/*
- * Copyright (c) 2005-2007 Chelsio, Inc. All rights reserved.
+ * 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
#include <linux/ip.h>
#include <linux/tcp.h>
#include <linux/dma-mapping.h>
+#include <net/arp.h>
#include "common.h"
#include "regs.h"
#include "sge_defs.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 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)
};
struct tx_desc {
- u64 flit[TX_DESC_FLITS];
+ __be64 flit[TX_DESC_FLITS];
};
struct rx_desc {
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 */
- struct sk_buff *skb;
- DECLARE_PCI_UNMAP_ADDR(dma_addr);
+struct rx_sw_desc { /* SW state per Rx descriptor */
+ union {
+ struct sk_buff *skb;
+ struct fl_pg_chunk pg_chunk;
+ };
+ DECLARE_PCI_UNMAP_ADDR(dma_addr);
};
struct rsp_desc { /* response queue descriptor */
u8 intr_gen;
};
-struct unmap_info { /* packet unmapping info, overlays skb->cb */
- int sflit; /* start flit of first SGL entry in Tx descriptor */
- u16 fragidx; /* first page fragment in current Tx descriptor */
- u16 addr_idx; /* buffer index of first SGL entry in descriptor */
- u32 len; /* mapped length of skb main body */
+/*
+ * 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];
};
/*
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));
}
*
* 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, we keep the information necessary to
- * unmap an sk_buff partly in the sk_buff itself (in its cb), and partly
- * in the Tx descriptors (the physical addresses of the various data
- * buffers). The send functions initialize the state in skb->cb so we
- * can unmap the buffers held in the first Tx descriptor here, and we
- * have enough information at this point to update the state for the next
- * Tx descriptor.
+ * 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 unmap_info *ui = (struct unmap_info *)skb->cb;
- int nfrags, frag_idx, curflit, j = ui->addr_idx;
+ 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[ui->sflit];
+ sgp = (struct sg_ent *)&q->desc[cidx].flit[d->sflit];
+ frag_idx = d->fragidx;
- if (ui->len) {
- pci_unmap_single(pdev, be64_to_cpu(sgp->addr[0]), ui->len,
- PCI_DMA_TODEVICE);
- ui->len = 0; /* so we know for next descriptor for this skb */
+ 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;
}
- frag_idx = ui->fragidx;
- curflit = ui->sflit + 1 + j;
+ curflit = d->sflit + 1 + j;
nfrags = skb_shinfo(skb)->nr_frags;
while (frag_idx < nfrags && curflit < WR_FLITS) {
frag_idx++;
}
- if (frag_idx < nfrags) { /* SGL continues into next Tx descriptor */
- ui->fragidx = frag_idx;
- ui->addr_idx = j;
- ui->sflit = curflit - WR_FLITS - j; /* sflit can be -1 */
+ 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 */
}
}
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_skb_unmap())
+ if (need_unmap)
unmap_skb(d->skb, q, cidx, pdev);
- if (d->skb->priority == cidx)
+ if (d->eop)
kfree_skb(d->skb);
}
++d;
* 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 void reclaim_completed_tx(struct adapter *adapter,
- struct sge_txq *q)
+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;
}
/**
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, pci_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
while (q->credits--) {
struct rx_sw_desc *d = &q->sdesc[cidx];
- pci_unmap_single(pdev, pci_unmap_addr(d, dma_addr),
- q->buf_size, PCI_DMA_FROMDEVICE);
- kfree_skb(d->skb);
- d->skb = NULL;
+
+ 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
- * @skb: the buffer to add
+ * @va: buffer start VA
* @len: the buffer length
* @d: the HW Rx descriptor to write
* @sd: the SW Rx descriptor to write
* Add a buffer of the given length to the supplied HW and SW Rx
* descriptors.
*/
-static inline void add_one_rx_buf(struct sk_buff *skb, unsigned int len,
- struct rx_desc *d, struct rx_sw_desc *sd,
- unsigned int gen, struct pci_dev *pdev)
+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;
- sd->skb = skb;
- mapping = pci_map_single(pdev, skb->data, len, PCI_DMA_FROMDEVICE);
+ mapping = pci_map_single(pdev, va, len, PCI_DMA_FROMDEVICE);
+ if (unlikely(pci_dma_mapping_error(pdev, mapping)))
+ return -ENOMEM;
+
pci_unmap_addr_set(sd, dma_addr, mapping);
d->addr_lo = cpu_to_be32(mapping);
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;
+ t3_write_reg(adap, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
+ }
}
/**
* allocated with the supplied gfp flags. The caller must assure that
* @n does not exceed the queue's capacity.
*/
-static void refill_fl(struct adapter *adap, struct sge_fl *q, int n, gfp_t gfp)
+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--) {
- struct sk_buff *skb = alloc_skb(q->buf_size, gfp);
+ dma_addr_t mapping;
+ int err;
- if (!skb)
- break;
+ 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;
+ pci_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;
+ }
+ }
- add_one_rx_buf(skb, q->buf_size, d, sd, q->gen, adap->pdev);
d++;
sd++;
if (++q->pidx == q->size) {
sd = q->sdesc;
d = q->desc;
}
- q->credits++;
+ count++;
}
- t3_write_reg(adap, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
+ 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(16U, fl->size - fl->credits), GFP_ATOMIC);
+ refill_fl(adap, fl, min(MAX_RX_REFILL, fl->size - fl->credits),
+ GFP_ATOMIC | __GFP_COMP);
}
/**
wmb();
to->len_gen = cpu_to_be32(V_FLD_GEN1(q->gen));
to->gen2 = cpu_to_be32(V_FLD_GEN2(q->gen));
- q->credits++;
if (++q->pidx == q->size) {
q->pidx = 0;
q->gen ^= 1;
}
- t3_write_reg(adap, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
+
+ q->credits++;
+ q->pend_cred++;
+ ring_fl_db(adap, q);
}
/**
* 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 sw_size, dma_addr_t * phys, void *metadata)
{
size_t len = nelem * elem_size;
void *s = NULL;
if (!p)
return NULL;
- if (sw_size) {
+ if (sw_size && metadata) {
s = kcalloc(nelem, sw_size, GFP_KERNEL);
if (!s) {
dma_free_coherent(&pdev->dev, len, p, *phys);
return NULL;
}
- }
- if (metadata)
*(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
* as HW contexts, packet buffers, and descriptor rings. Traffic to the
* queue set must be quiesced prior to calling this.
*/
-void t3_free_qset(struct adapter *adapter, struct sge_qset *q)
+static void t3_free_qset(struct adapter *adapter, struct sge_qset *q)
{
int i;
struct pci_dev *pdev = adapter->pdev;
- if (q->tx_reclaim_timer.function)
- del_timer_sync(&q->tx_reclaim_timer);
-
for (i = 0; i < SGE_RXQ_PER_SET; ++i)
if (q->fl[i].desc) {
- spin_lock(&adapter->sge.reg_lock);
+ spin_lock_irq(&adapter->sge.reg_lock);
t3_sge_disable_fl(adapter, q->fl[i].cntxt_id);
- spin_unlock(&adapter->sge.reg_lock);
+ spin_unlock_irq(&adapter->sge.reg_lock);
free_rx_bufs(pdev, &q->fl[i]);
kfree(q->fl[i].sdesc);
dma_free_coherent(&pdev->dev,
for (i = 0; i < SGE_TXQ_PER_SET; ++i)
if (q->txq[i].desc) {
- spin_lock(&adapter->sge.reg_lock);
+ spin_lock_irq(&adapter->sge.reg_lock);
t3_sge_enable_ecntxt(adapter, q->txq[i].cntxt_id, 0);
- spin_unlock(&adapter->sge.reg_lock);
+ spin_unlock_irq(&adapter->sge.reg_lock);
if (q->txq[i].sdesc) {
free_tx_desc(adapter, &q->txq[i],
q->txq[i].in_use);
}
if (q->rspq.desc) {
- spin_lock(&adapter->sge.reg_lock);
+ spin_lock_irq(&adapter->sge.reg_lock);
t3_sge_disable_rspcntxt(adapter, q->rspq.cntxt_id);
- spin_unlock(&adapter->sge.reg_lock);
+ 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);
}
- if (q->netdev)
- q->netdev->atalk_ptr = NULL;
-
- memset(q, 0, sizeof(*q));
+ t3_reset_qset(q);
}
/**
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,
- pci_unmap_addr(sd,
- dma_addr),
- len, PCI_DMA_FROMDEVICE);
+ pci_unmap_addr(sd, dma_addr), len,
+ PCI_DMA_FROMDEVICE);
memcpy(skb->data, sd->skb->data, len);
pci_dma_sync_single_for_device(adap->pdev,
- pci_unmap_addr(sd,
- dma_addr),
- len, PCI_DMA_FROMDEVICE);
+ pci_unmap_addr(sd, dma_addr), len,
+ PCI_DMA_FROMDEVICE);
} else if (!drop_thres)
goto use_orig_buf;
- recycle:
+recycle:
recycle_rx_buf(adap, fl, fl->cidx);
return skb;
}
- if (unlikely(fl->credits < drop_thres))
+ 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:
+use_orig_buf:
pci_unmap_single(adap->pdev, pci_unmap_addr(sd, dma_addr),
fl->buf_size, PCI_DMA_FROMDEVICE);
skb = sd->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 = pci_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)
+ 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
*
if (skb) {
__skb_put(skb, IMMED_PKT_SIZE);
- memcpy(skb->data, resp->imm_data, IMMED_PKT_SIZE);
+ skb_copy_to_linear_data(skb, resp->imm_data, IMMED_PKT_SIZE);
}
return skb;
}
const struct sge_txq *q,
const struct sg_ent *sgl,
unsigned int flits, unsigned int sgl_flits,
- unsigned int gen, unsigned int wr_hi,
- unsigned int wr_lo)
+ unsigned int gen, __be32 wr_hi,
+ __be32 wr_lo)
{
struct work_request_hdr *wrp = (struct work_request_hdr *)d;
struct tx_sw_desc *sd = &q->sdesc[pidx];
sd->skb = skb;
if (need_skb_unmap()) {
- struct unmap_info *ui = (struct unmap_info *)skb->cb;
-
- ui->fragidx = 0;
- ui->addr_idx = 0;
- ui->sflit = flits;
+ sd->fragidx = 0;
+ sd->addr_idx = 0;
+ sd->sflit = flits;
}
if (likely(ndesc == 1)) {
- skb->priority = pidx;
+ sd->eop = 1;
wrp->wr_hi = htonl(F_WR_SOP | F_WR_EOP | V_WR_DATATYPE(1) |
V_WR_SGLSFLT(flits)) | wr_hi;
wmb();
fp += avail;
d++;
+ sd->eop = 0;
sd++;
if (++pidx == q->size) {
pidx = 0;
wr_gen2(d, gen);
flits = 1;
}
- skb->priority = pidx;
+ sd->eop = 1;
wrp->wr_hi |= htonl(F_WR_EOP);
wmb();
wp->wr_lo = htonl(V_WR_LEN(WR_FLITS) | V_WR_GEN(ogen)) | wr_lo;
struct tx_desc *d = &q->desc[pidx];
struct cpl_tx_pkt *cpl = (struct cpl_tx_pkt *)d;
- cpl->len = htonl(skb->len | 0x80000000);
+ cpl->len = htonl(skb->len);
cntrl = V_TXPKT_INTF(pi->port_id);
if (vlan_tx_tag_present(skb) && pi->vlan_grp)
d->flit[2] = 0;
cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT_LSO);
hdr->cntrl = htonl(cntrl);
- eth_type = skb->nh.raw - skb->data == ETH_HLEN ?
+ eth_type = skb_network_offset(skb) == ETH_HLEN ?
CPL_ETH_II : CPL_ETH_II_VLAN;
tso_info |= V_LSO_ETH_TYPE(eth_type) |
- V_LSO_IPHDR_WORDS(skb->nh.iph->ihl) |
- V_LSO_TCPHDR_WORDS(skb->h.th->doff);
+ V_LSO_IPHDR_WORDS(ip_hdr(skb)->ihl) |
+ V_LSO_TCPHDR_WORDS(tcp_hdr(skb)->doff);
hdr->lso_info = htonl(tso_info);
flits = 3;
} else {
if (skb->len <= WR_LEN - sizeof(*cpl)) {
q->sdesc[pidx].skb = NULL;
if (!skb->data_len)
- memcpy(&d->flit[2], skb->data, skb->len);
+ skb_copy_from_linear_data(skb, &d->flit[2],
+ skb->len);
else
skb_copy_bits(skb, 0, &d->flit[2], skb->len);
sgp = ndesc == 1 ? (struct sg_ent *)&d->flit[flits] : sgl;
sgl_flits = make_sgl(skb, sgp, skb->data, skb_headlen(skb), adap->pdev);
- if (need_skb_unmap())
- ((struct unmap_info *)skb->cb)->len = skb_headlen(skb);
write_wr_hdr_sgl(ndesc, skb, d, pidx, q, sgl, flits, sgl_flits, gen,
htonl(V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) | compl),
htonl(V_WR_TID(q->token)));
}
+static inline void t3_stop_tx_queue(struct netdev_queue *txq,
+ struct sge_qset *qs, struct sge_txq *q)
+{
+ netif_tx_stop_queue(txq);
+ set_bit(TXQ_ETH, &qs->txq_stopped);
+ q->stops++;
+}
+
/**
* eth_xmit - add a packet to the Ethernet Tx queue
* @skb: the packet
*
* Add a packet to an SGE Tx queue. Runs with softirqs disabled.
*/
-int t3_eth_xmit(struct sk_buff *skb, struct net_device *dev)
+netdev_tx_t t3_eth_xmit(struct sk_buff *skb, struct net_device *dev)
{
+ int qidx;
unsigned int ndesc, pidx, credits, gen, compl;
const struct port_info *pi = netdev_priv(dev);
- struct adapter *adap = dev->priv;
- struct sge_qset *qs = dev2qset(dev);
- struct sge_txq *q = &qs->txq[TXQ_ETH];
+ struct adapter *adap = pi->adapter;
+ struct netdev_queue *txq;
+ struct sge_qset *qs;
+ struct sge_txq *q;
/*
* The chip min packet length is 9 octets but play safe and reject
return NETDEV_TX_OK;
}
- spin_lock(&q->lock);
- reclaim_completed_tx(adap, q);
+ qidx = skb_get_queue_mapping(skb);
+ qs = &pi->qs[qidx];
+ q = &qs->txq[TXQ_ETH];
+ txq = netdev_get_tx_queue(dev, qidx);
+
+ reclaim_completed_tx(adap, q, TX_RECLAIM_CHUNK);
credits = q->size - q->in_use;
ndesc = calc_tx_descs(skb);
if (unlikely(credits < ndesc)) {
- if (!netif_queue_stopped(dev)) {
- netif_stop_queue(dev);
- set_bit(TXQ_ETH, &qs->txq_stopped);
- q->stops++;
- dev_err(&adap->pdev->dev,
- "%s: Tx ring %u full while queue awake!\n",
- dev->name, q->cntxt_id & 7);
- }
- spin_unlock(&q->lock);
+ t3_stop_tx_queue(txq, qs, q);
+ dev_err(&adap->pdev->dev,
+ "%s: Tx ring %u full while queue awake!\n",
+ dev->name, q->cntxt_id & 7);
return NETDEV_TX_BUSY;
}
q->in_use += ndesc;
if (unlikely(credits - ndesc < q->stop_thres)) {
- q->stops++;
- netif_stop_queue(dev);
- set_bit(TXQ_ETH, &qs->txq_stopped);
-#if !USE_GTS
+ t3_stop_tx_queue(txq, qs, q);
+
if (should_restart_tx(q) &&
test_and_clear_bit(TXQ_ETH, &qs->txq_stopped)) {
q->restarts++;
- netif_wake_queue(dev);
+ netif_tx_wake_queue(txq);
}
-#endif
}
gen = q->gen;
if (vlan_tx_tag_present(skb) && pi->vlan_grp)
qs->port_stats[SGE_PSTAT_VLANINS]++;
- dev->trans_start = jiffies;
- spin_unlock(&q->lock);
-
/*
* We do not use Tx completion interrupts to free DMAd Tx packets.
- * This is good for performamce but means that we rely on new Tx
+ * This is good for performance but means that we rely on new Tx
* packets arriving to run the destructors of completed packets,
* which open up space in their sockets' send queues. Sometimes
* we do not get such new packets causing Tx to stall. A single
*
* Writes a packet as immediate data into a Tx descriptor. The packet
* contains a work request at its beginning. We must write the packet
- * carefully so the SGE doesn't read accidentally before it's written in
- * its entirety.
+ * carefully so the SGE doesn't read it accidentally before it's written
+ * in its entirety.
*/
static inline void write_imm(struct tx_desc *d, struct sk_buff *skb,
unsigned int len, unsigned int gen)
struct work_request_hdr *from = (struct work_request_hdr *)skb->data;
struct work_request_hdr *to = (struct work_request_hdr *)d;
- memcpy(&to[1], &from[1], len - sizeof(*from));
+ if (likely(!skb->data_len))
+ memcpy(&to[1], &from[1], len - sizeof(*from));
+ else
+ skb_copy_bits(skb, sizeof(*from), &to[1], len - sizeof(*from));
+
to->wr_hi = from->wr_hi | htonl(F_WR_SOP | F_WR_EOP |
V_WR_BCNTLFLT(len & 7));
wmb();
static inline int immediate(const struct sk_buff *skb)
{
- return skb->len <= WR_LEN && !skb->data_len;
+ return skb->len <= WR_LEN;
}
/**
struct sk_buff *skb;
struct sge_qset *qs = (struct sge_qset *)data;
struct sge_txq *q = &qs->txq[TXQ_CTRL];
- struct adapter *adap = qs->netdev->priv;
spin_lock(&q->lock);
again:reclaim_completed_tx_imm(q);
- while (q->in_use < q->size && (skb = __skb_dequeue(&q->sendq)) != NULL) {
+ while (q->in_use < q->size &&
+ (skb = __skb_dequeue(&q->sendq)) != NULL) {
write_imm(&q->desc[q->pidx], skb, skb->len, q->gen);
}
spin_unlock(&q->lock);
- t3_write_reg(adap, A_SG_KDOORBELL,
+ wmb();
+ t3_write_reg(qs->adap, A_SG_KDOORBELL,
F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
}
*/
int t3_mgmt_tx(struct adapter *adap, struct sk_buff *skb)
{
- return ctrl_xmit(adap, &adap->sge.qs[0].txq[TXQ_CTRL], skb);
+ int ret;
+ local_bh_disable();
+ ret = ctrl_xmit(adap, &adap->sge.qs[0].txq[TXQ_CTRL], skb);
+ local_bh_enable();
+
+ return ret;
+}
+
+/**
+ * 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)
+{
+ int i;
+ const dma_addr_t *p;
+ const struct skb_shared_info *si;
+ const struct deferred_unmap_info *dui;
+
+ dui = (struct deferred_unmap_info *)skb->head;
+ p = dui->addr;
+
+ if (skb->tail - skb->transport_header)
+ pci_unmap_single(dui->pdev, *p++,
+ skb->tail - skb->transport_header,
+ PCI_DMA_TODEVICE);
+
+ si = skb_shinfo(skb);
+ for (i = 0; i < si->nr_frags; i++)
+ pci_unmap_page(dui->pdev, *p++, si->frags[i].size,
+ PCI_DMA_TODEVICE);
+}
+
+static void setup_deferred_unmapping(struct sk_buff *skb, struct pci_dev *pdev,
+ const struct sg_ent *sgl, int sgl_flits)
+{
+ dma_addr_t *p;
+ struct deferred_unmap_info *dui;
+
+ dui = (struct deferred_unmap_info *)skb->head;
+ dui->pdev = pdev;
+ for (p = dui->addr; sgl_flits >= 3; sgl++, sgl_flits -= 3) {
+ *p++ = be64_to_cpu(sgl->addr[0]);
+ *p++ = be64_to_cpu(sgl->addr[1]);
+ }
+ if (sgl_flits)
+ *p = be64_to_cpu(sgl->addr[0]);
}
/**
/* Only TX_DATA builds SGLs */
from = (struct work_request_hdr *)skb->data;
- memcpy(&d->flit[1], &from[1], skb->h.raw - skb->data - sizeof(*from));
+ memcpy(&d->flit[1], &from[1],
+ skb_transport_offset(skb) - sizeof(*from));
- flits = (skb->h.raw - skb->data) / 8;
+ flits = skb_transport_offset(skb) / 8;
sgp = ndesc == 1 ? (struct sg_ent *)&d->flit[flits] : sgl;
- sgl_flits = make_sgl(skb, sgp, skb->h.raw, skb->tail - skb->h.raw,
+ sgl_flits = make_sgl(skb, sgp, skb_transport_header(skb),
+ skb->tail - skb->transport_header,
adap->pdev);
- if (need_skb_unmap())
- ((struct unmap_info *)skb->cb)->len = skb->tail - skb->h.raw;
+ if (need_skb_unmap()) {
+ setup_deferred_unmapping(skb, adap->pdev, sgp, sgl_flits);
+ skb->destructor = deferred_unmap_destructor;
+ }
write_wr_hdr_sgl(ndesc, skb, d, pidx, q, sgl, flits, sgl_flits,
gen, from->wr_hi, from->wr_lo);
*/
static inline unsigned int calc_tx_descs_ofld(const struct sk_buff *skb)
{
- unsigned int flits, cnt = skb_shinfo(skb)->nr_frags;
+ unsigned int flits, cnt;
- if (skb->len <= WR_LEN && cnt == 0)
+ if (skb->len <= WR_LEN)
return 1; /* packet fits as immediate data */
- flits = (skb->h.raw - skb->data) / 8; /* headers */
- if (skb->tail != skb->h.raw)
+ flits = skb_transport_offset(skb) / 8; /* headers */
+ cnt = skb_shinfo(skb)->nr_frags;
+ if (skb->tail != skb->transport_header)
cnt++;
return flits_to_desc(flits + sgl_len(cnt));
}
unsigned int ndesc = calc_tx_descs_ofld(skb), pidx, gen;
spin_lock(&q->lock);
- again:reclaim_completed_tx(adap, q);
+again: reclaim_completed_tx(adap, q, TX_RECLAIM_CHUNK);
ret = check_desc_avail(adap, q, skb, ndesc, TXQ_OFLD);
if (unlikely(ret)) {
struct sk_buff *skb;
struct sge_qset *qs = (struct sge_qset *)data;
struct sge_txq *q = &qs->txq[TXQ_OFLD];
- struct adapter *adap = qs->netdev->priv;
+ const struct port_info *pi = netdev_priv(qs->netdev);
+ struct adapter *adap = pi->adapter;
spin_lock(&q->lock);
- again:reclaim_completed_tx(adap, q);
+again: reclaim_completed_tx(adap, q, TX_RECLAIM_CHUNK);
while ((skb = skb_peek(&q->sendq)) != NULL) {
unsigned int gen, pidx;
set_bit(TXQ_RUNNING, &q->flags);
set_bit(TXQ_LAST_PKT_DB, &q->flags);
#endif
+ wmb();
t3_write_reg(adap, A_SG_KDOORBELL,
F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
}
*/
static inline void offload_enqueue(struct sge_rspq *q, struct sk_buff *skb)
{
- skb->next = skb->prev = NULL;
- if (q->rx_tail)
- q->rx_tail->next = skb;
- else {
+ int was_empty = skb_queue_empty(&q->rx_queue);
+
+ __skb_queue_tail(&q->rx_queue, skb);
+
+ if (was_empty) {
struct sge_qset *qs = rspq_to_qset(q);
- if (__netif_rx_schedule_prep(qs->netdev))
- __netif_rx_schedule(qs->netdev);
- q->rx_head = skb;
+ napi_schedule(&qs->napi);
}
- q->rx_tail = skb;
}
/**
* receive handler. Batches need to be of modest size as we do prefetches
* on the packets in each.
*/
-static int ofld_poll(struct net_device *dev, int *budget)
+static int ofld_poll(struct napi_struct *napi, int budget)
{
- struct adapter *adapter = dev->priv;
- struct sge_qset *qs = dev2qset(dev);
+ struct sge_qset *qs = container_of(napi, struct sge_qset, napi);
struct sge_rspq *q = &qs->rspq;
- int work_done, limit = min(*budget, dev->quota), avail = limit;
+ struct adapter *adapter = qs->adap;
+ int work_done = 0;
- while (avail) {
- struct sk_buff *head, *tail, *skbs[RX_BUNDLE_SIZE];
+ while (work_done < budget) {
+ struct sk_buff *skb, *tmp, *skbs[RX_BUNDLE_SIZE];
+ struct sk_buff_head queue;
int ngathered;
spin_lock_irq(&q->lock);
- head = q->rx_head;
- if (!head) {
- work_done = limit - avail;
- *budget -= work_done;
- dev->quota -= work_done;
- __netif_rx_complete(dev);
+ __skb_queue_head_init(&queue);
+ skb_queue_splice_init(&q->rx_queue, &queue);
+ if (skb_queue_empty(&queue)) {
+ napi_complete(napi);
spin_unlock_irq(&q->lock);
- return 0;
+ return work_done;
}
-
- tail = q->rx_tail;
- q->rx_head = q->rx_tail = NULL;
spin_unlock_irq(&q->lock);
- for (ngathered = 0; avail && head; avail--) {
- prefetch(head->data);
- skbs[ngathered] = head;
- head = head->next;
- skbs[ngathered]->next = NULL;
+ ngathered = 0;
+ skb_queue_walk_safe(&queue, skb, tmp) {
+ if (work_done >= budget)
+ break;
+ work_done++;
+
+ __skb_unlink(skb, &queue);
+ prefetch(skb->data);
+ skbs[ngathered] = skb;
if (++ngathered == RX_BUNDLE_SIZE) {
q->offload_bundles++;
adapter->tdev.recv(&adapter->tdev, skbs,
ngathered = 0;
}
}
- if (head) { /* splice remaining packets back onto Rx queue */
+ if (!skb_queue_empty(&queue)) {
+ /* splice remaining packets back onto Rx queue */
spin_lock_irq(&q->lock);
- tail->next = q->rx_head;
- if (!q->rx_head)
- q->rx_tail = tail;
- q->rx_head = head;
+ skb_queue_splice(&queue, &q->rx_queue);
spin_unlock_irq(&q->lock);
}
deliver_partial_bundle(&adapter->tdev, q, skbs, ngathered);
}
- work_done = limit - avail;
- *budget -= work_done;
- dev->quota -= work_done;
- return 1;
+
+ return work_done;
}
/**
struct sk_buff *skb, struct sk_buff *rx_gather[],
unsigned int gather_idx)
{
- rq->offload_pkts++;
- skb->mac.raw = skb->nh.raw = skb->h.raw = skb->data;
+ skb_reset_mac_header(skb);
+ skb_reset_network_header(skb);
+ skb_reset_transport_header(skb);
if (rq->polling) {
rx_gather[gather_idx++] = skb;
test_and_clear_bit(TXQ_ETH, &qs->txq_stopped)) {
qs->txq[TXQ_ETH].restarts++;
if (netif_running(qs->netdev))
- netif_wake_queue(qs->netdev);
+ netif_tx_wake_queue(qs->tx_q);
}
if (test_bit(TXQ_OFLD, &qs->txq_stopped) &&
}
/**
+ * cxgb3_arp_process - process an ARP request probing a private IP address
+ * @adapter: the adapter
+ * @skb: the skbuff containing the ARP request
+ *
+ * Check if the ARP request is probing the private IP address
+ * dedicated to iSCSI, generate an ARP reply if so.
+ */
+static void cxgb3_arp_process(struct adapter *adapter, struct sk_buff *skb)
+{
+ struct net_device *dev = skb->dev;
+ struct port_info *pi;
+ struct arphdr *arp;
+ unsigned char *arp_ptr;
+ unsigned char *sha;
+ __be32 sip, tip;
+
+ if (!dev)
+ return;
+
+ skb_reset_network_header(skb);
+ arp = arp_hdr(skb);
+
+ if (arp->ar_op != htons(ARPOP_REQUEST))
+ return;
+
+ arp_ptr = (unsigned char *)(arp + 1);
+ sha = arp_ptr;
+ arp_ptr += dev->addr_len;
+ memcpy(&sip, arp_ptr, sizeof(sip));
+ arp_ptr += sizeof(sip);
+ arp_ptr += dev->addr_len;
+ memcpy(&tip, arp_ptr, sizeof(tip));
+
+ pi = netdev_priv(dev);
+ if (tip != pi->iscsi_ipv4addr)
+ return;
+
+ arp_send(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha,
+ dev->dev_addr, sha);
+
+}
+
+static inline int is_arp(struct sk_buff *skb)
+{
+ return skb->protocol == htons(ETH_P_ARP);
+}
+
+/**
* rx_eth - process an ingress ethernet packet
* @adap: the adapter
* @rq: the response queue that received the packet
* if it was immediate data in a response.
*/
static void rx_eth(struct adapter *adap, struct sge_rspq *rq,
- struct sk_buff *skb, int pad)
+ struct sk_buff *skb, int pad, int lro)
{
struct cpl_rx_pkt *p = (struct cpl_rx_pkt *)(skb->data + pad);
+ struct sge_qset *qs = rspq_to_qset(rq);
struct port_info *pi;
- rq->eth_pkts++;
skb_pull(skb, sizeof(*p) + pad);
- skb->dev = adap->port[p->iff];
- skb->dev->last_rx = jiffies;
- skb->protocol = eth_type_trans(skb, skb->dev);
+ skb->protocol = eth_type_trans(skb, adap->port[p->iff]);
pi = netdev_priv(skb->dev);
- if (pi->rx_csum_offload && p->csum_valid && p->csum == 0xffff &&
- !p->fragment) {
- rspq_to_qset(rq)->port_stats[SGE_PSTAT_RX_CSUM_GOOD]++;
+ if ((pi->rx_offload & T3_RX_CSUM) && p->csum_valid &&
+ p->csum == htons(0xffff) && !p->fragment) {
+ qs->port_stats[SGE_PSTAT_RX_CSUM_GOOD]++;
skb->ip_summed = CHECKSUM_UNNECESSARY;
} else
skb->ip_summed = CHECKSUM_NONE;
+ skb_record_rx_queue(skb, qs - &adap->sge.qs[0]);
if (unlikely(p->vlan_valid)) {
struct vlan_group *grp = pi->vlan_grp;
- rspq_to_qset(rq)->port_stats[SGE_PSTAT_VLANEX]++;
+ qs->port_stats[SGE_PSTAT_VLANEX]++;
if (likely(grp))
- __vlan_hwaccel_rx(skb, grp, ntohs(p->vlan),
- rq->polling);
+ if (lro)
+ vlan_gro_receive(&qs->napi, grp,
+ ntohs(p->vlan), skb);
+ else {
+ if (unlikely(pi->iscsi_ipv4addr &&
+ is_arp(skb))) {
+ unsigned short vtag = ntohs(p->vlan) &
+ VLAN_VID_MASK;
+ skb->dev = vlan_group_get_device(grp,
+ vtag);
+ cxgb3_arp_process(adap, skb);
+ }
+ __vlan_hwaccel_rx(skb, grp, ntohs(p->vlan),
+ rq->polling);
+ }
else
dev_kfree_skb_any(skb);
- } else if (rq->polling)
- netif_receive_skb(skb);
- else
+ } else if (rq->polling) {
+ if (lro)
+ napi_gro_receive(&qs->napi, skb);
+ else {
+ if (unlikely(pi->iscsi_ipv4addr && is_arp(skb)))
+ cxgb3_arp_process(adap, skb);
+ netif_receive_skb(skb);
+ }
+ } else
netif_rx(skb);
}
+static inline int is_eth_tcp(u32 rss)
+{
+ return G_HASHTYPE(ntohl(rss)) == RSS_HASH_4_TUPLE;
+}
+
+/**
+ * lro_add_page - add a page chunk to an LRO session
+ * @adap: the adapter
+ * @qs: the associated queue set
+ * @fl: the free list containing the page chunk to add
+ * @len: packet length
+ * @complete: Indicates the last fragment of a frame
+ *
+ * Add a received packet contained in a page chunk to an existing LRO
+ * session.
+ */
+static void lro_add_page(struct adapter *adap, struct sge_qset *qs,
+ struct sge_fl *fl, int len, int complete)
+{
+ struct rx_sw_desc *sd = &fl->sdesc[fl->cidx];
+ struct sk_buff *skb = NULL;
+ struct cpl_rx_pkt *cpl;
+ struct skb_frag_struct *rx_frag;
+ int nr_frags;
+ int offset = 0;
+
+ if (!qs->nomem) {
+ skb = napi_get_frags(&qs->napi);
+ qs->nomem = !skb;
+ }
+
+ fl->credits--;
+
+ pci_dma_sync_single_for_cpu(adap->pdev,
+ pci_unmap_addr(sd, dma_addr),
+ fl->buf_size - SGE_PG_RSVD,
+ PCI_DMA_FROMDEVICE);
+
+ (*sd->pg_chunk.p_cnt)--;
+ if (!*sd->pg_chunk.p_cnt)
+ pci_unmap_page(adap->pdev,
+ sd->pg_chunk.mapping,
+ fl->alloc_size,
+ PCI_DMA_FROMDEVICE);
+
+ if (!skb) {
+ put_page(sd->pg_chunk.page);
+ if (complete)
+ qs->nomem = 0;
+ return;
+ }
+
+ rx_frag = skb_shinfo(skb)->frags;
+ nr_frags = skb_shinfo(skb)->nr_frags;
+
+ if (!nr_frags) {
+ offset = 2 + sizeof(struct cpl_rx_pkt);
+ qs->lro_va = sd->pg_chunk.va + 2;
+ }
+ len -= offset;
+
+ prefetch(qs->lro_va);
+
+ rx_frag += nr_frags;
+ rx_frag->page = sd->pg_chunk.page;
+ rx_frag->page_offset = sd->pg_chunk.offset + offset;
+ rx_frag->size = len;
+
+ skb->len += len;
+ skb->data_len += len;
+ skb->truesize += len;
+ skb_shinfo(skb)->nr_frags++;
+
+ if (!complete)
+ return;
+
+ skb->ip_summed = CHECKSUM_UNNECESSARY;
+ cpl = qs->lro_va;
+
+ if (unlikely(cpl->vlan_valid)) {
+ struct net_device *dev = qs->netdev;
+ struct port_info *pi = netdev_priv(dev);
+ struct vlan_group *grp = pi->vlan_grp;
+
+ if (likely(grp != NULL)) {
+ vlan_gro_frags(&qs->napi, grp, ntohs(cpl->vlan));
+ return;
+ }
+ }
+ napi_gro_frags(&qs->napi);
+}
+
/**
* handle_rsp_cntrl_info - handles control information in a response
* @qs: the queue set corresponding to the response
return (r->intr_gen & F_RSPD_GEN2) == q->gen;
}
+static inline void clear_rspq_bufstate(struct sge_rspq * const q)
+{
+ q->pg_skb = NULL;
+ q->rx_recycle_buf = 0;
+}
+
#define RSPD_GTS_MASK (F_RSPD_TXQ0_GTS | F_RSPD_TXQ1_GTS)
#define RSPD_CTRL_MASK (RSPD_GTS_MASK | \
V_RSPD_TXQ0_CR(M_RSPD_TXQ0_CR) | \
q->next_holdoff = q->holdoff_tmr;
while (likely(budget_left && is_new_response(r, q))) {
- int eth, ethpad = 0;
+ int packet_complete, eth, ethpad = 2, lro = qs->lro_enabled;
struct sk_buff *skb = NULL;
u32 len, flags = ntohl(r->flags);
- u32 rss_hi = *(const u32 *)r, rss_lo = r->rss_hdr.rss_hash_val;
+ __be32 rss_hi = *(const __be32 *)r,
+ rss_lo = r->rss_hdr.rss_hash_val;
eth = r->rss_hdr.opcode == CPL_RX_PKT;
} else if (flags & F_RSPD_IMM_DATA_VALID) {
skb = get_imm_packet(r);
if (unlikely(!skb)) {
- no_mem:
+no_mem:
q->next_holdoff = NOMEM_INTR_DELAY;
q->nomem++;
/* consume one credit since we tried */
break;
}
q->imm_data++;
+ ethpad = 0;
} else if ((len = ntohl(r->len_cq)) != 0) {
struct sge_fl *fl;
+ lro &= eth && is_eth_tcp(rss_hi);
+
fl = (len & F_RSPD_FLQ) ? &qs->fl[1] : &qs->fl[0];
- fl->credits--;
- skb = get_packet(adap, fl, G_RSPD_LEN(len),
- eth ? SGE_RX_DROP_THRES : 0);
- if (!skb)
+ if (fl->use_pages) {
+ void *addr = fl->sdesc[fl->cidx].pg_chunk.va;
+
+ prefetch(addr);
+#if L1_CACHE_BYTES < 128
+ prefetch(addr + L1_CACHE_BYTES);
+#endif
+ __refill_fl(adap, fl);
+ if (lro > 0) {
+ lro_add_page(adap, qs, fl,
+ G_RSPD_LEN(len),
+ flags & F_RSPD_EOP);
+ goto next_fl;
+ }
+
+ skb = get_packet_pg(adap, fl, q,
+ G_RSPD_LEN(len),
+ eth ?
+ SGE_RX_DROP_THRES : 0);
+ q->pg_skb = skb;
+ } else
+ skb = get_packet(adap, fl, G_RSPD_LEN(len),
+ eth ? SGE_RX_DROP_THRES : 0);
+ if (unlikely(!skb)) {
+ if (!eth)
+ goto no_mem;
q->rx_drops++;
- else if (r->rss_hdr.opcode == CPL_TRACE_PKT)
+ } else if (unlikely(r->rss_hdr.opcode == CPL_TRACE_PKT))
__skb_pull(skb, 2);
- ethpad = 2;
+next_fl:
if (++fl->cidx == fl->size)
fl->cidx = 0;
} else
q->credits = 0;
}
- if (likely(skb != NULL)) {
+ packet_complete = flags &
+ (F_RSPD_EOP | F_RSPD_IMM_DATA_VALID |
+ F_RSPD_ASYNC_NOTIF);
+
+ if (skb != NULL && packet_complete) {
if (eth)
- rx_eth(adap, q, skb, ethpad);
+ rx_eth(adap, q, skb, ethpad, lro);
else {
+ q->offload_pkts++;
/* Preserve the RSS info in csum & priority */
skb->csum = rss_hi;
skb->priority = rss_lo;
ngathered = rx_offload(&adap->tdev, q, skb,
- offload_skbs, ngathered);
+ offload_skbs,
+ ngathered);
}
- }
+ if (flags & F_RSPD_EOP)
+ clear_rspq_bufstate(q);
+ }
--budget_left;
}
deliver_partial_bundle(&adap->tdev, q, offload_skbs, ngathered);
+
if (sleeping)
check_ring_db(adap, qs, sleeping);
static inline int is_pure_response(const struct rsp_desc *r)
{
- u32 n = ntohl(r->flags) & (F_RSPD_ASYNC_NOTIF | F_RSPD_IMM_DATA_VALID);
+ __be32 n = r->flags & htonl(F_RSPD_ASYNC_NOTIF | F_RSPD_IMM_DATA_VALID);
return (n | r->len_cq) == 0;
}
/**
* napi_rx_handler - the NAPI handler for Rx processing
- * @dev: the net device
+ * @napi: the napi instance
* @budget: how many packets we can process in this round
*
* Handler for new data events when using NAPI.
*/
-static int napi_rx_handler(struct net_device *dev, int *budget)
+static int napi_rx_handler(struct napi_struct *napi, int budget)
{
- struct adapter *adap = dev->priv;
- struct sge_qset *qs = dev2qset(dev);
- int effective_budget = min(*budget, dev->quota);
-
- int work_done = process_responses(adap, qs, effective_budget);
- *budget -= work_done;
- dev->quota -= work_done;
-
- if (work_done >= effective_budget)
- return 1;
+ struct sge_qset *qs = container_of(napi, struct sge_qset, napi);
+ struct adapter *adap = qs->adap;
+ int work_done = process_responses(adap, qs, budget);
- netif_rx_complete(dev);
+ if (likely(work_done < budget)) {
+ napi_complete(napi);
- /*
- * Because we don't atomically flush the following write it is
- * possible that in very rare cases it can reach the device in a way
- * that races with a new response being written plus an error interrupt
- * causing the NAPI interrupt handler below to return unhandled status
- * to the OS. To protect against this would require flushing the write
- * and doing both the write and the flush with interrupts off. Way too
- * expensive and unjustifiable given the rarity of the race.
- *
- * The race cannot happen at all with MSI-X.
- */
- t3_write_reg(adap, A_SG_GTS, V_RSPQ(qs->rspq.cntxt_id) |
- V_NEWTIMER(qs->rspq.next_holdoff) |
- V_NEWINDEX(qs->rspq.cidx));
- return 0;
+ /*
+ * Because we don't atomically flush the following
+ * write it is possible that in very rare cases it can
+ * reach the device in a way that races with a new
+ * response being written plus an error interrupt
+ * causing the NAPI interrupt handler below to return
+ * unhandled status to the OS. To protect against
+ * this would require flushing the write and doing
+ * both the write and the flush with interrupts off.
+ * Way too expensive and unjustifiable given the
+ * rarity of the race.
+ *
+ * The race cannot happen at all with MSI-X.
+ */
+ t3_write_reg(adap, A_SG_GTS, V_RSPQ(qs->rspq.cntxt_id) |
+ V_NEWTIMER(qs->rspq.next_holdoff) |
+ V_NEWINDEX(qs->rspq.cidx));
+ }
+ return work_done;
}
/*
* Returns true if the device is already scheduled for polling.
*/
-static inline int napi_is_scheduled(struct net_device *dev)
+static inline int napi_is_scheduled(struct napi_struct *napi)
{
- return test_bit(__LINK_STATE_RX_SCHED, &dev->state);
+ return test_bit(NAPI_STATE_SCHED, &napi->state);
}
/**
V_NEWTIMER(q->holdoff_tmr) | V_NEWINDEX(q->cidx));
return 0;
}
- if (likely(__netif_rx_schedule_prep(qs->netdev)))
- __netif_rx_schedule(qs->netdev);
+ napi_schedule(&qs->napi);
return 1;
}
irqreturn_t t3_sge_intr_msix(int irq, void *cookie)
{
struct sge_qset *qs = cookie;
- struct adapter *adap = qs->netdev->priv;
+ struct adapter *adap = qs->adap;
struct sge_rspq *q = &qs->rspq;
spin_lock(&q->lock);
* The MSI-X interrupt handler for an SGE response queue for the NAPI case
* (i.e., response queue serviced by NAPI polling).
*/
-irqreturn_t t3_sge_intr_msix_napi(int irq, void *cookie)
+static irqreturn_t t3_sge_intr_msix_napi(int irq, void *cookie)
{
struct sge_qset *qs = cookie;
- struct adapter *adap = qs->netdev->priv;
struct sge_rspq *q = &qs->rspq;
spin_lock(&q->lock);
- BUG_ON(napi_is_scheduled(qs->netdev));
- if (handle_responses(adap, q) < 0)
+ if (handle_responses(qs->adap, q) < 0)
q->unhandled_irqs++;
spin_unlock(&q->lock);
return IRQ_HANDLED;
return IRQ_HANDLED;
}
-static int rspq_check_napi(struct net_device *dev, struct sge_rspq *q)
+static int rspq_check_napi(struct sge_qset *qs)
{
- if (!napi_is_scheduled(dev) && is_new_response(&q->desc[q->cidx], q)) {
- if (likely(__netif_rx_schedule_prep(dev)))
- __netif_rx_schedule(dev);
+ struct sge_rspq *q = &qs->rspq;
+
+ if (!napi_is_scheduled(&qs->napi) &&
+ is_new_response(&q->desc[q->cidx], q)) {
+ napi_schedule(&qs->napi);
return 1;
}
return 0;
* one SGE response queue per port in this mode and protect all response
* queues with queue 0's lock.
*/
-irqreturn_t t3_intr_msi_napi(int irq, void *cookie)
+static irqreturn_t t3_intr_msi_napi(int irq, void *cookie)
{
int new_packets;
struct adapter *adap = cookie;
spin_lock(&q->lock);
- new_packets = rspq_check_napi(adap->sge.qs[0].netdev, q);
+ new_packets = rspq_check_napi(&adap->sge.qs[0]);
if (adap->params.nports == 2)
- new_packets += rspq_check_napi(adap->sge.qs[1].netdev,
- &adap->sge.qs[1].rspq);
+ new_packets += rspq_check_napi(&adap->sge.qs[1]);
if (!new_packets && t3_slow_intr_handler(adap) == 0)
q->unhandled_irqs++;
static irqreturn_t t3b_intr_napi(int irq, void *cookie)
{
u32 map;
- struct net_device *dev;
struct adapter *adap = cookie;
- struct sge_rspq *q0 = &adap->sge.qs[0].rspq;
+ struct sge_qset *qs0 = &adap->sge.qs[0];
+ struct sge_rspq *q0 = &qs0->rspq;
t3_write_reg(adap, A_PL_CLI, 0);
map = t3_read_reg(adap, A_SG_DATA_INTR);
if (unlikely(map & F_ERRINTR))
t3_slow_intr_handler(adap);
- if (likely(map & 1)) {
- dev = adap->sge.qs[0].netdev;
-
- if (likely(__netif_rx_schedule_prep(dev)))
- __netif_rx_schedule(dev);
- }
- if (map & 2) {
- dev = adap->sge.qs[1].netdev;
+ if (likely(map & 1))
+ napi_schedule(&qs0->napi);
- if (likely(__netif_rx_schedule_prep(dev)))
- __netif_rx_schedule(dev);
- }
+ if (map & 2)
+ napi_schedule(&adap->sge.qs[1].napi);
spin_unlock(&q0->lock);
return IRQ_HANDLED;
* (MSI-X, MSI, or legacy) and whether NAPI will be used to service the
* response queues.
*/
-intr_handler_t t3_intr_handler(struct adapter *adap, int polling)
+irq_handler_t t3_intr_handler(struct adapter *adap, int polling)
{
if (adap->flags & USING_MSIX)
return polling ? t3_sge_intr_msix_napi : t3_sge_intr_msix;
return t3_intr;
}
+#define SGE_PARERR (F_CPPARITYERROR | F_OCPARITYERROR | F_RCPARITYERROR | \
+ F_IRPARITYERROR | V_ITPARITYERROR(M_ITPARITYERROR) | \
+ V_FLPARITYERROR(M_FLPARITYERROR) | F_LODRBPARITYERROR | \
+ F_HIDRBPARITYERROR | F_LORCQPARITYERROR | \
+ F_HIRCQPARITYERROR)
+#define SGE_FRAMINGERR (F_UC_REQ_FRAMINGERROR | F_R_REQ_FRAMINGERROR)
+#define SGE_FATALERR (SGE_PARERR | SGE_FRAMINGERR | F_RSPQCREDITOVERFOW | \
+ F_RSPQDISABLED)
+
/**
* t3_sge_err_intr_handler - SGE async event interrupt handler
* @adapter: the adapter
*/
void t3_sge_err_intr_handler(struct adapter *adapter)
{
- unsigned int v, status = t3_read_reg(adapter, A_SG_INT_CAUSE);
+ unsigned int v, status = t3_read_reg(adapter, A_SG_INT_CAUSE) &
+ ~F_FLEMPTY;
+
+ if (status & SGE_PARERR)
+ CH_ALERT(adapter, "SGE parity error (0x%x)\n",
+ status & SGE_PARERR);
+ if (status & SGE_FRAMINGERR)
+ CH_ALERT(adapter, "SGE framing error (0x%x)\n",
+ status & SGE_FRAMINGERR);
if (status & F_RSPQCREDITOVERFOW)
CH_ALERT(adapter, "SGE response queue credit overflow\n");
"(0x%x)\n", (v >> S_RSPQ0DISABLED) & 0xff);
}
+ if (status & (F_HIPIODRBDROPERR | F_LOPIODRBDROPERR))
+ CH_ALERT(adapter, "SGE dropped %s priority doorbell\n",
+ status & F_HIPIODRBDROPERR ? "high" : "lo");
+
t3_write_reg(adapter, A_SG_INT_CAUSE, status);
- if (status & (F_RSPQCREDITOVERFOW | F_RSPQDISABLED))
+ if (status & SGE_FATALERR)
t3_fatal_err(adapter);
}
/**
- * sge_timer_cb - perform periodic maintenance of an SGE qset
+ * sge_timer_tx - perform periodic maintenance of an SGE qset
* @data: the SGE queue set to maintain
*
* Runs periodically from a timer to perform maintenance of an SGE queue
* set. It performs two tasks:
*
- * a) Cleans up any completed Tx descriptors that may still be pending.
+ * Cleans up any completed Tx descriptors that may still be pending.
* Normal descriptor cleanup happens when new packets are added to a Tx
* queue so this timer is relatively infrequent and does any cleanup only
* if the Tx queue has not seen any new packets in a while. We make a
* up). Since control queues use immediate data exclusively we don't
* bother cleaning them up here.
*
- * b) Replenishes Rx queues that have run out due to memory shortage.
+ */
+static void sge_timer_tx(unsigned long data)
+{
+ struct sge_qset *qs = (struct sge_qset *)data;
+ struct port_info *pi = netdev_priv(qs->netdev);
+ struct adapter *adap = pi->adapter;
+ unsigned int tbd[SGE_TXQ_PER_SET] = {0, 0};
+ unsigned long next_period;
+
+ if (__netif_tx_trylock(qs->tx_q)) {
+ tbd[TXQ_ETH] = reclaim_completed_tx(adap, &qs->txq[TXQ_ETH],
+ TX_RECLAIM_TIMER_CHUNK);
+ __netif_tx_unlock(qs->tx_q);
+ }
+
+ if (spin_trylock(&qs->txq[TXQ_OFLD].lock)) {
+ tbd[TXQ_OFLD] = reclaim_completed_tx(adap, &qs->txq[TXQ_OFLD],
+ TX_RECLAIM_TIMER_CHUNK);
+ spin_unlock(&qs->txq[TXQ_OFLD].lock);
+ }
+
+ next_period = TX_RECLAIM_PERIOD >>
+ (max(tbd[TXQ_ETH], tbd[TXQ_OFLD]) /
+ TX_RECLAIM_TIMER_CHUNK);
+ mod_timer(&qs->tx_reclaim_timer, jiffies + next_period);
+}
+
+/*
+ * sge_timer_rx - perform periodic maintenance of an SGE qset
+ * @data: the SGE queue set to maintain
+ *
+ * a) Replenishes Rx queues that have run out due to memory shortage.
* Normally new Rx buffers are added when existing ones are consumed but
* when out of memory a queue can become empty. We try to add only a few
* buffers here, the queue will be replenished fully as these new buffers
* are used up if memory shortage has subsided.
+ *
+ * b) Return coalesced response queue credits in case a response queue is
+ * starved.
+ *
*/
-static void sge_timer_cb(unsigned long data)
+static void sge_timer_rx(unsigned long data)
{
spinlock_t *lock;
struct sge_qset *qs = (struct sge_qset *)data;
- struct adapter *adap = qs->netdev->priv;
-
- if (spin_trylock(&qs->txq[TXQ_ETH].lock)) {
- reclaim_completed_tx(adap, &qs->txq[TXQ_ETH]);
- spin_unlock(&qs->txq[TXQ_ETH].lock);
- }
- if (spin_trylock(&qs->txq[TXQ_OFLD].lock)) {
- reclaim_completed_tx(adap, &qs->txq[TXQ_OFLD]);
- spin_unlock(&qs->txq[TXQ_OFLD].lock);
- }
- lock = (adap->flags & USING_MSIX) ? &qs->rspq.lock :
- &adap->sge.qs[0].rspq.lock;
- if (spin_trylock_irq(lock)) {
- if (!napi_is_scheduled(qs->netdev)) {
- if (qs->fl[0].credits < qs->fl[0].size)
- __refill_fl(adap, &qs->fl[0]);
- if (qs->fl[1].credits < qs->fl[1].size)
- __refill_fl(adap, &qs->fl[1]);
+ struct port_info *pi = netdev_priv(qs->netdev);
+ struct adapter *adap = pi->adapter;
+ u32 status;
+
+ lock = adap->params.rev > 0 ?
+ &qs->rspq.lock : &adap->sge.qs[0].rspq.lock;
+
+ if (!spin_trylock_irq(lock))
+ goto out;
+
+ if (napi_is_scheduled(&qs->napi))
+ goto unlock;
+
+ if (adap->params.rev < 4) {
+ status = t3_read_reg(adap, A_SG_RSPQ_FL_STATUS);
+
+ if (status & (1 << qs->rspq.cntxt_id)) {
+ qs->rspq.starved++;
+ if (qs->rspq.credits) {
+ qs->rspq.credits--;
+ refill_rspq(adap, &qs->rspq, 1);
+ qs->rspq.restarted++;
+ t3_write_reg(adap, A_SG_RSPQ_FL_STATUS,
+ 1 << qs->rspq.cntxt_id);
+ }
}
- spin_unlock_irq(lock);
}
- mod_timer(&qs->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
+
+ if (qs->fl[0].credits < qs->fl[0].size)
+ __refill_fl(adap, &qs->fl[0]);
+ if (qs->fl[1].credits < qs->fl[1].size)
+ __refill_fl(adap, &qs->fl[1]);
+
+unlock:
+ spin_unlock_irq(lock);
+out:
+ mod_timer(&qs->rx_reclaim_timer, jiffies + RX_RECLAIM_PERIOD);
}
/**
*/
void t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p)
{
- if (!qs->netdev)
- return;
-
qs->rspq.holdoff_tmr = max(p->coalesce_usecs * 10, 1U);/* can't be 0 */
qs->rspq.polling = p->polling;
- qs->netdev->poll = p->polling ? napi_rx_handler : ofld_poll;
+ qs->napi.poll = p->polling ? napi_rx_handler : ofld_poll;
}
/**
* @p: configuration parameters for this queue set
* @ntxq: number of Tx queues for the queue set
* @netdev: net device associated with this queue set
+ * @netdevq: net device TX queue associated with this queue set
*
* Allocate resources and initialize an SGE queue set. A queue set
* comprises a response queue, two Rx free-buffer queues, and up to 3
*/
int t3_sge_alloc_qset(struct adapter *adapter, unsigned int id, int nports,
int irq_vec_idx, const struct qset_params *p,
- int ntxq, struct net_device *netdev)
+ int ntxq, struct net_device *dev,
+ struct netdev_queue *netdevq)
{
- int i, ret = -ENOMEM;
+ int i, avail, ret = -ENOMEM;
struct sge_qset *q = &adapter->sge.qs[id];
init_qset_cntxt(q, id);
- init_timer(&q->tx_reclaim_timer);
- q->tx_reclaim_timer.data = (unsigned long)q;
- q->tx_reclaim_timer.function = sge_timer_cb;
+ setup_timer(&q->tx_reclaim_timer, sge_timer_tx, (unsigned long)q);
+ setup_timer(&q->rx_reclaim_timer, sge_timer_rx, (unsigned long)q);
q->fl[0].desc = alloc_ring(adapter->pdev, p->fl_size,
sizeof(struct rx_desc),
q->rspq.gen = 1;
q->rspq.size = p->rspq_size;
spin_lock_init(&q->rspq.lock);
+ skb_queue_head_init(&q->rspq.rx_queue);
q->txq[TXQ_ETH].stop_thres = nports *
flits_to_desc(sgl_len(MAX_SKB_FRAGS + 1) + 3);
- if (ntxq == 1) {
- q->fl[0].buf_size = SGE_RX_SM_BUF_SIZE + 2 +
- sizeof(struct cpl_rx_pkt);
- q->fl[1].buf_size = MAX_FRAME_SIZE + 2 +
- sizeof(struct cpl_rx_pkt);
- } else {
- q->fl[0].buf_size = SGE_RX_SM_BUF_SIZE +
- sizeof(struct cpl_rx_data);
- q->fl[1].buf_size = (16 * 1024) -
- SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
- }
+#if FL0_PG_CHUNK_SIZE > 0
+ q->fl[0].buf_size = FL0_PG_CHUNK_SIZE;
+#else
+ q->fl[0].buf_size = SGE_RX_SM_BUF_SIZE + sizeof(struct cpl_rx_data);
+#endif
+#if FL1_PG_CHUNK_SIZE > 0
+ q->fl[1].buf_size = FL1_PG_CHUNK_SIZE;
+#else
+ q->fl[1].buf_size = is_offload(adapter) ?
+ (16 * 1024) - SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) :
+ MAX_FRAME_SIZE + 2 + sizeof(struct cpl_rx_pkt);
+#endif
- spin_lock(&adapter->sge.reg_lock);
+ q->fl[0].use_pages = FL0_PG_CHUNK_SIZE > 0;
+ q->fl[1].use_pages = FL1_PG_CHUNK_SIZE > 0;
+ q->fl[0].order = FL0_PG_ORDER;
+ q->fl[1].order = FL1_PG_ORDER;
+ q->fl[0].alloc_size = FL0_PG_ALLOC_SIZE;
+ q->fl[1].alloc_size = FL1_PG_ALLOC_SIZE;
+
+ spin_lock_irq(&adapter->sge.reg_lock);
/* FL threshold comparison uses < */
ret = t3_sge_init_rspcntxt(adapter, q->rspq.cntxt_id, irq_vec_idx,
q->rspq.phys_addr, q->rspq.size,
- q->fl[0].buf_size, 1, 0);
+ q->fl[0].buf_size - SGE_PG_RSVD, 1, 0);
if (ret)
goto err_unlock;
for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
ret = t3_sge_init_flcntxt(adapter, q->fl[i].cntxt_id, 0,
q->fl[i].phys_addr, q->fl[i].size,
- q->fl[i].buf_size, p->cong_thres, 1,
- 0);
+ q->fl[i].buf_size - SGE_PG_RSVD,
+ p->cong_thres, 1, 0);
if (ret)
goto err_unlock;
}
goto err_unlock;
}
- spin_unlock(&adapter->sge.reg_lock);
- q->netdev = netdev;
- t3_update_qset_coalesce(q, p);
+ spin_unlock_irq(&adapter->sge.reg_lock);
- /*
- * We use atalk_ptr as a backpointer to a qset. In case a device is
- * associated with multiple queue sets only the first one sets
- * atalk_ptr.
- */
- if (netdev->atalk_ptr == NULL)
- netdev->atalk_ptr = q;
+ q->adap = adapter;
+ q->netdev = dev;
+ q->tx_q = netdevq;
+ t3_update_qset_coalesce(q, p);
- refill_fl(adapter, &q->fl[0], q->fl[0].size, GFP_KERNEL);
- refill_fl(adapter, &q->fl[1], q->fl[1].size, GFP_KERNEL);
+ avail = refill_fl(adapter, &q->fl[0], q->fl[0].size,
+ GFP_KERNEL | __GFP_COMP);
+ if (!avail) {
+ CH_ALERT(adapter, "free list queue 0 initialization failed\n");
+ goto err;
+ }
+ if (avail < q->fl[0].size)
+ CH_WARN(adapter, "free list queue 0 enabled with %d credits\n",
+ avail);
+
+ avail = refill_fl(adapter, &q->fl[1], q->fl[1].size,
+ GFP_KERNEL | __GFP_COMP);
+ if (avail < q->fl[1].size)
+ CH_WARN(adapter, "free list queue 1 enabled with %d credits\n",
+ avail);
refill_rspq(adapter, &q->rspq, q->rspq.size - 1);
t3_write_reg(adapter, A_SG_GTS, V_RSPQ(q->rspq.cntxt_id) |
V_NEWTIMER(q->rspq.holdoff_tmr));
- mod_timer(&q->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
return 0;
- err_unlock:
- spin_unlock(&adapter->sge.reg_lock);
- err:
+err_unlock:
+ spin_unlock_irq(&adapter->sge.reg_lock);
+err:
t3_free_qset(adapter, q);
return ret;
}
/**
+ * t3_start_sge_timers - start SGE timer call backs
+ * @adap: the adapter
+ *
+ * Starts each SGE queue set's timer call back
+ */
+void t3_start_sge_timers(struct adapter *adap)
+{
+ int i;
+
+ for (i = 0; i < SGE_QSETS; ++i) {
+ struct sge_qset *q = &adap->sge.qs[i];
+
+ if (q->tx_reclaim_timer.function)
+ mod_timer(&q->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
+
+ if (q->rx_reclaim_timer.function)
+ mod_timer(&q->rx_reclaim_timer, jiffies + RX_RECLAIM_PERIOD);
+ }
+}
+
+/**
+ * t3_stop_sge_timers - stop SGE timer call backs
+ * @adap: the adapter
+ *
+ * Stops each SGE queue set's timer call back
+ */
+void t3_stop_sge_timers(struct adapter *adap)
+{
+ int i;
+
+ for (i = 0; i < SGE_QSETS; ++i) {
+ struct sge_qset *q = &adap->sge.qs[i];
+
+ if (q->tx_reclaim_timer.function)
+ del_timer_sync(&q->tx_reclaim_timer);
+ if (q->rx_reclaim_timer.function)
+ del_timer_sync(&q->rx_reclaim_timer);
+ }
+}
+
+/**
* t3_free_sge_resources - free SGE resources
* @adap: the adapter
*
unsigned int ctrl, ups = ffs(pci_resource_len(adap->pdev, 2) >> 12);
ctrl = F_DROPPKT | V_PKTSHIFT(2) | F_FLMODE | F_AVOIDCQOVFL |
- F_CQCRDTCTRL |
+ F_CQCRDTCTRL | F_CONGMODE | F_TNLFLMODE | F_FATLPERREN |
V_HOSTPAGESIZE(PAGE_SHIFT - 11) | F_BIGENDIANINGRESS |
V_USERSPACESIZE(ups ? ups - 1 : 0) | F_ISCSICOALESCING;
#if SGE_NUM_GENBITS == 1
if (adap->params.rev > 0) {
if (!(adap->flags & (USING_MSIX | USING_MSI)))
ctrl |= F_ONEINTMULTQ | F_OPTONEINTMULTQ;
- ctrl |= F_CQCRDTCTRL | F_AVOIDCQOVFL;
}
t3_write_reg(adap, A_SG_CONTROL, ctrl);
t3_write_reg(adap, A_SG_EGR_RCQ_DRB_THRSH, V_HIRCQDRBTHRSH(512) |
t3_write_reg(adap, A_SG_TIMER_TICK, core_ticks_per_usec(adap) / 10);
t3_write_reg(adap, A_SG_CMDQ_CREDIT_TH, V_THRESHOLD(32) |
V_TIMEOUT(200 * core_ticks_per_usec(adap)));
- t3_write_reg(adap, A_SG_HI_DRB_HI_THRSH, 1000);
+ t3_write_reg(adap, A_SG_HI_DRB_HI_THRSH,
+ adap->params.rev < T3_REV_C ? 1000 : 500);
t3_write_reg(adap, A_SG_HI_DRB_LO_THRSH, 256);
t3_write_reg(adap, A_SG_LO_DRB_HI_THRSH, 1000);
t3_write_reg(adap, A_SG_LO_DRB_LO_THRSH, 256);
* defaults for the assorted SGE parameters, which admins can change until
* they are used to initialize the SGE.
*/
-void __devinit t3_sge_prep(struct adapter *adap, struct sge_params *p)
+void t3_sge_prep(struct adapter *adap, struct sge_params *p)
{
int i;
q->polling = adap->params.rev > 0;
q->coalesce_usecs = 5;
q->rspq_size = 1024;
- q->fl_size = 4096;
- q->jumbo_size = 512;
+ q->fl_size = 1024;
+ q->jumbo_size = 512;
q->txq_size[TXQ_ETH] = 1024;
q->txq_size[TXQ_OFLD] = 1024;
q->txq_size[TXQ_CTRL] = 256;
git://ftp.safe.ca / safe / jmp / linux-2.6 / blobdiff