drivers/staging/octeon/ethernet-tx.c

   1 /*********************************************************************
   2  * Author: Cavium Networks
   3  *
   4  * Contact: support@caviumnetworks.com
   5  * This file is part of the OCTEON SDK
   6  *
   7  * Copyright (c) 2003-2010 Cavium Networks
   8  *
   9  * This file is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License, Version 2, as
  11  * published by the Free Software Foundation.
  12  *
  13  * This file is distributed in the hope that it will be useful, but
  14  * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty
  15  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or
  16  * NONINFRINGEMENT.  See the GNU General Public License for more
  17  * details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this file; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
  22  * or visit http://www.gnu.org/licenses/.
  23  *
  24  * This file may also be available under a different license from Cavium.
  25  * Contact Cavium Networks for more information
  26 *********************************************************************/
  27 #include <linux/module.h>
  28 #include <linux/kernel.h>
  29 #include <linux/netdevice.h>
  30 #include <linux/init.h>
  31 #include <linux/etherdevice.h>
  32 #include <linux/ip.h>
  33 #include <linux/string.h>
  34 #include <net/dst.h>
  35 #ifdef CONFIG_XFRM
  36 #include <linux/xfrm.h>
  37 #include <net/xfrm.h>
  38 #endif /* CONFIG_XFRM */
  39
  40 #include <asm/atomic.h>
  41
  42 #include <asm/octeon/octeon.h>
  43
  44 #include "ethernet-defines.h"
  45 #include "octeon-ethernet.h"
  46 #include "ethernet-tx.h"
  47 #include "ethernet-util.h"
  48
  49 #include "cvmx-wqe.h"
  50 #include "cvmx-fau.h"
  51 #include "cvmx-pip.h"
  52 #include "cvmx-pko.h"
  53 #include "cvmx-helper.h"
  54
  55 #include "cvmx-gmxx-defs.h"
  56
  57 #define CVM_OCT_SKB_CB(skb)     ((u64 *)((skb)->cb))
  58
  59 /*
  60  * You can define GET_SKBUFF_QOS() to override how the skbuff output
  61  * function determines which output queue is used. The default
  62  * implementation always uses the base queue for the port. If, for
  63  * example, you wanted to use the skb->priority fieid, define
  64  * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
  65  */
  66 #ifndef GET_SKBUFF_QOS
  67 #define GET_SKBUFF_QOS(skb) 0
  68 #endif
  69
  70 static void cvm_oct_tx_do_cleanup(unsigned long arg);
  71 static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
  72
  73 /* Maximum number of SKBs to try to free per xmit packet. */
  74 #define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
  75
  76 static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
  77 {
  78         int32_t undo;
  79         undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free + MAX_SKB_TO_FREE;
  80         if (undo > 0)
  81                 cvmx_fau_atomic_add32(fau, -undo);
  82         skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE : -skb_to_free;
  83         return skb_to_free;
  84 }
  85
  86 static void cvm_oct_kick_tx_poll_watchdog(void)
  87 {
  88         union cvmx_ciu_timx ciu_timx;
  89         ciu_timx.u64 = 0;
  90         ciu_timx.s.one_shot = 1;
  91         ciu_timx.s.len = cvm_oct_tx_poll_interval;
  92         cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
  93 }
  94
  95 void cvm_oct_free_tx_skbs(struct net_device *dev)
  96 {
  97         int32_t skb_to_free;
  98         int qos, queues_per_port;
  99         int total_freed = 0;
 100         int total_remaining = 0;
 101         unsigned long flags;
 102         struct octeon_ethernet *priv = netdev_priv(dev);
 103
 104         queues_per_port = cvmx_pko_get_num_queues(priv->port);
 105         /* Drain any pending packets in the free list */
 106         for (qos = 0; qos < queues_per_port; qos++) {
 107                 if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
 108                         continue;
 109                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau+qos*4, MAX_SKB_TO_FREE);
 110                 skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4);
 111
 112
 113                 total_freed += skb_to_free;
 114                 if (skb_to_free > 0) {
 115                         struct sk_buff *to_free_list = NULL;
 116                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 117                         while (skb_to_free > 0) {
 118                                 struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
 119                                 t->next = to_free_list;
 120                                 to_free_list = t;
 121                                 skb_to_free--;
 122                         }
 123                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 124                         /* Do the actual freeing outside of the lock. */
 125                         while (to_free_list) {
 126                                 struct sk_buff *t = to_free_list;
 127                                 to_free_list = to_free_list->next;
 128                                 dev_kfree_skb_any(t);
 129                         }
 130                 }
 131                 total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 132         }
 133         if (total_freed >= 0 && netif_queue_stopped(dev))
 134                 netif_wake_queue(dev);
 135         if (total_remaining)
 136                 cvm_oct_kick_tx_poll_watchdog();
 137 }
 138
 139 /**
 140  * cvm_oct_xmit - transmit a packet
 141  * @skb:    Packet to send
 142  * @dev:    Device info structure
 143  *
 144  * Returns Always returns NETDEV_TX_OK
 145  */
 146 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 147 {
 148         cvmx_pko_command_word0_t pko_command;
 149         union cvmx_buf_ptr hw_buffer;
 150         uint64_t old_scratch;
 151         uint64_t old_scratch2;
 152         int qos;
 153         int i;
 154         enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
 155         struct octeon_ethernet *priv = netdev_priv(dev);
 156         struct sk_buff *to_free_list;
 157         int32_t skb_to_free;
 158         int32_t buffers_to_free;
 159         u32 total_to_clean;
 160         unsigned long flags;
 161 #if REUSE_SKBUFFS_WITHOUT_FREE
 162         unsigned char *fpa_head;
 163 #endif
 164
 165         /*
 166          * Prefetch the private data structure.  It is larger that one
 167          * cache line.
 168          */
 169         prefetch(priv);
 170
 171         /*
 172          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
 173          * completely remove "qos" in the event neither interface
 174          * supports multiple queues per port.
 175          */
 176         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
 177             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
 178                 qos = GET_SKBUFF_QOS(skb);
 179                 if (qos <= 0)
 180                         qos = 0;
 181                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
 182                         qos = 0;
 183         } else
 184                 qos = 0;
 185
 186         if (USE_ASYNC_IOBDMA) {
 187                 /* Save scratch in case userspace is using it */
 188                 CVMX_SYNCIOBDMA;
 189                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 190                 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 191
 192                 /*
 193                  * Fetch and increment the number of packets to be
 194                  * freed.
 195                  */
 196                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
 197                                                FAU_NUM_PACKET_BUFFERS_TO_FREE,
 198                                                0);
 199                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
 200                                                priv->fau + qos * 4,
 201                                                MAX_SKB_TO_FREE);
 202         }
 203
 204         /*
 205          * We have space for 6 segment pointers, If there will be more
 206          * than that, we must linearize.
 207          */
 208         if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
 209                 if (unlikely(__skb_linearize(skb))) {
 210                         queue_type = QUEUE_DROP;
 211                         if (USE_ASYNC_IOBDMA) {
 212                                 /* Get the number of skbuffs in use by the hardware */
 213                                 CVMX_SYNCIOBDMA;
 214                                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 215                         } else {
 216                                 /* Get the number of skbuffs in use by the hardware */
 217                                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 218                                                                        MAX_SKB_TO_FREE);
 219                         }
 220                         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau + qos * 4);
 221                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 222                         goto skip_xmit;
 223                 }
 224         }
 225
 226         /*
 227          * The CN3XXX series of parts has an errata (GMX-401) which
 228          * causes the GMX block to hang if a collision occurs towards
 229          * the end of a <68 byte packet. As a workaround for this, we
 230          * pad packets to be 68 bytes whenever we are in half duplex
 231          * mode. We don't handle the case of having a small packet but
 232          * no room to add the padding.  The kernel should always give
 233          * us at least a cache line
 234          */
 235         if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
 236                 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
 237                 int interface = INTERFACE(priv->port);
 238                 int index = INDEX(priv->port);
 239
 240                 if (interface < 2) {
 241                         /* We only need to pad packet in half duplex mode */
 242                         gmx_prt_cfg.u64 =
 243                             cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
 244                         if (gmx_prt_cfg.s.duplex == 0) {
 245                                 int add_bytes = 64 - skb->len;
 246                                 if ((skb_tail_pointer(skb) + add_bytes) <=
 247                                     skb_end_pointer(skb))
 248                                         memset(__skb_put(skb, add_bytes), 0,
 249                                                add_bytes);
 250                         }
 251                 }
 252         }
 253
 254         /* Build the PKO command */
 255         pko_command.u64 = 0;
 256         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
 257         pko_command.s.segs = 1;
 258         pko_command.s.total_bytes = skb->len;
 259         pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
 260         pko_command.s.subone0 = 1;
 261
 262         pko_command.s.dontfree = 1;
 263
 264         /* Build the PKO buffer pointer */
 265         hw_buffer.u64 = 0;
 266         if (skb_shinfo(skb)->nr_frags == 0) {
 267                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
 268                 hw_buffer.s.pool = 0;
 269                 hw_buffer.s.size = skb->len;
 270         } else {
 271                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
 272                 hw_buffer.s.pool = 0;
 273                 hw_buffer.s.size = skb_headlen(skb);
 274                 CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
 275                 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 276                         struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i;
 277                         hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)(page_address(fs->page) + fs->page_offset));
 278                         hw_buffer.s.size = fs->size;
 279                         CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
 280                 }
 281                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb));
 282                 hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1;
 283                 pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1;
 284                 pko_command.s.gather = 1;
 285                 goto dont_put_skbuff_in_hw;
 286         }
 287
 288         /*
 289          * See if we can put this skb in the FPA pool. Any strange
 290          * behavior from the Linux networking stack will most likely
 291          * be caused by a bug in the following code. If some field is
 292          * in use by the network stack and get carried over when a
 293          * buffer is reused, bad thing may happen.  If in doubt and
 294          * you dont need the absolute best performance, disable the
 295          * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
 296          * shown a 25% increase in performance under some loads.
 297          */
 298 #if REUSE_SKBUFFS_WITHOUT_FREE
 299         fpa_head = skb->head + 256 - ((unsigned long)skb->head & 0x7f);
 300         if (unlikely(skb->data < fpa_head)) {
 301                 /*
 302                  * printk("TX buffer beginning can't meet FPA
 303                  * alignment constraints\n");
 304                  */
 305                 goto dont_put_skbuff_in_hw;
 306         }
 307         if (unlikely
 308             ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
 309                 /*
 310                    printk("TX buffer isn't large enough for the FPA\n");
 311                  */
 312                 goto dont_put_skbuff_in_hw;
 313         }
 314         if (unlikely(skb_shared(skb))) {
 315                 /*
 316                    printk("TX buffer sharing data with someone else\n");
 317                  */
 318                 goto dont_put_skbuff_in_hw;
 319         }
 320         if (unlikely(skb_cloned(skb))) {
 321                 /*
 322                    printk("TX buffer has been cloned\n");
 323                  */
 324                 goto dont_put_skbuff_in_hw;
 325         }
 326         if (unlikely(skb_header_cloned(skb))) {
 327                 /*
 328                    printk("TX buffer header has been cloned\n");
 329                  */
 330                 goto dont_put_skbuff_in_hw;
 331         }
 332         if (unlikely(skb->destructor)) {
 333                 /*
 334                    printk("TX buffer has a destructor\n");
 335                  */
 336                 goto dont_put_skbuff_in_hw;
 337         }
 338         if (unlikely(skb_shinfo(skb)->nr_frags)) {
 339                 /*
 340                    printk("TX buffer has fragments\n");
 341                  */
 342                 goto dont_put_skbuff_in_hw;
 343         }
 344         if (unlikely
 345             (skb->truesize !=
 346              sizeof(*skb) + skb_end_pointer(skb) - skb->head)) {
 347                 /*
 348                    printk("TX buffer truesize has been changed\n");
 349                  */
 350                 goto dont_put_skbuff_in_hw;
 351         }
 352
 353         /*
 354          * We can use this buffer in the FPA.  We don't need the FAU
 355          * update anymore
 356          */
 357         pko_command.s.dontfree = 0;
 358
 359         hw_buffer.s.back = ((unsigned long)skb->data >> 7) - ((unsigned long)fpa_head >> 7);
 360         *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
 361
 362         /*
 363          * The skbuff will be reused without ever being freed. We must
 364          * cleanup a bunch of core things.
 365          */
 366         dst_release(skb_dst(skb));
 367         skb_dst_set(skb, NULL);
 368 #ifdef CONFIG_XFRM
 369         secpath_put(skb->sp);
 370         skb->sp = NULL;
 371 #endif
 372         nf_reset(skb);
 373
 374 #ifdef CONFIG_NET_SCHED
 375         skb->tc_index = 0;
 376 #ifdef CONFIG_NET_CLS_ACT
 377         skb->tc_verd = 0;
 378 #endif /* CONFIG_NET_CLS_ACT */
 379 #endif /* CONFIG_NET_SCHED */
 380 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
 381
 382 dont_put_skbuff_in_hw:
 383
 384         /* Check if we can use the hardware checksumming */
 385         if (USE_HW_TCPUDP_CHECKSUM && (skb->protocol == htons(ETH_P_IP)) &&
 386             (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) &&
 387             ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == 1 << 14))
 388             && ((ip_hdr(skb)->protocol == IPPROTO_TCP)
 389                 || (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
 390                 /* Use hardware checksum calc */
 391                 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
 392         }
 393
 394         if (USE_ASYNC_IOBDMA) {
 395                 /* Get the number of skbuffs in use by the hardware */
 396                 CVMX_SYNCIOBDMA;
 397                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 398                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 399         } else {
 400                 /* Get the number of skbuffs in use by the hardware */
 401                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 402                                                        MAX_SKB_TO_FREE);
 403                 buffers_to_free =
 404                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 405         }
 406
 407         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4);
 408
 409         /*
 410          * If we're sending faster than the receive can free them then
 411          * don't do the HW free.
 412          */
 413         if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 414                 pko_command.s.dontfree = 1;
 415
 416         if (pko_command.s.dontfree) {
 417                 queue_type = QUEUE_CORE;
 418                 pko_command.s.reg0 = priv->fau+qos*4;
 419         } else {
 420                 queue_type = QUEUE_HW;
 421         }
 422         if (USE_ASYNC_IOBDMA)
 423                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
 424
 425         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 426
 427         /* Drop this packet if we have too many already queued to the HW */
 428         if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >= MAX_OUT_QUEUE_DEPTH)) {
 429                 if (dev->tx_queue_len != 0) {
 430                         /* Drop the lock when notifying the core.  */
 431                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 432                         netif_stop_queue(dev);
 433                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 434                 } else {
 435                         /* If not using normal queueing.  */
 436                         queue_type = QUEUE_DROP;
 437                         goto skip_xmit;
 438                 }
 439         }
 440
 441         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
 442                                      CVMX_PKO_LOCK_NONE);
 443
 444         /* Send the packet to the output queue */
 445         if (unlikely(cvmx_pko_send_packet_finish(priv->port,
 446                                                  priv->queue + qos,
 447                                                  pko_command, hw_buffer,
 448                                                  CVMX_PKO_LOCK_NONE))) {
 449                 DEBUGPRINT("%s: Failed to send the packet\n", dev->name);
 450                 queue_type = QUEUE_DROP;
 451         }
 452 skip_xmit:
 453         to_free_list = NULL;
 454
 455         switch (queue_type) {
 456         case QUEUE_DROP:
 457                 skb->next = to_free_list;
 458                 to_free_list = skb;
 459                 priv->stats.tx_dropped++;
 460                 break;
 461         case QUEUE_HW:
 462                 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
 463                 break;
 464         case QUEUE_CORE:
 465                 __skb_queue_tail(&priv->tx_free_list[qos], skb);
 466                 break;
 467         default:
 468                 BUG();
 469         }
 470
 471         while (skb_to_free > 0) {
 472                 struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
 473                 t->next = to_free_list;
 474                 to_free_list = t;
 475                 skb_to_free--;
 476         }
 477
 478         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 479
 480         /* Do the actual freeing outside of the lock. */
 481         while (to_free_list) {
 482                 struct sk_buff *t = to_free_list;
 483                 to_free_list = to_free_list->next;
 484                 dev_kfree_skb_any(t);
 485         }
 486
 487         if (USE_ASYNC_IOBDMA) {
 488                 CVMX_SYNCIOBDMA;
 489                 total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 490                 /* Restore the scratch area */
 491                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 492                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 493         } else {
 494                 total_to_clean = cvmx_fau_fetch_and_add32(FAU_TOTAL_TX_TO_CLEAN, 1);
 495         }
 496
 497         if (total_to_clean & 0x3ff) {
 498                 /*
 499                  * Schedule the cleanup tasklet every 1024 packets for
 500                  * the pathological case of high traffic on one port
 501                  * delaying clean up of packets on a different port
 502                  * that is blocked waiting for the cleanup.
 503                  */
 504                 tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 505         }
 506
 507         cvm_oct_kick_tx_poll_watchdog();
 508
 509         return NETDEV_TX_OK;
 510 }
 511
 512 /**
 513  * cvm_oct_xmit_pow - transmit a packet to the POW
 514  * @skb:    Packet to send
 515  * @dev:    Device info structure
 516
 517  * Returns Always returns zero
 518  */
 519 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
 520 {
 521         struct octeon_ethernet *priv = netdev_priv(dev);
 522         void *packet_buffer;
 523         void *copy_location;
 524
 525         /* Get a work queue entry */
 526         cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
 527         if (unlikely(work == NULL)) {
 528                 DEBUGPRINT("%s: Failed to allocate a work queue entry\n",
 529                            dev->name);
 530                 priv->stats.tx_dropped++;
 531                 dev_kfree_skb(skb);
 532                 return 0;
 533         }
 534
 535         /* Get a packet buffer */
 536         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
 537         if (unlikely(packet_buffer == NULL)) {
 538                 DEBUGPRINT("%s: Failed to allocate a packet buffer\n",
 539                            dev->name);
 540                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
 541                 priv->stats.tx_dropped++;
 542                 dev_kfree_skb(skb);
 543                 return 0;
 544         }
 545
 546         /*
 547          * Calculate where we need to copy the data to. We need to
 548          * leave 8 bytes for a next pointer (unused). We also need to
 549          * include any configure skip. Then we need to align the IP
 550          * packet src and dest into the same 64bit word. The below
 551          * calculation may add a little extra, but that doesn't
 552          * hurt.
 553          */
 554         copy_location = packet_buffer + sizeof(uint64_t);
 555         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
 556
 557         /*
 558          * We have to copy the packet since whoever processes this
 559          * packet will free it to a hardware pool. We can't use the
 560          * trick of counting outstanding packets like in
 561          * cvm_oct_xmit.
 562          */
 563         memcpy(copy_location, skb->data, skb->len);
 564
 565         /*
 566          * Fill in some of the work queue fields. We may need to add
 567          * more if the software at the other end needs them.
 568          */
 569         work->hw_chksum = skb->csum;
 570         work->len = skb->len;
 571         work->ipprt = priv->port;
 572         work->qos = priv->port & 0x7;
 573         work->grp = pow_send_group;
 574         work->tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
 575         work->tag = pow_send_group;     /* FIXME */
 576         /* Default to zero. Sets of zero later are commented out */
 577         work->word2.u64 = 0;
 578         work->word2.s.bufs = 1;
 579         work->packet_ptr.u64 = 0;
 580         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
 581         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
 582         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
 583         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
 584
 585         if (skb->protocol == htons(ETH_P_IP)) {
 586                 work->word2.s.ip_offset = 14;
 587 #if 0
 588                 work->word2.s.vlan_valid = 0;   /* FIXME */
 589                 work->word2.s.vlan_cfi = 0;     /* FIXME */
 590                 work->word2.s.vlan_id = 0;      /* FIXME */
 591                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
 592 #endif
 593                 work->word2.s.tcp_or_udp =
 594                     (ip_hdr(skb)->protocol == IPPROTO_TCP)
 595                     || (ip_hdr(skb)->protocol == IPPROTO_UDP);
 596 #if 0
 597                 /* FIXME */
 598                 work->word2.s.dec_ipsec = 0;
 599                 /* We only support IPv4 right now */
 600                 work->word2.s.is_v6 = 0;
 601                 /* Hardware would set to zero */
 602                 work->word2.s.software = 0;
 603                 /* No error, packet is internal */
 604                 work->word2.s.L4_error = 0;
 605 #endif
 606                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0)
 607                                           || (ip_hdr(skb)->frag_off ==
 608                                               1 << 14));
 609 #if 0
 610                 /* Assume Linux is sending a good packet */
 611                 work->word2.s.IP_exc = 0;
 612 #endif
 613                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
 614                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
 615 #if 0
 616                 /* This is an IP packet */
 617                 work->word2.s.not_IP = 0;
 618                 /* No error, packet is internal */
 619                 work->word2.s.rcv_error = 0;
 620                 /* No error, packet is internal */
 621                 work->word2.s.err_code = 0;
 622 #endif
 623
 624                 /*
 625                  * When copying the data, include 4 bytes of the
 626                  * ethernet header to align the same way hardware
 627                  * does.
 628                  */
 629                 memcpy(work->packet_data, skb->data + 10,
 630                        sizeof(work->packet_data));
 631         } else {
 632 #if 0
 633                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
 634                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
 635                 work->word2.snoip.vlan_id = 0;  /* FIXME */
 636                 work->word2.snoip.software = 0; /* Hardware would set to zero */
 637 #endif
 638                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
 639                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
 640                 work->word2.snoip.is_bcast =
 641                     (skb->pkt_type == PACKET_BROADCAST);
 642                 work->word2.snoip.is_mcast =
 643                     (skb->pkt_type == PACKET_MULTICAST);
 644                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
 645 #if 0
 646                 /* No error, packet is internal */
 647                 work->word2.snoip.rcv_error = 0;
 648                 /* No error, packet is internal */
 649                 work->word2.snoip.err_code = 0;
 650 #endif
 651                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
 652         }
 653
 654         /* Submit the packet to the POW */
 655         cvmx_pow_work_submit(work, work->tag, work->tag_type, work->qos,
 656                              work->grp);
 657         priv->stats.tx_packets++;
 658         priv->stats.tx_bytes += skb->len;
 659         dev_kfree_skb(skb);
 660         return 0;
 661 }
 662
 663 /**
 664  * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
 665  * @dev:    Device being shutdown
 666  *
 667  */
 668 void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 669 {
 670         struct octeon_ethernet *priv = netdev_priv(dev);
 671         unsigned long flags;
 672         int qos;
 673
 674         for (qos = 0; qos < 16; qos++) {
 675                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 676                 while (skb_queue_len(&priv->tx_free_list[qos]))
 677                         dev_kfree_skb_any(__skb_dequeue
 678                                           (&priv->tx_free_list[qos]));
 679                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 680         }
 681 }
 682
 683 static void cvm_oct_tx_do_cleanup(unsigned long arg)
 684 {
 685         int port;
 686
 687         for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
 688                 if (cvm_oct_device[port]) {
 689                         struct net_device *dev = cvm_oct_device[port];
 690                         cvm_oct_free_tx_skbs(dev);
 691                 }
 692         }
 693 }
 694
 695 static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
 696 {
 697         /* Disable the interrupt.  */
 698         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 699         /* Do the work in the tasklet.  */
 700         tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 701         return IRQ_HANDLED;
 702 }
 703
 704 void cvm_oct_tx_initialize(void)
 705 {
 706         int i;
 707
 708         /* Disable the interrupt.  */
 709         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 710         /* Register an IRQ hander for to receive CIU_TIMX(1) interrupts */
 711         i = request_irq(OCTEON_IRQ_TIMER1,
 712                         cvm_oct_tx_cleanup_watchdog, 0,
 713                         "Ethernet", cvm_oct_device);
 714
 715         if (i)
 716                 panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
 717 }
 718
 719 void cvm_oct_tx_shutdown(void)
 720 {
 721         /* Free the interrupt handler */
 722         free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
 723 }