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/ipv6.h>
40#include <linux/tcp.h>
41#include <linux/net_tstamp.h>
42#include <linux/ptp_clock_kernel.h>
43
44#include <asm/checksum.h>
45#include <asm/homecache.h>
46#include <gxio/mpipe.h>
47#include <arch/sim.h>
48
49/* Default transmit lockup timeout period, in jiffies. */
50#define TILE_NET_TIMEOUT (5 * HZ)
51
52/* The maximum number of distinct channels (idesc.channel is 5 bits). */
53#define TILE_NET_CHANNELS 32
54
55/* Maximum number of idescs to handle per "poll". */
56#define TILE_NET_BATCH 128
57
58/* Maximum number of packets to handle per "poll". */
59#define TILE_NET_WEIGHT 64
60
61/* Number of entries in each iqueue. */
62#define IQUEUE_ENTRIES 512
63
64/* Number of entries in each equeue. */
65#define EQUEUE_ENTRIES 2048
66
67/* Total header bytes per equeue slot.  Must be big enough for 2 bytes
68 * of NET_IP_ALIGN alignment, plus 14 bytes (?) of L2 header, plus up to
69 * 60 bytes of actual TCP header.  We round up to align to cache lines.
70 */
71#define HEADER_BYTES 128
72
73/* Maximum completions per cpu per device (must be a power of two).
74 * ISSUE: What is the right number here?  If this is too small, then
75 * egress might block waiting for free space in a completions array.
76 * ISSUE: At the least, allocate these only for initialized echannels.
77 */
78#define TILE_NET_MAX_COMPS 64
79
80#define MAX_FRAGS (MAX_SKB_FRAGS + 1)
81
82/* The "kinds" of buffer stacks (small/large/jumbo). */
83#define MAX_KINDS 3
84
85/* Size of completions data to allocate.
86 * ISSUE: Probably more than needed since we don't use all the channels.
87 */
88#define COMPS_SIZE (TILE_NET_CHANNELS * sizeof(struct tile_net_comps))
89
90/* Size of NotifRing data to allocate. */
91#define NOTIF_RING_SIZE (IQUEUE_ENTRIES * sizeof(gxio_mpipe_idesc_t))
92
93/* Timeout to wake the per-device TX timer after we stop the queue.
94 * We don't want the timeout too short (adds overhead, and might end
95 * up causing stop/wake/stop/wake cycles) or too long (affects performance).
96 * For the 10 Gb NIC, 30 usec means roughly 30+ 1500-byte packets.
97 */
98#define TX_TIMER_DELAY_USEC 30
99
100/* Timeout to wake the per-cpu egress timer to free completions. */
101#define EGRESS_TIMER_DELAY_USEC 1000
102
103MODULE_AUTHOR("Tilera Corporation");
104MODULE_LICENSE("GPL");
105
106/* A "packet fragment" (a chunk of memory). */
107struct frag {
108	void *buf;
109	size_t length;
110};
111
112/* A single completion. */
113struct tile_net_comp {
114	/* The "complete_count" when the completion will be complete. */
115	s64 when;
116	/* The buffer to be freed when the completion is complete. */
117	struct sk_buff *skb;
118};
119
120/* The completions for a given cpu and echannel. */
121struct tile_net_comps {
122	/* The completions. */
123	struct tile_net_comp comp_queue[TILE_NET_MAX_COMPS];
124	/* The number of completions used. */
125	unsigned long comp_next;
126	/* The number of completions freed. */
127	unsigned long comp_last;
128};
129
130/* The transmit wake timer for a given cpu and echannel. */
131struct tile_net_tx_wake {
132	int tx_queue_idx;
133	struct hrtimer timer;
134	struct net_device *dev;
135};
136
137/* Info for a specific cpu. */
138struct tile_net_info {
139	/* Our cpu. */
140	int my_cpu;
141	/* A timer for handling egress completions. */
142	struct hrtimer egress_timer;
143	/* True if "egress_timer" is scheduled. */
144	bool egress_timer_scheduled;
145	struct info_mpipe {
146		/* Packet queue. */
147		gxio_mpipe_iqueue_t iqueue;
148		/* The NAPI struct. */
149		struct napi_struct napi;
150		/* Number of buffers (by kind) which must still be provided. */
151		unsigned int num_needed_buffers[MAX_KINDS];
152		/* instance id. */
153		int instance;
154		/* True if iqueue is valid. */
155		bool has_iqueue;
156		/* NAPI flags. */
157		bool napi_added;
158		bool napi_enabled;
159		/* Comps for each egress channel. */
160		struct tile_net_comps *comps_for_echannel[TILE_NET_CHANNELS];
161		/* Transmit wake timer for each egress channel. */
162		struct tile_net_tx_wake tx_wake[TILE_NET_CHANNELS];
163	} mpipe[NR_MPIPE_MAX];
164};
165
166/* Info for egress on a particular egress channel. */
167struct tile_net_egress {
168	/* The "equeue". */
169	gxio_mpipe_equeue_t *equeue;
170	/* The headers for TSO. */
171	unsigned char *headers;
172};
173
174/* Info for a specific device. */
175struct tile_net_priv {
176	/* Our network device. */
177	struct net_device *dev;
178	/* The primary link. */
179	gxio_mpipe_link_t link;
180	/* The primary channel, if open, else -1. */
181	int channel;
182	/* The "loopify" egress link, if needed. */
183	gxio_mpipe_link_t loopify_link;
184	/* The "loopify" egress channel, if open, else -1. */
185	int loopify_channel;
186	/* The egress channel (channel or loopify_channel). */
187	int echannel;
188	/* mPIPE instance, 0 or 1. */
189	int instance;
190	/* The timestamp config. */
191	struct hwtstamp_config stamp_cfg;
192};
193
194static struct mpipe_data {
195	/* The ingress irq. */
196	int ingress_irq;
197
198	/* The "context" for all devices. */
199	gxio_mpipe_context_t context;
200
201	/* Egress info, indexed by "priv->echannel"
202	 * (lazily created as needed).
203	 */
204	struct tile_net_egress
205	egress_for_echannel[TILE_NET_CHANNELS];
206
207	/* Devices currently associated with each channel.
208	 * NOTE: The array entry can become NULL after ifconfig down, but
209	 * we do not free the underlying net_device structures, so it is
210	 * safe to use a pointer after reading it from this array.
211	 */
212	struct net_device
213	*tile_net_devs_for_channel[TILE_NET_CHANNELS];
214
215	/* The actual memory allocated for the buffer stacks. */
216	void *buffer_stack_vas[MAX_KINDS];
217
218	/* The amount of memory allocated for each buffer stack. */
219	size_t buffer_stack_bytes[MAX_KINDS];
220
221	/* The first buffer stack index
222	 * (small = +0, large = +1, jumbo = +2).
223	 */
224	int first_buffer_stack;
225
226	/* The buckets. */
227	int first_bucket;
228	int num_buckets;
229
230	/* PTP-specific data. */
231	struct ptp_clock *ptp_clock;
232	struct ptp_clock_info caps;
233
234	/* Lock for ptp accessors. */
235	struct mutex ptp_lock;
236
237} mpipe_data[NR_MPIPE_MAX] = {
238	[0 ... (NR_MPIPE_MAX - 1)] {
239		.ingress_irq = -1,
240		.first_buffer_stack = -1,
241		.first_bucket = -1,
242		.num_buckets = 1
243	}
244};
245
246/* A mutex for "tile_net_devs_for_channel". */
247static DEFINE_MUTEX(tile_net_devs_for_channel_mutex);
248
249/* The per-cpu info. */
250static DEFINE_PER_CPU(struct tile_net_info, per_cpu_info);
251
252
253/* The buffer size enums for each buffer stack.
254 * See arch/tile/include/gxio/mpipe.h for the set of possible values.
255 * We avoid the "10384" size because it can induce "false chaining"
256 * on "cut-through" jumbo packets.
257 */
258static gxio_mpipe_buffer_size_enum_t buffer_size_enums[MAX_KINDS] = {
259	GXIO_MPIPE_BUFFER_SIZE_128,
260	GXIO_MPIPE_BUFFER_SIZE_1664,
261	GXIO_MPIPE_BUFFER_SIZE_16384
262};
263
264/* Text value of tile_net.cpus if passed as a module parameter. */
265static char *network_cpus_string;
266
267/* The actual cpus in "network_cpus". */
268static struct cpumask network_cpus_map;
269
270/* If "tile_net.loopify=LINK" was specified, this is "LINK". */
271static char *loopify_link_name;
272
273/* If "tile_net.custom" was specified, this is true. */
274static bool custom_flag;
275
276/* If "tile_net.jumbo=NUM" was specified, this is "NUM". */
277static uint jumbo_num;
278
279/* Obtain mpipe instance from struct tile_net_priv given struct net_device. */
280static inline int mpipe_instance(struct net_device *dev)
281{
282	struct tile_net_priv *priv = netdev_priv(dev);
283	return priv->instance;
284}
285
286/* The "tile_net.cpus" argument specifies the cpus that are dedicated
287 * to handle ingress packets.
288 *
289 * The parameter should be in the form "tile_net.cpus=m-n[,x-y]", where
290 * m, n, x, y are integer numbers that represent the cpus that can be
291 * neither a dedicated cpu nor a dataplane cpu.
292 */
293static bool network_cpus_init(void)
294{
295	int rc;
296
297	if (network_cpus_string == NULL)
298		return false;
299
300	rc = cpulist_parse_crop(network_cpus_string, &network_cpus_map);
301	if (rc != 0) {
302		pr_warn("tile_net.cpus=%s: malformed cpu list\n",
303			network_cpus_string);
304		return false;
305	}
306
307	/* Remove dedicated cpus. */
308	cpumask_and(&network_cpus_map, &network_cpus_map, cpu_possible_mask);
309
310	if (cpumask_empty(&network_cpus_map)) {
311		pr_warn("Ignoring empty tile_net.cpus='%s'.\n",
312			network_cpus_string);
313		return false;
314	}
315
316	pr_info("Linux network CPUs: %*pbl\n",
317		cpumask_pr_args(&network_cpus_map));
318	return true;
319}
320
321module_param_named(cpus, network_cpus_string, charp, 0444);
322MODULE_PARM_DESC(cpus, "cpulist of cores that handle network interrupts");
323
324/* The "tile_net.loopify=LINK" argument causes the named device to
325 * actually use "loop0" for ingress, and "loop1" for egress.  This
326 * allows an app to sit between the actual link and linux, passing
327 * (some) packets along to linux, and forwarding (some) packets sent
328 * out by linux.
329 */
330module_param_named(loopify, loopify_link_name, charp, 0444);
331MODULE_PARM_DESC(loopify, "name the device to use loop0/1 for ingress/egress");
332
333/* The "tile_net.custom" argument causes us to ignore the "conventional"
334 * classifier metadata, in particular, the "l2_offset".
335 */
336module_param_named(custom, custom_flag, bool, 0444);
337MODULE_PARM_DESC(custom, "indicates a (heavily) customized classifier");
338
339/* The "tile_net.jumbo" argument causes us to support "jumbo" packets,
340 * and to allocate the given number of "jumbo" buffers.
341 */
342module_param_named(jumbo, jumbo_num, uint, 0444);
343MODULE_PARM_DESC(jumbo, "the number of buffers to support jumbo packets");
344
345/* Atomically update a statistics field.
346 * Note that on TILE-Gx, this operation is fire-and-forget on the
347 * issuing core (single-cycle dispatch) and takes only a few cycles
348 * longer than a regular store when the request reaches the home cache.
349 * No expensive bus management overhead is required.
350 */
351static void tile_net_stats_add(unsigned long value, unsigned long *field)
352{
353	BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(unsigned long));
354	atomic_long_add(value, (atomic_long_t *)field);
355}
356
357/* Allocate and push a buffer. */
358static bool tile_net_provide_buffer(int instance, int kind)
359{
360	struct mpipe_data *md = &mpipe_data[instance];
361	gxio_mpipe_buffer_size_enum_t bse = buffer_size_enums[kind];
362	size_t bs = gxio_mpipe_buffer_size_enum_to_buffer_size(bse);
363	const unsigned long buffer_alignment = 128;
364	struct sk_buff *skb;
365	int len;
366
367	len = sizeof(struct sk_buff **) + buffer_alignment + bs;
368	skb = dev_alloc_skb(len);
369	if (skb == NULL)
370		return false;
371
372	/* Make room for a back-pointer to 'skb' and guarantee alignment. */
373	skb_reserve(skb, sizeof(struct sk_buff **));
374	skb_reserve(skb, -(long)skb->data & (buffer_alignment - 1));
375
376	/* Save a back-pointer to 'skb'. */
377	*(struct sk_buff **)(skb->data - sizeof(struct sk_buff **)) = skb;
378
379	/* Make sure "skb" and the back-pointer have been flushed. */
380	wmb();
381
382	gxio_mpipe_push_buffer(&md->context, md->first_buffer_stack + kind,
383			       (void *)va_to_tile_io_addr(skb->data));
384
385	return true;
386}
387
388/* Convert a raw mpipe buffer to its matching skb pointer. */
389static struct sk_buff *mpipe_buf_to_skb(void *va)
390{
391	/* Acquire the associated "skb". */
392	struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
393	struct sk_buff *skb = *skb_ptr;
394
395	/* Paranoia. */
396	if (skb->data != va) {
397		/* Panic here since there's a reasonable chance
398		 * that corrupt buffers means generic memory
399		 * corruption, with unpredictable system effects.
400		 */
401		panic("Corrupt linux buffer! va=%p, skb=%p, skb->data=%p",
402		      va, skb, skb->data);
403	}
404
405	return skb;
406}
407
408static void tile_net_pop_all_buffers(int instance, int stack)
409{
410	struct mpipe_data *md = &mpipe_data[instance];
411
412	for (;;) {
413		tile_io_addr_t addr =
414			(tile_io_addr_t)gxio_mpipe_pop_buffer(&md->context,
415							      stack);
416		if (addr == 0)
417			break;
418		dev_kfree_skb_irq(mpipe_buf_to_skb(tile_io_addr_to_va(addr)));
419	}
420}
421
422/* Provide linux buffers to mPIPE. */
423static void tile_net_provide_needed_buffers(void)
424{
425	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
426	int instance, kind;
427	for (instance = 0; instance < NR_MPIPE_MAX &&
428		     info->mpipe[instance].has_iqueue; instance++)	{
429		for (kind = 0; kind < MAX_KINDS; kind++) {
430			while (info->mpipe[instance].num_needed_buffers[kind]
431			       != 0) {
432				if (!tile_net_provide_buffer(instance, kind)) {
433					pr_notice("Tile %d still needs"
434						  " some buffers\n",
435						  info->my_cpu);
436					return;
437				}
438				info->mpipe[instance].
439					num_needed_buffers[kind]--;
440			}
441		}
442	}
443}
444
445/* Get RX timestamp, and store it in the skb. */
446static void tile_rx_timestamp(struct tile_net_priv *priv, struct sk_buff *skb,
447			      gxio_mpipe_idesc_t *idesc)
448{
449	if (unlikely(priv->stamp_cfg.rx_filter != HWTSTAMP_FILTER_NONE)) {
450		struct skb_shared_hwtstamps *shhwtstamps = skb_hwtstamps(skb);
451		memset(shhwtstamps, 0, sizeof(*shhwtstamps));
452		shhwtstamps->hwtstamp = ktime_set(idesc->time_stamp_sec,
453						  idesc->time_stamp_ns);
454	}
455}
456
457/* Get TX timestamp, and store it in the skb. */
458static void tile_tx_timestamp(struct sk_buff *skb, int instance)
459{
460	struct skb_shared_info *shtx = skb_shinfo(skb);
461	if (unlikely((shtx->tx_flags & SKBTX_HW_TSTAMP) != 0)) {
462		struct mpipe_data *md = &mpipe_data[instance];
463		struct skb_shared_hwtstamps shhwtstamps;
464		struct timespec ts;
465
466		shtx->tx_flags |= SKBTX_IN_PROGRESS;
467		gxio_mpipe_get_timestamp(&md->context, &ts);
468		memset(&shhwtstamps, 0, sizeof(shhwtstamps));
469		shhwtstamps.hwtstamp = ktime_set(ts.tv_sec, ts.tv_nsec);
470		skb_tstamp_tx(skb, &shhwtstamps);
471	}
472}
473
474/* Use ioctl() to enable or disable TX or RX timestamping. */
475static int tile_hwtstamp_set(struct net_device *dev, struct ifreq *rq)
476{
477	struct hwtstamp_config config;
478	struct tile_net_priv *priv = netdev_priv(dev);
479
480	if (copy_from_user(&config, rq->ifr_data, sizeof(config)))
481		return -EFAULT;
482
483	if (config.flags)  /* reserved for future extensions */
484		return -EINVAL;
485
486	switch (config.tx_type) {
487	case HWTSTAMP_TX_OFF:
488	case HWTSTAMP_TX_ON:
489		break;
490	default:
491		return -ERANGE;
492	}
493
494	switch (config.rx_filter) {
495	case HWTSTAMP_FILTER_NONE:
496		break;
497	case HWTSTAMP_FILTER_ALL:
498	case HWTSTAMP_FILTER_SOME:
499	case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
500	case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
501	case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
502	case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
503	case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
504	case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
505	case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
506	case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
507	case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
508	case HWTSTAMP_FILTER_PTP_V2_EVENT:
509	case HWTSTAMP_FILTER_PTP_V2_SYNC:
510	case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
511		config.rx_filter = HWTSTAMP_FILTER_ALL;
512		break;
513	default:
514		return -ERANGE;
515	}
516
517	if (copy_to_user(rq->ifr_data, &config, sizeof(config)))
518		return -EFAULT;
519
520	priv->stamp_cfg = config;
521	return 0;
522}
523
524static int tile_hwtstamp_get(struct net_device *dev, struct ifreq *rq)
525{
526	struct tile_net_priv *priv = netdev_priv(dev);
527
528	if (copy_to_user(rq->ifr_data, &priv->stamp_cfg,
529			 sizeof(priv->stamp_cfg)))
530		return -EFAULT;
531
532	return 0;
533}
534
535static inline bool filter_packet(struct net_device *dev, void *buf)
536{
537	/* Filter packets received before we're up. */
538	if (dev == NULL || !(dev->flags & IFF_UP))
539		return true;
540
541	/* Filter out packets that aren't for us. */
542	if (!(dev->flags & IFF_PROMISC) &&
543	    !is_multicast_ether_addr(buf) &&
544	    !ether_addr_equal(dev->dev_addr, buf))
545		return true;
546
547	return false;
548}
549
550static void tile_net_receive_skb(struct net_device *dev, struct sk_buff *skb,
551				 gxio_mpipe_idesc_t *idesc, unsigned long len)
552{
553	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
554	struct tile_net_priv *priv = netdev_priv(dev);
555	int instance = priv->instance;
556
557	/* Encode the actual packet length. */
558	skb_put(skb, len);
559
560	skb->protocol = eth_type_trans(skb, dev);
561
562	/* Acknowledge "good" hardware checksums. */
563	if (idesc->cs && idesc->csum_seed_val == 0xFFFF)
564		skb->ip_summed = CHECKSUM_UNNECESSARY;
565
566	/* Get RX timestamp from idesc. */
567	tile_rx_timestamp(priv, skb, idesc);
568
569	napi_gro_receive(&info->mpipe[instance].napi, skb);
570
571	/* Update stats. */
572	tile_net_stats_add(1, &dev->stats.rx_packets);
573	tile_net_stats_add(len, &dev->stats.rx_bytes);
574
575	/* Need a new buffer. */
576	if (idesc->size == buffer_size_enums[0])
577		info->mpipe[instance].num_needed_buffers[0]++;
578	else if (idesc->size == buffer_size_enums[1])
579		info->mpipe[instance].num_needed_buffers[1]++;
580	else
581		info->mpipe[instance].num_needed_buffers[2]++;
582}
583
584/* Handle a packet.  Return true if "processed", false if "filtered". */
585static bool tile_net_handle_packet(int instance, gxio_mpipe_idesc_t *idesc)
586{
587	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
588	struct mpipe_data *md = &mpipe_data[instance];
589	struct net_device *dev = md->tile_net_devs_for_channel[idesc->channel];
590	uint8_t l2_offset;
591	void *va;
592	void *buf;
593	unsigned long len;
594	bool filter;
595
596	/* Drop packets for which no buffer was available (which can
597	 * happen under heavy load), or for which the me/tr/ce flags
598	 * are set (which can happen for jumbo cut-through packets,
599	 * or with a customized classifier).
600	 */
601	if (idesc->be || idesc->me || idesc->tr || idesc->ce) {
602		if (dev)
603			tile_net_stats_add(1, &dev->stats.rx_errors);
604		goto drop;
605	}
606
607	/* Get the "l2_offset", if allowed. */
608	l2_offset = custom_flag ? 0 : gxio_mpipe_idesc_get_l2_offset(idesc);
609
610	/* Get the VA (including NET_IP_ALIGN bytes of "headroom"). */
611	va = tile_io_addr_to_va((unsigned long)idesc->va);
612
613	/* Get the actual packet start/length. */
614	buf = va + l2_offset;
615	len = idesc->l2_size - l2_offset;
616
617	/* Point "va" at the raw buffer. */
618	va -= NET_IP_ALIGN;
619
620	filter = filter_packet(dev, buf);
621	if (filter) {
622		if (dev)
623			tile_net_stats_add(1, &dev->stats.rx_dropped);
624drop:
625		gxio_mpipe_iqueue_drop(&info->mpipe[instance].iqueue, idesc);
626	} else {
627		struct sk_buff *skb = mpipe_buf_to_skb(va);
628
629		/* Skip headroom, and any custom header. */
630		skb_reserve(skb, NET_IP_ALIGN + l2_offset);
631
632		tile_net_receive_skb(dev, skb, idesc, len);
633	}
634
635	gxio_mpipe_iqueue_consume(&info->mpipe[instance].iqueue, idesc);
636	return !filter;
637}
638
639/* Handle some packets for the current CPU.
640 *
641 * This function handles up to TILE_NET_BATCH idescs per call.
642 *
643 * ISSUE: Since we do not provide new buffers until this function is
644 * complete, we must initially provide enough buffers for each network
645 * cpu to fill its iqueue and also its batched idescs.
646 *
647 * ISSUE: The "rotting packet" race condition occurs if a packet
648 * arrives after the queue appears to be empty, and before the
649 * hypervisor interrupt is re-enabled.
650 */
651static int tile_net_poll(struct napi_struct *napi, int budget)
652{
653	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
654	unsigned int work = 0;
655	gxio_mpipe_idesc_t *idesc;
656	int instance, i, n;
657	struct mpipe_data *md;
658	struct info_mpipe *info_mpipe =
659		container_of(napi, struct info_mpipe, napi);
660
661	if (budget <= 0)
662		goto done;
663
664	instance = info_mpipe->instance;
665	while ((n = gxio_mpipe_iqueue_try_peek(
666			&info_mpipe->iqueue,
667			&idesc)) > 0) {
668		for (i = 0; i < n; i++) {
669			if (i == TILE_NET_BATCH)
670				goto done;
671			if (tile_net_handle_packet(instance,
672						   idesc + i)) {
673				if (++work >= budget)
674					goto done;
675			}
676		}
677	}
678
679	/* There are no packets left. */
680	napi_complete(&info_mpipe->napi);
681
682	md = &mpipe_data[instance];
683	/* Re-enable hypervisor interrupts. */
684	gxio_mpipe_enable_notif_ring_interrupt(
685		&md->context, info->mpipe[instance].iqueue.ring);
686
687	/* HACK: Avoid the "rotting packet" problem. */
688	if (gxio_mpipe_iqueue_try_peek(&info_mpipe->iqueue, &idesc) > 0)
689		napi_schedule(&info_mpipe->napi);
690
691	/* ISSUE: Handle completions? */
692
693done:
694	tile_net_provide_needed_buffers();
695
696	return work;
697}
698
699/* Handle an ingress interrupt from an instance on the current cpu. */
700static irqreturn_t tile_net_handle_ingress_irq(int irq, void *id)
701{
702	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
703	napi_schedule(&info->mpipe[(uint64_t)id].napi);
704	return IRQ_HANDLED;
705}
706
707/* Free some completions.  This must be called with interrupts blocked. */
708static int tile_net_free_comps(gxio_mpipe_equeue_t *equeue,
709				struct tile_net_comps *comps,
710				int limit, bool force_update)
711{
712	int n = 0;
713	while (comps->comp_last < comps->comp_next) {
714		unsigned int cid = comps->comp_last % TILE_NET_MAX_COMPS;
715		struct tile_net_comp *comp = &comps->comp_queue[cid];
716		if (!gxio_mpipe_equeue_is_complete(equeue, comp->when,
717						   force_update || n == 0))
718			break;
719		dev_kfree_skb_irq(comp->skb);
720		comps->comp_last++;
721		if (++n == limit)
722			break;
723	}
724	return n;
725}
726
727/* Add a completion.  This must be called with interrupts blocked.
728 * tile_net_equeue_try_reserve() will have ensured a free completion entry.
729 */
730static void add_comp(gxio_mpipe_equeue_t *equeue,
731		     struct tile_net_comps *comps,
732		     uint64_t when, struct sk_buff *skb)
733{
734	int cid = comps->comp_next % TILE_NET_MAX_COMPS;
735	comps->comp_queue[cid].when = when;
736	comps->comp_queue[cid].skb = skb;
737	comps->comp_next++;
738}
739
740static void tile_net_schedule_tx_wake_timer(struct net_device *dev,
741                                            int tx_queue_idx)
742{
743	struct tile_net_info *info = &per_cpu(per_cpu_info, tx_queue_idx);
744	struct tile_net_priv *priv = netdev_priv(dev);
745	int instance = priv->instance;
746	struct tile_net_tx_wake *tx_wake =
747		&info->mpipe[instance].tx_wake[priv->echannel];
748
749	hrtimer_start(&tx_wake->timer,
750		      ktime_set(0, TX_TIMER_DELAY_USEC * 1000UL),
751		      HRTIMER_MODE_REL_PINNED);
752}
753
754static enum hrtimer_restart tile_net_handle_tx_wake_timer(struct hrtimer *t)
755{
756	struct tile_net_tx_wake *tx_wake =
757		container_of(t, struct tile_net_tx_wake, timer);
758	netif_wake_subqueue(tx_wake->dev, tx_wake->tx_queue_idx);
759	return HRTIMER_NORESTART;
760}
761
762/* Make sure the egress timer is scheduled. */
763static void tile_net_schedule_egress_timer(void)
764{
765	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
766
767	if (!info->egress_timer_scheduled) {
768		hrtimer_start(&info->egress_timer,
769			      ktime_set(0, EGRESS_TIMER_DELAY_USEC * 1000UL),
770			      HRTIMER_MODE_REL_PINNED);
771		info->egress_timer_scheduled = true;
772	}
773}
774
775/* The "function" for "info->egress_timer".
776 *
777 * This timer will reschedule itself as long as there are any pending
778 * completions expected for this tile.
779 */
780static enum hrtimer_restart tile_net_handle_egress_timer(struct hrtimer *t)
781{
782	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
783	unsigned long irqflags;
784	bool pending = false;
785	int i, instance;
786
787	local_irq_save(irqflags);
788
789	/* The timer is no longer scheduled. */
790	info->egress_timer_scheduled = false;
791
792	/* Free all possible comps for this tile. */
793	for (instance = 0; instance < NR_MPIPE_MAX &&
794		     info->mpipe[instance].has_iqueue; instance++) {
795		for (i = 0; i < TILE_NET_CHANNELS; i++) {
796			struct tile_net_egress *egress =
797				&mpipe_data[instance].egress_for_echannel[i];
798			struct tile_net_comps *comps =
799				info->mpipe[instance].comps_for_echannel[i];
800			if (!egress || comps->comp_last >= comps->comp_next)
801				continue;
802			tile_net_free_comps(egress->equeue, comps, -1, true);
803			pending = pending ||
804				(comps->comp_last < comps->comp_next);
805		}
806	}
807
808	/* Reschedule timer if needed. */
809	if (pending)
810		tile_net_schedule_egress_timer();
811
812	local_irq_restore(irqflags);
813
814	return HRTIMER_NORESTART;
815}
816
817/* PTP clock operations. */
818
819static int ptp_mpipe_adjfreq(struct ptp_clock_info *ptp, s32 ppb)
820{
821	int ret = 0;
822	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
823	mutex_lock(&md->ptp_lock);
824	if (gxio_mpipe_adjust_timestamp_freq(&md->context, ppb))
825		ret = -EINVAL;
826	mutex_unlock(&md->ptp_lock);
827	return ret;
828}
829
830static int ptp_mpipe_adjtime(struct ptp_clock_info *ptp, s64 delta)
831{
832	int ret = 0;
833	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
834	mutex_lock(&md->ptp_lock);
835	if (gxio_mpipe_adjust_timestamp(&md->context, delta))
836		ret = -EBUSY;
837	mutex_unlock(&md->ptp_lock);
838	return ret;
839}
840
841static int ptp_mpipe_gettime(struct ptp_clock_info *ptp,
842			     struct timespec64 *ts)
843{
844	int ret = 0;
845	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
846	mutex_lock(&md->ptp_lock);
847	if (gxio_mpipe_get_timestamp(&md->context, ts))
848		ret = -EBUSY;
849	mutex_unlock(&md->ptp_lock);
850	return ret;
851}
852
853static int ptp_mpipe_settime(struct ptp_clock_info *ptp,
854			     const struct timespec64 *ts)
855{
856	int ret = 0;
857	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
858	mutex_lock(&md->ptp_lock);
859	if (gxio_mpipe_set_timestamp(&md->context, ts))
860		ret = -EBUSY;
861	mutex_unlock(&md->ptp_lock);
862	return ret;
863}
864
865static int ptp_mpipe_enable(struct ptp_clock_info *ptp,
866			    struct ptp_clock_request *request, int on)
867{
868	return -EOPNOTSUPP;
869}
870
871static struct ptp_clock_info ptp_mpipe_caps = {
872	.owner		= THIS_MODULE,
873	.name		= "mPIPE clock",
874	.max_adj	= 999999999,
875	.n_ext_ts	= 0,
876	.n_pins		= 0,
877	.pps		= 0,
878	.adjfreq	= ptp_mpipe_adjfreq,
879	.adjtime	= ptp_mpipe_adjtime,
880	.gettime64	= ptp_mpipe_gettime,
881	.settime64	= ptp_mpipe_settime,
882	.enable		= ptp_mpipe_enable,
883};
884
885/* Sync mPIPE's timestamp up with Linux system time and register PTP clock. */
886static void register_ptp_clock(struct net_device *dev, struct mpipe_data *md)
887{
888	struct timespec ts;
889
890	getnstimeofday(&ts);
891	gxio_mpipe_set_timestamp(&md->context, &ts);
892
893	mutex_init(&md->ptp_lock);
894	md->caps = ptp_mpipe_caps;
895	md->ptp_clock = ptp_clock_register(&md->caps, NULL);
896	if (IS_ERR(md->ptp_clock))
897		netdev_err(dev, "ptp_clock_register failed %ld\n",
898			   PTR_ERR(md->ptp_clock));
899}
900
901/* Initialize PTP fields in a new device. */
902static void init_ptp_dev(struct tile_net_priv *priv)
903{
904	priv->stamp_cfg.rx_filter = HWTSTAMP_FILTER_NONE;
905	priv->stamp_cfg.tx_type = HWTSTAMP_TX_OFF;
906}
907
908/* Helper functions for "tile_net_update()". */
909static void enable_ingress_irq(void *irq)
910{
911	enable_percpu_irq((long)irq, 0);
912}
913
914static void disable_ingress_irq(void *irq)
915{
916	disable_percpu_irq((long)irq);
917}
918
919/* Helper function for tile_net_open() and tile_net_stop().
920 * Always called under tile_net_devs_for_channel_mutex.
921 */
922static int tile_net_update(struct net_device *dev)
923{
924	static gxio_mpipe_rules_t rules;  /* too big to fit on the stack */
925	bool saw_channel = false;
926	int instance = mpipe_instance(dev);
927	struct mpipe_data *md = &mpipe_data[instance];
928	int channel;
929	int rc;
930	int cpu;
931
932	saw_channel = false;
933	gxio_mpipe_rules_init(&rules, &md->context);
934
935	for (channel = 0; channel < TILE_NET_CHANNELS; channel++) {
936		if (md->tile_net_devs_for_channel[channel] == NULL)
937			continue;
938		if (!saw_channel) {
939			saw_channel = true;
940			gxio_mpipe_rules_begin(&rules, md->first_bucket,
941					       md->num_buckets, NULL);
942			gxio_mpipe_rules_set_headroom(&rules, NET_IP_ALIGN);
943		}
944		gxio_mpipe_rules_add_channel(&rules, channel);
945	}
946
947	/* NOTE: This can fail if there is no classifier.
948	 * ISSUE: Can anything else cause it to fail?
949	 */
950	rc = gxio_mpipe_rules_commit(&rules);
951	if (rc != 0) {
952		netdev_warn(dev, "gxio_mpipe_rules_commit: mpipe[%d] %d\n",
953			    instance, rc);
954		return -EIO;
955	}
956
957	/* Update all cpus, sequentially (to protect "netif_napi_add()").
958	 * We use on_each_cpu to handle the IPI mask or unmask.
959	 */
960	if (!saw_channel)
961		on_each_cpu(disable_ingress_irq,
962			    (void *)(long)(md->ingress_irq), 1);
963	for_each_online_cpu(cpu) {
964		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
965
966		if (!info->mpipe[instance].has_iqueue)
967			continue;
968		if (saw_channel) {
969			if (!info->mpipe[instance].napi_added) {
970				netif_napi_add(dev, &info->mpipe[instance].napi,
971					       tile_net_poll, TILE_NET_WEIGHT);
972				info->mpipe[instance].napi_added = true;
973			}
974			if (!info->mpipe[instance].napi_enabled) {
975				napi_enable(&info->mpipe[instance].napi);
976				info->mpipe[instance].napi_enabled = true;
977			}
978		} else {
979			if (info->mpipe[instance].napi_enabled) {
980				napi_disable(&info->mpipe[instance].napi);
981				info->mpipe[instance].napi_enabled = false;
982			}
983			/* FIXME: Drain the iqueue. */
984		}
985	}
986	if (saw_channel)
987		on_each_cpu(enable_ingress_irq,
988			    (void *)(long)(md->ingress_irq), 1);
989
990	/* HACK: Allow packets to flow in the simulator. */
991	if (saw_channel)
992		sim_enable_mpipe_links(instance, -1);
993
994	return 0;
995}
996
997/* Initialize a buffer stack. */
998static int create_buffer_stack(struct net_device *dev,
999			       int kind, size_t num_buffers)
1000{
1001	pte_t hash_pte = pte_set_home((pte_t) { 0 }, PAGE_HOME_HASH);
1002	int instance = mpipe_instance(dev);
1003	struct mpipe_data *md = &mpipe_data[instance];
1004	size_t needed = gxio_mpipe_calc_buffer_stack_bytes(num_buffers);
1005	int stack_idx = md->first_buffer_stack + kind;
1006	void *va;
1007	int i, rc;
1008
1009	/* Round up to 64KB and then use alloc_pages() so we get the
1010	 * required 64KB alignment.
1011	 */
1012	md->buffer_stack_bytes[kind] =
1013		ALIGN(needed, 64 * 1024);
1014
1015	va = alloc_pages_exact(md->buffer_stack_bytes[kind], GFP_KERNEL);
1016	if (va == NULL) {
1017		netdev_err(dev,
1018			   "Could not alloc %zd bytes for buffer stack %d\n",
1019			   md->buffer_stack_bytes[kind], kind);
1020		return -ENOMEM;
1021	}
1022
1023	/* Initialize the buffer stack. */
1024	rc = gxio_mpipe_init_buffer_stack(&md->context, stack_idx,
1025					  buffer_size_enums[kind],  va,
1026					  md->buffer_stack_bytes[kind], 0);
1027	if (rc != 0) {
1028		netdev_err(dev, "gxio_mpipe_init_buffer_stack: mpipe[%d] %d\n",
1029			   instance, rc);
1030		free_pages_exact(va, md->buffer_stack_bytes[kind]);
1031		return rc;
1032	}
1033
1034	md->buffer_stack_vas[kind] = va;
1035
1036	rc = gxio_mpipe_register_client_memory(&md->context, stack_idx,
1037					       hash_pte, 0);
1038	if (rc != 0) {
1039		netdev_err(dev,
1040			   "gxio_mpipe_register_client_memory: mpipe[%d] %d\n",
1041			   instance, rc);
1042		return rc;
1043	}
1044
1045	/* Provide initial buffers. */
1046	for (i = 0; i < num_buffers; i++) {
1047		if (!tile_net_provide_buffer(instance, kind)) {
1048			netdev_err(dev, "Cannot allocate initial sk_bufs!\n");
1049			return -ENOMEM;
1050		}
1051	}
1052
1053	return 0;
1054}
1055
1056/* Allocate and initialize mpipe buffer stacks, and register them in
1057 * the mPIPE TLBs, for small, large, and (possibly) jumbo packet sizes.
1058 * This routine supports tile_net_init_mpipe(), below.
1059 */
1060static int init_buffer_stacks(struct net_device *dev,
1061			      int network_cpus_count)
1062{
1063	int num_kinds = MAX_KINDS - (jumbo_num == 0);
1064	size_t num_buffers;
1065	int rc;
1066	int instance = mpipe_instance(dev);
1067	struct mpipe_data *md = &mpipe_data[instance];
1068
1069	/* Allocate the buffer stacks. */
1070	rc = gxio_mpipe_alloc_buffer_stacks(&md->context, num_kinds, 0, 0);
1071	if (rc < 0) {
1072		netdev_err(dev,
1073			   "gxio_mpipe_alloc_buffer_stacks: mpipe[%d] %d\n",
1074			   instance, rc);
1075		return rc;
1076	}
1077	md->first_buffer_stack = rc;
1078
1079	/* Enough small/large buffers to (normally) avoid buffer errors. */
1080	num_buffers =
1081		network_cpus_count * (IQUEUE_ENTRIES + TILE_NET_BATCH);
1082
1083	/* Allocate the small memory stack. */
1084	if (rc >= 0)
1085		rc = create_buffer_stack(dev, 0, num_buffers);
1086
1087	/* Allocate the large buffer stack. */
1088	if (rc >= 0)
1089		rc = create_buffer_stack(dev, 1, num_buffers);
1090
1091	/* Allocate the jumbo buffer stack if needed. */
1092	if (rc >= 0 && jumbo_num != 0)
1093		rc = create_buffer_stack(dev, 2, jumbo_num);
1094
1095	return rc;
1096}
1097
1098/* Allocate per-cpu resources (memory for completions and idescs).
1099 * This routine supports tile_net_init_mpipe(), below.
1100 */
1101static int alloc_percpu_mpipe_resources(struct net_device *dev,
1102					int cpu, int ring)
1103{
1104	struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
1105	int order, i, rc;
1106	int instance = mpipe_instance(dev);
1107	struct mpipe_data *md = &mpipe_data[instance];
1108	struct page *page;
1109	void *addr;
1110
1111	/* Allocate the "comps". */
1112	order = get_order(COMPS_SIZE);
1113	page = homecache_alloc_pages(GFP_KERNEL, order, cpu);
1114	if (page == NULL) {
1115		netdev_err(dev, "Failed to alloc %zd bytes comps memory\n",
1116			   COMPS_SIZE);
1117		return -ENOMEM;
1118	}
1119	addr = pfn_to_kaddr(page_to_pfn(page));
1120	memset(addr, 0, COMPS_SIZE);
1121	for (i = 0; i < TILE_NET_CHANNELS; i++)
1122		info->mpipe[instance].comps_for_echannel[i] =
1123			addr + i * sizeof(struct tile_net_comps);
1124
1125	/* If this is a network cpu, create an iqueue. */
1126	if (cpumask_test_cpu(cpu, &network_cpus_map)) {
1127		order = get_order(NOTIF_RING_SIZE);
1128		page = homecache_alloc_pages(GFP_KERNEL, order, cpu);
1129		if (page == NULL) {
1130			netdev_err(dev,
1131				   "Failed to alloc %zd bytes iqueue memory\n",
1132				   NOTIF_RING_SIZE);
1133			return -ENOMEM;
1134		}
1135		addr = pfn_to_kaddr(page_to_pfn(page));
1136		rc = gxio_mpipe_iqueue_init(&info->mpipe[instance].iqueue,
1137					    &md->context, ring++, addr,
1138					    NOTIF_RING_SIZE, 0);
1139		if (rc < 0) {
1140			netdev_err(dev,
1141				   "gxio_mpipe_iqueue_init failed: %d\n", rc);
1142			return rc;
1143		}
1144		info->mpipe[instance].has_iqueue = true;
1145	}
1146
1147	return ring;
1148}
1149
1150/* Initialize NotifGroup and buckets.
1151 * This routine supports tile_net_init_mpipe(), below.
1152 */
1153static int init_notif_group_and_buckets(struct net_device *dev,
1154					int ring, int network_cpus_count)
1155{
1156	int group, rc;
1157	int instance = mpipe_instance(dev);
1158	struct mpipe_data *md = &mpipe_data[instance];
1159
1160	/* Allocate one NotifGroup. */
1161	rc = gxio_mpipe_alloc_notif_groups(&md->context, 1, 0, 0);
1162	if (rc < 0) {
1163		netdev_err(dev, "gxio_mpipe_alloc_notif_groups: mpipe[%d] %d\n",
1164			   instance, rc);
1165		return rc;
1166	}
1167	group = rc;
1168
1169	/* Initialize global num_buckets value. */
1170	if (network_cpus_count > 4)
1171		md->num_buckets = 256;
1172	else if (network_cpus_count > 1)
1173		md->num_buckets = 16;
1174
1175	/* Allocate some buckets, and set global first_bucket value. */
1176	rc = gxio_mpipe_alloc_buckets(&md->context, md->num_buckets, 0, 0);
1177	if (rc < 0) {
1178		netdev_err(dev, "gxio_mpipe_alloc_buckets: mpipe[%d] %d\n",
1179			   instance, rc);
1180		return rc;
1181	}
1182	md->first_bucket = rc;
1183
1184	/* Init group and buckets. */
1185	rc = gxio_mpipe_init_notif_group_and_buckets(
1186		&md->context, group, ring, network_cpus_count,
1187		md->first_bucket, md->num_buckets,
1188		GXIO_MPIPE_BUCKET_STICKY_FLOW_LOCALITY);
1189	if (rc != 0) {
1190		netdev_err(dev,	"gxio_mpipe_init_notif_group_and_buckets: "
1191			   "mpipe[%d] %d\n", instance, rc);
1192		return rc;
1193	}
1194
1195	return 0;
1196}
1197
1198/* Create an irq and register it, then activate the irq and request
1199 * interrupts on all cores.  Note that "ingress_irq" being initialized
1200 * is how we know not to call tile_net_init_mpipe() again.
1201 * This routine supports tile_net_init_mpipe(), below.
1202 */
1203static int tile_net_setup_interrupts(struct net_device *dev)
1204{
1205	int cpu, rc, irq;
1206	int instance = mpipe_instance(dev);
1207	struct mpipe_data *md = &mpipe_data[instance];
1208
1209	irq = md->ingress_irq;
1210	if (irq < 0) {
1211		irq = irq_alloc_hwirq(-1);
1212		if (!irq) {
1213			netdev_err(dev,
1214				   "create_irq failed: mpipe[%d] %d\n",
1215				   instance, irq);
1216			return irq;
1217		}
1218		tile_irq_activate(irq, TILE_IRQ_PERCPU);
1219
1220		rc = request_irq(irq, tile_net_handle_ingress_irq,
1221				 0, "tile_net", (void *)((uint64_t)instance));
1222
1223		if (rc != 0) {
1224			netdev_err(dev, "request_irq failed: mpipe[%d] %d\n",
1225				   instance, rc);
1226			irq_free_hwirq(irq);
1227			return rc;
1228		}
1229		md->ingress_irq = irq;
1230	}
1231
1232	for_each_online_cpu(cpu) {
1233		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
1234		if (info->mpipe[instance].has_iqueue) {
1235			gxio_mpipe_request_notif_ring_interrupt(&md->context,
1236				cpu_x(cpu), cpu_y(cpu), KERNEL_PL, irq,
1237				info->mpipe[instance].iqueue.ring);
1238		}
1239	}
1240
1241	return 0;
1242}
1243
1244/* Undo any state set up partially by a failed call to tile_net_init_mpipe. */
1245static void tile_net_init_mpipe_fail(int instance)
1246{
1247	int kind, cpu;
1248	struct mpipe_data *md = &mpipe_data[instance];
1249
1250	/* Do cleanups that require the mpipe context first. */
1251	for (kind = 0; kind < MAX_KINDS; kind++) {
1252		if (md->buffer_stack_vas[kind] != NULL) {
1253			tile_net_pop_all_buffers(instance,
1254						 md->first_buffer_stack +
1255						 kind);
1256		}
1257	}
1258
1259	/* Destroy mpipe context so the hardware no longer owns any memory. */
1260	gxio_mpipe_destroy(&md->context);
1261
1262	for_each_online_cpu(cpu) {
1263		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
1264		free_pages(
1265			(unsigned long)(
1266				info->mpipe[instance].comps_for_echannel[0]),
1267			get_order(COMPS_SIZE));
1268		info->mpipe[instance].comps_for_echannel[0] = NULL;
1269		free_pages((unsigned long)(info->mpipe[instance].iqueue.idescs),
1270			   get_order(NOTIF_RING_SIZE));
1271		info->mpipe[instance].iqueue.idescs = NULL;
1272	}
1273
1274	for (kind = 0; kind < MAX_KINDS; kind++) {
1275		if (md->buffer_stack_vas[kind] != NULL) {
1276			free_pages_exact(md->buffer_stack_vas[kind],
1277					 md->buffer_stack_bytes[kind]);
1278			md->buffer_stack_vas[kind] = NULL;
1279		}
1280	}
1281
1282	md->first_buffer_stack = -1;
1283	md->first_bucket = -1;
1284}
1285
1286/* The first time any tilegx network device is opened, we initialize
1287 * the global mpipe state.  If this step fails, we fail to open the
1288 * device, but if it succeeds, we never need to do it again, and since
1289 * tile_net can't be unloaded, we never undo it.
1290 *
1291 * Note that some resources in this path (buffer stack indices,
1292 * bindings from init_buffer_stack, etc.) are hypervisor resources
1293 * that are freed implicitly by gxio_mpipe_destroy().
1294 */
1295static int tile_net_init_mpipe(struct net_device *dev)
1296{
1297	int rc;
1298	int cpu;
1299	int first_ring, ring;
1300	int instance = mpipe_instance(dev);
1301	struct mpipe_data *md = &mpipe_data[instance];
1302	int network_cpus_count = cpumask_weight(&network_cpus_map);
1303
1304	if (!hash_default) {
1305		netdev_err(dev, "Networking requires hash_default!\n");
1306		return -EIO;
1307	}
1308
1309	rc = gxio_mpipe_init(&md->context, instance);
1310	if (rc != 0) {
1311		netdev_err(dev, "gxio_mpipe_init: mpipe[%d] %d\n",
1312			   instance, rc);
1313		return -EIO;
1314	}
1315
1316	/* Set up the buffer stacks. */
1317	rc = init_buffer_stacks(dev, network_cpus_count);
1318	if (rc != 0)
1319		goto fail;
1320
1321	/* Allocate one NotifRing for each network cpu. */
1322	rc = gxio_mpipe_alloc_notif_rings(&md->context,
1323					  network_cpus_count, 0, 0);
1324	if (rc < 0) {
1325		netdev_err(dev, "gxio_mpipe_alloc_notif_rings failed %d\n",
1326			   rc);
1327		goto fail;
1328	}
1329
1330	/* Init NotifRings per-cpu. */
1331	first_ring = rc;
1332	ring = first_ring;
1333	for_each_online_cpu(cpu) {
1334		rc = alloc_percpu_mpipe_resources(dev, cpu, ring);
1335		if (rc < 0)
1336			goto fail;
1337		ring = rc;
1338	}
1339
1340	/* Initialize NotifGroup and buckets. */
1341	rc = init_notif_group_and_buckets(dev, first_ring, network_cpus_count);
1342	if (rc != 0)
1343		goto fail;
1344
1345	/* Create and enable interrupts. */
1346	rc = tile_net_setup_interrupts(dev);
1347	if (rc != 0)
1348		goto fail;
1349
1350	/* Register PTP clock and set mPIPE timestamp, if configured. */
1351	register_ptp_clock(dev, md);
1352
1353	return 0;
1354
1355fail:
1356	tile_net_init_mpipe_fail(instance);
1357	return rc;
1358}
1359
1360/* Create persistent egress info for a given egress channel.
1361 * Note that this may be shared between, say, "gbe0" and "xgbe0".
1362 * ISSUE: Defer header allocation until TSO is actually needed?
1363 */
1364static int tile_net_init_egress(struct net_device *dev, int echannel)
1365{
1366	static int ering = -1;
1367	struct page *headers_page, *edescs_page, *equeue_page;
1368	gxio_mpipe_edesc_t *edescs;
1369	gxio_mpipe_equeue_t *equeue;
1370	unsigned char *headers;
1371	int headers_order, edescs_order, equeue_order;
1372	size_t edescs_size;
1373	int rc = -ENOMEM;
1374	int instance = mpipe_instance(dev);
1375	struct mpipe_data *md = &mpipe_data[instance];
1376
1377	/* Only initialize once. */
1378	if (md->egress_for_echannel[echannel].equeue != NULL)
1379		return 0;
1380
1381	/* Allocate memory for the "headers". */
1382	headers_order = get_order(EQUEUE_ENTRIES * HEADER_BYTES);
1383	headers_page = alloc_pages(GFP_KERNEL, headers_order);
1384	if (headers_page == NULL) {
1385		netdev_warn(dev,
1386			    "Could not alloc %zd bytes for TSO headers.\n",
1387			    PAGE_SIZE << headers_order);
1388		goto fail;
1389	}
1390	headers = pfn_to_kaddr(page_to_pfn(headers_page));
1391
1392	/* Allocate memory for the "edescs". */
1393	edescs_size = EQUEUE_ENTRIES * sizeof(*edescs);
1394	edescs_order = get_order(edescs_size);
1395	edescs_page = alloc_pages(GFP_KERNEL, edescs_order);
1396	if (edescs_page == NULL) {
1397		netdev_warn(dev,
1398			    "Could not alloc %zd bytes for eDMA ring.\n",
1399			    edescs_size);
1400		goto fail_headers;
1401	}
1402	edescs = pfn_to_kaddr(page_to_pfn(edescs_page));
1403
1404	/* Allocate memory for the "equeue". */
1405	equeue_order = get_order(sizeof(*equeue));
1406	equeue_page = alloc_pages(GFP_KERNEL, equeue_order);
1407	if (equeue_page == NULL) {
1408		netdev_warn(dev,
1409			    "Could not alloc %zd bytes for equeue info.\n",
1410			    PAGE_SIZE << equeue_order);
1411		goto fail_edescs;
1412	}
1413	equeue = pfn_to_kaddr(page_to_pfn(equeue_page));
1414
1415	/* Allocate an edma ring (using a one entry "free list"). */
1416	if (ering < 0) {
1417		rc = gxio_mpipe_alloc_edma_rings(&md->context, 1, 0, 0);
1418		if (rc < 0) {
1419			netdev_warn(dev, "gxio_mpipe_alloc_edma_rings: "
1420				    "mpipe[%d] %d\n", instance, rc);
1421			goto fail_equeue;
1422		}
1423		ering = rc;
1424	}
1425
1426	/* Initialize the equeue. */
1427	rc = gxio_mpipe_equeue_init(equeue, &md->context, ering, echannel,
1428				    edescs, edescs_size, 0);
1429	if (rc != 0) {
1430		netdev_err(dev, "gxio_mpipe_equeue_init: mpipe[%d] %d\n",
1431			   instance, rc);
1432		goto fail_equeue;
1433	}
1434
1435	/* Don't reuse the ering later. */
1436	ering = -1;
1437
1438	if (jumbo_num != 0) {
1439		/* Make sure "jumbo" packets can be egressed safely. */
1440		if (gxio_mpipe_equeue_set_snf_size(equeue, 10368) < 0) {
1441			/* ISSUE: There is no "gxio_mpipe_equeue_destroy()". */
1442			netdev_warn(dev, "Jumbo packets may not be egressed"
1443				    " properly on channel %d\n", echannel);
1444		}
1445	}
1446
1447	/* Done. */
1448	md->egress_for_echannel[echannel].equeue = equeue;
1449	md->egress_for_echannel[echannel].headers = headers;
1450	return 0;
1451
1452fail_equeue:
1453	__free_pages(equeue_page, equeue_order);
1454
1455fail_edescs:
1456	__free_pages(edescs_page, edescs_order);
1457
1458fail_headers:
1459	__free_pages(headers_page, headers_order);
1460
1461fail:
1462	return rc;
1463}
1464
1465/* Return channel number for a newly-opened link. */
1466static int tile_net_link_open(struct net_device *dev, gxio_mpipe_link_t *link,
1467			      const char *link_name)
1468{
1469	int instance = mpipe_instance(dev);
1470	struct mpipe_data *md = &mpipe_data[instance];
1471	int rc = gxio_mpipe_link_open(link, &md->context, link_name, 0);
1472	if (rc < 0) {
1473		netdev_err(dev, "Failed to open '%s', mpipe[%d], %d\n",
1474			   link_name, instance, rc);
1475		return rc;
1476	}
1477	if (jumbo_num != 0) {
1478		u32 attr = GXIO_MPIPE_LINK_RECEIVE_JUMBO;
1479		rc = gxio_mpipe_link_set_attr(link, attr, 1);
1480		if (rc != 0) {
1481			netdev_err(dev,
1482				   "Cannot receive jumbo packets on '%s'\n",
1483				   link_name);
1484			gxio_mpipe_link_close(link);
1485			return rc;
1486		}
1487	}
1488	rc = gxio_mpipe_link_channel(link);
1489	if (rc < 0 || rc >= TILE_NET_CHANNELS) {
1490		netdev_err(dev, "gxio_mpipe_link_channel bad value: %d\n", rc);
1491		gxio_mpipe_link_close(link);
1492		return -EINVAL;
1493	}
1494	return rc;
1495}
1496
1497/* Help the kernel activate the given network interface. */
1498static int tile_net_open(struct net_device *dev)
1499{
1500	struct tile_net_priv *priv = netdev_priv(dev);
1501	int cpu, rc, instance;
1502
1503	mutex_lock(&tile_net_devs_for_channel_mutex);
1504
1505	/* Get the instance info. */
1506	rc = gxio_mpipe_link_instance(dev->name);
1507	if (rc < 0 || rc >= NR_MPIPE_MAX) {
1508		mutex_unlock(&tile_net_devs_for_channel_mutex);
1509		return -EIO;
1510	}
1511
1512	priv->instance = rc;
1513	instance = rc;
1514	if (!mpipe_data[rc].context.mmio_fast_base) {
1515		/* Do one-time initialization per instance the first time
1516		 * any device is opened.
1517		 */
1518		rc = tile_net_init_mpipe(dev);
1519		if (rc != 0)
1520			goto fail;
1521	}
1522
1523	/* Determine if this is the "loopify" device. */
1524	if (unlikely((loopify_link_name != NULL) &&
1525		     !strcmp(dev->name, loopify_link_name))) {
1526		rc = tile_net_link_open(dev, &priv->link, "loop0");
1527		if (rc < 0)
1528			goto fail;
1529		priv->channel = rc;
1530		rc = tile_net_link_open(dev, &priv->loopify_link, "loop1");
1531		if (rc < 0)
1532			goto fail;
1533		priv->loopify_channel = rc;
1534		priv->echannel = rc;
1535	} else {
1536		rc = tile_net_link_open(dev, &priv->link, dev->name);
1537		if (rc < 0)
1538			goto fail;
1539		priv->channel = rc;
1540		priv->echannel = rc;
1541	}
1542
1543	/* Initialize egress info (if needed).  Once ever, per echannel. */
1544	rc = tile_net_init_egress(dev, priv->echannel);
1545	if (rc != 0)
1546		goto fail;
1547
1548	mpipe_data[instance].tile_net_devs_for_channel[priv->channel] = dev;
1549
1550	rc = tile_net_update(dev);
1551	if (rc != 0)
1552		goto fail;
1553
1554	mutex_unlock(&tile_net_devs_for_channel_mutex);
1555
1556	/* Initialize the transmit wake timer for this device for each cpu. */
1557	for_each_online_cpu(cpu) {
1558		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
1559		struct tile_net_tx_wake *tx_wake =
1560			&info->mpipe[instance].tx_wake[priv->echannel];
1561
1562		hrtimer_init(&tx_wake->timer, CLOCK_MONOTONIC,
1563			     HRTIMER_MODE_REL);
1564		tx_wake->tx_queue_idx = cpu;
1565		tx_wake->timer.function = tile_net_handle_tx_wake_timer;
1566		tx_wake->dev = dev;
1567	}
1568
1569	for_each_online_cpu(cpu)
1570		netif_start_subqueue(dev, cpu);
1571	netif_carrier_on(dev);
1572	return 0;
1573
1574fail:
1575	if (priv->loopify_channel >= 0) {
1576		if (gxio_mpipe_link_close(&priv->loopify_link) != 0)
1577			netdev_warn(dev, "Failed to close loopify link!\n");
1578		priv->loopify_channel = -1;
1579	}
1580	if (priv->channel >= 0) {
1581		if (gxio_mpipe_link_close(&priv->link) != 0)
1582			netdev_warn(dev, "Failed to close link!\n");
1583		priv->channel = -1;
1584	}
1585	priv->echannel = -1;
1586	mpipe_data[instance].tile_net_devs_for_channel[priv->channel] =	NULL;
1587	mutex_unlock(&tile_net_devs_for_channel_mutex);
1588
1589	/* Don't return raw gxio error codes to generic Linux. */
1590	return (rc > -512) ? rc : -EIO;
1591}
1592
1593/* Help the kernel deactivate the given network interface. */
1594static int tile_net_stop(struct net_device *dev)
1595{
1596	struct tile_net_priv *priv = netdev_priv(dev);
1597	int cpu;
1598	int instance = priv->instance;
1599	struct mpipe_data *md = &mpipe_data[instance];
1600
1601	for_each_online_cpu(cpu) {
1602		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
1603		struct tile_net_tx_wake *tx_wake =
1604			&info->mpipe[instance].tx_wake[priv->echannel];
1605
1606		hrtimer_cancel(&tx_wake->timer);
1607		netif_stop_subqueue(dev, cpu);
1608	}
1609
1610	mutex_lock(&tile_net_devs_for_channel_mutex);
1611	md->tile_net_devs_for_channel[priv->channel] = NULL;
1612	(void)tile_net_update(dev);
1613	if (priv->loopify_channel >= 0) {
1614		if (gxio_mpipe_link_close(&priv->loopify_link) != 0)
1615			netdev_warn(dev, "Failed to close loopify link!\n");
1616		priv->loopify_channel = -1;
1617	}
1618	if (priv->channel >= 0) {
1619		if (gxio_mpipe_link_close(&priv->link) != 0)
1620			netdev_warn(dev, "Failed to close link!\n");
1621		priv->channel = -1;
1622	}
1623	priv->echannel = -1;
1624	mutex_unlock(&tile_net_devs_for_channel_mutex);
1625
1626	return 0;
1627}
1628
1629/* Determine the VA for a fragment. */
1630static inline void *tile_net_frag_buf(skb_frag_t *f)
1631{
1632	unsigned long pfn = page_to_pfn(skb_frag_page(f));
1633	return pfn_to_kaddr(pfn) + f->page_offset;
1634}
1635
1636/* Acquire a completion entry and an egress slot, or if we can't,
1637 * stop the queue and schedule the tx_wake timer.
1638 */
1639static s64 tile_net_equeue_try_reserve(struct net_device *dev,
1640				       int tx_queue_idx,
1641				       struct tile_net_comps *comps,
1642				       gxio_mpipe_equeue_t *equeue,
1643				       int num_edescs)
1644{
1645	/* Try to acquire a completion entry. */
1646	if (comps->comp_next - comps->comp_last < TILE_NET_MAX_COMPS - 1 ||
1647	    tile_net_free_comps(equeue, comps, 32, false) != 0) {
1648
1649		/* Try to acquire an egress slot. */
1650		s64 slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs);
1651		if (slot >= 0)
1652			return slot;
1653
1654		/* Freeing some completions gives the equeue time to drain. */
1655		tile_net_free_comps(equeue, comps, TILE_NET_MAX_COMPS, false);
1656
1657		slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs);
1658		if (slot >= 0)
1659			return slot;
1660	}
1661
1662	/* Still nothing; give up and stop the queue for a short while. */
1663	netif_stop_subqueue(dev, tx_queue_idx);
1664	tile_net_schedule_tx_wake_timer(dev, tx_queue_idx);
1665	return -1;
1666}
1667
1668/* Determine how many edesc's are needed for TSO.
1669 *
1670 * Sometimes, if "sendfile()" requires copying, we will be called with
1671 * "data" containing the header and payload, with "frags" being empty.
1672 * Sometimes, for example when using NFS over TCP, a single segment can
1673 * span 3 fragments.  This requires special care.
1674 */
1675static int tso_count_edescs(struct sk_buff *skb)
1676{
1677	struct skb_shared_info *sh = skb_shinfo(skb);
1678	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1679	unsigned int data_len = skb->len - sh_len;
1680	unsigned int p_len = sh->gso_size;
1681	long f_id = -1;    /* id of the current fragment */
1682	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
1683	long f_used = 0;  /* bytes used from the current fragment */
1684	long n;            /* size of the current piece of payload */
1685	int num_edescs = 0;
1686	int segment;
1687
1688	for (segment = 0; segment < sh->gso_segs; segment++) {
1689
1690		unsigned int p_used = 0;
1691
1692		/* One edesc for header and for each piece of the payload. */
1693		for (num_edescs++; p_used < p_len; num_edescs++) {
1694
1695			/* Advance as needed. */
1696			while (f_used >= f_size) {
1697				f_id++;
1698				f_size = skb_frag_size(&sh->frags[f_id]);
1699				f_used = 0;
1700			}
1701
1702			/* Use bytes from the current fragment. */
1703			n = p_len - p_used;
1704			if (n > f_size - f_used)
1705				n = f_size - f_used;
1706			f_used += n;
1707			p_used += n;
1708		}
1709
1710		/* The last segment may be less than gso_size. */
1711		data_len -= p_len;
1712		if (data_len < p_len)
1713			p_len = data_len;
1714	}
1715
1716	return num_edescs;
1717}
1718
1719/* Prepare modified copies of the skbuff headers. */
1720static void tso_headers_prepare(struct sk_buff *skb, unsigned char *headers,
1721				s64 slot)
1722{
1723	struct skb_shared_info *sh = skb_shinfo(skb);
1724	struct iphdr *ih;
1725	struct ipv6hdr *ih6;
1726	struct tcphdr *th;
1727	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1728	unsigned int data_len = skb->len - sh_len;
1729	unsigned char *data = skb->data;
1730	unsigned int ih_off, th_off, p_len;
1731	unsigned int isum_seed, tsum_seed, seq;
1732	unsigned int uninitialized_var(id);
1733	int is_ipv6;
1734	long f_id = -1;    /* id of the current fragment */
1735	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
1736	long f_used = 0;  /* bytes used from the current fragment */
1737	long n;            /* size of the current piece of payload */
1738	int segment;
1739
1740	/* Locate original headers and compute various lengths. */
1741	is_ipv6 = skb_is_gso_v6(skb);
1742	if (is_ipv6) {
1743		ih6 = ipv6_hdr(skb);
1744		ih_off = skb_network_offset(skb);
1745	} else {
1746		ih = ip_hdr(skb);
1747		ih_off = skb_network_offset(skb);
1748		isum_seed = ((0xFFFF - ih->check) +
1749			     (0xFFFF - ih->tot_len) +
1750			     (0xFFFF - ih->id));
1751		id = ntohs(ih->id);
1752	}
1753
1754	th = tcp_hdr(skb);
1755	th_off = skb_transport_offset(skb);
1756	p_len = sh->gso_size;
1757
1758	tsum_seed = th->check + (0xFFFF ^ htons(skb->len));
1759	seq = ntohl(th->seq);
1760
1761	/* Prepare all the headers. */
1762	for (segment = 0; segment < sh->gso_segs; segment++) {
1763		unsigned char *buf;
1764		unsigned int p_used = 0;
1765
1766		/* Copy to the header memory for this segment. */
1767		buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES +
1768			NET_IP_ALIGN;
1769		memcpy(buf, data, sh_len);
1770
1771		/* Update copied ip header. */
1772		if (is_ipv6) {
1773			ih6 = (struct ipv6hdr *)(buf + ih_off);
1774			ih6->payload_len = htons(sh_len + p_len - ih_off -
1775						 sizeof(*ih6));
1776		} else {
1777			ih = (struct iphdr *)(buf + ih_off);
1778			ih->tot_len = htons(sh_len + p_len - ih_off);
1779			ih->id = htons(id++);
1780			ih->check = csum_long(isum_seed + ih->tot_len +
1781					      ih->id) ^ 0xffff;
1782		}
1783
1784		/* Update copied tcp header. */
1785		th = (struct tcphdr *)(buf + th_off);
1786		th->seq = htonl(seq);
1787		th->check = csum_long(tsum_seed + htons(sh_len + p_len));
1788		if (segment != sh->gso_segs - 1) {
1789			th->fin = 0;
1790			th->psh = 0;
1791		}
1792
1793		/* Skip past the header. */
1794		slot++;
1795
1796		/* Skip past the payload. */
1797		while (p_used < p_len) {
1798
1799			/* Advance as needed. */
1800			while (f_used >= f_size) {
1801				f_id++;
1802				f_size = skb_frag_size(&sh->frags[f_id]);
1803				f_used = 0;
1804			}
1805
1806			/* Use bytes from the current fragment. */
1807			n = p_len - p_used;
1808			if (n > f_size - f_used)
1809				n = f_size - f_used;
1810			f_used += n;
1811			p_used += n;
1812
1813			slot++;
1814		}
1815
1816		seq += p_len;
1817
1818		/* The last segment may be less than gso_size. */
1819		data_len -= p_len;
1820		if (data_len < p_len)
1821			p_len = data_len;
1822	}
1823
1824	/* Flush the headers so they are ready for hardware DMA. */
1825	wmb();
1826}
1827
1828/* Pass all the data to mpipe for egress. */
1829static void tso_egress(struct net_device *dev, gxio_mpipe_equeue_t *equeue,
1830		       struct sk_buff *skb, unsigned char *headers, s64 slot)
1831{
1832	struct skb_shared_info *sh = skb_shinfo(skb);
1833	int instance = mpipe_instance(dev);
1834	struct mpipe_data *md = &mpipe_data[instance];
1835	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1836	unsigned int data_len = skb->len - sh_len;
1837	unsigned int p_len = sh->gso_size;
1838	gxio_mpipe_edesc_t edesc_head = { { 0 } };
1839	gxio_mpipe_edesc_t edesc_body = { { 0 } };
1840	long f_id = -1;    /* id of the current fragment */
1841	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
1842	long f_used = 0;  /* bytes used from the current fragment */
1843	void *f_data = skb->data + sh_len;
1844	long n;            /* size of the current piece of payload */
1845	unsigned long tx_packets = 0, tx_bytes = 0;
1846	unsigned int csum_start;
1847	int segment;
1848
1849	/* Prepare to egress the headers: set up header edesc. */
1850	csum_start = skb_checksum_start_offset(skb);
1851	edesc_head.csum = 1;
1852	edesc_head.csum_start = csum_start;
1853	edesc_head.csum_dest = csum_start + skb->csum_offset;
1854	edesc_head.xfer_size = sh_len;
1855
1856	/* This is only used to specify the TLB. */
1857	edesc_head.stack_idx = md->first_buffer_stack;
1858	edesc_body.stack_idx = md->first_buffer_stack;
1859
1860	/* Egress all the edescs. */
1861	for (segment = 0; segment < sh->gso_segs; segment++) {
1862		unsigned char *buf;
1863		unsigned int p_used = 0;
1864
1865		/* Egress the header. */
1866		buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES +
1867			NET_IP_ALIGN;
1868		edesc_head.va = va_to_tile_io_addr(buf);
1869		gxio_mpipe_equeue_put_at(equeue, edesc_head, slot);
1870		slot++;
1871
1872		/* Egress the payload. */
1873		while (p_used < p_len) {
1874			void *va;
1875
1876			/* Advance as needed. */
1877			while (f_used >= f_size) {
1878				f_id++;
1879				f_size = skb_frag_size(&sh->frags[f_id]);
1880				f_data = tile_net_frag_buf(&sh->frags[f_id]);
1881				f_used = 0;
1882			}
1883
1884			va = f_data + f_used;
1885
1886			/* Use bytes from the current fragment. */
1887			n = p_len - p_used;
1888			if (n > f_size - f_used)
1889				n = f_size - f_used;
1890			f_used += n;
1891			p_used += n;
1892
1893			/* Egress a piece of the payload. */
1894			edesc_body.va = va_to_tile_io_addr(va);
1895			edesc_body.xfer_size = n;
1896			edesc_body.bound = !(p_used < p_len);
1897			gxio_mpipe_equeue_put_at(equeue, edesc_body, slot);
1898			slot++;
1899		}
1900
1901		tx_packets++;
1902		tx_bytes += sh_len + p_len;
1903
1904		/* The last segment may be less than gso_size. */
1905		data_len -= p_len;
1906		if (data_len < p_len)
1907			p_len = data_len;
1908	}
1909
1910	/* Update stats. */
1911	tile_net_stats_add(tx_packets, &dev->stats.tx_packets);
1912	tile_net_stats_add(tx_bytes, &dev->stats.tx_bytes);
1913}
1914
1915/* Do "TSO" handling for egress.
1916 *
1917 * Normally drivers set NETIF_F_TSO only to support hardware TSO;
1918 * otherwise the stack uses scatter-gather to implement GSO in software.
1919 * On our testing, enabling GSO support (via NETIF_F_SG) drops network
1920 * performance down to around 7.5 Gbps on the 10G interfaces, although
1921 * also dropping cpu utilization way down, to under 8%.  But
1922 * implementing "TSO" in the driver brings performance back up to line
1923 * rate, while dropping cpu usage even further, to less than 4%.  In
1924 * practice, profiling of GSO shows that skb_segment() is what causes
1925 * the performance overheads; we benefit in the driver from using
1926 * preallocated memory to duplicate the TCP/IP headers.
1927 */
1928static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
1929{
1930	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
1931	struct tile_net_priv *priv = netdev_priv(dev);
1932	int channel = priv->echannel;
1933	int instance = priv->instance;
1934	struct mpipe_data *md = &mpipe_data[instance];
1935	struct tile_net_egress *egress = &md->egress_for_echannel[channel];
1936	struct tile_net_comps *comps =
1937		info->mpipe[instance].comps_for_echannel[channel];
1938	gxio_mpipe_equeue_t *equeue = egress->equeue;
1939	unsigned long irqflags;
1940	int num_edescs;
1941	s64 slot;
1942
1943	/* Determine how many mpipe edesc's are needed. */
1944	num_edescs = tso_count_edescs(skb);
1945
1946	local_irq_save(irqflags);
1947
1948	/* Try to acquire a completion entry and an egress slot. */
1949	slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps,
1950					   equeue, num_edescs);
1951	if (slot < 0) {
1952		local_irq_restore(irqflags);
1953		return NETDEV_TX_BUSY;
1954	}
1955
1956	/* Set up copies of header data properly. */
1957	tso_headers_prepare(skb, egress->headers, slot);
1958
1959	/* Actually pass the data to the network hardware. */
1960	tso_egress(dev, equeue, skb, egress->headers, slot);
1961
1962	/* Add a completion record. */
1963	add_comp(equeue, comps, slot + num_edescs - 1, skb);
1964
1965	local_irq_restore(irqflags);
1966
1967	/* Make sure the egress timer is scheduled. */
1968	tile_net_schedule_egress_timer();
1969
1970	return NETDEV_TX_OK;
1971}
1972
1973/* Analyze the body and frags for a transmit request. */
1974static unsigned int tile_net_tx_frags(struct frag *frags,
1975				       struct sk_buff *skb,
1976				       void *b_data, unsigned int b_len)
1977{
1978	unsigned int i, n = 0;
1979
1980	struct skb_shared_info *sh = skb_shinfo(skb);
1981
1982	if (b_len != 0) {
1983		frags[n].buf = b_data;
1984		frags[n++].length = b_len;
1985	}
1986
1987	for (i = 0; i < sh->nr_frags; i++) {
1988		skb_frag_t *f = &sh->frags[i];
1989		frags[n].buf = tile_net_frag_buf(f);
1990		frags[n++].length = skb_frag_size(f);
1991	}
1992
1993	return n;
1994}
1995
1996/* Help the kernel transmit a packet. */
1997static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
1998{
1999	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
2000	struct tile_net_priv *priv = netdev_priv(dev);
2001	int instance = priv->instance;
2002	struct mpipe_data *md = &mpipe_data[instance];
2003	struct tile_net_egress *egress =
2004		&md->egress_for_echannel[priv->echannel];
2005	gxio_mpipe_equeue_t *equeue = egress->equeue;
2006	struct tile_net_comps *comps =
2007		info->mpipe[instance].comps_for_echannel[priv->echannel];
2008	unsigned int len = skb->len;
2009	unsigned char *data = skb->data;
2010	unsigned int num_edescs;
2011	struct frag frags[MAX_FRAGS];
2012	gxio_mpipe_edesc_t edescs[MAX_FRAGS];
2013	unsigned long irqflags;
2014	gxio_mpipe_edesc_t edesc = { { 0 } };
2015	unsigned int i;
2016	s64 slot;
2017
2018	if (skb_is_gso(skb))
2019		return tile_net_tx_tso(skb, dev);
2020
2021	num_edescs = tile_net_tx_frags(frags, skb, data, skb_headlen(skb));
2022
2023	/* This is only used to specify the TLB. */
2024	edesc.stack_idx = md->first_buffer_stack;
2025
2026	/* Prepare the edescs. */
2027	for (i = 0; i < num_edescs; i++) {
2028		edesc.xfer_size = frags[i].length;
2029		edesc.va = va_to_tile_io_addr(frags[i].buf);
2030		edescs[i] = edesc;
2031	}
2032
2033	/* Mark the final edesc. */
2034	edescs[num_edescs - 1].bound = 1;
2035
2036	/* Add checksum info to the initial edesc, if needed. */
2037	if (skb->ip_summed == CHECKSUM_PARTIAL) {
2038		unsigned int csum_start = skb_checksum_start_offset(skb);
2039		edescs[0].csum = 1;
2040		edescs[0].csum_start = csum_start;
2041		edescs[0].csum_dest = csum_start + skb->csum_offset;
2042	}
2043
2044	local_irq_save(irqflags);
2045
2046	/* Try to acquire a completion entry and an egress slot. */
2047	slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps,
2048					   equeue, num_edescs);
2049	if (slot < 0) {
2050		local_irq_restore(irqflags);
2051		return NETDEV_TX_BUSY;
2052	}
2053
2054	for (i = 0; i < num_edescs; i++)
2055		gxio_mpipe_equeue_put_at(equeue, edescs[i], slot++);
2056
2057	/* Store TX timestamp if needed. */
2058	tile_tx_timestamp(skb, instance);
2059
2060	/* Add a completion record. */
2061	add_comp(equeue, comps, slot - 1, skb);
2062
2063	/* NOTE: Use ETH_ZLEN for short packets (e.g. 42 < 60). */
2064	tile_net_stats_add(1, &dev->stats.tx_packets);
2065	tile_net_stats_add(max_t(unsigned int, len, ETH_ZLEN),
2066			   &dev->stats.tx_bytes);
2067
2068	local_irq_restore(irqflags);
2069
2070	/* Make sure the egress timer is scheduled. */
2071	tile_net_schedule_egress_timer();
2072
2073	return NETDEV_TX_OK;
2074}
2075
2076/* Return subqueue id on this core (one per core). */
2077static u16 tile_net_select_queue(struct net_device *dev, struct sk_buff *skb,
2078				 void *accel_priv, select_queue_fallback_t fallback)
2079{
2080	return smp_processor_id();
2081}
2082
2083/* Deal with a transmit timeout. */
2084static void tile_net_tx_timeout(struct net_device *dev)
2085{
2086	int cpu;
2087
2088	for_each_online_cpu(cpu)
2089		netif_wake_subqueue(dev, cpu);
2090}
2091
2092/* Ioctl commands. */
2093static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
2094{
2095	if (cmd == SIOCSHWTSTAMP)
2096		return tile_hwtstamp_set(dev, rq);
2097	if (cmd == SIOCGHWTSTAMP)
2098		return tile_hwtstamp_get(dev, rq);
2099
2100	return -EOPNOTSUPP;
2101}
2102
2103/* Change the MTU. */
2104static int tile_net_change_mtu(struct net_device *dev, int new_mtu)
2105{
2106	if (new_mtu < 68)
2107		return -EINVAL;
2108	if (new_mtu > ((jumbo_num != 0) ? 9000 : 1500))
2109		return -EINVAL;
2110	dev->mtu = new_mtu;
2111	return 0;
2112}
2113
2114/* Change the Ethernet address of the NIC.
2115 *
2116 * The hypervisor driver does not support changing MAC address.  However,
2117 * the hardware does not do anything with the MAC address, so the address
2118 * which gets used on outgoing packets, and which is accepted on incoming
2119 * packets, is completely up to us.
2120 *
2121 * Returns 0 on success, negative on failure.
2122 */
2123static int tile_net_set_mac_address(struct net_device *dev, void *p)
2124{
2125	struct sockaddr *addr = p;
2126
2127	if (!is_valid_ether_addr(addr->sa_data))
2128		return -EINVAL;
2129	memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
2130	return 0;
2131}
2132
2133#ifdef CONFIG_NET_POLL_CONTROLLER
2134/* Polling 'interrupt' - used by things like netconsole to send skbs
2135 * without having to re-enable interrupts. It's not called while
2136 * the interrupt routine is executing.
2137 */
2138static void tile_net_netpoll(struct net_device *dev)
2139{
2140	int instance = mpipe_instance(dev);
2141	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
2142	struct mpipe_data *md = &mpipe_data[instance];
2143
2144	disable_percpu_irq(md->ingress_irq);
2145	napi_schedule(&info->mpipe[instance].napi);
2146	enable_percpu_irq(md->ingress_irq, 0);
2147}
2148#endif
2149
2150static const struct net_device_ops tile_net_ops = {
2151	.ndo_open = tile_net_open,
2152	.ndo_stop = tile_net_stop,
2153	.ndo_start_xmit = tile_net_tx,
2154	.ndo_select_queue = tile_net_select_queue,
2155	.ndo_do_ioctl = tile_net_ioctl,
2156	.ndo_change_mtu = tile_net_change_mtu,
2157	.ndo_tx_timeout = tile_net_tx_timeout,
2158	.ndo_set_mac_address = tile_net_set_mac_address,
2159#ifdef CONFIG_NET_POLL_CONTROLLER
2160	.ndo_poll_controller = tile_net_netpoll,
2161#endif
2162};
2163
2164/* The setup function.
2165 *
2166 * This uses ether_setup() to assign various fields in dev, including
2167 * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields.
2168 */
2169static void tile_net_setup(struct net_device *dev)
2170{
2171	netdev_features_t features = 0;
2172
2173	ether_setup(dev);
2174	dev->netdev_ops = &tile_net_ops;
2175	dev->watchdog_timeo = TILE_NET_TIMEOUT;
2176	dev->mtu = 1500;
2177
2178	features |= NETIF_F_HW_CSUM;
2179	features |= NETIF_F_SG;
2180	features |= NETIF_F_TSO;
2181	features |= NETIF_F_TSO6;
2182
2183	dev->hw_features   |= features;
2184	dev->vlan_features |= features;
2185	dev->features      |= features;
2186}
2187
2188/* Allocate the device structure, register the device, and obtain the
2189 * MAC address from the hypervisor.
2190 */
2191static void tile_net_dev_init(const char *name, const uint8_t *mac)
2192{
2193	int ret;
2194	struct net_device *dev;
2195	struct tile_net_priv *priv;
2196
2197	/* HACK: Ignore "loop" links. */
2198	if (strncmp(name, "loop", 4) == 0)
2199		return;
2200
2201	/* Allocate the device structure.  Normally, "name" is a
2202	 * template, instantiated by register_netdev(), but not for us.
2203	 */
2204	dev = alloc_netdev_mqs(sizeof(*priv), name, NET_NAME_UNKNOWN,
2205			       tile_net_setup, NR_CPUS, 1);
2206	if (!dev) {
2207		pr_err("alloc_netdev_mqs(%s) failed\n", name);
2208		return;
2209	}
2210
2211	/* Initialize "priv". */
2212	priv = netdev_priv(dev);
2213	priv->dev = dev;
2214	priv->channel = -1;
2215	priv->loopify_channel = -1;
2216	priv->echannel = -1;
2217	init_ptp_dev(priv);
2218
2219	/* Get the MAC address and set it in the device struct; this must
2220	 * be done before the device is opened.  If the MAC is all zeroes,
2221	 * we use a random address, since we're probably on the simulator.
2222	 */
2223	if (!is_zero_ether_addr(mac))
2224		ether_addr_copy(dev->dev_addr, mac);
2225	else
2226		eth_hw_addr_random(dev);
2227
2228	/* Register the network device. */
2229	ret = register_netdev(dev);
2230	if (ret) {
2231		netdev_err(dev, "register_netdev failed %d\n", ret);
2232		free_netdev(dev);
2233		return;
2234	}
2235}
2236
2237/* Per-cpu module initialization. */
2238static void tile_net_init_module_percpu(void *unused)
2239{
2240	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
2241	int my_cpu = smp_processor_id();
2242	int instance;
2243
2244	for (instance = 0; instance < NR_MPIPE_MAX; instance++) {
2245		info->mpipe[instance].has_iqueue = false;
2246		info->mpipe[instance].instance = instance;
2247	}
2248	info->my_cpu = my_cpu;
2249
2250	/* Initialize the egress timer. */
2251	hrtimer_init(&info->egress_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2252	info->egress_timer.function = tile_net_handle_egress_timer;
2253}
2254
2255/* Module initialization. */
2256static int __init tile_net_init_module(void)
2257{
2258	int i;
2259	char name[GXIO_MPIPE_LINK_NAME_LEN];
2260	uint8_t mac[6];
2261
2262	pr_info("Tilera Network Driver\n");
2263
2264	BUILD_BUG_ON(NR_MPIPE_MAX != 2);
2265
2266	mutex_init(&tile_net_devs_for_channel_mutex);
2267
2268	/* Initialize each CPU. */
2269	on_each_cpu(tile_net_init_module_percpu, NULL, 1);
2270
2271	/* Find out what devices we have, and initialize them. */
2272	for (i = 0; gxio_mpipe_link_enumerate_mac(i, name, mac) >= 0; i++)
2273		tile_net_dev_init(name, mac);
2274
2275	if (!network_cpus_init())
2276		network_cpus_map = *cpu_online_mask;
2277
2278	return 0;
2279}
2280
2281module_init(tile_net_init_module);
2282