RDMA/vmw_pvrdma: Add shared receive queue support

Add the required functions needed to support SRQs. Currently, kernel
clients are not supported. SRQs will only be available in userspace.
Reviewed-by: NAdit Ranadive <aditr@vmware.com>
Reviewed-by: NAditya Sarwade <asarwade@vmware.com>
Reviewed-by: NJorgen Hansen <jhansen@vmware.com>
Reviewed-by: NNitish Bhat <bnitish@vmware.com>
Signed-off-by: NBryan Tan <bryantan@vmware.com>
Reviewed-by: NYuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

drivers/infiniband/hw/vmw_pvrdma/Makefile

 obj-$(CONFIG_INFINIBAND_VMWARE_PVRDMA) += vmw_pvrdma.o
 
-vmw_pvrdma-y := pvrdma_cmd.o pvrdma_cq.o pvrdma_doorbell.o pvrdma_main.o pvrdma_misc.o pvrdma_mr.o pvrdma_qp.o pvrdma_verbs.o
+vmw_pvrdma-y := pvrdma_cmd.o pvrdma_cq.o pvrdma_doorbell.o pvrdma_main.o pvrdma_misc.o pvrdma_mr.o pvrdma_qp.o pvrdma_srq.o pvrdma_verbs.o

drivers/infiniband/hw/vmw_pvrdma/pvrdma.h

@@ -162,6 +162,22 @@ struct pvrdma_ah {
 	struct pvrdma_av av;
 };
 
+struct pvrdma_srq {
+	struct ib_srq ibsrq;
+	int offset;
+	spinlock_t lock; /* SRQ lock. */
+	int wqe_cnt;
+	int wqe_size;
+	int max_gs;
+	struct ib_umem *umem;
+	struct pvrdma_ring_state *ring;
+	struct pvrdma_page_dir pdir;
+	u32 srq_handle;
+	int npages;
+	refcount_t refcnt;
+	wait_queue_head_t wait;
+};
+
 struct pvrdma_qp {
 	struct ib_qp ibqp;
 	u32 qp_handle;
@@ -171,6 +187,7 @@ struct pvrdma_qp {
 	struct ib_umem *rumem;
 	struct ib_umem *sumem;
 	struct pvrdma_page_dir pdir;
+	struct pvrdma_srq *srq;
 	int npages;
 	int npages_send;
 	int npages_recv;
@@ -210,6 +227,8 @@ struct pvrdma_dev {
 	struct pvrdma_page_dir cq_pdir;
 	struct pvrdma_cq **cq_tbl;
 	spinlock_t cq_tbl_lock;
+	struct pvrdma_srq **srq_tbl;
+	spinlock_t srq_tbl_lock;
 	struct pvrdma_qp **qp_tbl;
 	spinlock_t qp_tbl_lock;
 	struct pvrdma_uar_table uar_table;
@@ -221,6 +240,7 @@ struct pvrdma_dev {
 	bool ib_active;
 	atomic_t num_qps;
 	atomic_t num_cqs;
+	atomic_t num_srqs;
 	atomic_t num_pds;
 	atomic_t num_ahs;
 
@@ -256,6 +276,11 @@ static inline struct pvrdma_cq *to_vcq(struct ib_cq *ibcq)
 	return container_of(ibcq, struct pvrdma_cq, ibcq);
 }
 
+static inline struct pvrdma_srq *to_vsrq(struct ib_srq *ibsrq)
+{
+	return container_of(ibsrq, struct pvrdma_srq, ibsrq);
+}
+
 static inline struct pvrdma_user_mr *to_vmr(struct ib_mr *ibmr)
 {
 	return container_of(ibmr, struct pvrdma_user_mr, ibmr);

drivers/infiniband/hw/vmw_pvrdma/pvrdma_dev_api.h

@@ -339,6 +339,10 @@ enum {
 	PVRDMA_CMD_DESTROY_UC,
 	PVRDMA_CMD_CREATE_BIND,
 	PVRDMA_CMD_DESTROY_BIND,
+	PVRDMA_CMD_CREATE_SRQ,
+	PVRDMA_CMD_MODIFY_SRQ,
+	PVRDMA_CMD_QUERY_SRQ,
+	PVRDMA_CMD_DESTROY_SRQ,
 	PVRDMA_CMD_MAX,
 };
 
@@ -361,6 +365,10 @@ enum {
 	PVRDMA_CMD_DESTROY_UC_RESP_NOOP,
 	PVRDMA_CMD_CREATE_BIND_RESP_NOOP,
 	PVRDMA_CMD_DESTROY_BIND_RESP_NOOP,
+	PVRDMA_CMD_CREATE_SRQ_RESP,
+	PVRDMA_CMD_MODIFY_SRQ_RESP,
+	PVRDMA_CMD_QUERY_SRQ_RESP,
+	PVRDMA_CMD_DESTROY_SRQ_RESP,
 	PVRDMA_CMD_MAX_RESP,
 };
 
@@ -495,6 +503,46 @@ struct pvrdma_cmd_destroy_cq {
 	u8 reserved[4];
 };
 
+struct pvrdma_cmd_create_srq {
+	struct pvrdma_cmd_hdr hdr;
+	u64 pdir_dma;
+	u32 pd_handle;
+	u32 nchunks;
+	struct pvrdma_srq_attr attrs;
+	u8 srq_type;
+	u8 reserved[7];
+};
+
+struct pvrdma_cmd_create_srq_resp {
+	struct pvrdma_cmd_resp_hdr hdr;
+	u32 srqn;
+	u8 reserved[4];
+};
+
+struct pvrdma_cmd_modify_srq {
+	struct pvrdma_cmd_hdr hdr;
+	u32 srq_handle;
+	u32 attr_mask;
+	struct pvrdma_srq_attr attrs;
+};
+
+struct pvrdma_cmd_query_srq {
+	struct pvrdma_cmd_hdr hdr;
+	u32 srq_handle;
+	u8 reserved[4];
+};
+
+struct pvrdma_cmd_query_srq_resp {
+	struct pvrdma_cmd_resp_hdr hdr;
+	struct pvrdma_srq_attr attrs;
+};
+
+struct pvrdma_cmd_destroy_srq {
+	struct pvrdma_cmd_hdr hdr;
+	u32 srq_handle;
+	u8 reserved[4];
+};
+
 struct pvrdma_cmd_create_qp {
 	struct pvrdma_cmd_hdr hdr;
 	u64 pdir_dma;
@@ -594,6 +642,10 @@ union pvrdma_cmd_req {
 	struct pvrdma_cmd_destroy_qp destroy_qp;
 	struct pvrdma_cmd_create_bind create_bind;
 	struct pvrdma_cmd_destroy_bind destroy_bind;
+	struct pvrdma_cmd_create_srq create_srq;
+	struct pvrdma_cmd_modify_srq modify_srq;
+	struct pvrdma_cmd_query_srq query_srq;
+	struct pvrdma_cmd_destroy_srq destroy_srq;
 };
 
 union pvrdma_cmd_resp {
@@ -608,6 +660,8 @@ union pvrdma_cmd_resp {
 	struct pvrdma_cmd_create_qp_resp create_qp_resp;
 	struct pvrdma_cmd_query_qp_resp query_qp_resp;
 	struct pvrdma_cmd_destroy_qp_resp destroy_qp_resp;
+	struct pvrdma_cmd_create_srq_resp create_srq_resp;
+	struct pvrdma_cmd_query_srq_resp query_srq_resp;
 };
 
 #endif /* __PVRDMA_DEV_API_H__ */

drivers/infiniband/hw/vmw_pvrdma/pvrdma_main.c

@@ -118,6 +118,7 @@ static int pvrdma_init_device(struct pvrdma_dev *dev)
 	spin_lock_init(&dev->cmd_lock);
 	sema_init(&dev->cmd_sema, 1);
 	atomic_set(&dev->num_qps, 0);
+	atomic_set(&dev->num_srqs, 0);
 	atomic_set(&dev->num_cqs, 0);
 	atomic_set(&dev->num_pds, 0);
 	atomic_set(&dev->num_ahs, 0);
@@ -254,9 +255,32 @@ static int pvrdma_register_device(struct pvrdma_dev *dev)
 		goto err_cq_free;
 	spin_lock_init(&dev->qp_tbl_lock);
 
+	/* Check if SRQ is supported by backend */
+	if (dev->dsr->caps.max_srq) {
+		dev->ib_dev.uverbs_cmd_mask |=
+			(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)	|
+			(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)	|
+			(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)	|
+			(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)	|
+			(1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
+
+		dev->ib_dev.create_srq = pvrdma_create_srq;
+		dev->ib_dev.modify_srq = pvrdma_modify_srq;
+		dev->ib_dev.query_srq = pvrdma_query_srq;
+		dev->ib_dev.destroy_srq = pvrdma_destroy_srq;
+		dev->ib_dev.post_srq_recv = pvrdma_post_srq_recv;
+
+		dev->srq_tbl = kcalloc(dev->dsr->caps.max_srq,
+				       sizeof(struct pvrdma_srq *),
+				       GFP_KERNEL);
+		if (!dev->srq_tbl)
+			goto err_qp_free;
+	}
+	spin_lock_init(&dev->srq_tbl_lock);
+
 	ret = ib_register_device(&dev->ib_dev, NULL);
 	if (ret)
-		goto err_qp_free;
+		goto err_srq_free;
 
 	for (i = 0; i < ARRAY_SIZE(pvrdma_class_attributes); ++i) {
 		ret = device_create_file(&dev->ib_dev.dev,
@@ -271,6 +295,8 @@ static int pvrdma_register_device(struct pvrdma_dev *dev)
 
 err_class:
 	ib_unregister_device(&dev->ib_dev);
+err_srq_free:
+	kfree(dev->srq_tbl);
 err_qp_free:
 	kfree(dev->qp_tbl);
 err_cq_free:
@@ -353,6 +379,35 @@ static void pvrdma_cq_event(struct pvrdma_dev *dev, u32 cqn, int type)
 	}
 }
 
+static void pvrdma_srq_event(struct pvrdma_dev *dev, u32 srqn, int type)
+{
+	struct pvrdma_srq *srq;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dev->srq_tbl_lock, flags);
+	if (dev->srq_tbl)
+		srq = dev->srq_tbl[srqn % dev->dsr->caps.max_srq];
+	else
+		srq = NULL;
+	if (srq)
+		refcount_inc(&srq->refcnt);
+	spin_unlock_irqrestore(&dev->srq_tbl_lock, flags);
+
+	if (srq && srq->ibsrq.event_handler) {
+		struct ib_srq *ibsrq = &srq->ibsrq;
+		struct ib_event e;
+
+		e.device = ibsrq->device;
+		e.element.srq = ibsrq;
+		e.event = type; /* 1:1 mapping for now. */
+		ibsrq->event_handler(&e, ibsrq->srq_context);
+	}
+	if (srq) {
+		if (refcount_dec_and_test(&srq->refcnt))
+			wake_up(&srq->wait);
+	}
+}
+
 static void pvrdma_dispatch_event(struct pvrdma_dev *dev, int port,
 				  enum ib_event_type event)
 {
@@ -423,6 +478,7 @@ static irqreturn_t pvrdma_intr1_handler(int irq, void *dev_id)
 
 		case PVRDMA_EVENT_SRQ_ERR:
 		case PVRDMA_EVENT_SRQ_LIMIT_REACHED:
+			pvrdma_srq_event(dev, eqe->info, eqe->type);
 			break;
 
 		case PVRDMA_EVENT_PORT_ACTIVE:
@@ -1059,6 +1115,7 @@ static void pvrdma_pci_remove(struct pci_dev *pdev)
 	iounmap(dev->regs);
 	kfree(dev->sgid_tbl);
 	kfree(dev->cq_tbl);
+	kfree(dev->srq_tbl);
 	kfree(dev->qp_tbl);
 	pvrdma_uar_table_cleanup(dev);
 	iounmap(dev->driver_uar.map);

drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c

@@ -198,6 +198,7 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 	struct pvrdma_create_qp ucmd;
 	unsigned long flags;
 	int ret;
+	bool is_srq = !!init_attr->srq;
 
 	if (init_attr->create_flags) {
 		dev_warn(&dev->pdev->dev,
@@ -214,6 +215,12 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 		return ERR_PTR(-EINVAL);
 	}
 
+	if (is_srq && !dev->dsr->caps.max_srq) {
+		dev_warn(&dev->pdev->dev,
+			 "SRQs not supported by device\n");
+		return ERR_PTR(-EINVAL);
+	}
+
 	if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
 		return ERR_PTR(-ENOMEM);
 
@@ -252,26 +259,36 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 				goto err_qp;
 			}
 
-			/* set qp->sq.wqe_cnt, shift, buf_size.. */
-			qp->rumem = ib_umem_get(pd->uobject->context,
-						ucmd.rbuf_addr,
-						ucmd.rbuf_size, 0, 0);
-			if (IS_ERR(qp->rumem)) {
-				ret = PTR_ERR(qp->rumem);
-				goto err_qp;
+			if (!is_srq) {
+				/* set qp->sq.wqe_cnt, shift, buf_size.. */
+				qp->rumem = ib_umem_get(pd->uobject->context,
+							ucmd.rbuf_addr,
+							ucmd.rbuf_size, 0, 0);
+				if (IS_ERR(qp->rumem)) {
+					ret = PTR_ERR(qp->rumem);
+					goto err_qp;
+				}
+				qp->srq = NULL;
+			} else {
+				qp->rumem = NULL;
+				qp->srq = to_vsrq(init_attr->srq);
 			}
 
 			qp->sumem = ib_umem_get(pd->uobject->context,
 						ucmd.sbuf_addr,
 						ucmd.sbuf_size, 0, 0);
 			if (IS_ERR(qp->sumem)) {
-				ib_umem_release(qp->rumem);
+				if (!is_srq)
+					ib_umem_release(qp->rumem);
 				ret = PTR_ERR(qp->sumem);
 				goto err_qp;
 			}
 
 			qp->npages_send = ib_umem_page_count(qp->sumem);
-			qp->npages_recv = ib_umem_page_count(qp->rumem);
+			if (!is_srq)
+				qp->npages_recv = ib_umem_page_count(qp->rumem);
+			else
+				qp->npages_recv = 0;
 			qp->npages = qp->npages_send + qp->npages_recv;
 		} else {
 			qp->is_kernel = true;
@@ -312,12 +329,14 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 
 		if (!qp->is_kernel) {
 			pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
-			pvrdma_page_dir_insert_umem(&qp->pdir, qp->rumem,
-						    qp->npages_send);
+			if (!is_srq)
+				pvrdma_page_dir_insert_umem(&qp->pdir,
+							    qp->rumem,
+							    qp->npages_send);
 		} else {
 			/* Ring state is always the first page. */
 			qp->sq.ring = qp->pdir.pages[0];
-			qp->rq.ring = &qp->sq.ring[1];
+			qp->rq.ring = is_srq ? NULL : &qp->sq.ring[1];
 		}
 		break;
 	default:
@@ -333,6 +352,10 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 	cmd->pd_handle = to_vpd(pd)->pd_handle;
 	cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
 	cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
+	if (is_srq)
+		cmd->srq_handle = to_vsrq(init_attr->srq)->srq_handle;
+	else
+		cmd->srq_handle = 0;
 	cmd->max_send_wr = init_attr->cap.max_send_wr;
 	cmd->max_recv_wr = init_attr->cap.max_recv_wr;
 	cmd->max_send_sge = init_attr->cap.max_send_sge;
@@ -340,6 +363,8 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 	cmd->max_inline_data = init_attr->cap.max_inline_data;
 	cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 	cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
+	cmd->is_srq = is_srq;
+	cmd->lkey = 0;
 	cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
 	cmd->total_chunks = qp->npages;
 	cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES;
@@ -815,6 +840,12 @@ int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 		return -EINVAL;
 	}
 
+	if (qp->srq) {
+		dev_warn(&dev->pdev->dev, "QP associated with SRQ\n");
+		*bad_wr = wr;
+		return -EINVAL;
+	}
+
 	spin_lock_irqsave(&qp->rq.lock, flags);
 
 	while (wr) {

drivers/infiniband/hw/vmw_pvrdma/pvrdma_srq.c

+/*
+ * Copyright (c) 2016-2017 VMware, Inc.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of EITHER the GNU General Public License
+ * version 2 as published by the Free Software Foundation or the BSD
+ * 2-Clause License. This program is distributed in the hope that it
+ * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED
+ * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See the GNU General Public License version 2 for more details at
+ * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program available in the file COPYING in the main
+ * directory of this source tree.
+ *
+ * The BSD 2-Clause License
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
+ * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
+ * OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <asm/page.h>
+#include <linux/io.h>
+#include <linux/wait.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_user_verbs.h>
+
+#include "pvrdma.h"
+
+int pvrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
+			 struct ib_recv_wr **bad_wr)
+{
+	/* No support for kernel clients. */
+	return -EOPNOTSUPP;
+}
+
+/**
+ * pvrdma_query_srq - query shared receive queue
+ * @ibsrq: the shared receive queue to query
+ * @srq_attr: attributes to query and return to client
+ *
+ * @return: 0 for success, otherwise returns an errno.
+ */
+int pvrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
+{
+	struct pvrdma_dev *dev = to_vdev(ibsrq->device);
+	struct pvrdma_srq *srq = to_vsrq(ibsrq);
+	union pvrdma_cmd_req req;
+	union pvrdma_cmd_resp rsp;
+	struct pvrdma_cmd_query_srq *cmd = &req.query_srq;
+	struct pvrdma_cmd_query_srq_resp *resp = &rsp.query_srq_resp;
+	int ret;
+
+	memset(cmd, 0, sizeof(*cmd));
+	cmd->hdr.cmd = PVRDMA_CMD_QUERY_SRQ;
+	cmd->srq_handle = srq->srq_handle;
+
+	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_SRQ_RESP);
+	if (ret < 0) {
+		dev_warn(&dev->pdev->dev,
+			 "could not query shared receive queue, error: %d\n",
+			 ret);
+		return -EINVAL;
+	}
+
+	srq_attr->srq_limit = resp->attrs.srq_limit;
+	srq_attr->max_wr = resp->attrs.max_wr;
+	srq_attr->max_sge = resp->attrs.max_sge;
+
+	return 0;
+}
+
+/**
+ * pvrdma_create_srq - create shared receive queue
+ * @pd: protection domain
+ * @init_attr: shared receive queue attributes
+ * @udata: user data
+ *
+ * @return: the ib_srq pointer on success, otherwise returns an errno.
+ */
+struct ib_srq *pvrdma_create_srq(struct ib_pd *pd,
+				 struct ib_srq_init_attr *init_attr,
+				 struct ib_udata *udata)
+{
+	struct pvrdma_srq *srq = NULL;
+	struct pvrdma_dev *dev = to_vdev(pd->device);
+	union pvrdma_cmd_req req;
+	union pvrdma_cmd_resp rsp;
+	struct pvrdma_cmd_create_srq *cmd = &req.create_srq;
+	struct pvrdma_cmd_create_srq_resp *resp = &rsp.create_srq_resp;
+	struct pvrdma_create_srq ucmd;
+	unsigned long flags;
+	int ret;
+
+	if (!(pd->uobject && udata)) {
+		/* No support for kernel clients. */
+		dev_warn(&dev->pdev->dev,
+			 "no shared receive queue support for kernel client\n");
+		return ERR_PTR(-EOPNOTSUPP);
+	}
+
+	if (init_attr->srq_type != IB_SRQT_BASIC) {
+		dev_warn(&dev->pdev->dev,
+			 "shared receive queue type %d not supported\n",
+			 init_attr->srq_type);
+		return ERR_PTR(-EINVAL);
+	}
+
+	if (init_attr->attr.max_wr  > dev->dsr->caps.max_srq_wr ||
+	    init_attr->attr.max_sge > dev->dsr->caps.max_srq_sge) {
+		dev_warn(&dev->pdev->dev,
+			 "shared receive queue size invalid\n");
+		return ERR_PTR(-EINVAL);
+	}
+
+	if (!atomic_add_unless(&dev->num_srqs, 1, dev->dsr->caps.max_srq))
+		return ERR_PTR(-ENOMEM);
+
+	srq = kmalloc(sizeof(*srq), GFP_KERNEL);
+	if (!srq) {
+		ret = -ENOMEM;
+		goto err_srq;
+	}
+
+	spin_lock_init(&srq->lock);
+	refcount_set(&srq->refcnt, 1);
+	init_waitqueue_head(&srq->wait);
+
+	dev_dbg(&dev->pdev->dev,
+		"create shared receive queue from user space\n");
+
+	if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
+		ret = -EFAULT;
+		goto err_srq;
+	}
+
+	srq->umem = ib_umem_get(pd->uobject->context,
+				ucmd.buf_addr,
+				ucmd.buf_size, 0, 0);
+	if (IS_ERR(srq->umem)) {
+		ret = PTR_ERR(srq->umem);
+		goto err_srq;
+	}
+
+	srq->npages = ib_umem_page_count(srq->umem);
+
+	if (srq->npages < 0 || srq->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
+		dev_warn(&dev->pdev->dev,
+			 "overflow pages in shared receive queue\n");
+		ret = -EINVAL;
+		goto err_umem;
+	}
+
+	ret = pvrdma_page_dir_init(dev, &srq->pdir, srq->npages, false);
+	if (ret) {
+		dev_warn(&dev->pdev->dev,
+			 "could not allocate page directory\n");
+		goto err_umem;
+	}
+
+	pvrdma_page_dir_insert_umem(&srq->pdir, srq->umem, 0);
+
+	memset(cmd, 0, sizeof(*cmd));
+	cmd->hdr.cmd = PVRDMA_CMD_CREATE_SRQ;
+	cmd->srq_type = init_attr->srq_type;
+	cmd->nchunks = srq->npages;
+	cmd->pd_handle = to_vpd(pd)->pd_handle;
+	cmd->attrs.max_wr = init_attr->attr.max_wr;
+	cmd->attrs.max_sge = init_attr->attr.max_sge;
+	cmd->attrs.srq_limit = init_attr->attr.srq_limit;
+	cmd->pdir_dma = srq->pdir.dir_dma;
+
+	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_SRQ_RESP);
+	if (ret < 0) {
+		dev_warn(&dev->pdev->dev,
+			 "could not create shared receive queue, error: %d\n",
+			 ret);
+		goto err_page_dir;
+	}
+
+	srq->srq_handle = resp->srqn;
+	spin_lock_irqsave(&dev->srq_tbl_lock, flags);
+	dev->srq_tbl[srq->srq_handle % dev->dsr->caps.max_srq] = srq;
+	spin_unlock_irqrestore(&dev->srq_tbl_lock, flags);
+
+	/* Copy udata back. */
+	if (ib_copy_to_udata(udata, &srq->srq_handle, sizeof(__u32))) {
+		dev_warn(&dev->pdev->dev, "failed to copy back udata\n");
+		pvrdma_destroy_srq(&srq->ibsrq);
+		return ERR_PTR(-EINVAL);
+	}
+
+	return &srq->ibsrq;
+
+err_page_dir:
+	pvrdma_page_dir_cleanup(dev, &srq->pdir);
+err_umem:
+	ib_umem_release(srq->umem);
+err_srq:
+	kfree(srq);
+	atomic_dec(&dev->num_srqs);
+
+	return ERR_PTR(ret);
+}
+
+static void pvrdma_free_srq(struct pvrdma_dev *dev, struct pvrdma_srq *srq)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&dev->srq_tbl_lock, flags);
+	dev->srq_tbl[srq->srq_handle] = NULL;
+	spin_unlock_irqrestore(&dev->srq_tbl_lock, flags);
+
+	refcount_dec(&srq->refcnt);
+	wait_event(srq->wait, !refcount_read(&srq->refcnt));
+
+	/* There is no support for kernel clients, so this is safe. */
+	ib_umem_release(srq->umem);
+
+	pvrdma_page_dir_cleanup(dev, &srq->pdir);
+
+	kfree(srq);
+
+	atomic_dec(&dev->num_srqs);
+}
+
+/**
+ * pvrdma_destroy_srq - destroy shared receive queue
+ * @srq: the shared receive queue to destroy
+ *
+ * @return: 0 for success.
+ */
+int pvrdma_destroy_srq(struct ib_srq *srq)
+{
+	struct pvrdma_srq *vsrq = to_vsrq(srq);
+	union pvrdma_cmd_req req;
+	struct pvrdma_cmd_destroy_srq *cmd = &req.destroy_srq;
+	struct pvrdma_dev *dev = to_vdev(srq->device);
+	int ret;
+
+	memset(cmd, 0, sizeof(*cmd));
+	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_SRQ;
+	cmd->srq_handle = vsrq->srq_handle;
+
+	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
+	if (ret < 0)
+		dev_warn(&dev->pdev->dev,
+			 "destroy shared receive queue failed, error: %d\n",
+			 ret);
+
+	pvrdma_free_srq(dev, vsrq);
+
+	return 0;
+}
+
+/**
+ * pvrdma_modify_srq - modify shared receive queue attributes
+ * @ibsrq: the shared receive queue to modify
+ * @attr: the shared receive queue's new attributes
+ * @attr_mask: attributes mask
+ * @udata: user data
+ *
+ * @returns 0 on success, otherwise returns an errno.
+ */
+int pvrdma_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
+		      enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
+{
+	struct pvrdma_srq *vsrq = to_vsrq(ibsrq);
+	union pvrdma_cmd_req req;
+	struct pvrdma_cmd_modify_srq *cmd = &req.modify_srq;
+	struct pvrdma_dev *dev = to_vdev(ibsrq->device);
+	int ret;
+
+	/* Only support SRQ limit. */
+	if (!(attr_mask & IB_SRQ_LIMIT))
+		return -EINVAL;
+
+	memset(cmd, 0, sizeof(*cmd));
+	cmd->hdr.cmd = PVRDMA_CMD_MODIFY_SRQ;
+	cmd->srq_handle = vsrq->srq_handle;
+	cmd->attrs.srq_limit = attr->srq_limit;
+	cmd->attr_mask = attr_mask;
+
+	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
+	if (ret < 0) {
+		dev_warn(&dev->pdev->dev,
+			 "could not modify shared receive queue, error: %d\n",
+			 ret);
+
+		return -EINVAL;
+	}
+
+	return ret;
+}

drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.c

@@ -85,6 +85,9 @@ int pvrdma_query_device(struct ib_device *ibdev,
 	props->max_sge = dev->dsr->caps.max_sge;
 	props->max_sge_rd = PVRDMA_GET_CAP(dev, dev->dsr->caps.max_sge,
 					   dev->dsr->caps.max_sge_rd);
+	props->max_srq = dev->dsr->caps.max_srq;
+	props->max_srq_wr = dev->dsr->caps.max_srq_wr;
+	props->max_srq_sge = dev->dsr->caps.max_srq_sge;
 	props->max_cq = dev->dsr->caps.max_cq;
 	props->max_cqe = dev->dsr->caps.max_cqe;
 	props->max_mr = dev->dsr->caps.max_mr;

drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.h

@@ -324,6 +324,13 @@ enum pvrdma_mw_type {
 	PVRDMA_MW_TYPE_2 = 2,
 };
 
+struct pvrdma_srq_attr {
+	u32			max_wr;
+	u32			max_sge;
+	u32			srq_limit;
+	u32			reserved;
+};
+
 struct pvrdma_qp_attr {
 	enum pvrdma_qp_state	qp_state;
 	enum pvrdma_qp_state	cur_qp_state;
@@ -420,6 +427,17 @@ int pvrdma_req_notify_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags);
 struct ib_ah *pvrdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
 			       struct ib_udata *udata);
 int pvrdma_destroy_ah(struct ib_ah *ah);
+
+struct ib_srq *pvrdma_create_srq(struct ib_pd *pd,
+				 struct ib_srq_init_attr *init_attr,
+				 struct ib_udata *udata);
+int pvrdma_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
+		      enum ib_srq_attr_mask attr_mask, struct ib_udata *udata);
+int pvrdma_query_srq(struct ib_srq *srq, struct ib_srq_attr *srq_attr);
+int pvrdma_destroy_srq(struct ib_srq *srq);
+int pvrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
+			 struct ib_recv_wr **bad_wr);
+
 struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 			       struct ib_qp_init_attr *init_attr,
 			       struct ib_udata *udata);

include/uapi/rdma/vmw_pvrdma-abi.h

@@ -158,6 +158,8 @@ struct pvrdma_resize_cq {
 
 struct pvrdma_create_srq {
 	__u64 buf_addr;
+	__u32 buf_size;
+	__u32 reserved;
 };
 
 struct pvrdma_create_srq_resp {