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 int cvm_oct_adjust_skb_to_free(int skb_to_free, int fau)
  62 {
  63         int 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         int 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                 total_freed += skb_to_free;
 103                 if (skb_to_free > 0) {
 104                         struct sk_buff *to_free_list = NULL;
 105
 106                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 107                         while (skb_to_free > 0) {
 108                                 struct sk_buff *t;
 109
 110                                 t = __skb_dequeue(&priv->tx_free_list[qos]);
 111                                 t->next = to_free_list;
 112                                 to_free_list = t;
 113                                 skb_to_free--;
 114                         }
 115                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 116                                                flags);
 117                         /* Do the actual freeing outside of the lock. */
 118                         while (to_free_list) {
 119                                 struct sk_buff *t = to_free_list;
 120
 121                                 to_free_list = to_free_list->next;
 122                                 dev_kfree_skb_any(t);
 123                         }
 124                 }
 125                 total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 126         }
 127         if (total_remaining < MAX_OUT_QUEUE_DEPTH && netif_queue_stopped(dev))
 128                 netif_wake_queue(dev);
 129         if (total_remaining)
 130                 cvm_oct_kick_tx_poll_watchdog();
 131 }
 132
 133 /**
 134  * cvm_oct_xmit - transmit a packet
 135  * @skb:    Packet to send
 136  * @dev:    Device info structure
 137  *
 138  * Returns Always returns NETDEV_TX_OK
 139  */
 140 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 141 {
 142         cvmx_pko_command_word0_t pko_command;
 143         union cvmx_buf_ptr hw_buffer;
 144         u64 old_scratch;
 145         u64 old_scratch2;
 146         int qos;
 147         int i;
 148         enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
 149         struct octeon_ethernet *priv = netdev_priv(dev);
 150         struct sk_buff *to_free_list;
 151         int skb_to_free;
 152         int buffers_to_free;
 153         u32 total_to_clean;
 154         unsigned long flags;
 155 #if REUSE_SKBUFFS_WITHOUT_FREE
 156         unsigned char *fpa_head;
 157 #endif
 158
 159         /*
 160          * Prefetch the private data structure.  It is larger than the
 161          * one cache line.
 162          */
 163         prefetch(priv);
 164
 165         /*
 166          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
 167          * completely remove "qos" in the event neither interface
 168          * supports multiple queues per port.
 169          */
 170         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
 171             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
 172                 qos = GET_SKBUFF_QOS(skb);
 173                 if (qos <= 0)
 174                         qos = 0;
 175                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
 176                         qos = 0;
 177         } else {
 178                 qos = 0;
 179         }
 180
 181         if (USE_ASYNC_IOBDMA) {
 182                 /* Save scratch in case userspace is using it */
 183                 CVMX_SYNCIOBDMA;
 184                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 185                 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 186
 187                 /*
 188                  * Fetch and increment the number of packets to be
 189                  * freed.
 190                  */
 191                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
 192                                                FAU_NUM_PACKET_BUFFERS_TO_FREE,
 193                                                0);
 194                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
 195                                                priv->fau + qos * 4,
 196                                                MAX_SKB_TO_FREE);
 197         }
 198
 199         /*
 200          * We have space for 6 segment pointers, If there will be more
 201          * than that, we must linearize.
 202          */
 203         if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
 204                 if (unlikely(__skb_linearize(skb))) {
 205                         queue_type = QUEUE_DROP;
 206                         if (USE_ASYNC_IOBDMA) {
 207                                 /*
 208                                  * Get the number of skbuffs in use
 209                                  * by the hardware
 210                                  */
 211                                 CVMX_SYNCIOBDMA;
 212                                 skb_to_free =
 213                                         cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 214                         } else {
 215                                 /*
 216                                  * Get the number of skbuffs in use
 217                                  * by the hardware
 218                                  */
 219                                 skb_to_free = cvmx_fau_fetch_and_add32(
 220                                         priv->fau + qos * 4, MAX_SKB_TO_FREE);
 221                         }
 222                         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 223                                                                  priv->fau +
 224                                                                  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                 /* 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 #ifdef CONFIG_NET_CLS_ACT
 373         skb->tc_verd = 0;
 374 #endif /* CONFIG_NET_CLS_ACT */
 375 #endif /* CONFIG_NET_SCHED */
 376 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
 377
 378 dont_put_skbuff_in_hw:
 379
 380         /* Check if we can use the hardware checksumming */
 381         if ((skb->protocol == htons(ETH_P_IP)) &&
 382             (ip_hdr(skb)->version == 4) &&
 383             (ip_hdr(skb)->ihl == 5) &&
 384             ((ip_hdr(skb)->frag_off == 0) ||
 385              (ip_hdr(skb)->frag_off == htons(1 << 14))) &&
 386             ((ip_hdr(skb)->protocol == IPPROTO_TCP) ||
 387              (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
 388                 /* Use hardware checksum calc */
 389                 pko_command.s.ipoffp1 = skb_network_offset(skb) + 1;
 390         }
 391
 392         if (USE_ASYNC_IOBDMA) {
 393                 /* Get the number of skbuffs in use by the hardware */
 394                 CVMX_SYNCIOBDMA;
 395                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 396                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 397         } else {
 398                 /* Get the number of skbuffs in use by the hardware */
 399                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 400                                                        MAX_SKB_TO_FREE);
 401                 buffers_to_free =
 402                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 403         }
 404
 405         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 406                                                  priv->fau + qos * 4);
 407
 408         /*
 409          * If we're sending faster than the receive can free them then
 410          * don't do the HW free.
 411          */
 412         if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 413                 pko_command.s.dontfree = 1;
 414
 415         if (pko_command.s.dontfree) {
 416                 queue_type = QUEUE_CORE;
 417                 pko_command.s.reg0 = priv->fau + qos * 4;
 418         } else {
 419                 queue_type = QUEUE_HW;
 420         }
 421         if (USE_ASYNC_IOBDMA)
 422                 cvmx_fau_async_fetch_and_add32(
 423                                 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]) >=
 429                      MAX_OUT_QUEUE_DEPTH)) {
 430                 if (dev->tx_queue_len != 0) {
 431                         /* Drop the lock when notifying the core.  */
 432                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 433                                                flags);
 434                         netif_stop_queue(dev);
 435                         spin_lock_irqsave(&priv->tx_free_list[qos].lock,
 436                                           flags);
 437                 } else {
 438                         /* If not using normal queueing.  */
 439                         queue_type = QUEUE_DROP;
 440                         goto skip_xmit;
 441                 }
 442         }
 443
 444         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
 445                                      CVMX_PKO_LOCK_NONE);
 446
 447         /* Send the packet to the output queue */
 448         if (unlikely(cvmx_pko_send_packet_finish(priv->port,
 449                                                  priv->queue + qos,
 450                                                  pko_command, hw_buffer,
 451                                                  CVMX_PKO_LOCK_NONE))) {
 452                 printk_ratelimited("%s: Failed to send the packet\n",
 453                                    dev->name);
 454                 queue_type = QUEUE_DROP;
 455         }
 456 skip_xmit:
 457         to_free_list = NULL;
 458
 459         switch (queue_type) {
 460         case QUEUE_DROP:
 461                 skb->next = to_free_list;
 462                 to_free_list = skb;
 463                 priv->stats.tx_dropped++;
 464                 break;
 465         case QUEUE_HW:
 466                 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
 467                 break;
 468         case QUEUE_CORE:
 469                 __skb_queue_tail(&priv->tx_free_list[qos], skb);
 470                 break;
 471         default:
 472                 BUG();
 473         }
 474
 475         while (skb_to_free > 0) {
 476                 struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
 477
 478                 t->next = to_free_list;
 479                 to_free_list = t;
 480                 skb_to_free--;
 481         }
 482
 483         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 484
 485         /* Do the actual freeing outside of the lock. */
 486         while (to_free_list) {
 487                 struct sk_buff *t = to_free_list;
 488
 489                 to_free_list = to_free_list->next;
 490                 dev_kfree_skb_any(t);
 491         }
 492
 493         if (USE_ASYNC_IOBDMA) {
 494                 CVMX_SYNCIOBDMA;
 495                 total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 496                 /* Restore the scratch area */
 497                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 498                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 499         } else {
 500                 total_to_clean = cvmx_fau_fetch_and_add32(
 501                                                 FAU_TOTAL_TX_TO_CLEAN, 1);
 502         }
 503
 504         if (total_to_clean & 0x3ff) {
 505                 /*
 506                  * Schedule the cleanup tasklet every 1024 packets for
 507                  * the pathological case of high traffic on one port
 508                  * delaying clean up of packets on a different port
 509                  * that is blocked waiting for the cleanup.
 510                  */
 511                 tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 512         }
 513
 514         cvm_oct_kick_tx_poll_watchdog();
 515
 516         return NETDEV_TX_OK;
 517 }
 518
 519 /**
 520  * cvm_oct_xmit_pow - transmit a packet to the POW
 521  * @skb:    Packet to send
 522  * @dev:    Device info structure
 523
 524  * Returns Always returns zero
 525  */
 526 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
 527 {
 528         struct octeon_ethernet *priv = netdev_priv(dev);
 529         void *packet_buffer;
 530         void *copy_location;
 531
 532         /* Get a work queue entry */
 533         cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
 534
 535         if (unlikely(!work)) {
 536                 printk_ratelimited("%s: Failed to allocate a work queue entry\n",
 537                                    dev->name);
 538                 priv->stats.tx_dropped++;
 539                 dev_kfree_skb_any(skb);
 540                 return 0;
 541         }
 542
 543         /* Get a packet buffer */
 544         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
 545         if (unlikely(!packet_buffer)) {
 546                 printk_ratelimited("%s: Failed to allocate a packet buffer\n",
 547                                    dev->name);
 548                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 549                 priv->stats.tx_dropped++;
 550                 dev_kfree_skb_any(skb);
 551                 return 0;
 552         }
 553
 554         /*
 555          * Calculate where we need to copy the data to. We need to
 556          * leave 8 bytes for a next pointer (unused). We also need to
 557          * include any configure skip. Then we need to align the IP
 558          * packet src and dest into the same 64bit word. The below
 559          * calculation may add a little extra, but that doesn't
 560          * hurt.
 561          */
 562         copy_location = packet_buffer + sizeof(u64);
 563         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
 564
 565         /*
 566          * We have to copy the packet since whoever processes this
 567          * packet will free it to a hardware pool. We can't use the
 568          * trick of counting outstanding packets like in
 569          * cvm_oct_xmit.
 570          */
 571         memcpy(copy_location, skb->data, skb->len);
 572
 573         /*
 574          * Fill in some of the work queue fields. We may need to add
 575          * more if the software at the other end needs them.
 576          */
 577         if (!OCTEON_IS_MODEL(OCTEON_CN68XX))
 578                 work->word0.pip.cn38xx.hw_chksum = skb->csum;
 579         work->word1.len = skb->len;
 580         cvmx_wqe_set_port(work, priv->port);
 581         cvmx_wqe_set_qos(work, priv->port & 0x7);
 582         cvmx_wqe_set_grp(work, pow_send_group);
 583         work->word1.tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
 584         work->word1.tag = pow_send_group;       /* FIXME */
 585         /* Default to zero. Sets of zero later are commented out */
 586         work->word2.u64 = 0;
 587         work->word2.s.bufs = 1;
 588         work->packet_ptr.u64 = 0;
 589         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
 590         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
 591         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
 592         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
 593
 594         if (skb->protocol == htons(ETH_P_IP)) {
 595                 work->word2.s.ip_offset = 14;
 596 #if 0
 597                 work->word2.s.vlan_valid = 0;   /* FIXME */
 598                 work->word2.s.vlan_cfi = 0;     /* FIXME */
 599                 work->word2.s.vlan_id = 0;      /* FIXME */
 600                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
 601 #endif
 602                 work->word2.s.tcp_or_udp =
 603                     (ip_hdr(skb)->protocol == IPPROTO_TCP) ||
 604                     (ip_hdr(skb)->protocol == IPPROTO_UDP);
 605 #if 0
 606                 /* FIXME */
 607                 work->word2.s.dec_ipsec = 0;
 608                 /* We only support IPv4 right now */
 609                 work->word2.s.is_v6 = 0;
 610                 /* Hardware would set to zero */
 611                 work->word2.s.software = 0;
 612                 /* No error, packet is internal */
 613                 work->word2.s.L4_error = 0;
 614 #endif
 615                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0) ||
 616                                           (ip_hdr(skb)->frag_off ==
 617                                               1 << 14));
 618 #if 0
 619                 /* Assume Linux is sending a good packet */
 620                 work->word2.s.IP_exc = 0;
 621 #endif
 622                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
 623                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
 624 #if 0
 625                 /* This is an IP packet */
 626                 work->word2.s.not_IP = 0;
 627                 /* No error, packet is internal */
 628                 work->word2.s.rcv_error = 0;
 629                 /* No error, packet is internal */
 630                 work->word2.s.err_code = 0;
 631 #endif
 632
 633                 /*
 634                  * When copying the data, include 4 bytes of the
 635                  * ethernet header to align the same way hardware
 636                  * does.
 637                  */
 638                 memcpy(work->packet_data, skb->data + 10,
 639                        sizeof(work->packet_data));
 640         } else {
 641 #if 0
 642                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
 643                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
 644                 work->word2.snoip.vlan_id = 0;  /* FIXME */
 645                 work->word2.snoip.software = 0; /* Hardware would set to zero */
 646 #endif
 647                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
 648                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
 649                 work->word2.snoip.is_bcast =
 650                     (skb->pkt_type == PACKET_BROADCAST);
 651                 work->word2.snoip.is_mcast =
 652                     (skb->pkt_type == PACKET_MULTICAST);
 653                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
 654 #if 0
 655                 /* No error, packet is internal */
 656                 work->word2.snoip.rcv_error = 0;
 657                 /* No error, packet is internal */
 658                 work->word2.snoip.err_code = 0;
 659 #endif
 660                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
 661         }
 662
 663         /* Submit the packet to the POW */
 664         cvmx_pow_work_submit(work, work->word1.tag, work->word1.tag_type,
 665                              cvmx_wqe_get_qos(work), cvmx_wqe_get_grp(work));
 666         priv->stats.tx_packets++;
 667         priv->stats.tx_bytes += skb->len;
 668         dev_consume_skb_any(skb);
 669         return 0;
 670 }
 671
 672 /**
 673  * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
 674  * @dev:    Device being shutdown
 675  *
 676  */
 677 void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 678 {
 679         struct octeon_ethernet *priv = netdev_priv(dev);
 680         unsigned long flags;
 681         int qos;
 682
 683         for (qos = 0; qos < 16; qos++) {
 684                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 685                 while (skb_queue_len(&priv->tx_free_list[qos]))
 686                         dev_kfree_skb_any(__skb_dequeue
 687                                           (&priv->tx_free_list[qos]));
 688                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 689         }
 690 }
 691
 692 static void cvm_oct_tx_do_cleanup(unsigned long arg)
 693 {
 694         int port;
 695
 696         for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
 697                 if (cvm_oct_device[port]) {
 698                         struct net_device *dev = cvm_oct_device[port];
 699
 700                         cvm_oct_free_tx_skbs(dev);
 701                 }
 702         }
 703 }
 704
 705 static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
 706 {
 707         /* Disable the interrupt.  */
 708         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 709         /* Do the work in the tasklet.  */
 710         tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 711         return IRQ_HANDLED;
 712 }
 713
 714 void cvm_oct_tx_initialize(void)
 715 {
 716         int i;
 717
 718         /* Disable the interrupt.  */
 719         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 720         /* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
 721         i = request_irq(OCTEON_IRQ_TIMER1,
 722                         cvm_oct_tx_cleanup_watchdog, 0,
 723                         "Ethernet", cvm_oct_device);
 724
 725         if (i)
 726                 panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
 727 }
 728
 729 void cvm_oct_tx_shutdown(void)
 730 {
 731         /* Free the interrupt handler */
 732         free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
 733 }