]>
Commit | Line | Data |
---|---|---|
8ceee660 BH |
1 | /**************************************************************************** |
2 | * Driver for Solarflare Solarstorm network controllers and boards | |
3 | * Copyright 2005-2006 Fen Systems Ltd. | |
4 | * Copyright 2005-2008 Solarflare Communications Inc. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify it | |
7 | * under the terms of the GNU General Public License version 2 as published | |
8 | * by the Free Software Foundation, incorporated herein by reference. | |
9 | */ | |
10 | ||
11 | #include <linux/module.h> | |
12 | #include <linux/pci.h> | |
13 | #include <linux/netdevice.h> | |
14 | #include <linux/etherdevice.h> | |
15 | #include <linux/delay.h> | |
16 | #include <linux/notifier.h> | |
17 | #include <linux/ip.h> | |
18 | #include <linux/tcp.h> | |
19 | #include <linux/in.h> | |
20 | #include <linux/crc32.h> | |
21 | #include <linux/ethtool.h> | |
22 | #include "net_driver.h" | |
23 | #include "gmii.h" | |
24 | #include "ethtool.h" | |
25 | #include "tx.h" | |
26 | #include "rx.h" | |
27 | #include "efx.h" | |
28 | #include "mdio_10g.h" | |
29 | #include "falcon.h" | |
30 | #include "workarounds.h" | |
31 | #include "mac.h" | |
32 | ||
33 | #define EFX_MAX_MTU (9 * 1024) | |
34 | ||
35 | /* RX slow fill workqueue. If memory allocation fails in the fast path, | |
36 | * a work item is pushed onto this work queue to retry the allocation later, | |
37 | * to avoid the NIC being starved of RX buffers. Since this is a per cpu | |
38 | * workqueue, there is nothing to be gained in making it per NIC | |
39 | */ | |
40 | static struct workqueue_struct *refill_workqueue; | |
41 | ||
42 | /************************************************************************** | |
43 | * | |
44 | * Configurable values | |
45 | * | |
46 | *************************************************************************/ | |
47 | ||
48 | /* | |
49 | * Enable large receive offload (LRO) aka soft segment reassembly (SSR) | |
50 | * | |
51 | * This sets the default for new devices. It can be controlled later | |
52 | * using ethtool. | |
53 | */ | |
54 | static int lro = 1; | |
55 | module_param(lro, int, 0644); | |
56 | MODULE_PARM_DESC(lro, "Large receive offload acceleration"); | |
57 | ||
58 | /* | |
59 | * Use separate channels for TX and RX events | |
60 | * | |
61 | * Set this to 1 to use separate channels for TX and RX. It allows us to | |
62 | * apply a higher level of interrupt moderation to TX events. | |
63 | * | |
64 | * This is forced to 0 for MSI interrupt mode as the interrupt vector | |
65 | * is not written | |
66 | */ | |
67 | static unsigned int separate_tx_and_rx_channels = 1; | |
68 | ||
69 | /* This is the weight assigned to each of the (per-channel) virtual | |
70 | * NAPI devices. | |
71 | */ | |
72 | static int napi_weight = 64; | |
73 | ||
74 | /* This is the time (in jiffies) between invocations of the hardware | |
75 | * monitor, which checks for known hardware bugs and resets the | |
76 | * hardware and driver as necessary. | |
77 | */ | |
78 | unsigned int efx_monitor_interval = 1 * HZ; | |
79 | ||
80 | /* This controls whether or not the hardware monitor will trigger a | |
81 | * reset when it detects an error condition. | |
82 | */ | |
83 | static unsigned int monitor_reset = 1; | |
84 | ||
85 | /* This controls whether or not the driver will initialise devices | |
86 | * with invalid MAC addresses stored in the EEPROM or flash. If true, | |
87 | * such devices will be initialised with a random locally-generated | |
88 | * MAC address. This allows for loading the sfc_mtd driver to | |
89 | * reprogram the flash, even if the flash contents (including the MAC | |
90 | * address) have previously been erased. | |
91 | */ | |
92 | static unsigned int allow_bad_hwaddr; | |
93 | ||
94 | /* Initial interrupt moderation settings. They can be modified after | |
95 | * module load with ethtool. | |
96 | * | |
97 | * The default for RX should strike a balance between increasing the | |
98 | * round-trip latency and reducing overhead. | |
99 | */ | |
100 | static unsigned int rx_irq_mod_usec = 60; | |
101 | ||
102 | /* Initial interrupt moderation settings. They can be modified after | |
103 | * module load with ethtool. | |
104 | * | |
105 | * This default is chosen to ensure that a 10G link does not go idle | |
106 | * while a TX queue is stopped after it has become full. A queue is | |
107 | * restarted when it drops below half full. The time this takes (assuming | |
108 | * worst case 3 descriptors per packet and 1024 descriptors) is | |
109 | * 512 / 3 * 1.2 = 205 usec. | |
110 | */ | |
111 | static unsigned int tx_irq_mod_usec = 150; | |
112 | ||
113 | /* This is the first interrupt mode to try out of: | |
114 | * 0 => MSI-X | |
115 | * 1 => MSI | |
116 | * 2 => legacy | |
117 | */ | |
118 | static unsigned int interrupt_mode; | |
119 | ||
120 | /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS), | |
121 | * i.e. the number of CPUs among which we may distribute simultaneous | |
122 | * interrupt handling. | |
123 | * | |
124 | * Cards without MSI-X will only target one CPU via legacy or MSI interrupt. | |
125 | * The default (0) means to assign an interrupt to each package (level II cache) | |
126 | */ | |
127 | static unsigned int rss_cpus; | |
128 | module_param(rss_cpus, uint, 0444); | |
129 | MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling"); | |
130 | ||
131 | /************************************************************************** | |
132 | * | |
133 | * Utility functions and prototypes | |
134 | * | |
135 | *************************************************************************/ | |
136 | static void efx_remove_channel(struct efx_channel *channel); | |
137 | static void efx_remove_port(struct efx_nic *efx); | |
138 | static void efx_fini_napi(struct efx_nic *efx); | |
139 | static void efx_fini_channels(struct efx_nic *efx); | |
140 | ||
141 | #define EFX_ASSERT_RESET_SERIALISED(efx) \ | |
142 | do { \ | |
143 | if ((efx->state == STATE_RUNNING) || \ | |
144 | (efx->state == STATE_RESETTING)) \ | |
145 | ASSERT_RTNL(); \ | |
146 | } while (0) | |
147 | ||
148 | /************************************************************************** | |
149 | * | |
150 | * Event queue processing | |
151 | * | |
152 | *************************************************************************/ | |
153 | ||
154 | /* Process channel's event queue | |
155 | * | |
156 | * This function is responsible for processing the event queue of a | |
157 | * single channel. The caller must guarantee that this function will | |
158 | * never be concurrently called more than once on the same channel, | |
159 | * though different channels may be being processed concurrently. | |
160 | */ | |
161 | static inline int efx_process_channel(struct efx_channel *channel, int rx_quota) | |
162 | { | |
163 | int rxdmaqs; | |
164 | struct efx_rx_queue *rx_queue; | |
165 | ||
166 | if (unlikely(channel->efx->reset_pending != RESET_TYPE_NONE || | |
167 | !channel->enabled)) | |
168 | return rx_quota; | |
169 | ||
170 | rxdmaqs = falcon_process_eventq(channel, &rx_quota); | |
171 | ||
172 | /* Deliver last RX packet. */ | |
173 | if (channel->rx_pkt) { | |
174 | __efx_rx_packet(channel, channel->rx_pkt, | |
175 | channel->rx_pkt_csummed); | |
176 | channel->rx_pkt = NULL; | |
177 | } | |
178 | ||
179 | efx_flush_lro(channel); | |
180 | efx_rx_strategy(channel); | |
181 | ||
182 | /* Refill descriptor rings as necessary */ | |
183 | rx_queue = &channel->efx->rx_queue[0]; | |
184 | while (rxdmaqs) { | |
185 | if (rxdmaqs & 0x01) | |
186 | efx_fast_push_rx_descriptors(rx_queue); | |
187 | rx_queue++; | |
188 | rxdmaqs >>= 1; | |
189 | } | |
190 | ||
191 | return rx_quota; | |
192 | } | |
193 | ||
194 | /* Mark channel as finished processing | |
195 | * | |
196 | * Note that since we will not receive further interrupts for this | |
197 | * channel before we finish processing and call the eventq_read_ack() | |
198 | * method, there is no need to use the interrupt hold-off timers. | |
199 | */ | |
200 | static inline void efx_channel_processed(struct efx_channel *channel) | |
201 | { | |
202 | /* Write to EVQ_RPTR_REG. If a new event arrived in a race | |
203 | * with finishing processing, a new interrupt will be raised. | |
204 | */ | |
205 | channel->work_pending = 0; | |
206 | smp_wmb(); /* Ensure channel updated before any new interrupt. */ | |
207 | falcon_eventq_read_ack(channel); | |
208 | } | |
209 | ||
210 | /* NAPI poll handler | |
211 | * | |
212 | * NAPI guarantees serialisation of polls of the same device, which | |
213 | * provides the guarantee required by efx_process_channel(). | |
214 | */ | |
215 | static int efx_poll(struct napi_struct *napi, int budget) | |
216 | { | |
217 | struct efx_channel *channel = | |
218 | container_of(napi, struct efx_channel, napi_str); | |
219 | struct net_device *napi_dev = channel->napi_dev; | |
220 | int unused; | |
221 | int rx_packets; | |
222 | ||
223 | EFX_TRACE(channel->efx, "channel %d NAPI poll executing on CPU %d\n", | |
224 | channel->channel, raw_smp_processor_id()); | |
225 | ||
226 | unused = efx_process_channel(channel, budget); | |
227 | rx_packets = (budget - unused); | |
228 | ||
229 | if (rx_packets < budget) { | |
230 | /* There is no race here; although napi_disable() will | |
231 | * only wait for netif_rx_complete(), this isn't a problem | |
232 | * since efx_channel_processed() will have no effect if | |
233 | * interrupts have already been disabled. | |
234 | */ | |
235 | netif_rx_complete(napi_dev, napi); | |
236 | efx_channel_processed(channel); | |
237 | } | |
238 | ||
239 | return rx_packets; | |
240 | } | |
241 | ||
242 | /* Process the eventq of the specified channel immediately on this CPU | |
243 | * | |
244 | * Disable hardware generated interrupts, wait for any existing | |
245 | * processing to finish, then directly poll (and ack ) the eventq. | |
246 | * Finally reenable NAPI and interrupts. | |
247 | * | |
248 | * Since we are touching interrupts the caller should hold the suspend lock | |
249 | */ | |
250 | void efx_process_channel_now(struct efx_channel *channel) | |
251 | { | |
252 | struct efx_nic *efx = channel->efx; | |
253 | ||
254 | BUG_ON(!channel->used_flags); | |
255 | BUG_ON(!channel->enabled); | |
256 | ||
257 | /* Disable interrupts and wait for ISRs to complete */ | |
258 | falcon_disable_interrupts(efx); | |
259 | if (efx->legacy_irq) | |
260 | synchronize_irq(efx->legacy_irq); | |
261 | if (channel->has_interrupt && channel->irq) | |
262 | synchronize_irq(channel->irq); | |
263 | ||
264 | /* Wait for any NAPI processing to complete */ | |
265 | napi_disable(&channel->napi_str); | |
266 | ||
267 | /* Poll the channel */ | |
268 | (void) efx_process_channel(channel, efx->type->evq_size); | |
269 | ||
270 | /* Ack the eventq. This may cause an interrupt to be generated | |
271 | * when they are reenabled */ | |
272 | efx_channel_processed(channel); | |
273 | ||
274 | napi_enable(&channel->napi_str); | |
275 | falcon_enable_interrupts(efx); | |
276 | } | |
277 | ||
278 | /* Create event queue | |
279 | * Event queue memory allocations are done only once. If the channel | |
280 | * is reset, the memory buffer will be reused; this guards against | |
281 | * errors during channel reset and also simplifies interrupt handling. | |
282 | */ | |
283 | static int efx_probe_eventq(struct efx_channel *channel) | |
284 | { | |
285 | EFX_LOG(channel->efx, "chan %d create event queue\n", channel->channel); | |
286 | ||
287 | return falcon_probe_eventq(channel); | |
288 | } | |
289 | ||
290 | /* Prepare channel's event queue */ | |
291 | static int efx_init_eventq(struct efx_channel *channel) | |
292 | { | |
293 | EFX_LOG(channel->efx, "chan %d init event queue\n", channel->channel); | |
294 | ||
295 | channel->eventq_read_ptr = 0; | |
296 | ||
297 | return falcon_init_eventq(channel); | |
298 | } | |
299 | ||
300 | static void efx_fini_eventq(struct efx_channel *channel) | |
301 | { | |
302 | EFX_LOG(channel->efx, "chan %d fini event queue\n", channel->channel); | |
303 | ||
304 | falcon_fini_eventq(channel); | |
305 | } | |
306 | ||
307 | static void efx_remove_eventq(struct efx_channel *channel) | |
308 | { | |
309 | EFX_LOG(channel->efx, "chan %d remove event queue\n", channel->channel); | |
310 | ||
311 | falcon_remove_eventq(channel); | |
312 | } | |
313 | ||
314 | /************************************************************************** | |
315 | * | |
316 | * Channel handling | |
317 | * | |
318 | *************************************************************************/ | |
319 | ||
320 | /* Setup per-NIC RX buffer parameters. | |
321 | * Calculate the rx buffer allocation parameters required to support | |
322 | * the current MTU, including padding for header alignment and overruns. | |
323 | */ | |
324 | static void efx_calc_rx_buffer_params(struct efx_nic *efx) | |
325 | { | |
326 | unsigned int order, len; | |
327 | ||
328 | len = (max(EFX_PAGE_IP_ALIGN, NET_IP_ALIGN) + | |
329 | EFX_MAX_FRAME_LEN(efx->net_dev->mtu) + | |
330 | efx->type->rx_buffer_padding); | |
331 | ||
332 | /* Calculate page-order */ | |
333 | for (order = 0; ((1u << order) * PAGE_SIZE) < len; ++order) | |
334 | ; | |
335 | ||
336 | efx->rx_buffer_len = len; | |
337 | efx->rx_buffer_order = order; | |
338 | } | |
339 | ||
340 | static int efx_probe_channel(struct efx_channel *channel) | |
341 | { | |
342 | struct efx_tx_queue *tx_queue; | |
343 | struct efx_rx_queue *rx_queue; | |
344 | int rc; | |
345 | ||
346 | EFX_LOG(channel->efx, "creating channel %d\n", channel->channel); | |
347 | ||
348 | rc = efx_probe_eventq(channel); | |
349 | if (rc) | |
350 | goto fail1; | |
351 | ||
352 | efx_for_each_channel_tx_queue(tx_queue, channel) { | |
353 | rc = efx_probe_tx_queue(tx_queue); | |
354 | if (rc) | |
355 | goto fail2; | |
356 | } | |
357 | ||
358 | efx_for_each_channel_rx_queue(rx_queue, channel) { | |
359 | rc = efx_probe_rx_queue(rx_queue); | |
360 | if (rc) | |
361 | goto fail3; | |
362 | } | |
363 | ||
364 | channel->n_rx_frm_trunc = 0; | |
365 | ||
366 | return 0; | |
367 | ||
368 | fail3: | |
369 | efx_for_each_channel_rx_queue(rx_queue, channel) | |
370 | efx_remove_rx_queue(rx_queue); | |
371 | fail2: | |
372 | efx_for_each_channel_tx_queue(tx_queue, channel) | |
373 | efx_remove_tx_queue(tx_queue); | |
374 | fail1: | |
375 | return rc; | |
376 | } | |
377 | ||
378 | ||
379 | /* Channels are shutdown and reinitialised whilst the NIC is running | |
380 | * to propagate configuration changes (mtu, checksum offload), or | |
381 | * to clear hardware error conditions | |
382 | */ | |
383 | static int efx_init_channels(struct efx_nic *efx) | |
384 | { | |
385 | struct efx_tx_queue *tx_queue; | |
386 | struct efx_rx_queue *rx_queue; | |
387 | struct efx_channel *channel; | |
388 | int rc = 0; | |
389 | ||
390 | efx_calc_rx_buffer_params(efx); | |
391 | ||
392 | /* Initialise the channels */ | |
393 | efx_for_each_channel(channel, efx) { | |
394 | EFX_LOG(channel->efx, "init chan %d\n", channel->channel); | |
395 | ||
396 | rc = efx_init_eventq(channel); | |
397 | if (rc) | |
398 | goto err; | |
399 | ||
400 | efx_for_each_channel_tx_queue(tx_queue, channel) { | |
401 | rc = efx_init_tx_queue(tx_queue); | |
402 | if (rc) | |
403 | goto err; | |
404 | } | |
405 | ||
406 | /* The rx buffer allocation strategy is MTU dependent */ | |
407 | efx_rx_strategy(channel); | |
408 | ||
409 | efx_for_each_channel_rx_queue(rx_queue, channel) { | |
410 | rc = efx_init_rx_queue(rx_queue); | |
411 | if (rc) | |
412 | goto err; | |
413 | } | |
414 | ||
415 | WARN_ON(channel->rx_pkt != NULL); | |
416 | efx_rx_strategy(channel); | |
417 | } | |
418 | ||
419 | return 0; | |
420 | ||
421 | err: | |
422 | EFX_ERR(efx, "failed to initialise channel %d\n", | |
423 | channel ? channel->channel : -1); | |
424 | efx_fini_channels(efx); | |
425 | return rc; | |
426 | } | |
427 | ||
428 | /* This enables event queue processing and packet transmission. | |
429 | * | |
430 | * Note that this function is not allowed to fail, since that would | |
431 | * introduce too much complexity into the suspend/resume path. | |
432 | */ | |
433 | static void efx_start_channel(struct efx_channel *channel) | |
434 | { | |
435 | struct efx_rx_queue *rx_queue; | |
436 | ||
437 | EFX_LOG(channel->efx, "starting chan %d\n", channel->channel); | |
438 | ||
439 | if (!(channel->efx->net_dev->flags & IFF_UP)) | |
440 | netif_napi_add(channel->napi_dev, &channel->napi_str, | |
441 | efx_poll, napi_weight); | |
442 | ||
443 | channel->work_pending = 0; | |
444 | channel->enabled = 1; | |
445 | smp_wmb(); /* ensure channel updated before first interrupt */ | |
446 | ||
447 | napi_enable(&channel->napi_str); | |
448 | ||
449 | /* Load up RX descriptors */ | |
450 | efx_for_each_channel_rx_queue(rx_queue, channel) | |
451 | efx_fast_push_rx_descriptors(rx_queue); | |
452 | } | |
453 | ||
454 | /* This disables event queue processing and packet transmission. | |
455 | * This function does not guarantee that all queue processing | |
456 | * (e.g. RX refill) is complete. | |
457 | */ | |
458 | static void efx_stop_channel(struct efx_channel *channel) | |
459 | { | |
460 | struct efx_rx_queue *rx_queue; | |
461 | ||
462 | if (!channel->enabled) | |
463 | return; | |
464 | ||
465 | EFX_LOG(channel->efx, "stop chan %d\n", channel->channel); | |
466 | ||
467 | channel->enabled = 0; | |
468 | napi_disable(&channel->napi_str); | |
469 | ||
470 | /* Ensure that any worker threads have exited or will be no-ops */ | |
471 | efx_for_each_channel_rx_queue(rx_queue, channel) { | |
472 | spin_lock_bh(&rx_queue->add_lock); | |
473 | spin_unlock_bh(&rx_queue->add_lock); | |
474 | } | |
475 | } | |
476 | ||
477 | static void efx_fini_channels(struct efx_nic *efx) | |
478 | { | |
479 | struct efx_channel *channel; | |
480 | struct efx_tx_queue *tx_queue; | |
481 | struct efx_rx_queue *rx_queue; | |
482 | ||
483 | EFX_ASSERT_RESET_SERIALISED(efx); | |
484 | BUG_ON(efx->port_enabled); | |
485 | ||
486 | efx_for_each_channel(channel, efx) { | |
487 | EFX_LOG(channel->efx, "shut down chan %d\n", channel->channel); | |
488 | ||
489 | efx_for_each_channel_rx_queue(rx_queue, channel) | |
490 | efx_fini_rx_queue(rx_queue); | |
491 | efx_for_each_channel_tx_queue(tx_queue, channel) | |
492 | efx_fini_tx_queue(tx_queue); | |
493 | } | |
494 | ||
495 | /* Do the event queues last so that we can handle flush events | |
496 | * for all DMA queues. */ | |
497 | efx_for_each_channel(channel, efx) { | |
498 | EFX_LOG(channel->efx, "shut down evq %d\n", channel->channel); | |
499 | ||
500 | efx_fini_eventq(channel); | |
501 | } | |
502 | } | |
503 | ||
504 | static void efx_remove_channel(struct efx_channel *channel) | |
505 | { | |
506 | struct efx_tx_queue *tx_queue; | |
507 | struct efx_rx_queue *rx_queue; | |
508 | ||
509 | EFX_LOG(channel->efx, "destroy chan %d\n", channel->channel); | |
510 | ||
511 | efx_for_each_channel_rx_queue(rx_queue, channel) | |
512 | efx_remove_rx_queue(rx_queue); | |
513 | efx_for_each_channel_tx_queue(tx_queue, channel) | |
514 | efx_remove_tx_queue(tx_queue); | |
515 | efx_remove_eventq(channel); | |
516 | ||
517 | channel->used_flags = 0; | |
518 | } | |
519 | ||
520 | void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue, int delay) | |
521 | { | |
522 | queue_delayed_work(refill_workqueue, &rx_queue->work, delay); | |
523 | } | |
524 | ||
525 | /************************************************************************** | |
526 | * | |
527 | * Port handling | |
528 | * | |
529 | **************************************************************************/ | |
530 | ||
531 | /* This ensures that the kernel is kept informed (via | |
532 | * netif_carrier_on/off) of the link status, and also maintains the | |
533 | * link status's stop on the port's TX queue. | |
534 | */ | |
535 | static void efx_link_status_changed(struct efx_nic *efx) | |
536 | { | |
537 | int carrier_ok; | |
538 | ||
539 | /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure | |
540 | * that no events are triggered between unregister_netdev() and the | |
541 | * driver unloading. A more general condition is that NETDEV_CHANGE | |
542 | * can only be generated between NETDEV_UP and NETDEV_DOWN */ | |
543 | if (!netif_running(efx->net_dev)) | |
544 | return; | |
545 | ||
546 | carrier_ok = netif_carrier_ok(efx->net_dev) ? 1 : 0; | |
547 | if (efx->link_up != carrier_ok) { | |
548 | efx->n_link_state_changes++; | |
549 | ||
550 | if (efx->link_up) | |
551 | netif_carrier_on(efx->net_dev); | |
552 | else | |
553 | netif_carrier_off(efx->net_dev); | |
554 | } | |
555 | ||
556 | /* Status message for kernel log */ | |
557 | if (efx->link_up) { | |
558 | struct mii_if_info *gmii = &efx->mii; | |
559 | unsigned adv, lpa; | |
560 | /* NONE here means direct XAUI from the controller, with no | |
561 | * MDIO-attached device we can query. */ | |
562 | if (efx->phy_type != PHY_TYPE_NONE) { | |
563 | adv = gmii_advertised(gmii); | |
564 | lpa = gmii_lpa(gmii); | |
565 | } else { | |
566 | lpa = GM_LPA_10000 | LPA_DUPLEX; | |
567 | adv = lpa; | |
568 | } | |
569 | EFX_INFO(efx, "link up at %dMbps %s-duplex " | |
570 | "(adv %04x lpa %04x) (MTU %d)%s\n", | |
571 | (efx->link_options & GM_LPA_10000 ? 10000 : | |
572 | (efx->link_options & GM_LPA_1000 ? 1000 : | |
573 | (efx->link_options & GM_LPA_100 ? 100 : | |
574 | 10))), | |
575 | (efx->link_options & GM_LPA_DUPLEX ? | |
576 | "full" : "half"), | |
577 | adv, lpa, | |
578 | efx->net_dev->mtu, | |
579 | (efx->promiscuous ? " [PROMISC]" : "")); | |
580 | } else { | |
581 | EFX_INFO(efx, "link down\n"); | |
582 | } | |
583 | ||
584 | } | |
585 | ||
586 | /* This call reinitialises the MAC to pick up new PHY settings. The | |
587 | * caller must hold the mac_lock */ | |
588 | static void __efx_reconfigure_port(struct efx_nic *efx) | |
589 | { | |
590 | WARN_ON(!mutex_is_locked(&efx->mac_lock)); | |
591 | ||
592 | EFX_LOG(efx, "reconfiguring MAC from PHY settings on CPU %d\n", | |
593 | raw_smp_processor_id()); | |
594 | ||
595 | falcon_reconfigure_xmac(efx); | |
596 | ||
597 | /* Inform kernel of loss/gain of carrier */ | |
598 | efx_link_status_changed(efx); | |
599 | } | |
600 | ||
601 | /* Reinitialise the MAC to pick up new PHY settings, even if the port is | |
602 | * disabled. */ | |
603 | void efx_reconfigure_port(struct efx_nic *efx) | |
604 | { | |
605 | EFX_ASSERT_RESET_SERIALISED(efx); | |
606 | ||
607 | mutex_lock(&efx->mac_lock); | |
608 | __efx_reconfigure_port(efx); | |
609 | mutex_unlock(&efx->mac_lock); | |
610 | } | |
611 | ||
612 | /* Asynchronous efx_reconfigure_port work item. To speed up efx_flush_all() | |
613 | * we don't efx_reconfigure_port() if the port is disabled. Care is taken | |
614 | * in efx_stop_all() and efx_start_port() to prevent PHY events being lost */ | |
615 | static void efx_reconfigure_work(struct work_struct *data) | |
616 | { | |
617 | struct efx_nic *efx = container_of(data, struct efx_nic, | |
618 | reconfigure_work); | |
619 | ||
620 | mutex_lock(&efx->mac_lock); | |
621 | if (efx->port_enabled) | |
622 | __efx_reconfigure_port(efx); | |
623 | mutex_unlock(&efx->mac_lock); | |
624 | } | |
625 | ||
626 | static int efx_probe_port(struct efx_nic *efx) | |
627 | { | |
628 | int rc; | |
629 | ||
630 | EFX_LOG(efx, "create port\n"); | |
631 | ||
632 | /* Connect up MAC/PHY operations table and read MAC address */ | |
633 | rc = falcon_probe_port(efx); | |
634 | if (rc) | |
635 | goto err; | |
636 | ||
637 | /* Sanity check MAC address */ | |
638 | if (is_valid_ether_addr(efx->mac_address)) { | |
639 | memcpy(efx->net_dev->dev_addr, efx->mac_address, ETH_ALEN); | |
640 | } else { | |
641 | DECLARE_MAC_BUF(mac); | |
642 | ||
643 | EFX_ERR(efx, "invalid MAC address %s\n", | |
644 | print_mac(mac, efx->mac_address)); | |
645 | if (!allow_bad_hwaddr) { | |
646 | rc = -EINVAL; | |
647 | goto err; | |
648 | } | |
649 | random_ether_addr(efx->net_dev->dev_addr); | |
650 | EFX_INFO(efx, "using locally-generated MAC %s\n", | |
651 | print_mac(mac, efx->net_dev->dev_addr)); | |
652 | } | |
653 | ||
654 | return 0; | |
655 | ||
656 | err: | |
657 | efx_remove_port(efx); | |
658 | return rc; | |
659 | } | |
660 | ||
661 | static int efx_init_port(struct efx_nic *efx) | |
662 | { | |
663 | int rc; | |
664 | ||
665 | EFX_LOG(efx, "init port\n"); | |
666 | ||
667 | /* Initialise the MAC and PHY */ | |
668 | rc = falcon_init_xmac(efx); | |
669 | if (rc) | |
670 | return rc; | |
671 | ||
672 | efx->port_initialized = 1; | |
673 | ||
674 | /* Reconfigure port to program MAC registers */ | |
675 | falcon_reconfigure_xmac(efx); | |
676 | ||
677 | return 0; | |
678 | } | |
679 | ||
680 | /* Allow efx_reconfigure_port() to be scheduled, and close the window | |
681 | * between efx_stop_port and efx_flush_all whereby a previously scheduled | |
682 | * efx_reconfigure_port() may have been cancelled */ | |
683 | static void efx_start_port(struct efx_nic *efx) | |
684 | { | |
685 | EFX_LOG(efx, "start port\n"); | |
686 | BUG_ON(efx->port_enabled); | |
687 | ||
688 | mutex_lock(&efx->mac_lock); | |
689 | efx->port_enabled = 1; | |
690 | __efx_reconfigure_port(efx); | |
691 | mutex_unlock(&efx->mac_lock); | |
692 | } | |
693 | ||
694 | /* Prevent efx_reconfigure_work and efx_monitor() from executing, and | |
695 | * efx_set_multicast_list() from scheduling efx_reconfigure_work. | |
696 | * efx_reconfigure_work can still be scheduled via NAPI processing | |
697 | * until efx_flush_all() is called */ | |
698 | static void efx_stop_port(struct efx_nic *efx) | |
699 | { | |
700 | EFX_LOG(efx, "stop port\n"); | |
701 | ||
702 | mutex_lock(&efx->mac_lock); | |
703 | efx->port_enabled = 0; | |
704 | mutex_unlock(&efx->mac_lock); | |
705 | ||
706 | /* Serialise against efx_set_multicast_list() */ | |
707 | if (NET_DEV_REGISTERED(efx)) { | |
708 | netif_tx_lock_bh(efx->net_dev); | |
709 | netif_tx_unlock_bh(efx->net_dev); | |
710 | } | |
711 | } | |
712 | ||
713 | static void efx_fini_port(struct efx_nic *efx) | |
714 | { | |
715 | EFX_LOG(efx, "shut down port\n"); | |
716 | ||
717 | if (!efx->port_initialized) | |
718 | return; | |
719 | ||
720 | falcon_fini_xmac(efx); | |
721 | efx->port_initialized = 0; | |
722 | ||
723 | efx->link_up = 0; | |
724 | efx_link_status_changed(efx); | |
725 | } | |
726 | ||
727 | static void efx_remove_port(struct efx_nic *efx) | |
728 | { | |
729 | EFX_LOG(efx, "destroying port\n"); | |
730 | ||
731 | falcon_remove_port(efx); | |
732 | } | |
733 | ||
734 | /************************************************************************** | |
735 | * | |
736 | * NIC handling | |
737 | * | |
738 | **************************************************************************/ | |
739 | ||
740 | /* This configures the PCI device to enable I/O and DMA. */ | |
741 | static int efx_init_io(struct efx_nic *efx) | |
742 | { | |
743 | struct pci_dev *pci_dev = efx->pci_dev; | |
744 | dma_addr_t dma_mask = efx->type->max_dma_mask; | |
745 | int rc; | |
746 | ||
747 | EFX_LOG(efx, "initialising I/O\n"); | |
748 | ||
749 | rc = pci_enable_device(pci_dev); | |
750 | if (rc) { | |
751 | EFX_ERR(efx, "failed to enable PCI device\n"); | |
752 | goto fail1; | |
753 | } | |
754 | ||
755 | pci_set_master(pci_dev); | |
756 | ||
757 | /* Set the PCI DMA mask. Try all possibilities from our | |
758 | * genuine mask down to 32 bits, because some architectures | |
759 | * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit | |
760 | * masks event though they reject 46 bit masks. | |
761 | */ | |
762 | while (dma_mask > 0x7fffffffUL) { | |
763 | if (pci_dma_supported(pci_dev, dma_mask) && | |
764 | ((rc = pci_set_dma_mask(pci_dev, dma_mask)) == 0)) | |
765 | break; | |
766 | dma_mask >>= 1; | |
767 | } | |
768 | if (rc) { | |
769 | EFX_ERR(efx, "could not find a suitable DMA mask\n"); | |
770 | goto fail2; | |
771 | } | |
772 | EFX_LOG(efx, "using DMA mask %llx\n", (unsigned long long) dma_mask); | |
773 | rc = pci_set_consistent_dma_mask(pci_dev, dma_mask); | |
774 | if (rc) { | |
775 | /* pci_set_consistent_dma_mask() is not *allowed* to | |
776 | * fail with a mask that pci_set_dma_mask() accepted, | |
777 | * but just in case... | |
778 | */ | |
779 | EFX_ERR(efx, "failed to set consistent DMA mask\n"); | |
780 | goto fail2; | |
781 | } | |
782 | ||
783 | efx->membase_phys = pci_resource_start(efx->pci_dev, | |
784 | efx->type->mem_bar); | |
785 | rc = pci_request_region(pci_dev, efx->type->mem_bar, "sfc"); | |
786 | if (rc) { | |
787 | EFX_ERR(efx, "request for memory BAR failed\n"); | |
788 | rc = -EIO; | |
789 | goto fail3; | |
790 | } | |
791 | efx->membase = ioremap_nocache(efx->membase_phys, | |
792 | efx->type->mem_map_size); | |
793 | if (!efx->membase) { | |
794 | EFX_ERR(efx, "could not map memory BAR %d at %lx+%x\n", | |
795 | efx->type->mem_bar, efx->membase_phys, | |
796 | efx->type->mem_map_size); | |
797 | rc = -ENOMEM; | |
798 | goto fail4; | |
799 | } | |
800 | EFX_LOG(efx, "memory BAR %u at %lx+%x (virtual %p)\n", | |
801 | efx->type->mem_bar, efx->membase_phys, efx->type->mem_map_size, | |
802 | efx->membase); | |
803 | ||
804 | return 0; | |
805 | ||
806 | fail4: | |
807 | release_mem_region(efx->membase_phys, efx->type->mem_map_size); | |
808 | fail3: | |
809 | efx->membase_phys = 0UL; | |
810 | fail2: | |
811 | pci_disable_device(efx->pci_dev); | |
812 | fail1: | |
813 | return rc; | |
814 | } | |
815 | ||
816 | static void efx_fini_io(struct efx_nic *efx) | |
817 | { | |
818 | EFX_LOG(efx, "shutting down I/O\n"); | |
819 | ||
820 | if (efx->membase) { | |
821 | iounmap(efx->membase); | |
822 | efx->membase = NULL; | |
823 | } | |
824 | ||
825 | if (efx->membase_phys) { | |
826 | pci_release_region(efx->pci_dev, efx->type->mem_bar); | |
827 | efx->membase_phys = 0UL; | |
828 | } | |
829 | ||
830 | pci_disable_device(efx->pci_dev); | |
831 | } | |
832 | ||
833 | /* Probe the number and type of interrupts we are able to obtain. */ | |
834 | static void efx_probe_interrupts(struct efx_nic *efx) | |
835 | { | |
836 | int max_channel = efx->type->phys_addr_channels - 1; | |
837 | struct msix_entry xentries[EFX_MAX_CHANNELS]; | |
838 | int rc, i; | |
839 | ||
840 | if (efx->interrupt_mode == EFX_INT_MODE_MSIX) { | |
841 | BUG_ON(!pci_find_capability(efx->pci_dev, PCI_CAP_ID_MSIX)); | |
842 | ||
843 | efx->rss_queues = rss_cpus ? rss_cpus : num_online_cpus(); | |
844 | efx->rss_queues = min(efx->rss_queues, max_channel + 1); | |
845 | efx->rss_queues = min(efx->rss_queues, EFX_MAX_CHANNELS); | |
846 | ||
847 | /* Request maximum number of MSI interrupts, and fill out | |
848 | * the channel interrupt information the allowed allocation */ | |
849 | for (i = 0; i < efx->rss_queues; i++) | |
850 | xentries[i].entry = i; | |
851 | rc = pci_enable_msix(efx->pci_dev, xentries, efx->rss_queues); | |
852 | if (rc > 0) { | |
853 | EFX_BUG_ON_PARANOID(rc >= efx->rss_queues); | |
854 | efx->rss_queues = rc; | |
855 | rc = pci_enable_msix(efx->pci_dev, xentries, | |
856 | efx->rss_queues); | |
857 | } | |
858 | ||
859 | if (rc == 0) { | |
860 | for (i = 0; i < efx->rss_queues; i++) { | |
861 | efx->channel[i].has_interrupt = 1; | |
862 | efx->channel[i].irq = xentries[i].vector; | |
863 | } | |
864 | } else { | |
865 | /* Fall back to single channel MSI */ | |
866 | efx->interrupt_mode = EFX_INT_MODE_MSI; | |
867 | EFX_ERR(efx, "could not enable MSI-X\n"); | |
868 | } | |
869 | } | |
870 | ||
871 | /* Try single interrupt MSI */ | |
872 | if (efx->interrupt_mode == EFX_INT_MODE_MSI) { | |
873 | efx->rss_queues = 1; | |
874 | rc = pci_enable_msi(efx->pci_dev); | |
875 | if (rc == 0) { | |
876 | efx->channel[0].irq = efx->pci_dev->irq; | |
877 | efx->channel[0].has_interrupt = 1; | |
878 | } else { | |
879 | EFX_ERR(efx, "could not enable MSI\n"); | |
880 | efx->interrupt_mode = EFX_INT_MODE_LEGACY; | |
881 | } | |
882 | } | |
883 | ||
884 | /* Assume legacy interrupts */ | |
885 | if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) { | |
886 | efx->rss_queues = 1; | |
887 | /* Every channel is interruptible */ | |
888 | for (i = 0; i < EFX_MAX_CHANNELS; i++) | |
889 | efx->channel[i].has_interrupt = 1; | |
890 | efx->legacy_irq = efx->pci_dev->irq; | |
891 | } | |
892 | } | |
893 | ||
894 | static void efx_remove_interrupts(struct efx_nic *efx) | |
895 | { | |
896 | struct efx_channel *channel; | |
897 | ||
898 | /* Remove MSI/MSI-X interrupts */ | |
899 | efx_for_each_channel_with_interrupt(channel, efx) | |
900 | channel->irq = 0; | |
901 | pci_disable_msi(efx->pci_dev); | |
902 | pci_disable_msix(efx->pci_dev); | |
903 | ||
904 | /* Remove legacy interrupt */ | |
905 | efx->legacy_irq = 0; | |
906 | } | |
907 | ||
908 | /* Select number of used resources | |
909 | * Should be called after probe_interrupts() | |
910 | */ | |
911 | static void efx_select_used(struct efx_nic *efx) | |
912 | { | |
913 | struct efx_tx_queue *tx_queue; | |
914 | struct efx_rx_queue *rx_queue; | |
915 | int i; | |
916 | ||
917 | /* TX queues. One per port per channel with TX capability | |
918 | * (more than one per port won't work on Linux, due to out | |
919 | * of order issues... but will be fine on Solaris) | |
920 | */ | |
921 | tx_queue = &efx->tx_queue[0]; | |
922 | ||
923 | /* Perform this for each channel with TX capabilities. | |
924 | * At the moment, we only support a single TX queue | |
925 | */ | |
926 | tx_queue->used = 1; | |
927 | if ((!EFX_INT_MODE_USE_MSI(efx)) && separate_tx_and_rx_channels) | |
928 | tx_queue->channel = &efx->channel[1]; | |
929 | else | |
930 | tx_queue->channel = &efx->channel[0]; | |
931 | tx_queue->channel->used_flags |= EFX_USED_BY_TX; | |
932 | tx_queue++; | |
933 | ||
934 | /* RX queues. Each has a dedicated channel. */ | |
935 | for (i = 0; i < EFX_MAX_RX_QUEUES; i++) { | |
936 | rx_queue = &efx->rx_queue[i]; | |
937 | ||
938 | if (i < efx->rss_queues) { | |
939 | rx_queue->used = 1; | |
940 | /* If we allow multiple RX queues per channel | |
941 | * we need to decide that here | |
942 | */ | |
943 | rx_queue->channel = &efx->channel[rx_queue->queue]; | |
944 | rx_queue->channel->used_flags |= EFX_USED_BY_RX; | |
945 | rx_queue++; | |
946 | } | |
947 | } | |
948 | } | |
949 | ||
950 | static int efx_probe_nic(struct efx_nic *efx) | |
951 | { | |
952 | int rc; | |
953 | ||
954 | EFX_LOG(efx, "creating NIC\n"); | |
955 | ||
956 | /* Carry out hardware-type specific initialisation */ | |
957 | rc = falcon_probe_nic(efx); | |
958 | if (rc) | |
959 | return rc; | |
960 | ||
961 | /* Determine the number of channels and RX queues by trying to hook | |
962 | * in MSI-X interrupts. */ | |
963 | efx_probe_interrupts(efx); | |
964 | ||
965 | /* Determine number of RX queues and TX queues */ | |
966 | efx_select_used(efx); | |
967 | ||
968 | /* Initialise the interrupt moderation settings */ | |
969 | efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec); | |
970 | ||
971 | return 0; | |
972 | } | |
973 | ||
974 | static void efx_remove_nic(struct efx_nic *efx) | |
975 | { | |
976 | EFX_LOG(efx, "destroying NIC\n"); | |
977 | ||
978 | efx_remove_interrupts(efx); | |
979 | falcon_remove_nic(efx); | |
980 | } | |
981 | ||
982 | /************************************************************************** | |
983 | * | |
984 | * NIC startup/shutdown | |
985 | * | |
986 | *************************************************************************/ | |
987 | ||
988 | static int efx_probe_all(struct efx_nic *efx) | |
989 | { | |
990 | struct efx_channel *channel; | |
991 | int rc; | |
992 | ||
993 | /* Create NIC */ | |
994 | rc = efx_probe_nic(efx); | |
995 | if (rc) { | |
996 | EFX_ERR(efx, "failed to create NIC\n"); | |
997 | goto fail1; | |
998 | } | |
999 | ||
1000 | /* Create port */ | |
1001 | rc = efx_probe_port(efx); | |
1002 | if (rc) { | |
1003 | EFX_ERR(efx, "failed to create port\n"); | |
1004 | goto fail2; | |
1005 | } | |
1006 | ||
1007 | /* Create channels */ | |
1008 | efx_for_each_channel(channel, efx) { | |
1009 | rc = efx_probe_channel(channel); | |
1010 | if (rc) { | |
1011 | EFX_ERR(efx, "failed to create channel %d\n", | |
1012 | channel->channel); | |
1013 | goto fail3; | |
1014 | } | |
1015 | } | |
1016 | ||
1017 | return 0; | |
1018 | ||
1019 | fail3: | |
1020 | efx_for_each_channel(channel, efx) | |
1021 | efx_remove_channel(channel); | |
1022 | efx_remove_port(efx); | |
1023 | fail2: | |
1024 | efx_remove_nic(efx); | |
1025 | fail1: | |
1026 | return rc; | |
1027 | } | |
1028 | ||
1029 | /* Called after previous invocation(s) of efx_stop_all, restarts the | |
1030 | * port, kernel transmit queue, NAPI processing and hardware interrupts, | |
1031 | * and ensures that the port is scheduled to be reconfigured. | |
1032 | * This function is safe to call multiple times when the NIC is in any | |
1033 | * state. */ | |
1034 | static void efx_start_all(struct efx_nic *efx) | |
1035 | { | |
1036 | struct efx_channel *channel; | |
1037 | ||
1038 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1039 | ||
1040 | /* Check that it is appropriate to restart the interface. All | |
1041 | * of these flags are safe to read under just the rtnl lock */ | |
1042 | if (efx->port_enabled) | |
1043 | return; | |
1044 | if ((efx->state != STATE_RUNNING) && (efx->state != STATE_INIT)) | |
1045 | return; | |
1046 | if (NET_DEV_REGISTERED(efx) && !netif_running(efx->net_dev)) | |
1047 | return; | |
1048 | ||
1049 | /* Mark the port as enabled so port reconfigurations can start, then | |
1050 | * restart the transmit interface early so the watchdog timer stops */ | |
1051 | efx_start_port(efx); | |
1052 | efx_wake_queue(efx); | |
1053 | ||
1054 | efx_for_each_channel(channel, efx) | |
1055 | efx_start_channel(channel); | |
1056 | ||
1057 | falcon_enable_interrupts(efx); | |
1058 | ||
1059 | /* Start hardware monitor if we're in RUNNING */ | |
1060 | if (efx->state == STATE_RUNNING) | |
1061 | queue_delayed_work(efx->workqueue, &efx->monitor_work, | |
1062 | efx_monitor_interval); | |
1063 | } | |
1064 | ||
1065 | /* Flush all delayed work. Should only be called when no more delayed work | |
1066 | * will be scheduled. This doesn't flush pending online resets (efx_reset), | |
1067 | * since we're holding the rtnl_lock at this point. */ | |
1068 | static void efx_flush_all(struct efx_nic *efx) | |
1069 | { | |
1070 | struct efx_rx_queue *rx_queue; | |
1071 | ||
1072 | /* Make sure the hardware monitor is stopped */ | |
1073 | cancel_delayed_work_sync(&efx->monitor_work); | |
1074 | ||
1075 | /* Ensure that all RX slow refills are complete. */ | |
1076 | efx_for_each_rx_queue(rx_queue, efx) { | |
1077 | cancel_delayed_work_sync(&rx_queue->work); | |
1078 | } | |
1079 | ||
1080 | /* Stop scheduled port reconfigurations */ | |
1081 | cancel_work_sync(&efx->reconfigure_work); | |
1082 | ||
1083 | } | |
1084 | ||
1085 | /* Quiesce hardware and software without bringing the link down. | |
1086 | * Safe to call multiple times, when the nic and interface is in any | |
1087 | * state. The caller is guaranteed to subsequently be in a position | |
1088 | * to modify any hardware and software state they see fit without | |
1089 | * taking locks. */ | |
1090 | static void efx_stop_all(struct efx_nic *efx) | |
1091 | { | |
1092 | struct efx_channel *channel; | |
1093 | ||
1094 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1095 | ||
1096 | /* port_enabled can be read safely under the rtnl lock */ | |
1097 | if (!efx->port_enabled) | |
1098 | return; | |
1099 | ||
1100 | /* Disable interrupts and wait for ISR to complete */ | |
1101 | falcon_disable_interrupts(efx); | |
1102 | if (efx->legacy_irq) | |
1103 | synchronize_irq(efx->legacy_irq); | |
1104 | efx_for_each_channel_with_interrupt(channel, efx) | |
1105 | if (channel->irq) | |
1106 | synchronize_irq(channel->irq); | |
1107 | ||
1108 | /* Stop all NAPI processing and synchronous rx refills */ | |
1109 | efx_for_each_channel(channel, efx) | |
1110 | efx_stop_channel(channel); | |
1111 | ||
1112 | /* Stop all asynchronous port reconfigurations. Since all | |
1113 | * event processing has already been stopped, there is no | |
1114 | * window to loose phy events */ | |
1115 | efx_stop_port(efx); | |
1116 | ||
1117 | /* Flush reconfigure_work, refill_workqueue, monitor_work */ | |
1118 | efx_flush_all(efx); | |
1119 | ||
1120 | /* Isolate the MAC from the TX and RX engines, so that queue | |
1121 | * flushes will complete in a timely fashion. */ | |
1122 | falcon_deconfigure_mac_wrapper(efx); | |
1123 | falcon_drain_tx_fifo(efx); | |
1124 | ||
1125 | /* Stop the kernel transmit interface late, so the watchdog | |
1126 | * timer isn't ticking over the flush */ | |
1127 | efx_stop_queue(efx); | |
1128 | if (NET_DEV_REGISTERED(efx)) { | |
1129 | netif_tx_lock_bh(efx->net_dev); | |
1130 | netif_tx_unlock_bh(efx->net_dev); | |
1131 | } | |
1132 | } | |
1133 | ||
1134 | static void efx_remove_all(struct efx_nic *efx) | |
1135 | { | |
1136 | struct efx_channel *channel; | |
1137 | ||
1138 | efx_for_each_channel(channel, efx) | |
1139 | efx_remove_channel(channel); | |
1140 | efx_remove_port(efx); | |
1141 | efx_remove_nic(efx); | |
1142 | } | |
1143 | ||
1144 | /* A convinience function to safely flush all the queues */ | |
1145 | int efx_flush_queues(struct efx_nic *efx) | |
1146 | { | |
1147 | int rc; | |
1148 | ||
1149 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1150 | ||
1151 | efx_stop_all(efx); | |
1152 | ||
1153 | efx_fini_channels(efx); | |
1154 | rc = efx_init_channels(efx); | |
1155 | if (rc) { | |
1156 | efx_schedule_reset(efx, RESET_TYPE_DISABLE); | |
1157 | return rc; | |
1158 | } | |
1159 | ||
1160 | efx_start_all(efx); | |
1161 | ||
1162 | return 0; | |
1163 | } | |
1164 | ||
1165 | /************************************************************************** | |
1166 | * | |
1167 | * Interrupt moderation | |
1168 | * | |
1169 | **************************************************************************/ | |
1170 | ||
1171 | /* Set interrupt moderation parameters */ | |
1172 | void efx_init_irq_moderation(struct efx_nic *efx, int tx_usecs, int rx_usecs) | |
1173 | { | |
1174 | struct efx_tx_queue *tx_queue; | |
1175 | struct efx_rx_queue *rx_queue; | |
1176 | ||
1177 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1178 | ||
1179 | efx_for_each_tx_queue(tx_queue, efx) | |
1180 | tx_queue->channel->irq_moderation = tx_usecs; | |
1181 | ||
1182 | efx_for_each_rx_queue(rx_queue, efx) | |
1183 | rx_queue->channel->irq_moderation = rx_usecs; | |
1184 | } | |
1185 | ||
1186 | /************************************************************************** | |
1187 | * | |
1188 | * Hardware monitor | |
1189 | * | |
1190 | **************************************************************************/ | |
1191 | ||
1192 | /* Run periodically off the general workqueue. Serialised against | |
1193 | * efx_reconfigure_port via the mac_lock */ | |
1194 | static void efx_monitor(struct work_struct *data) | |
1195 | { | |
1196 | struct efx_nic *efx = container_of(data, struct efx_nic, | |
1197 | monitor_work.work); | |
1198 | int rc = 0; | |
1199 | ||
1200 | EFX_TRACE(efx, "hardware monitor executing on CPU %d\n", | |
1201 | raw_smp_processor_id()); | |
1202 | ||
1203 | ||
1204 | /* If the mac_lock is already held then it is likely a port | |
1205 | * reconfiguration is already in place, which will likely do | |
1206 | * most of the work of check_hw() anyway. */ | |
1207 | if (!mutex_trylock(&efx->mac_lock)) { | |
1208 | queue_delayed_work(efx->workqueue, &efx->monitor_work, | |
1209 | efx_monitor_interval); | |
1210 | return; | |
1211 | } | |
1212 | ||
1213 | if (efx->port_enabled) | |
1214 | rc = falcon_check_xmac(efx); | |
1215 | mutex_unlock(&efx->mac_lock); | |
1216 | ||
1217 | if (rc) { | |
1218 | if (monitor_reset) { | |
1219 | EFX_ERR(efx, "hardware monitor detected a fault: " | |
1220 | "triggering reset\n"); | |
1221 | efx_schedule_reset(efx, RESET_TYPE_MONITOR); | |
1222 | } else { | |
1223 | EFX_ERR(efx, "hardware monitor detected a fault, " | |
1224 | "skipping reset\n"); | |
1225 | } | |
1226 | } | |
1227 | ||
1228 | queue_delayed_work(efx->workqueue, &efx->monitor_work, | |
1229 | efx_monitor_interval); | |
1230 | } | |
1231 | ||
1232 | /************************************************************************** | |
1233 | * | |
1234 | * ioctls | |
1235 | * | |
1236 | *************************************************************************/ | |
1237 | ||
1238 | /* Net device ioctl | |
1239 | * Context: process, rtnl_lock() held. | |
1240 | */ | |
1241 | static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd) | |
1242 | { | |
1243 | struct efx_nic *efx = net_dev->priv; | |
1244 | ||
1245 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1246 | ||
1247 | return generic_mii_ioctl(&efx->mii, if_mii(ifr), cmd, NULL); | |
1248 | } | |
1249 | ||
1250 | /************************************************************************** | |
1251 | * | |
1252 | * NAPI interface | |
1253 | * | |
1254 | **************************************************************************/ | |
1255 | ||
1256 | static int efx_init_napi(struct efx_nic *efx) | |
1257 | { | |
1258 | struct efx_channel *channel; | |
1259 | int rc; | |
1260 | ||
1261 | efx_for_each_channel(channel, efx) { | |
1262 | channel->napi_dev = efx->net_dev; | |
1263 | rc = efx_lro_init(&channel->lro_mgr, efx); | |
1264 | if (rc) | |
1265 | goto err; | |
1266 | } | |
1267 | return 0; | |
1268 | err: | |
1269 | efx_fini_napi(efx); | |
1270 | return rc; | |
1271 | } | |
1272 | ||
1273 | static void efx_fini_napi(struct efx_nic *efx) | |
1274 | { | |
1275 | struct efx_channel *channel; | |
1276 | ||
1277 | efx_for_each_channel(channel, efx) { | |
1278 | efx_lro_fini(&channel->lro_mgr); | |
1279 | channel->napi_dev = NULL; | |
1280 | } | |
1281 | } | |
1282 | ||
1283 | /************************************************************************** | |
1284 | * | |
1285 | * Kernel netpoll interface | |
1286 | * | |
1287 | *************************************************************************/ | |
1288 | ||
1289 | #ifdef CONFIG_NET_POLL_CONTROLLER | |
1290 | ||
1291 | /* Although in the common case interrupts will be disabled, this is not | |
1292 | * guaranteed. However, all our work happens inside the NAPI callback, | |
1293 | * so no locking is required. | |
1294 | */ | |
1295 | static void efx_netpoll(struct net_device *net_dev) | |
1296 | { | |
1297 | struct efx_nic *efx = net_dev->priv; | |
1298 | struct efx_channel *channel; | |
1299 | ||
1300 | efx_for_each_channel_with_interrupt(channel, efx) | |
1301 | efx_schedule_channel(channel); | |
1302 | } | |
1303 | ||
1304 | #endif | |
1305 | ||
1306 | /************************************************************************** | |
1307 | * | |
1308 | * Kernel net device interface | |
1309 | * | |
1310 | *************************************************************************/ | |
1311 | ||
1312 | /* Context: process, rtnl_lock() held. */ | |
1313 | static int efx_net_open(struct net_device *net_dev) | |
1314 | { | |
1315 | struct efx_nic *efx = net_dev->priv; | |
1316 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1317 | ||
1318 | EFX_LOG(efx, "opening device %s on CPU %d\n", net_dev->name, | |
1319 | raw_smp_processor_id()); | |
1320 | ||
1321 | efx_start_all(efx); | |
1322 | return 0; | |
1323 | } | |
1324 | ||
1325 | /* Context: process, rtnl_lock() held. | |
1326 | * Note that the kernel will ignore our return code; this method | |
1327 | * should really be a void. | |
1328 | */ | |
1329 | static int efx_net_stop(struct net_device *net_dev) | |
1330 | { | |
1331 | struct efx_nic *efx = net_dev->priv; | |
1332 | int rc; | |
1333 | ||
1334 | EFX_LOG(efx, "closing %s on CPU %d\n", net_dev->name, | |
1335 | raw_smp_processor_id()); | |
1336 | ||
1337 | /* Stop the device and flush all the channels */ | |
1338 | efx_stop_all(efx); | |
1339 | efx_fini_channels(efx); | |
1340 | rc = efx_init_channels(efx); | |
1341 | if (rc) | |
1342 | efx_schedule_reset(efx, RESET_TYPE_DISABLE); | |
1343 | ||
1344 | return 0; | |
1345 | } | |
1346 | ||
1347 | /* Context: process, dev_base_lock held, non-blocking. */ | |
1348 | static struct net_device_stats *efx_net_stats(struct net_device *net_dev) | |
1349 | { | |
1350 | struct efx_nic *efx = net_dev->priv; | |
1351 | struct efx_mac_stats *mac_stats = &efx->mac_stats; | |
1352 | struct net_device_stats *stats = &net_dev->stats; | |
1353 | ||
1354 | if (!spin_trylock(&efx->stats_lock)) | |
1355 | return stats; | |
1356 | if (efx->state == STATE_RUNNING) { | |
1357 | falcon_update_stats_xmac(efx); | |
1358 | falcon_update_nic_stats(efx); | |
1359 | } | |
1360 | spin_unlock(&efx->stats_lock); | |
1361 | ||
1362 | stats->rx_packets = mac_stats->rx_packets; | |
1363 | stats->tx_packets = mac_stats->tx_packets; | |
1364 | stats->rx_bytes = mac_stats->rx_bytes; | |
1365 | stats->tx_bytes = mac_stats->tx_bytes; | |
1366 | stats->multicast = mac_stats->rx_multicast; | |
1367 | stats->collisions = mac_stats->tx_collision; | |
1368 | stats->rx_length_errors = (mac_stats->rx_gtjumbo + | |
1369 | mac_stats->rx_length_error); | |
1370 | stats->rx_over_errors = efx->n_rx_nodesc_drop_cnt; | |
1371 | stats->rx_crc_errors = mac_stats->rx_bad; | |
1372 | stats->rx_frame_errors = mac_stats->rx_align_error; | |
1373 | stats->rx_fifo_errors = mac_stats->rx_overflow; | |
1374 | stats->rx_missed_errors = mac_stats->rx_missed; | |
1375 | stats->tx_window_errors = mac_stats->tx_late_collision; | |
1376 | ||
1377 | stats->rx_errors = (stats->rx_length_errors + | |
1378 | stats->rx_over_errors + | |
1379 | stats->rx_crc_errors + | |
1380 | stats->rx_frame_errors + | |
1381 | stats->rx_fifo_errors + | |
1382 | stats->rx_missed_errors + | |
1383 | mac_stats->rx_symbol_error); | |
1384 | stats->tx_errors = (stats->tx_window_errors + | |
1385 | mac_stats->tx_bad); | |
1386 | ||
1387 | return stats; | |
1388 | } | |
1389 | ||
1390 | /* Context: netif_tx_lock held, BHs disabled. */ | |
1391 | static void efx_watchdog(struct net_device *net_dev) | |
1392 | { | |
1393 | struct efx_nic *efx = net_dev->priv; | |
1394 | ||
1395 | EFX_ERR(efx, "TX stuck with stop_count=%d port_enabled=%d: %s\n", | |
1396 | atomic_read(&efx->netif_stop_count), efx->port_enabled, | |
1397 | monitor_reset ? "resetting channels" : "skipping reset"); | |
1398 | ||
1399 | if (monitor_reset) | |
1400 | efx_schedule_reset(efx, RESET_TYPE_MONITOR); | |
1401 | } | |
1402 | ||
1403 | ||
1404 | /* Context: process, rtnl_lock() held. */ | |
1405 | static int efx_change_mtu(struct net_device *net_dev, int new_mtu) | |
1406 | { | |
1407 | struct efx_nic *efx = net_dev->priv; | |
1408 | int rc = 0; | |
1409 | ||
1410 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1411 | ||
1412 | if (new_mtu > EFX_MAX_MTU) | |
1413 | return -EINVAL; | |
1414 | ||
1415 | efx_stop_all(efx); | |
1416 | ||
1417 | EFX_LOG(efx, "changing MTU to %d\n", new_mtu); | |
1418 | ||
1419 | efx_fini_channels(efx); | |
1420 | net_dev->mtu = new_mtu; | |
1421 | rc = efx_init_channels(efx); | |
1422 | if (rc) | |
1423 | goto fail; | |
1424 | ||
1425 | efx_start_all(efx); | |
1426 | return rc; | |
1427 | ||
1428 | fail: | |
1429 | efx_schedule_reset(efx, RESET_TYPE_DISABLE); | |
1430 | return rc; | |
1431 | } | |
1432 | ||
1433 | static int efx_set_mac_address(struct net_device *net_dev, void *data) | |
1434 | { | |
1435 | struct efx_nic *efx = net_dev->priv; | |
1436 | struct sockaddr *addr = data; | |
1437 | char *new_addr = addr->sa_data; | |
1438 | ||
1439 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1440 | ||
1441 | if (!is_valid_ether_addr(new_addr)) { | |
1442 | DECLARE_MAC_BUF(mac); | |
1443 | EFX_ERR(efx, "invalid ethernet MAC address requested: %s\n", | |
1444 | print_mac(mac, new_addr)); | |
1445 | return -EINVAL; | |
1446 | } | |
1447 | ||
1448 | memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len); | |
1449 | ||
1450 | /* Reconfigure the MAC */ | |
1451 | efx_reconfigure_port(efx); | |
1452 | ||
1453 | return 0; | |
1454 | } | |
1455 | ||
1456 | /* Context: netif_tx_lock held, BHs disabled. */ | |
1457 | static void efx_set_multicast_list(struct net_device *net_dev) | |
1458 | { | |
1459 | struct efx_nic *efx = net_dev->priv; | |
1460 | struct dev_mc_list *mc_list = net_dev->mc_list; | |
1461 | union efx_multicast_hash *mc_hash = &efx->multicast_hash; | |
1462 | int promiscuous; | |
1463 | u32 crc; | |
1464 | int bit; | |
1465 | int i; | |
1466 | ||
1467 | /* Set per-MAC promiscuity flag and reconfigure MAC if necessary */ | |
1468 | promiscuous = (net_dev->flags & IFF_PROMISC) ? 1 : 0; | |
1469 | if (efx->promiscuous != promiscuous) { | |
1470 | efx->promiscuous = promiscuous; | |
1471 | /* Close the window between efx_stop_port() and efx_flush_all() | |
1472 | * by only queuing work when the port is enabled. */ | |
1473 | if (efx->port_enabled) | |
1474 | queue_work(efx->workqueue, &efx->reconfigure_work); | |
1475 | } | |
1476 | ||
1477 | /* Build multicast hash table */ | |
1478 | if (promiscuous || (net_dev->flags & IFF_ALLMULTI)) { | |
1479 | memset(mc_hash, 0xff, sizeof(*mc_hash)); | |
1480 | } else { | |
1481 | memset(mc_hash, 0x00, sizeof(*mc_hash)); | |
1482 | for (i = 0; i < net_dev->mc_count; i++) { | |
1483 | crc = ether_crc_le(ETH_ALEN, mc_list->dmi_addr); | |
1484 | bit = crc & (EFX_MCAST_HASH_ENTRIES - 1); | |
1485 | set_bit_le(bit, mc_hash->byte); | |
1486 | mc_list = mc_list->next; | |
1487 | } | |
1488 | } | |
1489 | ||
1490 | /* Create and activate new global multicast hash table */ | |
1491 | falcon_set_multicast_hash(efx); | |
1492 | } | |
1493 | ||
1494 | static int efx_netdev_event(struct notifier_block *this, | |
1495 | unsigned long event, void *ptr) | |
1496 | { | |
1497 | struct net_device *net_dev = (struct net_device *)ptr; | |
1498 | ||
1499 | if (net_dev->open == efx_net_open && event == NETDEV_CHANGENAME) { | |
1500 | struct efx_nic *efx = net_dev->priv; | |
1501 | ||
1502 | strcpy(efx->name, net_dev->name); | |
1503 | } | |
1504 | ||
1505 | return NOTIFY_DONE; | |
1506 | } | |
1507 | ||
1508 | static struct notifier_block efx_netdev_notifier = { | |
1509 | .notifier_call = efx_netdev_event, | |
1510 | }; | |
1511 | ||
1512 | static int efx_register_netdev(struct efx_nic *efx) | |
1513 | { | |
1514 | struct net_device *net_dev = efx->net_dev; | |
1515 | int rc; | |
1516 | ||
1517 | net_dev->watchdog_timeo = 5 * HZ; | |
1518 | net_dev->irq = efx->pci_dev->irq; | |
1519 | net_dev->open = efx_net_open; | |
1520 | net_dev->stop = efx_net_stop; | |
1521 | net_dev->get_stats = efx_net_stats; | |
1522 | net_dev->tx_timeout = &efx_watchdog; | |
1523 | net_dev->hard_start_xmit = efx_hard_start_xmit; | |
1524 | net_dev->do_ioctl = efx_ioctl; | |
1525 | net_dev->change_mtu = efx_change_mtu; | |
1526 | net_dev->set_mac_address = efx_set_mac_address; | |
1527 | net_dev->set_multicast_list = efx_set_multicast_list; | |
1528 | #ifdef CONFIG_NET_POLL_CONTROLLER | |
1529 | net_dev->poll_controller = efx_netpoll; | |
1530 | #endif | |
1531 | SET_NETDEV_DEV(net_dev, &efx->pci_dev->dev); | |
1532 | SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops); | |
1533 | ||
1534 | /* Always start with carrier off; PHY events will detect the link */ | |
1535 | netif_carrier_off(efx->net_dev); | |
1536 | ||
1537 | /* Clear MAC statistics */ | |
1538 | falcon_update_stats_xmac(efx); | |
1539 | memset(&efx->mac_stats, 0, sizeof(efx->mac_stats)); | |
1540 | ||
1541 | rc = register_netdev(net_dev); | |
1542 | if (rc) { | |
1543 | EFX_ERR(efx, "could not register net dev\n"); | |
1544 | return rc; | |
1545 | } | |
1546 | strcpy(efx->name, net_dev->name); | |
1547 | ||
1548 | return 0; | |
1549 | } | |
1550 | ||
1551 | static void efx_unregister_netdev(struct efx_nic *efx) | |
1552 | { | |
1553 | struct efx_tx_queue *tx_queue; | |
1554 | ||
1555 | if (!efx->net_dev) | |
1556 | return; | |
1557 | ||
1558 | BUG_ON(efx->net_dev->priv != efx); | |
1559 | ||
1560 | /* Free up any skbs still remaining. This has to happen before | |
1561 | * we try to unregister the netdev as running their destructors | |
1562 | * may be needed to get the device ref. count to 0. */ | |
1563 | efx_for_each_tx_queue(tx_queue, efx) | |
1564 | efx_release_tx_buffers(tx_queue); | |
1565 | ||
1566 | if (NET_DEV_REGISTERED(efx)) { | |
1567 | strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name)); | |
1568 | unregister_netdev(efx->net_dev); | |
1569 | } | |
1570 | } | |
1571 | ||
1572 | /************************************************************************** | |
1573 | * | |
1574 | * Device reset and suspend | |
1575 | * | |
1576 | **************************************************************************/ | |
1577 | ||
1578 | /* The final hardware and software finalisation before reset. */ | |
1579 | static int efx_reset_down(struct efx_nic *efx, struct ethtool_cmd *ecmd) | |
1580 | { | |
1581 | int rc; | |
1582 | ||
1583 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1584 | ||
1585 | rc = falcon_xmac_get_settings(efx, ecmd); | |
1586 | if (rc) { | |
1587 | EFX_ERR(efx, "could not back up PHY settings\n"); | |
1588 | goto fail; | |
1589 | } | |
1590 | ||
1591 | efx_fini_channels(efx); | |
1592 | return 0; | |
1593 | ||
1594 | fail: | |
1595 | return rc; | |
1596 | } | |
1597 | ||
1598 | /* The first part of software initialisation after a hardware reset | |
1599 | * This function does not handle serialisation with the kernel, it | |
1600 | * assumes the caller has done this */ | |
1601 | static int efx_reset_up(struct efx_nic *efx, struct ethtool_cmd *ecmd) | |
1602 | { | |
1603 | int rc; | |
1604 | ||
1605 | rc = efx_init_channels(efx); | |
1606 | if (rc) | |
1607 | goto fail1; | |
1608 | ||
1609 | /* Restore MAC and PHY settings. */ | |
1610 | rc = falcon_xmac_set_settings(efx, ecmd); | |
1611 | if (rc) { | |
1612 | EFX_ERR(efx, "could not restore PHY settings\n"); | |
1613 | goto fail2; | |
1614 | } | |
1615 | ||
1616 | return 0; | |
1617 | ||
1618 | fail2: | |
1619 | efx_fini_channels(efx); | |
1620 | fail1: | |
1621 | return rc; | |
1622 | } | |
1623 | ||
1624 | /* Reset the NIC as transparently as possible. Do not reset the PHY | |
1625 | * Note that the reset may fail, in which case the card will be left | |
1626 | * in a most-probably-unusable state. | |
1627 | * | |
1628 | * This function will sleep. You cannot reset from within an atomic | |
1629 | * state; use efx_schedule_reset() instead. | |
1630 | * | |
1631 | * Grabs the rtnl_lock. | |
1632 | */ | |
1633 | static int efx_reset(struct efx_nic *efx) | |
1634 | { | |
1635 | struct ethtool_cmd ecmd; | |
1636 | enum reset_type method = efx->reset_pending; | |
1637 | int rc; | |
1638 | ||
1639 | /* Serialise with kernel interfaces */ | |
1640 | rtnl_lock(); | |
1641 | ||
1642 | /* If we're not RUNNING then don't reset. Leave the reset_pending | |
1643 | * flag set so that efx_pci_probe_main will be retried */ | |
1644 | if (efx->state != STATE_RUNNING) { | |
1645 | EFX_INFO(efx, "scheduled reset quenched. NIC not RUNNING\n"); | |
1646 | goto unlock_rtnl; | |
1647 | } | |
1648 | ||
1649 | efx->state = STATE_RESETTING; | |
1650 | EFX_INFO(efx, "resetting (%d)\n", method); | |
1651 | ||
1652 | /* The net_dev->get_stats handler is quite slow, and will fail | |
1653 | * if a fetch is pending over reset. Serialise against it. */ | |
1654 | spin_lock(&efx->stats_lock); | |
1655 | spin_unlock(&efx->stats_lock); | |
1656 | ||
1657 | efx_stop_all(efx); | |
1658 | mutex_lock(&efx->mac_lock); | |
1659 | ||
1660 | rc = efx_reset_down(efx, &ecmd); | |
1661 | if (rc) | |
1662 | goto fail1; | |
1663 | ||
1664 | rc = falcon_reset_hw(efx, method); | |
1665 | if (rc) { | |
1666 | EFX_ERR(efx, "failed to reset hardware\n"); | |
1667 | goto fail2; | |
1668 | } | |
1669 | ||
1670 | /* Allow resets to be rescheduled. */ | |
1671 | efx->reset_pending = RESET_TYPE_NONE; | |
1672 | ||
1673 | /* Reinitialise bus-mastering, which may have been turned off before | |
1674 | * the reset was scheduled. This is still appropriate, even in the | |
1675 | * RESET_TYPE_DISABLE since this driver generally assumes the hardware | |
1676 | * can respond to requests. */ | |
1677 | pci_set_master(efx->pci_dev); | |
1678 | ||
1679 | /* Reinitialise device. This is appropriate in the RESET_TYPE_DISABLE | |
1680 | * case so the driver can talk to external SRAM */ | |
1681 | rc = falcon_init_nic(efx); | |
1682 | if (rc) { | |
1683 | EFX_ERR(efx, "failed to initialise NIC\n"); | |
1684 | goto fail3; | |
1685 | } | |
1686 | ||
1687 | /* Leave device stopped if necessary */ | |
1688 | if (method == RESET_TYPE_DISABLE) { | |
1689 | /* Reinitialise the device anyway so the driver unload sequence | |
1690 | * can talk to the external SRAM */ | |
1691 | (void) falcon_init_nic(efx); | |
1692 | rc = -EIO; | |
1693 | goto fail4; | |
1694 | } | |
1695 | ||
1696 | rc = efx_reset_up(efx, &ecmd); | |
1697 | if (rc) | |
1698 | goto fail5; | |
1699 | ||
1700 | mutex_unlock(&efx->mac_lock); | |
1701 | EFX_LOG(efx, "reset complete\n"); | |
1702 | ||
1703 | efx->state = STATE_RUNNING; | |
1704 | efx_start_all(efx); | |
1705 | ||
1706 | unlock_rtnl: | |
1707 | rtnl_unlock(); | |
1708 | return 0; | |
1709 | ||
1710 | fail5: | |
1711 | fail4: | |
1712 | fail3: | |
1713 | fail2: | |
1714 | fail1: | |
1715 | EFX_ERR(efx, "has been disabled\n"); | |
1716 | efx->state = STATE_DISABLED; | |
1717 | ||
1718 | mutex_unlock(&efx->mac_lock); | |
1719 | rtnl_unlock(); | |
1720 | efx_unregister_netdev(efx); | |
1721 | efx_fini_port(efx); | |
1722 | return rc; | |
1723 | } | |
1724 | ||
1725 | /* The worker thread exists so that code that cannot sleep can | |
1726 | * schedule a reset for later. | |
1727 | */ | |
1728 | static void efx_reset_work(struct work_struct *data) | |
1729 | { | |
1730 | struct efx_nic *nic = container_of(data, struct efx_nic, reset_work); | |
1731 | ||
1732 | efx_reset(nic); | |
1733 | } | |
1734 | ||
1735 | void efx_schedule_reset(struct efx_nic *efx, enum reset_type type) | |
1736 | { | |
1737 | enum reset_type method; | |
1738 | ||
1739 | if (efx->reset_pending != RESET_TYPE_NONE) { | |
1740 | EFX_INFO(efx, "quenching already scheduled reset\n"); | |
1741 | return; | |
1742 | } | |
1743 | ||
1744 | switch (type) { | |
1745 | case RESET_TYPE_INVISIBLE: | |
1746 | case RESET_TYPE_ALL: | |
1747 | case RESET_TYPE_WORLD: | |
1748 | case RESET_TYPE_DISABLE: | |
1749 | method = type; | |
1750 | break; | |
1751 | case RESET_TYPE_RX_RECOVERY: | |
1752 | case RESET_TYPE_RX_DESC_FETCH: | |
1753 | case RESET_TYPE_TX_DESC_FETCH: | |
1754 | case RESET_TYPE_TX_SKIP: | |
1755 | method = RESET_TYPE_INVISIBLE; | |
1756 | break; | |
1757 | default: | |
1758 | method = RESET_TYPE_ALL; | |
1759 | break; | |
1760 | } | |
1761 | ||
1762 | if (method != type) | |
1763 | EFX_LOG(efx, "scheduling reset (%d:%d)\n", type, method); | |
1764 | else | |
1765 | EFX_LOG(efx, "scheduling reset (%d)\n", method); | |
1766 | ||
1767 | efx->reset_pending = method; | |
1768 | ||
1769 | queue_work(efx->workqueue, &efx->reset_work); | |
1770 | } | |
1771 | ||
1772 | /************************************************************************** | |
1773 | * | |
1774 | * List of NICs we support | |
1775 | * | |
1776 | **************************************************************************/ | |
1777 | ||
1778 | /* PCI device ID table */ | |
1779 | static struct pci_device_id efx_pci_table[] __devinitdata = { | |
1780 | {PCI_DEVICE(EFX_VENDID_SFC, FALCON_A_P_DEVID), | |
1781 | .driver_data = (unsigned long) &falcon_a_nic_type}, | |
1782 | {PCI_DEVICE(EFX_VENDID_SFC, FALCON_B_P_DEVID), | |
1783 | .driver_data = (unsigned long) &falcon_b_nic_type}, | |
1784 | {0} /* end of list */ | |
1785 | }; | |
1786 | ||
1787 | /************************************************************************** | |
1788 | * | |
1789 | * Dummy PHY/MAC/Board operations | |
1790 | * | |
1791 | * Can be used where the MAC does not implement this operation | |
1792 | * Needed so all function pointers are valid and do not have to be tested | |
1793 | * before use | |
1794 | * | |
1795 | **************************************************************************/ | |
1796 | int efx_port_dummy_op_int(struct efx_nic *efx) | |
1797 | { | |
1798 | return 0; | |
1799 | } | |
1800 | void efx_port_dummy_op_void(struct efx_nic *efx) {} | |
1801 | void efx_port_dummy_op_blink(struct efx_nic *efx, int blink) {} | |
1802 | ||
1803 | static struct efx_phy_operations efx_dummy_phy_operations = { | |
1804 | .init = efx_port_dummy_op_int, | |
1805 | .reconfigure = efx_port_dummy_op_void, | |
1806 | .check_hw = efx_port_dummy_op_int, | |
1807 | .fini = efx_port_dummy_op_void, | |
1808 | .clear_interrupt = efx_port_dummy_op_void, | |
1809 | .reset_xaui = efx_port_dummy_op_void, | |
1810 | }; | |
1811 | ||
1812 | /* Dummy board operations */ | |
1813 | static int efx_nic_dummy_op_int(struct efx_nic *nic) | |
1814 | { | |
1815 | return 0; | |
1816 | } | |
1817 | ||
1818 | static struct efx_board efx_dummy_board_info = { | |
1819 | .init = efx_nic_dummy_op_int, | |
1820 | .init_leds = efx_port_dummy_op_int, | |
1821 | .set_fault_led = efx_port_dummy_op_blink, | |
1822 | }; | |
1823 | ||
1824 | /************************************************************************** | |
1825 | * | |
1826 | * Data housekeeping | |
1827 | * | |
1828 | **************************************************************************/ | |
1829 | ||
1830 | /* This zeroes out and then fills in the invariants in a struct | |
1831 | * efx_nic (including all sub-structures). | |
1832 | */ | |
1833 | static int efx_init_struct(struct efx_nic *efx, struct efx_nic_type *type, | |
1834 | struct pci_dev *pci_dev, struct net_device *net_dev) | |
1835 | { | |
1836 | struct efx_channel *channel; | |
1837 | struct efx_tx_queue *tx_queue; | |
1838 | struct efx_rx_queue *rx_queue; | |
1839 | int i, rc; | |
1840 | ||
1841 | /* Initialise common structures */ | |
1842 | memset(efx, 0, sizeof(*efx)); | |
1843 | spin_lock_init(&efx->biu_lock); | |
1844 | spin_lock_init(&efx->phy_lock); | |
1845 | INIT_WORK(&efx->reset_work, efx_reset_work); | |
1846 | INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor); | |
1847 | efx->pci_dev = pci_dev; | |
1848 | efx->state = STATE_INIT; | |
1849 | efx->reset_pending = RESET_TYPE_NONE; | |
1850 | strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name)); | |
1851 | efx->board_info = efx_dummy_board_info; | |
1852 | ||
1853 | efx->net_dev = net_dev; | |
1854 | efx->rx_checksum_enabled = 1; | |
1855 | spin_lock_init(&efx->netif_stop_lock); | |
1856 | spin_lock_init(&efx->stats_lock); | |
1857 | mutex_init(&efx->mac_lock); | |
1858 | efx->phy_op = &efx_dummy_phy_operations; | |
1859 | efx->mii.dev = net_dev; | |
1860 | INIT_WORK(&efx->reconfigure_work, efx_reconfigure_work); | |
1861 | atomic_set(&efx->netif_stop_count, 1); | |
1862 | ||
1863 | for (i = 0; i < EFX_MAX_CHANNELS; i++) { | |
1864 | channel = &efx->channel[i]; | |
1865 | channel->efx = efx; | |
1866 | channel->channel = i; | |
1867 | channel->evqnum = i; | |
1868 | channel->work_pending = 0; | |
1869 | } | |
1870 | for (i = 0; i < EFX_MAX_TX_QUEUES; i++) { | |
1871 | tx_queue = &efx->tx_queue[i]; | |
1872 | tx_queue->efx = efx; | |
1873 | tx_queue->queue = i; | |
1874 | tx_queue->buffer = NULL; | |
1875 | tx_queue->channel = &efx->channel[0]; /* for safety */ | |
1876 | } | |
1877 | for (i = 0; i < EFX_MAX_RX_QUEUES; i++) { | |
1878 | rx_queue = &efx->rx_queue[i]; | |
1879 | rx_queue->efx = efx; | |
1880 | rx_queue->queue = i; | |
1881 | rx_queue->channel = &efx->channel[0]; /* for safety */ | |
1882 | rx_queue->buffer = NULL; | |
1883 | spin_lock_init(&rx_queue->add_lock); | |
1884 | INIT_DELAYED_WORK(&rx_queue->work, efx_rx_work); | |
1885 | } | |
1886 | ||
1887 | efx->type = type; | |
1888 | ||
1889 | /* Sanity-check NIC type */ | |
1890 | EFX_BUG_ON_PARANOID(efx->type->txd_ring_mask & | |
1891 | (efx->type->txd_ring_mask + 1)); | |
1892 | EFX_BUG_ON_PARANOID(efx->type->rxd_ring_mask & | |
1893 | (efx->type->rxd_ring_mask + 1)); | |
1894 | EFX_BUG_ON_PARANOID(efx->type->evq_size & | |
1895 | (efx->type->evq_size - 1)); | |
1896 | /* As close as we can get to guaranteeing that we don't overflow */ | |
1897 | EFX_BUG_ON_PARANOID(efx->type->evq_size < | |
1898 | (efx->type->txd_ring_mask + 1 + | |
1899 | efx->type->rxd_ring_mask + 1)); | |
1900 | EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS); | |
1901 | ||
1902 | /* Higher numbered interrupt modes are less capable! */ | |
1903 | efx->interrupt_mode = max(efx->type->max_interrupt_mode, | |
1904 | interrupt_mode); | |
1905 | ||
1906 | efx->workqueue = create_singlethread_workqueue("sfc_work"); | |
1907 | if (!efx->workqueue) { | |
1908 | rc = -ENOMEM; | |
1909 | goto fail1; | |
1910 | } | |
1911 | ||
1912 | return 0; | |
1913 | ||
1914 | fail1: | |
1915 | return rc; | |
1916 | } | |
1917 | ||
1918 | static void efx_fini_struct(struct efx_nic *efx) | |
1919 | { | |
1920 | if (efx->workqueue) { | |
1921 | destroy_workqueue(efx->workqueue); | |
1922 | efx->workqueue = NULL; | |
1923 | } | |
1924 | } | |
1925 | ||
1926 | /************************************************************************** | |
1927 | * | |
1928 | * PCI interface | |
1929 | * | |
1930 | **************************************************************************/ | |
1931 | ||
1932 | /* Main body of final NIC shutdown code | |
1933 | * This is called only at module unload (or hotplug removal). | |
1934 | */ | |
1935 | static void efx_pci_remove_main(struct efx_nic *efx) | |
1936 | { | |
1937 | EFX_ASSERT_RESET_SERIALISED(efx); | |
1938 | ||
1939 | /* Skip everything if we never obtained a valid membase */ | |
1940 | if (!efx->membase) | |
1941 | return; | |
1942 | ||
1943 | efx_fini_channels(efx); | |
1944 | efx_fini_port(efx); | |
1945 | ||
1946 | /* Shutdown the board, then the NIC and board state */ | |
1947 | falcon_fini_interrupt(efx); | |
1948 | ||
1949 | efx_fini_napi(efx); | |
1950 | efx_remove_all(efx); | |
1951 | } | |
1952 | ||
1953 | /* Final NIC shutdown | |
1954 | * This is called only at module unload (or hotplug removal). | |
1955 | */ | |
1956 | static void efx_pci_remove(struct pci_dev *pci_dev) | |
1957 | { | |
1958 | struct efx_nic *efx; | |
1959 | ||
1960 | efx = pci_get_drvdata(pci_dev); | |
1961 | if (!efx) | |
1962 | return; | |
1963 | ||
1964 | /* Mark the NIC as fini, then stop the interface */ | |
1965 | rtnl_lock(); | |
1966 | efx->state = STATE_FINI; | |
1967 | dev_close(efx->net_dev); | |
1968 | ||
1969 | /* Allow any queued efx_resets() to complete */ | |
1970 | rtnl_unlock(); | |
1971 | ||
1972 | if (efx->membase == NULL) | |
1973 | goto out; | |
1974 | ||
1975 | efx_unregister_netdev(efx); | |
1976 | ||
1977 | /* Wait for any scheduled resets to complete. No more will be | |
1978 | * scheduled from this point because efx_stop_all() has been | |
1979 | * called, we are no longer registered with driverlink, and | |
1980 | * the net_device's have been removed. */ | |
1981 | flush_workqueue(efx->workqueue); | |
1982 | ||
1983 | efx_pci_remove_main(efx); | |
1984 | ||
1985 | out: | |
1986 | efx_fini_io(efx); | |
1987 | EFX_LOG(efx, "shutdown successful\n"); | |
1988 | ||
1989 | pci_set_drvdata(pci_dev, NULL); | |
1990 | efx_fini_struct(efx); | |
1991 | free_netdev(efx->net_dev); | |
1992 | }; | |
1993 | ||
1994 | /* Main body of NIC initialisation | |
1995 | * This is called at module load (or hotplug insertion, theoretically). | |
1996 | */ | |
1997 | static int efx_pci_probe_main(struct efx_nic *efx) | |
1998 | { | |
1999 | int rc; | |
2000 | ||
2001 | /* Do start-of-day initialisation */ | |
2002 | rc = efx_probe_all(efx); | |
2003 | if (rc) | |
2004 | goto fail1; | |
2005 | ||
2006 | rc = efx_init_napi(efx); | |
2007 | if (rc) | |
2008 | goto fail2; | |
2009 | ||
2010 | /* Initialise the board */ | |
2011 | rc = efx->board_info.init(efx); | |
2012 | if (rc) { | |
2013 | EFX_ERR(efx, "failed to initialise board\n"); | |
2014 | goto fail3; | |
2015 | } | |
2016 | ||
2017 | rc = falcon_init_nic(efx); | |
2018 | if (rc) { | |
2019 | EFX_ERR(efx, "failed to initialise NIC\n"); | |
2020 | goto fail4; | |
2021 | } | |
2022 | ||
2023 | rc = efx_init_port(efx); | |
2024 | if (rc) { | |
2025 | EFX_ERR(efx, "failed to initialise port\n"); | |
2026 | goto fail5; | |
2027 | } | |
2028 | ||
2029 | rc = efx_init_channels(efx); | |
2030 | if (rc) | |
2031 | goto fail6; | |
2032 | ||
2033 | rc = falcon_init_interrupt(efx); | |
2034 | if (rc) | |
2035 | goto fail7; | |
2036 | ||
2037 | return 0; | |
2038 | ||
2039 | fail7: | |
2040 | efx_fini_channels(efx); | |
2041 | fail6: | |
2042 | efx_fini_port(efx); | |
2043 | fail5: | |
2044 | fail4: | |
2045 | fail3: | |
2046 | efx_fini_napi(efx); | |
2047 | fail2: | |
2048 | efx_remove_all(efx); | |
2049 | fail1: | |
2050 | return rc; | |
2051 | } | |
2052 | ||
2053 | /* NIC initialisation | |
2054 | * | |
2055 | * This is called at module load (or hotplug insertion, | |
2056 | * theoretically). It sets up PCI mappings, tests and resets the NIC, | |
2057 | * sets up and registers the network devices with the kernel and hooks | |
2058 | * the interrupt service routine. It does not prepare the device for | |
2059 | * transmission; this is left to the first time one of the network | |
2060 | * interfaces is brought up (i.e. efx_net_open). | |
2061 | */ | |
2062 | static int __devinit efx_pci_probe(struct pci_dev *pci_dev, | |
2063 | const struct pci_device_id *entry) | |
2064 | { | |
2065 | struct efx_nic_type *type = (struct efx_nic_type *) entry->driver_data; | |
2066 | struct net_device *net_dev; | |
2067 | struct efx_nic *efx; | |
2068 | int i, rc; | |
2069 | ||
2070 | /* Allocate and initialise a struct net_device and struct efx_nic */ | |
2071 | net_dev = alloc_etherdev(sizeof(*efx)); | |
2072 | if (!net_dev) | |
2073 | return -ENOMEM; | |
2074 | net_dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_HIGHDMA; | |
2075 | if (lro) | |
2076 | net_dev->features |= NETIF_F_LRO; | |
2077 | efx = net_dev->priv; | |
2078 | pci_set_drvdata(pci_dev, efx); | |
2079 | rc = efx_init_struct(efx, type, pci_dev, net_dev); | |
2080 | if (rc) | |
2081 | goto fail1; | |
2082 | ||
2083 | EFX_INFO(efx, "Solarflare Communications NIC detected\n"); | |
2084 | ||
2085 | /* Set up basic I/O (BAR mappings etc) */ | |
2086 | rc = efx_init_io(efx); | |
2087 | if (rc) | |
2088 | goto fail2; | |
2089 | ||
2090 | /* No serialisation is required with the reset path because | |
2091 | * we're in STATE_INIT. */ | |
2092 | for (i = 0; i < 5; i++) { | |
2093 | rc = efx_pci_probe_main(efx); | |
2094 | if (rc == 0) | |
2095 | break; | |
2096 | ||
2097 | /* Serialise against efx_reset(). No more resets will be | |
2098 | * scheduled since efx_stop_all() has been called, and we | |
2099 | * have not and never have been registered with either | |
2100 | * the rtnetlink or driverlink layers. */ | |
2101 | cancel_work_sync(&efx->reset_work); | |
2102 | ||
2103 | /* Retry if a recoverably reset event has been scheduled */ | |
2104 | if ((efx->reset_pending != RESET_TYPE_INVISIBLE) && | |
2105 | (efx->reset_pending != RESET_TYPE_ALL)) | |
2106 | goto fail3; | |
2107 | ||
2108 | efx->reset_pending = RESET_TYPE_NONE; | |
2109 | } | |
2110 | ||
2111 | if (rc) { | |
2112 | EFX_ERR(efx, "Could not reset NIC\n"); | |
2113 | goto fail4; | |
2114 | } | |
2115 | ||
2116 | /* Switch to the running state before we expose the device to | |
2117 | * the OS. This is to ensure that the initial gathering of | |
2118 | * MAC stats succeeds. */ | |
2119 | rtnl_lock(); | |
2120 | efx->state = STATE_RUNNING; | |
2121 | rtnl_unlock(); | |
2122 | ||
2123 | rc = efx_register_netdev(efx); | |
2124 | if (rc) | |
2125 | goto fail5; | |
2126 | ||
2127 | EFX_LOG(efx, "initialisation successful\n"); | |
2128 | ||
2129 | return 0; | |
2130 | ||
2131 | fail5: | |
2132 | efx_pci_remove_main(efx); | |
2133 | fail4: | |
2134 | fail3: | |
2135 | efx_fini_io(efx); | |
2136 | fail2: | |
2137 | efx_fini_struct(efx); | |
2138 | fail1: | |
2139 | EFX_LOG(efx, "initialisation failed. rc=%d\n", rc); | |
2140 | free_netdev(net_dev); | |
2141 | return rc; | |
2142 | } | |
2143 | ||
2144 | static struct pci_driver efx_pci_driver = { | |
2145 | .name = EFX_DRIVER_NAME, | |
2146 | .id_table = efx_pci_table, | |
2147 | .probe = efx_pci_probe, | |
2148 | .remove = efx_pci_remove, | |
2149 | }; | |
2150 | ||
2151 | /************************************************************************** | |
2152 | * | |
2153 | * Kernel module interface | |
2154 | * | |
2155 | *************************************************************************/ | |
2156 | ||
2157 | module_param(interrupt_mode, uint, 0444); | |
2158 | MODULE_PARM_DESC(interrupt_mode, | |
2159 | "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)"); | |
2160 | ||
2161 | static int __init efx_init_module(void) | |
2162 | { | |
2163 | int rc; | |
2164 | ||
2165 | printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n"); | |
2166 | ||
2167 | rc = register_netdevice_notifier(&efx_netdev_notifier); | |
2168 | if (rc) | |
2169 | goto err_notifier; | |
2170 | ||
2171 | refill_workqueue = create_workqueue("sfc_refill"); | |
2172 | if (!refill_workqueue) { | |
2173 | rc = -ENOMEM; | |
2174 | goto err_refill; | |
2175 | } | |
2176 | ||
2177 | rc = pci_register_driver(&efx_pci_driver); | |
2178 | if (rc < 0) | |
2179 | goto err_pci; | |
2180 | ||
2181 | return 0; | |
2182 | ||
2183 | err_pci: | |
2184 | destroy_workqueue(refill_workqueue); | |
2185 | err_refill: | |
2186 | unregister_netdevice_notifier(&efx_netdev_notifier); | |
2187 | err_notifier: | |
2188 | return rc; | |
2189 | } | |
2190 | ||
2191 | static void __exit efx_exit_module(void) | |
2192 | { | |
2193 | printk(KERN_INFO "Solarflare NET driver unloading\n"); | |
2194 | ||
2195 | pci_unregister_driver(&efx_pci_driver); | |
2196 | destroy_workqueue(refill_workqueue); | |
2197 | unregister_netdevice_notifier(&efx_netdev_notifier); | |
2198 | ||
2199 | } | |
2200 | ||
2201 | module_init(efx_init_module); | |
2202 | module_exit(efx_exit_module); | |
2203 | ||
2204 | MODULE_AUTHOR("Michael Brown <mbrown@fensystems.co.uk> and " | |
2205 | "Solarflare Communications"); | |
2206 | MODULE_DESCRIPTION("Solarflare Communications network driver"); | |
2207 | MODULE_LICENSE("GPL"); | |
2208 | MODULE_DEVICE_TABLE(pci, efx_pci_table); |