[go: nahoru, domu]

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