1 /* 2 * Broadcom NetXtreme-E RoCE driver. 3 * 4 * Copyright (c) 2016 - 2017, Broadcom. All rights reserved. The term 5 * Broadcom refers to Broadcom Limited and/or its subsidiaries. 6 * 7 * This software is available to you under a choice of one of two 8 * licenses. You may choose to be licensed under the terms of the GNU 9 * General Public License (GPL) Version 2, available from the file 10 * COPYING in the main directory of this source tree, or the 11 * BSD license below: 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in 21 * the documentation and/or other materials provided with the 22 * distribution. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' 25 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, 26 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 27 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS 28 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 29 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 30 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR 31 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 32 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE 33 * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN 34 * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 35 * 36 * Description: Fast Path Operators 37 */ 38 39 #define dev_fmt(fmt) "QPLIB: " fmt 40 41 #include <linux/interrupt.h> 42 #include <linux/spinlock.h> 43 #include <linux/sched.h> 44 #include <linux/slab.h> 45 #include <linux/pci.h> 46 #include <linux/delay.h> 47 #include <linux/prefetch.h> 48 #include <linux/if_ether.h> 49 #include <rdma/ib_mad.h> 50 51 #include "roce_hsi.h" 52 53 #include "qplib_res.h" 54 #include "qplib_rcfw.h" 55 #include "qplib_sp.h" 56 #include "qplib_fp.h" 57 58 static void __clean_cq(struct bnxt_qplib_cq *cq, u64 qp); 59 60 static void bnxt_qplib_cancel_phantom_processing(struct bnxt_qplib_qp *qp) 61 { 62 qp->sq.condition = false; 63 qp->sq.send_phantom = false; 64 qp->sq.single = false; 65 } 66 67 /* Flush list */ 68 static void __bnxt_qplib_add_flush_qp(struct bnxt_qplib_qp *qp) 69 { 70 struct bnxt_qplib_cq *scq, *rcq; 71 72 scq = qp->scq; 73 rcq = qp->rcq; 74 75 if (!qp->sq.flushed) { 76 dev_dbg(&scq->hwq.pdev->dev, 77 "FP: Adding to SQ Flush list = %p\n", qp); 78 bnxt_qplib_cancel_phantom_processing(qp); 79 list_add_tail(&qp->sq_flush, &scq->sqf_head); 80 qp->sq.flushed = true; 81 } 82 if (!qp->srq) { 83 if (!qp->rq.flushed) { 84 dev_dbg(&rcq->hwq.pdev->dev, 85 "FP: Adding to RQ Flush list = %p\n", qp); 86 list_add_tail(&qp->rq_flush, &rcq->rqf_head); 87 qp->rq.flushed = true; 88 } 89 } 90 } 91 92 static void bnxt_qplib_acquire_cq_flush_locks(struct bnxt_qplib_qp *qp, 93 unsigned long *flags) 94 __acquires(&qp->scq->flush_lock) __acquires(&qp->rcq->flush_lock) 95 { 96 spin_lock_irqsave(&qp->scq->flush_lock, *flags); 97 if (qp->scq == qp->rcq) 98 __acquire(&qp->rcq->flush_lock); 99 else 100 spin_lock(&qp->rcq->flush_lock); 101 } 102 103 static void bnxt_qplib_release_cq_flush_locks(struct bnxt_qplib_qp *qp, 104 unsigned long *flags) 105 __releases(&qp->scq->flush_lock) __releases(&qp->rcq->flush_lock) 106 { 107 if (qp->scq == qp->rcq) 108 __release(&qp->rcq->flush_lock); 109 else 110 spin_unlock(&qp->rcq->flush_lock); 111 spin_unlock_irqrestore(&qp->scq->flush_lock, *flags); 112 } 113 114 void bnxt_qplib_add_flush_qp(struct bnxt_qplib_qp *qp) 115 { 116 unsigned long flags; 117 118 bnxt_qplib_acquire_cq_flush_locks(qp, &flags); 119 __bnxt_qplib_add_flush_qp(qp); 120 bnxt_qplib_release_cq_flush_locks(qp, &flags); 121 } 122 123 static void __bnxt_qplib_del_flush_qp(struct bnxt_qplib_qp *qp) 124 { 125 if (qp->sq.flushed) { 126 qp->sq.flushed = false; 127 list_del(&qp->sq_flush); 128 } 129 if (!qp->srq) { 130 if (qp->rq.flushed) { 131 qp->rq.flushed = false; 132 list_del(&qp->rq_flush); 133 } 134 } 135 } 136 137 void bnxt_qplib_clean_qp(struct bnxt_qplib_qp *qp) 138 { 139 unsigned long flags; 140 141 bnxt_qplib_acquire_cq_flush_locks(qp, &flags); 142 __clean_cq(qp->scq, (u64)(unsigned long)qp); 143 qp->sq.hwq.prod = 0; 144 qp->sq.hwq.cons = 0; 145 __clean_cq(qp->rcq, (u64)(unsigned long)qp); 146 qp->rq.hwq.prod = 0; 147 qp->rq.hwq.cons = 0; 148 149 __bnxt_qplib_del_flush_qp(qp); 150 bnxt_qplib_release_cq_flush_locks(qp, &flags); 151 } 152 153 static void bnxt_qpn_cqn_sched_task(struct work_struct *work) 154 { 155 struct bnxt_qplib_nq_work *nq_work = 156 container_of(work, struct bnxt_qplib_nq_work, work); 157 158 struct bnxt_qplib_cq *cq = nq_work->cq; 159 struct bnxt_qplib_nq *nq = nq_work->nq; 160 161 if (cq && nq) { 162 spin_lock_bh(&cq->compl_lock); 163 if (atomic_read(&cq->arm_state) && nq->cqn_handler) { 164 dev_dbg(&nq->pdev->dev, 165 "%s:Trigger cq = %p event nq = %p\n", 166 __func__, cq, nq); 167 nq->cqn_handler(nq, cq); 168 } 169 spin_unlock_bh(&cq->compl_lock); 170 } 171 kfree(nq_work); 172 } 173 174 static void bnxt_qplib_free_qp_hdr_buf(struct bnxt_qplib_res *res, 175 struct bnxt_qplib_qp *qp) 176 { 177 struct bnxt_qplib_q *rq = &qp->rq; 178 struct bnxt_qplib_q *sq = &qp->sq; 179 180 if (qp->rq_hdr_buf) 181 dma_free_coherent(&res->pdev->dev, 182 rq->max_wqe * qp->rq_hdr_buf_size, 183 qp->rq_hdr_buf, qp->rq_hdr_buf_map); 184 if (qp->sq_hdr_buf) 185 dma_free_coherent(&res->pdev->dev, 186 sq->max_wqe * qp->sq_hdr_buf_size, 187 qp->sq_hdr_buf, qp->sq_hdr_buf_map); 188 qp->rq_hdr_buf = NULL; 189 qp->sq_hdr_buf = NULL; 190 qp->rq_hdr_buf_map = 0; 191 qp->sq_hdr_buf_map = 0; 192 qp->sq_hdr_buf_size = 0; 193 qp->rq_hdr_buf_size = 0; 194 } 195 196 static int bnxt_qplib_alloc_qp_hdr_buf(struct bnxt_qplib_res *res, 197 struct bnxt_qplib_qp *qp) 198 { 199 struct bnxt_qplib_q *rq = &qp->rq; 200 struct bnxt_qplib_q *sq = &qp->sq; 201 int rc = 0; 202 203 if (qp->sq_hdr_buf_size && sq->max_wqe) { 204 qp->sq_hdr_buf = dma_alloc_coherent(&res->pdev->dev, 205 sq->max_wqe * qp->sq_hdr_buf_size, 206 &qp->sq_hdr_buf_map, GFP_KERNEL); 207 if (!qp->sq_hdr_buf) { 208 rc = -ENOMEM; 209 dev_err(&res->pdev->dev, 210 "Failed to create sq_hdr_buf\n"); 211 goto fail; 212 } 213 } 214 215 if (qp->rq_hdr_buf_size && rq->max_wqe) { 216 qp->rq_hdr_buf = dma_alloc_coherent(&res->pdev->dev, 217 rq->max_wqe * 218 qp->rq_hdr_buf_size, 219 &qp->rq_hdr_buf_map, 220 GFP_KERNEL); 221 if (!qp->rq_hdr_buf) { 222 rc = -ENOMEM; 223 dev_err(&res->pdev->dev, 224 "Failed to create rq_hdr_buf\n"); 225 goto fail; 226 } 227 } 228 return 0; 229 230 fail: 231 bnxt_qplib_free_qp_hdr_buf(res, qp); 232 return rc; 233 } 234 235 static void clean_nq(struct bnxt_qplib_nq *nq, struct bnxt_qplib_cq *cq) 236 { 237 struct bnxt_qplib_hwq *hwq = &nq->hwq; 238 struct nq_base *nqe, **nq_ptr; 239 int budget = nq->budget; 240 u32 sw_cons, raw_cons; 241 uintptr_t q_handle; 242 u16 type; 243 244 spin_lock_bh(&hwq->lock); 245 /* Service the NQ until empty */ 246 raw_cons = hwq->cons; 247 while (budget--) { 248 sw_cons = HWQ_CMP(raw_cons, hwq); 249 nq_ptr = (struct nq_base **)hwq->pbl_ptr; 250 nqe = &nq_ptr[NQE_PG(sw_cons)][NQE_IDX(sw_cons)]; 251 if (!NQE_CMP_VALID(nqe, raw_cons, hwq->max_elements)) 252 break; 253 254 /* 255 * The valid test of the entry must be done first before 256 * reading any further. 257 */ 258 dma_rmb(); 259 260 type = le16_to_cpu(nqe->info10_type) & NQ_BASE_TYPE_MASK; 261 switch (type) { 262 case NQ_BASE_TYPE_CQ_NOTIFICATION: 263 { 264 struct nq_cn *nqcne = (struct nq_cn *)nqe; 265 266 q_handle = le32_to_cpu(nqcne->cq_handle_low); 267 q_handle |= (u64)le32_to_cpu(nqcne->cq_handle_high) 268 << 32; 269 if ((unsigned long)cq == q_handle) { 270 nqcne->cq_handle_low = 0; 271 nqcne->cq_handle_high = 0; 272 cq->cnq_events++; 273 } 274 break; 275 } 276 default: 277 break; 278 } 279 raw_cons++; 280 } 281 spin_unlock_bh(&hwq->lock); 282 } 283 284 /* Wait for receiving all NQEs for this CQ and clean the NQEs associated with 285 * this CQ. 286 */ 287 static void __wait_for_all_nqes(struct bnxt_qplib_cq *cq, u16 cnq_events) 288 { 289 u32 retry_cnt = 100; 290 291 while (retry_cnt--) { 292 if (cnq_events == cq->cnq_events) 293 return; 294 usleep_range(50, 100); 295 clean_nq(cq->nq, cq); 296 } 297 } 298 299 static void bnxt_qplib_service_nq(struct tasklet_struct *t) 300 { 301 struct bnxt_qplib_nq *nq = from_tasklet(nq, t, nq_tasklet); 302 struct bnxt_qplib_hwq *hwq = &nq->hwq; 303 struct bnxt_qplib_cq *cq; 304 int budget = nq->budget; 305 u32 sw_cons, raw_cons; 306 struct nq_base *nqe; 307 uintptr_t q_handle; 308 u16 type; 309 310 spin_lock_bh(&hwq->lock); 311 /* Service the NQ until empty */ 312 raw_cons = hwq->cons; 313 while (budget--) { 314 sw_cons = HWQ_CMP(raw_cons, hwq); 315 nqe = bnxt_qplib_get_qe(hwq, sw_cons, NULL); 316 if (!NQE_CMP_VALID(nqe, raw_cons, hwq->max_elements)) 317 break; 318 319 /* 320 * The valid test of the entry must be done first before 321 * reading any further. 322 */ 323 dma_rmb(); 324 325 type = le16_to_cpu(nqe->info10_type) & NQ_BASE_TYPE_MASK; 326 switch (type) { 327 case NQ_BASE_TYPE_CQ_NOTIFICATION: 328 { 329 struct nq_cn *nqcne = (struct nq_cn *)nqe; 330 331 q_handle = le32_to_cpu(nqcne->cq_handle_low); 332 q_handle |= (u64)le32_to_cpu(nqcne->cq_handle_high) 333 << 32; 334 cq = (struct bnxt_qplib_cq *)(unsigned long)q_handle; 335 if (!cq) 336 break; 337 bnxt_qplib_armen_db(&cq->dbinfo, 338 DBC_DBC_TYPE_CQ_ARMENA); 339 spin_lock_bh(&cq->compl_lock); 340 atomic_set(&cq->arm_state, 0); 341 if (nq->cqn_handler(nq, (cq))) 342 dev_warn(&nq->pdev->dev, 343 "cqn - type 0x%x not handled\n", type); 344 cq->cnq_events++; 345 spin_unlock_bh(&cq->compl_lock); 346 break; 347 } 348 case NQ_BASE_TYPE_SRQ_EVENT: 349 { 350 struct bnxt_qplib_srq *srq; 351 struct nq_srq_event *nqsrqe = 352 (struct nq_srq_event *)nqe; 353 354 q_handle = le32_to_cpu(nqsrqe->srq_handle_low); 355 q_handle |= (u64)le32_to_cpu(nqsrqe->srq_handle_high) 356 << 32; 357 srq = (struct bnxt_qplib_srq *)q_handle; 358 bnxt_qplib_armen_db(&srq->dbinfo, 359 DBC_DBC_TYPE_SRQ_ARMENA); 360 if (nq->srqn_handler(nq, 361 (struct bnxt_qplib_srq *)q_handle, 362 nqsrqe->event)) 363 dev_warn(&nq->pdev->dev, 364 "SRQ event 0x%x not handled\n", 365 nqsrqe->event); 366 break; 367 } 368 case NQ_BASE_TYPE_DBQ_EVENT: 369 break; 370 default: 371 dev_warn(&nq->pdev->dev, 372 "nqe with type = 0x%x not handled\n", type); 373 break; 374 } 375 raw_cons++; 376 } 377 if (hwq->cons != raw_cons) { 378 hwq->cons = raw_cons; 379 bnxt_qplib_ring_nq_db(&nq->nq_db.dbinfo, nq->res->cctx, true); 380 } 381 spin_unlock_bh(&hwq->lock); 382 } 383 384 /* bnxt_re_synchronize_nq - self polling notification queue. 385 * @nq - notification queue pointer 386 * 387 * This function will start polling entries of a given notification queue 388 * for all pending entries. 389 * This function is useful to synchronize notification entries while resources 390 * are going away. 391 */ 392 393 void bnxt_re_synchronize_nq(struct bnxt_qplib_nq *nq) 394 { 395 int budget = nq->budget; 396 397 nq->budget = nq->hwq.max_elements; 398 bnxt_qplib_service_nq(&nq->nq_tasklet); 399 nq->budget = budget; 400 } 401 402 static irqreturn_t bnxt_qplib_nq_irq(int irq, void *dev_instance) 403 { 404 struct bnxt_qplib_nq *nq = dev_instance; 405 struct bnxt_qplib_hwq *hwq = &nq->hwq; 406 u32 sw_cons; 407 408 /* Prefetch the NQ element */ 409 sw_cons = HWQ_CMP(hwq->cons, hwq); 410 prefetch(bnxt_qplib_get_qe(hwq, sw_cons, NULL)); 411 412 /* Fan out to CPU affinitized kthreads? */ 413 tasklet_schedule(&nq->nq_tasklet); 414 415 return IRQ_HANDLED; 416 } 417 418 void bnxt_qplib_nq_stop_irq(struct bnxt_qplib_nq *nq, bool kill) 419 { 420 if (!nq->requested) 421 return; 422 423 nq->requested = false; 424 /* Mask h/w interrupt */ 425 bnxt_qplib_ring_nq_db(&nq->nq_db.dbinfo, nq->res->cctx, false); 426 /* Sync with last running IRQ handler */ 427 synchronize_irq(nq->msix_vec); 428 irq_set_affinity_hint(nq->msix_vec, NULL); 429 free_irq(nq->msix_vec, nq); 430 kfree(nq->name); 431 nq->name = NULL; 432 433 if (kill) 434 tasklet_kill(&nq->nq_tasklet); 435 tasklet_disable(&nq->nq_tasklet); 436 } 437 438 void bnxt_qplib_disable_nq(struct bnxt_qplib_nq *nq) 439 { 440 if (nq->cqn_wq) { 441 destroy_workqueue(nq->cqn_wq); 442 nq->cqn_wq = NULL; 443 } 444 445 /* Make sure the HW is stopped! */ 446 bnxt_qplib_nq_stop_irq(nq, true); 447 448 if (nq->nq_db.reg.bar_reg) { 449 iounmap(nq->nq_db.reg.bar_reg); 450 nq->nq_db.reg.bar_reg = NULL; 451 } 452 453 nq->cqn_handler = NULL; 454 nq->srqn_handler = NULL; 455 nq->msix_vec = 0; 456 } 457 458 int bnxt_qplib_nq_start_irq(struct bnxt_qplib_nq *nq, int nq_indx, 459 int msix_vector, bool need_init) 460 { 461 struct bnxt_qplib_res *res = nq->res; 462 int rc; 463 464 if (nq->requested) 465 return -EFAULT; 466 467 nq->msix_vec = msix_vector; 468 if (need_init) 469 tasklet_setup(&nq->nq_tasklet, bnxt_qplib_service_nq); 470 else 471 tasklet_enable(&nq->nq_tasklet); 472 473 nq->name = kasprintf(GFP_KERNEL, "bnxt_re-nq-%d@pci:%s", 474 nq_indx, pci_name(res->pdev)); 475 if (!nq->name) 476 return -ENOMEM; 477 rc = request_irq(nq->msix_vec, bnxt_qplib_nq_irq, 0, nq->name, nq); 478 if (rc) { 479 kfree(nq->name); 480 nq->name = NULL; 481 tasklet_disable(&nq->nq_tasklet); 482 return rc; 483 } 484 485 cpumask_clear(&nq->mask); 486 cpumask_set_cpu(nq_indx, &nq->mask); 487 rc = irq_set_affinity_hint(nq->msix_vec, &nq->mask); 488 if (rc) { 489 dev_warn(&nq->pdev->dev, 490 "set affinity failed; vector: %d nq_idx: %d\n", 491 nq->msix_vec, nq_indx); 492 } 493 nq->requested = true; 494 bnxt_qplib_ring_nq_db(&nq->nq_db.dbinfo, res->cctx, true); 495 496 return rc; 497 } 498 499 static int bnxt_qplib_map_nq_db(struct bnxt_qplib_nq *nq, u32 reg_offt) 500 { 501 resource_size_t reg_base; 502 struct bnxt_qplib_nq_db *nq_db; 503 struct pci_dev *pdev; 504 505 pdev = nq->pdev; 506 nq_db = &nq->nq_db; 507 508 nq_db->reg.bar_id = NQ_CONS_PCI_BAR_REGION; 509 nq_db->reg.bar_base = pci_resource_start(pdev, nq_db->reg.bar_id); 510 if (!nq_db->reg.bar_base) { 511 dev_err(&pdev->dev, "QPLIB: NQ BAR region %d resc start is 0!", 512 nq_db->reg.bar_id); 513 return -ENOMEM; 514 } 515 516 reg_base = nq_db->reg.bar_base + reg_offt; 517 /* Unconditionally map 8 bytes to support 57500 series */ 518 nq_db->reg.len = 8; 519 nq_db->reg.bar_reg = ioremap(reg_base, nq_db->reg.len); 520 if (!nq_db->reg.bar_reg) { 521 dev_err(&pdev->dev, "QPLIB: NQ BAR region %d mapping failed", 522 nq_db->reg.bar_id); 523 return -ENOMEM; 524 } 525 526 nq_db->dbinfo.db = nq_db->reg.bar_reg; 527 nq_db->dbinfo.hwq = &nq->hwq; 528 nq_db->dbinfo.xid = nq->ring_id; 529 530 return 0; 531 } 532 533 int bnxt_qplib_enable_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq, 534 int nq_idx, int msix_vector, int bar_reg_offset, 535 cqn_handler_t cqn_handler, 536 srqn_handler_t srqn_handler) 537 { 538 int rc; 539 540 nq->pdev = pdev; 541 nq->cqn_handler = cqn_handler; 542 nq->srqn_handler = srqn_handler; 543 544 /* Have a task to schedule CQ notifiers in post send case */ 545 nq->cqn_wq = create_singlethread_workqueue("bnxt_qplib_nq"); 546 if (!nq->cqn_wq) 547 return -ENOMEM; 548 549 rc = bnxt_qplib_map_nq_db(nq, bar_reg_offset); 550 if (rc) 551 goto fail; 552 553 rc = bnxt_qplib_nq_start_irq(nq, nq_idx, msix_vector, true); 554 if (rc) { 555 dev_err(&nq->pdev->dev, 556 "Failed to request irq for nq-idx %d\n", nq_idx); 557 goto fail; 558 } 559 560 return 0; 561 fail: 562 bnxt_qplib_disable_nq(nq); 563 return rc; 564 } 565 566 void bnxt_qplib_free_nq(struct bnxt_qplib_nq *nq) 567 { 568 if (nq->hwq.max_elements) { 569 bnxt_qplib_free_hwq(nq->res, &nq->hwq); 570 nq->hwq.max_elements = 0; 571 } 572 } 573 574 int bnxt_qplib_alloc_nq(struct bnxt_qplib_res *res, struct bnxt_qplib_nq *nq) 575 { 576 struct bnxt_qplib_hwq_attr hwq_attr = {}; 577 struct bnxt_qplib_sg_info sginfo = {}; 578 579 nq->pdev = res->pdev; 580 nq->res = res; 581 if (!nq->hwq.max_elements || 582 nq->hwq.max_elements > BNXT_QPLIB_NQE_MAX_CNT) 583 nq->hwq.max_elements = BNXT_QPLIB_NQE_MAX_CNT; 584 585 sginfo.pgsize = PAGE_SIZE; 586 sginfo.pgshft = PAGE_SHIFT; 587 hwq_attr.res = res; 588 hwq_attr.sginfo = &sginfo; 589 hwq_attr.depth = nq->hwq.max_elements; 590 hwq_attr.stride = sizeof(struct nq_base); 591 hwq_attr.type = bnxt_qplib_get_hwq_type(nq->res); 592 if (bnxt_qplib_alloc_init_hwq(&nq->hwq, &hwq_attr)) { 593 dev_err(&nq->pdev->dev, "FP NQ allocation failed"); 594 return -ENOMEM; 595 } 596 nq->budget = 8; 597 return 0; 598 } 599 600 /* SRQ */ 601 void bnxt_qplib_destroy_srq(struct bnxt_qplib_res *res, 602 struct bnxt_qplib_srq *srq) 603 { 604 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 605 struct creq_destroy_srq_resp resp = {}; 606 struct bnxt_qplib_cmdqmsg msg = {}; 607 struct cmdq_destroy_srq req = {}; 608 int rc; 609 610 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 611 CMDQ_BASE_OPCODE_DESTROY_SRQ, 612 sizeof(req)); 613 614 /* Configure the request */ 615 req.srq_cid = cpu_to_le32(srq->id); 616 617 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), sizeof(resp), 0); 618 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 619 kfree(srq->swq); 620 if (rc) 621 return; 622 bnxt_qplib_free_hwq(res, &srq->hwq); 623 } 624 625 int bnxt_qplib_create_srq(struct bnxt_qplib_res *res, 626 struct bnxt_qplib_srq *srq) 627 { 628 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 629 struct bnxt_qplib_hwq_attr hwq_attr = {}; 630 struct creq_create_srq_resp resp = {}; 631 struct bnxt_qplib_cmdqmsg msg = {}; 632 struct cmdq_create_srq req = {}; 633 struct bnxt_qplib_pbl *pbl; 634 u16 pg_sz_lvl; 635 int rc, idx; 636 637 hwq_attr.res = res; 638 hwq_attr.sginfo = &srq->sg_info; 639 hwq_attr.depth = srq->max_wqe; 640 hwq_attr.stride = srq->wqe_size; 641 hwq_attr.type = HWQ_TYPE_QUEUE; 642 rc = bnxt_qplib_alloc_init_hwq(&srq->hwq, &hwq_attr); 643 if (rc) 644 return rc; 645 646 srq->swq = kcalloc(srq->hwq.max_elements, sizeof(*srq->swq), 647 GFP_KERNEL); 648 if (!srq->swq) { 649 rc = -ENOMEM; 650 goto fail; 651 } 652 653 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 654 CMDQ_BASE_OPCODE_CREATE_SRQ, 655 sizeof(req)); 656 657 /* Configure the request */ 658 req.dpi = cpu_to_le32(srq->dpi->dpi); 659 req.srq_handle = cpu_to_le64((uintptr_t)srq); 660 661 req.srq_size = cpu_to_le16((u16)srq->hwq.max_elements); 662 pbl = &srq->hwq.pbl[PBL_LVL_0]; 663 pg_sz_lvl = ((u16)bnxt_qplib_base_pg_size(&srq->hwq) << 664 CMDQ_CREATE_SRQ_PG_SIZE_SFT); 665 pg_sz_lvl |= (srq->hwq.level & CMDQ_CREATE_SRQ_LVL_MASK) << 666 CMDQ_CREATE_SRQ_LVL_SFT; 667 req.pg_size_lvl = cpu_to_le16(pg_sz_lvl); 668 req.pbl = cpu_to_le64(pbl->pg_map_arr[0]); 669 req.pd_id = cpu_to_le32(srq->pd->id); 670 req.eventq_id = cpu_to_le16(srq->eventq_hw_ring_id); 671 672 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), sizeof(resp), 0); 673 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 674 if (rc) 675 goto fail; 676 677 spin_lock_init(&srq->lock); 678 srq->start_idx = 0; 679 srq->last_idx = srq->hwq.max_elements - 1; 680 for (idx = 0; idx < srq->hwq.max_elements; idx++) 681 srq->swq[idx].next_idx = idx + 1; 682 srq->swq[srq->last_idx].next_idx = -1; 683 684 srq->id = le32_to_cpu(resp.xid); 685 srq->dbinfo.hwq = &srq->hwq; 686 srq->dbinfo.xid = srq->id; 687 srq->dbinfo.db = srq->dpi->dbr; 688 srq->dbinfo.max_slot = 1; 689 srq->dbinfo.priv_db = res->dpi_tbl.priv_db; 690 if (srq->threshold) 691 bnxt_qplib_armen_db(&srq->dbinfo, DBC_DBC_TYPE_SRQ_ARMENA); 692 srq->arm_req = false; 693 694 return 0; 695 fail: 696 bnxt_qplib_free_hwq(res, &srq->hwq); 697 kfree(srq->swq); 698 699 return rc; 700 } 701 702 int bnxt_qplib_modify_srq(struct bnxt_qplib_res *res, 703 struct bnxt_qplib_srq *srq) 704 { 705 struct bnxt_qplib_hwq *srq_hwq = &srq->hwq; 706 u32 sw_prod, sw_cons, count = 0; 707 708 sw_prod = HWQ_CMP(srq_hwq->prod, srq_hwq); 709 sw_cons = HWQ_CMP(srq_hwq->cons, srq_hwq); 710 711 count = sw_prod > sw_cons ? sw_prod - sw_cons : 712 srq_hwq->max_elements - sw_cons + sw_prod; 713 if (count > srq->threshold) { 714 srq->arm_req = false; 715 bnxt_qplib_srq_arm_db(&srq->dbinfo, srq->threshold); 716 } else { 717 /* Deferred arming */ 718 srq->arm_req = true; 719 } 720 721 return 0; 722 } 723 724 int bnxt_qplib_query_srq(struct bnxt_qplib_res *res, 725 struct bnxt_qplib_srq *srq) 726 { 727 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 728 struct creq_query_srq_resp resp = {}; 729 struct bnxt_qplib_cmdqmsg msg = {}; 730 struct bnxt_qplib_rcfw_sbuf sbuf; 731 struct creq_query_srq_resp_sb *sb; 732 struct cmdq_query_srq req = {}; 733 int rc; 734 735 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 736 CMDQ_BASE_OPCODE_QUERY_SRQ, 737 sizeof(req)); 738 739 /* Configure the request */ 740 sbuf.size = ALIGN(sizeof(*sb), BNXT_QPLIB_CMDQE_UNITS); 741 sbuf.sb = dma_alloc_coherent(&rcfw->pdev->dev, sbuf.size, 742 &sbuf.dma_addr, GFP_KERNEL); 743 if (!sbuf.sb) 744 return -ENOMEM; 745 req.resp_size = sbuf.size / BNXT_QPLIB_CMDQE_UNITS; 746 req.srq_cid = cpu_to_le32(srq->id); 747 sb = sbuf.sb; 748 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, &sbuf, sizeof(req), 749 sizeof(resp), 0); 750 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 751 srq->threshold = le16_to_cpu(sb->srq_limit); 752 dma_free_coherent(&rcfw->pdev->dev, sbuf.size, 753 sbuf.sb, sbuf.dma_addr); 754 755 return rc; 756 } 757 758 int bnxt_qplib_post_srq_recv(struct bnxt_qplib_srq *srq, 759 struct bnxt_qplib_swqe *wqe) 760 { 761 struct bnxt_qplib_hwq *srq_hwq = &srq->hwq; 762 struct rq_wqe *srqe; 763 struct sq_sge *hw_sge; 764 u32 sw_prod, sw_cons, count = 0; 765 int i, next; 766 767 spin_lock(&srq_hwq->lock); 768 if (srq->start_idx == srq->last_idx) { 769 dev_err(&srq_hwq->pdev->dev, 770 "FP: SRQ (0x%x) is full!\n", srq->id); 771 spin_unlock(&srq_hwq->lock); 772 return -EINVAL; 773 } 774 next = srq->start_idx; 775 srq->start_idx = srq->swq[next].next_idx; 776 spin_unlock(&srq_hwq->lock); 777 778 sw_prod = HWQ_CMP(srq_hwq->prod, srq_hwq); 779 srqe = bnxt_qplib_get_qe(srq_hwq, sw_prod, NULL); 780 memset(srqe, 0, srq->wqe_size); 781 /* Calculate wqe_size16 and data_len */ 782 for (i = 0, hw_sge = (struct sq_sge *)srqe->data; 783 i < wqe->num_sge; i++, hw_sge++) { 784 hw_sge->va_or_pa = cpu_to_le64(wqe->sg_list[i].addr); 785 hw_sge->l_key = cpu_to_le32(wqe->sg_list[i].lkey); 786 hw_sge->size = cpu_to_le32(wqe->sg_list[i].size); 787 } 788 srqe->wqe_type = wqe->type; 789 srqe->flags = wqe->flags; 790 srqe->wqe_size = wqe->num_sge + 791 ((offsetof(typeof(*srqe), data) + 15) >> 4); 792 srqe->wr_id[0] = cpu_to_le32((u32)next); 793 srq->swq[next].wr_id = wqe->wr_id; 794 795 srq_hwq->prod++; 796 797 spin_lock(&srq_hwq->lock); 798 sw_prod = HWQ_CMP(srq_hwq->prod, srq_hwq); 799 /* retaining srq_hwq->cons for this logic 800 * actually the lock is only required to 801 * read srq_hwq->cons. 802 */ 803 sw_cons = HWQ_CMP(srq_hwq->cons, srq_hwq); 804 count = sw_prod > sw_cons ? sw_prod - sw_cons : 805 srq_hwq->max_elements - sw_cons + sw_prod; 806 spin_unlock(&srq_hwq->lock); 807 /* Ring DB */ 808 bnxt_qplib_ring_prod_db(&srq->dbinfo, DBC_DBC_TYPE_SRQ); 809 if (srq->arm_req == true && count > srq->threshold) { 810 srq->arm_req = false; 811 bnxt_qplib_srq_arm_db(&srq->dbinfo, srq->threshold); 812 } 813 814 return 0; 815 } 816 817 /* QP */ 818 819 static int bnxt_qplib_alloc_init_swq(struct bnxt_qplib_q *que) 820 { 821 int indx; 822 823 que->swq = kcalloc(que->max_wqe, sizeof(*que->swq), GFP_KERNEL); 824 if (!que->swq) 825 return -ENOMEM; 826 827 que->swq_start = 0; 828 que->swq_last = que->max_wqe - 1; 829 for (indx = 0; indx < que->max_wqe; indx++) 830 que->swq[indx].next_idx = indx + 1; 831 que->swq[que->swq_last].next_idx = 0; /* Make it circular */ 832 que->swq_last = 0; 833 834 return 0; 835 } 836 837 int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp) 838 { 839 struct bnxt_qplib_hwq_attr hwq_attr = {}; 840 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 841 struct creq_create_qp1_resp resp = {}; 842 struct bnxt_qplib_cmdqmsg msg = {}; 843 struct bnxt_qplib_q *sq = &qp->sq; 844 struct bnxt_qplib_q *rq = &qp->rq; 845 struct cmdq_create_qp1 req = {}; 846 struct bnxt_qplib_pbl *pbl; 847 u32 qp_flags = 0; 848 u8 pg_sz_lvl; 849 u32 tbl_indx; 850 int rc; 851 852 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 853 CMDQ_BASE_OPCODE_CREATE_QP1, 854 sizeof(req)); 855 /* General */ 856 req.type = qp->type; 857 req.dpi = cpu_to_le32(qp->dpi->dpi); 858 req.qp_handle = cpu_to_le64(qp->qp_handle); 859 860 /* SQ */ 861 hwq_attr.res = res; 862 hwq_attr.sginfo = &sq->sg_info; 863 hwq_attr.stride = sizeof(struct sq_sge); 864 hwq_attr.depth = bnxt_qplib_get_depth(sq); 865 hwq_attr.type = HWQ_TYPE_QUEUE; 866 rc = bnxt_qplib_alloc_init_hwq(&sq->hwq, &hwq_attr); 867 if (rc) 868 return rc; 869 870 rc = bnxt_qplib_alloc_init_swq(sq); 871 if (rc) 872 goto fail_sq; 873 874 req.sq_size = cpu_to_le32(bnxt_qplib_set_sq_size(sq, qp->wqe_mode)); 875 pbl = &sq->hwq.pbl[PBL_LVL_0]; 876 req.sq_pbl = cpu_to_le64(pbl->pg_map_arr[0]); 877 pg_sz_lvl = (bnxt_qplib_base_pg_size(&sq->hwq) << 878 CMDQ_CREATE_QP1_SQ_PG_SIZE_SFT); 879 pg_sz_lvl |= (sq->hwq.level & CMDQ_CREATE_QP1_SQ_LVL_MASK); 880 req.sq_pg_size_sq_lvl = pg_sz_lvl; 881 req.sq_fwo_sq_sge = 882 cpu_to_le16((sq->max_sge & CMDQ_CREATE_QP1_SQ_SGE_MASK) << 883 CMDQ_CREATE_QP1_SQ_SGE_SFT); 884 req.scq_cid = cpu_to_le32(qp->scq->id); 885 886 /* RQ */ 887 if (rq->max_wqe) { 888 hwq_attr.res = res; 889 hwq_attr.sginfo = &rq->sg_info; 890 hwq_attr.stride = sizeof(struct sq_sge); 891 hwq_attr.depth = bnxt_qplib_get_depth(rq); 892 hwq_attr.type = HWQ_TYPE_QUEUE; 893 rc = bnxt_qplib_alloc_init_hwq(&rq->hwq, &hwq_attr); 894 if (rc) 895 goto sq_swq; 896 rc = bnxt_qplib_alloc_init_swq(rq); 897 if (rc) 898 goto fail_rq; 899 req.rq_size = cpu_to_le32(rq->max_wqe); 900 pbl = &rq->hwq.pbl[PBL_LVL_0]; 901 req.rq_pbl = cpu_to_le64(pbl->pg_map_arr[0]); 902 pg_sz_lvl = (bnxt_qplib_base_pg_size(&rq->hwq) << 903 CMDQ_CREATE_QP1_RQ_PG_SIZE_SFT); 904 pg_sz_lvl |= (rq->hwq.level & CMDQ_CREATE_QP1_RQ_LVL_MASK); 905 req.rq_pg_size_rq_lvl = pg_sz_lvl; 906 req.rq_fwo_rq_sge = 907 cpu_to_le16((rq->max_sge & 908 CMDQ_CREATE_QP1_RQ_SGE_MASK) << 909 CMDQ_CREATE_QP1_RQ_SGE_SFT); 910 } 911 req.rcq_cid = cpu_to_le32(qp->rcq->id); 912 /* Header buffer - allow hdr_buf pass in */ 913 rc = bnxt_qplib_alloc_qp_hdr_buf(res, qp); 914 if (rc) { 915 rc = -ENOMEM; 916 goto rq_rwq; 917 } 918 qp_flags |= CMDQ_CREATE_QP1_QP_FLAGS_RESERVED_LKEY_ENABLE; 919 req.qp_flags = cpu_to_le32(qp_flags); 920 req.pd_id = cpu_to_le32(qp->pd->id); 921 922 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), sizeof(resp), 0); 923 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 924 if (rc) 925 goto fail; 926 927 qp->id = le32_to_cpu(resp.xid); 928 qp->cur_qp_state = CMDQ_MODIFY_QP_NEW_STATE_RESET; 929 qp->cctx = res->cctx; 930 sq->dbinfo.hwq = &sq->hwq; 931 sq->dbinfo.xid = qp->id; 932 sq->dbinfo.db = qp->dpi->dbr; 933 sq->dbinfo.max_slot = bnxt_qplib_set_sq_max_slot(qp->wqe_mode); 934 if (rq->max_wqe) { 935 rq->dbinfo.hwq = &rq->hwq; 936 rq->dbinfo.xid = qp->id; 937 rq->dbinfo.db = qp->dpi->dbr; 938 rq->dbinfo.max_slot = bnxt_qplib_set_rq_max_slot(rq->wqe_size); 939 } 940 tbl_indx = map_qp_id_to_tbl_indx(qp->id, rcfw); 941 rcfw->qp_tbl[tbl_indx].qp_id = qp->id; 942 rcfw->qp_tbl[tbl_indx].qp_handle = (void *)qp; 943 944 return 0; 945 946 fail: 947 bnxt_qplib_free_qp_hdr_buf(res, qp); 948 rq_rwq: 949 kfree(rq->swq); 950 fail_rq: 951 bnxt_qplib_free_hwq(res, &rq->hwq); 952 sq_swq: 953 kfree(sq->swq); 954 fail_sq: 955 bnxt_qplib_free_hwq(res, &sq->hwq); 956 return rc; 957 } 958 959 static void bnxt_qplib_init_psn_ptr(struct bnxt_qplib_qp *qp, int size) 960 { 961 struct bnxt_qplib_hwq *hwq; 962 struct bnxt_qplib_q *sq; 963 u64 fpsne, psn_pg; 964 u16 indx_pad = 0; 965 966 sq = &qp->sq; 967 hwq = &sq->hwq; 968 /* First psn entry */ 969 fpsne = (u64)bnxt_qplib_get_qe(hwq, hwq->depth, &psn_pg); 970 if (!IS_ALIGNED(fpsne, PAGE_SIZE)) 971 indx_pad = (fpsne & ~PAGE_MASK) / size; 972 hwq->pad_pgofft = indx_pad; 973 hwq->pad_pg = (u64 *)psn_pg; 974 hwq->pad_stride = size; 975 } 976 977 int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp) 978 { 979 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 980 struct bnxt_qplib_hwq_attr hwq_attr = {}; 981 struct bnxt_qplib_sg_info sginfo = {}; 982 struct creq_create_qp_resp resp = {}; 983 struct bnxt_qplib_cmdqmsg msg = {}; 984 struct bnxt_qplib_q *sq = &qp->sq; 985 struct bnxt_qplib_q *rq = &qp->rq; 986 struct cmdq_create_qp req = {}; 987 int rc, req_size, psn_sz = 0; 988 struct bnxt_qplib_hwq *xrrq; 989 struct bnxt_qplib_pbl *pbl; 990 u32 qp_flags = 0; 991 u8 pg_sz_lvl; 992 u32 tbl_indx; 993 u16 nsge; 994 995 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 996 CMDQ_BASE_OPCODE_CREATE_QP, 997 sizeof(req)); 998 999 /* General */ 1000 req.type = qp->type; 1001 req.dpi = cpu_to_le32(qp->dpi->dpi); 1002 req.qp_handle = cpu_to_le64(qp->qp_handle); 1003 1004 /* SQ */ 1005 if (qp->type == CMDQ_CREATE_QP_TYPE_RC) { 1006 psn_sz = bnxt_qplib_is_chip_gen_p5(res->cctx) ? 1007 sizeof(struct sq_psn_search_ext) : 1008 sizeof(struct sq_psn_search); 1009 } 1010 1011 hwq_attr.res = res; 1012 hwq_attr.sginfo = &sq->sg_info; 1013 hwq_attr.stride = sizeof(struct sq_sge); 1014 hwq_attr.depth = bnxt_qplib_get_depth(sq); 1015 hwq_attr.aux_stride = psn_sz; 1016 hwq_attr.aux_depth = bnxt_qplib_set_sq_size(sq, qp->wqe_mode); 1017 hwq_attr.type = HWQ_TYPE_QUEUE; 1018 rc = bnxt_qplib_alloc_init_hwq(&sq->hwq, &hwq_attr); 1019 if (rc) 1020 return rc; 1021 1022 rc = bnxt_qplib_alloc_init_swq(sq); 1023 if (rc) 1024 goto fail_sq; 1025 1026 if (psn_sz) 1027 bnxt_qplib_init_psn_ptr(qp, psn_sz); 1028 1029 req.sq_size = cpu_to_le32(bnxt_qplib_set_sq_size(sq, qp->wqe_mode)); 1030 pbl = &sq->hwq.pbl[PBL_LVL_0]; 1031 req.sq_pbl = cpu_to_le64(pbl->pg_map_arr[0]); 1032 pg_sz_lvl = (bnxt_qplib_base_pg_size(&sq->hwq) << 1033 CMDQ_CREATE_QP_SQ_PG_SIZE_SFT); 1034 pg_sz_lvl |= (sq->hwq.level & CMDQ_CREATE_QP_SQ_LVL_MASK); 1035 req.sq_pg_size_sq_lvl = pg_sz_lvl; 1036 req.sq_fwo_sq_sge = 1037 cpu_to_le16(((sq->max_sge & CMDQ_CREATE_QP_SQ_SGE_MASK) << 1038 CMDQ_CREATE_QP_SQ_SGE_SFT) | 0); 1039 req.scq_cid = cpu_to_le32(qp->scq->id); 1040 1041 /* RQ */ 1042 if (!qp->srq) { 1043 hwq_attr.res = res; 1044 hwq_attr.sginfo = &rq->sg_info; 1045 hwq_attr.stride = sizeof(struct sq_sge); 1046 hwq_attr.depth = bnxt_qplib_get_depth(rq); 1047 hwq_attr.aux_stride = 0; 1048 hwq_attr.aux_depth = 0; 1049 hwq_attr.type = HWQ_TYPE_QUEUE; 1050 rc = bnxt_qplib_alloc_init_hwq(&rq->hwq, &hwq_attr); 1051 if (rc) 1052 goto sq_swq; 1053 rc = bnxt_qplib_alloc_init_swq(rq); 1054 if (rc) 1055 goto fail_rq; 1056 1057 req.rq_size = cpu_to_le32(rq->max_wqe); 1058 pbl = &rq->hwq.pbl[PBL_LVL_0]; 1059 req.rq_pbl = cpu_to_le64(pbl->pg_map_arr[0]); 1060 pg_sz_lvl = (bnxt_qplib_base_pg_size(&rq->hwq) << 1061 CMDQ_CREATE_QP_RQ_PG_SIZE_SFT); 1062 pg_sz_lvl |= (rq->hwq.level & CMDQ_CREATE_QP_RQ_LVL_MASK); 1063 req.rq_pg_size_rq_lvl = pg_sz_lvl; 1064 nsge = (qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ? 1065 6 : rq->max_sge; 1066 req.rq_fwo_rq_sge = 1067 cpu_to_le16(((nsge & 1068 CMDQ_CREATE_QP_RQ_SGE_MASK) << 1069 CMDQ_CREATE_QP_RQ_SGE_SFT) | 0); 1070 } else { 1071 /* SRQ */ 1072 qp_flags |= CMDQ_CREATE_QP_QP_FLAGS_SRQ_USED; 1073 req.srq_cid = cpu_to_le32(qp->srq->id); 1074 } 1075 req.rcq_cid = cpu_to_le32(qp->rcq->id); 1076 1077 qp_flags |= CMDQ_CREATE_QP_QP_FLAGS_RESERVED_LKEY_ENABLE; 1078 qp_flags |= CMDQ_CREATE_QP_QP_FLAGS_FR_PMR_ENABLED; 1079 if (qp->sig_type) 1080 qp_flags |= CMDQ_CREATE_QP_QP_FLAGS_FORCE_COMPLETION; 1081 if (qp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) 1082 qp_flags |= CMDQ_CREATE_QP_QP_FLAGS_VARIABLE_SIZED_WQE_ENABLED; 1083 if (_is_ext_stats_supported(res->dattr->dev_cap_flags) && !res->is_vf) 1084 qp_flags |= CMDQ_CREATE_QP_QP_FLAGS_EXT_STATS_ENABLED; 1085 1086 req.qp_flags = cpu_to_le32(qp_flags); 1087 1088 /* ORRQ and IRRQ */ 1089 if (psn_sz) { 1090 xrrq = &qp->orrq; 1091 xrrq->max_elements = 1092 ORD_LIMIT_TO_ORRQ_SLOTS(qp->max_rd_atomic); 1093 req_size = xrrq->max_elements * 1094 BNXT_QPLIB_MAX_ORRQE_ENTRY_SIZE + PAGE_SIZE - 1; 1095 req_size &= ~(PAGE_SIZE - 1); 1096 sginfo.pgsize = req_size; 1097 sginfo.pgshft = PAGE_SHIFT; 1098 1099 hwq_attr.res = res; 1100 hwq_attr.sginfo = &sginfo; 1101 hwq_attr.depth = xrrq->max_elements; 1102 hwq_attr.stride = BNXT_QPLIB_MAX_ORRQE_ENTRY_SIZE; 1103 hwq_attr.aux_stride = 0; 1104 hwq_attr.aux_depth = 0; 1105 hwq_attr.type = HWQ_TYPE_CTX; 1106 rc = bnxt_qplib_alloc_init_hwq(xrrq, &hwq_attr); 1107 if (rc) 1108 goto rq_swq; 1109 pbl = &xrrq->pbl[PBL_LVL_0]; 1110 req.orrq_addr = cpu_to_le64(pbl->pg_map_arr[0]); 1111 1112 xrrq = &qp->irrq; 1113 xrrq->max_elements = IRD_LIMIT_TO_IRRQ_SLOTS( 1114 qp->max_dest_rd_atomic); 1115 req_size = xrrq->max_elements * 1116 BNXT_QPLIB_MAX_IRRQE_ENTRY_SIZE + PAGE_SIZE - 1; 1117 req_size &= ~(PAGE_SIZE - 1); 1118 sginfo.pgsize = req_size; 1119 hwq_attr.depth = xrrq->max_elements; 1120 hwq_attr.stride = BNXT_QPLIB_MAX_IRRQE_ENTRY_SIZE; 1121 rc = bnxt_qplib_alloc_init_hwq(xrrq, &hwq_attr); 1122 if (rc) 1123 goto fail_orrq; 1124 1125 pbl = &xrrq->pbl[PBL_LVL_0]; 1126 req.irrq_addr = cpu_to_le64(pbl->pg_map_arr[0]); 1127 } 1128 req.pd_id = cpu_to_le32(qp->pd->id); 1129 1130 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), 1131 sizeof(resp), 0); 1132 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 1133 if (rc) 1134 goto fail; 1135 1136 qp->id = le32_to_cpu(resp.xid); 1137 qp->cur_qp_state = CMDQ_MODIFY_QP_NEW_STATE_RESET; 1138 INIT_LIST_HEAD(&qp->sq_flush); 1139 INIT_LIST_HEAD(&qp->rq_flush); 1140 qp->cctx = res->cctx; 1141 sq->dbinfo.hwq = &sq->hwq; 1142 sq->dbinfo.xid = qp->id; 1143 sq->dbinfo.db = qp->dpi->dbr; 1144 sq->dbinfo.max_slot = bnxt_qplib_set_sq_max_slot(qp->wqe_mode); 1145 if (rq->max_wqe) { 1146 rq->dbinfo.hwq = &rq->hwq; 1147 rq->dbinfo.xid = qp->id; 1148 rq->dbinfo.db = qp->dpi->dbr; 1149 rq->dbinfo.max_slot = bnxt_qplib_set_rq_max_slot(rq->wqe_size); 1150 } 1151 tbl_indx = map_qp_id_to_tbl_indx(qp->id, rcfw); 1152 rcfw->qp_tbl[tbl_indx].qp_id = qp->id; 1153 rcfw->qp_tbl[tbl_indx].qp_handle = (void *)qp; 1154 1155 return 0; 1156 fail: 1157 bnxt_qplib_free_hwq(res, &qp->irrq); 1158 fail_orrq: 1159 bnxt_qplib_free_hwq(res, &qp->orrq); 1160 rq_swq: 1161 kfree(rq->swq); 1162 fail_rq: 1163 bnxt_qplib_free_hwq(res, &rq->hwq); 1164 sq_swq: 1165 kfree(sq->swq); 1166 fail_sq: 1167 bnxt_qplib_free_hwq(res, &sq->hwq); 1168 return rc; 1169 } 1170 1171 static void __modify_flags_from_init_state(struct bnxt_qplib_qp *qp) 1172 { 1173 switch (qp->state) { 1174 case CMDQ_MODIFY_QP_NEW_STATE_RTR: 1175 /* INIT->RTR, configure the path_mtu to the default 1176 * 2048 if not being requested 1177 */ 1178 if (!(qp->modify_flags & 1179 CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU)) { 1180 qp->modify_flags |= 1181 CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU; 1182 qp->path_mtu = 1183 CMDQ_MODIFY_QP_PATH_MTU_MTU_2048; 1184 } 1185 qp->modify_flags &= 1186 ~CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID; 1187 /* Bono FW require the max_dest_rd_atomic to be >= 1 */ 1188 if (qp->max_dest_rd_atomic < 1) 1189 qp->max_dest_rd_atomic = 1; 1190 qp->modify_flags &= ~CMDQ_MODIFY_QP_MODIFY_MASK_SRC_MAC; 1191 /* Bono FW 20.6.5 requires SGID_INDEX configuration */ 1192 if (!(qp->modify_flags & 1193 CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX)) { 1194 qp->modify_flags |= 1195 CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX; 1196 qp->ah.sgid_index = 0; 1197 } 1198 break; 1199 default: 1200 break; 1201 } 1202 } 1203 1204 static void __modify_flags_from_rtr_state(struct bnxt_qplib_qp *qp) 1205 { 1206 switch (qp->state) { 1207 case CMDQ_MODIFY_QP_NEW_STATE_RTS: 1208 /* Bono FW requires the max_rd_atomic to be >= 1 */ 1209 if (qp->max_rd_atomic < 1) 1210 qp->max_rd_atomic = 1; 1211 /* Bono FW does not allow PKEY_INDEX, 1212 * DGID, FLOW_LABEL, SGID_INDEX, HOP_LIMIT, 1213 * TRAFFIC_CLASS, DEST_MAC, PATH_MTU, RQ_PSN, 1214 * MIN_RNR_TIMER, MAX_DEST_RD_ATOMIC, DEST_QP_ID 1215 * modification 1216 */ 1217 qp->modify_flags &= 1218 ~(CMDQ_MODIFY_QP_MODIFY_MASK_PKEY | 1219 CMDQ_MODIFY_QP_MODIFY_MASK_DGID | 1220 CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL | 1221 CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX | 1222 CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT | 1223 CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS | 1224 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC | 1225 CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU | 1226 CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN | 1227 CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER | 1228 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC | 1229 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID); 1230 break; 1231 default: 1232 break; 1233 } 1234 } 1235 1236 static void __filter_modify_flags(struct bnxt_qplib_qp *qp) 1237 { 1238 switch (qp->cur_qp_state) { 1239 case CMDQ_MODIFY_QP_NEW_STATE_RESET: 1240 break; 1241 case CMDQ_MODIFY_QP_NEW_STATE_INIT: 1242 __modify_flags_from_init_state(qp); 1243 break; 1244 case CMDQ_MODIFY_QP_NEW_STATE_RTR: 1245 __modify_flags_from_rtr_state(qp); 1246 break; 1247 case CMDQ_MODIFY_QP_NEW_STATE_RTS: 1248 break; 1249 case CMDQ_MODIFY_QP_NEW_STATE_SQD: 1250 break; 1251 case CMDQ_MODIFY_QP_NEW_STATE_SQE: 1252 break; 1253 case CMDQ_MODIFY_QP_NEW_STATE_ERR: 1254 break; 1255 default: 1256 break; 1257 } 1258 } 1259 1260 int bnxt_qplib_modify_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp) 1261 { 1262 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 1263 struct creq_modify_qp_resp resp = {}; 1264 struct bnxt_qplib_cmdqmsg msg = {}; 1265 struct cmdq_modify_qp req = {}; 1266 u32 temp32[4]; 1267 u32 bmask; 1268 int rc; 1269 1270 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 1271 CMDQ_BASE_OPCODE_MODIFY_QP, 1272 sizeof(req)); 1273 1274 /* Filter out the qp_attr_mask based on the state->new transition */ 1275 __filter_modify_flags(qp); 1276 bmask = qp->modify_flags; 1277 req.modify_mask = cpu_to_le32(qp->modify_flags); 1278 req.qp_cid = cpu_to_le32(qp->id); 1279 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_STATE) { 1280 req.network_type_en_sqd_async_notify_new_state = 1281 (qp->state & CMDQ_MODIFY_QP_NEW_STATE_MASK) | 1282 (qp->en_sqd_async_notify ? 1283 CMDQ_MODIFY_QP_EN_SQD_ASYNC_NOTIFY : 0); 1284 } 1285 req.network_type_en_sqd_async_notify_new_state |= qp->nw_type; 1286 1287 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS) 1288 req.access = qp->access; 1289 1290 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_PKEY) 1291 req.pkey = cpu_to_le16(IB_DEFAULT_PKEY_FULL); 1292 1293 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_QKEY) 1294 req.qkey = cpu_to_le32(qp->qkey); 1295 1296 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_DGID) { 1297 memcpy(temp32, qp->ah.dgid.data, sizeof(struct bnxt_qplib_gid)); 1298 req.dgid[0] = cpu_to_le32(temp32[0]); 1299 req.dgid[1] = cpu_to_le32(temp32[1]); 1300 req.dgid[2] = cpu_to_le32(temp32[2]); 1301 req.dgid[3] = cpu_to_le32(temp32[3]); 1302 } 1303 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL) 1304 req.flow_label = cpu_to_le32(qp->ah.flow_label); 1305 1306 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX) 1307 req.sgid_index = cpu_to_le16(res->sgid_tbl.hw_id 1308 [qp->ah.sgid_index]); 1309 1310 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT) 1311 req.hop_limit = qp->ah.hop_limit; 1312 1313 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS) 1314 req.traffic_class = qp->ah.traffic_class; 1315 1316 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC) 1317 memcpy(req.dest_mac, qp->ah.dmac, 6); 1318 1319 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU) 1320 req.path_mtu_pingpong_push_enable |= qp->path_mtu; 1321 1322 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT) 1323 req.timeout = qp->timeout; 1324 1325 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT) 1326 req.retry_cnt = qp->retry_cnt; 1327 1328 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY) 1329 req.rnr_retry = qp->rnr_retry; 1330 1331 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER) 1332 req.min_rnr_timer = qp->min_rnr_timer; 1333 1334 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN) 1335 req.rq_psn = cpu_to_le32(qp->rq.psn); 1336 1337 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN) 1338 req.sq_psn = cpu_to_le32(qp->sq.psn); 1339 1340 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC) 1341 req.max_rd_atomic = 1342 ORD_LIMIT_TO_ORRQ_SLOTS(qp->max_rd_atomic); 1343 1344 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC) 1345 req.max_dest_rd_atomic = 1346 IRD_LIMIT_TO_IRRQ_SLOTS(qp->max_dest_rd_atomic); 1347 1348 req.sq_size = cpu_to_le32(qp->sq.hwq.max_elements); 1349 req.rq_size = cpu_to_le32(qp->rq.hwq.max_elements); 1350 req.sq_sge = cpu_to_le16(qp->sq.max_sge); 1351 req.rq_sge = cpu_to_le16(qp->rq.max_sge); 1352 req.max_inline_data = cpu_to_le32(qp->max_inline_data); 1353 if (bmask & CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID) 1354 req.dest_qp_id = cpu_to_le32(qp->dest_qpn); 1355 1356 req.vlan_pcp_vlan_dei_vlan_id = cpu_to_le16(qp->vlan_id); 1357 1358 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), sizeof(resp), 0); 1359 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 1360 if (rc) 1361 return rc; 1362 qp->cur_qp_state = qp->state; 1363 return 0; 1364 } 1365 1366 int bnxt_qplib_query_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp) 1367 { 1368 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 1369 struct creq_query_qp_resp resp = {}; 1370 struct bnxt_qplib_cmdqmsg msg = {}; 1371 struct bnxt_qplib_rcfw_sbuf sbuf; 1372 struct creq_query_qp_resp_sb *sb; 1373 struct cmdq_query_qp req = {}; 1374 u32 temp32[4]; 1375 int i, rc; 1376 1377 sbuf.size = ALIGN(sizeof(*sb), BNXT_QPLIB_CMDQE_UNITS); 1378 sbuf.sb = dma_alloc_coherent(&rcfw->pdev->dev, sbuf.size, 1379 &sbuf.dma_addr, GFP_KERNEL); 1380 if (!sbuf.sb) 1381 return -ENOMEM; 1382 sb = sbuf.sb; 1383 1384 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 1385 CMDQ_BASE_OPCODE_QUERY_QP, 1386 sizeof(req)); 1387 1388 req.qp_cid = cpu_to_le32(qp->id); 1389 req.resp_size = sbuf.size / BNXT_QPLIB_CMDQE_UNITS; 1390 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, &sbuf, sizeof(req), 1391 sizeof(resp), 0); 1392 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 1393 if (rc) 1394 goto bail; 1395 /* Extract the context from the side buffer */ 1396 qp->state = sb->en_sqd_async_notify_state & 1397 CREQ_QUERY_QP_RESP_SB_STATE_MASK; 1398 qp->en_sqd_async_notify = sb->en_sqd_async_notify_state & 1399 CREQ_QUERY_QP_RESP_SB_EN_SQD_ASYNC_NOTIFY; 1400 qp->access = sb->access; 1401 qp->pkey_index = le16_to_cpu(sb->pkey); 1402 qp->qkey = le32_to_cpu(sb->qkey); 1403 1404 temp32[0] = le32_to_cpu(sb->dgid[0]); 1405 temp32[1] = le32_to_cpu(sb->dgid[1]); 1406 temp32[2] = le32_to_cpu(sb->dgid[2]); 1407 temp32[3] = le32_to_cpu(sb->dgid[3]); 1408 memcpy(qp->ah.dgid.data, temp32, sizeof(qp->ah.dgid.data)); 1409 1410 qp->ah.flow_label = le32_to_cpu(sb->flow_label); 1411 1412 qp->ah.sgid_index = 0; 1413 for (i = 0; i < res->sgid_tbl.max; i++) { 1414 if (res->sgid_tbl.hw_id[i] == le16_to_cpu(sb->sgid_index)) { 1415 qp->ah.sgid_index = i; 1416 break; 1417 } 1418 } 1419 if (i == res->sgid_tbl.max) 1420 dev_warn(&res->pdev->dev, "SGID not found??\n"); 1421 1422 qp->ah.hop_limit = sb->hop_limit; 1423 qp->ah.traffic_class = sb->traffic_class; 1424 memcpy(qp->ah.dmac, sb->dest_mac, 6); 1425 qp->ah.vlan_id = (le16_to_cpu(sb->path_mtu_dest_vlan_id) & 1426 CREQ_QUERY_QP_RESP_SB_VLAN_ID_MASK) >> 1427 CREQ_QUERY_QP_RESP_SB_VLAN_ID_SFT; 1428 qp->path_mtu = (le16_to_cpu(sb->path_mtu_dest_vlan_id) & 1429 CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) >> 1430 CREQ_QUERY_QP_RESP_SB_PATH_MTU_SFT; 1431 qp->timeout = sb->timeout; 1432 qp->retry_cnt = sb->retry_cnt; 1433 qp->rnr_retry = sb->rnr_retry; 1434 qp->min_rnr_timer = sb->min_rnr_timer; 1435 qp->rq.psn = le32_to_cpu(sb->rq_psn); 1436 qp->max_rd_atomic = ORRQ_SLOTS_TO_ORD_LIMIT(sb->max_rd_atomic); 1437 qp->sq.psn = le32_to_cpu(sb->sq_psn); 1438 qp->max_dest_rd_atomic = 1439 IRRQ_SLOTS_TO_IRD_LIMIT(sb->max_dest_rd_atomic); 1440 qp->sq.max_wqe = qp->sq.hwq.max_elements; 1441 qp->rq.max_wqe = qp->rq.hwq.max_elements; 1442 qp->sq.max_sge = le16_to_cpu(sb->sq_sge); 1443 qp->rq.max_sge = le16_to_cpu(sb->rq_sge); 1444 qp->max_inline_data = le32_to_cpu(sb->max_inline_data); 1445 qp->dest_qpn = le32_to_cpu(sb->dest_qp_id); 1446 memcpy(qp->smac, sb->src_mac, 6); 1447 qp->vlan_id = le16_to_cpu(sb->vlan_pcp_vlan_dei_vlan_id); 1448 bail: 1449 dma_free_coherent(&rcfw->pdev->dev, sbuf.size, 1450 sbuf.sb, sbuf.dma_addr); 1451 return rc; 1452 } 1453 1454 static void __clean_cq(struct bnxt_qplib_cq *cq, u64 qp) 1455 { 1456 struct bnxt_qplib_hwq *cq_hwq = &cq->hwq; 1457 struct cq_base *hw_cqe; 1458 int i; 1459 1460 for (i = 0; i < cq_hwq->max_elements; i++) { 1461 hw_cqe = bnxt_qplib_get_qe(cq_hwq, i, NULL); 1462 if (!CQE_CMP_VALID(hw_cqe, i, cq_hwq->max_elements)) 1463 continue; 1464 /* 1465 * The valid test of the entry must be done first before 1466 * reading any further. 1467 */ 1468 dma_rmb(); 1469 switch (hw_cqe->cqe_type_toggle & CQ_BASE_CQE_TYPE_MASK) { 1470 case CQ_BASE_CQE_TYPE_REQ: 1471 case CQ_BASE_CQE_TYPE_TERMINAL: 1472 { 1473 struct cq_req *cqe = (struct cq_req *)hw_cqe; 1474 1475 if (qp == le64_to_cpu(cqe->qp_handle)) 1476 cqe->qp_handle = 0; 1477 break; 1478 } 1479 case CQ_BASE_CQE_TYPE_RES_RC: 1480 case CQ_BASE_CQE_TYPE_RES_UD: 1481 case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1: 1482 { 1483 struct cq_res_rc *cqe = (struct cq_res_rc *)hw_cqe; 1484 1485 if (qp == le64_to_cpu(cqe->qp_handle)) 1486 cqe->qp_handle = 0; 1487 break; 1488 } 1489 default: 1490 break; 1491 } 1492 } 1493 } 1494 1495 int bnxt_qplib_destroy_qp(struct bnxt_qplib_res *res, 1496 struct bnxt_qplib_qp *qp) 1497 { 1498 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 1499 struct creq_destroy_qp_resp resp = {}; 1500 struct bnxt_qplib_cmdqmsg msg = {}; 1501 struct cmdq_destroy_qp req = {}; 1502 u32 tbl_indx; 1503 int rc; 1504 1505 tbl_indx = map_qp_id_to_tbl_indx(qp->id, rcfw); 1506 rcfw->qp_tbl[tbl_indx].qp_id = BNXT_QPLIB_QP_ID_INVALID; 1507 rcfw->qp_tbl[tbl_indx].qp_handle = NULL; 1508 1509 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 1510 CMDQ_BASE_OPCODE_DESTROY_QP, 1511 sizeof(req)); 1512 1513 req.qp_cid = cpu_to_le32(qp->id); 1514 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), 1515 sizeof(resp), 0); 1516 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 1517 if (rc) { 1518 rcfw->qp_tbl[tbl_indx].qp_id = qp->id; 1519 rcfw->qp_tbl[tbl_indx].qp_handle = qp; 1520 return rc; 1521 } 1522 1523 return 0; 1524 } 1525 1526 void bnxt_qplib_free_qp_res(struct bnxt_qplib_res *res, 1527 struct bnxt_qplib_qp *qp) 1528 { 1529 bnxt_qplib_free_qp_hdr_buf(res, qp); 1530 bnxt_qplib_free_hwq(res, &qp->sq.hwq); 1531 kfree(qp->sq.swq); 1532 1533 bnxt_qplib_free_hwq(res, &qp->rq.hwq); 1534 kfree(qp->rq.swq); 1535 1536 if (qp->irrq.max_elements) 1537 bnxt_qplib_free_hwq(res, &qp->irrq); 1538 if (qp->orrq.max_elements) 1539 bnxt_qplib_free_hwq(res, &qp->orrq); 1540 1541 } 1542 1543 void *bnxt_qplib_get_qp1_sq_buf(struct bnxt_qplib_qp *qp, 1544 struct bnxt_qplib_sge *sge) 1545 { 1546 struct bnxt_qplib_q *sq = &qp->sq; 1547 u32 sw_prod; 1548 1549 memset(sge, 0, sizeof(*sge)); 1550 1551 if (qp->sq_hdr_buf) { 1552 sw_prod = sq->swq_start; 1553 sge->addr = (dma_addr_t)(qp->sq_hdr_buf_map + 1554 sw_prod * qp->sq_hdr_buf_size); 1555 sge->lkey = 0xFFFFFFFF; 1556 sge->size = qp->sq_hdr_buf_size; 1557 return qp->sq_hdr_buf + sw_prod * sge->size; 1558 } 1559 return NULL; 1560 } 1561 1562 u32 bnxt_qplib_get_rq_prod_index(struct bnxt_qplib_qp *qp) 1563 { 1564 struct bnxt_qplib_q *rq = &qp->rq; 1565 1566 return rq->swq_start; 1567 } 1568 1569 dma_addr_t bnxt_qplib_get_qp_buf_from_index(struct bnxt_qplib_qp *qp, u32 index) 1570 { 1571 return (qp->rq_hdr_buf_map + index * qp->rq_hdr_buf_size); 1572 } 1573 1574 void *bnxt_qplib_get_qp1_rq_buf(struct bnxt_qplib_qp *qp, 1575 struct bnxt_qplib_sge *sge) 1576 { 1577 struct bnxt_qplib_q *rq = &qp->rq; 1578 u32 sw_prod; 1579 1580 memset(sge, 0, sizeof(*sge)); 1581 1582 if (qp->rq_hdr_buf) { 1583 sw_prod = rq->swq_start; 1584 sge->addr = (dma_addr_t)(qp->rq_hdr_buf_map + 1585 sw_prod * qp->rq_hdr_buf_size); 1586 sge->lkey = 0xFFFFFFFF; 1587 sge->size = qp->rq_hdr_buf_size; 1588 return qp->rq_hdr_buf + sw_prod * sge->size; 1589 } 1590 return NULL; 1591 } 1592 1593 static void bnxt_qplib_fill_psn_search(struct bnxt_qplib_qp *qp, 1594 struct bnxt_qplib_swqe *wqe, 1595 struct bnxt_qplib_swq *swq) 1596 { 1597 struct sq_psn_search_ext *psns_ext; 1598 struct sq_psn_search *psns; 1599 u32 flg_npsn; 1600 u32 op_spsn; 1601 1602 if (!swq->psn_search) 1603 return; 1604 psns = swq->psn_search; 1605 psns_ext = swq->psn_ext; 1606 1607 op_spsn = ((swq->start_psn << SQ_PSN_SEARCH_START_PSN_SFT) & 1608 SQ_PSN_SEARCH_START_PSN_MASK); 1609 op_spsn |= ((wqe->type << SQ_PSN_SEARCH_OPCODE_SFT) & 1610 SQ_PSN_SEARCH_OPCODE_MASK); 1611 flg_npsn = ((swq->next_psn << SQ_PSN_SEARCH_NEXT_PSN_SFT) & 1612 SQ_PSN_SEARCH_NEXT_PSN_MASK); 1613 1614 if (bnxt_qplib_is_chip_gen_p5(qp->cctx)) { 1615 psns_ext->opcode_start_psn = cpu_to_le32(op_spsn); 1616 psns_ext->flags_next_psn = cpu_to_le32(flg_npsn); 1617 psns_ext->start_slot_idx = cpu_to_le16(swq->slot_idx); 1618 } else { 1619 psns->opcode_start_psn = cpu_to_le32(op_spsn); 1620 psns->flags_next_psn = cpu_to_le32(flg_npsn); 1621 } 1622 } 1623 1624 static int bnxt_qplib_put_inline(struct bnxt_qplib_qp *qp, 1625 struct bnxt_qplib_swqe *wqe, 1626 u16 *idx) 1627 { 1628 struct bnxt_qplib_hwq *hwq; 1629 int len, t_len, offt; 1630 bool pull_dst = true; 1631 void *il_dst = NULL; 1632 void *il_src = NULL; 1633 int t_cplen, cplen; 1634 int indx; 1635 1636 hwq = &qp->sq.hwq; 1637 t_len = 0; 1638 for (indx = 0; indx < wqe->num_sge; indx++) { 1639 len = wqe->sg_list[indx].size; 1640 il_src = (void *)wqe->sg_list[indx].addr; 1641 t_len += len; 1642 if (t_len > qp->max_inline_data) 1643 return -ENOMEM; 1644 while (len) { 1645 if (pull_dst) { 1646 pull_dst = false; 1647 il_dst = bnxt_qplib_get_prod_qe(hwq, *idx); 1648 (*idx)++; 1649 t_cplen = 0; 1650 offt = 0; 1651 } 1652 cplen = min_t(int, len, sizeof(struct sq_sge)); 1653 cplen = min_t(int, cplen, 1654 (sizeof(struct sq_sge) - offt)); 1655 memcpy(il_dst, il_src, cplen); 1656 t_cplen += cplen; 1657 il_src += cplen; 1658 il_dst += cplen; 1659 offt += cplen; 1660 len -= cplen; 1661 if (t_cplen == sizeof(struct sq_sge)) 1662 pull_dst = true; 1663 } 1664 } 1665 1666 return t_len; 1667 } 1668 1669 static u32 bnxt_qplib_put_sges(struct bnxt_qplib_hwq *hwq, 1670 struct bnxt_qplib_sge *ssge, 1671 u16 nsge, u16 *idx) 1672 { 1673 struct sq_sge *dsge; 1674 int indx, len = 0; 1675 1676 for (indx = 0; indx < nsge; indx++, (*idx)++) { 1677 dsge = bnxt_qplib_get_prod_qe(hwq, *idx); 1678 dsge->va_or_pa = cpu_to_le64(ssge[indx].addr); 1679 dsge->l_key = cpu_to_le32(ssge[indx].lkey); 1680 dsge->size = cpu_to_le32(ssge[indx].size); 1681 len += ssge[indx].size; 1682 } 1683 1684 return len; 1685 } 1686 1687 static u16 bnxt_qplib_required_slots(struct bnxt_qplib_qp *qp, 1688 struct bnxt_qplib_swqe *wqe, 1689 u16 *wqe_sz, u16 *qdf, u8 mode) 1690 { 1691 u32 ilsize, bytes; 1692 u16 nsge; 1693 u16 slot; 1694 1695 nsge = wqe->num_sge; 1696 /* Adding sq_send_hdr is a misnomer, for rq also hdr size is same. */ 1697 bytes = sizeof(struct sq_send_hdr) + nsge * sizeof(struct sq_sge); 1698 if (wqe->flags & BNXT_QPLIB_SWQE_FLAGS_INLINE) { 1699 ilsize = bnxt_qplib_calc_ilsize(wqe, qp->max_inline_data); 1700 bytes = ALIGN(ilsize, sizeof(struct sq_sge)); 1701 bytes += sizeof(struct sq_send_hdr); 1702 } 1703 1704 *qdf = __xlate_qfd(qp->sq.q_full_delta, bytes); 1705 slot = bytes >> 4; 1706 *wqe_sz = slot; 1707 if (mode == BNXT_QPLIB_WQE_MODE_STATIC) 1708 slot = 8; 1709 return slot; 1710 } 1711 1712 static void bnxt_qplib_pull_psn_buff(struct bnxt_qplib_q *sq, 1713 struct bnxt_qplib_swq *swq) 1714 { 1715 struct bnxt_qplib_hwq *hwq; 1716 u32 pg_num, pg_indx; 1717 void *buff; 1718 u32 tail; 1719 1720 hwq = &sq->hwq; 1721 if (!hwq->pad_pg) 1722 return; 1723 tail = swq->slot_idx / sq->dbinfo.max_slot; 1724 pg_num = (tail + hwq->pad_pgofft) / (PAGE_SIZE / hwq->pad_stride); 1725 pg_indx = (tail + hwq->pad_pgofft) % (PAGE_SIZE / hwq->pad_stride); 1726 buff = (void *)(hwq->pad_pg[pg_num] + pg_indx * hwq->pad_stride); 1727 swq->psn_ext = buff; 1728 swq->psn_search = buff; 1729 } 1730 1731 void bnxt_qplib_post_send_db(struct bnxt_qplib_qp *qp) 1732 { 1733 struct bnxt_qplib_q *sq = &qp->sq; 1734 1735 bnxt_qplib_ring_prod_db(&sq->dbinfo, DBC_DBC_TYPE_SQ); 1736 } 1737 1738 int bnxt_qplib_post_send(struct bnxt_qplib_qp *qp, 1739 struct bnxt_qplib_swqe *wqe) 1740 { 1741 struct bnxt_qplib_nq_work *nq_work = NULL; 1742 int i, rc = 0, data_len = 0, pkt_num = 0; 1743 struct bnxt_qplib_q *sq = &qp->sq; 1744 struct bnxt_qplib_hwq *hwq; 1745 struct bnxt_qplib_swq *swq; 1746 bool sch_handler = false; 1747 u16 wqe_sz, qdf = 0; 1748 void *base_hdr; 1749 void *ext_hdr; 1750 __le32 temp32; 1751 u32 wqe_idx; 1752 u32 slots; 1753 u16 idx; 1754 1755 hwq = &sq->hwq; 1756 if (qp->state != CMDQ_MODIFY_QP_NEW_STATE_RTS && 1757 qp->state != CMDQ_MODIFY_QP_NEW_STATE_ERR) { 1758 dev_err(&hwq->pdev->dev, 1759 "QPLIB: FP: QP (0x%x) is in the 0x%x state", 1760 qp->id, qp->state); 1761 rc = -EINVAL; 1762 goto done; 1763 } 1764 1765 slots = bnxt_qplib_required_slots(qp, wqe, &wqe_sz, &qdf, qp->wqe_mode); 1766 if (bnxt_qplib_queue_full(sq, slots + qdf)) { 1767 dev_err(&hwq->pdev->dev, 1768 "prod = %#x cons = %#x qdepth = %#x delta = %#x\n", 1769 hwq->prod, hwq->cons, hwq->depth, sq->q_full_delta); 1770 rc = -ENOMEM; 1771 goto done; 1772 } 1773 1774 swq = bnxt_qplib_get_swqe(sq, &wqe_idx); 1775 bnxt_qplib_pull_psn_buff(sq, swq); 1776 1777 idx = 0; 1778 swq->slot_idx = hwq->prod; 1779 swq->slots = slots; 1780 swq->wr_id = wqe->wr_id; 1781 swq->type = wqe->type; 1782 swq->flags = wqe->flags; 1783 swq->start_psn = sq->psn & BTH_PSN_MASK; 1784 if (qp->sig_type) 1785 swq->flags |= SQ_SEND_FLAGS_SIGNAL_COMP; 1786 1787 if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { 1788 sch_handler = true; 1789 dev_dbg(&hwq->pdev->dev, 1790 "%s Error QP. Scheduling for poll_cq\n", __func__); 1791 goto queue_err; 1792 } 1793 1794 base_hdr = bnxt_qplib_get_prod_qe(hwq, idx++); 1795 ext_hdr = bnxt_qplib_get_prod_qe(hwq, idx++); 1796 memset(base_hdr, 0, sizeof(struct sq_sge)); 1797 memset(ext_hdr, 0, sizeof(struct sq_sge)); 1798 1799 if (wqe->flags & BNXT_QPLIB_SWQE_FLAGS_INLINE) 1800 /* Copy the inline data */ 1801 data_len = bnxt_qplib_put_inline(qp, wqe, &idx); 1802 else 1803 data_len = bnxt_qplib_put_sges(hwq, wqe->sg_list, wqe->num_sge, 1804 &idx); 1805 if (data_len < 0) 1806 goto queue_err; 1807 /* Specifics */ 1808 switch (wqe->type) { 1809 case BNXT_QPLIB_SWQE_TYPE_SEND: 1810 if (qp->type == CMDQ_CREATE_QP1_TYPE_GSI) { 1811 struct sq_send_raweth_qp1_hdr *sqe = base_hdr; 1812 struct sq_raw_ext_hdr *ext_sqe = ext_hdr; 1813 /* Assemble info for Raw Ethertype QPs */ 1814 1815 sqe->wqe_type = wqe->type; 1816 sqe->flags = wqe->flags; 1817 sqe->wqe_size = wqe_sz; 1818 sqe->cfa_action = cpu_to_le16(wqe->rawqp1.cfa_action); 1819 sqe->lflags = cpu_to_le16(wqe->rawqp1.lflags); 1820 sqe->length = cpu_to_le32(data_len); 1821 ext_sqe->cfa_meta = cpu_to_le32((wqe->rawqp1.cfa_meta & 1822 SQ_SEND_RAWETH_QP1_CFA_META_VLAN_VID_MASK) << 1823 SQ_SEND_RAWETH_QP1_CFA_META_VLAN_VID_SFT); 1824 1825 break; 1826 } 1827 fallthrough; 1828 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM: 1829 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV: 1830 { 1831 struct sq_ud_ext_hdr *ext_sqe = ext_hdr; 1832 struct sq_send_hdr *sqe = base_hdr; 1833 1834 sqe->wqe_type = wqe->type; 1835 sqe->flags = wqe->flags; 1836 sqe->wqe_size = wqe_sz; 1837 sqe->inv_key_or_imm_data = cpu_to_le32(wqe->send.inv_key); 1838 if (qp->type == CMDQ_CREATE_QP_TYPE_UD || 1839 qp->type == CMDQ_CREATE_QP_TYPE_GSI) { 1840 sqe->q_key = cpu_to_le32(wqe->send.q_key); 1841 sqe->length = cpu_to_le32(data_len); 1842 sq->psn = (sq->psn + 1) & BTH_PSN_MASK; 1843 ext_sqe->dst_qp = cpu_to_le32(wqe->send.dst_qp & 1844 SQ_SEND_DST_QP_MASK); 1845 ext_sqe->avid = cpu_to_le32(wqe->send.avid & 1846 SQ_SEND_AVID_MASK); 1847 } else { 1848 sqe->length = cpu_to_le32(data_len); 1849 if (qp->mtu) 1850 pkt_num = (data_len + qp->mtu - 1) / qp->mtu; 1851 if (!pkt_num) 1852 pkt_num = 1; 1853 sq->psn = (sq->psn + pkt_num) & BTH_PSN_MASK; 1854 } 1855 break; 1856 } 1857 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE: 1858 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM: 1859 case BNXT_QPLIB_SWQE_TYPE_RDMA_READ: 1860 { 1861 struct sq_rdma_ext_hdr *ext_sqe = ext_hdr; 1862 struct sq_rdma_hdr *sqe = base_hdr; 1863 1864 sqe->wqe_type = wqe->type; 1865 sqe->flags = wqe->flags; 1866 sqe->wqe_size = wqe_sz; 1867 sqe->imm_data = cpu_to_le32(wqe->rdma.inv_key); 1868 sqe->length = cpu_to_le32((u32)data_len); 1869 ext_sqe->remote_va = cpu_to_le64(wqe->rdma.remote_va); 1870 ext_sqe->remote_key = cpu_to_le32(wqe->rdma.r_key); 1871 if (qp->mtu) 1872 pkt_num = (data_len + qp->mtu - 1) / qp->mtu; 1873 if (!pkt_num) 1874 pkt_num = 1; 1875 sq->psn = (sq->psn + pkt_num) & BTH_PSN_MASK; 1876 break; 1877 } 1878 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP: 1879 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD: 1880 { 1881 struct sq_atomic_ext_hdr *ext_sqe = ext_hdr; 1882 struct sq_atomic_hdr *sqe = base_hdr; 1883 1884 sqe->wqe_type = wqe->type; 1885 sqe->flags = wqe->flags; 1886 sqe->remote_key = cpu_to_le32(wqe->atomic.r_key); 1887 sqe->remote_va = cpu_to_le64(wqe->atomic.remote_va); 1888 ext_sqe->swap_data = cpu_to_le64(wqe->atomic.swap_data); 1889 ext_sqe->cmp_data = cpu_to_le64(wqe->atomic.cmp_data); 1890 if (qp->mtu) 1891 pkt_num = (data_len + qp->mtu - 1) / qp->mtu; 1892 if (!pkt_num) 1893 pkt_num = 1; 1894 sq->psn = (sq->psn + pkt_num) & BTH_PSN_MASK; 1895 break; 1896 } 1897 case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV: 1898 { 1899 struct sq_localinvalidate *sqe = base_hdr; 1900 1901 sqe->wqe_type = wqe->type; 1902 sqe->flags = wqe->flags; 1903 sqe->inv_l_key = cpu_to_le32(wqe->local_inv.inv_l_key); 1904 1905 break; 1906 } 1907 case BNXT_QPLIB_SWQE_TYPE_FAST_REG_MR: 1908 { 1909 struct sq_fr_pmr_ext_hdr *ext_sqe = ext_hdr; 1910 struct sq_fr_pmr_hdr *sqe = base_hdr; 1911 1912 sqe->wqe_type = wqe->type; 1913 sqe->flags = wqe->flags; 1914 sqe->access_cntl = wqe->frmr.access_cntl | 1915 SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE; 1916 sqe->zero_based_page_size_log = 1917 (wqe->frmr.pg_sz_log & SQ_FR_PMR_PAGE_SIZE_LOG_MASK) << 1918 SQ_FR_PMR_PAGE_SIZE_LOG_SFT | 1919 (wqe->frmr.zero_based ? SQ_FR_PMR_ZERO_BASED : 0); 1920 sqe->l_key = cpu_to_le32(wqe->frmr.l_key); 1921 temp32 = cpu_to_le32(wqe->frmr.length); 1922 memcpy(sqe->length, &temp32, sizeof(wqe->frmr.length)); 1923 sqe->numlevels_pbl_page_size_log = 1924 ((wqe->frmr.pbl_pg_sz_log << 1925 SQ_FR_PMR_PBL_PAGE_SIZE_LOG_SFT) & 1926 SQ_FR_PMR_PBL_PAGE_SIZE_LOG_MASK) | 1927 ((wqe->frmr.levels << SQ_FR_PMR_NUMLEVELS_SFT) & 1928 SQ_FR_PMR_NUMLEVELS_MASK); 1929 1930 for (i = 0; i < wqe->frmr.page_list_len; i++) 1931 wqe->frmr.pbl_ptr[i] = cpu_to_le64( 1932 wqe->frmr.page_list[i] | 1933 PTU_PTE_VALID); 1934 ext_sqe->pblptr = cpu_to_le64(wqe->frmr.pbl_dma_ptr); 1935 ext_sqe->va = cpu_to_le64(wqe->frmr.va); 1936 1937 break; 1938 } 1939 case BNXT_QPLIB_SWQE_TYPE_BIND_MW: 1940 { 1941 struct sq_bind_ext_hdr *ext_sqe = ext_hdr; 1942 struct sq_bind_hdr *sqe = base_hdr; 1943 1944 sqe->wqe_type = wqe->type; 1945 sqe->flags = wqe->flags; 1946 sqe->access_cntl = wqe->bind.access_cntl; 1947 sqe->mw_type_zero_based = wqe->bind.mw_type | 1948 (wqe->bind.zero_based ? SQ_BIND_ZERO_BASED : 0); 1949 sqe->parent_l_key = cpu_to_le32(wqe->bind.parent_l_key); 1950 sqe->l_key = cpu_to_le32(wqe->bind.r_key); 1951 ext_sqe->va = cpu_to_le64(wqe->bind.va); 1952 ext_sqe->length_lo = cpu_to_le32(wqe->bind.length); 1953 break; 1954 } 1955 default: 1956 /* Bad wqe, return error */ 1957 rc = -EINVAL; 1958 goto done; 1959 } 1960 swq->next_psn = sq->psn & BTH_PSN_MASK; 1961 bnxt_qplib_fill_psn_search(qp, wqe, swq); 1962 queue_err: 1963 bnxt_qplib_swq_mod_start(sq, wqe_idx); 1964 bnxt_qplib_hwq_incr_prod(hwq, swq->slots); 1965 qp->wqe_cnt++; 1966 done: 1967 if (sch_handler) { 1968 nq_work = kzalloc(sizeof(*nq_work), GFP_ATOMIC); 1969 if (nq_work) { 1970 nq_work->cq = qp->scq; 1971 nq_work->nq = qp->scq->nq; 1972 INIT_WORK(&nq_work->work, bnxt_qpn_cqn_sched_task); 1973 queue_work(qp->scq->nq->cqn_wq, &nq_work->work); 1974 } else { 1975 dev_err(&hwq->pdev->dev, 1976 "FP: Failed to allocate SQ nq_work!\n"); 1977 rc = -ENOMEM; 1978 } 1979 } 1980 return rc; 1981 } 1982 1983 void bnxt_qplib_post_recv_db(struct bnxt_qplib_qp *qp) 1984 { 1985 struct bnxt_qplib_q *rq = &qp->rq; 1986 1987 bnxt_qplib_ring_prod_db(&rq->dbinfo, DBC_DBC_TYPE_RQ); 1988 } 1989 1990 int bnxt_qplib_post_recv(struct bnxt_qplib_qp *qp, 1991 struct bnxt_qplib_swqe *wqe) 1992 { 1993 struct bnxt_qplib_nq_work *nq_work = NULL; 1994 struct bnxt_qplib_q *rq = &qp->rq; 1995 struct rq_wqe_hdr *base_hdr; 1996 struct rq_ext_hdr *ext_hdr; 1997 struct bnxt_qplib_hwq *hwq; 1998 struct bnxt_qplib_swq *swq; 1999 bool sch_handler = false; 2000 u16 wqe_sz, idx; 2001 u32 wqe_idx; 2002 int rc = 0; 2003 2004 hwq = &rq->hwq; 2005 if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_RESET) { 2006 dev_err(&hwq->pdev->dev, 2007 "QPLIB: FP: QP (0x%x) is in the 0x%x state", 2008 qp->id, qp->state); 2009 rc = -EINVAL; 2010 goto done; 2011 } 2012 2013 if (bnxt_qplib_queue_full(rq, rq->dbinfo.max_slot)) { 2014 dev_err(&hwq->pdev->dev, 2015 "FP: QP (0x%x) RQ is full!\n", qp->id); 2016 rc = -EINVAL; 2017 goto done; 2018 } 2019 2020 swq = bnxt_qplib_get_swqe(rq, &wqe_idx); 2021 swq->wr_id = wqe->wr_id; 2022 swq->slots = rq->dbinfo.max_slot; 2023 2024 if (qp->state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { 2025 sch_handler = true; 2026 dev_dbg(&hwq->pdev->dev, 2027 "%s: Error QP. Scheduling for poll_cq\n", __func__); 2028 goto queue_err; 2029 } 2030 2031 idx = 0; 2032 base_hdr = bnxt_qplib_get_prod_qe(hwq, idx++); 2033 ext_hdr = bnxt_qplib_get_prod_qe(hwq, idx++); 2034 memset(base_hdr, 0, sizeof(struct sq_sge)); 2035 memset(ext_hdr, 0, sizeof(struct sq_sge)); 2036 wqe_sz = (sizeof(struct rq_wqe_hdr) + 2037 wqe->num_sge * sizeof(struct sq_sge)) >> 4; 2038 bnxt_qplib_put_sges(hwq, wqe->sg_list, wqe->num_sge, &idx); 2039 if (!wqe->num_sge) { 2040 struct sq_sge *sge; 2041 2042 sge = bnxt_qplib_get_prod_qe(hwq, idx++); 2043 sge->size = 0; 2044 wqe_sz++; 2045 } 2046 base_hdr->wqe_type = wqe->type; 2047 base_hdr->flags = wqe->flags; 2048 base_hdr->wqe_size = wqe_sz; 2049 base_hdr->wr_id[0] = cpu_to_le32(wqe_idx); 2050 queue_err: 2051 bnxt_qplib_swq_mod_start(rq, wqe_idx); 2052 bnxt_qplib_hwq_incr_prod(hwq, swq->slots); 2053 done: 2054 if (sch_handler) { 2055 nq_work = kzalloc(sizeof(*nq_work), GFP_ATOMIC); 2056 if (nq_work) { 2057 nq_work->cq = qp->rcq; 2058 nq_work->nq = qp->rcq->nq; 2059 INIT_WORK(&nq_work->work, bnxt_qpn_cqn_sched_task); 2060 queue_work(qp->rcq->nq->cqn_wq, &nq_work->work); 2061 } else { 2062 dev_err(&hwq->pdev->dev, 2063 "FP: Failed to allocate RQ nq_work!\n"); 2064 rc = -ENOMEM; 2065 } 2066 } 2067 2068 return rc; 2069 } 2070 2071 /* CQ */ 2072 int bnxt_qplib_create_cq(struct bnxt_qplib_res *res, struct bnxt_qplib_cq *cq) 2073 { 2074 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 2075 struct bnxt_qplib_hwq_attr hwq_attr = {}; 2076 struct creq_create_cq_resp resp = {}; 2077 struct bnxt_qplib_cmdqmsg msg = {}; 2078 struct cmdq_create_cq req = {}; 2079 struct bnxt_qplib_pbl *pbl; 2080 u32 pg_sz_lvl; 2081 int rc; 2082 2083 if (!cq->dpi) { 2084 dev_err(&rcfw->pdev->dev, 2085 "FP: CREATE_CQ failed due to NULL DPI\n"); 2086 return -EINVAL; 2087 } 2088 2089 hwq_attr.res = res; 2090 hwq_attr.depth = cq->max_wqe; 2091 hwq_attr.stride = sizeof(struct cq_base); 2092 hwq_attr.type = HWQ_TYPE_QUEUE; 2093 hwq_attr.sginfo = &cq->sg_info; 2094 rc = bnxt_qplib_alloc_init_hwq(&cq->hwq, &hwq_attr); 2095 if (rc) 2096 return rc; 2097 2098 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 2099 CMDQ_BASE_OPCODE_CREATE_CQ, 2100 sizeof(req)); 2101 2102 req.dpi = cpu_to_le32(cq->dpi->dpi); 2103 req.cq_handle = cpu_to_le64(cq->cq_handle); 2104 req.cq_size = cpu_to_le32(cq->hwq.max_elements); 2105 pbl = &cq->hwq.pbl[PBL_LVL_0]; 2106 pg_sz_lvl = (bnxt_qplib_base_pg_size(&cq->hwq) << 2107 CMDQ_CREATE_CQ_PG_SIZE_SFT); 2108 pg_sz_lvl |= (cq->hwq.level & CMDQ_CREATE_CQ_LVL_MASK); 2109 req.pg_size_lvl = cpu_to_le32(pg_sz_lvl); 2110 req.pbl = cpu_to_le64(pbl->pg_map_arr[0]); 2111 req.cq_fco_cnq_id = cpu_to_le32( 2112 (cq->cnq_hw_ring_id & CMDQ_CREATE_CQ_CNQ_ID_MASK) << 2113 CMDQ_CREATE_CQ_CNQ_ID_SFT); 2114 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), 2115 sizeof(resp), 0); 2116 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 2117 if (rc) 2118 goto fail; 2119 2120 cq->id = le32_to_cpu(resp.xid); 2121 cq->period = BNXT_QPLIB_QUEUE_START_PERIOD; 2122 init_waitqueue_head(&cq->waitq); 2123 INIT_LIST_HEAD(&cq->sqf_head); 2124 INIT_LIST_HEAD(&cq->rqf_head); 2125 spin_lock_init(&cq->compl_lock); 2126 spin_lock_init(&cq->flush_lock); 2127 2128 cq->dbinfo.hwq = &cq->hwq; 2129 cq->dbinfo.xid = cq->id; 2130 cq->dbinfo.db = cq->dpi->dbr; 2131 cq->dbinfo.priv_db = res->dpi_tbl.priv_db; 2132 2133 bnxt_qplib_armen_db(&cq->dbinfo, DBC_DBC_TYPE_CQ_ARMENA); 2134 2135 return 0; 2136 2137 fail: 2138 bnxt_qplib_free_hwq(res, &cq->hwq); 2139 return rc; 2140 } 2141 2142 void bnxt_qplib_resize_cq_complete(struct bnxt_qplib_res *res, 2143 struct bnxt_qplib_cq *cq) 2144 { 2145 bnxt_qplib_free_hwq(res, &cq->hwq); 2146 memcpy(&cq->hwq, &cq->resize_hwq, sizeof(cq->hwq)); 2147 } 2148 2149 int bnxt_qplib_resize_cq(struct bnxt_qplib_res *res, struct bnxt_qplib_cq *cq, 2150 int new_cqes) 2151 { 2152 struct bnxt_qplib_hwq_attr hwq_attr = {}; 2153 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 2154 struct creq_resize_cq_resp resp = {}; 2155 struct bnxt_qplib_cmdqmsg msg = {}; 2156 struct cmdq_resize_cq req = {}; 2157 struct bnxt_qplib_pbl *pbl; 2158 u32 pg_sz, lvl, new_sz; 2159 int rc; 2160 2161 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 2162 CMDQ_BASE_OPCODE_RESIZE_CQ, 2163 sizeof(req)); 2164 hwq_attr.sginfo = &cq->sg_info; 2165 hwq_attr.res = res; 2166 hwq_attr.depth = new_cqes; 2167 hwq_attr.stride = sizeof(struct cq_base); 2168 hwq_attr.type = HWQ_TYPE_QUEUE; 2169 rc = bnxt_qplib_alloc_init_hwq(&cq->resize_hwq, &hwq_attr); 2170 if (rc) 2171 return rc; 2172 2173 req.cq_cid = cpu_to_le32(cq->id); 2174 pbl = &cq->resize_hwq.pbl[PBL_LVL_0]; 2175 pg_sz = bnxt_qplib_base_pg_size(&cq->resize_hwq); 2176 lvl = (cq->resize_hwq.level << CMDQ_RESIZE_CQ_LVL_SFT) & 2177 CMDQ_RESIZE_CQ_LVL_MASK; 2178 new_sz = (new_cqes << CMDQ_RESIZE_CQ_NEW_CQ_SIZE_SFT) & 2179 CMDQ_RESIZE_CQ_NEW_CQ_SIZE_MASK; 2180 req.new_cq_size_pg_size_lvl = cpu_to_le32(new_sz | pg_sz | lvl); 2181 req.new_pbl = cpu_to_le64(pbl->pg_map_arr[0]); 2182 2183 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), 2184 sizeof(resp), 0); 2185 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 2186 return rc; 2187 } 2188 2189 int bnxt_qplib_destroy_cq(struct bnxt_qplib_res *res, struct bnxt_qplib_cq *cq) 2190 { 2191 struct bnxt_qplib_rcfw *rcfw = res->rcfw; 2192 struct creq_destroy_cq_resp resp = {}; 2193 struct bnxt_qplib_cmdqmsg msg = {}; 2194 struct cmdq_destroy_cq req = {}; 2195 u16 total_cnq_events; 2196 int rc; 2197 2198 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req, 2199 CMDQ_BASE_OPCODE_DESTROY_CQ, 2200 sizeof(req)); 2201 2202 req.cq_cid = cpu_to_le32(cq->id); 2203 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), 2204 sizeof(resp), 0); 2205 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg); 2206 if (rc) 2207 return rc; 2208 total_cnq_events = le16_to_cpu(resp.total_cnq_events); 2209 __wait_for_all_nqes(cq, total_cnq_events); 2210 bnxt_qplib_free_hwq(res, &cq->hwq); 2211 return 0; 2212 } 2213 2214 static int __flush_sq(struct bnxt_qplib_q *sq, struct bnxt_qplib_qp *qp, 2215 struct bnxt_qplib_cqe **pcqe, int *budget) 2216 { 2217 struct bnxt_qplib_cqe *cqe; 2218 u32 start, last; 2219 int rc = 0; 2220 2221 /* Now complete all outstanding SQEs with FLUSHED_ERR */ 2222 start = sq->swq_start; 2223 cqe = *pcqe; 2224 while (*budget) { 2225 last = sq->swq_last; 2226 if (start == last) 2227 break; 2228 /* Skip the FENCE WQE completions */ 2229 if (sq->swq[last].wr_id == BNXT_QPLIB_FENCE_WRID) { 2230 bnxt_qplib_cancel_phantom_processing(qp); 2231 goto skip_compl; 2232 } 2233 memset(cqe, 0, sizeof(*cqe)); 2234 cqe->status = CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR; 2235 cqe->opcode = CQ_BASE_CQE_TYPE_REQ; 2236 cqe->qp_handle = (u64)(unsigned long)qp; 2237 cqe->wr_id = sq->swq[last].wr_id; 2238 cqe->src_qp = qp->id; 2239 cqe->type = sq->swq[last].type; 2240 cqe++; 2241 (*budget)--; 2242 skip_compl: 2243 bnxt_qplib_hwq_incr_cons(&sq->hwq, sq->swq[last].slots); 2244 sq->swq_last = sq->swq[last].next_idx; 2245 } 2246 *pcqe = cqe; 2247 if (!(*budget) && sq->swq_last != start) 2248 /* Out of budget */ 2249 rc = -EAGAIN; 2250 2251 return rc; 2252 } 2253 2254 static int __flush_rq(struct bnxt_qplib_q *rq, struct bnxt_qplib_qp *qp, 2255 struct bnxt_qplib_cqe **pcqe, int *budget) 2256 { 2257 struct bnxt_qplib_cqe *cqe; 2258 u32 start, last; 2259 int opcode = 0; 2260 int rc = 0; 2261 2262 switch (qp->type) { 2263 case CMDQ_CREATE_QP1_TYPE_GSI: 2264 opcode = CQ_BASE_CQE_TYPE_RES_RAWETH_QP1; 2265 break; 2266 case CMDQ_CREATE_QP_TYPE_RC: 2267 opcode = CQ_BASE_CQE_TYPE_RES_RC; 2268 break; 2269 case CMDQ_CREATE_QP_TYPE_UD: 2270 case CMDQ_CREATE_QP_TYPE_GSI: 2271 opcode = CQ_BASE_CQE_TYPE_RES_UD; 2272 break; 2273 } 2274 2275 /* Flush the rest of the RQ */ 2276 start = rq->swq_start; 2277 cqe = *pcqe; 2278 while (*budget) { 2279 last = rq->swq_last; 2280 if (last == start) 2281 break; 2282 memset(cqe, 0, sizeof(*cqe)); 2283 cqe->status = 2284 CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR; 2285 cqe->opcode = opcode; 2286 cqe->qp_handle = (unsigned long)qp; 2287 cqe->wr_id = rq->swq[last].wr_id; 2288 cqe++; 2289 (*budget)--; 2290 bnxt_qplib_hwq_incr_cons(&rq->hwq, rq->swq[last].slots); 2291 rq->swq_last = rq->swq[last].next_idx; 2292 } 2293 *pcqe = cqe; 2294 if (!*budget && rq->swq_last != start) 2295 /* Out of budget */ 2296 rc = -EAGAIN; 2297 2298 return rc; 2299 } 2300 2301 void bnxt_qplib_mark_qp_error(void *qp_handle) 2302 { 2303 struct bnxt_qplib_qp *qp = qp_handle; 2304 2305 if (!qp) 2306 return; 2307 2308 /* Must block new posting of SQ and RQ */ 2309 qp->state = CMDQ_MODIFY_QP_NEW_STATE_ERR; 2310 bnxt_qplib_cancel_phantom_processing(qp); 2311 } 2312 2313 /* Note: SQE is valid from sw_sq_cons up to cqe_sq_cons (exclusive) 2314 * CQE is track from sw_cq_cons to max_element but valid only if VALID=1 2315 */ 2316 static int do_wa9060(struct bnxt_qplib_qp *qp, struct bnxt_qplib_cq *cq, 2317 u32 cq_cons, u32 swq_last, u32 cqe_sq_cons) 2318 { 2319 u32 peek_sw_cq_cons, peek_raw_cq_cons, peek_sq_cons_idx; 2320 struct bnxt_qplib_q *sq = &qp->sq; 2321 struct cq_req *peek_req_hwcqe; 2322 struct bnxt_qplib_qp *peek_qp; 2323 struct bnxt_qplib_q *peek_sq; 2324 struct bnxt_qplib_swq *swq; 2325 struct cq_base *peek_hwcqe; 2326 int i, rc = 0; 2327 2328 /* Normal mode */ 2329 /* Check for the psn_search marking before completing */ 2330 swq = &sq->swq[swq_last]; 2331 if (swq->psn_search && 2332 le32_to_cpu(swq->psn_search->flags_next_psn) & 0x80000000) { 2333 /* Unmark */ 2334 swq->psn_search->flags_next_psn = cpu_to_le32 2335 (le32_to_cpu(swq->psn_search->flags_next_psn) 2336 & ~0x80000000); 2337 dev_dbg(&cq->hwq.pdev->dev, 2338 "FP: Process Req cq_cons=0x%x qp=0x%x sq cons sw=0x%x cqe=0x%x marked!\n", 2339 cq_cons, qp->id, swq_last, cqe_sq_cons); 2340 sq->condition = true; 2341 sq->send_phantom = true; 2342 2343 /* TODO: Only ARM if the previous SQE is ARMALL */ 2344 bnxt_qplib_ring_db(&cq->dbinfo, DBC_DBC_TYPE_CQ_ARMALL); 2345 rc = -EAGAIN; 2346 goto out; 2347 } 2348 if (sq->condition) { 2349 /* Peek at the completions */ 2350 peek_raw_cq_cons = cq->hwq.cons; 2351 peek_sw_cq_cons = cq_cons; 2352 i = cq->hwq.max_elements; 2353 while (i--) { 2354 peek_sw_cq_cons = HWQ_CMP((peek_sw_cq_cons), &cq->hwq); 2355 peek_hwcqe = bnxt_qplib_get_qe(&cq->hwq, 2356 peek_sw_cq_cons, NULL); 2357 /* If the next hwcqe is VALID */ 2358 if (CQE_CMP_VALID(peek_hwcqe, peek_raw_cq_cons, 2359 cq->hwq.max_elements)) { 2360 /* 2361 * The valid test of the entry must be done first before 2362 * reading any further. 2363 */ 2364 dma_rmb(); 2365 /* If the next hwcqe is a REQ */ 2366 if ((peek_hwcqe->cqe_type_toggle & 2367 CQ_BASE_CQE_TYPE_MASK) == 2368 CQ_BASE_CQE_TYPE_REQ) { 2369 peek_req_hwcqe = (struct cq_req *) 2370 peek_hwcqe; 2371 peek_qp = (struct bnxt_qplib_qp *) 2372 ((unsigned long) 2373 le64_to_cpu 2374 (peek_req_hwcqe->qp_handle)); 2375 peek_sq = &peek_qp->sq; 2376 peek_sq_cons_idx = 2377 ((le16_to_cpu( 2378 peek_req_hwcqe->sq_cons_idx) 2379 - 1) % sq->max_wqe); 2380 /* If the hwcqe's sq's wr_id matches */ 2381 if (peek_sq == sq && 2382 sq->swq[peek_sq_cons_idx].wr_id == 2383 BNXT_QPLIB_FENCE_WRID) { 2384 /* 2385 * Unbreak only if the phantom 2386 * comes back 2387 */ 2388 dev_dbg(&cq->hwq.pdev->dev, 2389 "FP: Got Phantom CQE\n"); 2390 sq->condition = false; 2391 sq->single = true; 2392 rc = 0; 2393 goto out; 2394 } 2395 } 2396 /* Valid but not the phantom, so keep looping */ 2397 } else { 2398 /* Not valid yet, just exit and wait */ 2399 rc = -EINVAL; 2400 goto out; 2401 } 2402 peek_sw_cq_cons++; 2403 peek_raw_cq_cons++; 2404 } 2405 dev_err(&cq->hwq.pdev->dev, 2406 "Should not have come here! cq_cons=0x%x qp=0x%x sq cons sw=0x%x hw=0x%x\n", 2407 cq_cons, qp->id, swq_last, cqe_sq_cons); 2408 rc = -EINVAL; 2409 } 2410 out: 2411 return rc; 2412 } 2413 2414 static int bnxt_qplib_cq_process_req(struct bnxt_qplib_cq *cq, 2415 struct cq_req *hwcqe, 2416 struct bnxt_qplib_cqe **pcqe, int *budget, 2417 u32 cq_cons, struct bnxt_qplib_qp **lib_qp) 2418 { 2419 struct bnxt_qplib_swq *swq; 2420 struct bnxt_qplib_cqe *cqe; 2421 struct bnxt_qplib_qp *qp; 2422 struct bnxt_qplib_q *sq; 2423 u32 cqe_sq_cons; 2424 int rc = 0; 2425 2426 qp = (struct bnxt_qplib_qp *)((unsigned long) 2427 le64_to_cpu(hwcqe->qp_handle)); 2428 if (!qp) { 2429 dev_err(&cq->hwq.pdev->dev, 2430 "FP: Process Req qp is NULL\n"); 2431 return -EINVAL; 2432 } 2433 sq = &qp->sq; 2434 2435 cqe_sq_cons = le16_to_cpu(hwcqe->sq_cons_idx) % sq->max_wqe; 2436 if (qp->sq.flushed) { 2437 dev_dbg(&cq->hwq.pdev->dev, 2438 "%s: QP in Flush QP = %p\n", __func__, qp); 2439 goto done; 2440 } 2441 /* Require to walk the sq's swq to fabricate CQEs for all previously 2442 * signaled SWQEs due to CQE aggregation from the current sq cons 2443 * to the cqe_sq_cons 2444 */ 2445 cqe = *pcqe; 2446 while (*budget) { 2447 if (sq->swq_last == cqe_sq_cons) 2448 /* Done */ 2449 break; 2450 2451 swq = &sq->swq[sq->swq_last]; 2452 memset(cqe, 0, sizeof(*cqe)); 2453 cqe->opcode = CQ_BASE_CQE_TYPE_REQ; 2454 cqe->qp_handle = (u64)(unsigned long)qp; 2455 cqe->src_qp = qp->id; 2456 cqe->wr_id = swq->wr_id; 2457 if (cqe->wr_id == BNXT_QPLIB_FENCE_WRID) 2458 goto skip; 2459 cqe->type = swq->type; 2460 2461 /* For the last CQE, check for status. For errors, regardless 2462 * of the request being signaled or not, it must complete with 2463 * the hwcqe error status 2464 */ 2465 if (swq->next_idx == cqe_sq_cons && 2466 hwcqe->status != CQ_REQ_STATUS_OK) { 2467 cqe->status = hwcqe->status; 2468 dev_err(&cq->hwq.pdev->dev, 2469 "FP: CQ Processed Req wr_id[%d] = 0x%llx with status 0x%x\n", 2470 sq->swq_last, cqe->wr_id, cqe->status); 2471 cqe++; 2472 (*budget)--; 2473 bnxt_qplib_mark_qp_error(qp); 2474 /* Add qp to flush list of the CQ */ 2475 bnxt_qplib_add_flush_qp(qp); 2476 } else { 2477 /* Before we complete, do WA 9060 */ 2478 if (do_wa9060(qp, cq, cq_cons, sq->swq_last, 2479 cqe_sq_cons)) { 2480 *lib_qp = qp; 2481 goto out; 2482 } 2483 if (swq->flags & SQ_SEND_FLAGS_SIGNAL_COMP) { 2484 cqe->status = CQ_REQ_STATUS_OK; 2485 cqe++; 2486 (*budget)--; 2487 } 2488 } 2489 skip: 2490 bnxt_qplib_hwq_incr_cons(&sq->hwq, swq->slots); 2491 sq->swq_last = swq->next_idx; 2492 if (sq->single) 2493 break; 2494 } 2495 out: 2496 *pcqe = cqe; 2497 if (sq->swq_last != cqe_sq_cons) { 2498 /* Out of budget */ 2499 rc = -EAGAIN; 2500 goto done; 2501 } 2502 /* 2503 * Back to normal completion mode only after it has completed all of 2504 * the WC for this CQE 2505 */ 2506 sq->single = false; 2507 done: 2508 return rc; 2509 } 2510 2511 static void bnxt_qplib_release_srqe(struct bnxt_qplib_srq *srq, u32 tag) 2512 { 2513 spin_lock(&srq->hwq.lock); 2514 srq->swq[srq->last_idx].next_idx = (int)tag; 2515 srq->last_idx = (int)tag; 2516 srq->swq[srq->last_idx].next_idx = -1; 2517 srq->hwq.cons++; /* Support for SRQE counter */ 2518 spin_unlock(&srq->hwq.lock); 2519 } 2520 2521 static int bnxt_qplib_cq_process_res_rc(struct bnxt_qplib_cq *cq, 2522 struct cq_res_rc *hwcqe, 2523 struct bnxt_qplib_cqe **pcqe, 2524 int *budget) 2525 { 2526 struct bnxt_qplib_srq *srq; 2527 struct bnxt_qplib_cqe *cqe; 2528 struct bnxt_qplib_qp *qp; 2529 struct bnxt_qplib_q *rq; 2530 u32 wr_id_idx; 2531 2532 qp = (struct bnxt_qplib_qp *)((unsigned long) 2533 le64_to_cpu(hwcqe->qp_handle)); 2534 if (!qp) { 2535 dev_err(&cq->hwq.pdev->dev, "process_cq RC qp is NULL\n"); 2536 return -EINVAL; 2537 } 2538 if (qp->rq.flushed) { 2539 dev_dbg(&cq->hwq.pdev->dev, 2540 "%s: QP in Flush QP = %p\n", __func__, qp); 2541 return 0; 2542 } 2543 2544 cqe = *pcqe; 2545 cqe->opcode = hwcqe->cqe_type_toggle & CQ_BASE_CQE_TYPE_MASK; 2546 cqe->length = le32_to_cpu(hwcqe->length); 2547 cqe->invrkey = le32_to_cpu(hwcqe->imm_data_or_inv_r_key); 2548 cqe->mr_handle = le64_to_cpu(hwcqe->mr_handle); 2549 cqe->flags = le16_to_cpu(hwcqe->flags); 2550 cqe->status = hwcqe->status; 2551 cqe->qp_handle = (u64)(unsigned long)qp; 2552 2553 wr_id_idx = le32_to_cpu(hwcqe->srq_or_rq_wr_id) & 2554 CQ_RES_RC_SRQ_OR_RQ_WR_ID_MASK; 2555 if (cqe->flags & CQ_RES_RC_FLAGS_SRQ_SRQ) { 2556 srq = qp->srq; 2557 if (!srq) 2558 return -EINVAL; 2559 if (wr_id_idx >= srq->hwq.max_elements) { 2560 dev_err(&cq->hwq.pdev->dev, 2561 "FP: CQ Process RC wr_id idx 0x%x exceeded SRQ max 0x%x\n", 2562 wr_id_idx, srq->hwq.max_elements); 2563 return -EINVAL; 2564 } 2565 cqe->wr_id = srq->swq[wr_id_idx].wr_id; 2566 bnxt_qplib_release_srqe(srq, wr_id_idx); 2567 cqe++; 2568 (*budget)--; 2569 *pcqe = cqe; 2570 } else { 2571 struct bnxt_qplib_swq *swq; 2572 2573 rq = &qp->rq; 2574 if (wr_id_idx > (rq->max_wqe - 1)) { 2575 dev_err(&cq->hwq.pdev->dev, 2576 "FP: CQ Process RC wr_id idx 0x%x exceeded RQ max 0x%x\n", 2577 wr_id_idx, rq->max_wqe); 2578 return -EINVAL; 2579 } 2580 if (wr_id_idx != rq->swq_last) 2581 return -EINVAL; 2582 swq = &rq->swq[rq->swq_last]; 2583 cqe->wr_id = swq->wr_id; 2584 cqe++; 2585 (*budget)--; 2586 bnxt_qplib_hwq_incr_cons(&rq->hwq, swq->slots); 2587 rq->swq_last = swq->next_idx; 2588 *pcqe = cqe; 2589 2590 if (hwcqe->status != CQ_RES_RC_STATUS_OK) { 2591 qp->state = CMDQ_MODIFY_QP_NEW_STATE_ERR; 2592 /* Add qp to flush list of the CQ */ 2593 bnxt_qplib_add_flush_qp(qp); 2594 } 2595 } 2596 2597 return 0; 2598 } 2599 2600 static int bnxt_qplib_cq_process_res_ud(struct bnxt_qplib_cq *cq, 2601 struct cq_res_ud *hwcqe, 2602 struct bnxt_qplib_cqe **pcqe, 2603 int *budget) 2604 { 2605 struct bnxt_qplib_srq *srq; 2606 struct bnxt_qplib_cqe *cqe; 2607 struct bnxt_qplib_qp *qp; 2608 struct bnxt_qplib_q *rq; 2609 u32 wr_id_idx; 2610 2611 qp = (struct bnxt_qplib_qp *)((unsigned long) 2612 le64_to_cpu(hwcqe->qp_handle)); 2613 if (!qp) { 2614 dev_err(&cq->hwq.pdev->dev, "process_cq UD qp is NULL\n"); 2615 return -EINVAL; 2616 } 2617 if (qp->rq.flushed) { 2618 dev_dbg(&cq->hwq.pdev->dev, 2619 "%s: QP in Flush QP = %p\n", __func__, qp); 2620 return 0; 2621 } 2622 cqe = *pcqe; 2623 cqe->opcode = hwcqe->cqe_type_toggle & CQ_BASE_CQE_TYPE_MASK; 2624 cqe->length = le16_to_cpu(hwcqe->length) & CQ_RES_UD_LENGTH_MASK; 2625 cqe->cfa_meta = le16_to_cpu(hwcqe->cfa_metadata); 2626 cqe->invrkey = le32_to_cpu(hwcqe->imm_data); 2627 cqe->flags = le16_to_cpu(hwcqe->flags); 2628 cqe->status = hwcqe->status; 2629 cqe->qp_handle = (u64)(unsigned long)qp; 2630 /*FIXME: Endianness fix needed for smace */ 2631 memcpy(cqe->smac, hwcqe->src_mac, ETH_ALEN); 2632 wr_id_idx = le32_to_cpu(hwcqe->src_qp_high_srq_or_rq_wr_id) 2633 & CQ_RES_UD_SRQ_OR_RQ_WR_ID_MASK; 2634 cqe->src_qp = le16_to_cpu(hwcqe->src_qp_low) | 2635 ((le32_to_cpu( 2636 hwcqe->src_qp_high_srq_or_rq_wr_id) & 2637 CQ_RES_UD_SRC_QP_HIGH_MASK) >> 8); 2638 2639 if (cqe->flags & CQ_RES_RC_FLAGS_SRQ_SRQ) { 2640 srq = qp->srq; 2641 if (!srq) 2642 return -EINVAL; 2643 2644 if (wr_id_idx >= srq->hwq.max_elements) { 2645 dev_err(&cq->hwq.pdev->dev, 2646 "FP: CQ Process UD wr_id idx 0x%x exceeded SRQ max 0x%x\n", 2647 wr_id_idx, srq->hwq.max_elements); 2648 return -EINVAL; 2649 } 2650 cqe->wr_id = srq->swq[wr_id_idx].wr_id; 2651 bnxt_qplib_release_srqe(srq, wr_id_idx); 2652 cqe++; 2653 (*budget)--; 2654 *pcqe = cqe; 2655 } else { 2656 struct bnxt_qplib_swq *swq; 2657 2658 rq = &qp->rq; 2659 if (wr_id_idx > (rq->max_wqe - 1)) { 2660 dev_err(&cq->hwq.pdev->dev, 2661 "FP: CQ Process UD wr_id idx 0x%x exceeded RQ max 0x%x\n", 2662 wr_id_idx, rq->max_wqe); 2663 return -EINVAL; 2664 } 2665 2666 if (rq->swq_last != wr_id_idx) 2667 return -EINVAL; 2668 swq = &rq->swq[rq->swq_last]; 2669 cqe->wr_id = swq->wr_id; 2670 cqe++; 2671 (*budget)--; 2672 bnxt_qplib_hwq_incr_cons(&rq->hwq, swq->slots); 2673 rq->swq_last = swq->next_idx; 2674 *pcqe = cqe; 2675 2676 if (hwcqe->status != CQ_RES_RC_STATUS_OK) { 2677 qp->state = CMDQ_MODIFY_QP_NEW_STATE_ERR; 2678 /* Add qp to flush list of the CQ */ 2679 bnxt_qplib_add_flush_qp(qp); 2680 } 2681 } 2682 2683 return 0; 2684 } 2685 2686 bool bnxt_qplib_is_cq_empty(struct bnxt_qplib_cq *cq) 2687 { 2688 struct cq_base *hw_cqe; 2689 u32 sw_cons, raw_cons; 2690 bool rc = true; 2691 2692 raw_cons = cq->hwq.cons; 2693 sw_cons = HWQ_CMP(raw_cons, &cq->hwq); 2694 hw_cqe = bnxt_qplib_get_qe(&cq->hwq, sw_cons, NULL); 2695 /* Check for Valid bit. If the CQE is valid, return false */ 2696 rc = !CQE_CMP_VALID(hw_cqe, raw_cons, cq->hwq.max_elements); 2697 return rc; 2698 } 2699 2700 static int bnxt_qplib_cq_process_res_raweth_qp1(struct bnxt_qplib_cq *cq, 2701 struct cq_res_raweth_qp1 *hwcqe, 2702 struct bnxt_qplib_cqe **pcqe, 2703 int *budget) 2704 { 2705 struct bnxt_qplib_qp *qp; 2706 struct bnxt_qplib_q *rq; 2707 struct bnxt_qplib_srq *srq; 2708 struct bnxt_qplib_cqe *cqe; 2709 u32 wr_id_idx; 2710 2711 qp = (struct bnxt_qplib_qp *)((unsigned long) 2712 le64_to_cpu(hwcqe->qp_handle)); 2713 if (!qp) { 2714 dev_err(&cq->hwq.pdev->dev, "process_cq Raw/QP1 qp is NULL\n"); 2715 return -EINVAL; 2716 } 2717 if (qp->rq.flushed) { 2718 dev_dbg(&cq->hwq.pdev->dev, 2719 "%s: QP in Flush QP = %p\n", __func__, qp); 2720 return 0; 2721 } 2722 cqe = *pcqe; 2723 cqe->opcode = hwcqe->cqe_type_toggle & CQ_BASE_CQE_TYPE_MASK; 2724 cqe->flags = le16_to_cpu(hwcqe->flags); 2725 cqe->qp_handle = (u64)(unsigned long)qp; 2726 2727 wr_id_idx = 2728 le32_to_cpu(hwcqe->raweth_qp1_payload_offset_srq_or_rq_wr_id) 2729 & CQ_RES_RAWETH_QP1_SRQ_OR_RQ_WR_ID_MASK; 2730 cqe->src_qp = qp->id; 2731 if (qp->id == 1 && !cqe->length) { 2732 /* Add workaround for the length misdetection */ 2733 cqe->length = 296; 2734 } else { 2735 cqe->length = le16_to_cpu(hwcqe->length); 2736 } 2737 cqe->pkey_index = qp->pkey_index; 2738 memcpy(cqe->smac, qp->smac, 6); 2739 2740 cqe->raweth_qp1_flags = le16_to_cpu(hwcqe->raweth_qp1_flags); 2741 cqe->raweth_qp1_flags2 = le32_to_cpu(hwcqe->raweth_qp1_flags2); 2742 cqe->raweth_qp1_metadata = le32_to_cpu(hwcqe->raweth_qp1_metadata); 2743 2744 if (cqe->flags & CQ_RES_RAWETH_QP1_FLAGS_SRQ_SRQ) { 2745 srq = qp->srq; 2746 if (!srq) { 2747 dev_err(&cq->hwq.pdev->dev, 2748 "FP: SRQ used but not defined??\n"); 2749 return -EINVAL; 2750 } 2751 if (wr_id_idx >= srq->hwq.max_elements) { 2752 dev_err(&cq->hwq.pdev->dev, 2753 "FP: CQ Process Raw/QP1 wr_id idx 0x%x exceeded SRQ max 0x%x\n", 2754 wr_id_idx, srq->hwq.max_elements); 2755 return -EINVAL; 2756 } 2757 cqe->wr_id = srq->swq[wr_id_idx].wr_id; 2758 bnxt_qplib_release_srqe(srq, wr_id_idx); 2759 cqe++; 2760 (*budget)--; 2761 *pcqe = cqe; 2762 } else { 2763 struct bnxt_qplib_swq *swq; 2764 2765 rq = &qp->rq; 2766 if (wr_id_idx > (rq->max_wqe - 1)) { 2767 dev_err(&cq->hwq.pdev->dev, 2768 "FP: CQ Process Raw/QP1 RQ wr_id idx 0x%x exceeded RQ max 0x%x\n", 2769 wr_id_idx, rq->max_wqe); 2770 return -EINVAL; 2771 } 2772 if (rq->swq_last != wr_id_idx) 2773 return -EINVAL; 2774 swq = &rq->swq[rq->swq_last]; 2775 cqe->wr_id = swq->wr_id; 2776 cqe++; 2777 (*budget)--; 2778 bnxt_qplib_hwq_incr_cons(&rq->hwq, swq->slots); 2779 rq->swq_last = swq->next_idx; 2780 *pcqe = cqe; 2781 2782 if (hwcqe->status != CQ_RES_RC_STATUS_OK) { 2783 qp->state = CMDQ_MODIFY_QP_NEW_STATE_ERR; 2784 /* Add qp to flush list of the CQ */ 2785 bnxt_qplib_add_flush_qp(qp); 2786 } 2787 } 2788 2789 return 0; 2790 } 2791 2792 static int bnxt_qplib_cq_process_terminal(struct bnxt_qplib_cq *cq, 2793 struct cq_terminal *hwcqe, 2794 struct bnxt_qplib_cqe **pcqe, 2795 int *budget) 2796 { 2797 struct bnxt_qplib_qp *qp; 2798 struct bnxt_qplib_q *sq, *rq; 2799 struct bnxt_qplib_cqe *cqe; 2800 u32 swq_last = 0, cqe_cons; 2801 int rc = 0; 2802 2803 /* Check the Status */ 2804 if (hwcqe->status != CQ_TERMINAL_STATUS_OK) 2805 dev_warn(&cq->hwq.pdev->dev, 2806 "FP: CQ Process Terminal Error status = 0x%x\n", 2807 hwcqe->status); 2808 2809 qp = (struct bnxt_qplib_qp *)((unsigned long) 2810 le64_to_cpu(hwcqe->qp_handle)); 2811 if (!qp) 2812 return -EINVAL; 2813 2814 /* Must block new posting of SQ and RQ */ 2815 qp->state = CMDQ_MODIFY_QP_NEW_STATE_ERR; 2816 2817 sq = &qp->sq; 2818 rq = &qp->rq; 2819 2820 cqe_cons = le16_to_cpu(hwcqe->sq_cons_idx); 2821 if (cqe_cons == 0xFFFF) 2822 goto do_rq; 2823 cqe_cons %= sq->max_wqe; 2824 2825 if (qp->sq.flushed) { 2826 dev_dbg(&cq->hwq.pdev->dev, 2827 "%s: QP in Flush QP = %p\n", __func__, qp); 2828 goto sq_done; 2829 } 2830 2831 /* Terminal CQE can also include aggregated successful CQEs prior. 2832 * So we must complete all CQEs from the current sq's cons to the 2833 * cq_cons with status OK 2834 */ 2835 cqe = *pcqe; 2836 while (*budget) { 2837 swq_last = sq->swq_last; 2838 if (swq_last == cqe_cons) 2839 break; 2840 if (sq->swq[swq_last].flags & SQ_SEND_FLAGS_SIGNAL_COMP) { 2841 memset(cqe, 0, sizeof(*cqe)); 2842 cqe->status = CQ_REQ_STATUS_OK; 2843 cqe->opcode = CQ_BASE_CQE_TYPE_REQ; 2844 cqe->qp_handle = (u64)(unsigned long)qp; 2845 cqe->src_qp = qp->id; 2846 cqe->wr_id = sq->swq[swq_last].wr_id; 2847 cqe->type = sq->swq[swq_last].type; 2848 cqe++; 2849 (*budget)--; 2850 } 2851 bnxt_qplib_hwq_incr_cons(&sq->hwq, sq->swq[swq_last].slots); 2852 sq->swq_last = sq->swq[swq_last].next_idx; 2853 } 2854 *pcqe = cqe; 2855 if (!(*budget) && swq_last != cqe_cons) { 2856 /* Out of budget */ 2857 rc = -EAGAIN; 2858 goto sq_done; 2859 } 2860 sq_done: 2861 if (rc) 2862 return rc; 2863 do_rq: 2864 cqe_cons = le16_to_cpu(hwcqe->rq_cons_idx); 2865 if (cqe_cons == 0xFFFF) { 2866 goto done; 2867 } else if (cqe_cons > rq->max_wqe - 1) { 2868 dev_err(&cq->hwq.pdev->dev, 2869 "FP: CQ Processed terminal reported rq_cons_idx 0x%x exceeds max 0x%x\n", 2870 cqe_cons, rq->max_wqe); 2871 rc = -EINVAL; 2872 goto done; 2873 } 2874 2875 if (qp->rq.flushed) { 2876 dev_dbg(&cq->hwq.pdev->dev, 2877 "%s: QP in Flush QP = %p\n", __func__, qp); 2878 rc = 0; 2879 goto done; 2880 } 2881 2882 /* Terminal CQE requires all posted RQEs to complete with FLUSHED_ERR 2883 * from the current rq->cons to the rq->prod regardless what the 2884 * rq->cons the terminal CQE indicates 2885 */ 2886 2887 /* Add qp to flush list of the CQ */ 2888 bnxt_qplib_add_flush_qp(qp); 2889 done: 2890 return rc; 2891 } 2892 2893 static int bnxt_qplib_cq_process_cutoff(struct bnxt_qplib_cq *cq, 2894 struct cq_cutoff *hwcqe) 2895 { 2896 /* Check the Status */ 2897 if (hwcqe->status != CQ_CUTOFF_STATUS_OK) { 2898 dev_err(&cq->hwq.pdev->dev, 2899 "FP: CQ Process Cutoff Error status = 0x%x\n", 2900 hwcqe->status); 2901 return -EINVAL; 2902 } 2903 clear_bit(CQ_FLAGS_RESIZE_IN_PROG, &cq->flags); 2904 wake_up_interruptible(&cq->waitq); 2905 2906 return 0; 2907 } 2908 2909 int bnxt_qplib_process_flush_list(struct bnxt_qplib_cq *cq, 2910 struct bnxt_qplib_cqe *cqe, 2911 int num_cqes) 2912 { 2913 struct bnxt_qplib_qp *qp = NULL; 2914 u32 budget = num_cqes; 2915 unsigned long flags; 2916 2917 spin_lock_irqsave(&cq->flush_lock, flags); 2918 list_for_each_entry(qp, &cq->sqf_head, sq_flush) { 2919 dev_dbg(&cq->hwq.pdev->dev, "FP: Flushing SQ QP= %p\n", qp); 2920 __flush_sq(&qp->sq, qp, &cqe, &budget); 2921 } 2922 2923 list_for_each_entry(qp, &cq->rqf_head, rq_flush) { 2924 dev_dbg(&cq->hwq.pdev->dev, "FP: Flushing RQ QP= %p\n", qp); 2925 __flush_rq(&qp->rq, qp, &cqe, &budget); 2926 } 2927 spin_unlock_irqrestore(&cq->flush_lock, flags); 2928 2929 return num_cqes - budget; 2930 } 2931 2932 int bnxt_qplib_poll_cq(struct bnxt_qplib_cq *cq, struct bnxt_qplib_cqe *cqe, 2933 int num_cqes, struct bnxt_qplib_qp **lib_qp) 2934 { 2935 struct cq_base *hw_cqe; 2936 u32 sw_cons, raw_cons; 2937 int budget, rc = 0; 2938 u8 type; 2939 2940 raw_cons = cq->hwq.cons; 2941 budget = num_cqes; 2942 2943 while (budget) { 2944 sw_cons = HWQ_CMP(raw_cons, &cq->hwq); 2945 hw_cqe = bnxt_qplib_get_qe(&cq->hwq, sw_cons, NULL); 2946 2947 /* Check for Valid bit */ 2948 if (!CQE_CMP_VALID(hw_cqe, raw_cons, cq->hwq.max_elements)) 2949 break; 2950 2951 /* 2952 * The valid test of the entry must be done first before 2953 * reading any further. 2954 */ 2955 dma_rmb(); 2956 /* From the device's respective CQE format to qplib_wc*/ 2957 type = hw_cqe->cqe_type_toggle & CQ_BASE_CQE_TYPE_MASK; 2958 switch (type) { 2959 case CQ_BASE_CQE_TYPE_REQ: 2960 rc = bnxt_qplib_cq_process_req(cq, 2961 (struct cq_req *)hw_cqe, 2962 &cqe, &budget, 2963 sw_cons, lib_qp); 2964 break; 2965 case CQ_BASE_CQE_TYPE_RES_RC: 2966 rc = bnxt_qplib_cq_process_res_rc(cq, 2967 (struct cq_res_rc *) 2968 hw_cqe, &cqe, 2969 &budget); 2970 break; 2971 case CQ_BASE_CQE_TYPE_RES_UD: 2972 rc = bnxt_qplib_cq_process_res_ud 2973 (cq, (struct cq_res_ud *)hw_cqe, &cqe, 2974 &budget); 2975 break; 2976 case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1: 2977 rc = bnxt_qplib_cq_process_res_raweth_qp1 2978 (cq, (struct cq_res_raweth_qp1 *) 2979 hw_cqe, &cqe, &budget); 2980 break; 2981 case CQ_BASE_CQE_TYPE_TERMINAL: 2982 rc = bnxt_qplib_cq_process_terminal 2983 (cq, (struct cq_terminal *)hw_cqe, 2984 &cqe, &budget); 2985 break; 2986 case CQ_BASE_CQE_TYPE_CUT_OFF: 2987 bnxt_qplib_cq_process_cutoff 2988 (cq, (struct cq_cutoff *)hw_cqe); 2989 /* Done processing this CQ */ 2990 goto exit; 2991 default: 2992 dev_err(&cq->hwq.pdev->dev, 2993 "process_cq unknown type 0x%lx\n", 2994 hw_cqe->cqe_type_toggle & 2995 CQ_BASE_CQE_TYPE_MASK); 2996 rc = -EINVAL; 2997 break; 2998 } 2999 if (rc < 0) { 3000 if (rc == -EAGAIN) 3001 break; 3002 /* Error while processing the CQE, just skip to the 3003 * next one 3004 */ 3005 if (type != CQ_BASE_CQE_TYPE_TERMINAL) 3006 dev_err(&cq->hwq.pdev->dev, 3007 "process_cqe error rc = 0x%x\n", rc); 3008 } 3009 raw_cons++; 3010 } 3011 if (cq->hwq.cons != raw_cons) { 3012 cq->hwq.cons = raw_cons; 3013 bnxt_qplib_ring_db(&cq->dbinfo, DBC_DBC_TYPE_CQ); 3014 } 3015 exit: 3016 return num_cqes - budget; 3017 } 3018 3019 void bnxt_qplib_req_notify_cq(struct bnxt_qplib_cq *cq, u32 arm_type) 3020 { 3021 if (arm_type) 3022 bnxt_qplib_ring_db(&cq->dbinfo, arm_type); 3023 /* Using cq->arm_state variable to track whether to issue cq handler */ 3024 atomic_set(&cq->arm_state, 1); 3025 } 3026 3027 void bnxt_qplib_flush_cqn_wq(struct bnxt_qplib_qp *qp) 3028 { 3029 flush_workqueue(qp->scq->nq->cqn_wq); 3030 if (qp->scq != qp->rcq) 3031 flush_workqueue(qp->rcq->nq->cqn_wq); 3032 } 3033