IB: Add vmw_pvrdma driver
authorAdit Ranadive <aditr@vmware.com>
Mon, 3 Oct 2016 02:10:22 +0000 (19:10 -0700)
committerDoug Ledford <dledford@redhat.com>
Wed, 14 Dec 2016 19:55:10 +0000 (14:55 -0500)
This patch series adds a driver for a paravirtual RDMA device. The
device is developed for VMware's Virtual Machines and allows existing RDMA
applications to continue to use existing Verbs API when deployed in VMs
on ESXi. We recently did a presentation in the OFA Workshop [1] regarding
this device.

Description and RDMA Support
============================
The virtual device is exposed as a dual function PCIe device. One part
is a virtual network device (VMXNet3) which provides networking properties
like MAC, IP addresses to the RDMA part of the device. The networking
properties are used to register GIDs required by RDMA applications to
communicate.

These patches add support and the all required infrastructure for
letting applications use such a device. We support the mandatory Verbs API as
well as the base memory management extensions (Local Inv, Send with Inv and
Fast Register Work Requests). We currently support both Reliable Connected
and Unreliable Datagram QPs but do not support Shared Receive Queues
(SRQs).

Also, we support the following types of Work Requests:
 o Send/Receive (with or without Immediate Data)
 o RDMA Write (with or without Immediate Data)
 o RDMA Read
 o Local Invalidate
 o Send with Invalidate
 o Fast Register Work Requests

This version only adds support for version 1 of RoCE. We will add RoCEv2
support in a future patch. We do support registration of both MAC-based
and IP-based GIDs. I have also created a git tree for our user-level driver
[2].

Testing
=======
We have tested this internally for various types of Guest OS - Red Hat,
Centos, Ubuntu 12.04/14.04/16.04, Oracle Enterprise Linux, SLES 12
using backported versions of this driver. The tests included several
runs of the performance tests (included with OFED), Intel MPI PingPong
benchmark on OpenMPI, krping for FRWRs. Mellanox has been kind enough
to test the backported version of the driver internally on their hardware
using a VMware provided ESX build. I have also applied and tested this
with Doug's k.o/for-4.9 branch (commit 5603910b). Note, that this patch
series should be applied all together. I split out the commits so that
it may be easier to review.

PVRDMA Resources
================
[1] OFA Workshop Presentation -
https://openfabrics.org/images/eventpresos/2016presentations/102parardma.pdf

[2] Libpvrdma User-level library -
http://git.openfabrics.org/?p=~aditr/libpvrdma.git;a=summary

Reviewed-by: Jorgen Hansen <jhansen@vmware.com>
Reviewed-by: George Zhang <georgezhang@vmware.com>
Reviewed-by: Aditya Sarwade <asarwade@vmware.com>
Reviewed-by: Bryan Tan <bryantan@vmware.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Adit Ranadive <aditr@vmware.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
19 files changed:
MAINTAINERS
drivers/infiniband/Kconfig
drivers/infiniband/hw/Makefile
drivers/infiniband/hw/vmw_pvrdma/Kconfig [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/Makefile [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma.h [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_cmd.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_dev_api.h [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_doorbell.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_main.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_misc.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_mr.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_ring.h [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.c [new file with mode: 0644]
drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.h [new file with mode: 0644]
include/uapi/rdma/Kbuild
include/uapi/rdma/vmw_pvrdma-abi.h [new file with mode: 0644]

index 3d838cf..75a68f0 100644 (file)
@@ -12928,6 +12928,13 @@ S:     Maintained
 F:     drivers/scsi/vmw_pvscsi.c
 F:     drivers/scsi/vmw_pvscsi.h
 
+VMWARE PVRDMA DRIVER
+M:     Adit Ranadive <aditr@vmware.com>
+M:     VMware PV-Drivers <pv-drivers@vmware.com>
+L:     linux-rdma@vger.kernel.org
+S:     Maintained
+F:     drivers/infiniband/hw/vmw_pvrdma/
+
 VOLTAGE AND CURRENT REGULATOR FRAMEWORK
 M:     Liam Girdwood <lgirdwood@gmail.com>
 M:     Mark Brown <broonie@kernel.org>
index fb3fb89..6709173 100644 (file)
@@ -73,6 +73,7 @@ source "drivers/infiniband/hw/mlx4/Kconfig"
 source "drivers/infiniband/hw/mlx5/Kconfig"
 source "drivers/infiniband/hw/nes/Kconfig"
 source "drivers/infiniband/hw/ocrdma/Kconfig"
+source "drivers/infiniband/hw/vmw_pvrdma/Kconfig"
 source "drivers/infiniband/hw/usnic/Kconfig"
 source "drivers/infiniband/hw/hns/Kconfig"
 
index e7a5ed9..ed553de 100644 (file)
@@ -7,6 +7,7 @@ obj-$(CONFIG_MLX4_INFINIBAND)           += mlx4/
 obj-$(CONFIG_MLX5_INFINIBAND)          += mlx5/
 obj-$(CONFIG_INFINIBAND_NES)           += nes/
 obj-$(CONFIG_INFINIBAND_OCRDMA)                += ocrdma/
+obj-$(CONFIG_INFINIBAND_VMWARE_PVRDMA) += vmw_pvrdma/
 obj-$(CONFIG_INFINIBAND_USNIC)         += usnic/
 obj-$(CONFIG_INFINIBAND_HFI1)          += hfi1/
 obj-$(CONFIG_INFINIBAND_HNS)           += hns/
diff --git a/drivers/infiniband/hw/vmw_pvrdma/Kconfig b/drivers/infiniband/hw/vmw_pvrdma/Kconfig
new file mode 100644 (file)
index 0000000..5a9790a
--- /dev/null
@@ -0,0 +1,7 @@
+config INFINIBAND_VMWARE_PVRDMA
+       tristate "VMware Paravirtualized RDMA Driver"
+       depends on NETDEVICES && ETHERNET && PCI && INET && VMXNET3
+       ---help---
+         This driver provides low-level support for VMware Paravirtual
+         RDMA adapter. It interacts with the VMXNet3 driver to provide
+         Ethernet capabilities.
diff --git a/drivers/infiniband/hw/vmw_pvrdma/Makefile b/drivers/infiniband/hw/vmw_pvrdma/Makefile
new file mode 100644 (file)
index 0000000..0194ed1
--- /dev/null
@@ -0,0 +1,3 @@
+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
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma.h b/drivers/infiniband/hw/vmw_pvrdma/pvrdma.h
new file mode 100644 (file)
index 0000000..71e1d55
--- /dev/null
@@ -0,0 +1,474 @@
+/*
+ * Copyright (c) 2012-2016 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.
+ */
+
+#ifndef __PVRDMA_H__
+#define __PVRDMA_H__
+
+#include <linux/compiler.h>
+#include <linux/interrupt.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
+#include <linux/pci.h>
+#include <linux/semaphore.h>
+#include <linux/workqueue.h>
+#include <rdma/ib_umem.h>
+#include <rdma/ib_verbs.h>
+#include <rdma/vmw_pvrdma-abi.h>
+
+#include "pvrdma_ring.h"
+#include "pvrdma_dev_api.h"
+#include "pvrdma_verbs.h"
+
+/* NOT the same as BIT_MASK(). */
+#define PVRDMA_MASK(n) ((n << 1) - 1)
+
+/*
+ * VMware PVRDMA PCI device id.
+ */
+#define PCI_DEVICE_ID_VMWARE_PVRDMA    0x0820
+
+struct pvrdma_dev;
+
+struct pvrdma_page_dir {
+       dma_addr_t dir_dma;
+       u64 *dir;
+       int ntables;
+       u64 **tables;
+       u64 npages;
+       void **pages;
+};
+
+struct pvrdma_cq {
+       struct ib_cq ibcq;
+       int offset;
+       spinlock_t cq_lock; /* Poll lock. */
+       struct pvrdma_uar_map *uar;
+       struct ib_umem *umem;
+       struct pvrdma_ring_state *ring_state;
+       struct pvrdma_page_dir pdir;
+       u32 cq_handle;
+       bool is_kernel;
+       atomic_t refcnt;
+       wait_queue_head_t wait;
+};
+
+struct pvrdma_id_table {
+       u32 last;
+       u32 top;
+       u32 max;
+       u32 mask;
+       spinlock_t lock; /* Table lock. */
+       unsigned long *table;
+};
+
+struct pvrdma_uar_map {
+       unsigned long pfn;
+       void __iomem *map;
+       int index;
+};
+
+struct pvrdma_uar_table {
+       struct pvrdma_id_table tbl;
+       int size;
+};
+
+struct pvrdma_ucontext {
+       struct ib_ucontext ibucontext;
+       struct pvrdma_dev *dev;
+       struct pvrdma_uar_map uar;
+       u64 ctx_handle;
+};
+
+struct pvrdma_pd {
+       struct ib_pd ibpd;
+       u32 pdn;
+       u32 pd_handle;
+       int privileged;
+};
+
+struct pvrdma_mr {
+       u32 mr_handle;
+       u64 iova;
+       u64 size;
+};
+
+struct pvrdma_user_mr {
+       struct ib_mr ibmr;
+       struct ib_umem *umem;
+       struct pvrdma_mr mmr;
+       struct pvrdma_page_dir pdir;
+       u64 *pages;
+       u32 npages;
+       u32 max_pages;
+       u32 page_shift;
+};
+
+struct pvrdma_wq {
+       struct pvrdma_ring *ring;
+       spinlock_t lock; /* Work queue lock. */
+       int wqe_cnt;
+       int wqe_size;
+       int max_sg;
+       int offset;
+};
+
+struct pvrdma_ah {
+       struct ib_ah ibah;
+       struct pvrdma_av av;
+};
+
+struct pvrdma_qp {
+       struct ib_qp ibqp;
+       u32 qp_handle;
+       u32 qkey;
+       struct pvrdma_wq sq;
+       struct pvrdma_wq rq;
+       struct ib_umem *rumem;
+       struct ib_umem *sumem;
+       struct pvrdma_page_dir pdir;
+       int npages;
+       int npages_send;
+       int npages_recv;
+       u32 flags;
+       u8 port;
+       u8 state;
+       bool is_kernel;
+       struct mutex mutex; /* QP state mutex. */
+       atomic_t refcnt;
+       wait_queue_head_t wait;
+};
+
+struct pvrdma_dev {
+       /* PCI device-related information. */
+       struct ib_device ib_dev;
+       struct pci_dev *pdev;
+       void __iomem *regs;
+       struct pvrdma_device_shared_region *dsr; /* Shared region pointer */
+       dma_addr_t dsrbase; /* Shared region base address */
+       void *cmd_slot;
+       void *resp_slot;
+       unsigned long flags;
+       struct list_head device_link;
+
+       /* Locking and interrupt information. */
+       spinlock_t cmd_lock; /* Command lock. */
+       struct semaphore cmd_sema;
+       struct completion cmd_done;
+       struct {
+               enum pvrdma_intr_type type; /* Intr type */
+               struct msix_entry msix_entry[PVRDMA_MAX_INTERRUPTS];
+               irq_handler_t handler[PVRDMA_MAX_INTERRUPTS];
+               u8 enabled[PVRDMA_MAX_INTERRUPTS];
+               u8 size;
+       } intr;
+
+       /* RDMA-related device information. */
+       union ib_gid *sgid_tbl;
+       struct pvrdma_ring_state *async_ring_state;
+       struct pvrdma_page_dir async_pdir;
+       struct pvrdma_ring_state *cq_ring_state;
+       struct pvrdma_page_dir cq_pdir;
+       struct pvrdma_cq **cq_tbl;
+       spinlock_t cq_tbl_lock;
+       struct pvrdma_qp **qp_tbl;
+       spinlock_t qp_tbl_lock;
+       struct pvrdma_uar_table uar_table;
+       struct pvrdma_uar_map driver_uar;
+       __be64 sys_image_guid;
+       spinlock_t desc_lock; /* Device modification lock. */
+       u32 port_cap_mask;
+       struct mutex port_mutex; /* Port modification mutex. */
+       bool ib_active;
+       atomic_t num_qps;
+       atomic_t num_cqs;
+       atomic_t num_pds;
+       atomic_t num_ahs;
+
+       /* Network device information. */
+       struct net_device *netdev;
+       struct notifier_block nb_netdev;
+};
+
+struct pvrdma_netdevice_work {
+       struct work_struct work;
+       struct net_device *event_netdev;
+       unsigned long event;
+};
+
+static inline struct pvrdma_dev *to_vdev(struct ib_device *ibdev)
+{
+       return container_of(ibdev, struct pvrdma_dev, ib_dev);
+}
+
+static inline struct
+pvrdma_ucontext *to_vucontext(struct ib_ucontext *ibucontext)
+{
+       return container_of(ibucontext, struct pvrdma_ucontext, ibucontext);
+}
+
+static inline struct pvrdma_pd *to_vpd(struct ib_pd *ibpd)
+{
+       return container_of(ibpd, struct pvrdma_pd, ibpd);
+}
+
+static inline struct pvrdma_cq *to_vcq(struct ib_cq *ibcq)
+{
+       return container_of(ibcq, struct pvrdma_cq, ibcq);
+}
+
+static inline struct pvrdma_user_mr *to_vmr(struct ib_mr *ibmr)
+{
+       return container_of(ibmr, struct pvrdma_user_mr, ibmr);
+}
+
+static inline struct pvrdma_qp *to_vqp(struct ib_qp *ibqp)
+{
+       return container_of(ibqp, struct pvrdma_qp, ibqp);
+}
+
+static inline struct pvrdma_ah *to_vah(struct ib_ah *ibah)
+{
+       return container_of(ibah, struct pvrdma_ah, ibah);
+}
+
+static inline void pvrdma_write_reg(struct pvrdma_dev *dev, u32 reg, u32 val)
+{
+       writel(cpu_to_le32(val), dev->regs + reg);
+}
+
+static inline u32 pvrdma_read_reg(struct pvrdma_dev *dev, u32 reg)
+{
+       return le32_to_cpu(readl(dev->regs + reg));
+}
+
+static inline void pvrdma_write_uar_cq(struct pvrdma_dev *dev, u32 val)
+{
+       writel(cpu_to_le32(val), dev->driver_uar.map + PVRDMA_UAR_CQ_OFFSET);
+}
+
+static inline void pvrdma_write_uar_qp(struct pvrdma_dev *dev, u32 val)
+{
+       writel(cpu_to_le32(val), dev->driver_uar.map + PVRDMA_UAR_QP_OFFSET);
+}
+
+static inline void *pvrdma_page_dir_get_ptr(struct pvrdma_page_dir *pdir,
+                                           u64 offset)
+{
+       return pdir->pages[offset / PAGE_SIZE] + (offset % PAGE_SIZE);
+}
+
+static inline enum pvrdma_mtu ib_mtu_to_pvrdma(enum ib_mtu mtu)
+{
+       return (enum pvrdma_mtu)mtu;
+}
+
+static inline enum ib_mtu pvrdma_mtu_to_ib(enum pvrdma_mtu mtu)
+{
+       return (enum ib_mtu)mtu;
+}
+
+static inline enum pvrdma_port_state ib_port_state_to_pvrdma(
+                                       enum ib_port_state state)
+{
+       return (enum pvrdma_port_state)state;
+}
+
+static inline enum ib_port_state pvrdma_port_state_to_ib(
+                                       enum pvrdma_port_state state)
+{
+       return (enum ib_port_state)state;
+}
+
+static inline int ib_port_cap_flags_to_pvrdma(int flags)
+{
+       return flags & PVRDMA_MASK(PVRDMA_PORT_CAP_FLAGS_MAX);
+}
+
+static inline int pvrdma_port_cap_flags_to_ib(int flags)
+{
+       return flags;
+}
+
+static inline enum pvrdma_port_width ib_port_width_to_pvrdma(
+                                       enum ib_port_width width)
+{
+       return (enum pvrdma_port_width)width;
+}
+
+static inline enum ib_port_width pvrdma_port_width_to_ib(
+                                       enum pvrdma_port_width width)
+{
+       return (enum ib_port_width)width;
+}
+
+static inline enum pvrdma_port_speed ib_port_speed_to_pvrdma(
+                                       enum ib_port_speed speed)
+{
+       return (enum pvrdma_port_speed)speed;
+}
+
+static inline enum ib_port_speed pvrdma_port_speed_to_ib(
+                                       enum pvrdma_port_speed speed)
+{
+       return (enum ib_port_speed)speed;
+}
+
+static inline int pvrdma_qp_attr_mask_to_ib(int attr_mask)
+{
+       return attr_mask;
+}
+
+static inline int ib_qp_attr_mask_to_pvrdma(int attr_mask)
+{
+       return attr_mask & PVRDMA_MASK(PVRDMA_QP_ATTR_MASK_MAX);
+}
+
+static inline enum pvrdma_mig_state ib_mig_state_to_pvrdma(
+                                       enum ib_mig_state state)
+{
+       return (enum pvrdma_mig_state)state;
+}
+
+static inline enum ib_mig_state pvrdma_mig_state_to_ib(
+                                       enum pvrdma_mig_state state)
+{
+       return (enum ib_mig_state)state;
+}
+
+static inline int ib_access_flags_to_pvrdma(int flags)
+{
+       return flags;
+}
+
+static inline int pvrdma_access_flags_to_ib(int flags)
+{
+       return flags & PVRDMA_MASK(PVRDMA_ACCESS_FLAGS_MAX);
+}
+
+static inline enum pvrdma_qp_type ib_qp_type_to_pvrdma(enum ib_qp_type type)
+{
+       return (enum pvrdma_qp_type)type;
+}
+
+static inline enum ib_qp_type pvrdma_qp_type_to_ib(enum pvrdma_qp_type type)
+{
+       return (enum ib_qp_type)type;
+}
+
+static inline enum pvrdma_qp_state ib_qp_state_to_pvrdma(enum ib_qp_state state)
+{
+       return (enum pvrdma_qp_state)state;
+}
+
+static inline enum ib_qp_state pvrdma_qp_state_to_ib(enum pvrdma_qp_state state)
+{
+       return (enum ib_qp_state)state;
+}
+
+static inline enum pvrdma_wr_opcode ib_wr_opcode_to_pvrdma(enum ib_wr_opcode op)
+{
+       return (enum pvrdma_wr_opcode)op;
+}
+
+static inline enum ib_wc_status pvrdma_wc_status_to_ib(
+                                       enum pvrdma_wc_status status)
+{
+       return (enum ib_wc_status)status;
+}
+
+static inline int pvrdma_wc_opcode_to_ib(int opcode)
+{
+       return opcode;
+}
+
+static inline int pvrdma_wc_flags_to_ib(int flags)
+{
+       return flags;
+}
+
+static inline int ib_send_flags_to_pvrdma(int flags)
+{
+       return flags & PVRDMA_MASK(PVRDMA_SEND_FLAGS_MAX);
+}
+
+void pvrdma_qp_cap_to_ib(struct ib_qp_cap *dst,
+                        const struct pvrdma_qp_cap *src);
+void ib_qp_cap_to_pvrdma(struct pvrdma_qp_cap *dst,
+                        const struct ib_qp_cap *src);
+void pvrdma_gid_to_ib(union ib_gid *dst, const union pvrdma_gid *src);
+void ib_gid_to_pvrdma(union pvrdma_gid *dst, const union ib_gid *src);
+void pvrdma_global_route_to_ib(struct ib_global_route *dst,
+                              const struct pvrdma_global_route *src);
+void ib_global_route_to_pvrdma(struct pvrdma_global_route *dst,
+                              const struct ib_global_route *src);
+void pvrdma_ah_attr_to_ib(struct ib_ah_attr *dst,
+                         const struct pvrdma_ah_attr *src);
+void ib_ah_attr_to_pvrdma(struct pvrdma_ah_attr *dst,
+                         const struct ib_ah_attr *src);
+
+int pvrdma_uar_table_init(struct pvrdma_dev *dev);
+void pvrdma_uar_table_cleanup(struct pvrdma_dev *dev);
+
+int pvrdma_uar_alloc(struct pvrdma_dev *dev, struct pvrdma_uar_map *uar);
+void pvrdma_uar_free(struct pvrdma_dev *dev, struct pvrdma_uar_map *uar);
+
+void _pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq);
+
+int pvrdma_page_dir_init(struct pvrdma_dev *dev, struct pvrdma_page_dir *pdir,
+                        u64 npages, bool alloc_pages);
+void pvrdma_page_dir_cleanup(struct pvrdma_dev *dev,
+                            struct pvrdma_page_dir *pdir);
+int pvrdma_page_dir_insert_dma(struct pvrdma_page_dir *pdir, u64 idx,
+                              dma_addr_t daddr);
+int pvrdma_page_dir_insert_umem(struct pvrdma_page_dir *pdir,
+                               struct ib_umem *umem, u64 offset);
+dma_addr_t pvrdma_page_dir_get_dma(struct pvrdma_page_dir *pdir, u64 idx);
+int pvrdma_page_dir_insert_page_list(struct pvrdma_page_dir *pdir,
+                                    u64 *page_list, int num_pages);
+
+int pvrdma_cmd_post(struct pvrdma_dev *dev, union pvrdma_cmd_req *req,
+                   union pvrdma_cmd_resp *rsp, unsigned resp_code);
+
+#endif /* __PVRDMA_H__ */
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_cmd.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_cmd.c
new file mode 100644 (file)
index 0000000..4a78c53
--- /dev/null
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2012-2016 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 <linux/list.h>
+
+#include "pvrdma.h"
+
+#define PVRDMA_CMD_TIMEOUT     10000 /* ms */
+
+static inline int pvrdma_cmd_recv(struct pvrdma_dev *dev,
+                                 union pvrdma_cmd_resp *resp,
+                                 unsigned resp_code)
+{
+       int err;
+
+       dev_dbg(&dev->pdev->dev, "receive response from device\n");
+
+       err = wait_for_completion_interruptible_timeout(&dev->cmd_done,
+                       msecs_to_jiffies(PVRDMA_CMD_TIMEOUT));
+       if (err == 0 || err == -ERESTARTSYS) {
+               dev_warn(&dev->pdev->dev,
+                        "completion timeout or interrupted\n");
+               return -ETIMEDOUT;
+       }
+
+       spin_lock(&dev->cmd_lock);
+       memcpy(resp, dev->resp_slot, sizeof(*resp));
+       spin_unlock(&dev->cmd_lock);
+
+       if (resp->hdr.ack != resp_code) {
+               dev_warn(&dev->pdev->dev,
+                        "unknown response %#x expected %#x\n",
+                        resp->hdr.ack, resp_code);
+               return -EFAULT;
+       }
+
+       return 0;
+}
+
+int
+pvrdma_cmd_post(struct pvrdma_dev *dev, union pvrdma_cmd_req *req,
+               union pvrdma_cmd_resp *resp, unsigned resp_code)
+{
+       int err;
+
+       dev_dbg(&dev->pdev->dev, "post request to device\n");
+
+       /* Serializiation */
+       down(&dev->cmd_sema);
+
+       BUILD_BUG_ON(sizeof(union pvrdma_cmd_req) !=
+                    sizeof(struct pvrdma_cmd_modify_qp));
+
+       spin_lock(&dev->cmd_lock);
+       memcpy(dev->cmd_slot, req, sizeof(*req));
+       spin_unlock(&dev->cmd_lock);
+
+       init_completion(&dev->cmd_done);
+       pvrdma_write_reg(dev, PVRDMA_REG_REQUEST, 0);
+
+       /* Make sure the request is written before reading status. */
+       mb();
+
+       err = pvrdma_read_reg(dev, PVRDMA_REG_ERR);
+       if (err == 0) {
+               if (resp != NULL)
+                       err = pvrdma_cmd_recv(dev, resp, resp_code);
+       } else {
+               dev_warn(&dev->pdev->dev,
+                        "failed to write request error reg: %d\n", err);
+               err = -EFAULT;
+       }
+
+       up(&dev->cmd_sema);
+
+       return err;
+}
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c
new file mode 100644 (file)
index 0000000..e429ca5
--- /dev/null
@@ -0,0 +1,425 @@
+/*
+ * Copyright (c) 2012-2016 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"
+
+/**
+ * pvrdma_req_notify_cq - request notification for a completion queue
+ * @ibcq: the completion queue
+ * @notify_flags: notification flags
+ *
+ * @return: 0 for success.
+ */
+int pvrdma_req_notify_cq(struct ib_cq *ibcq,
+                        enum ib_cq_notify_flags notify_flags)
+{
+       struct pvrdma_dev *dev = to_vdev(ibcq->device);
+       struct pvrdma_cq *cq = to_vcq(ibcq);
+       u32 val = cq->cq_handle;
+
+       val |= (notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
+               PVRDMA_UAR_CQ_ARM_SOL : PVRDMA_UAR_CQ_ARM;
+
+       pvrdma_write_uar_cq(dev, val);
+
+       return 0;
+}
+
+/**
+ * pvrdma_create_cq - create completion queue
+ * @ibdev: the device
+ * @attr: completion queue attributes
+ * @context: user context
+ * @udata: user data
+ *
+ * @return: ib_cq completion queue pointer on success,
+ *          otherwise returns negative errno.
+ */
+struct ib_cq *pvrdma_create_cq(struct ib_device *ibdev,
+                              const struct ib_cq_init_attr *attr,
+                              struct ib_ucontext *context,
+                              struct ib_udata *udata)
+{
+       int entries = attr->cqe;
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+       struct pvrdma_cq *cq;
+       int ret;
+       int npages;
+       unsigned long flags;
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_create_cq *cmd = &req.create_cq;
+       struct pvrdma_cmd_create_cq_resp *resp = &rsp.create_cq_resp;
+       struct pvrdma_create_cq ucmd;
+
+       BUILD_BUG_ON(sizeof(struct pvrdma_cqe) != 64);
+
+       entries = roundup_pow_of_two(entries);
+       if (entries < 1 || entries > dev->dsr->caps.max_cqe)
+               return ERR_PTR(-EINVAL);
+
+       if (!atomic_add_unless(&dev->num_cqs, 1, dev->dsr->caps.max_cq))
+               return ERR_PTR(-ENOMEM);
+
+       cq = kzalloc(sizeof(*cq), GFP_KERNEL);
+       if (!cq) {
+               atomic_dec(&dev->num_cqs);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       cq->ibcq.cqe = entries;
+
+       if (context) {
+               if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
+                       ret = -EFAULT;
+                       goto err_cq;
+               }
+
+               cq->umem = ib_umem_get(context, ucmd.buf_addr, ucmd.buf_size,
+                                      IB_ACCESS_LOCAL_WRITE, 1);
+               if (IS_ERR(cq->umem)) {
+                       ret = PTR_ERR(cq->umem);
+                       goto err_cq;
+               }
+
+               npages = ib_umem_page_count(cq->umem);
+       } else {
+               cq->is_kernel = true;
+
+               /* One extra page for shared ring state */
+               npages = 1 + (entries * sizeof(struct pvrdma_cqe) +
+                             PAGE_SIZE - 1) / PAGE_SIZE;
+
+               /* Skip header page. */
+               cq->offset = PAGE_SIZE;
+       }
+
+       if (npages < 0 || npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
+               dev_warn(&dev->pdev->dev,
+                        "overflow pages in completion queue\n");
+               ret = -EINVAL;
+               goto err_umem;
+       }
+
+       ret = pvrdma_page_dir_init(dev, &cq->pdir, npages, cq->is_kernel);
+       if (ret) {
+               dev_warn(&dev->pdev->dev,
+                        "could not allocate page directory\n");
+               goto err_umem;
+       }
+
+       /* Ring state is always the first page. Set in library for user cq. */
+       if (cq->is_kernel)
+               cq->ring_state = cq->pdir.pages[0];
+       else
+               pvrdma_page_dir_insert_umem(&cq->pdir, cq->umem, 0);
+
+       atomic_set(&cq->refcnt, 1);
+       init_waitqueue_head(&cq->wait);
+       spin_lock_init(&cq->cq_lock);
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_CREATE_CQ;
+       cmd->nchunks = npages;
+       cmd->ctx_handle = (context) ?
+               (u64)to_vucontext(context)->ctx_handle : 0;
+       cmd->cqe = entries;
+       cmd->pdir_dma = cq->pdir.dir_dma;
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_CQ_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not create completion queue, error: %d\n", ret);
+               goto err_page_dir;
+       }
+
+       cq->ibcq.cqe = resp->cqe;
+       cq->cq_handle = resp->cq_handle;
+       spin_lock_irqsave(&dev->cq_tbl_lock, flags);
+       dev->cq_tbl[cq->cq_handle % dev->dsr->caps.max_cq] = cq;
+       spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
+
+       if (context) {
+               cq->uar = &(to_vucontext(context)->uar);
+
+               /* Copy udata back. */
+               if (ib_copy_to_udata(udata, &cq->cq_handle, sizeof(__u32))) {
+                       dev_warn(&dev->pdev->dev,
+                                "failed to copy back udata\n");
+                       pvrdma_destroy_cq(&cq->ibcq);
+                       return ERR_PTR(-EINVAL);
+               }
+       }
+
+       return &cq->ibcq;
+
+err_page_dir:
+       pvrdma_page_dir_cleanup(dev, &cq->pdir);
+err_umem:
+       if (context)
+               ib_umem_release(cq->umem);
+err_cq:
+       atomic_dec(&dev->num_cqs);
+       kfree(cq);
+
+       return ERR_PTR(ret);
+}
+
+static void pvrdma_free_cq(struct pvrdma_dev *dev, struct pvrdma_cq *cq)
+{
+       atomic_dec(&cq->refcnt);
+       wait_event(cq->wait, !atomic_read(&cq->refcnt));
+
+       if (!cq->is_kernel)
+               ib_umem_release(cq->umem);
+
+       pvrdma_page_dir_cleanup(dev, &cq->pdir);
+       kfree(cq);
+}
+
+/**
+ * pvrdma_destroy_cq - destroy completion queue
+ * @cq: the completion queue to destroy.
+ *
+ * @return: 0 for success.
+ */
+int pvrdma_destroy_cq(struct ib_cq *cq)
+{
+       struct pvrdma_cq *vcq = to_vcq(cq);
+       union pvrdma_cmd_req req;
+       struct pvrdma_cmd_destroy_cq *cmd = &req.destroy_cq;
+       struct pvrdma_dev *dev = to_vdev(cq->device);
+       unsigned long flags;
+       int ret;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_DESTROY_CQ;
+       cmd->cq_handle = vcq->cq_handle;
+
+       ret = pvrdma_cmd_post(dev, &req, NULL, 0);
+       if (ret < 0)
+               dev_warn(&dev->pdev->dev,
+                        "could not destroy completion queue, error: %d\n",
+                        ret);
+
+       /* free cq's resources */
+       spin_lock_irqsave(&dev->cq_tbl_lock, flags);
+       dev->cq_tbl[vcq->cq_handle] = NULL;
+       spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
+
+       pvrdma_free_cq(dev, vcq);
+       atomic_dec(&dev->num_cqs);
+
+       return ret;
+}
+
+/**
+ * pvrdma_modify_cq - modify the CQ moderation parameters
+ * @ibcq: the CQ to modify
+ * @cq_count: number of CQEs that will trigger an event
+ * @cq_period: max period of time in usec before triggering an event
+ *
+ * @return: -EOPNOTSUPP as CQ resize is not supported.
+ */
+int pvrdma_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
+{
+       return -EOPNOTSUPP;
+}
+
+static inline struct pvrdma_cqe *get_cqe(struct pvrdma_cq *cq, int i)
+{
+       return (struct pvrdma_cqe *)pvrdma_page_dir_get_ptr(
+                                       &cq->pdir,
+                                       cq->offset +
+                                       sizeof(struct pvrdma_cqe) * i);
+}
+
+void _pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq)
+{
+       int head;
+       int has_data;
+
+       if (!cq->is_kernel)
+               return;
+
+       /* Lock held */
+       has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
+                                           cq->ibcq.cqe, &head);
+       if (unlikely(has_data > 0)) {
+               int items;
+               int curr;
+               int tail = pvrdma_idx(&cq->ring_state->rx.prod_tail,
+                                     cq->ibcq.cqe);
+               struct pvrdma_cqe *cqe;
+               struct pvrdma_cqe *curr_cqe;
+
+               items = (tail > head) ? (tail - head) :
+                       (cq->ibcq.cqe - head + tail);
+               curr = --tail;
+               while (items-- > 0) {
+                       if (curr < 0)
+                               curr = cq->ibcq.cqe - 1;
+                       if (tail < 0)
+                               tail = cq->ibcq.cqe - 1;
+                       curr_cqe = get_cqe(cq, curr);
+                       if ((curr_cqe->qp & 0xFFFF) != qp->qp_handle) {
+                               if (curr != tail) {
+                                       cqe = get_cqe(cq, tail);
+                                       *cqe = *curr_cqe;
+                               }
+                               tail--;
+                       } else {
+                               pvrdma_idx_ring_inc(
+                                       &cq->ring_state->rx.cons_head,
+                                       cq->ibcq.cqe);
+                       }
+                       curr--;
+               }
+       }
+}
+
+static int pvrdma_poll_one(struct pvrdma_cq *cq, struct pvrdma_qp **cur_qp,
+                          struct ib_wc *wc)
+{
+       struct pvrdma_dev *dev = to_vdev(cq->ibcq.device);
+       int has_data;
+       unsigned int head;
+       bool tried = false;
+       struct pvrdma_cqe *cqe;
+
+retry:
+       has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx,
+                                           cq->ibcq.cqe, &head);
+       if (has_data == 0) {
+               if (tried)
+                       return -EAGAIN;
+
+               pvrdma_write_uar_cq(dev, cq->cq_handle | PVRDMA_UAR_CQ_POLL);
+
+               tried = true;
+               goto retry;
+       } else if (has_data == PVRDMA_INVALID_IDX) {
+               dev_err(&dev->pdev->dev, "CQ ring state invalid\n");
+               return -EAGAIN;
+       }
+
+       cqe = get_cqe(cq, head);
+
+       /* Ensure cqe is valid. */
+       rmb();
+       if (dev->qp_tbl[cqe->qp & 0xffff])
+               *cur_qp = (struct pvrdma_qp *)dev->qp_tbl[cqe->qp & 0xffff];
+       else
+               return -EAGAIN;
+
+       wc->opcode = pvrdma_wc_opcode_to_ib(cqe->opcode);
+       wc->status = pvrdma_wc_status_to_ib(cqe->status);
+       wc->wr_id = cqe->wr_id;
+       wc->qp = &(*cur_qp)->ibqp;
+       wc->byte_len = cqe->byte_len;
+       wc->ex.imm_data = cqe->imm_data;
+       wc->src_qp = cqe->src_qp;
+       wc->wc_flags = pvrdma_wc_flags_to_ib(cqe->wc_flags);
+       wc->pkey_index = cqe->pkey_index;
+       wc->slid = cqe->slid;
+       wc->sl = cqe->sl;
+       wc->dlid_path_bits = cqe->dlid_path_bits;
+       wc->port_num = cqe->port_num;
+       wc->vendor_err = 0;
+
+       /* Update shared ring state */
+       pvrdma_idx_ring_inc(&cq->ring_state->rx.cons_head, cq->ibcq.cqe);
+
+       return 0;
+}
+
+/**
+ * pvrdma_poll_cq - poll for work completion queue entries
+ * @ibcq: completion queue
+ * @num_entries: the maximum number of entries
+ * @entry: pointer to work completion array
+ *
+ * @return: number of polled completion entries
+ */
+int pvrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
+{
+       struct pvrdma_cq *cq = to_vcq(ibcq);
+       struct pvrdma_qp *cur_qp = NULL;
+       unsigned long flags;
+       int npolled;
+
+       if (num_entries < 1 || wc == NULL)
+               return 0;
+
+       spin_lock_irqsave(&cq->cq_lock, flags);
+       for (npolled = 0; npolled < num_entries; ++npolled) {
+               if (pvrdma_poll_one(cq, &cur_qp, wc + npolled))
+                       break;
+       }
+
+       spin_unlock_irqrestore(&cq->cq_lock, flags);
+
+       /* Ensure we do not return errors from poll_cq */
+       return npolled;
+}
+
+/**
+ * pvrdma_resize_cq - resize CQ
+ * @ibcq: the completion queue
+ * @entries: CQ entries
+ * @udata: user data
+ *
+ * @return: -EOPNOTSUPP as CQ resize is not supported.
+ */
+int pvrdma_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
+{
+       return -EOPNOTSUPP;
+}
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_dev_api.h b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_dev_api.h
new file mode 100644 (file)
index 0000000..c067686
--- /dev/null
@@ -0,0 +1,586 @@
+/*
+ * Copyright (c) 2012-2016 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.
+ */
+
+#ifndef __PVRDMA_DEV_API_H__
+#define __PVRDMA_DEV_API_H__
+
+#include <linux/types.h>
+
+#include "pvrdma_verbs.h"
+
+#define PVRDMA_VERSION                 17
+#define PVRDMA_BOARD_ID                        1
+#define PVRDMA_REV_ID                  1
+
+/*
+ * Masks and accessors for page directory, which is a two-level lookup:
+ * page directory -> page table -> page. Only one directory for now, but we
+ * could expand that easily. 9 bits for tables, 9 bits for pages, gives one
+ * gigabyte for memory regions and so forth.
+ */
+
+#define PVRDMA_PDIR_SHIFT              18
+#define PVRDMA_PTABLE_SHIFT            9
+#define PVRDMA_PAGE_DIR_DIR(x)         (((x) >> PVRDMA_PDIR_SHIFT) & 0x1)
+#define PVRDMA_PAGE_DIR_TABLE(x)       (((x) >> PVRDMA_PTABLE_SHIFT) & 0x1ff)
+#define PVRDMA_PAGE_DIR_PAGE(x)                ((x) & 0x1ff)
+#define PVRDMA_PAGE_DIR_MAX_PAGES      (1 * 512 * 512)
+#define PVRDMA_MAX_FAST_REG_PAGES      128
+
+/*
+ * Max MSI-X vectors.
+ */
+
+#define PVRDMA_MAX_INTERRUPTS  3
+
+/* Register offsets within PCI resource on BAR1. */
+#define PVRDMA_REG_VERSION     0x00    /* R: Version of device. */
+#define PVRDMA_REG_DSRLOW      0x04    /* W: Device shared region low PA. */
+#define PVRDMA_REG_DSRHIGH     0x08    /* W: Device shared region high PA. */
+#define PVRDMA_REG_CTL         0x0c    /* W: PVRDMA_DEVICE_CTL */
+#define PVRDMA_REG_REQUEST     0x10    /* W: Indicate device request. */
+#define PVRDMA_REG_ERR         0x14    /* R: Device error. */
+#define PVRDMA_REG_ICR         0x18    /* R: Interrupt cause. */
+#define PVRDMA_REG_IMR         0x1c    /* R/W: Interrupt mask. */
+#define PVRDMA_REG_MACL                0x20    /* R/W: MAC address low. */
+#define PVRDMA_REG_MACH                0x24    /* R/W: MAC address high. */
+
+/* Object flags. */
+#define PVRDMA_CQ_FLAG_ARMED_SOL       BIT(0)  /* Armed for solicited-only. */
+#define PVRDMA_CQ_FLAG_ARMED           BIT(1)  /* Armed. */
+#define PVRDMA_MR_FLAG_DMA             BIT(0)  /* DMA region. */
+#define PVRDMA_MR_FLAG_FRMR            BIT(1)  /* Fast reg memory region. */
+
+/*
+ * Atomic operation capability (masked versions are extended atomic
+ * operations.
+ */
+
+#define PVRDMA_ATOMIC_OP_COMP_SWAP     BIT(0)  /* Compare and swap. */
+#define PVRDMA_ATOMIC_OP_FETCH_ADD     BIT(1)  /* Fetch and add. */
+#define PVRDMA_ATOMIC_OP_MASK_COMP_SWAP        BIT(2)  /* Masked compare and swap. */
+#define PVRDMA_ATOMIC_OP_MASK_FETCH_ADD        BIT(3)  /* Masked fetch and add. */
+
+/*
+ * Base Memory Management Extension flags to support Fast Reg Memory Regions
+ * and Fast Reg Work Requests. Each flag represents a verb operation and we
+ * must support all of them to qualify for the BMME device cap.
+ */
+
+#define PVRDMA_BMME_FLAG_LOCAL_INV     BIT(0)  /* Local Invalidate. */
+#define PVRDMA_BMME_FLAG_REMOTE_INV    BIT(1)  /* Remote Invalidate. */
+#define PVRDMA_BMME_FLAG_FAST_REG_WR   BIT(2)  /* Fast Reg Work Request. */
+
+/*
+ * GID types. The interpretation of the gid_types bit field in the device
+ * capabilities will depend on the device mode. For now, the device only
+ * supports RoCE as mode, so only the different GID types for RoCE are
+ * defined.
+ */
+
+#define PVRDMA_GID_TYPE_FLAG_ROCE_V1   BIT(0)
+#define PVRDMA_GID_TYPE_FLAG_ROCE_V2   BIT(1)
+
+enum pvrdma_pci_resource {
+       PVRDMA_PCI_RESOURCE_MSIX,       /* BAR0: MSI-X, MMIO. */
+       PVRDMA_PCI_RESOURCE_REG,        /* BAR1: Registers, MMIO. */
+       PVRDMA_PCI_RESOURCE_UAR,        /* BAR2: UAR pages, MMIO, 64-bit. */
+       PVRDMA_PCI_RESOURCE_LAST,       /* Last. */
+};
+
+enum pvrdma_device_ctl {
+       PVRDMA_DEVICE_CTL_ACTIVATE,     /* Activate device. */
+       PVRDMA_DEVICE_CTL_QUIESCE,      /* Quiesce device. */
+       PVRDMA_DEVICE_CTL_RESET,        /* Reset device. */
+};
+
+enum pvrdma_intr_vector {
+       PVRDMA_INTR_VECTOR_RESPONSE,    /* Command response. */
+       PVRDMA_INTR_VECTOR_ASYNC,       /* Async events. */
+       PVRDMA_INTR_VECTOR_CQ,          /* CQ notification. */
+       /* Additional CQ notification vectors. */
+};
+
+enum pvrdma_intr_cause {
+       PVRDMA_INTR_CAUSE_RESPONSE      = (1 << PVRDMA_INTR_VECTOR_RESPONSE),
+       PVRDMA_INTR_CAUSE_ASYNC         = (1 << PVRDMA_INTR_VECTOR_ASYNC),
+       PVRDMA_INTR_CAUSE_CQ            = (1 << PVRDMA_INTR_VECTOR_CQ),
+};
+
+enum pvrdma_intr_type {
+       PVRDMA_INTR_TYPE_INTX,          /* Legacy. */
+       PVRDMA_INTR_TYPE_MSI,           /* MSI. */
+       PVRDMA_INTR_TYPE_MSIX,          /* MSI-X. */
+};
+
+enum pvrdma_gos_bits {
+       PVRDMA_GOS_BITS_UNK,            /* Unknown. */
+       PVRDMA_GOS_BITS_32,             /* 32-bit. */
+       PVRDMA_GOS_BITS_64,             /* 64-bit. */
+};
+
+enum pvrdma_gos_type {
+       PVRDMA_GOS_TYPE_UNK,            /* Unknown. */
+       PVRDMA_GOS_TYPE_LINUX,          /* Linux. */
+};
+
+enum pvrdma_device_mode {
+       PVRDMA_DEVICE_MODE_ROCE,        /* RoCE. */
+       PVRDMA_DEVICE_MODE_IWARP,       /* iWarp. */
+       PVRDMA_DEVICE_MODE_IB,          /* InfiniBand. */
+};
+
+struct pvrdma_gos_info {
+       u32 gos_bits:2;                 /* W: PVRDMA_GOS_BITS_ */
+       u32 gos_type:4;                 /* W: PVRDMA_GOS_TYPE_ */
+       u32 gos_ver:16;                 /* W: Guest OS version. */
+       u32 gos_misc:10;                /* W: Other. */
+       u32 pad;                        /* Pad to 8-byte alignment. */
+};
+
+struct pvrdma_device_caps {
+       u64 fw_ver;                             /* R: Query device. */
+       __be64 node_guid;
+       __be64 sys_image_guid;
+       u64 max_mr_size;
+       u64 page_size_cap;
+       u64 atomic_arg_sizes;                   /* EX verbs. */
+       u32 ex_comp_mask;                       /* EX verbs. */
+       u32 device_cap_flags2;                  /* EX verbs. */
+       u32 max_fa_bit_boundary;                /* EX verbs. */
+       u32 log_max_atomic_inline_arg;          /* EX verbs. */
+       u32 vendor_id;
+       u32 vendor_part_id;
+       u32 hw_ver;
+       u32 max_qp;
+       u32 max_qp_wr;
+       u32 device_cap_flags;
+       u32 max_sge;
+       u32 max_sge_rd;
+       u32 max_cq;
+       u32 max_cqe;
+       u32 max_mr;
+       u32 max_pd;
+       u32 max_qp_rd_atom;
+       u32 max_ee_rd_atom;
+       u32 max_res_rd_atom;
+       u32 max_qp_init_rd_atom;
+       u32 max_ee_init_rd_atom;
+       u32 max_ee;
+       u32 max_rdd;
+       u32 max_mw;
+       u32 max_raw_ipv6_qp;
+       u32 max_raw_ethy_qp;
+       u32 max_mcast_grp;
+       u32 max_mcast_qp_attach;
+       u32 max_total_mcast_qp_attach;
+       u32 max_ah;
+       u32 max_fmr;
+       u32 max_map_per_fmr;
+       u32 max_srq;
+       u32 max_srq_wr;
+       u32 max_srq_sge;
+       u32 max_uar;
+       u32 gid_tbl_len;
+       u16 max_pkeys;
+       u8  local_ca_ack_delay;
+       u8  phys_port_cnt;
+       u8  mode;                               /* PVRDMA_DEVICE_MODE_ */
+       u8  atomic_ops;                         /* PVRDMA_ATOMIC_OP_* bits */
+       u8  bmme_flags;                         /* FRWR Mem Mgmt Extensions */
+       u8  gid_types;                          /* PVRDMA_GID_TYPE_FLAG_ */
+       u8  reserved[4];
+};
+
+struct pvrdma_ring_page_info {
+       u32 num_pages;                          /* Num pages incl. header. */
+       u32 reserved;                           /* Reserved. */
+       u64 pdir_dma;                           /* Page directory PA. */
+};
+
+#pragma pack(push, 1)
+
+struct pvrdma_device_shared_region {
+       u32 driver_version;                     /* W: Driver version. */
+       u32 pad;                                /* Pad to 8-byte align. */
+       struct pvrdma_gos_info gos_info;        /* W: Guest OS information. */
+       u64 cmd_slot_dma;                       /* W: Command slot address. */
+       u64 resp_slot_dma;                      /* W: Response slot address. */
+       struct pvrdma_ring_page_info async_ring_pages;
+                                               /* W: Async ring page info. */
+       struct pvrdma_ring_page_info cq_ring_pages;
+                                               /* W: CQ ring page info. */
+       u32 uar_pfn;                            /* W: UAR pageframe. */
+       u32 pad2;                               /* Pad to 8-byte align. */
+       struct pvrdma_device_caps caps;         /* R: Device capabilities. */
+};
+
+#pragma pack(pop)
+
+/* Event types. Currently a 1:1 mapping with enum ib_event. */
+enum pvrdma_eqe_type {
+       PVRDMA_EVENT_CQ_ERR,
+       PVRDMA_EVENT_QP_FATAL,
+       PVRDMA_EVENT_QP_REQ_ERR,
+       PVRDMA_EVENT_QP_ACCESS_ERR,
+       PVRDMA_EVENT_COMM_EST,
+       PVRDMA_EVENT_SQ_DRAINED,
+       PVRDMA_EVENT_PATH_MIG,
+       PVRDMA_EVENT_PATH_MIG_ERR,
+       PVRDMA_EVENT_DEVICE_FATAL,
+       PVRDMA_EVENT_PORT_ACTIVE,
+       PVRDMA_EVENT_PORT_ERR,
+       PVRDMA_EVENT_LID_CHANGE,
+       PVRDMA_EVENT_PKEY_CHANGE,
+       PVRDMA_EVENT_SM_CHANGE,
+       PVRDMA_EVENT_SRQ_ERR,
+       PVRDMA_EVENT_SRQ_LIMIT_REACHED,
+       PVRDMA_EVENT_QP_LAST_WQE_REACHED,
+       PVRDMA_EVENT_CLIENT_REREGISTER,
+       PVRDMA_EVENT_GID_CHANGE,
+};
+
+/* Event queue element. */
+struct pvrdma_eqe {
+       u32 type;       /* Event type. */
+       u32 info;       /* Handle, other. */
+};
+
+/* CQ notification queue element. */
+struct pvrdma_cqne {
+       u32 info;       /* Handle */
+};
+
+enum {
+       PVRDMA_CMD_FIRST,
+       PVRDMA_CMD_QUERY_PORT = PVRDMA_CMD_FIRST,
+       PVRDMA_CMD_QUERY_PKEY,
+       PVRDMA_CMD_CREATE_PD,
+       PVRDMA_CMD_DESTROY_PD,
+       PVRDMA_CMD_CREATE_MR,
+       PVRDMA_CMD_DESTROY_MR,
+       PVRDMA_CMD_CREATE_CQ,
+       PVRDMA_CMD_RESIZE_CQ,
+       PVRDMA_CMD_DESTROY_CQ,
+       PVRDMA_CMD_CREATE_QP,
+       PVRDMA_CMD_MODIFY_QP,
+       PVRDMA_CMD_QUERY_QP,
+       PVRDMA_CMD_DESTROY_QP,
+       PVRDMA_CMD_CREATE_UC,
+       PVRDMA_CMD_DESTROY_UC,
+       PVRDMA_CMD_CREATE_BIND,
+       PVRDMA_CMD_DESTROY_BIND,
+       PVRDMA_CMD_MAX,
+};
+
+enum {
+       PVRDMA_CMD_FIRST_RESP = (1 << 31),
+       PVRDMA_CMD_QUERY_PORT_RESP = PVRDMA_CMD_FIRST_RESP,
+       PVRDMA_CMD_QUERY_PKEY_RESP,
+       PVRDMA_CMD_CREATE_PD_RESP,
+       PVRDMA_CMD_DESTROY_PD_RESP_NOOP,
+       PVRDMA_CMD_CREATE_MR_RESP,
+       PVRDMA_CMD_DESTROY_MR_RESP_NOOP,
+       PVRDMA_CMD_CREATE_CQ_RESP,
+       PVRDMA_CMD_RESIZE_CQ_RESP,
+       PVRDMA_CMD_DESTROY_CQ_RESP_NOOP,
+       PVRDMA_CMD_CREATE_QP_RESP,
+       PVRDMA_CMD_MODIFY_QP_RESP,
+       PVRDMA_CMD_QUERY_QP_RESP,
+       PVRDMA_CMD_DESTROY_QP_RESP,
+       PVRDMA_CMD_CREATE_UC_RESP,
+       PVRDMA_CMD_DESTROY_UC_RESP_NOOP,
+       PVRDMA_CMD_CREATE_BIND_RESP_NOOP,
+       PVRDMA_CMD_DESTROY_BIND_RESP_NOOP,
+       PVRDMA_CMD_MAX_RESP,
+};
+
+struct pvrdma_cmd_hdr {
+       u64 response;           /* Key for response lookup. */
+       u32 cmd;                /* PVRDMA_CMD_ */
+       u32 reserved;           /* Reserved. */
+};
+
+struct pvrdma_cmd_resp_hdr {
+       u64 response;           /* From cmd hdr. */
+       u32 ack;                /* PVRDMA_CMD_XXX_RESP */
+       u8 err;                 /* Error. */
+       u8 reserved[3];         /* Reserved. */
+};
+
+struct pvrdma_cmd_query_port {
+       struct pvrdma_cmd_hdr hdr;
+       u8 port_num;
+       u8 reserved[7];
+};
+
+struct pvrdma_cmd_query_port_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       struct pvrdma_port_attr attrs;
+};
+
+struct pvrdma_cmd_query_pkey {
+       struct pvrdma_cmd_hdr hdr;
+       u8 port_num;
+       u8 index;
+       u8 reserved[6];
+};
+
+struct pvrdma_cmd_query_pkey_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u16 pkey;
+       u8 reserved[6];
+};
+
+struct pvrdma_cmd_create_uc {
+       struct pvrdma_cmd_hdr hdr;
+       u32 pfn; /* UAR page frame number */
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_uc_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u32 ctx_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_destroy_uc {
+       struct pvrdma_cmd_hdr hdr;
+       u32 ctx_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_pd {
+       struct pvrdma_cmd_hdr hdr;
+       u32 ctx_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_pd_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u32 pd_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_destroy_pd {
+       struct pvrdma_cmd_hdr hdr;
+       u32 pd_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_mr {
+       struct pvrdma_cmd_hdr hdr;
+       u64 start;
+       u64 length;
+       u64 pdir_dma;
+       u32 pd_handle;
+       u32 access_flags;
+       u32 flags;
+       u32 nchunks;
+};
+
+struct pvrdma_cmd_create_mr_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u32 mr_handle;
+       u32 lkey;
+       u32 rkey;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_destroy_mr {
+       struct pvrdma_cmd_hdr hdr;
+       u32 mr_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_cq {
+       struct pvrdma_cmd_hdr hdr;
+       u64 pdir_dma;
+       u32 ctx_handle;
+       u32 cqe;
+       u32 nchunks;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_cq_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u32 cq_handle;
+       u32 cqe;
+};
+
+struct pvrdma_cmd_resize_cq {
+       struct pvrdma_cmd_hdr hdr;
+       u32 cq_handle;
+       u32 cqe;
+};
+
+struct pvrdma_cmd_resize_cq_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u32 cqe;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_destroy_cq {
+       struct pvrdma_cmd_hdr hdr;
+       u32 cq_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_qp {
+       struct pvrdma_cmd_hdr hdr;
+       u64 pdir_dma;
+       u32 pd_handle;
+       u32 send_cq_handle;
+       u32 recv_cq_handle;
+       u32 srq_handle;
+       u32 max_send_wr;
+       u32 max_recv_wr;
+       u32 max_send_sge;
+       u32 max_recv_sge;
+       u32 max_inline_data;
+       u32 lkey;
+       u32 access_flags;
+       u16 total_chunks;
+       u16 send_chunks;
+       u16 max_atomic_arg;
+       u8 sq_sig_all;
+       u8 qp_type;
+       u8 is_srq;
+       u8 reserved[3];
+};
+
+struct pvrdma_cmd_create_qp_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u32 qpn;
+       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;
+       u32 attr_mask;
+       struct pvrdma_qp_attr attrs;
+};
+
+struct pvrdma_cmd_query_qp {
+       struct pvrdma_cmd_hdr hdr;
+       u32 qp_handle;
+       u32 attr_mask;
+};
+
+struct pvrdma_cmd_query_qp_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       struct pvrdma_qp_attr attrs;
+};
+
+struct pvrdma_cmd_destroy_qp {
+       struct pvrdma_cmd_hdr hdr;
+       u32 qp_handle;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_destroy_qp_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       u32 events_reported;
+       u8 reserved[4];
+};
+
+struct pvrdma_cmd_create_bind {
+       struct pvrdma_cmd_hdr hdr;
+       u32 mtu;
+       u32 vlan;
+       u32 index;
+       u8 new_gid[16];
+       u8 gid_type;
+       u8 reserved[3];
+};
+
+struct pvrdma_cmd_destroy_bind {
+       struct pvrdma_cmd_hdr hdr;
+       u32 index;
+       u8 dest_gid[16];
+       u8 reserved[4];
+};
+
+union pvrdma_cmd_req {
+       struct pvrdma_cmd_hdr hdr;
+       struct pvrdma_cmd_query_port query_port;
+       struct pvrdma_cmd_query_pkey query_pkey;
+       struct pvrdma_cmd_create_uc create_uc;
+       struct pvrdma_cmd_destroy_uc destroy_uc;
+       struct pvrdma_cmd_create_pd create_pd;
+       struct pvrdma_cmd_destroy_pd destroy_pd;
+       struct pvrdma_cmd_create_mr create_mr;
+       struct pvrdma_cmd_destroy_mr destroy_mr;
+       struct pvrdma_cmd_create_cq create_cq;
+       struct pvrdma_cmd_resize_cq resize_cq;
+       struct pvrdma_cmd_destroy_cq destroy_cq;
+       struct pvrdma_cmd_create_qp create_qp;
+       struct pvrdma_cmd_modify_qp modify_qp;
+       struct pvrdma_cmd_query_qp query_qp;
+       struct pvrdma_cmd_destroy_qp destroy_qp;
+       struct pvrdma_cmd_create_bind create_bind;
+       struct pvrdma_cmd_destroy_bind destroy_bind;
+};
+
+union pvrdma_cmd_resp {
+       struct pvrdma_cmd_resp_hdr hdr;
+       struct pvrdma_cmd_query_port_resp query_port_resp;
+       struct pvrdma_cmd_query_pkey_resp query_pkey_resp;
+       struct pvrdma_cmd_create_uc_resp create_uc_resp;
+       struct pvrdma_cmd_create_pd_resp create_pd_resp;
+       struct pvrdma_cmd_create_mr_resp create_mr_resp;
+       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_query_qp_resp query_qp_resp;
+       struct pvrdma_cmd_destroy_qp_resp destroy_qp_resp;
+};
+
+#endif /* __PVRDMA_DEV_API_H__ */
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_doorbell.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_doorbell.c
new file mode 100644 (file)
index 0000000..bf51357
--- /dev/null
@@ -0,0 +1,127 @@
+/*
+ * Copyright (c) 2012-2016 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 <linux/bitmap.h>
+#include <linux/errno.h>
+#include <linux/slab.h>
+
+#include "pvrdma.h"
+
+int pvrdma_uar_table_init(struct pvrdma_dev *dev)
+{
+       u32 num = dev->dsr->caps.max_uar;
+       u32 mask = num - 1;
+       struct pvrdma_id_table *tbl = &dev->uar_table.tbl;
+
+       if (!is_power_of_2(num))
+               return -EINVAL;
+
+       tbl->last = 0;
+       tbl->top = 0;
+       tbl->max = num;
+       tbl->mask = mask;
+       spin_lock_init(&tbl->lock);
+       tbl->table = kcalloc(BITS_TO_LONGS(num), sizeof(long), GFP_KERNEL);
+       if (!tbl->table)
+               return -ENOMEM;
+
+       /* 0th UAR is taken by the device. */
+       set_bit(0, tbl->table);
+
+       return 0;
+}
+
+void pvrdma_uar_table_cleanup(struct pvrdma_dev *dev)
+{
+       struct pvrdma_id_table *tbl = &dev->uar_table.tbl;
+
+       kfree(tbl->table);
+}
+
+int pvrdma_uar_alloc(struct pvrdma_dev *dev, struct pvrdma_uar_map *uar)
+{
+       struct pvrdma_id_table *tbl;
+       unsigned long flags;
+       u32 obj;
+
+       tbl = &dev->uar_table.tbl;
+
+       spin_lock_irqsave(&tbl->lock, flags);
+       obj = find_next_zero_bit(tbl->table, tbl->max, tbl->last);
+       if (obj >= tbl->max) {
+               tbl->top = (tbl->top + tbl->max) & tbl->mask;
+               obj = find_first_zero_bit(tbl->table, tbl->max);
+       }
+
+       if (obj >= tbl->max) {
+               spin_unlock_irqrestore(&tbl->lock, flags);
+               return -ENOMEM;
+       }
+
+       set_bit(obj, tbl->table);
+       obj |= tbl->top;
+
+       spin_unlock_irqrestore(&tbl->lock, flags);
+
+       uar->index = obj;
+       uar->pfn = (pci_resource_start(dev->pdev, PVRDMA_PCI_RESOURCE_UAR) >>
+                   PAGE_SHIFT) + uar->index;
+
+       return 0;
+}
+
+void pvrdma_uar_free(struct pvrdma_dev *dev, struct pvrdma_uar_map *uar)
+{
+       struct pvrdma_id_table *tbl = &dev->uar_table.tbl;
+       unsigned long flags;
+       u32 obj;
+
+       obj = uar->index & (tbl->max - 1);
+       spin_lock_irqsave(&tbl->lock, flags);
+       clear_bit(obj, tbl->table);
+       tbl->last = min(tbl->last, obj);
+       tbl->top = (tbl->top + tbl->max) & tbl->mask;
+       spin_unlock_irqrestore(&tbl->lock, flags);
+}
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_main.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_main.c
new file mode 100644 (file)
index 0000000..231a1ce
--- /dev/null
@@ -0,0 +1,1211 @@
+/*
+ * Copyright (c) 2012-2016 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 <linux/errno.h>
+#include <linux/inetdevice.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_user_verbs.h>
+#include <net/addrconf.h>
+
+#include "pvrdma.h"
+
+#define DRV_NAME       "vmw_pvrdma"
+#define DRV_VERSION    "1.0.0.0-k"
+
+static DEFINE_MUTEX(pvrdma_device_list_lock);
+static LIST_HEAD(pvrdma_device_list);
+static struct workqueue_struct *event_wq;
+
+static int pvrdma_add_gid(struct ib_device *ibdev,
+                         u8 port_num,
+                         unsigned int index,
+                         const union ib_gid *gid,
+                         const struct ib_gid_attr *attr,
+                         void **context);
+static int pvrdma_del_gid(struct ib_device *ibdev,
+                         u8 port_num,
+                         unsigned int index,
+                         void **context);
+
+
+static ssize_t show_hca(struct device *device, struct device_attribute *attr,
+                       char *buf)
+{
+       return sprintf(buf, "VMW_PVRDMA-%s\n", DRV_VERSION);
+}
+
+static ssize_t show_rev(struct device *device, struct device_attribute *attr,
+                       char *buf)
+{
+       return sprintf(buf, "%d\n", PVRDMA_REV_ID);
+}
+
+static ssize_t show_board(struct device *device, struct device_attribute *attr,
+                         char *buf)
+{
+       return sprintf(buf, "%d\n", PVRDMA_BOARD_ID);
+}
+
+static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,           NULL);
+static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,           NULL);
+static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
+
+static struct device_attribute *pvrdma_class_attributes[] = {
+       &dev_attr_hw_rev,
+       &dev_attr_hca_type,
+       &dev_attr_board_id
+};
+
+static void pvrdma_get_fw_ver_str(struct ib_device *device, char *str,
+                                 size_t str_len)
+{
+       struct pvrdma_dev *dev =
+               container_of(device, struct pvrdma_dev, ib_dev);
+       snprintf(str, str_len, "%d.%d.%d\n",
+                (int) (dev->dsr->caps.fw_ver >> 32),
+                (int) (dev->dsr->caps.fw_ver >> 16) & 0xffff,
+                (int) dev->dsr->caps.fw_ver & 0xffff);
+}
+
+static int pvrdma_init_device(struct pvrdma_dev *dev)
+{
+       /*  Initialize some device related stuff */
+       spin_lock_init(&dev->cmd_lock);
+       sema_init(&dev->cmd_sema, 1);
+       atomic_set(&dev->num_qps, 0);
+       atomic_set(&dev->num_cqs, 0);
+       atomic_set(&dev->num_pds, 0);
+       atomic_set(&dev->num_ahs, 0);
+
+       return 0;
+}
+
+static int pvrdma_port_immutable(struct ib_device *ibdev, u8 port_num,
+                                struct ib_port_immutable *immutable)
+{
+       struct ib_port_attr attr;
+       int err;
+
+       err = pvrdma_query_port(ibdev, port_num, &attr);
+       if (err)
+               return err;
+
+       immutable->pkey_tbl_len = attr.pkey_tbl_len;
+       immutable->gid_tbl_len = attr.gid_tbl_len;
+       immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
+       immutable->max_mad_size = IB_MGMT_MAD_SIZE;
+       return 0;
+}
+
+static struct net_device *pvrdma_get_netdev(struct ib_device *ibdev,
+                                           u8 port_num)
+{
+       struct net_device *netdev;
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+
+       if (port_num != 1)
+               return NULL;
+
+       rcu_read_lock();
+       netdev = dev->netdev;
+       if (netdev)
+               dev_hold(netdev);
+       rcu_read_unlock();
+
+       return netdev;
+}
+
+static int pvrdma_register_device(struct pvrdma_dev *dev)
+{
+       int ret = -1;
+       int i = 0;
+
+       strlcpy(dev->ib_dev.name, "vmw_pvrdma%d", IB_DEVICE_NAME_MAX);
+       dev->ib_dev.node_guid = dev->dsr->caps.node_guid;
+       dev->sys_image_guid = dev->dsr->caps.sys_image_guid;
+       dev->flags = 0;
+       dev->ib_dev.owner = THIS_MODULE;
+       dev->ib_dev.num_comp_vectors = 1;
+       dev->ib_dev.dma_device = &dev->pdev->dev;
+       dev->ib_dev.uverbs_abi_ver = PVRDMA_UVERBS_ABI_VERSION;
+       dev->ib_dev.uverbs_cmd_mask =
+               (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
+               (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
+               (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
+               (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
+               (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
+               (1ull << IB_USER_VERBS_CMD_REG_MR)              |
+               (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
+               (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
+               (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
+               (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
+               (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
+               (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
+               (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
+               (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
+               (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
+               (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
+               (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
+               (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
+               (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
+               (1ull << IB_USER_VERBS_CMD_DESTROY_AH);
+
+       dev->ib_dev.node_type = RDMA_NODE_IB_CA;
+       dev->ib_dev.phys_port_cnt = dev->dsr->caps.phys_port_cnt;
+
+       dev->ib_dev.query_device = pvrdma_query_device;
+       dev->ib_dev.query_port = pvrdma_query_port;
+       dev->ib_dev.query_gid = pvrdma_query_gid;
+       dev->ib_dev.query_pkey = pvrdma_query_pkey;
+       dev->ib_dev.modify_port = pvrdma_modify_port;
+       dev->ib_dev.alloc_ucontext = pvrdma_alloc_ucontext;
+       dev->ib_dev.dealloc_ucontext = pvrdma_dealloc_ucontext;
+       dev->ib_dev.mmap = pvrdma_mmap;
+       dev->ib_dev.alloc_pd = pvrdma_alloc_pd;
+       dev->ib_dev.dealloc_pd = pvrdma_dealloc_pd;
+       dev->ib_dev.create_ah = pvrdma_create_ah;
+       dev->ib_dev.destroy_ah = pvrdma_destroy_ah;
+       dev->ib_dev.create_qp = pvrdma_create_qp;
+       dev->ib_dev.modify_qp = pvrdma_modify_qp;
+       dev->ib_dev.query_qp = pvrdma_query_qp;
+       dev->ib_dev.destroy_qp = pvrdma_destroy_qp;
+       dev->ib_dev.post_send = pvrdma_post_send;
+       dev->ib_dev.post_recv = pvrdma_post_recv;
+       dev->ib_dev.create_cq = pvrdma_create_cq;
+       dev->ib_dev.modify_cq = pvrdma_modify_cq;
+       dev->ib_dev.resize_cq = pvrdma_resize_cq;
+       dev->ib_dev.destroy_cq = pvrdma_destroy_cq;
+       dev->ib_dev.poll_cq = pvrdma_poll_cq;
+       dev->ib_dev.req_notify_cq = pvrdma_req_notify_cq;
+       dev->ib_dev.get_dma_mr = pvrdma_get_dma_mr;
+       dev->ib_dev.reg_user_mr = pvrdma_reg_user_mr;
+       dev->ib_dev.dereg_mr = pvrdma_dereg_mr;
+       dev->ib_dev.alloc_mr = pvrdma_alloc_mr;
+       dev->ib_dev.map_mr_sg = pvrdma_map_mr_sg;
+       dev->ib_dev.add_gid = pvrdma_add_gid;
+       dev->ib_dev.del_gid = pvrdma_del_gid;
+       dev->ib_dev.get_netdev = pvrdma_get_netdev;
+       dev->ib_dev.get_port_immutable = pvrdma_port_immutable;
+       dev->ib_dev.get_link_layer = pvrdma_port_link_layer;
+       dev->ib_dev.get_dev_fw_str = pvrdma_get_fw_ver_str;
+
+       mutex_init(&dev->port_mutex);
+       spin_lock_init(&dev->desc_lock);
+
+       dev->cq_tbl = kcalloc(dev->dsr->caps.max_cq, sizeof(void *),
+                             GFP_KERNEL);
+       if (!dev->cq_tbl)
+               return ret;
+       spin_lock_init(&dev->cq_tbl_lock);
+
+       dev->qp_tbl = kcalloc(dev->dsr->caps.max_qp, sizeof(void *),
+                             GFP_KERNEL);
+       if (!dev->qp_tbl)
+               goto err_cq_free;
+       spin_lock_init(&dev->qp_tbl_lock);
+
+       ret = ib_register_device(&dev->ib_dev, NULL);
+       if (ret)
+               goto err_qp_free;
+
+       for (i = 0; i < ARRAY_SIZE(pvrdma_class_attributes); ++i) {
+               ret = device_create_file(&dev->ib_dev.dev,
+                                        pvrdma_class_attributes[i]);
+               if (ret)
+                       goto err_class;
+       }
+
+       dev->ib_active = true;
+
+       return 0;
+
+err_class:
+       ib_unregister_device(&dev->ib_dev);
+err_qp_free:
+       kfree(dev->qp_tbl);
+err_cq_free:
+       kfree(dev->cq_tbl);
+
+       return ret;
+}
+
+static irqreturn_t pvrdma_intr0_handler(int irq, void *dev_id)
+{
+       u32 icr = PVRDMA_INTR_CAUSE_RESPONSE;
+       struct pvrdma_dev *dev = dev_id;
+
+       dev_dbg(&dev->pdev->dev, "interrupt 0 (response) handler\n");
+
+       if (dev->intr.type != PVRDMA_INTR_TYPE_MSIX) {
+               /* Legacy intr */
+               icr = pvrdma_read_reg(dev, PVRDMA_REG_ICR);
+               if (icr == 0)
+                       return IRQ_NONE;
+       }
+
+       if (icr == PVRDMA_INTR_CAUSE_RESPONSE)
+               complete(&dev->cmd_done);
+
+       return IRQ_HANDLED;
+}
+
+static void pvrdma_qp_event(struct pvrdma_dev *dev, u32 qpn, int type)
+{
+       struct pvrdma_qp *qp;
+       unsigned long flags;
+
+       spin_lock_irqsave(&dev->qp_tbl_lock, flags);
+       qp = dev->qp_tbl[qpn % dev->dsr->caps.max_qp];
+       if (qp)
+               atomic_inc(&qp->refcnt);
+       spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
+
+       if (qp && qp->ibqp.event_handler) {
+               struct ib_qp *ibqp = &qp->ibqp;
+               struct ib_event e;
+
+               e.device = ibqp->device;
+               e.element.qp = ibqp;
+               e.event = type; /* 1:1 mapping for now. */
+               ibqp->event_handler(&e, ibqp->qp_context);
+       }
+       if (qp) {
+               atomic_dec(&qp->refcnt);
+               if (atomic_read(&qp->refcnt) == 0)
+                       wake_up(&qp->wait);
+       }
+}
+
+static void pvrdma_cq_event(struct pvrdma_dev *dev, u32 cqn, int type)
+{
+       struct pvrdma_cq *cq;
+       unsigned long flags;
+
+       spin_lock_irqsave(&dev->cq_tbl_lock, flags);
+       cq = dev->cq_tbl[cqn % dev->dsr->caps.max_cq];
+       if (cq)
+               atomic_inc(&cq->refcnt);
+       spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
+
+       if (cq && cq->ibcq.event_handler) {
+               struct ib_cq *ibcq = &cq->ibcq;
+               struct ib_event e;
+
+               e.device = ibcq->device;
+               e.element.cq = ibcq;
+               e.event = type; /* 1:1 mapping for now. */
+               ibcq->event_handler(&e, ibcq->cq_context);
+       }
+       if (cq) {
+               atomic_dec(&cq->refcnt);
+               if (atomic_read(&cq->refcnt) == 0)
+                       wake_up(&cq->wait);
+       }
+}
+
+static void pvrdma_dispatch_event(struct pvrdma_dev *dev, int port,
+                                 enum ib_event_type event)
+{
+       struct ib_event ib_event;
+
+       memset(&ib_event, 0, sizeof(ib_event));
+       ib_event.device = &dev->ib_dev;
+       ib_event.element.port_num = port;
+       ib_event.event = event;
+       ib_dispatch_event(&ib_event);
+}
+
+static void pvrdma_dev_event(struct pvrdma_dev *dev, u8 port, int type)
+{
+       if (port < 1 || port > dev->dsr->caps.phys_port_cnt) {
+               dev_warn(&dev->pdev->dev, "event on port %d\n", port);
+               return;
+       }
+
+       pvrdma_dispatch_event(dev, port, type);
+}
+
+static inline struct pvrdma_eqe *get_eqe(struct pvrdma_dev *dev, unsigned int i)
+{
+       return (struct pvrdma_eqe *)pvrdma_page_dir_get_ptr(
+                                       &dev->async_pdir,
+                                       PAGE_SIZE +
+                                       sizeof(struct pvrdma_eqe) * i);
+}
+
+static irqreturn_t pvrdma_intr1_handler(int irq, void *dev_id)
+{
+       struct pvrdma_dev *dev = dev_id;
+       struct pvrdma_ring *ring = &dev->async_ring_state->rx;
+       int ring_slots = (dev->dsr->async_ring_pages.num_pages - 1) *
+                        PAGE_SIZE / sizeof(struct pvrdma_eqe);
+       unsigned int head;
+
+       dev_dbg(&dev->pdev->dev, "interrupt 1 (async event) handler\n");
+
+       /*
+        * Don't process events until the IB device is registered. Otherwise
+        * we'll try to ib_dispatch_event() on an invalid device.
+        */
+       if (!dev->ib_active)
+               return IRQ_HANDLED;
+
+       while (pvrdma_idx_ring_has_data(ring, ring_slots, &head) > 0) {
+               struct pvrdma_eqe *eqe;
+
+               eqe = get_eqe(dev, head);
+
+               switch (eqe->type) {
+               case PVRDMA_EVENT_QP_FATAL:
+               case PVRDMA_EVENT_QP_REQ_ERR:
+               case PVRDMA_EVENT_QP_ACCESS_ERR:
+               case PVRDMA_EVENT_COMM_EST:
+               case PVRDMA_EVENT_SQ_DRAINED:
+               case PVRDMA_EVENT_PATH_MIG:
+               case PVRDMA_EVENT_PATH_MIG_ERR:
+               case PVRDMA_EVENT_QP_LAST_WQE_REACHED:
+                       pvrdma_qp_event(dev, eqe->info, eqe->type);
+                       break;
+
+               case PVRDMA_EVENT_CQ_ERR:
+                       pvrdma_cq_event(dev, eqe->info, eqe->type);
+                       break;
+
+               case PVRDMA_EVENT_SRQ_ERR:
+               case PVRDMA_EVENT_SRQ_LIMIT_REACHED:
+                       break;
+
+               case PVRDMA_EVENT_PORT_ACTIVE:
+               case PVRDMA_EVENT_PORT_ERR:
+               case PVRDMA_EVENT_LID_CHANGE:
+               case PVRDMA_EVENT_PKEY_CHANGE:
+               case PVRDMA_EVENT_SM_CHANGE:
+               case PVRDMA_EVENT_CLIENT_REREGISTER:
+               case PVRDMA_EVENT_GID_CHANGE:
+                       pvrdma_dev_event(dev, eqe->info, eqe->type);
+                       break;
+
+               case PVRDMA_EVENT_DEVICE_FATAL:
+                       pvrdma_dev_event(dev, 1, eqe->type);
+                       break;
+
+               default:
+                       break;
+               }
+
+               pvrdma_idx_ring_inc(&ring->cons_head, ring_slots);
+       }
+
+       return IRQ_HANDLED;
+}
+
+static inline struct pvrdma_cqne *get_cqne(struct pvrdma_dev *dev,
+                                          unsigned int i)
+{
+       return (struct pvrdma_cqne *)pvrdma_page_dir_get_ptr(
+                                       &dev->cq_pdir,
+                                       PAGE_SIZE +
+                                       sizeof(struct pvrdma_cqne) * i);
+}
+
+static irqreturn_t pvrdma_intrx_handler(int irq, void *dev_id)
+{
+       struct pvrdma_dev *dev = dev_id;
+       struct pvrdma_ring *ring = &dev->cq_ring_state->rx;
+       int ring_slots = (dev->dsr->cq_ring_pages.num_pages - 1) * PAGE_SIZE /
+                        sizeof(struct pvrdma_cqne);
+       unsigned int head;
+       unsigned long flags;
+
+       dev_dbg(&dev->pdev->dev, "interrupt x (completion) handler\n");
+
+       while (pvrdma_idx_ring_has_data(ring, ring_slots, &head) > 0) {
+               struct pvrdma_cqne *cqne;
+               struct pvrdma_cq *cq;
+
+               cqne = get_cqne(dev, head);
+               spin_lock_irqsave(&dev->cq_tbl_lock, flags);
+               cq = dev->cq_tbl[cqne->info % dev->dsr->caps.max_cq];
+               if (cq)
+                       atomic_inc(&cq->refcnt);
+               spin_unlock_irqrestore(&dev->cq_tbl_lock, flags);
+
+               if (cq && cq->ibcq.comp_handler)
+                       cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
+               if (cq) {
+                       atomic_dec(&cq->refcnt);
+                       if (atomic_read(&cq->refcnt))
+                               wake_up(&cq->wait);
+               }
+               pvrdma_idx_ring_inc(&ring->cons_head, ring_slots);
+       }
+
+       return IRQ_HANDLED;
+}
+
+static void pvrdma_disable_msi_all(struct pvrdma_dev *dev)
+{
+       if (dev->intr.type == PVRDMA_INTR_TYPE_MSIX)
+               pci_disable_msix(dev->pdev);
+       else if (dev->intr.type == PVRDMA_INTR_TYPE_MSI)
+               pci_disable_msi(dev->pdev);
+}
+
+static void pvrdma_free_irq(struct pvrdma_dev *dev)
+{
+       int i;
+
+       dev_dbg(&dev->pdev->dev, "freeing interrupts\n");
+
+       if (dev->intr.type == PVRDMA_INTR_TYPE_MSIX) {
+               for (i = 0; i < dev->intr.size; i++) {
+                       if (dev->intr.enabled[i]) {
+                               free_irq(dev->intr.msix_entry[i].vector, dev);
+                               dev->intr.enabled[i] = 0;
+                       }
+               }
+       } else if (dev->intr.type == PVRDMA_INTR_TYPE_INTX ||
+                  dev->intr.type == PVRDMA_INTR_TYPE_MSI) {
+               free_irq(dev->pdev->irq, dev);
+       }
+}
+
+static void pvrdma_enable_intrs(struct pvrdma_dev *dev)
+{
+       dev_dbg(&dev->pdev->dev, "enable interrupts\n");
+       pvrdma_write_reg(dev, PVRDMA_REG_IMR, 0);
+}
+
+static void pvrdma_disable_intrs(struct pvrdma_dev *dev)
+{
+       dev_dbg(&dev->pdev->dev, "disable interrupts\n");
+       pvrdma_write_reg(dev, PVRDMA_REG_IMR, ~0);
+}
+
+static int pvrdma_enable_msix(struct pci_dev *pdev, struct pvrdma_dev *dev)
+{
+       int i;
+       int ret;
+
+       for (i = 0; i < PVRDMA_MAX_INTERRUPTS; i++) {
+               dev->intr.msix_entry[i].entry = i;
+               dev->intr.msix_entry[i].vector = i;
+
+               switch (i) {
+               case 0:
+                       /* CMD ring handler */
+                       dev->intr.handler[i] = pvrdma_intr0_handler;
+                       break;
+               case 1:
+                       /* Async event ring handler */
+                       dev->intr.handler[i] = pvrdma_intr1_handler;
+                       break;
+               default:
+                       /* Completion queue handler */
+                       dev->intr.handler[i] = pvrdma_intrx_handler;
+                       break;
+               }
+       }
+
+       ret = pci_enable_msix(pdev, dev->intr.msix_entry,
+                             PVRDMA_MAX_INTERRUPTS);
+       if (!ret) {
+               dev->intr.type = PVRDMA_INTR_TYPE_MSIX;
+               dev->intr.size = PVRDMA_MAX_INTERRUPTS;
+       } else if (ret > 0) {
+               ret = pci_enable_msix(pdev, dev->intr.msix_entry, ret);
+               if (!ret) {
+                       dev->intr.type = PVRDMA_INTR_TYPE_MSIX;
+                       dev->intr.size = ret;
+               } else {
+                       dev->intr.size = 0;
+               }
+       }
+
+       dev_dbg(&pdev->dev, "using interrupt type %d, size %d\n",
+               dev->intr.type, dev->intr.size);
+
+       return ret;
+}
+
+static int pvrdma_alloc_intrs(struct pvrdma_dev *dev)
+{
+       int ret = 0;
+       int i;
+
+       if (pci_find_capability(dev->pdev, PCI_CAP_ID_MSIX) &&
+           pvrdma_enable_msix(dev->pdev, dev)) {
+               /* Try MSI */
+               ret = pci_enable_msi(dev->pdev);
+               if (!ret) {
+                       dev->intr.type = PVRDMA_INTR_TYPE_MSI;
+               } else {
+                       /* Legacy INTR */
+                       dev->intr.type = PVRDMA_INTR_TYPE_INTX;
+               }
+       }
+
+       /* Request First IRQ */
+       switch (dev->intr.type) {
+       case PVRDMA_INTR_TYPE_INTX:
+       case PVRDMA_INTR_TYPE_MSI:
+               ret = request_irq(dev->pdev->irq, pvrdma_intr0_handler,
+                                 IRQF_SHARED, DRV_NAME, dev);
+               if (ret) {
+                       dev_err(&dev->pdev->dev,
+                               "failed to request interrupt\n");
+                       goto disable_msi;
+               }
+               break;
+       case PVRDMA_INTR_TYPE_MSIX:
+               ret = request_irq(dev->intr.msix_entry[0].vector,
+                                 pvrdma_intr0_handler, 0, DRV_NAME, dev);
+               if (ret) {
+                       dev_err(&dev->pdev->dev,
+                               "failed to request interrupt 0\n");
+                       goto disable_msi;
+               }
+               dev->intr.enabled[0] = 1;
+               break;
+       default:
+               /* Not reached */
+               break;
+       }
+
+       /* For MSIX: request intr for each vector */
+       if (dev->intr.size > 1) {
+               ret = request_irq(dev->intr.msix_entry[1].vector,
+                                 pvrdma_intr1_handler, 0, DRV_NAME, dev);
+               if (ret) {
+                       dev_err(&dev->pdev->dev,
+                               "failed to request interrupt 1\n");
+                       goto free_irq;
+               }
+               dev->intr.enabled[1] = 1;
+
+               for (i = 2; i < dev->intr.size; i++) {
+                       ret = request_irq(dev->intr.msix_entry[i].vector,
+                                         pvrdma_intrx_handler, 0,
+                                         DRV_NAME, dev);
+                       if (ret) {
+                               dev_err(&dev->pdev->dev,
+                                       "failed to request interrupt %d\n", i);
+                               goto free_irq;
+                       }
+                       dev->intr.enabled[i] = 1;
+               }
+       }
+
+       return 0;
+
+free_irq:
+       pvrdma_free_irq(dev);
+disable_msi:
+       pvrdma_disable_msi_all(dev);
+       return ret;
+}
+
+static void pvrdma_free_slots(struct pvrdma_dev *dev)
+{
+       struct pci_dev *pdev = dev->pdev;
+
+       if (dev->resp_slot)
+               dma_free_coherent(&pdev->dev, PAGE_SIZE, dev->resp_slot,
+                                 dev->dsr->resp_slot_dma);
+       if (dev->cmd_slot)
+               dma_free_coherent(&pdev->dev, PAGE_SIZE, dev->cmd_slot,
+                                 dev->dsr->cmd_slot_dma);
+}
+
+static int pvrdma_add_gid_at_index(struct pvrdma_dev *dev,
+                                  const union ib_gid *gid,
+                                  int index)
+{
+       int ret;
+       union pvrdma_cmd_req req;
+       struct pvrdma_cmd_create_bind *cmd_bind = &req.create_bind;
+
+       if (!dev->sgid_tbl) {
+               dev_warn(&dev->pdev->dev, "sgid table not initialized\n");
+               return -EINVAL;
+       }
+
+       memset(cmd_bind, 0, sizeof(*cmd_bind));
+       cmd_bind->hdr.cmd = PVRDMA_CMD_CREATE_BIND;
+       memcpy(cmd_bind->new_gid, gid->raw, 16);
+       cmd_bind->mtu = ib_mtu_enum_to_int(IB_MTU_1024);
+       cmd_bind->vlan = 0xfff;
+       cmd_bind->index = index;
+       cmd_bind->gid_type = PVRDMA_GID_TYPE_FLAG_ROCE_V1;
+
+       ret = pvrdma_cmd_post(dev, &req, NULL, 0);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not create binding, error: %d\n", ret);
+               return -EFAULT;
+       }
+       memcpy(&dev->sgid_tbl[index], gid, sizeof(*gid));
+       return 0;
+}
+
+static int pvrdma_add_gid(struct ib_device *ibdev,
+                         u8 port_num,
+                         unsigned int index,
+                         const union ib_gid *gid,
+                         const struct ib_gid_attr *attr,
+                         void **context)
+{
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+
+       return pvrdma_add_gid_at_index(dev, gid, index);
+}
+
+static int pvrdma_del_gid_at_index(struct pvrdma_dev *dev, int index)
+{
+       int ret;
+       union pvrdma_cmd_req req;
+       struct pvrdma_cmd_destroy_bind *cmd_dest = &req.destroy_bind;
+
+       /* Update sgid table. */
+       if (!dev->sgid_tbl) {
+               dev_warn(&dev->pdev->dev, "sgid table not initialized\n");
+               return -EINVAL;
+       }
+
+       memset(cmd_dest, 0, sizeof(*cmd_dest));
+       cmd_dest->hdr.cmd = PVRDMA_CMD_DESTROY_BIND;
+       memcpy(cmd_dest->dest_gid, &dev->sgid_tbl[index], 16);
+       cmd_dest->index = index;
+
+       ret = pvrdma_cmd_post(dev, &req, NULL, 0);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not destroy binding, error: %d\n", ret);
+               return ret;
+       }
+       memset(&dev->sgid_tbl[index], 0, 16);
+       return 0;
+}
+
+static int pvrdma_del_gid(struct ib_device *ibdev,
+                         u8 port_num,
+                         unsigned int index,
+                         void **context)
+{
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+
+       dev_dbg(&dev->pdev->dev, "removing gid at index %u from %s",
+               index, dev->netdev->name);
+
+       return pvrdma_del_gid_at_index(dev, index);
+}
+
+static void pvrdma_netdevice_event_handle(struct pvrdma_dev *dev,
+                                         unsigned long event)
+{
+       switch (event) {
+       case NETDEV_REBOOT:
+       case NETDEV_DOWN:
+               pvrdma_dispatch_event(dev, 1, IB_EVENT_PORT_ERR);
+               break;
+       case NETDEV_UP:
+               pvrdma_dispatch_event(dev, 1, IB_EVENT_PORT_ACTIVE);
+               break;
+       default:
+               dev_dbg(&dev->pdev->dev, "ignore netdevice event %ld on %s\n",
+                       event, dev->ib_dev.name);
+               break;
+       }
+}
+
+static void pvrdma_netdevice_event_work(struct work_struct *work)
+{
+       struct pvrdma_netdevice_work *netdev_work;
+       struct pvrdma_dev *dev;
+
+       netdev_work = container_of(work, struct pvrdma_netdevice_work, work);
+
+       mutex_lock(&pvrdma_device_list_lock);
+       list_for_each_entry(dev, &pvrdma_device_list, device_link) {
+               if (dev->netdev == netdev_work->event_netdev) {
+                       pvrdma_netdevice_event_handle(dev, netdev_work->event);
+                       break;
+               }
+       }
+       mutex_unlock(&pvrdma_device_list_lock);
+
+       kfree(netdev_work);
+}
+
+static int pvrdma_netdevice_event(struct notifier_block *this,
+                                 unsigned long event, void *ptr)
+{
+       struct net_device *event_netdev = netdev_notifier_info_to_dev(ptr);
+       struct pvrdma_netdevice_work *netdev_work;
+
+       netdev_work = kmalloc(sizeof(*netdev_work), GFP_ATOMIC);
+       if (!netdev_work)
+               return NOTIFY_BAD;
+
+       INIT_WORK(&netdev_work->work, pvrdma_netdevice_event_work);
+       netdev_work->event_netdev = event_netdev;
+       netdev_work->event = event;
+       queue_work(event_wq, &netdev_work->work);
+
+       return NOTIFY_DONE;
+}
+
+static int pvrdma_pci_probe(struct pci_dev *pdev,
+                           const struct pci_device_id *id)
+{
+       struct pci_dev *pdev_net;
+       struct pvrdma_dev *dev;
+       int ret;
+       unsigned long start;
+       unsigned long len;
+       unsigned int version;
+       dma_addr_t slot_dma = 0;
+
+       dev_dbg(&pdev->dev, "initializing driver %s\n", pci_name(pdev));
+
+       /* Allocate zero-out device */
+       dev = (struct pvrdma_dev *)ib_alloc_device(sizeof(*dev));
+       if (!dev) {
+               dev_err(&pdev->dev, "failed to allocate IB device\n");
+               return -ENOMEM;
+       }
+
+       mutex_lock(&pvrdma_device_list_lock);
+       list_add(&dev->device_link, &pvrdma_device_list);
+       mutex_unlock(&pvrdma_device_list_lock);
+
+       ret = pvrdma_init_device(dev);
+       if (ret)
+               goto err_free_device;
+
+       dev->pdev = pdev;
+       pci_set_drvdata(pdev, dev);
+
+       ret = pci_enable_device(pdev);
+       if (ret) {
+               dev_err(&pdev->dev, "cannot enable PCI device\n");
+               goto err_free_device;
+       }
+
+       dev_dbg(&pdev->dev, "PCI resource flags BAR0 %#lx\n",
+               pci_resource_flags(pdev, 0));
+       dev_dbg(&pdev->dev, "PCI resource len %#llx\n",
+               (unsigned long long)pci_resource_len(pdev, 0));
+       dev_dbg(&pdev->dev, "PCI resource start %#llx\n",
+               (unsigned long long)pci_resource_start(pdev, 0));
+       dev_dbg(&pdev->dev, "PCI resource flags BAR1 %#lx\n",
+               pci_resource_flags(pdev, 1));
+       dev_dbg(&pdev->dev, "PCI resource len %#llx\n",
+               (unsigned long long)pci_resource_len(pdev, 1));
+       dev_dbg(&pdev->dev, "PCI resource start %#llx\n",
+               (unsigned long long)pci_resource_start(pdev, 1));
+
+       if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
+           !(pci_resource_flags(pdev, 1) & IORESOURCE_MEM)) {
+               dev_err(&pdev->dev, "PCI BAR region not MMIO\n");
+               ret = -ENOMEM;
+               goto err_free_device;
+       }
+
+       ret = pci_request_regions(pdev, DRV_NAME);
+       if (ret) {
+               dev_err(&pdev->dev, "cannot request PCI resources\n");
+               goto err_disable_pdev;
+       }
+
+       /* Enable 64-Bit DMA */
+       if (pci_set_dma_mask(pdev, DMA_BIT_MASK(64)) == 0) {
+               ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
+               if (ret != 0) {
+                       dev_err(&pdev->dev,
+                               "pci_set_consistent_dma_mask failed\n");
+                       goto err_free_resource;
+               }
+       } else {
+               ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
+               if (ret != 0) {
+                       dev_err(&pdev->dev,
+                               "pci_set_dma_mask failed\n");
+                       goto err_free_resource;
+               }
+       }
+
+       pci_set_master(pdev);
+
+       /* Map register space */
+       start = pci_resource_start(dev->pdev, PVRDMA_PCI_RESOURCE_REG);
+       len = pci_resource_len(dev->pdev, PVRDMA_PCI_RESOURCE_REG);
+       dev->regs = ioremap(start, len);
+       if (!dev->regs) {
+               dev_err(&pdev->dev, "register mapping failed\n");
+               ret = -ENOMEM;
+               goto err_free_resource;
+       }
+
+       /* Setup per-device UAR. */
+       dev->driver_uar.index = 0;
+       dev->driver_uar.pfn =
+               pci_resource_start(dev->pdev, PVRDMA_PCI_RESOURCE_UAR) >>
+               PAGE_SHIFT;
+       dev->driver_uar.map =
+               ioremap(dev->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
+       if (!dev->driver_uar.map) {
+               dev_err(&pdev->dev, "failed to remap UAR pages\n");
+               ret = -ENOMEM;
+               goto err_unmap_regs;
+       }
+
+       version = pvrdma_read_reg(dev, PVRDMA_REG_VERSION);
+       dev_info(&pdev->dev, "device version %d, driver version %d\n",
+                version, PVRDMA_VERSION);
+       if (version < PVRDMA_VERSION) {
+               dev_err(&pdev->dev, "incompatible device version\n");
+               goto err_uar_unmap;
+       }
+
+       dev->dsr = dma_alloc_coherent(&pdev->dev, sizeof(*dev->dsr),
+                                     &dev->dsrbase, GFP_KERNEL);
+       if (!dev->dsr) {
+               dev_err(&pdev->dev, "failed to allocate shared region\n");
+               ret = -ENOMEM;
+               goto err_uar_unmap;
+       }
+
+       /* Setup the shared region */
+       memset(dev->dsr, 0, sizeof(*dev->dsr));
+       dev->dsr->driver_version = PVRDMA_VERSION;
+       dev->dsr->gos_info.gos_bits = sizeof(void *) == 4 ?
+               PVRDMA_GOS_BITS_32 :
+               PVRDMA_GOS_BITS_64;
+       dev->dsr->gos_info.gos_type = PVRDMA_GOS_TYPE_LINUX;
+       dev->dsr->gos_info.gos_ver = 1;
+       dev->dsr->uar_pfn = dev->driver_uar.pfn;
+
+       /* Command slot. */
+       dev->cmd_slot = dma_alloc_coherent(&pdev->dev, PAGE_SIZE,
+                                          &slot_dma, GFP_KERNEL);
+       if (!dev->cmd_slot) {
+               ret = -ENOMEM;
+               goto err_free_dsr;
+       }
+
+       dev->dsr->cmd_slot_dma = (u64)slot_dma;
+
+       /* Response slot. */
+       dev->resp_slot = dma_alloc_coherent(&pdev->dev, PAGE_SIZE,
+                                           &slot_dma, GFP_KERNEL);
+       if (!dev->resp_slot) {
+               ret = -ENOMEM;
+               goto err_free_slots;
+       }
+
+       dev->dsr->resp_slot_dma = (u64)slot_dma;
+
+       /* Async event ring */
+       dev->dsr->async_ring_pages.num_pages = 4;
+       ret = pvrdma_page_dir_init(dev, &dev->async_pdir,
+                                  dev->dsr->async_ring_pages.num_pages, true);
+       if (ret)
+               goto err_free_slots;
+       dev->async_ring_state = dev->async_pdir.pages[0];
+       dev->dsr->async_ring_pages.pdir_dma = dev->async_pdir.dir_dma;
+
+       /* CQ notification ring */
+       dev->dsr->cq_ring_pages.num_pages = 4;
+       ret = pvrdma_page_dir_init(dev, &dev->cq_pdir,
+                                  dev->dsr->cq_ring_pages.num_pages, true);
+       if (ret)
+               goto err_free_async_ring;
+       dev->cq_ring_state = dev->cq_pdir.pages[0];
+       dev->dsr->cq_ring_pages.pdir_dma = dev->cq_pdir.dir_dma;
+
+       /*
+        * Write the PA of the shared region to the device. The writes must be
+        * ordered such that the high bits are written last. When the writes
+        * complete, the device will have filled out the capabilities.
+        */
+
+       pvrdma_write_reg(dev, PVRDMA_REG_DSRLOW, (u32)dev->dsrbase);
+       pvrdma_write_reg(dev, PVRDMA_REG_DSRHIGH,
+                        (u32)((u64)(dev->dsrbase) >> 32));
+
+       /* Make sure the write is complete before reading status. */
+       mb();
+
+       /* Currently, the driver only supports RoCE mode. */
+       if (dev->dsr->caps.mode != PVRDMA_DEVICE_MODE_ROCE) {
+               dev_err(&pdev->dev, "unsupported transport %d\n",
+                       dev->dsr->caps.mode);
+               ret = -EFAULT;
+               goto err_free_cq_ring;
+       }
+
+       /* Currently, the driver only supports RoCE V1. */
+       if (!(dev->dsr->caps.gid_types & PVRDMA_GID_TYPE_FLAG_ROCE_V1)) {
+               dev_err(&pdev->dev, "driver needs RoCE v1 support\n");
+               ret = -EFAULT;
+               goto err_free_cq_ring;
+       }
+
+       /* Paired vmxnet3 will have same bus, slot. But func will be 0 */
+       pdev_net = pci_get_slot(pdev->bus, PCI_DEVFN(PCI_SLOT(pdev->devfn), 0));
+       if (!pdev_net) {
+               dev_err(&pdev->dev, "failed to find paired net device\n");
+               ret = -ENODEV;
+               goto err_free_cq_ring;
+       }
+
+       if (pdev_net->vendor != PCI_VENDOR_ID_VMWARE ||
+           pdev_net->device != PCI_DEVICE_ID_VMWARE_VMXNET3) {
+               dev_err(&pdev->dev, "failed to find paired vmxnet3 device\n");
+               pci_dev_put(pdev_net);
+               ret = -ENODEV;
+               goto err_free_cq_ring;
+       }
+
+       dev->netdev = pci_get_drvdata(pdev_net);
+       pci_dev_put(pdev_net);
+       if (!dev->netdev) {
+               dev_err(&pdev->dev, "failed to get vmxnet3 device\n");
+               ret = -ENODEV;
+               goto err_free_cq_ring;
+       }
+
+       dev_info(&pdev->dev, "paired device to %s\n", dev->netdev->name);
+
+       /* Interrupt setup */
+       ret = pvrdma_alloc_intrs(dev);
+       if (ret) {
+               dev_err(&pdev->dev, "failed to allocate interrupts\n");
+               ret = -ENOMEM;
+               goto err_netdevice;
+       }
+
+       /* Allocate UAR table. */
+       ret = pvrdma_uar_table_init(dev);
+       if (ret) {
+               dev_err(&pdev->dev, "failed to allocate UAR table\n");
+               ret = -ENOMEM;
+               goto err_free_intrs;
+       }
+
+       /* Allocate GID table */
+       dev->sgid_tbl = kcalloc(dev->dsr->caps.gid_tbl_len,
+                               sizeof(union ib_gid), GFP_KERNEL);
+       if (!dev->sgid_tbl) {
+               ret = -ENOMEM;
+               goto err_free_uar_table;
+       }
+       dev_dbg(&pdev->dev, "gid table len %d\n", dev->dsr->caps.gid_tbl_len);
+
+       pvrdma_enable_intrs(dev);
+
+       /* Activate pvrdma device */
+       pvrdma_write_reg(dev, PVRDMA_REG_CTL, PVRDMA_DEVICE_CTL_ACTIVATE);
+
+       /* Make sure the write is complete before reading status. */
+       mb();
+
+       /* Check if device was successfully activated */
+       ret = pvrdma_read_reg(dev, PVRDMA_REG_ERR);
+       if (ret != 0) {
+               dev_err(&pdev->dev, "failed to activate device\n");
+               ret = -EFAULT;
+               goto err_disable_intr;
+       }
+
+       /* Register IB device */
+       ret = pvrdma_register_device(dev);
+       if (ret) {
+               dev_err(&pdev->dev, "failed to register IB device\n");
+               goto err_disable_intr;
+       }
+
+       dev->nb_netdev.notifier_call = pvrdma_netdevice_event;
+       ret = register_netdevice_notifier(&dev->nb_netdev);
+       if (ret) {
+               dev_err(&pdev->dev, "failed to register netdevice events\n");
+               goto err_unreg_ibdev;
+       }
+
+       dev_info(&pdev->dev, "attached to device\n");
+       return 0;
+
+err_unreg_ibdev:
+       ib_unregister_device(&dev->ib_dev);
+err_disable_intr:
+       pvrdma_disable_intrs(dev);
+       kfree(dev->sgid_tbl);
+err_free_uar_table:
+       pvrdma_uar_table_cleanup(dev);
+err_free_intrs:
+       pvrdma_free_irq(dev);
+       pvrdma_disable_msi_all(dev);
+err_netdevice:
+       unregister_netdevice_notifier(&dev->nb_netdev);
+err_free_cq_ring:
+       pvrdma_page_dir_cleanup(dev, &dev->cq_pdir);
+err_free_async_ring:
+       pvrdma_page_dir_cleanup(dev, &dev->async_pdir);
+err_free_slots:
+       pvrdma_free_slots(dev);
+err_free_dsr:
+       dma_free_coherent(&pdev->dev, sizeof(*dev->dsr), dev->dsr,
+                         dev->dsrbase);
+err_uar_unmap:
+       iounmap(dev->driver_uar.map);
+err_unmap_regs:
+       iounmap(dev->regs);
+err_free_resource:
+       pci_release_regions(pdev);
+err_disable_pdev:
+       pci_disable_device(pdev);
+       pci_set_drvdata(pdev, NULL);
+err_free_device:
+       mutex_lock(&pvrdma_device_list_lock);
+       list_del(&dev->device_link);
+       mutex_unlock(&pvrdma_device_list_lock);
+       ib_dealloc_device(&dev->ib_dev);
+       return ret;
+}
+
+static void pvrdma_pci_remove(struct pci_dev *pdev)
+{
+       struct pvrdma_dev *dev = pci_get_drvdata(pdev);
+
+       if (!dev)
+               return;
+
+       dev_info(&pdev->dev, "detaching from device\n");
+
+       unregister_netdevice_notifier(&dev->nb_netdev);
+       dev->nb_netdev.notifier_call = NULL;
+
+       flush_workqueue(event_wq);
+
+       /* Unregister ib device */
+       ib_unregister_device(&dev->ib_dev);
+
+       mutex_lock(&pvrdma_device_list_lock);
+       list_del(&dev->device_link);
+       mutex_unlock(&pvrdma_device_list_lock);
+
+       pvrdma_disable_intrs(dev);
+       pvrdma_free_irq(dev);
+       pvrdma_disable_msi_all(dev);
+
+       /* Deactivate pvrdma device */
+       pvrdma_write_reg(dev, PVRDMA_REG_CTL, PVRDMA_DEVICE_CTL_RESET);
+       pvrdma_page_dir_cleanup(dev, &dev->cq_pdir);
+       pvrdma_page_dir_cleanup(dev, &dev->async_pdir);
+       pvrdma_free_slots(dev);
+
+       iounmap(dev->regs);
+       kfree(dev->sgid_tbl);
+       kfree(dev->cq_tbl);
+       kfree(dev->qp_tbl);
+       pvrdma_uar_table_cleanup(dev);
+       iounmap(dev->driver_uar.map);
+
+       ib_dealloc_device(&dev->ib_dev);
+
+       /* Free pci resources */
+       pci_release_regions(pdev);
+       pci_disable_device(pdev);
+       pci_set_drvdata(pdev, NULL);
+}
+
+static struct pci_device_id pvrdma_pci_table[] = {
+       { PCI_DEVICE(PCI_VENDOR_ID_VMWARE, PCI_DEVICE_ID_VMWARE_PVRDMA), },
+       { 0 },
+};
+
+MODULE_DEVICE_TABLE(pci, pvrdma_pci_table);
+
+static struct pci_driver pvrdma_driver = {
+       .name           = DRV_NAME,
+       .id_table       = pvrdma_pci_table,
+       .probe          = pvrdma_pci_probe,
+       .remove         = pvrdma_pci_remove,
+};
+
+static int __init pvrdma_init(void)
+{
+       int err;
+
+       event_wq = alloc_ordered_workqueue("pvrdma_event_wq", WQ_MEM_RECLAIM);
+       if (!event_wq)
+               return -ENOMEM;
+
+       err = pci_register_driver(&pvrdma_driver);
+       if (err)
+               destroy_workqueue(event_wq);
+
+       return err;
+}
+
+static void __exit pvrdma_cleanup(void)
+{
+       pci_unregister_driver(&pvrdma_driver);
+
+       destroy_workqueue(event_wq);
+}
+
+module_init(pvrdma_init);
+module_exit(pvrdma_cleanup);
+
+MODULE_AUTHOR("VMware, Inc");
+MODULE_DESCRIPTION("VMware Paravirtual RDMA driver");
+MODULE_VERSION(DRV_VERSION);
+MODULE_LICENSE("Dual BSD/GPL");
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_misc.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_misc.c
new file mode 100644 (file)
index 0000000..948b5cc
--- /dev/null
@@ -0,0 +1,304 @@
+/*
+ * Copyright (c) 2012-2016 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 <linux/errno.h>
+#include <linux/slab.h>
+#include <linux/bitmap.h>
+
+#include "pvrdma.h"
+
+int pvrdma_page_dir_init(struct pvrdma_dev *dev, struct pvrdma_page_dir *pdir,
+                        u64 npages, bool alloc_pages)
+{
+       u64 i;
+
+       if (npages > PVRDMA_PAGE_DIR_MAX_PAGES)
+               return -EINVAL;
+
+       memset(pdir, 0, sizeof(*pdir));
+
+       pdir->dir = dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
+                                      &pdir->dir_dma, GFP_KERNEL);
+       if (!pdir->dir)
+               goto err;
+
+       pdir->ntables = PVRDMA_PAGE_DIR_TABLE(npages - 1) + 1;
+       pdir->tables = kcalloc(pdir->ntables, sizeof(*pdir->tables),
+                              GFP_KERNEL);
+       if (!pdir->tables)
+               goto err;
+
+       for (i = 0; i < pdir->ntables; i++) {
+               pdir->tables[i] = dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
+                                               (dma_addr_t *)&pdir->dir[i],
+                                               GFP_KERNEL);
+               if (!pdir->tables[i])
+                       goto err;
+       }
+
+       pdir->npages = npages;
+
+       if (alloc_pages) {
+               pdir->pages = kcalloc(npages, sizeof(*pdir->pages),
+                                     GFP_KERNEL);
+               if (!pdir->pages)
+                       goto err;
+
+               for (i = 0; i < pdir->npages; i++) {
+                       dma_addr_t page_dma;
+
+                       pdir->pages[i] = dma_alloc_coherent(&dev->pdev->dev,
+                                                           PAGE_SIZE,
+                                                           &page_dma,
+                                                           GFP_KERNEL);
+                       if (!pdir->pages[i])
+                               goto err;
+
+                       pvrdma_page_dir_insert_dma(pdir, i, page_dma);
+               }
+       }
+
+       return 0;
+
+err:
+       pvrdma_page_dir_cleanup(dev, pdir);
+
+       return -ENOMEM;
+}
+
+static u64 *pvrdma_page_dir_table(struct pvrdma_page_dir *pdir, u64 idx)
+{
+       return pdir->tables[PVRDMA_PAGE_DIR_TABLE(idx)];
+}
+
+dma_addr_t pvrdma_page_dir_get_dma(struct pvrdma_page_dir *pdir, u64 idx)
+{
+       return pvrdma_page_dir_table(pdir, idx)[PVRDMA_PAGE_DIR_PAGE(idx)];
+}
+
+static void pvrdma_page_dir_cleanup_pages(struct pvrdma_dev *dev,
+                                         struct pvrdma_page_dir *pdir)
+{
+       if (pdir->pages) {
+               u64 i;
+
+               for (i = 0; i < pdir->npages && pdir->pages[i]; i++) {
+                       dma_addr_t page_dma = pvrdma_page_dir_get_dma(pdir, i);
+
+                       dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+                                         pdir->pages[i], page_dma);
+               }
+
+               kfree(pdir->pages);
+       }
+}
+
+static void pvrdma_page_dir_cleanup_tables(struct pvrdma_dev *dev,
+                                          struct pvrdma_page_dir *pdir)
+{
+       if (pdir->tables) {
+               int i;
+
+               pvrdma_page_dir_cleanup_pages(dev, pdir);
+
+               for (i = 0; i < pdir->ntables; i++) {
+                       u64 *table = pdir->tables[i];
+
+                       if (table)
+                               dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+                                                 table, pdir->dir[i]);
+               }
+
+               kfree(pdir->tables);
+       }
+}
+
+void pvrdma_page_dir_cleanup(struct pvrdma_dev *dev,
+                            struct pvrdma_page_dir *pdir)
+{
+       if (pdir->dir) {
+               pvrdma_page_dir_cleanup_tables(dev, pdir);
+               dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+                                 pdir->dir, pdir->dir_dma);
+       }
+}
+
+int pvrdma_page_dir_insert_dma(struct pvrdma_page_dir *pdir, u64 idx,
+                              dma_addr_t daddr)
+{
+       u64 *table;
+
+       if (idx >= pdir->npages)
+               return -EINVAL;
+
+       table = pvrdma_page_dir_table(pdir, idx);
+       table[PVRDMA_PAGE_DIR_PAGE(idx)] = daddr;
+
+       return 0;
+}
+
+int pvrdma_page_dir_insert_umem(struct pvrdma_page_dir *pdir,
+                               struct ib_umem *umem, u64 offset)
+{
+       u64 i = offset;
+       int j, entry;
+       int ret = 0, len = 0;
+       struct scatterlist *sg;
+
+       if (offset >= pdir->npages)
+               return -EINVAL;
+
+       for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
+               len = sg_dma_len(sg) >> PAGE_SHIFT;
+               for (j = 0; j < len; j++) {
+                       dma_addr_t addr = sg_dma_address(sg) +
+                                         umem->page_size * j;
+
+                       ret = pvrdma_page_dir_insert_dma(pdir, i, addr);
+                       if (ret)
+                               goto exit;
+
+                       i++;
+               }
+       }
+
+exit:
+       return ret;
+}
+
+int pvrdma_page_dir_insert_page_list(struct pvrdma_page_dir *pdir,
+                                    u64 *page_list,
+                                    int num_pages)
+{
+       int i;
+       int ret;
+
+       if (num_pages > pdir->npages)
+               return -EINVAL;
+
+       for (i = 0; i < num_pages; i++) {
+               ret = pvrdma_page_dir_insert_dma(pdir, i, page_list[i]);
+               if (ret)
+                       return ret;
+       }
+
+       return 0;
+}
+
+void pvrdma_qp_cap_to_ib(struct ib_qp_cap *dst, const struct pvrdma_qp_cap *src)
+{
+       dst->max_send_wr = src->max_send_wr;
+       dst->max_recv_wr = src->max_recv_wr;
+       dst->max_send_sge = src->max_send_sge;
+       dst->max_recv_sge = src->max_recv_sge;
+       dst->max_inline_data = src->max_inline_data;
+}
+
+void ib_qp_cap_to_pvrdma(struct pvrdma_qp_cap *dst, const struct ib_qp_cap *src)
+{
+       dst->max_send_wr = src->max_send_wr;
+       dst->max_recv_wr = src->max_recv_wr;
+       dst->max_send_sge = src->max_send_sge;
+       dst->max_recv_sge = src->max_recv_sge;
+       dst->max_inline_data = src->max_inline_data;
+}
+
+void pvrdma_gid_to_ib(union ib_gid *dst, const union pvrdma_gid *src)
+{
+       BUILD_BUG_ON(sizeof(union pvrdma_gid) != sizeof(union ib_gid));
+       memcpy(dst, src, sizeof(*src));
+}
+
+void ib_gid_to_pvrdma(union pvrdma_gid *dst, const union ib_gid *src)
+{
+       BUILD_BUG_ON(sizeof(union pvrdma_gid) != sizeof(union ib_gid));
+       memcpy(dst, src, sizeof(*src));
+}
+
+void pvrdma_global_route_to_ib(struct ib_global_route *dst,
+                              const struct pvrdma_global_route *src)
+{
+       pvrdma_gid_to_ib(&dst->dgid, &src->dgid);
+       dst->flow_label = src->flow_label;
+       dst->sgid_index = src->sgid_index;
+       dst->hop_limit = src->hop_limit;
+       dst->traffic_class = src->traffic_class;
+}
+
+void ib_global_route_to_pvrdma(struct pvrdma_global_route *dst,
+                              const struct ib_global_route *src)
+{
+       ib_gid_to_pvrdma(&dst->dgid, &src->dgid);
+       dst->flow_label = src->flow_label;
+       dst->sgid_index = src->sgid_index;
+       dst->hop_limit = src->hop_limit;
+       dst->traffic_class = src->traffic_class;
+}
+
+void pvrdma_ah_attr_to_ib(struct ib_ah_attr *dst,
+                         const struct pvrdma_ah_attr *src)
+{
+       pvrdma_global_route_to_ib(&dst->grh, &src->grh);
+       dst->dlid = src->dlid;
+       dst->sl = src->sl;
+       dst->src_path_bits = src->src_path_bits;
+       dst->static_rate = src->static_rate;
+       dst->ah_flags = src->ah_flags;
+       dst->port_num = src->port_num;
+       memcpy(&dst->dmac, &src->dmac, sizeof(dst->dmac));
+}
+
+void ib_ah_attr_to_pvrdma(struct pvrdma_ah_attr *dst,
+                         const struct ib_ah_attr *src)
+{
+       ib_global_route_to_pvrdma(&dst->grh, &src->grh);
+       dst->dlid = src->dlid;
+       dst->sl = src->sl;
+       dst->src_path_bits = src->src_path_bits;
+       dst->static_rate = src->static_rate;
+       dst->ah_flags = src->ah_flags;
+       dst->port_num = src->port_num;
+       memcpy(&dst->dmac, &src->dmac, sizeof(dst->dmac));
+}
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_mr.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_mr.c
new file mode 100644 (file)
index 0000000..8519f32
--- /dev/null
@@ -0,0 +1,334 @@
+/*
+ * Copyright (c) 2012-2016 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 <linux/list.h>
+#include <linux/slab.h>
+
+#include "pvrdma.h"
+
+/**
+ * pvrdma_get_dma_mr - get a DMA memory region
+ * @pd: protection domain
+ * @acc: access flags
+ *
+ * @return: ib_mr pointer on success, otherwise returns an errno.
+ */
+struct ib_mr *pvrdma_get_dma_mr(struct ib_pd *pd, int acc)
+{
+       struct pvrdma_dev *dev = to_vdev(pd->device);
+       struct pvrdma_user_mr *mr;
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_create_mr *cmd = &req.create_mr;
+       struct pvrdma_cmd_create_mr_resp *resp = &rsp.create_mr_resp;
+       int ret;
+
+       /* Support only LOCAL_WRITE flag for DMA MRs */
+       if (acc & ~IB_ACCESS_LOCAL_WRITE) {
+               dev_warn(&dev->pdev->dev,
+                        "unsupported dma mr access flags %#x\n", acc);
+               return ERR_PTR(-EOPNOTSUPP);
+       }
+
+       mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+       if (!mr)
+               return ERR_PTR(-ENOMEM);
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_CREATE_MR;
+       cmd->pd_handle = to_vpd(pd)->pd_handle;
+       cmd->access_flags = acc;
+       cmd->flags = PVRDMA_MR_FLAG_DMA;
+
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_MR_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not get DMA mem region, error: %d\n", ret);
+               kfree(mr);
+               return ERR_PTR(ret);
+       }
+
+       mr->mmr.mr_handle = resp->mr_handle;
+       mr->ibmr.lkey = resp->lkey;
+       mr->ibmr.rkey = resp->rkey;
+
+       return &mr->ibmr;
+}
+
+/**
+ * pvrdma_reg_user_mr - register a userspace memory region
+ * @pd: protection domain
+ * @start: starting address
+ * @length: length of region
+ * @virt_addr: I/O virtual address
+ * @access_flags: access flags for memory region
+ * @udata: user data
+ *
+ * @return: ib_mr pointer on success, otherwise returns an errno.
+ */
+struct ib_mr *pvrdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+                                u64 virt_addr, int access_flags,
+                                struct ib_udata *udata)
+{
+       struct pvrdma_dev *dev = to_vdev(pd->device);
+       struct pvrdma_user_mr *mr = NULL;
+       struct ib_umem *umem;
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_create_mr *cmd = &req.create_mr;
+       struct pvrdma_cmd_create_mr_resp *resp = &rsp.create_mr_resp;
+       int nchunks;
+       int ret;
+       int entry;
+       struct scatterlist *sg;
+
+       if (length == 0 || length > dev->dsr->caps.max_mr_size) {
+               dev_warn(&dev->pdev->dev, "invalid mem region length\n");
+               return ERR_PTR(-EINVAL);
+       }
+
+       umem = ib_umem_get(pd->uobject->context, start,
+                          length, access_flags, 0);
+       if (IS_ERR(umem)) {
+               dev_warn(&dev->pdev->dev,
+                        "could not get umem for mem region\n");
+               return ERR_CAST(umem);
+       }
+
+       nchunks = 0;
+       for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry)
+               nchunks += sg_dma_len(sg) >> PAGE_SHIFT;
+
+       if (nchunks < 0 || nchunks > PVRDMA_PAGE_DIR_MAX_PAGES) {
+               dev_warn(&dev->pdev->dev, "overflow %d pages in mem region\n",
+                        nchunks);
+               ret = -EINVAL;
+               goto err_umem;
+       }
+
+       mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+       if (!mr) {
+               ret = -ENOMEM;
+               goto err_umem;
+       }
+
+       mr->mmr.iova = virt_addr;
+       mr->mmr.size = length;
+       mr->umem = umem;
+
+       ret = pvrdma_page_dir_init(dev, &mr->pdir, nchunks, false);
+       if (ret) {
+               dev_warn(&dev->pdev->dev,
+                        "could not allocate page directory\n");
+               goto err_umem;
+       }
+
+       ret = pvrdma_page_dir_insert_umem(&mr->pdir, mr->umem, 0);
+       if (ret)
+               goto err_pdir;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_CREATE_MR;
+       cmd->start = start;
+       cmd->length = length;
+       cmd->pd_handle = to_vpd(pd)->pd_handle;
+       cmd->access_flags = access_flags;
+       cmd->nchunks = nchunks;
+       cmd->pdir_dma = mr->pdir.dir_dma;
+
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_MR_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not register mem region, error: %d\n", ret);
+               goto err_pdir;
+       }
+
+       mr->mmr.mr_handle = resp->mr_handle;
+       mr->ibmr.lkey = resp->lkey;
+       mr->ibmr.rkey = resp->rkey;
+
+       return &mr->ibmr;
+
+err_pdir:
+       pvrdma_page_dir_cleanup(dev, &mr->pdir);
+err_umem:
+       ib_umem_release(umem);
+       kfree(mr);
+
+       return ERR_PTR(ret);
+}
+
+/**
+ * pvrdma_alloc_mr - allocate a memory region
+ * @pd: protection domain
+ * @mr_type: type of memory region
+ * @max_num_sg: maximum number of pages
+ *
+ * @return: ib_mr pointer on success, otherwise returns an errno.
+ */
+struct ib_mr *pvrdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
+                             u32 max_num_sg)
+{
+       struct pvrdma_dev *dev = to_vdev(pd->device);
+       struct pvrdma_user_mr *mr;
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_create_mr *cmd = &req.create_mr;
+       struct pvrdma_cmd_create_mr_resp *resp = &rsp.create_mr_resp;
+       int size = max_num_sg * sizeof(u64);
+       int ret;
+
+       if (mr_type != IB_MR_TYPE_MEM_REG ||
+           max_num_sg > PVRDMA_MAX_FAST_REG_PAGES)
+               return ERR_PTR(-EINVAL);
+
+       mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+       if (!mr)
+               return ERR_PTR(-ENOMEM);
+
+       mr->pages = kzalloc(size, GFP_KERNEL);
+       if (!mr->pages) {
+               ret = -ENOMEM;
+               goto freemr;
+       }
+
+       ret = pvrdma_page_dir_init(dev, &mr->pdir, max_num_sg, false);
+       if (ret) {
+               dev_warn(&dev->pdev->dev,
+                        "failed to allocate page dir for mr\n");
+               ret = -ENOMEM;
+               goto freepages;
+       }
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_CREATE_MR;
+       cmd->pd_handle = to_vpd(pd)->pd_handle;
+       cmd->access_flags = 0;
+       cmd->flags = PVRDMA_MR_FLAG_FRMR;
+       cmd->nchunks = max_num_sg;
+
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_MR_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not create FR mem region, error: %d\n", ret);
+               goto freepdir;
+       }
+
+       mr->max_pages = max_num_sg;
+       mr->mmr.mr_handle = resp->mr_handle;
+       mr->ibmr.lkey = resp->lkey;
+       mr->ibmr.rkey = resp->rkey;
+       mr->page_shift = PAGE_SHIFT;
+       mr->umem = NULL;
+
+       return &mr->ibmr;
+
+freepdir:
+       pvrdma_page_dir_cleanup(dev, &mr->pdir);
+freepages:
+       kfree(mr->pages);
+freemr:
+       kfree(mr);
+       return ERR_PTR(ret);
+}
+
+/**
+ * pvrdma_dereg_mr - deregister a memory region
+ * @ibmr: memory region
+ *
+ * @return: 0 on success.
+ */
+int pvrdma_dereg_mr(struct ib_mr *ibmr)
+{
+       struct pvrdma_user_mr *mr = to_vmr(ibmr);
+       struct pvrdma_dev *dev = to_vdev(ibmr->device);
+       union pvrdma_cmd_req req;
+       struct pvrdma_cmd_destroy_mr *cmd = &req.destroy_mr;
+       int ret;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_DESTROY_MR;
+       cmd->mr_handle = mr->mmr.mr_handle;
+       ret = pvrdma_cmd_post(dev, &req, NULL, 0);
+       if (ret < 0)
+               dev_warn(&dev->pdev->dev,
+                        "could not deregister mem region, error: %d\n", ret);
+
+       pvrdma_page_dir_cleanup(dev, &mr->pdir);
+       if (mr->umem)
+               ib_umem_release(mr->umem);
+
+       kfree(mr->pages);
+       kfree(mr);
+
+       return 0;
+}
+
+static int pvrdma_set_page(struct ib_mr *ibmr, u64 addr)
+{
+       struct pvrdma_user_mr *mr = to_vmr(ibmr);
+
+       if (mr->npages == mr->max_pages)
+               return -ENOMEM;
+
+       mr->pages[mr->npages++] = addr;
+       return 0;
+}
+
+int pvrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
+                    unsigned int *sg_offset)
+{
+       struct pvrdma_user_mr *mr = to_vmr(ibmr);
+       struct pvrdma_dev *dev = to_vdev(ibmr->device);
+       int ret;
+
+       mr->npages = 0;
+
+       ret = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, pvrdma_set_page);
+       if (ret < 0)
+               dev_warn(&dev->pdev->dev, "could not map sg to pages\n");
+
+       return ret;
+}
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c
new file mode 100644 (file)
index 0000000..c8c01e5
--- /dev/null
@@ -0,0 +1,972 @@
+/*
+ * Copyright (c) 2012-2016 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"
+
+static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
+                          struct pvrdma_cq **recv_cq)
+{
+       *send_cq = to_vcq(qp->ibqp.send_cq);
+       *recv_cq = to_vcq(qp->ibqp.recv_cq);
+}
+
+static void pvrdma_lock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
+                           unsigned long *scq_flags,
+                           unsigned long *rcq_flags)
+       __acquires(scq->cq_lock) __acquires(rcq->cq_lock)
+{
+       if (scq == rcq) {
+               spin_lock_irqsave(&scq->cq_lock, *scq_flags);
+               __acquire(rcq->cq_lock);
+       } else if (scq->cq_handle < rcq->cq_handle) {
+               spin_lock_irqsave(&scq->cq_lock, *scq_flags);
+               spin_lock_irqsave_nested(&rcq->cq_lock, *rcq_flags,
+                                        SINGLE_DEPTH_NESTING);
+       } else {
+               spin_lock_irqsave(&rcq->cq_lock, *rcq_flags);
+               spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags,
+                                        SINGLE_DEPTH_NESTING);
+       }
+}
+
+static void pvrdma_unlock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
+                             unsigned long *scq_flags,
+                             unsigned long *rcq_flags)
+       __releases(scq->cq_lock) __releases(rcq->cq_lock)
+{
+       if (scq == rcq) {
+               __release(rcq->cq_lock);
+               spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
+       } else if (scq->cq_handle < rcq->cq_handle) {
+               spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
+               spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
+       } else {
+               spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
+               spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
+       }
+}
+
+static void pvrdma_reset_qp(struct pvrdma_qp *qp)
+{
+       struct pvrdma_cq *scq, *rcq;
+       unsigned long scq_flags, rcq_flags;
+
+       /* Clean up cqes */
+       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);
+
+       pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
+
+       /*
+        * Reset queuepair. The checks are because usermode queuepairs won't
+        * have kernel ringstates.
+        */
+       if (qp->rq.ring) {
+               atomic_set(&qp->rq.ring->cons_head, 0);
+               atomic_set(&qp->rq.ring->prod_tail, 0);
+       }
+       if (qp->sq.ring) {
+               atomic_set(&qp->sq.ring->cons_head, 0);
+               atomic_set(&qp->sq.ring->prod_tail, 0);
+       }
+}
+
+static int pvrdma_set_rq_size(struct pvrdma_dev *dev,
+                             struct ib_qp_cap *req_cap,
+                             struct pvrdma_qp *qp)
+{
+       if (req_cap->max_recv_wr > dev->dsr->caps.max_qp_wr ||
+           req_cap->max_recv_sge > dev->dsr->caps.max_sge) {
+               dev_warn(&dev->pdev->dev, "recv queue size invalid\n");
+               return -EINVAL;
+       }
+
+       qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_recv_wr));
+       qp->rq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_recv_sge));
+
+       /* Write back */
+       req_cap->max_recv_wr = qp->rq.wqe_cnt;
+       req_cap->max_recv_sge = qp->rq.max_sg;
+
+       qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) +
+                                            sizeof(struct pvrdma_sge) *
+                                            qp->rq.max_sg);
+       qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) /
+                         PAGE_SIZE;
+
+       return 0;
+}
+
+static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap,
+                             enum ib_qp_type type, struct pvrdma_qp *qp)
+{
+       if (req_cap->max_send_wr > dev->dsr->caps.max_qp_wr ||
+           req_cap->max_send_sge > dev->dsr->caps.max_sge) {
+               dev_warn(&dev->pdev->dev, "send queue size invalid\n");
+               return -EINVAL;
+       }
+
+       qp->sq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_send_wr));
+       qp->sq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_send_sge));
+
+       /* Write back */
+       req_cap->max_send_wr = qp->sq.wqe_cnt;
+       req_cap->max_send_sge = qp->sq.max_sg;
+
+       qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) +
+                                            sizeof(struct pvrdma_sge) *
+                                            qp->sq.max_sg);
+       /* Note: one extra page for the header. */
+       qp->npages_send = 1 + (qp->sq.wqe_cnt * qp->sq.wqe_size +
+                              PAGE_SIZE - 1) / PAGE_SIZE;
+
+       return 0;
+}
+
+/**
+ * pvrdma_create_qp - create queue pair
+ * @pd: protection domain
+ * @init_attr: queue pair attributes
+ * @udata: user data
+ *
+ * @return: the ib_qp pointer on success, otherwise returns an errno.
+ */
+struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
+                              struct ib_qp_init_attr *init_attr,
+                              struct ib_udata *udata)
+{
+       struct pvrdma_qp *qp = NULL;
+       struct pvrdma_dev *dev = to_vdev(pd->device);
+       union pvrdma_cmd_req req;
+       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_create_qp ucmd;
+       unsigned long flags;
+       int ret;
+
+       if (init_attr->create_flags) {
+               dev_warn(&dev->pdev->dev,
+                        "invalid create queuepair flags %#x\n",
+                        init_attr->create_flags);
+               return ERR_PTR(-EINVAL);
+       }
+
+       if (init_attr->qp_type != IB_QPT_RC &&
+           init_attr->qp_type != IB_QPT_UD &&
+           init_attr->qp_type != IB_QPT_GSI) {
+               dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
+                        init_attr->qp_type);
+               return ERR_PTR(-EINVAL);
+       }
+
+       if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
+               return ERR_PTR(-ENOMEM);
+
+       switch (init_attr->qp_type) {
+       case IB_QPT_GSI:
+               if (init_attr->port_num == 0 ||
+                   init_attr->port_num > pd->device->phys_port_cnt ||
+                   udata) {
+                       dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
+                       ret = -EINVAL;
+                       goto err_qp;
+               }
+               /* fall through */
+       case IB_QPT_RC:
+       case IB_QPT_UD:
+               qp = kzalloc(sizeof(*qp), GFP_KERNEL);
+               if (!qp) {
+                       ret = -ENOMEM;
+                       goto err_qp;
+               }
+
+               spin_lock_init(&qp->sq.lock);
+               spin_lock_init(&qp->rq.lock);
+               mutex_init(&qp->mutex);
+               atomic_set(&qp->refcnt, 1);
+               init_waitqueue_head(&qp->wait);
+
+               qp->state = IB_QPS_RESET;
+
+               if (pd->uobject && udata) {
+                       dev_dbg(&dev->pdev->dev,
+                               "create queuepair from user space\n");
+
+                       if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
+                               ret = -EFAULT;
+                               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;
+                       }
+
+                       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);
+                               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);
+                       qp->npages = qp->npages_send + qp->npages_recv;
+               } else {
+                       qp->is_kernel = true;
+
+                       ret = pvrdma_set_sq_size(to_vdev(pd->device),
+                                                &init_attr->cap,
+                                                init_attr->qp_type, qp);
+                       if (ret)
+                               goto err_qp;
+
+                       ret = pvrdma_set_rq_size(to_vdev(pd->device),
+                                                &init_attr->cap, qp);
+                       if (ret)
+                               goto err_qp;
+
+                       qp->npages = qp->npages_send + qp->npages_recv;
+
+                       /* Skip header page. */
+                       qp->sq.offset = PAGE_SIZE;
+
+                       /* Recv queue pages are after send pages. */
+                       qp->rq.offset = qp->npages_send * PAGE_SIZE;
+               }
+
+               if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
+                       dev_warn(&dev->pdev->dev,
+                                "overflow pages in queuepair\n");
+                       ret = -EINVAL;
+                       goto err_umem;
+               }
+
+               ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
+                                          qp->is_kernel);
+               if (ret) {
+                       dev_warn(&dev->pdev->dev,
+                                "could not allocate page directory\n");
+                       goto err_umem;
+               }
+
+               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);
+               } else {
+                       /* Ring state is always the first page. */
+                       qp->sq.ring = qp->pdir.pages[0];
+                       qp->rq.ring = &qp->sq.ring[1];
+               }
+               break;
+       default:
+               ret = -EINVAL;
+               goto err_qp;
+       }
+
+       /* Not supported */
+       init_attr->cap.max_inline_data = 0;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
+       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;
+       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;
+       cmd->max_recv_sge = init_attr->cap.max_recv_sge;
+       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->access_flags = IB_ACCESS_LOCAL_WRITE;
+       cmd->total_chunks = qp->npages;
+       cmd->send_chunks = qp->npages_send - 1;
+       cmd->pdir_dma = qp->pdir.dir_dma;
+
+       dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
+               cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
+               cmd->max_recv_sge);
+
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not create queuepair, error: %d\n", ret);
+               goto err_pdir;
+       }
+
+       /* 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;
+       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);
+
+       return &qp->ibqp;
+
+err_pdir:
+       pvrdma_page_dir_cleanup(dev, &qp->pdir);
+err_umem:
+       if (pd->uobject && udata) {
+               if (qp->rumem)
+                       ib_umem_release(qp->rumem);
+               if (qp->sumem)
+                       ib_umem_release(qp->sumem);
+       }
+err_qp:
+       kfree(qp);
+       atomic_dec(&dev->num_qps);
+
+       return ERR_PTR(ret);
+}
+
+static void pvrdma_free_qp(struct pvrdma_qp *qp)
+{
+       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);
+
+       atomic_dec(&qp->refcnt);
+       wait_event(qp->wait, !atomic_read(&qp->refcnt));
+
+       pvrdma_page_dir_cleanup(dev, &qp->pdir);
+
+       kfree(qp);
+
+       atomic_dec(&dev->num_qps);
+}
+
+/**
+ * pvrdma_destroy_qp - destroy a queue pair
+ * @qp: the queue pair to destroy
+ *
+ * @return: 0 on success.
+ */
+int pvrdma_destroy_qp(struct ib_qp *qp)
+{
+       struct pvrdma_qp *vqp = to_vqp(qp);
+       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;
+
+       ret = pvrdma_cmd_post(to_vdev(qp->device), &req, NULL, 0);
+       if (ret < 0)
+               dev_warn(&to_vdev(qp->device)->pdev->dev,
+                        "destroy queuepair failed, error: %d\n", ret);
+
+       pvrdma_free_qp(vqp);
+
+       return 0;
+}
+
+/**
+ * pvrdma_modify_qp - modify queue pair attributes
+ * @ibqp: the queue pair
+ * @attr: the new queue pair's attributes
+ * @attr_mask: attributes mask
+ * @udata: user data
+ *
+ * @returns 0 on success, otherwise returns an errno.
+ */
+int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+                    int attr_mask, struct ib_udata *udata)
+{
+       struct pvrdma_dev *dev = to_vdev(ibqp->device);
+       struct pvrdma_qp *qp = to_vqp(ibqp);
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
+       int cur_state, next_state;
+       int ret;
+
+       /* Sanity checking. Should need lock here */
+       mutex_lock(&qp->mutex);
+       cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
+               qp->state;
+       next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
+
+       if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
+                               attr_mask, IB_LINK_LAYER_ETHERNET)) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       if (attr_mask & IB_QP_PORT) {
+               if (attr->port_num == 0 ||
+                   attr->port_num > ibqp->device->phys_port_cnt) {
+                       ret = -EINVAL;
+                       goto out;
+               }
+       }
+
+       if (attr_mask & IB_QP_MIN_RNR_TIMER) {
+               if (attr->min_rnr_timer > 31) {
+                       ret = -EINVAL;
+                       goto out;
+               }
+       }
+
+       if (attr_mask & IB_QP_PKEY_INDEX) {
+               if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
+                       ret = -EINVAL;
+                       goto out;
+               }
+       }
+
+       if (attr_mask & IB_QP_QKEY)
+               qp->qkey = attr->qkey;
+
+       if (cur_state == next_state && cur_state == IB_QPS_RESET) {
+               ret = 0;
+               goto out;
+       }
+
+       qp->state = next_state;
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
+       cmd->qp_handle = qp->qp_handle;
+       cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
+       cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
+       cmd->attrs.cur_qp_state =
+               ib_qp_state_to_pvrdma(attr->cur_qp_state);
+       cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
+       cmd->attrs.path_mig_state =
+               ib_mig_state_to_pvrdma(attr->path_mig_state);
+       cmd->attrs.qkey = attr->qkey;
+       cmd->attrs.rq_psn = attr->rq_psn;
+       cmd->attrs.sq_psn = attr->sq_psn;
+       cmd->attrs.dest_qp_num = attr->dest_qp_num;
+       cmd->attrs.qp_access_flags =
+               ib_access_flags_to_pvrdma(attr->qp_access_flags);
+       cmd->attrs.pkey_index = attr->pkey_index;
+       cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
+       cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
+       cmd->attrs.sq_draining = attr->sq_draining;
+       cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
+       cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
+       cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
+       cmd->attrs.port_num = attr->port_num;
+       cmd->attrs.timeout = attr->timeout;
+       cmd->attrs.retry_cnt = attr->retry_cnt;
+       cmd->attrs.rnr_retry = attr->rnr_retry;
+       cmd->attrs.alt_port_num = attr->alt_port_num;
+       cmd->attrs.alt_timeout = attr->alt_timeout;
+       ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
+       ib_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
+       ib_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
+
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not modify queuepair, error: %d\n", ret);
+       } else if (rsp.hdr.err > 0) {
+               dev_warn(&dev->pdev->dev,
+                        "cannot modify queuepair, error: %d\n", rsp.hdr.err);
+               ret = -EINVAL;
+       }
+
+       if (ret == 0 && next_state == IB_QPS_RESET)
+               pvrdma_reset_qp(qp);
+
+out:
+       mutex_unlock(&qp->mutex);
+
+       return ret;
+}
+
+static inline void *get_sq_wqe(struct pvrdma_qp *qp, int n)
+{
+       return pvrdma_page_dir_get_ptr(&qp->pdir,
+                                      qp->sq.offset + n * qp->sq.wqe_size);
+}
+
+static inline void *get_rq_wqe(struct pvrdma_qp *qp, int n)
+{
+       return pvrdma_page_dir_get_ptr(&qp->pdir,
+                                      qp->rq.offset + n * qp->rq.wqe_size);
+}
+
+static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr, struct ib_reg_wr *wr)
+{
+       struct pvrdma_user_mr *mr = to_vmr(wr->mr);
+
+       wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
+       wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
+       wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
+       wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
+       wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
+       wqe_hdr->wr.fast_reg.access_flags = wr->access;
+       wqe_hdr->wr.fast_reg.rkey = wr->key;
+
+       return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
+                                               mr->npages);
+}
+
+/**
+ * pvrdma_post_send - post send work request entries on a QP
+ * @ibqp: the QP
+ * @wr: work request list to post
+ * @bad_wr: the first bad WR returned
+ *
+ * @return: 0 on success, otherwise errno returned.
+ */
+int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+                    struct ib_send_wr **bad_wr)
+{
+       struct pvrdma_qp *qp = to_vqp(ibqp);
+       struct pvrdma_dev *dev = to_vdev(ibqp->device);
+       unsigned long flags;
+       struct pvrdma_sq_wqe_hdr *wqe_hdr;
+       struct pvrdma_sge *sge;
+       int i, index;
+       int nreq;
+       int ret;
+
+       /*
+        * In states lower than RTS, we can fail immediately. In other states,
+        * just post and let the device figure it out.
+        */
+       if (qp->state < IB_QPS_RTS) {
+               *bad_wr = wr;
+               return -EINVAL;
+       }
+
+       spin_lock_irqsave(&qp->sq.lock, flags);
+
+       index = pvrdma_idx(&qp->sq.ring->prod_tail, qp->sq.wqe_cnt);
+       for (nreq = 0; wr; nreq++, wr = wr->next) {
+               unsigned int tail;
+
+               if (unlikely(!pvrdma_idx_ring_has_space(
+                               qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
+                       dev_warn_ratelimited(&dev->pdev->dev,
+                                            "send queue is full\n");
+                       *bad_wr = wr;
+                       ret = -ENOMEM;
+                       goto out;
+               }
+
+               if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
+                       dev_warn_ratelimited(&dev->pdev->dev,
+                                            "send SGE overflow\n");
+                       *bad_wr = wr;
+                       ret = -EINVAL;
+                       goto out;
+               }
+
+               if (unlikely(wr->opcode < 0)) {
+                       dev_warn_ratelimited(&dev->pdev->dev,
+                                            "invalid send opcode\n");
+                       *bad_wr = wr;
+                       ret = -EINVAL;
+                       goto out;
+               }
+
+               /*
+                * Only support UD, RC.
+                * Need to check opcode table for thorough checking.
+                * opcode               _UD     _UC     _RC
+                * _SEND                x       x       x
+                * _SEND_WITH_IMM       x       x       x
+                * _RDMA_WRITE                  x       x
+                * _RDMA_WRITE_WITH_IMM         x       x
+                * _LOCAL_INV                   x       x
+                * _SEND_WITH_INV               x       x
+                * _RDMA_READ                           x
+                * _ATOMIC_CMP_AND_SWP                  x
+                * _ATOMIC_FETCH_AND_ADD                x
+                * _MASK_ATOMIC_CMP_AND_SWP             x
+                * _MASK_ATOMIC_FETCH_AND_ADD           x
+                * _REG_MR                              x
+                *
+                */
+               if (qp->ibqp.qp_type != IB_QPT_UD &&
+                   qp->ibqp.qp_type != IB_QPT_RC &&
+                       wr->opcode != IB_WR_SEND) {
+                       dev_warn_ratelimited(&dev->pdev->dev,
+                                            "unsupported queuepair type\n");
+                       *bad_wr = wr;
+                       ret = -EINVAL;
+                       goto out;
+               } else if (qp->ibqp.qp_type == IB_QPT_UD ||
+                          qp->ibqp.qp_type == IB_QPT_GSI) {
+                       if (wr->opcode != IB_WR_SEND &&
+                           wr->opcode != IB_WR_SEND_WITH_IMM) {
+                               dev_warn_ratelimited(&dev->pdev->dev,
+                                                    "invalid send opcode\n");
+                               *bad_wr = wr;
+                               ret = -EINVAL;
+                               goto out;
+                       }
+               }
+
+               wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, index);
+               memset(wqe_hdr, 0, sizeof(*wqe_hdr));
+               wqe_hdr->wr_id = wr->wr_id;
+               wqe_hdr->num_sge = wr->num_sge;
+               wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
+               wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
+               if (wr->opcode == IB_WR_SEND_WITH_IMM ||
+                   wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
+                       wqe_hdr->ex.imm_data = wr->ex.imm_data;
+
+               switch (qp->ibqp.qp_type) {
+               case IB_QPT_GSI:
+               case IB_QPT_UD:
+                       if (unlikely(!ud_wr(wr)->ah)) {
+                               dev_warn_ratelimited(&dev->pdev->dev,
+                                                    "invalid address handle\n");
+                               *bad_wr = wr;
+                               ret = -EINVAL;
+                               goto out;
+                       }
+
+                       /*
+                        * Use qkey from qp context if high order bit set,
+                        * otherwise from work request.
+                        */
+                       wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
+                       wqe_hdr->wr.ud.remote_qkey =
+                               ud_wr(wr)->remote_qkey & 0x80000000 ?
+                               qp->qkey : ud_wr(wr)->remote_qkey;
+                       wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
+
+                       break;
+               case IB_QPT_RC:
+                       switch (wr->opcode) {
+                       case IB_WR_RDMA_READ:
+                       case IB_WR_RDMA_WRITE:
+                       case IB_WR_RDMA_WRITE_WITH_IMM:
+                               wqe_hdr->wr.rdma.remote_addr =
+                                       rdma_wr(wr)->remote_addr;
+                               wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
+                               break;
+                       case IB_WR_LOCAL_INV:
+                       case IB_WR_SEND_WITH_INV:
+                               wqe_hdr->ex.invalidate_rkey =
+                                       wr->ex.invalidate_rkey;
+                               break;
+                       case IB_WR_ATOMIC_CMP_AND_SWP:
+                       case IB_WR_ATOMIC_FETCH_AND_ADD:
+                               wqe_hdr->wr.atomic.remote_addr =
+                                       atomic_wr(wr)->remote_addr;
+                               wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
+                               wqe_hdr->wr.atomic.compare_add =
+                                       atomic_wr(wr)->compare_add;
+                               if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
+                                       wqe_hdr->wr.atomic.swap =
+                                               atomic_wr(wr)->swap;
+                               break;
+                       case IB_WR_REG_MR:
+                               ret = set_reg_seg(wqe_hdr, reg_wr(wr));
+                               if (ret < 0) {
+                                       dev_warn_ratelimited(&dev->pdev->dev,
+                                                            "Failed to set fast register work request\n");
+                                       *bad_wr = wr;
+                                       goto out;
+                               }
+                               break;
+                       default:
+                               break;
+                       }
+
+                       break;
+               default:
+                       dev_warn_ratelimited(&dev->pdev->dev,
+                                            "invalid queuepair type\n");
+                       ret = -EINVAL;
+                       *bad_wr = wr;
+                       goto out;
+               }
+
+               sge = (struct pvrdma_sge *)(wqe_hdr + 1);
+               for (i = 0; i < wr->num_sge; i++) {
+                       /* Need to check wqe_size 0 or max size */
+                       sge->addr = wr->sg_list[i].addr;
+                       sge->length = wr->sg_list[i].length;
+                       sge->lkey = wr->sg_list[i].lkey;
+                       sge++;
+               }
+
+               /* Make sure wqe is written before index update */
+               smp_wmb();
+
+               index++;
+               if (unlikely(index >= qp->sq.wqe_cnt))
+                       index = 0;
+               /* Update shared sq ring */
+               pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
+                                   qp->sq.wqe_cnt);
+       }
+
+       ret = 0;
+
+out:
+       spin_unlock_irqrestore(&qp->sq.lock, flags);
+
+       if (!ret)
+               pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
+
+       return ret;
+}
+
+/**
+ * pvrdma_post_receive - post receive work request entries on a QP
+ * @ibqp: the QP
+ * @wr: the work request list to post
+ * @bad_wr: the first bad WR returned
+ *
+ * @return: 0 on success, otherwise errno returned.
+ */
+int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
+                    struct ib_recv_wr **bad_wr)
+{
+       struct pvrdma_dev *dev = to_vdev(ibqp->device);
+       unsigned long flags;
+       struct pvrdma_qp *qp = to_vqp(ibqp);
+       struct pvrdma_rq_wqe_hdr *wqe_hdr;
+       struct pvrdma_sge *sge;
+       int index, nreq;
+       int ret = 0;
+       int i;
+
+       /*
+        * In the RESET state, we can fail immediately. For other states,
+        * just post and let the device figure it out.
+        */
+       if (qp->state == IB_QPS_RESET) {
+               *bad_wr = wr;
+               return -EINVAL;
+       }
+
+       spin_lock_irqsave(&qp->rq.lock, flags);
+
+       index = pvrdma_idx(&qp->rq.ring->prod_tail, qp->rq.wqe_cnt);
+       for (nreq = 0; wr; nreq++, wr = wr->next) {
+               unsigned int tail;
+
+               if (unlikely(wr->num_sge > qp->rq.max_sg ||
+                            wr->num_sge < 0)) {
+                       ret = -EINVAL;
+                       *bad_wr = wr;
+                       dev_warn_ratelimited(&dev->pdev->dev,
+                                            "recv SGE overflow\n");
+                       goto out;
+               }
+
+               if (unlikely(!pvrdma_idx_ring_has_space(
+                               qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
+                       ret = -ENOMEM;
+                       *bad_wr = wr;
+                       dev_warn_ratelimited(&dev->pdev->dev,
+                                            "recv queue full\n");
+                       goto out;
+               }
+
+               wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, index);
+               wqe_hdr->wr_id = wr->wr_id;
+               wqe_hdr->num_sge = wr->num_sge;
+               wqe_hdr->total_len = 0;
+
+               sge = (struct pvrdma_sge *)(wqe_hdr + 1);
+               for (i = 0; i < wr->num_sge; i++) {
+                       sge->addr = wr->sg_list[i].addr;
+                       sge->length = wr->sg_list[i].length;
+                       sge->lkey = wr->sg_list[i].lkey;
+                       sge++;
+               }
+
+               /* Make sure wqe is written before index update */
+               smp_wmb();
+
+               index++;
+               if (unlikely(index >= qp->rq.wqe_cnt))
+                       index = 0;
+               /* Update shared rq ring */
+               pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
+                                   qp->rq.wqe_cnt);
+       }
+
+       spin_unlock_irqrestore(&qp->rq.lock, flags);
+
+       pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
+
+       return ret;
+
+out:
+       spin_unlock_irqrestore(&qp->rq.lock, flags);
+
+       return ret;
+}
+
+/**
+ * pvrdma_query_qp - query a queue pair's attributes
+ * @ibqp: the queue pair to query
+ * @attr: the queue pair's attributes
+ * @attr_mask: attributes mask
+ * @init_attr: initial queue pair attributes
+ *
+ * @returns 0 on success, otherwise returns an errno.
+ */
+int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+                   int attr_mask, struct ib_qp_init_attr *init_attr)
+{
+       struct pvrdma_dev *dev = to_vdev(ibqp->device);
+       struct pvrdma_qp *qp = to_vqp(ibqp);
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
+       struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
+       int ret = 0;
+
+       mutex_lock(&qp->mutex);
+
+       if (qp->state == IB_QPS_RESET) {
+               attr->qp_state = IB_QPS_RESET;
+               goto out;
+       }
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
+       cmd->qp_handle = qp->qp_handle;
+       cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
+
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not query queuepair, error: %d\n", ret);
+               goto out;
+       }
+
+       attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
+       attr->cur_qp_state =
+               pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
+       attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
+       attr->path_mig_state =
+               pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
+       attr->qkey = resp->attrs.qkey;
+       attr->rq_psn = resp->attrs.rq_psn;
+       attr->sq_psn = resp->attrs.sq_psn;
+       attr->dest_qp_num = resp->attrs.dest_qp_num;
+       attr->qp_access_flags =
+               pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
+       attr->pkey_index = resp->attrs.pkey_index;
+       attr->alt_pkey_index = resp->attrs.alt_pkey_index;
+       attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
+       attr->sq_draining = resp->attrs.sq_draining;
+       attr->max_rd_atomic = resp->attrs.max_rd_atomic;
+       attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
+       attr->min_rnr_timer = resp->attrs.min_rnr_timer;
+       attr->port_num = resp->attrs.port_num;
+       attr->timeout = resp->attrs.timeout;
+       attr->retry_cnt = resp->attrs.retry_cnt;
+       attr->rnr_retry = resp->attrs.rnr_retry;
+       attr->alt_port_num = resp->attrs.alt_port_num;
+       attr->alt_timeout = resp->attrs.alt_timeout;
+       pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
+       pvrdma_ah_attr_to_ib(&attr->ah_attr, &resp->attrs.ah_attr);
+       pvrdma_ah_attr_to_ib(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
+
+       qp->state = attr->qp_state;
+
+       ret = 0;
+
+out:
+       attr->cur_qp_state = attr->qp_state;
+
+       init_attr->event_handler = qp->ibqp.event_handler;
+       init_attr->qp_context = qp->ibqp.qp_context;
+       init_attr->send_cq = qp->ibqp.send_cq;
+       init_attr->recv_cq = qp->ibqp.recv_cq;
+       init_attr->srq = qp->ibqp.srq;
+       init_attr->xrcd = NULL;
+       init_attr->cap = attr->cap;
+       init_attr->sq_sig_type = 0;
+       init_attr->qp_type = qp->ibqp.qp_type;
+       init_attr->create_flags = 0;
+       init_attr->port_num = qp->port;
+
+       mutex_unlock(&qp->mutex);
+       return ret;
+}
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_ring.h b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_ring.h
new file mode 100644 (file)
index 0000000..ed9022a
--- /dev/null
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2012-2016 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.
+ */
+
+#ifndef __PVRDMA_RING_H__
+#define __PVRDMA_RING_H__
+
+#include <linux/types.h>
+
+#define PVRDMA_INVALID_IDX     -1      /* Invalid index. */
+
+struct pvrdma_ring {
+       atomic_t prod_tail;     /* Producer tail. */
+       atomic_t cons_head;     /* Consumer head. */
+};
+
+struct pvrdma_ring_state {
+       struct pvrdma_ring tx;  /* Tx ring. */
+       struct pvrdma_ring rx;  /* Rx ring. */
+};
+
+static inline int pvrdma_idx_valid(__u32 idx, __u32 max_elems)
+{
+       /* Generates fewer instructions than a less-than. */
+       return (idx & ~((max_elems << 1) - 1)) == 0;
+}
+
+static inline __s32 pvrdma_idx(atomic_t *var, __u32 max_elems)
+{
+       const unsigned int idx = atomic_read(var);
+
+       if (pvrdma_idx_valid(idx, max_elems))
+               return idx & (max_elems - 1);
+       return PVRDMA_INVALID_IDX;
+}
+
+static inline void pvrdma_idx_ring_inc(atomic_t *var, __u32 max_elems)
+{
+       __u32 idx = atomic_read(var) + 1;       /* Increment. */
+
+       idx &= (max_elems << 1) - 1;            /* Modulo size, flip gen. */
+       atomic_set(var, idx);
+}
+
+static inline __s32 pvrdma_idx_ring_has_space(const struct pvrdma_ring *r,
+                                             __u32 max_elems, __u32 *out_tail)
+{
+       const __u32 tail = atomic_read(&r->prod_tail);
+       const __u32 head = atomic_read(&r->cons_head);
+
+       if (pvrdma_idx_valid(tail, max_elems) &&
+           pvrdma_idx_valid(head, max_elems)) {
+               *out_tail = tail & (max_elems - 1);
+               return tail != (head ^ max_elems);
+       }
+       return PVRDMA_INVALID_IDX;
+}
+
+static inline __s32 pvrdma_idx_ring_has_data(const struct pvrdma_ring *r,
+                                            __u32 max_elems, __u32 *out_head)
+{
+       const __u32 tail = atomic_read(&r->prod_tail);
+       const __u32 head = atomic_read(&r->cons_head);
+
+       if (pvrdma_idx_valid(tail, max_elems) &&
+           pvrdma_idx_valid(head, max_elems)) {
+               *out_head = head & (max_elems - 1);
+               return tail != head;
+       }
+       return PVRDMA_INVALID_IDX;
+}
+
+static inline bool pvrdma_idx_ring_is_valid_idx(const struct pvrdma_ring *r,
+                                               __u32 max_elems, __u32 *idx)
+{
+       const __u32 tail = atomic_read(&r->prod_tail);
+       const __u32 head = atomic_read(&r->cons_head);
+
+       if (pvrdma_idx_valid(tail, max_elems) &&
+           pvrdma_idx_valid(head, max_elems) &&
+           pvrdma_idx_valid(*idx, max_elems)) {
+               if (tail > head && (*idx < tail && *idx >= head))
+                       return true;
+               else if (head > tail && (*idx >= head || *idx < tail))
+                       return true;
+       }
+       return false;
+}
+
+#endif /* __PVRDMA_RING_H__ */
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.c b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.c
new file mode 100644 (file)
index 0000000..5489137
--- /dev/null
@@ -0,0 +1,579 @@
+/*
+ * Copyright (c) 2012-2016 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/inet.h>
+#include <linux/io.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_user_verbs.h>
+#include <rdma/vmw_pvrdma-abi.h>
+
+#include "pvrdma.h"
+
+/**
+ * pvrdma_query_device - query device
+ * @ibdev: the device to query
+ * @props: the device properties
+ * @uhw: user data
+ *
+ * @return: 0 on success, otherwise negative errno
+ */
+int pvrdma_query_device(struct ib_device *ibdev,
+                       struct ib_device_attr *props,
+                       struct ib_udata *uhw)
+{
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+
+       if (uhw->inlen || uhw->outlen)
+               return -EINVAL;
+
+       memset(props, 0, sizeof(*props));
+
+       props->fw_ver = dev->dsr->caps.fw_ver;
+       props->sys_image_guid = dev->dsr->caps.sys_image_guid;
+       props->max_mr_size = dev->dsr->caps.max_mr_size;
+       props->page_size_cap = dev->dsr->caps.page_size_cap;
+       props->vendor_id = dev->dsr->caps.vendor_id;
+       props->vendor_part_id = dev->pdev->device;
+       props->hw_ver = dev->dsr->caps.hw_ver;
+       props->max_qp = dev->dsr->caps.max_qp;
+       props->max_qp_wr = dev->dsr->caps.max_qp_wr;
+       props->device_cap_flags = dev->dsr->caps.device_cap_flags;
+       props->max_sge = dev->dsr->caps.max_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;
+       props->max_pd = dev->dsr->caps.max_pd;
+       props->max_qp_rd_atom = dev->dsr->caps.max_qp_rd_atom;
+       props->max_qp_init_rd_atom = dev->dsr->caps.max_qp_init_rd_atom;
+       props->atomic_cap =
+               dev->dsr->caps.atomic_ops &
+               (PVRDMA_ATOMIC_OP_COMP_SWAP | PVRDMA_ATOMIC_OP_FETCH_ADD) ?
+               IB_ATOMIC_HCA : IB_ATOMIC_NONE;
+       props->masked_atomic_cap = props->atomic_cap;
+       props->max_ah = dev->dsr->caps.max_ah;
+       props->max_pkeys = dev->dsr->caps.max_pkeys;
+       props->local_ca_ack_delay = dev->dsr->caps.local_ca_ack_delay;
+       if ((dev->dsr->caps.bmme_flags & PVRDMA_BMME_FLAG_LOCAL_INV) &&
+           (dev->dsr->caps.bmme_flags & PVRDMA_BMME_FLAG_REMOTE_INV) &&
+           (dev->dsr->caps.bmme_flags & PVRDMA_BMME_FLAG_FAST_REG_WR)) {
+               props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
+       }
+
+       return 0;
+}
+
+/**
+ * pvrdma_query_port - query device port attributes
+ * @ibdev: the device to query
+ * @port: the port number
+ * @props: the device properties
+ *
+ * @return: 0 on success, otherwise negative errno
+ */
+int pvrdma_query_port(struct ib_device *ibdev, u8 port,
+                     struct ib_port_attr *props)
+{
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_query_port *cmd = &req.query_port;
+       struct pvrdma_cmd_query_port_resp *resp = &rsp.query_port_resp;
+       int err;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_QUERY_PORT;
+       cmd->port_num = port;
+
+       err = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_PORT_RESP);
+       if (err < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "could not query port, error: %d\n", err);
+               return err;
+       }
+
+       memset(props, 0, sizeof(*props));
+
+       props->state = pvrdma_port_state_to_ib(resp->attrs.state);
+       props->max_mtu = pvrdma_mtu_to_ib(resp->attrs.max_mtu);
+       props->active_mtu = pvrdma_mtu_to_ib(resp->attrs.active_mtu);
+       props->gid_tbl_len = resp->attrs.gid_tbl_len;
+       props->port_cap_flags =
+               pvrdma_port_cap_flags_to_ib(resp->attrs.port_cap_flags);
+       props->max_msg_sz = resp->attrs.max_msg_sz;
+       props->bad_pkey_cntr = resp->attrs.bad_pkey_cntr;
+       props->qkey_viol_cntr = resp->attrs.qkey_viol_cntr;
+       props->pkey_tbl_len = resp->attrs.pkey_tbl_len;
+       props->lid = resp->attrs.lid;
+       props->sm_lid = resp->attrs.sm_lid;
+       props->lmc = resp->attrs.lmc;
+       props->max_vl_num = resp->attrs.max_vl_num;
+       props->sm_sl = resp->attrs.sm_sl;
+       props->subnet_timeout = resp->attrs.subnet_timeout;
+       props->init_type_reply = resp->attrs.init_type_reply;
+       props->active_width = pvrdma_port_width_to_ib(resp->attrs.active_width);
+       props->active_speed = pvrdma_port_speed_to_ib(resp->attrs.active_speed);
+       props->phys_state = resp->attrs.phys_state;
+
+       return 0;
+}
+
+/**
+ * pvrdma_query_gid - query device gid
+ * @ibdev: the device to query
+ * @port: the port number
+ * @index: the index
+ * @gid: the device gid value
+ *
+ * @return: 0 on success, otherwise negative errno
+ */
+int pvrdma_query_gid(struct ib_device *ibdev, u8 port, int index,
+                    union ib_gid *gid)
+{
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+
+       if (index >= dev->dsr->caps.gid_tbl_len)
+               return -EINVAL;
+
+       memcpy(gid, &dev->sgid_tbl[index], sizeof(union ib_gid));
+
+       return 0;
+}
+
+/**
+ * pvrdma_query_pkey - query device port's P_Key table
+ * @ibdev: the device to query
+ * @port: the port number
+ * @index: the index
+ * @pkey: the device P_Key value
+ *
+ * @return: 0 on success, otherwise negative errno
+ */
+int pvrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
+                     u16 *pkey)
+{
+       int err = 0;
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_query_pkey *cmd = &req.query_pkey;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_QUERY_PKEY;
+       cmd->port_num = port;
+       cmd->index = index;
+
+       err = pvrdma_cmd_post(to_vdev(ibdev), &req, &rsp,
+                             PVRDMA_CMD_QUERY_PKEY_RESP);
+       if (err < 0) {
+               dev_warn(&to_vdev(ibdev)->pdev->dev,
+                        "could not query pkey, error: %d\n", err);
+               return err;
+       }
+
+       *pkey = rsp.query_pkey_resp.pkey;
+
+       return 0;
+}
+
+enum rdma_link_layer pvrdma_port_link_layer(struct ib_device *ibdev,
+                                           u8 port)
+{
+       return IB_LINK_LAYER_ETHERNET;
+}
+
+int pvrdma_modify_device(struct ib_device *ibdev, int mask,
+                        struct ib_device_modify *props)
+{
+       unsigned long flags;
+
+       if (mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
+                    IB_DEVICE_MODIFY_NODE_DESC)) {
+               dev_warn(&to_vdev(ibdev)->pdev->dev,
+                        "unsupported device modify mask %#x\n", mask);
+               return -EOPNOTSUPP;
+       }
+
+       if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
+               spin_lock_irqsave(&to_vdev(ibdev)->desc_lock, flags);
+               memcpy(ibdev->node_desc, props->node_desc, 64);
+               spin_unlock_irqrestore(&to_vdev(ibdev)->desc_lock, flags);
+       }
+
+       if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
+               mutex_lock(&to_vdev(ibdev)->port_mutex);
+               to_vdev(ibdev)->sys_image_guid =
+                       cpu_to_be64(props->sys_image_guid);
+               mutex_unlock(&to_vdev(ibdev)->port_mutex);
+       }
+
+       return 0;
+}
+
+/**
+ * pvrdma_modify_port - modify device port attributes
+ * @ibdev: the device to modify
+ * @port: the port number
+ * @mask: attributes to modify
+ * @props: the device properties
+ *
+ * @return: 0 on success, otherwise negative errno
+ */
+int pvrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
+                      struct ib_port_modify *props)
+{
+       struct ib_port_attr attr;
+       struct pvrdma_dev *vdev = to_vdev(ibdev);
+       int ret;
+
+       if (mask & ~IB_PORT_SHUTDOWN) {
+               dev_warn(&vdev->pdev->dev,
+                        "unsupported port modify mask %#x\n", mask);
+               return -EOPNOTSUPP;
+       }
+
+       mutex_lock(&vdev->port_mutex);
+       ret = pvrdma_query_port(ibdev, port, &attr);
+       if (ret)
+               goto out;
+
+       vdev->port_cap_mask |= props->set_port_cap_mask;
+       vdev->port_cap_mask &= ~props->clr_port_cap_mask;
+
+       if (mask & IB_PORT_SHUTDOWN)
+               vdev->ib_active = false;
+
+out:
+       mutex_unlock(&vdev->port_mutex);
+       return ret;
+}
+
+/**
+ * pvrdma_alloc_ucontext - allocate ucontext
+ * @ibdev: the IB device
+ * @udata: user data
+ *
+ * @return: the ib_ucontext pointer on success, otherwise errno.
+ */
+struct ib_ucontext *pvrdma_alloc_ucontext(struct ib_device *ibdev,
+                                         struct ib_udata *udata)
+{
+       struct pvrdma_dev *vdev = to_vdev(ibdev);
+       struct pvrdma_ucontext *context;
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_create_uc *cmd = &req.create_uc;
+       struct pvrdma_cmd_create_uc_resp *resp = &rsp.create_uc_resp;
+       struct pvrdma_alloc_ucontext_resp uresp;
+       int ret;
+       void *ptr;
+
+       if (!vdev->ib_active)
+               return ERR_PTR(-EAGAIN);
+
+       context = kmalloc(sizeof(*context), GFP_KERNEL);
+       if (!context)
+               return ERR_PTR(-ENOMEM);
+
+       context->dev = vdev;
+       ret = pvrdma_uar_alloc(vdev, &context->uar);
+       if (ret) {
+               kfree(context);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       /* get ctx_handle from host */
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->pfn = context->uar.pfn;
+       cmd->hdr.cmd = PVRDMA_CMD_CREATE_UC;
+       ret = pvrdma_cmd_post(vdev, &req, &rsp, PVRDMA_CMD_CREATE_UC_RESP);
+       if (ret < 0) {
+               dev_warn(&vdev->pdev->dev,
+                        "could not create ucontext, error: %d\n", ret);
+               ptr = ERR_PTR(ret);
+               goto err;
+       }
+
+       context->ctx_handle = resp->ctx_handle;
+
+       /* copy back to user */
+       uresp.qp_tab_size = vdev->dsr->caps.max_qp;
+       ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
+       if (ret) {
+               pvrdma_uar_free(vdev, &context->uar);
+               context->ibucontext.device = ibdev;
+               pvrdma_dealloc_ucontext(&context->ibucontext);
+               return ERR_PTR(-EFAULT);
+       }
+
+       return &context->ibucontext;
+
+err:
+       pvrdma_uar_free(vdev, &context->uar);
+       kfree(context);
+       return ptr;
+}
+
+/**
+ * pvrdma_dealloc_ucontext - deallocate ucontext
+ * @ibcontext: the ucontext
+ *
+ * @return: 0 on success, otherwise errno.
+ */
+int pvrdma_dealloc_ucontext(struct ib_ucontext *ibcontext)
+{
+       struct pvrdma_ucontext *context = to_vucontext(ibcontext);
+       union pvrdma_cmd_req req;
+       struct pvrdma_cmd_destroy_uc *cmd = &req.destroy_uc;
+       int ret;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_DESTROY_UC;
+       cmd->ctx_handle = context->ctx_handle;
+
+       ret = pvrdma_cmd_post(context->dev, &req, NULL, 0);
+       if (ret < 0)
+               dev_warn(&context->dev->pdev->dev,
+                        "destroy ucontext failed, error: %d\n", ret);
+
+       /* Free the UAR even if the device command failed */
+       pvrdma_uar_free(to_vdev(ibcontext->device), &context->uar);
+       kfree(context);
+
+       return ret;
+}
+
+/**
+ * pvrdma_mmap - create mmap region
+ * @ibcontext: the user context
+ * @vma: the VMA
+ *
+ * @return: 0 on success, otherwise errno.
+ */
+int pvrdma_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
+{
+       struct pvrdma_ucontext *context = to_vucontext(ibcontext);
+       unsigned long start = vma->vm_start;
+       unsigned long size = vma->vm_end - vma->vm_start;
+       unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
+
+       dev_dbg(&context->dev->pdev->dev, "create mmap region\n");
+
+       if ((size != PAGE_SIZE) || (offset & ~PAGE_MASK)) {
+               dev_warn(&context->dev->pdev->dev,
+                        "invalid params for mmap region\n");
+               return -EINVAL;
+       }
+
+       /* Map UAR to kernel space, VM_LOCKED? */
+       vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
+       vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+       if (io_remap_pfn_range(vma, start, context->uar.pfn, size,
+                              vma->vm_page_prot))
+               return -EAGAIN;
+
+       return 0;
+}
+
+/**
+ * pvrdma_alloc_pd - allocate protection domain
+ * @ibdev: the IB device
+ * @context: user context
+ * @udata: user data
+ *
+ * @return: the ib_pd protection domain pointer on success, otherwise errno.
+ */
+struct ib_pd *pvrdma_alloc_pd(struct ib_device *ibdev,
+                             struct ib_ucontext *context,
+                             struct ib_udata *udata)
+{
+       struct pvrdma_pd *pd;
+       struct pvrdma_dev *dev = to_vdev(ibdev);
+       union pvrdma_cmd_req req;
+       union pvrdma_cmd_resp rsp;
+       struct pvrdma_cmd_create_pd *cmd = &req.create_pd;
+       struct pvrdma_cmd_create_pd_resp *resp = &rsp.create_pd_resp;
+       int ret;
+       void *ptr;
+
+       /* Check allowed max pds */
+       if (!atomic_add_unless(&dev->num_pds, 1, dev->dsr->caps.max_pd))
+               return ERR_PTR(-ENOMEM);
+
+       pd = kmalloc(sizeof(*pd), GFP_KERNEL);
+       if (!pd) {
+               ptr = ERR_PTR(-ENOMEM);
+               goto err;
+       }
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_CREATE_PD;
+       cmd->ctx_handle = (context) ? to_vucontext(context)->ctx_handle : 0;
+       ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_PD_RESP);
+       if (ret < 0) {
+               dev_warn(&dev->pdev->dev,
+                        "failed to allocate protection domain, error: %d\n",
+                        ret);
+               ptr = ERR_PTR(ret);
+               goto freepd;
+       }
+
+       pd->privileged = !context;
+       pd->pd_handle = resp->pd_handle;
+       pd->pdn = resp->pd_handle;
+
+       if (context) {
+               if (ib_copy_to_udata(udata, &pd->pdn, sizeof(__u32))) {
+                       dev_warn(&dev->pdev->dev,
+                                "failed to copy back protection domain\n");
+                       pvrdma_dealloc_pd(&pd->ibpd);
+                       return ERR_PTR(-EFAULT);
+               }
+       }
+
+       /* u32 pd handle */
+       return &pd->ibpd;
+
+freepd:
+       kfree(pd);
+err:
+       atomic_dec(&dev->num_pds);
+       return ptr;
+}
+
+/**
+ * pvrdma_dealloc_pd - deallocate protection domain
+ * @pd: the protection domain to be released
+ *
+ * @return: 0 on success, otherwise errno.
+ */
+int pvrdma_dealloc_pd(struct ib_pd *pd)
+{
+       struct pvrdma_dev *dev = to_vdev(pd->device);
+       union pvrdma_cmd_req req;
+       struct pvrdma_cmd_destroy_pd *cmd = &req.destroy_pd;
+       int ret;
+
+       memset(cmd, 0, sizeof(*cmd));
+       cmd->hdr.cmd = PVRDMA_CMD_DESTROY_PD;
+       cmd->pd_handle = to_vpd(pd)->pd_handle;
+
+       ret = pvrdma_cmd_post(dev, &req, NULL, 0);
+       if (ret)
+               dev_warn(&dev->pdev->dev,
+                        "could not dealloc protection domain, error: %d\n",
+                        ret);
+
+       kfree(to_vpd(pd));
+       atomic_dec(&dev->num_pds);
+
+       return 0;
+}
+
+/**
+ * pvrdma_create_ah - create an address handle
+ * @pd: the protection domain
+ * @ah_attr: the attributes of the AH
+ * @udata: user data blob
+ *
+ * @return: the ib_ah pointer on success, otherwise errno.
+ */
+struct ib_ah *pvrdma_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr,
+                              struct ib_udata *udata)
+{
+       struct pvrdma_dev *dev = to_vdev(pd->device);
+       struct pvrdma_ah *ah;
+       enum rdma_link_layer ll;
+
+       if (!(ah_attr->ah_flags & IB_AH_GRH))
+               return ERR_PTR(-EINVAL);
+
+       ll = rdma_port_get_link_layer(pd->device, ah_attr->port_num);
+
+       if (ll != IB_LINK_LAYER_ETHERNET ||
+           rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw))
+               return ERR_PTR(-EINVAL);
+
+       if (!atomic_add_unless(&dev->num_ahs, 1, dev->dsr->caps.max_ah))
+               return ERR_PTR(-ENOMEM);
+
+       ah = kzalloc(sizeof(*ah), GFP_KERNEL);
+       if (!ah) {
+               atomic_dec(&dev->num_ahs);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       ah->av.port_pd = to_vpd(pd)->pd_handle | (ah_attr->port_num << 24);
+       ah->av.src_path_bits = ah_attr->src_path_bits;
+       ah->av.src_path_bits |= 0x80;
+       ah->av.gid_index = ah_attr->grh.sgid_index;
+       ah->av.hop_limit = ah_attr->grh.hop_limit;
+       ah->av.sl_tclass_flowlabel = (ah_attr->grh.traffic_class << 20) |
+                                     ah_attr->grh.flow_label;
+       memcpy(ah->av.dgid, ah_attr->grh.dgid.raw, 16);
+       memcpy(ah->av.dmac, ah_attr->dmac, 6);
+
+       ah->ibah.device = pd->device;
+       ah->ibah.pd = pd;
+       ah->ibah.uobject = NULL;
+
+       return &ah->ibah;
+}
+
+/**
+ * pvrdma_destroy_ah - destroy an address handle
+ * @ah: the address handle to destroyed
+ *
+ * @return: 0 on success.
+ */
+int pvrdma_destroy_ah(struct ib_ah *ah)
+{
+       struct pvrdma_dev *dev = to_vdev(ah->device);
+
+       kfree(to_vah(ah));
+       atomic_dec(&dev->num_ahs);
+
+       return 0;
+}
diff --git a/drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.h b/drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.h
new file mode 100644 (file)
index 0000000..bfbe96b
--- /dev/null
@@ -0,0 +1,436 @@
+/*
+ * Copyright (c) 2012-2016 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.
+ */
+
+#ifndef __PVRDMA_VERBS_H__
+#define __PVRDMA_VERBS_H__
+
+#include <linux/types.h>
+
+union pvrdma_gid {
+       u8      raw[16];
+       struct {
+               __be64  subnet_prefix;
+               __be64  interface_id;
+       } global;
+};
+
+enum pvrdma_link_layer {
+       PVRDMA_LINK_LAYER_UNSPECIFIED,
+       PVRDMA_LINK_LAYER_INFINIBAND,
+       PVRDMA_LINK_LAYER_ETHERNET,
+};
+
+enum pvrdma_mtu {
+       PVRDMA_MTU_256  = 1,
+       PVRDMA_MTU_512  = 2,
+       PVRDMA_MTU_1024 = 3,
+       PVRDMA_MTU_2048 = 4,
+       PVRDMA_MTU_4096 = 5,
+};
+
+static inline int pvrdma_mtu_enum_to_int(enum pvrdma_mtu mtu)
+{
+       switch (mtu) {
+       case PVRDMA_MTU_256:    return  256;
+       case PVRDMA_MTU_512:    return  512;
+       case PVRDMA_MTU_1024:   return 1024;
+       case PVRDMA_MTU_2048:   return 2048;
+       case PVRDMA_MTU_4096:   return 4096;
+       default:                return   -1;
+       }
+}
+
+static inline enum pvrdma_mtu pvrdma_mtu_int_to_enum(int mtu)
+{
+       switch (mtu) {
+       case 256:       return PVRDMA_MTU_256;
+       case 512:       return PVRDMA_MTU_512;
+       case 1024:      return PVRDMA_MTU_1024;
+       case 2048:      return PVRDMA_MTU_2048;
+       case 4096:
+       default:        return PVRDMA_MTU_4096;
+       }
+}
+
+enum pvrdma_port_state {
+       PVRDMA_PORT_NOP                 = 0,
+       PVRDMA_PORT_DOWN                = 1,
+       PVRDMA_PORT_INIT                = 2,
+       PVRDMA_PORT_ARMED               = 3,
+       PVRDMA_PORT_ACTIVE              = 4,
+       PVRDMA_PORT_ACTIVE_DEFER        = 5,
+};
+
+enum pvrdma_port_cap_flags {
+       PVRDMA_PORT_SM                          = 1 <<  1,
+       PVRDMA_PORT_NOTICE_SUP                  = 1 <<  2,
+       PVRDMA_PORT_TRAP_SUP                    = 1 <<  3,
+       PVRDMA_PORT_OPT_IPD_SUP                 = 1 <<  4,
+       PVRDMA_PORT_AUTO_MIGR_SUP               = 1 <<  5,
+       PVRDMA_PORT_SL_MAP_SUP                  = 1 <<  6,
+       PVRDMA_PORT_MKEY_NVRAM                  = 1 <<  7,
+       PVRDMA_PORT_PKEY_NVRAM                  = 1 <<  8,
+       PVRDMA_PORT_LED_INFO_SUP                = 1 <<  9,
+       PVRDMA_PORT_SM_DISABLED                 = 1 << 10,
+       PVRDMA_PORT_SYS_IMAGE_GUID_SUP          = 1 << 11,
+       PVRDMA_PORT_PKEY_SW_EXT_PORT_TRAP_SUP   = 1 << 12,
+       PVRDMA_PORT_EXTENDED_SPEEDS_SUP         = 1 << 14,
+       PVRDMA_PORT_CM_SUP                      = 1 << 16,
+       PVRDMA_PORT_SNMP_TUNNEL_SUP             = 1 << 17,
+       PVRDMA_PORT_REINIT_SUP                  = 1 << 18,
+       PVRDMA_PORT_DEVICE_MGMT_SUP             = 1 << 19,
+       PVRDMA_PORT_VENDOR_CLASS_SUP            = 1 << 20,
+       PVRDMA_PORT_DR_NOTICE_SUP               = 1 << 21,
+       PVRDMA_PORT_CAP_MASK_NOTICE_SUP         = 1 << 22,
+       PVRDMA_PORT_BOOT_MGMT_SUP               = 1 << 23,
+       PVRDMA_PORT_LINK_LATENCY_SUP            = 1 << 24,
+       PVRDMA_PORT_CLIENT_REG_SUP              = 1 << 25,
+       PVRDMA_PORT_IP_BASED_GIDS               = 1 << 26,
+       PVRDMA_PORT_CAP_FLAGS_MAX               = PVRDMA_PORT_IP_BASED_GIDS,
+};
+
+enum pvrdma_port_width {
+       PVRDMA_WIDTH_1X         = 1,
+       PVRDMA_WIDTH_4X         = 2,
+       PVRDMA_WIDTH_8X         = 4,
+       PVRDMA_WIDTH_12X        = 8,
+};
+
+static inline int pvrdma_width_enum_to_int(enum pvrdma_port_width width)
+{
+       switch (width) {
+       case PVRDMA_WIDTH_1X:   return  1;
+       case PVRDMA_WIDTH_4X:   return  4;
+       case PVRDMA_WIDTH_8X:   return  8;
+       case PVRDMA_WIDTH_12X:  return 12;
+       default:                return -1;
+       }
+}
+
+enum pvrdma_port_speed {
+       PVRDMA_SPEED_SDR        = 1,
+       PVRDMA_SPEED_DDR        = 2,
+       PVRDMA_SPEED_QDR        = 4,
+       PVRDMA_SPEED_FDR10      = 8,
+       PVRDMA_SPEED_FDR        = 16,
+       PVRDMA_SPEED_EDR        = 32,
+};
+
+struct pvrdma_port_attr {
+       enum pvrdma_port_state  state;
+       enum pvrdma_mtu         max_mtu;
+       enum pvrdma_mtu         active_mtu;
+       u32                     gid_tbl_len;
+       u32                     port_cap_flags;
+       u32                     max_msg_sz;
+       u32                     bad_pkey_cntr;
+       u32                     qkey_viol_cntr;
+       u16                     pkey_tbl_len;
+       u16                     lid;
+       u16                     sm_lid;
+       u8                      lmc;
+       u8                      max_vl_num;
+       u8                      sm_sl;
+       u8                      subnet_timeout;
+       u8                      init_type_reply;
+       u8                      active_width;
+       u8                      active_speed;
+       u8                      phys_state;
+       u8                      reserved[2];
+};
+
+struct pvrdma_global_route {
+       union pvrdma_gid        dgid;
+       u32                     flow_label;
+       u8                      sgid_index;
+       u8                      hop_limit;
+       u8                      traffic_class;
+       u8                      reserved;
+};
+
+struct pvrdma_grh {
+       __be32                  version_tclass_flow;
+       __be16                  paylen;
+       u8                      next_hdr;
+       u8                      hop_limit;
+       union pvrdma_gid        sgid;
+       union pvrdma_gid        dgid;
+};
+
+enum pvrdma_ah_flags {
+       PVRDMA_AH_GRH = 1,
+};
+
+enum pvrdma_rate {
+       PVRDMA_RATE_PORT_CURRENT        = 0,
+       PVRDMA_RATE_2_5_GBPS            = 2,
+       PVRDMA_RATE_5_GBPS              = 5,
+       PVRDMA_RATE_10_GBPS             = 3,
+       PVRDMA_RATE_20_GBPS             = 6,
+       PVRDMA_RATE_30_GBPS             = 4,
+       PVRDMA_RATE_40_GBPS             = 7,
+       PVRDMA_RATE_60_GBPS             = 8,
+       PVRDMA_RATE_80_GBPS             = 9,
+       PVRDMA_RATE_120_GBPS            = 10,
+       PVRDMA_RATE_14_GBPS             = 11,
+       PVRDMA_RATE_56_GBPS             = 12,
+       PVRDMA_RATE_112_GBPS            = 13,
+       PVRDMA_RATE_168_GBPS            = 14,
+       PVRDMA_RATE_25_GBPS             = 15,
+       PVRDMA_RATE_100_GBPS            = 16,
+       PVRDMA_RATE_200_GBPS            = 17,
+       PVRDMA_RATE_300_GBPS            = 18,
+};
+
+struct pvrdma_ah_attr {
+       struct pvrdma_global_route      grh;
+       u16                             dlid;
+       u16                             vlan_id;
+       u8                              sl;
+       u8                              src_path_bits;
+       u8                              static_rate;
+       u8                              ah_flags;
+       u8                              port_num;
+       u8                              dmac[6];
+       u8                              reserved;
+};
+
+enum pvrdma_cq_notify_flags {
+       PVRDMA_CQ_SOLICITED             = 1 << 0,
+       PVRDMA_CQ_NEXT_COMP             = 1 << 1,
+       PVRDMA_CQ_SOLICITED_MASK        = PVRDMA_CQ_SOLICITED |
+                                         PVRDMA_CQ_NEXT_COMP,
+       PVRDMA_CQ_REPORT_MISSED_EVENTS  = 1 << 2,
+};
+
+struct pvrdma_qp_cap {
+       u32     max_send_wr;
+       u32     max_recv_wr;
+       u32     max_send_sge;
+       u32     max_recv_sge;
+       u32     max_inline_data;
+       u32     reserved;
+};
+
+enum pvrdma_sig_type {
+       PVRDMA_SIGNAL_ALL_WR,
+       PVRDMA_SIGNAL_REQ_WR,
+};
+
+enum pvrdma_qp_type {
+       PVRDMA_QPT_SMI,
+       PVRDMA_QPT_GSI,
+       PVRDMA_QPT_RC,
+       PVRDMA_QPT_UC,
+       PVRDMA_QPT_UD,
+       PVRDMA_QPT_RAW_IPV6,
+       PVRDMA_QPT_RAW_ETHERTYPE,
+       PVRDMA_QPT_RAW_PACKET = 8,
+       PVRDMA_QPT_XRC_INI = 9,
+       PVRDMA_QPT_XRC_TGT,
+       PVRDMA_QPT_MAX,
+};
+
+enum pvrdma_qp_create_flags {
+       PVRDMA_QP_CREATE_IPOPVRDMA_UD_LSO               = 1 << 0,
+       PVRDMA_QP_CREATE_BLOCK_MULTICAST_LOOPBACK       = 1 << 1,
+};
+
+enum pvrdma_qp_attr_mask {
+       PVRDMA_QP_STATE                 = 1 << 0,
+       PVRDMA_QP_CUR_STATE             = 1 << 1,
+       PVRDMA_QP_EN_SQD_ASYNC_NOTIFY   = 1 << 2,
+       PVRDMA_QP_ACCESS_FLAGS          = 1 << 3,
+       PVRDMA_QP_PKEY_INDEX            = 1 << 4,
+       PVRDMA_QP_PORT                  = 1 << 5,
+       PVRDMA_QP_QKEY                  = 1 << 6,
+       PVRDMA_QP_AV                    = 1 << 7,
+       PVRDMA_QP_PATH_MTU              = 1 << 8,
+       PVRDMA_QP_TIMEOUT               = 1 << 9,
+       PVRDMA_QP_RETRY_CNT             = 1 << 10,
+       PVRDMA_QP_RNR_RETRY             = 1 << 11,
+       PVRDMA_QP_RQ_PSN                = 1 << 12,
+       PVRDMA_QP_MAX_QP_RD_ATOMIC      = 1 << 13,
+       PVRDMA_QP_ALT_PATH              = 1 << 14,
+       PVRDMA_QP_MIN_RNR_TIMER         = 1 << 15,
+       PVRDMA_QP_SQ_PSN                = 1 << 16,
+       PVRDMA_QP_MAX_DEST_RD_ATOMIC    = 1 << 17,
+       PVRDMA_QP_PATH_MIG_STATE        = 1 << 18,
+       PVRDMA_QP_CAP                   = 1 << 19,
+       PVRDMA_QP_DEST_QPN              = 1 << 20,
+       PVRDMA_QP_ATTR_MASK_MAX         = PVRDMA_QP_DEST_QPN,
+};
+
+enum pvrdma_qp_state {
+       PVRDMA_QPS_RESET,
+       PVRDMA_QPS_INIT,
+       PVRDMA_QPS_RTR,
+       PVRDMA_QPS_RTS,
+       PVRDMA_QPS_SQD,
+       PVRDMA_QPS_SQE,
+       PVRDMA_QPS_ERR,
+};
+
+enum pvrdma_mig_state {
+       PVRDMA_MIG_MIGRATED,
+       PVRDMA_MIG_REARM,
+       PVRDMA_MIG_ARMED,
+};
+
+enum pvrdma_mw_type {
+       PVRDMA_MW_TYPE_1 = 1,
+       PVRDMA_MW_TYPE_2 = 2,
+};
+
+struct pvrdma_qp_attr {
+       enum pvrdma_qp_state    qp_state;
+       enum pvrdma_qp_state    cur_qp_state;
+       enum pvrdma_mtu         path_mtu;
+       enum pvrdma_mig_state   path_mig_state;
+       u32                     qkey;
+       u32                     rq_psn;
+       u32                     sq_psn;
+       u32                     dest_qp_num;
+       u32                     qp_access_flags;
+       u16                     pkey_index;
+       u16                     alt_pkey_index;
+       u8                      en_sqd_async_notify;
+       u8                      sq_draining;
+       u8                      max_rd_atomic;
+       u8                      max_dest_rd_atomic;
+       u8                      min_rnr_timer;
+       u8                      port_num;
+       u8                      timeout;
+       u8                      retry_cnt;
+       u8                      rnr_retry;
+       u8                      alt_port_num;
+       u8                      alt_timeout;
+       u8                      reserved[5];
+       struct pvrdma_qp_cap    cap;
+       struct pvrdma_ah_attr   ah_attr;
+       struct pvrdma_ah_attr   alt_ah_attr;
+};
+
+enum pvrdma_send_flags {
+       PVRDMA_SEND_FENCE       = 1 << 0,
+       PVRDMA_SEND_SIGNALED    = 1 << 1,
+       PVRDMA_SEND_SOLICITED   = 1 << 2,
+       PVRDMA_SEND_INLINE      = 1 << 3,
+       PVRDMA_SEND_IP_CSUM     = 1 << 4,
+       PVRDMA_SEND_FLAGS_MAX   = PVRDMA_SEND_IP_CSUM,
+};
+
+enum pvrdma_access_flags {
+       PVRDMA_ACCESS_LOCAL_WRITE       = 1 << 0,
+       PVRDMA_ACCESS_REMOTE_WRITE      = 1 << 1,
+       PVRDMA_ACCESS_REMOTE_READ       = 1 << 2,
+       PVRDMA_ACCESS_REMOTE_ATOMIC     = 1 << 3,
+       PVRDMA_ACCESS_MW_BIND           = 1 << 4,
+       PVRDMA_ZERO_BASED               = 1 << 5,
+       PVRDMA_ACCESS_ON_DEMAND         = 1 << 6,
+       PVRDMA_ACCESS_FLAGS_MAX         = PVRDMA_ACCESS_ON_DEMAND,
+};
+
+int pvrdma_query_device(struct ib_device *ibdev,
+                       struct ib_device_attr *props,
+                       struct ib_udata *udata);
+int pvrdma_query_port(struct ib_device *ibdev, u8 port,
+                     struct ib_port_attr *props);
+int pvrdma_query_gid(struct ib_device *ibdev, u8 port,
+                    int index, union ib_gid *gid);
+int pvrdma_query_pkey(struct ib_device *ibdev, u8 port,
+                     u16 index, u16 *pkey);
+enum rdma_link_layer pvrdma_port_link_layer(struct ib_device *ibdev,
+                                           u8 port);
+int pvrdma_modify_device(struct ib_device *ibdev, int mask,
+                        struct ib_device_modify *props);
+int pvrdma_modify_port(struct ib_device *ibdev, u8 port,
+                      int mask, struct ib_port_modify *props);
+int pvrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma);
+struct ib_ucontext *pvrdma_alloc_ucontext(struct ib_device *ibdev,
+                                         struct ib_udata *udata);
+int pvrdma_dealloc_ucontext(struct ib_ucontext *context);
+struct ib_pd *pvrdma_alloc_pd(struct ib_device *ibdev,
+                             struct ib_ucontext *context,
+                             struct ib_udata *udata);
+int pvrdma_dealloc_pd(struct ib_pd *ibpd);
+struct ib_mr *pvrdma_get_dma_mr(struct ib_pd *pd, int acc);
+struct ib_mr *pvrdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+                                u64 virt_addr, int access_flags,
+                                struct ib_udata *udata);
+int pvrdma_dereg_mr(struct ib_mr *mr);
+struct ib_mr *pvrdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
+                             u32 max_num_sg);
+int pvrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
+                    int sg_nents, unsigned int *sg_offset);
+int pvrdma_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period);
+int pvrdma_resize_cq(struct ib_cq *ibcq, int entries,
+                    struct ib_udata *udata);
+struct ib_cq *pvrdma_create_cq(struct ib_device *ibdev,
+                              const struct ib_cq_init_attr *attr,
+                              struct ib_ucontext *context,
+                              struct ib_udata *udata);
+int pvrdma_resize_cq(struct ib_cq *ibcq, int entries,
+                    struct ib_udata *udata);
+int pvrdma_destroy_cq(struct ib_cq *cq);
+int pvrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
+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 ib_ah_attr *ah_attr,
+                              struct ib_udata *udata);
+int pvrdma_destroy_ah(struct ib_ah *ah);
+struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
+                              struct ib_qp_init_attr *init_attr,
+                              struct ib_udata *udata);
+int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+                    int attr_mask, struct ib_udata *udata);
+int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
+                   int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr);
+int pvrdma_destroy_qp(struct ib_qp *qp);
+int pvrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+                    struct ib_send_wr **bad_wr);
+int pvrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
+                    struct ib_recv_wr **bad_wr);
+
+#endif /* __PVRDMA_VERBS_H__ */
index f14ab7f..6a8a934 100644 (file)
@@ -14,3 +14,4 @@ header-y += mlx5-abi.h
 header-y += mthca-abi.h
 header-y += nes-abi.h
 header-y += ocrdma-abi.h
+header-y += vmw_pvrdma-abi.h
diff --git a/include/uapi/rdma/vmw_pvrdma-abi.h b/include/uapi/rdma/vmw_pvrdma-abi.h
new file mode 100644 (file)
index 0000000..5016abc
--- /dev/null
@@ -0,0 +1,289 @@
+/*
+ * Copyright (c) 2012-2016 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.
+ */
+
+#ifndef __VMW_PVRDMA_ABI_H__
+#define __VMW_PVRDMA_ABI_H__
+
+#include <linux/types.h>
+
+#define PVRDMA_UVERBS_ABI_VERSION      3               /* ABI Version. */
+#define PVRDMA_UAR_HANDLE_MASK         0x00FFFFFF      /* Bottom 24 bits. */
+#define PVRDMA_UAR_QP_OFFSET           0               /* QP doorbell. */
+#define PVRDMA_UAR_QP_SEND             BIT(30)         /* Send bit. */
+#define PVRDMA_UAR_QP_RECV             BIT(31)         /* Recv bit. */
+#define PVRDMA_UAR_CQ_OFFSET           4               /* CQ doorbell. */
+#define PVRDMA_UAR_CQ_ARM_SOL          BIT(29)         /* Arm solicited bit. */
+#define PVRDMA_UAR_CQ_ARM              BIT(30)         /* Arm bit. */
+#define PVRDMA_UAR_CQ_POLL             BIT(31)         /* Poll bit. */
+
+enum pvrdma_wr_opcode {
+       PVRDMA_WR_RDMA_WRITE,
+       PVRDMA_WR_RDMA_WRITE_WITH_IMM,
+       PVRDMA_WR_SEND,
+       PVRDMA_WR_SEND_WITH_IMM,
+       PVRDMA_WR_RDMA_READ,
+       PVRDMA_WR_ATOMIC_CMP_AND_SWP,
+       PVRDMA_WR_ATOMIC_FETCH_AND_ADD,
+       PVRDMA_WR_LSO,
+       PVRDMA_WR_SEND_WITH_INV,
+       PVRDMA_WR_RDMA_READ_WITH_INV,
+       PVRDMA_WR_LOCAL_INV,
+       PVRDMA_WR_FAST_REG_MR,
+       PVRDMA_WR_MASKED_ATOMIC_CMP_AND_SWP,
+       PVRDMA_WR_MASKED_ATOMIC_FETCH_AND_ADD,
+       PVRDMA_WR_BIND_MW,
+       PVRDMA_WR_REG_SIG_MR,
+};
+
+enum pvrdma_wc_status {
+       PVRDMA_WC_SUCCESS,
+       PVRDMA_WC_LOC_LEN_ERR,
+       PVRDMA_WC_LOC_QP_OP_ERR,
+       PVRDMA_WC_LOC_EEC_OP_ERR,
+       PVRDMA_WC_LOC_PROT_ERR,
+       PVRDMA_WC_WR_FLUSH_ERR,
+       PVRDMA_WC_MW_BIND_ERR,
+       PVRDMA_WC_BAD_RESP_ERR,
+       PVRDMA_WC_LOC_ACCESS_ERR,
+       PVRDMA_WC_REM_INV_REQ_ERR,
+       PVRDMA_WC_REM_ACCESS_ERR,
+       PVRDMA_WC_REM_OP_ERR,
+       PVRDMA_WC_RETRY_EXC_ERR,
+       PVRDMA_WC_RNR_RETRY_EXC_ERR,
+       PVRDMA_WC_LOC_RDD_VIOL_ERR,
+       PVRDMA_WC_REM_INV_RD_REQ_ERR,
+       PVRDMA_WC_REM_ABORT_ERR,
+       PVRDMA_WC_INV_EECN_ERR,
+       PVRDMA_WC_INV_EEC_STATE_ERR,
+       PVRDMA_WC_FATAL_ERR,
+       PVRDMA_WC_RESP_TIMEOUT_ERR,
+       PVRDMA_WC_GENERAL_ERR,
+};
+
+enum pvrdma_wc_opcode {
+       PVRDMA_WC_SEND,
+       PVRDMA_WC_RDMA_WRITE,
+       PVRDMA_WC_RDMA_READ,
+       PVRDMA_WC_COMP_SWAP,
+       PVRDMA_WC_FETCH_ADD,
+       PVRDMA_WC_BIND_MW,
+       PVRDMA_WC_LSO,
+       PVRDMA_WC_LOCAL_INV,
+       PVRDMA_WC_FAST_REG_MR,
+       PVRDMA_WC_MASKED_COMP_SWAP,
+       PVRDMA_WC_MASKED_FETCH_ADD,
+       PVRDMA_WC_RECV = 1 << 7,
+       PVRDMA_WC_RECV_RDMA_WITH_IMM,
+};
+
+enum pvrdma_wc_flags {
+       PVRDMA_WC_GRH                   = 1 << 0,
+       PVRDMA_WC_WITH_IMM              = 1 << 1,
+       PVRDMA_WC_WITH_INVALIDATE       = 1 << 2,
+       PVRDMA_WC_IP_CSUM_OK            = 1 << 3,
+       PVRDMA_WC_WITH_SMAC             = 1 << 4,
+       PVRDMA_WC_WITH_VLAN             = 1 << 5,
+       PVRDMA_WC_FLAGS_MAX             = PVRDMA_WC_WITH_VLAN,
+};
+
+struct pvrdma_alloc_ucontext_resp {
+       __u32 qp_tab_size;
+       __u32 reserved;
+};
+
+struct pvrdma_alloc_pd_resp {
+       __u32 pdn;
+       __u32 reserved;
+};
+
+struct pvrdma_create_cq {
+       __u64 buf_addr;
+       __u32 buf_size;
+       __u32 reserved;
+};
+
+struct pvrdma_create_cq_resp {
+       __u32 cqn;
+       __u32 reserved;
+};
+
+struct pvrdma_resize_cq {
+       __u64 buf_addr;
+       __u32 buf_size;
+       __u32 reserved;
+};
+
+struct pvrdma_create_srq {
+       __u64 buf_addr;
+};
+
+struct pvrdma_create_srq_resp {
+       __u32 srqn;
+       __u32 reserved;
+};
+
+struct pvrdma_create_qp {
+       __u64 rbuf_addr;
+       __u64 sbuf_addr;
+       __u32 rbuf_size;
+       __u32 sbuf_size;
+       __u64 qp_addr;
+};
+
+/* PVRDMA masked atomic compare and swap */
+struct pvrdma_ex_cmp_swap {
+       __u64 swap_val;
+       __u64 compare_val;
+       __u64 swap_mask;
+       __u64 compare_mask;
+};
+
+/* PVRDMA masked atomic fetch and add */
+struct pvrdma_ex_fetch_add {
+       __u64 add_val;
+       __u64 field_boundary;
+};
+
+/* PVRDMA address vector. */
+struct pvrdma_av {
+       __u32 port_pd;
+       __u32 sl_tclass_flowlabel;
+       __u8 dgid[16];
+       __u8 src_path_bits;
+       __u8 gid_index;
+       __u8 stat_rate;
+       __u8 hop_limit;
+       __u8 dmac[6];
+       __u8 reserved[6];
+};
+
+/* PVRDMA scatter/gather entry */
+struct pvrdma_sge {
+       __u64   addr;
+       __u32   length;
+       __u32   lkey;
+};
+
+/* PVRDMA receive queue work request */
+struct pvrdma_rq_wqe_hdr {
+       __u64 wr_id;            /* wr id */
+       __u32 num_sge;          /* size of s/g array */
+       __u32 total_len;        /* reserved */
+};
+/* Use pvrdma_sge (ib_sge) for receive queue s/g array elements. */
+
+/* PVRDMA send queue work request */
+struct pvrdma_sq_wqe_hdr {
+       __u64 wr_id;            /* wr id */
+       __u32 num_sge;          /* size of s/g array */
+       __u32 total_len;        /* reserved */
+       __u32 opcode;           /* operation type */
+       __u32 send_flags;       /* wr flags */
+       union {
+               __u32 imm_data;
+               __u32 invalidate_rkey;
+       } ex;
+       __u32 reserved;
+       union {
+               struct {
+                       __u64 remote_addr;
+                       __u32 rkey;
+                       __u8 reserved[4];
+               } rdma;
+               struct {
+                       __u64 remote_addr;
+                       __u64 compare_add;
+                       __u64 swap;
+                       __u32 rkey;
+                       __u32 reserved;
+               } atomic;
+               struct {
+                       __u64 remote_addr;
+                       __u32 log_arg_sz;
+                       __u32 rkey;
+                       union {
+                               struct pvrdma_ex_cmp_swap  cmp_swap;
+                               struct pvrdma_ex_fetch_add fetch_add;
+                       } wr_data;
+               } masked_atomics;
+               struct {
+                       __u64 iova_start;
+                       __u64 pl_pdir_dma;
+                       __u32 page_shift;
+                       __u32 page_list_len;
+                       __u32 length;
+                       __u32 access_flags;
+                       __u32 rkey;
+               } fast_reg;
+               struct {
+                       __u32 remote_qpn;
+                       __u32 remote_qkey;
+                       struct pvrdma_av av;
+               } ud;
+       } wr;
+};
+/* Use pvrdma_sge (ib_sge) for send queue s/g array elements. */
+
+/* Completion queue element. */
+struct pvrdma_cqe {
+       __u64 wr_id;
+       __u64 qp;
+       __u32 opcode;
+       __u32 status;
+       __u32 byte_len;
+       __u32 imm_data;
+       __u32 src_qp;
+       __u32 wc_flags;
+       __u32 vendor_err;
+       __u16 pkey_index;
+       __u16 slid;
+       __u8 sl;
+       __u8 dlid_path_bits;
+       __u8 port_num;
+       __u8 smac[6];
+       __u8 reserved2[7]; /* Pad to next power of 2 (64). */
+};
+
+#endif /* __VMW_PVRDMA_ABI_H__ */