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 #include <net/addrconf.h> 45 46 #include <rdma/ib_verbs.h> 47 #include <rdma/ib_user_verbs.h> 48 #include <rdma/ib_umem.h> 49 #include <rdma/ib_addr.h> 50 #include <rdma/ib_mad.h> 51 #include <rdma/ib_cache.h> 52 #include <rdma/uverbs_ioctl.h> 53 #include <linux/hashtable.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 "debugfs.h" 64 65 #include <rdma/uverbs_types.h> 66 #include <rdma/uverbs_std_types.h> 67 68 #include <rdma/ib_user_ioctl_cmds.h> 69 70 #define UVERBS_MODULE_NAME bnxt_re 71 #include <rdma/uverbs_named_ioctl.h> 72 73 #include <rdma/bnxt_re-abi.h> 74 75 static int __from_ib_access_flags(int iflags) 76 { 77 int qflags = 0; 78 79 if (iflags & IB_ACCESS_LOCAL_WRITE) 80 qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE; 81 if (iflags & IB_ACCESS_REMOTE_READ) 82 qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ; 83 if (iflags & IB_ACCESS_REMOTE_WRITE) 84 qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE; 85 if (iflags & IB_ACCESS_REMOTE_ATOMIC) 86 qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC; 87 if (iflags & IB_ACCESS_MW_BIND) 88 qflags |= BNXT_QPLIB_ACCESS_MW_BIND; 89 if (iflags & IB_ZERO_BASED) 90 qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED; 91 if (iflags & IB_ACCESS_ON_DEMAND) 92 qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND; 93 return qflags; 94 }; 95 96 static int __to_ib_access_flags(int qflags) 97 { 98 int iflags = 0; 99 100 if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE) 101 iflags |= IB_ACCESS_LOCAL_WRITE; 102 if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE) 103 iflags |= IB_ACCESS_REMOTE_WRITE; 104 if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ) 105 iflags |= IB_ACCESS_REMOTE_READ; 106 if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC) 107 iflags |= IB_ACCESS_REMOTE_ATOMIC; 108 if (qflags & BNXT_QPLIB_ACCESS_MW_BIND) 109 iflags |= IB_ACCESS_MW_BIND; 110 if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED) 111 iflags |= IB_ZERO_BASED; 112 if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND) 113 iflags |= IB_ACCESS_ON_DEMAND; 114 return iflags; 115 } 116 117 static u8 __qp_access_flags_from_ib(struct bnxt_qplib_chip_ctx *cctx, int iflags) 118 { 119 u8 qflags = 0; 120 121 if (!bnxt_qplib_is_chip_gen_p5_p7(cctx)) 122 /* For Wh+ */ 123 return (u8)__from_ib_access_flags(iflags); 124 125 /* For P5, P7 and later chips */ 126 if (iflags & IB_ACCESS_LOCAL_WRITE) 127 qflags |= CMDQ_MODIFY_QP_ACCESS_LOCAL_WRITE; 128 if (iflags & IB_ACCESS_REMOTE_WRITE) 129 qflags |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE; 130 if (iflags & IB_ACCESS_REMOTE_READ) 131 qflags |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ; 132 if (iflags & IB_ACCESS_REMOTE_ATOMIC) 133 qflags |= CMDQ_MODIFY_QP_ACCESS_REMOTE_ATOMIC; 134 135 return qflags; 136 } 137 138 static int __qp_access_flags_to_ib(struct bnxt_qplib_chip_ctx *cctx, u8 qflags) 139 { 140 int iflags = 0; 141 142 if (!bnxt_qplib_is_chip_gen_p5_p7(cctx)) 143 /* For Wh+ */ 144 return __to_ib_access_flags(qflags); 145 146 /* For P5, P7 and later chips */ 147 if (qflags & CMDQ_MODIFY_QP_ACCESS_LOCAL_WRITE) 148 iflags |= IB_ACCESS_LOCAL_WRITE; 149 if (qflags & CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE) 150 iflags |= IB_ACCESS_REMOTE_WRITE; 151 if (qflags & CMDQ_MODIFY_QP_ACCESS_REMOTE_READ) 152 iflags |= IB_ACCESS_REMOTE_READ; 153 if (qflags & CMDQ_MODIFY_QP_ACCESS_REMOTE_ATOMIC) 154 iflags |= IB_ACCESS_REMOTE_ATOMIC; 155 156 return iflags; 157 } 158 159 static void bnxt_re_check_and_set_relaxed_ordering(struct bnxt_re_dev *rdev, 160 struct bnxt_qplib_mrw *qplib_mr) 161 { 162 if (_is_relaxed_ordering_supported(rdev->dev_attr->dev_cap_flags2) && 163 pcie_relaxed_ordering_enabled(rdev->en_dev->pdev)) 164 qplib_mr->flags |= CMDQ_REGISTER_MR_FLAGS_ENABLE_RO; 165 } 166 167 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list, 168 struct bnxt_qplib_sge *sg_list, int num) 169 { 170 int i, total = 0; 171 172 for (i = 0; i < num; i++) { 173 sg_list[i].addr = ib_sg_list[i].addr; 174 sg_list[i].lkey = ib_sg_list[i].lkey; 175 sg_list[i].size = ib_sg_list[i].length; 176 total += sg_list[i].size; 177 } 178 return total; 179 } 180 181 /* Device */ 182 int bnxt_re_query_device(struct ib_device *ibdev, 183 struct ib_device_attr *ib_attr, 184 struct ib_udata *udata) 185 { 186 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 187 struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr; 188 189 memset(ib_attr, 0, sizeof(*ib_attr)); 190 memcpy(&ib_attr->fw_ver, dev_attr->fw_ver, 191 min(sizeof(dev_attr->fw_ver), 192 sizeof(ib_attr->fw_ver))); 193 addrconf_addr_eui48((u8 *)&ib_attr->sys_image_guid, 194 rdev->netdev->dev_addr); 195 ib_attr->max_mr_size = BNXT_RE_MAX_MR_SIZE; 196 ib_attr->page_size_cap = BNXT_RE_PAGE_SIZE_SUPPORTED; 197 198 ib_attr->vendor_id = rdev->en_dev->pdev->vendor; 199 ib_attr->vendor_part_id = rdev->en_dev->pdev->device; 200 ib_attr->hw_ver = rdev->en_dev->pdev->revision; 201 ib_attr->max_qp = dev_attr->max_qp; 202 ib_attr->max_qp_wr = dev_attr->max_qp_wqes; 203 ib_attr->device_cap_flags = 204 IB_DEVICE_CURR_QP_STATE_MOD 205 | IB_DEVICE_RC_RNR_NAK_GEN 206 | IB_DEVICE_SHUTDOWN_PORT 207 | IB_DEVICE_SYS_IMAGE_GUID 208 | IB_DEVICE_RESIZE_MAX_WR 209 | IB_DEVICE_PORT_ACTIVE_EVENT 210 | IB_DEVICE_N_NOTIFY_CQ 211 | IB_DEVICE_MEM_WINDOW 212 | IB_DEVICE_MEM_WINDOW_TYPE_2B 213 | IB_DEVICE_MEM_MGT_EXTENSIONS; 214 ib_attr->kernel_cap_flags = IBK_LOCAL_DMA_LKEY; 215 ib_attr->max_send_sge = dev_attr->max_qp_sges; 216 ib_attr->max_recv_sge = dev_attr->max_qp_sges; 217 ib_attr->max_sge_rd = dev_attr->max_qp_sges; 218 ib_attr->max_cq = dev_attr->max_cq; 219 ib_attr->max_cqe = dev_attr->max_cq_wqes; 220 ib_attr->max_mr = dev_attr->max_mr; 221 ib_attr->max_pd = dev_attr->max_pd; 222 ib_attr->max_qp_rd_atom = dev_attr->max_qp_rd_atom; 223 ib_attr->max_qp_init_rd_atom = dev_attr->max_qp_init_rd_atom; 224 ib_attr->atomic_cap = IB_ATOMIC_NONE; 225 ib_attr->masked_atomic_cap = IB_ATOMIC_NONE; 226 if (dev_attr->is_atomic) { 227 ib_attr->atomic_cap = IB_ATOMIC_GLOB; 228 ib_attr->masked_atomic_cap = IB_ATOMIC_GLOB; 229 } 230 231 ib_attr->max_ee_rd_atom = 0; 232 ib_attr->max_res_rd_atom = 0; 233 ib_attr->max_ee_init_rd_atom = 0; 234 ib_attr->max_ee = 0; 235 ib_attr->max_rdd = 0; 236 ib_attr->max_mw = dev_attr->max_mw; 237 ib_attr->max_raw_ipv6_qp = 0; 238 ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp; 239 ib_attr->max_mcast_grp = 0; 240 ib_attr->max_mcast_qp_attach = 0; 241 ib_attr->max_total_mcast_qp_attach = 0; 242 ib_attr->max_ah = dev_attr->max_ah; 243 244 ib_attr->max_srq = dev_attr->max_srq; 245 ib_attr->max_srq_wr = dev_attr->max_srq_wqes; 246 ib_attr->max_srq_sge = dev_attr->max_srq_sges; 247 248 ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS; 249 250 ib_attr->max_pkeys = 1; 251 ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY; 252 return 0; 253 } 254 255 int bnxt_re_modify_device(struct ib_device *ibdev, 256 int device_modify_mask, 257 struct ib_device_modify *device_modify) 258 { 259 ibdev_dbg(ibdev, "Modify device with mask 0x%x", device_modify_mask); 260 261 if (device_modify_mask & ~IB_DEVICE_MODIFY_NODE_DESC) 262 return -EOPNOTSUPP; 263 264 if (!(device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)) 265 return 0; 266 267 memcpy(ibdev->node_desc, device_modify->node_desc, IB_DEVICE_NODE_DESC_MAX); 268 return 0; 269 } 270 271 /* Port */ 272 int bnxt_re_query_port(struct ib_device *ibdev, u32 port_num, 273 struct ib_port_attr *port_attr) 274 { 275 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 276 struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr; 277 int rc; 278 279 memset(port_attr, 0, sizeof(*port_attr)); 280 281 if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) { 282 port_attr->state = IB_PORT_ACTIVE; 283 port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP; 284 } else { 285 port_attr->state = IB_PORT_DOWN; 286 port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED; 287 } 288 port_attr->max_mtu = IB_MTU_4096; 289 port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu); 290 port_attr->gid_tbl_len = dev_attr->max_sgid; 291 port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP | 292 IB_PORT_DEVICE_MGMT_SUP | 293 IB_PORT_VENDOR_CLASS_SUP; 294 port_attr->ip_gids = true; 295 296 port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW; 297 port_attr->bad_pkey_cntr = 0; 298 port_attr->qkey_viol_cntr = 0; 299 port_attr->pkey_tbl_len = dev_attr->max_pkey; 300 port_attr->lid = 0; 301 port_attr->sm_lid = 0; 302 port_attr->lmc = 0; 303 port_attr->max_vl_num = 4; 304 port_attr->sm_sl = 0; 305 port_attr->subnet_timeout = 0; 306 port_attr->init_type_reply = 0; 307 rc = ib_get_eth_speed(&rdev->ibdev, port_num, &port_attr->active_speed, 308 &port_attr->active_width); 309 310 return rc; 311 } 312 313 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u32 port_num, 314 struct ib_port_immutable *immutable) 315 { 316 struct ib_port_attr port_attr; 317 318 if (bnxt_re_query_port(ibdev, port_num, &port_attr)) 319 return -EINVAL; 320 321 immutable->pkey_tbl_len = port_attr.pkey_tbl_len; 322 immutable->gid_tbl_len = port_attr.gid_tbl_len; 323 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE; 324 immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP; 325 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 326 return 0; 327 } 328 329 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str) 330 { 331 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 332 333 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d", 334 rdev->dev_attr->fw_ver[0], rdev->dev_attr->fw_ver[1], 335 rdev->dev_attr->fw_ver[2], rdev->dev_attr->fw_ver[3]); 336 } 337 338 int bnxt_re_query_pkey(struct ib_device *ibdev, u32 port_num, 339 u16 index, u16 *pkey) 340 { 341 if (index > 0) 342 return -EINVAL; 343 344 *pkey = IB_DEFAULT_PKEY_FULL; 345 346 return 0; 347 } 348 349 int bnxt_re_query_gid(struct ib_device *ibdev, u32 port_num, 350 int index, union ib_gid *gid) 351 { 352 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 353 int rc; 354 355 /* Ignore port_num */ 356 memset(gid, 0, sizeof(*gid)); 357 rc = bnxt_qplib_get_sgid(&rdev->qplib_res, 358 &rdev->qplib_res.sgid_tbl, index, 359 (struct bnxt_qplib_gid *)gid); 360 return rc; 361 } 362 363 int bnxt_re_del_gid(const struct ib_gid_attr *attr, void **context) 364 { 365 int rc = 0; 366 struct bnxt_re_gid_ctx *ctx, **ctx_tbl; 367 struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev); 368 struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl; 369 struct bnxt_qplib_gid *gid_to_del; 370 u16 vlan_id = 0xFFFF; 371 372 /* Delete the entry from the hardware */ 373 ctx = *context; 374 if (!ctx) 375 return -EINVAL; 376 377 if (sgid_tbl && sgid_tbl->active) { 378 if (ctx->idx >= sgid_tbl->max) 379 return -EINVAL; 380 gid_to_del = &sgid_tbl->tbl[ctx->idx].gid; 381 vlan_id = sgid_tbl->tbl[ctx->idx].vlan_id; 382 /* DEL_GID is called in WQ context(netdevice_event_work_handler) 383 * or via the ib_unregister_device path. In the former case QP1 384 * may not be destroyed yet, in which case just return as FW 385 * needs that entry to be present and will fail it's deletion. 386 * We could get invoked again after QP1 is destroyed OR get an 387 * ADD_GID call with a different GID value for the same index 388 * where we issue MODIFY_GID cmd to update the GID entry -- TBD 389 */ 390 if (ctx->idx == 0 && 391 rdma_link_local_addr((struct in6_addr *)gid_to_del) && 392 ctx->refcnt == 1 && rdev->gsi_ctx.gsi_sqp) { 393 ibdev_dbg(&rdev->ibdev, 394 "Trying to delete GID0 while QP1 is alive\n"); 395 return -EFAULT; 396 } 397 ctx->refcnt--; 398 if (!ctx->refcnt) { 399 rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del, 400 vlan_id, true); 401 if (rc) { 402 ibdev_err(&rdev->ibdev, 403 "Failed to remove GID: %#x", rc); 404 } else { 405 ctx_tbl = sgid_tbl->ctx; 406 ctx_tbl[ctx->idx] = NULL; 407 kfree(ctx); 408 } 409 } 410 } else { 411 return -EINVAL; 412 } 413 return rc; 414 } 415 416 int bnxt_re_add_gid(const struct ib_gid_attr *attr, void **context) 417 { 418 int rc; 419 u32 tbl_idx = 0; 420 u16 vlan_id = 0xFFFF; 421 struct bnxt_re_gid_ctx *ctx, **ctx_tbl; 422 struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev); 423 struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl; 424 425 rc = rdma_read_gid_l2_fields(attr, &vlan_id, NULL); 426 if (rc) 427 return rc; 428 429 rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)&attr->gid, 430 rdev->qplib_res.netdev->dev_addr, 431 vlan_id, true, &tbl_idx); 432 if (rc == -EALREADY) { 433 ctx_tbl = sgid_tbl->ctx; 434 ctx_tbl[tbl_idx]->refcnt++; 435 *context = ctx_tbl[tbl_idx]; 436 return 0; 437 } 438 439 if (rc < 0) { 440 ibdev_err(&rdev->ibdev, "Failed to add GID: %#x", rc); 441 return rc; 442 } 443 444 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 445 if (!ctx) 446 return -ENOMEM; 447 ctx_tbl = sgid_tbl->ctx; 448 ctx->idx = tbl_idx; 449 ctx->refcnt = 1; 450 ctx_tbl[tbl_idx] = ctx; 451 *context = ctx; 452 453 return rc; 454 } 455 456 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev, 457 u32 port_num) 458 { 459 return IB_LINK_LAYER_ETHERNET; 460 } 461 462 #define BNXT_RE_FENCE_PBL_SIZE DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE) 463 464 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd) 465 { 466 struct bnxt_re_fence_data *fence = &pd->fence; 467 struct ib_mr *ib_mr = &fence->mr->ib_mr; 468 struct bnxt_qplib_swqe *wqe = &fence->bind_wqe; 469 struct bnxt_re_dev *rdev = pd->rdev; 470 471 if (bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx)) 472 return; 473 474 memset(wqe, 0, sizeof(*wqe)); 475 wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW; 476 wqe->wr_id = BNXT_QPLIB_FENCE_WRID; 477 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 478 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 479 wqe->bind.zero_based = false; 480 wqe->bind.parent_l_key = ib_mr->lkey; 481 wqe->bind.va = (u64)(unsigned long)fence->va; 482 wqe->bind.length = fence->size; 483 wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ); 484 wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1; 485 486 /* Save the initial rkey in fence structure for now; 487 * wqe->bind.r_key will be set at (re)bind time. 488 */ 489 fence->bind_rkey = ib_inc_rkey(fence->mw->rkey); 490 } 491 492 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp) 493 { 494 struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp, 495 qplib_qp); 496 struct ib_pd *ib_pd = qp->ib_qp.pd; 497 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 498 struct bnxt_re_fence_data *fence = &pd->fence; 499 struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe; 500 struct bnxt_qplib_swqe wqe; 501 int rc; 502 503 memcpy(&wqe, fence_wqe, sizeof(wqe)); 504 wqe.bind.r_key = fence->bind_rkey; 505 fence->bind_rkey = ib_inc_rkey(fence->bind_rkey); 506 507 ibdev_dbg(&qp->rdev->ibdev, 508 "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n", 509 wqe.bind.r_key, qp->qplib_qp.id, pd); 510 rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe); 511 if (rc) { 512 ibdev_err(&qp->rdev->ibdev, "Failed to bind fence-WQE\n"); 513 return rc; 514 } 515 bnxt_qplib_post_send_db(&qp->qplib_qp); 516 517 return rc; 518 } 519 520 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd) 521 { 522 struct bnxt_re_fence_data *fence = &pd->fence; 523 struct bnxt_re_dev *rdev = pd->rdev; 524 struct device *dev = &rdev->en_dev->pdev->dev; 525 struct bnxt_re_mr *mr = fence->mr; 526 527 if (bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx)) 528 return; 529 530 if (fence->mw) { 531 bnxt_re_dealloc_mw(fence->mw); 532 fence->mw = NULL; 533 } 534 if (mr) { 535 if (mr->ib_mr.rkey) 536 bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr, 537 true); 538 if (mr->ib_mr.lkey) 539 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 540 kfree(mr); 541 fence->mr = NULL; 542 } 543 if (fence->dma_addr) { 544 dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES, 545 DMA_BIDIRECTIONAL); 546 fence->dma_addr = 0; 547 } 548 } 549 550 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd) 551 { 552 int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND; 553 struct bnxt_re_fence_data *fence = &pd->fence; 554 struct bnxt_re_dev *rdev = pd->rdev; 555 struct device *dev = &rdev->en_dev->pdev->dev; 556 struct bnxt_re_mr *mr = NULL; 557 dma_addr_t dma_addr = 0; 558 struct ib_mw *mw; 559 int rc; 560 561 if (bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx)) 562 return 0; 563 564 dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES, 565 DMA_BIDIRECTIONAL); 566 rc = dma_mapping_error(dev, dma_addr); 567 if (rc) { 568 ibdev_err(&rdev->ibdev, "Failed to dma-map fence-MR-mem\n"); 569 rc = -EIO; 570 fence->dma_addr = 0; 571 goto fail; 572 } 573 fence->dma_addr = dma_addr; 574 575 /* Allocate a MR */ 576 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 577 if (!mr) { 578 rc = -ENOMEM; 579 goto fail; 580 } 581 fence->mr = mr; 582 mr->rdev = rdev; 583 mr->qplib_mr.pd = &pd->qplib_pd; 584 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR; 585 mr->qplib_mr.access_flags = __from_ib_access_flags(mr_access_flags); 586 if (!_is_alloc_mr_unified(rdev->dev_attr->dev_cap_flags)) { 587 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 588 if (rc) { 589 ibdev_err(&rdev->ibdev, "Failed to alloc fence-HW-MR\n"); 590 goto fail; 591 } 592 593 /* Register MR */ 594 mr->ib_mr.lkey = mr->qplib_mr.lkey; 595 } else { 596 mr->qplib_mr.flags = CMDQ_REGISTER_MR_FLAGS_ALLOC_MR; 597 } 598 mr->qplib_mr.va = (u64)(unsigned long)fence->va; 599 mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES; 600 rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL, 601 BNXT_RE_FENCE_PBL_SIZE, PAGE_SIZE); 602 if (rc) { 603 ibdev_err(&rdev->ibdev, "Failed to register fence-MR\n"); 604 goto fail; 605 } 606 mr->ib_mr.rkey = mr->qplib_mr.rkey; 607 608 /* Create a fence MW only for kernel consumers */ 609 mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL); 610 if (IS_ERR(mw)) { 611 ibdev_err(&rdev->ibdev, 612 "Failed to create fence-MW for PD: %p\n", pd); 613 rc = PTR_ERR(mw); 614 goto fail; 615 } 616 fence->mw = mw; 617 618 bnxt_re_create_fence_wqe(pd); 619 return 0; 620 621 fail: 622 bnxt_re_destroy_fence_mr(pd); 623 return rc; 624 } 625 626 static struct bnxt_re_user_mmap_entry* 627 bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset, 628 enum bnxt_re_mmap_flag mmap_flag, u64 *offset) 629 { 630 struct bnxt_re_user_mmap_entry *entry; 631 int ret; 632 633 entry = kzalloc(sizeof(*entry), GFP_KERNEL); 634 if (!entry) 635 return NULL; 636 637 entry->mem_offset = mem_offset; 638 entry->mmap_flag = mmap_flag; 639 entry->uctx = uctx; 640 641 switch (mmap_flag) { 642 case BNXT_RE_MMAP_SH_PAGE: 643 ret = rdma_user_mmap_entry_insert_exact(&uctx->ib_uctx, 644 &entry->rdma_entry, PAGE_SIZE, 0); 645 break; 646 case BNXT_RE_MMAP_UC_DB: 647 case BNXT_RE_MMAP_WC_DB: 648 case BNXT_RE_MMAP_DBR_BAR: 649 case BNXT_RE_MMAP_DBR_PAGE: 650 case BNXT_RE_MMAP_TOGGLE_PAGE: 651 ret = rdma_user_mmap_entry_insert(&uctx->ib_uctx, 652 &entry->rdma_entry, PAGE_SIZE); 653 break; 654 default: 655 ret = -EINVAL; 656 break; 657 } 658 659 if (ret) { 660 kfree(entry); 661 return NULL; 662 } 663 if (offset) 664 *offset = rdma_user_mmap_get_offset(&entry->rdma_entry); 665 666 return entry; 667 } 668 669 /* Protection Domains */ 670 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata) 671 { 672 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 673 struct bnxt_re_dev *rdev = pd->rdev; 674 675 if (udata) { 676 rdma_user_mmap_entry_remove(pd->pd_db_mmap); 677 pd->pd_db_mmap = NULL; 678 } 679 680 bnxt_re_destroy_fence_mr(pd); 681 682 if (pd->qplib_pd.id) { 683 if (!bnxt_qplib_dealloc_pd(&rdev->qplib_res, 684 &rdev->qplib_res.pd_tbl, 685 &pd->qplib_pd)) 686 atomic_dec(&rdev->stats.res.pd_count); 687 } 688 return 0; 689 } 690 691 int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 692 { 693 struct ib_device *ibdev = ibpd->device; 694 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 695 struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context( 696 udata, struct bnxt_re_ucontext, ib_uctx); 697 struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd); 698 struct bnxt_re_user_mmap_entry *entry = NULL; 699 u32 active_pds; 700 int rc = 0; 701 702 pd->rdev = rdev; 703 if (bnxt_qplib_alloc_pd(&rdev->qplib_res, &pd->qplib_pd)) { 704 ibdev_err(&rdev->ibdev, "Failed to allocate HW PD"); 705 rc = -ENOMEM; 706 goto fail; 707 } 708 709 if (udata) { 710 struct bnxt_re_pd_resp resp = {}; 711 712 if (!ucntx->dpi.dbr) { 713 /* Allocate DPI in alloc_pd to avoid failing of 714 * ibv_devinfo and family of application when DPIs 715 * are depleted. 716 */ 717 if (bnxt_qplib_alloc_dpi(&rdev->qplib_res, 718 &ucntx->dpi, ucntx, BNXT_QPLIB_DPI_TYPE_UC)) { 719 rc = -ENOMEM; 720 goto dbfail; 721 } 722 } 723 724 resp.pdid = pd->qplib_pd.id; 725 /* Still allow mapping this DBR to the new user PD. */ 726 resp.dpi = ucntx->dpi.dpi; 727 728 entry = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->dpi.umdbr, 729 BNXT_RE_MMAP_UC_DB, &resp.dbr); 730 731 if (!entry) { 732 rc = -ENOMEM; 733 goto dbfail; 734 } 735 736 pd->pd_db_mmap = &entry->rdma_entry; 737 738 rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen)); 739 if (rc) { 740 rdma_user_mmap_entry_remove(pd->pd_db_mmap); 741 rc = -EFAULT; 742 goto dbfail; 743 } 744 } 745 746 if (!udata) 747 if (bnxt_re_create_fence_mr(pd)) 748 ibdev_warn(&rdev->ibdev, 749 "Failed to create Fence-MR\n"); 750 active_pds = atomic_inc_return(&rdev->stats.res.pd_count); 751 if (active_pds > rdev->stats.res.pd_watermark) 752 rdev->stats.res.pd_watermark = active_pds; 753 754 return 0; 755 dbfail: 756 bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl, 757 &pd->qplib_pd); 758 fail: 759 return rc; 760 } 761 762 /* Address Handles */ 763 int bnxt_re_destroy_ah(struct ib_ah *ib_ah, u32 flags) 764 { 765 struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah); 766 struct bnxt_re_dev *rdev = ah->rdev; 767 bool block = true; 768 int rc; 769 770 block = !(flags & RDMA_DESTROY_AH_SLEEPABLE); 771 rc = bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah, block); 772 if (BNXT_RE_CHECK_RC(rc)) { 773 if (rc == -ETIMEDOUT) 774 rc = 0; 775 else 776 goto fail; 777 } 778 atomic_dec(&rdev->stats.res.ah_count); 779 fail: 780 return rc; 781 } 782 783 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype) 784 { 785 u8 nw_type; 786 787 switch (ntype) { 788 case RDMA_NETWORK_IPV4: 789 nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4; 790 break; 791 case RDMA_NETWORK_IPV6: 792 nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6; 793 break; 794 default: 795 nw_type = CMDQ_CREATE_AH_TYPE_V1; 796 break; 797 } 798 return nw_type; 799 } 800 801 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr, 802 struct ib_udata *udata) 803 { 804 struct ib_pd *ib_pd = ib_ah->pd; 805 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 806 struct rdma_ah_attr *ah_attr = init_attr->ah_attr; 807 const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr); 808 struct bnxt_re_dev *rdev = pd->rdev; 809 const struct ib_gid_attr *sgid_attr; 810 struct bnxt_re_gid_ctx *ctx; 811 struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah); 812 u32 active_ahs; 813 u8 nw_type; 814 int rc; 815 816 if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) { 817 ibdev_err(&rdev->ibdev, "Failed to alloc AH: GRH not set"); 818 return -EINVAL; 819 } 820 821 ah->rdev = rdev; 822 ah->qplib_ah.pd = &pd->qplib_pd; 823 824 /* Supply the configuration for the HW */ 825 memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw, 826 sizeof(union ib_gid)); 827 sgid_attr = grh->sgid_attr; 828 /* Get the HW context of the GID. The reference 829 * of GID table entry is already taken by the caller. 830 */ 831 ctx = rdma_read_gid_hw_context(sgid_attr); 832 ah->qplib_ah.sgid_index = ctx->idx; 833 ah->qplib_ah.host_sgid_index = grh->sgid_index; 834 ah->qplib_ah.traffic_class = grh->traffic_class; 835 ah->qplib_ah.flow_label = grh->flow_label; 836 ah->qplib_ah.hop_limit = grh->hop_limit; 837 ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr); 838 839 /* Get network header type for this GID */ 840 nw_type = rdma_gid_attr_network_type(sgid_attr); 841 ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type); 842 843 memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN); 844 rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, 845 !(init_attr->flags & 846 RDMA_CREATE_AH_SLEEPABLE)); 847 if (rc) { 848 ibdev_err(&rdev->ibdev, "Failed to allocate HW AH"); 849 return rc; 850 } 851 852 /* Write AVID to shared page. */ 853 if (udata) { 854 struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context( 855 udata, struct bnxt_re_ucontext, ib_uctx); 856 unsigned long flag; 857 u32 *wrptr; 858 859 spin_lock_irqsave(&uctx->sh_lock, flag); 860 wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT); 861 *wrptr = ah->qplib_ah.id; 862 wmb(); /* make sure cache is updated. */ 863 spin_unlock_irqrestore(&uctx->sh_lock, flag); 864 } 865 active_ahs = atomic_inc_return(&rdev->stats.res.ah_count); 866 if (active_ahs > rdev->stats.res.ah_watermark) 867 rdev->stats.res.ah_watermark = active_ahs; 868 869 return 0; 870 } 871 872 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr) 873 { 874 struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah); 875 876 ah_attr->type = ib_ah->type; 877 rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl); 878 memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN); 879 rdma_ah_set_grh(ah_attr, NULL, 0, 880 ah->qplib_ah.host_sgid_index, 881 0, ah->qplib_ah.traffic_class); 882 rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data); 883 rdma_ah_set_port_num(ah_attr, 1); 884 rdma_ah_set_static_rate(ah_attr, 0); 885 return 0; 886 } 887 888 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp) 889 __acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock) 890 { 891 unsigned long flags; 892 893 spin_lock_irqsave(&qp->scq->cq_lock, flags); 894 if (qp->rcq != qp->scq) 895 spin_lock(&qp->rcq->cq_lock); 896 else 897 __acquire(&qp->rcq->cq_lock); 898 899 return flags; 900 } 901 902 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp, 903 unsigned long flags) 904 __releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock) 905 { 906 if (qp->rcq != qp->scq) 907 spin_unlock(&qp->rcq->cq_lock); 908 else 909 __release(&qp->rcq->cq_lock); 910 spin_unlock_irqrestore(&qp->scq->cq_lock, flags); 911 } 912 913 static int bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp *qp) 914 { 915 struct bnxt_re_qp *gsi_sqp; 916 struct bnxt_re_ah *gsi_sah; 917 struct bnxt_re_dev *rdev; 918 int rc; 919 920 rdev = qp->rdev; 921 gsi_sqp = rdev->gsi_ctx.gsi_sqp; 922 gsi_sah = rdev->gsi_ctx.gsi_sah; 923 924 ibdev_dbg(&rdev->ibdev, "Destroy the shadow AH\n"); 925 bnxt_qplib_destroy_ah(&rdev->qplib_res, 926 &gsi_sah->qplib_ah, 927 true); 928 atomic_dec(&rdev->stats.res.ah_count); 929 bnxt_qplib_clean_qp(&qp->qplib_qp); 930 931 ibdev_dbg(&rdev->ibdev, "Destroy the shadow QP\n"); 932 rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &gsi_sqp->qplib_qp); 933 if (rc) { 934 ibdev_err(&rdev->ibdev, "Destroy Shadow QP failed"); 935 goto fail; 936 } 937 bnxt_qplib_free_qp_res(&rdev->qplib_res, &gsi_sqp->qplib_qp); 938 939 /* remove from active qp list */ 940 mutex_lock(&rdev->qp_lock); 941 list_del(&gsi_sqp->list); 942 mutex_unlock(&rdev->qp_lock); 943 atomic_dec(&rdev->stats.res.qp_count); 944 945 kfree(rdev->gsi_ctx.sqp_tbl); 946 kfree(gsi_sah); 947 kfree(gsi_sqp); 948 rdev->gsi_ctx.gsi_sqp = NULL; 949 rdev->gsi_ctx.gsi_sah = NULL; 950 rdev->gsi_ctx.sqp_tbl = NULL; 951 952 return 0; 953 fail: 954 return rc; 955 } 956 957 /* Queue Pairs */ 958 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata) 959 { 960 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 961 struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp; 962 struct bnxt_re_dev *rdev = qp->rdev; 963 struct bnxt_qplib_nq *scq_nq = NULL; 964 struct bnxt_qplib_nq *rcq_nq = NULL; 965 unsigned int flags; 966 int rc; 967 968 bnxt_re_debug_rem_qpinfo(rdev, qp); 969 970 bnxt_qplib_flush_cqn_wq(&qp->qplib_qp); 971 972 rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp); 973 if (rc) 974 ibdev_err(&rdev->ibdev, "Failed to destroy HW QP"); 975 976 if (rdma_is_kernel_res(&qp->ib_qp.res)) { 977 flags = bnxt_re_lock_cqs(qp); 978 bnxt_qplib_clean_qp(&qp->qplib_qp); 979 bnxt_re_unlock_cqs(qp, flags); 980 } 981 982 bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp); 983 984 if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) 985 bnxt_re_destroy_gsi_sqp(qp); 986 987 mutex_lock(&rdev->qp_lock); 988 list_del(&qp->list); 989 mutex_unlock(&rdev->qp_lock); 990 atomic_dec(&rdev->stats.res.qp_count); 991 if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_RC) 992 atomic_dec(&rdev->stats.res.rc_qp_count); 993 else if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD) 994 atomic_dec(&rdev->stats.res.ud_qp_count); 995 996 ib_umem_release(qp->rumem); 997 ib_umem_release(qp->sumem); 998 999 /* Flush all the entries of notification queue associated with 1000 * given qp. 1001 */ 1002 scq_nq = qplib_qp->scq->nq; 1003 rcq_nq = qplib_qp->rcq->nq; 1004 bnxt_re_synchronize_nq(scq_nq); 1005 if (scq_nq != rcq_nq) 1006 bnxt_re_synchronize_nq(rcq_nq); 1007 1008 return 0; 1009 } 1010 1011 static u8 __from_ib_qp_type(enum ib_qp_type type) 1012 { 1013 switch (type) { 1014 case IB_QPT_GSI: 1015 return CMDQ_CREATE_QP1_TYPE_GSI; 1016 case IB_QPT_RC: 1017 return CMDQ_CREATE_QP_TYPE_RC; 1018 case IB_QPT_UD: 1019 return CMDQ_CREATE_QP_TYPE_UD; 1020 default: 1021 return IB_QPT_MAX; 1022 } 1023 } 1024 1025 static u16 bnxt_re_setup_rwqe_size(struct bnxt_qplib_qp *qplqp, 1026 int rsge, int max) 1027 { 1028 if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) 1029 rsge = max; 1030 return bnxt_re_get_rwqe_size(rsge); 1031 } 1032 1033 static u16 bnxt_re_get_wqe_size(int ilsize, int nsge) 1034 { 1035 u16 wqe_size, calc_ils; 1036 1037 wqe_size = bnxt_re_get_swqe_size(nsge); 1038 if (ilsize) { 1039 calc_ils = sizeof(struct sq_send_hdr) + ilsize; 1040 wqe_size = max_t(u16, calc_ils, wqe_size); 1041 wqe_size = ALIGN(wqe_size, sizeof(struct sq_send_hdr)); 1042 } 1043 return wqe_size; 1044 } 1045 1046 static int bnxt_re_setup_swqe_size(struct bnxt_re_qp *qp, 1047 struct ib_qp_init_attr *init_attr) 1048 { 1049 struct bnxt_qplib_dev_attr *dev_attr; 1050 struct bnxt_qplib_qp *qplqp; 1051 struct bnxt_re_dev *rdev; 1052 struct bnxt_qplib_q *sq; 1053 int align, ilsize; 1054 1055 rdev = qp->rdev; 1056 qplqp = &qp->qplib_qp; 1057 sq = &qplqp->sq; 1058 dev_attr = rdev->dev_attr; 1059 1060 align = sizeof(struct sq_send_hdr); 1061 ilsize = ALIGN(init_attr->cap.max_inline_data, align); 1062 1063 /* For gen p4 and gen p5 fixed wqe compatibility mode 1064 * wqe size is fixed to 128 bytes - ie 6 SGEs 1065 */ 1066 if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) { 1067 sq->wqe_size = bnxt_re_get_swqe_size(BNXT_STATIC_MAX_SGE); 1068 sq->max_sge = BNXT_STATIC_MAX_SGE; 1069 } else { 1070 sq->wqe_size = bnxt_re_get_wqe_size(ilsize, sq->max_sge); 1071 if (sq->wqe_size > bnxt_re_get_swqe_size(dev_attr->max_qp_sges)) 1072 return -EINVAL; 1073 } 1074 1075 if (init_attr->cap.max_inline_data) { 1076 qplqp->max_inline_data = sq->wqe_size - 1077 sizeof(struct sq_send_hdr); 1078 init_attr->cap.max_inline_data = qplqp->max_inline_data; 1079 } 1080 1081 return 0; 1082 } 1083 1084 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd, 1085 struct bnxt_re_qp *qp, struct bnxt_re_ucontext *cntx, 1086 struct bnxt_re_qp_req *ureq) 1087 { 1088 struct bnxt_qplib_qp *qplib_qp; 1089 int bytes = 0, psn_sz; 1090 struct ib_umem *umem; 1091 int psn_nume; 1092 1093 qplib_qp = &qp->qplib_qp; 1094 1095 bytes = (qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size); 1096 /* Consider mapping PSN search memory only for RC QPs. */ 1097 if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) { 1098 psn_sz = bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx) ? 1099 sizeof(struct sq_psn_search_ext) : 1100 sizeof(struct sq_psn_search); 1101 if (cntx && bnxt_re_is_var_size_supported(rdev, cntx)) { 1102 psn_nume = ureq->sq_slots; 1103 } else { 1104 psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ? 1105 qplib_qp->sq.max_wqe : ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) / 1106 sizeof(struct bnxt_qplib_sge)); 1107 } 1108 if (_is_host_msn_table(rdev->qplib_res.dattr->dev_cap_flags2)) 1109 psn_nume = roundup_pow_of_two(psn_nume); 1110 bytes += (psn_nume * psn_sz); 1111 } 1112 1113 bytes = PAGE_ALIGN(bytes); 1114 umem = ib_umem_get(&rdev->ibdev, ureq->qpsva, bytes, 1115 IB_ACCESS_LOCAL_WRITE); 1116 if (IS_ERR(umem)) 1117 return PTR_ERR(umem); 1118 1119 qp->sumem = umem; 1120 qplib_qp->sq.sg_info.umem = umem; 1121 qplib_qp->sq.sg_info.pgsize = PAGE_SIZE; 1122 qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT; 1123 qplib_qp->qp_handle = ureq->qp_handle; 1124 1125 if (!qp->qplib_qp.srq) { 1126 bytes = (qplib_qp->rq.max_wqe * qplib_qp->rq.wqe_size); 1127 bytes = PAGE_ALIGN(bytes); 1128 umem = ib_umem_get(&rdev->ibdev, ureq->qprva, bytes, 1129 IB_ACCESS_LOCAL_WRITE); 1130 if (IS_ERR(umem)) 1131 goto rqfail; 1132 qp->rumem = umem; 1133 qplib_qp->rq.sg_info.umem = umem; 1134 qplib_qp->rq.sg_info.pgsize = PAGE_SIZE; 1135 qplib_qp->rq.sg_info.pgshft = PAGE_SHIFT; 1136 } 1137 1138 qplib_qp->dpi = &cntx->dpi; 1139 return 0; 1140 rqfail: 1141 ib_umem_release(qp->sumem); 1142 qp->sumem = NULL; 1143 memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info)); 1144 1145 return PTR_ERR(umem); 1146 } 1147 1148 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah 1149 (struct bnxt_re_pd *pd, 1150 struct bnxt_qplib_res *qp1_res, 1151 struct bnxt_qplib_qp *qp1_qp) 1152 { 1153 struct bnxt_re_dev *rdev = pd->rdev; 1154 struct bnxt_re_ah *ah; 1155 union ib_gid sgid; 1156 int rc; 1157 1158 ah = kzalloc(sizeof(*ah), GFP_KERNEL); 1159 if (!ah) 1160 return NULL; 1161 1162 ah->rdev = rdev; 1163 ah->qplib_ah.pd = &pd->qplib_pd; 1164 1165 rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid); 1166 if (rc) 1167 goto fail; 1168 1169 /* supply the dgid data same as sgid */ 1170 memcpy(ah->qplib_ah.dgid.data, &sgid.raw, 1171 sizeof(union ib_gid)); 1172 ah->qplib_ah.sgid_index = 0; 1173 1174 ah->qplib_ah.traffic_class = 0; 1175 ah->qplib_ah.flow_label = 0; 1176 ah->qplib_ah.hop_limit = 1; 1177 ah->qplib_ah.sl = 0; 1178 /* Have DMAC same as SMAC */ 1179 ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr); 1180 1181 rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false); 1182 if (rc) { 1183 ibdev_err(&rdev->ibdev, 1184 "Failed to allocate HW AH for Shadow QP"); 1185 goto fail; 1186 } 1187 atomic_inc(&rdev->stats.res.ah_count); 1188 1189 return ah; 1190 1191 fail: 1192 kfree(ah); 1193 return NULL; 1194 } 1195 1196 static struct bnxt_re_qp *bnxt_re_create_shadow_qp 1197 (struct bnxt_re_pd *pd, 1198 struct bnxt_qplib_res *qp1_res, 1199 struct bnxt_qplib_qp *qp1_qp) 1200 { 1201 struct bnxt_re_dev *rdev = pd->rdev; 1202 struct bnxt_re_qp *qp; 1203 int rc; 1204 1205 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 1206 if (!qp) 1207 return NULL; 1208 1209 qp->rdev = rdev; 1210 1211 /* Initialize the shadow QP structure from the QP1 values */ 1212 ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr); 1213 1214 qp->qplib_qp.pd = &pd->qplib_pd; 1215 qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp); 1216 qp->qplib_qp.type = IB_QPT_UD; 1217 1218 qp->qplib_qp.max_inline_data = 0; 1219 qp->qplib_qp.sig_type = true; 1220 1221 /* Shadow QP SQ depth should be same as QP1 RQ depth */ 1222 qp->qplib_qp.sq.wqe_size = bnxt_re_get_wqe_size(0, 6); 1223 qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe; 1224 qp->qplib_qp.sq.max_sw_wqe = qp1_qp->rq.max_wqe; 1225 qp->qplib_qp.sq.max_sge = 2; 1226 /* Q full delta can be 1 since it is internal QP */ 1227 qp->qplib_qp.sq.q_full_delta = 1; 1228 qp->qplib_qp.sq.sg_info.pgsize = PAGE_SIZE; 1229 qp->qplib_qp.sq.sg_info.pgshft = PAGE_SHIFT; 1230 1231 qp->qplib_qp.scq = qp1_qp->scq; 1232 qp->qplib_qp.rcq = qp1_qp->rcq; 1233 1234 qp->qplib_qp.rq.wqe_size = bnxt_re_get_rwqe_size(6); 1235 qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe; 1236 qp->qplib_qp.rq.max_sw_wqe = qp1_qp->rq.max_wqe; 1237 qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge; 1238 /* Q full delta can be 1 since it is internal QP */ 1239 qp->qplib_qp.rq.q_full_delta = 1; 1240 qp->qplib_qp.rq.sg_info.pgsize = PAGE_SIZE; 1241 qp->qplib_qp.rq.sg_info.pgshft = PAGE_SHIFT; 1242 1243 qp->qplib_qp.mtu = qp1_qp->mtu; 1244 1245 qp->qplib_qp.sq_hdr_buf_size = 0; 1246 qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6; 1247 qp->qplib_qp.dpi = &rdev->dpi_privileged; 1248 1249 rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp); 1250 if (rc) 1251 goto fail; 1252 1253 spin_lock_init(&qp->sq_lock); 1254 INIT_LIST_HEAD(&qp->list); 1255 mutex_lock(&rdev->qp_lock); 1256 list_add_tail(&qp->list, &rdev->qp_list); 1257 atomic_inc(&rdev->stats.res.qp_count); 1258 mutex_unlock(&rdev->qp_lock); 1259 return qp; 1260 fail: 1261 kfree(qp); 1262 return NULL; 1263 } 1264 1265 static int bnxt_re_init_rq_attr(struct bnxt_re_qp *qp, 1266 struct ib_qp_init_attr *init_attr, 1267 struct bnxt_re_ucontext *uctx) 1268 { 1269 struct bnxt_qplib_dev_attr *dev_attr; 1270 struct bnxt_qplib_qp *qplqp; 1271 struct bnxt_re_dev *rdev; 1272 struct bnxt_qplib_q *rq; 1273 int entries; 1274 1275 rdev = qp->rdev; 1276 qplqp = &qp->qplib_qp; 1277 rq = &qplqp->rq; 1278 dev_attr = rdev->dev_attr; 1279 1280 if (init_attr->srq) { 1281 struct bnxt_re_srq *srq; 1282 1283 srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq); 1284 qplqp->srq = &srq->qplib_srq; 1285 rq->max_wqe = 0; 1286 } else { 1287 rq->max_sge = init_attr->cap.max_recv_sge; 1288 if (rq->max_sge > dev_attr->max_qp_sges) 1289 rq->max_sge = dev_attr->max_qp_sges; 1290 init_attr->cap.max_recv_sge = rq->max_sge; 1291 rq->wqe_size = bnxt_re_setup_rwqe_size(qplqp, rq->max_sge, 1292 dev_attr->max_qp_sges); 1293 /* Allocate 1 more than what's provided so posting max doesn't 1294 * mean empty. 1295 */ 1296 entries = bnxt_re_init_depth(init_attr->cap.max_recv_wr + 1, uctx); 1297 rq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + 1); 1298 rq->max_sw_wqe = rq->max_wqe; 1299 rq->q_full_delta = 0; 1300 rq->sg_info.pgsize = PAGE_SIZE; 1301 rq->sg_info.pgshft = PAGE_SHIFT; 1302 } 1303 1304 return 0; 1305 } 1306 1307 static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp) 1308 { 1309 struct bnxt_qplib_dev_attr *dev_attr; 1310 struct bnxt_qplib_qp *qplqp; 1311 struct bnxt_re_dev *rdev; 1312 1313 rdev = qp->rdev; 1314 qplqp = &qp->qplib_qp; 1315 dev_attr = rdev->dev_attr; 1316 1317 if (!bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx)) { 1318 qplqp->rq.max_sge = dev_attr->max_qp_sges; 1319 if (qplqp->rq.max_sge > dev_attr->max_qp_sges) 1320 qplqp->rq.max_sge = dev_attr->max_qp_sges; 1321 qplqp->rq.max_sge = 6; 1322 } 1323 } 1324 1325 static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp, 1326 struct ib_qp_init_attr *init_attr, 1327 struct bnxt_re_ucontext *uctx, 1328 struct bnxt_re_qp_req *ureq) 1329 { 1330 struct bnxt_qplib_dev_attr *dev_attr; 1331 struct bnxt_qplib_qp *qplqp; 1332 struct bnxt_re_dev *rdev; 1333 struct bnxt_qplib_q *sq; 1334 int diff = 0; 1335 int entries; 1336 int rc; 1337 1338 rdev = qp->rdev; 1339 qplqp = &qp->qplib_qp; 1340 sq = &qplqp->sq; 1341 dev_attr = rdev->dev_attr; 1342 1343 sq->max_sge = init_attr->cap.max_send_sge; 1344 entries = init_attr->cap.max_send_wr; 1345 if (uctx && qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) { 1346 sq->max_wqe = ureq->sq_slots; 1347 sq->max_sw_wqe = ureq->sq_slots; 1348 sq->wqe_size = sizeof(struct sq_sge); 1349 } else { 1350 if (sq->max_sge > dev_attr->max_qp_sges) { 1351 sq->max_sge = dev_attr->max_qp_sges; 1352 init_attr->cap.max_send_sge = sq->max_sge; 1353 } 1354 1355 rc = bnxt_re_setup_swqe_size(qp, init_attr); 1356 if (rc) 1357 return rc; 1358 1359 /* Allocate 128 + 1 more than what's provided */ 1360 diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ? 1361 0 : BNXT_QPLIB_RESERVED_QP_WRS; 1362 entries = bnxt_re_init_depth(entries + diff + 1, uctx); 1363 sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1); 1364 if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) 1365 sq->max_sw_wqe = bnxt_qplib_get_depth(sq, qplqp->wqe_mode, true); 1366 else 1367 sq->max_sw_wqe = sq->max_wqe; 1368 1369 } 1370 sq->q_full_delta = diff + 1; 1371 /* 1372 * Reserving one slot for Phantom WQE. Application can 1373 * post one extra entry in this case. But allowing this to avoid 1374 * unexpected Queue full condition 1375 */ 1376 qplqp->sq.q_full_delta -= 1; 1377 qplqp->sq.sg_info.pgsize = PAGE_SIZE; 1378 qplqp->sq.sg_info.pgshft = PAGE_SHIFT; 1379 1380 return 0; 1381 } 1382 1383 static void bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp *qp, 1384 struct ib_qp_init_attr *init_attr, 1385 struct bnxt_re_ucontext *uctx) 1386 { 1387 struct bnxt_qplib_dev_attr *dev_attr; 1388 struct bnxt_qplib_qp *qplqp; 1389 struct bnxt_re_dev *rdev; 1390 int entries; 1391 1392 rdev = qp->rdev; 1393 qplqp = &qp->qplib_qp; 1394 dev_attr = rdev->dev_attr; 1395 1396 if (!bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx)) { 1397 entries = bnxt_re_init_depth(init_attr->cap.max_send_wr + 1, uctx); 1398 qplqp->sq.max_wqe = min_t(u32, entries, 1399 dev_attr->max_qp_wqes + 1); 1400 qplqp->sq.q_full_delta = qplqp->sq.max_wqe - 1401 init_attr->cap.max_send_wr; 1402 qplqp->sq.max_sge++; /* Need one extra sge to put UD header */ 1403 if (qplqp->sq.max_sge > dev_attr->max_qp_sges) 1404 qplqp->sq.max_sge = dev_attr->max_qp_sges; 1405 } 1406 } 1407 1408 static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev, 1409 struct ib_qp_init_attr *init_attr) 1410 { 1411 struct bnxt_qplib_chip_ctx *chip_ctx; 1412 int qptype; 1413 1414 chip_ctx = rdev->chip_ctx; 1415 1416 qptype = __from_ib_qp_type(init_attr->qp_type); 1417 if (qptype == IB_QPT_MAX) { 1418 ibdev_err(&rdev->ibdev, "QP type 0x%x not supported", qptype); 1419 qptype = -EOPNOTSUPP; 1420 goto out; 1421 } 1422 1423 if (bnxt_qplib_is_chip_gen_p5_p7(chip_ctx) && 1424 init_attr->qp_type == IB_QPT_GSI) 1425 qptype = CMDQ_CREATE_QP_TYPE_GSI; 1426 out: 1427 return qptype; 1428 } 1429 1430 static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, 1431 struct ib_qp_init_attr *init_attr, 1432 struct bnxt_re_ucontext *uctx, 1433 struct bnxt_re_qp_req *ureq) 1434 { 1435 struct bnxt_qplib_dev_attr *dev_attr; 1436 struct bnxt_qplib_qp *qplqp; 1437 struct bnxt_re_dev *rdev; 1438 struct bnxt_re_cq *cq; 1439 int rc = 0, qptype; 1440 1441 rdev = qp->rdev; 1442 qplqp = &qp->qplib_qp; 1443 dev_attr = rdev->dev_attr; 1444 1445 /* Setup misc params */ 1446 ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr); 1447 qplqp->pd = &pd->qplib_pd; 1448 qplqp->qp_handle = (u64)qplqp; 1449 qplqp->max_inline_data = init_attr->cap.max_inline_data; 1450 qplqp->sig_type = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR; 1451 qptype = bnxt_re_init_qp_type(rdev, init_attr); 1452 if (qptype < 0) { 1453 rc = qptype; 1454 goto out; 1455 } 1456 qplqp->type = (u8)qptype; 1457 qplqp->wqe_mode = bnxt_re_is_var_size_supported(rdev, uctx); 1458 if (init_attr->qp_type == IB_QPT_RC) { 1459 qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom; 1460 qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom; 1461 } 1462 qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu)); 1463 qplqp->dpi = &rdev->dpi_privileged; /* Doorbell page */ 1464 if (init_attr->create_flags) { 1465 ibdev_dbg(&rdev->ibdev, 1466 "QP create flags 0x%x not supported", 1467 init_attr->create_flags); 1468 return -EOPNOTSUPP; 1469 } 1470 1471 /* Setup CQs */ 1472 if (init_attr->send_cq) { 1473 cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq); 1474 qplqp->scq = &cq->qplib_cq; 1475 qp->scq = cq; 1476 } 1477 1478 if (init_attr->recv_cq) { 1479 cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq); 1480 qplqp->rcq = &cq->qplib_cq; 1481 qp->rcq = cq; 1482 } 1483 1484 /* Setup RQ/SRQ */ 1485 rc = bnxt_re_init_rq_attr(qp, init_attr, uctx); 1486 if (rc) 1487 goto out; 1488 if (init_attr->qp_type == IB_QPT_GSI) 1489 bnxt_re_adjust_gsi_rq_attr(qp); 1490 1491 /* Setup SQ */ 1492 rc = bnxt_re_init_sq_attr(qp, init_attr, uctx, ureq); 1493 if (rc) 1494 goto out; 1495 if (init_attr->qp_type == IB_QPT_GSI) 1496 bnxt_re_adjust_gsi_sq_attr(qp, init_attr, uctx); 1497 1498 if (uctx) /* This will update DPI and qp_handle */ 1499 rc = bnxt_re_init_user_qp(rdev, pd, qp, uctx, ureq); 1500 out: 1501 return rc; 1502 } 1503 1504 static int bnxt_re_create_shadow_gsi(struct bnxt_re_qp *qp, 1505 struct bnxt_re_pd *pd) 1506 { 1507 struct bnxt_re_sqp_entries *sqp_tbl; 1508 struct bnxt_re_dev *rdev; 1509 struct bnxt_re_qp *sqp; 1510 struct bnxt_re_ah *sah; 1511 int rc = 0; 1512 1513 rdev = qp->rdev; 1514 /* Create a shadow QP to handle the QP1 traffic */ 1515 sqp_tbl = kcalloc(BNXT_RE_MAX_GSI_SQP_ENTRIES, sizeof(*sqp_tbl), 1516 GFP_KERNEL); 1517 if (!sqp_tbl) 1518 return -ENOMEM; 1519 rdev->gsi_ctx.sqp_tbl = sqp_tbl; 1520 1521 sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res, &qp->qplib_qp); 1522 if (!sqp) { 1523 rc = -ENODEV; 1524 ibdev_err(&rdev->ibdev, "Failed to create Shadow QP for QP1"); 1525 goto out; 1526 } 1527 rdev->gsi_ctx.gsi_sqp = sqp; 1528 1529 sqp->rcq = qp->rcq; 1530 sqp->scq = qp->scq; 1531 sah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res, 1532 &qp->qplib_qp); 1533 if (!sah) { 1534 bnxt_qplib_destroy_qp(&rdev->qplib_res, 1535 &sqp->qplib_qp); 1536 rc = -ENODEV; 1537 ibdev_err(&rdev->ibdev, 1538 "Failed to create AH entry for ShadowQP"); 1539 goto out; 1540 } 1541 rdev->gsi_ctx.gsi_sah = sah; 1542 1543 return 0; 1544 out: 1545 kfree(sqp_tbl); 1546 return rc; 1547 } 1548 1549 static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, 1550 struct ib_qp_init_attr *init_attr) 1551 { 1552 struct bnxt_re_dev *rdev; 1553 struct bnxt_qplib_qp *qplqp; 1554 int rc; 1555 1556 rdev = qp->rdev; 1557 qplqp = &qp->qplib_qp; 1558 1559 qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2; 1560 qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2; 1561 1562 rc = bnxt_qplib_create_qp1(&rdev->qplib_res, qplqp); 1563 if (rc) { 1564 ibdev_err(&rdev->ibdev, "create HW QP1 failed!"); 1565 goto out; 1566 } 1567 1568 rc = bnxt_re_create_shadow_gsi(qp, pd); 1569 out: 1570 return rc; 1571 } 1572 1573 static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev, 1574 struct ib_qp_init_attr *init_attr, 1575 struct bnxt_qplib_dev_attr *dev_attr) 1576 { 1577 bool rc = true; 1578 1579 if (init_attr->cap.max_send_wr > dev_attr->max_qp_wqes || 1580 init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes || 1581 init_attr->cap.max_send_sge > dev_attr->max_qp_sges || 1582 init_attr->cap.max_recv_sge > dev_attr->max_qp_sges || 1583 init_attr->cap.max_inline_data > dev_attr->max_inline_data) { 1584 ibdev_err(&rdev->ibdev, 1585 "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", 1586 init_attr->cap.max_send_wr, dev_attr->max_qp_wqes, 1587 init_attr->cap.max_recv_wr, dev_attr->max_qp_wqes, 1588 init_attr->cap.max_send_sge, dev_attr->max_qp_sges, 1589 init_attr->cap.max_recv_sge, dev_attr->max_qp_sges, 1590 init_attr->cap.max_inline_data, 1591 dev_attr->max_inline_data); 1592 rc = false; 1593 } 1594 return rc; 1595 } 1596 1597 int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr, 1598 struct ib_udata *udata) 1599 { 1600 struct bnxt_qplib_dev_attr *dev_attr; 1601 struct bnxt_re_ucontext *uctx; 1602 struct bnxt_re_qp_req ureq; 1603 struct bnxt_re_dev *rdev; 1604 struct bnxt_re_pd *pd; 1605 struct bnxt_re_qp *qp; 1606 struct ib_pd *ib_pd; 1607 u32 active_qps; 1608 int rc; 1609 1610 ib_pd = ib_qp->pd; 1611 pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 1612 rdev = pd->rdev; 1613 dev_attr = rdev->dev_attr; 1614 qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 1615 1616 uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx); 1617 if (udata) 1618 if (ib_copy_from_udata(&ureq, udata, min(udata->inlen, sizeof(ureq)))) 1619 return -EFAULT; 1620 1621 rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr); 1622 if (!rc) { 1623 rc = -EINVAL; 1624 goto fail; 1625 } 1626 1627 qp->rdev = rdev; 1628 rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, uctx, &ureq); 1629 if (rc) 1630 goto fail; 1631 1632 if (qp_init_attr->qp_type == IB_QPT_GSI && 1633 !(bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx))) { 1634 rc = bnxt_re_create_gsi_qp(qp, pd, qp_init_attr); 1635 if (rc == -ENODEV) 1636 goto qp_destroy; 1637 if (rc) 1638 goto fail; 1639 } else { 1640 rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp); 1641 if (rc) { 1642 ibdev_err(&rdev->ibdev, "Failed to create HW QP"); 1643 goto free_umem; 1644 } 1645 if (udata) { 1646 struct bnxt_re_qp_resp resp; 1647 1648 resp.qpid = qp->qplib_qp.id; 1649 resp.rsvd = 0; 1650 rc = ib_copy_to_udata(udata, &resp, sizeof(resp)); 1651 if (rc) { 1652 ibdev_err(&rdev->ibdev, "Failed to copy QP udata"); 1653 goto qp_destroy; 1654 } 1655 } 1656 } 1657 1658 qp->ib_qp.qp_num = qp->qplib_qp.id; 1659 if (qp_init_attr->qp_type == IB_QPT_GSI) 1660 rdev->gsi_ctx.gsi_qp = qp; 1661 spin_lock_init(&qp->sq_lock); 1662 spin_lock_init(&qp->rq_lock); 1663 INIT_LIST_HEAD(&qp->list); 1664 mutex_lock(&rdev->qp_lock); 1665 list_add_tail(&qp->list, &rdev->qp_list); 1666 mutex_unlock(&rdev->qp_lock); 1667 active_qps = atomic_inc_return(&rdev->stats.res.qp_count); 1668 if (active_qps > rdev->stats.res.qp_watermark) 1669 rdev->stats.res.qp_watermark = active_qps; 1670 if (qp_init_attr->qp_type == IB_QPT_RC) { 1671 active_qps = atomic_inc_return(&rdev->stats.res.rc_qp_count); 1672 if (active_qps > rdev->stats.res.rc_qp_watermark) 1673 rdev->stats.res.rc_qp_watermark = active_qps; 1674 } else if (qp_init_attr->qp_type == IB_QPT_UD) { 1675 active_qps = atomic_inc_return(&rdev->stats.res.ud_qp_count); 1676 if (active_qps > rdev->stats.res.ud_qp_watermark) 1677 rdev->stats.res.ud_qp_watermark = active_qps; 1678 } 1679 bnxt_re_debug_add_qpinfo(rdev, qp); 1680 1681 return 0; 1682 qp_destroy: 1683 bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp); 1684 free_umem: 1685 ib_umem_release(qp->rumem); 1686 ib_umem_release(qp->sumem); 1687 fail: 1688 return rc; 1689 } 1690 1691 static u8 __from_ib_qp_state(enum ib_qp_state state) 1692 { 1693 switch (state) { 1694 case IB_QPS_RESET: 1695 return CMDQ_MODIFY_QP_NEW_STATE_RESET; 1696 case IB_QPS_INIT: 1697 return CMDQ_MODIFY_QP_NEW_STATE_INIT; 1698 case IB_QPS_RTR: 1699 return CMDQ_MODIFY_QP_NEW_STATE_RTR; 1700 case IB_QPS_RTS: 1701 return CMDQ_MODIFY_QP_NEW_STATE_RTS; 1702 case IB_QPS_SQD: 1703 return CMDQ_MODIFY_QP_NEW_STATE_SQD; 1704 case IB_QPS_SQE: 1705 return CMDQ_MODIFY_QP_NEW_STATE_SQE; 1706 case IB_QPS_ERR: 1707 default: 1708 return CMDQ_MODIFY_QP_NEW_STATE_ERR; 1709 } 1710 } 1711 1712 static enum ib_qp_state __to_ib_qp_state(u8 state) 1713 { 1714 switch (state) { 1715 case CMDQ_MODIFY_QP_NEW_STATE_RESET: 1716 return IB_QPS_RESET; 1717 case CMDQ_MODIFY_QP_NEW_STATE_INIT: 1718 return IB_QPS_INIT; 1719 case CMDQ_MODIFY_QP_NEW_STATE_RTR: 1720 return IB_QPS_RTR; 1721 case CMDQ_MODIFY_QP_NEW_STATE_RTS: 1722 return IB_QPS_RTS; 1723 case CMDQ_MODIFY_QP_NEW_STATE_SQD: 1724 return IB_QPS_SQD; 1725 case CMDQ_MODIFY_QP_NEW_STATE_SQE: 1726 return IB_QPS_SQE; 1727 case CMDQ_MODIFY_QP_NEW_STATE_ERR: 1728 default: 1729 return IB_QPS_ERR; 1730 } 1731 } 1732 1733 static u32 __from_ib_mtu(enum ib_mtu mtu) 1734 { 1735 switch (mtu) { 1736 case IB_MTU_256: 1737 return CMDQ_MODIFY_QP_PATH_MTU_MTU_256; 1738 case IB_MTU_512: 1739 return CMDQ_MODIFY_QP_PATH_MTU_MTU_512; 1740 case IB_MTU_1024: 1741 return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024; 1742 case IB_MTU_2048: 1743 return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048; 1744 case IB_MTU_4096: 1745 return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096; 1746 default: 1747 return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048; 1748 } 1749 } 1750 1751 static enum ib_mtu __to_ib_mtu(u32 mtu) 1752 { 1753 switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) { 1754 case CMDQ_MODIFY_QP_PATH_MTU_MTU_256: 1755 return IB_MTU_256; 1756 case CMDQ_MODIFY_QP_PATH_MTU_MTU_512: 1757 return IB_MTU_512; 1758 case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024: 1759 return IB_MTU_1024; 1760 case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048: 1761 return IB_MTU_2048; 1762 case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096: 1763 return IB_MTU_4096; 1764 default: 1765 return IB_MTU_2048; 1766 } 1767 } 1768 1769 /* Shared Receive Queues */ 1770 int bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata) 1771 { 1772 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1773 ib_srq); 1774 struct bnxt_re_dev *rdev = srq->rdev; 1775 struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq; 1776 struct bnxt_qplib_nq *nq = NULL; 1777 1778 if (qplib_srq->cq) 1779 nq = qplib_srq->cq->nq; 1780 if (rdev->chip_ctx->modes.toggle_bits & BNXT_QPLIB_SRQ_TOGGLE_BIT) { 1781 free_page((unsigned long)srq->uctx_srq_page); 1782 hash_del(&srq->hash_entry); 1783 } 1784 bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq); 1785 ib_umem_release(srq->umem); 1786 atomic_dec(&rdev->stats.res.srq_count); 1787 if (nq) 1788 nq->budget--; 1789 return 0; 1790 } 1791 1792 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev, 1793 struct bnxt_re_pd *pd, 1794 struct bnxt_re_srq *srq, 1795 struct ib_udata *udata) 1796 { 1797 struct bnxt_re_srq_req ureq; 1798 struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq; 1799 struct ib_umem *umem; 1800 int bytes = 0; 1801 struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context( 1802 udata, struct bnxt_re_ucontext, ib_uctx); 1803 1804 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) 1805 return -EFAULT; 1806 1807 bytes = (qplib_srq->max_wqe * qplib_srq->wqe_size); 1808 bytes = PAGE_ALIGN(bytes); 1809 umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes, 1810 IB_ACCESS_LOCAL_WRITE); 1811 if (IS_ERR(umem)) 1812 return PTR_ERR(umem); 1813 1814 srq->umem = umem; 1815 qplib_srq->sg_info.umem = umem; 1816 qplib_srq->sg_info.pgsize = PAGE_SIZE; 1817 qplib_srq->sg_info.pgshft = PAGE_SHIFT; 1818 qplib_srq->srq_handle = ureq.srq_handle; 1819 qplib_srq->dpi = &cntx->dpi; 1820 1821 return 0; 1822 } 1823 1824 int bnxt_re_create_srq(struct ib_srq *ib_srq, 1825 struct ib_srq_init_attr *srq_init_attr, 1826 struct ib_udata *udata) 1827 { 1828 struct bnxt_qplib_dev_attr *dev_attr; 1829 struct bnxt_qplib_nq *nq = NULL; 1830 struct bnxt_re_ucontext *uctx; 1831 struct bnxt_re_dev *rdev; 1832 struct bnxt_re_srq *srq; 1833 struct bnxt_re_pd *pd; 1834 struct ib_pd *ib_pd; 1835 u32 active_srqs; 1836 int rc, entries; 1837 1838 ib_pd = ib_srq->pd; 1839 pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 1840 rdev = pd->rdev; 1841 dev_attr = rdev->dev_attr; 1842 srq = container_of(ib_srq, struct bnxt_re_srq, ib_srq); 1843 1844 if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) { 1845 ibdev_err(&rdev->ibdev, "Create CQ failed - max exceeded"); 1846 rc = -EINVAL; 1847 goto exit; 1848 } 1849 1850 if (srq_init_attr->srq_type != IB_SRQT_BASIC) { 1851 rc = -EOPNOTSUPP; 1852 goto exit; 1853 } 1854 1855 uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx); 1856 srq->rdev = rdev; 1857 srq->qplib_srq.pd = &pd->qplib_pd; 1858 srq->qplib_srq.dpi = &rdev->dpi_privileged; 1859 /* Allocate 1 more than what's provided so posting max doesn't 1860 * mean empty 1861 */ 1862 entries = bnxt_re_init_depth(srq_init_attr->attr.max_wr + 1, uctx); 1863 if (entries > dev_attr->max_srq_wqes + 1) 1864 entries = dev_attr->max_srq_wqes + 1; 1865 srq->qplib_srq.max_wqe = entries; 1866 1867 srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge; 1868 /* 128 byte wqe size for SRQ . So use max sges */ 1869 srq->qplib_srq.wqe_size = bnxt_re_get_rwqe_size(dev_attr->max_srq_sges); 1870 srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit; 1871 srq->srq_limit = srq_init_attr->attr.srq_limit; 1872 srq->qplib_srq.eventq_hw_ring_id = rdev->nqr->nq[0].ring_id; 1873 srq->qplib_srq.sg_info.pgsize = PAGE_SIZE; 1874 srq->qplib_srq.sg_info.pgshft = PAGE_SHIFT; 1875 nq = &rdev->nqr->nq[0]; 1876 1877 if (udata) { 1878 rc = bnxt_re_init_user_srq(rdev, pd, srq, udata); 1879 if (rc) 1880 goto fail; 1881 } 1882 1883 rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq); 1884 if (rc) { 1885 ibdev_err(&rdev->ibdev, "Create HW SRQ failed!"); 1886 goto fail; 1887 } 1888 1889 if (udata) { 1890 struct bnxt_re_srq_resp resp = {}; 1891 1892 resp.srqid = srq->qplib_srq.id; 1893 if (rdev->chip_ctx->modes.toggle_bits & BNXT_QPLIB_SRQ_TOGGLE_BIT) { 1894 hash_add(rdev->srq_hash, &srq->hash_entry, srq->qplib_srq.id); 1895 srq->uctx_srq_page = (void *)get_zeroed_page(GFP_KERNEL); 1896 if (!srq->uctx_srq_page) { 1897 rc = -ENOMEM; 1898 goto fail; 1899 } 1900 resp.comp_mask |= BNXT_RE_SRQ_TOGGLE_PAGE_SUPPORT; 1901 } 1902 rc = ib_copy_to_udata(udata, &resp, sizeof(resp)); 1903 if (rc) { 1904 ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!"); 1905 bnxt_qplib_destroy_srq(&rdev->qplib_res, 1906 &srq->qplib_srq); 1907 goto fail; 1908 } 1909 } 1910 if (nq) 1911 nq->budget++; 1912 active_srqs = atomic_inc_return(&rdev->stats.res.srq_count); 1913 if (active_srqs > rdev->stats.res.srq_watermark) 1914 rdev->stats.res.srq_watermark = active_srqs; 1915 spin_lock_init(&srq->lock); 1916 1917 return 0; 1918 1919 fail: 1920 ib_umem_release(srq->umem); 1921 exit: 1922 return rc; 1923 } 1924 1925 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr, 1926 enum ib_srq_attr_mask srq_attr_mask, 1927 struct ib_udata *udata) 1928 { 1929 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1930 ib_srq); 1931 struct bnxt_re_dev *rdev = srq->rdev; 1932 int rc; 1933 1934 switch (srq_attr_mask) { 1935 case IB_SRQ_MAX_WR: 1936 /* SRQ resize is not supported */ 1937 return -EINVAL; 1938 case IB_SRQ_LIMIT: 1939 /* Change the SRQ threshold */ 1940 if (srq_attr->srq_limit > srq->qplib_srq.max_wqe) 1941 return -EINVAL; 1942 1943 srq->qplib_srq.threshold = srq_attr->srq_limit; 1944 rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq); 1945 if (rc) { 1946 ibdev_err(&rdev->ibdev, "Modify HW SRQ failed!"); 1947 return rc; 1948 } 1949 /* On success, update the shadow */ 1950 srq->srq_limit = srq_attr->srq_limit; 1951 /* No need to Build and send response back to udata */ 1952 return 0; 1953 default: 1954 ibdev_err(&rdev->ibdev, 1955 "Unsupported srq_attr_mask 0x%x", srq_attr_mask); 1956 return -EINVAL; 1957 } 1958 } 1959 1960 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr) 1961 { 1962 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1963 ib_srq); 1964 struct bnxt_re_srq tsrq; 1965 struct bnxt_re_dev *rdev = srq->rdev; 1966 int rc; 1967 1968 /* Get live SRQ attr */ 1969 tsrq.qplib_srq.id = srq->qplib_srq.id; 1970 rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq); 1971 if (rc) { 1972 ibdev_err(&rdev->ibdev, "Query HW SRQ failed!"); 1973 return rc; 1974 } 1975 srq_attr->max_wr = srq->qplib_srq.max_wqe; 1976 srq_attr->max_sge = srq->qplib_srq.max_sge; 1977 srq_attr->srq_limit = tsrq.qplib_srq.threshold; 1978 1979 return 0; 1980 } 1981 1982 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr, 1983 const struct ib_recv_wr **bad_wr) 1984 { 1985 struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq, 1986 ib_srq); 1987 struct bnxt_qplib_swqe wqe; 1988 unsigned long flags; 1989 int rc = 0; 1990 1991 spin_lock_irqsave(&srq->lock, flags); 1992 while (wr) { 1993 /* Transcribe each ib_recv_wr to qplib_swqe */ 1994 wqe.num_sge = wr->num_sge; 1995 bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge); 1996 wqe.wr_id = wr->wr_id; 1997 wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV; 1998 1999 rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe); 2000 if (rc) { 2001 *bad_wr = wr; 2002 break; 2003 } 2004 wr = wr->next; 2005 } 2006 spin_unlock_irqrestore(&srq->lock, flags); 2007 2008 return rc; 2009 } 2010 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev, 2011 struct bnxt_re_qp *qp1_qp, 2012 int qp_attr_mask) 2013 { 2014 struct bnxt_re_qp *qp = rdev->gsi_ctx.gsi_sqp; 2015 int rc; 2016 2017 if (qp_attr_mask & IB_QP_STATE) { 2018 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE; 2019 qp->qplib_qp.state = qp1_qp->qplib_qp.state; 2020 } 2021 if (qp_attr_mask & IB_QP_PKEY_INDEX) { 2022 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY; 2023 qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index; 2024 } 2025 2026 if (qp_attr_mask & IB_QP_QKEY) { 2027 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY; 2028 /* Using a Random QKEY */ 2029 qp->qplib_qp.qkey = 0x81818181; 2030 } 2031 if (qp_attr_mask & IB_QP_SQ_PSN) { 2032 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN; 2033 qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn; 2034 } 2035 2036 rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp); 2037 if (rc) 2038 ibdev_err(&rdev->ibdev, "Failed to modify Shadow QP for QP1"); 2039 return rc; 2040 } 2041 2042 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr, 2043 int qp_attr_mask, struct ib_udata *udata) 2044 { 2045 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 2046 struct bnxt_re_dev *rdev = qp->rdev; 2047 struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr; 2048 enum ib_qp_state curr_qp_state, new_qp_state; 2049 int rc, entries; 2050 unsigned int flags; 2051 u8 nw_type; 2052 2053 if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS) 2054 return -EOPNOTSUPP; 2055 2056 qp->qplib_qp.modify_flags = 0; 2057 if (qp_attr_mask & IB_QP_STATE) { 2058 curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state); 2059 new_qp_state = qp_attr->qp_state; 2060 if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state, 2061 ib_qp->qp_type, qp_attr_mask)) { 2062 ibdev_err(&rdev->ibdev, 2063 "Invalid attribute mask: %#x specified ", 2064 qp_attr_mask); 2065 ibdev_err(&rdev->ibdev, 2066 "for qpn: %#x type: %#x", 2067 ib_qp->qp_num, ib_qp->qp_type); 2068 ibdev_err(&rdev->ibdev, 2069 "curr_qp_state=0x%x, new_qp_state=0x%x\n", 2070 curr_qp_state, new_qp_state); 2071 return -EINVAL; 2072 } 2073 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE; 2074 qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state); 2075 2076 if (!qp->sumem && 2077 qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) { 2078 ibdev_dbg(&rdev->ibdev, 2079 "Move QP = %p to flush list\n", qp); 2080 flags = bnxt_re_lock_cqs(qp); 2081 bnxt_qplib_add_flush_qp(&qp->qplib_qp); 2082 bnxt_re_unlock_cqs(qp, flags); 2083 } 2084 if (!qp->sumem && 2085 qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) { 2086 ibdev_dbg(&rdev->ibdev, 2087 "Move QP = %p out of flush list\n", qp); 2088 flags = bnxt_re_lock_cqs(qp); 2089 bnxt_qplib_clean_qp(&qp->qplib_qp); 2090 bnxt_re_unlock_cqs(qp, flags); 2091 } 2092 } 2093 if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) { 2094 qp->qplib_qp.modify_flags |= 2095 CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY; 2096 qp->qplib_qp.en_sqd_async_notify = true; 2097 } 2098 if (qp_attr_mask & IB_QP_ACCESS_FLAGS) { 2099 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS; 2100 qp->qplib_qp.access = 2101 __qp_access_flags_from_ib(qp->qplib_qp.cctx, 2102 qp_attr->qp_access_flags); 2103 /* LOCAL_WRITE access must be set to allow RC receive */ 2104 qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_LOCAL_WRITE; 2105 } 2106 if (qp_attr_mask & IB_QP_PKEY_INDEX) { 2107 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY; 2108 qp->qplib_qp.pkey_index = qp_attr->pkey_index; 2109 } 2110 if (qp_attr_mask & IB_QP_QKEY) { 2111 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY; 2112 qp->qplib_qp.qkey = qp_attr->qkey; 2113 } 2114 if (qp_attr_mask & IB_QP_AV) { 2115 const struct ib_global_route *grh = 2116 rdma_ah_read_grh(&qp_attr->ah_attr); 2117 const struct ib_gid_attr *sgid_attr; 2118 struct bnxt_re_gid_ctx *ctx; 2119 2120 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID | 2121 CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL | 2122 CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX | 2123 CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT | 2124 CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS | 2125 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC | 2126 CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID; 2127 memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw, 2128 sizeof(qp->qplib_qp.ah.dgid.data)); 2129 qp->qplib_qp.ah.flow_label = grh->flow_label; 2130 sgid_attr = grh->sgid_attr; 2131 /* Get the HW context of the GID. The reference 2132 * of GID table entry is already taken by the caller. 2133 */ 2134 ctx = rdma_read_gid_hw_context(sgid_attr); 2135 qp->qplib_qp.ah.sgid_index = ctx->idx; 2136 qp->qplib_qp.ah.host_sgid_index = grh->sgid_index; 2137 qp->qplib_qp.ah.hop_limit = grh->hop_limit; 2138 qp->qplib_qp.ah.traffic_class = grh->traffic_class >> 2; 2139 qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr); 2140 ether_addr_copy(qp->qplib_qp.ah.dmac, 2141 qp_attr->ah_attr.roce.dmac); 2142 2143 rc = rdma_read_gid_l2_fields(sgid_attr, NULL, 2144 &qp->qplib_qp.smac[0]); 2145 if (rc) 2146 return rc; 2147 2148 nw_type = rdma_gid_attr_network_type(sgid_attr); 2149 switch (nw_type) { 2150 case RDMA_NETWORK_IPV4: 2151 qp->qplib_qp.nw_type = 2152 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4; 2153 break; 2154 case RDMA_NETWORK_IPV6: 2155 qp->qplib_qp.nw_type = 2156 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6; 2157 break; 2158 default: 2159 qp->qplib_qp.nw_type = 2160 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1; 2161 break; 2162 } 2163 } 2164 2165 if (qp_attr->qp_state == IB_QPS_RTR) { 2166 enum ib_mtu qpmtu; 2167 2168 qpmtu = iboe_get_mtu(rdev->netdev->mtu); 2169 if (qp_attr_mask & IB_QP_PATH_MTU) { 2170 if (ib_mtu_enum_to_int(qp_attr->path_mtu) > 2171 ib_mtu_enum_to_int(qpmtu)) 2172 return -EINVAL; 2173 qpmtu = qp_attr->path_mtu; 2174 } 2175 2176 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU; 2177 qp->qplib_qp.path_mtu = __from_ib_mtu(qpmtu); 2178 qp->qplib_qp.mtu = ib_mtu_enum_to_int(qpmtu); 2179 } 2180 2181 if (qp_attr_mask & IB_QP_TIMEOUT) { 2182 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT; 2183 qp->qplib_qp.timeout = qp_attr->timeout; 2184 } 2185 if (qp_attr_mask & IB_QP_RETRY_CNT) { 2186 qp->qplib_qp.modify_flags |= 2187 CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT; 2188 qp->qplib_qp.retry_cnt = qp_attr->retry_cnt; 2189 } 2190 if (qp_attr_mask & IB_QP_RNR_RETRY) { 2191 qp->qplib_qp.modify_flags |= 2192 CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY; 2193 qp->qplib_qp.rnr_retry = qp_attr->rnr_retry; 2194 } 2195 if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) { 2196 qp->qplib_qp.modify_flags |= 2197 CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER; 2198 qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer; 2199 } 2200 if (qp_attr_mask & IB_QP_RQ_PSN) { 2201 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN; 2202 qp->qplib_qp.rq.psn = qp_attr->rq_psn; 2203 } 2204 if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 2205 qp->qplib_qp.modify_flags |= 2206 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC; 2207 /* Cap the max_rd_atomic to device max */ 2208 qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic, 2209 dev_attr->max_qp_rd_atom); 2210 } 2211 if (qp_attr_mask & IB_QP_SQ_PSN) { 2212 qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN; 2213 qp->qplib_qp.sq.psn = qp_attr->sq_psn; 2214 } 2215 if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 2216 if (qp_attr->max_dest_rd_atomic > 2217 dev_attr->max_qp_init_rd_atom) { 2218 ibdev_err(&rdev->ibdev, 2219 "max_dest_rd_atomic requested%d is > dev_max%d", 2220 qp_attr->max_dest_rd_atomic, 2221 dev_attr->max_qp_init_rd_atom); 2222 return -EINVAL; 2223 } 2224 2225 qp->qplib_qp.modify_flags |= 2226 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC; 2227 qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic; 2228 } 2229 if (qp_attr_mask & IB_QP_CAP) { 2230 struct bnxt_re_ucontext *uctx = 2231 rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx); 2232 2233 qp->qplib_qp.modify_flags |= 2234 CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE | 2235 CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE | 2236 CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE | 2237 CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE | 2238 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA; 2239 if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) || 2240 (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) || 2241 (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) || 2242 (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) || 2243 (qp_attr->cap.max_inline_data >= 2244 dev_attr->max_inline_data)) { 2245 ibdev_err(&rdev->ibdev, 2246 "Create QP failed - max exceeded"); 2247 return -EINVAL; 2248 } 2249 entries = bnxt_re_init_depth(qp_attr->cap.max_send_wr, uctx); 2250 qp->qplib_qp.sq.max_wqe = min_t(u32, entries, 2251 dev_attr->max_qp_wqes + 1); 2252 qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe - 2253 qp_attr->cap.max_send_wr; 2254 /* 2255 * Reserving one slot for Phantom WQE. Some application can 2256 * post one extra entry in this case. Allowing this to avoid 2257 * unexpected Queue full condition 2258 */ 2259 qp->qplib_qp.sq.q_full_delta -= 1; 2260 qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge; 2261 if (qp->qplib_qp.rq.max_wqe) { 2262 entries = bnxt_re_init_depth(qp_attr->cap.max_recv_wr, uctx); 2263 qp->qplib_qp.rq.max_wqe = 2264 min_t(u32, entries, dev_attr->max_qp_wqes + 1); 2265 qp->qplib_qp.rq.max_sw_wqe = qp->qplib_qp.rq.max_wqe; 2266 qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe - 2267 qp_attr->cap.max_recv_wr; 2268 qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge; 2269 } else { 2270 /* SRQ was used prior, just ignore the RQ caps */ 2271 } 2272 } 2273 if (qp_attr_mask & IB_QP_DEST_QPN) { 2274 qp->qplib_qp.modify_flags |= 2275 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID; 2276 qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num; 2277 } 2278 rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp); 2279 if (rc) { 2280 ibdev_err(&rdev->ibdev, "Failed to modify HW QP"); 2281 return rc; 2282 } 2283 if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) 2284 rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask); 2285 return rc; 2286 } 2287 2288 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr, 2289 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr) 2290 { 2291 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 2292 struct bnxt_re_dev *rdev = qp->rdev; 2293 struct bnxt_qplib_qp *qplib_qp; 2294 int rc; 2295 2296 qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL); 2297 if (!qplib_qp) 2298 return -ENOMEM; 2299 2300 qplib_qp->id = qp->qplib_qp.id; 2301 qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index; 2302 2303 rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp); 2304 if (rc) { 2305 ibdev_err(&rdev->ibdev, "Failed to query HW QP"); 2306 goto out; 2307 } 2308 qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state); 2309 qp_attr->cur_qp_state = __to_ib_qp_state(qplib_qp->cur_qp_state); 2310 qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0; 2311 qp_attr->qp_access_flags = __qp_access_flags_to_ib(qp->qplib_qp.cctx, 2312 qplib_qp->access); 2313 qp_attr->pkey_index = qplib_qp->pkey_index; 2314 qp_attr->qkey = qplib_qp->qkey; 2315 qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE; 2316 rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label, 2317 qplib_qp->ah.host_sgid_index, 2318 qplib_qp->ah.hop_limit, 2319 qplib_qp->ah.traffic_class); 2320 rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data); 2321 rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl); 2322 ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac); 2323 qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu); 2324 qp_attr->timeout = qplib_qp->timeout; 2325 qp_attr->retry_cnt = qplib_qp->retry_cnt; 2326 qp_attr->rnr_retry = qplib_qp->rnr_retry; 2327 qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer; 2328 qp_attr->port_num = __to_ib_port_num(qplib_qp->port_id); 2329 qp_attr->rq_psn = qplib_qp->rq.psn; 2330 qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic; 2331 qp_attr->sq_psn = qplib_qp->sq.psn; 2332 qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic; 2333 qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR : 2334 IB_SIGNAL_REQ_WR; 2335 qp_attr->dest_qp_num = qplib_qp->dest_qpn; 2336 2337 qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe; 2338 qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge; 2339 qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe; 2340 qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge; 2341 qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data; 2342 qp_init_attr->cap = qp_attr->cap; 2343 2344 out: 2345 kfree(qplib_qp); 2346 return rc; 2347 } 2348 2349 /* Routine for sending QP1 packets for RoCE V1 an V2 2350 */ 2351 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp, 2352 const struct ib_send_wr *wr, 2353 struct bnxt_qplib_swqe *wqe, 2354 int payload_size) 2355 { 2356 struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, 2357 ib_ah); 2358 struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah; 2359 const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr; 2360 struct bnxt_qplib_sge sge; 2361 u8 nw_type; 2362 u16 ether_type; 2363 union ib_gid dgid; 2364 bool is_eth = false; 2365 bool is_vlan = false; 2366 bool is_grh = false; 2367 bool is_udp = false; 2368 u8 ip_version = 0; 2369 u16 vlan_id = 0xFFFF; 2370 void *buf; 2371 int i, rc; 2372 2373 memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr)); 2374 2375 rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL); 2376 if (rc) 2377 return rc; 2378 2379 /* Get network header type for this GID */ 2380 nw_type = rdma_gid_attr_network_type(sgid_attr); 2381 switch (nw_type) { 2382 case RDMA_NETWORK_IPV4: 2383 nw_type = BNXT_RE_ROCEV2_IPV4_PACKET; 2384 break; 2385 case RDMA_NETWORK_IPV6: 2386 nw_type = BNXT_RE_ROCEV2_IPV6_PACKET; 2387 break; 2388 default: 2389 nw_type = BNXT_RE_ROCE_V1_PACKET; 2390 break; 2391 } 2392 memcpy(&dgid.raw, &qplib_ah->dgid, 16); 2393 is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP; 2394 if (is_udp) { 2395 if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) { 2396 ip_version = 4; 2397 ether_type = ETH_P_IP; 2398 } else { 2399 ip_version = 6; 2400 ether_type = ETH_P_IPV6; 2401 } 2402 is_grh = false; 2403 } else { 2404 ether_type = ETH_P_IBOE; 2405 is_grh = true; 2406 } 2407 2408 is_eth = true; 2409 is_vlan = vlan_id && (vlan_id < 0x1000); 2410 2411 ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh, 2412 ip_version, is_udp, 0, &qp->qp1_hdr); 2413 2414 /* ETH */ 2415 ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac); 2416 ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac); 2417 2418 /* For vlan, check the sgid for vlan existence */ 2419 2420 if (!is_vlan) { 2421 qp->qp1_hdr.eth.type = cpu_to_be16(ether_type); 2422 } else { 2423 qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type); 2424 qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id); 2425 } 2426 2427 if (is_grh || (ip_version == 6)) { 2428 memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw, 2429 sizeof(sgid_attr->gid)); 2430 memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data, 2431 sizeof(sgid_attr->gid)); 2432 qp->qp1_hdr.grh.hop_limit = qplib_ah->hop_limit; 2433 } 2434 2435 if (ip_version == 4) { 2436 qp->qp1_hdr.ip4.tos = 0; 2437 qp->qp1_hdr.ip4.id = 0; 2438 qp->qp1_hdr.ip4.frag_off = htons(IP_DF); 2439 qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit; 2440 2441 memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4); 2442 memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4); 2443 qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr); 2444 } 2445 2446 if (is_udp) { 2447 qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT); 2448 qp->qp1_hdr.udp.sport = htons(0x8CD1); 2449 qp->qp1_hdr.udp.csum = 0; 2450 } 2451 2452 /* BTH */ 2453 if (wr->opcode == IB_WR_SEND_WITH_IMM) { 2454 qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 2455 qp->qp1_hdr.immediate_present = 1; 2456 } else { 2457 qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 2458 } 2459 if (wr->send_flags & IB_SEND_SOLICITED) 2460 qp->qp1_hdr.bth.solicited_event = 1; 2461 /* pad_count */ 2462 qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3; 2463 2464 /* P_key for QP1 is for all members */ 2465 qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF); 2466 qp->qp1_hdr.bth.destination_qpn = IB_QP1; 2467 qp->qp1_hdr.bth.ack_req = 0; 2468 qp->send_psn++; 2469 qp->send_psn &= BTH_PSN_MASK; 2470 qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn); 2471 /* DETH */ 2472 /* Use the priviledged Q_Key for QP1 */ 2473 qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY); 2474 qp->qp1_hdr.deth.source_qpn = IB_QP1; 2475 2476 /* Pack the QP1 to the transmit buffer */ 2477 buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge); 2478 if (buf) { 2479 ib_ud_header_pack(&qp->qp1_hdr, buf); 2480 for (i = wqe->num_sge; i; i--) { 2481 wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr; 2482 wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey; 2483 wqe->sg_list[i].size = wqe->sg_list[i - 1].size; 2484 } 2485 2486 /* 2487 * Max Header buf size for IPV6 RoCE V2 is 86, 2488 * which is same as the QP1 SQ header buffer. 2489 * Header buf size for IPV4 RoCE V2 can be 66. 2490 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20). 2491 * Subtract 20 bytes from QP1 SQ header buf size 2492 */ 2493 if (is_udp && ip_version == 4) 2494 sge.size -= 20; 2495 /* 2496 * Max Header buf size for RoCE V1 is 78. 2497 * ETH(14) + VLAN(4) + GRH(40) + BTH(20). 2498 * Subtract 8 bytes from QP1 SQ header buf size 2499 */ 2500 if (!is_udp) 2501 sge.size -= 8; 2502 2503 /* Subtract 4 bytes for non vlan packets */ 2504 if (!is_vlan) 2505 sge.size -= 4; 2506 2507 wqe->sg_list[0].addr = sge.addr; 2508 wqe->sg_list[0].lkey = sge.lkey; 2509 wqe->sg_list[0].size = sge.size; 2510 wqe->num_sge++; 2511 2512 } else { 2513 ibdev_err(&qp->rdev->ibdev, "QP1 buffer is empty!"); 2514 rc = -ENOMEM; 2515 } 2516 return rc; 2517 } 2518 2519 /* For the MAD layer, it only provides the recv SGE the size of 2520 * ib_grh + MAD datagram. No Ethernet headers, Ethertype, BTH, DETH, 2521 * nor RoCE iCRC. The Cu+ solution must provide buffer for the entire 2522 * receive packet (334 bytes) with no VLAN and then copy the GRH 2523 * and the MAD datagram out to the provided SGE. 2524 */ 2525 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp, 2526 const struct ib_recv_wr *wr, 2527 struct bnxt_qplib_swqe *wqe, 2528 int payload_size) 2529 { 2530 struct bnxt_re_sqp_entries *sqp_entry; 2531 struct bnxt_qplib_sge ref, sge; 2532 struct bnxt_re_dev *rdev; 2533 u32 rq_prod_index; 2534 2535 rdev = qp->rdev; 2536 2537 rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp); 2538 2539 if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge)) 2540 return -ENOMEM; 2541 2542 /* Create 1 SGE to receive the entire 2543 * ethernet packet 2544 */ 2545 /* Save the reference from ULP */ 2546 ref.addr = wqe->sg_list[0].addr; 2547 ref.lkey = wqe->sg_list[0].lkey; 2548 ref.size = wqe->sg_list[0].size; 2549 2550 sqp_entry = &rdev->gsi_ctx.sqp_tbl[rq_prod_index]; 2551 2552 /* SGE 1 */ 2553 wqe->sg_list[0].addr = sge.addr; 2554 wqe->sg_list[0].lkey = sge.lkey; 2555 wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2; 2556 sge.size -= wqe->sg_list[0].size; 2557 2558 sqp_entry->sge.addr = ref.addr; 2559 sqp_entry->sge.lkey = ref.lkey; 2560 sqp_entry->sge.size = ref.size; 2561 /* Store the wrid for reporting completion */ 2562 sqp_entry->wrid = wqe->wr_id; 2563 /* change the wqe->wrid to table index */ 2564 wqe->wr_id = rq_prod_index; 2565 return 0; 2566 } 2567 2568 static int is_ud_qp(struct bnxt_re_qp *qp) 2569 { 2570 return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD || 2571 qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI); 2572 } 2573 2574 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp, 2575 const struct ib_send_wr *wr, 2576 struct bnxt_qplib_swqe *wqe) 2577 { 2578 struct bnxt_re_ah *ah = NULL; 2579 2580 if (is_ud_qp(qp)) { 2581 ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah); 2582 wqe->send.q_key = ud_wr(wr)->remote_qkey; 2583 wqe->send.dst_qp = ud_wr(wr)->remote_qpn; 2584 wqe->send.avid = ah->qplib_ah.id; 2585 } 2586 switch (wr->opcode) { 2587 case IB_WR_SEND: 2588 wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND; 2589 break; 2590 case IB_WR_SEND_WITH_IMM: 2591 wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM; 2592 wqe->send.imm_data = be32_to_cpu(wr->ex.imm_data); 2593 break; 2594 case IB_WR_SEND_WITH_INV: 2595 wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV; 2596 wqe->send.inv_key = wr->ex.invalidate_rkey; 2597 break; 2598 default: 2599 return -EINVAL; 2600 } 2601 if (wr->send_flags & IB_SEND_SIGNALED) 2602 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2603 if (wr->send_flags & IB_SEND_FENCE) 2604 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2605 if (wr->send_flags & IB_SEND_SOLICITED) 2606 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2607 if (wr->send_flags & IB_SEND_INLINE) 2608 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE; 2609 2610 return 0; 2611 } 2612 2613 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr, 2614 struct bnxt_qplib_swqe *wqe) 2615 { 2616 switch (wr->opcode) { 2617 case IB_WR_RDMA_WRITE: 2618 wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE; 2619 break; 2620 case IB_WR_RDMA_WRITE_WITH_IMM: 2621 wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM; 2622 wqe->rdma.imm_data = be32_to_cpu(wr->ex.imm_data); 2623 break; 2624 case IB_WR_RDMA_READ: 2625 wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ; 2626 wqe->rdma.inv_key = wr->ex.invalidate_rkey; 2627 break; 2628 default: 2629 return -EINVAL; 2630 } 2631 wqe->rdma.remote_va = rdma_wr(wr)->remote_addr; 2632 wqe->rdma.r_key = rdma_wr(wr)->rkey; 2633 if (wr->send_flags & IB_SEND_SIGNALED) 2634 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2635 if (wr->send_flags & IB_SEND_FENCE) 2636 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2637 if (wr->send_flags & IB_SEND_SOLICITED) 2638 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2639 if (wr->send_flags & IB_SEND_INLINE) 2640 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE; 2641 2642 return 0; 2643 } 2644 2645 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr, 2646 struct bnxt_qplib_swqe *wqe) 2647 { 2648 switch (wr->opcode) { 2649 case IB_WR_ATOMIC_CMP_AND_SWP: 2650 wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP; 2651 wqe->atomic.cmp_data = atomic_wr(wr)->compare_add; 2652 wqe->atomic.swap_data = atomic_wr(wr)->swap; 2653 break; 2654 case IB_WR_ATOMIC_FETCH_AND_ADD: 2655 wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD; 2656 wqe->atomic.cmp_data = atomic_wr(wr)->compare_add; 2657 break; 2658 default: 2659 return -EINVAL; 2660 } 2661 wqe->atomic.remote_va = atomic_wr(wr)->remote_addr; 2662 wqe->atomic.r_key = atomic_wr(wr)->rkey; 2663 if (wr->send_flags & IB_SEND_SIGNALED) 2664 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2665 if (wr->send_flags & IB_SEND_FENCE) 2666 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2667 if (wr->send_flags & IB_SEND_SOLICITED) 2668 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2669 return 0; 2670 } 2671 2672 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr, 2673 struct bnxt_qplib_swqe *wqe) 2674 { 2675 wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV; 2676 wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey; 2677 2678 if (wr->send_flags & IB_SEND_SIGNALED) 2679 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2680 if (wr->send_flags & IB_SEND_SOLICITED) 2681 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT; 2682 2683 return 0; 2684 } 2685 2686 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr, 2687 struct bnxt_qplib_swqe *wqe) 2688 { 2689 struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr); 2690 struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl; 2691 int access = wr->access; 2692 2693 wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0]; 2694 wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0]; 2695 wqe->frmr.page_list = mr->pages; 2696 wqe->frmr.page_list_len = mr->npages; 2697 wqe->frmr.levels = qplib_frpl->hwq.level; 2698 wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR; 2699 2700 if (wr->wr.send_flags & IB_SEND_SIGNALED) 2701 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP; 2702 2703 if (access & IB_ACCESS_LOCAL_WRITE) 2704 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE; 2705 if (access & IB_ACCESS_REMOTE_READ) 2706 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ; 2707 if (access & IB_ACCESS_REMOTE_WRITE) 2708 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE; 2709 if (access & IB_ACCESS_REMOTE_ATOMIC) 2710 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC; 2711 if (access & IB_ACCESS_MW_BIND) 2712 wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND; 2713 2714 wqe->frmr.l_key = wr->key; 2715 wqe->frmr.length = wr->mr->length; 2716 wqe->frmr.pbl_pg_sz_log = ilog2(PAGE_SIZE >> PAGE_SHIFT_4K); 2717 wqe->frmr.pg_sz_log = ilog2(wr->mr->page_size >> PAGE_SHIFT_4K); 2718 wqe->frmr.va = wr->mr->iova; 2719 return 0; 2720 } 2721 2722 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev, 2723 const struct ib_send_wr *wr, 2724 struct bnxt_qplib_swqe *wqe) 2725 { 2726 /* Copy the inline data to the data field */ 2727 u8 *in_data; 2728 u32 i, sge_len; 2729 void *sge_addr; 2730 2731 in_data = wqe->inline_data; 2732 for (i = 0; i < wr->num_sge; i++) { 2733 sge_addr = (void *)(unsigned long) 2734 wr->sg_list[i].addr; 2735 sge_len = wr->sg_list[i].length; 2736 2737 if ((sge_len + wqe->inline_len) > 2738 BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) { 2739 ibdev_err(&rdev->ibdev, 2740 "Inline data size requested > supported value"); 2741 return -EINVAL; 2742 } 2743 sge_len = wr->sg_list[i].length; 2744 2745 memcpy(in_data, sge_addr, sge_len); 2746 in_data += wr->sg_list[i].length; 2747 wqe->inline_len += wr->sg_list[i].length; 2748 } 2749 return wqe->inline_len; 2750 } 2751 2752 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev, 2753 const struct ib_send_wr *wr, 2754 struct bnxt_qplib_swqe *wqe) 2755 { 2756 int payload_sz = 0; 2757 2758 if (wr->send_flags & IB_SEND_INLINE) 2759 payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe); 2760 else 2761 payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list, 2762 wqe->num_sge); 2763 2764 return payload_sz; 2765 } 2766 2767 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp) 2768 { 2769 if ((qp->ib_qp.qp_type == IB_QPT_UD || 2770 qp->ib_qp.qp_type == IB_QPT_GSI || 2771 qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) && 2772 qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) { 2773 int qp_attr_mask; 2774 struct ib_qp_attr qp_attr; 2775 2776 qp_attr_mask = IB_QP_STATE; 2777 qp_attr.qp_state = IB_QPS_RTS; 2778 bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL); 2779 qp->qplib_qp.wqe_cnt = 0; 2780 } 2781 } 2782 2783 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev, 2784 struct bnxt_re_qp *qp, 2785 const struct ib_send_wr *wr) 2786 { 2787 int rc = 0, payload_sz = 0; 2788 unsigned long flags; 2789 2790 spin_lock_irqsave(&qp->sq_lock, flags); 2791 while (wr) { 2792 struct bnxt_qplib_swqe wqe = {}; 2793 2794 /* Common */ 2795 wqe.num_sge = wr->num_sge; 2796 if (wr->num_sge > qp->qplib_qp.sq.max_sge) { 2797 ibdev_err(&rdev->ibdev, 2798 "Limit exceeded for Send SGEs"); 2799 rc = -EINVAL; 2800 goto bad; 2801 } 2802 2803 payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe); 2804 if (payload_sz < 0) { 2805 rc = -EINVAL; 2806 goto bad; 2807 } 2808 wqe.wr_id = wr->wr_id; 2809 2810 wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND; 2811 2812 rc = bnxt_re_build_send_wqe(qp, wr, &wqe); 2813 if (!rc) 2814 rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe); 2815 bad: 2816 if (rc) { 2817 ibdev_err(&rdev->ibdev, 2818 "Post send failed opcode = %#x rc = %d", 2819 wr->opcode, rc); 2820 break; 2821 } 2822 wr = wr->next; 2823 } 2824 bnxt_qplib_post_send_db(&qp->qplib_qp); 2825 if (!bnxt_qplib_is_chip_gen_p5_p7(qp->rdev->chip_ctx)) 2826 bnxt_ud_qp_hw_stall_workaround(qp); 2827 spin_unlock_irqrestore(&qp->sq_lock, flags); 2828 return rc; 2829 } 2830 2831 static void bnxt_re_legacy_set_uc_fence(struct bnxt_qplib_swqe *wqe) 2832 { 2833 /* Need unconditional fence for non-wire memory opcode 2834 * to work as expected. 2835 */ 2836 if (wqe->type == BNXT_QPLIB_SWQE_TYPE_LOCAL_INV || 2837 wqe->type == BNXT_QPLIB_SWQE_TYPE_FAST_REG_MR || 2838 wqe->type == BNXT_QPLIB_SWQE_TYPE_REG_MR || 2839 wqe->type == BNXT_QPLIB_SWQE_TYPE_BIND_MW) 2840 wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE; 2841 } 2842 2843 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr, 2844 const struct ib_send_wr **bad_wr) 2845 { 2846 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 2847 struct bnxt_qplib_swqe wqe; 2848 int rc = 0, payload_sz = 0; 2849 unsigned long flags; 2850 2851 spin_lock_irqsave(&qp->sq_lock, flags); 2852 while (wr) { 2853 /* House keeping */ 2854 memset(&wqe, 0, sizeof(wqe)); 2855 2856 /* Common */ 2857 wqe.num_sge = wr->num_sge; 2858 if (wr->num_sge > qp->qplib_qp.sq.max_sge) { 2859 ibdev_err(&qp->rdev->ibdev, 2860 "Limit exceeded for Send SGEs"); 2861 rc = -EINVAL; 2862 goto bad; 2863 } 2864 2865 payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe); 2866 if (payload_sz < 0) { 2867 rc = -EINVAL; 2868 goto bad; 2869 } 2870 wqe.wr_id = wr->wr_id; 2871 2872 switch (wr->opcode) { 2873 case IB_WR_SEND: 2874 case IB_WR_SEND_WITH_IMM: 2875 if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) { 2876 rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe, 2877 payload_sz); 2878 if (rc) 2879 goto bad; 2880 wqe.rawqp1.lflags |= 2881 SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC; 2882 } 2883 switch (wr->send_flags) { 2884 case IB_SEND_IP_CSUM: 2885 wqe.rawqp1.lflags |= 2886 SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM; 2887 break; 2888 default: 2889 break; 2890 } 2891 fallthrough; 2892 case IB_WR_SEND_WITH_INV: 2893 rc = bnxt_re_build_send_wqe(qp, wr, &wqe); 2894 break; 2895 case IB_WR_RDMA_WRITE: 2896 case IB_WR_RDMA_WRITE_WITH_IMM: 2897 case IB_WR_RDMA_READ: 2898 rc = bnxt_re_build_rdma_wqe(wr, &wqe); 2899 break; 2900 case IB_WR_ATOMIC_CMP_AND_SWP: 2901 case IB_WR_ATOMIC_FETCH_AND_ADD: 2902 rc = bnxt_re_build_atomic_wqe(wr, &wqe); 2903 break; 2904 case IB_WR_RDMA_READ_WITH_INV: 2905 ibdev_err(&qp->rdev->ibdev, 2906 "RDMA Read with Invalidate is not supported"); 2907 rc = -EINVAL; 2908 goto bad; 2909 case IB_WR_LOCAL_INV: 2910 rc = bnxt_re_build_inv_wqe(wr, &wqe); 2911 break; 2912 case IB_WR_REG_MR: 2913 rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe); 2914 break; 2915 default: 2916 /* Unsupported WRs */ 2917 ibdev_err(&qp->rdev->ibdev, 2918 "WR (%#x) is not supported", wr->opcode); 2919 rc = -EINVAL; 2920 goto bad; 2921 } 2922 if (!rc) { 2923 if (!bnxt_qplib_is_chip_gen_p5_p7(qp->rdev->chip_ctx)) 2924 bnxt_re_legacy_set_uc_fence(&wqe); 2925 rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe); 2926 } 2927 bad: 2928 if (rc) { 2929 ibdev_err(&qp->rdev->ibdev, 2930 "post_send failed op:%#x qps = %#x rc = %d\n", 2931 wr->opcode, qp->qplib_qp.state, rc); 2932 *bad_wr = wr; 2933 break; 2934 } 2935 wr = wr->next; 2936 } 2937 bnxt_qplib_post_send_db(&qp->qplib_qp); 2938 if (!bnxt_qplib_is_chip_gen_p5_p7(qp->rdev->chip_ctx)) 2939 bnxt_ud_qp_hw_stall_workaround(qp); 2940 spin_unlock_irqrestore(&qp->sq_lock, flags); 2941 2942 return rc; 2943 } 2944 2945 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev, 2946 struct bnxt_re_qp *qp, 2947 const struct ib_recv_wr *wr) 2948 { 2949 struct bnxt_qplib_swqe wqe; 2950 int rc = 0; 2951 2952 while (wr) { 2953 /* House keeping */ 2954 memset(&wqe, 0, sizeof(wqe)); 2955 2956 /* Common */ 2957 wqe.num_sge = wr->num_sge; 2958 if (wr->num_sge > qp->qplib_qp.rq.max_sge) { 2959 ibdev_err(&rdev->ibdev, 2960 "Limit exceeded for Receive SGEs"); 2961 rc = -EINVAL; 2962 break; 2963 } 2964 bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge); 2965 wqe.wr_id = wr->wr_id; 2966 wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV; 2967 2968 rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe); 2969 if (rc) 2970 break; 2971 2972 wr = wr->next; 2973 } 2974 if (!rc) 2975 bnxt_qplib_post_recv_db(&qp->qplib_qp); 2976 return rc; 2977 } 2978 2979 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr, 2980 const struct ib_recv_wr **bad_wr) 2981 { 2982 struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); 2983 struct bnxt_qplib_swqe wqe; 2984 int rc = 0, payload_sz = 0; 2985 unsigned long flags; 2986 u32 count = 0; 2987 2988 spin_lock_irqsave(&qp->rq_lock, flags); 2989 while (wr) { 2990 /* House keeping */ 2991 memset(&wqe, 0, sizeof(wqe)); 2992 2993 /* Common */ 2994 wqe.num_sge = wr->num_sge; 2995 if (wr->num_sge > qp->qplib_qp.rq.max_sge) { 2996 ibdev_err(&qp->rdev->ibdev, 2997 "Limit exceeded for Receive SGEs"); 2998 rc = -EINVAL; 2999 *bad_wr = wr; 3000 break; 3001 } 3002 3003 payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, 3004 wr->num_sge); 3005 wqe.wr_id = wr->wr_id; 3006 wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV; 3007 3008 if (ib_qp->qp_type == IB_QPT_GSI && 3009 qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI) 3010 rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe, 3011 payload_sz); 3012 if (!rc) 3013 rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe); 3014 if (rc) { 3015 *bad_wr = wr; 3016 break; 3017 } 3018 3019 /* Ring DB if the RQEs posted reaches a threshold value */ 3020 if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) { 3021 bnxt_qplib_post_recv_db(&qp->qplib_qp); 3022 count = 0; 3023 } 3024 3025 wr = wr->next; 3026 } 3027 3028 if (count) 3029 bnxt_qplib_post_recv_db(&qp->qplib_qp); 3030 3031 spin_unlock_irqrestore(&qp->rq_lock, flags); 3032 3033 return rc; 3034 } 3035 3036 static struct bnxt_qplib_nq *bnxt_re_get_nq(struct bnxt_re_dev *rdev) 3037 { 3038 int min, indx; 3039 3040 mutex_lock(&rdev->nqr->load_lock); 3041 for (indx = 0, min = 0; indx < (rdev->nqr->num_msix - 1); indx++) { 3042 if (rdev->nqr->nq[min].load > rdev->nqr->nq[indx].load) 3043 min = indx; 3044 } 3045 rdev->nqr->nq[min].load++; 3046 mutex_unlock(&rdev->nqr->load_lock); 3047 3048 return &rdev->nqr->nq[min]; 3049 } 3050 3051 static void bnxt_re_put_nq(struct bnxt_re_dev *rdev, struct bnxt_qplib_nq *nq) 3052 { 3053 mutex_lock(&rdev->nqr->load_lock); 3054 nq->load--; 3055 mutex_unlock(&rdev->nqr->load_lock); 3056 } 3057 3058 /* Completion Queues */ 3059 int bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata) 3060 { 3061 struct bnxt_qplib_chip_ctx *cctx; 3062 struct bnxt_qplib_nq *nq; 3063 struct bnxt_re_dev *rdev; 3064 struct bnxt_re_cq *cq; 3065 3066 cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq); 3067 rdev = cq->rdev; 3068 nq = cq->qplib_cq.nq; 3069 cctx = rdev->chip_ctx; 3070 3071 if (cctx->modes.toggle_bits & BNXT_QPLIB_CQ_TOGGLE_BIT) { 3072 free_page((unsigned long)cq->uctx_cq_page); 3073 hash_del(&cq->hash_entry); 3074 } 3075 bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq); 3076 3077 bnxt_re_put_nq(rdev, nq); 3078 ib_umem_release(cq->umem); 3079 3080 atomic_dec(&rdev->stats.res.cq_count); 3081 nq->budget--; 3082 kfree(cq->cql); 3083 return 0; 3084 } 3085 3086 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, 3087 struct uverbs_attr_bundle *attrs) 3088 { 3089 struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq); 3090 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev); 3091 struct ib_udata *udata = &attrs->driver_udata; 3092 struct bnxt_re_ucontext *uctx = 3093 rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx); 3094 struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr; 3095 struct bnxt_qplib_chip_ctx *cctx; 3096 int cqe = attr->cqe; 3097 int rc, entries; 3098 u32 active_cqs; 3099 3100 if (attr->flags) 3101 return -EOPNOTSUPP; 3102 3103 /* Validate CQ fields */ 3104 if (cqe < 1 || cqe > dev_attr->max_cq_wqes) { 3105 ibdev_err(&rdev->ibdev, "Failed to create CQ -max exceeded"); 3106 return -EINVAL; 3107 } 3108 3109 cq->rdev = rdev; 3110 cctx = rdev->chip_ctx; 3111 cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq); 3112 3113 entries = bnxt_re_init_depth(cqe + 1, uctx); 3114 if (entries > dev_attr->max_cq_wqes + 1) 3115 entries = dev_attr->max_cq_wqes + 1; 3116 3117 cq->qplib_cq.sg_info.pgsize = PAGE_SIZE; 3118 cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT; 3119 if (udata) { 3120 struct bnxt_re_cq_req req; 3121 if (ib_copy_from_udata(&req, udata, sizeof(req))) { 3122 rc = -EFAULT; 3123 goto fail; 3124 } 3125 3126 cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va, 3127 entries * sizeof(struct cq_base), 3128 IB_ACCESS_LOCAL_WRITE); 3129 if (IS_ERR(cq->umem)) { 3130 rc = PTR_ERR(cq->umem); 3131 goto fail; 3132 } 3133 cq->qplib_cq.sg_info.umem = cq->umem; 3134 cq->qplib_cq.dpi = &uctx->dpi; 3135 } else { 3136 cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL); 3137 cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe), 3138 GFP_KERNEL); 3139 if (!cq->cql) { 3140 rc = -ENOMEM; 3141 goto fail; 3142 } 3143 3144 cq->qplib_cq.dpi = &rdev->dpi_privileged; 3145 } 3146 cq->qplib_cq.max_wqe = entries; 3147 cq->qplib_cq.coalescing = &rdev->cq_coalescing; 3148 cq->qplib_cq.nq = bnxt_re_get_nq(rdev); 3149 cq->qplib_cq.cnq_hw_ring_id = cq->qplib_cq.nq->ring_id; 3150 3151 rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq); 3152 if (rc) { 3153 ibdev_err(&rdev->ibdev, "Failed to create HW CQ"); 3154 goto fail; 3155 } 3156 3157 cq->ib_cq.cqe = entries; 3158 cq->cq_period = cq->qplib_cq.period; 3159 3160 active_cqs = atomic_inc_return(&rdev->stats.res.cq_count); 3161 if (active_cqs > rdev->stats.res.cq_watermark) 3162 rdev->stats.res.cq_watermark = active_cqs; 3163 spin_lock_init(&cq->cq_lock); 3164 3165 if (udata) { 3166 struct bnxt_re_cq_resp resp = {}; 3167 3168 if (cctx->modes.toggle_bits & BNXT_QPLIB_CQ_TOGGLE_BIT) { 3169 hash_add(rdev->cq_hash, &cq->hash_entry, cq->qplib_cq.id); 3170 /* Allocate a page */ 3171 cq->uctx_cq_page = (void *)get_zeroed_page(GFP_KERNEL); 3172 if (!cq->uctx_cq_page) { 3173 rc = -ENOMEM; 3174 goto c2fail; 3175 } 3176 resp.comp_mask |= BNXT_RE_CQ_TOGGLE_PAGE_SUPPORT; 3177 } 3178 resp.cqid = cq->qplib_cq.id; 3179 resp.tail = cq->qplib_cq.hwq.cons; 3180 resp.phase = cq->qplib_cq.period; 3181 resp.rsvd = 0; 3182 rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen)); 3183 if (rc) { 3184 ibdev_err(&rdev->ibdev, "Failed to copy CQ udata"); 3185 bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq); 3186 goto free_mem; 3187 } 3188 } 3189 3190 return 0; 3191 3192 free_mem: 3193 free_page((unsigned long)cq->uctx_cq_page); 3194 c2fail: 3195 ib_umem_release(cq->umem); 3196 fail: 3197 kfree(cq->cql); 3198 return rc; 3199 } 3200 3201 static void bnxt_re_resize_cq_complete(struct bnxt_re_cq *cq) 3202 { 3203 struct bnxt_re_dev *rdev = cq->rdev; 3204 3205 bnxt_qplib_resize_cq_complete(&rdev->qplib_res, &cq->qplib_cq); 3206 3207 cq->qplib_cq.max_wqe = cq->resize_cqe; 3208 if (cq->resize_umem) { 3209 ib_umem_release(cq->umem); 3210 cq->umem = cq->resize_umem; 3211 cq->resize_umem = NULL; 3212 cq->resize_cqe = 0; 3213 } 3214 } 3215 3216 int bnxt_re_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) 3217 { 3218 struct bnxt_qplib_sg_info sg_info = {}; 3219 struct bnxt_qplib_dpi *orig_dpi = NULL; 3220 struct bnxt_qplib_dev_attr *dev_attr; 3221 struct bnxt_re_ucontext *uctx = NULL; 3222 struct bnxt_re_resize_cq_req req; 3223 struct bnxt_re_dev *rdev; 3224 struct bnxt_re_cq *cq; 3225 int rc, entries; 3226 3227 cq = container_of(ibcq, struct bnxt_re_cq, ib_cq); 3228 rdev = cq->rdev; 3229 dev_attr = rdev->dev_attr; 3230 if (!ibcq->uobject) { 3231 ibdev_err(&rdev->ibdev, "Kernel CQ Resize not supported"); 3232 return -EOPNOTSUPP; 3233 } 3234 3235 if (cq->resize_umem) { 3236 ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - Busy", 3237 cq->qplib_cq.id); 3238 return -EBUSY; 3239 } 3240 3241 /* Check the requested cq depth out of supported depth */ 3242 if (cqe < 1 || cqe > dev_attr->max_cq_wqes) { 3243 ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - out of range cqe %d", 3244 cq->qplib_cq.id, cqe); 3245 return -EINVAL; 3246 } 3247 3248 uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx); 3249 entries = bnxt_re_init_depth(cqe + 1, uctx); 3250 if (entries > dev_attr->max_cq_wqes + 1) 3251 entries = dev_attr->max_cq_wqes + 1; 3252 3253 /* uverbs consumer */ 3254 if (ib_copy_from_udata(&req, udata, sizeof(req))) { 3255 rc = -EFAULT; 3256 goto fail; 3257 } 3258 3259 cq->resize_umem = ib_umem_get(&rdev->ibdev, req.cq_va, 3260 entries * sizeof(struct cq_base), 3261 IB_ACCESS_LOCAL_WRITE); 3262 if (IS_ERR(cq->resize_umem)) { 3263 rc = PTR_ERR(cq->resize_umem); 3264 cq->resize_umem = NULL; 3265 ibdev_err(&rdev->ibdev, "%s: ib_umem_get failed! rc = %d\n", 3266 __func__, rc); 3267 goto fail; 3268 } 3269 cq->resize_cqe = entries; 3270 memcpy(&sg_info, &cq->qplib_cq.sg_info, sizeof(sg_info)); 3271 orig_dpi = cq->qplib_cq.dpi; 3272 3273 cq->qplib_cq.sg_info.umem = cq->resize_umem; 3274 cq->qplib_cq.sg_info.pgsize = PAGE_SIZE; 3275 cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT; 3276 cq->qplib_cq.dpi = &uctx->dpi; 3277 3278 rc = bnxt_qplib_resize_cq(&rdev->qplib_res, &cq->qplib_cq, entries); 3279 if (rc) { 3280 ibdev_err(&rdev->ibdev, "Resize HW CQ %#x failed!", 3281 cq->qplib_cq.id); 3282 goto fail; 3283 } 3284 3285 cq->ib_cq.cqe = cq->resize_cqe; 3286 atomic_inc(&rdev->stats.res.resize_count); 3287 3288 return 0; 3289 3290 fail: 3291 if (cq->resize_umem) { 3292 ib_umem_release(cq->resize_umem); 3293 cq->resize_umem = NULL; 3294 cq->resize_cqe = 0; 3295 memcpy(&cq->qplib_cq.sg_info, &sg_info, sizeof(sg_info)); 3296 cq->qplib_cq.dpi = orig_dpi; 3297 } 3298 return rc; 3299 } 3300 3301 static u8 __req_to_ib_wc_status(u8 qstatus) 3302 { 3303 switch (qstatus) { 3304 case CQ_REQ_STATUS_OK: 3305 return IB_WC_SUCCESS; 3306 case CQ_REQ_STATUS_BAD_RESPONSE_ERR: 3307 return IB_WC_BAD_RESP_ERR; 3308 case CQ_REQ_STATUS_LOCAL_LENGTH_ERR: 3309 return IB_WC_LOC_LEN_ERR; 3310 case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR: 3311 return IB_WC_LOC_QP_OP_ERR; 3312 case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR: 3313 return IB_WC_LOC_PROT_ERR; 3314 case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR: 3315 return IB_WC_GENERAL_ERR; 3316 case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR: 3317 return IB_WC_REM_INV_REQ_ERR; 3318 case CQ_REQ_STATUS_REMOTE_ACCESS_ERR: 3319 return IB_WC_REM_ACCESS_ERR; 3320 case CQ_REQ_STATUS_REMOTE_OPERATION_ERR: 3321 return IB_WC_REM_OP_ERR; 3322 case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR: 3323 return IB_WC_RNR_RETRY_EXC_ERR; 3324 case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR: 3325 return IB_WC_RETRY_EXC_ERR; 3326 case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR: 3327 return IB_WC_WR_FLUSH_ERR; 3328 default: 3329 return IB_WC_GENERAL_ERR; 3330 } 3331 return 0; 3332 } 3333 3334 static u8 __rawqp1_to_ib_wc_status(u8 qstatus) 3335 { 3336 switch (qstatus) { 3337 case CQ_RES_RAWETH_QP1_STATUS_OK: 3338 return IB_WC_SUCCESS; 3339 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR: 3340 return IB_WC_LOC_ACCESS_ERR; 3341 case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR: 3342 return IB_WC_LOC_LEN_ERR; 3343 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR: 3344 return IB_WC_LOC_PROT_ERR; 3345 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR: 3346 return IB_WC_LOC_QP_OP_ERR; 3347 case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR: 3348 return IB_WC_GENERAL_ERR; 3349 case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR: 3350 return IB_WC_WR_FLUSH_ERR; 3351 case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR: 3352 return IB_WC_WR_FLUSH_ERR; 3353 default: 3354 return IB_WC_GENERAL_ERR; 3355 } 3356 } 3357 3358 static u8 __rc_to_ib_wc_status(u8 qstatus) 3359 { 3360 switch (qstatus) { 3361 case CQ_RES_RC_STATUS_OK: 3362 return IB_WC_SUCCESS; 3363 case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR: 3364 return IB_WC_LOC_ACCESS_ERR; 3365 case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR: 3366 return IB_WC_LOC_LEN_ERR; 3367 case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR: 3368 return IB_WC_LOC_PROT_ERR; 3369 case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR: 3370 return IB_WC_LOC_QP_OP_ERR; 3371 case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR: 3372 return IB_WC_GENERAL_ERR; 3373 case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR: 3374 return IB_WC_REM_INV_REQ_ERR; 3375 case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR: 3376 return IB_WC_WR_FLUSH_ERR; 3377 case CQ_RES_RC_STATUS_HW_FLUSH_ERR: 3378 return IB_WC_WR_FLUSH_ERR; 3379 default: 3380 return IB_WC_GENERAL_ERR; 3381 } 3382 } 3383 3384 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe) 3385 { 3386 switch (cqe->type) { 3387 case BNXT_QPLIB_SWQE_TYPE_SEND: 3388 wc->opcode = IB_WC_SEND; 3389 break; 3390 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM: 3391 wc->opcode = IB_WC_SEND; 3392 wc->wc_flags |= IB_WC_WITH_IMM; 3393 break; 3394 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV: 3395 wc->opcode = IB_WC_SEND; 3396 wc->wc_flags |= IB_WC_WITH_INVALIDATE; 3397 break; 3398 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE: 3399 wc->opcode = IB_WC_RDMA_WRITE; 3400 break; 3401 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM: 3402 wc->opcode = IB_WC_RDMA_WRITE; 3403 wc->wc_flags |= IB_WC_WITH_IMM; 3404 break; 3405 case BNXT_QPLIB_SWQE_TYPE_RDMA_READ: 3406 wc->opcode = IB_WC_RDMA_READ; 3407 break; 3408 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP: 3409 wc->opcode = IB_WC_COMP_SWAP; 3410 break; 3411 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD: 3412 wc->opcode = IB_WC_FETCH_ADD; 3413 break; 3414 case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV: 3415 wc->opcode = IB_WC_LOCAL_INV; 3416 break; 3417 case BNXT_QPLIB_SWQE_TYPE_REG_MR: 3418 wc->opcode = IB_WC_REG_MR; 3419 break; 3420 default: 3421 wc->opcode = IB_WC_SEND; 3422 break; 3423 } 3424 3425 wc->status = __req_to_ib_wc_status(cqe->status); 3426 } 3427 3428 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags, 3429 u16 raweth_qp1_flags2) 3430 { 3431 bool is_ipv6 = false, is_ipv4 = false; 3432 3433 /* raweth_qp1_flags Bit 9-6 indicates itype */ 3434 if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE) 3435 != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE) 3436 return -1; 3437 3438 if (raweth_qp1_flags2 & 3439 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC && 3440 raweth_qp1_flags2 & 3441 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) { 3442 /* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */ 3443 (raweth_qp1_flags2 & 3444 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ? 3445 (is_ipv6 = true) : (is_ipv4 = true); 3446 return ((is_ipv6) ? 3447 BNXT_RE_ROCEV2_IPV6_PACKET : 3448 BNXT_RE_ROCEV2_IPV4_PACKET); 3449 } else { 3450 return BNXT_RE_ROCE_V1_PACKET; 3451 } 3452 } 3453 3454 static int bnxt_re_to_ib_nw_type(int nw_type) 3455 { 3456 u8 nw_hdr_type = 0xFF; 3457 3458 switch (nw_type) { 3459 case BNXT_RE_ROCE_V1_PACKET: 3460 nw_hdr_type = RDMA_NETWORK_ROCE_V1; 3461 break; 3462 case BNXT_RE_ROCEV2_IPV4_PACKET: 3463 nw_hdr_type = RDMA_NETWORK_IPV4; 3464 break; 3465 case BNXT_RE_ROCEV2_IPV6_PACKET: 3466 nw_hdr_type = RDMA_NETWORK_IPV6; 3467 break; 3468 } 3469 return nw_hdr_type; 3470 } 3471 3472 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev, 3473 void *rq_hdr_buf) 3474 { 3475 u8 *tmp_buf = NULL; 3476 struct ethhdr *eth_hdr; 3477 u16 eth_type; 3478 bool rc = false; 3479 3480 tmp_buf = (u8 *)rq_hdr_buf; 3481 /* 3482 * If dest mac is not same as I/F mac, this could be a 3483 * loopback address or multicast address, check whether 3484 * it is a loopback packet 3485 */ 3486 if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) { 3487 tmp_buf += 4; 3488 /* Check the ether type */ 3489 eth_hdr = (struct ethhdr *)tmp_buf; 3490 eth_type = ntohs(eth_hdr->h_proto); 3491 switch (eth_type) { 3492 case ETH_P_IBOE: 3493 rc = true; 3494 break; 3495 case ETH_P_IP: 3496 case ETH_P_IPV6: { 3497 u32 len; 3498 struct udphdr *udp_hdr; 3499 3500 len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) : 3501 sizeof(struct ipv6hdr)); 3502 tmp_buf += sizeof(struct ethhdr) + len; 3503 udp_hdr = (struct udphdr *)tmp_buf; 3504 if (ntohs(udp_hdr->dest) == 3505 ROCE_V2_UDP_DPORT) 3506 rc = true; 3507 break; 3508 } 3509 default: 3510 break; 3511 } 3512 } 3513 3514 return rc; 3515 } 3516 3517 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *gsi_qp, 3518 struct bnxt_qplib_cqe *cqe) 3519 { 3520 struct bnxt_re_dev *rdev = gsi_qp->rdev; 3521 struct bnxt_re_sqp_entries *sqp_entry = NULL; 3522 struct bnxt_re_qp *gsi_sqp = rdev->gsi_ctx.gsi_sqp; 3523 dma_addr_t shrq_hdr_buf_map; 3524 struct ib_sge s_sge[2] = {}; 3525 struct ib_sge r_sge[2] = {}; 3526 struct bnxt_re_ah *gsi_sah; 3527 struct ib_recv_wr rwr = {}; 3528 dma_addr_t rq_hdr_buf_map; 3529 struct ib_ud_wr udwr = {}; 3530 struct ib_send_wr *swr; 3531 u32 skip_bytes = 0; 3532 int pkt_type = 0; 3533 void *rq_hdr_buf; 3534 u32 offset = 0; 3535 u32 tbl_idx; 3536 int rc; 3537 3538 swr = &udwr.wr; 3539 tbl_idx = cqe->wr_id; 3540 3541 rq_hdr_buf = gsi_qp->qplib_qp.rq_hdr_buf + 3542 (tbl_idx * gsi_qp->qplib_qp.rq_hdr_buf_size); 3543 rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp, 3544 tbl_idx); 3545 3546 /* Shadow QP header buffer */ 3547 shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp, 3548 tbl_idx); 3549 sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx]; 3550 3551 /* Store this cqe */ 3552 memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe)); 3553 sqp_entry->qp1_qp = gsi_qp; 3554 3555 /* Find packet type from the cqe */ 3556 3557 pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags, 3558 cqe->raweth_qp1_flags2); 3559 if (pkt_type < 0) { 3560 ibdev_err(&rdev->ibdev, "Invalid packet\n"); 3561 return -EINVAL; 3562 } 3563 3564 /* Adjust the offset for the user buffer and post in the rq */ 3565 3566 if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET) 3567 offset = 20; 3568 3569 /* 3570 * QP1 loopback packet has 4 bytes of internal header before 3571 * ether header. Skip these four bytes. 3572 */ 3573 if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf)) 3574 skip_bytes = 4; 3575 3576 /* First send SGE . Skip the ether header*/ 3577 s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE 3578 + skip_bytes; 3579 s_sge[0].lkey = 0xFFFFFFFF; 3580 s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 : 3581 BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6; 3582 3583 /* Second Send SGE */ 3584 s_sge[1].addr = s_sge[0].addr + s_sge[0].length + 3585 BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE; 3586 if (pkt_type != BNXT_RE_ROCE_V1_PACKET) 3587 s_sge[1].addr += 8; 3588 s_sge[1].lkey = 0xFFFFFFFF; 3589 s_sge[1].length = 256; 3590 3591 /* First recv SGE */ 3592 3593 r_sge[0].addr = shrq_hdr_buf_map; 3594 r_sge[0].lkey = 0xFFFFFFFF; 3595 r_sge[0].length = 40; 3596 3597 r_sge[1].addr = sqp_entry->sge.addr + offset; 3598 r_sge[1].lkey = sqp_entry->sge.lkey; 3599 r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset; 3600 3601 /* Create receive work request */ 3602 rwr.num_sge = 2; 3603 rwr.sg_list = r_sge; 3604 rwr.wr_id = tbl_idx; 3605 rwr.next = NULL; 3606 3607 rc = bnxt_re_post_recv_shadow_qp(rdev, gsi_sqp, &rwr); 3608 if (rc) { 3609 ibdev_err(&rdev->ibdev, 3610 "Failed to post Rx buffers to shadow QP"); 3611 return -ENOMEM; 3612 } 3613 3614 swr->num_sge = 2; 3615 swr->sg_list = s_sge; 3616 swr->wr_id = tbl_idx; 3617 swr->opcode = IB_WR_SEND; 3618 swr->next = NULL; 3619 gsi_sah = rdev->gsi_ctx.gsi_sah; 3620 udwr.ah = &gsi_sah->ib_ah; 3621 udwr.remote_qpn = gsi_sqp->qplib_qp.id; 3622 udwr.remote_qkey = gsi_sqp->qplib_qp.qkey; 3623 3624 /* post data received in the send queue */ 3625 return bnxt_re_post_send_shadow_qp(rdev, gsi_sqp, swr); 3626 } 3627 3628 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc, 3629 struct bnxt_qplib_cqe *cqe) 3630 { 3631 wc->opcode = IB_WC_RECV; 3632 wc->status = __rawqp1_to_ib_wc_status(cqe->status); 3633 wc->wc_flags |= IB_WC_GRH; 3634 } 3635 3636 static bool bnxt_re_check_if_vlan_valid(struct bnxt_re_dev *rdev, 3637 u16 vlan_id) 3638 { 3639 /* 3640 * Check if the vlan is configured in the host. If not configured, it 3641 * can be a transparent VLAN. So dont report the vlan id. 3642 */ 3643 if (!__vlan_find_dev_deep_rcu(rdev->netdev, 3644 htons(ETH_P_8021Q), vlan_id)) 3645 return false; 3646 return true; 3647 } 3648 3649 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe, 3650 u16 *vid, u8 *sl) 3651 { 3652 bool ret = false; 3653 u32 metadata; 3654 u16 tpid; 3655 3656 metadata = orig_cqe->raweth_qp1_metadata; 3657 if (orig_cqe->raweth_qp1_flags2 & 3658 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) { 3659 tpid = ((metadata & 3660 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >> 3661 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT); 3662 if (tpid == ETH_P_8021Q) { 3663 *vid = metadata & 3664 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK; 3665 *sl = (metadata & 3666 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >> 3667 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT; 3668 ret = true; 3669 } 3670 } 3671 3672 return ret; 3673 } 3674 3675 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc, 3676 struct bnxt_qplib_cqe *cqe) 3677 { 3678 wc->opcode = IB_WC_RECV; 3679 wc->status = __rc_to_ib_wc_status(cqe->status); 3680 3681 if (cqe->flags & CQ_RES_RC_FLAGS_IMM) 3682 wc->wc_flags |= IB_WC_WITH_IMM; 3683 if (cqe->flags & CQ_RES_RC_FLAGS_INV) 3684 wc->wc_flags |= IB_WC_WITH_INVALIDATE; 3685 if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) == 3686 (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) 3687 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM; 3688 } 3689 3690 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *gsi_sqp, 3691 struct ib_wc *wc, 3692 struct bnxt_qplib_cqe *cqe) 3693 { 3694 struct bnxt_re_dev *rdev = gsi_sqp->rdev; 3695 struct bnxt_re_qp *gsi_qp = NULL; 3696 struct bnxt_qplib_cqe *orig_cqe = NULL; 3697 struct bnxt_re_sqp_entries *sqp_entry = NULL; 3698 int nw_type; 3699 u32 tbl_idx; 3700 u16 vlan_id; 3701 u8 sl; 3702 3703 tbl_idx = cqe->wr_id; 3704 3705 sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx]; 3706 gsi_qp = sqp_entry->qp1_qp; 3707 orig_cqe = &sqp_entry->cqe; 3708 3709 wc->wr_id = sqp_entry->wrid; 3710 wc->byte_len = orig_cqe->length; 3711 wc->qp = &gsi_qp->ib_qp; 3712 3713 wc->ex.imm_data = cpu_to_be32(orig_cqe->immdata); 3714 wc->src_qp = orig_cqe->src_qp; 3715 memcpy(wc->smac, orig_cqe->smac, ETH_ALEN); 3716 if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) { 3717 if (bnxt_re_check_if_vlan_valid(rdev, vlan_id)) { 3718 wc->vlan_id = vlan_id; 3719 wc->sl = sl; 3720 wc->wc_flags |= IB_WC_WITH_VLAN; 3721 } 3722 } 3723 wc->port_num = 1; 3724 wc->vendor_err = orig_cqe->status; 3725 3726 wc->opcode = IB_WC_RECV; 3727 wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status); 3728 wc->wc_flags |= IB_WC_GRH; 3729 3730 nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags, 3731 orig_cqe->raweth_qp1_flags2); 3732 if (nw_type >= 0) { 3733 wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type); 3734 wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE; 3735 } 3736 } 3737 3738 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp, 3739 struct ib_wc *wc, 3740 struct bnxt_qplib_cqe *cqe) 3741 { 3742 struct bnxt_re_dev *rdev; 3743 u16 vlan_id = 0; 3744 u8 nw_type; 3745 3746 rdev = qp->rdev; 3747 wc->opcode = IB_WC_RECV; 3748 wc->status = __rc_to_ib_wc_status(cqe->status); 3749 3750 if (cqe->flags & CQ_RES_UD_FLAGS_IMM) 3751 wc->wc_flags |= IB_WC_WITH_IMM; 3752 /* report only on GSI QP for Thor */ 3753 if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) { 3754 wc->wc_flags |= IB_WC_GRH; 3755 memcpy(wc->smac, cqe->smac, ETH_ALEN); 3756 wc->wc_flags |= IB_WC_WITH_SMAC; 3757 if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) { 3758 vlan_id = (cqe->cfa_meta & 0xFFF); 3759 } 3760 /* Mark only if vlan_id is non zero */ 3761 if (vlan_id && bnxt_re_check_if_vlan_valid(rdev, vlan_id)) { 3762 wc->vlan_id = vlan_id; 3763 wc->wc_flags |= IB_WC_WITH_VLAN; 3764 } 3765 nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >> 3766 CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT; 3767 wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type); 3768 wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE; 3769 } 3770 3771 } 3772 3773 static int send_phantom_wqe(struct bnxt_re_qp *qp) 3774 { 3775 struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp; 3776 unsigned long flags; 3777 int rc; 3778 3779 spin_lock_irqsave(&qp->sq_lock, flags); 3780 3781 rc = bnxt_re_bind_fence_mw(lib_qp); 3782 if (!rc) { 3783 lib_qp->sq.phantom_wqe_cnt++; 3784 ibdev_dbg(&qp->rdev->ibdev, 3785 "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n", 3786 lib_qp->id, lib_qp->sq.hwq.prod, 3787 HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq), 3788 lib_qp->sq.phantom_wqe_cnt); 3789 } 3790 3791 spin_unlock_irqrestore(&qp->sq_lock, flags); 3792 return rc; 3793 } 3794 3795 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc) 3796 { 3797 struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq); 3798 struct bnxt_re_qp *qp, *sh_qp; 3799 struct bnxt_qplib_cqe *cqe; 3800 int i, ncqe, budget; 3801 struct bnxt_qplib_q *sq; 3802 struct bnxt_qplib_qp *lib_qp; 3803 u32 tbl_idx; 3804 struct bnxt_re_sqp_entries *sqp_entry = NULL; 3805 unsigned long flags; 3806 3807 /* User CQ; the only processing we do is to 3808 * complete any pending CQ resize operation. 3809 */ 3810 if (cq->umem) { 3811 if (cq->resize_umem) 3812 bnxt_re_resize_cq_complete(cq); 3813 return 0; 3814 } 3815 3816 spin_lock_irqsave(&cq->cq_lock, flags); 3817 budget = min_t(u32, num_entries, cq->max_cql); 3818 num_entries = budget; 3819 if (!cq->cql) { 3820 ibdev_err(&cq->rdev->ibdev, "POLL CQ : no CQL to use"); 3821 goto exit; 3822 } 3823 cqe = &cq->cql[0]; 3824 while (budget) { 3825 lib_qp = NULL; 3826 ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp); 3827 if (lib_qp) { 3828 sq = &lib_qp->sq; 3829 if (sq->send_phantom) { 3830 qp = container_of(lib_qp, 3831 struct bnxt_re_qp, qplib_qp); 3832 if (send_phantom_wqe(qp) == -ENOMEM) 3833 ibdev_err(&cq->rdev->ibdev, 3834 "Phantom failed! Scheduled to send again\n"); 3835 else 3836 sq->send_phantom = false; 3837 } 3838 } 3839 if (ncqe < budget) 3840 ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq, 3841 cqe + ncqe, 3842 budget - ncqe); 3843 3844 if (!ncqe) 3845 break; 3846 3847 for (i = 0; i < ncqe; i++, cqe++) { 3848 /* Transcribe each qplib_wqe back to ib_wc */ 3849 memset(wc, 0, sizeof(*wc)); 3850 3851 wc->wr_id = cqe->wr_id; 3852 wc->byte_len = cqe->length; 3853 qp = container_of 3854 ((struct bnxt_qplib_qp *) 3855 (unsigned long)(cqe->qp_handle), 3856 struct bnxt_re_qp, qplib_qp); 3857 wc->qp = &qp->ib_qp; 3858 if (cqe->flags & CQ_RES_RC_FLAGS_IMM) 3859 wc->ex.imm_data = cpu_to_be32(cqe->immdata); 3860 else 3861 wc->ex.invalidate_rkey = cqe->invrkey; 3862 wc->src_qp = cqe->src_qp; 3863 memcpy(wc->smac, cqe->smac, ETH_ALEN); 3864 wc->port_num = 1; 3865 wc->vendor_err = cqe->status; 3866 3867 switch (cqe->opcode) { 3868 case CQ_BASE_CQE_TYPE_REQ: 3869 sh_qp = qp->rdev->gsi_ctx.gsi_sqp; 3870 if (sh_qp && 3871 qp->qplib_qp.id == sh_qp->qplib_qp.id) { 3872 /* Handle this completion with 3873 * the stored completion 3874 */ 3875 memset(wc, 0, sizeof(*wc)); 3876 continue; 3877 } 3878 bnxt_re_process_req_wc(wc, cqe); 3879 break; 3880 case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1: 3881 if (!cqe->status) { 3882 int rc = 0; 3883 3884 rc = bnxt_re_process_raw_qp_pkt_rx 3885 (qp, cqe); 3886 if (!rc) { 3887 memset(wc, 0, sizeof(*wc)); 3888 continue; 3889 } 3890 cqe->status = -1; 3891 } 3892 /* Errors need not be looped back. 3893 * But change the wr_id to the one 3894 * stored in the table 3895 */ 3896 tbl_idx = cqe->wr_id; 3897 sqp_entry = &cq->rdev->gsi_ctx.sqp_tbl[tbl_idx]; 3898 wc->wr_id = sqp_entry->wrid; 3899 bnxt_re_process_res_rawqp1_wc(wc, cqe); 3900 break; 3901 case CQ_BASE_CQE_TYPE_RES_RC: 3902 bnxt_re_process_res_rc_wc(wc, cqe); 3903 break; 3904 case CQ_BASE_CQE_TYPE_RES_UD: 3905 sh_qp = qp->rdev->gsi_ctx.gsi_sqp; 3906 if (sh_qp && 3907 qp->qplib_qp.id == sh_qp->qplib_qp.id) { 3908 /* Handle this completion with 3909 * the stored completion 3910 */ 3911 if (cqe->status) { 3912 continue; 3913 } else { 3914 bnxt_re_process_res_shadow_qp_wc 3915 (qp, wc, cqe); 3916 break; 3917 } 3918 } 3919 bnxt_re_process_res_ud_wc(qp, wc, cqe); 3920 break; 3921 default: 3922 ibdev_err(&cq->rdev->ibdev, 3923 "POLL CQ : type 0x%x not handled", 3924 cqe->opcode); 3925 continue; 3926 } 3927 wc++; 3928 budget--; 3929 } 3930 } 3931 exit: 3932 spin_unlock_irqrestore(&cq->cq_lock, flags); 3933 return num_entries - budget; 3934 } 3935 3936 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq, 3937 enum ib_cq_notify_flags ib_cqn_flags) 3938 { 3939 struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq); 3940 int type = 0, rc = 0; 3941 unsigned long flags; 3942 3943 spin_lock_irqsave(&cq->cq_lock, flags); 3944 /* Trigger on the very next completion */ 3945 if (ib_cqn_flags & IB_CQ_NEXT_COMP) 3946 type = DBC_DBC_TYPE_CQ_ARMALL; 3947 /* Trigger on the next solicited completion */ 3948 else if (ib_cqn_flags & IB_CQ_SOLICITED) 3949 type = DBC_DBC_TYPE_CQ_ARMSE; 3950 3951 /* Poll to see if there are missed events */ 3952 if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) && 3953 !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) { 3954 rc = 1; 3955 goto exit; 3956 } 3957 bnxt_qplib_req_notify_cq(&cq->qplib_cq, type); 3958 3959 exit: 3960 spin_unlock_irqrestore(&cq->cq_lock, flags); 3961 return rc; 3962 } 3963 3964 /* Memory Regions */ 3965 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags) 3966 { 3967 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 3968 struct bnxt_re_dev *rdev = pd->rdev; 3969 struct bnxt_re_mr *mr; 3970 u32 active_mrs; 3971 int rc; 3972 3973 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 3974 if (!mr) 3975 return ERR_PTR(-ENOMEM); 3976 3977 mr->rdev = rdev; 3978 mr->qplib_mr.pd = &pd->qplib_pd; 3979 mr->qplib_mr.access_flags = __from_ib_access_flags(mr_access_flags); 3980 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR; 3981 3982 if (mr_access_flags & IB_ACCESS_RELAXED_ORDERING) 3983 bnxt_re_check_and_set_relaxed_ordering(rdev, &mr->qplib_mr); 3984 3985 /* Allocate and register 0 as the address */ 3986 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 3987 if (rc) 3988 goto fail; 3989 3990 mr->qplib_mr.hwq.level = PBL_LVL_MAX; 3991 mr->qplib_mr.total_size = -1; /* Infinte length */ 3992 rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL, 0, 3993 PAGE_SIZE); 3994 if (rc) 3995 goto fail_mr; 3996 3997 mr->ib_mr.lkey = mr->qplib_mr.lkey; 3998 if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ | 3999 IB_ACCESS_REMOTE_ATOMIC)) 4000 mr->ib_mr.rkey = mr->ib_mr.lkey; 4001 active_mrs = atomic_inc_return(&rdev->stats.res.mr_count); 4002 if (active_mrs > rdev->stats.res.mr_watermark) 4003 rdev->stats.res.mr_watermark = active_mrs; 4004 4005 return &mr->ib_mr; 4006 4007 fail_mr: 4008 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 4009 fail: 4010 kfree(mr); 4011 return ERR_PTR(rc); 4012 } 4013 4014 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata) 4015 { 4016 struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr); 4017 struct bnxt_re_dev *rdev = mr->rdev; 4018 int rc; 4019 4020 rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 4021 if (rc) { 4022 ibdev_err(&rdev->ibdev, "Dereg MR failed: %#x\n", rc); 4023 return rc; 4024 } 4025 4026 if (mr->pages) { 4027 rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res, 4028 &mr->qplib_frpl); 4029 kfree(mr->pages); 4030 mr->npages = 0; 4031 mr->pages = NULL; 4032 } 4033 ib_umem_release(mr->ib_umem); 4034 4035 kfree(mr); 4036 atomic_dec(&rdev->stats.res.mr_count); 4037 return rc; 4038 } 4039 4040 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr) 4041 { 4042 struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr); 4043 4044 if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs)) 4045 return -ENOMEM; 4046 4047 mr->pages[mr->npages++] = addr; 4048 return 0; 4049 } 4050 4051 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents, 4052 unsigned int *sg_offset) 4053 { 4054 struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr); 4055 4056 mr->npages = 0; 4057 return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page); 4058 } 4059 4060 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type, 4061 u32 max_num_sg) 4062 { 4063 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 4064 struct bnxt_re_dev *rdev = pd->rdev; 4065 struct bnxt_re_mr *mr = NULL; 4066 u32 active_mrs; 4067 int rc; 4068 4069 if (type != IB_MR_TYPE_MEM_REG) { 4070 ibdev_dbg(&rdev->ibdev, "MR type 0x%x not supported", type); 4071 return ERR_PTR(-EINVAL); 4072 } 4073 if (max_num_sg > MAX_PBL_LVL_1_PGS) 4074 return ERR_PTR(-EINVAL); 4075 4076 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 4077 if (!mr) 4078 return ERR_PTR(-ENOMEM); 4079 4080 mr->rdev = rdev; 4081 mr->qplib_mr.pd = &pd->qplib_pd; 4082 mr->qplib_mr.access_flags = BNXT_QPLIB_FR_PMR; 4083 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR; 4084 4085 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 4086 if (rc) 4087 goto bail; 4088 4089 mr->ib_mr.lkey = mr->qplib_mr.lkey; 4090 mr->ib_mr.rkey = mr->ib_mr.lkey; 4091 4092 mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL); 4093 if (!mr->pages) { 4094 rc = -ENOMEM; 4095 goto fail; 4096 } 4097 rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res, 4098 &mr->qplib_frpl, max_num_sg); 4099 if (rc) { 4100 ibdev_err(&rdev->ibdev, 4101 "Failed to allocate HW FR page list"); 4102 goto fail_mr; 4103 } 4104 4105 active_mrs = atomic_inc_return(&rdev->stats.res.mr_count); 4106 if (active_mrs > rdev->stats.res.mr_watermark) 4107 rdev->stats.res.mr_watermark = active_mrs; 4108 return &mr->ib_mr; 4109 4110 fail_mr: 4111 kfree(mr->pages); 4112 fail: 4113 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 4114 bail: 4115 kfree(mr); 4116 return ERR_PTR(rc); 4117 } 4118 4119 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type, 4120 struct ib_udata *udata) 4121 { 4122 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 4123 struct bnxt_re_dev *rdev = pd->rdev; 4124 struct bnxt_re_mw *mw; 4125 u32 active_mws; 4126 int rc; 4127 4128 mw = kzalloc(sizeof(*mw), GFP_KERNEL); 4129 if (!mw) 4130 return ERR_PTR(-ENOMEM); 4131 mw->rdev = rdev; 4132 mw->qplib_mw.pd = &pd->qplib_pd; 4133 4134 mw->qplib_mw.type = (type == IB_MW_TYPE_1 ? 4135 CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 : 4136 CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B); 4137 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw); 4138 if (rc) { 4139 ibdev_err(&rdev->ibdev, "Allocate MW failed!"); 4140 goto fail; 4141 } 4142 mw->ib_mw.rkey = mw->qplib_mw.rkey; 4143 4144 active_mws = atomic_inc_return(&rdev->stats.res.mw_count); 4145 if (active_mws > rdev->stats.res.mw_watermark) 4146 rdev->stats.res.mw_watermark = active_mws; 4147 return &mw->ib_mw; 4148 4149 fail: 4150 kfree(mw); 4151 return ERR_PTR(rc); 4152 } 4153 4154 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw) 4155 { 4156 struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw); 4157 struct bnxt_re_dev *rdev = mw->rdev; 4158 int rc; 4159 4160 rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw); 4161 if (rc) { 4162 ibdev_err(&rdev->ibdev, "Free MW failed: %#x\n", rc); 4163 return rc; 4164 } 4165 4166 kfree(mw); 4167 atomic_dec(&rdev->stats.res.mw_count); 4168 return rc; 4169 } 4170 4171 static struct ib_mr *__bnxt_re_user_reg_mr(struct ib_pd *ib_pd, u64 length, u64 virt_addr, 4172 int mr_access_flags, struct ib_umem *umem) 4173 { 4174 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 4175 struct bnxt_re_dev *rdev = pd->rdev; 4176 unsigned long page_size; 4177 struct bnxt_re_mr *mr; 4178 int umem_pgs, rc; 4179 u32 active_mrs; 4180 4181 if (length > BNXT_RE_MAX_MR_SIZE) { 4182 ibdev_err(&rdev->ibdev, "MR Size: %lld > Max supported:%lld\n", 4183 length, BNXT_RE_MAX_MR_SIZE); 4184 return ERR_PTR(-ENOMEM); 4185 } 4186 4187 page_size = ib_umem_find_best_pgsz(umem, BNXT_RE_PAGE_SIZE_SUPPORTED, virt_addr); 4188 if (!page_size) { 4189 ibdev_err(&rdev->ibdev, "umem page size unsupported!"); 4190 return ERR_PTR(-EINVAL); 4191 } 4192 4193 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 4194 if (!mr) 4195 return ERR_PTR(-ENOMEM); 4196 4197 mr->rdev = rdev; 4198 mr->qplib_mr.pd = &pd->qplib_pd; 4199 mr->qplib_mr.access_flags = __from_ib_access_flags(mr_access_flags); 4200 mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR; 4201 4202 if (!_is_alloc_mr_unified(rdev->dev_attr->dev_cap_flags)) { 4203 rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr); 4204 if (rc) { 4205 ibdev_err(&rdev->ibdev, "Failed to allocate MR rc = %d", rc); 4206 rc = -EIO; 4207 goto free_mr; 4208 } 4209 /* The fixed portion of the rkey is the same as the lkey */ 4210 mr->ib_mr.rkey = mr->qplib_mr.rkey; 4211 } else { 4212 mr->qplib_mr.flags = CMDQ_REGISTER_MR_FLAGS_ALLOC_MR; 4213 } 4214 mr->ib_umem = umem; 4215 mr->qplib_mr.va = virt_addr; 4216 mr->qplib_mr.total_size = length; 4217 4218 if (mr_access_flags & IB_ACCESS_RELAXED_ORDERING) 4219 bnxt_re_check_and_set_relaxed_ordering(rdev, &mr->qplib_mr); 4220 4221 umem_pgs = ib_umem_num_dma_blocks(umem, page_size); 4222 rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, umem, 4223 umem_pgs, page_size); 4224 if (rc) { 4225 ibdev_err(&rdev->ibdev, "Failed to register user MR - rc = %d\n", rc); 4226 rc = -EIO; 4227 goto free_mrw; 4228 } 4229 4230 mr->ib_mr.lkey = mr->qplib_mr.lkey; 4231 mr->ib_mr.rkey = mr->qplib_mr.lkey; 4232 active_mrs = atomic_inc_return(&rdev->stats.res.mr_count); 4233 if (active_mrs > rdev->stats.res.mr_watermark) 4234 rdev->stats.res.mr_watermark = active_mrs; 4235 4236 return &mr->ib_mr; 4237 4238 free_mrw: 4239 bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr); 4240 free_mr: 4241 kfree(mr); 4242 return ERR_PTR(rc); 4243 } 4244 4245 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length, 4246 u64 virt_addr, int mr_access_flags, 4247 struct ib_udata *udata) 4248 { 4249 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 4250 struct bnxt_re_dev *rdev = pd->rdev; 4251 struct ib_umem *umem; 4252 struct ib_mr *ib_mr; 4253 4254 umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags); 4255 if (IS_ERR(umem)) 4256 return ERR_CAST(umem); 4257 4258 ib_mr = __bnxt_re_user_reg_mr(ib_pd, length, virt_addr, mr_access_flags, umem); 4259 if (IS_ERR(ib_mr)) 4260 ib_umem_release(umem); 4261 return ib_mr; 4262 } 4263 4264 struct ib_mr *bnxt_re_reg_user_mr_dmabuf(struct ib_pd *ib_pd, u64 start, 4265 u64 length, u64 virt_addr, int fd, 4266 int mr_access_flags, 4267 struct uverbs_attr_bundle *attrs) 4268 { 4269 struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); 4270 struct bnxt_re_dev *rdev = pd->rdev; 4271 struct ib_umem_dmabuf *umem_dmabuf; 4272 struct ib_umem *umem; 4273 struct ib_mr *ib_mr; 4274 4275 umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, start, length, 4276 fd, mr_access_flags); 4277 if (IS_ERR(umem_dmabuf)) 4278 return ERR_CAST(umem_dmabuf); 4279 4280 umem = &umem_dmabuf->umem; 4281 4282 ib_mr = __bnxt_re_user_reg_mr(ib_pd, length, virt_addr, mr_access_flags, umem); 4283 if (IS_ERR(ib_mr)) 4284 ib_umem_release(umem); 4285 return ib_mr; 4286 } 4287 4288 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata) 4289 { 4290 struct ib_device *ibdev = ctx->device; 4291 struct bnxt_re_ucontext *uctx = 4292 container_of(ctx, struct bnxt_re_ucontext, ib_uctx); 4293 struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev); 4294 struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr; 4295 struct bnxt_re_user_mmap_entry *entry; 4296 struct bnxt_re_uctx_resp resp = {}; 4297 struct bnxt_re_uctx_req ureq = {}; 4298 u32 chip_met_rev_num = 0; 4299 int rc; 4300 4301 ibdev_dbg(ibdev, "ABI version requested %u", ibdev->ops.uverbs_abi_ver); 4302 4303 if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) { 4304 ibdev_dbg(ibdev, " is different from the device %d ", 4305 BNXT_RE_ABI_VERSION); 4306 return -EPERM; 4307 } 4308 4309 uctx->rdev = rdev; 4310 4311 uctx->shpg = (void *)__get_free_page(GFP_KERNEL); 4312 if (!uctx->shpg) { 4313 rc = -ENOMEM; 4314 goto fail; 4315 } 4316 spin_lock_init(&uctx->sh_lock); 4317 4318 resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX; 4319 chip_met_rev_num = rdev->chip_ctx->chip_num; 4320 chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_rev & 0xFF) << 4321 BNXT_RE_CHIP_ID0_CHIP_REV_SFT; 4322 chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_metal & 0xFF) << 4323 BNXT_RE_CHIP_ID0_CHIP_MET_SFT; 4324 resp.chip_id0 = chip_met_rev_num; 4325 /*Temp, Use xa_alloc instead */ 4326 resp.dev_id = rdev->en_dev->pdev->devfn; 4327 resp.max_qp = rdev->qplib_ctx.qpc_count; 4328 resp.pg_size = PAGE_SIZE; 4329 resp.cqe_sz = sizeof(struct cq_base); 4330 resp.max_cqd = dev_attr->max_cq_wqes; 4331 4332 if (rdev->chip_ctx->modes.db_push) 4333 resp.comp_mask |= BNXT_RE_UCNTX_CMASK_WC_DPI_ENABLED; 4334 4335 entry = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_SH_PAGE, NULL); 4336 if (!entry) { 4337 rc = -ENOMEM; 4338 goto cfail; 4339 } 4340 uctx->shpage_mmap = &entry->rdma_entry; 4341 if (rdev->pacing.dbr_pacing) 4342 resp.comp_mask |= BNXT_RE_UCNTX_CMASK_DBR_PACING_ENABLED; 4343 4344 if (_is_host_msn_table(rdev->qplib_res.dattr->dev_cap_flags2)) 4345 resp.comp_mask |= BNXT_RE_UCNTX_CMASK_MSN_TABLE_ENABLED; 4346 4347 if (udata->inlen >= sizeof(ureq)) { 4348 rc = ib_copy_from_udata(&ureq, udata, min(udata->inlen, sizeof(ureq))); 4349 if (rc) 4350 goto cfail; 4351 if (ureq.comp_mask & BNXT_RE_COMP_MASK_REQ_UCNTX_POW2_SUPPORT) { 4352 resp.comp_mask |= BNXT_RE_UCNTX_CMASK_POW2_DISABLED; 4353 uctx->cmask |= BNXT_RE_UCNTX_CAP_POW2_DISABLED; 4354 } 4355 if (ureq.comp_mask & BNXT_RE_COMP_MASK_REQ_UCNTX_VAR_WQE_SUPPORT) { 4356 resp.comp_mask |= BNXT_RE_UCNTX_CMASK_HAVE_MODE; 4357 resp.mode = rdev->chip_ctx->modes.wqe_mode; 4358 if (resp.mode == BNXT_QPLIB_WQE_MODE_VARIABLE) 4359 uctx->cmask |= BNXT_RE_UCNTX_CAP_VAR_WQE_ENABLED; 4360 } 4361 } 4362 4363 rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp))); 4364 if (rc) { 4365 ibdev_err(ibdev, "Failed to copy user context"); 4366 rc = -EFAULT; 4367 goto cfail; 4368 } 4369 4370 return 0; 4371 cfail: 4372 free_page((unsigned long)uctx->shpg); 4373 uctx->shpg = NULL; 4374 fail: 4375 return rc; 4376 } 4377 4378 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx) 4379 { 4380 struct bnxt_re_ucontext *uctx = container_of(ib_uctx, 4381 struct bnxt_re_ucontext, 4382 ib_uctx); 4383 4384 struct bnxt_re_dev *rdev = uctx->rdev; 4385 4386 rdma_user_mmap_entry_remove(uctx->shpage_mmap); 4387 uctx->shpage_mmap = NULL; 4388 if (uctx->shpg) 4389 free_page((unsigned long)uctx->shpg); 4390 4391 if (uctx->dpi.dbr) { 4392 /* Free DPI only if this is the first PD allocated by the 4393 * application and mark the context dpi as NULL 4394 */ 4395 bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->dpi); 4396 uctx->dpi.dbr = NULL; 4397 } 4398 } 4399 4400 static struct bnxt_re_cq *bnxt_re_search_for_cq(struct bnxt_re_dev *rdev, u32 cq_id) 4401 { 4402 struct bnxt_re_cq *cq = NULL, *tmp_cq; 4403 4404 hash_for_each_possible(rdev->cq_hash, tmp_cq, hash_entry, cq_id) { 4405 if (tmp_cq->qplib_cq.id == cq_id) { 4406 cq = tmp_cq; 4407 break; 4408 } 4409 } 4410 return cq; 4411 } 4412 4413 static struct bnxt_re_srq *bnxt_re_search_for_srq(struct bnxt_re_dev *rdev, u32 srq_id) 4414 { 4415 struct bnxt_re_srq *srq = NULL, *tmp_srq; 4416 4417 hash_for_each_possible(rdev->srq_hash, tmp_srq, hash_entry, srq_id) { 4418 if (tmp_srq->qplib_srq.id == srq_id) { 4419 srq = tmp_srq; 4420 break; 4421 } 4422 } 4423 return srq; 4424 } 4425 4426 /* Helper function to mmap the virtual memory from user app */ 4427 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma) 4428 { 4429 struct bnxt_re_ucontext *uctx = container_of(ib_uctx, 4430 struct bnxt_re_ucontext, 4431 ib_uctx); 4432 struct bnxt_re_user_mmap_entry *bnxt_entry; 4433 struct rdma_user_mmap_entry *rdma_entry; 4434 int ret = 0; 4435 u64 pfn; 4436 4437 rdma_entry = rdma_user_mmap_entry_get(&uctx->ib_uctx, vma); 4438 if (!rdma_entry) 4439 return -EINVAL; 4440 4441 bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry, 4442 rdma_entry); 4443 4444 switch (bnxt_entry->mmap_flag) { 4445 case BNXT_RE_MMAP_WC_DB: 4446 pfn = bnxt_entry->mem_offset >> PAGE_SHIFT; 4447 ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE, 4448 pgprot_writecombine(vma->vm_page_prot), 4449 rdma_entry); 4450 break; 4451 case BNXT_RE_MMAP_UC_DB: 4452 pfn = bnxt_entry->mem_offset >> PAGE_SHIFT; 4453 ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE, 4454 pgprot_noncached(vma->vm_page_prot), 4455 rdma_entry); 4456 break; 4457 case BNXT_RE_MMAP_SH_PAGE: 4458 ret = vm_insert_page(vma, vma->vm_start, virt_to_page(uctx->shpg)); 4459 break; 4460 case BNXT_RE_MMAP_DBR_BAR: 4461 pfn = bnxt_entry->mem_offset >> PAGE_SHIFT; 4462 ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE, 4463 pgprot_noncached(vma->vm_page_prot), 4464 rdma_entry); 4465 break; 4466 case BNXT_RE_MMAP_DBR_PAGE: 4467 case BNXT_RE_MMAP_TOGGLE_PAGE: 4468 /* Driver doesn't expect write access for user space */ 4469 if (vma->vm_flags & VM_WRITE) 4470 ret = -EFAULT; 4471 else 4472 ret = vm_insert_page(vma, vma->vm_start, 4473 virt_to_page((void *)bnxt_entry->mem_offset)); 4474 break; 4475 default: 4476 ret = -EINVAL; 4477 break; 4478 } 4479 4480 rdma_user_mmap_entry_put(rdma_entry); 4481 return ret; 4482 } 4483 4484 void bnxt_re_mmap_free(struct rdma_user_mmap_entry *rdma_entry) 4485 { 4486 struct bnxt_re_user_mmap_entry *bnxt_entry; 4487 4488 bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry, 4489 rdma_entry); 4490 4491 kfree(bnxt_entry); 4492 } 4493 4494 static int UVERBS_HANDLER(BNXT_RE_METHOD_NOTIFY_DRV)(struct uverbs_attr_bundle *attrs) 4495 { 4496 struct bnxt_re_ucontext *uctx; 4497 4498 uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx); 4499 bnxt_re_pacing_alert(uctx->rdev); 4500 return 0; 4501 } 4502 4503 static int UVERBS_HANDLER(BNXT_RE_METHOD_ALLOC_PAGE)(struct uverbs_attr_bundle *attrs) 4504 { 4505 struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_ALLOC_PAGE_HANDLE); 4506 enum bnxt_re_alloc_page_type alloc_type; 4507 struct bnxt_re_user_mmap_entry *entry; 4508 enum bnxt_re_mmap_flag mmap_flag; 4509 struct bnxt_qplib_chip_ctx *cctx; 4510 struct bnxt_re_ucontext *uctx; 4511 struct bnxt_re_dev *rdev; 4512 u64 mmap_offset; 4513 u32 length; 4514 u32 dpi; 4515 u64 addr; 4516 int err; 4517 4518 uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx); 4519 if (IS_ERR(uctx)) 4520 return PTR_ERR(uctx); 4521 4522 err = uverbs_get_const(&alloc_type, attrs, BNXT_RE_ALLOC_PAGE_TYPE); 4523 if (err) 4524 return err; 4525 4526 rdev = uctx->rdev; 4527 cctx = rdev->chip_ctx; 4528 4529 switch (alloc_type) { 4530 case BNXT_RE_ALLOC_WC_PAGE: 4531 if (cctx->modes.db_push) { 4532 if (bnxt_qplib_alloc_dpi(&rdev->qplib_res, &uctx->wcdpi, 4533 uctx, BNXT_QPLIB_DPI_TYPE_WC)) 4534 return -ENOMEM; 4535 length = PAGE_SIZE; 4536 dpi = uctx->wcdpi.dpi; 4537 addr = (u64)uctx->wcdpi.umdbr; 4538 mmap_flag = BNXT_RE_MMAP_WC_DB; 4539 } else { 4540 return -EINVAL; 4541 } 4542 4543 break; 4544 case BNXT_RE_ALLOC_DBR_BAR_PAGE: 4545 length = PAGE_SIZE; 4546 addr = (u64)rdev->pacing.dbr_bar_addr; 4547 mmap_flag = BNXT_RE_MMAP_DBR_BAR; 4548 break; 4549 4550 case BNXT_RE_ALLOC_DBR_PAGE: 4551 length = PAGE_SIZE; 4552 addr = (u64)rdev->pacing.dbr_page; 4553 mmap_flag = BNXT_RE_MMAP_DBR_PAGE; 4554 break; 4555 4556 default: 4557 return -EOPNOTSUPP; 4558 } 4559 4560 entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mmap_offset); 4561 if (!entry) 4562 return -ENOMEM; 4563 4564 uobj->object = entry; 4565 uverbs_finalize_uobj_create(attrs, BNXT_RE_ALLOC_PAGE_HANDLE); 4566 err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_OFFSET, 4567 &mmap_offset, sizeof(mmap_offset)); 4568 if (err) 4569 return err; 4570 4571 err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_LENGTH, 4572 &length, sizeof(length)); 4573 if (err) 4574 return err; 4575 4576 err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_DPI, 4577 &dpi, sizeof(length)); 4578 if (err) 4579 return err; 4580 4581 return 0; 4582 } 4583 4584 static int alloc_page_obj_cleanup(struct ib_uobject *uobject, 4585 enum rdma_remove_reason why, 4586 struct uverbs_attr_bundle *attrs) 4587 { 4588 struct bnxt_re_user_mmap_entry *entry = uobject->object; 4589 struct bnxt_re_ucontext *uctx = entry->uctx; 4590 4591 switch (entry->mmap_flag) { 4592 case BNXT_RE_MMAP_WC_DB: 4593 if (uctx && uctx->wcdpi.dbr) { 4594 struct bnxt_re_dev *rdev = uctx->rdev; 4595 4596 bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->wcdpi); 4597 uctx->wcdpi.dbr = NULL; 4598 } 4599 break; 4600 case BNXT_RE_MMAP_DBR_BAR: 4601 case BNXT_RE_MMAP_DBR_PAGE: 4602 break; 4603 default: 4604 goto exit; 4605 } 4606 rdma_user_mmap_entry_remove(&entry->rdma_entry); 4607 exit: 4608 return 0; 4609 } 4610 4611 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_ALLOC_PAGE, 4612 UVERBS_ATTR_IDR(BNXT_RE_ALLOC_PAGE_HANDLE, 4613 BNXT_RE_OBJECT_ALLOC_PAGE, 4614 UVERBS_ACCESS_NEW, 4615 UA_MANDATORY), 4616 UVERBS_ATTR_CONST_IN(BNXT_RE_ALLOC_PAGE_TYPE, 4617 enum bnxt_re_alloc_page_type, 4618 UA_MANDATORY), 4619 UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_OFFSET, 4620 UVERBS_ATTR_TYPE(u64), 4621 UA_MANDATORY), 4622 UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_LENGTH, 4623 UVERBS_ATTR_TYPE(u32), 4624 UA_MANDATORY), 4625 UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_DPI, 4626 UVERBS_ATTR_TYPE(u32), 4627 UA_MANDATORY)); 4628 4629 DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_DESTROY_PAGE, 4630 UVERBS_ATTR_IDR(BNXT_RE_DESTROY_PAGE_HANDLE, 4631 BNXT_RE_OBJECT_ALLOC_PAGE, 4632 UVERBS_ACCESS_DESTROY, 4633 UA_MANDATORY)); 4634 4635 DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_ALLOC_PAGE, 4636 UVERBS_TYPE_ALLOC_IDR(alloc_page_obj_cleanup), 4637 &UVERBS_METHOD(BNXT_RE_METHOD_ALLOC_PAGE), 4638 &UVERBS_METHOD(BNXT_RE_METHOD_DESTROY_PAGE)); 4639 4640 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_NOTIFY_DRV); 4641 4642 DECLARE_UVERBS_GLOBAL_METHODS(BNXT_RE_OBJECT_NOTIFY_DRV, 4643 &UVERBS_METHOD(BNXT_RE_METHOD_NOTIFY_DRV)); 4644 4645 /* Toggle MEM */ 4646 static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_TOGGLE_MEM)(struct uverbs_attr_bundle *attrs) 4647 { 4648 struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_TOGGLE_MEM_HANDLE); 4649 enum bnxt_re_mmap_flag mmap_flag = BNXT_RE_MMAP_TOGGLE_PAGE; 4650 enum bnxt_re_get_toggle_mem_type res_type; 4651 struct bnxt_re_user_mmap_entry *entry; 4652 struct bnxt_re_ucontext *uctx; 4653 struct ib_ucontext *ib_uctx; 4654 struct bnxt_re_dev *rdev; 4655 struct bnxt_re_srq *srq; 4656 u32 length = PAGE_SIZE; 4657 struct bnxt_re_cq *cq; 4658 u64 mem_offset; 4659 u32 offset = 0; 4660 u64 addr = 0; 4661 u32 res_id; 4662 int err; 4663 4664 ib_uctx = ib_uverbs_get_ucontext(attrs); 4665 if (IS_ERR(ib_uctx)) 4666 return PTR_ERR(ib_uctx); 4667 4668 err = uverbs_get_const(&res_type, attrs, BNXT_RE_TOGGLE_MEM_TYPE); 4669 if (err) 4670 return err; 4671 4672 uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx); 4673 rdev = uctx->rdev; 4674 err = uverbs_copy_from(&res_id, attrs, BNXT_RE_TOGGLE_MEM_RES_ID); 4675 if (err) 4676 return err; 4677 4678 switch (res_type) { 4679 case BNXT_RE_CQ_TOGGLE_MEM: 4680 cq = bnxt_re_search_for_cq(rdev, res_id); 4681 if (!cq) 4682 return -EINVAL; 4683 4684 addr = (u64)cq->uctx_cq_page; 4685 break; 4686 case BNXT_RE_SRQ_TOGGLE_MEM: 4687 srq = bnxt_re_search_for_srq(rdev, res_id); 4688 if (!srq) 4689 return -EINVAL; 4690 4691 addr = (u64)srq->uctx_srq_page; 4692 break; 4693 4694 default: 4695 return -EOPNOTSUPP; 4696 } 4697 4698 entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mem_offset); 4699 if (!entry) 4700 return -ENOMEM; 4701 4702 uobj->object = entry; 4703 uverbs_finalize_uobj_create(attrs, BNXT_RE_TOGGLE_MEM_HANDLE); 4704 err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_PAGE, 4705 &mem_offset, sizeof(mem_offset)); 4706 if (err) 4707 return err; 4708 4709 err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_LENGTH, 4710 &length, sizeof(length)); 4711 if (err) 4712 return err; 4713 4714 err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_OFFSET, 4715 &offset, sizeof(length)); 4716 if (err) 4717 return err; 4718 4719 return 0; 4720 } 4721 4722 static int get_toggle_mem_obj_cleanup(struct ib_uobject *uobject, 4723 enum rdma_remove_reason why, 4724 struct uverbs_attr_bundle *attrs) 4725 { 4726 struct bnxt_re_user_mmap_entry *entry = uobject->object; 4727 4728 rdma_user_mmap_entry_remove(&entry->rdma_entry); 4729 return 0; 4730 } 4731 4732 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_GET_TOGGLE_MEM, 4733 UVERBS_ATTR_IDR(BNXT_RE_TOGGLE_MEM_HANDLE, 4734 BNXT_RE_OBJECT_GET_TOGGLE_MEM, 4735 UVERBS_ACCESS_NEW, 4736 UA_MANDATORY), 4737 UVERBS_ATTR_CONST_IN(BNXT_RE_TOGGLE_MEM_TYPE, 4738 enum bnxt_re_get_toggle_mem_type, 4739 UA_MANDATORY), 4740 UVERBS_ATTR_PTR_IN(BNXT_RE_TOGGLE_MEM_RES_ID, 4741 UVERBS_ATTR_TYPE(u32), 4742 UA_MANDATORY), 4743 UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_PAGE, 4744 UVERBS_ATTR_TYPE(u64), 4745 UA_MANDATORY), 4746 UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_OFFSET, 4747 UVERBS_ATTR_TYPE(u32), 4748 UA_MANDATORY), 4749 UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_LENGTH, 4750 UVERBS_ATTR_TYPE(u32), 4751 UA_MANDATORY)); 4752 4753 DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_RELEASE_TOGGLE_MEM, 4754 UVERBS_ATTR_IDR(BNXT_RE_RELEASE_TOGGLE_MEM_HANDLE, 4755 BNXT_RE_OBJECT_GET_TOGGLE_MEM, 4756 UVERBS_ACCESS_DESTROY, 4757 UA_MANDATORY)); 4758 4759 DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_GET_TOGGLE_MEM, 4760 UVERBS_TYPE_ALLOC_IDR(get_toggle_mem_obj_cleanup), 4761 &UVERBS_METHOD(BNXT_RE_METHOD_GET_TOGGLE_MEM), 4762 &UVERBS_METHOD(BNXT_RE_METHOD_RELEASE_TOGGLE_MEM)); 4763 4764 const struct uapi_definition bnxt_re_uapi_defs[] = { 4765 UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE), 4766 UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_NOTIFY_DRV), 4767 UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_GET_TOGGLE_MEM), 4768 {} 4769 }; 4770