1 /* 2 * Broadcom NetXtreme-E RoCE driver. 3 * 4 * Copyright (c) 2016 - 2017, Broadcom. All rights reserved. The term 5 * Broadcom refers to Broadcom Limited and/or its subsidiaries. 6 * 7 * This software is available to you under a choice of one of two 8 * licenses. You may choose to be licensed under the terms of the GNU 9 * General Public License (GPL) Version 2, available from the file 10 * COPYING in the main directory of this source tree, or the 11 * BSD license below: 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in 21 * the documentation and/or other materials provided with the 22 * distribution. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' 25 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, 26 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 27 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS 28 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 29 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 30 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR 31 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 32 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE 33 * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN 34 * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 35 * 36 * Description: IB Verbs interpreter 37 */ 38 39 #include <linux/interrupt.h> 40 #include <linux/types.h> 41 #include <linux/pci.h> 42 #include <linux/netdevice.h> 43 #include <linux/if_ether.h> 44 45 #include <rdma/ib_verbs.h> 46 #include <rdma/ib_user_verbs.h> 47 #include <rdma/ib_umem.h> 48 #include <rdma/ib_addr.h> 49 #include <rdma/ib_mad.h> 50 #include <rdma/ib_cache.h> 51 #include <rdma/uverbs_ioctl.h> 52 53 #include "bnxt_ulp.h" 54 55 #include "roce_hsi.h" 56 #include "qplib_res.h" 57 #include "qplib_sp.h" 58 #include "qplib_fp.h" 59 #include "qplib_rcfw.h" 60 61 #include "bnxt_re.h" 62 #include "ib_verbs.h" 63 #include <rdma/bnxt_re-abi.h> 64 65 static int __from_ib_access_flags(int iflags) 66 { 67 int qflags = 0; 68 69 if (iflags & IB_ACCESS_LOCAL_WRITE) 70 qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE; 71 if (iflags & IB_ACCESS_REMOTE_READ) 72 qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ; 73 if (iflags & IB_ACCESS_REMOTE_WRITE) 74 qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE; 75 if (iflags & IB_ACCESS_REMOTE_ATOMIC) 76 qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC; 77 if (iflags & IB_ACCESS_MW_BIND) 78 qflags |= BNXT_QPLIB_ACCESS_MW_BIND; 79 if (iflags & IB_ZERO_BASED) 80 qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED; 81 if (iflags & IB_ACCESS_ON_DEMAND) 82 qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND; 83 return qflags; 84 }; 85 86 static enum ib_access_flags __to_ib_access_flags(int qflags) 87 { 88 enum ib_access_flags iflags = 0; 89 90 if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE) 91 iflags |= IB_ACCESS_LOCAL_WRITE; 92 if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE) 93 iflags |= IB_ACCESS_REMOTE_WRITE; 94 if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ) 95 iflags |= IB_ACCESS_REMOTE_READ; 96 if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC) 97 iflags |= IB_ACCESS_REMOTE_ATOMIC; 98 if (qflags & BNXT_QPLIB_ACCESS_MW_BIND) 99 iflags |= IB_ACCESS_MW_BIND; 100 if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED) 101 iflags |= IB_ZERO_BASED; 102 if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND) 103 iflags |= IB_ACCESS_ON_DEMAND; 104 return iflags; 105 }; 106 107 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list, 108 struct bnxt_qplib_sge *sg_list, int num) 109 { 110 int i, total = 0; 111 112 for (i = 0; i < num; i++) { 113 sg_list[i].addr = ib_sg_list[i].addr; 114 sg_list[i].lkey = ib_sg_list[i].lkey; 115 sg_list[i].size = ib_sg_list[i].length; 116 total += sg_list[i].size; 117 } 118 return total; 119 } 120 121 /* Device */ 122 int bnxt_re_query_device(struct ib_device *ibdev, 123 struct ib_device_attr *ib_attr, 124 struct ib_udata *udata) 125 { 126 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 127 struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr; 128 129 memset(ib_attr, 0, sizeof(*ib_attr)); 130 memcpy(&ib_attr->fw_ver, dev_attr->fw_ver, 131 min(sizeof(dev_attr->fw_ver), 132 sizeof(ib_attr->fw_ver))); 133 bnxt_qplib_get_guid(rdev->netdev->dev_addr, 134 (u8 *)&ib_attr->sys_image_guid); 135 ib_attr->max_mr_size = BNXT_RE_MAX_MR_SIZE; 136 ib_attr->page_size_cap = BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M; 137 138 ib_attr->vendor_id = rdev->en_dev->pdev->vendor; 139 ib_attr->vendor_part_id = rdev->en_dev->pdev->device; 140 ib_attr->hw_ver = rdev->en_dev->pdev->subsystem_device; 141 ib_attr->max_qp = dev_attr->max_qp; 142 ib_attr->max_qp_wr = dev_attr->max_qp_wqes; 143 ib_attr->device_cap_flags = 144 IB_DEVICE_CURR_QP_STATE_MOD 145 | IB_DEVICE_RC_RNR_NAK_GEN 146 | IB_DEVICE_SHUTDOWN_PORT 147 | IB_DEVICE_SYS_IMAGE_GUID 148 | IB_DEVICE_LOCAL_DMA_LKEY 149 | IB_DEVICE_RESIZE_MAX_WR 150 | IB_DEVICE_PORT_ACTIVE_EVENT 151 | IB_DEVICE_N_NOTIFY_CQ 152 | IB_DEVICE_MEM_WINDOW 153 | IB_DEVICE_MEM_WINDOW_TYPE_2B 154 | IB_DEVICE_MEM_MGT_EXTENSIONS; 155 ib_attr->max_send_sge = dev_attr->max_qp_sges; 156 ib_attr->max_recv_sge = dev_attr->max_qp_sges; 157 ib_attr->max_sge_rd = dev_attr->max_qp_sges; 158 ib_attr->max_cq = dev_attr->max_cq; 159 ib_attr->max_cqe = dev_attr->max_cq_wqes; 160 ib_attr->max_mr = dev_attr->max_mr; 161 ib_attr->max_pd = dev_attr->max_pd; 162 ib_attr->max_qp_rd_atom = dev_attr->max_qp_rd_atom; 163 ib_attr->max_qp_init_rd_atom = dev_attr->max_qp_init_rd_atom; 164 ib_attr->atomic_cap = IB_ATOMIC_NONE; 165 ib_attr->masked_atomic_cap = IB_ATOMIC_NONE; 166 167 ib_attr->max_ee_rd_atom = 0; 168 ib_attr->max_res_rd_atom = 0; 169 ib_attr->max_ee_init_rd_atom = 0; 170 ib_attr->max_ee = 0; 171 ib_attr->max_rdd = 0; 172 ib_attr->max_mw = dev_attr->max_mw; 173 ib_attr->max_raw_ipv6_qp = 0; 174 ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp; 175 ib_attr->max_mcast_grp = 0; 176 ib_attr->max_mcast_qp_attach = 0; 177 ib_attr->max_total_mcast_qp_attach = 0; 178 ib_attr->max_ah = dev_attr->max_ah; 179 180 ib_attr->max_srq = dev_attr->max_srq; 181 ib_attr->max_srq_wr = dev_attr->max_srq_wqes; 182 ib_attr->max_srq_sge = dev_attr->max_srq_sges; 183 184 ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS; 185 186 ib_attr->max_pkeys = 1; 187 ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY; 188 return 0; 189 } 190 191 /* Port */ 192 int bnxt_re_query_port(struct ib_device *ibdev, u8 port_num, 193 struct ib_port_attr *port_attr) 194 { 195 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 196 struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr; 197 198 memset(port_attr, 0, sizeof(*port_attr)); 199 200 if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) { 201 port_attr->state = IB_PORT_ACTIVE; 202 port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP; 203 } else { 204 port_attr->state = IB_PORT_DOWN; 205 port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED; 206 } 207 port_attr->max_mtu = IB_MTU_4096; 208 port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu); 209 port_attr->gid_tbl_len = dev_attr->max_sgid; 210 port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP | 211 IB_PORT_DEVICE_MGMT_SUP | 212 IB_PORT_VENDOR_CLASS_SUP; 213 port_attr->ip_gids = true; 214 215 port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW; 216 port_attr->bad_pkey_cntr = 0; 217 port_attr->qkey_viol_cntr = 0; 218 port_attr->pkey_tbl_len = dev_attr->max_pkey; 219 port_attr->lid = 0; 220 port_attr->sm_lid = 0; 221 port_attr->lmc = 0; 222 port_attr->max_vl_num = 4; 223 port_attr->sm_sl = 0; 224 port_attr->subnet_timeout = 0; 225 port_attr->init_type_reply = 0; 226 port_attr->active_speed = rdev->active_speed; 227 port_attr->active_width = rdev->active_width; 228 229 return 0; 230 } 231 232 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u8 port_num, 233 struct ib_port_immutable *immutable) 234 { 235 struct ib_port_attr port_attr; 236 237 if (bnxt_re_query_port(ibdev, port_num, &port_attr)) 238 return -EINVAL; 239 240 immutable->pkey_tbl_len = port_attr.pkey_tbl_len; 241 immutable->gid_tbl_len = port_attr.gid_tbl_len; 242 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE; 243 immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP; 244 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 245 return 0; 246 } 247 248 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str) 249 { 250 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 251 252 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d", 253 rdev->dev_attr.fw_ver[0], rdev->dev_attr.fw_ver[1], 254 rdev->dev_attr.fw_ver[2], rdev->dev_attr.fw_ver[3]); 255 } 256 257 int bnxt_re_query_pkey(struct ib_device *ibdev, u8 port_num, 258 u16 index, u16 *pkey) 259 { 260 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 261 262 /* Ignore port_num */ 263 264 memset(pkey, 0, sizeof(*pkey)); 265 return bnxt_qplib_get_pkey(&rdev->qplib_res, 266 &rdev->qplib_res.pkey_tbl, index, pkey); 267 } 268 269 int bnxt_re_query_gid(struct ib_device *ibdev, u8 port_num, 270 int index, union ib_gid *gid) 271 { 272 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 273 int rc = 0; 274 275 /* Ignore port_num */ 276 memset(gid, 0, sizeof(*gid)); 277 rc = bnxt_qplib_get_sgid(&rdev->qplib_res, 278 &rdev->qplib_res.sgid_tbl, index, 279 (struct bnxt_qplib_gid *)gid); 280 return rc; 281 } 282 283 int bnxt_re_del_gid(const struct ib_gid_attr *attr, void **context) 284 { 285 int rc = 0; 286 struct bnxt_re_gid_ctx *ctx, **ctx_tbl; 287 struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev); 288 struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl; 289 struct bnxt_qplib_gid *gid_to_del; 290 u16 vlan_id = 0xFFFF; 291 292 /* Delete the entry from the hardware */ 293 ctx = *context; 294 if (!ctx) 295 return -EINVAL; 296 297 if (sgid_tbl && sgid_tbl->active) { 298 if (ctx->idx >= sgid_tbl->max) 299 return -EINVAL; 300 gid_to_del = &sgid_tbl->tbl[ctx->idx].gid; 301 vlan_id = sgid_tbl->tbl[ctx->idx].vlan_id; 302 /* DEL_GID is called in WQ context(netdevice_event_work_handler) 303 * or via the ib_unregister_device path. In the former case QP1 304 * may not be destroyed yet, in which case just return as FW 305 * needs that entry to be present and will fail it's deletion. 306 * We could get invoked again after QP1 is destroyed OR get an 307 * ADD_GID call with a different GID value for the same index 308 * where we issue MODIFY_GID cmd to update the GID entry -- TBD 309 */ 310 if (ctx->idx == 0 && 311 rdma_link_local_addr((struct in6_addr *)gid_to_del) && 312 ctx->refcnt == 1 && rdev->gsi_ctx.gsi_sqp) { 313 ibdev_dbg(&rdev->ibdev, 314 "Trying to delete GID0 while QP1 is alive\n"); 315 return -EFAULT; 316 } 317 ctx->refcnt--; 318 if (!ctx->refcnt) { 319 rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del, 320 vlan_id, true); 321 if (rc) { 322 ibdev_err(&rdev->ibdev, 323 "Failed to remove GID: %#x", rc); 324 } else { 325 ctx_tbl = sgid_tbl->ctx; 326 ctx_tbl[ctx->idx] = NULL; 327 kfree(ctx); 328 } 329 } 330 } else { 331 return -EINVAL; 332 } 333 return rc; 334 } 335 336 int bnxt_re_add_gid(const struct ib_gid_attr *attr, void **context) 337 { 338 int rc; 339 u32 tbl_idx = 0; 340 u16 vlan_id = 0xFFFF; 341 struct bnxt_re_gid_ctx *ctx, **ctx_tbl; 342 struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev); 343 struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl; 344 345 rc = rdma_read_gid_l2_fields(attr, &vlan_id, NULL); 346 if (rc) 347 return rc; 348 349 rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)&attr->gid, 350 rdev->qplib_res.netdev->dev_addr, 351 vlan_id, true, &tbl_idx); 352 if (rc == -EALREADY) { 353 ctx_tbl = sgid_tbl->ctx; 354 ctx_tbl[tbl_idx]->refcnt++; 355 *context = ctx_tbl[tbl_idx]; 356 return 0; 357 } 358 359 if (rc < 0) { 360 ibdev_err(&rdev->ibdev, "Failed to add GID: %#x", rc); 361 return rc; 362 } 363 364 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 365 if (!ctx) 366 return -ENOMEM; 367 ctx_tbl = sgid_tbl->ctx; 368 ctx->idx = tbl_idx; 369 ctx->refcnt = 1; 370 ctx_tbl[tbl_idx] = ctx; 371 *context = ctx; 372 373 return rc; 374 } 375 376 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev, 377 u8 port_num) 378 { 379 return IB_LINK_LAYER_ETHERNET; 380 } 381 382 #define BNXT_RE_FENCE_PBL_SIZE DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE) 383 384 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd) 385 { 386 struct bnxt_re_fence_data *fence = &pd->fence; 387 struct ib_mr *ib_mr = &fence->mr->ib_mr; 388 struct bnxt_qplib_swqe *wqe = &fence->bind_wqe; 389 390 memset(wqe, 0, sizeof(*wqe)); 391 wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW; 392 wqe->wr_id = BNXT_QPLIB_FENCE_WRID; 393 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 394 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 395 wqe->bind.zero_based = false; 396 wqe->bind.parent_l_key = ib_mr->lkey; 397 wqe->bind.va = (u64)(unsigned long)fence->va; 398 wqe->bind.length = fence->size; 399 wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ); 400 wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1; 401 402 /* Save the initial rkey in fence structure for now; 403 * wqe->bind.r_key will be set at (re)bind time. 404 */ 405 fence->bind_rkey = ib_inc_rkey(fence->mw->rkey); 406 } 407 408 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp) 409 { 410 struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp, 411 qplib_qp); 412 struct ib_pd *ib_pd = qp->ib_qp.pd; 413 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 414 struct bnxt_re_fence_data *fence = &pd->fence; 415 struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe; 416 struct bnxt_qplib_swqe wqe; 417 int rc; 418 419 memcpy(&wqe, fence_wqe, sizeof(wqe)); 420 wqe.bind.r_key = fence->bind_rkey; 421 fence->bind_rkey = ib_inc_rkey(fence->bind_rkey); 422 423 ibdev_dbg(&qp->rdev->ibdev, 424 "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n", 425 wqe.bind.r_key, qp->qplib_qp.id, pd); 426 rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe); 427 if (rc) { 428 ibdev_err(&qp->rdev->ibdev, "Failed to bind fence-WQE\n"); 429 return rc; 430 } 431 bnxt_qplib_post_send_db(&qp->qplib_qp); 432 433 return rc; 434 } 435 436 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd) 437 { 438 struct bnxt_re_fence_data *fence = &pd->fence; 439 struct bnxt_re_dev *rdev = pd->rdev; 440 struct device *dev = &rdev->en_dev->pdev->dev; 441 struct bnxt_re_mr *mr = fence->mr; 442 443 if (fence->mw) { 444 bnxt_re_dealloc_mw(fence->mw); 445 fence->mw = NULL; 446 } 447 if (mr) { 448 if (mr->ib_mr.rkey) 449 bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr, 450 true); 451 if (mr->ib_mr.lkey) 452 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 453 kfree(mr); 454 fence->mr = NULL; 455 } 456 if (fence->dma_addr) { 457 dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES, 458 DMA_BIDIRECTIONAL); 459 fence->dma_addr = 0; 460 } 461 } 462 463 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd) 464 { 465 int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND; 466 struct bnxt_re_fence_data *fence = &pd->fence; 467 struct bnxt_re_dev *rdev = pd->rdev; 468 struct device *dev = &rdev->en_dev->pdev->dev; 469 struct bnxt_re_mr *mr = NULL; 470 dma_addr_t dma_addr = 0; 471 struct ib_mw *mw; 472 u64 pbl_tbl; 473 int rc; 474 475 dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES, 476 DMA_BIDIRECTIONAL); 477 rc = dma_mapping_error(dev, dma_addr); 478 if (rc) { 479 ibdev_err(&rdev->ibdev, "Failed to dma-map fence-MR-mem\n"); 480 rc = -EIO; 481 fence->dma_addr = 0; 482 goto fail; 483 } 484 fence->dma_addr = dma_addr; 485 486 /* Allocate a MR */ 487 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 488 if (!mr) { 489 rc = -ENOMEM; 490 goto fail; 491 } 492 fence->mr = mr; 493 mr->rdev = rdev; 494 mr->qplib_mr.pd = &pd->qplib_pd; 495 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR; 496 mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags); 497 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 498 if (rc) { 499 ibdev_err(&rdev->ibdev, "Failed to alloc fence-HW-MR\n"); 500 goto fail; 501 } 502 503 /* Register MR */ 504 mr->ib_mr.lkey = mr->qplib_mr.lkey; 505 mr->qplib_mr.va = (u64)(unsigned long)fence->va; 506 mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES; 507 pbl_tbl = dma_addr; 508 rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl_tbl, 509 BNXT_RE_FENCE_PBL_SIZE, false, PAGE_SIZE); 510 if (rc) { 511 ibdev_err(&rdev->ibdev, "Failed to register fence-MR\n"); 512 goto fail; 513 } 514 mr->ib_mr.rkey = mr->qplib_mr.rkey; 515 516 /* Create a fence MW only for kernel consumers */ 517 mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL); 518 if (IS_ERR(mw)) { 519 ibdev_err(&rdev->ibdev, 520 "Failed to create fence-MW for PD: %p\n", pd); 521 rc = PTR_ERR(mw); 522 goto fail; 523 } 524 fence->mw = mw; 525 526 bnxt_re_create_fence_wqe(pd); 527 return 0; 528 529 fail: 530 bnxt_re_destroy_fence_mr(pd); 531 return rc; 532 } 533 534 /* Protection Domains */ 535 void bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata) 536 { 537 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 538 struct bnxt_re_dev *rdev = pd->rdev; 539 540 bnxt_re_destroy_fence_mr(pd); 541 542 if (pd->qplib_pd.id) 543 bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl, 544 &pd->qplib_pd); 545 } 546 547 int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 548 { 549 struct ib_device *ibdev = ibpd->device; 550 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 551 struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context( 552 udata, struct bnxt_re_ucontext, ib_uctx); 553 struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd); 554 int rc; 555 556 pd->rdev = rdev; 557 if (bnxt_qplib_alloc_pd(&rdev->qplib_res.pd_tbl, &pd->qplib_pd)) { 558 ibdev_err(&rdev->ibdev, "Failed to allocate HW PD"); 559 rc = -ENOMEM; 560 goto fail; 561 } 562 563 if (udata) { 564 struct bnxt_re_pd_resp resp; 565 566 if (!ucntx->dpi.dbr) { 567 /* Allocate DPI in alloc_pd to avoid failing of 568 * ibv_devinfo and family of application when DPIs 569 * are depleted. 570 */ 571 if (bnxt_qplib_alloc_dpi(&rdev->qplib_res.dpi_tbl, 572 &ucntx->dpi, ucntx)) { 573 rc = -ENOMEM; 574 goto dbfail; 575 } 576 } 577 578 resp.pdid = pd->qplib_pd.id; 579 /* Still allow mapping this DBR to the new user PD. */ 580 resp.dpi = ucntx->dpi.dpi; 581 resp.dbr = (u64)ucntx->dpi.umdbr; 582 583 rc = ib_copy_to_udata(udata, &resp, sizeof(resp)); 584 if (rc) { 585 ibdev_err(&rdev->ibdev, 586 "Failed to copy user response\n"); 587 goto dbfail; 588 } 589 } 590 591 if (!udata) 592 if (bnxt_re_create_fence_mr(pd)) 593 ibdev_warn(&rdev->ibdev, 594 "Failed to create Fence-MR\n"); 595 return 0; 596 dbfail: 597 bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl, 598 &pd->qplib_pd); 599 fail: 600 return rc; 601 } 602 603 /* Address Handles */ 604 void bnxt_re_destroy_ah(struct ib_ah *ib_ah, u32 flags) 605 { 606 struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah); 607 struct bnxt_re_dev *rdev = ah->rdev; 608 609 bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah, 610 !(flags & RDMA_DESTROY_AH_SLEEPABLE)); 611 } 612 613 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype) 614 { 615 u8 nw_type; 616 617 switch (ntype) { 618 case RDMA_NETWORK_IPV4: 619 nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4; 620 break; 621 case RDMA_NETWORK_IPV6: 622 nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6; 623 break; 624 default: 625 nw_type = CMDQ_CREATE_AH_TYPE_V1; 626 break; 627 } 628 return nw_type; 629 } 630 631 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr, 632 struct ib_udata *udata) 633 { 634 struct ib_pd *ib_pd = ib_ah->pd; 635 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 636 struct rdma_ah_attr *ah_attr = init_attr->ah_attr; 637 const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr); 638 struct bnxt_re_dev *rdev = pd->rdev; 639 const struct ib_gid_attr *sgid_attr; 640 struct bnxt_re_gid_ctx *ctx; 641 struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah); 642 u8 nw_type; 643 int rc; 644 645 if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) { 646 ibdev_err(&rdev->ibdev, "Failed to alloc AH: GRH not set"); 647 return -EINVAL; 648 } 649 650 ah->rdev = rdev; 651 ah->qplib_ah.pd = &pd->qplib_pd; 652 653 /* Supply the configuration for the HW */ 654 memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw, 655 sizeof(union ib_gid)); 656 sgid_attr = grh->sgid_attr; 657 /* Get the HW context of the GID. The reference 658 * of GID table entry is already taken by the caller. 659 */ 660 ctx = rdma_read_gid_hw_context(sgid_attr); 661 ah->qplib_ah.sgid_index = ctx->idx; 662 ah->qplib_ah.host_sgid_index = grh->sgid_index; 663 ah->qplib_ah.traffic_class = grh->traffic_class; 664 ah->qplib_ah.flow_label = grh->flow_label; 665 ah->qplib_ah.hop_limit = grh->hop_limit; 666 ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr); 667 668 /* Get network header type for this GID */ 669 nw_type = rdma_gid_attr_network_type(sgid_attr); 670 ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type); 671 672 memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN); 673 rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, 674 !(init_attr->flags & 675 RDMA_CREATE_AH_SLEEPABLE)); 676 if (rc) { 677 ibdev_err(&rdev->ibdev, "Failed to allocate HW AH"); 678 return rc; 679 } 680 681 /* Write AVID to shared page. */ 682 if (udata) { 683 struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context( 684 udata, struct bnxt_re_ucontext, ib_uctx); 685 unsigned long flag; 686 u32 *wrptr; 687 688 spin_lock_irqsave(&uctx->sh_lock, flag); 689 wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT); 690 *wrptr = ah->qplib_ah.id; 691 wmb(); /* make sure cache is updated. */ 692 spin_unlock_irqrestore(&uctx->sh_lock, flag); 693 } 694 695 return 0; 696 } 697 698 int bnxt_re_modify_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr) 699 { 700 return 0; 701 } 702 703 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr) 704 { 705 struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah); 706 707 ah_attr->type = ib_ah->type; 708 rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl); 709 memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN); 710 rdma_ah_set_grh(ah_attr, NULL, 0, 711 ah->qplib_ah.host_sgid_index, 712 0, ah->qplib_ah.traffic_class); 713 rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data); 714 rdma_ah_set_port_num(ah_attr, 1); 715 rdma_ah_set_static_rate(ah_attr, 0); 716 return 0; 717 } 718 719 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp) 720 __acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock) 721 { 722 unsigned long flags; 723 724 spin_lock_irqsave(&qp->scq->cq_lock, flags); 725 if (qp->rcq != qp->scq) 726 spin_lock(&qp->rcq->cq_lock); 727 else 728 __acquire(&qp->rcq->cq_lock); 729 730 return flags; 731 } 732 733 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp, 734 unsigned long flags) 735 __releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock) 736 { 737 if (qp->rcq != qp->scq) 738 spin_unlock(&qp->rcq->cq_lock); 739 else 740 __release(&qp->rcq->cq_lock); 741 spin_unlock_irqrestore(&qp->scq->cq_lock, flags); 742 } 743 744 static int bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp *qp) 745 { 746 struct bnxt_re_qp *gsi_sqp; 747 struct bnxt_re_ah *gsi_sah; 748 struct bnxt_re_dev *rdev; 749 int rc = 0; 750 751 rdev = qp->rdev; 752 gsi_sqp = rdev->gsi_ctx.gsi_sqp; 753 gsi_sah = rdev->gsi_ctx.gsi_sah; 754 755 ibdev_dbg(&rdev->ibdev, "Destroy the shadow AH\n"); 756 bnxt_qplib_destroy_ah(&rdev->qplib_res, 757 &gsi_sah->qplib_ah, 758 true); 759 bnxt_qplib_clean_qp(&qp->qplib_qp); 760 761 ibdev_dbg(&rdev->ibdev, "Destroy the shadow QP\n"); 762 rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &gsi_sqp->qplib_qp); 763 if (rc) { 764 ibdev_err(&rdev->ibdev, "Destroy Shadow QP failed"); 765 goto fail; 766 } 767 bnxt_qplib_free_qp_res(&rdev->qplib_res, &gsi_sqp->qplib_qp); 768 769 /* remove from active qp list */ 770 mutex_lock(&rdev->qp_lock); 771 list_del(&gsi_sqp->list); 772 mutex_unlock(&rdev->qp_lock); 773 atomic_dec(&rdev->qp_count); 774 775 kfree(rdev->gsi_ctx.sqp_tbl); 776 kfree(gsi_sah); 777 kfree(gsi_sqp); 778 rdev->gsi_ctx.gsi_sqp = NULL; 779 rdev->gsi_ctx.gsi_sah = NULL; 780 rdev->gsi_ctx.sqp_tbl = NULL; 781 782 return 0; 783 fail: 784 return rc; 785 } 786 787 /* Queue Pairs */ 788 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata) 789 { 790 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 791 struct bnxt_re_dev *rdev = qp->rdev; 792 unsigned int flags; 793 int rc; 794 795 bnxt_qplib_flush_cqn_wq(&qp->qplib_qp); 796 797 rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp); 798 if (rc) { 799 ibdev_err(&rdev->ibdev, "Failed to destroy HW QP"); 800 return rc; 801 } 802 803 if (rdma_is_kernel_res(&qp->ib_qp.res)) { 804 flags = bnxt_re_lock_cqs(qp); 805 bnxt_qplib_clean_qp(&qp->qplib_qp); 806 bnxt_re_unlock_cqs(qp, flags); 807 } 808 809 bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp); 810 811 if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) { 812 rc = bnxt_re_destroy_gsi_sqp(qp); 813 if (rc) 814 goto sh_fail; 815 } 816 817 mutex_lock(&rdev->qp_lock); 818 list_del(&qp->list); 819 mutex_unlock(&rdev->qp_lock); 820 atomic_dec(&rdev->qp_count); 821 822 ib_umem_release(qp->rumem); 823 ib_umem_release(qp->sumem); 824 825 kfree(qp); 826 return 0; 827 sh_fail: 828 return rc; 829 } 830 831 static u8 __from_ib_qp_type(enum ib_qp_type type) 832 { 833 switch (type) { 834 case IB_QPT_GSI: 835 return CMDQ_CREATE_QP1_TYPE_GSI; 836 case IB_QPT_RC: 837 return CMDQ_CREATE_QP_TYPE_RC; 838 case IB_QPT_UD: 839 return CMDQ_CREATE_QP_TYPE_UD; 840 default: 841 return IB_QPT_MAX; 842 } 843 } 844 845 static u16 bnxt_re_setup_rwqe_size(struct bnxt_qplib_qp *qplqp, 846 int rsge, int max) 847 { 848 if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) 849 rsge = max; 850 return bnxt_re_get_rwqe_size(rsge); 851 } 852 853 static u16 bnxt_re_get_wqe_size(int ilsize, int nsge) 854 { 855 u16 wqe_size, calc_ils; 856 857 wqe_size = bnxt_re_get_swqe_size(nsge); 858 if (ilsize) { 859 calc_ils = sizeof(struct sq_send_hdr) + ilsize; 860 wqe_size = max_t(u16, calc_ils, wqe_size); 861 wqe_size = ALIGN(wqe_size, sizeof(struct sq_send_hdr)); 862 } 863 return wqe_size; 864 } 865 866 static int bnxt_re_setup_swqe_size(struct bnxt_re_qp *qp, 867 struct ib_qp_init_attr *init_attr) 868 { 869 struct bnxt_qplib_dev_attr *dev_attr; 870 struct bnxt_qplib_qp *qplqp; 871 struct bnxt_re_dev *rdev; 872 struct bnxt_qplib_q *sq; 873 int align, ilsize; 874 875 rdev = qp->rdev; 876 qplqp = &qp->qplib_qp; 877 sq = &qplqp->sq; 878 dev_attr = &rdev->dev_attr; 879 880 align = sizeof(struct sq_send_hdr); 881 ilsize = ALIGN(init_attr->cap.max_inline_data, align); 882 883 sq->wqe_size = bnxt_re_get_wqe_size(ilsize, sq->max_sge); 884 if (sq->wqe_size > bnxt_re_get_swqe_size(dev_attr->max_qp_sges)) 885 return -EINVAL; 886 /* For gen p4 and gen p5 backward compatibility mode 887 * wqe size is fixed to 128 bytes 888 */ 889 if (sq->wqe_size < bnxt_re_get_swqe_size(dev_attr->max_qp_sges) && 890 qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) 891 sq->wqe_size = bnxt_re_get_swqe_size(dev_attr->max_qp_sges); 892 893 if (init_attr->cap.max_inline_data) { 894 qplqp->max_inline_data = sq->wqe_size - 895 sizeof(struct sq_send_hdr); 896 init_attr->cap.max_inline_data = qplqp->max_inline_data; 897 if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) 898 sq->max_sge = qplqp->max_inline_data / 899 sizeof(struct sq_sge); 900 } 901 902 return 0; 903 } 904 905 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd, 906 struct bnxt_re_qp *qp, struct ib_udata *udata) 907 { 908 struct bnxt_qplib_qp *qplib_qp; 909 struct bnxt_re_ucontext *cntx; 910 struct bnxt_re_qp_req ureq; 911 int bytes = 0, psn_sz; 912 struct ib_umem *umem; 913 int psn_nume; 914 915 qplib_qp = &qp->qplib_qp; 916 cntx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, 917 ib_uctx); 918 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) 919 return -EFAULT; 920 921 bytes = (qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size); 922 /* Consider mapping PSN search memory only for RC QPs. */ 923 if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) { 924 psn_sz = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ? 925 sizeof(struct sq_psn_search_ext) : 926 sizeof(struct sq_psn_search); 927 psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ? 928 qplib_qp->sq.max_wqe : 929 ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) / 930 sizeof(struct bnxt_qplib_sge)); 931 bytes += (psn_nume * psn_sz); 932 } 933 934 bytes = PAGE_ALIGN(bytes); 935 umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes, 936 IB_ACCESS_LOCAL_WRITE); 937 if (IS_ERR(umem)) 938 return PTR_ERR(umem); 939 940 qp->sumem = umem; 941 qplib_qp->sq.sg_info.sghead = umem->sg_head.sgl; 942 qplib_qp->sq.sg_info.npages = ib_umem_num_pages(umem); 943 qplib_qp->sq.sg_info.nmap = umem->nmap; 944 qplib_qp->sq.sg_info.pgsize = PAGE_SIZE; 945 qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT; 946 qplib_qp->qp_handle = ureq.qp_handle; 947 948 if (!qp->qplib_qp.srq) { 949 bytes = (qplib_qp->rq.max_wqe * qplib_qp->rq.wqe_size); 950 bytes = PAGE_ALIGN(bytes); 951 umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes, 952 IB_ACCESS_LOCAL_WRITE); 953 if (IS_ERR(umem)) 954 goto rqfail; 955 qp->rumem = umem; 956 qplib_qp->rq.sg_info.sghead = umem->sg_head.sgl; 957 qplib_qp->rq.sg_info.npages = ib_umem_num_pages(umem); 958 qplib_qp->rq.sg_info.nmap = umem->nmap; 959 qplib_qp->rq.sg_info.pgsize = PAGE_SIZE; 960 qplib_qp->rq.sg_info.pgshft = PAGE_SHIFT; 961 } 962 963 qplib_qp->dpi = &cntx->dpi; 964 return 0; 965 rqfail: 966 ib_umem_release(qp->sumem); 967 qp->sumem = NULL; 968 memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info)); 969 970 return PTR_ERR(umem); 971 } 972 973 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah 974 (struct bnxt_re_pd *pd, 975 struct bnxt_qplib_res *qp1_res, 976 struct bnxt_qplib_qp *qp1_qp) 977 { 978 struct bnxt_re_dev *rdev = pd->rdev; 979 struct bnxt_re_ah *ah; 980 union ib_gid sgid; 981 int rc; 982 983 ah = kzalloc(sizeof(*ah), GFP_KERNEL); 984 if (!ah) 985 return NULL; 986 987 ah->rdev = rdev; 988 ah->qplib_ah.pd = &pd->qplib_pd; 989 990 rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid); 991 if (rc) 992 goto fail; 993 994 /* supply the dgid data same as sgid */ 995 memcpy(ah->qplib_ah.dgid.data, &sgid.raw, 996 sizeof(union ib_gid)); 997 ah->qplib_ah.sgid_index = 0; 998 999 ah->qplib_ah.traffic_class = 0; 1000 ah->qplib_ah.flow_label = 0; 1001 ah->qplib_ah.hop_limit = 1; 1002 ah->qplib_ah.sl = 0; 1003 /* Have DMAC same as SMAC */ 1004 ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr); 1005 1006 rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false); 1007 if (rc) { 1008 ibdev_err(&rdev->ibdev, 1009 "Failed to allocate HW AH for Shadow QP"); 1010 goto fail; 1011 } 1012 1013 return ah; 1014 1015 fail: 1016 kfree(ah); 1017 return NULL; 1018 } 1019 1020 static struct bnxt_re_qp *bnxt_re_create_shadow_qp 1021 (struct bnxt_re_pd *pd, 1022 struct bnxt_qplib_res *qp1_res, 1023 struct bnxt_qplib_qp *qp1_qp) 1024 { 1025 struct bnxt_re_dev *rdev = pd->rdev; 1026 struct bnxt_re_qp *qp; 1027 int rc; 1028 1029 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 1030 if (!qp) 1031 return NULL; 1032 1033 qp->rdev = rdev; 1034 1035 /* Initialize the shadow QP structure from the QP1 values */ 1036 ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr); 1037 1038 qp->qplib_qp.pd = &pd->qplib_pd; 1039 qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp); 1040 qp->qplib_qp.type = IB_QPT_UD; 1041 1042 qp->qplib_qp.max_inline_data = 0; 1043 qp->qplib_qp.sig_type = true; 1044 1045 /* Shadow QP SQ depth should be same as QP1 RQ depth */ 1046 qp->qplib_qp.sq.wqe_size = bnxt_re_get_wqe_size(0, 6); 1047 qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe; 1048 qp->qplib_qp.sq.max_sge = 2; 1049 /* Q full delta can be 1 since it is internal QP */ 1050 qp->qplib_qp.sq.q_full_delta = 1; 1051 qp->qplib_qp.sq.sg_info.pgsize = PAGE_SIZE; 1052 qp->qplib_qp.sq.sg_info.pgshft = PAGE_SHIFT; 1053 1054 qp->qplib_qp.scq = qp1_qp->scq; 1055 qp->qplib_qp.rcq = qp1_qp->rcq; 1056 1057 qp->qplib_qp.rq.wqe_size = bnxt_re_get_rwqe_size(6); 1058 qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe; 1059 qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge; 1060 /* Q full delta can be 1 since it is internal QP */ 1061 qp->qplib_qp.rq.q_full_delta = 1; 1062 qp->qplib_qp.rq.sg_info.pgsize = PAGE_SIZE; 1063 qp->qplib_qp.rq.sg_info.pgshft = PAGE_SHIFT; 1064 1065 qp->qplib_qp.mtu = qp1_qp->mtu; 1066 1067 qp->qplib_qp.sq_hdr_buf_size = 0; 1068 qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6; 1069 qp->qplib_qp.dpi = &rdev->dpi_privileged; 1070 1071 rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp); 1072 if (rc) 1073 goto fail; 1074 1075 spin_lock_init(&qp->sq_lock); 1076 INIT_LIST_HEAD(&qp->list); 1077 mutex_lock(&rdev->qp_lock); 1078 list_add_tail(&qp->list, &rdev->qp_list); 1079 atomic_inc(&rdev->qp_count); 1080 mutex_unlock(&rdev->qp_lock); 1081 return qp; 1082 fail: 1083 kfree(qp); 1084 return NULL; 1085 } 1086 1087 static int bnxt_re_init_rq_attr(struct bnxt_re_qp *qp, 1088 struct ib_qp_init_attr *init_attr) 1089 { 1090 struct bnxt_qplib_dev_attr *dev_attr; 1091 struct bnxt_qplib_qp *qplqp; 1092 struct bnxt_re_dev *rdev; 1093 struct bnxt_qplib_q *rq; 1094 int entries; 1095 1096 rdev = qp->rdev; 1097 qplqp = &qp->qplib_qp; 1098 rq = &qplqp->rq; 1099 dev_attr = &rdev->dev_attr; 1100 1101 if (init_attr->srq) { 1102 struct bnxt_re_srq *srq; 1103 1104 srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq); 1105 if (!srq) { 1106 ibdev_err(&rdev->ibdev, "SRQ not found"); 1107 return -EINVAL; 1108 } 1109 qplqp->srq = &srq->qplib_srq; 1110 rq->max_wqe = 0; 1111 } else { 1112 rq->max_sge = init_attr->cap.max_recv_sge; 1113 if (rq->max_sge > dev_attr->max_qp_sges) 1114 rq->max_sge = dev_attr->max_qp_sges; 1115 init_attr->cap.max_recv_sge = rq->max_sge; 1116 rq->wqe_size = bnxt_re_setup_rwqe_size(qplqp, rq->max_sge, 1117 dev_attr->max_qp_sges); 1118 /* Allocate 1 more than what's provided so posting max doesn't 1119 * mean empty. 1120 */ 1121 entries = roundup_pow_of_two(init_attr->cap.max_recv_wr + 1); 1122 rq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + 1); 1123 rq->q_full_delta = 0; 1124 rq->sg_info.pgsize = PAGE_SIZE; 1125 rq->sg_info.pgshft = PAGE_SHIFT; 1126 } 1127 1128 return 0; 1129 } 1130 1131 static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp) 1132 { 1133 struct bnxt_qplib_dev_attr *dev_attr; 1134 struct bnxt_qplib_qp *qplqp; 1135 struct bnxt_re_dev *rdev; 1136 1137 rdev = qp->rdev; 1138 qplqp = &qp->qplib_qp; 1139 dev_attr = &rdev->dev_attr; 1140 1141 if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) { 1142 qplqp->rq.max_sge = dev_attr->max_qp_sges; 1143 if (qplqp->rq.max_sge > dev_attr->max_qp_sges) 1144 qplqp->rq.max_sge = dev_attr->max_qp_sges; 1145 qplqp->rq.max_sge = 6; 1146 } 1147 } 1148 1149 static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp, 1150 struct ib_qp_init_attr *init_attr, 1151 struct ib_udata *udata) 1152 { 1153 struct bnxt_qplib_dev_attr *dev_attr; 1154 struct bnxt_qplib_qp *qplqp; 1155 struct bnxt_re_dev *rdev; 1156 struct bnxt_qplib_q *sq; 1157 int entries; 1158 int diff; 1159 int rc; 1160 1161 rdev = qp->rdev; 1162 qplqp = &qp->qplib_qp; 1163 sq = &qplqp->sq; 1164 dev_attr = &rdev->dev_attr; 1165 1166 sq->max_sge = init_attr->cap.max_send_sge; 1167 if (sq->max_sge > dev_attr->max_qp_sges) { 1168 sq->max_sge = dev_attr->max_qp_sges; 1169 init_attr->cap.max_send_sge = sq->max_sge; 1170 } 1171 1172 rc = bnxt_re_setup_swqe_size(qp, init_attr); 1173 if (rc) 1174 return rc; 1175 1176 entries = init_attr->cap.max_send_wr; 1177 /* Allocate 128 + 1 more than what's provided */ 1178 diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ? 1179 0 : BNXT_QPLIB_RESERVED_QP_WRS; 1180 entries = roundup_pow_of_two(entries + diff + 1); 1181 sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1); 1182 sq->q_full_delta = diff + 1; 1183 /* 1184 * Reserving one slot for Phantom WQE. Application can 1185 * post one extra entry in this case. But allowing this to avoid 1186 * unexpected Queue full condition 1187 */ 1188 qplqp->sq.q_full_delta -= 1; 1189 qplqp->sq.sg_info.pgsize = PAGE_SIZE; 1190 qplqp->sq.sg_info.pgshft = PAGE_SHIFT; 1191 1192 return 0; 1193 } 1194 1195 static void bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp *qp, 1196 struct ib_qp_init_attr *init_attr) 1197 { 1198 struct bnxt_qplib_dev_attr *dev_attr; 1199 struct bnxt_qplib_qp *qplqp; 1200 struct bnxt_re_dev *rdev; 1201 int entries; 1202 1203 rdev = qp->rdev; 1204 qplqp = &qp->qplib_qp; 1205 dev_attr = &rdev->dev_attr; 1206 1207 if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) { 1208 entries = roundup_pow_of_two(init_attr->cap.max_send_wr + 1); 1209 qplqp->sq.max_wqe = min_t(u32, entries, 1210 dev_attr->max_qp_wqes + 1); 1211 qplqp->sq.q_full_delta = qplqp->sq.max_wqe - 1212 init_attr->cap.max_send_wr; 1213 qplqp->sq.max_sge++; /* Need one extra sge to put UD header */ 1214 if (qplqp->sq.max_sge > dev_attr->max_qp_sges) 1215 qplqp->sq.max_sge = dev_attr->max_qp_sges; 1216 } 1217 } 1218 1219 static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev, 1220 struct ib_qp_init_attr *init_attr) 1221 { 1222 struct bnxt_qplib_chip_ctx *chip_ctx; 1223 int qptype; 1224 1225 chip_ctx = rdev->chip_ctx; 1226 1227 qptype = __from_ib_qp_type(init_attr->qp_type); 1228 if (qptype == IB_QPT_MAX) { 1229 ibdev_err(&rdev->ibdev, "QP type 0x%x not supported", qptype); 1230 qptype = -EOPNOTSUPP; 1231 goto out; 1232 } 1233 1234 if (bnxt_qplib_is_chip_gen_p5(chip_ctx) && 1235 init_attr->qp_type == IB_QPT_GSI) 1236 qptype = CMDQ_CREATE_QP_TYPE_GSI; 1237 out: 1238 return qptype; 1239 } 1240 1241 static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, 1242 struct ib_qp_init_attr *init_attr, 1243 struct ib_udata *udata) 1244 { 1245 struct bnxt_qplib_dev_attr *dev_attr; 1246 struct bnxt_qplib_qp *qplqp; 1247 struct bnxt_re_dev *rdev; 1248 struct bnxt_re_cq *cq; 1249 int rc = 0, qptype; 1250 1251 rdev = qp->rdev; 1252 qplqp = &qp->qplib_qp; 1253 dev_attr = &rdev->dev_attr; 1254 1255 /* Setup misc params */ 1256 ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr); 1257 qplqp->pd = &pd->qplib_pd; 1258 qplqp->qp_handle = (u64)qplqp; 1259 qplqp->max_inline_data = init_attr->cap.max_inline_data; 1260 qplqp->sig_type = ((init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1261 true : false); 1262 qptype = bnxt_re_init_qp_type(rdev, init_attr); 1263 if (qptype < 0) { 1264 rc = qptype; 1265 goto out; 1266 } 1267 qplqp->type = (u8)qptype; 1268 qplqp->wqe_mode = rdev->chip_ctx->modes.wqe_mode; 1269 1270 if (init_attr->qp_type == IB_QPT_RC) { 1271 qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom; 1272 qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom; 1273 } 1274 qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu)); 1275 qplqp->dpi = &rdev->dpi_privileged; /* Doorbell page */ 1276 if (init_attr->create_flags) 1277 ibdev_dbg(&rdev->ibdev, 1278 "QP create flags 0x%x not supported", 1279 init_attr->create_flags); 1280 1281 /* Setup CQs */ 1282 if (init_attr->send_cq) { 1283 cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq); 1284 if (!cq) { 1285 ibdev_err(&rdev->ibdev, "Send CQ not found"); 1286 rc = -EINVAL; 1287 goto out; 1288 } 1289 qplqp->scq = &cq->qplib_cq; 1290 qp->scq = cq; 1291 } 1292 1293 if (init_attr->recv_cq) { 1294 cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq); 1295 if (!cq) { 1296 ibdev_err(&rdev->ibdev, "Receive CQ not found"); 1297 rc = -EINVAL; 1298 goto out; 1299 } 1300 qplqp->rcq = &cq->qplib_cq; 1301 qp->rcq = cq; 1302 } 1303 1304 /* Setup RQ/SRQ */ 1305 rc = bnxt_re_init_rq_attr(qp, init_attr); 1306 if (rc) 1307 goto out; 1308 if (init_attr->qp_type == IB_QPT_GSI) 1309 bnxt_re_adjust_gsi_rq_attr(qp); 1310 1311 /* Setup SQ */ 1312 rc = bnxt_re_init_sq_attr(qp, init_attr, udata); 1313 if (rc) 1314 goto out; 1315 if (init_attr->qp_type == IB_QPT_GSI) 1316 bnxt_re_adjust_gsi_sq_attr(qp, init_attr); 1317 1318 if (udata) /* This will update DPI and qp_handle */ 1319 rc = bnxt_re_init_user_qp(rdev, pd, qp, udata); 1320 out: 1321 return rc; 1322 } 1323 1324 static int bnxt_re_create_shadow_gsi(struct bnxt_re_qp *qp, 1325 struct bnxt_re_pd *pd) 1326 { 1327 struct bnxt_re_sqp_entries *sqp_tbl = NULL; 1328 struct bnxt_re_dev *rdev; 1329 struct bnxt_re_qp *sqp; 1330 struct bnxt_re_ah *sah; 1331 int rc = 0; 1332 1333 rdev = qp->rdev; 1334 /* Create a shadow QP to handle the QP1 traffic */ 1335 sqp_tbl = kzalloc(sizeof(*sqp_tbl) * BNXT_RE_MAX_GSI_SQP_ENTRIES, 1336 GFP_KERNEL); 1337 if (!sqp_tbl) 1338 return -ENOMEM; 1339 rdev->gsi_ctx.sqp_tbl = sqp_tbl; 1340 1341 sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res, &qp->qplib_qp); 1342 if (!sqp) { 1343 rc = -ENODEV; 1344 ibdev_err(&rdev->ibdev, "Failed to create Shadow QP for QP1"); 1345 goto out; 1346 } 1347 rdev->gsi_ctx.gsi_sqp = sqp; 1348 1349 sqp->rcq = qp->rcq; 1350 sqp->scq = qp->scq; 1351 sah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res, 1352 &qp->qplib_qp); 1353 if (!sah) { 1354 bnxt_qplib_destroy_qp(&rdev->qplib_res, 1355 &sqp->qplib_qp); 1356 rc = -ENODEV; 1357 ibdev_err(&rdev->ibdev, 1358 "Failed to create AH entry for ShadowQP"); 1359 goto out; 1360 } 1361 rdev->gsi_ctx.gsi_sah = sah; 1362 1363 return 0; 1364 out: 1365 kfree(sqp_tbl); 1366 return rc; 1367 } 1368 1369 static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, 1370 struct ib_qp_init_attr *init_attr) 1371 { 1372 struct bnxt_re_dev *rdev; 1373 struct bnxt_qplib_qp *qplqp; 1374 int rc = 0; 1375 1376 rdev = qp->rdev; 1377 qplqp = &qp->qplib_qp; 1378 1379 qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2; 1380 qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2; 1381 1382 rc = bnxt_qplib_create_qp1(&rdev->qplib_res, qplqp); 1383 if (rc) { 1384 ibdev_err(&rdev->ibdev, "create HW QP1 failed!"); 1385 goto out; 1386 } 1387 1388 rc = bnxt_re_create_shadow_gsi(qp, pd); 1389 out: 1390 return rc; 1391 } 1392 1393 static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev, 1394 struct ib_qp_init_attr *init_attr, 1395 struct bnxt_qplib_dev_attr *dev_attr) 1396 { 1397 bool rc = true; 1398 1399 if (init_attr->cap.max_send_wr > dev_attr->max_qp_wqes || 1400 init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes || 1401 init_attr->cap.max_send_sge > dev_attr->max_qp_sges || 1402 init_attr->cap.max_recv_sge > dev_attr->max_qp_sges || 1403 init_attr->cap.max_inline_data > dev_attr->max_inline_data) { 1404 ibdev_err(&rdev->ibdev, 1405 "Create QP failed - max exceeded! 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x", 1406 init_attr->cap.max_send_wr, dev_attr->max_qp_wqes, 1407 init_attr->cap.max_recv_wr, dev_attr->max_qp_wqes, 1408 init_attr->cap.max_send_sge, dev_attr->max_qp_sges, 1409 init_attr->cap.max_recv_sge, dev_attr->max_qp_sges, 1410 init_attr->cap.max_inline_data, 1411 dev_attr->max_inline_data); 1412 rc = false; 1413 } 1414 return rc; 1415 } 1416 1417 struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd, 1418 struct ib_qp_init_attr *qp_init_attr, 1419 struct ib_udata *udata) 1420 { 1421 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 1422 struct bnxt_re_dev *rdev = pd->rdev; 1423 struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr; 1424 struct bnxt_re_qp *qp; 1425 int rc; 1426 1427 rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr); 1428 if (!rc) { 1429 rc = -EINVAL; 1430 goto exit; 1431 } 1432 1433 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 1434 if (!qp) { 1435 rc = -ENOMEM; 1436 goto exit; 1437 } 1438 qp->rdev = rdev; 1439 rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, udata); 1440 if (rc) 1441 goto fail; 1442 1443 if (qp_init_attr->qp_type == IB_QPT_GSI && 1444 !(bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx))) { 1445 rc = bnxt_re_create_gsi_qp(qp, pd, qp_init_attr); 1446 if (rc == -ENODEV) 1447 goto qp_destroy; 1448 if (rc) 1449 goto fail; 1450 } else { 1451 rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp); 1452 if (rc) { 1453 ibdev_err(&rdev->ibdev, "Failed to create HW QP"); 1454 goto free_umem; 1455 } 1456 if (udata) { 1457 struct bnxt_re_qp_resp resp; 1458 1459 resp.qpid = qp->qplib_qp.id; 1460 resp.rsvd = 0; 1461 rc = ib_copy_to_udata(udata, &resp, sizeof(resp)); 1462 if (rc) { 1463 ibdev_err(&rdev->ibdev, "Failed to copy QP udata"); 1464 goto qp_destroy; 1465 } 1466 } 1467 } 1468 1469 qp->ib_qp.qp_num = qp->qplib_qp.id; 1470 if (qp_init_attr->qp_type == IB_QPT_GSI) 1471 rdev->gsi_ctx.gsi_qp = qp; 1472 spin_lock_init(&qp->sq_lock); 1473 spin_lock_init(&qp->rq_lock); 1474 INIT_LIST_HEAD(&qp->list); 1475 mutex_lock(&rdev->qp_lock); 1476 list_add_tail(&qp->list, &rdev->qp_list); 1477 mutex_unlock(&rdev->qp_lock); 1478 atomic_inc(&rdev->qp_count); 1479 1480 return &qp->ib_qp; 1481 qp_destroy: 1482 bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp); 1483 free_umem: 1484 ib_umem_release(qp->rumem); 1485 ib_umem_release(qp->sumem); 1486 fail: 1487 kfree(qp); 1488 exit: 1489 return ERR_PTR(rc); 1490 } 1491 1492 static u8 __from_ib_qp_state(enum ib_qp_state state) 1493 { 1494 switch (state) { 1495 case IB_QPS_RESET: 1496 return CMDQ_MODIFY_QP_NEW_STATE_RESET; 1497 case IB_QPS_INIT: 1498 return CMDQ_MODIFY_QP_NEW_STATE_INIT; 1499 case IB_QPS_RTR: 1500 return CMDQ_MODIFY_QP_NEW_STATE_RTR; 1501 case IB_QPS_RTS: 1502 return CMDQ_MODIFY_QP_NEW_STATE_RTS; 1503 case IB_QPS_SQD: 1504 return CMDQ_MODIFY_QP_NEW_STATE_SQD; 1505 case IB_QPS_SQE: 1506 return CMDQ_MODIFY_QP_NEW_STATE_SQE; 1507 case IB_QPS_ERR: 1508 default: 1509 return CMDQ_MODIFY_QP_NEW_STATE_ERR; 1510 } 1511 } 1512 1513 static enum ib_qp_state __to_ib_qp_state(u8 state) 1514 { 1515 switch (state) { 1516 case CMDQ_MODIFY_QP_NEW_STATE_RESET: 1517 return IB_QPS_RESET; 1518 case CMDQ_MODIFY_QP_NEW_STATE_INIT: 1519 return IB_QPS_INIT; 1520 case CMDQ_MODIFY_QP_NEW_STATE_RTR: 1521 return IB_QPS_RTR; 1522 case CMDQ_MODIFY_QP_NEW_STATE_RTS: 1523 return IB_QPS_RTS; 1524 case CMDQ_MODIFY_QP_NEW_STATE_SQD: 1525 return IB_QPS_SQD; 1526 case CMDQ_MODIFY_QP_NEW_STATE_SQE: 1527 return IB_QPS_SQE; 1528 case CMDQ_MODIFY_QP_NEW_STATE_ERR: 1529 default: 1530 return IB_QPS_ERR; 1531 } 1532 } 1533 1534 static u32 __from_ib_mtu(enum ib_mtu mtu) 1535 { 1536 switch (mtu) { 1537 case IB_MTU_256: 1538 return CMDQ_MODIFY_QP_PATH_MTU_MTU_256; 1539 case IB_MTU_512: 1540 return CMDQ_MODIFY_QP_PATH_MTU_MTU_512; 1541 case IB_MTU_1024: 1542 return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024; 1543 case IB_MTU_2048: 1544 return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048; 1545 case IB_MTU_4096: 1546 return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096; 1547 default: 1548 return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048; 1549 } 1550 } 1551 1552 static enum ib_mtu __to_ib_mtu(u32 mtu) 1553 { 1554 switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) { 1555 case CMDQ_MODIFY_QP_PATH_MTU_MTU_256: 1556 return IB_MTU_256; 1557 case CMDQ_MODIFY_QP_PATH_MTU_MTU_512: 1558 return IB_MTU_512; 1559 case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024: 1560 return IB_MTU_1024; 1561 case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048: 1562 return IB_MTU_2048; 1563 case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096: 1564 return IB_MTU_4096; 1565 default: 1566 return IB_MTU_2048; 1567 } 1568 } 1569 1570 /* Shared Receive Queues */ 1571 void bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata) 1572 { 1573 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1574 ib_srq); 1575 struct bnxt_re_dev *rdev = srq->rdev; 1576 struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq; 1577 struct bnxt_qplib_nq *nq = NULL; 1578 1579 if (qplib_srq->cq) 1580 nq = qplib_srq->cq->nq; 1581 bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq); 1582 ib_umem_release(srq->umem); 1583 atomic_dec(&rdev->srq_count); 1584 if (nq) 1585 nq->budget--; 1586 } 1587 1588 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev, 1589 struct bnxt_re_pd *pd, 1590 struct bnxt_re_srq *srq, 1591 struct ib_udata *udata) 1592 { 1593 struct bnxt_re_srq_req ureq; 1594 struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq; 1595 struct ib_umem *umem; 1596 int bytes = 0; 1597 struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context( 1598 udata, struct bnxt_re_ucontext, ib_uctx); 1599 1600 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) 1601 return -EFAULT; 1602 1603 bytes = (qplib_srq->max_wqe * qplib_srq->wqe_size); 1604 bytes = PAGE_ALIGN(bytes); 1605 umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes, 1606 IB_ACCESS_LOCAL_WRITE); 1607 if (IS_ERR(umem)) 1608 return PTR_ERR(umem); 1609 1610 srq->umem = umem; 1611 qplib_srq->sg_info.sghead = umem->sg_head.sgl; 1612 qplib_srq->sg_info.npages = ib_umem_num_pages(umem); 1613 qplib_srq->sg_info.nmap = umem->nmap; 1614 qplib_srq->sg_info.pgsize = PAGE_SIZE; 1615 qplib_srq->sg_info.pgshft = PAGE_SHIFT; 1616 qplib_srq->srq_handle = ureq.srq_handle; 1617 qplib_srq->dpi = &cntx->dpi; 1618 1619 return 0; 1620 } 1621 1622 int bnxt_re_create_srq(struct ib_srq *ib_srq, 1623 struct ib_srq_init_attr *srq_init_attr, 1624 struct ib_udata *udata) 1625 { 1626 struct bnxt_qplib_dev_attr *dev_attr; 1627 struct bnxt_qplib_nq *nq = NULL; 1628 struct bnxt_re_dev *rdev; 1629 struct bnxt_re_srq *srq; 1630 struct bnxt_re_pd *pd; 1631 struct ib_pd *ib_pd; 1632 int rc, entries; 1633 1634 ib_pd = ib_srq->pd; 1635 pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 1636 rdev = pd->rdev; 1637 dev_attr = &rdev->dev_attr; 1638 srq = container_of(ib_srq, struct bnxt_re_srq, ib_srq); 1639 1640 if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) { 1641 ibdev_err(&rdev->ibdev, "Create CQ failed - max exceeded"); 1642 rc = -EINVAL; 1643 goto exit; 1644 } 1645 1646 if (srq_init_attr->srq_type != IB_SRQT_BASIC) { 1647 rc = -EOPNOTSUPP; 1648 goto exit; 1649 } 1650 1651 srq->rdev = rdev; 1652 srq->qplib_srq.pd = &pd->qplib_pd; 1653 srq->qplib_srq.dpi = &rdev->dpi_privileged; 1654 /* Allocate 1 more than what's provided so posting max doesn't 1655 * mean empty 1656 */ 1657 entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1); 1658 if (entries > dev_attr->max_srq_wqes + 1) 1659 entries = dev_attr->max_srq_wqes + 1; 1660 srq->qplib_srq.max_wqe = entries; 1661 1662 srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge; 1663 srq->qplib_srq.wqe_size = 1664 bnxt_re_get_rwqe_size(srq->qplib_srq.max_sge); 1665 srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit; 1666 srq->srq_limit = srq_init_attr->attr.srq_limit; 1667 srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id; 1668 nq = &rdev->nq[0]; 1669 1670 if (udata) { 1671 rc = bnxt_re_init_user_srq(rdev, pd, srq, udata); 1672 if (rc) 1673 goto fail; 1674 } 1675 1676 rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq); 1677 if (rc) { 1678 ibdev_err(&rdev->ibdev, "Create HW SRQ failed!"); 1679 goto fail; 1680 } 1681 1682 if (udata) { 1683 struct bnxt_re_srq_resp resp; 1684 1685 resp.srqid = srq->qplib_srq.id; 1686 rc = ib_copy_to_udata(udata, &resp, sizeof(resp)); 1687 if (rc) { 1688 ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!"); 1689 bnxt_qplib_destroy_srq(&rdev->qplib_res, 1690 &srq->qplib_srq); 1691 goto fail; 1692 } 1693 } 1694 if (nq) 1695 nq->budget++; 1696 atomic_inc(&rdev->srq_count); 1697 1698 return 0; 1699 1700 fail: 1701 ib_umem_release(srq->umem); 1702 exit: 1703 return rc; 1704 } 1705 1706 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr, 1707 enum ib_srq_attr_mask srq_attr_mask, 1708 struct ib_udata *udata) 1709 { 1710 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1711 ib_srq); 1712 struct bnxt_re_dev *rdev = srq->rdev; 1713 int rc; 1714 1715 switch (srq_attr_mask) { 1716 case IB_SRQ_MAX_WR: 1717 /* SRQ resize is not supported */ 1718 break; 1719 case IB_SRQ_LIMIT: 1720 /* Change the SRQ threshold */ 1721 if (srq_attr->srq_limit > srq->qplib_srq.max_wqe) 1722 return -EINVAL; 1723 1724 srq->qplib_srq.threshold = srq_attr->srq_limit; 1725 rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq); 1726 if (rc) { 1727 ibdev_err(&rdev->ibdev, "Modify HW SRQ failed!"); 1728 return rc; 1729 } 1730 /* On success, update the shadow */ 1731 srq->srq_limit = srq_attr->srq_limit; 1732 /* No need to Build and send response back to udata */ 1733 break; 1734 default: 1735 ibdev_err(&rdev->ibdev, 1736 "Unsupported srq_attr_mask 0x%x", srq_attr_mask); 1737 return -EINVAL; 1738 } 1739 return 0; 1740 } 1741 1742 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr) 1743 { 1744 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1745 ib_srq); 1746 struct bnxt_re_srq tsrq; 1747 struct bnxt_re_dev *rdev = srq->rdev; 1748 int rc; 1749 1750 /* Get live SRQ attr */ 1751 tsrq.qplib_srq.id = srq->qplib_srq.id; 1752 rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq); 1753 if (rc) { 1754 ibdev_err(&rdev->ibdev, "Query HW SRQ failed!"); 1755 return rc; 1756 } 1757 srq_attr->max_wr = srq->qplib_srq.max_wqe; 1758 srq_attr->max_sge = srq->qplib_srq.max_sge; 1759 srq_attr->srq_limit = tsrq.qplib_srq.threshold; 1760 1761 return 0; 1762 } 1763 1764 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr, 1765 const struct ib_recv_wr **bad_wr) 1766 { 1767 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1768 ib_srq); 1769 struct bnxt_qplib_swqe wqe; 1770 unsigned long flags; 1771 int rc = 0; 1772 1773 spin_lock_irqsave(&srq->lock, flags); 1774 while (wr) { 1775 /* Transcribe each ib_recv_wr to qplib_swqe */ 1776 wqe.num_sge = wr->num_sge; 1777 bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge); 1778 wqe.wr_id = wr->wr_id; 1779 wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV; 1780 1781 rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe); 1782 if (rc) { 1783 *bad_wr = wr; 1784 break; 1785 } 1786 wr = wr->next; 1787 } 1788 spin_unlock_irqrestore(&srq->lock, flags); 1789 1790 return rc; 1791 } 1792 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev, 1793 struct bnxt_re_qp *qp1_qp, 1794 int qp_attr_mask) 1795 { 1796 struct bnxt_re_qp *qp = rdev->gsi_ctx.gsi_sqp; 1797 int rc = 0; 1798 1799 if (qp_attr_mask & IB_QP_STATE) { 1800 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE; 1801 qp->qplib_qp.state = qp1_qp->qplib_qp.state; 1802 } 1803 if (qp_attr_mask & IB_QP_PKEY_INDEX) { 1804 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY; 1805 qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index; 1806 } 1807 1808 if (qp_attr_mask & IB_QP_QKEY) { 1809 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY; 1810 /* Using a Random QKEY */ 1811 qp->qplib_qp.qkey = 0x81818181; 1812 } 1813 if (qp_attr_mask & IB_QP_SQ_PSN) { 1814 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN; 1815 qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn; 1816 } 1817 1818 rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp); 1819 if (rc) 1820 ibdev_err(&rdev->ibdev, "Failed to modify Shadow QP for QP1"); 1821 return rc; 1822 } 1823 1824 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr, 1825 int qp_attr_mask, struct ib_udata *udata) 1826 { 1827 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 1828 struct bnxt_re_dev *rdev = qp->rdev; 1829 struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr; 1830 enum ib_qp_state curr_qp_state, new_qp_state; 1831 int rc, entries; 1832 unsigned int flags; 1833 u8 nw_type; 1834 1835 qp->qplib_qp.modify_flags = 0; 1836 if (qp_attr_mask & IB_QP_STATE) { 1837 curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state); 1838 new_qp_state = qp_attr->qp_state; 1839 if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state, 1840 ib_qp->qp_type, qp_attr_mask)) { 1841 ibdev_err(&rdev->ibdev, 1842 "Invalid attribute mask: %#x specified ", 1843 qp_attr_mask); 1844 ibdev_err(&rdev->ibdev, 1845 "for qpn: %#x type: %#x", 1846 ib_qp->qp_num, ib_qp->qp_type); 1847 ibdev_err(&rdev->ibdev, 1848 "curr_qp_state=0x%x, new_qp_state=0x%x\n", 1849 curr_qp_state, new_qp_state); 1850 return -EINVAL; 1851 } 1852 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE; 1853 qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state); 1854 1855 if (!qp->sumem && 1856 qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { 1857 ibdev_dbg(&rdev->ibdev, 1858 "Move QP = %p to flush list\n", qp); 1859 flags = bnxt_re_lock_cqs(qp); 1860 bnxt_qplib_add_flush_qp(&qp->qplib_qp); 1861 bnxt_re_unlock_cqs(qp, flags); 1862 } 1863 if (!qp->sumem && 1864 qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) { 1865 ibdev_dbg(&rdev->ibdev, 1866 "Move QP = %p out of flush list\n", qp); 1867 flags = bnxt_re_lock_cqs(qp); 1868 bnxt_qplib_clean_qp(&qp->qplib_qp); 1869 bnxt_re_unlock_cqs(qp, flags); 1870 } 1871 } 1872 if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) { 1873 qp->qplib_qp.modify_flags |= 1874 CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY; 1875 qp->qplib_qp.en_sqd_async_notify = true; 1876 } 1877 if (qp_attr_mask & IB_QP_ACCESS_FLAGS) { 1878 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS; 1879 qp->qplib_qp.access = 1880 __from_ib_access_flags(qp_attr->qp_access_flags); 1881 /* LOCAL_WRITE access must be set to allow RC receive */ 1882 qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE; 1883 /* Temp: Set all params on QP as of now */ 1884 qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE; 1885 qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ; 1886 } 1887 if (qp_attr_mask & IB_QP_PKEY_INDEX) { 1888 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY; 1889 qp->qplib_qp.pkey_index = qp_attr->pkey_index; 1890 } 1891 if (qp_attr_mask & IB_QP_QKEY) { 1892 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY; 1893 qp->qplib_qp.qkey = qp_attr->qkey; 1894 } 1895 if (qp_attr_mask & IB_QP_AV) { 1896 const struct ib_global_route *grh = 1897 rdma_ah_read_grh(&qp_attr->ah_attr); 1898 const struct ib_gid_attr *sgid_attr; 1899 struct bnxt_re_gid_ctx *ctx; 1900 1901 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID | 1902 CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL | 1903 CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX | 1904 CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT | 1905 CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS | 1906 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC | 1907 CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID; 1908 memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw, 1909 sizeof(qp->qplib_qp.ah.dgid.data)); 1910 qp->qplib_qp.ah.flow_label = grh->flow_label; 1911 sgid_attr = grh->sgid_attr; 1912 /* Get the HW context of the GID. The reference 1913 * of GID table entry is already taken by the caller. 1914 */ 1915 ctx = rdma_read_gid_hw_context(sgid_attr); 1916 qp->qplib_qp.ah.sgid_index = ctx->idx; 1917 qp->qplib_qp.ah.host_sgid_index = grh->sgid_index; 1918 qp->qplib_qp.ah.hop_limit = grh->hop_limit; 1919 qp->qplib_qp.ah.traffic_class = grh->traffic_class; 1920 qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr); 1921 ether_addr_copy(qp->qplib_qp.ah.dmac, 1922 qp_attr->ah_attr.roce.dmac); 1923 1924 rc = rdma_read_gid_l2_fields(sgid_attr, NULL, 1925 &qp->qplib_qp.smac[0]); 1926 if (rc) 1927 return rc; 1928 1929 nw_type = rdma_gid_attr_network_type(sgid_attr); 1930 switch (nw_type) { 1931 case RDMA_NETWORK_IPV4: 1932 qp->qplib_qp.nw_type = 1933 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4; 1934 break; 1935 case RDMA_NETWORK_IPV6: 1936 qp->qplib_qp.nw_type = 1937 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6; 1938 break; 1939 default: 1940 qp->qplib_qp.nw_type = 1941 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1; 1942 break; 1943 } 1944 } 1945 1946 if (qp_attr_mask & IB_QP_PATH_MTU) { 1947 qp->qplib_qp.modify_flags |= 1948 CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU; 1949 qp->qplib_qp.path_mtu = __from_ib_mtu(qp_attr->path_mtu); 1950 qp->qplib_qp.mtu = ib_mtu_enum_to_int(qp_attr->path_mtu); 1951 } else if (qp_attr->qp_state == IB_QPS_RTR) { 1952 qp->qplib_qp.modify_flags |= 1953 CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU; 1954 qp->qplib_qp.path_mtu = 1955 __from_ib_mtu(iboe_get_mtu(rdev->netdev->mtu)); 1956 qp->qplib_qp.mtu = 1957 ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu)); 1958 } 1959 1960 if (qp_attr_mask & IB_QP_TIMEOUT) { 1961 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT; 1962 qp->qplib_qp.timeout = qp_attr->timeout; 1963 } 1964 if (qp_attr_mask & IB_QP_RETRY_CNT) { 1965 qp->qplib_qp.modify_flags |= 1966 CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT; 1967 qp->qplib_qp.retry_cnt = qp_attr->retry_cnt; 1968 } 1969 if (qp_attr_mask & IB_QP_RNR_RETRY) { 1970 qp->qplib_qp.modify_flags |= 1971 CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY; 1972 qp->qplib_qp.rnr_retry = qp_attr->rnr_retry; 1973 } 1974 if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) { 1975 qp->qplib_qp.modify_flags |= 1976 CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER; 1977 qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer; 1978 } 1979 if (qp_attr_mask & IB_QP_RQ_PSN) { 1980 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN; 1981 qp->qplib_qp.rq.psn = qp_attr->rq_psn; 1982 } 1983 if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 1984 qp->qplib_qp.modify_flags |= 1985 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC; 1986 /* Cap the max_rd_atomic to device max */ 1987 qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic, 1988 dev_attr->max_qp_rd_atom); 1989 } 1990 if (qp_attr_mask & IB_QP_SQ_PSN) { 1991 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN; 1992 qp->qplib_qp.sq.psn = qp_attr->sq_psn; 1993 } 1994 if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 1995 if (qp_attr->max_dest_rd_atomic > 1996 dev_attr->max_qp_init_rd_atom) { 1997 ibdev_err(&rdev->ibdev, 1998 "max_dest_rd_atomic requested%d is > dev_max%d", 1999 qp_attr->max_dest_rd_atomic, 2000 dev_attr->max_qp_init_rd_atom); 2001 return -EINVAL; 2002 } 2003 2004 qp->qplib_qp.modify_flags |= 2005 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC; 2006 qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic; 2007 } 2008 if (qp_attr_mask & IB_QP_CAP) { 2009 qp->qplib_qp.modify_flags |= 2010 CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE | 2011 CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE | 2012 CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE | 2013 CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE | 2014 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA; 2015 if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) || 2016 (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) || 2017 (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) || 2018 (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) || 2019 (qp_attr->cap.max_inline_data >= 2020 dev_attr->max_inline_data)) { 2021 ibdev_err(&rdev->ibdev, 2022 "Create QP failed - max exceeded"); 2023 return -EINVAL; 2024 } 2025 entries = roundup_pow_of_two(qp_attr->cap.max_send_wr); 2026 qp->qplib_qp.sq.max_wqe = min_t(u32, entries, 2027 dev_attr->max_qp_wqes + 1); 2028 qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe - 2029 qp_attr->cap.max_send_wr; 2030 /* 2031 * Reserving one slot for Phantom WQE. Some application can 2032 * post one extra entry in this case. Allowing this to avoid 2033 * unexpected Queue full condition 2034 */ 2035 qp->qplib_qp.sq.q_full_delta -= 1; 2036 qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge; 2037 if (qp->qplib_qp.rq.max_wqe) { 2038 entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr); 2039 qp->qplib_qp.rq.max_wqe = 2040 min_t(u32, entries, dev_attr->max_qp_wqes + 1); 2041 qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe - 2042 qp_attr->cap.max_recv_wr; 2043 qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge; 2044 } else { 2045 /* SRQ was used prior, just ignore the RQ caps */ 2046 } 2047 } 2048 if (qp_attr_mask & IB_QP_DEST_QPN) { 2049 qp->qplib_qp.modify_flags |= 2050 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID; 2051 qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num; 2052 } 2053 rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp); 2054 if (rc) { 2055 ibdev_err(&rdev->ibdev, "Failed to modify HW QP"); 2056 return rc; 2057 } 2058 if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) 2059 rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask); 2060 return rc; 2061 } 2062 2063 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr, 2064 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr) 2065 { 2066 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 2067 struct bnxt_re_dev *rdev = qp->rdev; 2068 struct bnxt_qplib_qp *qplib_qp; 2069 int rc; 2070 2071 qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL); 2072 if (!qplib_qp) 2073 return -ENOMEM; 2074 2075 qplib_qp->id = qp->qplib_qp.id; 2076 qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index; 2077 2078 rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp); 2079 if (rc) { 2080 ibdev_err(&rdev->ibdev, "Failed to query HW QP"); 2081 goto out; 2082 } 2083 qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state); 2084 qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0; 2085 qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access); 2086 qp_attr->pkey_index = qplib_qp->pkey_index; 2087 qp_attr->qkey = qplib_qp->qkey; 2088 qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE; 2089 rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label, 2090 qplib_qp->ah.host_sgid_index, 2091 qplib_qp->ah.hop_limit, 2092 qplib_qp->ah.traffic_class); 2093 rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data); 2094 rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl); 2095 ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac); 2096 qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu); 2097 qp_attr->timeout = qplib_qp->timeout; 2098 qp_attr->retry_cnt = qplib_qp->retry_cnt; 2099 qp_attr->rnr_retry = qplib_qp->rnr_retry; 2100 qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer; 2101 qp_attr->rq_psn = qplib_qp->rq.psn; 2102 qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic; 2103 qp_attr->sq_psn = qplib_qp->sq.psn; 2104 qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic; 2105 qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR : 2106 IB_SIGNAL_REQ_WR; 2107 qp_attr->dest_qp_num = qplib_qp->dest_qpn; 2108 2109 qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe; 2110 qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge; 2111 qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe; 2112 qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge; 2113 qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data; 2114 qp_init_attr->cap = qp_attr->cap; 2115 2116 out: 2117 kfree(qplib_qp); 2118 return rc; 2119 } 2120 2121 /* Routine for sending QP1 packets for RoCE V1 an V2 2122 */ 2123 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp, 2124 const struct ib_send_wr *wr, 2125 struct bnxt_qplib_swqe *wqe, 2126 int payload_size) 2127 { 2128 struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, 2129 ib_ah); 2130 struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah; 2131 const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr; 2132 struct bnxt_qplib_sge sge; 2133 u8 nw_type; 2134 u16 ether_type; 2135 union ib_gid dgid; 2136 bool is_eth = false; 2137 bool is_vlan = false; 2138 bool is_grh = false; 2139 bool is_udp = false; 2140 u8 ip_version = 0; 2141 u16 vlan_id = 0xFFFF; 2142 void *buf; 2143 int i, rc = 0; 2144 2145 memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr)); 2146 2147 rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL); 2148 if (rc) 2149 return rc; 2150 2151 /* Get network header type for this GID */ 2152 nw_type = rdma_gid_attr_network_type(sgid_attr); 2153 switch (nw_type) { 2154 case RDMA_NETWORK_IPV4: 2155 nw_type = BNXT_RE_ROCEV2_IPV4_PACKET; 2156 break; 2157 case RDMA_NETWORK_IPV6: 2158 nw_type = BNXT_RE_ROCEV2_IPV6_PACKET; 2159 break; 2160 default: 2161 nw_type = BNXT_RE_ROCE_V1_PACKET; 2162 break; 2163 } 2164 memcpy(&dgid.raw, &qplib_ah->dgid, 16); 2165 is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP; 2166 if (is_udp) { 2167 if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) { 2168 ip_version = 4; 2169 ether_type = ETH_P_IP; 2170 } else { 2171 ip_version = 6; 2172 ether_type = ETH_P_IPV6; 2173 } 2174 is_grh = false; 2175 } else { 2176 ether_type = ETH_P_IBOE; 2177 is_grh = true; 2178 } 2179 2180 is_eth = true; 2181 is_vlan = (vlan_id && (vlan_id < 0x1000)) ? true : false; 2182 2183 ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh, 2184 ip_version, is_udp, 0, &qp->qp1_hdr); 2185 2186 /* ETH */ 2187 ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac); 2188 ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac); 2189 2190 /* For vlan, check the sgid for vlan existence */ 2191 2192 if (!is_vlan) { 2193 qp->qp1_hdr.eth.type = cpu_to_be16(ether_type); 2194 } else { 2195 qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type); 2196 qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id); 2197 } 2198 2199 if (is_grh || (ip_version == 6)) { 2200 memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw, 2201 sizeof(sgid_attr->gid)); 2202 memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data, 2203 sizeof(sgid_attr->gid)); 2204 qp->qp1_hdr.grh.hop_limit = qplib_ah->hop_limit; 2205 } 2206 2207 if (ip_version == 4) { 2208 qp->qp1_hdr.ip4.tos = 0; 2209 qp->qp1_hdr.ip4.id = 0; 2210 qp->qp1_hdr.ip4.frag_off = htons(IP_DF); 2211 qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit; 2212 2213 memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4); 2214 memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4); 2215 qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr); 2216 } 2217 2218 if (is_udp) { 2219 qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT); 2220 qp->qp1_hdr.udp.sport = htons(0x8CD1); 2221 qp->qp1_hdr.udp.csum = 0; 2222 } 2223 2224 /* BTH */ 2225 if (wr->opcode == IB_WR_SEND_WITH_IMM) { 2226 qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 2227 qp->qp1_hdr.immediate_present = 1; 2228 } else { 2229 qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 2230 } 2231 if (wr->send_flags & IB_SEND_SOLICITED) 2232 qp->qp1_hdr.bth.solicited_event = 1; 2233 /* pad_count */ 2234 qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3; 2235 2236 /* P_key for QP1 is for all members */ 2237 qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF); 2238 qp->qp1_hdr.bth.destination_qpn = IB_QP1; 2239 qp->qp1_hdr.bth.ack_req = 0; 2240 qp->send_psn++; 2241 qp->send_psn &= BTH_PSN_MASK; 2242 qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn); 2243 /* DETH */ 2244 /* Use the priviledged Q_Key for QP1 */ 2245 qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY); 2246 qp->qp1_hdr.deth.source_qpn = IB_QP1; 2247 2248 /* Pack the QP1 to the transmit buffer */ 2249 buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge); 2250 if (buf) { 2251 ib_ud_header_pack(&qp->qp1_hdr, buf); 2252 for (i = wqe->num_sge; i; i--) { 2253 wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr; 2254 wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey; 2255 wqe->sg_list[i].size = wqe->sg_list[i - 1].size; 2256 } 2257 2258 /* 2259 * Max Header buf size for IPV6 RoCE V2 is 86, 2260 * which is same as the QP1 SQ header buffer. 2261 * Header buf size for IPV4 RoCE V2 can be 66. 2262 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20). 2263 * Subtract 20 bytes from QP1 SQ header buf size 2264 */ 2265 if (is_udp && ip_version == 4) 2266 sge.size -= 20; 2267 /* 2268 * Max Header buf size for RoCE V1 is 78. 2269 * ETH(14) + VLAN(4) + GRH(40) + BTH(20). 2270 * Subtract 8 bytes from QP1 SQ header buf size 2271 */ 2272 if (!is_udp) 2273 sge.size -= 8; 2274 2275 /* Subtract 4 bytes for non vlan packets */ 2276 if (!is_vlan) 2277 sge.size -= 4; 2278 2279 wqe->sg_list[0].addr = sge.addr; 2280 wqe->sg_list[0].lkey = sge.lkey; 2281 wqe->sg_list[0].size = sge.size; 2282 wqe->num_sge++; 2283 2284 } else { 2285 ibdev_err(&qp->rdev->ibdev, "QP1 buffer is empty!"); 2286 rc = -ENOMEM; 2287 } 2288 return rc; 2289 } 2290 2291 /* For the MAD layer, it only provides the recv SGE the size of 2292 * ib_grh + MAD datagram. No Ethernet headers, Ethertype, BTH, DETH, 2293 * nor RoCE iCRC. The Cu+ solution must provide buffer for the entire 2294 * receive packet (334 bytes) with no VLAN and then copy the GRH 2295 * and the MAD datagram out to the provided SGE. 2296 */ 2297 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp, 2298 const struct ib_recv_wr *wr, 2299 struct bnxt_qplib_swqe *wqe, 2300 int payload_size) 2301 { 2302 struct bnxt_re_sqp_entries *sqp_entry; 2303 struct bnxt_qplib_sge ref, sge; 2304 struct bnxt_re_dev *rdev; 2305 u32 rq_prod_index; 2306 2307 rdev = qp->rdev; 2308 2309 rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp); 2310 2311 if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge)) 2312 return -ENOMEM; 2313 2314 /* Create 1 SGE to receive the entire 2315 * ethernet packet 2316 */ 2317 /* Save the reference from ULP */ 2318 ref.addr = wqe->sg_list[0].addr; 2319 ref.lkey = wqe->sg_list[0].lkey; 2320 ref.size = wqe->sg_list[0].size; 2321 2322 sqp_entry = &rdev->gsi_ctx.sqp_tbl[rq_prod_index]; 2323 2324 /* SGE 1 */ 2325 wqe->sg_list[0].addr = sge.addr; 2326 wqe->sg_list[0].lkey = sge.lkey; 2327 wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2; 2328 sge.size -= wqe->sg_list[0].size; 2329 2330 sqp_entry->sge.addr = ref.addr; 2331 sqp_entry->sge.lkey = ref.lkey; 2332 sqp_entry->sge.size = ref.size; 2333 /* Store the wrid for reporting completion */ 2334 sqp_entry->wrid = wqe->wr_id; 2335 /* change the wqe->wrid to table index */ 2336 wqe->wr_id = rq_prod_index; 2337 return 0; 2338 } 2339 2340 static int is_ud_qp(struct bnxt_re_qp *qp) 2341 { 2342 return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD || 2343 qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI); 2344 } 2345 2346 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp, 2347 const struct ib_send_wr *wr, 2348 struct bnxt_qplib_swqe *wqe) 2349 { 2350 struct bnxt_re_ah *ah = NULL; 2351 2352 if (is_ud_qp(qp)) { 2353 ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah); 2354 wqe->send.q_key = ud_wr(wr)->remote_qkey; 2355 wqe->send.dst_qp = ud_wr(wr)->remote_qpn; 2356 wqe->send.avid = ah->qplib_ah.id; 2357 } 2358 switch (wr->opcode) { 2359 case IB_WR_SEND: 2360 wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND; 2361 break; 2362 case IB_WR_SEND_WITH_IMM: 2363 wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM; 2364 wqe->send.imm_data = wr->ex.imm_data; 2365 break; 2366 case IB_WR_SEND_WITH_INV: 2367 wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV; 2368 wqe->send.inv_key = wr->ex.invalidate_rkey; 2369 break; 2370 default: 2371 return -EINVAL; 2372 } 2373 if (wr->send_flags & IB_SEND_SIGNALED) 2374 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2375 if (wr->send_flags & IB_SEND_FENCE) 2376 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2377 if (wr->send_flags & IB_SEND_SOLICITED) 2378 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2379 if (wr->send_flags & IB_SEND_INLINE) 2380 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE; 2381 2382 return 0; 2383 } 2384 2385 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr, 2386 struct bnxt_qplib_swqe *wqe) 2387 { 2388 switch (wr->opcode) { 2389 case IB_WR_RDMA_WRITE: 2390 wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE; 2391 break; 2392 case IB_WR_RDMA_WRITE_WITH_IMM: 2393 wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM; 2394 wqe->rdma.imm_data = wr->ex.imm_data; 2395 break; 2396 case IB_WR_RDMA_READ: 2397 wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ; 2398 wqe->rdma.inv_key = wr->ex.invalidate_rkey; 2399 break; 2400 default: 2401 return -EINVAL; 2402 } 2403 wqe->rdma.remote_va = rdma_wr(wr)->remote_addr; 2404 wqe->rdma.r_key = rdma_wr(wr)->rkey; 2405 if (wr->send_flags & IB_SEND_SIGNALED) 2406 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2407 if (wr->send_flags & IB_SEND_FENCE) 2408 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2409 if (wr->send_flags & IB_SEND_SOLICITED) 2410 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2411 if (wr->send_flags & IB_SEND_INLINE) 2412 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE; 2413 2414 return 0; 2415 } 2416 2417 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr, 2418 struct bnxt_qplib_swqe *wqe) 2419 { 2420 switch (wr->opcode) { 2421 case IB_WR_ATOMIC_CMP_AND_SWP: 2422 wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP; 2423 wqe->atomic.cmp_data = atomic_wr(wr)->compare_add; 2424 wqe->atomic.swap_data = atomic_wr(wr)->swap; 2425 break; 2426 case IB_WR_ATOMIC_FETCH_AND_ADD: 2427 wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD; 2428 wqe->atomic.cmp_data = atomic_wr(wr)->compare_add; 2429 break; 2430 default: 2431 return -EINVAL; 2432 } 2433 wqe->atomic.remote_va = atomic_wr(wr)->remote_addr; 2434 wqe->atomic.r_key = atomic_wr(wr)->rkey; 2435 if (wr->send_flags & IB_SEND_SIGNALED) 2436 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2437 if (wr->send_flags & IB_SEND_FENCE) 2438 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2439 if (wr->send_flags & IB_SEND_SOLICITED) 2440 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2441 return 0; 2442 } 2443 2444 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr, 2445 struct bnxt_qplib_swqe *wqe) 2446 { 2447 wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV; 2448 wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey; 2449 2450 /* Need unconditional fence for local invalidate 2451 * opcode to work as expected. 2452 */ 2453 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2454 2455 if (wr->send_flags & IB_SEND_SIGNALED) 2456 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2457 if (wr->send_flags & IB_SEND_SOLICITED) 2458 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2459 2460 return 0; 2461 } 2462 2463 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr, 2464 struct bnxt_qplib_swqe *wqe) 2465 { 2466 struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr); 2467 struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl; 2468 int access = wr->access; 2469 2470 wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0]; 2471 wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0]; 2472 wqe->frmr.page_list = mr->pages; 2473 wqe->frmr.page_list_len = mr->npages; 2474 wqe->frmr.levels = qplib_frpl->hwq.level; 2475 wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR; 2476 2477 /* Need unconditional fence for reg_mr 2478 * opcode to function as expected. 2479 */ 2480 2481 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2482 2483 if (wr->wr.send_flags & IB_SEND_SIGNALED) 2484 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2485 2486 if (access & IB_ACCESS_LOCAL_WRITE) 2487 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE; 2488 if (access & IB_ACCESS_REMOTE_READ) 2489 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ; 2490 if (access & IB_ACCESS_REMOTE_WRITE) 2491 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE; 2492 if (access & IB_ACCESS_REMOTE_ATOMIC) 2493 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC; 2494 if (access & IB_ACCESS_MW_BIND) 2495 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND; 2496 2497 wqe->frmr.l_key = wr->key; 2498 wqe->frmr.length = wr->mr->length; 2499 wqe->frmr.pbl_pg_sz_log = (wr->mr->page_size >> PAGE_SHIFT_4K) - 1; 2500 wqe->frmr.va = wr->mr->iova; 2501 return 0; 2502 } 2503 2504 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev, 2505 const struct ib_send_wr *wr, 2506 struct bnxt_qplib_swqe *wqe) 2507 { 2508 /* Copy the inline data to the data field */ 2509 u8 *in_data; 2510 u32 i, sge_len; 2511 void *sge_addr; 2512 2513 in_data = wqe->inline_data; 2514 for (i = 0; i < wr->num_sge; i++) { 2515 sge_addr = (void *)(unsigned long) 2516 wr->sg_list[i].addr; 2517 sge_len = wr->sg_list[i].length; 2518 2519 if ((sge_len + wqe->inline_len) > 2520 BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) { 2521 ibdev_err(&rdev->ibdev, 2522 "Inline data size requested > supported value"); 2523 return -EINVAL; 2524 } 2525 sge_len = wr->sg_list[i].length; 2526 2527 memcpy(in_data, sge_addr, sge_len); 2528 in_data += wr->sg_list[i].length; 2529 wqe->inline_len += wr->sg_list[i].length; 2530 } 2531 return wqe->inline_len; 2532 } 2533 2534 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev, 2535 const struct ib_send_wr *wr, 2536 struct bnxt_qplib_swqe *wqe) 2537 { 2538 int payload_sz = 0; 2539 2540 if (wr->send_flags & IB_SEND_INLINE) 2541 payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe); 2542 else 2543 payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list, 2544 wqe->num_sge); 2545 2546 return payload_sz; 2547 } 2548 2549 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp) 2550 { 2551 if ((qp->ib_qp.qp_type == IB_QPT_UD || 2552 qp->ib_qp.qp_type == IB_QPT_GSI || 2553 qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) && 2554 qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) { 2555 int qp_attr_mask; 2556 struct ib_qp_attr qp_attr; 2557 2558 qp_attr_mask = IB_QP_STATE; 2559 qp_attr.qp_state = IB_QPS_RTS; 2560 bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL); 2561 qp->qplib_qp.wqe_cnt = 0; 2562 } 2563 } 2564 2565 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev, 2566 struct bnxt_re_qp *qp, 2567 const struct ib_send_wr *wr) 2568 { 2569 int rc = 0, payload_sz = 0; 2570 unsigned long flags; 2571 2572 spin_lock_irqsave(&qp->sq_lock, flags); 2573 while (wr) { 2574 struct bnxt_qplib_swqe wqe = {}; 2575 2576 /* Common */ 2577 wqe.num_sge = wr->num_sge; 2578 if (wr->num_sge > qp->qplib_qp.sq.max_sge) { 2579 ibdev_err(&rdev->ibdev, 2580 "Limit exceeded for Send SGEs"); 2581 rc = -EINVAL; 2582 goto bad; 2583 } 2584 2585 payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe); 2586 if (payload_sz < 0) { 2587 rc = -EINVAL; 2588 goto bad; 2589 } 2590 wqe.wr_id = wr->wr_id; 2591 2592 wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND; 2593 2594 rc = bnxt_re_build_send_wqe(qp, wr, &wqe); 2595 if (!rc) 2596 rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe); 2597 bad: 2598 if (rc) { 2599 ibdev_err(&rdev->ibdev, 2600 "Post send failed opcode = %#x rc = %d", 2601 wr->opcode, rc); 2602 break; 2603 } 2604 wr = wr->next; 2605 } 2606 bnxt_qplib_post_send_db(&qp->qplib_qp); 2607 bnxt_ud_qp_hw_stall_workaround(qp); 2608 spin_unlock_irqrestore(&qp->sq_lock, flags); 2609 return rc; 2610 } 2611 2612 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr, 2613 const struct ib_send_wr **bad_wr) 2614 { 2615 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 2616 struct bnxt_qplib_swqe wqe; 2617 int rc = 0, payload_sz = 0; 2618 unsigned long flags; 2619 2620 spin_lock_irqsave(&qp->sq_lock, flags); 2621 while (wr) { 2622 /* House keeping */ 2623 memset(&wqe, 0, sizeof(wqe)); 2624 2625 /* Common */ 2626 wqe.num_sge = wr->num_sge; 2627 if (wr->num_sge > qp->qplib_qp.sq.max_sge) { 2628 ibdev_err(&qp->rdev->ibdev, 2629 "Limit exceeded for Send SGEs"); 2630 rc = -EINVAL; 2631 goto bad; 2632 } 2633 2634 payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe); 2635 if (payload_sz < 0) { 2636 rc = -EINVAL; 2637 goto bad; 2638 } 2639 wqe.wr_id = wr->wr_id; 2640 2641 switch (wr->opcode) { 2642 case IB_WR_SEND: 2643 case IB_WR_SEND_WITH_IMM: 2644 if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) { 2645 rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe, 2646 payload_sz); 2647 if (rc) 2648 goto bad; 2649 wqe.rawqp1.lflags |= 2650 SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC; 2651 } 2652 switch (wr->send_flags) { 2653 case IB_SEND_IP_CSUM: 2654 wqe.rawqp1.lflags |= 2655 SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM; 2656 break; 2657 default: 2658 break; 2659 } 2660 fallthrough; 2661 case IB_WR_SEND_WITH_INV: 2662 rc = bnxt_re_build_send_wqe(qp, wr, &wqe); 2663 break; 2664 case IB_WR_RDMA_WRITE: 2665 case IB_WR_RDMA_WRITE_WITH_IMM: 2666 case IB_WR_RDMA_READ: 2667 rc = bnxt_re_build_rdma_wqe(wr, &wqe); 2668 break; 2669 case IB_WR_ATOMIC_CMP_AND_SWP: 2670 case IB_WR_ATOMIC_FETCH_AND_ADD: 2671 rc = bnxt_re_build_atomic_wqe(wr, &wqe); 2672 break; 2673 case IB_WR_RDMA_READ_WITH_INV: 2674 ibdev_err(&qp->rdev->ibdev, 2675 "RDMA Read with Invalidate is not supported"); 2676 rc = -EINVAL; 2677 goto bad; 2678 case IB_WR_LOCAL_INV: 2679 rc = bnxt_re_build_inv_wqe(wr, &wqe); 2680 break; 2681 case IB_WR_REG_MR: 2682 rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe); 2683 break; 2684 default: 2685 /* Unsupported WRs */ 2686 ibdev_err(&qp->rdev->ibdev, 2687 "WR (%#x) is not supported", wr->opcode); 2688 rc = -EINVAL; 2689 goto bad; 2690 } 2691 if (!rc) 2692 rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe); 2693 bad: 2694 if (rc) { 2695 ibdev_err(&qp->rdev->ibdev, 2696 "post_send failed op:%#x qps = %#x rc = %d\n", 2697 wr->opcode, qp->qplib_qp.state, rc); 2698 *bad_wr = wr; 2699 break; 2700 } 2701 wr = wr->next; 2702 } 2703 bnxt_qplib_post_send_db(&qp->qplib_qp); 2704 bnxt_ud_qp_hw_stall_workaround(qp); 2705 spin_unlock_irqrestore(&qp->sq_lock, flags); 2706 2707 return rc; 2708 } 2709 2710 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev, 2711 struct bnxt_re_qp *qp, 2712 const struct ib_recv_wr *wr) 2713 { 2714 struct bnxt_qplib_swqe wqe; 2715 int rc = 0; 2716 2717 memset(&wqe, 0, sizeof(wqe)); 2718 while (wr) { 2719 /* House keeping */ 2720 memset(&wqe, 0, sizeof(wqe)); 2721 2722 /* Common */ 2723 wqe.num_sge = wr->num_sge; 2724 if (wr->num_sge > qp->qplib_qp.rq.max_sge) { 2725 ibdev_err(&rdev->ibdev, 2726 "Limit exceeded for Receive SGEs"); 2727 rc = -EINVAL; 2728 break; 2729 } 2730 bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge); 2731 wqe.wr_id = wr->wr_id; 2732 wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV; 2733 2734 rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe); 2735 if (rc) 2736 break; 2737 2738 wr = wr->next; 2739 } 2740 if (!rc) 2741 bnxt_qplib_post_recv_db(&qp->qplib_qp); 2742 return rc; 2743 } 2744 2745 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr, 2746 const struct ib_recv_wr **bad_wr) 2747 { 2748 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 2749 struct bnxt_qplib_swqe wqe; 2750 int rc = 0, payload_sz = 0; 2751 unsigned long flags; 2752 u32 count = 0; 2753 2754 spin_lock_irqsave(&qp->rq_lock, flags); 2755 while (wr) { 2756 /* House keeping */ 2757 memset(&wqe, 0, sizeof(wqe)); 2758 2759 /* Common */ 2760 wqe.num_sge = wr->num_sge; 2761 if (wr->num_sge > qp->qplib_qp.rq.max_sge) { 2762 ibdev_err(&qp->rdev->ibdev, 2763 "Limit exceeded for Receive SGEs"); 2764 rc = -EINVAL; 2765 *bad_wr = wr; 2766 break; 2767 } 2768 2769 payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, 2770 wr->num_sge); 2771 wqe.wr_id = wr->wr_id; 2772 wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV; 2773 2774 if (ib_qp->qp_type == IB_QPT_GSI && 2775 qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI) 2776 rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe, 2777 payload_sz); 2778 if (!rc) 2779 rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe); 2780 if (rc) { 2781 *bad_wr = wr; 2782 break; 2783 } 2784 2785 /* Ring DB if the RQEs posted reaches a threshold value */ 2786 if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) { 2787 bnxt_qplib_post_recv_db(&qp->qplib_qp); 2788 count = 0; 2789 } 2790 2791 wr = wr->next; 2792 } 2793 2794 if (count) 2795 bnxt_qplib_post_recv_db(&qp->qplib_qp); 2796 2797 spin_unlock_irqrestore(&qp->rq_lock, flags); 2798 2799 return rc; 2800 } 2801 2802 /* Completion Queues */ 2803 void bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata) 2804 { 2805 struct bnxt_re_cq *cq; 2806 struct bnxt_qplib_nq *nq; 2807 struct bnxt_re_dev *rdev; 2808 2809 cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq); 2810 rdev = cq->rdev; 2811 nq = cq->qplib_cq.nq; 2812 2813 bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq); 2814 ib_umem_release(cq->umem); 2815 2816 atomic_dec(&rdev->cq_count); 2817 nq->budget--; 2818 kfree(cq->cql); 2819 } 2820 2821 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, 2822 struct ib_udata *udata) 2823 { 2824 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev); 2825 struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr; 2826 struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq); 2827 int rc, entries; 2828 int cqe = attr->cqe; 2829 struct bnxt_qplib_nq *nq = NULL; 2830 unsigned int nq_alloc_cnt; 2831 2832 /* Validate CQ fields */ 2833 if (cqe < 1 || cqe > dev_attr->max_cq_wqes) { 2834 ibdev_err(&rdev->ibdev, "Failed to create CQ -max exceeded"); 2835 return -EINVAL; 2836 } 2837 2838 cq->rdev = rdev; 2839 cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq); 2840 2841 entries = roundup_pow_of_two(cqe + 1); 2842 if (entries > dev_attr->max_cq_wqes + 1) 2843 entries = dev_attr->max_cq_wqes + 1; 2844 2845 cq->qplib_cq.sg_info.pgsize = PAGE_SIZE; 2846 cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT; 2847 if (udata) { 2848 struct bnxt_re_cq_req req; 2849 struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context( 2850 udata, struct bnxt_re_ucontext, ib_uctx); 2851 if (ib_copy_from_udata(&req, udata, sizeof(req))) { 2852 rc = -EFAULT; 2853 goto fail; 2854 } 2855 2856 cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va, 2857 entries * sizeof(struct cq_base), 2858 IB_ACCESS_LOCAL_WRITE); 2859 if (IS_ERR(cq->umem)) { 2860 rc = PTR_ERR(cq->umem); 2861 goto fail; 2862 } 2863 cq->qplib_cq.sg_info.sghead = cq->umem->sg_head.sgl; 2864 cq->qplib_cq.sg_info.npages = ib_umem_num_pages(cq->umem); 2865 cq->qplib_cq.sg_info.nmap = cq->umem->nmap; 2866 cq->qplib_cq.dpi = &uctx->dpi; 2867 } else { 2868 cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL); 2869 cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe), 2870 GFP_KERNEL); 2871 if (!cq->cql) { 2872 rc = -ENOMEM; 2873 goto fail; 2874 } 2875 2876 cq->qplib_cq.dpi = &rdev->dpi_privileged; 2877 } 2878 /* 2879 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a 2880 * used for getting the NQ index. 2881 */ 2882 nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt); 2883 nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)]; 2884 cq->qplib_cq.max_wqe = entries; 2885 cq->qplib_cq.cnq_hw_ring_id = nq->ring_id; 2886 cq->qplib_cq.nq = nq; 2887 2888 rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq); 2889 if (rc) { 2890 ibdev_err(&rdev->ibdev, "Failed to create HW CQ"); 2891 goto fail; 2892 } 2893 2894 cq->ib_cq.cqe = entries; 2895 cq->cq_period = cq->qplib_cq.period; 2896 nq->budget++; 2897 2898 atomic_inc(&rdev->cq_count); 2899 spin_lock_init(&cq->cq_lock); 2900 2901 if (udata) { 2902 struct bnxt_re_cq_resp resp; 2903 2904 resp.cqid = cq->qplib_cq.id; 2905 resp.tail = cq->qplib_cq.hwq.cons; 2906 resp.phase = cq->qplib_cq.period; 2907 resp.rsvd = 0; 2908 rc = ib_copy_to_udata(udata, &resp, sizeof(resp)); 2909 if (rc) { 2910 ibdev_err(&rdev->ibdev, "Failed to copy CQ udata"); 2911 bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq); 2912 goto c2fail; 2913 } 2914 } 2915 2916 return 0; 2917 2918 c2fail: 2919 ib_umem_release(cq->umem); 2920 fail: 2921 kfree(cq->cql); 2922 return rc; 2923 } 2924 2925 static u8 __req_to_ib_wc_status(u8 qstatus) 2926 { 2927 switch (qstatus) { 2928 case CQ_REQ_STATUS_OK: 2929 return IB_WC_SUCCESS; 2930 case CQ_REQ_STATUS_BAD_RESPONSE_ERR: 2931 return IB_WC_BAD_RESP_ERR; 2932 case CQ_REQ_STATUS_LOCAL_LENGTH_ERR: 2933 return IB_WC_LOC_LEN_ERR; 2934 case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR: 2935 return IB_WC_LOC_QP_OP_ERR; 2936 case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR: 2937 return IB_WC_LOC_PROT_ERR; 2938 case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR: 2939 return IB_WC_GENERAL_ERR; 2940 case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR: 2941 return IB_WC_REM_INV_REQ_ERR; 2942 case CQ_REQ_STATUS_REMOTE_ACCESS_ERR: 2943 return IB_WC_REM_ACCESS_ERR; 2944 case CQ_REQ_STATUS_REMOTE_OPERATION_ERR: 2945 return IB_WC_REM_OP_ERR; 2946 case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR: 2947 return IB_WC_RNR_RETRY_EXC_ERR; 2948 case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR: 2949 return IB_WC_RETRY_EXC_ERR; 2950 case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR: 2951 return IB_WC_WR_FLUSH_ERR; 2952 default: 2953 return IB_WC_GENERAL_ERR; 2954 } 2955 return 0; 2956 } 2957 2958 static u8 __rawqp1_to_ib_wc_status(u8 qstatus) 2959 { 2960 switch (qstatus) { 2961 case CQ_RES_RAWETH_QP1_STATUS_OK: 2962 return IB_WC_SUCCESS; 2963 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR: 2964 return IB_WC_LOC_ACCESS_ERR; 2965 case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR: 2966 return IB_WC_LOC_LEN_ERR; 2967 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR: 2968 return IB_WC_LOC_PROT_ERR; 2969 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR: 2970 return IB_WC_LOC_QP_OP_ERR; 2971 case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR: 2972 return IB_WC_GENERAL_ERR; 2973 case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR: 2974 return IB_WC_WR_FLUSH_ERR; 2975 case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR: 2976 return IB_WC_WR_FLUSH_ERR; 2977 default: 2978 return IB_WC_GENERAL_ERR; 2979 } 2980 } 2981 2982 static u8 __rc_to_ib_wc_status(u8 qstatus) 2983 { 2984 switch (qstatus) { 2985 case CQ_RES_RC_STATUS_OK: 2986 return IB_WC_SUCCESS; 2987 case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR: 2988 return IB_WC_LOC_ACCESS_ERR; 2989 case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR: 2990 return IB_WC_LOC_LEN_ERR; 2991 case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR: 2992 return IB_WC_LOC_PROT_ERR; 2993 case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR: 2994 return IB_WC_LOC_QP_OP_ERR; 2995 case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR: 2996 return IB_WC_GENERAL_ERR; 2997 case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR: 2998 return IB_WC_REM_INV_REQ_ERR; 2999 case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR: 3000 return IB_WC_WR_FLUSH_ERR; 3001 case CQ_RES_RC_STATUS_HW_FLUSH_ERR: 3002 return IB_WC_WR_FLUSH_ERR; 3003 default: 3004 return IB_WC_GENERAL_ERR; 3005 } 3006 } 3007 3008 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe) 3009 { 3010 switch (cqe->type) { 3011 case BNXT_QPLIB_SWQE_TYPE_SEND: 3012 wc->opcode = IB_WC_SEND; 3013 break; 3014 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM: 3015 wc->opcode = IB_WC_SEND; 3016 wc->wc_flags |= IB_WC_WITH_IMM; 3017 break; 3018 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV: 3019 wc->opcode = IB_WC_SEND; 3020 wc->wc_flags |= IB_WC_WITH_INVALIDATE; 3021 break; 3022 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE: 3023 wc->opcode = IB_WC_RDMA_WRITE; 3024 break; 3025 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM: 3026 wc->opcode = IB_WC_RDMA_WRITE; 3027 wc->wc_flags |= IB_WC_WITH_IMM; 3028 break; 3029 case BNXT_QPLIB_SWQE_TYPE_RDMA_READ: 3030 wc->opcode = IB_WC_RDMA_READ; 3031 break; 3032 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP: 3033 wc->opcode = IB_WC_COMP_SWAP; 3034 break; 3035 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD: 3036 wc->opcode = IB_WC_FETCH_ADD; 3037 break; 3038 case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV: 3039 wc->opcode = IB_WC_LOCAL_INV; 3040 break; 3041 case BNXT_QPLIB_SWQE_TYPE_REG_MR: 3042 wc->opcode = IB_WC_REG_MR; 3043 break; 3044 default: 3045 wc->opcode = IB_WC_SEND; 3046 break; 3047 } 3048 3049 wc->status = __req_to_ib_wc_status(cqe->status); 3050 } 3051 3052 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags, 3053 u16 raweth_qp1_flags2) 3054 { 3055 bool is_ipv6 = false, is_ipv4 = false; 3056 3057 /* raweth_qp1_flags Bit 9-6 indicates itype */ 3058 if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE) 3059 != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE) 3060 return -1; 3061 3062 if (raweth_qp1_flags2 & 3063 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC && 3064 raweth_qp1_flags2 & 3065 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) { 3066 /* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */ 3067 (raweth_qp1_flags2 & 3068 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ? 3069 (is_ipv6 = true) : (is_ipv4 = true); 3070 return ((is_ipv6) ? 3071 BNXT_RE_ROCEV2_IPV6_PACKET : 3072 BNXT_RE_ROCEV2_IPV4_PACKET); 3073 } else { 3074 return BNXT_RE_ROCE_V1_PACKET; 3075 } 3076 } 3077 3078 static int bnxt_re_to_ib_nw_type(int nw_type) 3079 { 3080 u8 nw_hdr_type = 0xFF; 3081 3082 switch (nw_type) { 3083 case BNXT_RE_ROCE_V1_PACKET: 3084 nw_hdr_type = RDMA_NETWORK_ROCE_V1; 3085 break; 3086 case BNXT_RE_ROCEV2_IPV4_PACKET: 3087 nw_hdr_type = RDMA_NETWORK_IPV4; 3088 break; 3089 case BNXT_RE_ROCEV2_IPV6_PACKET: 3090 nw_hdr_type = RDMA_NETWORK_IPV6; 3091 break; 3092 } 3093 return nw_hdr_type; 3094 } 3095 3096 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev, 3097 void *rq_hdr_buf) 3098 { 3099 u8 *tmp_buf = NULL; 3100 struct ethhdr *eth_hdr; 3101 u16 eth_type; 3102 bool rc = false; 3103 3104 tmp_buf = (u8 *)rq_hdr_buf; 3105 /* 3106 * If dest mac is not same as I/F mac, this could be a 3107 * loopback address or multicast address, check whether 3108 * it is a loopback packet 3109 */ 3110 if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) { 3111 tmp_buf += 4; 3112 /* Check the ether type */ 3113 eth_hdr = (struct ethhdr *)tmp_buf; 3114 eth_type = ntohs(eth_hdr->h_proto); 3115 switch (eth_type) { 3116 case ETH_P_IBOE: 3117 rc = true; 3118 break; 3119 case ETH_P_IP: 3120 case ETH_P_IPV6: { 3121 u32 len; 3122 struct udphdr *udp_hdr; 3123 3124 len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) : 3125 sizeof(struct ipv6hdr)); 3126 tmp_buf += sizeof(struct ethhdr) + len; 3127 udp_hdr = (struct udphdr *)tmp_buf; 3128 if (ntohs(udp_hdr->dest) == 3129 ROCE_V2_UDP_DPORT) 3130 rc = true; 3131 break; 3132 } 3133 default: 3134 break; 3135 } 3136 } 3137 3138 return rc; 3139 } 3140 3141 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *gsi_qp, 3142 struct bnxt_qplib_cqe *cqe) 3143 { 3144 struct bnxt_re_dev *rdev = gsi_qp->rdev; 3145 struct bnxt_re_sqp_entries *sqp_entry = NULL; 3146 struct bnxt_re_qp *gsi_sqp = rdev->gsi_ctx.gsi_sqp; 3147 struct bnxt_re_ah *gsi_sah; 3148 struct ib_send_wr *swr; 3149 struct ib_ud_wr udwr; 3150 struct ib_recv_wr rwr; 3151 int pkt_type = 0; 3152 u32 tbl_idx; 3153 void *rq_hdr_buf; 3154 dma_addr_t rq_hdr_buf_map; 3155 dma_addr_t shrq_hdr_buf_map; 3156 u32 offset = 0; 3157 u32 skip_bytes = 0; 3158 struct ib_sge s_sge[2]; 3159 struct ib_sge r_sge[2]; 3160 int rc; 3161 3162 memset(&udwr, 0, sizeof(udwr)); 3163 memset(&rwr, 0, sizeof(rwr)); 3164 memset(&s_sge, 0, sizeof(s_sge)); 3165 memset(&r_sge, 0, sizeof(r_sge)); 3166 3167 swr = &udwr.wr; 3168 tbl_idx = cqe->wr_id; 3169 3170 rq_hdr_buf = gsi_qp->qplib_qp.rq_hdr_buf + 3171 (tbl_idx * gsi_qp->qplib_qp.rq_hdr_buf_size); 3172 rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp, 3173 tbl_idx); 3174 3175 /* Shadow QP header buffer */ 3176 shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp, 3177 tbl_idx); 3178 sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx]; 3179 3180 /* Store this cqe */ 3181 memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe)); 3182 sqp_entry->qp1_qp = gsi_qp; 3183 3184 /* Find packet type from the cqe */ 3185 3186 pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags, 3187 cqe->raweth_qp1_flags2); 3188 if (pkt_type < 0) { 3189 ibdev_err(&rdev->ibdev, "Invalid packet\n"); 3190 return -EINVAL; 3191 } 3192 3193 /* Adjust the offset for the user buffer and post in the rq */ 3194 3195 if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET) 3196 offset = 20; 3197 3198 /* 3199 * QP1 loopback packet has 4 bytes of internal header before 3200 * ether header. Skip these four bytes. 3201 */ 3202 if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf)) 3203 skip_bytes = 4; 3204 3205 /* First send SGE . Skip the ether header*/ 3206 s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE 3207 + skip_bytes; 3208 s_sge[0].lkey = 0xFFFFFFFF; 3209 s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 : 3210 BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6; 3211 3212 /* Second Send SGE */ 3213 s_sge[1].addr = s_sge[0].addr + s_sge[0].length + 3214 BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE; 3215 if (pkt_type != BNXT_RE_ROCE_V1_PACKET) 3216 s_sge[1].addr += 8; 3217 s_sge[1].lkey = 0xFFFFFFFF; 3218 s_sge[1].length = 256; 3219 3220 /* First recv SGE */ 3221 3222 r_sge[0].addr = shrq_hdr_buf_map; 3223 r_sge[0].lkey = 0xFFFFFFFF; 3224 r_sge[0].length = 40; 3225 3226 r_sge[1].addr = sqp_entry->sge.addr + offset; 3227 r_sge[1].lkey = sqp_entry->sge.lkey; 3228 r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset; 3229 3230 /* Create receive work request */ 3231 rwr.num_sge = 2; 3232 rwr.sg_list = r_sge; 3233 rwr.wr_id = tbl_idx; 3234 rwr.next = NULL; 3235 3236 rc = bnxt_re_post_recv_shadow_qp(rdev, gsi_sqp, &rwr); 3237 if (rc) { 3238 ibdev_err(&rdev->ibdev, 3239 "Failed to post Rx buffers to shadow QP"); 3240 return -ENOMEM; 3241 } 3242 3243 swr->num_sge = 2; 3244 swr->sg_list = s_sge; 3245 swr->wr_id = tbl_idx; 3246 swr->opcode = IB_WR_SEND; 3247 swr->next = NULL; 3248 gsi_sah = rdev->gsi_ctx.gsi_sah; 3249 udwr.ah = &gsi_sah->ib_ah; 3250 udwr.remote_qpn = gsi_sqp->qplib_qp.id; 3251 udwr.remote_qkey = gsi_sqp->qplib_qp.qkey; 3252 3253 /* post data received in the send queue */ 3254 rc = bnxt_re_post_send_shadow_qp(rdev, gsi_sqp, swr); 3255 3256 return 0; 3257 } 3258 3259 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc, 3260 struct bnxt_qplib_cqe *cqe) 3261 { 3262 wc->opcode = IB_WC_RECV; 3263 wc->status = __rawqp1_to_ib_wc_status(cqe->status); 3264 wc->wc_flags |= IB_WC_GRH; 3265 } 3266 3267 static bool bnxt_re_check_if_vlan_valid(struct bnxt_re_dev *rdev, 3268 u16 vlan_id) 3269 { 3270 /* 3271 * Check if the vlan is configured in the host. If not configured, it 3272 * can be a transparent VLAN. So dont report the vlan id. 3273 */ 3274 if (!__vlan_find_dev_deep_rcu(rdev->netdev, 3275 htons(ETH_P_8021Q), vlan_id)) 3276 return false; 3277 return true; 3278 } 3279 3280 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe, 3281 u16 *vid, u8 *sl) 3282 { 3283 bool ret = false; 3284 u32 metadata; 3285 u16 tpid; 3286 3287 metadata = orig_cqe->raweth_qp1_metadata; 3288 if (orig_cqe->raweth_qp1_flags2 & 3289 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) { 3290 tpid = ((metadata & 3291 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >> 3292 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT); 3293 if (tpid == ETH_P_8021Q) { 3294 *vid = metadata & 3295 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK; 3296 *sl = (metadata & 3297 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >> 3298 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT; 3299 ret = true; 3300 } 3301 } 3302 3303 return ret; 3304 } 3305 3306 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc, 3307 struct bnxt_qplib_cqe *cqe) 3308 { 3309 wc->opcode = IB_WC_RECV; 3310 wc->status = __rc_to_ib_wc_status(cqe->status); 3311 3312 if (cqe->flags & CQ_RES_RC_FLAGS_IMM) 3313 wc->wc_flags |= IB_WC_WITH_IMM; 3314 if (cqe->flags & CQ_RES_RC_FLAGS_INV) 3315 wc->wc_flags |= IB_WC_WITH_INVALIDATE; 3316 if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) == 3317 (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) 3318 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM; 3319 } 3320 3321 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *gsi_sqp, 3322 struct ib_wc *wc, 3323 struct bnxt_qplib_cqe *cqe) 3324 { 3325 struct bnxt_re_dev *rdev = gsi_sqp->rdev; 3326 struct bnxt_re_qp *gsi_qp = NULL; 3327 struct bnxt_qplib_cqe *orig_cqe = NULL; 3328 struct bnxt_re_sqp_entries *sqp_entry = NULL; 3329 int nw_type; 3330 u32 tbl_idx; 3331 u16 vlan_id; 3332 u8 sl; 3333 3334 tbl_idx = cqe->wr_id; 3335 3336 sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx]; 3337 gsi_qp = sqp_entry->qp1_qp; 3338 orig_cqe = &sqp_entry->cqe; 3339 3340 wc->wr_id = sqp_entry->wrid; 3341 wc->byte_len = orig_cqe->length; 3342 wc->qp = &gsi_qp->ib_qp; 3343 3344 wc->ex.imm_data = orig_cqe->immdata; 3345 wc->src_qp = orig_cqe->src_qp; 3346 memcpy(wc->smac, orig_cqe->smac, ETH_ALEN); 3347 if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) { 3348 if (bnxt_re_check_if_vlan_valid(rdev, vlan_id)) { 3349 wc->vlan_id = vlan_id; 3350 wc->sl = sl; 3351 wc->wc_flags |= IB_WC_WITH_VLAN; 3352 } 3353 } 3354 wc->port_num = 1; 3355 wc->vendor_err = orig_cqe->status; 3356 3357 wc->opcode = IB_WC_RECV; 3358 wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status); 3359 wc->wc_flags |= IB_WC_GRH; 3360 3361 nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags, 3362 orig_cqe->raweth_qp1_flags2); 3363 if (nw_type >= 0) { 3364 wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type); 3365 wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE; 3366 } 3367 } 3368 3369 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp, 3370 struct ib_wc *wc, 3371 struct bnxt_qplib_cqe *cqe) 3372 { 3373 u8 nw_type; 3374 3375 wc->opcode = IB_WC_RECV; 3376 wc->status = __rc_to_ib_wc_status(cqe->status); 3377 3378 if (cqe->flags & CQ_RES_UD_FLAGS_IMM) 3379 wc->wc_flags |= IB_WC_WITH_IMM; 3380 /* report only on GSI QP for Thor */ 3381 if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) { 3382 wc->wc_flags |= IB_WC_GRH; 3383 memcpy(wc->smac, cqe->smac, ETH_ALEN); 3384 wc->wc_flags |= IB_WC_WITH_SMAC; 3385 if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) { 3386 wc->vlan_id = (cqe->cfa_meta & 0xFFF); 3387 if (wc->vlan_id < 0x1000) 3388 wc->wc_flags |= IB_WC_WITH_VLAN; 3389 } 3390 nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >> 3391 CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT; 3392 wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type); 3393 wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE; 3394 } 3395 3396 } 3397 3398 static int send_phantom_wqe(struct bnxt_re_qp *qp) 3399 { 3400 struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp; 3401 unsigned long flags; 3402 int rc = 0; 3403 3404 spin_lock_irqsave(&qp->sq_lock, flags); 3405 3406 rc = bnxt_re_bind_fence_mw(lib_qp); 3407 if (!rc) { 3408 lib_qp->sq.phantom_wqe_cnt++; 3409 ibdev_dbg(&qp->rdev->ibdev, 3410 "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n", 3411 lib_qp->id, lib_qp->sq.hwq.prod, 3412 HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq), 3413 lib_qp->sq.phantom_wqe_cnt); 3414 } 3415 3416 spin_unlock_irqrestore(&qp->sq_lock, flags); 3417 return rc; 3418 } 3419 3420 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc) 3421 { 3422 struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq); 3423 struct bnxt_re_qp *qp, *sh_qp; 3424 struct bnxt_qplib_cqe *cqe; 3425 int i, ncqe, budget; 3426 struct bnxt_qplib_q *sq; 3427 struct bnxt_qplib_qp *lib_qp; 3428 u32 tbl_idx; 3429 struct bnxt_re_sqp_entries *sqp_entry = NULL; 3430 unsigned long flags; 3431 3432 spin_lock_irqsave(&cq->cq_lock, flags); 3433 budget = min_t(u32, num_entries, cq->max_cql); 3434 num_entries = budget; 3435 if (!cq->cql) { 3436 ibdev_err(&cq->rdev->ibdev, "POLL CQ : no CQL to use"); 3437 goto exit; 3438 } 3439 cqe = &cq->cql[0]; 3440 while (budget) { 3441 lib_qp = NULL; 3442 ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp); 3443 if (lib_qp) { 3444 sq = &lib_qp->sq; 3445 if (sq->send_phantom) { 3446 qp = container_of(lib_qp, 3447 struct bnxt_re_qp, qplib_qp); 3448 if (send_phantom_wqe(qp) == -ENOMEM) 3449 ibdev_err(&cq->rdev->ibdev, 3450 "Phantom failed! Scheduled to send again\n"); 3451 else 3452 sq->send_phantom = false; 3453 } 3454 } 3455 if (ncqe < budget) 3456 ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq, 3457 cqe + ncqe, 3458 budget - ncqe); 3459 3460 if (!ncqe) 3461 break; 3462 3463 for (i = 0; i < ncqe; i++, cqe++) { 3464 /* Transcribe each qplib_wqe back to ib_wc */ 3465 memset(wc, 0, sizeof(*wc)); 3466 3467 wc->wr_id = cqe->wr_id; 3468 wc->byte_len = cqe->length; 3469 qp = container_of 3470 ((struct bnxt_qplib_qp *) 3471 (unsigned long)(cqe->qp_handle), 3472 struct bnxt_re_qp, qplib_qp); 3473 if (!qp) { 3474 ibdev_err(&cq->rdev->ibdev, "POLL CQ : bad QP handle"); 3475 continue; 3476 } 3477 wc->qp = &qp->ib_qp; 3478 wc->ex.imm_data = cqe->immdata; 3479 wc->src_qp = cqe->src_qp; 3480 memcpy(wc->smac, cqe->smac, ETH_ALEN); 3481 wc->port_num = 1; 3482 wc->vendor_err = cqe->status; 3483 3484 switch (cqe->opcode) { 3485 case CQ_BASE_CQE_TYPE_REQ: 3486 sh_qp = qp->rdev->gsi_ctx.gsi_sqp; 3487 if (sh_qp && 3488 qp->qplib_qp.id == sh_qp->qplib_qp.id) { 3489 /* Handle this completion with 3490 * the stored completion 3491 */ 3492 memset(wc, 0, sizeof(*wc)); 3493 continue; 3494 } 3495 bnxt_re_process_req_wc(wc, cqe); 3496 break; 3497 case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1: 3498 if (!cqe->status) { 3499 int rc = 0; 3500 3501 rc = bnxt_re_process_raw_qp_pkt_rx 3502 (qp, cqe); 3503 if (!rc) { 3504 memset(wc, 0, sizeof(*wc)); 3505 continue; 3506 } 3507 cqe->status = -1; 3508 } 3509 /* Errors need not be looped back. 3510 * But change the wr_id to the one 3511 * stored in the table 3512 */ 3513 tbl_idx = cqe->wr_id; 3514 sqp_entry = &cq->rdev->gsi_ctx.sqp_tbl[tbl_idx]; 3515 wc->wr_id = sqp_entry->wrid; 3516 bnxt_re_process_res_rawqp1_wc(wc, cqe); 3517 break; 3518 case CQ_BASE_CQE_TYPE_RES_RC: 3519 bnxt_re_process_res_rc_wc(wc, cqe); 3520 break; 3521 case CQ_BASE_CQE_TYPE_RES_UD: 3522 sh_qp = qp->rdev->gsi_ctx.gsi_sqp; 3523 if (sh_qp && 3524 qp->qplib_qp.id == sh_qp->qplib_qp.id) { 3525 /* Handle this completion with 3526 * the stored completion 3527 */ 3528 if (cqe->status) { 3529 continue; 3530 } else { 3531 bnxt_re_process_res_shadow_qp_wc 3532 (qp, wc, cqe); 3533 break; 3534 } 3535 } 3536 bnxt_re_process_res_ud_wc(qp, wc, cqe); 3537 break; 3538 default: 3539 ibdev_err(&cq->rdev->ibdev, 3540 "POLL CQ : type 0x%x not handled", 3541 cqe->opcode); 3542 continue; 3543 } 3544 wc++; 3545 budget--; 3546 } 3547 } 3548 exit: 3549 spin_unlock_irqrestore(&cq->cq_lock, flags); 3550 return num_entries - budget; 3551 } 3552 3553 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq, 3554 enum ib_cq_notify_flags ib_cqn_flags) 3555 { 3556 struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq); 3557 int type = 0, rc = 0; 3558 unsigned long flags; 3559 3560 spin_lock_irqsave(&cq->cq_lock, flags); 3561 /* Trigger on the very next completion */ 3562 if (ib_cqn_flags & IB_CQ_NEXT_COMP) 3563 type = DBC_DBC_TYPE_CQ_ARMALL; 3564 /* Trigger on the next solicited completion */ 3565 else if (ib_cqn_flags & IB_CQ_SOLICITED) 3566 type = DBC_DBC_TYPE_CQ_ARMSE; 3567 3568 /* Poll to see if there are missed events */ 3569 if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) && 3570 !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) { 3571 rc = 1; 3572 goto exit; 3573 } 3574 bnxt_qplib_req_notify_cq(&cq->qplib_cq, type); 3575 3576 exit: 3577 spin_unlock_irqrestore(&cq->cq_lock, flags); 3578 return rc; 3579 } 3580 3581 /* Memory Regions */ 3582 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags) 3583 { 3584 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 3585 struct bnxt_re_dev *rdev = pd->rdev; 3586 struct bnxt_re_mr *mr; 3587 u64 pbl = 0; 3588 int rc; 3589 3590 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 3591 if (!mr) 3592 return ERR_PTR(-ENOMEM); 3593 3594 mr->rdev = rdev; 3595 mr->qplib_mr.pd = &pd->qplib_pd; 3596 mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags); 3597 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR; 3598 3599 /* Allocate and register 0 as the address */ 3600 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 3601 if (rc) 3602 goto fail; 3603 3604 mr->qplib_mr.hwq.level = PBL_LVL_MAX; 3605 mr->qplib_mr.total_size = -1; /* Infinte length */ 3606 rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl, 0, false, 3607 PAGE_SIZE); 3608 if (rc) 3609 goto fail_mr; 3610 3611 mr->ib_mr.lkey = mr->qplib_mr.lkey; 3612 if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ | 3613 IB_ACCESS_REMOTE_ATOMIC)) 3614 mr->ib_mr.rkey = mr->ib_mr.lkey; 3615 atomic_inc(&rdev->mr_count); 3616 3617 return &mr->ib_mr; 3618 3619 fail_mr: 3620 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 3621 fail: 3622 kfree(mr); 3623 return ERR_PTR(rc); 3624 } 3625 3626 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata) 3627 { 3628 struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr); 3629 struct bnxt_re_dev *rdev = mr->rdev; 3630 int rc; 3631 3632 rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 3633 if (rc) { 3634 ibdev_err(&rdev->ibdev, "Dereg MR failed: %#x\n", rc); 3635 return rc; 3636 } 3637 3638 if (mr->pages) { 3639 rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res, 3640 &mr->qplib_frpl); 3641 kfree(mr->pages); 3642 mr->npages = 0; 3643 mr->pages = NULL; 3644 } 3645 ib_umem_release(mr->ib_umem); 3646 3647 kfree(mr); 3648 atomic_dec(&rdev->mr_count); 3649 return rc; 3650 } 3651 3652 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr) 3653 { 3654 struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr); 3655 3656 if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs)) 3657 return -ENOMEM; 3658 3659 mr->pages[mr->npages++] = addr; 3660 return 0; 3661 } 3662 3663 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents, 3664 unsigned int *sg_offset) 3665 { 3666 struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr); 3667 3668 mr->npages = 0; 3669 return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page); 3670 } 3671 3672 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type, 3673 u32 max_num_sg) 3674 { 3675 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 3676 struct bnxt_re_dev *rdev = pd->rdev; 3677 struct bnxt_re_mr *mr = NULL; 3678 int rc; 3679 3680 if (type != IB_MR_TYPE_MEM_REG) { 3681 ibdev_dbg(&rdev->ibdev, "MR type 0x%x not supported", type); 3682 return ERR_PTR(-EINVAL); 3683 } 3684 if (max_num_sg > MAX_PBL_LVL_1_PGS) 3685 return ERR_PTR(-EINVAL); 3686 3687 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 3688 if (!mr) 3689 return ERR_PTR(-ENOMEM); 3690 3691 mr->rdev = rdev; 3692 mr->qplib_mr.pd = &pd->qplib_pd; 3693 mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR; 3694 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR; 3695 3696 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 3697 if (rc) 3698 goto bail; 3699 3700 mr->ib_mr.lkey = mr->qplib_mr.lkey; 3701 mr->ib_mr.rkey = mr->ib_mr.lkey; 3702 3703 mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL); 3704 if (!mr->pages) { 3705 rc = -ENOMEM; 3706 goto fail; 3707 } 3708 rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res, 3709 &mr->qplib_frpl, max_num_sg); 3710 if (rc) { 3711 ibdev_err(&rdev->ibdev, 3712 "Failed to allocate HW FR page list"); 3713 goto fail_mr; 3714 } 3715 3716 atomic_inc(&rdev->mr_count); 3717 return &mr->ib_mr; 3718 3719 fail_mr: 3720 kfree(mr->pages); 3721 fail: 3722 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 3723 bail: 3724 kfree(mr); 3725 return ERR_PTR(rc); 3726 } 3727 3728 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type, 3729 struct ib_udata *udata) 3730 { 3731 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 3732 struct bnxt_re_dev *rdev = pd->rdev; 3733 struct bnxt_re_mw *mw; 3734 int rc; 3735 3736 mw = kzalloc(sizeof(*mw), GFP_KERNEL); 3737 if (!mw) 3738 return ERR_PTR(-ENOMEM); 3739 mw->rdev = rdev; 3740 mw->qplib_mw.pd = &pd->qplib_pd; 3741 3742 mw->qplib_mw.type = (type == IB_MW_TYPE_1 ? 3743 CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 : 3744 CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B); 3745 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw); 3746 if (rc) { 3747 ibdev_err(&rdev->ibdev, "Allocate MW failed!"); 3748 goto fail; 3749 } 3750 mw->ib_mw.rkey = mw->qplib_mw.rkey; 3751 3752 atomic_inc(&rdev->mw_count); 3753 return &mw->ib_mw; 3754 3755 fail: 3756 kfree(mw); 3757 return ERR_PTR(rc); 3758 } 3759 3760 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw) 3761 { 3762 struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw); 3763 struct bnxt_re_dev *rdev = mw->rdev; 3764 int rc; 3765 3766 rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw); 3767 if (rc) { 3768 ibdev_err(&rdev->ibdev, "Free MW failed: %#x\n", rc); 3769 return rc; 3770 } 3771 3772 kfree(mw); 3773 atomic_dec(&rdev->mw_count); 3774 return rc; 3775 } 3776 3777 static int bnxt_re_page_size_ok(int page_shift) 3778 { 3779 switch (page_shift) { 3780 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4K: 3781 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_8K: 3782 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_64K: 3783 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_2M: 3784 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_256K: 3785 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1M: 3786 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4M: 3787 case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1G: 3788 return 1; 3789 default: 3790 return 0; 3791 } 3792 } 3793 3794 static int fill_umem_pbl_tbl(struct ib_umem *umem, u64 *pbl_tbl_orig, 3795 int page_shift) 3796 { 3797 u64 *pbl_tbl = pbl_tbl_orig; 3798 u64 page_size = BIT_ULL(page_shift); 3799 struct ib_block_iter biter; 3800 3801 rdma_for_each_block(umem->sg_head.sgl, &biter, umem->nmap, page_size) 3802 *pbl_tbl++ = rdma_block_iter_dma_address(&biter); 3803 3804 return pbl_tbl - pbl_tbl_orig; 3805 } 3806 3807 /* uverbs */ 3808 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length, 3809 u64 virt_addr, int mr_access_flags, 3810 struct ib_udata *udata) 3811 { 3812 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 3813 struct bnxt_re_dev *rdev = pd->rdev; 3814 struct bnxt_re_mr *mr; 3815 struct ib_umem *umem; 3816 u64 *pbl_tbl = NULL; 3817 int umem_pgs, page_shift, rc; 3818 3819 if (length > BNXT_RE_MAX_MR_SIZE) { 3820 ibdev_err(&rdev->ibdev, "MR Size: %lld > Max supported:%lld\n", 3821 length, BNXT_RE_MAX_MR_SIZE); 3822 return ERR_PTR(-ENOMEM); 3823 } 3824 3825 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 3826 if (!mr) 3827 return ERR_PTR(-ENOMEM); 3828 3829 mr->rdev = rdev; 3830 mr->qplib_mr.pd = &pd->qplib_pd; 3831 mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags); 3832 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR; 3833 3834 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 3835 if (rc) { 3836 ibdev_err(&rdev->ibdev, "Failed to allocate MR"); 3837 goto free_mr; 3838 } 3839 /* The fixed portion of the rkey is the same as the lkey */ 3840 mr->ib_mr.rkey = mr->qplib_mr.rkey; 3841 3842 umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags); 3843 if (IS_ERR(umem)) { 3844 ibdev_err(&rdev->ibdev, "Failed to get umem"); 3845 rc = -EFAULT; 3846 goto free_mrw; 3847 } 3848 mr->ib_umem = umem; 3849 3850 mr->qplib_mr.va = virt_addr; 3851 umem_pgs = ib_umem_page_count(umem); 3852 if (!umem_pgs) { 3853 ibdev_err(&rdev->ibdev, "umem is invalid!"); 3854 rc = -EINVAL; 3855 goto free_umem; 3856 } 3857 mr->qplib_mr.total_size = length; 3858 3859 pbl_tbl = kcalloc(umem_pgs, sizeof(u64 *), GFP_KERNEL); 3860 if (!pbl_tbl) { 3861 rc = -ENOMEM; 3862 goto free_umem; 3863 } 3864 3865 page_shift = __ffs(ib_umem_find_best_pgsz(umem, 3866 BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M, 3867 virt_addr)); 3868 3869 if (!bnxt_re_page_size_ok(page_shift)) { 3870 ibdev_err(&rdev->ibdev, "umem page size unsupported!"); 3871 rc = -EFAULT; 3872 goto fail; 3873 } 3874 3875 if (page_shift == BNXT_RE_PAGE_SHIFT_4K && 3876 length > BNXT_RE_MAX_MR_SIZE_LOW) { 3877 ibdev_err(&rdev->ibdev, "Requested MR Sz:%llu Max sup:%llu", 3878 length, (u64)BNXT_RE_MAX_MR_SIZE_LOW); 3879 rc = -EINVAL; 3880 goto fail; 3881 } 3882 3883 /* Map umem buf ptrs to the PBL */ 3884 umem_pgs = fill_umem_pbl_tbl(umem, pbl_tbl, page_shift); 3885 rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, pbl_tbl, 3886 umem_pgs, false, 1 << page_shift); 3887 if (rc) { 3888 ibdev_err(&rdev->ibdev, "Failed to register user MR"); 3889 goto fail; 3890 } 3891 3892 kfree(pbl_tbl); 3893 3894 mr->ib_mr.lkey = mr->qplib_mr.lkey; 3895 mr->ib_mr.rkey = mr->qplib_mr.lkey; 3896 atomic_inc(&rdev->mr_count); 3897 3898 return &mr->ib_mr; 3899 fail: 3900 kfree(pbl_tbl); 3901 free_umem: 3902 ib_umem_release(umem); 3903 free_mrw: 3904 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 3905 free_mr: 3906 kfree(mr); 3907 return ERR_PTR(rc); 3908 } 3909 3910 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata) 3911 { 3912 struct ib_device *ibdev = ctx->device; 3913 struct bnxt_re_ucontext *uctx = 3914 container_of(ctx, struct bnxt_re_ucontext, ib_uctx); 3915 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 3916 struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr; 3917 struct bnxt_re_uctx_resp resp; 3918 u32 chip_met_rev_num = 0; 3919 int rc; 3920 3921 ibdev_dbg(ibdev, "ABI version requested %u", ibdev->ops.uverbs_abi_ver); 3922 3923 if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) { 3924 ibdev_dbg(ibdev, " is different from the device %d ", 3925 BNXT_RE_ABI_VERSION); 3926 return -EPERM; 3927 } 3928 3929 uctx->rdev = rdev; 3930 3931 uctx->shpg = (void *)__get_free_page(GFP_KERNEL); 3932 if (!uctx->shpg) { 3933 rc = -ENOMEM; 3934 goto fail; 3935 } 3936 spin_lock_init(&uctx->sh_lock); 3937 3938 resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX; 3939 chip_met_rev_num = rdev->chip_ctx->chip_num; 3940 chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_rev & 0xFF) << 3941 BNXT_RE_CHIP_ID0_CHIP_REV_SFT; 3942 chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_metal & 0xFF) << 3943 BNXT_RE_CHIP_ID0_CHIP_MET_SFT; 3944 resp.chip_id0 = chip_met_rev_num; 3945 /* Future extension of chip info */ 3946 resp.chip_id1 = 0; 3947 /*Temp, Use xa_alloc instead */ 3948 resp.dev_id = rdev->en_dev->pdev->devfn; 3949 resp.max_qp = rdev->qplib_ctx.qpc_count; 3950 resp.pg_size = PAGE_SIZE; 3951 resp.cqe_sz = sizeof(struct cq_base); 3952 resp.max_cqd = dev_attr->max_cq_wqes; 3953 resp.rsvd = 0; 3954 3955 rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp))); 3956 if (rc) { 3957 ibdev_err(ibdev, "Failed to copy user context"); 3958 rc = -EFAULT; 3959 goto cfail; 3960 } 3961 3962 return 0; 3963 cfail: 3964 free_page((unsigned long)uctx->shpg); 3965 uctx->shpg = NULL; 3966 fail: 3967 return rc; 3968 } 3969 3970 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx) 3971 { 3972 struct bnxt_re_ucontext *uctx = container_of(ib_uctx, 3973 struct bnxt_re_ucontext, 3974 ib_uctx); 3975 3976 struct bnxt_re_dev *rdev = uctx->rdev; 3977 3978 if (uctx->shpg) 3979 free_page((unsigned long)uctx->shpg); 3980 3981 if (uctx->dpi.dbr) { 3982 /* Free DPI only if this is the first PD allocated by the 3983 * application and mark the context dpi as NULL 3984 */ 3985 bnxt_qplib_dealloc_dpi(&rdev->qplib_res, 3986 &rdev->qplib_res.dpi_tbl, &uctx->dpi); 3987 uctx->dpi.dbr = NULL; 3988 } 3989 } 3990 3991 /* Helper function to mmap the virtual memory from user app */ 3992 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma) 3993 { 3994 struct bnxt_re_ucontext *uctx = container_of(ib_uctx, 3995 struct bnxt_re_ucontext, 3996 ib_uctx); 3997 struct bnxt_re_dev *rdev = uctx->rdev; 3998 u64 pfn; 3999 4000 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 4001 return -EINVAL; 4002 4003 if (vma->vm_pgoff) { 4004 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 4005 if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff, 4006 PAGE_SIZE, vma->vm_page_prot)) { 4007 ibdev_err(&rdev->ibdev, "Failed to map DPI"); 4008 return -EAGAIN; 4009 } 4010 } else { 4011 pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT; 4012 if (remap_pfn_range(vma, vma->vm_start, 4013 pfn, PAGE_SIZE, vma->vm_page_prot)) { 4014 ibdev_err(&rdev->ibdev, "Failed to map shared page"); 4015 return -EAGAIN; 4016 } 4017 } 4018 4019 return 0; 4020 } 4021