bfd3dd2fcef8d997e0324b690fd6ba6477d547bd
[linux-2.6.git] / 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) 2003-2007 Cavium Networks
8  *
9  * This file is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License, Version 2, as
11  * published by the Free Software Foundation.
12  *
13  * This file is distributed in the hope that it will be useful, but
14  * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty
15  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or
16  * NONINFRINGEMENT.  See the GNU General Public License for more
17  * details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this file; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
22  * or visit http://www.gnu.org/licenses/.
23  *
24  * This file may also be available under a different license from Cavium.
25  * Contact Cavium Networks for more information
26 *********************************************************************/
27 #include <linux/module.h>
28 #include <linux/kernel.h>
29 #include <linux/netdevice.h>
30 #include <linux/init.h>
31 #include <linux/etherdevice.h>
32 #include <linux/ip.h>
33 #include <linux/string.h>
34 #include <linux/ethtool.h>
35 #include <linux/mii.h>
36 #include <linux/seq_file.h>
37 #include <linux/proc_fs.h>
38 #include <net/dst.h>
39 #ifdef CONFIG_XFRM
40 #include <linux/xfrm.h>
41 #include <net/xfrm.h>
42 #endif /* CONFIG_XFRM */
43
44 #include <asm/atomic.h>
45
46 #include <asm/octeon/octeon.h>
47
48 #include "ethernet-defines.h"
49 #include "octeon-ethernet.h"
50 #include "ethernet-util.h"
51
52 #include "cvmx-wqe.h"
53 #include "cvmx-fau.h"
54 #include "cvmx-pko.h"
55 #include "cvmx-helper.h"
56
57 #include "cvmx-gmxx-defs.h"
58
59 /*
60  * You can define GET_SKBUFF_QOS() to override how the skbuff output
61  * function determines which output queue is used. The default
62  * implementation always uses the base queue for the port. If, for
63  * example, you wanted to use the skb->priority fieid, define
64  * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
65  */
66 #ifndef GET_SKBUFF_QOS
67 #define GET_SKBUFF_QOS(skb) 0
68 #endif
69
70 /**
71  * Packet transmit
72  *
73  * @skb:    Packet to send
74  * @dev:    Device info structure
75  * Returns Always returns zero
76  */
77 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
78 {
79         cvmx_pko_command_word0_t pko_command;
80         union cvmx_buf_ptr hw_buffer;
81         uint64_t old_scratch;
82         uint64_t old_scratch2;
83         int dropped;
84         int qos;
85         struct octeon_ethernet *priv = netdev_priv(dev);
86         int32_t in_use;
87         int32_t buffers_to_free;
88 #if REUSE_SKBUFFS_WITHOUT_FREE
89         unsigned char *fpa_head;
90 #endif
91
92         /*
93          * Prefetch the private data structure.  It is larger that one
94          * cache line.
95          */
96         prefetch(priv);
97
98         /* Start off assuming no drop */
99         dropped = 0;
100
101         /*
102          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
103          * completely remove "qos" in the event neither interface
104          * supports multiple queues per port.
105          */
106         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
107             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
108                 qos = GET_SKBUFF_QOS(skb);
109                 if (qos <= 0)
110                         qos = 0;
111                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
112                         qos = 0;
113         } else
114                 qos = 0;
115
116         if (USE_ASYNC_IOBDMA) {
117                 /* Save scratch in case userspace is using it */
118                 CVMX_SYNCIOBDMA;
119                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
120                 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
121
122                 /*
123                  * Assume we're going to be able t osend this
124                  * packet. Fetch and increment the number of pending
125                  * packets for output.
126                  */
127                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
128                                                FAU_NUM_PACKET_BUFFERS_TO_FREE,
129                                                0);
130                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
131                                                priv->fau + qos * 4, 1);
132         }
133
134         /*
135          * The CN3XXX series of parts has an errata (GMX-401) which
136          * causes the GMX block to hang if a collision occurs towards
137          * the end of a <68 byte packet. As a workaround for this, we
138          * pad packets to be 68 bytes whenever we are in half duplex
139          * mode. We don't handle the case of having a small packet but
140          * no room to add the padding.  The kernel should always give
141          * us at least a cache line
142          */
143         if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
144                 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
145                 int interface = INTERFACE(priv->port);
146                 int index = INDEX(priv->port);
147
148                 if (interface < 2) {
149                         /* We only need to pad packet in half duplex mode */
150                         gmx_prt_cfg.u64 =
151                             cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
152                         if (gmx_prt_cfg.s.duplex == 0) {
153                                 int add_bytes = 64 - skb->len;
154                                 if ((skb_tail_pointer(skb) + add_bytes) <=
155                                     skb_end_pointer(skb))
156                                         memset(__skb_put(skb, add_bytes), 0,
157                                                add_bytes);
158                         }
159                 }
160         }
161
162         /* Build the PKO buffer pointer */
163         hw_buffer.u64 = 0;
164         hw_buffer.s.addr = cvmx_ptr_to_phys(skb->data);
165         hw_buffer.s.pool = 0;
166         hw_buffer.s.size =
167             (unsigned long)skb_end_pointer(skb) - (unsigned long)skb->head;
168
169         /* Build the PKO command */
170         pko_command.u64 = 0;
171         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
172         pko_command.s.segs = 1;
173         pko_command.s.total_bytes = skb->len;
174         pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
175         pko_command.s.subone0 = 1;
176
177         pko_command.s.dontfree = 1;
178         pko_command.s.reg0 = priv->fau + qos * 4;
179         /*
180          * See if we can put this skb in the FPA pool. Any strange
181          * behavior from the Linux networking stack will most likely
182          * be caused by a bug in the following code. If some field is
183          * in use by the network stack and get carried over when a
184          * buffer is reused, bad thing may happen.  If in doubt and
185          * you dont need the absolute best performance, disable the
186          * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
187          * shown a 25% increase in performance under some loads.
188          */
189 #if REUSE_SKBUFFS_WITHOUT_FREE
190         fpa_head = skb->head + 128 - ((unsigned long)skb->head & 0x7f);
191         if (unlikely(skb->data < fpa_head)) {
192                 /*
193                  * printk("TX buffer beginning can't meet FPA
194                  * alignment constraints\n");
195                  */
196                 goto dont_put_skbuff_in_hw;
197         }
198         if (unlikely
199             ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
200                 /*
201                    printk("TX buffer isn't large enough for the FPA\n");
202                  */
203                 goto dont_put_skbuff_in_hw;
204         }
205         if (unlikely(skb_shared(skb))) {
206                 /*
207                    printk("TX buffer sharing data with someone else\n");
208                  */
209                 goto dont_put_skbuff_in_hw;
210         }
211         if (unlikely(skb_cloned(skb))) {
212                 /*
213                    printk("TX buffer has been cloned\n");
214                  */
215                 goto dont_put_skbuff_in_hw;
216         }
217         if (unlikely(skb_header_cloned(skb))) {
218                 /*
219                    printk("TX buffer header has been cloned\n");
220                  */
221                 goto dont_put_skbuff_in_hw;
222         }
223         if (unlikely(skb->destructor)) {
224                 /*
225                    printk("TX buffer has a destructor\n");
226                  */
227                 goto dont_put_skbuff_in_hw;
228         }
229         if (unlikely(skb_shinfo(skb)->nr_frags)) {
230                 /*
231                    printk("TX buffer has fragments\n");
232                  */
233                 goto dont_put_skbuff_in_hw;
234         }
235         if (unlikely
236             (skb->truesize !=
237              sizeof(*skb) + skb_end_pointer(skb) - skb->head)) {
238                 /*
239                    printk("TX buffer truesize has been changed\n");
240                  */
241                 goto dont_put_skbuff_in_hw;
242         }
243
244         /*
245          * We can use this buffer in the FPA.  We don't need the FAU
246          * update anymore
247          */
248         pko_command.s.reg0 = 0;
249         pko_command.s.dontfree = 0;
250
251         hw_buffer.s.back = (skb->data - fpa_head) >> 7;
252         *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
253
254         /*
255          * The skbuff will be reused without ever being freed. We must
256          * cleanup a bunch of core things.
257          */
258         dst_release(skb_dst(skb));
259         skb_dst_set(skb, NULL);
260 #ifdef CONFIG_XFRM
261         secpath_put(skb->sp);
262         skb->sp = NULL;
263 #endif
264         nf_reset(skb);
265
266 #ifdef CONFIG_NET_SCHED
267         skb->tc_index = 0;
268 #ifdef CONFIG_NET_CLS_ACT
269         skb->tc_verd = 0;
270 #endif /* CONFIG_NET_CLS_ACT */
271 #endif /* CONFIG_NET_SCHED */
272
273 dont_put_skbuff_in_hw:
274 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
275
276         /* Check if we can use the hardware checksumming */
277         if (USE_HW_TCPUDP_CHECKSUM && (skb->protocol == htons(ETH_P_IP)) &&
278             (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) &&
279             ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == 1 << 14))
280             && ((ip_hdr(skb)->protocol == IP_PROTOCOL_TCP)
281                 || (ip_hdr(skb)->protocol == IP_PROTOCOL_UDP))) {
282                 /* Use hardware checksum calc */
283                 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
284         }
285
286         if (USE_ASYNC_IOBDMA) {
287                 /* Get the number of skbuffs in use by the hardware */
288                 CVMX_SYNCIOBDMA;
289                 in_use = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
290                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
291         } else {
292                 /* Get the number of skbuffs in use by the hardware */
293                 in_use = cvmx_fau_fetch_and_add32(priv->fau + qos * 4, 1);
294                 buffers_to_free =
295                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
296         }
297
298         /*
299          * If we're sending faster than the receive can free them then
300          * don't do the HW free.
301          */
302         if ((buffers_to_free < -100) && !pko_command.s.dontfree) {
303                 pko_command.s.dontfree = 1;
304                 pko_command.s.reg0 = priv->fau + qos * 4;
305         }
306
307         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
308                                      CVMX_PKO_LOCK_CMD_QUEUE);
309
310         /* Drop this packet if we have too many already queued to the HW */
311         if (unlikely
312             (skb_queue_len(&priv->tx_free_list[qos]) >= MAX_OUT_QUEUE_DEPTH)) {
313                 /*
314                    DEBUGPRINT("%s: Tx dropped. Too many queued\n", dev->name);
315                  */
316                 dropped = 1;
317         }
318         /* Send the packet to the output queue */
319         else if (unlikely
320                  (cvmx_pko_send_packet_finish
321                   (priv->port, priv->queue + qos, pko_command, hw_buffer,
322                    CVMX_PKO_LOCK_CMD_QUEUE))) {
323                 DEBUGPRINT("%s: Failed to send the packet\n", dev->name);
324                 dropped = 1;
325         }
326
327         if (USE_ASYNC_IOBDMA) {
328                 /* Restore the scratch area */
329                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
330                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
331         }
332
333         if (unlikely(dropped)) {
334                 dev_kfree_skb_any(skb);
335                 cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
336                 priv->stats.tx_dropped++;
337         } else {
338                 if (USE_SKBUFFS_IN_HW) {
339                         /* Put this packet on the queue to be freed later */
340                         if (pko_command.s.dontfree)
341                                 skb_queue_tail(&priv->tx_free_list[qos], skb);
342                         else {
343                                 cvmx_fau_atomic_add32
344                                     (FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
345                                 cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
346                         }
347                 } else {
348                         /* Put this packet on the queue to be freed later */
349                         skb_queue_tail(&priv->tx_free_list[qos], skb);
350                 }
351         }
352
353         /* Free skbuffs not in use by the hardware, possibly two at a time */
354         if (skb_queue_len(&priv->tx_free_list[qos]) > in_use) {
355                 spin_lock(&priv->tx_free_list[qos].lock);
356                 /*
357                  * Check again now that we have the lock. It might
358                  * have changed.
359                  */
360                 if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
361                         dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
362                 if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
363                         dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
364                 spin_unlock(&priv->tx_free_list[qos].lock);
365         }
366
367         return 0;
368 }
369
370 /**
371  * Packet transmit to the POW
372  *
373  * @skb:    Packet to send
374  * @dev:    Device info structure
375  * Returns Always returns zero
376  */
377 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
378 {
379         struct octeon_ethernet *priv = netdev_priv(dev);
380         void *packet_buffer;
381         void *copy_location;
382
383         /* Get a work queue entry */
384         cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
385         if (unlikely(work == NULL)) {
386                 DEBUGPRINT("%s: Failed to allocate a work queue entry\n",
387                            dev->name);
388                 priv->stats.tx_dropped++;
389                 dev_kfree_skb(skb);
390                 return 0;
391         }
392
393         /* Get a packet buffer */
394         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
395         if (unlikely(packet_buffer == NULL)) {
396                 DEBUGPRINT("%s: Failed to allocate a packet buffer\n",
397                            dev->name);
398                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
399                 priv->stats.tx_dropped++;
400                 dev_kfree_skb(skb);
401                 return 0;
402         }
403
404         /*
405          * Calculate where we need to copy the data to. We need to
406          * leave 8 bytes for a next pointer (unused). We also need to
407          * include any configure skip. Then we need to align the IP
408          * packet src and dest into the same 64bit word. The below
409          * calculation may add a little extra, but that doesn't
410          * hurt.
411          */
412         copy_location = packet_buffer + sizeof(uint64_t);
413         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
414
415         /*
416          * We have to copy the packet since whoever processes this
417          * packet will free it to a hardware pool. We can't use the
418          * trick of counting outstanding packets like in
419          * cvm_oct_xmit.
420          */
421         memcpy(copy_location, skb->data, skb->len);
422
423         /*
424          * Fill in some of the work queue fields. We may need to add
425          * more if the software at the other end needs them.
426          */
427         work->hw_chksum = skb->csum;
428         work->len = skb->len;
429         work->ipprt = priv->port;
430         work->qos = priv->port & 0x7;
431         work->grp = pow_send_group;
432         work->tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
433         work->tag = pow_send_group;     /* FIXME */
434         /* Default to zero. Sets of zero later are commented out */
435         work->word2.u64 = 0;
436         work->word2.s.bufs = 1;
437         work->packet_ptr.u64 = 0;
438         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
439         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
440         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
441         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
442
443         if (skb->protocol == htons(ETH_P_IP)) {
444                 work->word2.s.ip_offset = 14;
445 #if 0
446                 work->word2.s.vlan_valid = 0;   /* FIXME */
447                 work->word2.s.vlan_cfi = 0;     /* FIXME */
448                 work->word2.s.vlan_id = 0;      /* FIXME */
449                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
450 #endif
451                 work->word2.s.tcp_or_udp =
452                     (ip_hdr(skb)->protocol == IP_PROTOCOL_TCP)
453                     || (ip_hdr(skb)->protocol == IP_PROTOCOL_UDP);
454 #if 0
455                 /* FIXME */
456                 work->word2.s.dec_ipsec = 0;
457                 /* We only support IPv4 right now */
458                 work->word2.s.is_v6 = 0;
459                 /* Hardware would set to zero */
460                 work->word2.s.software = 0;
461                 /* No error, packet is internal */
462                 work->word2.s.L4_error = 0;
463 #endif
464                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0)
465                                           || (ip_hdr(skb)->frag_off ==
466                                               1 << 14));
467 #if 0
468                 /* Assume Linux is sending a good packet */
469                 work->word2.s.IP_exc = 0;
470 #endif
471                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
472                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
473 #if 0
474                 /* This is an IP packet */
475                 work->word2.s.not_IP = 0;
476                 /* No error, packet is internal */
477                 work->word2.s.rcv_error = 0;
478                 /* No error, packet is internal */
479                 work->word2.s.err_code = 0;
480 #endif
481
482                 /*
483                  * When copying the data, include 4 bytes of the
484                  * ethernet header to align the same way hardware
485                  * does.
486                  */
487                 memcpy(work->packet_data, skb->data + 10,
488                        sizeof(work->packet_data));
489         } else {
490 #if 0
491                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
492                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
493                 work->word2.snoip.vlan_id = 0;  /* FIXME */
494                 work->word2.snoip.software = 0; /* Hardware would set to zero */
495 #endif
496                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
497                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
498                 work->word2.snoip.is_bcast =
499                     (skb->pkt_type == PACKET_BROADCAST);
500                 work->word2.snoip.is_mcast =
501                     (skb->pkt_type == PACKET_MULTICAST);
502                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
503 #if 0
504                 /* No error, packet is internal */
505                 work->word2.snoip.rcv_error = 0;
506                 /* No error, packet is internal */
507                 work->word2.snoip.err_code = 0;
508 #endif
509                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
510         }
511
512         /* Submit the packet to the POW */
513         cvmx_pow_work_submit(work, work->tag, work->tag_type, work->qos,
514                              work->grp);
515         priv->stats.tx_packets++;
516         priv->stats.tx_bytes += skb->len;
517         dev_kfree_skb(skb);
518         return 0;
519 }
520
521 /**
522  * Transmit a work queue entry out of the ethernet port. Both
523  * the work queue entry and the packet data can optionally be
524  * freed. The work will be freed on error as well.
525  *
526  * @dev:     Device to transmit out.
527  * @work_queue_entry:
528  *                Work queue entry to send
529  * @do_free: True if the work queue entry and packet data should be
530  *                freed. If false, neither will be freed.
531  * @qos:     Index into the queues for this port to transmit on. This
532  *                is used to implement QoS if their are multiple queues per
533  *                port. This parameter must be between 0 and the number of
534  *                queues per port minus 1. Values outside of this range will
535  *                be change to zero.
536  *
537  * Returns Zero on success, negative on failure.
538  */
539 int cvm_oct_transmit_qos(struct net_device *dev, void *work_queue_entry,
540                          int do_free, int qos)
541 {
542         unsigned long flags;
543         union cvmx_buf_ptr hw_buffer;
544         cvmx_pko_command_word0_t pko_command;
545         int dropped;
546         struct octeon_ethernet *priv = netdev_priv(dev);
547         cvmx_wqe_t *work = work_queue_entry;
548
549         if (!(dev->flags & IFF_UP)) {
550                 DEBUGPRINT("%s: Device not up\n", dev->name);
551                 if (do_free)
552                         cvm_oct_free_work(work);
553                 return -1;
554         }
555
556         /* The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to completely
557            remove "qos" in the event neither interface supports
558            multiple queues per port */
559         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
560             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
561                 if (qos <= 0)
562                         qos = 0;
563                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
564                         qos = 0;
565         } else
566                 qos = 0;
567
568         /* Start off assuming no drop */
569         dropped = 0;
570
571         local_irq_save(flags);
572         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
573                                      CVMX_PKO_LOCK_CMD_QUEUE);
574
575         /* Build the PKO buffer pointer */
576         hw_buffer.u64 = 0;
577         hw_buffer.s.addr = work->packet_ptr.s.addr;
578         hw_buffer.s.pool = CVMX_FPA_PACKET_POOL;
579         hw_buffer.s.size = CVMX_FPA_PACKET_POOL_SIZE;
580         hw_buffer.s.back = work->packet_ptr.s.back;
581
582         /* Build the PKO command */
583         pko_command.u64 = 0;
584         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
585         pko_command.s.dontfree = !do_free;
586         pko_command.s.segs = work->word2.s.bufs;
587         pko_command.s.total_bytes = work->len;
588
589         /* Check if we can use the hardware checksumming */
590         if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc))
591                 pko_command.s.ipoffp1 = 0;
592         else
593                 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
594
595         /* Send the packet to the output queue */
596         if (unlikely
597             (cvmx_pko_send_packet_finish
598              (priv->port, priv->queue + qos, pko_command, hw_buffer,
599               CVMX_PKO_LOCK_CMD_QUEUE))) {
600                 DEBUGPRINT("%s: Failed to send the packet\n", dev->name);
601                 dropped = -1;
602         }
603         local_irq_restore(flags);
604
605         if (unlikely(dropped)) {
606                 if (do_free)
607                         cvm_oct_free_work(work);
608                 priv->stats.tx_dropped++;
609         } else if (do_free)
610                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
611
612         return dropped;
613 }
614 EXPORT_SYMBOL(cvm_oct_transmit_qos);
615
616 /**
617  * This function frees all skb that are currenty queued for TX.
618  *
619  * @dev:    Device being shutdown
620  */
621 void cvm_oct_tx_shutdown(struct net_device *dev)
622 {
623         struct octeon_ethernet *priv = netdev_priv(dev);
624         unsigned long flags;
625         int qos;
626
627         for (qos = 0; qos < 16; qos++) {
628                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
629                 while (skb_queue_len(&priv->tx_free_list[qos]))
630                         dev_kfree_skb_any(__skb_dequeue
631                                           (&priv->tx_free_list[qos]));
632                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
633         }
634 }