sge.c

/*
 * Copyright (c) 2005-2008 Chelsio, 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/tcp.h>
#include <linux/dma-mapping.h>
#include <linux/slab.h>
#include <linux/prefetch.h>
#include <net/arp.h>
#include "common.h"
#include "regs.h"
#include "sge_defs.h"
#include "t3_cpl.h"
#include "firmware_exports.h"
#include "cxgb3_offload.h"

#define USE_GTS 0

#define SGE_RX_SM_BUF_SIZE 1536

#define SGE_RX_COPY_THRES  256
#define SGE_RX_PULL_LEN    128

#define SGE_PG_RSVD SMP_CACHE_BYTES
/*
 * Page chunk size for FL0 buffers if FL0 is to be populated with page chunks.
 * It must be a divisor of PAGE_SIZE.  If set to 0 FL0 will use sk_buffs
 * directly.
 */
#define FL0_PG_CHUNK_SIZE  2048
#define FL0_PG_ORDER 0
#define FL0_PG_ALLOC_SIZE (PAGE_SIZE << FL0_PG_ORDER)
#define FL1_PG_CHUNK_SIZE (PAGE_SIZE > 8192 ? 16384 : 8192)
#define FL1_PG_ORDER (PAGE_SIZE > 8192 ? 0 : 1)
#define FL1_PG_ALLOC_SIZE (PAGE_SIZE << FL1_PG_ORDER)

#define SGE_RX_DROP_THRES 16
#define RX_RECLAIM_PERIOD (HZ/4)

/*
 * Max number of Rx buffers we replenish at a time.
 */
#define MAX_RX_REFILL 16U
/*
 * Period of the Tx buffer reclaim timer.  This timer does not need to run
 * frequently as Tx buffers are usually reclaimed by new Tx packets.
 */
#define TX_RECLAIM_PERIOD (HZ / 4)
#define TX_RECLAIM_TIMER_CHUNK 64U
#define TX_RECLAIM_CHUNK 16U

/* WR size in bytes */
#define WR_LEN (WR_FLITS * 8)

/*
 * Types of Tx queues in each queue set.  Order here matters, do not change.
 */
enum { TXQ_ETH, TXQ_OFLD, TXQ_CTRL };

/* Values for sge_txq.flags */
enum {
	TXQ_RUNNING = 1 << 0,	/* fetch engine is running */
	TXQ_LAST_PKT_DB = 1 << 1,	/* last packet rang the doorbell */
};

struct tx_desc {
	__be64 flit[TX_DESC_FLITS];
};

struct rx_desc {
	__be32 addr_lo;
	__be32 len_gen;
	__be32 gen2;
	__be32 addr_hi;
};

struct tx_sw_desc {		/* SW state per Tx descriptor */
	struct sk_buff *skb;
	u8 eop;       /* set if last descriptor for packet */
	u8 addr_idx;  /* buffer index of first SGL entry in descriptor */
	u8 fragidx;   /* first page fragment associated with descriptor */
	s8 sflit;     /* start flit of first SGL entry in descriptor */
};

struct rx_sw_desc {                /* SW state per Rx descriptor */
	union {
		struct sk_buff *skb;
		struct fl_pg_chunk pg_chunk;
	};
	DEFINE_DMA_UNMAP_ADDR(dma_addr);
};

struct rsp_desc {		/* response queue descriptor */
	struct rss_header rss_hdr;
	__be32 flags;
	__be32 len_cq;
	u8 imm_data[47];
	u8 intr_gen;
};

/*
 * Holds unmapping information for Tx packets that need deferred unmapping.
 * This structure lives at skb->head and must be allocated by callers.
 */
struct deferred_unmap_info {
	struct pci_dev *pdev;
	dma_addr_t addr[MAX_SKB_FRAGS + 1];
};

/*
 * Maps a number of flits to the number of Tx descriptors that can hold them.
 * The formula is
 *
 * desc = 1 + (flits - 2) / (WR_FLITS - 1).
 *
 * HW allows up to 4 descriptors to be combined into a WR.
 */
static u8 flit_desc_map[] = {
	0,
#if SGE_NUM_GENBITS == 1
	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
	2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
	3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
	4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4
#elif SGE_NUM_GENBITS == 2
	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
	2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
	3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
	4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
#else
# error "SGE_NUM_GENBITS must be 1 or 2"
#endif
};

static inline struct sge_qset *fl_to_qset(const struct sge_fl *q, int qidx)
{
	return container_of(q, struct sge_qset, fl[qidx]);
}

static inline struct sge_qset *rspq_to_qset(const struct sge_rspq *q)
{
	return container_of(q, struct sge_qset, rspq);
}

static inline struct sge_qset *txq_to_qset(const struct sge_txq *q, int qidx)
{
	return container_of(q, struct sge_qset, txq[qidx]);
}

/**
 *	refill_rspq - replenish an SGE response queue
 *	@adapter: the adapter
 *	@q: the response queue to replenish
 *	@credits: how many new responses to make available
 *
 *	Replenishes a response queue by making the supplied number of responses
 *	available to HW.
 */
static inline void refill_rspq(struct adapter *adapter,
			       const struct sge_rspq *q, unsigned int credits)
{
	rmb();
	t3_write_reg(adapter, A_SG_RSPQ_CREDIT_RETURN,
		     V_RSPQ(q->cntxt_id) | V_CREDITS(credits));
}

/**
 *	need_skb_unmap - does the platform need unmapping of sk_buffs?
 *
 *	Returns true if the platform needs sk_buff unmapping.  The compiler
 *	optimizes away unnecessary code if this returns true.
 */
static inline int need_skb_unmap(void)
{
#ifdef CONFIG_NEED_DMA_MAP_STATE
	return 1;
#else
	return 0;
#endif
}

/**
 *	unmap_skb - unmap a packet main body and its page fragments
 *	@skb: the packet
 *	@q: the Tx queue containing Tx descriptors for the packet
 *	@cidx: index of Tx descriptor
 *	@pdev: the PCI device
 *
 *	Unmap the main body of an sk_buff and its page fragments, if any.
 *	Because of the fairly complicated structure of our SGLs and the desire
 *	to conserve space for metadata, the information necessary to unmap an
 *	sk_buff is spread across the sk_buff itself (buffer lengths), the HW Tx
 *	descriptors (the physical addresses of the various data buffers), and
 *	the SW descriptor state (assorted indices).  The send functions
 *	initialize the indices for the first packet descriptor so we can unmap
 *	the buffers held in the first Tx descriptor here, and we have enough
 *	information at this point to set the state for the next Tx descriptor.
 *
 *	Note that it is possible to clean up the first descriptor of a packet
 *	before the send routines have written the next descriptors, but this
 *	race does not cause any problem.  We just end up writing the unmapping
 *	info for the descriptor first.
 */
static inline void unmap_skb(struct sk_buff *skb, struct sge_txq *q,
			     unsigned int cidx, struct pci_dev *pdev)
{
	const struct sg_ent *sgp;
	struct tx_sw_desc *d = &q->sdesc[cidx];
	int nfrags, frag_idx, curflit, j = d->addr_idx;

	sgp = (struct sg_ent *)&q->desc[cidx].flit[d->sflit];
	frag_idx = d->fragidx;

	if (frag_idx == 0 && skb_headlen(skb)) {
		pci_unmap_single(pdev, be64_to_cpu(sgp->addr[0]),
				 skb_headlen(skb), PCI_DMA_TODEVICE);
		j = 1;
	}

	curflit = d->sflit + 1 + j;
	nfrags = skb_shinfo(skb)->nr_frags;

	while (frag_idx < nfrags && curflit < WR_FLITS) {
		pci_unmap_page(pdev, be64_to_cpu(sgp->addr[j]),
			       skb_shinfo(skb)->frags[frag_idx].size,
			       PCI_DMA_TODEVICE);
		j ^= 1;
		if (j == 0) {
			sgp++;
			curflit++;
		}
		curflit++;
		frag_idx++;
	}

	if (frag_idx < nfrags) {   /* SGL continues into next Tx descriptor */
		d = cidx + 1 == q->size ? q->sdesc : d + 1;
		d->fragidx = frag_idx;
		d->addr_idx = j;
		d->sflit = curflit - WR_FLITS - j; /* sflit can be -1 */
	}
}

/**
 *	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
 *
 *	Reclaims Tx descriptors from an SGE Tx queue and frees the associated
 *	Tx buffers.  Called with the Tx queue lock held.
 */
static void free_tx_desc(struct adapter *adapter, struct sge_txq *q,
			 unsigned int n)
{
	struct tx_sw_desc *d;
	struct pci_dev *pdev = adapter->pdev;
	unsigned int cidx = q->cidx;

	const int need_unmap = need_skb_unmap() &&
			       q->cntxt_id >= FW_TUNNEL_SGEEC_START;

	d = &q->sdesc[cidx];
	while (n--) {
		if (d->skb) {	/* an SGL is present */
			if (need_unmap)
				unmap_skb(d->skb, q, cidx, pdev);
			if (d->eop) {
				kfree_skb(d->skb);
				d->skb = NULL;
			}
		}
		++d;
		if (++cidx == q->size) {
			cidx = 0;
			d = q->sdesc;
		}
	}
	q->cidx = cidx;
}

/**
 *	reclaim_completed_tx - reclaims completed Tx descriptors
 *	@adapter: the adapter
 *	@q: the Tx queue to reclaim completed descriptors from
 *	@chunk: maximum number of descriptors to reclaim
 *
 *	Reclaims Tx descriptors that the SGE has indicated it has processed,
 *	and frees the associated buffers if possible.  Called with the Tx
 *	queue's lock held.
 */
static inline unsigned int reclaim_completed_tx(struct adapter *adapter,
						struct sge_txq *q,
						unsigned int chunk)
{
	unsigned int reclaim = q->processed - q->cleaned;

	reclaim = min(chunk, reclaim);
	if (reclaim) {
		free_tx_desc(adapter, q, reclaim);
		q->cleaned += reclaim;
		q->in_use -= reclaim;
	}
	return q->processed - q->cleaned;
}

/**
 *	should_restart_tx - are there enough resources to restart a Tx queue?
 *	@q: the Tx queue
 *
 *	Checks if there are enough descriptors to restart a suspended Tx queue.
 */
static inline int should_restart_tx(const struct sge_txq *q)
{
	unsigned int r = q->processed - q->cleaned;

	return q->in_use - r < (q->size >> 1);
}

static void clear_rx_desc(struct pci_dev *pdev, const struct sge_fl *q,
			  struct rx_sw_desc *d)
{
	if (q->use_pages && d->pg_chunk.page) {
		(*d->pg_chunk.p_cnt)--;
		if (!*d->pg_chunk.p_cnt)
			pci_unmap_page(pdev,
				       d->pg_chunk.mapping,
				       q->alloc_size, PCI_DMA_FROMDEVICE);

		put_page(d->pg_chunk.page);
		d->pg_chunk.page = NULL;
	} else {
		pci_unmap_single(pdev, dma_unmap_addr(d, dma_addr),
				 q->buf_size, PCI_DMA_FROMDEVICE);
		kfree_skb(d->skb);
		d->skb = NULL;
	}
}

/**
 *	free_rx_bufs - free the Rx buffers on an SGE free list
 *	@pdev: the PCI device associated with the adapter
 *	@rxq: the SGE free list to clean up
 *
 *	Release the buffers on an SGE free-buffer Rx queue.  HW fetching from
 *	this queue should be stopped before calling this function.
 */
static void free_rx_bufs(struct pci_dev *pdev, struct sge_fl *q)
{
	unsigned int cidx = q->cidx;

	while (q->credits--) {
		struct rx_sw_desc *d = &q->sdesc[cidx];


		clear_rx_desc(pdev, q, d);
		if (++cidx == q->size)
			cidx = 0;
	}

	if (q->pg_chunk.page) {
		__free_pages(q->pg_chunk.page, q->order);
		q->pg_chunk.page = NULL;
	}
}

/**
 *	add_one_rx_buf - add a packet buffer to a free-buffer list
 *	@va:  buffer start VA
 *	@len: the buffer length
 *	@d: the HW Rx descriptor to write
 *	@sd: the SW Rx descriptor to write
 *	@gen: the generation bit value
 *	@pdev: the PCI device associated with the adapter
 *
 *	Add a buffer of the given length to the supplied HW and SW Rx
 *	descriptors.
 */
static inline int add_one_rx_buf(void *va, unsigned int len,
				 struct rx_desc *d, struct rx_sw_desc *sd,
				 unsigned int gen, struct pci_dev *pdev)
{
	dma_addr_t mapping;

	mapping = pci_map_single(pdev, va, len, PCI_DMA_FROMDEVICE);
	if (unlikely(pci_dma_mapping_error(pdev, mapping)))
		return -ENOMEM;

	dma_unmap_addr_set(sd, dma_addr, mapping);

	d->addr_lo = cpu_to_be32(mapping);
	d->addr_hi = cpu_to_be32((u64) mapping >> 32);
	wmb();
	d->len_gen = cpu_to_be32(V_FLD_GEN1(gen));
	d->gen2 = cpu_to_be32(V_FLD_GEN2(gen));
	return 0;
}

static inline int add_one_rx_chunk(dma_addr_t mapping, struct rx_desc *d,
				   unsigned int gen)
{
	d->addr_lo = cpu_to_be32(mapping);
	d->addr_hi = cpu_to_be32((u64) mapping >> 32);
	wmb();
	d->len_gen = cpu_to_be32(V_FLD_GEN1(gen));
	d->gen2 = cpu_to_be32(V_FLD_GEN2(gen));
	return 0;
}

static int alloc_pg_chunk(struct adapter *adapter, struct sge_fl *q,
			  struct rx_sw_desc *sd, gfp_t gfp,
			  unsigned int order)
{
	if (!q->pg_chunk.page) {
		dma_addr_t mapping;

		q->pg_chunk.page = alloc_pages(gfp, order);
		if (unlikely(!q->pg_chunk.page))
			return -ENOMEM;
		q->pg_chunk.va = page_address(q->pg_chunk.page);
		q->pg_chunk.p_cnt = q->pg_chunk.va + (PAGE_SIZE << order) -
				    SGE_PG_RSVD;
		q->pg_chunk.offset = 0;
		mapping = pci_map_page(adapter->pdev, q->pg_chunk.page,
				       0, q->alloc_size, PCI_DMA_FROMDEVICE);
		q->pg_chunk.mapping = mapping;
	}
	sd->pg_chunk = q->pg_chunk;

	prefetch(sd->pg_chunk.p_cnt);

	q->pg_chunk.offset += q->buf_size;
	if (q->pg_chunk.offset == (PAGE_SIZE << order))
		q->pg_chunk.page = NULL;
	else {
		q->pg_chunk.va += q->buf_size;
		get_page(q->pg_chunk.page);
	}

	if (sd->pg_chunk.offset == 0)
		*sd->pg_chunk.p_cnt = 1;
	else
		*sd->pg_chunk.p_cnt += 1;

	return 0;
}

static inline void ring_fl_db(struct adapter *adap, struct sge_fl *q)
{
	if (q->pend_cred >= q->credits / 4) {
		q->pend_cred = 0;
		wmb();
		t3_write_reg(adap, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
	}
}

/**
 *	refill_fl - refill an SGE free-buffer list
 *	@adapter: the adapter
 *	@q: the free-list to refill
 *	@n: the number of new buffers to allocate
 *	@gfp: the gfp flags for allocating new buffers
 *
 *	(Re)populate an SGE free-buffer list 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.
 */
static int refill_fl(struct adapter *adap, struct sge_fl *q, int n, gfp_t gfp)
{
	struct rx_sw_desc *sd = &q->sdesc[q->pidx];
	struct rx_desc *d = &q->desc[q->pidx];
	unsigned int count = 0;

	while (n--) {
		dma_addr_t mapping;
		int err;

		if (q->use_pages) {
			if (unlikely(alloc_pg_chunk(adap, q, sd, gfp,
						    q->order))) {
nomem:				q->alloc_failed++;
				break;
			}
			mapping = sd->pg_chunk.mapping + sd->pg_chunk.offset;
			dma_unmap_addr_set(sd, dma_addr, mapping);

			add_one_rx_chunk(mapping, d, q->gen);
			pci_dma_sync_single_for_device(adap->pdev, mapping,
						q->buf_size - SGE_PG_RSVD,
						PCI_DMA_FROMDEVICE);
		} else {
			void *buf_start;

			struct sk_buff *skb = alloc_skb(q->buf_size, gfp);
			if (!skb)
				goto nomem;

			sd->skb = skb;
			buf_start = skb->data;
			err = add_one_rx_buf(buf_start, q->buf_size, d, sd,
					     q->gen, adap->pdev);
			if (unlikely(err)) {
				clear_rx_desc(adap->pdev, q, sd);
				break;
			}
		}

		d++;
		sd++;
		if (++q->pidx == q->size) {
			q->pidx = 0;
			q->gen ^= 1;
			sd = q->sdesc;
			d = q->desc;
		}
		count++;
	}

	q->credits += count;
	q->pend_cred += count;
	ring_fl_db(adap, q);

	return count;
}

static inline void __refill_fl(struct adapter *adap, struct sge_fl *fl)
{
	refill_fl(adap, fl, min(MAX_RX_REFILL, fl->size - fl->credits),
		  GFP_ATOMIC | __GFP_COMP);
}

/**
 *	recycle_rx_buf - recycle a receive buffer
 *	@adapter: the adapter
 *	@q: the SGE free list
 *	@idx: index of buffer to recycle
 *
 *	Recycles the specified buffer on the given free list by adding it at
 *	the next available slot on the list.
 */
static void recycle_rx_buf(struct adapter *adap, struct sge_fl *q,
			   unsigned int idx)
{
	struct rx_desc *from = &q->desc[idx];
	struct rx_desc *to = &q->desc[q->pidx];

	q->sdesc[q->pidx] = q->sdesc[idx];
	to->addr_lo = from->addr_lo;	/* already big endian */
	to->addr_hi = from->addr_hi;	/* likewise */
	wmb();
	to->len_gen = cpu_to_be32(V_FLD_GEN1(q->gen));
	to->gen2 = cpu_to_be32(V_FLD_GEN2(q->gen));

	if (++q->pidx == q->size) {
		q->pidx = 0;
		q->gen ^= 1;
	}

	q->credits++;
	q->pend_cred++;
	ring_fl_db(adap, q);
}

/**
 *	alloc_ring - allocate resources for an SGE descriptor ring
 *	@pdev: the PCI 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
 *
 *	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 physical address of the HW ring, and the address
 *	of the SW ring.
 */
static void *alloc_ring(struct pci_dev *pdev, size_t nelem, size_t elem_size,
			size_t sw_size, dma_addr_t * phys, void *metadata)
{
	size_t len = nelem * elem_size;
	void *s = NULL;
	void *p = dma_alloc_coherent(&pdev->dev, len, phys, GFP_KERNEL);

	if (!p)
		return NULL;
	if (sw_size && metadata) {
		s = kcalloc(nelem, sw_size, GFP_KERNEL);

		if (!s) {
			dma_free_coherent(&pdev->dev, len, p, *phys);
			return NULL;
		}
		*(void **)metadata = s;
	}
	memset(p, 0, len);
	return p;
}

/**
 *	t3_reset_qset - reset a sge qset
 *	@q: the queue set
 *
 *	Reset the qset structure.
 *	the NAPI structure is preserved in the event of
 *	the qset's reincarnation, for example during EEH recovery.
 */
static void t3_reset_qset(struct sge_qset *q)
{
	if (q->adap &&
	    !(q->adap->flags & NAPI_INIT)) {
		memset(q, 0, sizeof(*q));
		return;
	}

	q->adap = NULL;
	memset(&q->rspq, 0, sizeof(q->rspq));
	memset(q->fl, 0, sizeof(struct sge_fl) * SGE_RXQ_PER_SET);
	memset(q->txq, 0, sizeof(struct sge_txq) * SGE_TXQ_PER_SET);
	q->txq_stopped = 0;
	q->tx_reclaim_timer.function = NULL; /* for t3_stop_sge_timers() */
	q->rx_reclaim_timer.function = NULL;
	q->nomem = 0;
	napi_free_frags(&q->napi);
}


/**
 *	free_qset - free the resources of an SGE queue set
 *	@adapter: the adapter owning the queue set
 *	@q: the queue set
 *
 *	Release the HW and SW resources associated with an SGE queue set, such
 *	as HW contexts, packet buffers, and descriptor rings.  Traffic to the
 *	queue set must be quiesced prior to calling this.
 */
static void t3_free_qset(struct adapter *adapter, struct sge_qset *q)
{
	int i;
	struct pci_dev *pdev = adapter->pdev;

	for (i = 0; i < SGE_RXQ_PER_SET; ++i)
		if (q->fl[i].desc) {
			spin_lock_irq(&adapter->sge.reg_lock);
			t3_sge_disable_fl(adapter, q->fl[i].cntxt_id);
			spin_unlock_irq(&adapter->sge.reg_lock);
			free_rx_bufs(pdev, &q->fl[i]);
			kfree(q->fl[i].sdesc);
			dma_free_coherent(&pdev->dev,
					  q->fl[i].size *
					  sizeof(struct rx_desc), q->fl[i].desc,
					  q->fl[i].phys_addr);
		}

	for (i = 0; i < SGE_TXQ_PER_SET; ++i)
		if (q->txq[i].desc) {
			spin_lock_irq(&adapter->sge.reg_lock);
			t3_sge_enable_ecntxt(adapter, q->txq[i].cntxt_id, 0);
			spin_unlock_irq(&adapter->sge.reg_lock);
			if (q->txq[i].sdesc) {
				free_tx_desc(adapter, &q->txq[i],
					     q->txq[i].in_use);
				kfree(q->txq[i].sdesc);
			}
			dma_free_coherent(&pdev->dev,
					  q->txq[i].size *
					  sizeof(struct tx_desc),
					  q->txq[i].desc, q->txq[i].phys_addr);
			__skb_queue_purge(&q->txq[i].sendq);
		}

	if (q->rspq.desc) {
		spin_lock_irq(&adapter->sge.reg_lock);
		t3_sge_disable_rspcntxt(adapter, q->rspq.cntxt_id);
		spin_unlock_irq(&adapter->sge.reg_lock);
		dma_free_coherent(&pdev->dev,
				  q->rspq.size * sizeof(struct rsp_desc),
				  q->rspq.desc, q->rspq.phys_addr);
	}

	t3_reset_qset(q);
}

/**
 *	init_qset_cntxt - initialize an SGE queue set context info
 *	@qs: the queue set
 *	@id: the queue set id
 *
 *	Initializes the TIDs and context ids for the queues of a queue set.
 */
static void init_qset_cntxt(struct sge_qset *qs, unsigned int id)
{
	qs->rspq.cntxt_id = id;
	qs->fl[0].cntxt_id = 2 * id;
	qs->fl[1].cntxt_id = 2 * id + 1;
	qs->txq[TXQ_ETH].cntxt_id = FW_TUNNEL_SGEEC_START + id;
	qs->txq[TXQ_ETH].token = FW_TUNNEL_TID_START + id;
	qs->txq[TXQ_OFLD].cntxt_id = FW_OFLD_SGEEC_START + id;
	qs->txq[TXQ_CTRL].cntxt_id = FW_CTRL_SGEEC_START + id;
	qs->txq[TXQ_CTRL].token = FW_CTRL_TID_START + id;
}

/**
 *	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)
{
	/* alternatively: 3 * (n / 2) + 2 * (n & 1) */
	return (3 * n) / 2 + (n & 1);
}

/**
 *	flits_to_desc - returns the num of Tx descriptors for the given flits
 *	@n: the number of flits
 *
 *	Calculates 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 >= ARRAY_SIZE(flit_desc_map));
	return flit_desc_map[n];
}

/**
 *	get_packet - return the next ingress packet buffer from a free list
 *	@adap: the adapter that received the packet
 *	@fl: the SGE free list holding the packet
 *	@len: the packet length including any SGE padding
 *	@drop_thres: # of remaining buffers before we start dropping packets
 *
 *	Get the next packet from a free list and complete setup of the
 *	sk_buff.  If the packet is small we make a copy and recycle the
 *	original buffer, otherwise we use the original buffer itself.  If a
 *	positive drop threshold is supplied packets are dropped and their
 *	buffers recycled if (a) the number of remaining buffers is under the
 *	threshold and the packet is too big to copy, or (b) the packet should
 *	be copied but there is no memory for the copy.
 */
static struct sk_buff *get_packet(struct adapter *adap, struct sge_fl *fl,
				  unsigned int len, unsigned int drop_thres)
{
	struct sk_buff *skb = NULL;
	struct rx_sw_desc *sd = &fl->sdesc[fl->cidx];

	prefetch(sd->skb->data);
	fl->credits--;

	if (len <= SGE_RX_COPY_THRES) {
		skb = alloc_skb(len, GFP_ATOMIC);
		if (likely(skb != NULL)) {
			__skb_put(skb, len);
			pci_dma_sync_single_for_cpu(adap->pdev,
					    dma_unmap_addr(sd, dma_addr), len,
					    PCI_DMA_FROMDEVICE);
			memcpy(skb->data, sd->skb->data, len);
			pci_dma_sync_single_for_device(adap->pdev,
					    dma_unmap_addr(sd, dma_addr), len,
					    PCI_DMA_FROMDEVICE);
		} else if (!drop_thres)
			goto use_orig_buf;
recycle:
		recycle_rx_buf(adap, fl, fl->cidx);
		return skb;
	}

	if (unlikely(fl->credits < drop_thres) &&
	    refill_fl(adap, fl, min(MAX_RX_REFILL, fl->size - fl->credits - 1),
		      GFP_ATOMIC | __GFP_COMP) == 0)
		goto recycle;

use_orig_buf:
	pci_unmap_single(adap->pdev, dma_unmap_addr(sd, dma_addr),
			 fl->buf_size, PCI_DMA_FROMDEVICE);
	skb = sd->skb;
	skb_put(skb, len);
	__refill_fl(adap, fl);
	return skb;
}

/**
 *	get_packet_pg - return the next ingress packet buffer from a free list
 *	@adap: the adapter that received the packet
 *	@fl: the SGE free list holding the packet
 *	@len: the packet length including any SGE padding
 *	@drop_thres: # of remaining buffers before we start dropping packets
 *
 *	Get the next packet from a free list populated with page chunks.
 *	If the packet is small we make a copy and recycle the original buffer,
 *	otherwise we attach the original buffer as a page fragment to a fresh
 *	sk_buff.  If a positive drop threshold is supplied packets are dropped
 *	and their buffers recycled if (a) the number of remaining buffers is
 *	under the threshold and the packet is too big to copy, or (b) there's
 *	no system memory.
 *
 * 	Note: this function is similar to @get_packet but deals with Rx buffers
 * 	that are page chunks rather than sk_buffs.
 */
static struct sk_buff *get_packet_pg(struct adapter *adap, struct sge_fl *fl,
				     struct sge_rspq *q, unsigned int len,
				     unsigned int drop_thres)
{
	struct sk_buff *newskb, *skb;
	struct rx_sw_desc *sd = &fl->sdesc[fl->cidx];

	dma_addr_t dma_addr = dma_unmap_addr(sd, dma_addr);

	newskb = skb = q->pg_skb;
	if (!skb && (len <= SGE_RX_COPY_THRES)) {
		newskb = alloc_skb(len, GFP_ATOMIC);
		if (likely(newskb != NULL)) {
			__skb_put(newskb, len);
			pci_dma_sync_single_for_cpu(adap->pdev, dma_addr, len,
					    PCI_DMA_FROMDEVICE);
			memcpy(newskb->data, sd->pg_chunk.va, len);
			pci_dma_sync_single_for_device(adap->pdev, dma_addr,
						       len,
						       PCI_DMA_FROMDEVICE);
		} else if (!drop_thres)
			return NULL;
recycle:
		fl->credits--;
		recycle_rx_buf(adap, fl, fl->cidx);
		q->rx_recycle_buf++;
		return newskb;
	}

	if (unlikely(q->rx_recycle_buf || (!skb && fl->credits <= drop_thres)))
		goto recycle;

	prefetch(sd->pg_chunk.p_cnt);

	if (!skb)
		newskb = alloc_skb(SGE_RX_PULL_LEN, GFP_ATOMIC);

	if (unlikely(!newskb)) {
		if (!drop_thres)
			return NULL;
		goto recycle;
	}

	pci_dma_sync_single_for_cpu(adap->pdev, dma_addr, len,
				    PCI_DMA_FROMDEVICE);
	(*sd->pg_chunk.p_cnt)--;
	if (!*sd->pg_chunk.p_cnt && sd->pg_chunk.page != fl->pg_chunk.page)
		pci_unmap_page(adap->pdev,
			       sd->pg_chunk.mapping,
			       fl->alloc_size,
			       PCI_DMA_FROMDEVICE);
	if (!skb) {
		__skb_put(newskb, SGE_RX_PULL_LEN);
		memcpy(newskb->data, sd->pg_chunk.va, SGE_RX_PULL_LEN);
		skb_fill_page_desc(newskb, 0, sd->pg_chunk.page,
				   sd->pg_chunk.offset + SGE_RX_PULL_LEN,
				   len - SGE_RX_PULL_LEN);
		newskb->len = len;
		newskb->data_len = len - SGE_RX_PULL_LEN;
		newskb->truesize += newskb->data_len;
	} else {
		skb_fill_page_desc(newskb, skb_shinfo(newskb)->nr_frags,
				   sd->pg_chunk.page,
				   sd->pg_chunk.offset, len);
		newskb->len += len;
		newskb->data_len += len;
		newskb->truesize += len;
	}

	fl->credits--;
	/*
	 * We do not refill FLs here, we let the caller do it to overlap a
	 * prefetch.
	 */
	return newskb;
}

/**
 *	get_imm_packet - return the next ingress packet buffer from a response
 *	@resp: the response descriptor containing the packet data
 *
 *	Return a packet containing the immediate data of the given response.
 */
static inline struct sk_buff *get_imm_packet(const struct rsp_desc *resp)
{
	struct sk_buff *skb = alloc_skb(IMMED_PKT_SIZE, GFP_ATOMIC);

	if (skb) {
		__skb_put(skb, IMMED_PKT_SIZE);
		skb_copy_to_linear_data(skb, resp->imm_data, IMMED_PKT_SIZE);
	}
	return skb;
}

/**
 *	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.  Ethernet packets require addition of WR and CPL headers.
 */
static inline unsigned int calc_tx_descs(const struct sk_buff *skb)
{
	unsigned int flits;

	if (skb->len <= WR_LEN - sizeof(struct cpl_tx_pkt))
		return 1;

	flits = sgl_len(skb_shinfo(skb)->nr_frags + 1) + 2;
	if (skb_shinfo(skb)->gso_size)
		flits++;
	return flits_to_desc(flits);
}

/**
 *	make_sgl - populate a scatter/gather list for a packet
 *	@skb: the packet
 *	@sgp: the SGL to populate
 *	@start: start address of skb main body data to include in the SGL
 *	@len: length of skb main body data to include in the SGL
 *	@pdev: the PCI device
 *
 *	Generates a scatter/gather list for the buffers that make up a packet
 *	and returns the SGL size in 8-byte words.  The caller must size the SGL
 *	appropriately.
 */
static inline unsigned int make_sgl(const struct sk_buff *skb,
				    struct sg_ent *sgp, unsigned char *start,
				    unsigned int len, struct pci_dev *pdev)
{
	dma_addr_t mapping;
	unsigned int i, j = 0, nfrags;

	if (len) {
		mapping = pci_map_single(pdev, start, len, PCI_DMA_TODEVICE);
		sgp->len[0] = cpu_to_be32(len);
		sgp->addr[0] = cpu_to_be64(mapping);
		j = 1;
	}

	nfrags = skb_shinfo(skb)->nr_frags;
	for (i = 0; i < nfrags; i++) {
		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];

		mapping = pci_map_page(pdev, frag->page, frag->page_offset,
				       frag->size, PCI_DMA_TODEVICE);
		sgp->len[j] = cpu_to_be32(frag->size);
		sgp->addr[j] = cpu_to_be64(mapping);
		j ^= 1;
		if (j == 0)
			++sgp;
	}
	if (j)
		sgp->len[j] = 0;
	return ((nfrags + (len != 0)) * 3) / 2 + j;
}

/**
 *	check_ring_tx_db - check and potentially ring a Tx queue's doorbell
 *	@adap: the adapter
 *	@q: the Tx queue
 *
 *	Ring the doorbel if a Tx queue is asleep.  There is a natural race,