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 "ocrdma_abi.h"
56 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
65 int ocrdma_query_gid(struct ib_device *ibdev, u8 port,
66 int index, union ib_gid *sgid)
69 struct ocrdma_dev *dev;
71 dev = get_ocrdma_dev(ibdev);
72 memset(sgid, 0, sizeof(*sgid));
73 if (index >= OCRDMA_MAX_SGID)
76 ret = ib_get_cached_gid(ibdev, port, index, sgid, NULL);
78 memcpy(sgid, &zgid, sizeof(*sgid));
85 int ocrdma_add_gid(struct ib_device *device,
88 const union ib_gid *gid,
89 const struct ib_gid_attr *attr,
94 int ocrdma_del_gid(struct ib_device *device,
101 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
102 struct ib_udata *uhw)
104 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
106 if (uhw->inlen || uhw->outlen)
109 memset(attr, 0, sizeof *attr);
110 memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
111 min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
112 ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
113 attr->max_mr_size = dev->attr.max_mr_size;
114 attr->page_size_cap = 0xffff000;
115 attr->vendor_id = dev->nic_info.pdev->vendor;
116 attr->vendor_part_id = dev->nic_info.pdev->device;
117 attr->hw_ver = dev->asic_id;
118 attr->max_qp = dev->attr.max_qp;
119 attr->max_ah = OCRDMA_MAX_AH;
120 attr->max_qp_wr = dev->attr.max_wqe;
122 attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
123 IB_DEVICE_RC_RNR_NAK_GEN |
124 IB_DEVICE_SHUTDOWN_PORT |
125 IB_DEVICE_SYS_IMAGE_GUID |
126 IB_DEVICE_LOCAL_DMA_LKEY |
127 IB_DEVICE_MEM_MGT_EXTENSIONS;
128 attr->max_sge = min(dev->attr.max_send_sge, dev->attr.max_srq_sge);
129 attr->max_sge_rd = 0;
130 attr->max_cq = dev->attr.max_cq;
131 attr->max_cqe = dev->attr.max_cqe;
132 attr->max_mr = dev->attr.max_mr;
133 attr->max_mw = dev->attr.max_mw;
134 attr->max_pd = dev->attr.max_pd;
135 attr->atomic_cap = 0;
137 attr->max_map_per_fmr = 0;
138 attr->max_qp_rd_atom =
139 min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
140 attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
141 attr->max_srq = dev->attr.max_srq;
142 attr->max_srq_sge = dev->attr.max_srq_sge;
143 attr->max_srq_wr = dev->attr.max_rqe;
144 attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
145 attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
150 struct net_device *ocrdma_get_netdev(struct ib_device *ibdev, u8 port_num)
152 struct ocrdma_dev *dev;
153 struct net_device *ndev = NULL;
157 dev = get_ocrdma_dev(ibdev);
159 ndev = dev->nic_info.netdev;
168 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
169 u8 *ib_speed, u8 *ib_width)
174 status = ocrdma_mbx_get_link_speed(dev, &speed);
176 speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
179 case OCRDMA_PHYS_LINK_SPEED_1GBPS:
180 *ib_speed = IB_SPEED_SDR;
181 *ib_width = IB_WIDTH_1X;
184 case OCRDMA_PHYS_LINK_SPEED_10GBPS:
185 *ib_speed = IB_SPEED_QDR;
186 *ib_width = IB_WIDTH_1X;
189 case OCRDMA_PHYS_LINK_SPEED_20GBPS:
190 *ib_speed = IB_SPEED_DDR;
191 *ib_width = IB_WIDTH_4X;
194 case OCRDMA_PHYS_LINK_SPEED_40GBPS:
195 *ib_speed = IB_SPEED_QDR;
196 *ib_width = IB_WIDTH_4X;
201 *ib_speed = IB_SPEED_SDR;
202 *ib_width = IB_WIDTH_1X;
206 int ocrdma_query_port(struct ib_device *ibdev,
207 u8 port, struct ib_port_attr *props)
209 enum ib_port_state port_state;
210 struct ocrdma_dev *dev;
211 struct net_device *netdev;
213 dev = get_ocrdma_dev(ibdev);
215 pr_err("%s(%d) invalid_port=0x%x\n", __func__,
219 netdev = dev->nic_info.netdev;
220 if (netif_running(netdev) && netif_oper_up(netdev)) {
221 port_state = IB_PORT_ACTIVE;
222 props->phys_state = 5;
224 port_state = IB_PORT_DOWN;
225 props->phys_state = 3;
227 props->max_mtu = IB_MTU_4096;
228 props->active_mtu = iboe_get_mtu(netdev->mtu);
233 props->state = port_state;
234 props->port_cap_flags =
237 IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP |
238 IB_PORT_IP_BASED_GIDS;
239 props->gid_tbl_len = OCRDMA_MAX_SGID;
240 props->pkey_tbl_len = 1;
241 props->bad_pkey_cntr = 0;
242 props->qkey_viol_cntr = 0;
243 get_link_speed_and_width(dev, &props->active_speed,
244 &props->active_width);
245 props->max_msg_sz = 0x80000000;
246 props->max_vl_num = 4;
250 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
251 struct ib_port_modify *props)
253 struct ocrdma_dev *dev;
255 dev = get_ocrdma_dev(ibdev);
257 pr_err("%s(%d) invalid_port=0x%x\n", __func__, dev->id, port);
263 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
266 struct ocrdma_mm *mm;
268 mm = kzalloc(sizeof(*mm), GFP_KERNEL);
271 mm->key.phy_addr = phy_addr;
273 INIT_LIST_HEAD(&mm->entry);
275 mutex_lock(&uctx->mm_list_lock);
276 list_add_tail(&mm->entry, &uctx->mm_head);
277 mutex_unlock(&uctx->mm_list_lock);
281 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
284 struct ocrdma_mm *mm, *tmp;
286 mutex_lock(&uctx->mm_list_lock);
287 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
288 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
291 list_del(&mm->entry);
295 mutex_unlock(&uctx->mm_list_lock);
298 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
302 struct ocrdma_mm *mm;
304 mutex_lock(&uctx->mm_list_lock);
305 list_for_each_entry(mm, &uctx->mm_head, entry) {
306 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
312 mutex_unlock(&uctx->mm_list_lock);
317 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
319 u16 pd_bitmap_idx = 0;
320 const unsigned long *pd_bitmap;
323 pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
324 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
325 dev->pd_mgr->max_dpp_pd);
326 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
327 dev->pd_mgr->pd_dpp_count++;
328 if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
329 dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
331 pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
332 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
333 dev->pd_mgr->max_normal_pd);
334 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
335 dev->pd_mgr->pd_norm_count++;
336 if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
337 dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
339 return pd_bitmap_idx;
342 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
348 pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
349 dev->pd_mgr->pd_norm_count;
354 pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
355 if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
358 __clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
359 dev->pd_mgr->pd_dpp_count--;
362 pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
363 if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
366 __clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
367 dev->pd_mgr->pd_norm_count--;
374 static u8 ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
379 mutex_lock(&dev->dev_lock);
380 status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
381 mutex_unlock(&dev->dev_lock);
385 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
390 mutex_lock(&dev->dev_lock);
391 if (pd->dpp_enabled) {
392 /* try allocating DPP PD, if not available then normal PD */
393 if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
394 pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
395 pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
396 pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
397 } else if (dev->pd_mgr->pd_norm_count <
398 dev->pd_mgr->max_normal_pd) {
399 pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
400 pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
401 pd->dpp_enabled = false;
406 if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
407 pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
408 pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
413 mutex_unlock(&dev->dev_lock);
417 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
418 struct ocrdma_ucontext *uctx,
419 struct ib_udata *udata)
421 struct ocrdma_pd *pd = NULL;
424 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
426 return ERR_PTR(-ENOMEM);
428 if (udata && uctx && dev->attr.max_dpp_pds) {
430 ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
432 pd->dpp_enabled ? (dev->nic_info.db_page_size /
433 dev->attr.wqe_size) : 0;
436 if (dev->pd_mgr->pd_prealloc_valid) {
437 status = ocrdma_get_pd_num(dev, pd);
442 return ERR_PTR(status);
447 status = ocrdma_mbx_alloc_pd(dev, pd);
449 if (pd->dpp_enabled) {
450 pd->dpp_enabled = false;
455 return ERR_PTR(status);
462 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
463 struct ocrdma_pd *pd)
465 return (uctx->cntxt_pd == pd ? true : false);
468 static int _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
469 struct ocrdma_pd *pd)
473 if (dev->pd_mgr->pd_prealloc_valid)
474 status = ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
476 status = ocrdma_mbx_dealloc_pd(dev, pd);
482 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
483 struct ocrdma_ucontext *uctx,
484 struct ib_udata *udata)
488 uctx->cntxt_pd = _ocrdma_alloc_pd(dev, uctx, udata);
489 if (IS_ERR(uctx->cntxt_pd)) {
490 status = PTR_ERR(uctx->cntxt_pd);
491 uctx->cntxt_pd = NULL;
495 uctx->cntxt_pd->uctx = uctx;
496 uctx->cntxt_pd->ibpd.device = &dev->ibdev;
501 static int ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
503 struct ocrdma_pd *pd = uctx->cntxt_pd;
504 struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
506 if (uctx->pd_in_use) {
507 pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
508 __func__, dev->id, pd->id);
510 uctx->cntxt_pd = NULL;
511 (void)_ocrdma_dealloc_pd(dev, pd);
515 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
517 struct ocrdma_pd *pd = NULL;
519 mutex_lock(&uctx->mm_list_lock);
520 if (!uctx->pd_in_use) {
521 uctx->pd_in_use = true;
524 mutex_unlock(&uctx->mm_list_lock);
529 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
531 mutex_lock(&uctx->mm_list_lock);
532 uctx->pd_in_use = false;
533 mutex_unlock(&uctx->mm_list_lock);
536 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
537 struct ib_udata *udata)
540 struct ocrdma_ucontext *ctx;
541 struct ocrdma_alloc_ucontext_resp resp;
542 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
543 struct pci_dev *pdev = dev->nic_info.pdev;
544 u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
547 return ERR_PTR(-EFAULT);
548 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
550 return ERR_PTR(-ENOMEM);
551 INIT_LIST_HEAD(&ctx->mm_head);
552 mutex_init(&ctx->mm_list_lock);
554 ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
555 &ctx->ah_tbl.pa, GFP_KERNEL);
556 if (!ctx->ah_tbl.va) {
558 return ERR_PTR(-ENOMEM);
560 memset(ctx->ah_tbl.va, 0, map_len);
561 ctx->ah_tbl.len = map_len;
563 memset(&resp, 0, sizeof(resp));
564 resp.ah_tbl_len = ctx->ah_tbl.len;
565 resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
567 status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
571 status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
575 resp.dev_id = dev->id;
576 resp.max_inline_data = dev->attr.max_inline_data;
577 resp.wqe_size = dev->attr.wqe_size;
578 resp.rqe_size = dev->attr.rqe_size;
579 resp.dpp_wqe_size = dev->attr.wqe_size;
581 memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
582 status = ib_copy_to_udata(udata, &resp, sizeof(resp));
585 return &ctx->ibucontext;
589 ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
591 dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
594 return ERR_PTR(status);
597 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
600 struct ocrdma_mm *mm, *tmp;
601 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
602 struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
603 struct pci_dev *pdev = dev->nic_info.pdev;
605 status = ocrdma_dealloc_ucontext_pd(uctx);
607 ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
608 dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
611 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
612 list_del(&mm->entry);
619 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
621 struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
622 struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
623 unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
624 u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
625 unsigned long len = (vma->vm_end - vma->vm_start);
629 if (vma->vm_start & (PAGE_SIZE - 1))
631 found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
635 if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
636 dev->nic_info.db_total_size)) &&
637 (len <= dev->nic_info.db_page_size)) {
638 if (vma->vm_flags & VM_READ)
641 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
642 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
643 len, vma->vm_page_prot);
644 } else if (dev->nic_info.dpp_unmapped_len &&
645 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
646 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
647 dev->nic_info.dpp_unmapped_len)) &&
648 (len <= dev->nic_info.dpp_unmapped_len)) {
649 if (vma->vm_flags & VM_READ)
652 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
653 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
654 len, vma->vm_page_prot);
656 status = remap_pfn_range(vma, vma->vm_start,
657 vma->vm_pgoff, len, vma->vm_page_prot);
662 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
663 struct ib_ucontext *ib_ctx,
664 struct ib_udata *udata)
668 u64 dpp_page_addr = 0;
670 struct ocrdma_alloc_pd_uresp rsp;
671 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
673 memset(&rsp, 0, sizeof(rsp));
675 rsp.dpp_enabled = pd->dpp_enabled;
676 db_page_addr = ocrdma_get_db_addr(dev, pd->id);
677 db_page_size = dev->nic_info.db_page_size;
679 status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
683 if (pd->dpp_enabled) {
684 dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
685 (pd->id * PAGE_SIZE);
686 status = ocrdma_add_mmap(uctx, dpp_page_addr,
690 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
691 rsp.dpp_page_addr_lo = dpp_page_addr;
694 status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
703 ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
705 ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
709 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
710 struct ib_ucontext *context,
711 struct ib_udata *udata)
713 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
714 struct ocrdma_pd *pd;
715 struct ocrdma_ucontext *uctx = NULL;
717 u8 is_uctx_pd = false;
719 if (udata && context) {
720 uctx = get_ocrdma_ucontext(context);
721 pd = ocrdma_get_ucontext_pd(uctx);
728 pd = _ocrdma_alloc_pd(dev, uctx, udata);
730 status = PTR_ERR(pd);
735 if (udata && context) {
736 status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
744 ocrdma_release_ucontext_pd(uctx);
746 status = _ocrdma_dealloc_pd(dev, pd);
749 return ERR_PTR(status);
752 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
754 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
755 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
756 struct ocrdma_ucontext *uctx = NULL;
762 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
763 (pd->id * PAGE_SIZE);
765 ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
766 usr_db = ocrdma_get_db_addr(dev, pd->id);
767 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
769 if (is_ucontext_pd(uctx, pd)) {
770 ocrdma_release_ucontext_pd(uctx);
774 status = _ocrdma_dealloc_pd(dev, pd);
778 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
779 u32 pdid, int acc, u32 num_pbls, u32 addr_check)
784 mr->hwmr.local_rd = 1;
785 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
786 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
787 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
788 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
789 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
790 mr->hwmr.num_pbls = num_pbls;
792 status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
796 mr->ibmr.lkey = mr->hwmr.lkey;
797 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
798 mr->ibmr.rkey = mr->hwmr.lkey;
802 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
805 struct ocrdma_mr *mr;
806 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
807 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
809 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
810 pr_err("%s err, invalid access rights\n", __func__);
811 return ERR_PTR(-EINVAL);
814 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
816 return ERR_PTR(-ENOMEM);
818 status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
819 OCRDMA_ADDR_CHECK_DISABLE);
822 return ERR_PTR(status);
828 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
829 struct ocrdma_hw_mr *mr)
831 struct pci_dev *pdev = dev->nic_info.pdev;
835 for (i = 0; i < mr->num_pbls; i++) {
836 if (!mr->pbl_table[i].va)
838 dma_free_coherent(&pdev->dev, mr->pbl_size,
840 mr->pbl_table[i].pa);
842 kfree(mr->pbl_table);
843 mr->pbl_table = NULL;
847 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
856 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
857 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
861 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
862 num_pbls = num_pbls / (pbl_size / sizeof(u64));
864 } while (num_pbls >= dev->attr.max_num_mr_pbl);
866 mr->hwmr.num_pbes = num_pbes;
867 mr->hwmr.num_pbls = num_pbls;
868 mr->hwmr.pbl_size = pbl_size;
872 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
876 u32 dma_len = mr->pbl_size;
877 struct pci_dev *pdev = dev->nic_info.pdev;
881 mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
882 mr->num_pbls, GFP_KERNEL);
887 for (i = 0; i < mr->num_pbls; i++) {
888 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
890 ocrdma_free_mr_pbl_tbl(dev, mr);
894 memset(va, 0, dma_len);
895 mr->pbl_table[i].va = va;
896 mr->pbl_table[i].pa = pa;
901 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
904 struct ocrdma_pbe *pbe;
905 struct scatterlist *sg;
906 struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
907 struct ib_umem *umem = mr->umem;
908 int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
910 if (!mr->hwmr.num_pbes)
913 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
916 shift = ilog2(umem->page_size);
918 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
919 pages = sg_dma_len(sg) >> shift;
920 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
921 /* store the page address in pbe */
923 cpu_to_le32(sg_dma_address
925 (umem->page_size * pg_cnt));
927 cpu_to_le32(upper_32_bits
930 umem->page_size * pg_cnt)));
935 /* if done building pbes, issue the mbx cmd. */
936 if (total_num_pbes == num_pbes)
939 /* if the given pbl is full storing the pbes,
943 (mr->hwmr.pbl_size / sizeof(u64))) {
945 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
953 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
954 u64 usr_addr, int acc, struct ib_udata *udata)
956 int status = -ENOMEM;
957 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
958 struct ocrdma_mr *mr;
959 struct ocrdma_pd *pd;
962 pd = get_ocrdma_pd(ibpd);
964 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
965 return ERR_PTR(-EINVAL);
967 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
969 return ERR_PTR(status);
970 mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
971 if (IS_ERR(mr->umem)) {
975 num_pbes = ib_umem_page_count(mr->umem);
976 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
980 mr->hwmr.pbe_size = mr->umem->page_size;
981 mr->hwmr.fbo = ib_umem_offset(mr->umem);
982 mr->hwmr.va = usr_addr;
984 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
985 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
986 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
987 mr->hwmr.local_rd = 1;
988 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
989 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
992 build_user_pbes(dev, mr, num_pbes);
993 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
996 mr->ibmr.lkey = mr->hwmr.lkey;
997 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
998 mr->ibmr.rkey = mr->hwmr.lkey;
1003 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1006 return ERR_PTR(status);
1009 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
1011 struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
1012 struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
1014 (void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
1016 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1018 /* it could be user registered memory. */
1020 ib_umem_release(mr->umem);
1023 /* Don't stop cleanup, in case FW is unresponsive */
1024 if (dev->mqe_ctx.fw_error_state) {
1025 pr_err("%s(%d) fw not responding.\n",
1031 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
1032 struct ib_udata *udata,
1033 struct ib_ucontext *ib_ctx)
1036 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
1037 struct ocrdma_create_cq_uresp uresp;
1039 memset(&uresp, 0, sizeof(uresp));
1040 uresp.cq_id = cq->id;
1041 uresp.page_size = PAGE_ALIGN(cq->len);
1042 uresp.num_pages = 1;
1043 uresp.max_hw_cqe = cq->max_hw_cqe;
1044 uresp.page_addr[0] = virt_to_phys(cq->va);
1045 uresp.db_page_addr = ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
1046 uresp.db_page_size = dev->nic_info.db_page_size;
1047 uresp.phase_change = cq->phase_change ? 1 : 0;
1048 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1050 pr_err("%s(%d) copy error cqid=0x%x.\n",
1051 __func__, dev->id, cq->id);
1054 status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1057 status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
1059 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1062 cq->ucontext = uctx;
1067 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev,
1068 const struct ib_cq_init_attr *attr,
1069 struct ib_ucontext *ib_ctx,
1070 struct ib_udata *udata)
1072 int entries = attr->cqe;
1073 struct ocrdma_cq *cq;
1074 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
1075 struct ocrdma_ucontext *uctx = NULL;
1078 struct ocrdma_create_cq_ureq ureq;
1081 return ERR_PTR(-EINVAL);
1084 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1085 return ERR_PTR(-EFAULT);
1088 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
1090 return ERR_PTR(-ENOMEM);
1092 spin_lock_init(&cq->cq_lock);
1093 spin_lock_init(&cq->comp_handler_lock);
1094 INIT_LIST_HEAD(&cq->sq_head);
1095 INIT_LIST_HEAD(&cq->rq_head);
1096 cq->first_arm = true;
1099 uctx = get_ocrdma_ucontext(ib_ctx);
1100 pd_id = uctx->cntxt_pd->id;
1103 status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1106 return ERR_PTR(status);
1109 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
1113 cq->phase = OCRDMA_CQE_VALID;
1114 dev->cq_tbl[cq->id] = cq;
1118 ocrdma_mbx_destroy_cq(dev, cq);
1120 return ERR_PTR(status);
1123 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1124 struct ib_udata *udata)
1127 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1129 if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1133 ibcq->cqe = new_cnt;
1137 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1140 int valid_count = 0;
1141 unsigned long flags;
1143 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1144 struct ocrdma_cqe *cqe = NULL;
1147 cqe_cnt = cq->cqe_cnt;
1149 /* Last irq might have scheduled a polling thread
1150 * sync-up with it before hard flushing.
1152 spin_lock_irqsave(&cq->cq_lock, flags);
1154 if (is_cqe_valid(cq, cqe))
1159 ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1160 spin_unlock_irqrestore(&cq->cq_lock, flags);
1163 int ocrdma_destroy_cq(struct ib_cq *ibcq)
1165 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1166 struct ocrdma_eq *eq = NULL;
1167 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1171 dev->cq_tbl[cq->id] = NULL;
1172 indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1173 if (indx == -EINVAL)
1176 eq = &dev->eq_tbl[indx];
1177 irq = ocrdma_get_irq(dev, eq);
1178 synchronize_irq(irq);
1179 ocrdma_flush_cq(cq);
1181 (void)ocrdma_mbx_destroy_cq(dev, cq);
1183 pdid = cq->ucontext->cntxt_pd->id;
1184 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1185 PAGE_ALIGN(cq->len));
1186 ocrdma_del_mmap(cq->ucontext,
1187 ocrdma_get_db_addr(dev, pdid),
1188 dev->nic_info.db_page_size);
1195 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1197 int status = -EINVAL;
1199 if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1200 dev->qp_tbl[qp->id] = qp;
1206 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1208 dev->qp_tbl[qp->id] = NULL;
1211 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1212 struct ib_qp_init_attr *attrs)
1214 if ((attrs->qp_type != IB_QPT_GSI) &&
1215 (attrs->qp_type != IB_QPT_RC) &&
1216 (attrs->qp_type != IB_QPT_UC) &&
1217 (attrs->qp_type != IB_QPT_UD)) {
1218 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1219 __func__, dev->id, attrs->qp_type);
1222 /* Skip the check for QP1 to support CM size of 128 */
1223 if ((attrs->qp_type != IB_QPT_GSI) &&
1224 (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1225 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1226 __func__, dev->id, attrs->cap.max_send_wr);
1227 pr_err("%s(%d) supported send_wr=0x%x\n",
1228 __func__, dev->id, dev->attr.max_wqe);
1231 if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1232 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1233 __func__, dev->id, attrs->cap.max_recv_wr);
1234 pr_err("%s(%d) supported recv_wr=0x%x\n",
1235 __func__, dev->id, dev->attr.max_rqe);
1238 if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1239 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1240 __func__, dev->id, attrs->cap.max_inline_data);
1241 pr_err("%s(%d) supported inline data size=0x%x\n",
1242 __func__, dev->id, dev->attr.max_inline_data);
1245 if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1246 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1247 __func__, dev->id, attrs->cap.max_send_sge);
1248 pr_err("%s(%d) supported send_sge=0x%x\n",
1249 __func__, dev->id, dev->attr.max_send_sge);
1252 if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1253 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1254 __func__, dev->id, attrs->cap.max_recv_sge);
1255 pr_err("%s(%d) supported recv_sge=0x%x\n",
1256 __func__, dev->id, dev->attr.max_recv_sge);
1259 /* unprivileged user space cannot create special QP */
1260 if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1262 ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1263 __func__, dev->id, attrs->qp_type);
1266 /* allow creating only one GSI type of QP */
1267 if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1268 pr_err("%s(%d) GSI special QPs already created.\n",
1272 /* verify consumer QPs are not trying to use GSI QP's CQ */
1273 if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1274 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1275 (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1276 pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1284 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1285 struct ib_udata *udata, int dpp_offset,
1286 int dpp_credit_lmt, int srq)
1290 struct ocrdma_create_qp_uresp uresp;
1291 struct ocrdma_pd *pd = qp->pd;
1292 struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1294 memset(&uresp, 0, sizeof(uresp));
1295 usr_db = dev->nic_info.unmapped_db +
1296 (pd->id * dev->nic_info.db_page_size);
1297 uresp.qp_id = qp->id;
1298 uresp.sq_dbid = qp->sq.dbid;
1299 uresp.num_sq_pages = 1;
1300 uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1301 uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1302 uresp.num_wqe_allocated = qp->sq.max_cnt;
1304 uresp.rq_dbid = qp->rq.dbid;
1305 uresp.num_rq_pages = 1;
1306 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1307 uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1308 uresp.num_rqe_allocated = qp->rq.max_cnt;
1310 uresp.db_page_addr = usr_db;
1311 uresp.db_page_size = dev->nic_info.db_page_size;
1312 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1313 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1314 uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1316 if (qp->dpp_enabled) {
1317 uresp.dpp_credit = dpp_credit_lmt;
1318 uresp.dpp_offset = dpp_offset;
1320 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1322 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1325 status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1326 uresp.sq_page_size);
1331 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1332 uresp.rq_page_size);
1338 ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1343 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1344 struct ocrdma_pd *pd)
1346 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1347 qp->sq_db = dev->nic_info.db +
1348 (pd->id * dev->nic_info.db_page_size) +
1349 OCRDMA_DB_GEN2_SQ_OFFSET;
1350 qp->rq_db = dev->nic_info.db +
1351 (pd->id * dev->nic_info.db_page_size) +
1352 OCRDMA_DB_GEN2_RQ_OFFSET;
1354 qp->sq_db = dev->nic_info.db +
1355 (pd->id * dev->nic_info.db_page_size) +
1356 OCRDMA_DB_SQ_OFFSET;
1357 qp->rq_db = dev->nic_info.db +
1358 (pd->id * dev->nic_info.db_page_size) +
1359 OCRDMA_DB_RQ_OFFSET;
1363 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1366 kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1368 if (qp->wqe_wr_id_tbl == NULL)
1371 kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1372 if (qp->rqe_wr_id_tbl == NULL)
1378 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1379 struct ocrdma_pd *pd,
1380 struct ib_qp_init_attr *attrs)
1383 spin_lock_init(&qp->q_lock);
1384 INIT_LIST_HEAD(&qp->sq_entry);
1385 INIT_LIST_HEAD(&qp->rq_entry);
1387 qp->qp_type = attrs->qp_type;
1388 qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1389 qp->max_inline_data = attrs->cap.max_inline_data;
1390 qp->sq.max_sges = attrs->cap.max_send_sge;
1391 qp->rq.max_sges = attrs->cap.max_recv_sge;
1392 qp->state = OCRDMA_QPS_RST;
1393 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1396 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1397 struct ib_qp_init_attr *attrs)
1399 if (attrs->qp_type == IB_QPT_GSI) {
1400 dev->gsi_qp_created = 1;
1401 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1402 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1406 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1407 struct ib_qp_init_attr *attrs,
1408 struct ib_udata *udata)
1411 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1412 struct ocrdma_qp *qp;
1413 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1414 struct ocrdma_create_qp_ureq ureq;
1415 u16 dpp_credit_lmt, dpp_offset;
1417 status = ocrdma_check_qp_params(ibpd, dev, attrs);
1421 memset(&ureq, 0, sizeof(ureq));
1423 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1424 return ERR_PTR(-EFAULT);
1426 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1431 ocrdma_set_qp_init_params(qp, pd, attrs);
1433 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1434 OCRDMA_QP_FAST_REG);
1436 mutex_lock(&dev->dev_lock);
1437 status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1439 &dpp_offset, &dpp_credit_lmt);
1443 /* user space QP's wr_id table are managed in library */
1444 if (udata == NULL) {
1445 status = ocrdma_alloc_wr_id_tbl(qp);
1450 status = ocrdma_add_qpn_map(dev, qp);
1453 ocrdma_set_qp_db(dev, qp, pd);
1455 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1457 (attrs->srq != NULL));
1461 ocrdma_store_gsi_qp_cq(dev, attrs);
1462 qp->ibqp.qp_num = qp->id;
1463 mutex_unlock(&dev->dev_lock);
1467 ocrdma_del_qpn_map(dev, qp);
1469 ocrdma_mbx_destroy_qp(dev, qp);
1471 mutex_unlock(&dev->dev_lock);
1472 kfree(qp->wqe_wr_id_tbl);
1473 kfree(qp->rqe_wr_id_tbl);
1475 pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1477 return ERR_PTR(status);
1480 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1484 struct ocrdma_qp *qp;
1485 struct ocrdma_dev *dev;
1486 enum ib_qp_state old_qps;
1488 qp = get_ocrdma_qp(ibqp);
1489 dev = get_ocrdma_dev(ibqp->device);
1490 if (attr_mask & IB_QP_STATE)
1491 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1492 /* if new and previous states are same hw doesn't need to
1497 status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1502 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1503 int attr_mask, struct ib_udata *udata)
1505 unsigned long flags;
1506 int status = -EINVAL;
1507 struct ocrdma_qp *qp;
1508 struct ocrdma_dev *dev;
1509 enum ib_qp_state old_qps, new_qps;
1511 qp = get_ocrdma_qp(ibqp);
1512 dev = get_ocrdma_dev(ibqp->device);
1514 /* syncronize with multiple context trying to change, retrive qps */
1515 mutex_lock(&dev->dev_lock);
1516 /* syncronize with wqe, rqe posting and cqe processing contexts */
1517 spin_lock_irqsave(&qp->q_lock, flags);
1518 old_qps = get_ibqp_state(qp->state);
1519 if (attr_mask & IB_QP_STATE)
1520 new_qps = attr->qp_state;
1523 spin_unlock_irqrestore(&qp->q_lock, flags);
1525 if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1526 IB_LINK_LAYER_ETHERNET)) {
1527 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1528 "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1529 __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1534 status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1538 mutex_unlock(&dev->dev_lock);
1542 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1560 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1562 int ib_qp_acc_flags = 0;
1564 if (qp_cap_flags & OCRDMA_QP_INB_WR)
1565 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1566 if (qp_cap_flags & OCRDMA_QP_INB_RD)
1567 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1568 return ib_qp_acc_flags;
1571 int ocrdma_query_qp(struct ib_qp *ibqp,
1572 struct ib_qp_attr *qp_attr,
1573 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1577 struct ocrdma_qp_params params;
1578 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1579 struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1581 memset(¶ms, 0, sizeof(params));
1582 mutex_lock(&dev->dev_lock);
1583 status = ocrdma_mbx_query_qp(dev, qp, ¶ms);
1584 mutex_unlock(&dev->dev_lock);
1587 if (qp->qp_type == IB_QPT_UD)
1588 qp_attr->qkey = params.qkey;
1590 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1591 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1592 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1593 qp_attr->path_mig_state = IB_MIG_MIGRATED;
1594 qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1595 qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1596 qp_attr->dest_qp_num =
1597 params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1599 qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1600 qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1601 qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1602 qp_attr->cap.max_send_sge = qp->sq.max_sges;
1603 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1604 qp_attr->cap.max_inline_data = qp->max_inline_data;
1605 qp_init_attr->cap = qp_attr->cap;
1606 memcpy(&qp_attr->ah_attr.grh.dgid, ¶ms.dgid[0],
1607 sizeof(params.dgid));
1608 qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1609 OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1610 qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1611 qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1612 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1613 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1614 qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1615 OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1616 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1618 qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1619 qp_attr->ah_attr.port_num = 1;
1620 qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1621 OCRDMA_QP_PARAMS_SL_MASK) >>
1622 OCRDMA_QP_PARAMS_SL_SHIFT;
1623 qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1624 OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1625 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1626 qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1627 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1628 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1629 qp_attr->retry_cnt =
1630 (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1631 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1632 qp_attr->min_rnr_timer = 0;
1633 qp_attr->pkey_index = 0;
1634 qp_attr->port_num = 1;
1635 qp_attr->ah_attr.src_path_bits = 0;
1636 qp_attr->ah_attr.static_rate = 0;
1637 qp_attr->alt_pkey_index = 0;
1638 qp_attr->alt_port_num = 0;
1639 qp_attr->alt_timeout = 0;
1640 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1641 qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1642 OCRDMA_QP_PARAMS_STATE_SHIFT;
1643 qp_attr->qp_state = get_ibqp_state(qp_state);
1644 qp_attr->cur_qp_state = qp_attr->qp_state;
1645 qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1646 qp_attr->max_dest_rd_atomic =
1647 params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1648 qp_attr->max_rd_atomic =
1649 params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1650 qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1651 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1652 /* Sync driver QP state with FW */
1653 ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1658 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1660 unsigned int i = idx / 32;
1661 u32 mask = (1U << (idx % 32));
1663 srq->idx_bit_fields[i] ^= mask;
1666 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1668 return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1671 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1673 return (qp->sq.tail == qp->sq.head);
1676 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1678 return (qp->rq.tail == qp->rq.head);
1681 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1683 return q->va + (q->head * q->entry_size);
1686 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1689 return q->va + (idx * q->entry_size);
1692 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1694 q->head = (q->head + 1) & q->max_wqe_idx;
1697 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1699 q->tail = (q->tail + 1) & q->max_wqe_idx;
1702 /* discard the cqe for a given QP */
1703 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1705 unsigned long cq_flags;
1706 unsigned long flags;
1707 int discard_cnt = 0;
1708 u32 cur_getp, stop_getp;
1709 struct ocrdma_cqe *cqe;
1710 u32 qpn = 0, wqe_idx = 0;
1712 spin_lock_irqsave(&cq->cq_lock, cq_flags);
1714 /* traverse through the CQEs in the hw CQ,
1715 * find the matching CQE for a given qp,
1716 * mark the matching one discarded by clearing qpn.
1717 * ring the doorbell in the poll_cq() as
1718 * we don't complete out of order cqe.
1721 cur_getp = cq->getp;
1722 /* find upto when do we reap the cq. */
1723 stop_getp = cur_getp;
1725 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1728 cqe = cq->va + cur_getp;
1729 /* if (a) done reaping whole hw cq, or
1730 * (b) qp_xq becomes empty.
1733 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1734 /* if previously discarded cqe found, skip that too. */
1735 /* check for matching qp */
1736 if (qpn == 0 || qpn != qp->id)
1739 if (is_cqe_for_sq(cqe)) {
1740 ocrdma_hwq_inc_tail(&qp->sq);
1743 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1744 OCRDMA_CQE_BUFTAG_SHIFT) &
1745 qp->srq->rq.max_wqe_idx;
1748 spin_lock_irqsave(&qp->srq->q_lock, flags);
1749 ocrdma_hwq_inc_tail(&qp->srq->rq);
1750 ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1751 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1754 ocrdma_hwq_inc_tail(&qp->rq);
1757 /* mark cqe discarded so that it is not picked up later
1763 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1764 } while (cur_getp != stop_getp);
1765 spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1768 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1771 unsigned long flags;
1772 struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1773 /* sync with any active CQ poll */
1775 spin_lock_irqsave(&dev->flush_q_lock, flags);
1776 found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1778 list_del(&qp->sq_entry);
1780 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1782 list_del(&qp->rq_entry);
1784 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1787 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1789 struct ocrdma_pd *pd;
1790 struct ocrdma_qp *qp;
1791 struct ocrdma_dev *dev;
1792 struct ib_qp_attr attrs;
1794 unsigned long flags;
1796 qp = get_ocrdma_qp(ibqp);
1797 dev = get_ocrdma_dev(ibqp->device);
1801 /* change the QP state to ERROR */
1802 if (qp->state != OCRDMA_QPS_RST) {
1803 attrs.qp_state = IB_QPS_ERR;
1804 attr_mask = IB_QP_STATE;
1805 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1807 /* ensure that CQEs for newly created QP (whose id may be same with
1808 * one which just getting destroyed are same), dont get
1809 * discarded until the old CQEs are discarded.
1811 mutex_lock(&dev->dev_lock);
1812 (void) ocrdma_mbx_destroy_qp(dev, qp);
1815 * acquire CQ lock while destroy is in progress, in order to
1816 * protect against proessing in-flight CQEs for this QP.
1818 spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1819 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1820 spin_lock(&qp->rq_cq->cq_lock);
1822 ocrdma_del_qpn_map(dev, qp);
1824 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1825 spin_unlock(&qp->rq_cq->cq_lock);
1826 spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1829 ocrdma_discard_cqes(qp, qp->sq_cq);
1830 ocrdma_discard_cqes(qp, qp->rq_cq);
1832 mutex_unlock(&dev->dev_lock);
1835 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1836 PAGE_ALIGN(qp->sq.len));
1838 ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1839 PAGE_ALIGN(qp->rq.len));
1842 ocrdma_del_flush_qp(qp);
1844 kfree(qp->wqe_wr_id_tbl);
1845 kfree(qp->rqe_wr_id_tbl);
1850 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1851 struct ib_udata *udata)
1854 struct ocrdma_create_srq_uresp uresp;
1856 memset(&uresp, 0, sizeof(uresp));
1857 uresp.rq_dbid = srq->rq.dbid;
1858 uresp.num_rq_pages = 1;
1859 uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1860 uresp.rq_page_size = srq->rq.len;
1861 uresp.db_page_addr = dev->nic_info.unmapped_db +
1862 (srq->pd->id * dev->nic_info.db_page_size);
1863 uresp.db_page_size = dev->nic_info.db_page_size;
1864 uresp.num_rqe_allocated = srq->rq.max_cnt;
1865 if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1866 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1867 uresp.db_shift = 24;
1869 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1870 uresp.db_shift = 16;
1873 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1876 status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1877 uresp.rq_page_size);
1883 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1884 struct ib_srq_init_attr *init_attr,
1885 struct ib_udata *udata)
1887 int status = -ENOMEM;
1888 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1889 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1890 struct ocrdma_srq *srq;
1892 if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1893 return ERR_PTR(-EINVAL);
1894 if (init_attr->attr.max_wr > dev->attr.max_rqe)
1895 return ERR_PTR(-EINVAL);
1897 srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1899 return ERR_PTR(status);
1901 spin_lock_init(&srq->q_lock);
1903 srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1904 status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1908 if (udata == NULL) {
1909 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1911 if (srq->rqe_wr_id_tbl == NULL)
1914 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1915 (srq->rq.max_cnt % 32 ? 1 : 0);
1916 srq->idx_bit_fields =
1917 kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1918 if (srq->idx_bit_fields == NULL)
1920 memset(srq->idx_bit_fields, 0xff,
1921 srq->bit_fields_len * sizeof(u32));
1924 if (init_attr->attr.srq_limit) {
1925 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1931 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1939 ocrdma_mbx_destroy_srq(dev, srq);
1941 kfree(srq->rqe_wr_id_tbl);
1942 kfree(srq->idx_bit_fields);
1944 return ERR_PTR(status);
1947 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1948 struct ib_srq_attr *srq_attr,
1949 enum ib_srq_attr_mask srq_attr_mask,
1950 struct ib_udata *udata)
1953 struct ocrdma_srq *srq;
1955 srq = get_ocrdma_srq(ibsrq);
1956 if (srq_attr_mask & IB_SRQ_MAX_WR)
1959 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1963 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1966 struct ocrdma_srq *srq;
1968 srq = get_ocrdma_srq(ibsrq);
1969 status = ocrdma_mbx_query_srq(srq, srq_attr);
1973 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1976 struct ocrdma_srq *srq;
1977 struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1979 srq = get_ocrdma_srq(ibsrq);
1981 status = ocrdma_mbx_destroy_srq(dev, srq);
1984 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1985 PAGE_ALIGN(srq->rq.len));
1987 kfree(srq->idx_bit_fields);
1988 kfree(srq->rqe_wr_id_tbl);
1993 /* unprivileged verbs and their support functions. */
1994 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1995 struct ocrdma_hdr_wqe *hdr,
1996 struct ib_send_wr *wr)
1998 struct ocrdma_ewqe_ud_hdr *ud_hdr =
1999 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
2000 struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
2002 ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
2003 if (qp->qp_type == IB_QPT_GSI)
2004 ud_hdr->qkey = qp->qkey;
2006 ud_hdr->qkey = wr->wr.ud.remote_qkey;
2007 ud_hdr->rsvd_ahid = ah->id;
2008 if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
2009 hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
2012 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
2013 struct ocrdma_sge *sge, int num_sge,
2014 struct ib_sge *sg_list)
2018 for (i = 0; i < num_sge; i++) {
2019 sge[i].lrkey = sg_list[i].lkey;
2020 sge[i].addr_lo = sg_list[i].addr;
2021 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
2022 sge[i].len = sg_list[i].length;
2023 hdr->total_len += sg_list[i].length;
2026 memset(sge, 0, sizeof(*sge));
2029 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
2031 uint32_t total_len = 0, i;
2033 for (i = 0; i < num_sge; i++)
2034 total_len += sg_list[i].length;
2039 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
2040 struct ocrdma_hdr_wqe *hdr,
2041 struct ocrdma_sge *sge,
2042 struct ib_send_wr *wr, u32 wqe_size)
2047 if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
2048 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
2049 if (unlikely(hdr->total_len > qp->max_inline_data)) {
2050 pr_err("%s() supported_len=0x%x,\n"
2051 " unsupported len req=0x%x\n", __func__,
2052 qp->max_inline_data, hdr->total_len);
2055 dpp_addr = (char *)sge;
2056 for (i = 0; i < wr->num_sge; i++) {
2058 (void *)(unsigned long)wr->sg_list[i].addr,
2059 wr->sg_list[i].length);
2060 dpp_addr += wr->sg_list[i].length;
2063 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
2064 if (0 == hdr->total_len)
2065 wqe_size += sizeof(struct ocrdma_sge);
2066 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
2068 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2070 wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
2072 wqe_size += sizeof(struct ocrdma_sge);
2073 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2075 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2079 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2080 struct ib_send_wr *wr)
2083 struct ocrdma_sge *sge;
2084 u32 wqe_size = sizeof(*hdr);
2086 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2087 ocrdma_build_ud_hdr(qp, hdr, wr);
2088 sge = (struct ocrdma_sge *)(hdr + 2);
2089 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
2091 sge = (struct ocrdma_sge *)(hdr + 1);
2094 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2098 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2099 struct ib_send_wr *wr)
2102 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2103 struct ocrdma_sge *sge = ext_rw + 1;
2104 u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
2106 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2109 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
2110 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
2111 ext_rw->lrkey = wr->wr.rdma.rkey;
2112 ext_rw->len = hdr->total_len;
2116 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2117 struct ib_send_wr *wr)
2119 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2120 struct ocrdma_sge *sge = ext_rw + 1;
2121 u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2122 sizeof(struct ocrdma_hdr_wqe);
2124 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2125 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2126 hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2127 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2129 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
2130 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
2131 ext_rw->lrkey = wr->wr.rdma.rkey;
2132 ext_rw->len = hdr->total_len;
2135 static void build_frmr_pbes(struct ib_send_wr *wr, struct ocrdma_pbl *pbl_tbl,
2136 struct ocrdma_hw_mr *hwmr)
2141 struct ocrdma_pbe *pbe;
2143 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2146 /* go through the OS phy regions & fill hw pbe entries into pbls. */
2147 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
2148 /* number of pbes can be more for one OS buf, when
2149 * buffers are of different sizes.
2150 * split the ib_buf to one or more pbes.
2152 buf_addr = wr->wr.fast_reg.page_list->page_list[i];
2153 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2154 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2158 /* if the pbl is full storing the pbes,
2161 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
2163 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2169 static int get_encoded_page_size(int pg_sz)
2171 /* Max size is 256M 4096 << 16 */
2174 if (pg_sz == (4096 << i))
2180 static int ocrdma_build_fr(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2181 struct ib_send_wr *wr)
2184 struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2185 struct ocrdma_mr *mr;
2186 struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2187 u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2189 wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2191 if (wr->wr.fast_reg.page_list_len > dev->attr.max_pages_per_frmr)
2194 hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2195 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2197 if (wr->wr.fast_reg.page_list_len == 0)
2199 if (wr->wr.fast_reg.access_flags & IB_ACCESS_LOCAL_WRITE)
2200 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2201 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_WRITE)
2202 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2203 if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_READ)
2204 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2205 hdr->lkey = wr->wr.fast_reg.rkey;
2206 hdr->total_len = wr->wr.fast_reg.length;
2208 fbo = wr->wr.fast_reg.iova_start -
2209 (wr->wr.fast_reg.page_list->page_list[0] & PAGE_MASK);
2211 fast_reg->va_hi = upper_32_bits(wr->wr.fast_reg.iova_start);
2212 fast_reg->va_lo = (u32) (wr->wr.fast_reg.iova_start & 0xffffffff);
2213 fast_reg->fbo_hi = upper_32_bits(fbo);
2214 fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2215 fast_reg->num_sges = wr->wr.fast_reg.page_list_len;
2216 fast_reg->size_sge =
2217 get_encoded_page_size(1 << wr->wr.fast_reg.page_shift);
2218 mr = (struct ocrdma_mr *) (unsigned long)
2219 dev->stag_arr[(hdr->lkey >> 8) & (OCRDMA_MAX_STAG - 1)];
2220 build_frmr_pbes(wr, mr->hwmr.pbl_table, &mr->hwmr);
2224 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2226 u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2228 iowrite32(val, qp->sq_db);
2231 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2232 struct ib_send_wr **bad_wr)
2235 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2236 struct ocrdma_hdr_wqe *hdr;
2237 unsigned long flags;
2239 spin_lock_irqsave(&qp->q_lock, flags);
2240 if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2241 spin_unlock_irqrestore(&qp->q_lock, flags);
2247 if (qp->qp_type == IB_QPT_UD &&
2248 (wr->opcode != IB_WR_SEND &&
2249 wr->opcode != IB_WR_SEND_WITH_IMM)) {
2254 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2255 wr->num_sge > qp->sq.max_sges) {
2260 hdr = ocrdma_hwq_head(&qp->sq);
2262 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2263 hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2264 if (wr->send_flags & IB_SEND_FENCE)
2266 (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2267 if (wr->send_flags & IB_SEND_SOLICITED)
2269 (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2271 switch (wr->opcode) {
2272 case IB_WR_SEND_WITH_IMM:
2273 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2274 hdr->immdt = ntohl(wr->ex.imm_data);
2276 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2277 ocrdma_build_send(qp, hdr, wr);
2279 case IB_WR_SEND_WITH_INV:
2280 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2281 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2282 hdr->lkey = wr->ex.invalidate_rkey;
2283 status = ocrdma_build_send(qp, hdr, wr);
2285 case IB_WR_RDMA_WRITE_WITH_IMM:
2286 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2287 hdr->immdt = ntohl(wr->ex.imm_data);
2288 case IB_WR_RDMA_WRITE:
2289 hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2290 status = ocrdma_build_write(qp, hdr, wr);
2292 case IB_WR_RDMA_READ:
2293 ocrdma_build_read(qp, hdr, wr);
2295 case IB_WR_LOCAL_INV:
2297 (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2298 hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2299 sizeof(struct ocrdma_sge)) /
2300 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2301 hdr->lkey = wr->ex.invalidate_rkey;
2303 case IB_WR_FAST_REG_MR:
2304 status = ocrdma_build_fr(qp, hdr, wr);
2314 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2315 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2317 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2318 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2319 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2320 OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2321 /* make sure wqe is written before adapter can access it */
2323 /* inform hw to start processing it */
2324 ocrdma_ring_sq_db(qp);
2326 /* update pointer, counter for next wr */
2327 ocrdma_hwq_inc_head(&qp->sq);
2330 spin_unlock_irqrestore(&qp->q_lock, flags);
2334 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2336 u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2338 iowrite32(val, qp->rq_db);
2341 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2345 struct ocrdma_sge *sge;
2347 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2349 wqe_size = sizeof(*sge) + sizeof(*rqe);
2351 rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2352 OCRDMA_WQE_SIZE_SHIFT);
2353 rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2354 rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2356 rqe->rsvd_tag = tag;
2357 sge = (struct ocrdma_sge *)(rqe + 1);
2358 ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2359 ocrdma_cpu_to_le32(rqe, wqe_size);
2362 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2363 struct ib_recv_wr **bad_wr)
2366 unsigned long flags;
2367 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2368 struct ocrdma_hdr_wqe *rqe;
2370 spin_lock_irqsave(&qp->q_lock, flags);
2371 if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2372 spin_unlock_irqrestore(&qp->q_lock, flags);
2377 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2378 wr->num_sge > qp->rq.max_sges) {
2383 rqe = ocrdma_hwq_head(&qp->rq);
2384 ocrdma_build_rqe(rqe, wr, 0);
2386 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2387 /* make sure rqe is written before adapter can access it */
2390 /* inform hw to start processing it */
2391 ocrdma_ring_rq_db(qp);
2393 /* update pointer, counter for next wr */
2394 ocrdma_hwq_inc_head(&qp->rq);
2397 spin_unlock_irqrestore(&qp->q_lock, flags);
2401 /* cqe for srq's rqe can potentially arrive out of order.
2402 * index gives the entry in the shadow table where to store
2403 * the wr_id. tag/index is returned in cqe to reference back
2406 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2411 for (row = 0; row < srq->bit_fields_len; row++) {
2412 if (srq->idx_bit_fields[row]) {
2413 indx = ffs(srq->idx_bit_fields[row]);
2414 indx = (row * 32) + (indx - 1);
2415 if (indx >= srq->rq.max_cnt)
2417 ocrdma_srq_toggle_bit(srq, indx);
2422 if (row == srq->bit_fields_len)
2424 return indx + 1; /* Use from index 1 */
2427 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2429 u32 val = srq->rq.dbid | (1 << 16);
2431 iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2434 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2435 struct ib_recv_wr **bad_wr)
2438 unsigned long flags;
2439 struct ocrdma_srq *srq;
2440 struct ocrdma_hdr_wqe *rqe;
2443 srq = get_ocrdma_srq(ibsrq);
2445 spin_lock_irqsave(&srq->q_lock, flags);
2447 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2448 wr->num_sge > srq->rq.max_sges) {
2453 tag = ocrdma_srq_get_idx(srq);
2454 rqe = ocrdma_hwq_head(&srq->rq);
2455 ocrdma_build_rqe(rqe, wr, tag);
2457 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2458 /* make sure rqe is written before adapter can perform DMA */
2460 /* inform hw to start processing it */
2461 ocrdma_ring_srq_db(srq);
2462 /* update pointer, counter for next wr */
2463 ocrdma_hwq_inc_head(&srq->rq);
2466 spin_unlock_irqrestore(&srq->q_lock, flags);
2470 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2472 enum ib_wc_status ibwc_status;
2475 case OCRDMA_CQE_GENERAL_ERR:
2476 ibwc_status = IB_WC_GENERAL_ERR;
2478 case OCRDMA_CQE_LOC_LEN_ERR:
2479 ibwc_status = IB_WC_LOC_LEN_ERR;
2481 case OCRDMA_CQE_LOC_QP_OP_ERR:
2482 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2484 case OCRDMA_CQE_LOC_EEC_OP_ERR:
2485 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2487 case OCRDMA_CQE_LOC_PROT_ERR:
2488 ibwc_status = IB_WC_LOC_PROT_ERR;
2490 case OCRDMA_CQE_WR_FLUSH_ERR:
2491 ibwc_status = IB_WC_WR_FLUSH_ERR;
2493 case OCRDMA_CQE_MW_BIND_ERR:
2494 ibwc_status = IB_WC_MW_BIND_ERR;
2496 case OCRDMA_CQE_BAD_RESP_ERR:
2497 ibwc_status = IB_WC_BAD_RESP_ERR;
2499 case OCRDMA_CQE_LOC_ACCESS_ERR:
2500 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2502 case OCRDMA_CQE_REM_INV_REQ_ERR:
2503 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2505 case OCRDMA_CQE_REM_ACCESS_ERR:
2506 ibwc_status = IB_WC_REM_ACCESS_ERR;
2508 case OCRDMA_CQE_REM_OP_ERR:
2509 ibwc_status = IB_WC_REM_OP_ERR;
2511 case OCRDMA_CQE_RETRY_EXC_ERR:
2512 ibwc_status = IB_WC_RETRY_EXC_ERR;
2514 case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2515 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2517 case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2518 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2520 case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2521 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2523 case OCRDMA_CQE_REM_ABORT_ERR:
2524 ibwc_status = IB_WC_REM_ABORT_ERR;
2526 case OCRDMA_CQE_INV_EECN_ERR:
2527 ibwc_status = IB_WC_INV_EECN_ERR;
2529 case OCRDMA_CQE_INV_EEC_STATE_ERR:
2530 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2532 case OCRDMA_CQE_FATAL_ERR:
2533 ibwc_status = IB_WC_FATAL_ERR;
2535 case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2536 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2539 ibwc_status = IB_WC_GENERAL_ERR;
2545 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2548 struct ocrdma_hdr_wqe *hdr;
2549 struct ocrdma_sge *rw;
2552 hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2554 ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2555 /* Undo the hdr->cw swap */
2556 opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2559 ibwc->opcode = IB_WC_RDMA_WRITE;
2562 rw = (struct ocrdma_sge *)(hdr + 1);
2563 ibwc->opcode = IB_WC_RDMA_READ;
2564 ibwc->byte_len = rw->len;
2567 ibwc->opcode = IB_WC_SEND;
2570 ibwc->opcode = IB_WC_FAST_REG_MR;
2572 case OCRDMA_LKEY_INV:
2573 ibwc->opcode = IB_WC_LOCAL_INV;
2576 ibwc->status = IB_WC_GENERAL_ERR;
2577 pr_err("%s() invalid opcode received = 0x%x\n",
2578 __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2583 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2584 struct ocrdma_cqe *cqe)
2586 if (is_cqe_for_sq(cqe)) {
2587 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2588 cqe->flags_status_srcqpn) &
2589 ~OCRDMA_CQE_STATUS_MASK);
2590 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2591 cqe->flags_status_srcqpn) |
2592 (OCRDMA_CQE_WR_FLUSH_ERR <<
2593 OCRDMA_CQE_STATUS_SHIFT));
2595 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2596 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2597 cqe->flags_status_srcqpn) &
2598 ~OCRDMA_CQE_UD_STATUS_MASK);
2599 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2600 cqe->flags_status_srcqpn) |
2601 (OCRDMA_CQE_WR_FLUSH_ERR <<
2602 OCRDMA_CQE_UD_STATUS_SHIFT));
2604 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2605 cqe->flags_status_srcqpn) &
2606 ~OCRDMA_CQE_STATUS_MASK);
2607 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2608 cqe->flags_status_srcqpn) |
2609 (OCRDMA_CQE_WR_FLUSH_ERR <<
2610 OCRDMA_CQE_STATUS_SHIFT));
2615 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2616 struct ocrdma_qp *qp, int status)
2618 bool expand = false;
2621 ibwc->qp = &qp->ibqp;
2622 ibwc->status = ocrdma_to_ibwc_err(status);
2624 ocrdma_flush_qp(qp);
2625 ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2627 /* if wqe/rqe pending for which cqe needs to be returned,
2628 * trigger inflating it.
2630 if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2632 ocrdma_set_cqe_status_flushed(qp, cqe);
2637 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2638 struct ocrdma_qp *qp, int status)
2640 ibwc->opcode = IB_WC_RECV;
2641 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2642 ocrdma_hwq_inc_tail(&qp->rq);
2644 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2647 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2648 struct ocrdma_qp *qp, int status)
2650 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2651 ocrdma_hwq_inc_tail(&qp->sq);
2653 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2657 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2658 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2659 bool *polled, bool *stop)
2662 struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2663 int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2664 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2665 if (status < OCRDMA_MAX_CQE_ERR)
2666 atomic_inc(&dev->cqe_err_stats[status]);
2668 /* when hw sq is empty, but rq is not empty, so we continue
2669 * to keep the cqe in order to get the cq event again.
2671 if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2672 /* when cq for rq and sq is same, it is safe to return
2673 * flush cqe for RQEs.
2675 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2677 status = OCRDMA_CQE_WR_FLUSH_ERR;
2678 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2680 /* stop processing further cqe as this cqe is used for
2681 * triggering cq event on buddy cq of RQ.
2682 * When QP is destroyed, this cqe will be removed
2683 * from the cq's hardware q.
2689 } else if (is_hw_sq_empty(qp)) {
2696 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2701 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2702 struct ocrdma_cqe *cqe,
2703 struct ib_wc *ibwc, bool *polled)
2705 bool expand = false;
2706 int tail = qp->sq.tail;
2709 if (!qp->wqe_wr_id_tbl[tail].signaled) {
2710 *polled = false; /* WC cannot be consumed yet */
2712 ibwc->status = IB_WC_SUCCESS;
2714 ibwc->qp = &qp->ibqp;
2715 ocrdma_update_wc(qp, ibwc, tail);
2718 wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2719 OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2720 if (tail != wqe_idx)
2721 expand = true; /* Coalesced CQE can't be consumed yet */
2723 ocrdma_hwq_inc_tail(&qp->sq);
2727 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2728 struct ib_wc *ibwc, bool *polled, bool *stop)
2733 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2734 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2736 if (status == OCRDMA_CQE_SUCCESS)
2737 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2739 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2743 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2747 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2748 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2749 ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2750 OCRDMA_CQE_SRCQP_MASK;
2751 ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2752 OCRDMA_CQE_PKEY_MASK;
2753 ibwc->wc_flags = IB_WC_GRH;
2754 ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2755 OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2759 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2760 struct ocrdma_cqe *cqe,
2761 struct ocrdma_qp *qp)
2763 unsigned long flags;
2764 struct ocrdma_srq *srq;
2767 srq = get_ocrdma_srq(qp->ibqp.srq);
2768 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2769 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2773 ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2774 spin_lock_irqsave(&srq->q_lock, flags);
2775 ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2776 spin_unlock_irqrestore(&srq->q_lock, flags);
2777 ocrdma_hwq_inc_tail(&srq->rq);
2780 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2781 struct ib_wc *ibwc, bool *polled, bool *stop,
2785 struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2787 if (status < OCRDMA_MAX_CQE_ERR)
2788 atomic_inc(&dev->cqe_err_stats[status]);
2790 /* when hw_rq is empty, but wq is not empty, so continue
2791 * to keep the cqe to get the cq event again.
2793 if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2794 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2796 status = OCRDMA_CQE_WR_FLUSH_ERR;
2797 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2803 } else if (is_hw_rq_empty(qp)) {
2810 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2815 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2816 struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2818 ibwc->opcode = IB_WC_RECV;
2819 ibwc->qp = &qp->ibqp;
2820 ibwc->status = IB_WC_SUCCESS;
2822 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2823 ocrdma_update_ud_rcqe(ibwc, cqe);
2825 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2827 if (is_cqe_imm(cqe)) {
2828 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2829 ibwc->wc_flags |= IB_WC_WITH_IMM;
2830 } else if (is_cqe_wr_imm(cqe)) {
2831 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2832 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2833 ibwc->wc_flags |= IB_WC_WITH_IMM;
2834 } else if (is_cqe_invalidated(cqe)) {
2835 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2836 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2839 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2841 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2842 ocrdma_hwq_inc_tail(&qp->rq);
2846 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2847 struct ib_wc *ibwc, bool *polled, bool *stop)
2850 bool expand = false;
2853 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2854 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2855 OCRDMA_CQE_UD_STATUS_MASK) >>
2856 OCRDMA_CQE_UD_STATUS_SHIFT;
2858 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2859 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2862 if (status == OCRDMA_CQE_SUCCESS) {
2864 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2866 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2872 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2875 if (cq->phase_change) {
2877 cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2879 /* clear valid bit */
2880 cqe->flags_status_srcqpn = 0;
2884 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2889 bool expand = false;
2890 int polled_hw_cqes = 0;
2891 struct ocrdma_qp *qp = NULL;
2892 struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2893 struct ocrdma_cqe *cqe;
2894 u16 cur_getp; bool polled = false; bool stop = false;
2896 cur_getp = cq->getp;
2897 while (num_entries) {
2898 cqe = cq->va + cur_getp;
2899 /* check whether valid cqe or not */
2900 if (!is_cqe_valid(cq, cqe))
2902 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2903 /* ignore discarded cqe */
2906 qp = dev->qp_tbl[qpn];
2909 if (is_cqe_for_sq(cqe)) {
2910 expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2913 expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2920 /* clear qpn to avoid duplicate processing by discard_cqe() */
2923 polled_hw_cqes += 1;
2924 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2925 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2935 cq->getp = cur_getp;
2936 if (cq->deferred_arm || polled_hw_cqes) {
2937 ocrdma_ring_cq_db(dev, cq->id, cq->deferred_arm,
2938 cq->deferred_sol, polled_hw_cqes);
2939 cq->deferred_arm = false;
2940 cq->deferred_sol = false;
2946 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2947 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2948 struct ocrdma_qp *qp, struct ib_wc *ibwc)
2952 while (num_entries) {
2953 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2955 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2956 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2957 ocrdma_hwq_inc_tail(&qp->sq);
2958 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2959 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2960 ocrdma_hwq_inc_tail(&qp->rq);
2965 ibwc->status = IB_WC_WR_FLUSH_ERR;
2973 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2975 int cqes_to_poll = num_entries;
2976 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2977 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2978 int num_os_cqe = 0, err_cqes = 0;
2979 struct ocrdma_qp *qp;
2980 unsigned long flags;
2982 /* poll cqes from adapter CQ */
2983 spin_lock_irqsave(&cq->cq_lock, flags);
2984 num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2985 spin_unlock_irqrestore(&cq->cq_lock, flags);
2986 cqes_to_poll -= num_os_cqe;
2989 wc = wc + num_os_cqe;
2990 /* adapter returns single error cqe when qp moves to
2991 * error state. So insert error cqes with wc_status as
2992 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2993 * respectively which uses this CQ.
2995 spin_lock_irqsave(&dev->flush_q_lock, flags);
2996 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2997 if (cqes_to_poll == 0)
2999 err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
3000 cqes_to_poll -= err_cqes;
3001 num_os_cqe += err_cqes;
3004 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
3009 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
3011 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
3012 struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
3014 unsigned long flags;
3015 bool arm_needed = false, sol_needed = false;
3019 spin_lock_irqsave(&cq->cq_lock, flags);
3020 if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
3022 if (cq_flags & IB_CQ_SOLICITED)
3025 if (cq->first_arm) {
3026 ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
3027 cq->first_arm = false;
3030 cq->deferred_arm = true;
3031 cq->deferred_sol = sol_needed;
3032 spin_unlock_irqrestore(&cq->cq_lock, flags);
3037 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd,
3038 enum ib_mr_type mr_type,
3042 struct ocrdma_mr *mr;
3043 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3044 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3046 if (mr_type != IB_MR_TYPE_MEM_REG)
3047 return ERR_PTR(-EINVAL);
3049 if (max_num_sg > dev->attr.max_pages_per_frmr)
3050 return ERR_PTR(-EINVAL);
3052 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3054 return ERR_PTR(-ENOMEM);
3056 status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
3060 mr->hwmr.remote_rd = 0;
3061 mr->hwmr.remote_wr = 0;
3062 mr->hwmr.local_rd = 0;
3063 mr->hwmr.local_wr = 0;
3064 mr->hwmr.mw_bind = 0;
3065 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3068 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
3071 mr->ibmr.rkey = mr->hwmr.lkey;
3072 mr->ibmr.lkey = mr->hwmr.lkey;
3073 dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
3077 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3080 return ERR_PTR(-ENOMEM);
3083 struct ib_fast_reg_page_list *ocrdma_alloc_frmr_page_list(struct ib_device
3087 struct ib_fast_reg_page_list *frmr_list;
3090 size = sizeof(*frmr_list) + (page_list_len * sizeof(u64));
3091 frmr_list = kzalloc(size, GFP_KERNEL);
3093 return ERR_PTR(-ENOMEM);
3094 frmr_list->page_list = (u64 *)(frmr_list + 1);
3098 void ocrdma_free_frmr_page_list(struct ib_fast_reg_page_list *page_list)
3103 #define MAX_KERNEL_PBE_SIZE 65536
3104 static inline int count_kernel_pbes(struct ib_phys_buf *buf_list,
3105 int buf_cnt, u32 *pbe_size)
3110 *pbe_size = roundup(buf_list[0].size, PAGE_SIZE);
3111 *pbe_size = roundup_pow_of_two(*pbe_size);
3113 /* find the smallest PBE size that we can have */
3114 for (i = 0; i < buf_cnt; i++) {
3115 /* first addr may not be page aligned, so ignore checking */
3116 if ((i != 0) && ((buf_list[i].addr & ~PAGE_MASK) ||
3117 (buf_list[i].size & ~PAGE_MASK))) {
3121 /* if configured PBE size is greater then the chosen one,
3122 * reduce the PBE size.
3124 buf_size = roundup(buf_list[i].size, PAGE_SIZE);
3125 /* pbe_size has to be even multiple of 4K 1,2,4,8...*/
3126 buf_size = roundup_pow_of_two(buf_size);
3127 if (*pbe_size > buf_size)
3128 *pbe_size = buf_size;
3130 total_size += buf_size;
3132 *pbe_size = *pbe_size > MAX_KERNEL_PBE_SIZE ?
3133 (MAX_KERNEL_PBE_SIZE) : (*pbe_size);
3135 /* num_pbes = total_size / (*pbe_size); this is implemented below. */
3137 return total_size >> ilog2(*pbe_size);
3140 static void build_kernel_pbes(struct ib_phys_buf *buf_list, int ib_buf_cnt,
3141 u32 pbe_size, struct ocrdma_pbl *pbl_tbl,
3142 struct ocrdma_hw_mr *hwmr)
3146 int pbes_per_buf = 0;
3149 struct ocrdma_pbe *pbe;
3150 int total_num_pbes = 0;
3152 if (!hwmr->num_pbes)
3155 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
3158 /* go through the OS phy regions & fill hw pbe entries into pbls. */
3159 for (i = 0; i < ib_buf_cnt; i++) {
3160 buf_addr = buf_list[i].addr;
3162 roundup_pow_of_two(roundup(buf_list[i].size, PAGE_SIZE)) /
3164 hwmr->len += buf_list[i].size;
3165 /* number of pbes can be more for one OS buf, when
3166 * buffers are of different sizes.
3167 * split the ib_buf to one or more pbes.
3169 for (idx = 0; idx < pbes_per_buf; idx++) {
3170 /* we program always page aligned addresses,
3171 * first unaligned address is taken care by fbo.
3174 /* for non zero fbo, assign the
3175 * start of the page.
3178 cpu_to_le32((u32) (buf_addr & PAGE_MASK));
3180 cpu_to_le32((u32) upper_32_bits(buf_addr));
3183 cpu_to_le32((u32) (buf_addr & 0xffffffff));
3185 cpu_to_le32((u32) upper_32_bits(buf_addr));
3187 buf_addr += pbe_size;
3189 total_num_pbes += 1;
3192 if (total_num_pbes == hwmr->num_pbes)
3194 /* if the pbl is full storing the pbes,
3197 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
3199 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
3208 struct ib_mr *ocrdma_reg_kernel_mr(struct ib_pd *ibpd,
3209 struct ib_phys_buf *buf_list,
3210 int buf_cnt, int acc, u64 *iova_start)
3212 int status = -ENOMEM;
3213 struct ocrdma_mr *mr;
3214 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3215 struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3219 if ((acc & IB_ACCESS_REMOTE_WRITE) && !(acc & IB_ACCESS_LOCAL_WRITE))
3220 return ERR_PTR(-EINVAL);
3222 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3224 return ERR_PTR(status);
3226 num_pbes = count_kernel_pbes(buf_list, buf_cnt, &pbe_size);
3227 if (num_pbes == 0) {
3231 status = ocrdma_get_pbl_info(dev, mr, num_pbes);
3235 mr->hwmr.pbe_size = pbe_size;
3236 mr->hwmr.fbo = *iova_start - (buf_list[0].addr & PAGE_MASK);
3237 mr->hwmr.va = *iova_start;
3238 mr->hwmr.local_rd = 1;
3239 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3240 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3241 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3242 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3243 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
3245 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3248 build_kernel_pbes(buf_list, buf_cnt, pbe_size, mr->hwmr.pbl_table,
3250 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
3254 mr->ibmr.lkey = mr->hwmr.lkey;
3255 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
3256 mr->ibmr.rkey = mr->hwmr.lkey;
3260 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3263 return ERR_PTR(status);