aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/net/tile/tilepro.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/net/tile/tilepro.c')
-rw-r--r--drivers/net/tile/tilepro.c2406
1 files changed, 2406 insertions, 0 deletions
diff --git a/drivers/net/tile/tilepro.c b/drivers/net/tile/tilepro.c
new file mode 100644
index 000000000000..0e6bac5ec65b
--- /dev/null
+++ b/drivers/net/tile/tilepro.c
@@ -0,0 +1,2406 @@
1/*
2 * Copyright 2010 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/netdevice.h> /* struct device, and other headers */
26#include <linux/etherdevice.h> /* eth_type_trans */
27#include <linux/skbuff.h>
28#include <linux/ioctl.h>
29#include <linux/cdev.h>
30#include <linux/hugetlb.h>
31#include <linux/in6.h>
32#include <linux/timer.h>
33#include <linux/io.h>
34#include <asm/checksum.h>
35#include <asm/homecache.h>
36
37#include <hv/drv_xgbe_intf.h>
38#include <hv/drv_xgbe_impl.h>
39#include <hv/hypervisor.h>
40#include <hv/netio_intf.h>
41
42/* For TSO */
43#include <linux/ip.h>
44#include <linux/tcp.h>
45
46
47/* There is no singlethread_cpu, so schedule work on the current cpu. */
48#define singlethread_cpu -1
49
50
51/*
52 * First, "tile_net_init_module()" initializes all four "devices" which
53 * can be used by linux.
54 *
55 * Then, "ifconfig DEVICE up" calls "tile_net_open()", which analyzes
56 * the network cpus, then uses "tile_net_open_aux()" to initialize
57 * LIPP/LEPP, and then uses "tile_net_open_inner()" to register all
58 * the tiles, provide buffers to LIPP, allow ingress to start, and
59 * turn on hypervisor interrupt handling (and NAPI) on all tiles.
60 *
61 * If registration fails due to the link being down, then "retry_work"
62 * is used to keep calling "tile_net_open_inner()" until it succeeds.
63 *
64 * If "ifconfig DEVICE down" is called, it uses "tile_net_stop()" to
65 * stop egress, drain the LIPP buffers, unregister all the tiles, stop
66 * LIPP/LEPP, and wipe the LEPP queue.
67 *
68 * We start out with the ingress interrupt enabled on each CPU. When
69 * this interrupt fires, we disable it, and call "napi_schedule()".
70 * This will cause "tile_net_poll()" to be called, which will pull
71 * packets from the netio queue, filtering them out, or passing them
72 * to "netif_receive_skb()". If our budget is exhausted, we will
73 * return, knowing we will be called again later. Otherwise, we
74 * reenable the ingress interrupt, and call "napi_complete()".
75 *
76 *
77 * NOTE: The use of "native_driver" ensures that EPP exists, and that
78 * "epp_sendv" is legal, and that "LIPP" is being used.
79 *
80 * NOTE: Failing to free completions for an arbitrarily long time
81 * (which is defined to be illegal) does in fact cause bizarre
82 * problems. The "egress_timer" helps prevent this from happening.
83 *
84 * NOTE: The egress code can be interrupted by the interrupt handler.
85 */
86
87
88/* HACK: Allow use of "jumbo" packets. */
89/* This should be 1500 if "jumbo" is not set in LIPP. */
90/* This should be at most 10226 (10240 - 14) if "jumbo" is set in LIPP. */
91/* ISSUE: This has not been thoroughly tested (except at 1500). */
92#define TILE_NET_MTU 1500
93
94/* HACK: Define to support GSO. */
95/* ISSUE: This may actually hurt performance of the TCP blaster. */
96/* #define TILE_NET_GSO */
97
98/* Define this to collapse "duplicate" acks. */
99/* #define IGNORE_DUP_ACKS */
100
101/* HACK: Define this to verify incoming packets. */
102/* #define TILE_NET_VERIFY_INGRESS */
103
104/* Use 3000 to enable the Linux Traffic Control (QoS) layer, else 0. */
105#define TILE_NET_TX_QUEUE_LEN 0
106
107/* Define to dump packets (prints out the whole packet on tx and rx). */
108/* #define TILE_NET_DUMP_PACKETS */
109
110/* Define to enable debug spew (all PDEBUG's are enabled). */
111/* #define TILE_NET_DEBUG */
112
113
114/* Define to activate paranoia checks. */
115/* #define TILE_NET_PARANOIA */
116
117/* Default transmit lockup timeout period, in jiffies. */
118#define TILE_NET_TIMEOUT (5 * HZ)
119
120/* Default retry interval for bringing up the NetIO interface, in jiffies. */
121#define TILE_NET_RETRY_INTERVAL (5 * HZ)
122
123/* Number of ports (xgbe0, xgbe1, gbe0, gbe1). */
124#define TILE_NET_DEVS 4
125
126
127
128/* Paranoia. */
129#if NET_IP_ALIGN != LIPP_PACKET_PADDING
130#error "NET_IP_ALIGN must match LIPP_PACKET_PADDING."
131#endif
132
133
134/* Debug print. */
135#ifdef TILE_NET_DEBUG
136#define PDEBUG(fmt, args...) net_printk(fmt, ## args)
137#else
138#define PDEBUG(fmt, args...)
139#endif
140
141
142MODULE_AUTHOR("Tilera");
143MODULE_LICENSE("GPL");
144
145
146#define IS_MULTICAST(mac_addr) \
147 (((u8 *)(mac_addr))[0] & 0x01)
148
149#define IS_BROADCAST(mac_addr) \
150 (((u16 *)(mac_addr))[0] == 0xffff)
151
152
153/*
154 * Queue of incoming packets for a specific cpu and device.
155 *
156 * Includes a pointer to the "system" data, and the actual "user" data.
157 */
158struct tile_netio_queue {
159 netio_queue_impl_t *__system_part;
160 netio_queue_user_impl_t __user_part;
161
162};
163
164
165/*
166 * Statistics counters for a specific cpu and device.
167 */
168struct tile_net_stats_t {
169 u32 rx_packets;
170 u32 rx_bytes;
171 u32 tx_packets;
172 u32 tx_bytes;
173};
174
175
176/*
177 * Info for a specific cpu and device.
178 *
179 * ISSUE: There is a "dev" pointer in "napi" as well.
180 */
181struct tile_net_cpu {
182 /* The NAPI struct. */
183 struct napi_struct napi;
184 /* Packet queue. */
185 struct tile_netio_queue queue;
186 /* Statistics. */
187 struct tile_net_stats_t stats;
188 /* ISSUE: Is this needed? */
189 bool napi_enabled;
190 /* True if this tile has succcessfully registered with the IPP. */
191 bool registered;
192 /* True if the link was down last time we tried to register. */
193 bool link_down;
194 /* True if "egress_timer" is scheduled. */
195 bool egress_timer_scheduled;
196 /* Number of small sk_buffs which must still be provided. */
197 unsigned int num_needed_small_buffers;
198 /* Number of large sk_buffs which must still be provided. */
199 unsigned int num_needed_large_buffers;
200 /* A timer for handling egress completions. */
201 struct timer_list egress_timer;
202};
203
204
205/*
206 * Info for a specific device.
207 */
208struct tile_net_priv {
209 /* Our network device. */
210 struct net_device *dev;
211 /* The actual egress queue. */
212 lepp_queue_t *epp_queue;
213 /* Protects "epp_queue->cmd_tail" and "epp_queue->comp_tail" */
214 spinlock_t cmd_lock;
215 /* Protects "epp_queue->comp_head". */
216 spinlock_t comp_lock;
217 /* The hypervisor handle for this interface. */
218 int hv_devhdl;
219 /* The intr bit mask that IDs this device. */
220 u32 intr_id;
221 /* True iff "tile_net_open_aux()" has succeeded. */
222 int partly_opened;
223 /* True iff "tile_net_open_inner()" has succeeded. */
224 int fully_opened;
225 /* Effective network cpus. */
226 struct cpumask network_cpus_map;
227 /* Number of network cpus. */
228 int network_cpus_count;
229 /* Credits per network cpu. */
230 int network_cpus_credits;
231 /* Network stats. */
232 struct net_device_stats stats;
233 /* For NetIO bringup retries. */
234 struct delayed_work retry_work;
235 /* Quick access to per cpu data. */
236 struct tile_net_cpu *cpu[NR_CPUS];
237};
238
239
240/*
241 * The actual devices (xgbe0, xgbe1, gbe0, gbe1).
242 */
243static struct net_device *tile_net_devs[TILE_NET_DEVS];
244
245/*
246 * The "tile_net_cpu" structures for each device.
247 */
248static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe0);
249static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe1);
250static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe0);
251static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe1);
252
253
254/*
255 * True if "network_cpus" was specified.
256 */
257static bool network_cpus_used;
258
259/*
260 * The actual cpus in "network_cpus".
261 */
262static struct cpumask network_cpus_map;
263
264
265
266#ifdef TILE_NET_DEBUG
267/*
268 * printk with extra stuff.
269 *
270 * We print the CPU we're running in brackets.
271 */
272static void net_printk(char *fmt, ...)
273{
274 int i;
275 int len;
276 va_list args;
277 static char buf[256];
278
279 len = sprintf(buf, "tile_net[%2.2d]: ", smp_processor_id());
280 va_start(args, fmt);
281 i = vscnprintf(buf + len, sizeof(buf) - len - 1, fmt, args);
282 va_end(args);
283 buf[255] = '\0';
284 pr_notice(buf);
285}
286#endif
287
288
289#ifdef TILE_NET_DUMP_PACKETS
290/*
291 * Dump a packet.
292 */
293static void dump_packet(unsigned char *data, unsigned long length, char *s)
294{
295 unsigned long i;
296 static unsigned int count;
297
298 pr_info("dump_packet(data %p, length 0x%lx s %s count 0x%x)\n",
299 data, length, s, count++);
300
301 pr_info("\n");
302
303 for (i = 0; i < length; i++) {
304 if ((i & 0xf) == 0)
305 sprintf(buf, "%8.8lx:", i);
306 sprintf(buf + strlen(buf), " %2.2x", data[i]);
307 if ((i & 0xf) == 0xf || i == length - 1)
308 pr_info("%s\n", buf);
309 }
310}
311#endif
312
313
314/*
315 * Provide support for the __netio_fastio1() swint
316 * (see <hv/drv_xgbe_intf.h> for how it is used).
317 *
318 * The fastio swint2 call may clobber all the caller-saved registers.
319 * It rarely clobbers memory, but we allow for the possibility in
320 * the signature just to be on the safe side.
321 *
322 * Also, gcc doesn't seem to allow an input operand to be
323 * clobbered, so we fake it with dummy outputs.
324 *
325 * This function can't be static because of the way it is declared
326 * in the netio header.
327 */
328inline int __netio_fastio1(u32 fastio_index, u32 arg0)
329{
330 long result, clobber_r1, clobber_r10;
331 asm volatile("swint2"
332 : "=R00" (result),
333 "=R01" (clobber_r1), "=R10" (clobber_r10)
334 : "R10" (fastio_index), "R01" (arg0)
335 : "memory", "r2", "r3", "r4",
336 "r5", "r6", "r7", "r8", "r9",
337 "r11", "r12", "r13", "r14",
338 "r15", "r16", "r17", "r18", "r19",
339 "r20", "r21", "r22", "r23", "r24",
340 "r25", "r26", "r27", "r28", "r29");
341 return result;
342}
343
344
345/*
346 * Provide a linux buffer to LIPP.
347 */
348static void tile_net_provide_linux_buffer(struct tile_net_cpu *info,
349 void *va, bool small)
350{
351 struct tile_netio_queue *queue = &info->queue;
352
353 /* Convert "va" and "small" to "linux_buffer_t". */
354 unsigned int buffer = ((unsigned int)(__pa(va) >> 7) << 1) + small;
355
356 __netio_fastio_free_buffer(queue->__user_part.__fastio_index, buffer);
357}
358
359
360/*
361 * Provide a linux buffer for LIPP.
362 */
363static bool tile_net_provide_needed_buffer(struct tile_net_cpu *info,
364 bool small)
365{
366 /* ISSUE: What should we use here? */
367 unsigned int large_size = NET_IP_ALIGN + TILE_NET_MTU + 100;
368
369 /* Round up to ensure to avoid "false sharing" with last cache line. */
370 unsigned int buffer_size =
371 (((small ? LIPP_SMALL_PACKET_SIZE : large_size) +
372 CHIP_L2_LINE_SIZE() - 1) & -CHIP_L2_LINE_SIZE());
373
374 /*
375 * ISSUE: Since CPAs are 38 bits, and we can only encode the
376 * high 31 bits in a "linux_buffer_t", the low 7 bits must be
377 * zero, and thus, we must align the actual "va" mod 128.
378 */
379 const unsigned long align = 128;
380
381 struct sk_buff *skb;
382 void *va;
383
384 struct sk_buff **skb_ptr;
385
386 /* Note that "dev_alloc_skb()" adds NET_SKB_PAD more bytes, */
387 /* and also "reserves" that many bytes. */
388 /* ISSUE: Can we "share" the NET_SKB_PAD bytes with "skb_ptr"? */
389 int len = sizeof(*skb_ptr) + align + buffer_size;
390
391 while (1) {
392
393 /* Allocate (or fail). */
394 skb = dev_alloc_skb(len);
395 if (skb == NULL)
396 return false;
397
398 /* Make room for a back-pointer to 'skb'. */
399 skb_reserve(skb, sizeof(*skb_ptr));
400
401 /* Make sure we are aligned. */
402 skb_reserve(skb, -(long)skb->data & (align - 1));
403
404 /* This address is given to IPP. */
405 va = skb->data;
406
407 if (small)
408 break;
409
410 /* ISSUE: This has never been observed! */
411 /* Large buffers must not span a huge page. */
412 if (((((long)va & ~HPAGE_MASK) + 1535) & HPAGE_MASK) == 0)
413 break;
414 pr_err("Leaking unaligned linux buffer at %p.\n", va);
415 }
416
417 /* Skip two bytes to satisfy LIPP assumptions. */
418 /* Note that this aligns IP on a 16 byte boundary. */
419 /* ISSUE: Do this when the packet arrives? */
420 skb_reserve(skb, NET_IP_ALIGN);
421
422 /* Save a back-pointer to 'skb'. */
423 skb_ptr = va - sizeof(*skb_ptr);
424 *skb_ptr = skb;
425
426 /* Invalidate the packet buffer. */
427 if (!hash_default)
428 __inv_buffer(skb->data, buffer_size);
429
430 /* Make sure "skb_ptr" has been flushed. */
431 __insn_mf();
432
433#ifdef TILE_NET_PARANOIA
434#if CHIP_HAS_CBOX_HOME_MAP()
435 if (hash_default) {
436 HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)va);
437 if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
438 panic("Non-coherent ingress buffer!");
439 }
440#endif
441#endif
442
443 /* Provide the new buffer. */
444 tile_net_provide_linux_buffer(info, va, small);
445
446 return true;
447}
448
449
450/*
451 * Provide linux buffers for LIPP.
452 */
453static void tile_net_provide_needed_buffers(struct tile_net_cpu *info)
454{
455 while (info->num_needed_small_buffers != 0) {
456 if (!tile_net_provide_needed_buffer(info, true))
457 goto oops;
458 info->num_needed_small_buffers--;
459 }
460
461 while (info->num_needed_large_buffers != 0) {
462 if (!tile_net_provide_needed_buffer(info, false))
463 goto oops;
464 info->num_needed_large_buffers--;
465 }
466
467 return;
468
469oops:
470
471 /* Add a description to the page allocation failure dump. */
472 pr_notice("Could not provide a linux buffer to LIPP.\n");
473}
474
475
476/*
477 * Grab some LEPP completions, and store them in "comps", of size
478 * "comps_size", and return the number of completions which were
479 * stored, so the caller can free them.
480 *
481 * If "pending" is not NULL, it will be set to true if there might
482 * still be some pending completions caused by this tile, else false.
483 */
484static unsigned int tile_net_lepp_grab_comps(struct net_device *dev,
485 struct sk_buff *comps[],
486 unsigned int comps_size,
487 bool *pending)
488{
489 struct tile_net_priv *priv = netdev_priv(dev);
490
491 lepp_queue_t *eq = priv->epp_queue;
492
493 unsigned int n = 0;
494
495 unsigned int comp_head;
496 unsigned int comp_busy;
497 unsigned int comp_tail;
498
499 spin_lock(&priv->comp_lock);
500
501 comp_head = eq->comp_head;
502 comp_busy = eq->comp_busy;
503 comp_tail = eq->comp_tail;
504
505 while (comp_head != comp_busy && n < comps_size) {
506 comps[n++] = eq->comps[comp_head];
507 LEPP_QINC(comp_head);
508 }
509
510 if (pending != NULL)
511 *pending = (comp_head != comp_tail);
512
513 eq->comp_head = comp_head;
514
515 spin_unlock(&priv->comp_lock);
516
517 return n;
518}
519
520
521/*
522 * Make sure the egress timer is scheduled.
523 *
524 * Note that we use "schedule if not scheduled" logic instead of the more
525 * obvious "reschedule" logic, because "reschedule" is fairly expensive.
526 */
527static void tile_net_schedule_egress_timer(struct tile_net_cpu *info)
528{
529 if (!info->egress_timer_scheduled) {
530 mod_timer_pinned(&info->egress_timer, jiffies + 1);
531 info->egress_timer_scheduled = true;
532 }
533}
534
535
536/*
537 * The "function" for "info->egress_timer".
538 *
539 * This timer will reschedule itself as long as there are any pending
540 * completions expected (on behalf of any tile).
541 *
542 * ISSUE: Realistically, will the timer ever stop scheduling itself?
543 *
544 * ISSUE: This timer is almost never actually needed, so just use a global
545 * timer that can run on any tile.
546 *
547 * ISSUE: Maybe instead track number of expected completions, and free
548 * only that many, resetting to zero if "pending" is ever false.
549 */
550static void tile_net_handle_egress_timer(unsigned long arg)
551{
552 struct tile_net_cpu *info = (struct tile_net_cpu *)arg;
553 struct net_device *dev = info->napi.dev;
554
555 struct sk_buff *olds[32];
556 unsigned int wanted = 32;
557 unsigned int i, nolds = 0;
558 bool pending;
559
560 /* The timer is no longer scheduled. */
561 info->egress_timer_scheduled = false;
562
563 nolds = tile_net_lepp_grab_comps(dev, olds, wanted, &pending);
564
565 for (i = 0; i < nolds; i++)
566 kfree_skb(olds[i]);
567
568 /* Reschedule timer if needed. */
569 if (pending)
570 tile_net_schedule_egress_timer(info);
571}
572
573
574#ifdef IGNORE_DUP_ACKS
575
576/*
577 * Help detect "duplicate" ACKs. These are sequential packets (for a
578 * given flow) which are exactly 66 bytes long, sharing everything but
579 * ID=2@0x12, Hsum=2@0x18, Ack=4@0x2a, WinSize=2@0x30, Csum=2@0x32,
580 * Tstamps=10@0x38. The ID's are +1, the Hsum's are -1, the Ack's are
581 * +N, and the Tstamps are usually identical.
582 *
583 * NOTE: Apparently truly duplicate acks (with identical "ack" values),
584 * should not be collapsed, as they are used for some kind of flow control.
585 */
586static bool is_dup_ack(char *s1, char *s2, unsigned int len)
587{
588 int i;
589
590 unsigned long long ignorable = 0;
591
592 /* Identification. */
593 ignorable |= (1ULL << 0x12);
594 ignorable |= (1ULL << 0x13);
595
596 /* Header checksum. */
597 ignorable |= (1ULL << 0x18);
598 ignorable |= (1ULL << 0x19);
599
600 /* ACK. */
601 ignorable |= (1ULL << 0x2a);
602 ignorable |= (1ULL << 0x2b);
603 ignorable |= (1ULL << 0x2c);
604 ignorable |= (1ULL << 0x2d);
605
606 /* WinSize. */
607 ignorable |= (1ULL << 0x30);
608 ignorable |= (1ULL << 0x31);
609
610 /* Checksum. */
611 ignorable |= (1ULL << 0x32);
612 ignorable |= (1ULL << 0x33);
613
614 for (i = 0; i < len; i++, ignorable >>= 1) {
615
616 if ((ignorable & 1) || (s1[i] == s2[i]))
617 continue;
618
619#ifdef TILE_NET_DEBUG
620 /* HACK: Mention non-timestamp diffs. */
621 if (i < 0x38 && i != 0x2f &&
622 net_ratelimit())
623 pr_info("Diff at 0x%x\n", i);
624#endif
625
626 return false;
627 }
628
629#ifdef TILE_NET_NO_SUPPRESS_DUP_ACKS
630 /* HACK: Do not suppress truly duplicate ACKs. */
631 /* ISSUE: Is this actually necessary or helpful? */
632 if (s1[0x2a] == s2[0x2a] &&
633 s1[0x2b] == s2[0x2b] &&
634 s1[0x2c] == s2[0x2c] &&
635 s1[0x2d] == s2[0x2d]) {
636 return false;
637 }
638#endif
639
640 return true;
641}
642
643#endif
644
645
646
647/*
648 * Like "tile_net_handle_packets()", but just discard packets.
649 */
650static void tile_net_discard_packets(struct net_device *dev)
651{
652 struct tile_net_priv *priv = netdev_priv(dev);
653 int my_cpu = smp_processor_id();
654 struct tile_net_cpu *info = priv->cpu[my_cpu];
655 struct tile_netio_queue *queue = &info->queue;
656 netio_queue_impl_t *qsp = queue->__system_part;
657 netio_queue_user_impl_t *qup = &queue->__user_part;
658
659 while (qup->__packet_receive_read !=
660 qsp->__packet_receive_queue.__packet_write) {
661
662 int index = qup->__packet_receive_read;
663
664 int index2_aux = index + sizeof(netio_pkt_t);
665 int index2 =
666 ((index2_aux ==
667 qsp->__packet_receive_queue.__last_packet_plus_one) ?
668 0 : index2_aux);
669
670 netio_pkt_t *pkt = (netio_pkt_t *)
671 ((unsigned long) &qsp[1] + index);
672
673 /* Extract the "linux_buffer_t". */
674 unsigned int buffer = pkt->__packet.word;
675
676 /* Convert "linux_buffer_t" to "va". */
677 void *va = __va((phys_addr_t)(buffer >> 1) << 7);
678
679 /* Acquire the associated "skb". */
680 struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
681 struct sk_buff *skb = *skb_ptr;
682
683 kfree_skb(skb);
684
685 /* Consume this packet. */
686 qup->__packet_receive_read = index2;
687 }
688}
689
690
691/*
692 * Handle the next packet. Return true if "processed", false if "filtered".
693 */
694static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
695{
696 struct net_device *dev = info->napi.dev;
697
698 struct tile_netio_queue *queue = &info->queue;
699 netio_queue_impl_t *qsp = queue->__system_part;
700 netio_queue_user_impl_t *qup = &queue->__user_part;
701 struct tile_net_stats_t *stats = &info->stats;
702
703 int filter;
704
705 int index2_aux = index + sizeof(netio_pkt_t);
706 int index2 =
707 ((index2_aux ==
708 qsp->__packet_receive_queue.__last_packet_plus_one) ?
709 0 : index2_aux);
710
711 netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
712
713 netio_pkt_metadata_t *metadata = NETIO_PKT_METADATA(pkt);
714
715 /* Extract the packet size. */
716 unsigned long len =
717 (NETIO_PKT_CUSTOM_LENGTH(pkt) +
718 NET_IP_ALIGN - NETIO_PACKET_PADDING);
719
720 /* Extract the "linux_buffer_t". */
721 unsigned int buffer = pkt->__packet.word;
722
723 /* Extract "small" (vs "large"). */
724 bool small = ((buffer & 1) != 0);
725
726 /* Convert "linux_buffer_t" to "va". */
727 void *va = __va((phys_addr_t)(buffer >> 1) << 7);
728
729 /* Extract the packet data pointer. */
730 /* Compare to "NETIO_PKT_CUSTOM_DATA(pkt)". */
731 unsigned char *buf = va + NET_IP_ALIGN;
732
733#ifdef IGNORE_DUP_ACKS
734
735 static int other;
736 static int final;
737 static int keep;
738 static int skip;
739
740#endif
741
742 /* Invalidate the packet buffer. */
743 if (!hash_default)
744 __inv_buffer(buf, len);
745
746 /* ISSUE: Is this needed? */
747 dev->last_rx = jiffies;
748
749#ifdef TILE_NET_DUMP_PACKETS
750 dump_packet(buf, len, "rx");
751#endif /* TILE_NET_DUMP_PACKETS */
752
753#ifdef TILE_NET_VERIFY_INGRESS
754 if (!NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt) &&
755 NETIO_PKT_L4_CSUM_CALCULATED_M(metadata, pkt)) {
756 /*
757 * FIXME: This complains about UDP packets
758 * with a "zero" checksum (bug 6624).
759 */
760#ifdef TILE_NET_PANIC_ON_BAD
761 dump_packet(buf, len, "rx");
762 panic("Bad L4 checksum.");
763#else
764 pr_warning("Bad L4 checksum on %d byte packet.\n", len);
765#endif
766 }
767 if (!NETIO_PKT_L3_CSUM_CORRECT_M(metadata, pkt) &&
768 NETIO_PKT_L3_CSUM_CALCULATED_M(metadata, pkt)) {
769 dump_packet(buf, len, "rx");
770 panic("Bad L3 checksum.");
771 }
772 switch (NETIO_PKT_STATUS_M(metadata, pkt)) {
773 case NETIO_PKT_STATUS_OVERSIZE:
774 if (len >= 64) {
775 dump_packet(buf, len, "rx");
776 panic("Unexpected OVERSIZE.");
777 }
778 break;
779 case NETIO_PKT_STATUS_BAD:
780#ifdef TILE_NET_PANIC_ON_BAD
781 dump_packet(buf, len, "rx");
782 panic("Unexpected BAD packet.");
783#else
784 pr_warning("Unexpected BAD %d byte packet.\n", len);
785#endif
786 }
787#endif
788
789 filter = 0;
790
791 if (!(dev->flags & IFF_UP)) {
792 /* Filter packets received before we're up. */
793 filter = 1;
794 } else if (!(dev->flags & IFF_PROMISC)) {
795 /*
796 * FIXME: Implement HW multicast filter.
797 */
798 if (!IS_MULTICAST(buf) && !IS_BROADCAST(buf)) {
799 /* Filter packets not for our address. */
800 const u8 *mine = dev->dev_addr;
801 filter = compare_ether_addr(mine, buf);
802 }
803 }
804
805#ifdef IGNORE_DUP_ACKS
806
807 if (len != 66) {
808 /* FIXME: Must check "is_tcp_ack(buf, len)" somehow. */
809
810 other++;
811
812 } else if (index2 ==
813 qsp->__packet_receive_queue.__packet_write) {
814
815 final++;
816
817 } else {
818
819 netio_pkt_t *pkt2 = (netio_pkt_t *)
820 ((unsigned long) &qsp[1] + index2);
821
822 netio_pkt_metadata_t *metadata2 =
823 NETIO_PKT_METADATA(pkt2);
824
825 /* Extract the packet size. */
826 unsigned long len2 =
827 (NETIO_PKT_CUSTOM_LENGTH(pkt2) +
828 NET_IP_ALIGN - NETIO_PACKET_PADDING);
829
830 if (len2 == 66 &&
831 NETIO_PKT_FLOW_HASH_M(metadata, pkt) ==
832 NETIO_PKT_FLOW_HASH_M(metadata2, pkt2)) {
833
834 /* Extract the "linux_buffer_t". */
835 unsigned int buffer2 = pkt2->__packet.word;
836
837 /* Convert "linux_buffer_t" to "va". */
838 void *va2 =
839 __va((phys_addr_t)(buffer2 >> 1) << 7);
840
841 /* Extract the packet data pointer. */
842 /* Compare to "NETIO_PKT_CUSTOM_DATA(pkt)". */
843 unsigned char *buf2 = va2 + NET_IP_ALIGN;
844
845 /* Invalidate the packet buffer. */
846 if (!hash_default)
847 __inv_buffer(buf2, len2);
848
849 if (is_dup_ack(buf, buf2, len)) {
850 skip++;
851 filter = 1;
852 } else {
853 keep++;
854 }
855 }
856 }
857
858 if (net_ratelimit())
859 pr_info("Other %d Final %d Keep %d Skip %d.\n",
860 other, final, keep, skip);
861
862#endif
863
864 if (filter) {
865
866 /* ISSUE: Update "drop" statistics? */
867
868 tile_net_provide_linux_buffer(info, va, small);
869
870 } else {
871
872 /* Acquire the associated "skb". */
873 struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
874 struct sk_buff *skb = *skb_ptr;
875
876 /* Paranoia. */
877 if (skb->data != buf)
878 panic("Corrupt linux buffer from LIPP! "
879 "VA=%p, skb=%p, skb->data=%p\n",
880 va, skb, skb->data);
881
882 /* Encode the actual packet length. */
883 skb_put(skb, len);
884
885 /* NOTE: This call also sets "skb->dev = dev". */
886 skb->protocol = eth_type_trans(skb, dev);
887
888 /* ISSUE: Discard corrupt packets? */
889 /* ISSUE: Discard packets with bad checksums? */
890
891 /* Avoid recomputing TCP/UDP checksums. */
892 if (NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt))
893 skb->ip_summed = CHECKSUM_UNNECESSARY;
894
895 netif_receive_skb(skb);
896
897 stats->rx_packets++;
898 stats->rx_bytes += len;
899
900 if (small)
901 info->num_needed_small_buffers++;
902 else
903 info->num_needed_large_buffers++;
904 }
905
906 /* Return four credits after every fourth packet. */
907 if (--qup->__receive_credit_remaining == 0) {
908 u32 interval = qup->__receive_credit_interval;
909 qup->__receive_credit_remaining = interval;
910 __netio_fastio_return_credits(qup->__fastio_index, interval);
911 }
912
913 /* Consume this packet. */
914 qup->__packet_receive_read = index2;
915
916 return !filter;
917}
918
919
920/*
921 * Handle some packets for the given device on the current CPU.
922 *
923 * ISSUE: The "rotting packet" race condition occurs if a packet
924 * arrives after the queue appears to be empty, and before the
925 * hypervisor interrupt is re-enabled.
926 */
927static int tile_net_poll(struct napi_struct *napi, int budget)
928{
929 struct net_device *dev = napi->dev;
930 struct tile_net_priv *priv = netdev_priv(dev);
931 int my_cpu = smp_processor_id();
932 struct tile_net_cpu *info = priv->cpu[my_cpu];
933 struct tile_netio_queue *queue = &info->queue;
934 netio_queue_impl_t *qsp = queue->__system_part;
935 netio_queue_user_impl_t *qup = &queue->__user_part;
936
937 unsigned int work = 0;
938
939 while (1) {
940 int index = qup->__packet_receive_read;
941 if (index == qsp->__packet_receive_queue.__packet_write)
942 break;
943
944 if (tile_net_poll_aux(info, index)) {
945 if (++work >= budget)
946 goto done;
947 }
948 }
949
950 napi_complete(&info->napi);
951
952 /* Re-enable hypervisor interrupts. */
953 enable_percpu_irq(priv->intr_id);
954
955 /* HACK: Avoid the "rotting packet" problem. */
956 if (qup->__packet_receive_read !=
957 qsp->__packet_receive_queue.__packet_write)
958 napi_schedule(&info->napi);
959
960 /* ISSUE: Handle completions? */
961
962done:
963
964 tile_net_provide_needed_buffers(info);
965
966 return work;
967}
968
969
970/*
971 * Handle an ingress interrupt for the given device on the current cpu.
972 */
973static irqreturn_t tile_net_handle_ingress_interrupt(int irq, void *dev_ptr)
974{
975 struct net_device *dev = (struct net_device *)dev_ptr;
976 struct tile_net_priv *priv = netdev_priv(dev);
977 int my_cpu = smp_processor_id();
978 struct tile_net_cpu *info = priv->cpu[my_cpu];
979
980 /* Disable hypervisor interrupt. */
981 disable_percpu_irq(priv->intr_id);
982
983 napi_schedule(&info->napi);
984
985 return IRQ_HANDLED;
986}
987
988
989/*
990 * One time initialization per interface.
991 */
992static int tile_net_open_aux(struct net_device *dev)
993{
994 struct tile_net_priv *priv = netdev_priv(dev);
995
996 int ret;
997 int dummy;
998 unsigned int epp_lotar;
999
1000 /*
1001 * Find out where EPP memory should be homed.
1002 */
1003 ret = hv_dev_pread(priv->hv_devhdl, 0,
1004 (HV_VirtAddr)&epp_lotar, sizeof(epp_lotar),
1005 NETIO_EPP_SHM_OFF);
1006 if (ret < 0) {
1007 pr_err("could not read epp_shm_queue lotar.\n");
1008 return -EIO;
1009 }
1010
1011 /*
1012 * Home the page on the EPP.
1013 */
1014 {
1015 int epp_home = hv_lotar_to_cpu(epp_lotar);
1016 struct page *page = virt_to_page(priv->epp_queue);
1017 homecache_change_page_home(page, 0, epp_home);
1018 }
1019
1020 /*
1021 * Register the EPP shared memory queue.
1022 */
1023 {
1024 netio_ipp_address_t ea = {
1025 .va = 0,
1026 .pa = __pa(priv->epp_queue),
1027 .pte = hv_pte(0),
1028 .size = PAGE_SIZE,
1029 };
1030 ea.pte = hv_pte_set_lotar(ea.pte, epp_lotar);
1031 ea.pte = hv_pte_set_mode(ea.pte, HV_PTE_MODE_CACHE_TILE_L3);
1032 ret = hv_dev_pwrite(priv->hv_devhdl, 0,
1033 (HV_VirtAddr)&ea,
1034 sizeof(ea),
1035 NETIO_EPP_SHM_OFF);
1036 if (ret < 0)
1037 return -EIO;
1038 }
1039
1040 /*
1041 * Start LIPP/LEPP.
1042 */
1043 if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1044 sizeof(dummy), NETIO_IPP_START_SHIM_OFF) < 0) {
1045 pr_warning("Failed to start LIPP/LEPP.\n");
1046 return -EIO;
1047 }
1048
1049 return 0;
1050}
1051
1052
1053/*
1054 * Register with hypervisor on each CPU.
1055 *
1056 * Strangely, this function does important things even if it "fails",
1057 * which is especially common if the link is not up yet. Hopefully
1058 * these things are all "harmless" if done twice!
1059 */
1060static void tile_net_register(void *dev_ptr)
1061{
1062 struct net_device *dev = (struct net_device *)dev_ptr;
1063 struct tile_net_priv *priv = netdev_priv(dev);
1064 int my_cpu = smp_processor_id();
1065 struct tile_net_cpu *info;
1066
1067 struct tile_netio_queue *queue;
1068
1069 /* Only network cpus can receive packets. */
1070 int queue_id =
1071 cpumask_test_cpu(my_cpu, &priv->network_cpus_map) ? 0 : 255;
1072
1073 netio_input_config_t config = {
1074 .flags = 0,
1075 .num_receive_packets = priv->network_cpus_credits,
1076 .queue_id = queue_id
1077 };
1078
1079 int ret = 0;
1080 netio_queue_impl_t *queuep;
1081
1082 PDEBUG("tile_net_register(queue_id %d)\n", queue_id);
1083
1084 if (!strcmp(dev->name, "xgbe0"))
1085 info = &__get_cpu_var(hv_xgbe0);
1086 else if (!strcmp(dev->name, "xgbe1"))
1087 info = &__get_cpu_var(hv_xgbe1);
1088 else if (!strcmp(dev->name, "gbe0"))
1089 info = &__get_cpu_var(hv_gbe0);
1090 else if (!strcmp(dev->name, "gbe1"))
1091 info = &__get_cpu_var(hv_gbe1);
1092 else
1093 BUG();
1094
1095 /* Initialize the egress timer. */
1096 init_timer(&info->egress_timer);
1097 info->egress_timer.data = (long)info;
1098 info->egress_timer.function = tile_net_handle_egress_timer;
1099
1100 priv->cpu[my_cpu] = info;
1101
1102 /*
1103 * Register ourselves with the IPP.
1104 */
1105 ret = hv_dev_pwrite(priv->hv_devhdl, 0,
1106 (HV_VirtAddr)&config,
1107 sizeof(netio_input_config_t),
1108 NETIO_IPP_INPUT_REGISTER_OFF);
1109 PDEBUG("hv_dev_pwrite(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
1110 ret);
1111 if (ret < 0) {
1112 printk(KERN_DEBUG "hv_dev_pwrite NETIO_IPP_INPUT_REGISTER_OFF"
1113 " failure %d\n", ret);
1114 info->link_down = (ret == NETIO_LINK_DOWN);
1115 return;
1116 }
1117
1118 /*
1119 * Get the pointer to our queue's system part.
1120 */
1121
1122 ret = hv_dev_pread(priv->hv_devhdl, 0,
1123 (HV_VirtAddr)&queuep,
1124 sizeof(netio_queue_impl_t *),
1125 NETIO_IPP_INPUT_REGISTER_OFF);
1126 PDEBUG("hv_dev_pread(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
1127 ret);
1128 PDEBUG("queuep %p\n", queuep);
1129 if (ret <= 0) {
1130 /* ISSUE: Shouldn't this be a fatal error? */
1131 pr_err("hv_dev_pread NETIO_IPP_INPUT_REGISTER_OFF failure\n");
1132 return;
1133 }
1134
1135 queue = &info->queue;
1136
1137 queue->__system_part = queuep;
1138
1139 memset(&queue->__user_part, 0, sizeof(netio_queue_user_impl_t));
1140
1141 /* This is traditionally "config.num_receive_packets / 2". */
1142 queue->__user_part.__receive_credit_interval = 4;
1143 queue->__user_part.__receive_credit_remaining =
1144 queue->__user_part.__receive_credit_interval;
1145
1146 /*
1147 * Get a fastio index from the hypervisor.
1148 * ISSUE: Shouldn't this check the result?
1149 */
1150 ret = hv_dev_pread(priv->hv_devhdl, 0,
1151 (HV_VirtAddr)&queue->__user_part.__fastio_index,
1152 sizeof(queue->__user_part.__fastio_index),
1153 NETIO_IPP_GET_FASTIO_OFF);
1154 PDEBUG("hv_dev_pread(NETIO_IPP_GET_FASTIO_OFF) returned %d\n", ret);
1155
1156 netif_napi_add(dev, &info->napi, tile_net_poll, 64);
1157
1158 /* Now we are registered. */
1159 info->registered = true;
1160}
1161
1162
1163/*
1164 * Unregister with hypervisor on each CPU.
1165 */
1166static void tile_net_unregister(void *dev_ptr)
1167{
1168 struct net_device *dev = (struct net_device *)dev_ptr;
1169 struct tile_net_priv *priv = netdev_priv(dev);
1170 int my_cpu = smp_processor_id();
1171 struct tile_net_cpu *info = priv->cpu[my_cpu];
1172
1173 int ret = 0;
1174 int dummy = 0;
1175
1176 /* Do nothing if never registered. */
1177 if (info == NULL)
1178 return;
1179
1180 /* Do nothing if already unregistered. */
1181 if (!info->registered)
1182 return;
1183
1184 /*
1185 * Unregister ourselves with LIPP.
1186 */
1187 ret = hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1188 sizeof(dummy), NETIO_IPP_INPUT_UNREGISTER_OFF);
1189 PDEBUG("hv_dev_pwrite(NETIO_IPP_INPUT_UNREGISTER_OFF) returned %d\n",
1190 ret);
1191 if (ret < 0) {
1192 /* FIXME: Just panic? */
1193 pr_err("hv_dev_pwrite NETIO_IPP_INPUT_UNREGISTER_OFF"
1194 " failure %d\n", ret);
1195 }
1196
1197 /*
1198 * Discard all packets still in our NetIO queue. Hopefully,
1199 * once the unregister call is complete, there will be no
1200 * packets still in flight on the IDN.
1201 */
1202 tile_net_discard_packets(dev);
1203
1204 /* Reset state. */
1205 info->num_needed_small_buffers = 0;
1206 info->num_needed_large_buffers = 0;
1207
1208 /* Cancel egress timer. */
1209 del_timer(&info->egress_timer);
1210 info->egress_timer_scheduled = false;
1211
1212 netif_napi_del(&info->napi);
1213
1214 /* Now we are unregistered. */
1215 info->registered = false;
1216}
1217
1218
1219/*
1220 * Helper function for "tile_net_stop()".
1221 *
1222 * Also used to handle registration failure in "tile_net_open_inner()",
1223 * when "fully_opened" is known to be false, and the various extra
1224 * steps in "tile_net_stop()" are not necessary. ISSUE: It might be
1225 * simpler if we could just call "tile_net_stop()" anyway.
1226 */
1227static void tile_net_stop_aux(struct net_device *dev)
1228{
1229 struct tile_net_priv *priv = netdev_priv(dev);
1230
1231 int dummy = 0;
1232
1233 /* Unregister all tiles, so LIPP will stop delivering packets. */
1234 on_each_cpu(tile_net_unregister, (void *)dev, 1);
1235
1236 /* Stop LIPP/LEPP. */
1237 if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1238 sizeof(dummy), NETIO_IPP_STOP_SHIM_OFF) < 0)
1239 panic("Failed to stop LIPP/LEPP!\n");
1240
1241 priv->partly_opened = 0;
1242}
1243
1244
1245/*
1246 * Disable ingress interrupts for the given device on the current cpu.
1247 */
1248static void tile_net_disable_intr(void *dev_ptr)
1249{
1250 struct net_device *dev = (struct net_device *)dev_ptr;
1251 struct tile_net_priv *priv = netdev_priv(dev);
1252 int my_cpu = smp_processor_id();
1253 struct tile_net_cpu *info = priv->cpu[my_cpu];
1254
1255 /* Disable hypervisor interrupt. */
1256 disable_percpu_irq(priv->intr_id);
1257
1258 /* Disable NAPI if needed. */
1259 if (info != NULL && info->napi_enabled) {
1260 napi_disable(&info->napi);
1261 info->napi_enabled = false;
1262 }
1263}
1264
1265
1266/*
1267 * Enable ingress interrupts for the given device on the current cpu.
1268 */
1269static void tile_net_enable_intr(void *dev_ptr)
1270{
1271 struct net_device *dev = (struct net_device *)dev_ptr;
1272 struct tile_net_priv *priv = netdev_priv(dev);
1273 int my_cpu = smp_processor_id();
1274 struct tile_net_cpu *info = priv->cpu[my_cpu];
1275
1276 /* Enable hypervisor interrupt. */
1277 enable_percpu_irq(priv->intr_id);
1278
1279 /* Enable NAPI. */
1280 napi_enable(&info->napi);
1281 info->napi_enabled = true;
1282}
1283
1284
1285/*
1286 * tile_net_open_inner does most of the work of bringing up the interface.
1287 * It's called from tile_net_open(), and also from tile_net_retry_open().
1288 * The return value is 0 if the interface was brought up, < 0 if
1289 * tile_net_open() should return the return value as an error, and > 0 if
1290 * tile_net_open() should return success and schedule a work item to
1291 * periodically retry the bringup.
1292 */
1293static int tile_net_open_inner(struct net_device *dev)
1294{
1295 struct tile_net_priv *priv = netdev_priv(dev);
1296 int my_cpu = smp_processor_id();
1297 struct tile_net_cpu *info;
1298 struct tile_netio_queue *queue;
1299 unsigned int irq;
1300 int i;
1301
1302 /*
1303 * First try to register just on the local CPU, and handle any
1304 * semi-expected "link down" failure specially. Note that we
1305 * do NOT call "tile_net_stop_aux()", unlike below.
1306 */
1307 tile_net_register(dev);
1308 info = priv->cpu[my_cpu];
1309 if (!info->registered) {
1310 if (info->link_down)
1311 return 1;
1312 return -EAGAIN;
1313 }
1314
1315 /*
1316 * Now register everywhere else. If any registration fails,
1317 * even for "link down" (which might not be possible), we
1318 * clean up using "tile_net_stop_aux()".
1319 */
1320 smp_call_function(tile_net_register, (void *)dev, 1);
1321 for_each_online_cpu(i) {
1322 if (!priv->cpu[i]->registered) {
1323 tile_net_stop_aux(dev);
1324 return -EAGAIN;
1325 }
1326 }
1327
1328 queue = &info->queue;
1329
1330 /*
1331 * Set the device intr bit mask.
1332 * The tile_net_register above sets per tile __intr_id.
1333 */
1334 priv->intr_id = queue->__system_part->__intr_id;
1335 BUG_ON(!priv->intr_id);
1336
1337 /*
1338 * Register the device interrupt handler.
1339 * The __ffs() function returns the index into the interrupt handler
1340 * table from the interrupt bit mask which should have one bit
1341 * and one bit only set.
1342 */
1343 irq = __ffs(priv->intr_id);
1344 tile_irq_activate(irq, TILE_IRQ_PERCPU);
1345 BUG_ON(request_irq(irq, tile_net_handle_ingress_interrupt,
1346 0, dev->name, (void *)dev) != 0);
1347
1348 /* ISSUE: How could "priv->fully_opened" ever be "true" here? */
1349
1350 if (!priv->fully_opened) {
1351
1352 int dummy = 0;
1353
1354 /* Allocate initial buffers. */
1355
1356 int max_buffers =
1357 priv->network_cpus_count * priv->network_cpus_credits;
1358
1359 info->num_needed_small_buffers =
1360 min(LIPP_SMALL_BUFFERS, max_buffers);
1361
1362 info->num_needed_large_buffers =
1363 min(LIPP_LARGE_BUFFERS, max_buffers);
1364
1365 tile_net_provide_needed_buffers(info);
1366
1367 if (info->num_needed_small_buffers != 0 ||
1368 info->num_needed_large_buffers != 0)
1369 panic("Insufficient memory for buffer stack!");
1370
1371 /* Start LIPP/LEPP and activate "ingress" at the shim. */
1372 if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1373 sizeof(dummy), NETIO_IPP_INPUT_INIT_OFF) < 0)
1374 panic("Failed to activate the LIPP Shim!\n");
1375
1376 priv->fully_opened = 1;
1377 }
1378
1379 /* On each tile, enable the hypervisor to trigger interrupts. */
1380 /* ISSUE: Do this before starting LIPP/LEPP? */
1381 on_each_cpu(tile_net_enable_intr, (void *)dev, 1);
1382
1383 /* Start our transmit queue. */
1384 netif_start_queue(dev);
1385
1386 return 0;
1387}
1388
1389
1390/*
1391 * Called periodically to retry bringing up the NetIO interface,
1392 * if it doesn't come up cleanly during tile_net_open().
1393 */
1394static void tile_net_open_retry(struct work_struct *w)
1395{
1396 struct delayed_work *dw =
1397 container_of(w, struct delayed_work, work);
1398
1399 struct tile_net_priv *priv =
1400 container_of(dw, struct tile_net_priv, retry_work);
1401
1402 /*
1403 * Try to bring the NetIO interface up. If it fails, reschedule
1404 * ourselves to try again later; otherwise, tell Linux we now have
1405 * a working link. ISSUE: What if the return value is negative?
1406 */
1407 if (tile_net_open_inner(priv->dev))
1408 schedule_delayed_work_on(singlethread_cpu, &priv->retry_work,
1409 TILE_NET_RETRY_INTERVAL);
1410 else
1411 netif_carrier_on(priv->dev);
1412}
1413
1414
1415/*
1416 * Called when a network interface is made active.
1417 *
1418 * Returns 0 on success, negative value on failure.
1419 *
1420 * The open entry point is called when a network interface is made
1421 * active by the system (IFF_UP). At this point all resources needed
1422 * for transmit and receive operations are allocated, the interrupt
1423 * handler is registered with the OS, the watchdog timer is started,
1424 * and the stack is notified that the interface is ready.
1425 *
1426 * If the actual link is not available yet, then we tell Linux that
1427 * we have no carrier, and we keep checking until the link comes up.
1428 */
1429static int tile_net_open(struct net_device *dev)
1430{
1431 int ret = 0;
1432 struct tile_net_priv *priv = netdev_priv(dev);
1433
1434 /*
1435 * We rely on priv->partly_opened to tell us if this is the
1436 * first time this interface is being brought up. If it is
1437 * set, the IPP was already initialized and should not be
1438 * initialized again.
1439 */
1440 if (!priv->partly_opened) {
1441
1442 int count;
1443 int credits;
1444
1445 /* Initialize LIPP/LEPP, and start the Shim. */
1446 ret = tile_net_open_aux(dev);
1447 if (ret < 0) {
1448 pr_err("tile_net_open_aux failed: %d\n", ret);
1449 return ret;
1450 }
1451
1452 /* Analyze the network cpus. */
1453
1454 if (network_cpus_used)
1455 cpumask_copy(&priv->network_cpus_map,
1456 &network_cpus_map);
1457 else
1458 cpumask_copy(&priv->network_cpus_map, cpu_online_mask);
1459
1460
1461 count = cpumask_weight(&priv->network_cpus_map);
1462
1463 /* Limit credits to available buffers, and apply min. */
1464 credits = max(16, (LIPP_LARGE_BUFFERS / count) & ~1);
1465
1466 /* Apply "GBE" max limit. */
1467 /* ISSUE: Use higher limit for XGBE? */
1468 credits = min(NETIO_MAX_RECEIVE_PKTS, credits);
1469
1470 priv->network_cpus_count = count;
1471 priv->network_cpus_credits = credits;
1472
1473#ifdef TILE_NET_DEBUG
1474 pr_info("Using %d network cpus, with %d credits each\n",
1475 priv->network_cpus_count, priv->network_cpus_credits);
1476#endif
1477
1478 priv->partly_opened = 1;
1479 }
1480
1481 /*
1482 * Attempt to bring up the link.
1483 */
1484 ret = tile_net_open_inner(dev);
1485 if (ret <= 0) {
1486 if (ret == 0)
1487 netif_carrier_on(dev);
1488 return ret;
1489 }
1490
1491 /*
1492 * We were unable to bring up the NetIO interface, but we want to
1493 * try again in a little bit. Tell Linux that we have no carrier
1494 * so it doesn't try to use the interface before the link comes up
1495 * and then remember to try again later.
1496 */
1497 netif_carrier_off(dev);
1498 schedule_delayed_work_on(singlethread_cpu, &priv->retry_work,
1499 TILE_NET_RETRY_INTERVAL);
1500
1501 return 0;
1502}
1503
1504
1505/*
1506 * Disables a network interface.
1507 *
1508 * Returns 0, this is not allowed to fail.
1509 *
1510 * The close entry point is called when an interface is de-activated
1511 * by the OS. The hardware is still under the drivers control, but
1512 * needs to be disabled. A global MAC reset is issued to stop the
1513 * hardware, and all transmit and receive resources are freed.
1514 *
1515 * ISSUE: Can this can be called while "tile_net_poll()" is running?
1516 */
1517static int tile_net_stop(struct net_device *dev)
1518{
1519 struct tile_net_priv *priv = netdev_priv(dev);
1520
1521 bool pending = true;
1522
1523 PDEBUG("tile_net_stop()\n");
1524
1525 /* ISSUE: Only needed if not yet fully open. */
1526 cancel_delayed_work_sync(&priv->retry_work);
1527
1528 /* Can't transmit any more. */
1529 netif_stop_queue(dev);
1530
1531 /*
1532 * Disable hypervisor interrupts on each tile.
1533 */
1534 on_each_cpu(tile_net_disable_intr, (void *)dev, 1);
1535
1536 /*
1537 * Unregister the interrupt handler.
1538 * The __ffs() function returns the index into the interrupt handler
1539 * table from the interrupt bit mask which should have one bit
1540 * and one bit only set.
1541 */
1542 if (priv->intr_id)
1543 free_irq(__ffs(priv->intr_id), dev);
1544
1545 /*
1546 * Drain all the LIPP buffers.
1547 */
1548
1549 while (true) {
1550 int buffer;
1551
1552 /* NOTE: This should never fail. */
1553 if (hv_dev_pread(priv->hv_devhdl, 0, (HV_VirtAddr)&buffer,
1554 sizeof(buffer), NETIO_IPP_DRAIN_OFF) < 0)
1555 break;
1556
1557 /* Stop when done. */
1558 if (buffer == 0)
1559 break;
1560
1561 {
1562 /* Convert "linux_buffer_t" to "va". */
1563 void *va = __va((phys_addr_t)(buffer >> 1) << 7);
1564
1565 /* Acquire the associated "skb". */
1566 struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
1567 struct sk_buff *skb = *skb_ptr;
1568
1569 kfree_skb(skb);
1570 }
1571 }
1572
1573 /* Stop LIPP/LEPP. */
1574 tile_net_stop_aux(dev);
1575
1576
1577 priv->fully_opened = 0;
1578
1579
1580 /*
1581 * XXX: ISSUE: It appears that, in practice anyway, by the
1582 * time we get here, there are no pending completions.
1583 */
1584 while (pending) {
1585
1586 struct sk_buff *olds[32];
1587 unsigned int wanted = 32;
1588 unsigned int i, nolds = 0;
1589
1590 nolds = tile_net_lepp_grab_comps(dev, olds,
1591 wanted, &pending);
1592
1593 /* ISSUE: We have never actually seen this debug spew. */
1594 if (nolds != 0)
1595 pr_info("During tile_net_stop(), grabbed %d comps.\n",
1596 nolds);
1597
1598 for (i = 0; i < nolds; i++)
1599 kfree_skb(olds[i]);
1600 }
1601
1602
1603 /* Wipe the EPP queue. */
1604 memset(priv->epp_queue, 0, sizeof(lepp_queue_t));
1605
1606 /* Evict the EPP queue. */
1607 finv_buffer(priv->epp_queue, PAGE_SIZE);
1608
1609 return 0;
1610}
1611
1612
1613/*
1614 * Prepare the "frags" info for the resulting LEPP command.
1615 *
1616 * If needed, flush the memory used by the frags.
1617 */
1618static unsigned int tile_net_tx_frags(lepp_frag_t *frags,
1619 struct sk_buff *skb,
1620 void *b_data, unsigned int b_len)
1621{
1622 unsigned int i, n = 0;
1623
1624 struct skb_shared_info *sh = skb_shinfo(skb);
1625
1626 phys_addr_t cpa;
1627
1628 if (b_len != 0) {
1629
1630 if (!hash_default)
1631 finv_buffer_remote(b_data, b_len);
1632
1633 cpa = __pa(b_data);
1634 frags[n].cpa_lo = cpa;
1635 frags[n].cpa_hi = cpa >> 32;
1636 frags[n].length = b_len;
1637 frags[n].hash_for_home = hash_default;
1638 n++;
1639 }
1640
1641 for (i = 0; i < sh->nr_frags; i++) {
1642
1643 skb_frag_t *f = &sh->frags[i];
1644 unsigned long pfn = page_to_pfn(f->page);
1645
1646 /* FIXME: Compute "hash_for_home" properly. */
1647 /* ISSUE: The hypervisor checks CHIP_HAS_REV1_DMA_PACKETS(). */
1648 int hash_for_home = hash_default;
1649
1650 /* FIXME: Hmmm. */
1651 if (!hash_default) {
1652 void *va = pfn_to_kaddr(pfn) + f->page_offset;
1653 BUG_ON(PageHighMem(f->page));
1654 finv_buffer_remote(va, f->size);
1655 }
1656
1657 cpa = ((phys_addr_t)pfn << PAGE_SHIFT) + f->page_offset;
1658 frags[n].cpa_lo = cpa;
1659 frags[n].cpa_hi = cpa >> 32;
1660 frags[n].length = f->size;
1661 frags[n].hash_for_home = hash_for_home;
1662 n++;
1663 }
1664
1665 return n;
1666}
1667
1668
1669/*
1670 * This function takes "skb", consisting of a header template and a
1671 * payload, and hands it to LEPP, to emit as one or more segments,
1672 * each consisting of a possibly modified header, plus a piece of the
1673 * payload, via a process known as "tcp segmentation offload".
1674 *
1675 * Usually, "data" will contain the header template, of size "sh_len",
1676 * and "sh->frags" will contain "skb->data_len" bytes of payload, and
1677 * there will be "sh->gso_segs" segments.
1678 *
1679 * Sometimes, if "sendfile()" requires copying, we will be called with
1680 * "data" containing the header and payload, with "frags" being empty.
1681 *
1682 * In theory, "sh->nr_frags" could be 3, but in practice, it seems
1683 * that this will never actually happen.
1684 *
1685 * See "emulate_large_send_offload()" for some reference code, which
1686 * does not handle checksumming.
1687 *
1688 * ISSUE: How do we make sure that high memory DMA does not migrate?
1689 */
1690static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
1691{
1692 struct tile_net_priv *priv = netdev_priv(dev);
1693 int my_cpu = smp_processor_id();
1694 struct tile_net_cpu *info = priv->cpu[my_cpu];
1695 struct tile_net_stats_t *stats = &info->stats;
1696
1697 struct skb_shared_info *sh = skb_shinfo(skb);
1698
1699 unsigned char *data = skb->data;
1700
1701 /* The ip header follows the ethernet header. */
1702 struct iphdr *ih = ip_hdr(skb);
1703 unsigned int ih_len = ih->ihl * 4;
1704
1705 /* Note that "nh == ih", by definition. */
1706 unsigned char *nh = skb_network_header(skb);
1707 unsigned int eh_len = nh - data;
1708
1709 /* The tcp header follows the ip header. */
1710 struct tcphdr *th = (struct tcphdr *)(nh + ih_len);
1711 unsigned int th_len = th->doff * 4;
1712
1713 /* The total number of header bytes. */
1714 /* NOTE: This may be less than skb_headlen(skb). */
1715 unsigned int sh_len = eh_len + ih_len + th_len;
1716
1717 /* The number of payload bytes at "skb->data + sh_len". */
1718 /* This is non-zero for sendfile() without HIGHDMA. */
1719 unsigned int b_len = skb_headlen(skb) - sh_len;
1720
1721 /* The total number of payload bytes. */
1722 unsigned int d_len = b_len + skb->data_len;
1723
1724 /* The maximum payload size. */
1725 unsigned int p_len = sh->gso_size;
1726
1727 /* The total number of segments. */
1728 unsigned int num_segs = sh->gso_segs;
1729
1730 /* The temporary copy of the command. */
1731 u32 cmd_body[(LEPP_MAX_CMD_SIZE + 3) / 4];
1732 lepp_tso_cmd_t *cmd = (lepp_tso_cmd_t *)cmd_body;
1733
1734 /* Analyze the "frags". */
1735 unsigned int num_frags =
1736 tile_net_tx_frags(cmd->frags, skb, data + sh_len, b_len);
1737
1738 /* The size of the command, including frags and header. */
1739 size_t cmd_size = LEPP_TSO_CMD_SIZE(num_frags, sh_len);
1740
1741 /* The command header. */
1742 lepp_tso_cmd_t cmd_init = {
1743 .tso = true,
1744 .header_size = sh_len,
1745 .ip_offset = eh_len,
1746 .tcp_offset = eh_len + ih_len,
1747 .payload_size = p_len,
1748 .num_frags = num_frags,
1749 };
1750
1751 unsigned long irqflags;
1752
1753 lepp_queue_t *eq = priv->epp_queue;
1754
1755 struct sk_buff *olds[4];
1756 unsigned int wanted = 4;
1757 unsigned int i, nolds = 0;
1758
1759 unsigned int cmd_head, cmd_tail, cmd_next;
1760 unsigned int comp_tail;
1761
1762 unsigned int free_slots;
1763
1764
1765 /* Paranoia. */
1766 BUG_ON(skb->protocol != htons(ETH_P_IP));
1767 BUG_ON(ih->protocol != IPPROTO_TCP);
1768 BUG_ON(skb->ip_summed != CHECKSUM_PARTIAL);
1769 BUG_ON(num_frags > LEPP_MAX_FRAGS);
1770 /*--BUG_ON(num_segs != (d_len + (p_len - 1)) / p_len); */
1771 BUG_ON(num_segs <= 1);
1772
1773
1774 /* Finish preparing the command. */
1775
1776 /* Copy the command header. */
1777 *cmd = cmd_init;
1778
1779 /* Copy the "header". */
1780 memcpy(&cmd->frags[num_frags], data, sh_len);
1781
1782
1783 /* Prefetch and wait, to minimize time spent holding the spinlock. */
1784 prefetch_L1(&eq->comp_tail);
1785 prefetch_L1(&eq->cmd_tail);
1786 mb();
1787
1788
1789 /* Enqueue the command. */
1790
1791 spin_lock_irqsave(&priv->cmd_lock, irqflags);
1792
1793 /*
1794 * Handle completions if needed to make room.
1795 * HACK: Spin until there is sufficient room.
1796 */
1797 free_slots = lepp_num_free_comp_slots(eq);
1798 if (free_slots < 1) {
1799spin:
1800 nolds += tile_net_lepp_grab_comps(dev, olds + nolds,
1801 wanted - nolds, NULL);
1802 if (lepp_num_free_comp_slots(eq) < 1)
1803 goto spin;
1804 }
1805
1806 cmd_head = eq->cmd_head;
1807 cmd_tail = eq->cmd_tail;
1808
1809 /* NOTE: The "gotos" below are untested. */
1810
1811 /* Prepare to advance, detecting full queue. */
1812 cmd_next = cmd_tail + cmd_size;
1813 if (cmd_tail < cmd_head && cmd_next >= cmd_head)
1814 goto spin;
1815 if (cmd_next > LEPP_CMD_LIMIT) {
1816 cmd_next = 0;
1817 if (cmd_next == cmd_head)
1818 goto spin;
1819 }
1820
1821 /* Copy the command. */
1822 memcpy(&eq->cmds[cmd_tail], cmd, cmd_size);
1823
1824 /* Advance. */
1825 cmd_tail = cmd_next;
1826
1827 /* Record "skb" for eventual freeing. */
1828 comp_tail = eq->comp_tail;
1829 eq->comps[comp_tail] = skb;
1830 LEPP_QINC(comp_tail);
1831 eq->comp_tail = comp_tail;
1832
1833 /* Flush before allowing LEPP to handle the command. */
1834 __insn_mf();
1835
1836 eq->cmd_tail = cmd_tail;
1837
1838 spin_unlock_irqrestore(&priv->cmd_lock, irqflags);
1839
1840 if (nolds == 0)
1841 nolds = tile_net_lepp_grab_comps(dev, olds, wanted, NULL);
1842
1843 /* Handle completions. */
1844 for (i = 0; i < nolds; i++)
1845 kfree_skb(olds[i]);
1846
1847 /* Update stats. */
1848 stats->tx_packets += num_segs;
1849 stats->tx_bytes += (num_segs * sh_len) + d_len;
1850
1851 /* Make sure the egress timer is scheduled. */
1852 tile_net_schedule_egress_timer(info);
1853
1854 return NETDEV_TX_OK;
1855}
1856
1857
1858/*
1859 * Transmit a packet (called by the kernel via "hard_start_xmit" hook).
1860 */
1861static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
1862{
1863 struct tile_net_priv *priv = netdev_priv(dev);
1864 int my_cpu = smp_processor_id();
1865 struct tile_net_cpu *info = priv->cpu[my_cpu];
1866 struct tile_net_stats_t *stats = &info->stats;
1867
1868 unsigned long irqflags;
1869
1870 struct skb_shared_info *sh = skb_shinfo(skb);
1871
1872 unsigned int len = skb->len;
1873 unsigned char *data = skb->data;
1874
1875 unsigned int csum_start = skb->csum_start - skb_headroom(skb);
1876
1877 lepp_frag_t frags[LEPP_MAX_FRAGS];
1878
1879 unsigned int num_frags;
1880
1881 lepp_queue_t *eq = priv->epp_queue;
1882
1883 struct sk_buff *olds[4];
1884 unsigned int wanted = 4;
1885 unsigned int i, nolds = 0;
1886
1887 unsigned int cmd_size = sizeof(lepp_cmd_t);
1888
1889 unsigned int cmd_head, cmd_tail, cmd_next;
1890 unsigned int comp_tail;
1891
1892 lepp_cmd_t cmds[LEPP_MAX_FRAGS];
1893
1894 unsigned int free_slots;
1895
1896
1897 /*
1898 * This is paranoia, since we think that if the link doesn't come
1899 * up, telling Linux we have no carrier will keep it from trying
1900 * to transmit. If it does, though, we can't execute this routine,
1901 * since data structures we depend on aren't set up yet.
1902 */
1903 if (!info->registered)
1904 return NETDEV_TX_BUSY;
1905
1906
1907 /* Save the timestamp. */
1908 dev->trans_start = jiffies;
1909
1910
1911#ifdef TILE_NET_PARANOIA
1912#if CHIP_HAS_CBOX_HOME_MAP()
1913 if (hash_default) {
1914 HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)data);
1915 if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
1916 panic("Non-coherent egress buffer!");
1917 }
1918#endif
1919#endif
1920
1921
1922#ifdef TILE_NET_DUMP_PACKETS
1923 /* ISSUE: Does not dump the "frags". */
1924 dump_packet(data, skb_headlen(skb), "tx");
1925#endif /* TILE_NET_DUMP_PACKETS */
1926
1927
1928 if (sh->gso_size != 0)
1929 return tile_net_tx_tso(skb, dev);
1930
1931
1932 /* Prepare the commands. */
1933
1934 num_frags = tile_net_tx_frags(frags, skb, data, skb_headlen(skb));
1935
1936 for (i = 0; i < num_frags; i++) {
1937
1938 bool final = (i == num_frags - 1);
1939
1940 lepp_cmd_t cmd = {
1941 .cpa_lo = frags[i].cpa_lo,
1942 .cpa_hi = frags[i].cpa_hi,
1943 .length = frags[i].length,
1944 .hash_for_home = frags[i].hash_for_home,
1945 .send_completion = final,
1946 .end_of_packet = final
1947 };
1948
1949 if (i == 0 && skb->ip_summed == CHECKSUM_PARTIAL) {
1950 cmd.compute_checksum = 1;
1951 cmd.checksum_data.bits.start_byte = csum_start;
1952 cmd.checksum_data.bits.count = len - csum_start;
1953 cmd.checksum_data.bits.destination_byte =
1954 csum_start + skb->csum_offset;
1955 }
1956
1957 cmds[i] = cmd;
1958 }
1959
1960
1961 /* Prefetch and wait, to minimize time spent holding the spinlock. */
1962 prefetch_L1(&eq->comp_tail);
1963 prefetch_L1(&eq->cmd_tail);
1964 mb();
1965
1966
1967 /* Enqueue the commands. */
1968
1969 spin_lock_irqsave(&priv->cmd_lock, irqflags);
1970
1971 /*
1972 * Handle completions if needed to make room.
1973 * HACK: Spin until there is sufficient room.
1974 */
1975 free_slots = lepp_num_free_comp_slots(eq);
1976 if (free_slots < 1) {
1977spin:
1978 nolds += tile_net_lepp_grab_comps(dev, olds + nolds,
1979 wanted - nolds, NULL);
1980 if (lepp_num_free_comp_slots(eq) < 1)
1981 goto spin;
1982 }
1983
1984 cmd_head = eq->cmd_head;
1985 cmd_tail = eq->cmd_tail;
1986
1987 /* NOTE: The "gotos" below are untested. */
1988
1989 /* Copy the commands, or fail. */
1990 for (i = 0; i < num_frags; i++) {
1991
1992 /* Prepare to advance, detecting full queue. */
1993 cmd_next = cmd_tail + cmd_size;
1994 if (cmd_tail < cmd_head && cmd_next >= cmd_head)
1995 goto spin;
1996 if (cmd_next > LEPP_CMD_LIMIT) {
1997 cmd_next = 0;
1998 if (cmd_next == cmd_head)
1999 goto spin;
2000 }
2001
2002 /* Copy the command. */
2003 *(lepp_cmd_t *)&eq->cmds[cmd_tail] = cmds[i];
2004
2005 /* Advance. */
2006 cmd_tail = cmd_next;
2007 }
2008
2009 /* Record "skb" for eventual freeing. */
2010 comp_tail = eq->comp_tail;
2011 eq->comps[comp_tail] = skb;
2012 LEPP_QINC(comp_tail);
2013 eq->comp_tail = comp_tail;
2014
2015 /* Flush before allowing LEPP to handle the command. */
2016 __insn_mf();
2017
2018 eq->cmd_tail = cmd_tail;
2019
2020 spin_unlock_irqrestore(&priv->cmd_lock, irqflags);
2021
2022 if (nolds == 0)
2023 nolds = tile_net_lepp_grab_comps(dev, olds, wanted, NULL);
2024
2025 /* Handle completions. */
2026 for (i = 0; i < nolds; i++)
2027 kfree_skb(olds[i]);
2028
2029 /* HACK: Track "expanded" size for short packets (e.g. 42 < 60). */
2030 stats->tx_packets++;
2031 stats->tx_bytes += ((len >= ETH_ZLEN) ? len : ETH_ZLEN);
2032
2033 /* Make sure the egress timer is scheduled. */
2034 tile_net_schedule_egress_timer(info);
2035
2036 return NETDEV_TX_OK;
2037}
2038
2039
2040/*
2041 * Deal with a transmit timeout.
2042 */
2043static void tile_net_tx_timeout(struct net_device *dev)
2044{
2045 PDEBUG("tile_net_tx_timeout()\n");
2046 PDEBUG("Transmit timeout at %ld, latency %ld\n", jiffies,
2047 jiffies - dev->trans_start);
2048
2049 /* XXX: ISSUE: This doesn't seem useful for us. */
2050 netif_wake_queue(dev);
2051}
2052
2053
2054/*
2055 * Ioctl commands.
2056 */
2057static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
2058{
2059 return -EOPNOTSUPP;
2060}
2061
2062
2063/*
2064 * Get System Network Statistics.
2065 *
2066 * Returns the address of the device statistics structure.
2067 */
2068static struct net_device_stats *tile_net_get_stats(struct net_device *dev)
2069{
2070 struct tile_net_priv *priv = netdev_priv(dev);
2071 u32 rx_packets = 0;
2072 u32 tx_packets = 0;
2073 u32 rx_bytes = 0;
2074 u32 tx_bytes = 0;
2075 int i;
2076
2077 for_each_online_cpu(i) {
2078 if (priv->cpu[i]) {
2079 rx_packets += priv->cpu[i]->stats.rx_packets;
2080 rx_bytes += priv->cpu[i]->stats.rx_bytes;
2081 tx_packets += priv->cpu[i]->stats.tx_packets;
2082 tx_bytes += priv->cpu[i]->stats.tx_bytes;
2083 }
2084 }
2085
2086 priv->stats.rx_packets = rx_packets;
2087 priv->stats.rx_bytes = rx_bytes;
2088 priv->stats.tx_packets = tx_packets;
2089 priv->stats.tx_bytes = tx_bytes;
2090
2091 return &priv->stats;
2092}
2093
2094
2095/*
2096 * Change the "mtu".
2097 *
2098 * The "change_mtu" method is usually not needed.
2099 * If you need it, it must be like this.
2100 */
2101static int tile_net_change_mtu(struct net_device *dev, int new_mtu)
2102{
2103 PDEBUG("tile_net_change_mtu()\n");
2104
2105 /* Check ranges. */
2106 if ((new_mtu < 68) || (new_mtu > 1500))
2107 return -EINVAL;
2108
2109 /* Accept the value. */
2110 dev->mtu = new_mtu;
2111
2112 return 0;
2113}
2114
2115
2116/*
2117 * Change the Ethernet Address of the NIC.
2118 *
2119 * The hypervisor driver does not support changing MAC address. However,
2120 * the IPP does not do anything with the MAC address, so the address which
2121 * gets used on outgoing packets, and which is accepted on incoming packets,
2122 * is completely up to the NetIO program or kernel driver which is actually
2123 * handling them.
2124 *
2125 * Returns 0 on success, negative on failure.
2126 */
2127static int tile_net_set_mac_address(struct net_device *dev, void *p)
2128{
2129 struct sockaddr *addr = p;
2130
2131 if (!is_valid_ether_addr(addr->sa_data))
2132 return -EINVAL;
2133
2134 /* ISSUE: Note that "dev_addr" is now a pointer. */
2135 memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
2136
2137 return 0;
2138}
2139
2140
2141/*
2142 * Obtain the MAC address from the hypervisor.
2143 * This must be done before opening the device.
2144 */
2145static int tile_net_get_mac(struct net_device *dev)
2146{
2147 struct tile_net_priv *priv = netdev_priv(dev);
2148
2149 char hv_dev_name[32];
2150 int len;
2151
2152 __netio_getset_offset_t offset = { .word = NETIO_IPP_PARAM_OFF };
2153
2154 int ret;
2155
2156 /* For example, "xgbe0". */
2157 strcpy(hv_dev_name, dev->name);
2158 len = strlen(hv_dev_name);
2159
2160 /* For example, "xgbe/0". */
2161 hv_dev_name[len] = hv_dev_name[len - 1];
2162 hv_dev_name[len - 1] = '/';
2163 len++;
2164
2165 /* For example, "xgbe/0/native_hash". */
2166 strcpy(hv_dev_name + len, hash_default ? "/native_hash" : "/native");
2167
2168 /* Get the hypervisor handle for this device. */
2169 priv->hv_devhdl = hv_dev_open((HV_VirtAddr)hv_dev_name, 0);
2170 PDEBUG("hv_dev_open(%s) returned %d %p\n",
2171 hv_dev_name, priv->hv_devhdl, &priv->hv_devhdl);
2172 if (priv->hv_devhdl < 0) {
2173 if (priv->hv_devhdl == HV_ENODEV)
2174 printk(KERN_DEBUG "Ignoring unconfigured device %s\n",
2175 hv_dev_name);
2176 else
2177 printk(KERN_DEBUG "hv_dev_open(%s) returned %d\n",
2178 hv_dev_name, priv->hv_devhdl);
2179 return -1;
2180 }
2181
2182 /*
2183 * Read the hardware address from the hypervisor.
2184 * ISSUE: Note that "dev_addr" is now a pointer.
2185 */
2186 offset.bits.class = NETIO_PARAM;
2187 offset.bits.addr = NETIO_PARAM_MAC;
2188 ret = hv_dev_pread(priv->hv_devhdl, 0,
2189 (HV_VirtAddr)dev->dev_addr, dev->addr_len,
2190 offset.word);
2191 PDEBUG("hv_dev_pread(NETIO_PARAM_MAC) returned %d\n", ret);
2192 if (ret <= 0) {
2193 printk(KERN_DEBUG "hv_dev_pread(NETIO_PARAM_MAC) %s failed\n",
2194 dev->name);
2195 /*
2196 * Since the device is configured by the hypervisor but we
2197 * can't get its MAC address, we are most likely running
2198 * the simulator, so let's generate a random MAC address.
2199 */
2200 random_ether_addr(dev->dev_addr);
2201 }
2202
2203 return 0;
2204}
2205
2206
2207static struct net_device_ops tile_net_ops = {
2208 .ndo_open = tile_net_open,
2209 .ndo_stop = tile_net_stop,
2210 .ndo_start_xmit = tile_net_tx,
2211 .ndo_do_ioctl = tile_net_ioctl,
2212 .ndo_get_stats = tile_net_get_stats,
2213 .ndo_change_mtu = tile_net_change_mtu,
2214 .ndo_tx_timeout = tile_net_tx_timeout,
2215 .ndo_set_mac_address = tile_net_set_mac_address
2216};
2217
2218
2219/*
2220 * The setup function.
2221 *
2222 * This uses ether_setup() to assign various fields in dev, including
2223 * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields.
2224 */
2225static void tile_net_setup(struct net_device *dev)
2226{
2227 PDEBUG("tile_net_setup()\n");
2228
2229 ether_setup(dev);
2230
2231 dev->netdev_ops = &tile_net_ops;
2232
2233 dev->watchdog_timeo = TILE_NET_TIMEOUT;
2234
2235 /* We want lockless xmit. */
2236 dev->features |= NETIF_F_LLTX;
2237
2238 /* We support hardware tx checksums. */
2239 dev->features |= NETIF_F_HW_CSUM;
2240
2241 /* We support scatter/gather. */
2242 dev->features |= NETIF_F_SG;
2243
2244 /* We support TSO. */
2245 dev->features |= NETIF_F_TSO;
2246
2247#ifdef TILE_NET_GSO
2248 /* We support GSO. */
2249 dev->features |= NETIF_F_GSO;
2250#endif
2251
2252 if (hash_default)
2253 dev->features |= NETIF_F_HIGHDMA;
2254
2255 /* ISSUE: We should support NETIF_F_UFO. */
2256
2257 dev->tx_queue_len = TILE_NET_TX_QUEUE_LEN;
2258
2259 dev->mtu = TILE_NET_MTU;
2260}
2261
2262
2263/*
2264 * Allocate the device structure, register the device, and obtain the
2265 * MAC address from the hypervisor.
2266 */
2267static struct net_device *tile_net_dev_init(const char *name)
2268{
2269 int ret;
2270 struct net_device *dev;
2271 struct tile_net_priv *priv;
2272 struct page *page;
2273
2274 /*
2275 * Allocate the device structure. This allocates "priv", calls
2276 * tile_net_setup(), and saves "name". Normally, "name" is a
2277 * template, instantiated by register_netdev(), but not for us.
2278 */
2279 dev = alloc_netdev(sizeof(*priv), name, tile_net_setup);
2280 if (!dev) {
2281 pr_err("alloc_netdev(%s) failed\n", name);
2282 return NULL;
2283 }
2284
2285 priv = netdev_priv(dev);
2286
2287 /* Initialize "priv". */
2288
2289 memset(priv, 0, sizeof(*priv));
2290
2291 /* Save "dev" for "tile_net_open_retry()". */
2292 priv->dev = dev;
2293
2294 INIT_DELAYED_WORK(&priv->retry_work, tile_net_open_retry);
2295
2296 spin_lock_init(&priv->cmd_lock);
2297 spin_lock_init(&priv->comp_lock);
2298
2299 /* Allocate "epp_queue". */
2300 BUG_ON(get_order(sizeof(lepp_queue_t)) != 0);
2301 page = alloc_pages(GFP_KERNEL | __GFP_ZERO, 0);
2302 if (!page) {
2303 free_netdev(dev);
2304 return NULL;
2305 }
2306 priv->epp_queue = page_address(page);
2307
2308 /* Register the network device. */
2309 ret = register_netdev(dev);
2310 if (ret) {
2311 pr_err("register_netdev %s failed %d\n", dev->name, ret);
2312 free_page((unsigned long)priv->epp_queue);
2313 free_netdev(dev);
2314 return NULL;
2315 }
2316
2317 /* Get the MAC address. */
2318 ret = tile_net_get_mac(dev);
2319 if (ret < 0) {
2320 unregister_netdev(dev);
2321 free_page((unsigned long)priv->epp_queue);
2322 free_netdev(dev);
2323 return NULL;
2324 }
2325
2326 return dev;
2327}
2328
2329
2330/*
2331 * Module cleanup.
2332 */
2333static void tile_net_cleanup(void)
2334{
2335 int i;
2336
2337 for (i = 0; i < TILE_NET_DEVS; i++) {
2338 if (tile_net_devs[i]) {
2339 struct net_device *dev = tile_net_devs[i];
2340 struct tile_net_priv *priv = netdev_priv(dev);
2341 unregister_netdev(dev);
2342 finv_buffer(priv->epp_queue, PAGE_SIZE);
2343 free_page((unsigned long)priv->epp_queue);
2344 free_netdev(dev);
2345 }
2346 }
2347}
2348
2349
2350/*
2351 * Module initialization.
2352 */
2353static int tile_net_init_module(void)
2354{
2355 pr_info("Tilera IPP Net Driver\n");
2356
2357 tile_net_devs[0] = tile_net_dev_init("xgbe0");
2358 tile_net_devs[1] = tile_net_dev_init("xgbe1");
2359 tile_net_devs[2] = tile_net_dev_init("gbe0");
2360 tile_net_devs[3] = tile_net_dev_init("gbe1");
2361
2362 return 0;
2363}
2364
2365
2366#ifndef MODULE
2367/*
2368 * The "network_cpus" boot argument specifies the cpus that are dedicated
2369 * to handle ingress packets.
2370 *
2371 * The parameter should be in the form "network_cpus=m-n[,x-y]", where
2372 * m, n, x, y are integer numbers that represent the cpus that can be
2373 * neither a dedicated cpu nor a dataplane cpu.
2374 */
2375static int __init network_cpus_setup(char *str)
2376{
2377 int rc = cpulist_parse_crop(str, &network_cpus_map);
2378 if (rc != 0) {
2379 pr_warning("network_cpus=%s: malformed cpu list\n",
2380 str);
2381 } else {
2382
2383 /* Remove dedicated cpus. */
2384 cpumask_and(&network_cpus_map, &network_cpus_map,
2385 cpu_possible_mask);
2386
2387
2388 if (cpumask_empty(&network_cpus_map)) {
2389 pr_warning("Ignoring network_cpus='%s'.\n",
2390 str);
2391 } else {
2392 char buf[1024];
2393 cpulist_scnprintf(buf, sizeof(buf), &network_cpus_map);
2394 pr_info("Linux network CPUs: %s\n", buf);
2395 network_cpus_used = true;
2396 }
2397 }
2398
2399 return 0;
2400}
2401__setup("network_cpus=", network_cpus_setup);
2402#endif
2403
2404
2405module_init(tile_net_init_module);
2406module_exit(tile_net_cleanup);
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-28 18:45:48 -0500