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
  27 #include <asm/octeon/octeon.h>
  28
  29 #include "ethernet-defines.h"
  30 #include "octeon-ethernet.h"
  31 #include "ethernet-tx.h"
  32 #include "ethernet-util.h"
  33
  34 #include <asm/octeon/cvmx-wqe.h>
  35 #include <asm/octeon/cvmx-fau.h>
  36 #include <asm/octeon/cvmx-pip.h>
  37 #include <asm/octeon/cvmx-pko.h>
  38 #include <asm/octeon/cvmx-helper.h>
  39
  40 #include <asm/octeon/cvmx-gmxx-defs.h>
  41
  42 #define CVM_OCT_SKB_CB(skb)     ((u64 *)((skb)->cb))
  43
  44 /*
  45  * You can define GET_SKBUFF_QOS() to override how the skbuff output
  46  * function determines which output queue is used. The default
  47  * implementation always uses the base queue for the port. If, for
  48  * example, you wanted to use the skb->priority field, define
  49  * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
  50  */
  51 #ifndef GET_SKBUFF_QOS
  52 #define GET_SKBUFF_QOS(skb) 0
  53 #endif
  54
  55 static void cvm_oct_tx_do_cleanup(unsigned long arg);
  56 static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
  57
  58 /* Maximum number of SKBs to try to free per xmit packet. */
  59 #define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
  60
  61 static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
  62 {
  63         int32_t undo;
  64
  65         undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free +
  66                                                    MAX_SKB_TO_FREE;
  67         if (undo > 0)
  68                 cvmx_fau_atomic_add32(fau, -undo);
  69         skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE :
  70                                                        -skb_to_free;
  71         return skb_to_free;
  72 }
  73
  74 static void cvm_oct_kick_tx_poll_watchdog(void)
  75 {
  76         union cvmx_ciu_timx ciu_timx;
  77
  78         ciu_timx.u64 = 0;
  79         ciu_timx.s.one_shot = 1;
  80         ciu_timx.s.len = cvm_oct_tx_poll_interval;
  81         cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
  82 }
  83
  84 static void cvm_oct_free_tx_skbs(struct net_device *dev)
  85 {
  86         int32_t skb_to_free;
  87         int qos, queues_per_port;
  88         int total_freed = 0;
  89         int total_remaining = 0;
  90         unsigned long flags;
  91         struct octeon_ethernet *priv = netdev_priv(dev);
  92
  93         queues_per_port = cvmx_pko_get_num_queues(priv->port);
  94         /* Drain any pending packets in the free list */
  95         for (qos = 0; qos < queues_per_port; qos++) {
  96                 if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
  97                         continue;
  98                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau+qos*4,
  99                                                        MAX_SKB_TO_FREE);
 100                 skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 101                                                          priv->fau+qos*4);
 102
 103
 104                 total_freed += skb_to_free;
 105                 if (skb_to_free > 0) {
 106                         struct sk_buff *to_free_list = NULL;
 107
 108                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 109                         while (skb_to_free > 0) {
 110                                 struct sk_buff *t;
 111
 112                                 t = __skb_dequeue(&priv->tx_free_list[qos]);
 113                                 t->next = to_free_list;
 114                                 to_free_list = t;
 115                                 skb_to_free--;
 116                         }
 117                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 118                                                flags);
 119                         /* Do the actual freeing outside of the lock. */
 120                         while (to_free_list) {
 121                                 struct sk_buff *t = to_free_list;
 122
 123                                 to_free_list = to_free_list->next;
 124                                 dev_kfree_skb_any(t);
 125                         }
 126                 }
 127                 total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 128         }
 129         if (total_freed >= 0 && netif_queue_stopped(dev))
 130                 netif_wake_queue(dev);
 131         if (total_remaining)
 132                 cvm_oct_kick_tx_poll_watchdog();
 133 }
 134
 135 /**
 136  * cvm_oct_xmit - transmit a packet
 137  * @skb:    Packet to send
 138  * @dev:    Device info structure
 139  *
 140  * Returns Always returns NETDEV_TX_OK
 141  */
 142 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 143 {
 144         cvmx_pko_command_word0_t pko_command;
 145         union cvmx_buf_ptr hw_buffer;
 146         uint64_t old_scratch;
 147         uint64_t old_scratch2;
 148         int qos;
 149         int i;
 150         enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
 151         struct octeon_ethernet *priv = netdev_priv(dev);
 152         struct sk_buff *to_free_list;
 153         int32_t skb_to_free;
 154         int32_t buffers_to_free;
 155         u32 total_to_clean;
 156         unsigned long flags;
 157 #if REUSE_SKBUFFS_WITHOUT_FREE
 158         unsigned char *fpa_head;
 159 #endif
 160
 161         /*
 162          * Prefetch the private data structure.  It is larger than the
 163          * one cache line.
 164          */
 165         prefetch(priv);
 166
 167         /*
 168          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
 169          * completely remove "qos" in the event neither interface
 170          * supports multiple queues per port.
 171          */
 172         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
 173             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
 174                 qos = GET_SKBUFF_QOS(skb);
 175                 if (qos <= 0)
 176                         qos = 0;
 177                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
 178                         qos = 0;
 179         } else
 180                 qos = 0;
 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 + qos * 4);
 225                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 226                         goto skip_xmit;
 227                 }
 228         }
 229
 230         /*
 231          * The CN3XXX series of parts has an errata (GMX-401) which
 232          * causes the GMX block to hang if a collision occurs towards
 233          * the end of a <68 byte packet. As a workaround for this, we
 234          * pad packets to be 68 bytes whenever we are in half duplex
 235          * mode. We don't handle the case of having a small packet but
 236          * no room to add the padding.  The kernel should always give
 237          * us at least a cache line
 238          */
 239         if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
 240                 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
 241                 int interface = INTERFACE(priv->port);
 242                 int index = INDEX(priv->port);
 243
 244                 if (interface < 2) {
 245                         /* We only need to pad packet in half duplex mode */
 246                         gmx_prt_cfg.u64 =
 247                             cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
 248                         if (gmx_prt_cfg.s.duplex == 0) {
 249                                 int add_bytes = 64 - skb->len;
 250
 251                                 if ((skb_tail_pointer(skb) + add_bytes) <=
 252                                     skb_end_pointer(skb))
 253                                         memset(__skb_put(skb, add_bytes), 0,
 254                                                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                 /*
 313                  * printk("TX buffer beginning can't meet FPA
 314                  * alignment constraints\n");
 315                  */
 316                 goto dont_put_skbuff_in_hw;
 317         }
 318         if (unlikely
 319             ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
 320                 /*
 321                    printk("TX buffer isn't large enough for the FPA\n");
 322                  */
 323                 goto dont_put_skbuff_in_hw;
 324         }
 325         if (unlikely(skb_shared(skb))) {
 326                 /*
 327                    printk("TX buffer sharing data with someone else\n");
 328                  */
 329                 goto dont_put_skbuff_in_hw;
 330         }
 331         if (unlikely(skb_cloned(skb))) {
 332                 /*
 333                    printk("TX buffer has been cloned\n");
 334                  */
 335                 goto dont_put_skbuff_in_hw;
 336         }
 337         if (unlikely(skb_header_cloned(skb))) {
 338                 /*
 339                    printk("TX buffer header has been cloned\n");
 340                  */
 341                 goto dont_put_skbuff_in_hw;
 342         }
 343         if (unlikely(skb->destructor)) {
 344                 /*
 345                    printk("TX buffer has a destructor\n");
 346                  */
 347                 goto dont_put_skbuff_in_hw;
 348         }
 349         if (unlikely(skb_shinfo(skb)->nr_frags)) {
 350                 /*
 351                    printk("TX buffer has fragments\n");
 352                  */
 353                 goto dont_put_skbuff_in_hw;
 354         }
 355         if (unlikely
 356             (skb->truesize !=
 357              sizeof(*skb) + skb_end_offset(skb))) {
 358                 /*
 359                    printk("TX buffer truesize has been changed\n");
 360                  */
 361                 goto dont_put_skbuff_in_hw;
 362         }
 363
 364         /*
 365          * We can use this buffer in the FPA.  We don't need the FAU
 366          * update anymore
 367          */
 368         pko_command.s.dontfree = 0;
 369
 370         hw_buffer.s.back = ((unsigned long)skb->data >> 7) -
 371                            ((unsigned long)fpa_head >> 7);
 372
 373         *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
 374
 375         /*
 376          * The skbuff will be reused without ever being freed. We must
 377          * cleanup a bunch of core things.
 378          */
 379         dst_release(skb_dst(skb));
 380         skb_dst_set(skb, NULL);
 381 #ifdef CONFIG_XFRM
 382         secpath_put(skb->sp);
 383         skb->sp = NULL;
 384 #endif
 385         nf_reset(skb);
 386
 387 #ifdef CONFIG_NET_SCHED
 388         skb->tc_index = 0;
 389 #ifdef CONFIG_NET_CLS_ACT
 390         skb->tc_verd = 0;
 391 #endif /* CONFIG_NET_CLS_ACT */
 392 #endif /* CONFIG_NET_SCHED */
 393 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
 394
 395 dont_put_skbuff_in_hw:
 396
 397         /* Check if we can use the hardware checksumming */
 398         if ((skb->protocol == htons(ETH_P_IP)) &&
 399             (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) &&
 400             ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == htons(1 << 14)))
 401             && ((ip_hdr(skb)->protocol == IPPROTO_TCP)
 402                 || (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
 403                 /* Use hardware checksum calc */
 404                 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
 405         }
 406
 407         if (USE_ASYNC_IOBDMA) {
 408                 /* Get the number of skbuffs in use by the hardware */
 409                 CVMX_SYNCIOBDMA;
 410                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 411                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 412         } else {
 413                 /* Get the number of skbuffs in use by the hardware */
 414                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 415                                                        MAX_SKB_TO_FREE);
 416                 buffers_to_free =
 417                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 418         }
 419
 420         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4);
 421
 422         /*
 423          * If we're sending faster than the receive can free them then
 424          * don't do the HW free.
 425          */
 426         if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 427                 pko_command.s.dontfree = 1;
 428
 429         if (pko_command.s.dontfree) {
 430                 queue_type = QUEUE_CORE;
 431                 pko_command.s.reg0 = priv->fau+qos*4;
 432         } else {
 433                 queue_type = QUEUE_HW;
 434         }
 435         if (USE_ASYNC_IOBDMA)
 436                 cvmx_fau_async_fetch_and_add32(
 437                                 CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
 438
 439         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 440
 441         /* Drop this packet if we have too many already queued to the HW */
 442         if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >=
 443                      MAX_OUT_QUEUE_DEPTH)) {
 444
 445                 if (dev->tx_queue_len != 0) {
 446                         /* Drop the lock when notifying the core.  */
 447                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 448                                                flags);
 449                         netif_stop_queue(dev);
 450                         spin_lock_irqsave(&priv->tx_free_list[qos].lock,
 451                                           flags);
 452                 } else {
 453                         /* If not using normal queueing.  */
 454                         queue_type = QUEUE_DROP;
 455                         goto skip_xmit;
 456                 }
 457         }
 458
 459         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
 460                                      CVMX_PKO_LOCK_NONE);
 461
 462         /* Send the packet to the output queue */
 463         if (unlikely(cvmx_pko_send_packet_finish(priv->port,
 464                                                  priv->queue + qos,
 465                                                  pko_command, hw_buffer,
 466                                                  CVMX_PKO_LOCK_NONE))) {
 467                 printk_ratelimited("%s: Failed to send the packet\n",
 468                                    dev->name);
 469                 queue_type = QUEUE_DROP;
 470         }
 471 skip_xmit:
 472         to_free_list = NULL;
 473
 474         switch (queue_type) {
 475         case QUEUE_DROP:
 476                 skb->next = to_free_list;
 477                 to_free_list = skb;
 478                 priv->stats.tx_dropped++;
 479                 break;
 480         case QUEUE_HW:
 481                 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
 482                 break;
 483         case QUEUE_CORE:
 484                 __skb_queue_tail(&priv->tx_free_list[qos], skb);
 485                 break;
 486         default:
 487                 BUG();
 488         }
 489
 490         while (skb_to_free > 0) {
 491                 struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
 492
 493                 t->next = to_free_list;
 494                 to_free_list = t;
 495                 skb_to_free--;
 496         }
 497
 498         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 499
 500         /* Do the actual freeing outside of the lock. */
 501         while (to_free_list) {
 502                 struct sk_buff *t = to_free_list;
 503
 504                 to_free_list = to_free_list->next;
 505                 dev_kfree_skb_any(t);
 506         }
 507
 508         if (USE_ASYNC_IOBDMA) {
 509                 CVMX_SYNCIOBDMA;
 510                 total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 511                 /* Restore the scratch area */
 512                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 513                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 514         } else {
 515                 total_to_clean = cvmx_fau_fetch_and_add32(
 516                                                 FAU_TOTAL_TX_TO_CLEAN, 1);
 517         }
 518
 519         if (total_to_clean & 0x3ff) {
 520                 /*
 521                  * Schedule the cleanup tasklet every 1024 packets for
 522                  * the pathological case of high traffic on one port
 523                  * delaying clean up of packets on a different port
 524                  * that is blocked waiting for the cleanup.
 525                  */
 526                 tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 527         }
 528
 529         cvm_oct_kick_tx_poll_watchdog();
 530
 531         return NETDEV_TX_OK;
 532 }
 533
 534 /**
 535  * cvm_oct_xmit_pow - transmit a packet to the POW
 536  * @skb:    Packet to send
 537  * @dev:    Device info structure
 538
 539  * Returns Always returns zero
 540  */
 541 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
 542 {
 543         struct octeon_ethernet *priv = netdev_priv(dev);
 544         void *packet_buffer;
 545         void *copy_location;
 546
 547         /* Get a work queue entry */
 548         cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
 549
 550         if (unlikely(work == NULL)) {
 551                 printk_ratelimited("%s: Failed to allocate a work queue entry\n",
 552                                    dev->name);
 553                 priv->stats.tx_dropped++;
 554                 dev_kfree_skb_any(skb);
 555                 return 0;
 556         }
 557
 558         /* Get a packet buffer */
 559         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
 560         if (unlikely(packet_buffer == NULL)) {
 561                 printk_ratelimited("%s: Failed to allocate a packet buffer\n",
 562                                    dev->name);
 563                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 564                 priv->stats.tx_dropped++;
 565                 dev_kfree_skb_any(skb);
 566                 return 0;
 567         }
 568
 569         /*
 570          * Calculate where we need to copy the data to. We need to
 571          * leave 8 bytes for a next pointer (unused). We also need to
 572          * include any configure skip. Then we need to align the IP
 573          * packet src and dest into the same 64bit word. The below
 574          * calculation may add a little extra, but that doesn't
 575          * hurt.
 576          */
 577         copy_location = packet_buffer + sizeof(uint64_t);
 578         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
 579
 580         /*
 581          * We have to copy the packet since whoever processes this
 582          * packet will free it to a hardware pool. We can't use the
 583          * trick of counting outstanding packets like in
 584          * cvm_oct_xmit.
 585          */
 586         memcpy(copy_location, skb->data, skb->len);
 587
 588         /*
 589          * Fill in some of the work queue fields. We may need to add
 590          * more if the software at the other end needs them.
 591          */
 592         if (!OCTEON_IS_MODEL(OCTEON_CN68XX))
 593                 work->word0.pip.cn38xx.hw_chksum = skb->csum;
 594         work->word1.len = skb->len;
 595         cvmx_wqe_set_port(work, priv->port);
 596         cvmx_wqe_set_qos(work, priv->port & 0x7);
 597         cvmx_wqe_set_grp(work, pow_send_group);
 598         work->word1.tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
 599         work->word1.tag = pow_send_group;       /* FIXME */
 600         /* Default to zero. Sets of zero later are commented out */
 601         work->word2.u64 = 0;
 602         work->word2.s.bufs = 1;
 603         work->packet_ptr.u64 = 0;
 604         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
 605         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
 606         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
 607         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
 608
 609         if (skb->protocol == htons(ETH_P_IP)) {
 610                 work->word2.s.ip_offset = 14;
 611 #if 0
 612                 work->word2.s.vlan_valid = 0;   /* FIXME */
 613                 work->word2.s.vlan_cfi = 0;     /* FIXME */
 614                 work->word2.s.vlan_id = 0;      /* FIXME */
 615                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
 616 #endif
 617                 work->word2.s.tcp_or_udp =
 618                     (ip_hdr(skb)->protocol == IPPROTO_TCP)
 619                     || (ip_hdr(skb)->protocol == IPPROTO_UDP);
 620 #if 0
 621                 /* FIXME */
 622                 work->word2.s.dec_ipsec = 0;
 623                 /* We only support IPv4 right now */
 624                 work->word2.s.is_v6 = 0;
 625                 /* Hardware would set to zero */
 626                 work->word2.s.software = 0;
 627                 /* No error, packet is internal */
 628                 work->word2.s.L4_error = 0;
 629 #endif
 630                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0)
 631                                           || (ip_hdr(skb)->frag_off ==
 632                                               1 << 14));
 633 #if 0
 634                 /* Assume Linux is sending a good packet */
 635                 work->word2.s.IP_exc = 0;
 636 #endif
 637                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
 638                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
 639 #if 0
 640                 /* This is an IP packet */
 641                 work->word2.s.not_IP = 0;
 642                 /* No error, packet is internal */
 643                 work->word2.s.rcv_error = 0;
 644                 /* No error, packet is internal */
 645                 work->word2.s.err_code = 0;
 646 #endif
 647
 648                 /*
 649                  * When copying the data, include 4 bytes of the
 650                  * ethernet header to align the same way hardware
 651                  * does.
 652                  */
 653                 memcpy(work->packet_data, skb->data + 10,
 654                        sizeof(work->packet_data));
 655         } else {
 656 #if 0
 657                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
 658                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
 659                 work->word2.snoip.vlan_id = 0;  /* FIXME */
 660                 work->word2.snoip.software = 0; /* Hardware would set to zero */
 661 #endif
 662                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
 663                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
 664                 work->word2.snoip.is_bcast =
 665                     (skb->pkt_type == PACKET_BROADCAST);
 666                 work->word2.snoip.is_mcast =
 667                     (skb->pkt_type == PACKET_MULTICAST);
 668                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
 669 #if 0
 670                 /* No error, packet is internal */
 671                 work->word2.snoip.rcv_error = 0;
 672                 /* No error, packet is internal */
 673                 work->word2.snoip.err_code = 0;
 674 #endif
 675                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
 676         }
 677
 678         /* Submit the packet to the POW */
 679         cvmx_pow_work_submit(work, work->word1.tag, work->word1.tag_type,
 680                              cvmx_wqe_get_qos(work), cvmx_wqe_get_grp(work));
 681         priv->stats.tx_packets++;
 682         priv->stats.tx_bytes += skb->len;
 683         dev_consume_skb_any(skb);
 684         return 0;
 685 }
 686
 687 /**
 688  * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
 689  * @dev:    Device being shutdown
 690  *
 691  */
 692 void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 693 {
 694         struct octeon_ethernet *priv = netdev_priv(dev);
 695         unsigned long flags;
 696         int qos;
 697
 698         for (qos = 0; qos < 16; qos++) {
 699                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 700                 while (skb_queue_len(&priv->tx_free_list[qos]))
 701                         dev_kfree_skb_any(__skb_dequeue
 702                                           (&priv->tx_free_list[qos]));
 703                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 704         }
 705 }
 706
 707 static void cvm_oct_tx_do_cleanup(unsigned long arg)
 708 {
 709         int port;
 710
 711         for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
 712                 if (cvm_oct_device[port]) {
 713                         struct net_device *dev = cvm_oct_device[port];
 714
 715                         cvm_oct_free_tx_skbs(dev);
 716                 }
 717         }
 718 }
 719
 720 static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
 721 {
 722         /* Disable the interrupt.  */
 723         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 724         /* Do the work in the tasklet.  */
 725         tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 726         return IRQ_HANDLED;
 727 }
 728
 729 void cvm_oct_tx_initialize(void)
 730 {
 731         int i;
 732
 733         /* Disable the interrupt.  */
 734         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 735         /* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
 736         i = request_irq(OCTEON_IRQ_TIMER1,
 737                         cvm_oct_tx_cleanup_watchdog, 0,
 738                         "Ethernet", cvm_oct_device);
 739
 740         if (i)
 741                 panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
 742 }
 743
 744 void cvm_oct_tx_shutdown(void)
 745 {
 746         /* Free the interrupt handler */
 747         free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
 748 }