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 = ~0ull;
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;
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 = 0;
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 BUG_ON(uctx->pd_in_use);
333 uctx->cntxt_pd = NULL;
334 status = _ocrdma_dealloc_pd(dev, pd);
338 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
340 struct ocrdma_pd *pd = NULL;
342 mutex_lock(&uctx->mm_list_lock);
343 if (!uctx->pd_in_use) {
344 uctx->pd_in_use = true;
347 mutex_unlock(&uctx->mm_list_lock);
352 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
354 mutex_lock(&uctx->mm_list_lock);
355 uctx->pd_in_use = false;
356 mutex_unlock(&uctx->mm_list_lock);
359 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
360 struct ib_udata *udata)
363 struct ocrdma_ucontext *ctx;
364 struct ocrdma_alloc_ucontext_resp resp;
365 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
366 struct pci_dev *pdev = dev->nic_info.pdev;
367 u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
370 return ERR_PTR(-EFAULT);
371 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
373 return ERR_PTR(-ENOMEM);
374 INIT_LIST_HEAD(&ctx->mm_head);
375 mutex_init(&ctx->mm_list_lock);
377 ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
378 &ctx->ah_tbl.pa, GFP_KERNEL);
379 if (!ctx->ah_tbl.va) {
381 return ERR_PTR(-ENOMEM);
383 memset(ctx->ah_tbl.va, 0, map_len);
384 ctx->ah_tbl.len = map_len;
386 memset(&resp, 0, sizeof(resp));
387 resp.ah_tbl_len = ctx->ah_tbl.len;
388 resp.ah_tbl_page = ctx->ah_tbl.pa;
390 status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
394 status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
398 resp.dev_id = dev->id;
399 resp.max_inline_data = dev->attr.max_inline_data;
400 resp.wqe_size = dev->attr.wqe_size;
401 resp.rqe_size = dev->attr.rqe_size;
402 resp.dpp_wqe_size = dev->attr.wqe_size;
404 memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
405 status = ib_copy_to_udata(udata, &resp, sizeof(resp));
408 return &ctx->ibucontext;
412 ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
414 dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
417 return ERR_PTR(status);
420 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
423 struct ocrdma_mm *mm, *tmp;
424 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
425 struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
426 struct pci_dev *pdev = dev->nic_info.pdev;
428 status = ocrdma_dealloc_ucontext_pd(uctx);
430 ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
431 dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
434 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
435 list_del(&mm->entry);
442 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
444 struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
445 struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
446 unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
447 u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
448 unsigned long len = (vma->vm_end - vma->vm_start);
452 if (vma->vm_start & (PAGE_SIZE - 1))
454 found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
458 if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
459 dev->nic_info.db_total_size)) &&
460 (len <= dev->nic_info.db_page_size)) {
461 if (vma->vm_flags & VM_READ)
464 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
465 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
466 len, vma->vm_page_prot);
467 } else if (dev->nic_info.dpp_unmapped_len &&
468 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
469 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
470 dev->nic_info.dpp_unmapped_len)) &&
471 (len <= dev->nic_info.dpp_unmapped_len)) {
472 if (vma->vm_flags & VM_READ)
475 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
476 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
477 len, vma->vm_page_prot);
479 status = remap_pfn_range(vma, vma->vm_start,
480 vma->vm_pgoff, len, vma->vm_page_prot);
485 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
486 struct ib_ucontext *ib_ctx,
487 struct ib_udata *udata)
491 u64 dpp_page_addr = 0;
493 struct ocrdma_alloc_pd_uresp rsp;
494 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
496 memset(&rsp, 0, sizeof(rsp));
498 rsp.dpp_enabled = pd->dpp_enabled;
499 db_page_addr = ocrdma_get_db_addr(dev, pd->id);
500 db_page_size = dev->nic_info.db_page_size;
502 status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
506 if (pd->dpp_enabled) {
507 dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
508 (pd->id * PAGE_SIZE);
509 status = ocrdma_add_mmap(uctx, dpp_page_addr,
513 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
514 rsp.dpp_page_addr_lo = dpp_page_addr;
517 status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
526 ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
528 ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
532 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
533 struct ib_ucontext *context,
534 struct ib_udata *udata)
536 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
537 struct ocrdma_pd *pd;
538 struct ocrdma_ucontext *uctx = NULL;
540 u8 is_uctx_pd = false;
542 if (udata && context) {
543 uctx = get_ocrdma_ucontext(context);
544 pd = ocrdma_get_ucontext_pd(uctx);
551 pd = _ocrdma_alloc_pd(dev, uctx, udata);
553 status = PTR_ERR(pd);
558 if (udata && context) {
559 status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
567 ocrdma_release_ucontext_pd(uctx);
569 status = ocrdma_mbx_dealloc_pd(dev, pd);
573 return ERR_PTR(status);
576 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
578 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
579 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
580 struct ocrdma_ucontext *uctx = NULL;
586 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
587 (pd->id * PAGE_SIZE);
589 ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
590 usr_db = ocrdma_get_db_addr(dev, pd->id);
591 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
593 if (is_ucontext_pd(uctx, pd)) {
594 ocrdma_release_ucontext_pd(uctx);
598 status = _ocrdma_dealloc_pd(dev, pd);
602 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
603 u32 pdid, int acc, u32 num_pbls, u32 addr_check)
608 mr->hwmr.local_rd = 1;
609 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
610 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
611 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
612 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
613 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
614 mr->hwmr.num_pbls = num_pbls;
616 status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
620 mr->ibmr.lkey = mr->hwmr.lkey;
621 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
622 mr->ibmr.rkey = mr->hwmr.lkey;
626 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
629 struct ocrdma_mr *mr;
630 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
631 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
633 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
634 pr_err("%s err, invalid access rights\n", __func__);
635 return ERR_PTR(-EINVAL);
638 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
640 return ERR_PTR(-ENOMEM);
642 status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
643 OCRDMA_ADDR_CHECK_DISABLE);
646 return ERR_PTR(status);
652 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
653 struct ocrdma_hw_mr *mr)
655 struct pci_dev *pdev = dev->nic_info.pdev;
659 for (i = 0; i < mr->num_pbls; i++) {
660 if (!mr->pbl_table[i].va)
662 dma_free_coherent(&pdev->dev, mr->pbl_size,
664 mr->pbl_table[i].pa);
666 kfree(mr->pbl_table);
667 mr->pbl_table = NULL;
671 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
680 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
681 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
685 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
686 num_pbls = num_pbls / (pbl_size / sizeof(u64));
688 } while (num_pbls >= dev->attr.max_num_mr_pbl);
690 mr->hwmr.num_pbes = num_pbes;
691 mr->hwmr.num_pbls = num_pbls;
692 mr->hwmr.pbl_size = pbl_size;
696 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
700 u32 dma_len = mr->pbl_size;
701 struct pci_dev *pdev = dev->nic_info.pdev;
705 mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
706 mr->num_pbls, GFP_KERNEL);
711 for (i = 0; i < mr->num_pbls; i++) {
712 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
714 ocrdma_free_mr_pbl_tbl(dev, mr);
718 memset(va, 0, dma_len);
719 mr->pbl_table[i].va = va;
720 mr->pbl_table[i].pa = pa;
725 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
728 struct ocrdma_pbe *pbe;
729 struct scatterlist *sg;
730 struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
731 struct ib_umem *umem = mr->umem;
732 int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
734 if (!mr->hwmr.num_pbes)
737 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
740 shift = ilog2(umem->page_size);
742 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
743 pages = sg_dma_len(sg) >> shift;
744 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
745 /* store the page address in pbe */
747 cpu_to_le32(sg_dma_address
749 (umem->page_size * pg_cnt));
751 cpu_to_le32(upper_32_bits
754 umem->page_size * pg_cnt)));
759 /* if done building pbes, issue the mbx cmd. */
760 if (total_num_pbes == num_pbes)
763 /* if the given pbl is full storing the pbes,
767 (mr->hwmr.pbl_size / sizeof(u64))) {
769 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
777 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
778 u64 usr_addr, int acc, struct ib_udata *udata)
780 int status = -ENOMEM;
781 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
782 struct ocrdma_mr *mr;
783 struct ocrdma_pd *pd;
786 pd = get_ocrdma_pd(ibpd);
788 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
789 return ERR_PTR(-EINVAL);
791 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
793 return ERR_PTR(status);
794 mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
795 if (IS_ERR(mr->umem)) {
799 num_pbes = ib_umem_page_count(mr->umem);
800 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
804 mr->hwmr.pbe_size = mr->umem->page_size;
805 mr->hwmr.fbo = mr->umem->offset;
806 mr->hwmr.va = usr_addr;
808 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
809 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
810 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
811 mr->hwmr.local_rd = 1;
812 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
813 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
816 build_user_pbes(dev, mr, num_pbes);
817 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
820 mr->ibmr.lkey = mr->hwmr.lkey;
821 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
822 mr->ibmr.rkey = mr->hwmr.lkey;
827 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
830 return ERR_PTR(status);
833 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
835 struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
836 struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
839 status = ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
841 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
843 /* it could be user registered memory. */
845 ib_umem_release(mr->umem);
850 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
851 struct ib_udata *udata,
852 struct ib_ucontext *ib_ctx)
855 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
856 struct ocrdma_create_cq_uresp uresp;
858 memset(&uresp, 0, sizeof(uresp));
859 uresp.cq_id = cq->id;
860 uresp.page_size = PAGE_ALIGN(cq->len);
862 uresp.max_hw_cqe = cq->max_hw_cqe;
863 uresp.page_addr[0] = cq->pa;
864 uresp.db_page_addr = ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
865 uresp.db_page_size = dev->nic_info.db_page_size;
866 uresp.phase_change = cq->phase_change ? 1 : 0;
867 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
869 pr_err("%s(%d) copy error cqid=0x%x.\n",
870 __func__, dev->id, cq->id);
873 status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
876 status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
878 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
886 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev, int entries, int vector,
887 struct ib_ucontext *ib_ctx,
888 struct ib_udata *udata)
890 struct ocrdma_cq *cq;
891 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
892 struct ocrdma_ucontext *uctx = NULL;
895 struct ocrdma_create_cq_ureq ureq;
898 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
899 return ERR_PTR(-EFAULT);
902 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
904 return ERR_PTR(-ENOMEM);
906 spin_lock_init(&cq->cq_lock);
907 spin_lock_init(&cq->comp_handler_lock);
908 INIT_LIST_HEAD(&cq->sq_head);
909 INIT_LIST_HEAD(&cq->rq_head);
910 cq->first_arm = true;
913 uctx = get_ocrdma_ucontext(ib_ctx);
914 pd_id = uctx->cntxt_pd->id;
917 status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
920 return ERR_PTR(status);
923 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
927 cq->phase = OCRDMA_CQE_VALID;
928 dev->cq_tbl[cq->id] = cq;
932 ocrdma_mbx_destroy_cq(dev, cq);
934 return ERR_PTR(status);
937 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
938 struct ib_udata *udata)
941 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
943 if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
951 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
957 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
958 struct ocrdma_cqe *cqe = NULL;
961 cqe_cnt = cq->cqe_cnt;
963 /* Last irq might have scheduled a polling thread
964 * sync-up with it before hard flushing.
966 spin_lock_irqsave(&cq->cq_lock, flags);
968 if (is_cqe_valid(cq, cqe))
973 ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
974 spin_unlock_irqrestore(&cq->cq_lock, flags);
977 int ocrdma_destroy_cq(struct ib_cq *ibcq)
980 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
981 struct ocrdma_eq *eq = NULL;
982 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
986 dev->cq_tbl[cq->id] = NULL;
987 indx = ocrdma_get_eq_table_index(dev, cq->eqn);
991 eq = &dev->eq_tbl[indx];
992 irq = ocrdma_get_irq(dev, eq);
993 synchronize_irq(irq);
996 status = ocrdma_mbx_destroy_cq(dev, cq);
998 pdid = cq->ucontext->cntxt_pd->id;
999 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1000 PAGE_ALIGN(cq->len));
1001 ocrdma_del_mmap(cq->ucontext,
1002 ocrdma_get_db_addr(dev, pdid),
1003 dev->nic_info.db_page_size);
1010 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1012 int status = -EINVAL;
1014 if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1015 dev->qp_tbl[qp->id] = qp;
1021 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1023 dev->qp_tbl[qp->id] = NULL;
1026 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1027 struct ib_qp_init_attr *attrs)
1029 if ((attrs->qp_type != IB_QPT_GSI) &&
1030 (attrs->qp_type != IB_QPT_RC) &&
1031 (attrs->qp_type != IB_QPT_UC) &&
1032 (attrs->qp_type != IB_QPT_UD)) {
1033 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1034 __func__, dev->id, attrs->qp_type);
1037 /* Skip the check for QP1 to support CM size of 128 */
1038 if ((attrs->qp_type != IB_QPT_GSI) &&
1039 (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1040 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1041 __func__, dev->id, attrs->cap.max_send_wr);
1042 pr_err("%s(%d) supported send_wr=0x%x\n",
1043 __func__, dev->id, dev->attr.max_wqe);
1046 if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1047 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1048 __func__, dev->id, attrs->cap.max_recv_wr);
1049 pr_err("%s(%d) supported recv_wr=0x%x\n",
1050 __func__, dev->id, dev->attr.max_rqe);
1053 if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1054 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1055 __func__, dev->id, attrs->cap.max_inline_data);
1056 pr_err("%s(%d) supported inline data size=0x%x\n",
1057 __func__, dev->id, dev->attr.max_inline_data);
1060 if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1061 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1062 __func__, dev->id, attrs->cap.max_send_sge);
1063 pr_err("%s(%d) supported send_sge=0x%x\n",
1064 __func__, dev->id, dev->attr.max_send_sge);
1067 if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1068 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1069 __func__, dev->id, attrs->cap.max_recv_sge);
1070 pr_err("%s(%d) supported recv_sge=0x%x\n",
1071 __func__, dev->id, dev->attr.max_recv_sge);
1074 /* unprivileged user space cannot create special QP */
1075 if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1077 ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1078 __func__, dev->id, attrs->qp_type);
1081 /* allow creating only one GSI type of QP */
1082 if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1083 pr_err("%s(%d) GSI special QPs already created.\n",
1087 /* verify consumer QPs are not trying to use GSI QP's CQ */
1088 if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1089 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1090 (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1091 pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1099 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1100 struct ib_udata *udata, int dpp_offset,
1101 int dpp_credit_lmt, int srq)
1105 struct ocrdma_create_qp_uresp uresp;
1106 struct ocrdma_dev *dev = qp->dev;
1107 struct ocrdma_pd *pd = qp->pd;
1109 memset(&uresp, 0, sizeof(uresp));
1110 usr_db = dev->nic_info.unmapped_db +
1111 (pd->id * dev->nic_info.db_page_size);
1112 uresp.qp_id = qp->id;
1113 uresp.sq_dbid = qp->sq.dbid;
1114 uresp.num_sq_pages = 1;
1115 uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1116 uresp.sq_page_addr[0] = qp->sq.pa;
1117 uresp.num_wqe_allocated = qp->sq.max_cnt;
1119 uresp.rq_dbid = qp->rq.dbid;
1120 uresp.num_rq_pages = 1;
1121 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1122 uresp.rq_page_addr[0] = qp->rq.pa;
1123 uresp.num_rqe_allocated = qp->rq.max_cnt;
1125 uresp.db_page_addr = usr_db;
1126 uresp.db_page_size = dev->nic_info.db_page_size;
1127 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1128 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1129 uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1131 if (qp->dpp_enabled) {
1132 uresp.dpp_credit = dpp_credit_lmt;
1133 uresp.dpp_offset = dpp_offset;
1135 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1137 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1140 status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1141 uresp.sq_page_size);
1146 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1147 uresp.rq_page_size);
1153 ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1158 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1159 struct ocrdma_pd *pd)
1161 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1162 qp->sq_db = dev->nic_info.db +
1163 (pd->id * dev->nic_info.db_page_size) +
1164 OCRDMA_DB_GEN2_SQ_OFFSET;
1165 qp->rq_db = dev->nic_info.db +
1166 (pd->id * dev->nic_info.db_page_size) +
1167 OCRDMA_DB_GEN2_RQ_OFFSET;
1169 qp->sq_db = dev->nic_info.db +
1170 (pd->id * dev->nic_info.db_page_size) +
1171 OCRDMA_DB_SQ_OFFSET;
1172 qp->rq_db = dev->nic_info.db +
1173 (pd->id * dev->nic_info.db_page_size) +
1174 OCRDMA_DB_RQ_OFFSET;
1178 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1181 kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1183 if (qp->wqe_wr_id_tbl == NULL)
1186 kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1187 if (qp->rqe_wr_id_tbl == NULL)
1193 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1194 struct ocrdma_pd *pd,
1195 struct ib_qp_init_attr *attrs)
1198 spin_lock_init(&qp->q_lock);
1199 INIT_LIST_HEAD(&qp->sq_entry);
1200 INIT_LIST_HEAD(&qp->rq_entry);
1202 qp->qp_type = attrs->qp_type;
1203 qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1204 qp->max_inline_data = attrs->cap.max_inline_data;
1205 qp->sq.max_sges = attrs->cap.max_send_sge;
1206 qp->rq.max_sges = attrs->cap.max_recv_sge;
1207 qp->state = OCRDMA_QPS_RST;
1208 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1211 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1212 struct ib_qp_init_attr *attrs)
1214 if (attrs->qp_type == IB_QPT_GSI) {
1215 dev->gsi_qp_created = 1;
1216 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1217 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1221 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1222 struct ib_qp_init_attr *attrs,
1223 struct ib_udata *udata)
1226 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1227 struct ocrdma_qp *qp;
1228 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1229 struct ocrdma_create_qp_ureq ureq;
1230 u16 dpp_credit_lmt, dpp_offset;
1232 status = ocrdma_check_qp_params(ibpd, dev, attrs);
1236 memset(&ureq, 0, sizeof(ureq));
1238 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1239 return ERR_PTR(-EFAULT);
1241 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1247 ocrdma_set_qp_init_params(qp, pd, attrs);
1249 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1250 OCRDMA_QP_FAST_REG);
1252 mutex_lock(&dev->dev_lock);
1253 status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1255 &dpp_offset, &dpp_credit_lmt);
1259 /* user space QP's wr_id table are managed in library */
1260 if (udata == NULL) {
1261 status = ocrdma_alloc_wr_id_tbl(qp);
1266 status = ocrdma_add_qpn_map(dev, qp);
1269 ocrdma_set_qp_db(dev, qp, pd);
1271 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1273 (attrs->srq != NULL));
1277 ocrdma_store_gsi_qp_cq(dev, attrs);
1278 qp->ibqp.qp_num = qp->id;
1279 mutex_unlock(&dev->dev_lock);
1283 ocrdma_del_qpn_map(dev, qp);
1285 ocrdma_mbx_destroy_qp(dev, qp);
1287 mutex_unlock(&dev->dev_lock);
1288 kfree(qp->wqe_wr_id_tbl);
1289 kfree(qp->rqe_wr_id_tbl);
1291 pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1293 return ERR_PTR(status);
1296 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1300 struct ocrdma_qp *qp;
1301 struct ocrdma_dev *dev;
1302 enum ib_qp_state old_qps;
1304 qp = get_ocrdma_qp(ibqp);
1306 if (attr_mask & IB_QP_STATE)
1307 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1308 /* if new and previous states are same hw doesn't need to
1313 status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1318 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1319 int attr_mask, struct ib_udata *udata)
1321 unsigned long flags;
1322 int status = -EINVAL;
1323 struct ocrdma_qp *qp;
1324 struct ocrdma_dev *dev;
1325 enum ib_qp_state old_qps, new_qps;
1327 qp = get_ocrdma_qp(ibqp);
1330 /* syncronize with multiple context trying to change, retrive qps */
1331 mutex_lock(&dev->dev_lock);
1332 /* syncronize with wqe, rqe posting and cqe processing contexts */
1333 spin_lock_irqsave(&qp->q_lock, flags);
1334 old_qps = get_ibqp_state(qp->state);
1335 if (attr_mask & IB_QP_STATE)
1336 new_qps = attr->qp_state;
1339 spin_unlock_irqrestore(&qp->q_lock, flags);
1341 if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1342 IB_LINK_LAYER_ETHERNET)) {
1343 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1344 "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1345 __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1350 status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1354 mutex_unlock(&dev->dev_lock);
1358 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1376 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1378 int ib_qp_acc_flags = 0;
1380 if (qp_cap_flags & OCRDMA_QP_INB_WR)
1381 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1382 if (qp_cap_flags & OCRDMA_QP_INB_RD)
1383 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1384 return ib_qp_acc_flags;
1387 int ocrdma_query_qp(struct ib_qp *ibqp,
1388 struct ib_qp_attr *qp_attr,
1389 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1393 struct ocrdma_qp_params params;
1394 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1395 struct ocrdma_dev *dev = qp->dev;
1397 memset(¶ms, 0, sizeof(params));
1398 mutex_lock(&dev->dev_lock);
1399 status = ocrdma_mbx_query_qp(dev, qp, ¶ms);
1400 mutex_unlock(&dev->dev_lock);
1403 qp_attr->qp_state = get_ibqp_state(IB_QPS_INIT);
1404 qp_attr->cur_qp_state = get_ibqp_state(IB_QPS_INIT);
1406 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1407 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1408 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1409 qp_attr->path_mig_state = IB_MIG_MIGRATED;
1410 qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1411 qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1412 qp_attr->dest_qp_num =
1413 params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1415 qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1416 qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1417 qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1418 qp_attr->cap.max_send_sge = qp->sq.max_sges;
1419 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1420 qp_attr->cap.max_inline_data = qp->max_inline_data;
1421 qp_init_attr->cap = qp_attr->cap;
1422 memcpy(&qp_attr->ah_attr.grh.dgid, ¶ms.dgid[0],
1423 sizeof(params.dgid));
1424 qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1425 OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1426 qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1427 qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1428 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1429 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1430 qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1431 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1432 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1434 qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1435 qp_attr->ah_attr.port_num = 1;
1436 qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1437 OCRDMA_QP_PARAMS_SL_MASK) >>
1438 OCRDMA_QP_PARAMS_SL_SHIFT;
1439 qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1440 OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1441 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1442 qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1443 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1444 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1445 qp_attr->retry_cnt =
1446 (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1447 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1448 qp_attr->min_rnr_timer = 0;
1449 qp_attr->pkey_index = 0;
1450 qp_attr->port_num = 1;
1451 qp_attr->ah_attr.src_path_bits = 0;
1452 qp_attr->ah_attr.static_rate = 0;
1453 qp_attr->alt_pkey_index = 0;
1454 qp_attr->alt_port_num = 0;
1455 qp_attr->alt_timeout = 0;
1456 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1457 qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1458 OCRDMA_QP_PARAMS_STATE_SHIFT;
1459 qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1460 qp_attr->max_dest_rd_atomic =
1461 params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1462 qp_attr->max_rd_atomic =
1463 params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1464 qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1465 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1470 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, int idx)
1473 unsigned int mask = (1 << (idx % 32));
1475 if (srq->idx_bit_fields[i] & mask)
1476 srq->idx_bit_fields[i] &= ~mask;
1478 srq->idx_bit_fields[i] |= mask;
1481 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1483 return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1486 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1488 return (qp->sq.tail == qp->sq.head);
1491 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1493 return (qp->rq.tail == qp->rq.head);
1496 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1498 return q->va + (q->head * q->entry_size);
1501 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1504 return q->va + (idx * q->entry_size);
1507 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1509 q->head = (q->head + 1) & q->max_wqe_idx;
1512 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1514 q->tail = (q->tail + 1) & q->max_wqe_idx;
1517 /* discard the cqe for a given QP */
1518 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1520 unsigned long cq_flags;
1521 unsigned long flags;
1522 int discard_cnt = 0;
1523 u32 cur_getp, stop_getp;
1524 struct ocrdma_cqe *cqe;
1525 u32 qpn = 0, wqe_idx = 0;
1527 spin_lock_irqsave(&cq->cq_lock, cq_flags);
1529 /* traverse through the CQEs in the hw CQ,
1530 * find the matching CQE for a given qp,
1531 * mark the matching one discarded by clearing qpn.
1532 * ring the doorbell in the poll_cq() as
1533 * we don't complete out of order cqe.
1536 cur_getp = cq->getp;
1537 /* find upto when do we reap the cq. */
1538 stop_getp = cur_getp;
1540 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1543 cqe = cq->va + cur_getp;
1544 /* if (a) done reaping whole hw cq, or
1545 * (b) qp_xq becomes empty.
1548 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1549 /* if previously discarded cqe found, skip that too. */
1550 /* check for matching qp */
1551 if (qpn == 0 || qpn != qp->id)
1554 if (is_cqe_for_sq(cqe)) {
1555 ocrdma_hwq_inc_tail(&qp->sq);
1558 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1559 OCRDMA_CQE_BUFTAG_SHIFT) &
1560 qp->srq->rq.max_wqe_idx;
1563 spin_lock_irqsave(&qp->srq->q_lock, flags);
1564 ocrdma_hwq_inc_tail(&qp->srq->rq);
1565 ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1566 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1569 ocrdma_hwq_inc_tail(&qp->rq);
1572 /* mark cqe discarded so that it is not picked up later
1578 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1579 } while (cur_getp != stop_getp);
1580 spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1583 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1586 unsigned long flags;
1587 struct ocrdma_dev *dev = qp->dev;
1588 /* sync with any active CQ poll */
1590 spin_lock_irqsave(&dev->flush_q_lock, flags);
1591 found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1593 list_del(&qp->sq_entry);
1595 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1597 list_del(&qp->rq_entry);
1599 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1602 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1605 struct ocrdma_pd *pd;
1606 struct ocrdma_qp *qp;
1607 struct ocrdma_dev *dev;
1608 struct ib_qp_attr attrs;
1609 int attr_mask = IB_QP_STATE;
1610 unsigned long flags;
1612 qp = get_ocrdma_qp(ibqp);
1615 attrs.qp_state = IB_QPS_ERR;
1618 /* change the QP state to ERROR */
1619 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1621 /* ensure that CQEs for newly created QP (whose id may be same with
1622 * one which just getting destroyed are same), dont get
1623 * discarded until the old CQEs are discarded.
1625 mutex_lock(&dev->dev_lock);
1626 status = ocrdma_mbx_destroy_qp(dev, qp);
1629 * acquire CQ lock while destroy is in progress, in order to
1630 * protect against proessing in-flight CQEs for this QP.
1632 spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1633 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1634 spin_lock(&qp->rq_cq->cq_lock);
1636 ocrdma_del_qpn_map(dev, qp);
1638 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1639 spin_unlock(&qp->rq_cq->cq_lock);
1640 spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1643 ocrdma_discard_cqes(qp, qp->sq_cq);
1644 ocrdma_discard_cqes(qp, qp->rq_cq);
1646 mutex_unlock(&dev->dev_lock);
1649 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1650 PAGE_ALIGN(qp->sq.len));
1652 ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1653 PAGE_ALIGN(qp->rq.len));
1656 ocrdma_del_flush_qp(qp);
1658 kfree(qp->wqe_wr_id_tbl);
1659 kfree(qp->rqe_wr_id_tbl);
1664 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1665 struct ib_udata *udata)
1668 struct ocrdma_create_srq_uresp uresp;
1670 memset(&uresp, 0, sizeof(uresp));
1671 uresp.rq_dbid = srq->rq.dbid;
1672 uresp.num_rq_pages = 1;
1673 uresp.rq_page_addr[0] = srq->rq.pa;
1674 uresp.rq_page_size = srq->rq.len;
1675 uresp.db_page_addr = dev->nic_info.unmapped_db +
1676 (srq->pd->id * dev->nic_info.db_page_size);
1677 uresp.db_page_size = dev->nic_info.db_page_size;
1678 uresp.num_rqe_allocated = srq->rq.max_cnt;
1679 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1680 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1681 uresp.db_shift = 24;
1683 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1684 uresp.db_shift = 16;
1687 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1690 status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1691 uresp.rq_page_size);
1697 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1698 struct ib_srq_init_attr *init_attr,
1699 struct ib_udata *udata)
1701 int status = -ENOMEM;
1702 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1703 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1704 struct ocrdma_srq *srq;
1706 if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1707 return ERR_PTR(-EINVAL);
1708 if (init_attr->attr.max_wr > dev->attr.max_rqe)
1709 return ERR_PTR(-EINVAL);
1711 srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1713 return ERR_PTR(status);
1715 spin_lock_init(&srq->q_lock);
1717 srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1718 status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1722 if (udata == NULL) {
1723 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1725 if (srq->rqe_wr_id_tbl == NULL)
1728 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1729 (srq->rq.max_cnt % 32 ? 1 : 0);
1730 srq->idx_bit_fields =
1731 kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1732 if (srq->idx_bit_fields == NULL)
1734 memset(srq->idx_bit_fields, 0xff,
1735 srq->bit_fields_len * sizeof(u32));
1738 if (init_attr->attr.srq_limit) {
1739 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1745 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1753 ocrdma_mbx_destroy_srq(dev, srq);
1755 kfree(srq->rqe_wr_id_tbl);
1756 kfree(srq->idx_bit_fields);
1758 return ERR_PTR(status);
1761 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1762 struct ib_srq_attr *srq_attr,
1763 enum ib_srq_attr_mask srq_attr_mask,
1764 struct ib_udata *udata)
1767 struct ocrdma_srq *srq;
1769 srq = get_ocrdma_srq(ibsrq);
1770 if (srq_attr_mask & IB_SRQ_MAX_WR)
1773 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1777 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1780 struct ocrdma_srq *srq;
1782 srq = get_ocrdma_srq(ibsrq);
1783 status = ocrdma_mbx_query_srq(srq, srq_attr);
1787 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1790 struct ocrdma_srq *srq;
1791 struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1793 srq = get_ocrdma_srq(ibsrq);
1795 status = ocrdma_mbx_destroy_srq(dev, srq);
1798 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1799 PAGE_ALIGN(srq->rq.len));
1801 kfree(srq->idx_bit_fields);
1802 kfree(srq->rqe_wr_id_tbl);
1807 /* unprivileged verbs and their support functions. */
1808 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1809 struct ocrdma_hdr_wqe *hdr,
1810 struct ib_send_wr *wr)
1812 struct ocrdma_ewqe_ud_hdr *ud_hdr =
1813 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1814 struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
1816 ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
1817 if (qp->qp_type == IB_QPT_GSI)
1818 ud_hdr->qkey = qp->qkey;
1820 ud_hdr->qkey = wr->wr.ud.remote_qkey;
1821 ud_hdr->rsvd_ahid = ah->id;
1824 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1825 struct ocrdma_sge *sge, int num_sge,
1826 struct ib_sge *sg_list)
1830 for (i = 0; i < num_sge; i++) {
1831 sge[i].lrkey = sg_list[i].lkey;
1832 sge[i].addr_lo = sg_list[i].addr;
1833 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1834 sge[i].len = sg_list[i].length;
1835 hdr->total_len += sg_list[i].length;
1838 memset(sge, 0, sizeof(*sge));
1841 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1843 uint32_t total_len = 0, i;
1845 for (i = 0; i < num_sge; i++)
1846 total_len += sg_list[i].length;
1851 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1852 struct ocrdma_hdr_wqe *hdr,
1853 struct ocrdma_sge *sge,
1854 struct ib_send_wr *wr, u32 wqe_size)
1859 if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1860 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1861 if (unlikely(hdr->total_len > qp->max_inline_data)) {
1862 pr_err("%s() supported_len=0x%x,\n"
1863 " unspported len req=0x%x\n", __func__,
1864 qp->max_inline_data, hdr->total_len);
1867 dpp_addr = (char *)sge;
1868 for (i = 0; i < wr->num_sge; i++) {
1870 (void *)(unsigned long)wr->sg_list[i].addr,
1871 wr->sg_list[i].length);
1872 dpp_addr += wr->sg_list[i].length;
1875 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1876 if (0 == hdr->total_len)
1877 wqe_size += sizeof(struct ocrdma_sge);
1878 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1880 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1882 wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1884 wqe_size += sizeof(struct ocrdma_sge);
1885 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1887 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1891 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1892 struct ib_send_wr *wr)
1895 struct ocrdma_sge *sge;
1896 u32 wqe_size = sizeof(*hdr);
1898 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1899 ocrdma_build_ud_hdr(qp, hdr, wr);
1900 sge = (struct ocrdma_sge *)(hdr + 2);
1901 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1903 sge = (struct ocrdma_sge *)(hdr + 1);
1906 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1910 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1911 struct ib_send_wr *wr)
1914 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1915 struct ocrdma_sge *sge = ext_rw + 1;
1916 u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1918 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1921 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1922 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1923 ext_rw->lrkey = wr->wr.rdma.rkey;
1924 ext_rw->len = hdr->total_len;
1928 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1929 struct ib_send_wr *wr)
1931 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1932 struct ocrdma_sge *sge = ext_rw + 1;
1933 u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
1934 sizeof(struct ocrdma_hdr_wqe);
1936 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1937 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1938 hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
1939 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1941 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1942 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1943 ext_rw->lrkey = wr->wr.rdma.rkey;
1944 ext_rw->len = hdr->total_len;
1947 static void build_frmr_pbes(struct ib_send_wr *wr, struct ocrdma_pbl *pbl_tbl,
1948 struct ocrdma_hw_mr *hwmr)
1953 struct ocrdma_pbe *pbe;
1955 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1958 /* go through the OS phy regions & fill hw pbe entries into pbls. */
1959 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
1960 /* number of pbes can be more for one OS buf, when
1961 * buffers are of different sizes.
1962 * split the ib_buf to one or more pbes.
1964 buf_addr = wr->wr.fast_reg.page_list->page_list[i];
1965 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
1966 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
1970 /* if the pbl is full storing the pbes,
1973 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
1975 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1981 static int get_encoded_page_size(int pg_sz)
1983 /* Max size is 256M 4096 << 16 */
1986 if (pg_sz == (4096 << i))
1992 static int ocrdma_build_fr(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1993 struct ib_send_wr *wr)
1996 struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
1997 struct ocrdma_mr *mr;
1998 u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2000 wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2002 if (wr->wr.fast_reg.page_list_len > qp->dev->attr.max_pages_per_frmr)
2005 hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2006 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2008 if (wr->wr.fast_reg.page_list_len == 0)
2010 if (wr->wr.fast_reg.access_flags & IB_ACCESS_LOCAL_WRITE)
2011 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2012 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_WRITE)
2013 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2014 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_READ)
2015 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2016 hdr->lkey = wr->wr.fast_reg.rkey;
2017 hdr->total_len = wr->wr.fast_reg.length;
2019 fbo = wr->wr.fast_reg.iova_start -
2020 (wr->wr.fast_reg.page_list->page_list[0] & PAGE_MASK);
2022 fast_reg->va_hi = upper_32_bits(wr->wr.fast_reg.iova_start);
2023 fast_reg->va_lo = (u32) (wr->wr.fast_reg.iova_start & 0xffffffff);
2024 fast_reg->fbo_hi = upper_32_bits(fbo);
2025 fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2026 fast_reg->num_sges = wr->wr.fast_reg.page_list_len;
2027 fast_reg->size_sge =
2028 get_encoded_page_size(1 << wr->wr.fast_reg.page_shift);
2029 mr = (struct ocrdma_mr *) (unsigned long)
2030 qp->dev->stag_arr[(hdr->lkey >> 8) & (OCRDMA_MAX_STAG - 1)];
2031 build_frmr_pbes(wr, mr->hwmr.pbl_table, &mr->hwmr);
2035 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2037 u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2039 iowrite32(val, qp->sq_db);
2042 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2043 struct ib_send_wr **bad_wr)
2046 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2047 struct ocrdma_hdr_wqe *hdr;
2048 unsigned long flags;
2050 spin_lock_irqsave(&qp->q_lock, flags);
2051 if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2052 spin_unlock_irqrestore(&qp->q_lock, flags);
2058 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2059 wr->num_sge > qp->sq.max_sges) {
2064 hdr = ocrdma_hwq_head(&qp->sq);
2066 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2067 hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2068 if (wr->send_flags & IB_SEND_FENCE)
2070 (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2071 if (wr->send_flags & IB_SEND_SOLICITED)
2073 (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2075 switch (wr->opcode) {
2076 case IB_WR_SEND_WITH_IMM:
2077 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2078 hdr->immdt = ntohl(wr->ex.imm_data);
2080 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2081 ocrdma_build_send(qp, hdr, wr);
2083 case IB_WR_SEND_WITH_INV:
2084 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2085 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2086 hdr->lkey = wr->ex.invalidate_rkey;
2087 status = ocrdma_build_send(qp, hdr, wr);
2089 case IB_WR_RDMA_WRITE_WITH_IMM:
2090 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2091 hdr->immdt = ntohl(wr->ex.imm_data);
2092 case IB_WR_RDMA_WRITE:
2093 hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2094 status = ocrdma_build_write(qp, hdr, wr);
2096 case IB_WR_RDMA_READ_WITH_INV:
2097 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2098 case IB_WR_RDMA_READ:
2099 ocrdma_build_read(qp, hdr, wr);
2101 case IB_WR_LOCAL_INV:
2103 (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2104 hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2105 sizeof(struct ocrdma_sge)) /
2106 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2107 hdr->lkey = wr->ex.invalidate_rkey;
2109 case IB_WR_FAST_REG_MR:
2110 status = ocrdma_build_fr(qp, hdr, wr);
2120 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2121 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2123 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2124 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2125 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2126 OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2127 /* make sure wqe is written before adapter can access it */
2129 /* inform hw to start processing it */
2130 ocrdma_ring_sq_db(qp);
2132 /* update pointer, counter for next wr */
2133 ocrdma_hwq_inc_head(&qp->sq);
2136 spin_unlock_irqrestore(&qp->q_lock, flags);
2140 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2142 u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2144 iowrite32(val, qp->rq_db);
2147 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2151 struct ocrdma_sge *sge;
2153 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2155 wqe_size = sizeof(*sge) + sizeof(*rqe);
2157 rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2158 OCRDMA_WQE_SIZE_SHIFT);
2159 rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2160 rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2162 rqe->rsvd_tag = tag;
2163 sge = (struct ocrdma_sge *)(rqe + 1);
2164 ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2165 ocrdma_cpu_to_le32(rqe, wqe_size);
2168 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2169 struct ib_recv_wr **bad_wr)
2172 unsigned long flags;
2173 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2174 struct ocrdma_hdr_wqe *rqe;
2176 spin_lock_irqsave(&qp->q_lock, flags);
2177 if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2178 spin_unlock_irqrestore(&qp->q_lock, flags);
2183 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2184 wr->num_sge > qp->rq.max_sges) {
2189 rqe = ocrdma_hwq_head(&qp->rq);
2190 ocrdma_build_rqe(rqe, wr, 0);
2192 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2193 /* make sure rqe is written before adapter can access it */
2196 /* inform hw to start processing it */
2197 ocrdma_ring_rq_db(qp);
2199 /* update pointer, counter for next wr */
2200 ocrdma_hwq_inc_head(&qp->rq);
2203 spin_unlock_irqrestore(&qp->q_lock, flags);
2207 /* cqe for srq's rqe can potentially arrive out of order.
2208 * index gives the entry in the shadow table where to store
2209 * the wr_id. tag/index is returned in cqe to reference back
2212 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2217 for (row = 0; row < srq->bit_fields_len; row++) {
2218 if (srq->idx_bit_fields[row]) {
2219 indx = ffs(srq->idx_bit_fields[row]);
2220 indx = (row * 32) + (indx - 1);
2221 if (indx >= srq->rq.max_cnt)
2223 ocrdma_srq_toggle_bit(srq, indx);
2228 if (row == srq->bit_fields_len)
2230 return indx + 1; /* Use from index 1 */
2233 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2235 u32 val = srq->rq.dbid | (1 << 16);
2237 iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2240 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2241 struct ib_recv_wr **bad_wr)
2244 unsigned long flags;
2245 struct ocrdma_srq *srq;
2246 struct ocrdma_hdr_wqe *rqe;
2249 srq = get_ocrdma_srq(ibsrq);
2251 spin_lock_irqsave(&srq->q_lock, flags);
2253 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2254 wr->num_sge > srq->rq.max_sges) {
2259 tag = ocrdma_srq_get_idx(srq);
2260 rqe = ocrdma_hwq_head(&srq->rq);
2261 ocrdma_build_rqe(rqe, wr, tag);
2263 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2264 /* make sure rqe is written before adapter can perform DMA */
2266 /* inform hw to start processing it */
2267 ocrdma_ring_srq_db(srq);
2268 /* update pointer, counter for next wr */
2269 ocrdma_hwq_inc_head(&srq->rq);
2272 spin_unlock_irqrestore(&srq->q_lock, flags);
2276 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2278 enum ib_wc_status ibwc_status;
2281 case OCRDMA_CQE_GENERAL_ERR:
2282 ibwc_status = IB_WC_GENERAL_ERR;
2284 case OCRDMA_CQE_LOC_LEN_ERR:
2285 ibwc_status = IB_WC_LOC_LEN_ERR;
2287 case OCRDMA_CQE_LOC_QP_OP_ERR:
2288 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2290 case OCRDMA_CQE_LOC_EEC_OP_ERR:
2291 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2293 case OCRDMA_CQE_LOC_PROT_ERR:
2294 ibwc_status = IB_WC_LOC_PROT_ERR;
2296 case OCRDMA_CQE_WR_FLUSH_ERR:
2297 ibwc_status = IB_WC_WR_FLUSH_ERR;
2299 case OCRDMA_CQE_MW_BIND_ERR:
2300 ibwc_status = IB_WC_MW_BIND_ERR;
2302 case OCRDMA_CQE_BAD_RESP_ERR:
2303 ibwc_status = IB_WC_BAD_RESP_ERR;
2305 case OCRDMA_CQE_LOC_ACCESS_ERR:
2306 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2308 case OCRDMA_CQE_REM_INV_REQ_ERR:
2309 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2311 case OCRDMA_CQE_REM_ACCESS_ERR:
2312 ibwc_status = IB_WC_REM_ACCESS_ERR;
2314 case OCRDMA_CQE_REM_OP_ERR:
2315 ibwc_status = IB_WC_REM_OP_ERR;
2317 case OCRDMA_CQE_RETRY_EXC_ERR:
2318 ibwc_status = IB_WC_RETRY_EXC_ERR;
2320 case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2321 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2323 case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2324 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2326 case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2327 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2329 case OCRDMA_CQE_REM_ABORT_ERR:
2330 ibwc_status = IB_WC_REM_ABORT_ERR;
2332 case OCRDMA_CQE_INV_EECN_ERR:
2333 ibwc_status = IB_WC_INV_EECN_ERR;
2335 case OCRDMA_CQE_INV_EEC_STATE_ERR:
2336 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2338 case OCRDMA_CQE_FATAL_ERR:
2339 ibwc_status = IB_WC_FATAL_ERR;
2341 case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2342 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2345 ibwc_status = IB_WC_GENERAL_ERR;
2351 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2354 struct ocrdma_hdr_wqe *hdr;
2355 struct ocrdma_sge *rw;
2358 hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2360 ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2361 /* Undo the hdr->cw swap */
2362 opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2365 ibwc->opcode = IB_WC_RDMA_WRITE;
2368 rw = (struct ocrdma_sge *)(hdr + 1);
2369 ibwc->opcode = IB_WC_RDMA_READ;
2370 ibwc->byte_len = rw->len;
2373 ibwc->opcode = IB_WC_SEND;
2376 ibwc->opcode = IB_WC_FAST_REG_MR;
2378 case OCRDMA_LKEY_INV:
2379 ibwc->opcode = IB_WC_LOCAL_INV;
2382 ibwc->status = IB_WC_GENERAL_ERR;
2383 pr_err("%s() invalid opcode received = 0x%x\n",
2384 __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2389 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2390 struct ocrdma_cqe *cqe)
2392 if (is_cqe_for_sq(cqe)) {
2393 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2394 cqe->flags_status_srcqpn) &
2395 ~OCRDMA_CQE_STATUS_MASK);
2396 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2397 cqe->flags_status_srcqpn) |
2398 (OCRDMA_CQE_WR_FLUSH_ERR <<
2399 OCRDMA_CQE_STATUS_SHIFT));
2401 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2402 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2403 cqe->flags_status_srcqpn) &
2404 ~OCRDMA_CQE_UD_STATUS_MASK);
2405 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2406 cqe->flags_status_srcqpn) |
2407 (OCRDMA_CQE_WR_FLUSH_ERR <<
2408 OCRDMA_CQE_UD_STATUS_SHIFT));
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));
2421 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2422 struct ocrdma_qp *qp, int status)
2424 bool expand = false;
2427 ibwc->qp = &qp->ibqp;
2428 ibwc->status = ocrdma_to_ibwc_err(status);
2430 ocrdma_flush_qp(qp);
2431 ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2433 /* if wqe/rqe pending for which cqe needs to be returned,
2434 * trigger inflating it.
2436 if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2438 ocrdma_set_cqe_status_flushed(qp, cqe);
2443 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2444 struct ocrdma_qp *qp, int status)
2446 ibwc->opcode = IB_WC_RECV;
2447 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2448 ocrdma_hwq_inc_tail(&qp->rq);
2450 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2453 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2454 struct ocrdma_qp *qp, int status)
2456 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2457 ocrdma_hwq_inc_tail(&qp->sq);
2459 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2463 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2464 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2465 bool *polled, bool *stop)
2468 int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2469 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2471 /* when hw sq is empty, but rq is not empty, so we continue
2472 * to keep the cqe in order to get the cq event again.
2474 if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2475 /* when cq for rq and sq is same, it is safe to return
2476 * flush cqe for RQEs.
2478 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2480 status = OCRDMA_CQE_WR_FLUSH_ERR;
2481 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2483 /* stop processing further cqe as this cqe is used for
2484 * triggering cq event on buddy cq of RQ.
2485 * When QP is destroyed, this cqe will be removed
2486 * from the cq's hardware q.
2494 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2499 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2500 struct ocrdma_cqe *cqe,
2501 struct ib_wc *ibwc, bool *polled)
2503 bool expand = false;
2504 int tail = qp->sq.tail;
2507 if (!qp->wqe_wr_id_tbl[tail].signaled) {
2508 *polled = false; /* WC cannot be consumed yet */
2510 ibwc->status = IB_WC_SUCCESS;
2512 ibwc->qp = &qp->ibqp;
2513 ocrdma_update_wc(qp, ibwc, tail);
2516 wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2517 OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2518 if (tail != wqe_idx)
2519 expand = true; /* Coalesced CQE can't be consumed yet */
2521 ocrdma_hwq_inc_tail(&qp->sq);
2525 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2526 struct ib_wc *ibwc, bool *polled, bool *stop)
2531 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2532 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2534 if (status == OCRDMA_CQE_SUCCESS)
2535 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2537 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2541 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2545 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2546 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2547 ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2548 OCRDMA_CQE_SRCQP_MASK;
2549 ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2550 OCRDMA_CQE_PKEY_MASK;
2551 ibwc->wc_flags = IB_WC_GRH;
2552 ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2553 OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2557 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2558 struct ocrdma_cqe *cqe,
2559 struct ocrdma_qp *qp)
2561 unsigned long flags;
2562 struct ocrdma_srq *srq;
2565 srq = get_ocrdma_srq(qp->ibqp.srq);
2566 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2567 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2571 ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2572 spin_lock_irqsave(&srq->q_lock, flags);
2573 ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2574 spin_unlock_irqrestore(&srq->q_lock, flags);
2575 ocrdma_hwq_inc_tail(&srq->rq);
2578 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2579 struct ib_wc *ibwc, bool *polled, bool *stop,
2584 /* when hw_rq is empty, but wq is not empty, so continue
2585 * to keep the cqe to get the cq event again.
2587 if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2588 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2590 status = OCRDMA_CQE_WR_FLUSH_ERR;
2591 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2599 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2604 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2605 struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2607 ibwc->opcode = IB_WC_RECV;
2608 ibwc->qp = &qp->ibqp;
2609 ibwc->status = IB_WC_SUCCESS;
2611 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2612 ocrdma_update_ud_rcqe(ibwc, cqe);
2614 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2616 if (is_cqe_imm(cqe)) {
2617 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2618 ibwc->wc_flags |= IB_WC_WITH_IMM;
2619 } else if (is_cqe_wr_imm(cqe)) {
2620 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2621 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2622 ibwc->wc_flags |= IB_WC_WITH_IMM;
2623 } else if (is_cqe_invalidated(cqe)) {
2624 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2625 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2628 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2630 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2631 ocrdma_hwq_inc_tail(&qp->rq);
2635 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2636 struct ib_wc *ibwc, bool *polled, bool *stop)
2639 bool expand = false;
2642 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2643 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2644 OCRDMA_CQE_UD_STATUS_MASK) >>
2645 OCRDMA_CQE_UD_STATUS_SHIFT;
2647 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2648 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2651 if (status == OCRDMA_CQE_SUCCESS) {
2653 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2655 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2661 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2664 if (cq->phase_change) {
2666 cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2668 /* clear valid bit */
2669 cqe->flags_status_srcqpn = 0;
2673 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2678 bool expand = false;
2679 int polled_hw_cqes = 0;
2680 struct ocrdma_qp *qp = NULL;
2681 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2682 struct ocrdma_cqe *cqe;
2683 u16 cur_getp; bool polled = false; bool stop = false;
2685 cur_getp = cq->getp;
2686 while (num_entries) {
2687 cqe = cq->va + cur_getp;
2688 /* check whether valid cqe or not */
2689 if (!is_cqe_valid(cq, cqe))
2691 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2692 /* ignore discarded cqe */
2695 qp = dev->qp_tbl[qpn];
2698 if (is_cqe_for_sq(cqe)) {
2699 expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2702 expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2709 /* clear qpn to avoid duplicate processing by discard_cqe() */
2712 polled_hw_cqes += 1;
2713 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2714 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2724 cq->getp = cur_getp;
2725 if (cq->deferred_arm) {
2726 ocrdma_ring_cq_db(dev, cq->id, true, cq->deferred_sol,
2728 cq->deferred_arm = false;
2729 cq->deferred_sol = false;
2731 /* We need to pop the CQE. No need to arm */
2732 ocrdma_ring_cq_db(dev, cq->id, false, cq->deferred_sol,
2734 cq->deferred_sol = false;
2740 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2741 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2742 struct ocrdma_qp *qp, struct ib_wc *ibwc)
2746 while (num_entries) {
2747 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2749 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2750 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2751 ocrdma_hwq_inc_tail(&qp->sq);
2752 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2753 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2754 ocrdma_hwq_inc_tail(&qp->rq);
2759 ibwc->status = IB_WC_WR_FLUSH_ERR;
2767 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2769 int cqes_to_poll = num_entries;
2770 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2771 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2772 int num_os_cqe = 0, err_cqes = 0;
2773 struct ocrdma_qp *qp;
2774 unsigned long flags;
2776 /* poll cqes from adapter CQ */
2777 spin_lock_irqsave(&cq->cq_lock, flags);
2778 num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2779 spin_unlock_irqrestore(&cq->cq_lock, flags);
2780 cqes_to_poll -= num_os_cqe;
2783 wc = wc + num_os_cqe;
2784 /* adapter returns single error cqe when qp moves to
2785 * error state. So insert error cqes with wc_status as
2786 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2787 * respectively which uses this CQ.
2789 spin_lock_irqsave(&dev->flush_q_lock, flags);
2790 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2791 if (cqes_to_poll == 0)
2793 err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2794 cqes_to_poll -= err_cqes;
2795 num_os_cqe += err_cqes;
2798 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2803 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2805 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2806 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2808 unsigned long flags;
2809 bool arm_needed = false, sol_needed = false;
2813 spin_lock_irqsave(&cq->cq_lock, flags);
2814 if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2816 if (cq_flags & IB_CQ_SOLICITED)
2819 if (cq->first_arm) {
2820 ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2821 cq->first_arm = false;
2824 cq->deferred_arm = true;
2827 cq->deferred_sol = sol_needed;
2828 spin_unlock_irqrestore(&cq->cq_lock, flags);
2833 struct ib_mr *ocrdma_alloc_frmr(struct ib_pd *ibpd, int max_page_list_len)
2836 struct ocrdma_mr *mr;
2837 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2838 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2840 if (max_page_list_len > dev->attr.max_pages_per_frmr)
2841 return ERR_PTR(-EINVAL);
2843 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2845 return ERR_PTR(-ENOMEM);
2847 status = ocrdma_get_pbl_info(dev, mr, max_page_list_len);
2851 mr->hwmr.remote_rd = 0;
2852 mr->hwmr.remote_wr = 0;
2853 mr->hwmr.local_rd = 0;
2854 mr->hwmr.local_wr = 0;
2855 mr->hwmr.mw_bind = 0;
2856 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2859 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2862 mr->ibmr.rkey = mr->hwmr.lkey;
2863 mr->ibmr.lkey = mr->hwmr.lkey;
2864 dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
2868 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2871 return ERR_PTR(-ENOMEM);
2874 struct ib_fast_reg_page_list *ocrdma_alloc_frmr_page_list(struct ib_device
2878 struct ib_fast_reg_page_list *frmr_list;
2881 size = sizeof(*frmr_list) + (page_list_len * sizeof(u64));
2882 frmr_list = kzalloc(size, GFP_KERNEL);
2884 return ERR_PTR(-ENOMEM);
2885 frmr_list->page_list = (u64 *)(frmr_list + 1);
2889 void ocrdma_free_frmr_page_list(struct ib_fast_reg_page_list *page_list)
2894 #define MAX_KERNEL_PBE_SIZE 65536
2895 static inline int count_kernel_pbes(struct ib_phys_buf *buf_list,
2896 int buf_cnt, u32 *pbe_size)
2901 *pbe_size = roundup(buf_list[0].size, PAGE_SIZE);
2902 *pbe_size = roundup_pow_of_two(*pbe_size);
2904 /* find the smallest PBE size that we can have */
2905 for (i = 0; i < buf_cnt; i++) {
2906 /* first addr may not be page aligned, so ignore checking */
2907 if ((i != 0) && ((buf_list[i].addr & ~PAGE_MASK) ||
2908 (buf_list[i].size & ~PAGE_MASK))) {
2912 /* if configured PBE size is greater then the chosen one,
2913 * reduce the PBE size.
2915 buf_size = roundup(buf_list[i].size, PAGE_SIZE);
2916 /* pbe_size has to be even multiple of 4K 1,2,4,8...*/
2917 buf_size = roundup_pow_of_two(buf_size);
2918 if (*pbe_size > buf_size)
2919 *pbe_size = buf_size;
2921 total_size += buf_size;
2923 *pbe_size = *pbe_size > MAX_KERNEL_PBE_SIZE ?
2924 (MAX_KERNEL_PBE_SIZE) : (*pbe_size);
2926 /* num_pbes = total_size / (*pbe_size); this is implemented below. */
2928 return total_size >> ilog2(*pbe_size);
2931 static void build_kernel_pbes(struct ib_phys_buf *buf_list, int ib_buf_cnt,
2932 u32 pbe_size, struct ocrdma_pbl *pbl_tbl,
2933 struct ocrdma_hw_mr *hwmr)
2937 int pbes_per_buf = 0;
2940 struct ocrdma_pbe *pbe;
2941 int total_num_pbes = 0;
2943 if (!hwmr->num_pbes)
2946 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2949 /* go through the OS phy regions & fill hw pbe entries into pbls. */
2950 for (i = 0; i < ib_buf_cnt; i++) {
2951 buf_addr = buf_list[i].addr;
2953 roundup_pow_of_two(roundup(buf_list[i].size, PAGE_SIZE)) /
2955 hwmr->len += buf_list[i].size;
2956 /* number of pbes can be more for one OS buf, when
2957 * buffers are of different sizes.
2958 * split the ib_buf to one or more pbes.
2960 for (idx = 0; idx < pbes_per_buf; idx++) {
2961 /* we program always page aligned addresses,
2962 * first unaligned address is taken care by fbo.
2965 /* for non zero fbo, assign the
2966 * start of the page.
2969 cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2971 cpu_to_le32((u32) upper_32_bits(buf_addr));
2974 cpu_to_le32((u32) (buf_addr & 0xffffffff));
2976 cpu_to_le32((u32) upper_32_bits(buf_addr));
2978 buf_addr += pbe_size;
2980 total_num_pbes += 1;
2983 if (total_num_pbes == hwmr->num_pbes)
2985 /* if the pbl is full storing the pbes,
2988 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
2990 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2999 struct ib_mr *ocrdma_reg_kernel_mr(struct ib_pd *ibpd,
3000 struct ib_phys_buf *buf_list,
3001 int buf_cnt, int acc, u64 *iova_start)
3003 int status = -ENOMEM;
3004 struct ocrdma_mr *mr;
3005 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3006 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3010 if ((acc & IB_ACCESS_REMOTE_WRITE) && !(acc & IB_ACCESS_LOCAL_WRITE))
3011 return ERR_PTR(-EINVAL);
3013 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3015 return ERR_PTR(status);
3017 num_pbes = count_kernel_pbes(buf_list, buf_cnt, &pbe_size);
3018 if (num_pbes == 0) {
3022 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
3026 mr->hwmr.pbe_size = pbe_size;
3027 mr->hwmr.fbo = *iova_start - (buf_list[0].addr & PAGE_MASK);
3028 mr->hwmr.va = *iova_start;
3029 mr->hwmr.local_rd = 1;
3030 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3031 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3032 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3033 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3034 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
3036 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3039 build_kernel_pbes(buf_list, buf_cnt, pbe_size, mr->hwmr.pbl_table,
3041 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
3045 mr->ibmr.lkey = mr->hwmr.lkey;
3046 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
3047 mr->ibmr.rkey = mr->hwmr.lkey;
3051 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3054 return ERR_PTR(status);