]> git.karo-electronics.de Git - linux-beck.git/commitdiff
ib_srpt: Initial SRP Target merge for v3.3-rc1
authorBart Van Assche <bvanassche@acm.org>
Fri, 14 Oct 2011 01:30:46 +0000 (01:30 +0000)
committerNicholas Bellinger <nab@linux-iscsi.org>
Fri, 16 Dec 2011 06:33:56 +0000 (06:33 +0000)
This patch adds the kernel module ib_srpt SCSI RDMA Protocol (SRP) target
implementation conforming to the SRP r16a specification for the mainline
drivers/target infrastructure.

This driver was originally developed by Vu Pham and has been optimized by
Bart Van Assche and merged into upstream LIO based on his srpt-lio-4.1
branch here:

   https://github.com/bvanassche/srpt-lio/commits/srpt-lio-4.1/

This updated patch also contains the following two changes from
lio-core-2.6.git/master.  One is to fix a bug with 1 >= task->task_sg[]
chained mappings in ib_srpt, and the other to convert the configfs control
plane to reference IB Port GUID and struct srpt_port directly following
mainline v4.x target_core_fabric_configfs.c convertion for ib_srpt
to work with rtslib/rtsadmin v2 code.

These seperate patches can be found here:

ib_srpt: Fix bug with chainged SGLs in srpt_map_sg_to_ib_sge
http://www.risingtidesystems.com/git/?p=lio-core-2.6.git;a=commitdiff;h=ea485147563b6555a97dbf811825fbb586519252

ib_srpt: Convert se_wwn endpoint reference to struct srpt_port->port_wwn
http://www.risingtidesystems.com/git/?p=lio-core-2.6.git;a=commitdiff;h=4e544a210acb227df1bb4ca5086e65bdf4e648ea

This also includes the following recent v1 -> v2 review changes:

ib_srpt: Fix potential out-of-bounds array access
ib_srpt: Avoid failed multipart RDMA transfers
ib_srpt: Fix srpt_alloc_fabric_acl failure case return value
ib_srpt: Update comments to reference $driver/$port layout
ib_srpt: Fix sport->port_guid formatting code
ib_srpt: Remove legacy use_port_guid_in_session_name module parameter
ib_srpt: Convert srp_max_rdma_size into per port configfs attribute
ib_srpt: Convert srp_max_rsp_size into per port configfs attribute
ib_srpt: Convert srpt_sq_size into per port configfs attribute

and v2 -> v3 review changes:

ib_srpt: Fix possible race with srp_sq_size in srpt_create_ch_ib
ib_srpt: Fix possible race with srp_max_rsp_size in srpt_release_channel_work
ib_srpt: Fix up MAX_SRPT_RDMA_SIZE define
ib_srpt: Make srpt_map_sg_to_ib_sge() failure case return -EAGAIN
ib_srpt: Convert port_guid to use subnet_prefix + interface_id formatting
ib_srpt: Make srpt_check_stop_free return kref_put status
ib_srpt: Make compilation with BUG=n proceed`
ib_srpt: Use new target_core_fabric.h include
ib_srpt: Check hex2bin() return code to silence build warning

Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Roland Dreier <roland@purestorage.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Vu Pham <vu@mellanox.com>
Cc: David Dillow <dillowda@ornl.gov>
Signed-off-by: Nicholas A. Bellinger <nab@risingtidesystems.com>
drivers/infiniband/Kconfig
drivers/infiniband/Makefile
drivers/infiniband/ulp/srpt/Kconfig [new file with mode: 0644]
drivers/infiniband/ulp/srpt/Makefile [new file with mode: 0644]
drivers/infiniband/ulp/srpt/ib_dm_mad.h [new file with mode: 0644]
drivers/infiniband/ulp/srpt/ib_srpt.c [new file with mode: 0644]
drivers/infiniband/ulp/srpt/ib_srpt.h [new file with mode: 0644]

index 0f9a84c1046ab38776abe891ef0ff98ea0c19c98..eb0add311dc8be40ff8876f8b1bcd7d3f2e0b161 100644 (file)
@@ -55,6 +55,7 @@ source "drivers/infiniband/hw/nes/Kconfig"
 source "drivers/infiniband/ulp/ipoib/Kconfig"
 
 source "drivers/infiniband/ulp/srp/Kconfig"
+source "drivers/infiniband/ulp/srpt/Kconfig"
 
 source "drivers/infiniband/ulp/iser/Kconfig"
 
index 9cc7a47d3e6737f22379384ac60366ea183d65dc..a3b2d8eac86ef0089a41dff9686c4dd752b0ca4a 100644 (file)
@@ -10,4 +10,5 @@ obj-$(CONFIG_MLX4_INFINIBAND)         += hw/mlx4/
 obj-$(CONFIG_INFINIBAND_NES)           += hw/nes/
 obj-$(CONFIG_INFINIBAND_IPOIB)         += ulp/ipoib/
 obj-$(CONFIG_INFINIBAND_SRP)           += ulp/srp/
+obj-$(CONFIG_INFINIBAND_SRPT)          += ulp/srpt/
 obj-$(CONFIG_INFINIBAND_ISER)          += ulp/iser/
diff --git a/drivers/infiniband/ulp/srpt/Kconfig b/drivers/infiniband/ulp/srpt/Kconfig
new file mode 100644 (file)
index 0000000..31ee83d
--- /dev/null
@@ -0,0 +1,12 @@
+config INFINIBAND_SRPT
+       tristate "InfiniBand SCSI RDMA Protocol target support"
+       depends on INFINIBAND && TARGET_CORE
+       ---help---
+
+         Support for the SCSI RDMA Protocol (SRP) Target driver. The
+         SRP protocol is a protocol that allows an initiator to access
+         a block storage device on another host (target) over a network
+         that supports the RDMA protocol. Currently the RDMA protocol is
+         supported by InfiniBand and by iWarp network hardware. More
+         information about the SRP protocol can be found on the website
+         of the INCITS T10 technical committee (http://www.t10.org/).
diff --git a/drivers/infiniband/ulp/srpt/Makefile b/drivers/infiniband/ulp/srpt/Makefile
new file mode 100644 (file)
index 0000000..e3ee4bd
--- /dev/null
@@ -0,0 +1,2 @@
+ccflags-y                      := -Idrivers/target
+obj-$(CONFIG_INFINIBAND_SRPT)  += ib_srpt.o
diff --git a/drivers/infiniband/ulp/srpt/ib_dm_mad.h b/drivers/infiniband/ulp/srpt/ib_dm_mad.h
new file mode 100644 (file)
index 0000000..fb1de1f
--- /dev/null
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2006 - 2009 Mellanox Technology Inc.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+
+#ifndef IB_DM_MAD_H
+#define IB_DM_MAD_H
+
+#include <linux/types.h>
+
+#include <rdma/ib_mad.h>
+
+enum {
+       /*
+        * See also section 13.4.7 Status Field, table 115 MAD Common Status
+        * Field Bit Values and also section 16.3.1.1 Status Field in the
+        * InfiniBand Architecture Specification.
+        */
+       DM_MAD_STATUS_UNSUP_METHOD = 0x0008,
+       DM_MAD_STATUS_UNSUP_METHOD_ATTR = 0x000c,
+       DM_MAD_STATUS_INVALID_FIELD = 0x001c,
+       DM_MAD_STATUS_NO_IOC = 0x0100,
+
+       /*
+        * See also the Device Management chapter, section 16.3.3 Attributes,
+        * table 279 Device Management Attributes in the InfiniBand
+        * Architecture Specification.
+        */
+       DM_ATTR_CLASS_PORT_INFO = 0x01,
+       DM_ATTR_IOU_INFO = 0x10,
+       DM_ATTR_IOC_PROFILE = 0x11,
+       DM_ATTR_SVC_ENTRIES = 0x12
+};
+
+struct ib_dm_hdr {
+       u8 reserved[28];
+};
+
+/*
+ * Structure of management datagram sent by the SRP target implementation.
+ * Contains a management datagram header, reliable multi-packet transaction
+ * protocol (RMPP) header and ib_dm_hdr. Notes:
+ * - The SRP target implementation does not use RMPP or ib_dm_hdr when sending
+ *   management datagrams.
+ * - The header size must be exactly 64 bytes (IB_MGMT_DEVICE_HDR), since this
+ *   is the header size that is passed to ib_create_send_mad() in ib_srpt.c.
+ * - The maximum supported size for a management datagram when not using RMPP
+ *   is 256 bytes -- 64 bytes header and 192 (IB_MGMT_DEVICE_DATA) bytes data.
+ */
+struct ib_dm_mad {
+       struct ib_mad_hdr mad_hdr;
+       struct ib_rmpp_hdr rmpp_hdr;
+       struct ib_dm_hdr dm_hdr;
+       u8 data[IB_MGMT_DEVICE_DATA];
+};
+
+/*
+ * IOUnitInfo as defined in section 16.3.3.3 IOUnitInfo of the InfiniBand
+ * Architecture Specification.
+ */
+struct ib_dm_iou_info {
+       __be16 change_id;
+       u8 max_controllers;
+       u8 op_rom;
+       u8 controller_list[128];
+};
+
+/*
+ * IOControllerprofile as defined in section 16.3.3.4 IOControllerProfile of
+ * the InfiniBand Architecture Specification.
+ */
+struct ib_dm_ioc_profile {
+       __be64 guid;
+       __be32 vendor_id;
+       __be32 device_id;
+       __be16 device_version;
+       __be16 reserved1;
+       __be32 subsys_vendor_id;
+       __be32 subsys_device_id;
+       __be16 io_class;
+       __be16 io_subclass;
+       __be16 protocol;
+       __be16 protocol_version;
+       __be16 service_conn;
+       __be16 initiators_supported;
+       __be16 send_queue_depth;
+       u8 reserved2;
+       u8 rdma_read_depth;
+       __be32 send_size;
+       __be32 rdma_size;
+       u8 op_cap_mask;
+       u8 svc_cap_mask;
+       u8 num_svc_entries;
+       u8 reserved3[9];
+       u8 id_string[64];
+};
+
+struct ib_dm_svc_entry {
+       u8 name[40];
+       __be64 id;
+};
+
+/*
+ * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
+ * Specification. See also section B.7, table B.8 in the T10 SRP r16a document.
+ */
+struct ib_dm_svc_entries {
+       struct ib_dm_svc_entry service_entries[4];
+};
+
+#endif
diff --git a/drivers/infiniband/ulp/srpt/ib_srpt.c b/drivers/infiniband/ulp/srpt/ib_srpt.c
new file mode 100644 (file)
index 0000000..cd5d05e
--- /dev/null
@@ -0,0 +1,4073 @@
+/*
+ * Copyright (c) 2006 - 2009 Mellanox Technology Inc.  All rights reserved.
+ * Copyright (C) 2008 - 2011 Bart Van Assche <bvanassche@acm.org>.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/slab.h>
+#include <linux/err.h>
+#include <linux/ctype.h>
+#include <linux/kthread.h>
+#include <linux/string.h>
+#include <linux/delay.h>
+#include <linux/atomic.h>
+#include <scsi/scsi_tcq.h>
+#include <target/configfs_macros.h>
+#include <target/target_core_base.h>
+#include <target/target_core_fabric_configfs.h>
+#include <target/target_core_fabric.h>
+#include <target/target_core_configfs.h>
+#include "ib_srpt.h"
+
+/* Name of this kernel module. */
+#define DRV_NAME               "ib_srpt"
+#define DRV_VERSION            "2.0.0"
+#define DRV_RELDATE            "2011-02-14"
+
+#define SRPT_ID_STRING "Linux SRP target"
+
+#undef pr_fmt
+#define pr_fmt(fmt) DRV_NAME " " fmt
+
+MODULE_AUTHOR("Vu Pham and Bart Van Assche");
+MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol target "
+                  "v" DRV_VERSION " (" DRV_RELDATE ")");
+MODULE_LICENSE("Dual BSD/GPL");
+
+/*
+ * Global Variables
+ */
+
+static u64 srpt_service_guid;
+static spinlock_t srpt_dev_lock;       /* Protects srpt_dev_list. */
+static struct list_head srpt_dev_list; /* List of srpt_device structures. */
+
+static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
+module_param(srp_max_req_size, int, 0444);
+MODULE_PARM_DESC(srp_max_req_size,
+                "Maximum size of SRP request messages in bytes.");
+
+static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
+module_param(srpt_srq_size, int, 0444);
+MODULE_PARM_DESC(srpt_srq_size,
+                "Shared receive queue (SRQ) size.");
+
+static int srpt_get_u64_x(char *buffer, struct kernel_param *kp)
+{
+       return sprintf(buffer, "0x%016llx", *(u64 *)kp->arg);
+}
+module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid,
+                 0444);
+MODULE_PARM_DESC(srpt_service_guid,
+                "Using this value for ioc_guid, id_ext, and cm_listen_id"
+                " instead of using the node_guid of the first HCA.");
+
+static struct ib_client srpt_client;
+static struct target_fabric_configfs *srpt_target;
+static void srpt_release_channel(struct srpt_rdma_ch *ch);
+static int srpt_queue_status(struct se_cmd *cmd);
+
+/**
+ * opposite_dma_dir() - Swap DMA_TO_DEVICE and DMA_FROM_DEVICE.
+ */
+static inline
+enum dma_data_direction opposite_dma_dir(enum dma_data_direction dir)
+{
+       switch (dir) {
+       case DMA_TO_DEVICE:     return DMA_FROM_DEVICE;
+       case DMA_FROM_DEVICE:   return DMA_TO_DEVICE;
+       default:                return dir;
+       }
+}
+
+/**
+ * srpt_sdev_name() - Return the name associated with the HCA.
+ *
+ * Examples are ib0, ib1, ...
+ */
+static inline const char *srpt_sdev_name(struct srpt_device *sdev)
+{
+       return sdev->device->name;
+}
+
+static enum rdma_ch_state srpt_get_ch_state(struct srpt_rdma_ch *ch)
+{
+       unsigned long flags;
+       enum rdma_ch_state state;
+
+       spin_lock_irqsave(&ch->spinlock, flags);
+       state = ch->state;
+       spin_unlock_irqrestore(&ch->spinlock, flags);
+       return state;
+}
+
+static enum rdma_ch_state
+srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new_state)
+{
+       unsigned long flags;
+       enum rdma_ch_state prev;
+
+       spin_lock_irqsave(&ch->spinlock, flags);
+       prev = ch->state;
+       ch->state = new_state;
+       spin_unlock_irqrestore(&ch->spinlock, flags);
+       return prev;
+}
+
+/**
+ * srpt_test_and_set_ch_state() - Test and set the channel state.
+ *
+ * Returns true if and only if the channel state has been set to the new state.
+ */
+static bool
+srpt_test_and_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state old,
+                          enum rdma_ch_state new)
+{
+       unsigned long flags;
+       enum rdma_ch_state prev;
+
+       spin_lock_irqsave(&ch->spinlock, flags);
+       prev = ch->state;
+       if (prev == old)
+               ch->state = new;
+       spin_unlock_irqrestore(&ch->spinlock, flags);
+       return prev == old;
+}
+
+/**
+ * srpt_event_handler() - Asynchronous IB event callback function.
+ *
+ * Callback function called by the InfiniBand core when an asynchronous IB
+ * event occurs. This callback may occur in interrupt context. See also
+ * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
+ * Architecture Specification.
+ */
+static void srpt_event_handler(struct ib_event_handler *handler,
+                              struct ib_event *event)
+{
+       struct srpt_device *sdev;
+       struct srpt_port *sport;
+
+       sdev = ib_get_client_data(event->device, &srpt_client);
+       if (!sdev || sdev->device != event->device)
+               return;
+
+       pr_debug("ASYNC event= %d on device= %s\n", event->event,
+                srpt_sdev_name(sdev));
+
+       switch (event->event) {
+       case IB_EVENT_PORT_ERR:
+               if (event->element.port_num <= sdev->device->phys_port_cnt) {
+                       sport = &sdev->port[event->element.port_num - 1];
+                       sport->lid = 0;
+                       sport->sm_lid = 0;
+               }
+               break;
+       case IB_EVENT_PORT_ACTIVE:
+       case IB_EVENT_LID_CHANGE:
+       case IB_EVENT_PKEY_CHANGE:
+       case IB_EVENT_SM_CHANGE:
+       case IB_EVENT_CLIENT_REREGISTER:
+               /* Refresh port data asynchronously. */
+               if (event->element.port_num <= sdev->device->phys_port_cnt) {
+                       sport = &sdev->port[event->element.port_num - 1];
+                       if (!sport->lid && !sport->sm_lid)
+                               schedule_work(&sport->work);
+               }
+               break;
+       default:
+               printk(KERN_ERR "received unrecognized IB event %d\n",
+                      event->event);
+               break;
+       }
+}
+
+/**
+ * srpt_srq_event() - SRQ event callback function.
+ */
+static void srpt_srq_event(struct ib_event *event, void *ctx)
+{
+       printk(KERN_INFO "SRQ event %d\n", event->event);
+}
+
+/**
+ * srpt_qp_event() - QP event callback function.
+ */
+static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
+{
+       pr_debug("QP event %d on cm_id=%p sess_name=%s state=%d\n",
+                event->event, ch->cm_id, ch->sess_name, srpt_get_ch_state(ch));
+
+       switch (event->event) {
+       case IB_EVENT_COMM_EST:
+               ib_cm_notify(ch->cm_id, event->event);
+               break;
+       case IB_EVENT_QP_LAST_WQE_REACHED:
+               if (srpt_test_and_set_ch_state(ch, CH_DRAINING,
+                                              CH_RELEASING))
+                       srpt_release_channel(ch);
+               else
+                       pr_debug("%s: state %d - ignored LAST_WQE.\n",
+                                ch->sess_name, srpt_get_ch_state(ch));
+               break;
+       default:
+               printk(KERN_ERR "received unrecognized IB QP event %d\n",
+                      event->event);
+               break;
+       }
+}
+
+/**
+ * srpt_set_ioc() - Helper function for initializing an IOUnitInfo structure.
+ *
+ * @slot: one-based slot number.
+ * @value: four-bit value.
+ *
+ * Copies the lowest four bits of value in element slot of the array of four
+ * bit elements called c_list (controller list). The index slot is one-based.
+ */
+static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
+{
+       u16 id;
+       u8 tmp;
+
+       id = (slot - 1) / 2;
+       if (slot & 0x1) {
+               tmp = c_list[id] & 0xf;
+               c_list[id] = (value << 4) | tmp;
+       } else {
+               tmp = c_list[id] & 0xf0;
+               c_list[id] = (value & 0xf) | tmp;
+       }
+}
+
+/**
+ * srpt_get_class_port_info() - Copy ClassPortInfo to a management datagram.
+ *
+ * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
+ * Specification.
+ */
+static void srpt_get_class_port_info(struct ib_dm_mad *mad)
+{
+       struct ib_class_port_info *cif;
+
+       cif = (struct ib_class_port_info *)mad->data;
+       memset(cif, 0, sizeof *cif);
+       cif->base_version = 1;
+       cif->class_version = 1;
+       cif->resp_time_value = 20;
+
+       mad->mad_hdr.status = 0;
+}
+
+/**
+ * srpt_get_iou() - Write IOUnitInfo to a management datagram.
+ *
+ * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
+ * Specification. See also section B.7, table B.6 in the SRP r16a document.
+ */
+static void srpt_get_iou(struct ib_dm_mad *mad)
+{
+       struct ib_dm_iou_info *ioui;
+       u8 slot;
+       int i;
+
+       ioui = (struct ib_dm_iou_info *)mad->data;
+       ioui->change_id = __constant_cpu_to_be16(1);
+       ioui->max_controllers = 16;
+
+       /* set present for slot 1 and empty for the rest */
+       srpt_set_ioc(ioui->controller_list, 1, 1);
+       for (i = 1, slot = 2; i < 16; i++, slot++)
+               srpt_set_ioc(ioui->controller_list, slot, 0);
+
+       mad->mad_hdr.status = 0;
+}
+
+/**
+ * srpt_get_ioc() - Write IOControllerprofile to a management datagram.
+ *
+ * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
+ * Architecture Specification. See also section B.7, table B.7 in the SRP
+ * r16a document.
+ */
+static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
+                        struct ib_dm_mad *mad)
+{
+       struct srpt_device *sdev = sport->sdev;
+       struct ib_dm_ioc_profile *iocp;
+
+       iocp = (struct ib_dm_ioc_profile *)mad->data;
+
+       if (!slot || slot > 16) {
+               mad->mad_hdr.status
+                       = __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
+               return;
+       }
+
+       if (slot > 2) {
+               mad->mad_hdr.status
+                       = __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
+               return;
+       }
+
+       memset(iocp, 0, sizeof *iocp);
+       strcpy(iocp->id_string, SRPT_ID_STRING);
+       iocp->guid = cpu_to_be64(srpt_service_guid);
+       iocp->vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
+       iocp->device_id = cpu_to_be32(sdev->dev_attr.vendor_part_id);
+       iocp->device_version = cpu_to_be16(sdev->dev_attr.hw_ver);
+       iocp->subsys_vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
+       iocp->subsys_device_id = 0x0;
+       iocp->io_class = __constant_cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
+       iocp->io_subclass = __constant_cpu_to_be16(SRP_IO_SUBCLASS);
+       iocp->protocol = __constant_cpu_to_be16(SRP_PROTOCOL);
+       iocp->protocol_version = __constant_cpu_to_be16(SRP_PROTOCOL_VERSION);
+       iocp->send_queue_depth = cpu_to_be16(sdev->srq_size);
+       iocp->rdma_read_depth = 4;
+       iocp->send_size = cpu_to_be32(srp_max_req_size);
+       iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size,
+                                         1U << 24));
+       iocp->num_svc_entries = 1;
+       iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC |
+               SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC;
+
+       mad->mad_hdr.status = 0;
+}
+
+/**
+ * srpt_get_svc_entries() - Write ServiceEntries to a management datagram.
+ *
+ * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
+ * Specification. See also section B.7, table B.8 in the SRP r16a document.
+ */
+static void srpt_get_svc_entries(u64 ioc_guid,
+                                u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
+{
+       struct ib_dm_svc_entries *svc_entries;
+
+       WARN_ON(!ioc_guid);
+
+       if (!slot || slot > 16) {
+               mad->mad_hdr.status
+                       = __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
+               return;
+       }
+
+       if (slot > 2 || lo > hi || hi > 1) {
+               mad->mad_hdr.status
+                       = __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
+               return;
+       }
+
+       svc_entries = (struct ib_dm_svc_entries *)mad->data;
+       memset(svc_entries, 0, sizeof *svc_entries);
+       svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
+       snprintf(svc_entries->service_entries[0].name,
+                sizeof(svc_entries->service_entries[0].name),
+                "%s%016llx",
+                SRP_SERVICE_NAME_PREFIX,
+                ioc_guid);
+
+       mad->mad_hdr.status = 0;
+}
+
+/**
+ * srpt_mgmt_method_get() - Process a received management datagram.
+ * @sp:      source port through which the MAD has been received.
+ * @rq_mad:  received MAD.
+ * @rsp_mad: response MAD.
+ */
+static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
+                                struct ib_dm_mad *rsp_mad)
+{
+       u16 attr_id;
+       u32 slot;
+       u8 hi, lo;
+
+       attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
+       switch (attr_id) {
+       case DM_ATTR_CLASS_PORT_INFO:
+               srpt_get_class_port_info(rsp_mad);
+               break;
+       case DM_ATTR_IOU_INFO:
+               srpt_get_iou(rsp_mad);
+               break;
+       case DM_ATTR_IOC_PROFILE:
+               slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
+               srpt_get_ioc(sp, slot, rsp_mad);
+               break;
+       case DM_ATTR_SVC_ENTRIES:
+               slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
+               hi = (u8) ((slot >> 8) & 0xff);
+               lo = (u8) (slot & 0xff);
+               slot = (u16) ((slot >> 16) & 0xffff);
+               srpt_get_svc_entries(srpt_service_guid,
+                                    slot, hi, lo, rsp_mad);
+               break;
+       default:
+               rsp_mad->mad_hdr.status =
+                   __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
+               break;
+       }
+}
+
+/**
+ * srpt_mad_send_handler() - Post MAD-send callback function.
+ */
+static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
+                                 struct ib_mad_send_wc *mad_wc)
+{
+       ib_destroy_ah(mad_wc->send_buf->ah);
+       ib_free_send_mad(mad_wc->send_buf);
+}
+
+/**
+ * srpt_mad_recv_handler() - MAD reception callback function.
+ */
+static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
+                                 struct ib_mad_recv_wc *mad_wc)
+{
+       struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
+       struct ib_ah *ah;
+       struct ib_mad_send_buf *rsp;
+       struct ib_dm_mad *dm_mad;
+
+       if (!mad_wc || !mad_wc->recv_buf.mad)
+               return;
+
+       ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
+                                 mad_wc->recv_buf.grh, mad_agent->port_num);
+       if (IS_ERR(ah))
+               goto err;
+
+       BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
+
+       rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
+                                mad_wc->wc->pkey_index, 0,
+                                IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
+                                GFP_KERNEL);
+       if (IS_ERR(rsp))
+               goto err_rsp;
+
+       rsp->ah = ah;
+
+       dm_mad = rsp->mad;
+       memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof *dm_mad);
+       dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
+       dm_mad->mad_hdr.status = 0;
+
+       switch (mad_wc->recv_buf.mad->mad_hdr.method) {
+       case IB_MGMT_METHOD_GET:
+               srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
+               break;
+       case IB_MGMT_METHOD_SET:
+               dm_mad->mad_hdr.status =
+                   __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
+               break;
+       default:
+               dm_mad->mad_hdr.status =
+                   __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
+               break;
+       }
+
+       if (!ib_post_send_mad(rsp, NULL)) {
+               ib_free_recv_mad(mad_wc);
+               /* will destroy_ah & free_send_mad in send completion */
+               return;
+       }
+
+       ib_free_send_mad(rsp);
+
+err_rsp:
+       ib_destroy_ah(ah);
+err:
+       ib_free_recv_mad(mad_wc);
+}
+
+/**
+ * srpt_refresh_port() - Configure a HCA port.
+ *
+ * Enable InfiniBand management datagram processing, update the cached sm_lid,
+ * lid and gid values, and register a callback function for processing MADs
+ * on the specified port.
+ *
+ * Note: It is safe to call this function more than once for the same port.
+ */
+static int srpt_refresh_port(struct srpt_port *sport)
+{
+       struct ib_mad_reg_req reg_req;
+       struct ib_port_modify port_modify;
+       struct ib_port_attr port_attr;
+       int ret;
+
+       memset(&port_modify, 0, sizeof port_modify);
+       port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
+       port_modify.clr_port_cap_mask = 0;
+
+       ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
+       if (ret)
+               goto err_mod_port;
+
+       ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
+       if (ret)
+               goto err_query_port;
+
+       sport->sm_lid = port_attr.sm_lid;
+       sport->lid = port_attr.lid;
+
+       ret = ib_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
+       if (ret)
+               goto err_query_port;
+
+       if (!sport->mad_agent) {
+               memset(&reg_req, 0, sizeof reg_req);
+               reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
+               reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
+               set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
+               set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
+
+               sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
+                                                        sport->port,
+                                                        IB_QPT_GSI,
+                                                        &reg_req, 0,
+                                                        srpt_mad_send_handler,
+                                                        srpt_mad_recv_handler,
+                                                        sport);
+               if (IS_ERR(sport->mad_agent)) {
+                       ret = PTR_ERR(sport->mad_agent);
+                       sport->mad_agent = NULL;
+                       goto err_query_port;
+               }
+       }
+
+       return 0;
+
+err_query_port:
+
+       port_modify.set_port_cap_mask = 0;
+       port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
+       ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
+
+err_mod_port:
+
+       return ret;
+}
+
+/**
+ * srpt_unregister_mad_agent() - Unregister MAD callback functions.
+ *
+ * Note: It is safe to call this function more than once for the same device.
+ */
+static void srpt_unregister_mad_agent(struct srpt_device *sdev)
+{
+       struct ib_port_modify port_modify = {
+               .clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
+       };
+       struct srpt_port *sport;
+       int i;
+
+       for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
+               sport = &sdev->port[i - 1];
+               WARN_ON(sport->port != i);
+               if (ib_modify_port(sdev->device, i, 0, &port_modify) < 0)
+                       printk(KERN_ERR "disabling MAD processing failed.\n");
+               if (sport->mad_agent) {
+                       ib_unregister_mad_agent(sport->mad_agent);
+                       sport->mad_agent = NULL;
+               }
+       }
+}
+
+/**
+ * srpt_alloc_ioctx() - Allocate an SRPT I/O context structure.
+ */
+static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
+                                          int ioctx_size, int dma_size,
+                                          enum dma_data_direction dir)
+{
+       struct srpt_ioctx *ioctx;
+
+       ioctx = kmalloc(ioctx_size, GFP_KERNEL);
+       if (!ioctx)
+               goto err;
+
+       ioctx->buf = kmalloc(dma_size, GFP_KERNEL);
+       if (!ioctx->buf)
+               goto err_free_ioctx;
+
+       ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, dma_size, dir);
+       if (ib_dma_mapping_error(sdev->device, ioctx->dma))
+               goto err_free_buf;
+
+       return ioctx;
+
+err_free_buf:
+       kfree(ioctx->buf);
+err_free_ioctx:
+       kfree(ioctx);
+err:
+       return NULL;
+}
+
+/**
+ * srpt_free_ioctx() - Free an SRPT I/O context structure.
+ */
+static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
+                           int dma_size, enum dma_data_direction dir)
+{
+       if (!ioctx)
+               return;
+
+       ib_dma_unmap_single(sdev->device, ioctx->dma, dma_size, dir);
+       kfree(ioctx->buf);
+       kfree(ioctx);
+}
+
+/**
+ * srpt_alloc_ioctx_ring() - Allocate a ring of SRPT I/O context structures.
+ * @sdev:       Device to allocate the I/O context ring for.
+ * @ring_size:  Number of elements in the I/O context ring.
+ * @ioctx_size: I/O context size.
+ * @dma_size:   DMA buffer size.
+ * @dir:        DMA data direction.
+ */
+static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
+                               int ring_size, int ioctx_size,
+                               int dma_size, enum dma_data_direction dir)
+{
+       struct srpt_ioctx **ring;
+       int i;
+
+       WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx)
+               && ioctx_size != sizeof(struct srpt_send_ioctx));
+
+       ring = kmalloc(ring_size * sizeof(ring[0]), GFP_KERNEL);
+       if (!ring)
+               goto out;
+       for (i = 0; i < ring_size; ++i) {
+               ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, dma_size, dir);
+               if (!ring[i])
+                       goto err;
+               ring[i]->index = i;
+       }
+       goto out;
+
+err:
+       while (--i >= 0)
+               srpt_free_ioctx(sdev, ring[i], dma_size, dir);
+       kfree(ring);
+out:
+       return ring;
+}
+
+/**
+ * srpt_free_ioctx_ring() - Free the ring of SRPT I/O context structures.
+ */
+static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
+                                struct srpt_device *sdev, int ring_size,
+                                int dma_size, enum dma_data_direction dir)
+{
+       int i;
+
+       for (i = 0; i < ring_size; ++i)
+               srpt_free_ioctx(sdev, ioctx_ring[i], dma_size, dir);
+       kfree(ioctx_ring);
+}
+
+/**
+ * srpt_get_cmd_state() - Get the state of a SCSI command.
+ */
+static enum srpt_command_state srpt_get_cmd_state(struct srpt_send_ioctx *ioctx)
+{
+       enum srpt_command_state state;
+       unsigned long flags;
+
+       BUG_ON(!ioctx);
+
+       spin_lock_irqsave(&ioctx->spinlock, flags);
+       state = ioctx->state;
+       spin_unlock_irqrestore(&ioctx->spinlock, flags);
+       return state;
+}
+
+/**
+ * srpt_set_cmd_state() - Set the state of a SCSI command.
+ *
+ * Does not modify the state of aborted commands. Returns the previous command
+ * state.
+ */
+static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
+                                                 enum srpt_command_state new)
+{
+       enum srpt_command_state previous;
+       unsigned long flags;
+
+       BUG_ON(!ioctx);
+
+       spin_lock_irqsave(&ioctx->spinlock, flags);
+       previous = ioctx->state;
+       if (previous != SRPT_STATE_DONE)
+               ioctx->state = new;
+       spin_unlock_irqrestore(&ioctx->spinlock, flags);
+
+       return previous;
+}
+
+/**
+ * srpt_test_and_set_cmd_state() - Test and set the state of a command.
+ *
+ * Returns true if and only if the previous command state was equal to 'old'.
+ */
+static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
+                                       enum srpt_command_state old,
+                                       enum srpt_command_state new)
+{
+       enum srpt_command_state previous;
+       unsigned long flags;
+
+       WARN_ON(!ioctx);
+       WARN_ON(old == SRPT_STATE_DONE);
+       WARN_ON(new == SRPT_STATE_NEW);
+
+       spin_lock_irqsave(&ioctx->spinlock, flags);
+       previous = ioctx->state;
+       if (previous == old)
+               ioctx->state = new;
+       spin_unlock_irqrestore(&ioctx->spinlock, flags);
+       return previous == old;
+}
+
+/**
+ * srpt_post_recv() - Post an IB receive request.
+ */
+static int srpt_post_recv(struct srpt_device *sdev,
+                         struct srpt_recv_ioctx *ioctx)
+{
+       struct ib_sge list;
+       struct ib_recv_wr wr, *bad_wr;
+
+       BUG_ON(!sdev);
+       wr.wr_id = encode_wr_id(SRPT_RECV, ioctx->ioctx.index);
+
+       list.addr = ioctx->ioctx.dma;
+       list.length = srp_max_req_size;
+       list.lkey = sdev->mr->lkey;
+
+       wr.next = NULL;
+       wr.sg_list = &list;
+       wr.num_sge = 1;
+
+       return ib_post_srq_recv(sdev->srq, &wr, &bad_wr);
+}
+
+/**
+ * srpt_post_send() - Post an IB send request.
+ *
+ * Returns zero upon success and a non-zero value upon failure.
+ */
+static int srpt_post_send(struct srpt_rdma_ch *ch,
+                         struct srpt_send_ioctx *ioctx, int len)
+{
+       struct ib_sge list;
+       struct ib_send_wr wr, *bad_wr;
+       struct srpt_device *sdev = ch->sport->sdev;
+       int ret;
+
+       atomic_inc(&ch->req_lim);
+
+       ret = -ENOMEM;
+       if (unlikely(atomic_dec_return(&ch->sq_wr_avail) < 0)) {
+               printk(KERN_WARNING "IB send queue full (needed 1)\n");
+               goto out;
+       }
+
+       ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, len,
+                                     DMA_TO_DEVICE);
+
+       list.addr = ioctx->ioctx.dma;
+       list.length = len;
+       list.lkey = sdev->mr->lkey;
+
+       wr.next = NULL;
+       wr.wr_id = encode_wr_id(SRPT_SEND, ioctx->ioctx.index);
+       wr.sg_list = &list;
+       wr.num_sge = 1;
+       wr.opcode = IB_WR_SEND;
+       wr.send_flags = IB_SEND_SIGNALED;
+
+       ret = ib_post_send(ch->qp, &wr, &bad_wr);
+
+out:
+       if (ret < 0) {
+               atomic_inc(&ch->sq_wr_avail);
+               atomic_dec(&ch->req_lim);
+       }
+       return ret;
+}
+
+/**
+ * srpt_get_desc_tbl() - Parse the data descriptors of an SRP_CMD request.
+ * @ioctx: Pointer to the I/O context associated with the request.
+ * @srp_cmd: Pointer to the SRP_CMD request data.
+ * @dir: Pointer to the variable to which the transfer direction will be
+ *   written.
+ * @data_len: Pointer to the variable to which the total data length of all
+ *   descriptors in the SRP_CMD request will be written.
+ *
+ * This function initializes ioctx->nrbuf and ioctx->r_bufs.
+ *
+ * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
+ * -ENOMEM when memory allocation fails and zero upon success.
+ */
+static int srpt_get_desc_tbl(struct srpt_send_ioctx *ioctx,
+                            struct srp_cmd *srp_cmd,
+                            enum dma_data_direction *dir, u64 *data_len)
+{
+       struct srp_indirect_buf *idb;
+       struct srp_direct_buf *db;
+       unsigned add_cdb_offset;
+       int ret;
+
+       /*
+        * The pointer computations below will only be compiled correctly
+        * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
+        * whether srp_cmd::add_data has been declared as a byte pointer.
+        */
+       BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0)
+                    && !__same_type(srp_cmd->add_data[0], (u8)0));
+
+       BUG_ON(!dir);
+       BUG_ON(!data_len);
+
+       ret = 0;
+       *data_len = 0;
+
+       /*
+        * The lower four bits of the buffer format field contain the DATA-IN
+        * buffer descriptor format, and the highest four bits contain the
+        * DATA-OUT buffer descriptor format.
+        */
+       *dir = DMA_NONE;
+       if (srp_cmd->buf_fmt & 0xf)
+               /* DATA-IN: transfer data from target to initiator (read). */
+               *dir = DMA_FROM_DEVICE;
+       else if (srp_cmd->buf_fmt >> 4)
+               /* DATA-OUT: transfer data from initiator to target (write). */
+               *dir = DMA_TO_DEVICE;
+
+       /*
+        * According to the SRP spec, the lower two bits of the 'ADDITIONAL
+        * CDB LENGTH' field are reserved and the size in bytes of this field
+        * is four times the value specified in bits 3..7. Hence the "& ~3".
+        */
+       add_cdb_offset = srp_cmd->add_cdb_len & ~3;
+       if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
+           ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
+               ioctx->n_rbuf = 1;
+               ioctx->rbufs = &ioctx->single_rbuf;
+
+               db = (struct srp_direct_buf *)(srp_cmd->add_data
+                                              + add_cdb_offset);
+               memcpy(ioctx->rbufs, db, sizeof *db);
+               *data_len = be32_to_cpu(db->len);
+       } else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
+                  ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
+               idb = (struct srp_indirect_buf *)(srp_cmd->add_data
+                                                 + add_cdb_offset);
+
+               ioctx->n_rbuf = be32_to_cpu(idb->table_desc.len) / sizeof *db;
+
+               if (ioctx->n_rbuf >
+                   (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
+                       printk(KERN_ERR "received unsupported SRP_CMD request"
+                              " type (%u out + %u in != %u / %zu)\n",
+                              srp_cmd->data_out_desc_cnt,
+                              srp_cmd->data_in_desc_cnt,
+                              be32_to_cpu(idb->table_desc.len),
+                              sizeof(*db));
+                       ioctx->n_rbuf = 0;
+                       ret = -EINVAL;
+                       goto out;
+               }
+
+               if (ioctx->n_rbuf == 1)
+                       ioctx->rbufs = &ioctx->single_rbuf;
+               else {
+                       ioctx->rbufs =
+                               kmalloc(ioctx->n_rbuf * sizeof *db, GFP_ATOMIC);
+                       if (!ioctx->rbufs) {
+                               ioctx->n_rbuf = 0;
+                               ret = -ENOMEM;
+                               goto out;
+                       }
+               }
+
+               db = idb->desc_list;
+               memcpy(ioctx->rbufs, db, ioctx->n_rbuf * sizeof *db);
+               *data_len = be32_to_cpu(idb->len);
+       }
+out:
+       return ret;
+}
+
+/**
+ * srpt_init_ch_qp() - Initialize queue pair attributes.
+ *
+ * Initialized the attributes of queue pair 'qp' by allowing local write,
+ * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
+ */
+static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
+{
+       struct ib_qp_attr *attr;
+       int ret;
+
+       attr = kzalloc(sizeof *attr, GFP_KERNEL);
+       if (!attr)
+               return -ENOMEM;
+
+       attr->qp_state = IB_QPS_INIT;
+       attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
+           IB_ACCESS_REMOTE_WRITE;
+       attr->port_num = ch->sport->port;
+       attr->pkey_index = 0;
+
+       ret = ib_modify_qp(qp, attr,
+                          IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
+                          IB_QP_PKEY_INDEX);
+
+       kfree(attr);
+       return ret;
+}
+
+/**
+ * srpt_ch_qp_rtr() - Change the state of a channel to 'ready to receive' (RTR).
+ * @ch: channel of the queue pair.
+ * @qp: queue pair to change the state of.
+ *
+ * Returns zero upon success and a negative value upon failure.
+ *
+ * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
+ * If this structure ever becomes larger, it might be necessary to allocate
+ * it dynamically instead of on the stack.
+ */
+static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
+{
+       struct ib_qp_attr qp_attr;
+       int attr_mask;
+       int ret;
+
+       qp_attr.qp_state = IB_QPS_RTR;
+       ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
+       if (ret)
+               goto out;
+
+       qp_attr.max_dest_rd_atomic = 4;
+
+       ret = ib_modify_qp(qp, &qp_attr, attr_mask);
+
+out:
+       return ret;
+}
+
+/**
+ * srpt_ch_qp_rts() - Change the state of a channel to 'ready to send' (RTS).
+ * @ch: channel of the queue pair.
+ * @qp: queue pair to change the state of.
+ *
+ * Returns zero upon success and a negative value upon failure.
+ *
+ * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
+ * If this structure ever becomes larger, it might be necessary to allocate
+ * it dynamically instead of on the stack.
+ */
+static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
+{
+       struct ib_qp_attr qp_attr;
+       int attr_mask;
+       int ret;
+
+       qp_attr.qp_state = IB_QPS_RTS;
+       ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
+       if (ret)
+               goto out;
+
+       qp_attr.max_rd_atomic = 4;
+
+       ret = ib_modify_qp(qp, &qp_attr, attr_mask);
+
+out:
+       return ret;
+}
+
+/**
+ * srpt_ch_qp_err() - Set the channel queue pair state to 'error'.
+ */
+static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
+{
+       struct ib_qp_attr qp_attr;
+
+       qp_attr.qp_state = IB_QPS_ERR;
+       return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
+}
+
+/**
+ * srpt_unmap_sg_to_ib_sge() - Unmap an IB SGE list.
+ */
+static void srpt_unmap_sg_to_ib_sge(struct srpt_rdma_ch *ch,
+                                   struct srpt_send_ioctx *ioctx)
+{
+       struct scatterlist *sg;
+       enum dma_data_direction dir;
+
+       BUG_ON(!ch);
+       BUG_ON(!ioctx);
+       BUG_ON(ioctx->n_rdma && !ioctx->rdma_ius);
+
+       while (ioctx->n_rdma)
+               kfree(ioctx->rdma_ius[--ioctx->n_rdma].sge);
+
+       kfree(ioctx->rdma_ius);
+       ioctx->rdma_ius = NULL;
+
+       if (ioctx->mapped_sg_count) {
+               sg = ioctx->sg;
+               WARN_ON(!sg);
+               dir = ioctx->cmd.data_direction;
+               BUG_ON(dir == DMA_NONE);
+               ib_dma_unmap_sg(ch->sport->sdev->device, sg, ioctx->sg_cnt,
+                               opposite_dma_dir(dir));
+               ioctx->mapped_sg_count = 0;
+       }
+}
+
+/**
+ * srpt_map_sg_to_ib_sge() - Map an SG list to an IB SGE list.
+ */
+static int srpt_map_sg_to_ib_sge(struct srpt_rdma_ch *ch,
+                                struct srpt_send_ioctx *ioctx)
+{
+       struct se_cmd *cmd;
+       struct scatterlist *sg, *sg_orig;
+       int sg_cnt;
+       enum dma_data_direction dir;
+       struct rdma_iu *riu;
+       struct srp_direct_buf *db;
+       dma_addr_t dma_addr;
+       struct ib_sge *sge;
+       u64 raddr;
+       u32 rsize;
+       u32 tsize;
+       u32 dma_len;
+       int count, nrdma;
+       int i, j, k;
+
+       BUG_ON(!ch);
+       BUG_ON(!ioctx);
+       cmd = &ioctx->cmd;
+       dir = cmd->data_direction;
+       BUG_ON(dir == DMA_NONE);
+
+       transport_do_task_sg_chain(cmd);
+       ioctx->sg = sg = sg_orig = cmd->t_tasks_sg_chained;
+       ioctx->sg_cnt = sg_cnt = cmd->t_tasks_sg_chained_no;
+
+       count = ib_dma_map_sg(ch->sport->sdev->device, sg, sg_cnt,
+                             opposite_dma_dir(dir));
+       if (unlikely(!count))
+               return -EAGAIN;
+
+       ioctx->mapped_sg_count = count;
+
+       if (ioctx->rdma_ius && ioctx->n_rdma_ius)
+               nrdma = ioctx->n_rdma_ius;
+       else {
+               nrdma = (count + SRPT_DEF_SG_PER_WQE - 1) / SRPT_DEF_SG_PER_WQE
+                       + ioctx->n_rbuf;
+
+               ioctx->rdma_ius = kzalloc(nrdma * sizeof *riu, GFP_KERNEL);
+               if (!ioctx->rdma_ius)
+                       goto free_mem;
+
+               ioctx->n_rdma_ius = nrdma;
+       }
+
+       db = ioctx->rbufs;
+       tsize = cmd->data_length;
+       dma_len = sg_dma_len(&sg[0]);
+       riu = ioctx->rdma_ius;
+
+       /*
+        * For each remote desc - calculate the #ib_sge.
+        * If #ib_sge < SRPT_DEF_SG_PER_WQE per rdma operation then
+        *      each remote desc rdma_iu is required a rdma wr;
+        * else
+        *      we need to allocate extra rdma_iu to carry extra #ib_sge in
+        *      another rdma wr
+        */
+       for (i = 0, j = 0;
+            j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
+               rsize = be32_to_cpu(db->len);
+               raddr = be64_to_cpu(db->va);
+               riu->raddr = raddr;
+               riu->rkey = be32_to_cpu(db->key);
+               riu->sge_cnt = 0;
+
+               /* calculate how many sge required for this remote_buf */
+               while (rsize > 0 && tsize > 0) {
+
+                       if (rsize >= dma_len) {
+                               tsize -= dma_len;
+                               rsize -= dma_len;
+                               raddr += dma_len;
+
+                               if (tsize > 0) {
+                                       ++j;
+                                       if (j < count) {
+                                               sg = sg_next(sg);
+                                               dma_len = sg_dma_len(sg);
+                                       }
+                               }
+                       } else {
+                               tsize -= rsize;
+                               dma_len -= rsize;
+                               rsize = 0;
+                       }
+
+                       ++riu->sge_cnt;
+
+                       if (rsize > 0 && riu->sge_cnt == SRPT_DEF_SG_PER_WQE) {
+                               ++ioctx->n_rdma;
+                               riu->sge =
+                                   kmalloc(riu->sge_cnt * sizeof *riu->sge,
+                                           GFP_KERNEL);
+                               if (!riu->sge)
+                                       goto free_mem;
+
+                               ++riu;
+                               riu->sge_cnt = 0;
+                               riu->raddr = raddr;
+                               riu->rkey = be32_to_cpu(db->key);
+                       }
+               }
+
+               ++ioctx->n_rdma;
+               riu->sge = kmalloc(riu->sge_cnt * sizeof *riu->sge,
+                                  GFP_KERNEL);
+               if (!riu->sge)
+                       goto free_mem;
+       }
+
+       db = ioctx->rbufs;
+       tsize = cmd->data_length;
+       riu = ioctx->rdma_ius;
+       sg = sg_orig;
+       dma_len = sg_dma_len(&sg[0]);
+       dma_addr = sg_dma_address(&sg[0]);
+
+       /* this second loop is really mapped sg_addres to rdma_iu->ib_sge */
+       for (i = 0, j = 0;
+            j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
+               rsize = be32_to_cpu(db->len);
+               sge = riu->sge;
+               k = 0;
+
+               while (rsize > 0 && tsize > 0) {
+                       sge->addr = dma_addr;
+                       sge->lkey = ch->sport->sdev->mr->lkey;
+
+                       if (rsize >= dma_len) {
+                               sge->length =
+                                       (tsize < dma_len) ? tsize : dma_len;
+                               tsize -= dma_len;
+                               rsize -= dma_len;
+
+                               if (tsize > 0) {
+                                       ++j;
+                                       if (j < count) {
+                                               sg = sg_next(sg);
+                                               dma_len = sg_dma_len(sg);
+                                               dma_addr = sg_dma_address(sg);
+                                       }
+                               }
+                       } else {
+                               sge->length = (tsize < rsize) ? tsize : rsize;
+                               tsize -= rsize;
+                               dma_len -= rsize;
+                               dma_addr += rsize;
+                               rsize = 0;
+                       }
+
+                       ++k;
+                       if (k == riu->sge_cnt && rsize > 0 && tsize > 0) {
+                               ++riu;
+                               sge = riu->sge;
+                               k = 0;
+                       } else if (rsize > 0 && tsize > 0)
+                               ++sge;
+               }
+       }
+
+       return 0;
+
+free_mem:
+       srpt_unmap_sg_to_ib_sge(ch, ioctx);
+
+       return -ENOMEM;
+}
+
+/**
+ * srpt_get_send_ioctx() - Obtain an I/O context for sending to the initiator.
+ */
+static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
+{
+       struct srpt_send_ioctx *ioctx;
+       unsigned long flags;
+
+       BUG_ON(!ch);
+
+       ioctx = NULL;
+       spin_lock_irqsave(&ch->spinlock, flags);
+       if (!list_empty(&ch->free_list)) {
+               ioctx = list_first_entry(&ch->free_list,
+                                        struct srpt_send_ioctx, free_list);
+               list_del(&ioctx->free_list);
+       }
+       spin_unlock_irqrestore(&ch->spinlock, flags);
+
+       if (!ioctx)
+               return ioctx;
+
+       BUG_ON(ioctx->ch != ch);
+       kref_init(&ioctx->kref);
+       spin_lock_init(&ioctx->spinlock);
+       ioctx->state = SRPT_STATE_NEW;
+       ioctx->n_rbuf = 0;
+       ioctx->rbufs = NULL;
+       ioctx->n_rdma = 0;
+       ioctx->n_rdma_ius = 0;
+       ioctx->rdma_ius = NULL;
+       ioctx->mapped_sg_count = 0;
+       init_completion(&ioctx->tx_done);
+       ioctx->queue_status_only = false;
+       /*
+        * transport_init_se_cmd() does not initialize all fields, so do it
+        * here.
+        */
+       memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
+       memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
+
+       return ioctx;
+}
+
+/**
+ * srpt_put_send_ioctx() - Free up resources.
+ */
+static void srpt_put_send_ioctx(struct srpt_send_ioctx *ioctx)
+{
+       struct srpt_rdma_ch *ch;
+       unsigned long flags;
+
+       BUG_ON(!ioctx);
+       ch = ioctx->ch;
+       BUG_ON(!ch);
+
+       WARN_ON(srpt_get_cmd_state(ioctx) != SRPT_STATE_DONE);
+
+       srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
+       transport_generic_free_cmd(&ioctx->cmd, 0);
+
+       if (ioctx->n_rbuf > 1) {
+               kfree(ioctx->rbufs);
+               ioctx->rbufs = NULL;
+               ioctx->n_rbuf = 0;
+       }
+
+       spin_lock_irqsave(&ch->spinlock, flags);
+       list_add(&ioctx->free_list, &ch->free_list);
+       spin_unlock_irqrestore(&ch->spinlock, flags);
+}
+
+static void srpt_put_send_ioctx_kref(struct kref *kref)
+{
+       srpt_put_send_ioctx(container_of(kref, struct srpt_send_ioctx, kref));
+}
+
+/**
+ * srpt_abort_cmd() - Abort a SCSI command.
+ * @ioctx:   I/O context associated with the SCSI command.
+ * @context: Preferred execution context.
+ */
+static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
+{
+       enum srpt_command_state state;
+       unsigned long flags;
+
+       BUG_ON(!ioctx);
+
+       /*
+        * If the command is in a state where the target core is waiting for
+        * the ib_srpt driver, change the state to the next state. Changing
+        * the state of the command from SRPT_STATE_NEED_DATA to
+        * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
+        * function a second time.
+        */
+
+       spin_lock_irqsave(&ioctx->spinlock, flags);
+       state = ioctx->state;
+       switch (state) {
+       case SRPT_STATE_NEED_DATA:
+               ioctx->state = SRPT_STATE_DATA_IN;
+               break;
+       case SRPT_STATE_DATA_IN:
+       case SRPT_STATE_CMD_RSP_SENT:
+       case SRPT_STATE_MGMT_RSP_SENT:
+               ioctx->state = SRPT_STATE_DONE;
+               break;
+       default:
+               break;
+       }
+       spin_unlock_irqrestore(&ioctx->spinlock, flags);
+
+       if (state == SRPT_STATE_DONE)
+               goto out;
+
+       pr_debug("Aborting cmd with state %d and tag %lld\n", state,
+                ioctx->tag);
+
+       switch (state) {
+       case SRPT_STATE_NEW:
+       case SRPT_STATE_DATA_IN:
+       case SRPT_STATE_MGMT:
+               /*
+                * Do nothing - defer abort processing until
+                * srpt_queue_response() is invoked.
+                */
+               WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
+               break;
+       case SRPT_STATE_NEED_DATA:
+               /* DMA_TO_DEVICE (write) - RDMA read error. */
+               atomic_set(&ioctx->cmd.transport_lun_stop, 1);
+               transport_generic_handle_data(&ioctx->cmd);
+               break;
+       case SRPT_STATE_CMD_RSP_SENT:
+               /*
+                * SRP_RSP sending failed or the SRP_RSP send completion has
+                * not been received in time.
+                */
+               srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
+               atomic_set(&ioctx->cmd.transport_lun_stop, 1);
+               kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
+               break;
+       case SRPT_STATE_MGMT_RSP_SENT:
+               srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
+               kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
+               break;
+       default:
+               WARN_ON("ERROR: unexpected command state");
+               break;
+       }
+
+out:
+       return state;
+}
+
+/**
+ * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
+ */
+static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
+{
+       struct srpt_send_ioctx *ioctx;
+       enum srpt_command_state state;
+       struct se_cmd *cmd;
+       u32 index;
+
+       atomic_inc(&ch->sq_wr_avail);
+
+       index = idx_from_wr_id(wr_id);
+       ioctx = ch->ioctx_ring[index];
+       state = srpt_get_cmd_state(ioctx);
+       cmd = &ioctx->cmd;
+
+       WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
+               && state != SRPT_STATE_MGMT_RSP_SENT
+               && state != SRPT_STATE_NEED_DATA
+               && state != SRPT_STATE_DONE);
+
+       /* If SRP_RSP sending failed, undo the ch->req_lim change. */
+       if (state == SRPT_STATE_CMD_RSP_SENT
+           || state == SRPT_STATE_MGMT_RSP_SENT)
+               atomic_dec(&ch->req_lim);
+
+       srpt_abort_cmd(ioctx);
+}
+
+/**
+ * srpt_handle_send_comp() - Process an IB send completion notification.
+ */
+static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
+                                 struct srpt_send_ioctx *ioctx)
+{
+       enum srpt_command_state state;
+
+       atomic_inc(&ch->sq_wr_avail);
+
+       state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
+
+       if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
+                   && state != SRPT_STATE_MGMT_RSP_SENT
+                   && state != SRPT_STATE_DONE))
+               pr_debug("state = %d\n", state);
+
+       if (state != SRPT_STATE_DONE)
+               kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
+       else
+               printk(KERN_ERR "IB completion has been received too late for"
+                      " wr_id = %u.\n", ioctx->ioctx.index);
+}
+
+/**
+ * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
+ *
+ * Note: transport_generic_handle_data() is asynchronous so unmapping the
+ * data that has been transferred via IB RDMA must be postponed until the
+ * check_stop_free() callback.
+ */
+static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
+                                 struct srpt_send_ioctx *ioctx,
+                                 enum srpt_opcode opcode)
+{
+       WARN_ON(ioctx->n_rdma <= 0);
+       atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
+
+       if (opcode == SRPT_RDMA_READ_LAST) {
+               if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
+                                               SRPT_STATE_DATA_IN))
+                       transport_generic_handle_data(&ioctx->cmd);
+               else
+                       printk(KERN_ERR "%s[%d]: wrong state = %d\n", __func__,
+                              __LINE__, srpt_get_cmd_state(ioctx));
+       } else if (opcode == SRPT_RDMA_ABORT) {
+               ioctx->rdma_aborted = true;
+       } else {
+               WARN(true, "unexpected opcode %d\n", opcode);
+       }
+}
+
+/**
+ * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
+ */
+static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
+                                     struct srpt_send_ioctx *ioctx,
+                                     enum srpt_opcode opcode)
+{
+       struct se_cmd *cmd;
+       enum srpt_command_state state;
+
+       cmd = &ioctx->cmd;
+       state = srpt_get_cmd_state(ioctx);
+       switch (opcode) {
+       case SRPT_RDMA_READ_LAST:
+               if (ioctx->n_rdma <= 0) {
+                       printk(KERN_ERR "Received invalid RDMA read"
+                              " error completion with idx %d\n",
+                              ioctx->ioctx.index);
+                       break;
+               }
+               atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
+               if (state == SRPT_STATE_NEED_DATA)
+                       srpt_abort_cmd(ioctx);
+               else
+                       printk(KERN_ERR "%s[%d]: wrong state = %d\n",
+                              __func__, __LINE__, state);
+               break;
+       case SRPT_RDMA_WRITE_LAST:
+               atomic_set(&ioctx->cmd.transport_lun_stop, 1);
+               break;
+       default:
+               printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__,
+                      __LINE__, opcode);
+               break;
+       }
+}
+
+/**
+ * srpt_build_cmd_rsp() - Build an SRP_RSP response.
+ * @ch: RDMA channel through which the request has been received.
+ * @ioctx: I/O context associated with the SRP_CMD request. The response will
+ *   be built in the buffer ioctx->buf points at and hence this function will
+ *   overwrite the request data.
+ * @tag: tag of the request for which this response is being generated.
+ * @status: value for the STATUS field of the SRP_RSP information unit.
+ *
+ * Returns the size in bytes of the SRP_RSP response.
+ *
+ * An SRP_RSP response contains a SCSI status or service response. See also
+ * section 6.9 in the SRP r16a document for the format of an SRP_RSP
+ * response. See also SPC-2 for more information about sense data.
+ */
+static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
+                             struct srpt_send_ioctx *ioctx, u64 tag,
+                             int status)
+{
+       struct srp_rsp *srp_rsp;
+       const u8 *sense_data;
+       int sense_data_len, max_sense_len;
+
+       /*
+        * The lowest bit of all SAM-3 status codes is zero (see also
+        * paragraph 5.3 in SAM-3).
+        */
+       WARN_ON(status & 1);
+
+       srp_rsp = ioctx->ioctx.buf;
+       BUG_ON(!srp_rsp);
+
+       sense_data = ioctx->sense_data;
+       sense_data_len = ioctx->cmd.scsi_sense_length;
+       WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
+
+       memset(srp_rsp, 0, sizeof *srp_rsp);
+       srp_rsp->opcode = SRP_RSP;
+       srp_rsp->req_lim_delta =
+               __constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
+       srp_rsp->tag = tag;
+       srp_rsp->status = status;
+
+       if (sense_data_len) {
+               BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
+               max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
+               if (sense_data_len > max_sense_len) {
+                       printk(KERN_WARNING "truncated sense data from %d to %d"
+                              " bytes\n", sense_data_len, max_sense_len);
+                       sense_data_len = max_sense_len;
+               }
+
+               srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
+               srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
+               memcpy(srp_rsp + 1, sense_data, sense_data_len);
+       }
+
+       return sizeof(*srp_rsp) + sense_data_len;
+}
+
+/**
+ * srpt_build_tskmgmt_rsp() - Build a task management response.
+ * @ch:       RDMA channel through which the request has been received.
+ * @ioctx:    I/O context in which the SRP_RSP response will be built.
+ * @rsp_code: RSP_CODE that will be stored in the response.
+ * @tag:      Tag of the request for which this response is being generated.
+ *
+ * Returns the size in bytes of the SRP_RSP response.
+ *
+ * An SRP_RSP response contains a SCSI status or service response. See also
+ * section 6.9 in the SRP r16a document for the format of an SRP_RSP
+ * response.
+ */
+static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
+                                 struct srpt_send_ioctx *ioctx,
+                                 u8 rsp_code, u64 tag)
+{
+       struct srp_rsp *srp_rsp;
+       int resp_data_len;
+       int resp_len;
+
+       resp_data_len = (rsp_code == SRP_TSK_MGMT_SUCCESS) ? 0 : 4;
+       resp_len = sizeof(*srp_rsp) + resp_data_len;
+
+       srp_rsp = ioctx->ioctx.buf;
+       BUG_ON(!srp_rsp);
+       memset(srp_rsp, 0, sizeof *srp_rsp);
+
+       srp_rsp->opcode = SRP_RSP;
+       srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
+                                   + atomic_xchg(&ch->req_lim_delta, 0));
+       srp_rsp->tag = tag;
+
+       if (rsp_code != SRP_TSK_MGMT_SUCCESS) {
+               srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
+               srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
+               srp_rsp->data[3] = rsp_code;
+       }
+
+       return resp_len;
+}
+
+#define NO_SUCH_LUN ((uint64_t)-1LL)
+
+/*
+ * SCSI LUN addressing method. See also SAM-2 and the section about
+ * eight byte LUNs.
+ */
+enum scsi_lun_addr_method {
+       SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
+       SCSI_LUN_ADDR_METHOD_FLAT         = 1,
+       SCSI_LUN_ADDR_METHOD_LUN          = 2,
+       SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
+};
+
+/*
+ * srpt_unpack_lun() - Convert from network LUN to linear LUN.
+ *
+ * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
+ * order (big endian) to a linear LUN. Supports three LUN addressing methods:
+ * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
+ */
+static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
+{
+       uint64_t res = NO_SUCH_LUN;
+       int addressing_method;
+
+       if (unlikely(len < 2)) {
+               printk(KERN_ERR "Illegal LUN length %d, expected 2 bytes or "
+                      "more", len);
+               goto out;
+       }
+
+       switch (len) {
+       case 8:
+               if ((*((__be64 *)lun) &
+                    __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
+                       goto out_err;
+               break;
+       case 4:
+               if (*((__be16 *)&lun[2]) != 0)
+                       goto out_err;
+               break;
+       case 6:
+               if (*((__be32 *)&lun[2]) != 0)
+                       goto out_err;
+               break;
+       case 2:
+               break;
+       default:
+               goto out_err;
+       }
+
+       addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
+       switch (addressing_method) {
+       case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
+       case SCSI_LUN_ADDR_METHOD_FLAT:
+       case SCSI_LUN_ADDR_METHOD_LUN:
+               res = *(lun + 1) | (((*lun) & 0x3f) << 8);
+               break;
+
+       case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
+       default:
+               printk(KERN_ERR "Unimplemented LUN addressing method %u",
+                      addressing_method);
+               break;
+       }
+
+out:
+       return res;
+
+out_err:
+       printk(KERN_ERR "Support for multi-level LUNs has not yet been"
+              " implemented");
+       goto out;
+}
+
+static int srpt_check_stop_free(struct se_cmd *cmd)
+{
+       struct srpt_send_ioctx *ioctx;
+
+       ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
+       return kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
+}
+
+/**
+ * srpt_handle_cmd() - Process SRP_CMD.
+ */
+static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
+                          struct srpt_recv_ioctx *recv_ioctx,
+                          struct srpt_send_ioctx *send_ioctx)
+{
+       struct se_cmd *cmd;
+       struct srp_cmd *srp_cmd;
+       uint64_t unpacked_lun;
+       u64 data_len;
+       enum dma_data_direction dir;
+       int ret;
+
+       BUG_ON(!send_ioctx);
+
+       srp_cmd = recv_ioctx->ioctx.buf;
+       kref_get(&send_ioctx->kref);
+       cmd = &send_ioctx->cmd;
+       send_ioctx->tag = srp_cmd->tag;
+
+       switch (srp_cmd->task_attr) {
+       case SRP_CMD_SIMPLE_Q:
+               cmd->sam_task_attr = MSG_SIMPLE_TAG;
+               break;
+       case SRP_CMD_ORDERED_Q:
+       default:
+               cmd->sam_task_attr = MSG_ORDERED_TAG;
+               break;
+       case SRP_CMD_HEAD_OF_Q:
+               cmd->sam_task_attr = MSG_HEAD_TAG;
+               break;
+       case SRP_CMD_ACA:
+               cmd->sam_task_attr = MSG_ACA_TAG;
+               break;
+       }
+
+       ret = srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len);
+       if (ret) {
+               printk(KERN_ERR "0x%llx: parsing SRP descriptor table failed.\n",
+                      srp_cmd->tag);
+               cmd->se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
+               cmd->scsi_sense_reason = TCM_INVALID_CDB_FIELD;
+               goto send_sense;
+       }
+
+       cmd->data_length = data_len;
+       cmd->data_direction = dir;
+       unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
+                                      sizeof(srp_cmd->lun));
+       if (transport_lookup_cmd_lun(cmd, unpacked_lun) < 0)
+               goto send_sense;
+       ret = transport_generic_allocate_tasks(cmd, srp_cmd->cdb);
+       if (cmd->se_cmd_flags & SCF_SCSI_RESERVATION_CONFLICT)
+               srpt_queue_status(cmd);
+       else if (cmd->se_cmd_flags & SCF_SCSI_CDB_EXCEPTION)
+               goto send_sense;
+       else
+               WARN_ON_ONCE(ret);
+
+       transport_handle_cdb_direct(cmd);
+       return 0;
+
+send_sense:
+       transport_send_check_condition_and_sense(cmd, cmd->scsi_sense_reason,
+                                                0);
+       return -1;
+}
+
+/**
+ * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
+ * @ch: RDMA channel of the task management request.
+ * @fn: Task management function to perform.
+ * @req_tag: Tag of the SRP task management request.
+ * @mgmt_ioctx: I/O context of the task management request.
+ *
+ * Returns zero if the target core will process the task management
+ * request asynchronously.
+ *
+ * Note: It is assumed that the initiator serializes tag-based task management
+ * requests.
+ */
+static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
+{
+       struct srpt_device *sdev;
+       struct srpt_rdma_ch *ch;
+       struct srpt_send_ioctx *target;
+       int ret, i;
+
+       ret = -EINVAL;
+       ch = ioctx->ch;
+       BUG_ON(!ch);
+       BUG_ON(!ch->sport);
+       sdev = ch->sport->sdev;
+       BUG_ON(!sdev);
+       spin_lock_irq(&sdev->spinlock);
+       for (i = 0; i < ch->rq_size; ++i) {
+               target = ch->ioctx_ring[i];
+               if (target->cmd.se_lun == ioctx->cmd.se_lun &&
+                   target->tag == tag &&
+                   srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
+                       ret = 0;
+                       /* now let the target core abort &target->cmd; */
+                       break;
+               }
+       }
+       spin_unlock_irq(&sdev->spinlock);
+       return ret;
+}
+
+static int srp_tmr_to_tcm(int fn)
+{
+       switch (fn) {
+       case SRP_TSK_ABORT_TASK:
+               return TMR_ABORT_TASK;
+       case SRP_TSK_ABORT_TASK_SET:
+               return TMR_ABORT_TASK_SET;
+       case SRP_TSK_CLEAR_TASK_SET:
+               return TMR_CLEAR_TASK_SET;
+       case SRP_TSK_LUN_RESET:
+               return TMR_LUN_RESET;
+       case SRP_TSK_CLEAR_ACA:
+               return TMR_CLEAR_ACA;
+       default:
+               return -1;
+       }
+}
+
+/**
+ * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
+ *
+ * Returns 0 if and only if the request will be processed by the target core.
+ *
+ * For more information about SRP_TSK_MGMT information units, see also section
+ * 6.7 in the SRP r16a document.
+ */
+static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
+                                struct srpt_recv_ioctx *recv_ioctx,
+                                struct srpt_send_ioctx *send_ioctx)
+{
+       struct srp_tsk_mgmt *srp_tsk;
+       struct se_cmd *cmd;
+       uint64_t unpacked_lun;
+       int tcm_tmr;
+       int res;
+
+       BUG_ON(!send_ioctx);
+
+       srp_tsk = recv_ioctx->ioctx.buf;
+       cmd = &send_ioctx->cmd;
+
+       pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
+                " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
+                srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
+
+       srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
+       send_ioctx->tag = srp_tsk->tag;
+       tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
+       if (tcm_tmr < 0) {
+               send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
+               send_ioctx->cmd.se_tmr_req->response =
+                       TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
+               goto process_tmr;
+       }
+       cmd->se_tmr_req = core_tmr_alloc_req(cmd, NULL, tcm_tmr, GFP_KERNEL);
+       if (!cmd->se_tmr_req) {
+               send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
+               send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
+               goto process_tmr;
+       }
+
+       unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
+                                      sizeof(srp_tsk->lun));
+       res = transport_lookup_tmr_lun(&send_ioctx->cmd, unpacked_lun);
+       if (res) {
+               pr_debug("rejecting TMR for LUN %lld\n", unpacked_lun);
+               send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
+               send_ioctx->cmd.se_tmr_req->response = TMR_LUN_DOES_NOT_EXIST;
+               goto process_tmr;
+       }
+
+       if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK)
+               srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
+
+process_tmr:
+       kref_get(&send_ioctx->kref);
+       if (!(send_ioctx->cmd.se_cmd_flags & SCF_SCSI_CDB_EXCEPTION))
+               transport_generic_handle_tmr(&send_ioctx->cmd);
+       else
+               transport_send_check_condition_and_sense(cmd,
+                                               cmd->scsi_sense_reason, 0);
+
+}
+
+/**
+ * srpt_handle_new_iu() - Process a newly received information unit.
+ * @ch:    RDMA channel through which the information unit has been received.
+ * @ioctx: SRPT I/O context associated with the information unit.
+ */
+static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
+                              struct srpt_recv_ioctx *recv_ioctx,
+                              struct srpt_send_ioctx *send_ioctx)
+{
+       struct srp_cmd *srp_cmd;
+       enum rdma_ch_state ch_state;
+
+       BUG_ON(!ch);
+       BUG_ON(!recv_ioctx);
+
+       ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
+                                  recv_ioctx->ioctx.dma, srp_max_req_size,
+                                  DMA_FROM_DEVICE);
+
+       ch_state = srpt_get_ch_state(ch);
+       if (unlikely(ch_state == CH_CONNECTING)) {
+               list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
+               goto out;
+       }
+
+       if (unlikely(ch_state != CH_LIVE))
+               goto out;
+
+       srp_cmd = recv_ioctx->ioctx.buf;
+       if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
+               if (!send_ioctx)
+                       send_ioctx = srpt_get_send_ioctx(ch);
+               if (unlikely(!send_ioctx)) {
+                       list_add_tail(&recv_ioctx->wait_list,
+                                     &ch->cmd_wait_list);
+                       goto out;
+               }
+       }
+
+       transport_init_se_cmd(&send_ioctx->cmd, &srpt_target->tf_ops, ch->sess,
+                             0, DMA_NONE, MSG_SIMPLE_TAG,
+                             send_ioctx->sense_data);
+
+       switch (srp_cmd->opcode) {
+       case SRP_CMD:
+               srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
+               break;
+       case SRP_TSK_MGMT:
+               srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
+               break;
+       case SRP_I_LOGOUT:
+               printk(KERN_ERR "Not yet implemented: SRP_I_LOGOUT\n");
+               break;
+       case SRP_CRED_RSP:
+               pr_debug("received SRP_CRED_RSP\n");
+               break;
+       case SRP_AER_RSP:
+               pr_debug("received SRP_AER_RSP\n");
+               break;
+       case SRP_RSP:
+               printk(KERN_ERR "Received SRP_RSP\n");
+               break;
+       default:
+               printk(KERN_ERR "received IU with unknown opcode 0x%x\n",
+                      srp_cmd->opcode);
+               break;
+       }
+
+       srpt_post_recv(ch->sport->sdev, recv_ioctx);
+out:
+       return;
+}
+
+static void srpt_process_rcv_completion(struct ib_cq *cq,
+                                       struct srpt_rdma_ch *ch,
+                                       struct ib_wc *wc)
+{
+       struct srpt_device *sdev = ch->sport->sdev;
+       struct srpt_recv_ioctx *ioctx;
+       u32 index;
+
+       index = idx_from_wr_id(wc->wr_id);
+       if (wc->status == IB_WC_SUCCESS) {
+               int req_lim;
+
+               req_lim = atomic_dec_return(&ch->req_lim);
+               if (unlikely(req_lim < 0))
+                       printk(KERN_ERR "req_lim = %d < 0\n", req_lim);
+               ioctx = sdev->ioctx_ring[index];
+               srpt_handle_new_iu(ch, ioctx, NULL);
+       } else {
+               printk(KERN_INFO "receiving failed for idx %u with status %d\n",
+                      index, wc->status);
+       }
+}
+
+/**
+ * srpt_process_send_completion() - Process an IB send completion.
+ *
+ * Note: Although this has not yet been observed during tests, at least in
+ * theory it is possible that the srpt_get_send_ioctx() call invoked by
+ * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
+ * value in each response is set to one, and it is possible that this response
+ * makes the initiator send a new request before the send completion for that
+ * response has been processed. This could e.g. happen if the call to
+ * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
+ * if IB retransmission causes generation of the send completion to be
+ * delayed. Incoming information units for which srpt_get_send_ioctx() fails
+ * are queued on cmd_wait_list. The code below processes these delayed
+ * requests one at a time.
+ */
+static void srpt_process_send_completion(struct ib_cq *cq,
+                                        struct srpt_rdma_ch *ch,
+                                        struct ib_wc *wc)
+{
+       struct srpt_send_ioctx *send_ioctx;
+       uint32_t index;
+       enum srpt_opcode opcode;
+
+       index = idx_from_wr_id(wc->wr_id);
+       opcode = opcode_from_wr_id(wc->wr_id);
+       send_ioctx = ch->ioctx_ring[index];
+       if (wc->status == IB_WC_SUCCESS) {
+               if (opcode == SRPT_SEND)
+                       srpt_handle_send_comp(ch, send_ioctx);
+               else {
+                       WARN_ON(opcode != SRPT_RDMA_ABORT &&
+                               wc->opcode != IB_WC_RDMA_READ);
+                       srpt_handle_rdma_comp(ch, send_ioctx, opcode);
+               }
+       } else {
+               if (opcode == SRPT_SEND) {
+                       printk(KERN_INFO "sending response for idx %u failed"
+                              " with status %d\n", index, wc->status);
+                       srpt_handle_send_err_comp(ch, wc->wr_id);
+               } else if (opcode != SRPT_RDMA_MID) {
+                       printk(KERN_INFO "RDMA t %d for idx %u failed with"
+                               " status %d", opcode, index, wc->status);
+                       srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
+               }
+       }
+
+       while (unlikely(opcode == SRPT_SEND
+                       && !list_empty(&ch->cmd_wait_list)
+                       && srpt_get_ch_state(ch) == CH_LIVE
+                       && (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
+               struct srpt_recv_ioctx *recv_ioctx;
+
+               recv_ioctx = list_first_entry(&ch->cmd_wait_list,
+                                             struct srpt_recv_ioctx,
+                                             wait_list);
+               list_del(&recv_ioctx->wait_list);
+               srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
+       }
+}
+
+static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
+{
+       struct ib_wc *const wc = ch->wc;
+       int i, n;
+
+       WARN_ON(cq != ch->cq);
+
+       ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
+       while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
+               for (i = 0; i < n; i++) {
+                       if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
+                               srpt_process_rcv_completion(cq, ch, &wc[i]);
+                       else
+                               srpt_process_send_completion(cq, ch, &wc[i]);
+               }
+       }
+}
+
+/**
+ * srpt_completion() - IB completion queue callback function.
+ *
+ * Notes:
+ * - It is guaranteed that a completion handler will never be invoked
+ *   concurrently on two different CPUs for the same completion queue. See also
+ *   Documentation/infiniband/core_locking.txt and the implementation of
+ *   handle_edge_irq() in kernel/irq/chip.c.
+ * - When threaded IRQs are enabled, completion handlers are invoked in thread
+ *   context instead of interrupt context.
+ */
+static void srpt_completion(struct ib_cq *cq, void *ctx)
+{
+       struct srpt_rdma_ch *ch = ctx;
+
+       wake_up_interruptible(&ch->wait_queue);
+}
+
+static int srpt_compl_thread(void *arg)
+{
+       struct srpt_rdma_ch *ch;
+
+       /* Hibernation / freezing of the SRPT kernel thread is not supported. */
+       current->flags |= PF_NOFREEZE;
+
+       ch = arg;
+       BUG_ON(!ch);
+       printk(KERN_INFO "Session %s: kernel thread %s (PID %d) started\n",
+              ch->sess_name, ch->thread->comm, current->pid);
+       while (!kthread_should_stop()) {
+               wait_event_interruptible(ch->wait_queue,
+                       (srpt_process_completion(ch->cq, ch),
+                        kthread_should_stop()));
+       }
+       printk(KERN_INFO "Session %s: kernel thread %s (PID %d) stopped\n",
+              ch->sess_name, ch->thread->comm, current->pid);
+       return 0;
+}
+
+/**
+ * srpt_create_ch_ib() - Create receive and send completion queues.
+ */
+static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
+{
+       struct ib_qp_init_attr *qp_init;
+       struct srpt_port *sport = ch->sport;
+       struct srpt_device *sdev = sport->sdev;
+       u32 srp_sq_size = sport->port_attrib.srp_sq_size;
+       int ret;
+
+       WARN_ON(ch->rq_size < 1);
+
+       ret = -ENOMEM;
+       qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
+       if (!qp_init)
+               goto out;
+
+       ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
+                             ch->rq_size + srp_sq_size, 0);
+       if (IS_ERR(ch->cq)) {
+               ret = PTR_ERR(ch->cq);
+               printk(KERN_ERR "failed to create CQ cqe= %d ret= %d\n",
+                      ch->rq_size + srp_sq_size, ret);
+               goto out;
+       }
+
+       qp_init->qp_context = (void *)ch;
+       qp_init->event_handler
+               = (void(*)(struct ib_event *, void*))srpt_qp_event;
+       qp_init->send_cq = ch->cq;
+       qp_init->recv_cq = ch->cq;
+       qp_init->srq = sdev->srq;
+       qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
+       qp_init->qp_type = IB_QPT_RC;
+       qp_init->cap.max_send_wr = srp_sq_size;
+       qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
+
+       ch->qp = ib_create_qp(sdev->pd, qp_init);
+       if (IS_ERR(ch->qp)) {
+               ret = PTR_ERR(ch->qp);
+               printk(KERN_ERR "failed to create_qp ret= %d\n", ret);
+               goto err_destroy_cq;
+       }
+
+       atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
+
+       pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
+                __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
+                qp_init->cap.max_send_wr, ch->cm_id);
+
+       ret = srpt_init_ch_qp(ch, ch->qp);
+       if (ret)
+               goto err_destroy_qp;
+
+       init_waitqueue_head(&ch->wait_queue);
+
+       pr_debug("creating thread for session %s\n", ch->sess_name);
+
+       ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
+       if (IS_ERR(ch->thread)) {
+               printk(KERN_ERR "failed to create kernel thread %ld\n",
+                      PTR_ERR(ch->thread));
+               ch->thread = NULL;
+               goto err_destroy_qp;
+       }
+
+out:
+       kfree(qp_init);
+       return ret;
+
+err_destroy_qp:
+       ib_destroy_qp(ch->qp);
+err_destroy_cq:
+       ib_destroy_cq(ch->cq);
+       goto out;
+}
+
+static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
+{
+       if (ch->thread)
+               kthread_stop(ch->thread);
+
+       ib_destroy_qp(ch->qp);
+       ib_destroy_cq(ch->cq);
+}
+
+/**
+ * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
+ *
+ * Reset the QP and make sure all resources associated with the channel will
+ * be deallocated at an appropriate time.
+ *
+ * Note: The caller must hold ch->sport->sdev->spinlock.
+ */
+static void __srpt_close_ch(struct srpt_rdma_ch *ch)
+{
+       struct srpt_device *sdev;
+       enum rdma_ch_state prev_state;
+       unsigned long flags;
+
+       sdev = ch->sport->sdev;
+
+       spin_lock_irqsave(&ch->spinlock, flags);
+       prev_state = ch->state;
+       switch (prev_state) {
+       case CH_CONNECTING:
+       case CH_LIVE:
+               ch->state = CH_DISCONNECTING;
+               break;
+       default:
+               break;
+       }
+       spin_unlock_irqrestore(&ch->spinlock, flags);
+
+       switch (prev_state) {
+       case CH_CONNECTING:
+               ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
+                              NULL, 0);
+               /* fall through */
+       case CH_LIVE:
+               if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
+                       printk(KERN_ERR "sending CM DREQ failed.\n");
+               break;
+       case CH_DISCONNECTING:
+               break;
+       case CH_DRAINING:
+       case CH_RELEASING:
+               break;
+       }
+}
+
+/**
+ * srpt_close_ch() - Close an RDMA channel.
+ */
+static void srpt_close_ch(struct srpt_rdma_ch *ch)
+{
+       struct srpt_device *sdev;
+
+       sdev = ch->sport->sdev;
+       spin_lock_irq(&sdev->spinlock);
+       __srpt_close_ch(ch);
+       spin_unlock_irq(&sdev->spinlock);
+}
+
+/**
+ * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
+ * @cm_id: Pointer to the CM ID of the channel to be drained.
+ *
+ * Note: Must be called from inside srpt_cm_handler to avoid a race between
+ * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
+ * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
+ * waits until all target sessions for the associated IB device have been
+ * unregistered and target session registration involves a call to
+ * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
+ * this function has finished).
+ */
+static void srpt_drain_channel(struct ib_cm_id *cm_id)
+{
+       struct srpt_device *sdev;
+       struct srpt_rdma_ch *ch;
+       int ret;
+       bool do_reset = false;
+
+       WARN_ON_ONCE(irqs_disabled());
+
+       sdev = cm_id->context;
+       BUG_ON(!sdev);
+       spin_lock_irq(&sdev->spinlock);
+       list_for_each_entry(ch, &sdev->rch_list, list) {
+               if (ch->cm_id == cm_id) {
+                       do_reset = srpt_test_and_set_ch_state(ch,
+                                       CH_CONNECTING, CH_DRAINING) ||
+                                  srpt_test_and_set_ch_state(ch,
+                                       CH_LIVE, CH_DRAINING) ||
+                                  srpt_test_and_set_ch_state(ch,
+                                       CH_DISCONNECTING, CH_DRAINING);
+                       break;
+               }
+       }
+       spin_unlock_irq(&sdev->spinlock);
+
+       if (do_reset) {
+               ret = srpt_ch_qp_err(ch);
+               if (ret < 0)
+                       printk(KERN_ERR "Setting queue pair in error state"
+                              " failed: %d\n", ret);
+       }
+}
+
+/**
+ * srpt_find_channel() - Look up an RDMA channel.
+ * @cm_id: Pointer to the CM ID of the channel to be looked up.
+ *
+ * Return NULL if no matching RDMA channel has been found.
+ */
+static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
+                                             struct ib_cm_id *cm_id)
+{
+       struct srpt_rdma_ch *ch;
+       bool found;
+
+       WARN_ON_ONCE(irqs_disabled());
+       BUG_ON(!sdev);
+
+       found = false;
+       spin_lock_irq(&sdev->spinlock);
+       list_for_each_entry(ch, &sdev->rch_list, list) {
+               if (ch->cm_id == cm_id) {
+                       found = true;
+                       break;
+               }
+       }
+       spin_unlock_irq(&sdev->spinlock);
+
+       return found ? ch : NULL;
+}
+
+/**
+ * srpt_release_channel() - Release channel resources.
+ *
+ * Schedules the actual release because:
+ * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
+ *   trigger a deadlock.
+ * - It is not safe to call TCM transport_* functions from interrupt context.
+ */
+static void srpt_release_channel(struct srpt_rdma_ch *ch)
+{
+       schedule_work(&ch->release_work);
+}
+
+static void srpt_release_channel_work(struct work_struct *w)
+{
+       struct srpt_rdma_ch *ch;
+       struct srpt_device *sdev;
+
+       ch = container_of(w, struct srpt_rdma_ch, release_work);
+       pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
+                ch->release_done);
+
+       sdev = ch->sport->sdev;
+       BUG_ON(!sdev);
+
+       transport_deregister_session_configfs(ch->sess);
+       transport_deregister_session(ch->sess);
+       ch->sess = NULL;
+
+       srpt_destroy_ch_ib(ch);
+
+       srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
+                            ch->sport->sdev, ch->rq_size,
+                            ch->rsp_size, DMA_TO_DEVICE);
+
+       spin_lock_irq(&sdev->spinlock);
+       list_del(&ch->list);
+       spin_unlock_irq(&sdev->spinlock);
+
+       ib_destroy_cm_id(ch->cm_id);
+
+       if (ch->release_done)
+               complete(ch->release_done);
+
+       wake_up(&sdev->ch_releaseQ);
+
+       kfree(ch);
+}
+
+static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
+                                              u8 i_port_id[16])
+{
+       struct srpt_node_acl *nacl;
+
+       list_for_each_entry(nacl, &sport->port_acl_list, list)
+               if (memcmp(nacl->i_port_id, i_port_id,
+                          sizeof(nacl->i_port_id)) == 0)
+                       return nacl;
+
+       return NULL;
+}
+
+static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
+                                            u8 i_port_id[16])
+{
+       struct srpt_node_acl *nacl;
+
+       spin_lock_irq(&sport->port_acl_lock);
+       nacl = __srpt_lookup_acl(sport, i_port_id);
+       spin_unlock_irq(&sport->port_acl_lock);
+
+       return nacl;
+}
+
+/**
+ * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
+ *
+ * Ownership of the cm_id is transferred to the target session if this
+ * functions returns zero. Otherwise the caller remains the owner of cm_id.
+ */
+static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
+                           struct ib_cm_req_event_param *param,
+                           void *private_data)
+{
+       struct srpt_device *sdev = cm_id->context;
+       struct srpt_port *sport = &sdev->port[param->port - 1];
+       struct srp_login_req *req;
+       struct srp_login_rsp *rsp;
+       struct srp_login_rej *rej;
+       struct ib_cm_rep_param *rep_param;
+       struct srpt_rdma_ch *ch, *tmp_ch;
+       struct srpt_node_acl *nacl;
+       u32 it_iu_len;
+       int i;
+       int ret = 0;
+
+       WARN_ON_ONCE(irqs_disabled());
+
+       if (WARN_ON(!sdev || !private_data))
+               return -EINVAL;
+
+       req = (struct srp_login_req *)private_data;
+
+       it_iu_len = be32_to_cpu(req->req_it_iu_len);
+
+       printk(KERN_INFO "Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
+              " t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
+              " (guid=0x%llx:0x%llx)\n",
+              be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
+              be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
+              be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
+              be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
+              it_iu_len,
+              param->port,
+              be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
+              be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
+
+       rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
+       rej = kzalloc(sizeof *rej, GFP_KERNEL);
+       rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
+
+       if (!rsp || !rej || !rep_param) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
+               rej->reason = __constant_cpu_to_be32(
+                               SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
+               ret = -EINVAL;
+               printk(KERN_ERR "rejected SRP_LOGIN_REQ because its"
+                      " length (%d bytes) is out of range (%d .. %d)\n",
+                      it_iu_len, 64, srp_max_req_size);
+               goto reject;
+       }
+
+       if (!sport->enabled) {
+               rej->reason = __constant_cpu_to_be32(
+                            SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
+               ret = -EINVAL;
+               printk(KERN_ERR "rejected SRP_LOGIN_REQ because the target port"
+                      " has not yet been enabled\n");
+               goto reject;
+       }
+
+       if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
+               rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
+
+               spin_lock_irq(&sdev->spinlock);
+
+               list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
+                       if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
+                           && !memcmp(ch->t_port_id, req->target_port_id, 16)
+                           && param->port == ch->sport->port
+                           && param->listen_id == ch->sport->sdev->cm_id
+                           && ch->cm_id) {
+                               enum rdma_ch_state ch_state;
+
+                               ch_state = srpt_get_ch_state(ch);
+                               if (ch_state != CH_CONNECTING
+                                   && ch_state != CH_LIVE)
+                                       continue;
+
+                               /* found an existing channel */
+                               pr_debug("Found existing channel %s"
+                                        " cm_id= %p state= %d\n",
+                                        ch->sess_name, ch->cm_id, ch_state);
+
+                               __srpt_close_ch(ch);
+
+                               rsp->rsp_flags =
+                                       SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
+                       }
+               }
+
+               spin_unlock_irq(&sdev->spinlock);
+
+       } else
+               rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
+
+       if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
+           || *(__be64 *)(req->target_port_id + 8) !=
+              cpu_to_be64(srpt_service_guid)) {
+               rej->reason = __constant_cpu_to_be32(
+                               SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
+               ret = -ENOMEM;
+               printk(KERN_ERR "rejected SRP_LOGIN_REQ because it"
+                      " has an invalid target port identifier.\n");
+               goto reject;
+       }
+
+       ch = kzalloc(sizeof *ch, GFP_KERNEL);
+       if (!ch) {
+               rej->reason = __constant_cpu_to_be32(
+                                       SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
+               printk(KERN_ERR "rejected SRP_LOGIN_REQ because no memory.\n");
+               ret = -ENOMEM;
+               goto reject;
+       }
+
+       INIT_WORK(&ch->release_work, srpt_release_channel_work);
+       memcpy(ch->i_port_id, req->initiator_port_id, 16);
+       memcpy(ch->t_port_id, req->target_port_id, 16);
+       ch->sport = &sdev->port[param->port - 1];
+       ch->cm_id = cm_id;
+       /*
+        * Avoid QUEUE_FULL conditions by limiting the number of buffers used
+        * for the SRP protocol to the command queue size.
+        */
+       ch->rq_size = SRPT_RQ_SIZE;
+       spin_lock_init(&ch->spinlock);
+       ch->state = CH_CONNECTING;
+       INIT_LIST_HEAD(&ch->cmd_wait_list);
+       ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
+
+       ch->ioctx_ring = (struct srpt_send_ioctx **)
+               srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
+                                     sizeof(*ch->ioctx_ring[0]),
+                                     ch->rsp_size, DMA_TO_DEVICE);
+       if (!ch->ioctx_ring)
+               goto free_ch;
+
+       INIT_LIST_HEAD(&ch->free_list);
+       for (i = 0; i < ch->rq_size; i++) {
+               ch->ioctx_ring[i]->ch = ch;
+               list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
+       }
+
+       ret = srpt_create_ch_ib(ch);
+       if (ret) {
+               rej->reason = __constant_cpu_to_be32(
+                               SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
+               printk(KERN_ERR "rejected SRP_LOGIN_REQ because creating"
+                      " a new RDMA channel failed.\n");
+               goto free_ring;
+       }
+
+       ret = srpt_ch_qp_rtr(ch, ch->qp);
+       if (ret) {
+               rej->reason = __constant_cpu_to_be32(
+                               SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
+               printk(KERN_ERR "rejected SRP_LOGIN_REQ because enabling"
+                      " RTR failed (error code = %d)\n", ret);
+               goto destroy_ib;
+       }
+       /*
+        * Use the initator port identifier as the session name.
+        */
+       snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
+                       be64_to_cpu(*(__be64 *)ch->i_port_id),
+                       be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
+
+       pr_debug("registering session %s\n", ch->sess_name);
+
+       nacl = srpt_lookup_acl(sport, ch->i_port_id);
+       if (!nacl) {
+               printk(KERN_INFO "Rejected login because no ACL has been"
+                      " configured yet for initiator %s.\n", ch->sess_name);
+               rej->reason = __constant_cpu_to_be32(
+                               SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
+               goto destroy_ib;
+       }
+
+       ch->sess = transport_init_session();
+       if (!ch->sess) {
+               rej->reason = __constant_cpu_to_be32(
+                               SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
+               pr_debug("Failed to create session\n");
+               goto deregister_session;
+       }
+       ch->sess->se_node_acl = &nacl->nacl;
+       transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
+
+       pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
+                ch->sess_name, ch->cm_id);
+
+       /* create srp_login_response */
+       rsp->opcode = SRP_LOGIN_RSP;
+       rsp->tag = req->tag;
+       rsp->max_it_iu_len = req->req_it_iu_len;
+       rsp->max_ti_iu_len = req->req_it_iu_len;
+       ch->max_ti_iu_len = it_iu_len;
+       rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
+                                             | SRP_BUF_FORMAT_INDIRECT);
+       rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
+       atomic_set(&ch->req_lim, ch->rq_size);
+       atomic_set(&ch->req_lim_delta, 0);
+
+       /* create cm reply */
+       rep_param->qp_num = ch->qp->qp_num;
+       rep_param->private_data = (void *)rsp;
+       rep_param->private_data_len = sizeof *rsp;
+       rep_param->rnr_retry_count = 7;
+       rep_param->flow_control = 1;
+       rep_param->failover_accepted = 0;
+       rep_param->srq = 1;
+       rep_param->responder_resources = 4;
+       rep_param->initiator_depth = 4;
+
+       ret = ib_send_cm_rep(cm_id, rep_param);
+       if (ret) {
+               printk(KERN_ERR "sending SRP_LOGIN_REQ response failed"
+                      " (error code = %d)\n", ret);
+               goto release_channel;
+       }
+
+       spin_lock_irq(&sdev->spinlock);
+       list_add_tail(&ch->list, &sdev->rch_list);
+       spin_unlock_irq(&sdev->spinlock);
+
+       goto out;
+
+release_channel:
+       srpt_set_ch_state(ch, CH_RELEASING);
+       transport_deregister_session_configfs(ch->sess);
+
+deregister_session:
+       transport_deregister_session(ch->sess);
+       ch->sess = NULL;
+
+destroy_ib:
+       srpt_destroy_ch_ib(ch);
+
+free_ring:
+       srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
+                            ch->sport->sdev, ch->rq_size,
+                            ch->rsp_size, DMA_TO_DEVICE);
+free_ch:
+       kfree(ch);
+
+reject:
+       rej->opcode = SRP_LOGIN_REJ;
+       rej->tag = req->tag;
+       rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
+                                             | SRP_BUF_FORMAT_INDIRECT);
+
+       ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
+                            (void *)rej, sizeof *rej);
+
+out:
+       kfree(rep_param);
+       kfree(rsp);
+       kfree(rej);
+
+       return ret;
+}
+
+static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
+{
+       printk(KERN_INFO "Received IB REJ for cm_id %p.\n", cm_id);
+       srpt_drain_channel(cm_id);
+}
+
+/**
+ * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
+ *
+ * An IB_CM_RTU_RECEIVED message indicates that the connection is established
+ * and that the recipient may begin transmitting (RTU = ready to use).
+ */
+static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
+{
+       struct srpt_rdma_ch *ch;
+       int ret;
+
+       ch = srpt_find_channel(cm_id->context, cm_id);
+       BUG_ON(!ch);
+
+       if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
+               struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
+
+               ret = srpt_ch_qp_rts(ch, ch->qp);
+
+               list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
+                                        wait_list) {
+                       list_del(&ioctx->wait_list);
+                       srpt_handle_new_iu(ch, ioctx, NULL);
+               }
+               if (ret)
+                       srpt_close_ch(ch);
+       }
+}
+
+static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
+{
+       printk(KERN_INFO "Received IB TimeWait exit for cm_id %p.\n", cm_id);
+       srpt_drain_channel(cm_id);
+}
+
+static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
+{
+       printk(KERN_INFO "Received IB REP error for cm_id %p.\n", cm_id);
+       srpt_drain_channel(cm_id);
+}
+
+/**
+ * srpt_cm_dreq_recv() - Process reception of a DREQ message.
+ */
+static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
+{
+       struct srpt_rdma_ch *ch;
+       unsigned long flags;
+       bool send_drep = false;
+
+       ch = srpt_find_channel(cm_id->context, cm_id);
+       BUG_ON(!ch);
+
+       pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
+
+       spin_lock_irqsave(&ch->spinlock, flags);
+       switch (ch->state) {
+       case CH_CONNECTING:
+       case CH_LIVE:
+               send_drep = true;
+               ch->state = CH_DISCONNECTING;
+               break;
+       case CH_DISCONNECTING:
+       case CH_DRAINING:
+       case CH_RELEASING:
+               WARN(true, "unexpected channel state %d\n", ch->state);
+               break;
+       }
+       spin_unlock_irqrestore(&ch->spinlock, flags);
+
+       if (send_drep) {
+               if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
+                       printk(KERN_ERR "Sending IB DREP failed.\n");
+               printk(KERN_INFO "Received DREQ and sent DREP for session %s.\n",
+                      ch->sess_name);
+       }
+}
+
+/**
+ * srpt_cm_drep_recv() - Process reception of a DREP message.
+ */
+static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
+{
+       printk(KERN_INFO "Received InfiniBand DREP message for cm_id %p.\n",
+              cm_id);
+       srpt_drain_channel(cm_id);
+}
+
+/**
+ * srpt_cm_handler() - IB connection manager callback function.
+ *
+ * A non-zero return value will cause the caller destroy the CM ID.
+ *
+ * Note: srpt_cm_handler() must only return a non-zero value when transferring
+ * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
+ * a non-zero value in any other case will trigger a race with the
+ * ib_destroy_cm_id() call in srpt_release_channel().
+ */
+static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
+{
+       int ret;
+
+       ret = 0;
+       switch (event->event) {
+       case IB_CM_REQ_RECEIVED:
+               ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
+                                      event->private_data);
+               break;
+       case IB_CM_REJ_RECEIVED:
+               srpt_cm_rej_recv(cm_id);
+               break;
+       case IB_CM_RTU_RECEIVED:
+       case IB_CM_USER_ESTABLISHED:
+               srpt_cm_rtu_recv(cm_id);
+               break;
+       case IB_CM_DREQ_RECEIVED:
+               srpt_cm_dreq_recv(cm_id);
+               break;
+       case IB_CM_DREP_RECEIVED:
+               srpt_cm_drep_recv(cm_id);
+               break;
+       case IB_CM_TIMEWAIT_EXIT:
+               srpt_cm_timewait_exit(cm_id);
+               break;
+       case IB_CM_REP_ERROR:
+               srpt_cm_rep_error(cm_id);
+               break;
+       case IB_CM_DREQ_ERROR:
+               printk(KERN_INFO "Received IB DREQ ERROR event.\n");
+               break;
+       case IB_CM_MRA_RECEIVED:
+               printk(KERN_INFO "Received IB MRA event\n");
+               break;
+       default:
+               printk(KERN_ERR "received unrecognized IB CM event %d\n",
+                      event->event);
+               break;
+       }
+
+       return ret;
+}
+
+/**
+ * srpt_perform_rdmas() - Perform IB RDMA.
+ *
+ * Returns zero upon success or a negative number upon failure.
+ */
+static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
+                             struct srpt_send_ioctx *ioctx)
+{
+       struct ib_send_wr wr;
+       struct ib_send_wr *bad_wr;
+       struct rdma_iu *riu;
+       int i;
+       int ret;
+       int sq_wr_avail;
+       enum dma_data_direction dir;
+       const int n_rdma = ioctx->n_rdma;
+
+       dir = ioctx->cmd.data_direction;
+       if (dir == DMA_TO_DEVICE) {
+               /* write */
+               ret = -ENOMEM;
+               sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
+               if (sq_wr_avail < 0) {
+                       printk(KERN_WARNING "IB send queue full (needed %d)\n",
+                              n_rdma);
+                       goto out;
+               }
+       }
+
+       ioctx->rdma_aborted = false;
+       ret = 0;
+       riu = ioctx->rdma_ius;
+       memset(&wr, 0, sizeof wr);
+
+       for (i = 0; i < n_rdma; ++i, ++riu) {
+               if (dir == DMA_FROM_DEVICE) {
+                       wr.opcode = IB_WR_RDMA_WRITE;
+                       wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
+                                               SRPT_RDMA_WRITE_LAST :
+                                               SRPT_RDMA_MID,
+                                               ioctx->ioctx.index);
+               } else {
+                       wr.opcode = IB_WR_RDMA_READ;
+                       wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
+                                               SRPT_RDMA_READ_LAST :
+                                               SRPT_RDMA_MID,
+                                               ioctx->ioctx.index);
+               }
+               wr.next = NULL;
+               wr.wr.rdma.remote_addr = riu->raddr;
+               wr.wr.rdma.rkey = riu->rkey;
+               wr.num_sge = riu->sge_cnt;
+               wr.sg_list = riu->sge;
+
+               /* only get completion event for the last rdma write */
+               if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
+                       wr.send_flags = IB_SEND_SIGNALED;
+
+               ret = ib_post_send(ch->qp, &wr, &bad_wr);
+               if (ret)
+                       break;
+       }
+
+       if (ret)
+               printk(KERN_ERR "%s[%d]: ib_post_send() returned %d for %d/%d",
+                                __func__, __LINE__, ret, i, n_rdma);
+       if (ret && i > 0) {
+               wr.num_sge = 0;
+               wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
+               wr.send_flags = IB_SEND_SIGNALED;
+               while (ch->state == CH_LIVE &&
+                       ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
+                       printk(KERN_INFO "Trying to abort failed RDMA transfer [%d]",
+                               ioctx->ioctx.index);
+                       msleep(1000);
+               }
+               while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
+                       printk(KERN_INFO "Waiting until RDMA abort finished [%d]",
+                               ioctx->ioctx.index);
+                       msleep(1000);
+               }
+       }
+out:
+       if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
+               atomic_add(n_rdma, &ch->sq_wr_avail);
+       return ret;
+}
+
+/**
+ * srpt_xfer_data() - Start data transfer from initiator to target.
+ */
+static int srpt_xfer_data(struct srpt_rdma_ch *ch,
+                         struct srpt_send_ioctx *ioctx)
+{
+       int ret;
+
+       ret = srpt_map_sg_to_ib_sge(ch, ioctx);
+       if (ret) {
+               printk(KERN_ERR "%s[%d] ret=%d\n", __func__, __LINE__, ret);
+               goto out;
+       }
+
+       ret = srpt_perform_rdmas(ch, ioctx);
+       if (ret) {
+               if (ret == -EAGAIN || ret == -ENOMEM)
+                       printk(KERN_INFO "%s[%d] queue full -- ret=%d\n",
+                                  __func__, __LINE__, ret);
+               else
+                       printk(KERN_ERR "%s[%d] fatal error -- ret=%d\n",
+                              __func__, __LINE__, ret);
+               goto out_unmap;
+       }
+
+out:
+       return ret;
+out_unmap:
+       srpt_unmap_sg_to_ib_sge(ch, ioctx);
+       goto out;
+}
+
+static int srpt_write_pending_status(struct se_cmd *se_cmd)
+{
+       struct srpt_send_ioctx *ioctx;
+
+       ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
+       return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
+}
+
+/*
+ * srpt_write_pending() - Start data transfer from initiator to target (write).
+ */
+static int srpt_write_pending(struct se_cmd *se_cmd)
+{
+       struct srpt_rdma_ch *ch;
+       struct srpt_send_ioctx *ioctx;
+       enum srpt_command_state new_state;
+       enum rdma_ch_state ch_state;
+       int ret;
+
+       ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
+
+       new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
+       WARN_ON(new_state == SRPT_STATE_DONE);
+
+       ch = ioctx->ch;
+       BUG_ON(!ch);
+
+       ch_state = srpt_get_ch_state(ch);
+       switch (ch_state) {
+       case CH_CONNECTING:
+               WARN(true, "unexpected channel state %d\n", ch_state);
+               ret = -EINVAL;
+               goto out;
+       case CH_LIVE:
+               break;
+       case CH_DISCONNECTING:
+       case CH_DRAINING:
+       case CH_RELEASING:
+               pr_debug("cmd with tag %lld: channel disconnecting\n",
+                        ioctx->tag);
+               srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
+               ret = -EINVAL;
+               goto out;
+       }
+       ret = srpt_xfer_data(ch, ioctx);
+
+out:
+       return ret;
+}
+
+static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
+{
+       switch (tcm_mgmt_status) {
+       case TMR_FUNCTION_COMPLETE:
+               return SRP_TSK_MGMT_SUCCESS;
+       case TMR_FUNCTION_REJECTED:
+               return SRP_TSK_MGMT_FUNC_NOT_SUPP;
+       }
+       return SRP_TSK_MGMT_FAILED;
+}
+
+/**
+ * srpt_queue_response() - Transmits the response to a SCSI command.
+ *
+ * Callback function called by the TCM core. Must not block since it can be
+ * invoked on the context of the IB completion handler.
+ */
+static int srpt_queue_response(struct se_cmd *cmd)
+{
+       struct srpt_rdma_ch *ch;
+       struct srpt_send_ioctx *ioctx;
+       enum srpt_command_state state;
+       unsigned long flags;
+       int ret;
+       enum dma_data_direction dir;
+       int resp_len;
+       u8 srp_tm_status;
+
+       ret = 0;
+
+       ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
+       ch = ioctx->ch;
+       BUG_ON(!ch);
+
+       spin_lock_irqsave(&ioctx->spinlock, flags);
+       state = ioctx->state;
+       switch (state) {
+       case SRPT_STATE_NEW:
+       case SRPT_STATE_DATA_IN:
+               ioctx->state = SRPT_STATE_CMD_RSP_SENT;
+               break;
+       case SRPT_STATE_MGMT:
+               ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
+               break;
+       default:
+               WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
+                       ch, ioctx->ioctx.index, ioctx->state);
+               break;
+       }
+       spin_unlock_irqrestore(&ioctx->spinlock, flags);
+
+       if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
+                    || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
+               atomic_inc(&ch->req_lim_delta);
+               srpt_abort_cmd(ioctx);
+               goto out;
+       }
+
+       dir = ioctx->cmd.data_direction;
+
+       /* For read commands, transfer the data to the initiator. */
+       if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
+           !ioctx->queue_status_only) {
+               ret = srpt_xfer_data(ch, ioctx);
+               if (ret) {
+                       printk(KERN_ERR "xfer_data failed for tag %llu\n",
+                              ioctx->tag);
+                       goto out;
+               }
+       }
+
+       if (state != SRPT_STATE_MGMT)
+               resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
+                                             cmd->scsi_status);
+       else {
+               srp_tm_status
+                       = tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
+               resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
+                                                ioctx->tag);
+       }
+       ret = srpt_post_send(ch, ioctx, resp_len);
+       if (ret) {
+               printk(KERN_ERR "sending cmd response failed for tag %llu\n",
+                      ioctx->tag);
+               srpt_unmap_sg_to_ib_sge(ch, ioctx);
+               srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
+               kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
+       }
+
+out:
+       return ret;
+}
+
+static int srpt_queue_status(struct se_cmd *cmd)
+{
+       struct srpt_send_ioctx *ioctx;
+
+       ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
+       BUG_ON(ioctx->sense_data != cmd->sense_buffer);
+       if (cmd->se_cmd_flags &
+           (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
+               WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
+       ioctx->queue_status_only = true;
+       return srpt_queue_response(cmd);
+}
+
+static void srpt_refresh_port_work(struct work_struct *work)
+{
+       struct srpt_port *sport = container_of(work, struct srpt_port, work);
+
+       srpt_refresh_port(sport);
+}
+
+static int srpt_ch_list_empty(struct srpt_device *sdev)
+{
+       int res;
+
+       spin_lock_irq(&sdev->spinlock);
+       res = list_empty(&sdev->rch_list);
+       spin_unlock_irq(&sdev->spinlock);
+
+       return res;
+}
+
+/**
+ * srpt_release_sdev() - Free the channel resources associated with a target.
+ */
+static int srpt_release_sdev(struct srpt_device *sdev)
+{
+       struct srpt_rdma_ch *ch, *tmp_ch;
+       int res;
+
+       WARN_ON_ONCE(irqs_disabled());
+
+       BUG_ON(!sdev);
+
+       spin_lock_irq(&sdev->spinlock);
+       list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
+               __srpt_close_ch(ch);
+       spin_unlock_irq(&sdev->spinlock);
+
+       res = wait_event_interruptible(sdev->ch_releaseQ,
+                                      srpt_ch_list_empty(sdev));
+       if (res)
+               printk(KERN_ERR "%s: interrupted.\n", __func__);
+
+       return 0;
+}
+
+static struct srpt_port *__srpt_lookup_port(const char *name)
+{
+       struct ib_device *dev;
+       struct srpt_device *sdev;
+       struct srpt_port *sport;
+       int i;
+
+       list_for_each_entry(sdev, &srpt_dev_list, list) {
+               dev = sdev->device;
+               if (!dev)
+                       continue;
+
+               for (i = 0; i < dev->phys_port_cnt; i++) {
+                       sport = &sdev->port[i];
+
+                       if (!strcmp(sport->port_guid, name))
+                               return sport;
+               }
+       }
+
+       return NULL;
+}
+
+static struct srpt_port *srpt_lookup_port(const char *name)
+{
+       struct srpt_port *sport;
+
+       spin_lock(&srpt_dev_lock);
+       sport = __srpt_lookup_port(name);
+       spin_unlock(&srpt_dev_lock);
+
+       return sport;
+}
+
+/**
+ * srpt_add_one() - Infiniband device addition callback function.
+ */
+static void srpt_add_one(struct ib_device *device)
+{
+       struct srpt_device *sdev;
+       struct srpt_port *sport;
+       struct ib_srq_init_attr srq_attr;
+       int i;
+
+       pr_debug("device = %p, device->dma_ops = %p\n", device,
+                device->dma_ops);
+
+       sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
+       if (!sdev)
+               goto err;
+
+       sdev->device = device;
+       INIT_LIST_HEAD(&sdev->rch_list);
+       init_waitqueue_head(&sdev->ch_releaseQ);
+       spin_lock_init(&sdev->spinlock);
+
+       if (ib_query_device(device, &sdev->dev_attr))
+               goto free_dev;
+
+       sdev->pd = ib_alloc_pd(device);
+       if (IS_ERR(sdev->pd))
+               goto free_dev;
+
+       sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
+       if (IS_ERR(sdev->mr))
+               goto err_pd;
+
+       sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
+
+       srq_attr.event_handler = srpt_srq_event;
+       srq_attr.srq_context = (void *)sdev;
+       srq_attr.attr.max_wr = sdev->srq_size;
+       srq_attr.attr.max_sge = 1;
+       srq_attr.attr.srq_limit = 0;
+
+       sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
+       if (IS_ERR(sdev->srq))
+               goto err_mr;
+
+       pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
+                __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
+                device->name);
+
+       if (!srpt_service_guid)
+               srpt_service_guid = be64_to_cpu(device->node_guid);
+
+       sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
+       if (IS_ERR(sdev->cm_id))
+               goto err_srq;
+
+       /* print out target login information */
+       pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
+                "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
+                srpt_service_guid, srpt_service_guid);
+
+       /*
+        * We do not have a consistent service_id (ie. also id_ext of target_id)
+        * to identify this target. We currently use the guid of the first HCA
+        * in the system as service_id; therefore, the target_id will change
+        * if this HCA is gone bad and replaced by different HCA
+        */
+       if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
+               goto err_cm;
+
+       INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
+                             srpt_event_handler);
+       if (ib_register_event_handler(&sdev->event_handler))
+               goto err_cm;
+
+       sdev->ioctx_ring = (struct srpt_recv_ioctx **)
+               srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
+                                     sizeof(*sdev->ioctx_ring[0]),
+                                     srp_max_req_size, DMA_FROM_DEVICE);
+       if (!sdev->ioctx_ring)
+               goto err_event;
+
+       for (i = 0; i < sdev->srq_size; ++i)
+               srpt_post_recv(sdev, sdev->ioctx_ring[i]);
+
+       WARN_ON(sdev->device->phys_port_cnt
+               > sizeof(sdev->port)/sizeof(sdev->port[0]));
+
+       for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
+               sport = &sdev->port[i - 1];
+               sport->sdev = sdev;
+               sport->port = i;
+               sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
+               sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
+               sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
+               INIT_WORK(&sport->work, srpt_refresh_port_work);
+               INIT_LIST_HEAD(&sport->port_acl_list);
+               spin_lock_init(&sport->port_acl_lock);
+
+               if (srpt_refresh_port(sport)) {
+                       printk(KERN_ERR "MAD registration failed for %s-%d.\n",
+                              srpt_sdev_name(sdev), i);
+                       goto err_ring;
+               }
+               snprintf(sport->port_guid, sizeof(sport->port_guid),
+                       "0x%016llx%016llx",
+                       be64_to_cpu(sport->gid.global.subnet_prefix),
+                       be64_to_cpu(sport->gid.global.interface_id));
+       }
+
+       spin_lock(&srpt_dev_lock);
+       list_add_tail(&sdev->list, &srpt_dev_list);
+       spin_unlock(&srpt_dev_lock);
+
+out:
+       ib_set_client_data(device, &srpt_client, sdev);
+       pr_debug("added %s.\n", device->name);
+       return;
+
+err_ring:
+       srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
+                            sdev->srq_size, srp_max_req_size,
+                            DMA_FROM_DEVICE);
+err_event:
+       ib_unregister_event_handler(&sdev->event_handler);
+err_cm:
+       ib_destroy_cm_id(sdev->cm_id);
+err_srq:
+       ib_destroy_srq(sdev->srq);
+err_mr:
+       ib_dereg_mr(sdev->mr);
+err_pd:
+       ib_dealloc_pd(sdev->pd);
+free_dev:
+       kfree(sdev);
+err:
+       sdev = NULL;
+       printk(KERN_INFO "%s(%s) failed.\n", __func__, device->name);
+       goto out;
+}
+
+/**
+ * srpt_remove_one() - InfiniBand device removal callback function.
+ */
+static void srpt_remove_one(struct ib_device *device)
+{
+       struct srpt_device *sdev;
+       int i;
+
+       sdev = ib_get_client_data(device, &srpt_client);
+       if (!sdev) {
+               printk(KERN_INFO "%s(%s): nothing to do.\n", __func__,
+                      device->name);
+               return;
+       }
+
+       srpt_unregister_mad_agent(sdev);
+
+       ib_unregister_event_handler(&sdev->event_handler);
+
+       /* Cancel any work queued by the just unregistered IB event handler. */
+       for (i = 0; i < sdev->device->phys_port_cnt; i++)
+               cancel_work_sync(&sdev->port[i].work);
+
+       ib_destroy_cm_id(sdev->cm_id);
+
+       /*
+        * Unregistering a target must happen after destroying sdev->cm_id
+        * such that no new SRP_LOGIN_REQ information units can arrive while
+        * destroying the target.
+        */
+       spin_lock(&srpt_dev_lock);
+       list_del(&sdev->list);
+       spin_unlock(&srpt_dev_lock);
+       srpt_release_sdev(sdev);
+
+       ib_destroy_srq(sdev->srq);
+       ib_dereg_mr(sdev->mr);
+       ib_dealloc_pd(sdev->pd);
+
+       srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
+                            sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
+       sdev->ioctx_ring = NULL;
+       kfree(sdev);
+}
+
+static struct ib_client srpt_client = {
+       .name = DRV_NAME,
+       .add = srpt_add_one,
+       .remove = srpt_remove_one
+};
+
+static int srpt_check_true(struct se_portal_group *se_tpg)
+{
+       return 1;
+}
+
+static int srpt_check_false(struct se_portal_group *se_tpg)
+{
+       return 0;
+}
+
+static char *srpt_get_fabric_name(void)
+{
+       return "srpt";
+}
+
+static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
+{
+       return SCSI_TRANSPORTID_PROTOCOLID_SRP;
+}
+
+static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
+{
+       struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
+
+       return sport->port_guid;
+}
+
+static u16 srpt_get_tag(struct se_portal_group *tpg)
+{
+       return 1;
+}
+
+static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
+{
+       return 1;
+}
+
+static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
+                                   struct se_node_acl *se_nacl,
+                                   struct t10_pr_registration *pr_reg,
+                                   int *format_code, unsigned char *buf)
+{
+       struct srpt_node_acl *nacl;
+       struct spc_rdma_transport_id *tr_id;
+
+       nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
+       tr_id = (void *)buf;
+       tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
+       memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
+       return sizeof(*tr_id);
+}
+
+static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
+                                       struct se_node_acl *se_nacl,
+                                       struct t10_pr_registration *pr_reg,
+                                       int *format_code)
+{
+       *format_code = 0;
+       return sizeof(struct spc_rdma_transport_id);
+}
+
+static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
+                                           const char *buf, u32 *out_tid_len,
+                                           char **port_nexus_ptr)
+{
+       struct spc_rdma_transport_id *tr_id;
+
+       *port_nexus_ptr = NULL;
+       *out_tid_len = sizeof(struct spc_rdma_transport_id);
+       tr_id = (void *)buf;
+       return (char *)tr_id->i_port_id;
+}
+
+static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
+{
+       struct srpt_node_acl *nacl;
+
+       nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
+       if (!nacl) {
+               printk(KERN_ERR "Unable to alocate struct srpt_node_acl\n");
+               return NULL;
+       }
+
+       return &nacl->nacl;
+}
+
+static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
+                                   struct se_node_acl *se_nacl)
+{
+       struct srpt_node_acl *nacl;
+
+       nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
+       kfree(nacl);
+}
+
+static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
+{
+       return 1;
+}
+
+static void srpt_release_cmd(struct se_cmd *se_cmd)
+{
+}
+
+/**
+ * srpt_shutdown_session() - Whether or not a session may be shut down.
+ */
+static int srpt_shutdown_session(struct se_session *se_sess)
+{
+       return true;
+}
+
+/**
+ * srpt_close_session() - Forcibly close a session.
+ *
+ * Callback function invoked by the TCM core to clean up sessions associated
+ * with a node ACL when the user invokes
+ * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
+ */
+static void srpt_close_session(struct se_session *se_sess)
+{
+       DECLARE_COMPLETION_ONSTACK(release_done);
+       struct srpt_rdma_ch *ch;
+       struct srpt_device *sdev;
+       int res;
+
+       ch = se_sess->fabric_sess_ptr;
+       WARN_ON(ch->sess != se_sess);
+
+       pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
+
+       sdev = ch->sport->sdev;
+       spin_lock_irq(&sdev->spinlock);
+       BUG_ON(ch->release_done);
+       ch->release_done = &release_done;
+       __srpt_close_ch(ch);
+       spin_unlock_irq(&sdev->spinlock);
+
+       res = wait_for_completion_timeout(&release_done, 60 * HZ);
+       WARN_ON(res <= 0);
+}
+
+/**
+ * To do: Find out whether stop_session() has a meaning for transports
+ * other than iSCSI.
+ */
+static void srpt_stop_session(struct se_session *se_sess, int sess_sleep,
+                             int conn_sleep)
+{
+}
+
+static void srpt_reset_nexus(struct se_session *sess)
+{
+       printk(KERN_ERR "This is the SRP protocol, not iSCSI\n");
+}
+
+static int srpt_sess_logged_in(struct se_session *se_sess)
+{
+       return true;
+}
+
+/**
+ * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
+ *
+ * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
+ * This object represents an arbitrary integer used to uniquely identify a
+ * particular attached remote initiator port to a particular SCSI target port
+ * within a particular SCSI target device within a particular SCSI instance.
+ */
+static u32 srpt_sess_get_index(struct se_session *se_sess)
+{
+       return 0;
+}
+
+static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
+{
+}
+
+static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
+{
+       struct srpt_send_ioctx *ioctx;
+
+       ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
+       return ioctx->tag;
+}
+
+/* Note: only used from inside debug printk's by the TCM core. */
+static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
+{
+       struct srpt_send_ioctx *ioctx;
+
+       ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
+       return srpt_get_cmd_state(ioctx);
+}
+
+static u16 srpt_set_fabric_sense_len(struct se_cmd *cmd, u32 sense_length)
+{
+       return 0;
+}
+
+static u16 srpt_get_fabric_sense_len(void)
+{
+       return 0;
+}
+
+static int srpt_is_state_remove(struct se_cmd *se_cmd)
+{
+       return 0;
+}
+
+/**
+ * srpt_parse_i_port_id() - Parse an initiator port ID.
+ * @name: ASCII representation of a 128-bit initiator port ID.
+ * @i_port_id: Binary 128-bit port ID.
+ */
+static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
+{
+       const char *p;
+       unsigned len, count, leading_zero_bytes;
+       int ret, rc;
+
+       p = name;
+       if (strnicmp(p, "0x", 2) == 0)
+               p += 2;
+       ret = -EINVAL;
+       len = strlen(p);
+       if (len % 2)
+               goto out;
+       count = min(len / 2, 16U);
+       leading_zero_bytes = 16 - count;
+       memset(i_port_id, 0, leading_zero_bytes);
+       rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
+       if (rc < 0)
+               pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
+       ret = 0;
+out:
+       return ret;
+}
+
+/*
+ * configfs callback function invoked for
+ * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
+ */
+static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
+                                            struct config_group *group,
+                                            const char *name)
+{
+       struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
+       struct se_node_acl *se_nacl, *se_nacl_new;
+       struct srpt_node_acl *nacl;
+       int ret = 0;
+       u32 nexus_depth = 1;
+       u8 i_port_id[16];
+
+       if (srpt_parse_i_port_id(i_port_id, name) < 0) {
+               printk(KERN_ERR "invalid initiator port ID %s\n", name);
+               ret = -EINVAL;
+               goto err;
+       }
+
+       se_nacl_new = srpt_alloc_fabric_acl(tpg);
+       if (!se_nacl_new) {
+               ret = -ENOMEM;
+               goto err;
+       }
+       /*
+        * nacl_new may be released by core_tpg_add_initiator_node_acl()
+        * when converting a node ACL from demo mode to explict
+        */
+       se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
+                                                 nexus_depth);
+       if (IS_ERR(se_nacl)) {
+               ret = PTR_ERR(se_nacl);
+               goto err;
+       }
+       /* Locate our struct srpt_node_acl and set sdev and i_port_id. */
+       nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
+       memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
+       nacl->sport = sport;
+
+       spin_lock_irq(&sport->port_acl_lock);
+       list_add_tail(&nacl->list, &sport->port_acl_list);
+       spin_unlock_irq(&sport->port_acl_lock);
+
+       return se_nacl;
+err:
+       return ERR_PTR(ret);
+}
+
+/*
+ * configfs callback function invoked for
+ * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
+ */
+static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
+{
+       struct srpt_node_acl *nacl;
+       struct srpt_device *sdev;
+       struct srpt_port *sport;
+
+       nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
+       sport = nacl->sport;
+       sdev = sport->sdev;
+       spin_lock_irq(&sport->port_acl_lock);
+       list_del(&nacl->list);
+       spin_unlock_irq(&sport->port_acl_lock);
+       core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
+       srpt_release_fabric_acl(NULL, se_nacl);
+}
+
+static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
+       struct se_portal_group *se_tpg,
+       char *page)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+
+       return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
+}
+
+static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
+       struct se_portal_group *se_tpg,
+       const char *page,
+       size_t count)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+       unsigned long val;
+       int ret;
+
+       ret = strict_strtoul(page, 0, &val);
+       if (ret < 0) {
+               pr_err("strict_strtoul() failed with ret: %d\n", ret);
+               return -EINVAL;
+       }
+       if (val > MAX_SRPT_RDMA_SIZE) {
+               pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
+                       MAX_SRPT_RDMA_SIZE);
+               return -EINVAL;
+       }
+       if (val < DEFAULT_MAX_RDMA_SIZE) {
+               pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
+                       val, DEFAULT_MAX_RDMA_SIZE);
+               return -EINVAL;
+       }
+       sport->port_attrib.srp_max_rdma_size = val;
+
+       return count;
+}
+
+TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
+
+static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
+       struct se_portal_group *se_tpg,
+       char *page)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+
+       return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
+}
+
+static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
+       struct se_portal_group *se_tpg,
+       const char *page,
+       size_t count)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+       unsigned long val;
+       int ret;
+
+       ret = strict_strtoul(page, 0, &val);
+       if (ret < 0) {
+               pr_err("strict_strtoul() failed with ret: %d\n", ret);
+               return -EINVAL;
+       }
+       if (val > MAX_SRPT_RSP_SIZE) {
+               pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
+                       MAX_SRPT_RSP_SIZE);
+               return -EINVAL;
+       }
+       if (val < MIN_MAX_RSP_SIZE) {
+               pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
+                       MIN_MAX_RSP_SIZE);
+               return -EINVAL;
+       }
+       sport->port_attrib.srp_max_rsp_size = val;
+
+       return count;
+}
+
+TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
+
+static ssize_t srpt_tpg_attrib_show_srp_sq_size(
+       struct se_portal_group *se_tpg,
+       char *page)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+
+       return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
+}
+
+static ssize_t srpt_tpg_attrib_store_srp_sq_size(
+       struct se_portal_group *se_tpg,
+       const char *page,
+       size_t count)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+       unsigned long val;
+       int ret;
+
+       ret = strict_strtoul(page, 0, &val);
+       if (ret < 0) {
+               pr_err("strict_strtoul() failed with ret: %d\n", ret);
+               return -EINVAL;
+       }
+       if (val > MAX_SRPT_SRQ_SIZE) {
+               pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
+                       MAX_SRPT_SRQ_SIZE);
+               return -EINVAL;
+       }
+       if (val < MIN_SRPT_SRQ_SIZE) {
+               pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
+                       MIN_SRPT_SRQ_SIZE);
+               return -EINVAL;
+       }
+       sport->port_attrib.srp_sq_size = val;
+
+       return count;
+}
+
+TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
+
+static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
+       &srpt_tpg_attrib_srp_max_rdma_size.attr,
+       &srpt_tpg_attrib_srp_max_rsp_size.attr,
+       &srpt_tpg_attrib_srp_sq_size.attr,
+       NULL,
+};
+
+static ssize_t srpt_tpg_show_enable(
+       struct se_portal_group *se_tpg,
+       char *page)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+
+       return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
+}
+
+static ssize_t srpt_tpg_store_enable(
+       struct se_portal_group *se_tpg,
+       const char *page,
+       size_t count)
+{
+       struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
+       unsigned long tmp;
+        int ret;
+
+       ret = strict_strtoul(page, 0, &tmp);
+       if (ret < 0) {
+               printk(KERN_ERR "Unable to extract srpt_tpg_store_enable\n");
+               return -EINVAL;
+       }
+
+       if ((tmp != 0) && (tmp != 1)) {
+               printk(KERN_ERR "Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
+               return -EINVAL;
+       }
+       if (tmp == 1)
+               sport->enabled = true;
+       else
+               sport->enabled = false;
+
+       return count;
+}
+
+TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
+
+static struct configfs_attribute *srpt_tpg_attrs[] = {
+       &srpt_tpg_enable.attr,
+       NULL,
+};
+
+/**
+ * configfs callback invoked for
+ * mkdir /sys/kernel/config/target/$driver/$port/$tpg
+ */
+static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
+                                            struct config_group *group,
+                                            const char *name)
+{
+       struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
+       int res;
+
+       /* Initialize sport->port_wwn and sport->port_tpg_1 */
+       res = core_tpg_register(&srpt_target->tf_ops, &sport->port_wwn,
+                       &sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
+       if (res)
+               return ERR_PTR(res);
+
+       return &sport->port_tpg_1;
+}
+
+/**
+ * configfs callback invoked for
+ * rmdir /sys/kernel/config/target/$driver/$port/$tpg
+ */
+static void srpt_drop_tpg(struct se_portal_group *tpg)
+{
+       struct srpt_port *sport = container_of(tpg,
+                               struct srpt_port, port_tpg_1);
+
+       sport->enabled = false;
+       core_tpg_deregister(&sport->port_tpg_1);
+}
+
+/**
+ * configfs callback invoked for
+ * mkdir /sys/kernel/config/target/$driver/$port
+ */
+static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
+                                     struct config_group *group,
+                                     const char *name)
+{
+       struct srpt_port *sport;
+       int ret;
+
+       sport = srpt_lookup_port(name);
+       pr_debug("make_tport(%s)\n", name);
+       ret = -EINVAL;
+       if (!sport)
+               goto err;
+
+       return &sport->port_wwn;
+
+err:
+       return ERR_PTR(ret);
+}
+
+/**
+ * configfs callback invoked for
+ * rmdir /sys/kernel/config/target/$driver/$port
+ */
+static void srpt_drop_tport(struct se_wwn *wwn)
+{
+       struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
+
+       pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
+}
+
+static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
+                                             char *buf)
+{
+       return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
+}
+
+TF_WWN_ATTR_RO(srpt, version);
+
+static struct configfs_attribute *srpt_wwn_attrs[] = {
+       &srpt_wwn_version.attr,
+       NULL,
+};
+
+static struct target_core_fabric_ops srpt_template = {
+       .get_fabric_name                = srpt_get_fabric_name,
+       .get_fabric_proto_ident         = srpt_get_fabric_proto_ident,
+       .tpg_get_wwn                    = srpt_get_fabric_wwn,
+       .tpg_get_tag                    = srpt_get_tag,
+       .tpg_get_default_depth          = srpt_get_default_depth,
+       .tpg_get_pr_transport_id        = srpt_get_pr_transport_id,
+       .tpg_get_pr_transport_id_len    = srpt_get_pr_transport_id_len,
+       .tpg_parse_pr_out_transport_id  = srpt_parse_pr_out_transport_id,
+       .tpg_check_demo_mode            = srpt_check_false,
+       .tpg_check_demo_mode_cache      = srpt_check_true,
+       .tpg_check_demo_mode_write_protect = srpt_check_true,
+       .tpg_check_prod_mode_write_protect = srpt_check_false,
+       .tpg_alloc_fabric_acl           = srpt_alloc_fabric_acl,
+       .tpg_release_fabric_acl         = srpt_release_fabric_acl,
+       .tpg_get_inst_index             = srpt_tpg_get_inst_index,
+       .release_cmd                    = srpt_release_cmd,
+       .check_stop_free                = srpt_check_stop_free,
+       .shutdown_session               = srpt_shutdown_session,
+       .close_session                  = srpt_close_session,
+       .stop_session                   = srpt_stop_session,
+       .fall_back_to_erl0              = srpt_reset_nexus,
+       .sess_logged_in                 = srpt_sess_logged_in,
+       .sess_get_index                 = srpt_sess_get_index,
+       .sess_get_initiator_sid         = NULL,
+       .write_pending                  = srpt_write_pending,
+       .write_pending_status           = srpt_write_pending_status,
+       .set_default_node_attributes    = srpt_set_default_node_attrs,
+       .get_task_tag                   = srpt_get_task_tag,
+       .get_cmd_state                  = srpt_get_tcm_cmd_state,
+       .queue_data_in                  = srpt_queue_response,
+       .queue_status                   = srpt_queue_status,
+       .queue_tm_rsp                   = srpt_queue_response,
+       .get_fabric_sense_len           = srpt_get_fabric_sense_len,
+       .set_fabric_sense_len           = srpt_set_fabric_sense_len,
+       .is_state_remove                = srpt_is_state_remove,
+       /*
+        * Setup function pointers for generic logic in
+        * target_core_fabric_configfs.c
+        */
+       .fabric_make_wwn                = srpt_make_tport,
+       .fabric_drop_wwn                = srpt_drop_tport,
+       .fabric_make_tpg                = srpt_make_tpg,
+       .fabric_drop_tpg                = srpt_drop_tpg,
+       .fabric_post_link               = NULL,
+       .fabric_pre_unlink              = NULL,
+       .fabric_make_np                 = NULL,
+       .fabric_drop_np                 = NULL,
+       .fabric_make_nodeacl            = srpt_make_nodeacl,
+       .fabric_drop_nodeacl            = srpt_drop_nodeacl,
+};
+
+/**
+ * srpt_init_module() - Kernel module initialization.
+ *
+ * Note: Since ib_register_client() registers callback functions, and since at
+ * least one of these callback functions (srpt_add_one()) calls target core
+ * functions, this driver must be registered with the target core before
+ * ib_register_client() is called.
+ */
+static int __init srpt_init_module(void)
+{
+       int ret;
+
+       ret = -EINVAL;
+       if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
+               printk(KERN_ERR "invalid value %d for kernel module parameter"
+                      " srp_max_req_size -- must be at least %d.\n",
+                      srp_max_req_size, MIN_MAX_REQ_SIZE);
+               goto out;
+       }
+
+       if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
+           || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
+               printk(KERN_ERR "invalid value %d for kernel module parameter"
+                      " srpt_srq_size -- must be in the range [%d..%d].\n",
+                      srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
+               goto out;
+       }
+
+       spin_lock_init(&srpt_dev_lock);
+       INIT_LIST_HEAD(&srpt_dev_list);
+
+       ret = -ENODEV;
+       srpt_target = target_fabric_configfs_init(THIS_MODULE, "srpt");
+       if (!srpt_target) {
+               printk(KERN_ERR "couldn't register\n");
+               goto out;
+       }
+
+       srpt_target->tf_ops = srpt_template;
+
+       /* Enable SG chaining */
+       srpt_target->tf_ops.task_sg_chaining = true;
+
+       /*
+        * Set up default attribute lists.
+        */
+       srpt_target->tf_cit_tmpl.tfc_wwn_cit.ct_attrs = srpt_wwn_attrs;
+       srpt_target->tf_cit_tmpl.tfc_tpg_base_cit.ct_attrs = srpt_tpg_attrs;
+       srpt_target->tf_cit_tmpl.tfc_tpg_attrib_cit.ct_attrs = srpt_tpg_attrib_attrs;
+       srpt_target->tf_cit_tmpl.tfc_tpg_param_cit.ct_attrs = NULL;
+       srpt_target->tf_cit_tmpl.tfc_tpg_np_base_cit.ct_attrs = NULL;
+       srpt_target->tf_cit_tmpl.tfc_tpg_nacl_base_cit.ct_attrs = NULL;
+       srpt_target->tf_cit_tmpl.tfc_tpg_nacl_attrib_cit.ct_attrs = NULL;
+       srpt_target->tf_cit_tmpl.tfc_tpg_nacl_auth_cit.ct_attrs = NULL;
+       srpt_target->tf_cit_tmpl.tfc_tpg_nacl_param_cit.ct_attrs = NULL;
+
+       ret = target_fabric_configfs_register(srpt_target);
+       if (ret < 0) {
+               printk(KERN_ERR "couldn't register\n");
+               goto out_free_target;
+       }
+
+       ret = ib_register_client(&srpt_client);
+       if (ret) {
+               printk(KERN_ERR "couldn't register IB client\n");
+               goto out_unregister_target;
+       }
+
+       return 0;
+
+out_unregister_target:
+       target_fabric_configfs_deregister(srpt_target);
+       srpt_target = NULL;
+out_free_target:
+       if (srpt_target)
+               target_fabric_configfs_free(srpt_target);
+out:
+       return ret;
+}
+
+static void __exit srpt_cleanup_module(void)
+{
+       ib_unregister_client(&srpt_client);
+       target_fabric_configfs_deregister(srpt_target);
+       srpt_target = NULL;
+}
+
+module_init(srpt_init_module);
+module_exit(srpt_cleanup_module);
diff --git a/drivers/infiniband/ulp/srpt/ib_srpt.h b/drivers/infiniband/ulp/srpt/ib_srpt.h
new file mode 100644 (file)
index 0000000..b4b4bbc
--- /dev/null
@@ -0,0 +1,444 @@
+/*
+ * Copyright (c) 2006 - 2009 Mellanox Technology Inc.  All rights reserved.
+ * Copyright (C) 2009 - 2010 Bart Van Assche <bvanassche@acm.org>.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     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.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+
+#ifndef IB_SRPT_H
+#define IB_SRPT_H
+
+#include <linux/version.h>
+#include <linux/types.h>
+#include <linux/list.h>
+#include <linux/wait.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_sa.h>
+#include <rdma/ib_cm.h>
+
+#include <scsi/srp.h>
+
+#include "ib_dm_mad.h"
+
+/*
+ * The prefix the ServiceName field must start with in the device management
+ * ServiceEntries attribute pair. See also the SRP specification.
+ */
+#define SRP_SERVICE_NAME_PREFIX                "SRP.T10:"
+
+enum {
+       /*
+        * SRP IOControllerProfile attributes for SRP target ports that have
+        * not been defined in <scsi/srp.h>. Source: section B.7, table B.7
+        * in the SRP specification.
+        */
+       SRP_PROTOCOL = 0x0108,
+       SRP_PROTOCOL_VERSION = 0x0001,
+       SRP_IO_SUBCLASS = 0x609e,
+       SRP_SEND_TO_IOC = 0x01,
+       SRP_SEND_FROM_IOC = 0x02,
+       SRP_RDMA_READ_FROM_IOC = 0x08,
+       SRP_RDMA_WRITE_FROM_IOC = 0x20,
+
+       /*
+        * srp_login_cmd.req_flags bitmasks. See also table 9 in the SRP
+        * specification.
+        */
+       SRP_MTCH_ACTION = 0x03, /* MULTI-CHANNEL ACTION */
+       SRP_LOSOLNT = 0x10, /* logout solicited notification */
+       SRP_CRSOLNT = 0x20, /* credit request solicited notification */
+       SRP_AESOLNT = 0x40, /* asynchronous event solicited notification */
+
+       /*
+        * srp_cmd.sol_nt / srp_tsk_mgmt.sol_not bitmasks. See also tables
+        * 18 and 20 in the SRP specification.
+        */
+       SRP_SCSOLNT = 0x02, /* SCSOLNT = successful solicited notification */
+       SRP_UCSOLNT = 0x04, /* UCSOLNT = unsuccessful solicited notification */
+
+       /*
+        * srp_rsp.sol_not / srp_t_logout.sol_not bitmasks. See also tables
+        * 16 and 22 in the SRP specification.
+        */
+       SRP_SOLNT = 0x01, /* SOLNT = solicited notification */
+
+       /* See also table 24 in the SRP specification. */
+       SRP_TSK_MGMT_SUCCESS = 0x00,
+       SRP_TSK_MGMT_FUNC_NOT_SUPP = 0x04,
+       SRP_TSK_MGMT_FAILED = 0x05,
+
+       /* See also table 21 in the SRP specification. */
+       SRP_CMD_SIMPLE_Q = 0x0,
+       SRP_CMD_HEAD_OF_Q = 0x1,
+       SRP_CMD_ORDERED_Q = 0x2,
+       SRP_CMD_ACA = 0x4,
+
+       SRP_LOGIN_RSP_MULTICHAN_NO_CHAN = 0x0,
+       SRP_LOGIN_RSP_MULTICHAN_TERMINATED = 0x1,
+       SRP_LOGIN_RSP_MULTICHAN_MAINTAINED = 0x2,
+
+       SRPT_DEF_SG_TABLESIZE = 128,
+       SRPT_DEF_SG_PER_WQE = 16,
+
+       MIN_SRPT_SQ_SIZE = 16,
+       DEF_SRPT_SQ_SIZE = 4096,
+       SRPT_RQ_SIZE = 128,
+       MIN_SRPT_SRQ_SIZE = 4,
+       DEFAULT_SRPT_SRQ_SIZE = 4095,
+       MAX_SRPT_SRQ_SIZE = 65535,
+       MAX_SRPT_RDMA_SIZE = 1U << 24,
+       MAX_SRPT_RSP_SIZE = 1024,
+
+       MIN_MAX_REQ_SIZE = 996,
+       DEFAULT_MAX_REQ_SIZE
+               = sizeof(struct srp_cmd)/*48*/
+               + sizeof(struct srp_indirect_buf)/*20*/
+               + 128 * sizeof(struct srp_direct_buf)/*16*/,
+
+       MIN_MAX_RSP_SIZE = sizeof(struct srp_rsp)/*36*/ + 4,
+       DEFAULT_MAX_RSP_SIZE = 256, /* leaves 220 bytes for sense data */
+
+       DEFAULT_MAX_RDMA_SIZE = 65536,
+};
+
+enum srpt_opcode {
+       SRPT_RECV,
+       SRPT_SEND,
+       SRPT_RDMA_MID,
+       SRPT_RDMA_ABORT,
+       SRPT_RDMA_READ_LAST,
+       SRPT_RDMA_WRITE_LAST,
+};
+
+static inline u64 encode_wr_id(u8 opcode, u32 idx)
+{
+       return ((u64)opcode << 32) | idx;
+}
+static inline enum srpt_opcode opcode_from_wr_id(u64 wr_id)
+{
+       return wr_id >> 32;
+}
+static inline u32 idx_from_wr_id(u64 wr_id)
+{
+       return (u32)wr_id;
+}
+
+struct rdma_iu {
+       u64             raddr;
+       u32             rkey;
+       struct ib_sge   *sge;
+       u32             sge_cnt;
+       int             mem_id;
+};
+
+/**
+ * enum srpt_command_state - SCSI command state managed by SRPT.
+ * @SRPT_STATE_NEW:           New command arrived and is being processed.
+ * @SRPT_STATE_NEED_DATA:     Processing a write or bidir command and waiting
+ *                            for data arrival.
+ * @SRPT_STATE_DATA_IN:       Data for the write or bidir command arrived and is
+ *                            being processed.
+ * @SRPT_STATE_CMD_RSP_SENT:  SRP_RSP for SRP_CMD has been sent.
+ * @SRPT_STATE_MGMT:          Processing a SCSI task management command.
+ * @SRPT_STATE_MGMT_RSP_SENT: SRP_RSP for SRP_TSK_MGMT has been sent.
+ * @SRPT_STATE_DONE:          Command processing finished successfully, command
+ *                            processing has been aborted or command processing
+ *                            failed.
+ */
+enum srpt_command_state {
+       SRPT_STATE_NEW           = 0,
+       SRPT_STATE_NEED_DATA     = 1,
+       SRPT_STATE_DATA_IN       = 2,
+       SRPT_STATE_CMD_RSP_SENT  = 3,
+       SRPT_STATE_MGMT          = 4,
+       SRPT_STATE_MGMT_RSP_SENT = 5,
+       SRPT_STATE_DONE          = 6,
+};
+
+/**
+ * struct srpt_ioctx - Shared SRPT I/O context information.
+ * @buf:   Pointer to the buffer.
+ * @dma:   DMA address of the buffer.
+ * @index: Index of the I/O context in its ioctx_ring array.
+ */
+struct srpt_ioctx {
+       void                    *buf;
+       dma_addr_t              dma;
+       uint32_t                index;
+};
+
+/**
+ * struct srpt_recv_ioctx - SRPT receive I/O context.
+ * @ioctx:     See above.
+ * @wait_list: Node for insertion in srpt_rdma_ch.cmd_wait_list.
+ */
+struct srpt_recv_ioctx {
+       struct srpt_ioctx       ioctx;
+       struct list_head        wait_list;
+};
+
+/**
+ * struct srpt_send_ioctx - SRPT send I/O context.
+ * @ioctx:       See above.
+ * @ch:          Channel pointer.
+ * @free_list:   Node in srpt_rdma_ch.free_list.
+ * @n_rbuf:      Number of data buffers in the received SRP command.
+ * @rbufs:       Pointer to SRP data buffer array.
+ * @single_rbuf: SRP data buffer if the command has only a single buffer.
+ * @sg:          Pointer to sg-list associated with this I/O context.
+ * @sg_cnt:      SG-list size.
+ * @mapped_sg_count: ib_dma_map_sg() return value.
+ * @n_rdma_ius:  Number of elements in the rdma_ius array.
+ * @rdma_ius:    Array with information about the RDMA mapping.
+ * @tag:         Tag of the received SRP information unit.
+ * @spinlock:    Protects 'state'.
+ * @state:       I/O context state.
+ * @rdma_aborted: If initiating a multipart RDMA transfer failed, whether
+ *              the already initiated transfers have finished.
+ * @cmd:         Target core command data structure.
+ * @sense_data:  SCSI sense data.
+ */
+struct srpt_send_ioctx {
+       struct srpt_ioctx       ioctx;
+       struct srpt_rdma_ch     *ch;
+       struct kref              kref;
+       struct rdma_iu          *rdma_ius;
+       struct srp_direct_buf   *rbufs;
+       struct srp_direct_buf   single_rbuf;
+       struct scatterlist      *sg;
+       struct list_head        free_list;
+       spinlock_t              spinlock;
+       enum srpt_command_state state;
+       bool                    rdma_aborted;
+       struct se_cmd           cmd;
+       struct completion       tx_done;
+       u64                     tag;
+       int                     sg_cnt;
+       int                     mapped_sg_count;
+       u16                     n_rdma_ius;
+       u8                      n_rdma;
+       u8                      n_rbuf;
+       bool                    queue_status_only;
+       u8                      sense_data[SCSI_SENSE_BUFFERSIZE];
+};
+
+/**
+ * enum rdma_ch_state - SRP channel state.
+ * @CH_CONNECTING:      QP is in RTR state; waiting for RTU.
+ * @CH_LIVE:            QP is in RTS state.
+ * @CH_DISCONNECTING:    DREQ has been received; waiting for DREP
+ *                       or DREQ has been send and waiting for DREP
+ *                       or .
+ * @CH_DRAINING:        QP is in ERR state; waiting for last WQE event.
+ * @CH_RELEASING:       Last WQE event has been received; releasing resources.
+ */
+enum rdma_ch_state {
+       CH_CONNECTING,
+       CH_LIVE,
+       CH_DISCONNECTING,
+       CH_DRAINING,
+       CH_RELEASING
+};
+
+/**
+ * struct srpt_rdma_ch - RDMA channel.
+ * @wait_queue:    Allows the kernel thread to wait for more work.
+ * @thread:        Kernel thread that processes the IB queues associated with
+ *                 the channel.
+ * @cm_id:         IB CM ID associated with the channel.
+ * @qp:            IB queue pair used for communicating over this channel.
+ * @cq:            IB completion queue for this channel.
+ * @rq_size:       IB receive queue size.
+ * @rsp_size      IB response message size in bytes.
+ * @sq_wr_avail:   number of work requests available in the send queue.
+ * @sport:         pointer to the information of the HCA port used by this
+ *                 channel.
+ * @i_port_id:     128-bit initiator port identifier copied from SRP_LOGIN_REQ.
+ * @t_port_id:     128-bit target port identifier copied from SRP_LOGIN_REQ.
+ * @max_ti_iu_len: maximum target-to-initiator information unit length.
+ * @req_lim:       request limit: maximum number of requests that may be sent
+ *                 by the initiator without having received a response.
+ * @req_lim_delta: Number of credits not yet sent back to the initiator.
+ * @spinlock:      Protects free_list and state.
+ * @free_list:     Head of list with free send I/O contexts.
+ * @state:         channel state. See also enum rdma_ch_state.
+ * @ioctx_ring:    Send ring.
+ * @wc:            IB work completion array for srpt_process_completion().
+ * @list:          Node for insertion in the srpt_device.rch_list list.
+ * @cmd_wait_list: List of SCSI commands that arrived before the RTU event. This
+ *                 list contains struct srpt_ioctx elements and is protected
+ *                 against concurrent modification by the cm_id spinlock.
+ * @sess:          Session information associated with this SRP channel.
+ * @sess_name:     Session name.
+ * @release_work:  Allows scheduling of srpt_release_channel().
+ * @release_done:  Enables waiting for srpt_release_channel() completion.
+ */
+struct srpt_rdma_ch {
+       wait_queue_head_t       wait_queue;
+       struct task_struct      *thread;
+       struct ib_cm_id         *cm_id;
+       struct ib_qp            *qp;
+       struct ib_cq            *cq;
+       int                     rq_size;
+       u32                     rsp_size;
+       atomic_t                sq_wr_avail;
+       struct srpt_port        *sport;
+       u8                      i_port_id[16];
+       u8                      t_port_id[16];
+       int                     max_ti_iu_len;
+       atomic_t                req_lim;
+       atomic_t                req_lim_delta;
+       spinlock_t              spinlock;
+       struct list_head        free_list;
+       enum rdma_ch_state      state;
+       struct srpt_send_ioctx  **ioctx_ring;
+       struct ib_wc            wc[16];
+       struct list_head        list;
+       struct list_head        cmd_wait_list;
+       struct se_session       *sess;
+       u8                      sess_name[36];
+       struct work_struct      release_work;
+       struct completion       *release_done;
+};
+
+/**
+ * struct srpt_port_attib - Attributes for SRPT port
+ * @srp_max_rdma_size: Maximum size of SRP RDMA transfers for new connections.
+ * @srp_max_rsp_size: Maximum size of SRP response messages in bytes.
+ * @srp_sq_size: Shared receive queue (SRQ) size.
+ */
+struct srpt_port_attrib {
+       u32                     srp_max_rdma_size;
+       u32                     srp_max_rsp_size;
+       u32                     srp_sq_size;
+};
+
+/**
+ * struct srpt_port - Information associated by SRPT with a single IB port.
+ * @sdev:      backpointer to the HCA information.
+ * @mad_agent: per-port management datagram processing information.
+ * @enabled:   Whether or not this target port is enabled.
+ * @port_guid: ASCII representation of Port GUID
+ * @port:      one-based port number.
+ * @sm_lid:    cached value of the port's sm_lid.
+ * @lid:       cached value of the port's lid.
+ * @gid:       cached value of the port's gid.
+ * @port_acl_lock spinlock for port_acl_list:
+ * @work:      work structure for refreshing the aforementioned cached values.
+ * @port_tpg_1 Target portal group = 1 data.
+ * @port_wwn:  Target core WWN data.
+ * @port_acl_list: Head of the list with all node ACLs for this port.
+ */
+struct srpt_port {
+       struct srpt_device      *sdev;
+       struct ib_mad_agent     *mad_agent;
+       bool                    enabled;
+       u8                      port_guid[64];
+       u8                      port;
+       u16                     sm_lid;
+       u16                     lid;
+       union ib_gid            gid;
+       spinlock_t              port_acl_lock;
+       struct work_struct      work;
+       struct se_portal_group  port_tpg_1;
+       struct se_wwn           port_wwn;
+       struct list_head        port_acl_list;
+       struct srpt_port_attrib port_attrib;
+};
+
+/**
+ * struct srpt_device - Information associated by SRPT with a single HCA.
+ * @device:        Backpointer to the struct ib_device managed by the IB core.
+ * @pd:            IB protection domain.
+ * @mr:            L_Key (local key) with write access to all local memory.
+ * @srq:           Per-HCA SRQ (shared receive queue).
+ * @cm_id:         Connection identifier.
+ * @dev_attr:      Attributes of the InfiniBand device as obtained during the
+ *                 ib_client.add() callback.
+ * @srq_size:      SRQ size.
+ * @ioctx_ring:    Per-HCA SRQ.
+ * @rch_list:      Per-device channel list -- see also srpt_rdma_ch.list.
+ * @ch_releaseQ:   Enables waiting for removal from rch_list.
+ * @spinlock:      Protects rch_list and tpg.
+ * @port:          Information about the ports owned by this HCA.
+ * @event_handler: Per-HCA asynchronous IB event handler.
+ * @list:          Node in srpt_dev_list.
+ */
+struct srpt_device {
+       struct ib_device        *device;
+       struct ib_pd            *pd;
+       struct ib_mr            *mr;
+       struct ib_srq           *srq;
+       struct ib_cm_id         *cm_id;
+       struct ib_device_attr   dev_attr;
+       int                     srq_size;
+       struct srpt_recv_ioctx  **ioctx_ring;
+       struct list_head        rch_list;
+       wait_queue_head_t       ch_releaseQ;
+       spinlock_t              spinlock;
+       struct srpt_port        port[2];
+       struct ib_event_handler event_handler;
+       struct list_head        list;
+};
+
+/**
+ * struct srpt_node_acl - Per-initiator ACL data (managed via configfs).
+ * @i_port_id: 128-bit SRP initiator port ID.
+ * @sport:     port information.
+ * @nacl:      Target core node ACL information.
+ * @list:      Element of the per-HCA ACL list.
+ */
+struct srpt_node_acl {
+       u8                      i_port_id[16];
+       struct srpt_port        *sport;
+       struct se_node_acl      nacl;
+       struct list_head        list;
+};
+
+/*
+ * SRP-releated SCSI persistent reservation definitions.
+ *
+ * See also SPC4r28, section 7.6.1 (Protocol specific parameters introduction).
+ * See also SPC4r28, section 7.6.4.5 (TransportID for initiator ports using
+ * SCSI over an RDMA interface).
+ */
+
+enum {
+       SCSI_TRANSPORTID_PROTOCOLID_SRP = 4,
+};
+
+struct spc_rdma_transport_id {
+       uint8_t protocol_identifier;
+       uint8_t reserved[7];
+       uint8_t i_port_id[16];
+};
+
+#endif                         /* IB_SRPT_H */