]>
Commit | Line | Data |
---|---|---|
e3d62d7e CM |
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> | |
2c7d04a9 | 39 | #include <linux/ipv6.h> |
e3d62d7e | 40 | #include <linux/tcp.h> |
9ab5ec59 CM |
41 | #include <linux/net_tstamp.h> |
42 | #include <linux/ptp_clock_kernel.h> | |
03f6199a | 43 | #include <linux/tick.h> |
e3d62d7e CM |
44 | |
45 | #include <asm/checksum.h> | |
46 | #include <asm/homecache.h> | |
47 | #include <gxio/mpipe.h> | |
48 | #include <arch/sim.h> | |
49 | ||
50 | /* Default transmit lockup timeout period, in jiffies. */ | |
51 | #define TILE_NET_TIMEOUT (5 * HZ) | |
52 | ||
53 | /* The maximum number of distinct channels (idesc.channel is 5 bits). */ | |
54 | #define TILE_NET_CHANNELS 32 | |
55 | ||
56 | /* Maximum number of idescs to handle per "poll". */ | |
57 | #define TILE_NET_BATCH 128 | |
58 | ||
59 | /* Maximum number of packets to handle per "poll". */ | |
60 | #define TILE_NET_WEIGHT 64 | |
61 | ||
f58f0817 JW |
62 | /* Maximum Jumbo Packet MTU */ |
63 | #define TILE_JUMBO_MAX_MTU 9000 | |
64 | ||
e3d62d7e CM |
65 | /* Number of entries in each iqueue. */ |
66 | #define IQUEUE_ENTRIES 512 | |
67 | ||
68 | /* Number of entries in each equeue. */ | |
69 | #define EQUEUE_ENTRIES 2048 | |
70 | ||
71 | /* Total header bytes per equeue slot. Must be big enough for 2 bytes | |
72 | * of NET_IP_ALIGN alignment, plus 14 bytes (?) of L2 header, plus up to | |
73 | * 60 bytes of actual TCP header. We round up to align to cache lines. | |
74 | */ | |
75 | #define HEADER_BYTES 128 | |
76 | ||
77 | /* Maximum completions per cpu per device (must be a power of two). | |
78 | * ISSUE: What is the right number here? If this is too small, then | |
79 | * egress might block waiting for free space in a completions array. | |
80 | * ISSUE: At the least, allocate these only for initialized echannels. | |
81 | */ | |
82 | #define TILE_NET_MAX_COMPS 64 | |
83 | ||
84 | #define MAX_FRAGS (MAX_SKB_FRAGS + 1) | |
85 | ||
2628e8af CM |
86 | /* The "kinds" of buffer stacks (small/large/jumbo). */ |
87 | #define MAX_KINDS 3 | |
88 | ||
e3d62d7e CM |
89 | /* Size of completions data to allocate. |
90 | * ISSUE: Probably more than needed since we don't use all the channels. | |
91 | */ | |
92 | #define COMPS_SIZE (TILE_NET_CHANNELS * sizeof(struct tile_net_comps)) | |
93 | ||
94 | /* Size of NotifRing data to allocate. */ | |
95 | #define NOTIF_RING_SIZE (IQUEUE_ENTRIES * sizeof(gxio_mpipe_idesc_t)) | |
96 | ||
97 | /* Timeout to wake the per-device TX timer after we stop the queue. | |
98 | * We don't want the timeout too short (adds overhead, and might end | |
99 | * up causing stop/wake/stop/wake cycles) or too long (affects performance). | |
100 | * For the 10 Gb NIC, 30 usec means roughly 30+ 1500-byte packets. | |
101 | */ | |
102 | #define TX_TIMER_DELAY_USEC 30 | |
103 | ||
104 | /* Timeout to wake the per-cpu egress timer to free completions. */ | |
105 | #define EGRESS_TIMER_DELAY_USEC 1000 | |
106 | ||
107 | MODULE_AUTHOR("Tilera Corporation"); | |
108 | MODULE_LICENSE("GPL"); | |
109 | ||
110 | /* A "packet fragment" (a chunk of memory). */ | |
111 | struct frag { | |
112 | void *buf; | |
113 | size_t length; | |
114 | }; | |
115 | ||
116 | /* A single completion. */ | |
117 | struct tile_net_comp { | |
118 | /* The "complete_count" when the completion will be complete. */ | |
119 | s64 when; | |
120 | /* The buffer to be freed when the completion is complete. */ | |
121 | struct sk_buff *skb; | |
122 | }; | |
123 | ||
124 | /* The completions for a given cpu and echannel. */ | |
125 | struct tile_net_comps { | |
126 | /* The completions. */ | |
127 | struct tile_net_comp comp_queue[TILE_NET_MAX_COMPS]; | |
128 | /* The number of completions used. */ | |
129 | unsigned long comp_next; | |
130 | /* The number of completions freed. */ | |
131 | unsigned long comp_last; | |
132 | }; | |
133 | ||
134 | /* The transmit wake timer for a given cpu and echannel. */ | |
135 | struct tile_net_tx_wake { | |
9b4c341b | 136 | int tx_queue_idx; |
e3d62d7e CM |
137 | struct hrtimer timer; |
138 | struct net_device *dev; | |
139 | }; | |
140 | ||
141 | /* Info for a specific cpu. */ | |
142 | struct tile_net_info { | |
e3d62d7e CM |
143 | /* Our cpu. */ |
144 | int my_cpu; | |
e3d62d7e CM |
145 | /* A timer for handling egress completions. */ |
146 | struct hrtimer egress_timer; | |
147 | /* True if "egress_timer" is scheduled. */ | |
148 | bool egress_timer_scheduled; | |
f3286a3a CM |
149 | struct info_mpipe { |
150 | /* Packet queue. */ | |
151 | gxio_mpipe_iqueue_t iqueue; | |
152 | /* The NAPI struct. */ | |
153 | struct napi_struct napi; | |
154 | /* Number of buffers (by kind) which must still be provided. */ | |
155 | unsigned int num_needed_buffers[MAX_KINDS]; | |
156 | /* instance id. */ | |
157 | int instance; | |
158 | /* True if iqueue is valid. */ | |
159 | bool has_iqueue; | |
160 | /* NAPI flags. */ | |
161 | bool napi_added; | |
162 | bool napi_enabled; | |
163 | /* Comps for each egress channel. */ | |
164 | struct tile_net_comps *comps_for_echannel[TILE_NET_CHANNELS]; | |
165 | /* Transmit wake timer for each egress channel. */ | |
166 | struct tile_net_tx_wake tx_wake[TILE_NET_CHANNELS]; | |
167 | } mpipe[NR_MPIPE_MAX]; | |
e3d62d7e CM |
168 | }; |
169 | ||
170 | /* Info for egress on a particular egress channel. */ | |
171 | struct tile_net_egress { | |
172 | /* The "equeue". */ | |
173 | gxio_mpipe_equeue_t *equeue; | |
174 | /* The headers for TSO. */ | |
175 | unsigned char *headers; | |
176 | }; | |
177 | ||
178 | /* Info for a specific device. */ | |
179 | struct tile_net_priv { | |
180 | /* Our network device. */ | |
181 | struct net_device *dev; | |
182 | /* The primary link. */ | |
183 | gxio_mpipe_link_t link; | |
184 | /* The primary channel, if open, else -1. */ | |
185 | int channel; | |
186 | /* The "loopify" egress link, if needed. */ | |
187 | gxio_mpipe_link_t loopify_link; | |
188 | /* The "loopify" egress channel, if open, else -1. */ | |
189 | int loopify_channel; | |
190 | /* The egress channel (channel or loopify_channel). */ | |
191 | int echannel; | |
f3286a3a CM |
192 | /* mPIPE instance, 0 or 1. */ |
193 | int instance; | |
9ab5ec59 CM |
194 | /* The timestamp config. */ |
195 | struct hwtstamp_config stamp_cfg; | |
e3d62d7e CM |
196 | }; |
197 | ||
f3286a3a CM |
198 | static struct mpipe_data { |
199 | /* The ingress irq. */ | |
200 | int ingress_irq; | |
e3d62d7e | 201 | |
f3286a3a CM |
202 | /* The "context" for all devices. */ |
203 | gxio_mpipe_context_t context; | |
204 | ||
205 | /* Egress info, indexed by "priv->echannel" | |
206 | * (lazily created as needed). | |
207 | */ | |
208 | struct tile_net_egress | |
209 | egress_for_echannel[TILE_NET_CHANNELS]; | |
210 | ||
211 | /* Devices currently associated with each channel. | |
212 | * NOTE: The array entry can become NULL after ifconfig down, but | |
213 | * we do not free the underlying net_device structures, so it is | |
214 | * safe to use a pointer after reading it from this array. | |
215 | */ | |
216 | struct net_device | |
217 | *tile_net_devs_for_channel[TILE_NET_CHANNELS]; | |
218 | ||
219 | /* The actual memory allocated for the buffer stacks. */ | |
220 | void *buffer_stack_vas[MAX_KINDS]; | |
221 | ||
222 | /* The amount of memory allocated for each buffer stack. */ | |
223 | size_t buffer_stack_bytes[MAX_KINDS]; | |
224 | ||
225 | /* The first buffer stack index | |
226 | * (small = +0, large = +1, jumbo = +2). | |
227 | */ | |
228 | int first_buffer_stack; | |
229 | ||
230 | /* The buckets. */ | |
231 | int first_bucket; | |
232 | int num_buckets; | |
233 | ||
9ab5ec59 CM |
234 | /* PTP-specific data. */ |
235 | struct ptp_clock *ptp_clock; | |
236 | struct ptp_clock_info caps; | |
237 | ||
238 | /* Lock for ptp accessors. */ | |
239 | struct mutex ptp_lock; | |
9ab5ec59 | 240 | |
f3286a3a CM |
241 | } mpipe_data[NR_MPIPE_MAX] = { |
242 | [0 ... (NR_MPIPE_MAX - 1)] { | |
243 | .ingress_irq = -1, | |
244 | .first_buffer_stack = -1, | |
245 | .first_bucket = -1, | |
246 | .num_buckets = 1 | |
247 | } | |
248 | }; | |
e3d62d7e CM |
249 | |
250 | /* A mutex for "tile_net_devs_for_channel". */ | |
251 | static DEFINE_MUTEX(tile_net_devs_for_channel_mutex); | |
252 | ||
253 | /* The per-cpu info. */ | |
254 | static DEFINE_PER_CPU(struct tile_net_info, per_cpu_info); | |
255 | ||
e3d62d7e | 256 | |
2628e8af | 257 | /* The buffer size enums for each buffer stack. |
e3d62d7e | 258 | * See arch/tile/include/gxio/mpipe.h for the set of possible values. |
2628e8af CM |
259 | * We avoid the "10384" size because it can induce "false chaining" |
260 | * on "cut-through" jumbo packets. | |
e3d62d7e | 261 | */ |
2628e8af CM |
262 | static gxio_mpipe_buffer_size_enum_t buffer_size_enums[MAX_KINDS] = { |
263 | GXIO_MPIPE_BUFFER_SIZE_128, | |
264 | GXIO_MPIPE_BUFFER_SIZE_1664, | |
265 | GXIO_MPIPE_BUFFER_SIZE_16384 | |
266 | }; | |
e3d62d7e | 267 | |
e3d62d7e CM |
268 | /* Text value of tile_net.cpus if passed as a module parameter. */ |
269 | static char *network_cpus_string; | |
270 | ||
271 | /* The actual cpus in "network_cpus". */ | |
272 | static struct cpumask network_cpus_map; | |
273 | ||
4aa02644 | 274 | /* If "tile_net.loopify=LINK" was specified, this is "LINK". */ |
e3d62d7e CM |
275 | static char *loopify_link_name; |
276 | ||
4aa02644 CM |
277 | /* If "tile_net.custom" was specified, this is true. */ |
278 | static bool custom_flag; | |
e3d62d7e | 279 | |
2628e8af CM |
280 | /* If "tile_net.jumbo=NUM" was specified, this is "NUM". */ |
281 | static uint jumbo_num; | |
282 | ||
f3286a3a CM |
283 | /* Obtain mpipe instance from struct tile_net_priv given struct net_device. */ |
284 | static inline int mpipe_instance(struct net_device *dev) | |
285 | { | |
286 | struct tile_net_priv *priv = netdev_priv(dev); | |
287 | return priv->instance; | |
288 | } | |
289 | ||
e3d62d7e CM |
290 | /* The "tile_net.cpus" argument specifies the cpus that are dedicated |
291 | * to handle ingress packets. | |
292 | * | |
293 | * The parameter should be in the form "tile_net.cpus=m-n[,x-y]", where | |
294 | * m, n, x, y are integer numbers that represent the cpus that can be | |
295 | * neither a dedicated cpu nor a dataplane cpu. | |
296 | */ | |
297 | static bool network_cpus_init(void) | |
298 | { | |
e3d62d7e CM |
299 | int rc; |
300 | ||
301 | if (network_cpus_string == NULL) | |
302 | return false; | |
303 | ||
304 | rc = cpulist_parse_crop(network_cpus_string, &network_cpus_map); | |
305 | if (rc != 0) { | |
306 | pr_warn("tile_net.cpus=%s: malformed cpu list\n", | |
307 | network_cpus_string); | |
308 | return false; | |
309 | } | |
310 | ||
311 | /* Remove dedicated cpus. */ | |
312 | cpumask_and(&network_cpus_map, &network_cpus_map, cpu_possible_mask); | |
313 | ||
314 | if (cpumask_empty(&network_cpus_map)) { | |
315 | pr_warn("Ignoring empty tile_net.cpus='%s'.\n", | |
316 | network_cpus_string); | |
317 | return false; | |
318 | } | |
319 | ||
839b2680 TH |
320 | pr_info("Linux network CPUs: %*pbl\n", |
321 | cpumask_pr_args(&network_cpus_map)); | |
e3d62d7e CM |
322 | return true; |
323 | } | |
324 | ||
325 | module_param_named(cpus, network_cpus_string, charp, 0444); | |
326 | MODULE_PARM_DESC(cpus, "cpulist of cores that handle network interrupts"); | |
327 | ||
328 | /* The "tile_net.loopify=LINK" argument causes the named device to | |
329 | * actually use "loop0" for ingress, and "loop1" for egress. This | |
330 | * allows an app to sit between the actual link and linux, passing | |
331 | * (some) packets along to linux, and forwarding (some) packets sent | |
332 | * out by linux. | |
333 | */ | |
334 | module_param_named(loopify, loopify_link_name, charp, 0444); | |
335 | MODULE_PARM_DESC(loopify, "name the device to use loop0/1 for ingress/egress"); | |
336 | ||
337 | /* The "tile_net.custom" argument causes us to ignore the "conventional" | |
338 | * classifier metadata, in particular, the "l2_offset". | |
339 | */ | |
4aa02644 | 340 | module_param_named(custom, custom_flag, bool, 0444); |
e3d62d7e CM |
341 | MODULE_PARM_DESC(custom, "indicates a (heavily) customized classifier"); |
342 | ||
2628e8af CM |
343 | /* The "tile_net.jumbo" argument causes us to support "jumbo" packets, |
344 | * and to allocate the given number of "jumbo" buffers. | |
345 | */ | |
346 | module_param_named(jumbo, jumbo_num, uint, 0444); | |
347 | MODULE_PARM_DESC(jumbo, "the number of buffers to support jumbo packets"); | |
348 | ||
e3d62d7e CM |
349 | /* Atomically update a statistics field. |
350 | * Note that on TILE-Gx, this operation is fire-and-forget on the | |
351 | * issuing core (single-cycle dispatch) and takes only a few cycles | |
352 | * longer than a regular store when the request reaches the home cache. | |
353 | * No expensive bus management overhead is required. | |
354 | */ | |
355 | static void tile_net_stats_add(unsigned long value, unsigned long *field) | |
356 | { | |
357 | BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(unsigned long)); | |
358 | atomic_long_add(value, (atomic_long_t *)field); | |
359 | } | |
360 | ||
361 | /* Allocate and push a buffer. */ | |
f3286a3a | 362 | static bool tile_net_provide_buffer(int instance, int kind) |
e3d62d7e | 363 | { |
f3286a3a | 364 | struct mpipe_data *md = &mpipe_data[instance]; |
2628e8af CM |
365 | gxio_mpipe_buffer_size_enum_t bse = buffer_size_enums[kind]; |
366 | size_t bs = gxio_mpipe_buffer_size_enum_to_buffer_size(bse); | |
e3d62d7e CM |
367 | const unsigned long buffer_alignment = 128; |
368 | struct sk_buff *skb; | |
369 | int len; | |
370 | ||
2628e8af | 371 | len = sizeof(struct sk_buff **) + buffer_alignment + bs; |
e3d62d7e CM |
372 | skb = dev_alloc_skb(len); |
373 | if (skb == NULL) | |
374 | return false; | |
375 | ||
376 | /* Make room for a back-pointer to 'skb' and guarantee alignment. */ | |
377 | skb_reserve(skb, sizeof(struct sk_buff **)); | |
378 | skb_reserve(skb, -(long)skb->data & (buffer_alignment - 1)); | |
379 | ||
380 | /* Save a back-pointer to 'skb'. */ | |
381 | *(struct sk_buff **)(skb->data - sizeof(struct sk_buff **)) = skb; | |
382 | ||
383 | /* Make sure "skb" and the back-pointer have been flushed. */ | |
384 | wmb(); | |
385 | ||
f3286a3a | 386 | gxio_mpipe_push_buffer(&md->context, md->first_buffer_stack + kind, |
e3d62d7e CM |
387 | (void *)va_to_tile_io_addr(skb->data)); |
388 | ||
389 | return true; | |
390 | } | |
391 | ||
392 | /* Convert a raw mpipe buffer to its matching skb pointer. */ | |
393 | static struct sk_buff *mpipe_buf_to_skb(void *va) | |
394 | { | |
395 | /* Acquire the associated "skb". */ | |
396 | struct sk_buff **skb_ptr = va - sizeof(*skb_ptr); | |
397 | struct sk_buff *skb = *skb_ptr; | |
398 | ||
399 | /* Paranoia. */ | |
400 | if (skb->data != va) { | |
401 | /* Panic here since there's a reasonable chance | |
402 | * that corrupt buffers means generic memory | |
403 | * corruption, with unpredictable system effects. | |
404 | */ | |
405 | panic("Corrupt linux buffer! va=%p, skb=%p, skb->data=%p", | |
406 | va, skb, skb->data); | |
407 | } | |
408 | ||
409 | return skb; | |
410 | } | |
411 | ||
f3286a3a | 412 | static void tile_net_pop_all_buffers(int instance, int stack) |
e3d62d7e | 413 | { |
f3286a3a CM |
414 | struct mpipe_data *md = &mpipe_data[instance]; |
415 | ||
e3d62d7e CM |
416 | for (;;) { |
417 | tile_io_addr_t addr = | |
f3286a3a CM |
418 | (tile_io_addr_t)gxio_mpipe_pop_buffer(&md->context, |
419 | stack); | |
e3d62d7e CM |
420 | if (addr == 0) |
421 | break; | |
422 | dev_kfree_skb_irq(mpipe_buf_to_skb(tile_io_addr_to_va(addr))); | |
423 | } | |
424 | } | |
425 | ||
426 | /* Provide linux buffers to mPIPE. */ | |
427 | static void tile_net_provide_needed_buffers(void) | |
428 | { | |
70b2776a | 429 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
f3286a3a CM |
430 | int instance, kind; |
431 | for (instance = 0; instance < NR_MPIPE_MAX && | |
432 | info->mpipe[instance].has_iqueue; instance++) { | |
433 | for (kind = 0; kind < MAX_KINDS; kind++) { | |
434 | while (info->mpipe[instance].num_needed_buffers[kind] | |
435 | != 0) { | |
436 | if (!tile_net_provide_buffer(instance, kind)) { | |
437 | pr_notice("Tile %d still needs" | |
438 | " some buffers\n", | |
439 | info->my_cpu); | |
440 | return; | |
441 | } | |
442 | info->mpipe[instance]. | |
443 | num_needed_buffers[kind]--; | |
2628e8af | 444 | } |
2628e8af | 445 | } |
e3d62d7e | 446 | } |
e3d62d7e CM |
447 | } |
448 | ||
9ab5ec59 CM |
449 | /* Get RX timestamp, and store it in the skb. */ |
450 | static void tile_rx_timestamp(struct tile_net_priv *priv, struct sk_buff *skb, | |
451 | gxio_mpipe_idesc_t *idesc) | |
452 | { | |
9ab5ec59 CM |
453 | if (unlikely(priv->stamp_cfg.rx_filter != HWTSTAMP_FILTER_NONE)) { |
454 | struct skb_shared_hwtstamps *shhwtstamps = skb_hwtstamps(skb); | |
455 | memset(shhwtstamps, 0, sizeof(*shhwtstamps)); | |
456 | shhwtstamps->hwtstamp = ktime_set(idesc->time_stamp_sec, | |
457 | idesc->time_stamp_ns); | |
458 | } | |
9ab5ec59 CM |
459 | } |
460 | ||
461 | /* Get TX timestamp, and store it in the skb. */ | |
462 | static void tile_tx_timestamp(struct sk_buff *skb, int instance) | |
463 | { | |
9ab5ec59 CM |
464 | struct skb_shared_info *shtx = skb_shinfo(skb); |
465 | if (unlikely((shtx->tx_flags & SKBTX_HW_TSTAMP) != 0)) { | |
466 | struct mpipe_data *md = &mpipe_data[instance]; | |
467 | struct skb_shared_hwtstamps shhwtstamps; | |
a9836cbb | 468 | struct timespec64 ts; |
9ab5ec59 CM |
469 | |
470 | shtx->tx_flags |= SKBTX_IN_PROGRESS; | |
471 | gxio_mpipe_get_timestamp(&md->context, &ts); | |
472 | memset(&shhwtstamps, 0, sizeof(shhwtstamps)); | |
473 | shhwtstamps.hwtstamp = ktime_set(ts.tv_sec, ts.tv_nsec); | |
474 | skb_tstamp_tx(skb, &shhwtstamps); | |
475 | } | |
9ab5ec59 CM |
476 | } |
477 | ||
478 | /* Use ioctl() to enable or disable TX or RX timestamping. */ | |
6ab96d1e | 479 | static int tile_hwtstamp_set(struct net_device *dev, struct ifreq *rq) |
9ab5ec59 | 480 | { |
9ab5ec59 CM |
481 | struct hwtstamp_config config; |
482 | struct tile_net_priv *priv = netdev_priv(dev); | |
483 | ||
484 | if (copy_from_user(&config, rq->ifr_data, sizeof(config))) | |
485 | return -EFAULT; | |
486 | ||
487 | if (config.flags) /* reserved for future extensions */ | |
488 | return -EINVAL; | |
489 | ||
490 | switch (config.tx_type) { | |
491 | case HWTSTAMP_TX_OFF: | |
492 | case HWTSTAMP_TX_ON: | |
493 | break; | |
494 | default: | |
495 | return -ERANGE; | |
496 | } | |
497 | ||
498 | switch (config.rx_filter) { | |
499 | case HWTSTAMP_FILTER_NONE: | |
500 | break; | |
501 | case HWTSTAMP_FILTER_ALL: | |
502 | case HWTSTAMP_FILTER_SOME: | |
503 | case HWTSTAMP_FILTER_PTP_V1_L4_EVENT: | |
504 | case HWTSTAMP_FILTER_PTP_V1_L4_SYNC: | |
505 | case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ: | |
506 | case HWTSTAMP_FILTER_PTP_V2_L4_EVENT: | |
507 | case HWTSTAMP_FILTER_PTP_V2_L4_SYNC: | |
508 | case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ: | |
509 | case HWTSTAMP_FILTER_PTP_V2_L2_EVENT: | |
510 | case HWTSTAMP_FILTER_PTP_V2_L2_SYNC: | |
511 | case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ: | |
512 | case HWTSTAMP_FILTER_PTP_V2_EVENT: | |
513 | case HWTSTAMP_FILTER_PTP_V2_SYNC: | |
514 | case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ: | |
e3412575 | 515 | case HWTSTAMP_FILTER_NTP_ALL: |
9ab5ec59 CM |
516 | config.rx_filter = HWTSTAMP_FILTER_ALL; |
517 | break; | |
518 | default: | |
519 | return -ERANGE; | |
520 | } | |
521 | ||
522 | if (copy_to_user(rq->ifr_data, &config, sizeof(config))) | |
523 | return -EFAULT; | |
524 | ||
525 | priv->stamp_cfg = config; | |
526 | return 0; | |
9ab5ec59 CM |
527 | } |
528 | ||
6ab96d1e BH |
529 | static int tile_hwtstamp_get(struct net_device *dev, struct ifreq *rq) |
530 | { | |
6ab96d1e BH |
531 | struct tile_net_priv *priv = netdev_priv(dev); |
532 | ||
533 | if (copy_to_user(rq->ifr_data, &priv->stamp_cfg, | |
534 | sizeof(priv->stamp_cfg))) | |
535 | return -EFAULT; | |
536 | ||
537 | return 0; | |
6ab96d1e BH |
538 | } |
539 | ||
e3d62d7e CM |
540 | static inline bool filter_packet(struct net_device *dev, void *buf) |
541 | { | |
542 | /* Filter packets received before we're up. */ | |
543 | if (dev == NULL || !(dev->flags & IFF_UP)) | |
544 | return true; | |
545 | ||
546 | /* Filter out packets that aren't for us. */ | |
547 | if (!(dev->flags & IFF_PROMISC) && | |
548 | !is_multicast_ether_addr(buf) && | |
7367d0b5 | 549 | !ether_addr_equal(dev->dev_addr, buf)) |
e3d62d7e CM |
550 | return true; |
551 | ||
552 | return false; | |
553 | } | |
554 | ||
555 | static void tile_net_receive_skb(struct net_device *dev, struct sk_buff *skb, | |
556 | gxio_mpipe_idesc_t *idesc, unsigned long len) | |
557 | { | |
eee8492d | 558 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
9ab5ec59 CM |
559 | struct tile_net_priv *priv = netdev_priv(dev); |
560 | int instance = priv->instance; | |
e3d62d7e CM |
561 | |
562 | /* Encode the actual packet length. */ | |
563 | skb_put(skb, len); | |
564 | ||
565 | skb->protocol = eth_type_trans(skb, dev); | |
566 | ||
567 | /* Acknowledge "good" hardware checksums. */ | |
568 | if (idesc->cs && idesc->csum_seed_val == 0xFFFF) | |
569 | skb->ip_summed = CHECKSUM_UNNECESSARY; | |
570 | ||
9ab5ec59 CM |
571 | /* Get RX timestamp from idesc. */ |
572 | tile_rx_timestamp(priv, skb, idesc); | |
573 | ||
f3286a3a | 574 | napi_gro_receive(&info->mpipe[instance].napi, skb); |
e3d62d7e CM |
575 | |
576 | /* Update stats. */ | |
ad018185 CM |
577 | tile_net_stats_add(1, &dev->stats.rx_packets); |
578 | tile_net_stats_add(len, &dev->stats.rx_bytes); | |
e3d62d7e CM |
579 | |
580 | /* Need a new buffer. */ | |
2628e8af | 581 | if (idesc->size == buffer_size_enums[0]) |
f3286a3a | 582 | info->mpipe[instance].num_needed_buffers[0]++; |
2628e8af | 583 | else if (idesc->size == buffer_size_enums[1]) |
f3286a3a | 584 | info->mpipe[instance].num_needed_buffers[1]++; |
e3d62d7e | 585 | else |
f3286a3a | 586 | info->mpipe[instance].num_needed_buffers[2]++; |
e3d62d7e CM |
587 | } |
588 | ||
589 | /* Handle a packet. Return true if "processed", false if "filtered". */ | |
f3286a3a | 590 | static bool tile_net_handle_packet(int instance, gxio_mpipe_idesc_t *idesc) |
e3d62d7e | 591 | { |
70b2776a | 592 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
f3286a3a CM |
593 | struct mpipe_data *md = &mpipe_data[instance]; |
594 | struct net_device *dev = md->tile_net_devs_for_channel[idesc->channel]; | |
e3d62d7e CM |
595 | uint8_t l2_offset; |
596 | void *va; | |
597 | void *buf; | |
598 | unsigned long len; | |
599 | bool filter; | |
600 | ||
2628e8af CM |
601 | /* Drop packets for which no buffer was available (which can |
602 | * happen under heavy load), or for which the me/tr/ce flags | |
603 | * are set (which can happen for jumbo cut-through packets, | |
604 | * or with a customized classifier). | |
e3d62d7e | 605 | */ |
2628e8af CM |
606 | if (idesc->be || idesc->me || idesc->tr || idesc->ce) { |
607 | if (dev) | |
ad018185 | 608 | tile_net_stats_add(1, &dev->stats.rx_errors); |
2628e8af | 609 | goto drop; |
e3d62d7e CM |
610 | } |
611 | ||
612 | /* Get the "l2_offset", if allowed. */ | |
4aa02644 | 613 | l2_offset = custom_flag ? 0 : gxio_mpipe_idesc_get_l2_offset(idesc); |
e3d62d7e | 614 | |
2628e8af CM |
615 | /* Get the VA (including NET_IP_ALIGN bytes of "headroom"). */ |
616 | va = tile_io_addr_to_va((unsigned long)idesc->va); | |
e3d62d7e CM |
617 | |
618 | /* Get the actual packet start/length. */ | |
619 | buf = va + l2_offset; | |
620 | len = idesc->l2_size - l2_offset; | |
621 | ||
622 | /* Point "va" at the raw buffer. */ | |
623 | va -= NET_IP_ALIGN; | |
624 | ||
625 | filter = filter_packet(dev, buf); | |
626 | if (filter) { | |
2628e8af | 627 | if (dev) |
ad018185 | 628 | tile_net_stats_add(1, &dev->stats.rx_dropped); |
2628e8af | 629 | drop: |
f3286a3a | 630 | gxio_mpipe_iqueue_drop(&info->mpipe[instance].iqueue, idesc); |
e3d62d7e CM |
631 | } else { |
632 | struct sk_buff *skb = mpipe_buf_to_skb(va); | |
633 | ||
634 | /* Skip headroom, and any custom header. */ | |
635 | skb_reserve(skb, NET_IP_ALIGN + l2_offset); | |
636 | ||
637 | tile_net_receive_skb(dev, skb, idesc, len); | |
638 | } | |
639 | ||
f3286a3a | 640 | gxio_mpipe_iqueue_consume(&info->mpipe[instance].iqueue, idesc); |
e3d62d7e CM |
641 | return !filter; |
642 | } | |
643 | ||
644 | /* Handle some packets for the current CPU. | |
645 | * | |
646 | * This function handles up to TILE_NET_BATCH idescs per call. | |
647 | * | |
648 | * ISSUE: Since we do not provide new buffers until this function is | |
649 | * complete, we must initially provide enough buffers for each network | |
650 | * cpu to fill its iqueue and also its batched idescs. | |
651 | * | |
652 | * ISSUE: The "rotting packet" race condition occurs if a packet | |
653 | * arrives after the queue appears to be empty, and before the | |
654 | * hypervisor interrupt is re-enabled. | |
655 | */ | |
656 | static int tile_net_poll(struct napi_struct *napi, int budget) | |
657 | { | |
70b2776a | 658 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
e3d62d7e CM |
659 | unsigned int work = 0; |
660 | gxio_mpipe_idesc_t *idesc; | |
f3286a3a CM |
661 | int instance, i, n; |
662 | struct mpipe_data *md; | |
663 | struct info_mpipe *info_mpipe = | |
664 | container_of(napi, struct info_mpipe, napi); | |
665 | ||
d110ec45 EB |
666 | if (budget <= 0) |
667 | goto done; | |
668 | ||
f3286a3a CM |
669 | instance = info_mpipe->instance; |
670 | while ((n = gxio_mpipe_iqueue_try_peek( | |
671 | &info_mpipe->iqueue, | |
672 | &idesc)) > 0) { | |
e3d62d7e CM |
673 | for (i = 0; i < n; i++) { |
674 | if (i == TILE_NET_BATCH) | |
675 | goto done; | |
f3286a3a CM |
676 | if (tile_net_handle_packet(instance, |
677 | idesc + i)) { | |
e3d62d7e CM |
678 | if (++work >= budget) |
679 | goto done; | |
680 | } | |
681 | } | |
682 | } | |
683 | ||
684 | /* There are no packets left. */ | |
6ad20165 | 685 | napi_complete_done(&info_mpipe->napi, work); |
e3d62d7e | 686 | |
f3286a3a | 687 | md = &mpipe_data[instance]; |
e3d62d7e | 688 | /* Re-enable hypervisor interrupts. */ |
f3286a3a CM |
689 | gxio_mpipe_enable_notif_ring_interrupt( |
690 | &md->context, info->mpipe[instance].iqueue.ring); | |
e3d62d7e CM |
691 | |
692 | /* HACK: Avoid the "rotting packet" problem. */ | |
f3286a3a CM |
693 | if (gxio_mpipe_iqueue_try_peek(&info_mpipe->iqueue, &idesc) > 0) |
694 | napi_schedule(&info_mpipe->napi); | |
e3d62d7e CM |
695 | |
696 | /* ISSUE: Handle completions? */ | |
697 | ||
698 | done: | |
699 | tile_net_provide_needed_buffers(); | |
700 | ||
701 | return work; | |
702 | } | |
703 | ||
f3286a3a CM |
704 | /* Handle an ingress interrupt from an instance on the current cpu. */ |
705 | static irqreturn_t tile_net_handle_ingress_irq(int irq, void *id) | |
e3d62d7e | 706 | { |
70b2776a | 707 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
f3286a3a | 708 | napi_schedule(&info->mpipe[(uint64_t)id].napi); |
e3d62d7e CM |
709 | return IRQ_HANDLED; |
710 | } | |
711 | ||
712 | /* Free some completions. This must be called with interrupts blocked. */ | |
713 | static int tile_net_free_comps(gxio_mpipe_equeue_t *equeue, | |
714 | struct tile_net_comps *comps, | |
715 | int limit, bool force_update) | |
716 | { | |
717 | int n = 0; | |
718 | while (comps->comp_last < comps->comp_next) { | |
719 | unsigned int cid = comps->comp_last % TILE_NET_MAX_COMPS; | |
720 | struct tile_net_comp *comp = &comps->comp_queue[cid]; | |
721 | if (!gxio_mpipe_equeue_is_complete(equeue, comp->when, | |
722 | force_update || n == 0)) | |
723 | break; | |
724 | dev_kfree_skb_irq(comp->skb); | |
725 | comps->comp_last++; | |
726 | if (++n == limit) | |
727 | break; | |
728 | } | |
729 | return n; | |
730 | } | |
731 | ||
732 | /* Add a completion. This must be called with interrupts blocked. | |
733 | * tile_net_equeue_try_reserve() will have ensured a free completion entry. | |
734 | */ | |
735 | static void add_comp(gxio_mpipe_equeue_t *equeue, | |
736 | struct tile_net_comps *comps, | |
737 | uint64_t when, struct sk_buff *skb) | |
738 | { | |
739 | int cid = comps->comp_next % TILE_NET_MAX_COMPS; | |
740 | comps->comp_queue[cid].when = when; | |
741 | comps->comp_queue[cid].skb = skb; | |
742 | comps->comp_next++; | |
743 | } | |
744 | ||
9b4c341b CM |
745 | static void tile_net_schedule_tx_wake_timer(struct net_device *dev, |
746 | int tx_queue_idx) | |
e3d62d7e | 747 | { |
9b4c341b | 748 | struct tile_net_info *info = &per_cpu(per_cpu_info, tx_queue_idx); |
e3d62d7e | 749 | struct tile_net_priv *priv = netdev_priv(dev); |
f3286a3a CM |
750 | int instance = priv->instance; |
751 | struct tile_net_tx_wake *tx_wake = | |
752 | &info->mpipe[instance].tx_wake[priv->echannel]; | |
e3d62d7e | 753 | |
9b4c341b | 754 | hrtimer_start(&tx_wake->timer, |
8b0e1953 | 755 | TX_TIMER_DELAY_USEC * 1000UL, |
e3d62d7e CM |
756 | HRTIMER_MODE_REL_PINNED); |
757 | } | |
758 | ||
759 | static enum hrtimer_restart tile_net_handle_tx_wake_timer(struct hrtimer *t) | |
760 | { | |
761 | struct tile_net_tx_wake *tx_wake = | |
762 | container_of(t, struct tile_net_tx_wake, timer); | |
9b4c341b | 763 | netif_wake_subqueue(tx_wake->dev, tx_wake->tx_queue_idx); |
e3d62d7e CM |
764 | return HRTIMER_NORESTART; |
765 | } | |
766 | ||
767 | /* Make sure the egress timer is scheduled. */ | |
768 | static void tile_net_schedule_egress_timer(void) | |
769 | { | |
70b2776a | 770 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
e3d62d7e CM |
771 | |
772 | if (!info->egress_timer_scheduled) { | |
773 | hrtimer_start(&info->egress_timer, | |
8b0e1953 | 774 | EGRESS_TIMER_DELAY_USEC * 1000UL, |
e3d62d7e CM |
775 | HRTIMER_MODE_REL_PINNED); |
776 | info->egress_timer_scheduled = true; | |
777 | } | |
778 | } | |
779 | ||
780 | /* The "function" for "info->egress_timer". | |
781 | * | |
782 | * This timer will reschedule itself as long as there are any pending | |
783 | * completions expected for this tile. | |
784 | */ | |
785 | static enum hrtimer_restart tile_net_handle_egress_timer(struct hrtimer *t) | |
786 | { | |
70b2776a | 787 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
e3d62d7e CM |
788 | unsigned long irqflags; |
789 | bool pending = false; | |
f3286a3a | 790 | int i, instance; |
e3d62d7e CM |
791 | |
792 | local_irq_save(irqflags); | |
793 | ||
794 | /* The timer is no longer scheduled. */ | |
795 | info->egress_timer_scheduled = false; | |
796 | ||
797 | /* Free all possible comps for this tile. */ | |
f3286a3a CM |
798 | for (instance = 0; instance < NR_MPIPE_MAX && |
799 | info->mpipe[instance].has_iqueue; instance++) { | |
800 | for (i = 0; i < TILE_NET_CHANNELS; i++) { | |
801 | struct tile_net_egress *egress = | |
802 | &mpipe_data[instance].egress_for_echannel[i]; | |
803 | struct tile_net_comps *comps = | |
804 | info->mpipe[instance].comps_for_echannel[i]; | |
805 | if (!egress || comps->comp_last >= comps->comp_next) | |
806 | continue; | |
807 | tile_net_free_comps(egress->equeue, comps, -1, true); | |
808 | pending = pending || | |
809 | (comps->comp_last < comps->comp_next); | |
810 | } | |
e3d62d7e CM |
811 | } |
812 | ||
813 | /* Reschedule timer if needed. */ | |
814 | if (pending) | |
815 | tile_net_schedule_egress_timer(); | |
816 | ||
817 | local_irq_restore(irqflags); | |
818 | ||
819 | return HRTIMER_NORESTART; | |
820 | } | |
821 | ||
9ab5ec59 CM |
822 | /* PTP clock operations. */ |
823 | ||
824 | static int ptp_mpipe_adjfreq(struct ptp_clock_info *ptp, s32 ppb) | |
825 | { | |
826 | int ret = 0; | |
827 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
828 | mutex_lock(&md->ptp_lock); | |
829 | if (gxio_mpipe_adjust_timestamp_freq(&md->context, ppb)) | |
830 | ret = -EINVAL; | |
831 | mutex_unlock(&md->ptp_lock); | |
832 | return ret; | |
833 | } | |
834 | ||
835 | static int ptp_mpipe_adjtime(struct ptp_clock_info *ptp, s64 delta) | |
836 | { | |
837 | int ret = 0; | |
838 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
839 | mutex_lock(&md->ptp_lock); | |
840 | if (gxio_mpipe_adjust_timestamp(&md->context, delta)) | |
841 | ret = -EBUSY; | |
842 | mutex_unlock(&md->ptp_lock); | |
843 | return ret; | |
844 | } | |
845 | ||
b9acf24f RC |
846 | static int ptp_mpipe_gettime(struct ptp_clock_info *ptp, |
847 | struct timespec64 *ts) | |
9ab5ec59 CM |
848 | { |
849 | int ret = 0; | |
850 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
851 | mutex_lock(&md->ptp_lock); | |
852 | if (gxio_mpipe_get_timestamp(&md->context, ts)) | |
853 | ret = -EBUSY; | |
854 | mutex_unlock(&md->ptp_lock); | |
855 | return ret; | |
856 | } | |
857 | ||
858 | static int ptp_mpipe_settime(struct ptp_clock_info *ptp, | |
b9acf24f | 859 | const struct timespec64 *ts) |
9ab5ec59 CM |
860 | { |
861 | int ret = 0; | |
862 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
863 | mutex_lock(&md->ptp_lock); | |
864 | if (gxio_mpipe_set_timestamp(&md->context, ts)) | |
865 | ret = -EBUSY; | |
866 | mutex_unlock(&md->ptp_lock); | |
867 | return ret; | |
868 | } | |
869 | ||
870 | static int ptp_mpipe_enable(struct ptp_clock_info *ptp, | |
871 | struct ptp_clock_request *request, int on) | |
872 | { | |
873 | return -EOPNOTSUPP; | |
874 | } | |
875 | ||
876 | static struct ptp_clock_info ptp_mpipe_caps = { | |
877 | .owner = THIS_MODULE, | |
878 | .name = "mPIPE clock", | |
879 | .max_adj = 999999999, | |
880 | .n_ext_ts = 0, | |
4986b4f0 | 881 | .n_pins = 0, |
9ab5ec59 CM |
882 | .pps = 0, |
883 | .adjfreq = ptp_mpipe_adjfreq, | |
884 | .adjtime = ptp_mpipe_adjtime, | |
b9acf24f RC |
885 | .gettime64 = ptp_mpipe_gettime, |
886 | .settime64 = ptp_mpipe_settime, | |
9ab5ec59 CM |
887 | .enable = ptp_mpipe_enable, |
888 | }; | |
889 | ||
9ab5ec59 CM |
890 | /* Sync mPIPE's timestamp up with Linux system time and register PTP clock. */ |
891 | static void register_ptp_clock(struct net_device *dev, struct mpipe_data *md) | |
892 | { | |
a9836cbb | 893 | struct timespec64 ts; |
9ab5ec59 | 894 | |
a9836cbb | 895 | ktime_get_ts64(&ts); |
9ab5ec59 CM |
896 | gxio_mpipe_set_timestamp(&md->context, &ts); |
897 | ||
898 | mutex_init(&md->ptp_lock); | |
899 | md->caps = ptp_mpipe_caps; | |
900 | md->ptp_clock = ptp_clock_register(&md->caps, NULL); | |
901 | if (IS_ERR(md->ptp_clock)) | |
902 | netdev_err(dev, "ptp_clock_register failed %ld\n", | |
903 | PTR_ERR(md->ptp_clock)); | |
9ab5ec59 CM |
904 | } |
905 | ||
906 | /* Initialize PTP fields in a new device. */ | |
907 | static void init_ptp_dev(struct tile_net_priv *priv) | |
908 | { | |
9ab5ec59 CM |
909 | priv->stamp_cfg.rx_filter = HWTSTAMP_FILTER_NONE; |
910 | priv->stamp_cfg.tx_type = HWTSTAMP_TX_OFF; | |
9ab5ec59 CM |
911 | } |
912 | ||
f3286a3a CM |
913 | /* Helper functions for "tile_net_update()". */ |
914 | static void enable_ingress_irq(void *irq) | |
e3d62d7e | 915 | { |
f3286a3a CM |
916 | enable_percpu_irq((long)irq, 0); |
917 | } | |
918 | ||
919 | static void disable_ingress_irq(void *irq) | |
920 | { | |
921 | disable_percpu_irq((long)irq); | |
e3d62d7e CM |
922 | } |
923 | ||
924 | /* Helper function for tile_net_open() and tile_net_stop(). | |
925 | * Always called under tile_net_devs_for_channel_mutex. | |
926 | */ | |
927 | static int tile_net_update(struct net_device *dev) | |
928 | { | |
929 | static gxio_mpipe_rules_t rules; /* too big to fit on the stack */ | |
930 | bool saw_channel = false; | |
f3286a3a CM |
931 | int instance = mpipe_instance(dev); |
932 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
933 | int channel; |
934 | int rc; | |
935 | int cpu; | |
936 | ||
f3286a3a CM |
937 | saw_channel = false; |
938 | gxio_mpipe_rules_init(&rules, &md->context); | |
e3d62d7e CM |
939 | |
940 | for (channel = 0; channel < TILE_NET_CHANNELS; channel++) { | |
f3286a3a | 941 | if (md->tile_net_devs_for_channel[channel] == NULL) |
e3d62d7e CM |
942 | continue; |
943 | if (!saw_channel) { | |
944 | saw_channel = true; | |
f3286a3a CM |
945 | gxio_mpipe_rules_begin(&rules, md->first_bucket, |
946 | md->num_buckets, NULL); | |
e3d62d7e CM |
947 | gxio_mpipe_rules_set_headroom(&rules, NET_IP_ALIGN); |
948 | } | |
949 | gxio_mpipe_rules_add_channel(&rules, channel); | |
950 | } | |
951 | ||
952 | /* NOTE: This can fail if there is no classifier. | |
953 | * ISSUE: Can anything else cause it to fail? | |
954 | */ | |
955 | rc = gxio_mpipe_rules_commit(&rules); | |
956 | if (rc != 0) { | |
f3286a3a CM |
957 | netdev_warn(dev, "gxio_mpipe_rules_commit: mpipe[%d] %d\n", |
958 | instance, rc); | |
e3d62d7e CM |
959 | return -EIO; |
960 | } | |
961 | ||
5e7a54a2 CM |
962 | /* Update all cpus, sequentially (to protect "netif_napi_add()"). |
963 | * We use on_each_cpu to handle the IPI mask or unmask. | |
964 | */ | |
965 | if (!saw_channel) | |
f3286a3a CM |
966 | on_each_cpu(disable_ingress_irq, |
967 | (void *)(long)(md->ingress_irq), 1); | |
5e7a54a2 CM |
968 | for_each_online_cpu(cpu) { |
969 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
f3286a3a CM |
970 | |
971 | if (!info->mpipe[instance].has_iqueue) | |
5e7a54a2 CM |
972 | continue; |
973 | if (saw_channel) { | |
f3286a3a CM |
974 | if (!info->mpipe[instance].napi_added) { |
975 | netif_napi_add(dev, &info->mpipe[instance].napi, | |
5e7a54a2 | 976 | tile_net_poll, TILE_NET_WEIGHT); |
f3286a3a | 977 | info->mpipe[instance].napi_added = true; |
5e7a54a2 | 978 | } |
f3286a3a CM |
979 | if (!info->mpipe[instance].napi_enabled) { |
980 | napi_enable(&info->mpipe[instance].napi); | |
981 | info->mpipe[instance].napi_enabled = true; | |
5e7a54a2 CM |
982 | } |
983 | } else { | |
f3286a3a CM |
984 | if (info->mpipe[instance].napi_enabled) { |
985 | napi_disable(&info->mpipe[instance].napi); | |
986 | info->mpipe[instance].napi_enabled = false; | |
5e7a54a2 CM |
987 | } |
988 | /* FIXME: Drain the iqueue. */ | |
989 | } | |
990 | } | |
991 | if (saw_channel) | |
f3286a3a CM |
992 | on_each_cpu(enable_ingress_irq, |
993 | (void *)(long)(md->ingress_irq), 1); | |
e3d62d7e CM |
994 | |
995 | /* HACK: Allow packets to flow in the simulator. */ | |
996 | if (saw_channel) | |
f3286a3a | 997 | sim_enable_mpipe_links(instance, -1); |
e3d62d7e CM |
998 | |
999 | return 0; | |
1000 | } | |
1001 | ||
2628e8af CM |
1002 | /* Initialize a buffer stack. */ |
1003 | static int create_buffer_stack(struct net_device *dev, | |
1004 | int kind, size_t num_buffers) | |
e3d62d7e CM |
1005 | { |
1006 | pte_t hash_pte = pte_set_home((pte_t) { 0 }, PAGE_HOME_HASH); | |
f3286a3a CM |
1007 | int instance = mpipe_instance(dev); |
1008 | struct mpipe_data *md = &mpipe_data[instance]; | |
2628e8af | 1009 | size_t needed = gxio_mpipe_calc_buffer_stack_bytes(num_buffers); |
f3286a3a | 1010 | int stack_idx = md->first_buffer_stack + kind; |
2628e8af CM |
1011 | void *va; |
1012 | int i, rc; | |
e3d62d7e | 1013 | |
2628e8af CM |
1014 | /* Round up to 64KB and then use alloc_pages() so we get the |
1015 | * required 64KB alignment. | |
e3d62d7e | 1016 | */ |
f3286a3a CM |
1017 | md->buffer_stack_bytes[kind] = |
1018 | ALIGN(needed, 64 * 1024); | |
e3d62d7e | 1019 | |
f3286a3a | 1020 | va = alloc_pages_exact(md->buffer_stack_bytes[kind], GFP_KERNEL); |
2628e8af | 1021 | if (va == NULL) { |
e3d62d7e | 1022 | netdev_err(dev, |
2628e8af | 1023 | "Could not alloc %zd bytes for buffer stack %d\n", |
f3286a3a | 1024 | md->buffer_stack_bytes[kind], kind); |
e3d62d7e CM |
1025 | return -ENOMEM; |
1026 | } | |
2628e8af CM |
1027 | |
1028 | /* Initialize the buffer stack. */ | |
f3286a3a CM |
1029 | rc = gxio_mpipe_init_buffer_stack(&md->context, stack_idx, |
1030 | buffer_size_enums[kind], va, | |
1031 | md->buffer_stack_bytes[kind], 0); | |
e3d62d7e | 1032 | if (rc != 0) { |
f3286a3a CM |
1033 | netdev_err(dev, "gxio_mpipe_init_buffer_stack: mpipe[%d] %d\n", |
1034 | instance, rc); | |
1035 | free_pages_exact(va, md->buffer_stack_bytes[kind]); | |
e3d62d7e CM |
1036 | return rc; |
1037 | } | |
2628e8af | 1038 | |
f3286a3a | 1039 | md->buffer_stack_vas[kind] = va; |
2628e8af | 1040 | |
f3286a3a | 1041 | rc = gxio_mpipe_register_client_memory(&md->context, stack_idx, |
e3d62d7e CM |
1042 | hash_pte, 0); |
1043 | if (rc != 0) { | |
f3286a3a CM |
1044 | netdev_err(dev, |
1045 | "gxio_mpipe_register_client_memory: mpipe[%d] %d\n", | |
1046 | instance, rc); | |
e3d62d7e CM |
1047 | return rc; |
1048 | } | |
1049 | ||
2628e8af CM |
1050 | /* Provide initial buffers. */ |
1051 | for (i = 0; i < num_buffers; i++) { | |
f3286a3a | 1052 | if (!tile_net_provide_buffer(instance, kind)) { |
2628e8af CM |
1053 | netdev_err(dev, "Cannot allocate initial sk_bufs!\n"); |
1054 | return -ENOMEM; | |
1055 | } | |
e3d62d7e | 1056 | } |
2628e8af CM |
1057 | |
1058 | return 0; | |
1059 | } | |
1060 | ||
1061 | /* Allocate and initialize mpipe buffer stacks, and register them in | |
1062 | * the mPIPE TLBs, for small, large, and (possibly) jumbo packet sizes. | |
1063 | * This routine supports tile_net_init_mpipe(), below. | |
1064 | */ | |
1065 | static int init_buffer_stacks(struct net_device *dev, | |
1066 | int network_cpus_count) | |
1067 | { | |
1068 | int num_kinds = MAX_KINDS - (jumbo_num == 0); | |
1069 | size_t num_buffers; | |
1070 | int rc; | |
f3286a3a CM |
1071 | int instance = mpipe_instance(dev); |
1072 | struct mpipe_data *md = &mpipe_data[instance]; | |
2628e8af CM |
1073 | |
1074 | /* Allocate the buffer stacks. */ | |
f3286a3a | 1075 | rc = gxio_mpipe_alloc_buffer_stacks(&md->context, num_kinds, 0, 0); |
2628e8af | 1076 | if (rc < 0) { |
f3286a3a CM |
1077 | netdev_err(dev, |
1078 | "gxio_mpipe_alloc_buffer_stacks: mpipe[%d] %d\n", | |
1079 | instance, rc); | |
e3d62d7e CM |
1080 | return rc; |
1081 | } | |
f3286a3a | 1082 | md->first_buffer_stack = rc; |
e3d62d7e | 1083 | |
2628e8af CM |
1084 | /* Enough small/large buffers to (normally) avoid buffer errors. */ |
1085 | num_buffers = | |
1086 | network_cpus_count * (IQUEUE_ENTRIES + TILE_NET_BATCH); | |
1087 | ||
1088 | /* Allocate the small memory stack. */ | |
1089 | if (rc >= 0) | |
1090 | rc = create_buffer_stack(dev, 0, num_buffers); | |
1091 | ||
1092 | /* Allocate the large buffer stack. */ | |
1093 | if (rc >= 0) | |
1094 | rc = create_buffer_stack(dev, 1, num_buffers); | |
1095 | ||
1096 | /* Allocate the jumbo buffer stack if needed. */ | |
1097 | if (rc >= 0 && jumbo_num != 0) | |
1098 | rc = create_buffer_stack(dev, 2, jumbo_num); | |
1099 | ||
1100 | return rc; | |
e3d62d7e CM |
1101 | } |
1102 | ||
1103 | /* Allocate per-cpu resources (memory for completions and idescs). | |
1104 | * This routine supports tile_net_init_mpipe(), below. | |
1105 | */ | |
1106 | static int alloc_percpu_mpipe_resources(struct net_device *dev, | |
1107 | int cpu, int ring) | |
1108 | { | |
1109 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
1110 | int order, i, rc; | |
f3286a3a CM |
1111 | int instance = mpipe_instance(dev); |
1112 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1113 | struct page *page; |
1114 | void *addr; | |
1115 | ||
1116 | /* Allocate the "comps". */ | |
1117 | order = get_order(COMPS_SIZE); | |
1118 | page = homecache_alloc_pages(GFP_KERNEL, order, cpu); | |
1119 | if (page == NULL) { | |
1120 | netdev_err(dev, "Failed to alloc %zd bytes comps memory\n", | |
1121 | COMPS_SIZE); | |
1122 | return -ENOMEM; | |
1123 | } | |
1124 | addr = pfn_to_kaddr(page_to_pfn(page)); | |
1125 | memset(addr, 0, COMPS_SIZE); | |
1126 | for (i = 0; i < TILE_NET_CHANNELS; i++) | |
f3286a3a | 1127 | info->mpipe[instance].comps_for_echannel[i] = |
e3d62d7e CM |
1128 | addr + i * sizeof(struct tile_net_comps); |
1129 | ||
1130 | /* If this is a network cpu, create an iqueue. */ | |
f9b531fe | 1131 | if (cpumask_test_cpu(cpu, &network_cpus_map)) { |
e3d62d7e CM |
1132 | order = get_order(NOTIF_RING_SIZE); |
1133 | page = homecache_alloc_pages(GFP_KERNEL, order, cpu); | |
1134 | if (page == NULL) { | |
1135 | netdev_err(dev, | |
1136 | "Failed to alloc %zd bytes iqueue memory\n", | |
1137 | NOTIF_RING_SIZE); | |
1138 | return -ENOMEM; | |
1139 | } | |
1140 | addr = pfn_to_kaddr(page_to_pfn(page)); | |
f3286a3a CM |
1141 | rc = gxio_mpipe_iqueue_init(&info->mpipe[instance].iqueue, |
1142 | &md->context, ring++, addr, | |
1143 | NOTIF_RING_SIZE, 0); | |
e3d62d7e CM |
1144 | if (rc < 0) { |
1145 | netdev_err(dev, | |
1146 | "gxio_mpipe_iqueue_init failed: %d\n", rc); | |
1147 | return rc; | |
1148 | } | |
f3286a3a | 1149 | info->mpipe[instance].has_iqueue = true; |
e3d62d7e CM |
1150 | } |
1151 | ||
1152 | return ring; | |
1153 | } | |
1154 | ||
1155 | /* Initialize NotifGroup and buckets. | |
1156 | * This routine supports tile_net_init_mpipe(), below. | |
1157 | */ | |
1158 | static int init_notif_group_and_buckets(struct net_device *dev, | |
1159 | int ring, int network_cpus_count) | |
1160 | { | |
1161 | int group, rc; | |
f3286a3a CM |
1162 | int instance = mpipe_instance(dev); |
1163 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1164 | |
1165 | /* Allocate one NotifGroup. */ | |
f3286a3a | 1166 | rc = gxio_mpipe_alloc_notif_groups(&md->context, 1, 0, 0); |
e3d62d7e | 1167 | if (rc < 0) { |
f3286a3a CM |
1168 | netdev_err(dev, "gxio_mpipe_alloc_notif_groups: mpipe[%d] %d\n", |
1169 | instance, rc); | |
e3d62d7e CM |
1170 | return rc; |
1171 | } | |
1172 | group = rc; | |
1173 | ||
1174 | /* Initialize global num_buckets value. */ | |
1175 | if (network_cpus_count > 4) | |
f3286a3a | 1176 | md->num_buckets = 256; |
e3d62d7e | 1177 | else if (network_cpus_count > 1) |
f3286a3a | 1178 | md->num_buckets = 16; |
e3d62d7e CM |
1179 | |
1180 | /* Allocate some buckets, and set global first_bucket value. */ | |
f3286a3a | 1181 | rc = gxio_mpipe_alloc_buckets(&md->context, md->num_buckets, 0, 0); |
e3d62d7e | 1182 | if (rc < 0) { |
f3286a3a CM |
1183 | netdev_err(dev, "gxio_mpipe_alloc_buckets: mpipe[%d] %d\n", |
1184 | instance, rc); | |
e3d62d7e CM |
1185 | return rc; |
1186 | } | |
f3286a3a | 1187 | md->first_bucket = rc; |
e3d62d7e CM |
1188 | |
1189 | /* Init group and buckets. */ | |
1190 | rc = gxio_mpipe_init_notif_group_and_buckets( | |
f3286a3a CM |
1191 | &md->context, group, ring, network_cpus_count, |
1192 | md->first_bucket, md->num_buckets, | |
e3d62d7e CM |
1193 | GXIO_MPIPE_BUCKET_STICKY_FLOW_LOCALITY); |
1194 | if (rc != 0) { | |
f3286a3a CM |
1195 | netdev_err(dev, "gxio_mpipe_init_notif_group_and_buckets: " |
1196 | "mpipe[%d] %d\n", instance, rc); | |
e3d62d7e CM |
1197 | return rc; |
1198 | } | |
1199 | ||
1200 | return 0; | |
1201 | } | |
1202 | ||
1203 | /* Create an irq and register it, then activate the irq and request | |
1204 | * interrupts on all cores. Note that "ingress_irq" being initialized | |
1205 | * is how we know not to call tile_net_init_mpipe() again. | |
1206 | * This routine supports tile_net_init_mpipe(), below. | |
1207 | */ | |
1208 | static int tile_net_setup_interrupts(struct net_device *dev) | |
1209 | { | |
f3286a3a CM |
1210 | int cpu, rc, irq; |
1211 | int instance = mpipe_instance(dev); | |
1212 | struct mpipe_data *md = &mpipe_data[instance]; | |
1213 | ||
1214 | irq = md->ingress_irq; | |
1215 | if (irq < 0) { | |
094e873c TG |
1216 | irq = irq_alloc_hwirq(-1); |
1217 | if (!irq) { | |
f3286a3a CM |
1218 | netdev_err(dev, |
1219 | "create_irq failed: mpipe[%d] %d\n", | |
1220 | instance, irq); | |
1221 | return irq; | |
1222 | } | |
1223 | tile_irq_activate(irq, TILE_IRQ_PERCPU); | |
e3d62d7e | 1224 | |
f3286a3a CM |
1225 | rc = request_irq(irq, tile_net_handle_ingress_irq, |
1226 | 0, "tile_net", (void *)((uint64_t)instance)); | |
1227 | ||
1228 | if (rc != 0) { | |
1229 | netdev_err(dev, "request_irq failed: mpipe[%d] %d\n", | |
1230 | instance, rc); | |
094e873c | 1231 | irq_free_hwirq(irq); |
f3286a3a CM |
1232 | return rc; |
1233 | } | |
1234 | md->ingress_irq = irq; | |
e3d62d7e CM |
1235 | } |
1236 | ||
1237 | for_each_online_cpu(cpu) { | |
1238 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
f3286a3a CM |
1239 | if (info->mpipe[instance].has_iqueue) { |
1240 | gxio_mpipe_request_notif_ring_interrupt(&md->context, | |
1241 | cpu_x(cpu), cpu_y(cpu), KERNEL_PL, irq, | |
1242 | info->mpipe[instance].iqueue.ring); | |
e3d62d7e CM |
1243 | } |
1244 | } | |
1245 | ||
1246 | return 0; | |
1247 | } | |
1248 | ||
1249 | /* Undo any state set up partially by a failed call to tile_net_init_mpipe. */ | |
f3286a3a | 1250 | static void tile_net_init_mpipe_fail(int instance) |
e3d62d7e | 1251 | { |
2628e8af | 1252 | int kind, cpu; |
f3286a3a | 1253 | struct mpipe_data *md = &mpipe_data[instance]; |
e3d62d7e CM |
1254 | |
1255 | /* Do cleanups that require the mpipe context first. */ | |
2628e8af | 1256 | for (kind = 0; kind < MAX_KINDS; kind++) { |
f3286a3a CM |
1257 | if (md->buffer_stack_vas[kind] != NULL) { |
1258 | tile_net_pop_all_buffers(instance, | |
1259 | md->first_buffer_stack + | |
1260 | kind); | |
2628e8af CM |
1261 | } |
1262 | } | |
e3d62d7e CM |
1263 | |
1264 | /* Destroy mpipe context so the hardware no longer owns any memory. */ | |
f3286a3a | 1265 | gxio_mpipe_destroy(&md->context); |
e3d62d7e CM |
1266 | |
1267 | for_each_online_cpu(cpu) { | |
1268 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
f3286a3a CM |
1269 | free_pages( |
1270 | (unsigned long)( | |
1271 | info->mpipe[instance].comps_for_echannel[0]), | |
1272 | get_order(COMPS_SIZE)); | |
1273 | info->mpipe[instance].comps_for_echannel[0] = NULL; | |
1274 | free_pages((unsigned long)(info->mpipe[instance].iqueue.idescs), | |
e3d62d7e | 1275 | get_order(NOTIF_RING_SIZE)); |
f3286a3a | 1276 | info->mpipe[instance].iqueue.idescs = NULL; |
e3d62d7e CM |
1277 | } |
1278 | ||
2628e8af | 1279 | for (kind = 0; kind < MAX_KINDS; kind++) { |
f3286a3a CM |
1280 | if (md->buffer_stack_vas[kind] != NULL) { |
1281 | free_pages_exact(md->buffer_stack_vas[kind], | |
1282 | md->buffer_stack_bytes[kind]); | |
1283 | md->buffer_stack_vas[kind] = NULL; | |
2628e8af CM |
1284 | } |
1285 | } | |
e3d62d7e | 1286 | |
f3286a3a CM |
1287 | md->first_buffer_stack = -1; |
1288 | md->first_bucket = -1; | |
e3d62d7e CM |
1289 | } |
1290 | ||
1291 | /* The first time any tilegx network device is opened, we initialize | |
1292 | * the global mpipe state. If this step fails, we fail to open the | |
1293 | * device, but if it succeeds, we never need to do it again, and since | |
1294 | * tile_net can't be unloaded, we never undo it. | |
1295 | * | |
1296 | * Note that some resources in this path (buffer stack indices, | |
1297 | * bindings from init_buffer_stack, etc.) are hypervisor resources | |
1298 | * that are freed implicitly by gxio_mpipe_destroy(). | |
1299 | */ | |
1300 | static int tile_net_init_mpipe(struct net_device *dev) | |
1301 | { | |
2628e8af | 1302 | int rc; |
e3d62d7e CM |
1303 | int cpu; |
1304 | int first_ring, ring; | |
f3286a3a CM |
1305 | int instance = mpipe_instance(dev); |
1306 | struct mpipe_data *md = &mpipe_data[instance]; | |
f9b531fe | 1307 | int network_cpus_count = cpumask_weight(&network_cpus_map); |
e3d62d7e CM |
1308 | |
1309 | if (!hash_default) { | |
1310 | netdev_err(dev, "Networking requires hash_default!\n"); | |
1311 | return -EIO; | |
1312 | } | |
1313 | ||
f3286a3a | 1314 | rc = gxio_mpipe_init(&md->context, instance); |
e3d62d7e | 1315 | if (rc != 0) { |
f3286a3a CM |
1316 | netdev_err(dev, "gxio_mpipe_init: mpipe[%d] %d\n", |
1317 | instance, rc); | |
e3d62d7e CM |
1318 | return -EIO; |
1319 | } | |
1320 | ||
1321 | /* Set up the buffer stacks. */ | |
2628e8af | 1322 | rc = init_buffer_stacks(dev, network_cpus_count); |
e3d62d7e CM |
1323 | if (rc != 0) |
1324 | goto fail; | |
1325 | ||
e3d62d7e | 1326 | /* Allocate one NotifRing for each network cpu. */ |
f3286a3a CM |
1327 | rc = gxio_mpipe_alloc_notif_rings(&md->context, |
1328 | network_cpus_count, 0, 0); | |
e3d62d7e CM |
1329 | if (rc < 0) { |
1330 | netdev_err(dev, "gxio_mpipe_alloc_notif_rings failed %d\n", | |
1331 | rc); | |
1332 | goto fail; | |
1333 | } | |
1334 | ||
1335 | /* Init NotifRings per-cpu. */ | |
1336 | first_ring = rc; | |
1337 | ring = first_ring; | |
1338 | for_each_online_cpu(cpu) { | |
1339 | rc = alloc_percpu_mpipe_resources(dev, cpu, ring); | |
1340 | if (rc < 0) | |
1341 | goto fail; | |
1342 | ring = rc; | |
1343 | } | |
1344 | ||
1345 | /* Initialize NotifGroup and buckets. */ | |
1346 | rc = init_notif_group_and_buckets(dev, first_ring, network_cpus_count); | |
1347 | if (rc != 0) | |
1348 | goto fail; | |
1349 | ||
1350 | /* Create and enable interrupts. */ | |
1351 | rc = tile_net_setup_interrupts(dev); | |
1352 | if (rc != 0) | |
1353 | goto fail; | |
1354 | ||
9ab5ec59 CM |
1355 | /* Register PTP clock and set mPIPE timestamp, if configured. */ |
1356 | register_ptp_clock(dev, md); | |
1357 | ||
e3d62d7e CM |
1358 | return 0; |
1359 | ||
1360 | fail: | |
f3286a3a | 1361 | tile_net_init_mpipe_fail(instance); |
e3d62d7e CM |
1362 | return rc; |
1363 | } | |
1364 | ||
1365 | /* Create persistent egress info for a given egress channel. | |
1366 | * Note that this may be shared between, say, "gbe0" and "xgbe0". | |
1367 | * ISSUE: Defer header allocation until TSO is actually needed? | |
1368 | */ | |
1369 | static int tile_net_init_egress(struct net_device *dev, int echannel) | |
1370 | { | |
2628e8af | 1371 | static int ering = -1; |
e3d62d7e CM |
1372 | struct page *headers_page, *edescs_page, *equeue_page; |
1373 | gxio_mpipe_edesc_t *edescs; | |
1374 | gxio_mpipe_equeue_t *equeue; | |
1375 | unsigned char *headers; | |
1376 | int headers_order, edescs_order, equeue_order; | |
1377 | size_t edescs_size; | |
e3d62d7e | 1378 | int rc = -ENOMEM; |
f3286a3a CM |
1379 | int instance = mpipe_instance(dev); |
1380 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1381 | |
1382 | /* Only initialize once. */ | |
f3286a3a | 1383 | if (md->egress_for_echannel[echannel].equeue != NULL) |
e3d62d7e CM |
1384 | return 0; |
1385 | ||
1386 | /* Allocate memory for the "headers". */ | |
1387 | headers_order = get_order(EQUEUE_ENTRIES * HEADER_BYTES); | |
1388 | headers_page = alloc_pages(GFP_KERNEL, headers_order); | |
1389 | if (headers_page == NULL) { | |
1390 | netdev_warn(dev, | |
1391 | "Could not alloc %zd bytes for TSO headers.\n", | |
1392 | PAGE_SIZE << headers_order); | |
1393 | goto fail; | |
1394 | } | |
1395 | headers = pfn_to_kaddr(page_to_pfn(headers_page)); | |
1396 | ||
1397 | /* Allocate memory for the "edescs". */ | |
1398 | edescs_size = EQUEUE_ENTRIES * sizeof(*edescs); | |
1399 | edescs_order = get_order(edescs_size); | |
1400 | edescs_page = alloc_pages(GFP_KERNEL, edescs_order); | |
1401 | if (edescs_page == NULL) { | |
1402 | netdev_warn(dev, | |
1403 | "Could not alloc %zd bytes for eDMA ring.\n", | |
1404 | edescs_size); | |
1405 | goto fail_headers; | |
1406 | } | |
1407 | edescs = pfn_to_kaddr(page_to_pfn(edescs_page)); | |
1408 | ||
1409 | /* Allocate memory for the "equeue". */ | |
1410 | equeue_order = get_order(sizeof(*equeue)); | |
1411 | equeue_page = alloc_pages(GFP_KERNEL, equeue_order); | |
1412 | if (equeue_page == NULL) { | |
1413 | netdev_warn(dev, | |
1414 | "Could not alloc %zd bytes for equeue info.\n", | |
1415 | PAGE_SIZE << equeue_order); | |
1416 | goto fail_edescs; | |
1417 | } | |
1418 | equeue = pfn_to_kaddr(page_to_pfn(equeue_page)); | |
1419 | ||
2628e8af CM |
1420 | /* Allocate an edma ring (using a one entry "free list"). */ |
1421 | if (ering < 0) { | |
f3286a3a | 1422 | rc = gxio_mpipe_alloc_edma_rings(&md->context, 1, 0, 0); |
2628e8af | 1423 | if (rc < 0) { |
f3286a3a CM |
1424 | netdev_warn(dev, "gxio_mpipe_alloc_edma_rings: " |
1425 | "mpipe[%d] %d\n", instance, rc); | |
2628e8af CM |
1426 | goto fail_equeue; |
1427 | } | |
1428 | ering = rc; | |
e3d62d7e | 1429 | } |
e3d62d7e CM |
1430 | |
1431 | /* Initialize the equeue. */ | |
f3286a3a | 1432 | rc = gxio_mpipe_equeue_init(equeue, &md->context, ering, echannel, |
e3d62d7e CM |
1433 | edescs, edescs_size, 0); |
1434 | if (rc != 0) { | |
f3286a3a CM |
1435 | netdev_err(dev, "gxio_mpipe_equeue_init: mpipe[%d] %d\n", |
1436 | instance, rc); | |
e3d62d7e CM |
1437 | goto fail_equeue; |
1438 | } | |
1439 | ||
2628e8af CM |
1440 | /* Don't reuse the ering later. */ |
1441 | ering = -1; | |
1442 | ||
1443 | if (jumbo_num != 0) { | |
1444 | /* Make sure "jumbo" packets can be egressed safely. */ | |
1445 | if (gxio_mpipe_equeue_set_snf_size(equeue, 10368) < 0) { | |
1446 | /* ISSUE: There is no "gxio_mpipe_equeue_destroy()". */ | |
1447 | netdev_warn(dev, "Jumbo packets may not be egressed" | |
1448 | " properly on channel %d\n", echannel); | |
1449 | } | |
1450 | } | |
1451 | ||
e3d62d7e | 1452 | /* Done. */ |
f3286a3a CM |
1453 | md->egress_for_echannel[echannel].equeue = equeue; |
1454 | md->egress_for_echannel[echannel].headers = headers; | |
e3d62d7e CM |
1455 | return 0; |
1456 | ||
1457 | fail_equeue: | |
1458 | __free_pages(equeue_page, equeue_order); | |
1459 | ||
1460 | fail_edescs: | |
1461 | __free_pages(edescs_page, edescs_order); | |
1462 | ||
1463 | fail_headers: | |
1464 | __free_pages(headers_page, headers_order); | |
1465 | ||
1466 | fail: | |
1467 | return rc; | |
1468 | } | |
1469 | ||
1470 | /* Return channel number for a newly-opened link. */ | |
1471 | static int tile_net_link_open(struct net_device *dev, gxio_mpipe_link_t *link, | |
1472 | const char *link_name) | |
1473 | { | |
f3286a3a CM |
1474 | int instance = mpipe_instance(dev); |
1475 | struct mpipe_data *md = &mpipe_data[instance]; | |
1476 | int rc = gxio_mpipe_link_open(link, &md->context, link_name, 0); | |
e3d62d7e | 1477 | if (rc < 0) { |
f3286a3a CM |
1478 | netdev_err(dev, "Failed to open '%s', mpipe[%d], %d\n", |
1479 | link_name, instance, rc); | |
e3d62d7e CM |
1480 | return rc; |
1481 | } | |
2628e8af CM |
1482 | if (jumbo_num != 0) { |
1483 | u32 attr = GXIO_MPIPE_LINK_RECEIVE_JUMBO; | |
1484 | rc = gxio_mpipe_link_set_attr(link, attr, 1); | |
1485 | if (rc != 0) { | |
1486 | netdev_err(dev, | |
1487 | "Cannot receive jumbo packets on '%s'\n", | |
1488 | link_name); | |
1489 | gxio_mpipe_link_close(link); | |
1490 | return rc; | |
1491 | } | |
1492 | } | |
e3d62d7e CM |
1493 | rc = gxio_mpipe_link_channel(link); |
1494 | if (rc < 0 || rc >= TILE_NET_CHANNELS) { | |
1495 | netdev_err(dev, "gxio_mpipe_link_channel bad value: %d\n", rc); | |
1496 | gxio_mpipe_link_close(link); | |
1497 | return -EINVAL; | |
1498 | } | |
1499 | return rc; | |
1500 | } | |
1501 | ||
1502 | /* Help the kernel activate the given network interface. */ | |
1503 | static int tile_net_open(struct net_device *dev) | |
1504 | { | |
1505 | struct tile_net_priv *priv = netdev_priv(dev); | |
f3286a3a | 1506 | int cpu, rc, instance; |
e3d62d7e CM |
1507 | |
1508 | mutex_lock(&tile_net_devs_for_channel_mutex); | |
1509 | ||
f3286a3a CM |
1510 | /* Get the instance info. */ |
1511 | rc = gxio_mpipe_link_instance(dev->name); | |
1155e964 WY |
1512 | if (rc < 0 || rc >= NR_MPIPE_MAX) { |
1513 | mutex_unlock(&tile_net_devs_for_channel_mutex); | |
f3286a3a | 1514 | return -EIO; |
1155e964 | 1515 | } |
f3286a3a CM |
1516 | |
1517 | priv->instance = rc; | |
1518 | instance = rc; | |
1519 | if (!mpipe_data[rc].context.mmio_fast_base) { | |
1520 | /* Do one-time initialization per instance the first time | |
1521 | * any device is opened. | |
1522 | */ | |
e3d62d7e CM |
1523 | rc = tile_net_init_mpipe(dev); |
1524 | if (rc != 0) | |
1525 | goto fail; | |
1526 | } | |
1527 | ||
1528 | /* Determine if this is the "loopify" device. */ | |
1529 | if (unlikely((loopify_link_name != NULL) && | |
1530 | !strcmp(dev->name, loopify_link_name))) { | |
1531 | rc = tile_net_link_open(dev, &priv->link, "loop0"); | |
1532 | if (rc < 0) | |
1533 | goto fail; | |
1534 | priv->channel = rc; | |
1535 | rc = tile_net_link_open(dev, &priv->loopify_link, "loop1"); | |
1536 | if (rc < 0) | |
1537 | goto fail; | |
1538 | priv->loopify_channel = rc; | |
1539 | priv->echannel = rc; | |
1540 | } else { | |
1541 | rc = tile_net_link_open(dev, &priv->link, dev->name); | |
1542 | if (rc < 0) | |
1543 | goto fail; | |
1544 | priv->channel = rc; | |
1545 | priv->echannel = rc; | |
1546 | } | |
1547 | ||
1548 | /* Initialize egress info (if needed). Once ever, per echannel. */ | |
1549 | rc = tile_net_init_egress(dev, priv->echannel); | |
1550 | if (rc != 0) | |
1551 | goto fail; | |
1552 | ||
f3286a3a | 1553 | mpipe_data[instance].tile_net_devs_for_channel[priv->channel] = dev; |
e3d62d7e CM |
1554 | |
1555 | rc = tile_net_update(dev); | |
1556 | if (rc != 0) | |
1557 | goto fail; | |
1558 | ||
1559 | mutex_unlock(&tile_net_devs_for_channel_mutex); | |
1560 | ||
1561 | /* Initialize the transmit wake timer for this device for each cpu. */ | |
1562 | for_each_online_cpu(cpu) { | |
1563 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
1564 | struct tile_net_tx_wake *tx_wake = | |
f3286a3a | 1565 | &info->mpipe[instance].tx_wake[priv->echannel]; |
e3d62d7e CM |
1566 | |
1567 | hrtimer_init(&tx_wake->timer, CLOCK_MONOTONIC, | |
1568 | HRTIMER_MODE_REL); | |
9b4c341b | 1569 | tx_wake->tx_queue_idx = cpu; |
e3d62d7e CM |
1570 | tx_wake->timer.function = tile_net_handle_tx_wake_timer; |
1571 | tx_wake->dev = dev; | |
1572 | } | |
1573 | ||
1574 | for_each_online_cpu(cpu) | |
1575 | netif_start_subqueue(dev, cpu); | |
1576 | netif_carrier_on(dev); | |
1577 | return 0; | |
1578 | ||
1579 | fail: | |
1580 | if (priv->loopify_channel >= 0) { | |
1581 | if (gxio_mpipe_link_close(&priv->loopify_link) != 0) | |
1582 | netdev_warn(dev, "Failed to close loopify link!\n"); | |
1583 | priv->loopify_channel = -1; | |
1584 | } | |
1585 | if (priv->channel >= 0) { | |
1586 | if (gxio_mpipe_link_close(&priv->link) != 0) | |
1587 | netdev_warn(dev, "Failed to close link!\n"); | |
1588 | priv->channel = -1; | |
1589 | } | |
1590 | priv->echannel = -1; | |
f3286a3a | 1591 | mpipe_data[instance].tile_net_devs_for_channel[priv->channel] = NULL; |
e3d62d7e CM |
1592 | mutex_unlock(&tile_net_devs_for_channel_mutex); |
1593 | ||
1594 | /* Don't return raw gxio error codes to generic Linux. */ | |
1595 | return (rc > -512) ? rc : -EIO; | |
1596 | } | |
1597 | ||
1598 | /* Help the kernel deactivate the given network interface. */ | |
1599 | static int tile_net_stop(struct net_device *dev) | |
1600 | { | |
1601 | struct tile_net_priv *priv = netdev_priv(dev); | |
1602 | int cpu; | |
f3286a3a CM |
1603 | int instance = priv->instance; |
1604 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1605 | |
1606 | for_each_online_cpu(cpu) { | |
1607 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
1608 | struct tile_net_tx_wake *tx_wake = | |
f3286a3a | 1609 | &info->mpipe[instance].tx_wake[priv->echannel]; |
e3d62d7e CM |
1610 | |
1611 | hrtimer_cancel(&tx_wake->timer); | |
1612 | netif_stop_subqueue(dev, cpu); | |
1613 | } | |
1614 | ||
1615 | mutex_lock(&tile_net_devs_for_channel_mutex); | |
f3286a3a | 1616 | md->tile_net_devs_for_channel[priv->channel] = NULL; |
e3d62d7e CM |
1617 | (void)tile_net_update(dev); |
1618 | if (priv->loopify_channel >= 0) { | |
1619 | if (gxio_mpipe_link_close(&priv->loopify_link) != 0) | |
1620 | netdev_warn(dev, "Failed to close loopify link!\n"); | |
1621 | priv->loopify_channel = -1; | |
1622 | } | |
1623 | if (priv->channel >= 0) { | |
1624 | if (gxio_mpipe_link_close(&priv->link) != 0) | |
1625 | netdev_warn(dev, "Failed to close link!\n"); | |
1626 | priv->channel = -1; | |
1627 | } | |
1628 | priv->echannel = -1; | |
1629 | mutex_unlock(&tile_net_devs_for_channel_mutex); | |
1630 | ||
1631 | return 0; | |
1632 | } | |
1633 | ||
1634 | /* Determine the VA for a fragment. */ | |
1635 | static inline void *tile_net_frag_buf(skb_frag_t *f) | |
1636 | { | |
1637 | unsigned long pfn = page_to_pfn(skb_frag_page(f)); | |
1638 | return pfn_to_kaddr(pfn) + f->page_offset; | |
1639 | } | |
1640 | ||
1641 | /* Acquire a completion entry and an egress slot, or if we can't, | |
1642 | * stop the queue and schedule the tx_wake timer. | |
1643 | */ | |
1644 | static s64 tile_net_equeue_try_reserve(struct net_device *dev, | |
9b4c341b | 1645 | int tx_queue_idx, |
e3d62d7e CM |
1646 | struct tile_net_comps *comps, |
1647 | gxio_mpipe_equeue_t *equeue, | |
1648 | int num_edescs) | |
1649 | { | |
1650 | /* Try to acquire a completion entry. */ | |
1651 | if (comps->comp_next - comps->comp_last < TILE_NET_MAX_COMPS - 1 || | |
1652 | tile_net_free_comps(equeue, comps, 32, false) != 0) { | |
1653 | ||
1654 | /* Try to acquire an egress slot. */ | |
1655 | s64 slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs); | |
1656 | if (slot >= 0) | |
1657 | return slot; | |
1658 | ||
1659 | /* Freeing some completions gives the equeue time to drain. */ | |
1660 | tile_net_free_comps(equeue, comps, TILE_NET_MAX_COMPS, false); | |
1661 | ||
1662 | slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs); | |
1663 | if (slot >= 0) | |
1664 | return slot; | |
1665 | } | |
1666 | ||
1667 | /* Still nothing; give up and stop the queue for a short while. */ | |
9b4c341b CM |
1668 | netif_stop_subqueue(dev, tx_queue_idx); |
1669 | tile_net_schedule_tx_wake_timer(dev, tx_queue_idx); | |
e3d62d7e CM |
1670 | return -1; |
1671 | } | |
1672 | ||
1673 | /* Determine how many edesc's are needed for TSO. | |
1674 | * | |
1675 | * Sometimes, if "sendfile()" requires copying, we will be called with | |
1676 | * "data" containing the header and payload, with "frags" being empty. | |
1677 | * Sometimes, for example when using NFS over TCP, a single segment can | |
1678 | * span 3 fragments. This requires special care. | |
1679 | */ | |
1680 | static int tso_count_edescs(struct sk_buff *skb) | |
1681 | { | |
1682 | struct skb_shared_info *sh = skb_shinfo(skb); | |
8388546e | 1683 | unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb); |
3da3fff8 | 1684 | unsigned int data_len = skb->len - sh_len; |
e3d62d7e CM |
1685 | unsigned int p_len = sh->gso_size; |
1686 | long f_id = -1; /* id of the current fragment */ | |
3da3fff8 CM |
1687 | long f_size = skb_headlen(skb) - sh_len; /* current fragment size */ |
1688 | long f_used = 0; /* bytes used from the current fragment */ | |
e3d62d7e CM |
1689 | long n; /* size of the current piece of payload */ |
1690 | int num_edescs = 0; | |
1691 | int segment; | |
1692 | ||
1693 | for (segment = 0; segment < sh->gso_segs; segment++) { | |
1694 | ||
1695 | unsigned int p_used = 0; | |
1696 | ||
1697 | /* One edesc for header and for each piece of the payload. */ | |
1698 | for (num_edescs++; p_used < p_len; num_edescs++) { | |
1699 | ||
1700 | /* Advance as needed. */ | |
1701 | while (f_used >= f_size) { | |
1702 | f_id++; | |
3da3fff8 | 1703 | f_size = skb_frag_size(&sh->frags[f_id]); |
e3d62d7e CM |
1704 | f_used = 0; |
1705 | } | |
1706 | ||
1707 | /* Use bytes from the current fragment. */ | |
1708 | n = p_len - p_used; | |
1709 | if (n > f_size - f_used) | |
1710 | n = f_size - f_used; | |
1711 | f_used += n; | |
1712 | p_used += n; | |
1713 | } | |
1714 | ||
1715 | /* The last segment may be less than gso_size. */ | |
1716 | data_len -= p_len; | |
1717 | if (data_len < p_len) | |
1718 | p_len = data_len; | |
1719 | } | |
1720 | ||
1721 | return num_edescs; | |
1722 | } | |
1723 | ||
2c7d04a9 | 1724 | /* Prepare modified copies of the skbuff headers. */ |
e3d62d7e CM |
1725 | static void tso_headers_prepare(struct sk_buff *skb, unsigned char *headers, |
1726 | s64 slot) | |
1727 | { | |
1728 | struct skb_shared_info *sh = skb_shinfo(skb); | |
1729 | struct iphdr *ih; | |
2c7d04a9 | 1730 | struct ipv6hdr *ih6; |
e3d62d7e | 1731 | struct tcphdr *th; |
8388546e | 1732 | unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb); |
3da3fff8 | 1733 | unsigned int data_len = skb->len - sh_len; |
e3d62d7e | 1734 | unsigned char *data = skb->data; |
8388546e | 1735 | unsigned int ih_off, th_off, p_len; |
444fa88a CM |
1736 | unsigned int isum_seed, tsum_seed, seq; |
1737 | unsigned int uninitialized_var(id); | |
2c7d04a9 | 1738 | int is_ipv6; |
e3d62d7e | 1739 | long f_id = -1; /* id of the current fragment */ |
3da3fff8 CM |
1740 | long f_size = skb_headlen(skb) - sh_len; /* current fragment size */ |
1741 | long f_used = 0; /* bytes used from the current fragment */ | |
e3d62d7e CM |
1742 | long n; /* size of the current piece of payload */ |
1743 | int segment; | |
1744 | ||
1745 | /* Locate original headers and compute various lengths. */ | |
2c7d04a9 CM |
1746 | is_ipv6 = skb_is_gso_v6(skb); |
1747 | if (is_ipv6) { | |
1748 | ih6 = ipv6_hdr(skb); | |
1749 | ih_off = skb_network_offset(skb); | |
1750 | } else { | |
1751 | ih = ip_hdr(skb); | |
1752 | ih_off = skb_network_offset(skb); | |
1753 | isum_seed = ((0xFFFF - ih->check) + | |
1754 | (0xFFFF - ih->tot_len) + | |
1755 | (0xFFFF - ih->id)); | |
1756 | id = ntohs(ih->id); | |
1757 | } | |
1758 | ||
e3d62d7e | 1759 | th = tcp_hdr(skb); |
e3d62d7e | 1760 | th_off = skb_transport_offset(skb); |
e3d62d7e CM |
1761 | p_len = sh->gso_size; |
1762 | ||
3da3fff8 | 1763 | tsum_seed = th->check + (0xFFFF ^ htons(skb->len)); |
e3d62d7e CM |
1764 | seq = ntohl(th->seq); |
1765 | ||
1766 | /* Prepare all the headers. */ | |
1767 | for (segment = 0; segment < sh->gso_segs; segment++) { | |
1768 | unsigned char *buf; | |
1769 | unsigned int p_used = 0; | |
1770 | ||
1771 | /* Copy to the header memory for this segment. */ | |
1772 | buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES + | |
1773 | NET_IP_ALIGN; | |
1774 | memcpy(buf, data, sh_len); | |
1775 | ||
1776 | /* Update copied ip header. */ | |
2c7d04a9 CM |
1777 | if (is_ipv6) { |
1778 | ih6 = (struct ipv6hdr *)(buf + ih_off); | |
1779 | ih6->payload_len = htons(sh_len + p_len - ih_off - | |
1780 | sizeof(*ih6)); | |
1781 | } else { | |
1782 | ih = (struct iphdr *)(buf + ih_off); | |
1783 | ih->tot_len = htons(sh_len + p_len - ih_off); | |
444fa88a | 1784 | ih->id = htons(id++); |
2c7d04a9 CM |
1785 | ih->check = csum_long(isum_seed + ih->tot_len + |
1786 | ih->id) ^ 0xffff; | |
1787 | } | |
e3d62d7e CM |
1788 | |
1789 | /* Update copied tcp header. */ | |
1790 | th = (struct tcphdr *)(buf + th_off); | |
1791 | th->seq = htonl(seq); | |
1792 | th->check = csum_long(tsum_seed + htons(sh_len + p_len)); | |
1793 | if (segment != sh->gso_segs - 1) { | |
1794 | th->fin = 0; | |
1795 | th->psh = 0; | |
1796 | } | |
1797 | ||
1798 | /* Skip past the header. */ | |
1799 | slot++; | |
1800 | ||
1801 | /* Skip past the payload. */ | |
1802 | while (p_used < p_len) { | |
1803 | ||
1804 | /* Advance as needed. */ | |
1805 | while (f_used >= f_size) { | |
1806 | f_id++; | |
3da3fff8 | 1807 | f_size = skb_frag_size(&sh->frags[f_id]); |
e3d62d7e CM |
1808 | f_used = 0; |
1809 | } | |
1810 | ||
1811 | /* Use bytes from the current fragment. */ | |
1812 | n = p_len - p_used; | |
1813 | if (n > f_size - f_used) | |
1814 | n = f_size - f_used; | |
1815 | f_used += n; | |
1816 | p_used += n; | |
1817 | ||
1818 | slot++; | |
1819 | } | |
1820 | ||
e3d62d7e CM |
1821 | seq += p_len; |
1822 | ||
1823 | /* The last segment may be less than gso_size. */ | |
1824 | data_len -= p_len; | |
1825 | if (data_len < p_len) | |
1826 | p_len = data_len; | |
1827 | } | |
1828 | ||
1829 | /* Flush the headers so they are ready for hardware DMA. */ | |
1830 | wmb(); | |
1831 | } | |
1832 | ||
1833 | /* Pass all the data to mpipe for egress. */ | |
1834 | static void tso_egress(struct net_device *dev, gxio_mpipe_equeue_t *equeue, | |
1835 | struct sk_buff *skb, unsigned char *headers, s64 slot) | |
1836 | { | |
e3d62d7e | 1837 | struct skb_shared_info *sh = skb_shinfo(skb); |
f3286a3a CM |
1838 | int instance = mpipe_instance(dev); |
1839 | struct mpipe_data *md = &mpipe_data[instance]; | |
8388546e | 1840 | unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb); |
3da3fff8 | 1841 | unsigned int data_len = skb->len - sh_len; |
e3d62d7e CM |
1842 | unsigned int p_len = sh->gso_size; |
1843 | gxio_mpipe_edesc_t edesc_head = { { 0 } }; | |
1844 | gxio_mpipe_edesc_t edesc_body = { { 0 } }; | |
1845 | long f_id = -1; /* id of the current fragment */ | |
3da3fff8 CM |
1846 | long f_size = skb_headlen(skb) - sh_len; /* current fragment size */ |
1847 | long f_used = 0; /* bytes used from the current fragment */ | |
1848 | void *f_data = skb->data + sh_len; | |
e3d62d7e CM |
1849 | long n; /* size of the current piece of payload */ |
1850 | unsigned long tx_packets = 0, tx_bytes = 0; | |
8388546e | 1851 | unsigned int csum_start; |
e3d62d7e CM |
1852 | int segment; |
1853 | ||
1854 | /* Prepare to egress the headers: set up header edesc. */ | |
1855 | csum_start = skb_checksum_start_offset(skb); | |
e3d62d7e CM |
1856 | edesc_head.csum = 1; |
1857 | edesc_head.csum_start = csum_start; | |
1858 | edesc_head.csum_dest = csum_start + skb->csum_offset; | |
1859 | edesc_head.xfer_size = sh_len; | |
1860 | ||
1861 | /* This is only used to specify the TLB. */ | |
f3286a3a CM |
1862 | edesc_head.stack_idx = md->first_buffer_stack; |
1863 | edesc_body.stack_idx = md->first_buffer_stack; | |
e3d62d7e CM |
1864 | |
1865 | /* Egress all the edescs. */ | |
1866 | for (segment = 0; segment < sh->gso_segs; segment++) { | |
e3d62d7e CM |
1867 | unsigned char *buf; |
1868 | unsigned int p_used = 0; | |
1869 | ||
1870 | /* Egress the header. */ | |
1871 | buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES + | |
1872 | NET_IP_ALIGN; | |
1873 | edesc_head.va = va_to_tile_io_addr(buf); | |
1874 | gxio_mpipe_equeue_put_at(equeue, edesc_head, slot); | |
1875 | slot++; | |
1876 | ||
1877 | /* Egress the payload. */ | |
1878 | while (p_used < p_len) { | |
3da3fff8 | 1879 | void *va; |
e3d62d7e CM |
1880 | |
1881 | /* Advance as needed. */ | |
1882 | while (f_used >= f_size) { | |
1883 | f_id++; | |
3da3fff8 | 1884 | f_size = skb_frag_size(&sh->frags[f_id]); |
8388546e | 1885 | f_data = tile_net_frag_buf(&sh->frags[f_id]); |
3da3fff8 | 1886 | f_used = 0; |
e3d62d7e CM |
1887 | } |
1888 | ||
3da3fff8 CM |
1889 | va = f_data + f_used; |
1890 | ||
e3d62d7e CM |
1891 | /* Use bytes from the current fragment. */ |
1892 | n = p_len - p_used; | |
1893 | if (n > f_size - f_used) | |
1894 | n = f_size - f_used; | |
1895 | f_used += n; | |
1896 | p_used += n; | |
1897 | ||
1898 | /* Egress a piece of the payload. */ | |
3da3fff8 | 1899 | edesc_body.va = va_to_tile_io_addr(va); |
e3d62d7e CM |
1900 | edesc_body.xfer_size = n; |
1901 | edesc_body.bound = !(p_used < p_len); | |
1902 | gxio_mpipe_equeue_put_at(equeue, edesc_body, slot); | |
1903 | slot++; | |
1904 | } | |
1905 | ||
1906 | tx_packets++; | |
1907 | tx_bytes += sh_len + p_len; | |
1908 | ||
1909 | /* The last segment may be less than gso_size. */ | |
1910 | data_len -= p_len; | |
1911 | if (data_len < p_len) | |
1912 | p_len = data_len; | |
1913 | } | |
1914 | ||
1915 | /* Update stats. */ | |
ad018185 CM |
1916 | tile_net_stats_add(tx_packets, &dev->stats.tx_packets); |
1917 | tile_net_stats_add(tx_bytes, &dev->stats.tx_bytes); | |
e3d62d7e CM |
1918 | } |
1919 | ||
1920 | /* Do "TSO" handling for egress. | |
1921 | * | |
1922 | * Normally drivers set NETIF_F_TSO only to support hardware TSO; | |
1923 | * otherwise the stack uses scatter-gather to implement GSO in software. | |
1924 | * On our testing, enabling GSO support (via NETIF_F_SG) drops network | |
1925 | * performance down to around 7.5 Gbps on the 10G interfaces, although | |
1926 | * also dropping cpu utilization way down, to under 8%. But | |
1927 | * implementing "TSO" in the driver brings performance back up to line | |
1928 | * rate, while dropping cpu usage even further, to less than 4%. In | |
1929 | * practice, profiling of GSO shows that skb_segment() is what causes | |
1930 | * the performance overheads; we benefit in the driver from using | |
1931 | * preallocated memory to duplicate the TCP/IP headers. | |
1932 | */ | |
1933 | static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev) | |
1934 | { | |
eee8492d | 1935 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
e3d62d7e CM |
1936 | struct tile_net_priv *priv = netdev_priv(dev); |
1937 | int channel = priv->echannel; | |
f3286a3a CM |
1938 | int instance = priv->instance; |
1939 | struct mpipe_data *md = &mpipe_data[instance]; | |
1940 | struct tile_net_egress *egress = &md->egress_for_echannel[channel]; | |
1941 | struct tile_net_comps *comps = | |
1942 | info->mpipe[instance].comps_for_echannel[channel]; | |
e3d62d7e CM |
1943 | gxio_mpipe_equeue_t *equeue = egress->equeue; |
1944 | unsigned long irqflags; | |
1945 | int num_edescs; | |
1946 | s64 slot; | |
1947 | ||
1948 | /* Determine how many mpipe edesc's are needed. */ | |
1949 | num_edescs = tso_count_edescs(skb); | |
1950 | ||
1951 | local_irq_save(irqflags); | |
1952 | ||
1953 | /* Try to acquire a completion entry and an egress slot. */ | |
9b4c341b CM |
1954 | slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps, |
1955 | equeue, num_edescs); | |
e3d62d7e CM |
1956 | if (slot < 0) { |
1957 | local_irq_restore(irqflags); | |
1958 | return NETDEV_TX_BUSY; | |
1959 | } | |
1960 | ||
1961 | /* Set up copies of header data properly. */ | |
1962 | tso_headers_prepare(skb, egress->headers, slot); | |
1963 | ||
1964 | /* Actually pass the data to the network hardware. */ | |
1965 | tso_egress(dev, equeue, skb, egress->headers, slot); | |
1966 | ||
1967 | /* Add a completion record. */ | |
1968 | add_comp(equeue, comps, slot + num_edescs - 1, skb); | |
1969 | ||
1970 | local_irq_restore(irqflags); | |
1971 | ||
1972 | /* Make sure the egress timer is scheduled. */ | |
1973 | tile_net_schedule_egress_timer(); | |
1974 | ||
1975 | return NETDEV_TX_OK; | |
1976 | } | |
1977 | ||
1978 | /* Analyze the body and frags for a transmit request. */ | |
1979 | static unsigned int tile_net_tx_frags(struct frag *frags, | |
1980 | struct sk_buff *skb, | |
1981 | void *b_data, unsigned int b_len) | |
1982 | { | |
1983 | unsigned int i, n = 0; | |
1984 | ||
1985 | struct skb_shared_info *sh = skb_shinfo(skb); | |
1986 | ||
1987 | if (b_len != 0) { | |
1988 | frags[n].buf = b_data; | |
1989 | frags[n++].length = b_len; | |
1990 | } | |
1991 | ||
1992 | for (i = 0; i < sh->nr_frags; i++) { | |
1993 | skb_frag_t *f = &sh->frags[i]; | |
1994 | frags[n].buf = tile_net_frag_buf(f); | |
1995 | frags[n++].length = skb_frag_size(f); | |
1996 | } | |
1997 | ||
1998 | return n; | |
1999 | } | |
2000 | ||
2001 | /* Help the kernel transmit a packet. */ | |
2002 | static int tile_net_tx(struct sk_buff *skb, struct net_device *dev) | |
2003 | { | |
70b2776a | 2004 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
e3d62d7e | 2005 | struct tile_net_priv *priv = netdev_priv(dev); |
f3286a3a CM |
2006 | int instance = priv->instance; |
2007 | struct mpipe_data *md = &mpipe_data[instance]; | |
2008 | struct tile_net_egress *egress = | |
2009 | &md->egress_for_echannel[priv->echannel]; | |
e3d62d7e CM |
2010 | gxio_mpipe_equeue_t *equeue = egress->equeue; |
2011 | struct tile_net_comps *comps = | |
f3286a3a | 2012 | info->mpipe[instance].comps_for_echannel[priv->echannel]; |
e3d62d7e CM |
2013 | unsigned int len = skb->len; |
2014 | unsigned char *data = skb->data; | |
2015 | unsigned int num_edescs; | |
2016 | struct frag frags[MAX_FRAGS]; | |
2017 | gxio_mpipe_edesc_t edescs[MAX_FRAGS]; | |
2018 | unsigned long irqflags; | |
2019 | gxio_mpipe_edesc_t edesc = { { 0 } }; | |
2020 | unsigned int i; | |
2021 | s64 slot; | |
2022 | ||
2023 | if (skb_is_gso(skb)) | |
2024 | return tile_net_tx_tso(skb, dev); | |
2025 | ||
2026 | num_edescs = tile_net_tx_frags(frags, skb, data, skb_headlen(skb)); | |
2027 | ||
2028 | /* This is only used to specify the TLB. */ | |
f3286a3a | 2029 | edesc.stack_idx = md->first_buffer_stack; |
e3d62d7e CM |
2030 | |
2031 | /* Prepare the edescs. */ | |
2032 | for (i = 0; i < num_edescs; i++) { | |
2033 | edesc.xfer_size = frags[i].length; | |
2034 | edesc.va = va_to_tile_io_addr(frags[i].buf); | |
2035 | edescs[i] = edesc; | |
2036 | } | |
2037 | ||
2038 | /* Mark the final edesc. */ | |
2039 | edescs[num_edescs - 1].bound = 1; | |
2040 | ||
2041 | /* Add checksum info to the initial edesc, if needed. */ | |
2042 | if (skb->ip_summed == CHECKSUM_PARTIAL) { | |
2043 | unsigned int csum_start = skb_checksum_start_offset(skb); | |
2044 | edescs[0].csum = 1; | |
2045 | edescs[0].csum_start = csum_start; | |
2046 | edescs[0].csum_dest = csum_start + skb->csum_offset; | |
2047 | } | |
2048 | ||
2049 | local_irq_save(irqflags); | |
2050 | ||
2051 | /* Try to acquire a completion entry and an egress slot. */ | |
9b4c341b CM |
2052 | slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps, |
2053 | equeue, num_edescs); | |
e3d62d7e CM |
2054 | if (slot < 0) { |
2055 | local_irq_restore(irqflags); | |
2056 | return NETDEV_TX_BUSY; | |
2057 | } | |
2058 | ||
2059 | for (i = 0; i < num_edescs; i++) | |
2060 | gxio_mpipe_equeue_put_at(equeue, edescs[i], slot++); | |
2061 | ||
9ab5ec59 CM |
2062 | /* Store TX timestamp if needed. */ |
2063 | tile_tx_timestamp(skb, instance); | |
2064 | ||
e3d62d7e CM |
2065 | /* Add a completion record. */ |
2066 | add_comp(equeue, comps, slot - 1, skb); | |
2067 | ||
2068 | /* NOTE: Use ETH_ZLEN for short packets (e.g. 42 < 60). */ | |
ad018185 | 2069 | tile_net_stats_add(1, &dev->stats.tx_packets); |
e3d62d7e | 2070 | tile_net_stats_add(max_t(unsigned int, len, ETH_ZLEN), |
ad018185 | 2071 | &dev->stats.tx_bytes); |
e3d62d7e CM |
2072 | |
2073 | local_irq_restore(irqflags); | |
2074 | ||
2075 | /* Make sure the egress timer is scheduled. */ | |
2076 | tile_net_schedule_egress_timer(); | |
2077 | ||
2078 | return NETDEV_TX_OK; | |
2079 | } | |
2080 | ||
2081 | /* Return subqueue id on this core (one per core). */ | |
f663dd9a | 2082 | static u16 tile_net_select_queue(struct net_device *dev, struct sk_buff *skb, |
99932d4f | 2083 | void *accel_priv, select_queue_fallback_t fallback) |
e3d62d7e CM |
2084 | { |
2085 | return smp_processor_id(); | |
2086 | } | |
2087 | ||
2088 | /* Deal with a transmit timeout. */ | |
2089 | static void tile_net_tx_timeout(struct net_device *dev) | |
2090 | { | |
2091 | int cpu; | |
2092 | ||
2093 | for_each_online_cpu(cpu) | |
2094 | netif_wake_subqueue(dev, cpu); | |
2095 | } | |
2096 | ||
2097 | /* Ioctl commands. */ | |
2098 | static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd) | |
2099 | { | |
9ab5ec59 | 2100 | if (cmd == SIOCSHWTSTAMP) |
6ab96d1e BH |
2101 | return tile_hwtstamp_set(dev, rq); |
2102 | if (cmd == SIOCGHWTSTAMP) | |
2103 | return tile_hwtstamp_get(dev, rq); | |
9ab5ec59 | 2104 | |
e3d62d7e CM |
2105 | return -EOPNOTSUPP; |
2106 | } | |
2107 | ||
e3d62d7e CM |
2108 | /* Change the Ethernet address of the NIC. |
2109 | * | |
2110 | * The hypervisor driver does not support changing MAC address. However, | |
2111 | * the hardware does not do anything with the MAC address, so the address | |
2112 | * which gets used on outgoing packets, and which is accepted on incoming | |
2113 | * packets, is completely up to us. | |
2114 | * | |
2115 | * Returns 0 on success, negative on failure. | |
2116 | */ | |
2117 | static int tile_net_set_mac_address(struct net_device *dev, void *p) | |
2118 | { | |
2119 | struct sockaddr *addr = p; | |
2120 | ||
2121 | if (!is_valid_ether_addr(addr->sa_data)) | |
2122 | return -EINVAL; | |
2123 | memcpy(dev->dev_addr, addr->sa_data, dev->addr_len); | |
2124 | return 0; | |
2125 | } | |
2126 | ||
2127 | #ifdef CONFIG_NET_POLL_CONTROLLER | |
2128 | /* Polling 'interrupt' - used by things like netconsole to send skbs | |
2129 | * without having to re-enable interrupts. It's not called while | |
2130 | * the interrupt routine is executing. | |
2131 | */ | |
2132 | static void tile_net_netpoll(struct net_device *dev) | |
2133 | { | |
f3286a3a | 2134 | int instance = mpipe_instance(dev); |
70b2776a | 2135 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
f3286a3a CM |
2136 | struct mpipe_data *md = &mpipe_data[instance]; |
2137 | ||
2138 | disable_percpu_irq(md->ingress_irq); | |
2139 | napi_schedule(&info->mpipe[instance].napi); | |
2140 | enable_percpu_irq(md->ingress_irq, 0); | |
e3d62d7e CM |
2141 | } |
2142 | #endif | |
2143 | ||
2144 | static const struct net_device_ops tile_net_ops = { | |
2145 | .ndo_open = tile_net_open, | |
2146 | .ndo_stop = tile_net_stop, | |
2147 | .ndo_start_xmit = tile_net_tx, | |
2148 | .ndo_select_queue = tile_net_select_queue, | |
2149 | .ndo_do_ioctl = tile_net_ioctl, | |
e3d62d7e CM |
2150 | .ndo_tx_timeout = tile_net_tx_timeout, |
2151 | .ndo_set_mac_address = tile_net_set_mac_address, | |
2152 | #ifdef CONFIG_NET_POLL_CONTROLLER | |
2153 | .ndo_poll_controller = tile_net_netpoll, | |
2154 | #endif | |
2155 | }; | |
2156 | ||
2157 | /* The setup function. | |
2158 | * | |
2159 | * This uses ether_setup() to assign various fields in dev, including | |
2160 | * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields. | |
2161 | */ | |
2162 | static void tile_net_setup(struct net_device *dev) | |
2163 | { | |
a8eaed55 CM |
2164 | netdev_features_t features = 0; |
2165 | ||
e3d62d7e CM |
2166 | ether_setup(dev); |
2167 | dev->netdev_ops = &tile_net_ops; | |
2168 | dev->watchdog_timeo = TILE_NET_TIMEOUT; | |
f58f0817 JW |
2169 | |
2170 | /* MTU range: 68 - 1500 or 9000 */ | |
2171 | dev->mtu = ETH_DATA_LEN; | |
2172 | dev->min_mtu = ETH_MIN_MTU; | |
2173 | dev->max_mtu = jumbo_num ? TILE_JUMBO_MAX_MTU : ETH_DATA_LEN; | |
a8eaed55 | 2174 | |
a8eaed55 CM |
2175 | features |= NETIF_F_HW_CSUM; |
2176 | features |= NETIF_F_SG; | |
2177 | features |= NETIF_F_TSO; | |
2c7d04a9 | 2178 | features |= NETIF_F_TSO6; |
a8eaed55 CM |
2179 | |
2180 | dev->hw_features |= features; | |
2181 | dev->vlan_features |= features; | |
2182 | dev->features |= features; | |
e3d62d7e CM |
2183 | } |
2184 | ||
2185 | /* Allocate the device structure, register the device, and obtain the | |
2186 | * MAC address from the hypervisor. | |
2187 | */ | |
2188 | static void tile_net_dev_init(const char *name, const uint8_t *mac) | |
2189 | { | |
2190 | int ret; | |
e3d62d7e CM |
2191 | struct net_device *dev; |
2192 | struct tile_net_priv *priv; | |
2193 | ||
2194 | /* HACK: Ignore "loop" links. */ | |
2195 | if (strncmp(name, "loop", 4) == 0) | |
2196 | return; | |
2197 | ||
2198 | /* Allocate the device structure. Normally, "name" is a | |
2199 | * template, instantiated by register_netdev(), but not for us. | |
2200 | */ | |
c835a677 TG |
2201 | dev = alloc_netdev_mqs(sizeof(*priv), name, NET_NAME_UNKNOWN, |
2202 | tile_net_setup, NR_CPUS, 1); | |
e3d62d7e CM |
2203 | if (!dev) { |
2204 | pr_err("alloc_netdev_mqs(%s) failed\n", name); | |
2205 | return; | |
2206 | } | |
2207 | ||
2208 | /* Initialize "priv". */ | |
2209 | priv = netdev_priv(dev); | |
e3d62d7e CM |
2210 | priv->dev = dev; |
2211 | priv->channel = -1; | |
2212 | priv->loopify_channel = -1; | |
2213 | priv->echannel = -1; | |
9ab5ec59 | 2214 | init_ptp_dev(priv); |
e3d62d7e CM |
2215 | |
2216 | /* Get the MAC address and set it in the device struct; this must | |
2217 | * be done before the device is opened. If the MAC is all zeroes, | |
2218 | * we use a random address, since we're probably on the simulator. | |
2219 | */ | |
d581ebf5 TK |
2220 | if (!is_zero_ether_addr(mac)) |
2221 | ether_addr_copy(dev->dev_addr, mac); | |
2222 | else | |
c8ab13fb | 2223 | eth_hw_addr_random(dev); |
e3d62d7e CM |
2224 | |
2225 | /* Register the network device. */ | |
2226 | ret = register_netdev(dev); | |
2227 | if (ret) { | |
2228 | netdev_err(dev, "register_netdev failed %d\n", ret); | |
2229 | free_netdev(dev); | |
2230 | return; | |
2231 | } | |
2232 | } | |
2233 | ||
2234 | /* Per-cpu module initialization. */ | |
2235 | static void tile_net_init_module_percpu(void *unused) | |
2236 | { | |
70b2776a | 2237 | struct tile_net_info *info = this_cpu_ptr(&per_cpu_info); |
e3d62d7e | 2238 | int my_cpu = smp_processor_id(); |
f3286a3a | 2239 | int instance; |
e3d62d7e | 2240 | |
f3286a3a CM |
2241 | for (instance = 0; instance < NR_MPIPE_MAX; instance++) { |
2242 | info->mpipe[instance].has_iqueue = false; | |
2243 | info->mpipe[instance].instance = instance; | |
2244 | } | |
e3d62d7e CM |
2245 | info->my_cpu = my_cpu; |
2246 | ||
2247 | /* Initialize the egress timer. */ | |
2248 | hrtimer_init(&info->egress_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2249 | info->egress_timer.function = tile_net_handle_egress_timer; | |
2250 | } | |
2251 | ||
2252 | /* Module initialization. */ | |
2253 | static int __init tile_net_init_module(void) | |
2254 | { | |
2255 | int i; | |
2256 | char name[GXIO_MPIPE_LINK_NAME_LEN]; | |
2257 | uint8_t mac[6]; | |
2258 | ||
2259 | pr_info("Tilera Network Driver\n"); | |
2260 | ||
f3286a3a CM |
2261 | BUILD_BUG_ON(NR_MPIPE_MAX != 2); |
2262 | ||
e3d62d7e CM |
2263 | mutex_init(&tile_net_devs_for_channel_mutex); |
2264 | ||
2265 | /* Initialize each CPU. */ | |
2266 | on_each_cpu(tile_net_init_module_percpu, NULL, 1); | |
2267 | ||
2268 | /* Find out what devices we have, and initialize them. */ | |
2269 | for (i = 0; gxio_mpipe_link_enumerate_mac(i, name, mac) >= 0; i++) | |
2270 | tile_net_dev_init(name, mac); | |
2271 | ||
2272 | if (!network_cpus_init()) | |
03f6199a CM |
2273 | cpumask_and(&network_cpus_map, housekeeping_cpumask(), |
2274 | cpu_online_mask); | |
e3d62d7e CM |
2275 | |
2276 | return 0; | |
2277 | } | |
2278 | ||
2279 | module_init(tile_net_init_module); |