1 /* This file is part of the Emulex RoCE Device Driver for
2 * RoCE (RDMA over Converged Ethernet) adapters.
3 * Copyright (C) 2012-2015 Emulex. All rights reserved.
4 * EMULEX and SLI are trademarks of Emulex.
7 * This software is available to you under a choice of one of two licenses.
8 * You may choose to be licensed under the terms of the GNU General Public
9 * License (GPL) Version 2, available from the file COPYING in the main
10 * directory of this source tree, or the BSD license below:
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * - Redistributions of source code must retain the above copyright notice,
17 * this list of conditions and the following disclaimer.
19 * - Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in
21 * the documentation and/or other materials provided with the distribution.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
24 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
27 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
28 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
29 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
30 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
31 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
32 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
33 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 * Contact Information:
36 * linux-drivers@emulex.com
40 * Costa Mesa, CA 92626
43 #include <linux/dma-mapping.h>
44 #include <rdma/ib_verbs.h>
45 #include <rdma/ib_user_verbs.h>
46 #include <rdma/iw_cm.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_cache.h>
52 #include "ocrdma_hw.h"
53 #include "ocrdma_verbs.h"
54 #include <rdma/ocrdma-abi.h>
56 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
65 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
68 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
70 if (uhw->inlen || uhw->outlen)
73 memset(attr, 0, sizeof *attr);
74 memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
75 min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
76 ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
77 attr->max_mr_size = dev->attr.max_mr_size;
78 attr->page_size_cap = 0xffff000;
79 attr->vendor_id = dev->nic_info.pdev->vendor;
80 attr->vendor_part_id = dev->nic_info.pdev->device;
81 attr->hw_ver = dev->asic_id;
82 attr->max_qp = dev->attr.max_qp;
83 attr->max_ah = OCRDMA_MAX_AH;
84 attr->max_qp_wr = dev->attr.max_wqe;
86 attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
87 IB_DEVICE_RC_RNR_NAK_GEN |
88 IB_DEVICE_SHUTDOWN_PORT |
89 IB_DEVICE_SYS_IMAGE_GUID |
90 IB_DEVICE_LOCAL_DMA_LKEY |
91 IB_DEVICE_MEM_MGT_EXTENSIONS;
92 attr->max_send_sge = dev->attr.max_send_sge;
93 attr->max_recv_sge = dev->attr.max_recv_sge;
94 attr->max_sge_rd = dev->attr.max_rdma_sge;
95 attr->max_cq = dev->attr.max_cq;
96 attr->max_cqe = dev->attr.max_cqe;
97 attr->max_mr = dev->attr.max_mr;
98 attr->max_mw = dev->attr.max_mw;
99 attr->max_pd = dev->attr.max_pd;
100 attr->atomic_cap = 0;
102 attr->max_map_per_fmr = 0;
103 attr->max_qp_rd_atom =
104 min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
105 attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
106 attr->max_srq = dev->attr.max_srq;
107 attr->max_srq_sge = dev->attr.max_srq_sge;
108 attr->max_srq_wr = dev->attr.max_rqe;
109 attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
110 attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
115 struct net_device *ocrdma_get_netdev(struct ib_device *ibdev, u8 port_num)
117 struct ocrdma_dev *dev;
118 struct net_device *ndev = NULL;
122 dev = get_ocrdma_dev(ibdev);
124 ndev = dev->nic_info.netdev;
133 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
134 u8 *ib_speed, u8 *ib_width)
139 status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
141 speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
144 case OCRDMA_PHYS_LINK_SPEED_1GBPS:
145 *ib_speed = IB_SPEED_SDR;
146 *ib_width = IB_WIDTH_1X;
149 case OCRDMA_PHYS_LINK_SPEED_10GBPS:
150 *ib_speed = IB_SPEED_QDR;
151 *ib_width = IB_WIDTH_1X;
154 case OCRDMA_PHYS_LINK_SPEED_20GBPS:
155 *ib_speed = IB_SPEED_DDR;
156 *ib_width = IB_WIDTH_4X;
159 case OCRDMA_PHYS_LINK_SPEED_40GBPS:
160 *ib_speed = IB_SPEED_QDR;
161 *ib_width = IB_WIDTH_4X;
166 *ib_speed = IB_SPEED_SDR;
167 *ib_width = IB_WIDTH_1X;
171 int ocrdma_query_port(struct ib_device *ibdev,
172 u8 port, struct ib_port_attr *props)
174 enum ib_port_state port_state;
175 struct ocrdma_dev *dev;
176 struct net_device *netdev;
178 /* props being zeroed by the caller, avoid zeroing it here */
179 dev = get_ocrdma_dev(ibdev);
180 netdev = dev->nic_info.netdev;
181 if (netif_running(netdev) && netif_oper_up(netdev)) {
182 port_state = IB_PORT_ACTIVE;
183 props->phys_state = 5;
185 port_state = IB_PORT_DOWN;
186 props->phys_state = 3;
188 props->max_mtu = IB_MTU_4096;
189 props->active_mtu = iboe_get_mtu(netdev->mtu);
194 props->state = port_state;
195 props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
196 IB_PORT_DEVICE_MGMT_SUP |
197 IB_PORT_VENDOR_CLASS_SUP;
198 props->ip_gids = true;
199 props->gid_tbl_len = OCRDMA_MAX_SGID;
200 props->pkey_tbl_len = 1;
201 props->bad_pkey_cntr = 0;
202 props->qkey_viol_cntr = 0;
203 get_link_speed_and_width(dev, &props->active_speed,
204 &props->active_width);
205 props->max_msg_sz = 0x80000000;
206 props->max_vl_num = 4;
210 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
211 struct ib_port_modify *props)
216 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
219 struct ocrdma_mm *mm;
221 mm = kzalloc(sizeof(*mm), GFP_KERNEL);
224 mm->key.phy_addr = phy_addr;
226 INIT_LIST_HEAD(&mm->entry);
228 mutex_lock(&uctx->mm_list_lock);
229 list_add_tail(&mm->entry, &uctx->mm_head);
230 mutex_unlock(&uctx->mm_list_lock);
234 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
237 struct ocrdma_mm *mm, *tmp;
239 mutex_lock(&uctx->mm_list_lock);
240 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
241 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
244 list_del(&mm->entry);
248 mutex_unlock(&uctx->mm_list_lock);
251 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
255 struct ocrdma_mm *mm;
257 mutex_lock(&uctx->mm_list_lock);
258 list_for_each_entry(mm, &uctx->mm_head, entry) {
259 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
265 mutex_unlock(&uctx->mm_list_lock);
270 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
272 u16 pd_bitmap_idx = 0;
273 const unsigned long *pd_bitmap;
276 pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
277 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
278 dev->pd_mgr->max_dpp_pd);
279 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
280 dev->pd_mgr->pd_dpp_count++;
281 if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
282 dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
284 pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
285 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
286 dev->pd_mgr->max_normal_pd);
287 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
288 dev->pd_mgr->pd_norm_count++;
289 if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
290 dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
292 return pd_bitmap_idx;
295 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
301 pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
302 dev->pd_mgr->pd_norm_count;
307 pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
308 if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
311 __clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
312 dev->pd_mgr->pd_dpp_count--;
315 pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
316 if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
319 __clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
320 dev->pd_mgr->pd_norm_count--;
327 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
332 mutex_lock(&dev->dev_lock);
333 status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
334 mutex_unlock(&dev->dev_lock);
338 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
343 mutex_lock(&dev->dev_lock);
344 if (pd->dpp_enabled) {
345 /* try allocating DPP PD, if not available then normal PD */
346 if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
347 pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
348 pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
349 pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
350 } else if (dev->pd_mgr->pd_norm_count <
351 dev->pd_mgr->max_normal_pd) {
352 pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
353 pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
354 pd->dpp_enabled = false;
359 if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
360 pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
361 pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
366 mutex_unlock(&dev->dev_lock);
370 static int _ocrdma_alloc_pd(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
371 struct ocrdma_ucontext *uctx,
372 struct ib_udata *udata)
376 if (udata && uctx && dev->attr.max_dpp_pds) {
378 ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
380 pd->dpp_enabled ? (dev->nic_info.db_page_size /
381 dev->attr.wqe_size) : 0;
384 if (dev->pd_mgr->pd_prealloc_valid)
385 return ocrdma_get_pd_num(dev, pd);
388 status = ocrdma_mbx_alloc_pd(dev, pd);
390 if (pd->dpp_enabled) {
391 pd->dpp_enabled = false;
401 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
402 struct ocrdma_pd *pd)
404 return (uctx->cntxt_pd == pd);
407 static void _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
408 struct ocrdma_pd *pd)
410 if (dev->pd_mgr->pd_prealloc_valid)
411 ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
413 ocrdma_mbx_dealloc_pd(dev, pd);
416 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
417 struct ocrdma_ucontext *uctx,
418 struct ib_udata *udata)
420 struct ib_device *ibdev = &dev->ibdev;
424 pd = rdma_zalloc_drv_obj(ibdev, ib_pd);
429 uctx->cntxt_pd = get_ocrdma_pd(pd);
431 status = _ocrdma_alloc_pd(dev, uctx->cntxt_pd, uctx, udata);
433 kfree(uctx->cntxt_pd);
437 uctx->cntxt_pd->uctx = uctx;
438 uctx->cntxt_pd->ibpd.device = &dev->ibdev;
443 static void ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
445 struct ocrdma_pd *pd = uctx->cntxt_pd;
446 struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
448 if (uctx->pd_in_use) {
449 pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
450 __func__, dev->id, pd->id);
452 kfree(uctx->cntxt_pd);
453 uctx->cntxt_pd = NULL;
454 _ocrdma_dealloc_pd(dev, pd);
457 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
459 struct ocrdma_pd *pd = NULL;
461 mutex_lock(&uctx->mm_list_lock);
462 if (!uctx->pd_in_use) {
463 uctx->pd_in_use = true;
466 mutex_unlock(&uctx->mm_list_lock);
471 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
473 mutex_lock(&uctx->mm_list_lock);
474 uctx->pd_in_use = false;
475 mutex_unlock(&uctx->mm_list_lock);
478 int ocrdma_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
480 struct ib_device *ibdev = uctx->device;
482 struct ocrdma_ucontext *ctx = get_ocrdma_ucontext(uctx);
483 struct ocrdma_alloc_ucontext_resp resp = {};
484 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
485 struct pci_dev *pdev = dev->nic_info.pdev;
486 u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
490 INIT_LIST_HEAD(&ctx->mm_head);
491 mutex_init(&ctx->mm_list_lock);
493 ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
494 &ctx->ah_tbl.pa, GFP_KERNEL);
498 ctx->ah_tbl.len = map_len;
500 resp.ah_tbl_len = ctx->ah_tbl.len;
501 resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
503 status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
507 status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
511 resp.dev_id = dev->id;
512 resp.max_inline_data = dev->attr.max_inline_data;
513 resp.wqe_size = dev->attr.wqe_size;
514 resp.rqe_size = dev->attr.rqe_size;
515 resp.dpp_wqe_size = dev->attr.wqe_size;
517 memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
518 status = ib_copy_to_udata(udata, &resp, sizeof(resp));
524 ocrdma_dealloc_ucontext_pd(ctx);
526 ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
528 dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
533 void ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
535 struct ocrdma_mm *mm, *tmp;
536 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
537 struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
538 struct pci_dev *pdev = dev->nic_info.pdev;
540 ocrdma_dealloc_ucontext_pd(uctx);
542 ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
543 dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
546 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
547 list_del(&mm->entry);
552 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
554 struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
555 struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
556 unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
557 u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
558 unsigned long len = (vma->vm_end - vma->vm_start);
562 if (vma->vm_start & (PAGE_SIZE - 1))
564 found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
568 if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
569 dev->nic_info.db_total_size)) &&
570 (len <= dev->nic_info.db_page_size)) {
571 if (vma->vm_flags & VM_READ)
574 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
575 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
576 len, vma->vm_page_prot);
577 } else if (dev->nic_info.dpp_unmapped_len &&
578 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
579 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
580 dev->nic_info.dpp_unmapped_len)) &&
581 (len <= dev->nic_info.dpp_unmapped_len)) {
582 if (vma->vm_flags & VM_READ)
585 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
586 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
587 len, vma->vm_page_prot);
589 status = remap_pfn_range(vma, vma->vm_start,
590 vma->vm_pgoff, len, vma->vm_page_prot);
595 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
596 struct ib_ucontext *ib_ctx,
597 struct ib_udata *udata)
601 u64 dpp_page_addr = 0;
603 struct ocrdma_alloc_pd_uresp rsp;
604 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
606 memset(&rsp, 0, sizeof(rsp));
608 rsp.dpp_enabled = pd->dpp_enabled;
609 db_page_addr = ocrdma_get_db_addr(dev, pd->id);
610 db_page_size = dev->nic_info.db_page_size;
612 status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
616 if (pd->dpp_enabled) {
617 dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
618 (pd->id * PAGE_SIZE);
619 status = ocrdma_add_mmap(uctx, dpp_page_addr,
623 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
624 rsp.dpp_page_addr_lo = dpp_page_addr;
627 status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
636 ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
638 ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
642 int ocrdma_alloc_pd(struct ib_pd *ibpd, struct ib_ucontext *context,
643 struct ib_udata *udata)
645 struct ib_device *ibdev = ibpd->device;
646 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
647 struct ocrdma_pd *pd;
648 struct ocrdma_ucontext *uctx = NULL;
650 u8 is_uctx_pd = false;
652 if (udata && context) {
653 uctx = get_ocrdma_ucontext(context);
654 pd = ocrdma_get_ucontext_pd(uctx);
661 pd = get_ocrdma_pd(ibpd);
662 status = _ocrdma_alloc_pd(dev, pd, uctx, udata);
667 if (udata && context) {
668 status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
676 ocrdma_release_ucontext_pd(uctx);
678 _ocrdma_dealloc_pd(dev, pd);
683 void ocrdma_dealloc_pd(struct ib_pd *ibpd)
685 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
686 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
687 struct ocrdma_ucontext *uctx = NULL;
692 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
693 (pd->id * PAGE_SIZE);
695 ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
696 usr_db = ocrdma_get_db_addr(dev, pd->id);
697 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
699 if (is_ucontext_pd(uctx, pd)) {
700 ocrdma_release_ucontext_pd(uctx);
704 _ocrdma_dealloc_pd(dev, pd);
707 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
708 u32 pdid, int acc, u32 num_pbls, u32 addr_check)
713 mr->hwmr.local_rd = 1;
714 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
715 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
716 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
717 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
718 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
719 mr->hwmr.num_pbls = num_pbls;
721 status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
725 mr->ibmr.lkey = mr->hwmr.lkey;
726 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
727 mr->ibmr.rkey = mr->hwmr.lkey;
731 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
734 struct ocrdma_mr *mr;
735 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
736 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
738 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
739 pr_err("%s err, invalid access rights\n", __func__);
740 return ERR_PTR(-EINVAL);
743 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
745 return ERR_PTR(-ENOMEM);
747 status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
748 OCRDMA_ADDR_CHECK_DISABLE);
751 return ERR_PTR(status);
757 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
758 struct ocrdma_hw_mr *mr)
760 struct pci_dev *pdev = dev->nic_info.pdev;
764 for (i = 0; i < mr->num_pbls; i++) {
765 if (!mr->pbl_table[i].va)
767 dma_free_coherent(&pdev->dev, mr->pbl_size,
769 mr->pbl_table[i].pa);
771 kfree(mr->pbl_table);
772 mr->pbl_table = NULL;
776 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
785 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
786 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
790 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
791 num_pbls = num_pbls / (pbl_size / sizeof(u64));
793 } while (num_pbls >= dev->attr.max_num_mr_pbl);
795 mr->hwmr.num_pbes = num_pbes;
796 mr->hwmr.num_pbls = num_pbls;
797 mr->hwmr.pbl_size = pbl_size;
801 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
805 u32 dma_len = mr->pbl_size;
806 struct pci_dev *pdev = dev->nic_info.pdev;
810 mr->pbl_table = kcalloc(mr->num_pbls, sizeof(struct ocrdma_pbl),
816 for (i = 0; i < mr->num_pbls; i++) {
817 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
819 ocrdma_free_mr_pbl_tbl(dev, mr);
823 mr->pbl_table[i].va = va;
824 mr->pbl_table[i].pa = pa;
829 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
832 struct ocrdma_pbe *pbe;
833 struct sg_dma_page_iter sg_iter;
834 struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
835 struct ib_umem *umem = mr->umem;
836 int pbe_cnt, total_num_pbes = 0;
839 if (!mr->hwmr.num_pbes)
842 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
845 for_each_sg_dma_page (umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
846 /* store the page address in pbe */
847 pg_addr = sg_page_iter_dma_address(&sg_iter);
848 pbe->pa_lo = cpu_to_le32(pg_addr);
849 pbe->pa_hi = cpu_to_le32(upper_32_bits(pg_addr));
854 /* if done building pbes, issue the mbx cmd. */
855 if (total_num_pbes == num_pbes)
858 /* if the given pbl is full storing the pbes,
861 if (pbe_cnt == (mr->hwmr.pbl_size / sizeof(u64))) {
863 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
869 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
870 u64 usr_addr, int acc, struct ib_udata *udata)
872 int status = -ENOMEM;
873 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
874 struct ocrdma_mr *mr;
875 struct ocrdma_pd *pd;
878 pd = get_ocrdma_pd(ibpd);
880 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
881 return ERR_PTR(-EINVAL);
883 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
885 return ERR_PTR(status);
886 mr->umem = ib_umem_get(udata, start, len, acc, 0);
887 if (IS_ERR(mr->umem)) {
891 num_pbes = ib_umem_page_count(mr->umem);
892 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
896 mr->hwmr.pbe_size = PAGE_SIZE;
897 mr->hwmr.fbo = ib_umem_offset(mr->umem);
898 mr->hwmr.va = usr_addr;
900 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
901 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
902 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
903 mr->hwmr.local_rd = 1;
904 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
905 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
908 build_user_pbes(dev, mr, num_pbes);
909 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
912 mr->ibmr.lkey = mr->hwmr.lkey;
913 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
914 mr->ibmr.rkey = mr->hwmr.lkey;
919 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
922 return ERR_PTR(status);
925 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
927 struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
928 struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
930 (void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
933 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
935 /* it could be user registered memory. */
937 ib_umem_release(mr->umem);
940 /* Don't stop cleanup, in case FW is unresponsive */
941 if (dev->mqe_ctx.fw_error_state) {
942 pr_err("%s(%d) fw not responding.\n",
948 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
949 struct ib_udata *udata,
950 struct ib_ucontext *ib_ctx)
953 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
954 struct ocrdma_create_cq_uresp uresp;
956 memset(&uresp, 0, sizeof(uresp));
957 uresp.cq_id = cq->id;
958 uresp.page_size = PAGE_ALIGN(cq->len);
960 uresp.max_hw_cqe = cq->max_hw_cqe;
961 uresp.page_addr[0] = virt_to_phys(cq->va);
962 uresp.db_page_addr = ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
963 uresp.db_page_size = dev->nic_info.db_page_size;
964 uresp.phase_change = cq->phase_change ? 1 : 0;
965 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
967 pr_err("%s(%d) copy error cqid=0x%x.\n",
968 __func__, dev->id, cq->id);
971 status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
974 status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
976 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
984 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev,
985 const struct ib_cq_init_attr *attr,
986 struct ib_ucontext *ib_ctx,
987 struct ib_udata *udata)
989 int entries = attr->cqe;
990 struct ocrdma_cq *cq;
991 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
992 struct ocrdma_ucontext *uctx = NULL;
995 struct ocrdma_create_cq_ureq ureq;
998 return ERR_PTR(-EINVAL);
1001 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1002 return ERR_PTR(-EFAULT);
1005 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
1007 return ERR_PTR(-ENOMEM);
1009 spin_lock_init(&cq->cq_lock);
1010 spin_lock_init(&cq->comp_handler_lock);
1011 INIT_LIST_HEAD(&cq->sq_head);
1012 INIT_LIST_HEAD(&cq->rq_head);
1015 uctx = get_ocrdma_ucontext(ib_ctx);
1016 pd_id = uctx->cntxt_pd->id;
1019 status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1022 return ERR_PTR(status);
1025 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
1029 cq->phase = OCRDMA_CQE_VALID;
1030 dev->cq_tbl[cq->id] = cq;
1034 ocrdma_mbx_destroy_cq(dev, cq);
1036 return ERR_PTR(status);
1039 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1040 struct ib_udata *udata)
1043 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1045 if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1049 ibcq->cqe = new_cnt;
1053 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1056 int valid_count = 0;
1057 unsigned long flags;
1059 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1060 struct ocrdma_cqe *cqe = NULL;
1063 cqe_cnt = cq->cqe_cnt;
1065 /* Last irq might have scheduled a polling thread
1066 * sync-up with it before hard flushing.
1068 spin_lock_irqsave(&cq->cq_lock, flags);
1070 if (is_cqe_valid(cq, cqe))
1075 ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1076 spin_unlock_irqrestore(&cq->cq_lock, flags);
1079 int ocrdma_destroy_cq(struct ib_cq *ibcq)
1081 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1082 struct ocrdma_eq *eq = NULL;
1083 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1087 dev->cq_tbl[cq->id] = NULL;
1088 indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1089 BUG_ON(indx == -EINVAL);
1091 eq = &dev->eq_tbl[indx];
1092 irq = ocrdma_get_irq(dev, eq);
1093 synchronize_irq(irq);
1094 ocrdma_flush_cq(cq);
1096 (void)ocrdma_mbx_destroy_cq(dev, cq);
1098 pdid = cq->ucontext->cntxt_pd->id;
1099 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1100 PAGE_ALIGN(cq->len));
1101 ocrdma_del_mmap(cq->ucontext,
1102 ocrdma_get_db_addr(dev, pdid),
1103 dev->nic_info.db_page_size);
1110 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1112 int status = -EINVAL;
1114 if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1115 dev->qp_tbl[qp->id] = qp;
1121 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1123 dev->qp_tbl[qp->id] = NULL;
1126 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1127 struct ib_qp_init_attr *attrs,
1128 struct ib_udata *udata)
1130 if ((attrs->qp_type != IB_QPT_GSI) &&
1131 (attrs->qp_type != IB_QPT_RC) &&
1132 (attrs->qp_type != IB_QPT_UC) &&
1133 (attrs->qp_type != IB_QPT_UD)) {
1134 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1135 __func__, dev->id, attrs->qp_type);
1138 /* Skip the check for QP1 to support CM size of 128 */
1139 if ((attrs->qp_type != IB_QPT_GSI) &&
1140 (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1141 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1142 __func__, dev->id, attrs->cap.max_send_wr);
1143 pr_err("%s(%d) supported send_wr=0x%x\n",
1144 __func__, dev->id, dev->attr.max_wqe);
1147 if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1148 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1149 __func__, dev->id, attrs->cap.max_recv_wr);
1150 pr_err("%s(%d) supported recv_wr=0x%x\n",
1151 __func__, dev->id, dev->attr.max_rqe);
1154 if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1155 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1156 __func__, dev->id, attrs->cap.max_inline_data);
1157 pr_err("%s(%d) supported inline data size=0x%x\n",
1158 __func__, dev->id, dev->attr.max_inline_data);
1161 if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1162 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1163 __func__, dev->id, attrs->cap.max_send_sge);
1164 pr_err("%s(%d) supported send_sge=0x%x\n",
1165 __func__, dev->id, dev->attr.max_send_sge);
1168 if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1169 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1170 __func__, dev->id, attrs->cap.max_recv_sge);
1171 pr_err("%s(%d) supported recv_sge=0x%x\n",
1172 __func__, dev->id, dev->attr.max_recv_sge);
1175 /* unprivileged user space cannot create special QP */
1176 if (udata && attrs->qp_type == IB_QPT_GSI) {
1178 ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1179 __func__, dev->id, attrs->qp_type);
1182 /* allow creating only one GSI type of QP */
1183 if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1184 pr_err("%s(%d) GSI special QPs already created.\n",
1188 /* verify consumer QPs are not trying to use GSI QP's CQ */
1189 if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1190 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1191 (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1192 pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1200 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1201 struct ib_udata *udata, int dpp_offset,
1202 int dpp_credit_lmt, int srq)
1206 struct ocrdma_create_qp_uresp uresp;
1207 struct ocrdma_pd *pd = qp->pd;
1208 struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1210 memset(&uresp, 0, sizeof(uresp));
1211 usr_db = dev->nic_info.unmapped_db +
1212 (pd->id * dev->nic_info.db_page_size);
1213 uresp.qp_id = qp->id;
1214 uresp.sq_dbid = qp->sq.dbid;
1215 uresp.num_sq_pages = 1;
1216 uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1217 uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1218 uresp.num_wqe_allocated = qp->sq.max_cnt;
1220 uresp.rq_dbid = qp->rq.dbid;
1221 uresp.num_rq_pages = 1;
1222 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1223 uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1224 uresp.num_rqe_allocated = qp->rq.max_cnt;
1226 uresp.db_page_addr = usr_db;
1227 uresp.db_page_size = dev->nic_info.db_page_size;
1228 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1229 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1230 uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1232 if (qp->dpp_enabled) {
1233 uresp.dpp_credit = dpp_credit_lmt;
1234 uresp.dpp_offset = dpp_offset;
1236 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1238 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1241 status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1242 uresp.sq_page_size);
1247 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1248 uresp.rq_page_size);
1254 ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1259 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1260 struct ocrdma_pd *pd)
1262 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1263 qp->sq_db = dev->nic_info.db +
1264 (pd->id * dev->nic_info.db_page_size) +
1265 OCRDMA_DB_GEN2_SQ_OFFSET;
1266 qp->rq_db = dev->nic_info.db +
1267 (pd->id * dev->nic_info.db_page_size) +
1268 OCRDMA_DB_GEN2_RQ_OFFSET;
1270 qp->sq_db = dev->nic_info.db +
1271 (pd->id * dev->nic_info.db_page_size) +
1272 OCRDMA_DB_SQ_OFFSET;
1273 qp->rq_db = dev->nic_info.db +
1274 (pd->id * dev->nic_info.db_page_size) +
1275 OCRDMA_DB_RQ_OFFSET;
1279 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1282 kcalloc(qp->sq.max_cnt, sizeof(*(qp->wqe_wr_id_tbl)),
1284 if (qp->wqe_wr_id_tbl == NULL)
1287 kcalloc(qp->rq.max_cnt, sizeof(u64), GFP_KERNEL);
1288 if (qp->rqe_wr_id_tbl == NULL)
1294 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1295 struct ocrdma_pd *pd,
1296 struct ib_qp_init_attr *attrs)
1299 spin_lock_init(&qp->q_lock);
1300 INIT_LIST_HEAD(&qp->sq_entry);
1301 INIT_LIST_HEAD(&qp->rq_entry);
1303 qp->qp_type = attrs->qp_type;
1304 qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1305 qp->max_inline_data = attrs->cap.max_inline_data;
1306 qp->sq.max_sges = attrs->cap.max_send_sge;
1307 qp->rq.max_sges = attrs->cap.max_recv_sge;
1308 qp->state = OCRDMA_QPS_RST;
1309 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1312 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1313 struct ib_qp_init_attr *attrs)
1315 if (attrs->qp_type == IB_QPT_GSI) {
1316 dev->gsi_qp_created = 1;
1317 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1318 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1322 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1323 struct ib_qp_init_attr *attrs,
1324 struct ib_udata *udata)
1327 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1328 struct ocrdma_qp *qp;
1329 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1330 struct ocrdma_create_qp_ureq ureq;
1331 u16 dpp_credit_lmt, dpp_offset;
1333 status = ocrdma_check_qp_params(ibpd, dev, attrs, udata);
1337 memset(&ureq, 0, sizeof(ureq));
1339 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1340 return ERR_PTR(-EFAULT);
1342 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1347 ocrdma_set_qp_init_params(qp, pd, attrs);
1349 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1350 OCRDMA_QP_FAST_REG);
1352 mutex_lock(&dev->dev_lock);
1353 status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1355 &dpp_offset, &dpp_credit_lmt);
1359 /* user space QP's wr_id table are managed in library */
1360 if (udata == NULL) {
1361 status = ocrdma_alloc_wr_id_tbl(qp);
1366 status = ocrdma_add_qpn_map(dev, qp);
1369 ocrdma_set_qp_db(dev, qp, pd);
1371 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1373 (attrs->srq != NULL));
1377 ocrdma_store_gsi_qp_cq(dev, attrs);
1378 qp->ibqp.qp_num = qp->id;
1379 mutex_unlock(&dev->dev_lock);
1383 ocrdma_del_qpn_map(dev, qp);
1385 ocrdma_mbx_destroy_qp(dev, qp);
1387 mutex_unlock(&dev->dev_lock);
1388 kfree(qp->wqe_wr_id_tbl);
1389 kfree(qp->rqe_wr_id_tbl);
1391 pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1393 return ERR_PTR(status);
1396 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1400 struct ocrdma_qp *qp;
1401 struct ocrdma_dev *dev;
1402 enum ib_qp_state old_qps;
1404 qp = get_ocrdma_qp(ibqp);
1405 dev = get_ocrdma_dev(ibqp->device);
1406 if (attr_mask & IB_QP_STATE)
1407 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1408 /* if new and previous states are same hw doesn't need to
1413 return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1416 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1417 int attr_mask, struct ib_udata *udata)
1419 unsigned long flags;
1420 int status = -EINVAL;
1421 struct ocrdma_qp *qp;
1422 struct ocrdma_dev *dev;
1423 enum ib_qp_state old_qps, new_qps;
1425 qp = get_ocrdma_qp(ibqp);
1426 dev = get_ocrdma_dev(ibqp->device);
1428 /* syncronize with multiple context trying to change, retrive qps */
1429 mutex_lock(&dev->dev_lock);
1430 /* syncronize with wqe, rqe posting and cqe processing contexts */
1431 spin_lock_irqsave(&qp->q_lock, flags);
1432 old_qps = get_ibqp_state(qp->state);
1433 if (attr_mask & IB_QP_STATE)
1434 new_qps = attr->qp_state;
1437 spin_unlock_irqrestore(&qp->q_lock, flags);
1439 if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1440 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1441 "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1442 __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1447 status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1451 mutex_unlock(&dev->dev_lock);
1455 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1473 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1475 int ib_qp_acc_flags = 0;
1477 if (qp_cap_flags & OCRDMA_QP_INB_WR)
1478 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1479 if (qp_cap_flags & OCRDMA_QP_INB_RD)
1480 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1481 return ib_qp_acc_flags;
1484 int ocrdma_query_qp(struct ib_qp *ibqp,
1485 struct ib_qp_attr *qp_attr,
1486 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1490 struct ocrdma_qp_params params;
1491 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1492 struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1494 memset(¶ms, 0, sizeof(params));
1495 mutex_lock(&dev->dev_lock);
1496 status = ocrdma_mbx_query_qp(dev, qp, ¶ms);
1497 mutex_unlock(&dev->dev_lock);
1500 if (qp->qp_type == IB_QPT_UD)
1501 qp_attr->qkey = params.qkey;
1503 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1504 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1505 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1506 qp_attr->path_mig_state = IB_MIG_MIGRATED;
1507 qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1508 qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1509 qp_attr->dest_qp_num =
1510 params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1512 qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1513 qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1514 qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1515 qp_attr->cap.max_send_sge = qp->sq.max_sges;
1516 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1517 qp_attr->cap.max_inline_data = qp->max_inline_data;
1518 qp_init_attr->cap = qp_attr->cap;
1519 qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1521 rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1522 params.rnt_rc_sl_fl &
1523 OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1525 (params.hop_lmt_rq_psn &
1526 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1527 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1528 (params.tclass_sq_psn &
1529 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1530 OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1531 rdma_ah_set_dgid_raw(&qp_attr->ah_attr, ¶ms.dgid[0]);
1533 rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1534 rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1535 OCRDMA_QP_PARAMS_SL_MASK) >>
1536 OCRDMA_QP_PARAMS_SL_SHIFT);
1537 qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1538 OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1539 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1540 qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1541 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1542 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1543 qp_attr->retry_cnt =
1544 (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1545 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1546 qp_attr->min_rnr_timer = 0;
1547 qp_attr->pkey_index = 0;
1548 qp_attr->port_num = 1;
1549 rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1550 rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1551 qp_attr->alt_pkey_index = 0;
1552 qp_attr->alt_port_num = 0;
1553 qp_attr->alt_timeout = 0;
1554 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1555 qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1556 OCRDMA_QP_PARAMS_STATE_SHIFT;
1557 qp_attr->qp_state = get_ibqp_state(qp_state);
1558 qp_attr->cur_qp_state = qp_attr->qp_state;
1559 qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1560 qp_attr->max_dest_rd_atomic =
1561 params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1562 qp_attr->max_rd_atomic =
1563 params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1564 qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1565 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1566 /* Sync driver QP state with FW */
1567 ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1572 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1574 unsigned int i = idx / 32;
1575 u32 mask = (1U << (idx % 32));
1577 srq->idx_bit_fields[i] ^= mask;
1580 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1582 return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1585 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1587 return (qp->sq.tail == qp->sq.head);
1590 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1592 return (qp->rq.tail == qp->rq.head);
1595 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1597 return q->va + (q->head * q->entry_size);
1600 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1603 return q->va + (idx * q->entry_size);
1606 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1608 q->head = (q->head + 1) & q->max_wqe_idx;
1611 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1613 q->tail = (q->tail + 1) & q->max_wqe_idx;
1616 /* discard the cqe for a given QP */
1617 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1619 unsigned long cq_flags;
1620 unsigned long flags;
1621 int discard_cnt = 0;
1622 u32 cur_getp, stop_getp;
1623 struct ocrdma_cqe *cqe;
1624 u32 qpn = 0, wqe_idx = 0;
1626 spin_lock_irqsave(&cq->cq_lock, cq_flags);
1628 /* traverse through the CQEs in the hw CQ,
1629 * find the matching CQE for a given qp,
1630 * mark the matching one discarded by clearing qpn.
1631 * ring the doorbell in the poll_cq() as
1632 * we don't complete out of order cqe.
1635 cur_getp = cq->getp;
1636 /* find upto when do we reap the cq. */
1637 stop_getp = cur_getp;
1639 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1642 cqe = cq->va + cur_getp;
1643 /* if (a) done reaping whole hw cq, or
1644 * (b) qp_xq becomes empty.
1647 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1648 /* if previously discarded cqe found, skip that too. */
1649 /* check for matching qp */
1650 if (qpn == 0 || qpn != qp->id)
1653 if (is_cqe_for_sq(cqe)) {
1654 ocrdma_hwq_inc_tail(&qp->sq);
1657 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1658 OCRDMA_CQE_BUFTAG_SHIFT) &
1659 qp->srq->rq.max_wqe_idx;
1660 BUG_ON(wqe_idx < 1);
1661 spin_lock_irqsave(&qp->srq->q_lock, flags);
1662 ocrdma_hwq_inc_tail(&qp->srq->rq);
1663 ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1664 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1667 ocrdma_hwq_inc_tail(&qp->rq);
1670 /* mark cqe discarded so that it is not picked up later
1676 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1677 } while (cur_getp != stop_getp);
1678 spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1681 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1684 unsigned long flags;
1685 struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1686 /* sync with any active CQ poll */
1688 spin_lock_irqsave(&dev->flush_q_lock, flags);
1689 found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1691 list_del(&qp->sq_entry);
1693 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1695 list_del(&qp->rq_entry);
1697 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1700 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1702 struct ocrdma_pd *pd;
1703 struct ocrdma_qp *qp;
1704 struct ocrdma_dev *dev;
1705 struct ib_qp_attr attrs;
1707 unsigned long flags;
1709 qp = get_ocrdma_qp(ibqp);
1710 dev = get_ocrdma_dev(ibqp->device);
1714 /* change the QP state to ERROR */
1715 if (qp->state != OCRDMA_QPS_RST) {
1716 attrs.qp_state = IB_QPS_ERR;
1717 attr_mask = IB_QP_STATE;
1718 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1720 /* ensure that CQEs for newly created QP (whose id may be same with
1721 * one which just getting destroyed are same), dont get
1722 * discarded until the old CQEs are discarded.
1724 mutex_lock(&dev->dev_lock);
1725 (void) ocrdma_mbx_destroy_qp(dev, qp);
1728 * acquire CQ lock while destroy is in progress, in order to
1729 * protect against proessing in-flight CQEs for this QP.
1731 spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1732 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq)) {
1733 spin_lock(&qp->rq_cq->cq_lock);
1734 ocrdma_del_qpn_map(dev, qp);
1735 spin_unlock(&qp->rq_cq->cq_lock);
1737 ocrdma_del_qpn_map(dev, qp);
1739 spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1742 ocrdma_discard_cqes(qp, qp->sq_cq);
1743 ocrdma_discard_cqes(qp, qp->rq_cq);
1745 mutex_unlock(&dev->dev_lock);
1748 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1749 PAGE_ALIGN(qp->sq.len));
1751 ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1752 PAGE_ALIGN(qp->rq.len));
1755 ocrdma_del_flush_qp(qp);
1757 kfree(qp->wqe_wr_id_tbl);
1758 kfree(qp->rqe_wr_id_tbl);
1763 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1764 struct ib_udata *udata)
1767 struct ocrdma_create_srq_uresp uresp;
1769 memset(&uresp, 0, sizeof(uresp));
1770 uresp.rq_dbid = srq->rq.dbid;
1771 uresp.num_rq_pages = 1;
1772 uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1773 uresp.rq_page_size = srq->rq.len;
1774 uresp.db_page_addr = dev->nic_info.unmapped_db +
1775 (srq->pd->id * dev->nic_info.db_page_size);
1776 uresp.db_page_size = dev->nic_info.db_page_size;
1777 uresp.num_rqe_allocated = srq->rq.max_cnt;
1778 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1779 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1780 uresp.db_shift = 24;
1782 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1783 uresp.db_shift = 16;
1786 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1789 status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1790 uresp.rq_page_size);
1796 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1797 struct ib_srq_init_attr *init_attr,
1798 struct ib_udata *udata)
1800 int status = -ENOMEM;
1801 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1802 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1803 struct ocrdma_srq *srq;
1805 if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1806 return ERR_PTR(-EINVAL);
1807 if (init_attr->attr.max_wr > dev->attr.max_rqe)
1808 return ERR_PTR(-EINVAL);
1810 srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1812 return ERR_PTR(status);
1814 spin_lock_init(&srq->q_lock);
1816 srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1817 status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1821 if (udata == NULL) {
1823 srq->rqe_wr_id_tbl = kcalloc(srq->rq.max_cnt, sizeof(u64),
1825 if (srq->rqe_wr_id_tbl == NULL)
1828 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1829 (srq->rq.max_cnt % 32 ? 1 : 0);
1830 srq->idx_bit_fields =
1831 kmalloc_array(srq->bit_fields_len, sizeof(u32),
1833 if (srq->idx_bit_fields == NULL)
1835 memset(srq->idx_bit_fields, 0xff,
1836 srq->bit_fields_len * sizeof(u32));
1839 if (init_attr->attr.srq_limit) {
1840 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1846 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1854 ocrdma_mbx_destroy_srq(dev, srq);
1856 kfree(srq->rqe_wr_id_tbl);
1857 kfree(srq->idx_bit_fields);
1859 return ERR_PTR(status);
1862 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1863 struct ib_srq_attr *srq_attr,
1864 enum ib_srq_attr_mask srq_attr_mask,
1865 struct ib_udata *udata)
1868 struct ocrdma_srq *srq;
1870 srq = get_ocrdma_srq(ibsrq);
1871 if (srq_attr_mask & IB_SRQ_MAX_WR)
1874 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1878 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1881 struct ocrdma_srq *srq;
1883 srq = get_ocrdma_srq(ibsrq);
1884 status = ocrdma_mbx_query_srq(srq, srq_attr);
1888 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1891 struct ocrdma_srq *srq;
1892 struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1894 srq = get_ocrdma_srq(ibsrq);
1896 status = ocrdma_mbx_destroy_srq(dev, srq);
1899 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1900 PAGE_ALIGN(srq->rq.len));
1902 kfree(srq->idx_bit_fields);
1903 kfree(srq->rqe_wr_id_tbl);
1908 /* unprivileged verbs and their support functions. */
1909 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1910 struct ocrdma_hdr_wqe *hdr,
1911 const struct ib_send_wr *wr)
1913 struct ocrdma_ewqe_ud_hdr *ud_hdr =
1914 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1915 struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1917 ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1918 if (qp->qp_type == IB_QPT_GSI)
1919 ud_hdr->qkey = qp->qkey;
1921 ud_hdr->qkey = ud_wr(wr)->remote_qkey;
1922 ud_hdr->rsvd_ahid = ah->id;
1923 ud_hdr->hdr_type = ah->hdr_type;
1924 if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
1925 hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
1928 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1929 struct ocrdma_sge *sge, int num_sge,
1930 struct ib_sge *sg_list)
1934 for (i = 0; i < num_sge; i++) {
1935 sge[i].lrkey = sg_list[i].lkey;
1936 sge[i].addr_lo = sg_list[i].addr;
1937 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1938 sge[i].len = sg_list[i].length;
1939 hdr->total_len += sg_list[i].length;
1942 memset(sge, 0, sizeof(*sge));
1945 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1947 uint32_t total_len = 0, i;
1949 for (i = 0; i < num_sge; i++)
1950 total_len += sg_list[i].length;
1955 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1956 struct ocrdma_hdr_wqe *hdr,
1957 struct ocrdma_sge *sge,
1958 const struct ib_send_wr *wr, u32 wqe_size)
1963 if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1964 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1965 if (unlikely(hdr->total_len > qp->max_inline_data)) {
1966 pr_err("%s() supported_len=0x%x,\n"
1967 " unsupported len req=0x%x\n", __func__,
1968 qp->max_inline_data, hdr->total_len);
1971 dpp_addr = (char *)sge;
1972 for (i = 0; i < wr->num_sge; i++) {
1974 (void *)(unsigned long)wr->sg_list[i].addr,
1975 wr->sg_list[i].length);
1976 dpp_addr += wr->sg_list[i].length;
1979 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1980 if (0 == hdr->total_len)
1981 wqe_size += sizeof(struct ocrdma_sge);
1982 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1984 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1986 wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1988 wqe_size += sizeof(struct ocrdma_sge);
1989 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1991 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1995 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1996 const struct ib_send_wr *wr)
1999 struct ocrdma_sge *sge;
2000 u32 wqe_size = sizeof(*hdr);
2002 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2003 ocrdma_build_ud_hdr(qp, hdr, wr);
2004 sge = (struct ocrdma_sge *)(hdr + 2);
2005 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
2007 sge = (struct ocrdma_sge *)(hdr + 1);
2010 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2014 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2015 const struct ib_send_wr *wr)
2018 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2019 struct ocrdma_sge *sge = ext_rw + 1;
2020 u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
2022 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2025 ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2026 ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2027 ext_rw->lrkey = rdma_wr(wr)->rkey;
2028 ext_rw->len = hdr->total_len;
2032 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2033 const struct ib_send_wr *wr)
2035 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2036 struct ocrdma_sge *sge = ext_rw + 1;
2037 u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2038 sizeof(struct ocrdma_hdr_wqe);
2040 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2041 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2042 hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2043 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2045 ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2046 ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2047 ext_rw->lrkey = rdma_wr(wr)->rkey;
2048 ext_rw->len = hdr->total_len;
2051 static int get_encoded_page_size(int pg_sz)
2053 /* Max size is 256M 4096 << 16 */
2056 if (pg_sz == (4096 << i))
2061 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2062 struct ocrdma_hdr_wqe *hdr,
2063 const struct ib_reg_wr *wr)
2066 struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2067 struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2068 struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2069 struct ocrdma_pbe *pbe;
2070 u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2071 int num_pbes = 0, i;
2073 wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2075 hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2076 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2078 if (wr->access & IB_ACCESS_LOCAL_WRITE)
2079 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2080 if (wr->access & IB_ACCESS_REMOTE_WRITE)
2081 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2082 if (wr->access & IB_ACCESS_REMOTE_READ)
2083 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2084 hdr->lkey = wr->key;
2085 hdr->total_len = mr->ibmr.length;
2087 fbo = mr->ibmr.iova - mr->pages[0];
2089 fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2090 fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2091 fast_reg->fbo_hi = upper_32_bits(fbo);
2092 fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2093 fast_reg->num_sges = mr->npages;
2094 fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2097 for (i = 0; i < mr->npages; i++) {
2098 u64 buf_addr = mr->pages[i];
2100 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2101 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2105 /* if the pbl is full storing the pbes,
2108 if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2110 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2117 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2119 u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2121 iowrite32(val, qp->sq_db);
2124 int ocrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
2125 const struct ib_send_wr **bad_wr)
2128 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2129 struct ocrdma_hdr_wqe *hdr;
2130 unsigned long flags;
2132 spin_lock_irqsave(&qp->q_lock, flags);
2133 if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2134 spin_unlock_irqrestore(&qp->q_lock, flags);
2140 if (qp->qp_type == IB_QPT_UD &&
2141 (wr->opcode != IB_WR_SEND &&
2142 wr->opcode != IB_WR_SEND_WITH_IMM)) {
2147 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2148 wr->num_sge > qp->sq.max_sges) {
2153 hdr = ocrdma_hwq_head(&qp->sq);
2155 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2156 hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2157 if (wr->send_flags & IB_SEND_FENCE)
2159 (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2160 if (wr->send_flags & IB_SEND_SOLICITED)
2162 (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2164 switch (wr->opcode) {
2165 case IB_WR_SEND_WITH_IMM:
2166 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2167 hdr->immdt = ntohl(wr->ex.imm_data);
2170 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2171 ocrdma_build_send(qp, hdr, wr);
2173 case IB_WR_SEND_WITH_INV:
2174 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2175 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2176 hdr->lkey = wr->ex.invalidate_rkey;
2177 status = ocrdma_build_send(qp, hdr, wr);
2179 case IB_WR_RDMA_WRITE_WITH_IMM:
2180 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2181 hdr->immdt = ntohl(wr->ex.imm_data);
2183 case IB_WR_RDMA_WRITE:
2184 hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2185 status = ocrdma_build_write(qp, hdr, wr);
2187 case IB_WR_RDMA_READ:
2188 ocrdma_build_read(qp, hdr, wr);
2190 case IB_WR_LOCAL_INV:
2192 (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2193 hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2194 sizeof(struct ocrdma_sge)) /
2195 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2196 hdr->lkey = wr->ex.invalidate_rkey;
2199 status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2209 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2210 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2212 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2213 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2214 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2215 OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2216 /* make sure wqe is written before adapter can access it */
2218 /* inform hw to start processing it */
2219 ocrdma_ring_sq_db(qp);
2221 /* update pointer, counter for next wr */
2222 ocrdma_hwq_inc_head(&qp->sq);
2225 spin_unlock_irqrestore(&qp->q_lock, flags);
2229 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2231 u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2233 iowrite32(val, qp->rq_db);
2236 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe,
2237 const struct ib_recv_wr *wr, u16 tag)
2240 struct ocrdma_sge *sge;
2242 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2244 wqe_size = sizeof(*sge) + sizeof(*rqe);
2246 rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2247 OCRDMA_WQE_SIZE_SHIFT);
2248 rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2249 rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2251 rqe->rsvd_tag = tag;
2252 sge = (struct ocrdma_sge *)(rqe + 1);
2253 ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2254 ocrdma_cpu_to_le32(rqe, wqe_size);
2257 int ocrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2258 const struct ib_recv_wr **bad_wr)
2261 unsigned long flags;
2262 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2263 struct ocrdma_hdr_wqe *rqe;
2265 spin_lock_irqsave(&qp->q_lock, flags);
2266 if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2267 spin_unlock_irqrestore(&qp->q_lock, flags);
2272 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2273 wr->num_sge > qp->rq.max_sges) {
2278 rqe = ocrdma_hwq_head(&qp->rq);
2279 ocrdma_build_rqe(rqe, wr, 0);
2281 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2282 /* make sure rqe is written before adapter can access it */
2285 /* inform hw to start processing it */
2286 ocrdma_ring_rq_db(qp);
2288 /* update pointer, counter for next wr */
2289 ocrdma_hwq_inc_head(&qp->rq);
2292 spin_unlock_irqrestore(&qp->q_lock, flags);
2296 /* cqe for srq's rqe can potentially arrive out of order.
2297 * index gives the entry in the shadow table where to store
2298 * the wr_id. tag/index is returned in cqe to reference back
2301 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2306 for (row = 0; row < srq->bit_fields_len; row++) {
2307 if (srq->idx_bit_fields[row]) {
2308 indx = ffs(srq->idx_bit_fields[row]);
2309 indx = (row * 32) + (indx - 1);
2310 BUG_ON(indx >= srq->rq.max_cnt);
2311 ocrdma_srq_toggle_bit(srq, indx);
2316 BUG_ON(row == srq->bit_fields_len);
2317 return indx + 1; /* Use from index 1 */
2320 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2322 u32 val = srq->rq.dbid | (1 << 16);
2324 iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2327 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
2328 const struct ib_recv_wr **bad_wr)
2331 unsigned long flags;
2332 struct ocrdma_srq *srq;
2333 struct ocrdma_hdr_wqe *rqe;
2336 srq = get_ocrdma_srq(ibsrq);
2338 spin_lock_irqsave(&srq->q_lock, flags);
2340 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2341 wr->num_sge > srq->rq.max_sges) {
2346 tag = ocrdma_srq_get_idx(srq);
2347 rqe = ocrdma_hwq_head(&srq->rq);
2348 ocrdma_build_rqe(rqe, wr, tag);
2350 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2351 /* make sure rqe is written before adapter can perform DMA */
2353 /* inform hw to start processing it */
2354 ocrdma_ring_srq_db(srq);
2355 /* update pointer, counter for next wr */
2356 ocrdma_hwq_inc_head(&srq->rq);
2359 spin_unlock_irqrestore(&srq->q_lock, flags);
2363 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2365 enum ib_wc_status ibwc_status;
2368 case OCRDMA_CQE_GENERAL_ERR:
2369 ibwc_status = IB_WC_GENERAL_ERR;
2371 case OCRDMA_CQE_LOC_LEN_ERR:
2372 ibwc_status = IB_WC_LOC_LEN_ERR;
2374 case OCRDMA_CQE_LOC_QP_OP_ERR:
2375 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2377 case OCRDMA_CQE_LOC_EEC_OP_ERR:
2378 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2380 case OCRDMA_CQE_LOC_PROT_ERR:
2381 ibwc_status = IB_WC_LOC_PROT_ERR;
2383 case OCRDMA_CQE_WR_FLUSH_ERR:
2384 ibwc_status = IB_WC_WR_FLUSH_ERR;
2386 case OCRDMA_CQE_MW_BIND_ERR:
2387 ibwc_status = IB_WC_MW_BIND_ERR;
2389 case OCRDMA_CQE_BAD_RESP_ERR:
2390 ibwc_status = IB_WC_BAD_RESP_ERR;
2392 case OCRDMA_CQE_LOC_ACCESS_ERR:
2393 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2395 case OCRDMA_CQE_REM_INV_REQ_ERR:
2396 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2398 case OCRDMA_CQE_REM_ACCESS_ERR:
2399 ibwc_status = IB_WC_REM_ACCESS_ERR;
2401 case OCRDMA_CQE_REM_OP_ERR:
2402 ibwc_status = IB_WC_REM_OP_ERR;
2404 case OCRDMA_CQE_RETRY_EXC_ERR:
2405 ibwc_status = IB_WC_RETRY_EXC_ERR;
2407 case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2408 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2410 case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2411 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2413 case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2414 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2416 case OCRDMA_CQE_REM_ABORT_ERR:
2417 ibwc_status = IB_WC_REM_ABORT_ERR;
2419 case OCRDMA_CQE_INV_EECN_ERR:
2420 ibwc_status = IB_WC_INV_EECN_ERR;
2422 case OCRDMA_CQE_INV_EEC_STATE_ERR:
2423 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2425 case OCRDMA_CQE_FATAL_ERR:
2426 ibwc_status = IB_WC_FATAL_ERR;
2428 case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2429 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2432 ibwc_status = IB_WC_GENERAL_ERR;
2438 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2441 struct ocrdma_hdr_wqe *hdr;
2442 struct ocrdma_sge *rw;
2445 hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2447 ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2448 /* Undo the hdr->cw swap */
2449 opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2452 ibwc->opcode = IB_WC_RDMA_WRITE;
2455 rw = (struct ocrdma_sge *)(hdr + 1);
2456 ibwc->opcode = IB_WC_RDMA_READ;
2457 ibwc->byte_len = rw->len;
2460 ibwc->opcode = IB_WC_SEND;
2463 ibwc->opcode = IB_WC_REG_MR;
2465 case OCRDMA_LKEY_INV:
2466 ibwc->opcode = IB_WC_LOCAL_INV;
2469 ibwc->status = IB_WC_GENERAL_ERR;
2470 pr_err("%s() invalid opcode received = 0x%x\n",
2471 __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2476 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2477 struct ocrdma_cqe *cqe)
2479 if (is_cqe_for_sq(cqe)) {
2480 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2481 cqe->flags_status_srcqpn) &
2482 ~OCRDMA_CQE_STATUS_MASK);
2483 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2484 cqe->flags_status_srcqpn) |
2485 (OCRDMA_CQE_WR_FLUSH_ERR <<
2486 OCRDMA_CQE_STATUS_SHIFT));
2488 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2489 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2490 cqe->flags_status_srcqpn) &
2491 ~OCRDMA_CQE_UD_STATUS_MASK);
2492 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2493 cqe->flags_status_srcqpn) |
2494 (OCRDMA_CQE_WR_FLUSH_ERR <<
2495 OCRDMA_CQE_UD_STATUS_SHIFT));
2497 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2498 cqe->flags_status_srcqpn) &
2499 ~OCRDMA_CQE_STATUS_MASK);
2500 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2501 cqe->flags_status_srcqpn) |
2502 (OCRDMA_CQE_WR_FLUSH_ERR <<
2503 OCRDMA_CQE_STATUS_SHIFT));
2508 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2509 struct ocrdma_qp *qp, int status)
2511 bool expand = false;
2514 ibwc->qp = &qp->ibqp;
2515 ibwc->status = ocrdma_to_ibwc_err(status);
2517 ocrdma_flush_qp(qp);
2518 ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2520 /* if wqe/rqe pending for which cqe needs to be returned,
2521 * trigger inflating it.
2523 if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2525 ocrdma_set_cqe_status_flushed(qp, cqe);
2530 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2531 struct ocrdma_qp *qp, int status)
2533 ibwc->opcode = IB_WC_RECV;
2534 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2535 ocrdma_hwq_inc_tail(&qp->rq);
2537 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2540 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2541 struct ocrdma_qp *qp, int status)
2543 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2544 ocrdma_hwq_inc_tail(&qp->sq);
2546 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2550 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2551 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2552 bool *polled, bool *stop)
2555 struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2556 int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2557 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2558 if (status < OCRDMA_MAX_CQE_ERR)
2559 atomic_inc(&dev->cqe_err_stats[status]);
2561 /* when hw sq is empty, but rq is not empty, so we continue
2562 * to keep the cqe in order to get the cq event again.
2564 if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2565 /* when cq for rq and sq is same, it is safe to return
2566 * flush cqe for RQEs.
2568 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2570 status = OCRDMA_CQE_WR_FLUSH_ERR;
2571 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2573 /* stop processing further cqe as this cqe is used for
2574 * triggering cq event on buddy cq of RQ.
2575 * When QP is destroyed, this cqe will be removed
2576 * from the cq's hardware q.
2582 } else if (is_hw_sq_empty(qp)) {
2589 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2594 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2595 struct ocrdma_cqe *cqe,
2596 struct ib_wc *ibwc, bool *polled)
2598 bool expand = false;
2599 int tail = qp->sq.tail;
2602 if (!qp->wqe_wr_id_tbl[tail].signaled) {
2603 *polled = false; /* WC cannot be consumed yet */
2605 ibwc->status = IB_WC_SUCCESS;
2607 ibwc->qp = &qp->ibqp;
2608 ocrdma_update_wc(qp, ibwc, tail);
2611 wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2612 OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2613 if (tail != wqe_idx)
2614 expand = true; /* Coalesced CQE can't be consumed yet */
2616 ocrdma_hwq_inc_tail(&qp->sq);
2620 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2621 struct ib_wc *ibwc, bool *polled, bool *stop)
2626 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2627 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2629 if (status == OCRDMA_CQE_SUCCESS)
2630 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2632 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2636 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2637 struct ocrdma_cqe *cqe)
2642 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2643 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2644 ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2645 OCRDMA_CQE_SRCQP_MASK;
2646 ibwc->pkey_index = 0;
2647 ibwc->wc_flags = IB_WC_GRH;
2648 ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2649 OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2650 OCRDMA_CQE_UD_XFER_LEN_MASK;
2652 if (ocrdma_is_udp_encap_supported(dev)) {
2653 hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2654 OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2655 OCRDMA_CQE_UD_L3TYPE_MASK;
2656 ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2657 ibwc->network_hdr_type = hdr_type;
2663 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2664 struct ocrdma_cqe *cqe,
2665 struct ocrdma_qp *qp)
2667 unsigned long flags;
2668 struct ocrdma_srq *srq;
2671 srq = get_ocrdma_srq(qp->ibqp.srq);
2672 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2673 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2674 BUG_ON(wqe_idx < 1);
2676 ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2677 spin_lock_irqsave(&srq->q_lock, flags);
2678 ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2679 spin_unlock_irqrestore(&srq->q_lock, flags);
2680 ocrdma_hwq_inc_tail(&srq->rq);
2683 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2684 struct ib_wc *ibwc, bool *polled, bool *stop,
2688 struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2690 if (status < OCRDMA_MAX_CQE_ERR)
2691 atomic_inc(&dev->cqe_err_stats[status]);
2693 /* when hw_rq is empty, but wq is not empty, so continue
2694 * to keep the cqe to get the cq event again.
2696 if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2697 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2699 status = OCRDMA_CQE_WR_FLUSH_ERR;
2700 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2706 } else if (is_hw_rq_empty(qp)) {
2713 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2718 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2719 struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2721 struct ocrdma_dev *dev;
2723 dev = get_ocrdma_dev(qp->ibqp.device);
2724 ibwc->opcode = IB_WC_RECV;
2725 ibwc->qp = &qp->ibqp;
2726 ibwc->status = IB_WC_SUCCESS;
2728 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2729 ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2731 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2733 if (is_cqe_imm(cqe)) {
2734 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2735 ibwc->wc_flags |= IB_WC_WITH_IMM;
2736 } else if (is_cqe_wr_imm(cqe)) {
2737 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2738 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2739 ibwc->wc_flags |= IB_WC_WITH_IMM;
2740 } else if (is_cqe_invalidated(cqe)) {
2741 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2742 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2745 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2747 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2748 ocrdma_hwq_inc_tail(&qp->rq);
2752 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2753 struct ib_wc *ibwc, bool *polled, bool *stop)
2756 bool expand = false;
2759 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2760 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2761 OCRDMA_CQE_UD_STATUS_MASK) >>
2762 OCRDMA_CQE_UD_STATUS_SHIFT;
2764 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2765 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2768 if (status == OCRDMA_CQE_SUCCESS) {
2770 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2772 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2778 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2781 if (cq->phase_change) {
2783 cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2785 /* clear valid bit */
2786 cqe->flags_status_srcqpn = 0;
2790 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2795 bool expand = false;
2796 int polled_hw_cqes = 0;
2797 struct ocrdma_qp *qp = NULL;
2798 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2799 struct ocrdma_cqe *cqe;
2800 u16 cur_getp; bool polled = false; bool stop = false;
2802 cur_getp = cq->getp;
2803 while (num_entries) {
2804 cqe = cq->va + cur_getp;
2805 /* check whether valid cqe or not */
2806 if (!is_cqe_valid(cq, cqe))
2808 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2809 /* ignore discarded cqe */
2812 qp = dev->qp_tbl[qpn];
2815 if (is_cqe_for_sq(cqe)) {
2816 expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2819 expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2826 /* clear qpn to avoid duplicate processing by discard_cqe() */
2829 polled_hw_cqes += 1;
2830 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2831 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2841 cq->getp = cur_getp;
2844 ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2849 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2850 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2851 struct ocrdma_qp *qp, struct ib_wc *ibwc)
2855 while (num_entries) {
2856 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2858 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2859 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2860 ocrdma_hwq_inc_tail(&qp->sq);
2861 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2862 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2863 ocrdma_hwq_inc_tail(&qp->rq);
2868 ibwc->status = IB_WC_WR_FLUSH_ERR;
2876 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2878 int cqes_to_poll = num_entries;
2879 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2880 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2881 int num_os_cqe = 0, err_cqes = 0;
2882 struct ocrdma_qp *qp;
2883 unsigned long flags;
2885 /* poll cqes from adapter CQ */
2886 spin_lock_irqsave(&cq->cq_lock, flags);
2887 num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2888 spin_unlock_irqrestore(&cq->cq_lock, flags);
2889 cqes_to_poll -= num_os_cqe;
2892 wc = wc + num_os_cqe;
2893 /* adapter returns single error cqe when qp moves to
2894 * error state. So insert error cqes with wc_status as
2895 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2896 * respectively which uses this CQ.
2898 spin_lock_irqsave(&dev->flush_q_lock, flags);
2899 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2900 if (cqes_to_poll == 0)
2902 err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2903 cqes_to_poll -= err_cqes;
2904 num_os_cqe += err_cqes;
2907 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2912 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2914 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2915 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2917 unsigned long flags;
2918 bool arm_needed = false, sol_needed = false;
2922 spin_lock_irqsave(&cq->cq_lock, flags);
2923 if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2925 if (cq_flags & IB_CQ_SOLICITED)
2928 ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2929 spin_unlock_irqrestore(&cq->cq_lock, flags);
2934 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd,
2935 enum ib_mr_type mr_type,
2939 struct ocrdma_mr *mr;
2940 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2941 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2943 if (mr_type != IB_MR_TYPE_MEM_REG)
2944 return ERR_PTR(-EINVAL);
2946 if (max_num_sg > dev->attr.max_pages_per_frmr)
2947 return ERR_PTR(-EINVAL);
2949 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2951 return ERR_PTR(-ENOMEM);
2953 mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
2959 status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
2963 mr->hwmr.remote_rd = 0;
2964 mr->hwmr.remote_wr = 0;
2965 mr->hwmr.local_rd = 0;
2966 mr->hwmr.local_wr = 0;
2967 mr->hwmr.mw_bind = 0;
2968 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2971 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2974 mr->ibmr.rkey = mr->hwmr.lkey;
2975 mr->ibmr.lkey = mr->hwmr.lkey;
2976 dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
2980 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2985 return ERR_PTR(-ENOMEM);
2988 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
2990 struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
2992 if (unlikely(mr->npages == mr->hwmr.num_pbes))
2995 mr->pages[mr->npages++] = addr;
3000 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
3001 unsigned int *sg_offset)
3003 struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3007 return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);