Optimize cxgb3 xmit path (a bit)
[safe/jmp/linux-2.6] / drivers / net / cxgb3 / sge.c
index d38b1bc..979f3fc 100644 (file)
 
 #define SGE_RX_SM_BUF_SIZE 1536
 
-/*
- * If USE_RX_PAGE is defined, the small freelist populated with (partial)
- * pages instead of skbs. Pages are carved up into RX_PAGE_SIZE chunks (must
- * be a multiple of the host page size).
- */
-#define USE_RX_PAGE
-#define RX_PAGE_SIZE 2048
-
-/*
- * skb freelist packets are copied into a new skb (and the freelist one is 
- * reused) if their len is <= 
- */
 #define SGE_RX_COPY_THRES  256
+#define SGE_RX_PULL_LEN    128
 
 /*
- * Minimum number of freelist entries before we start dropping TUNNEL frames.
+ * 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 SGE_RX_DROP_THRES 16
 
 /*
@@ -86,7 +79,7 @@ enum {
 };
 
 struct tx_desc {
-       u64 flit[TX_DESC_FLITS];
+       __be64 flit[TX_DESC_FLITS];
 };
 
 struct rx_desc {
@@ -98,14 +91,18 @@ 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 rx_sw_desc {                /* SW state per Rx descriptor */
        union {
                struct sk_buff *skb;
-               struct sge_fl_page page;
-       } t;
-        DECLARE_PCI_UNMAP_ADDR(dma_addr);
+               struct fl_pg_chunk pg_chunk;
+       };
+       DECLARE_PCI_UNMAP_ADDR(dma_addr);
 };
 
 struct rsp_desc {              /* response queue descriptor */
@@ -116,13 +113,6 @@ 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.
@@ -184,6 +174,7 @@ static inline struct sge_qset *txq_to_qset(const struct sge_txq *q, int qidx)
 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));
 }
@@ -216,32 +207,36 @@ static inline int need_skb_unmap(void)
  *
  *     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) {
@@ -257,10 +252,11 @@ static inline void unmap_skb(struct sk_buff *skb, struct sge_txq *q,
                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 */
        }
 }
 
@@ -288,7 +284,7 @@ static void free_tx_desc(struct adapter *adapter, struct sge_txq *q,
                if (d->skb) {   /* an SGL is present */
                        if (need_unmap)
                                unmap_skb(d->skb, q, cidx, pdev);
-                       if (d->skb->priority == cidx)
+                       if (d->eop)
                                kfree_skb(d->skb);
                }
                ++d;
@@ -351,27 +347,26 @@ static void free_rx_bufs(struct pci_dev *pdev, struct sge_fl *q)
 
                pci_unmap_single(pdev, pci_unmap_addr(d, dma_addr),
                                 q->buf_size, PCI_DMA_FROMDEVICE);
-
-               if (q->buf_size != RX_PAGE_SIZE) {
-                       kfree_skb(d->t.skb);
-                       d->t.skb = NULL;
+               if (q->use_pages) {
+                       put_page(d->pg_chunk.page);
+                       d->pg_chunk.page = NULL;
                } else {
-                       if (d->t.page.frag.page)
-                               put_page(d->t.page.frag.page);
-                       d->t.page.frag.page = NULL;
+                       kfree_skb(d->skb);
+                       d->skb = NULL;
                }
                if (++cidx == q->size)
                        cidx = 0;
        }
 
-       if (q->page.frag.page)
-               put_page(q->page.frag.page);
-       q->page.frag.page = NULL;
+       if (q->pg_chunk.page) {
+               __free_page(q->pg_chunk.page);
+               q->pg_chunk.page = NULL;
+       }
 }
 
 /**
  *     add_one_rx_buf - add a packet buffer to a free-buffer list
- *     @va: va of 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
@@ -381,7 +376,7 @@ static void free_rx_bufs(struct pci_dev *pdev, struct sge_fl *q)
  *     Add a buffer of the given length to the supplied HW and SW Rx
  *     descriptors.
  */
-static inline void add_one_rx_buf(unsigned char *va, unsigned int len,
+static inline void 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)
 {
@@ -397,6 +392,27 @@ static inline void add_one_rx_buf(unsigned char *va, unsigned int len,
        d->gen2 = cpu_to_be32(V_FLD_GEN2(gen));
 }
 
+static int alloc_pg_chunk(struct sge_fl *q, struct rx_sw_desc *sd, gfp_t gfp)
+{
+       if (!q->pg_chunk.page) {
+               q->pg_chunk.page = alloc_page(gfp);
+               if (unlikely(!q->pg_chunk.page))
+                       return -ENOMEM;
+               q->pg_chunk.va = page_address(q->pg_chunk.page);
+               q->pg_chunk.offset = 0;
+       }
+       sd->pg_chunk = q->pg_chunk;
+
+       q->pg_chunk.offset += q->buf_size;
+       if (q->pg_chunk.offset == PAGE_SIZE)
+               q->pg_chunk.page = NULL;
+       else {
+               q->pg_chunk.va += q->buf_size;
+               get_page(q->pg_chunk.page);
+       }
+       return 0;
+}
+
 /**
  *     refill_fl - refill an SGE free-buffer list
  *     @adapter: the adapter
@@ -410,49 +426,29 @@ static inline void add_one_rx_buf(unsigned char *va, unsigned int len,
  */
 static void refill_fl(struct adapter *adap, struct sge_fl *q, int n, gfp_t gfp)
 {
+       void *buf_start;
        struct rx_sw_desc *sd = &q->sdesc[q->pidx];
        struct rx_desc *d = &q->desc[q->pidx];
-       struct sge_fl_page *p = &q->page;
 
        while (n--) {
-               unsigned char *va;
-
-               if (unlikely(q->buf_size != RX_PAGE_SIZE)) {
-                       struct sk_buff *skb = alloc_skb(q->buf_size, gfp);
-
-                       if (!skb) {
-                               q->alloc_failed++;
+               if (q->use_pages) {
+                       if (unlikely(alloc_pg_chunk(q, sd, gfp))) {
+nomem:                         q->alloc_failed++;
                                break;
                        }
-                       va = skb->data;
-                       sd->t.skb = skb;
+                       buf_start = sd->pg_chunk.va;
                } else {
-                       if (!p->frag.page) {
-                               p->frag.page = alloc_pages(gfp, 0);
-                               if (unlikely(!p->frag.page)) {
-                                       q->alloc_failed++;
-                                       break;
-                               } else {
-                                       p->frag.size = RX_PAGE_SIZE;
-                                       p->frag.page_offset = 0;
-                                       p->va = page_address(p->frag.page);
-                               }
-                       }
+                       struct sk_buff *skb = alloc_skb(q->buf_size, gfp);
 
-                       memcpy(&sd->t, p, sizeof(*p));
-                       va = p->va;
+                       if (!skb)
+                               goto nomem;
 
-                       p->frag.page_offset += RX_PAGE_SIZE;
-                       BUG_ON(p->frag.page_offset > PAGE_SIZE);
-                       p->va += RX_PAGE_SIZE;
-                       if (p->frag.page_offset == PAGE_SIZE)
-                               p->frag.page = NULL;
-                       else
-                               get_page(p->frag.page);
+                       sd->skb = skb;
+                       buf_start = skb->data;
                }
 
-               add_one_rx_buf(va, q->buf_size, d, sd, q->gen, adap->pdev);
-
+               add_one_rx_buf(buf_start, q->buf_size, d, sd, q->gen,
+                              adap->pdev);
                d++;
                sd++;
                if (++q->pidx == q->size) {
@@ -463,7 +459,7 @@ static void refill_fl(struct adapter *adap, struct sge_fl *q, int n, gfp_t gfp)
                }
                q->credits++;
        }
-
+       wmb();
        t3_write_reg(adap, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
 }
 
@@ -487,7 +483,7 @@ static void recycle_rx_buf(struct adapter *adap, struct sge_fl *q,
        struct rx_desc *from = &q->desc[idx];
        struct rx_desc *to = &q->desc[q->pidx];
 
-       memcpy(&q->sdesc[q->pidx], &q->sdesc[idx], sizeof(struct rx_sw_desc));
+       q->sdesc[q->pidx] = q->sdesc[idx];
        to->addr_lo = from->addr_lo;    /* already big endian */
        to->addr_hi = from->addr_hi;    /* likewise */
        wmb();
@@ -551,7 +547,7 @@ static void *alloc_ring(struct pci_dev *pdev, size_t nelem, size_t elem_size,
  *     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;
@@ -598,9 +594,6 @@ void t3_free_qset(struct adapter *adapter, struct sge_qset *q)
                                  q->rspq.desc, q->rspq.phys_addr);
        }
 
-       if (q->netdev)
-               q->netdev->atalk_ptr = NULL;
-
        memset(q, 0, sizeof(*q));
 }
 
@@ -650,6 +643,132 @@ static inline unsigned int flits_to_desc(unsigned int n)
 }
 
 /**
+ *     get_packet - return the next ingress packet buffer from a free list
+ *     @adap: the adapter that received the packet
+ *     @fl: the SGE free list holding the packet
+ *     @len: the packet length including any SGE padding
+ *     @drop_thres: # of remaining buffers before we start dropping packets
+ *
+ *     Get the next packet from a free list and complete setup of the
+ *     sk_buff.  If the packet is small we make a copy and recycle the
+ *     original buffer, otherwise we use the original buffer itself.  If a
+ *     positive drop threshold is supplied packets are dropped and their
+ *     buffers recycled if (a) the number of remaining buffers is under the
+ *     threshold and the packet is too big to copy, or (b) the packet should
+ *     be copied but there is no memory for the copy.
+ */
+static struct sk_buff *get_packet(struct adapter *adap, struct sge_fl *fl,
+                                 unsigned int len, unsigned int drop_thres)
+{
+       struct sk_buff *skb = NULL;
+       struct rx_sw_desc *sd = &fl->sdesc[fl->cidx];
+
+       prefetch(sd->skb->data);
+       fl->credits--;
+
+       if (len <= SGE_RX_COPY_THRES) {
+               skb = alloc_skb(len, GFP_ATOMIC);
+               if (likely(skb != NULL)) {
+                       __skb_put(skb, len);
+                       pci_dma_sync_single_for_cpu(adap->pdev,
+                                           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);
+               } else if (!drop_thres)
+                       goto use_orig_buf;
+recycle:
+               recycle_rx_buf(adap, fl, fl->cidx);
+               return skb;
+       }
+
+       if (unlikely(fl->credits < drop_thres))
+               goto recycle;
+
+use_orig_buf:
+       pci_unmap_single(adap->pdev, pci_unmap_addr(sd, dma_addr),
+                        fl->buf_size, PCI_DMA_FROMDEVICE);
+       skb = sd->skb;
+       skb_put(skb, len);
+       __refill_fl(adap, fl);
+       return skb;
+}
+
+/**
+ *     get_packet_pg - return the next ingress packet buffer from a free list
+ *     @adap: the adapter that received the packet
+ *     @fl: the SGE free list holding the packet
+ *     @len: the packet length including any SGE padding
+ *     @drop_thres: # of remaining buffers before we start dropping packets
+ *
+ *     Get the next packet from a free list populated with page chunks.
+ *     If the packet is small we make a copy and recycle the original buffer,
+ *     otherwise we attach the original buffer as a page fragment to a fresh
+ *     sk_buff.  If a positive drop threshold is supplied packets are dropped
+ *     and their buffers recycled if (a) the number of remaining buffers is
+ *     under the threshold and the packet is too big to copy, or (b) there's
+ *     no system memory.
+ *
+ *     Note: this function is similar to @get_packet but deals with Rx buffers
+ *     that are page chunks rather than sk_buffs.
+ */
+static struct sk_buff *get_packet_pg(struct adapter *adap, struct sge_fl *fl,
+                                    unsigned int len, unsigned int drop_thres)
+{
+       struct sk_buff *skb = NULL;
+       struct rx_sw_desc *sd = &fl->sdesc[fl->cidx];
+
+       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);
+                       memcpy(skb->data, sd->pg_chunk.va, len);
+                       pci_dma_sync_single_for_device(adap->pdev,
+                                           pci_unmap_addr(sd, dma_addr), len,
+                                           PCI_DMA_FROMDEVICE);
+               } else if (!drop_thres)
+                       return NULL;
+recycle:
+               fl->credits--;
+               recycle_rx_buf(adap, fl, fl->cidx);
+               return skb;
+       }
+
+       if (unlikely(fl->credits <= drop_thres))
+               goto recycle;
+
+       skb = alloc_skb(SGE_RX_PULL_LEN, GFP_ATOMIC);
+       if (unlikely(!skb)) {
+               if (!drop_thres)
+                       return NULL;
+               goto recycle;
+       }
+
+       pci_unmap_single(adap->pdev, pci_unmap_addr(sd, dma_addr),
+                        fl->buf_size, PCI_DMA_FROMDEVICE);
+       __skb_put(skb, SGE_RX_PULL_LEN);
+       memcpy(skb->data, sd->pg_chunk.va, SGE_RX_PULL_LEN);
+       skb_fill_page_desc(skb, 0, sd->pg_chunk.page,
+                          sd->pg_chunk.offset + SGE_RX_PULL_LEN,
+                          len - SGE_RX_PULL_LEN);
+       skb->len = len;
+       skb->data_len = len - SGE_RX_PULL_LEN;
+       skb->truesize += skb->data_len;
+
+       fl->credits--;
+       /*
+        * We do not refill FLs here, we let the caller do it to overlap a
+        * prefetch.
+        */
+       return skb;
+}
+
+/**
  *     get_imm_packet - return the next ingress packet buffer from a response
  *     @resp: the response descriptor containing the packet data
  *
@@ -661,7 +780,7 @@ static inline struct sk_buff *get_imm_packet(const struct rsp_desc *resp)
 
        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;
 }
@@ -788,23 +907,21 @@ static void write_wr_hdr_sgl(unsigned int ndesc, struct sk_buff *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();
@@ -832,6 +949,7 @@ static void write_wr_hdr_sgl(unsigned int ndesc, struct sk_buff *skb,
 
                        fp += avail;
                        d++;
+                       sd->eop = 0;
                        sd++;
                        if (++pidx == q->size) {
                                pidx = 0;
@@ -850,7 +968,7 @@ static void write_wr_hdr_sgl(unsigned int ndesc, struct sk_buff *skb,
                        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;
@@ -901,7 +1019,7 @@ static void write_tx_pkt_wr(struct adapter *adap, struct sk_buff *skb,
                    CPL_ETH_II : CPL_ETH_II_VLAN;
                tso_info |= V_LSO_ETH_TYPE(eth_type) |
                    V_LSO_IPHDR_WORDS(ip_hdr(skb)->ihl) |
-                   V_LSO_TCPHDR_WORDS(skb->h.th->doff);
+                   V_LSO_TCPHDR_WORDS(tcp_hdr(skb)->doff);
                hdr->lso_info = htonl(tso_info);
                flits = 3;
        } else {
@@ -913,7 +1031,8 @@ static void write_tx_pkt_wr(struct adapter *adap, struct sk_buff *skb,
                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);
 
@@ -934,14 +1053,20 @@ static void write_tx_pkt_wr(struct adapter *adap, struct sk_buff *skb,
 
        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_queue(struct net_device *dev, struct sge_qset *qs,
+                                struct sge_txq *q)
+{
+       netif_stop_queue(dev);
+       set_bit(TXQ_ETH, &qs->txq_stopped);
+       q->stops++;
+}
+
 /**
  *     eth_xmit - add a packet to the Ethernet Tx queue
  *     @skb: the packet
@@ -953,8 +1078,8 @@ int t3_eth_xmit(struct sk_buff *skb, struct net_device *dev)
 {
        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 adapter *adap = pi->adapter;
+       struct sge_qset *qs = pi->qs;
        struct sge_txq *q = &qs->txq[TXQ_ETH];
 
        /*
@@ -973,31 +1098,18 @@ int t3_eth_xmit(struct sk_buff *skb, struct net_device *dev)
        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);
-               }
+               t3_stop_queue(dev, qs, q);
+               dev_err(&adap->pdev->dev,
+                       "%s: Tx ring %u full while queue awake!\n",
+                       dev->name, q->cntxt_id & 7);
                spin_unlock(&q->lock);
                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
-               if (should_restart_tx(q) &&
-                   test_and_clear_bit(TXQ_ETH, &qs->txq_stopped)) {
-                       q->restarts++;
-                       netif_wake_queue(dev);
-               }
-#endif
-       }
+       if (unlikely(credits - ndesc < q->stop_thres))
+               if (USE_GTS || !should_restart_tx(q))
+                       t3_stop_queue(dev, qs, q);
 
        gen = q->gen;
        q->unacked += ndesc;
@@ -1062,8 +1174,8 @@ int t3_eth_xmit(struct sk_buff *skb, struct net_device *dev)
  *
  *     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)
@@ -1071,7 +1183,11 @@ static inline void write_imm(struct tx_desc *d, struct sk_buff *skb,
        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();
@@ -1141,7 +1257,7 @@ static inline void reclaim_completed_tx_imm(struct sge_txq *q)
 
 static inline int immediate(const struct sk_buff *skb)
 {
-       return skb->len <= WR_LEN && !skb->data_len;
+       return skb->len <= WR_LEN;
 }
 
 /**
@@ -1206,12 +1322,12 @@ static void restart_ctrlq(unsigned long data)
        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);
 
@@ -1233,7 +1349,8 @@ static void restart_ctrlq(unsigned long data)
        }
 
        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));
 }
 
@@ -1242,7 +1359,12 @@ static void restart_ctrlq(unsigned long data)
  */
 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;
 }
 
 /**
@@ -1259,13 +1381,14 @@ static void deferred_unmap_destructor(struct sk_buff *skb)
        const dma_addr_t *p;
        const struct skb_shared_info *si;
        const struct deferred_unmap_info *dui;
-       const struct unmap_info *ui = (struct unmap_info *)skb->cb;
 
        dui = (struct deferred_unmap_info *)skb->head;
        p = dui->addr;
 
-       if (ui->len)
-               pci_unmap_single(dui->pdev, *p++, ui->len, PCI_DMA_TODEVICE);
+       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++)
@@ -1324,12 +1447,12 @@ static void write_ofld_wr(struct adapter *adap, struct sk_buff *skb,
 
        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()) {
                setup_deferred_unmapping(skb, adap->pdev, sgp, sgl_flits);
                skb->destructor = deferred_unmap_destructor;
-               ((struct unmap_info *)skb->cb)->len = skb->tail - skb->h.raw;
        }
 
        write_wr_hdr_sgl(ndesc, skb, d, pidx, q, sgl, flits, sgl_flits,
@@ -1345,13 +1468,14 @@ static void write_ofld_wr(struct adapter *adap, struct sk_buff *skb,
  */
 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_transport_offset(skb) / 8;  /* headers */
-       if (skb->tail != skb->h.raw)
+       cnt = skb_shinfo(skb)->nr_frags;
+       if (skb->tail != skb->transport_header)
                cnt++;
        return flits_to_desc(flits + sgl_len(cnt));
 }
@@ -1409,7 +1533,8 @@ static void restart_offloadq(unsigned long data)
        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);
@@ -1449,6 +1574,7 @@ static void restart_offloadq(unsigned long data)
        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));
 }
@@ -1514,8 +1640,7 @@ static inline void offload_enqueue(struct sge_rspq *q, struct sk_buff *skb)
        else {
                struct sge_qset *qs = rspq_to_qset(q);
 
-               if (__netif_rx_schedule_prep(qs->netdev))
-                       __netif_rx_schedule(qs->netdev);
+               napi_schedule(&qs->napi);
                q->rx_head = skb;
        }
        q->rx_tail = skb;
@@ -1551,33 +1676,30 @@ static inline void deliver_partial_bundle(struct t3cdev *tdev,
  *     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) {
+       while (work_done < budget) {
                struct sk_buff *head, *tail, *skbs[RX_BUNDLE_SIZE];
                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);
+                       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--) {
+               for (ngathered = 0; work_done < budget && head; work_done++) {
                        prefetch(head->data);
                        skbs[ngathered] = head;
                        head = head->next;
@@ -1599,10 +1721,8 @@ static int ofld_poll(struct net_device *dev, int *budget)
                }
                deliver_partial_bundle(&adapter->tdev, q, skbs, ngathered);
        }
-       work_done = limit - avail;
-       *budget -= work_done;
-       dev->quota -= work_done;
-       return 1;
+
+       return work_done;
 }
 
 /**
@@ -1620,7 +1740,6 @@ static inline int rx_offload(struct t3cdev *tdev, struct sge_rspq *rq,
                             struct sk_buff *skb, struct sk_buff *rx_gather[],
                             unsigned int gather_idx)
 {
-       rq->offload_pkts++;
        skb_reset_mac_header(skb);
        skb_reset_network_header(skb);
        skb_reset_transport_header(skb);
@@ -1687,10 +1806,10 @@ static void rx_eth(struct adapter *adap, struct sge_rspq *rq,
        struct port_info *pi;
 
        skb_pull(skb, sizeof(*p) + pad);
-       skb->dev->last_rx = jiffies;
        skb->protocol = eth_type_trans(skb, adap->port[p->iff]);
+       skb->dev->last_rx = jiffies;
        pi = netdev_priv(skb->dev);
-       if (pi->rx_csum_offload && p->csum_valid && p->csum == 0xffff &&
+       if (pi->rx_csum_offload && p->csum_valid && p->csum == htons(0xffff) &&
            !p->fragment) {
                rspq_to_qset(rq)->port_stats[SGE_PSTAT_RX_CSUM_GOOD]++;
                skb->ip_summed = CHECKSUM_UNNECESSARY;
@@ -1712,85 +1831,6 @@ static void rx_eth(struct adapter *adap, struct sge_rspq *rq,
                netif_rx(skb);
 }
 
-#define SKB_DATA_SIZE 128
-
-static void skb_data_init(struct sk_buff *skb, struct sge_fl_page *p,
-                         unsigned int len)
-{
-       skb->len = len;
-       if (len <= SKB_DATA_SIZE) {
-               memcpy(skb->data, p->va, len);
-               skb->tail += len;
-               put_page(p->frag.page);
-       } else {
-               memcpy(skb->data, p->va, SKB_DATA_SIZE);
-               skb_shinfo(skb)->frags[0].page = p->frag.page;
-               skb_shinfo(skb)->frags[0].page_offset =
-                   p->frag.page_offset + SKB_DATA_SIZE;
-               skb_shinfo(skb)->frags[0].size = len - SKB_DATA_SIZE;
-               skb_shinfo(skb)->nr_frags = 1;
-               skb->data_len = len - SKB_DATA_SIZE;
-               skb->tail += SKB_DATA_SIZE;
-               skb->truesize += skb->data_len;
-       }
-}
-
-/**
-*      get_packet - return the next ingress packet buffer from a free list
-*      @adap: the adapter that received the packet
-*      @fl: the SGE free list holding the packet
-*      @len: the packet length including any SGE padding
-*      @drop_thres: # of remaining buffers before we start dropping packets
-*
-*      Get the next packet from a free list and complete setup of the
-*      sk_buff.  If the packet is small we make a copy and recycle the
-*      original buffer, otherwise we use the original buffer itself.  If a
-*      positive drop threshold is supplied packets are dropped and their
-*      buffers recycled if (a) the number of remaining buffers is under the
-*      threshold and the packet is too big to copy, or (b) the packet should
-*      be copied but there is no memory for the copy.
-*/
-static struct sk_buff *get_packet(struct adapter *adap, struct sge_fl *fl,
-                                 unsigned int len, unsigned int drop_thres)
-{
-       struct sk_buff *skb = NULL;
-       struct rx_sw_desc *sd = &fl->sdesc[fl->cidx];
-
-       prefetch(sd->t.skb->data);
-
-       if (len <= SGE_RX_COPY_THRES) {
-               skb = alloc_skb(len, GFP_ATOMIC);
-               if (likely(skb != NULL)) {
-                       struct rx_desc *d = &fl->desc[fl->cidx];
-                       dma_addr_t mapping =
-                           (dma_addr_t)((u64) be32_to_cpu(d->addr_hi) << 32 |
-                                        be32_to_cpu(d->addr_lo));
-
-                       __skb_put(skb, len);
-                       pci_dma_sync_single_for_cpu(adap->pdev, mapping, len,
-                                                   PCI_DMA_FROMDEVICE);
-                       memcpy(skb->data, sd->t.skb->data, len);
-                       pci_dma_sync_single_for_device(adap->pdev, mapping, len,
-                                                      PCI_DMA_FROMDEVICE);
-               } else if (!drop_thres)
-                       goto use_orig_buf;
-recycle:
-               recycle_rx_buf(adap, fl, fl->cidx);
-               return skb;
-       }
-
-       if (unlikely(fl->credits < drop_thres))
-               goto recycle;
-
-use_orig_buf:
-       pci_unmap_single(adap->pdev, pci_unmap_addr(sd, dma_addr),
-                        fl->buf_size, PCI_DMA_FROMDEVICE);
-       skb = sd->t.skb;
-       skb_put(skb, len);
-       __refill_fl(adap, fl);
-       return skb;
-}
-
 /**
  *     handle_rsp_cntrl_info - handles control information in a response
  *     @qs: the queue set corresponding to the response
@@ -1916,7 +1956,7 @@ static int process_responses(struct adapter *adap, struct sge_qset *qs,
                int eth, ethpad = 2;
                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;
 
@@ -1932,7 +1972,7 @@ static int process_responses(struct adapter *adap, struct sge_qset *qs,
                } 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 */
@@ -1942,53 +1982,29 @@ static int process_responses(struct adapter *adap, struct sge_qset *qs,
                        q->imm_data++;
                        ethpad = 0;
                } else if ((len = ntohl(r->len_cq)) != 0) {
-                       struct sge_fl *fl =
-                           (len & F_RSPD_FLQ) ? &qs->fl[1] : &qs->fl[0];
-
-                       if (fl->buf_size == RX_PAGE_SIZE) {
-                               struct rx_sw_desc *sd = &fl->sdesc[fl->cidx];
-                               struct sge_fl_page *p = &sd->t.page;
+                       struct sge_fl *fl;
 
-                               prefetch(p->va);
-                               prefetch(p->va + L1_CACHE_BYTES);
+                       fl = (len & F_RSPD_FLQ) ? &qs->fl[1] : &qs->fl[0];
+                       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);
 
-                               pci_unmap_single(adap->pdev,
-                                                pci_unmap_addr(sd, dma_addr),
-                                                fl->buf_size,
-                                                PCI_DMA_FROMDEVICE);
-
-                               if (eth) {
-                                       if (unlikely(fl->credits <
-                                                    SGE_RX_DROP_THRES))
-                                               goto eth_recycle;
-
-                                       skb = alloc_skb(SKB_DATA_SIZE,
-                                                       GFP_ATOMIC);
-                                       if (unlikely(!skb)) {
-eth_recycle:
-                                               q->rx_drops++;
-                                               recycle_rx_buf(adap, fl,
-                                                              fl->cidx);
-                                               goto eth_done;
-                                       }
-                               } else {
-                                       skb = alloc_skb(SKB_DATA_SIZE,
-                                                       GFP_ATOMIC);
-                                       if (unlikely(!skb))
-                                               goto no_mem;
-                               }
-
-                               skb_data_init(skb, p, G_RSPD_LEN(len));
-eth_done:
-                               fl->credits--;
-                               q->eth_pkts++;
-                       } else {
-                               fl->credits--;
+                               skb = get_packet_pg(adap, fl, G_RSPD_LEN(len),
+                                                eth ? SGE_RX_DROP_THRES : 0);
+                       } 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 (unlikely(r->rss_hdr.opcode == CPL_TRACE_PKT))
+                               __skb_pull(skb, 2);
 
                        if (++fl->cidx == fl->size)
                                fl->cidx = 0;
@@ -2013,20 +2029,16 @@ eth_done:
                        q->credits = 0;
                }
 
-               if (skb) {
-                       /* Preserve the RSS info in csum & priority */
-                       skb->csum = rss_hi;
-                       skb->priority = rss_lo;
-
+               if (likely(skb != NULL)) {
                        if (eth)
                                rx_eth(adap, q, skb, ethpad);
                        else {
-                               if (unlikely(r->rss_hdr.opcode ==
-                                            CPL_TRACE_PKT))
-                                       __skb_pull(skb, ethpad);
-
-                               ngathered = rx_offload(&adap->tdev, q,
-                                                      skb, offload_skbs,
+                               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);
                        }
                }
@@ -2054,49 +2066,47 @@ static inline int is_pure_response(const struct rsp_desc *r)
 
 /**
  *     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;
+       struct sge_qset *qs = container_of(napi, struct sge_qset, napi);
+       struct adapter *adap = qs->adap;
+       int work_done = process_responses(adap, qs, budget);
 
-       if (work_done >= effective_budget)
-               return 1;
-
-       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);
 }
 
 /**
@@ -2179,8 +2189,7 @@ static inline int handle_responses(struct adapter *adap, struct sge_rspq *q)
                             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;
 }
 
@@ -2191,7 +2200,7 @@ static inline int handle_responses(struct adapter *adap, struct sge_rspq *q)
 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);
@@ -2207,16 +2216,14 @@ irqreturn_t t3_sge_intr_msix(int irq, void *cookie)
  * 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;
@@ -2259,11 +2266,13 @@ static irqreturn_t t3_intr_msi(int irq, void *cookie)
        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;
@@ -2276,7 +2285,7 @@ static int rspq_check_napi(struct net_device *dev, struct sge_rspq *q)
  * 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;
@@ -2284,10 +2293,9 @@ irqreturn_t t3_intr_msi_napi(int irq, void *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++;
 
@@ -2390,9 +2398,9 @@ static irqreturn_t t3b_intr(int irq, void *cookie)
 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);
@@ -2405,18 +2413,11 @@ static irqreturn_t t3b_intr_napi(int irq, void *cookie)
        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;
@@ -2431,7 +2432,7 @@ static irqreturn_t t3b_intr_napi(int irq, void *cookie)
  *     (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;
@@ -2442,6 +2443,15 @@ intr_handler_t t3_intr_handler(struct adapter *adap, int polling)
        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
@@ -2452,6 +2462,13 @@ void t3_sge_err_intr_handler(struct adapter *adapter)
 {
        unsigned int v, status = t3_read_reg(adapter, A_SG_INT_CAUSE);
 
+       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");
 
@@ -2463,8 +2480,12 @@ void t3_sge_err_intr_handler(struct adapter *adapter)
                         "(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);
 }
 
@@ -2495,7 +2516,7 @@ static void sge_timer_cb(unsigned long data)
 {
        spinlock_t *lock;
        struct sge_qset *qs = (struct sge_qset *)data;
-       struct adapter *adap = qs->netdev->priv;
+       struct adapter *adap = qs->adap;
 
        if (spin_trylock(&qs->txq[TXQ_ETH].lock)) {
                reclaim_completed_tx(adap, &qs->txq[TXQ_ETH]);
@@ -2506,9 +2527,9 @@ static void sge_timer_cb(unsigned long data)
                spin_unlock(&qs->txq[TXQ_OFLD].lock);
        }
        lock = (adap->flags & USING_MSIX) ? &qs->rspq.lock :
-           &adap->sge.qs[0].rspq.lock;
+                                           &adap->sge.qs[0].rspq.lock;
        if (spin_trylock_irq(lock)) {
-               if (!napi_is_scheduled(qs->netdev)) {
+               if (!napi_is_scheduled(&qs->napi)) {
                        u32 status = t3_read_reg(adap, A_SG_RSPQ_FL_STATUS);
 
                        if (qs->fl[0].credits < qs->fl[0].size)
@@ -2542,12 +2563,9 @@ static void sge_timer_cb(unsigned long data)
  */
 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;
 }
 
 /**
@@ -2567,7 +2585,7 @@ void t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p)
  */
 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)
 {
        int i, ret = -ENOMEM;
        struct sge_qset *q = &adapter->sge.qs[id];
@@ -2633,25 +2651,15 @@ int t3_sge_alloc_qset(struct adapter *adapter, unsigned int id, int nports,
        q->txq[TXQ_ETH].stop_thres = nports *
            flits_to_desc(sgl_len(MAX_SKB_FRAGS + 1) + 3);
 
-       if (!is_offload(adapter)) {
-#ifdef USE_RX_PAGE
-               q->fl[0].buf_size = RX_PAGE_SIZE;
-#else
-               q->fl[0].buf_size = SGE_RX_SM_BUF_SIZE + 2 +
-                   sizeof(struct cpl_rx_pkt);
-#endif
-               q->fl[1].buf_size = MAX_FRAME_SIZE + 2 +
-                   sizeof(struct cpl_rx_pkt);
-       } else {
-#ifdef USE_RX_PAGE
-               q->fl[0].buf_size = RX_PAGE_SIZE;
+#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);
+       q->fl[0].buf_size = SGE_RX_SM_BUF_SIZE + sizeof(struct cpl_rx_data);
 #endif
-               q->fl[1].buf_size = (16 * 1024) -
-                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-       }
+       q->fl[0].use_pages = FL0_PG_CHUNK_SIZE > 0;
+       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);
 
        spin_lock(&adapter->sge.reg_lock);
 
@@ -2698,16 +2706,10 @@ int t3_sge_alloc_qset(struct adapter *adapter, unsigned int id, int nports,
        }
 
        spin_unlock(&adapter->sge.reg_lock);
-       q->netdev = netdev;
-       t3_update_qset_coalesce(q, p);
 
-       /*
-        * 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;
+       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);
@@ -2795,7 +2797,7 @@ void t3_sge_init(struct adapter *adap, struct sge_params *p)
        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
@@ -2804,7 +2806,6 @@ void t3_sge_init(struct adapter *adap, struct sge_params *p)
        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) |
@@ -2812,7 +2813,8 @@ void t3_sge_init(struct adapter *adap, struct sge_params *p)
        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);
@@ -2829,7 +2831,7 @@ void t3_sge_init(struct adapter *adap, struct sge_params *p)
  *     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;