xref: /linux/drivers/net/ethernet/sfc/efx.c (revision f2ee442115c9b6219083c019939a9cc0c9abb2f8)
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2011 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 <linux/topology.h>
23 #include <linux/gfp.h>
24 #include <linux/cpu_rmap.h>
25 #include "net_driver.h"
26 #include "efx.h"
27 #include "nic.h"
28 
29 #include "mcdi.h"
30 #include "workarounds.h"
31 
32 /**************************************************************************
33  *
34  * Type name strings
35  *
36  **************************************************************************
37  */
38 
39 /* Loopback mode names (see LOOPBACK_MODE()) */
40 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
41 const char *efx_loopback_mode_names[] = {
42 	[LOOPBACK_NONE]		= "NONE",
43 	[LOOPBACK_DATA]		= "DATAPATH",
44 	[LOOPBACK_GMAC]		= "GMAC",
45 	[LOOPBACK_XGMII]	= "XGMII",
46 	[LOOPBACK_XGXS]		= "XGXS",
47 	[LOOPBACK_XAUI]  	= "XAUI",
48 	[LOOPBACK_GMII] 	= "GMII",
49 	[LOOPBACK_SGMII] 	= "SGMII",
50 	[LOOPBACK_XGBR]		= "XGBR",
51 	[LOOPBACK_XFI]		= "XFI",
52 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
53 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
54 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
55 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
56 	[LOOPBACK_GPHY]		= "GPHY",
57 	[LOOPBACK_PHYXS]	= "PHYXS",
58 	[LOOPBACK_PCS]	 	= "PCS",
59 	[LOOPBACK_PMAPMD] 	= "PMA/PMD",
60 	[LOOPBACK_XPORT]	= "XPORT",
61 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
62 	[LOOPBACK_XAUI_WS]  	= "XAUI_WS",
63 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
64 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
65 	[LOOPBACK_GMII_WS] 	= "GMII_WS",
66 	[LOOPBACK_XFI_WS]	= "XFI_WS",
67 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
68 	[LOOPBACK_PHYXS_WS]  	= "PHYXS_WS",
69 };
70 
71 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
72 const char *efx_reset_type_names[] = {
73 	[RESET_TYPE_INVISIBLE]     = "INVISIBLE",
74 	[RESET_TYPE_ALL]           = "ALL",
75 	[RESET_TYPE_WORLD]         = "WORLD",
76 	[RESET_TYPE_DISABLE]       = "DISABLE",
77 	[RESET_TYPE_TX_WATCHDOG]   = "TX_WATCHDOG",
78 	[RESET_TYPE_INT_ERROR]     = "INT_ERROR",
79 	[RESET_TYPE_RX_RECOVERY]   = "RX_RECOVERY",
80 	[RESET_TYPE_RX_DESC_FETCH] = "RX_DESC_FETCH",
81 	[RESET_TYPE_TX_DESC_FETCH] = "TX_DESC_FETCH",
82 	[RESET_TYPE_TX_SKIP]       = "TX_SKIP",
83 	[RESET_TYPE_MC_FAILURE]    = "MC_FAILURE",
84 };
85 
86 #define EFX_MAX_MTU (9 * 1024)
87 
88 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
89  * queued onto this work queue. This is not a per-nic work queue, because
90  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
91  */
92 static struct workqueue_struct *reset_workqueue;
93 
94 /**************************************************************************
95  *
96  * Configurable values
97  *
98  *************************************************************************/
99 
100 /*
101  * Use separate channels for TX and RX events
102  *
103  * Set this to 1 to use separate channels for TX and RX. It allows us
104  * to control interrupt affinity separately for TX and RX.
105  *
106  * This is only used in MSI-X interrupt mode
107  */
108 static unsigned int separate_tx_channels;
109 module_param(separate_tx_channels, uint, 0444);
110 MODULE_PARM_DESC(separate_tx_channels,
111 		 "Use separate channels for TX and RX");
112 
113 /* This is the weight assigned to each of the (per-channel) virtual
114  * NAPI devices.
115  */
116 static int napi_weight = 64;
117 
118 /* This is the time (in jiffies) between invocations of the hardware
119  * monitor.  On Falcon-based NICs, this will:
120  * - Check the on-board hardware monitor;
121  * - Poll the link state and reconfigure the hardware as necessary.
122  */
123 static unsigned int efx_monitor_interval = 1 * HZ;
124 
125 /* This controls whether or not the driver will initialise devices
126  * with invalid MAC addresses stored in the EEPROM or flash.  If true,
127  * such devices will be initialised with a random locally-generated
128  * MAC address.  This allows for loading the sfc_mtd driver to
129  * reprogram the flash, even if the flash contents (including the MAC
130  * address) have previously been erased.
131  */
132 static unsigned int allow_bad_hwaddr;
133 
134 /* Initial interrupt moderation settings.  They can be modified after
135  * module load with ethtool.
136  *
137  * The default for RX should strike a balance between increasing the
138  * round-trip latency and reducing overhead.
139  */
140 static unsigned int rx_irq_mod_usec = 60;
141 
142 /* Initial interrupt moderation settings.  They can be modified after
143  * module load with ethtool.
144  *
145  * This default is chosen to ensure that a 10G link does not go idle
146  * while a TX queue is stopped after it has become full.  A queue is
147  * restarted when it drops below half full.  The time this takes (assuming
148  * worst case 3 descriptors per packet and 1024 descriptors) is
149  *   512 / 3 * 1.2 = 205 usec.
150  */
151 static unsigned int tx_irq_mod_usec = 150;
152 
153 /* This is the first interrupt mode to try out of:
154  * 0 => MSI-X
155  * 1 => MSI
156  * 2 => legacy
157  */
158 static unsigned int interrupt_mode;
159 
160 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
161  * i.e. the number of CPUs among which we may distribute simultaneous
162  * interrupt handling.
163  *
164  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
165  * The default (0) means to assign an interrupt to each package (level II cache)
166  */
167 static unsigned int rss_cpus;
168 module_param(rss_cpus, uint, 0444);
169 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
170 
171 static int phy_flash_cfg;
172 module_param(phy_flash_cfg, int, 0644);
173 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
174 
175 static unsigned irq_adapt_low_thresh = 10000;
176 module_param(irq_adapt_low_thresh, uint, 0644);
177 MODULE_PARM_DESC(irq_adapt_low_thresh,
178 		 "Threshold score for reducing IRQ moderation");
179 
180 static unsigned irq_adapt_high_thresh = 20000;
181 module_param(irq_adapt_high_thresh, uint, 0644);
182 MODULE_PARM_DESC(irq_adapt_high_thresh,
183 		 "Threshold score for increasing IRQ moderation");
184 
185 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
186 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
187 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
188 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
189 module_param(debug, uint, 0);
190 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
191 
192 /**************************************************************************
193  *
194  * Utility functions and prototypes
195  *
196  *************************************************************************/
197 
198 static void efx_remove_channels(struct efx_nic *efx);
199 static void efx_remove_port(struct efx_nic *efx);
200 static void efx_init_napi(struct efx_nic *efx);
201 static void efx_fini_napi(struct efx_nic *efx);
202 static void efx_fini_napi_channel(struct efx_channel *channel);
203 static void efx_fini_struct(struct efx_nic *efx);
204 static void efx_start_all(struct efx_nic *efx);
205 static void efx_stop_all(struct efx_nic *efx);
206 
207 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
208 	do {						\
209 		if ((efx->state == STATE_RUNNING) ||	\
210 		    (efx->state == STATE_DISABLED))	\
211 			ASSERT_RTNL();			\
212 	} while (0)
213 
214 /**************************************************************************
215  *
216  * Event queue processing
217  *
218  *************************************************************************/
219 
220 /* Process channel's event queue
221  *
222  * This function is responsible for processing the event queue of a
223  * single channel.  The caller must guarantee that this function will
224  * never be concurrently called more than once on the same channel,
225  * though different channels may be being processed concurrently.
226  */
227 static int efx_process_channel(struct efx_channel *channel, int budget)
228 {
229 	struct efx_nic *efx = channel->efx;
230 	int spent;
231 
232 	if (unlikely(efx->reset_pending || !channel->enabled))
233 		return 0;
234 
235 	spent = efx_nic_process_eventq(channel, budget);
236 	if (spent == 0)
237 		return 0;
238 
239 	/* Deliver last RX packet. */
240 	if (channel->rx_pkt) {
241 		__efx_rx_packet(channel, channel->rx_pkt,
242 				channel->rx_pkt_csummed);
243 		channel->rx_pkt = NULL;
244 	}
245 
246 	efx_rx_strategy(channel);
247 
248 	efx_fast_push_rx_descriptors(efx_channel_get_rx_queue(channel));
249 
250 	return spent;
251 }
252 
253 /* Mark channel as finished processing
254  *
255  * Note that since we will not receive further interrupts for this
256  * channel before we finish processing and call the eventq_read_ack()
257  * method, there is no need to use the interrupt hold-off timers.
258  */
259 static inline void efx_channel_processed(struct efx_channel *channel)
260 {
261 	/* The interrupt handler for this channel may set work_pending
262 	 * as soon as we acknowledge the events we've seen.  Make sure
263 	 * it's cleared before then. */
264 	channel->work_pending = false;
265 	smp_wmb();
266 
267 	efx_nic_eventq_read_ack(channel);
268 }
269 
270 /* NAPI poll handler
271  *
272  * NAPI guarantees serialisation of polls of the same device, which
273  * provides the guarantee required by efx_process_channel().
274  */
275 static int efx_poll(struct napi_struct *napi, int budget)
276 {
277 	struct efx_channel *channel =
278 		container_of(napi, struct efx_channel, napi_str);
279 	struct efx_nic *efx = channel->efx;
280 	int spent;
281 
282 	netif_vdbg(efx, intr, efx->net_dev,
283 		   "channel %d NAPI poll executing on CPU %d\n",
284 		   channel->channel, raw_smp_processor_id());
285 
286 	spent = efx_process_channel(channel, budget);
287 
288 	if (spent < budget) {
289 		if (channel->channel < efx->n_rx_channels &&
290 		    efx->irq_rx_adaptive &&
291 		    unlikely(++channel->irq_count == 1000)) {
292 			if (unlikely(channel->irq_mod_score <
293 				     irq_adapt_low_thresh)) {
294 				if (channel->irq_moderation > 1) {
295 					channel->irq_moderation -= 1;
296 					efx->type->push_irq_moderation(channel);
297 				}
298 			} else if (unlikely(channel->irq_mod_score >
299 					    irq_adapt_high_thresh)) {
300 				if (channel->irq_moderation <
301 				    efx->irq_rx_moderation) {
302 					channel->irq_moderation += 1;
303 					efx->type->push_irq_moderation(channel);
304 				}
305 			}
306 			channel->irq_count = 0;
307 			channel->irq_mod_score = 0;
308 		}
309 
310 		efx_filter_rfs_expire(channel);
311 
312 		/* There is no race here; although napi_disable() will
313 		 * only wait for napi_complete(), this isn't a problem
314 		 * since efx_channel_processed() will have no effect if
315 		 * interrupts have already been disabled.
316 		 */
317 		napi_complete(napi);
318 		efx_channel_processed(channel);
319 	}
320 
321 	return spent;
322 }
323 
324 /* Process the eventq of the specified channel immediately on this CPU
325  *
326  * Disable hardware generated interrupts, wait for any existing
327  * processing to finish, then directly poll (and ack ) the eventq.
328  * Finally reenable NAPI and interrupts.
329  *
330  * This is for use only during a loopback self-test.  It must not
331  * deliver any packets up the stack as this can result in deadlock.
332  */
333 void efx_process_channel_now(struct efx_channel *channel)
334 {
335 	struct efx_nic *efx = channel->efx;
336 
337 	BUG_ON(channel->channel >= efx->n_channels);
338 	BUG_ON(!channel->enabled);
339 	BUG_ON(!efx->loopback_selftest);
340 
341 	/* Disable interrupts and wait for ISRs to complete */
342 	efx_nic_disable_interrupts(efx);
343 	if (efx->legacy_irq) {
344 		synchronize_irq(efx->legacy_irq);
345 		efx->legacy_irq_enabled = false;
346 	}
347 	if (channel->irq)
348 		synchronize_irq(channel->irq);
349 
350 	/* Wait for any NAPI processing to complete */
351 	napi_disable(&channel->napi_str);
352 
353 	/* Poll the channel */
354 	efx_process_channel(channel, channel->eventq_mask + 1);
355 
356 	/* Ack the eventq. This may cause an interrupt to be generated
357 	 * when they are reenabled */
358 	efx_channel_processed(channel);
359 
360 	napi_enable(&channel->napi_str);
361 	if (efx->legacy_irq)
362 		efx->legacy_irq_enabled = true;
363 	efx_nic_enable_interrupts(efx);
364 }
365 
366 /* Create event queue
367  * Event queue memory allocations are done only once.  If the channel
368  * is reset, the memory buffer will be reused; this guards against
369  * errors during channel reset and also simplifies interrupt handling.
370  */
371 static int efx_probe_eventq(struct efx_channel *channel)
372 {
373 	struct efx_nic *efx = channel->efx;
374 	unsigned long entries;
375 
376 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
377 		  "chan %d create event queue\n", channel->channel);
378 
379 	/* Build an event queue with room for one event per tx and rx buffer,
380 	 * plus some extra for link state events and MCDI completions. */
381 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
382 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
383 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
384 
385 	return efx_nic_probe_eventq(channel);
386 }
387 
388 /* Prepare channel's event queue */
389 static void efx_init_eventq(struct efx_channel *channel)
390 {
391 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
392 		  "chan %d init event queue\n", channel->channel);
393 
394 	channel->eventq_read_ptr = 0;
395 
396 	efx_nic_init_eventq(channel);
397 }
398 
399 static void efx_fini_eventq(struct efx_channel *channel)
400 {
401 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
402 		  "chan %d fini event queue\n", channel->channel);
403 
404 	efx_nic_fini_eventq(channel);
405 }
406 
407 static void efx_remove_eventq(struct efx_channel *channel)
408 {
409 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
410 		  "chan %d remove event queue\n", channel->channel);
411 
412 	efx_nic_remove_eventq(channel);
413 }
414 
415 /**************************************************************************
416  *
417  * Channel handling
418  *
419  *************************************************************************/
420 
421 /* Allocate and initialise a channel structure, optionally copying
422  * parameters (but not resources) from an old channel structure. */
423 static struct efx_channel *
424 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
425 {
426 	struct efx_channel *channel;
427 	struct efx_rx_queue *rx_queue;
428 	struct efx_tx_queue *tx_queue;
429 	int j;
430 
431 	if (old_channel) {
432 		channel = kmalloc(sizeof(*channel), GFP_KERNEL);
433 		if (!channel)
434 			return NULL;
435 
436 		*channel = *old_channel;
437 
438 		channel->napi_dev = NULL;
439 		memset(&channel->eventq, 0, sizeof(channel->eventq));
440 
441 		rx_queue = &channel->rx_queue;
442 		rx_queue->buffer = NULL;
443 		memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
444 
445 		for (j = 0; j < EFX_TXQ_TYPES; j++) {
446 			tx_queue = &channel->tx_queue[j];
447 			if (tx_queue->channel)
448 				tx_queue->channel = channel;
449 			tx_queue->buffer = NULL;
450 			memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
451 		}
452 	} else {
453 		channel = kzalloc(sizeof(*channel), GFP_KERNEL);
454 		if (!channel)
455 			return NULL;
456 
457 		channel->efx = efx;
458 		channel->channel = i;
459 
460 		for (j = 0; j < EFX_TXQ_TYPES; j++) {
461 			tx_queue = &channel->tx_queue[j];
462 			tx_queue->efx = efx;
463 			tx_queue->queue = i * EFX_TXQ_TYPES + j;
464 			tx_queue->channel = channel;
465 		}
466 	}
467 
468 	rx_queue = &channel->rx_queue;
469 	rx_queue->efx = efx;
470 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
471 		    (unsigned long)rx_queue);
472 
473 	return channel;
474 }
475 
476 static int efx_probe_channel(struct efx_channel *channel)
477 {
478 	struct efx_tx_queue *tx_queue;
479 	struct efx_rx_queue *rx_queue;
480 	int rc;
481 
482 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
483 		  "creating channel %d\n", channel->channel);
484 
485 	rc = efx_probe_eventq(channel);
486 	if (rc)
487 		goto fail1;
488 
489 	efx_for_each_channel_tx_queue(tx_queue, channel) {
490 		rc = efx_probe_tx_queue(tx_queue);
491 		if (rc)
492 			goto fail2;
493 	}
494 
495 	efx_for_each_channel_rx_queue(rx_queue, channel) {
496 		rc = efx_probe_rx_queue(rx_queue);
497 		if (rc)
498 			goto fail3;
499 	}
500 
501 	channel->n_rx_frm_trunc = 0;
502 
503 	return 0;
504 
505  fail3:
506 	efx_for_each_channel_rx_queue(rx_queue, channel)
507 		efx_remove_rx_queue(rx_queue);
508  fail2:
509 	efx_for_each_channel_tx_queue(tx_queue, channel)
510 		efx_remove_tx_queue(tx_queue);
511  fail1:
512 	return rc;
513 }
514 
515 
516 static void efx_set_channel_names(struct efx_nic *efx)
517 {
518 	struct efx_channel *channel;
519 	const char *type = "";
520 	int number;
521 
522 	efx_for_each_channel(channel, efx) {
523 		number = channel->channel;
524 		if (efx->n_channels > efx->n_rx_channels) {
525 			if (channel->channel < efx->n_rx_channels) {
526 				type = "-rx";
527 			} else {
528 				type = "-tx";
529 				number -= efx->n_rx_channels;
530 			}
531 		}
532 		snprintf(efx->channel_name[channel->channel],
533 			 sizeof(efx->channel_name[0]),
534 			 "%s%s-%d", efx->name, type, number);
535 	}
536 }
537 
538 static int efx_probe_channels(struct efx_nic *efx)
539 {
540 	struct efx_channel *channel;
541 	int rc;
542 
543 	/* Restart special buffer allocation */
544 	efx->next_buffer_table = 0;
545 
546 	efx_for_each_channel(channel, efx) {
547 		rc = efx_probe_channel(channel);
548 		if (rc) {
549 			netif_err(efx, probe, efx->net_dev,
550 				  "failed to create channel %d\n",
551 				  channel->channel);
552 			goto fail;
553 		}
554 	}
555 	efx_set_channel_names(efx);
556 
557 	return 0;
558 
559 fail:
560 	efx_remove_channels(efx);
561 	return rc;
562 }
563 
564 /* Channels are shutdown and reinitialised whilst the NIC is running
565  * to propagate configuration changes (mtu, checksum offload), or
566  * to clear hardware error conditions
567  */
568 static void efx_init_channels(struct efx_nic *efx)
569 {
570 	struct efx_tx_queue *tx_queue;
571 	struct efx_rx_queue *rx_queue;
572 	struct efx_channel *channel;
573 
574 	/* Calculate the rx buffer allocation parameters required to
575 	 * support the current MTU, including padding for header
576 	 * alignment and overruns.
577 	 */
578 	efx->rx_buffer_len = (max(EFX_PAGE_IP_ALIGN, NET_IP_ALIGN) +
579 			      EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
580 			      efx->type->rx_buffer_hash_size +
581 			      efx->type->rx_buffer_padding);
582 	efx->rx_buffer_order = get_order(efx->rx_buffer_len +
583 					 sizeof(struct efx_rx_page_state));
584 
585 	/* Initialise the channels */
586 	efx_for_each_channel(channel, efx) {
587 		netif_dbg(channel->efx, drv, channel->efx->net_dev,
588 			  "init chan %d\n", channel->channel);
589 
590 		efx_init_eventq(channel);
591 
592 		efx_for_each_channel_tx_queue(tx_queue, channel)
593 			efx_init_tx_queue(tx_queue);
594 
595 		/* The rx buffer allocation strategy is MTU dependent */
596 		efx_rx_strategy(channel);
597 
598 		efx_for_each_channel_rx_queue(rx_queue, channel)
599 			efx_init_rx_queue(rx_queue);
600 
601 		WARN_ON(channel->rx_pkt != NULL);
602 		efx_rx_strategy(channel);
603 	}
604 }
605 
606 /* This enables event queue processing and packet transmission.
607  *
608  * Note that this function is not allowed to fail, since that would
609  * introduce too much complexity into the suspend/resume path.
610  */
611 static void efx_start_channel(struct efx_channel *channel)
612 {
613 	struct efx_rx_queue *rx_queue;
614 
615 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
616 		  "starting chan %d\n", channel->channel);
617 
618 	/* The interrupt handler for this channel may set work_pending
619 	 * as soon as we enable it.  Make sure it's cleared before
620 	 * then.  Similarly, make sure it sees the enabled flag set. */
621 	channel->work_pending = false;
622 	channel->enabled = true;
623 	smp_wmb();
624 
625 	/* Fill the queues before enabling NAPI */
626 	efx_for_each_channel_rx_queue(rx_queue, channel)
627 		efx_fast_push_rx_descriptors(rx_queue);
628 
629 	napi_enable(&channel->napi_str);
630 }
631 
632 /* This disables event queue processing and packet transmission.
633  * This function does not guarantee that all queue processing
634  * (e.g. RX refill) is complete.
635  */
636 static void efx_stop_channel(struct efx_channel *channel)
637 {
638 	if (!channel->enabled)
639 		return;
640 
641 	netif_dbg(channel->efx, ifdown, channel->efx->net_dev,
642 		  "stop chan %d\n", channel->channel);
643 
644 	channel->enabled = false;
645 	napi_disable(&channel->napi_str);
646 }
647 
648 static void efx_fini_channels(struct efx_nic *efx)
649 {
650 	struct efx_channel *channel;
651 	struct efx_tx_queue *tx_queue;
652 	struct efx_rx_queue *rx_queue;
653 	int rc;
654 
655 	EFX_ASSERT_RESET_SERIALISED(efx);
656 	BUG_ON(efx->port_enabled);
657 
658 	rc = efx_nic_flush_queues(efx);
659 	if (rc && EFX_WORKAROUND_7803(efx)) {
660 		/* Schedule a reset to recover from the flush failure. The
661 		 * descriptor caches reference memory we're about to free,
662 		 * but falcon_reconfigure_mac_wrapper() won't reconnect
663 		 * the MACs because of the pending reset. */
664 		netif_err(efx, drv, efx->net_dev,
665 			  "Resetting to recover from flush failure\n");
666 		efx_schedule_reset(efx, RESET_TYPE_ALL);
667 	} else if (rc) {
668 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
669 	} else {
670 		netif_dbg(efx, drv, efx->net_dev,
671 			  "successfully flushed all queues\n");
672 	}
673 
674 	efx_for_each_channel(channel, efx) {
675 		netif_dbg(channel->efx, drv, channel->efx->net_dev,
676 			  "shut down chan %d\n", channel->channel);
677 
678 		efx_for_each_channel_rx_queue(rx_queue, channel)
679 			efx_fini_rx_queue(rx_queue);
680 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
681 			efx_fini_tx_queue(tx_queue);
682 		efx_fini_eventq(channel);
683 	}
684 }
685 
686 static void efx_remove_channel(struct efx_channel *channel)
687 {
688 	struct efx_tx_queue *tx_queue;
689 	struct efx_rx_queue *rx_queue;
690 
691 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
692 		  "destroy chan %d\n", channel->channel);
693 
694 	efx_for_each_channel_rx_queue(rx_queue, channel)
695 		efx_remove_rx_queue(rx_queue);
696 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
697 		efx_remove_tx_queue(tx_queue);
698 	efx_remove_eventq(channel);
699 }
700 
701 static void efx_remove_channels(struct efx_nic *efx)
702 {
703 	struct efx_channel *channel;
704 
705 	efx_for_each_channel(channel, efx)
706 		efx_remove_channel(channel);
707 }
708 
709 int
710 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
711 {
712 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
713 	u32 old_rxq_entries, old_txq_entries;
714 	unsigned i;
715 	int rc;
716 
717 	efx_stop_all(efx);
718 	efx_fini_channels(efx);
719 
720 	/* Clone channels */
721 	memset(other_channel, 0, sizeof(other_channel));
722 	for (i = 0; i < efx->n_channels; i++) {
723 		channel = efx_alloc_channel(efx, i, efx->channel[i]);
724 		if (!channel) {
725 			rc = -ENOMEM;
726 			goto out;
727 		}
728 		other_channel[i] = channel;
729 	}
730 
731 	/* Swap entry counts and channel pointers */
732 	old_rxq_entries = efx->rxq_entries;
733 	old_txq_entries = efx->txq_entries;
734 	efx->rxq_entries = rxq_entries;
735 	efx->txq_entries = txq_entries;
736 	for (i = 0; i < efx->n_channels; i++) {
737 		channel = efx->channel[i];
738 		efx->channel[i] = other_channel[i];
739 		other_channel[i] = channel;
740 	}
741 
742 	rc = efx_probe_channels(efx);
743 	if (rc)
744 		goto rollback;
745 
746 	efx_init_napi(efx);
747 
748 	/* Destroy old channels */
749 	for (i = 0; i < efx->n_channels; i++) {
750 		efx_fini_napi_channel(other_channel[i]);
751 		efx_remove_channel(other_channel[i]);
752 	}
753 out:
754 	/* Free unused channel structures */
755 	for (i = 0; i < efx->n_channels; i++)
756 		kfree(other_channel[i]);
757 
758 	efx_init_channels(efx);
759 	efx_start_all(efx);
760 	return rc;
761 
762 rollback:
763 	/* Swap back */
764 	efx->rxq_entries = old_rxq_entries;
765 	efx->txq_entries = old_txq_entries;
766 	for (i = 0; i < efx->n_channels; i++) {
767 		channel = efx->channel[i];
768 		efx->channel[i] = other_channel[i];
769 		other_channel[i] = channel;
770 	}
771 	goto out;
772 }
773 
774 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
775 {
776 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
777 }
778 
779 /**************************************************************************
780  *
781  * Port handling
782  *
783  **************************************************************************/
784 
785 /* This ensures that the kernel is kept informed (via
786  * netif_carrier_on/off) of the link status, and also maintains the
787  * link status's stop on the port's TX queue.
788  */
789 void efx_link_status_changed(struct efx_nic *efx)
790 {
791 	struct efx_link_state *link_state = &efx->link_state;
792 
793 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
794 	 * that no events are triggered between unregister_netdev() and the
795 	 * driver unloading. A more general condition is that NETDEV_CHANGE
796 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
797 	if (!netif_running(efx->net_dev))
798 		return;
799 
800 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
801 		efx->n_link_state_changes++;
802 
803 		if (link_state->up)
804 			netif_carrier_on(efx->net_dev);
805 		else
806 			netif_carrier_off(efx->net_dev);
807 	}
808 
809 	/* Status message for kernel log */
810 	if (link_state->up) {
811 		netif_info(efx, link, efx->net_dev,
812 			   "link up at %uMbps %s-duplex (MTU %d)%s\n",
813 			   link_state->speed, link_state->fd ? "full" : "half",
814 			   efx->net_dev->mtu,
815 			   (efx->promiscuous ? " [PROMISC]" : ""));
816 	} else {
817 		netif_info(efx, link, efx->net_dev, "link down\n");
818 	}
819 
820 }
821 
822 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
823 {
824 	efx->link_advertising = advertising;
825 	if (advertising) {
826 		if (advertising & ADVERTISED_Pause)
827 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
828 		else
829 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
830 		if (advertising & ADVERTISED_Asym_Pause)
831 			efx->wanted_fc ^= EFX_FC_TX;
832 	}
833 }
834 
835 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
836 {
837 	efx->wanted_fc = wanted_fc;
838 	if (efx->link_advertising) {
839 		if (wanted_fc & EFX_FC_RX)
840 			efx->link_advertising |= (ADVERTISED_Pause |
841 						  ADVERTISED_Asym_Pause);
842 		else
843 			efx->link_advertising &= ~(ADVERTISED_Pause |
844 						   ADVERTISED_Asym_Pause);
845 		if (wanted_fc & EFX_FC_TX)
846 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
847 	}
848 }
849 
850 static void efx_fini_port(struct efx_nic *efx);
851 
852 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
853  * the MAC appropriately. All other PHY configuration changes are pushed
854  * through phy_op->set_settings(), and pushed asynchronously to the MAC
855  * through efx_monitor().
856  *
857  * Callers must hold the mac_lock
858  */
859 int __efx_reconfigure_port(struct efx_nic *efx)
860 {
861 	enum efx_phy_mode phy_mode;
862 	int rc;
863 
864 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
865 
866 	/* Serialise the promiscuous flag with efx_set_multicast_list. */
867 	if (efx_dev_registered(efx)) {
868 		netif_addr_lock_bh(efx->net_dev);
869 		netif_addr_unlock_bh(efx->net_dev);
870 	}
871 
872 	/* Disable PHY transmit in mac level loopbacks */
873 	phy_mode = efx->phy_mode;
874 	if (LOOPBACK_INTERNAL(efx))
875 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
876 	else
877 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
878 
879 	rc = efx->type->reconfigure_port(efx);
880 
881 	if (rc)
882 		efx->phy_mode = phy_mode;
883 
884 	return rc;
885 }
886 
887 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
888  * disabled. */
889 int efx_reconfigure_port(struct efx_nic *efx)
890 {
891 	int rc;
892 
893 	EFX_ASSERT_RESET_SERIALISED(efx);
894 
895 	mutex_lock(&efx->mac_lock);
896 	rc = __efx_reconfigure_port(efx);
897 	mutex_unlock(&efx->mac_lock);
898 
899 	return rc;
900 }
901 
902 /* Asynchronous work item for changing MAC promiscuity and multicast
903  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
904  * MAC directly. */
905 static void efx_mac_work(struct work_struct *data)
906 {
907 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
908 
909 	mutex_lock(&efx->mac_lock);
910 	if (efx->port_enabled) {
911 		efx->type->push_multicast_hash(efx);
912 		efx->mac_op->reconfigure(efx);
913 	}
914 	mutex_unlock(&efx->mac_lock);
915 }
916 
917 static int efx_probe_port(struct efx_nic *efx)
918 {
919 	unsigned char *perm_addr;
920 	int rc;
921 
922 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
923 
924 	if (phy_flash_cfg)
925 		efx->phy_mode = PHY_MODE_SPECIAL;
926 
927 	/* Connect up MAC/PHY operations table */
928 	rc = efx->type->probe_port(efx);
929 	if (rc)
930 		return rc;
931 
932 	/* Sanity check MAC address */
933 	perm_addr = efx->net_dev->perm_addr;
934 	if (is_valid_ether_addr(perm_addr)) {
935 		memcpy(efx->net_dev->dev_addr, perm_addr, ETH_ALEN);
936 	} else {
937 		netif_err(efx, probe, efx->net_dev, "invalid MAC address %pM\n",
938 			  perm_addr);
939 		if (!allow_bad_hwaddr) {
940 			rc = -EINVAL;
941 			goto err;
942 		}
943 		random_ether_addr(efx->net_dev->dev_addr);
944 		netif_info(efx, probe, efx->net_dev,
945 			   "using locally-generated MAC %pM\n",
946 			   efx->net_dev->dev_addr);
947 	}
948 
949 	return 0;
950 
951  err:
952 	efx->type->remove_port(efx);
953 	return rc;
954 }
955 
956 static int efx_init_port(struct efx_nic *efx)
957 {
958 	int rc;
959 
960 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
961 
962 	mutex_lock(&efx->mac_lock);
963 
964 	rc = efx->phy_op->init(efx);
965 	if (rc)
966 		goto fail1;
967 
968 	efx->port_initialized = true;
969 
970 	/* Reconfigure the MAC before creating dma queues (required for
971 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
972 	efx->mac_op->reconfigure(efx);
973 
974 	/* Ensure the PHY advertises the correct flow control settings */
975 	rc = efx->phy_op->reconfigure(efx);
976 	if (rc)
977 		goto fail2;
978 
979 	mutex_unlock(&efx->mac_lock);
980 	return 0;
981 
982 fail2:
983 	efx->phy_op->fini(efx);
984 fail1:
985 	mutex_unlock(&efx->mac_lock);
986 	return rc;
987 }
988 
989 static void efx_start_port(struct efx_nic *efx)
990 {
991 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
992 	BUG_ON(efx->port_enabled);
993 
994 	mutex_lock(&efx->mac_lock);
995 	efx->port_enabled = true;
996 
997 	/* efx_mac_work() might have been scheduled after efx_stop_port(),
998 	 * and then cancelled by efx_flush_all() */
999 	efx->type->push_multicast_hash(efx);
1000 	efx->mac_op->reconfigure(efx);
1001 
1002 	mutex_unlock(&efx->mac_lock);
1003 }
1004 
1005 /* Prevent efx_mac_work() and efx_monitor() from working */
1006 static void efx_stop_port(struct efx_nic *efx)
1007 {
1008 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1009 
1010 	mutex_lock(&efx->mac_lock);
1011 	efx->port_enabled = false;
1012 	mutex_unlock(&efx->mac_lock);
1013 
1014 	/* Serialise against efx_set_multicast_list() */
1015 	if (efx_dev_registered(efx)) {
1016 		netif_addr_lock_bh(efx->net_dev);
1017 		netif_addr_unlock_bh(efx->net_dev);
1018 	}
1019 }
1020 
1021 static void efx_fini_port(struct efx_nic *efx)
1022 {
1023 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1024 
1025 	if (!efx->port_initialized)
1026 		return;
1027 
1028 	efx->phy_op->fini(efx);
1029 	efx->port_initialized = false;
1030 
1031 	efx->link_state.up = false;
1032 	efx_link_status_changed(efx);
1033 }
1034 
1035 static void efx_remove_port(struct efx_nic *efx)
1036 {
1037 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1038 
1039 	efx->type->remove_port(efx);
1040 }
1041 
1042 /**************************************************************************
1043  *
1044  * NIC handling
1045  *
1046  **************************************************************************/
1047 
1048 /* This configures the PCI device to enable I/O and DMA. */
1049 static int efx_init_io(struct efx_nic *efx)
1050 {
1051 	struct pci_dev *pci_dev = efx->pci_dev;
1052 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1053 	int rc;
1054 
1055 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1056 
1057 	rc = pci_enable_device(pci_dev);
1058 	if (rc) {
1059 		netif_err(efx, probe, efx->net_dev,
1060 			  "failed to enable PCI device\n");
1061 		goto fail1;
1062 	}
1063 
1064 	pci_set_master(pci_dev);
1065 
1066 	/* Set the PCI DMA mask.  Try all possibilities from our
1067 	 * genuine mask down to 32 bits, because some architectures
1068 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1069 	 * masks event though they reject 46 bit masks.
1070 	 */
1071 	while (dma_mask > 0x7fffffffUL) {
1072 		if (pci_dma_supported(pci_dev, dma_mask) &&
1073 		    ((rc = pci_set_dma_mask(pci_dev, dma_mask)) == 0))
1074 			break;
1075 		dma_mask >>= 1;
1076 	}
1077 	if (rc) {
1078 		netif_err(efx, probe, efx->net_dev,
1079 			  "could not find a suitable DMA mask\n");
1080 		goto fail2;
1081 	}
1082 	netif_dbg(efx, probe, efx->net_dev,
1083 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1084 	rc = pci_set_consistent_dma_mask(pci_dev, dma_mask);
1085 	if (rc) {
1086 		/* pci_set_consistent_dma_mask() is not *allowed* to
1087 		 * fail with a mask that pci_set_dma_mask() accepted,
1088 		 * but just in case...
1089 		 */
1090 		netif_err(efx, probe, efx->net_dev,
1091 			  "failed to set consistent DMA mask\n");
1092 		goto fail2;
1093 	}
1094 
1095 	efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1096 	rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1097 	if (rc) {
1098 		netif_err(efx, probe, efx->net_dev,
1099 			  "request for memory BAR failed\n");
1100 		rc = -EIO;
1101 		goto fail3;
1102 	}
1103 	efx->membase = ioremap_nocache(efx->membase_phys,
1104 				       efx->type->mem_map_size);
1105 	if (!efx->membase) {
1106 		netif_err(efx, probe, efx->net_dev,
1107 			  "could not map memory BAR at %llx+%x\n",
1108 			  (unsigned long long)efx->membase_phys,
1109 			  efx->type->mem_map_size);
1110 		rc = -ENOMEM;
1111 		goto fail4;
1112 	}
1113 	netif_dbg(efx, probe, efx->net_dev,
1114 		  "memory BAR at %llx+%x (virtual %p)\n",
1115 		  (unsigned long long)efx->membase_phys,
1116 		  efx->type->mem_map_size, efx->membase);
1117 
1118 	return 0;
1119 
1120  fail4:
1121 	pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1122  fail3:
1123 	efx->membase_phys = 0;
1124  fail2:
1125 	pci_disable_device(efx->pci_dev);
1126  fail1:
1127 	return rc;
1128 }
1129 
1130 static void efx_fini_io(struct efx_nic *efx)
1131 {
1132 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1133 
1134 	if (efx->membase) {
1135 		iounmap(efx->membase);
1136 		efx->membase = NULL;
1137 	}
1138 
1139 	if (efx->membase_phys) {
1140 		pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1141 		efx->membase_phys = 0;
1142 	}
1143 
1144 	pci_disable_device(efx->pci_dev);
1145 }
1146 
1147 /* Get number of channels wanted.  Each channel will have its own IRQ,
1148  * 1 RX queue and/or 2 TX queues. */
1149 static int efx_wanted_channels(void)
1150 {
1151 	cpumask_var_t core_mask;
1152 	int count;
1153 	int cpu;
1154 
1155 	if (rss_cpus)
1156 		return rss_cpus;
1157 
1158 	if (unlikely(!zalloc_cpumask_var(&core_mask, GFP_KERNEL))) {
1159 		printk(KERN_WARNING
1160 		       "sfc: RSS disabled due to allocation failure\n");
1161 		return 1;
1162 	}
1163 
1164 	count = 0;
1165 	for_each_online_cpu(cpu) {
1166 		if (!cpumask_test_cpu(cpu, core_mask)) {
1167 			++count;
1168 			cpumask_or(core_mask, core_mask,
1169 				   topology_core_cpumask(cpu));
1170 		}
1171 	}
1172 
1173 	free_cpumask_var(core_mask);
1174 	return count;
1175 }
1176 
1177 static int
1178 efx_init_rx_cpu_rmap(struct efx_nic *efx, struct msix_entry *xentries)
1179 {
1180 #ifdef CONFIG_RFS_ACCEL
1181 	int i, rc;
1182 
1183 	efx->net_dev->rx_cpu_rmap = alloc_irq_cpu_rmap(efx->n_rx_channels);
1184 	if (!efx->net_dev->rx_cpu_rmap)
1185 		return -ENOMEM;
1186 	for (i = 0; i < efx->n_rx_channels; i++) {
1187 		rc = irq_cpu_rmap_add(efx->net_dev->rx_cpu_rmap,
1188 				      xentries[i].vector);
1189 		if (rc) {
1190 			free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1191 			efx->net_dev->rx_cpu_rmap = NULL;
1192 			return rc;
1193 		}
1194 	}
1195 #endif
1196 	return 0;
1197 }
1198 
1199 /* Probe the number and type of interrupts we are able to obtain, and
1200  * the resulting numbers of channels and RX queues.
1201  */
1202 static int efx_probe_interrupts(struct efx_nic *efx)
1203 {
1204 	int max_channels =
1205 		min_t(int, efx->type->phys_addr_channels, EFX_MAX_CHANNELS);
1206 	int rc, i;
1207 
1208 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1209 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1210 		int n_channels;
1211 
1212 		n_channels = efx_wanted_channels();
1213 		if (separate_tx_channels)
1214 			n_channels *= 2;
1215 		n_channels = min(n_channels, max_channels);
1216 
1217 		for (i = 0; i < n_channels; i++)
1218 			xentries[i].entry = i;
1219 		rc = pci_enable_msix(efx->pci_dev, xentries, n_channels);
1220 		if (rc > 0) {
1221 			netif_err(efx, drv, efx->net_dev,
1222 				  "WARNING: Insufficient MSI-X vectors"
1223 				  " available (%d < %d).\n", rc, n_channels);
1224 			netif_err(efx, drv, efx->net_dev,
1225 				  "WARNING: Performance may be reduced.\n");
1226 			EFX_BUG_ON_PARANOID(rc >= n_channels);
1227 			n_channels = rc;
1228 			rc = pci_enable_msix(efx->pci_dev, xentries,
1229 					     n_channels);
1230 		}
1231 
1232 		if (rc == 0) {
1233 			efx->n_channels = n_channels;
1234 			if (separate_tx_channels) {
1235 				efx->n_tx_channels =
1236 					max(efx->n_channels / 2, 1U);
1237 				efx->n_rx_channels =
1238 					max(efx->n_channels -
1239 					    efx->n_tx_channels, 1U);
1240 			} else {
1241 				efx->n_tx_channels = efx->n_channels;
1242 				efx->n_rx_channels = efx->n_channels;
1243 			}
1244 			rc = efx_init_rx_cpu_rmap(efx, xentries);
1245 			if (rc) {
1246 				pci_disable_msix(efx->pci_dev);
1247 				return rc;
1248 			}
1249 			for (i = 0; i < n_channels; i++)
1250 				efx_get_channel(efx, i)->irq =
1251 					xentries[i].vector;
1252 		} else {
1253 			/* Fall back to single channel MSI */
1254 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1255 			netif_err(efx, drv, efx->net_dev,
1256 				  "could not enable MSI-X\n");
1257 		}
1258 	}
1259 
1260 	/* Try single interrupt MSI */
1261 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1262 		efx->n_channels = 1;
1263 		efx->n_rx_channels = 1;
1264 		efx->n_tx_channels = 1;
1265 		rc = pci_enable_msi(efx->pci_dev);
1266 		if (rc == 0) {
1267 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1268 		} else {
1269 			netif_err(efx, drv, efx->net_dev,
1270 				  "could not enable MSI\n");
1271 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1272 		}
1273 	}
1274 
1275 	/* Assume legacy interrupts */
1276 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1277 		efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1278 		efx->n_rx_channels = 1;
1279 		efx->n_tx_channels = 1;
1280 		efx->legacy_irq = efx->pci_dev->irq;
1281 	}
1282 
1283 	return 0;
1284 }
1285 
1286 static void efx_remove_interrupts(struct efx_nic *efx)
1287 {
1288 	struct efx_channel *channel;
1289 
1290 	/* Remove MSI/MSI-X interrupts */
1291 	efx_for_each_channel(channel, efx)
1292 		channel->irq = 0;
1293 	pci_disable_msi(efx->pci_dev);
1294 	pci_disable_msix(efx->pci_dev);
1295 
1296 	/* Remove legacy interrupt */
1297 	efx->legacy_irq = 0;
1298 }
1299 
1300 static void efx_set_channels(struct efx_nic *efx)
1301 {
1302 	struct efx_channel *channel;
1303 	struct efx_tx_queue *tx_queue;
1304 
1305 	efx->tx_channel_offset =
1306 		separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1307 
1308 	/* We need to adjust the TX queue numbers if we have separate
1309 	 * RX-only and TX-only channels.
1310 	 */
1311 	efx_for_each_channel(channel, efx) {
1312 		efx_for_each_channel_tx_queue(tx_queue, channel)
1313 			tx_queue->queue -= (efx->tx_channel_offset *
1314 					    EFX_TXQ_TYPES);
1315 	}
1316 }
1317 
1318 static int efx_probe_nic(struct efx_nic *efx)
1319 {
1320 	size_t i;
1321 	int rc;
1322 
1323 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1324 
1325 	/* Carry out hardware-type specific initialisation */
1326 	rc = efx->type->probe(efx);
1327 	if (rc)
1328 		return rc;
1329 
1330 	/* Determine the number of channels and queues by trying to hook
1331 	 * in MSI-X interrupts. */
1332 	rc = efx_probe_interrupts(efx);
1333 	if (rc)
1334 		goto fail;
1335 
1336 	if (efx->n_channels > 1)
1337 		get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1338 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1339 		efx->rx_indir_table[i] = i % efx->n_rx_channels;
1340 
1341 	efx_set_channels(efx);
1342 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1343 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1344 
1345 	/* Initialise the interrupt moderation settings */
1346 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1347 				true);
1348 
1349 	return 0;
1350 
1351 fail:
1352 	efx->type->remove(efx);
1353 	return rc;
1354 }
1355 
1356 static void efx_remove_nic(struct efx_nic *efx)
1357 {
1358 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1359 
1360 	efx_remove_interrupts(efx);
1361 	efx->type->remove(efx);
1362 }
1363 
1364 /**************************************************************************
1365  *
1366  * NIC startup/shutdown
1367  *
1368  *************************************************************************/
1369 
1370 static int efx_probe_all(struct efx_nic *efx)
1371 {
1372 	int rc;
1373 
1374 	rc = efx_probe_nic(efx);
1375 	if (rc) {
1376 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1377 		goto fail1;
1378 	}
1379 
1380 	rc = efx_probe_port(efx);
1381 	if (rc) {
1382 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1383 		goto fail2;
1384 	}
1385 
1386 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1387 	rc = efx_probe_channels(efx);
1388 	if (rc)
1389 		goto fail3;
1390 
1391 	rc = efx_probe_filters(efx);
1392 	if (rc) {
1393 		netif_err(efx, probe, efx->net_dev,
1394 			  "failed to create filter tables\n");
1395 		goto fail4;
1396 	}
1397 
1398 	return 0;
1399 
1400  fail4:
1401 	efx_remove_channels(efx);
1402  fail3:
1403 	efx_remove_port(efx);
1404  fail2:
1405 	efx_remove_nic(efx);
1406  fail1:
1407 	return rc;
1408 }
1409 
1410 /* Called after previous invocation(s) of efx_stop_all, restarts the
1411  * port, kernel transmit queue, NAPI processing and hardware interrupts,
1412  * and ensures that the port is scheduled to be reconfigured.
1413  * This function is safe to call multiple times when the NIC is in any
1414  * state. */
1415 static void efx_start_all(struct efx_nic *efx)
1416 {
1417 	struct efx_channel *channel;
1418 
1419 	EFX_ASSERT_RESET_SERIALISED(efx);
1420 
1421 	/* Check that it is appropriate to restart the interface. All
1422 	 * of these flags are safe to read under just the rtnl lock */
1423 	if (efx->port_enabled)
1424 		return;
1425 	if ((efx->state != STATE_RUNNING) && (efx->state != STATE_INIT))
1426 		return;
1427 	if (efx_dev_registered(efx) && !netif_running(efx->net_dev))
1428 		return;
1429 
1430 	/* Mark the port as enabled so port reconfigurations can start, then
1431 	 * restart the transmit interface early so the watchdog timer stops */
1432 	efx_start_port(efx);
1433 
1434 	if (efx_dev_registered(efx) && netif_device_present(efx->net_dev))
1435 		netif_tx_wake_all_queues(efx->net_dev);
1436 
1437 	efx_for_each_channel(channel, efx)
1438 		efx_start_channel(channel);
1439 
1440 	if (efx->legacy_irq)
1441 		efx->legacy_irq_enabled = true;
1442 	efx_nic_enable_interrupts(efx);
1443 
1444 	/* Switch to event based MCDI completions after enabling interrupts.
1445 	 * If a reset has been scheduled, then we need to stay in polled mode.
1446 	 * Rather than serialising efx_mcdi_mode_event() [which sleeps] and
1447 	 * reset_pending [modified from an atomic context], we instead guarantee
1448 	 * that efx_mcdi_mode_poll() isn't reverted erroneously */
1449 	efx_mcdi_mode_event(efx);
1450 	if (efx->reset_pending)
1451 		efx_mcdi_mode_poll(efx);
1452 
1453 	/* Start the hardware monitor if there is one. Otherwise (we're link
1454 	 * event driven), we have to poll the PHY because after an event queue
1455 	 * flush, we could have a missed a link state change */
1456 	if (efx->type->monitor != NULL) {
1457 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1458 				   efx_monitor_interval);
1459 	} else {
1460 		mutex_lock(&efx->mac_lock);
1461 		if (efx->phy_op->poll(efx))
1462 			efx_link_status_changed(efx);
1463 		mutex_unlock(&efx->mac_lock);
1464 	}
1465 
1466 	efx->type->start_stats(efx);
1467 }
1468 
1469 /* Flush all delayed work. Should only be called when no more delayed work
1470  * will be scheduled. This doesn't flush pending online resets (efx_reset),
1471  * since we're holding the rtnl_lock at this point. */
1472 static void efx_flush_all(struct efx_nic *efx)
1473 {
1474 	/* Make sure the hardware monitor is stopped */
1475 	cancel_delayed_work_sync(&efx->monitor_work);
1476 	/* Stop scheduled port reconfigurations */
1477 	cancel_work_sync(&efx->mac_work);
1478 }
1479 
1480 /* Quiesce hardware and software without bringing the link down.
1481  * Safe to call multiple times, when the nic and interface is in any
1482  * state. The caller is guaranteed to subsequently be in a position
1483  * to modify any hardware and software state they see fit without
1484  * taking locks. */
1485 static void efx_stop_all(struct efx_nic *efx)
1486 {
1487 	struct efx_channel *channel;
1488 
1489 	EFX_ASSERT_RESET_SERIALISED(efx);
1490 
1491 	/* port_enabled can be read safely under the rtnl lock */
1492 	if (!efx->port_enabled)
1493 		return;
1494 
1495 	efx->type->stop_stats(efx);
1496 
1497 	/* Switch to MCDI polling on Siena before disabling interrupts */
1498 	efx_mcdi_mode_poll(efx);
1499 
1500 	/* Disable interrupts and wait for ISR to complete */
1501 	efx_nic_disable_interrupts(efx);
1502 	if (efx->legacy_irq) {
1503 		synchronize_irq(efx->legacy_irq);
1504 		efx->legacy_irq_enabled = false;
1505 	}
1506 	efx_for_each_channel(channel, efx) {
1507 		if (channel->irq)
1508 			synchronize_irq(channel->irq);
1509 	}
1510 
1511 	/* Stop all NAPI processing and synchronous rx refills */
1512 	efx_for_each_channel(channel, efx)
1513 		efx_stop_channel(channel);
1514 
1515 	/* Stop all asynchronous port reconfigurations. Since all
1516 	 * event processing has already been stopped, there is no
1517 	 * window to loose phy events */
1518 	efx_stop_port(efx);
1519 
1520 	/* Flush efx_mac_work(), refill_workqueue, monitor_work */
1521 	efx_flush_all(efx);
1522 
1523 	/* Stop the kernel transmit interface late, so the watchdog
1524 	 * timer isn't ticking over the flush */
1525 	if (efx_dev_registered(efx)) {
1526 		netif_tx_stop_all_queues(efx->net_dev);
1527 		netif_tx_lock_bh(efx->net_dev);
1528 		netif_tx_unlock_bh(efx->net_dev);
1529 	}
1530 }
1531 
1532 static void efx_remove_all(struct efx_nic *efx)
1533 {
1534 	efx_remove_filters(efx);
1535 	efx_remove_channels(efx);
1536 	efx_remove_port(efx);
1537 	efx_remove_nic(efx);
1538 }
1539 
1540 /**************************************************************************
1541  *
1542  * Interrupt moderation
1543  *
1544  **************************************************************************/
1545 
1546 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int resolution)
1547 {
1548 	if (usecs == 0)
1549 		return 0;
1550 	if (usecs < resolution)
1551 		return 1; /* never round down to 0 */
1552 	return usecs / resolution;
1553 }
1554 
1555 /* Set interrupt moderation parameters */
1556 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1557 			    unsigned int rx_usecs, bool rx_adaptive,
1558 			    bool rx_may_override_tx)
1559 {
1560 	struct efx_channel *channel;
1561 	unsigned tx_ticks = irq_mod_ticks(tx_usecs, EFX_IRQ_MOD_RESOLUTION);
1562 	unsigned rx_ticks = irq_mod_ticks(rx_usecs, EFX_IRQ_MOD_RESOLUTION);
1563 
1564 	EFX_ASSERT_RESET_SERIALISED(efx);
1565 
1566 	if (tx_ticks > EFX_IRQ_MOD_MAX || rx_ticks > EFX_IRQ_MOD_MAX)
1567 		return -EINVAL;
1568 
1569 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1570 	    !rx_may_override_tx) {
1571 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1572 			  "RX and TX IRQ moderation must be equal\n");
1573 		return -EINVAL;
1574 	}
1575 
1576 	efx->irq_rx_adaptive = rx_adaptive;
1577 	efx->irq_rx_moderation = rx_ticks;
1578 	efx_for_each_channel(channel, efx) {
1579 		if (efx_channel_has_rx_queue(channel))
1580 			channel->irq_moderation = rx_ticks;
1581 		else if (efx_channel_has_tx_queues(channel))
1582 			channel->irq_moderation = tx_ticks;
1583 	}
1584 
1585 	return 0;
1586 }
1587 
1588 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1589 			    unsigned int *rx_usecs, bool *rx_adaptive)
1590 {
1591 	*rx_adaptive = efx->irq_rx_adaptive;
1592 	*rx_usecs = efx->irq_rx_moderation * EFX_IRQ_MOD_RESOLUTION;
1593 
1594 	/* If channels are shared between RX and TX, so is IRQ
1595 	 * moderation.  Otherwise, IRQ moderation is the same for all
1596 	 * TX channels and is not adaptive.
1597 	 */
1598 	if (efx->tx_channel_offset == 0)
1599 		*tx_usecs = *rx_usecs;
1600 	else
1601 		*tx_usecs =
1602 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1603 			EFX_IRQ_MOD_RESOLUTION;
1604 }
1605 
1606 /**************************************************************************
1607  *
1608  * Hardware monitor
1609  *
1610  **************************************************************************/
1611 
1612 /* Run periodically off the general workqueue */
1613 static void efx_monitor(struct work_struct *data)
1614 {
1615 	struct efx_nic *efx = container_of(data, struct efx_nic,
1616 					   monitor_work.work);
1617 
1618 	netif_vdbg(efx, timer, efx->net_dev,
1619 		   "hardware monitor executing on CPU %d\n",
1620 		   raw_smp_processor_id());
1621 	BUG_ON(efx->type->monitor == NULL);
1622 
1623 	/* If the mac_lock is already held then it is likely a port
1624 	 * reconfiguration is already in place, which will likely do
1625 	 * most of the work of monitor() anyway. */
1626 	if (mutex_trylock(&efx->mac_lock)) {
1627 		if (efx->port_enabled)
1628 			efx->type->monitor(efx);
1629 		mutex_unlock(&efx->mac_lock);
1630 	}
1631 
1632 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1633 			   efx_monitor_interval);
1634 }
1635 
1636 /**************************************************************************
1637  *
1638  * ioctls
1639  *
1640  *************************************************************************/
1641 
1642 /* Net device ioctl
1643  * Context: process, rtnl_lock() held.
1644  */
1645 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1646 {
1647 	struct efx_nic *efx = netdev_priv(net_dev);
1648 	struct mii_ioctl_data *data = if_mii(ifr);
1649 
1650 	EFX_ASSERT_RESET_SERIALISED(efx);
1651 
1652 	/* Convert phy_id from older PRTAD/DEVAD format */
1653 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1654 	    (data->phy_id & 0xfc00) == 0x0400)
1655 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1656 
1657 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
1658 }
1659 
1660 /**************************************************************************
1661  *
1662  * NAPI interface
1663  *
1664  **************************************************************************/
1665 
1666 static void efx_init_napi(struct efx_nic *efx)
1667 {
1668 	struct efx_channel *channel;
1669 
1670 	efx_for_each_channel(channel, efx) {
1671 		channel->napi_dev = efx->net_dev;
1672 		netif_napi_add(channel->napi_dev, &channel->napi_str,
1673 			       efx_poll, napi_weight);
1674 	}
1675 }
1676 
1677 static void efx_fini_napi_channel(struct efx_channel *channel)
1678 {
1679 	if (channel->napi_dev)
1680 		netif_napi_del(&channel->napi_str);
1681 	channel->napi_dev = NULL;
1682 }
1683 
1684 static void efx_fini_napi(struct efx_nic *efx)
1685 {
1686 	struct efx_channel *channel;
1687 
1688 	efx_for_each_channel(channel, efx)
1689 		efx_fini_napi_channel(channel);
1690 }
1691 
1692 /**************************************************************************
1693  *
1694  * Kernel netpoll interface
1695  *
1696  *************************************************************************/
1697 
1698 #ifdef CONFIG_NET_POLL_CONTROLLER
1699 
1700 /* Although in the common case interrupts will be disabled, this is not
1701  * guaranteed. However, all our work happens inside the NAPI callback,
1702  * so no locking is required.
1703  */
1704 static void efx_netpoll(struct net_device *net_dev)
1705 {
1706 	struct efx_nic *efx = netdev_priv(net_dev);
1707 	struct efx_channel *channel;
1708 
1709 	efx_for_each_channel(channel, efx)
1710 		efx_schedule_channel(channel);
1711 }
1712 
1713 #endif
1714 
1715 /**************************************************************************
1716  *
1717  * Kernel net device interface
1718  *
1719  *************************************************************************/
1720 
1721 /* Context: process, rtnl_lock() held. */
1722 static int efx_net_open(struct net_device *net_dev)
1723 {
1724 	struct efx_nic *efx = netdev_priv(net_dev);
1725 	EFX_ASSERT_RESET_SERIALISED(efx);
1726 
1727 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
1728 		  raw_smp_processor_id());
1729 
1730 	if (efx->state == STATE_DISABLED)
1731 		return -EIO;
1732 	if (efx->phy_mode & PHY_MODE_SPECIAL)
1733 		return -EBUSY;
1734 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
1735 		return -EIO;
1736 
1737 	/* Notify the kernel of the link state polled during driver load,
1738 	 * before the monitor starts running */
1739 	efx_link_status_changed(efx);
1740 
1741 	efx_start_all(efx);
1742 	return 0;
1743 }
1744 
1745 /* Context: process, rtnl_lock() held.
1746  * Note that the kernel will ignore our return code; this method
1747  * should really be a void.
1748  */
1749 static int efx_net_stop(struct net_device *net_dev)
1750 {
1751 	struct efx_nic *efx = netdev_priv(net_dev);
1752 
1753 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
1754 		  raw_smp_processor_id());
1755 
1756 	if (efx->state != STATE_DISABLED) {
1757 		/* Stop the device and flush all the channels */
1758 		efx_stop_all(efx);
1759 		efx_fini_channels(efx);
1760 		efx_init_channels(efx);
1761 	}
1762 
1763 	return 0;
1764 }
1765 
1766 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1767 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev, struct rtnl_link_stats64 *stats)
1768 {
1769 	struct efx_nic *efx = netdev_priv(net_dev);
1770 	struct efx_mac_stats *mac_stats = &efx->mac_stats;
1771 
1772 	spin_lock_bh(&efx->stats_lock);
1773 	efx->type->update_stats(efx);
1774 	spin_unlock_bh(&efx->stats_lock);
1775 
1776 	stats->rx_packets = mac_stats->rx_packets;
1777 	stats->tx_packets = mac_stats->tx_packets;
1778 	stats->rx_bytes = mac_stats->rx_bytes;
1779 	stats->tx_bytes = mac_stats->tx_bytes;
1780 	stats->rx_dropped = efx->n_rx_nodesc_drop_cnt;
1781 	stats->multicast = mac_stats->rx_multicast;
1782 	stats->collisions = mac_stats->tx_collision;
1783 	stats->rx_length_errors = (mac_stats->rx_gtjumbo +
1784 				   mac_stats->rx_length_error);
1785 	stats->rx_crc_errors = mac_stats->rx_bad;
1786 	stats->rx_frame_errors = mac_stats->rx_align_error;
1787 	stats->rx_fifo_errors = mac_stats->rx_overflow;
1788 	stats->rx_missed_errors = mac_stats->rx_missed;
1789 	stats->tx_window_errors = mac_stats->tx_late_collision;
1790 
1791 	stats->rx_errors = (stats->rx_length_errors +
1792 			    stats->rx_crc_errors +
1793 			    stats->rx_frame_errors +
1794 			    mac_stats->rx_symbol_error);
1795 	stats->tx_errors = (stats->tx_window_errors +
1796 			    mac_stats->tx_bad);
1797 
1798 	return stats;
1799 }
1800 
1801 /* Context: netif_tx_lock held, BHs disabled. */
1802 static void efx_watchdog(struct net_device *net_dev)
1803 {
1804 	struct efx_nic *efx = netdev_priv(net_dev);
1805 
1806 	netif_err(efx, tx_err, efx->net_dev,
1807 		  "TX stuck with port_enabled=%d: resetting channels\n",
1808 		  efx->port_enabled);
1809 
1810 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
1811 }
1812 
1813 
1814 /* Context: process, rtnl_lock() held. */
1815 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
1816 {
1817 	struct efx_nic *efx = netdev_priv(net_dev);
1818 	int rc = 0;
1819 
1820 	EFX_ASSERT_RESET_SERIALISED(efx);
1821 
1822 	if (new_mtu > EFX_MAX_MTU)
1823 		return -EINVAL;
1824 
1825 	efx_stop_all(efx);
1826 
1827 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
1828 
1829 	efx_fini_channels(efx);
1830 
1831 	mutex_lock(&efx->mac_lock);
1832 	/* Reconfigure the MAC before enabling the dma queues so that
1833 	 * the RX buffers don't overflow */
1834 	net_dev->mtu = new_mtu;
1835 	efx->mac_op->reconfigure(efx);
1836 	mutex_unlock(&efx->mac_lock);
1837 
1838 	efx_init_channels(efx);
1839 
1840 	efx_start_all(efx);
1841 	return rc;
1842 }
1843 
1844 static int efx_set_mac_address(struct net_device *net_dev, void *data)
1845 {
1846 	struct efx_nic *efx = netdev_priv(net_dev);
1847 	struct sockaddr *addr = data;
1848 	char *new_addr = addr->sa_data;
1849 
1850 	EFX_ASSERT_RESET_SERIALISED(efx);
1851 
1852 	if (!is_valid_ether_addr(new_addr)) {
1853 		netif_err(efx, drv, efx->net_dev,
1854 			  "invalid ethernet MAC address requested: %pM\n",
1855 			  new_addr);
1856 		return -EINVAL;
1857 	}
1858 
1859 	memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
1860 
1861 	/* Reconfigure the MAC */
1862 	mutex_lock(&efx->mac_lock);
1863 	efx->mac_op->reconfigure(efx);
1864 	mutex_unlock(&efx->mac_lock);
1865 
1866 	return 0;
1867 }
1868 
1869 /* Context: netif_addr_lock held, BHs disabled. */
1870 static void efx_set_multicast_list(struct net_device *net_dev)
1871 {
1872 	struct efx_nic *efx = netdev_priv(net_dev);
1873 	struct netdev_hw_addr *ha;
1874 	union efx_multicast_hash *mc_hash = &efx->multicast_hash;
1875 	u32 crc;
1876 	int bit;
1877 
1878 	efx->promiscuous = !!(net_dev->flags & IFF_PROMISC);
1879 
1880 	/* Build multicast hash table */
1881 	if (efx->promiscuous || (net_dev->flags & IFF_ALLMULTI)) {
1882 		memset(mc_hash, 0xff, sizeof(*mc_hash));
1883 	} else {
1884 		memset(mc_hash, 0x00, sizeof(*mc_hash));
1885 		netdev_for_each_mc_addr(ha, net_dev) {
1886 			crc = ether_crc_le(ETH_ALEN, ha->addr);
1887 			bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
1888 			set_bit_le(bit, mc_hash->byte);
1889 		}
1890 
1891 		/* Broadcast packets go through the multicast hash filter.
1892 		 * ether_crc_le() of the broadcast address is 0xbe2612ff
1893 		 * so we always add bit 0xff to the mask.
1894 		 */
1895 		set_bit_le(0xff, mc_hash->byte);
1896 	}
1897 
1898 	if (efx->port_enabled)
1899 		queue_work(efx->workqueue, &efx->mac_work);
1900 	/* Otherwise efx_start_port() will do this */
1901 }
1902 
1903 static int efx_set_features(struct net_device *net_dev, u32 data)
1904 {
1905 	struct efx_nic *efx = netdev_priv(net_dev);
1906 
1907 	/* If disabling RX n-tuple filtering, clear existing filters */
1908 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
1909 		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
1910 
1911 	return 0;
1912 }
1913 
1914 static const struct net_device_ops efx_netdev_ops = {
1915 	.ndo_open		= efx_net_open,
1916 	.ndo_stop		= efx_net_stop,
1917 	.ndo_get_stats64	= efx_net_stats,
1918 	.ndo_tx_timeout		= efx_watchdog,
1919 	.ndo_start_xmit		= efx_hard_start_xmit,
1920 	.ndo_validate_addr	= eth_validate_addr,
1921 	.ndo_do_ioctl		= efx_ioctl,
1922 	.ndo_change_mtu		= efx_change_mtu,
1923 	.ndo_set_mac_address	= efx_set_mac_address,
1924 	.ndo_set_rx_mode	= efx_set_multicast_list,
1925 	.ndo_set_features	= efx_set_features,
1926 #ifdef CONFIG_NET_POLL_CONTROLLER
1927 	.ndo_poll_controller = efx_netpoll,
1928 #endif
1929 	.ndo_setup_tc		= efx_setup_tc,
1930 #ifdef CONFIG_RFS_ACCEL
1931 	.ndo_rx_flow_steer	= efx_filter_rfs,
1932 #endif
1933 };
1934 
1935 static void efx_update_name(struct efx_nic *efx)
1936 {
1937 	strcpy(efx->name, efx->net_dev->name);
1938 	efx_mtd_rename(efx);
1939 	efx_set_channel_names(efx);
1940 }
1941 
1942 static int efx_netdev_event(struct notifier_block *this,
1943 			    unsigned long event, void *ptr)
1944 {
1945 	struct net_device *net_dev = ptr;
1946 
1947 	if (net_dev->netdev_ops == &efx_netdev_ops &&
1948 	    event == NETDEV_CHANGENAME)
1949 		efx_update_name(netdev_priv(net_dev));
1950 
1951 	return NOTIFY_DONE;
1952 }
1953 
1954 static struct notifier_block efx_netdev_notifier = {
1955 	.notifier_call = efx_netdev_event,
1956 };
1957 
1958 static ssize_t
1959 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
1960 {
1961 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
1962 	return sprintf(buf, "%d\n", efx->phy_type);
1963 }
1964 static DEVICE_ATTR(phy_type, 0644, show_phy_type, NULL);
1965 
1966 static int efx_register_netdev(struct efx_nic *efx)
1967 {
1968 	struct net_device *net_dev = efx->net_dev;
1969 	struct efx_channel *channel;
1970 	int rc;
1971 
1972 	net_dev->watchdog_timeo = 5 * HZ;
1973 	net_dev->irq = efx->pci_dev->irq;
1974 	net_dev->netdev_ops = &efx_netdev_ops;
1975 	SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
1976 
1977 	/* Clear MAC statistics */
1978 	efx->mac_op->update_stats(efx);
1979 	memset(&efx->mac_stats, 0, sizeof(efx->mac_stats));
1980 
1981 	rtnl_lock();
1982 
1983 	rc = dev_alloc_name(net_dev, net_dev->name);
1984 	if (rc < 0)
1985 		goto fail_locked;
1986 	efx_update_name(efx);
1987 
1988 	rc = register_netdevice(net_dev);
1989 	if (rc)
1990 		goto fail_locked;
1991 
1992 	efx_for_each_channel(channel, efx) {
1993 		struct efx_tx_queue *tx_queue;
1994 		efx_for_each_channel_tx_queue(tx_queue, channel)
1995 			efx_init_tx_queue_core_txq(tx_queue);
1996 	}
1997 
1998 	/* Always start with carrier off; PHY events will detect the link */
1999 	netif_carrier_off(efx->net_dev);
2000 
2001 	rtnl_unlock();
2002 
2003 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2004 	if (rc) {
2005 		netif_err(efx, drv, efx->net_dev,
2006 			  "failed to init net dev attributes\n");
2007 		goto fail_registered;
2008 	}
2009 
2010 	return 0;
2011 
2012 fail_locked:
2013 	rtnl_unlock();
2014 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2015 	return rc;
2016 
2017 fail_registered:
2018 	unregister_netdev(net_dev);
2019 	return rc;
2020 }
2021 
2022 static void efx_unregister_netdev(struct efx_nic *efx)
2023 {
2024 	struct efx_channel *channel;
2025 	struct efx_tx_queue *tx_queue;
2026 
2027 	if (!efx->net_dev)
2028 		return;
2029 
2030 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2031 
2032 	/* Free up any skbs still remaining. This has to happen before
2033 	 * we try to unregister the netdev as running their destructors
2034 	 * may be needed to get the device ref. count to 0. */
2035 	efx_for_each_channel(channel, efx) {
2036 		efx_for_each_channel_tx_queue(tx_queue, channel)
2037 			efx_release_tx_buffers(tx_queue);
2038 	}
2039 
2040 	if (efx_dev_registered(efx)) {
2041 		strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2042 		device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2043 		unregister_netdev(efx->net_dev);
2044 	}
2045 }
2046 
2047 /**************************************************************************
2048  *
2049  * Device reset and suspend
2050  *
2051  **************************************************************************/
2052 
2053 /* Tears down the entire software state and most of the hardware state
2054  * before reset.  */
2055 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2056 {
2057 	EFX_ASSERT_RESET_SERIALISED(efx);
2058 
2059 	efx_stop_all(efx);
2060 	mutex_lock(&efx->mac_lock);
2061 
2062 	efx_fini_channels(efx);
2063 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2064 		efx->phy_op->fini(efx);
2065 	efx->type->fini(efx);
2066 }
2067 
2068 /* This function will always ensure that the locks acquired in
2069  * efx_reset_down() are released. A failure return code indicates
2070  * that we were unable to reinitialise the hardware, and the
2071  * driver should be disabled. If ok is false, then the rx and tx
2072  * engines are not restarted, pending a RESET_DISABLE. */
2073 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2074 {
2075 	int rc;
2076 
2077 	EFX_ASSERT_RESET_SERIALISED(efx);
2078 
2079 	rc = efx->type->init(efx);
2080 	if (rc) {
2081 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2082 		goto fail;
2083 	}
2084 
2085 	if (!ok)
2086 		goto fail;
2087 
2088 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2089 		rc = efx->phy_op->init(efx);
2090 		if (rc)
2091 			goto fail;
2092 		if (efx->phy_op->reconfigure(efx))
2093 			netif_err(efx, drv, efx->net_dev,
2094 				  "could not restore PHY settings\n");
2095 	}
2096 
2097 	efx->mac_op->reconfigure(efx);
2098 
2099 	efx_init_channels(efx);
2100 	efx_restore_filters(efx);
2101 
2102 	mutex_unlock(&efx->mac_lock);
2103 
2104 	efx_start_all(efx);
2105 
2106 	return 0;
2107 
2108 fail:
2109 	efx->port_initialized = false;
2110 
2111 	mutex_unlock(&efx->mac_lock);
2112 
2113 	return rc;
2114 }
2115 
2116 /* Reset the NIC using the specified method.  Note that the reset may
2117  * fail, in which case the card will be left in an unusable state.
2118  *
2119  * Caller must hold the rtnl_lock.
2120  */
2121 int efx_reset(struct efx_nic *efx, enum reset_type method)
2122 {
2123 	int rc, rc2;
2124 	bool disabled;
2125 
2126 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2127 		   RESET_TYPE(method));
2128 
2129 	netif_device_detach(efx->net_dev);
2130 	efx_reset_down(efx, method);
2131 
2132 	rc = efx->type->reset(efx, method);
2133 	if (rc) {
2134 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2135 		goto out;
2136 	}
2137 
2138 	/* Clear flags for the scopes we covered.  We assume the NIC and
2139 	 * driver are now quiescent so that there is no race here.
2140 	 */
2141 	efx->reset_pending &= -(1 << (method + 1));
2142 
2143 	/* Reinitialise bus-mastering, which may have been turned off before
2144 	 * the reset was scheduled. This is still appropriate, even in the
2145 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2146 	 * can respond to requests. */
2147 	pci_set_master(efx->pci_dev);
2148 
2149 out:
2150 	/* Leave device stopped if necessary */
2151 	disabled = rc || method == RESET_TYPE_DISABLE;
2152 	rc2 = efx_reset_up(efx, method, !disabled);
2153 	if (rc2) {
2154 		disabled = true;
2155 		if (!rc)
2156 			rc = rc2;
2157 	}
2158 
2159 	if (disabled) {
2160 		dev_close(efx->net_dev);
2161 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2162 		efx->state = STATE_DISABLED;
2163 	} else {
2164 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2165 		netif_device_attach(efx->net_dev);
2166 	}
2167 	return rc;
2168 }
2169 
2170 /* The worker thread exists so that code that cannot sleep can
2171  * schedule a reset for later.
2172  */
2173 static void efx_reset_work(struct work_struct *data)
2174 {
2175 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2176 	unsigned long pending = ACCESS_ONCE(efx->reset_pending);
2177 
2178 	if (!pending)
2179 		return;
2180 
2181 	/* If we're not RUNNING then don't reset. Leave the reset_pending
2182 	 * flags set so that efx_pci_probe_main will be retried */
2183 	if (efx->state != STATE_RUNNING) {
2184 		netif_info(efx, drv, efx->net_dev,
2185 			   "scheduled reset quenched. NIC not RUNNING\n");
2186 		return;
2187 	}
2188 
2189 	rtnl_lock();
2190 	(void)efx_reset(efx, fls(pending) - 1);
2191 	rtnl_unlock();
2192 }
2193 
2194 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2195 {
2196 	enum reset_type method;
2197 
2198 	switch (type) {
2199 	case RESET_TYPE_INVISIBLE:
2200 	case RESET_TYPE_ALL:
2201 	case RESET_TYPE_WORLD:
2202 	case RESET_TYPE_DISABLE:
2203 		method = type;
2204 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2205 			  RESET_TYPE(method));
2206 		break;
2207 	default:
2208 		method = efx->type->map_reset_reason(type);
2209 		netif_dbg(efx, drv, efx->net_dev,
2210 			  "scheduling %s reset for %s\n",
2211 			  RESET_TYPE(method), RESET_TYPE(type));
2212 		break;
2213 	}
2214 
2215 	set_bit(method, &efx->reset_pending);
2216 
2217 	/* efx_process_channel() will no longer read events once a
2218 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2219 	efx_mcdi_mode_poll(efx);
2220 
2221 	queue_work(reset_workqueue, &efx->reset_work);
2222 }
2223 
2224 /**************************************************************************
2225  *
2226  * List of NICs we support
2227  *
2228  **************************************************************************/
2229 
2230 /* PCI device ID table */
2231 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2232 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2233 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2234 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2235 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2236 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2237 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2238 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, BETHPAGE_A_P_DEVID),
2239 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2240 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, SIENA_A_P_DEVID),
2241 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2242 	{0}			/* end of list */
2243 };
2244 
2245 /**************************************************************************
2246  *
2247  * Dummy PHY/MAC operations
2248  *
2249  * Can be used for some unimplemented operations
2250  * Needed so all function pointers are valid and do not have to be tested
2251  * before use
2252  *
2253  **************************************************************************/
2254 int efx_port_dummy_op_int(struct efx_nic *efx)
2255 {
2256 	return 0;
2257 }
2258 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2259 
2260 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2261 {
2262 	return false;
2263 }
2264 
2265 static const struct efx_phy_operations efx_dummy_phy_operations = {
2266 	.init		 = efx_port_dummy_op_int,
2267 	.reconfigure	 = efx_port_dummy_op_int,
2268 	.poll		 = efx_port_dummy_op_poll,
2269 	.fini		 = efx_port_dummy_op_void,
2270 };
2271 
2272 /**************************************************************************
2273  *
2274  * Data housekeeping
2275  *
2276  **************************************************************************/
2277 
2278 /* This zeroes out and then fills in the invariants in a struct
2279  * efx_nic (including all sub-structures).
2280  */
2281 static int efx_init_struct(struct efx_nic *efx, const struct efx_nic_type *type,
2282 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2283 {
2284 	int i;
2285 
2286 	/* Initialise common structures */
2287 	memset(efx, 0, sizeof(*efx));
2288 	spin_lock_init(&efx->biu_lock);
2289 #ifdef CONFIG_SFC_MTD
2290 	INIT_LIST_HEAD(&efx->mtd_list);
2291 #endif
2292 	INIT_WORK(&efx->reset_work, efx_reset_work);
2293 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2294 	efx->pci_dev = pci_dev;
2295 	efx->msg_enable = debug;
2296 	efx->state = STATE_INIT;
2297 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2298 
2299 	efx->net_dev = net_dev;
2300 	spin_lock_init(&efx->stats_lock);
2301 	mutex_init(&efx->mac_lock);
2302 	efx->mac_op = type->default_mac_ops;
2303 	efx->phy_op = &efx_dummy_phy_operations;
2304 	efx->mdio.dev = net_dev;
2305 	INIT_WORK(&efx->mac_work, efx_mac_work);
2306 
2307 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2308 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2309 		if (!efx->channel[i])
2310 			goto fail;
2311 	}
2312 
2313 	efx->type = type;
2314 
2315 	EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS);
2316 
2317 	/* Higher numbered interrupt modes are less capable! */
2318 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2319 				  interrupt_mode);
2320 
2321 	/* Would be good to use the net_dev name, but we're too early */
2322 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2323 		 pci_name(pci_dev));
2324 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2325 	if (!efx->workqueue)
2326 		goto fail;
2327 
2328 	return 0;
2329 
2330 fail:
2331 	efx_fini_struct(efx);
2332 	return -ENOMEM;
2333 }
2334 
2335 static void efx_fini_struct(struct efx_nic *efx)
2336 {
2337 	int i;
2338 
2339 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2340 		kfree(efx->channel[i]);
2341 
2342 	if (efx->workqueue) {
2343 		destroy_workqueue(efx->workqueue);
2344 		efx->workqueue = NULL;
2345 	}
2346 }
2347 
2348 /**************************************************************************
2349  *
2350  * PCI interface
2351  *
2352  **************************************************************************/
2353 
2354 /* Main body of final NIC shutdown code
2355  * This is called only at module unload (or hotplug removal).
2356  */
2357 static void efx_pci_remove_main(struct efx_nic *efx)
2358 {
2359 #ifdef CONFIG_RFS_ACCEL
2360 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
2361 	efx->net_dev->rx_cpu_rmap = NULL;
2362 #endif
2363 	efx_nic_fini_interrupt(efx);
2364 	efx_fini_channels(efx);
2365 	efx_fini_port(efx);
2366 	efx->type->fini(efx);
2367 	efx_fini_napi(efx);
2368 	efx_remove_all(efx);
2369 }
2370 
2371 /* Final NIC shutdown
2372  * This is called only at module unload (or hotplug removal).
2373  */
2374 static void efx_pci_remove(struct pci_dev *pci_dev)
2375 {
2376 	struct efx_nic *efx;
2377 
2378 	efx = pci_get_drvdata(pci_dev);
2379 	if (!efx)
2380 		return;
2381 
2382 	/* Mark the NIC as fini, then stop the interface */
2383 	rtnl_lock();
2384 	efx->state = STATE_FINI;
2385 	dev_close(efx->net_dev);
2386 
2387 	/* Allow any queued efx_resets() to complete */
2388 	rtnl_unlock();
2389 
2390 	efx_unregister_netdev(efx);
2391 
2392 	efx_mtd_remove(efx);
2393 
2394 	/* Wait for any scheduled resets to complete. No more will be
2395 	 * scheduled from this point because efx_stop_all() has been
2396 	 * called, we are no longer registered with driverlink, and
2397 	 * the net_device's have been removed. */
2398 	cancel_work_sync(&efx->reset_work);
2399 
2400 	efx_pci_remove_main(efx);
2401 
2402 	efx_fini_io(efx);
2403 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2404 
2405 	pci_set_drvdata(pci_dev, NULL);
2406 	efx_fini_struct(efx);
2407 	free_netdev(efx->net_dev);
2408 };
2409 
2410 /* Main body of NIC initialisation
2411  * This is called at module load (or hotplug insertion, theoretically).
2412  */
2413 static int efx_pci_probe_main(struct efx_nic *efx)
2414 {
2415 	int rc;
2416 
2417 	/* Do start-of-day initialisation */
2418 	rc = efx_probe_all(efx);
2419 	if (rc)
2420 		goto fail1;
2421 
2422 	efx_init_napi(efx);
2423 
2424 	rc = efx->type->init(efx);
2425 	if (rc) {
2426 		netif_err(efx, probe, efx->net_dev,
2427 			  "failed to initialise NIC\n");
2428 		goto fail3;
2429 	}
2430 
2431 	rc = efx_init_port(efx);
2432 	if (rc) {
2433 		netif_err(efx, probe, efx->net_dev,
2434 			  "failed to initialise port\n");
2435 		goto fail4;
2436 	}
2437 
2438 	efx_init_channels(efx);
2439 
2440 	rc = efx_nic_init_interrupt(efx);
2441 	if (rc)
2442 		goto fail5;
2443 
2444 	return 0;
2445 
2446  fail5:
2447 	efx_fini_channels(efx);
2448 	efx_fini_port(efx);
2449  fail4:
2450 	efx->type->fini(efx);
2451  fail3:
2452 	efx_fini_napi(efx);
2453 	efx_remove_all(efx);
2454  fail1:
2455 	return rc;
2456 }
2457 
2458 /* NIC initialisation
2459  *
2460  * This is called at module load (or hotplug insertion,
2461  * theoretically).  It sets up PCI mappings, tests and resets the NIC,
2462  * sets up and registers the network devices with the kernel and hooks
2463  * the interrupt service routine.  It does not prepare the device for
2464  * transmission; this is left to the first time one of the network
2465  * interfaces is brought up (i.e. efx_net_open).
2466  */
2467 static int __devinit efx_pci_probe(struct pci_dev *pci_dev,
2468 				   const struct pci_device_id *entry)
2469 {
2470 	const struct efx_nic_type *type = (const struct efx_nic_type *) entry->driver_data;
2471 	struct net_device *net_dev;
2472 	struct efx_nic *efx;
2473 	int i, rc;
2474 
2475 	/* Allocate and initialise a struct net_device and struct efx_nic */
2476 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2477 				     EFX_MAX_RX_QUEUES);
2478 	if (!net_dev)
2479 		return -ENOMEM;
2480 	net_dev->features |= (type->offload_features | NETIF_F_SG |
2481 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
2482 			      NETIF_F_RXCSUM);
2483 	if (type->offload_features & NETIF_F_V6_CSUM)
2484 		net_dev->features |= NETIF_F_TSO6;
2485 	/* Mask for features that also apply to VLAN devices */
2486 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2487 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2488 				   NETIF_F_RXCSUM);
2489 	/* All offloads can be toggled */
2490 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2491 	efx = netdev_priv(net_dev);
2492 	pci_set_drvdata(pci_dev, efx);
2493 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2494 	rc = efx_init_struct(efx, type, pci_dev, net_dev);
2495 	if (rc)
2496 		goto fail1;
2497 
2498 	netif_info(efx, probe, efx->net_dev,
2499 		   "Solarflare NIC detected\n");
2500 
2501 	/* Set up basic I/O (BAR mappings etc) */
2502 	rc = efx_init_io(efx);
2503 	if (rc)
2504 		goto fail2;
2505 
2506 	/* No serialisation is required with the reset path because
2507 	 * we're in STATE_INIT. */
2508 	for (i = 0; i < 5; i++) {
2509 		rc = efx_pci_probe_main(efx);
2510 
2511 		/* Serialise against efx_reset(). No more resets will be
2512 		 * scheduled since efx_stop_all() has been called, and we
2513 		 * have not and never have been registered with either
2514 		 * the rtnetlink or driverlink layers. */
2515 		cancel_work_sync(&efx->reset_work);
2516 
2517 		if (rc == 0) {
2518 			if (efx->reset_pending) {
2519 				/* If there was a scheduled reset during
2520 				 * probe, the NIC is probably hosed anyway */
2521 				efx_pci_remove_main(efx);
2522 				rc = -EIO;
2523 			} else {
2524 				break;
2525 			}
2526 		}
2527 
2528 		/* Retry if a recoverably reset event has been scheduled */
2529 		if (efx->reset_pending &
2530 		    ~(1 << RESET_TYPE_INVISIBLE | 1 << RESET_TYPE_ALL) ||
2531 		    !efx->reset_pending)
2532 			goto fail3;
2533 
2534 		efx->reset_pending = 0;
2535 	}
2536 
2537 	if (rc) {
2538 		netif_err(efx, probe, efx->net_dev, "Could not reset NIC\n");
2539 		goto fail4;
2540 	}
2541 
2542 	/* Switch to the running state before we expose the device to the OS,
2543 	 * so that dev_open()|efx_start_all() will actually start the device */
2544 	efx->state = STATE_RUNNING;
2545 
2546 	rc = efx_register_netdev(efx);
2547 	if (rc)
2548 		goto fail5;
2549 
2550 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2551 
2552 	rtnl_lock();
2553 	efx_mtd_probe(efx); /* allowed to fail */
2554 	rtnl_unlock();
2555 	return 0;
2556 
2557  fail5:
2558 	efx_pci_remove_main(efx);
2559  fail4:
2560  fail3:
2561 	efx_fini_io(efx);
2562  fail2:
2563 	efx_fini_struct(efx);
2564  fail1:
2565 	WARN_ON(rc > 0);
2566 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2567 	free_netdev(net_dev);
2568 	return rc;
2569 }
2570 
2571 static int efx_pm_freeze(struct device *dev)
2572 {
2573 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2574 
2575 	efx->state = STATE_FINI;
2576 
2577 	netif_device_detach(efx->net_dev);
2578 
2579 	efx_stop_all(efx);
2580 	efx_fini_channels(efx);
2581 
2582 	return 0;
2583 }
2584 
2585 static int efx_pm_thaw(struct device *dev)
2586 {
2587 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2588 
2589 	efx->state = STATE_INIT;
2590 
2591 	efx_init_channels(efx);
2592 
2593 	mutex_lock(&efx->mac_lock);
2594 	efx->phy_op->reconfigure(efx);
2595 	mutex_unlock(&efx->mac_lock);
2596 
2597 	efx_start_all(efx);
2598 
2599 	netif_device_attach(efx->net_dev);
2600 
2601 	efx->state = STATE_RUNNING;
2602 
2603 	efx->type->resume_wol(efx);
2604 
2605 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2606 	queue_work(reset_workqueue, &efx->reset_work);
2607 
2608 	return 0;
2609 }
2610 
2611 static int efx_pm_poweroff(struct device *dev)
2612 {
2613 	struct pci_dev *pci_dev = to_pci_dev(dev);
2614 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2615 
2616 	efx->type->fini(efx);
2617 
2618 	efx->reset_pending = 0;
2619 
2620 	pci_save_state(pci_dev);
2621 	return pci_set_power_state(pci_dev, PCI_D3hot);
2622 }
2623 
2624 /* Used for both resume and restore */
2625 static int efx_pm_resume(struct device *dev)
2626 {
2627 	struct pci_dev *pci_dev = to_pci_dev(dev);
2628 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2629 	int rc;
2630 
2631 	rc = pci_set_power_state(pci_dev, PCI_D0);
2632 	if (rc)
2633 		return rc;
2634 	pci_restore_state(pci_dev);
2635 	rc = pci_enable_device(pci_dev);
2636 	if (rc)
2637 		return rc;
2638 	pci_set_master(efx->pci_dev);
2639 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
2640 	if (rc)
2641 		return rc;
2642 	rc = efx->type->init(efx);
2643 	if (rc)
2644 		return rc;
2645 	efx_pm_thaw(dev);
2646 	return 0;
2647 }
2648 
2649 static int efx_pm_suspend(struct device *dev)
2650 {
2651 	int rc;
2652 
2653 	efx_pm_freeze(dev);
2654 	rc = efx_pm_poweroff(dev);
2655 	if (rc)
2656 		efx_pm_resume(dev);
2657 	return rc;
2658 }
2659 
2660 static struct dev_pm_ops efx_pm_ops = {
2661 	.suspend	= efx_pm_suspend,
2662 	.resume		= efx_pm_resume,
2663 	.freeze		= efx_pm_freeze,
2664 	.thaw		= efx_pm_thaw,
2665 	.poweroff	= efx_pm_poweroff,
2666 	.restore	= efx_pm_resume,
2667 };
2668 
2669 static struct pci_driver efx_pci_driver = {
2670 	.name		= KBUILD_MODNAME,
2671 	.id_table	= efx_pci_table,
2672 	.probe		= efx_pci_probe,
2673 	.remove		= efx_pci_remove,
2674 	.driver.pm	= &efx_pm_ops,
2675 };
2676 
2677 /**************************************************************************
2678  *
2679  * Kernel module interface
2680  *
2681  *************************************************************************/
2682 
2683 module_param(interrupt_mode, uint, 0444);
2684 MODULE_PARM_DESC(interrupt_mode,
2685 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2686 
2687 static int __init efx_init_module(void)
2688 {
2689 	int rc;
2690 
2691 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
2692 
2693 	rc = register_netdevice_notifier(&efx_netdev_notifier);
2694 	if (rc)
2695 		goto err_notifier;
2696 
2697 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
2698 	if (!reset_workqueue) {
2699 		rc = -ENOMEM;
2700 		goto err_reset;
2701 	}
2702 
2703 	rc = pci_register_driver(&efx_pci_driver);
2704 	if (rc < 0)
2705 		goto err_pci;
2706 
2707 	return 0;
2708 
2709  err_pci:
2710 	destroy_workqueue(reset_workqueue);
2711  err_reset:
2712 	unregister_netdevice_notifier(&efx_netdev_notifier);
2713  err_notifier:
2714 	return rc;
2715 }
2716 
2717 static void __exit efx_exit_module(void)
2718 {
2719 	printk(KERN_INFO "Solarflare NET driver unloading\n");
2720 
2721 	pci_unregister_driver(&efx_pci_driver);
2722 	destroy_workqueue(reset_workqueue);
2723 	unregister_netdevice_notifier(&efx_netdev_notifier);
2724 
2725 }
2726 
2727 module_init(efx_init_module);
2728 module_exit(efx_exit_module);
2729 
2730 MODULE_AUTHOR("Solarflare Communications and "
2731 	      "Michael Brown <mbrown@fensystems.co.uk>");
2732 MODULE_DESCRIPTION("Solarflare Communications network driver");
2733 MODULE_LICENSE("GPL");
2734 MODULE_DEVICE_TABLE(pci, efx_pci_table);
2735