Initial import

7 files changed, 8361 insertions, 0 deletions

diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/Makefile b/drivers/net/ethernet/chelsio/cxgb4vf/Makefile
new file mode 100644
index 000000000..d72ee26cb
--- /dev/null
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/Makefile
@@ -0,0 +1,7 @@
+#
+# Chelsio T4 SR-IOV Virtual Function Driver
+#
+
+obj-$(CONFIG_CHELSIO_T4VF) += cxgb4vf.o
+
+cxgb4vf-objs := cxgb4vf_main.o t4vf_hw.o sge.o
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/adapter.h b/drivers/net/ethernet/chelsio/cxgb4vf/adapter.h
new file mode 100644
index 000000000..6049f70e1
--- /dev/null
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/adapter.h
@@ -0,0 +1,552 @@
+/*
+ * This file is part of the Chelsio T4 PCI-E SR-IOV Virtual Function Ethernet
+ * driver for Linux.
+ *
+ * Copyright (c) 2009-2010 Chelsio Communications, 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.
+ */
+
+/*
+ * This file should not be included directly.  Include t4vf_common.h instead.
+ */
+
+#ifndef __CXGB4VF_ADAPTER_H__
+#define __CXGB4VF_ADAPTER_H__
+
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/spinlock.h>
+#include <linux/skbuff.h>
+#include <linux/if_ether.h>
+#include <linux/netdevice.h>
+
+#include "../cxgb4/t4_hw.h"
+
+/*
+ * Constants of the implementation.
+ */
+enum {
+	MAX_NPORTS	= 1,		/* max # of "ports" */
+	MAX_PORT_QSETS	= 8,		/* max # of Queue Sets / "port" */
+	MAX_ETH_QSETS	= MAX_NPORTS*MAX_PORT_QSETS,
+
+	/*
+	 * MSI-X interrupt index usage.
+	 */
+	MSIX_FW		= 0,		/* MSI-X index for firmware Q */
+	MSIX_IQFLINT	= 1,		/* MSI-X index base for Ingress Qs */
+	MSIX_EXTRAS	= 1,
+	MSIX_ENTRIES	= MAX_ETH_QSETS + MSIX_EXTRAS,
+
+	/*
+	 * The maximum number of Ingress and Egress Queues is determined by
+	 * the maximum number of "Queue Sets" which we support plus any
+	 * ancillary queues.  Each "Queue Set" requires one Ingress Queue
+	 * for RX Packet Ingress Event notifications and two Egress Queues for
+	 * a Free List and an Ethernet TX list.
+	 */
+	INGQ_EXTRAS	= 2,		/* firmware event queue and */
+					/*   forwarded interrupts */
+	MAX_INGQ	= MAX_ETH_QSETS+INGQ_EXTRAS,
+	MAX_EGRQ	= MAX_ETH_QSETS*2,
+};
+
+/*
+ * Forward structure definition references.
+ */
+struct adapter;
+struct sge_eth_rxq;
+struct sge_rspq;
+
+/*
+ * Per-"port" information.  This is really per-Virtual Interface information
+ * but the use of the "port" nomanclature makes it easier to go back and forth
+ * between the PF and VF drivers ...
+ */
+struct port_info {
+	struct adapter *adapter;	/* our adapter */
+	u16 viid;			/* virtual interface ID */
+	s16 xact_addr_filt;		/* index of our MAC address filter */
+	u16 rss_size;			/* size of VI's RSS table slice */
+	u8 pidx;			/* index into adapter port[] */
+	s8 mdio_addr;
+	u8 port_type;			/* firmware port type */
+	u8 mod_type;			/* firmware module type */
+	u8 port_id;			/* physical port ID */
+	u8 nqsets;			/* # of "Queue Sets" */
+	u8 first_qset;			/* index of first "Queue Set" */
+	struct link_config link_cfg;	/* physical port configuration */
+};
+
+/*
+ * Scatter Gather Engine resources for the "adapter".  Our ingress and egress
+ * queues are organized into "Queue Sets" with one ingress and one egress
+ * queue per Queue Set.  These Queue Sets are aportionable between the "ports"
+ * (Virtual Interfaces).  One extra ingress queue is used to receive
+ * asynchronous messages from the firmware.  Note that the "Queue IDs" that we
+ * use here are really "Relative Queue IDs" which are returned as part of the
+ * firmware command to allocate queues.  These queue IDs are relative to the
+ * absolute Queue ID base of the section of the Queue ID space allocated to
+ * the PF/VF.
+ */
+
+/*
+ * SGE free-list queue state.
+ */
+struct rx_sw_desc;
+struct sge_fl {
+	unsigned int avail;		/* # of available RX buffers */
+	unsigned int pend_cred;		/* new buffers since last FL DB ring */
+	unsigned int cidx;		/* consumer index */
+	unsigned int pidx;		/* producer index */
+	unsigned long alloc_failed;	/* # of buffer allocation failures */
+	unsigned long large_alloc_failed;
+	unsigned long starving;		/* # of times FL was found starving */
+
+	/*
+	 * Write-once/infrequently fields.
+	 * -------------------------------
+	 */
+
+	unsigned int cntxt_id;		/* SGE relative QID for the free list */
+	unsigned int abs_id;		/* SGE absolute QID for the free list */
+	unsigned int size;		/* capacity of free list */
+	struct rx_sw_desc *sdesc;	/* address of SW RX descriptor ring */
+	__be64 *desc;			/* address of HW RX descriptor ring */
+	dma_addr_t addr;		/* PCI bus address of hardware ring */
+	void __iomem *bar2_addr;	/* address of BAR2 Queue registers */
+	unsigned int bar2_qid;		/* Queue ID for BAR2 Queue registers */
+};
+
+/*
+ * An ingress packet gather list.
+ */
+struct pkt_gl {
+	struct page_frag frags[MAX_SKB_FRAGS];
+	void *va;			/* virtual address of first byte */
+	unsigned int nfrags;		/* # of fragments */
+	unsigned int tot_len;		/* total length of fragments */
+};
+
+typedef int (*rspq_handler_t)(struct sge_rspq *, const __be64 *,
+			      const struct pkt_gl *);
+
+/*
+ * State for an SGE Response Queue.
+ */
+struct sge_rspq {
+	struct napi_struct napi;	/* NAPI scheduling control */
+	const __be64 *cur_desc;		/* current descriptor in queue */
+	unsigned int cidx;		/* consumer index */
+	u8 gen;				/* current generation bit */
+	u8 next_intr_params;		/* holdoff params for next interrupt */
+	int offset;			/* offset into current FL buffer */
+
+	unsigned int unhandled_irqs;	/* bogus interrupts */
+
+	/*
+	 * Write-once/infrequently fields.
+	 * -------------------------------
+	 */
+
+	u8 intr_params;			/* interrupt holdoff parameters */
+	u8 pktcnt_idx;			/* interrupt packet threshold */
+	u8 idx;				/* queue index within its group */
+	u16 cntxt_id;			/* SGE rel QID for the response Q */
+	u16 abs_id;			/* SGE abs QID for the response Q */
+	__be64 *desc;			/* address of hardware response ring */
+	dma_addr_t phys_addr;		/* PCI bus address of ring */
+	void __iomem *bar2_addr;	/* address of BAR2 Queue registers */
+	unsigned int bar2_qid;		/* Queue ID for BAR2 Queue registers */
+	unsigned int iqe_len;		/* entry size */
+	unsigned int size;		/* capcity of response Q */
+	struct adapter *adapter;	/* our adapter */
+	struct net_device *netdev;	/* associated net device */
+	rspq_handler_t handler;		/* the handler for this response Q */
+};
+
+/*
+ * Ethernet queue statistics
+ */
+struct sge_eth_stats {
+	unsigned long pkts;		/* # of ethernet packets */
+	unsigned long lro_pkts;		/* # of LRO super packets */
+	unsigned long lro_merged;	/* # of wire packets merged by LRO */
+	unsigned long rx_cso;		/* # of Rx checksum offloads */
+	unsigned long vlan_ex;		/* # of Rx VLAN extractions */
+	unsigned long rx_drops;		/* # of packets dropped due to no mem */
+};
+
+/*
+ * State for an Ethernet Receive Queue.
+ */
+struct sge_eth_rxq {
+	struct sge_rspq rspq;		/* Response Queue */
+	struct sge_fl fl;		/* Free List */
+	struct sge_eth_stats stats;	/* receive statistics */
+};
+
+/*
+ * SGE Transmit Queue state.  This contains all of the resources associated
+ * with the hardware status of a TX Queue which is a circular ring of hardware
+ * TX Descriptors.  For convenience, it also contains a pointer to a parallel
+ * "Software Descriptor" array but we don't know anything about it here other
+ * than its type name.
+ */
+struct tx_desc {
+	/*
+	 * Egress Queues are measured in units of SGE_EQ_IDXSIZE by the
+	 * hardware: Sizes, Producer and Consumer indices, etc.
+	 */
+	__be64 flit[SGE_EQ_IDXSIZE/sizeof(__be64)];
+};
+struct tx_sw_desc;
+struct sge_txq {
+	unsigned int in_use;		/* # of in-use TX descriptors */
+	unsigned int size;		/* # of descriptors */
+	unsigned int cidx;		/* SW consumer index */
+	unsigned int pidx;		/* producer index */
+	unsigned long stops;		/* # of times queue has been stopped */
+	unsigned long restarts;		/* # of queue restarts */
+
+	/*
+	 * Write-once/infrequently fields.
+	 * -------------------------------
+	 */
+
+	unsigned int cntxt_id;		/* SGE relative QID for the TX Q */
+	unsigned int abs_id;		/* SGE absolute QID for the TX Q */
+	struct tx_desc *desc;		/* address of HW TX descriptor ring */
+	struct tx_sw_desc *sdesc;	/* address of SW TX descriptor ring */
+	struct sge_qstat *stat;		/* queue status entry */
+	dma_addr_t phys_addr;		/* PCI bus address of hardware ring */
+	void __iomem *bar2_addr;	/* address of BAR2 Queue registers */
+	unsigned int bar2_qid;		/* Queue ID for BAR2 Queue registers */
+};
+
+/*
+ * State for an Ethernet Transmit Queue.
+ */
+struct sge_eth_txq {
+	struct sge_txq q;		/* SGE TX Queue */
+	struct netdev_queue *txq;	/* associated netdev TX queue */
+	unsigned long tso;		/* # of TSO requests */
+	unsigned long tx_cso;		/* # of TX checksum offloads */
+	unsigned long vlan_ins;		/* # of TX VLAN insertions */
+	unsigned long mapping_err;	/* # of I/O MMU packet mapping errors */
+};
+
+/*
+ * The complete set of Scatter/Gather Engine resources.
+ */
+struct sge {
+	/*
+	 * Our "Queue Sets" ...
+	 */
+	struct sge_eth_txq ethtxq[MAX_ETH_QSETS];
+	struct sge_eth_rxq ethrxq[MAX_ETH_QSETS];
+
+	/*
+	 * Extra ingress queues for asynchronous firmware events and
+	 * forwarded interrupts (when in MSI mode).
+	 */
+	struct sge_rspq fw_evtq ____cacheline_aligned_in_smp;
+
+	struct sge_rspq intrq ____cacheline_aligned_in_smp;
+	spinlock_t intrq_lock;
+
+	/*
+	 * State for managing "starving Free Lists" -- Free Lists which have
+	 * fallen below a certain threshold of buffers available to the
+	 * hardware and attempts to refill them up to that threshold have
+	 * failed.  We have a regular "slow tick" timer process which will
+	 * make periodic attempts to refill these starving Free Lists ...
+	 */
+	DECLARE_BITMAP(starving_fl, MAX_EGRQ);
+	struct timer_list rx_timer;
+
+	/*
+	 * State for cleaning up completed TX descriptors.
+	 */
+	struct timer_list tx_timer;
+
+	/*
+	 * Write-once/infrequently fields.
+	 * -------------------------------
+	 */
+
+	u16 max_ethqsets;		/* # of available Ethernet queue sets */
+	u16 ethqsets;			/* # of active Ethernet queue sets */
+	u16 ethtxq_rover;		/* Tx queue to clean up next */
+	u16 timer_val[SGE_NTIMERS];	/* interrupt holdoff timer array */
+	u8 counter_val[SGE_NCOUNTERS];	/* interrupt RX threshold array */
+
+	/* Decoded Adapter Parameters.
+	 */
+	u32 fl_pg_order;		/* large page allocation size */
+	u32 stat_len;			/* length of status page at ring end */
+	u32 pktshift;			/* padding between CPL & packet data */
+	u32 fl_align;			/* response queue message alignment */
+	u32 fl_starve_thres;		/* Free List starvation threshold */
+
+	/*
+	 * Reverse maps from Absolute Queue IDs to associated queue pointers.
+	 * The absolute Queue IDs are in a compact range which start at a
+	 * [potentially large] Base Queue ID.  We perform the reverse map by
+	 * first converting the Absolute Queue ID into a Relative Queue ID by
+	 * subtracting off the Base Queue ID and then use a Relative Queue ID
+	 * indexed table to get the pointer to the corresponding software
+	 * queue structure.
+	 */
+	unsigned int egr_base;
+	unsigned int ingr_base;
+	void *egr_map[MAX_EGRQ];
+	struct sge_rspq *ingr_map[MAX_INGQ];
+};
+
+/*
+ * Utility macros to convert Absolute- to Relative-Queue indices and Egress-
+ * and Ingress-Queues.  The EQ_MAP() and IQ_MAP() macros which provide
+ * pointers to Ingress- and Egress-Queues can be used as both L- and R-values
+ */
+#define EQ_IDX(s, abs_id) ((unsigned int)((abs_id) - (s)->egr_base))
+#define IQ_IDX(s, abs_id) ((unsigned int)((abs_id) - (s)->ingr_base))
+
+#define EQ_MAP(s, abs_id) ((s)->egr_map[EQ_IDX(s, abs_id)])
+#define IQ_MAP(s, abs_id) ((s)->ingr_map[IQ_IDX(s, abs_id)])
+
+/*
+ * Macro to iterate across Queue Sets ("rxq" is a historic misnomer).
+ */
+#define for_each_ethrxq(sge, iter) \
+	for (iter = 0; iter < (sge)->ethqsets; iter++)
+
+/*
+ * Per-"adapter" (Virtual Function) information.
+ */
+struct adapter {
+	/* PCI resources */
+	void __iomem *regs;
+	void __iomem *bar2;
+	struct pci_dev *pdev;
+	struct device *pdev_dev;
+
+	/* "adapter" resources */
+	unsigned long registered_device_map;
+	unsigned long open_device_map;
+	unsigned long flags;
+	struct adapter_params params;
+
+	/* queue and interrupt resources */
+	struct {
+		unsigned short vec;
+		char desc[22];
+	} msix_info[MSIX_ENTRIES];
+	struct sge sge;
+
+	/* Linux network device resources */
+	struct net_device *port[MAX_NPORTS];
+	const char *name;
+	unsigned int msg_enable;
+
+	/* debugfs resources */
+	struct dentry *debugfs_root;
+
+	/* various locks */
+	spinlock_t stats_lock;
+};
+
+enum { /* adapter flags */
+	FULL_INIT_DONE     = (1UL << 0),
+	USING_MSI          = (1UL << 1),
+	USING_MSIX         = (1UL << 2),
+	QUEUES_BOUND       = (1UL << 3),
+};
+
+/*
+ * The following register read/write routine definitions are required by
+ * the common code.
+ */
+
+/**
+ * t4_read_reg - read a HW register
+ * @adapter: the adapter
+ * @reg_addr: the register address
+ *
+ * Returns the 32-bit value of the given HW register.
+ */
+static inline u32 t4_read_reg(struct adapter *adapter, u32 reg_addr)
+{
+	return readl(adapter->regs + reg_addr);
+}
+
+/**
+ * t4_write_reg - write a HW register
+ * @adapter: the adapter
+ * @reg_addr: the register address
+ * @val: the value to write
+ *
+ * Write a 32-bit value into the given HW register.
+ */
+static inline void t4_write_reg(struct adapter *adapter, u32 reg_addr, u32 val)
+{
+	writel(val, adapter->regs + reg_addr);
+}
+
+#ifndef readq
+static inline u64 readq(const volatile void __iomem *addr)
+{
+	return readl(addr) + ((u64)readl(addr + 4) << 32);
+}
+
+static inline void writeq(u64 val, volatile void __iomem *addr)
+{
+	writel(val, addr);
+	writel(val >> 32, addr + 4);
+}
+#endif
+
+/**
+ * t4_read_reg64 - read a 64-bit HW register
+ * @adapter: the adapter
+ * @reg_addr: the register address
+ *
+ * Returns the 64-bit value of the given HW register.
+ */
+static inline u64 t4_read_reg64(struct adapter *adapter, u32 reg_addr)
+{
+	return readq(adapter->regs + reg_addr);
+}
+
+/**
+ * t4_write_reg64 - write a 64-bit HW register
+ * @adapter: the adapter
+ * @reg_addr: the register address
+ * @val: the value to write
+ *
+ * Write a 64-bit value into the given HW register.
+ */
+static inline void t4_write_reg64(struct adapter *adapter, u32 reg_addr,
+				  u64 val)
+{
+	writeq(val, adapter->regs + reg_addr);
+}
+
+/**
+ * port_name - return the string name of a port
+ * @adapter: the adapter
+ * @pidx: the port index
+ *
+ * Return the string name of the selected port.
+ */
+static inline const char *port_name(struct adapter *adapter, int pidx)
+{
+	return adapter->port[pidx]->name;
+}
+
+/**
+ * t4_os_set_hw_addr - store a port's MAC address in SW
+ * @adapter: the adapter
+ * @pidx: the port index
+ * @hw_addr: the Ethernet address
+ *
+ * Store the Ethernet address of the given port in SW.  Called by the common
+ * code when it retrieves a port's Ethernet address from EEPROM.
+ */
+static inline void t4_os_set_hw_addr(struct adapter *adapter, int pidx,
+				     u8 hw_addr[])
+{
+	memcpy(adapter->port[pidx]->dev_addr, hw_addr, ETH_ALEN);
+}
+
+/**
+ * netdev2pinfo - return the port_info structure associated with a net_device
+ * @dev: the netdev
+ *
+ * Return the struct port_info associated with a net_device
+ */
+static inline struct port_info *netdev2pinfo(const struct net_device *dev)
+{
+	return netdev_priv(dev);
+}
+
+/**
+ * adap2pinfo - return the port_info of a port
+ * @adap: the adapter
+ * @pidx: the port index
+ *
+ * Return the port_info structure for the adapter.
+ */
+static inline struct port_info *adap2pinfo(struct adapter *adapter, int pidx)
+{
+	return netdev_priv(adapter->port[pidx]);
+}
+
+/**
+ * netdev2adap - return the adapter structure associated with a net_device
+ * @dev: the netdev
+ *
+ * Return the struct adapter associated with a net_device
+ */
+static inline struct adapter *netdev2adap(const struct net_device *dev)
+{
+	return netdev2pinfo(dev)->adapter;
+}
+
+/*
+ * OS "Callback" function declarations.  These are functions that the OS code
+ * is "contracted" to provide for the common code.
+ */
+void t4vf_os_link_changed(struct adapter *, int, int);
+void t4vf_os_portmod_changed(struct adapter *, int);
+
+/*
+ * SGE function prototype declarations.
+ */
+int t4vf_sge_alloc_rxq(struct adapter *, struct sge_rspq *, bool,
+		       struct net_device *, int,
+		       struct sge_fl *, rspq_handler_t);
+int t4vf_sge_alloc_eth_txq(struct adapter *, struct sge_eth_txq *,
+			   struct net_device *, struct netdev_queue *,
+			   unsigned int);
+void t4vf_free_sge_resources(struct adapter *);
+
+int t4vf_eth_xmit(struct sk_buff *, struct net_device *);
+int t4vf_ethrx_handler(struct sge_rspq *, const __be64 *,
+		       const struct pkt_gl *);
+
+irq_handler_t t4vf_intr_handler(struct adapter *);
+irqreturn_t t4vf_sge_intr_msix(int, void *);
+
+int t4vf_sge_init(struct adapter *);
+void t4vf_sge_start(struct adapter *);
+void t4vf_sge_stop(struct adapter *);
+
+#endif /* __CXGB4VF_ADAPTER_H__ */
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/cxgb4vf_main.c b/drivers/net/ethernet/chelsio/cxgb4vf/cxgb4vf_main.c
new file mode 100644
index 000000000..1d893b0b7
--- /dev/null
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/cxgb4vf_main.c
@@ -0,0 +1,3098 @@
+/*
+ * This file is part of the Chelsio T4 PCI-E SR-IOV Virtual Function Ethernet
+ * driver for Linux.
+ *
+ * Copyright (c) 2009-2010 Chelsio Communications, 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.
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/module.h>
+#include <linux/moduleparam.h>
+#include <linux/init.h>
+#include <linux/pci.h>
+#include <linux/dma-mapping.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/debugfs.h>
+#include <linux/ethtool.h>
+#include <linux/mdio.h>
+
+#include "t4vf_common.h"
+#include "t4vf_defs.h"
+
+#include "../cxgb4/t4_regs.h"
+#include "../cxgb4/t4_msg.h"
+
+/*
+ * Generic information about the driver.
+ */
+#define DRV_VERSION "2.0.0-ko"
+#define DRV_DESC "Chelsio T4/T5 Virtual Function (VF) Network Driver"
+
+/*
+ * Module Parameters.
+ * ==================
+ */
+
+/*
+ * Default ethtool "message level" for adapters.
+ */
+#define DFLT_MSG_ENABLE (NETIF_MSG_DRV | NETIF_MSG_PROBE | NETIF_MSG_LINK | \
+			 NETIF_MSG_TIMER | NETIF_MSG_IFDOWN | NETIF_MSG_IFUP |\
+			 NETIF_MSG_RX_ERR | NETIF_MSG_TX_ERR)
+
+static int dflt_msg_enable = DFLT_MSG_ENABLE;
+
+module_param(dflt_msg_enable, int, 0644);
+MODULE_PARM_DESC(dflt_msg_enable,
+		 "default adapter ethtool message level bitmap");
+
+/*
+ * The driver uses the best interrupt scheme available on a platform in the
+ * order MSI-X then MSI.  This parameter determines which of these schemes the
+ * driver may consider as follows:
+ *
+ *     msi = 2: choose from among MSI-X and MSI
+ *     msi = 1: only consider MSI interrupts
+ *
+ * Note that unlike the Physical Function driver, this Virtual Function driver
+ * does _not_ support legacy INTx interrupts (this limitation is mandated by
+ * the PCI-E SR-IOV standard).
+ */
+#define MSI_MSIX	2
+#define MSI_MSI		1
+#define MSI_DEFAULT	MSI_MSIX
+
+static int msi = MSI_DEFAULT;
+
+module_param(msi, int, 0644);
+MODULE_PARM_DESC(msi, "whether to use MSI-X or MSI");
+
+/*
+ * Fundamental constants.
+ * ======================
+ */
+
+enum {
+	MAX_TXQ_ENTRIES		= 16384,
+	MAX_RSPQ_ENTRIES	= 16384,
+	MAX_RX_BUFFERS		= 16384,
+
+	MIN_TXQ_ENTRIES		= 32,
+	MIN_RSPQ_ENTRIES	= 128,
+	MIN_FL_ENTRIES		= 16,
+
+	/*
+	 * For purposes of manipulating the Free List size we need to
+	 * recognize that Free Lists are actually Egress Queues (the host
+	 * produces free buffers which the hardware consumes), Egress Queues
+	 * indices are all in units of Egress Context Units bytes, and free
+	 * list entries are 64-bit PCI DMA addresses.  And since the state of
+	 * the Producer Index == the Consumer Index implies an EMPTY list, we
+	 * always have at least one Egress Unit's worth of Free List entries
+	 * unused.  See sge.c for more details ...
+	 */
+	EQ_UNIT = SGE_EQ_IDXSIZE,
+	FL_PER_EQ_UNIT = EQ_UNIT / sizeof(__be64),
+	MIN_FL_RESID = FL_PER_EQ_UNIT,
+};
+
+/*
+ * Global driver state.
+ * ====================
+ */
+
+static struct dentry *cxgb4vf_debugfs_root;
+
+/*
+ * OS "Callback" functions.
+ * ========================
+ */
+
+/*
+ * The link status has changed on the indicated "port" (Virtual Interface).
+ */
+void t4vf_os_link_changed(struct adapter *adapter, int pidx, int link_ok)
+{
+	struct net_device *dev = adapter->port[pidx];
+
+	/*
+	 * If the port is disabled or the current recorded "link up"
+	 * status matches the new status, just return.
+	 */
+	if (!netif_running(dev) || link_ok == netif_carrier_ok(dev))
+		return;
+
+	/*
+	 * Tell the OS that the link status has changed and print a short
+	 * informative message on the console about the event.
+	 */
+	if (link_ok) {
+		const char *s;
+		const char *fc;
+		const struct port_info *pi = netdev_priv(dev);
+
+		netif_carrier_on(dev);
+
+		switch (pi->link_cfg.speed) {
+		case 40000:
+			s = "40Gbps";
+			break;
+
+		case 10000:
+			s = "10Gbps";
+			break;
+
+		case 1000:
+			s = "1000Mbps";
+			break;
+
+		case 100:
+			s = "100Mbps";
+			break;
+
+		default:
+			s = "unknown";
+			break;
+		}
+
+		switch (pi->link_cfg.fc) {
+		case PAUSE_RX:
+			fc = "RX";
+			break;
+
+		case PAUSE_TX:
+			fc = "TX";
+			break;
+
+		case PAUSE_RX|PAUSE_TX:
+			fc = "RX/TX";
+			break;
+
+		default:
+			fc = "no";
+			break;
+		}
+
+		netdev_info(dev, "link up, %s, full-duplex, %s PAUSE\n", s, fc);
+	} else {
+		netif_carrier_off(dev);
+		netdev_info(dev, "link down\n");
+	}
+}
+
+/*
+ * THe port module type has changed on the indicated "port" (Virtual
+ * Interface).
+ */
+void t4vf_os_portmod_changed(struct adapter *adapter, int pidx)
+{
+	static const char * const mod_str[] = {
+		NULL, "LR", "SR", "ER", "passive DA", "active DA", "LRM"
+	};
+	const struct net_device *dev = adapter->port[pidx];
+	const struct port_info *pi = netdev_priv(dev);
+
+	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
+		dev_info(adapter->pdev_dev, "%s: port module unplugged\n",
+			 dev->name);
+	else if (pi->mod_type < ARRAY_SIZE(mod_str))
+		dev_info(adapter->pdev_dev, "%s: %s port module inserted\n",
+			 dev->name, mod_str[pi->mod_type]);
+	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
+		dev_info(adapter->pdev_dev, "%s: unsupported optical port "
+			 "module inserted\n", dev->name);
+	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
+		dev_info(adapter->pdev_dev, "%s: unknown port module inserted,"
+			 "forcing TWINAX\n", dev->name);
+	else if (pi->mod_type == FW_PORT_MOD_TYPE_ERROR)
+		dev_info(adapter->pdev_dev, "%s: transceiver module error\n",
+			 dev->name);
+	else
+		dev_info(adapter->pdev_dev, "%s: unknown module type %d "
+			 "inserted\n", dev->name, pi->mod_type);
+}
+
+/*
+ * Net device operations.
+ * ======================
+ */
+
+
+
+
+/*
+ * Perform the MAC and PHY actions needed to enable a "port" (Virtual
+ * Interface).
+ */
+static int link_start(struct net_device *dev)
+{
+	int ret;
+	struct port_info *pi = netdev_priv(dev);
+
+	/*
+	 * We do not set address filters and promiscuity here, the stack does
+	 * that step explicitly. Enable vlan accel.
+	 */
+	ret = t4vf_set_rxmode(pi->adapter, pi->viid, dev->mtu, -1, -1, -1, 1,
+			      true);
+	if (ret == 0) {
+		ret = t4vf_change_mac(pi->adapter, pi->viid,
+				      pi->xact_addr_filt, dev->dev_addr, true);
+		if (ret >= 0) {
+			pi->xact_addr_filt = ret;
+			ret = 0;
+		}
+	}
+
+	/*
+	 * We don't need to actually "start the link" itself since the
+	 * firmware will do that for us when the first Virtual Interface
+	 * is enabled on a port.
+	 */
+	if (ret == 0)
+		ret = t4vf_enable_vi(pi->adapter, pi->viid, true, true);
+	return ret;
+}
+
+/*
+ * Name the MSI-X interrupts.
+ */
+static void name_msix_vecs(struct adapter *adapter)
+{
+	int namelen = sizeof(adapter->msix_info[0].desc) - 1;
+	int pidx;
+
+	/*
+	 * Firmware events.
+	 */
+	snprintf(adapter->msix_info[MSIX_FW].desc, namelen,
+		 "%s-FWeventq", adapter->name);
+	adapter->msix_info[MSIX_FW].desc[namelen] = 0;
+
+	/*
+	 * Ethernet queues.
+	 */
+	for_each_port(adapter, pidx) {
+		struct net_device *dev = adapter->port[pidx];
+		const struct port_info *pi = netdev_priv(dev);
+		int qs, msi;
+
+		for (qs = 0, msi = MSIX_IQFLINT; qs < pi->nqsets; qs++, msi++) {
+			snprintf(adapter->msix_info[msi].desc, namelen,
+				 "%s-%d", dev->name, qs);
+			adapter->msix_info[msi].desc[namelen] = 0;
+		}
+	}
+}
+
+/*
+ * Request all of our MSI-X resources.
+ */
+static int request_msix_queue_irqs(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+	int rxq, msi, err;
+
+	/*
+	 * Firmware events.
+	 */
+	err = request_irq(adapter->msix_info[MSIX_FW].vec, t4vf_sge_intr_msix,
+			  0, adapter->msix_info[MSIX_FW].desc, &s->fw_evtq);
+	if (err)
+		return err;
+
+	/*
+	 * Ethernet queues.
+	 */
+	msi = MSIX_IQFLINT;
+	for_each_ethrxq(s, rxq) {
+		err = request_irq(adapter->msix_info[msi].vec,
+				  t4vf_sge_intr_msix, 0,
+				  adapter->msix_info[msi].desc,
+				  &s->ethrxq[rxq].rspq);
+		if (err)
+			goto err_free_irqs;
+		msi++;
+	}
+	return 0;
+
+err_free_irqs:
+	while (--rxq >= 0)
+		free_irq(adapter->msix_info[--msi].vec, &s->ethrxq[rxq].rspq);
+	free_irq(adapter->msix_info[MSIX_FW].vec, &s->fw_evtq);
+	return err;
+}
+
+/*
+ * Free our MSI-X resources.
+ */
+static void free_msix_queue_irqs(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+	int rxq, msi;
+
+	free_irq(adapter->msix_info[MSIX_FW].vec, &s->fw_evtq);
+	msi = MSIX_IQFLINT;
+	for_each_ethrxq(s, rxq)
+		free_irq(adapter->msix_info[msi++].vec,
+			 &s->ethrxq[rxq].rspq);
+}
+
+/*
+ * Turn on NAPI and start up interrupts on a response queue.
+ */
+static void qenable(struct sge_rspq *rspq)
+{
+	napi_enable(&rspq->napi);
+
+	/*
+	 * 0-increment the Going To Sleep register to start the timer and
+	 * enable interrupts.
+	 */
+	t4_write_reg(rspq->adapter, T4VF_SGE_BASE_ADDR + SGE_VF_GTS,
+		     CIDXINC_V(0) |
+		     SEINTARM_V(rspq->intr_params) |
+		     INGRESSQID_V(rspq->cntxt_id));
+}
+
+/*
+ * Enable NAPI scheduling and interrupt generation for all Receive Queues.
+ */
+static void enable_rx(struct adapter *adapter)
+{
+	int rxq;
+	struct sge *s = &adapter->sge;
+
+	for_each_ethrxq(s, rxq)
+		qenable(&s->ethrxq[rxq].rspq);
+	qenable(&s->fw_evtq);
+
+	/*
+	 * The interrupt queue doesn't use NAPI so we do the 0-increment of
+	 * its Going To Sleep register here to get it started.
+	 */
+	if (adapter->flags & USING_MSI)
+		t4_write_reg(adapter, T4VF_SGE_BASE_ADDR + SGE_VF_GTS,
+			     CIDXINC_V(0) |
+			     SEINTARM_V(s->intrq.intr_params) |
+			     INGRESSQID_V(s->intrq.cntxt_id));
+
+}
+
+/*
+ * Wait until all NAPI handlers are descheduled.
+ */
+static void quiesce_rx(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+	int rxq;
+
+	for_each_ethrxq(s, rxq)
+		napi_disable(&s->ethrxq[rxq].rspq.napi);
+	napi_disable(&s->fw_evtq.napi);
+}
+
+/*
+ * Response queue handler for the firmware event queue.
+ */
+static int fwevtq_handler(struct sge_rspq *rspq, const __be64 *rsp,
+			  const struct pkt_gl *gl)
+{
+	/*
+	 * Extract response opcode and get pointer to CPL message body.
+	 */
+	struct adapter *adapter = rspq->adapter;
+	u8 opcode = ((const struct rss_header *)rsp)->opcode;
+	void *cpl = (void *)(rsp + 1);
+
+	switch (opcode) {
+	case CPL_FW6_MSG: {
+		/*
+		 * We've received an asynchronous message from the firmware.
+		 */
+		const struct cpl_fw6_msg *fw_msg = cpl;
+		if (fw_msg->type == FW6_TYPE_CMD_RPL)
+			t4vf_handle_fw_rpl(adapter, fw_msg->data);
+		break;
+	}
+
+	case CPL_FW4_MSG: {
+		/* FW can send EGR_UPDATEs encapsulated in a CPL_FW4_MSG.
+		 */
+		const struct cpl_sge_egr_update *p = (void *)(rsp + 3);
+		opcode = CPL_OPCODE_G(ntohl(p->opcode_qid));
+		if (opcode != CPL_SGE_EGR_UPDATE) {
+			dev_err(adapter->pdev_dev, "unexpected FW4/CPL %#x on FW event queue\n"
+				, opcode);
+			break;
+		}
+		cpl = (void *)p;
+		/*FALLTHROUGH*/
+	}
+
+	case CPL_SGE_EGR_UPDATE: {
+		/*
+		 * We've received an Egress Queue Status Update message.  We
+		 * get these, if the SGE is configured to send these when the
+		 * firmware passes certain points in processing our TX
+		 * Ethernet Queue or if we make an explicit request for one.
+		 * We use these updates to determine when we may need to
+		 * restart a TX Ethernet Queue which was stopped for lack of
+		 * free TX Queue Descriptors ...
+		 */
+		const struct cpl_sge_egr_update *p = cpl;
+		unsigned int qid = EGR_QID_G(be32_to_cpu(p->opcode_qid));
+		struct sge *s = &adapter->sge;
+		struct sge_txq *tq;
+		struct sge_eth_txq *txq;
+		unsigned int eq_idx;
+
+		/*
+		 * Perform sanity checking on the Queue ID to make sure it
+		 * really refers to one of our TX Ethernet Egress Queues which
+		 * is active and matches the queue's ID.  None of these error
+		 * conditions should ever happen so we may want to either make
+		 * them fatal and/or conditionalized under DEBUG.
+		 */
+		eq_idx = EQ_IDX(s, qid);
+		if (unlikely(eq_idx >= MAX_EGRQ)) {
+			dev_err(adapter->pdev_dev,
+				"Egress Update QID %d out of range\n", qid);
+			break;
+		}
+		tq = s->egr_map[eq_idx];
+		if (unlikely(tq == NULL)) {
+			dev_err(adapter->pdev_dev,
+				"Egress Update QID %d TXQ=NULL\n", qid);
+			break;
+		}
+		txq = container_of(tq, struct sge_eth_txq, q);
+		if (unlikely(tq->abs_id != qid)) {
+			dev_err(adapter->pdev_dev,
+				"Egress Update QID %d refers to TXQ %d\n",
+				qid, tq->abs_id);
+			break;
+		}
+
+		/*
+		 * Restart a stopped TX Queue which has less than half of its
+		 * TX ring in use ...
+		 */
+		txq->q.restarts++;
+		netif_tx_wake_queue(txq->txq);
+		break;
+	}
+
+	default:
+		dev_err(adapter->pdev_dev,
+			"unexpected CPL %#x on FW event queue\n", opcode);
+	}
+
+	return 0;
+}
+
+/*
+ * Allocate SGE TX/RX response queues.  Determine how many sets of SGE queues
+ * to use and initializes them.  We support multiple "Queue Sets" per port if
+ * we have MSI-X, otherwise just one queue set per port.
+ */
+static int setup_sge_queues(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+	int err, pidx, msix;
+
+	/*
+	 * Clear "Queue Set" Free List Starving and TX Queue Mapping Error
+	 * state.
+	 */
+	bitmap_zero(s->starving_fl, MAX_EGRQ);
+
+	/*
+	 * If we're using MSI interrupt mode we need to set up a "forwarded
+	 * interrupt" queue which we'll set up with our MSI vector.  The rest
+	 * of the ingress queues will be set up to forward their interrupts to
+	 * this queue ...  This must be first since t4vf_sge_alloc_rxq() uses
+	 * the intrq's queue ID as the interrupt forwarding queue for the
+	 * subsequent calls ...
+	 */
+	if (adapter->flags & USING_MSI) {
+		err = t4vf_sge_alloc_rxq(adapter, &s->intrq, false,
+					 adapter->port[0], 0, NULL, NULL);
+		if (err)
+			goto err_free_queues;
+	}
+
+	/*
+	 * Allocate our ingress queue for asynchronous firmware messages.
+	 */
+	err = t4vf_sge_alloc_rxq(adapter, &s->fw_evtq, true, adapter->port[0],
+				 MSIX_FW, NULL, fwevtq_handler);
+	if (err)
+		goto err_free_queues;
+
+	/*
+	 * Allocate each "port"'s initial Queue Sets.  These can be changed
+	 * later on ... up to the point where any interface on the adapter is
+	 * brought up at which point lots of things get nailed down
+	 * permanently ...
+	 */
+	msix = MSIX_IQFLINT;
+	for_each_port(adapter, pidx) {
+		struct net_device *dev = adapter->port[pidx];
+		struct port_info *pi = netdev_priv(dev);
+		struct sge_eth_rxq *rxq = &s->ethrxq[pi->first_qset];
+		struct sge_eth_txq *txq = &s->ethtxq[pi->first_qset];
+		int qs;
+
+		for (qs = 0; qs < pi->nqsets; qs++, rxq++, txq++) {
+			err = t4vf_sge_alloc_rxq(adapter, &rxq->rspq, false,
+						 dev, msix++,
+						 &rxq->fl, t4vf_ethrx_handler);
+			if (err)
+				goto err_free_queues;
+
+			err = t4vf_sge_alloc_eth_txq(adapter, txq, dev,
+					     netdev_get_tx_queue(dev, qs),
+					     s->fw_evtq.cntxt_id);
+			if (err)
+				goto err_free_queues;
+
+			rxq->rspq.idx = qs;
+			memset(&rxq->stats, 0, sizeof(rxq->stats));
+		}
+	}
+
+	/*
+	 * Create the reverse mappings for the queues.
+	 */
+	s->egr_base = s->ethtxq[0].q.abs_id - s->ethtxq[0].q.cntxt_id;
+	s->ingr_base = s->ethrxq[0].rspq.abs_id - s->ethrxq[0].rspq.cntxt_id;
+	IQ_MAP(s, s->fw_evtq.abs_id) = &s->fw_evtq;
+	for_each_port(adapter, pidx) {
+		struct net_device *dev = adapter->port[pidx];
+		struct port_info *pi = netdev_priv(dev);
+		struct sge_eth_rxq *rxq = &s->ethrxq[pi->first_qset];
+		struct sge_eth_txq *txq = &s->ethtxq[pi->first_qset];
+		int qs;
+
+		for (qs = 0; qs < pi->nqsets; qs++, rxq++, txq++) {
+			IQ_MAP(s, rxq->rspq.abs_id) = &rxq->rspq;
+			EQ_MAP(s, txq->q.abs_id) = &txq->q;
+
+			/*
+			 * The FW_IQ_CMD doesn't return the Absolute Queue IDs
+			 * for Free Lists but since all of the Egress Queues
+			 * (including Free Lists) have Relative Queue IDs
+			 * which are computed as Absolute - Base Queue ID, we
+			 * can synthesize the Absolute Queue IDs for the Free
+			 * Lists.  This is useful for debugging purposes when
+			 * we want to dump Queue Contexts via the PF Driver.
+			 */
+			rxq->fl.abs_id = rxq->fl.cntxt_id + s->egr_base;
+			EQ_MAP(s, rxq->fl.abs_id) = &rxq->fl;
+		}
+	}
+	return 0;
+
+err_free_queues:
+	t4vf_free_sge_resources(adapter);
+	return err;
+}
+
+/*
+ * Set up Receive Side Scaling (RSS) to distribute packets to multiple receive
+ * queues.  We configure the RSS CPU lookup table to distribute to the number
+ * of HW receive queues, and the response queue lookup table to narrow that
+ * down to the response queues actually configured for each "port" (Virtual
+ * Interface).  We always configure the RSS mapping for all ports since the
+ * mapping table has plenty of entries.
+ */
+static int setup_rss(struct adapter *adapter)
+{
+	int pidx;
+
+	for_each_port(adapter, pidx) {
+		struct port_info *pi = adap2pinfo(adapter, pidx);
+		struct sge_eth_rxq *rxq = &adapter->sge.ethrxq[pi->first_qset];
+		u16 rss[MAX_PORT_QSETS];
+		int qs, err;
+
+		for (qs = 0; qs < pi->nqsets; qs++)
+			rss[qs] = rxq[qs].rspq.abs_id;
+
+		err = t4vf_config_rss_range(adapter, pi->viid,
+					    0, pi->rss_size, rss, pi->nqsets);
+		if (err)
+			return err;
+
+		/*
+		 * Perform Global RSS Mode-specific initialization.
+		 */
+		switch (adapter->params.rss.mode) {
+		case FW_RSS_GLB_CONFIG_CMD_MODE_BASICVIRTUAL:
+			/*
+			 * If Tunnel All Lookup isn't specified in the global
+			 * RSS Configuration, then we need to specify a
+			 * default Ingress Queue for any ingress packets which
+			 * aren't hashed.  We'll use our first ingress queue
+			 * ...
+			 */
+			if (!adapter->params.rss.u.basicvirtual.tnlalllookup) {
+				union rss_vi_config config;
+				err = t4vf_read_rss_vi_config(adapter,
+							      pi->viid,
+							      &config);
+				if (err)
+					return err;
+				config.basicvirtual.defaultq =
+					rxq[0].rspq.abs_id;
+				err = t4vf_write_rss_vi_config(adapter,
+							       pi->viid,
+							       &config);
+				if (err)
+					return err;
+			}
+			break;
+		}
+	}
+
+	return 0;
+}
+
+/*
+ * Bring the adapter up.  Called whenever we go from no "ports" open to having
+ * one open.  This function performs the actions necessary to make an adapter
+ * operational, such as completing the initialization of HW modules, and
+ * enabling interrupts.  Must be called with the rtnl lock held.  (Note that
+ * this is called "cxgb_up" in the PF Driver.)
+ */
+static int adapter_up(struct adapter *adapter)
+{
+	int err;
+
+	/*
+	 * If this is the first time we've been called, perform basic
+	 * adapter setup.  Once we've done this, many of our adapter
+	 * parameters can no longer be changed ...
+	 */
+	if ((adapter->flags & FULL_INIT_DONE) == 0) {
+		err = setup_sge_queues(adapter);
+		if (err)
+			return err;
+		err = setup_rss(adapter);
+		if (err) {
+			t4vf_free_sge_resources(adapter);
+			return err;
+		}
+
+		if (adapter->flags & USING_MSIX)
+			name_msix_vecs(adapter);
+		adapter->flags |= FULL_INIT_DONE;
+	}
+
+	/*
+	 * Acquire our interrupt resources.  We only support MSI-X and MSI.
+	 */
+	BUG_ON((adapter->flags & (USING_MSIX|USING_MSI)) == 0);
+	if (adapter->flags & USING_MSIX)
+		err = request_msix_queue_irqs(adapter);
+	else
+		err = request_irq(adapter->pdev->irq,
+				  t4vf_intr_handler(adapter), 0,
+				  adapter->name, adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "request_irq failed, err %d\n",
+			err);
+		return err;
+	}
+
+	/*
+	 * Enable NAPI ingress processing and return success.
+	 */
+	enable_rx(adapter);
+	t4vf_sge_start(adapter);
+	return 0;
+}
+
+/*
+ * Bring the adapter down.  Called whenever the last "port" (Virtual
+ * Interface) closed.  (Note that this routine is called "cxgb_down" in the PF
+ * Driver.)
+ */
+static void adapter_down(struct adapter *adapter)
+{
+	/*
+	 * Free interrupt resources.
+	 */
+	if (adapter->flags & USING_MSIX)
+		free_msix_queue_irqs(adapter);
+	else
+		free_irq(adapter->pdev->irq, adapter);
+
+	/*
+	 * Wait for NAPI handlers to finish.
+	 */
+	quiesce_rx(adapter);
+}
+
+/*
+ * Start up a net device.
+ */
+static int cxgb4vf_open(struct net_device *dev)
+{
+	int err;
+	struct port_info *pi = netdev_priv(dev);
+	struct adapter *adapter = pi->adapter;
+
+	/*
+	 * If this is the first interface that we're opening on the "adapter",
+	 * bring the "adapter" up now.
+	 */
+	if (adapter->open_device_map == 0) {
+		err = adapter_up(adapter);
+		if (err)
+			return err;
+	}
+
+	/*
+	 * Note that this interface is up and start everything up ...
+	 */
+	netif_set_real_num_tx_queues(dev, pi->nqsets);
+	err = netif_set_real_num_rx_queues(dev, pi->nqsets);
+	if (err)
+		goto err_unwind;
+	err = link_start(dev);
+	if (err)
+		goto err_unwind;
+
+	netif_tx_start_all_queues(dev);
+	set_bit(pi->port_id, &adapter->open_device_map);
+	return 0;
+
+err_unwind:
+	if (adapter->open_device_map == 0)
+		adapter_down(adapter);
+	return err;
+}
+
+/*
+ * Shut down a net device.  This routine is called "cxgb_close" in the PF
+ * Driver ...
+ */
+static int cxgb4vf_stop(struct net_device *dev)
+{
+	struct port_info *pi = netdev_priv(dev);
+	struct adapter *adapter = pi->adapter;
+
+	netif_tx_stop_all_queues(dev);
+	netif_carrier_off(dev);
+	t4vf_enable_vi(adapter, pi->viid, false, false);
+	pi->link_cfg.link_ok = 0;
+
+	clear_bit(pi->port_id, &adapter->open_device_map);
+	if (adapter->open_device_map == 0)
+		adapter_down(adapter);
+	return 0;
+}
+
+/*
+ * Translate our basic statistics into the standard "ifconfig" statistics.
+ */
+static struct net_device_stats *cxgb4vf_get_stats(struct net_device *dev)
+{
+	struct t4vf_port_stats stats;
+	struct port_info *pi = netdev2pinfo(dev);
+	struct adapter *adapter = pi->adapter;
+	struct net_device_stats *ns = &dev->stats;
+	int err;
+
+	spin_lock(&adapter->stats_lock);
+	err = t4vf_get_port_stats(adapter, pi->pidx, &stats);
+	spin_unlock(&adapter->stats_lock);
+
+	memset(ns, 0, sizeof(*ns));
+	if (err)
+		return ns;
+
+	ns->tx_bytes = (stats.tx_bcast_bytes + stats.tx_mcast_bytes +
+			stats.tx_ucast_bytes + stats.tx_offload_bytes);
+	ns->tx_packets = (stats.tx_bcast_frames + stats.tx_mcast_frames +
+			  stats.tx_ucast_frames + stats.tx_offload_frames);
+	ns->rx_bytes = (stats.rx_bcast_bytes + stats.rx_mcast_bytes +
+			stats.rx_ucast_bytes);
+	ns->rx_packets = (stats.rx_bcast_frames + stats.rx_mcast_frames +
+			  stats.rx_ucast_frames);
+	ns->multicast = stats.rx_mcast_frames;
+	ns->tx_errors = stats.tx_drop_frames;
+	ns->rx_errors = stats.rx_err_frames;
+
+	return ns;
+}
+
+/*
+ * Collect up to maxaddrs worth of a netdevice's unicast addresses, starting
+ * at a specified offset within the list, into an array of addrss pointers and
+ * return the number collected.
+ */
+static inline unsigned int collect_netdev_uc_list_addrs(const struct net_device *dev,
+							const u8 **addr,
+							unsigned int offset,
+							unsigned int maxaddrs)
+{
+	unsigned int index = 0;
+	unsigned int naddr = 0;
+	const struct netdev_hw_addr *ha;
+
+	for_each_dev_addr(dev, ha)
+		if (index++ >= offset) {
+			addr[naddr++] = ha->addr;
+			if (naddr >= maxaddrs)
+				break;
+		}
+	return naddr;
+}
+
+/*
+ * Collect up to maxaddrs worth of a netdevice's multicast addresses, starting
+ * at a specified offset within the list, into an array of addrss pointers and
+ * return the number collected.
+ */
+static inline unsigned int collect_netdev_mc_list_addrs(const struct net_device *dev,
+							const u8 **addr,
+							unsigned int offset,
+							unsigned int maxaddrs)
+{
+	unsigned int index = 0;
+	unsigned int naddr = 0;
+	const struct netdev_hw_addr *ha;
+
+	netdev_for_each_mc_addr(ha, dev)
+		if (index++ >= offset) {
+			addr[naddr++] = ha->addr;
+			if (naddr >= maxaddrs)
+				break;
+		}
+	return naddr;
+}
+
+/*
+ * Configure the exact and hash address filters to handle a port's multicast
+ * and secondary unicast MAC addresses.
+ */
+static int set_addr_filters(const struct net_device *dev, bool sleep)
+{
+	u64 mhash = 0;
+	u64 uhash = 0;
+	bool free = true;
+	unsigned int offset, naddr;
+	const u8 *addr[7];
+	int ret;
+	const struct port_info *pi = netdev_priv(dev);
+
+	/* first do the secondary unicast addresses */
+	for (offset = 0; ; offset += naddr) {
+		naddr = collect_netdev_uc_list_addrs(dev, addr, offset,
+						     ARRAY_SIZE(addr));
+		if (naddr == 0)
+			break;
+
+		ret = t4vf_alloc_mac_filt(pi->adapter, pi->viid, free,
+					  naddr, addr, NULL, &uhash, sleep);
+		if (ret < 0)
+			return ret;
+
+		free = false;
+	}
+
+	/* next set up the multicast addresses */
+	for (offset = 0; ; offset += naddr) {
+		naddr = collect_netdev_mc_list_addrs(dev, addr, offset,
+						     ARRAY_SIZE(addr));
+		if (naddr == 0)
+			break;
+
+		ret = t4vf_alloc_mac_filt(pi->adapter, pi->viid, free,
+					  naddr, addr, NULL, &mhash, sleep);
+		if (ret < 0)
+			return ret;
+		free = false;
+	}
+
+	return t4vf_set_addr_hash(pi->adapter, pi->viid, uhash != 0,
+				  uhash | mhash, sleep);
+}
+
+/*
+ * Set RX properties of a port, such as promiscruity, address filters, and MTU.
+ * If @mtu is -1 it is left unchanged.
+ */
+static int set_rxmode(struct net_device *dev, int mtu, bool sleep_ok)
+{
+	int ret;
+	struct port_info *pi = netdev_priv(dev);
+
+	ret = set_addr_filters(dev, sleep_ok);
+	if (ret == 0)
+		ret = t4vf_set_rxmode(pi->adapter, pi->viid, -1,
+				      (dev->flags & IFF_PROMISC) != 0,
+				      (dev->flags & IFF_ALLMULTI) != 0,
+				      1, -1, sleep_ok);
+	return ret;
+}
+
+/*
+ * Set the current receive modes on the device.
+ */
+static void cxgb4vf_set_rxmode(struct net_device *dev)
+{
+	/* unfortunately we can't return errors to the stack */
+	set_rxmode(dev, -1, false);
+}
+
+/*
+ * Find the entry in the interrupt holdoff timer value array which comes
+ * closest to the specified interrupt holdoff value.
+ */
+static int closest_timer(const struct sge *s, int us)
+{
+	int i, timer_idx = 0, min_delta = INT_MAX;
+
+	for (i = 0; i < ARRAY_SIZE(s->timer_val); i++) {
+		int delta = us - s->timer_val[i];
+		if (delta < 0)
+			delta = -delta;
+		if (delta < min_delta) {
+			min_delta = delta;
+			timer_idx = i;
+		}
+	}
+	return timer_idx;
+}
+
+static int closest_thres(const struct sge *s, int thres)
+{
+	int i, delta, pktcnt_idx = 0, min_delta = INT_MAX;
+
+	for (i = 0; i < ARRAY_SIZE(s->counter_val); i++) {
+		delta = thres - s->counter_val[i];
+		if (delta < 0)
+			delta = -delta;
+		if (delta < min_delta) {
+			min_delta = delta;
+			pktcnt_idx = i;
+		}
+	}
+	return pktcnt_idx;
+}
+
+/*
+ * Return a queue's interrupt hold-off time in us.  0 means no timer.
+ */
+static unsigned int qtimer_val(const struct adapter *adapter,
+			       const struct sge_rspq *rspq)
+{
+	unsigned int timer_idx = QINTR_TIMER_IDX_GET(rspq->intr_params);
+
+	return timer_idx < SGE_NTIMERS
+		? adapter->sge.timer_val[timer_idx]
+		: 0;
+}
+
+/**
+ *	set_rxq_intr_params - set a queue's interrupt holdoff parameters
+ *	@adapter: the adapter
+ *	@rspq: the RX response queue
+ *	@us: the hold-off time in us, or 0 to disable timer
+ *	@cnt: the hold-off packet count, or 0 to disable counter
+ *
+ *	Sets an RX response queue's interrupt hold-off time and packet count.
+ *	At least one of the two needs to be enabled for the queue to generate
+ *	interrupts.
+ */
+static int set_rxq_intr_params(struct adapter *adapter, struct sge_rspq *rspq,
+			       unsigned int us, unsigned int cnt)
+{
+	unsigned int timer_idx;
+
+	/*
+	 * If both the interrupt holdoff timer and count are specified as
+	 * zero, default to a holdoff count of 1 ...
+	 */
+	if ((us | cnt) == 0)
+		cnt = 1;
+
+	/*
+	 * If an interrupt holdoff count has been specified, then find the
+	 * closest configured holdoff count and use that.  If the response
+	 * queue has already been created, then update its queue context
+	 * parameters ...
+	 */
+	if (cnt) {
+		int err;
+		u32 v, pktcnt_idx;
+
+		pktcnt_idx = closest_thres(&adapter->sge, cnt);
+		if (rspq->desc && rspq->pktcnt_idx != pktcnt_idx) {
+			v = FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_DMAQ) |
+			    FW_PARAMS_PARAM_X_V(
+					FW_PARAMS_PARAM_DMAQ_IQ_INTCNTTHRESH) |
+			    FW_PARAMS_PARAM_YZ_V(rspq->cntxt_id);
+			err = t4vf_set_params(adapter, 1, &v, &pktcnt_idx);
+			if (err)
+				return err;
+		}
+		rspq->pktcnt_idx = pktcnt_idx;
+	}
+
+	/*
+	 * Compute the closest holdoff timer index from the supplied holdoff
+	 * timer value.
+	 */
+	timer_idx = (us == 0
+		     ? SGE_TIMER_RSTRT_CNTR
+		     : closest_timer(&adapter->sge, us));
+
+	/*
+	 * Update the response queue's interrupt coalescing parameters and
+	 * return success.
+	 */
+	rspq->intr_params = (QINTR_TIMER_IDX(timer_idx) |
+			     (cnt > 0 ? QINTR_CNT_EN : 0));
+	return 0;
+}
+
+/*
+ * Return a version number to identify the type of adapter.  The scheme is:
+ * - bits 0..9: chip version
+ * - bits 10..15: chip revision
+ */
+static inline unsigned int mk_adap_vers(const struct adapter *adapter)
+{
+	/*
+	 * Chip version 4, revision 0x3f (cxgb4vf).
+	 */
+	return CHELSIO_CHIP_VERSION(adapter->params.chip) | (0x3f << 10);
+}
+
+/*
+ * Execute the specified ioctl command.
+ */
+static int cxgb4vf_do_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
+{
+	int ret = 0;
+
+	switch (cmd) {
+	    /*
+	     * The VF Driver doesn't have access to any of the other
+	     * common Ethernet device ioctl()'s (like reading/writing
+	     * PHY registers, etc.
+	     */
+
+	default:
+		ret = -EOPNOTSUPP;
+		break;
+	}
+	return ret;
+}
+
+/*
+ * Change the device's MTU.
+ */
+static int cxgb4vf_change_mtu(struct net_device *dev, int new_mtu)
+{
+	int ret;
+	struct port_info *pi = netdev_priv(dev);
+
+	/* accommodate SACK */
+	if (new_mtu < 81)
+		return -EINVAL;
+
+	ret = t4vf_set_rxmode(pi->adapter, pi->viid, new_mtu,
+			      -1, -1, -1, -1, true);
+	if (!ret)
+		dev->mtu = new_mtu;
+	return ret;
+}
+
+static netdev_features_t cxgb4vf_fix_features(struct net_device *dev,
+	netdev_features_t features)
+{
+	/*
+	 * Since there is no support for separate rx/tx vlan accel
+	 * enable/disable make sure tx flag is always in same state as rx.
+	 */
+	if (features & NETIF_F_HW_VLAN_CTAG_RX)
+		features |= NETIF_F_HW_VLAN_CTAG_TX;
+	else
+		features &= ~NETIF_F_HW_VLAN_CTAG_TX;
+
+	return features;
+}
+
+static int cxgb4vf_set_features(struct net_device *dev,
+	netdev_features_t features)
+{
+	struct port_info *pi = netdev_priv(dev);
+	netdev_features_t changed = dev->features ^ features;
+
+	if (changed & NETIF_F_HW_VLAN_CTAG_RX)
+		t4vf_set_rxmode(pi->adapter, pi->viid, -1, -1, -1, -1,
+				features & NETIF_F_HW_VLAN_CTAG_TX, 0);
+
+	return 0;
+}
+
+/*
+ * Change the devices MAC address.
+ */
+static int cxgb4vf_set_mac_addr(struct net_device *dev, void *_addr)
+{
+	int ret;
+	struct sockaddr *addr = _addr;
+	struct port_info *pi = netdev_priv(dev);
+
+	if (!is_valid_ether_addr(addr->sa_data))
+		return -EADDRNOTAVAIL;
+
+	ret = t4vf_change_mac(pi->adapter, pi->viid, pi->xact_addr_filt,
+			      addr->sa_data, true);
+	if (ret < 0)
+		return ret;
+
+	memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
+	pi->xact_addr_filt = ret;
+	return 0;
+}
+
+#ifdef CONFIG_NET_POLL_CONTROLLER
+/*
+ * Poll all of our receive queues.  This is called outside of normal interrupt
+ * context.
+ */
+static void cxgb4vf_poll_controller(struct net_device *dev)
+{
+	struct port_info *pi = netdev_priv(dev);
+	struct adapter *adapter = pi->adapter;
+
+	if (adapter->flags & USING_MSIX) {
+		struct sge_eth_rxq *rxq;
+		int nqsets;
+
+		rxq = &adapter->sge.ethrxq[pi->first_qset];
+		for (nqsets = pi->nqsets; nqsets; nqsets--) {
+			t4vf_sge_intr_msix(0, &rxq->rspq);
+			rxq++;
+		}
+	} else
+		t4vf_intr_handler(adapter)(0, adapter);
+}
+#endif
+
+/*
+ * Ethtool operations.
+ * ===================
+ *
+ * Note that we don't support any ethtool operations which change the physical
+ * state of the port to which we're linked.
+ */
+
+static unsigned int t4vf_from_fw_linkcaps(enum fw_port_type type,
+					  unsigned int caps)
+{
+	unsigned int v = 0;
+
+	if (type == FW_PORT_TYPE_BT_SGMII || type == FW_PORT_TYPE_BT_XFI ||
+	    type == FW_PORT_TYPE_BT_XAUI) {
+		v |= SUPPORTED_TP;
+		if (caps & FW_PORT_CAP_SPEED_100M)
+			v |= SUPPORTED_100baseT_Full;
+		if (caps & FW_PORT_CAP_SPEED_1G)
+			v |= SUPPORTED_1000baseT_Full;
+		if (caps & FW_PORT_CAP_SPEED_10G)
+			v |= SUPPORTED_10000baseT_Full;
+	} else if (type == FW_PORT_TYPE_KX4 || type == FW_PORT_TYPE_KX) {
+		v |= SUPPORTED_Backplane;
+		if (caps & FW_PORT_CAP_SPEED_1G)
+			v |= SUPPORTED_1000baseKX_Full;
+		if (caps & FW_PORT_CAP_SPEED_10G)
+			v |= SUPPORTED_10000baseKX4_Full;
+	} else if (type == FW_PORT_TYPE_KR)
+		v |= SUPPORTED_Backplane | SUPPORTED_10000baseKR_Full;
+	else if (type == FW_PORT_TYPE_BP_AP)
+		v |= SUPPORTED_Backplane | SUPPORTED_10000baseR_FEC |
+		     SUPPORTED_10000baseKR_Full | SUPPORTED_1000baseKX_Full;
+	else if (type == FW_PORT_TYPE_BP4_AP)
+		v |= SUPPORTED_Backplane | SUPPORTED_10000baseR_FEC |
+		     SUPPORTED_10000baseKR_Full | SUPPORTED_1000baseKX_Full |
+		     SUPPORTED_10000baseKX4_Full;
+	else if (type == FW_PORT_TYPE_FIBER_XFI ||
+		 type == FW_PORT_TYPE_FIBER_XAUI ||
+		 type == FW_PORT_TYPE_SFP ||
+		 type == FW_PORT_TYPE_QSFP_10G ||
+		 type == FW_PORT_TYPE_QSA) {
+		v |= SUPPORTED_FIBRE;
+		if (caps & FW_PORT_CAP_SPEED_1G)
+			v |= SUPPORTED_1000baseT_Full;
+		if (caps & FW_PORT_CAP_SPEED_10G)
+			v |= SUPPORTED_10000baseT_Full;
+	} else if (type == FW_PORT_TYPE_BP40_BA ||
+		   type == FW_PORT_TYPE_QSFP) {
+		v |= SUPPORTED_40000baseSR4_Full;
+		v |= SUPPORTED_FIBRE;
+	}
+
+	if (caps & FW_PORT_CAP_ANEG)
+		v |= SUPPORTED_Autoneg;
+	return v;
+}
+
+static int cxgb4vf_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
+{
+	const struct port_info *p = netdev_priv(dev);
+
+	if (p->port_type == FW_PORT_TYPE_BT_SGMII ||
+	    p->port_type == FW_PORT_TYPE_BT_XFI ||
+	    p->port_type == FW_PORT_TYPE_BT_XAUI)
+		cmd->port = PORT_TP;
+	else if (p->port_type == FW_PORT_TYPE_FIBER_XFI ||
+		 p->port_type == FW_PORT_TYPE_FIBER_XAUI)
+		cmd->port = PORT_FIBRE;
+	else if (p->port_type == FW_PORT_TYPE_SFP ||
+		 p->port_type == FW_PORT_TYPE_QSFP_10G ||
+		 p->port_type == FW_PORT_TYPE_QSA ||
+		 p->port_type == FW_PORT_TYPE_QSFP) {
+		if (p->mod_type == FW_PORT_MOD_TYPE_LR ||
+		    p->mod_type == FW_PORT_MOD_TYPE_SR ||
+		    p->mod_type == FW_PORT_MOD_TYPE_ER ||
+		    p->mod_type == FW_PORT_MOD_TYPE_LRM)
+			cmd->port = PORT_FIBRE;
+		else if (p->mod_type == FW_PORT_MOD_TYPE_TWINAX_PASSIVE ||
+			 p->mod_type == FW_PORT_MOD_TYPE_TWINAX_ACTIVE)
+			cmd->port = PORT_DA;
+		else
+			cmd->port = PORT_OTHER;
+	} else
+		cmd->port = PORT_OTHER;
+
+	if (p->mdio_addr >= 0) {
+		cmd->phy_address = p->mdio_addr;
+		cmd->transceiver = XCVR_EXTERNAL;
+		cmd->mdio_support = p->port_type == FW_PORT_TYPE_BT_SGMII ?
+			MDIO_SUPPORTS_C22 : MDIO_SUPPORTS_C45;
+	} else {
+		cmd->phy_address = 0;  /* not really, but no better option */
+		cmd->transceiver = XCVR_INTERNAL;
+		cmd->mdio_support = 0;
+	}
+
+	cmd->supported = t4vf_from_fw_linkcaps(p->port_type,
+					       p->link_cfg.supported);
+	cmd->advertising = t4vf_from_fw_linkcaps(p->port_type,
+					    p->link_cfg.advertising);
+	ethtool_cmd_speed_set(cmd,
+			      netif_carrier_ok(dev) ? p->link_cfg.speed : 0);
+	cmd->duplex = DUPLEX_FULL;
+	cmd->autoneg = p->link_cfg.autoneg;
+	cmd->maxtxpkt = 0;
+	cmd->maxrxpkt = 0;
+	return 0;
+}
+
+/*
+ * Return our driver information.
+ */
+static void cxgb4vf_get_drvinfo(struct net_device *dev,
+				struct ethtool_drvinfo *drvinfo)
+{
+	struct adapter *adapter = netdev2adap(dev);
+
+	strlcpy(drvinfo->driver, KBUILD_MODNAME, sizeof(drvinfo->driver));
+	strlcpy(drvinfo->version, DRV_VERSION, sizeof(drvinfo->version));
+	strlcpy(drvinfo->bus_info, pci_name(to_pci_dev(dev->dev.parent)),
+		sizeof(drvinfo->bus_info));
+	snprintf(drvinfo->fw_version, sizeof(drvinfo->fw_version),
+		 "%u.%u.%u.%u, TP %u.%u.%u.%u",
+		 FW_HDR_FW_VER_MAJOR_G(adapter->params.dev.fwrev),
+		 FW_HDR_FW_VER_MINOR_G(adapter->params.dev.fwrev),
+		 FW_HDR_FW_VER_MICRO_G(adapter->params.dev.fwrev),
+		 FW_HDR_FW_VER_BUILD_G(adapter->params.dev.fwrev),
+		 FW_HDR_FW_VER_MAJOR_G(adapter->params.dev.tprev),
+		 FW_HDR_FW_VER_MINOR_G(adapter->params.dev.tprev),
+		 FW_HDR_FW_VER_MICRO_G(adapter->params.dev.tprev),
+		 FW_HDR_FW_VER_BUILD_G(adapter->params.dev.tprev));
+}
+
+/*
+ * Return current adapter message level.
+ */
+static u32 cxgb4vf_get_msglevel(struct net_device *dev)
+{
+	return netdev2adap(dev)->msg_enable;
+}
+
+/*
+ * Set current adapter message level.
+ */
+static void cxgb4vf_set_msglevel(struct net_device *dev, u32 msglevel)
+{
+	netdev2adap(dev)->msg_enable = msglevel;
+}
+
+/*
+ * Return the device's current Queue Set ring size parameters along with the
+ * allowed maximum values.  Since ethtool doesn't understand the concept of
+ * multi-queue devices, we just return the current values associated with the
+ * first Queue Set.
+ */
+static void cxgb4vf_get_ringparam(struct net_device *dev,
+				  struct ethtool_ringparam *rp)
+{
+	const struct port_info *pi = netdev_priv(dev);
+	const struct sge *s = &pi->adapter->sge;
+
+	rp->rx_max_pending = MAX_RX_BUFFERS;
+	rp->rx_mini_max_pending = MAX_RSPQ_ENTRIES;
+	rp->rx_jumbo_max_pending = 0;
+	rp->tx_max_pending = MAX_TXQ_ENTRIES;
+
+	rp->rx_pending = s->ethrxq[pi->first_qset].fl.size - MIN_FL_RESID;
+	rp->rx_mini_pending = s->ethrxq[pi->first_qset].rspq.size;
+	rp->rx_jumbo_pending = 0;
+	rp->tx_pending = s->ethtxq[pi->first_qset].q.size;
+}
+
+/*
+ * Set the Queue Set ring size parameters for the device.  Again, since
+ * ethtool doesn't allow for the concept of multiple queues per device, we'll
+ * apply these new values across all of the Queue Sets associated with the
+ * device -- after vetting them of course!
+ */
+static int cxgb4vf_set_ringparam(struct net_device *dev,
+				 struct ethtool_ringparam *rp)
+{
+	const struct port_info *pi = netdev_priv(dev);
+	struct adapter *adapter = pi->adapter;
+	struct sge *s = &adapter->sge;
+	int qs;
+
+	if (rp->rx_pending > MAX_RX_BUFFERS ||
+	    rp->rx_jumbo_pending ||
+	    rp->tx_pending > MAX_TXQ_ENTRIES ||
+	    rp->rx_mini_pending > MAX_RSPQ_ENTRIES ||
+	    rp->rx_mini_pending < MIN_RSPQ_ENTRIES ||
+	    rp->rx_pending < MIN_FL_ENTRIES ||
+	    rp->tx_pending < MIN_TXQ_ENTRIES)
+		return -EINVAL;
+
+	if (adapter->flags & FULL_INIT_DONE)
+		return -EBUSY;
+
+	for (qs = pi->first_qset; qs < pi->first_qset + pi->nqsets; qs++) {
+		s->ethrxq[qs].fl.size = rp->rx_pending + MIN_FL_RESID;
+		s->ethrxq[qs].rspq.size = rp->rx_mini_pending;
+		s->ethtxq[qs].q.size = rp->tx_pending;
+	}
+	return 0;
+}
+
+/*
+ * Return the interrupt holdoff timer and count for the first Queue Set on the
+ * device.  Our extension ioctl() (the cxgbtool interface) allows the
+ * interrupt holdoff timer to be read on all of the device's Queue Sets.
+ */
+static int cxgb4vf_get_coalesce(struct net_device *dev,
+				struct ethtool_coalesce *coalesce)
+{
+	const struct port_info *pi = netdev_priv(dev);
+	const struct adapter *adapter = pi->adapter;
+	const struct sge_rspq *rspq = &adapter->sge.ethrxq[pi->first_qset].rspq;
+
+	coalesce->rx_coalesce_usecs = qtimer_val(adapter, rspq);
+	coalesce->rx_max_coalesced_frames =
+		((rspq->intr_params & QINTR_CNT_EN)
+		 ? adapter->sge.counter_val[rspq->pktcnt_idx]
+		 : 0);
+	return 0;
+}
+
+/*
+ * Set the RX interrupt holdoff timer and count for the first Queue Set on the
+ * interface.  Our extension ioctl() (the cxgbtool interface) allows us to set
+ * the interrupt holdoff timer on any of the device's Queue Sets.
+ */
+static int cxgb4vf_set_coalesce(struct net_device *dev,
+				struct ethtool_coalesce *coalesce)
+{
+	const struct port_info *pi = netdev_priv(dev);
+	struct adapter *adapter = pi->adapter;
+
+	return set_rxq_intr_params(adapter,
+				   &adapter->sge.ethrxq[pi->first_qset].rspq,
+				   coalesce->rx_coalesce_usecs,
+				   coalesce->rx_max_coalesced_frames);
+}
+
+/*
+ * Report current port link pause parameter settings.
+ */
+static void cxgb4vf_get_pauseparam(struct net_device *dev,
+				   struct ethtool_pauseparam *pauseparam)
+{
+	struct port_info *pi = netdev_priv(dev);
+
+	pauseparam->autoneg = (pi->link_cfg.requested_fc & PAUSE_AUTONEG) != 0;
+	pauseparam->rx_pause = (pi->link_cfg.fc & PAUSE_RX) != 0;
+	pauseparam->tx_pause = (pi->link_cfg.fc & PAUSE_TX) != 0;
+}
+
+/*
+ * Identify the port by blinking the port's LED.
+ */
+static int cxgb4vf_phys_id(struct net_device *dev,
+			   enum ethtool_phys_id_state state)
+{
+	unsigned int val;
+	struct port_info *pi = netdev_priv(dev);
+
+	if (state == ETHTOOL_ID_ACTIVE)
+		val = 0xffff;
+	else if (state == ETHTOOL_ID_INACTIVE)
+		val = 0;
+	else
+		return -EINVAL;
+
+	return t4vf_identify_port(pi->adapter, pi->viid, val);
+}
+
+/*
+ * Port stats maintained per queue of the port.
+ */
+struct queue_port_stats {
+	u64 tso;
+	u64 tx_csum;
+	u64 rx_csum;
+	u64 vlan_ex;
+	u64 vlan_ins;
+	u64 lro_pkts;
+	u64 lro_merged;
+};
+
+/*
+ * Strings for the ETH_SS_STATS statistics set ("ethtool -S").  Note that
+ * these need to match the order of statistics returned by
+ * t4vf_get_port_stats().
+ */
+static const char stats_strings[][ETH_GSTRING_LEN] = {
+	/*
+	 * These must match the layout of the t4vf_port_stats structure.
+	 */
+	"TxBroadcastBytes  ",
+	"TxBroadcastFrames ",
+	"TxMulticastBytes  ",
+	"TxMulticastFrames ",
+	"TxUnicastBytes    ",
+	"TxUnicastFrames   ",
+	"TxDroppedFrames   ",
+	"TxOffloadBytes    ",
+	"TxOffloadFrames   ",
+	"RxBroadcastBytes  ",
+	"RxBroadcastFrames ",
+	"RxMulticastBytes  ",
+	"RxMulticastFrames ",
+	"RxUnicastBytes    ",
+	"RxUnicastFrames   ",
+	"RxErrorFrames     ",
+
+	/*
+	 * These are accumulated per-queue statistics and must match the
+	 * order of the fields in the queue_port_stats structure.
+	 */
+	"TSO               ",
+	"TxCsumOffload     ",
+	"RxCsumGood        ",
+	"VLANextractions   ",
+	"VLANinsertions    ",
+	"GROPackets        ",
+	"GROMerged         ",
+};
+
+/*
+ * Return the number of statistics in the specified statistics set.
+ */
+static int cxgb4vf_get_sset_count(struct net_device *dev, int sset)
+{
+	switch (sset) {
+	case ETH_SS_STATS:
+		return ARRAY_SIZE(stats_strings);
+	default:
+		return -EOPNOTSUPP;
+	}
+	/*NOTREACHED*/
+}
+
+/*
+ * Return the strings for the specified statistics set.
+ */
+static void cxgb4vf_get_strings(struct net_device *dev,
+				u32 sset,
+				u8 *data)
+{
+	switch (sset) {
+	case ETH_SS_STATS:
+		memcpy(data, stats_strings, sizeof(stats_strings));
+		break;
+	}
+}
+
+/*
+ * Small utility routine to accumulate queue statistics across the queues of
+ * a "port".
+ */
+static void collect_sge_port_stats(const struct adapter *adapter,
+				   const struct port_info *pi,
+				   struct queue_port_stats *stats)
+{
+	const struct sge_eth_txq *txq = &adapter->sge.ethtxq[pi->first_qset];
+	const struct sge_eth_rxq *rxq = &adapter->sge.ethrxq[pi->first_qset];
+	int qs;
+
+	memset(stats, 0, sizeof(*stats));
+	for (qs = 0; qs < pi->nqsets; qs++, rxq++, txq++) {
+		stats->tso += txq->tso;
+		stats->tx_csum += txq->tx_cso;
+		stats->rx_csum += rxq->stats.rx_cso;
+		stats->vlan_ex += rxq->stats.vlan_ex;
+		stats->vlan_ins += txq->vlan_ins;
+		stats->lro_pkts += rxq->stats.lro_pkts;
+		stats->lro_merged += rxq->stats.lro_merged;
+	}
+}
+
+/*
+ * Return the ETH_SS_STATS statistics set.
+ */
+static void cxgb4vf_get_ethtool_stats(struct net_device *dev,
+				      struct ethtool_stats *stats,
+				      u64 *data)
+{
+	struct port_info *pi = netdev2pinfo(dev);
+	struct adapter *adapter = pi->adapter;
+	int err = t4vf_get_port_stats(adapter, pi->pidx,
+				      (struct t4vf_port_stats *)data);
+	if (err)
+		memset(data, 0, sizeof(struct t4vf_port_stats));
+
+	data += sizeof(struct t4vf_port_stats) / sizeof(u64);
+	collect_sge_port_stats(adapter, pi, (struct queue_port_stats *)data);
+}
+
+/*
+ * Return the size of our register map.
+ */
+static int cxgb4vf_get_regs_len(struct net_device *dev)
+{
+	return T4VF_REGMAP_SIZE;
+}
+
+/*
+ * Dump a block of registers, start to end inclusive, into a buffer.
+ */
+static void reg_block_dump(struct adapter *adapter, void *regbuf,
+			   unsigned int start, unsigned int end)
+{
+	u32 *bp = regbuf + start - T4VF_REGMAP_START;
+
+	for ( ; start <= end; start += sizeof(u32)) {
+		/*
+		 * Avoid reading the Mailbox Control register since that
+		 * can trigger a Mailbox Ownership Arbitration cycle and
+		 * interfere with communication with the firmware.
+		 */
+		if (start == T4VF_CIM_BASE_ADDR + CIM_VF_EXT_MAILBOX_CTRL)
+			*bp++ = 0xffff;
+		else
+			*bp++ = t4_read_reg(adapter, start);
+	}
+}
+
+/*
+ * Copy our entire register map into the provided buffer.
+ */
+static void cxgb4vf_get_regs(struct net_device *dev,
+			     struct ethtool_regs *regs,
+			     void *regbuf)
+{
+	struct adapter *adapter = netdev2adap(dev);
+
+	regs->version = mk_adap_vers(adapter);
+
+	/*
+	 * Fill in register buffer with our register map.
+	 */
+	memset(regbuf, 0, T4VF_REGMAP_SIZE);
+
+	reg_block_dump(adapter, regbuf,
+		       T4VF_SGE_BASE_ADDR + T4VF_MOD_MAP_SGE_FIRST,
+		       T4VF_SGE_BASE_ADDR + T4VF_MOD_MAP_SGE_LAST);
+	reg_block_dump(adapter, regbuf,
+		       T4VF_MPS_BASE_ADDR + T4VF_MOD_MAP_MPS_FIRST,
+		       T4VF_MPS_BASE_ADDR + T4VF_MOD_MAP_MPS_LAST);
+
+	/* T5 adds new registers in the PL Register map.
+	 */
+	reg_block_dump(adapter, regbuf,
+		       T4VF_PL_BASE_ADDR + T4VF_MOD_MAP_PL_FIRST,
+		       T4VF_PL_BASE_ADDR + (is_t4(adapter->params.chip)
+		       ? PL_VF_WHOAMI_A : PL_VF_REVISION_A));
+	reg_block_dump(adapter, regbuf,
+		       T4VF_CIM_BASE_ADDR + T4VF_MOD_MAP_CIM_FIRST,
+		       T4VF_CIM_BASE_ADDR + T4VF_MOD_MAP_CIM_LAST);
+
+	reg_block_dump(adapter, regbuf,
+		       T4VF_MBDATA_BASE_ADDR + T4VF_MBDATA_FIRST,
+		       T4VF_MBDATA_BASE_ADDR + T4VF_MBDATA_LAST);
+}
+
+/*
+ * Report current Wake On LAN settings.
+ */
+static void cxgb4vf_get_wol(struct net_device *dev,
+			    struct ethtool_wolinfo *wol)
+{
+	wol->supported = 0;
+	wol->wolopts = 0;
+	memset(&wol->sopass, 0, sizeof(wol->sopass));
+}
+
+/*
+ * TCP Segmentation Offload flags which we support.
+ */
+#define TSO_FLAGS (NETIF_F_TSO | NETIF_F_TSO6 | NETIF_F_TSO_ECN)
+
+static const struct ethtool_ops cxgb4vf_ethtool_ops = {
+	.get_settings		= cxgb4vf_get_settings,
+	.get_drvinfo		= cxgb4vf_get_drvinfo,
+	.get_msglevel		= cxgb4vf_get_msglevel,
+	.set_msglevel		= cxgb4vf_set_msglevel,
+	.get_ringparam		= cxgb4vf_get_ringparam,
+	.set_ringparam		= cxgb4vf_set_ringparam,
+	.get_coalesce		= cxgb4vf_get_coalesce,
+	.set_coalesce		= cxgb4vf_set_coalesce,
+	.get_pauseparam		= cxgb4vf_get_pauseparam,
+	.get_link		= ethtool_op_get_link,
+	.get_strings		= cxgb4vf_get_strings,
+	.set_phys_id		= cxgb4vf_phys_id,
+	.get_sset_count		= cxgb4vf_get_sset_count,
+	.get_ethtool_stats	= cxgb4vf_get_ethtool_stats,
+	.get_regs_len		= cxgb4vf_get_regs_len,
+	.get_regs		= cxgb4vf_get_regs,
+	.get_wol		= cxgb4vf_get_wol,
+};
+
+/*
+ * /sys/kernel/debug/cxgb4vf support code and data.
+ * ================================================
+ */
+
+/*
+ * Show SGE Queue Set information.  We display QPL Queues Sets per line.
+ */
+#define QPL	4
+
+static int sge_qinfo_show(struct seq_file *seq, void *v)
+{
+	struct adapter *adapter = seq->private;
+	int eth_entries = DIV_ROUND_UP(adapter->sge.ethqsets, QPL);
+	int qs, r = (uintptr_t)v - 1;
+
+	if (r)
+		seq_putc(seq, '\n');
+
+	#define S3(fmt_spec, s, v) \
+		do {\
+			seq_printf(seq, "%-12s", s); \
+			for (qs = 0; qs < n; ++qs) \
+				seq_printf(seq, " %16" fmt_spec, v); \
+			seq_putc(seq, '\n'); \
+		} while (0)
+	#define S(s, v)		S3("s", s, v)
+	#define T(s, v)		S3("u", s, txq[qs].v)
+	#define R(s, v)		S3("u", s, rxq[qs].v)
+
+	if (r < eth_entries) {
+		const struct sge_eth_rxq *rxq = &adapter->sge.ethrxq[r * QPL];
+		const struct sge_eth_txq *txq = &adapter->sge.ethtxq[r * QPL];
+		int n = min(QPL, adapter->sge.ethqsets - QPL * r);
+
+		S("QType:", "Ethernet");
+		S("Interface:",
+		  (rxq[qs].rspq.netdev
+		   ? rxq[qs].rspq.netdev->name
+		   : "N/A"));
+		S3("d", "Port:",
+		   (rxq[qs].rspq.netdev
+		    ? ((struct port_info *)
+		       netdev_priv(rxq[qs].rspq.netdev))->port_id
+		    : -1));
+		T("TxQ ID:", q.abs_id);
+		T("TxQ size:", q.size);
+		T("TxQ inuse:", q.in_use);
+		T("TxQ PIdx:", q.pidx);
+		T("TxQ CIdx:", q.cidx);
+		R("RspQ ID:", rspq.abs_id);
+		R("RspQ size:", rspq.size);
+		R("RspQE size:", rspq.iqe_len);
+		S3("u", "Intr delay:", qtimer_val(adapter, &rxq[qs].rspq));
+		S3("u", "Intr pktcnt:",
+		   adapter->sge.counter_val[rxq[qs].rspq.pktcnt_idx]);
+		R("RspQ CIdx:", rspq.cidx);
+		R("RspQ Gen:", rspq.gen);
+		R("FL ID:", fl.abs_id);
+		R("FL size:", fl.size - MIN_FL_RESID);
+		R("FL avail:", fl.avail);
+		R("FL PIdx:", fl.pidx);
+		R("FL CIdx:", fl.cidx);
+		return 0;
+	}
+
+	r -= eth_entries;
+	if (r == 0) {
+		const struct sge_rspq *evtq = &adapter->sge.fw_evtq;
+
+		seq_printf(seq, "%-12s %16s\n", "QType:", "FW event queue");
+		seq_printf(seq, "%-12s %16u\n", "RspQ ID:", evtq->abs_id);
+		seq_printf(seq, "%-12s %16u\n", "Intr delay:",
+			   qtimer_val(adapter, evtq));
+		seq_printf(seq, "%-12s %16u\n", "Intr pktcnt:",
+			   adapter->sge.counter_val[evtq->pktcnt_idx]);
+		seq_printf(seq, "%-12s %16u\n", "RspQ Cidx:", evtq->cidx);
+		seq_printf(seq, "%-12s %16u\n", "RspQ Gen:", evtq->gen);
+	} else if (r == 1) {
+		const struct sge_rspq *intrq = &adapter->sge.intrq;
+
+		seq_printf(seq, "%-12s %16s\n", "QType:", "Interrupt Queue");
+		seq_printf(seq, "%-12s %16u\n", "RspQ ID:", intrq->abs_id);
+		seq_printf(seq, "%-12s %16u\n", "Intr delay:",
+			   qtimer_val(adapter, intrq));
+		seq_printf(seq, "%-12s %16u\n", "Intr pktcnt:",
+			   adapter->sge.counter_val[intrq->pktcnt_idx]);
+		seq_printf(seq, "%-12s %16u\n", "RspQ Cidx:", intrq->cidx);
+		seq_printf(seq, "%-12s %16u\n", "RspQ Gen:", intrq->gen);
+	}
+
+	#undef R
+	#undef T
+	#undef S
+	#undef S3
+
+	return 0;
+}
+
+/*
+ * Return the number of "entries" in our "file".  We group the multi-Queue
+ * sections with QPL Queue Sets per "entry".  The sections of the output are:
+ *
+ *     Ethernet RX/TX Queue Sets
+ *     Firmware Event Queue
+ *     Forwarded Interrupt Queue (if in MSI mode)
+ */
+static int sge_queue_entries(const struct adapter *adapter)
+{
+	return DIV_ROUND_UP(adapter->sge.ethqsets, QPL) + 1 +
+		((adapter->flags & USING_MSI) != 0);
+}
+
+static void *sge_queue_start(struct seq_file *seq, loff_t *pos)
+{
+	int entries = sge_queue_entries(seq->private);
+
+	return *pos < entries ? (void *)((uintptr_t)*pos + 1) : NULL;
+}
+
+static void sge_queue_stop(struct seq_file *seq, void *v)
+{
+}
+
+static void *sge_queue_next(struct seq_file *seq, void *v, loff_t *pos)
+{
+	int entries = sge_queue_entries(seq->private);
+
+	++*pos;
+	return *pos < entries ? (void *)((uintptr_t)*pos + 1) : NULL;
+}
+
+static const struct seq_operations sge_qinfo_seq_ops = {
+	.start = sge_queue_start,
+	.next  = sge_queue_next,
+	.stop  = sge_queue_stop,
+	.show  = sge_qinfo_show
+};
+
+static int sge_qinfo_open(struct inode *inode, struct file *file)
+{
+	int res = seq_open(file, &sge_qinfo_seq_ops);
+
+	if (!res) {
+		struct seq_file *seq = file->private_data;
+		seq->private = inode->i_private;
+	}
+	return res;
+}
+
+static const struct file_operations sge_qinfo_debugfs_fops = {
+	.owner   = THIS_MODULE,
+	.open    = sge_qinfo_open,
+	.read    = seq_read,
+	.llseek  = seq_lseek,
+	.release = seq_release,
+};
+
+/*
+ * Show SGE Queue Set statistics.  We display QPL Queues Sets per line.
+ */
+#define QPL	4
+
+static int sge_qstats_show(struct seq_file *seq, void *v)
+{
+	struct adapter *adapter = seq->private;
+	int eth_entries = DIV_ROUND_UP(adapter->sge.ethqsets, QPL);
+	int qs, r = (uintptr_t)v - 1;
+
+	if (r)
+		seq_putc(seq, '\n');
+
+	#define S3(fmt, s, v) \
+		do { \
+			seq_printf(seq, "%-16s", s); \
+			for (qs = 0; qs < n; ++qs) \
+				seq_printf(seq, " %8" fmt, v); \
+			seq_putc(seq, '\n'); \
+		} while (0)
+	#define S(s, v)		S3("s", s, v)
+
+	#define T3(fmt, s, v)	S3(fmt, s, txq[qs].v)
+	#define T(s, v)		T3("lu", s, v)
+
+	#define R3(fmt, s, v)	S3(fmt, s, rxq[qs].v)
+	#define R(s, v)		R3("lu", s, v)
+
+	if (r < eth_entries) {
+		const struct sge_eth_rxq *rxq = &adapter->sge.ethrxq[r * QPL];
+		const struct sge_eth_txq *txq = &adapter->sge.ethtxq[r * QPL];
+		int n = min(QPL, adapter->sge.ethqsets - QPL * r);
+
+		S("QType:", "Ethernet");
+		S("Interface:",
+		  (rxq[qs].rspq.netdev
+		   ? rxq[qs].rspq.netdev->name
+		   : "N/A"));
+		R3("u", "RspQNullInts:", rspq.unhandled_irqs);
+		R("RxPackets:", stats.pkts);
+		R("RxCSO:", stats.rx_cso);
+		R("VLANxtract:", stats.vlan_ex);
+		R("LROmerged:", stats.lro_merged);
+		R("LROpackets:", stats.lro_pkts);
+		R("RxDrops:", stats.rx_drops);
+		T("TSO:", tso);
+		T("TxCSO:", tx_cso);
+		T("VLANins:", vlan_ins);
+		T("TxQFull:", q.stops);
+		T("TxQRestarts:", q.restarts);
+		T("TxMapErr:", mapping_err);
+		R("FLAllocErr:", fl.alloc_failed);
+		R("FLLrgAlcErr:", fl.large_alloc_failed);
+		R("FLStarving:", fl.starving);
+		return 0;
+	}
+
+	r -= eth_entries;
+	if (r == 0) {
+		const struct sge_rspq *evtq = &adapter->sge.fw_evtq;
+
+		seq_printf(seq, "%-8s %16s\n", "QType:", "FW event queue");
+		seq_printf(seq, "%-16s %8u\n", "RspQNullInts:",
+			   evtq->unhandled_irqs);
+		seq_printf(seq, "%-16s %8u\n", "RspQ CIdx:", evtq->cidx);
+		seq_printf(seq, "%-16s %8u\n", "RspQ Gen:", evtq->gen);
+	} else if (r == 1) {
+		const struct sge_rspq *intrq = &adapter->sge.intrq;
+
+		seq_printf(seq, "%-8s %16s\n", "QType:", "Interrupt Queue");
+		seq_printf(seq, "%-16s %8u\n", "RspQNullInts:",
+			   intrq->unhandled_irqs);
+		seq_printf(seq, "%-16s %8u\n", "RspQ CIdx:", intrq->cidx);
+		seq_printf(seq, "%-16s %8u\n", "RspQ Gen:", intrq->gen);
+	}
+
+	#undef R
+	#undef T
+	#undef S
+	#undef R3
+	#undef T3
+	#undef S3
+
+	return 0;
+}
+
+/*
+ * Return the number of "entries" in our "file".  We group the multi-Queue
+ * sections with QPL Queue Sets per "entry".  The sections of the output are:
+ *
+ *     Ethernet RX/TX Queue Sets
+ *     Firmware Event Queue
+ *     Forwarded Interrupt Queue (if in MSI mode)
+ */
+static int sge_qstats_entries(const struct adapter *adapter)
+{
+	return DIV_ROUND_UP(adapter->sge.ethqsets, QPL) + 1 +
+		((adapter->flags & USING_MSI) != 0);
+}
+
+static void *sge_qstats_start(struct seq_file *seq, loff_t *pos)
+{
+	int entries = sge_qstats_entries(seq->private);
+
+	return *pos < entries ? (void *)((uintptr_t)*pos + 1) : NULL;
+}
+
+static void sge_qstats_stop(struct seq_file *seq, void *v)
+{
+}
+
+static void *sge_qstats_next(struct seq_file *seq, void *v, loff_t *pos)
+{
+	int entries = sge_qstats_entries(seq->private);
+
+	(*pos)++;
+	return *pos < entries ? (void *)((uintptr_t)*pos + 1) : NULL;
+}
+
+static const struct seq_operations sge_qstats_seq_ops = {
+	.start = sge_qstats_start,
+	.next  = sge_qstats_next,
+	.stop  = sge_qstats_stop,
+	.show  = sge_qstats_show
+};
+
+static int sge_qstats_open(struct inode *inode, struct file *file)
+{
+	int res = seq_open(file, &sge_qstats_seq_ops);
+
+	if (res == 0) {
+		struct seq_file *seq = file->private_data;
+		seq->private = inode->i_private;
+	}
+	return res;
+}
+
+static const struct file_operations sge_qstats_proc_fops = {
+	.owner   = THIS_MODULE,
+	.open    = sge_qstats_open,
+	.read    = seq_read,
+	.llseek  = seq_lseek,
+	.release = seq_release,
+};
+
+/*
+ * Show PCI-E SR-IOV Virtual Function Resource Limits.
+ */
+static int resources_show(struct seq_file *seq, void *v)
+{
+	struct adapter *adapter = seq->private;
+	struct vf_resources *vfres = &adapter->params.vfres;
+
+	#define S(desc, fmt, var) \
+		seq_printf(seq, "%-60s " fmt "\n", \
+			   desc " (" #var "):", vfres->var)
+
+	S("Virtual Interfaces", "%d", nvi);
+	S("Egress Queues", "%d", neq);
+	S("Ethernet Control", "%d", nethctrl);
+	S("Ingress Queues/w Free Lists/Interrupts", "%d", niqflint);
+	S("Ingress Queues", "%d", niq);
+	S("Traffic Class", "%d", tc);
+	S("Port Access Rights Mask", "%#x", pmask);
+	S("MAC Address Filters", "%d", nexactf);
+	S("Firmware Command Read Capabilities", "%#x", r_caps);
+	S("Firmware Command Write/Execute Capabilities", "%#x", wx_caps);
+
+	#undef S
+
+	return 0;
+}
+
+static int resources_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, resources_show, inode->i_private);
+}
+
+static const struct file_operations resources_proc_fops = {
+	.owner   = THIS_MODULE,
+	.open    = resources_open,
+	.read    = seq_read,
+	.llseek  = seq_lseek,
+	.release = single_release,
+};
+
+/*
+ * Show Virtual Interfaces.
+ */
+static int interfaces_show(struct seq_file *seq, void *v)
+{
+	if (v == SEQ_START_TOKEN) {
+		seq_puts(seq, "Interface  Port   VIID\n");
+	} else {
+		struct adapter *adapter = seq->private;
+		int pidx = (uintptr_t)v - 2;
+		struct net_device *dev = adapter->port[pidx];
+		struct port_info *pi = netdev_priv(dev);
+
+		seq_printf(seq, "%9s  %4d  %#5x\n",
+			   dev->name, pi->port_id, pi->viid);
+	}
+	return 0;
+}
+
+static inline void *interfaces_get_idx(struct adapter *adapter, loff_t pos)
+{
+	return pos <= adapter->params.nports
+		? (void *)(uintptr_t)(pos + 1)
+		: NULL;
+}
+
+static void *interfaces_start(struct seq_file *seq, loff_t *pos)
+{
+	return *pos
+		? interfaces_get_idx(seq->private, *pos)
+		: SEQ_START_TOKEN;
+}
+
+static void *interfaces_next(struct seq_file *seq, void *v, loff_t *pos)
+{
+	(*pos)++;
+	return interfaces_get_idx(seq->private, *pos);
+}
+
+static void interfaces_stop(struct seq_file *seq, void *v)
+{
+}
+
+static const struct seq_operations interfaces_seq_ops = {
+	.start = interfaces_start,
+	.next  = interfaces_next,
+	.stop  = interfaces_stop,
+	.show  = interfaces_show
+};
+
+static int interfaces_open(struct inode *inode, struct file *file)
+{
+	int res = seq_open(file, &interfaces_seq_ops);
+
+	if (res == 0) {
+		struct seq_file *seq = file->private_data;
+		seq->private = inode->i_private;
+	}
+	return res;
+}
+
+static const struct file_operations interfaces_proc_fops = {
+	.owner   = THIS_MODULE,
+	.open    = interfaces_open,
+	.read    = seq_read,
+	.llseek  = seq_lseek,
+	.release = seq_release,
+};
+
+/*
+ * /sys/kernel/debugfs/cxgb4vf/ files list.
+ */
+struct cxgb4vf_debugfs_entry {
+	const char *name;		/* name of debugfs node */
+	umode_t mode;			/* file system mode */
+	const struct file_operations *fops;
+};
+
+static struct cxgb4vf_debugfs_entry debugfs_files[] = {
+	{ "sge_qinfo",  S_IRUGO, &sge_qinfo_debugfs_fops },
+	{ "sge_qstats", S_IRUGO, &sge_qstats_proc_fops },
+	{ "resources",  S_IRUGO, &resources_proc_fops },
+	{ "interfaces", S_IRUGO, &interfaces_proc_fops },
+};
+
+/*
+ * Module and device initialization and cleanup code.
+ * ==================================================
+ */
+
+/*
+ * Set up out /sys/kernel/debug/cxgb4vf sub-nodes.  We assume that the
+ * directory (debugfs_root) has already been set up.
+ */
+static int setup_debugfs(struct adapter *adapter)
+{
+	int i;
+
+	BUG_ON(IS_ERR_OR_NULL(adapter->debugfs_root));
+
+	/*
+	 * Debugfs support is best effort.
+	 */
+	for (i = 0; i < ARRAY_SIZE(debugfs_files); i++)
+		(void)debugfs_create_file(debugfs_files[i].name,
+				  debugfs_files[i].mode,
+				  adapter->debugfs_root,
+				  (void *)adapter,
+				  debugfs_files[i].fops);
+
+	return 0;
+}
+
+/*
+ * Tear down the /sys/kernel/debug/cxgb4vf sub-nodes created above.  We leave
+ * it to our caller to tear down the directory (debugfs_root).
+ */
+static void cleanup_debugfs(struct adapter *adapter)
+{
+	BUG_ON(IS_ERR_OR_NULL(adapter->debugfs_root));
+
+	/*
+	 * Unlike our sister routine cleanup_proc(), we don't need to remove
+	 * individual entries because a call will be made to
+	 * debugfs_remove_recursive().  We just need to clean up any ancillary
+	 * persistent state.
+	 */
+	/* nothing to do */
+}
+
+/*
+ * Perform early "adapter" initialization.  This is where we discover what
+ * adapter parameters we're going to be using and initialize basic adapter
+ * hardware support.
+ */
+static int adap_init0(struct adapter *adapter)
+{
+	struct vf_resources *vfres = &adapter->params.vfres;
+	struct sge_params *sge_params = &adapter->params.sge;
+	struct sge *s = &adapter->sge;
+	unsigned int ethqsets;
+	int err;
+	u32 param, val = 0;
+
+	/*
+	 * Wait for the device to become ready before proceeding ...
+	 */
+	err = t4vf_wait_dev_ready(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "device didn't become ready:"
+			" err=%d\n", err);
+		return err;
+	}
+
+	/*
+	 * Some environments do not properly handle PCIE FLRs -- e.g. in Linux
+	 * 2.6.31 and later we can't call pci_reset_function() in order to
+	 * issue an FLR because of a self- deadlock on the device semaphore.
+	 * Meanwhile, the OS infrastructure doesn't issue FLRs in all the
+	 * cases where they're needed -- for instance, some versions of KVM
+	 * fail to reset "Assigned Devices" when the VM reboots.  Therefore we
+	 * use the firmware based reset in order to reset any per function
+	 * state.
+	 */
+	err = t4vf_fw_reset(adapter);
+	if (err < 0) {
+		dev_err(adapter->pdev_dev, "FW reset failed: err=%d\n", err);
+		return err;
+	}
+
+	/*
+	 * Grab basic operational parameters.  These will predominantly have
+	 * been set up by the Physical Function Driver or will be hard coded
+	 * into the adapter.  We just have to live with them ...  Note that
+	 * we _must_ get our VPD parameters before our SGE parameters because
+	 * we need to know the adapter's core clock from the VPD in order to
+	 * properly decode the SGE Timer Values.
+	 */
+	err = t4vf_get_dev_params(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "unable to retrieve adapter"
+			" device parameters: err=%d\n", err);
+		return err;
+	}
+	err = t4vf_get_vpd_params(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "unable to retrieve adapter"
+			" VPD parameters: err=%d\n", err);
+		return err;
+	}
+	err = t4vf_get_sge_params(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "unable to retrieve adapter"
+			" SGE parameters: err=%d\n", err);
+		return err;
+	}
+	err = t4vf_get_rss_glb_config(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "unable to retrieve adapter"
+			" RSS parameters: err=%d\n", err);
+		return err;
+	}
+	if (adapter->params.rss.mode !=
+	    FW_RSS_GLB_CONFIG_CMD_MODE_BASICVIRTUAL) {
+		dev_err(adapter->pdev_dev, "unable to operate with global RSS"
+			" mode %d\n", adapter->params.rss.mode);
+		return -EINVAL;
+	}
+	err = t4vf_sge_init(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "unable to use adapter parameters:"
+			" err=%d\n", err);
+		return err;
+	}
+
+	/* If we're running on newer firmware, let it know that we're
+	 * prepared to deal with encapsulated CPL messages.  Older
+	 * firmware won't understand this and we'll just get
+	 * unencapsulated messages ...
+	 */
+	param = FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_PFVF) |
+		FW_PARAMS_PARAM_X_V(FW_PARAMS_PARAM_PFVF_CPLFW4MSG_ENCAP);
+	val = 1;
+	(void) t4vf_set_params(adapter, 1, &param, &val);
+
+	/*
+	 * Retrieve our RX interrupt holdoff timer values and counter
+	 * threshold values from the SGE parameters.
+	 */
+	s->timer_val[0] = core_ticks_to_us(adapter,
+		TIMERVALUE0_G(sge_params->sge_timer_value_0_and_1));
+	s->timer_val[1] = core_ticks_to_us(adapter,
+		TIMERVALUE1_G(sge_params->sge_timer_value_0_and_1));
+	s->timer_val[2] = core_ticks_to_us(adapter,
+		TIMERVALUE0_G(sge_params->sge_timer_value_2_and_3));
+	s->timer_val[3] = core_ticks_to_us(adapter,
+		TIMERVALUE1_G(sge_params->sge_timer_value_2_and_3));
+	s->timer_val[4] = core_ticks_to_us(adapter,
+		TIMERVALUE0_G(sge_params->sge_timer_value_4_and_5));
+	s->timer_val[5] = core_ticks_to_us(adapter,
+		TIMERVALUE1_G(sge_params->sge_timer_value_4_and_5));
+
+	s->counter_val[0] = THRESHOLD_0_G(sge_params->sge_ingress_rx_threshold);
+	s->counter_val[1] = THRESHOLD_1_G(sge_params->sge_ingress_rx_threshold);
+	s->counter_val[2] = THRESHOLD_2_G(sge_params->sge_ingress_rx_threshold);
+	s->counter_val[3] = THRESHOLD_3_G(sge_params->sge_ingress_rx_threshold);
+
+	/*
+	 * Grab our Virtual Interface resource allocation, extract the
+	 * features that we're interested in and do a bit of sanity testing on
+	 * what we discover.
+	 */
+	err = t4vf_get_vfres(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "unable to get virtual interface"
+			" resources: err=%d\n", err);
+		return err;
+	}
+
+	/*
+	 * The number of "ports" which we support is equal to the number of
+	 * Virtual Interfaces with which we've been provisioned.
+	 */
+	adapter->params.nports = vfres->nvi;
+	if (adapter->params.nports > MAX_NPORTS) {
+		dev_warn(adapter->pdev_dev, "only using %d of %d allowed"
+			 " virtual interfaces\n", MAX_NPORTS,
+			 adapter->params.nports);
+		adapter->params.nports = MAX_NPORTS;
+	}
+
+	/*
+	 * We need to reserve a number of the ingress queues with Free List
+	 * and Interrupt capabilities for special interrupt purposes (like
+	 * asynchronous firmware messages, or forwarded interrupts if we're
+	 * using MSI).  The rest of the FL/Intr-capable ingress queues will be
+	 * matched up one-for-one with Ethernet/Control egress queues in order
+	 * to form "Queue Sets" which will be aportioned between the "ports".
+	 * For each Queue Set, we'll need the ability to allocate two Egress
+	 * Contexts -- one for the Ingress Queue Free List and one for the TX
+	 * Ethernet Queue.
+	 */
+	ethqsets = vfres->niqflint - INGQ_EXTRAS;
+	if (vfres->nethctrl != ethqsets) {
+		dev_warn(adapter->pdev_dev, "unequal number of [available]"
+			 " ingress/egress queues (%d/%d); using minimum for"
+			 " number of Queue Sets\n", ethqsets, vfres->nethctrl);
+		ethqsets = min(vfres->nethctrl, ethqsets);
+	}
+	if (vfres->neq < ethqsets*2) {
+		dev_warn(adapter->pdev_dev, "Not enough Egress Contexts (%d)"
+			 " to support Queue Sets (%d); reducing allowed Queue"
+			 " Sets\n", vfres->neq, ethqsets);
+		ethqsets = vfres->neq/2;
+	}
+	if (ethqsets > MAX_ETH_QSETS) {
+		dev_warn(adapter->pdev_dev, "only using %d of %d allowed Queue"
+			 " Sets\n", MAX_ETH_QSETS, adapter->sge.max_ethqsets);
+		ethqsets = MAX_ETH_QSETS;
+	}
+	if (vfres->niq != 0 || vfres->neq > ethqsets*2) {
+		dev_warn(adapter->pdev_dev, "unused resources niq/neq (%d/%d)"
+			 " ignored\n", vfres->niq, vfres->neq - ethqsets*2);
+	}
+	adapter->sge.max_ethqsets = ethqsets;
+
+	/*
+	 * Check for various parameter sanity issues.  Most checks simply
+	 * result in us using fewer resources than our provissioning but we
+	 * do need at least  one "port" with which to work ...
+	 */
+	if (adapter->sge.max_ethqsets < adapter->params.nports) {
+		dev_warn(adapter->pdev_dev, "only using %d of %d available"
+			 " virtual interfaces (too few Queue Sets)\n",
+			 adapter->sge.max_ethqsets, adapter->params.nports);
+		adapter->params.nports = adapter->sge.max_ethqsets;
+	}
+	if (adapter->params.nports == 0) {
+		dev_err(adapter->pdev_dev, "no virtual interfaces configured/"
+			"usable!\n");
+		return -EINVAL;
+	}
+	return 0;
+}
+
+static inline void init_rspq(struct sge_rspq *rspq, u8 timer_idx,
+			     u8 pkt_cnt_idx, unsigned int size,
+			     unsigned int iqe_size)
+{
+	rspq->intr_params = (QINTR_TIMER_IDX(timer_idx) |
+			     (pkt_cnt_idx < SGE_NCOUNTERS ? QINTR_CNT_EN : 0));
+	rspq->pktcnt_idx = (pkt_cnt_idx < SGE_NCOUNTERS
+			    ? pkt_cnt_idx
+			    : 0);
+	rspq->iqe_len = iqe_size;
+	rspq->size = size;
+}
+
+/*
+ * Perform default configuration of DMA queues depending on the number and
+ * type of ports we found and the number of available CPUs.  Most settings can
+ * be modified by the admin via ethtool and cxgbtool prior to the adapter
+ * being brought up for the first time.
+ */
+static void cfg_queues(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+	int q10g, n10g, qidx, pidx, qs;
+	size_t iqe_size;
+
+	/*
+	 * We should not be called till we know how many Queue Sets we can
+	 * support.  In particular, this means that we need to know what kind
+	 * of interrupts we'll be using ...
+	 */
+	BUG_ON((adapter->flags & (USING_MSIX|USING_MSI)) == 0);
+
+	/*
+	 * Count the number of 10GbE Virtual Interfaces that we have.
+	 */
+	n10g = 0;
+	for_each_port(adapter, pidx)
+		n10g += is_x_10g_port(&adap2pinfo(adapter, pidx)->link_cfg);
+
+	/*
+	 * We default to 1 queue per non-10G port and up to # of cores queues
+	 * per 10G port.
+	 */
+	if (n10g == 0)
+		q10g = 0;
+	else {
+		int n1g = (adapter->params.nports - n10g);
+		q10g = (adapter->sge.max_ethqsets - n1g) / n10g;
+		if (q10g > num_online_cpus())
+			q10g = num_online_cpus();
+	}
+
+	/*
+	 * Allocate the "Queue Sets" to the various Virtual Interfaces.
+	 * The layout will be established in setup_sge_queues() when the
+	 * adapter is brough up for the first time.
+	 */
+	qidx = 0;
+	for_each_port(adapter, pidx) {
+		struct port_info *pi = adap2pinfo(adapter, pidx);
+
+		pi->first_qset = qidx;
+		pi->nqsets = is_x_10g_port(&pi->link_cfg) ? q10g : 1;
+		qidx += pi->nqsets;
+	}
+	s->ethqsets = qidx;
+
+	/*
+	 * The Ingress Queue Entry Size for our various Response Queues needs
+	 * to be big enough to accommodate the largest message we can receive
+	 * from the chip/firmware; which is 64 bytes ...
+	 */
+	iqe_size = 64;
+
+	/*
+	 * Set up default Queue Set parameters ...  Start off with the
+	 * shortest interrupt holdoff timer.
+	 */
+	for (qs = 0; qs < s->max_ethqsets; qs++) {
+		struct sge_eth_rxq *rxq = &s->ethrxq[qs];
+		struct sge_eth_txq *txq = &s->ethtxq[qs];
+
+		init_rspq(&rxq->rspq, 0, 0, 1024, iqe_size);
+		rxq->fl.size = 72;
+		txq->q.size = 1024;
+	}
+
+	/*
+	 * The firmware event queue is used for link state changes and
+	 * notifications of TX DMA completions.
+	 */
+	init_rspq(&s->fw_evtq, SGE_TIMER_RSTRT_CNTR, 0, 512, iqe_size);
+
+	/*
+	 * The forwarded interrupt queue is used when we're in MSI interrupt
+	 * mode.  In this mode all interrupts associated with RX queues will
+	 * be forwarded to a single queue which we'll associate with our MSI
+	 * interrupt vector.  The messages dropped in the forwarded interrupt
+	 * queue will indicate which ingress queue needs servicing ...  This
+	 * queue needs to be large enough to accommodate all of the ingress
+	 * queues which are forwarding their interrupt (+1 to prevent the PIDX
+	 * from equalling the CIDX if every ingress queue has an outstanding
+	 * interrupt).  The queue doesn't need to be any larger because no
+	 * ingress queue will ever have more than one outstanding interrupt at
+	 * any time ...
+	 */
+	init_rspq(&s->intrq, SGE_TIMER_RSTRT_CNTR, 0, MSIX_ENTRIES + 1,
+		  iqe_size);
+}
+
+/*
+ * Reduce the number of Ethernet queues across all ports to at most n.
+ * n provides at least one queue per port.
+ */
+static void reduce_ethqs(struct adapter *adapter, int n)
+{
+	int i;
+	struct port_info *pi;
+
+	/*
+	 * While we have too many active Ether Queue Sets, interate across the
+	 * "ports" and reduce their individual Queue Set allocations.
+	 */
+	BUG_ON(n < adapter->params.nports);
+	while (n < adapter->sge.ethqsets)
+		for_each_port(adapter, i) {
+			pi = adap2pinfo(adapter, i);
+			if (pi->nqsets > 1) {
+				pi->nqsets--;
+				adapter->sge.ethqsets--;
+				if (adapter->sge.ethqsets <= n)
+					break;
+			}
+		}
+
+	/*
+	 * Reassign the starting Queue Sets for each of the "ports" ...
+	 */
+	n = 0;
+	for_each_port(adapter, i) {
+		pi = adap2pinfo(adapter, i);
+		pi->first_qset = n;
+		n += pi->nqsets;
+	}
+}
+
+/*
+ * We need to grab enough MSI-X vectors to cover our interrupt needs.  Ideally
+ * we get a separate MSI-X vector for every "Queue Set" plus any extras we
+ * need.  Minimally we need one for every Virtual Interface plus those needed
+ * for our "extras".  Note that this process may lower the maximum number of
+ * allowed Queue Sets ...
+ */
+static int enable_msix(struct adapter *adapter)
+{
+	int i, want, need, nqsets;
+	struct msix_entry entries[MSIX_ENTRIES];
+	struct sge *s = &adapter->sge;
+
+	for (i = 0; i < MSIX_ENTRIES; ++i)
+		entries[i].entry = i;
+
+	/*
+	 * We _want_ enough MSI-X interrupts to cover all of our "Queue Sets"
+	 * plus those needed for our "extras" (for example, the firmware
+	 * message queue).  We _need_ at least one "Queue Set" per Virtual
+	 * Interface plus those needed for our "extras".  So now we get to see
+	 * if the song is right ...
+	 */
+	want = s->max_ethqsets + MSIX_EXTRAS;
+	need = adapter->params.nports + MSIX_EXTRAS;
+
+	want = pci_enable_msix_range(adapter->pdev, entries, need, want);
+	if (want < 0)
+		return want;
+
+	nqsets = want - MSIX_EXTRAS;
+	if (nqsets < s->max_ethqsets) {
+		dev_warn(adapter->pdev_dev, "only enough MSI-X vectors"
+			 " for %d Queue Sets\n", nqsets);
+		s->max_ethqsets = nqsets;
+		if (nqsets < s->ethqsets)
+			reduce_ethqs(adapter, nqsets);
+	}
+	for (i = 0; i < want; ++i)
+		adapter->msix_info[i].vec = entries[i].vector;
+
+	return 0;
+}
+
+static const struct net_device_ops cxgb4vf_netdev_ops	= {
+	.ndo_open		= cxgb4vf_open,
+	.ndo_stop		= cxgb4vf_stop,
+	.ndo_start_xmit		= t4vf_eth_xmit,
+	.ndo_get_stats		= cxgb4vf_get_stats,
+	.ndo_set_rx_mode	= cxgb4vf_set_rxmode,
+	.ndo_set_mac_address	= cxgb4vf_set_mac_addr,
+	.ndo_validate_addr	= eth_validate_addr,
+	.ndo_do_ioctl		= cxgb4vf_do_ioctl,
+	.ndo_change_mtu		= cxgb4vf_change_mtu,
+	.ndo_fix_features	= cxgb4vf_fix_features,
+	.ndo_set_features	= cxgb4vf_set_features,
+#ifdef CONFIG_NET_POLL_CONTROLLER
+	.ndo_poll_controller	= cxgb4vf_poll_controller,
+#endif
+};
+
+/*
+ * "Probe" a device: initialize a device and construct all kernel and driver
+ * state needed to manage the device.  This routine is called "init_one" in
+ * the PF Driver ...
+ */
+static int cxgb4vf_pci_probe(struct pci_dev *pdev,
+			     const struct pci_device_id *ent)
+{
+	int pci_using_dac;
+	int err, pidx;
+	unsigned int pmask;
+	struct adapter *adapter;
+	struct port_info *pi;
+	struct net_device *netdev;
+
+	/*
+	 * Print our driver banner the first time we're called to initialize a
+	 * device.
+	 */
+	pr_info_once("%s - version %s\n", DRV_DESC, DRV_VERSION);
+
+	/*
+	 * Initialize generic PCI device state.
+	 */
+	err = pci_enable_device(pdev);
+	if (err) {
+		dev_err(&pdev->dev, "cannot enable PCI device\n");
+		return err;
+	}
+
+	/*
+	 * Reserve PCI resources for the device.  If we can't get them some
+	 * other driver may have already claimed the device ...
+	 */
+	err = pci_request_regions(pdev, KBUILD_MODNAME);
+	if (err) {
+		dev_err(&pdev->dev, "cannot obtain PCI resources\n");
+		goto err_disable_device;
+	}
+
+	/*
+	 * Set up our DMA mask: try for 64-bit address masking first and
+	 * fall back to 32-bit if we can't get 64 bits ...
+	 */
+	err = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
+	if (err == 0) {
+		err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
+		if (err) {
+			dev_err(&pdev->dev, "unable to obtain 64-bit DMA for"
+				" coherent allocations\n");
+			goto err_release_regions;
+		}
+		pci_using_dac = 1;
+	} else {
+		err = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
+		if (err != 0) {
+			dev_err(&pdev->dev, "no usable DMA configuration\n");
+			goto err_release_regions;
+		}
+		pci_using_dac = 0;
+	}
+
+	/*
+	 * Enable bus mastering for the device ...
+	 */
+	pci_set_master(pdev);
+
+	/*
+	 * Allocate our adapter data structure and attach it to the device.
+	 */
+	adapter = kzalloc(sizeof(*adapter), GFP_KERNEL);
+	if (!adapter) {
+		err = -ENOMEM;
+		goto err_release_regions;
+	}
+	pci_set_drvdata(pdev, adapter);
+	adapter->pdev = pdev;
+	adapter->pdev_dev = &pdev->dev;
+
+	/*
+	 * Initialize SMP data synchronization resources.
+	 */
+	spin_lock_init(&adapter->stats_lock);
+
+	/*
+	 * Map our I/O registers in BAR0.
+	 */
+	adapter->regs = pci_ioremap_bar(pdev, 0);
+	if (!adapter->regs) {
+		dev_err(&pdev->dev, "cannot map device registers\n");
+		err = -ENOMEM;
+		goto err_free_adapter;
+	}
+
+	/* Wait for the device to become ready before proceeding ...
+	 */
+	err = t4vf_prep_adapter(adapter);
+	if (err) {
+		dev_err(adapter->pdev_dev, "device didn't become ready:"
+			" err=%d\n", err);
+		goto err_unmap_bar0;
+	}
+
+	/* For T5 and later we want to use the new BAR-based User Doorbells,
+	 * so we need to map BAR2 here ...
+	 */
+	if (!is_t4(adapter->params.chip)) {
+		adapter->bar2 = ioremap_wc(pci_resource_start(pdev, 2),
+					   pci_resource_len(pdev, 2));
+		if (!adapter->bar2) {
+			dev_err(adapter->pdev_dev, "cannot map BAR2 doorbells\n");
+			err = -ENOMEM;
+			goto err_unmap_bar0;
+		}
+	}
+	/*
+	 * Initialize adapter level features.
+	 */
+	adapter->name = pci_name(pdev);
+	adapter->msg_enable = dflt_msg_enable;
+	err = adap_init0(adapter);
+	if (err)
+		goto err_unmap_bar;
+
+	/*
+	 * Allocate our "adapter ports" and stitch everything together.
+	 */
+	pmask = adapter->params.vfres.pmask;
+	for_each_port(adapter, pidx) {
+		int port_id, viid;
+
+		/*
+		 * We simplistically allocate our virtual interfaces
+		 * sequentially across the port numbers to which we have
+		 * access rights.  This should be configurable in some manner
+		 * ...
+		 */
+		if (pmask == 0)
+			break;
+		port_id = ffs(pmask) - 1;
+		pmask &= ~(1 << port_id);
+		viid = t4vf_alloc_vi(adapter, port_id);
+		if (viid < 0) {
+			dev_err(&pdev->dev, "cannot allocate VI for port %d:"
+				" err=%d\n", port_id, viid);
+			err = viid;
+			goto err_free_dev;
+		}
+
+		/*
+		 * Allocate our network device and stitch things together.
+		 */
+		netdev = alloc_etherdev_mq(sizeof(struct port_info),
+					   MAX_PORT_QSETS);
+		if (netdev == NULL) {
+			t4vf_free_vi(adapter, viid);
+			err = -ENOMEM;
+			goto err_free_dev;
+		}
+		adapter->port[pidx] = netdev;
+		SET_NETDEV_DEV(netdev, &pdev->dev);
+		pi = netdev_priv(netdev);
+		pi->adapter = adapter;
+		pi->pidx = pidx;
+		pi->port_id = port_id;
+		pi->viid = viid;
+
+		/*
+		 * Initialize the starting state of our "port" and register
+		 * it.
+		 */
+		pi->xact_addr_filt = -1;
+		netif_carrier_off(netdev);
+		netdev->irq = pdev->irq;
+
+		netdev->hw_features = NETIF_F_SG | TSO_FLAGS |
+			NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
+			NETIF_F_HW_VLAN_CTAG_RX | NETIF_F_RXCSUM;
+		netdev->vlan_features = NETIF_F_SG | TSO_FLAGS |
+			NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
+			NETIF_F_HIGHDMA;
+		netdev->features = netdev->hw_features |
+				   NETIF_F_HW_VLAN_CTAG_TX;
+		if (pci_using_dac)
+			netdev->features |= NETIF_F_HIGHDMA;
+
+		netdev->priv_flags |= IFF_UNICAST_FLT;
+
+		netdev->netdev_ops = &cxgb4vf_netdev_ops;
+		netdev->ethtool_ops = &cxgb4vf_ethtool_ops;
+
+		/*
+		 * Initialize the hardware/software state for the port.
+		 */
+		err = t4vf_port_init(adapter, pidx);
+		if (err) {
+			dev_err(&pdev->dev, "cannot initialize port %d\n",
+				pidx);
+			goto err_free_dev;
+		}
+	}
+
+	/*
+	 * The "card" is now ready to go.  If any errors occur during device
+	 * registration we do not fail the whole "card" but rather proceed
+	 * only with the ports we manage to register successfully.  However we
+	 * must register at least one net device.
+	 */
+	for_each_port(adapter, pidx) {
+		netdev = adapter->port[pidx];
+		if (netdev == NULL)
+			continue;
+
+		err = register_netdev(netdev);
+		if (err) {
+			dev_warn(&pdev->dev, "cannot register net device %s,"
+				 " skipping\n", netdev->name);
+			continue;
+		}
+
+		set_bit(pidx, &adapter->registered_device_map);
+	}
+	if (adapter->registered_device_map == 0) {
+		dev_err(&pdev->dev, "could not register any net devices\n");
+		goto err_free_dev;
+	}
+
+	/*
+	 * Set up our debugfs entries.
+	 */
+	if (!IS_ERR_OR_NULL(cxgb4vf_debugfs_root)) {
+		adapter->debugfs_root =
+			debugfs_create_dir(pci_name(pdev),
+					   cxgb4vf_debugfs_root);
+		if (IS_ERR_OR_NULL(adapter->debugfs_root))
+			dev_warn(&pdev->dev, "could not create debugfs"
+				 " directory");
+		else
+			setup_debugfs(adapter);
+	}
+
+	/*
+	 * See what interrupts we'll be using.  If we've been configured to
+	 * use MSI-X interrupts, try to enable them but fall back to using
+	 * MSI interrupts if we can't enable MSI-X interrupts.  If we can't
+	 * get MSI interrupts we bail with the error.
+	 */
+	if (msi == MSI_MSIX && enable_msix(adapter) == 0)
+		adapter->flags |= USING_MSIX;
+	else {
+		err = pci_enable_msi(pdev);
+		if (err) {
+			dev_err(&pdev->dev, "Unable to allocate %s interrupts;"
+				" err=%d\n",
+				msi == MSI_MSIX ? "MSI-X or MSI" : "MSI", err);
+			goto err_free_debugfs;
+		}
+		adapter->flags |= USING_MSI;
+	}
+
+	/*
+	 * Now that we know how many "ports" we have and what their types are,
+	 * and how many Queue Sets we can support, we can configure our queue
+	 * resources.
+	 */
+	cfg_queues(adapter);
+
+	/*
+	 * Print a short notice on the existence and configuration of the new
+	 * VF network device ...
+	 */
+	for_each_port(adapter, pidx) {
+		dev_info(adapter->pdev_dev, "%s: Chelsio VF NIC PCIe %s\n",
+			 adapter->port[pidx]->name,
+			 (adapter->flags & USING_MSIX) ? "MSI-X" :
+			 (adapter->flags & USING_MSI)  ? "MSI" : "");
+	}
+
+	/*
+	 * Return success!
+	 */
+	return 0;
+
+	/*
+	 * Error recovery and exit code.  Unwind state that's been created
+	 * so far and return the error.
+	 */
+
+err_free_debugfs:
+	if (!IS_ERR_OR_NULL(adapter->debugfs_root)) {
+		cleanup_debugfs(adapter);
+		debugfs_remove_recursive(adapter->debugfs_root);
+	}
+
+err_free_dev:
+	for_each_port(adapter, pidx) {
+		netdev = adapter->port[pidx];
+		if (netdev == NULL)
+			continue;
+		pi = netdev_priv(netdev);
+		t4vf_free_vi(adapter, pi->viid);
+		if (test_bit(pidx, &adapter->registered_device_map))
+			unregister_netdev(netdev);
+		free_netdev(netdev);
+	}
+
+err_unmap_bar:
+	if (!is_t4(adapter->params.chip))
+		iounmap(adapter->bar2);
+
+err_unmap_bar0:
+	iounmap(adapter->regs);
+
+err_free_adapter:
+	kfree(adapter);
+
+err_release_regions:
+	pci_release_regions(pdev);
+	pci_clear_master(pdev);
+
+err_disable_device:
+	pci_disable_device(pdev);
+
+	return err;
+}
+
+/*
+ * "Remove" a device: tear down all kernel and driver state created in the
+ * "probe" routine and quiesce the device (disable interrupts, etc.).  (Note
+ * that this is called "remove_one" in the PF Driver.)
+ */
+static void cxgb4vf_pci_remove(struct pci_dev *pdev)
+{
+	struct adapter *adapter = pci_get_drvdata(pdev);
+
+	/*
+	 * Tear down driver state associated with device.
+	 */
+	if (adapter) {
+		int pidx;
+
+		/*
+		 * Stop all of our activity.  Unregister network port,
+		 * disable interrupts, etc.
+		 */
+		for_each_port(adapter, pidx)
+			if (test_bit(pidx, &adapter->registered_device_map))
+				unregister_netdev(adapter->port[pidx]);
+		t4vf_sge_stop(adapter);
+		if (adapter->flags & USING_MSIX) {
+			pci_disable_msix(adapter->pdev);
+			adapter->flags &= ~USING_MSIX;
+		} else if (adapter->flags & USING_MSI) {
+			pci_disable_msi(adapter->pdev);
+			adapter->flags &= ~USING_MSI;
+		}
+
+		/*
+		 * Tear down our debugfs entries.
+		 */
+		if (!IS_ERR_OR_NULL(adapter->debugfs_root)) {
+			cleanup_debugfs(adapter);
+			debugfs_remove_recursive(adapter->debugfs_root);
+		}
+
+		/*
+		 * Free all of the various resources which we've acquired ...
+		 */
+		t4vf_free_sge_resources(adapter);
+		for_each_port(adapter, pidx) {
+			struct net_device *netdev = adapter->port[pidx];
+			struct port_info *pi;
+
+			if (netdev == NULL)
+				continue;
+
+			pi = netdev_priv(netdev);
+			t4vf_free_vi(adapter, pi->viid);
+			free_netdev(netdev);
+		}
+		iounmap(adapter->regs);
+		if (!is_t4(adapter->params.chip))
+			iounmap(adapter->bar2);
+		kfree(adapter);
+	}
+
+	/*
+	 * Disable the device and release its PCI resources.
+	 */
+	pci_disable_device(pdev);
+	pci_clear_master(pdev);
+	pci_release_regions(pdev);
+}
+
+/*
+ * "Shutdown" quiesce the device, stopping Ingress Packet and Interrupt
+ * delivery.
+ */
+static void cxgb4vf_pci_shutdown(struct pci_dev *pdev)
+{
+	struct adapter *adapter;
+	int pidx;
+
+	adapter = pci_get_drvdata(pdev);
+	if (!adapter)
+		return;
+
+	/* Disable all Virtual Interfaces.  This will shut down the
+	 * delivery of all ingress packets into the chip for these
+	 * Virtual Interfaces.
+	 */
+	for_each_port(adapter, pidx)
+		if (test_bit(pidx, &adapter->registered_device_map))
+			unregister_netdev(adapter->port[pidx]);
+
+	/* Free up all Queues which will prevent further DMA and
+	 * Interrupts allowing various internal pathways to drain.
+	 */
+	t4vf_sge_stop(adapter);
+	if (adapter->flags & USING_MSIX) {
+		pci_disable_msix(adapter->pdev);
+		adapter->flags &= ~USING_MSIX;
+	} else if (adapter->flags & USING_MSI) {
+		pci_disable_msi(adapter->pdev);
+		adapter->flags &= ~USING_MSI;
+	}
+
+	/*
+	 * Free up all Queues which will prevent further DMA and
+	 * Interrupts allowing various internal pathways to drain.
+	 */
+	t4vf_free_sge_resources(adapter);
+	pci_set_drvdata(pdev, NULL);
+}
+
+/* Macros needed to support the PCI Device ID Table ...
+ */
+#define CH_PCI_DEVICE_ID_TABLE_DEFINE_BEGIN \
+	static const struct pci_device_id cxgb4vf_pci_tbl[] = {
+#define CH_PCI_DEVICE_ID_FUNCTION	0x8
+
+#define CH_PCI_ID_TABLE_ENTRY(devid) \
+		{ PCI_VDEVICE(CHELSIO, (devid)), 0 }
+
+#define CH_PCI_DEVICE_ID_TABLE_DEFINE_END { 0, } }
+
+#include "../cxgb4/t4_pci_id_tbl.h"
+
+MODULE_DESCRIPTION(DRV_DESC);
+MODULE_AUTHOR("Chelsio Communications");
+MODULE_LICENSE("Dual BSD/GPL");
+MODULE_VERSION(DRV_VERSION);
+MODULE_DEVICE_TABLE(pci, cxgb4vf_pci_tbl);
+
+static struct pci_driver cxgb4vf_driver = {
+	.name		= KBUILD_MODNAME,
+	.id_table	= cxgb4vf_pci_tbl,
+	.probe		= cxgb4vf_pci_probe,
+	.remove		= cxgb4vf_pci_remove,
+	.shutdown	= cxgb4vf_pci_shutdown,
+};
+
+/*
+ * Initialize global driver state.
+ */
+static int __init cxgb4vf_module_init(void)
+{
+	int ret;
+
+	/*
+	 * Vet our module parameters.
+	 */
+	if (msi != MSI_MSIX && msi != MSI_MSI) {
+		pr_warn("bad module parameter msi=%d; must be %d (MSI-X or MSI) or %d (MSI)\n",
+			msi, MSI_MSIX, MSI_MSI);
+		return -EINVAL;
+	}
+
+	/* Debugfs support is optional, just warn if this fails */
+	cxgb4vf_debugfs_root = debugfs_create_dir(KBUILD_MODNAME, NULL);
+	if (IS_ERR_OR_NULL(cxgb4vf_debugfs_root))
+		pr_warn("could not create debugfs entry, continuing\n");
+
+	ret = pci_register_driver(&cxgb4vf_driver);
+	if (ret < 0 && !IS_ERR_OR_NULL(cxgb4vf_debugfs_root))
+		debugfs_remove(cxgb4vf_debugfs_root);
+	return ret;
+}
+
+/*
+ * Tear down global driver state.
+ */
+static void __exit cxgb4vf_module_exit(void)
+{
+	pci_unregister_driver(&cxgb4vf_driver);
+	debugfs_remove(cxgb4vf_debugfs_root);
+}
+
+module_init(cxgb4vf_module_init);
+module_exit(cxgb4vf_module_exit);
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/sge.c b/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
new file mode 100644
index 000000000..482f6de68
--- /dev/null
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
@@ -0,0 +1,2655 @@
+/*
+ * This file is part of the Chelsio T4 PCI-E SR-IOV Virtual Function Ethernet
+ * driver for Linux.
+ *
+ * Copyright (c) 2009-2010 Chelsio Communications, 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.
+ */
+
+#include <linux/skbuff.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/if_vlan.h>
+#include <linux/ip.h>
+#include <net/ipv6.h>
+#include <net/tcp.h>
+#include <linux/dma-mapping.h>
+#include <linux/prefetch.h>
+
+#include "t4vf_common.h"
+#include "t4vf_defs.h"
+
+#include "../cxgb4/t4_regs.h"
+#include "../cxgb4/t4_values.h"
+#include "../cxgb4/t4fw_api.h"
+#include "../cxgb4/t4_msg.h"
+
+/*
+ * Constants ...
+ */
+enum {
+	/*
+	 * Egress Queue sizes, producer and consumer indices are all in units
+	 * of Egress Context Units bytes.  Note that as far as the hardware is
+	 * concerned, the free list is an Egress Queue (the host produces free
+	 * buffers which the hardware consumes) and free list entries are
+	 * 64-bit PCI DMA addresses.
+	 */
+	EQ_UNIT = SGE_EQ_IDXSIZE,
+	FL_PER_EQ_UNIT = EQ_UNIT / sizeof(__be64),
+	TXD_PER_EQ_UNIT = EQ_UNIT / sizeof(__be64),
+
+	/*
+	 * Max number of TX descriptors we clean up at a time.  Should be
+	 * modest as freeing skbs isn't cheap and it happens while holding
+	 * locks.  We just need to free packets faster than they arrive, we
+	 * eventually catch up and keep the amortized cost reasonable.
+	 */
+	MAX_TX_RECLAIM = 16,
+
+	/*
+	 * Max number of Rx buffers we replenish at a time.  Again keep this
+	 * modest, allocating buffers isn't cheap either.
+	 */
+	MAX_RX_REFILL = 16,
+
+	/*
+	 * Period of the Rx queue check timer.  This timer is infrequent as it
+	 * has something to do only when the system experiences severe memory
+	 * shortage.
+	 */
+	RX_QCHECK_PERIOD = (HZ / 2),
+
+	/*
+	 * Period of the TX queue check timer and the maximum number of TX
+	 * descriptors to be reclaimed by the TX timer.
+	 */
+	TX_QCHECK_PERIOD = (HZ / 2),
+	MAX_TIMER_TX_RECLAIM = 100,
+
+	/*
+	 * Suspend an Ethernet TX queue with fewer available descriptors than
+	 * this.  We always want to have room for a maximum sized packet:
+	 * inline immediate data + MAX_SKB_FRAGS. This is the same as
+	 * calc_tx_flits() for a TSO packet with nr_frags == MAX_SKB_FRAGS
+	 * (see that function and its helpers for a description of the
+	 * calculation).
+	 */
+	ETHTXQ_MAX_FRAGS = MAX_SKB_FRAGS + 1,
+	ETHTXQ_MAX_SGL_LEN = ((3 * (ETHTXQ_MAX_FRAGS-1))/2 +
+				   ((ETHTXQ_MAX_FRAGS-1) & 1) +
+				   2),
+	ETHTXQ_MAX_HDR = (sizeof(struct fw_eth_tx_pkt_vm_wr) +
+			  sizeof(struct cpl_tx_pkt_lso_core) +
+			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64),
+	ETHTXQ_MAX_FLITS = ETHTXQ_MAX_SGL_LEN + ETHTXQ_MAX_HDR,
+
+	ETHTXQ_STOP_THRES = 1 + DIV_ROUND_UP(ETHTXQ_MAX_FLITS, TXD_PER_EQ_UNIT),
+
+	/*
+	 * Max TX descriptor space we allow for an Ethernet packet to be
+	 * inlined into a WR.  This is limited by the maximum value which
+	 * we can specify for immediate data in the firmware Ethernet TX
+	 * Work Request.
+	 */
+	MAX_IMM_TX_PKT_LEN = FW_WR_IMMDLEN_M,
+
+	/*
+	 * Max size of a WR sent through a control TX queue.
+	 */
+	MAX_CTRL_WR_LEN = 256,
+
+	/*
+	 * Maximum amount of data which we'll ever need to inline into a
+	 * TX ring: max(MAX_IMM_TX_PKT_LEN, MAX_CTRL_WR_LEN).
+	 */
+	MAX_IMM_TX_LEN = (MAX_IMM_TX_PKT_LEN > MAX_CTRL_WR_LEN
+			  ? MAX_IMM_TX_PKT_LEN
+			  : MAX_CTRL_WR_LEN),
+
+	/*
+	 * For incoming packets less than RX_COPY_THRES, we copy the data into
+	 * an skb rather than referencing the data.  We allocate enough
+	 * in-line room in skb's to accommodate pulling in RX_PULL_LEN bytes
+	 * of the data (header).
+	 */
+	RX_COPY_THRES = 256,
+	RX_PULL_LEN = 128,
+
+	/*
+	 * Main body length for sk_buffs used for RX Ethernet packets with
+	 * fragments.  Should be >= RX_PULL_LEN but possibly bigger to give
+	 * pskb_may_pull() some room.
+	 */
+	RX_SKB_LEN = 512,
+};
+
+/*
+ * Software state per TX descriptor.
+ */
+struct tx_sw_desc {
+	struct sk_buff *skb;		/* socket buffer of TX data source */
+	struct ulptx_sgl *sgl;		/* scatter/gather list in TX Queue */
+};
+
+/*
+ * Software state per RX Free List descriptor.  We keep track of the allocated
+ * FL page, its size, and its PCI DMA address (if the page is mapped).  The FL
+ * page size and its PCI DMA mapped state are stored in the low bits of the
+ * PCI DMA address as per below.
+ */
+struct rx_sw_desc {
+	struct page *page;		/* Free List page buffer */
+	dma_addr_t dma_addr;		/* PCI DMA address (if mapped) */
+					/*   and flags (see below) */
+};
+
+/*
+ * The low bits of rx_sw_desc.dma_addr have special meaning.  Note that the
+ * SGE also uses the low 4 bits to determine the size of the buffer.  It uses
+ * those bits to index into the SGE_FL_BUFFER_SIZE[index] register array.
+ * Since we only use SGE_FL_BUFFER_SIZE0 and SGE_FL_BUFFER_SIZE1, these low 4
+ * bits can only contain a 0 or a 1 to indicate which size buffer we're giving
+ * to the SGE.  Thus, our software state of "is the buffer mapped for DMA" is
+ * maintained in an inverse sense so the hardware never sees that bit high.
+ */
+enum {
+	RX_LARGE_BUF    = 1 << 0,	/* buffer is SGE_FL_BUFFER_SIZE[1] */
+	RX_UNMAPPED_BUF = 1 << 1,	/* buffer is not mapped */
+};
+
+/**
+ *	get_buf_addr - return DMA buffer address of software descriptor
+ *	@sdesc: pointer to the software buffer descriptor
+ *
+ *	Return the DMA buffer address of a software descriptor (stripping out
+ *	our low-order flag bits).
+ */
+static inline dma_addr_t get_buf_addr(const struct rx_sw_desc *sdesc)
+{
+	return sdesc->dma_addr & ~(dma_addr_t)(RX_LARGE_BUF | RX_UNMAPPED_BUF);
+}
+
+/**
+ *	is_buf_mapped - is buffer mapped for DMA?
+ *	@sdesc: pointer to the software buffer descriptor
+ *
+ *	Determine whether the buffer associated with a software descriptor in
+ *	mapped for DMA or not.
+ */
+static inline bool is_buf_mapped(const struct rx_sw_desc *sdesc)
+{
+	return !(sdesc->dma_addr & RX_UNMAPPED_BUF);
+}
+
+/**
+ *	need_skb_unmap - does the platform need unmapping of sk_buffs?
+ *
+ *	Returns true if the platform needs sk_buff unmapping.  The compiler
+ *	optimizes away unnecessary code if this returns true.
+ */
+static inline int need_skb_unmap(void)
+{
+#ifdef CONFIG_NEED_DMA_MAP_STATE
+	return 1;
+#else
+	return 0;
+#endif
+}
+
+/**
+ *	txq_avail - return the number of available slots in a TX queue
+ *	@tq: the TX queue
+ *
+ *	Returns the number of available descriptors in a TX queue.
+ */
+static inline unsigned int txq_avail(const struct sge_txq *tq)
+{
+	return tq->size - 1 - tq->in_use;
+}
+
+/**
+ *	fl_cap - return the capacity of a Free List
+ *	@fl: the Free List
+ *
+ *	Returns the capacity of a Free List.  The capacity is less than the
+ *	size because an Egress Queue Index Unit worth of descriptors needs to
+ *	be left unpopulated, otherwise the Producer and Consumer indices PIDX
+ *	and CIDX will match and the hardware will think the FL is empty.
+ */
+static inline unsigned int fl_cap(const struct sge_fl *fl)
+{
+	return fl->size - FL_PER_EQ_UNIT;
+}
+
+/**
+ *	fl_starving - return whether a Free List is starving.
+ *	@adapter: pointer to the adapter
+ *	@fl: the Free List
+ *
+ *	Tests specified Free List to see whether the number of buffers
+ *	available to the hardware has falled below our "starvation"
+ *	threshold.
+ */
+static inline bool fl_starving(const struct adapter *adapter,
+			       const struct sge_fl *fl)
+{
+	const struct sge *s = &adapter->sge;
+
+	return fl->avail - fl->pend_cred <= s->fl_starve_thres;
+}
+
+/**
+ *	map_skb -  map an skb for DMA to the device
+ *	@dev: the egress net device
+ *	@skb: the packet to map
+ *	@addr: a pointer to the base of the DMA mapping array
+ *
+ *	Map an skb for DMA to the device and return an array of DMA addresses.
+ */
+static int map_skb(struct device *dev, const struct sk_buff *skb,
+		   dma_addr_t *addr)
+{
+	const skb_frag_t *fp, *end;
+	const struct skb_shared_info *si;
+
+	*addr = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE);
+	if (dma_mapping_error(dev, *addr))
+		goto out_err;
+
+	si = skb_shinfo(skb);
+	end = &si->frags[si->nr_frags];
+	for (fp = si->frags; fp < end; fp++) {
+		*++addr = skb_frag_dma_map(dev, fp, 0, skb_frag_size(fp),
+					   DMA_TO_DEVICE);
+		if (dma_mapping_error(dev, *addr))
+			goto unwind;
+	}
+	return 0;
+
+unwind:
+	while (fp-- > si->frags)
+		dma_unmap_page(dev, *--addr, skb_frag_size(fp), DMA_TO_DEVICE);
+	dma_unmap_single(dev, addr[-1], skb_headlen(skb), DMA_TO_DEVICE);
+
+out_err:
+	return -ENOMEM;
+}
+
+static void unmap_sgl(struct device *dev, const struct sk_buff *skb,
+		      const struct ulptx_sgl *sgl, const struct sge_txq *tq)
+{
+	const struct ulptx_sge_pair *p;
+	unsigned int nfrags = skb_shinfo(skb)->nr_frags;
+
+	if (likely(skb_headlen(skb)))
+		dma_unmap_single(dev, be64_to_cpu(sgl->addr0),
+				 be32_to_cpu(sgl->len0), DMA_TO_DEVICE);
+	else {
+		dma_unmap_page(dev, be64_to_cpu(sgl->addr0),
+			       be32_to_cpu(sgl->len0), DMA_TO_DEVICE);
+		nfrags--;
+	}
+
+	/*
+	 * the complexity below is because of the possibility of a wrap-around
+	 * in the middle of an SGL
+	 */
+	for (p = sgl->sge; nfrags >= 2; nfrags -= 2) {
+		if (likely((u8 *)(p + 1) <= (u8 *)tq->stat)) {
+unmap:
+			dma_unmap_page(dev, be64_to_cpu(p->addr[0]),
+				       be32_to_cpu(p->len[0]), DMA_TO_DEVICE);
+			dma_unmap_page(dev, be64_to_cpu(p->addr[1]),
+				       be32_to_cpu(p->len[1]), DMA_TO_DEVICE);
+			p++;
+		} else if ((u8 *)p == (u8 *)tq->stat) {
+			p = (const struct ulptx_sge_pair *)tq->desc;
+			goto unmap;
+		} else if ((u8 *)p + 8 == (u8 *)tq->stat) {
+			const __be64 *addr = (const __be64 *)tq->desc;
+
+			dma_unmap_page(dev, be64_to_cpu(addr[0]),
+				       be32_to_cpu(p->len[0]), DMA_TO_DEVICE);
+			dma_unmap_page(dev, be64_to_cpu(addr[1]),
+				       be32_to_cpu(p->len[1]), DMA_TO_DEVICE);
+			p = (const struct ulptx_sge_pair *)&addr[2];
+		} else {
+			const __be64 *addr = (const __be64 *)tq->desc;
+
+			dma_unmap_page(dev, be64_to_cpu(p->addr[0]),
+				       be32_to_cpu(p->len[0]), DMA_TO_DEVICE);
+			dma_unmap_page(dev, be64_to_cpu(addr[0]),
+				       be32_to_cpu(p->len[1]), DMA_TO_DEVICE);
+			p = (const struct ulptx_sge_pair *)&addr[1];
+		}
+	}
+	if (nfrags) {
+		__be64 addr;
+
+		if ((u8 *)p == (u8 *)tq->stat)
+			p = (const struct ulptx_sge_pair *)tq->desc;
+		addr = ((u8 *)p + 16 <= (u8 *)tq->stat
+			? p->addr[0]
+			: *(const __be64 *)tq->desc);
+		dma_unmap_page(dev, be64_to_cpu(addr), be32_to_cpu(p->len[0]),
+			       DMA_TO_DEVICE);
+	}
+}
+
+/**
+ *	free_tx_desc - reclaims TX descriptors and their buffers
+ *	@adapter: the adapter
+ *	@tq: the TX queue to reclaim descriptors from
+ *	@n: the number of descriptors to reclaim
+ *	@unmap: whether the buffers should be unmapped for DMA
+ *
+ *	Reclaims TX descriptors from an SGE TX queue and frees the associated
+ *	TX buffers.  Called with the TX queue lock held.
+ */
+static void free_tx_desc(struct adapter *adapter, struct sge_txq *tq,
+			 unsigned int n, bool unmap)
+{
+	struct tx_sw_desc *sdesc;
+	unsigned int cidx = tq->cidx;
+	struct device *dev = adapter->pdev_dev;
+
+	const int need_unmap = need_skb_unmap() && unmap;
+
+	sdesc = &tq->sdesc[cidx];
+	while (n--) {
+		/*
+		 * If we kept a reference to the original TX skb, we need to
+		 * unmap it from PCI DMA space (if required) and free it.
+		 */
+		if (sdesc->skb) {
+			if (need_unmap)
+				unmap_sgl(dev, sdesc->skb, sdesc->sgl, tq);
+			dev_consume_skb_any(sdesc->skb);
+			sdesc->skb = NULL;
+		}
+
+		sdesc++;
+		if (++cidx == tq->size) {
+			cidx = 0;
+			sdesc = tq->sdesc;
+		}
+	}
+	tq->cidx = cidx;
+}
+
+/*
+ * Return the number of reclaimable descriptors in a TX queue.
+ */
+static inline int reclaimable(const struct sge_txq *tq)
+{
+	int hw_cidx = be16_to_cpu(tq->stat->cidx);
+	int reclaimable = hw_cidx - tq->cidx;
+	if (reclaimable < 0)
+		reclaimable += tq->size;
+	return reclaimable;
+}
+
+/**
+ *	reclaim_completed_tx - reclaims completed TX descriptors
+ *	@adapter: the adapter
+ *	@tq: the TX queue to reclaim completed descriptors from
+ *	@unmap: whether the buffers should be unmapped for DMA
+ *
+ *	Reclaims TX descriptors that the SGE has indicated it has processed,
+ *	and frees the associated buffers if possible.  Called with the TX
+ *	queue locked.
+ */
+static inline void reclaim_completed_tx(struct adapter *adapter,
+					struct sge_txq *tq,
+					bool unmap)
+{
+	int avail = reclaimable(tq);
+
+	if (avail) {
+		/*
+		 * Limit the amount of clean up work we do at a time to keep
+		 * the TX lock hold time O(1).
+		 */
+		if (avail > MAX_TX_RECLAIM)
+			avail = MAX_TX_RECLAIM;
+
+		free_tx_desc(adapter, tq, avail, unmap);
+		tq->in_use -= avail;
+	}
+}
+
+/**
+ *	get_buf_size - return the size of an RX Free List buffer.
+ *	@adapter: pointer to the associated adapter
+ *	@sdesc: pointer to the software buffer descriptor
+ */
+static inline int get_buf_size(const struct adapter *adapter,
+			       const struct rx_sw_desc *sdesc)
+{
+	const struct sge *s = &adapter->sge;
+
+	return (s->fl_pg_order > 0 && (sdesc->dma_addr & RX_LARGE_BUF)
+		? (PAGE_SIZE << s->fl_pg_order) : PAGE_SIZE);
+}
+
+/**
+ *	free_rx_bufs - free RX buffers on an SGE Free List
+ *	@adapter: the adapter
+ *	@fl: the SGE Free List to free buffers from
+ *	@n: how many buffers to free
+ *
+ *	Release the next @n buffers on an SGE Free List RX queue.   The
+ *	buffers must be made inaccessible to hardware before calling this
+ *	function.
+ */
+static void free_rx_bufs(struct adapter *adapter, struct sge_fl *fl, int n)
+{
+	while (n--) {
+		struct rx_sw_desc *sdesc = &fl->sdesc[fl->cidx];
+
+		if (is_buf_mapped(sdesc))
+			dma_unmap_page(adapter->pdev_dev, get_buf_addr(sdesc),
+				       get_buf_size(adapter, sdesc),
+				       PCI_DMA_FROMDEVICE);
+		put_page(sdesc->page);
+		sdesc->page = NULL;
+		if (++fl->cidx == fl->size)
+			fl->cidx = 0;
+		fl->avail--;
+	}
+}
+
+/**
+ *	unmap_rx_buf - unmap the current RX buffer on an SGE Free List
+ *	@adapter: the adapter
+ *	@fl: the SGE Free List
+ *
+ *	Unmap the current buffer on an SGE Free List RX queue.   The
+ *	buffer must be made inaccessible to HW before calling this function.
+ *
+ *	This is similar to @free_rx_bufs above but does not free the buffer.
+ *	Do note that the FL still loses any further access to the buffer.
+ *	This is used predominantly to "transfer ownership" of an FL buffer
+ *	to another entity (typically an skb's fragment list).
+ */
+static void unmap_rx_buf(struct adapter *adapter, struct sge_fl *fl)
+{
+	struct rx_sw_desc *sdesc = &fl->sdesc[fl->cidx];
+
+	if (is_buf_mapped(sdesc))
+		dma_unmap_page(adapter->pdev_dev, get_buf_addr(sdesc),
+			       get_buf_size(adapter, sdesc),
+			       PCI_DMA_FROMDEVICE);
+	sdesc->page = NULL;
+	if (++fl->cidx == fl->size)
+		fl->cidx = 0;
+	fl->avail--;
+}
+
+/**
+ *	ring_fl_db - righ doorbell on free list
+ *	@adapter: the adapter
+ *	@fl: the Free List whose doorbell should be rung ...
+ *
+ *	Tell the Scatter Gather Engine that there are new free list entries
+ *	available.
+ */
+static inline void ring_fl_db(struct adapter *adapter, struct sge_fl *fl)
+{
+	u32 val;
+
+	/* The SGE keeps track of its Producer and Consumer Indices in terms
+	 * of Egress Queue Units so we can only tell it about integral numbers
+	 * of multiples of Free List Entries per Egress Queue Units ...
+	 */
+	if (fl->pend_cred >= FL_PER_EQ_UNIT) {
+		if (is_t4(adapter->params.chip))
+			val = PIDX_V(fl->pend_cred / FL_PER_EQ_UNIT);
+		else
+			val = PIDX_T5_V(fl->pend_cred / FL_PER_EQ_UNIT) |
+			      DBTYPE_F;
+		val |= DBPRIO_F;
+
+		/* Make sure all memory writes to the Free List queue are
+		 * committed before we tell the hardware about them.
+		 */
+		wmb();
+
+		/* If we don't have access to the new User Doorbell (T5+), use
+		 * the old doorbell mechanism; otherwise use the new BAR2
+		 * mechanism.
+		 */
+		if (unlikely(fl->bar2_addr == NULL)) {
+			t4_write_reg(adapter,
+				     T4VF_SGE_BASE_ADDR + SGE_VF_KDOORBELL,
+				     QID_V(fl->cntxt_id) | val);
+		} else {
+			writel(val | QID_V(fl->bar2_qid),
+			       fl->bar2_addr + SGE_UDB_KDOORBELL);
+
+			/* This Write memory Barrier will force the write to
+			 * the User Doorbell area to be flushed.
+			 */
+			wmb();
+		}
+		fl->pend_cred %= FL_PER_EQ_UNIT;
+	}
+}
+
+/**
+ *	set_rx_sw_desc - initialize software RX buffer descriptor
+ *	@sdesc: pointer to the softwore RX buffer descriptor
+ *	@page: pointer to the page data structure backing the RX buffer
+ *	@dma_addr: PCI DMA address (possibly with low-bit flags)
+ */
+static inline void set_rx_sw_desc(struct rx_sw_desc *sdesc, struct page *page,
+				  dma_addr_t dma_addr)
+{
+	sdesc->page = page;
+	sdesc->dma_addr = dma_addr;
+}
+
+/*
+ * Support for poisoning RX buffers ...
+ */
+#define POISON_BUF_VAL -1
+
+static inline void poison_buf(struct page *page, size_t sz)
+{
+#if POISON_BUF_VAL >= 0
+	memset(page_address(page), POISON_BUF_VAL, sz);
+#endif
+}
+
+/**
+ *	refill_fl - refill an SGE RX buffer ring
+ *	@adapter: the adapter
+ *	@fl: the Free List ring to refill
+ *	@n: the number of new buffers to allocate
+ *	@gfp: the gfp flags for the allocations
+ *
+ *	(Re)populate an SGE free-buffer queue with up to @n new packet buffers,
+ *	allocated with the supplied gfp flags.  The caller must assure that
+ *	@n does not exceed the queue's capacity -- i.e. (cidx == pidx) _IN
+ *	EGRESS QUEUE UNITS_ indicates an empty Free List!  Returns the number
+ *	of buffers allocated.  If afterwards the queue is found critically low,
+ *	mark it as starving in the bitmap of starving FLs.
+ */
+static unsigned int refill_fl(struct adapter *adapter, struct sge_fl *fl,
+			      int n, gfp_t gfp)
+{
+	struct sge *s = &adapter->sge;
+	struct page *page;
+	dma_addr_t dma_addr;
+	unsigned int cred = fl->avail;
+	__be64 *d = &fl->desc[fl->pidx];
+	struct rx_sw_desc *sdesc = &fl->sdesc[fl->pidx];
+
+	/*
+	 * Sanity: ensure that the result of adding n Free List buffers
+	 * won't result in wrapping the SGE's Producer Index around to
+	 * it's Consumer Index thereby indicating an empty Free List ...
+	 */
+	BUG_ON(fl->avail + n > fl->size - FL_PER_EQ_UNIT);
+
+	gfp |= __GFP_NOWARN;
+
+	/*
+	 * If we support large pages, prefer large buffers and fail over to
+	 * small pages if we can't allocate large pages to satisfy the refill.
+	 * If we don't support large pages, drop directly into the small page
+	 * allocation code.
+	 */
+	if (s->fl_pg_order == 0)
+		goto alloc_small_pages;
+
+	while (n) {
+		page = __dev_alloc_pages(gfp, s->fl_pg_order);
+		if (unlikely(!page)) {
+			/*
+			 * We've failed inour attempt to allocate a "large
+			 * page".  Fail over to the "small page" allocation
+			 * below.
+			 */
+			fl->large_alloc_failed++;
+			break;
+		}
+		poison_buf(page, PAGE_SIZE << s->fl_pg_order);
+
+		dma_addr = dma_map_page(adapter->pdev_dev, page, 0,
+					PAGE_SIZE << s->fl_pg_order,
+					PCI_DMA_FROMDEVICE);
+		if (unlikely(dma_mapping_error(adapter->pdev_dev, dma_addr))) {
+			/*
+			 * We've run out of DMA mapping space.  Free up the
+			 * buffer and return with what we've managed to put
+			 * into the free list.  We don't want to fail over to
+			 * the small page allocation below in this case
+			 * because DMA mapping resources are typically
+			 * critical resources once they become scarse.
+			 */
+			__free_pages(page, s->fl_pg_order);
+			goto out;
+		}
+		dma_addr |= RX_LARGE_BUF;
+		*d++ = cpu_to_be64(dma_addr);
+
+		set_rx_sw_desc(sdesc, page, dma_addr);
+		sdesc++;
+
+		fl->avail++;
+		if (++fl->pidx == fl->size) {
+			fl->pidx = 0;
+			sdesc = fl->sdesc;
+			d = fl->desc;
+		}
+		n--;
+	}
+
+alloc_small_pages:
+	while (n--) {
+		page = __dev_alloc_page(gfp);
+		if (unlikely(!page)) {
+			fl->alloc_failed++;
+			break;
+		}
+		poison_buf(page, PAGE_SIZE);
+
+		dma_addr = dma_map_page(adapter->pdev_dev, page, 0, PAGE_SIZE,
+				       PCI_DMA_FROMDEVICE);
+		if (unlikely(dma_mapping_error(adapter->pdev_dev, dma_addr))) {
+			put_page(page);
+			break;
+		}
+		*d++ = cpu_to_be64(dma_addr);
+
+		set_rx_sw_desc(sdesc, page, dma_addr);
+		sdesc++;
+
+		fl->avail++;
+		if (++fl->pidx == fl->size) {
+			fl->pidx = 0;
+			sdesc = fl->sdesc;
+			d = fl->desc;
+		}
+	}
+
+out:
+	/*
+	 * Update our accounting state to incorporate the new Free List
+	 * buffers, tell the hardware about them and return the number of
+	 * buffers which we were able to allocate.
+	 */
+	cred = fl->avail - cred;
+	fl->pend_cred += cred;
+	ring_fl_db(adapter, fl);
+
+	if (unlikely(fl_starving(adapter, fl))) {
+		smp_wmb();
+		set_bit(fl->cntxt_id, adapter->sge.starving_fl);
+	}
+
+	return cred;
+}
+
+/*
+ * Refill a Free List to its capacity or the Maximum Refill Increment,
+ * whichever is smaller ...
+ */
+static inline void __refill_fl(struct adapter *adapter, struct sge_fl *fl)
+{
+	refill_fl(adapter, fl,
+		  min((unsigned int)MAX_RX_REFILL, fl_cap(fl) - fl->avail),
+		  GFP_ATOMIC);
+}
+
+/**
+ *	alloc_ring - allocate resources for an SGE descriptor ring
+ *	@dev: the PCI device's core device
+ *	@nelem: the number of descriptors
+ *	@hwsize: the size of each hardware descriptor
+ *	@swsize: the size of each software descriptor
+ *	@busaddrp: the physical PCI bus address of the allocated ring
+ *	@swringp: return address pointer for software ring
+ *	@stat_size: extra space in hardware ring for status information
+ *
+ *	Allocates resources for an SGE descriptor ring, such as TX queues,
+ *	free buffer lists, response queues, etc.  Each SGE ring requires
+ *	space for its hardware descriptors plus, optionally, space for software
+ *	state associated with each hardware entry (the metadata).  The function
+ *	returns three values: the virtual address for the hardware ring (the
+ *	return value of the function), the PCI bus address of the hardware
+ *	ring (in *busaddrp), and the address of the software ring (in swringp).
+ *	Both the hardware and software rings are returned zeroed out.
+ */
+static void *alloc_ring(struct device *dev, size_t nelem, size_t hwsize,
+			size_t swsize, dma_addr_t *busaddrp, void *swringp,
+			size_t stat_size)
+{
+	/*
+	 * Allocate the hardware ring and PCI DMA bus address space for said.
+	 */
+	size_t hwlen = nelem * hwsize + stat_size;
+	void *hwring = dma_alloc_coherent(dev, hwlen, busaddrp, GFP_KERNEL);
+
+	if (!hwring)
+		return NULL;
+
+	/*
+	 * If the caller wants a software ring, allocate it and return a
+	 * pointer to it in *swringp.
+	 */
+	BUG_ON((swsize != 0) != (swringp != NULL));
+	if (swsize) {
+		void *swring = kcalloc(nelem, swsize, GFP_KERNEL);
+
+		if (!swring) {
+			dma_free_coherent(dev, hwlen, hwring, *busaddrp);
+			return NULL;
+		}
+		*(void **)swringp = swring;
+	}
+
+	/*
+	 * Zero out the hardware ring and return its address as our function
+	 * value.
+	 */
+	memset(hwring, 0, hwlen);
+	return hwring;
+}
+
+/**
+ *	sgl_len - calculates the size of an SGL of the given capacity
+ *	@n: the number of SGL entries
+ *
+ *	Calculates the number of flits (8-byte units) needed for a Direct
+ *	Scatter/Gather List that can hold the given number of entries.
+ */
+static inline unsigned int sgl_len(unsigned int n)
+{
+	/*
+	 * A Direct Scatter Gather List uses 32-bit lengths and 64-bit PCI DMA
+	 * addresses.  The DSGL Work Request starts off with a 32-bit DSGL
+	 * ULPTX header, then Length0, then Address0, then, for 1 <= i <= N,
+	 * repeated sequences of { Length[i], Length[i+1], Address[i],
+	 * Address[i+1] } (this ensures that all addresses are on 64-bit
+	 * boundaries).  If N is even, then Length[N+1] should be set to 0 and
+	 * Address[N+1] is omitted.
+	 *
+	 * The following calculation incorporates all of the above.  It's
+	 * somewhat hard to follow but, briefly: the "+2" accounts for the
+	 * first two flits which include the DSGL header, Length0 and
+	 * Address0; the "(3*(n-1))/2" covers the main body of list entries (3
+	 * flits for every pair of the remaining N) +1 if (n-1) is odd; and
+	 * finally the "+((n-1)&1)" adds the one remaining flit needed if
+	 * (n-1) is odd ...
+	 */
+	n--;
+	return (3 * n) / 2 + (n & 1) + 2;
+}
+
+/**
+ *	flits_to_desc - returns the num of TX descriptors for the given flits
+ *	@flits: the number of flits
+ *
+ *	Returns the number of TX descriptors needed for the supplied number
+ *	of flits.
+ */
+static inline unsigned int flits_to_desc(unsigned int flits)
+{
+	BUG_ON(flits > SGE_MAX_WR_LEN / sizeof(__be64));
+	return DIV_ROUND_UP(flits, TXD_PER_EQ_UNIT);
+}
+
+/**
+ *	is_eth_imm - can an Ethernet packet be sent as immediate data?
+ *	@skb: the packet
+ *
+ *	Returns whether an Ethernet packet is small enough to fit completely as
+ *	immediate data.
+ */
+static inline int is_eth_imm(const struct sk_buff *skb)
+{
+	/*
+	 * The VF Driver uses the FW_ETH_TX_PKT_VM_WR firmware Work Request
+	 * which does not accommodate immediate data.  We could dike out all
+	 * of the support code for immediate data but that would tie our hands
+	 * too much if we ever want to enhace the firmware.  It would also
+	 * create more differences between the PF and VF Drivers.
+	 */
+	return false;
+}
+
+/**
+ *	calc_tx_flits - calculate the number of flits for a packet TX WR
+ *	@skb: the packet
+ *
+ *	Returns the number of flits needed for a TX Work Request for the
+ *	given Ethernet packet, including the needed WR and CPL headers.
+ */
+static inline unsigned int calc_tx_flits(const struct sk_buff *skb)
+{
+	unsigned int flits;
+
+	/*
+	 * If the skb is small enough, we can pump it out as a work request
+	 * with only immediate data.  In that case we just have to have the
+	 * TX Packet header plus the skb data in the Work Request.
+	 */
+	if (is_eth_imm(skb))
+		return DIV_ROUND_UP(skb->len + sizeof(struct cpl_tx_pkt),
+				    sizeof(__be64));
+
+	/*
+	 * Otherwise, we're going to have to construct a Scatter gather list
+	 * of the skb body and fragments.  We also include the flits necessary
+	 * for the TX Packet Work Request and CPL.  We always have a firmware
+	 * Write Header (incorporated as part of the cpl_tx_pkt_lso and
+	 * cpl_tx_pkt structures), followed by either a TX Packet Write CPL
+	 * message or, if we're doing a Large Send Offload, an LSO CPL message
+	 * with an embedded TX Packet Write CPL message.
+	 */
+	flits = sgl_len(skb_shinfo(skb)->nr_frags + 1);
+	if (skb_shinfo(skb)->gso_size)
+		flits += (sizeof(struct fw_eth_tx_pkt_vm_wr) +
+			  sizeof(struct cpl_tx_pkt_lso_core) +
+			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
+	else
+		flits += (sizeof(struct fw_eth_tx_pkt_vm_wr) +
+			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
+	return flits;
+}
+
+/**
+ *	write_sgl - populate a Scatter/Gather List for a packet
+ *	@skb: the packet
+ *	@tq: the TX queue we are writing into
+ *	@sgl: starting location for writing the SGL
+ *	@end: points right after the end of the SGL
+ *	@start: start offset into skb main-body data to include in the SGL
+ *	@addr: the list of DMA bus addresses for the SGL elements
+ *
+ *	Generates a Scatter/Gather List for the buffers that make up a packet.
+ *	The caller must provide adequate space for the SGL that will be written.
+ *	The SGL includes all of the packet's page fragments and the data in its
+ *	main body except for the first @start bytes.  @pos must be 16-byte
+ *	aligned and within a TX descriptor with available space.  @end points
+ *	write after the end of the SGL but does not account for any potential
+ *	wrap around, i.e., @end > @tq->stat.
+ */
+static void write_sgl(const struct sk_buff *skb, struct sge_txq *tq,
+		      struct ulptx_sgl *sgl, u64 *end, unsigned int start,
+		      const dma_addr_t *addr)
+{
+	unsigned int i, len;
+	struct ulptx_sge_pair *to;
+	const struct skb_shared_info *si = skb_shinfo(skb);
+	unsigned int nfrags = si->nr_frags;
+	struct ulptx_sge_pair buf[MAX_SKB_FRAGS / 2 + 1];
+
+	len = skb_headlen(skb) - start;
+	if (likely(len)) {
+		sgl->len0 = htonl(len);
+		sgl->addr0 = cpu_to_be64(addr[0] + start);
+		nfrags++;
+	} else {
+		sgl->len0 = htonl(skb_frag_size(&si->frags[0]));
+		sgl->addr0 = cpu_to_be64(addr[1]);
+	}
+
+	sgl->cmd_nsge = htonl(ULPTX_CMD_V(ULP_TX_SC_DSGL) |
+			      ULPTX_NSGE_V(nfrags));
+	if (likely(--nfrags == 0))
+		return;
+	/*
+	 * Most of the complexity below deals with the possibility we hit the
+	 * end of the queue in the middle of writing the SGL.  For this case
+	 * only we create the SGL in a temporary buffer and then copy it.
+	 */
+	to = (u8 *)end > (u8 *)tq->stat ? buf : sgl->sge;
+
+	for (i = (nfrags != si->nr_frags); nfrags >= 2; nfrags -= 2, to++) {
+		to->len[0] = cpu_to_be32(skb_frag_size(&si->frags[i]));
+		to->len[1] = cpu_to_be32(skb_frag_size(&si->frags[++i]));
+		to->addr[0] = cpu_to_be64(addr[i]);
+		to->addr[1] = cpu_to_be64(addr[++i]);
+	}
+	if (nfrags) {
+		to->len[0] = cpu_to_be32(skb_frag_size(&si->frags[i]));
+		to->len[1] = cpu_to_be32(0);
+		to->addr[0] = cpu_to_be64(addr[i + 1]);
+	}
+	if (unlikely((u8 *)end > (u8 *)tq->stat)) {
+		unsigned int part0 = (u8 *)tq->stat - (u8 *)sgl->sge, part1;
+
+		if (likely(part0))
+			memcpy(sgl->sge, buf, part0);
+		part1 = (u8 *)end - (u8 *)tq->stat;
+		memcpy(tq->desc, (u8 *)buf + part0, part1);
+		end = (void *)tq->desc + part1;
+	}
+	if ((uintptr_t)end & 8)           /* 0-pad to multiple of 16 */
+		*end = 0;
+}
+
+/**
+ *	check_ring_tx_db - check and potentially ring a TX queue's doorbell
+ *	@adapter: the adapter
+ *	@tq: the TX queue
+ *	@n: number of new descriptors to give to HW
+ *
+ *	Ring the doorbel for a TX queue.
+ */
+static inline void ring_tx_db(struct adapter *adapter, struct sge_txq *tq,
+			      int n)
+{
+	/* Make sure that all writes to the TX Descriptors are committed
+	 * before we tell the hardware about them.
+	 */
+	wmb();
+
+	/* If we don't have access to the new User Doorbell (T5+), use the old
+	 * doorbell mechanism; otherwise use the new BAR2 mechanism.
+	 */
+	if (unlikely(tq->bar2_addr == NULL)) {
+		u32 val = PIDX_V(n);
+
+		t4_write_reg(adapter, T4VF_SGE_BASE_ADDR + SGE_VF_KDOORBELL,
+			     QID_V(tq->cntxt_id) | val);
+	} else {
+		u32 val = PIDX_T5_V(n);
+
+		/* T4 and later chips share the same PIDX field offset within
+		 * the doorbell, but T5 and later shrank the field in order to
+		 * gain a bit for Doorbell Priority.  The field was absurdly
+		 * large in the first place (14 bits) so we just use the T5
+		 * and later limits and warn if a Queue ID is too large.
+		 */
+		WARN_ON(val & DBPRIO_F);
+
+		/* If we're only writing a single Egress Unit and the BAR2
+		 * Queue ID is 0, we can use the Write Combining Doorbell
+		 * Gather Buffer; otherwise we use the simple doorbell.
+		 */
+		if (n == 1 && tq->bar2_qid == 0) {
+			unsigned int index = (tq->pidx
+					      ? (tq->pidx - 1)
+					      : (tq->size - 1));
+			__be64 *src = (__be64 *)&tq->desc[index];
+			__be64 __iomem *dst = (__be64 __iomem *)(tq->bar2_addr +
+							 SGE_UDB_WCDOORBELL);
+			unsigned int count = EQ_UNIT / sizeof(__be64);
+
+			/* Copy the TX Descriptor in a tight loop in order to
+			 * try to get it to the adapter in a single Write
+			 * Combined transfer on the PCI-E Bus.  If the Write
+			 * Combine fails (say because of an interrupt, etc.)
+			 * the hardware will simply take the last write as a
+			 * simple doorbell write with a PIDX Increment of 1
+			 * and will fetch the TX Descriptor from memory via
+			 * DMA.
+			 */
+			while (count) {
+				/* the (__force u64) is because the compiler
+				 * doesn't understand the endian swizzling
+				 * going on
+				 */
+				writeq((__force u64)*src, dst);
+				src++;
+				dst++;
+				count--;
+			}
+		} else
+			writel(val | QID_V(tq->bar2_qid),
+			       tq->bar2_addr + SGE_UDB_KDOORBELL);
+
+		/* This Write Memory Barrier will force the write to the User
+		 * Doorbell area to be flushed.  This is needed to prevent
+		 * writes on different CPUs for the same queue from hitting
+		 * the adapter out of order.  This is required when some Work
+		 * Requests take the Write Combine Gather Buffer path (user
+		 * doorbell area offset [SGE_UDB_WCDOORBELL..+63]) and some
+		 * take the traditional path where we simply increment the
+		 * PIDX (User Doorbell area SGE_UDB_KDOORBELL) and have the
+		 * hardware DMA read the actual Work Request.
+		 */
+		wmb();
+	}
+}
+
+/**
+ *	inline_tx_skb - inline a packet's data into TX descriptors
+ *	@skb: the packet
+ *	@tq: the TX queue where the packet will be inlined
+ *	@pos: starting position in the TX queue to inline the packet
+ *
+ *	Inline a packet's contents directly into TX descriptors, starting at
+ *	the given position within the TX DMA ring.
+ *	Most of the complexity of this operation is dealing with wrap arounds
+ *	in the middle of the packet we want to inline.
+ */
+static void inline_tx_skb(const struct sk_buff *skb, const struct sge_txq *tq,
+			  void *pos)
+{
+	u64 *p;
+	int left = (void *)tq->stat - pos;
+
+	if (likely(skb->len <= left)) {
+		if (likely(!skb->data_len))
+			skb_copy_from_linear_data(skb, pos, skb->len);
+		else
+			skb_copy_bits(skb, 0, pos, skb->len);
+		pos += skb->len;
+	} else {
+		skb_copy_bits(skb, 0, pos, left);
+		skb_copy_bits(skb, left, tq->desc, skb->len - left);
+		pos = (void *)tq->desc + (skb->len - left);
+	}
+
+	/* 0-pad to multiple of 16 */
+	p = PTR_ALIGN(pos, 8);
+	if ((uintptr_t)p & 8)
+		*p = 0;
+}
+
+/*
+ * Figure out what HW csum a packet wants and return the appropriate control
+ * bits.
+ */
+static u64 hwcsum(const struct sk_buff *skb)
+{
+	int csum_type;
+	const struct iphdr *iph = ip_hdr(skb);
+
+	if (iph->version == 4) {
+		if (iph->protocol == IPPROTO_TCP)
+			csum_type = TX_CSUM_TCPIP;
+		else if (iph->protocol == IPPROTO_UDP)
+			csum_type = TX_CSUM_UDPIP;
+		else {
+nocsum:
+			/*
+			 * unknown protocol, disable HW csum
+			 * and hope a bad packet is detected
+			 */
+			return TXPKT_L4CSUM_DIS;
+		}
+	} else {
+		/*
+		 * this doesn't work with extension headers
+		 */
+		const struct ipv6hdr *ip6h = (const struct ipv6hdr *)iph;
+
+		if (ip6h->nexthdr == IPPROTO_TCP)
+			csum_type = TX_CSUM_TCPIP6;
+		else if (ip6h->nexthdr == IPPROTO_UDP)
+			csum_type = TX_CSUM_UDPIP6;
+		else
+			goto nocsum;
+	}
+
+	if (likely(csum_type >= TX_CSUM_TCPIP))
+		return TXPKT_CSUM_TYPE(csum_type) |
+			TXPKT_IPHDR_LEN(skb_network_header_len(skb)) |
+			TXPKT_ETHHDR_LEN(skb_network_offset(skb) - ETH_HLEN);
+	else {
+		int start = skb_transport_offset(skb);
+
+		return TXPKT_CSUM_TYPE(csum_type) |
+			TXPKT_CSUM_START(start) |
+			TXPKT_CSUM_LOC(start + skb->csum_offset);
+	}
+}
+
+/*
+ * Stop an Ethernet TX queue and record that state change.
+ */
+static void txq_stop(struct sge_eth_txq *txq)
+{
+	netif_tx_stop_queue(txq->txq);
+	txq->q.stops++;
+}
+
+/*
+ * Advance our software state for a TX queue by adding n in use descriptors.
+ */
+static inline void txq_advance(struct sge_txq *tq, unsigned int n)
+{
+	tq->in_use += n;
+	tq->pidx += n;
+	if (tq->pidx >= tq->size)
+		tq->pidx -= tq->size;
+}
+
+/**
+ *	t4vf_eth_xmit - add a packet to an Ethernet TX queue
+ *	@skb: the packet
+ *	@dev: the egress net device
+ *
+ *	Add a packet to an SGE Ethernet TX queue.  Runs with softirqs disabled.
+ */
+int t4vf_eth_xmit(struct sk_buff *skb, struct net_device *dev)
+{
+	u32 wr_mid;
+	u64 cntrl, *end;
+	int qidx, credits;
+	unsigned int flits, ndesc;
+	struct adapter *adapter;
+	struct sge_eth_txq *txq;
+	const struct port_info *pi;
+	struct fw_eth_tx_pkt_vm_wr *wr;
+	struct cpl_tx_pkt_core *cpl;
+	const struct skb_shared_info *ssi;
+	dma_addr_t addr[MAX_SKB_FRAGS + 1];
+	const size_t fw_hdr_copy_len = (sizeof(wr->ethmacdst) +
+					sizeof(wr->ethmacsrc) +
+					sizeof(wr->ethtype) +
+					sizeof(wr->vlantci));
+
+	/*
+	 * The chip minimum packet length is 10 octets but the firmware
+	 * command that we are using requires that we copy the Ethernet header
+	 * (including the VLAN tag) into the header so we reject anything
+	 * smaller than that ...
+	 */
+	if (unlikely(skb->len < fw_hdr_copy_len))
+		goto out_free;
+
+	/*
+	 * Figure out which TX Queue we're going to use.
+	 */
+	pi = netdev_priv(dev);
+	adapter = pi->adapter;
+	qidx = skb_get_queue_mapping(skb);
+	BUG_ON(qidx >= pi->nqsets);
+	txq = &adapter->sge.ethtxq[pi->first_qset + qidx];
+
+	/*
+	 * Take this opportunity to reclaim any TX Descriptors whose DMA
+	 * transfers have completed.
+	 */
+	reclaim_completed_tx(adapter, &txq->q, true);
+
+	/*
+	 * Calculate the number of flits and TX Descriptors we're going to
+	 * need along with how many TX Descriptors will be left over after
+	 * we inject our Work Request.
+	 */
+	flits = calc_tx_flits(skb);
+	ndesc = flits_to_desc(flits);
+	credits = txq_avail(&txq->q) - ndesc;
+
+	if (unlikely(credits < 0)) {
+		/*
+		 * Not enough room for this packet's Work Request.  Stop the
+		 * TX Queue and return a "busy" condition.  The queue will get
+		 * started later on when the firmware informs us that space
+		 * has opened up.
+		 */
+		txq_stop(txq);
+		dev_err(adapter->pdev_dev,
+			"%s: TX ring %u full while queue awake!\n",
+			dev->name, qidx);
+		return NETDEV_TX_BUSY;
+	}
+
+	if (!is_eth_imm(skb) &&
+	    unlikely(map_skb(adapter->pdev_dev, skb, addr) < 0)) {
+		/*
+		 * We need to map the skb into PCI DMA space (because it can't
+		 * be in-lined directly into the Work Request) and the mapping
+		 * operation failed.  Record the error and drop the packet.
+		 */
+		txq->mapping_err++;
+		goto out_free;
+	}
+
+	wr_mid = FW_WR_LEN16_V(DIV_ROUND_UP(flits, 2));
+	if (unlikely(credits < ETHTXQ_STOP_THRES)) {
+		/*
+		 * After we're done injecting the Work Request for this
+		 * packet, we'll be below our "stop threshold" so stop the TX
+		 * Queue now and schedule a request for an SGE Egress Queue
+		 * Update message.  The queue will get started later on when
+		 * the firmware processes this Work Request and sends us an
+		 * Egress Queue Status Update message indicating that space
+		 * has opened up.
+		 */
+		txq_stop(txq);
+		wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
+	}
+
+	/*
+	 * Start filling in our Work Request.  Note that we do _not_ handle
+	 * the WR Header wrapping around the TX Descriptor Ring.  If our
+	 * maximum header size ever exceeds one TX Descriptor, we'll need to
+	 * do something else here.
+	 */
+	BUG_ON(DIV_ROUND_UP(ETHTXQ_MAX_HDR, TXD_PER_EQ_UNIT) > 1);
+	wr = (void *)&txq->q.desc[txq->q.pidx];
+	wr->equiq_to_len16 = cpu_to_be32(wr_mid);
+	wr->r3[0] = cpu_to_be32(0);
+	wr->r3[1] = cpu_to_be32(0);
+	skb_copy_from_linear_data(skb, (void *)wr->ethmacdst, fw_hdr_copy_len);
+	end = (u64 *)wr + flits;
+
+	/*
+	 * If this is a Large Send Offload packet we'll put in an LSO CPL
+	 * message with an encapsulated TX Packet CPL message.  Otherwise we
+	 * just use a TX Packet CPL message.
+	 */
+	ssi = skb_shinfo(skb);
+	if (ssi->gso_size) {
+		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
+		bool v6 = (ssi->gso_type & SKB_GSO_TCPV6) != 0;
+		int l3hdr_len = skb_network_header_len(skb);
+		int eth_xtra_len = skb_network_offset(skb) - ETH_HLEN;
+
+		wr->op_immdlen =
+			cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_PKT_VM_WR) |
+				    FW_WR_IMMDLEN_V(sizeof(*lso) +
+						    sizeof(*cpl)));
+		/*
+		 * Fill in the LSO CPL message.
+		 */
+		lso->lso_ctrl =
+			cpu_to_be32(LSO_OPCODE(CPL_TX_PKT_LSO) |
+				    LSO_FIRST_SLICE |
+				    LSO_LAST_SLICE |
+				    LSO_IPV6(v6) |
+				    LSO_ETHHDR_LEN(eth_xtra_len/4) |
+				    LSO_IPHDR_LEN(l3hdr_len/4) |
+				    LSO_TCPHDR_LEN(tcp_hdr(skb)->doff));
+		lso->ipid_ofst = cpu_to_be16(0);
+		lso->mss = cpu_to_be16(ssi->gso_size);
+		lso->seqno_offset = cpu_to_be32(0);
+		if (is_t4(adapter->params.chip))
+			lso->len = cpu_to_be32(skb->len);
+		else
+			lso->len = cpu_to_be32(LSO_T5_XFER_SIZE(skb->len));
+
+		/*
+		 * Set up TX Packet CPL pointer, control word and perform
+		 * accounting.
+		 */
+		cpl = (void *)(lso + 1);
+		cntrl = (TXPKT_CSUM_TYPE(v6 ? TX_CSUM_TCPIP6 : TX_CSUM_TCPIP) |
+			 TXPKT_IPHDR_LEN(l3hdr_len) |
+			 TXPKT_ETHHDR_LEN(eth_xtra_len));
+		txq->tso++;
+		txq->tx_cso += ssi->gso_segs;
+	} else {
+		int len;
+
+		len = is_eth_imm(skb) ? skb->len + sizeof(*cpl) : sizeof(*cpl);
+		wr->op_immdlen =
+			cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_PKT_VM_WR) |
+				    FW_WR_IMMDLEN_V(len));
+
+		/*
+		 * Set up TX Packet CPL pointer, control word and perform
+		 * accounting.
+		 */
+		cpl = (void *)(wr + 1);
+		if (skb->ip_summed == CHECKSUM_PARTIAL) {
+			cntrl = hwcsum(skb) | TXPKT_IPCSUM_DIS;
+			txq->tx_cso++;
+		} else
+			cntrl = TXPKT_L4CSUM_DIS | TXPKT_IPCSUM_DIS;
+	}
+
+	/*
+	 * If there's a VLAN tag present, add that to the list of things to
+	 * do in this Work Request.
+	 */
+	if (skb_vlan_tag_present(skb)) {
+		txq->vlan_ins++;
+		cntrl |= TXPKT_VLAN_VLD | TXPKT_VLAN(skb_vlan_tag_get(skb));
+	}
+
+	/*
+	 * Fill in the TX Packet CPL message header.
+	 */
+	cpl->ctrl0 = cpu_to_be32(TXPKT_OPCODE(CPL_TX_PKT_XT) |
+				 TXPKT_INTF(pi->port_id) |
+				 TXPKT_PF(0));
+	cpl->pack = cpu_to_be16(0);
+	cpl->len = cpu_to_be16(skb->len);
+	cpl->ctrl1 = cpu_to_be64(cntrl);
+
+#ifdef T4_TRACE
+	T4_TRACE5(adapter->tb[txq->q.cntxt_id & 7],
+		  "eth_xmit: ndesc %u, credits %u, pidx %u, len %u, frags %u",
+		  ndesc, credits, txq->q.pidx, skb->len, ssi->nr_frags);
+#endif
+
+	/*
+	 * Fill in the body of the TX Packet CPL message with either in-lined
+	 * data or a Scatter/Gather List.
+	 */
+	if (is_eth_imm(skb)) {
+		/*
+		 * In-line the packet's data and free the skb since we don't
+		 * need it any longer.
+		 */
+		inline_tx_skb(skb, &txq->q, cpl + 1);
+		dev_consume_skb_any(skb);
+	} else {
+		/*
+		 * Write the skb's Scatter/Gather list into the TX Packet CPL
+		 * message and retain a pointer to the skb so we can free it
+		 * later when its DMA completes.  (We store the skb pointer
+		 * in the Software Descriptor corresponding to the last TX
+		 * Descriptor used by the Work Request.)
+		 *
+		 * The retained skb will be freed when the corresponding TX
+		 * Descriptors are reclaimed after their DMAs complete.
+		 * However, this could take quite a while since, in general,
+		 * the hardware is set up to be lazy about sending DMA
+		 * completion notifications to us and we mostly perform TX
+		 * reclaims in the transmit routine.
+		 *
+		 * This is good for performamce but means that we rely on new
+		 * TX packets arriving to run the destructors of completed
+		 * packets, which open up space in their sockets' send queues.
+		 * Sometimes we do not get such new packets causing TX to
+		 * stall.  A single UDP transmitter is a good example of this
+		 * situation.  We have a clean up timer that periodically
+		 * reclaims completed packets but it doesn't run often enough
+		 * (nor do we want it to) to prevent lengthy stalls.  A
+		 * solution to this problem is to run the destructor early,
+		 * after the packet is queued but before it's DMAd.  A con is
+		 * that we lie to socket memory accounting, but the amount of
+		 * extra memory is reasonable (limited by the number of TX
+		 * descriptors), the packets do actually get freed quickly by
+		 * new packets almost always, and for protocols like TCP that
+		 * wait for acks to really free up the data the extra memory
+		 * is even less.  On the positive side we run the destructors
+		 * on the sending CPU rather than on a potentially different
+		 * completing CPU, usually a good thing.
+		 *
+		 * Run the destructor before telling the DMA engine about the
+		 * packet to make sure it doesn't complete and get freed
+		 * prematurely.
+		 */
+		struct ulptx_sgl *sgl = (struct ulptx_sgl *)(cpl + 1);
+		struct sge_txq *tq = &txq->q;
+		int last_desc;
+
+		/*
+		 * If the Work Request header was an exact multiple of our TX
+		 * Descriptor length, then it's possible that the starting SGL
+		 * pointer lines up exactly with the end of our TX Descriptor
+		 * ring.  If that's the case, wrap around to the beginning
+		 * here ...
+		 */
+		if (unlikely((void *)sgl == (void *)tq->stat)) {
+			sgl = (void *)tq->desc;
+			end = ((void *)tq->desc + ((void *)end - (void *)tq->stat));
+		}
+
+		write_sgl(skb, tq, sgl, end, 0, addr);
+		skb_orphan(skb);
+
+		last_desc = tq->pidx + ndesc - 1;
+		if (last_desc >= tq->size)
+			last_desc -= tq->size;
+		tq->sdesc[last_desc].skb = skb;
+		tq->sdesc[last_desc].sgl = sgl;
+	}
+
+	/*
+	 * Advance our internal TX Queue state, tell the hardware about
+	 * the new TX descriptors and return success.
+	 */
+	txq_advance(&txq->q, ndesc);
+	dev->trans_start = jiffies;
+	ring_tx_db(adapter, &txq->q, ndesc);
+	return NETDEV_TX_OK;
+
+out_free:
+	/*
+	 * An error of some sort happened.  Free the TX skb and tell the
+	 * OS that we've "dealt" with the packet ...
+	 */
+	dev_kfree_skb_any(skb);
+	return NETDEV_TX_OK;
+}
+
+/**
+ *	copy_frags - copy fragments from gather list into skb_shared_info
+ *	@skb: destination skb
+ *	@gl: source internal packet gather list
+ *	@offset: packet start offset in first page
+ *
+ *	Copy an internal packet gather list into a Linux skb_shared_info
+ *	structure.
+ */
+static inline void copy_frags(struct sk_buff *skb,
+			      const struct pkt_gl *gl,
+			      unsigned int offset)
+{
+	int i;
+
+	/* usually there's just one frag */
+	__skb_fill_page_desc(skb, 0, gl->frags[0].page,
+			     gl->frags[0].offset + offset,
+			     gl->frags[0].size - offset);
+	skb_shinfo(skb)->nr_frags = gl->nfrags;
+	for (i = 1; i < gl->nfrags; i++)
+		__skb_fill_page_desc(skb, i, gl->frags[i].page,
+				     gl->frags[i].offset,
+				     gl->frags[i].size);
+
+	/* get a reference to the last page, we don't own it */
+	get_page(gl->frags[gl->nfrags - 1].page);
+}
+
+/**
+ *	t4vf_pktgl_to_skb - build an sk_buff from a packet gather list
+ *	@gl: the gather list
+ *	@skb_len: size of sk_buff main body if it carries fragments
+ *	@pull_len: amount of data to move to the sk_buff's main body
+ *
+ *	Builds an sk_buff from the given packet gather list.  Returns the
+ *	sk_buff or %NULL if sk_buff allocation failed.
+ */
+static struct sk_buff *t4vf_pktgl_to_skb(const struct pkt_gl *gl,
+					 unsigned int skb_len,
+					 unsigned int pull_len)
+{
+	struct sk_buff *skb;
+
+	/*
+	 * If the ingress packet is small enough, allocate an skb large enough
+	 * for all of the data and copy it inline.  Otherwise, allocate an skb
+	 * with enough room to pull in the header and reference the rest of
+	 * the data via the skb fragment list.
+	 *
+	 * Below we rely on RX_COPY_THRES being less than the smallest Rx
+	 * buff!  size, which is expected since buffers are at least
+	 * PAGE_SIZEd.  In this case packets up to RX_COPY_THRES have only one
+	 * fragment.
+	 */
+	if (gl->tot_len <= RX_COPY_THRES) {
+		/* small packets have only one fragment */
+		skb = alloc_skb(gl->tot_len, GFP_ATOMIC);
+		if (unlikely(!skb))
+			goto out;
+		__skb_put(skb, gl->tot_len);
+		skb_copy_to_linear_data(skb, gl->va, gl->tot_len);
+	} else {
+		skb = alloc_skb(skb_len, GFP_ATOMIC);
+		if (unlikely(!skb))
+			goto out;
+		__skb_put(skb, pull_len);
+		skb_copy_to_linear_data(skb, gl->va, pull_len);
+
+		copy_frags(skb, gl, pull_len);
+		skb->len = gl->tot_len;
+		skb->data_len = skb->len - pull_len;
+		skb->truesize += skb->data_len;
+	}
+
+out:
+	return skb;
+}
+
+/**
+ *	t4vf_pktgl_free - free a packet gather list
+ *	@gl: the gather list
+ *
+ *	Releases the pages of a packet gather list.  We do not own the last
+ *	page on the list and do not free it.
+ */
+static void t4vf_pktgl_free(const struct pkt_gl *gl)
+{
+	int frag;
+
+	frag = gl->nfrags - 1;
+	while (frag--)
+		put_page(gl->frags[frag].page);
+}
+
+/**
+ *	do_gro - perform Generic Receive Offload ingress packet processing
+ *	@rxq: ingress RX Ethernet Queue
+ *	@gl: gather list for ingress packet
+ *	@pkt: CPL header for last packet fragment
+ *
+ *	Perform Generic Receive Offload (GRO) ingress packet processing.
+ *	We use the standard Linux GRO interfaces for this.
+ */
+static void do_gro(struct sge_eth_rxq *rxq, const struct pkt_gl *gl,
+		   const struct cpl_rx_pkt *pkt)
+{
+	struct adapter *adapter = rxq->rspq.adapter;
+	struct sge *s = &adapter->sge;
+	int ret;
+	struct sk_buff *skb;
+
+	skb = napi_get_frags(&rxq->rspq.napi);
+	if (unlikely(!skb)) {
+		t4vf_pktgl_free(gl);
+		rxq->stats.rx_drops++;
+		return;
+	}
+
+	copy_frags(skb, gl, s->pktshift);
+	skb->len = gl->tot_len - s->pktshift;
+	skb->data_len = skb->len;
+	skb->truesize += skb->data_len;
+	skb->ip_summed = CHECKSUM_UNNECESSARY;
+	skb_record_rx_queue(skb, rxq->rspq.idx);
+
+	if (pkt->vlan_ex) {
+		__vlan_hwaccel_put_tag(skb, cpu_to_be16(ETH_P_8021Q),
+					be16_to_cpu(pkt->vlan));
+		rxq->stats.vlan_ex++;
+	}
+	ret = napi_gro_frags(&rxq->rspq.napi);
+
+	if (ret == GRO_HELD)
+		rxq->stats.lro_pkts++;
+	else if (ret == GRO_MERGED || ret == GRO_MERGED_FREE)
+		rxq->stats.lro_merged++;
+	rxq->stats.pkts++;
+	rxq->stats.rx_cso++;
+}
+
+/**
+ *	t4vf_ethrx_handler - process an ingress ethernet packet
+ *	@rspq: the response queue that received the packet
+ *	@rsp: the response queue descriptor holding the RX_PKT message
+ *	@gl: the gather list of packet fragments
+ *
+ *	Process an ingress ethernet packet and deliver it to the stack.
+ */
+int t4vf_ethrx_handler(struct sge_rspq *rspq, const __be64 *rsp,
+		       const struct pkt_gl *gl)
+{
+	struct sk_buff *skb;
+	const struct cpl_rx_pkt *pkt = (void *)rsp;
+	bool csum_ok = pkt->csum_calc && !pkt->err_vec &&
+		       (rspq->netdev->features & NETIF_F_RXCSUM);
+	struct sge_eth_rxq *rxq = container_of(rspq, struct sge_eth_rxq, rspq);
+	struct adapter *adapter = rspq->adapter;
+	struct sge *s = &adapter->sge;
+
+	/*
+	 * If this is a good TCP packet and we have Generic Receive Offload
+	 * enabled, handle the packet in the GRO path.
+	 */
+	if ((pkt->l2info & cpu_to_be32(RXF_TCP_F)) &&
+	    (rspq->netdev->features & NETIF_F_GRO) && csum_ok &&
+	    !pkt->ip_frag) {
+		do_gro(rxq, gl, pkt);
+		return 0;
+	}
+
+	/*
+	 * Convert the Packet Gather List into an skb.
+	 */
+	skb = t4vf_pktgl_to_skb(gl, RX_SKB_LEN, RX_PULL_LEN);
+	if (unlikely(!skb)) {
+		t4vf_pktgl_free(gl);
+		rxq->stats.rx_drops++;
+		return 0;
+	}
+	__skb_pull(skb, s->pktshift);
+	skb->protocol = eth_type_trans(skb, rspq->netdev);
+	skb_record_rx_queue(skb, rspq->idx);
+	rxq->stats.pkts++;
+
+	if (csum_ok && !pkt->err_vec &&
+	    (be32_to_cpu(pkt->l2info) & (RXF_UDP_F | RXF_TCP_F))) {
+		if (!pkt->ip_frag)
+			skb->ip_summed = CHECKSUM_UNNECESSARY;
+		else {
+			__sum16 c = (__force __sum16)pkt->csum;
+			skb->csum = csum_unfold(c);
+			skb->ip_summed = CHECKSUM_COMPLETE;
+		}
+		rxq->stats.rx_cso++;
+	} else
+		skb_checksum_none_assert(skb);
+
+	if (pkt->vlan_ex) {
+		rxq->stats.vlan_ex++;
+		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), be16_to_cpu(pkt->vlan));
+	}
+
+	netif_receive_skb(skb);
+
+	return 0;
+}
+
+/**
+ *	is_new_response - check if a response is newly written
+ *	@rc: the response control descriptor
+ *	@rspq: the response queue
+ *
+ *	Returns true if a response descriptor contains a yet unprocessed
+ *	response.
+ */
+static inline bool is_new_response(const struct rsp_ctrl *rc,
+				   const struct sge_rspq *rspq)
+{
+	return RSPD_GEN(rc->type_gen) == rspq->gen;
+}
+
+/**
+ *	restore_rx_bufs - put back a packet's RX buffers
+ *	@gl: the packet gather list
+ *	@fl: the SGE Free List
+ *	@nfrags: how many fragments in @si
+ *
+ *	Called when we find out that the current packet, @si, can't be
+ *	processed right away for some reason.  This is a very rare event and
+ *	there's no effort to make this suspension/resumption process
+ *	particularly efficient.
+ *
+ *	We implement the suspension by putting all of the RX buffers associated
+ *	with the current packet back on the original Free List.  The buffers
+ *	have already been unmapped and are left unmapped, we mark them as
+ *	unmapped in order to prevent further unmapping attempts.  (Effectively
+ *	this function undoes the series of @unmap_rx_buf calls which were done
+ *	to create the current packet's gather list.)  This leaves us ready to
+ *	restart processing of the packet the next time we start processing the
+ *	RX Queue ...
+ */
+static void restore_rx_bufs(const struct pkt_gl *gl, struct sge_fl *fl,
+			    int frags)
+{
+	struct rx_sw_desc *sdesc;
+
+	while (frags--) {
+		if (fl->cidx == 0)
+			fl->cidx = fl->size - 1;
+		else
+			fl->cidx--;
+		sdesc = &fl->sdesc[fl->cidx];
+		sdesc->page = gl->frags[frags].page;
+		sdesc->dma_addr |= RX_UNMAPPED_BUF;
+		fl->avail++;
+	}
+}
+
+/**
+ *	rspq_next - advance to the next entry in a response queue
+ *	@rspq: the queue
+ *
+ *	Updates the state of a response queue to advance it to the next entry.
+ */
+static inline void rspq_next(struct sge_rspq *rspq)
+{
+	rspq->cur_desc = (void *)rspq->cur_desc + rspq->iqe_len;
+	if (unlikely(++rspq->cidx == rspq->size)) {
+		rspq->cidx = 0;
+		rspq->gen ^= 1;
+		rspq->cur_desc = rspq->desc;
+	}
+}
+
+/**
+ *	process_responses - process responses from an SGE response queue
+ *	@rspq: the ingress response queue to process
+ *	@budget: how many responses can be processed in this round
+ *
+ *	Process responses from a Scatter Gather Engine response queue up to
+ *	the supplied budget.  Responses include received packets as well as
+ *	control messages from firmware or hardware.
+ *
+ *	Additionally choose the interrupt holdoff time for the next interrupt
+ *	on this queue.  If the system is under memory shortage use a fairly
+ *	long delay to help recovery.
+ */
+static int process_responses(struct sge_rspq *rspq, int budget)
+{
+	struct sge_eth_rxq *rxq = container_of(rspq, struct sge_eth_rxq, rspq);
+	struct adapter *adapter = rspq->adapter;
+	struct sge *s = &adapter->sge;
+	int budget_left = budget;
+
+	while (likely(budget_left)) {
+		int ret, rsp_type;
+		const struct rsp_ctrl *rc;
+
+		rc = (void *)rspq->cur_desc + (rspq->iqe_len - sizeof(*rc));
+		if (!is_new_response(rc, rspq))
+			break;
+
+		/*
+		 * Figure out what kind of response we've received from the
+		 * SGE.
+		 */
+		dma_rmb();
+		rsp_type = RSPD_TYPE(rc->type_gen);
+		if (likely(rsp_type == RSP_TYPE_FLBUF)) {
+			struct page_frag *fp;
+			struct pkt_gl gl;
+			const struct rx_sw_desc *sdesc;
+			u32 bufsz, frag;
+			u32 len = be32_to_cpu(rc->pldbuflen_qid);
+
+			/*
+			 * If we get a "new buffer" message from the SGE we
+			 * need to move on to the next Free List buffer.
+			 */
+			if (len & RSPD_NEWBUF) {
+				/*
+				 * We get one "new buffer" message when we
+				 * first start up a queue so we need to ignore
+				 * it when our offset into the buffer is 0.
+				 */
+				if (likely(rspq->offset > 0)) {
+					free_rx_bufs(rspq->adapter, &rxq->fl,
+						     1);
+					rspq->offset = 0;
+				}
+				len = RSPD_LEN(len);
+			}
+			gl.tot_len = len;
+
+			/*
+			 * Gather packet fragments.
+			 */
+			for (frag = 0, fp = gl.frags; /**/; frag++, fp++) {
+				BUG_ON(frag >= MAX_SKB_FRAGS);
+				BUG_ON(rxq->fl.avail == 0);
+				sdesc = &rxq->fl.sdesc[rxq->fl.cidx];
+				bufsz = get_buf_size(adapter, sdesc);
+				fp->page = sdesc->page;
+				fp->offset = rspq->offset;
+				fp->size = min(bufsz, len);
+				len -= fp->size;
+				if (!len)
+					break;
+				unmap_rx_buf(rspq->adapter, &rxq->fl);
+			}
+			gl.nfrags = frag+1;
+
+			/*
+			 * Last buffer remains mapped so explicitly make it
+			 * coherent for CPU access and start preloading first
+			 * cache line ...
+			 */
+			dma_sync_single_for_cpu(rspq->adapter->pdev_dev,
+						get_buf_addr(sdesc),
+						fp->size, DMA_FROM_DEVICE);
+			gl.va = (page_address(gl.frags[0].page) +
+				 gl.frags[0].offset);
+			prefetch(gl.va);
+
+			/*
+			 * Hand the new ingress packet to the handler for
+			 * this Response Queue.
+			 */
+			ret = rspq->handler(rspq, rspq->cur_desc, &gl);
+			if (likely(ret == 0))
+				rspq->offset += ALIGN(fp->size, s->fl_align);
+			else
+				restore_rx_bufs(&gl, &rxq->fl, frag);
+		} else if (likely(rsp_type == RSP_TYPE_CPL)) {
+			ret = rspq->handler(rspq, rspq->cur_desc, NULL);
+		} else {
+			WARN_ON(rsp_type > RSP_TYPE_CPL);
+			ret = 0;
+		}
+
+		if (unlikely(ret)) {
+			/*
+			 * Couldn't process descriptor, back off for recovery.
+			 * We use the SGE's last timer which has the longest
+			 * interrupt coalescing value ...
+			 */
+			const int NOMEM_TIMER_IDX = SGE_NTIMERS-1;
+			rspq->next_intr_params =
+				QINTR_TIMER_IDX(NOMEM_TIMER_IDX);
+			break;
+		}
+
+		rspq_next(rspq);
+		budget_left--;
+	}
+
+	/*
+	 * If this is a Response Queue with an associated Free List and
+	 * at least two Egress Queue units available in the Free List
+	 * for new buffer pointers, refill the Free List.
+	 */
+	if (rspq->offset >= 0 &&
+	    rxq->fl.size - rxq->fl.avail >= 2*FL_PER_EQ_UNIT)
+		__refill_fl(rspq->adapter, &rxq->fl);
+	return budget - budget_left;
+}
+
+/**
+ *	napi_rx_handler - the NAPI handler for RX processing
+ *	@napi: the napi instance
+ *	@budget: how many packets we can process in this round
+ *
+ *	Handler for new data events when using NAPI.  This does not need any
+ *	locking or protection from interrupts as data interrupts are off at
+ *	this point and other adapter interrupts do not interfere (the latter
+ *	in not a concern at all with MSI-X as non-data interrupts then have
+ *	a separate handler).
+ */
+static int napi_rx_handler(struct napi_struct *napi, int budget)
+{
+	unsigned int intr_params;
+	struct sge_rspq *rspq = container_of(napi, struct sge_rspq, napi);
+	int work_done = process_responses(rspq, budget);
+	u32 val;
+
+	if (likely(work_done < budget)) {
+		napi_complete(napi);
+		intr_params = rspq->next_intr_params;
+		rspq->next_intr_params = rspq->intr_params;
+	} else
+		intr_params = QINTR_TIMER_IDX(SGE_TIMER_UPD_CIDX);
+
+	if (unlikely(work_done == 0))
+		rspq->unhandled_irqs++;
+
+	val = CIDXINC_V(work_done) | SEINTARM_V(intr_params);
+	if (is_t4(rspq->adapter->params.chip)) {
+		t4_write_reg(rspq->adapter,
+			     T4VF_SGE_BASE_ADDR + SGE_VF_GTS,
+			     val | INGRESSQID_V((u32)rspq->cntxt_id));
+	} else {
+		writel(val | INGRESSQID_V(rspq->bar2_qid),
+		       rspq->bar2_addr + SGE_UDB_GTS);
+		wmb();
+	}
+	return work_done;
+}
+
+/*
+ * The MSI-X interrupt handler for an SGE response queue for the NAPI case
+ * (i.e., response queue serviced by NAPI polling).
+ */
+irqreturn_t t4vf_sge_intr_msix(int irq, void *cookie)
+{
+	struct sge_rspq *rspq = cookie;
+
+	napi_schedule(&rspq->napi);
+	return IRQ_HANDLED;
+}
+
+/*
+ * Process the indirect interrupt entries in the interrupt queue and kick off
+ * NAPI for each queue that has generated an entry.
+ */
+static unsigned int process_intrq(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+	struct sge_rspq *intrq = &s->intrq;
+	unsigned int work_done;
+	u32 val;
+
+	spin_lock(&adapter->sge.intrq_lock);
+	for (work_done = 0; ; work_done++) {
+		const struct rsp_ctrl *rc;
+		unsigned int qid, iq_idx;
+		struct sge_rspq *rspq;
+
+		/*
+		 * Grab the next response from the interrupt queue and bail
+		 * out if it's not a new response.
+		 */
+		rc = (void *)intrq->cur_desc + (intrq->iqe_len - sizeof(*rc));
+		if (!is_new_response(rc, intrq))
+			break;
+
+		/*
+		 * If the response isn't a forwarded interrupt message issue a
+		 * error and go on to the next response message.  This should
+		 * never happen ...
+		 */
+		dma_rmb();
+		if (unlikely(RSPD_TYPE(rc->type_gen) != RSP_TYPE_INTR)) {
+			dev_err(adapter->pdev_dev,
+				"Unexpected INTRQ response type %d\n",
+				RSPD_TYPE(rc->type_gen));
+			continue;
+		}
+
+		/*
+		 * Extract the Queue ID from the interrupt message and perform
+		 * sanity checking to make sure it really refers to one of our
+		 * Ingress Queues which is active and matches the queue's ID.
+		 * None of these error conditions should ever happen so we may
+		 * want to either make them fatal and/or conditionalized under
+		 * DEBUG.
+		 */
+		qid = RSPD_QID(be32_to_cpu(rc->pldbuflen_qid));
+		iq_idx = IQ_IDX(s, qid);
+		if (unlikely(iq_idx >= MAX_INGQ)) {
+			dev_err(adapter->pdev_dev,
+				"Ingress QID %d out of range\n", qid);
+			continue;
+		}
+		rspq = s->ingr_map[iq_idx];
+		if (unlikely(rspq == NULL)) {
+			dev_err(adapter->pdev_dev,
+				"Ingress QID %d RSPQ=NULL\n", qid);
+			continue;
+		}
+		if (unlikely(rspq->abs_id != qid)) {
+			dev_err(adapter->pdev_dev,
+				"Ingress QID %d refers to RSPQ %d\n",
+				qid, rspq->abs_id);
+			continue;
+		}
+
+		/*
+		 * Schedule NAPI processing on the indicated Response Queue
+		 * and move on to the next entry in the Forwarded Interrupt
+		 * Queue.
+		 */
+		napi_schedule(&rspq->napi);
+		rspq_next(intrq);
+	}
+
+	val = CIDXINC_V(work_done) | SEINTARM_V(intrq->intr_params);
+	if (is_t4(adapter->params.chip))
+		t4_write_reg(adapter, T4VF_SGE_BASE_ADDR + SGE_VF_GTS,
+			     val | INGRESSQID_V(intrq->cntxt_id));
+	else {
+		writel(val | INGRESSQID_V(intrq->bar2_qid),
+		       intrq->bar2_addr + SGE_UDB_GTS);
+		wmb();
+	}
+
+	spin_unlock(&adapter->sge.intrq_lock);
+
+	return work_done;
+}
+
+/*
+ * The MSI interrupt handler handles data events from SGE response queues as
+ * well as error and other async events as they all use the same MSI vector.
+ */
+static irqreturn_t t4vf_intr_msi(int irq, void *cookie)
+{
+	struct adapter *adapter = cookie;
+
+	process_intrq(adapter);
+	return IRQ_HANDLED;
+}
+
+/**
+ *	t4vf_intr_handler - select the top-level interrupt handler
+ *	@adapter: the adapter
+ *
+ *	Selects the top-level interrupt handler based on the type of interrupts
+ *	(MSI-X or MSI).
+ */
+irq_handler_t t4vf_intr_handler(struct adapter *adapter)
+{
+	BUG_ON((adapter->flags & (USING_MSIX|USING_MSI)) == 0);
+	if (adapter->flags & USING_MSIX)
+		return t4vf_sge_intr_msix;
+	else
+		return t4vf_intr_msi;
+}
+
+/**
+ *	sge_rx_timer_cb - perform periodic maintenance of SGE RX queues
+ *	@data: the adapter
+ *
+ *	Runs periodically from a timer to perform maintenance of SGE RX queues.
+ *
+ *	a) Replenishes RX queues that have run out due to memory shortage.
+ *	Normally new RX buffers are added when existing ones are consumed but
+ *	when out of memory a queue can become empty.  We schedule NAPI to do
+ *	the actual refill.
+ */
+static void sge_rx_timer_cb(unsigned long data)
+{
+	struct adapter *adapter = (struct adapter *)data;
+	struct sge *s = &adapter->sge;
+	unsigned int i;
+
+	/*
+	 * Scan the "Starving Free Lists" flag array looking for any Free
+	 * Lists in need of more free buffers.  If we find one and it's not
+	 * being actively polled, then bump its "starving" counter and attempt
+	 * to refill it.  If we're successful in adding enough buffers to push
+	 * the Free List over the starving threshold, then we can clear its
+	 * "starving" status.
+	 */
+	for (i = 0; i < ARRAY_SIZE(s->starving_fl); i++) {
+		unsigned long m;
+
+		for (m = s->starving_fl[i]; m; m &= m - 1) {
+			unsigned int id = __ffs(m) + i * BITS_PER_LONG;
+			struct sge_fl *fl = s->egr_map[id];
+
+			clear_bit(id, s->starving_fl);
+			smp_mb__after_atomic();
+
+			/*
+			 * Since we are accessing fl without a lock there's a
+			 * small probability of a false positive where we
+			 * schedule napi but the FL is no longer starving.
+			 * No biggie.
+			 */
+			if (fl_starving(adapter, fl)) {
+				struct sge_eth_rxq *rxq;
+
+				rxq = container_of(fl, struct sge_eth_rxq, fl);
+				if (napi_reschedule(&rxq->rspq.napi))
+					fl->starving++;
+				else
+					set_bit(id, s->starving_fl);
+			}
+		}
+	}
+
+	/*
+	 * Reschedule the next scan for starving Free Lists ...
+	 */
+	mod_timer(&s->rx_timer, jiffies + RX_QCHECK_PERIOD);
+}
+
+/**
+ *	sge_tx_timer_cb - perform periodic maintenance of SGE Tx queues
+ *	@data: the adapter
+ *
+ *	Runs periodically from a timer to perform maintenance of SGE TX queues.
+ *
+ *	b) Reclaims completed Tx packets for the Ethernet queues.  Normally
+ *	packets are cleaned up by new Tx packets, this timer cleans up packets
+ *	when no new packets are being submitted.  This is essential for pktgen,
+ *	at least.
+ */
+static void sge_tx_timer_cb(unsigned long data)
+{
+	struct adapter *adapter = (struct adapter *)data;
+	struct sge *s = &adapter->sge;
+	unsigned int i, budget;
+
+	budget = MAX_TIMER_TX_RECLAIM;
+	i = s->ethtxq_rover;
+	do {
+		struct sge_eth_txq *txq = &s->ethtxq[i];
+
+		if (reclaimable(&txq->q) && __netif_tx_trylock(txq->txq)) {
+			int avail = reclaimable(&txq->q);
+
+			if (avail > budget)
+				avail = budget;
+
+			free_tx_desc(adapter, &txq->q, avail, true);
+			txq->q.in_use -= avail;
+			__netif_tx_unlock(txq->txq);
+
+			budget -= avail;
+			if (!budget)
+				break;
+		}
+
+		i++;
+		if (i >= s->ethqsets)
+			i = 0;
+	} while (i != s->ethtxq_rover);
+	s->ethtxq_rover = i;
+
+	/*
+	 * If we found too many reclaimable packets schedule a timer in the
+	 * near future to continue where we left off.  Otherwise the next timer
+	 * will be at its normal interval.
+	 */
+	mod_timer(&s->tx_timer, jiffies + (budget ? TX_QCHECK_PERIOD : 2));
+}
+
+/**
+ *	bar2_address - return the BAR2 address for an SGE Queue's Registers
+ *	@adapter: the adapter
+ *	@qid: the SGE Queue ID
+ *	@qtype: the SGE Queue Type (Egress or Ingress)
+ *	@pbar2_qid: BAR2 Queue ID or 0 for Queue ID inferred SGE Queues
+ *
+ *	Returns the BAR2 address for the SGE Queue Registers associated with
+ *	@qid.  If BAR2 SGE Registers aren't available, returns NULL.  Also
+ *	returns the BAR2 Queue ID to be used with writes to the BAR2 SGE
+ *	Queue Registers.  If the BAR2 Queue ID is 0, then "Inferred Queue ID"
+ *	Registers are supported (e.g. the Write Combining Doorbell Buffer).
+ */
+static void __iomem *bar2_address(struct adapter *adapter,
+				  unsigned int qid,
+				  enum t4_bar2_qtype qtype,
+				  unsigned int *pbar2_qid)
+{
+	u64 bar2_qoffset;
+	int ret;
+
+	ret = t4_bar2_sge_qregs(adapter, qid, qtype,
+				&bar2_qoffset, pbar2_qid);
+	if (ret)
+		return NULL;
+
+	return adapter->bar2 + bar2_qoffset;
+}
+
+/**
+ *	t4vf_sge_alloc_rxq - allocate an SGE RX Queue
+ *	@adapter: the adapter
+ *	@rspq: pointer to to the new rxq's Response Queue to be filled in
+ *	@iqasynch: if 0, a normal rspq; if 1, an asynchronous event queue
+ *	@dev: the network device associated with the new rspq
+ *	@intr_dest: MSI-X vector index (overriden in MSI mode)
+ *	@fl: pointer to the new rxq's Free List to be filled in
+ *	@hnd: the interrupt handler to invoke for the rspq
+ */
+int t4vf_sge_alloc_rxq(struct adapter *adapter, struct sge_rspq *rspq,
+		       bool iqasynch, struct net_device *dev,
+		       int intr_dest,
+		       struct sge_fl *fl, rspq_handler_t hnd)
+{
+	struct sge *s = &adapter->sge;
+	struct port_info *pi = netdev_priv(dev);
+	struct fw_iq_cmd cmd, rpl;
+	int ret, iqandst, flsz = 0;
+
+	/*
+	 * If we're using MSI interrupts and we're not initializing the
+	 * Forwarded Interrupt Queue itself, then set up this queue for
+	 * indirect interrupts to the Forwarded Interrupt Queue.  Obviously
+	 * the Forwarded Interrupt Queue must be set up before any other
+	 * ingress queue ...
+	 */
+	if ((adapter->flags & USING_MSI) && rspq != &adapter->sge.intrq) {
+		iqandst = SGE_INTRDST_IQ;
+		intr_dest = adapter->sge.intrq.abs_id;
+	} else
+		iqandst = SGE_INTRDST_PCI;
+
+	/*
+	 * Allocate the hardware ring for the Response Queue.  The size needs
+	 * to be a multiple of 16 which includes the mandatory status entry
+	 * (regardless of whether the Status Page capabilities are enabled or
+	 * not).
+	 */
+	rspq->size = roundup(rspq->size, 16);
+	rspq->desc = alloc_ring(adapter->pdev_dev, rspq->size, rspq->iqe_len,
+				0, &rspq->phys_addr, NULL, 0);
+	if (!rspq->desc)
+		return -ENOMEM;
+
+	/*
+	 * Fill in the Ingress Queue Command.  Note: Ideally this code would
+	 * be in t4vf_hw.c but there are so many parameters and dependencies
+	 * on our Linux SGE state that we would end up having to pass tons of
+	 * parameters.  We'll have to think about how this might be migrated
+	 * into OS-independent common code ...
+	 */
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_IQ_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_WRITE_F |
+				    FW_CMD_EXEC_F);
+	cmd.alloc_to_len16 = cpu_to_be32(FW_IQ_CMD_ALLOC_F |
+					 FW_IQ_CMD_IQSTART_F |
+					 FW_LEN16(cmd));
+	cmd.type_to_iqandstindex =
+		cpu_to_be32(FW_IQ_CMD_TYPE_V(FW_IQ_TYPE_FL_INT_CAP) |
+			    FW_IQ_CMD_IQASYNCH_V(iqasynch) |
+			    FW_IQ_CMD_VIID_V(pi->viid) |
+			    FW_IQ_CMD_IQANDST_V(iqandst) |
+			    FW_IQ_CMD_IQANUS_V(1) |
+			    FW_IQ_CMD_IQANUD_V(SGE_UPDATEDEL_INTR) |
+			    FW_IQ_CMD_IQANDSTINDEX_V(intr_dest));
+	cmd.iqdroprss_to_iqesize =
+		cpu_to_be16(FW_IQ_CMD_IQPCIECH_V(pi->port_id) |
+			    FW_IQ_CMD_IQGTSMODE_F |
+			    FW_IQ_CMD_IQINTCNTTHRESH_V(rspq->pktcnt_idx) |
+			    FW_IQ_CMD_IQESIZE_V(ilog2(rspq->iqe_len) - 4));
+	cmd.iqsize = cpu_to_be16(rspq->size);
+	cmd.iqaddr = cpu_to_be64(rspq->phys_addr);
+
+	if (fl) {
+		/*
+		 * Allocate the ring for the hardware free list (with space
+		 * for its status page) along with the associated software
+		 * descriptor ring.  The free list size needs to be a multiple
+		 * of the Egress Queue Unit.
+		 */
+		fl->size = roundup(fl->size, FL_PER_EQ_UNIT);
+		fl->desc = alloc_ring(adapter->pdev_dev, fl->size,
+				      sizeof(__be64), sizeof(struct rx_sw_desc),
+				      &fl->addr, &fl->sdesc, s->stat_len);
+		if (!fl->desc) {
+			ret = -ENOMEM;
+			goto err;
+		}
+
+		/*
+		 * Calculate the size of the hardware free list ring plus
+		 * Status Page (which the SGE will place after the end of the
+		 * free list ring) in Egress Queue Units.
+		 */
+		flsz = (fl->size / FL_PER_EQ_UNIT +
+			s->stat_len / EQ_UNIT);
+
+		/*
+		 * Fill in all the relevant firmware Ingress Queue Command
+		 * fields for the free list.
+		 */
+		cmd.iqns_to_fl0congen =
+			cpu_to_be32(
+				FW_IQ_CMD_FL0HOSTFCMODE_V(SGE_HOSTFCMODE_NONE) |
+				FW_IQ_CMD_FL0PACKEN_F |
+				FW_IQ_CMD_FL0PADEN_F);
+		cmd.fl0dcaen_to_fl0cidxfthresh =
+			cpu_to_be16(
+				FW_IQ_CMD_FL0FBMIN_V(SGE_FETCHBURSTMIN_64B) |
+				FW_IQ_CMD_FL0FBMAX_V(SGE_FETCHBURSTMAX_512B));
+		cmd.fl0size = cpu_to_be16(flsz);
+		cmd.fl0addr = cpu_to_be64(fl->addr);
+	}
+
+	/*
+	 * Issue the firmware Ingress Queue Command and extract the results if
+	 * it completes successfully.
+	 */
+	ret = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (ret)
+		goto err;
+
+	netif_napi_add(dev, &rspq->napi, napi_rx_handler, 64);
+	rspq->cur_desc = rspq->desc;
+	rspq->cidx = 0;
+	rspq->gen = 1;
+	rspq->next_intr_params = rspq->intr_params;
+	rspq->cntxt_id = be16_to_cpu(rpl.iqid);
+	rspq->bar2_addr = bar2_address(adapter,
+				       rspq->cntxt_id,
+				       T4_BAR2_QTYPE_INGRESS,
+				       &rspq->bar2_qid);
+	rspq->abs_id = be16_to_cpu(rpl.physiqid);
+	rspq->size--;			/* subtract status entry */
+	rspq->adapter = adapter;
+	rspq->netdev = dev;
+	rspq->handler = hnd;
+
+	/* set offset to -1 to distinguish ingress queues without FL */
+	rspq->offset = fl ? 0 : -1;
+
+	if (fl) {
+		fl->cntxt_id = be16_to_cpu(rpl.fl0id);
+		fl->avail = 0;
+		fl->pend_cred = 0;
+		fl->pidx = 0;
+		fl->cidx = 0;
+		fl->alloc_failed = 0;
+		fl->large_alloc_failed = 0;
+		fl->starving = 0;
+
+		/* Note, we must initialize the BAR2 Free List User Doorbell
+		 * information before refilling the Free List!
+		 */
+		fl->bar2_addr = bar2_address(adapter,
+					     fl->cntxt_id,
+					     T4_BAR2_QTYPE_EGRESS,
+					     &fl->bar2_qid);
+
+		refill_fl(adapter, fl, fl_cap(fl), GFP_KERNEL);
+	}
+
+	return 0;
+
+err:
+	/*
+	 * An error occurred.  Clean up our partial allocation state and
+	 * return the error.
+	 */
+	if (rspq->desc) {
+		dma_free_coherent(adapter->pdev_dev, rspq->size * rspq->iqe_len,
+				  rspq->desc, rspq->phys_addr);
+		rspq->desc = NULL;
+	}
+	if (fl && fl->desc) {
+		kfree(fl->sdesc);
+		fl->sdesc = NULL;
+		dma_free_coherent(adapter->pdev_dev, flsz * EQ_UNIT,
+				  fl->desc, fl->addr);
+		fl->desc = NULL;
+	}
+	return ret;
+}
+
+/**
+ *	t4vf_sge_alloc_eth_txq - allocate an SGE Ethernet TX Queue
+ *	@adapter: the adapter
+ *	@txq: pointer to the new txq to be filled in
+ *	@devq: the network TX queue associated with the new txq
+ *	@iqid: the relative ingress queue ID to which events relating to
+ *		the new txq should be directed
+ */
+int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
+			   struct net_device *dev, struct netdev_queue *devq,
+			   unsigned int iqid)
+{
+	struct sge *s = &adapter->sge;
+	int ret, nentries;
+	struct fw_eq_eth_cmd cmd, rpl;
+	struct port_info *pi = netdev_priv(dev);
+
+	/*
+	 * Calculate the size of the hardware TX Queue (including the Status
+	 * Page on the end of the TX Queue) in units of TX Descriptors.
+	 */
+	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
+
+	/*
+	 * Allocate the hardware ring for the TX ring (with space for its
+	 * status page) along with the associated software descriptor ring.
+	 */
+	txq->q.desc = alloc_ring(adapter->pdev_dev, txq->q.size,
+				 sizeof(struct tx_desc),
+				 sizeof(struct tx_sw_desc),
+				 &txq->q.phys_addr, &txq->q.sdesc, s->stat_len);
+	if (!txq->q.desc)
+		return -ENOMEM;
+
+	/*
+	 * Fill in the Egress Queue Command.  Note: As with the direct use of
+	 * the firmware Ingress Queue COmmand above in our RXQ allocation
+	 * routine, ideally, this code would be in t4vf_hw.c.  Again, we'll
+	 * have to see if there's some reasonable way to parameterize it
+	 * into the common code ...
+	 */
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_EQ_ETH_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_WRITE_F |
+				    FW_CMD_EXEC_F);
+	cmd.alloc_to_len16 = cpu_to_be32(FW_EQ_ETH_CMD_ALLOC_F |
+					 FW_EQ_ETH_CMD_EQSTART_F |
+					 FW_LEN16(cmd));
+	cmd.viid_pkd = cpu_to_be32(FW_EQ_ETH_CMD_AUTOEQUEQE_F |
+				   FW_EQ_ETH_CMD_VIID_V(pi->viid));
+	cmd.fetchszm_to_iqid =
+		cpu_to_be32(FW_EQ_ETH_CMD_HOSTFCMODE_V(SGE_HOSTFCMODE_STPG) |
+			    FW_EQ_ETH_CMD_PCIECHN_V(pi->port_id) |
+			    FW_EQ_ETH_CMD_IQID_V(iqid));
+	cmd.dcaen_to_eqsize =
+		cpu_to_be32(FW_EQ_ETH_CMD_FBMIN_V(SGE_FETCHBURSTMIN_64B) |
+			    FW_EQ_ETH_CMD_FBMAX_V(SGE_FETCHBURSTMAX_512B) |
+			    FW_EQ_ETH_CMD_CIDXFTHRESH_V(
+						SGE_CIDXFLUSHTHRESH_32) |
+			    FW_EQ_ETH_CMD_EQSIZE_V(nentries));
+	cmd.eqaddr = cpu_to_be64(txq->q.phys_addr);
+
+	/*
+	 * Issue the firmware Egress Queue Command and extract the results if
+	 * it completes successfully.
+	 */
+	ret = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (ret) {
+		/*
+		 * The girmware Ingress Queue Command failed for some reason.
+		 * Free up our partial allocation state and return the error.
+		 */
+		kfree(txq->q.sdesc);
+		txq->q.sdesc = NULL;
+		dma_free_coherent(adapter->pdev_dev,
+				  nentries * sizeof(struct tx_desc),
+				  txq->q.desc, txq->q.phys_addr);
+		txq->q.desc = NULL;
+		return ret;
+	}
+
+	txq->q.in_use = 0;
+	txq->q.cidx = 0;
+	txq->q.pidx = 0;
+	txq->q.stat = (void *)&txq->q.desc[txq->q.size];
+	txq->q.cntxt_id = FW_EQ_ETH_CMD_EQID_G(be32_to_cpu(rpl.eqid_pkd));
+	txq->q.bar2_addr = bar2_address(adapter,
+					txq->q.cntxt_id,
+					T4_BAR2_QTYPE_EGRESS,
+					&txq->q.bar2_qid);
+	txq->q.abs_id =
+		FW_EQ_ETH_CMD_PHYSEQID_G(be32_to_cpu(rpl.physeqid_pkd));
+	txq->txq = devq;
+	txq->tso = 0;
+	txq->tx_cso = 0;
+	txq->vlan_ins = 0;
+	txq->q.stops = 0;
+	txq->q.restarts = 0;
+	txq->mapping_err = 0;
+	return 0;
+}
+
+/*
+ * Free the DMA map resources associated with a TX queue.
+ */
+static void free_txq(struct adapter *adapter, struct sge_txq *tq)
+{
+	struct sge *s = &adapter->sge;
+
+	dma_free_coherent(adapter->pdev_dev,
+			  tq->size * sizeof(*tq->desc) + s->stat_len,
+			  tq->desc, tq->phys_addr);
+	tq->cntxt_id = 0;
+	tq->sdesc = NULL;
+	tq->desc = NULL;
+}
+
+/*
+ * Free the resources associated with a response queue (possibly including a
+ * free list).
+ */
+static void free_rspq_fl(struct adapter *adapter, struct sge_rspq *rspq,
+			 struct sge_fl *fl)
+{
+	struct sge *s = &adapter->sge;
+	unsigned int flid = fl ? fl->cntxt_id : 0xffff;
+
+	t4vf_iq_free(adapter, FW_IQ_TYPE_FL_INT_CAP,
+		     rspq->cntxt_id, flid, 0xffff);
+	dma_free_coherent(adapter->pdev_dev, (rspq->size + 1) * rspq->iqe_len,
+			  rspq->desc, rspq->phys_addr);
+	netif_napi_del(&rspq->napi);
+	rspq->netdev = NULL;
+	rspq->cntxt_id = 0;
+	rspq->abs_id = 0;
+	rspq->desc = NULL;
+
+	if (fl) {
+		free_rx_bufs(adapter, fl, fl->avail);
+		dma_free_coherent(adapter->pdev_dev,
+				  fl->size * sizeof(*fl->desc) + s->stat_len,
+				  fl->desc, fl->addr);
+		kfree(fl->sdesc);
+		fl->sdesc = NULL;
+		fl->cntxt_id = 0;
+		fl->desc = NULL;
+	}
+}
+
+/**
+ *	t4vf_free_sge_resources - free SGE resources
+ *	@adapter: the adapter
+ *
+ *	Frees resources used by the SGE queue sets.
+ */
+void t4vf_free_sge_resources(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+	struct sge_eth_rxq *rxq = s->ethrxq;
+	struct sge_eth_txq *txq = s->ethtxq;
+	struct sge_rspq *evtq = &s->fw_evtq;
+	struct sge_rspq *intrq = &s->intrq;
+	int qs;
+
+	for (qs = 0; qs < adapter->sge.ethqsets; qs++, rxq++, txq++) {
+		if (rxq->rspq.desc)
+			free_rspq_fl(adapter, &rxq->rspq, &rxq->fl);
+		if (txq->q.desc) {
+			t4vf_eth_eq_free(adapter, txq->q.cntxt_id);
+			free_tx_desc(adapter, &txq->q, txq->q.in_use, true);
+			kfree(txq->q.sdesc);
+			free_txq(adapter, &txq->q);
+		}
+	}
+	if (evtq->desc)
+		free_rspq_fl(adapter, evtq, NULL);
+	if (intrq->desc)
+		free_rspq_fl(adapter, intrq, NULL);
+}
+
+/**
+ *	t4vf_sge_start - enable SGE operation
+ *	@adapter: the adapter
+ *
+ *	Start tasklets and timers associated with the DMA engine.
+ */
+void t4vf_sge_start(struct adapter *adapter)
+{
+	adapter->sge.ethtxq_rover = 0;
+	mod_timer(&adapter->sge.rx_timer, jiffies + RX_QCHECK_PERIOD);
+	mod_timer(&adapter->sge.tx_timer, jiffies + TX_QCHECK_PERIOD);
+}
+
+/**
+ *	t4vf_sge_stop - disable SGE operation
+ *	@adapter: the adapter
+ *
+ *	Stop tasklets and timers associated with the DMA engine.  Note that
+ *	this is effective only if measures have been taken to disable any HW
+ *	events that may restart them.
+ */
+void t4vf_sge_stop(struct adapter *adapter)
+{
+	struct sge *s = &adapter->sge;
+
+	if (s->rx_timer.function)
+		del_timer_sync(&s->rx_timer);
+	if (s->tx_timer.function)
+		del_timer_sync(&s->tx_timer);
+}
+
+/**
+ *	t4vf_sge_init - initialize SGE
+ *	@adapter: the adapter
+ *
+ *	Performs SGE initialization needed every time after a chip reset.
+ *	We do not initialize any of the queue sets here, instead the driver
+ *	top-level must request those individually.  We also do not enable DMA
+ *	here, that should be done after the queues have been set up.
+ */
+int t4vf_sge_init(struct adapter *adapter)
+{
+	struct sge_params *sge_params = &adapter->params.sge;
+	u32 fl0 = sge_params->sge_fl_buffer_size[0];
+	u32 fl1 = sge_params->sge_fl_buffer_size[1];
+	struct sge *s = &adapter->sge;
+	unsigned int ingpadboundary, ingpackboundary;
+
+	/*
+	 * Start by vetting the basic SGE parameters which have been set up by
+	 * the Physical Function Driver.  Ideally we should be able to deal
+	 * with _any_ configuration.  Practice is different ...
+	 */
+	if (fl0 != PAGE_SIZE || (fl1 != 0 && fl1 <= fl0)) {
+		dev_err(adapter->pdev_dev, "bad SGE FL buffer sizes [%d, %d]\n",
+			fl0, fl1);
+		return -EINVAL;
+	}
+	if ((sge_params->sge_control & RXPKTCPLMODE_F) == 0) {
+		dev_err(adapter->pdev_dev, "bad SGE CPL MODE\n");
+		return -EINVAL;
+	}
+
+	/*
+	 * Now translate the adapter parameters into our internal forms.
+	 */
+	if (fl1)
+		s->fl_pg_order = ilog2(fl1) - PAGE_SHIFT;
+	s->stat_len = ((sge_params->sge_control & EGRSTATUSPAGESIZE_F)
+			? 128 : 64);
+	s->pktshift = PKTSHIFT_G(sge_params->sge_control);
+
+	/* T4 uses a single control field to specify both the PCIe Padding and
+	 * Packing Boundary.  T5 introduced the ability to specify these
+	 * separately.  The actual Ingress Packet Data alignment boundary
+	 * within Packed Buffer Mode is the maximum of these two
+	 * specifications.  (Note that it makes no real practical sense to
+	 * have the Pading Boudary be larger than the Packing Boundary but you
+	 * could set the chip up that way and, in fact, legacy T4 code would
+	 * end doing this because it would initialize the Padding Boundary and
+	 * leave the Packing Boundary initialized to 0 (16 bytes).)
+	 */
+	ingpadboundary = 1 << (INGPADBOUNDARY_G(sge_params->sge_control) +
+			       INGPADBOUNDARY_SHIFT_X);
+	if (is_t4(adapter->params.chip)) {
+		s->fl_align = ingpadboundary;
+	} else {
+		/* T5 has a different interpretation of one of the PCIe Packing
+		 * Boundary values.
+		 */
+		ingpackboundary = INGPACKBOUNDARY_G(sge_params->sge_control2);
+		if (ingpackboundary == INGPACKBOUNDARY_16B_X)
+			ingpackboundary = 16;
+		else
+			ingpackboundary = 1 << (ingpackboundary +
+						INGPACKBOUNDARY_SHIFT_X);
+
+		s->fl_align = max(ingpadboundary, ingpackboundary);
+	}
+
+	/* A FL with <= fl_starve_thres buffers is starving and a periodic
+	 * timer will attempt to refill it.  This needs to be larger than the
+	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
+	 * stuck waiting for new packets while the SGE is waiting for us to
+	 * give it more Free List entries.  (Note that the SGE's Egress
+	 * Congestion Threshold is in units of 2 Free List pointers.)
+	 */
+	s->fl_starve_thres
+		= EGRTHRESHOLD_G(sge_params->sge_congestion_control)*2 + 1;
+
+	/*
+	 * Set up tasklet timers.
+	 */
+	setup_timer(&s->rx_timer, sge_rx_timer_cb, (unsigned long)adapter);
+	setup_timer(&s->tx_timer, sge_tx_timer_cb, (unsigned long)adapter);
+
+	/*
+	 * Initialize Forwarded Interrupt Queue lock.
+	 */
+	spin_lock_init(&s->intrq_lock);
+
+	return 0;
+}
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h
new file mode 100644
index 000000000..b9debb4f2
--- /dev/null
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h
@@ -0,0 +1,326 @@
+/*
+ * This file is part of the Chelsio T4 PCI-E SR-IOV Virtual Function Ethernet
+ * driver for Linux.
+ *
+ * Copyright (c) 2009-2010 Chelsio Communications, 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 __T4VF_COMMON_H__
+#define __T4VF_COMMON_H__
+
+#include "../cxgb4/t4fw_api.h"
+
+#define CHELSIO_CHIP_CODE(version, revision) (((version) << 4) | (revision))
+#define CHELSIO_CHIP_VERSION(code) (((code) >> 4) & 0xf)
+#define CHELSIO_CHIP_RELEASE(code) ((code) & 0xf)
+
+/* All T4 and later chips have their PCI-E Device IDs encoded as 0xVFPP where:
+ *
+ *   V  = "4" for T4; "5" for T5, etc. or
+ *      = "a" for T4 FPGA; "b" for T4 FPGA, etc.
+ *   F  = "0" for PF 0..3; "4".."7" for PF4..7; and "8" for VFs
+ *   PP = adapter product designation
+ */
+#define CHELSIO_T4		0x4
+#define CHELSIO_T5		0x5
+
+enum chip_type {
+	T4_A1 = CHELSIO_CHIP_CODE(CHELSIO_T4, 1),
+	T4_A2 = CHELSIO_CHIP_CODE(CHELSIO_T4, 2),
+	T4_FIRST_REV	= T4_A1,
+	T4_LAST_REV	= T4_A2,
+
+	T5_A0 = CHELSIO_CHIP_CODE(CHELSIO_T5, 0),
+	T5_A1 = CHELSIO_CHIP_CODE(CHELSIO_T5, 1),
+	T5_FIRST_REV	= T5_A0,
+	T5_LAST_REV	= T5_A1,
+};
+
+/*
+ * The "len16" field of a Firmware Command Structure ...
+ */
+#define FW_LEN16(fw_struct) FW_CMD_LEN16_V(sizeof(fw_struct) / 16)
+
+/*
+ * Per-VF statistics.
+ */
+struct t4vf_port_stats {
+	/*
+	 * TX statistics.
+	 */
+	u64 tx_bcast_bytes;		/* broadcast */
+	u64 tx_bcast_frames;
+	u64 tx_mcast_bytes;		/* multicast */
+	u64 tx_mcast_frames;
+	u64 tx_ucast_bytes;		/* unicast */
+	u64 tx_ucast_frames;
+	u64 tx_drop_frames;		/* TX dropped frames */
+	u64 tx_offload_bytes;		/* offload */
+	u64 tx_offload_frames;
+
+	/*
+	 * RX statistics.
+	 */
+	u64 rx_bcast_bytes;		/* broadcast */
+	u64 rx_bcast_frames;
+	u64 rx_mcast_bytes;		/* multicast */
+	u64 rx_mcast_frames;
+	u64 rx_ucast_bytes;
+	u64 rx_ucast_frames;		/* unicast */
+
+	u64 rx_err_frames;		/* RX error frames */
+};
+
+/*
+ * Per-"port" (Virtual Interface) link configuration ...
+ */
+struct link_config {
+	unsigned int   supported;        /* link capabilities */
+	unsigned int   advertising;      /* advertised capabilities */
+	unsigned short requested_speed;  /* speed user has requested */
+	unsigned short speed;            /* actual link speed */
+	unsigned char  requested_fc;     /* flow control user has requested */
+	unsigned char  fc;               /* actual link flow control */
+	unsigned char  autoneg;          /* autonegotiating? */
+	unsigned char  link_ok;          /* link up? */
+};
+
+enum {
+	PAUSE_RX      = 1 << 0,
+	PAUSE_TX      = 1 << 1,
+	PAUSE_AUTONEG = 1 << 2
+};
+
+/*
+ * General device parameters ...
+ */
+struct dev_params {
+	u32 fwrev;			/* firmware version */
+	u32 tprev;			/* TP Microcode Version */
+};
+
+/*
+ * Scatter Gather Engine parameters.  These are almost all determined by the
+ * Physical Function Driver.  We just need to grab them to see within which
+ * environment we're playing ...
+ */
+struct sge_params {
+	u32 sge_control;		/* padding, boundaries, lengths, etc. */
+	u32 sge_control2;		/* T5: more of the same */
+	u32 sge_host_page_size;		/* PF0-7 page sizes */
+	u32 sge_egress_queues_per_page;	/* PF0-7 egress queues/page */
+	u32 sge_ingress_queues_per_page;/* PF0-7 ingress queues/page */
+	u32 sge_vf_hps;                 /* host page size for our vf */
+	u32 sge_vf_eq_qpp;		/* egress queues/page for our VF */
+	u32 sge_vf_iq_qpp;		/* ingress queues/page for our VF */
+	u32 sge_fl_buffer_size[16];	/* free list buffer sizes */
+	u32 sge_ingress_rx_threshold;	/* RX counter interrupt threshold[4] */
+	u32 sge_congestion_control;     /* congestion thresholds, etc. */
+	u32 sge_timer_value_0_and_1;	/* interrupt coalescing timer values */
+	u32 sge_timer_value_2_and_3;
+	u32 sge_timer_value_4_and_5;
+};
+
+/*
+ * Vital Product Data parameters.
+ */
+struct vpd_params {
+	u32 cclk;			/* Core Clock (KHz) */
+};
+
+/*
+ * Global Receive Side Scaling (RSS) parameters in host-native format.
+ */
+struct rss_params {
+	unsigned int mode;		/* RSS mode */
+	union {
+	    struct {
+		unsigned int synmapen:1;	/* SYN Map Enable */
+		unsigned int syn4tupenipv6:1;	/* enable hashing 4-tuple IPv6 SYNs */
+		unsigned int syn2tupenipv6:1;	/* enable hashing 2-tuple IPv6 SYNs */
+		unsigned int syn4tupenipv4:1;	/* enable hashing 4-tuple IPv4 SYNs */
+		unsigned int syn2tupenipv4:1;	/* enable hashing 2-tuple IPv4 SYNs */
+		unsigned int ofdmapen:1;	/* Offload Map Enable */
+		unsigned int tnlmapen:1;	/* Tunnel Map Enable */
+		unsigned int tnlalllookup:1;	/* Tunnel All Lookup */
+		unsigned int hashtoeplitz:1;	/* use Toeplitz hash */
+	    } basicvirtual;
+	} u;
+};
+
+/*
+ * Virtual Interface RSS Configuration in host-native format.
+ */
+union rss_vi_config {
+    struct {
+	u16 defaultq;			/* Ingress Queue ID for !tnlalllookup */
+	unsigned int ip6fourtupen:1;	/* hash 4-tuple IPv6 ingress packets */
+	unsigned int ip6twotupen:1;	/* hash 2-tuple IPv6 ingress packets */
+	unsigned int ip4fourtupen:1;	/* hash 4-tuple IPv4 ingress packets */
+	unsigned int ip4twotupen:1;	/* hash 2-tuple IPv4 ingress packets */
+	int udpen;			/* hash 4-tuple UDP ingress packets */
+    } basicvirtual;
+};
+
+/*
+ * Maximum resources provisioned for a PCI VF.
+ */
+struct vf_resources {
+	unsigned int nvi;		/* N virtual interfaces */
+	unsigned int neq;		/* N egress Qs */
+	unsigned int nethctrl;		/* N egress ETH or CTRL Qs */
+	unsigned int niqflint;		/* N ingress Qs/w free list(s) & intr */
+	unsigned int niq;		/* N ingress Qs */
+	unsigned int tc;		/* PCI-E traffic class */
+	unsigned int pmask;		/* port access rights mask */
+	unsigned int nexactf;		/* N exact MPS filters */
+	unsigned int r_caps;		/* read capabilities */
+	unsigned int wx_caps;		/* write/execute capabilities */
+};
+
+/*
+ * Per-"adapter" (Virtual Function) parameters.
+ */
+struct adapter_params {
+	struct dev_params dev;		/* general device parameters */
+	struct sge_params sge;		/* Scatter Gather Engine */
+	struct vpd_params vpd;		/* Vital Product Data */
+	struct rss_params rss;		/* Receive Side Scaling */
+	struct vf_resources vfres;	/* Virtual Function Resource limits */
+	enum chip_type chip;		/* chip code */
+	u8 nports;			/* # of Ethernet "ports" */
+};
+
+#include "adapter.h"
+
+#ifndef PCI_VENDOR_ID_CHELSIO
+# define PCI_VENDOR_ID_CHELSIO 0x1425
+#endif
+
+#define for_each_port(adapter, iter) \
+	for (iter = 0; iter < (adapter)->params.nports; iter++)
+
+static inline bool is_10g_port(const struct link_config *lc)
+{
+	return (lc->supported & FW_PORT_CAP_SPEED_10G) != 0;
+}
+
+static inline bool is_x_10g_port(const struct link_config *lc)
+{
+	return (lc->supported & FW_PORT_CAP_SPEED_10G) != 0 ||
+		(lc->supported & FW_PORT_CAP_SPEED_40G) != 0;
+}
+
+static inline unsigned int core_ticks_per_usec(const struct adapter *adapter)
+{
+	return adapter->params.vpd.cclk / 1000;
+}
+
+static inline unsigned int us_to_core_ticks(const struct adapter *adapter,
+					    unsigned int us)
+{
+	return (us * adapter->params.vpd.cclk) / 1000;
+}
+
+static inline unsigned int core_ticks_to_us(const struct adapter *adapter,
+					    unsigned int ticks)
+{
+	return (ticks * 1000) / adapter->params.vpd.cclk;
+}
+
+int t4vf_wr_mbox_core(struct adapter *, const void *, int, void *, bool);
+
+static inline int t4vf_wr_mbox(struct adapter *adapter, const void *cmd,
+			       int size, void *rpl)
+{
+	return t4vf_wr_mbox_core(adapter, cmd, size, rpl, true);
+}
+
+static inline int t4vf_wr_mbox_ns(struct adapter *adapter, const void *cmd,
+				  int size, void *rpl)
+{
+	return t4vf_wr_mbox_core(adapter, cmd, size, rpl, false);
+}
+
+#define CHELSIO_PCI_ID_VER(dev_id)  ((dev_id) >> 12)
+
+static inline int is_t4(enum chip_type chip)
+{
+	return CHELSIO_CHIP_VERSION(chip) == CHELSIO_T4;
+}
+
+int t4vf_wait_dev_ready(struct adapter *);
+int t4vf_port_init(struct adapter *, int);
+
+int t4vf_fw_reset(struct adapter *);
+int t4vf_set_params(struct adapter *, unsigned int, const u32 *, const u32 *);
+
+enum t4_bar2_qtype { T4_BAR2_QTYPE_EGRESS, T4_BAR2_QTYPE_INGRESS };
+int t4_bar2_sge_qregs(struct adapter *adapter,
+		      unsigned int qid,
+		      enum t4_bar2_qtype qtype,
+		      u64 *pbar2_qoffset,
+		      unsigned int *pbar2_qid);
+
+int t4vf_get_sge_params(struct adapter *);
+int t4vf_get_vpd_params(struct adapter *);
+int t4vf_get_dev_params(struct adapter *);
+int t4vf_get_rss_glb_config(struct adapter *);
+int t4vf_get_vfres(struct adapter *);
+
+int t4vf_read_rss_vi_config(struct adapter *, unsigned int,
+			    union rss_vi_config *);
+int t4vf_write_rss_vi_config(struct adapter *, unsigned int,
+			     union rss_vi_config *);
+int t4vf_config_rss_range(struct adapter *, unsigned int, int, int,
+			  const u16 *, int);
+
+int t4vf_alloc_vi(struct adapter *, int);
+int t4vf_free_vi(struct adapter *, int);
+int t4vf_enable_vi(struct adapter *, unsigned int, bool, bool);
+int t4vf_identify_port(struct adapter *, unsigned int, unsigned int);
+
+int t4vf_set_rxmode(struct adapter *, unsigned int, int, int, int, int, int,
+		    bool);
+int t4vf_alloc_mac_filt(struct adapter *, unsigned int, bool, unsigned int,
+			const u8 **, u16 *, u64 *, bool);
+int t4vf_change_mac(struct adapter *, unsigned int, int, const u8 *, bool);
+int t4vf_set_addr_hash(struct adapter *, unsigned int, bool, u64, bool);
+int t4vf_get_port_stats(struct adapter *, int, struct t4vf_port_stats *);
+
+int t4vf_iq_free(struct adapter *, unsigned int, unsigned int, unsigned int,
+		 unsigned int);
+int t4vf_eth_eq_free(struct adapter *, unsigned int);
+
+int t4vf_handle_fw_rpl(struct adapter *, const __be64 *);
+int t4vf_prep_adapter(struct adapter *);
+
+#endif /* __T4VF_COMMON_H__ */
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_defs.h b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_defs.h
new file mode 100644
index 000000000..b516b12b1
--- /dev/null
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_defs.h
@@ -0,0 +1,121 @@
+/*
+ * This file is part of the Chelsio T4 PCI-E SR-IOV Virtual Function Ethernet
+ * driver for Linux.
+ *
+ * Copyright (c) 2009-2010 Chelsio Communications, 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 __T4VF_DEFS_H__
+#define __T4VF_DEFS_H__
+
+#include "../cxgb4/t4_regs.h"
+
+/*
+ * The VF Register Map.
+ *
+ * The Scatter Gather Engine (SGE), Multiport Support module (MPS), PIO Local
+ * bus module (PL) and CPU Interface Module (CIM) components are mapped via
+ * the Slice to Module Map Table (see below) in the Physical Function Register
+ * Map.  The Mail Box Data (MBDATA) range is mapped via the PCI-E Mailbox Base
+ * and Offset registers in the PF Register Map.  The MBDATA base address is
+ * quite constrained as it determines the Mailbox Data addresses for both PFs
+ * and VFs, and therefore must fit in both the VF and PF Register Maps without
+ * overlapping other registers.
+ */
+#define T4VF_SGE_BASE_ADDR	0x0000
+#define T4VF_MPS_BASE_ADDR	0x0100
+#define T4VF_PL_BASE_ADDR	0x0200
+#define T4VF_MBDATA_BASE_ADDR	0x0240
+#define T4VF_CIM_BASE_ADDR	0x0300
+
+#define T4VF_REGMAP_START	0x0000
+#define T4VF_REGMAP_SIZE	0x0400
+
+/*
+ * There's no hardware limitation which requires that the addresses of the
+ * Mailbox Data in the fixed CIM PF map and the programmable VF map must
+ * match.  However, it's a useful convention ...
+ */
+#if T4VF_MBDATA_BASE_ADDR != CIM_PF_MAILBOX_DATA_A
+#error T4VF_MBDATA_BASE_ADDR must match CIM_PF_MAILBOX_DATA_A!
+#endif
+
+/*
+ * Virtual Function "Slice to Module Map Table" definitions.
+ *
+ * This table allows us to map subsets of the various module register sets
+ * into the T4VF Register Map.  Each table entry identifies the index of the
+ * module whose registers are being mapped, the offset within the module's
+ * register set that the mapping should start at, the limit of the mapping,
+ * and the offset within the T4VF Register Map to which the module's registers
+ * are being mapped.  All addresses and qualtities are in terms of 32-bit
+ * words.  The "limit" value is also in terms of 32-bit words and is equal to
+ * the last address mapped in the T4VF Register Map 1 (i.e. it's a "<="
+ * relation rather than a "<").
+ */
+#define T4VF_MOD_MAP(module, index, first, last) \
+	T4VF_MOD_MAP_##module##_INDEX  = (index), \
+	T4VF_MOD_MAP_##module##_FIRST  = (first), \
+	T4VF_MOD_MAP_##module##_LAST   = (last), \
+	T4VF_MOD_MAP_##module##_OFFSET = ((first)/4), \
+	T4VF_MOD_MAP_##module##_BASE = \
+		(T4VF_##module##_BASE_ADDR/4 + (first)/4), \
+	T4VF_MOD_MAP_##module##_LIMIT = \
+		(T4VF_##module##_BASE_ADDR/4 + (last)/4),
+
+#define SGE_VF_KDOORBELL 0x0
+#define SGE_VF_GTS 0x4
+#define MPS_VF_CTL 0x0
+#define MPS_VF_STAT_RX_VF_ERR_FRAMES_H 0xfc
+#define PL_VF_WHOAMI 0x0
+#define CIM_VF_EXT_MAILBOX_CTRL 0x0
+#define CIM_VF_EXT_MAILBOX_STATUS 0x4
+
+enum {
+    T4VF_MOD_MAP(SGE, 2, SGE_VF_KDOORBELL, SGE_VF_GTS)
+    T4VF_MOD_MAP(MPS, 0, MPS_VF_CTL, MPS_VF_STAT_RX_VF_ERR_FRAMES_H)
+    T4VF_MOD_MAP(PL,  3, PL_VF_WHOAMI, PL_VF_WHOAMI)
+    T4VF_MOD_MAP(CIM, 1, CIM_VF_EXT_MAILBOX_CTRL, CIM_VF_EXT_MAILBOX_STATUS)
+};
+
+/*
+ * There isn't a Slice to Module Map Table entry for the Mailbox Data
+ * registers, but it's convenient to use similar names as above.  There are 8
+ * little-endian 64-bit Mailbox Data registers.  Note that the "instances"
+ * value below is in terms of 32-bit words which matches the "word" addressing
+ * space we use above for the Slice to Module Map Space.
+ */
+#define NUM_CIM_VF_MAILBOX_DATA_INSTANCES 16
+
+#define T4VF_MBDATA_FIRST	0
+#define T4VF_MBDATA_LAST	((NUM_CIM_VF_MAILBOX_DATA_INSTANCES-1)*4)
+
+#endif /* __T4T4VF_DEFS_H__ */
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c
new file mode 100644
index 000000000..966ee900e
--- /dev/null
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c
@@ -0,0 +1,1602 @@
+/*
+ * This file is part of the Chelsio T4 PCI-E SR-IOV Virtual Function Ethernet
+ * driver for Linux.
+ *
+ * Copyright (c) 2009-2010 Chelsio Communications, 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.
+ */
+
+#include <linux/pci.h>
+
+#include "t4vf_common.h"
+#include "t4vf_defs.h"
+
+#include "../cxgb4/t4_regs.h"
+#include "../cxgb4/t4_values.h"
+#include "../cxgb4/t4fw_api.h"
+
+/*
+ * Wait for the device to become ready (signified by our "who am I" register
+ * returning a value other than all 1's).  Return an error if it doesn't
+ * become ready ...
+ */
+int t4vf_wait_dev_ready(struct adapter *adapter)
+{
+	const u32 whoami = T4VF_PL_BASE_ADDR + PL_VF_WHOAMI;
+	const u32 notready1 = 0xffffffff;
+	const u32 notready2 = 0xeeeeeeee;
+	u32 val;
+
+	val = t4_read_reg(adapter, whoami);
+	if (val != notready1 && val != notready2)
+		return 0;
+	msleep(500);
+	val = t4_read_reg(adapter, whoami);
+	if (val != notready1 && val != notready2)
+		return 0;
+	else
+		return -EIO;
+}
+
+/*
+ * Get the reply to a mailbox command and store it in @rpl in big-endian order
+ * (since the firmware data structures are specified in a big-endian layout).
+ */
+static void get_mbox_rpl(struct adapter *adapter, __be64 *rpl, int size,
+			 u32 mbox_data)
+{
+	for ( ; size; size -= 8, mbox_data += 8)
+		*rpl++ = cpu_to_be64(t4_read_reg64(adapter, mbox_data));
+}
+
+/*
+ * Dump contents of mailbox with a leading tag.
+ */
+static void dump_mbox(struct adapter *adapter, const char *tag, u32 mbox_data)
+{
+	dev_err(adapter->pdev_dev,
+		"mbox %s: %llx %llx %llx %llx %llx %llx %llx %llx\n", tag,
+		(unsigned long long)t4_read_reg64(adapter, mbox_data +  0),
+		(unsigned long long)t4_read_reg64(adapter, mbox_data +  8),
+		(unsigned long long)t4_read_reg64(adapter, mbox_data + 16),
+		(unsigned long long)t4_read_reg64(adapter, mbox_data + 24),
+		(unsigned long long)t4_read_reg64(adapter, mbox_data + 32),
+		(unsigned long long)t4_read_reg64(adapter, mbox_data + 40),
+		(unsigned long long)t4_read_reg64(adapter, mbox_data + 48),
+		(unsigned long long)t4_read_reg64(adapter, mbox_data + 56));
+}
+
+/**
+ *	t4vf_wr_mbox_core - send a command to FW through the mailbox
+ *	@adapter: the adapter
+ *	@cmd: the command to write
+ *	@size: command length in bytes
+ *	@rpl: where to optionally store the reply
+ *	@sleep_ok: if true we may sleep while awaiting command completion
+ *
+ *	Sends the given command to FW through the mailbox and waits for the
+ *	FW to execute the command.  If @rpl is not %NULL it is used to store
+ *	the FW's reply to the command.  The command and its optional reply
+ *	are of the same length.  FW can take up to 500 ms to respond.
+ *	@sleep_ok determines whether we may sleep while awaiting the response.
+ *	If sleeping is allowed we use progressive backoff otherwise we spin.
+ *
+ *	The return value is 0 on success or a negative errno on failure.  A
+ *	failure can happen either because we are not able to execute the
+ *	command or FW executes it but signals an error.  In the latter case
+ *	the return value is the error code indicated by FW (negated).
+ */
+int t4vf_wr_mbox_core(struct adapter *adapter, const void *cmd, int size,
+		      void *rpl, bool sleep_ok)
+{
+	static const int delay[] = {
+		1, 1, 3, 5, 10, 10, 20, 50, 100
+	};
+
+	u32 v;
+	int i, ms, delay_idx;
+	const __be64 *p;
+	u32 mbox_data = T4VF_MBDATA_BASE_ADDR;
+	u32 mbox_ctl = T4VF_CIM_BASE_ADDR + CIM_VF_EXT_MAILBOX_CTRL;
+
+	/*
+	 * Commands must be multiples of 16 bytes in length and may not be
+	 * larger than the size of the Mailbox Data register array.
+	 */
+	if ((size % 16) != 0 ||
+	    size > NUM_CIM_VF_MAILBOX_DATA_INSTANCES * 4)
+		return -EINVAL;
+
+	/*
+	 * Loop trying to get ownership of the mailbox.  Return an error
+	 * if we can't gain ownership.
+	 */
+	v = MBOWNER_G(t4_read_reg(adapter, mbox_ctl));
+	for (i = 0; v == MBOX_OWNER_NONE && i < 3; i++)
+		v = MBOWNER_G(t4_read_reg(adapter, mbox_ctl));
+	if (v != MBOX_OWNER_DRV)
+		return v == MBOX_OWNER_FW ? -EBUSY : -ETIMEDOUT;
+
+	/*
+	 * Write the command array into the Mailbox Data register array and
+	 * transfer ownership of the mailbox to the firmware.
+	 *
+	 * For the VFs, the Mailbox Data "registers" are actually backed by
+	 * T4's "MA" interface rather than PL Registers (as is the case for
+	 * the PFs).  Because these are in different coherency domains, the
+	 * write to the VF's PL-register-backed Mailbox Control can race in
+	 * front of the writes to the MA-backed VF Mailbox Data "registers".
+	 * So we need to do a read-back on at least one byte of the VF Mailbox
+	 * Data registers before doing the write to the VF Mailbox Control
+	 * register.
+	 */
+	for (i = 0, p = cmd; i < size; i += 8)
+		t4_write_reg64(adapter, mbox_data + i, be64_to_cpu(*p++));
+	t4_read_reg(adapter, mbox_data);         /* flush write */
+
+	t4_write_reg(adapter, mbox_ctl,
+		     MBMSGVALID_F | MBOWNER_V(MBOX_OWNER_FW));
+	t4_read_reg(adapter, mbox_ctl);          /* flush write */
+
+	/*
+	 * Spin waiting for firmware to acknowledge processing our command.
+	 */
+	delay_idx = 0;
+	ms = delay[0];
+
+	for (i = 0; i < FW_CMD_MAX_TIMEOUT; i += ms) {
+		if (sleep_ok) {
+			ms = delay[delay_idx];
+			if (delay_idx < ARRAY_SIZE(delay) - 1)
+				delay_idx++;
+			msleep(ms);
+		} else
+			mdelay(ms);
+
+		/*
+		 * If we're the owner, see if this is the reply we wanted.
+		 */
+		v = t4_read_reg(adapter, mbox_ctl);
+		if (MBOWNER_G(v) == MBOX_OWNER_DRV) {
+			/*
+			 * If the Message Valid bit isn't on, revoke ownership
+			 * of the mailbox and continue waiting for our reply.
+			 */
+			if ((v & MBMSGVALID_F) == 0) {
+				t4_write_reg(adapter, mbox_ctl,
+					     MBOWNER_V(MBOX_OWNER_NONE));
+				continue;
+			}
+
+			/*
+			 * We now have our reply.  Extract the command return
+			 * value, copy the reply back to our caller's buffer
+			 * (if specified) and revoke ownership of the mailbox.
+			 * We return the (negated) firmware command return
+			 * code (this depends on FW_SUCCESS == 0).
+			 */
+
+			/* return value in low-order little-endian word */
+			v = t4_read_reg(adapter, mbox_data);
+			if (FW_CMD_RETVAL_G(v))
+				dump_mbox(adapter, "FW Error", mbox_data);
+
+			if (rpl) {
+				/* request bit in high-order BE word */
+				WARN_ON((be32_to_cpu(*(const __be32 *)cmd)
+					 & FW_CMD_REQUEST_F) == 0);
+				get_mbox_rpl(adapter, rpl, size, mbox_data);
+				WARN_ON((be32_to_cpu(*(__be32 *)rpl)
+					 & FW_CMD_REQUEST_F) != 0);
+			}
+			t4_write_reg(adapter, mbox_ctl,
+				     MBOWNER_V(MBOX_OWNER_NONE));
+			return -FW_CMD_RETVAL_G(v);
+		}
+	}
+
+	/*
+	 * We timed out.  Return the error ...
+	 */
+	dump_mbox(adapter, "FW Timeout", mbox_data);
+	return -ETIMEDOUT;
+}
+
+/**
+ *	hash_mac_addr - return the hash value of a MAC address
+ *	@addr: the 48-bit Ethernet MAC address
+ *
+ *	Hashes a MAC address according to the hash function used by hardware
+ *	inexact (hash) address matching.
+ */
+static int hash_mac_addr(const u8 *addr)
+{
+	u32 a = ((u32)addr[0] << 16) | ((u32)addr[1] << 8) | addr[2];
+	u32 b = ((u32)addr[3] << 16) | ((u32)addr[4] << 8) | addr[5];
+	a ^= b;
+	a ^= (a >> 12);
+	a ^= (a >> 6);
+	return a & 0x3f;
+}
+
+#define ADVERT_MASK (FW_PORT_CAP_SPEED_100M | FW_PORT_CAP_SPEED_1G |\
+		     FW_PORT_CAP_SPEED_10G | FW_PORT_CAP_SPEED_40G | \
+		     FW_PORT_CAP_SPEED_100G | FW_PORT_CAP_ANEG)
+
+/**
+ *	init_link_config - initialize a link's SW state
+ *	@lc: structure holding the link state
+ *	@caps: link capabilities
+ *
+ *	Initializes the SW state maintained for each link, including the link's
+ *	capabilities and default speed/flow-control/autonegotiation settings.
+ */
+static void init_link_config(struct link_config *lc, unsigned int caps)
+{
+	lc->supported = caps;
+	lc->requested_speed = 0;
+	lc->speed = 0;
+	lc->requested_fc = lc->fc = PAUSE_RX | PAUSE_TX;
+	if (lc->supported & FW_PORT_CAP_ANEG) {
+		lc->advertising = lc->supported & ADVERT_MASK;
+		lc->autoneg = AUTONEG_ENABLE;
+		lc->requested_fc |= PAUSE_AUTONEG;
+	} else {
+		lc->advertising = 0;
+		lc->autoneg = AUTONEG_DISABLE;
+	}
+}
+
+/**
+ *	t4vf_port_init - initialize port hardware/software state
+ *	@adapter: the adapter
+ *	@pidx: the adapter port index
+ */
+int t4vf_port_init(struct adapter *adapter, int pidx)
+{
+	struct port_info *pi = adap2pinfo(adapter, pidx);
+	struct fw_vi_cmd vi_cmd, vi_rpl;
+	struct fw_port_cmd port_cmd, port_rpl;
+	int v;
+
+	/*
+	 * Execute a VI Read command to get our Virtual Interface information
+	 * like MAC address, etc.
+	 */
+	memset(&vi_cmd, 0, sizeof(vi_cmd));
+	vi_cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_VI_CMD) |
+				       FW_CMD_REQUEST_F |
+				       FW_CMD_READ_F);
+	vi_cmd.alloc_to_len16 = cpu_to_be32(FW_LEN16(vi_cmd));
+	vi_cmd.type_viid = cpu_to_be16(FW_VI_CMD_VIID_V(pi->viid));
+	v = t4vf_wr_mbox(adapter, &vi_cmd, sizeof(vi_cmd), &vi_rpl);
+	if (v)
+		return v;
+
+	BUG_ON(pi->port_id != FW_VI_CMD_PORTID_G(vi_rpl.portid_pkd));
+	pi->rss_size = FW_VI_CMD_RSSSIZE_G(be16_to_cpu(vi_rpl.rsssize_pkd));
+	t4_os_set_hw_addr(adapter, pidx, vi_rpl.mac);
+
+	/*
+	 * If we don't have read access to our port information, we're done
+	 * now.  Otherwise, execute a PORT Read command to get it ...
+	 */
+	if (!(adapter->params.vfres.r_caps & FW_CMD_CAP_PORT))
+		return 0;
+
+	memset(&port_cmd, 0, sizeof(port_cmd));
+	port_cmd.op_to_portid = cpu_to_be32(FW_CMD_OP_V(FW_PORT_CMD) |
+					    FW_CMD_REQUEST_F |
+					    FW_CMD_READ_F |
+					    FW_PORT_CMD_PORTID_V(pi->port_id));
+	port_cmd.action_to_len16 =
+		cpu_to_be32(FW_PORT_CMD_ACTION_V(FW_PORT_ACTION_GET_PORT_INFO) |
+			    FW_LEN16(port_cmd));
+	v = t4vf_wr_mbox(adapter, &port_cmd, sizeof(port_cmd), &port_rpl);
+	if (v)
+		return v;
+
+	v = be32_to_cpu(port_rpl.u.info.lstatus_to_modtype);
+	pi->mdio_addr = (v & FW_PORT_CMD_MDIOCAP_F) ?
+			FW_PORT_CMD_MDIOADDR_G(v) : -1;
+	pi->port_type = FW_PORT_CMD_PTYPE_G(v);
+	pi->mod_type = FW_PORT_MOD_TYPE_NA;
+
+	init_link_config(&pi->link_cfg, be16_to_cpu(port_rpl.u.info.pcap));
+
+	return 0;
+}
+
+/**
+ *      t4vf_fw_reset - issue a reset to FW
+ *      @adapter: the adapter
+ *
+ *	Issues a reset command to FW.  For a Physical Function this would
+ *	result in the Firmware resetting all of its state.  For a Virtual
+ *	Function this just resets the state associated with the VF.
+ */
+int t4vf_fw_reset(struct adapter *adapter)
+{
+	struct fw_reset_cmd cmd;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_write = cpu_to_be32(FW_CMD_OP_V(FW_RESET_CMD) |
+				      FW_CMD_WRITE_F);
+	cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd));
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+}
+
+/**
+ *	t4vf_query_params - query FW or device parameters
+ *	@adapter: the adapter
+ *	@nparams: the number of parameters
+ *	@params: the parameter names
+ *	@vals: the parameter values
+ *
+ *	Reads the values of firmware or device parameters.  Up to 7 parameters
+ *	can be queried at once.
+ */
+static int t4vf_query_params(struct adapter *adapter, unsigned int nparams,
+			     const u32 *params, u32 *vals)
+{
+	int i, ret;
+	struct fw_params_cmd cmd, rpl;
+	struct fw_params_param *p;
+	size_t len16;
+
+	if (nparams > 7)
+		return -EINVAL;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_PARAMS_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_READ_F);
+	len16 = DIV_ROUND_UP(offsetof(struct fw_params_cmd,
+				      param[nparams].mnem), 16);
+	cmd.retval_len16 = cpu_to_be32(FW_CMD_LEN16_V(len16));
+	for (i = 0, p = &cmd.param[0]; i < nparams; i++, p++)
+		p->mnem = htonl(*params++);
+
+	ret = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (ret == 0)
+		for (i = 0, p = &rpl.param[0]; i < nparams; i++, p++)
+			*vals++ = be32_to_cpu(p->val);
+	return ret;
+}
+
+/**
+ *	t4vf_set_params - sets FW or device parameters
+ *	@adapter: the adapter
+ *	@nparams: the number of parameters
+ *	@params: the parameter names
+ *	@vals: the parameter values
+ *
+ *	Sets the values of firmware or device parameters.  Up to 7 parameters
+ *	can be specified at once.
+ */
+int t4vf_set_params(struct adapter *adapter, unsigned int nparams,
+		    const u32 *params, const u32 *vals)
+{
+	int i;
+	struct fw_params_cmd cmd;
+	struct fw_params_param *p;
+	size_t len16;
+
+	if (nparams > 7)
+		return -EINVAL;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_PARAMS_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_WRITE_F);
+	len16 = DIV_ROUND_UP(offsetof(struct fw_params_cmd,
+				      param[nparams]), 16);
+	cmd.retval_len16 = cpu_to_be32(FW_CMD_LEN16_V(len16));
+	for (i = 0, p = &cmd.param[0]; i < nparams; i++, p++) {
+		p->mnem = cpu_to_be32(*params++);
+		p->val = cpu_to_be32(*vals++);
+	}
+
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+}
+
+/**
+ *	t4_bar2_sge_qregs - return BAR2 SGE Queue register information
+ *	@adapter: the adapter
+ *	@qid: the Queue ID
+ *	@qtype: the Ingress or Egress type for @qid
+ *	@pbar2_qoffset: BAR2 Queue Offset
+ *	@pbar2_qid: BAR2 Queue ID or 0 for Queue ID inferred SGE Queues
+ *
+ *	Returns the BAR2 SGE Queue Registers information associated with the
+ *	indicated Absolute Queue ID.  These are passed back in return value
+ *	pointers.  @qtype should be T4_BAR2_QTYPE_EGRESS for Egress Queue
+ *	and T4_BAR2_QTYPE_INGRESS for Ingress Queues.
+ *
+ *	This may return an error which indicates that BAR2 SGE Queue
+ *	registers aren't available.  If an error is not returned, then the
+ *	following values are returned:
+ *
+ *	  *@pbar2_qoffset: the BAR2 Offset of the @qid Registers
+ *	  *@pbar2_qid: the BAR2 SGE Queue ID or 0 of @qid
+ *
+ *	If the returned BAR2 Queue ID is 0, then BAR2 SGE registers which
+ *	require the "Inferred Queue ID" ability may be used.  E.g. the
+ *	Write Combining Doorbell Buffer. If the BAR2 Queue ID is not 0,
+ *	then these "Inferred Queue ID" register may not be used.
+ */
+int t4_bar2_sge_qregs(struct adapter *adapter,
+		      unsigned int qid,
+		      enum t4_bar2_qtype qtype,
+		      u64 *pbar2_qoffset,
+		      unsigned int *pbar2_qid)
+{
+	unsigned int page_shift, page_size, qpp_shift, qpp_mask;
+	u64 bar2_page_offset, bar2_qoffset;
+	unsigned int bar2_qid, bar2_qid_offset, bar2_qinferred;
+
+	/* T4 doesn't support BAR2 SGE Queue registers.
+	 */
+	if (is_t4(adapter->params.chip))
+		return -EINVAL;
+
+	/* Get our SGE Page Size parameters.
+	 */
+	page_shift = adapter->params.sge.sge_vf_hps + 10;
+	page_size = 1 << page_shift;
+
+	/* Get the right Queues per Page parameters for our Queue.
+	 */
+	qpp_shift = (qtype == T4_BAR2_QTYPE_EGRESS
+		     ? adapter->params.sge.sge_vf_eq_qpp
+		     : adapter->params.sge.sge_vf_iq_qpp);
+	qpp_mask = (1 << qpp_shift) - 1;
+
+	/* Calculate the basics of the BAR2 SGE Queue register area:
+	 *  o The BAR2 page the Queue registers will be in.
+	 *  o The BAR2 Queue ID.
+	 *  o The BAR2 Queue ID Offset into the BAR2 page.
+	 */
+	bar2_page_offset = ((u64)(qid >> qpp_shift) << page_shift);
+	bar2_qid = qid & qpp_mask;
+	bar2_qid_offset = bar2_qid * SGE_UDB_SIZE;
+
+	/* If the BAR2 Queue ID Offset is less than the Page Size, then the
+	 * hardware will infer the Absolute Queue ID simply from the writes to
+	 * the BAR2 Queue ID Offset within the BAR2 Page (and we need to use a
+	 * BAR2 Queue ID of 0 for those writes).  Otherwise, we'll simply
+	 * write to the first BAR2 SGE Queue Area within the BAR2 Page with
+	 * the BAR2 Queue ID and the hardware will infer the Absolute Queue ID
+	 * from the BAR2 Page and BAR2 Queue ID.
+	 *
+	 * One important censequence of this is that some BAR2 SGE registers
+	 * have a "Queue ID" field and we can write the BAR2 SGE Queue ID
+	 * there.  But other registers synthesize the SGE Queue ID purely
+	 * from the writes to the registers -- the Write Combined Doorbell
+	 * Buffer is a good example.  These BAR2 SGE Registers are only
+	 * available for those BAR2 SGE Register areas where the SGE Absolute
+	 * Queue ID can be inferred from simple writes.
+	 */
+	bar2_qoffset = bar2_page_offset;
+	bar2_qinferred = (bar2_qid_offset < page_size);
+	if (bar2_qinferred) {
+		bar2_qoffset += bar2_qid_offset;
+		bar2_qid = 0;
+	}
+
+	*pbar2_qoffset = bar2_qoffset;
+	*pbar2_qid = bar2_qid;
+	return 0;
+}
+
+/**
+ *	t4vf_get_sge_params - retrieve adapter Scatter gather Engine parameters
+ *	@adapter: the adapter
+ *
+ *	Retrieves various core SGE parameters in the form of hardware SGE
+ *	register values.  The caller is responsible for decoding these as
+ *	needed.  The SGE parameters are stored in @adapter->params.sge.
+ */
+int t4vf_get_sge_params(struct adapter *adapter)
+{
+	struct sge_params *sge_params = &adapter->params.sge;
+	u32 params[7], vals[7];
+	int v;
+
+	params[0] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_CONTROL_A));
+	params[1] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_HOST_PAGE_SIZE_A));
+	params[2] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_FL_BUFFER_SIZE0_A));
+	params[3] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_FL_BUFFER_SIZE1_A));
+	params[4] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_TIMER_VALUE_0_AND_1_A));
+	params[5] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_TIMER_VALUE_2_AND_3_A));
+	params[6] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_TIMER_VALUE_4_AND_5_A));
+	v = t4vf_query_params(adapter, 7, params, vals);
+	if (v)
+		return v;
+	sge_params->sge_control = vals[0];
+	sge_params->sge_host_page_size = vals[1];
+	sge_params->sge_fl_buffer_size[0] = vals[2];
+	sge_params->sge_fl_buffer_size[1] = vals[3];
+	sge_params->sge_timer_value_0_and_1 = vals[4];
+	sge_params->sge_timer_value_2_and_3 = vals[5];
+	sge_params->sge_timer_value_4_and_5 = vals[6];
+
+	/* T4 uses a single control field to specify both the PCIe Padding and
+	 * Packing Boundary.  T5 introduced the ability to specify these
+	 * separately with the Padding Boundary in SGE_CONTROL and and Packing
+	 * Boundary in SGE_CONTROL2.  So for T5 and later we need to grab
+	 * SGE_CONTROL in order to determine how ingress packet data will be
+	 * laid out in Packed Buffer Mode.  Unfortunately, older versions of
+	 * the firmware won't let us retrieve SGE_CONTROL2 so if we get a
+	 * failure grabbing it we throw an error since we can't figure out the
+	 * right value.
+	 */
+	if (!is_t4(adapter->params.chip)) {
+		params[0] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+			     FW_PARAMS_PARAM_XYZ_V(SGE_CONTROL2_A));
+		v = t4vf_query_params(adapter, 1, params, vals);
+		if (v != FW_SUCCESS) {
+			dev_err(adapter->pdev_dev,
+				"Unable to get SGE Control2; "
+				"probably old firmware.\n");
+			return v;
+		}
+		sge_params->sge_control2 = vals[0];
+	}
+
+	params[0] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_INGRESS_RX_THRESHOLD_A));
+	params[1] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ_V(SGE_CONM_CTRL_A));
+	v = t4vf_query_params(adapter, 2, params, vals);
+	if (v)
+		return v;
+	sge_params->sge_ingress_rx_threshold = vals[0];
+	sge_params->sge_congestion_control = vals[1];
+
+	/* For T5 and later we want to use the new BAR2 Doorbells.
+	 * Unfortunately, older firmware didn't allow the this register to be
+	 * read.
+	 */
+	if (!is_t4(adapter->params.chip)) {
+		u32 whoami;
+		unsigned int pf, s_hps, s_qpp;
+
+		params[0] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+			     FW_PARAMS_PARAM_XYZ_V(
+				     SGE_EGRESS_QUEUES_PER_PAGE_VF_A));
+		params[1] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_REG) |
+			     FW_PARAMS_PARAM_XYZ_V(
+				     SGE_INGRESS_QUEUES_PER_PAGE_VF_A));
+		v = t4vf_query_params(adapter, 2, params, vals);
+		if (v != FW_SUCCESS) {
+			dev_warn(adapter->pdev_dev,
+				 "Unable to get VF SGE Queues/Page; "
+				 "probably old firmware.\n");
+			return v;
+		}
+		sge_params->sge_egress_queues_per_page = vals[0];
+		sge_params->sge_ingress_queues_per_page = vals[1];
+
+		/* We need the Queues/Page for our VF.  This is based on the
+		 * PF from which we're instantiated and is indexed in the
+		 * register we just read. Do it once here so other code in
+		 * the driver can just use it.
+		 */
+		whoami = t4_read_reg(adapter,
+				     T4VF_PL_BASE_ADDR + PL_VF_WHOAMI_A);
+		pf = SOURCEPF_G(whoami);
+
+		s_hps = (HOSTPAGESIZEPF0_S +
+			 (HOSTPAGESIZEPF1_S - HOSTPAGESIZEPF0_S) * pf);
+		sge_params->sge_vf_hps =
+			((sge_params->sge_host_page_size >> s_hps)
+			 & HOSTPAGESIZEPF0_M);
+
+		s_qpp = (QUEUESPERPAGEPF0_S +
+			 (QUEUESPERPAGEPF1_S - QUEUESPERPAGEPF0_S) * pf);
+		sge_params->sge_vf_eq_qpp =
+			((sge_params->sge_egress_queues_per_page >> s_qpp)
+			 & QUEUESPERPAGEPF0_M);
+		sge_params->sge_vf_iq_qpp =
+			((sge_params->sge_ingress_queues_per_page >> s_qpp)
+			 & QUEUESPERPAGEPF0_M);
+	}
+
+	return 0;
+}
+
+/**
+ *	t4vf_get_vpd_params - retrieve device VPD paremeters
+ *	@adapter: the adapter
+ *
+ *	Retrives various device Vital Product Data parameters.  The parameters
+ *	are stored in @adapter->params.vpd.
+ */
+int t4vf_get_vpd_params(struct adapter *adapter)
+{
+	struct vpd_params *vpd_params = &adapter->params.vpd;
+	u32 params[7], vals[7];
+	int v;
+
+	params[0] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_DEV) |
+		     FW_PARAMS_PARAM_X_V(FW_PARAMS_PARAM_DEV_CCLK));
+	v = t4vf_query_params(adapter, 1, params, vals);
+	if (v)
+		return v;
+	vpd_params->cclk = vals[0];
+
+	return 0;
+}
+
+/**
+ *	t4vf_get_dev_params - retrieve device paremeters
+ *	@adapter: the adapter
+ *
+ *	Retrives various device parameters.  The parameters are stored in
+ *	@adapter->params.dev.
+ */
+int t4vf_get_dev_params(struct adapter *adapter)
+{
+	struct dev_params *dev_params = &adapter->params.dev;
+	u32 params[7], vals[7];
+	int v;
+
+	params[0] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_DEV) |
+		     FW_PARAMS_PARAM_X_V(FW_PARAMS_PARAM_DEV_FWREV));
+	params[1] = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_DEV) |
+		     FW_PARAMS_PARAM_X_V(FW_PARAMS_PARAM_DEV_TPREV));
+	v = t4vf_query_params(adapter, 2, params, vals);
+	if (v)
+		return v;
+	dev_params->fwrev = vals[0];
+	dev_params->tprev = vals[1];
+
+	return 0;
+}
+
+/**
+ *	t4vf_get_rss_glb_config - retrieve adapter RSS Global Configuration
+ *	@adapter: the adapter
+ *
+ *	Retrieves global RSS mode and parameters with which we have to live
+ *	and stores them in the @adapter's RSS parameters.
+ */
+int t4vf_get_rss_glb_config(struct adapter *adapter)
+{
+	struct rss_params *rss = &adapter->params.rss;
+	struct fw_rss_glb_config_cmd cmd, rpl;
+	int v;
+
+	/*
+	 * Execute an RSS Global Configuration read command to retrieve
+	 * our RSS configuration.
+	 */
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_write = cpu_to_be32(FW_CMD_OP_V(FW_RSS_GLB_CONFIG_CMD) |
+				      FW_CMD_REQUEST_F |
+				      FW_CMD_READ_F);
+	cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd));
+	v = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (v)
+		return v;
+
+	/*
+	 * Transate the big-endian RSS Global Configuration into our
+	 * cpu-endian format based on the RSS mode.  We also do first level
+	 * filtering at this point to weed out modes which don't support
+	 * VF Drivers ...
+	 */
+	rss->mode = FW_RSS_GLB_CONFIG_CMD_MODE_G(
+			be32_to_cpu(rpl.u.manual.mode_pkd));
+	switch (rss->mode) {
+	case FW_RSS_GLB_CONFIG_CMD_MODE_BASICVIRTUAL: {
+		u32 word = be32_to_cpu(
+				rpl.u.basicvirtual.synmapen_to_hashtoeplitz);
+
+		rss->u.basicvirtual.synmapen =
+			((word & FW_RSS_GLB_CONFIG_CMD_SYNMAPEN_F) != 0);
+		rss->u.basicvirtual.syn4tupenipv6 =
+			((word & FW_RSS_GLB_CONFIG_CMD_SYN4TUPENIPV6_F) != 0);
+		rss->u.basicvirtual.syn2tupenipv6 =
+			((word & FW_RSS_GLB_CONFIG_CMD_SYN2TUPENIPV6_F) != 0);
+		rss->u.basicvirtual.syn4tupenipv4 =
+			((word & FW_RSS_GLB_CONFIG_CMD_SYN4TUPENIPV4_F) != 0);
+		rss->u.basicvirtual.syn2tupenipv4 =
+			((word & FW_RSS_GLB_CONFIG_CMD_SYN2TUPENIPV4_F) != 0);
+
+		rss->u.basicvirtual.ofdmapen =
+			((word & FW_RSS_GLB_CONFIG_CMD_OFDMAPEN_F) != 0);
+
+		rss->u.basicvirtual.tnlmapen =
+			((word & FW_RSS_GLB_CONFIG_CMD_TNLMAPEN_F) != 0);
+		rss->u.basicvirtual.tnlalllookup =
+			((word  & FW_RSS_GLB_CONFIG_CMD_TNLALLLKP_F) != 0);
+
+		rss->u.basicvirtual.hashtoeplitz =
+			((word & FW_RSS_GLB_CONFIG_CMD_HASHTOEPLITZ_F) != 0);
+
+		/* we need at least Tunnel Map Enable to be set */
+		if (!rss->u.basicvirtual.tnlmapen)
+			return -EINVAL;
+		break;
+	}
+
+	default:
+		/* all unknown/unsupported RSS modes result in an error */
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+/**
+ *	t4vf_get_vfres - retrieve VF resource limits
+ *	@adapter: the adapter
+ *
+ *	Retrieves configured resource limits and capabilities for a virtual
+ *	function.  The results are stored in @adapter->vfres.
+ */
+int t4vf_get_vfres(struct adapter *adapter)
+{
+	struct vf_resources *vfres = &adapter->params.vfres;
+	struct fw_pfvf_cmd cmd, rpl;
+	int v;
+	u32 word;
+
+	/*
+	 * Execute PFVF Read command to get VF resource limits; bail out early
+	 * with error on command failure.
+	 */
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_PFVF_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_READ_F);
+	cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd));
+	v = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (v)
+		return v;
+
+	/*
+	 * Extract VF resource limits and return success.
+	 */
+	word = be32_to_cpu(rpl.niqflint_niq);
+	vfres->niqflint = FW_PFVF_CMD_NIQFLINT_G(word);
+	vfres->niq = FW_PFVF_CMD_NIQ_G(word);
+
+	word = be32_to_cpu(rpl.type_to_neq);
+	vfres->neq = FW_PFVF_CMD_NEQ_G(word);
+	vfres->pmask = FW_PFVF_CMD_PMASK_G(word);
+
+	word = be32_to_cpu(rpl.tc_to_nexactf);
+	vfres->tc = FW_PFVF_CMD_TC_G(word);
+	vfres->nvi = FW_PFVF_CMD_NVI_G(word);
+	vfres->nexactf = FW_PFVF_CMD_NEXACTF_G(word);
+
+	word = be32_to_cpu(rpl.r_caps_to_nethctrl);
+	vfres->r_caps = FW_PFVF_CMD_R_CAPS_G(word);
+	vfres->wx_caps = FW_PFVF_CMD_WX_CAPS_G(word);
+	vfres->nethctrl = FW_PFVF_CMD_NETHCTRL_G(word);
+
+	return 0;
+}
+
+/**
+ *	t4vf_read_rss_vi_config - read a VI's RSS configuration
+ *	@adapter: the adapter
+ *	@viid: Virtual Interface ID
+ *	@config: pointer to host-native VI RSS Configuration buffer
+ *
+ *	Reads the Virtual Interface's RSS configuration information and
+ *	translates it into CPU-native format.
+ */
+int t4vf_read_rss_vi_config(struct adapter *adapter, unsigned int viid,
+			    union rss_vi_config *config)
+{
+	struct fw_rss_vi_config_cmd cmd, rpl;
+	int v;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_RSS_VI_CONFIG_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_READ_F |
+				     FW_RSS_VI_CONFIG_CMD_VIID(viid));
+	cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd));
+	v = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (v)
+		return v;
+
+	switch (adapter->params.rss.mode) {
+	case FW_RSS_GLB_CONFIG_CMD_MODE_BASICVIRTUAL: {
+		u32 word = be32_to_cpu(rpl.u.basicvirtual.defaultq_to_udpen);
+
+		config->basicvirtual.ip6fourtupen =
+			((word & FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN_F) != 0);
+		config->basicvirtual.ip6twotupen =
+			((word & FW_RSS_VI_CONFIG_CMD_IP6TWOTUPEN_F) != 0);
+		config->basicvirtual.ip4fourtupen =
+			((word & FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN_F) != 0);
+		config->basicvirtual.ip4twotupen =
+			((word & FW_RSS_VI_CONFIG_CMD_IP4TWOTUPEN_F) != 0);
+		config->basicvirtual.udpen =
+			((word & FW_RSS_VI_CONFIG_CMD_UDPEN_F) != 0);
+		config->basicvirtual.defaultq =
+			FW_RSS_VI_CONFIG_CMD_DEFAULTQ_G(word);
+		break;
+	}
+
+	default:
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+/**
+ *	t4vf_write_rss_vi_config - write a VI's RSS configuration
+ *	@adapter: the adapter
+ *	@viid: Virtual Interface ID
+ *	@config: pointer to host-native VI RSS Configuration buffer
+ *
+ *	Write the Virtual Interface's RSS configuration information
+ *	(translating it into firmware-native format before writing).
+ */
+int t4vf_write_rss_vi_config(struct adapter *adapter, unsigned int viid,
+			     union rss_vi_config *config)
+{
+	struct fw_rss_vi_config_cmd cmd, rpl;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_RSS_VI_CONFIG_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_WRITE_F |
+				     FW_RSS_VI_CONFIG_CMD_VIID(viid));
+	cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd));
+	switch (adapter->params.rss.mode) {
+	case FW_RSS_GLB_CONFIG_CMD_MODE_BASICVIRTUAL: {
+		u32 word = 0;
+
+		if (config->basicvirtual.ip6fourtupen)
+			word |= FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN_F;
+		if (config->basicvirtual.ip6twotupen)
+			word |= FW_RSS_VI_CONFIG_CMD_IP6TWOTUPEN_F;
+		if (config->basicvirtual.ip4fourtupen)
+			word |= FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN_F;
+		if (config->basicvirtual.ip4twotupen)
+			word |= FW_RSS_VI_CONFIG_CMD_IP4TWOTUPEN_F;
+		if (config->basicvirtual.udpen)
+			word |= FW_RSS_VI_CONFIG_CMD_UDPEN_F;
+		word |= FW_RSS_VI_CONFIG_CMD_DEFAULTQ_V(
+				config->basicvirtual.defaultq);
+		cmd.u.basicvirtual.defaultq_to_udpen = cpu_to_be32(word);
+		break;
+	}
+
+	default:
+		return -EINVAL;
+	}
+
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+}
+
+/**
+ *	t4vf_config_rss_range - configure a portion of the RSS mapping table
+ *	@adapter: the adapter
+ *	@viid: Virtual Interface of RSS Table Slice
+ *	@start: starting entry in the table to write
+ *	@n: how many table entries to write
+ *	@rspq: values for the "Response Queue" (Ingress Queue) lookup table
+ *	@nrspq: number of values in @rspq
+ *
+ *	Programs the selected part of the VI's RSS mapping table with the
+ *	provided values.  If @nrspq < @n the supplied values are used repeatedly
+ *	until the full table range is populated.
+ *
+ *	The caller must ensure the values in @rspq are in the range 0..1023.
+ */
+int t4vf_config_rss_range(struct adapter *adapter, unsigned int viid,
+			  int start, int n, const u16 *rspq, int nrspq)
+{
+	const u16 *rsp = rspq;
+	const u16 *rsp_end = rspq+nrspq;
+	struct fw_rss_ind_tbl_cmd cmd;
+
+	/*
+	 * Initialize firmware command template to write the RSS table.
+	 */
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_RSS_IND_TBL_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_WRITE_F |
+				     FW_RSS_IND_TBL_CMD_VIID_V(viid));
+	cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd));
+
+	/*
+	 * Each firmware RSS command can accommodate up to 32 RSS Ingress
+	 * Queue Identifiers.  These Ingress Queue IDs are packed three to
+	 * a 32-bit word as 10-bit values with the upper remaining 2 bits
+	 * reserved.
+	 */
+	while (n > 0) {
+		__be32 *qp = &cmd.iq0_to_iq2;
+		int nq = min(n, 32);
+		int ret;
+
+		/*
+		 * Set up the firmware RSS command header to send the next
+		 * "nq" Ingress Queue IDs to the firmware.
+		 */
+		cmd.niqid = cpu_to_be16(nq);
+		cmd.startidx = cpu_to_be16(start);
+
+		/*
+		 * "nq" more done for the start of the next loop.
+		 */
+		start += nq;
+		n -= nq;
+
+		/*
+		 * While there are still Ingress Queue IDs to stuff into the
+		 * current firmware RSS command, retrieve them from the
+		 * Ingress Queue ID array and insert them into the command.
+		 */
+		while (nq > 0) {
+			/*
+			 * Grab up to the next 3 Ingress Queue IDs (wrapping
+			 * around the Ingress Queue ID array if necessary) and
+			 * insert them into the firmware RSS command at the
+			 * current 3-tuple position within the commad.
+			 */
+			u16 qbuf[3];
+			u16 *qbp = qbuf;
+			int nqbuf = min(3, nq);
+
+			nq -= nqbuf;
+			qbuf[0] = qbuf[1] = qbuf[2] = 0;
+			while (nqbuf) {
+				nqbuf--;
+				*qbp++ = *rsp++;
+				if (rsp >= rsp_end)
+					rsp = rspq;
+			}
+			*qp++ = cpu_to_be32(FW_RSS_IND_TBL_CMD_IQ0_V(qbuf[0]) |
+					    FW_RSS_IND_TBL_CMD_IQ1_V(qbuf[1]) |
+					    FW_RSS_IND_TBL_CMD_IQ2_V(qbuf[2]));
+		}
+
+		/*
+		 * Send this portion of the RRS table update to the firmware;
+		 * bail out on any errors.
+		 */
+		ret = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+		if (ret)
+			return ret;
+	}
+	return 0;
+}
+
+/**
+ *	t4vf_alloc_vi - allocate a virtual interface on a port
+ *	@adapter: the adapter
+ *	@port_id: physical port associated with the VI
+ *
+ *	Allocate a new Virtual Interface and bind it to the indicated
+ *	physical port.  Return the new Virtual Interface Identifier on
+ *	success, or a [negative] error number on failure.
+ */
+int t4vf_alloc_vi(struct adapter *adapter, int port_id)
+{
+	struct fw_vi_cmd cmd, rpl;
+	int v;
+
+	/*
+	 * Execute a VI command to allocate Virtual Interface and return its
+	 * VIID.
+	 */
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_VI_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_WRITE_F |
+				    FW_CMD_EXEC_F);
+	cmd.alloc_to_len16 = cpu_to_be32(FW_LEN16(cmd) |
+					 FW_VI_CMD_ALLOC_F);
+	cmd.portid_pkd = FW_VI_CMD_PORTID_V(port_id);
+	v = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (v)
+		return v;
+
+	return FW_VI_CMD_VIID_G(be16_to_cpu(rpl.type_viid));
+}
+
+/**
+ *	t4vf_free_vi -- free a virtual interface
+ *	@adapter: the adapter
+ *	@viid: the virtual interface identifier
+ *
+ *	Free a previously allocated Virtual Interface.  Return an error on
+ *	failure.
+ */
+int t4vf_free_vi(struct adapter *adapter, int viid)
+{
+	struct fw_vi_cmd cmd;
+
+	/*
+	 * Execute a VI command to free the Virtual Interface.
+	 */
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_VI_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_EXEC_F);
+	cmd.alloc_to_len16 = cpu_to_be32(FW_LEN16(cmd) |
+					 FW_VI_CMD_FREE_F);
+	cmd.type_viid = cpu_to_be16(FW_VI_CMD_VIID_V(viid));
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+}
+
+/**
+ *	t4vf_enable_vi - enable/disable a virtual interface
+ *	@adapter: the adapter
+ *	@viid: the Virtual Interface ID
+ *	@rx_en: 1=enable Rx, 0=disable Rx
+ *	@tx_en: 1=enable Tx, 0=disable Tx
+ *
+ *	Enables/disables a virtual interface.
+ */
+int t4vf_enable_vi(struct adapter *adapter, unsigned int viid,
+		   bool rx_en, bool tx_en)
+{
+	struct fw_vi_enable_cmd cmd;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_VI_ENABLE_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_EXEC_F |
+				     FW_VI_ENABLE_CMD_VIID_V(viid));
+	cmd.ien_to_len16 = cpu_to_be32(FW_VI_ENABLE_CMD_IEN_V(rx_en) |
+				       FW_VI_ENABLE_CMD_EEN_V(tx_en) |
+				       FW_LEN16(cmd));
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+}
+
+/**
+ *	t4vf_identify_port - identify a VI's port by blinking its LED
+ *	@adapter: the adapter
+ *	@viid: the Virtual Interface ID
+ *	@nblinks: how many times to blink LED at 2.5 Hz
+ *
+ *	Identifies a VI's port by blinking its LED.
+ */
+int t4vf_identify_port(struct adapter *adapter, unsigned int viid,
+		       unsigned int nblinks)
+{
+	struct fw_vi_enable_cmd cmd;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_VI_ENABLE_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_EXEC_F |
+				     FW_VI_ENABLE_CMD_VIID_V(viid));
+	cmd.ien_to_len16 = cpu_to_be32(FW_VI_ENABLE_CMD_LED_F |
+				       FW_LEN16(cmd));
+	cmd.blinkdur = cpu_to_be16(nblinks);
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+}
+
+/**
+ *	t4vf_set_rxmode - set Rx properties of a virtual interface
+ *	@adapter: the adapter
+ *	@viid: the VI id
+ *	@mtu: the new MTU or -1 for no change
+ *	@promisc: 1 to enable promiscuous mode, 0 to disable it, -1 no change
+ *	@all_multi: 1 to enable all-multi mode, 0 to disable it, -1 no change
+ *	@bcast: 1 to enable broadcast Rx, 0 to disable it, -1 no change
+ *	@vlanex: 1 to enable hardware VLAN Tag extraction, 0 to disable it,
+ *		-1 no change
+ *
+ *	Sets Rx properties of a virtual interface.
+ */
+int t4vf_set_rxmode(struct adapter *adapter, unsigned int viid,
+		    int mtu, int promisc, int all_multi, int bcast, int vlanex,
+		    bool sleep_ok)
+{
+	struct fw_vi_rxmode_cmd cmd;
+
+	/* convert to FW values */
+	if (mtu < 0)
+		mtu = FW_VI_RXMODE_CMD_MTU_M;
+	if (promisc < 0)
+		promisc = FW_VI_RXMODE_CMD_PROMISCEN_M;
+	if (all_multi < 0)
+		all_multi = FW_VI_RXMODE_CMD_ALLMULTIEN_M;
+	if (bcast < 0)
+		bcast = FW_VI_RXMODE_CMD_BROADCASTEN_M;
+	if (vlanex < 0)
+		vlanex = FW_VI_RXMODE_CMD_VLANEXEN_M;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_VI_RXMODE_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_WRITE_F |
+				     FW_VI_RXMODE_CMD_VIID_V(viid));
+	cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd));
+	cmd.mtu_to_vlanexen =
+		cpu_to_be32(FW_VI_RXMODE_CMD_MTU_V(mtu) |
+			    FW_VI_RXMODE_CMD_PROMISCEN_V(promisc) |
+			    FW_VI_RXMODE_CMD_ALLMULTIEN_V(all_multi) |
+			    FW_VI_RXMODE_CMD_BROADCASTEN_V(bcast) |
+			    FW_VI_RXMODE_CMD_VLANEXEN_V(vlanex));
+	return t4vf_wr_mbox_core(adapter, &cmd, sizeof(cmd), NULL, sleep_ok);
+}
+
+/**
+ *	t4vf_alloc_mac_filt - allocates exact-match filters for MAC addresses
+ *	@adapter: the adapter
+ *	@viid: the Virtual Interface Identifier
+ *	@free: if true any existing filters for this VI id are first removed
+ *	@naddr: the number of MAC addresses to allocate filters for (up to 7)
+ *	@addr: the MAC address(es)
+ *	@idx: where to store the index of each allocated filter
+ *	@hash: pointer to hash address filter bitmap
+ *	@sleep_ok: call is allowed to sleep
+ *
+ *	Allocates an exact-match filter for each of the supplied addresses and
+ *	sets it to the corresponding address.  If @idx is not %NULL it should
+ *	have at least @naddr entries, each of which will be set to the index of
+ *	the filter allocated for the corresponding MAC address.  If a filter
+ *	could not be allocated for an address its index is set to 0xffff.
+ *	If @hash is not %NULL addresses that fail to allocate an exact filter
+ *	are hashed and update the hash filter bitmap pointed at by @hash.
+ *
+ *	Returns a negative error number or the number of filters allocated.
+ */
+int t4vf_alloc_mac_filt(struct adapter *adapter, unsigned int viid, bool free,
+			unsigned int naddr, const u8 **addr, u16 *idx,
+			u64 *hash, bool sleep_ok)
+{
+	int offset, ret = 0;
+	unsigned nfilters = 0;
+	unsigned int rem = naddr;
+	struct fw_vi_mac_cmd cmd, rpl;
+	unsigned int max_naddr = is_t4(adapter->params.chip) ?
+				 NUM_MPS_CLS_SRAM_L_INSTANCES :
+				 NUM_MPS_T5_CLS_SRAM_L_INSTANCES;
+
+	if (naddr > max_naddr)
+		return -EINVAL;
+
+	for (offset = 0; offset < naddr; /**/) {
+		unsigned int fw_naddr = (rem < ARRAY_SIZE(cmd.u.exact)
+					 ? rem
+					 : ARRAY_SIZE(cmd.u.exact));
+		size_t len16 = DIV_ROUND_UP(offsetof(struct fw_vi_mac_cmd,
+						     u.exact[fw_naddr]), 16);
+		struct fw_vi_mac_exact *p;
+		int i;
+
+		memset(&cmd, 0, sizeof(cmd));
+		cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_VI_MAC_CMD) |
+					     FW_CMD_REQUEST_F |
+					     FW_CMD_WRITE_F |
+					     (free ? FW_CMD_EXEC_F : 0) |
+					     FW_VI_MAC_CMD_VIID_V(viid));
+		cmd.freemacs_to_len16 =
+			cpu_to_be32(FW_VI_MAC_CMD_FREEMACS_V(free) |
+				    FW_CMD_LEN16_V(len16));
+
+		for (i = 0, p = cmd.u.exact; i < fw_naddr; i++, p++) {
+			p->valid_to_idx = cpu_to_be16(
+				FW_VI_MAC_CMD_VALID_F |
+				FW_VI_MAC_CMD_IDX_V(FW_VI_MAC_ADD_MAC));
+			memcpy(p->macaddr, addr[offset+i], sizeof(p->macaddr));
+		}
+
+
+		ret = t4vf_wr_mbox_core(adapter, &cmd, sizeof(cmd), &rpl,
+					sleep_ok);
+		if (ret && ret != -ENOMEM)
+			break;
+
+		for (i = 0, p = rpl.u.exact; i < fw_naddr; i++, p++) {
+			u16 index = FW_VI_MAC_CMD_IDX_G(
+				be16_to_cpu(p->valid_to_idx));
+
+			if (idx)
+				idx[offset+i] =
+					(index >= max_naddr
+					 ? 0xffff
+					 : index);
+			if (index < max_naddr)
+				nfilters++;
+			else if (hash)
+				*hash |= (1ULL << hash_mac_addr(addr[offset+i]));
+		}
+
+		free = false;
+		offset += fw_naddr;
+		rem -= fw_naddr;
+	}
+
+	/*
+	 * If there were no errors or we merely ran out of room in our MAC
+	 * address arena, return the number of filters actually written.
+	 */
+	if (ret == 0 || ret == -ENOMEM)
+		ret = nfilters;
+	return ret;
+}
+
+/**
+ *	t4vf_change_mac - modifies the exact-match filter for a MAC address
+ *	@adapter: the adapter
+ *	@viid: the Virtual Interface ID
+ *	@idx: index of existing filter for old value of MAC address, or -1
+ *	@addr: the new MAC address value
+ *	@persist: if idx < 0, the new MAC allocation should be persistent
+ *
+ *	Modifies an exact-match filter and sets it to the new MAC address.
+ *	Note that in general it is not possible to modify the value of a given
+ *	filter so the generic way to modify an address filter is to free the
+ *	one being used by the old address value and allocate a new filter for
+ *	the new address value.  @idx can be -1 if the address is a new
+ *	addition.
+ *
+ *	Returns a negative error number or the index of the filter with the new
+ *	MAC value.
+ */
+int t4vf_change_mac(struct adapter *adapter, unsigned int viid,
+		    int idx, const u8 *addr, bool persist)
+{
+	int ret;
+	struct fw_vi_mac_cmd cmd, rpl;
+	struct fw_vi_mac_exact *p = &cmd.u.exact[0];
+	size_t len16 = DIV_ROUND_UP(offsetof(struct fw_vi_mac_cmd,
+					     u.exact[1]), 16);
+	unsigned int max_naddr = is_t4(adapter->params.chip) ?
+				 NUM_MPS_CLS_SRAM_L_INSTANCES :
+				 NUM_MPS_T5_CLS_SRAM_L_INSTANCES;
+
+	/*
+	 * If this is a new allocation, determine whether it should be
+	 * persistent (across a "freemacs" operation) or not.
+	 */
+	if (idx < 0)
+		idx = persist ? FW_VI_MAC_ADD_PERSIST_MAC : FW_VI_MAC_ADD_MAC;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_VI_MAC_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_WRITE_F |
+				     FW_VI_MAC_CMD_VIID_V(viid));
+	cmd.freemacs_to_len16 = cpu_to_be32(FW_CMD_LEN16_V(len16));
+	p->valid_to_idx = cpu_to_be16(FW_VI_MAC_CMD_VALID_F |
+				      FW_VI_MAC_CMD_IDX_V(idx));
+	memcpy(p->macaddr, addr, sizeof(p->macaddr));
+
+	ret = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
+	if (ret == 0) {
+		p = &rpl.u.exact[0];
+		ret = FW_VI_MAC_CMD_IDX_G(be16_to_cpu(p->valid_to_idx));
+		if (ret >= max_naddr)
+			ret = -ENOMEM;
+	}
+	return ret;
+}
+
+/**
+ *	t4vf_set_addr_hash - program the MAC inexact-match hash filter
+ *	@adapter: the adapter
+ *	@viid: the Virtual Interface Identifier
+ *	@ucast: whether the hash filter should also match unicast addresses
+ *	@vec: the value to be written to the hash filter
+ *	@sleep_ok: call is allowed to sleep
+ *
+ *	Sets the 64-bit inexact-match hash filter for a virtual interface.
+ */
+int t4vf_set_addr_hash(struct adapter *adapter, unsigned int viid,
+		       bool ucast, u64 vec, bool sleep_ok)
+{
+	struct fw_vi_mac_cmd cmd;
+	size_t len16 = DIV_ROUND_UP(offsetof(struct fw_vi_mac_cmd,
+					     u.exact[0]), 16);
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_VI_MAC_CMD) |
+				     FW_CMD_REQUEST_F |
+				     FW_CMD_WRITE_F |
+				     FW_VI_ENABLE_CMD_VIID_V(viid));
+	cmd.freemacs_to_len16 = cpu_to_be32(FW_VI_MAC_CMD_HASHVECEN_F |
+					    FW_VI_MAC_CMD_HASHUNIEN_V(ucast) |
+					    FW_CMD_LEN16_V(len16));
+	cmd.u.hash.hashvec = cpu_to_be64(vec);
+	return t4vf_wr_mbox_core(adapter, &cmd, sizeof(cmd), NULL, sleep_ok);
+}
+
+/**
+ *	t4vf_get_port_stats - collect "port" statistics
+ *	@adapter: the adapter
+ *	@pidx: the port index
+ *	@s: the stats structure to fill
+ *
+ *	Collect statistics for the "port"'s Virtual Interface.
+ */
+int t4vf_get_port_stats(struct adapter *adapter, int pidx,
+			struct t4vf_port_stats *s)
+{
+	struct port_info *pi = adap2pinfo(adapter, pidx);
+	struct fw_vi_stats_vf fwstats;
+	unsigned int rem = VI_VF_NUM_STATS;
+	__be64 *fwsp = (__be64 *)&fwstats;
+
+	/*
+	 * Grab the Virtual Interface statistics a chunk at a time via mailbox
+	 * commands.  We could use a Work Request and get all of them at once
+	 * but that's an asynchronous interface which is awkward to use.
+	 */
+	while (rem) {
+		unsigned int ix = VI_VF_NUM_STATS - rem;
+		unsigned int nstats = min(6U, rem);
+		struct fw_vi_stats_cmd cmd, rpl;
+		size_t len = (offsetof(struct fw_vi_stats_cmd, u) +
+			      sizeof(struct fw_vi_stats_ctl));
+		size_t len16 = DIV_ROUND_UP(len, 16);
+		int ret;
+
+		memset(&cmd, 0, sizeof(cmd));
+		cmd.op_to_viid = cpu_to_be32(FW_CMD_OP_V(FW_VI_STATS_CMD) |
+					     FW_VI_STATS_CMD_VIID_V(pi->viid) |
+					     FW_CMD_REQUEST_F |
+					     FW_CMD_READ_F);
+		cmd.retval_len16 = cpu_to_be32(FW_CMD_LEN16_V(len16));
+		cmd.u.ctl.nstats_ix =
+			cpu_to_be16(FW_VI_STATS_CMD_IX_V(ix) |
+				    FW_VI_STATS_CMD_NSTATS_V(nstats));
+		ret = t4vf_wr_mbox_ns(adapter, &cmd, len, &rpl);
+		if (ret)
+			return ret;
+
+		memcpy(fwsp, &rpl.u.ctl.stat0, sizeof(__be64) * nstats);
+
+		rem -= nstats;
+		fwsp += nstats;
+	}
+
+	/*
+	 * Translate firmware statistics into host native statistics.
+	 */
+	s->tx_bcast_bytes = be64_to_cpu(fwstats.tx_bcast_bytes);
+	s->tx_bcast_frames = be64_to_cpu(fwstats.tx_bcast_frames);
+	s->tx_mcast_bytes = be64_to_cpu(fwstats.tx_mcast_bytes);
+	s->tx_mcast_frames = be64_to_cpu(fwstats.tx_mcast_frames);
+	s->tx_ucast_bytes = be64_to_cpu(fwstats.tx_ucast_bytes);
+	s->tx_ucast_frames = be64_to_cpu(fwstats.tx_ucast_frames);
+	s->tx_drop_frames = be64_to_cpu(fwstats.tx_drop_frames);
+	s->tx_offload_bytes = be64_to_cpu(fwstats.tx_offload_bytes);
+	s->tx_offload_frames = be64_to_cpu(fwstats.tx_offload_frames);
+
+	s->rx_bcast_bytes = be64_to_cpu(fwstats.rx_bcast_bytes);
+	s->rx_bcast_frames = be64_to_cpu(fwstats.rx_bcast_frames);
+	s->rx_mcast_bytes = be64_to_cpu(fwstats.rx_mcast_bytes);
+	s->rx_mcast_frames = be64_to_cpu(fwstats.rx_mcast_frames);
+	s->rx_ucast_bytes = be64_to_cpu(fwstats.rx_ucast_bytes);
+	s->rx_ucast_frames = be64_to_cpu(fwstats.rx_ucast_frames);
+
+	s->rx_err_frames = be64_to_cpu(fwstats.rx_err_frames);
+
+	return 0;
+}
+
+/**
+ *	t4vf_iq_free - free an ingress queue and its free lists
+ *	@adapter: the adapter
+ *	@iqtype: the ingress queue type (FW_IQ_TYPE_FL_INT_CAP, etc.)
+ *	@iqid: ingress queue ID
+ *	@fl0id: FL0 queue ID or 0xffff if no attached FL0
+ *	@fl1id: FL1 queue ID or 0xffff if no attached FL1
+ *
+ *	Frees an ingress queue and its associated free lists, if any.
+ */
+int t4vf_iq_free(struct adapter *adapter, unsigned int iqtype,
+		 unsigned int iqid, unsigned int fl0id, unsigned int fl1id)
+{
+	struct fw_iq_cmd cmd;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_IQ_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_EXEC_F);
+	cmd.alloc_to_len16 = cpu_to_be32(FW_IQ_CMD_FREE_F |
+					 FW_LEN16(cmd));
+	cmd.type_to_iqandstindex =
+		cpu_to_be32(FW_IQ_CMD_TYPE_V(iqtype));
+
+	cmd.iqid = cpu_to_be16(iqid);
+	cmd.fl0id = cpu_to_be16(fl0id);
+	cmd.fl1id = cpu_to_be16(fl1id);
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+}
+
+/**
+ *	t4vf_eth_eq_free - free an Ethernet egress queue
+ *	@adapter: the adapter
+ *	@eqid: egress queue ID
+ *
+ *	Frees an Ethernet egress queue.
+ */
+int t4vf_eth_eq_free(struct adapter *adapter, unsigned int eqid)
+{
+	struct fw_eq_eth_cmd cmd;
+
+	memset(&cmd, 0, sizeof(cmd));
+	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_EQ_ETH_CMD) |
+				    FW_CMD_REQUEST_F |
+				    FW_CMD_EXEC_F);
+	cmd.alloc_to_len16 = cpu_to_be32(FW_EQ_ETH_CMD_FREE_F |
+					 FW_LEN16(cmd));
+	cmd.eqid_pkd = cpu_to_be32(FW_EQ_ETH_CMD_EQID_V(eqid));
+	return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL);
+}
+
+/**
+ *	t4vf_handle_fw_rpl - process a firmware reply message
+ *	@adapter: the adapter
+ *	@rpl: start of the firmware message
+ *
+ *	Processes a firmware message, such as link state change messages.
+ */
+int t4vf_handle_fw_rpl(struct adapter *adapter, const __be64 *rpl)
+{
+	const struct fw_cmd_hdr *cmd_hdr = (const struct fw_cmd_hdr *)rpl;
+	u8 opcode = FW_CMD_OP_G(be32_to_cpu(cmd_hdr->hi));
+
+	switch (opcode) {
+	case FW_PORT_CMD: {
+		/*
+		 * Link/module state change message.
+		 */
+		const struct fw_port_cmd *port_cmd =
+			(const struct fw_port_cmd *)rpl;
+		u32 stat, mod;
+		int action, port_id, link_ok, speed, fc, pidx;
+
+		/*
+		 * Extract various fields from port status change message.
+		 */
+		action = FW_PORT_CMD_ACTION_G(
+			be32_to_cpu(port_cmd->action_to_len16));
+		if (action != FW_PORT_ACTION_GET_PORT_INFO) {
+			dev_err(adapter->pdev_dev,
+				"Unknown firmware PORT reply action %x\n",
+				action);
+			break;
+		}
+
+		port_id = FW_PORT_CMD_PORTID_G(
+			be32_to_cpu(port_cmd->op_to_portid));
+
+		stat = be32_to_cpu(port_cmd->u.info.lstatus_to_modtype);
+		link_ok = (stat & FW_PORT_CMD_LSTATUS_F) != 0;
+		speed = 0;
+		fc = 0;
+		if (stat & FW_PORT_CMD_RXPAUSE_F)
+			fc |= PAUSE_RX;
+		if (stat & FW_PORT_CMD_TXPAUSE_F)
+			fc |= PAUSE_TX;
+		if (stat & FW_PORT_CMD_LSPEED_V(FW_PORT_CAP_SPEED_100M))
+			speed = 100;
+		else if (stat & FW_PORT_CMD_LSPEED_V(FW_PORT_CAP_SPEED_1G))
+			speed = 1000;
+		else if (stat & FW_PORT_CMD_LSPEED_V(FW_PORT_CAP_SPEED_10G))
+			speed = 10000;
+		else if (stat & FW_PORT_CMD_LSPEED_V(FW_PORT_CAP_SPEED_40G))
+			speed = 40000;
+
+		/*
+		 * Scan all of our "ports" (Virtual Interfaces) looking for
+		 * those bound to the physical port which has changed.  If
+		 * our recorded state doesn't match the current state,
+		 * signal that change to the OS code.
+		 */
+		for_each_port(adapter, pidx) {
+			struct port_info *pi = adap2pinfo(adapter, pidx);
+			struct link_config *lc;
+
+			if (pi->port_id != port_id)
+				continue;
+
+			lc = &pi->link_cfg;
+
+			mod = FW_PORT_CMD_MODTYPE_G(stat);
+			if (mod != pi->mod_type) {
+				pi->mod_type = mod;
+				t4vf_os_portmod_changed(adapter, pidx);
+			}
+
+			if (link_ok != lc->link_ok || speed != lc->speed ||
+			    fc != lc->fc) {
+				/* something changed */
+				lc->link_ok = link_ok;
+				lc->speed = speed;
+				lc->fc = fc;
+				lc->supported =
+					be16_to_cpu(port_cmd->u.info.pcap);
+				t4vf_os_link_changed(adapter, pidx, link_ok);
+			}
+		}
+		break;
+	}
+
+	default:
+		dev_err(adapter->pdev_dev, "Unknown firmware reply %X\n",
+			opcode);
+	}
+	return 0;
+}
+
+/**
+ */
+int t4vf_prep_adapter(struct adapter *adapter)
+{
+	int err;
+	unsigned int chipid;
+
+	/* Wait for the device to become ready before proceeding ...
+	 */
+	err = t4vf_wait_dev_ready(adapter);
+	if (err)
+		return err;
+
+	/* Default port and clock for debugging in case we can't reach
+	 * firmware.
+	 */
+	adapter->params.nports = 1;
+	adapter->params.vfres.pmask = 1;
+	adapter->params.vpd.cclk = 50000;
+
+	adapter->params.chip = 0;
+	switch (CHELSIO_PCI_ID_VER(adapter->pdev->device)) {
+	case CHELSIO_T4:
+		adapter->params.chip |= CHELSIO_CHIP_CODE(CHELSIO_T4, 0);
+		break;
+
+	case CHELSIO_T5:
+		chipid = REV_G(t4_read_reg(adapter, PL_VF_REV_A));
+		adapter->params.chip |= CHELSIO_CHIP_CODE(CHELSIO_T5, chipid);
+		break;
+	}
+
+	return 0;
+}