2 * This file is based on code from OCTEON SDK by Cavium Networks.
4 * Copyright (c) 2003-2010 Cavium Networks
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.
11 #include <linux/module.h>
12 #include <linux/kernel.h>
13 #include <linux/cache.h>
14 #include <linux/cpumask.h>
15 #include <linux/netdevice.h>
16 #include <linux/etherdevice.h>
18 #include <linux/string.h>
19 #include <linux/prefetch.h>
20 #include <linux/ratelimit.h>
21 #include <linux/smp.h>
22 #include <linux/interrupt.h>
25 #include <linux/xfrm.h>
27 #endif /* CONFIG_XFRM */
29 #include <asm/octeon/octeon.h>
31 #include "ethernet-defines.h"
32 #include "ethernet-mem.h"
33 #include "ethernet-rx.h"
34 #include "octeon-ethernet.h"
35 #include "ethernet-util.h"
37 #include <asm/octeon/cvmx-helper.h>
38 #include <asm/octeon/cvmx-wqe.h>
39 #include <asm/octeon/cvmx-fau.h>
40 #include <asm/octeon/cvmx-pow.h>
41 #include <asm/octeon/cvmx-pip.h>
42 #include <asm/octeon/cvmx-scratch.h>
44 #include <asm/octeon/cvmx-gmxx-defs.h>
46 static struct napi_struct cvm_oct_napi;
49 * cvm_oct_do_interrupt - interrupt handler.
50 * @cpl: Interrupt number. Unused
51 * @dev_id: Cookie to identify the device. Unused
53 * The interrupt occurs whenever the POW has packets in our group.
56 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
58 /* Disable the IRQ and start napi_poll. */
59 disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
60 napi_schedule(&cvm_oct_napi);
66 * cvm_oct_check_rcv_error - process receive errors
67 * @work: Work queue entry pointing to the packet.
69 * Returns Non-zero if the packet can be dropped, zero otherwise.
71 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
75 if (octeon_has_feature(OCTEON_FEATURE_PKND))
76 port = work->word0.pip.cn68xx.pknd;
78 port = work->word1.cn38xx.ipprt;
80 if ((work->word2.snoip.err_code == 10) && (work->word1.len <= 64)) {
82 * Ignore length errors on min size packets. Some
83 * equipment incorrectly pads packets to 64+4FCS
84 * instead of 60+4FCS. Note these packets still get
85 * counted as frame errors.
87 } else if (work->word2.snoip.err_code == 5 ||
88 work->word2.snoip.err_code == 7) {
90 * We received a packet with either an alignment error
91 * or a FCS error. This may be signalling that we are
92 * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
93 * off. If this is the case we need to parse the
94 * packet to determine if we can remove a non spec
95 * preamble and generate a correct packet.
97 int interface = cvmx_helper_get_interface_num(port);
98 int index = cvmx_helper_get_interface_index_num(port);
99 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
101 gmxx_rxx_frm_ctl.u64 =
102 cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
103 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
105 cvmx_phys_to_ptr(work->packet_ptr.s.addr);
108 while (i < work->word1.len - 1) {
116 /* Port received 0xd5 preamble */
117 work->packet_ptr.s.addr += i + 1;
118 work->word1.len -= i + 5;
119 } else if ((*ptr & 0xf) == 0xd) {
120 /* Port received 0xd preamble */
121 work->packet_ptr.s.addr += i;
122 work->word1.len -= i + 4;
123 for (i = 0; i < work->word1.len; i++) {
125 ((*ptr & 0xf0) >> 4) |
126 ((*(ptr + 1) & 0xf) << 4);
130 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
132 cvm_oct_free_work(work);
137 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
138 port, work->word2.snoip.err_code);
139 cvm_oct_free_work(work);
147 * cvm_oct_napi_poll - the NAPI poll function.
148 * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
149 * @budget: Maximum number of packets to receive.
151 * Returns the number of packets processed.
153 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
155 const int coreid = cvmx_get_core_num();
159 int did_work_request = 0;
160 int packet_not_copied;
162 /* Prefetch cvm_oct_device since we know we need it soon */
163 prefetch(cvm_oct_device);
165 if (USE_ASYNC_IOBDMA) {
166 /* Save scratch in case userspace is using it */
168 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
171 /* Only allow work for our group (and preserve priorities) */
172 if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
173 old_group_mask = cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid));
174 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid),
175 1ull << pow_receive_group);
176 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
178 old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
179 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
180 (old_group_mask & ~0xFFFFull) |
181 1 << pow_receive_group);
184 if (USE_ASYNC_IOBDMA) {
185 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
186 did_work_request = 1;
189 while (rx_count < budget) {
190 struct sk_buff *skb = NULL;
191 struct sk_buff **pskb = NULL;
196 if (USE_ASYNC_IOBDMA && did_work_request)
197 work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
199 work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
202 did_work_request = 0;
204 if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
205 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS,
206 1ull << pow_receive_group);
207 cvmx_write_csr(CVMX_SSO_WQ_INT,
208 1ull << pow_receive_group);
210 union cvmx_pow_wq_int wq_int;
213 wq_int.s.iq_dis = 1 << pow_receive_group;
214 wq_int.s.wq_int = 1 << pow_receive_group;
215 cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
219 pskb = (struct sk_buff **)
220 (cvm_oct_get_buffer_ptr(work->packet_ptr) -
224 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
225 cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
227 did_work_request = 1;
231 skb_in_hw = work->word2.s.bufs == 1;
232 if (likely(skb_in_hw)) {
234 prefetch(&skb->head);
238 if (octeon_has_feature(OCTEON_FEATURE_PKND))
239 port = work->word0.pip.cn68xx.pknd;
241 port = work->word1.cn38xx.ipprt;
243 prefetch(cvm_oct_device[port]);
245 /* Immediately throw away all packets with receive errors */
246 if (unlikely(work->word2.snoip.rcv_error)) {
247 if (cvm_oct_check_rcv_error(work))
252 * We can only use the zero copy path if skbuffs are
253 * in the FPA pool and the packet fits in a single
256 if (likely(skb_in_hw)) {
257 skb->data = skb->head + work->packet_ptr.s.addr -
258 cvmx_ptr_to_phys(skb->head);
260 skb->len = work->word1.len;
261 skb_set_tail_pointer(skb, skb->len);
262 packet_not_copied = 1;
265 * We have to copy the packet. First allocate
268 skb = dev_alloc_skb(work->word1.len);
270 cvm_oct_free_work(work);
275 * Check if we've received a packet that was
276 * entirely stored in the work entry.
278 if (unlikely(work->word2.s.bufs == 0)) {
279 u8 *ptr = work->packet_data;
281 if (likely(!work->word2.s.not_IP)) {
283 * The beginning of the packet
284 * moves for IP packets.
286 if (work->word2.s.is_v6)
291 memcpy(skb_put(skb, work->word1.len), ptr,
293 /* No packet buffers to free */
295 int segments = work->word2.s.bufs;
296 union cvmx_buf_ptr segment_ptr =
298 int len = work->word1.len;
301 union cvmx_buf_ptr next_ptr =
302 *(union cvmx_buf_ptr *)
304 segment_ptr.s.addr - 8);
307 * Octeon Errata PKI-100: The segment size is
308 * wrong. Until it is fixed, calculate the
309 * segment size based on the packet pool
310 * buffer size. When it is fixed, the
311 * following line should be replaced with this
312 * one: int segment_size =
313 * segment_ptr.s.size;
316 CVMX_FPA_PACKET_POOL_SIZE -
317 (segment_ptr.s.addr -
318 (((segment_ptr.s.addr >> 7) -
319 segment_ptr.s.back) << 7));
321 * Don't copy more than what
322 * is left in the packet.
324 if (segment_size > len)
326 /* Copy the data into the packet */
327 memcpy(skb_put(skb, segment_size),
332 segment_ptr = next_ptr;
335 packet_not_copied = 0;
337 if (likely((port < TOTAL_NUMBER_OF_PORTS) &&
338 cvm_oct_device[port])) {
339 struct net_device *dev = cvm_oct_device[port];
340 struct octeon_ethernet *priv = netdev_priv(dev);
343 * Only accept packets for devices that are
346 if (likely(dev->flags & IFF_UP)) {
347 skb->protocol = eth_type_trans(skb, dev);
350 if (unlikely(work->word2.s.not_IP ||
351 work->word2.s.IP_exc ||
352 work->word2.s.L4_error ||
353 !work->word2.s.tcp_or_udp))
354 skb->ip_summed = CHECKSUM_NONE;
356 skb->ip_summed = CHECKSUM_UNNECESSARY;
358 /* Increment RX stats for virtual ports */
359 if (port >= CVMX_PIP_NUM_INPUT_PORTS) {
360 priv->stats.rx_packets++;
361 priv->stats.rx_bytes += skb->len;
363 netif_receive_skb(skb);
366 * Drop any packet received for a device that
369 priv->stats.rx_dropped++;
370 dev_kfree_skb_irq(skb);
374 * Drop any packet received for a device that
377 printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
379 dev_kfree_skb_irq(skb);
382 * Check to see if the skbuff and work share the same
385 if (likely(packet_not_copied)) {
387 * This buffer needs to be replaced, increment
388 * the number of buffers we need to free by
391 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
394 cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
396 cvm_oct_free_work(work);
399 /* Restore the original POW group mask */
400 if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
401 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid), old_group_mask);
402 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
404 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
407 if (USE_ASYNC_IOBDMA) {
408 /* Restore the scratch area */
409 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
411 cvm_oct_rx_refill_pool(0);
413 if (rx_count < budget && napi) {
416 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
421 #ifdef CONFIG_NET_POLL_CONTROLLER
423 * cvm_oct_poll_controller - poll for receive packets
426 * @dev: Device to poll. Unused
428 void cvm_oct_poll_controller(struct net_device *dev)
430 cvm_oct_napi_poll(NULL, 16);
434 void cvm_oct_rx_initialize(void)
437 struct net_device *dev_for_napi = NULL;
439 for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
440 if (cvm_oct_device[i]) {
441 dev_for_napi = cvm_oct_device[i];
447 panic("No net_devices were allocated.");
449 netif_napi_add(dev_for_napi, &cvm_oct_napi, cvm_oct_napi_poll,
451 napi_enable(&cvm_oct_napi);
453 /* Register an IRQ handler to receive POW interrupts */
454 i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
455 cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
458 panic("Could not acquire Ethernet IRQ %d\n",
459 OCTEON_IRQ_WORKQ0 + pow_receive_group);
461 disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
463 /* Enable POW interrupt when our port has at least one packet */
464 if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
465 union cvmx_sso_wq_int_thrx int_thr;
466 union cvmx_pow_wq_int_pc int_pc;
470 int_thr.s.tc_thr = 1;
471 cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(pow_receive_group),
476 cvmx_write_csr(CVMX_SSO_WQ_INT_PC, int_pc.u64);
478 union cvmx_pow_wq_int_thrx int_thr;
479 union cvmx_pow_wq_int_pc int_pc;
483 int_thr.s.tc_thr = 1;
484 cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group),
489 cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
492 /* Schedule NAPI now. This will indirectly enable the interrupt. */
493 napi_schedule(&cvm_oct_napi);
496 void cvm_oct_rx_shutdown(void)
498 netif_napi_del(&cvm_oct_napi);