Merge branch 'for-next' into for-linus
[safe/jmp/linux-2.6] / drivers / net / chelsio / sge.c
index ac7c46b..55d99ca 100644 (file)
 #define SGE_FREEL_REFILL_THRESH        16
 #define SGE_RESPQ_E_N          1024
 #define SGE_INTRTIMER_NRES     1000
-#define SGE_RX_COPY_THRES      256
 #define SGE_RX_SM_BUF_SIZE     1536
 #define SGE_TX_DESC_MAX_PLEN   16384
 
-# define SGE_RX_DROP_THRES 2
-
 #define SGE_RESPQ_REPLENISH_THRES (SGE_RESPQ_E_N / 4)
 
 /*
  */
 #define TX_RECLAIM_PERIOD (HZ / 4)
 
-#ifndef NET_IP_ALIGN
-# define NET_IP_ALIGN 2
-#endif
-
 #define M_CMD_LEN       0x7fffffff
 #define V_CMD_LEN(v)    (v)
 #define G_CMD_LEN(v)    ((v) & M_CMD_LEN)
@@ -255,7 +248,7 @@ static void restart_sched(unsigned long);
  *
  * Interrupts are handled by a single CPU and it is likely that on a MP system
  * the application is migrated to another CPU. In that scenario, we try to
- * seperate the RX(in irq context) and TX state in order to decrease memory
+ * separate the RX(in irq context) and TX state in order to decrease memory
  * contention.
  */
 struct sge {
@@ -274,7 +267,7 @@ struct sge {
        struct sk_buff  *espibug_skb[MAX_NPORTS];
        u32             sge_control;    /* shadow value of sge control reg */
        struct sge_intr_counts stats;
-       struct sge_port_stats *port_stats[MAX_NPORTS];
+       struct sge_port_stats __percpu *port_stats[MAX_NPORTS];
        struct sched    *tx_sched;
        struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned_in_smp;
 };
@@ -337,6 +330,8 @@ unsigned int t1_sched_update_parms(struct sge *sge, unsigned int port,
        return max_avail_segs * (p->mtu - 40);
 }
 
+#if 0
+
 /*
  * t1_sched_max_avail_bytes() tells the scheduler the maximum amount of
  * data that can be pushed per port.
@@ -364,6 +359,8 @@ void t1_sched_set_drain_bits_per_us(struct sge *sge, unsigned int port,
        t1_sched_update_parms(sge, port, 0, 0);
 }
 
+#endif  /*  0  */
+
 
 /*
  * get_clock() implements a ns clock (see ktime_get)
@@ -575,11 +572,10 @@ static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
                q->size = p->freelQ_size[i];
                q->dma_offset = sge->rx_pkt_pad ? 0 : NET_IP_ALIGN;
                size = sizeof(struct freelQ_e) * q->size;
-               q->entries = (struct freelQ_e *)
-                             pci_alloc_consistent(pdev, size, &q->dma_addr);
+               q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
                if (!q->entries)
                        goto err_no_mem;
-               memset(q->entries, 0, size);
+
                size = sizeof(struct freelQ_ce) * q->size;
                q->centries = kzalloc(size, GFP_KERNEL);
                if (!q->centries)
@@ -613,11 +609,10 @@ static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
        sge->respQ.size = SGE_RESPQ_E_N;
        sge->respQ.credits = 0;
        size = sizeof(struct respQ_e) * sge->respQ.size;
-       sge->respQ.entries = (struct respQ_e *)
+       sge->respQ.entries =
                pci_alloc_consistent(pdev, size, &sge->respQ.dma_addr);
        if (!sge->respQ.entries)
                goto err_no_mem;
-       memset(sge->respQ.entries, 0, size);
        return 0;
 
 err_no_mem:
@@ -637,20 +632,12 @@ static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *q, unsigned int n)
        q->in_use -= n;
        ce = &q->centries[cidx];
        while (n--) {
-               if (q->sop) {
-                       if (likely(pci_unmap_len(ce, dma_len))) {
-                               pci_unmap_single(pdev,
-                                                pci_unmap_addr(ce, dma_addr),
-                                                pci_unmap_len(ce, dma_len),
-                                                PCI_DMA_TODEVICE);
+               if (likely(pci_unmap_len(ce, dma_len))) {
+                       pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
+                                        pci_unmap_len(ce, dma_len),
+                                        PCI_DMA_TODEVICE);
+                       if (q->sop)
                                q->sop = 0;
-                       }
-               } else {
-                       if (likely(pci_unmap_len(ce, dma_len))) {
-                               pci_unmap_page(pdev, pci_unmap_addr(ce, dma_addr),
-                                              pci_unmap_len(ce, dma_len),
-                                              PCI_DMA_TODEVICE);
-                       }
                }
                if (ce->skb) {
                        dev_kfree_skb_any(ce->skb);
@@ -711,11 +698,10 @@ static int alloc_tx_resources(struct sge *sge, struct sge_params *p)
                q->stop_thres = 0;
                spin_lock_init(&q->lock);
                size = sizeof(struct cmdQ_e) * q->size;
-               q->entries = (struct cmdQ_e *)
-                             pci_alloc_consistent(pdev, size, &q->dma_addr);
+               q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
                if (!q->entries)
                        goto err_no_mem;
-               memset(q->entries, 0, size);
+
                size = sizeof(struct cmdQ_ce) * q->size;
                q->centries = kzalloc(size, GFP_KERNEL);
                if (!q->centries)
@@ -861,6 +847,8 @@ static void refill_free_list(struct sge *sge, struct freelQ *q)
                skb_reserve(skb, q->dma_offset);
                mapping = pci_map_single(pdev, skb->data, dma_len,
                                         PCI_DMA_FROMDEVICE);
+               skb_reserve(skb, sge->rx_pkt_pad);
+
                ce->skb = skb;
                pci_unmap_addr_set(ce, dma_addr, mapping);
                pci_unmap_len_set(ce, dma_len, dma_len);
@@ -965,7 +953,7 @@ int t1_sge_intr_error_handler(struct sge *sge)
                sge->stats.respQ_empty++;
        if (cause & F_RESPQ_OVERFLOW) {
                sge->stats.respQ_overflow++;
-               CH_ALERT("%s: SGE response queue overflow\n",
+               pr_alert("%s: SGE response queue overflow\n",
                         adapter->name);
        }
        if (cause & F_FL_EXHAUSTED) {
@@ -974,12 +962,12 @@ int t1_sge_intr_error_handler(struct sge *sge)
        }
        if (cause & F_PACKET_TOO_BIG) {
                sge->stats.pkt_too_big++;
-               CH_ALERT("%s: SGE max packet size exceeded\n",
+               pr_alert("%s: SGE max packet size exceeded\n",
                         adapter->name);
        }
        if (cause & F_PACKET_MISMATCH) {
                sge->stats.pkt_mismatch++;
-               CH_ALERT("%s: SGE packet mismatch\n", adapter->name);
+               pr_alert("%s: SGE packet mismatch\n", adapter->name);
        }
        if (cause & SGE_INT_FATAL)
                t1_fatal_err(adapter);
@@ -1002,11 +990,10 @@ void t1_sge_get_port_stats(const struct sge *sge, int port,
        for_each_possible_cpu(cpu) {
                struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[port], cpu);
 
-               ss->rx_packets += st->rx_packets;
                ss->rx_cso_good += st->rx_cso_good;
-               ss->tx_packets += st->tx_packets;
                ss->tx_cso += st->tx_cso;
                ss->tx_tso += st->tx_tso;
+               ss->tx_need_hdrroom += st->tx_need_hdrroom;
                ss->vlan_xtract += st->vlan_xtract;
                ss->vlan_insert += st->vlan_insert;
        }
@@ -1039,15 +1026,15 @@ static void recycle_fl_buf(struct freelQ *fl, int idx)
        }
 }
 
+static int copybreak __read_mostly = 256;
+module_param(copybreak, int, 0);
+MODULE_PARM_DESC(copybreak, "Receive copy threshold");
+
 /**
  *     get_packet - return the next ingress packet buffer
  *     @pdev: the PCI device that received the packet
  *     @fl: the SGE free list holding the packet
  *     @len: the actual packet length, excluding any SGE padding
- *     @dma_pad: padding at beginning of buffer left by SGE DMA
- *     @skb_pad: padding to be used if the packet is copied
- *     @copy_thres: length threshold under which a packet should be copied
- *     @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
@@ -1058,45 +1045,42 @@ static void recycle_fl_buf(struct freelQ *fl, int idx)
  *     be copied but there is no memory for the copy.
  */
 static inline struct sk_buff *get_packet(struct pci_dev *pdev,
-                                        struct freelQ *fl, unsigned int len,
-                                        int dma_pad, int skb_pad,
-                                        unsigned int copy_thres,
-                                        unsigned int drop_thres)
+                                        struct freelQ *fl, unsigned int len)
 {
        struct sk_buff *skb;
-       struct freelQ_ce *ce = &fl->centries[fl->cidx];
+       const struct freelQ_ce *ce = &fl->centries[fl->cidx];
 
-       if (len < copy_thres) {
-               skb = alloc_skb(len + skb_pad, GFP_ATOMIC);
-               if (likely(skb != NULL)) {
-                       skb_reserve(skb, skb_pad);
-                       skb_put(skb, len);
-                       pci_dma_sync_single_for_cpu(pdev,
-                                           pci_unmap_addr(ce, dma_addr),
-                                           pci_unmap_len(ce, dma_len),
-                                           PCI_DMA_FROMDEVICE);
-                       memcpy(skb->data, ce->skb->data + dma_pad, len);
-                       pci_dma_sync_single_for_device(pdev,
+       if (len < copybreak) {
+               skb = alloc_skb(len + 2, GFP_ATOMIC);
+               if (!skb)
+                       goto use_orig_buf;
+
+               skb_reserve(skb, 2);    /* align IP header */
+               skb_put(skb, len);
+               pci_dma_sync_single_for_cpu(pdev,
                                            pci_unmap_addr(ce, dma_addr),
                                            pci_unmap_len(ce, dma_len),
                                            PCI_DMA_FROMDEVICE);
-               } else if (!drop_thres)
-                       goto use_orig_buf;
-
+               skb_copy_from_linear_data(ce->skb, skb->data, len);
+               pci_dma_sync_single_for_device(pdev,
+                                              pci_unmap_addr(ce, dma_addr),
+                                              pci_unmap_len(ce, dma_len),
+                                              PCI_DMA_FROMDEVICE);
                recycle_fl_buf(fl, fl->cidx);
                return skb;
        }
 
-       if (fl->credits < drop_thres) {
+use_orig_buf:
+       if (fl->credits < 2) {
                recycle_fl_buf(fl, fl->cidx);
                return NULL;
        }
 
-use_orig_buf:
        pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
                         pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
        skb = ce->skb;
-       skb_reserve(skb, dma_pad);
+       prefetch(skb->data);
+
        skb_put(skb, len);
        return skb;
 }
@@ -1117,7 +1101,7 @@ static void unexpected_offload(struct adapter *adapter, struct freelQ *fl)
 
        pci_dma_sync_single_for_cpu(adapter->pdev, pci_unmap_addr(ce, dma_addr),
                            pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
-       CH_ERR("%s: unexpected offload packet, cmd %u\n",
+       pr_err("%s: unexpected offload packet, cmd %u\n",
               adapter->name, *skb->data);
        recycle_fl_buf(fl, fl->cidx);
 }
@@ -1165,8 +1149,8 @@ static inline void write_tx_desc(struct cmdQ_e *e, dma_addr_t mapping,
                                 unsigned int len, unsigned int gen,
                                 unsigned int eop)
 {
-       if (unlikely(len > SGE_TX_DESC_MAX_PLEN))
-               BUG();
+       BUG_ON(len > SGE_TX_DESC_MAX_PLEN);
+
        e->addr_lo = (u32)mapping;
        e->addr_hi = (u64)mapping >> 32;
        e->len_gen = V_CMD_LEN(len) | V_CMD_GEN1(gen);
@@ -1374,34 +1358,29 @@ static void restart_sched(unsigned long arg)
  *
  *     Process an ingress ethernet pakcet and deliver it to the stack.
  */
-static int sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
+static void sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
 {
        struct sk_buff *skb;
-       struct cpl_rx_pkt *p;
+       const struct cpl_rx_pkt *p;
        struct adapter *adapter = sge->adapter;
        struct sge_port_stats *st;
 
-       skb = get_packet(adapter->pdev, fl, len - sge->rx_pkt_pad,
-                        sge->rx_pkt_pad, 2, SGE_RX_COPY_THRES,
-                        SGE_RX_DROP_THRES);
+       skb = get_packet(adapter->pdev, fl, len - sge->rx_pkt_pad);
        if (unlikely(!skb)) {
                sge->stats.rx_drops++;
-               return 0;
+               return;
        }
 
-       p = (struct cpl_rx_pkt *)skb->data;
-       skb_pull(skb, sizeof(*p));
+       p = (const struct cpl_rx_pkt *) skb->data;
        if (p->iff >= adapter->params.nports) {
                kfree_skb(skb);
-               return 0;
+               return;
        }
+       __skb_pull(skb, sizeof(*p));
 
-       skb->dev = adapter->port[p->iff].dev;
-       skb->dev->last_rx = jiffies;
-       st = per_cpu_ptr(sge->port_stats[p->iff], smp_processor_id());
-       st->rx_packets++;
+       st = this_cpu_ptr(sge->port_stats[p->iff]);
 
-       skb->protocol = eth_type_trans(skb, skb->dev);
+       skb->protocol = eth_type_trans(skb, adapter->port[p->iff].dev);
        if ((adapter->flags & RX_CSUM_ENABLED) && p->csum == 0xffff &&
            skb->protocol == htons(ETH_P_IP) &&
            (skb->data[9] == IPPROTO_TCP || skb->data[9] == IPPROTO_UDP)) {
@@ -1412,21 +1391,10 @@ static int sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
 
        if (unlikely(adapter->vlan_grp && p->vlan_valid)) {
                st->vlan_xtract++;
-#ifdef CONFIG_CHELSIO_T1_NAPI
-                       vlan_hwaccel_receive_skb(skb, adapter->vlan_grp,
-                                                ntohs(p->vlan));
-#else
-                       vlan_hwaccel_rx(skb, adapter->vlan_grp,
-                                       ntohs(p->vlan));
-#endif
-       } else {
-#ifdef CONFIG_CHELSIO_T1_NAPI
+               vlan_hwaccel_receive_skb(skb, adapter->vlan_grp,
+                                        ntohs(p->vlan));
+       } else
                netif_receive_skb(skb);
-#else
-               netif_rx(skb);
-#endif
-       }
-       return 0;
 }
 
 /*
@@ -1447,19 +1415,18 @@ static inline int enough_free_Tx_descs(const struct cmdQ *q)
 static void restart_tx_queues(struct sge *sge)
 {
        struct adapter *adap = sge->adapter;
+       int i;
 
-       if (enough_free_Tx_descs(&sge->cmdQ[0])) {
-               int i;
+       if (!enough_free_Tx_descs(&sge->cmdQ[0]))
+               return;
 
-               for_each_port(adap, i) {
-                       struct net_device *nd = adap->port[i].dev;
+       for_each_port(adap, i) {
+               struct net_device *nd = adap->port[i].dev;
 
-                       if (test_and_clear_bit(nd->if_port,
-                                              &sge->stopped_tx_queues) &&
-                           netif_running(nd)) {
-                               sge->stats.cmdQ_restarted[2]++;
-                               netif_wake_queue(nd);
-                       }
+               if (test_and_clear_bit(nd->if_port, &sge->stopped_tx_queues) &&
+                   netif_running(nd)) {
+                       sge->stats.cmdQ_restarted[2]++;
+                       netif_wake_queue(nd);
                }
        }
 }
@@ -1509,12 +1476,11 @@ static int process_responses(struct adapter *adapter, int budget)
        struct sge *sge = adapter->sge;
        struct respQ *q = &sge->respQ;
        struct respQ_e *e = &q->entries[q->cidx];
-       int budget_left = budget;
+       int done = 0;
        unsigned int flags = 0;
        unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
 
-
-       while (likely(budget_left && e->GenerationBit == q->genbit)) {
+       while (done < budget && e->GenerationBit == q->genbit) {
                flags |= e->Qsleeping;
 
                cmdq_processed[0] += e->Cmdq0CreditReturn;
@@ -1524,14 +1490,16 @@ static int process_responses(struct adapter *adapter, int budget)
                 * ping-pong of TX state information on MP where the sender
                 * might run on a different CPU than this function...
                 */
-               if (unlikely(flags & F_CMDQ0_ENABLE || cmdq_processed[0] > 64)) {
+               if (unlikely((flags & F_CMDQ0_ENABLE) || cmdq_processed[0] > 64)) {
                        flags = update_tx_info(adapter, flags, cmdq_processed[0]);
                        cmdq_processed[0] = 0;
                }
+
                if (unlikely(cmdq_processed[1] > 16)) {
                        sge->cmdQ[1].processed += cmdq_processed[1];
                        cmdq_processed[1] = 0;
                }
+
                if (likely(e->DataValid)) {
                        struct freelQ *fl = &sge->freelQ[e->FreelistQid];
 
@@ -1541,12 +1509,16 @@ static int process_responses(struct adapter *adapter, int budget)
                        else
                                sge_rx(sge, fl, e->BufferLength);
 
+                       ++done;
+
                        /*
                         * Note: this depends on each packet consuming a
                         * single free-list buffer; cf. the BUG above.
                         */
                        if (++fl->cidx == fl->size)
                                fl->cidx = 0;
+                       prefetch(fl->centries[fl->cidx].skb);
+
                        if (unlikely(--fl->credits <
                                     fl->size - SGE_FREEL_REFILL_THRESH))
                                refill_free_list(sge, fl);
@@ -1565,17 +1537,22 @@ static int process_responses(struct adapter *adapter, int budget)
                        writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
                        q->credits = 0;
                }
-               --budget_left;
        }
 
        flags = update_tx_info(adapter, flags, cmdq_processed[0]);
        sge->cmdQ[1].processed += cmdq_processed[1];
 
-       budget -= budget_left;
-       return budget;
+       return done;
+}
+
+static inline int responses_pending(const struct adapter *adapter)
+{
+       const struct respQ *Q = &adapter->sge->respQ;
+       const struct respQ_e *e = &Q->entries[Q->cidx];
+
+       return (e->GenerationBit == Q->genbit);
 }
 
-#ifdef CONFIG_CHELSIO_T1_NAPI
 /*
  * A simpler version of process_responses() that handles only pure (i.e.,
  * non data-carrying) responses.  Such respones are too light-weight to justify
@@ -1584,13 +1561,19 @@ static int process_responses(struct adapter *adapter, int budget)
  * which the caller must ensure is a valid pure response.  Returns 1 if it
  * encounters a valid data-carrying response, 0 otherwise.
  */
-static int process_pure_responses(struct adapter *adapter, struct respQ_e *e)
+static int process_pure_responses(struct adapter *adapter)
 {
        struct sge *sge = adapter->sge;
        struct respQ *q = &sge->respQ;
+       struct respQ_e *e = &q->entries[q->cidx];
+       const struct freelQ *fl = &sge->freelQ[e->FreelistQid];
        unsigned int flags = 0;
        unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
 
+       prefetch(fl->centries[fl->cidx].skb);
+       if (e->DataValid)
+               return 1;
+
        do {
                flags |= e->Qsleeping;
 
@@ -1623,124 +1606,50 @@ static int process_pure_responses(struct adapter *adapter, struct respQ_e *e)
  * or protection from interrupts as data interrupts are off at this point and
  * other adapter interrupts do not interfere.
  */
-int t1_poll(struct net_device *dev, int *budget)
+int t1_poll(struct napi_struct *napi, int budget)
 {
-       struct adapter *adapter = dev->priv;
-       int effective_budget = min(*budget, dev->quota);
-       int work_done = process_responses(adapter, effective_budget);
-
-       *budget -= work_done;
-       dev->quota -= work_done;
-
-       if (work_done >= effective_budget)
-               return 1;
-
-       spin_lock_irq(&adapter->async_lock);
-       __netif_rx_complete(dev);
-       writel(adapter->sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
-       writel(adapter->slow_intr_mask | F_PL_INTR_SGE_DATA,
-              adapter->regs + A_PL_ENABLE);
-       spin_unlock_irq(&adapter->async_lock);
+       struct adapter *adapter = container_of(napi, struct adapter, napi);
+       int work_done = process_responses(adapter, budget);
 
-       return 0;
+       if (likely(work_done < budget)) {
+               napi_complete(napi);
+               writel(adapter->sge->respQ.cidx,
+                      adapter->regs + A_SG_SLEEPING);
+       }
+       return work_done;
 }
 
-/*
- * NAPI version of the main interrupt handler.
- */
 irqreturn_t t1_interrupt(int irq, void *data)
 {
        struct adapter *adapter = data;
-       struct net_device *dev = adapter->sge->netdev;
        struct sge *sge = adapter->sge;
-       u32 cause;
-       int handled = 0;
-
-       cause = readl(adapter->regs + A_PL_CAUSE);
-       if (cause == 0 || cause == ~0)
-               return IRQ_NONE;
-
-       spin_lock(&adapter->async_lock);
-       if (cause & F_PL_INTR_SGE_DATA) {
-               struct respQ *q = &adapter->sge->respQ;
-               struct respQ_e *e = &q->entries[q->cidx];
+       int handled;
 
-               handled = 1;
+       if (likely(responses_pending(adapter))) {
                writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
 
-               if (e->GenerationBit == q->genbit &&
-                   __netif_rx_schedule_prep(dev)) {
-                       if (e->DataValid || process_pure_responses(adapter, e)) {
-                               /* mask off data IRQ */
-                               writel(adapter->slow_intr_mask,
-                                      adapter->regs + A_PL_ENABLE);
-                               __netif_rx_schedule(sge->netdev);
-                               goto unlock;
+               if (napi_schedule_prep(&adapter->napi)) {
+                       if (process_pure_responses(adapter))
+                               __napi_schedule(&adapter->napi);
+                       else {
+                               /* no data, no NAPI needed */
+                               writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
+                               /* undo schedule_prep */
+                               napi_enable(&adapter->napi);
                        }
-                       /* no data, no NAPI needed */
-                       netif_poll_enable(dev);
-
                }
-               writel(q->cidx, adapter->regs + A_SG_SLEEPING);
-       } else
-               handled = t1_slow_intr_handler(adapter);
-
-       if (!handled)
-               sge->stats.unhandled_irqs++;
-unlock:
-       spin_unlock(&adapter->async_lock);
-       return IRQ_RETVAL(handled != 0);
-}
-
-#else
-/*
- * Main interrupt handler, optimized assuming that we took a 'DATA'
- * interrupt.
- *
- * 1. Clear the interrupt
- * 2. Loop while we find valid descriptors and process them; accumulate
- *      information that can be processed after the loop
- * 3. Tell the SGE at which index we stopped processing descriptors
- * 4. Bookkeeping; free TX buffers, ring doorbell if there are any
- *      outstanding TX buffers waiting, replenish RX buffers, potentially
- *      reenable upper layers if they were turned off due to lack of TX
- *      resources which are available again.
- * 5. If we took an interrupt, but no valid respQ descriptors was found we
- *      let the slow_intr_handler run and do error handling.
- */
-irqreturn_t t1_interrupt(int irq, void *cookie)
-{
-       int work_done;
-       struct respQ_e *e;
-       struct adapter *adapter = cookie;
-       struct respQ *Q = &adapter->sge->respQ;
+               return IRQ_HANDLED;
+       }
 
        spin_lock(&adapter->async_lock);
-       e = &Q->entries[Q->cidx];
-       prefetch(e);
-
-       writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
-
-       if (likely(e->GenerationBit == Q->genbit))
-               work_done = process_responses(adapter, -1);
-       else
-               work_done = t1_slow_intr_handler(adapter);
+       handled = t1_slow_intr_handler(adapter);
+       spin_unlock(&adapter->async_lock);
 
-       /*
-        * The unconditional clearing of the PL_CAUSE above may have raced
-        * with DMA completion and the corresponding generation of a response
-        * to cause us to miss the resulting data interrupt.  The next write
-        * is also unconditional to recover the missed interrupt and render
-        * this race harmless.
-        */
-       writel(Q->cidx, adapter->regs + A_SG_SLEEPING);
+       if (!handled)
+               sge->stats.unhandled_irqs++;
 
-       if (!work_done)
-               adapter->sge->stats.unhandled_irqs++;
-       spin_unlock(&adapter->async_lock);
-       return IRQ_RETVAL(work_done != 0);
+       return IRQ_RETVAL(handled != 0);
 }
-#endif
 
 /*
  * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
@@ -1778,7 +1687,7 @@ static int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
                        netif_stop_queue(dev);
                        set_bit(dev->if_port, &sge->stopped_tx_queues);
                        sge->stats.cmdQ_full[2]++;
-                       CH_ERR("%s: Tx ring full while queue awake!\n",
+                       pr_err("%s: Tx ring full while queue awake!\n",
                               adapter->name);
                }
                spin_unlock(&q->lock);
@@ -1867,11 +1776,11 @@ static inline int eth_hdr_len(const void *data)
 /*
  * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
  */
-int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
+netdev_tx_t t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
-       struct adapter *adapter = dev->priv;
+       struct adapter *adapter = dev->ml_priv;
        struct sge *sge = adapter->sge;
-       struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[dev->if_port], smp_processor_id());
+       struct sge_port_stats *st = this_cpu_ptr(sge->port_stats[dev->if_port]);
        struct cpl_tx_pkt *cpl;
        struct sk_buff *orig_skb = skb;
        int ret;
@@ -1879,20 +1788,32 @@ int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
        if (skb->protocol == htons(ETH_P_CPL5))
                goto send;
 
+       /*
+        * We are using a non-standard hard_header_len.
+        * Allocate more header room in the rare cases it is not big enough.
+        */
+       if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
+               skb = skb_realloc_headroom(skb, sizeof(struct cpl_tx_pkt_lso));
+               ++st->tx_need_hdrroom;
+               dev_kfree_skb_any(orig_skb);
+               if (!skb)
+                       return NETDEV_TX_OK;
+       }
+
        if (skb_shinfo(skb)->gso_size) {
                int eth_type;
                struct cpl_tx_pkt_lso *hdr;
 
                ++st->tx_tso;
 
-               eth_type = skb->nh.raw - skb->data == ETH_HLEN ?
+               eth_type = skb_network_offset(skb) == ETH_HLEN ?
                        CPL_ETH_II : CPL_ETH_II_VLAN;
 
                hdr = (struct cpl_tx_pkt_lso *)skb_push(skb, sizeof(*hdr));
                hdr->opcode = CPL_TX_PKT_LSO;
                hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
-               hdr->ip_hdr_words = skb->nh.iph->ihl;
-               hdr->tcp_hdr_words = skb->h.th->doff;
+               hdr->ip_hdr_words = ip_hdr(skb)->ihl;
+               hdr->tcp_hdr_words = tcp_hdr(skb)->doff;
                hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
                                                          skb_shinfo(skb)->gso_size));
                hdr->len = htonl(skb->len - sizeof(*hdr));
@@ -1912,27 +1833,9 @@ int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
                        return NETDEV_TX_OK;
                }
 
-               /*
-                * We are using a non-standard hard_header_len and some kernel
-                * components, such as pktgen, do not handle it right.
-                * Complain when this happens but try to fix things up.
-                */
-               if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
-                       pr_debug("%s: headroom %d header_len %d\n", dev->name,
-                                skb_headroom(skb), dev->hard_header_len);
-
-                       if (net_ratelimit())
-                               printk(KERN_ERR "%s: inadequate headroom in "
-                                      "Tx packet\n", dev->name);
-                       skb = skb_realloc_headroom(skb, sizeof(*cpl));
-                       dev_kfree_skb_any(orig_skb);
-                       if (!skb)
-                               return NETDEV_TX_OK;
-               }
-
                if (!(adapter->flags & UDP_CSUM_CAPABLE) &&
                    skb->ip_summed == CHECKSUM_PARTIAL &&
-                   skb->nh.iph->protocol == IPPROTO_UDP) {
+                   ip_hdr(skb)->protocol == IPPROTO_UDP) {
                        if (unlikely(skb_checksum_help(skb))) {
                                pr_debug("%s: unable to do udp checksum\n", dev->name);
                                dev_kfree_skb_any(skb);
@@ -1945,7 +1848,7 @@ int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
                 */
                if ((unlikely(!adapter->sge->espibug_skb[dev->if_port]))) {
                        if (skb->protocol == htons(ETH_P_ARP) &&
-                           skb->nh.arph->ar_op == htons(ARPOP_REQUEST)) {
+                           arp_hdr(skb)->ar_op == htons(ARPOP_REQUEST)) {
                                adapter->sge->espibug_skb[dev->if_port] = skb;
                                /* We want to re-use this skb later. We
                                 * simply bump the reference count and it
@@ -1975,8 +1878,6 @@ int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
                cpl->vlan_valid = 0;
 
 send:
-       st->tx_packets++;
-       dev->trans_start = jiffies;
        ret = t1_sge_tx(skb, adapter, 0, dev);
 
        /* If transmit busy, and we reallocated skb's due to headroom limit,
@@ -2064,8 +1965,7 @@ void t1_sge_stop(struct sge *sge)
                tx_sched_stop(sge);
 
        for (i = 0; i < MAX_NPORTS; i++)
-               if (sge->espibug_skb[i])
-                       kfree_skb(sge->espibug_skb[i]);
+               kfree_skb(sge->espibug_skb[i]);
 }
 
 /*
@@ -2115,10 +2015,14 @@ static void espibug_workaround_t204(unsigned long data)
                                        0x0, 0x7, 0x43, 0x0, 0x0, 0x0
                                };
 
-                               memcpy(skb->data + sizeof(struct cpl_tx_pkt),
-                                       ch_mac_addr, ETH_ALEN);
-                               memcpy(skb->data + skb->len - 10,
-                                       ch_mac_addr, ETH_ALEN);
+                               skb_copy_to_linear_data_offset(skb,
+                                                   sizeof(struct cpl_tx_pkt),
+                                                              ch_mac_addr,
+                                                              ETH_ALEN);
+                               skb_copy_to_linear_data_offset(skb,
+                                                              skb->len - 10,
+                                                              ch_mac_addr,
+                                                              ETH_ALEN);
                                skb->cb[0] = 0xff;
                        }
 
@@ -2145,10 +2049,14 @@ static void espibug_workaround(unsigned long data)
                        if (!skb->cb[0]) {
                                u8 ch_mac_addr[ETH_ALEN] =
                                    {0x0, 0x7, 0x43, 0x0, 0x0, 0x0};
-                               memcpy(skb->data + sizeof(struct cpl_tx_pkt),
-                                      ch_mac_addr, ETH_ALEN);
-                               memcpy(skb->data + skb->len - 10, ch_mac_addr,
-                                      ETH_ALEN);
+                               skb_copy_to_linear_data_offset(skb,
+                                                    sizeof(struct cpl_tx_pkt),
+                                                              ch_mac_addr,
+                                                              ETH_ALEN);
+                               skb_copy_to_linear_data_offset(skb,
+                                                              skb->len - 10,
+                                                              ch_mac_addr,
+                                                              ETH_ALEN);
                                skb->cb[0] = 0xff;
                        }