SGISEEQ: use cached memory access to make driver work on IP28
authorThomas Bogendoerfer <tsbogend@alpha.franken.de>
Sat, 24 Nov 2007 12:29:19 +0000 (13:29 +0100)
committerDavid S. Miller <davem@davemloft.net>
Mon, 28 Jan 2008 23:04:28 +0000 (15:04 -0800)
SGI IP28 machines would need special treatment (enable adding addtional
wait states) when accessing memory uncached. To avoid this pain I changed
the driver to use only cached access to memory.

Signed-off-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
drivers/net/sgiseeq.c

index ff4056310356a8e04b2b33634d1098e474347b2c..3145ca13d2559d3df95341d93a32a9fd097c407e 100644 (file)
@@ -12,7 +12,6 @@
 #include <linux/init.h>
 #include <linux/types.h>
 #include <linux/interrupt.h>
-#include <linux/slab.h>
 #include <linux/string.h>
 #include <linux/delay.h>
 #include <linux/netdevice.h>
@@ -53,14 +52,35 @@ static char *sgiseeqstr = "SGI Seeq8003";
                            sp->tx_old + (SEEQ_TX_BUFFERS - 1) - sp->tx_new : \
                            sp->tx_old - sp->tx_new - 1)
 
+#define VIRT_TO_DMA(sp, v) ((sp)->srings_dma +                                 \
+                                 (dma_addr_t)((unsigned long)(v) -            \
+                                              (unsigned long)((sp)->rx_desc)))
+
+#define DMA_SYNC_DESC_CPU(dev, addr) \
+       do { dma_cache_sync((dev)->dev.parent, (void *)addr, \
+            sizeof(struct sgiseeq_rx_desc), DMA_FROM_DEVICE); } while (0)
+
+#define DMA_SYNC_DESC_DEV(dev, addr) \
+       do { dma_cache_sync((dev)->dev.parent, (void *)addr, \
+            sizeof(struct sgiseeq_rx_desc), DMA_TO_DEVICE); } while (0)
+
+/* Copy frames shorter than rx_copybreak, otherwise pass on up in
+ * a full sized sk_buff.  Value of 100 stolen from tulip.c (!alpha).
+ */
+static int rx_copybreak = 100;
+
+#define PAD_SIZE    (128 - sizeof(struct hpc_dma_desc) - sizeof(void *))
+
 struct sgiseeq_rx_desc {
        volatile struct hpc_dma_desc rdma;
-       volatile signed int buf_vaddr;
+       u8 padding[PAD_SIZE];
+       struct sk_buff *skb;
 };
 
 struct sgiseeq_tx_desc {
        volatile struct hpc_dma_desc tdma;
-       volatile signed int buf_vaddr;
+       u8 padding[PAD_SIZE];
+       struct sk_buff *skb;
 };
 
 /*
@@ -163,35 +183,55 @@ static int seeq_init_ring(struct net_device *dev)
 
        /* Setup tx ring. */
        for(i = 0; i < SEEQ_TX_BUFFERS; i++) {
-               if (!sp->tx_desc[i].tdma.pbuf) {
-                       unsigned long buffer;
-
-                       buffer = (unsigned long) kmalloc(PKT_BUF_SZ, GFP_KERNEL);
-                       if (!buffer)
-                               return -ENOMEM;
-                       sp->tx_desc[i].buf_vaddr = CKSEG1ADDR(buffer);
-                       sp->tx_desc[i].tdma.pbuf = CPHYSADDR(buffer);
-               }
                sp->tx_desc[i].tdma.cntinfo = TCNTINFO_INIT;
+               DMA_SYNC_DESC_DEV(dev, &sp->tx_desc[i]);
        }
 
        /* And now the rx ring. */
        for (i = 0; i < SEEQ_RX_BUFFERS; i++) {
                if (!sp->rx_desc[i].rdma.pbuf) {
-                       unsigned long buffer;
+                       dma_addr_t dma_addr;
+                       struct sk_buff *skb = netdev_alloc_skb(dev, PKT_BUF_SZ);
 
-                       buffer = (unsigned long) kmalloc(PKT_BUF_SZ, GFP_KERNEL);
-                       if (!buffer)
+                       if (skb == NULL)
                                return -ENOMEM;
-                       sp->rx_desc[i].buf_vaddr = CKSEG1ADDR(buffer);
-                       sp->rx_desc[i].rdma.pbuf = CPHYSADDR(buffer);
+                       skb_reserve(skb, 2);
+                       dma_addr = dma_map_single(dev->dev.parent,
+                                                 skb->data - 2,
+                                                 PKT_BUF_SZ, DMA_FROM_DEVICE);
+                       sp->rx_desc[i].skb = skb;
+                       sp->rx_desc[i].rdma.pbuf = dma_addr;
                }
                sp->rx_desc[i].rdma.cntinfo = RCNTINFO_INIT;
+               DMA_SYNC_DESC_DEV(dev, &sp->rx_desc[i]);
        }
        sp->rx_desc[i - 1].rdma.cntinfo |= HPCDMA_EOR;
+       DMA_SYNC_DESC_DEV(dev, &sp->rx_desc[i - 1]);
        return 0;
 }
 
+static void seeq_purge_ring(struct net_device *dev)
+{
+       struct sgiseeq_private *sp = netdev_priv(dev);
+       int i;
+
+       /* clear tx ring. */
+       for (i = 0; i < SEEQ_TX_BUFFERS; i++) {
+               if (sp->tx_desc[i].skb) {
+                       dev_kfree_skb(sp->tx_desc[i].skb);
+                       sp->tx_desc[i].skb = NULL;
+               }
+       }
+
+       /* And now the rx ring. */
+       for (i = 0; i < SEEQ_RX_BUFFERS; i++) {
+               if (sp->rx_desc[i].skb) {
+                       dev_kfree_skb(sp->rx_desc[i].skb);
+                       sp->rx_desc[i].skb = NULL;
+               }
+       }
+}
+
 #ifdef DEBUG
 static struct sgiseeq_private *gpriv;
 static struct net_device *gdev;
@@ -258,8 +298,8 @@ static int init_seeq(struct net_device *dev, struct sgiseeq_private *sp,
                sregs->tstat = TSTAT_INIT_SEEQ;
        }
 
-       hregs->rx_ndptr = CPHYSADDR(sp->rx_desc);
-       hregs->tx_ndptr = CPHYSADDR(sp->tx_desc);
+       hregs->rx_ndptr = VIRT_TO_DMA(sp, sp->rx_desc);
+       hregs->tx_ndptr = VIRT_TO_DMA(sp, sp->tx_desc);
 
        seeq_go(sp, hregs, sregs);
        return 0;
@@ -283,69 +323,90 @@ static inline void rx_maybe_restart(struct sgiseeq_private *sp,
                                    struct sgiseeq_regs *sregs)
 {
        if (!(hregs->rx_ctrl & HPC3_ERXCTRL_ACTIVE)) {
-               hregs->rx_ndptr = CPHYSADDR(sp->rx_desc + sp->rx_new);
+               hregs->rx_ndptr = VIRT_TO_DMA(sp, sp->rx_desc + sp->rx_new);
                seeq_go(sp, hregs, sregs);
        }
 }
 
-#define for_each_rx(rd, sp) for((rd) = &(sp)->rx_desc[(sp)->rx_new]; \
-                               !((rd)->rdma.cntinfo & HPCDMA_OWN); \
-                               (rd) = &(sp)->rx_desc[(sp)->rx_new])
-
 static inline void sgiseeq_rx(struct net_device *dev, struct sgiseeq_private *sp,
                              struct hpc3_ethregs *hregs,
                              struct sgiseeq_regs *sregs)
 {
        struct sgiseeq_rx_desc *rd;
        struct sk_buff *skb = NULL;
+       struct sk_buff *newskb;
        unsigned char pkt_status;
-       unsigned char *pkt_pointer = NULL;
        int len = 0;
        unsigned int orig_end = PREV_RX(sp->rx_new);
 
        /* Service every received packet. */
-       for_each_rx(rd, sp) {
+       rd = &sp->rx_desc[sp->rx_new];
+       DMA_SYNC_DESC_CPU(dev, rd);
+       while (!(rd->rdma.cntinfo & HPCDMA_OWN)) {
                len = PKT_BUF_SZ - (rd->rdma.cntinfo & HPCDMA_BCNT) - 3;
-               pkt_pointer = (unsigned char *)(long)rd->buf_vaddr;
-               pkt_status = pkt_pointer[len + 2];
-
+               dma_unmap_single(dev->dev.parent, rd->rdma.pbuf,
+                                PKT_BUF_SZ, DMA_FROM_DEVICE);
+               pkt_status = rd->skb->data[len];
                if (pkt_status & SEEQ_RSTAT_FIG) {
                        /* Packet is OK. */
-                       skb = dev_alloc_skb(len + 2);
-
-                       if (skb) {
-                               skb_reserve(skb, 2);
-                               skb_put(skb, len);
-
-                               /* Copy out of kseg1 to avoid silly cache flush. */
-                               skb_copy_to_linear_data(skb, pkt_pointer + 2, len);
-                               skb->protocol = eth_type_trans(skb, dev);
-
-                               /* We don't want to receive our own packets */
-                               if (memcmp(eth_hdr(skb)->h_source, dev->dev_addr, ETH_ALEN)) {
+                       /* We don't want to receive our own packets */
+                       if (memcmp(rd->skb->data + 6, dev->dev_addr, ETH_ALEN)) {
+                               if (len > rx_copybreak) {
+                                       skb = rd->skb;
+                                       newskb = netdev_alloc_skb(dev, PKT_BUF_SZ);
+                                       if (!newskb) {
+                                               newskb = skb;
+                                               skb = NULL;
+                                               goto memory_squeeze;
+                                       }
+                                       skb_reserve(newskb, 2);
+                               } else {
+                                       skb = netdev_alloc_skb(dev, len + 2);
+                                       if (skb) {
+                                               skb_reserve(skb, 2);
+                                               skb_copy_to_linear_data(skb, rd->skb->data, len);
+                                       }
+                                       newskb = rd->skb;
+                               }
+memory_squeeze:
+                               if (skb) {
+                                       skb_put(skb, len);
+                                       skb->protocol = eth_type_trans(skb, dev);
                                        netif_rx(skb);
                                        dev->last_rx = jiffies;
                                        dev->stats.rx_packets++;
                                        dev->stats.rx_bytes += len;
                                } else {
-                                       /* Silently drop my own packets */
-                                       dev_kfree_skb_irq(skb);
+                                       printk(KERN_NOTICE "%s: Memory squeeze, deferring packet.\n",
+                                               dev->name);
+                                       dev->stats.rx_dropped++;
                                }
                        } else {
-                               printk (KERN_NOTICE "%s: Memory squeeze, deferring packet.\n",
-                                       dev->name);
-                               dev->stats.rx_dropped++;
+                               /* Silently drop my own packets */
+                               newskb = rd->skb;
                        }
                } else {
                        record_rx_errors(dev, pkt_status);
+                       newskb = rd->skb;
                }
+               rd->skb = newskb;
+               rd->rdma.pbuf = dma_map_single(dev->dev.parent,
+                                              newskb->data - 2,
+                                              PKT_BUF_SZ, DMA_FROM_DEVICE);
 
                /* Return the entry to the ring pool. */
                rd->rdma.cntinfo = RCNTINFO_INIT;
                sp->rx_new = NEXT_RX(sp->rx_new);
+               DMA_SYNC_DESC_DEV(dev, rd);
+               rd = &sp->rx_desc[sp->rx_new];
+               DMA_SYNC_DESC_CPU(dev, rd);
        }
+       DMA_SYNC_DESC_CPU(dev, &sp->rx_desc[orig_end]);
        sp->rx_desc[orig_end].rdma.cntinfo &= ~(HPCDMA_EOR);
+       DMA_SYNC_DESC_DEV(dev, &sp->rx_desc[orig_end]);
+       DMA_SYNC_DESC_CPU(dev, &sp->rx_desc[PREV_RX(sp->rx_new)]);
        sp->rx_desc[PREV_RX(sp->rx_new)].rdma.cntinfo |= HPCDMA_EOR;
+       DMA_SYNC_DESC_DEV(dev, &sp->rx_desc[PREV_RX(sp->rx_new)]);
        rx_maybe_restart(sp, hregs, sregs);
 }
 
@@ -358,20 +419,29 @@ static inline void tx_maybe_reset_collisions(struct sgiseeq_private *sp,
        }
 }
 
-static inline void kick_tx(struct sgiseeq_tx_desc *td,
+static inline void kick_tx(struct net_device *dev,
+                          struct sgiseeq_private *sp,
                           struct hpc3_ethregs *hregs)
 {
+       struct sgiseeq_tx_desc *td;
+       int i = sp->tx_old;
+
        /* If the HPC aint doin nothin, and there are more packets
         * with ETXD cleared and XIU set we must make very certain
         * that we restart the HPC else we risk locking up the
         * adapter.  The following code is only safe iff the HPCDMA
         * is not active!
         */
+       td = &sp->tx_desc[i];
+       DMA_SYNC_DESC_CPU(dev, td);
        while ((td->tdma.cntinfo & (HPCDMA_XIU | HPCDMA_ETXD)) ==
-             (HPCDMA_XIU | HPCDMA_ETXD))
-               td = (struct sgiseeq_tx_desc *)(long) CKSEG1ADDR(td->tdma.pnext);
+             (HPCDMA_XIU | HPCDMA_ETXD)) {
+               i = NEXT_TX(i);
+               td = &sp->tx_desc[i];
+               DMA_SYNC_DESC_CPU(dev, td);
+       }
        if (td->tdma.cntinfo & HPCDMA_XIU) {
-               hregs->tx_ndptr = CPHYSADDR(td);
+               hregs->tx_ndptr = VIRT_TO_DMA(sp, td);
                hregs->tx_ctrl = HPC3_ETXCTRL_ACTIVE;
        }
 }
@@ -400,11 +470,12 @@ static inline void sgiseeq_tx(struct net_device *dev, struct sgiseeq_private *sp
        for (j = sp->tx_old; j != sp->tx_new; j = NEXT_TX(j)) {
                td = &sp->tx_desc[j];
 
+               DMA_SYNC_DESC_CPU(dev, td);
                if (!(td->tdma.cntinfo & (HPCDMA_XIU)))
                        break;
                if (!(td->tdma.cntinfo & (HPCDMA_ETXD))) {
                        if (!(status & HPC3_ETXCTRL_ACTIVE)) {
-                               hregs->tx_ndptr = CPHYSADDR(td);
+                               hregs->tx_ndptr = VIRT_TO_DMA(sp, td);
                                hregs->tx_ctrl = HPC3_ETXCTRL_ACTIVE;
                        }
                        break;
@@ -413,6 +484,11 @@ static inline void sgiseeq_tx(struct net_device *dev, struct sgiseeq_private *sp
                sp->tx_old = NEXT_TX(sp->tx_old);
                td->tdma.cntinfo &= ~(HPCDMA_XIU | HPCDMA_XIE);
                td->tdma.cntinfo |= HPCDMA_EOX;
+               if (td->skb) {
+                       dev_kfree_skb_any(td->skb);
+                       td->skb = NULL;
+               }
+               DMA_SYNC_DESC_DEV(dev, td);
        }
 }
 
@@ -480,6 +556,7 @@ static int sgiseeq_close(struct net_device *dev)
        /* Shutdown the Seeq. */
        reset_hpc3_and_seeq(sp->hregs, sregs);
        free_irq(irq, dev);
+       seeq_purge_ring(dev);
 
        return 0;
 }
@@ -506,16 +583,22 @@ static int sgiseeq_start_xmit(struct sk_buff *skb, struct net_device *dev)
        struct hpc3_ethregs *hregs = sp->hregs;
        unsigned long flags;
        struct sgiseeq_tx_desc *td;
-       int skblen, len, entry;
+       int len, entry;
 
        spin_lock_irqsave(&sp->tx_lock, flags);
 
        /* Setup... */
-       skblen = skb->len;
-       len = (skblen <= ETH_ZLEN) ? ETH_ZLEN : skblen;
+       len = skb->len;
+       if (len < ETH_ZLEN) {
+               if (skb_padto(skb, ETH_ZLEN))
+                       return 0;
+               len = ETH_ZLEN;
+       }
+
        dev->stats.tx_bytes += len;
        entry = sp->tx_new;
        td = &sp->tx_desc[entry];
+       DMA_SYNC_DESC_CPU(dev, td);
 
        /* Create entry.  There are so many races with adding a new
         * descriptor to the chain:
@@ -530,25 +613,27 @@ static int sgiseeq_start_xmit(struct sk_buff *skb, struct net_device *dev)
         *    entry and the HPC got to the end of the chain before we
         *    added this new entry and restarted it.
         */
-       skb_copy_from_linear_data(skb, (char *)(long)td->buf_vaddr, skblen);
-       if (len != skblen)
-               memset((char *)(long)td->buf_vaddr + skb->len, 0, len-skblen);
+       td->skb = skb;
+       td->tdma.pbuf = dma_map_single(dev->dev.parent, skb->data,
+                                      len, DMA_TO_DEVICE);
        td->tdma.cntinfo = (len & HPCDMA_BCNT) |
                           HPCDMA_XIU | HPCDMA_EOXP | HPCDMA_XIE | HPCDMA_EOX;
+       DMA_SYNC_DESC_DEV(dev, td);
        if (sp->tx_old != sp->tx_new) {
                struct sgiseeq_tx_desc *backend;
 
                backend = &sp->tx_desc[PREV_TX(sp->tx_new)];
+               DMA_SYNC_DESC_CPU(dev, backend);
                backend->tdma.cntinfo &= ~HPCDMA_EOX;
+               DMA_SYNC_DESC_DEV(dev, backend);
        }
        sp->tx_new = NEXT_TX(sp->tx_new); /* Advance. */
 
        /* Maybe kick the HPC back into motion. */
        if (!(hregs->tx_ctrl & HPC3_ETXCTRL_ACTIVE))
-               kick_tx(&sp->tx_desc[sp->tx_old], hregs);
+               kick_tx(dev, sp, hregs);
 
        dev->trans_start = jiffies;
-       dev_kfree_skb(skb);
 
        if (!TX_BUFFS_AVAIL(sp))
                netif_stop_queue(dev);
@@ -586,33 +671,41 @@ static void sgiseeq_set_multicast(struct net_device *dev)
                sgiseeq_reset(dev);
 }
 
-static inline void setup_tx_ring(struct sgiseeq_tx_desc *buf, int nbufs)
+static inline void setup_tx_ring(struct net_device *dev,
+                                struct sgiseeq_tx_desc *buf,
+                                int nbufs)
 {
+       struct sgiseeq_private *sp = netdev_priv(dev);
        int i = 0;
 
        while (i < (nbufs - 1)) {
-               buf[i].tdma.pnext = CPHYSADDR(buf + i + 1);
+               buf[i].tdma.pnext = VIRT_TO_DMA(sp, buf + i + 1);
                buf[i].tdma.pbuf = 0;
+               DMA_SYNC_DESC_DEV(dev, &buf[i]);
                i++;
        }
-       buf[i].tdma.pnext = CPHYSADDR(buf);
+       buf[i].tdma.pnext = VIRT_TO_DMA(sp, buf);
+       DMA_SYNC_DESC_DEV(dev, &buf[i]);
 }
 
-static inline void setup_rx_ring(struct sgiseeq_rx_desc *buf, int nbufs)
+static inline void setup_rx_ring(struct net_device *dev,
+                                struct sgiseeq_rx_desc *buf,
+                                int nbufs)
 {
+       struct sgiseeq_private *sp = netdev_priv(dev);
        int i = 0;
 
        while (i < (nbufs - 1)) {
-               buf[i].rdma.pnext = CPHYSADDR(buf + i + 1);
+               buf[i].rdma.pnext = VIRT_TO_DMA(sp, buf + i + 1);
                buf[i].rdma.pbuf = 0;
+               DMA_SYNC_DESC_DEV(dev, &buf[i]);
                i++;
        }
        buf[i].rdma.pbuf = 0;
-       buf[i].rdma.pnext = CPHYSADDR(buf);
+       buf[i].rdma.pnext = VIRT_TO_DMA(sp, buf);
+       DMA_SYNC_DESC_DEV(dev, &buf[i]);
 }
 
-#define ALIGNED(x)  ((((unsigned long)(x)) + 0xf) & ~(0xf))
-
 static int __init sgiseeq_probe(struct platform_device *pdev)
 {
        struct sgiseeq_platform_data *pd = pdev->dev.platform_data;
@@ -621,7 +714,7 @@ static int __init sgiseeq_probe(struct platform_device *pdev)
        unsigned int irq = pd->irq;
        struct sgiseeq_private *sp;
        struct net_device *dev;
-       int err, i;
+       int err;
        DECLARE_MAC_BUF(mac);
 
        dev = alloc_etherdev(sizeof (struct sgiseeq_private));
@@ -635,7 +728,7 @@ static int __init sgiseeq_probe(struct platform_device *pdev)
        sp = netdev_priv(dev);
 
        /* Make private data page aligned */
-       sr = dma_alloc_coherent(&pdev->dev, sizeof(*sp->srings),
+       sr = dma_alloc_noncoherent(&pdev->dev, sizeof(*sp->srings),
                                &sp->srings_dma, GFP_KERNEL);
        if (!sr) {
                printk(KERN_ERR "Sgiseeq: Page alloc failed, aborting.\n");
@@ -647,8 +740,8 @@ static int __init sgiseeq_probe(struct platform_device *pdev)
        sp->tx_desc = sp->srings->txvector;
 
        /* A couple calculations now, saves many cycles later. */
-       setup_rx_ring(sp->rx_desc, SEEQ_RX_BUFFERS);
-       setup_tx_ring(sp->tx_desc, SEEQ_TX_BUFFERS);
+       setup_rx_ring(dev, sp->rx_desc, SEEQ_RX_BUFFERS);
+       setup_tx_ring(dev, sp->tx_desc, SEEQ_TX_BUFFERS);
 
        memcpy(dev->dev_addr, pd->mac, ETH_ALEN);
 
@@ -716,8 +809,8 @@ static int __exit sgiseeq_remove(struct platform_device *pdev)
        struct sgiseeq_private *sp = netdev_priv(dev);
 
        unregister_netdev(dev);
-       dma_free_coherent(&pdev->dev, sizeof(*sp->srings), sp->srings,
-                         sp->srings_dma);
+       dma_free_noncoherent(&pdev->dev, sizeof(*sp->srings), sp->srings,
+                            sp->srings_dma);
        free_netdev(dev);
        platform_set_drvdata(pdev, NULL);