]>
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> | |
e3d62d7e CM |
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 | ||
2628e8af CM |
82 | /* The "kinds" of buffer stacks (small/large/jumbo). */ |
83 | #define MAX_KINDS 3 | |
84 | ||
e3d62d7e CM |
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 | ||
103 | MODULE_AUTHOR("Tilera Corporation"); | |
104 | MODULE_LICENSE("GPL"); | |
105 | ||
106 | /* A "packet fragment" (a chunk of memory). */ | |
107 | struct frag { | |
108 | void *buf; | |
109 | size_t length; | |
110 | }; | |
111 | ||
112 | /* A single completion. */ | |
113 | struct 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. */ | |
121 | struct 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. */ | |
131 | struct tile_net_tx_wake { | |
9b4c341b | 132 | int tx_queue_idx; |
e3d62d7e CM |
133 | struct hrtimer timer; |
134 | struct net_device *dev; | |
135 | }; | |
136 | ||
137 | /* Info for a specific cpu. */ | |
138 | struct tile_net_info { | |
e3d62d7e CM |
139 | /* Our cpu. */ |
140 | int my_cpu; | |
e3d62d7e CM |
141 | /* A timer for handling egress completions. */ |
142 | struct hrtimer egress_timer; | |
143 | /* True if "egress_timer" is scheduled. */ | |
144 | bool egress_timer_scheduled; | |
f3286a3a CM |
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]; | |
e3d62d7e CM |
164 | }; |
165 | ||
166 | /* Info for egress on a particular egress channel. */ | |
167 | struct 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. */ | |
175 | struct 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; | |
f3286a3a CM |
188 | /* mPIPE instance, 0 or 1. */ |
189 | int instance; | |
9ab5ec59 CM |
190 | /* The timestamp config. */ |
191 | struct hwtstamp_config stamp_cfg; | |
e3d62d7e CM |
192 | }; |
193 | ||
f3286a3a CM |
194 | static struct mpipe_data { |
195 | /* The ingress irq. */ | |
196 | int ingress_irq; | |
e3d62d7e | 197 | |
f3286a3a CM |
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 | ||
9ab5ec59 CM |
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; | |
9ab5ec59 | 236 | |
f3286a3a CM |
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 | }; | |
e3d62d7e CM |
245 | |
246 | /* A mutex for "tile_net_devs_for_channel". */ | |
247 | static DEFINE_MUTEX(tile_net_devs_for_channel_mutex); | |
248 | ||
249 | /* The per-cpu info. */ | |
250 | static DEFINE_PER_CPU(struct tile_net_info, per_cpu_info); | |
251 | ||
e3d62d7e | 252 | |
2628e8af | 253 | /* The buffer size enums for each buffer stack. |
e3d62d7e | 254 | * See arch/tile/include/gxio/mpipe.h for the set of possible values. |
2628e8af CM |
255 | * We avoid the "10384" size because it can induce "false chaining" |
256 | * on "cut-through" jumbo packets. | |
e3d62d7e | 257 | */ |
2628e8af CM |
258 | static 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 | }; | |
e3d62d7e | 263 | |
e3d62d7e CM |
264 | /* Text value of tile_net.cpus if passed as a module parameter. */ |
265 | static char *network_cpus_string; | |
266 | ||
267 | /* The actual cpus in "network_cpus". */ | |
268 | static struct cpumask network_cpus_map; | |
269 | ||
4aa02644 | 270 | /* If "tile_net.loopify=LINK" was specified, this is "LINK". */ |
e3d62d7e CM |
271 | static char *loopify_link_name; |
272 | ||
4aa02644 CM |
273 | /* If "tile_net.custom" was specified, this is true. */ |
274 | static bool custom_flag; | |
e3d62d7e | 275 | |
2628e8af CM |
276 | /* If "tile_net.jumbo=NUM" was specified, this is "NUM". */ |
277 | static uint jumbo_num; | |
278 | ||
f3286a3a CM |
279 | /* Obtain mpipe instance from struct tile_net_priv given struct net_device. */ |
280 | static inline int mpipe_instance(struct net_device *dev) | |
281 | { | |
282 | struct tile_net_priv *priv = netdev_priv(dev); | |
283 | return priv->instance; | |
284 | } | |
285 | ||
e3d62d7e CM |
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 | */ | |
293 | static bool network_cpus_init(void) | |
294 | { | |
295 | char buf[1024]; | |
296 | int rc; | |
297 | ||
298 | if (network_cpus_string == NULL) | |
299 | return false; | |
300 | ||
301 | rc = cpulist_parse_crop(network_cpus_string, &network_cpus_map); | |
302 | if (rc != 0) { | |
303 | pr_warn("tile_net.cpus=%s: malformed cpu list\n", | |
304 | network_cpus_string); | |
305 | return false; | |
306 | } | |
307 | ||
308 | /* Remove dedicated cpus. */ | |
309 | cpumask_and(&network_cpus_map, &network_cpus_map, cpu_possible_mask); | |
310 | ||
311 | if (cpumask_empty(&network_cpus_map)) { | |
312 | pr_warn("Ignoring empty tile_net.cpus='%s'.\n", | |
313 | network_cpus_string); | |
314 | return false; | |
315 | } | |
316 | ||
317 | cpulist_scnprintf(buf, sizeof(buf), &network_cpus_map); | |
318 | pr_info("Linux network CPUs: %s\n", buf); | |
319 | return true; | |
320 | } | |
321 | ||
322 | module_param_named(cpus, network_cpus_string, charp, 0444); | |
323 | MODULE_PARM_DESC(cpus, "cpulist of cores that handle network interrupts"); | |
324 | ||
325 | /* The "tile_net.loopify=LINK" argument causes the named device to | |
326 | * actually use "loop0" for ingress, and "loop1" for egress. This | |
327 | * allows an app to sit between the actual link and linux, passing | |
328 | * (some) packets along to linux, and forwarding (some) packets sent | |
329 | * out by linux. | |
330 | */ | |
331 | module_param_named(loopify, loopify_link_name, charp, 0444); | |
332 | MODULE_PARM_DESC(loopify, "name the device to use loop0/1 for ingress/egress"); | |
333 | ||
334 | /* The "tile_net.custom" argument causes us to ignore the "conventional" | |
335 | * classifier metadata, in particular, the "l2_offset". | |
336 | */ | |
4aa02644 | 337 | module_param_named(custom, custom_flag, bool, 0444); |
e3d62d7e CM |
338 | MODULE_PARM_DESC(custom, "indicates a (heavily) customized classifier"); |
339 | ||
2628e8af CM |
340 | /* The "tile_net.jumbo" argument causes us to support "jumbo" packets, |
341 | * and to allocate the given number of "jumbo" buffers. | |
342 | */ | |
343 | module_param_named(jumbo, jumbo_num, uint, 0444); | |
344 | MODULE_PARM_DESC(jumbo, "the number of buffers to support jumbo packets"); | |
345 | ||
e3d62d7e CM |
346 | /* Atomically update a statistics field. |
347 | * Note that on TILE-Gx, this operation is fire-and-forget on the | |
348 | * issuing core (single-cycle dispatch) and takes only a few cycles | |
349 | * longer than a regular store when the request reaches the home cache. | |
350 | * No expensive bus management overhead is required. | |
351 | */ | |
352 | static void tile_net_stats_add(unsigned long value, unsigned long *field) | |
353 | { | |
354 | BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(unsigned long)); | |
355 | atomic_long_add(value, (atomic_long_t *)field); | |
356 | } | |
357 | ||
358 | /* Allocate and push a buffer. */ | |
f3286a3a | 359 | static bool tile_net_provide_buffer(int instance, int kind) |
e3d62d7e | 360 | { |
f3286a3a | 361 | struct mpipe_data *md = &mpipe_data[instance]; |
2628e8af CM |
362 | gxio_mpipe_buffer_size_enum_t bse = buffer_size_enums[kind]; |
363 | size_t bs = gxio_mpipe_buffer_size_enum_to_buffer_size(bse); | |
e3d62d7e CM |
364 | const unsigned long buffer_alignment = 128; |
365 | struct sk_buff *skb; | |
366 | int len; | |
367 | ||
2628e8af | 368 | len = sizeof(struct sk_buff **) + buffer_alignment + bs; |
e3d62d7e CM |
369 | skb = dev_alloc_skb(len); |
370 | if (skb == NULL) | |
371 | return false; | |
372 | ||
373 | /* Make room for a back-pointer to 'skb' and guarantee alignment. */ | |
374 | skb_reserve(skb, sizeof(struct sk_buff **)); | |
375 | skb_reserve(skb, -(long)skb->data & (buffer_alignment - 1)); | |
376 | ||
377 | /* Save a back-pointer to 'skb'. */ | |
378 | *(struct sk_buff **)(skb->data - sizeof(struct sk_buff **)) = skb; | |
379 | ||
380 | /* Make sure "skb" and the back-pointer have been flushed. */ | |
381 | wmb(); | |
382 | ||
f3286a3a | 383 | gxio_mpipe_push_buffer(&md->context, md->first_buffer_stack + kind, |
e3d62d7e CM |
384 | (void *)va_to_tile_io_addr(skb->data)); |
385 | ||
386 | return true; | |
387 | } | |
388 | ||
389 | /* Convert a raw mpipe buffer to its matching skb pointer. */ | |
390 | static struct sk_buff *mpipe_buf_to_skb(void *va) | |
391 | { | |
392 | /* Acquire the associated "skb". */ | |
393 | struct sk_buff **skb_ptr = va - sizeof(*skb_ptr); | |
394 | struct sk_buff *skb = *skb_ptr; | |
395 | ||
396 | /* Paranoia. */ | |
397 | if (skb->data != va) { | |
398 | /* Panic here since there's a reasonable chance | |
399 | * that corrupt buffers means generic memory | |
400 | * corruption, with unpredictable system effects. | |
401 | */ | |
402 | panic("Corrupt linux buffer! va=%p, skb=%p, skb->data=%p", | |
403 | va, skb, skb->data); | |
404 | } | |
405 | ||
406 | return skb; | |
407 | } | |
408 | ||
f3286a3a | 409 | static void tile_net_pop_all_buffers(int instance, int stack) |
e3d62d7e | 410 | { |
f3286a3a CM |
411 | struct mpipe_data *md = &mpipe_data[instance]; |
412 | ||
e3d62d7e CM |
413 | for (;;) { |
414 | tile_io_addr_t addr = | |
f3286a3a CM |
415 | (tile_io_addr_t)gxio_mpipe_pop_buffer(&md->context, |
416 | stack); | |
e3d62d7e CM |
417 | if (addr == 0) |
418 | break; | |
419 | dev_kfree_skb_irq(mpipe_buf_to_skb(tile_io_addr_to_va(addr))); | |
420 | } | |
421 | } | |
422 | ||
423 | /* Provide linux buffers to mPIPE. */ | |
424 | static void tile_net_provide_needed_buffers(void) | |
425 | { | |
426 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
f3286a3a CM |
427 | int instance, kind; |
428 | for (instance = 0; instance < NR_MPIPE_MAX && | |
429 | info->mpipe[instance].has_iqueue; instance++) { | |
430 | for (kind = 0; kind < MAX_KINDS; kind++) { | |
431 | while (info->mpipe[instance].num_needed_buffers[kind] | |
432 | != 0) { | |
433 | if (!tile_net_provide_buffer(instance, kind)) { | |
434 | pr_notice("Tile %d still needs" | |
435 | " some buffers\n", | |
436 | info->my_cpu); | |
437 | return; | |
438 | } | |
439 | info->mpipe[instance]. | |
440 | num_needed_buffers[kind]--; | |
2628e8af | 441 | } |
2628e8af | 442 | } |
e3d62d7e | 443 | } |
e3d62d7e CM |
444 | } |
445 | ||
9ab5ec59 CM |
446 | /* Get RX timestamp, and store it in the skb. */ |
447 | static void tile_rx_timestamp(struct tile_net_priv *priv, struct sk_buff *skb, | |
448 | gxio_mpipe_idesc_t *idesc) | |
449 | { | |
9ab5ec59 CM |
450 | if (unlikely(priv->stamp_cfg.rx_filter != HWTSTAMP_FILTER_NONE)) { |
451 | struct skb_shared_hwtstamps *shhwtstamps = skb_hwtstamps(skb); | |
452 | memset(shhwtstamps, 0, sizeof(*shhwtstamps)); | |
453 | shhwtstamps->hwtstamp = ktime_set(idesc->time_stamp_sec, | |
454 | idesc->time_stamp_ns); | |
455 | } | |
9ab5ec59 CM |
456 | } |
457 | ||
458 | /* Get TX timestamp, and store it in the skb. */ | |
459 | static void tile_tx_timestamp(struct sk_buff *skb, int instance) | |
460 | { | |
9ab5ec59 CM |
461 | struct skb_shared_info *shtx = skb_shinfo(skb); |
462 | if (unlikely((shtx->tx_flags & SKBTX_HW_TSTAMP) != 0)) { | |
463 | struct mpipe_data *md = &mpipe_data[instance]; | |
464 | struct skb_shared_hwtstamps shhwtstamps; | |
465 | struct timespec ts; | |
466 | ||
467 | shtx->tx_flags |= SKBTX_IN_PROGRESS; | |
468 | gxio_mpipe_get_timestamp(&md->context, &ts); | |
469 | memset(&shhwtstamps, 0, sizeof(shhwtstamps)); | |
470 | shhwtstamps.hwtstamp = ktime_set(ts.tv_sec, ts.tv_nsec); | |
471 | skb_tstamp_tx(skb, &shhwtstamps); | |
472 | } | |
9ab5ec59 CM |
473 | } |
474 | ||
475 | /* Use ioctl() to enable or disable TX or RX timestamping. */ | |
6ab96d1e | 476 | static int tile_hwtstamp_set(struct net_device *dev, struct ifreq *rq) |
9ab5ec59 | 477 | { |
9ab5ec59 CM |
478 | struct hwtstamp_config config; |
479 | struct tile_net_priv *priv = netdev_priv(dev); | |
480 | ||
481 | if (copy_from_user(&config, rq->ifr_data, sizeof(config))) | |
482 | return -EFAULT; | |
483 | ||
484 | if (config.flags) /* reserved for future extensions */ | |
485 | return -EINVAL; | |
486 | ||
487 | switch (config.tx_type) { | |
488 | case HWTSTAMP_TX_OFF: | |
489 | case HWTSTAMP_TX_ON: | |
490 | break; | |
491 | default: | |
492 | return -ERANGE; | |
493 | } | |
494 | ||
495 | switch (config.rx_filter) { | |
496 | case HWTSTAMP_FILTER_NONE: | |
497 | break; | |
498 | case HWTSTAMP_FILTER_ALL: | |
499 | case HWTSTAMP_FILTER_SOME: | |
500 | case HWTSTAMP_FILTER_PTP_V1_L4_EVENT: | |
501 | case HWTSTAMP_FILTER_PTP_V1_L4_SYNC: | |
502 | case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ: | |
503 | case HWTSTAMP_FILTER_PTP_V2_L4_EVENT: | |
504 | case HWTSTAMP_FILTER_PTP_V2_L4_SYNC: | |
505 | case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ: | |
506 | case HWTSTAMP_FILTER_PTP_V2_L2_EVENT: | |
507 | case HWTSTAMP_FILTER_PTP_V2_L2_SYNC: | |
508 | case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ: | |
509 | case HWTSTAMP_FILTER_PTP_V2_EVENT: | |
510 | case HWTSTAMP_FILTER_PTP_V2_SYNC: | |
511 | case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ: | |
512 | config.rx_filter = HWTSTAMP_FILTER_ALL; | |
513 | break; | |
514 | default: | |
515 | return -ERANGE; | |
516 | } | |
517 | ||
518 | if (copy_to_user(rq->ifr_data, &config, sizeof(config))) | |
519 | return -EFAULT; | |
520 | ||
521 | priv->stamp_cfg = config; | |
522 | return 0; | |
9ab5ec59 CM |
523 | } |
524 | ||
6ab96d1e BH |
525 | static int tile_hwtstamp_get(struct net_device *dev, struct ifreq *rq) |
526 | { | |
6ab96d1e BH |
527 | struct tile_net_priv *priv = netdev_priv(dev); |
528 | ||
529 | if (copy_to_user(rq->ifr_data, &priv->stamp_cfg, | |
530 | sizeof(priv->stamp_cfg))) | |
531 | return -EFAULT; | |
532 | ||
533 | return 0; | |
6ab96d1e BH |
534 | } |
535 | ||
e3d62d7e CM |
536 | static inline bool filter_packet(struct net_device *dev, void *buf) |
537 | { | |
538 | /* Filter packets received before we're up. */ | |
539 | if (dev == NULL || !(dev->flags & IFF_UP)) | |
540 | return true; | |
541 | ||
542 | /* Filter out packets that aren't for us. */ | |
543 | if (!(dev->flags & IFF_PROMISC) && | |
544 | !is_multicast_ether_addr(buf) && | |
7367d0b5 | 545 | !ether_addr_equal(dev->dev_addr, buf)) |
e3d62d7e CM |
546 | return true; |
547 | ||
548 | return false; | |
549 | } | |
550 | ||
551 | static void tile_net_receive_skb(struct net_device *dev, struct sk_buff *skb, | |
552 | gxio_mpipe_idesc_t *idesc, unsigned long len) | |
553 | { | |
554 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
9ab5ec59 CM |
555 | struct tile_net_priv *priv = netdev_priv(dev); |
556 | int instance = priv->instance; | |
e3d62d7e CM |
557 | |
558 | /* Encode the actual packet length. */ | |
559 | skb_put(skb, len); | |
560 | ||
561 | skb->protocol = eth_type_trans(skb, dev); | |
562 | ||
563 | /* Acknowledge "good" hardware checksums. */ | |
564 | if (idesc->cs && idesc->csum_seed_val == 0xFFFF) | |
565 | skb->ip_summed = CHECKSUM_UNNECESSARY; | |
566 | ||
9ab5ec59 CM |
567 | /* Get RX timestamp from idesc. */ |
568 | tile_rx_timestamp(priv, skb, idesc); | |
569 | ||
f3286a3a | 570 | napi_gro_receive(&info->mpipe[instance].napi, skb); |
e3d62d7e CM |
571 | |
572 | /* Update stats. */ | |
ad018185 CM |
573 | tile_net_stats_add(1, &dev->stats.rx_packets); |
574 | tile_net_stats_add(len, &dev->stats.rx_bytes); | |
e3d62d7e CM |
575 | |
576 | /* Need a new buffer. */ | |
2628e8af | 577 | if (idesc->size == buffer_size_enums[0]) |
f3286a3a | 578 | info->mpipe[instance].num_needed_buffers[0]++; |
2628e8af | 579 | else if (idesc->size == buffer_size_enums[1]) |
f3286a3a | 580 | info->mpipe[instance].num_needed_buffers[1]++; |
e3d62d7e | 581 | else |
f3286a3a | 582 | info->mpipe[instance].num_needed_buffers[2]++; |
e3d62d7e CM |
583 | } |
584 | ||
585 | /* Handle a packet. Return true if "processed", false if "filtered". */ | |
f3286a3a | 586 | static bool tile_net_handle_packet(int instance, gxio_mpipe_idesc_t *idesc) |
e3d62d7e CM |
587 | { |
588 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
f3286a3a CM |
589 | struct mpipe_data *md = &mpipe_data[instance]; |
590 | struct net_device *dev = md->tile_net_devs_for_channel[idesc->channel]; | |
e3d62d7e CM |
591 | uint8_t l2_offset; |
592 | void *va; | |
593 | void *buf; | |
594 | unsigned long len; | |
595 | bool filter; | |
596 | ||
2628e8af CM |
597 | /* Drop packets for which no buffer was available (which can |
598 | * happen under heavy load), or for which the me/tr/ce flags | |
599 | * are set (which can happen for jumbo cut-through packets, | |
600 | * or with a customized classifier). | |
e3d62d7e | 601 | */ |
2628e8af CM |
602 | if (idesc->be || idesc->me || idesc->tr || idesc->ce) { |
603 | if (dev) | |
ad018185 | 604 | tile_net_stats_add(1, &dev->stats.rx_errors); |
2628e8af | 605 | goto drop; |
e3d62d7e CM |
606 | } |
607 | ||
608 | /* Get the "l2_offset", if allowed. */ | |
4aa02644 | 609 | l2_offset = custom_flag ? 0 : gxio_mpipe_idesc_get_l2_offset(idesc); |
e3d62d7e | 610 | |
2628e8af CM |
611 | /* Get the VA (including NET_IP_ALIGN bytes of "headroom"). */ |
612 | va = tile_io_addr_to_va((unsigned long)idesc->va); | |
e3d62d7e CM |
613 | |
614 | /* Get the actual packet start/length. */ | |
615 | buf = va + l2_offset; | |
616 | len = idesc->l2_size - l2_offset; | |
617 | ||
618 | /* Point "va" at the raw buffer. */ | |
619 | va -= NET_IP_ALIGN; | |
620 | ||
621 | filter = filter_packet(dev, buf); | |
622 | if (filter) { | |
2628e8af | 623 | if (dev) |
ad018185 | 624 | tile_net_stats_add(1, &dev->stats.rx_dropped); |
2628e8af | 625 | drop: |
f3286a3a | 626 | gxio_mpipe_iqueue_drop(&info->mpipe[instance].iqueue, idesc); |
e3d62d7e CM |
627 | } else { |
628 | struct sk_buff *skb = mpipe_buf_to_skb(va); | |
629 | ||
630 | /* Skip headroom, and any custom header. */ | |
631 | skb_reserve(skb, NET_IP_ALIGN + l2_offset); | |
632 | ||
633 | tile_net_receive_skb(dev, skb, idesc, len); | |
634 | } | |
635 | ||
f3286a3a | 636 | gxio_mpipe_iqueue_consume(&info->mpipe[instance].iqueue, idesc); |
e3d62d7e CM |
637 | return !filter; |
638 | } | |
639 | ||
640 | /* Handle some packets for the current CPU. | |
641 | * | |
642 | * This function handles up to TILE_NET_BATCH idescs per call. | |
643 | * | |
644 | * ISSUE: Since we do not provide new buffers until this function is | |
645 | * complete, we must initially provide enough buffers for each network | |
646 | * cpu to fill its iqueue and also its batched idescs. | |
647 | * | |
648 | * ISSUE: The "rotting packet" race condition occurs if a packet | |
649 | * arrives after the queue appears to be empty, and before the | |
650 | * hypervisor interrupt is re-enabled. | |
651 | */ | |
652 | static int tile_net_poll(struct napi_struct *napi, int budget) | |
653 | { | |
654 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
655 | unsigned int work = 0; | |
656 | gxio_mpipe_idesc_t *idesc; | |
f3286a3a CM |
657 | int instance, i, n; |
658 | struct mpipe_data *md; | |
659 | struct info_mpipe *info_mpipe = | |
660 | container_of(napi, struct info_mpipe, napi); | |
661 | ||
662 | instance = info_mpipe->instance; | |
663 | while ((n = gxio_mpipe_iqueue_try_peek( | |
664 | &info_mpipe->iqueue, | |
665 | &idesc)) > 0) { | |
e3d62d7e CM |
666 | for (i = 0; i < n; i++) { |
667 | if (i == TILE_NET_BATCH) | |
668 | goto done; | |
f3286a3a CM |
669 | if (tile_net_handle_packet(instance, |
670 | idesc + i)) { | |
e3d62d7e CM |
671 | if (++work >= budget) |
672 | goto done; | |
673 | } | |
674 | } | |
675 | } | |
676 | ||
677 | /* There are no packets left. */ | |
f3286a3a | 678 | napi_complete(&info_mpipe->napi); |
e3d62d7e | 679 | |
f3286a3a | 680 | md = &mpipe_data[instance]; |
e3d62d7e | 681 | /* Re-enable hypervisor interrupts. */ |
f3286a3a CM |
682 | gxio_mpipe_enable_notif_ring_interrupt( |
683 | &md->context, info->mpipe[instance].iqueue.ring); | |
e3d62d7e CM |
684 | |
685 | /* HACK: Avoid the "rotting packet" problem. */ | |
f3286a3a CM |
686 | if (gxio_mpipe_iqueue_try_peek(&info_mpipe->iqueue, &idesc) > 0) |
687 | napi_schedule(&info_mpipe->napi); | |
e3d62d7e CM |
688 | |
689 | /* ISSUE: Handle completions? */ | |
690 | ||
691 | done: | |
692 | tile_net_provide_needed_buffers(); | |
693 | ||
694 | return work; | |
695 | } | |
696 | ||
f3286a3a CM |
697 | /* Handle an ingress interrupt from an instance on the current cpu. */ |
698 | static irqreturn_t tile_net_handle_ingress_irq(int irq, void *id) | |
e3d62d7e CM |
699 | { |
700 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
f3286a3a | 701 | napi_schedule(&info->mpipe[(uint64_t)id].napi); |
e3d62d7e CM |
702 | return IRQ_HANDLED; |
703 | } | |
704 | ||
705 | /* Free some completions. This must be called with interrupts blocked. */ | |
706 | static int tile_net_free_comps(gxio_mpipe_equeue_t *equeue, | |
707 | struct tile_net_comps *comps, | |
708 | int limit, bool force_update) | |
709 | { | |
710 | int n = 0; | |
711 | while (comps->comp_last < comps->comp_next) { | |
712 | unsigned int cid = comps->comp_last % TILE_NET_MAX_COMPS; | |
713 | struct tile_net_comp *comp = &comps->comp_queue[cid]; | |
714 | if (!gxio_mpipe_equeue_is_complete(equeue, comp->when, | |
715 | force_update || n == 0)) | |
716 | break; | |
717 | dev_kfree_skb_irq(comp->skb); | |
718 | comps->comp_last++; | |
719 | if (++n == limit) | |
720 | break; | |
721 | } | |
722 | return n; | |
723 | } | |
724 | ||
725 | /* Add a completion. This must be called with interrupts blocked. | |
726 | * tile_net_equeue_try_reserve() will have ensured a free completion entry. | |
727 | */ | |
728 | static void add_comp(gxio_mpipe_equeue_t *equeue, | |
729 | struct tile_net_comps *comps, | |
730 | uint64_t when, struct sk_buff *skb) | |
731 | { | |
732 | int cid = comps->comp_next % TILE_NET_MAX_COMPS; | |
733 | comps->comp_queue[cid].when = when; | |
734 | comps->comp_queue[cid].skb = skb; | |
735 | comps->comp_next++; | |
736 | } | |
737 | ||
9b4c341b CM |
738 | static void tile_net_schedule_tx_wake_timer(struct net_device *dev, |
739 | int tx_queue_idx) | |
e3d62d7e | 740 | { |
9b4c341b | 741 | struct tile_net_info *info = &per_cpu(per_cpu_info, tx_queue_idx); |
e3d62d7e | 742 | struct tile_net_priv *priv = netdev_priv(dev); |
f3286a3a CM |
743 | int instance = priv->instance; |
744 | struct tile_net_tx_wake *tx_wake = | |
745 | &info->mpipe[instance].tx_wake[priv->echannel]; | |
e3d62d7e | 746 | |
9b4c341b | 747 | hrtimer_start(&tx_wake->timer, |
e3d62d7e CM |
748 | ktime_set(0, TX_TIMER_DELAY_USEC * 1000UL), |
749 | HRTIMER_MODE_REL_PINNED); | |
750 | } | |
751 | ||
752 | static enum hrtimer_restart tile_net_handle_tx_wake_timer(struct hrtimer *t) | |
753 | { | |
754 | struct tile_net_tx_wake *tx_wake = | |
755 | container_of(t, struct tile_net_tx_wake, timer); | |
9b4c341b | 756 | netif_wake_subqueue(tx_wake->dev, tx_wake->tx_queue_idx); |
e3d62d7e CM |
757 | return HRTIMER_NORESTART; |
758 | } | |
759 | ||
760 | /* Make sure the egress timer is scheduled. */ | |
761 | static void tile_net_schedule_egress_timer(void) | |
762 | { | |
763 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
764 | ||
765 | if (!info->egress_timer_scheduled) { | |
766 | hrtimer_start(&info->egress_timer, | |
767 | ktime_set(0, EGRESS_TIMER_DELAY_USEC * 1000UL), | |
768 | HRTIMER_MODE_REL_PINNED); | |
769 | info->egress_timer_scheduled = true; | |
770 | } | |
771 | } | |
772 | ||
773 | /* The "function" for "info->egress_timer". | |
774 | * | |
775 | * This timer will reschedule itself as long as there are any pending | |
776 | * completions expected for this tile. | |
777 | */ | |
778 | static enum hrtimer_restart tile_net_handle_egress_timer(struct hrtimer *t) | |
779 | { | |
780 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
781 | unsigned long irqflags; | |
782 | bool pending = false; | |
f3286a3a | 783 | int i, instance; |
e3d62d7e CM |
784 | |
785 | local_irq_save(irqflags); | |
786 | ||
787 | /* The timer is no longer scheduled. */ | |
788 | info->egress_timer_scheduled = false; | |
789 | ||
790 | /* Free all possible comps for this tile. */ | |
f3286a3a CM |
791 | for (instance = 0; instance < NR_MPIPE_MAX && |
792 | info->mpipe[instance].has_iqueue; instance++) { | |
793 | for (i = 0; i < TILE_NET_CHANNELS; i++) { | |
794 | struct tile_net_egress *egress = | |
795 | &mpipe_data[instance].egress_for_echannel[i]; | |
796 | struct tile_net_comps *comps = | |
797 | info->mpipe[instance].comps_for_echannel[i]; | |
798 | if (!egress || comps->comp_last >= comps->comp_next) | |
799 | continue; | |
800 | tile_net_free_comps(egress->equeue, comps, -1, true); | |
801 | pending = pending || | |
802 | (comps->comp_last < comps->comp_next); | |
803 | } | |
e3d62d7e CM |
804 | } |
805 | ||
806 | /* Reschedule timer if needed. */ | |
807 | if (pending) | |
808 | tile_net_schedule_egress_timer(); | |
809 | ||
810 | local_irq_restore(irqflags); | |
811 | ||
812 | return HRTIMER_NORESTART; | |
813 | } | |
814 | ||
9ab5ec59 CM |
815 | /* PTP clock operations. */ |
816 | ||
817 | static int ptp_mpipe_adjfreq(struct ptp_clock_info *ptp, s32 ppb) | |
818 | { | |
819 | int ret = 0; | |
820 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
821 | mutex_lock(&md->ptp_lock); | |
822 | if (gxio_mpipe_adjust_timestamp_freq(&md->context, ppb)) | |
823 | ret = -EINVAL; | |
824 | mutex_unlock(&md->ptp_lock); | |
825 | return ret; | |
826 | } | |
827 | ||
828 | static int ptp_mpipe_adjtime(struct ptp_clock_info *ptp, s64 delta) | |
829 | { | |
830 | int ret = 0; | |
831 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
832 | mutex_lock(&md->ptp_lock); | |
833 | if (gxio_mpipe_adjust_timestamp(&md->context, delta)) | |
834 | ret = -EBUSY; | |
835 | mutex_unlock(&md->ptp_lock); | |
836 | return ret; | |
837 | } | |
838 | ||
839 | static int ptp_mpipe_gettime(struct ptp_clock_info *ptp, struct timespec *ts) | |
840 | { | |
841 | int ret = 0; | |
842 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
843 | mutex_lock(&md->ptp_lock); | |
844 | if (gxio_mpipe_get_timestamp(&md->context, ts)) | |
845 | ret = -EBUSY; | |
846 | mutex_unlock(&md->ptp_lock); | |
847 | return ret; | |
848 | } | |
849 | ||
850 | static int ptp_mpipe_settime(struct ptp_clock_info *ptp, | |
851 | const struct timespec *ts) | |
852 | { | |
853 | int ret = 0; | |
854 | struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps); | |
855 | mutex_lock(&md->ptp_lock); | |
856 | if (gxio_mpipe_set_timestamp(&md->context, ts)) | |
857 | ret = -EBUSY; | |
858 | mutex_unlock(&md->ptp_lock); | |
859 | return ret; | |
860 | } | |
861 | ||
862 | static int ptp_mpipe_enable(struct ptp_clock_info *ptp, | |
863 | struct ptp_clock_request *request, int on) | |
864 | { | |
865 | return -EOPNOTSUPP; | |
866 | } | |
867 | ||
868 | static struct ptp_clock_info ptp_mpipe_caps = { | |
869 | .owner = THIS_MODULE, | |
870 | .name = "mPIPE clock", | |
871 | .max_adj = 999999999, | |
872 | .n_ext_ts = 0, | |
873 | .pps = 0, | |
874 | .adjfreq = ptp_mpipe_adjfreq, | |
875 | .adjtime = ptp_mpipe_adjtime, | |
876 | .gettime = ptp_mpipe_gettime, | |
877 | .settime = ptp_mpipe_settime, | |
878 | .enable = ptp_mpipe_enable, | |
879 | }; | |
880 | ||
9ab5ec59 CM |
881 | /* Sync mPIPE's timestamp up with Linux system time and register PTP clock. */ |
882 | static void register_ptp_clock(struct net_device *dev, struct mpipe_data *md) | |
883 | { | |
9ab5ec59 CM |
884 | struct timespec ts; |
885 | ||
886 | getnstimeofday(&ts); | |
887 | gxio_mpipe_set_timestamp(&md->context, &ts); | |
888 | ||
889 | mutex_init(&md->ptp_lock); | |
890 | md->caps = ptp_mpipe_caps; | |
891 | md->ptp_clock = ptp_clock_register(&md->caps, NULL); | |
892 | if (IS_ERR(md->ptp_clock)) | |
893 | netdev_err(dev, "ptp_clock_register failed %ld\n", | |
894 | PTR_ERR(md->ptp_clock)); | |
9ab5ec59 CM |
895 | } |
896 | ||
897 | /* Initialize PTP fields in a new device. */ | |
898 | static void init_ptp_dev(struct tile_net_priv *priv) | |
899 | { | |
9ab5ec59 CM |
900 | priv->stamp_cfg.rx_filter = HWTSTAMP_FILTER_NONE; |
901 | priv->stamp_cfg.tx_type = HWTSTAMP_TX_OFF; | |
9ab5ec59 CM |
902 | } |
903 | ||
f3286a3a CM |
904 | /* Helper functions for "tile_net_update()". */ |
905 | static void enable_ingress_irq(void *irq) | |
e3d62d7e | 906 | { |
f3286a3a CM |
907 | enable_percpu_irq((long)irq, 0); |
908 | } | |
909 | ||
910 | static void disable_ingress_irq(void *irq) | |
911 | { | |
912 | disable_percpu_irq((long)irq); | |
e3d62d7e CM |
913 | } |
914 | ||
915 | /* Helper function for tile_net_open() and tile_net_stop(). | |
916 | * Always called under tile_net_devs_for_channel_mutex. | |
917 | */ | |
918 | static int tile_net_update(struct net_device *dev) | |
919 | { | |
920 | static gxio_mpipe_rules_t rules; /* too big to fit on the stack */ | |
921 | bool saw_channel = false; | |
f3286a3a CM |
922 | int instance = mpipe_instance(dev); |
923 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
924 | int channel; |
925 | int rc; | |
926 | int cpu; | |
927 | ||
f3286a3a CM |
928 | saw_channel = false; |
929 | gxio_mpipe_rules_init(&rules, &md->context); | |
e3d62d7e CM |
930 | |
931 | for (channel = 0; channel < TILE_NET_CHANNELS; channel++) { | |
f3286a3a | 932 | if (md->tile_net_devs_for_channel[channel] == NULL) |
e3d62d7e CM |
933 | continue; |
934 | if (!saw_channel) { | |
935 | saw_channel = true; | |
f3286a3a CM |
936 | gxio_mpipe_rules_begin(&rules, md->first_bucket, |
937 | md->num_buckets, NULL); | |
e3d62d7e CM |
938 | gxio_mpipe_rules_set_headroom(&rules, NET_IP_ALIGN); |
939 | } | |
940 | gxio_mpipe_rules_add_channel(&rules, channel); | |
941 | } | |
942 | ||
943 | /* NOTE: This can fail if there is no classifier. | |
944 | * ISSUE: Can anything else cause it to fail? | |
945 | */ | |
946 | rc = gxio_mpipe_rules_commit(&rules); | |
947 | if (rc != 0) { | |
f3286a3a CM |
948 | netdev_warn(dev, "gxio_mpipe_rules_commit: mpipe[%d] %d\n", |
949 | instance, rc); | |
e3d62d7e CM |
950 | return -EIO; |
951 | } | |
952 | ||
5e7a54a2 CM |
953 | /* Update all cpus, sequentially (to protect "netif_napi_add()"). |
954 | * We use on_each_cpu to handle the IPI mask or unmask. | |
955 | */ | |
956 | if (!saw_channel) | |
f3286a3a CM |
957 | on_each_cpu(disable_ingress_irq, |
958 | (void *)(long)(md->ingress_irq), 1); | |
5e7a54a2 CM |
959 | for_each_online_cpu(cpu) { |
960 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
f3286a3a CM |
961 | |
962 | if (!info->mpipe[instance].has_iqueue) | |
5e7a54a2 CM |
963 | continue; |
964 | if (saw_channel) { | |
f3286a3a CM |
965 | if (!info->mpipe[instance].napi_added) { |
966 | netif_napi_add(dev, &info->mpipe[instance].napi, | |
5e7a54a2 | 967 | tile_net_poll, TILE_NET_WEIGHT); |
f3286a3a | 968 | info->mpipe[instance].napi_added = true; |
5e7a54a2 | 969 | } |
f3286a3a CM |
970 | if (!info->mpipe[instance].napi_enabled) { |
971 | napi_enable(&info->mpipe[instance].napi); | |
972 | info->mpipe[instance].napi_enabled = true; | |
5e7a54a2 CM |
973 | } |
974 | } else { | |
f3286a3a CM |
975 | if (info->mpipe[instance].napi_enabled) { |
976 | napi_disable(&info->mpipe[instance].napi); | |
977 | info->mpipe[instance].napi_enabled = false; | |
5e7a54a2 CM |
978 | } |
979 | /* FIXME: Drain the iqueue. */ | |
980 | } | |
981 | } | |
982 | if (saw_channel) | |
f3286a3a CM |
983 | on_each_cpu(enable_ingress_irq, |
984 | (void *)(long)(md->ingress_irq), 1); | |
e3d62d7e CM |
985 | |
986 | /* HACK: Allow packets to flow in the simulator. */ | |
987 | if (saw_channel) | |
f3286a3a | 988 | sim_enable_mpipe_links(instance, -1); |
e3d62d7e CM |
989 | |
990 | return 0; | |
991 | } | |
992 | ||
2628e8af CM |
993 | /* Initialize a buffer stack. */ |
994 | static int create_buffer_stack(struct net_device *dev, | |
995 | int kind, size_t num_buffers) | |
e3d62d7e CM |
996 | { |
997 | pte_t hash_pte = pte_set_home((pte_t) { 0 }, PAGE_HOME_HASH); | |
f3286a3a CM |
998 | int instance = mpipe_instance(dev); |
999 | struct mpipe_data *md = &mpipe_data[instance]; | |
2628e8af | 1000 | size_t needed = gxio_mpipe_calc_buffer_stack_bytes(num_buffers); |
f3286a3a | 1001 | int stack_idx = md->first_buffer_stack + kind; |
2628e8af CM |
1002 | void *va; |
1003 | int i, rc; | |
e3d62d7e | 1004 | |
2628e8af CM |
1005 | /* Round up to 64KB and then use alloc_pages() so we get the |
1006 | * required 64KB alignment. | |
e3d62d7e | 1007 | */ |
f3286a3a CM |
1008 | md->buffer_stack_bytes[kind] = |
1009 | ALIGN(needed, 64 * 1024); | |
e3d62d7e | 1010 | |
f3286a3a | 1011 | va = alloc_pages_exact(md->buffer_stack_bytes[kind], GFP_KERNEL); |
2628e8af | 1012 | if (va == NULL) { |
e3d62d7e | 1013 | netdev_err(dev, |
2628e8af | 1014 | "Could not alloc %zd bytes for buffer stack %d\n", |
f3286a3a | 1015 | md->buffer_stack_bytes[kind], kind); |
e3d62d7e CM |
1016 | return -ENOMEM; |
1017 | } | |
2628e8af CM |
1018 | |
1019 | /* Initialize the buffer stack. */ | |
f3286a3a CM |
1020 | rc = gxio_mpipe_init_buffer_stack(&md->context, stack_idx, |
1021 | buffer_size_enums[kind], va, | |
1022 | md->buffer_stack_bytes[kind], 0); | |
e3d62d7e | 1023 | if (rc != 0) { |
f3286a3a CM |
1024 | netdev_err(dev, "gxio_mpipe_init_buffer_stack: mpipe[%d] %d\n", |
1025 | instance, rc); | |
1026 | free_pages_exact(va, md->buffer_stack_bytes[kind]); | |
e3d62d7e CM |
1027 | return rc; |
1028 | } | |
2628e8af | 1029 | |
f3286a3a | 1030 | md->buffer_stack_vas[kind] = va; |
2628e8af | 1031 | |
f3286a3a | 1032 | rc = gxio_mpipe_register_client_memory(&md->context, stack_idx, |
e3d62d7e CM |
1033 | hash_pte, 0); |
1034 | if (rc != 0) { | |
f3286a3a CM |
1035 | netdev_err(dev, |
1036 | "gxio_mpipe_register_client_memory: mpipe[%d] %d\n", | |
1037 | instance, rc); | |
e3d62d7e CM |
1038 | return rc; |
1039 | } | |
1040 | ||
2628e8af CM |
1041 | /* Provide initial buffers. */ |
1042 | for (i = 0; i < num_buffers; i++) { | |
f3286a3a | 1043 | if (!tile_net_provide_buffer(instance, kind)) { |
2628e8af CM |
1044 | netdev_err(dev, "Cannot allocate initial sk_bufs!\n"); |
1045 | return -ENOMEM; | |
1046 | } | |
e3d62d7e | 1047 | } |
2628e8af CM |
1048 | |
1049 | return 0; | |
1050 | } | |
1051 | ||
1052 | /* Allocate and initialize mpipe buffer stacks, and register them in | |
1053 | * the mPIPE TLBs, for small, large, and (possibly) jumbo packet sizes. | |
1054 | * This routine supports tile_net_init_mpipe(), below. | |
1055 | */ | |
1056 | static int init_buffer_stacks(struct net_device *dev, | |
1057 | int network_cpus_count) | |
1058 | { | |
1059 | int num_kinds = MAX_KINDS - (jumbo_num == 0); | |
1060 | size_t num_buffers; | |
1061 | int rc; | |
f3286a3a CM |
1062 | int instance = mpipe_instance(dev); |
1063 | struct mpipe_data *md = &mpipe_data[instance]; | |
2628e8af CM |
1064 | |
1065 | /* Allocate the buffer stacks. */ | |
f3286a3a | 1066 | rc = gxio_mpipe_alloc_buffer_stacks(&md->context, num_kinds, 0, 0); |
2628e8af | 1067 | if (rc < 0) { |
f3286a3a CM |
1068 | netdev_err(dev, |
1069 | "gxio_mpipe_alloc_buffer_stacks: mpipe[%d] %d\n", | |
1070 | instance, rc); | |
e3d62d7e CM |
1071 | return rc; |
1072 | } | |
f3286a3a | 1073 | md->first_buffer_stack = rc; |
e3d62d7e | 1074 | |
2628e8af CM |
1075 | /* Enough small/large buffers to (normally) avoid buffer errors. */ |
1076 | num_buffers = | |
1077 | network_cpus_count * (IQUEUE_ENTRIES + TILE_NET_BATCH); | |
1078 | ||
1079 | /* Allocate the small memory stack. */ | |
1080 | if (rc >= 0) | |
1081 | rc = create_buffer_stack(dev, 0, num_buffers); | |
1082 | ||
1083 | /* Allocate the large buffer stack. */ | |
1084 | if (rc >= 0) | |
1085 | rc = create_buffer_stack(dev, 1, num_buffers); | |
1086 | ||
1087 | /* Allocate the jumbo buffer stack if needed. */ | |
1088 | if (rc >= 0 && jumbo_num != 0) | |
1089 | rc = create_buffer_stack(dev, 2, jumbo_num); | |
1090 | ||
1091 | return rc; | |
e3d62d7e CM |
1092 | } |
1093 | ||
1094 | /* Allocate per-cpu resources (memory for completions and idescs). | |
1095 | * This routine supports tile_net_init_mpipe(), below. | |
1096 | */ | |
1097 | static int alloc_percpu_mpipe_resources(struct net_device *dev, | |
1098 | int cpu, int ring) | |
1099 | { | |
1100 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
1101 | int order, i, rc; | |
f3286a3a CM |
1102 | int instance = mpipe_instance(dev); |
1103 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1104 | struct page *page; |
1105 | void *addr; | |
1106 | ||
1107 | /* Allocate the "comps". */ | |
1108 | order = get_order(COMPS_SIZE); | |
1109 | page = homecache_alloc_pages(GFP_KERNEL, order, cpu); | |
1110 | if (page == NULL) { | |
1111 | netdev_err(dev, "Failed to alloc %zd bytes comps memory\n", | |
1112 | COMPS_SIZE); | |
1113 | return -ENOMEM; | |
1114 | } | |
1115 | addr = pfn_to_kaddr(page_to_pfn(page)); | |
1116 | memset(addr, 0, COMPS_SIZE); | |
1117 | for (i = 0; i < TILE_NET_CHANNELS; i++) | |
f3286a3a | 1118 | info->mpipe[instance].comps_for_echannel[i] = |
e3d62d7e CM |
1119 | addr + i * sizeof(struct tile_net_comps); |
1120 | ||
1121 | /* If this is a network cpu, create an iqueue. */ | |
1122 | if (cpu_isset(cpu, network_cpus_map)) { | |
1123 | order = get_order(NOTIF_RING_SIZE); | |
1124 | page = homecache_alloc_pages(GFP_KERNEL, order, cpu); | |
1125 | if (page == NULL) { | |
1126 | netdev_err(dev, | |
1127 | "Failed to alloc %zd bytes iqueue memory\n", | |
1128 | NOTIF_RING_SIZE); | |
1129 | return -ENOMEM; | |
1130 | } | |
1131 | addr = pfn_to_kaddr(page_to_pfn(page)); | |
f3286a3a CM |
1132 | rc = gxio_mpipe_iqueue_init(&info->mpipe[instance].iqueue, |
1133 | &md->context, ring++, addr, | |
1134 | NOTIF_RING_SIZE, 0); | |
e3d62d7e CM |
1135 | if (rc < 0) { |
1136 | netdev_err(dev, | |
1137 | "gxio_mpipe_iqueue_init failed: %d\n", rc); | |
1138 | return rc; | |
1139 | } | |
f3286a3a | 1140 | info->mpipe[instance].has_iqueue = true; |
e3d62d7e CM |
1141 | } |
1142 | ||
1143 | return ring; | |
1144 | } | |
1145 | ||
1146 | /* Initialize NotifGroup and buckets. | |
1147 | * This routine supports tile_net_init_mpipe(), below. | |
1148 | */ | |
1149 | static int init_notif_group_and_buckets(struct net_device *dev, | |
1150 | int ring, int network_cpus_count) | |
1151 | { | |
1152 | int group, rc; | |
f3286a3a CM |
1153 | int instance = mpipe_instance(dev); |
1154 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1155 | |
1156 | /* Allocate one NotifGroup. */ | |
f3286a3a | 1157 | rc = gxio_mpipe_alloc_notif_groups(&md->context, 1, 0, 0); |
e3d62d7e | 1158 | if (rc < 0) { |
f3286a3a CM |
1159 | netdev_err(dev, "gxio_mpipe_alloc_notif_groups: mpipe[%d] %d\n", |
1160 | instance, rc); | |
e3d62d7e CM |
1161 | return rc; |
1162 | } | |
1163 | group = rc; | |
1164 | ||
1165 | /* Initialize global num_buckets value. */ | |
1166 | if (network_cpus_count > 4) | |
f3286a3a | 1167 | md->num_buckets = 256; |
e3d62d7e | 1168 | else if (network_cpus_count > 1) |
f3286a3a | 1169 | md->num_buckets = 16; |
e3d62d7e CM |
1170 | |
1171 | /* Allocate some buckets, and set global first_bucket value. */ | |
f3286a3a | 1172 | rc = gxio_mpipe_alloc_buckets(&md->context, md->num_buckets, 0, 0); |
e3d62d7e | 1173 | if (rc < 0) { |
f3286a3a CM |
1174 | netdev_err(dev, "gxio_mpipe_alloc_buckets: mpipe[%d] %d\n", |
1175 | instance, rc); | |
e3d62d7e CM |
1176 | return rc; |
1177 | } | |
f3286a3a | 1178 | md->first_bucket = rc; |
e3d62d7e CM |
1179 | |
1180 | /* Init group and buckets. */ | |
1181 | rc = gxio_mpipe_init_notif_group_and_buckets( | |
f3286a3a CM |
1182 | &md->context, group, ring, network_cpus_count, |
1183 | md->first_bucket, md->num_buckets, | |
e3d62d7e CM |
1184 | GXIO_MPIPE_BUCKET_STICKY_FLOW_LOCALITY); |
1185 | if (rc != 0) { | |
f3286a3a CM |
1186 | netdev_err(dev, "gxio_mpipe_init_notif_group_and_buckets: " |
1187 | "mpipe[%d] %d\n", instance, rc); | |
e3d62d7e CM |
1188 | return rc; |
1189 | } | |
1190 | ||
1191 | return 0; | |
1192 | } | |
1193 | ||
1194 | /* Create an irq and register it, then activate the irq and request | |
1195 | * interrupts on all cores. Note that "ingress_irq" being initialized | |
1196 | * is how we know not to call tile_net_init_mpipe() again. | |
1197 | * This routine supports tile_net_init_mpipe(), below. | |
1198 | */ | |
1199 | static int tile_net_setup_interrupts(struct net_device *dev) | |
1200 | { | |
f3286a3a CM |
1201 | int cpu, rc, irq; |
1202 | int instance = mpipe_instance(dev); | |
1203 | struct mpipe_data *md = &mpipe_data[instance]; | |
1204 | ||
1205 | irq = md->ingress_irq; | |
1206 | if (irq < 0) { | |
1207 | irq = create_irq(); | |
1208 | if (irq < 0) { | |
1209 | netdev_err(dev, | |
1210 | "create_irq failed: mpipe[%d] %d\n", | |
1211 | instance, irq); | |
1212 | return irq; | |
1213 | } | |
1214 | tile_irq_activate(irq, TILE_IRQ_PERCPU); | |
e3d62d7e | 1215 | |
f3286a3a CM |
1216 | rc = request_irq(irq, tile_net_handle_ingress_irq, |
1217 | 0, "tile_net", (void *)((uint64_t)instance)); | |
1218 | ||
1219 | if (rc != 0) { | |
1220 | netdev_err(dev, "request_irq failed: mpipe[%d] %d\n", | |
1221 | instance, rc); | |
1222 | destroy_irq(irq); | |
1223 | return rc; | |
1224 | } | |
1225 | md->ingress_irq = irq; | |
e3d62d7e CM |
1226 | } |
1227 | ||
1228 | for_each_online_cpu(cpu) { | |
1229 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
f3286a3a CM |
1230 | if (info->mpipe[instance].has_iqueue) { |
1231 | gxio_mpipe_request_notif_ring_interrupt(&md->context, | |
1232 | cpu_x(cpu), cpu_y(cpu), KERNEL_PL, irq, | |
1233 | info->mpipe[instance].iqueue.ring); | |
e3d62d7e CM |
1234 | } |
1235 | } | |
1236 | ||
1237 | return 0; | |
1238 | } | |
1239 | ||
1240 | /* Undo any state set up partially by a failed call to tile_net_init_mpipe. */ | |
f3286a3a | 1241 | static void tile_net_init_mpipe_fail(int instance) |
e3d62d7e | 1242 | { |
2628e8af | 1243 | int kind, cpu; |
f3286a3a | 1244 | struct mpipe_data *md = &mpipe_data[instance]; |
e3d62d7e CM |
1245 | |
1246 | /* Do cleanups that require the mpipe context first. */ | |
2628e8af | 1247 | for (kind = 0; kind < MAX_KINDS; kind++) { |
f3286a3a CM |
1248 | if (md->buffer_stack_vas[kind] != NULL) { |
1249 | tile_net_pop_all_buffers(instance, | |
1250 | md->first_buffer_stack + | |
1251 | kind); | |
2628e8af CM |
1252 | } |
1253 | } | |
e3d62d7e CM |
1254 | |
1255 | /* Destroy mpipe context so the hardware no longer owns any memory. */ | |
f3286a3a | 1256 | gxio_mpipe_destroy(&md->context); |
e3d62d7e CM |
1257 | |
1258 | for_each_online_cpu(cpu) { | |
1259 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
f3286a3a CM |
1260 | free_pages( |
1261 | (unsigned long)( | |
1262 | info->mpipe[instance].comps_for_echannel[0]), | |
1263 | get_order(COMPS_SIZE)); | |
1264 | info->mpipe[instance].comps_for_echannel[0] = NULL; | |
1265 | free_pages((unsigned long)(info->mpipe[instance].iqueue.idescs), | |
e3d62d7e | 1266 | get_order(NOTIF_RING_SIZE)); |
f3286a3a | 1267 | info->mpipe[instance].iqueue.idescs = NULL; |
e3d62d7e CM |
1268 | } |
1269 | ||
2628e8af | 1270 | for (kind = 0; kind < MAX_KINDS; kind++) { |
f3286a3a CM |
1271 | if (md->buffer_stack_vas[kind] != NULL) { |
1272 | free_pages_exact(md->buffer_stack_vas[kind], | |
1273 | md->buffer_stack_bytes[kind]); | |
1274 | md->buffer_stack_vas[kind] = NULL; | |
2628e8af CM |
1275 | } |
1276 | } | |
e3d62d7e | 1277 | |
f3286a3a CM |
1278 | md->first_buffer_stack = -1; |
1279 | md->first_bucket = -1; | |
e3d62d7e CM |
1280 | } |
1281 | ||
1282 | /* The first time any tilegx network device is opened, we initialize | |
1283 | * the global mpipe state. If this step fails, we fail to open the | |
1284 | * device, but if it succeeds, we never need to do it again, and since | |
1285 | * tile_net can't be unloaded, we never undo it. | |
1286 | * | |
1287 | * Note that some resources in this path (buffer stack indices, | |
1288 | * bindings from init_buffer_stack, etc.) are hypervisor resources | |
1289 | * that are freed implicitly by gxio_mpipe_destroy(). | |
1290 | */ | |
1291 | static int tile_net_init_mpipe(struct net_device *dev) | |
1292 | { | |
2628e8af | 1293 | int rc; |
e3d62d7e CM |
1294 | int cpu; |
1295 | int first_ring, ring; | |
f3286a3a CM |
1296 | int instance = mpipe_instance(dev); |
1297 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1298 | int network_cpus_count = cpus_weight(network_cpus_map); |
1299 | ||
1300 | if (!hash_default) { | |
1301 | netdev_err(dev, "Networking requires hash_default!\n"); | |
1302 | return -EIO; | |
1303 | } | |
1304 | ||
f3286a3a | 1305 | rc = gxio_mpipe_init(&md->context, instance); |
e3d62d7e | 1306 | if (rc != 0) { |
f3286a3a CM |
1307 | netdev_err(dev, "gxio_mpipe_init: mpipe[%d] %d\n", |
1308 | instance, rc); | |
e3d62d7e CM |
1309 | return -EIO; |
1310 | } | |
1311 | ||
1312 | /* Set up the buffer stacks. */ | |
2628e8af | 1313 | rc = init_buffer_stacks(dev, network_cpus_count); |
e3d62d7e CM |
1314 | if (rc != 0) |
1315 | goto fail; | |
1316 | ||
e3d62d7e | 1317 | /* Allocate one NotifRing for each network cpu. */ |
f3286a3a CM |
1318 | rc = gxio_mpipe_alloc_notif_rings(&md->context, |
1319 | network_cpus_count, 0, 0); | |
e3d62d7e CM |
1320 | if (rc < 0) { |
1321 | netdev_err(dev, "gxio_mpipe_alloc_notif_rings failed %d\n", | |
1322 | rc); | |
1323 | goto fail; | |
1324 | } | |
1325 | ||
1326 | /* Init NotifRings per-cpu. */ | |
1327 | first_ring = rc; | |
1328 | ring = first_ring; | |
1329 | for_each_online_cpu(cpu) { | |
1330 | rc = alloc_percpu_mpipe_resources(dev, cpu, ring); | |
1331 | if (rc < 0) | |
1332 | goto fail; | |
1333 | ring = rc; | |
1334 | } | |
1335 | ||
1336 | /* Initialize NotifGroup and buckets. */ | |
1337 | rc = init_notif_group_and_buckets(dev, first_ring, network_cpus_count); | |
1338 | if (rc != 0) | |
1339 | goto fail; | |
1340 | ||
1341 | /* Create and enable interrupts. */ | |
1342 | rc = tile_net_setup_interrupts(dev); | |
1343 | if (rc != 0) | |
1344 | goto fail; | |
1345 | ||
9ab5ec59 CM |
1346 | /* Register PTP clock and set mPIPE timestamp, if configured. */ |
1347 | register_ptp_clock(dev, md); | |
1348 | ||
e3d62d7e CM |
1349 | return 0; |
1350 | ||
1351 | fail: | |
f3286a3a | 1352 | tile_net_init_mpipe_fail(instance); |
e3d62d7e CM |
1353 | return rc; |
1354 | } | |
1355 | ||
1356 | /* Create persistent egress info for a given egress channel. | |
1357 | * Note that this may be shared between, say, "gbe0" and "xgbe0". | |
1358 | * ISSUE: Defer header allocation until TSO is actually needed? | |
1359 | */ | |
1360 | static int tile_net_init_egress(struct net_device *dev, int echannel) | |
1361 | { | |
2628e8af | 1362 | static int ering = -1; |
e3d62d7e CM |
1363 | struct page *headers_page, *edescs_page, *equeue_page; |
1364 | gxio_mpipe_edesc_t *edescs; | |
1365 | gxio_mpipe_equeue_t *equeue; | |
1366 | unsigned char *headers; | |
1367 | int headers_order, edescs_order, equeue_order; | |
1368 | size_t edescs_size; | |
e3d62d7e | 1369 | int rc = -ENOMEM; |
f3286a3a CM |
1370 | int instance = mpipe_instance(dev); |
1371 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1372 | |
1373 | /* Only initialize once. */ | |
f3286a3a | 1374 | if (md->egress_for_echannel[echannel].equeue != NULL) |
e3d62d7e CM |
1375 | return 0; |
1376 | ||
1377 | /* Allocate memory for the "headers". */ | |
1378 | headers_order = get_order(EQUEUE_ENTRIES * HEADER_BYTES); | |
1379 | headers_page = alloc_pages(GFP_KERNEL, headers_order); | |
1380 | if (headers_page == NULL) { | |
1381 | netdev_warn(dev, | |
1382 | "Could not alloc %zd bytes for TSO headers.\n", | |
1383 | PAGE_SIZE << headers_order); | |
1384 | goto fail; | |
1385 | } | |
1386 | headers = pfn_to_kaddr(page_to_pfn(headers_page)); | |
1387 | ||
1388 | /* Allocate memory for the "edescs". */ | |
1389 | edescs_size = EQUEUE_ENTRIES * sizeof(*edescs); | |
1390 | edescs_order = get_order(edescs_size); | |
1391 | edescs_page = alloc_pages(GFP_KERNEL, edescs_order); | |
1392 | if (edescs_page == NULL) { | |
1393 | netdev_warn(dev, | |
1394 | "Could not alloc %zd bytes for eDMA ring.\n", | |
1395 | edescs_size); | |
1396 | goto fail_headers; | |
1397 | } | |
1398 | edescs = pfn_to_kaddr(page_to_pfn(edescs_page)); | |
1399 | ||
1400 | /* Allocate memory for the "equeue". */ | |
1401 | equeue_order = get_order(sizeof(*equeue)); | |
1402 | equeue_page = alloc_pages(GFP_KERNEL, equeue_order); | |
1403 | if (equeue_page == NULL) { | |
1404 | netdev_warn(dev, | |
1405 | "Could not alloc %zd bytes for equeue info.\n", | |
1406 | PAGE_SIZE << equeue_order); | |
1407 | goto fail_edescs; | |
1408 | } | |
1409 | equeue = pfn_to_kaddr(page_to_pfn(equeue_page)); | |
1410 | ||
2628e8af CM |
1411 | /* Allocate an edma ring (using a one entry "free list"). */ |
1412 | if (ering < 0) { | |
f3286a3a | 1413 | rc = gxio_mpipe_alloc_edma_rings(&md->context, 1, 0, 0); |
2628e8af | 1414 | if (rc < 0) { |
f3286a3a CM |
1415 | netdev_warn(dev, "gxio_mpipe_alloc_edma_rings: " |
1416 | "mpipe[%d] %d\n", instance, rc); | |
2628e8af CM |
1417 | goto fail_equeue; |
1418 | } | |
1419 | ering = rc; | |
e3d62d7e | 1420 | } |
e3d62d7e CM |
1421 | |
1422 | /* Initialize the equeue. */ | |
f3286a3a | 1423 | rc = gxio_mpipe_equeue_init(equeue, &md->context, ering, echannel, |
e3d62d7e CM |
1424 | edescs, edescs_size, 0); |
1425 | if (rc != 0) { | |
f3286a3a CM |
1426 | netdev_err(dev, "gxio_mpipe_equeue_init: mpipe[%d] %d\n", |
1427 | instance, rc); | |
e3d62d7e CM |
1428 | goto fail_equeue; |
1429 | } | |
1430 | ||
2628e8af CM |
1431 | /* Don't reuse the ering later. */ |
1432 | ering = -1; | |
1433 | ||
1434 | if (jumbo_num != 0) { | |
1435 | /* Make sure "jumbo" packets can be egressed safely. */ | |
1436 | if (gxio_mpipe_equeue_set_snf_size(equeue, 10368) < 0) { | |
1437 | /* ISSUE: There is no "gxio_mpipe_equeue_destroy()". */ | |
1438 | netdev_warn(dev, "Jumbo packets may not be egressed" | |
1439 | " properly on channel %d\n", echannel); | |
1440 | } | |
1441 | } | |
1442 | ||
e3d62d7e | 1443 | /* Done. */ |
f3286a3a CM |
1444 | md->egress_for_echannel[echannel].equeue = equeue; |
1445 | md->egress_for_echannel[echannel].headers = headers; | |
e3d62d7e CM |
1446 | return 0; |
1447 | ||
1448 | fail_equeue: | |
1449 | __free_pages(equeue_page, equeue_order); | |
1450 | ||
1451 | fail_edescs: | |
1452 | __free_pages(edescs_page, edescs_order); | |
1453 | ||
1454 | fail_headers: | |
1455 | __free_pages(headers_page, headers_order); | |
1456 | ||
1457 | fail: | |
1458 | return rc; | |
1459 | } | |
1460 | ||
1461 | /* Return channel number for a newly-opened link. */ | |
1462 | static int tile_net_link_open(struct net_device *dev, gxio_mpipe_link_t *link, | |
1463 | const char *link_name) | |
1464 | { | |
f3286a3a CM |
1465 | int instance = mpipe_instance(dev); |
1466 | struct mpipe_data *md = &mpipe_data[instance]; | |
1467 | int rc = gxio_mpipe_link_open(link, &md->context, link_name, 0); | |
e3d62d7e | 1468 | if (rc < 0) { |
f3286a3a CM |
1469 | netdev_err(dev, "Failed to open '%s', mpipe[%d], %d\n", |
1470 | link_name, instance, rc); | |
e3d62d7e CM |
1471 | return rc; |
1472 | } | |
2628e8af CM |
1473 | if (jumbo_num != 0) { |
1474 | u32 attr = GXIO_MPIPE_LINK_RECEIVE_JUMBO; | |
1475 | rc = gxio_mpipe_link_set_attr(link, attr, 1); | |
1476 | if (rc != 0) { | |
1477 | netdev_err(dev, | |
1478 | "Cannot receive jumbo packets on '%s'\n", | |
1479 | link_name); | |
1480 | gxio_mpipe_link_close(link); | |
1481 | return rc; | |
1482 | } | |
1483 | } | |
e3d62d7e CM |
1484 | rc = gxio_mpipe_link_channel(link); |
1485 | if (rc < 0 || rc >= TILE_NET_CHANNELS) { | |
1486 | netdev_err(dev, "gxio_mpipe_link_channel bad value: %d\n", rc); | |
1487 | gxio_mpipe_link_close(link); | |
1488 | return -EINVAL; | |
1489 | } | |
1490 | return rc; | |
1491 | } | |
1492 | ||
1493 | /* Help the kernel activate the given network interface. */ | |
1494 | static int tile_net_open(struct net_device *dev) | |
1495 | { | |
1496 | struct tile_net_priv *priv = netdev_priv(dev); | |
f3286a3a | 1497 | int cpu, rc, instance; |
e3d62d7e CM |
1498 | |
1499 | mutex_lock(&tile_net_devs_for_channel_mutex); | |
1500 | ||
f3286a3a CM |
1501 | /* Get the instance info. */ |
1502 | rc = gxio_mpipe_link_instance(dev->name); | |
1155e964 WY |
1503 | if (rc < 0 || rc >= NR_MPIPE_MAX) { |
1504 | mutex_unlock(&tile_net_devs_for_channel_mutex); | |
f3286a3a | 1505 | return -EIO; |
1155e964 | 1506 | } |
f3286a3a CM |
1507 | |
1508 | priv->instance = rc; | |
1509 | instance = rc; | |
1510 | if (!mpipe_data[rc].context.mmio_fast_base) { | |
1511 | /* Do one-time initialization per instance the first time | |
1512 | * any device is opened. | |
1513 | */ | |
e3d62d7e CM |
1514 | rc = tile_net_init_mpipe(dev); |
1515 | if (rc != 0) | |
1516 | goto fail; | |
1517 | } | |
1518 | ||
1519 | /* Determine if this is the "loopify" device. */ | |
1520 | if (unlikely((loopify_link_name != NULL) && | |
1521 | !strcmp(dev->name, loopify_link_name))) { | |
1522 | rc = tile_net_link_open(dev, &priv->link, "loop0"); | |
1523 | if (rc < 0) | |
1524 | goto fail; | |
1525 | priv->channel = rc; | |
1526 | rc = tile_net_link_open(dev, &priv->loopify_link, "loop1"); | |
1527 | if (rc < 0) | |
1528 | goto fail; | |
1529 | priv->loopify_channel = rc; | |
1530 | priv->echannel = rc; | |
1531 | } else { | |
1532 | rc = tile_net_link_open(dev, &priv->link, dev->name); | |
1533 | if (rc < 0) | |
1534 | goto fail; | |
1535 | priv->channel = rc; | |
1536 | priv->echannel = rc; | |
1537 | } | |
1538 | ||
1539 | /* Initialize egress info (if needed). Once ever, per echannel. */ | |
1540 | rc = tile_net_init_egress(dev, priv->echannel); | |
1541 | if (rc != 0) | |
1542 | goto fail; | |
1543 | ||
f3286a3a | 1544 | mpipe_data[instance].tile_net_devs_for_channel[priv->channel] = dev; |
e3d62d7e CM |
1545 | |
1546 | rc = tile_net_update(dev); | |
1547 | if (rc != 0) | |
1548 | goto fail; | |
1549 | ||
1550 | mutex_unlock(&tile_net_devs_for_channel_mutex); | |
1551 | ||
1552 | /* Initialize the transmit wake timer for this device for each cpu. */ | |
1553 | for_each_online_cpu(cpu) { | |
1554 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
1555 | struct tile_net_tx_wake *tx_wake = | |
f3286a3a | 1556 | &info->mpipe[instance].tx_wake[priv->echannel]; |
e3d62d7e CM |
1557 | |
1558 | hrtimer_init(&tx_wake->timer, CLOCK_MONOTONIC, | |
1559 | HRTIMER_MODE_REL); | |
9b4c341b | 1560 | tx_wake->tx_queue_idx = cpu; |
e3d62d7e CM |
1561 | tx_wake->timer.function = tile_net_handle_tx_wake_timer; |
1562 | tx_wake->dev = dev; | |
1563 | } | |
1564 | ||
1565 | for_each_online_cpu(cpu) | |
1566 | netif_start_subqueue(dev, cpu); | |
1567 | netif_carrier_on(dev); | |
1568 | return 0; | |
1569 | ||
1570 | fail: | |
1571 | if (priv->loopify_channel >= 0) { | |
1572 | if (gxio_mpipe_link_close(&priv->loopify_link) != 0) | |
1573 | netdev_warn(dev, "Failed to close loopify link!\n"); | |
1574 | priv->loopify_channel = -1; | |
1575 | } | |
1576 | if (priv->channel >= 0) { | |
1577 | if (gxio_mpipe_link_close(&priv->link) != 0) | |
1578 | netdev_warn(dev, "Failed to close link!\n"); | |
1579 | priv->channel = -1; | |
1580 | } | |
1581 | priv->echannel = -1; | |
f3286a3a | 1582 | mpipe_data[instance].tile_net_devs_for_channel[priv->channel] = NULL; |
e3d62d7e CM |
1583 | mutex_unlock(&tile_net_devs_for_channel_mutex); |
1584 | ||
1585 | /* Don't return raw gxio error codes to generic Linux. */ | |
1586 | return (rc > -512) ? rc : -EIO; | |
1587 | } | |
1588 | ||
1589 | /* Help the kernel deactivate the given network interface. */ | |
1590 | static int tile_net_stop(struct net_device *dev) | |
1591 | { | |
1592 | struct tile_net_priv *priv = netdev_priv(dev); | |
1593 | int cpu; | |
f3286a3a CM |
1594 | int instance = priv->instance; |
1595 | struct mpipe_data *md = &mpipe_data[instance]; | |
e3d62d7e CM |
1596 | |
1597 | for_each_online_cpu(cpu) { | |
1598 | struct tile_net_info *info = &per_cpu(per_cpu_info, cpu); | |
1599 | struct tile_net_tx_wake *tx_wake = | |
f3286a3a | 1600 | &info->mpipe[instance].tx_wake[priv->echannel]; |
e3d62d7e CM |
1601 | |
1602 | hrtimer_cancel(&tx_wake->timer); | |
1603 | netif_stop_subqueue(dev, cpu); | |
1604 | } | |
1605 | ||
1606 | mutex_lock(&tile_net_devs_for_channel_mutex); | |
f3286a3a | 1607 | md->tile_net_devs_for_channel[priv->channel] = NULL; |
e3d62d7e CM |
1608 | (void)tile_net_update(dev); |
1609 | if (priv->loopify_channel >= 0) { | |
1610 | if (gxio_mpipe_link_close(&priv->loopify_link) != 0) | |
1611 | netdev_warn(dev, "Failed to close loopify link!\n"); | |
1612 | priv->loopify_channel = -1; | |
1613 | } | |
1614 | if (priv->channel >= 0) { | |
1615 | if (gxio_mpipe_link_close(&priv->link) != 0) | |
1616 | netdev_warn(dev, "Failed to close link!\n"); | |
1617 | priv->channel = -1; | |
1618 | } | |
1619 | priv->echannel = -1; | |
1620 | mutex_unlock(&tile_net_devs_for_channel_mutex); | |
1621 | ||
1622 | return 0; | |
1623 | } | |
1624 | ||
1625 | /* Determine the VA for a fragment. */ | |
1626 | static inline void *tile_net_frag_buf(skb_frag_t *f) | |
1627 | { | |
1628 | unsigned long pfn = page_to_pfn(skb_frag_page(f)); | |
1629 | return pfn_to_kaddr(pfn) + f->page_offset; | |
1630 | } | |
1631 | ||
1632 | /* Acquire a completion entry and an egress slot, or if we can't, | |
1633 | * stop the queue and schedule the tx_wake timer. | |
1634 | */ | |
1635 | static s64 tile_net_equeue_try_reserve(struct net_device *dev, | |
9b4c341b | 1636 | int tx_queue_idx, |
e3d62d7e CM |
1637 | struct tile_net_comps *comps, |
1638 | gxio_mpipe_equeue_t *equeue, | |
1639 | int num_edescs) | |
1640 | { | |
1641 | /* Try to acquire a completion entry. */ | |
1642 | if (comps->comp_next - comps->comp_last < TILE_NET_MAX_COMPS - 1 || | |
1643 | tile_net_free_comps(equeue, comps, 32, false) != 0) { | |
1644 | ||
1645 | /* Try to acquire an egress slot. */ | |
1646 | s64 slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs); | |
1647 | if (slot >= 0) | |
1648 | return slot; | |
1649 | ||
1650 | /* Freeing some completions gives the equeue time to drain. */ | |
1651 | tile_net_free_comps(equeue, comps, TILE_NET_MAX_COMPS, false); | |
1652 | ||
1653 | slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs); | |
1654 | if (slot >= 0) | |
1655 | return slot; | |
1656 | } | |
1657 | ||
1658 | /* Still nothing; give up and stop the queue for a short while. */ | |
9b4c341b CM |
1659 | netif_stop_subqueue(dev, tx_queue_idx); |
1660 | tile_net_schedule_tx_wake_timer(dev, tx_queue_idx); | |
e3d62d7e CM |
1661 | return -1; |
1662 | } | |
1663 | ||
1664 | /* Determine how many edesc's are needed for TSO. | |
1665 | * | |
1666 | * Sometimes, if "sendfile()" requires copying, we will be called with | |
1667 | * "data" containing the header and payload, with "frags" being empty. | |
1668 | * Sometimes, for example when using NFS over TCP, a single segment can | |
1669 | * span 3 fragments. This requires special care. | |
1670 | */ | |
1671 | static int tso_count_edescs(struct sk_buff *skb) | |
1672 | { | |
1673 | struct skb_shared_info *sh = skb_shinfo(skb); | |
8388546e | 1674 | unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb); |
3da3fff8 | 1675 | unsigned int data_len = skb->len - sh_len; |
e3d62d7e CM |
1676 | unsigned int p_len = sh->gso_size; |
1677 | long f_id = -1; /* id of the current fragment */ | |
3da3fff8 CM |
1678 | long f_size = skb_headlen(skb) - sh_len; /* current fragment size */ |
1679 | long f_used = 0; /* bytes used from the current fragment */ | |
e3d62d7e CM |
1680 | long n; /* size of the current piece of payload */ |
1681 | int num_edescs = 0; | |
1682 | int segment; | |
1683 | ||
1684 | for (segment = 0; segment < sh->gso_segs; segment++) { | |
1685 | ||
1686 | unsigned int p_used = 0; | |
1687 | ||
1688 | /* One edesc for header and for each piece of the payload. */ | |
1689 | for (num_edescs++; p_used < p_len; num_edescs++) { | |
1690 | ||
1691 | /* Advance as needed. */ | |
1692 | while (f_used >= f_size) { | |
1693 | f_id++; | |
3da3fff8 | 1694 | f_size = skb_frag_size(&sh->frags[f_id]); |
e3d62d7e CM |
1695 | f_used = 0; |
1696 | } | |
1697 | ||
1698 | /* Use bytes from the current fragment. */ | |
1699 | n = p_len - p_used; | |
1700 | if (n > f_size - f_used) | |
1701 | n = f_size - f_used; | |
1702 | f_used += n; | |
1703 | p_used += n; | |
1704 | } | |
1705 | ||
1706 | /* The last segment may be less than gso_size. */ | |
1707 | data_len -= p_len; | |
1708 | if (data_len < p_len) | |
1709 | p_len = data_len; | |
1710 | } | |
1711 | ||
1712 | return num_edescs; | |
1713 | } | |
1714 | ||
2c7d04a9 | 1715 | /* Prepare modified copies of the skbuff headers. */ |
e3d62d7e CM |
1716 | static void tso_headers_prepare(struct sk_buff *skb, unsigned char *headers, |
1717 | s64 slot) | |
1718 | { | |
1719 | struct skb_shared_info *sh = skb_shinfo(skb); | |
1720 | struct iphdr *ih; | |
2c7d04a9 | 1721 | struct ipv6hdr *ih6; |
e3d62d7e | 1722 | struct tcphdr *th; |
8388546e | 1723 | unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb); |
3da3fff8 | 1724 | unsigned int data_len = skb->len - sh_len; |
e3d62d7e | 1725 | unsigned char *data = skb->data; |
8388546e | 1726 | unsigned int ih_off, th_off, p_len; |
444fa88a CM |
1727 | unsigned int isum_seed, tsum_seed, seq; |
1728 | unsigned int uninitialized_var(id); | |
2c7d04a9 | 1729 | int is_ipv6; |
e3d62d7e | 1730 | long f_id = -1; /* id of the current fragment */ |
3da3fff8 CM |
1731 | long f_size = skb_headlen(skb) - sh_len; /* current fragment size */ |
1732 | long f_used = 0; /* bytes used from the current fragment */ | |
e3d62d7e CM |
1733 | long n; /* size of the current piece of payload */ |
1734 | int segment; | |
1735 | ||
1736 | /* Locate original headers and compute various lengths. */ | |
2c7d04a9 CM |
1737 | is_ipv6 = skb_is_gso_v6(skb); |
1738 | if (is_ipv6) { | |
1739 | ih6 = ipv6_hdr(skb); | |
1740 | ih_off = skb_network_offset(skb); | |
1741 | } else { | |
1742 | ih = ip_hdr(skb); | |
1743 | ih_off = skb_network_offset(skb); | |
1744 | isum_seed = ((0xFFFF - ih->check) + | |
1745 | (0xFFFF - ih->tot_len) + | |
1746 | (0xFFFF - ih->id)); | |
1747 | id = ntohs(ih->id); | |
1748 | } | |
1749 | ||
e3d62d7e | 1750 | th = tcp_hdr(skb); |
e3d62d7e | 1751 | th_off = skb_transport_offset(skb); |
e3d62d7e CM |
1752 | p_len = sh->gso_size; |
1753 | ||
3da3fff8 | 1754 | tsum_seed = th->check + (0xFFFF ^ htons(skb->len)); |
e3d62d7e CM |
1755 | seq = ntohl(th->seq); |
1756 | ||
1757 | /* Prepare all the headers. */ | |
1758 | for (segment = 0; segment < sh->gso_segs; segment++) { | |
1759 | unsigned char *buf; | |
1760 | unsigned int p_used = 0; | |
1761 | ||
1762 | /* Copy to the header memory for this segment. */ | |
1763 | buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES + | |
1764 | NET_IP_ALIGN; | |
1765 | memcpy(buf, data, sh_len); | |
1766 | ||
1767 | /* Update copied ip header. */ | |
2c7d04a9 CM |
1768 | if (is_ipv6) { |
1769 | ih6 = (struct ipv6hdr *)(buf + ih_off); | |
1770 | ih6->payload_len = htons(sh_len + p_len - ih_off - | |
1771 | sizeof(*ih6)); | |
1772 | } else { | |
1773 | ih = (struct iphdr *)(buf + ih_off); | |
1774 | ih->tot_len = htons(sh_len + p_len - ih_off); | |
444fa88a | 1775 | ih->id = htons(id++); |
2c7d04a9 CM |
1776 | ih->check = csum_long(isum_seed + ih->tot_len + |
1777 | ih->id) ^ 0xffff; | |
1778 | } | |
e3d62d7e CM |
1779 | |
1780 | /* Update copied tcp header. */ | |
1781 | th = (struct tcphdr *)(buf + th_off); | |
1782 | th->seq = htonl(seq); | |
1783 | th->check = csum_long(tsum_seed + htons(sh_len + p_len)); | |
1784 | if (segment != sh->gso_segs - 1) { | |
1785 | th->fin = 0; | |
1786 | th->psh = 0; | |
1787 | } | |
1788 | ||
1789 | /* Skip past the header. */ | |
1790 | slot++; | |
1791 | ||
1792 | /* Skip past the payload. */ | |
1793 | while (p_used < p_len) { | |
1794 | ||
1795 | /* Advance as needed. */ | |
1796 | while (f_used >= f_size) { | |
1797 | f_id++; | |
3da3fff8 | 1798 | f_size = skb_frag_size(&sh->frags[f_id]); |
e3d62d7e CM |
1799 | f_used = 0; |
1800 | } | |
1801 | ||
1802 | /* Use bytes from the current fragment. */ | |
1803 | n = p_len - p_used; | |
1804 | if (n > f_size - f_used) | |
1805 | n = f_size - f_used; | |
1806 | f_used += n; | |
1807 | p_used += n; | |
1808 | ||
1809 | slot++; | |
1810 | } | |
1811 | ||
e3d62d7e CM |
1812 | seq += p_len; |
1813 | ||
1814 | /* The last segment may be less than gso_size. */ | |
1815 | data_len -= p_len; | |
1816 | if (data_len < p_len) | |
1817 | p_len = data_len; | |
1818 | } | |
1819 | ||
1820 | /* Flush the headers so they are ready for hardware DMA. */ | |
1821 | wmb(); | |
1822 | } | |
1823 | ||
1824 | /* Pass all the data to mpipe for egress. */ | |
1825 | static void tso_egress(struct net_device *dev, gxio_mpipe_equeue_t *equeue, | |
1826 | struct sk_buff *skb, unsigned char *headers, s64 slot) | |
1827 | { | |
e3d62d7e | 1828 | struct skb_shared_info *sh = skb_shinfo(skb); |
f3286a3a CM |
1829 | int instance = mpipe_instance(dev); |
1830 | struct mpipe_data *md = &mpipe_data[instance]; | |
8388546e | 1831 | unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb); |
3da3fff8 | 1832 | unsigned int data_len = skb->len - sh_len; |
e3d62d7e CM |
1833 | unsigned int p_len = sh->gso_size; |
1834 | gxio_mpipe_edesc_t edesc_head = { { 0 } }; | |
1835 | gxio_mpipe_edesc_t edesc_body = { { 0 } }; | |
1836 | long f_id = -1; /* id of the current fragment */ | |
3da3fff8 CM |
1837 | long f_size = skb_headlen(skb) - sh_len; /* current fragment size */ |
1838 | long f_used = 0; /* bytes used from the current fragment */ | |
1839 | void *f_data = skb->data + sh_len; | |
e3d62d7e CM |
1840 | long n; /* size of the current piece of payload */ |
1841 | unsigned long tx_packets = 0, tx_bytes = 0; | |
8388546e | 1842 | unsigned int csum_start; |
e3d62d7e CM |
1843 | int segment; |
1844 | ||
1845 | /* Prepare to egress the headers: set up header edesc. */ | |
1846 | csum_start = skb_checksum_start_offset(skb); | |
e3d62d7e CM |
1847 | edesc_head.csum = 1; |
1848 | edesc_head.csum_start = csum_start; | |
1849 | edesc_head.csum_dest = csum_start + skb->csum_offset; | |
1850 | edesc_head.xfer_size = sh_len; | |
1851 | ||
1852 | /* This is only used to specify the TLB. */ | |
f3286a3a CM |
1853 | edesc_head.stack_idx = md->first_buffer_stack; |
1854 | edesc_body.stack_idx = md->first_buffer_stack; | |
e3d62d7e CM |
1855 | |
1856 | /* Egress all the edescs. */ | |
1857 | for (segment = 0; segment < sh->gso_segs; segment++) { | |
e3d62d7e CM |
1858 | unsigned char *buf; |
1859 | unsigned int p_used = 0; | |
1860 | ||
1861 | /* Egress the header. */ | |
1862 | buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES + | |
1863 | NET_IP_ALIGN; | |
1864 | edesc_head.va = va_to_tile_io_addr(buf); | |
1865 | gxio_mpipe_equeue_put_at(equeue, edesc_head, slot); | |
1866 | slot++; | |
1867 | ||
1868 | /* Egress the payload. */ | |
1869 | while (p_used < p_len) { | |
3da3fff8 | 1870 | void *va; |
e3d62d7e CM |
1871 | |
1872 | /* Advance as needed. */ | |
1873 | while (f_used >= f_size) { | |
1874 | f_id++; | |
3da3fff8 | 1875 | f_size = skb_frag_size(&sh->frags[f_id]); |
8388546e | 1876 | f_data = tile_net_frag_buf(&sh->frags[f_id]); |
3da3fff8 | 1877 | f_used = 0; |
e3d62d7e CM |
1878 | } |
1879 | ||
3da3fff8 CM |
1880 | va = f_data + f_used; |
1881 | ||
e3d62d7e CM |
1882 | /* Use bytes from the current fragment. */ |
1883 | n = p_len - p_used; | |
1884 | if (n > f_size - f_used) | |
1885 | n = f_size - f_used; | |
1886 | f_used += n; | |
1887 | p_used += n; | |
1888 | ||
1889 | /* Egress a piece of the payload. */ | |
3da3fff8 | 1890 | edesc_body.va = va_to_tile_io_addr(va); |
e3d62d7e CM |
1891 | edesc_body.xfer_size = n; |
1892 | edesc_body.bound = !(p_used < p_len); | |
1893 | gxio_mpipe_equeue_put_at(equeue, edesc_body, slot); | |
1894 | slot++; | |
1895 | } | |
1896 | ||
1897 | tx_packets++; | |
1898 | tx_bytes += sh_len + p_len; | |
1899 | ||
1900 | /* The last segment may be less than gso_size. */ | |
1901 | data_len -= p_len; | |
1902 | if (data_len < p_len) | |
1903 | p_len = data_len; | |
1904 | } | |
1905 | ||
1906 | /* Update stats. */ | |
ad018185 CM |
1907 | tile_net_stats_add(tx_packets, &dev->stats.tx_packets); |
1908 | tile_net_stats_add(tx_bytes, &dev->stats.tx_bytes); | |
e3d62d7e CM |
1909 | } |
1910 | ||
1911 | /* Do "TSO" handling for egress. | |
1912 | * | |
1913 | * Normally drivers set NETIF_F_TSO only to support hardware TSO; | |
1914 | * otherwise the stack uses scatter-gather to implement GSO in software. | |
1915 | * On our testing, enabling GSO support (via NETIF_F_SG) drops network | |
1916 | * performance down to around 7.5 Gbps on the 10G interfaces, although | |
1917 | * also dropping cpu utilization way down, to under 8%. But | |
1918 | * implementing "TSO" in the driver brings performance back up to line | |
1919 | * rate, while dropping cpu usage even further, to less than 4%. In | |
1920 | * practice, profiling of GSO shows that skb_segment() is what causes | |
1921 | * the performance overheads; we benefit in the driver from using | |
1922 | * preallocated memory to duplicate the TCP/IP headers. | |
1923 | */ | |
1924 | static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev) | |
1925 | { | |
1926 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
1927 | struct tile_net_priv *priv = netdev_priv(dev); | |
1928 | int channel = priv->echannel; | |
f3286a3a CM |
1929 | int instance = priv->instance; |
1930 | struct mpipe_data *md = &mpipe_data[instance]; | |
1931 | struct tile_net_egress *egress = &md->egress_for_echannel[channel]; | |
1932 | struct tile_net_comps *comps = | |
1933 | info->mpipe[instance].comps_for_echannel[channel]; | |
e3d62d7e CM |
1934 | gxio_mpipe_equeue_t *equeue = egress->equeue; |
1935 | unsigned long irqflags; | |
1936 | int num_edescs; | |
1937 | s64 slot; | |
1938 | ||
1939 | /* Determine how many mpipe edesc's are needed. */ | |
1940 | num_edescs = tso_count_edescs(skb); | |
1941 | ||
1942 | local_irq_save(irqflags); | |
1943 | ||
1944 | /* Try to acquire a completion entry and an egress slot. */ | |
9b4c341b CM |
1945 | slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps, |
1946 | equeue, num_edescs); | |
e3d62d7e CM |
1947 | if (slot < 0) { |
1948 | local_irq_restore(irqflags); | |
1949 | return NETDEV_TX_BUSY; | |
1950 | } | |
1951 | ||
1952 | /* Set up copies of header data properly. */ | |
1953 | tso_headers_prepare(skb, egress->headers, slot); | |
1954 | ||
1955 | /* Actually pass the data to the network hardware. */ | |
1956 | tso_egress(dev, equeue, skb, egress->headers, slot); | |
1957 | ||
1958 | /* Add a completion record. */ | |
1959 | add_comp(equeue, comps, slot + num_edescs - 1, skb); | |
1960 | ||
1961 | local_irq_restore(irqflags); | |
1962 | ||
1963 | /* Make sure the egress timer is scheduled. */ | |
1964 | tile_net_schedule_egress_timer(); | |
1965 | ||
1966 | return NETDEV_TX_OK; | |
1967 | } | |
1968 | ||
1969 | /* Analyze the body and frags for a transmit request. */ | |
1970 | static unsigned int tile_net_tx_frags(struct frag *frags, | |
1971 | struct sk_buff *skb, | |
1972 | void *b_data, unsigned int b_len) | |
1973 | { | |
1974 | unsigned int i, n = 0; | |
1975 | ||
1976 | struct skb_shared_info *sh = skb_shinfo(skb); | |
1977 | ||
1978 | if (b_len != 0) { | |
1979 | frags[n].buf = b_data; | |
1980 | frags[n++].length = b_len; | |
1981 | } | |
1982 | ||
1983 | for (i = 0; i < sh->nr_frags; i++) { | |
1984 | skb_frag_t *f = &sh->frags[i]; | |
1985 | frags[n].buf = tile_net_frag_buf(f); | |
1986 | frags[n++].length = skb_frag_size(f); | |
1987 | } | |
1988 | ||
1989 | return n; | |
1990 | } | |
1991 | ||
1992 | /* Help the kernel transmit a packet. */ | |
1993 | static int tile_net_tx(struct sk_buff *skb, struct net_device *dev) | |
1994 | { | |
1995 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
1996 | struct tile_net_priv *priv = netdev_priv(dev); | |
f3286a3a CM |
1997 | int instance = priv->instance; |
1998 | struct mpipe_data *md = &mpipe_data[instance]; | |
1999 | struct tile_net_egress *egress = | |
2000 | &md->egress_for_echannel[priv->echannel]; | |
e3d62d7e CM |
2001 | gxio_mpipe_equeue_t *equeue = egress->equeue; |
2002 | struct tile_net_comps *comps = | |
f3286a3a | 2003 | info->mpipe[instance].comps_for_echannel[priv->echannel]; |
e3d62d7e CM |
2004 | unsigned int len = skb->len; |
2005 | unsigned char *data = skb->data; | |
2006 | unsigned int num_edescs; | |
2007 | struct frag frags[MAX_FRAGS]; | |
2008 | gxio_mpipe_edesc_t edescs[MAX_FRAGS]; | |
2009 | unsigned long irqflags; | |
2010 | gxio_mpipe_edesc_t edesc = { { 0 } }; | |
2011 | unsigned int i; | |
2012 | s64 slot; | |
2013 | ||
2014 | if (skb_is_gso(skb)) | |
2015 | return tile_net_tx_tso(skb, dev); | |
2016 | ||
2017 | num_edescs = tile_net_tx_frags(frags, skb, data, skb_headlen(skb)); | |
2018 | ||
2019 | /* This is only used to specify the TLB. */ | |
f3286a3a | 2020 | edesc.stack_idx = md->first_buffer_stack; |
e3d62d7e CM |
2021 | |
2022 | /* Prepare the edescs. */ | |
2023 | for (i = 0; i < num_edescs; i++) { | |
2024 | edesc.xfer_size = frags[i].length; | |
2025 | edesc.va = va_to_tile_io_addr(frags[i].buf); | |
2026 | edescs[i] = edesc; | |
2027 | } | |
2028 | ||
2029 | /* Mark the final edesc. */ | |
2030 | edescs[num_edescs - 1].bound = 1; | |
2031 | ||
2032 | /* Add checksum info to the initial edesc, if needed. */ | |
2033 | if (skb->ip_summed == CHECKSUM_PARTIAL) { | |
2034 | unsigned int csum_start = skb_checksum_start_offset(skb); | |
2035 | edescs[0].csum = 1; | |
2036 | edescs[0].csum_start = csum_start; | |
2037 | edescs[0].csum_dest = csum_start + skb->csum_offset; | |
2038 | } | |
2039 | ||
2040 | local_irq_save(irqflags); | |
2041 | ||
2042 | /* Try to acquire a completion entry and an egress slot. */ | |
9b4c341b CM |
2043 | slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps, |
2044 | equeue, num_edescs); | |
e3d62d7e CM |
2045 | if (slot < 0) { |
2046 | local_irq_restore(irqflags); | |
2047 | return NETDEV_TX_BUSY; | |
2048 | } | |
2049 | ||
2050 | for (i = 0; i < num_edescs; i++) | |
2051 | gxio_mpipe_equeue_put_at(equeue, edescs[i], slot++); | |
2052 | ||
9ab5ec59 CM |
2053 | /* Store TX timestamp if needed. */ |
2054 | tile_tx_timestamp(skb, instance); | |
2055 | ||
e3d62d7e CM |
2056 | /* Add a completion record. */ |
2057 | add_comp(equeue, comps, slot - 1, skb); | |
2058 | ||
2059 | /* NOTE: Use ETH_ZLEN for short packets (e.g. 42 < 60). */ | |
ad018185 | 2060 | tile_net_stats_add(1, &dev->stats.tx_packets); |
e3d62d7e | 2061 | tile_net_stats_add(max_t(unsigned int, len, ETH_ZLEN), |
ad018185 | 2062 | &dev->stats.tx_bytes); |
e3d62d7e CM |
2063 | |
2064 | local_irq_restore(irqflags); | |
2065 | ||
2066 | /* Make sure the egress timer is scheduled. */ | |
2067 | tile_net_schedule_egress_timer(); | |
2068 | ||
2069 | return NETDEV_TX_OK; | |
2070 | } | |
2071 | ||
2072 | /* Return subqueue id on this core (one per core). */ | |
f663dd9a | 2073 | static u16 tile_net_select_queue(struct net_device *dev, struct sk_buff *skb, |
99932d4f | 2074 | void *accel_priv, select_queue_fallback_t fallback) |
e3d62d7e CM |
2075 | { |
2076 | return smp_processor_id(); | |
2077 | } | |
2078 | ||
2079 | /* Deal with a transmit timeout. */ | |
2080 | static void tile_net_tx_timeout(struct net_device *dev) | |
2081 | { | |
2082 | int cpu; | |
2083 | ||
2084 | for_each_online_cpu(cpu) | |
2085 | netif_wake_subqueue(dev, cpu); | |
2086 | } | |
2087 | ||
2088 | /* Ioctl commands. */ | |
2089 | static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd) | |
2090 | { | |
9ab5ec59 | 2091 | if (cmd == SIOCSHWTSTAMP) |
6ab96d1e BH |
2092 | return tile_hwtstamp_set(dev, rq); |
2093 | if (cmd == SIOCGHWTSTAMP) | |
2094 | return tile_hwtstamp_get(dev, rq); | |
9ab5ec59 | 2095 | |
e3d62d7e CM |
2096 | return -EOPNOTSUPP; |
2097 | } | |
2098 | ||
e3d62d7e CM |
2099 | /* Change the MTU. */ |
2100 | static int tile_net_change_mtu(struct net_device *dev, int new_mtu) | |
2101 | { | |
2628e8af CM |
2102 | if (new_mtu < 68) |
2103 | return -EINVAL; | |
2104 | if (new_mtu > ((jumbo_num != 0) ? 9000 : 1500)) | |
e3d62d7e CM |
2105 | return -EINVAL; |
2106 | dev->mtu = new_mtu; | |
2107 | return 0; | |
2108 | } | |
2109 | ||
2110 | /* Change the Ethernet address of the NIC. | |
2111 | * | |
2112 | * The hypervisor driver does not support changing MAC address. However, | |
2113 | * the hardware does not do anything with the MAC address, so the address | |
2114 | * which gets used on outgoing packets, and which is accepted on incoming | |
2115 | * packets, is completely up to us. | |
2116 | * | |
2117 | * Returns 0 on success, negative on failure. | |
2118 | */ | |
2119 | static int tile_net_set_mac_address(struct net_device *dev, void *p) | |
2120 | { | |
2121 | struct sockaddr *addr = p; | |
2122 | ||
2123 | if (!is_valid_ether_addr(addr->sa_data)) | |
2124 | return -EINVAL; | |
2125 | memcpy(dev->dev_addr, addr->sa_data, dev->addr_len); | |
2126 | return 0; | |
2127 | } | |
2128 | ||
2129 | #ifdef CONFIG_NET_POLL_CONTROLLER | |
2130 | /* Polling 'interrupt' - used by things like netconsole to send skbs | |
2131 | * without having to re-enable interrupts. It's not called while | |
2132 | * the interrupt routine is executing. | |
2133 | */ | |
2134 | static void tile_net_netpoll(struct net_device *dev) | |
2135 | { | |
f3286a3a CM |
2136 | int instance = mpipe_instance(dev); |
2137 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
2138 | struct mpipe_data *md = &mpipe_data[instance]; | |
2139 | ||
2140 | disable_percpu_irq(md->ingress_irq); | |
2141 | napi_schedule(&info->mpipe[instance].napi); | |
2142 | enable_percpu_irq(md->ingress_irq, 0); | |
e3d62d7e CM |
2143 | } |
2144 | #endif | |
2145 | ||
2146 | static const struct net_device_ops tile_net_ops = { | |
2147 | .ndo_open = tile_net_open, | |
2148 | .ndo_stop = tile_net_stop, | |
2149 | .ndo_start_xmit = tile_net_tx, | |
2150 | .ndo_select_queue = tile_net_select_queue, | |
2151 | .ndo_do_ioctl = tile_net_ioctl, | |
e3d62d7e CM |
2152 | .ndo_change_mtu = tile_net_change_mtu, |
2153 | .ndo_tx_timeout = tile_net_tx_timeout, | |
2154 | .ndo_set_mac_address = tile_net_set_mac_address, | |
2155 | #ifdef CONFIG_NET_POLL_CONTROLLER | |
2156 | .ndo_poll_controller = tile_net_netpoll, | |
2157 | #endif | |
2158 | }; | |
2159 | ||
2160 | /* The setup function. | |
2161 | * | |
2162 | * This uses ether_setup() to assign various fields in dev, including | |
2163 | * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields. | |
2164 | */ | |
2165 | static void tile_net_setup(struct net_device *dev) | |
2166 | { | |
a8eaed55 CM |
2167 | netdev_features_t features = 0; |
2168 | ||
e3d62d7e CM |
2169 | ether_setup(dev); |
2170 | dev->netdev_ops = &tile_net_ops; | |
2171 | dev->watchdog_timeo = TILE_NET_TIMEOUT; | |
e3d62d7e | 2172 | dev->mtu = 1500; |
a8eaed55 | 2173 | |
a8eaed55 CM |
2174 | features |= NETIF_F_HW_CSUM; |
2175 | features |= NETIF_F_SG; | |
2176 | features |= NETIF_F_TSO; | |
2c7d04a9 | 2177 | features |= NETIF_F_TSO6; |
a8eaed55 CM |
2178 | |
2179 | dev->hw_features |= features; | |
2180 | dev->vlan_features |= features; | |
2181 | dev->features |= features; | |
e3d62d7e CM |
2182 | } |
2183 | ||
2184 | /* Allocate the device structure, register the device, and obtain the | |
2185 | * MAC address from the hypervisor. | |
2186 | */ | |
2187 | static void tile_net_dev_init(const char *name, const uint8_t *mac) | |
2188 | { | |
2189 | int ret; | |
2190 | int i; | |
2191 | int nz_addr = 0; | |
2192 | struct net_device *dev; | |
2193 | struct tile_net_priv *priv; | |
2194 | ||
2195 | /* HACK: Ignore "loop" links. */ | |
2196 | if (strncmp(name, "loop", 4) == 0) | |
2197 | return; | |
2198 | ||
2199 | /* Allocate the device structure. Normally, "name" is a | |
2200 | * template, instantiated by register_netdev(), but not for us. | |
2201 | */ | |
2202 | dev = alloc_netdev_mqs(sizeof(*priv), name, tile_net_setup, | |
2203 | NR_CPUS, 1); | |
2204 | if (!dev) { | |
2205 | pr_err("alloc_netdev_mqs(%s) failed\n", name); | |
2206 | return; | |
2207 | } | |
2208 | ||
2209 | /* Initialize "priv". */ | |
2210 | priv = netdev_priv(dev); | |
2211 | memset(priv, 0, sizeof(*priv)); | |
2212 | priv->dev = dev; | |
2213 | priv->channel = -1; | |
2214 | priv->loopify_channel = -1; | |
2215 | priv->echannel = -1; | |
9ab5ec59 | 2216 | init_ptp_dev(priv); |
e3d62d7e CM |
2217 | |
2218 | /* Get the MAC address and set it in the device struct; this must | |
2219 | * be done before the device is opened. If the MAC is all zeroes, | |
2220 | * we use a random address, since we're probably on the simulator. | |
2221 | */ | |
2222 | for (i = 0; i < 6; i++) | |
2223 | nz_addr |= mac[i]; | |
2224 | ||
2225 | if (nz_addr) { | |
d458cdf7 | 2226 | memcpy(dev->dev_addr, mac, ETH_ALEN); |
e3d62d7e CM |
2227 | dev->addr_len = 6; |
2228 | } else { | |
c8ab13fb | 2229 | eth_hw_addr_random(dev); |
e3d62d7e CM |
2230 | } |
2231 | ||
2232 | /* Register the network device. */ | |
2233 | ret = register_netdev(dev); | |
2234 | if (ret) { | |
2235 | netdev_err(dev, "register_netdev failed %d\n", ret); | |
2236 | free_netdev(dev); | |
2237 | return; | |
2238 | } | |
2239 | } | |
2240 | ||
2241 | /* Per-cpu module initialization. */ | |
2242 | static void tile_net_init_module_percpu(void *unused) | |
2243 | { | |
2244 | struct tile_net_info *info = &__get_cpu_var(per_cpu_info); | |
2245 | int my_cpu = smp_processor_id(); | |
f3286a3a | 2246 | int instance; |
e3d62d7e | 2247 | |
f3286a3a CM |
2248 | for (instance = 0; instance < NR_MPIPE_MAX; instance++) { |
2249 | info->mpipe[instance].has_iqueue = false; | |
2250 | info->mpipe[instance].instance = instance; | |
2251 | } | |
e3d62d7e CM |
2252 | info->my_cpu = my_cpu; |
2253 | ||
2254 | /* Initialize the egress timer. */ | |
2255 | hrtimer_init(&info->egress_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2256 | info->egress_timer.function = tile_net_handle_egress_timer; | |
2257 | } | |
2258 | ||
2259 | /* Module initialization. */ | |
2260 | static int __init tile_net_init_module(void) | |
2261 | { | |
2262 | int i; | |
2263 | char name[GXIO_MPIPE_LINK_NAME_LEN]; | |
2264 | uint8_t mac[6]; | |
2265 | ||
2266 | pr_info("Tilera Network Driver\n"); | |
2267 | ||
f3286a3a CM |
2268 | BUILD_BUG_ON(NR_MPIPE_MAX != 2); |
2269 | ||
e3d62d7e CM |
2270 | mutex_init(&tile_net_devs_for_channel_mutex); |
2271 | ||
2272 | /* Initialize each CPU. */ | |
2273 | on_each_cpu(tile_net_init_module_percpu, NULL, 1); | |
2274 | ||
2275 | /* Find out what devices we have, and initialize them. */ | |
2276 | for (i = 0; gxio_mpipe_link_enumerate_mac(i, name, mac) >= 0; i++) | |
2277 | tile_net_dev_init(name, mac); | |
2278 | ||
2279 | if (!network_cpus_init()) | |
2280 | network_cpus_map = *cpu_online_mask; | |
2281 | ||
2282 | return 0; | |
2283 | } | |
2284 | ||
2285 | module_init(tile_net_init_module); |