Linux Audio

Check our new training course

Loading...
v4.6
 
  1/*
  2 * This file is based on code from OCTEON SDK by Cavium Networks.
  3 *
  4 * Copyright (c) 2003-2010 Cavium Networks
  5 *
  6 * This file is free software; you can redistribute it and/or modify
  7 * it under the terms of the GNU General Public License, Version 2, as
  8 * published by the Free Software Foundation.
  9 */
 10
 11#include <linux/module.h>
 12#include <linux/kernel.h>
 13#include <linux/netdevice.h>
 14#include <linux/etherdevice.h>
 15#include <linux/ip.h>
 16#include <linux/ratelimit.h>
 17#include <linux/string.h>
 18#include <linux/interrupt.h>
 19#include <net/dst.h>
 20#ifdef CONFIG_XFRM
 21#include <linux/xfrm.h>
 22#include <net/xfrm.h>
 23#endif /* CONFIG_XFRM */
 24
 25#include <linux/atomic.h>
 
 26
 27#include <asm/octeon/octeon.h>
 28
 29#include "ethernet-defines.h"
 30#include "octeon-ethernet.h"
 31#include "ethernet-tx.h"
 32#include "ethernet-util.h"
 33
 34#include <asm/octeon/cvmx-wqe.h>
 35#include <asm/octeon/cvmx-fau.h>
 36#include <asm/octeon/cvmx-pip.h>
 37#include <asm/octeon/cvmx-pko.h>
 38#include <asm/octeon/cvmx-helper.h>
 39
 40#include <asm/octeon/cvmx-gmxx-defs.h>
 41
 42#define CVM_OCT_SKB_CB(skb)	((u64 *)((skb)->cb))
 43
 44/*
 45 * You can define GET_SKBUFF_QOS() to override how the skbuff output
 46 * function determines which output queue is used. The default
 47 * implementation always uses the base queue for the port. If, for
 48 * example, you wanted to use the skb->priority field, define
 49 * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
 50 */
 51#ifndef GET_SKBUFF_QOS
 52#define GET_SKBUFF_QOS(skb) 0
 53#endif
 54
 55static void cvm_oct_tx_do_cleanup(unsigned long arg);
 56static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
 57
 58/* Maximum number of SKBs to try to free per xmit packet. */
 59#define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
 60
 61static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
 62{
 63	int32_t undo;
 64
 65	undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free +
 66						   MAX_SKB_TO_FREE;
 67	if (undo > 0)
 68		cvmx_fau_atomic_add32(fau, -undo);
 69	skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE :
 70						       -skb_to_free;
 71	return skb_to_free;
 72}
 73
 74static void cvm_oct_kick_tx_poll_watchdog(void)
 75{
 76	union cvmx_ciu_timx ciu_timx;
 77
 78	ciu_timx.u64 = 0;
 79	ciu_timx.s.one_shot = 1;
 80	ciu_timx.s.len = cvm_oct_tx_poll_interval;
 81	cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
 82}
 83
 84static void cvm_oct_free_tx_skbs(struct net_device *dev)
 85{
 86	int32_t skb_to_free;
 87	int qos, queues_per_port;
 88	int total_freed = 0;
 89	int total_remaining = 0;
 90	unsigned long flags;
 91	struct octeon_ethernet *priv = netdev_priv(dev);
 92
 93	queues_per_port = cvmx_pko_get_num_queues(priv->port);
 94	/* Drain any pending packets in the free list */
 95	for (qos = 0; qos < queues_per_port; qos++) {
 96		if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
 97			continue;
 98		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 99						       MAX_SKB_TO_FREE);
100		skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
101							 priv->fau + qos * 4);
102		total_freed += skb_to_free;
103		if (skb_to_free > 0) {
104			struct sk_buff *to_free_list = NULL;
105
106			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
107			while (skb_to_free > 0) {
108				struct sk_buff *t;
109
110				t = __skb_dequeue(&priv->tx_free_list[qos]);
111				t->next = to_free_list;
112				to_free_list = t;
113				skb_to_free--;
114			}
115			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
116					       flags);
117			/* Do the actual freeing outside of the lock. */
118			while (to_free_list) {
119				struct sk_buff *t = to_free_list;
120
121				to_free_list = to_free_list->next;
122				dev_kfree_skb_any(t);
123			}
124		}
125		total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
126	}
127	if (total_remaining < MAX_OUT_QUEUE_DEPTH && netif_queue_stopped(dev))
128		netif_wake_queue(dev);
129	if (total_remaining)
130		cvm_oct_kick_tx_poll_watchdog();
131}
132
133/**
134 * cvm_oct_xmit - transmit a packet
135 * @skb:    Packet to send
136 * @dev:    Device info structure
137 *
138 * Returns Always returns NETDEV_TX_OK
139 */
140int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
141{
142	cvmx_pko_command_word0_t pko_command;
143	union cvmx_buf_ptr hw_buffer;
144	u64 old_scratch;
145	u64 old_scratch2;
146	int qos;
147	int i;
148	enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
149	struct octeon_ethernet *priv = netdev_priv(dev);
150	struct sk_buff *to_free_list;
151	int32_t skb_to_free;
152	int32_t buffers_to_free;
153	u32 total_to_clean;
154	unsigned long flags;
155#if REUSE_SKBUFFS_WITHOUT_FREE
156	unsigned char *fpa_head;
157#endif
158
159	/*
160	 * Prefetch the private data structure.  It is larger than the
161	 * one cache line.
162	 */
163	prefetch(priv);
164
165	/*
166	 * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
167	 * completely remove "qos" in the event neither interface
168	 * supports multiple queues per port.
169	 */
170	if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
171	    (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
172		qos = GET_SKBUFF_QOS(skb);
173		if (qos <= 0)
174			qos = 0;
175		else if (qos >= cvmx_pko_get_num_queues(priv->port))
176			qos = 0;
177	} else {
178		qos = 0;
179	}
180
181	if (USE_ASYNC_IOBDMA) {
182		/* Save scratch in case userspace is using it */
183		CVMX_SYNCIOBDMA;
184		old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
185		old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
186
187		/*
188		 * Fetch and increment the number of packets to be
189		 * freed.
190		 */
191		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
192					       FAU_NUM_PACKET_BUFFERS_TO_FREE,
193					       0);
194		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
195					       priv->fau + qos * 4,
196					       MAX_SKB_TO_FREE);
197	}
198
199	/*
200	 * We have space for 6 segment pointers, If there will be more
201	 * than that, we must linearize.
202	 */
203	if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
204		if (unlikely(__skb_linearize(skb))) {
205			queue_type = QUEUE_DROP;
206			if (USE_ASYNC_IOBDMA) {
207				/*
208				 * Get the number of skbuffs in use
209				 * by the hardware
210				 */
211				CVMX_SYNCIOBDMA;
212				skb_to_free =
213					cvmx_scratch_read64(CVMX_SCR_SCRATCH);
214			} else {
215				/*
216				 * Get the number of skbuffs in use
217				 * by the hardware
218				 */
219				skb_to_free = cvmx_fau_fetch_and_add32(
220					priv->fau + qos * 4, MAX_SKB_TO_FREE);
221			}
222			skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
223							priv->fau + qos * 4);
 
224			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
225			goto skip_xmit;
226		}
227	}
228
229	/*
230	 * The CN3XXX series of parts has an errata (GMX-401) which
231	 * causes the GMX block to hang if a collision occurs towards
232	 * the end of a <68 byte packet. As a workaround for this, we
233	 * pad packets to be 68 bytes whenever we are in half duplex
234	 * mode. We don't handle the case of having a small packet but
235	 * no room to add the padding.  The kernel should always give
236	 * us at least a cache line
237	 */
238	if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
239		union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
240		int interface = INTERFACE(priv->port);
241		int index = INDEX(priv->port);
242
243		if (interface < 2) {
244			/* We only need to pad packet in half duplex mode */
245			gmx_prt_cfg.u64 =
246			    cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
247			if (gmx_prt_cfg.s.duplex == 0) {
248				int add_bytes = 64 - skb->len;
249
250				if ((skb_tail_pointer(skb) + add_bytes) <=
251				    skb_end_pointer(skb))
252					memset(__skb_put(skb, add_bytes), 0,
253					       add_bytes);
254			}
255		}
256	}
257
258	/* Build the PKO command */
259	pko_command.u64 = 0;
260#ifdef __LITTLE_ENDIAN
261	pko_command.s.le = 1;
262#endif
263	pko_command.s.n2 = 1;	/* Don't pollute L2 with the outgoing packet */
264	pko_command.s.segs = 1;
265	pko_command.s.total_bytes = skb->len;
266	pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
267	pko_command.s.subone0 = 1;
268
269	pko_command.s.dontfree = 1;
270
271	/* Build the PKO buffer pointer */
272	hw_buffer.u64 = 0;
273	if (skb_shinfo(skb)->nr_frags == 0) {
274		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
275		hw_buffer.s.pool = 0;
276		hw_buffer.s.size = skb->len;
277	} else {
278		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
279		hw_buffer.s.pool = 0;
280		hw_buffer.s.size = skb_headlen(skb);
281		CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
282		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
283			struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i;
284
285			hw_buffer.s.addr = XKPHYS_TO_PHYS(
286				(u64)(page_address(fs->page.p) +
287				fs->page_offset));
288			hw_buffer.s.size = fs->size;
289			CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
290		}
291		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb));
292		hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1;
293		pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1;
294		pko_command.s.gather = 1;
295		goto dont_put_skbuff_in_hw;
296	}
297
298	/*
299	 * See if we can put this skb in the FPA pool. Any strange
300	 * behavior from the Linux networking stack will most likely
301	 * be caused by a bug in the following code. If some field is
302	 * in use by the network stack and gets carried over when a
303	 * buffer is reused, bad things may happen.  If in doubt and
304	 * you dont need the absolute best performance, disable the
305	 * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
306	 * shown a 25% increase in performance under some loads.
307	 */
308#if REUSE_SKBUFFS_WITHOUT_FREE
309	fpa_head = skb->head + 256 - ((unsigned long)skb->head & 0x7f);
310	if (unlikely(skb->data < fpa_head)) {
311		/* TX buffer beginning can't meet FPA alignment constraints */
312		goto dont_put_skbuff_in_hw;
313	}
314	if (unlikely
315	    ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
316		/* TX buffer isn't large enough for the FPA */
317		goto dont_put_skbuff_in_hw;
318	}
319	if (unlikely(skb_shared(skb))) {
320		/* TX buffer sharing data with someone else */
321		goto dont_put_skbuff_in_hw;
322	}
323	if (unlikely(skb_cloned(skb))) {
324		/* TX buffer has been cloned */
325		goto dont_put_skbuff_in_hw;
326	}
327	if (unlikely(skb_header_cloned(skb))) {
328		/* TX buffer header has been cloned */
329		goto dont_put_skbuff_in_hw;
330	}
331	if (unlikely(skb->destructor)) {
332		/* TX buffer has a destructor */
333		goto dont_put_skbuff_in_hw;
334	}
335	if (unlikely(skb_shinfo(skb)->nr_frags)) {
336		/* TX buffer has fragments */
337		goto dont_put_skbuff_in_hw;
338	}
339	if (unlikely
340	    (skb->truesize !=
341	     sizeof(*skb) + skb_end_offset(skb))) {
342		/* TX buffer truesize has been changed */
343		goto dont_put_skbuff_in_hw;
344	}
345
346	/*
347	 * We can use this buffer in the FPA.  We don't need the FAU
348	 * update anymore
349	 */
350	pko_command.s.dontfree = 0;
351
352	hw_buffer.s.back = ((unsigned long)skb->data >> 7) -
353			   ((unsigned long)fpa_head >> 7);
354
355	*(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
356
357	/*
358	 * The skbuff will be reused without ever being freed. We must
359	 * cleanup a bunch of core things.
360	 */
361	dst_release(skb_dst(skb));
362	skb_dst_set(skb, NULL);
363#ifdef CONFIG_XFRM
364	secpath_put(skb->sp);
365	skb->sp = NULL;
366#endif
367	nf_reset(skb);
368
369#ifdef CONFIG_NET_SCHED
370	skb->tc_index = 0;
371#ifdef CONFIG_NET_CLS_ACT
372	skb->tc_verd = 0;
373#endif /* CONFIG_NET_CLS_ACT */
374#endif /* CONFIG_NET_SCHED */
375#endif /* REUSE_SKBUFFS_WITHOUT_FREE */
376
377dont_put_skbuff_in_hw:
378
379	/* Check if we can use the hardware checksumming */
380	if ((skb->protocol == htons(ETH_P_IP)) &&
381	    (ip_hdr(skb)->version == 4) &&
382	    (ip_hdr(skb)->ihl == 5) &&
383	    ((ip_hdr(skb)->frag_off == 0) ||
384	     (ip_hdr(skb)->frag_off == htons(1 << 14))) &&
385	    ((ip_hdr(skb)->protocol == IPPROTO_TCP) ||
386	     (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
387		/* Use hardware checksum calc */
388		pko_command.s.ipoffp1 = skb_network_offset(skb) + 1;
389	}
390
391	if (USE_ASYNC_IOBDMA) {
392		/* Get the number of skbuffs in use by the hardware */
393		CVMX_SYNCIOBDMA;
394		skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
395		buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
396	} else {
397		/* Get the number of skbuffs in use by the hardware */
398		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
399						       MAX_SKB_TO_FREE);
400		buffers_to_free =
401		    cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
402	}
403
404	skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
405						priv->fau + qos * 4);
406
407	/*
408	 * If we're sending faster than the receive can free them then
409	 * don't do the HW free.
410	 */
411	if ((buffers_to_free < -100) && !pko_command.s.dontfree)
412		pko_command.s.dontfree = 1;
413
414	if (pko_command.s.dontfree) {
415		queue_type = QUEUE_CORE;
416		pko_command.s.reg0 = priv->fau + qos * 4;
417	} else {
418		queue_type = QUEUE_HW;
419	}
420	if (USE_ASYNC_IOBDMA)
421		cvmx_fau_async_fetch_and_add32(
422				CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
423
424	spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
425
426	/* Drop this packet if we have too many already queued to the HW */
427	if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >=
428		     MAX_OUT_QUEUE_DEPTH)) {
429		if (dev->tx_queue_len != 0) {
430			/* Drop the lock when notifying the core.  */
431			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
432					       flags);
433			netif_stop_queue(dev);
434			spin_lock_irqsave(&priv->tx_free_list[qos].lock,
435					  flags);
436		} else {
437			/* If not using normal queueing.  */
438			queue_type = QUEUE_DROP;
439			goto skip_xmit;
440		}
441	}
442
443	cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
444				     CVMX_PKO_LOCK_NONE);
445
446	/* Send the packet to the output queue */
447	if (unlikely(cvmx_pko_send_packet_finish(priv->port,
448						 priv->queue + qos,
449						 pko_command, hw_buffer,
450						 CVMX_PKO_LOCK_NONE))) {
451		printk_ratelimited("%s: Failed to send the packet\n",
452				   dev->name);
453		queue_type = QUEUE_DROP;
454	}
455skip_xmit:
456	to_free_list = NULL;
457
458	switch (queue_type) {
459	case QUEUE_DROP:
460		skb->next = to_free_list;
461		to_free_list = skb;
462		priv->stats.tx_dropped++;
463		break;
464	case QUEUE_HW:
465		cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
466		break;
467	case QUEUE_CORE:
468		__skb_queue_tail(&priv->tx_free_list[qos], skb);
469		break;
470	default:
471		BUG();
472	}
473
474	while (skb_to_free > 0) {
475		struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
476
477		t->next = to_free_list;
478		to_free_list = t;
479		skb_to_free--;
480	}
481
482	spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
483
484	/* Do the actual freeing outside of the lock. */
485	while (to_free_list) {
486		struct sk_buff *t = to_free_list;
487
488		to_free_list = to_free_list->next;
489		dev_kfree_skb_any(t);
490	}
491
492	if (USE_ASYNC_IOBDMA) {
493		CVMX_SYNCIOBDMA;
494		total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
495		/* Restore the scratch area */
496		cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
497		cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
498	} else {
499		total_to_clean = cvmx_fau_fetch_and_add32(
500						FAU_TOTAL_TX_TO_CLEAN, 1);
501	}
502
503	if (total_to_clean & 0x3ff) {
504		/*
505		 * Schedule the cleanup tasklet every 1024 packets for
506		 * the pathological case of high traffic on one port
507		 * delaying clean up of packets on a different port
508		 * that is blocked waiting for the cleanup.
509		 */
510		tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
511	}
512
513	cvm_oct_kick_tx_poll_watchdog();
514
515	return NETDEV_TX_OK;
516}
517
518/**
519 * cvm_oct_xmit_pow - transmit a packet to the POW
520 * @skb:    Packet to send
521 * @dev:    Device info structure
522
523 * Returns Always returns zero
524 */
525int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
526{
527	struct octeon_ethernet *priv = netdev_priv(dev);
528	void *packet_buffer;
529	void *copy_location;
530
531	/* Get a work queue entry */
532	cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
533
534	if (unlikely(!work)) {
535		printk_ratelimited("%s: Failed to allocate a work queue entry\n",
536				   dev->name);
537		priv->stats.tx_dropped++;
538		dev_kfree_skb_any(skb);
539		return 0;
540	}
541
542	/* Get a packet buffer */
543	packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
544	if (unlikely(!packet_buffer)) {
545		printk_ratelimited("%s: Failed to allocate a packet buffer\n",
546				   dev->name);
547		cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
548		priv->stats.tx_dropped++;
549		dev_kfree_skb_any(skb);
550		return 0;
551	}
552
553	/*
554	 * Calculate where we need to copy the data to. We need to
555	 * leave 8 bytes for a next pointer (unused). We also need to
556	 * include any configure skip. Then we need to align the IP
557	 * packet src and dest into the same 64bit word. The below
558	 * calculation may add a little extra, but that doesn't
559	 * hurt.
560	 */
561	copy_location = packet_buffer + sizeof(u64);
562	copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
563
564	/*
565	 * We have to copy the packet since whoever processes this
566	 * packet will free it to a hardware pool. We can't use the
567	 * trick of counting outstanding packets like in
568	 * cvm_oct_xmit.
569	 */
570	memcpy(copy_location, skb->data, skb->len);
571
572	/*
573	 * Fill in some of the work queue fields. We may need to add
574	 * more if the software at the other end needs them.
575	 */
576	if (!OCTEON_IS_MODEL(OCTEON_CN68XX))
577		work->word0.pip.cn38xx.hw_chksum = skb->csum;
578	work->word1.len = skb->len;
579	cvmx_wqe_set_port(work, priv->port);
580	cvmx_wqe_set_qos(work, priv->port & 0x7);
581	cvmx_wqe_set_grp(work, pow_send_group);
582	work->word1.tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
583	work->word1.tag = pow_send_group;	/* FIXME */
584	/* Default to zero. Sets of zero later are commented out */
585	work->word2.u64 = 0;
586	work->word2.s.bufs = 1;
587	work->packet_ptr.u64 = 0;
588	work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
589	work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
590	work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
591	work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
592
593	if (skb->protocol == htons(ETH_P_IP)) {
594		work->word2.s.ip_offset = 14;
595#if 0
596		work->word2.s.vlan_valid = 0;	/* FIXME */
597		work->word2.s.vlan_cfi = 0;	/* FIXME */
598		work->word2.s.vlan_id = 0;	/* FIXME */
599		work->word2.s.dec_ipcomp = 0;	/* FIXME */
600#endif
601		work->word2.s.tcp_or_udp =
602		    (ip_hdr(skb)->protocol == IPPROTO_TCP) ||
603		    (ip_hdr(skb)->protocol == IPPROTO_UDP);
604#if 0
605		/* FIXME */
606		work->word2.s.dec_ipsec = 0;
607		/* We only support IPv4 right now */
608		work->word2.s.is_v6 = 0;
609		/* Hardware would set to zero */
610		work->word2.s.software = 0;
611		/* No error, packet is internal */
612		work->word2.s.L4_error = 0;
613#endif
614		work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0) ||
615					  (ip_hdr(skb)->frag_off ==
616					      1 << 14));
617#if 0
618		/* Assume Linux is sending a good packet */
619		work->word2.s.IP_exc = 0;
620#endif
621		work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
622		work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
623#if 0
624		/* This is an IP packet */
625		work->word2.s.not_IP = 0;
626		/* No error, packet is internal */
627		work->word2.s.rcv_error = 0;
628		/* No error, packet is internal */
629		work->word2.s.err_code = 0;
630#endif
631
632		/*
633		 * When copying the data, include 4 bytes of the
634		 * ethernet header to align the same way hardware
635		 * does.
636		 */
637		memcpy(work->packet_data, skb->data + 10,
638		       sizeof(work->packet_data));
639	} else {
640#if 0
641		work->word2.snoip.vlan_valid = 0;	/* FIXME */
642		work->word2.snoip.vlan_cfi = 0;	/* FIXME */
643		work->word2.snoip.vlan_id = 0;	/* FIXME */
644		work->word2.snoip.software = 0;	/* Hardware would set to zero */
645#endif
646		work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
647		work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
648		work->word2.snoip.is_bcast =
649		    (skb->pkt_type == PACKET_BROADCAST);
650		work->word2.snoip.is_mcast =
651		    (skb->pkt_type == PACKET_MULTICAST);
652		work->word2.snoip.not_IP = 1;	/* IP was done up above */
653#if 0
654		/* No error, packet is internal */
655		work->word2.snoip.rcv_error = 0;
656		/* No error, packet is internal */
657		work->word2.snoip.err_code = 0;
658#endif
659		memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
660	}
661
662	/* Submit the packet to the POW */
663	cvmx_pow_work_submit(work, work->word1.tag, work->word1.tag_type,
664			     cvmx_wqe_get_qos(work), cvmx_wqe_get_grp(work));
665	priv->stats.tx_packets++;
666	priv->stats.tx_bytes += skb->len;
667	dev_consume_skb_any(skb);
668	return 0;
669}
670
671/**
672 * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
673 * @dev:    Device being shutdown
674 *
675 */
676void cvm_oct_tx_shutdown_dev(struct net_device *dev)
677{
678	struct octeon_ethernet *priv = netdev_priv(dev);
679	unsigned long flags;
680	int qos;
681
682	for (qos = 0; qos < 16; qos++) {
683		spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
684		while (skb_queue_len(&priv->tx_free_list[qos]))
685			dev_kfree_skb_any(__skb_dequeue
686					  (&priv->tx_free_list[qos]));
687		spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
688	}
689}
690
691static void cvm_oct_tx_do_cleanup(unsigned long arg)
692{
693	int port;
694
695	for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
696		if (cvm_oct_device[port]) {
697			struct net_device *dev = cvm_oct_device[port];
698
699			cvm_oct_free_tx_skbs(dev);
700		}
701	}
702}
703
704static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
705{
706	/* Disable the interrupt.  */
707	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
708	/* Do the work in the tasklet.  */
709	tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
710	return IRQ_HANDLED;
711}
712
713void cvm_oct_tx_initialize(void)
714{
715	int i;
716
717	/* Disable the interrupt.  */
718	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
719	/* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
720	i = request_irq(OCTEON_IRQ_TIMER1,
721			cvm_oct_tx_cleanup_watchdog, 0,
722			"Ethernet", cvm_oct_device);
723
724	if (i)
725		panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
726}
727
728void cvm_oct_tx_shutdown(void)
729{
730	/* Free the interrupt handler */
731	free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
732}
v4.17
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * This file is based on code from OCTEON SDK by Cavium Networks.
  4 *
  5 * Copyright (c) 2003-2010 Cavium Networks
 
 
 
 
  6 */
  7
  8#include <linux/module.h>
  9#include <linux/kernel.h>
 10#include <linux/netdevice.h>
 11#include <linux/etherdevice.h>
 12#include <linux/ip.h>
 13#include <linux/ratelimit.h>
 14#include <linux/string.h>
 15#include <linux/interrupt.h>
 16#include <net/dst.h>
 17#ifdef CONFIG_XFRM
 18#include <linux/xfrm.h>
 19#include <net/xfrm.h>
 20#endif /* CONFIG_XFRM */
 21
 22#include <linux/atomic.h>
 23#include <net/sch_generic.h>
 24
 25#include <asm/octeon/octeon.h>
 26
 27#include "ethernet-defines.h"
 28#include "octeon-ethernet.h"
 29#include "ethernet-tx.h"
 30#include "ethernet-util.h"
 31
 32#include <asm/octeon/cvmx-wqe.h>
 33#include <asm/octeon/cvmx-fau.h>
 34#include <asm/octeon/cvmx-pip.h>
 35#include <asm/octeon/cvmx-pko.h>
 36#include <asm/octeon/cvmx-helper.h>
 37
 38#include <asm/octeon/cvmx-gmxx-defs.h>
 39
 40#define CVM_OCT_SKB_CB(skb)	((u64 *)((skb)->cb))
 41
 42/*
 43 * You can define GET_SKBUFF_QOS() to override how the skbuff output
 44 * function determines which output queue is used. The default
 45 * implementation always uses the base queue for the port. If, for
 46 * example, you wanted to use the skb->priority field, define
 47 * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
 48 */
 49#ifndef GET_SKBUFF_QOS
 50#define GET_SKBUFF_QOS(skb) 0
 51#endif
 52
 53static void cvm_oct_tx_do_cleanup(unsigned long arg);
 54static DECLARE_TASKLET(cvm_oct_tx_cleanup_tasklet, cvm_oct_tx_do_cleanup, 0);
 55
 56/* Maximum number of SKBs to try to free per xmit packet. */
 57#define MAX_SKB_TO_FREE (MAX_OUT_QUEUE_DEPTH * 2)
 58
 59static inline int cvm_oct_adjust_skb_to_free(int skb_to_free, int fau)
 60{
 61	int undo;
 62
 63	undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free +
 64						   MAX_SKB_TO_FREE;
 65	if (undo > 0)
 66		cvmx_fau_atomic_add32(fau, -undo);
 67	skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE :
 68						       -skb_to_free;
 69	return skb_to_free;
 70}
 71
 72static void cvm_oct_kick_tx_poll_watchdog(void)
 73{
 74	union cvmx_ciu_timx ciu_timx;
 75
 76	ciu_timx.u64 = 0;
 77	ciu_timx.s.one_shot = 1;
 78	ciu_timx.s.len = cvm_oct_tx_poll_interval;
 79	cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
 80}
 81
 82static void cvm_oct_free_tx_skbs(struct net_device *dev)
 83{
 84	int skb_to_free;
 85	int qos, queues_per_port;
 86	int total_freed = 0;
 87	int total_remaining = 0;
 88	unsigned long flags;
 89	struct octeon_ethernet *priv = netdev_priv(dev);
 90
 91	queues_per_port = cvmx_pko_get_num_queues(priv->port);
 92	/* Drain any pending packets in the free list */
 93	for (qos = 0; qos < queues_per_port; qos++) {
 94		if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
 95			continue;
 96		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
 97						       MAX_SKB_TO_FREE);
 98		skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
 99							 priv->fau + qos * 4);
100		total_freed += skb_to_free;
101		if (skb_to_free > 0) {
102			struct sk_buff *to_free_list = NULL;
103
104			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
105			while (skb_to_free > 0) {
106				struct sk_buff *t;
107
108				t = __skb_dequeue(&priv->tx_free_list[qos]);
109				t->next = to_free_list;
110				to_free_list = t;
111				skb_to_free--;
112			}
113			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
114					       flags);
115			/* Do the actual freeing outside of the lock. */
116			while (to_free_list) {
117				struct sk_buff *t = to_free_list;
118
119				to_free_list = to_free_list->next;
120				dev_kfree_skb_any(t);
121			}
122		}
123		total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
124	}
125	if (total_remaining < MAX_OUT_QUEUE_DEPTH && netif_queue_stopped(dev))
126		netif_wake_queue(dev);
127	if (total_remaining)
128		cvm_oct_kick_tx_poll_watchdog();
129}
130
131/**
132 * cvm_oct_xmit - transmit a packet
133 * @skb:    Packet to send
134 * @dev:    Device info structure
135 *
136 * Returns Always returns NETDEV_TX_OK
137 */
138int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
139{
140	cvmx_pko_command_word0_t pko_command;
141	union cvmx_buf_ptr hw_buffer;
142	u64 old_scratch;
143	u64 old_scratch2;
144	int qos;
145	int i;
146	enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
147	struct octeon_ethernet *priv = netdev_priv(dev);
148	struct sk_buff *to_free_list;
149	int skb_to_free;
150	int buffers_to_free;
151	u32 total_to_clean;
152	unsigned long flags;
153#if REUSE_SKBUFFS_WITHOUT_FREE
154	unsigned char *fpa_head;
155#endif
156
157	/*
158	 * Prefetch the private data structure.  It is larger than the
159	 * one cache line.
160	 */
161	prefetch(priv);
162
163	/*
164	 * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
165	 * completely remove "qos" in the event neither interface
166	 * supports multiple queues per port.
167	 */
168	if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
169	    (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
170		qos = GET_SKBUFF_QOS(skb);
171		if (qos <= 0)
172			qos = 0;
173		else if (qos >= cvmx_pko_get_num_queues(priv->port))
174			qos = 0;
175	} else {
176		qos = 0;
177	}
178
179	if (USE_ASYNC_IOBDMA) {
180		/* Save scratch in case userspace is using it */
181		CVMX_SYNCIOBDMA;
182		old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
183		old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
184
185		/*
186		 * Fetch and increment the number of packets to be
187		 * freed.
188		 */
189		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
190					       FAU_NUM_PACKET_BUFFERS_TO_FREE,
191					       0);
192		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
193					       priv->fau + qos * 4,
194					       MAX_SKB_TO_FREE);
195	}
196
197	/*
198	 * We have space for 6 segment pointers, If there will be more
199	 * than that, we must linearize.
200	 */
201	if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
202		if (unlikely(__skb_linearize(skb))) {
203			queue_type = QUEUE_DROP;
204			if (USE_ASYNC_IOBDMA) {
205				/*
206				 * Get the number of skbuffs in use
207				 * by the hardware
208				 */
209				CVMX_SYNCIOBDMA;
210				skb_to_free =
211					cvmx_scratch_read64(CVMX_SCR_SCRATCH);
212			} else {
213				/*
214				 * Get the number of skbuffs in use
215				 * by the hardware
216				 */
217				skb_to_free = cvmx_fau_fetch_and_add32(
218					priv->fau + qos * 4, MAX_SKB_TO_FREE);
219			}
220			skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
221								 priv->fau +
222								 qos * 4);
223			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
224			goto skip_xmit;
225		}
226	}
227
228	/*
229	 * The CN3XXX series of parts has an errata (GMX-401) which
230	 * causes the GMX block to hang if a collision occurs towards
231	 * the end of a <68 byte packet. As a workaround for this, we
232	 * pad packets to be 68 bytes whenever we are in half duplex
233	 * mode. We don't handle the case of having a small packet but
234	 * no room to add the padding.  The kernel should always give
235	 * us at least a cache line
236	 */
237	if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
238		union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
239		int interface = INTERFACE(priv->port);
240		int index = INDEX(priv->port);
241
242		if (interface < 2) {
243			/* We only need to pad packet in half duplex mode */
244			gmx_prt_cfg.u64 =
245			    cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
246			if (gmx_prt_cfg.s.duplex == 0) {
247				int add_bytes = 64 - skb->len;
248
249				if ((skb_tail_pointer(skb) + add_bytes) <=
250				    skb_end_pointer(skb))
251					__skb_put_zero(skb, add_bytes);
 
252			}
253		}
254	}
255
256	/* Build the PKO command */
257	pko_command.u64 = 0;
258#ifdef __LITTLE_ENDIAN
259	pko_command.s.le = 1;
260#endif
261	pko_command.s.n2 = 1;	/* Don't pollute L2 with the outgoing packet */
262	pko_command.s.segs = 1;
263	pko_command.s.total_bytes = skb->len;
264	pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
265	pko_command.s.subone0 = 1;
266
267	pko_command.s.dontfree = 1;
268
269	/* Build the PKO buffer pointer */
270	hw_buffer.u64 = 0;
271	if (skb_shinfo(skb)->nr_frags == 0) {
272		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
273		hw_buffer.s.pool = 0;
274		hw_buffer.s.size = skb->len;
275	} else {
276		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
277		hw_buffer.s.pool = 0;
278		hw_buffer.s.size = skb_headlen(skb);
279		CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
280		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
281			struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i;
282
283			hw_buffer.s.addr = XKPHYS_TO_PHYS(
284				(u64)(page_address(fs->page.p) +
285				fs->page_offset));
286			hw_buffer.s.size = fs->size;
287			CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
288		}
289		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb));
290		hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1;
291		pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1;
292		pko_command.s.gather = 1;
293		goto dont_put_skbuff_in_hw;
294	}
295
296	/*
297	 * See if we can put this skb in the FPA pool. Any strange
298	 * behavior from the Linux networking stack will most likely
299	 * be caused by a bug in the following code. If some field is
300	 * in use by the network stack and gets carried over when a
301	 * buffer is reused, bad things may happen.  If in doubt and
302	 * you dont need the absolute best performance, disable the
303	 * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
304	 * shown a 25% increase in performance under some loads.
305	 */
306#if REUSE_SKBUFFS_WITHOUT_FREE
307	fpa_head = skb->head + 256 - ((unsigned long)skb->head & 0x7f);
308	if (unlikely(skb->data < fpa_head)) {
309		/* TX buffer beginning can't meet FPA alignment constraints */
310		goto dont_put_skbuff_in_hw;
311	}
312	if (unlikely
313	    ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
314		/* TX buffer isn't large enough for the FPA */
315		goto dont_put_skbuff_in_hw;
316	}
317	if (unlikely(skb_shared(skb))) {
318		/* TX buffer sharing data with someone else */
319		goto dont_put_skbuff_in_hw;
320	}
321	if (unlikely(skb_cloned(skb))) {
322		/* TX buffer has been cloned */
323		goto dont_put_skbuff_in_hw;
324	}
325	if (unlikely(skb_header_cloned(skb))) {
326		/* TX buffer header has been cloned */
327		goto dont_put_skbuff_in_hw;
328	}
329	if (unlikely(skb->destructor)) {
330		/* TX buffer has a destructor */
331		goto dont_put_skbuff_in_hw;
332	}
333	if (unlikely(skb_shinfo(skb)->nr_frags)) {
334		/* TX buffer has fragments */
335		goto dont_put_skbuff_in_hw;
336	}
337	if (unlikely
338	    (skb->truesize !=
339	     sizeof(*skb) + skb_end_offset(skb))) {
340		/* TX buffer truesize has been changed */
341		goto dont_put_skbuff_in_hw;
342	}
343
344	/*
345	 * We can use this buffer in the FPA.  We don't need the FAU
346	 * update anymore
347	 */
348	pko_command.s.dontfree = 0;
349
350	hw_buffer.s.back = ((unsigned long)skb->data >> 7) -
351			   ((unsigned long)fpa_head >> 7);
352
353	*(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
354
355	/*
356	 * The skbuff will be reused without ever being freed. We must
357	 * cleanup a bunch of core things.
358	 */
359	dst_release(skb_dst(skb));
360	skb_dst_set(skb, NULL);
361#ifdef CONFIG_XFRM
362	secpath_put(skb->sp);
363	skb->sp = NULL;
364#endif
365	nf_reset(skb);
366
367#ifdef CONFIG_NET_SCHED
368	skb->tc_index = 0;
369	skb_reset_tc(skb);
 
 
370#endif /* CONFIG_NET_SCHED */
371#endif /* REUSE_SKBUFFS_WITHOUT_FREE */
372
373dont_put_skbuff_in_hw:
374
375	/* Check if we can use the hardware checksumming */
376	if ((skb->protocol == htons(ETH_P_IP)) &&
377	    (ip_hdr(skb)->version == 4) &&
378	    (ip_hdr(skb)->ihl == 5) &&
379	    ((ip_hdr(skb)->frag_off == 0) ||
380	     (ip_hdr(skb)->frag_off == htons(1 << 14))) &&
381	    ((ip_hdr(skb)->protocol == IPPROTO_TCP) ||
382	     (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
383		/* Use hardware checksum calc */
384		pko_command.s.ipoffp1 = skb_network_offset(skb) + 1;
385	}
386
387	if (USE_ASYNC_IOBDMA) {
388		/* Get the number of skbuffs in use by the hardware */
389		CVMX_SYNCIOBDMA;
390		skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
391		buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
392	} else {
393		/* Get the number of skbuffs in use by the hardware */
394		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
395						       MAX_SKB_TO_FREE);
396		buffers_to_free =
397		    cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
398	}
399
400	skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
401						 priv->fau + qos * 4);
402
403	/*
404	 * If we're sending faster than the receive can free them then
405	 * don't do the HW free.
406	 */
407	if ((buffers_to_free < -100) && !pko_command.s.dontfree)
408		pko_command.s.dontfree = 1;
409
410	if (pko_command.s.dontfree) {
411		queue_type = QUEUE_CORE;
412		pko_command.s.reg0 = priv->fau + qos * 4;
413	} else {
414		queue_type = QUEUE_HW;
415	}
416	if (USE_ASYNC_IOBDMA)
417		cvmx_fau_async_fetch_and_add32(
418				CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
419
420	spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
421
422	/* Drop this packet if we have too many already queued to the HW */
423	if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >=
424		     MAX_OUT_QUEUE_DEPTH)) {
425		if (dev->tx_queue_len != 0) {
426			/* Drop the lock when notifying the core.  */
427			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
428					       flags);
429			netif_stop_queue(dev);
430			spin_lock_irqsave(&priv->tx_free_list[qos].lock,
431					  flags);
432		} else {
433			/* If not using normal queueing.  */
434			queue_type = QUEUE_DROP;
435			goto skip_xmit;
436		}
437	}
438
439	cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
440				     CVMX_PKO_LOCK_NONE);
441
442	/* Send the packet to the output queue */
443	if (unlikely(cvmx_pko_send_packet_finish(priv->port,
444						 priv->queue + qos,
445						 pko_command, hw_buffer,
446						 CVMX_PKO_LOCK_NONE))) {
447		printk_ratelimited("%s: Failed to send the packet\n",
448				   dev->name);
449		queue_type = QUEUE_DROP;
450	}
451skip_xmit:
452	to_free_list = NULL;
453
454	switch (queue_type) {
455	case QUEUE_DROP:
456		skb->next = to_free_list;
457		to_free_list = skb;
458		dev->stats.tx_dropped++;
459		break;
460	case QUEUE_HW:
461		cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
462		break;
463	case QUEUE_CORE:
464		__skb_queue_tail(&priv->tx_free_list[qos], skb);
465		break;
466	default:
467		BUG();
468	}
469
470	while (skb_to_free > 0) {
471		struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
472
473		t->next = to_free_list;
474		to_free_list = t;
475		skb_to_free--;
476	}
477
478	spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
479
480	/* Do the actual freeing outside of the lock. */
481	while (to_free_list) {
482		struct sk_buff *t = to_free_list;
483
484		to_free_list = to_free_list->next;
485		dev_kfree_skb_any(t);
486	}
487
488	if (USE_ASYNC_IOBDMA) {
489		CVMX_SYNCIOBDMA;
490		total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
491		/* Restore the scratch area */
492		cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
493		cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
494	} else {
495		total_to_clean = cvmx_fau_fetch_and_add32(
496						FAU_TOTAL_TX_TO_CLEAN, 1);
497	}
498
499	if (total_to_clean & 0x3ff) {
500		/*
501		 * Schedule the cleanup tasklet every 1024 packets for
502		 * the pathological case of high traffic on one port
503		 * delaying clean up of packets on a different port
504		 * that is blocked waiting for the cleanup.
505		 */
506		tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
507	}
508
509	cvm_oct_kick_tx_poll_watchdog();
510
511	return NETDEV_TX_OK;
512}
513
514/**
515 * cvm_oct_xmit_pow - transmit a packet to the POW
516 * @skb:    Packet to send
517 * @dev:    Device info structure
518
519 * Returns Always returns zero
520 */
521int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
522{
523	struct octeon_ethernet *priv = netdev_priv(dev);
524	void *packet_buffer;
525	void *copy_location;
526
527	/* Get a work queue entry */
528	cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
529
530	if (unlikely(!work)) {
531		printk_ratelimited("%s: Failed to allocate a work queue entry\n",
532				   dev->name);
533		dev->stats.tx_dropped++;
534		dev_kfree_skb_any(skb);
535		return 0;
536	}
537
538	/* Get a packet buffer */
539	packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
540	if (unlikely(!packet_buffer)) {
541		printk_ratelimited("%s: Failed to allocate a packet buffer\n",
542				   dev->name);
543		cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
544		dev->stats.tx_dropped++;
545		dev_kfree_skb_any(skb);
546		return 0;
547	}
548
549	/*
550	 * Calculate where we need to copy the data to. We need to
551	 * leave 8 bytes for a next pointer (unused). We also need to
552	 * include any configure skip. Then we need to align the IP
553	 * packet src and dest into the same 64bit word. The below
554	 * calculation may add a little extra, but that doesn't
555	 * hurt.
556	 */
557	copy_location = packet_buffer + sizeof(u64);
558	copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
559
560	/*
561	 * We have to copy the packet since whoever processes this
562	 * packet will free it to a hardware pool. We can't use the
563	 * trick of counting outstanding packets like in
564	 * cvm_oct_xmit.
565	 */
566	memcpy(copy_location, skb->data, skb->len);
567
568	/*
569	 * Fill in some of the work queue fields. We may need to add
570	 * more if the software at the other end needs them.
571	 */
572	if (!OCTEON_IS_MODEL(OCTEON_CN68XX))
573		work->word0.pip.cn38xx.hw_chksum = skb->csum;
574	work->word1.len = skb->len;
575	cvmx_wqe_set_port(work, priv->port);
576	cvmx_wqe_set_qos(work, priv->port & 0x7);
577	cvmx_wqe_set_grp(work, pow_send_group);
578	work->word1.tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
579	work->word1.tag = pow_send_group;	/* FIXME */
580	/* Default to zero. Sets of zero later are commented out */
581	work->word2.u64 = 0;
582	work->word2.s.bufs = 1;
583	work->packet_ptr.u64 = 0;
584	work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
585	work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
586	work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
587	work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
588
589	if (skb->protocol == htons(ETH_P_IP)) {
590		work->word2.s.ip_offset = 14;
591#if 0
592		work->word2.s.vlan_valid = 0;	/* FIXME */
593		work->word2.s.vlan_cfi = 0;	/* FIXME */
594		work->word2.s.vlan_id = 0;	/* FIXME */
595		work->word2.s.dec_ipcomp = 0;	/* FIXME */
596#endif
597		work->word2.s.tcp_or_udp =
598		    (ip_hdr(skb)->protocol == IPPROTO_TCP) ||
599		    (ip_hdr(skb)->protocol == IPPROTO_UDP);
600#if 0
601		/* FIXME */
602		work->word2.s.dec_ipsec = 0;
603		/* We only support IPv4 right now */
604		work->word2.s.is_v6 = 0;
605		/* Hardware would set to zero */
606		work->word2.s.software = 0;
607		/* No error, packet is internal */
608		work->word2.s.L4_error = 0;
609#endif
610		work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0) ||
611					  (ip_hdr(skb)->frag_off ==
612					      1 << 14));
613#if 0
614		/* Assume Linux is sending a good packet */
615		work->word2.s.IP_exc = 0;
616#endif
617		work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
618		work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
619#if 0
620		/* This is an IP packet */
621		work->word2.s.not_IP = 0;
622		/* No error, packet is internal */
623		work->word2.s.rcv_error = 0;
624		/* No error, packet is internal */
625		work->word2.s.err_code = 0;
626#endif
627
628		/*
629		 * When copying the data, include 4 bytes of the
630		 * ethernet header to align the same way hardware
631		 * does.
632		 */
633		memcpy(work->packet_data, skb->data + 10,
634		       sizeof(work->packet_data));
635	} else {
636#if 0
637		work->word2.snoip.vlan_valid = 0;	/* FIXME */
638		work->word2.snoip.vlan_cfi = 0;	/* FIXME */
639		work->word2.snoip.vlan_id = 0;	/* FIXME */
640		work->word2.snoip.software = 0;	/* Hardware would set to zero */
641#endif
642		work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
643		work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
644		work->word2.snoip.is_bcast =
645		    (skb->pkt_type == PACKET_BROADCAST);
646		work->word2.snoip.is_mcast =
647		    (skb->pkt_type == PACKET_MULTICAST);
648		work->word2.snoip.not_IP = 1;	/* IP was done up above */
649#if 0
650		/* No error, packet is internal */
651		work->word2.snoip.rcv_error = 0;
652		/* No error, packet is internal */
653		work->word2.snoip.err_code = 0;
654#endif
655		memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
656	}
657
658	/* Submit the packet to the POW */
659	cvmx_pow_work_submit(work, work->word1.tag, work->word1.tag_type,
660			     cvmx_wqe_get_qos(work), cvmx_wqe_get_grp(work));
661	dev->stats.tx_packets++;
662	dev->stats.tx_bytes += skb->len;
663	dev_consume_skb_any(skb);
664	return 0;
665}
666
667/**
668 * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
669 * @dev:    Device being shutdown
670 *
671 */
672void cvm_oct_tx_shutdown_dev(struct net_device *dev)
673{
674	struct octeon_ethernet *priv = netdev_priv(dev);
675	unsigned long flags;
676	int qos;
677
678	for (qos = 0; qos < 16; qos++) {
679		spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
680		while (skb_queue_len(&priv->tx_free_list[qos]))
681			dev_kfree_skb_any(__skb_dequeue
682					  (&priv->tx_free_list[qos]));
683		spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
684	}
685}
686
687static void cvm_oct_tx_do_cleanup(unsigned long arg)
688{
689	int port;
690
691	for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
692		if (cvm_oct_device[port]) {
693			struct net_device *dev = cvm_oct_device[port];
694
695			cvm_oct_free_tx_skbs(dev);
696		}
697	}
698}
699
700static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
701{
702	/* Disable the interrupt.  */
703	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
704	/* Do the work in the tasklet.  */
705	tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
706	return IRQ_HANDLED;
707}
708
709void cvm_oct_tx_initialize(void)
710{
711	int i;
712
713	/* Disable the interrupt.  */
714	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
715	/* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
716	i = request_irq(OCTEON_IRQ_TIMER1,
717			cvm_oct_tx_cleanup_watchdog, 0,
718			"Ethernet", cvm_oct_device);
719
720	if (i)
721		panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
722}
723
724void cvm_oct_tx_shutdown(void)
725{
726	/* Free the interrupt handler */
727	free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
728}