1 /*********************************************************************
2 * Author: Cavium Networks
4 * Contact: support@caviumnetworks.com
5 * This file is part of the OCTEON SDK
7 * Copyright (c) 2003-2007 Cavium Networks
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.
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
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/.
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>
33 #include <linux/string.h>
34 #include <linux/ethtool.h>
35 #include <linux/mii.h>
36 #include <linux/seq_file.h>
37 #include <linux/proc_fs.h>
40 #include <linux/xfrm.h>
42 #endif /* CONFIG_XFRM */
44 #include <asm/atomic.h>
46 #include <asm/octeon/octeon.h>
48 #include "ethernet-defines.h"
49 #include "octeon-ethernet.h"
50 #include "ethernet-util.h"
55 #include "cvmx-helper.h"
57 #include "cvmx-gmxx-defs.h"
60 * You can define GET_SKBUFF_QOS() to override how the skbuff output
61 * function determines which output queue is used. The default
62 * implementation always uses the base queue for the port. If, for
63 * example, you wanted to use the skb->priority fieid, define
64 * GET_SKBUFF_QOS as: #define GET_SKBUFF_QOS(skb) ((skb)->priority)
66 #ifndef GET_SKBUFF_QOS
67 #define GET_SKBUFF_QOS(skb) 0
73 * @skb: Packet to send
74 * @dev: Device info structure
75 * Returns Always returns zero
77 int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
79 cvmx_pko_command_word0_t pko_command;
80 union cvmx_buf_ptr hw_buffer;
82 uint64_t old_scratch2;
85 struct octeon_ethernet *priv = netdev_priv(dev);
87 int32_t buffers_to_free;
88 #if REUSE_SKBUFFS_WITHOUT_FREE
89 unsigned char *fpa_head;
93 * Prefetch the private data structure. It is larger that one
98 /* Start off assuming no drop */
102 * The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to
103 * completely remove "qos" in the event neither interface
104 * supports multiple queues per port.
106 if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
107 (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
108 qos = GET_SKBUFF_QOS(skb);
111 else if (qos >= cvmx_pko_get_num_queues(priv->port))
116 if (USE_ASYNC_IOBDMA) {
117 /* Save scratch in case userspace is using it */
119 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
120 old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
123 * Assume we're going to be able t osend this
124 * packet. Fetch and increment the number of pending
125 * packets for output.
127 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
128 FAU_NUM_PACKET_BUFFERS_TO_FREE,
130 cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
131 priv->fau + qos * 4, 1);
135 * The CN3XXX series of parts has an errata (GMX-401) which
136 * causes the GMX block to hang if a collision occurs towards
137 * the end of a <68 byte packet. As a workaround for this, we
138 * pad packets to be 68 bytes whenever we are in half duplex
139 * mode. We don't handle the case of having a small packet but
140 * no room to add the padding. The kernel should always give
141 * us at least a cache line
143 if ((skb->len < 64) && OCTEON_IS_MODEL(OCTEON_CN3XXX)) {
144 union cvmx_gmxx_prtx_cfg gmx_prt_cfg;
145 int interface = INTERFACE(priv->port);
146 int index = INDEX(priv->port);
149 /* We only need to pad packet in half duplex mode */
151 cvmx_read_csr(CVMX_GMXX_PRTX_CFG(index, interface));
152 if (gmx_prt_cfg.s.duplex == 0) {
153 int add_bytes = 64 - skb->len;
154 if ((skb_tail_pointer(skb) + add_bytes) <=
155 skb_end_pointer(skb))
156 memset(__skb_put(skb, add_bytes), 0,
162 /* Build the PKO buffer pointer */
164 hw_buffer.s.addr = cvmx_ptr_to_phys(skb->data);
165 hw_buffer.s.pool = 0;
167 (unsigned long)skb_end_pointer(skb) - (unsigned long)skb->head;
169 /* Build the PKO command */
171 pko_command.s.n2 = 1; /* Don't pollute L2 with the outgoing packet */
172 pko_command.s.segs = 1;
173 pko_command.s.total_bytes = skb->len;
174 pko_command.s.size0 = CVMX_FAU_OP_SIZE_32;
175 pko_command.s.subone0 = 1;
177 pko_command.s.dontfree = 1;
178 pko_command.s.reg0 = priv->fau + qos * 4;
180 * See if we can put this skb in the FPA pool. Any strange
181 * behavior from the Linux networking stack will most likely
182 * be caused by a bug in the following code. If some field is
183 * in use by the network stack and get carried over when a
184 * buffer is reused, bad thing may happen. If in doubt and
185 * you dont need the absolute best performance, disable the
186 * define REUSE_SKBUFFS_WITHOUT_FREE. The reuse of buffers has
187 * shown a 25% increase in performance under some loads.
189 #if REUSE_SKBUFFS_WITHOUT_FREE
190 fpa_head = skb->head + 128 - ((unsigned long)skb->head & 0x7f);
191 if (unlikely(skb->data < fpa_head)) {
193 * printk("TX buffer beginning can't meet FPA
194 * alignment constraints\n");
196 goto dont_put_skbuff_in_hw;
199 ((skb_end_pointer(skb) - fpa_head) < CVMX_FPA_PACKET_POOL_SIZE)) {
201 printk("TX buffer isn't large enough for the FPA\n");
203 goto dont_put_skbuff_in_hw;
205 if (unlikely(skb_shared(skb))) {
207 printk("TX buffer sharing data with someone else\n");
209 goto dont_put_skbuff_in_hw;
211 if (unlikely(skb_cloned(skb))) {
213 printk("TX buffer has been cloned\n");
215 goto dont_put_skbuff_in_hw;
217 if (unlikely(skb_header_cloned(skb))) {
219 printk("TX buffer header has been cloned\n");
221 goto dont_put_skbuff_in_hw;
223 if (unlikely(skb->destructor)) {
225 printk("TX buffer has a destructor\n");
227 goto dont_put_skbuff_in_hw;
229 if (unlikely(skb_shinfo(skb)->nr_frags)) {
231 printk("TX buffer has fragments\n");
233 goto dont_put_skbuff_in_hw;
237 sizeof(*skb) + skb_end_pointer(skb) - skb->head)) {
239 printk("TX buffer truesize has been changed\n");
241 goto dont_put_skbuff_in_hw;
245 * We can use this buffer in the FPA. We don't need the FAU
248 pko_command.s.reg0 = 0;
249 pko_command.s.dontfree = 0;
251 hw_buffer.s.back = (skb->data - fpa_head) >> 7;
252 *(struct sk_buff **)(fpa_head - sizeof(void *)) = skb;
255 * The skbuff will be reused without ever being freed. We must
256 * cleanup a bunch of Linux stuff.
258 dst_release(skb->dst);
261 secpath_put(skb->sp);
266 #ifdef CONFIG_NET_SCHED
268 #ifdef CONFIG_NET_CLS_ACT
270 #endif /* CONFIG_NET_CLS_ACT */
271 #endif /* CONFIG_NET_SCHED */
273 dont_put_skbuff_in_hw:
274 #endif /* REUSE_SKBUFFS_WITHOUT_FREE */
276 /* Check if we can use the hardware checksumming */
277 if (USE_HW_TCPUDP_CHECKSUM && (skb->protocol == htons(ETH_P_IP)) &&
278 (ip_hdr(skb)->version == 4) && (ip_hdr(skb)->ihl == 5) &&
279 ((ip_hdr(skb)->frag_off == 0) || (ip_hdr(skb)->frag_off == 1 << 14))
280 && ((ip_hdr(skb)->protocol == IP_PROTOCOL_TCP)
281 || (ip_hdr(skb)->protocol == IP_PROTOCOL_UDP))) {
282 /* Use hardware checksum calc */
283 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
286 if (USE_ASYNC_IOBDMA) {
287 /* Get the number of skbuffs in use by the hardware */
289 in_use = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
290 buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
292 /* Get the number of skbuffs in use by the hardware */
293 in_use = cvmx_fau_fetch_and_add32(priv->fau + qos * 4, 1);
295 cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
299 * If we're sending faster than the receive can free them then
300 * don't do the HW free.
302 if ((buffers_to_free < -100) && !pko_command.s.dontfree) {
303 pko_command.s.dontfree = 1;
304 pko_command.s.reg0 = priv->fau + qos * 4;
307 cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
308 CVMX_PKO_LOCK_CMD_QUEUE);
310 /* Drop this packet if we have too many already queued to the HW */
312 (skb_queue_len(&priv->tx_free_list[qos]) >= MAX_OUT_QUEUE_DEPTH)) {
314 DEBUGPRINT("%s: Tx dropped. Too many queued\n", dev->name);
318 /* Send the packet to the output queue */
320 (cvmx_pko_send_packet_finish
321 (priv->port, priv->queue + qos, pko_command, hw_buffer,
322 CVMX_PKO_LOCK_CMD_QUEUE))) {
323 DEBUGPRINT("%s: Failed to send the packet\n", dev->name);
327 if (USE_ASYNC_IOBDMA) {
328 /* Restore the scratch area */
329 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
330 cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
333 if (unlikely(dropped)) {
334 dev_kfree_skb_any(skb);
335 cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
336 priv->stats.tx_dropped++;
338 if (USE_SKBUFFS_IN_HW) {
339 /* Put this packet on the queue to be freed later */
340 if (pko_command.s.dontfree)
341 skb_queue_tail(&priv->tx_free_list[qos], skb);
343 cvmx_fau_atomic_add32
344 (FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
345 cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
348 /* Put this packet on the queue to be freed later */
349 skb_queue_tail(&priv->tx_free_list[qos], skb);
353 /* Free skbuffs not in use by the hardware, possibly two at a time */
354 if (skb_queue_len(&priv->tx_free_list[qos]) > in_use) {
355 spin_lock(&priv->tx_free_list[qos].lock);
357 * Check again now that we have the lock. It might
360 if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
361 dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
362 if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
363 dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
364 spin_unlock(&priv->tx_free_list[qos].lock);
371 * Packet transmit to the POW
373 * @skb: Packet to send
374 * @dev: Device info structure
375 * Returns Always returns zero
377 int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
379 struct octeon_ethernet *priv = netdev_priv(dev);
383 /* Get a work queue entry */
384 cvmx_wqe_t *work = cvmx_fpa_alloc(CVMX_FPA_WQE_POOL);
385 if (unlikely(work == NULL)) {
386 DEBUGPRINT("%s: Failed to allocate a work queue entry\n",
388 priv->stats.tx_dropped++;
393 /* Get a packet buffer */
394 packet_buffer = cvmx_fpa_alloc(CVMX_FPA_PACKET_POOL);
395 if (unlikely(packet_buffer == NULL)) {
396 DEBUGPRINT("%s: Failed to allocate a packet buffer\n",
398 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
399 priv->stats.tx_dropped++;
405 * Calculate where we need to copy the data to. We need to
406 * leave 8 bytes for a next pointer (unused). We also need to
407 * include any configure skip. Then we need to align the IP
408 * packet src and dest into the same 64bit word. The below
409 * calculation may add a little extra, but that doesn't
412 copy_location = packet_buffer + sizeof(uint64_t);
413 copy_location += ((CVMX_HELPER_FIRST_MBUFF_SKIP + 7) & 0xfff8) + 6;
416 * We have to copy the packet since whoever processes this
417 * packet will free it to a hardware pool. We can't use the
418 * trick of counting outstanding packets like in
421 memcpy(copy_location, skb->data, skb->len);
424 * Fill in some of the work queue fields. We may need to add
425 * more if the software at the other end needs them.
427 work->hw_chksum = skb->csum;
428 work->len = skb->len;
429 work->ipprt = priv->port;
430 work->qos = priv->port & 0x7;
431 work->grp = pow_send_group;
432 work->tag_type = CVMX_HELPER_INPUT_TAG_TYPE;
433 work->tag = pow_send_group; /* FIXME */
434 /* Default to zero. Sets of zero later are commented out */
436 work->word2.s.bufs = 1;
437 work->packet_ptr.u64 = 0;
438 work->packet_ptr.s.addr = cvmx_ptr_to_phys(copy_location);
439 work->packet_ptr.s.pool = CVMX_FPA_PACKET_POOL;
440 work->packet_ptr.s.size = CVMX_FPA_PACKET_POOL_SIZE;
441 work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;
443 if (skb->protocol == htons(ETH_P_IP)) {
444 work->word2.s.ip_offset = 14;
446 work->word2.s.vlan_valid = 0; /* FIXME */
447 work->word2.s.vlan_cfi = 0; /* FIXME */
448 work->word2.s.vlan_id = 0; /* FIXME */
449 work->word2.s.dec_ipcomp = 0; /* FIXME */
451 work->word2.s.tcp_or_udp =
452 (ip_hdr(skb)->protocol == IP_PROTOCOL_TCP)
453 || (ip_hdr(skb)->protocol == IP_PROTOCOL_UDP);
456 work->word2.s.dec_ipsec = 0;
457 /* We only support IPv4 right now */
458 work->word2.s.is_v6 = 0;
459 /* Hardware would set to zero */
460 work->word2.s.software = 0;
461 /* No error, packet is internal */
462 work->word2.s.L4_error = 0;
464 work->word2.s.is_frag = !((ip_hdr(skb)->frag_off == 0)
465 || (ip_hdr(skb)->frag_off ==
468 /* Assume Linux is sending a good packet */
469 work->word2.s.IP_exc = 0;
471 work->word2.s.is_bcast = (skb->pkt_type == PACKET_BROADCAST);
472 work->word2.s.is_mcast = (skb->pkt_type == PACKET_MULTICAST);
474 /* This is an IP packet */
475 work->word2.s.not_IP = 0;
476 /* No error, packet is internal */
477 work->word2.s.rcv_error = 0;
478 /* No error, packet is internal */
479 work->word2.s.err_code = 0;
483 * When copying the data, include 4 bytes of the
484 * ethernet header to align the same way hardware
487 memcpy(work->packet_data, skb->data + 10,
488 sizeof(work->packet_data));
491 work->word2.snoip.vlan_valid = 0; /* FIXME */
492 work->word2.snoip.vlan_cfi = 0; /* FIXME */
493 work->word2.snoip.vlan_id = 0; /* FIXME */
494 work->word2.snoip.software = 0; /* Hardware would set to zero */
496 work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
497 work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
498 work->word2.snoip.is_bcast =
499 (skb->pkt_type == PACKET_BROADCAST);
500 work->word2.snoip.is_mcast =
501 (skb->pkt_type == PACKET_MULTICAST);
502 work->word2.snoip.not_IP = 1; /* IP was done up above */
504 /* No error, packet is internal */
505 work->word2.snoip.rcv_error = 0;
506 /* No error, packet is internal */
507 work->word2.snoip.err_code = 0;
509 memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
512 /* Submit the packet to the POW */
513 cvmx_pow_work_submit(work, work->tag, work->tag_type, work->qos,
515 priv->stats.tx_packets++;
516 priv->stats.tx_bytes += skb->len;
522 * Transmit a work queue entry out of the ethernet port. Both
523 * the work queue entry and the packet data can optionally be
524 * freed. The work will be freed on error as well.
526 * @dev: Device to transmit out.
528 * Work queue entry to send
529 * @do_free: True if the work queue entry and packet data should be
530 * freed. If false, neither will be freed.
531 * @qos: Index into the queues for this port to transmit on. This
532 * is used to implement QoS if their are multiple queues per
533 * port. This parameter must be between 0 and the number of
534 * queues per port minus 1. Values outside of this range will
537 * Returns Zero on success, negative on failure.
539 int cvm_oct_transmit_qos(struct net_device *dev, void *work_queue_entry,
540 int do_free, int qos)
543 union cvmx_buf_ptr hw_buffer;
544 cvmx_pko_command_word0_t pko_command;
546 struct octeon_ethernet *priv = netdev_priv(dev);
547 cvmx_wqe_t *work = work_queue_entry;
549 if (!(dev->flags & IFF_UP)) {
550 DEBUGPRINT("%s: Device not up\n", dev->name);
552 cvm_oct_free_work(work);
556 /* The check on CVMX_PKO_QUEUES_PER_PORT_* is designed to completely
557 remove "qos" in the event neither interface supports
558 multiple queues per port */
559 if ((CVMX_PKO_QUEUES_PER_PORT_INTERFACE0 > 1) ||
560 (CVMX_PKO_QUEUES_PER_PORT_INTERFACE1 > 1)) {
563 else if (qos >= cvmx_pko_get_num_queues(priv->port))
568 /* Start off assuming no drop */
571 local_irq_save(flags);
572 cvmx_pko_send_packet_prepare(priv->port, priv->queue + qos,
573 CVMX_PKO_LOCK_CMD_QUEUE);
575 /* Build the PKO buffer pointer */
577 hw_buffer.s.addr = work->packet_ptr.s.addr;
578 hw_buffer.s.pool = CVMX_FPA_PACKET_POOL;
579 hw_buffer.s.size = CVMX_FPA_PACKET_POOL_SIZE;
580 hw_buffer.s.back = work->packet_ptr.s.back;
582 /* Build the PKO command */
584 pko_command.s.n2 = 1; /* Don't pollute L2 with the outgoing packet */
585 pko_command.s.dontfree = !do_free;
586 pko_command.s.segs = work->word2.s.bufs;
587 pko_command.s.total_bytes = work->len;
589 /* Check if we can use the hardware checksumming */
590 if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc))
591 pko_command.s.ipoffp1 = 0;
593 pko_command.s.ipoffp1 = sizeof(struct ethhdr) + 1;
595 /* Send the packet to the output queue */
597 (cvmx_pko_send_packet_finish
598 (priv->port, priv->queue + qos, pko_command, hw_buffer,
599 CVMX_PKO_LOCK_CMD_QUEUE))) {
600 DEBUGPRINT("%s: Failed to send the packet\n", dev->name);
603 local_irq_restore(flags);
605 if (unlikely(dropped)) {
607 cvm_oct_free_work(work);
608 priv->stats.tx_dropped++;
610 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, DONT_WRITEBACK(1));
614 EXPORT_SYMBOL(cvm_oct_transmit_qos);
617 * This function frees all skb that are currenty queued for TX.
619 * @dev: Device being shutdown
621 void cvm_oct_tx_shutdown(struct net_device *dev)
623 struct octeon_ethernet *priv = netdev_priv(dev);
627 for (qos = 0; qos < 16; qos++) {
628 spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
629 while (skb_queue_len(&priv->tx_free_list[qos]))
630 dev_kfree_skb_any(__skb_dequeue
631 (&priv->tx_free_list[qos]));
632 spin_unlock_irqrestore(&priv->tx_free_list[qos].lock, flags);