Staging: octeon-ethernet: Fix race freeing transmit buffers.
[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-tx.h"
51 #include "ethernet-util.h"
52
53 #include "cvmx-wqe.h"
54 #include "cvmx-fau.h"
55 #include "cvmx-pko.h"
56 #include "cvmx-helper.h"
57
58 #include "cvmx-gmxx-defs.h"
59
60 /*
61  * You can define GET_SKBUFF_QOS() to override how the skbuff output
62  * function determines which output queue is used. The default
63  * implementation always uses the base queue for the port. If, for
64  * example, you wanted to use the skb->priority fieid, define
65  * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
66  */
67 #ifndef GET_SKBUFF_QOS
68 #define GET_SKBUFF_QOS(skb) 0
69 #endif
70
71 /**
72  * Packet transmit
73  *
74  * @skb:    Packet to send
75  * @dev:    Device info structure
76  * Returns Always returns zero
77  */
78 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
79 {
80         cvmx_pko_command_word0_t pko_command;
81         union cvmx_buf_ptr hw_buffer;
82         uint64_t old_scratch;
83         uint64_t old_scratch2;
84         int dropped;
85         int qos;
86         int queue_it_up;
87         struct octeon_ethernet *priv = netdev_priv(dev);
88         int32_t skb_to_free;
89         int32_t undo;
90         int32_t buffers_to_free;
91 #if REUSE_SKBUFFS_WITHOUT_FREE
92         unsigned char *fpa_head;
93 #endif
94
95         /*
96          * Prefetch the private data structure.  It is larger that one
97          * cache line.
98          */
99         prefetch(priv);
100
101         /* Start off assuming no drop */
102         dropped = 0;
103
104         /*
105          * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
106          * completely remove "qos" in the event neither interface
107          * supports multiple queues per port.
108          */
109         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
110             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
111                 qos = GET_SKBUFF_QOS(skb);
112                 if (qos <= 0)
113                         qos = 0;
114                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
115                         qos = 0;
116         } else
117                 qos = 0;
118
119         if (USE_ASYNC_IOBDMA) {
120                 /* Save scratch in case userspace is using it */
121                 CVMX_SYNCIOBDMA;
122                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
123                 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
124
125                 /*
126                  * Fetch and increment the number of packets to be
127                  * freed.
128                  */
129                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
130                                                FAU_NUM_PACKET_BUFFERS_TO_FREE,
131                                                0);
132                 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
133                                                priv->fau + qos * 4,
134                                                MAX_SKB_TO_FREE);
135         }
136
137         /*
138          * The CN3XXX series of parts has an errata (GMX-401) which
139          * causes the GMX block to hang if a collision occurs towards
140          * the end of a <68 byte packet. As a workaround for this, we
141          * pad packets to be 68 bytes whenever we are in half duplex
142          * mode. We don't handle the case of having a small packet but
143          * no room to add the padding.  The kernel should always give
144          * us at least a cache line
145          */
146         if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
147                 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
148                 int interface = INTERFACE(priv->port);
149                 int index = INDEX(priv->port);
150
151                 if (interface < 2) {
152                         /* We only need to pad packet in half duplex mode */
153                         gmx_prt_cfg.u64 =
154                             cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
155                         if (gmx_prt_cfg.s.duplex == 0) {
156                                 int add_bytes = 64 - skb->len;
157                                 if ((skb_tail_pointer(skb) + add_bytes) <=
158                                     skb_end_pointer(skb))
159                                         memset(__skb_put(skb, add_bytes), 0,
160                                                add_bytes);
161                         }
162                 }
163         }
164
165         /* Build the PKO buffer pointer */
166         hw_buffer.u64 = 0;
167         hw_buffer.s.addr = cvmx_ptr_to_phys(skb->data);
168         hw_buffer.s.pool = 0;
169         hw_buffer.s.size =
170             (unsigned long)skb_end_pointer(skb) - (unsigned long)skb->head;
171
172         /* Build the PKO command */
173         pko_command.u64 = 0;
174         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
175         pko_command.s.segs = 1;
176         pko_command.s.total_bytes = skb->len;
177         pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
178         pko_command.s.subone0 = 1;
179
180         pko_command.s.dontfree = 1;
181         pko_command.s.reg0 = priv->fau + qos * 4;
182         /*
183          * See if we can put this skb in the FPA pool. Any strange
184          * behavior from the Linux networking stack will most likely
185          * be caused by a bug in the following code. If some field is
186          * in use by the network stack and get carried over when a
187          * buffer is reused, bad thing may happen.  If in doubt and
188          * you dont need the absolute best performance, disable the
189          * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
190          * shown a 25% increase in performance under some loads.
191          */
192 #if REUSE_SKBUFFS_WITHOUT_FREE
193         fpa_head = skb->head + 128 - ((unsigned long)skb->head & 0x7f);
194         if (unlikely(skb->data < fpa_head)) {
195                 /*
196                  * printk("TX buffer beginning can't meet FPA
197                  * alignment constraints\n");
198                  */
199                 goto dont_put_skbuff_in_hw;
200         }
201         if (unlikely
202             ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
203                 /*
204                    printk("TX buffer isn't large enough for the FPA\n");
205                  */
206                 goto dont_put_skbuff_in_hw;
207         }
208         if (unlikely(skb_shared(skb))) {
209                 /*
210                    printk("TX buffer sharing data with someone else\n");
211                  */
212                 goto dont_put_skbuff_in_hw;
213         }
214         if (unlikely(skb_cloned(skb))) {
215                 /*
216                    printk("TX buffer has been cloned\n");
217                  */
218                 goto dont_put_skbuff_in_hw;
219         }
220         if (unlikely(skb_header_cloned(skb))) {
221                 /*
222                    printk("TX buffer header has been cloned\n");
223                  */
224                 goto dont_put_skbuff_in_hw;
225         }
226         if (unlikely(skb->destructor)) {
227                 /*
228                    printk("TX buffer has a destructor\n");
229                  */
230                 goto dont_put_skbuff_in_hw;
231         }
232         if (unlikely(skb_shinfo(skb)->nr_frags)) {
233                 /*
234                    printk("TX buffer has fragments\n");
235                  */
236                 goto dont_put_skbuff_in_hw;
237         }
238         if (unlikely
239             (skb->truesize !=
240              sizeof(*skb) + skb_end_pointer(skb) - skb->head)) {
241                 /*
242                    printk("TX buffer truesize has been changed\n");
243                  */
244                 goto dont_put_skbuff_in_hw;
245         }
246
247         /*
248          * We can use this buffer in the FPA.  We don't need the FAU
249          * update anymore
250          */
251         pko_command.s.reg0 = 0;
252         pko_command.s.dontfree = 0;
253
254         hw_buffer.s.back = (skb->data - fpa_head) >> 7;
255         *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
256
257         /*
258          * The skbuff will be reused without ever being freed. We must
259          * cleanup a bunch of core things.
260          */
261         dst_release(skb_dst(skb));
262         skb_dst_set(skb, NULL);
263 #ifdef CONFIG_XFRM
264         secpath_put(skb->sp);
265         skb->sp = NULL;
266 #endif
267         nf_reset(skb);
268
269 #ifdef CONFIG_NET_SCHED
270         skb->tc_index = 0;
271 #ifdef CONFIG_NET_CLS_ACT
272         skb->tc_verd = 0;
273 #endif /* CONFIG_NET_CLS_ACT */
274 #endif /* CONFIG_NET_SCHED */
275
276 dont_put_skbuff_in_hw:
277 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
278
279         /* Check if we can use the hardware checksumming */
280         if (USE_HW_TCPUDP_CHECKSUM && (skb->protocol == htons(ETH_P_IP)) &&
281             (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) &&
282             ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == 1 << 14))
283             && ((ip_hdr(skb)->protocol == IP_PROTOCOL_TCP)
284                 || (ip_hdr(skb)->protocol == IP_PROTOCOL_UDP))) {
285                 /* Use hardware checksum calc */
286                 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
287         }
288
289         if (USE_ASYNC_IOBDMA) {
290                 /* Get the number of skbuffs in use by the hardware */
291                 CVMX_SYNCIOBDMA;
292                 skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
293                 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
294         } else {
295                 /* Get the number of skbuffs in use by the hardware */
296                 skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
297                                                        MAX_SKB_TO_FREE);
298                 buffers_to_free =
299                     cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
300         }
301
302         /*
303          * We try to claim MAX_SKB_TO_FREE buffers.  If there were not
304          * that many available, we have to un-claim (undo) any that
305          * were in excess.  If skb_to_free is positive we will free
306          * that many buffers.
307          */
308         undo = skb_to_free > 0 ?
309                 MAX_SKB_TO_FREE : skb_to_free + MAX_SKB_TO_FREE;
310         if (undo > 0)
311                 cvmx_fau_atomic_add32(priv->fau+qos*4, -undo);
312         skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ?
313                 MAX_SKB_TO_FREE : -skb_to_free;
314
315         /*
316          * If we're sending faster than the receive can free them then
317          * don't do the HW free.
318          */
319         if ((buffers_to_free < -100) && !pko_command.s.dontfree) {
320                 pko_command.s.dontfree = 1;
321                 pko_command.s.reg0 = priv->fau + qos * 4;
322         }
323
324         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
325                                      CVMX_PKO_LOCK_CMD_QUEUE);
326
327         /* Drop this packet if we have too many already queued to the HW */
328         if (unlikely
329             (skb_queue_len(&priv->tx_free_list[qos]) >= MAX_OUT_QUEUE_DEPTH)) {
330                 /*
331                    DEBUGPRINT("%s: Tx dropped. Too many queued\n", dev->name);
332                  */
333                 dropped = 1;
334         }
335         /* Send the packet to the output queue */
336         else if (unlikely
337                  (cvmx_pko_send_packet_finish
338                   (priv->port, priv->queue + qos, pko_command, hw_buffer,
339                    CVMX_PKO_LOCK_CMD_QUEUE))) {
340                 DEBUGPRINT("%s: Failed to send the packet\n", dev->name);
341                 dropped = 1;
342         }
343
344         if (USE_ASYNC_IOBDMA) {
345                 /* Restore the scratch area */
346                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
347                 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
348         }
349
350         queue_it_up = 0;
351         if (unlikely(dropped)) {
352                 dev_kfree_skb_any(skb);
353                 priv->stats.tx_dropped++;
354         } else {
355                 if (USE_SKBUFFS_IN_HW) {
356                         /* Put this packet on the queue to be freed later */
357                         if (pko_command.s.dontfree)
358                                 queue_it_up = 1;
359                         else
360                                 cvmx_fau_atomic_add32
361                                     (FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
362                 } else {
363                         /* Put this packet on the queue to be freed later */
364                         queue_it_up = 1;
365                 }
366         }
367
368         if (queue_it_up) {
369                 spin_lock(&priv->tx_free_list[qos].lock);
370                 __skb_queue_tail(&priv->tx_free_list[qos], skb);
371                 cvm_oct_free_tx_skbs(priv, skb_to_free, qos, 0);
372                 spin_unlock(&priv->tx_free_list[qos].lock);
373         } else {
374                 cvm_oct_free_tx_skbs(priv, skb_to_free, qos, 1);
375         }
376
377         return 0;
378 }
379
380 /**
381  * Packet transmit to the POW
382  *
383  * @skb:    Packet to send
384  * @dev:    Device info structure
385  * Returns Always returns zero
386  */
387 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
388 {
389         struct octeon_ethernet *priv = netdev_priv(dev);
390         void *packet_buffer;
391         void *copy_location;
392
393         /* Get a work queue entry */
394         cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
395         if (unlikely(work == NULL)) {
396                 DEBUGPRINT("%s: Failed to allocate a work queue entry\n",
397                            dev->name);
398                 priv->stats.tx_dropped++;
399                 dev_kfree_skb(skb);
400                 return 0;
401         }
402
403         /* Get a packet buffer */
404         packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
405         if (unlikely(packet_buffer == NULL)) {
406                 DEBUGPRINT("%s: Failed to allocate a packet buffer\n",
407                            dev->name);
408                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
409                 priv->stats.tx_dropped++;
410                 dev_kfree_skb(skb);
411                 return 0;
412         }
413
414         /*
415          * Calculate where we need to copy the data to. We need to
416          * leave 8 bytes for a next pointer (unused). We also need to
417          * include any configure skip. Then we need to align the IP
418          * packet src and dest into the same 64bit word. The below
419          * calculation may add a little extra, but that doesn't
420          * hurt.
421          */
422         copy_location = packet_buffer + sizeof(uint64_t);
423         copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
424
425         /*
426          * We have to copy the packet since whoever processes this
427          * packet will free it to a hardware pool. We can't use the
428          * trick of counting outstanding packets like in
429          * cvm_oct_xmit.
430          */
431         memcpy(copy_location, skb->data, skb->len);
432
433         /*
434          * Fill in some of the work queue fields. We may need to add
435          * more if the software at the other end needs them.
436          */
437         work->hw_chksum = skb->csum;
438         work->len = skb->len;
439         work->ipprt = priv->port;
440         work->qos = priv->port & 0x7;
441         work->grp = pow_send_group;
442         work->tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
443         work->tag = pow_send_group;     /* FIXME */
444         /* Default to zero. Sets of zero later are commented out */
445         work->word2.u64 = 0;
446         work->word2.s.bufs = 1;
447         work->packet_ptr.u64 = 0;
448         work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
449         work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
450         work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
451         work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
452
453         if (skb->protocol == htons(ETH_P_IP)) {
454                 work->word2.s.ip_offset = 14;
455 #if 0
456                 work->word2.s.vlan_valid = 0;   /* FIXME */
457                 work->word2.s.vlan_cfi = 0;     /* FIXME */
458                 work->word2.s.vlan_id = 0;      /* FIXME */
459                 work->word2.s.dec_ipcomp = 0;   /* FIXME */
460 #endif
461                 work->word2.s.tcp_or_udp =
462                     (ip_hdr(skb)->protocol == IP_PROTOCOL_TCP)
463                     || (ip_hdr(skb)->protocol == IP_PROTOCOL_UDP);
464 #if 0
465                 /* FIXME */
466                 work->word2.s.dec_ipsec = 0;
467                 /* We only support IPv4 right now */
468                 work->word2.s.is_v6 = 0;
469                 /* Hardware would set to zero */
470                 work->word2.s.software = 0;
471                 /* No error, packet is internal */
472                 work->word2.s.L4_error = 0;
473 #endif
474                 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0)
475                                           || (ip_hdr(skb)->frag_off ==
476                                               1 << 14));
477 #if 0
478                 /* Assume Linux is sending a good packet */
479                 work->word2.s.IP_exc = 0;
480 #endif
481                 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
482                 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
483 #if 0
484                 /* This is an IP packet */
485                 work->word2.s.not_IP = 0;
486                 /* No error, packet is internal */
487                 work->word2.s.rcv_error = 0;
488                 /* No error, packet is internal */
489                 work->word2.s.err_code = 0;
490 #endif
491
492                 /*
493                  * When copying the data, include 4 bytes of the
494                  * ethernet header to align the same way hardware
495                  * does.
496                  */
497                 memcpy(work->packet_data, skb->data + 10,
498                        sizeof(work->packet_data));
499         } else {
500 #if 0
501                 work->word2.snoip.vlan_valid = 0;       /* FIXME */
502                 work->word2.snoip.vlan_cfi = 0; /* FIXME */
503                 work->word2.snoip.vlan_id = 0;  /* FIXME */
504                 work->word2.snoip.software = 0; /* Hardware would set to zero */
505 #endif
506                 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
507                 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
508                 work->word2.snoip.is_bcast =
509                     (skb->pkt_type == PACKET_BROADCAST);
510                 work->word2.snoip.is_mcast =
511                     (skb->pkt_type == PACKET_MULTICAST);
512                 work->word2.snoip.not_IP = 1;   /* IP was done up above */
513 #if 0
514                 /* No error, packet is internal */
515                 work->word2.snoip.rcv_error = 0;
516                 /* No error, packet is internal */
517                 work->word2.snoip.err_code = 0;
518 #endif
519                 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
520         }
521
522         /* Submit the packet to the POW */
523         cvmx_pow_work_submit(work, work->tag, work->tag_type, work->qos,
524                              work->grp);
525         priv->stats.tx_packets++;
526         priv->stats.tx_bytes += skb->len;
527         dev_kfree_skb(skb);
528         return 0;
529 }
530
531 /**
532  * Transmit a work queue entry out of the ethernet port. Both
533  * the work queue entry and the packet data can optionally be
534  * freed. The work will be freed on error as well.
535  *
536  * @dev:     Device to transmit out.
537  * @work_queue_entry:
538  *                Work queue entry to send
539  * @do_free: True if the work queue entry and packet data should be
540  *                freed. If false, neither will be freed.
541  * @qos:     Index into the queues for this port to transmit on. This
542  *                is used to implement QoS if their are multiple queues per
543  *                port. This parameter must be between 0 and the number of
544  *                queues per port minus 1. Values outside of this range will
545  *                be change to zero.
546  *
547  * Returns Zero on success, negative on failure.
548  */
549 int cvm_oct_transmit_qos(struct net_device *dev, void *work_queue_entry,
550                          int do_free, int qos)
551 {
552         unsigned long flags;
553         union cvmx_buf_ptr hw_buffer;
554         cvmx_pko_command_word0_t pko_command;
555         int dropped;
556         struct octeon_ethernet *priv = netdev_priv(dev);
557         cvmx_wqe_t *work = work_queue_entry;
558
559         if (!(dev->flags & IFF_UP)) {
560                 DEBUGPRINT("%s: Device not up\n", dev->name);
561                 if (do_free)
562                         cvm_oct_free_work(work);
563                 return -1;
564         }
565
566         /* The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to completely
567            remove "qos" in the event neither interface supports
568            multiple queues per port */
569         if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
570             (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
571                 if (qos <= 0)
572                         qos = 0;
573                 else if (qos >= cvmx_pko_get_num_queues(priv->port))
574                         qos = 0;
575         } else
576                 qos = 0;
577
578         /* Start off assuming no drop */
579         dropped = 0;
580
581         local_irq_save(flags);
582         cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
583                                      CVMX_PKO_LOCK_CMD_QUEUE);
584
585         /* Build the PKO buffer pointer */
586         hw_buffer.u64 = 0;
587         hw_buffer.s.addr = work->packet_ptr.s.addr;
588         hw_buffer.s.pool = CVMX_FPA_PACKET_POOL;
589         hw_buffer.s.size = CVMX_FPA_PACKET_POOL_SIZE;
590         hw_buffer.s.back = work->packet_ptr.s.back;
591
592         /* Build the PKO command */
593         pko_command.u64 = 0;
594         pko_command.s.n2 = 1;   /* Don't pollute L2 with the outgoing packet */
595         pko_command.s.dontfree = !do_free;
596         pko_command.s.segs = work->word2.s.bufs;
597         pko_command.s.total_bytes = work->len;
598
599         /* Check if we can use the hardware checksumming */
600         if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc))
601                 pko_command.s.ipoffp1 = 0;
602         else
603                 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
604
605         /* Send the packet to the output queue */
606         if (unlikely
607             (cvmx_pko_send_packet_finish
608              (priv->port, priv->queue + qos, pko_command, hw_buffer,
609               CVMX_PKO_LOCK_CMD_QUEUE))) {
610                 DEBUGPRINT("%s: Failed to send the packet\n", dev->name);
611                 dropped = -1;
612         }
613         local_irq_restore(flags);
614
615         if (unlikely(dropped)) {
616                 if (do_free)
617                         cvm_oct_free_work(work);
618                 priv->stats.tx_dropped++;
619         } else if (do_free)
620                 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
621
622         return dropped;
623 }
624 EXPORT_SYMBOL(cvm_oct_transmit_qos);
625
626 /**
627  * This function frees all skb that are currenty queued for TX.
628  *
629  * @dev:    Device being shutdown
630  */
631 void cvm_oct_tx_shutdown(struct net_device *dev)
632 {
633         struct octeon_ethernet *priv = netdev_priv(dev);
634         unsigned long flags;
635         int qos;
636
637         for (qos = 0; qos < 16; qos++) {
638                 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
639                 while (skb_queue_len(&priv->tx_free_list[qos]))
640                         dev_kfree_skb_any(__skb_dequeue
641                                           (&priv->tx_free_list[qos]));
642                 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
643         }
644 }