1 /* QLogic qedr NIC Driver
2 * Copyright (c) 2015-2016 QLogic Corporation
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and /or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 #include <linux/dma-mapping.h>
33 #include <linux/crc32.h>
37 #include <linux/iommu.h>
39 #include <rdma/ib_verbs.h>
40 #include <rdma/ib_user_verbs.h>
41 #include <rdma/iw_cm.h>
42 #include <rdma/ib_umem.h>
43 #include <rdma/ib_addr.h>
44 #include <rdma/ib_cache.h>
47 #include <linux/qed/qed_if.h>
50 #include <rdma/qedr-abi.h>
53 #define DB_ADDR_SHIFT(addr) ((addr) << DB_PWM_ADDR_OFFSET_SHIFT)
55 int qedr_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
57 if (index > QEDR_ROCE_PKEY_TABLE_LEN)
60 *pkey = QEDR_ROCE_PKEY_DEFAULT;
64 int qedr_query_gid(struct ib_device *ibdev, u8 port, int index,
67 struct qedr_dev *dev = get_qedr_dev(ibdev);
70 if (!rdma_cap_roce_gid_table(ibdev, port))
73 rc = ib_get_cached_gid(ibdev, port, index, sgid, NULL);
75 memcpy(sgid, &zgid, sizeof(*sgid));
79 DP_DEBUG(dev, QEDR_MSG_INIT, "query gid: index=%d %llx:%llx\n", index,
80 sgid->global.interface_id, sgid->global.subnet_prefix);
85 int qedr_add_gid(struct ib_device *device, u8 port_num,
86 unsigned int index, const union ib_gid *gid,
87 const struct ib_gid_attr *attr, void **context)
89 if (!rdma_cap_roce_gid_table(device, port_num))
92 if (port_num > QEDR_MAX_PORT)
101 int qedr_del_gid(struct ib_device *device, u8 port_num,
102 unsigned int index, void **context)
104 if (!rdma_cap_roce_gid_table(device, port_num))
107 if (port_num > QEDR_MAX_PORT)
116 int qedr_query_device(struct ib_device *ibdev,
117 struct ib_device_attr *attr, struct ib_udata *udata)
119 struct qedr_dev *dev = get_qedr_dev(ibdev);
120 struct qedr_device_attr *qattr = &dev->attr;
122 if (!dev->rdma_ctx) {
124 "qedr_query_device called with invalid params rdma_ctx=%p\n",
129 memset(attr, 0, sizeof(*attr));
131 attr->fw_ver = qattr->fw_ver;
132 attr->sys_image_guid = qattr->sys_image_guid;
133 attr->max_mr_size = qattr->max_mr_size;
134 attr->page_size_cap = qattr->page_size_caps;
135 attr->vendor_id = qattr->vendor_id;
136 attr->vendor_part_id = qattr->vendor_part_id;
137 attr->hw_ver = qattr->hw_ver;
138 attr->max_qp = qattr->max_qp;
139 attr->max_qp_wr = max_t(u32, qattr->max_sqe, qattr->max_rqe);
140 attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
141 IB_DEVICE_RC_RNR_NAK_GEN |
142 IB_DEVICE_LOCAL_DMA_LKEY | IB_DEVICE_MEM_MGT_EXTENSIONS;
144 attr->max_sge = qattr->max_sge;
145 attr->max_sge_rd = qattr->max_sge;
146 attr->max_cq = qattr->max_cq;
147 attr->max_cqe = qattr->max_cqe;
148 attr->max_mr = qattr->max_mr;
149 attr->max_mw = qattr->max_mw;
150 attr->max_pd = qattr->max_pd;
151 attr->atomic_cap = dev->atomic_cap;
152 attr->max_fmr = qattr->max_fmr;
153 attr->max_map_per_fmr = 16;
154 attr->max_qp_init_rd_atom =
155 1 << (fls(qattr->max_qp_req_rd_atomic_resc) - 1);
156 attr->max_qp_rd_atom =
157 min(1 << (fls(qattr->max_qp_resp_rd_atomic_resc) - 1),
158 attr->max_qp_init_rd_atom);
160 attr->max_srq = qattr->max_srq;
161 attr->max_srq_sge = qattr->max_srq_sge;
162 attr->max_srq_wr = qattr->max_srq_wr;
164 attr->local_ca_ack_delay = qattr->dev_ack_delay;
165 attr->max_fast_reg_page_list_len = qattr->max_mr / 8;
166 attr->max_pkeys = QEDR_ROCE_PKEY_MAX;
167 attr->max_ah = qattr->max_ah;
172 #define QEDR_SPEED_SDR (1)
173 #define QEDR_SPEED_DDR (2)
174 #define QEDR_SPEED_QDR (4)
175 #define QEDR_SPEED_FDR10 (8)
176 #define QEDR_SPEED_FDR (16)
177 #define QEDR_SPEED_EDR (32)
179 static inline void get_link_speed_and_width(int speed, u8 *ib_speed,
184 *ib_speed = QEDR_SPEED_SDR;
185 *ib_width = IB_WIDTH_1X;
188 *ib_speed = QEDR_SPEED_QDR;
189 *ib_width = IB_WIDTH_1X;
193 *ib_speed = QEDR_SPEED_DDR;
194 *ib_width = IB_WIDTH_4X;
198 *ib_speed = QEDR_SPEED_EDR;
199 *ib_width = IB_WIDTH_1X;
203 *ib_speed = QEDR_SPEED_QDR;
204 *ib_width = IB_WIDTH_4X;
208 *ib_speed = QEDR_SPEED_QDR;
209 *ib_width = IB_WIDTH_4X;
213 *ib_speed = QEDR_SPEED_EDR;
214 *ib_width = IB_WIDTH_4X;
219 *ib_speed = QEDR_SPEED_SDR;
220 *ib_width = IB_WIDTH_1X;
224 int qedr_query_port(struct ib_device *ibdev, u8 port, struct ib_port_attr *attr)
226 struct qedr_dev *dev;
227 struct qed_rdma_port *rdma_port;
229 dev = get_qedr_dev(ibdev);
231 DP_ERR(dev, "invalid_port=0x%x\n", port);
235 if (!dev->rdma_ctx) {
236 DP_ERR(dev, "rdma_ctx is NULL\n");
240 rdma_port = dev->ops->rdma_query_port(dev->rdma_ctx);
241 memset(attr, 0, sizeof(*attr));
243 if (rdma_port->port_state == QED_RDMA_PORT_UP) {
244 attr->state = IB_PORT_ACTIVE;
245 attr->phys_state = 5;
247 attr->state = IB_PORT_DOWN;
248 attr->phys_state = 3;
250 attr->max_mtu = IB_MTU_4096;
251 attr->active_mtu = iboe_get_mtu(dev->ndev->mtu);
256 attr->port_cap_flags = IB_PORT_IP_BASED_GIDS;
257 attr->gid_tbl_len = QEDR_MAX_SGID;
258 attr->pkey_tbl_len = QEDR_ROCE_PKEY_TABLE_LEN;
259 attr->bad_pkey_cntr = rdma_port->pkey_bad_counter;
260 attr->qkey_viol_cntr = 0;
261 get_link_speed_and_width(rdma_port->link_speed,
262 &attr->active_speed, &attr->active_width);
263 attr->max_msg_sz = rdma_port->max_msg_size;
264 attr->max_vl_num = 4;
269 int qedr_modify_port(struct ib_device *ibdev, u8 port, int mask,
270 struct ib_port_modify *props)
272 struct qedr_dev *dev;
274 dev = get_qedr_dev(ibdev);
276 DP_ERR(dev, "invalid_port=0x%x\n", port);
283 static int qedr_add_mmap(struct qedr_ucontext *uctx, u64 phy_addr,
288 mm = kzalloc(sizeof(*mm), GFP_KERNEL);
292 mm->key.phy_addr = phy_addr;
293 /* This function might be called with a length which is not a multiple
294 * of PAGE_SIZE, while the mapping is PAGE_SIZE grained and the kernel
295 * forces this granularity by increasing the requested size if needed.
296 * When qedr_mmap is called, it will search the list with the updated
297 * length as a key. To prevent search failures, the length is rounded up
298 * in advance to PAGE_SIZE.
300 mm->key.len = roundup(len, PAGE_SIZE);
301 INIT_LIST_HEAD(&mm->entry);
303 mutex_lock(&uctx->mm_list_lock);
304 list_add(&mm->entry, &uctx->mm_head);
305 mutex_unlock(&uctx->mm_list_lock);
307 DP_DEBUG(uctx->dev, QEDR_MSG_MISC,
308 "added (addr=0x%llx,len=0x%lx) for ctx=%p\n",
309 (unsigned long long)mm->key.phy_addr,
310 (unsigned long)mm->key.len, uctx);
315 static bool qedr_search_mmap(struct qedr_ucontext *uctx, u64 phy_addr,
321 mutex_lock(&uctx->mm_list_lock);
322 list_for_each_entry(mm, &uctx->mm_head, entry) {
323 if (len != mm->key.len || phy_addr != mm->key.phy_addr)
329 mutex_unlock(&uctx->mm_list_lock);
330 DP_DEBUG(uctx->dev, QEDR_MSG_MISC,
331 "searched for (addr=0x%llx,len=0x%lx) for ctx=%p, result=%d\n",
332 mm->key.phy_addr, mm->key.len, uctx, found);
337 struct ib_ucontext *qedr_alloc_ucontext(struct ib_device *ibdev,
338 struct ib_udata *udata)
341 struct qedr_ucontext *ctx;
342 struct qedr_alloc_ucontext_resp uresp;
343 struct qedr_dev *dev = get_qedr_dev(ibdev);
344 struct qed_rdma_add_user_out_params oparams;
347 return ERR_PTR(-EFAULT);
349 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
351 return ERR_PTR(-ENOMEM);
353 rc = dev->ops->rdma_add_user(dev->rdma_ctx, &oparams);
356 "failed to allocate a DPI for a new RoCE application, rc=%d. To overcome this consider to increase the number of DPIs, increase the doorbell BAR size or just close unnecessary RoCE applications. In order to increase the number of DPIs consult the qedr readme\n",
361 ctx->dpi = oparams.dpi;
362 ctx->dpi_addr = oparams.dpi_addr;
363 ctx->dpi_phys_addr = oparams.dpi_phys_addr;
364 ctx->dpi_size = oparams.dpi_size;
365 INIT_LIST_HEAD(&ctx->mm_head);
366 mutex_init(&ctx->mm_list_lock);
368 memset(&uresp, 0, sizeof(uresp));
370 uresp.db_pa = ctx->dpi_phys_addr;
371 uresp.db_size = ctx->dpi_size;
372 uresp.max_send_wr = dev->attr.max_sqe;
373 uresp.max_recv_wr = dev->attr.max_rqe;
374 uresp.max_srq_wr = dev->attr.max_srq_wr;
375 uresp.sges_per_send_wr = QEDR_MAX_SQE_ELEMENTS_PER_SQE;
376 uresp.sges_per_recv_wr = QEDR_MAX_RQE_ELEMENTS_PER_RQE;
377 uresp.sges_per_srq_wr = dev->attr.max_srq_sge;
378 uresp.max_cqes = QEDR_MAX_CQES;
380 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
386 rc = qedr_add_mmap(ctx, ctx->dpi_phys_addr, ctx->dpi_size);
390 DP_DEBUG(dev, QEDR_MSG_INIT, "Allocating user context %p\n",
392 return &ctx->ibucontext;
399 int qedr_dealloc_ucontext(struct ib_ucontext *ibctx)
401 struct qedr_ucontext *uctx = get_qedr_ucontext(ibctx);
402 struct qedr_mm *mm, *tmp;
405 DP_DEBUG(uctx->dev, QEDR_MSG_INIT, "Deallocating user context %p\n",
407 uctx->dev->ops->rdma_remove_user(uctx->dev->rdma_ctx, uctx->dpi);
409 list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
410 DP_DEBUG(uctx->dev, QEDR_MSG_MISC,
411 "deleted (addr=0x%llx,len=0x%lx) for ctx=%p\n",
412 mm->key.phy_addr, mm->key.len, uctx);
413 list_del(&mm->entry);
421 int qedr_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
423 struct qedr_ucontext *ucontext = get_qedr_ucontext(context);
424 struct qedr_dev *dev = get_qedr_dev(context->device);
425 unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
426 u64 unmapped_db = dev->db_phys_addr;
427 unsigned long len = (vma->vm_end - vma->vm_start);
431 DP_DEBUG(dev, QEDR_MSG_INIT,
432 "qedr_mmap called vm_page=0x%lx vm_pgoff=0x%lx unmapped_db=0x%llx db_size=%x, len=%lx\n",
433 vm_page, vma->vm_pgoff, unmapped_db, dev->db_size, len);
434 if (vma->vm_start & (PAGE_SIZE - 1)) {
435 DP_ERR(dev, "Vma_start not page aligned = %ld\n",
440 found = qedr_search_mmap(ucontext, vm_page, len);
442 DP_ERR(dev, "Vma_pgoff not found in mapped array = %ld\n",
447 DP_DEBUG(dev, QEDR_MSG_INIT, "Mapping doorbell bar\n");
449 if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
451 DP_DEBUG(dev, QEDR_MSG_INIT, "Mapping doorbell bar\n");
452 if (vma->vm_flags & VM_READ) {
453 DP_ERR(dev, "Trying to map doorbell bar for read\n");
457 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
459 rc = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
460 PAGE_SIZE, vma->vm_page_prot);
462 DP_DEBUG(dev, QEDR_MSG_INIT, "Mapping chains\n");
463 rc = remap_pfn_range(vma, vma->vm_start,
464 vma->vm_pgoff, len, vma->vm_page_prot);
466 DP_DEBUG(dev, QEDR_MSG_INIT, "qedr_mmap return code: %d\n", rc);
470 struct ib_pd *qedr_alloc_pd(struct ib_device *ibdev,
471 struct ib_ucontext *context, struct ib_udata *udata)
473 struct qedr_dev *dev = get_qedr_dev(ibdev);
474 struct qedr_ucontext *uctx = NULL;
475 struct qedr_alloc_pd_uresp uresp;
480 DP_DEBUG(dev, QEDR_MSG_INIT, "Function called from: %s\n",
481 (udata && context) ? "User Lib" : "Kernel");
483 if (!dev->rdma_ctx) {
484 DP_ERR(dev, "invlaid RDMA context\n");
485 return ERR_PTR(-EINVAL);
488 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
490 return ERR_PTR(-ENOMEM);
492 dev->ops->rdma_alloc_pd(dev->rdma_ctx, &pd_id);
497 if (udata && context) {
498 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
500 DP_ERR(dev, "copy error pd_id=0x%x.\n", pd_id);
501 uctx = get_qedr_ucontext(context);
509 int qedr_dealloc_pd(struct ib_pd *ibpd)
511 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
512 struct qedr_pd *pd = get_qedr_pd(ibpd);
515 pr_err("Invalid PD received in dealloc_pd\n");
519 DP_DEBUG(dev, QEDR_MSG_INIT, "Deallocating PD %d\n", pd->pd_id);
520 dev->ops->rdma_dealloc_pd(dev->rdma_ctx, pd->pd_id);
527 static void qedr_free_pbl(struct qedr_dev *dev,
528 struct qedr_pbl_info *pbl_info, struct qedr_pbl *pbl)
530 struct pci_dev *pdev = dev->pdev;
533 for (i = 0; i < pbl_info->num_pbls; i++) {
536 dma_free_coherent(&pdev->dev, pbl_info->pbl_size,
537 pbl[i].va, pbl[i].pa);
543 #define MIN_FW_PBL_PAGE_SIZE (4 * 1024)
544 #define MAX_FW_PBL_PAGE_SIZE (64 * 1024)
546 #define NUM_PBES_ON_PAGE(_page_size) (_page_size / sizeof(u64))
547 #define MAX_PBES_ON_PAGE NUM_PBES_ON_PAGE(MAX_FW_PBL_PAGE_SIZE)
548 #define MAX_PBES_TWO_LAYER (MAX_PBES_ON_PAGE * MAX_PBES_ON_PAGE)
550 static struct qedr_pbl *qedr_alloc_pbl_tbl(struct qedr_dev *dev,
551 struct qedr_pbl_info *pbl_info,
554 struct pci_dev *pdev = dev->pdev;
555 struct qedr_pbl *pbl_table;
556 dma_addr_t *pbl_main_tbl;
561 pbl_table = kcalloc(pbl_info->num_pbls, sizeof(*pbl_table), flags);
563 return ERR_PTR(-ENOMEM);
565 for (i = 0; i < pbl_info->num_pbls; i++) {
566 va = dma_alloc_coherent(&pdev->dev, pbl_info->pbl_size,
571 memset(va, 0, pbl_info->pbl_size);
572 pbl_table[i].va = va;
573 pbl_table[i].pa = pa;
576 /* Two-Layer PBLs, if we have more than one pbl we need to initialize
577 * the first one with physical pointers to all of the rest
579 pbl_main_tbl = (dma_addr_t *)pbl_table[0].va;
580 for (i = 0; i < pbl_info->num_pbls - 1; i++)
581 pbl_main_tbl[i] = pbl_table[i + 1].pa;
586 for (i--; i >= 0; i--)
587 dma_free_coherent(&pdev->dev, pbl_info->pbl_size,
588 pbl_table[i].va, pbl_table[i].pa);
590 qedr_free_pbl(dev, pbl_info, pbl_table);
592 return ERR_PTR(-ENOMEM);
595 static int qedr_prepare_pbl_tbl(struct qedr_dev *dev,
596 struct qedr_pbl_info *pbl_info,
597 u32 num_pbes, int two_layer_capable)
603 if ((num_pbes > MAX_PBES_ON_PAGE) && two_layer_capable) {
604 if (num_pbes > MAX_PBES_TWO_LAYER) {
605 DP_ERR(dev, "prepare pbl table: too many pages %d\n",
610 /* calculate required pbl page size */
611 pbl_size = MIN_FW_PBL_PAGE_SIZE;
612 pbl_capacity = NUM_PBES_ON_PAGE(pbl_size) *
613 NUM_PBES_ON_PAGE(pbl_size);
615 while (pbl_capacity < num_pbes) {
617 pbl_capacity = pbl_size / sizeof(u64);
618 pbl_capacity = pbl_capacity * pbl_capacity;
621 num_pbls = DIV_ROUND_UP(num_pbes, NUM_PBES_ON_PAGE(pbl_size));
622 num_pbls++; /* One for the layer0 ( points to the pbls) */
623 pbl_info->two_layered = true;
625 /* One layered PBL */
627 pbl_size = max_t(u32, MIN_FW_PBL_PAGE_SIZE,
628 roundup_pow_of_two((num_pbes * sizeof(u64))));
629 pbl_info->two_layered = false;
632 pbl_info->num_pbls = num_pbls;
633 pbl_info->pbl_size = pbl_size;
634 pbl_info->num_pbes = num_pbes;
636 DP_DEBUG(dev, QEDR_MSG_MR,
637 "prepare pbl table: num_pbes=%d, num_pbls=%d, pbl_size=%d\n",
638 pbl_info->num_pbes, pbl_info->num_pbls, pbl_info->pbl_size);
643 static void qedr_populate_pbls(struct qedr_dev *dev, struct ib_umem *umem,
644 struct qedr_pbl *pbl,
645 struct qedr_pbl_info *pbl_info)
647 int shift, pg_cnt, pages, pbe_cnt, total_num_pbes = 0;
648 struct qedr_pbl *pbl_tbl;
649 struct scatterlist *sg;
654 if (!pbl_info->num_pbes)
657 /* If we have a two layered pbl, the first pbl points to the rest
658 * of the pbls and the first entry lays on the second pbl in the table
660 if (pbl_info->two_layered)
665 pbe = (struct regpair *)pbl_tbl->va;
667 DP_ERR(dev, "cannot populate PBL due to a NULL PBE\n");
673 shift = ilog2(umem->page_size);
675 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
676 pages = sg_dma_len(sg) >> shift;
677 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
678 /* store the page address in pbe */
679 pbe->lo = cpu_to_le32(sg_dma_address(sg) +
680 umem->page_size * pg_cnt);
681 addr = upper_32_bits(sg_dma_address(sg) +
682 umem->page_size * pg_cnt);
683 pbe->hi = cpu_to_le32(addr);
688 if (total_num_pbes == pbl_info->num_pbes)
691 /* If the given pbl is full storing the pbes,
694 if (pbe_cnt == (pbl_info->pbl_size / sizeof(u64))) {
696 pbe = (struct regpair *)pbl_tbl->va;
703 static int qedr_copy_cq_uresp(struct qedr_dev *dev,
704 struct qedr_cq *cq, struct ib_udata *udata)
706 struct qedr_create_cq_uresp uresp;
709 memset(&uresp, 0, sizeof(uresp));
711 uresp.db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT);
712 uresp.icid = cq->icid;
714 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
716 DP_ERR(dev, "copy error cqid=0x%x.\n", cq->icid);
721 static void consume_cqe(struct qedr_cq *cq)
723 if (cq->latest_cqe == cq->toggle_cqe)
724 cq->pbl_toggle ^= RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK;
726 cq->latest_cqe = qed_chain_consume(&cq->pbl);
729 static inline int qedr_align_cq_entries(int entries)
731 u64 size, aligned_size;
733 /* We allocate an extra entry that we don't report to the FW. */
734 size = (entries + 1) * QEDR_CQE_SIZE;
735 aligned_size = ALIGN(size, PAGE_SIZE);
737 return aligned_size / QEDR_CQE_SIZE;
740 static inline int qedr_init_user_queue(struct ib_ucontext *ib_ctx,
741 struct qedr_dev *dev,
742 struct qedr_userq *q,
743 u64 buf_addr, size_t buf_len,
744 int access, int dmasync)
749 q->buf_addr = buf_addr;
750 q->buf_len = buf_len;
751 q->umem = ib_umem_get(ib_ctx, q->buf_addr, q->buf_len, access, dmasync);
752 if (IS_ERR(q->umem)) {
753 DP_ERR(dev, "create user queue: failed ib_umem_get, got %ld\n",
755 return PTR_ERR(q->umem);
758 page_cnt = ib_umem_page_count(q->umem);
759 rc = qedr_prepare_pbl_tbl(dev, &q->pbl_info, page_cnt, 0);
763 q->pbl_tbl = qedr_alloc_pbl_tbl(dev, &q->pbl_info, GFP_KERNEL);
764 if (IS_ERR_OR_NULL(q->pbl_tbl))
767 qedr_populate_pbls(dev, q->umem, q->pbl_tbl, &q->pbl_info);
772 ib_umem_release(q->umem);
777 static inline void qedr_init_cq_params(struct qedr_cq *cq,
778 struct qedr_ucontext *ctx,
779 struct qedr_dev *dev, int vector,
780 int chain_entries, int page_cnt,
782 struct qed_rdma_create_cq_in_params
785 memset(params, 0, sizeof(*params));
786 params->cq_handle_hi = upper_32_bits((uintptr_t)cq);
787 params->cq_handle_lo = lower_32_bits((uintptr_t)cq);
788 params->cnq_id = vector;
789 params->cq_size = chain_entries - 1;
790 params->dpi = (ctx) ? ctx->dpi : dev->dpi;
791 params->pbl_num_pages = page_cnt;
792 params->pbl_ptr = pbl_ptr;
793 params->pbl_two_level = 0;
796 static void doorbell_cq(struct qedr_cq *cq, u32 cons, u8 flags)
798 /* Flush data before signalling doorbell */
800 cq->db.data.agg_flags = flags;
801 cq->db.data.value = cpu_to_le32(cons);
802 writeq(cq->db.raw, cq->db_addr);
804 /* Make sure write would stick */
808 int qedr_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
810 struct qedr_cq *cq = get_qedr_cq(ibcq);
811 unsigned long sflags;
813 if (cq->cq_type == QEDR_CQ_TYPE_GSI)
816 spin_lock_irqsave(&cq->cq_lock, sflags);
820 if (flags & IB_CQ_SOLICITED)
821 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_SE_CF_CMD;
823 if (flags & IB_CQ_NEXT_COMP)
824 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_CF_CMD;
826 doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags);
828 spin_unlock_irqrestore(&cq->cq_lock, sflags);
833 struct ib_cq *qedr_create_cq(struct ib_device *ibdev,
834 const struct ib_cq_init_attr *attr,
835 struct ib_ucontext *ib_ctx, struct ib_udata *udata)
837 struct qedr_ucontext *ctx = get_qedr_ucontext(ib_ctx);
838 struct qed_rdma_destroy_cq_out_params destroy_oparams;
839 struct qed_rdma_destroy_cq_in_params destroy_iparams;
840 struct qedr_dev *dev = get_qedr_dev(ibdev);
841 struct qed_rdma_create_cq_in_params params;
842 struct qedr_create_cq_ureq ureq;
843 int vector = attr->comp_vector;
844 int entries = attr->cqe;
852 DP_DEBUG(dev, QEDR_MSG_INIT,
853 "create_cq: called from %s. entries=%d, vector=%d\n",
854 udata ? "User Lib" : "Kernel", entries, vector);
856 if (entries > QEDR_MAX_CQES) {
858 "create cq: the number of entries %d is too high. Must be equal or below %d.\n",
859 entries, QEDR_MAX_CQES);
860 return ERR_PTR(-EINVAL);
863 chain_entries = qedr_align_cq_entries(entries);
864 chain_entries = min_t(int, chain_entries, QEDR_MAX_CQES);
866 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
868 return ERR_PTR(-ENOMEM);
871 memset(&ureq, 0, sizeof(ureq));
872 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) {
874 "create cq: problem copying data from user space\n");
880 "create cq: cannot create a cq with 0 entries\n");
884 cq->cq_type = QEDR_CQ_TYPE_USER;
886 rc = qedr_init_user_queue(ib_ctx, dev, &cq->q, ureq.addr,
887 ureq.len, IB_ACCESS_LOCAL_WRITE, 1);
891 pbl_ptr = cq->q.pbl_tbl->pa;
892 page_cnt = cq->q.pbl_info.num_pbes;
894 cq->ibcq.cqe = chain_entries;
896 cq->cq_type = QEDR_CQ_TYPE_KERNEL;
898 rc = dev->ops->common->chain_alloc(dev->cdev,
899 QED_CHAIN_USE_TO_CONSUME,
901 QED_CHAIN_CNT_TYPE_U32,
903 sizeof(union rdma_cqe),
908 page_cnt = qed_chain_get_page_cnt(&cq->pbl);
909 pbl_ptr = qed_chain_get_pbl_phys(&cq->pbl);
910 cq->ibcq.cqe = cq->pbl.capacity;
913 qedr_init_cq_params(cq, ctx, dev, vector, chain_entries, page_cnt,
916 rc = dev->ops->rdma_create_cq(dev->rdma_ctx, ¶ms, &icid);
921 cq->sig = QEDR_CQ_MAGIC_NUMBER;
922 spin_lock_init(&cq->cq_lock);
925 rc = qedr_copy_cq_uresp(dev, cq, udata);
929 /* Generate doorbell address. */
930 cq->db_addr = dev->db_addr +
931 DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT);
932 cq->db.data.icid = cq->icid;
933 cq->db.data.params = DB_AGG_CMD_SET <<
934 RDMA_PWM_VAL32_DATA_AGG_CMD_SHIFT;
936 /* point to the very last element, passing it we will toggle */
937 cq->toggle_cqe = qed_chain_get_last_elem(&cq->pbl);
938 cq->pbl_toggle = RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK;
939 cq->latest_cqe = NULL;
941 cq->cq_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
944 DP_DEBUG(dev, QEDR_MSG_CQ,
945 "create cq: icid=0x%0x, addr=%p, size(entries)=0x%0x\n",
946 cq->icid, cq, params.cq_size);
951 destroy_iparams.icid = cq->icid;
952 dev->ops->rdma_destroy_cq(dev->rdma_ctx, &destroy_iparams,
956 qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
958 dev->ops->common->chain_free(dev->cdev, &cq->pbl);
961 ib_umem_release(cq->q.umem);
964 return ERR_PTR(-EINVAL);
967 int qedr_resize_cq(struct ib_cq *ibcq, int new_cnt, struct ib_udata *udata)
969 struct qedr_dev *dev = get_qedr_dev(ibcq->device);
970 struct qedr_cq *cq = get_qedr_cq(ibcq);
972 DP_ERR(dev, "cq %p RESIZE NOT SUPPORTED\n", cq);
977 int qedr_destroy_cq(struct ib_cq *ibcq)
979 struct qedr_dev *dev = get_qedr_dev(ibcq->device);
980 struct qed_rdma_destroy_cq_out_params oparams;
981 struct qed_rdma_destroy_cq_in_params iparams;
982 struct qedr_cq *cq = get_qedr_cq(ibcq);
984 DP_DEBUG(dev, QEDR_MSG_CQ, "destroy cq: cq_id %d", cq->icid);
986 /* GSIs CQs are handled by driver, so they don't exist in the FW */
987 if (cq->cq_type != QEDR_CQ_TYPE_GSI) {
990 iparams.icid = cq->icid;
991 rc = dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams,
995 dev->ops->common->chain_free(dev->cdev, &cq->pbl);
998 if (ibcq->uobject && ibcq->uobject->context) {
999 qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
1000 ib_umem_release(cq->q.umem);
1008 static inline int get_gid_info_from_table(struct ib_qp *ibqp,
1009 struct ib_qp_attr *attr,
1011 struct qed_rdma_modify_qp_in_params
1014 enum rdma_network_type nw_type;
1015 struct ib_gid_attr gid_attr;
1021 rc = ib_get_cached_gid(ibqp->device, attr->ah_attr.port_num,
1022 attr->ah_attr.grh.sgid_index, &gid, &gid_attr);
1026 if (!memcmp(&gid, &zgid, sizeof(gid)))
1029 if (gid_attr.ndev) {
1030 qp_params->vlan_id = rdma_vlan_dev_vlan_id(gid_attr.ndev);
1032 dev_put(gid_attr.ndev);
1033 nw_type = ib_gid_to_network_type(gid_attr.gid_type, &gid);
1035 case RDMA_NETWORK_IPV6:
1036 memcpy(&qp_params->sgid.bytes[0], &gid.raw[0],
1037 sizeof(qp_params->sgid));
1038 memcpy(&qp_params->dgid.bytes[0],
1039 &attr->ah_attr.grh.dgid,
1040 sizeof(qp_params->dgid));
1041 qp_params->roce_mode = ROCE_V2_IPV6;
1042 SET_FIELD(qp_params->modify_flags,
1043 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
1045 case RDMA_NETWORK_IB:
1046 memcpy(&qp_params->sgid.bytes[0], &gid.raw[0],
1047 sizeof(qp_params->sgid));
1048 memcpy(&qp_params->dgid.bytes[0],
1049 &attr->ah_attr.grh.dgid,
1050 sizeof(qp_params->dgid));
1051 qp_params->roce_mode = ROCE_V1;
1053 case RDMA_NETWORK_IPV4:
1054 memset(&qp_params->sgid, 0, sizeof(qp_params->sgid));
1055 memset(&qp_params->dgid, 0, sizeof(qp_params->dgid));
1056 ipv4_addr = qedr_get_ipv4_from_gid(gid.raw);
1057 qp_params->sgid.ipv4_addr = ipv4_addr;
1059 qedr_get_ipv4_from_gid(attr->ah_attr.grh.dgid.raw);
1060 qp_params->dgid.ipv4_addr = ipv4_addr;
1061 SET_FIELD(qp_params->modify_flags,
1062 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
1063 qp_params->roce_mode = ROCE_V2_IPV4;
1068 for (i = 0; i < 4; i++) {
1069 qp_params->sgid.dwords[i] = ntohl(qp_params->sgid.dwords[i]);
1070 qp_params->dgid.dwords[i] = ntohl(qp_params->dgid.dwords[i]);
1073 if (qp_params->vlan_id >= VLAN_CFI_MASK)
1074 qp_params->vlan_id = 0;
1079 static void qedr_cleanup_user_sq(struct qedr_dev *dev, struct qedr_qp *qp)
1081 qedr_free_pbl(dev, &qp->usq.pbl_info, qp->usq.pbl_tbl);
1082 ib_umem_release(qp->usq.umem);
1085 static void qedr_cleanup_user_rq(struct qedr_dev *dev, struct qedr_qp *qp)
1087 qedr_free_pbl(dev, &qp->urq.pbl_info, qp->urq.pbl_tbl);
1088 ib_umem_release(qp->urq.umem);
1091 static void qedr_cleanup_kernel_sq(struct qedr_dev *dev, struct qedr_qp *qp)
1093 dev->ops->common->chain_free(dev->cdev, &qp->sq.pbl);
1094 kfree(qp->wqe_wr_id);
1097 static void qedr_cleanup_kernel_rq(struct qedr_dev *dev, struct qedr_qp *qp)
1099 dev->ops->common->chain_free(dev->cdev, &qp->rq.pbl);
1100 kfree(qp->rqe_wr_id);
1103 static int qedr_check_qp_attrs(struct ib_pd *ibpd, struct qedr_dev *dev,
1104 struct ib_qp_init_attr *attrs)
1106 struct qedr_device_attr *qattr = &dev->attr;
1108 /* QP0... attrs->qp_type == IB_QPT_GSI */
1109 if (attrs->qp_type != IB_QPT_RC && attrs->qp_type != IB_QPT_GSI) {
1110 DP_DEBUG(dev, QEDR_MSG_QP,
1111 "create qp: unsupported qp type=0x%x requested\n",
1116 if (attrs->cap.max_send_wr > qattr->max_sqe) {
1118 "create qp: cannot create a SQ with %d elements (max_send_wr=0x%x)\n",
1119 attrs->cap.max_send_wr, qattr->max_sqe);
1123 if (attrs->cap.max_inline_data > qattr->max_inline) {
1125 "create qp: unsupported inline data size=0x%x requested (max_inline=0x%x)\n",
1126 attrs->cap.max_inline_data, qattr->max_inline);
1130 if (attrs->cap.max_send_sge > qattr->max_sge) {
1132 "create qp: unsupported send_sge=0x%x requested (max_send_sge=0x%x)\n",
1133 attrs->cap.max_send_sge, qattr->max_sge);
1137 if (attrs->cap.max_recv_sge > qattr->max_sge) {
1139 "create qp: unsupported recv_sge=0x%x requested (max_recv_sge=0x%x)\n",
1140 attrs->cap.max_recv_sge, qattr->max_sge);
1144 /* Unprivileged user space cannot create special QP */
1145 if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1147 "create qp: userspace can't create special QPs of type=0x%x\n",
1155 static void qedr_copy_rq_uresp(struct qedr_create_qp_uresp *uresp,
1158 uresp->rq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
1159 uresp->rq_icid = qp->icid;
1162 static void qedr_copy_sq_uresp(struct qedr_create_qp_uresp *uresp,
1165 uresp->sq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1166 uresp->sq_icid = qp->icid + 1;
1169 static int qedr_copy_qp_uresp(struct qedr_dev *dev,
1170 struct qedr_qp *qp, struct ib_udata *udata)
1172 struct qedr_create_qp_uresp uresp;
1175 memset(&uresp, 0, sizeof(uresp));
1176 qedr_copy_sq_uresp(&uresp, qp);
1177 qedr_copy_rq_uresp(&uresp, qp);
1179 uresp.atomic_supported = dev->atomic_cap != IB_ATOMIC_NONE;
1180 uresp.qp_id = qp->qp_id;
1182 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1185 "create qp: failed a copy to user space with qp icid=0x%x.\n",
1191 static void qedr_set_qp_init_params(struct qedr_dev *dev,
1194 struct ib_qp_init_attr *attrs)
1198 spin_lock_init(&qp->q_lock);
1200 qp->qp_type = attrs->qp_type;
1201 qp->max_inline_data = attrs->cap.max_inline_data;
1202 qp->sq.max_sges = attrs->cap.max_send_sge;
1203 qp->state = QED_ROCE_QP_STATE_RESET;
1204 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1205 qp->sq_cq = get_qedr_cq(attrs->send_cq);
1206 qp->rq_cq = get_qedr_cq(attrs->recv_cq);
1209 DP_DEBUG(dev, QEDR_MSG_QP,
1210 "QP params:\tpd = %d, qp_type = %d, max_inline_data = %d, state = %d, signaled = %d, use_srq=%d\n",
1211 pd->pd_id, qp->qp_type, qp->max_inline_data,
1212 qp->state, qp->signaled, (attrs->srq) ? 1 : 0);
1213 DP_DEBUG(dev, QEDR_MSG_QP,
1214 "SQ params:\tsq_max_sges = %d, sq_cq_id = %d\n",
1215 qp->sq.max_sges, qp->sq_cq->icid);
1216 qp->rq.max_sges = attrs->cap.max_recv_sge;
1217 DP_DEBUG(dev, QEDR_MSG_QP,
1218 "RQ params:\trq_max_sges = %d, rq_cq_id = %d\n",
1219 qp->rq.max_sges, qp->rq_cq->icid);
1223 qedr_init_qp_user_params(struct qed_rdma_create_qp_in_params *params,
1224 struct qedr_create_qp_ureq *ureq)
1226 /* QP handle to be written in CQE */
1227 params->qp_handle_lo = ureq->qp_handle_lo;
1228 params->qp_handle_hi = ureq->qp_handle_hi;
1232 qedr_init_qp_kernel_doorbell_sq(struct qedr_dev *dev, struct qedr_qp *qp)
1234 qp->sq.db = dev->db_addr +
1235 DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1236 qp->sq.db_data.data.icid = qp->icid + 1;
1240 qedr_init_qp_kernel_doorbell_rq(struct qedr_dev *dev, struct qedr_qp *qp)
1242 qp->rq.db = dev->db_addr +
1243 DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
1244 qp->rq.db_data.data.icid = qp->icid;
1248 qedr_init_qp_kernel_params_rq(struct qedr_dev *dev,
1249 struct qedr_qp *qp, struct ib_qp_init_attr *attrs)
1251 /* Allocate driver internal RQ array */
1252 qp->rqe_wr_id = kcalloc(qp->rq.max_wr, sizeof(*qp->rqe_wr_id),
1257 DP_DEBUG(dev, QEDR_MSG_QP, "RQ max_wr set to %d.\n", qp->rq.max_wr);
1263 qedr_init_qp_kernel_params_sq(struct qedr_dev *dev,
1265 struct ib_qp_init_attr *attrs,
1266 struct qed_rdma_create_qp_in_params *params)
1270 /* Allocate driver internal SQ array */
1271 temp_max_wr = attrs->cap.max_send_wr * dev->wq_multiplier;
1272 temp_max_wr = min_t(u32, temp_max_wr, dev->attr.max_sqe);
1274 /* temp_max_wr < attr->max_sqe < u16 so the casting is safe */
1275 qp->sq.max_wr = (u16)temp_max_wr;
1276 qp->wqe_wr_id = kcalloc(qp->sq.max_wr, sizeof(*qp->wqe_wr_id),
1281 DP_DEBUG(dev, QEDR_MSG_QP, "SQ max_wr set to %d.\n", qp->sq.max_wr);
1283 /* QP handle to be written in CQE */
1284 params->qp_handle_lo = lower_32_bits((uintptr_t)qp);
1285 params->qp_handle_hi = upper_32_bits((uintptr_t)qp);
1290 static inline int qedr_init_qp_kernel_sq(struct qedr_dev *dev,
1292 struct ib_qp_init_attr *attrs)
1294 u32 n_sq_elems, n_sq_entries;
1297 /* A single work request may take up to QEDR_MAX_SQ_WQE_SIZE elements in
1298 * the ring. The ring should allow at least a single WR, even if the
1299 * user requested none, due to allocation issues.
1301 n_sq_entries = attrs->cap.max_send_wr;
1302 n_sq_entries = min_t(u32, n_sq_entries, dev->attr.max_sqe);
1303 n_sq_entries = max_t(u32, n_sq_entries, 1);
1304 n_sq_elems = n_sq_entries * QEDR_MAX_SQE_ELEMENTS_PER_SQE;
1305 rc = dev->ops->common->chain_alloc(dev->cdev,
1306 QED_CHAIN_USE_TO_PRODUCE,
1308 QED_CHAIN_CNT_TYPE_U32,
1310 QEDR_SQE_ELEMENT_SIZE,
1313 DP_ERR(dev, "failed to allocate QP %p SQ\n", qp);
1317 DP_DEBUG(dev, QEDR_MSG_SQ,
1318 "SQ Pbl base addr = %llx max_send_wr=%d max_wr=%d capacity=%d, rc=%d\n",
1319 qed_chain_get_pbl_phys(&qp->sq.pbl), attrs->cap.max_send_wr,
1320 n_sq_entries, qed_chain_get_capacity(&qp->sq.pbl), rc);
1324 static inline int qedr_init_qp_kernel_rq(struct qedr_dev *dev,
1326 struct ib_qp_init_attr *attrs)
1328 u32 n_rq_elems, n_rq_entries;
1331 /* A single work request may take up to QEDR_MAX_RQ_WQE_SIZE elements in
1332 * the ring. There ring should allow at least a single WR, even if the
1333 * user requested none, due to allocation issues.
1335 n_rq_entries = max_t(u32, attrs->cap.max_recv_wr, 1);
1336 n_rq_elems = n_rq_entries * QEDR_MAX_RQE_ELEMENTS_PER_RQE;
1337 rc = dev->ops->common->chain_alloc(dev->cdev,
1338 QED_CHAIN_USE_TO_CONSUME_PRODUCE,
1340 QED_CHAIN_CNT_TYPE_U32,
1342 QEDR_RQE_ELEMENT_SIZE,
1346 DP_ERR(dev, "failed to allocate memory for QP %p RQ\n", qp);
1350 DP_DEBUG(dev, QEDR_MSG_RQ,
1351 "RQ Pbl base addr = %llx max_recv_wr=%d max_wr=%d capacity=%d, rc=%d\n",
1352 qed_chain_get_pbl_phys(&qp->rq.pbl), attrs->cap.max_recv_wr,
1353 n_rq_entries, qed_chain_get_capacity(&qp->rq.pbl), rc);
1355 /* n_rq_entries < u16 so the casting is safe */
1356 qp->rq.max_wr = (u16)n_rq_entries;
1362 qedr_init_qp_in_params_sq(struct qedr_dev *dev,
1365 struct ib_qp_init_attr *attrs,
1366 struct ib_udata *udata,
1367 struct qed_rdma_create_qp_in_params *params)
1369 /* QP handle to be written in an async event */
1370 params->qp_handle_async_lo = lower_32_bits((uintptr_t)qp);
1371 params->qp_handle_async_hi = upper_32_bits((uintptr_t)qp);
1373 params->signal_all = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR);
1374 params->fmr_and_reserved_lkey = !udata;
1375 params->pd = pd->pd_id;
1376 params->dpi = pd->uctx ? pd->uctx->dpi : dev->dpi;
1377 params->sq_cq_id = get_qedr_cq(attrs->send_cq)->icid;
1378 params->max_sq_sges = 0;
1379 params->stats_queue = 0;
1382 params->sq_num_pages = qp->usq.pbl_info.num_pbes;
1383 params->sq_pbl_ptr = qp->usq.pbl_tbl->pa;
1385 params->sq_num_pages = qed_chain_get_page_cnt(&qp->sq.pbl);
1386 params->sq_pbl_ptr = qed_chain_get_pbl_phys(&qp->sq.pbl);
1391 qedr_init_qp_in_params_rq(struct qedr_qp *qp,
1392 struct ib_qp_init_attr *attrs,
1393 struct ib_udata *udata,
1394 struct qed_rdma_create_qp_in_params *params)
1396 params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
1398 params->use_srq = false;
1401 params->rq_num_pages = qp->urq.pbl_info.num_pbes;
1402 params->rq_pbl_ptr = qp->urq.pbl_tbl->pa;
1404 params->rq_num_pages = qed_chain_get_page_cnt(&qp->rq.pbl);
1405 params->rq_pbl_ptr = qed_chain_get_pbl_phys(&qp->rq.pbl);
1409 static inline void qedr_qp_user_print(struct qedr_dev *dev, struct qedr_qp *qp)
1411 DP_DEBUG(dev, QEDR_MSG_QP,
1412 "create qp: successfully created user QP. qp=%p, sq_addr=0x%llx, sq_len=%zd, rq_addr=0x%llx, rq_len=%zd\n",
1413 qp, qp->usq.buf_addr, qp->usq.buf_len, qp->urq.buf_addr,
1417 static inline int qedr_init_user_qp(struct ib_ucontext *ib_ctx,
1418 struct qedr_dev *dev,
1420 struct qedr_create_qp_ureq *ureq)
1424 /* SQ - read access only (0), dma sync not required (0) */
1425 rc = qedr_init_user_queue(ib_ctx, dev, &qp->usq, ureq->sq_addr,
1426 ureq->sq_len, 0, 0);
1430 /* RQ - read access only (0), dma sync not required (0) */
1431 rc = qedr_init_user_queue(ib_ctx, dev, &qp->urq, ureq->rq_addr,
1432 ureq->rq_len, 0, 0);
1435 qedr_cleanup_user_sq(dev, qp);
1440 qedr_init_kernel_qp(struct qedr_dev *dev,
1442 struct ib_qp_init_attr *attrs,
1443 struct qed_rdma_create_qp_in_params *params)
1447 rc = qedr_init_qp_kernel_sq(dev, qp, attrs);
1449 DP_ERR(dev, "failed to init kernel QP %p SQ\n", qp);
1453 rc = qedr_init_qp_kernel_params_sq(dev, qp, attrs, params);
1455 dev->ops->common->chain_free(dev->cdev, &qp->sq.pbl);
1456 DP_ERR(dev, "failed to init kernel QP %p SQ params\n", qp);
1460 rc = qedr_init_qp_kernel_rq(dev, qp, attrs);
1462 qedr_cleanup_kernel_sq(dev, qp);
1463 DP_ERR(dev, "failed to init kernel QP %p RQ\n", qp);
1467 rc = qedr_init_qp_kernel_params_rq(dev, qp, attrs);
1469 DP_ERR(dev, "failed to init kernel QP %p RQ params\n", qp);
1470 qedr_cleanup_kernel_sq(dev, qp);
1471 dev->ops->common->chain_free(dev->cdev, &qp->rq.pbl);
1478 struct ib_qp *qedr_create_qp(struct ib_pd *ibpd,
1479 struct ib_qp_init_attr *attrs,
1480 struct ib_udata *udata)
1482 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
1483 struct qed_rdma_create_qp_out_params out_params;
1484 struct qed_rdma_create_qp_in_params in_params;
1485 struct qedr_pd *pd = get_qedr_pd(ibpd);
1486 struct ib_ucontext *ib_ctx = NULL;
1487 struct qedr_ucontext *ctx = NULL;
1488 struct qedr_create_qp_ureq ureq;
1493 DP_DEBUG(dev, QEDR_MSG_QP, "create qp: called from %s, pd=%p\n",
1494 udata ? "user library" : "kernel", pd);
1496 rc = qedr_check_qp_attrs(ibpd, dev, attrs);
1501 return ERR_PTR(-EINVAL);
1503 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1505 return ERR_PTR(-ENOMEM);
1507 DP_DEBUG(dev, QEDR_MSG_QP,
1508 "create qp: sq_cq=%p, sq_icid=%d, rq_cq=%p, rq_icid=%d\n",
1509 get_qedr_cq(attrs->send_cq),
1510 get_qedr_cq(attrs->send_cq)->icid,
1511 get_qedr_cq(attrs->recv_cq),
1512 get_qedr_cq(attrs->recv_cq)->icid);
1514 qedr_set_qp_init_params(dev, qp, pd, attrs);
1516 if (attrs->qp_type == IB_QPT_GSI) {
1519 "create qp: unexpected udata when creating GSI QP\n");
1522 ibqp = qedr_create_gsi_qp(dev, attrs, qp);
1528 memset(&in_params, 0, sizeof(in_params));
1531 if (!(udata && ibpd->uobject && ibpd->uobject->context))
1534 ib_ctx = ibpd->uobject->context;
1535 ctx = get_qedr_ucontext(ib_ctx);
1537 memset(&ureq, 0, sizeof(ureq));
1538 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) {
1540 "create qp: problem copying data from user space\n");
1544 rc = qedr_init_user_qp(ib_ctx, dev, qp, &ureq);
1548 qedr_init_qp_user_params(&in_params, &ureq);
1550 rc = qedr_init_kernel_qp(dev, qp, attrs, &in_params);
1555 qedr_init_qp_in_params_sq(dev, pd, qp, attrs, udata, &in_params);
1556 qedr_init_qp_in_params_rq(qp, attrs, udata, &in_params);
1558 qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
1559 &in_params, &out_params);
1564 qp->qp_id = out_params.qp_id;
1565 qp->icid = out_params.icid;
1566 qp->ibqp.qp_num = qp->qp_id;
1569 rc = qedr_copy_qp_uresp(dev, qp, udata);
1573 qedr_qp_user_print(dev, qp);
1575 qedr_init_qp_kernel_doorbell_sq(dev, qp);
1576 qedr_init_qp_kernel_doorbell_rq(dev, qp);
1579 DP_DEBUG(dev, QEDR_MSG_QP, "created %s space QP %p\n",
1580 udata ? "user" : "kernel", qp);
1585 rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
1587 DP_ERR(dev, "create qp: fatal fault. rc=%d", rc);
1590 qedr_cleanup_user_sq(dev, qp);
1591 qedr_cleanup_user_rq(dev, qp);
1593 qedr_cleanup_kernel_sq(dev, qp);
1594 qedr_cleanup_kernel_rq(dev, qp);
1600 return ERR_PTR(-EFAULT);
1603 enum ib_qp_state qedr_get_ibqp_state(enum qed_roce_qp_state qp_state)
1606 case QED_ROCE_QP_STATE_RESET:
1607 return IB_QPS_RESET;
1608 case QED_ROCE_QP_STATE_INIT:
1610 case QED_ROCE_QP_STATE_RTR:
1612 case QED_ROCE_QP_STATE_RTS:
1614 case QED_ROCE_QP_STATE_SQD:
1616 case QED_ROCE_QP_STATE_ERR:
1618 case QED_ROCE_QP_STATE_SQE:
1624 enum qed_roce_qp_state qedr_get_state_from_ibqp(enum ib_qp_state qp_state)
1628 return QED_ROCE_QP_STATE_RESET;
1630 return QED_ROCE_QP_STATE_INIT;
1632 return QED_ROCE_QP_STATE_RTR;
1634 return QED_ROCE_QP_STATE_RTS;
1636 return QED_ROCE_QP_STATE_SQD;
1638 return QED_ROCE_QP_STATE_ERR;
1640 return QED_ROCE_QP_STATE_ERR;
1644 static void qedr_reset_qp_hwq_info(struct qedr_qp_hwq_info *qph)
1646 qed_chain_reset(&qph->pbl);
1650 qph->db_data.data.value = cpu_to_le16(0);
1653 static int qedr_update_qp_state(struct qedr_dev *dev,
1655 enum qed_roce_qp_state new_state)
1659 if (new_state == qp->state)
1662 switch (qp->state) {
1663 case QED_ROCE_QP_STATE_RESET:
1664 switch (new_state) {
1665 case QED_ROCE_QP_STATE_INIT:
1666 qp->prev_wqe_size = 0;
1667 qedr_reset_qp_hwq_info(&qp->sq);
1668 qedr_reset_qp_hwq_info(&qp->rq);
1675 case QED_ROCE_QP_STATE_INIT:
1676 switch (new_state) {
1677 case QED_ROCE_QP_STATE_RTR:
1678 /* Update doorbell (in case post_recv was
1679 * done before move to RTR)
1682 writel(qp->rq.db_data.raw, qp->rq.db);
1683 /* Make sure write takes effect */
1686 case QED_ROCE_QP_STATE_ERR:
1689 /* Invalid state change. */
1694 case QED_ROCE_QP_STATE_RTR:
1696 switch (new_state) {
1697 case QED_ROCE_QP_STATE_RTS:
1699 case QED_ROCE_QP_STATE_ERR:
1702 /* Invalid state change. */
1707 case QED_ROCE_QP_STATE_RTS:
1709 switch (new_state) {
1710 case QED_ROCE_QP_STATE_SQD:
1712 case QED_ROCE_QP_STATE_ERR:
1715 /* Invalid state change. */
1720 case QED_ROCE_QP_STATE_SQD:
1722 switch (new_state) {
1723 case QED_ROCE_QP_STATE_RTS:
1724 case QED_ROCE_QP_STATE_ERR:
1727 /* Invalid state change. */
1732 case QED_ROCE_QP_STATE_ERR:
1734 switch (new_state) {
1735 case QED_ROCE_QP_STATE_RESET:
1750 int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1751 int attr_mask, struct ib_udata *udata)
1753 struct qedr_qp *qp = get_qedr_qp(ibqp);
1754 struct qed_rdma_modify_qp_in_params qp_params = { 0 };
1755 struct qedr_dev *dev = get_qedr_dev(&qp->dev->ibdev);
1756 enum ib_qp_state old_qp_state, new_qp_state;
1759 DP_DEBUG(dev, QEDR_MSG_QP,
1760 "modify qp: qp %p attr_mask=0x%x, state=%d", qp, attr_mask,
1763 old_qp_state = qedr_get_ibqp_state(qp->state);
1764 if (attr_mask & IB_QP_STATE)
1765 new_qp_state = attr->qp_state;
1767 new_qp_state = old_qp_state;
1769 if (!ib_modify_qp_is_ok
1770 (old_qp_state, new_qp_state, ibqp->qp_type, attr_mask,
1771 IB_LINK_LAYER_ETHERNET)) {
1773 "modify qp: invalid attribute mask=0x%x specified for\n"
1774 "qpn=0x%x of type=0x%x old_qp_state=0x%x, new_qp_state=0x%x\n",
1775 attr_mask, qp->qp_id, ibqp->qp_type, old_qp_state,
1781 /* Translate the masks... */
1782 if (attr_mask & IB_QP_STATE) {
1783 SET_FIELD(qp_params.modify_flags,
1784 QED_RDMA_MODIFY_QP_VALID_NEW_STATE, 1);
1785 qp_params.new_state = qedr_get_state_from_ibqp(attr->qp_state);
1788 if (attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
1789 qp_params.sqd_async = true;
1791 if (attr_mask & IB_QP_PKEY_INDEX) {
1792 SET_FIELD(qp_params.modify_flags,
1793 QED_ROCE_MODIFY_QP_VALID_PKEY, 1);
1794 if (attr->pkey_index >= QEDR_ROCE_PKEY_TABLE_LEN) {
1799 qp_params.pkey = QEDR_ROCE_PKEY_DEFAULT;
1802 if (attr_mask & IB_QP_QKEY)
1803 qp->qkey = attr->qkey;
1805 if (attr_mask & IB_QP_ACCESS_FLAGS) {
1806 SET_FIELD(qp_params.modify_flags,
1807 QED_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN, 1);
1808 qp_params.incoming_rdma_read_en = attr->qp_access_flags &
1809 IB_ACCESS_REMOTE_READ;
1810 qp_params.incoming_rdma_write_en = attr->qp_access_flags &
1811 IB_ACCESS_REMOTE_WRITE;
1812 qp_params.incoming_atomic_en = attr->qp_access_flags &
1813 IB_ACCESS_REMOTE_ATOMIC;
1816 if (attr_mask & (IB_QP_AV | IB_QP_PATH_MTU)) {
1817 if (attr_mask & IB_QP_PATH_MTU) {
1818 if (attr->path_mtu < IB_MTU_256 ||
1819 attr->path_mtu > IB_MTU_4096) {
1820 pr_err("error: Only MTU sizes of 256, 512, 1024, 2048 and 4096 are supported by RoCE\n");
1824 qp->mtu = min(ib_mtu_enum_to_int(attr->path_mtu),
1825 ib_mtu_enum_to_int(iboe_get_mtu
1831 ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
1832 pr_err("Fixing zeroed MTU to qp->mtu = %d\n", qp->mtu);
1835 SET_FIELD(qp_params.modify_flags,
1836 QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR, 1);
1838 qp_params.traffic_class_tos = attr->ah_attr.grh.traffic_class;
1839 qp_params.flow_label = attr->ah_attr.grh.flow_label;
1840 qp_params.hop_limit_ttl = attr->ah_attr.grh.hop_limit;
1842 qp->sgid_idx = attr->ah_attr.grh.sgid_index;
1844 rc = get_gid_info_from_table(ibqp, attr, attr_mask, &qp_params);
1847 "modify qp: problems with GID index %d (rc=%d)\n",
1848 attr->ah_attr.grh.sgid_index, rc);
1852 rc = qedr_get_dmac(dev, &attr->ah_attr,
1853 qp_params.remote_mac_addr);
1857 qp_params.use_local_mac = true;
1858 ether_addr_copy(qp_params.local_mac_addr, dev->ndev->dev_addr);
1860 DP_DEBUG(dev, QEDR_MSG_QP, "dgid=%x:%x:%x:%x\n",
1861 qp_params.dgid.dwords[0], qp_params.dgid.dwords[1],
1862 qp_params.dgid.dwords[2], qp_params.dgid.dwords[3]);
1863 DP_DEBUG(dev, QEDR_MSG_QP, "sgid=%x:%x:%x:%x\n",
1864 qp_params.sgid.dwords[0], qp_params.sgid.dwords[1],
1865 qp_params.sgid.dwords[2], qp_params.sgid.dwords[3]);
1866 DP_DEBUG(dev, QEDR_MSG_QP, "remote_mac=[%pM]\n",
1867 qp_params.remote_mac_addr);
1870 qp_params.mtu = qp->mtu;
1871 qp_params.lb_indication = false;
1874 if (!qp_params.mtu) {
1875 /* Stay with current MTU */
1877 qp_params.mtu = qp->mtu;
1880 ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
1883 if (attr_mask & IB_QP_TIMEOUT) {
1884 SET_FIELD(qp_params.modify_flags,
1885 QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1);
1887 qp_params.ack_timeout = attr->timeout;
1888 if (attr->timeout) {
1891 temp = 4096 * (1UL << attr->timeout) / 1000 / 1000;
1892 /* FW requires [msec] */
1893 qp_params.ack_timeout = temp;
1896 qp_params.ack_timeout = 0;
1899 if (attr_mask & IB_QP_RETRY_CNT) {
1900 SET_FIELD(qp_params.modify_flags,
1901 QED_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1);
1902 qp_params.retry_cnt = attr->retry_cnt;
1905 if (attr_mask & IB_QP_RNR_RETRY) {
1906 SET_FIELD(qp_params.modify_flags,
1907 QED_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT, 1);
1908 qp_params.rnr_retry_cnt = attr->rnr_retry;
1911 if (attr_mask & IB_QP_RQ_PSN) {
1912 SET_FIELD(qp_params.modify_flags,
1913 QED_ROCE_MODIFY_QP_VALID_RQ_PSN, 1);
1914 qp_params.rq_psn = attr->rq_psn;
1915 qp->rq_psn = attr->rq_psn;
1918 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1919 if (attr->max_rd_atomic > dev->attr.max_qp_req_rd_atomic_resc) {
1922 "unsupported max_rd_atomic=%d, supported=%d\n",
1923 attr->max_rd_atomic,
1924 dev->attr.max_qp_req_rd_atomic_resc);
1928 SET_FIELD(qp_params.modify_flags,
1929 QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ, 1);
1930 qp_params.max_rd_atomic_req = attr->max_rd_atomic;
1933 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1934 SET_FIELD(qp_params.modify_flags,
1935 QED_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER, 1);
1936 qp_params.min_rnr_nak_timer = attr->min_rnr_timer;
1939 if (attr_mask & IB_QP_SQ_PSN) {
1940 SET_FIELD(qp_params.modify_flags,
1941 QED_ROCE_MODIFY_QP_VALID_SQ_PSN, 1);
1942 qp_params.sq_psn = attr->sq_psn;
1943 qp->sq_psn = attr->sq_psn;
1946 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1947 if (attr->max_dest_rd_atomic >
1948 dev->attr.max_qp_resp_rd_atomic_resc) {
1950 "unsupported max_dest_rd_atomic=%d, supported=%d\n",
1951 attr->max_dest_rd_atomic,
1952 dev->attr.max_qp_resp_rd_atomic_resc);
1958 SET_FIELD(qp_params.modify_flags,
1959 QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP, 1);
1960 qp_params.max_rd_atomic_resp = attr->max_dest_rd_atomic;
1963 if (attr_mask & IB_QP_DEST_QPN) {
1964 SET_FIELD(qp_params.modify_flags,
1965 QED_ROCE_MODIFY_QP_VALID_DEST_QP, 1);
1967 qp_params.dest_qp = attr->dest_qp_num;
1968 qp->dest_qp_num = attr->dest_qp_num;
1971 if (qp->qp_type != IB_QPT_GSI)
1972 rc = dev->ops->rdma_modify_qp(dev->rdma_ctx,
1973 qp->qed_qp, &qp_params);
1975 if (attr_mask & IB_QP_STATE) {
1976 if ((qp->qp_type != IB_QPT_GSI) && (!udata))
1977 qedr_update_qp_state(dev, qp, qp_params.new_state);
1978 qp->state = qp_params.new_state;
1985 static int qedr_to_ib_qp_acc_flags(struct qed_rdma_query_qp_out_params *params)
1987 int ib_qp_acc_flags = 0;
1989 if (params->incoming_rdma_write_en)
1990 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1991 if (params->incoming_rdma_read_en)
1992 ib_qp_acc_flags |= IB_ACCESS_REMOTE_READ;
1993 if (params->incoming_atomic_en)
1994 ib_qp_acc_flags |= IB_ACCESS_REMOTE_ATOMIC;
1995 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1996 return ib_qp_acc_flags;
1999 int qedr_query_qp(struct ib_qp *ibqp,
2000 struct ib_qp_attr *qp_attr,
2001 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
2003 struct qed_rdma_query_qp_out_params params;
2004 struct qedr_qp *qp = get_qedr_qp(ibqp);
2005 struct qedr_dev *dev = qp->dev;
2008 memset(¶ms, 0, sizeof(params));
2010 rc = dev->ops->rdma_query_qp(dev->rdma_ctx, qp->qed_qp, ¶ms);
2014 memset(qp_attr, 0, sizeof(*qp_attr));
2015 memset(qp_init_attr, 0, sizeof(*qp_init_attr));
2017 qp_attr->qp_state = qedr_get_ibqp_state(params.state);
2018 qp_attr->cur_qp_state = qedr_get_ibqp_state(params.state);
2019 qp_attr->path_mtu = iboe_get_mtu(params.mtu);
2020 qp_attr->path_mig_state = IB_MIG_MIGRATED;
2021 qp_attr->rq_psn = params.rq_psn;
2022 qp_attr->sq_psn = params.sq_psn;
2023 qp_attr->dest_qp_num = params.dest_qp;
2025 qp_attr->qp_access_flags = qedr_to_ib_qp_acc_flags(¶ms);
2027 qp_attr->cap.max_send_wr = qp->sq.max_wr;
2028 qp_attr->cap.max_recv_wr = qp->rq.max_wr;
2029 qp_attr->cap.max_send_sge = qp->sq.max_sges;
2030 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
2031 qp_attr->cap.max_inline_data = qp->max_inline_data;
2032 qp_init_attr->cap = qp_attr->cap;
2034 memcpy(&qp_attr->ah_attr.grh.dgid.raw[0], ¶ms.dgid.bytes[0],
2035 sizeof(qp_attr->ah_attr.grh.dgid.raw));
2037 qp_attr->ah_attr.grh.flow_label = params.flow_label;
2038 qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
2039 qp_attr->ah_attr.grh.hop_limit = params.hop_limit_ttl;
2040 qp_attr->ah_attr.grh.traffic_class = params.traffic_class_tos;
2042 qp_attr->ah_attr.ah_flags = IB_AH_GRH;
2043 qp_attr->ah_attr.port_num = 1;
2044 qp_attr->ah_attr.sl = 0;
2045 qp_attr->timeout = params.timeout;
2046 qp_attr->rnr_retry = params.rnr_retry;
2047 qp_attr->retry_cnt = params.retry_cnt;
2048 qp_attr->min_rnr_timer = params.min_rnr_nak_timer;
2049 qp_attr->pkey_index = params.pkey_index;
2050 qp_attr->port_num = 1;
2051 qp_attr->ah_attr.src_path_bits = 0;
2052 qp_attr->ah_attr.static_rate = 0;
2053 qp_attr->alt_pkey_index = 0;
2054 qp_attr->alt_port_num = 0;
2055 qp_attr->alt_timeout = 0;
2056 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
2058 qp_attr->sq_draining = (params.state == QED_ROCE_QP_STATE_SQD) ? 1 : 0;
2059 qp_attr->max_dest_rd_atomic = params.max_dest_rd_atomic;
2060 qp_attr->max_rd_atomic = params.max_rd_atomic;
2061 qp_attr->en_sqd_async_notify = (params.sqd_async) ? 1 : 0;
2063 DP_DEBUG(dev, QEDR_MSG_QP, "QEDR_QUERY_QP: max_inline_data=%d\n",
2064 qp_attr->cap.max_inline_data);
2070 int qedr_destroy_qp(struct ib_qp *ibqp)
2072 struct qedr_qp *qp = get_qedr_qp(ibqp);
2073 struct qedr_dev *dev = qp->dev;
2074 struct ib_qp_attr attr;
2078 DP_DEBUG(dev, QEDR_MSG_QP, "destroy qp: destroying %p, qp type=%d\n",
2081 if (qp->state != (QED_ROCE_QP_STATE_RESET | QED_ROCE_QP_STATE_ERR |
2082 QED_ROCE_QP_STATE_INIT)) {
2083 attr.qp_state = IB_QPS_ERR;
2084 attr_mask |= IB_QP_STATE;
2086 /* Change the QP state to ERROR */
2087 qedr_modify_qp(ibqp, &attr, attr_mask, NULL);
2090 if (qp->qp_type != IB_QPT_GSI) {
2091 rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
2095 qedr_destroy_gsi_qp(dev);
2098 if (ibqp->uobject && ibqp->uobject->context) {
2099 qedr_cleanup_user_sq(dev, qp);
2100 qedr_cleanup_user_rq(dev, qp);
2102 qedr_cleanup_kernel_sq(dev, qp);
2103 qedr_cleanup_kernel_rq(dev, qp);
2111 struct ib_ah *qedr_create_ah(struct ib_pd *ibpd, struct ib_ah_attr *attr,
2112 struct ib_udata *udata)
2116 ah = kzalloc(sizeof(*ah), GFP_ATOMIC);
2118 return ERR_PTR(-ENOMEM);
2125 int qedr_destroy_ah(struct ib_ah *ibah)
2127 struct qedr_ah *ah = get_qedr_ah(ibah);
2133 static void free_mr_info(struct qedr_dev *dev, struct mr_info *info)
2135 struct qedr_pbl *pbl, *tmp;
2137 if (info->pbl_table)
2138 list_add_tail(&info->pbl_table->list_entry,
2139 &info->free_pbl_list);
2141 if (!list_empty(&info->inuse_pbl_list))
2142 list_splice(&info->inuse_pbl_list, &info->free_pbl_list);
2144 list_for_each_entry_safe(pbl, tmp, &info->free_pbl_list, list_entry) {
2145 list_del(&pbl->list_entry);
2146 qedr_free_pbl(dev, &info->pbl_info, pbl);
2150 static int init_mr_info(struct qedr_dev *dev, struct mr_info *info,
2151 size_t page_list_len, bool two_layered)
2153 struct qedr_pbl *tmp;
2156 INIT_LIST_HEAD(&info->free_pbl_list);
2157 INIT_LIST_HEAD(&info->inuse_pbl_list);
2159 rc = qedr_prepare_pbl_tbl(dev, &info->pbl_info,
2160 page_list_len, two_layered);
2164 info->pbl_table = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
2165 if (!info->pbl_table) {
2170 DP_DEBUG(dev, QEDR_MSG_MR, "pbl_table_pa = %pa\n",
2171 &info->pbl_table->pa);
2173 /* in usual case we use 2 PBLs, so we add one to free
2174 * list and allocating another one
2176 tmp = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
2178 DP_DEBUG(dev, QEDR_MSG_MR, "Extra PBL is not allocated\n");
2182 list_add_tail(&tmp->list_entry, &info->free_pbl_list);
2184 DP_DEBUG(dev, QEDR_MSG_MR, "extra pbl_table_pa = %pa\n", &tmp->pa);
2188 free_mr_info(dev, info);
2193 struct ib_mr *qedr_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
2194 u64 usr_addr, int acc, struct ib_udata *udata)
2196 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
2201 pd = get_qedr_pd(ibpd);
2202 DP_DEBUG(dev, QEDR_MSG_MR,
2203 "qedr_register user mr pd = %d start = %lld, len = %lld, usr_addr = %lld, acc = %d\n",
2204 pd->pd_id, start, len, usr_addr, acc);
2206 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
2207 return ERR_PTR(-EINVAL);
2209 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2213 mr->type = QEDR_MR_USER;
2215 mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
2216 if (IS_ERR(mr->umem)) {
2221 rc = init_mr_info(dev, &mr->info, ib_umem_page_count(mr->umem), 1);
2225 qedr_populate_pbls(dev, mr->umem, mr->info.pbl_table,
2226 &mr->info.pbl_info);
2228 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
2230 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
2234 /* Index only, 18 bit long, lkey = itid << 8 | key */
2235 mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
2237 mr->hw_mr.pd = pd->pd_id;
2238 mr->hw_mr.local_read = 1;
2239 mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
2240 mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
2241 mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
2242 mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
2243 mr->hw_mr.mw_bind = false;
2244 mr->hw_mr.pbl_ptr = mr->info.pbl_table[0].pa;
2245 mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
2246 mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
2247 mr->hw_mr.page_size_log = ilog2(mr->umem->page_size);
2248 mr->hw_mr.fbo = ib_umem_offset(mr->umem);
2249 mr->hw_mr.length = len;
2250 mr->hw_mr.vaddr = usr_addr;
2251 mr->hw_mr.zbva = false;
2252 mr->hw_mr.phy_mr = false;
2253 mr->hw_mr.dma_mr = false;
2255 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
2257 DP_ERR(dev, "roce register tid returned an error %d\n", rc);
2261 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2262 if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
2263 mr->hw_mr.remote_atomic)
2264 mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2266 DP_DEBUG(dev, QEDR_MSG_MR, "register user mr lkey: %x\n",
2271 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
2273 qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
2279 int qedr_dereg_mr(struct ib_mr *ib_mr)
2281 struct qedr_mr *mr = get_qedr_mr(ib_mr);
2282 struct qedr_dev *dev = get_qedr_dev(ib_mr->device);
2285 rc = dev->ops->rdma_deregister_tid(dev->rdma_ctx, mr->hw_mr.itid);
2289 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
2291 if ((mr->type != QEDR_MR_DMA) && (mr->type != QEDR_MR_FRMR))
2292 qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
2294 /* it could be user registered memory. */
2296 ib_umem_release(mr->umem);
2303 struct qedr_mr *__qedr_alloc_mr(struct ib_pd *ibpd, int max_page_list_len)
2305 struct qedr_pd *pd = get_qedr_pd(ibpd);
2306 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
2310 DP_DEBUG(dev, QEDR_MSG_MR,
2311 "qedr_alloc_frmr pd = %d max_page_list_len= %d\n", pd->pd_id,
2314 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2319 mr->type = QEDR_MR_FRMR;
2321 rc = init_mr_info(dev, &mr->info, max_page_list_len, 1);
2325 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
2327 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
2331 /* Index only, 18 bit long, lkey = itid << 8 | key */
2332 mr->hw_mr.tid_type = QED_RDMA_TID_FMR;
2334 mr->hw_mr.pd = pd->pd_id;
2335 mr->hw_mr.local_read = 1;
2336 mr->hw_mr.local_write = 0;
2337 mr->hw_mr.remote_read = 0;
2338 mr->hw_mr.remote_write = 0;
2339 mr->hw_mr.remote_atomic = 0;
2340 mr->hw_mr.mw_bind = false;
2341 mr->hw_mr.pbl_ptr = 0;
2342 mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
2343 mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
2345 mr->hw_mr.length = 0;
2346 mr->hw_mr.vaddr = 0;
2347 mr->hw_mr.zbva = false;
2348 mr->hw_mr.phy_mr = true;
2349 mr->hw_mr.dma_mr = false;
2351 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
2353 DP_ERR(dev, "roce register tid returned an error %d\n", rc);
2357 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2358 mr->ibmr.rkey = mr->ibmr.lkey;
2360 DP_DEBUG(dev, QEDR_MSG_MR, "alloc frmr: %x\n", mr->ibmr.lkey);
2364 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
2370 struct ib_mr *qedr_alloc_mr(struct ib_pd *ibpd,
2371 enum ib_mr_type mr_type, u32 max_num_sg)
2373 struct qedr_dev *dev;
2376 if (mr_type != IB_MR_TYPE_MEM_REG)
2377 return ERR_PTR(-EINVAL);
2379 mr = __qedr_alloc_mr(ibpd, max_num_sg);
2382 return ERR_PTR(-EINVAL);
2389 static int qedr_set_page(struct ib_mr *ibmr, u64 addr)
2391 struct qedr_mr *mr = get_qedr_mr(ibmr);
2392 struct qedr_pbl *pbl_table;
2393 struct regpair *pbe;
2396 if (unlikely(mr->npages == mr->info.pbl_info.num_pbes)) {
2397 DP_ERR(mr->dev, "qedr_set_page failes when %d\n", mr->npages);
2401 DP_DEBUG(mr->dev, QEDR_MSG_MR, "qedr_set_page pages[%d] = 0x%llx\n",
2404 pbes_in_page = mr->info.pbl_info.pbl_size / sizeof(u64);
2405 pbl_table = mr->info.pbl_table + (mr->npages / pbes_in_page);
2406 pbe = (struct regpair *)pbl_table->va;
2407 pbe += mr->npages % pbes_in_page;
2408 pbe->lo = cpu_to_le32((u32)addr);
2409 pbe->hi = cpu_to_le32((u32)upper_32_bits(addr));
2416 static void handle_completed_mrs(struct qedr_dev *dev, struct mr_info *info)
2418 int work = info->completed - info->completed_handled - 1;
2420 DP_DEBUG(dev, QEDR_MSG_MR, "Special FMR work = %d\n", work);
2421 while (work-- > 0 && !list_empty(&info->inuse_pbl_list)) {
2422 struct qedr_pbl *pbl;
2424 /* Free all the page list that are possible to be freed
2425 * (all the ones that were invalidated), under the assumption
2426 * that if an FMR was completed successfully that means that
2427 * if there was an invalidate operation before it also ended
2429 pbl = list_first_entry(&info->inuse_pbl_list,
2430 struct qedr_pbl, list_entry);
2431 list_move_tail(&pbl->list_entry, &info->free_pbl_list);
2432 info->completed_handled++;
2436 int qedr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2437 int sg_nents, unsigned int *sg_offset)
2439 struct qedr_mr *mr = get_qedr_mr(ibmr);
2443 handle_completed_mrs(mr->dev, &mr->info);
2444 return ib_sg_to_pages(ibmr, sg, sg_nents, NULL, qedr_set_page);
2447 struct ib_mr *qedr_get_dma_mr(struct ib_pd *ibpd, int acc)
2449 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
2450 struct qedr_pd *pd = get_qedr_pd(ibpd);
2454 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2456 return ERR_PTR(-ENOMEM);
2458 mr->type = QEDR_MR_DMA;
2460 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
2462 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
2466 /* index only, 18 bit long, lkey = itid << 8 | key */
2467 mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
2468 mr->hw_mr.pd = pd->pd_id;
2469 mr->hw_mr.local_read = 1;
2470 mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
2471 mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
2472 mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
2473 mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
2474 mr->hw_mr.dma_mr = true;
2476 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
2478 DP_ERR(dev, "roce register tid returned an error %d\n", rc);
2482 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2483 if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
2484 mr->hw_mr.remote_atomic)
2485 mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
2487 DP_DEBUG(dev, QEDR_MSG_MR, "get dma mr: lkey = %x\n", mr->ibmr.lkey);
2491 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
2497 static inline int qedr_wq_is_full(struct qedr_qp_hwq_info *wq)
2499 return (((wq->prod + 1) % wq->max_wr) == wq->cons);
2502 static int sge_data_len(struct ib_sge *sg_list, int num_sge)
2506 for (i = 0; i < num_sge; i++)
2507 len += sg_list[i].length;
2512 static void swap_wqe_data64(u64 *p)
2516 for (i = 0; i < QEDR_SQE_ELEMENT_SIZE / sizeof(u64); i++, p++)
2517 *p = cpu_to_be64(cpu_to_le64(*p));
2520 static u32 qedr_prepare_sq_inline_data(struct qedr_dev *dev,
2521 struct qedr_qp *qp, u8 *wqe_size,
2522 struct ib_send_wr *wr,
2523 struct ib_send_wr **bad_wr, u8 *bits,
2526 u32 data_size = sge_data_len(wr->sg_list, wr->num_sge);
2527 char *seg_prt, *wqe;
2530 if (data_size > ROCE_REQ_MAX_INLINE_DATA_SIZE) {
2531 DP_ERR(dev, "Too much inline data in WR: %d\n", data_size);
2545 /* Copy data inline */
2546 for (i = 0; i < wr->num_sge; i++) {
2547 u32 len = wr->sg_list[i].length;
2548 void *src = (void *)(uintptr_t)wr->sg_list[i].addr;
2553 /* New segment required */
2555 wqe = (char *)qed_chain_produce(&qp->sq.pbl);
2557 seg_siz = sizeof(struct rdma_sq_common_wqe);
2561 /* Calculate currently allowed length */
2562 cur = min_t(u32, len, seg_siz);
2563 memcpy(seg_prt, src, cur);
2565 /* Update segment variables */
2569 /* Update sge variables */
2573 /* Swap fully-completed segments */
2575 swap_wqe_data64((u64 *)wqe);
2579 /* swap last not completed segment */
2581 swap_wqe_data64((u64 *)wqe);
2586 #define RQ_SGE_SET(sge, vaddr, vlength, vflags) \
2588 DMA_REGPAIR_LE(sge->addr, vaddr); \
2589 (sge)->length = cpu_to_le32(vlength); \
2590 (sge)->flags = cpu_to_le32(vflags); \
2593 #define SRQ_HDR_SET(hdr, vwr_id, num_sge) \
2595 DMA_REGPAIR_LE(hdr->wr_id, vwr_id); \
2596 (hdr)->num_sges = num_sge; \
2599 #define SRQ_SGE_SET(sge, vaddr, vlength, vlkey) \
2601 DMA_REGPAIR_LE(sge->addr, vaddr); \
2602 (sge)->length = cpu_to_le32(vlength); \
2603 (sge)->l_key = cpu_to_le32(vlkey); \
2606 static u32 qedr_prepare_sq_sges(struct qedr_qp *qp, u8 *wqe_size,
2607 struct ib_send_wr *wr)
2612 for (i = 0; i < wr->num_sge; i++) {
2613 struct rdma_sq_sge *sge = qed_chain_produce(&qp->sq.pbl);
2615 DMA_REGPAIR_LE(sge->addr, wr->sg_list[i].addr);
2616 sge->l_key = cpu_to_le32(wr->sg_list[i].lkey);
2617 sge->length = cpu_to_le32(wr->sg_list[i].length);
2618 data_size += wr->sg_list[i].length;
2622 *wqe_size += wr->num_sge;
2627 static u32 qedr_prepare_sq_rdma_data(struct qedr_dev *dev,
2629 struct rdma_sq_rdma_wqe_1st *rwqe,
2630 struct rdma_sq_rdma_wqe_2nd *rwqe2,
2631 struct ib_send_wr *wr,
2632 struct ib_send_wr **bad_wr)
2634 rwqe2->r_key = cpu_to_le32(rdma_wr(wr)->rkey);
2635 DMA_REGPAIR_LE(rwqe2->remote_va, rdma_wr(wr)->remote_addr);
2637 if (wr->send_flags & IB_SEND_INLINE) {
2640 SET_FIELD2(flags, RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG, 1);
2641 return qedr_prepare_sq_inline_data(dev, qp, &rwqe->wqe_size, wr,
2642 bad_wr, &rwqe->flags, flags);
2645 return qedr_prepare_sq_sges(qp, &rwqe->wqe_size, wr);
2648 static u32 qedr_prepare_sq_send_data(struct qedr_dev *dev,
2650 struct rdma_sq_send_wqe_1st *swqe,
2651 struct rdma_sq_send_wqe_2st *swqe2,
2652 struct ib_send_wr *wr,
2653 struct ib_send_wr **bad_wr)
2655 memset(swqe2, 0, sizeof(*swqe2));
2656 if (wr->send_flags & IB_SEND_INLINE) {
2659 SET_FIELD2(flags, RDMA_SQ_SEND_WQE_INLINE_FLG, 1);
2660 return qedr_prepare_sq_inline_data(dev, qp, &swqe->wqe_size, wr,
2661 bad_wr, &swqe->flags, flags);
2664 return qedr_prepare_sq_sges(qp, &swqe->wqe_size, wr);
2667 static int qedr_prepare_reg(struct qedr_qp *qp,
2668 struct rdma_sq_fmr_wqe_1st *fwqe1,
2669 struct ib_reg_wr *wr)
2671 struct qedr_mr *mr = get_qedr_mr(wr->mr);
2672 struct rdma_sq_fmr_wqe_2nd *fwqe2;
2674 fwqe2 = (struct rdma_sq_fmr_wqe_2nd *)qed_chain_produce(&qp->sq.pbl);
2675 fwqe1->addr.hi = upper_32_bits(mr->ibmr.iova);
2676 fwqe1->addr.lo = lower_32_bits(mr->ibmr.iova);
2677 fwqe1->l_key = wr->key;
2679 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_READ,
2680 !!(wr->access & IB_ACCESS_REMOTE_READ));
2681 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE,
2682 !!(wr->access & IB_ACCESS_REMOTE_WRITE));
2683 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC,
2684 !!(wr->access & IB_ACCESS_REMOTE_ATOMIC));
2685 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_READ, 1);
2686 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE,
2687 !!(wr->access & IB_ACCESS_LOCAL_WRITE));
2688 fwqe2->fmr_ctrl = 0;
2690 SET_FIELD2(fwqe2->fmr_ctrl, RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG,
2691 ilog2(mr->ibmr.page_size) - 12);
2693 fwqe2->length_hi = 0;
2694 fwqe2->length_lo = mr->ibmr.length;
2695 fwqe2->pbl_addr.hi = upper_32_bits(mr->info.pbl_table->pa);
2696 fwqe2->pbl_addr.lo = lower_32_bits(mr->info.pbl_table->pa);
2698 qp->wqe_wr_id[qp->sq.prod].mr = mr;
2703 enum ib_wc_opcode qedr_ib_to_wc_opcode(enum ib_wr_opcode opcode)
2706 case IB_WR_RDMA_WRITE:
2707 case IB_WR_RDMA_WRITE_WITH_IMM:
2708 return IB_WC_RDMA_WRITE;
2709 case IB_WR_SEND_WITH_IMM:
2711 case IB_WR_SEND_WITH_INV:
2713 case IB_WR_RDMA_READ:
2714 return IB_WC_RDMA_READ;
2715 case IB_WR_ATOMIC_CMP_AND_SWP:
2716 return IB_WC_COMP_SWAP;
2717 case IB_WR_ATOMIC_FETCH_AND_ADD:
2718 return IB_WC_FETCH_ADD;
2720 return IB_WC_REG_MR;
2721 case IB_WR_LOCAL_INV:
2722 return IB_WC_LOCAL_INV;
2728 inline bool qedr_can_post_send(struct qedr_qp *qp, struct ib_send_wr *wr)
2730 int wq_is_full, err_wr, pbl_is_full;
2731 struct qedr_dev *dev = qp->dev;
2733 /* prevent SQ overflow and/or processing of a bad WR */
2734 err_wr = wr->num_sge > qp->sq.max_sges;
2735 wq_is_full = qedr_wq_is_full(&qp->sq);
2736 pbl_is_full = qed_chain_get_elem_left_u32(&qp->sq.pbl) <
2737 QEDR_MAX_SQE_ELEMENTS_PER_SQE;
2738 if (wq_is_full || err_wr || pbl_is_full) {
2739 if (wq_is_full && !(qp->err_bitmap & QEDR_QP_ERR_SQ_FULL)) {
2741 "error: WQ is full. Post send on QP %p failed (this error appears only once)\n",
2743 qp->err_bitmap |= QEDR_QP_ERR_SQ_FULL;
2746 if (err_wr && !(qp->err_bitmap & QEDR_QP_ERR_BAD_SR)) {
2748 "error: WR is bad. Post send on QP %p failed (this error appears only once)\n",
2750 qp->err_bitmap |= QEDR_QP_ERR_BAD_SR;
2754 !(qp->err_bitmap & QEDR_QP_ERR_SQ_PBL_FULL)) {
2756 "error: WQ PBL is full. Post send on QP %p failed (this error appears only once)\n",
2758 qp->err_bitmap |= QEDR_QP_ERR_SQ_PBL_FULL;
2765 int __qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2766 struct ib_send_wr **bad_wr)
2768 struct qedr_dev *dev = get_qedr_dev(ibqp->device);
2769 struct qedr_qp *qp = get_qedr_qp(ibqp);
2770 struct rdma_sq_atomic_wqe_1st *awqe1;
2771 struct rdma_sq_atomic_wqe_2nd *awqe2;
2772 struct rdma_sq_atomic_wqe_3rd *awqe3;
2773 struct rdma_sq_send_wqe_2st *swqe2;
2774 struct rdma_sq_local_inv_wqe *iwqe;
2775 struct rdma_sq_rdma_wqe_2nd *rwqe2;
2776 struct rdma_sq_send_wqe_1st *swqe;
2777 struct rdma_sq_rdma_wqe_1st *rwqe;
2778 struct rdma_sq_fmr_wqe_1st *fwqe1;
2779 struct rdma_sq_common_wqe *wqe;
2784 if (!qedr_can_post_send(qp, wr)) {
2789 wqe = qed_chain_produce(&qp->sq.pbl);
2790 qp->wqe_wr_id[qp->sq.prod].signaled =
2791 !!(wr->send_flags & IB_SEND_SIGNALED) || qp->signaled;
2794 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_SE_FLG,
2795 !!(wr->send_flags & IB_SEND_SOLICITED));
2796 comp = (!!(wr->send_flags & IB_SEND_SIGNALED)) || qp->signaled;
2797 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_COMP_FLG, comp);
2798 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_RD_FENCE_FLG,
2799 !!(wr->send_flags & IB_SEND_FENCE));
2800 wqe->prev_wqe_size = qp->prev_wqe_size;
2802 qp->wqe_wr_id[qp->sq.prod].opcode = qedr_ib_to_wc_opcode(wr->opcode);
2804 switch (wr->opcode) {
2805 case IB_WR_SEND_WITH_IMM:
2806 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_IMM;
2807 swqe = (struct rdma_sq_send_wqe_1st *)wqe;
2809 swqe2 = qed_chain_produce(&qp->sq.pbl);
2811 swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.imm_data);
2812 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
2814 swqe->length = cpu_to_le32(length);
2815 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
2816 qp->prev_wqe_size = swqe->wqe_size;
2817 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
2820 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND;
2821 swqe = (struct rdma_sq_send_wqe_1st *)wqe;
2824 swqe2 = qed_chain_produce(&qp->sq.pbl);
2825 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
2827 swqe->length = cpu_to_le32(length);
2828 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
2829 qp->prev_wqe_size = swqe->wqe_size;
2830 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
2832 case IB_WR_SEND_WITH_INV:
2833 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_INVALIDATE;
2834 swqe = (struct rdma_sq_send_wqe_1st *)wqe;
2835 swqe2 = qed_chain_produce(&qp->sq.pbl);
2837 swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.invalidate_rkey);
2838 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
2840 swqe->length = cpu_to_le32(length);
2841 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
2842 qp->prev_wqe_size = swqe->wqe_size;
2843 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
2846 case IB_WR_RDMA_WRITE_WITH_IMM:
2847 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR_WITH_IMM;
2848 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
2851 rwqe->imm_data = htonl(cpu_to_le32(wr->ex.imm_data));
2852 rwqe2 = qed_chain_produce(&qp->sq.pbl);
2853 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
2855 rwqe->length = cpu_to_le32(length);
2856 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
2857 qp->prev_wqe_size = rwqe->wqe_size;
2858 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
2860 case IB_WR_RDMA_WRITE:
2861 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR;
2862 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
2865 rwqe2 = qed_chain_produce(&qp->sq.pbl);
2866 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
2868 rwqe->length = cpu_to_le32(length);
2869 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
2870 qp->prev_wqe_size = rwqe->wqe_size;
2871 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
2873 case IB_WR_RDMA_READ_WITH_INV:
2875 "RDMA READ WITH INVALIDATE not supported\n");
2880 case IB_WR_RDMA_READ:
2881 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_RD;
2882 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
2885 rwqe2 = qed_chain_produce(&qp->sq.pbl);
2886 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
2888 rwqe->length = cpu_to_le32(length);
2889 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
2890 qp->prev_wqe_size = rwqe->wqe_size;
2891 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
2894 case IB_WR_ATOMIC_CMP_AND_SWP:
2895 case IB_WR_ATOMIC_FETCH_AND_ADD:
2896 awqe1 = (struct rdma_sq_atomic_wqe_1st *)wqe;
2897 awqe1->wqe_size = 4;
2899 awqe2 = qed_chain_produce(&qp->sq.pbl);
2900 DMA_REGPAIR_LE(awqe2->remote_va, atomic_wr(wr)->remote_addr);
2901 awqe2->r_key = cpu_to_le32(atomic_wr(wr)->rkey);
2903 awqe3 = qed_chain_produce(&qp->sq.pbl);
2905 if (wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2906 wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_ADD;
2907 DMA_REGPAIR_LE(awqe3->swap_data,
2908 atomic_wr(wr)->compare_add);
2910 wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_CMP_AND_SWAP;
2911 DMA_REGPAIR_LE(awqe3->swap_data,
2912 atomic_wr(wr)->swap);
2913 DMA_REGPAIR_LE(awqe3->cmp_data,
2914 atomic_wr(wr)->compare_add);
2917 qedr_prepare_sq_sges(qp, NULL, wr);
2919 qp->wqe_wr_id[qp->sq.prod].wqe_size = awqe1->wqe_size;
2920 qp->prev_wqe_size = awqe1->wqe_size;
2923 case IB_WR_LOCAL_INV:
2924 iwqe = (struct rdma_sq_local_inv_wqe *)wqe;
2927 iwqe->req_type = RDMA_SQ_REQ_TYPE_LOCAL_INVALIDATE;
2928 iwqe->inv_l_key = wr->ex.invalidate_rkey;
2929 qp->wqe_wr_id[qp->sq.prod].wqe_size = iwqe->wqe_size;
2930 qp->prev_wqe_size = iwqe->wqe_size;
2933 DP_DEBUG(dev, QEDR_MSG_CQ, "REG_MR\n");
2934 wqe->req_type = RDMA_SQ_REQ_TYPE_FAST_MR;
2935 fwqe1 = (struct rdma_sq_fmr_wqe_1st *)wqe;
2936 fwqe1->wqe_size = 2;
2938 rc = qedr_prepare_reg(qp, fwqe1, reg_wr(wr));
2940 DP_ERR(dev, "IB_REG_MR failed rc=%d\n", rc);
2945 qp->wqe_wr_id[qp->sq.prod].wqe_size = fwqe1->wqe_size;
2946 qp->prev_wqe_size = fwqe1->wqe_size;
2949 DP_ERR(dev, "invalid opcode 0x%x!\n", wr->opcode);
2958 /* Restore prod to its position before
2959 * this WR was processed
2961 value = le16_to_cpu(qp->sq.db_data.data.value);
2962 qed_chain_set_prod(&qp->sq.pbl, value, wqe);
2964 /* Restore prev_wqe_size */
2965 qp->prev_wqe_size = wqe->prev_wqe_size;
2967 DP_ERR(dev, "POST SEND FAILED\n");
2973 int qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2974 struct ib_send_wr **bad_wr)
2976 struct qedr_dev *dev = get_qedr_dev(ibqp->device);
2977 struct qedr_qp *qp = get_qedr_qp(ibqp);
2978 unsigned long flags;
2983 if (qp->qp_type == IB_QPT_GSI)
2984 return qedr_gsi_post_send(ibqp, wr, bad_wr);
2986 spin_lock_irqsave(&qp->q_lock, flags);
2988 if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
2989 (qp->state == QED_ROCE_QP_STATE_ERR)) {
2990 spin_unlock_irqrestore(&qp->q_lock, flags);
2992 DP_DEBUG(dev, QEDR_MSG_CQ,
2993 "QP in wrong state! QP icid=0x%x state %d\n",
2994 qp->icid, qp->state);
2999 rc = __qedr_post_send(ibqp, wr, bad_wr);
3003 qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id;
3005 qedr_inc_sw_prod(&qp->sq);
3007 qp->sq.db_data.data.value++;
3013 * If there was a failure in the first WR then it will be triggered in
3014 * vane. However this is not harmful (as long as the producer value is
3015 * unchanged). For performance reasons we avoid checking for this
3016 * redundant doorbell.
3019 writel(qp->sq.db_data.raw, qp->sq.db);
3021 /* Make sure write sticks */
3024 spin_unlock_irqrestore(&qp->q_lock, flags);
3029 int qedr_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3030 struct ib_recv_wr **bad_wr)
3032 struct qedr_qp *qp = get_qedr_qp(ibqp);
3033 struct qedr_dev *dev = qp->dev;
3034 unsigned long flags;
3037 if (qp->qp_type == IB_QPT_GSI)
3038 return qedr_gsi_post_recv(ibqp, wr, bad_wr);
3040 spin_lock_irqsave(&qp->q_lock, flags);
3042 if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
3043 (qp->state == QED_ROCE_QP_STATE_ERR)) {
3044 spin_unlock_irqrestore(&qp->q_lock, flags);
3052 if (qed_chain_get_elem_left_u32(&qp->rq.pbl) <
3053 QEDR_MAX_RQE_ELEMENTS_PER_RQE ||
3054 wr->num_sge > qp->rq.max_sges) {
3055 DP_ERR(dev, "Can't post WR (%d < %d) || (%d > %d)\n",
3056 qed_chain_get_elem_left_u32(&qp->rq.pbl),
3057 QEDR_MAX_RQE_ELEMENTS_PER_RQE, wr->num_sge,
3063 for (i = 0; i < wr->num_sge; i++) {
3065 struct rdma_rq_sge *rqe =
3066 qed_chain_produce(&qp->rq.pbl);
3068 /* First one must include the number
3069 * of SGE in the list
3072 SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES,
3075 SET_FIELD(flags, RDMA_RQ_SGE_L_KEY,
3076 wr->sg_list[i].lkey);
3078 RQ_SGE_SET(rqe, wr->sg_list[i].addr,
3079 wr->sg_list[i].length, flags);
3082 /* Special case of no sges. FW requires between 1-4 sges...
3083 * in this case we need to post 1 sge with length zero. this is
3084 * because rdma write with immediate consumes an RQ.
3088 struct rdma_rq_sge *rqe =
3089 qed_chain_produce(&qp->rq.pbl);
3091 /* First one must include the number
3092 * of SGE in the list
3094 SET_FIELD(flags, RDMA_RQ_SGE_L_KEY, 0);
3095 SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, 1);
3097 RQ_SGE_SET(rqe, 0, 0, flags);
3101 qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id;
3102 qp->rqe_wr_id[qp->rq.prod].wqe_size = i;
3104 qedr_inc_sw_prod(&qp->rq);
3106 /* Flush all the writes before signalling doorbell */
3109 qp->rq.db_data.data.value++;
3111 writel(qp->rq.db_data.raw, qp->rq.db);
3113 /* Make sure write sticks */
3119 spin_unlock_irqrestore(&qp->q_lock, flags);
3124 static int is_valid_cqe(struct qedr_cq *cq, union rdma_cqe *cqe)
3126 struct rdma_cqe_requester *resp_cqe = &cqe->req;
3128 return (resp_cqe->flags & RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK) ==
3132 static struct qedr_qp *cqe_get_qp(union rdma_cqe *cqe)
3134 struct rdma_cqe_requester *resp_cqe = &cqe->req;
3137 qp = (struct qedr_qp *)(uintptr_t)HILO_GEN(resp_cqe->qp_handle.hi,
3138 resp_cqe->qp_handle.lo,
3143 static enum rdma_cqe_type cqe_get_type(union rdma_cqe *cqe)
3145 struct rdma_cqe_requester *resp_cqe = &cqe->req;
3147 return GET_FIELD(resp_cqe->flags, RDMA_CQE_REQUESTER_TYPE);
3150 /* Return latest CQE (needs processing) */
3151 static union rdma_cqe *get_cqe(struct qedr_cq *cq)
3153 return cq->latest_cqe;
3156 /* In fmr we need to increase the number of fmr completed counter for the fmr
3157 * algorithm determining whether we can free a pbl or not.
3158 * we need to perform this whether the work request was signaled or not. for
3159 * this purpose we call this function from the condition that checks if a wr
3160 * should be skipped, to make sure we don't miss it ( possibly this fmr
3161 * operation was not signalted)
3163 static inline void qedr_chk_if_fmr(struct qedr_qp *qp)
3165 if (qp->wqe_wr_id[qp->sq.cons].opcode == IB_WC_REG_MR)
3166 qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
3169 static int process_req(struct qedr_dev *dev, struct qedr_qp *qp,
3170 struct qedr_cq *cq, int num_entries,
3171 struct ib_wc *wc, u16 hw_cons, enum ib_wc_status status,
3176 while (num_entries && qp->sq.wqe_cons != hw_cons) {
3177 if (!qp->wqe_wr_id[qp->sq.cons].signaled && !force) {
3178 qedr_chk_if_fmr(qp);
3184 wc->status = status;
3186 wc->src_qp = qp->id;
3189 wc->wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id;
3190 wc->opcode = qp->wqe_wr_id[qp->sq.cons].opcode;
3192 switch (wc->opcode) {
3193 case IB_WC_RDMA_WRITE:
3194 wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len;
3196 case IB_WC_COMP_SWAP:
3197 case IB_WC_FETCH_ADD:
3201 qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
3211 while (qp->wqe_wr_id[qp->sq.cons].wqe_size--)
3212 qed_chain_consume(&qp->sq.pbl);
3213 qedr_inc_sw_cons(&qp->sq);
3219 static int qedr_poll_cq_req(struct qedr_dev *dev,
3220 struct qedr_qp *qp, struct qedr_cq *cq,
3221 int num_entries, struct ib_wc *wc,
3222 struct rdma_cqe_requester *req)
3226 switch (req->status) {
3227 case RDMA_CQE_REQ_STS_OK:
3228 cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
3231 case RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR:
3233 "Error: POLL CQ with RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3234 cq->icid, qp->icid);
3235 cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
3236 IB_WC_WR_FLUSH_ERR, 0);
3239 /* process all WQE before the cosumer */
3240 qp->state = QED_ROCE_QP_STATE_ERR;
3241 cnt = process_req(dev, qp, cq, num_entries, wc,
3242 req->sq_cons - 1, IB_WC_SUCCESS, 0);
3244 /* if we have extra WC fill it with actual error info */
3245 if (cnt < num_entries) {
3246 enum ib_wc_status wc_status;
3248 switch (req->status) {
3249 case RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR:
3251 "Error: POLL CQ with RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3252 cq->icid, qp->icid);
3253 wc_status = IB_WC_BAD_RESP_ERR;
3255 case RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR:
3257 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3258 cq->icid, qp->icid);
3259 wc_status = IB_WC_LOC_LEN_ERR;
3261 case RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR:
3263 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3264 cq->icid, qp->icid);
3265 wc_status = IB_WC_LOC_QP_OP_ERR;
3267 case RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR:
3269 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3270 cq->icid, qp->icid);
3271 wc_status = IB_WC_LOC_PROT_ERR;
3273 case RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR:
3275 "Error: POLL CQ with RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3276 cq->icid, qp->icid);
3277 wc_status = IB_WC_MW_BIND_ERR;
3279 case RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR:
3281 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3282 cq->icid, qp->icid);
3283 wc_status = IB_WC_REM_INV_REQ_ERR;
3285 case RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR:
3287 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3288 cq->icid, qp->icid);
3289 wc_status = IB_WC_REM_ACCESS_ERR;
3291 case RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR:
3293 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3294 cq->icid, qp->icid);
3295 wc_status = IB_WC_REM_OP_ERR;
3297 case RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR:
3299 "Error: POLL CQ with RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3300 cq->icid, qp->icid);
3301 wc_status = IB_WC_RNR_RETRY_EXC_ERR;
3303 case RDMA_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR:
3305 "Error: POLL CQ with ROCE_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3306 cq->icid, qp->icid);
3307 wc_status = IB_WC_RETRY_EXC_ERR;
3311 "Error: POLL CQ with IB_WC_GENERAL_ERR. CQ icid=0x%x, QP icid=0x%x\n",
3312 cq->icid, qp->icid);
3313 wc_status = IB_WC_GENERAL_ERR;
3315 cnt += process_req(dev, qp, cq, 1, wc, req->sq_cons,
3323 static void __process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
3324 struct qedr_cq *cq, struct ib_wc *wc,
3325 struct rdma_cqe_responder *resp, u64 wr_id)
3327 enum ib_wc_status wc_status = IB_WC_SUCCESS;
3330 wc->opcode = IB_WC_RECV;
3333 switch (resp->status) {
3334 case RDMA_CQE_RESP_STS_LOCAL_ACCESS_ERR:
3335 wc_status = IB_WC_LOC_ACCESS_ERR;
3337 case RDMA_CQE_RESP_STS_LOCAL_LENGTH_ERR:
3338 wc_status = IB_WC_LOC_LEN_ERR;
3340 case RDMA_CQE_RESP_STS_LOCAL_QP_OPERATION_ERR:
3341 wc_status = IB_WC_LOC_QP_OP_ERR;
3343 case RDMA_CQE_RESP_STS_LOCAL_PROTECTION_ERR:
3344 wc_status = IB_WC_LOC_PROT_ERR;
3346 case RDMA_CQE_RESP_STS_MEMORY_MGT_OPERATION_ERR:
3347 wc_status = IB_WC_MW_BIND_ERR;
3349 case RDMA_CQE_RESP_STS_REMOTE_INVALID_REQUEST_ERR:
3350 wc_status = IB_WC_REM_INV_RD_REQ_ERR;
3352 case RDMA_CQE_RESP_STS_OK:
3353 wc_status = IB_WC_SUCCESS;
3354 wc->byte_len = le32_to_cpu(resp->length);
3356 flags = resp->flags & QEDR_RESP_RDMA_IMM;
3358 if (flags == QEDR_RESP_RDMA_IMM)
3359 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3361 if (flags == QEDR_RESP_RDMA_IMM || flags == QEDR_RESP_IMM) {
3363 le32_to_cpu(resp->imm_data_or_inv_r_Key);
3364 wc->wc_flags |= IB_WC_WITH_IMM;
3368 wc->status = IB_WC_GENERAL_ERR;
3369 DP_ERR(dev, "Invalid CQE status detected\n");
3373 wc->status = wc_status;
3374 wc->src_qp = qp->id;
3379 static int process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
3380 struct qedr_cq *cq, struct ib_wc *wc,
3381 struct rdma_cqe_responder *resp)
3383 u64 wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
3385 __process_resp_one(dev, qp, cq, wc, resp, wr_id);
3387 while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
3388 qed_chain_consume(&qp->rq.pbl);
3389 qedr_inc_sw_cons(&qp->rq);
3394 static int process_resp_flush(struct qedr_qp *qp, struct qedr_cq *cq,
3395 int num_entries, struct ib_wc *wc, u16 hw_cons)
3399 while (num_entries && qp->rq.wqe_cons != hw_cons) {
3401 wc->status = IB_WC_WR_FLUSH_ERR;
3403 wc->src_qp = qp->id;
3405 wc->wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
3410 while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
3411 qed_chain_consume(&qp->rq.pbl);
3412 qedr_inc_sw_cons(&qp->rq);
3418 static void try_consume_resp_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
3419 struct rdma_cqe_responder *resp, int *update)
3421 if (le16_to_cpu(resp->rq_cons) == qp->rq.wqe_cons) {
3427 static int qedr_poll_cq_resp(struct qedr_dev *dev, struct qedr_qp *qp,
3428 struct qedr_cq *cq, int num_entries,
3429 struct ib_wc *wc, struct rdma_cqe_responder *resp,
3434 if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
3435 cnt = process_resp_flush(qp, cq, num_entries, wc,
3437 try_consume_resp_cqe(cq, qp, resp, update);
3439 cnt = process_resp_one(dev, qp, cq, wc, resp);
3447 static void try_consume_req_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
3448 struct rdma_cqe_requester *req, int *update)
3450 if (le16_to_cpu(req->sq_cons) == qp->sq.wqe_cons) {
3456 int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
3458 struct qedr_dev *dev = get_qedr_dev(ibcq->device);
3459 struct qedr_cq *cq = get_qedr_cq(ibcq);
3460 union rdma_cqe *cqe = cq->latest_cqe;
3461 u32 old_cons, new_cons;
3462 unsigned long flags;
3466 if (cq->cq_type == QEDR_CQ_TYPE_GSI)
3467 return qedr_gsi_poll_cq(ibcq, num_entries, wc);
3469 spin_lock_irqsave(&cq->cq_lock, flags);
3470 old_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
3471 while (num_entries && is_valid_cqe(cq, cqe)) {
3475 /* prevent speculative reads of any field of CQE */
3478 qp = cqe_get_qp(cqe);
3480 WARN(1, "Error: CQE QP pointer is NULL. CQE=%p\n", cqe);
3486 switch (cqe_get_type(cqe)) {
3487 case RDMA_CQE_TYPE_REQUESTER:
3488 cnt = qedr_poll_cq_req(dev, qp, cq, num_entries, wc,
3490 try_consume_req_cqe(cq, qp, &cqe->req, &update);
3492 case RDMA_CQE_TYPE_RESPONDER_RQ:
3493 cnt = qedr_poll_cq_resp(dev, qp, cq, num_entries, wc,
3494 &cqe->resp, &update);
3496 case RDMA_CQE_TYPE_INVALID:
3498 DP_ERR(dev, "Error: invalid CQE type = %d\n",
3507 new_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
3509 cq->cq_cons += new_cons - old_cons;
3512 /* doorbell notifies abount latest VALID entry,
3513 * but chain already point to the next INVALID one
3515 doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags);
3517 spin_unlock_irqrestore(&cq->cq_lock, flags);
3521 int qedr_process_mad(struct ib_device *ibdev, int process_mad_flags,
3523 const struct ib_wc *in_wc,
3524 const struct ib_grh *in_grh,
3525 const struct ib_mad_hdr *mad_hdr,
3526 size_t in_mad_size, struct ib_mad_hdr *out_mad,
3527 size_t *out_mad_size, u16 *out_mad_pkey_index)
3529 struct qedr_dev *dev = get_qedr_dev(ibdev);
3531 DP_DEBUG(dev, QEDR_MSG_GSI,
3532 "QEDR_PROCESS_MAD in_mad %x %x %x %x %x %x %x %x\n",
3533 mad_hdr->attr_id, mad_hdr->base_version, mad_hdr->attr_mod,
3534 mad_hdr->class_specific, mad_hdr->class_version,
3535 mad_hdr->method, mad_hdr->mgmt_class, mad_hdr->status);
3536 return IB_MAD_RESULT_SUCCESS;
3539 int qedr_port_immutable(struct ib_device *ibdev, u8 port_num,
3540 struct ib_port_immutable *immutable)
3542 struct ib_port_attr attr;
3545 err = qedr_query_port(ibdev, port_num, &attr);
3549 immutable->pkey_tbl_len = attr.pkey_tbl_len;
3550 immutable->gid_tbl_len = attr.gid_tbl_len;
3551 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE |
3552 RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3553 immutable->max_mad_size = IB_MGMT_MAD_SIZE;