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/ratelimit.h>
  34 #include <linux/string.h>
  35 #include <linux/interrupt.h>
  36 #include <net/dst.h>
  37 #ifdef CONFIG_XFRM
  38 #include <linux/xfrm.h>
  39 #include <net/xfrm.h>
  40 #endif /* CONFIG_XFRM */
  41
  42 #include <linux/atomic.h>
  43
  44 #include <asm/octeon/octeon.h>
  45
  46 #include "ethernet-defines.h"
  47 #include "octeon-ethernet.h"
  48 #include "ethernet-tx.h"
  49 #include "ethernet-util.h"
  50
  51 #include <asm/octeon/cvmx-wqe.h>
  52 #include <asm/octeon/cvmx-fau.h>
  53 #include <asm/octeon/cvmx-pip.h>
  54 #include <asm/octeon/cvmx-pko.h>
  55 #include <asm/octeon/cvmx-helper.h>
  56
  57 #include <asm/octeon/cvmx-gmxx-defs.h>
  58
  59 #define CVM_OCT_SKB_CB(skb)     ((u64 *)((skb)->cb))
  60
  61 /*
  62  * You can define GET_SKBUFF_QOS() to override how the skbuff output
  63  * function determines which output queue is used. The default
  64  * implementation always uses the base queue for the port. If, for
  65  * example, you wanted to use the skb->priority field, define
  66  * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
  67  */
  68 #ifndef GET_SKBUFF_QOS
  69 #define GET_SKBUFF_QOS(skb) 0
  70 #endif
  71
  72 static void cvm_oct_tx_do_cleanup(unsigned long arg);
  73 static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
  74
  75 /* Maximum number of SKBs to try to free per xmit packet. */
  76 #define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
  77
  78 static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
  79 {
  80         int32_t undo;
  81         undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free +
  82                                                    MAX_SKB_TO_FREE;
  83         if (undo > 0)
  84                 cvmx_fau_atomic_add32(fau, -undo);
  85         skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE :
  86                                                        -skb_to_free;
  87         return skb_to_free;
  88 }
  89
  90 static void cvm_oct_kick_tx_poll_watchdog(void)
  91 {
  92         union cvmx_ciu_timx ciu_timx;
  93         ciu_timx.u64 = 0;
  94         ciu_timx.s.one_shot = 1;
  95         ciu_timx.s.len = cvm_oct_tx_poll_interval;
  96         cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
  97 }
  98
  99 void cvm_oct_free_tx_skbs(struct net_device *dev)
 100 {
 101         int32_t skb_to_free;
 102         int qos, queues_per_port;
 103         int total_freed = 0;
 104         int total_remaining = 0;
 105         unsigned long flags;
 106         struct octeon_ethernet *priv = netdev_priv(dev);
 107
 108         queues_per_port = cvmx_pko_get_num_queues(priv->port);
 109         /* Drain any pending packets in the free list */
 110         for (qos = 0; qos < queues_per_port; qos++) {
 111                 if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
 112                         continue;
 113                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau+qos*4,
 114                                                        MAX_SKB_TO_FREE);
 115                 skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 116                                                          priv->fau+qos*4);
 117
 118
 119                 total_freed += skb_to_free;
 120                 if (skb_to_free > 0) {
 121                         struct sk_buff *to_free_list = NULL;
 122                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 123                         while (skb_to_free > 0) {
 124                                 struct sk_buff *t;
 125                                 t = __skb_dequeue(&priv->tx_free_list[qos]);
 126                                 t->next = to_free_list;
 127                                 to_free_list = t;
 128                                 skb_to_free--;
 129                         }
 130                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 131                                                flags);
 132                         /* Do the actual freeing outside of the lock. */
 133                         while (to_free_list) {
 134                                 struct sk_buff *t = to_free_list;
 135                                 to_free_list = to_free_list->next;
 136                                 dev_kfree_skb_any(t);
 137                         }
 138                 }
 139                 total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
 140         }
 141         if (total_freed >= 0 && netif_queue_stopped(dev))
 142                 netif_wake_queue(dev);
 143         if (total_remaining)
 144                 cvm_oct_kick_tx_poll_watchdog();
 145 }
 146
 147 /**
 148  * cvm_oct_xmit - transmit a packet
 149  * @skb:    Packet to send
 150  * @dev:    Device info structure
 151  *
 152  * Returns Always returns NETDEV_TX_OK
 153  */
 154 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 155 {
 156         cvmx_pko_command_word0_t pko_command;
 157         union cvmx_buf_ptr hw_buffer;
 158         uint64_t old_scratch;
 159         uint64_t old_scratch2;
 160         int qos;
 161         int i;
 162         enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
 163         struct octeon_ethernet *priv = netdev_priv(dev);
 164         struct sk_buff *to_free_list;
 165         int32_t skb_to_free;
 166         int32_t buffers_to_free;
 167         u32 total_to_clean;
 168         unsigned long flags;
 169 #if REUSE_SKBUFFS_WITHOUT_FREE
 170         unsigned char *fpa_head;
 171 #endif
 172
 173         /*
 174          * Prefetch the private data structure.  It is larger than the
 175          * one cache line.
 176          */
 177         prefetch(priv);
 178
 179         /*
 180          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
 181          * completely remove "qos" in the event neither interface
 182          * supports multiple queues per port.
 183          */
 184         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
 185             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
 186                 qos = GET_SKBUFF_QOS(skb);
 187                 if (qos <= 0)
 188                         qos = 0;
 189                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
 190                         qos = 0;
 191         } else
 192                 qos = 0;
 193
 194         if (USE_ASYNC_IOBDMA) {
 195                 /* Save scratch in case userspace is using it */
 196                 CVMX_SYNCIOBDMA;
 197                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 198                 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 199
 200                 /*
 201                  * Fetch and increment the number of packets to be
 202                  * freed.
 203                  */
 204                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
 205                                                FAU_NUM_PACKET_BUFFERS_TO_FREE,
 206                                                0);
 207                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
 208                                                priv->fau + qos * 4,
 209                                                MAX_SKB_TO_FREE);
 210         }
 211
 212         /*
 213          * We have space for 6 segment pointers, If there will be more
 214          * than that, we must linearize.
 215          */
 216         if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
 217                 if (unlikely(__skb_linearize(skb))) {
 218                         queue_type = QUEUE_DROP;
 219                         if (USE_ASYNC_IOBDMA) {
 220                                 /*
 221                                  * Get the number of skbuffs in use
 222                                  * by the hardware
 223                                  */
 224                                 CVMX_SYNCIOBDMA;
 225                                 skb_to_free =
 226                                         cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 227                         } else {
 228                                 /*
 229                                  * Get the number of skbuffs in use
 230                                  * by the hardware
 231                                  */
 232                                 skb_to_free = cvmx_fau_fetch_and_add32(
 233                                         priv->fau + qos * 4, MAX_SKB_TO_FREE);
 234                         }
 235                         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 236                                                         priv->fau + qos * 4);
 237                         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 238                         goto skip_xmit;
 239                 }
 240         }
 241
 242         /*
 243          * The CN3XXX series of parts has an errata (GMX-401) which
 244          * causes the GMX block to hang if a collision occurs towards
 245          * the end of a <68 byte packet. As a workaround for this, we
 246          * pad packets to be 68 bytes whenever we are in half duplex
 247          * mode. We don't handle the case of having a small packet but
 248          * no room to add the padding.  The kernel should always give
 249          * us at least a cache line
 250          */
 251         if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
 252                 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
 253                 int interface = INTERFACE(priv->port);
 254                 int index = INDEX(priv->port);
 255
 256                 if (interface < 2) {
 257                         /* We only need to pad packet in half duplex mode */
 258                         gmx_prt_cfg.u64 =
 259                             cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
 260                         if (gmx_prt_cfg.s.duplex == 0) {
 261                                 int add_bytes = 64 - skb->len;
 262                                 if ((skb_tail_pointer(skb) + add_bytes) <=
 263                                     skb_end_pointer(skb))
 264                                         memset(__skb_put(skb, add_bytes), 0,
 265                                                add_bytes);
 266                         }
 267                 }
 268         }
 269
 270         /* Build the PKO command */
 271         pko_command.u64 = 0;
 272         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
 273         pko_command.s.segs = 1;
 274         pko_command.s.total_bytes = skb->len;
 275         pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
 276         pko_command.s.subone0 = 1;
 277
 278         pko_command.s.dontfree = 1;
 279
 280         /* Build the PKO buffer pointer */
 281         hw_buffer.u64 = 0;
 282         if (skb_shinfo(skb)->nr_frags == 0) {
 283                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
 284                 hw_buffer.s.pool = 0;
 285                 hw_buffer.s.size = skb->len;
 286         } else {
 287                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
 288                 hw_buffer.s.pool = 0;
 289                 hw_buffer.s.size = skb_headlen(skb);
 290                 CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
 291                 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 292                         struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i;
 293                         hw_buffer.s.addr = XKPHYS_TO_PHYS(
 294                                 (u64)(page_address(fs->page.p) +
 295                                 fs->page_offset));
 296                         hw_buffer.s.size = fs->size;
 297                         CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
 298                 }
 299                 hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb));
 300                 hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1;
 301                 pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1;
 302                 pko_command.s.gather = 1;
 303                 goto dont_put_skbuff_in_hw;
 304         }
 305
 306         /*
 307          * See if we can put this skb in the FPA pool. Any strange
 308          * behavior from the Linux networking stack will most likely
 309          * be caused by a bug in the following code. If some field is
 310          * in use by the network stack and gets carried over when a
 311          * buffer is reused, bad things may happen.  If in doubt and
 312          * you dont need the absolute best performance, disable the
 313          * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
 314          * shown a 25% increase in performance under some loads.
 315          */
 316 #if REUSE_SKBUFFS_WITHOUT_FREE
 317         fpa_head = skb->head + 256 - ((unsigned long)skb->head & 0x7f);
 318         if (unlikely(skb->data < fpa_head)) {
 319                 /*
 320                  * printk("TX buffer beginning can't meet FPA
 321                  * alignment constraints\n");
 322                  */
 323                 goto dont_put_skbuff_in_hw;
 324         }
 325         if (unlikely
 326             ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
 327                 /*
 328                    printk("TX buffer isn't large enough for the FPA\n");
 329                  */
 330                 goto dont_put_skbuff_in_hw;
 331         }
 332         if (unlikely(skb_shared(skb))) {
 333                 /*
 334                    printk("TX buffer sharing data with someone else\n");
 335                  */
 336                 goto dont_put_skbuff_in_hw;
 337         }
 338         if (unlikely(skb_cloned(skb))) {
 339                 /*
 340                    printk("TX buffer has been cloned\n");
 341                  */
 342                 goto dont_put_skbuff_in_hw;
 343         }
 344         if (unlikely(skb_header_cloned(skb))) {
 345                 /*
 346                    printk("TX buffer header has been cloned\n");
 347                  */
 348                 goto dont_put_skbuff_in_hw;
 349         }
 350         if (unlikely(skb->destructor)) {
 351                 /*
 352                    printk("TX buffer has a destructor\n");
 353                  */
 354                 goto dont_put_skbuff_in_hw;
 355         }
 356         if (unlikely(skb_shinfo(skb)->nr_frags)) {
 357                 /*
 358                    printk("TX buffer has fragments\n");
 359                  */
 360                 goto dont_put_skbuff_in_hw;
 361         }
 362         if (unlikely
 363             (skb->truesize !=
 364              sizeof(*skb) + skb_end_offset(skb))) {
 365                 /*
 366                    printk("TX buffer truesize has been changed\n");
 367                  */
 368                 goto dont_put_skbuff_in_hw;
 369         }
 370
 371         /*
 372          * We can use this buffer in the FPA.  We don't need the FAU
 373          * update anymore
 374          */
 375         pko_command.s.dontfree = 0;
 376
 377         hw_buffer.s.back = ((unsigned long)skb->data >> 7) -
 378                            ((unsigned long)fpa_head >> 7);
 379
 380         *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
 381
 382         /*
 383          * The skbuff will be reused without ever being freed. We must
 384          * cleanup a bunch of core things.
 385          */
 386         dst_release(skb_dst(skb));
 387         skb_dst_set(skb, NULL);
 388 #ifdef CONFIG_XFRM
 389         secpath_put(skb->sp);
 390         skb->sp = NULL;
 391 #endif
 392         nf_reset(skb);
 393
 394 #ifdef CONFIG_NET_SCHED
 395         skb->tc_index = 0;
 396 #ifdef CONFIG_NET_CLS_ACT
 397         skb->tc_verd = 0;
 398 #endif /* CONFIG_NET_CLS_ACT */
 399 #endif /* CONFIG_NET_SCHED */
 400 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
 401
 402 dont_put_skbuff_in_hw:
 403
 404         /* Check if we can use the hardware checksumming */
 405         if (USE_HW_TCPUDP_CHECKSUM && (skb->protocol == htons(ETH_P_IP)) &&
 406             (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) &&
 407             ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == 1 << 14))
 408             && ((ip_hdr(skb)->protocol == IPPROTO_TCP)
 409                 || (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
 410                 /* Use hardware checksum calc */
 411                 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
 412         }
 413
 414         if (USE_ASYNC_IOBDMA) {
 415                 /* Get the number of skbuffs in use by the hardware */
 416                 CVMX_SYNCIOBDMA;
 417                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 418                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 419         } else {
 420                 /* Get the number of skbuffs in use by the hardware */
 421                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 422                                                        MAX_SKB_TO_FREE);
 423                 buffers_to_free =
 424                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 425         }
 426
 427         skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4);
 428
 429         /*
 430          * If we're sending faster than the receive can free them then
 431          * don't do the HW free.
 432          */
 433         if ((buffers_to_free < -100) && !pko_command.s.dontfree)
 434                 pko_command.s.dontfree = 1;
 435
 436         if (pko_command.s.dontfree) {
 437                 queue_type = QUEUE_CORE;
 438                 pko_command.s.reg0 = priv->fau+qos*4;
 439         } else {
 440                 queue_type = QUEUE_HW;
 441         }
 442         if (USE_ASYNC_IOBDMA)
 443                 cvmx_fau_async_fetch_and_add32(
 444                                 CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
 445
 446         spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 447
 448         /* Drop this packet if we have too many already queued to the HW */
 449         if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >=
 450                      MAX_OUT_QUEUE_DEPTH)) {
 451
 452                 if (dev->tx_queue_len != 0) {
 453                         /* Drop the lock when notifying the core.  */
 454                         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
 455                                                flags);
 456                         netif_stop_queue(dev);
 457                         spin_lock_irqsave(&priv->tx_free_list[qos].lock,
 458                                           flags);
 459                 } else {
 460                         /* If not using normal queueing.  */
 461                         queue_type = QUEUE_DROP;
 462                         goto skip_xmit;
 463                 }
 464         }
 465
 466         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
 467                                      CVMX_PKO_LOCK_NONE);
 468
 469         /* Send the packet to the output queue */
 470         if (unlikely(cvmx_pko_send_packet_finish(priv->port,
 471                                                  priv->queue + qos,
 472                                                  pko_command, hw_buffer,
 473                                                  CVMX_PKO_LOCK_NONE))) {
 474                 printk_ratelimited("%s: Failed to send the packet\n",
 475                                    dev->name);
 476                 queue_type = QUEUE_DROP;
 477         }
 478 skip_xmit:
 479         to_free_list = NULL;
 480
 481         switch (queue_type) {
 482         case QUEUE_DROP:
 483                 skb->next = to_free_list;
 484                 to_free_list = skb;
 485                 priv->stats.tx_dropped++;
 486                 break;
 487         case QUEUE_HW:
 488                 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
 489                 break;
 490         case QUEUE_CORE:
 491                 __skb_queue_tail(&priv->tx_free_list[qos], skb);
 492                 break;
 493         default:
 494                 BUG();
 495         }
 496
 497         while (skb_to_free > 0) {
 498                 struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
 499                 t->next = to_free_list;
 500                 to_free_list = t;
 501                 skb_to_free--;
 502         }
 503
 504         spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 505
 506         /* Do the actual freeing outside of the lock. */
 507         while (to_free_list) {
 508                 struct sk_buff *t = to_free_list;
 509                 to_free_list = to_free_list->next;
 510                 dev_kfree_skb_any(t);
 511         }
 512
 513         if (USE_ASYNC_IOBDMA) {
 514                 CVMX_SYNCIOBDMA;
 515                 total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 516                 /* Restore the scratch area */
 517                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 518                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 519         } else {
 520                 total_to_clean = cvmx_fau_fetch_and_add32(
 521                                                 FAU_TOTAL_TX_TO_CLEAN, 1);
 522         }
 523
 524         if (total_to_clean & 0x3ff) {
 525                 /*
 526                  * Schedule the cleanup tasklet every 1024 packets for
 527                  * the pathological case of high traffic on one port
 528                  * delaying clean up of packets on a different port
 529                  * that is blocked waiting for the cleanup.
 530                  */
 531                 tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 532         }
 533
 534         cvm_oct_kick_tx_poll_watchdog();
 535
 536         return NETDEV_TX_OK;
 537 }
 538
 539 /**
 540  * cvm_oct_xmit_pow - transmit a packet to the POW
 541  * @skb:    Packet to send
 542  * @dev:    Device info structure
 543
 544  * Returns Always returns zero
 545  */
 546 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
 547 {
 548         struct octeon_ethernet *priv = netdev_priv(dev);
 549         void *packet_buffer;
 550         void *copy_location;
 551
 552         /* Get a work queue entry */
 553         cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
 554         if (unlikely(work == NULL)) {
 555                 printk_ratelimited("%s: Failed to allocate a work queue entry\n",
 556                                    dev->name);
 557                 priv->stats.tx_dropped++;
 558                 dev_kfree_skb(skb);
 559                 return 0;
 560         }
 561
 562         /* Get a packet buffer */
 563         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
 564         if (unlikely(packet_buffer == NULL)) {
 565                 printk_ratelimited("%s: Failed to allocate a packet buffer\n",
 566                                    dev->name);
 567                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
 568                 priv->stats.tx_dropped++;
 569                 dev_kfree_skb(skb);
 570                 return 0;
 571         }
 572
 573         /*
 574          * Calculate where we need to copy the data to. We need to
 575          * leave 8 bytes for a next pointer (unused). We also need to
 576          * include any configure skip. Then we need to align the IP
 577          * packet src and dest into the same 64bit word. The below
 578          * calculation may add a little extra, but that doesn't
 579          * hurt.
 580          */
 581         copy_location = packet_buffer + sizeof(uint64_t);
 582         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
 583
 584         /*
 585          * We have to copy the packet since whoever processes this
 586          * packet will free it to a hardware pool. We can't use the
 587          * trick of counting outstanding packets like in
 588          * cvm_oct_xmit.
 589          */
 590         memcpy(copy_location, skb->data, skb->len);
 591
 592         /*
 593          * Fill in some of the work queue fields. We may need to add
 594          * more if the software at the other end needs them.
 595          */
 596         work->hw_chksum = skb->csum;
 597         work->len = skb->len;
 598         work->ipprt = priv->port;
 599         work->qos = priv->port & 0x7;
 600         work->grp = pow_send_group;
 601         work->tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
 602         work->tag = pow_send_group;     /* FIXME */
 603         /* Default to zero. Sets of zero later are commented out */
 604         work->word2.u64 = 0;
 605         work->word2.s.bufs = 1;
 606         work->packet_ptr.u64 = 0;
 607         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
 608         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
 609         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
 610         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
 611
 612         if (skb->protocol == htons(ETH_P_IP)) {
 613                 work->word2.s.ip_offset = 14;
 614 #if 0
 615                 work->word2.s.vlan_valid = 0;   /* FIXME */
 616                 work->word2.s.vlan_cfi = 0;     /* FIXME */
 617                 work->word2.s.vlan_id = 0;      /* FIXME */
 618                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
 619 #endif
 620                 work->word2.s.tcp_or_udp =
 621                     (ip_hdr(skb)->protocol == IPPROTO_TCP)
 622                     || (ip_hdr(skb)->protocol == IPPROTO_UDP);
 623 #if 0
 624                 /* FIXME */
 625                 work->word2.s.dec_ipsec = 0;
 626                 /* We only support IPv4 right now */
 627                 work->word2.s.is_v6 = 0;
 628                 /* Hardware would set to zero */
 629                 work->word2.s.software = 0;
 630                 /* No error, packet is internal */
 631                 work->word2.s.L4_error = 0;
 632 #endif
 633                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0)
 634                                           || (ip_hdr(skb)->frag_off ==
 635                                               1 << 14));
 636 #if 0
 637                 /* Assume Linux is sending a good packet */
 638                 work->word2.s.IP_exc = 0;
 639 #endif
 640                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
 641                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
 642 #if 0
 643                 /* This is an IP packet */
 644                 work->word2.s.not_IP = 0;
 645                 /* No error, packet is internal */
 646                 work->word2.s.rcv_error = 0;
 647                 /* No error, packet is internal */
 648                 work->word2.s.err_code = 0;
 649 #endif
 650
 651                 /*
 652                  * When copying the data, include 4 bytes of the
 653                  * ethernet header to align the same way hardware
 654                  * does.
 655                  */
 656                 memcpy(work->packet_data, skb->data + 10,
 657                        sizeof(work->packet_data));
 658         } else {
 659 #if 0
 660                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
 661                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
 662                 work->word2.snoip.vlan_id = 0;  /* FIXME */
 663                 work->word2.snoip.software = 0; /* Hardware would set to zero */
 664 #endif
 665                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
 666                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
 667                 work->word2.snoip.is_bcast =
 668                     (skb->pkt_type == PACKET_BROADCAST);
 669                 work->word2.snoip.is_mcast =
 670                     (skb->pkt_type == PACKET_MULTICAST);
 671                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
 672 #if 0
 673                 /* No error, packet is internal */
 674                 work->word2.snoip.rcv_error = 0;
 675                 /* No error, packet is internal */
 676                 work->word2.snoip.err_code = 0;
 677 #endif
 678                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
 679         }
 680
 681         /* Submit the packet to the POW */
 682         cvmx_pow_work_submit(work, work->tag, work->tag_type, work->qos,
 683                              work->grp);
 684         priv->stats.tx_packets++;
 685         priv->stats.tx_bytes += skb->len;
 686         dev_kfree_skb(skb);
 687         return 0;
 688 }
 689
 690 /**
 691  * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
 692  * @dev:    Device being shutdown
 693  *
 694  */
 695 void cvm_oct_tx_shutdown_dev(struct net_device *dev)
 696 {
 697         struct octeon_ethernet *priv = netdev_priv(dev);
 698         unsigned long flags;
 699         int qos;
 700
 701         for (qos = 0; qos < 16; qos++) {
 702                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
 703                 while (skb_queue_len(&priv->tx_free_list[qos]))
 704                         dev_kfree_skb_any(__skb_dequeue
 705                                           (&priv->tx_free_list[qos]));
 706                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
 707         }
 708 }
 709
 710 static void cvm_oct_tx_do_cleanup(unsigned long arg)
 711 {
 712         int port;
 713
 714         for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
 715                 if (cvm_oct_device[port]) {
 716                         struct net_device *dev = cvm_oct_device[port];
 717                         cvm_oct_free_tx_skbs(dev);
 718                 }
 719         }
 720 }
 721
 722 static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
 723 {
 724         /* Disable the interrupt.  */
 725         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 726         /* Do the work in the tasklet.  */
 727         tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
 728         return IRQ_HANDLED;
 729 }
 730
 731 void cvm_oct_tx_initialize(void)
 732 {
 733         int i;
 734
 735         /* Disable the interrupt.  */
 736         cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
 737         /* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
 738         i = request_irq(OCTEON_IRQ_TIMER1,
 739                         cvm_oct_tx_cleanup_watchdog, 0,
 740                         "Ethernet", cvm_oct_device);
 741
 742         if (i)
 743                 panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
 744 }
 745
 746 void cvm_oct_tx_shutdown(void)
 747 {
 748         /* Free the interrupt handler */
 749         free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
 750 }