it needs to for a qp 535 * that has been reset 536 */ 537 rdi->driver_f.notify_qp_reset(qp); 538 539 qp->remote_qpn = 0; 540 qp->qkey = 0; 541 qp->qp_access_flags = 0; 542 qp->s_flags &= RVT_S_SIGNAL_REQ_WR; 543 qp->s_hdrwords = 0; 544 qp->s_wqe = NULL; 545 qp->s_draining = 0; 546 qp->s_next_psn = 0; 547 qp->s_last_psn = 0; 548 qp->s_sending_psn = 0; 549 qp->s_sending_hpsn = 0; 550 qp->s_psn = 0; 551 qp->r_psn = 0; 552 qp->r_msn = 0; 553 if (type == IB_QPT_RC) { 554 qp->s_state = IB_OPCODE_RC_SEND_LAST; 555 qp->r_state = IB_OPCODE_RC_SEND_LAST; 556 } else { 557 qp->s_state = IB_OPCODE_UC_SEND_LAST; 558 qp->r_state = IB_OPCODE_UC_SEND_LAST; 559 } 560 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 561 qp->r_nak_state = 0; 562 qp->r_aflags = 0; 563 qp->r_flags = 0; 564 qp->s_head = 0; 565 qp->s_tail = 0; 566 qp->s_cur = 0; 567 qp->s_acked = 0; 568 qp->s_last = 0; 569 qp->s_ssn = 1; 570 qp->s_lsn = 0; 571 qp->s_mig_state = IB_MIG_MIGRATED; 572 memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue)); 573 qp->r_head_ack_queue = 0; 574 qp->s_tail_ack_queue = 0; 575 qp->s_num_rd_atomic = 0; 576 if (qp->r_rq.wq) { 577 qp->r_rq.wq->head = 0; 578 qp->r_rq.wq->tail = 0; 579 } 580 qp->r_sge.num_sge = 0; 581 } 582 583 /** 584 * rvt_create_qp - create a queue pair for a device 585 * @ibpd: the protection domain who's device we create the queue pair for 586 * @init_attr: the attributes of the queue pair 587 * @udata: user data for libibverbs.so 588 * 589 * Queue pair creation is mostly an rvt issue. However, drivers have their own 590 * unique idea of what queue pair numbers mean. For instance there is a reserved 591 * range for PSM. 592 * 593 * Return: the queue pair on success, otherwise returns an errno. 594 * 595 * Called by the ib_create_qp() core verbs function. 596 */ 597 struct ib_qp *rvt_create_qp(struct ib_pd *ibpd, 598 struct ib_qp_init_attr *init_attr, 599 struct ib_udata *udata) 600 { 601 struct rvt_qp *qp; 602 int err; 603 struct rvt_swqe *swq = NULL; 604 size_t sz; 605 size_t sg_list_sz; 606 struct ib_qp *ret = ERR_PTR(-ENOMEM); 607 struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device); 608 void *priv = NULL; 609 gfp_t gfp; 610 611 if (!rdi) 612 return ERR_PTR(-EINVAL); 613 614 if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge || 615 init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr || 616 init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO)) 617 return ERR_PTR(-EINVAL); 618 619 /* GFP_NOIO is applicable to RC QP's only */ 620 621 if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO && 622 init_attr->qp_type != IB_QPT_RC) 623 return ERR_PTR(-EINVAL); 624 625 gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ? 626 GFP_NOIO : GFP_KERNEL; 627 628 /* Check receive queue parameters if no SRQ is specified. */ 629 if (!init_attr->srq) { 630 if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge || 631 init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr) 632 return ERR_PTR(-EINVAL); 633 634 if (init_attr->cap.max_send_sge + 635 init_attr->cap.max_send_wr + 636 init_attr->cap.max_recv_sge + 637 init_attr->cap.max_recv_wr == 0) 638 return ERR_PTR(-EINVAL); 639 } 640 641 switch (init_attr->qp_type) { 642 case IB_QPT_SMI: 643 case IB_QPT_GSI: 644 if (init_attr->port_num == 0 || 645 init_attr->port_num > ibpd->device->phys_port_cnt) 646 return ERR_PTR(-EINVAL); 647 case IB_QPT_UC: 648 case IB_QPT_RC: 649 case IB_QPT_UD: 650 sz = sizeof(struct rvt_sge) * 651 init_attr->cap.max_send_sge + 652 sizeof(struct rvt_swqe); 653 if (gfp == GFP_NOIO) 654 swq = __vmalloc( 655 (init_attr->cap.max_send_wr + 1) * sz, 656 gfp, PAGE_KERNEL); 657 else 658 swq = vmalloc_node( 659 (init_attr->cap.max_send_wr + 1) * sz, 660 rdi->dparms.node); 661 if (!swq) 662 return ERR_PTR(-ENOMEM); 663 664 sz = sizeof(*qp); 665 sg_list_sz = 0; 666 if (init_attr->srq) { 667 struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq); 668 669 if (srq->rq.max_sge > 1) 670 sg_list_sz = sizeof(*qp->r_sg_list) * 671 (srq->rq.max_sge - 1); 672 } else if (init_attr->cap.max_recv_sge > 1) 673 sg_list_sz = sizeof(*qp->r_sg_list) * 674 (init_attr->cap.max_recv_sge - 1); 675 qp = kzalloc_node(sz + sg_list_sz, gfp, rdi->dparms.node); 676 if (!qp) 677 goto bail_swq; 678 679 RCU_INIT_POINTER(qp->next, NULL); 680 681 /* 682 * Driver needs to set up it's private QP structure and do any 683 * initialization that is needed. 684 */ 685 priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp); 686 if (!priv) 687 goto bail_qp; 688 qp->priv = priv; 689 qp->timeout_jiffies = 690 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / 691 1000UL); 692 if (init_attr->srq) { 693 sz = 0; 694 } else { 695 qp->r_rq.size = init_attr->cap.max_recv_wr + 1; 696 qp->r_rq.max_sge = init_attr->cap.max_recv_sge; 697 sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + 698 sizeof(struct rvt_rwqe); 699 if (udata) 700 qp->r_rq.wq = vmalloc_user( 701 sizeof(struct rvt_rwq) + 702 qp->r_rq.size * sz); 703 else if (gfp == GFP_NOIO) 704 qp->r_rq.wq = __vmalloc( 705 sizeof(struct rvt_rwq) + 706 qp->r_rq.size * sz, 707 gfp, PAGE_KERNEL); 708 else 709 qp->r_rq.wq = vmalloc_node( 710 sizeof(struct rvt_rwq) + 711 qp->r_rq.size * sz, 712 rdi->dparms.node); 713 if (!qp->r_rq.wq) 714 goto bail_driver_priv; 715 } 716 717 /* 718 * ib_create_qp() will initialize qp->ibqp 719 * except for qp->ibqp.qp_num. 720 */ 721 spin_lock_init(&qp->r_lock); 722 spin_lock_init(&qp->s_hlock); 723 spin_lock_init(&qp->s_lock); 724 spin_lock_init(&qp->r_rq.lock); 725 atomic_set(&qp->refcount, 0); 726 init_waitqueue_head(&qp->wait); 727 init_timer(&qp->s_timer); 728 qp->s_timer.data = (unsigned long)qp; 729 INIT_LIST_HEAD(&qp->rspwait); 730 qp->state = IB_QPS_RESET; 731 qp->s_wq = swq; 732 qp->s_size = init_attr->cap.max_send_wr + 1; 733 qp->s_avail = init_attr->cap.max_send_wr; 734 qp->s_max_sge = init_attr->cap.max_send_sge; 735 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) 736 qp->s_flags = RVT_S_SIGNAL_REQ_WR; 737 738 err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table, 739 init_attr->qp_type, 740 init_attr->port_num, gfp); 741 if (err < 0) { 742 ret = ERR_PTR(err); 743 goto bail_rq_wq; 744 } 745 qp->ibqp.qp_num = err; 746 qp->port_num = init_attr->port_num; 747 rvt_reset_qp(rdi, qp, init_attr->qp_type); 748 break; 749 750 default: 751 /* Don't support raw QPs */ 752 return ERR_PTR(-EINVAL); 753 } 754 755 init_attr->cap.max_inline_data = 0; 756 757 /* 758 * Return the address of the RWQ as the offset to mmap. 759 * See rvt_mmap() for details. 760 */ 761 if (udata && udata->outlen >= sizeof(__u64)) { 762 if (!qp->r_rq.wq) { 763 __u64 offset = 0; 764 765 err = ib_copy_to_udata(udata, &offset, 766 sizeof(offset)); 767 if (err) { 768 ret = ERR_PTR(err); 769 goto bail_qpn; 770 } 771 } else { 772 u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz; 773 774 qp->ip = rvt_create_mmap_info(rdi, s, 775 ibpd->uobject->context, 776 qp->r_rq.wq); 777 if (!qp->ip) { 778 ret = ERR_PTR(-ENOMEM); 779 goto bail_qpn; 780 } 781 782 err = ib_copy_to_udata(udata, &qp->ip->offset, 783 sizeof(qp->ip->offset)); 784 if (err) { 785 ret = ERR_PTR(err); 786 goto bail_ip; 787 } 788 } 789 qp->pid = current->pid; 790 } 791 792 spin_lock(&rdi->n_qps_lock); 793 if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) { 794 spin_unlock(&rdi->n_qps_lock); 795 ret = ERR_PTR(-ENOMEM); 796 goto bail_ip; 797 } 798 799 rdi->n_qps_allocated++; 800 /* 801 * Maintain a busy_jiffies variable that will be added to the timeout 802 * period in mod_retry_timer and add_retry_timer. This busy jiffies 803 * is scaled by the number of rc qps created for the device to reduce 804 * the number of timeouts occurring when there is a large number of 805 * qps. busy_jiffies is incremented every rc qp scaling interval. 806 * The scaling interval is selected based on extensive performance 807 * evaluation of targeted workloads. 808 */ 809 if (init_attr->qp_type == IB_QPT_RC) { 810 rdi->n_rc_qps++; 811 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL; 812 } 813 spin_unlock(&rdi->n_qps_lock); 814 815 if (qp->ip) { 816 spin_lock_irq(&rdi->pending_lock); 817 list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps); 818 spin_unlock_irq(&rdi->pending_lock); 819 } 820 821 ret = &qp->ibqp; 822 823 /* 824 * We have our QP and its good, now keep track of what types of opcodes 825 * can be processed on this QP. We do this by keeping track of what the 826 * 3 high order bits of the opcode are. 827 */ 828 switch (init_attr->qp_type) { 829 case IB_QPT_SMI: 830 case IB_QPT_GSI: 831 case IB_QPT_UD: 832 qp->allowed_ops = IB_OPCODE_UD_SEND_ONLY & RVT_OPCODE_QP_MASK; 833 break; 834 case IB_QPT_RC: 835 qp->allowed_ops = IB_OPCODE_RC_SEND_ONLY & RVT_OPCODE_QP_MASK; 836 break; 837 case IB_QPT_UC: 838 qp->allowed_ops = IB_OPCODE_UC_SEND_ONLY & RVT_OPCODE_QP_MASK; 839 break; 840 default: 841 ret = ERR_PTR(-EINVAL); 842 goto bail_ip; 843 } 844 845 return ret; 846 847 bail_ip: 848 kref_put(&qp->ip->ref, rvt_release_mmap_info); 849 850 bail_qpn: 851 free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); 852 853 bail_rq_wq: 854 vfree(qp->r_rq.wq); 855 856 bail_driver_priv: 857 rdi->driver_f.qp_priv_free(rdi, qp); 858 859 bail_qp: 860 kfree(qp); 861 862 bail_swq: 863 vfree(swq); 864 865 return ret; 866 } 867 868 /** 869 * rvt_error_qp - put a QP into the error state 870 * @qp: the QP to put into the error state 871 * @err: the receive completion error to signal if a RWQE is active 872 * 873 * Flushes both send and receive work queues. 874 * 875 * Return: true if last WQE event should be generated. 876 * The QP r_lock and s_lock should be held and interrupts disabled. 877 * If we are already in error state, just return. 878 */ 879 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err) 880 { 881 struct ib_wc wc; 882 int ret = 0; 883 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 884 885 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) 886 goto bail; 887 888 qp->state = IB_QPS_ERR; 889 890 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { 891 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); 892 del_timer(&qp->s_timer); 893 } 894 895 if (qp->s_flags & RVT_S_ANY_WAIT_SEND) 896 qp->s_flags &= ~RVT_S_ANY_WAIT_SEND; 897 898 rdi->driver_f.notify_error_qp(qp); 899 900 /* Schedule the sending tasklet to drain the send work queue. */ 901 if (ACCESS_ONCE(qp->s_last) != qp->s_head) 902 rdi->driver_f.schedule_send(qp); 903 904 rvt_clear_mr_refs(qp, 0); 905 906 memset(&wc, 0, sizeof(wc)); 907 wc.qp = &qp->ibqp; 908 wc.opcode = IB_WC_RECV; 909 910 if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) { 911 wc.wr_id = qp->r_wr_id; 912 wc.status = err; 913 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 914 } 915 wc.status = IB_WC_WR_FLUSH_ERR; 916 917 if (qp->r_rq.wq) { 918 struct rvt_rwq *wq; 919 u32 head; 920 u32 tail; 921 922 spin_lock(&qp->r_rq.lock); 923 924 /* sanity check pointers before trusting them */ 925 wq = qp->r_rq.wq; 926 head = wq->head; 927 if (head >= qp->r_rq.size) 928 head = 0; 929 tail = wq->tail; 930 if (tail >= qp->r_rq.size) 931 tail = 0; 932 while (tail != head) { 933 wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id; 934 if (++tail >= qp->r_rq.size) 935 tail = 0; 936 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 937 } 938 wq->tail = tail; 939 940 spin_unlock(&qp->r_rq.lock); 941 } else if (qp->ibqp.event_handler) { 942 ret = 1; 943 } 944 945 bail: 946 return ret; 947 } 948 EXPORT_SYMBOL(rvt_error_qp); 949 950 /* 951 * Put the QP into the hash table. 952 * The hash table holds a reference to the QP. 953 */ 954 static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) 955 { 956 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; 957 unsigned long flags; 958 959 atomic_inc(&qp->refcount); 960 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); 961 962 if (qp->ibqp.qp_num <= 1) { 963 rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp); 964 } else { 965 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); 966 967 qp->next = rdi->qp_dev->qp_table[n]; 968 rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp); 969 trace_rvt_qpinsert(qp, n); 970 } 971 972 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); 973 } 974 975 /** 976 * qib_modify_qp - modify the attributes of a queue pair 977 * @ibqp: the queue pair who's attributes we're modifying 978 * @attr: the new attributes 979 * @attr_mask: the mask of attributes to modify 980 * @udata: user data for libibverbs.so 981 * 982 * Return: 0 on success, otherwise returns an errno. 983 */ 984 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 985 int attr_mask, struct ib_udata *udata) 986 { 987 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 988 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 989 enum ib_qp_state cur_state, new_state; 990 struct ib_event ev; 991 int lastwqe = 0; 992 int mig = 0; 993 int pmtu = 0; /* for gcc warning only */ 994 enum rdma_link_layer link; 995 996 link = rdma_port_get_link_layer(ibqp->device, qp->port_num); 997 998 spin_lock_irq(&qp->r_lock); 999 spin_lock(&qp->s_hlock); 1000 spin_lock(&qp->s_lock); 1001 1002 cur_state = attr_mask & IB_QP_CUR_STATE ? 1003 attr->cur_qp_state : qp->state; 1004 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 1005 1006 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 1007 attr_mask, link)) 1008 goto inval; 1009 1010 if (rdi->driver_f.check_modify_qp && 1011 rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata)) 1012 goto inval; 1013 1014 if (attr_mask & IB_QP_AV) { 1015 if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) 1016 goto inval; 1017 if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr)) 1018 goto inval; 1019 } 1020 1021 if (attr_mask & IB_QP_ALT_PATH) { 1022 if (attr->alt_ah_attr.dlid >= 1023 be16_to_cpu(IB_MULTICAST_LID_BASE)) 1024 goto inval; 1025 if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr)) 1026 goto inval; 1027 if (attr->alt_pkey_index >= rvt_get_npkeys(rdi)) 1028 goto inval; 1029 } 1030 1031 if (attr_mask & IB_QP_PKEY_INDEX) 1032 if (attr->pkey_index >= rvt_get_npkeys(rdi)) 1033 goto inval; 1034 1035 if (attr_mask & IB_QP_MIN_RNR_TIMER) 1036 if (attr->min_rnr_timer > 31) 1037 goto inval; 1038 1039 if (attr_mask & IB_QP_PORT) 1040 if (qp->ibqp.qp_type == IB_QPT_SMI || 1041 qp->ibqp.qp_type == IB_QPT_GSI || 1042 attr->port_num == 0 || 1043 attr->port_num > ibqp->device->phys_port_cnt) 1044 goto inval; 1045 1046 if (attr_mask & IB_QP_DEST_QPN) 1047 if (attr->dest_qp_num > RVT_QPN_MASK) 1048 goto inval; 1049 1050 if (attr_mask & IB_QP_RETRY_CNT) 1051 if (attr->retry_cnt > 7) 1052 goto inval; 1053 1054 if (attr_mask & IB_QP_RNR_RETRY) 1055 if (attr->rnr_retry > 7) 1056 goto inval; 1057 1058 /* 1059 * Don't allow invalid path_mtu values. OK to set greater 1060 * than the active mtu (or even the max_cap, if we have tuned 1061 * that to a small mtu. We'll set qp->path_mtu 1062 * to the lesser of requested attribute mtu and active, 1063 * for packetizing messages. 1064 * Note that the QP port has to be set in INIT and MTU in RTR. 1065 */ 1066 if (attr_mask & IB_QP_PATH_MTU) { 1067 pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr); 1068 if (pmtu < 0) 1069 goto inval; 1070 } 1071 1072 if (attr_mask & IB_QP_PATH_MIG_STATE) { 1073 if (attr->path_mig_state == IB_MIG_REARM) { 1074 if (qp->s_mig_state == IB_MIG_ARMED) 1075 goto inval; 1076 if (new_state != IB_QPS_RTS) 1077 goto inval; 1078 } else if (attr->path_mig_state == IB_MIG_MIGRATED) { 1079 if (qp->s_mig_state == IB_MIG_REARM) 1080 goto inval; 1081 if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD) 1082 goto inval; 1083 if (qp->s_mig_state == IB_MIG_ARMED) 1084 mig = 1; 1085 } else { 1086 goto inval; 1087 } 1088 } 1089 1090 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1091 if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic) 1092 goto inval; 1093 1094 switch (new_state) { 1095 case IB_QPS_RESET: 1096 if (qp->state != IB_QPS_RESET) 1097 rvt_reset_qp(rdi, qp, ibqp->qp_type); 1098 break; 1099 1100 case IB_QPS_RTR: 1101 /* Allow event to re-trigger if QP set to RTR more than once */ 1102 qp->r_flags &= ~RVT_R_COMM_EST; 1103 qp->state = new_state; 1104 break; 1105 1106 case IB_QPS_SQD: 1107 qp->s_draining = qp->s_last != qp->s_cur; 1108 qp->state = new_state; 1109 break; 1110 1111 case IB_QPS_SQE: 1112 if (qp->ibqp.qp_type == IB_QPT_RC) 1113 goto inval; 1114 qp->state = new_state; 1115 break; 1116 1117 case IB_QPS_ERR: 1118 lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1119 break; 1120 1121 default: 1122 qp->state = new_state; 1123 break; 1124 } 1125 1126 if (attr_mask & IB_QP_PKEY_INDEX) 1127 qp->s_pkey_index = attr->pkey_index; 1128 1129 if (attr_mask & IB_QP_PORT) 1130 qp->port_num = attr->port_num; 1131 1132 if (attr_mask & IB_QP_DEST_QPN) 1133 qp->remote_qpn = attr->dest_qp_num; 1134 1135 if (attr_mask & IB_QP_SQ_PSN) { 1136 qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask; 1137 qp->s_psn = qp->s_next_psn; 1138 qp->s_sending_psn = qp->s_next_psn; 1139 qp->s_last_psn = qp->s_next_psn - 1; 1140 qp->s_sending_hpsn = qp->s_last_psn; 1141 } 1142 1143 if (attr_mask & IB_QP_RQ_PSN) 1144 qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask; 1145 1146 if (attr_mask & IB_QP_ACCESS_FLAGS) 1147 qp->qp_access_flags = attr->qp_access_flags; 1148 1149 if (attr_mask & IB_QP_AV) { 1150 qp->remote_ah_attr = attr->ah_attr; 1151 qp->s_srate = attr->ah_attr.static_rate; 1152 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); 1153 } 1154 1155 if (attr_mask & IB_QP_ALT_PATH) { 1156 qp->alt_ah_attr = attr->alt_ah_attr; 1157 qp->s_alt_pkey_index = attr->alt_pkey_index; 1158 } 1159 1160 if (attr_mask & IB_QP_PATH_MIG_STATE) { 1161 qp->s_mig_state = attr->path_mig_state; 1162 if (mig) { 1163 qp->remote_ah_attr = qp->alt_ah_attr; 1164 qp->port_num = qp->alt_ah_attr.port_num; 1165 qp->s_pkey_index = qp->s_alt_pkey_index; 1166 } 1167 } 1168 1169 if (attr_mask & IB_QP_PATH_MTU) { 1170 qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu); 1171 qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu); 1172 qp->log_pmtu = ilog2(qp->pmtu); 1173 } 1174 1175 if (attr_mask & IB_QP_RETRY_CNT) { 1176 qp->s_retry_cnt = attr->retry_cnt; 1177 qp->s_retry = attr->retry_cnt; 1178 } 1179 1180 if (attr_mask & IB_QP_RNR_RETRY) { 1181 qp->s_rnr_retry_cnt = attr->rnr_retry; 1182 qp->s_rnr_retry = attr->rnr_retry; 1183 } 1184 1185 if (attr_mask & IB_QP_MIN_RNR_TIMER) 1186 qp->r_min_rnr_timer = attr->min_rnr_timer; 1187 1188 if (attr_mask & IB_QP_TIMEOUT) { 1189 qp->timeout = attr->timeout; 1190 qp->timeout_jiffies = 1191 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / 1192 1000UL); 1193 } 1194 1195 if (attr_mask & IB_QP_QKEY) 1196 qp->qkey = attr->qkey; 1197 1198 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1199 qp->r_max_rd_atomic = attr->max_dest_rd_atomic; 1200 1201 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) 1202 qp->s_max_rd_atomic = attr->max_rd_atomic; 1203 1204 if (rdi->driver_f.modify_qp) 1205 rdi->driver_f.modify_qp(qp, attr, attr_mask, udata); 1206 1207 spin_unlock(&qp->s_lock); 1208 spin_unlock(&qp->s_hlock); 1209 spin_unlock_irq(&qp->r_lock); 1210 1211 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 1212 rvt_insert_qp(rdi, qp); 1213 1214 if (lastwqe) { 1215 ev.device = qp->ibqp.device; 1216 ev.element.qp = &qp->ibqp; 1217 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 1218 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1219 } 1220 if (mig) { 1221 ev.device = qp->ibqp.device; 1222 ev.element.qp = &qp->ibqp; 1223 ev.event = IB_EVENT_PATH_MIG; 1224 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1225 } 1226 return 0; 1227 1228 inval: 1229 spin_unlock(&qp->s_lock); 1230 spin_unlock(&qp->s_hlock); 1231 spin_unlock_irq(&qp->r_lock); 1232 return -EINVAL; 1233 } 1234 1235 /** rvt_free_qpn - Free a qpn from the bit map 1236 * @qpt: QP table 1237 * @qpn: queue pair number to free 1238 */ 1239 static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn) 1240 { 1241 struct rvt_qpn_map *map; 1242 1243 map = qpt->map + qpn / RVT_BITS_PER_PAGE; 1244 if (map->page) 1245 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); 1246 } 1247 1248 /** 1249 * rvt_destroy_qp - destroy a queue pair 1250 * @ibqp: the queue pair to destroy 1251 * 1252 * Note that this can be called while the QP is actively sending or 1253 * receiving! 1254 * 1255 * Return: 0 on success. 1256 */ 1257 int rvt_destroy_qp(struct ib_qp *ibqp) 1258 { 1259 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1260 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1261 1262 spin_lock_irq(&qp->r_lock); 1263 spin_lock(&qp->s_hlock); 1264 spin_lock(&qp->s_lock); 1265 rvt_reset_qp(rdi, qp, ibqp->qp_type); 1266 spin_unlock(&qp->s_lock); 1267 spin_unlock(&qp->s_hlock); 1268 spin_unlock_irq(&qp->r_lock); 1269 1270 /* qpn is now available for use again */ 1271 rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); 1272 1273 spin_lock(&rdi->n_qps_lock); 1274 rdi->n_qps_allocated--; 1275 if (qp->ibqp.qp_type == IB_QPT_RC) { 1276 rdi->n_rc_qps--; 1277 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL; 1278 } 1279 spin_unlock(&rdi->n_qps_lock); 1280 1281 if (qp->ip) 1282 kref_put(&qp->ip->ref, rvt_release_mmap_info); 1283 else 1284 vfree(qp->r_rq.wq); 1285 vfree(qp->s_wq); 1286 rdi->driver_f.qp_priv_free(rdi, qp); 1287 kfree(qp); 1288 return 0; 1289 } 1290 1291 /** 1292 * rvt_query_qp - query an ipbq 1293 * @ibqp: IB qp to query 1294 * @attr: attr struct to fill in 1295 * @attr_mask: attr mask ignored 1296 * @init_attr: struct to fill in 1297 * 1298 * Return: always 0 1299 */ 1300 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 1301 int attr_mask, struct ib_qp_init_attr *init_attr) 1302 { 1303 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1304 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1305 1306 attr->qp_state = qp->state; 1307 attr->cur_qp_state = attr->qp_state; 1308 attr->path_mtu = qp->path_mtu; 1309 attr->path_mig_state = qp->s_mig_state; 1310 attr->qkey = qp->qkey; 1311 attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask; 1312 attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask; 1313 attr->dest_qp_num = qp->remote_qpn; 1314 attr->qp_access_flags = qp->qp_access_flags; 1315 attr->cap.max_send_wr = qp->s_size - 1; 1316 attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1; 1317 attr->cap.max_send_sge = qp->s_max_sge; 1318 attr->cap.max_recv_sge = qp->r_rq.max_sge; 1319 attr->cap.max_inline_data = 0; 1320 attr->ah_attr = qp->remote_ah_attr; 1321 attr->alt_ah_attr = qp->alt_ah_attr; 1322 attr->pkey_index = qp->s_pkey_index; 1323 attr->alt_pkey_index = qp->s_alt_pkey_index; 1324 attr->en_sqd_async_notify = 0; 1325 attr->sq_draining = qp->s_draining; 1326 attr->max_rd_atomic = qp->s_max_rd_atomic; 1327 attr->max_dest_rd_atomic = qp->r_max_rd_atomic; 1328 attr->min_rnr_timer = qp->r_min_rnr_timer; 1329 attr->port_num = qp->port_num; 1330 attr->timeout = qp->timeout; 1331 attr->retry_cnt = qp->s_retry_cnt; 1332 attr->rnr_retry = qp->s_rnr_retry_cnt; 1333 attr->alt_port_num = qp->alt_ah_attr.port_num; 1334 attr->alt_timeout = qp->alt_timeout; 1335 1336 init_attr->event_handler = qp->ibqp.event_handler; 1337 init_attr->qp_context = qp->ibqp.qp_context; 1338 init_attr->send_cq = qp->ibqp.send_cq; 1339 init_attr->recv_cq = qp->ibqp.recv_cq; 1340 init_attr->srq = qp->ibqp.srq; 1341 init_attr->cap = attr->cap; 1342 if (qp->s_flags & RVT_S_SIGNAL_REQ_WR) 1343 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR; 1344 else 1345 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR; 1346 init_attr->qp_type = qp->ibqp.qp_type; 1347 init_attr->port_num = qp->port_num; 1348 return 0; 1349 } 1350 1351 /** 1352 * rvt_post_receive - post a receive on a QP 1353 * @ibqp: the QP to post the receive on 1354 * @wr: the WR to post 1355 * @bad_wr: the first bad WR is put here 1356 * 1357 * This may be called from interrupt context. 1358 * 1359 * Return: 0 on success otherwise errno 1360 */ 1361 int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 1362 struct ib_recv_wr **bad_wr) 1363 { 1364 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1365 struct rvt_rwq *wq = qp->r_rq.wq; 1366 unsigned long flags; 1367 int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) && 1368 !qp->ibqp.srq; 1369 1370 /* Check that state is OK to post receive. */ 1371 if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) { 1372 *bad_wr = wr; 1373 return -EINVAL; 1374 } 1375 1376 for (; wr; wr = wr->next) { 1377 struct rvt_rwqe *wqe; 1378 u32 next; 1379 int i; 1380 1381 if ((unsigned)wr->num_sge > qp->r_rq.max_sge) { 1382 *bad_wr = wr; 1383 return -EINVAL; 1384 } 1385 1386 spin_lock_irqsave(&qp->r_rq.lock, flags); 1387 next = wq->head + 1; 1388 if (next >= qp->r_rq.size) 1389 next = 0; 1390 if (next == wq->tail) { 1391 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 1392 *bad_wr = wr; 1393 return -ENOMEM; 1394 } 1395 if (unlikely(qp_err_flush)) { 1396 struct ib_wc wc; 1397 1398 memset(&wc, 0, sizeof(wc)); 1399 wc.qp = &qp->ibqp; 1400 wc.opcode = IB_WC_RECV; 1401 wc.wr_id = wr->wr_id; 1402 wc.status = IB_WC_WR_FLUSH_ERR; 1403 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 1404 } else { 1405 wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head); 1406 wqe->wr_id = wr->wr_id; 1407 wqe->num_sge = wr->num_sge; 1408 for (i = 0; i < wr->num_sge; i++) 1409 wqe->sg_list[i] = wr->sg_list[i]; 1410 /* 1411 * Make sure queue entry is written 1412 * before the head index. 1413 */ 1414 smp_wmb(); 1415 wq->head = next; 1416 } 1417 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 1418 } 1419 return 0; 1420 } 1421 1422 /** 1423 * qp_get_savail - return number of avail send entries 1424 * 1425 * @qp - the qp 1426 * 1427 * This assumes the s_hlock is held but the s_last 1428 * qp variable is uncontrolled. 1429 */ 1430 static inline u32 qp_get_savail(struct rvt_qp *qp) 1431 { 1432 u32 slast; 1433 u32 ret; 1434 1435 smp_read_barrier_depends(); /* see rc.c */ 1436 slast = ACCESS_ONCE(qp->s_last); 1437 if (qp->s_head >= slast) 1438 ret = qp->s_size - (qp->s_head - slast); 1439 else 1440 ret = slast - qp->s_head; 1441 return ret - 1; 1442 } 1443 1444 /** 1445 * rvt_post_one_wr - post one RC, UC, or UD send work request 1446 * @qp: the QP to post on 1447 * @wr: the work request to send 1448 */ 1449 static int rvt_post_one_wr(struct rvt_qp *qp, 1450 struct ib_send_wr *wr, 1451 int *call_send) 1452 { 1453 struct rvt_swqe *wqe; 1454 u32 next; 1455 int i; 1456 int j; 1457 int acc; 1458 struct rvt_lkey_table *rkt; 1459 struct rvt_pd *pd; 1460 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 1461 u8 log_pmtu; 1462 int ret; 1463 1464 /* IB spec says that num_sge == 0 is OK. */ 1465 if (unlikely(wr->num_sge > qp->s_max_sge)) 1466 return -EINVAL; 1467 1468 /* 1469 * Don't allow RDMA reads or atomic operations on UC or 1470 * undefined operations. 1471 * Make sure buffer is large enough to hold the result for atomics. 1472 */ 1473 if (qp->ibqp.qp_type == IB_QPT_UC) { 1474 if ((unsigned)wr->opcode >= IB_WR_RDMA_READ) 1475 return -EINVAL; 1476 } else if (qp->ibqp.qp_type != IB_QPT_RC) { 1477 /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */ 1478 if (wr->opcode != IB_WR_SEND && 1479 wr->opcode != IB_WR_SEND_WITH_IMM) 1480 return -EINVAL; 1481 /* Check UD destination address PD */ 1482 if (qp->ibqp.pd != ud_wr(wr)->ah->pd) 1483 return -EINVAL; 1484 } else if ((unsigned)wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) { 1485 return -EINVAL; 1486 } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP && 1487 (wr->num_sge == 0 || 1488 wr->sg_list[0].length < sizeof(u64) || 1489 wr->sg_list[0].addr & (sizeof(u64) - 1))) { 1490 return -EINVAL; 1491 } else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) { 1492 return -EINVAL; 1493 } 1494 /* check for avail */ 1495 if (unlikely(!qp->s_avail)) { 1496 qp->s_avail = qp_get_savail(qp); 1497 if (WARN_ON(qp->s_avail > (qp->s_size - 1))) 1498 rvt_pr_err(rdi, 1499 "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u", 1500 qp->ibqp.qp_num, qp->s_size, qp->s_avail, 1501 qp->s_head, qp->s_tail, qp->s_cur, 1502 qp->s_acked, qp->s_last); 1503 if (!qp->s_avail) 1504 return -ENOMEM; 1505 } 1506 next = qp->s_head + 1; 1507 if (next >= qp->s_size) 1508 next = 0; 1509 1510 rkt = &rdi->lkey_table; 1511 pd = ibpd_to_rvtpd(qp->ibqp.pd); 1512 wqe = rvt_get_swqe_ptr(qp, qp->s_head); 1513 1514 if (qp->ibqp.qp_type != IB_QPT_UC && 1515 qp->ibqp.qp_type != IB_QPT_RC) 1516 memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr)); 1517 else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM || 1518 wr->opcode == IB_WR_RDMA_WRITE || 1519 wr->opcode == IB_WR_RDMA_READ) 1520 memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr)); 1521 else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1522 wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) 1523 memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr)); 1524 else 1525 memcpy(&wqe->wr, wr, sizeof(wqe->wr)); 1526 1527 wqe->length = 0; 1528 j = 0; 1529 if (wr->num_sge) { 1530 acc = wr->opcode >= IB_WR_RDMA_READ ? 1531 IB_ACCESS_LOCAL_WRITE : 0; 1532 for (i = 0; i < wr->num_sge; i++) { 1533 u32 length = wr->sg_list[i].length; 1534 int ok; 1535 1536 if (length == 0) 1537 continue; 1538 ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j], 1539 &wr->sg_list[i], acc); 1540 if (!ok) { 1541 ret = -EINVAL; 1542 goto bail_inval_free; 1543 } 1544 wqe->length += length; 1545 j++; 1546 } 1547 wqe->wr.num_sge = j; 1548 } 1549 1550 /* general part of wqe valid - allow for driver checks */ 1551 if (rdi->driver_f.check_send_wqe) { 1552 ret = rdi->driver_f.check_send_wqe(qp, wqe); 1553 if (ret < 0) 1554 goto bail_inval_free; 1555 if (ret) 1556 *call_send = ret; 1557 } 1558 1559 log_pmtu = qp->log_pmtu; 1560 if (qp->ibqp.qp_type != IB_QPT_UC && 1561 qp->ibqp.qp_type != IB_QPT_RC) { 1562 struct rvt_ah *ah = ibah_to_rvtah(wqe->ud_wr.ah); 1563 1564 log_pmtu = ah->log_pmtu; 1565 atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount); 1566 } 1567 1568 wqe->ssn = qp->s_ssn++; 1569 wqe->psn = qp->s_next_psn; 1570 wqe->lpsn = wqe->psn + 1571 (wqe->length ? ((wqe->length - 1) >> log_pmtu) : 0); 1572 qp->s_next_psn = wqe->lpsn + 1; 1573 trace_rvt_post_one_wr(qp, wqe); 1574 smp_wmb(); /* see request builders */ 1575 qp->s_avail--; 1576 qp->s_head = next; 1577 1578 return 0; 1579 1580 bail_inval_free: 1581 /* release mr holds */ 1582 while (j) { 1583 struct rvt_sge *sge = &wqe->sg_list[--j]; 1584 1585 rvt_put_mr(sge->mr); 1586 } 1587 return ret; 1588 } 1589 1590 /** 1591 * rvt_post_send - post a send on a QP 1592 * @ibqp: the QP to post the send on 1593 * @wr: the list of work requests to post 1594 * @bad_wr: the first bad WR is put here 1595 * 1596 * This may be called from interrupt context. 1597 * 1598 * Return: 0 on success else errno 1599 */ 1600 int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 1601 struct ib_send_wr **bad_wr) 1602 { 1603 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1604 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1605 unsigned long flags = 0; 1606 int call_send; 1607 unsigned nreq = 0; 1608 int err = 0; 1609 1610 spin_lock_irqsave(&qp->s_hlock, flags); 1611 1612 /* 1613 * Ensure QP state is such that we can send. If not bail out early, 1614 * there is no need to do this every time we post a send. 1615 */ 1616 if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) { 1617 spin_unlock_irqrestore(&qp->s_hlock, flags); 1618 return -EINVAL; 1619 } 1620 1621 /* 1622 * If the send queue is empty, and we only have a single WR then just go 1623 * ahead and kick the send engine into gear. Otherwise we will always 1624 * just schedule the send to happen later. 1625 */ 1626 call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next; 1627 1628 for (; wr; wr = wr->next) { 1629 err = rvt_post_one_wr(qp, wr, &call_send); 1630 if (unlikely(err)) { 1631 *bad_wr = wr; 1632 goto bail; 1633 } 1634 nreq++; 1635 } 1636 bail: 1637 spin_unlock_irqrestore(&qp->s_hlock, flags); 1638 if (nreq) { 1639 if (call_send) 1640 rdi->driver_f.do_send(qp); 1641 else 1642 rdi->driver_f.schedule_send_no_lock(qp); 1643 } 1644 return err; 1645 } 1646 1647 /** 1648 * rvt_post_srq_receive - post a receive on a shared receive queue 1649 * @ibsrq: the SRQ to post the receive on 1650 * @wr: the list of work requests to post 1651 * @bad_wr: A pointer to the first WR to cause a problem is put here 1652 * 1653 * This may be called from interrupt context. 1654 * 1655 * Return: 0 on success else errno 1656 */ 1657 int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 1658 struct ib_recv_wr **bad_wr) 1659 { 1660 struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq); 1661 struct rvt_rwq *wq; 1662 unsigned long flags; 1663 1664 for (; wr; wr = wr->next) { 1665 struct rvt_rwqe *wqe; 1666 u32 next; 1667 int i; 1668 1669 if ((unsigned)wr->num_sge > srq->rq.max_sge) { 1670 *bad_wr = wr; 1671 return -EINVAL; 1672 } 1673 1674 spin_lock_irqsave(&srq->rq.lock, flags); 1675 wq = srq->rq.wq; 1676 next = wq->head + 1; 1677 if (next >= srq->rq.size) 1678 next = 0; 1679 if (next == wq->tail) { 1680 spin_unlock_irqrestore(&srq->rq.lock, flags); 1681 *bad_wr = wr; 1682 return -ENOMEM; 1683 } 1684 1685 wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head); 1686 wqe->wr_id = wr->wr_id; 1687 wqe->num_sge = wr->num_sge; 1688 for (i = 0; i < wr->num_sge; i++) 1689 wqe->sg_list[i] = wr->sg_list[i]; 1690 /* Make sure queue entry is written before the head index. */ 1691 smp_wmb(); 1692 wq->head = next; 1693 spin_unlock_irqrestore(&srq->rq.lock, flags); 1694 } 1695 return 0; 1696 } 1697