Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
ethernet-rx.c
Go to the documentation of this file.
1 /**********************************************************************
2  * Author: Cavium Networks
3  *
4  * Contact: [email protected]
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/cache.h>
30 #include <linux/cpumask.h>
31 #include <linux/netdevice.h>
32 #include <linux/init.h>
33 #include <linux/etherdevice.h>
34 #include <linux/ip.h>
35 #include <linux/string.h>
36 #include <linux/prefetch.h>
37 #include <linux/ratelimit.h>
38 #include <linux/smp.h>
39 #include <linux/interrupt.h>
40 #include <net/dst.h>
41 #ifdef CONFIG_XFRM
42 #include <linux/xfrm.h>
43 #include <net/xfrm.h>
44 #endif /* CONFIG_XFRM */
45 
46 #include <linux/atomic.h>
47 
48 #include <asm/octeon/octeon.h>
49 
50 #include "ethernet-defines.h"
51 #include "ethernet-mem.h"
52 #include "ethernet-rx.h"
53 #include "octeon-ethernet.h"
54 #include "ethernet-util.h"
55 
56 #include <asm/octeon/cvmx-helper.h>
57 #include <asm/octeon/cvmx-wqe.h>
58 #include <asm/octeon/cvmx-fau.h>
59 #include <asm/octeon/cvmx-pow.h>
60 #include <asm/octeon/cvmx-pip.h>
62 
64 
66  struct napi_struct napi;
68 
69 static struct cvm_napi_wrapper cvm_oct_napi[NR_CPUS] __cacheline_aligned_in_smp;
70 
73  /*
74  * The number of additional cores that could be processing
75  * input packtes.
76  */
80 
82 
83 static void cvm_oct_enable_napi(void *_)
84 {
85  int cpu = smp_processor_id();
86  napi_schedule(&cvm_oct_napi[cpu].napi);
87 }
88 
89 static void cvm_oct_enable_one_cpu(void)
90 {
91  int v;
92  int cpu;
93 
94  /* Check to see if more CPUs are available for receive processing... */
95  v = atomic_sub_if_positive(1, &core_state.available_cores);
96  if (v < 0)
97  return;
98 
99  /* ... if a CPU is available, Turn on NAPI polling for that CPU. */
100  for_each_online_cpu(cpu) {
101  if (!cpu_test_and_set(cpu, core_state.cpu_state)) {
102  v = smp_call_function_single(cpu, cvm_oct_enable_napi,
103  NULL, 0);
104  if (v)
105  panic("Can't enable NAPI.");
106  break;
107  }
108  }
109 }
110 
111 static void cvm_oct_no_more_work(void)
112 {
113  int cpu = smp_processor_id();
114 
115  /*
116  * CPU zero is special. It always has the irq enabled when
117  * waiting for incoming packets.
118  */
119  if (cpu == 0) {
121  return;
122  }
123 
124  cpu_clear(cpu, core_state.cpu_state);
125  atomic_add(1, &core_state.available_cores);
126 }
127 
134 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
135 {
136  /* Disable the IRQ and start napi_poll. */
138  cvm_oct_enable_napi(NULL);
139 
140  return IRQ_HANDLED;
141 }
142 
149 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
150 {
151  if ((work->word2.snoip.err_code == 10) && (work->len <= 64)) {
152  /*
153  * Ignore length errors on min size packets. Some
154  * equipment incorrectly pads packets to 64+4FCS
155  * instead of 60+4FCS. Note these packets still get
156  * counted as frame errors.
157  */
158  } else
160  && ((work->word2.snoip.err_code == 5)
161  || (work->word2.snoip.err_code == 7))) {
162 
163  /*
164  * We received a packet with either an alignment error
165  * or a FCS error. This may be signalling that we are
166  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
167  * off. If this is the case we need to parse the
168  * packet to determine if we can remove a non spec
169  * preamble and generate a correct packet.
170  */
171  int interface = cvmx_helper_get_interface_num(work->ipprt);
172  int index = cvmx_helper_get_interface_index_num(work->ipprt);
173  union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
174  gmxx_rxx_frm_ctl.u64 =
175  cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
176  if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
177 
178  uint8_t *ptr =
179  cvmx_phys_to_ptr(work->packet_ptr.s.addr);
180  int i = 0;
181 
182  while (i < work->len - 1) {
183  if (*ptr != 0x55)
184  break;
185  ptr++;
186  i++;
187  }
188 
189  if (*ptr == 0xd5) {
190  /*
191  printk_ratelimited("Port %d received 0xd5 preamble\n", work->ipprt);
192  */
193  work->packet_ptr.s.addr += i + 1;
194  work->len -= i + 5;
195  } else if ((*ptr & 0xf) == 0xd) {
196  /*
197  printk_ratelimited("Port %d received 0x?d preamble\n", work->ipprt);
198  */
199  work->packet_ptr.s.addr += i;
200  work->len -= i + 4;
201  for (i = 0; i < work->len; i++) {
202  *ptr =
203  ((*ptr & 0xf0) >> 4) |
204  ((*(ptr + 1) & 0xf) << 4);
205  ptr++;
206  }
207  } else {
208  printk_ratelimited("Port %d unknown preamble, packet "
209  "dropped\n",
210  work->ipprt);
211  /*
212  cvmx_helper_dump_packet(work);
213  */
214  cvm_oct_free_work(work);
215  return 1;
216  }
217  }
218  } else {
219  printk_ratelimited("Port %d receive error code %d, packet dropped\n",
220  work->ipprt, work->word2.snoip.err_code);
221  cvm_oct_free_work(work);
222  return 1;
223  }
224 
225  return 0;
226 }
227 
235 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
236 {
237  const int coreid = cvmx_get_core_num();
238  uint64_t old_group_mask;
239  uint64_t old_scratch;
240  int rx_count = 0;
241  int did_work_request = 0;
242  int packet_not_copied;
243 
244  /* Prefetch cvm_oct_device since we know we need it soon */
246 
247  if (USE_ASYNC_IOBDMA) {
248  /* Save scratch in case userspace is using it */
250  old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
251  }
252 
253  /* Only allow work for our group (and preserve priorities) */
254  old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
255  cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
256  (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
257 
258  if (USE_ASYNC_IOBDMA) {
259  cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
260  did_work_request = 1;
261  }
262 
263  while (rx_count < budget) {
264  struct sk_buff *skb = NULL;
265  struct sk_buff **pskb = NULL;
266  int skb_in_hw;
267  cvmx_wqe_t *work;
268 
269  if (USE_ASYNC_IOBDMA && did_work_request)
270  work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
271  else
272  work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
273 
274  prefetch(work);
275  did_work_request = 0;
276  if (work == NULL) {
277  union cvmx_pow_wq_int wq_int;
278  wq_int.u64 = 0;
279  wq_int.s.iq_dis = 1 << pow_receive_group;
280  wq_int.s.wq_int = 1 << pow_receive_group;
281  cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
282  break;
283  }
284  pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) - sizeof(void *));
285  prefetch(pskb);
286 
287  if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
288  cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
289  did_work_request = 1;
290  }
291 
292  if (rx_count == 0) {
293  /*
294  * First time through, see if there is enough
295  * work waiting to merit waking another
296  * CPU.
297  */
298  union cvmx_pow_wq_int_cntx counts;
299  int backlog;
300  int cores_in_use = core_state.baseline_cores - atomic_read(&core_state.available_cores);
301  counts.u64 = cvmx_read_csr(CVMX_POW_WQ_INT_CNTX(pow_receive_group));
302  backlog = counts.s.iq_cnt + counts.s.ds_cnt;
303  if (backlog > budget * cores_in_use && napi != NULL)
304  cvm_oct_enable_one_cpu();
305  }
306 
307  skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
308  if (likely(skb_in_hw)) {
309  skb = *pskb;
310  prefetch(&skb->head);
311  prefetch(&skb->len);
312  }
313  prefetch(cvm_oct_device[work->ipprt]);
314 
315  /* Immediately throw away all packets with receive errors */
316  if (unlikely(work->word2.snoip.rcv_error)) {
317  if (cvm_oct_check_rcv_error(work))
318  continue;
319  }
320 
321  /*
322  * We can only use the zero copy path if skbuffs are
323  * in the FPA pool and the packet fits in a single
324  * buffer.
325  */
326  if (likely(skb_in_hw)) {
327  skb->data = skb->head + work->packet_ptr.s.addr - cvmx_ptr_to_phys(skb->head);
328  prefetch(skb->data);
329  skb->len = work->len;
330  skb_set_tail_pointer(skb, skb->len);
331  packet_not_copied = 1;
332  } else {
333  /*
334  * We have to copy the packet. First allocate
335  * an skbuff for it.
336  */
337  skb = dev_alloc_skb(work->len);
338  if (!skb) {
339  printk_ratelimited("Port %d failed to allocate "
340  "skbuff, packet dropped\n",
341  work->ipprt);
342  cvm_oct_free_work(work);
343  continue;
344  }
345 
346  /*
347  * Check if we've received a packet that was
348  * entirely stored in the work entry.
349  */
350  if (unlikely(work->word2.s.bufs == 0)) {
351  uint8_t *ptr = work->packet_data;
352 
353  if (likely(!work->word2.s.not_IP)) {
354  /*
355  * The beginning of the packet
356  * moves for IP packets.
357  */
358  if (work->word2.s.is_v6)
359  ptr += 2;
360  else
361  ptr += 6;
362  }
363  memcpy(skb_put(skb, work->len), ptr, work->len);
364  /* No packet buffers to free */
365  } else {
366  int segments = work->word2.s.bufs;
367  union cvmx_buf_ptr segment_ptr = work->packet_ptr;
368  int len = work->len;
369 
370  while (segments--) {
371  union cvmx_buf_ptr next_ptr =
372  *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
373 
374  /*
375  * Octeon Errata PKI-100: The segment size is
376  * wrong. Until it is fixed, calculate the
377  * segment size based on the packet pool
378  * buffer size. When it is fixed, the
379  * following line should be replaced with this
380  * one: int segment_size =
381  * segment_ptr.s.size;
382  */
383  int segment_size = CVMX_FPA_PACKET_POOL_SIZE -
384  (segment_ptr.s.addr - (((segment_ptr.s.addr >> 7) - segment_ptr.s.back) << 7));
385  /*
386  * Don't copy more than what
387  * is left in the packet.
388  */
389  if (segment_size > len)
390  segment_size = len;
391  /* Copy the data into the packet */
392  memcpy(skb_put(skb, segment_size),
393  cvmx_phys_to_ptr(segment_ptr.s.addr),
394  segment_size);
395  len -= segment_size;
396  segment_ptr = next_ptr;
397  }
398  }
399  packet_not_copied = 0;
400  }
401 
402  if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
403  cvm_oct_device[work->ipprt])) {
404  struct net_device *dev = cvm_oct_device[work->ipprt];
405  struct octeon_ethernet *priv = netdev_priv(dev);
406 
407  /*
408  * Only accept packets for devices that are
409  * currently up.
410  */
411  if (likely(dev->flags & IFF_UP)) {
412  skb->protocol = eth_type_trans(skb, dev);
413  skb->dev = dev;
414 
415  if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc ||
416  work->word2.s.L4_error || !work->word2.s.tcp_or_udp))
417  skb->ip_summed = CHECKSUM_NONE;
418  else
420 
421  /* Increment RX stats for virtual ports */
422  if (work->ipprt >= CVMX_PIP_NUM_INPUT_PORTS) {
423 #ifdef CONFIG_64BIT
424  atomic64_add(1, (atomic64_t *)&priv->stats.rx_packets);
425  atomic64_add(skb->len, (atomic64_t *)&priv->stats.rx_bytes);
426 #else
427  atomic_add(1, (atomic_t *)&priv->stats.rx_packets);
428  atomic_add(skb->len, (atomic_t *)&priv->stats.rx_bytes);
429 #endif
430  }
431  netif_receive_skb(skb);
432  rx_count++;
433  } else {
434  /* Drop any packet received for a device that isn't up */
435  /*
436  printk_ratelimited("%s: Device not up, packet dropped\n",
437  dev->name);
438  */
439 #ifdef CONFIG_64BIT
440  atomic64_add(1, (atomic64_t *)&priv->stats.rx_dropped);
441 #else
442  atomic_add(1, (atomic_t *)&priv->stats.rx_dropped);
443 #endif
444  dev_kfree_skb_irq(skb);
445  }
446  } else {
447  /*
448  * Drop any packet received for a device that
449  * doesn't exist.
450  */
451  printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
452  work->ipprt);
453  dev_kfree_skb_irq(skb);
454  }
455  /*
456  * Check to see if the skbuff and work share the same
457  * packet buffer.
458  */
459  if (USE_SKBUFFS_IN_HW && likely(packet_not_copied)) {
460  /*
461  * This buffer needs to be replaced, increment
462  * the number of buffers we need to free by
463  * one.
464  */
465  cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
466  1);
467 
468  cvmx_fpa_free(work, CVMX_FPA_WQE_POOL,
469  DONT_WRITEBACK(1));
470  } else {
471  cvm_oct_free_work(work);
472  }
473  }
474  /* Restore the original POW group mask */
475  cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
476  if (USE_ASYNC_IOBDMA) {
477  /* Restore the scratch area */
478  cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
479  }
480  cvm_oct_rx_refill_pool(0);
481 
482  if (rx_count < budget && napi != NULL) {
483  /* No more work */
484  napi_complete(napi);
485  cvm_oct_no_more_work();
486  }
487  return rx_count;
488 }
489 
490 #ifdef CONFIG_NET_POLL_CONTROLLER
491 
497 void cvm_oct_poll_controller(struct net_device *dev)
498 {
499  cvm_oct_napi_poll(NULL, 16);
500 }
501 #endif
502 
504 {
505  int i;
506  struct net_device *dev_for_napi = NULL;
507  union cvmx_pow_wq_int_thrx int_thr;
508  union cvmx_pow_wq_int_pc int_pc;
509 
510  for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
511  if (cvm_oct_device[i]) {
512  dev_for_napi = cvm_oct_device[i];
513  break;
514  }
515  }
516 
517  if (NULL == dev_for_napi)
518  panic("No net_devices were allocated.");
519 
520  if (max_rx_cpus > 1 && max_rx_cpus < num_online_cpus())
521  atomic_set(&core_state.available_cores, max_rx_cpus);
522  else
523  atomic_set(&core_state.available_cores, num_online_cpus());
524  core_state.baseline_cores = atomic_read(&core_state.available_cores);
525 
526  core_state.cpu_state = CPU_MASK_NONE;
528  netif_napi_add(dev_for_napi, &cvm_oct_napi[i].napi,
529  cvm_oct_napi_poll, rx_napi_weight);
530  napi_enable(&cvm_oct_napi[i].napi);
531  }
532  /* Register an IRQ hander for to receive POW interrupts */
534  cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
535 
536  if (i)
537  panic("Could not acquire Ethernet IRQ %d\n",
539 
541 
542  int_thr.u64 = 0;
543  int_thr.s.tc_en = 1;
544  int_thr.s.tc_thr = 1;
545  /* Enable POW interrupt when our port has at least one packet */
546  cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
547 
548  int_pc.u64 = 0;
549  int_pc.s.pc_thr = 5;
550  cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
551 
552 
553  /* Scheduld NAPI now. This will indirectly enable interrupts. */
554  cvm_oct_enable_one_cpu();
555 }
556 
558 {
559  int i;
560  /* Shutdown all of the NAPIs */
562  netif_napi_del(&cvm_oct_napi[i].napi);
563 }