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 
71 static void cvm_oct_tx_do_cleanup(unsigned long arg);
72 static 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 
cvm_oct_adjust_skb_to_free(int32_t skb_to_free,int fau)77 static inline int32_t cvm_oct_adjust_skb_to_free(int32_t skb_to_free, int fau)
78 {
79 	int32_t undo;
80 
81 	undo = skb_to_free > 0 ? MAX_SKB_TO_FREE : skb_to_free +
82 						   MAX_SKB_TO_FREE;
83 	if (undo > 0)
84 		cvmx_fau_atomic_add32(fau, -undo);
85 	skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE :
86 						       -skb_to_free;
87 	return skb_to_free;
88 }
89 
cvm_oct_kick_tx_poll_watchdog(void)90 static void cvm_oct_kick_tx_poll_watchdog(void)
91 {
92 	union cvmx_ciu_timx ciu_timx;
93 
94 	ciu_timx.u64 = 0;
95 	ciu_timx.s.one_shot = 1;
96 	ciu_timx.s.len = cvm_oct_tx_poll_interval;
97 	cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64);
98 }
99 
cvm_oct_free_tx_skbs(struct net_device * dev)100 static void cvm_oct_free_tx_skbs(struct net_device *dev)
101 {
102 	int32_t skb_to_free;
103 	int qos, queues_per_port;
104 	int total_freed = 0;
105 	int total_remaining = 0;
106 	unsigned long flags;
107 	struct octeon_ethernet *priv = netdev_priv(dev);
108 
109 	queues_per_port = cvmx_pko_get_num_queues(priv->port);
110 	/* Drain any pending packets in the free list */
111 	for (qos = 0; qos < queues_per_port; qos++) {
112 		if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
113 			continue;
114 		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau+qos*4,
115 						       MAX_SKB_TO_FREE);
116 		skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
117 							 priv->fau+qos*4);
118 
119 
120 		total_freed += skb_to_free;
121 		if (skb_to_free > 0) {
122 			struct sk_buff *to_free_list = NULL;
123 
124 			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
125 			while (skb_to_free > 0) {
126 				struct sk_buff *t;
127 
128 				t = __skb_dequeue(&priv->tx_free_list[qos]);
129 				t->next = to_free_list;
130 				to_free_list = t;
131 				skb_to_free--;
132 			}
133 			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
134 					       flags);
135 			/* Do the actual freeing outside of the lock. */
136 			while (to_free_list) {
137 				struct sk_buff *t = to_free_list;
138 
139 				to_free_list = to_free_list->next;
140 				dev_kfree_skb_any(t);
141 			}
142 		}
143 		total_remaining += skb_queue_len(&priv->tx_free_list[qos]);
144 	}
145 	if (total_freed >= 0 && netif_queue_stopped(dev))
146 		netif_wake_queue(dev);
147 	if (total_remaining)
148 		cvm_oct_kick_tx_poll_watchdog();
149 }
150 
151 /**
152  * cvm_oct_xmit - transmit a packet
153  * @skb:    Packet to send
154  * @dev:    Device info structure
155  *
156  * Returns Always returns NETDEV_TX_OK
157  */
cvm_oct_xmit(struct sk_buff * skb,struct net_device * dev)158 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
159 {
160 	cvmx_pko_command_word0_t pko_command;
161 	union cvmx_buf_ptr hw_buffer;
162 	uint64_t old_scratch;
163 	uint64_t old_scratch2;
164 	int qos;
165 	int i;
166 	enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
167 	struct octeon_ethernet *priv = netdev_priv(dev);
168 	struct sk_buff *to_free_list;
169 	int32_t skb_to_free;
170 	int32_t buffers_to_free;
171 	u32 total_to_clean;
172 	unsigned long flags;
173 #if REUSE_SKBUFFS_WITHOUT_FREE
174 	unsigned char *fpa_head;
175 #endif
176 
177 	/*
178 	 * Prefetch the private data structure.  It is larger than the
179 	 * one cache line.
180 	 */
181 	prefetch(priv);
182 
183 	/*
184 	 * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
185 	 * completely remove "qos" in the event neither interface
186 	 * supports multiple queues per port.
187 	 */
188 	if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
189 	    (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
190 		qos = GET_SKBUFF_QOS(skb);
191 		if (qos <= 0)
192 			qos = 0;
193 		else if (qos >= cvmx_pko_get_num_queues(priv->port))
194 			qos = 0;
195 	} else
196 		qos = 0;
197 
198 	if (USE_ASYNC_IOBDMA) {
199 		/* Save scratch in case userspace is using it */
200 		CVMX_SYNCIOBDMA;
201 		old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
202 		old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
203 
204 		/*
205 		 * Fetch and increment the number of packets to be
206 		 * freed.
207 		 */
208 		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
209 					       FAU_NUM_PACKET_BUFFERS_TO_FREE,
210 					       0);
211 		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
212 					       priv->fau + qos * 4,
213 					       MAX_SKB_TO_FREE);
214 	}
215 
216 	/*
217 	 * We have space for 6 segment pointers, If there will be more
218 	 * than that, we must linearize.
219 	 */
220 	if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
221 		if (unlikely(__skb_linearize(skb))) {
222 			queue_type = QUEUE_DROP;
223 			if (USE_ASYNC_IOBDMA) {
224 				/*
225 				 * Get the number of skbuffs in use
226 				 * by the hardware
227 				 */
228 				CVMX_SYNCIOBDMA;
229 				skb_to_free =
230 					cvmx_scratch_read64(CVMX_SCR_SCRATCH);
231 			} else {
232 				/*
233 				 * Get the number of skbuffs in use
234 				 * by the hardware
235 				 */
236 				skb_to_free = cvmx_fau_fetch_and_add32(
237 					priv->fau + qos * 4, MAX_SKB_TO_FREE);
238 			}
239 			skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free,
240 							priv->fau + qos * 4);
241 			spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
242 			goto skip_xmit;
243 		}
244 	}
245 
246 	/*
247 	 * The CN3XXX series of parts has an errata (GMX-401) which
248 	 * causes the GMX block to hang if a collision occurs towards
249 	 * the end of a <68 byte packet. As a workaround for this, we
250 	 * pad packets to be 68 bytes whenever we are in half duplex
251 	 * mode. We don't handle the case of having a small packet but
252 	 * no room to add the padding.  The kernel should always give
253 	 * us at least a cache line
254 	 */
255 	if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
256 		union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
257 		int interface = INTERFACE(priv->port);
258 		int index = INDEX(priv->port);
259 
260 		if (interface < 2) {
261 			/* We only need to pad packet in half duplex mode */
262 			gmx_prt_cfg.u64 =
263 			    cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
264 			if (gmx_prt_cfg.s.duplex == 0) {
265 				int add_bytes = 64 - skb->len;
266 
267 				if ((skb_tail_pointer(skb) + add_bytes) <=
268 				    skb_end_pointer(skb))
269 					memset(__skb_put(skb, add_bytes), 0,
270 					       add_bytes);
271 			}
272 		}
273 	}
274 
275 	/* Build the PKO command */
276 	pko_command.u64 = 0;
277 #ifdef __LITTLE_ENDIAN
278 	pko_command.s.le = 1;
279 #endif
280 	pko_command.s.n2 = 1;	/* Don't pollute L2 with the outgoing packet */
281 	pko_command.s.segs = 1;
282 	pko_command.s.total_bytes = skb->len;
283 	pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
284 	pko_command.s.subone0 = 1;
285 
286 	pko_command.s.dontfree = 1;
287 
288 	/* Build the PKO buffer pointer */
289 	hw_buffer.u64 = 0;
290 	if (skb_shinfo(skb)->nr_frags == 0) {
291 		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
292 		hw_buffer.s.pool = 0;
293 		hw_buffer.s.size = skb->len;
294 	} else {
295 		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
296 		hw_buffer.s.pool = 0;
297 		hw_buffer.s.size = skb_headlen(skb);
298 		CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
299 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
300 			struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i;
301 
302 			hw_buffer.s.addr = XKPHYS_TO_PHYS(
303 				(u64)(page_address(fs->page.p) +
304 				fs->page_offset));
305 			hw_buffer.s.size = fs->size;
306 			CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
307 		}
308 		hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb));
309 		hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1;
310 		pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1;
311 		pko_command.s.gather = 1;
312 		goto dont_put_skbuff_in_hw;
313 	}
314 
315 	/*
316 	 * See if we can put this skb in the FPA pool. Any strange
317 	 * behavior from the Linux networking stack will most likely
318 	 * be caused by a bug in the following code. If some field is
319 	 * in use by the network stack and gets carried over when a
320 	 * buffer is reused, bad things may happen.  If in doubt and
321 	 * you dont need the absolute best performance, disable the
322 	 * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
323 	 * shown a 25% increase in performance under some loads.
324 	 */
325 #if REUSE_SKBUFFS_WITHOUT_FREE
326 	fpa_head = skb->head + 256 - ((unsigned long)skb->head & 0x7f);
327 	if (unlikely(skb->data < fpa_head)) {
328 		/*
329 		 * printk("TX buffer beginning can't meet FPA
330 		 * alignment constraints\n");
331 		 */
332 		goto dont_put_skbuff_in_hw;
333 	}
334 	if (unlikely
335 	    ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
336 		/*
337 		   printk("TX buffer isn't large enough for the FPA\n");
338 		 */
339 		goto dont_put_skbuff_in_hw;
340 	}
341 	if (unlikely(skb_shared(skb))) {
342 		/*
343 		   printk("TX buffer sharing data with someone else\n");
344 		 */
345 		goto dont_put_skbuff_in_hw;
346 	}
347 	if (unlikely(skb_cloned(skb))) {
348 		/*
349 		   printk("TX buffer has been cloned\n");
350 		 */
351 		goto dont_put_skbuff_in_hw;
352 	}
353 	if (unlikely(skb_header_cloned(skb))) {
354 		/*
355 		   printk("TX buffer header has been cloned\n");
356 		 */
357 		goto dont_put_skbuff_in_hw;
358 	}
359 	if (unlikely(skb->destructor)) {
360 		/*
361 		   printk("TX buffer has a destructor\n");
362 		 */
363 		goto dont_put_skbuff_in_hw;
364 	}
365 	if (unlikely(skb_shinfo(skb)->nr_frags)) {
366 		/*
367 		   printk("TX buffer has fragments\n");
368 		 */
369 		goto dont_put_skbuff_in_hw;
370 	}
371 	if (unlikely
372 	    (skb->truesize !=
373 	     sizeof(*skb) + skb_end_offset(skb))) {
374 		/*
375 		   printk("TX buffer truesize has been changed\n");
376 		 */
377 		goto dont_put_skbuff_in_hw;
378 	}
379 
380 	/*
381 	 * We can use this buffer in the FPA.  We don't need the FAU
382 	 * update anymore
383 	 */
384 	pko_command.s.dontfree = 0;
385 
386 	hw_buffer.s.back = ((unsigned long)skb->data >> 7) -
387 			   ((unsigned long)fpa_head >> 7);
388 
389 	*(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
390 
391 	/*
392 	 * The skbuff will be reused without ever being freed. We must
393 	 * cleanup a bunch of core things.
394 	 */
395 	dst_release(skb_dst(skb));
396 	skb_dst_set(skb, NULL);
397 #ifdef CONFIG_XFRM
398 	secpath_put(skb->sp);
399 	skb->sp = NULL;
400 #endif
401 	nf_reset(skb);
402 
403 #ifdef CONFIG_NET_SCHED
404 	skb->tc_index = 0;
405 #ifdef CONFIG_NET_CLS_ACT
406 	skb->tc_verd = 0;
407 #endif /* CONFIG_NET_CLS_ACT */
408 #endif /* CONFIG_NET_SCHED */
409 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
410 
411 dont_put_skbuff_in_hw:
412 
413 	/* Check if we can use the hardware checksumming */
414 	if (USE_HW_TCPUDP_CHECKSUM && (skb->protocol == htons(ETH_P_IP)) &&
415 	    (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) &&
416 	    ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == htons(1 << 14)))
417 	    && ((ip_hdr(skb)->protocol == IPPROTO_TCP)
418 		|| (ip_hdr(skb)->protocol == IPPROTO_UDP))) {
419 		/* Use hardware checksum calc */
420 		pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
421 	}
422 
423 	if (USE_ASYNC_IOBDMA) {
424 		/* Get the number of skbuffs in use by the hardware */
425 		CVMX_SYNCIOBDMA;
426 		skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
427 		buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
428 	} else {
429 		/* Get the number of skbuffs in use by the hardware */
430 		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
431 						       MAX_SKB_TO_FREE);
432 		buffers_to_free =
433 		    cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
434 	}
435 
436 	skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4);
437 
438 	/*
439 	 * If we're sending faster than the receive can free them then
440 	 * don't do the HW free.
441 	 */
442 	if ((buffers_to_free < -100) && !pko_command.s.dontfree)
443 		pko_command.s.dontfree = 1;
444 
445 	if (pko_command.s.dontfree) {
446 		queue_type = QUEUE_CORE;
447 		pko_command.s.reg0 = priv->fau+qos*4;
448 	} else {
449 		queue_type = QUEUE_HW;
450 	}
451 	if (USE_ASYNC_IOBDMA)
452 		cvmx_fau_async_fetch_and_add32(
453 				CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1);
454 
455 	spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
456 
457 	/* Drop this packet if we have too many already queued to the HW */
458 	if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >=
459 		     MAX_OUT_QUEUE_DEPTH)) {
460 
461 		if (dev->tx_queue_len != 0) {
462 			/* Drop the lock when notifying the core.  */
463 			spin_unlock_irqrestore(&priv->tx_free_list[qos].lock,
464 					       flags);
465 			netif_stop_queue(dev);
466 			spin_lock_irqsave(&priv->tx_free_list[qos].lock,
467 					  flags);
468 		} else {
469 			/* If not using normal queueing.  */
470 			queue_type = QUEUE_DROP;
471 			goto skip_xmit;
472 		}
473 	}
474 
475 	cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
476 				     CVMX_PKO_LOCK_NONE);
477 
478 	/* Send the packet to the output queue */
479 	if (unlikely(cvmx_pko_send_packet_finish(priv->port,
480 						 priv->queue + qos,
481 						 pko_command, hw_buffer,
482 						 CVMX_PKO_LOCK_NONE))) {
483 		printk_ratelimited("%s: Failed to send the packet\n",
484 				   dev->name);
485 		queue_type = QUEUE_DROP;
486 	}
487 skip_xmit:
488 	to_free_list = NULL;
489 
490 	switch (queue_type) {
491 	case QUEUE_DROP:
492 		skb->next = to_free_list;
493 		to_free_list = skb;
494 		priv->stats.tx_dropped++;
495 		break;
496 	case QUEUE_HW:
497 		cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
498 		break;
499 	case QUEUE_CORE:
500 		__skb_queue_tail(&priv->tx_free_list[qos], skb);
501 		break;
502 	default:
503 		BUG();
504 	}
505 
506 	while (skb_to_free > 0) {
507 		struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]);
508 
509 		t->next = to_free_list;
510 		to_free_list = t;
511 		skb_to_free--;
512 	}
513 
514 	spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
515 
516 	/* Do the actual freeing outside of the lock. */
517 	while (to_free_list) {
518 		struct sk_buff *t = to_free_list;
519 
520 		to_free_list = to_free_list->next;
521 		dev_kfree_skb_any(t);
522 	}
523 
524 	if (USE_ASYNC_IOBDMA) {
525 		CVMX_SYNCIOBDMA;
526 		total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
527 		/* Restore the scratch area */
528 		cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
529 		cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
530 	} else {
531 		total_to_clean = cvmx_fau_fetch_and_add32(
532 						FAU_TOTAL_TX_TO_CLEAN, 1);
533 	}
534 
535 	if (total_to_clean & 0x3ff) {
536 		/*
537 		 * Schedule the cleanup tasklet every 1024 packets for
538 		 * the pathological case of high traffic on one port
539 		 * delaying clean up of packets on a different port
540 		 * that is blocked waiting for the cleanup.
541 		 */
542 		tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
543 	}
544 
545 	cvm_oct_kick_tx_poll_watchdog();
546 
547 	return NETDEV_TX_OK;
548 }
549 
550 /**
551  * cvm_oct_xmit_pow - transmit a packet to the POW
552  * @skb:    Packet to send
553  * @dev:    Device info structure
554 
555  * Returns Always returns zero
556  */
cvm_oct_xmit_pow(struct sk_buff * skb,struct net_device * dev)557 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
558 {
559 	struct octeon_ethernet *priv = netdev_priv(dev);
560 	void *packet_buffer;
561 	void *copy_location;
562 
563 	/* Get a work queue entry */
564 	cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
565 
566 	if (unlikely(work == NULL)) {
567 		printk_ratelimited("%s: Failed to allocate a work queue entry\n",
568 				   dev->name);
569 		priv->stats.tx_dropped++;
570 		dev_kfree_skb_any(skb);
571 		return 0;
572 	}
573 
574 	/* Get a packet buffer */
575 	packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
576 	if (unlikely(packet_buffer == NULL)) {
577 		printk_ratelimited("%s: Failed to allocate a packet buffer\n",
578 				   dev->name);
579 		cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
580 		priv->stats.tx_dropped++;
581 		dev_kfree_skb_any(skb);
582 		return 0;
583 	}
584 
585 	/*
586 	 * Calculate where we need to copy the data to. We need to
587 	 * leave 8 bytes for a next pointer (unused). We also need to
588 	 * include any configure skip. Then we need to align the IP
589 	 * packet src and dest into the same 64bit word. The below
590 	 * calculation may add a little extra, but that doesn't
591 	 * hurt.
592 	 */
593 	copy_location = packet_buffer + sizeof(uint64_t);
594 	copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
595 
596 	/*
597 	 * We have to copy the packet since whoever processes this
598 	 * packet will free it to a hardware pool. We can't use the
599 	 * trick of counting outstanding packets like in
600 	 * cvm_oct_xmit.
601 	 */
602 	memcpy(copy_location, skb->data, skb->len);
603 
604 	/*
605 	 * Fill in some of the work queue fields. We may need to add
606 	 * more if the software at the other end needs them.
607 	 */
608 	work->hw_chksum = skb->csum;
609 	work->len = skb->len;
610 	work->ipprt = priv->port;
611 	work->qos = priv->port & 0x7;
612 	work->grp = pow_send_group;
613 	work->tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
614 	work->tag = pow_send_group;	/* FIXME */
615 	/* Default to zero. Sets of zero later are commented out */
616 	work->word2.u64 = 0;
617 	work->word2.s.bufs = 1;
618 	work->packet_ptr.u64 = 0;
619 	work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
620 	work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
621 	work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
622 	work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
623 
624 	if (skb->protocol == htons(ETH_P_IP)) {
625 		work->word2.s.ip_offset = 14;
626 #if 0
627 		work->word2.s.vlan_valid = 0;	/* FIXME */
628 		work->word2.s.vlan_cfi = 0;	/* FIXME */
629 		work->word2.s.vlan_id = 0;	/* FIXME */
630 		work->word2.s.dec_ipcomp = 0;	/* FIXME */
631 #endif
632 		work->word2.s.tcp_or_udp =
633 		    (ip_hdr(skb)->protocol == IPPROTO_TCP)
634 		    || (ip_hdr(skb)->protocol == IPPROTO_UDP);
635 #if 0
636 		/* FIXME */
637 		work->word2.s.dec_ipsec = 0;
638 		/* We only support IPv4 right now */
639 		work->word2.s.is_v6 = 0;
640 		/* Hardware would set to zero */
641 		work->word2.s.software = 0;
642 		/* No error, packet is internal */
643 		work->word2.s.L4_error = 0;
644 #endif
645 		work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0)
646 					  || (ip_hdr(skb)->frag_off ==
647 					      1 << 14));
648 #if 0
649 		/* Assume Linux is sending a good packet */
650 		work->word2.s.IP_exc = 0;
651 #endif
652 		work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
653 		work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
654 #if 0
655 		/* This is an IP packet */
656 		work->word2.s.not_IP = 0;
657 		/* No error, packet is internal */
658 		work->word2.s.rcv_error = 0;
659 		/* No error, packet is internal */
660 		work->word2.s.err_code = 0;
661 #endif
662 
663 		/*
664 		 * When copying the data, include 4 bytes of the
665 		 * ethernet header to align the same way hardware
666 		 * does.
667 		 */
668 		memcpy(work->packet_data, skb->data + 10,
669 		       sizeof(work->packet_data));
670 	} else {
671 #if 0
672 		work->word2.snoip.vlan_valid = 0;	/* FIXME */
673 		work->word2.snoip.vlan_cfi = 0;	/* FIXME */
674 		work->word2.snoip.vlan_id = 0;	/* FIXME */
675 		work->word2.snoip.software = 0;	/* Hardware would set to zero */
676 #endif
677 		work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
678 		work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
679 		work->word2.snoip.is_bcast =
680 		    (skb->pkt_type == PACKET_BROADCAST);
681 		work->word2.snoip.is_mcast =
682 		    (skb->pkt_type == PACKET_MULTICAST);
683 		work->word2.snoip.not_IP = 1;	/* IP was done up above */
684 #if 0
685 		/* No error, packet is internal */
686 		work->word2.snoip.rcv_error = 0;
687 		/* No error, packet is internal */
688 		work->word2.snoip.err_code = 0;
689 #endif
690 		memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
691 	}
692 
693 	/* Submit the packet to the POW */
694 	cvmx_pow_work_submit(work, work->tag, work->tag_type, work->qos,
695 			     work->grp);
696 	priv->stats.tx_packets++;
697 	priv->stats.tx_bytes += skb->len;
698 	dev_consume_skb_any(skb);
699 	return 0;
700 }
701 
702 /**
703  * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX.
704  * @dev:    Device being shutdown
705  *
706  */
cvm_oct_tx_shutdown_dev(struct net_device * dev)707 void cvm_oct_tx_shutdown_dev(struct net_device *dev)
708 {
709 	struct octeon_ethernet *priv = netdev_priv(dev);
710 	unsigned long flags;
711 	int qos;
712 
713 	for (qos = 0; qos < 16; qos++) {
714 		spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
715 		while (skb_queue_len(&priv->tx_free_list[qos]))
716 			dev_kfree_skb_any(__skb_dequeue
717 					  (&priv->tx_free_list[qos]));
718 		spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);
719 	}
720 }
721 
cvm_oct_tx_do_cleanup(unsigned long arg)722 static void cvm_oct_tx_do_cleanup(unsigned long arg)
723 {
724 	int port;
725 
726 	for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) {
727 		if (cvm_oct_device[port]) {
728 			struct net_device *dev = cvm_oct_device[port];
729 
730 			cvm_oct_free_tx_skbs(dev);
731 		}
732 	}
733 }
734 
cvm_oct_tx_cleanup_watchdog(int cpl,void * dev_id)735 static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id)
736 {
737 	/* Disable the interrupt.  */
738 	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
739 	/* Do the work in the tasklet.  */
740 	tasklet_schedule(&cvm_oct_tx_cleanup_tasklet);
741 	return IRQ_HANDLED;
742 }
743 
cvm_oct_tx_initialize(void)744 void cvm_oct_tx_initialize(void)
745 {
746 	int i;
747 
748 	/* Disable the interrupt.  */
749 	cvmx_write_csr(CVMX_CIU_TIMX(1), 0);
750 	/* Register an IRQ handler to receive CIU_TIMX(1) interrupts */
751 	i = request_irq(OCTEON_IRQ_TIMER1,
752 			cvm_oct_tx_cleanup_watchdog, 0,
753 			"Ethernet", cvm_oct_device);
754 
755 	if (i)
756 		panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1);
757 }
758 
cvm_oct_tx_shutdown(void)759 void cvm_oct_tx_shutdown(void)
760 {
761 	/* Free the interrupt handler */
762 	free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device);
763 }
764