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;
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;
148 int ocrdma_query_port(struct ib_device *ibdev,
149 u8 port, struct ib_port_attr *props)
151 enum ib_port_state port_state;
152 struct ocrdma_dev *dev;
153 struct net_device *netdev;
155 dev = get_ocrdma_dev(ibdev);
157 pr_err("%s(%d) invalid_port=0x%x\n", __func__,
161 netdev = dev->nic_info.netdev;
162 if (netif_running(netdev) && netif_oper_up(netdev)) {
163 port_state = IB_PORT_ACTIVE;
164 props->phys_state = 5;
166 port_state = IB_PORT_DOWN;
167 props->phys_state = 3;
169 props->max_mtu = IB_MTU_4096;
170 props->active_mtu = iboe_get_mtu(netdev->mtu);
175 props->state = port_state;
176 props->port_cap_flags =
179 IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP | IB_PORT_IP_BASED_GIDS;
180 props->gid_tbl_len = OCRDMA_MAX_SGID;
181 props->pkey_tbl_len = 1;
182 props->bad_pkey_cntr = 0;
183 props->qkey_viol_cntr = 0;
184 get_link_speed_and_width(dev, &props->active_speed,
185 &props->active_width);
186 props->max_msg_sz = 0x80000000;
187 props->max_vl_num = 4;
191 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
192 struct ib_port_modify *props)
194 struct ocrdma_dev *dev;
196 dev = get_ocrdma_dev(ibdev);
198 pr_err("%s(%d) invalid_port=0x%x\n", __func__, dev->id, port);
204 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
207 struct ocrdma_mm *mm;
209 mm = kzalloc(sizeof(*mm), GFP_KERNEL);
212 mm->key.phy_addr = phy_addr;
214 INIT_LIST_HEAD(&mm->entry);
216 mutex_lock(&uctx->mm_list_lock);
217 list_add_tail(&mm->entry, &uctx->mm_head);
218 mutex_unlock(&uctx->mm_list_lock);
222 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
225 struct ocrdma_mm *mm, *tmp;
227 mutex_lock(&uctx->mm_list_lock);
228 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
229 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
232 list_del(&mm->entry);
236 mutex_unlock(&uctx->mm_list_lock);
239 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
243 struct ocrdma_mm *mm;
245 mutex_lock(&uctx->mm_list_lock);
246 list_for_each_entry(mm, &uctx->mm_head, entry) {
247 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
253 mutex_unlock(&uctx->mm_list_lock);
257 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
258 struct ocrdma_ucontext *uctx,
259 struct ib_udata *udata)
261 struct ocrdma_pd *pd = NULL;
264 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
266 return ERR_PTR(-ENOMEM);
270 ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
272 pd->dpp_enabled ? OCRDMA_PD_MAX_DPP_ENABLED_QP : 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 ib_umem_chunk *chunk;
730 struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
731 struct ib_umem *umem = mr->umem;
732 int i, shift, pg_cnt, pages, pbe_cnt, 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 list_for_each_entry(chunk, &umem->chunk_list, list) {
743 /* get all the dma regions from the chunk. */
744 for (i = 0; i < chunk->nmap; i++) {
745 pages = sg_dma_len(&chunk->page_list[i]) >> shift;
746 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
747 /* store the page address in pbe */
749 cpu_to_le32(sg_dma_address
750 (&chunk->page_list[i]) +
751 (umem->page_size * pg_cnt));
753 cpu_to_le32(upper_32_bits
755 (&chunk->page_list[i]) +
756 umem->page_size * pg_cnt)));
761 /* if done building pbes, issue the mbx cmd. */
762 if (total_num_pbes == num_pbes)
765 /* if the given pbl is full storing the pbes,
769 (mr->hwmr.pbl_size / sizeof(u64))) {
771 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
779 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
780 u64 usr_addr, int acc, struct ib_udata *udata)
782 int status = -ENOMEM;
783 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
784 struct ocrdma_mr *mr;
785 struct ocrdma_pd *pd;
788 pd = get_ocrdma_pd(ibpd);
790 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
791 return ERR_PTR(-EINVAL);
793 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
795 return ERR_PTR(status);
796 mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
797 if (IS_ERR(mr->umem)) {
801 num_pbes = ib_umem_page_count(mr->umem);
802 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
806 mr->hwmr.pbe_size = mr->umem->page_size;
807 mr->hwmr.fbo = mr->umem->offset;
808 mr->hwmr.va = usr_addr;
810 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
811 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
812 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
813 mr->hwmr.local_rd = 1;
814 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
815 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
818 build_user_pbes(dev, mr, num_pbes);
819 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
822 mr->ibmr.lkey = mr->hwmr.lkey;
823 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
824 mr->ibmr.rkey = mr->hwmr.lkey;
829 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
832 return ERR_PTR(status);
835 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
837 struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
838 struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
841 status = ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
843 if (mr->hwmr.fr_mr == 0)
844 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
846 /* it could be user registered memory. */
848 ib_umem_release(mr->umem);
853 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
854 struct ib_udata *udata,
855 struct ib_ucontext *ib_ctx)
858 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
859 struct ocrdma_create_cq_uresp uresp;
861 memset(&uresp, 0, sizeof(uresp));
862 uresp.cq_id = cq->id;
863 uresp.page_size = PAGE_ALIGN(cq->len);
865 uresp.max_hw_cqe = cq->max_hw_cqe;
866 uresp.page_addr[0] = cq->pa;
867 uresp.db_page_addr = ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
868 uresp.db_page_size = dev->nic_info.db_page_size;
869 uresp.phase_change = cq->phase_change ? 1 : 0;
870 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
872 pr_err("%s(%d) copy error cqid=0x%x.\n",
873 __func__, dev->id, cq->id);
876 status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
879 status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
881 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
889 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev, int entries, int vector,
890 struct ib_ucontext *ib_ctx,
891 struct ib_udata *udata)
893 struct ocrdma_cq *cq;
894 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
895 struct ocrdma_ucontext *uctx = NULL;
898 struct ocrdma_create_cq_ureq ureq;
901 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
902 return ERR_PTR(-EFAULT);
905 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
907 return ERR_PTR(-ENOMEM);
909 spin_lock_init(&cq->cq_lock);
910 spin_lock_init(&cq->comp_handler_lock);
911 INIT_LIST_HEAD(&cq->sq_head);
912 INIT_LIST_HEAD(&cq->rq_head);
913 cq->first_arm = true;
916 uctx = get_ocrdma_ucontext(ib_ctx);
917 pd_id = uctx->cntxt_pd->id;
920 status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
923 return ERR_PTR(status);
926 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
930 cq->phase = OCRDMA_CQE_VALID;
931 dev->cq_tbl[cq->id] = cq;
935 ocrdma_mbx_destroy_cq(dev, cq);
937 return ERR_PTR(status);
940 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
941 struct ib_udata *udata)
944 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
946 if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
954 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
960 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
961 struct ocrdma_cqe *cqe = NULL;
964 cqe_cnt = cq->cqe_cnt;
966 /* Last irq might have scheduled a polling thread
967 * sync-up with it before hard flushing.
969 spin_lock_irqsave(&cq->cq_lock, flags);
971 if (is_cqe_valid(cq, cqe))
976 ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
977 spin_unlock_irqrestore(&cq->cq_lock, flags);
980 int ocrdma_destroy_cq(struct ib_cq *ibcq)
983 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
984 struct ocrdma_eq *eq = NULL;
985 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
989 dev->cq_tbl[cq->id] = NULL;
990 indx = ocrdma_get_eq_table_index(dev, cq->eqn);
994 eq = &dev->eq_tbl[indx];
995 irq = ocrdma_get_irq(dev, eq);
996 synchronize_irq(irq);
999 status = ocrdma_mbx_destroy_cq(dev, cq);
1001 pdid = cq->ucontext->cntxt_pd->id;
1002 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1003 PAGE_ALIGN(cq->len));
1004 ocrdma_del_mmap(cq->ucontext,
1005 ocrdma_get_db_addr(dev, pdid),
1006 dev->nic_info.db_page_size);
1013 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1015 int status = -EINVAL;
1017 if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1018 dev->qp_tbl[qp->id] = qp;
1024 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1026 dev->qp_tbl[qp->id] = NULL;
1029 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1030 struct ib_qp_init_attr *attrs)
1032 if ((attrs->qp_type != IB_QPT_GSI) &&
1033 (attrs->qp_type != IB_QPT_RC) &&
1034 (attrs->qp_type != IB_QPT_UC) &&
1035 (attrs->qp_type != IB_QPT_UD)) {
1036 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1037 __func__, dev->id, attrs->qp_type);
1040 /* Skip the check for QP1 to support CM size of 128 */
1041 if ((attrs->qp_type != IB_QPT_GSI) &&
1042 (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1043 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1044 __func__, dev->id, attrs->cap.max_send_wr);
1045 pr_err("%s(%d) supported send_wr=0x%x\n",
1046 __func__, dev->id, dev->attr.max_wqe);
1049 if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1050 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1051 __func__, dev->id, attrs->cap.max_recv_wr);
1052 pr_err("%s(%d) supported recv_wr=0x%x\n",
1053 __func__, dev->id, dev->attr.max_rqe);
1056 if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1057 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1058 __func__, dev->id, attrs->cap.max_inline_data);
1059 pr_err("%s(%d) supported inline data size=0x%x\n",
1060 __func__, dev->id, dev->attr.max_inline_data);
1063 if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1064 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1065 __func__, dev->id, attrs->cap.max_send_sge);
1066 pr_err("%s(%d) supported send_sge=0x%x\n",
1067 __func__, dev->id, dev->attr.max_send_sge);
1070 if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1071 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1072 __func__, dev->id, attrs->cap.max_recv_sge);
1073 pr_err("%s(%d) supported recv_sge=0x%x\n",
1074 __func__, dev->id, dev->attr.max_recv_sge);
1077 /* unprivileged user space cannot create special QP */
1078 if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1080 ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1081 __func__, dev->id, attrs->qp_type);
1084 /* allow creating only one GSI type of QP */
1085 if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1086 pr_err("%s(%d) GSI special QPs already created.\n",
1090 /* verify consumer QPs are not trying to use GSI QP's CQ */
1091 if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1092 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1093 (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1094 pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1102 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1103 struct ib_udata *udata, int dpp_offset,
1104 int dpp_credit_lmt, int srq)
1108 struct ocrdma_create_qp_uresp uresp;
1109 struct ocrdma_dev *dev = qp->dev;
1110 struct ocrdma_pd *pd = qp->pd;
1112 memset(&uresp, 0, sizeof(uresp));
1113 usr_db = dev->nic_info.unmapped_db +
1114 (pd->id * dev->nic_info.db_page_size);
1115 uresp.qp_id = qp->id;
1116 uresp.sq_dbid = qp->sq.dbid;
1117 uresp.num_sq_pages = 1;
1118 uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1119 uresp.sq_page_addr[0] = qp->sq.pa;
1120 uresp.num_wqe_allocated = qp->sq.max_cnt;
1122 uresp.rq_dbid = qp->rq.dbid;
1123 uresp.num_rq_pages = 1;
1124 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1125 uresp.rq_page_addr[0] = qp->rq.pa;
1126 uresp.num_rqe_allocated = qp->rq.max_cnt;
1128 uresp.db_page_addr = usr_db;
1129 uresp.db_page_size = dev->nic_info.db_page_size;
1130 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1131 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1132 uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1134 if (qp->dpp_enabled) {
1135 uresp.dpp_credit = dpp_credit_lmt;
1136 uresp.dpp_offset = dpp_offset;
1138 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1140 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1143 status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1144 uresp.sq_page_size);
1149 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1150 uresp.rq_page_size);
1156 ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1161 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1162 struct ocrdma_pd *pd)
1164 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1165 qp->sq_db = dev->nic_info.db +
1166 (pd->id * dev->nic_info.db_page_size) +
1167 OCRDMA_DB_GEN2_SQ_OFFSET;
1168 qp->rq_db = dev->nic_info.db +
1169 (pd->id * dev->nic_info.db_page_size) +
1170 OCRDMA_DB_GEN2_RQ_OFFSET;
1172 qp->sq_db = dev->nic_info.db +
1173 (pd->id * dev->nic_info.db_page_size) +
1174 OCRDMA_DB_SQ_OFFSET;
1175 qp->rq_db = dev->nic_info.db +
1176 (pd->id * dev->nic_info.db_page_size) +
1177 OCRDMA_DB_RQ_OFFSET;
1181 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1184 kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1186 if (qp->wqe_wr_id_tbl == NULL)
1189 kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1190 if (qp->rqe_wr_id_tbl == NULL)
1196 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1197 struct ocrdma_pd *pd,
1198 struct ib_qp_init_attr *attrs)
1201 spin_lock_init(&qp->q_lock);
1202 INIT_LIST_HEAD(&qp->sq_entry);
1203 INIT_LIST_HEAD(&qp->rq_entry);
1205 qp->qp_type = attrs->qp_type;
1206 qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1207 qp->max_inline_data = attrs->cap.max_inline_data;
1208 qp->sq.max_sges = attrs->cap.max_send_sge;
1209 qp->rq.max_sges = attrs->cap.max_recv_sge;
1210 qp->state = OCRDMA_QPS_RST;
1211 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1215 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1216 struct ib_qp_init_attr *attrs)
1218 if (attrs->qp_type == IB_QPT_GSI) {
1219 dev->gsi_qp_created = 1;
1220 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1221 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1225 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1226 struct ib_qp_init_attr *attrs,
1227 struct ib_udata *udata)
1230 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1231 struct ocrdma_qp *qp;
1232 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1233 struct ocrdma_create_qp_ureq ureq;
1234 u16 dpp_credit_lmt, dpp_offset;
1236 status = ocrdma_check_qp_params(ibpd, dev, attrs);
1240 memset(&ureq, 0, sizeof(ureq));
1242 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1243 return ERR_PTR(-EFAULT);
1245 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1251 ocrdma_set_qp_init_params(qp, pd, attrs);
1253 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1254 OCRDMA_QP_FAST_REG);
1256 mutex_lock(&dev->dev_lock);
1257 status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1259 &dpp_offset, &dpp_credit_lmt);
1263 /* user space QP's wr_id table are managed in library */
1264 if (udata == NULL) {
1265 status = ocrdma_alloc_wr_id_tbl(qp);
1270 status = ocrdma_add_qpn_map(dev, qp);
1273 ocrdma_set_qp_db(dev, qp, pd);
1275 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1277 (attrs->srq != NULL));
1281 ocrdma_store_gsi_qp_cq(dev, attrs);
1282 qp->ibqp.qp_num = qp->id;
1283 mutex_unlock(&dev->dev_lock);
1287 ocrdma_del_qpn_map(dev, qp);
1289 ocrdma_mbx_destroy_qp(dev, qp);
1291 mutex_unlock(&dev->dev_lock);
1292 kfree(qp->wqe_wr_id_tbl);
1293 kfree(qp->rqe_wr_id_tbl);
1295 pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1297 return ERR_PTR(status);
1301 static void ocrdma_flush_rq_db(struct ocrdma_qp *qp)
1304 u32 val = qp->rq.dbid | (qp->db_cache <<
1305 OCRDMA_DB_RQ_SHIFT);
1306 iowrite32(val, qp->rq_db);
1311 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1315 struct ocrdma_qp *qp;
1316 struct ocrdma_dev *dev;
1317 enum ib_qp_state old_qps;
1319 qp = get_ocrdma_qp(ibqp);
1321 if (attr_mask & IB_QP_STATE)
1322 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1323 /* if new and previous states are same hw doesn't need to
1328 status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1329 if (!status && attr_mask & IB_QP_STATE && attr->qp_state == IB_QPS_RTR)
1330 ocrdma_flush_rq_db(qp);
1335 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1336 int attr_mask, struct ib_udata *udata)
1338 unsigned long flags;
1339 int status = -EINVAL;
1340 struct ocrdma_qp *qp;
1341 struct ocrdma_dev *dev;
1342 enum ib_qp_state old_qps, new_qps;
1344 qp = get_ocrdma_qp(ibqp);
1347 /* syncronize with multiple context trying to change, retrive qps */
1348 mutex_lock(&dev->dev_lock);
1349 /* syncronize with wqe, rqe posting and cqe processing contexts */
1350 spin_lock_irqsave(&qp->q_lock, flags);
1351 old_qps = get_ibqp_state(qp->state);
1352 if (attr_mask & IB_QP_STATE)
1353 new_qps = attr->qp_state;
1356 spin_unlock_irqrestore(&qp->q_lock, flags);
1358 if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1359 IB_LINK_LAYER_ETHERNET)) {
1360 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1361 "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1362 __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1367 status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1371 mutex_unlock(&dev->dev_lock);
1375 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1393 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1395 int ib_qp_acc_flags = 0;
1397 if (qp_cap_flags & OCRDMA_QP_INB_WR)
1398 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1399 if (qp_cap_flags & OCRDMA_QP_INB_RD)
1400 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1401 return ib_qp_acc_flags;
1404 int ocrdma_query_qp(struct ib_qp *ibqp,
1405 struct ib_qp_attr *qp_attr,
1406 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1410 struct ocrdma_qp_params params;
1411 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1412 struct ocrdma_dev *dev = qp->dev;
1414 memset(¶ms, 0, sizeof(params));
1415 mutex_lock(&dev->dev_lock);
1416 status = ocrdma_mbx_query_qp(dev, qp, ¶ms);
1417 mutex_unlock(&dev->dev_lock);
1420 qp_attr->qp_state = get_ibqp_state(IB_QPS_INIT);
1421 qp_attr->cur_qp_state = get_ibqp_state(IB_QPS_INIT);
1423 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1424 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1425 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1426 qp_attr->path_mig_state = IB_MIG_MIGRATED;
1427 qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1428 qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1429 qp_attr->dest_qp_num =
1430 params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1432 qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1433 qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1434 qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1435 qp_attr->cap.max_send_sge = qp->sq.max_sges;
1436 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1437 qp_attr->cap.max_inline_data = qp->max_inline_data;
1438 qp_init_attr->cap = qp_attr->cap;
1439 memcpy(&qp_attr->ah_attr.grh.dgid, ¶ms.dgid[0],
1440 sizeof(params.dgid));
1441 qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1442 OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1443 qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1444 qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1445 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1446 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1447 qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1448 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1449 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1451 qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1452 qp_attr->ah_attr.port_num = 1;
1453 qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1454 OCRDMA_QP_PARAMS_SL_MASK) >>
1455 OCRDMA_QP_PARAMS_SL_SHIFT;
1456 qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1457 OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1458 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1459 qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1460 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1461 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1462 qp_attr->retry_cnt =
1463 (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1464 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1465 qp_attr->min_rnr_timer = 0;
1466 qp_attr->pkey_index = 0;
1467 qp_attr->port_num = 1;
1468 qp_attr->ah_attr.src_path_bits = 0;
1469 qp_attr->ah_attr.static_rate = 0;
1470 qp_attr->alt_pkey_index = 0;
1471 qp_attr->alt_port_num = 0;
1472 qp_attr->alt_timeout = 0;
1473 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1474 qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1475 OCRDMA_QP_PARAMS_STATE_SHIFT;
1476 qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1477 qp_attr->max_dest_rd_atomic =
1478 params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1479 qp_attr->max_rd_atomic =
1480 params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1481 qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1482 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1487 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, int idx)
1490 unsigned int mask = (1 << (idx % 32));
1492 if (srq->idx_bit_fields[i] & mask)
1493 srq->idx_bit_fields[i] &= ~mask;
1495 srq->idx_bit_fields[i] |= mask;
1498 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1500 return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1503 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1505 return (qp->sq.tail == qp->sq.head);
1508 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1510 return (qp->rq.tail == qp->rq.head);
1513 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1515 return q->va + (q->head * q->entry_size);
1518 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1521 return q->va + (idx * q->entry_size);
1524 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1526 q->head = (q->head + 1) & q->max_wqe_idx;
1529 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1531 q->tail = (q->tail + 1) & q->max_wqe_idx;
1534 /* discard the cqe for a given QP */
1535 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1537 unsigned long cq_flags;
1538 unsigned long flags;
1539 int discard_cnt = 0;
1540 u32 cur_getp, stop_getp;
1541 struct ocrdma_cqe *cqe;
1544 spin_lock_irqsave(&cq->cq_lock, cq_flags);
1546 /* traverse through the CQEs in the hw CQ,
1547 * find the matching CQE for a given qp,
1548 * mark the matching one discarded by clearing qpn.
1549 * ring the doorbell in the poll_cq() as
1550 * we don't complete out of order cqe.
1553 cur_getp = cq->getp;
1554 /* find upto when do we reap the cq. */
1555 stop_getp = cur_getp;
1557 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1560 cqe = cq->va + cur_getp;
1561 /* if (a) done reaping whole hw cq, or
1562 * (b) qp_xq becomes empty.
1565 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1566 /* if previously discarded cqe found, skip that too. */
1567 /* check for matching qp */
1568 if (qpn == 0 || qpn != qp->id)
1571 /* mark cqe discarded so that it is not picked up later
1576 if (is_cqe_for_sq(cqe)) {
1577 ocrdma_hwq_inc_tail(&qp->sq);
1580 spin_lock_irqsave(&qp->srq->q_lock, flags);
1581 ocrdma_hwq_inc_tail(&qp->srq->rq);
1582 ocrdma_srq_toggle_bit(qp->srq, cur_getp);
1583 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1586 ocrdma_hwq_inc_tail(&qp->rq);
1590 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1591 } while (cur_getp != stop_getp);
1592 spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1595 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1598 unsigned long flags;
1599 struct ocrdma_dev *dev = qp->dev;
1600 /* sync with any active CQ poll */
1602 spin_lock_irqsave(&dev->flush_q_lock, flags);
1603 found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1605 list_del(&qp->sq_entry);
1607 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1609 list_del(&qp->rq_entry);
1611 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1614 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1617 struct ocrdma_pd *pd;
1618 struct ocrdma_qp *qp;
1619 struct ocrdma_dev *dev;
1620 struct ib_qp_attr attrs;
1621 int attr_mask = IB_QP_STATE;
1622 unsigned long flags;
1624 qp = get_ocrdma_qp(ibqp);
1627 attrs.qp_state = IB_QPS_ERR;
1630 /* change the QP state to ERROR */
1631 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1633 /* ensure that CQEs for newly created QP (whose id may be same with
1634 * one which just getting destroyed are same), dont get
1635 * discarded until the old CQEs are discarded.
1637 mutex_lock(&dev->dev_lock);
1638 status = ocrdma_mbx_destroy_qp(dev, qp);
1641 * acquire CQ lock while destroy is in progress, in order to
1642 * protect against proessing in-flight CQEs for this QP.
1644 spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1645 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1646 spin_lock(&qp->rq_cq->cq_lock);
1648 ocrdma_del_qpn_map(dev, qp);
1650 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1651 spin_unlock(&qp->rq_cq->cq_lock);
1652 spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1655 ocrdma_discard_cqes(qp, qp->sq_cq);
1656 ocrdma_discard_cqes(qp, qp->rq_cq);
1658 mutex_unlock(&dev->dev_lock);
1661 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1662 PAGE_ALIGN(qp->sq.len));
1664 ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1665 PAGE_ALIGN(qp->rq.len));
1668 ocrdma_del_flush_qp(qp);
1670 kfree(qp->wqe_wr_id_tbl);
1671 kfree(qp->rqe_wr_id_tbl);
1676 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1677 struct ib_udata *udata)
1680 struct ocrdma_create_srq_uresp uresp;
1682 memset(&uresp, 0, sizeof(uresp));
1683 uresp.rq_dbid = srq->rq.dbid;
1684 uresp.num_rq_pages = 1;
1685 uresp.rq_page_addr[0] = srq->rq.pa;
1686 uresp.rq_page_size = srq->rq.len;
1687 uresp.db_page_addr = dev->nic_info.unmapped_db +
1688 (srq->pd->id * dev->nic_info.db_page_size);
1689 uresp.db_page_size = dev->nic_info.db_page_size;
1690 uresp.num_rqe_allocated = srq->rq.max_cnt;
1691 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1692 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1693 uresp.db_shift = 24;
1695 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1696 uresp.db_shift = 16;
1699 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1702 status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1703 uresp.rq_page_size);
1709 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1710 struct ib_srq_init_attr *init_attr,
1711 struct ib_udata *udata)
1713 int status = -ENOMEM;
1714 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1715 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1716 struct ocrdma_srq *srq;
1718 if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1719 return ERR_PTR(-EINVAL);
1720 if (init_attr->attr.max_wr > dev->attr.max_rqe)
1721 return ERR_PTR(-EINVAL);
1723 srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1725 return ERR_PTR(status);
1727 spin_lock_init(&srq->q_lock);
1729 srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1730 status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1734 if (udata == NULL) {
1735 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1737 if (srq->rqe_wr_id_tbl == NULL)
1740 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1741 (srq->rq.max_cnt % 32 ? 1 : 0);
1742 srq->idx_bit_fields =
1743 kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1744 if (srq->idx_bit_fields == NULL)
1746 memset(srq->idx_bit_fields, 0xff,
1747 srq->bit_fields_len * sizeof(u32));
1750 if (init_attr->attr.srq_limit) {
1751 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1757 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1765 ocrdma_mbx_destroy_srq(dev, srq);
1767 kfree(srq->rqe_wr_id_tbl);
1768 kfree(srq->idx_bit_fields);
1770 return ERR_PTR(status);
1773 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1774 struct ib_srq_attr *srq_attr,
1775 enum ib_srq_attr_mask srq_attr_mask,
1776 struct ib_udata *udata)
1779 struct ocrdma_srq *srq;
1781 srq = get_ocrdma_srq(ibsrq);
1782 if (srq_attr_mask & IB_SRQ_MAX_WR)
1785 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1789 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1792 struct ocrdma_srq *srq;
1794 srq = get_ocrdma_srq(ibsrq);
1795 status = ocrdma_mbx_query_srq(srq, srq_attr);
1799 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1802 struct ocrdma_srq *srq;
1803 struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1805 srq = get_ocrdma_srq(ibsrq);
1807 status = ocrdma_mbx_destroy_srq(dev, srq);
1810 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1811 PAGE_ALIGN(srq->rq.len));
1813 kfree(srq->idx_bit_fields);
1814 kfree(srq->rqe_wr_id_tbl);
1819 /* unprivileged verbs and their support functions. */
1820 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1821 struct ocrdma_hdr_wqe *hdr,
1822 struct ib_send_wr *wr)
1824 struct ocrdma_ewqe_ud_hdr *ud_hdr =
1825 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1826 struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
1828 ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
1829 if (qp->qp_type == IB_QPT_GSI)
1830 ud_hdr->qkey = qp->qkey;
1832 ud_hdr->qkey = wr->wr.ud.remote_qkey;
1833 ud_hdr->rsvd_ahid = ah->id;
1836 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1837 struct ocrdma_sge *sge, int num_sge,
1838 struct ib_sge *sg_list)
1842 for (i = 0; i < num_sge; i++) {
1843 sge[i].lrkey = sg_list[i].lkey;
1844 sge[i].addr_lo = sg_list[i].addr;
1845 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1846 sge[i].len = sg_list[i].length;
1847 hdr->total_len += sg_list[i].length;
1850 memset(sge, 0, sizeof(*sge));
1853 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1855 uint32_t total_len = 0, i;
1857 for (i = 0; i < num_sge; i++)
1858 total_len += sg_list[i].length;
1863 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1864 struct ocrdma_hdr_wqe *hdr,
1865 struct ocrdma_sge *sge,
1866 struct ib_send_wr *wr, u32 wqe_size)
1871 if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1872 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1873 if (unlikely(hdr->total_len > qp->max_inline_data)) {
1874 pr_err("%s() supported_len=0x%x,\n"
1875 " unspported len req=0x%x\n", __func__,
1876 qp->max_inline_data, hdr->total_len);
1879 dpp_addr = (char *)sge;
1880 for (i = 0; i < wr->num_sge; i++) {
1882 (void *)(unsigned long)wr->sg_list[i].addr,
1883 wr->sg_list[i].length);
1884 dpp_addr += wr->sg_list[i].length;
1887 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1888 if (0 == hdr->total_len)
1889 wqe_size += sizeof(struct ocrdma_sge);
1890 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1892 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1894 wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1896 wqe_size += sizeof(struct ocrdma_sge);
1897 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1899 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1903 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1904 struct ib_send_wr *wr)
1907 struct ocrdma_sge *sge;
1908 u32 wqe_size = sizeof(*hdr);
1910 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1911 ocrdma_build_ud_hdr(qp, hdr, wr);
1912 sge = (struct ocrdma_sge *)(hdr + 2);
1913 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1915 sge = (struct ocrdma_sge *)(hdr + 1);
1918 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1922 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1923 struct ib_send_wr *wr)
1926 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1927 struct ocrdma_sge *sge = ext_rw + 1;
1928 u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1930 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1933 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1934 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1935 ext_rw->lrkey = wr->wr.rdma.rkey;
1936 ext_rw->len = hdr->total_len;
1940 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1941 struct ib_send_wr *wr)
1943 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1944 struct ocrdma_sge *sge = ext_rw + 1;
1945 u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
1946 sizeof(struct ocrdma_hdr_wqe);
1948 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1949 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1950 hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
1951 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1953 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1954 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1955 ext_rw->lrkey = wr->wr.rdma.rkey;
1956 ext_rw->len = hdr->total_len;
1959 static void build_frmr_pbes(struct ib_send_wr *wr, struct ocrdma_pbl *pbl_tbl,
1960 struct ocrdma_hw_mr *hwmr)
1965 struct ocrdma_pbe *pbe;
1967 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1970 /* go through the OS phy regions & fill hw pbe entries into pbls. */
1971 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
1972 /* number of pbes can be more for one OS buf, when
1973 * buffers are of different sizes.
1974 * split the ib_buf to one or more pbes.
1976 buf_addr = wr->wr.fast_reg.page_list->page_list[i];
1977 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
1978 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
1982 /* if the pbl is full storing the pbes,
1985 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
1987 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1993 static int get_encoded_page_size(int pg_sz)
1995 /* Max size is 256M 4096 << 16 */
1998 if (pg_sz == (4096 << i))
2004 static int ocrdma_build_fr(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2005 struct ib_send_wr *wr)
2008 struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2009 struct ocrdma_mr *mr;
2010 u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2012 wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2014 if (wr->wr.fast_reg.page_list_len > qp->dev->attr.max_pages_per_frmr)
2017 hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2018 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2020 if (wr->wr.fast_reg.page_list_len == 0)
2022 if (wr->wr.fast_reg.access_flags & IB_ACCESS_LOCAL_WRITE)
2023 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2024 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_WRITE)
2025 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2026 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_READ)
2027 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2028 hdr->lkey = wr->wr.fast_reg.rkey;
2029 hdr->total_len = wr->wr.fast_reg.length;
2031 fbo = wr->wr.fast_reg.iova_start -
2032 (wr->wr.fast_reg.page_list->page_list[0] & PAGE_MASK);
2034 fast_reg->va_hi = upper_32_bits(wr->wr.fast_reg.iova_start);
2035 fast_reg->va_lo = (u32) (wr->wr.fast_reg.iova_start & 0xffffffff);
2036 fast_reg->fbo_hi = upper_32_bits(fbo);
2037 fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2038 fast_reg->num_sges = wr->wr.fast_reg.page_list_len;
2039 fast_reg->size_sge =
2040 get_encoded_page_size(1 << wr->wr.fast_reg.page_shift);
2041 mr = (struct ocrdma_mr *) (unsigned long) qp->dev->stag_arr[(hdr->lkey >> 8) &
2042 (OCRDMA_MAX_STAG - 1)];
2043 build_frmr_pbes(wr, mr->hwmr.pbl_table, &mr->hwmr);
2047 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2049 u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2051 iowrite32(val, qp->sq_db);
2054 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2055 struct ib_send_wr **bad_wr)
2058 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2059 struct ocrdma_hdr_wqe *hdr;
2060 unsigned long flags;
2062 spin_lock_irqsave(&qp->q_lock, flags);
2063 if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2064 spin_unlock_irqrestore(&qp->q_lock, flags);
2070 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2071 wr->num_sge > qp->sq.max_sges) {
2076 hdr = ocrdma_hwq_head(&qp->sq);
2078 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2079 hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2080 if (wr->send_flags & IB_SEND_FENCE)
2082 (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2083 if (wr->send_flags & IB_SEND_SOLICITED)
2085 (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2087 switch (wr->opcode) {
2088 case IB_WR_SEND_WITH_IMM:
2089 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2090 hdr->immdt = ntohl(wr->ex.imm_data);
2092 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2093 ocrdma_build_send(qp, hdr, wr);
2095 case IB_WR_SEND_WITH_INV:
2096 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2097 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2098 hdr->lkey = wr->ex.invalidate_rkey;
2099 status = ocrdma_build_send(qp, hdr, wr);
2101 case IB_WR_RDMA_WRITE_WITH_IMM:
2102 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2103 hdr->immdt = ntohl(wr->ex.imm_data);
2104 case IB_WR_RDMA_WRITE:
2105 hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2106 status = ocrdma_build_write(qp, hdr, wr);
2108 case IB_WR_RDMA_READ_WITH_INV:
2109 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2110 case IB_WR_RDMA_READ:
2111 ocrdma_build_read(qp, hdr, wr);
2113 case IB_WR_LOCAL_INV:
2115 (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2116 hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2117 sizeof(struct ocrdma_sge)) /
2118 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2119 hdr->lkey = wr->ex.invalidate_rkey;
2121 case IB_WR_FAST_REG_MR:
2122 status = ocrdma_build_fr(qp, hdr, wr);
2132 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2133 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2135 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2136 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2137 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2138 OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2139 /* make sure wqe is written before adapter can access it */
2141 /* inform hw to start processing it */
2142 ocrdma_ring_sq_db(qp);
2144 /* update pointer, counter for next wr */
2145 ocrdma_hwq_inc_head(&qp->sq);
2148 spin_unlock_irqrestore(&qp->q_lock, flags);
2152 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2154 u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2156 iowrite32(val, qp->rq_db);
2159 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2163 struct ocrdma_sge *sge;
2165 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2167 wqe_size = sizeof(*sge) + sizeof(*rqe);
2169 rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2170 OCRDMA_WQE_SIZE_SHIFT);
2171 rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2172 rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2174 rqe->rsvd_tag = tag;
2175 sge = (struct ocrdma_sge *)(rqe + 1);
2176 ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2177 ocrdma_cpu_to_le32(rqe, wqe_size);
2180 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2181 struct ib_recv_wr **bad_wr)
2184 unsigned long flags;
2185 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2186 struct ocrdma_hdr_wqe *rqe;
2188 spin_lock_irqsave(&qp->q_lock, flags);
2189 if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2190 spin_unlock_irqrestore(&qp->q_lock, flags);
2195 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2196 wr->num_sge > qp->rq.max_sges) {
2201 rqe = ocrdma_hwq_head(&qp->rq);
2202 ocrdma_build_rqe(rqe, wr, 0);
2204 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2205 /* make sure rqe is written before adapter can access it */
2208 /* inform hw to start processing it */
2209 ocrdma_ring_rq_db(qp);
2211 /* update pointer, counter for next wr */
2212 ocrdma_hwq_inc_head(&qp->rq);
2215 spin_unlock_irqrestore(&qp->q_lock, flags);
2219 /* cqe for srq's rqe can potentially arrive out of order.
2220 * index gives the entry in the shadow table where to store
2221 * the wr_id. tag/index is returned in cqe to reference back
2224 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2229 for (row = 0; row < srq->bit_fields_len; row++) {
2230 if (srq->idx_bit_fields[row]) {
2231 indx = ffs(srq->idx_bit_fields[row]);
2232 indx = (row * 32) + (indx - 1);
2233 if (indx >= srq->rq.max_cnt)
2235 ocrdma_srq_toggle_bit(srq, indx);
2240 if (row == srq->bit_fields_len)
2245 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2247 u32 val = srq->rq.dbid | (1 << 16);
2249 iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2252 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2253 struct ib_recv_wr **bad_wr)
2256 unsigned long flags;
2257 struct ocrdma_srq *srq;
2258 struct ocrdma_hdr_wqe *rqe;
2261 srq = get_ocrdma_srq(ibsrq);
2263 spin_lock_irqsave(&srq->q_lock, flags);
2265 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2266 wr->num_sge > srq->rq.max_sges) {
2271 tag = ocrdma_srq_get_idx(srq);
2272 rqe = ocrdma_hwq_head(&srq->rq);
2273 ocrdma_build_rqe(rqe, wr, tag);
2275 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2276 /* make sure rqe is written before adapter can perform DMA */
2278 /* inform hw to start processing it */
2279 ocrdma_ring_srq_db(srq);
2280 /* update pointer, counter for next wr */
2281 ocrdma_hwq_inc_head(&srq->rq);
2284 spin_unlock_irqrestore(&srq->q_lock, flags);
2288 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2290 enum ib_wc_status ibwc_status;
2293 case OCRDMA_CQE_GENERAL_ERR:
2294 ibwc_status = IB_WC_GENERAL_ERR;
2296 case OCRDMA_CQE_LOC_LEN_ERR:
2297 ibwc_status = IB_WC_LOC_LEN_ERR;
2299 case OCRDMA_CQE_LOC_QP_OP_ERR:
2300 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2302 case OCRDMA_CQE_LOC_EEC_OP_ERR:
2303 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2305 case OCRDMA_CQE_LOC_PROT_ERR:
2306 ibwc_status = IB_WC_LOC_PROT_ERR;
2308 case OCRDMA_CQE_WR_FLUSH_ERR:
2309 ibwc_status = IB_WC_WR_FLUSH_ERR;
2311 case OCRDMA_CQE_MW_BIND_ERR:
2312 ibwc_status = IB_WC_MW_BIND_ERR;
2314 case OCRDMA_CQE_BAD_RESP_ERR:
2315 ibwc_status = IB_WC_BAD_RESP_ERR;
2317 case OCRDMA_CQE_LOC_ACCESS_ERR:
2318 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2320 case OCRDMA_CQE_REM_INV_REQ_ERR:
2321 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2323 case OCRDMA_CQE_REM_ACCESS_ERR:
2324 ibwc_status = IB_WC_REM_ACCESS_ERR;
2326 case OCRDMA_CQE_REM_OP_ERR:
2327 ibwc_status = IB_WC_REM_OP_ERR;
2329 case OCRDMA_CQE_RETRY_EXC_ERR:
2330 ibwc_status = IB_WC_RETRY_EXC_ERR;
2332 case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2333 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2335 case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2336 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2338 case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2339 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2341 case OCRDMA_CQE_REM_ABORT_ERR:
2342 ibwc_status = IB_WC_REM_ABORT_ERR;
2344 case OCRDMA_CQE_INV_EECN_ERR:
2345 ibwc_status = IB_WC_INV_EECN_ERR;
2347 case OCRDMA_CQE_INV_EEC_STATE_ERR:
2348 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2350 case OCRDMA_CQE_FATAL_ERR:
2351 ibwc_status = IB_WC_FATAL_ERR;
2353 case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2354 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2357 ibwc_status = IB_WC_GENERAL_ERR;
2363 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2366 struct ocrdma_hdr_wqe *hdr;
2367 struct ocrdma_sge *rw;
2370 hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2372 ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2373 /* Undo the hdr->cw swap */
2374 opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2377 ibwc->opcode = IB_WC_RDMA_WRITE;
2380 rw = (struct ocrdma_sge *)(hdr + 1);
2381 ibwc->opcode = IB_WC_RDMA_READ;
2382 ibwc->byte_len = rw->len;
2385 ibwc->opcode = IB_WC_SEND;
2388 ibwc->opcode = IB_WC_FAST_REG_MR;
2390 case OCRDMA_LKEY_INV:
2391 ibwc->opcode = IB_WC_LOCAL_INV;
2394 ibwc->status = IB_WC_GENERAL_ERR;
2395 pr_err("%s() invalid opcode received = 0x%x\n",
2396 __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2401 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2402 struct ocrdma_cqe *cqe)
2404 if (is_cqe_for_sq(cqe)) {
2405 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2406 cqe->flags_status_srcqpn) &
2407 ~OCRDMA_CQE_STATUS_MASK);
2408 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2409 cqe->flags_status_srcqpn) |
2410 (OCRDMA_CQE_WR_FLUSH_ERR <<
2411 OCRDMA_CQE_STATUS_SHIFT));
2413 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2414 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2415 cqe->flags_status_srcqpn) &
2416 ~OCRDMA_CQE_UD_STATUS_MASK);
2417 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2418 cqe->flags_status_srcqpn) |
2419 (OCRDMA_CQE_WR_FLUSH_ERR <<
2420 OCRDMA_CQE_UD_STATUS_SHIFT));
2422 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2423 cqe->flags_status_srcqpn) &
2424 ~OCRDMA_CQE_STATUS_MASK);
2425 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2426 cqe->flags_status_srcqpn) |
2427 (OCRDMA_CQE_WR_FLUSH_ERR <<
2428 OCRDMA_CQE_STATUS_SHIFT));
2433 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2434 struct ocrdma_qp *qp, int status)
2436 bool expand = false;
2439 ibwc->qp = &qp->ibqp;
2440 ibwc->status = ocrdma_to_ibwc_err(status);
2442 ocrdma_flush_qp(qp);
2443 ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2445 /* if wqe/rqe pending for which cqe needs to be returned,
2446 * trigger inflating it.
2448 if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2450 ocrdma_set_cqe_status_flushed(qp, cqe);
2455 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2456 struct ocrdma_qp *qp, int status)
2458 ibwc->opcode = IB_WC_RECV;
2459 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2460 ocrdma_hwq_inc_tail(&qp->rq);
2462 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2465 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2466 struct ocrdma_qp *qp, int status)
2468 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2469 ocrdma_hwq_inc_tail(&qp->sq);
2471 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2475 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2476 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2477 bool *polled, bool *stop)
2480 int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2481 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2483 /* when hw sq is empty, but rq is not empty, so we continue
2484 * to keep the cqe in order to get the cq event again.
2486 if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2487 /* when cq for rq and sq is same, it is safe to return
2488 * flush cqe for RQEs.
2490 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2492 status = OCRDMA_CQE_WR_FLUSH_ERR;
2493 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2495 /* stop processing further cqe as this cqe is used for
2496 * triggering cq event on buddy cq of RQ.
2497 * When QP is destroyed, this cqe will be removed
2498 * from the cq's hardware q.
2506 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2511 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2512 struct ocrdma_cqe *cqe,
2513 struct ib_wc *ibwc, bool *polled)
2515 bool expand = false;
2516 int tail = qp->sq.tail;
2519 if (!qp->wqe_wr_id_tbl[tail].signaled) {
2520 *polled = false; /* WC cannot be consumed yet */
2522 ibwc->status = IB_WC_SUCCESS;
2524 ibwc->qp = &qp->ibqp;
2525 ocrdma_update_wc(qp, ibwc, tail);
2528 wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2529 OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2530 if (tail != wqe_idx)
2531 expand = true; /* Coalesced CQE can't be consumed yet */
2533 ocrdma_hwq_inc_tail(&qp->sq);
2537 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2538 struct ib_wc *ibwc, bool *polled, bool *stop)
2543 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2544 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2546 if (status == OCRDMA_CQE_SUCCESS)
2547 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2549 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2553 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2557 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2558 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2559 ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2560 OCRDMA_CQE_SRCQP_MASK;
2561 ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2562 OCRDMA_CQE_PKEY_MASK;
2563 ibwc->wc_flags = IB_WC_GRH;
2564 ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2565 OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2569 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2570 struct ocrdma_cqe *cqe,
2571 struct ocrdma_qp *qp)
2573 unsigned long flags;
2574 struct ocrdma_srq *srq;
2577 srq = get_ocrdma_srq(qp->ibqp.srq);
2578 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2579 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2580 ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2581 spin_lock_irqsave(&srq->q_lock, flags);
2582 ocrdma_srq_toggle_bit(srq, wqe_idx);
2583 spin_unlock_irqrestore(&srq->q_lock, flags);
2584 ocrdma_hwq_inc_tail(&srq->rq);
2587 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2588 struct ib_wc *ibwc, bool *polled, bool *stop,
2593 /* when hw_rq is empty, but wq is not empty, so continue
2594 * to keep the cqe to get the cq event again.
2596 if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2597 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2599 status = OCRDMA_CQE_WR_FLUSH_ERR;
2600 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2608 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2613 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2614 struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2616 ibwc->opcode = IB_WC_RECV;
2617 ibwc->qp = &qp->ibqp;
2618 ibwc->status = IB_WC_SUCCESS;
2620 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2621 ocrdma_update_ud_rcqe(ibwc, cqe);
2623 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2625 if (is_cqe_imm(cqe)) {
2626 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2627 ibwc->wc_flags |= IB_WC_WITH_IMM;
2628 } else if (is_cqe_wr_imm(cqe)) {
2629 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2630 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2631 ibwc->wc_flags |= IB_WC_WITH_IMM;
2632 } else if (is_cqe_invalidated(cqe)) {
2633 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2634 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2637 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2639 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2640 ocrdma_hwq_inc_tail(&qp->rq);
2644 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2645 struct ib_wc *ibwc, bool *polled, bool *stop)
2648 bool expand = false;
2651 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2652 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2653 OCRDMA_CQE_UD_STATUS_MASK) >>
2654 OCRDMA_CQE_UD_STATUS_SHIFT;
2656 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2657 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2660 if (status == OCRDMA_CQE_SUCCESS) {
2662 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2664 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2670 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2673 if (cq->phase_change) {
2675 cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2677 /* clear valid bit */
2678 cqe->flags_status_srcqpn = 0;
2682 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2687 bool expand = false;
2688 int polled_hw_cqes = 0;
2689 struct ocrdma_qp *qp = NULL;
2690 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2691 struct ocrdma_cqe *cqe;
2692 u16 cur_getp; bool polled = false; bool stop = false;
2694 cur_getp = cq->getp;
2695 while (num_entries) {
2696 cqe = cq->va + cur_getp;
2697 /* check whether valid cqe or not */
2698 if (!is_cqe_valid(cq, cqe))
2700 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2701 /* ignore discarded cqe */
2704 qp = dev->qp_tbl[qpn];
2707 if (is_cqe_for_sq(cqe)) {
2708 expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2711 expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2718 /* clear qpn to avoid duplicate processing by discard_cqe() */
2721 polled_hw_cqes += 1;
2722 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2723 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2733 cq->getp = cur_getp;
2734 if (cq->deferred_arm) {
2735 ocrdma_ring_cq_db(dev, cq->id, true, cq->deferred_sol,
2737 cq->deferred_arm = false;
2738 cq->deferred_sol = false;
2740 /* We need to pop the CQE. No need to arm */
2741 ocrdma_ring_cq_db(dev, cq->id, false, cq->deferred_sol,
2743 cq->deferred_sol = false;
2749 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2750 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2751 struct ocrdma_qp *qp, struct ib_wc *ibwc)
2755 while (num_entries) {
2756 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2758 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2759 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2760 ocrdma_hwq_inc_tail(&qp->sq);
2761 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2762 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2763 ocrdma_hwq_inc_tail(&qp->rq);
2768 ibwc->status = IB_WC_WR_FLUSH_ERR;
2776 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2778 int cqes_to_poll = num_entries;
2779 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2780 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2781 int num_os_cqe = 0, err_cqes = 0;
2782 struct ocrdma_qp *qp;
2783 unsigned long flags;
2785 /* poll cqes from adapter CQ */
2786 spin_lock_irqsave(&cq->cq_lock, flags);
2787 num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2788 spin_unlock_irqrestore(&cq->cq_lock, flags);
2789 cqes_to_poll -= num_os_cqe;
2792 wc = wc + num_os_cqe;
2793 /* adapter returns single error cqe when qp moves to
2794 * error state. So insert error cqes with wc_status as
2795 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2796 * respectively which uses this CQ.
2798 spin_lock_irqsave(&dev->flush_q_lock, flags);
2799 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2800 if (cqes_to_poll == 0)
2802 err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2803 cqes_to_poll -= err_cqes;
2804 num_os_cqe += err_cqes;
2807 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2812 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2814 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2815 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2817 unsigned long flags;
2818 bool arm_needed = false, sol_needed = false;
2822 spin_lock_irqsave(&cq->cq_lock, flags);
2823 if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2825 if (cq_flags & IB_CQ_SOLICITED)
2828 if (cq->first_arm) {
2829 ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2830 cq->first_arm = false;
2833 cq->deferred_arm = true;
2836 cq->deferred_sol = sol_needed;
2837 spin_unlock_irqrestore(&cq->cq_lock, flags);
2842 struct ib_mr *ocrdma_alloc_frmr(struct ib_pd *ibpd, int max_page_list_len)
2845 struct ocrdma_mr *mr;
2846 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2847 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2849 if (max_page_list_len > dev->attr.max_pages_per_frmr)
2850 return ERR_PTR(-EINVAL);
2852 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2854 return ERR_PTR(-ENOMEM);
2856 status = ocrdma_get_pbl_info(dev, mr, max_page_list_len);
2860 mr->hwmr.remote_rd = 0;
2861 mr->hwmr.remote_wr = 0;
2862 mr->hwmr.local_rd = 0;
2863 mr->hwmr.local_wr = 0;
2864 mr->hwmr.mw_bind = 0;
2865 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2868 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2871 mr->ibmr.rkey = mr->hwmr.lkey;
2872 mr->ibmr.lkey = mr->hwmr.lkey;
2873 dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] = mr;
2876 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2879 return ERR_PTR(-ENOMEM);
2882 struct ib_fast_reg_page_list *ocrdma_alloc_frmr_page_list(struct ib_device
2886 struct ib_fast_reg_page_list *frmr_list;
2889 size = sizeof(*frmr_list) + (page_list_len * sizeof(u64));
2890 frmr_list = kzalloc(size, GFP_KERNEL);
2892 return ERR_PTR(-ENOMEM);
2893 frmr_list->page_list = (u64 *)(frmr_list + 1);
2897 void ocrdma_free_frmr_page_list(struct ib_fast_reg_page_list *page_list)
2902 #define MAX_KERNEL_PBE_SIZE 65536
2903 static inline int count_kernel_pbes(struct ib_phys_buf *buf_list,
2904 int buf_cnt, u32 *pbe_size)
2909 *pbe_size = roundup(buf_list[0].size, PAGE_SIZE);
2910 *pbe_size = roundup_pow_of_two(*pbe_size);
2912 /* find the smallest PBE size that we can have */
2913 for (i = 0; i < buf_cnt; i++) {
2914 /* first addr may not be page aligned, so ignore checking */
2915 if ((i != 0) && ((buf_list[i].addr & ~PAGE_MASK) ||
2916 (buf_list[i].size & ~PAGE_MASK))) {
2920 /* if configured PBE size is greater then the chosen one,
2921 * reduce the PBE size.
2923 buf_size = roundup(buf_list[i].size, PAGE_SIZE);
2924 /* pbe_size has to be even multiple of 4K 1,2,4,8...*/
2925 buf_size = roundup_pow_of_two(buf_size);
2926 if (*pbe_size > buf_size)
2927 *pbe_size = buf_size;
2929 total_size += buf_size;
2931 *pbe_size = *pbe_size > MAX_KERNEL_PBE_SIZE ?
2932 (MAX_KERNEL_PBE_SIZE) : (*pbe_size);
2934 /* num_pbes = total_size / (*pbe_size); this is implemented below. */
2936 return total_size >> ilog2(*pbe_size);
2939 static void build_kernel_pbes(struct ib_phys_buf *buf_list, int ib_buf_cnt,
2940 u32 pbe_size, struct ocrdma_pbl *pbl_tbl,
2941 struct ocrdma_hw_mr *hwmr)
2945 int pbes_per_buf = 0;
2948 struct ocrdma_pbe *pbe;
2949 int total_num_pbes = 0;
2951 if (!hwmr->num_pbes)
2954 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2957 /* go through the OS phy regions & fill hw pbe entries into pbls. */
2958 for (i = 0; i < ib_buf_cnt; i++) {
2959 buf_addr = buf_list[i].addr;
2961 roundup_pow_of_two(roundup(buf_list[i].size, PAGE_SIZE)) /
2963 hwmr->len += buf_list[i].size;
2964 /* number of pbes can be more for one OS buf, when
2965 * buffers are of different sizes.
2966 * split the ib_buf to one or more pbes.
2968 for (idx = 0; idx < pbes_per_buf; idx++) {
2969 /* we program always page aligned addresses,
2970 * first unaligned address is taken care by fbo.
2973 /* for non zero fbo, assign the
2974 * start of the page.
2977 cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2979 cpu_to_le32((u32) upper_32_bits(buf_addr));
2982 cpu_to_le32((u32) (buf_addr & 0xffffffff));
2984 cpu_to_le32((u32) upper_32_bits(buf_addr));
2986 buf_addr += pbe_size;
2988 total_num_pbes += 1;
2991 if (total_num_pbes == hwmr->num_pbes)
2993 /* if the pbl is full storing the pbes,
2996 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
2998 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
3007 struct ib_mr *ocrdma_reg_kernel_mr(struct ib_pd *ibpd,
3008 struct ib_phys_buf *buf_list,
3009 int buf_cnt, int acc, u64 *iova_start)
3011 int status = -ENOMEM;
3012 struct ocrdma_mr *mr;
3013 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3014 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3018 if ((acc & IB_ACCESS_REMOTE_WRITE) && !(acc & IB_ACCESS_LOCAL_WRITE))
3019 return ERR_PTR(-EINVAL);
3021 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3023 return ERR_PTR(status);
3025 num_pbes = count_kernel_pbes(buf_list, buf_cnt, &pbe_size);
3026 if (num_pbes == 0) {
3030 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
3034 mr->hwmr.pbe_size = pbe_size;
3035 mr->hwmr.fbo = *iova_start - (buf_list[0].addr & PAGE_MASK);
3036 mr->hwmr.va = *iova_start;
3037 mr->hwmr.local_rd = 1;
3038 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3039 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3040 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3041 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3042 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
3044 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3047 build_kernel_pbes(buf_list, buf_cnt, pbe_size, mr->hwmr.pbl_table,
3049 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
3053 mr->ibmr.lkey = mr->hwmr.lkey;
3054 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
3055 mr->ibmr.rkey = mr->hwmr.lkey;
3059 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3062 return ERR_PTR(status);