]>
Commit | Line | Data |
---|---|---|
80ff0fd3 DD |
1 | /********************************************************************* |
2 | * Author: Cavium Networks | |
3 | * | |
4 | * Contact: support@caviumnetworks.com | |
5 | * This file is part of the OCTEON SDK | |
6 | * | |
166bdaa9 | 7 | * Copyright (c) 2003-2010 Cavium Networks |
80ff0fd3 DD |
8 | * |
9 | * This file is free software; you can redistribute it and/or modify | |
10 | * it under the terms of the GNU General Public License, Version 2, as | |
11 | * published by the Free Software Foundation. | |
12 | * | |
13 | * This file is distributed in the hope that it will be useful, but | |
14 | * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty | |
15 | * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or | |
16 | * NONINFRINGEMENT. See the GNU General Public License for more | |
17 | * details. | |
18 | * | |
19 | * You should have received a copy of the GNU General Public License | |
20 | * along with this file; if not, write to the Free Software | |
21 | * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA | |
22 | * or visit http://www.gnu.org/licenses/. | |
23 | * | |
24 | * This file may also be available under a different license from Cavium. | |
25 | * Contact Cavium Networks for more information | |
26 | *********************************************************************/ | |
27 | #include <linux/module.h> | |
28 | #include <linux/kernel.h> | |
29 | #include <linux/netdevice.h> | |
30 | #include <linux/init.h> | |
31 | #include <linux/etherdevice.h> | |
32 | #include <linux/ip.h> | |
7a2eaf93 | 33 | #include <linux/ratelimit.h> |
80ff0fd3 | 34 | #include <linux/string.h> |
80ff0fd3 DD |
35 | #include <net/dst.h> |
36 | #ifdef CONFIG_XFRM | |
37 | #include <linux/xfrm.h> | |
38 | #include <net/xfrm.h> | |
39 | #endif /* CONFIG_XFRM */ | |
40 | ||
60063497 | 41 | #include <linux/atomic.h> |
80ff0fd3 DD |
42 | |
43 | #include <asm/octeon/octeon.h> | |
44 | ||
45 | #include "ethernet-defines.h" | |
46 | #include "octeon-ethernet.h" | |
a620c163 | 47 | #include "ethernet-tx.h" |
80ff0fd3 DD |
48 | #include "ethernet-util.h" |
49 | ||
50 | #include "cvmx-wqe.h" | |
51 | #include "cvmx-fau.h" | |
4898c560 | 52 | #include "cvmx-pip.h" |
80ff0fd3 DD |
53 | #include "cvmx-pko.h" |
54 | #include "cvmx-helper.h" | |
55 | ||
56 | #include "cvmx-gmxx-defs.h" | |
57 | ||
924cc268 DD |
58 | #define CVM_OCT_SKB_CB(skb) ((u64 *)((skb)->cb)) |
59 | ||
80ff0fd3 DD |
60 | /* |
61 | * You can define GET_SKBUFF_QOS() to override how the skbuff output | |
62 | * function determines which output queue is used. The default | |
63 | * implementation always uses the base queue for the port. If, for | |
64 | * example, you wanted to use the skb->priority fieid, define | |
65 | * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority) | |
66 | */ | |
67 | #ifndef GET_SKBUFF_QOS | |
68 | #define GET_SKBUFF_QOS(skb) 0 | |
69 | #endif | |
70 | ||
4898c560 DD |
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) | |
6888fc87 DD |
76 | |
77 | static 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 + MAX_SKB_TO_FREE; | |
81 | if (undo > 0) | |
82 | cvmx_fau_atomic_add32(fau, -undo); | |
83 | skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ? MAX_SKB_TO_FREE : -skb_to_free; | |
84 | return skb_to_free; | |
85 | } | |
86 | ||
4898c560 DD |
87 | static void cvm_oct_kick_tx_poll_watchdog(void) |
88 | { | |
89 | union cvmx_ciu_timx ciu_timx; | |
90 | ciu_timx.u64 = 0; | |
91 | ciu_timx.s.one_shot = 1; | |
92 | ciu_timx.s.len = cvm_oct_tx_poll_interval; | |
93 | cvmx_write_csr(CVMX_CIU_TIMX(1), ciu_timx.u64); | |
94 | } | |
95 | ||
96 | void cvm_oct_free_tx_skbs(struct net_device *dev) | |
6888fc87 DD |
97 | { |
98 | int32_t skb_to_free; | |
99 | int qos, queues_per_port; | |
4898c560 DD |
100 | int total_freed = 0; |
101 | int total_remaining = 0; | |
102 | unsigned long flags; | |
103 | struct octeon_ethernet *priv = netdev_priv(dev); | |
104 | ||
6888fc87 DD |
105 | queues_per_port = cvmx_pko_get_num_queues(priv->port); |
106 | /* Drain any pending packets in the free list */ | |
107 | for (qos = 0; qos < queues_per_port; qos++) { | |
108 | if (skb_queue_len(&priv->tx_free_list[qos]) == 0) | |
109 | continue; | |
110 | skb_to_free = cvmx_fau_fetch_and_add32(priv->fau+qos*4, MAX_SKB_TO_FREE); | |
111 | skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4); | |
112 | ||
4898c560 DD |
113 | |
114 | total_freed += skb_to_free; | |
115 | if (skb_to_free > 0) { | |
116 | struct sk_buff *to_free_list = NULL; | |
117 | spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags); | |
118 | while (skb_to_free > 0) { | |
119 | struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]); | |
120 | t->next = to_free_list; | |
121 | to_free_list = t; | |
122 | skb_to_free--; | |
123 | } | |
124 | spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags); | |
125 | /* Do the actual freeing outside of the lock. */ | |
126 | while (to_free_list) { | |
127 | struct sk_buff *t = to_free_list; | |
128 | to_free_list = to_free_list->next; | |
129 | dev_kfree_skb_any(t); | |
130 | } | |
6888fc87 | 131 | } |
4898c560 | 132 | total_remaining += skb_queue_len(&priv->tx_free_list[qos]); |
6888fc87 | 133 | } |
4898c560 | 134 | if (total_freed >= 0 && netif_queue_stopped(dev)) |
6888fc87 | 135 | netif_wake_queue(dev); |
4898c560 DD |
136 | if (total_remaining) |
137 | cvm_oct_kick_tx_poll_watchdog(); | |
6888fc87 DD |
138 | } |
139 | ||
80ff0fd3 | 140 | /** |
ec977c5b | 141 | * cvm_oct_xmit - transmit a packet |
80ff0fd3 DD |
142 | * @skb: Packet to send |
143 | * @dev: Device info structure | |
ec977c5b DD |
144 | * |
145 | * Returns Always returns NETDEV_TX_OK | |
80ff0fd3 DD |
146 | */ |
147 | int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev) | |
148 | { | |
149 | cvmx_pko_command_word0_t pko_command; | |
150 | union cvmx_buf_ptr hw_buffer; | |
151 | uint64_t old_scratch; | |
152 | uint64_t old_scratch2; | |
80ff0fd3 | 153 | int qos; |
924cc268 | 154 | int i; |
6888fc87 | 155 | enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type; |
80ff0fd3 | 156 | struct octeon_ethernet *priv = netdev_priv(dev); |
6888fc87 | 157 | struct sk_buff *to_free_list; |
a620c163 | 158 | int32_t skb_to_free; |
80ff0fd3 | 159 | int32_t buffers_to_free; |
4898c560 | 160 | u32 total_to_clean; |
6888fc87 | 161 | unsigned long flags; |
80ff0fd3 DD |
162 | #if REUSE_SKBUFFS_WITHOUT_FREE |
163 | unsigned char *fpa_head; | |
164 | #endif | |
165 | ||
166 | /* | |
167 | * Prefetch the private data structure. It is larger that one | |
168 | * cache line. | |
169 | */ | |
170 | prefetch(priv); | |
171 | ||
80ff0fd3 DD |
172 | /* |
173 | * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to | |
174 | * completely remove "qos" in the event neither interface | |
175 | * supports multiple queues per port. | |
176 | */ | |
177 | if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) || | |
178 | (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) { | |
179 | qos = GET_SKBUFF_QOS(skb); | |
180 | if (qos <= 0) | |
181 | qos = 0; | |
182 | else if (qos >= cvmx_pko_get_num_queues(priv->port)) | |
183 | qos = 0; | |
184 | } else | |
185 | qos = 0; | |
186 | ||
187 | if (USE_ASYNC_IOBDMA) { | |
188 | /* Save scratch in case userspace is using it */ | |
189 | CVMX_SYNCIOBDMA; | |
190 | old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH); | |
191 | old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8); | |
192 | ||
193 | /* | |
a620c163 DD |
194 | * Fetch and increment the number of packets to be |
195 | * freed. | |
80ff0fd3 DD |
196 | */ |
197 | cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8, | |
198 | FAU_NUM_PACKET_BUFFERS_TO_FREE, | |
199 | 0); | |
200 | cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH, | |
a620c163 DD |
201 | priv->fau + qos * 4, |
202 | MAX_SKB_TO_FREE); | |
80ff0fd3 DD |
203 | } |
204 | ||
924cc268 DD |
205 | /* |
206 | * We have space for 6 segment pointers, If there will be more | |
207 | * than that, we must linearize. | |
208 | */ | |
209 | if (unlikely(skb_shinfo(skb)->nr_frags > 5)) { | |
210 | if (unlikely(__skb_linearize(skb))) { | |
211 | queue_type = QUEUE_DROP; | |
212 | if (USE_ASYNC_IOBDMA) { | |
213 | /* Get the number of skbuffs in use by the hardware */ | |
214 | CVMX_SYNCIOBDMA; | |
215 | skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH); | |
216 | } else { | |
217 | /* Get the number of skbuffs in use by the hardware */ | |
218 | skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4, | |
219 | MAX_SKB_TO_FREE); | |
220 | } | |
221 | skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau + qos * 4); | |
222 | spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags); | |
223 | goto skip_xmit; | |
224 | } | |
225 | } | |
226 | ||
80ff0fd3 DD |
227 | /* |
228 | * The CN3XXX series of parts has an errata (GMX-401) which | |
229 | * causes the GMX block to hang if a collision occurs towards | |
230 | * the end of a <68 byte packet. As a workaround for this, we | |
231 | * pad packets to be 68 bytes whenever we are in half duplex | |
232 | * mode. We don't handle the case of having a small packet but | |
233 | * no room to add the padding. The kernel should always give | |
234 | * us at least a cache line | |
235 | */ | |
236 | if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) { | |
237 | union cvmx_gmxx_prtx_cfg gmx_prt_cfg; | |
238 | int interface = INTERFACE(priv->port); | |
239 | int index = INDEX(priv->port); | |
240 | ||
241 | if (interface < 2) { | |
242 | /* We only need to pad packet in half duplex mode */ | |
243 | gmx_prt_cfg.u64 = | |
244 | cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface)); | |
245 | if (gmx_prt_cfg.s.duplex == 0) { | |
246 | int add_bytes = 64 - skb->len; | |
247 | if ((skb_tail_pointer(skb) + add_bytes) <= | |
248 | skb_end_pointer(skb)) | |
249 | memset(__skb_put(skb, add_bytes), 0, | |
250 | add_bytes); | |
251 | } | |
252 | } | |
253 | } | |
254 | ||
80ff0fd3 DD |
255 | /* Build the PKO command */ |
256 | pko_command.u64 = 0; | |
257 | pko_command.s.n2 = 1; /* Don't pollute L2 with the outgoing packet */ | |
258 | pko_command.s.segs = 1; | |
259 | pko_command.s.total_bytes = skb->len; | |
260 | pko_command.s.size0 = CVMX_FAU_OP_SIZE_32; | |
261 | pko_command.s.subone0 = 1; | |
262 | ||
263 | pko_command.s.dontfree = 1; | |
924cc268 DD |
264 | |
265 | /* Build the PKO buffer pointer */ | |
266 | hw_buffer.u64 = 0; | |
267 | if (skb_shinfo(skb)->nr_frags == 0) { | |
268 | hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data); | |
269 | hw_buffer.s.pool = 0; | |
270 | hw_buffer.s.size = skb->len; | |
271 | } else { | |
272 | hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data); | |
273 | hw_buffer.s.pool = 0; | |
274 | hw_buffer.s.size = skb_headlen(skb); | |
275 | CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64; | |
276 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { | |
277 | struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i; | |
278 | hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)(page_address(fs->page) + fs->page_offset)); | |
279 | hw_buffer.s.size = fs->size; | |
280 | CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64; | |
281 | } | |
282 | hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb)); | |
283 | hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1; | |
284 | pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1; | |
285 | pko_command.s.gather = 1; | |
286 | goto dont_put_skbuff_in_hw; | |
287 | } | |
288 | ||
80ff0fd3 DD |
289 | /* |
290 | * See if we can put this skb in the FPA pool. Any strange | |
291 | * behavior from the Linux networking stack will most likely | |
292 | * be caused by a bug in the following code. If some field is | |
293 | * in use by the network stack and get carried over when a | |
294 | * buffer is reused, bad thing may happen. If in doubt and | |
295 | * you dont need the absolute best performance, disable the | |
296 | * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has | |
297 | * shown a 25% increase in performance under some loads. | |
298 | */ | |
299 | #if REUSE_SKBUFFS_WITHOUT_FREE | |
166bdaa9 | 300 | fpa_head = skb->head + 256 - ((unsigned long)skb->head & 0x7f); |
80ff0fd3 DD |
301 | if (unlikely(skb->data < fpa_head)) { |
302 | /* | |
303 | * printk("TX buffer beginning can't meet FPA | |
304 | * alignment constraints\n"); | |
305 | */ | |
306 | goto dont_put_skbuff_in_hw; | |
307 | } | |
308 | if (unlikely | |
309 | ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) { | |
310 | /* | |
311 | printk("TX buffer isn't large enough for the FPA\n"); | |
312 | */ | |
313 | goto dont_put_skbuff_in_hw; | |
314 | } | |
315 | if (unlikely(skb_shared(skb))) { | |
316 | /* | |
317 | printk("TX buffer sharing data with someone else\n"); | |
318 | */ | |
319 | goto dont_put_skbuff_in_hw; | |
320 | } | |
321 | if (unlikely(skb_cloned(skb))) { | |
322 | /* | |
323 | printk("TX buffer has been cloned\n"); | |
324 | */ | |
325 | goto dont_put_skbuff_in_hw; | |
326 | } | |
327 | if (unlikely(skb_header_cloned(skb))) { | |
328 | /* | |
329 | printk("TX buffer header has been cloned\n"); | |
330 | */ | |
331 | goto dont_put_skbuff_in_hw; | |
332 | } | |
333 | if (unlikely(skb->destructor)) { | |
334 | /* | |
335 | printk("TX buffer has a destructor\n"); | |
336 | */ | |
337 | goto dont_put_skbuff_in_hw; | |
338 | } | |
339 | if (unlikely(skb_shinfo(skb)->nr_frags)) { | |
340 | /* | |
341 | printk("TX buffer has fragments\n"); | |
342 | */ | |
343 | goto dont_put_skbuff_in_hw; | |
344 | } | |
345 | if (unlikely | |
346 | (skb->truesize != | |
347 | sizeof(*skb) + skb_end_pointer(skb) - skb->head)) { | |
348 | /* | |
349 | printk("TX buffer truesize has been changed\n"); | |
350 | */ | |
351 | goto dont_put_skbuff_in_hw; | |
352 | } | |
353 | ||
354 | /* | |
355 | * We can use this buffer in the FPA. We don't need the FAU | |
356 | * update anymore | |
357 | */ | |
80ff0fd3 DD |
358 | pko_command.s.dontfree = 0; |
359 | ||
166bdaa9 | 360 | hw_buffer.s.back = ((unsigned long)skb->data >> 7) - ((unsigned long)fpa_head >> 7); |
80ff0fd3 DD |
361 | *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb; |
362 | ||
363 | /* | |
364 | * The skbuff will be reused without ever being freed. We must | |
f696a108 | 365 | * cleanup a bunch of core things. |
80ff0fd3 | 366 | */ |
f696a108 DD |
367 | dst_release(skb_dst(skb)); |
368 | skb_dst_set(skb, NULL); | |
80ff0fd3 DD |
369 | #ifdef CONFIG_XFRM |
370 | secpath_put(skb->sp); | |
371 | skb->sp = NULL; | |
372 | #endif | |
373 | nf_reset(skb); | |
374 | ||
375 | #ifdef CONFIG_NET_SCHED | |
376 | skb->tc_index = 0; | |
377 | #ifdef CONFIG_NET_CLS_ACT | |
378 | skb->tc_verd = 0; | |
379 | #endif /* CONFIG_NET_CLS_ACT */ | |
380 | #endif /* CONFIG_NET_SCHED */ | |
6888fc87 | 381 | #endif /* REUSE_SKBUFFS_WITHOUT_FREE */ |
80ff0fd3 DD |
382 | |
383 | dont_put_skbuff_in_hw: | |
80ff0fd3 DD |
384 | |
385 | /* Check if we can use the hardware checksumming */ | |
386 | if (USE_HW_TCPUDP_CHECKSUM && (skb->protocol == htons(ETH_P_IP)) && | |
387 | (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) && | |
388 | ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == 1 << 14)) | |
081f6749 DD |
389 | && ((ip_hdr(skb)->protocol == IPPROTO_TCP) |
390 | || (ip_hdr(skb)->protocol == IPPROTO_UDP))) { | |
80ff0fd3 DD |
391 | /* Use hardware checksum calc */ |
392 | pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1; | |
393 | } | |
394 | ||
395 | if (USE_ASYNC_IOBDMA) { | |
396 | /* Get the number of skbuffs in use by the hardware */ | |
397 | CVMX_SYNCIOBDMA; | |
a620c163 | 398 | skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH); |
80ff0fd3 DD |
399 | buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8); |
400 | } else { | |
401 | /* Get the number of skbuffs in use by the hardware */ | |
a620c163 DD |
402 | skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4, |
403 | MAX_SKB_TO_FREE); | |
80ff0fd3 DD |
404 | buffers_to_free = |
405 | cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0); | |
406 | } | |
407 | ||
6888fc87 | 408 | skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau+qos*4); |
a620c163 | 409 | |
80ff0fd3 DD |
410 | /* |
411 | * If we're sending faster than the receive can free them then | |
412 | * don't do the HW free. | |
413 | */ | |
4898c560 | 414 | if ((buffers_to_free < -100) && !pko_command.s.dontfree) |
80ff0fd3 | 415 | pko_command.s.dontfree = 1; |
80ff0fd3 | 416 | |
4898c560 | 417 | if (pko_command.s.dontfree) { |
6888fc87 | 418 | queue_type = QUEUE_CORE; |
4898c560 DD |
419 | pko_command.s.reg0 = priv->fau+qos*4; |
420 | } else { | |
6888fc87 | 421 | queue_type = QUEUE_HW; |
4898c560 DD |
422 | } |
423 | if (USE_ASYNC_IOBDMA) | |
424 | cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH, FAU_TOTAL_TX_TO_CLEAN, 1); | |
6888fc87 DD |
425 | |
426 | spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags); | |
80ff0fd3 DD |
427 | |
428 | /* Drop this packet if we have too many already queued to the HW */ | |
6888fc87 DD |
429 | if (unlikely(skb_queue_len(&priv->tx_free_list[qos]) >= MAX_OUT_QUEUE_DEPTH)) { |
430 | if (dev->tx_queue_len != 0) { | |
431 | /* Drop the lock when notifying the core. */ | |
432 | spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags); | |
433 | netif_stop_queue(dev); | |
6888fc87 | 434 | spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags); |
6888fc87 DD |
435 | } else { |
436 | /* If not using normal queueing. */ | |
437 | queue_type = QUEUE_DROP; | |
438 | goto skip_xmit; | |
439 | } | |
80ff0fd3 | 440 | } |
6888fc87 DD |
441 | |
442 | cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos, | |
443 | CVMX_PKO_LOCK_NONE); | |
444 | ||
80ff0fd3 | 445 | /* Send the packet to the output queue */ |
6888fc87 DD |
446 | if (unlikely(cvmx_pko_send_packet_finish(priv->port, |
447 | priv->queue + qos, | |
448 | pko_command, hw_buffer, | |
449 | CVMX_PKO_LOCK_NONE))) { | |
7a2eaf93 | 450 | printk_ratelimited("%s: Failed to send the packet\n", dev->name); |
6888fc87 | 451 | queue_type = QUEUE_DROP; |
80ff0fd3 | 452 | } |
6888fc87 DD |
453 | skip_xmit: |
454 | to_free_list = NULL; | |
80ff0fd3 | 455 | |
6888fc87 DD |
456 | switch (queue_type) { |
457 | case QUEUE_DROP: | |
458 | skb->next = to_free_list; | |
459 | to_free_list = skb; | |
460 | priv->stats.tx_dropped++; | |
461 | break; | |
462 | case QUEUE_HW: | |
463 | cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, -1); | |
464 | break; | |
465 | case QUEUE_CORE: | |
466 | __skb_queue_tail(&priv->tx_free_list[qos], skb); | |
467 | break; | |
468 | default: | |
469 | BUG(); | |
80ff0fd3 DD |
470 | } |
471 | ||
6888fc87 DD |
472 | while (skb_to_free > 0) { |
473 | struct sk_buff *t = __skb_dequeue(&priv->tx_free_list[qos]); | |
474 | t->next = to_free_list; | |
475 | to_free_list = t; | |
476 | skb_to_free--; | |
80ff0fd3 DD |
477 | } |
478 | ||
6888fc87 DD |
479 | spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags); |
480 | ||
481 | /* Do the actual freeing outside of the lock. */ | |
482 | while (to_free_list) { | |
483 | struct sk_buff *t = to_free_list; | |
484 | to_free_list = to_free_list->next; | |
485 | dev_kfree_skb_any(t); | |
486 | } | |
487 | ||
488 | if (USE_ASYNC_IOBDMA) { | |
4898c560 DD |
489 | CVMX_SYNCIOBDMA; |
490 | total_to_clean = cvmx_scratch_read64(CVMX_SCR_SCRATCH); | |
6888fc87 DD |
491 | /* Restore the scratch area */ |
492 | cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch); | |
493 | cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2); | |
4898c560 DD |
494 | } else { |
495 | total_to_clean = cvmx_fau_fetch_and_add32(FAU_TOTAL_TX_TO_CLEAN, 1); | |
80ff0fd3 DD |
496 | } |
497 | ||
4898c560 DD |
498 | if (total_to_clean & 0x3ff) { |
499 | /* | |
500 | * Schedule the cleanup tasklet every 1024 packets for | |
501 | * the pathological case of high traffic on one port | |
502 | * delaying clean up of packets on a different port | |
503 | * that is blocked waiting for the cleanup. | |
504 | */ | |
505 | tasklet_schedule(&cvm_oct_tx_cleanup_tasklet); | |
506 | } | |
507 | ||
508 | cvm_oct_kick_tx_poll_watchdog(); | |
509 | ||
6888fc87 | 510 | return NETDEV_TX_OK; |
80ff0fd3 DD |
511 | } |
512 | ||
513 | /** | |
ec977c5b | 514 | * cvm_oct_xmit_pow - transmit a packet to the POW |
80ff0fd3 DD |
515 | * @skb: Packet to send |
516 | * @dev: Device info structure | |
ec977c5b | 517 | |
80ff0fd3 DD |
518 | * Returns Always returns zero |
519 | */ | |
520 | int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev) | |
521 | { | |
522 | struct octeon_ethernet *priv = netdev_priv(dev); | |
523 | void *packet_buffer; | |
524 | void *copy_location; | |
525 | ||
526 | /* Get a work queue entry */ | |
527 | cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL); | |
528 | if (unlikely(work == NULL)) { | |
7a2eaf93 CD |
529 | printk_ratelimited("%s: Failed to allocate a work " |
530 | "queue entry\n", dev->name); | |
80ff0fd3 DD |
531 | priv->stats.tx_dropped++; |
532 | dev_kfree_skb(skb); | |
533 | return 0; | |
534 | } | |
535 | ||
536 | /* Get a packet buffer */ | |
537 | packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL); | |
538 | if (unlikely(packet_buffer == NULL)) { | |
7a2eaf93 CD |
539 | printk_ratelimited("%s: Failed to allocate a packet buffer\n", |
540 | dev->name); | |
80ff0fd3 DD |
541 | cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1)); |
542 | priv->stats.tx_dropped++; | |
543 | dev_kfree_skb(skb); | |
544 | return 0; | |
545 | } | |
546 | ||
547 | /* | |
548 | * Calculate where we need to copy the data to. We need to | |
549 | * leave 8 bytes for a next pointer (unused). We also need to | |
550 | * include any configure skip. Then we need to align the IP | |
551 | * packet src and dest into the same 64bit word. The below | |
552 | * calculation may add a little extra, but that doesn't | |
553 | * hurt. | |
554 | */ | |
555 | copy_location = packet_buffer + sizeof(uint64_t); | |
556 | copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6; | |
557 | ||
558 | /* | |
559 | * We have to copy the packet since whoever processes this | |
560 | * packet will free it to a hardware pool. We can't use the | |
561 | * trick of counting outstanding packets like in | |
562 | * cvm_oct_xmit. | |
563 | */ | |
564 | memcpy(copy_location, skb->data, skb->len); | |
565 | ||
566 | /* | |
567 | * Fill in some of the work queue fields. We may need to add | |
568 | * more if the software at the other end needs them. | |
569 | */ | |
570 | work->hw_chksum = skb->csum; | |
571 | work->len = skb->len; | |
572 | work->ipprt = priv->port; | |
573 | work->qos = priv->port & 0x7; | |
574 | work->grp = pow_send_group; | |
575 | work->tag_type = CVMX_HELPER_INPUT_TAG_TYPE; | |
576 | work->tag = pow_send_group; /* FIXME */ | |
577 | /* Default to zero. Sets of zero later are commented out */ | |
578 | work->word2.u64 = 0; | |
579 | work->word2.s.bufs = 1; | |
580 | work->packet_ptr.u64 = 0; | |
581 | work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location); | |
582 | work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL; | |
583 | work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE; | |
584 | work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7; | |
585 | ||
586 | if (skb->protocol == htons(ETH_P_IP)) { | |
587 | work->word2.s.ip_offset = 14; | |
588 | #if 0 | |
589 | work->word2.s.vlan_valid = 0; /* FIXME */ | |
590 | work->word2.s.vlan_cfi = 0; /* FIXME */ | |
591 | work->word2.s.vlan_id = 0; /* FIXME */ | |
592 | work->word2.s.dec_ipcomp = 0; /* FIXME */ | |
593 | #endif | |
594 | work->word2.s.tcp_or_udp = | |
081f6749 DD |
595 | (ip_hdr(skb)->protocol == IPPROTO_TCP) |
596 | || (ip_hdr(skb)->protocol == IPPROTO_UDP); | |
80ff0fd3 DD |
597 | #if 0 |
598 | /* FIXME */ | |
599 | work->word2.s.dec_ipsec = 0; | |
600 | /* We only support IPv4 right now */ | |
601 | work->word2.s.is_v6 = 0; | |
602 | /* Hardware would set to zero */ | |
603 | work->word2.s.software = 0; | |
604 | /* No error, packet is internal */ | |
605 | work->word2.s.L4_error = 0; | |
606 | #endif | |
607 | work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0) | |
608 | || (ip_hdr(skb)->frag_off == | |
609 | 1 << 14)); | |
610 | #if 0 | |
611 | /* Assume Linux is sending a good packet */ | |
612 | work->word2.s.IP_exc = 0; | |
613 | #endif | |
614 | work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST); | |
615 | work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST); | |
616 | #if 0 | |
617 | /* This is an IP packet */ | |
618 | work->word2.s.not_IP = 0; | |
619 | /* No error, packet is internal */ | |
620 | work->word2.s.rcv_error = 0; | |
621 | /* No error, packet is internal */ | |
622 | work->word2.s.err_code = 0; | |
623 | #endif | |
624 | ||
625 | /* | |
626 | * When copying the data, include 4 bytes of the | |
627 | * ethernet header to align the same way hardware | |
628 | * does. | |
629 | */ | |
630 | memcpy(work->packet_data, skb->data + 10, | |
631 | sizeof(work->packet_data)); | |
632 | } else { | |
633 | #if 0 | |
634 | work->word2.snoip.vlan_valid = 0; /* FIXME */ | |
635 | work->word2.snoip.vlan_cfi = 0; /* FIXME */ | |
636 | work->word2.snoip.vlan_id = 0; /* FIXME */ | |
637 | work->word2.snoip.software = 0; /* Hardware would set to zero */ | |
638 | #endif | |
639 | work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP); | |
640 | work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP); | |
641 | work->word2.snoip.is_bcast = | |
642 | (skb->pkt_type == PACKET_BROADCAST); | |
643 | work->word2.snoip.is_mcast = | |
644 | (skb->pkt_type == PACKET_MULTICAST); | |
645 | work->word2.snoip.not_IP = 1; /* IP was done up above */ | |
646 | #if 0 | |
647 | /* No error, packet is internal */ | |
648 | work->word2.snoip.rcv_error = 0; | |
649 | /* No error, packet is internal */ | |
650 | work->word2.snoip.err_code = 0; | |
651 | #endif | |
652 | memcpy(work->packet_data, skb->data, sizeof(work->packet_data)); | |
653 | } | |
654 | ||
655 | /* Submit the packet to the POW */ | |
656 | cvmx_pow_work_submit(work, work->tag, work->tag_type, work->qos, | |
657 | work->grp); | |
658 | priv->stats.tx_packets++; | |
659 | priv->stats.tx_bytes += skb->len; | |
660 | dev_kfree_skb(skb); | |
661 | return 0; | |
662 | } | |
663 | ||
80ff0fd3 | 664 | /** |
ec977c5b | 665 | * cvm_oct_tx_shutdown_dev - free all skb that are currently queued for TX. |
80ff0fd3 | 666 | * @dev: Device being shutdown |
ec977c5b | 667 | * |
80ff0fd3 | 668 | */ |
4898c560 | 669 | void cvm_oct_tx_shutdown_dev(struct net_device *dev) |
80ff0fd3 DD |
670 | { |
671 | struct octeon_ethernet *priv = netdev_priv(dev); | |
672 | unsigned long flags; | |
673 | int qos; | |
674 | ||
675 | for (qos = 0; qos < 16; qos++) { | |
676 | spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags); | |
677 | while (skb_queue_len(&priv->tx_free_list[qos])) | |
678 | dev_kfree_skb_any(__skb_dequeue | |
679 | (&priv->tx_free_list[qos])); | |
680 | spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags); | |
681 | } | |
682 | } | |
4898c560 DD |
683 | |
684 | static void cvm_oct_tx_do_cleanup(unsigned long arg) | |
685 | { | |
686 | int port; | |
687 | ||
688 | for (port = 0; port < TOTAL_NUMBER_OF_PORTS; port++) { | |
689 | if (cvm_oct_device[port]) { | |
690 | struct net_device *dev = cvm_oct_device[port]; | |
691 | cvm_oct_free_tx_skbs(dev); | |
692 | } | |
693 | } | |
694 | } | |
695 | ||
696 | static irqreturn_t cvm_oct_tx_cleanup_watchdog(int cpl, void *dev_id) | |
697 | { | |
698 | /* Disable the interrupt. */ | |
699 | cvmx_write_csr(CVMX_CIU_TIMX(1), 0); | |
700 | /* Do the work in the tasklet. */ | |
701 | tasklet_schedule(&cvm_oct_tx_cleanup_tasklet); | |
702 | return IRQ_HANDLED; | |
703 | } | |
704 | ||
705 | void cvm_oct_tx_initialize(void) | |
706 | { | |
707 | int i; | |
708 | ||
709 | /* Disable the interrupt. */ | |
710 | cvmx_write_csr(CVMX_CIU_TIMX(1), 0); | |
711 | /* Register an IRQ hander for to receive CIU_TIMX(1) interrupts */ | |
712 | i = request_irq(OCTEON_IRQ_TIMER1, | |
713 | cvm_oct_tx_cleanup_watchdog, 0, | |
714 | "Ethernet", cvm_oct_device); | |
715 | ||
716 | if (i) | |
717 | panic("Could not acquire Ethernet IRQ %d\n", OCTEON_IRQ_TIMER1); | |
718 | } | |
719 | ||
720 | void cvm_oct_tx_shutdown(void) | |
721 | { | |
722 | /* Free the interrupt handler */ | |
723 | free_irq(OCTEON_IRQ_TIMER1, cvm_oct_device); | |
724 | } |