cxgb4: Add packet queues and packet DMA code

Signed-off-by: Dimitris Michailidis <dm@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 2431 insertions, 0 deletions

diff --git a/drivers/net/cxgb4/sge.c b/drivers/net/cxgb4/sge.c
new file mode 100644
index 000000000000..14adc58e71c3
--- /dev/null
+++ b/drivers/net/cxgb4/sge.c
@@ -0,0 +1,2431 @@
+/*
+ * This file is part of the Chelsio T4 Ethernet driver for Linux.
+ *
+ * Copyright (c) 2003-2010 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 <linux/dma-mapping.h>
+#include <linux/jiffies.h>
+#include <net/ipv6.h>
+#include <net/tcp.h>
+#include "cxgb4.h"
+#include "t4_regs.h"
+#include "t4_msg.h"
+#include "t4fw_api.h"
+
+/*
+ * Rx buffer size.  We use largish buffers if possible but settle for single
+ * pages under memory shortage.
+ */
+#if PAGE_SHIFT >= 16
+# define FL_PG_ORDER 0
+#else
+# define FL_PG_ORDER (16 - PAGE_SHIFT)
+#endif
+
+/* RX_PULL_LEN should be <= RX_COPY_THRES */
+#define RX_COPY_THRES    256
+#define 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.
+ */
+#define RX_PKT_SKB_LEN   512
+
+/* Ethernet header padding prepended to RX_PKTs */
+#define RX_PKT_PAD 2
+
+/*
+ * 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.  Must be >= 2 * TXQ_STOP_THRES.
+ */
+#define MAX_TX_RECLAIM 16
+
+/*
+ * Max number of Rx buffers we replenish at a time.  Again keep this modest,
+ * allocating buffers isn't cheap either.
+ */
+#define MAX_RX_REFILL 16U
+
+/*
+ * 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.
+ */
+#define RX_QCHECK_PERIOD (HZ / 2)
+
+/*
+ * Period of the Tx queue check timer.
+ */
+#define TX_QCHECK_PERIOD (HZ / 2)
+
+/*
+ * Max number of Tx descriptors to be reclaimed by the Tx timer.
+ */
+#define MAX_TIMER_TX_RECLAIM 100
+
+/*
+ * Timer index used when backing off due to memory shortage.
+ */
+#define NOMEM_TMR_IDX (SGE_NTIMERS - 1)
+
+/*
+ * An FL with <= FL_STARVE_THRES buffers is starving and a periodic timer will
+ * attempt to refill it.
+ */
+#define FL_STARVE_THRES 4
+
+/*
+ * Suspend an Ethernet Tx queue with fewer available descriptors than this.
+ * This is the same as calc_tx_descs() for a TSO packet with
+ * nr_frags == MAX_SKB_FRAGS.
+ */
+#define ETHTXQ_STOP_THRES \
+	(1 + DIV_ROUND_UP((3 * MAX_SKB_FRAGS) / 2 + (MAX_SKB_FRAGS & 1), 8))
+
+/*
+ * Suspension threshold for non-Ethernet Tx queues.  We require enough room
+ * for a full sized WR.
+ */
+#define TXQ_STOP_THRES (SGE_MAX_WR_LEN / sizeof(struct tx_desc))
+
+/*
+ * Max Tx descriptor space we allow for an Ethernet packet to be inlined
+ * into a WR.
+ */
+#define MAX_IMM_TX_PKT_LEN 128
+
+/*
+ * Max size of a WR sent through a control Tx queue.
+ */
+#define MAX_CTRL_WR_LEN SGE_MAX_WR_LEN
+
+enum {
+	/* packet alignment in FL buffers */
+	FL_ALIGN = L1_CACHE_BYTES < 32 ? 32 : L1_CACHE_BYTES,
+	/* egress status entry size */
+	STAT_LEN = L1_CACHE_BYTES > 64 ? 128 : 64
+};
+
+struct tx_sw_desc {                /* SW state per Tx descriptor */
+	struct sk_buff *skb;
+	struct ulptx_sgl *sgl;
+};
+
+struct rx_sw_desc {                /* SW state per Rx descriptor */
+	struct page *page;
+	dma_addr_t dma_addr;
+};
+
+/*
+ * The low bits of rx_sw_desc.dma_addr have special meaning.
+ */
+enum {
+	RX_LARGE_BUF    = 1 << 0, /* buffer is larger than PAGE_SIZE */
+	RX_UNMAPPED_BUF = 1 << 1, /* buffer is not mapped */
+};
+
+static inline dma_addr_t get_buf_addr(const struct rx_sw_desc *d)
+{
+	return d->dma_addr & ~(dma_addr_t)(RX_LARGE_BUF | RX_UNMAPPED_BUF);
+}
+
+static inline bool is_buf_mapped(const struct rx_sw_desc *d)
+{
+	return !(d->dma_addr & RX_UNMAPPED_BUF);
+}
+
+/**
+ *	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;
+}
+
+/**
+ *	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 */
+}
+
+static inline bool fl_starving(const struct sge_fl *fl)
+{
+	return fl->avail - fl->pend_cred <= FL_STARVE_THRES;
+}
+
+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 = dma_map_page(dev, fp->page, fp->page_offset, fp->size,
+				       DMA_TO_DEVICE);
+		if (dma_mapping_error(dev, *addr))
+			goto unwind;
+	}
+	return 0;
+
+unwind:
+	while (fp-- > si->frags)
+		dma_unmap_page(dev, *--addr, fp->size, DMA_TO_DEVICE);
+
+	dma_unmap_single(dev, addr[-1], skb_headlen(skb), DMA_TO_DEVICE);
+out_err:
+	return -ENOMEM;
+}
+
+#ifdef CONFIG_NEED_DMA_MAP_STATE
+static void unmap_skb(struct device *dev, const struct sk_buff *skb,
+		      const dma_addr_t *addr)
+{
+	const skb_frag_t *fp, *end;
+	const struct skb_shared_info *si;
+
+	dma_unmap_single(dev, *addr++, skb_headlen(skb), DMA_TO_DEVICE);
+
+	si = skb_shinfo(skb);
+	end = &si->frags[si->nr_frags];
+	for (fp = si->frags; fp < end; fp++)
+		dma_unmap_page(dev, *addr++, fp->size, DMA_TO_DEVICE);
+}
+
+/**
+ *	deferred_unmap_destructor - unmap a packet when it is freed
+ *	@skb: the packet
+ *
+ *	This is the packet destructor used for Tx packets that need to remain
+ *	mapped until they are freed rather than until their Tx descriptors are
+ *	freed.
+ */
+static void deferred_unmap_destructor(struct sk_buff *skb)
+{
+	unmap_skb(skb->dev->dev.parent, skb, (dma_addr_t *)skb->head);
+}
+#endif
+
+static void unmap_sgl(struct device *dev, const struct sk_buff *skb,
+		      const struct ulptx_sgl *sgl, const struct sge_txq *q)
+{
+	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), ntohl(sgl->len0),
+				 DMA_TO_DEVICE);
+	else {
+		dma_unmap_page(dev, be64_to_cpu(sgl->addr0), ntohl(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 *)q->stat)) {
+unmap:			dma_unmap_page(dev, be64_to_cpu(p->addr[0]),
+				       ntohl(p->len[0]), DMA_TO_DEVICE);
+			dma_unmap_page(dev, be64_to_cpu(p->addr[1]),
+				       ntohl(p->len[1]), DMA_TO_DEVICE);
+			p++;
+		} else if ((u8 *)p == (u8 *)q->stat) {
+			p = (const struct ulptx_sge_pair *)q->desc;
+			goto unmap;
+		} else if ((u8 *)p + 8 == (u8 *)q->stat) {
+			const __be64 *addr = (const __be64 *)q->desc;
+
+			dma_unmap_page(dev, be64_to_cpu(addr[0]),
+				       ntohl(p->len[0]), DMA_TO_DEVICE);
+			dma_unmap_page(dev, be64_to_cpu(addr[1]),
+				       ntohl(p->len[1]), DMA_TO_DEVICE);
+			p = (const struct ulptx_sge_pair *)&addr[2];
+		} else {
+			const __be64 *addr = (const __be64 *)q->desc;
+
+			dma_unmap_page(dev, be64_to_cpu(p->addr[0]),
+				       ntohl(p->len[0]), DMA_TO_DEVICE);
+			dma_unmap_page(dev, be64_to_cpu(addr[0]),
+				       ntohl(p->len[1]), DMA_TO_DEVICE);
+			p = (const struct ulptx_sge_pair *)&addr[1];
+		}
+	}
+	if (nfrags) {
+		__be64 addr;
+
+		if ((u8 *)p == (u8 *)q->stat)
+			p = (const struct ulptx_sge_pair *)q->desc;
+		addr = (u8 *)p + 16 <= (u8 *)q->stat ? p->addr[0] :
+						       *(const __be64 *)q->desc;
+		dma_unmap_page(dev, be64_to_cpu(addr), ntohl(p->len[0]),
+			       DMA_TO_DEVICE);
+	}
+}
+
+/**
+ *	free_tx_desc - reclaims Tx descriptors and their buffers
+ *	@adapter: the adapter
+ *	@q: 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 *adap, struct sge_txq *q,
+			 unsigned int n, bool unmap)
+{
+	struct tx_sw_desc *d;
+	unsigned int cidx = q->cidx;
+	struct device *dev = adap->pdev_dev;
+
+	d = &q->sdesc[cidx];
+	while (n--) {
+		if (d->skb) {                       /* an SGL is present */
+			if (unmap)
+				unmap_sgl(dev, d->skb, d->sgl, q);
+			kfree_skb(d->skb);
+			d->skb = NULL;
+		}
+		++d;
+		if (++cidx == q->size) {
+			cidx = 0;
+			d = q->sdesc;
+		}
+	}
+	q->cidx = cidx;
+}
+
+/*
+ * 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;
+	return hw_cidx < 0 ? hw_cidx + q->size : hw_cidx;
+}
+
+/**
+ *	reclaim_completed_tx - reclaims completed Tx descriptors
+ *	@adap: the adapter
+ *	@q: 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 *adap, struct sge_txq *q,
+					bool unmap)
+{
+	int avail = reclaimable(q);
+
+	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(adap, q, avail, unmap);
+		q->in_use -= avail;
+	}
+}
+
+static inline int get_buf_size(const struct rx_sw_desc *d)
+{
+#if FL_PG_ORDER > 0
+	return (d->dma_addr & RX_LARGE_BUF) ? (PAGE_SIZE << FL_PG_ORDER) :
+					      PAGE_SIZE;
+#else
+	return PAGE_SIZE;
+#endif
+}
+
+/**
+ *	free_rx_bufs - free the Rx buffers on an SGE free list
+ *	@adap: the adapter
+ *	@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 adapter *adap, struct sge_fl *q, int n)
+{
+	while (n--) {
+		struct rx_sw_desc *d = &q->sdesc[q->cidx];
+
+		if (is_buf_mapped(d))
+			dma_unmap_page(adap->pdev_dev, get_buf_addr(d),
+				       get_buf_size(d), PCI_DMA_FROMDEVICE);
+		put_page(d->page);
+		d->page = NULL;
+		if (++q->cidx == q->size)
+			q->cidx = 0;
+		q->avail--;
+	}
+}
+
+/**
+ *	unmap_rx_buf - unmap the current Rx buffer on an SGE free list
+ *	@adap: the adapter
+ *	@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 adapter *adap, struct sge_fl *q)
+{
+	struct rx_sw_desc *d = &q->sdesc[q->cidx];
+
+	if (is_buf_mapped(d))
+		dma_unmap_page(adap->pdev_dev, get_buf_addr(d),
+			       get_buf_size(d), PCI_DMA_FROMDEVICE);
+	d->page = NULL;
+	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 >= 8) {
+		wmb();
+		t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL), DBPRIO |
+			     QID(q->cntxt_id) | PIDX(q->pend_cred / 8));
+		q->pend_cred &= 7;
+	}
+}
+
+static inline void set_rx_sw_desc(struct rx_sw_desc *sd, struct page *pg,
+				  dma_addr_t mapping)
+{
+	sd->page = pg;
+	sd->dma_addr = mapping;      /* includes size low bits */
+}
+
+/**
+ *	refill_fl - refill an SGE Rx buffer ring
+ *	@adap: the adapter
+ *	@q: the 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.  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(struct adapter *adap, struct sge_fl *q, int n,
+			      gfp_t gfp)
+{
+	struct page *pg;
+	dma_addr_t mapping;
+	unsigned int cred = q->avail;
+	__be64 *d = &q->desc[q->pidx];
+	struct rx_sw_desc *sd = &q->sdesc[q->pidx];
+
+	gfp |= __GFP_NOWARN;         /* failures are expected */
+
+#if FL_PG_ORDER > 0
+	/*
+	 * Prefer large buffers
+	 */
+	while (n) {
+		pg = alloc_pages(gfp | __GFP_COMP, FL_PG_ORDER);
+		if (unlikely(!pg)) {
+			q->large_alloc_failed++;
+			break;       /* fall back to single pages */
+		}
+
+		mapping = dma_map_page(adap->pdev_dev, pg, 0,
+				       PAGE_SIZE << FL_PG_ORDER,
+				       PCI_DMA_FROMDEVICE);
+		if (unlikely(dma_mapping_error(adap->pdev_dev, mapping))) {
+			__free_pages(pg, FL_PG_ORDER);
+			goto out;   /* do not try small pages for this error */
+		}
+		mapping |= RX_LARGE_BUF;
+		*d++ = cpu_to_be64(mapping);
+
+		set_rx_sw_desc(sd, pg, mapping);
+		sd++;
+
+		q->avail++;
+		if (++q->pidx == q->size) {
+			q->pidx = 0;
+			sd = q->sdesc;
+			d = q->desc;
+		}
+		n--;
+	}
+#endif
+
+	while (n--) {
+		pg = __netdev_alloc_page(adap->port[0], gfp);
+		if (unlikely(!pg)) {
+			q->alloc_failed++;
+			break;
+		}
+
+		mapping = dma_map_page(adap->pdev_dev, pg, 0, PAGE_SIZE,
+				       PCI_DMA_FROMDEVICE);
+		if (unlikely(dma_mapping_error(adap->pdev_dev, mapping))) {
+			netdev_free_page(adap->port[0], pg);
+			goto out;
+		}
+		*d++ = cpu_to_be64(mapping);
+
+		set_rx_sw_desc(sd, pg, 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(q))) {
+		smp_wmb();
+		set_bit(q->cntxt_id, adap->sge.starving_fl);
+	}
+
+	return cred;
+}
+
+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),
+		  GFP_ATOMIC);
+}
+
+/**
+ *	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
+ *
+ *	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(struct device *dev, size_t nelem, size_t elem_size,
+			size_t sw_size, dma_addr_t *phys, void *metadata,
+			size_t stat_size)
+{
+	size_t len = nelem * elem_size + stat_size;
+	void *s = NULL;
+	void *p = dma_alloc_coherent(dev, len, phys, GFP_KERNEL);
+
+	if (!p)
+		return NULL;
+	if (sw_size) {
+		s = kcalloc(nelem, sw_size, GFP_KERNEL);
+
+		if (!s) {
+			dma_free_coherent(dev, len, p, *phys);
+			return NULL;
+		}
+	}
+	if (metadata)
+		*(void **)metadata = s;
+	memset(p, 0, len);
+	return p;
+}
+
+/**
+ *	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)
+{
+	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)
+{
+	BUG_ON(n > SGE_MAX_WR_LEN / 8);
+	return DIV_ROUND_UP(n, 8);
+}
+
+/**
+ *	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 as
+ *	immediate data.
+ */
+static inline int is_eth_imm(const struct sk_buff *skb)
+{
+	return skb->len <= MAX_IMM_TX_PKT_LEN - sizeof(struct cpl_tx_pkt);
+}
+
+/**
+ *	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 WR 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 (is_eth_imm(skb))
+		return DIV_ROUND_UP(skb->len + sizeof(struct cpl_tx_pkt), 8);
+
+	flits = sgl_len(skb_shinfo(skb)->nr_frags + 1) + 4;
+	if (skb_shinfo(skb)->gso_size)
+		flits += 2;
+	return flits;
+}
+
+/**
+ *	calc_tx_descs - calculate the number of Tx descriptors for a packet
+ *	@skb: the packet
+ *
+ *	Returns the number of Tx descriptors needed for the given Ethernet
+ *	packet, including the needed WR and CPL headers.
+ */
+static inline unsigned int calc_tx_descs(const struct sk_buff *skb)
+{
+	return flits_to_desc(calc_tx_flits(skb));
+}
+
+/**
+ *	write_sgl - populate a scatter/gather list for a packet
+ *	@skb: 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 skb main-body data to include in the SGL
+ *	@addr: the list of bus addresses for the SGL elements
+ *
+ *	Generates a 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
+ *	right after the end of the SGL but does not account for any potential
+ *	wrap around, i.e., @end > @sgl.
+ */
+static void write_sgl(const struct sk_buff *skb, 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;
+	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(si->frags[0].size);
+		sgl->addr0 = cpu_to_be64(addr[1]);
+	}
+
+	sgl->cmd_nsge = htonl(ULPTX_CMD(ULP_TX_SC_DSGL) | 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 = (nfrags != si->nr_frags); nfrags >= 2; nfrags -= 2, to++) {
+		to->len[0] = cpu_to_be32(si->frags[i].size);
+		to->len[1] = cpu_to_be32(si->frags[++i].size);
+		to->addr[0] = cpu_to_be64(addr[i]);
+		to->addr[1] = cpu_to_be64(addr[++i]);
+	}
+	if (nfrags) {
+		to->len[0] = cpu_to_be32(si->frags[i].size);
+		to->len[1] = cpu_to_be32(0);
+		to->addr[0] = cpu_to_be64(addr[i + 1]);
+	}
+	if (unlikely((u8 *)end > (u8 *)q->stat)) {
+		unsigned int part0 = (u8 *)q->stat - (u8 *)sgl->sge, part1;
+
+		if (likely(part0))
+			memcpy(sgl->sge, buf, part0);
+		part1 = (u8 *)end - (u8 *)q->stat;
+		memcpy(q->desc, (u8 *)buf + part0, part1);
+		end = (void *)q->desc + part1;
+	}
+	if ((uintptr_t)end & 8)           /* 0-pad to multiple of 16 */
+		*(u64 *)end = 0;
+}
+
+/**
+ *	ring_tx_db - check and potentially 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)
+{
+	wmb();            /* write descriptors before telling HW */
+	t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL),
+		     QID(q->cntxt_id) | PIDX(n));
+}
+
+/**
+ *	inline_tx_skb - inline a packet's data into Tx descriptors
+ *	@skb: the packet
+ *	@q: the Tx queue where the packet will be inlined
+ *	@pos: starting position in the Tx queue where 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 *q,
+			  void *pos)
+{
+	u64 *p;
+	int left = (void *)q->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, q->desc, skb->len - left);
+		pos = (void *)q->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);
+	}
+}
+
+static void eth_txq_stop(struct sge_eth_txq *q)
+{
+	netif_tx_stop_queue(q->txq);
+	q->q.stops++;
+}
+
+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;
+}
+
+/**
+ *	t4_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.
+ */
+netdev_tx_t t4_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 *adap;
+	struct sge_eth_txq *q;
+	const struct port_info *pi;
+	struct fw_eth_tx_pkt_wr *wr;
+	struct cpl_tx_pkt_core *cpl;
+	const struct skb_shared_info *ssi;
+	dma_addr_t addr[MAX_SKB_FRAGS + 1];
+
+	/*
+	 * The chip min packet length is 10 octets but play safe and reject
+	 * anything shorter than an Ethernet header.
+	 */
+	if (unlikely(skb->len < ETH_HLEN)) {
+out_free:	dev_kfree_skb(skb);
+		return NETDEV_TX_OK;
+	}
+
+	pi = netdev_priv(dev);
+	adap = pi->adapter;
+	qidx = skb_get_queue_mapping(skb);
+	q = &adap->sge.ethtxq[qidx + pi->first_qset];
+
+	reclaim_completed_tx(adap, &q->q, true);
+
+	flits = calc_tx_flits(skb);
+	ndesc = flits_to_desc(flits);
+	credits = txq_avail(&q->q) - ndesc;
+
+	if (unlikely(credits < 0)) {
+		eth_txq_stop(q);
+		dev_err(adap->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(adap->pdev_dev, skb, addr) < 0)) {
+		q->mapping_err++;
+		goto out_free;
+	}
+
+	wr_mid = FW_WR_LEN16(DIV_ROUND_UP(flits, 2));
+	if (unlikely(credits < ETHTXQ_STOP_THRES)) {
+		eth_txq_stop(q);
+		wr_mid |= FW_WR_EQUEQ | FW_WR_EQUIQ;
+	}
+
+	wr = (void *)&q->q.desc[q->q.pidx];
+	wr->equiq_to_len16 = htonl(wr_mid);
+	wr->r3 = cpu_to_be64(0);
+	end = (u64 *)wr + flits;
+
+	ssi = skb_shinfo(skb);
+	if (ssi->gso_size) {
+		struct cpl_tx_pkt_lso *lso = (void *)wr;
+		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 = htonl(FW_WR_OP(FW_ETH_TX_PKT_WR) |
+				       FW_WR_IMMDLEN(sizeof(*lso)));
+		lso->lso_ctrl = htonl(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 = htons(0);
+		lso->mss = htons(ssi->gso_size);
+		lso->seqno_offset = htonl(0);
+		lso->len = htonl(skb->len);
+		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);
+		q->tso++;
+		q->tx_cso += ssi->gso_segs;
+	} else {
+		int len;
+
+		len = is_eth_imm(skb) ? skb->len + sizeof(*cpl) : sizeof(*cpl);
+		wr->op_immdlen = htonl(FW_WR_OP(FW_ETH_TX_PKT_WR) |
+				       FW_WR_IMMDLEN(len));
+		cpl = (void *)(wr + 1);
+		if (skb->ip_summed == CHECKSUM_PARTIAL) {
+			cntrl = hwcsum(skb) | TXPKT_IPCSUM_DIS;
+			q->tx_cso++;
+		} else
+			cntrl = TXPKT_L4CSUM_DIS | TXPKT_IPCSUM_DIS;
+	}
+
+	if (vlan_tx_tag_present(skb)) {
+		q->vlan_ins++;
+		cntrl |= TXPKT_VLAN_VLD | TXPKT_VLAN(vlan_tx_tag_get(skb));
+	}
+
+	cpl->ctrl0 = htonl(TXPKT_OPCODE(CPL_TX_PKT_XT) |
+			   TXPKT_INTF(pi->tx_chan) | TXPKT_PF(0));
+	cpl->pack = htons(0);
+	cpl->len = htons(skb->len);
+	cpl->ctrl1 = cpu_to_be64(cntrl);
+
+	if (is_eth_imm(skb)) {
+		inline_tx_skb(skb, &q->q, cpl + 1);
+		dev_kfree_skb(skb);
+	} else {
+		int last_desc;
+
+		write_sgl(skb, &q->q, (struct ulptx_sgl *)(cpl + 1), end, 0,
+			  addr);
+		skb_orphan(skb);
+
+		last_desc = q->q.pidx + ndesc - 1;
+		if (last_desc >= q->q.size)
+			last_desc -= q->q.size;
+		q->q.sdesc[last_desc].skb = skb;
+		q->q.sdesc[last_desc].sgl = (struct ulptx_sgl *)(cpl + 1);
+	}
+
+	txq_advance(&q->q, ndesc);
+
+	ring_tx_db(adap, &q->q, ndesc);
+	return NETDEV_TX_OK;
+}
+
+/**
+ *	reclaim_completed_tx_imm - reclaim completed control-queue Tx descs
+ *	@q: the SGE control Tx queue
+ *
+ *	This is a variant of reclaim_completed_tx() that is used for Tx queues
+ *	that send only immediate data (presently just the control queues) and
+ *	thus do not have any sk_buffs to release.
+ */
+static inline void reclaim_completed_tx_imm(struct sge_txq *q)
+{
+	int hw_cidx = ntohs(q->stat->cidx);
+	int reclaim = hw_cidx - q->cidx;
+
+	if (reclaim < 0)
+		reclaim += q->size;
+
+	q->in_use -= reclaim;
+	q->cidx = hw_cidx;
+}
+
+/**
+ *	is_imm - check whether a packet can be sent as immediate data
+ *	@skb: the packet
+ *
+ *	Returns true if a packet can be sent as a WR with immediate data.
+ */
+static inline int is_imm(const struct sk_buff *skb)
+{
+	return skb->len <= MAX_CTRL_WR_LEN;
+}
+
+/**
+ *	ctrlq_check_stop - check if a control queue is full and should stop
+ *	@q: the queue
+ *	@wr: most recent WR written to the queue
+ *
+ *	Check if a control queue has become full and should be stopped.
+ *	We clean up control queue descriptors very lazily, only when we are out.
+ *	If the queue is still full after reclaiming any completed descriptors
+ *	we suspend it and have the last WR wake it up.
+ */
+static void ctrlq_check_stop(struct sge_ctrl_txq *q, struct fw_wr_hdr *wr)
+{
+	reclaim_completed_tx_imm(&q->q);
+	if (unlikely(txq_avail(&q->q) < TXQ_STOP_THRES)) {
+		wr->lo |= htonl(FW_WR_EQUEQ | FW_WR_EQUIQ);
+		q->q.stops++;
+		q->full = 1;
+	}
+}
+
+/**
+ *	ctrl_xmit - send a packet through an SGE control Tx queue
+ *	@q: the control queue
+ *	@skb: the packet
+ *
+ *	Send a packet through an SGE control Tx queue.  Packets sent through
+ *	a control queue must fit entirely as immediate data.
+ */
+static int ctrl_xmit(struct sge_ctrl_txq *q, struct sk_buff *skb)
+{
+	unsigned int ndesc;
+	struct fw_wr_hdr *wr;
+
+	if (unlikely(!is_imm(skb))) {
+		WARN_ON(1);
+		dev_kfree_skb(skb);
+		return NET_XMIT_DROP;
+	}
+
+	ndesc = DIV_ROUND_UP(skb->len, sizeof(struct tx_desc));
+	spin_lock(&q->sendq.lock);
+
+	if (unlikely(q->full)) {
+		skb->priority = ndesc;                  /* save for restart */
+		__skb_queue_tail(&q->sendq, skb);
+		spin_unlock(&q->sendq.lock);
+		return NET_XMIT_CN;
+	}
+
+	wr = (struct fw_wr_hdr *)&q->q.desc[q->q.pidx];
+	inline_tx_skb(skb, &q->q, wr);
+
+	txq_advance(&q->q, ndesc);
+	if (unlikely(txq_avail(&q->q) < TXQ_STOP_THRES))
+		ctrlq_check_stop(q, wr);
+
+	ring_tx_db(q->adap, &q->q, ndesc);
+	spin_unlock(&q->sendq.lock);
+
+	kfree_skb(skb);
+	return NET_XMIT_SUCCESS;
+}
+
+/**
+ *	restart_ctrlq - restart a suspended control queue
+ *	@data: the control queue to restart
+ *
+ *	Resumes transmission on a suspended Tx control queue.
+ */
+static void restart_ctrlq(unsigned long data)
+{
+	struct sk_buff *skb;
+	unsigned int written = 0;
+	struct sge_ctrl_txq *q = (struct sge_ctrl_txq *)data;
+
+	spin_lock(&q->sendq.lock);
+	reclaim_completed_tx_imm(&q->q);
+	BUG_ON(txq_avail(&q->q) < TXQ_STOP_THRES);  /* q should be empty */
+
+	while ((skb = __skb_dequeue(&q->sendq)) != NULL) {
+		struct fw_wr_hdr *wr;
+		unsigned int ndesc = skb->priority;     /* previously saved */
+
+		/*
+		 * Write descriptors and free skbs outside the lock to limit
+		 * wait times.  q->full is still set so new skbs will be queued.
+		 */
+		spin_unlock(&q->sendq.lock);
+
+		wr = (struct fw_wr_hdr *)&q->q.desc[q->q.pidx];
+		inline_tx_skb(skb, &q->q, wr);
+		kfree_skb(skb);
+
+		written += ndesc;
+		txq_advance(&q->q, ndesc);
+		if (unlikely(txq_avail(&q->q) < TXQ_STOP_THRES)) {
+			unsigned long old = q->q.stops;
+
+			ctrlq_check_stop(q, wr);
+			if (q->q.stops != old) {          /* suspended anew */