1 /*******************************************************************
2 * This file is part of the Emulex RoCE Device Driver for *
3 * RoCE (RDMA over Converged Ethernet) adapters. *
4 * Copyright (C) 2008-2012 Emulex. All rights reserved. *
5 * EMULEX and SLI are trademarks of Emulex. *
8 * This program is free software; you can redistribute it and/or *
9 * modify it under the terms of version 2 of the GNU General *
10 * Public License as published by the Free Software Foundation. *
11 * This program is distributed in the hope that it will be useful. *
12 * ALL EXPRESS OR IMPLIED CONDITIONS, REPRESENTATIONS AND *
13 * WARRANTIES, INCLUDING ANY IMPLIED WARRANTY OF MERCHANTABILITY, *
14 * FITNESS FOR A PARTICULAR PURPOSE, OR NON-INFRINGEMENT, ARE *
15 * DISCLAIMED, EXCEPT TO THE EXTENT THAT SUCH DISCLAIMERS ARE HELD *
16 * TO BE LEGALLY INVALID. See the GNU General Public License for *
17 * more details, a copy of which can be found in the file COPYING *
18 * included with this package. *
20 * Contact Information:
21 * linux-drivers@emulex.com
25 * Costa Mesa, CA 92626
26 *******************************************************************/
28 #include <linux/dma-mapping.h>
29 #include <rdma/ib_verbs.h>
30 #include <rdma/ib_user_verbs.h>
31 #include <rdma/iw_cm.h>
32 #include <rdma/ib_umem.h>
33 #include <rdma/ib_addr.h>
36 #include "ocrdma_hw.h"
37 #include "ocrdma_verbs.h"
38 #include "ocrdma_abi.h"
40 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
49 int ocrdma_query_gid(struct ib_device *ibdev, u8 port,
50 int index, union ib_gid *sgid)
52 struct ocrdma_dev *dev;
54 dev = get_ocrdma_dev(ibdev);
55 memset(sgid, 0, sizeof(*sgid));
56 if (index > OCRDMA_MAX_SGID)
59 memcpy(sgid, &dev->sgid_tbl[index], sizeof(*sgid));
64 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr)
66 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
68 memset(attr, 0, sizeof *attr);
69 memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
70 min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
71 ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
72 attr->max_mr_size = dev->attr.max_mr_size;
73 attr->page_size_cap = 0xffff000;
74 attr->vendor_id = dev->nic_info.pdev->vendor;
75 attr->vendor_part_id = dev->nic_info.pdev->device;
76 attr->hw_ver = dev->asic_id;
77 attr->max_qp = dev->attr.max_qp;
78 attr->max_ah = OCRDMA_MAX_AH;
79 attr->max_qp_wr = dev->attr.max_wqe;
81 attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
82 IB_DEVICE_RC_RNR_NAK_GEN |
83 IB_DEVICE_SHUTDOWN_PORT |
84 IB_DEVICE_SYS_IMAGE_GUID |
85 IB_DEVICE_LOCAL_DMA_LKEY |
86 IB_DEVICE_MEM_MGT_EXTENSIONS;
87 attr->max_sge = min(dev->attr.max_send_sge, dev->attr.max_srq_sge);
89 attr->max_cq = dev->attr.max_cq;
90 attr->max_cqe = dev->attr.max_cqe;
91 attr->max_mr = dev->attr.max_mr;
92 attr->max_mw = dev->attr.max_mw;
93 attr->max_pd = dev->attr.max_pd;
96 attr->max_map_per_fmr = 0;
97 attr->max_qp_rd_atom =
98 min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
99 attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
100 attr->max_srq = dev->attr.max_srq;
101 attr->max_srq_sge = dev->attr.max_srq_sge;
102 attr->max_srq_wr = dev->attr.max_rqe;
103 attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
104 attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
109 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
110 u8 *ib_speed, u8 *ib_width)
115 status = ocrdma_mbx_get_link_speed(dev, &speed);
117 speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
120 case OCRDMA_PHYS_LINK_SPEED_1GBPS:
121 *ib_speed = IB_SPEED_SDR;
122 *ib_width = IB_WIDTH_1X;
125 case OCRDMA_PHYS_LINK_SPEED_10GBPS:
126 *ib_speed = IB_SPEED_QDR;
127 *ib_width = IB_WIDTH_1X;
130 case OCRDMA_PHYS_LINK_SPEED_20GBPS:
131 *ib_speed = IB_SPEED_DDR;
132 *ib_width = IB_WIDTH_4X;
135 case OCRDMA_PHYS_LINK_SPEED_40GBPS:
136 *ib_speed = IB_SPEED_QDR;
137 *ib_width = IB_WIDTH_4X;
142 *ib_speed = IB_SPEED_SDR;
143 *ib_width = IB_WIDTH_1X;
147 int ocrdma_query_port(struct ib_device *ibdev,
148 u8 port, struct ib_port_attr *props)
150 enum ib_port_state port_state;
151 struct ocrdma_dev *dev;
152 struct net_device *netdev;
154 dev = get_ocrdma_dev(ibdev);
156 pr_err("%s(%d) invalid_port=0x%x\n", __func__,
160 netdev = dev->nic_info.netdev;
161 if (netif_running(netdev) && netif_oper_up(netdev)) {
162 port_state = IB_PORT_ACTIVE;
163 props->phys_state = 5;
165 port_state = IB_PORT_DOWN;
166 props->phys_state = 3;
168 props->max_mtu = IB_MTU_4096;
169 props->active_mtu = iboe_get_mtu(netdev->mtu);
174 props->state = port_state;
175 props->port_cap_flags =
178 IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP | IB_PORT_IP_BASED_GIDS;
179 props->gid_tbl_len = OCRDMA_MAX_SGID;
180 props->pkey_tbl_len = 1;
181 props->bad_pkey_cntr = 0;
182 props->qkey_viol_cntr = 0;
183 get_link_speed_and_width(dev, &props->active_speed,
184 &props->active_width);
185 props->max_msg_sz = 0x80000000;
186 props->max_vl_num = 4;
190 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
191 struct ib_port_modify *props)
193 struct ocrdma_dev *dev;
195 dev = get_ocrdma_dev(ibdev);
197 pr_err("%s(%d) invalid_port=0x%x\n", __func__, dev->id, port);
203 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
206 struct ocrdma_mm *mm;
208 mm = kzalloc(sizeof(*mm), GFP_KERNEL);
211 mm->key.phy_addr = phy_addr;
213 INIT_LIST_HEAD(&mm->entry);
215 mutex_lock(&uctx->mm_list_lock);
216 list_add_tail(&mm->entry, &uctx->mm_head);
217 mutex_unlock(&uctx->mm_list_lock);
221 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
224 struct ocrdma_mm *mm, *tmp;
226 mutex_lock(&uctx->mm_list_lock);
227 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
228 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
231 list_del(&mm->entry);
235 mutex_unlock(&uctx->mm_list_lock);
238 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
242 struct ocrdma_mm *mm;
244 mutex_lock(&uctx->mm_list_lock);
245 list_for_each_entry(mm, &uctx->mm_head, entry) {
246 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
252 mutex_unlock(&uctx->mm_list_lock);
256 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
257 struct ocrdma_ucontext *uctx,
258 struct ib_udata *udata)
260 struct ocrdma_pd *pd = NULL;
263 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
265 return ERR_PTR(-ENOMEM);
269 ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
271 pd->dpp_enabled ? (dev->nic_info.db_page_size /
272 dev->attr.wqe_size) : 0;
276 status = ocrdma_mbx_alloc_pd(dev, pd);
278 if (pd->dpp_enabled) {
279 pd->dpp_enabled = false;
284 return ERR_PTR(status);
291 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
292 struct ocrdma_pd *pd)
294 return (uctx->cntxt_pd == pd ? true : false);
297 static int _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
298 struct ocrdma_pd *pd)
302 status = ocrdma_mbx_dealloc_pd(dev, pd);
307 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
308 struct ocrdma_ucontext *uctx,
309 struct ib_udata *udata)
313 uctx->cntxt_pd = _ocrdma_alloc_pd(dev, uctx, udata);
314 if (IS_ERR(uctx->cntxt_pd)) {
315 status = PTR_ERR(uctx->cntxt_pd);
316 uctx->cntxt_pd = NULL;
320 uctx->cntxt_pd->uctx = uctx;
321 uctx->cntxt_pd->ibpd.device = &dev->ibdev;
326 static int ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
329 struct ocrdma_pd *pd = uctx->cntxt_pd;
330 struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
332 if (uctx->pd_in_use) {
333 pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
334 __func__, dev->id, pd->id);
336 uctx->cntxt_pd = NULL;
337 status = _ocrdma_dealloc_pd(dev, pd);
341 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
343 struct ocrdma_pd *pd = NULL;
345 mutex_lock(&uctx->mm_list_lock);
346 if (!uctx->pd_in_use) {
347 uctx->pd_in_use = true;
350 mutex_unlock(&uctx->mm_list_lock);
355 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
357 mutex_lock(&uctx->mm_list_lock);
358 uctx->pd_in_use = false;
359 mutex_unlock(&uctx->mm_list_lock);
362 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
363 struct ib_udata *udata)
366 struct ocrdma_ucontext *ctx;
367 struct ocrdma_alloc_ucontext_resp resp;
368 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
369 struct pci_dev *pdev = dev->nic_info.pdev;
370 u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
373 return ERR_PTR(-EFAULT);
374 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
376 return ERR_PTR(-ENOMEM);
377 INIT_LIST_HEAD(&ctx->mm_head);
378 mutex_init(&ctx->mm_list_lock);
380 ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
381 &ctx->ah_tbl.pa, GFP_KERNEL);
382 if (!ctx->ah_tbl.va) {
384 return ERR_PTR(-ENOMEM);
386 memset(ctx->ah_tbl.va, 0, map_len);
387 ctx->ah_tbl.len = map_len;
389 memset(&resp, 0, sizeof(resp));
390 resp.ah_tbl_len = ctx->ah_tbl.len;
391 resp.ah_tbl_page = ctx->ah_tbl.pa;
393 status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
397 status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
401 resp.dev_id = dev->id;
402 resp.max_inline_data = dev->attr.max_inline_data;
403 resp.wqe_size = dev->attr.wqe_size;
404 resp.rqe_size = dev->attr.rqe_size;
405 resp.dpp_wqe_size = dev->attr.wqe_size;
407 memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
408 status = ib_copy_to_udata(udata, &resp, sizeof(resp));
411 return &ctx->ibucontext;
415 ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
417 dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
420 return ERR_PTR(status);
423 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
426 struct ocrdma_mm *mm, *tmp;
427 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
428 struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
429 struct pci_dev *pdev = dev->nic_info.pdev;
431 status = ocrdma_dealloc_ucontext_pd(uctx);
433 ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
434 dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
437 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
438 list_del(&mm->entry);
445 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
447 struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
448 struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
449 unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
450 u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
451 unsigned long len = (vma->vm_end - vma->vm_start);
455 if (vma->vm_start & (PAGE_SIZE - 1))
457 found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
461 if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
462 dev->nic_info.db_total_size)) &&
463 (len <= dev->nic_info.db_page_size)) {
464 if (vma->vm_flags & VM_READ)
467 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
468 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
469 len, vma->vm_page_prot);
470 } else if (dev->nic_info.dpp_unmapped_len &&
471 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
472 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
473 dev->nic_info.dpp_unmapped_len)) &&
474 (len <= dev->nic_info.dpp_unmapped_len)) {
475 if (vma->vm_flags & VM_READ)
478 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
479 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
480 len, vma->vm_page_prot);
482 status = remap_pfn_range(vma, vma->vm_start,
483 vma->vm_pgoff, len, vma->vm_page_prot);
488 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
489 struct ib_ucontext *ib_ctx,
490 struct ib_udata *udata)
494 u64 dpp_page_addr = 0;
496 struct ocrdma_alloc_pd_uresp rsp;
497 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
499 memset(&rsp, 0, sizeof(rsp));
501 rsp.dpp_enabled = pd->dpp_enabled;
502 db_page_addr = ocrdma_get_db_addr(dev, pd->id);
503 db_page_size = dev->nic_info.db_page_size;
505 status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
509 if (pd->dpp_enabled) {
510 dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
511 (pd->id * PAGE_SIZE);
512 status = ocrdma_add_mmap(uctx, dpp_page_addr,
516 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
517 rsp.dpp_page_addr_lo = dpp_page_addr;
520 status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
529 ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
531 ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
535 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
536 struct ib_ucontext *context,
537 struct ib_udata *udata)
539 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
540 struct ocrdma_pd *pd;
541 struct ocrdma_ucontext *uctx = NULL;
543 u8 is_uctx_pd = false;
545 if (udata && context) {
546 uctx = get_ocrdma_ucontext(context);
547 pd = ocrdma_get_ucontext_pd(uctx);
554 pd = _ocrdma_alloc_pd(dev, uctx, udata);
556 status = PTR_ERR(pd);
561 if (udata && context) {
562 status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
570 ocrdma_release_ucontext_pd(uctx);
572 status = ocrdma_mbx_dealloc_pd(dev, pd);
576 return ERR_PTR(status);
579 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
581 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
582 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
583 struct ocrdma_ucontext *uctx = NULL;
589 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
590 (pd->id * PAGE_SIZE);
592 ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
593 usr_db = ocrdma_get_db_addr(dev, pd->id);
594 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
596 if (is_ucontext_pd(uctx, pd)) {
597 ocrdma_release_ucontext_pd(uctx);
601 status = _ocrdma_dealloc_pd(dev, pd);
605 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
606 u32 pdid, int acc, u32 num_pbls, u32 addr_check)
611 mr->hwmr.local_rd = 1;
612 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
613 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
614 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
615 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
616 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
617 mr->hwmr.num_pbls = num_pbls;
619 status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
623 mr->ibmr.lkey = mr->hwmr.lkey;
624 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
625 mr->ibmr.rkey = mr->hwmr.lkey;
629 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
632 struct ocrdma_mr *mr;
633 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
634 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
636 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
637 pr_err("%s err, invalid access rights\n", __func__);
638 return ERR_PTR(-EINVAL);
641 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
643 return ERR_PTR(-ENOMEM);
645 status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
646 OCRDMA_ADDR_CHECK_DISABLE);
649 return ERR_PTR(status);
655 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
656 struct ocrdma_hw_mr *mr)
658 struct pci_dev *pdev = dev->nic_info.pdev;
662 for (i = 0; i < mr->num_pbls; i++) {
663 if (!mr->pbl_table[i].va)
665 dma_free_coherent(&pdev->dev, mr->pbl_size,
667 mr->pbl_table[i].pa);
669 kfree(mr->pbl_table);
670 mr->pbl_table = NULL;
674 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
683 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
684 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
688 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
689 num_pbls = num_pbls / (pbl_size / sizeof(u64));
691 } while (num_pbls >= dev->attr.max_num_mr_pbl);
693 mr->hwmr.num_pbes = num_pbes;
694 mr->hwmr.num_pbls = num_pbls;
695 mr->hwmr.pbl_size = pbl_size;
699 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
703 u32 dma_len = mr->pbl_size;
704 struct pci_dev *pdev = dev->nic_info.pdev;
708 mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
709 mr->num_pbls, GFP_KERNEL);
714 for (i = 0; i < mr->num_pbls; i++) {
715 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
717 ocrdma_free_mr_pbl_tbl(dev, mr);
721 memset(va, 0, dma_len);
722 mr->pbl_table[i].va = va;
723 mr->pbl_table[i].pa = pa;
728 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
731 struct ocrdma_pbe *pbe;
732 struct scatterlist *sg;
733 struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
734 struct ib_umem *umem = mr->umem;
735 int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
737 if (!mr->hwmr.num_pbes)
740 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
743 shift = ilog2(umem->page_size);
745 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
746 pages = sg_dma_len(sg) >> shift;
747 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
748 /* store the page address in pbe */
750 cpu_to_le32(sg_dma_address
752 (umem->page_size * pg_cnt));
754 cpu_to_le32(upper_32_bits
757 umem->page_size * pg_cnt)));
762 /* if done building pbes, issue the mbx cmd. */
763 if (total_num_pbes == num_pbes)
766 /* if the given pbl is full storing the pbes,
770 (mr->hwmr.pbl_size / sizeof(u64))) {
772 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
780 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
781 u64 usr_addr, int acc, struct ib_udata *udata)
783 int status = -ENOMEM;
784 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
785 struct ocrdma_mr *mr;
786 struct ocrdma_pd *pd;
789 pd = get_ocrdma_pd(ibpd);
791 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
792 return ERR_PTR(-EINVAL);
794 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
796 return ERR_PTR(status);
797 mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
798 if (IS_ERR(mr->umem)) {
802 num_pbes = ib_umem_page_count(mr->umem);
803 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
807 mr->hwmr.pbe_size = mr->umem->page_size;
808 mr->hwmr.fbo = mr->umem->offset;
809 mr->hwmr.va = usr_addr;
811 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
812 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
813 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
814 mr->hwmr.local_rd = 1;
815 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
816 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
819 build_user_pbes(dev, mr, num_pbes);
820 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
823 mr->ibmr.lkey = mr->hwmr.lkey;
824 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
825 mr->ibmr.rkey = mr->hwmr.lkey;
830 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
833 return ERR_PTR(status);
836 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
838 struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
839 struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
842 status = ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
844 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
846 /* it could be user registered memory. */
848 ib_umem_release(mr->umem);
851 /* Don't stop cleanup, in case FW is unresponsive */
852 if (dev->mqe_ctx.fw_error_state) {
854 pr_err("%s(%d) fw not responding.\n",
860 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
861 struct ib_udata *udata,
862 struct ib_ucontext *ib_ctx)
865 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
866 struct ocrdma_create_cq_uresp uresp;
868 memset(&uresp, 0, sizeof(uresp));
869 uresp.cq_id = cq->id;
870 uresp.page_size = PAGE_ALIGN(cq->len);
872 uresp.max_hw_cqe = cq->max_hw_cqe;
873 uresp.page_addr[0] = cq->pa;
874 uresp.db_page_addr = ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
875 uresp.db_page_size = dev->nic_info.db_page_size;
876 uresp.phase_change = cq->phase_change ? 1 : 0;
877 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
879 pr_err("%s(%d) copy error cqid=0x%x.\n",
880 __func__, dev->id, cq->id);
883 status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
886 status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
888 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
896 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev, int entries, int vector,
897 struct ib_ucontext *ib_ctx,
898 struct ib_udata *udata)
900 struct ocrdma_cq *cq;
901 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
902 struct ocrdma_ucontext *uctx = NULL;
905 struct ocrdma_create_cq_ureq ureq;
908 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
909 return ERR_PTR(-EFAULT);
912 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
914 return ERR_PTR(-ENOMEM);
916 spin_lock_init(&cq->cq_lock);
917 spin_lock_init(&cq->comp_handler_lock);
918 INIT_LIST_HEAD(&cq->sq_head);
919 INIT_LIST_HEAD(&cq->rq_head);
920 cq->first_arm = true;
923 uctx = get_ocrdma_ucontext(ib_ctx);
924 pd_id = uctx->cntxt_pd->id;
927 status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
930 return ERR_PTR(status);
933 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
937 cq->phase = OCRDMA_CQE_VALID;
938 dev->cq_tbl[cq->id] = cq;
942 ocrdma_mbx_destroy_cq(dev, cq);
944 return ERR_PTR(status);
947 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
948 struct ib_udata *udata)
951 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
953 if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
961 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
967 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
968 struct ocrdma_cqe *cqe = NULL;
971 cqe_cnt = cq->cqe_cnt;
973 /* Last irq might have scheduled a polling thread
974 * sync-up with it before hard flushing.
976 spin_lock_irqsave(&cq->cq_lock, flags);
978 if (is_cqe_valid(cq, cqe))
983 ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
984 spin_unlock_irqrestore(&cq->cq_lock, flags);
987 int ocrdma_destroy_cq(struct ib_cq *ibcq)
990 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
991 struct ocrdma_eq *eq = NULL;
992 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
996 dev->cq_tbl[cq->id] = NULL;
997 indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1001 eq = &dev->eq_tbl[indx];
1002 irq = ocrdma_get_irq(dev, eq);
1003 synchronize_irq(irq);
1004 ocrdma_flush_cq(cq);
1006 status = ocrdma_mbx_destroy_cq(dev, cq);
1008 pdid = cq->ucontext->cntxt_pd->id;
1009 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1010 PAGE_ALIGN(cq->len));
1011 ocrdma_del_mmap(cq->ucontext,
1012 ocrdma_get_db_addr(dev, pdid),
1013 dev->nic_info.db_page_size);
1020 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1022 int status = -EINVAL;
1024 if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1025 dev->qp_tbl[qp->id] = qp;
1031 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1033 dev->qp_tbl[qp->id] = NULL;
1036 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1037 struct ib_qp_init_attr *attrs)
1039 if ((attrs->qp_type != IB_QPT_GSI) &&
1040 (attrs->qp_type != IB_QPT_RC) &&
1041 (attrs->qp_type != IB_QPT_UC) &&
1042 (attrs->qp_type != IB_QPT_UD)) {
1043 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1044 __func__, dev->id, attrs->qp_type);
1047 /* Skip the check for QP1 to support CM size of 128 */
1048 if ((attrs->qp_type != IB_QPT_GSI) &&
1049 (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1050 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1051 __func__, dev->id, attrs->cap.max_send_wr);
1052 pr_err("%s(%d) supported send_wr=0x%x\n",
1053 __func__, dev->id, dev->attr.max_wqe);
1056 if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1057 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1058 __func__, dev->id, attrs->cap.max_recv_wr);
1059 pr_err("%s(%d) supported recv_wr=0x%x\n",
1060 __func__, dev->id, dev->attr.max_rqe);
1063 if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1064 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1065 __func__, dev->id, attrs->cap.max_inline_data);
1066 pr_err("%s(%d) supported inline data size=0x%x\n",
1067 __func__, dev->id, dev->attr.max_inline_data);
1070 if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1071 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1072 __func__, dev->id, attrs->cap.max_send_sge);
1073 pr_err("%s(%d) supported send_sge=0x%x\n",
1074 __func__, dev->id, dev->attr.max_send_sge);
1077 if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1078 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1079 __func__, dev->id, attrs->cap.max_recv_sge);
1080 pr_err("%s(%d) supported recv_sge=0x%x\n",
1081 __func__, dev->id, dev->attr.max_recv_sge);
1084 /* unprivileged user space cannot create special QP */
1085 if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1087 ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1088 __func__, dev->id, attrs->qp_type);
1091 /* allow creating only one GSI type of QP */
1092 if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1093 pr_err("%s(%d) GSI special QPs already created.\n",
1097 /* verify consumer QPs are not trying to use GSI QP's CQ */
1098 if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1099 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1100 (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1101 pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1109 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1110 struct ib_udata *udata, int dpp_offset,
1111 int dpp_credit_lmt, int srq)
1115 struct ocrdma_create_qp_uresp uresp;
1116 struct ocrdma_dev *dev = qp->dev;
1117 struct ocrdma_pd *pd = qp->pd;
1119 memset(&uresp, 0, sizeof(uresp));
1120 usr_db = dev->nic_info.unmapped_db +
1121 (pd->id * dev->nic_info.db_page_size);
1122 uresp.qp_id = qp->id;
1123 uresp.sq_dbid = qp->sq.dbid;
1124 uresp.num_sq_pages = 1;
1125 uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1126 uresp.sq_page_addr[0] = qp->sq.pa;
1127 uresp.num_wqe_allocated = qp->sq.max_cnt;
1129 uresp.rq_dbid = qp->rq.dbid;
1130 uresp.num_rq_pages = 1;
1131 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1132 uresp.rq_page_addr[0] = qp->rq.pa;
1133 uresp.num_rqe_allocated = qp->rq.max_cnt;
1135 uresp.db_page_addr = usr_db;
1136 uresp.db_page_size = dev->nic_info.db_page_size;
1137 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1138 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1139 uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1141 if (qp->dpp_enabled) {
1142 uresp.dpp_credit = dpp_credit_lmt;
1143 uresp.dpp_offset = dpp_offset;
1145 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1147 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1150 status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1151 uresp.sq_page_size);
1156 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1157 uresp.rq_page_size);
1163 ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1168 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1169 struct ocrdma_pd *pd)
1171 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1172 qp->sq_db = dev->nic_info.db +
1173 (pd->id * dev->nic_info.db_page_size) +
1174 OCRDMA_DB_GEN2_SQ_OFFSET;
1175 qp->rq_db = dev->nic_info.db +
1176 (pd->id * dev->nic_info.db_page_size) +
1177 OCRDMA_DB_GEN2_RQ_OFFSET;
1179 qp->sq_db = dev->nic_info.db +
1180 (pd->id * dev->nic_info.db_page_size) +
1181 OCRDMA_DB_SQ_OFFSET;
1182 qp->rq_db = dev->nic_info.db +
1183 (pd->id * dev->nic_info.db_page_size) +
1184 OCRDMA_DB_RQ_OFFSET;
1188 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1191 kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1193 if (qp->wqe_wr_id_tbl == NULL)
1196 kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1197 if (qp->rqe_wr_id_tbl == NULL)
1203 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1204 struct ocrdma_pd *pd,
1205 struct ib_qp_init_attr *attrs)
1208 spin_lock_init(&qp->q_lock);
1209 INIT_LIST_HEAD(&qp->sq_entry);
1210 INIT_LIST_HEAD(&qp->rq_entry);
1212 qp->qp_type = attrs->qp_type;
1213 qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1214 qp->max_inline_data = attrs->cap.max_inline_data;
1215 qp->sq.max_sges = attrs->cap.max_send_sge;
1216 qp->rq.max_sges = attrs->cap.max_recv_sge;
1217 qp->state = OCRDMA_QPS_RST;
1218 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1221 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1222 struct ib_qp_init_attr *attrs)
1224 if (attrs->qp_type == IB_QPT_GSI) {
1225 dev->gsi_qp_created = 1;
1226 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1227 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1231 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1232 struct ib_qp_init_attr *attrs,
1233 struct ib_udata *udata)
1236 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1237 struct ocrdma_qp *qp;
1238 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1239 struct ocrdma_create_qp_ureq ureq;
1240 u16 dpp_credit_lmt, dpp_offset;
1242 status = ocrdma_check_qp_params(ibpd, dev, attrs);
1246 memset(&ureq, 0, sizeof(ureq));
1248 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1249 return ERR_PTR(-EFAULT);
1251 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1257 ocrdma_set_qp_init_params(qp, pd, attrs);
1259 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1260 OCRDMA_QP_FAST_REG);
1262 mutex_lock(&dev->dev_lock);
1263 status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1265 &dpp_offset, &dpp_credit_lmt);
1269 /* user space QP's wr_id table are managed in library */
1270 if (udata == NULL) {
1271 status = ocrdma_alloc_wr_id_tbl(qp);
1276 status = ocrdma_add_qpn_map(dev, qp);
1279 ocrdma_set_qp_db(dev, qp, pd);
1281 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1283 (attrs->srq != NULL));
1287 ocrdma_store_gsi_qp_cq(dev, attrs);
1288 qp->ibqp.qp_num = qp->id;
1289 mutex_unlock(&dev->dev_lock);
1293 ocrdma_del_qpn_map(dev, qp);
1295 ocrdma_mbx_destroy_qp(dev, qp);
1297 mutex_unlock(&dev->dev_lock);
1298 kfree(qp->wqe_wr_id_tbl);
1299 kfree(qp->rqe_wr_id_tbl);
1301 pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1303 return ERR_PTR(status);
1306 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1310 struct ocrdma_qp *qp;
1311 struct ocrdma_dev *dev;
1312 enum ib_qp_state old_qps;
1314 qp = get_ocrdma_qp(ibqp);
1316 if (attr_mask & IB_QP_STATE)
1317 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1318 /* if new and previous states are same hw doesn't need to
1323 status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1328 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1329 int attr_mask, struct ib_udata *udata)
1331 unsigned long flags;
1332 int status = -EINVAL;
1333 struct ocrdma_qp *qp;
1334 struct ocrdma_dev *dev;
1335 enum ib_qp_state old_qps, new_qps;
1337 qp = get_ocrdma_qp(ibqp);
1340 /* syncronize with multiple context trying to change, retrive qps */
1341 mutex_lock(&dev->dev_lock);
1342 /* syncronize with wqe, rqe posting and cqe processing contexts */
1343 spin_lock_irqsave(&qp->q_lock, flags);
1344 old_qps = get_ibqp_state(qp->state);
1345 if (attr_mask & IB_QP_STATE)
1346 new_qps = attr->qp_state;
1349 spin_unlock_irqrestore(&qp->q_lock, flags);
1351 if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1352 IB_LINK_LAYER_ETHERNET)) {
1353 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1354 "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1355 __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1360 status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1364 mutex_unlock(&dev->dev_lock);
1368 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1386 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1388 int ib_qp_acc_flags = 0;
1390 if (qp_cap_flags & OCRDMA_QP_INB_WR)
1391 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1392 if (qp_cap_flags & OCRDMA_QP_INB_RD)
1393 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1394 return ib_qp_acc_flags;
1397 int ocrdma_query_qp(struct ib_qp *ibqp,
1398 struct ib_qp_attr *qp_attr,
1399 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1403 struct ocrdma_qp_params params;
1404 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1405 struct ocrdma_dev *dev = qp->dev;
1407 memset(¶ms, 0, sizeof(params));
1408 mutex_lock(&dev->dev_lock);
1409 status = ocrdma_mbx_query_qp(dev, qp, ¶ms);
1410 mutex_unlock(&dev->dev_lock);
1413 qp_attr->qp_state = get_ibqp_state(IB_QPS_INIT);
1414 qp_attr->cur_qp_state = get_ibqp_state(IB_QPS_INIT);
1416 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1417 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1418 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1419 qp_attr->path_mig_state = IB_MIG_MIGRATED;
1420 qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1421 qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1422 qp_attr->dest_qp_num =
1423 params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1425 qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1426 qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1427 qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1428 qp_attr->cap.max_send_sge = qp->sq.max_sges;
1429 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1430 qp_attr->cap.max_inline_data = qp->max_inline_data;
1431 qp_init_attr->cap = qp_attr->cap;
1432 memcpy(&qp_attr->ah_attr.grh.dgid, ¶ms.dgid[0],
1433 sizeof(params.dgid));
1434 qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1435 OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1436 qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1437 qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1438 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1439 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1440 qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1441 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1442 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1444 qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1445 qp_attr->ah_attr.port_num = 1;
1446 qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1447 OCRDMA_QP_PARAMS_SL_MASK) >>
1448 OCRDMA_QP_PARAMS_SL_SHIFT;
1449 qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1450 OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1451 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1452 qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1453 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1454 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1455 qp_attr->retry_cnt =
1456 (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1457 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1458 qp_attr->min_rnr_timer = 0;
1459 qp_attr->pkey_index = 0;
1460 qp_attr->port_num = 1;
1461 qp_attr->ah_attr.src_path_bits = 0;
1462 qp_attr->ah_attr.static_rate = 0;
1463 qp_attr->alt_pkey_index = 0;
1464 qp_attr->alt_port_num = 0;
1465 qp_attr->alt_timeout = 0;
1466 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1467 qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1468 OCRDMA_QP_PARAMS_STATE_SHIFT;
1469 qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1470 qp_attr->max_dest_rd_atomic =
1471 params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1472 qp_attr->max_rd_atomic =
1473 params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1474 qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1475 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1480 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, int idx)
1483 unsigned int mask = (1 << (idx % 32));
1485 if (srq->idx_bit_fields[i] & mask)
1486 srq->idx_bit_fields[i] &= ~mask;
1488 srq->idx_bit_fields[i] |= mask;
1491 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1493 return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1496 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1498 return (qp->sq.tail == qp->sq.head);
1501 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1503 return (qp->rq.tail == qp->rq.head);
1506 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1508 return q->va + (q->head * q->entry_size);
1511 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1514 return q->va + (idx * q->entry_size);
1517 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1519 q->head = (q->head + 1) & q->max_wqe_idx;
1522 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1524 q->tail = (q->tail + 1) & q->max_wqe_idx;
1527 /* discard the cqe for a given QP */
1528 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1530 unsigned long cq_flags;
1531 unsigned long flags;
1532 int discard_cnt = 0;
1533 u32 cur_getp, stop_getp;
1534 struct ocrdma_cqe *cqe;
1535 u32 qpn = 0, wqe_idx = 0;
1537 spin_lock_irqsave(&cq->cq_lock, cq_flags);
1539 /* traverse through the CQEs in the hw CQ,
1540 * find the matching CQE for a given qp,
1541 * mark the matching one discarded by clearing qpn.
1542 * ring the doorbell in the poll_cq() as
1543 * we don't complete out of order cqe.
1546 cur_getp = cq->getp;
1547 /* find upto when do we reap the cq. */
1548 stop_getp = cur_getp;
1550 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1553 cqe = cq->va + cur_getp;
1554 /* if (a) done reaping whole hw cq, or
1555 * (b) qp_xq becomes empty.
1558 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1559 /* if previously discarded cqe found, skip that too. */
1560 /* check for matching qp */
1561 if (qpn == 0 || qpn != qp->id)
1564 if (is_cqe_for_sq(cqe)) {
1565 ocrdma_hwq_inc_tail(&qp->sq);
1568 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1569 OCRDMA_CQE_BUFTAG_SHIFT) &
1570 qp->srq->rq.max_wqe_idx;
1573 spin_lock_irqsave(&qp->srq->q_lock, flags);
1574 ocrdma_hwq_inc_tail(&qp->srq->rq);
1575 ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1576 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1579 ocrdma_hwq_inc_tail(&qp->rq);
1582 /* mark cqe discarded so that it is not picked up later
1588 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1589 } while (cur_getp != stop_getp);
1590 spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1593 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1596 unsigned long flags;
1597 struct ocrdma_dev *dev = qp->dev;
1598 /* sync with any active CQ poll */
1600 spin_lock_irqsave(&dev->flush_q_lock, flags);
1601 found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1603 list_del(&qp->sq_entry);
1605 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1607 list_del(&qp->rq_entry);
1609 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1612 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1615 struct ocrdma_pd *pd;
1616 struct ocrdma_qp *qp;
1617 struct ocrdma_dev *dev;
1618 struct ib_qp_attr attrs;
1619 int attr_mask = IB_QP_STATE;
1620 unsigned long flags;
1622 qp = get_ocrdma_qp(ibqp);
1625 attrs.qp_state = IB_QPS_ERR;
1628 /* change the QP state to ERROR */
1629 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1631 /* ensure that CQEs for newly created QP (whose id may be same with
1632 * one which just getting destroyed are same), dont get
1633 * discarded until the old CQEs are discarded.
1635 mutex_lock(&dev->dev_lock);
1636 status = ocrdma_mbx_destroy_qp(dev, qp);
1639 * acquire CQ lock while destroy is in progress, in order to
1640 * protect against proessing in-flight CQEs for this QP.
1642 spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1643 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1644 spin_lock(&qp->rq_cq->cq_lock);
1646 ocrdma_del_qpn_map(dev, qp);
1648 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1649 spin_unlock(&qp->rq_cq->cq_lock);
1650 spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1653 ocrdma_discard_cqes(qp, qp->sq_cq);
1654 ocrdma_discard_cqes(qp, qp->rq_cq);
1656 mutex_unlock(&dev->dev_lock);
1659 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1660 PAGE_ALIGN(qp->sq.len));
1662 ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1663 PAGE_ALIGN(qp->rq.len));
1666 ocrdma_del_flush_qp(qp);
1668 kfree(qp->wqe_wr_id_tbl);
1669 kfree(qp->rqe_wr_id_tbl);
1674 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1675 struct ib_udata *udata)
1678 struct ocrdma_create_srq_uresp uresp;
1680 memset(&uresp, 0, sizeof(uresp));
1681 uresp.rq_dbid = srq->rq.dbid;
1682 uresp.num_rq_pages = 1;
1683 uresp.rq_page_addr[0] = srq->rq.pa;
1684 uresp.rq_page_size = srq->rq.len;
1685 uresp.db_page_addr = dev->nic_info.unmapped_db +
1686 (srq->pd->id * dev->nic_info.db_page_size);
1687 uresp.db_page_size = dev->nic_info.db_page_size;
1688 uresp.num_rqe_allocated = srq->rq.max_cnt;
1689 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1690 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1691 uresp.db_shift = 24;
1693 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1694 uresp.db_shift = 16;
1697 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1700 status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1701 uresp.rq_page_size);
1707 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1708 struct ib_srq_init_attr *init_attr,
1709 struct ib_udata *udata)
1711 int status = -ENOMEM;
1712 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1713 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1714 struct ocrdma_srq *srq;
1716 if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1717 return ERR_PTR(-EINVAL);
1718 if (init_attr->attr.max_wr > dev->attr.max_rqe)
1719 return ERR_PTR(-EINVAL);
1721 srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1723 return ERR_PTR(status);
1725 spin_lock_init(&srq->q_lock);
1727 srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1728 status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1732 if (udata == NULL) {
1733 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1735 if (srq->rqe_wr_id_tbl == NULL)
1738 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1739 (srq->rq.max_cnt % 32 ? 1 : 0);
1740 srq->idx_bit_fields =
1741 kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1742 if (srq->idx_bit_fields == NULL)
1744 memset(srq->idx_bit_fields, 0xff,
1745 srq->bit_fields_len * sizeof(u32));
1748 if (init_attr->attr.srq_limit) {
1749 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1755 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1763 ocrdma_mbx_destroy_srq(dev, srq);
1765 kfree(srq->rqe_wr_id_tbl);
1766 kfree(srq->idx_bit_fields);
1768 return ERR_PTR(status);
1771 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1772 struct ib_srq_attr *srq_attr,
1773 enum ib_srq_attr_mask srq_attr_mask,
1774 struct ib_udata *udata)
1777 struct ocrdma_srq *srq;
1779 srq = get_ocrdma_srq(ibsrq);
1780 if (srq_attr_mask & IB_SRQ_MAX_WR)
1783 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1787 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1790 struct ocrdma_srq *srq;
1792 srq = get_ocrdma_srq(ibsrq);
1793 status = ocrdma_mbx_query_srq(srq, srq_attr);
1797 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1800 struct ocrdma_srq *srq;
1801 struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1803 srq = get_ocrdma_srq(ibsrq);
1805 status = ocrdma_mbx_destroy_srq(dev, srq);
1808 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1809 PAGE_ALIGN(srq->rq.len));
1811 kfree(srq->idx_bit_fields);
1812 kfree(srq->rqe_wr_id_tbl);
1817 /* unprivileged verbs and their support functions. */
1818 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1819 struct ocrdma_hdr_wqe *hdr,
1820 struct ib_send_wr *wr)
1822 struct ocrdma_ewqe_ud_hdr *ud_hdr =
1823 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1824 struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
1826 ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
1827 if (qp->qp_type == IB_QPT_GSI)
1828 ud_hdr->qkey = qp->qkey;
1830 ud_hdr->qkey = wr->wr.ud.remote_qkey;
1831 ud_hdr->rsvd_ahid = ah->id;
1834 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1835 struct ocrdma_sge *sge, int num_sge,
1836 struct ib_sge *sg_list)
1840 for (i = 0; i < num_sge; i++) {
1841 sge[i].lrkey = sg_list[i].lkey;
1842 sge[i].addr_lo = sg_list[i].addr;
1843 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1844 sge[i].len = sg_list[i].length;
1845 hdr->total_len += sg_list[i].length;
1848 memset(sge, 0, sizeof(*sge));
1851 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1853 uint32_t total_len = 0, i;
1855 for (i = 0; i < num_sge; i++)
1856 total_len += sg_list[i].length;
1861 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1862 struct ocrdma_hdr_wqe *hdr,
1863 struct ocrdma_sge *sge,
1864 struct ib_send_wr *wr, u32 wqe_size)
1869 if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1870 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1871 if (unlikely(hdr->total_len > qp->max_inline_data)) {
1872 pr_err("%s() supported_len=0x%x,\n"
1873 " unspported len req=0x%x\n", __func__,
1874 qp->max_inline_data, hdr->total_len);
1877 dpp_addr = (char *)sge;
1878 for (i = 0; i < wr->num_sge; i++) {
1880 (void *)(unsigned long)wr->sg_list[i].addr,
1881 wr->sg_list[i].length);
1882 dpp_addr += wr->sg_list[i].length;
1885 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1886 if (0 == hdr->total_len)
1887 wqe_size += sizeof(struct ocrdma_sge);
1888 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1890 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1892 wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1894 wqe_size += sizeof(struct ocrdma_sge);
1895 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1897 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1901 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1902 struct ib_send_wr *wr)
1905 struct ocrdma_sge *sge;
1906 u32 wqe_size = sizeof(*hdr);
1908 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1909 ocrdma_build_ud_hdr(qp, hdr, wr);
1910 sge = (struct ocrdma_sge *)(hdr + 2);
1911 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1913 sge = (struct ocrdma_sge *)(hdr + 1);
1916 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1920 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1921 struct ib_send_wr *wr)
1924 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1925 struct ocrdma_sge *sge = ext_rw + 1;
1926 u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1928 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1931 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1932 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1933 ext_rw->lrkey = wr->wr.rdma.rkey;
1934 ext_rw->len = hdr->total_len;
1938 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1939 struct ib_send_wr *wr)
1941 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1942 struct ocrdma_sge *sge = ext_rw + 1;
1943 u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
1944 sizeof(struct ocrdma_hdr_wqe);
1946 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1947 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1948 hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
1949 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1951 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1952 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1953 ext_rw->lrkey = wr->wr.rdma.rkey;
1954 ext_rw->len = hdr->total_len;
1957 static void build_frmr_pbes(struct ib_send_wr *wr, struct ocrdma_pbl *pbl_tbl,
1958 struct ocrdma_hw_mr *hwmr)
1963 struct ocrdma_pbe *pbe;
1965 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1968 /* go through the OS phy regions & fill hw pbe entries into pbls. */
1969 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
1970 /* number of pbes can be more for one OS buf, when
1971 * buffers are of different sizes.
1972 * split the ib_buf to one or more pbes.
1974 buf_addr = wr->wr.fast_reg.page_list->page_list[i];
1975 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
1976 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
1980 /* if the pbl is full storing the pbes,
1983 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
1985 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1991 static int get_encoded_page_size(int pg_sz)
1993 /* Max size is 256M 4096 << 16 */
1996 if (pg_sz == (4096 << i))
2002 static int ocrdma_build_fr(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2003 struct ib_send_wr *wr)
2006 struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2007 struct ocrdma_mr *mr;
2008 u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2010 wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2012 if (wr->wr.fast_reg.page_list_len > qp->dev->attr.max_pages_per_frmr)
2015 hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2016 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2018 if (wr->wr.fast_reg.page_list_len == 0)
2020 if (wr->wr.fast_reg.access_flags & IB_ACCESS_LOCAL_WRITE)
2021 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2022 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_WRITE)
2023 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2024 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_READ)
2025 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2026 hdr->lkey = wr->wr.fast_reg.rkey;
2027 hdr->total_len = wr->wr.fast_reg.length;
2029 fbo = wr->wr.fast_reg.iova_start -
2030 (wr->wr.fast_reg.page_list->page_list[0] & PAGE_MASK);
2032 fast_reg->va_hi = upper_32_bits(wr->wr.fast_reg.iova_start);
2033 fast_reg->va_lo = (u32) (wr->wr.fast_reg.iova_start & 0xffffffff);
2034 fast_reg->fbo_hi = upper_32_bits(fbo);
2035 fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2036 fast_reg->num_sges = wr->wr.fast_reg.page_list_len;
2037 fast_reg->size_sge =
2038 get_encoded_page_size(1 << wr->wr.fast_reg.page_shift);
2039 mr = (struct ocrdma_mr *) (unsigned long)
2040 qp->dev->stag_arr[(hdr->lkey >> 8) & (OCRDMA_MAX_STAG - 1)];
2041 build_frmr_pbes(wr, mr->hwmr.pbl_table, &mr->hwmr);
2045 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2047 u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2049 iowrite32(val, qp->sq_db);
2052 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2053 struct ib_send_wr **bad_wr)
2056 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2057 struct ocrdma_hdr_wqe *hdr;
2058 unsigned long flags;
2060 spin_lock_irqsave(&qp->q_lock, flags);
2061 if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2062 spin_unlock_irqrestore(&qp->q_lock, flags);
2068 if (qp->qp_type == IB_QPT_UD &&
2069 (wr->opcode != IB_WR_SEND &&
2070 wr->opcode != IB_WR_SEND_WITH_IMM)) {
2075 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2076 wr->num_sge > qp->sq.max_sges) {
2081 hdr = ocrdma_hwq_head(&qp->sq);
2083 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2084 hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2085 if (wr->send_flags & IB_SEND_FENCE)
2087 (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2088 if (wr->send_flags & IB_SEND_SOLICITED)
2090 (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2092 switch (wr->opcode) {
2093 case IB_WR_SEND_WITH_IMM:
2094 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2095 hdr->immdt = ntohl(wr->ex.imm_data);
2097 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2098 ocrdma_build_send(qp, hdr, wr);
2100 case IB_WR_SEND_WITH_INV:
2101 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2102 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2103 hdr->lkey = wr->ex.invalidate_rkey;
2104 status = ocrdma_build_send(qp, hdr, wr);
2106 case IB_WR_RDMA_WRITE_WITH_IMM:
2107 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2108 hdr->immdt = ntohl(wr->ex.imm_data);
2109 case IB_WR_RDMA_WRITE:
2110 hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2111 status = ocrdma_build_write(qp, hdr, wr);
2113 case IB_WR_RDMA_READ_WITH_INV:
2114 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2115 case IB_WR_RDMA_READ:
2116 ocrdma_build_read(qp, hdr, wr);
2118 case IB_WR_LOCAL_INV:
2120 (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2121 hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2122 sizeof(struct ocrdma_sge)) /
2123 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2124 hdr->lkey = wr->ex.invalidate_rkey;
2126 case IB_WR_FAST_REG_MR:
2127 status = ocrdma_build_fr(qp, hdr, wr);
2137 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2138 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2140 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2141 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2142 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2143 OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2144 /* make sure wqe is written before adapter can access it */
2146 /* inform hw to start processing it */
2147 ocrdma_ring_sq_db(qp);
2149 /* update pointer, counter for next wr */
2150 ocrdma_hwq_inc_head(&qp->sq);
2153 spin_unlock_irqrestore(&qp->q_lock, flags);
2157 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2159 u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2161 iowrite32(val, qp->rq_db);
2164 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2168 struct ocrdma_sge *sge;
2170 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2172 wqe_size = sizeof(*sge) + sizeof(*rqe);
2174 rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2175 OCRDMA_WQE_SIZE_SHIFT);
2176 rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2177 rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2179 rqe->rsvd_tag = tag;
2180 sge = (struct ocrdma_sge *)(rqe + 1);
2181 ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2182 ocrdma_cpu_to_le32(rqe, wqe_size);
2185 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2186 struct ib_recv_wr **bad_wr)
2189 unsigned long flags;
2190 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2191 struct ocrdma_hdr_wqe *rqe;
2193 spin_lock_irqsave(&qp->q_lock, flags);
2194 if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2195 spin_unlock_irqrestore(&qp->q_lock, flags);
2200 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2201 wr->num_sge > qp->rq.max_sges) {
2206 rqe = ocrdma_hwq_head(&qp->rq);
2207 ocrdma_build_rqe(rqe, wr, 0);
2209 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2210 /* make sure rqe is written before adapter can access it */
2213 /* inform hw to start processing it */
2214 ocrdma_ring_rq_db(qp);
2216 /* update pointer, counter for next wr */
2217 ocrdma_hwq_inc_head(&qp->rq);
2220 spin_unlock_irqrestore(&qp->q_lock, flags);
2224 /* cqe for srq's rqe can potentially arrive out of order.
2225 * index gives the entry in the shadow table where to store
2226 * the wr_id. tag/index is returned in cqe to reference back
2229 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2234 for (row = 0; row < srq->bit_fields_len; row++) {
2235 if (srq->idx_bit_fields[row]) {
2236 indx = ffs(srq->idx_bit_fields[row]);
2237 indx = (row * 32) + (indx - 1);
2238 if (indx >= srq->rq.max_cnt)
2240 ocrdma_srq_toggle_bit(srq, indx);
2245 if (row == srq->bit_fields_len)
2247 return indx + 1; /* Use from index 1 */
2250 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2252 u32 val = srq->rq.dbid | (1 << 16);
2254 iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2257 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2258 struct ib_recv_wr **bad_wr)
2261 unsigned long flags;
2262 struct ocrdma_srq *srq;
2263 struct ocrdma_hdr_wqe *rqe;
2266 srq = get_ocrdma_srq(ibsrq);
2268 spin_lock_irqsave(&srq->q_lock, flags);
2270 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2271 wr->num_sge > srq->rq.max_sges) {
2276 tag = ocrdma_srq_get_idx(srq);
2277 rqe = ocrdma_hwq_head(&srq->rq);
2278 ocrdma_build_rqe(rqe, wr, tag);
2280 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2281 /* make sure rqe is written before adapter can perform DMA */
2283 /* inform hw to start processing it */
2284 ocrdma_ring_srq_db(srq);
2285 /* update pointer, counter for next wr */
2286 ocrdma_hwq_inc_head(&srq->rq);
2289 spin_unlock_irqrestore(&srq->q_lock, flags);
2293 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2295 enum ib_wc_status ibwc_status;
2298 case OCRDMA_CQE_GENERAL_ERR:
2299 ibwc_status = IB_WC_GENERAL_ERR;
2301 case OCRDMA_CQE_LOC_LEN_ERR:
2302 ibwc_status = IB_WC_LOC_LEN_ERR;
2304 case OCRDMA_CQE_LOC_QP_OP_ERR:
2305 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2307 case OCRDMA_CQE_LOC_EEC_OP_ERR:
2308 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2310 case OCRDMA_CQE_LOC_PROT_ERR:
2311 ibwc_status = IB_WC_LOC_PROT_ERR;
2313 case OCRDMA_CQE_WR_FLUSH_ERR:
2314 ibwc_status = IB_WC_WR_FLUSH_ERR;
2316 case OCRDMA_CQE_MW_BIND_ERR:
2317 ibwc_status = IB_WC_MW_BIND_ERR;
2319 case OCRDMA_CQE_BAD_RESP_ERR:
2320 ibwc_status = IB_WC_BAD_RESP_ERR;
2322 case OCRDMA_CQE_LOC_ACCESS_ERR:
2323 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2325 case OCRDMA_CQE_REM_INV_REQ_ERR:
2326 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2328 case OCRDMA_CQE_REM_ACCESS_ERR:
2329 ibwc_status = IB_WC_REM_ACCESS_ERR;
2331 case OCRDMA_CQE_REM_OP_ERR:
2332 ibwc_status = IB_WC_REM_OP_ERR;
2334 case OCRDMA_CQE_RETRY_EXC_ERR:
2335 ibwc_status = IB_WC_RETRY_EXC_ERR;
2337 case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2338 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2340 case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2341 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2343 case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2344 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2346 case OCRDMA_CQE_REM_ABORT_ERR:
2347 ibwc_status = IB_WC_REM_ABORT_ERR;
2349 case OCRDMA_CQE_INV_EECN_ERR:
2350 ibwc_status = IB_WC_INV_EECN_ERR;
2352 case OCRDMA_CQE_INV_EEC_STATE_ERR:
2353 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2355 case OCRDMA_CQE_FATAL_ERR:
2356 ibwc_status = IB_WC_FATAL_ERR;
2358 case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2359 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2362 ibwc_status = IB_WC_GENERAL_ERR;
2368 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2371 struct ocrdma_hdr_wqe *hdr;
2372 struct ocrdma_sge *rw;
2375 hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2377 ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2378 /* Undo the hdr->cw swap */
2379 opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2382 ibwc->opcode = IB_WC_RDMA_WRITE;
2385 rw = (struct ocrdma_sge *)(hdr + 1);
2386 ibwc->opcode = IB_WC_RDMA_READ;
2387 ibwc->byte_len = rw->len;
2390 ibwc->opcode = IB_WC_SEND;
2393 ibwc->opcode = IB_WC_FAST_REG_MR;
2395 case OCRDMA_LKEY_INV:
2396 ibwc->opcode = IB_WC_LOCAL_INV;
2399 ibwc->status = IB_WC_GENERAL_ERR;
2400 pr_err("%s() invalid opcode received = 0x%x\n",
2401 __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2406 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2407 struct ocrdma_cqe *cqe)
2409 if (is_cqe_for_sq(cqe)) {
2410 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2411 cqe->flags_status_srcqpn) &
2412 ~OCRDMA_CQE_STATUS_MASK);
2413 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2414 cqe->flags_status_srcqpn) |
2415 (OCRDMA_CQE_WR_FLUSH_ERR <<
2416 OCRDMA_CQE_STATUS_SHIFT));
2418 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2419 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2420 cqe->flags_status_srcqpn) &
2421 ~OCRDMA_CQE_UD_STATUS_MASK);
2422 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2423 cqe->flags_status_srcqpn) |
2424 (OCRDMA_CQE_WR_FLUSH_ERR <<
2425 OCRDMA_CQE_UD_STATUS_SHIFT));
2427 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2428 cqe->flags_status_srcqpn) &
2429 ~OCRDMA_CQE_STATUS_MASK);
2430 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2431 cqe->flags_status_srcqpn) |
2432 (OCRDMA_CQE_WR_FLUSH_ERR <<
2433 OCRDMA_CQE_STATUS_SHIFT));
2438 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2439 struct ocrdma_qp *qp, int status)
2441 bool expand = false;
2444 ibwc->qp = &qp->ibqp;
2445 ibwc->status = ocrdma_to_ibwc_err(status);
2447 ocrdma_flush_qp(qp);
2448 ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2450 /* if wqe/rqe pending for which cqe needs to be returned,
2451 * trigger inflating it.
2453 if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2455 ocrdma_set_cqe_status_flushed(qp, cqe);
2460 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2461 struct ocrdma_qp *qp, int status)
2463 ibwc->opcode = IB_WC_RECV;
2464 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2465 ocrdma_hwq_inc_tail(&qp->rq);
2467 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2470 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2471 struct ocrdma_qp *qp, int status)
2473 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2474 ocrdma_hwq_inc_tail(&qp->sq);
2476 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2480 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2481 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2482 bool *polled, bool *stop)
2485 int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2486 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2488 /* when hw sq is empty, but rq is not empty, so we continue
2489 * to keep the cqe in order to get the cq event again.
2491 if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2492 /* when cq for rq and sq is same, it is safe to return
2493 * flush cqe for RQEs.
2495 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2497 status = OCRDMA_CQE_WR_FLUSH_ERR;
2498 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2500 /* stop processing further cqe as this cqe is used for
2501 * triggering cq event on buddy cq of RQ.
2502 * When QP is destroyed, this cqe will be removed
2503 * from the cq's hardware q.
2509 } else if (is_hw_sq_empty(qp)) {
2516 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2521 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2522 struct ocrdma_cqe *cqe,
2523 struct ib_wc *ibwc, bool *polled)
2525 bool expand = false;
2526 int tail = qp->sq.tail;
2529 if (!qp->wqe_wr_id_tbl[tail].signaled) {
2530 *polled = false; /* WC cannot be consumed yet */
2532 ibwc->status = IB_WC_SUCCESS;
2534 ibwc->qp = &qp->ibqp;
2535 ocrdma_update_wc(qp, ibwc, tail);
2538 wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2539 OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2540 if (tail != wqe_idx)
2541 expand = true; /* Coalesced CQE can't be consumed yet */
2543 ocrdma_hwq_inc_tail(&qp->sq);
2547 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2548 struct ib_wc *ibwc, bool *polled, bool *stop)
2553 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2554 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2556 if (status == OCRDMA_CQE_SUCCESS)
2557 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2559 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2563 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2567 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2568 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2569 ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2570 OCRDMA_CQE_SRCQP_MASK;
2571 ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2572 OCRDMA_CQE_PKEY_MASK;
2573 ibwc->wc_flags = IB_WC_GRH;
2574 ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2575 OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2579 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2580 struct ocrdma_cqe *cqe,
2581 struct ocrdma_qp *qp)
2583 unsigned long flags;
2584 struct ocrdma_srq *srq;
2587 srq = get_ocrdma_srq(qp->ibqp.srq);
2588 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2589 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2593 ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2594 spin_lock_irqsave(&srq->q_lock, flags);
2595 ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2596 spin_unlock_irqrestore(&srq->q_lock, flags);
2597 ocrdma_hwq_inc_tail(&srq->rq);
2600 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2601 struct ib_wc *ibwc, bool *polled, bool *stop,
2606 /* when hw_rq is empty, but wq is not empty, so continue
2607 * to keep the cqe to get the cq event again.
2609 if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2610 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2612 status = OCRDMA_CQE_WR_FLUSH_ERR;
2613 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2619 } else if (is_hw_rq_empty(qp)) {
2626 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2631 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2632 struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2634 ibwc->opcode = IB_WC_RECV;
2635 ibwc->qp = &qp->ibqp;
2636 ibwc->status = IB_WC_SUCCESS;
2638 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2639 ocrdma_update_ud_rcqe(ibwc, cqe);
2641 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2643 if (is_cqe_imm(cqe)) {
2644 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2645 ibwc->wc_flags |= IB_WC_WITH_IMM;
2646 } else if (is_cqe_wr_imm(cqe)) {
2647 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2648 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2649 ibwc->wc_flags |= IB_WC_WITH_IMM;
2650 } else if (is_cqe_invalidated(cqe)) {
2651 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2652 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2655 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2657 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2658 ocrdma_hwq_inc_tail(&qp->rq);
2662 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2663 struct ib_wc *ibwc, bool *polled, bool *stop)
2666 bool expand = false;
2669 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2670 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2671 OCRDMA_CQE_UD_STATUS_MASK) >>
2672 OCRDMA_CQE_UD_STATUS_SHIFT;
2674 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2675 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2678 if (status == OCRDMA_CQE_SUCCESS) {
2680 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2682 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2688 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2691 if (cq->phase_change) {
2693 cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2695 /* clear valid bit */
2696 cqe->flags_status_srcqpn = 0;
2700 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2705 bool expand = false;
2706 int polled_hw_cqes = 0;
2707 struct ocrdma_qp *qp = NULL;
2708 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2709 struct ocrdma_cqe *cqe;
2710 u16 cur_getp; bool polled = false; bool stop = false;
2712 cur_getp = cq->getp;
2713 while (num_entries) {
2714 cqe = cq->va + cur_getp;
2715 /* check whether valid cqe or not */
2716 if (!is_cqe_valid(cq, cqe))
2718 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2719 /* ignore discarded cqe */
2722 qp = dev->qp_tbl[qpn];
2725 if (is_cqe_for_sq(cqe)) {
2726 expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2729 expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2736 /* clear qpn to avoid duplicate processing by discard_cqe() */
2739 polled_hw_cqes += 1;
2740 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2741 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2751 cq->getp = cur_getp;
2752 if (cq->deferred_arm) {
2753 ocrdma_ring_cq_db(dev, cq->id, true, cq->deferred_sol,
2755 cq->deferred_arm = false;
2756 cq->deferred_sol = false;
2758 /* We need to pop the CQE. No need to arm */
2759 ocrdma_ring_cq_db(dev, cq->id, false, cq->deferred_sol,
2761 cq->deferred_sol = false;
2767 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2768 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2769 struct ocrdma_qp *qp, struct ib_wc *ibwc)
2773 while (num_entries) {
2774 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2776 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2777 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2778 ocrdma_hwq_inc_tail(&qp->sq);
2779 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2780 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2781 ocrdma_hwq_inc_tail(&qp->rq);
2786 ibwc->status = IB_WC_WR_FLUSH_ERR;
2794 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2796 int cqes_to_poll = num_entries;
2797 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2798 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2799 int num_os_cqe = 0, err_cqes = 0;
2800 struct ocrdma_qp *qp;
2801 unsigned long flags;
2803 /* poll cqes from adapter CQ */
2804 spin_lock_irqsave(&cq->cq_lock, flags);
2805 num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2806 spin_unlock_irqrestore(&cq->cq_lock, flags);
2807 cqes_to_poll -= num_os_cqe;
2810 wc = wc + num_os_cqe;
2811 /* adapter returns single error cqe when qp moves to
2812 * error state. So insert error cqes with wc_status as
2813 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2814 * respectively which uses this CQ.
2816 spin_lock_irqsave(&dev->flush_q_lock, flags);
2817 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2818 if (cqes_to_poll == 0)
2820 err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2821 cqes_to_poll -= err_cqes;
2822 num_os_cqe += err_cqes;
2825 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2830 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2832 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2833 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2835 unsigned long flags;
2836 bool arm_needed = false, sol_needed = false;
2840 spin_lock_irqsave(&cq->cq_lock, flags);
2841 if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2843 if (cq_flags & IB_CQ_SOLICITED)
2846 if (cq->first_arm) {
2847 ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2848 cq->first_arm = false;
2851 cq->deferred_arm = true;
2852 cq->deferred_sol = sol_needed;
2853 spin_unlock_irqrestore(&cq->cq_lock, flags);
2858 struct ib_mr *ocrdma_alloc_frmr(struct ib_pd *ibpd, int max_page_list_len)
2861 struct ocrdma_mr *mr;
2862 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2863 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2865 if (max_page_list_len > dev->attr.max_pages_per_frmr)
2866 return ERR_PTR(-EINVAL);
2868 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2870 return ERR_PTR(-ENOMEM);
2872 status = ocrdma_get_pbl_info(dev, mr, max_page_list_len);
2876 mr->hwmr.remote_rd = 0;
2877 mr->hwmr.remote_wr = 0;
2878 mr->hwmr.local_rd = 0;
2879 mr->hwmr.local_wr = 0;
2880 mr->hwmr.mw_bind = 0;
2881 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2884 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2887 mr->ibmr.rkey = mr->hwmr.lkey;
2888 mr->ibmr.lkey = mr->hwmr.lkey;
2889 dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
2893 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2896 return ERR_PTR(-ENOMEM);
2899 struct ib_fast_reg_page_list *ocrdma_alloc_frmr_page_list(struct ib_device
2903 struct ib_fast_reg_page_list *frmr_list;
2906 size = sizeof(*frmr_list) + (page_list_len * sizeof(u64));
2907 frmr_list = kzalloc(size, GFP_KERNEL);
2909 return ERR_PTR(-ENOMEM);
2910 frmr_list->page_list = (u64 *)(frmr_list + 1);
2914 void ocrdma_free_frmr_page_list(struct ib_fast_reg_page_list *page_list)
2919 #define MAX_KERNEL_PBE_SIZE 65536
2920 static inline int count_kernel_pbes(struct ib_phys_buf *buf_list,
2921 int buf_cnt, u32 *pbe_size)
2926 *pbe_size = roundup(buf_list[0].size, PAGE_SIZE);
2927 *pbe_size = roundup_pow_of_two(*pbe_size);
2929 /* find the smallest PBE size that we can have */
2930 for (i = 0; i < buf_cnt; i++) {
2931 /* first addr may not be page aligned, so ignore checking */
2932 if ((i != 0) && ((buf_list[i].addr & ~PAGE_MASK) ||
2933 (buf_list[i].size & ~PAGE_MASK))) {
2937 /* if configured PBE size is greater then the chosen one,
2938 * reduce the PBE size.
2940 buf_size = roundup(buf_list[i].size, PAGE_SIZE);
2941 /* pbe_size has to be even multiple of 4K 1,2,4,8...*/
2942 buf_size = roundup_pow_of_two(buf_size);
2943 if (*pbe_size > buf_size)
2944 *pbe_size = buf_size;
2946 total_size += buf_size;
2948 *pbe_size = *pbe_size > MAX_KERNEL_PBE_SIZE ?
2949 (MAX_KERNEL_PBE_SIZE) : (*pbe_size);
2951 /* num_pbes = total_size / (*pbe_size); this is implemented below. */
2953 return total_size >> ilog2(*pbe_size);
2956 static void build_kernel_pbes(struct ib_phys_buf *buf_list, int ib_buf_cnt,
2957 u32 pbe_size, struct ocrdma_pbl *pbl_tbl,
2958 struct ocrdma_hw_mr *hwmr)
2962 int pbes_per_buf = 0;
2965 struct ocrdma_pbe *pbe;
2966 int total_num_pbes = 0;
2968 if (!hwmr->num_pbes)
2971 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2974 /* go through the OS phy regions & fill hw pbe entries into pbls. */
2975 for (i = 0; i < ib_buf_cnt; i++) {
2976 buf_addr = buf_list[i].addr;
2978 roundup_pow_of_two(roundup(buf_list[i].size, PAGE_SIZE)) /
2980 hwmr->len += buf_list[i].size;
2981 /* number of pbes can be more for one OS buf, when
2982 * buffers are of different sizes.
2983 * split the ib_buf to one or more pbes.
2985 for (idx = 0; idx < pbes_per_buf; idx++) {
2986 /* we program always page aligned addresses,
2987 * first unaligned address is taken care by fbo.
2990 /* for non zero fbo, assign the
2991 * start of the page.
2994 cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2996 cpu_to_le32((u32) upper_32_bits(buf_addr));
2999 cpu_to_le32((u32) (buf_addr & 0xffffffff));
3001 cpu_to_le32((u32) upper_32_bits(buf_addr));
3003 buf_addr += pbe_size;
3005 total_num_pbes += 1;
3008 if (total_num_pbes == hwmr->num_pbes)
3010 /* if the pbl is full storing the pbes,
3013 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
3015 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
3024 struct ib_mr *ocrdma_reg_kernel_mr(struct ib_pd *ibpd,
3025 struct ib_phys_buf *buf_list,
3026 int buf_cnt, int acc, u64 *iova_start)
3028 int status = -ENOMEM;
3029 struct ocrdma_mr *mr;
3030 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3031 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3035 if ((acc & IB_ACCESS_REMOTE_WRITE) && !(acc & IB_ACCESS_LOCAL_WRITE))
3036 return ERR_PTR(-EINVAL);
3038 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3040 return ERR_PTR(status);
3042 num_pbes = count_kernel_pbes(buf_list, buf_cnt, &pbe_size);
3043 if (num_pbes == 0) {
3047 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
3051 mr->hwmr.pbe_size = pbe_size;
3052 mr->hwmr.fbo = *iova_start - (buf_list[0].addr & PAGE_MASK);
3053 mr->hwmr.va = *iova_start;
3054 mr->hwmr.local_rd = 1;
3055 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3056 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3057 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3058 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3059 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
3061 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3064 build_kernel_pbes(buf_list, buf_cnt, pbe_size, mr->hwmr.pbl_table,
3066 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
3070 mr->ibmr.lkey = mr->hwmr.lkey;
3071 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
3072 mr->ibmr.rkey = mr->hwmr.lkey;
3076 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3079 return ERR_PTR(status);