Imported Upstream version 16.04

Change-Id: I77eadcd8538a9122e4773cbe55b24033dc451757
Signed-off-by: C.J. Collier <cjcollier@linuxfoundation.org>

1 files changed, 2255 insertions, 0 deletions

diff --git a/drivers/net/cxgbe/sge.c b/drivers/net/cxgbe/sge.c
new file mode 100644
index 00000000..ab5a842a
--- /dev/null
+++ b/drivers/net/cxgbe/sge.c
@@ -0,0 +1,2255 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2014-2015 Chelsio Communications.
+ *   All rights reserved.
+ *
+ *   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.
+ *     * Neither the name of Chelsio Communications nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <sys/queue.h>
+#include <stdio.h>
+#include <errno.h>
+#include <stdint.h>
+#include <string.h>
+#include <unistd.h>
+#include <stdarg.h>
+#include <inttypes.h>
+#include <netinet/in.h>
+
+#include <rte_byteorder.h>
+#include <rte_common.h>
+#include <rte_cycles.h>
+#include <rte_interrupts.h>
+#include <rte_log.h>
+#include <rte_debug.h>
+#include <rte_pci.h>
+#include <rte_atomic.h>
+#include <rte_branch_prediction.h>
+#include <rte_memory.h>
+#include <rte_memzone.h>
+#include <rte_tailq.h>
+#include <rte_eal.h>
+#include <rte_alarm.h>
+#include <rte_ether.h>
+#include <rte_ethdev.h>
+#include <rte_atomic.h>
+#include <rte_malloc.h>
+#include <rte_random.h>
+#include <rte_dev.h>
+
+#include "common.h"
+#include "t4_regs.h"
+#include "t4_msg.h"
+#include "cxgbe.h"
+
+static inline void ship_tx_pkt_coalesce_wr(struct adapter *adap,
+					   struct sge_eth_txq *txq);
+
+/*
+ * Max number of Rx buffers we replenish at a time.
+ */
+#define MAX_RX_REFILL 64U
+
+#define NOMEM_TMR_IDX (SGE_NTIMERS - 1)
+
+/*
+ * Max Tx descriptor space we allow for an Ethernet packet to be inlined
+ * into a WR.
+ */
+#define MAX_IMM_TX_PKT_LEN 256
+
+/*
+ * Rx buffer sizes for "usembufs" Free List buffers (one ingress packet
+ * per mbuf buffer).  We currently only support two sizes for 1500- and
+ * 9000-byte MTUs. We could easily support more but there doesn't seem to be
+ * much need for that ...
+ */
+#define FL_MTU_SMALL 1500
+#define FL_MTU_LARGE 9000
+
+static inline unsigned int fl_mtu_bufsize(struct adapter *adapter,
+					  unsigned int mtu)
+{
+	struct sge *s = &adapter->sge;
+
+	return CXGBE_ALIGN(s->pktshift + ETHER_HDR_LEN + VLAN_HLEN + mtu,
+			   s->fl_align);
+}
+
+#define FL_MTU_SMALL_BUFSIZE(adapter) fl_mtu_bufsize(adapter, FL_MTU_SMALL)
+#define FL_MTU_LARGE_BUFSIZE(adapter) fl_mtu_bufsize(adapter, FL_MTU_LARGE)
+
+/*
+ * Bits 0..3 of rx_sw_desc.dma_addr have special meaning.  The hardware uses
+ * these to specify the buffer size as an index into the SGE Free List Buffer
+ * Size register array.  We also use bit 4, when the buffer has been unmapped
+ * for DMA, but this is of course never sent to the hardware and is only used
+ * to prevent double unmappings.  All of the above requires that the Free List
+ * Buffers which we allocate have the bottom 5 bits free (0) -- i.e. are
+ * 32-byte or or a power of 2 greater in alignment.  Since the SGE's minimal
+ * Free List Buffer alignment is 32 bytes, this works out for us ...
+ */
+enum {
+	RX_BUF_FLAGS     = 0x1f,   /* bottom five bits are special */
+	RX_BUF_SIZE      = 0x0f,   /* bottom three bits are for buf sizes */
+	RX_UNMAPPED_BUF  = 0x10,   /* buffer is not mapped */
+
+	/*
+	 * XXX We shouldn't depend on being able to use these indices.
+	 * XXX Especially when some other Master PF has initialized the
+	 * XXX adapter or we use the Firmware Configuration File.  We
+	 * XXX should really search through the Host Buffer Size register
+	 * XXX array for the appropriately sized buffer indices.
+	 */
+	RX_SMALL_PG_BUF  = 0x0,   /* small (PAGE_SIZE) page buffer */
+	RX_LARGE_PG_BUF  = 0x1,   /* buffer large page buffer */
+
+	RX_SMALL_MTU_BUF = 0x2,   /* small MTU buffer */
+	RX_LARGE_MTU_BUF = 0x3,   /* large MTU buffer */
+};
+
+/**
+ * txq_avail - return the number of available slots in a Tx queue
+ * @q: the Tx queue
+ *
+ * Returns the number of descriptors in a Tx queue available to write new
+ * packets.
+ */
+static inline unsigned int txq_avail(const struct sge_txq *q)
+{
+	return q->size - 1 - q->in_use;
+}
+
+static int map_mbuf(struct rte_mbuf *mbuf, dma_addr_t *addr)
+{
+	struct rte_mbuf *m = mbuf;
+
+	for (; m; m = m->next, addr++) {
+		*addr = m->buf_physaddr + rte_pktmbuf_headroom(m);
+		if (*addr == 0)
+			goto out_err;
+	}
+	return 0;
+
+out_err:
+	return -ENOMEM;
+}
+
+/**
+ * free_tx_desc - reclaims Tx descriptors and their buffers
+ * @q: the Tx queue to reclaim descriptors from
+ * @n: the number of descriptors to reclaim
+ *
+ * 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 sge_txq *q, unsigned int n)
+{
+	struct tx_sw_desc *d;
+	unsigned int cidx = 0;
+
+	d = &q->sdesc[cidx];
+	while (n--) {
+		if (d->mbuf) {                       /* an SGL is present */
+			rte_pktmbuf_free(d->mbuf);
+			d->mbuf = NULL;
+		}
+		if (d->coalesce.idx) {
+			int i;
+
+			for (i = 0; i < d->coalesce.idx; i++) {
+				rte_pktmbuf_free(d->coalesce.mbuf[i]);
+				d->coalesce.mbuf[i] = NULL;
+			}
+			d->coalesce.idx = 0;
+		}
+		++d;
+		if (++cidx == q->size) {
+			cidx = 0;
+			d = q->sdesc;
+		}
+		RTE_MBUF_PREFETCH_TO_FREE(&q->sdesc->mbuf->pool);
+	}
+}
+
+static void reclaim_tx_desc(struct sge_txq *q, unsigned int n)
+{
+	struct tx_sw_desc *d;
+	unsigned int cidx = q->cidx;
+
+	d = &q->sdesc[cidx];
+	while (n--) {
+		if (d->mbuf) {                       /* an SGL is present */
+			rte_pktmbuf_free(d->mbuf);
+			d->mbuf = NULL;
+		}
+		++d;
+		if (++cidx == q->size) {
+			cidx = 0;
+			d = q->sdesc;
+		}
+	}
+	q->cidx = cidx;
+}
+
+/**
+ * fl_cap - return the capacity of a free-buffer list
+ * @fl: the FL
+ *
+ * Returns the capacity of a free-buffer list.  The capacity is less than
+ * the size because one descriptor needs to be left unpopulated, otherwise
+ * HW will think the FL is empty.
+ */
+static inline unsigned int fl_cap(const struct sge_fl *fl)
+{
+	return fl->size - 8;   /* 1 descriptor = 8 buffers */
+}
+
+/**
+ * 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;
+}
+
+static inline unsigned int get_buf_size(struct adapter *adapter,
+					const struct rx_sw_desc *d)
+{
+	unsigned int rx_buf_size_idx = d->dma_addr & RX_BUF_SIZE;
+	unsigned int buf_size = 0;
+
+	switch (rx_buf_size_idx) {
+	case RX_SMALL_MTU_BUF:
+		buf_size = FL_MTU_SMALL_BUFSIZE(adapter);
+		break;
+
+	case RX_LARGE_MTU_BUF:
+		buf_size = FL_MTU_LARGE_BUFSIZE(adapter);
+		break;
+
+	default:
+		BUG_ON(1);
+		/* NOT REACHED */
+	}
+
+	return buf_size;
+}
+
+/**
+ * free_rx_bufs - free the Rx buffers on an SGE free list
+ * @q: the SGE free list to free buffers from
+ * @n: how many buffers to free
+ *
+ * Release the next @n buffers on an SGE free-buffer Rx queue.   The
+ * buffers must be made inaccessible to HW before calling this function.
+ */
+static void free_rx_bufs(struct sge_fl *q, int n)
+{
+	unsigned int cidx = q->cidx;
+	struct rx_sw_desc *d;
+
+	d = &q->sdesc[cidx];
+	while (n--) {
+		if (d->buf) {
+			rte_pktmbuf_free(d->buf);
+			d->buf = NULL;
+		}
+		++d;
+		if (++cidx == q->size) {
+			cidx = 0;
+			d = q->sdesc;
+		}
+		q->avail--;
+	}
+	q->cidx = cidx;
+}
+
+/**
+ * unmap_rx_buf - unmap the current Rx buffer on an SGE free list
+ * @q: the SGE free list
+ *
+ * Unmap the current buffer on an SGE free-buffer 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.
+ */
+static void unmap_rx_buf(struct sge_fl *q)
+{
+	if (++q->cidx == q->size)
+		q->cidx = 0;
+	q->avail--;
+}
+
+static inline void ring_fl_db(struct adapter *adap, struct sge_fl *q)
+{
+	if (q->pend_cred >= 64) {
+		u32 val = adap->params.arch.sge_fl_db;
+
+		if (is_t4(adap->params.chip))
+			val |= V_PIDX(q->pend_cred / 8);
+		else
+			val |= V_PIDX_T5(q->pend_cred / 8);
+
+		/*
+		 * 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(!q->bar2_addr)) {
+			t4_write_reg(adap, MYPF_REG(A_SGE_PF_KDOORBELL),
+				     val | V_QID(q->cntxt_id));
+		} else {
+			writel(val | V_QID(q->bar2_qid),
+			       (void *)((uintptr_t)q->bar2_addr +
+			       SGE_UDB_KDOORBELL));
+
+			/*
+			 * This Write memory Barrier will force the write to
+			 * the User Doorbell area to be flushed.
+			 */
+			wmb();
+		}
+		q->pend_cred &= 7;
+	}
+}
+
+static inline void set_rx_sw_desc(struct rx_sw_desc *sd, void *buf,
+				  dma_addr_t mapping)
+{
+	sd->buf = buf;
+	sd->dma_addr = mapping;      /* includes size low bits */
+}
+
+/**
+ * refill_fl_usembufs - refill an SGE Rx buffer ring with mbufs
+ * @adap: the adapter
+ * @q: the ring to refill
+ * @n: the number of new buffers to allocate
+ *
+ * (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.  If afterwards the queue is
+ * found critically low mark it as starving in the bitmap of starving FLs.
+ *
+ * Returns the number of buffers allocated.
+ */
+static unsigned int refill_fl_usembufs(struct adapter *adap, struct sge_fl *q,
+				       int n)
+{
+	struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, fl);
+	unsigned int cred = q->avail;
+	__be64 *d = &q->desc[q->pidx];
+	struct rx_sw_desc *sd = &q->sdesc[q->pidx];
+	unsigned int buf_size_idx = RX_SMALL_MTU_BUF;
+	struct rte_mbuf *buf_bulk[n];
+	int ret, i;
+	struct rte_pktmbuf_pool_private *mbp_priv;
+	u8 jumbo_en = rxq->rspq.eth_dev->data->dev_conf.rxmode.jumbo_frame;
+
+	/* Use jumbo mtu buffers iff mbuf data room size can fit jumbo data. */
+	mbp_priv = rte_mempool_get_priv(rxq->rspq.mb_pool);
+	if (jumbo_en &&
+	    ((mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM) >= 9000))
+		buf_size_idx = RX_LARGE_MTU_BUF;
+
+	ret = rte_mempool_get_bulk(rxq->rspq.mb_pool, (void *)buf_bulk, n);
+	if (unlikely(ret != 0)) {
+		dev_debug(adap, "%s: failed to allocated fl entries in bulk ..\n",
+			  __func__);
+		q->alloc_failed++;
+		rxq->rspq.eth_dev->data->rx_mbuf_alloc_failed++;
+		goto out;
+	}
+
+	for (i = 0; i < n; i++) {
+		struct rte_mbuf *mbuf = buf_bulk[i];
+		dma_addr_t mapping;
+
+		if (!mbuf) {
+			dev_debug(adap, "%s: mbuf alloc failed\n", __func__);
+			q->alloc_failed++;
+			rxq->rspq.eth_dev->data->rx_mbuf_alloc_failed++;
+			goto out;
+		}
+
+		rte_mbuf_refcnt_set(mbuf, 1);
+		mbuf->data_off = RTE_PKTMBUF_HEADROOM;
+		mbuf->next = NULL;
+		mbuf->nb_segs = 1;
+		mbuf->port = rxq->rspq.port_id;
+
+		mapping = (dma_addr_t)(mbuf->buf_physaddr + mbuf->data_off);
+		mapping |= buf_size_idx;
+		*d++ = cpu_to_be64(mapping);
+		set_rx_sw_desc(sd, mbuf, mapping);
+		sd++;
+
+		q->avail++;
+		if (++q->pidx == q->size) {
+			q->pidx = 0;
+			sd = q->sdesc;
+			d = q->desc;
+		}
+	}
+
+out:    cred = q->avail - cred;
+	q->pend_cred += cred;
+	ring_fl_db(adap, q);
+
+	if (unlikely(fl_starving(adap, q))) {
+		/*
+		 * Make sure data has been written to free list
+		 */
+		wmb();
+		q->low++;
+	}
+
+	return cred;
+}
+
+/**
+ * refill_fl - refill an SGE Rx buffer ring with mbufs
+ * @adap: the adapter
+ * @q: the ring to refill
+ * @n: the number of new buffers to allocate
+ *
+ * (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.  Returns the number of buffers
+ * allocated.
+ */
+static unsigned int refill_fl(struct adapter *adap, struct sge_fl *q, int n)
+{
+	return refill_fl_usembufs(adap, q, n);
+}
+
+static inline void __refill_fl(struct adapter *adap, struct sge_fl *fl)
+{
+	refill_fl(adap, fl, min(MAX_RX_REFILL, fl_cap(fl) - fl->avail));
+}
+
+/*
+ * Return the number of reclaimable descriptors in a Tx queue.
+ */
+static inline int reclaimable(const struct sge_txq *q)
+{
+	int hw_cidx = ntohs(q->stat->cidx);
+
+	hw_cidx -= q->cidx;
+	if (hw_cidx < 0)
+		return hw_cidx + q->size;
+	return hw_cidx;
+}
+
+/**
+ * reclaim_completed_tx - reclaims completed Tx descriptors
+ * @q: the Tx queue to reclaim completed descriptors from
+ *
+ * Reclaims Tx descriptors that the SGE has indicated it has processed.
+ */
+void reclaim_completed_tx(struct sge_txq *q)
+{
+	unsigned int avail = reclaimable(q);
+
+	do {
+		/* reclaim as much as possible */
+		reclaim_tx_desc(q, avail);
+		q->in_use -= avail;
+		avail = reclaimable(q);
+	} while (avail);
+}
+
+/**
+ * sgl_len - calculates the size of an SGL of the given capacity
+ * @n: the number of SGL entries
+ *
+ * Calculates the number of flits needed for a 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
+ * @n: 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 n)
+{
+	return DIV_ROUND_UP(n, 8);
+}
+
+/**
+ * is_eth_imm - can an Ethernet packet be sent as immediate data?
+ * @m: the packet
+ *
+ * Returns whether an Ethernet packet is small enough to fit as
+ * immediate data. Return value corresponds to the headroom required.
+ */
+static inline int is_eth_imm(const struct rte_mbuf *m)
+{
+	unsigned int hdrlen = (m->ol_flags & PKT_TX_TCP_SEG) ?
+			      sizeof(struct cpl_tx_pkt_lso_core) : 0;
+
+	hdrlen += sizeof(struct cpl_tx_pkt);
+	if (m->pkt_len <= MAX_IMM_TX_PKT_LEN - hdrlen)
+		return hdrlen;
+
+	return 0;
+}
+
+/**
+ * calc_tx_flits - calculate the number of flits for a packet Tx WR
+ * @m: the packet
+ *
+ * Returns the number of flits needed for a Tx WR for the given Ethernet
+ * packet, including the needed WR and CPL headers.
+ */
+static inline unsigned int calc_tx_flits(const struct rte_mbuf *m)
+{
+	unsigned int flits;
+	int hdrlen;
+
+	/*
+	 * If the mbuf 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 mbuf data in the Work Request.
+	 */
+
+	hdrlen = is_eth_imm(m);
+	if (hdrlen)
+		return DIV_ROUND_UP(m->pkt_len + hdrlen, sizeof(__be64));
+
+	/*
+	 * Otherwise, we're going to have to construct a Scatter gather list
+	 * of the mbuf 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 embeded TX Packet Write CPL message.
+	 */
+	flits = sgl_len(m->nb_segs);
+	if (m->tso_segsz)
+		flits += (sizeof(struct fw_eth_tx_pkt_wr) +
+			  sizeof(struct cpl_tx_pkt_lso_core) +
+			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
+	else
+		flits += (sizeof(struct fw_eth_tx_pkt_wr) +
+			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
+	return flits;
+}
+
+/**
+ * write_sgl - populate a scatter/gather list for a packet
+ * @mbuf: the packet
+ * @q: 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 mbuf main-body data to include in the SGL
+ * @addr: address of mapped region
+ *
+ * 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.  @sgl 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 > @sgl.
+ */
+static void write_sgl(struct rte_mbuf *mbuf, struct sge_txq *q,
+		      struct ulptx_sgl *sgl, u64 *end, unsigned int start,
+		      const dma_addr_t *addr)
+{
+	unsigned int i, len;
+	struct ulptx_sge_pair *to;
+	struct rte_mbuf *m = mbuf;
+	unsigned int nfrags = m->nb_segs;
+	struct ulptx_sge_pair buf[nfrags / 2];
+
+	len = m->data_len - start;
+	sgl->len0 = htonl(len);
+	sgl->addr0 = rte_cpu_to_be_64(addr[0]);
+
+	sgl->cmd_nsge = htonl(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
+			      V_ULPTX_NSGE(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 *)q->stat ? buf : sgl->sge;
+
+	for (i = 0; nfrags >= 2; nfrags -= 2, to++) {
+		m = m->next;
+		to->len[0] = rte_cpu_to_be_32(m->data_len);
+		to->addr[0] = rte_cpu_to_be_64(addr[++i]);
+		m = m->next;
+		to->len[1] = rte_cpu_to_be_32(m->data_len);
+		to->addr[1] = rte_cpu_to_be_64(addr[++i]);
+	}
+	if (nfrags) {
+		m = m->next;
+		to->len[0] = rte_cpu_to_be_32(m->data_len);
+		to->len[1] = rte_cpu_to_be_32(0);
+		to->addr[0] = rte_cpu_to_be_64(addr[i + 1]);
+	}
+	if (unlikely((u8 *)end > (u8 *)q->stat)) {
+		unsigned int part0 = RTE_PTR_DIFF((u8 *)q->stat,
+						  (u8 *)sgl->sge);
+		unsigned int part1;
+
+		if (likely(part0))
+			memcpy(sgl->sge, buf, part0);
+		part1 = RTE_PTR_DIFF((u8 *)end, (u8 *)q->stat);
+		rte_memcpy(q->desc, RTE_PTR_ADD((u8 *)buf, part0), part1);
+		end = RTE_PTR_ADD((void *)q->desc, part1);
+	}
+	if ((uintptr_t)end & 8)           /* 0-pad to multiple of 16 */
+		*(u64 *)end = 0;
+}
+
+#define IDXDIFF(head, tail, wrap) \
+	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
+
+#define Q_IDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->size)
+#define R_IDXDIFF(q, idx) IDXDIFF((q)->cidx, (q)->idx, (q)->size)
+
+/**
+ * ring_tx_db - ring a Tx queue's doorbell
+ * @adap: the adapter
+ * @q: 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 *adap, struct sge_txq *q)
+{
+	int n = Q_IDXDIFF(q, dbidx);
+
+	/*
+	 * Make sure that all writes to the TX Descriptors are committed
+	 * before we tell the hardware about them.
+	 */
+	rte_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(!q->bar2_addr)) {
+		u32 val = V_PIDX(n);
+
+		/*
+		 * For T4 we need to participate in the Doorbell Recovery
+		 * mechanism.
+		 */
+		if (!q->db_disabled)
+			t4_write_reg(adap, MYPF_REG(A_SGE_PF_KDOORBELL),
+				     V_QID(q->cntxt_id) | val);
+		else
+			q->db_pidx_inc += n;
+		q->db_pidx = q->pidx;
+	} else {
+		u32 val = V_PIDX_T5(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 & F_DBPRIO);
+
+		writel(val | V_QID(q->bar2_qid),
+		       (void *)((uintptr_t)q->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.
+		 */
+		rte_wmb();
+	}
+	q->dbidx = q->pidx;
+}
+
+/*
+ * Figure out what HW csum a packet wants and return the appropriate control
+ * bits.
+ */
+static u64 hwcsum(enum chip_type chip, const struct rte_mbuf *m)
+{
+	int csum_type;
+
+	if (m->ol_flags & PKT_TX_IP_CKSUM) {
+		switch (m->ol_flags & PKT_TX_L4_MASK) {
+		case PKT_TX_TCP_CKSUM:
+			csum_type = TX_CSUM_TCPIP;
+			break;
+		case PKT_TX_UDP_CKSUM:
+			csum_type = TX_CSUM_UDPIP;
+			break;
+		default:
+			goto nocsum;
+		}
+	} else {
+		goto nocsum;
+	}
+
+	if (likely(csum_type >= TX_CSUM_TCPIP)) {
+		int hdr_len = V_TXPKT_IPHDR_LEN(m->l3_len);
+		int eth_hdr_len = m->l2_len;
+
+		if (CHELSIO_CHIP_VERSION(chip) <= CHELSIO_T5)
+			hdr_len |= V_TXPKT_ETHHDR_LEN(eth_hdr_len);
+		else
+			hdr_len |= V_T6_TXPKT_ETHHDR_LEN(eth_hdr_len);
+		return V_TXPKT_CSUM_TYPE(csum_type) | hdr_len;
+	}
+nocsum:
+	/*
+	 * unknown protocol, disable HW csum
+	 * and hope a bad packet is detected
+	 */
+	return F_TXPKT_L4CSUM_DIS;
+}
+
+static inline void txq_advance(struct sge_txq *q, unsigned int n)
+{
+	q->in_use += n;
+	q->pidx += n;
+	if (q->pidx >= q->size)
+		q->pidx -= q->size;
+}
+
+#define MAX_COALESCE_LEN 64000
+
+static inline int wraps_around(struct sge_txq *q, int ndesc)
+{
+	return (q->pidx + ndesc) > q->size ? 1 : 0;
+}
+
+static void tx_timer_cb(void *data)
+{
+	struct adapter *adap = (struct adapter *)data;
+	struct sge_eth_txq *txq = &adap->sge.ethtxq[0];
+	int i;
+	unsigned int coal_idx;
+
+	/* monitor any pending tx */
+	for (i = 0; i < adap->sge.max_ethqsets; i++, txq++) {
+		if (t4_os_trylock(&txq->txq_lock)) {
+			coal_idx = txq->q.coalesce.idx;
+			if (coal_idx) {
+				if (coal_idx == txq->q.last_coal_idx &&
+				    txq->q.pidx == txq->q.last_pidx) {
+					ship_tx_pkt_coalesce_wr(adap, txq);
+				} else {
+					txq->q.last_coal_idx = coal_idx;
+					txq->q.last_pidx = txq->q.pidx;
+				}
+			}
+			t4_os_unlock(&txq->txq_lock);
+		}
+	}
+	rte_eal_alarm_set(50, tx_timer_cb, (void *)adap);
+}
+
+/**
+ * ship_tx_pkt_coalesce_wr - finalizes and ships a coalesce WR
+ * @ adap: adapter structure
+ * @txq: tx queue
+ *
+ * writes the different fields of the pkts WR and sends it.
+ */
+static inline void ship_tx_pkt_coalesce_wr(struct adapter *adap,
+					   struct sge_eth_txq *txq)
+{
+	u32 wr_mid;
+	struct sge_txq *q = &txq->q;
+	struct fw_eth_tx_pkts_wr *wr;
+	unsigned int ndesc;
+
+	/* fill the pkts WR header */
+	wr = (void *)&q->desc[q->pidx];
+	wr->op_pkd = htonl(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
+
+	wr_mid = V_FW_WR_LEN16(DIV_ROUND_UP(q->coalesce.flits, 2));
+	ndesc = flits_to_desc(q->coalesce.flits);
+	wr->equiq_to_len16 = htonl(wr_mid);
+	wr->plen = cpu_to_be16(q->coalesce.len);
+	wr->npkt = q->coalesce.idx;
+	wr->r3 = 0;
+	wr->type = q->coalesce.type;
+
+	/* zero out coalesce structure members */
+	q->coalesce.idx = 0;
+	q->coalesce.flits = 0;
+	q->coalesce.len = 0;
+
+	txq_advance(q, ndesc);
+	txq->stats.coal_wr++;
+	txq->stats.coal_pkts += wr->npkt;
+
+	if (Q_IDXDIFF(q, equeidx) >= q->size / 2) {
+		q->equeidx = q->pidx;
+		wr_mid |= F_FW_WR_EQUEQ;
+		wr->equiq_to_len16 = htonl(wr_mid);
+	}
+	ring_tx_db(adap, q);
+}
+
+/**
+ * should_tx_packet_coalesce - decides wether to coalesce an mbuf or not
+ * @txq: tx queue where the mbuf is sent
+ * @mbuf: mbuf to be sent
+ * @nflits: return value for number of flits needed
+ * @adap: adapter structure
+ *
+ * This function decides if a packet should be coalesced or not.
+ */
+static inline int should_tx_packet_coalesce(struct sge_eth_txq *txq,
+					    struct rte_mbuf *mbuf,
+					    unsigned int *nflits,
+					    struct adapter *adap)
+{
+	struct sge_txq *q = &txq->q;
+	unsigned int flits, ndesc;
+	unsigned char type = 0;
+	int credits, hw_cidx = ntohs(q->stat->cidx);
+	int in_use = q->pidx - hw_cidx + flits_to_desc(q->coalesce.flits);
+
+	/* use coal WR type 1 when no frags are present */
+	type = (mbuf->nb_segs == 1) ? 1 : 0;
+
+	if (in_use < 0)
+		in_use += q->size;
+
+	if (unlikely(type != q->coalesce.type && q->coalesce.idx))
+		ship_tx_pkt_coalesce_wr(adap, txq);
+
+	/* calculate the number of flits required for coalescing this packet
+	 * without the 2 flits of the WR header. These are added further down
+	 * if we are just starting in new PKTS WR. sgl_len doesn't account for
+	 * the possible 16 bytes alignment ULP TX commands so we do it here.
+	 */
+	flits = (sgl_len(mbuf->nb_segs) + 1) & ~1U;
+	if (type == 0)
+		flits += (sizeof(struct ulp_txpkt) +
+			  sizeof(struct ulptx_idata)) / sizeof(__be64);
+	flits += sizeof(struct cpl_tx_pkt_core) / sizeof(__be64);
+	*nflits = flits;
+
+	/* If coalescing is on, the mbuf is added to a pkts WR */
+	if (q->coalesce.idx) {
+		ndesc = DIV_ROUND_UP(q->coalesce.flits + flits, 8);
+		credits = txq_avail(q) - ndesc;
+
+		/* If we are wrapping or this is last mbuf then, send the
+		 * already coalesced mbufs and let the non-coalesce pass
+		 * handle the mbuf.
+		 */
+		if (unlikely(credits < 0 || wraps_around(q, ndesc))) {
+			ship_tx_pkt_coalesce_wr(adap, txq);
+			return 0;
+		}
+
+		/* If the max coalesce len or the max WR len is reached
+		 * ship the WR and keep coalescing on.
+		 */
+		if (unlikely((q->coalesce.len + mbuf->pkt_len >
+						MAX_COALESCE_LEN) ||
+			     (q->coalesce.flits + flits >
+			      q->coalesce.max))) {
+			ship_tx_pkt_coalesce_wr(adap, txq);
+			goto new;
+		}
+		return 1;
+	}
+
+new:
+	/* start a new pkts WR, the WR header is not filled below */
+	flits += sizeof(struct fw_eth_tx_pkts_wr) / sizeof(__be64);
+	ndesc = flits_to_desc(q->coalesce.flits + flits);
+	credits = txq_avail(q) - ndesc;
+
+	if (unlikely(credits < 0 || wraps_around(q, ndesc)))
+		return 0;
+	q->coalesce.flits += 2;
+	q->coalesce.type = type;
+	q->coalesce.ptr = (unsigned char *)&q->desc[q->pidx] +
+			   2 * sizeof(__be64);
+	return 1;
+}
+
+/**
+ * tx_do_packet_coalesce - add an mbuf to a coalesce WR
+ * @txq: sge_eth_txq used send the mbuf
+ * @mbuf: mbuf to be sent
+ * @flits: flits needed for this mbuf
+ * @adap: adapter structure
+ * @pi: port_info structure
+ * @addr: mapped address of the mbuf
+ *
+ * Adds an mbuf to be sent as part of a coalesce WR by filling a
+ * ulp_tx_pkt command, ulp_tx_sc_imm command, cpl message and
+ * ulp_tx_sc_dsgl command.
+ */
+static inline int tx_do_packet_coalesce(struct sge_eth_txq *txq,
+					struct rte_mbuf *mbuf,
+					int flits, struct adapter *adap,
+					const struct port_info *pi,
+					dma_addr_t *addr)
+{
+	u64 cntrl, *end;
+	struct sge_txq *q = &txq->q;
+	struct ulp_txpkt *mc;
+	struct ulptx_idata *sc_imm;
+	struct cpl_tx_pkt_core *cpl;
+	struct tx_sw_desc *sd;
+	unsigned int idx = q->coalesce.idx, len = mbuf->pkt_len;
+
+	if (q->coalesce.type == 0) {
+		mc = (struct ulp_txpkt *)q->coalesce.ptr;
+		mc->cmd_dest = htonl(V_ULPTX_CMD(4) | V_ULP_TXPKT_DEST(0) |
+				     V_ULP_TXPKT_FID(adap->sge.fw_evtq.cntxt_id) |
+				     F_ULP_TXPKT_RO);
+		mc->len = htonl(DIV_ROUND_UP(flits, 2));
+		sc_imm = (struct ulptx_idata *)(mc + 1);
+		sc_imm->cmd_more = htonl(V_ULPTX_CMD(ULP_TX_SC_IMM) |
+					 F_ULP_TX_SC_MORE);
+		sc_imm->len = htonl(sizeof(*cpl));
+		end = (u64 *)mc + flits;
+		cpl = (struct cpl_tx_pkt_core *)(sc_imm + 1);
+	} else {
+		end = (u64 *)q->coalesce.ptr + flits;
+		cpl = (struct cpl_tx_pkt_core *)q->coalesce.ptr;
+	}
+
+	/* update coalesce structure for this txq */
+	q->coalesce.flits += flits;
+	q->coalesce.ptr += flits * sizeof(__be64);
+	q->coalesce.len += mbuf->pkt_len;
+
+	/* fill the cpl message, same as in t4_eth_xmit, this should be kept
+	 * similar to t4_eth_xmit
+	 */
+	if (mbuf->ol_flags & PKT_TX_IP_CKSUM) {
+		cntrl = hwcsum(adap->params.chip, mbuf) |
+			       F_TXPKT_IPCSUM_DIS;
+		txq->stats.tx_cso++;
+	} else {
+		cntrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
+	}
+
+	if (mbuf->ol_flags & PKT_TX_VLAN_PKT) {
+		txq->stats.vlan_ins++;
+		cntrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(mbuf->vlan_tci);
+	}
+
+	cpl->ctrl0 = htonl(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
+			   V_TXPKT_INTF(pi->tx_chan) |
+			   V_TXPKT_PF(adap->pf));
+	cpl->pack = htons(0);
+	cpl->len = htons(len);
+	cpl->ctrl1 = cpu_to_be64(cntrl);
+	write_sgl(mbuf, q, (struct ulptx_sgl *)(cpl + 1), end, 0,  addr);
+	txq->stats.pkts++;
+	txq->stats.tx_bytes += len;
+
+	sd = &q->sdesc[q->pidx + (idx >> 1)];
+	if (!(idx & 1)) {
+		if (sd->coalesce.idx) {
+			int i;
+
+			for (i = 0; i < sd->coalesce.idx; i++) {
+				rte_pktmbuf_free(sd->coalesce.mbuf[i]);
+				sd->coalesce.mbuf[i] = NULL;
+			}
+		}
+	}
+
+	/* store pointers to the mbuf and the sgl used in free_tx_desc.
+	 * each tx desc can hold two pointers corresponding to the value
+	 * of ETH_COALESCE_PKT_PER_DESC
+	 */
+	sd->coalesce.mbuf[idx & 1] = mbuf;
+	sd->coalesce.sgl[idx & 1] = (struct ulptx_sgl *)(cpl + 1);
+	sd->coalesce.idx = (idx & 1) + 1;
+
+	/* send the coaelsced work request if max reached */
+	if (++q->coalesce.idx == ETH_COALESCE_PKT_NUM)
+		ship_tx_pkt_coalesce_wr(adap, txq);
+	return 0;
+}
+
+/**
+ * t4_eth_xmit - add a packet to an Ethernet Tx queue
+ * @txq: the egress queue
+ * @mbuf: the packet
+ *
+ * Add a packet to an SGE Ethernet Tx queue.  Runs with softirqs disabled.
+ */
+int t4_eth_xmit(struct sge_eth_txq *txq, struct rte_mbuf *mbuf)
+{
+	const struct port_info *pi;
+	struct cpl_tx_pkt_lso_core *lso;
+	struct adapter *adap;
+	struct rte_mbuf *m = mbuf;
+	struct fw_eth_tx_pkt_wr *wr;
+	struct cpl_tx_pkt_core *cpl;
+	struct tx_sw_desc *d;
+	dma_addr_t addr[m->nb_segs];
+	unsigned int flits, ndesc, cflits;
+	int l3hdr_len, l4hdr_len, eth_xtra_len;
+	int len, last_desc;
+	int credits;
+	u32 wr_mid;
+	u64 cntrl, *end;
+	bool v6;
+	u32 max_pkt_len = txq->eth_dev->data->dev_conf.rxmode.max_rx_pkt_len;
+
+	/* Reject xmit if queue is stopped */
+	if (unlikely(txq->flags & EQ_STOPPED))
+		return -(EBUSY);
+
+	/*
+	 * The chip min packet length is 10 octets but play safe and reject
+	 * anything shorter than an Ethernet header.
+	 */
+	if (unlikely(m->pkt_len < ETHER_HDR_LEN)) {
+out_free:
+		rte_pktmbuf_free(m);
+		return 0;
+	}
+
+	if ((!(m->ol_flags & PKT_TX_TCP_SEG)) &&
+	    (unlikely(m->pkt_len > max_pkt_len)))
+		goto out_free;
+
+	pi = (struct port_info *)txq->eth_dev->data->dev_private;
+	adap = pi->adapter;
+
+	cntrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
+	/* align the end of coalesce WR to a 512 byte boundary */
+	txq->q.coalesce.max = (8 - (txq->q.pidx & 7)) * 8;
+
+	if (!((m->ol_flags & PKT_TX_TCP_SEG) || (m->pkt_len > ETHER_MAX_LEN))) {
+		if (should_tx_packet_coalesce(txq, mbuf, &cflits, adap)) {
+			if (unlikely(map_mbuf(mbuf, addr) < 0)) {
+				dev_warn(adap, "%s: mapping err for coalesce\n",
+					 __func__);
+				txq->stats.mapping_err++;
+				goto out_free;
+			}
+			rte_prefetch0((volatile void *)addr);
+			return tx_do_packet_coalesce(txq, mbuf, cflits, adap,
+						     pi, addr);
+		} else {
+			return -EBUSY;
+		}
+	}
+
+	if (txq->q.coalesce.idx)
+		ship_tx_pkt_coalesce_wr(adap, txq);
+
+	flits = calc_tx_flits(m);
+	ndesc = flits_to_desc(flits);
+	credits = txq_avail(&txq->q) - ndesc;
+
+	if (unlikely(credits < 0)) {
+		dev_debug(adap, "%s: Tx ring %u full; credits = %d\n",
+			  __func__, txq->q.cntxt_id, credits);
+		return -EBUSY;
+	}
+
+	if (unlikely(map_mbuf(m, addr) < 0)) {
+		txq->stats.mapping_err++;
+		goto out_free;
+	}
+
+	wr_mid = V_FW_WR_LEN16(DIV_ROUND_UP(flits, 2));
+	if (Q_IDXDIFF(&txq->q, equeidx)  >= 64) {
+		txq->q.equeidx = txq->q.pidx;
+		wr_mid |= F_FW_WR_EQUEQ;
+	}
+
+	wr = (void *)&txq->q.desc[txq->q.pidx];
+	wr->equiq_to_len16 = htonl(wr_mid);
+	wr->r3 = rte_cpu_to_be_64(0);
+	end = (u64 *)wr + flits;
+
+	len = 0;
+	len += sizeof(*cpl);
+
+	/* Coalescing skipped and we send through normal path */
+	if (!(m->ol_flags & PKT_TX_TCP_SEG)) {
+		wr->op_immdlen = htonl(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
+				       V_FW_WR_IMMDLEN(len));
+		cpl = (void *)(wr + 1);
+		if (m->ol_flags & PKT_TX_IP_CKSUM) {
+			cntrl = hwcsum(adap->params.chip, m) |
+				F_TXPKT_IPCSUM_DIS;
+			txq->stats.tx_cso++;
+		}
+	} else {
+		lso = (void *)(wr + 1);
+		v6 = (m->ol_flags & PKT_TX_IPV6) != 0;
+		l3hdr_len = m->l3_len;
+		l4hdr_len = m->l4_len;
+		eth_xtra_len = m->l2_len - ETHER_HDR_LEN;
+		len += sizeof(*lso);
+		wr->op_immdlen = htonl(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
+				       V_FW_WR_IMMDLEN(len));
+		lso->lso_ctrl = htonl(V_LSO_OPCODE(CPL_TX_PKT_LSO) |
+				      F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
+				      V_LSO_IPV6(v6) |
+				      V_LSO_ETHHDR_LEN(eth_xtra_len / 4) |
+				      V_LSO_IPHDR_LEN(l3hdr_len / 4) |
+				      V_LSO_TCPHDR_LEN(l4hdr_len / 4));
+		lso->ipid_ofst = htons(0);
+		lso->mss = htons(m->tso_segsz);
+		lso->seqno_offset = htonl(0);
+		if (is_t4(adap->params.chip))
+			lso->len = htonl(m->pkt_len);
+		else
+			lso->len = htonl(V_LSO_T5_XFER_SIZE(m->pkt_len));
+		cpl = (void *)(lso + 1);
+		cntrl = V_TXPKT_CSUM_TYPE(v6 ? TX_CSUM_TCPIP6 : TX_CSUM_TCPIP) |
+			V_TXPKT_IPHDR_LEN(l3hdr_len) |
+			V_TXPKT_ETHHDR_LEN(eth_xtra_len);
+		txq->stats.tso++;
+		txq->stats.tx_cso += m->tso_segsz;
+	}
+
+	if (m->ol_flags & PKT_TX_VLAN_PKT) {
+		txq->stats.vlan_ins++;
+		cntrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->vlan_tci);
+	}
+
+	cpl->ctrl0 = htonl(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
+			   V_TXPKT_INTF(pi->tx_chan) |
+			   V_TXPKT_PF(adap->pf));
+	cpl->pack = htons(0);
+	cpl->len = htons(m->pkt_len);
+	cpl->ctrl1 = cpu_to_be64(cntrl);
+
+	txq->stats.pkts++;
+	txq->stats.tx_bytes += m->pkt_len;
+	last_desc = txq->q.pidx + ndesc - 1;
+	if (last_desc >= (int)txq->q.size)
+		last_desc -= txq->q.size;
+
+	d = &txq->q.sdesc[last_desc];
+	if (d->coalesce.idx) {
+		int i;
+
+		for (i = 0; i < d->coalesce.idx; i++) {
+			rte_pktmbuf_free(d->coalesce.mbuf[i]);
+			d->coalesce.mbuf[i] = NULL;
+		}
+		d->coalesce.idx = 0;
+	}
+	write_sgl(m, &txq->q, (struct ulptx_sgl *)(cpl + 1), end, 0,
+		  addr);
+	txq->q.sdesc[last_desc].mbuf = m;
+	txq->q.sdesc[last_desc].sgl = (struct ulptx_sgl *)(cpl + 1);
+	txq_advance(&txq->q, ndesc);
+	ring_tx_db(adap, &txq->q);
+	return 0;
+}
+
+/**
+ * alloc_ring - allocate resources for an SGE descriptor ring
+ * @dev: the PCI device's core device
+ * @nelem: the number of descriptors
+ * @elem_size: the size of each descriptor
+ * @sw_size: the size of the SW state associated with each ring element
+ * @phys: the physical address of the allocated ring
+ * @metadata: address of the array holding the SW state for the ring
+ * @stat_size: extra space in HW ring for status information
+ * @node: preferred node for memory allocations
+ *
+ * Allocates resources for an SGE descriptor ring, such as Tx queues,
+ * free buffer lists, or response queues.  Each SGE ring requires
+ * space for its HW descriptors plus, optionally, space for the SW state
+ * associated with each HW entry (the metadata).  The function returns
+ * three values: the virtual address for the HW ring (the return value
+ * of the function), the bus address of the HW ring, and the address
+ * of the SW ring.
+ */
+static void *alloc_ring(size_t nelem, size_t elem_size,
+			size_t sw_size, dma_addr_t *phys, void *metadata,
+			size_t stat_size, __rte_unused uint16_t queue_id,
+			int socket_id, const char *z_name,
+			const char *z_name_sw)
+{
+	size_t len = CXGBE_MAX_RING_DESC_SIZE * elem_size + stat_size;
+	const struct rte_memzone *tz;
+	void *s = NULL;
+
+	dev_debug(adapter, "%s: nelem = %zu; elem_size = %zu; sw_size = %zu; "
+		  "stat_size = %zu; queue_id = %u; socket_id = %d; z_name = %s;"
+		  " z_name_sw = %s\n", __func__, nelem, elem_size, sw_size,
+		  stat_size, queue_id, socket_id, z_name, z_name_sw);
+
+	tz = rte_memzone_lookup(z_name);
+	if (tz) {
+		dev_debug(adapter, "%s: tz exists...returning existing..\n",
+			  __func__);
+		goto alloc_sw_ring;
+	}
+
+	/*
+	 * Allocate TX/RX ring hardware descriptors. A memzone large enough to
+	 * handle the maximum ring size is allocated in order to allow for
+	 * resizing in later calls to the queue setup function.
+	 */
+	tz = rte_memzone_reserve_aligned(z_name, len, socket_id, 0, 4096);
+	if (!tz)
+		return NULL;
+
+alloc_sw_ring:
+	memset(tz->addr, 0, len);
+	if (sw_size) {
+		s = rte_zmalloc_socket(z_name_sw, nelem * sw_size,
+				       RTE_CACHE_LINE_SIZE, socket_id);
+
+		if (!s) {
+			dev_err(adapter, "%s: failed to get sw_ring memory\n",
+				__func__);
+			return NULL;
+		}
+	}
+	if (metadata)
+		*(void **)metadata = s;
+
+	*phys = (uint64_t)tz->phys_addr;
+	return tz->addr;
+}
+
+/**
+ * t4_pktgl_to_mbuf_usembufs - build an mbuf from a packet gather list
+ * @gl: the gather list
+ *
+ * Builds an mbuf from the given packet gather list.  Returns the mbuf or
+ * %NULL if mbuf allocation failed.
+ */
+static struct rte_mbuf *t4_pktgl_to_mbuf_usembufs(const struct pkt_gl *gl)
+{
+	/*
+	 * If there's only one mbuf fragment, just return that.
+	 */
+	if (likely(gl->nfrags == 1))
+		return gl->mbufs[0];
+
+	return NULL;
+}
+
+/**
+ * t4_pktgl_to_mbuf - build an mbuf from a packet gather list
+ * @gl: the gather list
+ *
+ * Builds an mbuf from the given packet gather list.  Returns the mbuf or
+ * %NULL if mbuf allocation failed.
+ */
+static struct rte_mbuf *t4_pktgl_to_mbuf(const struct pkt_gl *gl)
+{
+	return t4_pktgl_to_mbuf_usembufs(gl);
+}
+
+/**
+ * t4_ethrx_handler - process an ingress ethernet packet
+ * @q: the response queue that received the packet
+ * @rsp: the response queue descriptor holding the RX_PKT message
+ * @si: the gather list of packet fragments
+ *
+ * Process an ingress ethernet packet and deliver it to the stack.
+ */
+int t4_ethrx_handler(struct sge_rspq *q, const __be64 *rsp,
+		     const struct pkt_gl *si)
+{
+	struct rte_mbuf *mbuf;
+	const struct cpl_rx_pkt *pkt;
+	const struct rss_header *rss_hdr;
+	bool csum_ok;
+	struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, rspq);
+
+	rss_hdr = (const void *)rsp;
+	pkt = (const void *)&rsp[1];
+	csum_ok = pkt->csum_calc && !pkt->err_vec;
+
+	mbuf = t4_pktgl_to_mbuf(si);
+	if (unlikely(!mbuf)) {
+		rxq->stats.rx_drops++;
+		return 0;
+	}
+
+	mbuf->port = pkt->iff;
+	if (pkt->l2info & htonl(F_RXF_IP)) {
+		mbuf->packet_type = RTE_PTYPE_L3_IPV4;
+		if (unlikely(!csum_ok))
+			mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
+
+		if ((pkt->l2info & htonl(F_RXF_UDP | F_RXF_TCP)) && !csum_ok)
+			mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
+	} else if (pkt->l2info & htonl(F_RXF_IP6)) {
+		mbuf->packet_type = RTE_PTYPE_L3_IPV6;
+	}
+
+	mbuf->port = pkt->iff;
+
+	if (!rss_hdr->filter_tid && rss_hdr->hash_type) {
+		mbuf->ol_flags |= PKT_RX_RSS_HASH;
+		mbuf->hash.rss = ntohl(rss_hdr->hash_val);
+	}
+
+	if (pkt->vlan_ex) {
+		mbuf->ol_flags |= PKT_RX_VLAN_PKT;
+		mbuf->vlan_tci = ntohs(pkt->vlan);
+	}
+	rxq->stats.pkts++;
+	rxq->stats.rx_bytes += mbuf->pkt_len;
+
+	return 0;
+}
+
+/**
+ * is_new_response - check if a response is newly written
+ * @r: the response descriptor
+ * @q: the response queue
+ *
+ * Returns true if a response descriptor contains a yet unprocessed
+ * response.
+ */
+static inline bool is_new_response(const struct rsp_ctrl *r,
+				   const struct sge_rspq *q)
+{
+	return (r->u.type_gen >> S_RSPD_GEN) == q->gen;
+}
+
+#define CXGB4_MSG_AN ((void *)1)
+
+/**
+ * rspq_next - advance to the next entry in a response queue
+ * @q: the queue
+ *
+ * Updates the state of a response queue to advance it to the next entry.
+ */
+static inline void rspq_next(struct sge_rspq *q)
+{
+	q->cur_desc = (const __be64 *)((const char *)q->cur_desc + q->iqe_len);
+	if (unlikely(++q->cidx == q->size)) {
+		q->cidx = 0;
+		q->gen ^= 1;
+		q->cur_desc = q->desc;
+	}
+}
+
+/**
+ * process_responses - process responses from an SGE response queue
+ * @q: the ingress queue to process
+ * @budget: how many responses can be processed in this round
+ * @rx_pkts: mbuf to put the pkts
+ *
+ * Process responses from an SGE response queue up to the supplied budget.
+ * Responses include received packets as well as control messages from FW
+ * or HW.
+ *
+ * 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 *q, int budget,
+			     struct rte_mbuf **rx_pkts)
+{
+	int ret = 0, rsp_type;
+	int budget_left = budget;
+	const struct rsp_ctrl *rc;
+	struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, rspq);
+
+	while (likely(budget_left)) {
+		rc = (const struct rsp_ctrl *)
+		     ((const char *)q->cur_desc + (q->iqe_len - sizeof(*rc)));
+
+		if (!is_new_response(rc, q))
+			break;
+
+		/*
+		 * Ensure response has been read
+		 */
+		rmb();
+		rsp_type = G_RSPD_TYPE(rc->u.type_gen);
+
+		if (likely(rsp_type == X_RSPD_TYPE_FLBUF)) {
+			const struct rx_sw_desc *rsd =
+						&rxq->fl.sdesc[rxq->fl.cidx];
+			const struct rss_header *rss_hdr =
+						(const void *)q->cur_desc;
+			const struct cpl_rx_pkt *cpl =
+						(const void *)&q->cur_desc[1];
+			bool csum_ok = cpl->csum_calc && !cpl->err_vec;
+			struct rte_mbuf *pkt, *npkt;
+			u32 len, bufsz;
+
+			len = ntohl(rc->pldbuflen_qid);
+			BUG_ON(!(len & F_RSPD_NEWBUF));
+			pkt = rsd->buf;
+			npkt = pkt;
+			len = G_RSPD_LEN(len);
+			pkt->pkt_len = len;
+
+			/* Chain mbufs into len if necessary */
+			while (len) {
+				struct rte_mbuf *new_pkt = rsd->buf;
+
+				bufsz = min(get_buf_size(q->adapter, rsd), len);
+				new_pkt->data_len = bufsz;
+				unmap_rx_buf(&rxq->fl);
+				len -= bufsz;
+				npkt->next = new_pkt;
+				npkt = new_pkt;
+				pkt->nb_segs++;
+				rsd = &rxq->fl.sdesc[rxq->fl.cidx];
+			}
+			npkt->next = NULL;
+			pkt->nb_segs--;
+
+			if (cpl->l2info & htonl(F_RXF_IP)) {
+				pkt->packet_type = RTE_PTYPE_L3_IPV4;
+				if (unlikely(!csum_ok))
+					pkt->ol_flags |= PKT_RX_IP_CKSUM_BAD;
+
+				if ((cpl->l2info &
+				     htonl(F_RXF_UDP | F_RXF_TCP)) && !csum_ok)
+					pkt->ol_flags |= PKT_RX_L4_CKSUM_BAD;
+			} else if (cpl->l2info & htonl(F_RXF_IP6)) {
+				pkt->packet_type = RTE_PTYPE_L3_IPV6;
+			}
+
+			if (!rss_hdr->filter_tid && rss_hdr->hash_type) {
+				pkt->ol_flags |= PKT_RX_RSS_HASH;
+				pkt->hash.rss = ntohl(rss_hdr->hash_val);
+			}
+
+			if (cpl->vlan_ex) {
+				pkt->ol_flags |= PKT_RX_VLAN_PKT;
+				pkt->vlan_tci = ntohs(cpl->vlan);
+			}
+			rxq->stats.pkts++;
+			rxq->stats.rx_bytes += pkt->pkt_len;
+			rx_pkts[budget - budget_left] = pkt;
+		} else if (likely(rsp_type == X_RSPD_TYPE_CPL)) {
+			ret = q->handler(q, q->cur_desc, NULL);
+		} else {
+			ret = q->handler(q, (const __be64 *)rc, CXGB4_MSG_AN);
+		}
+
+		if (unlikely(ret)) {
+			/* couldn't process descriptor, back off for recovery */
+			q->next_intr_params = V_QINTR_TIMER_IDX(NOMEM_TMR_IDX);
+			break;
+		}
+
+		rspq_next(q);
+		budget_left--;
+
+		if (R_IDXDIFF(q, gts_idx) >= 64) {
+			unsigned int cidx_inc = R_IDXDIFF(q, gts_idx);
+			unsigned int params;
+			u32 val;
+
+			if (fl_cap(&rxq->fl) - rxq->fl.avail >= 64)
+				__refill_fl(q->adapter, &rxq->fl);
+			params = V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX);
+			q->next_intr_params = params;
+			val = V_CIDXINC(cidx_inc) | V_SEINTARM(params);
+
+			if (unlikely(!q->bar2_addr))
+				t4_write_reg(q->adapter, MYPF_REG(A_SGE_PF_GTS),
+					     val |
+					     V_INGRESSQID((u32)q->cntxt_id));
+			else {
+				writel(val | V_INGRESSQID(q->bar2_qid),
+				       (void *)((uintptr_t)q->bar2_addr +
+				       SGE_UDB_GTS));
+				/*
+				 * This Write memory Barrier will force the
+				 * write to the User Doorbell area to be
+				 * flushed.
+				 */
+				wmb();
+			}
+			q->gts_idx = q->cidx;
+		}
+	}
+
+	/*
+	 * If this is a Response Queue with an associated Free List and
+	 * there's room for another chunk of new Free List buffer pointers,
+	 * refill the Free List.
+	 */
+
+	if (q->offset >= 0 && fl_cap(&rxq->fl) - rxq->fl.avail >= 64)
+		__refill_fl(q->adapter, &rxq->fl);
+
+	return budget - budget_left;
+}
+
+int cxgbe_poll(struct sge_rspq *q, struct rte_mbuf **rx_pkts,
+	       unsigned int budget, unsigned int *work_done)
+{
+	int err = 0;
+
+	*work_done = process_responses(q, budget, rx_pkts);
+	return err;
+}
+
+/**
+ * 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;
+}
+
+int t4_sge_eth_rxq_start(struct adapter *adap, struct sge_rspq *rq)
+{
+	struct sge_eth_rxq *rxq = container_of(rq, struct sge_eth_rxq, rspq);
+	unsigned int fl_id = rxq->fl.size ? rxq->fl.cntxt_id : 0xffff;
+
+	return t4_iq_start_stop(adap, adap->mbox, true, adap->pf, 0,
+				rq->cntxt_id, fl_id, 0xffff);
+}
+
+int t4_sge_eth_rxq_stop(struct adapter *adap, struct sge_rspq *rq)
+{
+	struct sge_eth_rxq *rxq = container_of(rq, struct sge_eth_rxq, rspq);
+	unsigned int fl_id = rxq->fl.size ? rxq->fl.cntxt_id : 0xffff;
+
+	return t4_iq_start_stop(adap, adap->mbox, false, adap->pf, 0,
+				rq->cntxt_id, fl_id, 0xffff);
+}
+
+/*
+ * @intr_idx: MSI/MSI-X vector if >=0, -(absolute qid + 1) if < 0
+ * @cong: < 0 -> no congestion feedback, >= 0 -> congestion channel map
+ */
+int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
+		     struct rte_eth_dev *eth_dev, int intr_idx,
+		     struct sge_fl *fl, rspq_handler_t hnd, int cong,
+		     struct rte_mempool *mp, int queue_id, int socket_id)
+{
+	int ret, flsz = 0;
+	struct fw_iq_cmd c;
+	struct sge *s = &adap->sge;
+	struct port_info *pi = (struct port_info *)(eth_dev->data->dev_private);
+	char z_name[RTE_MEMZONE_NAMESIZE];
+	char z_name_sw[RTE_MEMZONE_NAMESIZE];
+	unsigned int nb_refill;
+
+	/* Size needs to be multiple of 16, including status entry. */
+	iq->size = cxgbe_roundup(iq->size, 16);
+
+	snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
+		 eth_dev->driver->pci_drv.name, fwevtq ? "fwq_ring" : "rx_ring",
+		 eth_dev->data->port_id, queue_id);
+	snprintf(z_name_sw, sizeof(z_name_sw), "%s_sw_ring", z_name);
+
+	iq->desc = alloc_ring(iq->size, iq->iqe_len, 0, &iq->phys_addr, NULL, 0,
+			      queue_id, socket_id, z_name, z_name_sw);
+	if (!iq->desc)
+		return -ENOMEM;
+
+	memset(&c, 0, sizeof(c));
+	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
+			    F_FW_CMD_WRITE | F_FW_CMD_EXEC |
+			    V_FW_IQ_CMD_PFN(adap->pf) | V_FW_IQ_CMD_VFN(0));
+	c.alloc_to_len16 = htonl(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
+				 (sizeof(c) / 16));
+	c.type_to_iqandstindex =
+		htonl(V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
+		      V_FW_IQ_CMD_IQASYNCH(fwevtq) |
+		      V_FW_IQ_CMD_VIID(pi->viid) |
+		      V_FW_IQ_CMD_IQANDST(intr_idx < 0) |
+		      V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT) |
+		      V_FW_IQ_CMD_IQANDSTINDEX(intr_idx >= 0 ? intr_idx :
+							       -intr_idx - 1));
+	c.iqdroprss_to_iqesize =
+		htons(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
+		      F_FW_IQ_CMD_IQGTSMODE |
+		      V_FW_IQ_CMD_IQINTCNTTHRESH(iq->pktcnt_idx) |
+		      V_FW_IQ_CMD_IQESIZE(ilog2(iq->iqe_len) - 4));
+	c.iqsize = htons(iq->size);
+	c.iqaddr = cpu_to_be64(iq->phys_addr);
+	if (cong >= 0)
+		c.iqns_to_fl0congen = htonl(F_FW_IQ_CMD_IQFLINTCONGEN);
+
+	if (fl) {
+		struct sge_eth_rxq *rxq = container_of(fl, struct sge_eth_rxq,
+						       fl);
+		enum chip_type chip = (enum chip_type)CHELSIO_CHIP_VERSION(
+				adap->params.chip);
+
+		/*
+		 * 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 and at least 2 Egress Units larger
+		 * than the SGE's Egress Congrestion Threshold
+		 * (fl_starve_thres - 1).
+		 */
+		if (fl->size < s->fl_starve_thres - 1 + 2 * 8)
+			fl->size = s->fl_starve_thres - 1 + 2 * 8;
+		fl->size = cxgbe_roundup(fl->size, 8);
+
+		snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
+			 eth_dev->driver->pci_drv.name,
+			 fwevtq ? "fwq_ring" : "fl_ring",
+			 eth_dev->data->port_id, queue_id);
+		snprintf(z_name_sw, sizeof(z_name_sw), "%s_sw_ring", z_name);
+
+		fl->desc = alloc_ring(fl->size, sizeof(__be64),
+				      sizeof(struct rx_sw_desc),
+				      &fl->addr, &fl->sdesc, s->stat_len,
+				      queue_id, socket_id, z_name, z_name_sw);
+
+		if (!fl->desc)
+			goto fl_nomem;
+
+		flsz = fl->size / 8 + s->stat_len / sizeof(struct tx_desc);
+		c.iqns_to_fl0congen |=
+			htonl(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
+			      (unlikely(rxq->usembufs) ?
+			       0 : F_FW_IQ_CMD_FL0PACKEN) |
+			      F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
+			      F_FW_IQ_CMD_FL0PADEN);
+		if (cong >= 0)
+			c.iqns_to_fl0congen |=
+				htonl(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
+				      F_FW_IQ_CMD_FL0CONGCIF |
+				      F_FW_IQ_CMD_FL0CONGEN);
+
+		/* In T6, for egress queue type FL there is internal overhead
+		 * of 16B for header going into FLM module.
+		 * Hence maximum allowed burst size will be 448 bytes.
+		 */
+		c.fl0dcaen_to_fl0cidxfthresh =
+			htons(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
+			      V_FW_IQ_CMD_FL0FBMAX((chip <= CHELSIO_T5) ?
+			      X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
+		c.fl0size = htons(flsz);
+		c.fl0addr = cpu_to_be64(fl->addr);
+	}
+
+	ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
+	if (ret)
+		goto err;
+
+	iq->cur_desc = iq->desc;
+	iq->cidx = 0;
+	iq->gts_idx = 0;
+	iq->gen = 1;
+	iq->next_intr_params = iq->intr_params;
+	iq->cntxt_id = ntohs(c.iqid);
+	iq->abs_id = ntohs(c.physiqid);
+	iq->bar2_addr = bar2_address(adap, iq->cntxt_id, T4_BAR2_QTYPE_INGRESS,
+				     &iq->bar2_qid);
+	iq->size--;                           /* subtract status entry */
+	iq->eth_dev = eth_dev;
+	iq->handler = hnd;
+	iq->port_id = pi->port_id;
+	iq->mb_pool = mp;
+
+	/* set offset to -1 to distinguish ingress queues without FL */
+	iq->offset = fl ? 0 : -1;
+
+	if (fl) {
+		fl->cntxt_id = ntohs(c.fl0id);
+		fl->avail = 0;
+		fl->pend_cred = 0;
+		fl->pidx = 0;
+		fl->cidx = 0;
+		fl->alloc_failed = 0;
+
+		/*
+		 * Note, we must initialize the BAR2 Free List User Doorbell
+		 * information before refilling the Free List!
+		 */
+		fl->bar2_addr = bar2_address(adap, fl->cntxt_id,
+					     T4_BAR2_QTYPE_EGRESS,
+					     &fl->bar2_qid);
+
+		nb_refill = refill_fl(adap, fl, fl_cap(fl));
+		if (nb_refill != fl_cap(fl)) {
+			ret = -ENOMEM;
+			dev_err(adap, "%s: mbuf alloc failed with error: %d\n",
+				__func__, ret);
+			goto refill_fl_err;
+		}
+	}
+
+	/*
+	 * For T5 and later we attempt to set up the Congestion Manager values
+	 * of the new RX Ethernet Queue.  This should really be handled by
+	 * firmware because it's more complex than any host driver wants to
+	 * get involved with and it's different per chip and this is almost
+	 * certainly wrong.  Formware would be wrong as well, but it would be
+	 * a lot easier to fix in one place ...  For now we do something very
+	 * simple (and hopefully less wrong).
+	 */
+	if (!is_t4(adap->params.chip) && cong >= 0) {
+		u32 param, val;
+		int i;
+
+		param = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
+			 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
+			 V_FW_PARAMS_PARAM_YZ(iq->cntxt_id));
+		if (cong == 0) {
+			val = V_CONMCTXT_CNGTPMODE(X_CONMCTXT_CNGTPMODE_QUEUE);
+		} else {
+			val = V_CONMCTXT_CNGTPMODE(
+					X_CONMCTXT_CNGTPMODE_CHANNEL);
+			for (i = 0; i < 4; i++) {
+				if (cong & (1 << i))
+					val |= V_CONMCTXT_CNGCHMAP(1 <<
+								   (i << 2));
+			}
+		}
+		ret = t4_set_params(adap, adap->mbox, adap->pf, 0, 1,
+				    &param, &val);
+		if (ret)
+			dev_warn(adap->pdev_dev, "Failed to set Congestion Manager Context for Ingress Queue %d: %d\n",
+				 iq->cntxt_id, -ret);
+	}
+
+	return 0;
+
+refill_fl_err:
+	t4_iq_free(adap, adap->mbox, adap->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
+		   iq->cntxt_id, fl->cntxt_id, 0xffff);
+fl_nomem:
+	ret = -ENOMEM;
+err:
+	iq->cntxt_id = 0;
+	iq->abs_id = 0;
+	if (iq->desc)
+		iq->desc = NULL;
+
+	if (fl && fl->desc) {
+		rte_free(fl->sdesc);
+		fl->cntxt_id = 0;
+		fl->sdesc = NULL;
+		fl->desc = NULL;
+	}
+	return ret;
+}
+
+static void init_txq(struct adapter *adap, struct sge_txq *q, unsigned int id)
+{
+	q->cntxt_id = id;
+	q->bar2_addr = bar2_address(adap, q->cntxt_id, T4_BAR2_QTYPE_EGRESS,
+				    &q->bar2_qid);
+	q->cidx = 0;
+	q->pidx = 0;
+	q->dbidx = 0;
+	q->in_use = 0;
+	q->equeidx = 0;
+	q->coalesce.idx = 0;
+	q->coalesce.len = 0;
+	q->coalesce.flits = 0;
+	q->last_coal_idx = 0;
+	q->last_pidx = 0;
+	q->stat = (void *)&q->desc[q->size];
+}
+
+int t4_sge_eth_txq_start(struct sge_eth_txq *txq)
+{
+	/*
+	 *  TODO: For flow-control, queue may be stopped waiting to reclaim
+	 *  credits.
+	 *  Ensure queue is in EQ_STOPPED state before starting it.
+	 */
+	if (!(txq->flags & EQ_STOPPED))
+		return -(EBUSY);
+
+	txq->flags &= ~EQ_STOPPED;
+
+	return 0;
+}
+
+int t4_sge_eth_txq_stop(struct sge_eth_txq *txq)
+{
+	txq->flags |= EQ_STOPPED;
+
+	return 0;
+}
+
+int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
+			 struct rte_eth_dev *eth_dev, uint16_t queue_id,
+			 unsigned int iqid, int socket_id)
+{
+	int ret, nentries;
+	struct fw_eq_eth_cmd c;
+	struct sge *s = &adap->sge;
+	struct port_info *pi = (struct port_info *)(eth_dev->data->dev_private);
+	char z_name[RTE_MEMZONE_NAMESIZE];
+	char z_name_sw[RTE_MEMZONE_NAMESIZE];
+
+	/* Add status entries */
+	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
+
+	snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
+		 eth_dev->driver->pci_drv.name, "tx_ring",
+		 eth_dev->data->port_id, queue_id);
+	snprintf(z_name_sw, sizeof(z_name_sw), "%s_sw_ring", z_name);
+
+	txq->q.desc = alloc_ring(txq->q.size, sizeof(struct tx_desc),
+				 sizeof(struct tx_sw_desc), &txq->q.phys_addr,
+				 &txq->q.sdesc, s->stat_len, queue_id,
+				 socket_id, z_name, z_name_sw);
+	if (!txq->q.desc)
+		return -ENOMEM;
+
+	memset(&c, 0, sizeof(c));
+	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
+			    F_FW_CMD_WRITE | F_FW_CMD_EXEC |
+			    V_FW_EQ_ETH_CMD_PFN(adap->pf) |
+			    V_FW_EQ_ETH_CMD_VFN(0));
+	c.alloc_to_len16 = htonl(F_FW_EQ_ETH_CMD_ALLOC |
+				 F_FW_EQ_ETH_CMD_EQSTART | (sizeof(c) / 16));
+	c.autoequiqe_to_viid = htonl(F_FW_EQ_ETH_CMD_AUTOEQUEQE |
+				     V_FW_EQ_ETH_CMD_VIID(pi->viid));
+	c.fetchszm_to_iqid =
+		htonl(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
+		      V_FW_EQ_ETH_CMD_PCIECHN(pi->tx_chan) |
+		      F_FW_EQ_ETH_CMD_FETCHRO | V_FW_EQ_ETH_CMD_IQID(iqid));
+	c.dcaen_to_eqsize =
+		htonl(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
+		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
+		      V_FW_EQ_ETH_CMD_EQSIZE(nentries));
+	c.eqaddr = rte_cpu_to_be_64(txq->q.phys_addr);
+
+	ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
+	if (ret) {
+		rte_free(txq->q.sdesc);
+		txq->q.sdesc = NULL;
+		txq->q.desc = NULL;
+		return ret;
+	}
+
+	init_txq(adap, &txq->q, G_FW_EQ_ETH_CMD_EQID(ntohl(c.eqid_pkd)));
+	txq->stats.tso = 0;
+	txq->stats.pkts = 0;
+	txq->stats.tx_cso = 0;
+	txq->stats.coal_wr = 0;
+	txq->stats.vlan_ins = 0;
+	txq->stats.tx_bytes = 0;
+	txq->stats.coal_pkts = 0;
+	txq->stats.mapping_err = 0;
+	txq->flags |= EQ_STOPPED;
+	txq->eth_dev = eth_dev;
+	t4_os_lock_init(&txq->txq_lock);
+	return 0;
+}
+
+static void free_txq(struct sge_txq *q)
+{
+	q->cntxt_id = 0;
+	q->sdesc = NULL;
+	q->desc = NULL;
+}
+
+static void free_rspq_fl(struct adapter *adap, struct sge_rspq *rq,
+			 struct sge_fl *fl)
+{
+	unsigned int fl_id = fl ? fl->cntxt_id : 0xffff;
+
+	t4_iq_free(adap, adap->mbox, adap->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
+		   rq->cntxt_id, fl_id, 0xffff);
+	rq->cntxt_id = 0;
+	rq->abs_id = 0;
+	rq->desc = NULL;
+
+	if (fl) {
+		free_rx_bufs(fl, fl->avail);
+		rte_free(fl->sdesc);
+		fl->sdesc = NULL;
+		fl->cntxt_id = 0;
+		fl->desc = NULL;
+	}
+}
+
+/*
+ * Clear all queues of the port
+ *
+ * Note:  This function must only be called after rx and tx path
+ * of the port have been disabled.
+ */
+void t4_sge_eth_clear_queues(struct port_info *pi)
+{
+	int i;
+	struct adapter *adap = pi->adapter;
+	struct sge_eth_rxq *rxq = &adap->sge.ethrxq[pi->first_qset];
+	struct sge_eth_txq *txq = &adap->sge.ethtxq[pi->first_qset];
+
+	for (i = 0; i < pi->n_rx_qsets; i++, rxq++) {
+		if (rxq->rspq.desc)
+			t4_sge_eth_rxq_stop(adap, &rxq->rspq);
+	}
+	for (i = 0; i < pi->n_tx_qsets; i++, txq++) {
+		if (txq->q.desc) {
+			struct sge_txq *q = &txq->q;
+
+			t4_sge_eth_txq_stop(txq);
+			reclaim_completed_tx(q);
+			free_tx_desc(q, q->size);
+			q->equeidx = q->pidx;
+		}
+	}
+}
+
+void t4_sge_eth_rxq_release(struct adapter *adap, struct sge_eth_rxq *rxq)
+{
+	if (rxq->rspq.desc) {
+		t4_sge_eth_rxq_stop(adap, &rxq->rspq);
+		free_rspq_fl(adap, &rxq->rspq, rxq->fl.size ? &rxq->fl : NULL);
+	}
+}
+
+void t4_sge_eth_txq_release(struct adapter *adap, struct sge_eth_txq *txq)
+{
+	if (txq->q.desc) {
+		t4_sge_eth_txq_stop(txq);
+		reclaim_completed_tx(&txq->q);
+		t4_eth_eq_free(adap, adap->mbox, adap->pf, 0, txq->q.cntxt_id);
+		free_tx_desc(&txq->q, txq->q.size);
+		rte_free(txq->q.sdesc);
+		free_txq(&txq->q);
+	}
+}
+
+void t4_sge_tx_monitor_start(struct adapter *adap)
+{
+	rte_eal_alarm_set(50, tx_timer_cb, (void *)adap);
+}
+
+void t4_sge_tx_monitor_stop(struct adapter *adap)
+{
+	rte_eal_alarm_cancel(tx_timer_cb, (void *)adap);
+}
+
+/**
+ * t4_free_sge_resources - free SGE resources
+ * @adap: the adapter
+ *
+ * Frees resources used by the SGE queue sets.
+ */
+void t4_free_sge_resources(struct adapter *adap)
+{
+	int i;
+	struct sge_eth_rxq *rxq = &adap->sge.ethrxq[0];
+	struct sge_eth_txq *txq = &adap->sge.ethtxq[0];
+
+	/* clean up Ethernet Tx/Rx queues */
+	for (i = 0; i < adap->sge.max_ethqsets; i++, rxq++, txq++) {
+		/* Free only the queues allocated */
+		if (rxq->rspq.desc) {
+			t4_sge_eth_rxq_release(adap, rxq);
+			rxq->rspq.eth_dev = NULL;
+		}
+		if (txq->q.desc) {
+			t4_sge_eth_txq_release(adap, txq);
+			txq->eth_dev = NULL;
+		}
+	}
+
+	if (adap->sge.fw_evtq.desc)
+		free_rspq_fl(adap, &adap->sge.fw_evtq, NULL);
+}
+
+/**
+ * t4_sge_init - initialize SGE
+ * @adap: the adapter
+ *
+ * Performs SGE initialization needed every time after a chip reset.
+ * We do not initialize any of the queues here, instead the driver
+ * top-level must request those individually.
+ *
+ * Called in two different modes:
+ *
+ *  1. Perform actual hardware initialization and record hard-coded
+ *     parameters which were used.  This gets used when we're the
+ *     Master PF and the Firmware Configuration File support didn't
+ *     work for some reason.
+ *
+ *  2. We're not the Master PF or initialization was performed with
+ *     a Firmware Configuration File.  In this case we need to grab
+ *     any of the SGE operating parameters that we need to have in
+ *     order to do our job and make sure we can live with them ...
+ */
+static int t4_sge_init_soft(struct adapter *adap)
+{
+	struct sge *s = &adap->sge;
+	u32 fl_small_pg, fl_large_pg, fl_small_mtu, fl_large_mtu;
+	u32 timer_value_0_and_1, timer_value_2_and_3, timer_value_4_and_5;
+	u32 ingress_rx_threshold;
+
+	/*
+	 * Verify that CPL messages are going to the Ingress Queue for
+	 * process_responses() and that only packet data is going to the
+	 * Free Lists.
+	 */
+	if ((t4_read_reg(adap, A_SGE_CONTROL) & F_RXPKTCPLMODE) !=
+	    V_RXPKTCPLMODE(X_RXPKTCPLMODE_SPLIT)) {
+		dev_err(adap, "bad SGE CPL MODE\n");
+		return -EINVAL;
+	}
+
+	/*
+	 * Validate the Host Buffer Register Array indices that we want to
+	 * use ...
+	 *
+	 * XXX Note that we should really read through the Host Buffer Size
+	 * XXX register array and find the indices of the Buffer Sizes which
+	 * XXX meet our needs!
+	 */
+#define READ_FL_BUF(x) \
+	t4_read_reg(adap, A_SGE_FL_BUFFER_SIZE0 + (x) * sizeof(u32))
+
+	fl_small_pg = READ_FL_BUF(RX_SMALL_PG_BUF);
+	fl_large_pg = READ_FL_BUF(RX_LARGE_PG_BUF);
+	fl_small_mtu = READ_FL_BUF(RX_SMALL_MTU_BUF);
+	fl_large_mtu = READ_FL_BUF(RX_LARGE_MTU_BUF);
+
+	/*
+	 * We only bother using the Large Page logic if the Large Page Buffer
+	 * is larger than our Page Size Buffer.
+	 */
+	if (fl_large_pg <= fl_small_pg)
+		fl_large_pg = 0;
+
+#undef READ_FL_BUF
+
+	/*
+	 * The Page Size Buffer must be exactly equal to our Page Size and the
+	 * Large Page Size Buffer should be 0 (per above) or a power of 2.
+	 */
+	if (fl_small_pg != CXGBE_PAGE_SIZE ||
+	    (fl_large_pg & (fl_large_pg - 1)) != 0) {
+		dev_err(adap, "bad SGE FL page buffer sizes [%d, %d]\n",
+			fl_small_pg, fl_large_pg);
+		return -EINVAL;
+	}
+	if (fl_large_pg)
+		s->fl_pg_order = ilog2(fl_large_pg) - PAGE_SHIFT;
+
+	if (adap->use_unpacked_mode) {
+		int err = 0;
+
+		if (fl_small_mtu < FL_MTU_SMALL_BUFSIZE(adap)) {
+			dev_err(adap, "bad SGE FL small MTU %d\n",
+				fl_small_mtu);
+			err = -EINVAL;
+		}
+		if (fl_large_mtu < FL_MTU_LARGE_BUFSIZE(adap)) {
+			dev_err(adap, "bad SGE FL large MTU %d\n",
+				fl_large_mtu);
+			err = -EINVAL;
+		}
+		if (err)
+			return err;
+	}
+
+	/*
+	 * Retrieve our RX interrupt holdoff timer values and counter
+	 * threshold values from the SGE parameters.
+	 */
+	timer_value_0_and_1 = t4_read_reg(adap, A_SGE_TIMER_VALUE_0_AND_1);
+	timer_value_2_and_3 = t4_read_reg(adap, A_SGE_TIMER_VALUE_2_AND_3);
+	timer_value_4_and_5 = t4_read_reg(adap, A_SGE_TIMER_VALUE_4_AND_5);
+	s->timer_val[0] = core_ticks_to_us(adap,
+					   G_TIMERVALUE0(timer_value_0_and_1));
+	s->timer_val[1] = core_ticks_to_us(adap,
+					   G_TIMERVALUE1(timer_value_0_and_1));
+	s->timer_val[2] = core_ticks_to_us(adap,
+					   G_TIMERVALUE2(timer_value_2_and_3));
+	s->timer_val[3] = core_ticks_to_us(adap,
+					   G_TIMERVALUE3(timer_value_2_and_3));
+	s->timer_val[4] = core_ticks_to_us(adap,
+					   G_TIMERVALUE4(timer_value_4_and_5));
+	s->timer_val[5] = core_ticks_to_us(adap,
+					   G_TIMERVALUE5(timer_value_4_and_5));
+
+	ingress_rx_threshold = t4_read_reg(adap, A_SGE_INGRESS_RX_THRESHOLD);
+	s->counter_val[0] = G_THRESHOLD_0(ingress_rx_threshold);
+	s->counter_val[1] = G_THRESHOLD_1(ingress_rx_threshold);
+	s->counter_val[2] = G_THRESHOLD_2(ingress_rx_threshold);
+	s->counter_val[3] = G_THRESHOLD_3(ingress_rx_threshold);
+
+	return 0;
+}
+
+int t4_sge_init(struct adapter *adap)
+{
+	struct sge *s = &adap->sge;
+	u32 sge_control, sge_control2, sge_conm_ctrl;
+	unsigned int ingpadboundary, ingpackboundary;
+	int ret, egress_threshold;
+
+	/*
+	 * Ingress Padding Boundary and Egress Status Page Size are set up by
+	 * t4_fixup_host_params().
+	 */
+	sge_control = t4_read_reg(adap, A_SGE_CONTROL);
+	s->pktshift = G_PKTSHIFT(sge_control);
+	s->stat_len = (sge_control & F_EGRSTATUSPAGESIZE) ? 128 : 64;
+
+	/*
+	 * 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.
+	 */
+	ingpadboundary = 1 << (G_INGPADBOUNDARY(sge_control) +
+			 X_INGPADBOUNDARY_SHIFT);
+	s->fl_align = ingpadboundary;
+
+	if (!is_t4(adap->params.chip) && !adap->use_unpacked_mode) {
+		/*
+		 * T5 has a weird interpretation of one of the PCIe Packing
+		 * Boundary values.  No idea why ...
+		 */
+		sge_control2 = t4_read_reg(adap, A_SGE_CONTROL2);
+		ingpackboundary = G_INGPACKBOUNDARY(sge_control2);
+		if (ingpackboundary == X_INGPACKBOUNDARY_16B)
+			ingpackboundary = 16;
+		else
+			ingpackboundary = 1 << (ingpackboundary +
+					  X_INGPACKBOUNDARY_SHIFT);
+
+		s->fl_align = max(ingpadboundary, ingpackboundary);
+	}
+
+	ret = t4_sge_init_soft(adap);
+	if (ret < 0) {
+		dev_err(adap, "%s: t4_sge_init_soft failed, error %d\n",
+			__func__, -ret);
+		return ret;
+	}
+
+	/*
+	 * 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.)  For T4,
+	 * there was only a single field to control this.  For T5 there's the
+	 * original field which now only applies to Unpacked Mode Free List
+	 * buffers and a new field which only applies to Packed Mode Free List
+	 * buffers.
+	 */
+	sge_conm_ctrl = t4_read_reg(adap, A_SGE_CONM_CTRL);
+	if (is_t4(adap->params.chip) || adap->use_unpacked_mode)
+		egress_threshold = G_EGRTHRESHOLD(sge_conm_ctrl);
+	else
+		egress_threshold = G_EGRTHRESHOLDPACKING(sge_conm_ctrl);
+	s->fl_starve_thres = 2 * egress_threshold + 1;
+
+	return 0;
+}