drivers/staging/octeon/ethernet-tx.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This file is based on code from OCTEON SDK by Cavium Networks.
   4  *
   5  * Copyright (c) 2003-2010 Cavium Networks
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/kernel.h>
  10 #include <linux/netdevice.h>
  11 #include <linux/etherdevice.h>
  12 #include <linux/ip.h>
  13 #include <linux/ratelimit.h>
  14 #include <linux/string.h>
  15 #include <linux/interrupt.h>
  16 #include <net/dst.h>
  17 #ifdef CONFIG_XFRM
  18 #include <linux/xfrm.h>
  19 #include <net/xfrm.h>
  20 #endif /* CONFIG_XFRM */
  21
  22 #include <linux/atomic.h>
  23 #include <net/sch_generic.h>
  24
  25 #include "octeon-ethernet.h"
  26 #include "ethernet-defines.h"
  27 #include "ethernet-tx.h"
  28 #include "ethernet-util.h"
  29
  30 #define CVM_OCT_SKB_CB(skb)     ((u64 *)((skb)->cb))
  31
  32 /*
  33  * You can define GET_SKBUFF_QOS() to override how the skbuff output
  34  * function determines which output queue is used. The default
  35  * implementation always uses the base queue for the port. If, for
  36  * example, you wanted to use the skb->priority field, define
  37  * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
  38  */
  39 #ifndef GET_SKBUFF_QOS
  40 #define GET_SKBUFF_QOS(skb) 0
  41 #endif
  42
  43 static void cvm_oct_tx_do_cleanup(unsigned long arg);
  44 static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
  45
  46 /* Maximum number of SKBs to try to free per xmit packet. */
  47 #define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
  48
  49 static inline int cvm_oct_adjust_skb_to_free(int skb_to_free, int fau)
  50 {
  51         int undo;
  52
  53         undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free +
  54                                                    MAX_SKB_TO_FREE;
  55         if (undo > 0)
  56                 cvmx_fau_atomic_add32(fau, -undo);
  57         skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE :
  58                                                        -skb_to_free;
  59         return skb_to_free;
  60 }
  61
  62 static void cvm_oct_kick_tx_poll_watchdog(void)
  63 {
  64         union cvmx_ciu_timx ciu_timx;
  65
  66         ciu_timx.u64 = 0;
  67         ciu_timx.s.one_shot = 1;
  68         ciu_timx.s.len = cvm_oct_tx_poll_interval;
  69         cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
  70 }
  71
  72 static void cvm_oct_free_tx_skbs(struct net_device *dev)
  73 {
  74         int skb_to_free;
  75         int qos, queues_per_port;
  76         int total_freed = 0;
  77         int total_remaining = 0;
  78         unsigned long flags;
  79         struct octeon_ethernet *priv = netdev_priv(dev);
  80
  81         queues_per_port = cvmx_pko_get_num_queues(priv->port);
  82         /* Drain any pending packets in the free list */
  83         for (qos = 0; qos < queues_per_port; qos++) {
  84                 if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
  85                         continue;
  86                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
  87                                                        MAX_SKB_TO_FREE);
  88                 skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
  89                                                          priv->fau + qos * 4);
  90                 total_freed += skb_to_free;
  91                 if (skb_to_free > 0) {
  92                         struct sk_buff *to_free_list = NULL;
  93
  94                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
  95                         while (skb_to_free > 0) {
  96                                 struct sk_buff *t;
  97
  98                                 t = __skb_dequeue(&priv->tx_free_list[qos]);
  99                                 t->next = to_free_list;
 100                                 to_free_list = t;
 101                                 skb_to_free--;
 102                         }
 103                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 104                                                flags);
 105                         /* Do the actual freeing outside of the lock. */
 106                         while (to_free_list) {
 107                                 struct sk_buff *t = to_free_list;
 108
 109                                 to_free_list = to_free_list->next;
 110                                 dev_kfree_skb_any(t);
 111                         }
 112                 }
 113                 total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 114         }
 115         if (total_remaining < MAX_OUT_QUEUE_DEPTH && netif_queue_stopped(dev))
 116                 netif_wake_queue(dev);
 117         if (total_remaining)
 118                 cvm_oct_kick_tx_poll_watchdog();
 119 }
 120
 121 /**
 122  * cvm_oct_xmit - transmit a packet
 123  * @skb:    Packet to send
 124  * @dev:    Device info structure
 125  *
 126  * Returns Always returns NETDEV_TX_OK
 127  */
 128 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 129 {
 130         union cvmx_pko_command_word0 pko_command;
 131         union cvmx_buf_ptr hw_buffer;
 132         u64 old_scratch;
 133         u64 old_scratch2;
 134         int qos;
 135         int i;
 136         enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
 137         struct octeon_ethernet *priv = netdev_priv(dev);
 138         struct sk_buff *to_free_list;
 139         int skb_to_free;
 140         int buffers_to_free;
 141         u32 total_to_clean;
 142         unsigned long flags;
 143 #if REUSE_SKBUFFS_WITHOUT_FREE
 144         unsigned char *fpa_head;
 145 #endif
 146
 147         /*
 148          * Prefetch the private data structure.  It is larger than the
 149          * one cache line.
 150          */
 151         prefetch(priv);
 152
 153         /*
 154          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
 155          * completely remove "qos" in the event neither interface
 156          * supports multiple queues per port.
 157          */
 158         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
 159             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
 160                 qos = GET_SKBUFF_QOS(skb);
 161                 if (qos <= 0)
 162                         qos = 0;
 163                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
 164                         qos = 0;
 165         } else {
 166                 qos = 0;
 167         }
 168
 169         if (USE_ASYNC_IOBDMA) {
 170                 /* Save scratch in case userspace is using it */
 171                 CVMX_SYNCIOBDMA;
 172                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 173                 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 174
 175                 /*
 176                  * Fetch and increment the number of packets to be
 177                  * freed.
 178                  */
 179                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
 180                                                FAU_NUM_PACKET_BUFFERS_TO_FREE,
 181                                                0);
 182                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
 183                                                priv->fau + qos * 4,
 184                                                MAX_SKB_TO_FREE);
 185         }
 186
 187         /*
 188          * We have space for 6 segment pointers, If there will be more
 189          * than that, we must linearize.
 190          */
 191         if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
 192                 if (unlikely(__skb_linearize(skb))) {
 193                         queue_type = QUEUE_DROP;
 194                         if (USE_ASYNC_IOBDMA) {
 195                                 /*
 196                                  * Get the number of skbuffs in use
 197                                  * by the hardware
 198                                  */
 199                                 CVMX_SYNCIOBDMA;
 200                                 skb_to_free =
 201                                         cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 202                         } else {
 203                                 /*
 204                                  * Get the number of skbuffs in use
 205                                  * by the hardware
 206                                  */
 207                                 skb_to_free =
 208                                      cvmx_fau_fetch_and_add32(priv->fau +
 209                                                               qos * 4,
 210                                                               MAX_SKB_TO_FREE);
 211                         }
 212                         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 213                                                                  priv->fau +
 214                                                                  qos * 4);
 215                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 216                         goto skip_xmit;
 217                 }
 218         }
 219
 220         /*
 221          * The CN3XXX series of parts has an errata (GMX-401) which
 222          * causes the GMX block to hang if a collision occurs towards
 223          * the end of a <68 byte packet. As a workaround for this, we
 224          * pad packets to be 68 bytes whenever we are in half duplex
 225          * mode. We don't handle the case of having a small packet but
 226          * no room to add the padding.  The kernel should always give
 227          * us at least a cache line
 228          */
 229         if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
 230                 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
 231                 int interface = INTERFACE(priv->port);
 232                 int index = INDEX(priv->port);
 233
 234                 if (interface < 2) {
 235                         /* We only need to pad packet in half duplex mode */
 236                         gmx_prt_cfg.u64 =
 237                             cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
 238                         if (gmx_prt_cfg.s.duplex == 0) {
 239                                 int add_bytes = 64 - skb->len;
 240
 241                                 if ((skb_tail_pointer(skb) + add_bytes) <=
 242                                     skb_end_pointer(skb))
 243                                         __skb_put_zero(skb, add_bytes);
 244                         }
 245                 }
 246         }
 247
 248         /* Build the PKO command */
 249         pko_command.u64 = 0;
 250 #ifdef __LITTLE_ENDIAN
 251         pko_command.s.le = 1;
 252 #endif
 253         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
 254         pko_command.s.segs = 1;
 255         pko_command.s.total_bytes = skb->len;
 256         pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
 257         pko_command.s.subone0 = 1;
 258
 259         pko_command.s.dontfree = 1;
 260
 261         /* Build the PKO buffer pointer */
 262         hw_buffer.u64 = 0;
 263         if (skb_shinfo(skb)->nr_frags == 0) {
 264                 hw_buffer.s.addr = XKPHYS_TO_PHYS((uintptr_t)skb->data);
 265                 hw_buffer.s.pool = 0;
 266                 hw_buffer.s.size = skb->len;
 267         } else {
 268                 hw_buffer.s.addr = XKPHYS_TO_PHYS((uintptr_t)skb->data);
 269                 hw_buffer.s.pool = 0;
 270                 hw_buffer.s.size = skb_headlen(skb);
 271                 CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
 272                 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 273                         skb_frag_t *fs = skb_shinfo(skb)->frags + i;
 274
 275                         hw_buffer.s.addr =
 276                                 XKPHYS_TO_PHYS((uintptr_t)skb_frag_address(fs));
 277                         hw_buffer.s.size = skb_frag_size(fs);
 278                         CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
 279                 }
 280                 hw_buffer.s.addr =
 281                         XKPHYS_TO_PHYS((uintptr_t)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 gets carried over when a
 293          * buffer is reused, bad things 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                 /* TX buffer beginning can't meet FPA alignment constraints */
 302                 goto dont_put_skbuff_in_hw;
 303         }
 304         if (unlikely
 305             ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
 306                 /* TX buffer isn't large enough for the FPA */
 307                 goto dont_put_skbuff_in_hw;
 308         }
 309         if (unlikely(skb_shared(skb))) {
 310                 /* TX buffer sharing data with someone else */
 311                 goto dont_put_skbuff_in_hw;
 312         }
 313         if (unlikely(skb_cloned(skb))) {
 314                 /* TX buffer has been cloned */
 315                 goto dont_put_skbuff_in_hw;
 316         }
 317         if (unlikely(skb_header_cloned(skb))) {
 318                 /* TX buffer header has been cloned */
 319                 goto dont_put_skbuff_in_hw;
 320         }
 321         if (unlikely(skb->destructor)) {
 322                 /* TX buffer has a destructor */
 323                 goto dont_put_skbuff_in_hw;
 324         }
 325         if (unlikely(skb_shinfo(skb)->nr_frags)) {
 326                 /* TX buffer has fragments */
 327                 goto dont_put_skbuff_in_hw;
 328         }
 329         if (unlikely
 330             (skb->truesize !=
 331              sizeof(*skb) + skb_end_offset(skb))) {
 332                 /* TX buffer truesize has been changed */
 333                 goto dont_put_skbuff_in_hw;
 334         }
 335
 336         /*
 337          * We can use this buffer in the FPA.  We don't need the FAU
 338          * update anymore
 339          */
 340         pko_command.s.dontfree = 0;
 341
 342         hw_buffer.s.back = ((unsigned long)skb->data >> 7) -
 343                            ((unsigned long)fpa_head >> 7);
 344
 345         *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
 346
 347         /*
 348          * The skbuff will be reused without ever being freed. We must
 349          * cleanup a bunch of core things.
 350          */
 351         dst_release(skb_dst(skb));
 352         skb_dst_set(skb, NULL);
 353         skb_ext_reset(skb);
 354         nf_reset_ct(skb);
 355         skb_reset_redirect(skb);
 356
 357 #ifdef CONFIG_NET_SCHED
 358         skb->tc_index = 0;
 359 #endif /* CONFIG_NET_SCHED */
 360 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
 361
 362 dont_put_skbuff_in_hw:
 363
 364         /* Check if we can use the hardware checksumming */
 365         if ((skb->protocol == htons(ETH_P_IP)) &&
 366             (ip_hdr(skb)->version == 4) &&
 367             (ip_hdr(skb)->ihl == 5) &&
 368             ((ip_hdr(skb)->frag_off == 0) ||
 369              (ip_hdr(skb)->frag_off == htons(1 << 14))) &&
 370             ((ip_hdr(skb)->protocol == IPPROTO_TCP) ||
 371              (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
 372                 /* Use hardware checksum calc */
 373                 pko_command.s.ipoffp1 = skb_network_offset(skb) + 1;
 374         }
 375
 376         if (USE_ASYNC_IOBDMA) {
 377                 /* Get the number of skbuffs in use by the hardware */
 378                 CVMX_SYNCIOBDMA;
 379                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 380                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 381         } else {
 382                 /* Get the number of skbuffs in use by the hardware */
 383                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 384                                                        MAX_SKB_TO_FREE);
 385                 buffers_to_free =
 386                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 387         }
 388
 389         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 390                                                  priv->fau + qos * 4);
 391
 392         /*
 393          * If we're sending faster than the receive can free them then
 394          * don't do the HW free.
 395          */
 396         if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 397                 pko_command.s.dontfree = 1;
 398
 399         if (pko_command.s.dontfree) {
 400                 queue_type = QUEUE_CORE;
 401                 pko_command.s.reg0 = priv->fau + qos * 4;
 402         } else {
 403                 queue_type = QUEUE_HW;
 404         }
 405         if (USE_ASYNC_IOBDMA)
 406                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
 407                                                FAU_TOTAL_TX_TO_CLEAN, 1);
 408
 409         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 410
 411         /* Drop this packet if we have too many already queued to the HW */
 412         if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >=
 413                      MAX_OUT_QUEUE_DEPTH)) {
 414                 if (dev->tx_queue_len != 0) {
 415                         /* Drop the lock when notifying the core.  */
 416                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 417                                                flags);
 418                         netif_stop_queue(dev);
 419                         spin_lock_irqsave(&priv->tx_free_list[qos].lock,
 420                                           flags);
 421                 } else {
 422                         /* If not using normal queueing.  */
 423                         queue_type = QUEUE_DROP;
 424                         goto skip_xmit;
 425                 }
 426         }
 427
 428         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
 429                                      CVMX_PKO_LOCK_NONE);
 430
 431         /* Send the packet to the output queue */
 432         if (unlikely(cvmx_pko_send_packet_finish(priv->port,
 433                                                  priv->queue + qos,
 434                                                  pko_command, hw_buffer,
 435                                                  CVMX_PKO_LOCK_NONE))) {
 436                 printk_ratelimited("%s: Failed to send the packet\n",
 437                                    dev->name);
 438                 queue_type = QUEUE_DROP;
 439         }
 440 skip_xmit:
 441         to_free_list = NULL;
 442
 443         switch (queue_type) {
 444         case QUEUE_DROP:
 445                 skb->next = to_free_list;
 446                 to_free_list = skb;
 447                 dev->stats.tx_dropped++;
 448                 break;
 449         case QUEUE_HW:
 450                 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
 451                 break;
 452         case QUEUE_CORE:
 453                 __skb_queue_tail(&priv->tx_free_list[qos], skb);
 454                 break;
 455         default:
 456                 BUG();
 457         }
 458
 459         while (skb_to_free > 0) {
 460                 struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
 461
 462                 t->next = to_free_list;
 463                 to_free_list = t;
 464                 skb_to_free--;
 465         }
 466
 467         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 468
 469         /* Do the actual freeing outside of the lock. */
 470         while (to_free_list) {
 471                 struct sk_buff *t = to_free_list;
 472
 473                 to_free_list = to_free_list->next;
 474                 dev_kfree_skb_any(t);
 475         }
 476
 477         if (USE_ASYNC_IOBDMA) {
 478                 CVMX_SYNCIOBDMA;
 479                 total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 480                 /* Restore the scratch area */
 481                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 482                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 483         } else {
 484                 total_to_clean =
 485                         cvmx_fau_fetch_and_add32(FAU_TOTAL_TX_TO_CLEAN, 1);
 486         }
 487
 488         if (total_to_clean & 0x3ff) {
 489                 /*
 490                  * Schedule the cleanup tasklet every 1024 packets for
 491                  * the pathological case of high traffic on one port
 492                  * delaying clean up of packets on a different port
 493                  * that is blocked waiting for the cleanup.
 494                  */
 495                 tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 496         }
 497
 498         cvm_oct_kick_tx_poll_watchdog();
 499
 500         return NETDEV_TX_OK;
 501 }
 502
 503 /**
 504  * cvm_oct_xmit_pow - transmit a packet to the POW
 505  * @skb:    Packet to send
 506  * @dev:    Device info structure
 507
 508  * Returns Always returns zero
 509  */
 510 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
 511 {
 512         struct octeon_ethernet *priv = netdev_priv(dev);
 513         void *packet_buffer;
 514         void *copy_location;
 515
 516         /* Get a work queue entry */
 517         struct cvmx_wqe *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
 518
 519         if (unlikely(!work)) {
 520                 printk_ratelimited("%s: Failed to allocate a work queue entry\n",
 521                                    dev->name);
 522                 dev->stats.tx_dropped++;
 523                 dev_kfree_skb_any(skb);
 524                 return 0;
 525         }
 526
 527         /* Get a packet buffer */
 528         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
 529         if (unlikely(!packet_buffer)) {
 530                 printk_ratelimited("%s: Failed to allocate a packet buffer\n",
 531                                    dev->name);
 532                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 533                 dev->stats.tx_dropped++;
 534                 dev_kfree_skb_any(skb);
 535                 return 0;
 536         }
 537
 538         /*
 539          * Calculate where we need to copy the data to. We need to
 540          * leave 8 bytes for a next pointer (unused). We also need to
 541          * include any configure skip. Then we need to align the IP
 542          * packet src and dest into the same 64bit word. The below
 543          * calculation may add a little extra, but that doesn't
 544          * hurt.
 545          */
 546         copy_location = packet_buffer + sizeof(u64);
 547         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
 548
 549         /*
 550          * We have to copy the packet since whoever processes this
 551          * packet will free it to a hardware pool. We can't use the
 552          * trick of counting outstanding packets like in
 553          * cvm_oct_xmit.
 554          */
 555         memcpy(copy_location, skb->data, skb->len);
 556
 557         /*
 558          * Fill in some of the work queue fields. We may need to add
 559          * more if the software at the other end needs them.
 560          */
 561         if (!OCTEON_IS_MODEL(OCTEON_CN68XX))
 562                 work->word0.pip.cn38xx.hw_chksum = skb->csum;
 563         work->word1.len = skb->len;
 564         cvmx_wqe_set_port(work, priv->port);
 565         cvmx_wqe_set_qos(work, priv->port & 0x7);
 566         cvmx_wqe_set_grp(work, pow_send_group);
 567         work->word1.tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
 568         work->word1.tag = pow_send_group;       /* FIXME */
 569         /* Default to zero. Sets of zero later are commented out */
 570         work->word2.u64 = 0;
 571         work->word2.s.bufs = 1;
 572         work->packet_ptr.u64 = 0;
 573         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
 574         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
 575         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
 576         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
 577
 578         if (skb->protocol == htons(ETH_P_IP)) {
 579                 work->word2.s.ip_offset = 14;
 580 #if 0
 581                 work->word2.s.vlan_valid = 0;   /* FIXME */
 582                 work->word2.s.vlan_cfi = 0;     /* FIXME */
 583                 work->word2.s.vlan_id = 0;      /* FIXME */
 584                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
 585 #endif
 586                 work->word2.s.tcp_or_udp =
 587                     (ip_hdr(skb)->protocol == IPPROTO_TCP) ||
 588                     (ip_hdr(skb)->protocol == IPPROTO_UDP);
 589 #if 0
 590                 /* FIXME */
 591                 work->word2.s.dec_ipsec = 0;
 592                 /* We only support IPv4 right now */
 593                 work->word2.s.is_v6 = 0;
 594                 /* Hardware would set to zero */
 595                 work->word2.s.software = 0;
 596                 /* No error, packet is internal */
 597                 work->word2.s.L4_error = 0;
 598 #endif
 599                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0) ||
 600                                           (ip_hdr(skb)->frag_off ==
 601                                               cpu_to_be16(1 << 14)));
 602 #if 0
 603                 /* Assume Linux is sending a good packet */
 604                 work->word2.s.IP_exc = 0;
 605 #endif
 606                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
 607                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
 608 #if 0
 609                 /* This is an IP packet */
 610                 work->word2.s.not_IP = 0;
 611                 /* No error, packet is internal */
 612                 work->word2.s.rcv_error = 0;
 613                 /* No error, packet is internal */
 614                 work->word2.s.err_code = 0;
 615 #endif
 616
 617                 /*
 618                  * When copying the data, include 4 bytes of the
 619                  * ethernet header to align the same way hardware
 620                  * does.
 621                  */
 622                 memcpy(work->packet_data, skb->data + 10,
 623                        sizeof(work->packet_data));
 624         } else {
 625 #if 0
 626                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
 627                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
 628                 work->word2.snoip.vlan_id = 0;  /* FIXME */
 629                 work->word2.snoip.software = 0; /* Hardware would set to zero */
 630 #endif
 631                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
 632                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
 633                 work->word2.snoip.is_bcast =
 634                     (skb->pkt_type == PACKET_BROADCAST);
 635                 work->word2.snoip.is_mcast =
 636                     (skb->pkt_type == PACKET_MULTICAST);
 637                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
 638 #if 0
 639                 /* No error, packet is internal */
 640                 work->word2.snoip.rcv_error = 0;
 641                 /* No error, packet is internal */
 642                 work->word2.snoip.err_code = 0;
 643 #endif
 644                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
 645         }
 646
 647         /* Submit the packet to the POW */
 648         cvmx_pow_work_submit(work, work->word1.tag, work->word1.tag_type,
 649                              cvmx_wqe_get_qos(work), cvmx_wqe_get_grp(work));
 650         dev->stats.tx_packets++;
 651         dev->stats.tx_bytes += skb->len;
 652         dev_consume_skb_any(skb);
 653         return 0;
 654 }
 655
 656 /**
 657  * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
 658  * @dev:    Device being shutdown
 659  *
 660  */
 661 void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 662 {
 663         struct octeon_ethernet *priv = netdev_priv(dev);
 664         unsigned long flags;
 665         int qos;
 666
 667         for (qos = 0; qos < 16; qos++) {
 668                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 669                 while (skb_queue_len(&priv->tx_free_list[qos]))
 670                         dev_kfree_skb_any(__skb_dequeue
 671                                           (&priv->tx_free_list[qos]));
 672                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 673         }
 674 }
 675
 676 static void cvm_oct_tx_do_cleanup(unsigned long arg)
 677 {
 678         int port;
 679
 680         for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
 681                 if (cvm_oct_device[port]) {
 682                         struct net_device *dev = cvm_oct_device[port];
 683
 684                         cvm_oct_free_tx_skbs(dev);
 685                 }
 686         }
 687 }
 688
 689 static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
 690 {
 691         /* Disable the interrupt.  */
 692         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 693         /* Do the work in the tasklet.  */
 694         tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 695         return IRQ_HANDLED;
 696 }
 697
 698 void cvm_oct_tx_initialize(void)
 699 {
 700         int i;
 701
 702         /* Disable the interrupt.  */
 703         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 704         /* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
 705         i = request_irq(OCTEON_IRQ_TIMER1,
 706                         cvm_oct_tx_cleanup_watchdog, 0,
 707                         "Ethernet", cvm_oct_device);
 708
 709         if (i)
 710                 panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
 711 }
 712
 713 void cvm_oct_tx_shutdown(void)
 714 {
 715         /* Free the interrupt handler */
 716         free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
 717 }