drivers/staging/octeon/ethernet-tx.c

   1 /*
   2  * This file is based on code from OCTEON SDK by Cavium Networks.
   3  *
   4  * Copyright (c) 2003-2010 Cavium Networks
   5  *
   6  * This file is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License, Version 2, as
   8  * published by the Free Software Foundation.
   9  */
  10
  11 #include <linux/module.h>
  12 #include <linux/kernel.h>
  13 #include <linux/netdevice.h>
  14 #include <linux/etherdevice.h>
  15 #include <linux/ip.h>
  16 #include <linux/ratelimit.h>
  17 #include <linux/string.h>
  18 #include <linux/interrupt.h>
  19 #include <net/dst.h>
  20 #ifdef CONFIG_XFRM
  21 #include <linux/xfrm.h>
  22 #include <net/xfrm.h>
  23 #endif /* CONFIG_XFRM */
  24
  25 #include <linux/atomic.h>
  26 #include <net/sch_generic.h>
  27
  28 #include <asm/octeon/octeon.h>
  29
  30 #include "ethernet-defines.h"
  31 #include "octeon-ethernet.h"
  32 #include "ethernet-tx.h"
  33 #include "ethernet-util.h"
  34
  35 #include <asm/octeon/cvmx-wqe.h>
  36 #include <asm/octeon/cvmx-fau.h>
  37 #include <asm/octeon/cvmx-pip.h>
  38 #include <asm/octeon/cvmx-pko.h>
  39 #include <asm/octeon/cvmx-helper.h>
  40
  41 #include <asm/octeon/cvmx-gmxx-defs.h>
  42
  43 #define CVM_OCT_SKB_CB(skb)     ((u64 *)((skb)->cb))
  44
  45 /*
  46  * You can define GET_SKBUFF_QOS() to override how the skbuff output
  47  * function determines which output queue is used. The default
  48  * implementation always uses the base queue for the port. If, for
  49  * example, you wanted to use the skb->priority field, define
  50  * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
  51  */
  52 #ifndef GET_SKBUFF_QOS
  53 #define GET_SKBUFF_QOS(skb) 0
  54 #endif
  55
  56 static void cvm_oct_tx_do_cleanup(unsigned long arg);
  57 static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
  58
  59 /* Maximum number of SKBs to try to free per xmit packet. */
  60 #define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
  61
  62 static inline int cvm_oct_adjust_skb_to_free(int skb_to_free, int fau)
  63 {
  64         int undo;
  65
  66         undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free +
  67                                                    MAX_SKB_TO_FREE;
  68         if (undo > 0)
  69                 cvmx_fau_atomic_add32(fau, -undo);
  70         skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE :
  71                                                        -skb_to_free;
  72         return skb_to_free;
  73 }
  74
  75 static void cvm_oct_kick_tx_poll_watchdog(void)
  76 {
  77         union cvmx_ciu_timx ciu_timx;
  78
  79         ciu_timx.u64 = 0;
  80         ciu_timx.s.one_shot = 1;
  81         ciu_timx.s.len = cvm_oct_tx_poll_interval;
  82         cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
  83 }
  84
  85 static void cvm_oct_free_tx_skbs(struct net_device *dev)
  86 {
  87         int skb_to_free;
  88         int qos, queues_per_port;
  89         int total_freed = 0;
  90         int total_remaining = 0;
  91         unsigned long flags;
  92         struct octeon_ethernet *priv = netdev_priv(dev);
  93
  94         queues_per_port = cvmx_pko_get_num_queues(priv->port);
  95         /* Drain any pending packets in the free list */
  96         for (qos = 0; qos < queues_per_port; qos++) {
  97                 if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
  98                         continue;
  99                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 100                                                        MAX_SKB_TO_FREE);
 101                 skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 102                                                          priv->fau + qos * 4);
 103                 total_freed += skb_to_free;
 104                 if (skb_to_free > 0) {
 105                         struct sk_buff *to_free_list = NULL;
 106
 107                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 108                         while (skb_to_free > 0) {
 109                                 struct sk_buff *t;
 110
 111                                 t = __skb_dequeue(&priv->tx_free_list[qos]);
 112                                 t->next = to_free_list;
 113                                 to_free_list = t;
 114                                 skb_to_free--;
 115                         }
 116                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 117                                                flags);
 118                         /* Do the actual freeing outside of the lock. */
 119                         while (to_free_list) {
 120                                 struct sk_buff *t = to_free_list;
 121
 122                                 to_free_list = to_free_list->next;
 123                                 dev_kfree_skb_any(t);
 124                         }
 125                 }
 126                 total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 127         }
 128         if (total_remaining < MAX_OUT_QUEUE_DEPTH && netif_queue_stopped(dev))
 129                 netif_wake_queue(dev);
 130         if (total_remaining)
 131                 cvm_oct_kick_tx_poll_watchdog();
 132 }
 133
 134 /**
 135  * cvm_oct_xmit - transmit a packet
 136  * @skb:    Packet to send
 137  * @dev:    Device info structure
 138  *
 139  * Returns Always returns NETDEV_TX_OK
 140  */
 141 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 142 {
 143         cvmx_pko_command_word0_t pko_command;
 144         union cvmx_buf_ptr hw_buffer;
 145         u64 old_scratch;
 146         u64 old_scratch2;
 147         int qos;
 148         int i;
 149         enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
 150         struct octeon_ethernet *priv = netdev_priv(dev);
 151         struct sk_buff *to_free_list;
 152         int skb_to_free;
 153         int buffers_to_free;
 154         u32 total_to_clean;
 155         unsigned long flags;
 156 #if REUSE_SKBUFFS_WITHOUT_FREE
 157         unsigned char *fpa_head;
 158 #endif
 159
 160         /*
 161          * Prefetch the private data structure.  It is larger than the
 162          * one cache line.
 163          */
 164         prefetch(priv);
 165
 166         /*
 167          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
 168          * completely remove "qos" in the event neither interface
 169          * supports multiple queues per port.
 170          */
 171         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
 172             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
 173                 qos = GET_SKBUFF_QOS(skb);
 174                 if (qos <= 0)
 175                         qos = 0;
 176                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
 177                         qos = 0;
 178         } else {
 179                 qos = 0;
 180         }
 181
 182         if (USE_ASYNC_IOBDMA) {
 183                 /* Save scratch in case userspace is using it */
 184                 CVMX_SYNCIOBDMA;
 185                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 186                 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 187
 188                 /*
 189                  * Fetch and increment the number of packets to be
 190                  * freed.
 191                  */
 192                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
 193                                                FAU_NUM_PACKET_BUFFERS_TO_FREE,
 194                                                0);
 195                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
 196                                                priv->fau + qos * 4,
 197                                                MAX_SKB_TO_FREE);
 198         }
 199
 200         /*
 201          * We have space for 6 segment pointers, If there will be more
 202          * than that, we must linearize.
 203          */
 204         if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
 205                 if (unlikely(__skb_linearize(skb))) {
 206                         queue_type = QUEUE_DROP;
 207                         if (USE_ASYNC_IOBDMA) {
 208                                 /*
 209                                  * Get the number of skbuffs in use
 210                                  * by the hardware
 211                                  */
 212                                 CVMX_SYNCIOBDMA;
 213                                 skb_to_free =
 214                                         cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 215                         } else {
 216                                 /*
 217                                  * Get the number of skbuffs in use
 218                                  * by the hardware
 219                                  */
 220                                 skb_to_free = cvmx_fau_fetch_and_add32(
 221                                         priv->fau + qos * 4, MAX_SKB_TO_FREE);
 222                         }
 223                         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 224                                                                  priv->fau +
 225                                                                  qos * 4);
 226                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 227                         goto skip_xmit;
 228                 }
 229         }
 230
 231         /*
 232          * The CN3XXX series of parts has an errata (GMX-401) which
 233          * causes the GMX block to hang if a collision occurs towards
 234          * the end of a <68 byte packet. As a workaround for this, we
 235          * pad packets to be 68 bytes whenever we are in half duplex
 236          * mode. We don't handle the case of having a small packet but
 237          * no room to add the padding.  The kernel should always give
 238          * us at least a cache line
 239          */
 240         if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
 241                 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
 242                 int interface = INTERFACE(priv->port);
 243                 int index = INDEX(priv->port);
 244
 245                 if (interface < 2) {
 246                         /* We only need to pad packet in half duplex mode */
 247                         gmx_prt_cfg.u64 =
 248                             cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
 249                         if (gmx_prt_cfg.s.duplex == 0) {
 250                                 int add_bytes = 64 - skb->len;
 251
 252                                 if ((skb_tail_pointer(skb) + add_bytes) <=
 253                                     skb_end_pointer(skb))
 254                                         __skb_put_zero(skb, add_bytes);
 255                         }
 256                 }
 257         }
 258
 259         /* Build the PKO command */
 260         pko_command.u64 = 0;
 261 #ifdef __LITTLE_ENDIAN
 262         pko_command.s.le = 1;
 263 #endif
 264         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
 265         pko_command.s.segs = 1;
 266         pko_command.s.total_bytes = skb->len;
 267         pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
 268         pko_command.s.subone0 = 1;
 269
 270         pko_command.s.dontfree = 1;
 271
 272         /* Build the PKO buffer pointer */
 273         hw_buffer.u64 = 0;
 274         if (skb_shinfo(skb)->nr_frags == 0) {
 275                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
 276                 hw_buffer.s.pool = 0;
 277                 hw_buffer.s.size = skb->len;
 278         } else {
 279                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
 280                 hw_buffer.s.pool = 0;
 281                 hw_buffer.s.size = skb_headlen(skb);
 282                 CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
 283                 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 284                         struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i;
 285
 286                         hw_buffer.s.addr = XKPHYS_TO_PHYS(
 287                                 (u64)(page_address(fs->page.p) +
 288                                 fs->page_offset));
 289                         hw_buffer.s.size = fs->size;
 290                         CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
 291                 }
 292                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb));
 293                 hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1;
 294                 pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1;
 295                 pko_command.s.gather = 1;
 296                 goto dont_put_skbuff_in_hw;
 297         }
 298
 299         /*
 300          * See if we can put this skb in the FPA pool. Any strange
 301          * behavior from the Linux networking stack will most likely
 302          * be caused by a bug in the following code. If some field is
 303          * in use by the network stack and gets carried over when a
 304          * buffer is reused, bad things may happen.  If in doubt and
 305          * you dont need the absolute best performance, disable the
 306          * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
 307          * shown a 25% increase in performance under some loads.
 308          */
 309 #if REUSE_SKBUFFS_WITHOUT_FREE
 310         fpa_head = skb->head + 256 - ((unsigned long)skb->head & 0x7f);
 311         if (unlikely(skb->data < fpa_head)) {
 312                 /* TX buffer beginning can't meet FPA alignment constraints */
 313                 goto dont_put_skbuff_in_hw;
 314         }
 315         if (unlikely
 316             ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
 317                 /* TX buffer isn't large enough for the FPA */
 318                 goto dont_put_skbuff_in_hw;
 319         }
 320         if (unlikely(skb_shared(skb))) {
 321                 /* TX buffer sharing data with someone else */
 322                 goto dont_put_skbuff_in_hw;
 323         }
 324         if (unlikely(skb_cloned(skb))) {
 325                 /* TX buffer has been cloned */
 326                 goto dont_put_skbuff_in_hw;
 327         }
 328         if (unlikely(skb_header_cloned(skb))) {
 329                 /* TX buffer header has been cloned */
 330                 goto dont_put_skbuff_in_hw;
 331         }
 332         if (unlikely(skb->destructor)) {
 333                 /* TX buffer has a destructor */
 334                 goto dont_put_skbuff_in_hw;
 335         }
 336         if (unlikely(skb_shinfo(skb)->nr_frags)) {
 337                 /* TX buffer has fragments */
 338                 goto dont_put_skbuff_in_hw;
 339         }
 340         if (unlikely
 341             (skb->truesize !=
 342              sizeof(*skb) + skb_end_offset(skb))) {
 343                 /* TX buffer truesize has been changed */
 344                 goto dont_put_skbuff_in_hw;
 345         }
 346
 347         /*
 348          * We can use this buffer in the FPA.  We don't need the FAU
 349          * update anymore
 350          */
 351         pko_command.s.dontfree = 0;
 352
 353         hw_buffer.s.back = ((unsigned long)skb->data >> 7) -
 354                            ((unsigned long)fpa_head >> 7);
 355
 356         *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
 357
 358         /*
 359          * The skbuff will be reused without ever being freed. We must
 360          * cleanup a bunch of core things.
 361          */
 362         dst_release(skb_dst(skb));
 363         skb_dst_set(skb, NULL);
 364 #ifdef CONFIG_XFRM
 365         secpath_put(skb->sp);
 366         skb->sp = NULL;
 367 #endif
 368         nf_reset(skb);
 369
 370 #ifdef CONFIG_NET_SCHED
 371         skb->tc_index = 0;
 372         skb_reset_tc(skb);
 373 #endif /* CONFIG_NET_SCHED */
 374 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
 375
 376 dont_put_skbuff_in_hw:
 377
 378         /* Check if we can use the hardware checksumming */
 379         if ((skb->protocol == htons(ETH_P_IP)) &&
 380             (ip_hdr(skb)->version == 4) &&
 381             (ip_hdr(skb)->ihl == 5) &&
 382             ((ip_hdr(skb)->frag_off == 0) ||
 383              (ip_hdr(skb)->frag_off == htons(1 << 14))) &&
 384             ((ip_hdr(skb)->protocol == IPPROTO_TCP) ||
 385              (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
 386                 /* Use hardware checksum calc */
 387                 pko_command.s.ipoffp1 = skb_network_offset(skb) + 1;
 388         }
 389
 390         if (USE_ASYNC_IOBDMA) {
 391                 /* Get the number of skbuffs in use by the hardware */
 392                 CVMX_SYNCIOBDMA;
 393                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 394                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 395         } else {
 396                 /* Get the number of skbuffs in use by the hardware */
 397                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 398                                                        MAX_SKB_TO_FREE);
 399                 buffers_to_free =
 400                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 401         }
 402
 403         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 404                                                  priv->fau + qos * 4);
 405
 406         /*
 407          * If we're sending faster than the receive can free them then
 408          * don't do the HW free.
 409          */
 410         if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 411                 pko_command.s.dontfree = 1;
 412
 413         if (pko_command.s.dontfree) {
 414                 queue_type = QUEUE_CORE;
 415                 pko_command.s.reg0 = priv->fau + qos * 4;
 416         } else {
 417                 queue_type = QUEUE_HW;
 418         }
 419         if (USE_ASYNC_IOBDMA)
 420                 cvmx_fau_async_fetch_and_add32(
 421                                 CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
 422
 423         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 424
 425         /* Drop this packet if we have too many already queued to the HW */
 426         if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >=
 427                      MAX_OUT_QUEUE_DEPTH)) {
 428                 if (dev->tx_queue_len != 0) {
 429                         /* Drop the lock when notifying the core.  */
 430                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 431                                                flags);
 432                         netif_stop_queue(dev);
 433                         spin_lock_irqsave(&priv->tx_free_list[qos].lock,
 434                                           flags);
 435                 } else {
 436                         /* If not using normal queueing.  */
 437                         queue_type = QUEUE_DROP;
 438                         goto skip_xmit;
 439                 }
 440         }
 441
 442         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
 443                                      CVMX_PKO_LOCK_NONE);
 444
 445         /* Send the packet to the output queue */
 446         if (unlikely(cvmx_pko_send_packet_finish(priv->port,
 447                                                  priv->queue + qos,
 448                                                  pko_command, hw_buffer,
 449                                                  CVMX_PKO_LOCK_NONE))) {
 450                 printk_ratelimited("%s: Failed to send the packet\n",
 451                                    dev->name);
 452                 queue_type = QUEUE_DROP;
 453         }
 454 skip_xmit:
 455         to_free_list = NULL;
 456
 457         switch (queue_type) {
 458         case QUEUE_DROP:
 459                 skb->next = to_free_list;
 460                 to_free_list = skb;
 461                 dev->stats.tx_dropped++;
 462                 break;
 463         case QUEUE_HW:
 464                 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
 465                 break;
 466         case QUEUE_CORE:
 467                 __skb_queue_tail(&priv->tx_free_list[qos], skb);
 468                 break;
 469         default:
 470                 BUG();
 471         }
 472
 473         while (skb_to_free > 0) {
 474                 struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
 475
 476                 t->next = to_free_list;
 477                 to_free_list = t;
 478                 skb_to_free--;
 479         }
 480
 481         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 482
 483         /* Do the actual freeing outside of the lock. */
 484         while (to_free_list) {
 485                 struct sk_buff *t = to_free_list;
 486
 487                 to_free_list = to_free_list->next;
 488                 dev_kfree_skb_any(t);
 489         }
 490
 491         if (USE_ASYNC_IOBDMA) {
 492                 CVMX_SYNCIOBDMA;
 493                 total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 494                 /* Restore the scratch area */
 495                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 496                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 497         } else {
 498                 total_to_clean = cvmx_fau_fetch_and_add32(
 499                                                 FAU_TOTAL_TX_TO_CLEAN, 1);
 500         }
 501
 502         if (total_to_clean & 0x3ff) {
 503                 /*
 504                  * Schedule the cleanup tasklet every 1024 packets for
 505                  * the pathological case of high traffic on one port
 506                  * delaying clean up of packets on a different port
 507                  * that is blocked waiting for the cleanup.
 508                  */
 509                 tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 510         }
 511
 512         cvm_oct_kick_tx_poll_watchdog();
 513
 514         return NETDEV_TX_OK;
 515 }
 516
 517 /**
 518  * cvm_oct_xmit_pow - transmit a packet to the POW
 519  * @skb:    Packet to send
 520  * @dev:    Device info structure
 521
 522  * Returns Always returns zero
 523  */
 524 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
 525 {
 526         struct octeon_ethernet *priv = netdev_priv(dev);
 527         void *packet_buffer;
 528         void *copy_location;
 529
 530         /* Get a work queue entry */
 531         cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
 532
 533         if (unlikely(!work)) {
 534                 printk_ratelimited("%s: Failed to allocate a work queue entry\n",
 535                                    dev->name);
 536                 dev->stats.tx_dropped++;
 537                 dev_kfree_skb_any(skb);
 538                 return 0;
 539         }
 540
 541         /* Get a packet buffer */
 542         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
 543         if (unlikely(!packet_buffer)) {
 544                 printk_ratelimited("%s: Failed to allocate a packet buffer\n",
 545                                    dev->name);
 546                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 547                 dev->stats.tx_dropped++;
 548                 dev_kfree_skb_any(skb);
 549                 return 0;
 550         }
 551
 552         /*
 553          * Calculate where we need to copy the data to. We need to
 554          * leave 8 bytes for a next pointer (unused). We also need to
 555          * include any configure skip. Then we need to align the IP
 556          * packet src and dest into the same 64bit word. The below
 557          * calculation may add a little extra, but that doesn't
 558          * hurt.
 559          */
 560         copy_location = packet_buffer + sizeof(u64);
 561         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
 562
 563         /*
 564          * We have to copy the packet since whoever processes this
 565          * packet will free it to a hardware pool. We can't use the
 566          * trick of counting outstanding packets like in
 567          * cvm_oct_xmit.
 568          */
 569         memcpy(copy_location, skb->data, skb->len);
 570
 571         /*
 572          * Fill in some of the work queue fields. We may need to add
 573          * more if the software at the other end needs them.
 574          */
 575         if (!OCTEON_IS_MODEL(OCTEON_CN68XX))
 576                 work->word0.pip.cn38xx.hw_chksum = skb->csum;
 577         work->word1.len = skb->len;
 578         cvmx_wqe_set_port(work, priv->port);
 579         cvmx_wqe_set_qos(work, priv->port & 0x7);
 580         cvmx_wqe_set_grp(work, pow_send_group);
 581         work->word1.tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
 582         work->word1.tag = pow_send_group;       /* FIXME */
 583         /* Default to zero. Sets of zero later are commented out */
 584         work->word2.u64 = 0;
 585         work->word2.s.bufs = 1;
 586         work->packet_ptr.u64 = 0;
 587         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
 588         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
 589         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
 590         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
 591
 592         if (skb->protocol == htons(ETH_P_IP)) {
 593                 work->word2.s.ip_offset = 14;
 594 #if 0
 595                 work->word2.s.vlan_valid = 0;   /* FIXME */
 596                 work->word2.s.vlan_cfi = 0;     /* FIXME */
 597                 work->word2.s.vlan_id = 0;      /* FIXME */
 598                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
 599 #endif
 600                 work->word2.s.tcp_or_udp =
 601                     (ip_hdr(skb)->protocol == IPPROTO_TCP) ||
 602                     (ip_hdr(skb)->protocol == IPPROTO_UDP);
 603 #if 0
 604                 /* FIXME */
 605                 work->word2.s.dec_ipsec = 0;
 606                 /* We only support IPv4 right now */
 607                 work->word2.s.is_v6 = 0;
 608                 /* Hardware would set to zero */
 609                 work->word2.s.software = 0;
 610                 /* No error, packet is internal */
 611                 work->word2.s.L4_error = 0;
 612 #endif
 613                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0) ||
 614                                           (ip_hdr(skb)->frag_off ==
 615                                               1 << 14));
 616 #if 0
 617                 /* Assume Linux is sending a good packet */
 618                 work->word2.s.IP_exc = 0;
 619 #endif
 620                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
 621                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
 622 #if 0
 623                 /* This is an IP packet */
 624                 work->word2.s.not_IP = 0;
 625                 /* No error, packet is internal */
 626                 work->word2.s.rcv_error = 0;
 627                 /* No error, packet is internal */
 628                 work->word2.s.err_code = 0;
 629 #endif
 630
 631                 /*
 632                  * When copying the data, include 4 bytes of the
 633                  * ethernet header to align the same way hardware
 634                  * does.
 635                  */
 636                 memcpy(work->packet_data, skb->data + 10,
 637                        sizeof(work->packet_data));
 638         } else {
 639 #if 0
 640                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
 641                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
 642                 work->word2.snoip.vlan_id = 0;  /* FIXME */
 643                 work->word2.snoip.software = 0; /* Hardware would set to zero */
 644 #endif
 645                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
 646                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
 647                 work->word2.snoip.is_bcast =
 648                     (skb->pkt_type == PACKET_BROADCAST);
 649                 work->word2.snoip.is_mcast =
 650                     (skb->pkt_type == PACKET_MULTICAST);
 651                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
 652 #if 0
 653                 /* No error, packet is internal */
 654                 work->word2.snoip.rcv_error = 0;
 655                 /* No error, packet is internal */
 656                 work->word2.snoip.err_code = 0;
 657 #endif
 658                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
 659         }
 660
 661         /* Submit the packet to the POW */
 662         cvmx_pow_work_submit(work, work->word1.tag, work->word1.tag_type,
 663                              cvmx_wqe_get_qos(work), cvmx_wqe_get_grp(work));
 664         dev->stats.tx_packets++;
 665         dev->stats.tx_bytes += skb->len;
 666         dev_consume_skb_any(skb);
 667         return 0;
 668 }
 669
 670 /**
 671  * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
 672  * @dev:    Device being shutdown
 673  *
 674  */
 675 void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 676 {
 677         struct octeon_ethernet *priv = netdev_priv(dev);
 678         unsigned long flags;
 679         int qos;
 680
 681         for (qos = 0; qos < 16; qos++) {
 682                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 683                 while (skb_queue_len(&priv->tx_free_list[qos]))
 684                         dev_kfree_skb_any(__skb_dequeue
 685                                           (&priv->tx_free_list[qos]));
 686                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 687         }
 688 }
 689
 690 static void cvm_oct_tx_do_cleanup(unsigned long arg)
 691 {
 692         int port;
 693
 694         for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
 695                 if (cvm_oct_device[port]) {
 696                         struct net_device *dev = cvm_oct_device[port];
 697
 698                         cvm_oct_free_tx_skbs(dev);
 699                 }
 700         }
 701 }
 702
 703 static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
 704 {
 705         /* Disable the interrupt.  */
 706         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 707         /* Do the work in the tasklet.  */
 708         tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 709         return IRQ_HANDLED;
 710 }
 711
 712 void cvm_oct_tx_initialize(void)
 713 {
 714         int i;
 715
 716         /* Disable the interrupt.  */
 717         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 718         /* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
 719         i = request_irq(OCTEON_IRQ_TIMER1,
 720                         cvm_oct_tx_cleanup_watchdog, 0,
 721                         "Ethernet", cvm_oct_device);
 722
 723         if (i)
 724                 panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
 725 }
 726
 727 void cvm_oct_tx_shutdown(void)
 728 {
 729         /* Free the interrupt handler */
 730         free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
 731 }