RDMA/vmw_pvrdma: Use resource ids from physical device if available

This change allows the RDMA stack to use physical resource numbers if they
are passed up from the device. This is accomplished by separating the
concept of the QP number from the QP handle. Previously, the two were the
same, as the QP number was exposed to the guest and also used to reference
a virtual QP in the device backend.

With physical resource numbers exposed, the QP number given to the guest
is the number assigned from the physical HCA's QP, while the QP handle is
still the internal handle used to reference a virtual QP. Regardless of
whether the device is exposing physical ids, the driver will still try to
pick up the QP handle from the backend if possible. The MR keys exposed to
the guest will also be the MR keys created by the physical HCA, instead of
virtual MR keys. The distinction between handle and keys is already
present for MRs so there is no need to do anything special here.

A new version of the create QP response has been added to the device API
to pass up the QP number and handle. The driver will also report these to
userspace in the udata response if userspace supports it or not create the
queuepair if not. I also had to do a refactor of the destroy qp code to
reuse it if we fail to copy to userspace.

Link: https://lore.kernel.org/r/[e-mail]
Reviewed-by: Jorgen Hansen <[e-mail]>
Signed-off-by: Adit Ranadive <[e-mail]>
Signed-off-by: Bryan Tan <[e-mail]>
Signed-off-by: Jason Gunthorpe <[e-mail]>

drivers/infiniband/hw/vmw_pvrdma/pvrdma_dev_api.h

···
 #define PVRDMA_ROCEV1_VERSION		17
 #define PVRDMA_ROCEV2_VERSION		18
 #define PVRDMA_PPN64_VERSION		19
-#define PVRDMA_VERSION			PVRDMA_PPN64_VERSION
+#define PVRDMA_QPHANDLE_VERSION		20
+#define PVRDMA_VERSION			PVRDMA_QPHANDLE_VERSION
 
 #define PVRDMA_BOARD_ID			1
 #define PVRDMA_REV_ID			1
···
 	u32 max_inline_data;
 };
 
+struct pvrdma_cmd_create_qp_resp_v2 {
+	struct pvrdma_cmd_resp_hdr hdr;
+	u32 qpn;
+	u32 qp_handle;
+	u32 max_send_wr;
+	u32 max_recv_wr;
+	u32 max_send_sge;
+	u32 max_recv_sge;
+	u32 max_inline_data;
+};
+
 struct pvrdma_cmd_modify_qp {
 	struct pvrdma_cmd_hdr hdr;
 	u32 qp_handle;
···
 	struct pvrdma_cmd_create_cq_resp create_cq_resp;
 	struct pvrdma_cmd_resize_cq_resp resize_cq_resp;
 	struct pvrdma_cmd_create_qp_resp create_qp_resp;
+	struct pvrdma_cmd_create_qp_resp_v2 create_qp_resp_v2;
 	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;

drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c

···
 
 #include "pvrdma.h"
 
+static void __pvrdma_destroy_qp(struct pvrdma_dev *dev,
+				struct pvrdma_qp *qp);
+
 static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
 			   struct pvrdma_cq **recv_cq)
 {
···
 	union pvrdma_cmd_resp rsp;
 	struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
 	struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
+	struct pvrdma_cmd_create_qp_resp_v2 *resp_v2 = &rsp.create_qp_resp_v2;
 	struct pvrdma_create_qp ucmd;
+	struct pvrdma_create_qp_resp qp_resp = {};
 	unsigned long flags;
 	int ret;
 	bool is_srq = !!init_attr->srq;
···
 
 			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
 				ret = -EFAULT;
+				goto err_qp;
+			}
+
+			/* Userspace supports qpn and qp handles? */
+			if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION &&
+			    udata->outlen < sizeof(qp_resp)) {
+				dev_warn(&dev->pdev->dev,
+					 "create queuepair not supported\n");
+				ret = -EOPNOTSUPP;
 				goto err_qp;
 			}
 
···
 	}
 
 	/* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
-	qp->qp_handle = resp->qpn;
 	qp->port = init_attr->port_num;
-	qp->ibqp.qp_num = resp->qpn;
+
+	if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION) {
+		qp->ibqp.qp_num = resp_v2->qpn;
+		qp->qp_handle = resp_v2->qp_handle;
+	} else {
+		qp->ibqp.qp_num = resp->qpn;
+		qp->qp_handle = resp->qpn;
+	}
+
 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
 	dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
+
+	if (udata) {
+		qp_resp.qpn = qp->ibqp.qp_num;
+		qp_resp.qp_handle = qp->qp_handle;
+
+		if (ib_copy_to_udata(udata, &qp_resp,
+				     min(udata->outlen, sizeof(qp_resp)))) {
+			dev_warn(&dev->pdev->dev,
+				 "failed to copy back udata\n");
+			__pvrdma_destroy_qp(dev, qp);
+			return ERR_PTR(-EINVAL);
+		}
+	}
 
 	return &qp->ibqp;
 
···
 	return ERR_PTR(ret);
 }
 
-static void pvrdma_free_qp(struct pvrdma_qp *qp)
+static void _pvrdma_free_qp(struct pvrdma_qp *qp)
 {
+	unsigned long flags;
 	struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
-	struct pvrdma_cq *scq;
-	struct pvrdma_cq *rcq;
-	unsigned long flags, scq_flags, rcq_flags;
-
-	/* In case cq is polling */
-	get_cqs(qp, &scq, &rcq);
-	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
-
-	_pvrdma_flush_cqe(qp, scq);
-	if (scq != rcq)
-		_pvrdma_flush_cqe(qp, rcq);
 
 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
 	dev->qp_tbl[qp->qp_handle] = NULL;
 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
-
-	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
 
 	if (refcount_dec_and_test(&qp->refcnt))
 		complete(&qp->free);
···
 	atomic_dec(&dev->num_qps);
 }
 
-/**
- * pvrdma_destroy_qp - destroy a queue pair
- * @qp: the queue pair to destroy
- * @udata: user data or null for kernel object
- *
- * @return: 0 on success.
- */
-int pvrdma_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
+static void pvrdma_free_qp(struct pvrdma_qp *qp)
 {
-	struct pvrdma_qp *vqp = to_vqp(qp);
+	struct pvrdma_cq *scq;
+	struct pvrdma_cq *rcq;
+	unsigned long scq_flags, rcq_flags;
+
+	/* In case cq is polling */
+	get_cqs(qp, &scq, &rcq);
+	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
+
+	_pvrdma_flush_cqe(qp, scq);
+	if (scq != rcq)
+		_pvrdma_flush_cqe(qp, rcq);
+
+	/*
+	 * We're now unlocking the CQs before clearing out the qp handle this
+	 * should still be safe. We have destroyed the backend QP and flushed
+	 * the CQEs so there should be no other completions for this QP.
+	 */
+	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
+
+	_pvrdma_free_qp(qp);
+}
+
+static inline void _pvrdma_destroy_qp_work(struct pvrdma_dev *dev,
+					   u32 qp_handle)
+{
 	union pvrdma_cmd_req req;
 	struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
 	int ret;
 
 	memset(cmd, 0, sizeof(*cmd));
 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
-	cmd->qp_handle = vqp->qp_handle;
+	cmd->qp_handle = qp_handle;
 
-	ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0);
+	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
 	if (ret < 0)
-		dev_warn(&to_vdev(qp->device)->pdev->dev,
+		dev_warn(&dev->pdev->dev,
 			 "destroy queuepair failed, error: %d\n", ret);
+}
 
+/**
+ * pvrdma_destroy_qp - destroy a queue pair
+ * @qp: the queue pair to destroy
+ * @udata: user data or null for kernel object
+ *
+ * @return: always 0.
+ */
+int pvrdma_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
+{
+	struct pvrdma_qp *vqp = to_vqp(qp);
+
+	_pvrdma_destroy_qp_work(to_vdev(qp->device), vqp->qp_handle);
 	pvrdma_free_qp(vqp);
 
 	return 0;
+}
+
+static void __pvrdma_destroy_qp(struct pvrdma_dev *dev,
+				struct pvrdma_qp *qp)
+{
+	_pvrdma_destroy_qp_work(dev, qp->qp_handle);
+	_pvrdma_free_qp(qp);
 }
 
 /**

include/uapi/rdma/vmw_pvrdma-abi.h

···
 	__aligned_u64 qp_addr;
 };
 
+struct pvrdma_create_qp_resp {
+	__u32 qpn;
+	__u32 qp_handle;
+};
+
 /* PVRDMA masked atomic compare and swap */
 struct pvrdma_ex_cmp_swap {
 	__aligned_u64 swap_val;