1 /* QLogic qedr NIC Driver 2 * Copyright (c) 2015-2016 QLogic Corporation 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and /or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 #include <linux/dma-mapping.h> 33 #include <linux/crc32.h> 34 #include <net/ip.h> 35 #include <net/ipv6.h> 36 #include <net/udp.h> 37 #include <linux/iommu.h> 38 39 #include <rdma/ib_verbs.h> 40 #include <rdma/ib_user_verbs.h> 41 #include <rdma/iw_cm.h> 42 #include <rdma/ib_umem.h> 43 #include <rdma/ib_addr.h> 44 #include <rdma/ib_cache.h> 45 46 #include <linux/qed/common_hsi.h> 47 #include "qedr_hsi_rdma.h" 48 #include <linux/qed/qed_if.h> 49 #include "qedr.h" 50 #include "verbs.h" 51 #include <rdma/qedr-abi.h> 52 #include "qedr_cm.h" 53 54 #define DB_ADDR_SHIFT(addr) ((addr) << DB_PWM_ADDR_OFFSET_SHIFT) 55 56 int qedr_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey) 57 { 58 if (index > QEDR_ROCE_PKEY_TABLE_LEN) 59 return -EINVAL; 60 61 *pkey = QEDR_ROCE_PKEY_DEFAULT; 62 return 0; 63 } 64 65 int qedr_query_gid(struct ib_device *ibdev, u8 port, int index, 66 union ib_gid *sgid) 67 { 68 struct qedr_dev *dev = get_qedr_dev(ibdev); 69 int rc = 0; 70 71 if (!rdma_cap_roce_gid_table(ibdev, port)) 72 return -ENODEV; 73 74 rc = ib_get_cached_gid(ibdev, port, index, sgid, NULL); 75 if (rc == -EAGAIN) { 76 memcpy(sgid, &zgid, sizeof(*sgid)); 77 return 0; 78 } 79 80 DP_DEBUG(dev, QEDR_MSG_INIT, "query gid: index=%d %llx:%llx\n", index, 81 sgid->global.interface_id, sgid->global.subnet_prefix); 82 83 return rc; 84 } 85 86 int qedr_add_gid(struct ib_device *device, u8 port_num, 87 unsigned int index, const union ib_gid *gid, 88 const struct ib_gid_attr *attr, void **context) 89 { 90 if (!rdma_cap_roce_gid_table(device, port_num)) 91 return -EINVAL; 92 93 if (port_num > QEDR_MAX_PORT) 94 return -EINVAL; 95 96 if (!context) 97 return -EINVAL; 98 99 return 0; 100 } 101 102 int qedr_del_gid(struct ib_device *device, u8 port_num, 103 unsigned int index, void **context) 104 { 105 if (!rdma_cap_roce_gid_table(device, port_num)) 106 return -EINVAL; 107 108 if (port_num > QEDR_MAX_PORT) 109 return -EINVAL; 110 111 if (!context) 112 return -EINVAL; 113 114 return 0; 115 } 116 117 int qedr_query_device(struct ib_device *ibdev, 118 struct ib_device_attr *attr, struct ib_udata *udata) 119 { 120 struct qedr_dev *dev = get_qedr_dev(ibdev); 121 struct qedr_device_attr *qattr = &dev->attr; 122 123 if (!dev->rdma_ctx) { 124 DP_ERR(dev, 125 "qedr_query_device called with invalid params rdma_ctx=%p\n", 126 dev->rdma_ctx); 127 return -EINVAL; 128 } 129 130 memset(attr, 0, sizeof(*attr)); 131 132 attr->fw_ver = qattr->fw_ver; 133 attr->sys_image_guid = qattr->sys_image_guid; 134 attr->max_mr_size = qattr->max_mr_size; 135 attr->page_size_cap = qattr->page_size_caps; 136 attr->vendor_id = qattr->vendor_id; 137 attr->vendor_part_id = qattr->vendor_part_id; 138 attr->hw_ver = qattr->hw_ver; 139 attr->max_qp = qattr->max_qp; 140 attr->max_qp_wr = max_t(u32, qattr->max_sqe, qattr->max_rqe); 141 attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD | 142 IB_DEVICE_RC_RNR_NAK_GEN | 143 IB_DEVICE_LOCAL_DMA_LKEY | IB_DEVICE_MEM_MGT_EXTENSIONS; 144 145 attr->max_sge = qattr->max_sge; 146 attr->max_sge_rd = qattr->max_sge; 147 attr->max_cq = qattr->max_cq; 148 attr->max_cqe = qattr->max_cqe; 149 attr->max_mr = qattr->max_mr; 150 attr->max_mw = qattr->max_mw; 151 attr->max_pd = qattr->max_pd; 152 attr->atomic_cap = dev->atomic_cap; 153 attr->max_fmr = qattr->max_fmr; 154 attr->max_map_per_fmr = 16; 155 attr->max_qp_init_rd_atom = 156 1 << (fls(qattr->max_qp_req_rd_atomic_resc) - 1); 157 attr->max_qp_rd_atom = 158 min(1 << (fls(qattr->max_qp_resp_rd_atomic_resc) - 1), 159 attr->max_qp_init_rd_atom); 160 161 attr->max_srq = qattr->max_srq; 162 attr->max_srq_sge = qattr->max_srq_sge; 163 attr->max_srq_wr = qattr->max_srq_wr; 164 165 attr->local_ca_ack_delay = qattr->dev_ack_delay; 166 attr->max_fast_reg_page_list_len = qattr->max_mr / 8; 167 attr->max_pkeys = QEDR_ROCE_PKEY_MAX; 168 attr->max_ah = qattr->max_ah; 169 170 return 0; 171 } 172 173 #define QEDR_SPEED_SDR (1) 174 #define QEDR_SPEED_DDR (2) 175 #define QEDR_SPEED_QDR (4) 176 #define QEDR_SPEED_FDR10 (8) 177 #define QEDR_SPEED_FDR (16) 178 #define QEDR_SPEED_EDR (32) 179 180 static inline void get_link_speed_and_width(int speed, u8 *ib_speed, 181 u8 *ib_width) 182 { 183 switch (speed) { 184 case 1000: 185 *ib_speed = QEDR_SPEED_SDR; 186 *ib_width = IB_WIDTH_1X; 187 break; 188 case 10000: 189 *ib_speed = QEDR_SPEED_QDR; 190 *ib_width = IB_WIDTH_1X; 191 break; 192 193 case 20000: 194 *ib_speed = QEDR_SPEED_DDR; 195 *ib_width = IB_WIDTH_4X; 196 break; 197 198 case 25000: 199 *ib_speed = QEDR_SPEED_EDR; 200 *ib_width = IB_WIDTH_1X; 201 break; 202 203 case 40000: 204 *ib_speed = QEDR_SPEED_QDR; 205 *ib_width = IB_WIDTH_4X; 206 break; 207 208 case 50000: 209 *ib_speed = QEDR_SPEED_QDR; 210 *ib_width = IB_WIDTH_4X; 211 break; 212 213 case 100000: 214 *ib_speed = QEDR_SPEED_EDR; 215 *ib_width = IB_WIDTH_4X; 216 break; 217 218 default: 219 /* Unsupported */ 220 *ib_speed = QEDR_SPEED_SDR; 221 *ib_width = IB_WIDTH_1X; 222 } 223 } 224 225 int qedr_query_port(struct ib_device *ibdev, u8 port, struct ib_port_attr *attr) 226 { 227 struct qedr_dev *dev; 228 struct qed_rdma_port *rdma_port; 229 230 dev = get_qedr_dev(ibdev); 231 if (port > 1) { 232 DP_ERR(dev, "invalid_port=0x%x\n", port); 233 return -EINVAL; 234 } 235 236 if (!dev->rdma_ctx) { 237 DP_ERR(dev, "rdma_ctx is NULL\n"); 238 return -EINVAL; 239 } 240 241 rdma_port = dev->ops->rdma_query_port(dev->rdma_ctx); 242 243 /* *attr being zeroed by the caller, avoid zeroing it here */ 244 if (rdma_port->port_state == QED_RDMA_PORT_UP) { 245 attr->state = IB_PORT_ACTIVE; 246 attr->phys_state = 5; 247 } else { 248 attr->state = IB_PORT_DOWN; 249 attr->phys_state = 3; 250 } 251 attr->max_mtu = IB_MTU_4096; 252 attr->active_mtu = iboe_get_mtu(dev->ndev->mtu); 253 attr->lid = 0; 254 attr->lmc = 0; 255 attr->sm_lid = 0; 256 attr->sm_sl = 0; 257 attr->port_cap_flags = IB_PORT_IP_BASED_GIDS; 258 attr->gid_tbl_len = QEDR_MAX_SGID; 259 attr->pkey_tbl_len = QEDR_ROCE_PKEY_TABLE_LEN; 260 attr->bad_pkey_cntr = rdma_port->pkey_bad_counter; 261 attr->qkey_viol_cntr = 0; 262 get_link_speed_and_width(rdma_port->link_speed, 263 &attr->active_speed, &attr->active_width); 264 attr->max_msg_sz = rdma_port->max_msg_size; 265 attr->max_vl_num = 4; 266 267 return 0; 268 } 269 270 int qedr_modify_port(struct ib_device *ibdev, u8 port, int mask, 271 struct ib_port_modify *props) 272 { 273 struct qedr_dev *dev; 274 275 dev = get_qedr_dev(ibdev); 276 if (port > 1) { 277 DP_ERR(dev, "invalid_port=0x%x\n", port); 278 return -EINVAL; 279 } 280 281 return 0; 282 } 283 284 static int qedr_add_mmap(struct qedr_ucontext *uctx, u64 phy_addr, 285 unsigned long len) 286 { 287 struct qedr_mm *mm; 288 289 mm = kzalloc(sizeof(*mm), GFP_KERNEL); 290 if (!mm) 291 return -ENOMEM; 292 293 mm->key.phy_addr = phy_addr; 294 /* This function might be called with a length which is not a multiple 295 * of PAGE_SIZE, while the mapping is PAGE_SIZE grained and the kernel 296 * forces this granularity by increasing the requested size if needed. 297 * When qedr_mmap is called, it will search the list with the updated 298 * length as a key. To prevent search failures, the length is rounded up 299 * in advance to PAGE_SIZE. 300 */ 301 mm->key.len = roundup(len, PAGE_SIZE); 302 INIT_LIST_HEAD(&mm->entry); 303 304 mutex_lock(&uctx->mm_list_lock); 305 list_add(&mm->entry, &uctx->mm_head); 306 mutex_unlock(&uctx->mm_list_lock); 307 308 DP_DEBUG(uctx->dev, QEDR_MSG_MISC, 309 "added (addr=0x%llx,len=0x%lx) for ctx=%p\n", 310 (unsigned long long)mm->key.phy_addr, 311 (unsigned long)mm->key.len, uctx); 312 313 return 0; 314 } 315 316 static bool qedr_search_mmap(struct qedr_ucontext *uctx, u64 phy_addr, 317 unsigned long len) 318 { 319 bool found = false; 320 struct qedr_mm *mm; 321 322 mutex_lock(&uctx->mm_list_lock); 323 list_for_each_entry(mm, &uctx->mm_head, entry) { 324 if (len != mm->key.len || phy_addr != mm->key.phy_addr) 325 continue; 326 327 found = true; 328 break; 329 } 330 mutex_unlock(&uctx->mm_list_lock); 331 DP_DEBUG(uctx->dev, QEDR_MSG_MISC, 332 "searched for (addr=0x%llx,len=0x%lx) for ctx=%p, result=%d\n", 333 mm->key.phy_addr, mm->key.len, uctx, found); 334 335 return found; 336 } 337 338 struct ib_ucontext *qedr_alloc_ucontext(struct ib_device *ibdev, 339 struct ib_udata *udata) 340 { 341 int rc; 342 struct qedr_ucontext *ctx; 343 struct qedr_alloc_ucontext_resp uresp; 344 struct qedr_dev *dev = get_qedr_dev(ibdev); 345 struct qed_rdma_add_user_out_params oparams; 346 347 if (!udata) 348 return ERR_PTR(-EFAULT); 349 350 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 351 if (!ctx) 352 return ERR_PTR(-ENOMEM); 353 354 rc = dev->ops->rdma_add_user(dev->rdma_ctx, &oparams); 355 if (rc) { 356 DP_ERR(dev, 357 "failed to allocate a DPI for a new RoCE application, rc=%d. To overcome this consider to increase the number of DPIs, increase the doorbell BAR size or just close unnecessary RoCE applications. In order to increase the number of DPIs consult the qedr readme\n", 358 rc); 359 goto err; 360 } 361 362 ctx->dpi = oparams.dpi; 363 ctx->dpi_addr = oparams.dpi_addr; 364 ctx->dpi_phys_addr = oparams.dpi_phys_addr; 365 ctx->dpi_size = oparams.dpi_size; 366 INIT_LIST_HEAD(&ctx->mm_head); 367 mutex_init(&ctx->mm_list_lock); 368 369 memset(&uresp, 0, sizeof(uresp)); 370 371 uresp.db_pa = ctx->dpi_phys_addr; 372 uresp.db_size = ctx->dpi_size; 373 uresp.max_send_wr = dev->attr.max_sqe; 374 uresp.max_recv_wr = dev->attr.max_rqe; 375 uresp.max_srq_wr = dev->attr.max_srq_wr; 376 uresp.sges_per_send_wr = QEDR_MAX_SQE_ELEMENTS_PER_SQE; 377 uresp.sges_per_recv_wr = QEDR_MAX_RQE_ELEMENTS_PER_RQE; 378 uresp.sges_per_srq_wr = dev->attr.max_srq_sge; 379 uresp.max_cqes = QEDR_MAX_CQES; 380 381 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp)); 382 if (rc) 383 goto err; 384 385 ctx->dev = dev; 386 387 rc = qedr_add_mmap(ctx, ctx->dpi_phys_addr, ctx->dpi_size); 388 if (rc) 389 goto err; 390 391 DP_DEBUG(dev, QEDR_MSG_INIT, "Allocating user context %p\n", 392 &ctx->ibucontext); 393 return &ctx->ibucontext; 394 395 err: 396 kfree(ctx); 397 return ERR_PTR(rc); 398 } 399 400 int qedr_dealloc_ucontext(struct ib_ucontext *ibctx) 401 { 402 struct qedr_ucontext *uctx = get_qedr_ucontext(ibctx); 403 struct qedr_mm *mm, *tmp; 404 int status = 0; 405 406 DP_DEBUG(uctx->dev, QEDR_MSG_INIT, "Deallocating user context %p\n", 407 uctx); 408 uctx->dev->ops->rdma_remove_user(uctx->dev->rdma_ctx, uctx->dpi); 409 410 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) { 411 DP_DEBUG(uctx->dev, QEDR_MSG_MISC, 412 "deleted (addr=0x%llx,len=0x%lx) for ctx=%p\n", 413 mm->key.phy_addr, mm->key.len, uctx); 414 list_del(&mm->entry); 415 kfree(mm); 416 } 417 418 kfree(uctx); 419 return status; 420 } 421 422 int qedr_mmap(struct ib_ucontext *context, struct vm_area_struct *vma) 423 { 424 struct qedr_ucontext *ucontext = get_qedr_ucontext(context); 425 struct qedr_dev *dev = get_qedr_dev(context->device); 426 unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT; 427 u64 unmapped_db = dev->db_phys_addr; 428 unsigned long len = (vma->vm_end - vma->vm_start); 429 int rc = 0; 430 bool found; 431 432 DP_DEBUG(dev, QEDR_MSG_INIT, 433 "qedr_mmap called vm_page=0x%lx vm_pgoff=0x%lx unmapped_db=0x%llx db_size=%x, len=%lx\n", 434 vm_page, vma->vm_pgoff, unmapped_db, dev->db_size, len); 435 if (vma->vm_start & (PAGE_SIZE - 1)) { 436 DP_ERR(dev, "Vma_start not page aligned = %ld\n", 437 vma->vm_start); 438 return -EINVAL; 439 } 440 441 found = qedr_search_mmap(ucontext, vm_page, len); 442 if (!found) { 443 DP_ERR(dev, "Vma_pgoff not found in mapped array = %ld\n", 444 vma->vm_pgoff); 445 return -EINVAL; 446 } 447 448 DP_DEBUG(dev, QEDR_MSG_INIT, "Mapping doorbell bar\n"); 449 450 if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db + 451 dev->db_size))) { 452 DP_DEBUG(dev, QEDR_MSG_INIT, "Mapping doorbell bar\n"); 453 if (vma->vm_flags & VM_READ) { 454 DP_ERR(dev, "Trying to map doorbell bar for read\n"); 455 return -EPERM; 456 } 457 458 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot); 459 460 rc = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff, 461 PAGE_SIZE, vma->vm_page_prot); 462 } else { 463 DP_DEBUG(dev, QEDR_MSG_INIT, "Mapping chains\n"); 464 rc = remap_pfn_range(vma, vma->vm_start, 465 vma->vm_pgoff, len, vma->vm_page_prot); 466 } 467 DP_DEBUG(dev, QEDR_MSG_INIT, "qedr_mmap return code: %d\n", rc); 468 return rc; 469 } 470 471 struct ib_pd *qedr_alloc_pd(struct ib_device *ibdev, 472 struct ib_ucontext *context, struct ib_udata *udata) 473 { 474 struct qedr_dev *dev = get_qedr_dev(ibdev); 475 struct qedr_pd *pd; 476 u16 pd_id; 477 int rc; 478 479 DP_DEBUG(dev, QEDR_MSG_INIT, "Function called from: %s\n", 480 (udata && context) ? "User Lib" : "Kernel"); 481 482 if (!dev->rdma_ctx) { 483 DP_ERR(dev, "invlaid RDMA context\n"); 484 return ERR_PTR(-EINVAL); 485 } 486 487 pd = kzalloc(sizeof(*pd), GFP_KERNEL); 488 if (!pd) 489 return ERR_PTR(-ENOMEM); 490 491 rc = dev->ops->rdma_alloc_pd(dev->rdma_ctx, &pd_id); 492 if (rc) 493 goto err; 494 495 pd->pd_id = pd_id; 496 497 if (udata && context) { 498 struct qedr_alloc_pd_uresp uresp; 499 500 uresp.pd_id = pd_id; 501 502 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp)); 503 if (rc) { 504 DP_ERR(dev, "copy error pd_id=0x%x.\n", pd_id); 505 dev->ops->rdma_dealloc_pd(dev->rdma_ctx, pd_id); 506 goto err; 507 } 508 509 pd->uctx = get_qedr_ucontext(context); 510 pd->uctx->pd = pd; 511 } 512 513 return &pd->ibpd; 514 515 err: 516 kfree(pd); 517 return ERR_PTR(rc); 518 } 519 520 int qedr_dealloc_pd(struct ib_pd *ibpd) 521 { 522 struct qedr_dev *dev = get_qedr_dev(ibpd->device); 523 struct qedr_pd *pd = get_qedr_pd(ibpd); 524 525 if (!pd) { 526 pr_err("Invalid PD received in dealloc_pd\n"); 527 return -EINVAL; 528 } 529 530 DP_DEBUG(dev, QEDR_MSG_INIT, "Deallocating PD %d\n", pd->pd_id); 531 dev->ops->rdma_dealloc_pd(dev->rdma_ctx, pd->pd_id); 532 533 kfree(pd); 534 535 return 0; 536 } 537 538 static void qedr_free_pbl(struct qedr_dev *dev, 539 struct qedr_pbl_info *pbl_info, struct qedr_pbl *pbl) 540 { 541 struct pci_dev *pdev = dev->pdev; 542 int i; 543 544 for (i = 0; i < pbl_info->num_pbls; i++) { 545 if (!pbl[i].va) 546 continue; 547 dma_free_coherent(&pdev->dev, pbl_info->pbl_size, 548 pbl[i].va, pbl[i].pa); 549 } 550 551 kfree(pbl); 552 } 553 554 #define MIN_FW_PBL_PAGE_SIZE (4 * 1024) 555 #define MAX_FW_PBL_PAGE_SIZE (64 * 1024) 556 557 #define NUM_PBES_ON_PAGE(_page_size) (_page_size / sizeof(u64)) 558 #define MAX_PBES_ON_PAGE NUM_PBES_ON_PAGE(MAX_FW_PBL_PAGE_SIZE) 559 #define MAX_PBES_TWO_LAYER (MAX_PBES_ON_PAGE * MAX_PBES_ON_PAGE) 560 561 static struct qedr_pbl *qedr_alloc_pbl_tbl(struct qedr_dev *dev, 562 struct qedr_pbl_info *pbl_info, 563 gfp_t flags) 564 { 565 struct pci_dev *pdev = dev->pdev; 566 struct qedr_pbl *pbl_table; 567 dma_addr_t *pbl_main_tbl; 568 dma_addr_t pa; 569 void *va; 570 int i; 571 572 pbl_table = kcalloc(pbl_info->num_pbls, sizeof(*pbl_table), flags); 573 if (!pbl_table) 574 return ERR_PTR(-ENOMEM); 575 576 for (i = 0; i < pbl_info->num_pbls; i++) { 577 va = dma_alloc_coherent(&pdev->dev, pbl_info->pbl_size, 578 &pa, flags); 579 if (!va) 580 goto err; 581 582 memset(va, 0, pbl_info->pbl_size); 583 pbl_table[i].va = va; 584 pbl_table[i].pa = pa; 585 } 586 587 /* Two-Layer PBLs, if we have more than one pbl we need to initialize 588 * the first one with physical pointers to all of the rest 589 */ 590 pbl_main_tbl = (dma_addr_t *)pbl_table[0].va; 591 for (i = 0; i < pbl_info->num_pbls - 1; i++) 592 pbl_main_tbl[i] = pbl_table[i + 1].pa; 593 594 return pbl_table; 595 596 err: 597 for (i--; i >= 0; i--) 598 dma_free_coherent(&pdev->dev, pbl_info->pbl_size, 599 pbl_table[i].va, pbl_table[i].pa); 600 601 qedr_free_pbl(dev, pbl_info, pbl_table); 602 603 return ERR_PTR(-ENOMEM); 604 } 605 606 static int qedr_prepare_pbl_tbl(struct qedr_dev *dev, 607 struct qedr_pbl_info *pbl_info, 608 u32 num_pbes, int two_layer_capable) 609 { 610 u32 pbl_capacity; 611 u32 pbl_size; 612 u32 num_pbls; 613 614 if ((num_pbes > MAX_PBES_ON_PAGE) && two_layer_capable) { 615 if (num_pbes > MAX_PBES_TWO_LAYER) { 616 DP_ERR(dev, "prepare pbl table: too many pages %d\n", 617 num_pbes); 618 return -EINVAL; 619 } 620 621 /* calculate required pbl page size */ 622 pbl_size = MIN_FW_PBL_PAGE_SIZE; 623 pbl_capacity = NUM_PBES_ON_PAGE(pbl_size) * 624 NUM_PBES_ON_PAGE(pbl_size); 625 626 while (pbl_capacity < num_pbes) { 627 pbl_size *= 2; 628 pbl_capacity = pbl_size / sizeof(u64); 629 pbl_capacity = pbl_capacity * pbl_capacity; 630 } 631 632 num_pbls = DIV_ROUND_UP(num_pbes, NUM_PBES_ON_PAGE(pbl_size)); 633 num_pbls++; /* One for the layer0 ( points to the pbls) */ 634 pbl_info->two_layered = true; 635 } else { 636 /* One layered PBL */ 637 num_pbls = 1; 638 pbl_size = max_t(u32, MIN_FW_PBL_PAGE_SIZE, 639 roundup_pow_of_two((num_pbes * sizeof(u64)))); 640 pbl_info->two_layered = false; 641 } 642 643 pbl_info->num_pbls = num_pbls; 644 pbl_info->pbl_size = pbl_size; 645 pbl_info->num_pbes = num_pbes; 646 647 DP_DEBUG(dev, QEDR_MSG_MR, 648 "prepare pbl table: num_pbes=%d, num_pbls=%d, pbl_size=%d\n", 649 pbl_info->num_pbes, pbl_info->num_pbls, pbl_info->pbl_size); 650 651 return 0; 652 } 653 654 static void qedr_populate_pbls(struct qedr_dev *dev, struct ib_umem *umem, 655 struct qedr_pbl *pbl, 656 struct qedr_pbl_info *pbl_info) 657 { 658 int shift, pg_cnt, pages, pbe_cnt, total_num_pbes = 0; 659 struct qedr_pbl *pbl_tbl; 660 struct scatterlist *sg; 661 struct regpair *pbe; 662 int entry; 663 u32 addr; 664 665 if (!pbl_info->num_pbes) 666 return; 667 668 /* If we have a two layered pbl, the first pbl points to the rest 669 * of the pbls and the first entry lays on the second pbl in the table 670 */ 671 if (pbl_info->two_layered) 672 pbl_tbl = &pbl[1]; 673 else 674 pbl_tbl = pbl; 675 676 pbe = (struct regpair *)pbl_tbl->va; 677 if (!pbe) { 678 DP_ERR(dev, "cannot populate PBL due to a NULL PBE\n"); 679 return; 680 } 681 682 pbe_cnt = 0; 683 684 shift = umem->page_shift; 685 686 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) { 687 pages = sg_dma_len(sg) >> shift; 688 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) { 689 /* store the page address in pbe */ 690 pbe->lo = cpu_to_le32(sg_dma_address(sg) + 691 (pg_cnt << shift)); 692 addr = upper_32_bits(sg_dma_address(sg) + 693 (pg_cnt << shift)); 694 pbe->hi = cpu_to_le32(addr); 695 pbe_cnt++; 696 total_num_pbes++; 697 pbe++; 698 699 if (total_num_pbes == pbl_info->num_pbes) 700 return; 701 702 /* If the given pbl is full storing the pbes, 703 * move to next pbl. 704 */ 705 if (pbe_cnt == (pbl_info->pbl_size / sizeof(u64))) { 706 pbl_tbl++; 707 pbe = (struct regpair *)pbl_tbl->va; 708 pbe_cnt = 0; 709 } 710 } 711 } 712 } 713 714 static int qedr_copy_cq_uresp(struct qedr_dev *dev, 715 struct qedr_cq *cq, struct ib_udata *udata) 716 { 717 struct qedr_create_cq_uresp uresp; 718 int rc; 719 720 memset(&uresp, 0, sizeof(uresp)); 721 722 uresp.db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT); 723 uresp.icid = cq->icid; 724 725 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp)); 726 if (rc) 727 DP_ERR(dev, "copy error cqid=0x%x.\n", cq->icid); 728 729 return rc; 730 } 731 732 static void consume_cqe(struct qedr_cq *cq) 733 { 734 if (cq->latest_cqe == cq->toggle_cqe) 735 cq->pbl_toggle ^= RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK; 736 737 cq->latest_cqe = qed_chain_consume(&cq->pbl); 738 } 739 740 static inline int qedr_align_cq_entries(int entries) 741 { 742 u64 size, aligned_size; 743 744 /* We allocate an extra entry that we don't report to the FW. */ 745 size = (entries + 1) * QEDR_CQE_SIZE; 746 aligned_size = ALIGN(size, PAGE_SIZE); 747 748 return aligned_size / QEDR_CQE_SIZE; 749 } 750 751 static inline int qedr_init_user_queue(struct ib_ucontext *ib_ctx, 752 struct qedr_dev *dev, 753 struct qedr_userq *q, 754 u64 buf_addr, size_t buf_len, 755 int access, int dmasync) 756 { 757 int page_cnt; 758 int rc; 759 760 q->buf_addr = buf_addr; 761 q->buf_len = buf_len; 762 q->umem = ib_umem_get(ib_ctx, q->buf_addr, q->buf_len, access, dmasync); 763 if (IS_ERR(q->umem)) { 764 DP_ERR(dev, "create user queue: failed ib_umem_get, got %ld\n", 765 PTR_ERR(q->umem)); 766 return PTR_ERR(q->umem); 767 } 768 769 page_cnt = ib_umem_page_count(q->umem); 770 rc = qedr_prepare_pbl_tbl(dev, &q->pbl_info, page_cnt, 0); 771 if (rc) 772 goto err0; 773 774 q->pbl_tbl = qedr_alloc_pbl_tbl(dev, &q->pbl_info, GFP_KERNEL); 775 if (IS_ERR(q->pbl_tbl)) { 776 rc = PTR_ERR(q->pbl_tbl); 777 goto err0; 778 } 779 780 qedr_populate_pbls(dev, q->umem, q->pbl_tbl, &q->pbl_info); 781 782 return 0; 783 784 err0: 785 ib_umem_release(q->umem); 786 787 return rc; 788 } 789 790 static inline void qedr_init_cq_params(struct qedr_cq *cq, 791 struct qedr_ucontext *ctx, 792 struct qedr_dev *dev, int vector, 793 int chain_entries, int page_cnt, 794 u64 pbl_ptr, 795 struct qed_rdma_create_cq_in_params 796 *params) 797 { 798 memset(params, 0, sizeof(*params)); 799 params->cq_handle_hi = upper_32_bits((uintptr_t)cq); 800 params->cq_handle_lo = lower_32_bits((uintptr_t)cq); 801 params->cnq_id = vector; 802 params->cq_size = chain_entries - 1; 803 params->dpi = (ctx) ? ctx->dpi : dev->dpi; 804 params->pbl_num_pages = page_cnt; 805 params->pbl_ptr = pbl_ptr; 806 params->pbl_two_level = 0; 807 } 808 809 static void doorbell_cq(struct qedr_cq *cq, u32 cons, u8 flags) 810 { 811 /* Flush data before signalling doorbell */ 812 wmb(); 813 cq->db.data.agg_flags = flags; 814 cq->db.data.value = cpu_to_le32(cons); 815 writeq(cq->db.raw, cq->db_addr); 816 817 /* Make sure write would stick */ 818 mmiowb(); 819 } 820 821 int qedr_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 822 { 823 struct qedr_cq *cq = get_qedr_cq(ibcq); 824 unsigned long sflags; 825 struct qedr_dev *dev; 826 827 dev = get_qedr_dev(ibcq->device); 828 829 if (cq->destroyed) { 830 DP_ERR(dev, 831 "warning: arm was invoked after destroy for cq %p (icid=%d)\n", 832 cq, cq->icid); 833 return -EINVAL; 834 } 835 836 837 if (cq->cq_type == QEDR_CQ_TYPE_GSI) 838 return 0; 839 840 spin_lock_irqsave(&cq->cq_lock, sflags); 841 842 cq->arm_flags = 0; 843 844 if (flags & IB_CQ_SOLICITED) 845 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_SE_CF_CMD; 846 847 if (flags & IB_CQ_NEXT_COMP) 848 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_CF_CMD; 849 850 doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags); 851 852 spin_unlock_irqrestore(&cq->cq_lock, sflags); 853 854 return 0; 855 } 856 857 struct ib_cq *qedr_create_cq(struct ib_device *ibdev, 858 const struct ib_cq_init_attr *attr, 859 struct ib_ucontext *ib_ctx, struct ib_udata *udata) 860 { 861 struct qedr_ucontext *ctx = get_qedr_ucontext(ib_ctx); 862 struct qed_rdma_destroy_cq_out_params destroy_oparams; 863 struct qed_rdma_destroy_cq_in_params destroy_iparams; 864 struct qedr_dev *dev = get_qedr_dev(ibdev); 865 struct qed_rdma_create_cq_in_params params; 866 struct qedr_create_cq_ureq ureq; 867 int vector = attr->comp_vector; 868 int entries = attr->cqe; 869 struct qedr_cq *cq; 870 int chain_entries; 871 int page_cnt; 872 u64 pbl_ptr; 873 u16 icid; 874 int rc; 875 876 DP_DEBUG(dev, QEDR_MSG_INIT, 877 "create_cq: called from %s. entries=%d, vector=%d\n", 878 udata ? "User Lib" : "Kernel", entries, vector); 879 880 if (entries > QEDR_MAX_CQES) { 881 DP_ERR(dev, 882 "create cq: the number of entries %d is too high. Must be equal or below %d.\n", 883 entries, QEDR_MAX_CQES); 884 return ERR_PTR(-EINVAL); 885 } 886 887 chain_entries = qedr_align_cq_entries(entries); 888 chain_entries = min_t(int, chain_entries, QEDR_MAX_CQES); 889 890 cq = kzalloc(sizeof(*cq), GFP_KERNEL); 891 if (!cq) 892 return ERR_PTR(-ENOMEM); 893 894 if (udata) { 895 memset(&ureq, 0, sizeof(ureq)); 896 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) { 897 DP_ERR(dev, 898 "create cq: problem copying data from user space\n"); 899 goto err0; 900 } 901 902 if (!ureq.len) { 903 DP_ERR(dev, 904 "create cq: cannot create a cq with 0 entries\n"); 905 goto err0; 906 } 907 908 cq->cq_type = QEDR_CQ_TYPE_USER; 909 910 rc = qedr_init_user_queue(ib_ctx, dev, &cq->q, ureq.addr, 911 ureq.len, IB_ACCESS_LOCAL_WRITE, 1); 912 if (rc) 913 goto err0; 914 915 pbl_ptr = cq->q.pbl_tbl->pa; 916 page_cnt = cq->q.pbl_info.num_pbes; 917 918 cq->ibcq.cqe = chain_entries; 919 } else { 920 cq->cq_type = QEDR_CQ_TYPE_KERNEL; 921 922 rc = dev->ops->common->chain_alloc(dev->cdev, 923 QED_CHAIN_USE_TO_CONSUME, 924 QED_CHAIN_MODE_PBL, 925 QED_CHAIN_CNT_TYPE_U32, 926 chain_entries, 927 sizeof(union rdma_cqe), 928 &cq->pbl); 929 if (rc) 930 goto err1; 931 932 page_cnt = qed_chain_get_page_cnt(&cq->pbl); 933 pbl_ptr = qed_chain_get_pbl_phys(&cq->pbl); 934 cq->ibcq.cqe = cq->pbl.capacity; 935 } 936 937 qedr_init_cq_params(cq, ctx, dev, vector, chain_entries, page_cnt, 938 pbl_ptr, ¶ms); 939 940 rc = dev->ops->rdma_create_cq(dev->rdma_ctx, ¶ms, &icid); 941 if (rc) 942 goto err2; 943 944 cq->icid = icid; 945 cq->sig = QEDR_CQ_MAGIC_NUMBER; 946 spin_lock_init(&cq->cq_lock); 947 948 if (ib_ctx) { 949 rc = qedr_copy_cq_uresp(dev, cq, udata); 950 if (rc) 951 goto err3; 952 } else { 953 /* Generate doorbell address. */ 954 cq->db_addr = dev->db_addr + 955 DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT); 956 cq->db.data.icid = cq->icid; 957 cq->db.data.params = DB_AGG_CMD_SET << 958 RDMA_PWM_VAL32_DATA_AGG_CMD_SHIFT; 959 960 /* point to the very last element, passing it we will toggle */ 961 cq->toggle_cqe = qed_chain_get_last_elem(&cq->pbl); 962 cq->pbl_toggle = RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK; 963 cq->latest_cqe = NULL; 964 consume_cqe(cq); 965 cq->cq_cons = qed_chain_get_cons_idx_u32(&cq->pbl); 966 } 967 968 DP_DEBUG(dev, QEDR_MSG_CQ, 969 "create cq: icid=0x%0x, addr=%p, size(entries)=0x%0x\n", 970 cq->icid, cq, params.cq_size); 971 972 return &cq->ibcq; 973 974 err3: 975 destroy_iparams.icid = cq->icid; 976 dev->ops->rdma_destroy_cq(dev->rdma_ctx, &destroy_iparams, 977 &destroy_oparams); 978 err2: 979 if (udata) 980 qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl); 981 else 982 dev->ops->common->chain_free(dev->cdev, &cq->pbl); 983 err1: 984 if (udata) 985 ib_umem_release(cq->q.umem); 986 err0: 987 kfree(cq); 988 return ERR_PTR(-EINVAL); 989 } 990 991 int qedr_resize_cq(struct ib_cq *ibcq, int new_cnt, struct ib_udata *udata) 992 { 993 struct qedr_dev *dev = get_qedr_dev(ibcq->device); 994 struct qedr_cq *cq = get_qedr_cq(ibcq); 995 996 DP_ERR(dev, "cq %p RESIZE NOT SUPPORTED\n", cq); 997 998 return 0; 999 } 1000 1001 #define QEDR_DESTROY_CQ_MAX_ITERATIONS (10) 1002 #define QEDR_DESTROY_CQ_ITER_DURATION (10) 1003 1004 int qedr_destroy_cq(struct ib_cq *ibcq) 1005 { 1006 struct qedr_dev *dev = get_qedr_dev(ibcq->device); 1007 struct qed_rdma_destroy_cq_out_params oparams; 1008 struct qed_rdma_destroy_cq_in_params iparams; 1009 struct qedr_cq *cq = get_qedr_cq(ibcq); 1010 int iter; 1011 int rc; 1012 1013 DP_DEBUG(dev, QEDR_MSG_CQ, "destroy cq %p (icid=%d)\n", cq, cq->icid); 1014 1015 cq->destroyed = 1; 1016 1017 /* GSIs CQs are handled by driver, so they don't exist in the FW */ 1018 if (cq->cq_type == QEDR_CQ_TYPE_GSI) 1019 goto done; 1020 1021 iparams.icid = cq->icid; 1022 rc = dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams); 1023 if (rc) 1024 return rc; 1025 1026 dev->ops->common->chain_free(dev->cdev, &cq->pbl); 1027 1028 if (ibcq->uobject && ibcq->uobject->context) { 1029 qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl); 1030 ib_umem_release(cq->q.umem); 1031 } 1032 1033 /* We don't want the IRQ handler to handle a non-existing CQ so we 1034 * wait until all CNQ interrupts, if any, are received. This will always 1035 * happen and will always happen very fast. If not, then a serious error 1036 * has occured. That is why we can use a long delay. 1037 * We spin for a short time so we don’t lose time on context switching 1038 * in case all the completions are handled in that span. Otherwise 1039 * we sleep for a while and check again. Since the CNQ may be 1040 * associated with (only) the current CPU we use msleep to allow the 1041 * current CPU to be freed. 1042 * The CNQ notification is increased in qedr_irq_handler(). 1043 */ 1044 iter = QEDR_DESTROY_CQ_MAX_ITERATIONS; 1045 while (oparams.num_cq_notif != READ_ONCE(cq->cnq_notif) && iter) { 1046 udelay(QEDR_DESTROY_CQ_ITER_DURATION); 1047 iter--; 1048 } 1049 1050 iter = QEDR_DESTROY_CQ_MAX_ITERATIONS; 1051 while (oparams.num_cq_notif != READ_ONCE(cq->cnq_notif) && iter) { 1052 msleep(QEDR_DESTROY_CQ_ITER_DURATION); 1053 iter--; 1054 } 1055 1056 if (oparams.num_cq_notif != cq->cnq_notif) 1057 goto err; 1058 1059 /* Note that we don't need to have explicit code to wait for the 1060 * completion of the event handler because it is invoked from the EQ. 1061 * Since the destroy CQ ramrod has also been received on the EQ we can 1062 * be certain that there's no event handler in process. 1063 */ 1064 done: 1065 cq->sig = ~cq->sig; 1066 1067 kfree(cq); 1068 1069 return 0; 1070 1071 err: 1072 DP_ERR(dev, 1073 "CQ %p (icid=%d) not freed, expecting %d ints but got %d ints\n", 1074 cq, cq->icid, oparams.num_cq_notif, cq->cnq_notif); 1075 1076 return -EINVAL; 1077 } 1078 1079 static inline int get_gid_info_from_table(struct ib_qp *ibqp, 1080 struct ib_qp_attr *attr, 1081 int attr_mask, 1082 struct qed_rdma_modify_qp_in_params 1083 *qp_params) 1084 { 1085 enum rdma_network_type nw_type; 1086 struct ib_gid_attr gid_attr; 1087 const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr); 1088 union ib_gid gid; 1089 u32 ipv4_addr; 1090 int rc = 0; 1091 int i; 1092 1093 rc = ib_get_cached_gid(ibqp->device, 1094 rdma_ah_get_port_num(&attr->ah_attr), 1095 grh->sgid_index, &gid, &gid_attr); 1096 if (rc) 1097 return rc; 1098 1099 if (!memcmp(&gid, &zgid, sizeof(gid))) 1100 return -ENOENT; 1101 1102 if (gid_attr.ndev) { 1103 qp_params->vlan_id = rdma_vlan_dev_vlan_id(gid_attr.ndev); 1104 1105 dev_put(gid_attr.ndev); 1106 nw_type = ib_gid_to_network_type(gid_attr.gid_type, &gid); 1107 switch (nw_type) { 1108 case RDMA_NETWORK_IPV6: 1109 memcpy(&qp_params->sgid.bytes[0], &gid.raw[0], 1110 sizeof(qp_params->sgid)); 1111 memcpy(&qp_params->dgid.bytes[0], 1112 &grh->dgid, 1113 sizeof(qp_params->dgid)); 1114 qp_params->roce_mode = ROCE_V2_IPV6; 1115 SET_FIELD(qp_params->modify_flags, 1116 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1); 1117 break; 1118 case RDMA_NETWORK_IB: 1119 memcpy(&qp_params->sgid.bytes[0], &gid.raw[0], 1120 sizeof(qp_params->sgid)); 1121 memcpy(&qp_params->dgid.bytes[0], 1122 &grh->dgid, 1123 sizeof(qp_params->dgid)); 1124 qp_params->roce_mode = ROCE_V1; 1125 break; 1126 case RDMA_NETWORK_IPV4: 1127 memset(&qp_params->sgid, 0, sizeof(qp_params->sgid)); 1128 memset(&qp_params->dgid, 0, sizeof(qp_params->dgid)); 1129 ipv4_addr = qedr_get_ipv4_from_gid(gid.raw); 1130 qp_params->sgid.ipv4_addr = ipv4_addr; 1131 ipv4_addr = 1132 qedr_get_ipv4_from_gid(grh->dgid.raw); 1133 qp_params->dgid.ipv4_addr = ipv4_addr; 1134 SET_FIELD(qp_params->modify_flags, 1135 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1); 1136 qp_params->roce_mode = ROCE_V2_IPV4; 1137 break; 1138 } 1139 } 1140 1141 for (i = 0; i < 4; i++) { 1142 qp_params->sgid.dwords[i] = ntohl(qp_params->sgid.dwords[i]); 1143 qp_params->dgid.dwords[i] = ntohl(qp_params->dgid.dwords[i]); 1144 } 1145 1146 if (qp_params->vlan_id >= VLAN_CFI_MASK) 1147 qp_params->vlan_id = 0; 1148 1149 return 0; 1150 } 1151 1152 static int qedr_check_qp_attrs(struct ib_pd *ibpd, struct qedr_dev *dev, 1153 struct ib_qp_init_attr *attrs) 1154 { 1155 struct qedr_device_attr *qattr = &dev->attr; 1156 1157 /* QP0... attrs->qp_type == IB_QPT_GSI */ 1158 if (attrs->qp_type != IB_QPT_RC && attrs->qp_type != IB_QPT_GSI) { 1159 DP_DEBUG(dev, QEDR_MSG_QP, 1160 "create qp: unsupported qp type=0x%x requested\n", 1161 attrs->qp_type); 1162 return -EINVAL; 1163 } 1164 1165 if (attrs->cap.max_send_wr > qattr->max_sqe) { 1166 DP_ERR(dev, 1167 "create qp: cannot create a SQ with %d elements (max_send_wr=0x%x)\n", 1168 attrs->cap.max_send_wr, qattr->max_sqe); 1169 return -EINVAL; 1170 } 1171 1172 if (attrs->cap.max_inline_data > qattr->max_inline) { 1173 DP_ERR(dev, 1174 "create qp: unsupported inline data size=0x%x requested (max_inline=0x%x)\n", 1175 attrs->cap.max_inline_data, qattr->max_inline); 1176 return -EINVAL; 1177 } 1178 1179 if (attrs->cap.max_send_sge > qattr->max_sge) { 1180 DP_ERR(dev, 1181 "create qp: unsupported send_sge=0x%x requested (max_send_sge=0x%x)\n", 1182 attrs->cap.max_send_sge, qattr->max_sge); 1183 return -EINVAL; 1184 } 1185 1186 if (attrs->cap.max_recv_sge > qattr->max_sge) { 1187 DP_ERR(dev, 1188 "create qp: unsupported recv_sge=0x%x requested (max_recv_sge=0x%x)\n", 1189 attrs->cap.max_recv_sge, qattr->max_sge); 1190 return -EINVAL; 1191 } 1192 1193 /* Unprivileged user space cannot create special QP */ 1194 if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) { 1195 DP_ERR(dev, 1196 "create qp: userspace can't create special QPs of type=0x%x\n", 1197 attrs->qp_type); 1198 return -EINVAL; 1199 } 1200 1201 return 0; 1202 } 1203 1204 static void qedr_copy_rq_uresp(struct qedr_create_qp_uresp *uresp, 1205 struct qedr_qp *qp) 1206 { 1207 uresp->rq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD); 1208 uresp->rq_icid = qp->icid; 1209 } 1210 1211 static void qedr_copy_sq_uresp(struct qedr_create_qp_uresp *uresp, 1212 struct qedr_qp *qp) 1213 { 1214 uresp->sq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD); 1215 uresp->sq_icid = qp->icid + 1; 1216 } 1217 1218 static int qedr_copy_qp_uresp(struct qedr_dev *dev, 1219 struct qedr_qp *qp, struct ib_udata *udata) 1220 { 1221 struct qedr_create_qp_uresp uresp; 1222 int rc; 1223 1224 memset(&uresp, 0, sizeof(uresp)); 1225 qedr_copy_sq_uresp(&uresp, qp); 1226 qedr_copy_rq_uresp(&uresp, qp); 1227 1228 uresp.atomic_supported = dev->atomic_cap != IB_ATOMIC_NONE; 1229 uresp.qp_id = qp->qp_id; 1230 1231 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp)); 1232 if (rc) 1233 DP_ERR(dev, 1234 "create qp: failed a copy to user space with qp icid=0x%x.\n", 1235 qp->icid); 1236 1237 return rc; 1238 } 1239 1240 static void qedr_set_common_qp_params(struct qedr_dev *dev, 1241 struct qedr_qp *qp, 1242 struct qedr_pd *pd, 1243 struct ib_qp_init_attr *attrs) 1244 { 1245 spin_lock_init(&qp->q_lock); 1246 qp->pd = pd; 1247 qp->qp_type = attrs->qp_type; 1248 qp->max_inline_data = attrs->cap.max_inline_data; 1249 qp->sq.max_sges = attrs->cap.max_send_sge; 1250 qp->state = QED_ROCE_QP_STATE_RESET; 1251 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false; 1252 qp->sq_cq = get_qedr_cq(attrs->send_cq); 1253 qp->rq_cq = get_qedr_cq(attrs->recv_cq); 1254 qp->dev = dev; 1255 qp->rq.max_sges = attrs->cap.max_recv_sge; 1256 1257 DP_DEBUG(dev, QEDR_MSG_QP, 1258 "RQ params:\trq_max_sges = %d, rq_cq_id = %d\n", 1259 qp->rq.max_sges, qp->rq_cq->icid); 1260 DP_DEBUG(dev, QEDR_MSG_QP, 1261 "QP params:\tpd = %d, qp_type = %d, max_inline_data = %d, state = %d, signaled = %d, use_srq=%d\n", 1262 pd->pd_id, qp->qp_type, qp->max_inline_data, 1263 qp->state, qp->signaled, (attrs->srq) ? 1 : 0); 1264 DP_DEBUG(dev, QEDR_MSG_QP, 1265 "SQ params:\tsq_max_sges = %d, sq_cq_id = %d\n", 1266 qp->sq.max_sges, qp->sq_cq->icid); 1267 } 1268 1269 static void qedr_set_roce_db_info(struct qedr_dev *dev, struct qedr_qp *qp) 1270 { 1271 qp->sq.db = dev->db_addr + 1272 DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD); 1273 qp->sq.db_data.data.icid = qp->icid + 1; 1274 qp->rq.db = dev->db_addr + 1275 DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD); 1276 qp->rq.db_data.data.icid = qp->icid; 1277 } 1278 1279 static inline void 1280 qedr_init_common_qp_in_params(struct qedr_dev *dev, 1281 struct qedr_pd *pd, 1282 struct qedr_qp *qp, 1283 struct ib_qp_init_attr *attrs, 1284 bool fmr_and_reserved_lkey, 1285 struct qed_rdma_create_qp_in_params *params) 1286 { 1287 /* QP handle to be written in an async event */ 1288 params->qp_handle_async_lo = lower_32_bits((uintptr_t) qp); 1289 params->qp_handle_async_hi = upper_32_bits((uintptr_t) qp); 1290 1291 params->signal_all = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR); 1292 params->fmr_and_reserved_lkey = fmr_and_reserved_lkey; 1293 params->pd = pd->pd_id; 1294 params->dpi = pd->uctx ? pd->uctx->dpi : dev->dpi; 1295 params->sq_cq_id = get_qedr_cq(attrs->send_cq)->icid; 1296 params->stats_queue = 0; 1297 params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid; 1298 params->srq_id = 0; 1299 params->use_srq = false; 1300 } 1301 1302 static inline void qedr_qp_user_print(struct qedr_dev *dev, struct qedr_qp *qp) 1303 { 1304 DP_DEBUG(dev, QEDR_MSG_QP, "create qp: successfully created user QP. " 1305 "qp=%p. " 1306 "sq_addr=0x%llx, " 1307 "sq_len=%zd, " 1308 "rq_addr=0x%llx, " 1309 "rq_len=%zd" 1310 "\n", 1311 qp, 1312 qp->usq.buf_addr, 1313 qp->usq.buf_len, qp->urq.buf_addr, qp->urq.buf_len); 1314 } 1315 1316 static void qedr_cleanup_user(struct qedr_dev *dev, struct qedr_qp *qp) 1317 { 1318 if (qp->usq.umem) 1319 ib_umem_release(qp->usq.umem); 1320 qp->usq.umem = NULL; 1321 1322 if (qp->urq.umem) 1323 ib_umem_release(qp->urq.umem); 1324 qp->urq.umem = NULL; 1325 } 1326 1327 static int qedr_create_user_qp(struct qedr_dev *dev, 1328 struct qedr_qp *qp, 1329 struct ib_pd *ibpd, 1330 struct ib_udata *udata, 1331 struct ib_qp_init_attr *attrs) 1332 { 1333 struct qed_rdma_create_qp_in_params in_params; 1334 struct qed_rdma_create_qp_out_params out_params; 1335 struct qedr_pd *pd = get_qedr_pd(ibpd); 1336 struct ib_ucontext *ib_ctx = NULL; 1337 struct qedr_ucontext *ctx = NULL; 1338 struct qedr_create_qp_ureq ureq; 1339 int rc = -EINVAL; 1340 1341 ib_ctx = ibpd->uobject->context; 1342 ctx = get_qedr_ucontext(ib_ctx); 1343 1344 memset(&ureq, 0, sizeof(ureq)); 1345 rc = ib_copy_from_udata(&ureq, udata, sizeof(ureq)); 1346 if (rc) { 1347 DP_ERR(dev, "Problem copying data from user space\n"); 1348 return rc; 1349 } 1350 1351 /* SQ - read access only (0), dma sync not required (0) */ 1352 rc = qedr_init_user_queue(ib_ctx, dev, &qp->usq, ureq.sq_addr, 1353 ureq.sq_len, 0, 0); 1354 if (rc) 1355 return rc; 1356 1357 /* RQ - read access only (0), dma sync not required (0) */ 1358 rc = qedr_init_user_queue(ib_ctx, dev, &qp->urq, ureq.rq_addr, 1359 ureq.rq_len, 0, 0); 1360 1361 if (rc) 1362 return rc; 1363 1364 memset(&in_params, 0, sizeof(in_params)); 1365 qedr_init_common_qp_in_params(dev, pd, qp, attrs, false, &in_params); 1366 in_params.qp_handle_lo = ureq.qp_handle_lo; 1367 in_params.qp_handle_hi = ureq.qp_handle_hi; 1368 in_params.sq_num_pages = qp->usq.pbl_info.num_pbes; 1369 in_params.sq_pbl_ptr = qp->usq.pbl_tbl->pa; 1370 in_params.rq_num_pages = qp->urq.pbl_info.num_pbes; 1371 in_params.rq_pbl_ptr = qp->urq.pbl_tbl->pa; 1372 1373 qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx, 1374 &in_params, &out_params); 1375 1376 if (!qp->qed_qp) { 1377 rc = -ENOMEM; 1378 goto err1; 1379 } 1380 1381 qp->qp_id = out_params.qp_id; 1382 qp->icid = out_params.icid; 1383 1384 rc = qedr_copy_qp_uresp(dev, qp, udata); 1385 if (rc) 1386 goto err; 1387 1388 qedr_qp_user_print(dev, qp); 1389 1390 return 0; 1391 err: 1392 rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp); 1393 if (rc) 1394 DP_ERR(dev, "create qp: fatal fault. rc=%d", rc); 1395 1396 err1: 1397 qedr_cleanup_user(dev, qp); 1398 return rc; 1399 } 1400 1401 static int 1402 qedr_roce_create_kernel_qp(struct qedr_dev *dev, 1403 struct qedr_qp *qp, 1404 struct qed_rdma_create_qp_in_params *in_params, 1405 u32 n_sq_elems, u32 n_rq_elems) 1406 { 1407 struct qed_rdma_create_qp_out_params out_params; 1408 int rc; 1409 1410 rc = dev->ops->common->chain_alloc(dev->cdev, 1411 QED_CHAIN_USE_TO_PRODUCE, 1412 QED_CHAIN_MODE_PBL, 1413 QED_CHAIN_CNT_TYPE_U32, 1414 n_sq_elems, 1415 QEDR_SQE_ELEMENT_SIZE, 1416 &qp->sq.pbl); 1417 1418 if (rc) 1419 return rc; 1420 1421 in_params->sq_num_pages = qed_chain_get_page_cnt(&qp->sq.pbl); 1422 in_params->sq_pbl_ptr = qed_chain_get_pbl_phys(&qp->sq.pbl); 1423 1424 rc = dev->ops->common->chain_alloc(dev->cdev, 1425 QED_CHAIN_USE_TO_CONSUME_PRODUCE, 1426 QED_CHAIN_MODE_PBL, 1427 QED_CHAIN_CNT_TYPE_U32, 1428 n_rq_elems, 1429 QEDR_RQE_ELEMENT_SIZE, 1430 &qp->rq.pbl); 1431 if (rc) 1432 return rc; 1433 1434 in_params->rq_num_pages = qed_chain_get_page_cnt(&qp->rq.pbl); 1435 in_params->rq_pbl_ptr = qed_chain_get_pbl_phys(&qp->rq.pbl); 1436 1437 qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx, 1438 in_params, &out_params); 1439 1440 if (!qp->qed_qp) 1441 return -EINVAL; 1442 1443 qp->qp_id = out_params.qp_id; 1444 qp->icid = out_params.icid; 1445 1446 qedr_set_roce_db_info(dev, qp); 1447 1448 return 0; 1449 } 1450 1451 static void qedr_cleanup_kernel(struct qedr_dev *dev, struct qedr_qp *qp) 1452 { 1453 dev->ops->common->chain_free(dev->cdev, &qp->sq.pbl); 1454 kfree(qp->wqe_wr_id); 1455 1456 dev->ops->common->chain_free(dev->cdev, &qp->rq.pbl); 1457 kfree(qp->rqe_wr_id); 1458 } 1459 1460 static int qedr_create_kernel_qp(struct qedr_dev *dev, 1461 struct qedr_qp *qp, 1462 struct ib_pd *ibpd, 1463 struct ib_qp_init_attr *attrs) 1464 { 1465 struct qed_rdma_create_qp_in_params in_params; 1466 struct qedr_pd *pd = get_qedr_pd(ibpd); 1467 int rc = -EINVAL; 1468 u32 n_rq_elems; 1469 u32 n_sq_elems; 1470 u32 n_sq_entries; 1471 1472 memset(&in_params, 0, sizeof(in_params)); 1473 1474 /* A single work request may take up to QEDR_MAX_SQ_WQE_SIZE elements in 1475 * the ring. The ring should allow at least a single WR, even if the 1476 * user requested none, due to allocation issues. 1477 * We should add an extra WR since the prod and cons indices of 1478 * wqe_wr_id are managed in such a way that the WQ is considered full 1479 * when (prod+1)%max_wr==cons. We currently don't do that because we 1480 * double the number of entries due an iSER issue that pushes far more 1481 * WRs than indicated. If we decline its ib_post_send() then we get 1482 * error prints in the dmesg we'd like to avoid. 1483 */ 1484 qp->sq.max_wr = min_t(u32, attrs->cap.max_send_wr * dev->wq_multiplier, 1485 dev->attr.max_sqe); 1486 1487 qp->wqe_wr_id = kzalloc(qp->sq.max_wr * sizeof(*qp->wqe_wr_id), 1488 GFP_KERNEL); 1489 if (!qp->wqe_wr_id) { 1490 DP_ERR(dev, "create qp: failed SQ shadow memory allocation\n"); 1491 return -ENOMEM; 1492 } 1493 1494 /* QP handle to be written in CQE */ 1495 in_params.qp_handle_lo = lower_32_bits((uintptr_t) qp); 1496 in_params.qp_handle_hi = upper_32_bits((uintptr_t) qp); 1497 1498 /* A single work request may take up to QEDR_MAX_RQ_WQE_SIZE elements in 1499 * the ring. There ring should allow at least a single WR, even if the 1500 * user requested none, due to allocation issues. 1501 */ 1502 qp->rq.max_wr = (u16) max_t(u32, attrs->cap.max_recv_wr, 1); 1503 1504 /* Allocate driver internal RQ array */ 1505 qp->rqe_wr_id = kzalloc(qp->rq.max_wr * sizeof(*qp->rqe_wr_id), 1506 GFP_KERNEL); 1507 if (!qp->rqe_wr_id) { 1508 DP_ERR(dev, 1509 "create qp: failed RQ shadow memory allocation\n"); 1510 kfree(qp->wqe_wr_id); 1511 return -ENOMEM; 1512 } 1513 1514 qedr_init_common_qp_in_params(dev, pd, qp, attrs, true, &in_params); 1515 1516 n_sq_entries = attrs->cap.max_send_wr; 1517 n_sq_entries = min_t(u32, n_sq_entries, dev->attr.max_sqe); 1518 n_sq_entries = max_t(u32, n_sq_entries, 1); 1519 n_sq_elems = n_sq_entries * QEDR_MAX_SQE_ELEMENTS_PER_SQE; 1520 1521 n_rq_elems = qp->rq.max_wr * QEDR_MAX_RQE_ELEMENTS_PER_RQE; 1522 1523 rc = qedr_roce_create_kernel_qp(dev, qp, &in_params, 1524 n_sq_elems, n_rq_elems); 1525 if (rc) 1526 qedr_cleanup_kernel(dev, qp); 1527 1528 return rc; 1529 } 1530 1531 struct ib_qp *qedr_create_qp(struct ib_pd *ibpd, 1532 struct ib_qp_init_attr *attrs, 1533 struct ib_udata *udata) 1534 { 1535 struct qedr_dev *dev = get_qedr_dev(ibpd->device); 1536 struct qedr_pd *pd = get_qedr_pd(ibpd); 1537 struct qedr_qp *qp; 1538 struct ib_qp *ibqp; 1539 int rc = 0; 1540 1541 DP_DEBUG(dev, QEDR_MSG_QP, "create qp: called from %s, pd=%p\n", 1542 udata ? "user library" : "kernel", pd); 1543 1544 rc = qedr_check_qp_attrs(ibpd, dev, attrs); 1545 if (rc) 1546 return ERR_PTR(rc); 1547 1548 if (attrs->srq) 1549 return ERR_PTR(-EINVAL); 1550 1551 DP_DEBUG(dev, QEDR_MSG_QP, 1552 "create qp: called from %s, event_handler=%p, eepd=%p sq_cq=%p, sq_icid=%d, rq_cq=%p, rq_icid=%d\n", 1553 udata ? "user library" : "kernel", attrs->event_handler, pd, 1554 get_qedr_cq(attrs->send_cq), 1555 get_qedr_cq(attrs->send_cq)->icid, 1556 get_qedr_cq(attrs->recv_cq), 1557 get_qedr_cq(attrs->recv_cq)->icid); 1558 1559 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 1560 if (!qp) { 1561 DP_ERR(dev, "create qp: failed allocating memory\n"); 1562 return ERR_PTR(-ENOMEM); 1563 } 1564 1565 qedr_set_common_qp_params(dev, qp, pd, attrs); 1566 1567 if (attrs->qp_type == IB_QPT_GSI) { 1568 ibqp = qedr_create_gsi_qp(dev, attrs, qp); 1569 if (IS_ERR(ibqp)) 1570 kfree(qp); 1571 return ibqp; 1572 } 1573 1574 if (udata) 1575 rc = qedr_create_user_qp(dev, qp, ibpd, udata, attrs); 1576 else 1577 rc = qedr_create_kernel_qp(dev, qp, ibpd, attrs); 1578 1579 if (rc) 1580 goto err; 1581 1582 qp->ibqp.qp_num = qp->qp_id; 1583 1584 return &qp->ibqp; 1585 1586 err: 1587 kfree(qp); 1588 1589 return ERR_PTR(-EFAULT); 1590 } 1591 1592 static enum ib_qp_state qedr_get_ibqp_state(enum qed_roce_qp_state qp_state) 1593 { 1594 switch (qp_state) { 1595 case QED_ROCE_QP_STATE_RESET: 1596 return IB_QPS_RESET; 1597 case QED_ROCE_QP_STATE_INIT: 1598 return IB_QPS_INIT; 1599 case QED_ROCE_QP_STATE_RTR: 1600 return IB_QPS_RTR; 1601 case QED_ROCE_QP_STATE_RTS: 1602 return IB_QPS_RTS; 1603 case QED_ROCE_QP_STATE_SQD: 1604 return IB_QPS_SQD; 1605 case QED_ROCE_QP_STATE_ERR: 1606 return IB_QPS_ERR; 1607 case QED_ROCE_QP_STATE_SQE: 1608 return IB_QPS_SQE; 1609 } 1610 return IB_QPS_ERR; 1611 } 1612 1613 static enum qed_roce_qp_state qedr_get_state_from_ibqp( 1614 enum ib_qp_state qp_state) 1615 { 1616 switch (qp_state) { 1617 case IB_QPS_RESET: 1618 return QED_ROCE_QP_STATE_RESET; 1619 case IB_QPS_INIT: 1620 return QED_ROCE_QP_STATE_INIT; 1621 case IB_QPS_RTR: 1622 return QED_ROCE_QP_STATE_RTR; 1623 case IB_QPS_RTS: 1624 return QED_ROCE_QP_STATE_RTS; 1625 case IB_QPS_SQD: 1626 return QED_ROCE_QP_STATE_SQD; 1627 case IB_QPS_ERR: 1628 return QED_ROCE_QP_STATE_ERR; 1629 default: 1630 return QED_ROCE_QP_STATE_ERR; 1631 } 1632 } 1633 1634 static void qedr_reset_qp_hwq_info(struct qedr_qp_hwq_info *qph) 1635 { 1636 qed_chain_reset(&qph->pbl); 1637 qph->prod = 0; 1638 qph->cons = 0; 1639 qph->wqe_cons = 0; 1640 qph->db_data.data.value = cpu_to_le16(0); 1641 } 1642 1643 static int qedr_update_qp_state(struct qedr_dev *dev, 1644 struct qedr_qp *qp, 1645 enum qed_roce_qp_state new_state) 1646 { 1647 int status = 0; 1648 1649 if (new_state == qp->state) 1650 return 0; 1651 1652 switch (qp->state) { 1653 case QED_ROCE_QP_STATE_RESET: 1654 switch (new_state) { 1655 case QED_ROCE_QP_STATE_INIT: 1656 qp->prev_wqe_size = 0; 1657 qedr_reset_qp_hwq_info(&qp->sq); 1658 qedr_reset_qp_hwq_info(&qp->rq); 1659 break; 1660 default: 1661 status = -EINVAL; 1662 break; 1663 }; 1664 break; 1665 case QED_ROCE_QP_STATE_INIT: 1666 switch (new_state) { 1667 case QED_ROCE_QP_STATE_RTR: 1668 /* Update doorbell (in case post_recv was 1669 * done before move to RTR) 1670 */ 1671 wmb(); 1672 writel(qp->rq.db_data.raw, qp->rq.db); 1673 /* Make sure write takes effect */ 1674 mmiowb(); 1675 break; 1676 case QED_ROCE_QP_STATE_ERR: 1677 break; 1678 default: 1679 /* Invalid state change. */ 1680 status = -EINVAL; 1681 break; 1682 }; 1683 break; 1684 case QED_ROCE_QP_STATE_RTR: 1685 /* RTR->XXX */ 1686 switch (new_state) { 1687 case QED_ROCE_QP_STATE_RTS: 1688 break; 1689 case QED_ROCE_QP_STATE_ERR: 1690 break; 1691 default: 1692 /* Invalid state change. */ 1693 status = -EINVAL; 1694 break; 1695 }; 1696 break; 1697 case QED_ROCE_QP_STATE_RTS: 1698 /* RTS->XXX */ 1699 switch (new_state) { 1700 case QED_ROCE_QP_STATE_SQD: 1701 break; 1702 case QED_ROCE_QP_STATE_ERR: 1703 break; 1704 default: 1705 /* Invalid state change. */ 1706 status = -EINVAL; 1707 break; 1708 }; 1709 break; 1710 case QED_ROCE_QP_STATE_SQD: 1711 /* SQD->XXX */ 1712 switch (new_state) { 1713 case QED_ROCE_QP_STATE_RTS: 1714 case QED_ROCE_QP_STATE_ERR: 1715 break; 1716 default: 1717 /* Invalid state change. */ 1718 status = -EINVAL; 1719 break; 1720 }; 1721 break; 1722 case QED_ROCE_QP_STATE_ERR: 1723 /* ERR->XXX */ 1724 switch (new_state) { 1725 case QED_ROCE_QP_STATE_RESET: 1726 if ((qp->rq.prod != qp->rq.cons) || 1727 (qp->sq.prod != qp->sq.cons)) { 1728 DP_NOTICE(dev, 1729 "Error->Reset with rq/sq not empty rq.prod=%x rq.cons=%x sq.prod=%x sq.cons=%x\n", 1730 qp->rq.prod, qp->rq.cons, qp->sq.prod, 1731 qp->sq.cons); 1732 status = -EINVAL; 1733 } 1734 break; 1735 default: 1736 status = -EINVAL; 1737 break; 1738 }; 1739 break; 1740 default: 1741 status = -EINVAL; 1742 break; 1743 }; 1744 1745 return status; 1746 } 1747 1748 int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 1749 int attr_mask, struct ib_udata *udata) 1750 { 1751 struct qedr_qp *qp = get_qedr_qp(ibqp); 1752 struct qed_rdma_modify_qp_in_params qp_params = { 0 }; 1753 struct qedr_dev *dev = get_qedr_dev(&qp->dev->ibdev); 1754 const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr); 1755 enum ib_qp_state old_qp_state, new_qp_state; 1756 int rc = 0; 1757 1758 DP_DEBUG(dev, QEDR_MSG_QP, 1759 "modify qp: qp %p attr_mask=0x%x, state=%d", qp, attr_mask, 1760 attr->qp_state); 1761 1762 old_qp_state = qedr_get_ibqp_state(qp->state); 1763 if (attr_mask & IB_QP_STATE) 1764 new_qp_state = attr->qp_state; 1765 else 1766 new_qp_state = old_qp_state; 1767 1768 if (!ib_modify_qp_is_ok 1769 (old_qp_state, new_qp_state, ibqp->qp_type, attr_mask, 1770 IB_LINK_LAYER_ETHERNET)) { 1771 DP_ERR(dev, 1772 "modify qp: invalid attribute mask=0x%x specified for\n" 1773 "qpn=0x%x of type=0x%x old_qp_state=0x%x, new_qp_state=0x%x\n", 1774 attr_mask, qp->qp_id, ibqp->qp_type, old_qp_state, 1775 new_qp_state); 1776 rc = -EINVAL; 1777 goto err; 1778 } 1779 1780 /* Translate the masks... */ 1781 if (attr_mask & IB_QP_STATE) { 1782 SET_FIELD(qp_params.modify_flags, 1783 QED_RDMA_MODIFY_QP_VALID_NEW_STATE, 1); 1784 qp_params.new_state = qedr_get_state_from_ibqp(attr->qp_state); 1785 } 1786 1787 if (attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) 1788 qp_params.sqd_async = true; 1789 1790 if (attr_mask & IB_QP_PKEY_INDEX) { 1791 SET_FIELD(qp_params.modify_flags, 1792 QED_ROCE_MODIFY_QP_VALID_PKEY, 1); 1793 if (attr->pkey_index >= QEDR_ROCE_PKEY_TABLE_LEN) { 1794 rc = -EINVAL; 1795 goto err; 1796 } 1797 1798 qp_params.pkey = QEDR_ROCE_PKEY_DEFAULT; 1799 } 1800 1801 if (attr_mask & IB_QP_QKEY) 1802 qp->qkey = attr->qkey; 1803 1804 if (attr_mask & IB_QP_ACCESS_FLAGS) { 1805 SET_FIELD(qp_params.modify_flags, 1806 QED_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN, 1); 1807 qp_params.incoming_rdma_read_en = attr->qp_access_flags & 1808 IB_ACCESS_REMOTE_READ; 1809 qp_params.incoming_rdma_write_en = attr->qp_access_flags & 1810 IB_ACCESS_REMOTE_WRITE; 1811 qp_params.incoming_atomic_en = attr->qp_access_flags & 1812 IB_ACCESS_REMOTE_ATOMIC; 1813 } 1814 1815 if (attr_mask & (IB_QP_AV | IB_QP_PATH_MTU)) { 1816 if (attr_mask & IB_QP_PATH_MTU) { 1817 if (attr->path_mtu < IB_MTU_256 || 1818 attr->path_mtu > IB_MTU_4096) { 1819 pr_err("error: Only MTU sizes of 256, 512, 1024, 2048 and 4096 are supported by RoCE\n"); 1820 rc = -EINVAL; 1821 goto err; 1822 } 1823 qp->mtu = min(ib_mtu_enum_to_int(attr->path_mtu), 1824 ib_mtu_enum_to_int(iboe_get_mtu 1825 (dev->ndev->mtu))); 1826 } 1827 1828 if (!qp->mtu) { 1829 qp->mtu = 1830 ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu)); 1831 pr_err("Fixing zeroed MTU to qp->mtu = %d\n", qp->mtu); 1832 } 1833 1834 SET_FIELD(qp_params.modify_flags, 1835 QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR, 1); 1836 1837 qp_params.traffic_class_tos = grh->traffic_class; 1838 qp_params.flow_label = grh->flow_label; 1839 qp_params.hop_limit_ttl = grh->hop_limit; 1840 1841 qp->sgid_idx = grh->sgid_index; 1842 1843 rc = get_gid_info_from_table(ibqp, attr, attr_mask, &qp_params); 1844 if (rc) { 1845 DP_ERR(dev, 1846 "modify qp: problems with GID index %d (rc=%d)\n", 1847 grh->sgid_index, rc); 1848 return rc; 1849 } 1850 1851 rc = qedr_get_dmac(dev, &attr->ah_attr, 1852 qp_params.remote_mac_addr); 1853 if (rc) 1854 return rc; 1855 1856 qp_params.use_local_mac = true; 1857 ether_addr_copy(qp_params.local_mac_addr, dev->ndev->dev_addr); 1858 1859 DP_DEBUG(dev, QEDR_MSG_QP, "dgid=%x:%x:%x:%x\n", 1860 qp_params.dgid.dwords[0], qp_params.dgid.dwords[1], 1861 qp_params.dgid.dwords[2], qp_params.dgid.dwords[3]); 1862 DP_DEBUG(dev, QEDR_MSG_QP, "sgid=%x:%x:%x:%x\n", 1863 qp_params.sgid.dwords[0], qp_params.sgid.dwords[1], 1864 qp_params.sgid.dwords[2], qp_params.sgid.dwords[3]); 1865 DP_DEBUG(dev, QEDR_MSG_QP, "remote_mac=[%pM]\n", 1866 qp_params.remote_mac_addr); 1867 1868 qp_params.mtu = qp->mtu; 1869 qp_params.lb_indication = false; 1870 } 1871 1872 if (!qp_params.mtu) { 1873 /* Stay with current MTU */ 1874 if (qp->mtu) 1875 qp_params.mtu = qp->mtu; 1876 else 1877 qp_params.mtu = 1878 ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu)); 1879 } 1880 1881 if (attr_mask & IB_QP_TIMEOUT) { 1882 SET_FIELD(qp_params.modify_flags, 1883 QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1); 1884 1885 qp_params.ack_timeout = attr->timeout; 1886 if (attr->timeout) { 1887 u32 temp; 1888 1889 temp = 4096 * (1UL << attr->timeout) / 1000 / 1000; 1890 /* FW requires [msec] */ 1891 qp_params.ack_timeout = temp; 1892 } else { 1893 /* Infinite */ 1894 qp_params.ack_timeout = 0; 1895 } 1896 } 1897 if (attr_mask & IB_QP_RETRY_CNT) { 1898 SET_FIELD(qp_params.modify_flags, 1899 QED_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1); 1900 qp_params.retry_cnt = attr->retry_cnt; 1901 } 1902 1903 if (attr_mask & IB_QP_RNR_RETRY) { 1904 SET_FIELD(qp_params.modify_flags, 1905 QED_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT, 1); 1906 qp_params.rnr_retry_cnt = attr->rnr_retry; 1907 } 1908 1909 if (attr_mask & IB_QP_RQ_PSN) { 1910 SET_FIELD(qp_params.modify_flags, 1911 QED_ROCE_MODIFY_QP_VALID_RQ_PSN, 1); 1912 qp_params.rq_psn = attr->rq_psn; 1913 qp->rq_psn = attr->rq_psn; 1914 } 1915 1916 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 1917 if (attr->max_rd_atomic > dev->attr.max_qp_req_rd_atomic_resc) { 1918 rc = -EINVAL; 1919 DP_ERR(dev, 1920 "unsupported max_rd_atomic=%d, supported=%d\n", 1921 attr->max_rd_atomic, 1922 dev->attr.max_qp_req_rd_atomic_resc); 1923 goto err; 1924 } 1925 1926 SET_FIELD(qp_params.modify_flags, 1927 QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ, 1); 1928 qp_params.max_rd_atomic_req = attr->max_rd_atomic; 1929 } 1930 1931 if (attr_mask & IB_QP_MIN_RNR_TIMER) { 1932 SET_FIELD(qp_params.modify_flags, 1933 QED_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER, 1); 1934 qp_params.min_rnr_nak_timer = attr->min_rnr_timer; 1935 } 1936 1937 if (attr_mask & IB_QP_SQ_PSN) { 1938 SET_FIELD(qp_params.modify_flags, 1939 QED_ROCE_MODIFY_QP_VALID_SQ_PSN, 1); 1940 qp_params.sq_psn = attr->sq_psn; 1941 qp->sq_psn = attr->sq_psn; 1942 } 1943 1944 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 1945 if (attr->max_dest_rd_atomic > 1946 dev->attr.max_qp_resp_rd_atomic_resc) { 1947 DP_ERR(dev, 1948 "unsupported max_dest_rd_atomic=%d, supported=%d\n", 1949 attr->max_dest_rd_atomic, 1950 dev->attr.max_qp_resp_rd_atomic_resc); 1951 1952 rc = -EINVAL; 1953 goto err; 1954 } 1955 1956 SET_FIELD(qp_params.modify_flags, 1957 QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP, 1); 1958 qp_params.max_rd_atomic_resp = attr->max_dest_rd_atomic; 1959 } 1960 1961 if (attr_mask & IB_QP_DEST_QPN) { 1962 SET_FIELD(qp_params.modify_flags, 1963 QED_ROCE_MODIFY_QP_VALID_DEST_QP, 1); 1964 1965 qp_params.dest_qp = attr->dest_qp_num; 1966 qp->dest_qp_num = attr->dest_qp_num; 1967 } 1968 1969 if (qp->qp_type != IB_QPT_GSI) 1970 rc = dev->ops->rdma_modify_qp(dev->rdma_ctx, 1971 qp->qed_qp, &qp_params); 1972 1973 if (attr_mask & IB_QP_STATE) { 1974 if ((qp->qp_type != IB_QPT_GSI) && (!udata)) 1975 rc = qedr_update_qp_state(dev, qp, qp_params.new_state); 1976 qp->state = qp_params.new_state; 1977 } 1978 1979 err: 1980 return rc; 1981 } 1982 1983 static int qedr_to_ib_qp_acc_flags(struct qed_rdma_query_qp_out_params *params) 1984 { 1985 int ib_qp_acc_flags = 0; 1986 1987 if (params->incoming_rdma_write_en) 1988 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE; 1989 if (params->incoming_rdma_read_en) 1990 ib_qp_acc_flags |= IB_ACCESS_REMOTE_READ; 1991 if (params->incoming_atomic_en) 1992 ib_qp_acc_flags |= IB_ACCESS_REMOTE_ATOMIC; 1993 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE; 1994 return ib_qp_acc_flags; 1995 } 1996 1997 int qedr_query_qp(struct ib_qp *ibqp, 1998 struct ib_qp_attr *qp_attr, 1999 int attr_mask, struct ib_qp_init_attr *qp_init_attr) 2000 { 2001 struct qed_rdma_query_qp_out_params params; 2002 struct qedr_qp *qp = get_qedr_qp(ibqp); 2003 struct qedr_dev *dev = qp->dev; 2004 int rc = 0; 2005 2006 memset(¶ms, 0, sizeof(params)); 2007 2008 rc = dev->ops->rdma_query_qp(dev->rdma_ctx, qp->qed_qp, ¶ms); 2009 if (rc) 2010 goto err; 2011 2012 memset(qp_attr, 0, sizeof(*qp_attr)); 2013 memset(qp_init_attr, 0, sizeof(*qp_init_attr)); 2014 2015 qp_attr->qp_state = qedr_get_ibqp_state(params.state); 2016 qp_attr->cur_qp_state = qedr_get_ibqp_state(params.state); 2017 qp_attr->path_mtu = ib_mtu_int_to_enum(params.mtu); 2018 qp_attr->path_mig_state = IB_MIG_MIGRATED; 2019 qp_attr->rq_psn = params.rq_psn; 2020 qp_attr->sq_psn = params.sq_psn; 2021 qp_attr->dest_qp_num = params.dest_qp; 2022 2023 qp_attr->qp_access_flags = qedr_to_ib_qp_acc_flags(¶ms); 2024 2025 qp_attr->cap.max_send_wr = qp->sq.max_wr; 2026 qp_attr->cap.max_recv_wr = qp->rq.max_wr; 2027 qp_attr->cap.max_send_sge = qp->sq.max_sges; 2028 qp_attr->cap.max_recv_sge = qp->rq.max_sges; 2029 qp_attr->cap.max_inline_data = ROCE_REQ_MAX_INLINE_DATA_SIZE; 2030 qp_init_attr->cap = qp_attr->cap; 2031 2032 qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE; 2033 rdma_ah_set_grh(&qp_attr->ah_attr, NULL, 2034 params.flow_label, qp->sgid_idx, 2035 params.hop_limit_ttl, params.traffic_class_tos); 2036 rdma_ah_set_dgid_raw(&qp_attr->ah_attr, ¶ms.dgid.bytes[0]); 2037 rdma_ah_set_port_num(&qp_attr->ah_attr, 1); 2038 rdma_ah_set_sl(&qp_attr->ah_attr, 0); 2039 qp_attr->timeout = params.timeout; 2040 qp_attr->rnr_retry = params.rnr_retry; 2041 qp_attr->retry_cnt = params.retry_cnt; 2042 qp_attr->min_rnr_timer = params.min_rnr_nak_timer; 2043 qp_attr->pkey_index = params.pkey_index; 2044 qp_attr->port_num = 1; 2045 rdma_ah_set_path_bits(&qp_attr->ah_attr, 0); 2046 rdma_ah_set_static_rate(&qp_attr->ah_attr, 0); 2047 qp_attr->alt_pkey_index = 0; 2048 qp_attr->alt_port_num = 0; 2049 qp_attr->alt_timeout = 0; 2050 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr)); 2051 2052 qp_attr->sq_draining = (params.state == QED_ROCE_QP_STATE_SQD) ? 1 : 0; 2053 qp_attr->max_dest_rd_atomic = params.max_dest_rd_atomic; 2054 qp_attr->max_rd_atomic = params.max_rd_atomic; 2055 qp_attr->en_sqd_async_notify = (params.sqd_async) ? 1 : 0; 2056 2057 DP_DEBUG(dev, QEDR_MSG_QP, "QEDR_QUERY_QP: max_inline_data=%d\n", 2058 qp_attr->cap.max_inline_data); 2059 2060 err: 2061 return rc; 2062 } 2063 2064 int qedr_free_qp_resources(struct qedr_dev *dev, struct qedr_qp *qp) 2065 { 2066 int rc = 0; 2067 2068 if (qp->qp_type != IB_QPT_GSI) { 2069 rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp); 2070 if (rc) 2071 return rc; 2072 } 2073 2074 if (qp->ibqp.uobject && qp->ibqp.uobject->context) 2075 qedr_cleanup_user(dev, qp); 2076 else 2077 qedr_cleanup_kernel(dev, qp); 2078 2079 return 0; 2080 } 2081 2082 int qedr_destroy_qp(struct ib_qp *ibqp) 2083 { 2084 struct qedr_qp *qp = get_qedr_qp(ibqp); 2085 struct qedr_dev *dev = qp->dev; 2086 struct ib_qp_attr attr; 2087 int attr_mask = 0; 2088 int rc = 0; 2089 2090 DP_DEBUG(dev, QEDR_MSG_QP, "destroy qp: destroying %p, qp type=%d\n", 2091 qp, qp->qp_type); 2092 2093 if ((qp->state != QED_ROCE_QP_STATE_RESET) && 2094 (qp->state != QED_ROCE_QP_STATE_ERR) && 2095 (qp->state != QED_ROCE_QP_STATE_INIT)) { 2096 2097 attr.qp_state = IB_QPS_ERR; 2098 attr_mask |= IB_QP_STATE; 2099 2100 /* Change the QP state to ERROR */ 2101 qedr_modify_qp(ibqp, &attr, attr_mask, NULL); 2102 } 2103 2104 if (qp->qp_type == IB_QPT_GSI) 2105 qedr_destroy_gsi_qp(dev); 2106 2107 qedr_free_qp_resources(dev, qp); 2108 2109 kfree(qp); 2110 2111 return rc; 2112 } 2113 2114 struct ib_ah *qedr_create_ah(struct ib_pd *ibpd, struct rdma_ah_attr *attr, 2115 struct ib_udata *udata) 2116 { 2117 struct qedr_ah *ah; 2118 2119 ah = kzalloc(sizeof(*ah), GFP_ATOMIC); 2120 if (!ah) 2121 return ERR_PTR(-ENOMEM); 2122 2123 ah->attr = *attr; 2124 2125 return &ah->ibah; 2126 } 2127 2128 int qedr_destroy_ah(struct ib_ah *ibah) 2129 { 2130 struct qedr_ah *ah = get_qedr_ah(ibah); 2131 2132 kfree(ah); 2133 return 0; 2134 } 2135 2136 static void free_mr_info(struct qedr_dev *dev, struct mr_info *info) 2137 { 2138 struct qedr_pbl *pbl, *tmp; 2139 2140 if (info->pbl_table) 2141 list_add_tail(&info->pbl_table->list_entry, 2142 &info->free_pbl_list); 2143 2144 if (!list_empty(&info->inuse_pbl_list)) 2145 list_splice(&info->inuse_pbl_list, &info->free_pbl_list); 2146 2147 list_for_each_entry_safe(pbl, tmp, &info->free_pbl_list, list_entry) { 2148 list_del(&pbl->list_entry); 2149 qedr_free_pbl(dev, &info->pbl_info, pbl); 2150 } 2151 } 2152 2153 static int init_mr_info(struct qedr_dev *dev, struct mr_info *info, 2154 size_t page_list_len, bool two_layered) 2155 { 2156 struct qedr_pbl *tmp; 2157 int rc; 2158 2159 INIT_LIST_HEAD(&info->free_pbl_list); 2160 INIT_LIST_HEAD(&info->inuse_pbl_list); 2161 2162 rc = qedr_prepare_pbl_tbl(dev, &info->pbl_info, 2163 page_list_len, two_layered); 2164 if (rc) 2165 goto done; 2166 2167 info->pbl_table = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL); 2168 if (IS_ERR(info->pbl_table)) { 2169 rc = PTR_ERR(info->pbl_table); 2170 goto done; 2171 } 2172 2173 DP_DEBUG(dev, QEDR_MSG_MR, "pbl_table_pa = %pa\n", 2174 &info->pbl_table->pa); 2175 2176 /* in usual case we use 2 PBLs, so we add one to free 2177 * list and allocating another one 2178 */ 2179 tmp = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL); 2180 if (IS_ERR(tmp)) { 2181 DP_DEBUG(dev, QEDR_MSG_MR, "Extra PBL is not allocated\n"); 2182 goto done; 2183 } 2184 2185 list_add_tail(&tmp->list_entry, &info->free_pbl_list); 2186 2187 DP_DEBUG(dev, QEDR_MSG_MR, "extra pbl_table_pa = %pa\n", &tmp->pa); 2188 2189 done: 2190 if (rc) 2191 free_mr_info(dev, info); 2192 2193 return rc; 2194 } 2195 2196 struct ib_mr *qedr_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len, 2197 u64 usr_addr, int acc, struct ib_udata *udata) 2198 { 2199 struct qedr_dev *dev = get_qedr_dev(ibpd->device); 2200 struct qedr_mr *mr; 2201 struct qedr_pd *pd; 2202 int rc = -ENOMEM; 2203 2204 pd = get_qedr_pd(ibpd); 2205 DP_DEBUG(dev, QEDR_MSG_MR, 2206 "qedr_register user mr pd = %d start = %lld, len = %lld, usr_addr = %lld, acc = %d\n", 2207 pd->pd_id, start, len, usr_addr, acc); 2208 2209 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) 2210 return ERR_PTR(-EINVAL); 2211 2212 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 2213 if (!mr) 2214 return ERR_PTR(rc); 2215 2216 mr->type = QEDR_MR_USER; 2217 2218 mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0); 2219 if (IS_ERR(mr->umem)) { 2220 rc = -EFAULT; 2221 goto err0; 2222 } 2223 2224 rc = init_mr_info(dev, &mr->info, ib_umem_page_count(mr->umem), 1); 2225 if (rc) 2226 goto err1; 2227 2228 qedr_populate_pbls(dev, mr->umem, mr->info.pbl_table, 2229 &mr->info.pbl_info); 2230 2231 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid); 2232 if (rc) { 2233 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc); 2234 goto err1; 2235 } 2236 2237 /* Index only, 18 bit long, lkey = itid << 8 | key */ 2238 mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR; 2239 mr->hw_mr.key = 0; 2240 mr->hw_mr.pd = pd->pd_id; 2241 mr->hw_mr.local_read = 1; 2242 mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0; 2243 mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0; 2244 mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0; 2245 mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0; 2246 mr->hw_mr.mw_bind = false; 2247 mr->hw_mr.pbl_ptr = mr->info.pbl_table[0].pa; 2248 mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered; 2249 mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size); 2250 mr->hw_mr.page_size_log = mr->umem->page_shift; 2251 mr->hw_mr.fbo = ib_umem_offset(mr->umem); 2252 mr->hw_mr.length = len; 2253 mr->hw_mr.vaddr = usr_addr; 2254 mr->hw_mr.zbva = false; 2255 mr->hw_mr.phy_mr = false; 2256 mr->hw_mr.dma_mr = false; 2257 2258 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr); 2259 if (rc) { 2260 DP_ERR(dev, "roce register tid returned an error %d\n", rc); 2261 goto err2; 2262 } 2263 2264 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key; 2265 if (mr->hw_mr.remote_write || mr->hw_mr.remote_read || 2266 mr->hw_mr.remote_atomic) 2267 mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key; 2268 2269 DP_DEBUG(dev, QEDR_MSG_MR, "register user mr lkey: %x\n", 2270 mr->ibmr.lkey); 2271 return &mr->ibmr; 2272 2273 err2: 2274 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid); 2275 err1: 2276 qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table); 2277 err0: 2278 kfree(mr); 2279 return ERR_PTR(rc); 2280 } 2281 2282 int qedr_dereg_mr(struct ib_mr *ib_mr) 2283 { 2284 struct qedr_mr *mr = get_qedr_mr(ib_mr); 2285 struct qedr_dev *dev = get_qedr_dev(ib_mr->device); 2286 int rc = 0; 2287 2288 rc = dev->ops->rdma_deregister_tid(dev->rdma_ctx, mr->hw_mr.itid); 2289 if (rc) 2290 return rc; 2291 2292 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid); 2293 2294 if ((mr->type != QEDR_MR_DMA) && (mr->type != QEDR_MR_FRMR)) 2295 qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table); 2296 2297 /* it could be user registered memory. */ 2298 if (mr->umem) 2299 ib_umem_release(mr->umem); 2300 2301 kfree(mr); 2302 2303 return rc; 2304 } 2305 2306 static struct qedr_mr *__qedr_alloc_mr(struct ib_pd *ibpd, 2307 int max_page_list_len) 2308 { 2309 struct qedr_pd *pd = get_qedr_pd(ibpd); 2310 struct qedr_dev *dev = get_qedr_dev(ibpd->device); 2311 struct qedr_mr *mr; 2312 int rc = -ENOMEM; 2313 2314 DP_DEBUG(dev, QEDR_MSG_MR, 2315 "qedr_alloc_frmr pd = %d max_page_list_len= %d\n", pd->pd_id, 2316 max_page_list_len); 2317 2318 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 2319 if (!mr) 2320 return ERR_PTR(rc); 2321 2322 mr->dev = dev; 2323 mr->type = QEDR_MR_FRMR; 2324 2325 rc = init_mr_info(dev, &mr->info, max_page_list_len, 1); 2326 if (rc) 2327 goto err0; 2328 2329 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid); 2330 if (rc) { 2331 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc); 2332 goto err0; 2333 } 2334 2335 /* Index only, 18 bit long, lkey = itid << 8 | key */ 2336 mr->hw_mr.tid_type = QED_RDMA_TID_FMR; 2337 mr->hw_mr.key = 0; 2338 mr->hw_mr.pd = pd->pd_id; 2339 mr->hw_mr.local_read = 1; 2340 mr->hw_mr.local_write = 0; 2341 mr->hw_mr.remote_read = 0; 2342 mr->hw_mr.remote_write = 0; 2343 mr->hw_mr.remote_atomic = 0; 2344 mr->hw_mr.mw_bind = false; 2345 mr->hw_mr.pbl_ptr = 0; 2346 mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered; 2347 mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size); 2348 mr->hw_mr.fbo = 0; 2349 mr->hw_mr.length = 0; 2350 mr->hw_mr.vaddr = 0; 2351 mr->hw_mr.zbva = false; 2352 mr->hw_mr.phy_mr = true; 2353 mr->hw_mr.dma_mr = false; 2354 2355 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr); 2356 if (rc) { 2357 DP_ERR(dev, "roce register tid returned an error %d\n", rc); 2358 goto err1; 2359 } 2360 2361 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key; 2362 mr->ibmr.rkey = mr->ibmr.lkey; 2363 2364 DP_DEBUG(dev, QEDR_MSG_MR, "alloc frmr: %x\n", mr->ibmr.lkey); 2365 return mr; 2366 2367 err1: 2368 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid); 2369 err0: 2370 kfree(mr); 2371 return ERR_PTR(rc); 2372 } 2373 2374 struct ib_mr *qedr_alloc_mr(struct ib_pd *ibpd, 2375 enum ib_mr_type mr_type, u32 max_num_sg) 2376 { 2377 struct qedr_dev *dev; 2378 struct qedr_mr *mr; 2379 2380 if (mr_type != IB_MR_TYPE_MEM_REG) 2381 return ERR_PTR(-EINVAL); 2382 2383 mr = __qedr_alloc_mr(ibpd, max_num_sg); 2384 2385 if (IS_ERR(mr)) 2386 return ERR_PTR(-EINVAL); 2387 2388 dev = mr->dev; 2389 2390 return &mr->ibmr; 2391 } 2392 2393 static int qedr_set_page(struct ib_mr *ibmr, u64 addr) 2394 { 2395 struct qedr_mr *mr = get_qedr_mr(ibmr); 2396 struct qedr_pbl *pbl_table; 2397 struct regpair *pbe; 2398 u32 pbes_in_page; 2399 2400 if (unlikely(mr->npages == mr->info.pbl_info.num_pbes)) { 2401 DP_ERR(mr->dev, "qedr_set_page failes when %d\n", mr->npages); 2402 return -ENOMEM; 2403 } 2404 2405 DP_DEBUG(mr->dev, QEDR_MSG_MR, "qedr_set_page pages[%d] = 0x%llx\n", 2406 mr->npages, addr); 2407 2408 pbes_in_page = mr->info.pbl_info.pbl_size / sizeof(u64); 2409 pbl_table = mr->info.pbl_table + (mr->npages / pbes_in_page); 2410 pbe = (struct regpair *)pbl_table->va; 2411 pbe += mr->npages % pbes_in_page; 2412 pbe->lo = cpu_to_le32((u32)addr); 2413 pbe->hi = cpu_to_le32((u32)upper_32_bits(addr)); 2414 2415 mr->npages++; 2416 2417 return 0; 2418 } 2419 2420 static void handle_completed_mrs(struct qedr_dev *dev, struct mr_info *info) 2421 { 2422 int work = info->completed - info->completed_handled - 1; 2423 2424 DP_DEBUG(dev, QEDR_MSG_MR, "Special FMR work = %d\n", work); 2425 while (work-- > 0 && !list_empty(&info->inuse_pbl_list)) { 2426 struct qedr_pbl *pbl; 2427 2428 /* Free all the page list that are possible to be freed 2429 * (all the ones that were invalidated), under the assumption 2430 * that if an FMR was completed successfully that means that 2431 * if there was an invalidate operation before it also ended 2432 */ 2433 pbl = list_first_entry(&info->inuse_pbl_list, 2434 struct qedr_pbl, list_entry); 2435 list_move_tail(&pbl->list_entry, &info->free_pbl_list); 2436 info->completed_handled++; 2437 } 2438 } 2439 2440 int qedr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, 2441 int sg_nents, unsigned int *sg_offset) 2442 { 2443 struct qedr_mr *mr = get_qedr_mr(ibmr); 2444 2445 mr->npages = 0; 2446 2447 handle_completed_mrs(mr->dev, &mr->info); 2448 return ib_sg_to_pages(ibmr, sg, sg_nents, NULL, qedr_set_page); 2449 } 2450 2451 struct ib_mr *qedr_get_dma_mr(struct ib_pd *ibpd, int acc) 2452 { 2453 struct qedr_dev *dev = get_qedr_dev(ibpd->device); 2454 struct qedr_pd *pd = get_qedr_pd(ibpd); 2455 struct qedr_mr *mr; 2456 int rc; 2457 2458 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 2459 if (!mr) 2460 return ERR_PTR(-ENOMEM); 2461 2462 mr->type = QEDR_MR_DMA; 2463 2464 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid); 2465 if (rc) { 2466 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc); 2467 goto err1; 2468 } 2469 2470 /* index only, 18 bit long, lkey = itid << 8 | key */ 2471 mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR; 2472 mr->hw_mr.pd = pd->pd_id; 2473 mr->hw_mr.local_read = 1; 2474 mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0; 2475 mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0; 2476 mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0; 2477 mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0; 2478 mr->hw_mr.dma_mr = true; 2479 2480 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr); 2481 if (rc) { 2482 DP_ERR(dev, "roce register tid returned an error %d\n", rc); 2483 goto err2; 2484 } 2485 2486 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key; 2487 if (mr->hw_mr.remote_write || mr->hw_mr.remote_read || 2488 mr->hw_mr.remote_atomic) 2489 mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key; 2490 2491 DP_DEBUG(dev, QEDR_MSG_MR, "get dma mr: lkey = %x\n", mr->ibmr.lkey); 2492 return &mr->ibmr; 2493 2494 err2: 2495 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid); 2496 err1: 2497 kfree(mr); 2498 return ERR_PTR(rc); 2499 } 2500 2501 static inline int qedr_wq_is_full(struct qedr_qp_hwq_info *wq) 2502 { 2503 return (((wq->prod + 1) % wq->max_wr) == wq->cons); 2504 } 2505 2506 static int sge_data_len(struct ib_sge *sg_list, int num_sge) 2507 { 2508 int i, len = 0; 2509 2510 for (i = 0; i < num_sge; i++) 2511 len += sg_list[i].length; 2512 2513 return len; 2514 } 2515 2516 static void swap_wqe_data64(u64 *p) 2517 { 2518 int i; 2519 2520 for (i = 0; i < QEDR_SQE_ELEMENT_SIZE / sizeof(u64); i++, p++) 2521 *p = cpu_to_be64(cpu_to_le64(*p)); 2522 } 2523 2524 static u32 qedr_prepare_sq_inline_data(struct qedr_dev *dev, 2525 struct qedr_qp *qp, u8 *wqe_size, 2526 struct ib_send_wr *wr, 2527 struct ib_send_wr **bad_wr, u8 *bits, 2528 u8 bit) 2529 { 2530 u32 data_size = sge_data_len(wr->sg_list, wr->num_sge); 2531 char *seg_prt, *wqe; 2532 int i, seg_siz; 2533 2534 if (data_size > ROCE_REQ_MAX_INLINE_DATA_SIZE) { 2535 DP_ERR(dev, "Too much inline data in WR: %d\n", data_size); 2536 *bad_wr = wr; 2537 return 0; 2538 } 2539 2540 if (!data_size) 2541 return data_size; 2542 2543 *bits |= bit; 2544 2545 seg_prt = NULL; 2546 wqe = NULL; 2547 seg_siz = 0; 2548 2549 /* Copy data inline */ 2550 for (i = 0; i < wr->num_sge; i++) { 2551 u32 len = wr->sg_list[i].length; 2552 void *src = (void *)(uintptr_t)wr->sg_list[i].addr; 2553 2554 while (len > 0) { 2555 u32 cur; 2556 2557 /* New segment required */ 2558 if (!seg_siz) { 2559 wqe = (char *)qed_chain_produce(&qp->sq.pbl); 2560 seg_prt = wqe; 2561 seg_siz = sizeof(struct rdma_sq_common_wqe); 2562 (*wqe_size)++; 2563 } 2564 2565 /* Calculate currently allowed length */ 2566 cur = min_t(u32, len, seg_siz); 2567 memcpy(seg_prt, src, cur); 2568 2569 /* Update segment variables */ 2570 seg_prt += cur; 2571 seg_siz -= cur; 2572 2573 /* Update sge variables */ 2574 src += cur; 2575 len -= cur; 2576 2577 /* Swap fully-completed segments */ 2578 if (!seg_siz) 2579 swap_wqe_data64((u64 *)wqe); 2580 } 2581 } 2582 2583 /* swap last not completed segment */ 2584 if (seg_siz) 2585 swap_wqe_data64((u64 *)wqe); 2586 2587 return data_size; 2588 } 2589 2590 #define RQ_SGE_SET(sge, vaddr, vlength, vflags) \ 2591 do { \ 2592 DMA_REGPAIR_LE(sge->addr, vaddr); \ 2593 (sge)->length = cpu_to_le32(vlength); \ 2594 (sge)->flags = cpu_to_le32(vflags); \ 2595 } while (0) 2596 2597 #define SRQ_HDR_SET(hdr, vwr_id, num_sge) \ 2598 do { \ 2599 DMA_REGPAIR_LE(hdr->wr_id, vwr_id); \ 2600 (hdr)->num_sges = num_sge; \ 2601 } while (0) 2602 2603 #define SRQ_SGE_SET(sge, vaddr, vlength, vlkey) \ 2604 do { \ 2605 DMA_REGPAIR_LE(sge->addr, vaddr); \ 2606 (sge)->length = cpu_to_le32(vlength); \ 2607 (sge)->l_key = cpu_to_le32(vlkey); \ 2608 } while (0) 2609 2610 static u32 qedr_prepare_sq_sges(struct qedr_qp *qp, u8 *wqe_size, 2611 struct ib_send_wr *wr) 2612 { 2613 u32 data_size = 0; 2614 int i; 2615 2616 for (i = 0; i < wr->num_sge; i++) { 2617 struct rdma_sq_sge *sge = qed_chain_produce(&qp->sq.pbl); 2618 2619 DMA_REGPAIR_LE(sge->addr, wr->sg_list[i].addr); 2620 sge->l_key = cpu_to_le32(wr->sg_list[i].lkey); 2621 sge->length = cpu_to_le32(wr->sg_list[i].length); 2622 data_size += wr->sg_list[i].length; 2623 } 2624 2625 if (wqe_size) 2626 *wqe_size += wr->num_sge; 2627 2628 return data_size; 2629 } 2630 2631 static u32 qedr_prepare_sq_rdma_data(struct qedr_dev *dev, 2632 struct qedr_qp *qp, 2633 struct rdma_sq_rdma_wqe_1st *rwqe, 2634 struct rdma_sq_rdma_wqe_2nd *rwqe2, 2635 struct ib_send_wr *wr, 2636 struct ib_send_wr **bad_wr) 2637 { 2638 rwqe2->r_key = cpu_to_le32(rdma_wr(wr)->rkey); 2639 DMA_REGPAIR_LE(rwqe2->remote_va, rdma_wr(wr)->remote_addr); 2640 2641 if (wr->send_flags & IB_SEND_INLINE && 2642 (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM || 2643 wr->opcode == IB_WR_RDMA_WRITE)) { 2644 u8 flags = 0; 2645 2646 SET_FIELD2(flags, RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG, 1); 2647 return qedr_prepare_sq_inline_data(dev, qp, &rwqe->wqe_size, wr, 2648 bad_wr, &rwqe->flags, flags); 2649 } 2650 2651 return qedr_prepare_sq_sges(qp, &rwqe->wqe_size, wr); 2652 } 2653 2654 static u32 qedr_prepare_sq_send_data(struct qedr_dev *dev, 2655 struct qedr_qp *qp, 2656 struct rdma_sq_send_wqe_1st *swqe, 2657 struct rdma_sq_send_wqe_2st *swqe2, 2658 struct ib_send_wr *wr, 2659 struct ib_send_wr **bad_wr) 2660 { 2661 memset(swqe2, 0, sizeof(*swqe2)); 2662 if (wr->send_flags & IB_SEND_INLINE) { 2663 u8 flags = 0; 2664 2665 SET_FIELD2(flags, RDMA_SQ_SEND_WQE_INLINE_FLG, 1); 2666 return qedr_prepare_sq_inline_data(dev, qp, &swqe->wqe_size, wr, 2667 bad_wr, &swqe->flags, flags); 2668 } 2669 2670 return qedr_prepare_sq_sges(qp, &swqe->wqe_size, wr); 2671 } 2672 2673 static int qedr_prepare_reg(struct qedr_qp *qp, 2674 struct rdma_sq_fmr_wqe_1st *fwqe1, 2675 struct ib_reg_wr *wr) 2676 { 2677 struct qedr_mr *mr = get_qedr_mr(wr->mr); 2678 struct rdma_sq_fmr_wqe_2nd *fwqe2; 2679 2680 fwqe2 = (struct rdma_sq_fmr_wqe_2nd *)qed_chain_produce(&qp->sq.pbl); 2681 fwqe1->addr.hi = upper_32_bits(mr->ibmr.iova); 2682 fwqe1->addr.lo = lower_32_bits(mr->ibmr.iova); 2683 fwqe1->l_key = wr->key; 2684 2685 fwqe2->access_ctrl = 0; 2686 2687 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_READ, 2688 !!(wr->access & IB_ACCESS_REMOTE_READ)); 2689 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE, 2690 !!(wr->access & IB_ACCESS_REMOTE_WRITE)); 2691 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC, 2692 !!(wr->access & IB_ACCESS_REMOTE_ATOMIC)); 2693 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_READ, 1); 2694 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE, 2695 !!(wr->access & IB_ACCESS_LOCAL_WRITE)); 2696 fwqe2->fmr_ctrl = 0; 2697 2698 SET_FIELD2(fwqe2->fmr_ctrl, RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG, 2699 ilog2(mr->ibmr.page_size) - 12); 2700 2701 fwqe2->length_hi = 0; 2702 fwqe2->length_lo = mr->ibmr.length; 2703 fwqe2->pbl_addr.hi = upper_32_bits(mr->info.pbl_table->pa); 2704 fwqe2->pbl_addr.lo = lower_32_bits(mr->info.pbl_table->pa); 2705 2706 qp->wqe_wr_id[qp->sq.prod].mr = mr; 2707 2708 return 0; 2709 } 2710 2711 static enum ib_wc_opcode qedr_ib_to_wc_opcode(enum ib_wr_opcode opcode) 2712 { 2713 switch (opcode) { 2714 case IB_WR_RDMA_WRITE: 2715 case IB_WR_RDMA_WRITE_WITH_IMM: 2716 return IB_WC_RDMA_WRITE; 2717 case IB_WR_SEND_WITH_IMM: 2718 case IB_WR_SEND: 2719 case IB_WR_SEND_WITH_INV: 2720 return IB_WC_SEND; 2721 case IB_WR_RDMA_READ: 2722 return IB_WC_RDMA_READ; 2723 case IB_WR_ATOMIC_CMP_AND_SWP: 2724 return IB_WC_COMP_SWAP; 2725 case IB_WR_ATOMIC_FETCH_AND_ADD: 2726 return IB_WC_FETCH_ADD; 2727 case IB_WR_REG_MR: 2728 return IB_WC_REG_MR; 2729 case IB_WR_LOCAL_INV: 2730 return IB_WC_LOCAL_INV; 2731 default: 2732 return IB_WC_SEND; 2733 } 2734 } 2735 2736 static inline bool qedr_can_post_send(struct qedr_qp *qp, struct ib_send_wr *wr) 2737 { 2738 int wq_is_full, err_wr, pbl_is_full; 2739 struct qedr_dev *dev = qp->dev; 2740 2741 /* prevent SQ overflow and/or processing of a bad WR */ 2742 err_wr = wr->num_sge > qp->sq.max_sges; 2743 wq_is_full = qedr_wq_is_full(&qp->sq); 2744 pbl_is_full = qed_chain_get_elem_left_u32(&qp->sq.pbl) < 2745 QEDR_MAX_SQE_ELEMENTS_PER_SQE; 2746 if (wq_is_full || err_wr || pbl_is_full) { 2747 if (wq_is_full && !(qp->err_bitmap & QEDR_QP_ERR_SQ_FULL)) { 2748 DP_ERR(dev, 2749 "error: WQ is full. Post send on QP %p failed (this error appears only once)\n", 2750 qp); 2751 qp->err_bitmap |= QEDR_QP_ERR_SQ_FULL; 2752 } 2753 2754 if (err_wr && !(qp->err_bitmap & QEDR_QP_ERR_BAD_SR)) { 2755 DP_ERR(dev, 2756 "error: WR is bad. Post send on QP %p failed (this error appears only once)\n", 2757 qp); 2758 qp->err_bitmap |= QEDR_QP_ERR_BAD_SR; 2759 } 2760 2761 if (pbl_is_full && 2762 !(qp->err_bitmap & QEDR_QP_ERR_SQ_PBL_FULL)) { 2763 DP_ERR(dev, 2764 "error: WQ PBL is full. Post send on QP %p failed (this error appears only once)\n", 2765 qp); 2766 qp->err_bitmap |= QEDR_QP_ERR_SQ_PBL_FULL; 2767 } 2768 return false; 2769 } 2770 return true; 2771 } 2772 2773 static int __qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 2774 struct ib_send_wr **bad_wr) 2775 { 2776 struct qedr_dev *dev = get_qedr_dev(ibqp->device); 2777 struct qedr_qp *qp = get_qedr_qp(ibqp); 2778 struct rdma_sq_atomic_wqe_1st *awqe1; 2779 struct rdma_sq_atomic_wqe_2nd *awqe2; 2780 struct rdma_sq_atomic_wqe_3rd *awqe3; 2781 struct rdma_sq_send_wqe_2st *swqe2; 2782 struct rdma_sq_local_inv_wqe *iwqe; 2783 struct rdma_sq_rdma_wqe_2nd *rwqe2; 2784 struct rdma_sq_send_wqe_1st *swqe; 2785 struct rdma_sq_rdma_wqe_1st *rwqe; 2786 struct rdma_sq_fmr_wqe_1st *fwqe1; 2787 struct rdma_sq_common_wqe *wqe; 2788 u32 length; 2789 int rc = 0; 2790 bool comp; 2791 2792 if (!qedr_can_post_send(qp, wr)) { 2793 *bad_wr = wr; 2794 return -ENOMEM; 2795 } 2796 2797 wqe = qed_chain_produce(&qp->sq.pbl); 2798 qp->wqe_wr_id[qp->sq.prod].signaled = 2799 !!(wr->send_flags & IB_SEND_SIGNALED) || qp->signaled; 2800 2801 wqe->flags = 0; 2802 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_SE_FLG, 2803 !!(wr->send_flags & IB_SEND_SOLICITED)); 2804 comp = (!!(wr->send_flags & IB_SEND_SIGNALED)) || qp->signaled; 2805 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_COMP_FLG, comp); 2806 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_RD_FENCE_FLG, 2807 !!(wr->send_flags & IB_SEND_FENCE)); 2808 wqe->prev_wqe_size = qp->prev_wqe_size; 2809 2810 qp->wqe_wr_id[qp->sq.prod].opcode = qedr_ib_to_wc_opcode(wr->opcode); 2811 2812 switch (wr->opcode) { 2813 case IB_WR_SEND_WITH_IMM: 2814 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_IMM; 2815 swqe = (struct rdma_sq_send_wqe_1st *)wqe; 2816 swqe->wqe_size = 2; 2817 swqe2 = qed_chain_produce(&qp->sq.pbl); 2818 2819 swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.imm_data); 2820 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2, 2821 wr, bad_wr); 2822 swqe->length = cpu_to_le32(length); 2823 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size; 2824 qp->prev_wqe_size = swqe->wqe_size; 2825 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length; 2826 break; 2827 case IB_WR_SEND: 2828 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND; 2829 swqe = (struct rdma_sq_send_wqe_1st *)wqe; 2830 2831 swqe->wqe_size = 2; 2832 swqe2 = qed_chain_produce(&qp->sq.pbl); 2833 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2, 2834 wr, bad_wr); 2835 swqe->length = cpu_to_le32(length); 2836 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size; 2837 qp->prev_wqe_size = swqe->wqe_size; 2838 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length; 2839 break; 2840 case IB_WR_SEND_WITH_INV: 2841 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_INVALIDATE; 2842 swqe = (struct rdma_sq_send_wqe_1st *)wqe; 2843 swqe2 = qed_chain_produce(&qp->sq.pbl); 2844 swqe->wqe_size = 2; 2845 swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.invalidate_rkey); 2846 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2, 2847 wr, bad_wr); 2848 swqe->length = cpu_to_le32(length); 2849 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size; 2850 qp->prev_wqe_size = swqe->wqe_size; 2851 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length; 2852 break; 2853 2854 case IB_WR_RDMA_WRITE_WITH_IMM: 2855 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR_WITH_IMM; 2856 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe; 2857 2858 rwqe->wqe_size = 2; 2859 rwqe->imm_data = htonl(cpu_to_le32(wr->ex.imm_data)); 2860 rwqe2 = qed_chain_produce(&qp->sq.pbl); 2861 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2, 2862 wr, bad_wr); 2863 rwqe->length = cpu_to_le32(length); 2864 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size; 2865 qp->prev_wqe_size = rwqe->wqe_size; 2866 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length; 2867 break; 2868 case IB_WR_RDMA_WRITE: 2869 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR; 2870 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe; 2871 2872 rwqe->wqe_size = 2; 2873 rwqe2 = qed_chain_produce(&qp->sq.pbl); 2874 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2, 2875 wr, bad_wr); 2876 rwqe->length = cpu_to_le32(length); 2877 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size; 2878 qp->prev_wqe_size = rwqe->wqe_size; 2879 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length; 2880 break; 2881 case IB_WR_RDMA_READ_WITH_INV: 2882 DP_ERR(dev, 2883 "RDMA READ WITH INVALIDATE not supported\n"); 2884 *bad_wr = wr; 2885 rc = -EINVAL; 2886 break; 2887 2888 case IB_WR_RDMA_READ: 2889 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_RD; 2890 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe; 2891 2892 rwqe->wqe_size = 2; 2893 rwqe2 = qed_chain_produce(&qp->sq.pbl); 2894 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2, 2895 wr, bad_wr); 2896 rwqe->length = cpu_to_le32(length); 2897 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size; 2898 qp->prev_wqe_size = rwqe->wqe_size; 2899 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length; 2900 break; 2901 2902 case IB_WR_ATOMIC_CMP_AND_SWP: 2903 case IB_WR_ATOMIC_FETCH_AND_ADD: 2904 awqe1 = (struct rdma_sq_atomic_wqe_1st *)wqe; 2905 awqe1->wqe_size = 4; 2906 2907 awqe2 = qed_chain_produce(&qp->sq.pbl); 2908 DMA_REGPAIR_LE(awqe2->remote_va, atomic_wr(wr)->remote_addr); 2909 awqe2->r_key = cpu_to_le32(atomic_wr(wr)->rkey); 2910 2911 awqe3 = qed_chain_produce(&qp->sq.pbl); 2912 2913 if (wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) { 2914 wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_ADD; 2915 DMA_REGPAIR_LE(awqe3->swap_data, 2916 atomic_wr(wr)->compare_add); 2917 } else { 2918 wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_CMP_AND_SWAP; 2919 DMA_REGPAIR_LE(awqe3->swap_data, 2920 atomic_wr(wr)->swap); 2921 DMA_REGPAIR_LE(awqe3->cmp_data, 2922 atomic_wr(wr)->compare_add); 2923 } 2924 2925 qedr_prepare_sq_sges(qp, NULL, wr); 2926 2927 qp->wqe_wr_id[qp->sq.prod].wqe_size = awqe1->wqe_size; 2928 qp->prev_wqe_size = awqe1->wqe_size; 2929 break; 2930 2931 case IB_WR_LOCAL_INV: 2932 iwqe = (struct rdma_sq_local_inv_wqe *)wqe; 2933 iwqe->wqe_size = 1; 2934 2935 iwqe->req_type = RDMA_SQ_REQ_TYPE_LOCAL_INVALIDATE; 2936 iwqe->inv_l_key = wr->ex.invalidate_rkey; 2937 qp->wqe_wr_id[qp->sq.prod].wqe_size = iwqe->wqe_size; 2938 qp->prev_wqe_size = iwqe->wqe_size; 2939 break; 2940 case IB_WR_REG_MR: 2941 DP_DEBUG(dev, QEDR_MSG_CQ, "REG_MR\n"); 2942 wqe->req_type = RDMA_SQ_REQ_TYPE_FAST_MR; 2943 fwqe1 = (struct rdma_sq_fmr_wqe_1st *)wqe; 2944 fwqe1->wqe_size = 2; 2945 2946 rc = qedr_prepare_reg(qp, fwqe1, reg_wr(wr)); 2947 if (rc) { 2948 DP_ERR(dev, "IB_REG_MR failed rc=%d\n", rc); 2949 *bad_wr = wr; 2950 break; 2951 } 2952 2953 qp->wqe_wr_id[qp->sq.prod].wqe_size = fwqe1->wqe_size; 2954 qp->prev_wqe_size = fwqe1->wqe_size; 2955 break; 2956 default: 2957 DP_ERR(dev, "invalid opcode 0x%x!\n", wr->opcode); 2958 rc = -EINVAL; 2959 *bad_wr = wr; 2960 break; 2961 } 2962 2963 if (*bad_wr) { 2964 u16 value; 2965 2966 /* Restore prod to its position before 2967 * this WR was processed 2968 */ 2969 value = le16_to_cpu(qp->sq.db_data.data.value); 2970 qed_chain_set_prod(&qp->sq.pbl, value, wqe); 2971 2972 /* Restore prev_wqe_size */ 2973 qp->prev_wqe_size = wqe->prev_wqe_size; 2974 rc = -EINVAL; 2975 DP_ERR(dev, "POST SEND FAILED\n"); 2976 } 2977 2978 return rc; 2979 } 2980 2981 int qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 2982 struct ib_send_wr **bad_wr) 2983 { 2984 struct qedr_dev *dev = get_qedr_dev(ibqp->device); 2985 struct qedr_qp *qp = get_qedr_qp(ibqp); 2986 unsigned long flags; 2987 int rc = 0; 2988 2989 *bad_wr = NULL; 2990 2991 if (qp->qp_type == IB_QPT_GSI) 2992 return qedr_gsi_post_send(ibqp, wr, bad_wr); 2993 2994 spin_lock_irqsave(&qp->q_lock, flags); 2995 2996 if ((qp->state != QED_ROCE_QP_STATE_RTS) && 2997 (qp->state != QED_ROCE_QP_STATE_ERR) && 2998 (qp->state != QED_ROCE_QP_STATE_SQD)) { 2999 spin_unlock_irqrestore(&qp->q_lock, flags); 3000 *bad_wr = wr; 3001 DP_DEBUG(dev, QEDR_MSG_CQ, 3002 "QP in wrong state! QP icid=0x%x state %d\n", 3003 qp->icid, qp->state); 3004 return -EINVAL; 3005 } 3006 3007 while (wr) { 3008 rc = __qedr_post_send(ibqp, wr, bad_wr); 3009 if (rc) 3010 break; 3011 3012 qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id; 3013 3014 qedr_inc_sw_prod(&qp->sq); 3015 3016 qp->sq.db_data.data.value++; 3017 3018 wr = wr->next; 3019 } 3020 3021 /* Trigger doorbell 3022 * If there was a failure in the first WR then it will be triggered in 3023 * vane. However this is not harmful (as long as the producer value is 3024 * unchanged). For performance reasons we avoid checking for this 3025 * redundant doorbell. 3026 */ 3027 wmb(); 3028 writel(qp->sq.db_data.raw, qp->sq.db); 3029 3030 /* Make sure write sticks */ 3031 mmiowb(); 3032 3033 spin_unlock_irqrestore(&qp->q_lock, flags); 3034 3035 return rc; 3036 } 3037 3038 int qedr_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 3039 struct ib_recv_wr **bad_wr) 3040 { 3041 struct qedr_qp *qp = get_qedr_qp(ibqp); 3042 struct qedr_dev *dev = qp->dev; 3043 unsigned long flags; 3044 int status = 0; 3045 3046 if (qp->qp_type == IB_QPT_GSI) 3047 return qedr_gsi_post_recv(ibqp, wr, bad_wr); 3048 3049 spin_lock_irqsave(&qp->q_lock, flags); 3050 3051 if (qp->state == QED_ROCE_QP_STATE_RESET) { 3052 spin_unlock_irqrestore(&qp->q_lock, flags); 3053 *bad_wr = wr; 3054 return -EINVAL; 3055 } 3056 3057 while (wr) { 3058 int i; 3059 3060 if (qed_chain_get_elem_left_u32(&qp->rq.pbl) < 3061 QEDR_MAX_RQE_ELEMENTS_PER_RQE || 3062 wr->num_sge > qp->rq.max_sges) { 3063 DP_ERR(dev, "Can't post WR (%d < %d) || (%d > %d)\n", 3064 qed_chain_get_elem_left_u32(&qp->rq.pbl), 3065 QEDR_MAX_RQE_ELEMENTS_PER_RQE, wr->num_sge, 3066 qp->rq.max_sges); 3067 status = -ENOMEM; 3068 *bad_wr = wr; 3069 break; 3070 } 3071 for (i = 0; i < wr->num_sge; i++) { 3072 u32 flags = 0; 3073 struct rdma_rq_sge *rqe = 3074 qed_chain_produce(&qp->rq.pbl); 3075 3076 /* First one must include the number 3077 * of SGE in the list 3078 */ 3079 if (!i) 3080 SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, 3081 wr->num_sge); 3082 3083 SET_FIELD(flags, RDMA_RQ_SGE_L_KEY, 3084 wr->sg_list[i].lkey); 3085 3086 RQ_SGE_SET(rqe, wr->sg_list[i].addr, 3087 wr->sg_list[i].length, flags); 3088 } 3089 3090 /* Special case of no sges. FW requires between 1-4 sges... 3091 * in this case we need to post 1 sge with length zero. this is 3092 * because rdma write with immediate consumes an RQ. 3093 */ 3094 if (!wr->num_sge) { 3095 u32 flags = 0; 3096 struct rdma_rq_sge *rqe = 3097 qed_chain_produce(&qp->rq.pbl); 3098 3099 /* First one must include the number 3100 * of SGE in the list 3101 */ 3102 SET_FIELD(flags, RDMA_RQ_SGE_L_KEY, 0); 3103 SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, 1); 3104 3105 RQ_SGE_SET(rqe, 0, 0, flags); 3106 i = 1; 3107 } 3108 3109 qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id; 3110 qp->rqe_wr_id[qp->rq.prod].wqe_size = i; 3111 3112 qedr_inc_sw_prod(&qp->rq); 3113 3114 /* Flush all the writes before signalling doorbell */ 3115 wmb(); 3116 3117 qp->rq.db_data.data.value++; 3118 3119 writel(qp->rq.db_data.raw, qp->rq.db); 3120 3121 /* Make sure write sticks */ 3122 mmiowb(); 3123 3124 wr = wr->next; 3125 } 3126 3127 spin_unlock_irqrestore(&qp->q_lock, flags); 3128 3129 return status; 3130 } 3131 3132 static int is_valid_cqe(struct qedr_cq *cq, union rdma_cqe *cqe) 3133 { 3134 struct rdma_cqe_requester *resp_cqe = &cqe->req; 3135 3136 return (resp_cqe->flags & RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK) == 3137 cq->pbl_toggle; 3138 } 3139 3140 static struct qedr_qp *cqe_get_qp(union rdma_cqe *cqe) 3141 { 3142 struct rdma_cqe_requester *resp_cqe = &cqe->req; 3143 struct qedr_qp *qp; 3144 3145 qp = (struct qedr_qp *)(uintptr_t)HILO_GEN(resp_cqe->qp_handle.hi, 3146 resp_cqe->qp_handle.lo, 3147 u64); 3148 return qp; 3149 } 3150 3151 static enum rdma_cqe_type cqe_get_type(union rdma_cqe *cqe) 3152 { 3153 struct rdma_cqe_requester *resp_cqe = &cqe->req; 3154 3155 return GET_FIELD(resp_cqe->flags, RDMA_CQE_REQUESTER_TYPE); 3156 } 3157 3158 /* Return latest CQE (needs processing) */ 3159 static union rdma_cqe *get_cqe(struct qedr_cq *cq) 3160 { 3161 return cq->latest_cqe; 3162 } 3163 3164 /* In fmr we need to increase the number of fmr completed counter for the fmr 3165 * algorithm determining whether we can free a pbl or not. 3166 * we need to perform this whether the work request was signaled or not. for 3167 * this purpose we call this function from the condition that checks if a wr 3168 * should be skipped, to make sure we don't miss it ( possibly this fmr 3169 * operation was not signalted) 3170 */ 3171 static inline void qedr_chk_if_fmr(struct qedr_qp *qp) 3172 { 3173 if (qp->wqe_wr_id[qp->sq.cons].opcode == IB_WC_REG_MR) 3174 qp->wqe_wr_id[qp->sq.cons].mr->info.completed++; 3175 } 3176 3177 static int process_req(struct qedr_dev *dev, struct qedr_qp *qp, 3178 struct qedr_cq *cq, int num_entries, 3179 struct ib_wc *wc, u16 hw_cons, enum ib_wc_status status, 3180 int force) 3181 { 3182 u16 cnt = 0; 3183 3184 while (num_entries && qp->sq.wqe_cons != hw_cons) { 3185 if (!qp->wqe_wr_id[qp->sq.cons].signaled && !force) { 3186 qedr_chk_if_fmr(qp); 3187 /* skip WC */ 3188 goto next_cqe; 3189 } 3190 3191 /* fill WC */ 3192 wc->status = status; 3193 wc->vendor_err = 0; 3194 wc->wc_flags = 0; 3195 wc->src_qp = qp->id; 3196 wc->qp = &qp->ibqp; 3197 3198 wc->wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id; 3199 wc->opcode = qp->wqe_wr_id[qp->sq.cons].opcode; 3200 3201 switch (wc->opcode) { 3202 case IB_WC_RDMA_WRITE: 3203 wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len; 3204 break; 3205 case IB_WC_COMP_SWAP: 3206 case IB_WC_FETCH_ADD: 3207 wc->byte_len = 8; 3208 break; 3209 case IB_WC_REG_MR: 3210 qp->wqe_wr_id[qp->sq.cons].mr->info.completed++; 3211 break; 3212 default: 3213 break; 3214 } 3215 3216 num_entries--; 3217 wc++; 3218 cnt++; 3219 next_cqe: 3220 while (qp->wqe_wr_id[qp->sq.cons].wqe_size--) 3221 qed_chain_consume(&qp->sq.pbl); 3222 qedr_inc_sw_cons(&qp->sq); 3223 } 3224 3225 return cnt; 3226 } 3227 3228 static int qedr_poll_cq_req(struct qedr_dev *dev, 3229 struct qedr_qp *qp, struct qedr_cq *cq, 3230 int num_entries, struct ib_wc *wc, 3231 struct rdma_cqe_requester *req) 3232 { 3233 int cnt = 0; 3234 3235 switch (req->status) { 3236 case RDMA_CQE_REQ_STS_OK: 3237 cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons, 3238 IB_WC_SUCCESS, 0); 3239 break; 3240 case RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR: 3241 if (qp->state != QED_ROCE_QP_STATE_ERR) 3242 DP_ERR(dev, 3243 "Error: POLL CQ with RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3244 cq->icid, qp->icid); 3245 cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons, 3246 IB_WC_WR_FLUSH_ERR, 1); 3247 break; 3248 default: 3249 /* process all WQE before the cosumer */ 3250 qp->state = QED_ROCE_QP_STATE_ERR; 3251 cnt = process_req(dev, qp, cq, num_entries, wc, 3252 req->sq_cons - 1, IB_WC_SUCCESS, 0); 3253 wc += cnt; 3254 /* if we have extra WC fill it with actual error info */ 3255 if (cnt < num_entries) { 3256 enum ib_wc_status wc_status; 3257 3258 switch (req->status) { 3259 case RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR: 3260 DP_ERR(dev, 3261 "Error: POLL CQ with RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3262 cq->icid, qp->icid); 3263 wc_status = IB_WC_BAD_RESP_ERR; 3264 break; 3265 case RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR: 3266 DP_ERR(dev, 3267 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3268 cq->icid, qp->icid); 3269 wc_status = IB_WC_LOC_LEN_ERR; 3270 break; 3271 case RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR: 3272 DP_ERR(dev, 3273 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3274 cq->icid, qp->icid); 3275 wc_status = IB_WC_LOC_QP_OP_ERR; 3276 break; 3277 case RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR: 3278 DP_ERR(dev, 3279 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3280 cq->icid, qp->icid); 3281 wc_status = IB_WC_LOC_PROT_ERR; 3282 break; 3283 case RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR: 3284 DP_ERR(dev, 3285 "Error: POLL CQ with RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3286 cq->icid, qp->icid); 3287 wc_status = IB_WC_MW_BIND_ERR; 3288 break; 3289 case RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR: 3290 DP_ERR(dev, 3291 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3292 cq->icid, qp->icid); 3293 wc_status = IB_WC_REM_INV_REQ_ERR; 3294 break; 3295 case RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR: 3296 DP_ERR(dev, 3297 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3298 cq->icid, qp->icid); 3299 wc_status = IB_WC_REM_ACCESS_ERR; 3300 break; 3301 case RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR: 3302 DP_ERR(dev, 3303 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3304 cq->icid, qp->icid); 3305 wc_status = IB_WC_REM_OP_ERR; 3306 break; 3307 case RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR: 3308 DP_ERR(dev, 3309 "Error: POLL CQ with RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3310 cq->icid, qp->icid); 3311 wc_status = IB_WC_RNR_RETRY_EXC_ERR; 3312 break; 3313 case RDMA_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR: 3314 DP_ERR(dev, 3315 "Error: POLL CQ with ROCE_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3316 cq->icid, qp->icid); 3317 wc_status = IB_WC_RETRY_EXC_ERR; 3318 break; 3319 default: 3320 DP_ERR(dev, 3321 "Error: POLL CQ with IB_WC_GENERAL_ERR. CQ icid=0x%x, QP icid=0x%x\n", 3322 cq->icid, qp->icid); 3323 wc_status = IB_WC_GENERAL_ERR; 3324 } 3325 cnt += process_req(dev, qp, cq, 1, wc, req->sq_cons, 3326 wc_status, 1); 3327 } 3328 } 3329 3330 return cnt; 3331 } 3332 3333 static inline int qedr_cqe_resp_status_to_ib(u8 status) 3334 { 3335 switch (status) { 3336 case RDMA_CQE_RESP_STS_LOCAL_ACCESS_ERR: 3337 return IB_WC_LOC_ACCESS_ERR; 3338 case RDMA_CQE_RESP_STS_LOCAL_LENGTH_ERR: 3339 return IB_WC_LOC_LEN_ERR; 3340 case RDMA_CQE_RESP_STS_LOCAL_QP_OPERATION_ERR: 3341 return IB_WC_LOC_QP_OP_ERR; 3342 case RDMA_CQE_RESP_STS_LOCAL_PROTECTION_ERR: 3343 return IB_WC_LOC_PROT_ERR; 3344 case RDMA_CQE_RESP_STS_MEMORY_MGT_OPERATION_ERR: 3345 return IB_WC_MW_BIND_ERR; 3346 case RDMA_CQE_RESP_STS_REMOTE_INVALID_REQUEST_ERR: 3347 return IB_WC_REM_INV_RD_REQ_ERR; 3348 case RDMA_CQE_RESP_STS_OK: 3349 return IB_WC_SUCCESS; 3350 default: 3351 return IB_WC_GENERAL_ERR; 3352 } 3353 } 3354 3355 static inline int qedr_set_ok_cqe_resp_wc(struct rdma_cqe_responder *resp, 3356 struct ib_wc *wc) 3357 { 3358 wc->status = IB_WC_SUCCESS; 3359 wc->byte_len = le32_to_cpu(resp->length); 3360 3361 if (resp->flags & QEDR_RESP_IMM) { 3362 wc->ex.imm_data = le32_to_cpu(resp->imm_data_or_inv_r_Key); 3363 wc->wc_flags |= IB_WC_WITH_IMM; 3364 3365 if (resp->flags & QEDR_RESP_RDMA) 3366 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM; 3367 3368 if (resp->flags & QEDR_RESP_INV) 3369 return -EINVAL; 3370 3371 } else if (resp->flags & QEDR_RESP_INV) { 3372 wc->ex.imm_data = le32_to_cpu(resp->imm_data_or_inv_r_Key); 3373 wc->wc_flags |= IB_WC_WITH_INVALIDATE; 3374 3375 if (resp->flags & QEDR_RESP_RDMA) 3376 return -EINVAL; 3377 3378 } else if (resp->flags & QEDR_RESP_RDMA) { 3379 return -EINVAL; 3380 } 3381 3382 return 0; 3383 } 3384 3385 static void __process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp, 3386 struct qedr_cq *cq, struct ib_wc *wc, 3387 struct rdma_cqe_responder *resp, u64 wr_id) 3388 { 3389 /* Must fill fields before qedr_set_ok_cqe_resp_wc() */ 3390 wc->opcode = IB_WC_RECV; 3391 wc->wc_flags = 0; 3392 3393 if (likely(resp->status == RDMA_CQE_RESP_STS_OK)) { 3394 if (qedr_set_ok_cqe_resp_wc(resp, wc)) 3395 DP_ERR(dev, 3396 "CQ %p (icid=%d) has invalid CQE responder flags=0x%x\n", 3397 cq, cq->icid, resp->flags); 3398 3399 } else { 3400 wc->status = qedr_cqe_resp_status_to_ib(resp->status); 3401 if (wc->status == IB_WC_GENERAL_ERR) 3402 DP_ERR(dev, 3403 "CQ %p (icid=%d) contains an invalid CQE status %d\n", 3404 cq, cq->icid, resp->status); 3405 } 3406 3407 /* Fill the rest of the WC */ 3408 wc->vendor_err = 0; 3409 wc->src_qp = qp->id; 3410 wc->qp = &qp->ibqp; 3411 wc->wr_id = wr_id; 3412 } 3413 3414 static int process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp, 3415 struct qedr_cq *cq, struct ib_wc *wc, 3416 struct rdma_cqe_responder *resp) 3417 { 3418 u64 wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id; 3419 3420 __process_resp_one(dev, qp, cq, wc, resp, wr_id); 3421 3422 while (qp->rqe_wr_id[qp->rq.cons].wqe_size--) 3423 qed_chain_consume(&qp->rq.pbl); 3424 qedr_inc_sw_cons(&qp->rq); 3425 3426 return 1; 3427 } 3428 3429 static int process_resp_flush(struct qedr_qp *qp, struct qedr_cq *cq, 3430 int num_entries, struct ib_wc *wc, u16 hw_cons) 3431 { 3432 u16 cnt = 0; 3433 3434 while (num_entries && qp->rq.wqe_cons != hw_cons) { 3435 /* fill WC */ 3436 wc->status = IB_WC_WR_FLUSH_ERR; 3437 wc->vendor_err = 0; 3438 wc->wc_flags = 0; 3439 wc->src_qp = qp->id; 3440 wc->byte_len = 0; 3441 wc->wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id; 3442 wc->qp = &qp->ibqp; 3443 num_entries--; 3444 wc++; 3445 cnt++; 3446 while (qp->rqe_wr_id[qp->rq.cons].wqe_size--) 3447 qed_chain_consume(&qp->rq.pbl); 3448 qedr_inc_sw_cons(&qp->rq); 3449 } 3450 3451 return cnt; 3452 } 3453 3454 static void try_consume_resp_cqe(struct qedr_cq *cq, struct qedr_qp *qp, 3455 struct rdma_cqe_responder *resp, int *update) 3456 { 3457 if (le16_to_cpu(resp->rq_cons) == qp->rq.wqe_cons) { 3458 consume_cqe(cq); 3459 *update |= 1; 3460 } 3461 } 3462 3463 static int qedr_poll_cq_resp(struct qedr_dev *dev, struct qedr_qp *qp, 3464 struct qedr_cq *cq, int num_entries, 3465 struct ib_wc *wc, struct rdma_cqe_responder *resp, 3466 int *update) 3467 { 3468 int cnt; 3469 3470 if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) { 3471 cnt = process_resp_flush(qp, cq, num_entries, wc, 3472 resp->rq_cons); 3473 try_consume_resp_cqe(cq, qp, resp, update); 3474 } else { 3475 cnt = process_resp_one(dev, qp, cq, wc, resp); 3476 consume_cqe(cq); 3477 *update |= 1; 3478 } 3479 3480 return cnt; 3481 } 3482 3483 static void try_consume_req_cqe(struct qedr_cq *cq, struct qedr_qp *qp, 3484 struct rdma_cqe_requester *req, int *update) 3485 { 3486 if (le16_to_cpu(req->sq_cons) == qp->sq.wqe_cons) { 3487 consume_cqe(cq); 3488 *update |= 1; 3489 } 3490 } 3491 3492 int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 3493 { 3494 struct qedr_dev *dev = get_qedr_dev(ibcq->device); 3495 struct qedr_cq *cq = get_qedr_cq(ibcq); 3496 union rdma_cqe *cqe = cq->latest_cqe; 3497 u32 old_cons, new_cons; 3498 unsigned long flags; 3499 int update = 0; 3500 int done = 0; 3501 3502 if (cq->destroyed) { 3503 DP_ERR(dev, 3504 "warning: poll was invoked after destroy for cq %p (icid=%d)\n", 3505 cq, cq->icid); 3506 return 0; 3507 } 3508 3509 if (cq->cq_type == QEDR_CQ_TYPE_GSI) 3510 return qedr_gsi_poll_cq(ibcq, num_entries, wc); 3511 3512 spin_lock_irqsave(&cq->cq_lock, flags); 3513 old_cons = qed_chain_get_cons_idx_u32(&cq->pbl); 3514 while (num_entries && is_valid_cqe(cq, cqe)) { 3515 struct qedr_qp *qp; 3516 int cnt = 0; 3517 3518 /* prevent speculative reads of any field of CQE */ 3519 rmb(); 3520 3521 qp = cqe_get_qp(cqe); 3522 if (!qp) { 3523 WARN(1, "Error: CQE QP pointer is NULL. CQE=%p\n", cqe); 3524 break; 3525 } 3526 3527 wc->qp = &qp->ibqp; 3528 3529 switch (cqe_get_type(cqe)) { 3530 case RDMA_CQE_TYPE_REQUESTER: 3531 cnt = qedr_poll_cq_req(dev, qp, cq, num_entries, wc, 3532 &cqe->req); 3533 try_consume_req_cqe(cq, qp, &cqe->req, &update); 3534 break; 3535 case RDMA_CQE_TYPE_RESPONDER_RQ: 3536 cnt = qedr_poll_cq_resp(dev, qp, cq, num_entries, wc, 3537 &cqe->resp, &update); 3538 break; 3539 case RDMA_CQE_TYPE_INVALID: 3540 default: 3541 DP_ERR(dev, "Error: invalid CQE type = %d\n", 3542 cqe_get_type(cqe)); 3543 } 3544 num_entries -= cnt; 3545 wc += cnt; 3546 done += cnt; 3547 3548 cqe = get_cqe(cq); 3549 } 3550 new_cons = qed_chain_get_cons_idx_u32(&cq->pbl); 3551 3552 cq->cq_cons += new_cons - old_cons; 3553 3554 if (update) 3555 /* doorbell notifies abount latest VALID entry, 3556 * but chain already point to the next INVALID one 3557 */ 3558 doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags); 3559 3560 spin_unlock_irqrestore(&cq->cq_lock, flags); 3561 return done; 3562 } 3563 3564 int qedr_process_mad(struct ib_device *ibdev, int process_mad_flags, 3565 u8 port_num, 3566 const struct ib_wc *in_wc, 3567 const struct ib_grh *in_grh, 3568 const struct ib_mad_hdr *mad_hdr, 3569 size_t in_mad_size, struct ib_mad_hdr *out_mad, 3570 size_t *out_mad_size, u16 *out_mad_pkey_index) 3571 { 3572 struct qedr_dev *dev = get_qedr_dev(ibdev); 3573 3574 DP_DEBUG(dev, QEDR_MSG_GSI, 3575 "QEDR_PROCESS_MAD in_mad %x %x %x %x %x %x %x %x\n", 3576 mad_hdr->attr_id, mad_hdr->base_version, mad_hdr->attr_mod, 3577 mad_hdr->class_specific, mad_hdr->class_version, 3578 mad_hdr->method, mad_hdr->mgmt_class, mad_hdr->status); 3579 return IB_MAD_RESULT_SUCCESS; 3580 } 3581 3582 int qedr_port_immutable(struct ib_device *ibdev, u8 port_num, 3583 struct ib_port_immutable *immutable) 3584 { 3585 struct ib_port_attr attr; 3586 int err; 3587 3588 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE | 3589 RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 3590 3591 err = ib_query_port(ibdev, port_num, &attr); 3592 if (err) 3593 return err; 3594 3595 immutable->pkey_tbl_len = attr.pkey_tbl_len; 3596 immutable->gid_tbl_len = attr.gid_tbl_len; 3597 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 3598 3599 return 0; 3600 } 3601