target/linux/rb532-2.6/patches/240-via_rhine_performance.patch

   1 diff -ur linux.old/drivers/net/via-rhine.c linux.dev/drivers/net/via-rhine.c
   2 --- linux.old/drivers/net/via-rhine.c   2006-12-07 05:53:39.000000000 +0100
   3 +++ linux.dev/drivers/net/via-rhine.c   2006-12-07 07:06:52.000000000 +0100
   4 @@ -131,6 +131,10 @@
   5         - Fix Tx engine race for good
   6         - Craig Brind: Zero padded aligned buffers for short packets.
   7
   8 +       OpenWrt Version (Felix Fietkau <nbd@openwrt.org>)
   9 +       - Performance improvements
  10 +       - NAPI polling
  11 +
  12  */
  13
  14  #define DRV_NAME       "via-rhine"
  15 @@ -142,7 +146,6 @@
  16     These may be modified when a driver module is loaded. */
  17
  18  static int debug = 1;  /* 1 normal messages, 0 quiet .. 7 verbose. */
  19 -static int max_interrupt_work = 20;
  20
  21  /* Set the copy breakpoint for the copy-only-tiny-frames scheme.
  22     Setting to > 1518 effectively disables this feature. */
  23 @@ -165,9 +168,9 @@
  24     Making the Tx ring too large decreases the effectiveness of channel
  25     bonding and packet priority.
  26     There are no ill effects from too-large receive rings. */
  27 -#define TX_RING_SIZE   16
  28 -#define TX_QUEUE_LEN   10      /* Limit ring entries actually used. */
  29 -#define RX_RING_SIZE   16
  30 +#define TX_RING_SIZE   64
  31 +#define TX_QUEUE_LEN   60      /* Limit ring entries actually used. */
  32 +#define RX_RING_SIZE   64
  33
  34
  35  /* Operational parameters that usually are not changed. */
  36 @@ -201,6 +204,7 @@
  37  #include <asm/io.h>
  38  #include <asm/irq.h>
  39  #include <asm/uaccess.h>
  40 +#include <asm/unaligned.h>
  41
  42  /* These identify the driver base version and may not be removed. */
  43  static char version[] __devinitdata =
  44 @@ -217,10 +221,8 @@
  45  MODULE_DESCRIPTION("VIA Rhine PCI Fast Ethernet driver");
  46  MODULE_LICENSE("GPL");
  47
  48 -module_param(max_interrupt_work, int, 0);
  49  module_param(debug, int, 0);
  50  module_param(rx_copybreak, int, 0);
  51 -MODULE_PARM_DESC(max_interrupt_work, "VIA Rhine maximum events handled per interrupt");
  52  MODULE_PARM_DESC(debug, "VIA Rhine debug level (0-7)");
  53  MODULE_PARM_DESC(rx_copybreak, "VIA Rhine copy breakpoint for copy-only-tiny-frames");
  54
  55 @@ -461,6 +463,8 @@
  56         struct tx_desc *tx_ring;
  57         dma_addr_t rx_ring_dma;
  58         dma_addr_t tx_ring_dma;
  59 +       u32 istat;
  60 +       u32 imask;
  61
  62         /* The addresses of receive-in-place skbuffs. */
  63         struct sk_buff *rx_skbuff[RX_RING_SIZE];
  64 @@ -500,9 +504,10 @@
  65  static void rhine_tx_timeout(struct net_device *dev);
  66  static int  rhine_start_tx(struct sk_buff *skb, struct net_device *dev);
  67  static irqreturn_t rhine_interrupt(int irq, void *dev_instance, struct pt_regs *regs);
  68 -static void rhine_tx(struct net_device *dev);
  69 -static void rhine_rx(struct net_device *dev);
  70 -static void rhine_error(struct net_device *dev, int intr_status);
  71 +static int rhine_poll(struct net_device *dev, int *budget);
  72 +static int rhine_tx(struct net_device *dev);
  73 +static int rhine_rx(struct net_device *dev, int max_work);
  74 +static void rhine_error(struct net_device *dev);
  75  static void rhine_set_rx_mode(struct net_device *dev);
  76  static struct net_device_stats *rhine_get_stats(struct net_device *dev);
  77  static int netdev_ioctl(struct net_device *dev, struct ifreq *rq, int cmd);
  78 @@ -597,6 +602,7 @@
  79         struct rhine_private *rp = netdev_priv(dev);
  80         void __iomem *ioaddr = rp->base;
  81
  82 +       pci_enable_device(rp->pdev);
  83         iowrite8(Cmd1Reset, ioaddr + ChipCmd1);
  84         IOSYNC;
  85
  86 @@ -618,6 +624,28 @@
  87                         "failed" : "succeeded");
  88  }
  89
  90 +static inline void rhine_intr_enable(struct net_device *dev)
  91 +{
  92 +       struct rhine_private *rp = netdev_priv(dev);
  93 +       void __iomem *ioaddr = rp->base;
  94 +
  95 +       iowrite16(rp->imask = (IntrRxDone | IntrRxErr | IntrRxEmpty| IntrRxOverflow |
  96 +              IntrRxDropped | IntrRxNoBuf | IntrTxAborted |
  97 +              IntrTxDone | IntrTxError | IntrTxUnderrun |
  98 +              IntrPCIErr | IntrStatsMax | IntrLinkChange),
  99 +              ioaddr + IntrEnable);
 100 +}
 101 +
 102 +static inline void rhine_intr_disable(struct net_device *dev)
 103 +{
 104 +       struct rhine_private *rp = netdev_priv(dev);
 105 +       void __iomem *ioaddr = rp->base;
 106 +
 107 +       iowrite16(rp->imask = (IntrRxOverflow | IntrRxNoBuf | IntrTxAborted |
 108 +              IntrTxError | IntrTxUnderrun | IntrPCIErr | IntrStatsMax | IntrLinkChange),
 109 +              ioaddr + IntrEnable);
 110 +}
 111 +
 112  #ifdef USE_MMIO
 113  static void enable_mmio(long pioaddr, u32 quirks)
 114  {
 115 @@ -660,14 +688,26 @@
 116
 117  }
 118
 119 -#ifdef CONFIG_NET_POLL_CONTROLLER
 120 -static void rhine_poll(struct net_device *dev)
 121 +static int rhine_poll(struct net_device *dev, int *budget)
 122  {
 123 -       disable_irq(dev->irq);
 124 -       rhine_interrupt(dev->irq, (void *)dev, NULL);
 125 -       enable_irq(dev->irq);
 126 +       unsigned int work_done, work_to_do = min(*budget, dev->quota);
 127 +       struct rhine_private *rp = netdev_priv(dev);
 128 +
 129 +       work_done = rhine_rx(dev, (*budget < dev->quota ? *budget : dev->quota));
 130 +
 131 +       if (rp->istat & (IntrTxErrSummary | IntrTxDone))
 132 +               rhine_tx(dev);
 133 +
 134 +       *budget -= work_done;
 135 +       dev->quota -= work_done;
 136 +
 137 +       if (work_done < work_to_do) {
 138 +               netif_rx_complete(dev);
 139 +               rhine_intr_enable(dev);
 140 +       }
 141 +
 142 +       return (work_done >= work_to_do);
 143  }
 144 -#endif
 145
 146  static void rhine_hw_init(struct net_device *dev, long pioaddr)
 147  {
 148 @@ -846,11 +886,10 @@
 149         dev->ethtool_ops = &netdev_ethtool_ops;
 150         dev->tx_timeout = rhine_tx_timeout;
 151         dev->watchdog_timeo = TX_TIMEOUT;
 152 -#ifdef CONFIG_NET_POLL_CONTROLLER
 153 -       dev->poll_controller = rhine_poll;
 154 -#endif
 155 -       if (rp->quirks & rqRhineI)
 156 -               dev->features |= NETIF_F_SG|NETIF_F_HW_CSUM;
 157 +       dev->poll = rhine_poll;
 158 +       dev->weight = 64;
 159 +
 160 +       dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM;
 161
 162         /* dev->name not defined before register_netdev()! */
 163         rc = register_netdev(dev);
 164 @@ -894,6 +933,10 @@
 165                 }
 166         }
 167         rp->mii_if.phy_id = phy_id;
 168 +
 169 +       // shut down until somebody really needs it
 170 +       iowrite8(0x80, ioaddr + 0xa1);
 171 +       pci_set_power_state(rp->pdev, 3);
 172
 173         return 0;
 174
 175 @@ -985,7 +1028,7 @@
 176
 177         /* Fill in the Rx buffers.  Handle allocation failure gracefully. */
 178         for (i = 0; i < RX_RING_SIZE; i++) {
 179 -               struct sk_buff *skb = dev_alloc_skb(rp->rx_buf_sz);
 180 +               struct sk_buff *skb = dev_alloc_skb(rp->rx_buf_sz + 4);
 181                 rp->rx_skbuff[i] = skb;
 182                 if (skb == NULL)
 183                         break;
 184 @@ -1120,11 +1163,7 @@
 185         rhine_set_rx_mode(dev);
 186
 187         /* Enable interrupts by setting the interrupt mask. */
 188 -       iowrite16(IntrRxDone | IntrRxErr | IntrRxEmpty| IntrRxOverflow |
 189 -              IntrRxDropped | IntrRxNoBuf | IntrTxAborted |
 190 -              IntrTxDone | IntrTxError | IntrTxUnderrun |
 191 -              IntrPCIErr | IntrStatsMax | IntrLinkChange,
 192 -              ioaddr + IntrEnable);
 193 +       rhine_intr_enable(dev);
 194
 195         iowrite16(CmdStart | CmdTxOn | CmdRxOn | (Cmd1NoTxPoll << 8),
 196                ioaddr + ChipCmd);
 197 @@ -1235,6 +1274,7 @@
 198                        mdio_read(dev, rp->mii_if.phy_id, MII_BMSR));
 199
 200         netif_start_queue(dev);
 201 +       netif_poll_enable(dev);
 202
 203         return 0;
 204  }
 205 @@ -1263,8 +1303,8 @@
 206         /* Reinitialize the hardware. */
 207         rhine_chip_reset(dev);
 208         init_registers(dev);
 209 -
 210         spin_unlock(&rp->lock);
 211 +
 212         enable_irq(rp->pdev->irq);
 213
 214         dev->trans_start = jiffies;
 215 @@ -1358,77 +1398,66 @@
 216         struct net_device *dev = dev_instance;
 217         struct rhine_private *rp = netdev_priv(dev);
 218         void __iomem *ioaddr = rp->base;
 219 -       u32 intr_status;
 220 -       int boguscnt = max_interrupt_work;
 221         int handled = 0;
 222
 223 -       while ((intr_status = get_intr_status(dev))) {
 224 +       if ((rp->istat = (get_intr_status(dev) & rp->imask))) {
 225                 handled = 1;
 226
 227                 /* Acknowledge all of the current interrupt sources ASAP. */
 228 -               if (intr_status & IntrTxDescRace)
 229 +               if (rp->istat & IntrTxDescRace)
 230                         iowrite8(0x08, ioaddr + IntrStatus2);
 231 -               iowrite16(intr_status & 0xffff, ioaddr + IntrStatus);
 232 +               iowrite16(rp->istat & 0xffff, ioaddr + IntrStatus);
 233                 IOSYNC;
 234
 235 -               if (debug > 4)
 236 -                       printk(KERN_DEBUG "%s: Interrupt, status %8.8x.\n",
 237 -                              dev->name, intr_status);
 238 +               if (likely(rp->istat & ((IntrRxDone | IntrRxErr | IntrRxDropped |
 239 +                               IntrRxWakeUp | IntrRxEmpty | IntrRxNoBuf |
 240 +                               IntrTxErrSummary | IntrTxDone)))) {
 241 +
 242 +                       rhine_intr_disable(dev);
 243
 244 -               if (intr_status & (IntrRxDone | IntrRxErr | IntrRxDropped |
 245 -                   IntrRxWakeUp | IntrRxEmpty | IntrRxNoBuf))
 246 -                       rhine_rx(dev);
 247 -
 248 -               if (intr_status & (IntrTxErrSummary | IntrTxDone)) {
 249 -                       if (intr_status & IntrTxErrSummary) {
 250 -                               /* Avoid scavenging before Tx engine turned off */
 251 -                               RHINE_WAIT_FOR(!(ioread8(ioaddr+ChipCmd) & CmdTxOn));
 252 -                               if (debug > 2 &&
 253 -                                   ioread8(ioaddr+ChipCmd) & CmdTxOn)
 254 -                                       printk(KERN_WARNING "%s: "
 255 -                                              "rhine_interrupt() Tx engine"
 256 -                                              "still on.\n", dev->name);
 257 -                       }
 258 -                       rhine_tx(dev);
 259 +                       if (likely(netif_rx_schedule_prep(dev)))
 260 +                               __netif_rx_schedule(dev);
 261                 }
 262
 263                 /* Abnormal error summary/uncommon events handlers. */
 264 -               if (intr_status & (IntrPCIErr | IntrLinkChange |
 265 +               if (unlikely(rp->istat & (IntrPCIErr | IntrLinkChange |
 266                                    IntrStatsMax | IntrTxError | IntrTxAborted |
 267 -                                  IntrTxUnderrun | IntrTxDescRace))
 268 -                       rhine_error(dev, intr_status);
 269 -
 270 -               if (--boguscnt < 0) {
 271 -                       printk(KERN_WARNING "%s: Too much work at interrupt, "
 272 -                              "status=%#8.8x.\n",
 273 -                              dev->name, intr_status);
 274 -                       break;
 275 -               }
 276 +                                  IntrTxUnderrun | IntrTxDescRace)))
 277 +                       rhine_error(dev);
 278         }
 279
 280 -       if (debug > 3)
 281 -               printk(KERN_DEBUG "%s: exiting interrupt, status=%8.8x.\n",
 282 -                      dev->name, ioread16(ioaddr + IntrStatus));
 283         return IRQ_RETVAL(handled);
 284  }
 285
 286  /* This routine is logically part of the interrupt handler, but isolated
 287     for clarity. */
 288 -static void rhine_tx(struct net_device *dev)
 289 +static int rhine_tx(struct net_device *dev)
 290  {
 291         struct rhine_private *rp = netdev_priv(dev);
 292         int txstatus = 0, entry = rp->dirty_tx % TX_RING_SIZE;
 293 +       void __iomem *ioaddr = rp->base;
 294 +       int done = 0;
 295 +
 296 +       /* Avoid scavenging before Tx engine turned off */
 297 +       RHINE_WAIT_FOR(!(ioread8(ioaddr+ChipCmd) & CmdTxOn));
 298 +       if (debug > 2 &&
 299 +               ioread8(ioaddr+ChipCmd) & CmdTxOn)
 300 +               printk(KERN_WARNING "%s: "
 301 +                          "rhine_interrupt() Tx engine"
 302 +                          "still on.\n", dev->name);
 303
 304 -       spin_lock(&rp->lock);
 305
 306         /* find and cleanup dirty tx descriptors */
 307         while (rp->dirty_tx != rp->cur_tx) {
 308 +               spin_lock(&rp->lock);
 309                 txstatus = le32_to_cpu(rp->tx_ring[entry].tx_status);
 310                 if (debug > 6)
 311                         printk(KERN_DEBUG "Tx scavenge %d status %8.8x.\n",
 312                                entry, txstatus);
 313 -               if (txstatus & DescOwn)
 314 +               if (txstatus & DescOwn) {
 315 +                       spin_unlock(&rp->lock);
 316                         break;
 317 +               }
 318                 if (txstatus & 0x8000) {
 319                         if (debug > 1)
 320                                 printk(KERN_DEBUG "%s: Transmit error, "
 321 @@ -1443,6 +1472,7 @@
 322                             (txstatus & 0x0800) || (txstatus & 0x1000)) {
 323                                 rp->stats.tx_fifo_errors++;
 324                                 rp->tx_ring[entry].tx_status = cpu_to_le32(DescOwn);
 325 +                               spin_unlock(&rp->lock);
 326                                 break; /* Keep the skb - we try again */
 327                         }
 328                         /* Transmitter restarted in 'abnormal' handler. */
 329 @@ -1457,6 +1487,7 @@
 330                                        txstatus & 0xF);
 331                         rp->stats.tx_bytes += rp->tx_skbuff[entry]->len;
 332                         rp->stats.tx_packets++;
 333 +                       done++;
 334                 }
 335                 /* Free the original skb. */
 336                 if (rp->tx_skbuff_dma[entry]) {
 337 @@ -1465,23 +1496,25 @@
 338                                          rp->tx_skbuff[entry]->len,
 339                                          PCI_DMA_TODEVICE);
 340                 }
 341 -               dev_kfree_skb_irq(rp->tx_skbuff[entry]);
 342 +               dev_kfree_skb_any(rp->tx_skbuff[entry]);
 343                 rp->tx_skbuff[entry] = NULL;
 344                 entry = (++rp->dirty_tx) % TX_RING_SIZE;
 345 +               spin_unlock(&rp->lock);
 346         }
 347 +
 348         if ((rp->cur_tx - rp->dirty_tx) < TX_QUEUE_LEN - 4)
 349                 netif_wake_queue(dev);
 350
 351 -       spin_unlock(&rp->lock);
 352 +       return done;
 353  }
 354
 355  /* This routine is logically part of the interrupt handler, but isolated
 356     for clarity and better register allocation. */
 357 -static void rhine_rx(struct net_device *dev)
 358 +static int rhine_rx(struct net_device *dev, int max_work)
 359  {
 360         struct rhine_private *rp = netdev_priv(dev);
 361         int entry = rp->cur_rx % RX_RING_SIZE;
 362 -       int boguscnt = rp->dirty_rx + RX_RING_SIZE - rp->cur_rx;
 363 +       int done = 0;
 364
 365         if (debug > 4) {
 366                 printk(KERN_DEBUG "%s: rhine_rx(), entry %d status %8.8x.\n",
 367 @@ -1498,7 +1531,7 @@
 368                 if (debug > 4)
 369                         printk(KERN_DEBUG "rhine_rx() status is %8.8x.\n",
 370                                desc_status);
 371 -               if (--boguscnt < 0)
 372 +               if (--max_work < 0)
 373                         break;
 374                 if ((desc_status & (RxWholePkt | RxErr)) != RxWholePkt) {
 375                         if ((desc_status & RxWholePkt) != RxWholePkt) {
 376 @@ -1523,9 +1556,7 @@
 377                                 if (desc_status & 0x0004) rp->stats.rx_frame_errors++;
 378                                 if (desc_status & 0x0002) {
 379                                         /* this can also be updated outside the interrupt handler */
 380 -                                       spin_lock(&rp->lock);
 381                                         rp->stats.rx_crc_errors++;
 382 -                                       spin_unlock(&rp->lock);
 383                                 }
 384                         }
 385                 } else {
 386 @@ -1553,6 +1584,7 @@
 387                                                                rp->rx_buf_sz,
 388                                                                PCI_DMA_FROMDEVICE);
 389                         } else {
 390 +                               int i;
 391                                 skb = rp->rx_skbuff[entry];
 392                                 if (skb == NULL) {
 393                                         printk(KERN_ERR "%s: Inconsistent Rx "
 394 @@ -1561,6 +1593,14 @@
 395                                         break;
 396                                 }
 397                                 rp->rx_skbuff[entry] = NULL;
 398 +
 399 +                               /* align the data to the ip header - should be faster than using rx_copybreak */
 400 +                               for (i = pkt_len - (pkt_len % 4); i >= 0; i -= 4) {
 401 +                                       put_unaligned(*((u32 *) (skb->data + i)), (u32 *) (skb->data + i + 2));
 402 +                               }
 403 +                               skb->data += 2;
 404 +                               skb->tail += 2;
 405 +
 406                                 skb_put(skb, pkt_len);
 407                                 pci_unmap_single(rp->pdev,
 408                                                  rp->rx_skbuff_dma[entry],
 409 @@ -1568,10 +1608,11 @@
 410                                                  PCI_DMA_FROMDEVICE);
 411                         }
 412                         skb->protocol = eth_type_trans(skb, dev);
 413 -                       netif_rx(skb);
 414 +                       netif_receive_skb(skb);
 415                         dev->last_rx = jiffies;
 416                         rp->stats.rx_bytes += pkt_len;
 417                         rp->stats.rx_packets++;
 418 +                       done++;
 419                 }
 420                 entry = (++rp->cur_rx) % RX_RING_SIZE;
 421                 rp->rx_head_desc = &rp->rx_ring[entry];
 422 @@ -1582,7 +1623,7 @@
 423                 struct sk_buff *skb;
 424                 entry = rp->dirty_rx % RX_RING_SIZE;
 425                 if (rp->rx_skbuff[entry] == NULL) {
 426 -                       skb = dev_alloc_skb(rp->rx_buf_sz);
 427 +                       skb = dev_alloc_skb(rp->rx_buf_sz + 4);
 428                         rp->rx_skbuff[entry] = skb;
 429                         if (skb == NULL)
 430                                 break;  /* Better luck next round. */
 431 @@ -1595,6 +1636,8 @@
 432                 }
 433                 rp->rx_ring[entry].rx_status = cpu_to_le32(DescOwn);
 434         }
 435 +
 436 +       return done;
 437  }
 438
 439  /*
 440 @@ -1644,11 +1687,11 @@
 441
 442  }
 443
 444 -static void rhine_error(struct net_device *dev, int intr_status)
 445 +static void rhine_error(struct net_device *dev)
 446  {
 447         struct rhine_private *rp = netdev_priv(dev);
 448         void __iomem *ioaddr = rp->base;
 449 -
 450 +       u32 intr_status = rp->istat;
 451         spin_lock(&rp->lock);
 452
 453         if (intr_status & IntrLinkChange)
 454 @@ -1895,6 +1938,7 @@
 455
 456         /* Disable interrupts by clearing the interrupt mask. */
 457         iowrite16(0x0000, ioaddr + IntrEnable);
 458 +       rp->imask = 0;
 459
 460         /* Stop the chip's Tx and Rx processes. */
 461         iowrite16(CmdStop, ioaddr + ChipCmd);
 462 @@ -1906,6 +1950,9 @@
 463         free_tbufs(dev);
 464         free_ring(dev);
 465
 466 +       writeb(0x80, ioaddr + 0xa1);
 467 +       pci_set_power_state(rp->pdev, 3);
 468 +
 469         return 0;
 470  }
 471
 472 @@ -1935,6 +1982,7 @@
 473                 return; /* Nothing to do for non-WOL adapters */
 474
 475         rhine_power_init(dev);
 476 +       netif_poll_disable(dev);
 477
 478         /* Make sure we use pattern 0, 1 and not 4, 5 */
 479         if (rp->quirks & rq6patterns)
 480 Only in linux.dev/drivers/net: .via-rhine.c.swp