Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
tilegx.c
Go to the documentation of this file.
1 /*
2  * Copyright 2012 Tilera Corporation. All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License
6  * as published by the Free Software Foundation, version 2.
7  *
8  * This program is distributed in the hope that it will be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11  * NON INFRINGEMENT. See the GNU General Public License for
12  * more details.
13  */
14 
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/moduleparam.h>
18 #include <linux/sched.h>
19 #include <linux/kernel.h> /* printk() */
20 #include <linux/slab.h> /* kmalloc() */
21 #include <linux/errno.h> /* error codes */
22 #include <linux/types.h> /* size_t */
23 #include <linux/interrupt.h>
24 #include <linux/in.h>
25 #include <linux/irq.h>
26 #include <linux/netdevice.h> /* struct device, and other headers */
27 #include <linux/etherdevice.h> /* eth_type_trans */
28 #include <linux/skbuff.h>
29 #include <linux/ioctl.h>
30 #include <linux/cdev.h>
31 #include <linux/hugetlb.h>
32 #include <linux/in6.h>
33 #include <linux/timer.h>
34 #include <linux/hrtimer.h>
35 #include <linux/ktime.h>
36 #include <linux/io.h>
37 #include <linux/ctype.h>
38 #include <linux/ip.h>
39 #include <linux/tcp.h>
40 
41 #include <asm/checksum.h>
42 #include <asm/homecache.h>
43 #include <gxio/mpipe.h>
44 #include <arch/sim.h>
45 
46 /* Default transmit lockup timeout period, in jiffies. */
47 #define TILE_NET_TIMEOUT (5 * HZ)
48 
49 /* The maximum number of distinct channels (idesc.channel is 5 bits). */
50 #define TILE_NET_CHANNELS 32
51 
52 /* Maximum number of idescs to handle per "poll". */
53 #define TILE_NET_BATCH 128
54 
55 /* Maximum number of packets to handle per "poll". */
56 #define TILE_NET_WEIGHT 64
57 
58 /* Number of entries in each iqueue. */
59 #define IQUEUE_ENTRIES 512
60 
61 /* Number of entries in each equeue. */
62 #define EQUEUE_ENTRIES 2048
63 
64 /* Total header bytes per equeue slot. Must be big enough for 2 bytes
65  * of NET_IP_ALIGN alignment, plus 14 bytes (?) of L2 header, plus up to
66  * 60 bytes of actual TCP header. We round up to align to cache lines.
67  */
68 #define HEADER_BYTES 128
69 
70 /* Maximum completions per cpu per device (must be a power of two).
71  * ISSUE: What is the right number here? If this is too small, then
72  * egress might block waiting for free space in a completions array.
73  * ISSUE: At the least, allocate these only for initialized echannels.
74  */
75 #define TILE_NET_MAX_COMPS 64
76 
77 #define MAX_FRAGS (MAX_SKB_FRAGS + 1)
78 
79 /* Size of completions data to allocate.
80  * ISSUE: Probably more than needed since we don't use all the channels.
81  */
82 #define COMPS_SIZE (TILE_NET_CHANNELS * sizeof(struct tile_net_comps))
83 
84 /* Size of NotifRing data to allocate. */
85 #define NOTIF_RING_SIZE (IQUEUE_ENTRIES * sizeof(gxio_mpipe_idesc_t))
86 
87 /* Timeout to wake the per-device TX timer after we stop the queue.
88  * We don't want the timeout too short (adds overhead, and might end
89  * up causing stop/wake/stop/wake cycles) or too long (affects performance).
90  * For the 10 Gb NIC, 30 usec means roughly 30+ 1500-byte packets.
91  */
92 #define TX_TIMER_DELAY_USEC 30
93 
94 /* Timeout to wake the per-cpu egress timer to free completions. */
95 #define EGRESS_TIMER_DELAY_USEC 1000
96 
97 MODULE_AUTHOR("Tilera Corporation");
98 MODULE_LICENSE("GPL");
99 
100 /* A "packet fragment" (a chunk of memory). */
101 struct frag {
102  void *buf;
103  size_t length;
104 };
105 
106 /* A single completion. */
108  /* The "complete_count" when the completion will be complete. */
110  /* The buffer to be freed when the completion is complete. */
111  struct sk_buff *skb;
112 };
113 
114 /* The completions for a given cpu and echannel. */
116  /* The completions. */
118  /* The number of completions used. */
119  unsigned long comp_next;
120  /* The number of completions freed. */
121  unsigned long comp_last;
122 };
123 
124 /* The transmit wake timer for a given cpu and echannel. */
127  struct hrtimer timer;
128  struct net_device *dev;
129 };
130 
131 /* Info for a specific cpu. */
133  /* The NAPI struct. */
135  /* Packet queue. */
137  /* Our cpu. */
138  int my_cpu;
139  /* True if iqueue is valid. */
141  /* NAPI flags. */
144  /* Number of small sk_buffs which must still be provided. */
146  /* Number of large sk_buffs which must still be provided. */
148  /* A timer for handling egress completions. */
150  /* True if "egress_timer" is scheduled. */
152  /* Comps for each egress channel. */
154  /* Transmit wake timer for each egress channel. */
156 };
157 
158 /* Info for egress on a particular egress channel. */
160  /* The "equeue". */
162  /* The headers for TSO. */
163  unsigned char *headers;
164 };
165 
166 /* Info for a specific device. */
168  /* Our network device. */
169  struct net_device *dev;
170  /* The primary link. */
172  /* The primary channel, if open, else -1. */
173  int channel;
174  /* The "loopify" egress link, if needed. */
176  /* The "loopify" egress channel, if open, else -1. */
178  /* The egress channel (channel or loopify_channel). */
179  int echannel;
180  /* Total stats. */
182 };
183 
184 /* Egress info, indexed by "priv->echannel" (lazily created as needed). */
185 static struct tile_net_egress egress_for_echannel[TILE_NET_CHANNELS];
186 
187 /* Devices currently associated with each channel.
188  * NOTE: The array entry can become NULL after ifconfig down, but
189  * we do not free the underlying net_device structures, so it is
190  * safe to use a pointer after reading it from this array.
191  */
192 static struct net_device *tile_net_devs_for_channel[TILE_NET_CHANNELS];
193 
194 /* A mutex for "tile_net_devs_for_channel". */
195 static DEFINE_MUTEX(tile_net_devs_for_channel_mutex);
196 
197 /* The per-cpu info. */
198 static DEFINE_PER_CPU(struct tile_net_info, per_cpu_info);
199 
200 /* The "context" for all devices. */
202 
203 /* Buffer sizes and mpipe enum codes for buffer stacks.
204  * See arch/tile/include/gxio/mpipe.h for the set of possible values.
205  */
206 #define BUFFER_SIZE_SMALL_ENUM GXIO_MPIPE_BUFFER_SIZE_128
207 #define BUFFER_SIZE_SMALL 128
208 #define BUFFER_SIZE_LARGE_ENUM GXIO_MPIPE_BUFFER_SIZE_1664
209 #define BUFFER_SIZE_LARGE 1664
210 
211 /* The small/large "buffer stacks". */
212 static int small_buffer_stack = -1;
213 static int large_buffer_stack = -1;
214 
215 /* Amount of memory allocated for each buffer stack. */
216 static size_t buffer_stack_size;
217 
218 /* The actual memory allocated for the buffer stacks. */
219 static void *small_buffer_stack_va;
220 static void *large_buffer_stack_va;
221 
222 /* The buckets. */
223 static int first_bucket = -1;
224 static int num_buckets = 1;
225 
226 /* The ingress irq. */
227 static int ingress_irq = -1;
228 
229 /* Text value of tile_net.cpus if passed as a module parameter. */
230 static char *network_cpus_string;
231 
232 /* The actual cpus in "network_cpus". */
233 static struct cpumask network_cpus_map;
234 
235 /* If "loopify=LINK" was specified, this is "LINK". */
236 static char *loopify_link_name;
237 
238 /* If "tile_net.custom" was specified, this is non-NULL. */
239 static char *custom_str;
240 
241 /* The "tile_net.cpus" argument specifies the cpus that are dedicated
242  * to handle ingress packets.
243  *
244  * The parameter should be in the form "tile_net.cpus=m-n[,x-y]", where
245  * m, n, x, y are integer numbers that represent the cpus that can be
246  * neither a dedicated cpu nor a dataplane cpu.
247  */
248 static bool network_cpus_init(void)
249 {
250  char buf[1024];
251  int rc;
252 
253  if (network_cpus_string == NULL)
254  return false;
255 
256  rc = cpulist_parse_crop(network_cpus_string, &network_cpus_map);
257  if (rc != 0) {
258  pr_warn("tile_net.cpus=%s: malformed cpu list\n",
259  network_cpus_string);
260  return false;
261  }
262 
263  /* Remove dedicated cpus. */
264  cpumask_and(&network_cpus_map, &network_cpus_map, cpu_possible_mask);
265 
266  if (cpumask_empty(&network_cpus_map)) {
267  pr_warn("Ignoring empty tile_net.cpus='%s'.\n",
268  network_cpus_string);
269  return false;
270  }
271 
272  cpulist_scnprintf(buf, sizeof(buf), &network_cpus_map);
273  pr_info("Linux network CPUs: %s\n", buf);
274  return true;
275 }
276 
277 module_param_named(cpus, network_cpus_string, charp, 0444);
278 MODULE_PARM_DESC(cpus, "cpulist of cores that handle network interrupts");
279 
280 /* The "tile_net.loopify=LINK" argument causes the named device to
281  * actually use "loop0" for ingress, and "loop1" for egress. This
282  * allows an app to sit between the actual link and linux, passing
283  * (some) packets along to linux, and forwarding (some) packets sent
284  * out by linux.
285  */
286 module_param_named(loopify, loopify_link_name, charp, 0444);
287 MODULE_PARM_DESC(loopify, "name the device to use loop0/1 for ingress/egress");
288 
289 /* The "tile_net.custom" argument causes us to ignore the "conventional"
290  * classifier metadata, in particular, the "l2_offset".
291  */
292 module_param_named(custom, custom_str, charp, 0444);
293 MODULE_PARM_DESC(custom, "indicates a (heavily) customized classifier");
294 
295 /* Atomically update a statistics field.
296  * Note that on TILE-Gx, this operation is fire-and-forget on the
297  * issuing core (single-cycle dispatch) and takes only a few cycles
298  * longer than a regular store when the request reaches the home cache.
299  * No expensive bus management overhead is required.
300  */
301 static void tile_net_stats_add(unsigned long value, unsigned long *field)
302 {
303  BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(unsigned long));
304  atomic_long_add(value, (atomic_long_t *)field);
305 }
306 
307 /* Allocate and push a buffer. */
308 static bool tile_net_provide_buffer(bool small)
309 {
310  int stack = small ? small_buffer_stack : large_buffer_stack;
311  const unsigned long buffer_alignment = 128;
312  struct sk_buff *skb;
313  int len;
314 
315  len = sizeof(struct sk_buff **) + buffer_alignment;
316  len += (small ? BUFFER_SIZE_SMALL : BUFFER_SIZE_LARGE);
317  skb = dev_alloc_skb(len);
318  if (skb == NULL)
319  return false;
320 
321  /* Make room for a back-pointer to 'skb' and guarantee alignment. */
322  skb_reserve(skb, sizeof(struct sk_buff **));
323  skb_reserve(skb, -(long)skb->data & (buffer_alignment - 1));
324 
325  /* Save a back-pointer to 'skb'. */
326  *(struct sk_buff **)(skb->data - sizeof(struct sk_buff **)) = skb;
327 
328  /* Make sure "skb" and the back-pointer have been flushed. */
329  wmb();
330 
331  gxio_mpipe_push_buffer(&context, stack,
332  (void *)va_to_tile_io_addr(skb->data));
333 
334  return true;
335 }
336 
337 /* Convert a raw mpipe buffer to its matching skb pointer. */
338 static struct sk_buff *mpipe_buf_to_skb(void *va)
339 {
340  /* Acquire the associated "skb". */
341  struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
342  struct sk_buff *skb = *skb_ptr;
343 
344  /* Paranoia. */
345  if (skb->data != va) {
346  /* Panic here since there's a reasonable chance
347  * that corrupt buffers means generic memory
348  * corruption, with unpredictable system effects.
349  */
350  panic("Corrupt linux buffer! va=%p, skb=%p, skb->data=%p",
351  va, skb, skb->data);
352  }
353 
354  return skb;
355 }
356 
357 static void tile_net_pop_all_buffers(int stack)
358 {
359  for (;;) {
360  tile_io_addr_t addr =
361  (tile_io_addr_t)gxio_mpipe_pop_buffer(&context, stack);
362  if (addr == 0)
363  break;
364  dev_kfree_skb_irq(mpipe_buf_to_skb(tile_io_addr_to_va(addr)));
365  }
366 }
367 
368 /* Provide linux buffers to mPIPE. */
369 static void tile_net_provide_needed_buffers(void)
370 {
371  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
372 
373  while (info->num_needed_small_buffers != 0) {
374  if (!tile_net_provide_buffer(true))
375  goto oops;
376  info->num_needed_small_buffers--;
377  }
378 
379  while (info->num_needed_large_buffers != 0) {
380  if (!tile_net_provide_buffer(false))
381  goto oops;
382  info->num_needed_large_buffers--;
383  }
384 
385  return;
386 
387 oops:
388  /* Add a description to the page allocation failure dump. */
389  pr_notice("Tile %d still needs some buffers\n", info->my_cpu);
390 }
391 
392 static inline bool filter_packet(struct net_device *dev, void *buf)
393 {
394  /* Filter packets received before we're up. */
395  if (dev == NULL || !(dev->flags & IFF_UP))
396  return true;
397 
398  /* Filter out packets that aren't for us. */
399  if (!(dev->flags & IFF_PROMISC) &&
400  !is_multicast_ether_addr(buf) &&
401  compare_ether_addr(dev->dev_addr, buf) != 0)
402  return true;
403 
404  return false;
405 }
406 
407 static void tile_net_receive_skb(struct net_device *dev, struct sk_buff *skb,
408  gxio_mpipe_idesc_t *idesc, unsigned long len)
409 {
410  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
411  struct tile_net_priv *priv = netdev_priv(dev);
412 
413  /* Encode the actual packet length. */
414  skb_put(skb, len);
415 
416  skb->protocol = eth_type_trans(skb, dev);
417 
418  /* Acknowledge "good" hardware checksums. */
419  if (idesc->cs && idesc->csum_seed_val == 0xFFFF)
421 
422  netif_receive_skb(skb);
423 
424  /* Update stats. */
425  tile_net_stats_add(1, &priv->stats.rx_packets);
426  tile_net_stats_add(len, &priv->stats.rx_bytes);
427 
428  /* Need a new buffer. */
429  if (idesc->size == BUFFER_SIZE_SMALL_ENUM)
430  info->num_needed_small_buffers++;
431  else
432  info->num_needed_large_buffers++;
433 }
434 
435 /* Handle a packet. Return true if "processed", false if "filtered". */
436 static bool tile_net_handle_packet(gxio_mpipe_idesc_t *idesc)
437 {
438  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
439  struct net_device *dev = tile_net_devs_for_channel[idesc->channel];
440  uint8_t l2_offset;
441  void *va;
442  void *buf;
443  unsigned long len;
444  bool filter;
445 
446  /* Drop packets for which no buffer was available.
447  * NOTE: This happens under heavy load.
448  */
449  if (idesc->be) {
450  struct tile_net_priv *priv = netdev_priv(dev);
451  tile_net_stats_add(1, &priv->stats.rx_dropped);
452  gxio_mpipe_iqueue_consume(&info->iqueue, idesc);
453  if (net_ratelimit())
454  pr_info("Dropping packet (insufficient buffers).\n");
455  return false;
456  }
457 
458  /* Get the "l2_offset", if allowed. */
459  l2_offset = custom_str ? 0 : gxio_mpipe_idesc_get_l2_offset(idesc);
460 
461  /* Get the raw buffer VA (includes "headroom"). */
462  va = tile_io_addr_to_va((unsigned long)(long)idesc->va);
463 
464  /* Get the actual packet start/length. */
465  buf = va + l2_offset;
466  len = idesc->l2_size - l2_offset;
467 
468  /* Point "va" at the raw buffer. */
469  va -= NET_IP_ALIGN;
470 
471  filter = filter_packet(dev, buf);
472  if (filter) {
473  gxio_mpipe_iqueue_drop(&info->iqueue, idesc);
474  } else {
475  struct sk_buff *skb = mpipe_buf_to_skb(va);
476 
477  /* Skip headroom, and any custom header. */
478  skb_reserve(skb, NET_IP_ALIGN + l2_offset);
479 
480  tile_net_receive_skb(dev, skb, idesc, len);
481  }
482 
483  gxio_mpipe_iqueue_consume(&info->iqueue, idesc);
484  return !filter;
485 }
486 
487 /* Handle some packets for the current CPU.
488  *
489  * This function handles up to TILE_NET_BATCH idescs per call.
490  *
491  * ISSUE: Since we do not provide new buffers until this function is
492  * complete, we must initially provide enough buffers for each network
493  * cpu to fill its iqueue and also its batched idescs.
494  *
495  * ISSUE: The "rotting packet" race condition occurs if a packet
496  * arrives after the queue appears to be empty, and before the
497  * hypervisor interrupt is re-enabled.
498  */
499 static int tile_net_poll(struct napi_struct *napi, int budget)
500 {
501  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
502  unsigned int work = 0;
503  gxio_mpipe_idesc_t *idesc;
504  int i, n;
505 
506  /* Process packets. */
507  while ((n = gxio_mpipe_iqueue_try_peek(&info->iqueue, &idesc)) > 0) {
508  for (i = 0; i < n; i++) {
509  if (i == TILE_NET_BATCH)
510  goto done;
511  if (tile_net_handle_packet(idesc + i)) {
512  if (++work >= budget)
513  goto done;
514  }
515  }
516  }
517 
518  /* There are no packets left. */
519  napi_complete(&info->napi);
520 
521  /* Re-enable hypervisor interrupts. */
523 
524  /* HACK: Avoid the "rotting packet" problem. */
525  if (gxio_mpipe_iqueue_try_peek(&info->iqueue, &idesc) > 0)
526  napi_schedule(&info->napi);
527 
528  /* ISSUE: Handle completions? */
529 
530 done:
531  tile_net_provide_needed_buffers();
532 
533  return work;
534 }
535 
536 /* Handle an ingress interrupt on the current cpu. */
537 static irqreturn_t tile_net_handle_ingress_irq(int irq, void *unused)
538 {
539  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
540  napi_schedule(&info->napi);
541  return IRQ_HANDLED;
542 }
543 
544 /* Free some completions. This must be called with interrupts blocked. */
545 static int tile_net_free_comps(gxio_mpipe_equeue_t *equeue,
546  struct tile_net_comps *comps,
547  int limit, bool force_update)
548 {
549  int n = 0;
550  while (comps->comp_last < comps->comp_next) {
551  unsigned int cid = comps->comp_last % TILE_NET_MAX_COMPS;
552  struct tile_net_comp *comp = &comps->comp_queue[cid];
553  if (!gxio_mpipe_equeue_is_complete(equeue, comp->when,
554  force_update || n == 0))
555  break;
556  dev_kfree_skb_irq(comp->skb);
557  comps->comp_last++;
558  if (++n == limit)
559  break;
560  }
561  return n;
562 }
563 
564 /* Add a completion. This must be called with interrupts blocked.
565  * tile_net_equeue_try_reserve() will have ensured a free completion entry.
566  */
567 static void add_comp(gxio_mpipe_equeue_t *equeue,
568  struct tile_net_comps *comps,
569  uint64_t when, struct sk_buff *skb)
570 {
571  int cid = comps->comp_next % TILE_NET_MAX_COMPS;
572  comps->comp_queue[cid].when = when;
573  comps->comp_queue[cid].skb = skb;
574  comps->comp_next++;
575 }
576 
577 static void tile_net_schedule_tx_wake_timer(struct net_device *dev,
578  int tx_queue_idx)
579 {
580  struct tile_net_info *info = &per_cpu(per_cpu_info, tx_queue_idx);
581  struct tile_net_priv *priv = netdev_priv(dev);
582  struct tile_net_tx_wake *tx_wake = &info->tx_wake[priv->echannel];
583 
584  hrtimer_start(&tx_wake->timer,
585  ktime_set(0, TX_TIMER_DELAY_USEC * 1000UL),
587 }
588 
589 static enum hrtimer_restart tile_net_handle_tx_wake_timer(struct hrtimer *t)
590 {
591  struct tile_net_tx_wake *tx_wake =
592  container_of(t, struct tile_net_tx_wake, timer);
593  netif_wake_subqueue(tx_wake->dev, tx_wake->tx_queue_idx);
594  return HRTIMER_NORESTART;
595 }
596 
597 /* Make sure the egress timer is scheduled. */
598 static void tile_net_schedule_egress_timer(void)
599 {
600  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
601 
602  if (!info->egress_timer_scheduled) {
604  ktime_set(0, EGRESS_TIMER_DELAY_USEC * 1000UL),
606  info->egress_timer_scheduled = true;
607  }
608 }
609 
610 /* The "function" for "info->egress_timer".
611  *
612  * This timer will reschedule itself as long as there are any pending
613  * completions expected for this tile.
614  */
615 static enum hrtimer_restart tile_net_handle_egress_timer(struct hrtimer *t)
616 {
617  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
618  unsigned long irqflags;
619  bool pending = false;
620  int i;
621 
622  local_irq_save(irqflags);
623 
624  /* The timer is no longer scheduled. */
625  info->egress_timer_scheduled = false;
626 
627  /* Free all possible comps for this tile. */
628  for (i = 0; i < TILE_NET_CHANNELS; i++) {
629  struct tile_net_egress *egress = &egress_for_echannel[i];
630  struct tile_net_comps *comps = info->comps_for_echannel[i];
631  if (comps->comp_last >= comps->comp_next)
632  continue;
633  tile_net_free_comps(egress->equeue, comps, -1, true);
634  pending = pending || (comps->comp_last < comps->comp_next);
635  }
636 
637  /* Reschedule timer if needed. */
638  if (pending)
639  tile_net_schedule_egress_timer();
640 
641  local_irq_restore(irqflags);
642 
643  return HRTIMER_NORESTART;
644 }
645 
646 /* Helper function for "tile_net_update()".
647  * "dev" (i.e. arg) is the device being brought up or down,
648  * or NULL if all devices are now down.
649  */
650 static void tile_net_update_cpu(void *arg)
651 {
652  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
653  struct net_device *dev = arg;
654 
655  if (!info->has_iqueue)
656  return;
657 
658  if (dev != NULL) {
659  if (!info->napi_added) {
660  netif_napi_add(dev, &info->napi,
661  tile_net_poll, TILE_NET_WEIGHT);
662  info->napi_added = true;
663  }
664  if (!info->napi_enabled) {
665  napi_enable(&info->napi);
666  info->napi_enabled = true;
667  }
668  enable_percpu_irq(ingress_irq, 0);
669  } else {
670  disable_percpu_irq(ingress_irq);
671  if (info->napi_enabled) {
672  napi_disable(&info->napi);
673  info->napi_enabled = false;
674  }
675  /* FIXME: Drain the iqueue. */
676  }
677 }
678 
679 /* Helper function for tile_net_open() and tile_net_stop().
680  * Always called under tile_net_devs_for_channel_mutex.
681  */
682 static int tile_net_update(struct net_device *dev)
683 {
684  static gxio_mpipe_rules_t rules; /* too big to fit on the stack */
685  bool saw_channel = false;
686  int channel;
687  int rc;
688  int cpu;
689 
690  gxio_mpipe_rules_init(&rules, &context);
691 
692  for (channel = 0; channel < TILE_NET_CHANNELS; channel++) {
693  if (tile_net_devs_for_channel[channel] == NULL)
694  continue;
695  if (!saw_channel) {
696  saw_channel = true;
697  gxio_mpipe_rules_begin(&rules, first_bucket,
698  num_buckets, NULL);
700  }
701  gxio_mpipe_rules_add_channel(&rules, channel);
702  }
703 
704  /* NOTE: This can fail if there is no classifier.
705  * ISSUE: Can anything else cause it to fail?
706  */
707  rc = gxio_mpipe_rules_commit(&rules);
708  if (rc != 0) {
709  netdev_warn(dev, "gxio_mpipe_rules_commit failed: %d\n", rc);
710  return -EIO;
711  }
712 
713  /* Update all cpus, sequentially (to protect "netif_napi_add()"). */
715  smp_call_function_single(cpu, tile_net_update_cpu,
716  (saw_channel ? dev : NULL), 1);
717 
718  /* HACK: Allow packets to flow in the simulator. */
719  if (saw_channel)
720  sim_enable_mpipe_links(0, -1);
721 
722  return 0;
723 }
724 
725 /* Allocate and initialize mpipe buffer stacks, and register them in
726  * the mPIPE TLBs, for both small and large packet sizes.
727  * This routine supports tile_net_init_mpipe(), below.
728  */
729 static int init_buffer_stacks(struct net_device *dev, int num_buffers)
730 {
731  pte_t hash_pte = pte_set_home((pte_t) { 0 }, PAGE_HOME_HASH);
732  int rc;
733 
734  /* Compute stack bytes; we round up to 64KB and then use
735  * alloc_pages() so we get the required 64KB alignment as well.
736  */
737  buffer_stack_size =
739  64 * 1024);
740 
741  /* Allocate two buffer stack indices. */
742  rc = gxio_mpipe_alloc_buffer_stacks(&context, 2, 0, 0);
743  if (rc < 0) {
744  netdev_err(dev, "gxio_mpipe_alloc_buffer_stacks failed: %d\n",
745  rc);
746  return rc;
747  }
748  small_buffer_stack = rc;
749  large_buffer_stack = rc + 1;
750 
751  /* Allocate the small memory stack. */
752  small_buffer_stack_va =
753  alloc_pages_exact(buffer_stack_size, GFP_KERNEL);
754  if (small_buffer_stack_va == NULL) {
755  netdev_err(dev,
756  "Could not alloc %zd bytes for buffer stacks\n",
757  buffer_stack_size);
758  return -ENOMEM;
759  }
760  rc = gxio_mpipe_init_buffer_stack(&context, small_buffer_stack,
762  small_buffer_stack_va,
763  buffer_stack_size, 0);
764  if (rc != 0) {
765  netdev_err(dev, "gxio_mpipe_init_buffer_stack: %d\n", rc);
766  return rc;
767  }
768  rc = gxio_mpipe_register_client_memory(&context, small_buffer_stack,
769  hash_pte, 0);
770  if (rc != 0) {
771  netdev_err(dev,
772  "gxio_mpipe_register_buffer_memory failed: %d\n",
773  rc);
774  return rc;
775  }
776 
777  /* Allocate the large buffer stack. */
778  large_buffer_stack_va =
779  alloc_pages_exact(buffer_stack_size, GFP_KERNEL);
780  if (large_buffer_stack_va == NULL) {
781  netdev_err(dev,
782  "Could not alloc %zd bytes for buffer stacks\n",
783  buffer_stack_size);
784  return -ENOMEM;
785  }
786  rc = gxio_mpipe_init_buffer_stack(&context, large_buffer_stack,
788  large_buffer_stack_va,
789  buffer_stack_size, 0);
790  if (rc != 0) {
791  netdev_err(dev, "gxio_mpipe_init_buffer_stack failed: %d\n",
792  rc);
793  return rc;
794  }
795  rc = gxio_mpipe_register_client_memory(&context, large_buffer_stack,
796  hash_pte, 0);
797  if (rc != 0) {
798  netdev_err(dev,
799  "gxio_mpipe_register_buffer_memory failed: %d\n",
800  rc);
801  return rc;
802  }
803 
804  return 0;
805 }
806 
807 /* Allocate per-cpu resources (memory for completions and idescs).
808  * This routine supports tile_net_init_mpipe(), below.
809  */
810 static int alloc_percpu_mpipe_resources(struct net_device *dev,
811  int cpu, int ring)
812 {
813  struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
814  int order, i, rc;
815  struct page *page;
816  void *addr;
817 
818  /* Allocate the "comps". */
819  order = get_order(COMPS_SIZE);
820  page = homecache_alloc_pages(GFP_KERNEL, order, cpu);
821  if (page == NULL) {
822  netdev_err(dev, "Failed to alloc %zd bytes comps memory\n",
823  COMPS_SIZE);
824  return -ENOMEM;
825  }
826  addr = pfn_to_kaddr(page_to_pfn(page));
827  memset(addr, 0, COMPS_SIZE);
828  for (i = 0; i < TILE_NET_CHANNELS; i++)
829  info->comps_for_echannel[i] =
830  addr + i * sizeof(struct tile_net_comps);
831 
832  /* If this is a network cpu, create an iqueue. */
833  if (cpu_isset(cpu, network_cpus_map)) {
834  order = get_order(NOTIF_RING_SIZE);
835  page = homecache_alloc_pages(GFP_KERNEL, order, cpu);
836  if (page == NULL) {
837  netdev_err(dev,
838  "Failed to alloc %zd bytes iqueue memory\n",
840  return -ENOMEM;
841  }
842  addr = pfn_to_kaddr(page_to_pfn(page));
843  rc = gxio_mpipe_iqueue_init(&info->iqueue, &context, ring++,
844  addr, NOTIF_RING_SIZE, 0);
845  if (rc < 0) {
846  netdev_err(dev,
847  "gxio_mpipe_iqueue_init failed: %d\n", rc);
848  return rc;
849  }
850  info->has_iqueue = true;
851  }
852 
853  return ring;
854 }
855 
856 /* Initialize NotifGroup and buckets.
857  * This routine supports tile_net_init_mpipe(), below.
858  */
859 static int init_notif_group_and_buckets(struct net_device *dev,
860  int ring, int network_cpus_count)
861 {
862  int group, rc;
863 
864  /* Allocate one NotifGroup. */
865  rc = gxio_mpipe_alloc_notif_groups(&context, 1, 0, 0);
866  if (rc < 0) {
867  netdev_err(dev, "gxio_mpipe_alloc_notif_groups failed: %d\n",
868  rc);
869  return rc;
870  }
871  group = rc;
872 
873  /* Initialize global num_buckets value. */
874  if (network_cpus_count > 4)
875  num_buckets = 256;
876  else if (network_cpus_count > 1)
877  num_buckets = 16;
878 
879  /* Allocate some buckets, and set global first_bucket value. */
880  rc = gxio_mpipe_alloc_buckets(&context, num_buckets, 0, 0);
881  if (rc < 0) {
882  netdev_err(dev, "gxio_mpipe_alloc_buckets failed: %d\n", rc);
883  return rc;
884  }
885  first_bucket = rc;
886 
887  /* Init group and buckets. */
889  &context, group, ring, network_cpus_count,
890  first_bucket, num_buckets,
892  if (rc != 0) {
893  netdev_err(
894  dev,
895  "gxio_mpipe_init_notif_group_and_buckets failed: %d\n",
896  rc);
897  return rc;
898  }
899 
900  return 0;
901 }
902 
903 /* Create an irq and register it, then activate the irq and request
904  * interrupts on all cores. Note that "ingress_irq" being initialized
905  * is how we know not to call tile_net_init_mpipe() again.
906  * This routine supports tile_net_init_mpipe(), below.
907  */
908 static int tile_net_setup_interrupts(struct net_device *dev)
909 {
910  int cpu, rc;
911 
912  rc = create_irq();
913  if (rc < 0) {
914  netdev_err(dev, "create_irq failed: %d\n", rc);
915  return rc;
916  }
917  ingress_irq = rc;
918  tile_irq_activate(ingress_irq, TILE_IRQ_PERCPU);
919  rc = request_irq(ingress_irq, tile_net_handle_ingress_irq,
920  0, "tile_net", NULL);
921  if (rc != 0) {
922  netdev_err(dev, "request_irq failed: %d\n", rc);
923  destroy_irq(ingress_irq);
924  ingress_irq = -1;
925  return rc;
926  }
927 
928  for_each_online_cpu(cpu) {
929  struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
930  if (info->has_iqueue) {
932  &context, cpu_x(cpu), cpu_y(cpu),
933  1, ingress_irq, info->iqueue.ring);
934  }
935  }
936 
937  return 0;
938 }
939 
940 /* Undo any state set up partially by a failed call to tile_net_init_mpipe. */
941 static void tile_net_init_mpipe_fail(void)
942 {
943  int cpu;
944 
945  /* Do cleanups that require the mpipe context first. */
946  if (small_buffer_stack >= 0)
947  tile_net_pop_all_buffers(small_buffer_stack);
948  if (large_buffer_stack >= 0)
949  tile_net_pop_all_buffers(large_buffer_stack);
950 
951  /* Destroy mpipe context so the hardware no longer owns any memory. */
952  gxio_mpipe_destroy(&context);
953 
954  for_each_online_cpu(cpu) {
955  struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
956  free_pages((unsigned long)(info->comps_for_echannel[0]),
958  info->comps_for_echannel[0] = NULL;
959  free_pages((unsigned long)(info->iqueue.idescs),
961  info->iqueue.idescs = NULL;
962  }
963 
964  if (small_buffer_stack_va)
965  free_pages_exact(small_buffer_stack_va, buffer_stack_size);
966  if (large_buffer_stack_va)
967  free_pages_exact(large_buffer_stack_va, buffer_stack_size);
968 
969  small_buffer_stack_va = NULL;
970  large_buffer_stack_va = NULL;
971  large_buffer_stack = -1;
972  small_buffer_stack = -1;
973  first_bucket = -1;
974 }
975 
976 /* The first time any tilegx network device is opened, we initialize
977  * the global mpipe state. If this step fails, we fail to open the
978  * device, but if it succeeds, we never need to do it again, and since
979  * tile_net can't be unloaded, we never undo it.
980  *
981  * Note that some resources in this path (buffer stack indices,
982  * bindings from init_buffer_stack, etc.) are hypervisor resources
983  * that are freed implicitly by gxio_mpipe_destroy().
984  */
985 static int tile_net_init_mpipe(struct net_device *dev)
986 {
987  int i, num_buffers, rc;
988  int cpu;
989  int first_ring, ring;
990  int network_cpus_count = cpus_weight(network_cpus_map);
991 
992  if (!hash_default) {
993  netdev_err(dev, "Networking requires hash_default!\n");
994  return -EIO;
995  }
996 
997  rc = gxio_mpipe_init(&context, 0);
998  if (rc != 0) {
999  netdev_err(dev, "gxio_mpipe_init failed: %d\n", rc);
1000  return -EIO;
1001  }
1002 
1003  /* Set up the buffer stacks. */
1004  num_buffers =
1005  network_cpus_count * (IQUEUE_ENTRIES + TILE_NET_BATCH);
1006  rc = init_buffer_stacks(dev, num_buffers);
1007  if (rc != 0)
1008  goto fail;
1009 
1010  /* Provide initial buffers. */
1011  rc = -ENOMEM;
1012  for (i = 0; i < num_buffers; i++) {
1013  if (!tile_net_provide_buffer(true)) {
1014  netdev_err(dev, "Cannot allocate initial sk_bufs!\n");
1015  goto fail;
1016  }
1017  }
1018  for (i = 0; i < num_buffers; i++) {
1019  if (!tile_net_provide_buffer(false)) {
1020  netdev_err(dev, "Cannot allocate initial sk_bufs!\n");
1021  goto fail;
1022  }
1023  }
1024 
1025  /* Allocate one NotifRing for each network cpu. */
1026  rc = gxio_mpipe_alloc_notif_rings(&context, network_cpus_count, 0, 0);
1027  if (rc < 0) {
1028  netdev_err(dev, "gxio_mpipe_alloc_notif_rings failed %d\n",
1029  rc);
1030  goto fail;
1031  }
1032 
1033  /* Init NotifRings per-cpu. */
1034  first_ring = rc;
1035  ring = first_ring;
1036  for_each_online_cpu(cpu) {
1037  rc = alloc_percpu_mpipe_resources(dev, cpu, ring);
1038  if (rc < 0)
1039  goto fail;
1040  ring = rc;
1041  }
1042 
1043  /* Initialize NotifGroup and buckets. */
1044  rc = init_notif_group_and_buckets(dev, first_ring, network_cpus_count);
1045  if (rc != 0)
1046  goto fail;
1047 
1048  /* Create and enable interrupts. */
1049  rc = tile_net_setup_interrupts(dev);
1050  if (rc != 0)
1051  goto fail;
1052 
1053  return 0;
1054 
1055 fail:
1056  tile_net_init_mpipe_fail();
1057  return rc;
1058 }
1059 
1060 /* Create persistent egress info for a given egress channel.
1061  * Note that this may be shared between, say, "gbe0" and "xgbe0".
1062  * ISSUE: Defer header allocation until TSO is actually needed?
1063  */
1064 static int tile_net_init_egress(struct net_device *dev, int echannel)
1065 {
1066  struct page *headers_page, *edescs_page, *equeue_page;
1067  gxio_mpipe_edesc_t *edescs;
1068  gxio_mpipe_equeue_t *equeue;
1069  unsigned char *headers;
1070  int headers_order, edescs_order, equeue_order;
1071  size_t edescs_size;
1072  int edma;
1073  int rc = -ENOMEM;
1074 
1075  /* Only initialize once. */
1076  if (egress_for_echannel[echannel].equeue != NULL)
1077  return 0;
1078 
1079  /* Allocate memory for the "headers". */
1080  headers_order = get_order(EQUEUE_ENTRIES * HEADER_BYTES);
1081  headers_page = alloc_pages(GFP_KERNEL, headers_order);
1082  if (headers_page == NULL) {
1083  netdev_warn(dev,
1084  "Could not alloc %zd bytes for TSO headers.\n",
1085  PAGE_SIZE << headers_order);
1086  goto fail;
1087  }
1088  headers = pfn_to_kaddr(page_to_pfn(headers_page));
1089 
1090  /* Allocate memory for the "edescs". */
1091  edescs_size = EQUEUE_ENTRIES * sizeof(*edescs);
1092  edescs_order = get_order(edescs_size);
1093  edescs_page = alloc_pages(GFP_KERNEL, edescs_order);
1094  if (edescs_page == NULL) {
1095  netdev_warn(dev,
1096  "Could not alloc %zd bytes for eDMA ring.\n",
1097  edescs_size);
1098  goto fail_headers;
1099  }
1100  edescs = pfn_to_kaddr(page_to_pfn(edescs_page));
1101 
1102  /* Allocate memory for the "equeue". */
1103  equeue_order = get_order(sizeof(*equeue));
1104  equeue_page = alloc_pages(GFP_KERNEL, equeue_order);
1105  if (equeue_page == NULL) {
1106  netdev_warn(dev,
1107  "Could not alloc %zd bytes for equeue info.\n",
1108  PAGE_SIZE << equeue_order);
1109  goto fail_edescs;
1110  }
1111  equeue = pfn_to_kaddr(page_to_pfn(equeue_page));
1112 
1113  /* Allocate an edma ring. Note that in practice this can't
1114  * fail, which is good, because we will leak an edma ring if so.
1115  */
1116  rc = gxio_mpipe_alloc_edma_rings(&context, 1, 0, 0);
1117  if (rc < 0) {
1118  netdev_warn(dev, "gxio_mpipe_alloc_edma_rings failed: %d\n",
1119  rc);
1120  goto fail_equeue;
1121  }
1122  edma = rc;
1123 
1124  /* Initialize the equeue. */
1125  rc = gxio_mpipe_equeue_init(equeue, &context, edma, echannel,
1126  edescs, edescs_size, 0);
1127  if (rc != 0) {
1128  netdev_err(dev, "gxio_mpipe_equeue_init failed: %d\n", rc);
1129  goto fail_equeue;
1130  }
1131 
1132  /* Done. */
1133  egress_for_echannel[echannel].equeue = equeue;
1134  egress_for_echannel[echannel].headers = headers;
1135  return 0;
1136 
1137 fail_equeue:
1138  __free_pages(equeue_page, equeue_order);
1139 
1140 fail_edescs:
1141  __free_pages(edescs_page, edescs_order);
1142 
1143 fail_headers:
1144  __free_pages(headers_page, headers_order);
1145 
1146 fail:
1147  return rc;
1148 }
1149 
1150 /* Return channel number for a newly-opened link. */
1151 static int tile_net_link_open(struct net_device *dev, gxio_mpipe_link_t *link,
1152  const char *link_name)
1153 {
1154  int rc = gxio_mpipe_link_open(link, &context, link_name, 0);
1155  if (rc < 0) {
1156  netdev_err(dev, "Failed to open '%s'\n", link_name);
1157  return rc;
1158  }
1159  rc = gxio_mpipe_link_channel(link);
1160  if (rc < 0 || rc >= TILE_NET_CHANNELS) {
1161  netdev_err(dev, "gxio_mpipe_link_channel bad value: %d\n", rc);
1162  gxio_mpipe_link_close(link);
1163  return -EINVAL;
1164  }
1165  return rc;
1166 }
1167 
1168 /* Help the kernel activate the given network interface. */
1169 static int tile_net_open(struct net_device *dev)
1170 {
1171  struct tile_net_priv *priv = netdev_priv(dev);
1172  int cpu, rc;
1173 
1174  mutex_lock(&tile_net_devs_for_channel_mutex);
1175 
1176  /* Do one-time initialization the first time any device is opened. */
1177  if (ingress_irq < 0) {
1178  rc = tile_net_init_mpipe(dev);
1179  if (rc != 0)
1180  goto fail;
1181  }
1182 
1183  /* Determine if this is the "loopify" device. */
1184  if (unlikely((loopify_link_name != NULL) &&
1185  !strcmp(dev->name, loopify_link_name))) {
1186  rc = tile_net_link_open(dev, &priv->link, "loop0");
1187  if (rc < 0)
1188  goto fail;
1189  priv->channel = rc;
1190  rc = tile_net_link_open(dev, &priv->loopify_link, "loop1");
1191  if (rc < 0)
1192  goto fail;
1193  priv->loopify_channel = rc;
1194  priv->echannel = rc;
1195  } else {
1196  rc = tile_net_link_open(dev, &priv->link, dev->name);
1197  if (rc < 0)
1198  goto fail;
1199  priv->channel = rc;
1200  priv->echannel = rc;
1201  }
1202 
1203  /* Initialize egress info (if needed). Once ever, per echannel. */
1204  rc = tile_net_init_egress(dev, priv->echannel);
1205  if (rc != 0)
1206  goto fail;
1207 
1208  tile_net_devs_for_channel[priv->channel] = dev;
1209 
1210  rc = tile_net_update(dev);
1211  if (rc != 0)
1212  goto fail;
1213 
1214  mutex_unlock(&tile_net_devs_for_channel_mutex);
1215 
1216  /* Initialize the transmit wake timer for this device for each cpu. */
1217  for_each_online_cpu(cpu) {
1218  struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
1219  struct tile_net_tx_wake *tx_wake =
1220  &info->tx_wake[priv->echannel];
1221 
1222  hrtimer_init(&tx_wake->timer, CLOCK_MONOTONIC,
1224  tx_wake->tx_queue_idx = cpu;
1225  tx_wake->timer.function = tile_net_handle_tx_wake_timer;
1226  tx_wake->dev = dev;
1227  }
1228 
1229  for_each_online_cpu(cpu)
1230  netif_start_subqueue(dev, cpu);
1231  netif_carrier_on(dev);
1232  return 0;
1233 
1234 fail:
1235  if (priv->loopify_channel >= 0) {
1236  if (gxio_mpipe_link_close(&priv->loopify_link) != 0)
1237  netdev_warn(dev, "Failed to close loopify link!\n");
1238  priv->loopify_channel = -1;
1239  }
1240  if (priv->channel >= 0) {
1241  if (gxio_mpipe_link_close(&priv->link) != 0)
1242  netdev_warn(dev, "Failed to close link!\n");
1243  priv->channel = -1;
1244  }
1245  priv->echannel = -1;
1246  tile_net_devs_for_channel[priv->channel] = NULL;
1247  mutex_unlock(&tile_net_devs_for_channel_mutex);
1248 
1249  /* Don't return raw gxio error codes to generic Linux. */
1250  return (rc > -512) ? rc : -EIO;
1251 }
1252 
1253 /* Help the kernel deactivate the given network interface. */
1254 static int tile_net_stop(struct net_device *dev)
1255 {
1256  struct tile_net_priv *priv = netdev_priv(dev);
1257  int cpu;
1258 
1259  for_each_online_cpu(cpu) {
1260  struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
1261  struct tile_net_tx_wake *tx_wake =
1262  &info->tx_wake[priv->echannel];
1263 
1264  hrtimer_cancel(&tx_wake->timer);
1265  netif_stop_subqueue(dev, cpu);
1266  }
1267 
1268  mutex_lock(&tile_net_devs_for_channel_mutex);
1269  tile_net_devs_for_channel[priv->channel] = NULL;
1270  (void)tile_net_update(dev);
1271  if (priv->loopify_channel >= 0) {
1272  if (gxio_mpipe_link_close(&priv->loopify_link) != 0)
1273  netdev_warn(dev, "Failed to close loopify link!\n");
1274  priv->loopify_channel = -1;
1275  }
1276  if (priv->channel >= 0) {
1277  if (gxio_mpipe_link_close(&priv->link) != 0)
1278  netdev_warn(dev, "Failed to close link!\n");
1279  priv->channel = -1;
1280  }
1281  priv->echannel = -1;
1282  mutex_unlock(&tile_net_devs_for_channel_mutex);
1283 
1284  return 0;
1285 }
1286 
1287 /* Determine the VA for a fragment. */
1288 static inline void *tile_net_frag_buf(skb_frag_t *f)
1289 {
1290  unsigned long pfn = page_to_pfn(skb_frag_page(f));
1291  return pfn_to_kaddr(pfn) + f->page_offset;
1292 }
1293 
1294 /* Acquire a completion entry and an egress slot, or if we can't,
1295  * stop the queue and schedule the tx_wake timer.
1296  */
1297 static s64 tile_net_equeue_try_reserve(struct net_device *dev,
1298  int tx_queue_idx,
1299  struct tile_net_comps *comps,
1300  gxio_mpipe_equeue_t *equeue,
1301  int num_edescs)
1302 {
1303  /* Try to acquire a completion entry. */
1304  if (comps->comp_next - comps->comp_last < TILE_NET_MAX_COMPS - 1 ||
1305  tile_net_free_comps(equeue, comps, 32, false) != 0) {
1306 
1307  /* Try to acquire an egress slot. */
1308  s64 slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs);
1309  if (slot >= 0)
1310  return slot;
1311 
1312  /* Freeing some completions gives the equeue time to drain. */
1313  tile_net_free_comps(equeue, comps, TILE_NET_MAX_COMPS, false);
1314 
1315  slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs);
1316  if (slot >= 0)
1317  return slot;
1318  }
1319 
1320  /* Still nothing; give up and stop the queue for a short while. */
1321  netif_stop_subqueue(dev, tx_queue_idx);
1322  tile_net_schedule_tx_wake_timer(dev, tx_queue_idx);
1323  return -1;
1324 }
1325 
1326 /* Determine how many edesc's are needed for TSO.
1327  *
1328  * Sometimes, if "sendfile()" requires copying, we will be called with
1329  * "data" containing the header and payload, with "frags" being empty.
1330  * Sometimes, for example when using NFS over TCP, a single segment can
1331  * span 3 fragments. This requires special care.
1332  */
1333 static int tso_count_edescs(struct sk_buff *skb)
1334 {
1335  struct skb_shared_info *sh = skb_shinfo(skb);
1336  unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1337  unsigned int data_len = skb->len - sh_len;
1338  unsigned int p_len = sh->gso_size;
1339  long f_id = -1; /* id of the current fragment */
1340  long f_size = skb_headlen(skb) - sh_len; /* current fragment size */
1341  long f_used = 0; /* bytes used from the current fragment */
1342  long n; /* size of the current piece of payload */
1343  int num_edescs = 0;
1344  int segment;
1345 
1346  for (segment = 0; segment < sh->gso_segs; segment++) {
1347 
1348  unsigned int p_used = 0;
1349 
1350  /* One edesc for header and for each piece of the payload. */
1351  for (num_edescs++; p_used < p_len; num_edescs++) {
1352 
1353  /* Advance as needed. */
1354  while (f_used >= f_size) {
1355  f_id++;
1356  f_size = skb_frag_size(&sh->frags[f_id]);
1357  f_used = 0;
1358  }
1359 
1360  /* Use bytes from the current fragment. */
1361  n = p_len - p_used;
1362  if (n > f_size - f_used)
1363  n = f_size - f_used;
1364  f_used += n;
1365  p_used += n;
1366  }
1367 
1368  /* The last segment may be less than gso_size. */
1369  data_len -= p_len;
1370  if (data_len < p_len)
1371  p_len = data_len;
1372  }
1373 
1374  return num_edescs;
1375 }
1376 
1377 /* Prepare modified copies of the skbuff headers.
1378  * FIXME: add support for IPv6.
1379  */
1380 static void tso_headers_prepare(struct sk_buff *skb, unsigned char *headers,
1381  s64 slot)
1382 {
1383  struct skb_shared_info *sh = skb_shinfo(skb);
1384  struct iphdr *ih;
1385  struct tcphdr *th;
1386  unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1387  unsigned int data_len = skb->len - sh_len;
1388  unsigned char *data = skb->data;
1389  unsigned int ih_off, th_off, p_len;
1390  unsigned int isum_seed, tsum_seed, id, seq;
1391  long f_id = -1; /* id of the current fragment */
1392  long f_size = skb_headlen(skb) - sh_len; /* current fragment size */
1393  long f_used = 0; /* bytes used from the current fragment */
1394  long n; /* size of the current piece of payload */
1395  int segment;
1396 
1397  /* Locate original headers and compute various lengths. */
1398  ih = ip_hdr(skb);
1399  th = tcp_hdr(skb);
1400  ih_off = skb_network_offset(skb);
1401  th_off = skb_transport_offset(skb);
1402  p_len = sh->gso_size;
1403 
1404  /* Set up seed values for IP and TCP csum and initialize id and seq. */
1405  isum_seed = ((0xFFFF - ih->check) +
1406  (0xFFFF - ih->tot_len) +
1407  (0xFFFF - ih->id));
1408  tsum_seed = th->check + (0xFFFF ^ htons(skb->len));
1409  id = ntohs(ih->id);
1410  seq = ntohl(th->seq);
1411 
1412  /* Prepare all the headers. */
1413  for (segment = 0; segment < sh->gso_segs; segment++) {
1414  unsigned char *buf;
1415  unsigned int p_used = 0;
1416 
1417  /* Copy to the header memory for this segment. */
1418  buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES +
1419  NET_IP_ALIGN;
1420  memcpy(buf, data, sh_len);
1421 
1422  /* Update copied ip header. */
1423  ih = (struct iphdr *)(buf + ih_off);
1424  ih->tot_len = htons(sh_len + p_len - ih_off);
1425  ih->id = htons(id);
1426  ih->check = csum_long(isum_seed + ih->tot_len +
1427  ih->id) ^ 0xffff;
1428 
1429  /* Update copied tcp header. */
1430  th = (struct tcphdr *)(buf + th_off);
1431  th->seq = htonl(seq);
1432  th->check = csum_long(tsum_seed + htons(sh_len + p_len));
1433  if (segment != sh->gso_segs - 1) {
1434  th->fin = 0;
1435  th->psh = 0;
1436  }
1437 
1438  /* Skip past the header. */
1439  slot++;
1440 
1441  /* Skip past the payload. */
1442  while (p_used < p_len) {
1443 
1444  /* Advance as needed. */
1445  while (f_used >= f_size) {
1446  f_id++;
1447  f_size = skb_frag_size(&sh->frags[f_id]);
1448  f_used = 0;
1449  }
1450 
1451  /* Use bytes from the current fragment. */
1452  n = p_len - p_used;
1453  if (n > f_size - f_used)
1454  n = f_size - f_used;
1455  f_used += n;
1456  p_used += n;
1457 
1458  slot++;
1459  }
1460 
1461  id++;
1462  seq += p_len;
1463 
1464  /* The last segment may be less than gso_size. */
1465  data_len -= p_len;
1466  if (data_len < p_len)
1467  p_len = data_len;
1468  }
1469 
1470  /* Flush the headers so they are ready for hardware DMA. */
1471  wmb();
1472 }
1473 
1474 /* Pass all the data to mpipe for egress. */
1475 static void tso_egress(struct net_device *dev, gxio_mpipe_equeue_t *equeue,
1476  struct sk_buff *skb, unsigned char *headers, s64 slot)
1477 {
1478  struct tile_net_priv *priv = netdev_priv(dev);
1479  struct skb_shared_info *sh = skb_shinfo(skb);
1480  unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1481  unsigned int data_len = skb->len - sh_len;
1482  unsigned int p_len = sh->gso_size;
1483  gxio_mpipe_edesc_t edesc_head = { { 0 } };
1484  gxio_mpipe_edesc_t edesc_body = { { 0 } };
1485  long f_id = -1; /* id of the current fragment */
1486  long f_size = skb_headlen(skb) - sh_len; /* current fragment size */
1487  long f_used = 0; /* bytes used from the current fragment */
1488  void *f_data = skb->data + sh_len;
1489  long n; /* size of the current piece of payload */
1490  unsigned long tx_packets = 0, tx_bytes = 0;
1491  unsigned int csum_start;
1492  int segment;
1493 
1494  /* Prepare to egress the headers: set up header edesc. */
1495  csum_start = skb_checksum_start_offset(skb);
1496  edesc_head.csum = 1;
1497  edesc_head.csum_start = csum_start;
1498  edesc_head.csum_dest = csum_start + skb->csum_offset;
1499  edesc_head.xfer_size = sh_len;
1500 
1501  /* This is only used to specify the TLB. */
1502  edesc_head.stack_idx = large_buffer_stack;
1503  edesc_body.stack_idx = large_buffer_stack;
1504 
1505  /* Egress all the edescs. */
1506  for (segment = 0; segment < sh->gso_segs; segment++) {
1507  unsigned char *buf;
1508  unsigned int p_used = 0;
1509 
1510  /* Egress the header. */
1511  buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES +
1512  NET_IP_ALIGN;
1513  edesc_head.va = va_to_tile_io_addr(buf);
1514  gxio_mpipe_equeue_put_at(equeue, edesc_head, slot);
1515  slot++;
1516 
1517  /* Egress the payload. */
1518  while (p_used < p_len) {
1519  void *va;
1520 
1521  /* Advance as needed. */
1522  while (f_used >= f_size) {
1523  f_id++;
1524  f_size = skb_frag_size(&sh->frags[f_id]);
1525  f_data = tile_net_frag_buf(&sh->frags[f_id]);
1526  f_used = 0;
1527  }
1528 
1529  va = f_data + f_used;
1530 
1531  /* Use bytes from the current fragment. */
1532  n = p_len - p_used;
1533  if (n > f_size - f_used)
1534  n = f_size - f_used;
1535  f_used += n;
1536  p_used += n;
1537 
1538  /* Egress a piece of the payload. */
1539  edesc_body.va = va_to_tile_io_addr(va);
1540  edesc_body.xfer_size = n;
1541  edesc_body.bound = !(p_used < p_len);
1542  gxio_mpipe_equeue_put_at(equeue, edesc_body, slot);
1543  slot++;
1544  }
1545 
1546  tx_packets++;
1547  tx_bytes += sh_len + p_len;
1548 
1549  /* The last segment may be less than gso_size. */
1550  data_len -= p_len;
1551  if (data_len < p_len)
1552  p_len = data_len;
1553  }
1554 
1555  /* Update stats. */
1556  tile_net_stats_add(tx_packets, &priv->stats.tx_packets);
1557  tile_net_stats_add(tx_bytes, &priv->stats.tx_bytes);
1558 }
1559 
1560 /* Do "TSO" handling for egress.
1561  *
1562  * Normally drivers set NETIF_F_TSO only to support hardware TSO;
1563  * otherwise the stack uses scatter-gather to implement GSO in software.
1564  * On our testing, enabling GSO support (via NETIF_F_SG) drops network
1565  * performance down to around 7.5 Gbps on the 10G interfaces, although
1566  * also dropping cpu utilization way down, to under 8%. But
1567  * implementing "TSO" in the driver brings performance back up to line
1568  * rate, while dropping cpu usage even further, to less than 4%. In
1569  * practice, profiling of GSO shows that skb_segment() is what causes
1570  * the performance overheads; we benefit in the driver from using
1571  * preallocated memory to duplicate the TCP/IP headers.
1572  */
1573 static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
1574 {
1575  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
1576  struct tile_net_priv *priv = netdev_priv(dev);
1577  int channel = priv->echannel;
1578  struct tile_net_egress *egress = &egress_for_echannel[channel];
1579  struct tile_net_comps *comps = info->comps_for_echannel[channel];
1580  gxio_mpipe_equeue_t *equeue = egress->equeue;
1581  unsigned long irqflags;
1582  int num_edescs;
1583  s64 slot;
1584 
1585  /* Determine how many mpipe edesc's are needed. */
1586  num_edescs = tso_count_edescs(skb);
1587 
1588  local_irq_save(irqflags);
1589 
1590  /* Try to acquire a completion entry and an egress slot. */
1591  slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps,
1592  equeue, num_edescs);
1593  if (slot < 0) {
1594  local_irq_restore(irqflags);
1595  return NETDEV_TX_BUSY;
1596  }
1597 
1598  /* Set up copies of header data properly. */
1599  tso_headers_prepare(skb, egress->headers, slot);
1600 
1601  /* Actually pass the data to the network hardware. */
1602  tso_egress(dev, equeue, skb, egress->headers, slot);
1603 
1604  /* Add a completion record. */
1605  add_comp(equeue, comps, slot + num_edescs - 1, skb);
1606 
1607  local_irq_restore(irqflags);
1608 
1609  /* Make sure the egress timer is scheduled. */
1610  tile_net_schedule_egress_timer();
1611 
1612  return NETDEV_TX_OK;
1613 }
1614 
1615 /* Analyze the body and frags for a transmit request. */
1616 static unsigned int tile_net_tx_frags(struct frag *frags,
1617  struct sk_buff *skb,
1618  void *b_data, unsigned int b_len)
1619 {
1620  unsigned int i, n = 0;
1621 
1622  struct skb_shared_info *sh = skb_shinfo(skb);
1623 
1624  if (b_len != 0) {
1625  frags[n].buf = b_data;
1626  frags[n++].length = b_len;
1627  }
1628 
1629  for (i = 0; i < sh->nr_frags; i++) {
1630  skb_frag_t *f = &sh->frags[i];
1631  frags[n].buf = tile_net_frag_buf(f);
1632  frags[n++].length = skb_frag_size(f);
1633  }
1634 
1635  return n;
1636 }
1637 
1638 /* Help the kernel transmit a packet. */
1639 static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
1640 {
1641  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
1642  struct tile_net_priv *priv = netdev_priv(dev);
1643  struct tile_net_egress *egress = &egress_for_echannel[priv->echannel];
1644  gxio_mpipe_equeue_t *equeue = egress->equeue;
1645  struct tile_net_comps *comps =
1646  info->comps_for_echannel[priv->echannel];
1647  unsigned int len = skb->len;
1648  unsigned char *data = skb->data;
1649  unsigned int num_edescs;
1650  struct frag frags[MAX_FRAGS];
1651  gxio_mpipe_edesc_t edescs[MAX_FRAGS];
1652  unsigned long irqflags;
1653  gxio_mpipe_edesc_t edesc = { { 0 } };
1654  unsigned int i;
1655  s64 slot;
1656 
1657  if (skb_is_gso(skb))
1658  return tile_net_tx_tso(skb, dev);
1659 
1660  num_edescs = tile_net_tx_frags(frags, skb, data, skb_headlen(skb));
1661 
1662  /* This is only used to specify the TLB. */
1663  edesc.stack_idx = large_buffer_stack;
1664 
1665  /* Prepare the edescs. */
1666  for (i = 0; i < num_edescs; i++) {
1667  edesc.xfer_size = frags[i].length;
1668  edesc.va = va_to_tile_io_addr(frags[i].buf);
1669  edescs[i] = edesc;
1670  }
1671 
1672  /* Mark the final edesc. */
1673  edescs[num_edescs - 1].bound = 1;
1674 
1675  /* Add checksum info to the initial edesc, if needed. */
1676  if (skb->ip_summed == CHECKSUM_PARTIAL) {
1677  unsigned int csum_start = skb_checksum_start_offset(skb);
1678  edescs[0].csum = 1;
1679  edescs[0].csum_start = csum_start;
1680  edescs[0].csum_dest = csum_start + skb->csum_offset;
1681  }
1682 
1683  local_irq_save(irqflags);
1684 
1685  /* Try to acquire a completion entry and an egress slot. */
1686  slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps,
1687  equeue, num_edescs);
1688  if (slot < 0) {
1689  local_irq_restore(irqflags);
1690  return NETDEV_TX_BUSY;
1691  }
1692 
1693  for (i = 0; i < num_edescs; i++)
1694  gxio_mpipe_equeue_put_at(equeue, edescs[i], slot++);
1695 
1696  /* Add a completion record. */
1697  add_comp(equeue, comps, slot - 1, skb);
1698 
1699  /* NOTE: Use ETH_ZLEN for short packets (e.g. 42 < 60). */
1700  tile_net_stats_add(1, &priv->stats.tx_packets);
1701  tile_net_stats_add(max_t(unsigned int, len, ETH_ZLEN),
1702  &priv->stats.tx_bytes);
1703 
1704  local_irq_restore(irqflags);
1705 
1706  /* Make sure the egress timer is scheduled. */
1707  tile_net_schedule_egress_timer();
1708 
1709  return NETDEV_TX_OK;
1710 }
1711 
1712 /* Return subqueue id on this core (one per core). */
1713 static u16 tile_net_select_queue(struct net_device *dev, struct sk_buff *skb)
1714 {
1715  return smp_processor_id();
1716 }
1717 
1718 /* Deal with a transmit timeout. */
1719 static void tile_net_tx_timeout(struct net_device *dev)
1720 {
1721  int cpu;
1722 
1723  for_each_online_cpu(cpu)
1724  netif_wake_subqueue(dev, cpu);
1725 }
1726 
1727 /* Ioctl commands. */
1728 static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
1729 {
1730  return -EOPNOTSUPP;
1731 }
1732 
1733 /* Get system network statistics for device. */
1734 static struct net_device_stats *tile_net_get_stats(struct net_device *dev)
1735 {
1736  struct tile_net_priv *priv = netdev_priv(dev);
1737  return &priv->stats;
1738 }
1739 
1740 /* Change the MTU. */
1741 static int tile_net_change_mtu(struct net_device *dev, int new_mtu)
1742 {
1743  if ((new_mtu < 68) || (new_mtu > 1500))
1744  return -EINVAL;
1745  dev->mtu = new_mtu;
1746  return 0;
1747 }
1748 
1749 /* Change the Ethernet address of the NIC.
1750  *
1751  * The hypervisor driver does not support changing MAC address. However,
1752  * the hardware does not do anything with the MAC address, so the address
1753  * which gets used on outgoing packets, and which is accepted on incoming
1754  * packets, is completely up to us.
1755  *
1756  * Returns 0 on success, negative on failure.
1757  */
1758 static int tile_net_set_mac_address(struct net_device *dev, void *p)
1759 {
1760  struct sockaddr *addr = p;
1761 
1762  if (!is_valid_ether_addr(addr->sa_data))
1763  return -EINVAL;
1764  memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
1765  return 0;
1766 }
1767 
1768 #ifdef CONFIG_NET_POLL_CONTROLLER
1769 /* Polling 'interrupt' - used by things like netconsole to send skbs
1770  * without having to re-enable interrupts. It's not called while
1771  * the interrupt routine is executing.
1772  */
1773 static void tile_net_netpoll(struct net_device *dev)
1774 {
1775  disable_percpu_irq(ingress_irq);
1776  tile_net_handle_ingress_irq(ingress_irq, NULL);
1777  enable_percpu_irq(ingress_irq, 0);
1778 }
1779 #endif
1780 
1781 static const struct net_device_ops tile_net_ops = {
1782  .ndo_open = tile_net_open,
1783  .ndo_stop = tile_net_stop,
1784  .ndo_start_xmit = tile_net_tx,
1785  .ndo_select_queue = tile_net_select_queue,
1786  .ndo_do_ioctl = tile_net_ioctl,
1787  .ndo_get_stats = tile_net_get_stats,
1788  .ndo_change_mtu = tile_net_change_mtu,
1789  .ndo_tx_timeout = tile_net_tx_timeout,
1790  .ndo_set_mac_address = tile_net_set_mac_address,
1791 #ifdef CONFIG_NET_POLL_CONTROLLER
1792  .ndo_poll_controller = tile_net_netpoll,
1793 #endif
1794 };
1795 
1796 /* The setup function.
1797  *
1798  * This uses ether_setup() to assign various fields in dev, including
1799  * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields.
1800  */
1801 static void tile_net_setup(struct net_device *dev)
1802 {
1803  ether_setup(dev);
1804  dev->netdev_ops = &tile_net_ops;
1806  dev->features |= NETIF_F_LLTX;
1807  dev->features |= NETIF_F_HW_CSUM;
1808  dev->features |= NETIF_F_SG;
1809  dev->features |= NETIF_F_TSO;
1810  dev->mtu = 1500;
1811 }
1812 
1813 /* Allocate the device structure, register the device, and obtain the
1814  * MAC address from the hypervisor.
1815  */
1816 static void tile_net_dev_init(const char *name, const uint8_t *mac)
1817 {
1818  int ret;
1819  int i;
1820  int nz_addr = 0;
1821  struct net_device *dev;
1822  struct tile_net_priv *priv;
1823 
1824  /* HACK: Ignore "loop" links. */
1825  if (strncmp(name, "loop", 4) == 0)
1826  return;
1827 
1828  /* Allocate the device structure. Normally, "name" is a
1829  * template, instantiated by register_netdev(), but not for us.
1830  */
1831  dev = alloc_netdev_mqs(sizeof(*priv), name, tile_net_setup,
1832  NR_CPUS, 1);
1833  if (!dev) {
1834  pr_err("alloc_netdev_mqs(%s) failed\n", name);
1835  return;
1836  }
1837 
1838  /* Initialize "priv". */
1839  priv = netdev_priv(dev);
1840  memset(priv, 0, sizeof(*priv));
1841  priv->dev = dev;
1842  priv->channel = -1;
1843  priv->loopify_channel = -1;
1844  priv->echannel = -1;
1845 
1846  /* Get the MAC address and set it in the device struct; this must
1847  * be done before the device is opened. If the MAC is all zeroes,
1848  * we use a random address, since we're probably on the simulator.
1849  */
1850  for (i = 0; i < 6; i++)
1851  nz_addr |= mac[i];
1852 
1853  if (nz_addr) {
1854  memcpy(dev->dev_addr, mac, 6);
1855  dev->addr_len = 6;
1856  } else {
1857  eth_hw_addr_random(dev);
1858  }
1859 
1860  /* Register the network device. */
1861  ret = register_netdev(dev);
1862  if (ret) {
1863  netdev_err(dev, "register_netdev failed %d\n", ret);
1864  free_netdev(dev);
1865  return;
1866  }
1867 }
1868 
1869 /* Per-cpu module initialization. */
1870 static void tile_net_init_module_percpu(void *unused)
1871 {
1872  struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
1873  int my_cpu = smp_processor_id();
1874 
1875  info->has_iqueue = false;
1876 
1877  info->my_cpu = my_cpu;
1878 
1879  /* Initialize the egress timer. */
1881  info->egress_timer.function = tile_net_handle_egress_timer;
1882 }
1883 
1884 /* Module initialization. */
1885 static int __init tile_net_init_module(void)
1886 {
1887  int i;
1888  char name[GXIO_MPIPE_LINK_NAME_LEN];
1889  uint8_t mac[6];
1890 
1891  pr_info("Tilera Network Driver\n");
1892 
1893  mutex_init(&tile_net_devs_for_channel_mutex);
1894 
1895  /* Initialize each CPU. */
1896  on_each_cpu(tile_net_init_module_percpu, NULL, 1);
1897 
1898  /* Find out what devices we have, and initialize them. */
1899  for (i = 0; gxio_mpipe_link_enumerate_mac(i, name, mac) >= 0; i++)
1900  tile_net_dev_init(name, mac);
1901 
1902  if (!network_cpus_init())
1903  network_cpus_map = *cpu_online_mask;
1904 
1905  return 0;
1906 }
1907 
1908 module_init(tile_net_init_module);