1 /*******************************************************************
2 * This file is part of the Emulex RoCE Device Driver for *
3 * RoCE (RDMA over Converged Ethernet) adapters. *
4 * Copyright (C) 2008-2012 Emulex. All rights reserved. *
5 * EMULEX and SLI are trademarks of Emulex. *
8 * This program is free software; you can redistribute it and/or *
9 * modify it under the terms of version 2 of the GNU General *
10 * Public License as published by the Free Software Foundation. *
11 * This program is distributed in the hope that it will be useful. *
12 * ALL EXPRESS OR IMPLIED CONDITIONS, REPRESENTATIONS AND *
13 * WARRANTIES, INCLUDING ANY IMPLIED WARRANTY OF MERCHANTABILITY, *
14 * FITNESS FOR A PARTICULAR PURPOSE, OR NON-INFRINGEMENT, ARE *
15 * DISCLAIMED, EXCEPT TO THE EXTENT THAT SUCH DISCLAIMERS ARE HELD *
16 * TO BE LEGALLY INVALID. See the GNU General Public License for *
17 * more details, a copy of which can be found in the file COPYING *
18 * included with this package. *
20 * Contact Information:
21 * linux-drivers@emulex.com
25 * Costa Mesa, CA 92626
26 *******************************************************************/
28 #include <linux/dma-mapping.h>
29 #include <rdma/ib_verbs.h>
30 #include <rdma/ib_user_verbs.h>
31 #include <rdma/iw_cm.h>
32 #include <rdma/ib_umem.h>
33 #include <rdma/ib_addr.h>
36 #include "ocrdma_hw.h"
37 #include "ocrdma_verbs.h"
38 #include "ocrdma_abi.h"
40 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
49 int ocrdma_query_gid(struct ib_device *ibdev, u8 port,
50 int index, union ib_gid *sgid)
52 struct ocrdma_dev *dev;
54 dev = get_ocrdma_dev(ibdev);
55 memset(sgid, 0, sizeof(*sgid));
56 if (index >= OCRDMA_MAX_SGID)
59 memcpy(sgid, &dev->sgid_tbl[index], sizeof(*sgid));
64 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr)
66 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
68 memset(attr, 0, sizeof *attr);
69 memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
70 min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
71 ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
72 attr->max_mr_size = ~0ull;
73 attr->page_size_cap = 0xffff000;
74 attr->vendor_id = dev->nic_info.pdev->vendor;
75 attr->vendor_part_id = dev->nic_info.pdev->device;
77 attr->max_qp = dev->attr.max_qp;
78 attr->max_ah = dev->attr.max_qp;
79 attr->max_qp_wr = dev->attr.max_wqe;
81 attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
82 IB_DEVICE_RC_RNR_NAK_GEN |
83 IB_DEVICE_SHUTDOWN_PORT |
84 IB_DEVICE_SYS_IMAGE_GUID |
85 IB_DEVICE_LOCAL_DMA_LKEY;
86 attr->max_sge = min(dev->attr.max_send_sge, dev->attr.max_srq_sge);
88 attr->max_cq = dev->attr.max_cq;
89 attr->max_cqe = dev->attr.max_cqe;
90 attr->max_mr = dev->attr.max_mr;
92 attr->max_pd = dev->attr.max_pd;
95 attr->max_map_per_fmr = 0;
96 attr->max_qp_rd_atom =
97 min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
98 attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
99 attr->max_srq = (dev->attr.max_qp - 1);
100 attr->max_srq_sge = attr->max_srq_sge;
101 attr->max_srq_wr = dev->attr.max_rqe;
102 attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
103 attr->max_fast_reg_page_list_len = 0;
108 int ocrdma_query_port(struct ib_device *ibdev,
109 u8 port, struct ib_port_attr *props)
111 enum ib_port_state port_state;
112 struct ocrdma_dev *dev;
113 struct net_device *netdev;
115 dev = get_ocrdma_dev(ibdev);
117 ocrdma_err("%s(%d) invalid_port=0x%x\n", __func__,
121 netdev = dev->nic_info.netdev;
122 if (netif_running(netdev) && netif_oper_up(netdev)) {
123 port_state = IB_PORT_ACTIVE;
124 props->phys_state = 5;
126 port_state = IB_PORT_DOWN;
127 props->phys_state = 3;
129 props->max_mtu = IB_MTU_4096;
130 props->active_mtu = iboe_get_mtu(netdev->mtu);
135 props->state = port_state;
136 props->port_cap_flags =
139 IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP;
140 props->gid_tbl_len = OCRDMA_MAX_SGID;
141 props->pkey_tbl_len = 1;
142 props->bad_pkey_cntr = 0;
143 props->qkey_viol_cntr = 0;
144 props->active_width = IB_WIDTH_1X;
145 props->active_speed = 4;
146 props->max_msg_sz = 0x80000000;
147 props->max_vl_num = 4;
151 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
152 struct ib_port_modify *props)
154 struct ocrdma_dev *dev;
156 dev = get_ocrdma_dev(ibdev);
158 ocrdma_err("%s(%d) invalid_port=0x%x\n", __func__,
165 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
168 struct ocrdma_mm *mm;
170 mm = kzalloc(sizeof(*mm), GFP_KERNEL);
173 mm->key.phy_addr = phy_addr;
175 INIT_LIST_HEAD(&mm->entry);
177 mutex_lock(&uctx->mm_list_lock);
178 list_add_tail(&mm->entry, &uctx->mm_head);
179 mutex_unlock(&uctx->mm_list_lock);
183 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
186 struct ocrdma_mm *mm, *tmp;
188 mutex_lock(&uctx->mm_list_lock);
189 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
190 if (len != mm->key.len || phy_addr != mm->key.phy_addr)
193 list_del(&mm->entry);
197 mutex_unlock(&uctx->mm_list_lock);
200 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
204 struct ocrdma_mm *mm;
206 mutex_lock(&uctx->mm_list_lock);
207 list_for_each_entry(mm, &uctx->mm_head, entry) {
208 if (len != mm->key.len || phy_addr != mm->key.phy_addr)
214 mutex_unlock(&uctx->mm_list_lock);
218 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
219 struct ib_udata *udata)
222 struct ocrdma_ucontext *ctx;
223 struct ocrdma_alloc_ucontext_resp resp;
224 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
225 struct pci_dev *pdev = dev->nic_info.pdev;
226 u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
229 return ERR_PTR(-EFAULT);
230 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
232 return ERR_PTR(-ENOMEM);
234 INIT_LIST_HEAD(&ctx->mm_head);
235 mutex_init(&ctx->mm_list_lock);
237 ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
238 &ctx->ah_tbl.pa, GFP_KERNEL);
239 if (!ctx->ah_tbl.va) {
241 return ERR_PTR(-ENOMEM);
243 memset(ctx->ah_tbl.va, 0, map_len);
244 ctx->ah_tbl.len = map_len;
246 resp.ah_tbl_len = ctx->ah_tbl.len;
247 resp.ah_tbl_page = ctx->ah_tbl.pa;
249 status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
252 resp.dev_id = dev->id;
253 resp.max_inline_data = dev->attr.max_inline_data;
254 resp.wqe_size = dev->attr.wqe_size;
255 resp.rqe_size = dev->attr.rqe_size;
256 resp.dpp_wqe_size = dev->attr.wqe_size;
259 memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
260 status = ib_copy_to_udata(udata, &resp, sizeof(resp));
263 return &ctx->ibucontext;
266 ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
268 dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
271 return ERR_PTR(status);
274 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
276 struct ocrdma_mm *mm, *tmp;
277 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
278 struct pci_dev *pdev = uctx->dev->nic_info.pdev;
280 ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
281 dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
284 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
285 list_del(&mm->entry);
292 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
294 struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
295 struct ocrdma_dev *dev = ucontext->dev;
296 unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
297 u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
298 unsigned long len = (vma->vm_end - vma->vm_start);
302 if (vma->vm_start & (PAGE_SIZE - 1))
304 found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
308 if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
309 dev->nic_info.db_total_size)) &&
310 (len <= dev->nic_info.db_page_size)) {
311 /* doorbell mapping */
312 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
313 len, vma->vm_page_prot);
314 } else if (dev->nic_info.dpp_unmapped_len &&
315 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
316 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
317 dev->nic_info.dpp_unmapped_len)) &&
318 (len <= dev->nic_info.dpp_unmapped_len)) {
319 /* dpp area mapping */
320 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
321 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
322 len, vma->vm_page_prot);
324 /* queue memory mapping */
325 status = remap_pfn_range(vma, vma->vm_start,
326 vma->vm_pgoff, len, vma->vm_page_prot);
331 static int ocrdma_copy_pd_uresp(struct ocrdma_pd *pd,
332 struct ib_ucontext *ib_ctx,
333 struct ib_udata *udata)
337 u64 dpp_page_addr = 0;
339 struct ocrdma_alloc_pd_uresp rsp;
340 struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
343 rsp.dpp_enabled = pd->dpp_enabled;
344 db_page_addr = pd->dev->nic_info.unmapped_db +
345 (pd->id * pd->dev->nic_info.db_page_size);
346 db_page_size = pd->dev->nic_info.db_page_size;
348 status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
352 if (pd->dpp_enabled) {
353 dpp_page_addr = pd->dev->nic_info.dpp_unmapped_addr +
354 (pd->id * OCRDMA_DPP_PAGE_SIZE);
355 status = ocrdma_add_mmap(uctx, dpp_page_addr,
356 OCRDMA_DPP_PAGE_SIZE);
359 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
360 rsp.dpp_page_addr_lo = dpp_page_addr;
363 status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
372 ocrdma_del_mmap(pd->uctx, dpp_page_addr, OCRDMA_DPP_PAGE_SIZE);
374 ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
378 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
379 struct ib_ucontext *context,
380 struct ib_udata *udata)
382 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
383 struct ocrdma_pd *pd;
386 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
388 return ERR_PTR(-ENOMEM);
390 if (udata && context) {
391 pd->dpp_enabled = (dev->nic_info.dev_family ==
392 OCRDMA_GEN2_FAMILY) ? true : false;
394 pd->dpp_enabled ? OCRDMA_PD_MAX_DPP_ENABLED_QP : 0;
396 status = ocrdma_mbx_alloc_pd(dev, pd);
399 return ERR_PTR(status);
401 atomic_set(&pd->use_cnt, 0);
403 if (udata && context) {
404 status = ocrdma_copy_pd_uresp(pd, context, udata);
411 ocrdma_dealloc_pd(&pd->ibpd);
412 return ERR_PTR(status);
415 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
417 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
418 struct ocrdma_dev *dev = pd->dev;
422 if (atomic_read(&pd->use_cnt)) {
423 ocrdma_err("%s(%d) pd=0x%x is in use.\n",
424 __func__, dev->id, pd->id);
428 status = ocrdma_mbx_dealloc_pd(dev, pd);
430 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
431 (pd->id * OCRDMA_DPP_PAGE_SIZE);
433 ocrdma_del_mmap(pd->uctx, dpp_db, OCRDMA_DPP_PAGE_SIZE);
434 usr_db = dev->nic_info.unmapped_db +
435 (pd->id * dev->nic_info.db_page_size);
436 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
443 static struct ocrdma_mr *ocrdma_alloc_lkey(struct ib_pd *ibpd,
444 int acc, u32 num_pbls,
448 struct ocrdma_mr *mr;
449 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
450 struct ocrdma_dev *dev = pd->dev;
452 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
453 ocrdma_err("%s(%d) leaving err, invalid access rights\n",
455 return ERR_PTR(-EINVAL);
458 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
460 return ERR_PTR(-ENOMEM);
463 mr->hwmr.local_rd = 1;
464 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
465 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
466 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
467 mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
468 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
469 mr->hwmr.num_pbls = num_pbls;
471 status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pd->id, addr_check);
474 return ERR_PTR(-ENOMEM);
477 atomic_inc(&pd->use_cnt);
478 mr->ibmr.lkey = mr->hwmr.lkey;
479 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
480 mr->ibmr.rkey = mr->hwmr.lkey;
484 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
486 struct ocrdma_mr *mr;
488 mr = ocrdma_alloc_lkey(ibpd, acc, 0, OCRDMA_ADDR_CHECK_DISABLE);
495 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
496 struct ocrdma_hw_mr *mr)
498 struct pci_dev *pdev = dev->nic_info.pdev;
502 for (i = 0; i < mr->num_pbls; i++) {
503 if (!mr->pbl_table[i].va)
505 dma_free_coherent(&pdev->dev, mr->pbl_size,
507 mr->pbl_table[i].pa);
509 kfree(mr->pbl_table);
510 mr->pbl_table = NULL;
514 static int ocrdma_get_pbl_info(struct ocrdma_mr *mr, u32 num_pbes)
522 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
523 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
527 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
528 num_pbls = num_pbls / (pbl_size / sizeof(u64));
530 } while (num_pbls >= mr->hwmr.dev->attr.max_num_mr_pbl);
532 mr->hwmr.num_pbes = num_pbes;
533 mr->hwmr.num_pbls = num_pbls;
534 mr->hwmr.pbl_size = pbl_size;
538 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
542 u32 dma_len = mr->pbl_size;
543 struct pci_dev *pdev = dev->nic_info.pdev;
547 mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
548 mr->num_pbls, GFP_KERNEL);
553 for (i = 0; i < mr->num_pbls; i++) {
554 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
556 ocrdma_free_mr_pbl_tbl(dev, mr);
560 memset(va, 0, dma_len);
561 mr->pbl_table[i].va = va;
562 mr->pbl_table[i].pa = pa;
567 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
570 struct ocrdma_pbe *pbe;
571 struct ib_umem_chunk *chunk;
572 struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
573 struct ib_umem *umem = mr->umem;
574 int i, shift, pg_cnt, pages, pbe_cnt, total_num_pbes = 0;
576 if (!mr->hwmr.num_pbes)
579 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
582 shift = ilog2(umem->page_size);
584 list_for_each_entry(chunk, &umem->chunk_list, list) {
585 /* get all the dma regions from the chunk. */
586 for (i = 0; i < chunk->nmap; i++) {
587 pages = sg_dma_len(&chunk->page_list[i]) >> shift;
588 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
589 /* store the page address in pbe */
591 cpu_to_le32(sg_dma_address
592 (&chunk->page_list[i]) +
593 (umem->page_size * pg_cnt));
595 cpu_to_le32(upper_32_bits
597 (&chunk->page_list[i]) +
598 umem->page_size * pg_cnt)));
603 /* if done building pbes, issue the mbx cmd. */
604 if (total_num_pbes == num_pbes)
607 /* if the given pbl is full storing the pbes,
611 (mr->hwmr.pbl_size / sizeof(u64))) {
613 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
621 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
622 u64 usr_addr, int acc, struct ib_udata *udata)
624 int status = -ENOMEM;
625 struct ocrdma_dev *dev;
626 struct ocrdma_mr *mr;
627 struct ocrdma_pd *pd;
630 pd = get_ocrdma_pd(ibpd);
633 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
634 return ERR_PTR(-EINVAL);
636 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
638 return ERR_PTR(status);
640 mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
641 if (IS_ERR(mr->umem)) {
645 num_pbes = ib_umem_page_count(mr->umem);
646 status = ocrdma_get_pbl_info(mr, num_pbes);
650 mr->hwmr.pbe_size = mr->umem->page_size;
651 mr->hwmr.fbo = mr->umem->offset;
652 mr->hwmr.va = usr_addr;
654 mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
655 mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
656 mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
657 mr->hwmr.local_rd = 1;
658 mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
659 status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
662 build_user_pbes(dev, mr, num_pbes);
663 status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
667 atomic_inc(&pd->use_cnt);
668 mr->ibmr.lkey = mr->hwmr.lkey;
669 if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
670 mr->ibmr.rkey = mr->hwmr.lkey;
675 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
678 return ERR_PTR(status);
681 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
683 struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
684 struct ocrdma_dev *dev = mr->hwmr.dev;
687 status = ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
689 if (mr->hwmr.fr_mr == 0)
690 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
692 atomic_dec(&mr->pd->use_cnt);
693 /* it could be user registered memory. */
695 ib_umem_release(mr->umem);
700 static int ocrdma_copy_cq_uresp(struct ocrdma_cq *cq, struct ib_udata *udata,
701 struct ib_ucontext *ib_ctx)
704 struct ocrdma_ucontext *uctx;
705 struct ocrdma_create_cq_uresp uresp;
707 uresp.cq_id = cq->id;
708 uresp.page_size = cq->len;
710 uresp.max_hw_cqe = cq->max_hw_cqe;
711 uresp.page_addr[0] = cq->pa;
712 uresp.db_page_addr = cq->dev->nic_info.unmapped_db;
713 uresp.db_page_size = cq->dev->nic_info.db_page_size;
714 uresp.phase_change = cq->phase_change ? 1 : 0;
715 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
717 ocrdma_err("%s(%d) copy error cqid=0x%x.\n",
718 __func__, cq->dev->id, cq->id);
721 uctx = get_ocrdma_ucontext(ib_ctx);
722 status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
725 status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
727 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
735 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev, int entries, int vector,
736 struct ib_ucontext *ib_ctx,
737 struct ib_udata *udata)
739 struct ocrdma_cq *cq;
740 struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
742 struct ocrdma_create_cq_ureq ureq;
745 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
746 return ERR_PTR(-EFAULT);
749 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
751 return ERR_PTR(-ENOMEM);
753 spin_lock_init(&cq->cq_lock);
754 spin_lock_init(&cq->comp_handler_lock);
755 atomic_set(&cq->use_cnt, 0);
756 INIT_LIST_HEAD(&cq->sq_head);
757 INIT_LIST_HEAD(&cq->rq_head);
760 status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq);
763 return ERR_PTR(status);
766 status = ocrdma_copy_cq_uresp(cq, udata, ib_ctx);
770 cq->phase = OCRDMA_CQE_VALID;
771 cq->arm_needed = true;
772 dev->cq_tbl[cq->id] = cq;
777 ocrdma_mbx_destroy_cq(dev, cq);
779 return ERR_PTR(status);
782 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
783 struct ib_udata *udata)
786 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
788 if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
796 int ocrdma_destroy_cq(struct ib_cq *ibcq)
799 struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
800 struct ocrdma_dev *dev = cq->dev;
802 if (atomic_read(&cq->use_cnt))
805 status = ocrdma_mbx_destroy_cq(dev, cq);
808 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa, cq->len);
809 ocrdma_del_mmap(cq->ucontext, dev->nic_info.unmapped_db,
810 dev->nic_info.db_page_size);
812 dev->cq_tbl[cq->id] = NULL;
818 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
820 int status = -EINVAL;
822 if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
823 dev->qp_tbl[qp->id] = qp;
829 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
831 dev->qp_tbl[qp->id] = NULL;
834 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
835 struct ib_qp_init_attr *attrs)
837 if (attrs->qp_type != IB_QPT_GSI &&
838 attrs->qp_type != IB_QPT_RC &&
839 attrs->qp_type != IB_QPT_UD) {
840 ocrdma_err("%s(%d) unsupported qp type=0x%x requested\n",
841 __func__, dev->id, attrs->qp_type);
844 if (attrs->cap.max_send_wr > dev->attr.max_wqe) {
845 ocrdma_err("%s(%d) unsupported send_wr=0x%x requested\n",
846 __func__, dev->id, attrs->cap.max_send_wr);
847 ocrdma_err("%s(%d) supported send_wr=0x%x\n",
848 __func__, dev->id, dev->attr.max_wqe);
851 if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
852 ocrdma_err("%s(%d) unsupported recv_wr=0x%x requested\n",
853 __func__, dev->id, attrs->cap.max_recv_wr);
854 ocrdma_err("%s(%d) supported recv_wr=0x%x\n",
855 __func__, dev->id, dev->attr.max_rqe);
858 if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
859 ocrdma_err("%s(%d) unsupported inline data size=0x%x"
860 " requested\n", __func__, dev->id,
861 attrs->cap.max_inline_data);
862 ocrdma_err("%s(%d) supported inline data size=0x%x\n",
863 __func__, dev->id, dev->attr.max_inline_data);
866 if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
867 ocrdma_err("%s(%d) unsupported send_sge=0x%x requested\n",
868 __func__, dev->id, attrs->cap.max_send_sge);
869 ocrdma_err("%s(%d) supported send_sge=0x%x\n",
870 __func__, dev->id, dev->attr.max_send_sge);
873 if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
874 ocrdma_err("%s(%d) unsupported recv_sge=0x%x requested\n",
875 __func__, dev->id, attrs->cap.max_recv_sge);
876 ocrdma_err("%s(%d) supported recv_sge=0x%x\n",
877 __func__, dev->id, dev->attr.max_recv_sge);
880 /* unprivileged user space cannot create special QP */
881 if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
883 ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
884 __func__, dev->id, attrs->qp_type);
887 /* allow creating only one GSI type of QP */
888 if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
889 ocrdma_err("%s(%d) GSI special QPs already created.\n",
893 /* verify consumer QPs are not trying to use GSI QP's CQ */
894 if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
895 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
896 (dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq))) {
897 ocrdma_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
905 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
906 struct ib_udata *udata, int dpp_offset,
907 int dpp_credit_lmt, int srq)
911 struct ocrdma_create_qp_uresp uresp;
912 struct ocrdma_dev *dev = qp->dev;
913 struct ocrdma_pd *pd = qp->pd;
915 memset(&uresp, 0, sizeof(uresp));
916 usr_db = dev->nic_info.unmapped_db +
917 (pd->id * dev->nic_info.db_page_size);
918 uresp.qp_id = qp->id;
919 uresp.sq_dbid = qp->sq.dbid;
920 uresp.num_sq_pages = 1;
921 uresp.sq_page_size = qp->sq.len;
922 uresp.sq_page_addr[0] = qp->sq.pa;
923 uresp.num_wqe_allocated = qp->sq.max_cnt;
925 uresp.rq_dbid = qp->rq.dbid;
926 uresp.num_rq_pages = 1;
927 uresp.rq_page_size = qp->rq.len;
928 uresp.rq_page_addr[0] = qp->rq.pa;
929 uresp.num_rqe_allocated = qp->rq.max_cnt;
931 uresp.db_page_addr = usr_db;
932 uresp.db_page_size = dev->nic_info.db_page_size;
933 if (dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
934 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
935 uresp.db_rq_offset = ((qp->id & 0xFFFF) < 128) ?
936 OCRDMA_DB_GEN2_RQ1_OFFSET : OCRDMA_DB_GEN2_RQ2_OFFSET;
937 uresp.db_shift = (qp->id < 128) ? 24 : 16;
939 uresp.db_sq_offset = OCRDMA_DB_SQ_OFFSET;
940 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
944 if (qp->dpp_enabled) {
945 uresp.dpp_credit = dpp_credit_lmt;
946 uresp.dpp_offset = dpp_offset;
948 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
950 ocrdma_err("%s(%d) user copy error.\n", __func__, dev->id);
953 status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
959 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
966 ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
971 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
972 struct ocrdma_pd *pd)
974 if (dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
975 qp->sq_db = dev->nic_info.db +
976 (pd->id * dev->nic_info.db_page_size) +
977 OCRDMA_DB_GEN2_SQ_OFFSET;
978 qp->rq_db = dev->nic_info.db +
979 (pd->id * dev->nic_info.db_page_size) +
981 OCRDMA_DB_GEN2_RQ1_OFFSET : OCRDMA_DB_GEN2_RQ2_OFFSET);
983 qp->sq_db = dev->nic_info.db +
984 (pd->id * dev->nic_info.db_page_size) +
986 qp->rq_db = dev->nic_info.db +
987 (pd->id * dev->nic_info.db_page_size) +
992 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
995 kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
997 if (qp->wqe_wr_id_tbl == NULL)
1000 kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1001 if (qp->rqe_wr_id_tbl == NULL)
1007 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1008 struct ocrdma_pd *pd,
1009 struct ib_qp_init_attr *attrs)
1012 spin_lock_init(&qp->q_lock);
1013 INIT_LIST_HEAD(&qp->sq_entry);
1014 INIT_LIST_HEAD(&qp->rq_entry);
1016 qp->qp_type = attrs->qp_type;
1017 qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1018 qp->max_inline_data = attrs->cap.max_inline_data;
1019 qp->sq.max_sges = attrs->cap.max_send_sge;
1020 qp->rq.max_sges = attrs->cap.max_recv_sge;
1021 qp->state = OCRDMA_QPS_RST;
1024 static void ocrdma_set_qp_use_cnt(struct ocrdma_qp *qp, struct ocrdma_pd *pd)
1026 atomic_inc(&pd->use_cnt);
1027 atomic_inc(&qp->sq_cq->use_cnt);
1028 atomic_inc(&qp->rq_cq->use_cnt);
1030 atomic_inc(&qp->srq->use_cnt);
1031 qp->ibqp.qp_num = qp->id;
1034 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1035 struct ib_qp_init_attr *attrs)
1037 if (attrs->qp_type == IB_QPT_GSI) {
1038 dev->gsi_qp_created = 1;
1039 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1040 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1044 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1045 struct ib_qp_init_attr *attrs,
1046 struct ib_udata *udata)
1049 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1050 struct ocrdma_qp *qp;
1051 struct ocrdma_dev *dev = pd->dev;
1052 struct ocrdma_create_qp_ureq ureq;
1053 u16 dpp_credit_lmt, dpp_offset;
1055 status = ocrdma_check_qp_params(ibpd, dev, attrs);
1059 memset(&ureq, 0, sizeof(ureq));
1061 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1062 return ERR_PTR(-EFAULT);
1064 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1070 ocrdma_set_qp_init_params(qp, pd, attrs);
1072 mutex_lock(&dev->dev_lock);
1073 status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1075 &dpp_offset, &dpp_credit_lmt);
1079 /* user space QP's wr_id table are managed in library */
1080 if (udata == NULL) {
1081 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1082 OCRDMA_QP_FAST_REG);
1083 status = ocrdma_alloc_wr_id_tbl(qp);
1088 status = ocrdma_add_qpn_map(dev, qp);
1091 ocrdma_set_qp_db(dev, qp, pd);
1093 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1095 (attrs->srq != NULL));
1099 ocrdma_store_gsi_qp_cq(dev, attrs);
1100 ocrdma_set_qp_use_cnt(qp, pd);
1101 mutex_unlock(&dev->dev_lock);
1105 ocrdma_del_qpn_map(dev, qp);
1107 ocrdma_mbx_destroy_qp(dev, qp);
1109 mutex_unlock(&dev->dev_lock);
1110 kfree(qp->wqe_wr_id_tbl);
1111 kfree(qp->rqe_wr_id_tbl);
1113 ocrdma_err("%s(%d) error=%d\n", __func__, dev->id, status);
1115 return ERR_PTR(status);
1118 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1122 struct ocrdma_qp *qp;
1123 struct ocrdma_dev *dev;
1124 enum ib_qp_state old_qps;
1126 qp = get_ocrdma_qp(ibqp);
1128 if (attr_mask & IB_QP_STATE)
1129 status = ocrdma_qp_state_machine(qp, attr->qp_state, &old_qps);
1130 /* if new and previous states are same hw doesn't need to
1135 status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask, old_qps);
1139 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1140 int attr_mask, struct ib_udata *udata)
1142 unsigned long flags;
1143 int status = -EINVAL;
1144 struct ocrdma_qp *qp;
1145 struct ocrdma_dev *dev;
1146 enum ib_qp_state old_qps, new_qps;
1148 qp = get_ocrdma_qp(ibqp);
1151 /* syncronize with multiple context trying to change, retrive qps */
1152 mutex_lock(&dev->dev_lock);
1153 /* syncronize with wqe, rqe posting and cqe processing contexts */
1154 spin_lock_irqsave(&qp->q_lock, flags);
1155 old_qps = get_ibqp_state(qp->state);
1156 if (attr_mask & IB_QP_STATE)
1157 new_qps = attr->qp_state;
1160 spin_unlock_irqrestore(&qp->q_lock, flags);
1162 if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1163 ocrdma_err("%s(%d) invalid attribute mask=0x%x specified for "
1164 "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1165 __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1170 status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1174 mutex_unlock(&dev->dev_lock);
1178 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1196 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1198 int ib_qp_acc_flags = 0;
1200 if (qp_cap_flags & OCRDMA_QP_INB_WR)
1201 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1202 if (qp_cap_flags & OCRDMA_QP_INB_RD)
1203 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1204 return ib_qp_acc_flags;
1207 int ocrdma_query_qp(struct ib_qp *ibqp,
1208 struct ib_qp_attr *qp_attr,
1209 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1213 struct ocrdma_qp_params params;
1214 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1215 struct ocrdma_dev *dev = qp->dev;
1217 memset(¶ms, 0, sizeof(params));
1218 mutex_lock(&dev->dev_lock);
1219 status = ocrdma_mbx_query_qp(dev, qp, ¶ms);
1220 mutex_unlock(&dev->dev_lock);
1223 qp_attr->qp_state = get_ibqp_state(IB_QPS_INIT);
1224 qp_attr->cur_qp_state = get_ibqp_state(IB_QPS_INIT);
1226 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1227 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1228 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1229 qp_attr->path_mig_state = IB_MIG_MIGRATED;
1230 qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1231 qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1232 qp_attr->dest_qp_num =
1233 params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1235 qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1236 qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1237 qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1238 qp_attr->cap.max_send_sge = qp->sq.max_sges;
1239 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1240 qp_attr->cap.max_inline_data = dev->attr.max_inline_data;
1241 qp_init_attr->cap = qp_attr->cap;
1242 memcpy(&qp_attr->ah_attr.grh.dgid, ¶ms.dgid[0],
1243 sizeof(params.dgid));
1244 qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1245 OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1246 qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1247 qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1248 OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1249 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1250 qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1251 OCRDMA_QP_PARAMS_SQ_PSN_MASK) >>
1252 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1254 qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1255 qp_attr->ah_attr.port_num = 1;
1256 qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1257 OCRDMA_QP_PARAMS_SL_MASK) >>
1258 OCRDMA_QP_PARAMS_SL_SHIFT;
1259 qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1260 OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1261 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1262 qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1263 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1264 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1265 qp_attr->retry_cnt =
1266 (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1267 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1268 qp_attr->min_rnr_timer = 0;
1269 qp_attr->pkey_index = 0;
1270 qp_attr->port_num = 1;
1271 qp_attr->ah_attr.src_path_bits = 0;
1272 qp_attr->ah_attr.static_rate = 0;
1273 qp_attr->alt_pkey_index = 0;
1274 qp_attr->alt_port_num = 0;
1275 qp_attr->alt_timeout = 0;
1276 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1277 qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1278 OCRDMA_QP_PARAMS_STATE_SHIFT;
1279 qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1280 qp_attr->max_dest_rd_atomic =
1281 params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1282 qp_attr->max_rd_atomic =
1283 params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1284 qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1285 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1290 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, int idx)
1293 unsigned int mask = (1 << (idx % 32));
1295 if (srq->idx_bit_fields[i] & mask)
1296 srq->idx_bit_fields[i] &= ~mask;
1298 srq->idx_bit_fields[i] |= mask;
1301 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1304 if (q->head >= q->tail)
1305 free_cnt = (q->max_cnt - q->head) + q->tail;
1307 free_cnt = q->tail - q->head;
1311 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1313 return (qp->sq.tail == qp->sq.head &&
1314 ocrdma_hwq_free_cnt(&qp->sq) ? 1 : 0);
1317 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1319 return (qp->rq.tail == qp->rq.head) ? 1 : 0;
1322 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1324 return q->va + (q->head * q->entry_size);
1327 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1330 return q->va + (idx * q->entry_size);
1333 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1335 q->head = (q->head + 1) & q->max_wqe_idx;
1338 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1340 q->tail = (q->tail + 1) & q->max_wqe_idx;
1343 /* discard the cqe for a given QP */
1344 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1346 unsigned long cq_flags;
1347 unsigned long flags;
1348 int discard_cnt = 0;
1349 u32 cur_getp, stop_getp;
1350 struct ocrdma_cqe *cqe;
1353 spin_lock_irqsave(&cq->cq_lock, cq_flags);
1355 /* traverse through the CQEs in the hw CQ,
1356 * find the matching CQE for a given qp,
1357 * mark the matching one discarded by clearing qpn.
1358 * ring the doorbell in the poll_cq() as
1359 * we don't complete out of order cqe.
1362 cur_getp = cq->getp;
1363 /* find upto when do we reap the cq. */
1364 stop_getp = cur_getp;
1366 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1369 cqe = cq->va + cur_getp;
1370 /* if (a) done reaping whole hw cq, or
1371 * (b) qp_xq becomes empty.
1374 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1375 /* if previously discarded cqe found, skip that too. */
1376 /* check for matching qp */
1377 if (qpn == 0 || qpn != qp->id)
1380 /* mark cqe discarded so that it is not picked up later
1385 if (is_cqe_for_sq(cqe))
1386 ocrdma_hwq_inc_tail(&qp->sq);
1389 spin_lock_irqsave(&qp->srq->q_lock, flags);
1390 ocrdma_hwq_inc_tail(&qp->srq->rq);
1391 ocrdma_srq_toggle_bit(qp->srq, cur_getp);
1392 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1395 ocrdma_hwq_inc_tail(&qp->rq);
1398 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1399 } while (cur_getp != stop_getp);
1400 spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1403 static void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1406 unsigned long flags;
1407 struct ocrdma_dev *dev = qp->dev;
1408 /* sync with any active CQ poll */
1410 spin_lock_irqsave(&dev->flush_q_lock, flags);
1411 found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1413 list_del(&qp->sq_entry);
1415 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1417 list_del(&qp->rq_entry);
1419 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1422 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1425 struct ocrdma_pd *pd;
1426 struct ocrdma_qp *qp;
1427 struct ocrdma_dev *dev;
1428 struct ib_qp_attr attrs;
1429 int attr_mask = IB_QP_STATE;
1430 unsigned long flags;
1432 qp = get_ocrdma_qp(ibqp);
1435 attrs.qp_state = IB_QPS_ERR;
1438 /* change the QP state to ERROR */
1439 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1441 /* ensure that CQEs for newly created QP (whose id may be same with
1442 * one which just getting destroyed are same), dont get
1443 * discarded until the old CQEs are discarded.
1445 mutex_lock(&dev->dev_lock);
1446 status = ocrdma_mbx_destroy_qp(dev, qp);
1449 * acquire CQ lock while destroy is in progress, in order to
1450 * protect against proessing in-flight CQEs for this QP.
1452 spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1453 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1454 spin_lock(&qp->rq_cq->cq_lock);
1456 ocrdma_del_qpn_map(dev, qp);
1458 if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1459 spin_unlock(&qp->rq_cq->cq_lock);
1460 spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1463 ocrdma_discard_cqes(qp, qp->sq_cq);
1464 ocrdma_discard_cqes(qp, qp->rq_cq);
1466 mutex_unlock(&dev->dev_lock);
1469 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa, qp->sq.len);
1471 ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa, qp->rq.len);
1474 ocrdma_del_flush_qp(qp);
1476 atomic_dec(&qp->pd->use_cnt);
1477 atomic_dec(&qp->sq_cq->use_cnt);
1478 atomic_dec(&qp->rq_cq->use_cnt);
1480 atomic_dec(&qp->srq->use_cnt);
1481 kfree(qp->wqe_wr_id_tbl);
1482 kfree(qp->rqe_wr_id_tbl);
1487 static int ocrdma_copy_srq_uresp(struct ocrdma_srq *srq, struct ib_udata *udata)
1490 struct ocrdma_create_srq_uresp uresp;
1492 uresp.rq_dbid = srq->rq.dbid;
1493 uresp.num_rq_pages = 1;
1494 uresp.rq_page_addr[0] = srq->rq.pa;
1495 uresp.rq_page_size = srq->rq.len;
1496 uresp.db_page_addr = srq->dev->nic_info.unmapped_db +
1497 (srq->pd->id * srq->dev->nic_info.db_page_size);
1498 uresp.db_page_size = srq->dev->nic_info.db_page_size;
1499 uresp.num_rqe_allocated = srq->rq.max_cnt;
1500 if (srq->dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
1501 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ1_OFFSET;
1502 uresp.db_shift = 24;
1504 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1505 uresp.db_shift = 16;
1508 status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1511 status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1512 uresp.rq_page_size);
1518 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1519 struct ib_srq_init_attr *init_attr,
1520 struct ib_udata *udata)
1522 int status = -ENOMEM;
1523 struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1524 struct ocrdma_dev *dev = pd->dev;
1525 struct ocrdma_srq *srq;
1527 if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1528 return ERR_PTR(-EINVAL);
1529 if (init_attr->attr.max_wr > dev->attr.max_rqe)
1530 return ERR_PTR(-EINVAL);
1532 srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1534 return ERR_PTR(status);
1536 spin_lock_init(&srq->q_lock);
1539 srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1540 status = ocrdma_mbx_create_srq(srq, init_attr, pd);
1544 if (udata == NULL) {
1545 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1547 if (srq->rqe_wr_id_tbl == NULL)
1550 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1551 (srq->rq.max_cnt % 32 ? 1 : 0);
1552 srq->idx_bit_fields =
1553 kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1554 if (srq->idx_bit_fields == NULL)
1556 memset(srq->idx_bit_fields, 0xff,
1557 srq->bit_fields_len * sizeof(u32));
1560 if (init_attr->attr.srq_limit) {
1561 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1566 atomic_set(&srq->use_cnt, 0);
1568 status = ocrdma_copy_srq_uresp(srq, udata);
1573 atomic_inc(&pd->use_cnt);
1577 ocrdma_mbx_destroy_srq(dev, srq);
1579 kfree(srq->rqe_wr_id_tbl);
1580 kfree(srq->idx_bit_fields);
1582 return ERR_PTR(status);
1585 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1586 struct ib_srq_attr *srq_attr,
1587 enum ib_srq_attr_mask srq_attr_mask,
1588 struct ib_udata *udata)
1591 struct ocrdma_srq *srq;
1593 srq = get_ocrdma_srq(ibsrq);
1594 if (srq_attr_mask & IB_SRQ_MAX_WR)
1597 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1601 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1604 struct ocrdma_srq *srq;
1606 srq = get_ocrdma_srq(ibsrq);
1607 status = ocrdma_mbx_query_srq(srq, srq_attr);
1611 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1614 struct ocrdma_srq *srq;
1615 struct ocrdma_dev *dev;
1617 srq = get_ocrdma_srq(ibsrq);
1619 if (atomic_read(&srq->use_cnt)) {
1620 ocrdma_err("%s(%d) err, srq=0x%x in use\n",
1621 __func__, dev->id, srq->id);
1625 status = ocrdma_mbx_destroy_srq(dev, srq);
1628 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa, srq->rq.len);
1630 atomic_dec(&srq->pd->use_cnt);
1631 kfree(srq->idx_bit_fields);
1632 kfree(srq->rqe_wr_id_tbl);
1637 /* unprivileged verbs and their support functions. */
1638 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1639 struct ocrdma_hdr_wqe *hdr,
1640 struct ib_send_wr *wr)
1642 struct ocrdma_ewqe_ud_hdr *ud_hdr =
1643 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1644 struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
1646 ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
1647 if (qp->qp_type == IB_QPT_GSI)
1648 ud_hdr->qkey = qp->qkey;
1650 ud_hdr->qkey = wr->wr.ud.remote_qkey;
1651 ud_hdr->rsvd_ahid = ah->id;
1654 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1655 struct ocrdma_sge *sge, int num_sge,
1656 struct ib_sge *sg_list)
1660 for (i = 0; i < num_sge; i++) {
1661 sge[i].lrkey = sg_list[i].lkey;
1662 sge[i].addr_lo = sg_list[i].addr;
1663 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1664 sge[i].len = sg_list[i].length;
1665 hdr->total_len += sg_list[i].length;
1668 memset(sge, 0, sizeof(*sge));
1671 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1672 struct ocrdma_hdr_wqe *hdr,
1673 struct ocrdma_sge *sge,
1674 struct ib_send_wr *wr, u32 wqe_size)
1676 if (wr->send_flags & IB_SEND_INLINE) {
1677 if (wr->sg_list[0].length > qp->max_inline_data) {
1678 ocrdma_err("%s() supported_len=0x%x,"
1679 " unspported len req=0x%x\n", __func__,
1680 qp->max_inline_data, wr->sg_list[0].length);
1684 (void *)(unsigned long)wr->sg_list[0].addr,
1685 wr->sg_list[0].length);
1686 hdr->total_len = wr->sg_list[0].length;
1687 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1688 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1690 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1692 wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1694 wqe_size += sizeof(struct ocrdma_sge);
1695 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1697 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1701 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1702 struct ib_send_wr *wr)
1705 struct ocrdma_sge *sge;
1706 u32 wqe_size = sizeof(*hdr);
1708 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1709 ocrdma_build_ud_hdr(qp, hdr, wr);
1710 sge = (struct ocrdma_sge *)(hdr + 2);
1711 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1713 sge = (struct ocrdma_sge *)(hdr + 1);
1715 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1719 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1720 struct ib_send_wr *wr)
1723 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1724 struct ocrdma_sge *sge = ext_rw + 1;
1725 u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1727 status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1730 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1731 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1732 ext_rw->lrkey = wr->wr.rdma.rkey;
1733 ext_rw->len = hdr->total_len;
1737 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1738 struct ib_send_wr *wr)
1740 struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1741 struct ocrdma_sge *sge = ext_rw + 1;
1742 u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
1743 sizeof(struct ocrdma_hdr_wqe);
1745 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1746 hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1747 hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
1748 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1750 ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1751 ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1752 ext_rw->lrkey = wr->wr.rdma.rkey;
1753 ext_rw->len = hdr->total_len;
1756 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
1758 u32 val = qp->sq.dbid | (1 << 16);
1760 iowrite32(val, qp->sq_db);
1763 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1764 struct ib_send_wr **bad_wr)
1767 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1768 struct ocrdma_hdr_wqe *hdr;
1769 unsigned long flags;
1771 spin_lock_irqsave(&qp->q_lock, flags);
1772 if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
1773 spin_unlock_irqrestore(&qp->q_lock, flags);
1778 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
1779 wr->num_sge > qp->sq.max_sges) {
1783 hdr = ocrdma_hwq_head(&qp->sq);
1785 if (wr->send_flags & IB_SEND_SIGNALED)
1786 hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
1787 if (wr->send_flags & IB_SEND_FENCE)
1789 (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
1790 if (wr->send_flags & IB_SEND_SOLICITED)
1792 (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
1794 switch (wr->opcode) {
1795 case IB_WR_SEND_WITH_IMM:
1796 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
1797 hdr->immdt = ntohl(wr->ex.imm_data);
1799 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
1800 ocrdma_build_send(qp, hdr, wr);
1802 case IB_WR_SEND_WITH_INV:
1803 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
1804 hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
1805 hdr->lkey = wr->ex.invalidate_rkey;
1806 status = ocrdma_build_send(qp, hdr, wr);
1808 case IB_WR_RDMA_WRITE_WITH_IMM:
1809 hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
1810 hdr->immdt = ntohl(wr->ex.imm_data);
1811 case IB_WR_RDMA_WRITE:
1812 hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
1813 status = ocrdma_build_write(qp, hdr, wr);
1815 case IB_WR_RDMA_READ_WITH_INV:
1816 hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
1817 case IB_WR_RDMA_READ:
1818 ocrdma_build_read(qp, hdr, wr);
1820 case IB_WR_LOCAL_INV:
1822 (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
1823 hdr->cw |= (sizeof(struct ocrdma_hdr_wqe) /
1824 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
1825 hdr->lkey = wr->ex.invalidate_rkey;
1835 if (wr->send_flags & IB_SEND_SIGNALED)
1836 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
1838 qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
1839 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
1840 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
1841 OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
1842 /* make sure wqe is written before adapter can access it */
1844 /* inform hw to start processing it */
1845 ocrdma_ring_sq_db(qp);
1847 /* update pointer, counter for next wr */
1848 ocrdma_hwq_inc_head(&qp->sq);
1851 spin_unlock_irqrestore(&qp->q_lock, flags);
1855 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
1857 u32 val = qp->rq.dbid | (1 << OCRDMA_GET_NUM_POSTED_SHIFT_VAL(qp));
1859 iowrite32(val, qp->rq_db);
1862 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
1866 struct ocrdma_sge *sge;
1868 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
1870 wqe_size = sizeof(*sge) + sizeof(*rqe);
1872 rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
1873 OCRDMA_WQE_SIZE_SHIFT);
1874 rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
1875 rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1877 rqe->rsvd_tag = tag;
1878 sge = (struct ocrdma_sge *)(rqe + 1);
1879 ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
1880 ocrdma_cpu_to_le32(rqe, wqe_size);
1883 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1884 struct ib_recv_wr **bad_wr)
1887 unsigned long flags;
1888 struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1889 struct ocrdma_hdr_wqe *rqe;
1891 spin_lock_irqsave(&qp->q_lock, flags);
1892 if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
1893 spin_unlock_irqrestore(&qp->q_lock, flags);
1898 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
1899 wr->num_sge > qp->rq.max_sges) {
1904 rqe = ocrdma_hwq_head(&qp->rq);
1905 ocrdma_build_rqe(rqe, wr, 0);
1907 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
1908 /* make sure rqe is written before adapter can access it */
1911 /* inform hw to start processing it */
1912 ocrdma_ring_rq_db(qp);
1914 /* update pointer, counter for next wr */
1915 ocrdma_hwq_inc_head(&qp->rq);
1918 spin_unlock_irqrestore(&qp->q_lock, flags);
1922 /* cqe for srq's rqe can potentially arrive out of order.
1923 * index gives the entry in the shadow table where to store
1924 * the wr_id. tag/index is returned in cqe to reference back
1927 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
1932 for (row = 0; row < srq->bit_fields_len; row++) {
1933 if (srq->idx_bit_fields[row]) {
1934 indx = ffs(srq->idx_bit_fields[row]);
1935 indx = (row * 32) + (indx - 1);
1936 if (indx >= srq->rq.max_cnt)
1938 ocrdma_srq_toggle_bit(srq, indx);
1943 if (row == srq->bit_fields_len)
1948 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
1950 u32 val = srq->rq.dbid | (1 << 16);
1952 iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
1955 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
1956 struct ib_recv_wr **bad_wr)
1959 unsigned long flags;
1960 struct ocrdma_srq *srq;
1961 struct ocrdma_hdr_wqe *rqe;
1964 srq = get_ocrdma_srq(ibsrq);
1966 spin_lock_irqsave(&srq->q_lock, flags);
1968 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
1969 wr->num_sge > srq->rq.max_sges) {
1974 tag = ocrdma_srq_get_idx(srq);
1975 rqe = ocrdma_hwq_head(&srq->rq);
1976 ocrdma_build_rqe(rqe, wr, tag);
1978 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
1979 /* make sure rqe is written before adapter can perform DMA */
1981 /* inform hw to start processing it */
1982 ocrdma_ring_srq_db(srq);
1983 /* update pointer, counter for next wr */
1984 ocrdma_hwq_inc_head(&srq->rq);
1987 spin_unlock_irqrestore(&srq->q_lock, flags);
1991 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
1993 enum ib_wc_status ibwc_status = IB_WC_GENERAL_ERR;
1996 case OCRDMA_CQE_GENERAL_ERR:
1997 ibwc_status = IB_WC_GENERAL_ERR;
1999 case OCRDMA_CQE_LOC_LEN_ERR:
2000 ibwc_status = IB_WC_LOC_LEN_ERR;
2002 case OCRDMA_CQE_LOC_QP_OP_ERR:
2003 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2005 case OCRDMA_CQE_LOC_EEC_OP_ERR:
2006 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2008 case OCRDMA_CQE_LOC_PROT_ERR:
2009 ibwc_status = IB_WC_LOC_PROT_ERR;
2011 case OCRDMA_CQE_WR_FLUSH_ERR:
2012 ibwc_status = IB_WC_WR_FLUSH_ERR;
2014 case OCRDMA_CQE_MW_BIND_ERR:
2015 ibwc_status = IB_WC_MW_BIND_ERR;
2017 case OCRDMA_CQE_BAD_RESP_ERR:
2018 ibwc_status = IB_WC_BAD_RESP_ERR;
2020 case OCRDMA_CQE_LOC_ACCESS_ERR:
2021 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2023 case OCRDMA_CQE_REM_INV_REQ_ERR:
2024 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2026 case OCRDMA_CQE_REM_ACCESS_ERR:
2027 ibwc_status = IB_WC_REM_ACCESS_ERR;
2029 case OCRDMA_CQE_REM_OP_ERR:
2030 ibwc_status = IB_WC_REM_OP_ERR;
2032 case OCRDMA_CQE_RETRY_EXC_ERR:
2033 ibwc_status = IB_WC_RETRY_EXC_ERR;
2035 case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2036 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2038 case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2039 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2041 case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2042 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2044 case OCRDMA_CQE_REM_ABORT_ERR:
2045 ibwc_status = IB_WC_REM_ABORT_ERR;
2047 case OCRDMA_CQE_INV_EECN_ERR:
2048 ibwc_status = IB_WC_INV_EECN_ERR;
2050 case OCRDMA_CQE_INV_EEC_STATE_ERR:
2051 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2053 case OCRDMA_CQE_FATAL_ERR:
2054 ibwc_status = IB_WC_FATAL_ERR;
2056 case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2057 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2060 ibwc_status = IB_WC_GENERAL_ERR;
2066 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2069 struct ocrdma_hdr_wqe *hdr;
2070 struct ocrdma_sge *rw;
2073 hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2075 ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2076 /* Undo the hdr->cw swap */
2077 opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2080 ibwc->opcode = IB_WC_RDMA_WRITE;
2083 rw = (struct ocrdma_sge *)(hdr + 1);
2084 ibwc->opcode = IB_WC_RDMA_READ;
2085 ibwc->byte_len = rw->len;
2088 ibwc->opcode = IB_WC_SEND;
2090 case OCRDMA_LKEY_INV:
2091 ibwc->opcode = IB_WC_LOCAL_INV;
2094 ibwc->status = IB_WC_GENERAL_ERR;
2095 ocrdma_err("%s() invalid opcode received = 0x%x\n",
2096 __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2101 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2102 struct ocrdma_cqe *cqe)
2104 if (is_cqe_for_sq(cqe)) {
2105 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2106 cqe->flags_status_srcqpn) &
2107 ~OCRDMA_CQE_STATUS_MASK);
2108 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2109 cqe->flags_status_srcqpn) |
2110 (OCRDMA_CQE_WR_FLUSH_ERR <<
2111 OCRDMA_CQE_STATUS_SHIFT));
2113 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2114 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2115 cqe->flags_status_srcqpn) &
2116 ~OCRDMA_CQE_UD_STATUS_MASK);
2117 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2118 cqe->flags_status_srcqpn) |
2119 (OCRDMA_CQE_WR_FLUSH_ERR <<
2120 OCRDMA_CQE_UD_STATUS_SHIFT));
2122 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2123 cqe->flags_status_srcqpn) &
2124 ~OCRDMA_CQE_STATUS_MASK);
2125 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2126 cqe->flags_status_srcqpn) |
2127 (OCRDMA_CQE_WR_FLUSH_ERR <<
2128 OCRDMA_CQE_STATUS_SHIFT));
2133 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2134 struct ocrdma_qp *qp, int status)
2136 bool expand = false;
2139 ibwc->qp = &qp->ibqp;
2140 ibwc->status = ocrdma_to_ibwc_err(status);
2142 ocrdma_flush_qp(qp);
2143 ocrdma_qp_state_machine(qp, IB_QPS_ERR, NULL);
2145 /* if wqe/rqe pending for which cqe needs to be returned,
2146 * trigger inflating it.
2148 if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2150 ocrdma_set_cqe_status_flushed(qp, cqe);
2155 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2156 struct ocrdma_qp *qp, int status)
2158 ibwc->opcode = IB_WC_RECV;
2159 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2160 ocrdma_hwq_inc_tail(&qp->rq);
2162 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2165 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2166 struct ocrdma_qp *qp, int status)
2168 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2169 ocrdma_hwq_inc_tail(&qp->sq);
2171 return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2175 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2176 struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2177 bool *polled, bool *stop)
2180 int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2181 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2183 /* when hw sq is empty, but rq is not empty, so we continue
2184 * to keep the cqe in order to get the cq event again.
2186 if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2187 /* when cq for rq and sq is same, it is safe to return
2188 * flush cqe for RQEs.
2190 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2192 status = OCRDMA_CQE_WR_FLUSH_ERR;
2193 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2195 /* stop processing further cqe as this cqe is used for
2196 * triggering cq event on buddy cq of RQ.
2197 * When QP is destroyed, this cqe will be removed
2198 * from the cq's hardware q.
2206 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2211 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2212 struct ocrdma_cqe *cqe,
2213 struct ib_wc *ibwc, bool *polled)
2215 bool expand = false;
2216 int tail = qp->sq.tail;
2219 if (!qp->wqe_wr_id_tbl[tail].signaled) {
2220 expand = true; /* CQE cannot be consumed yet */
2221 *polled = false; /* WC cannot be consumed yet */
2223 ibwc->status = IB_WC_SUCCESS;
2225 ibwc->qp = &qp->ibqp;
2226 ocrdma_update_wc(qp, ibwc, tail);
2228 wqe_idx = le32_to_cpu(cqe->wq.wqeidx) & OCRDMA_CQE_WQEIDX_MASK;
2229 if (tail != wqe_idx)
2230 expand = true; /* Coalesced CQE can't be consumed yet */
2232 ocrdma_hwq_inc_tail(&qp->sq);
2236 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2237 struct ib_wc *ibwc, bool *polled, bool *stop)
2242 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2243 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2245 if (status == OCRDMA_CQE_SUCCESS)
2246 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2248 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2252 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2256 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2257 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2258 ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2259 OCRDMA_CQE_SRCQP_MASK;
2260 ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2261 OCRDMA_CQE_PKEY_MASK;
2262 ibwc->wc_flags = IB_WC_GRH;
2263 ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2264 OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2268 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2269 struct ocrdma_cqe *cqe,
2270 struct ocrdma_qp *qp)
2272 unsigned long flags;
2273 struct ocrdma_srq *srq;
2276 srq = get_ocrdma_srq(qp->ibqp.srq);
2277 wqe_idx = le32_to_cpu(cqe->rq.buftag_qpn) >> OCRDMA_CQE_BUFTAG_SHIFT;
2278 ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2279 spin_lock_irqsave(&srq->q_lock, flags);
2280 ocrdma_srq_toggle_bit(srq, wqe_idx);
2281 spin_unlock_irqrestore(&srq->q_lock, flags);
2282 ocrdma_hwq_inc_tail(&srq->rq);
2285 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2286 struct ib_wc *ibwc, bool *polled, bool *stop,
2291 /* when hw_rq is empty, but wq is not empty, so continue
2292 * to keep the cqe to get the cq event again.
2294 if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2295 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2297 status = OCRDMA_CQE_WR_FLUSH_ERR;
2298 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2306 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2311 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2312 struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2314 ibwc->opcode = IB_WC_RECV;
2315 ibwc->qp = &qp->ibqp;
2316 ibwc->status = IB_WC_SUCCESS;
2318 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2319 ocrdma_update_ud_rcqe(ibwc, cqe);
2321 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2323 if (is_cqe_imm(cqe)) {
2324 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2325 ibwc->wc_flags |= IB_WC_WITH_IMM;
2326 } else if (is_cqe_wr_imm(cqe)) {
2327 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2328 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2329 ibwc->wc_flags |= IB_WC_WITH_IMM;
2330 } else if (is_cqe_invalidated(cqe)) {
2331 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2332 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2335 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2337 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2338 ocrdma_hwq_inc_tail(&qp->rq);
2342 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2343 struct ib_wc *ibwc, bool *polled, bool *stop)
2346 bool expand = false;
2349 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2350 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2351 OCRDMA_CQE_UD_STATUS_MASK) >>
2352 OCRDMA_CQE_UD_STATUS_SHIFT;
2354 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2355 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2357 if (status == OCRDMA_CQE_SUCCESS) {
2359 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2361 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2367 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2370 if (cq->phase_change) {
2372 cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2374 /* clear valid bit */
2375 cqe->flags_status_srcqpn = 0;
2378 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2383 bool expand = false;
2384 int polled_hw_cqes = 0;
2385 struct ocrdma_qp *qp = NULL;
2386 struct ocrdma_dev *dev = cq->dev;
2387 struct ocrdma_cqe *cqe;
2388 u16 cur_getp; bool polled = false; bool stop = false;
2390 cur_getp = cq->getp;
2391 while (num_entries) {
2392 cqe = cq->va + cur_getp;
2393 /* check whether valid cqe or not */
2394 if (!is_cqe_valid(cq, cqe))
2396 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2397 /* ignore discarded cqe */
2400 qp = dev->qp_tbl[qpn];
2403 if (is_cqe_for_sq(cqe)) {
2404 expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2407 expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2414 /* clear qpn to avoid duplicate processing by discard_cqe() */
2417 polled_hw_cqes += 1;
2418 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2419 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2429 cq->getp = cur_getp;
2430 if (polled_hw_cqes || expand || stop) {
2431 ocrdma_ring_cq_db(dev, cq->id, cq->armed, cq->solicited,
2437 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2438 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2439 struct ocrdma_qp *qp, struct ib_wc *ibwc)
2443 while (num_entries) {
2444 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2446 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2447 ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2448 ocrdma_hwq_inc_tail(&qp->sq);
2449 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2450 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2451 ocrdma_hwq_inc_tail(&qp->rq);
2455 ibwc->status = IB_WC_WR_FLUSH_ERR;
2463 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2465 int cqes_to_poll = num_entries;
2466 struct ocrdma_cq *cq = NULL;
2467 unsigned long flags;
2468 struct ocrdma_dev *dev;
2469 int num_os_cqe = 0, err_cqes = 0;
2470 struct ocrdma_qp *qp;
2472 cq = get_ocrdma_cq(ibcq);
2475 /* poll cqes from adapter CQ */
2476 spin_lock_irqsave(&cq->cq_lock, flags);
2477 num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2478 spin_unlock_irqrestore(&cq->cq_lock, flags);
2479 cqes_to_poll -= num_os_cqe;
2482 wc = wc + num_os_cqe;
2483 /* adapter returns single error cqe when qp moves to
2484 * error state. So insert error cqes with wc_status as
2485 * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2486 * respectively which uses this CQ.
2488 spin_lock_irqsave(&dev->flush_q_lock, flags);
2489 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2490 if (cqes_to_poll == 0)
2492 err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2493 cqes_to_poll -= err_cqes;
2494 num_os_cqe += err_cqes;
2497 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2502 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2504 struct ocrdma_cq *cq;
2505 unsigned long flags;
2506 struct ocrdma_dev *dev;
2509 struct ocrdma_cqe *cqe;
2511 cq = get_ocrdma_cq(ibcq);
2515 spin_lock_irqsave(&cq->cq_lock, flags);
2516 if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2518 if (cq_flags & IB_CQ_SOLICITED)
2519 cq->solicited = true;
2521 cur_getp = cq->getp;
2522 cqe = cq->va + cur_getp;
2524 /* check whether any valid cqe exist or not, if not then safe to
2525 * arm. If cqe is not yet consumed, then let it get consumed and then
2526 * we arm it to avoid false interrupts.
2528 if (!is_cqe_valid(cq, cqe) || cq->arm_needed) {
2529 cq->arm_needed = false;
2530 ocrdma_ring_cq_db(dev, cq_id, cq->armed, cq->solicited, 0);
2532 spin_unlock_irqrestore(&cq->cq_lock, flags);