xref: /linux/drivers/net/ethernet/sfc/efx.c (revision c75c5ab575af7db707689cdbb5a5c458e9a034bb)
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 #include "selftest.h"
29 
30 #include "mcdi.h"
31 #include "workarounds.h"
32 
33 /**************************************************************************
34  *
35  * Type name strings
36  *
37  **************************************************************************
38  */
39 
40 /* Loopback mode names (see LOOPBACK_MODE()) */
41 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
42 const char *const efx_loopback_mode_names[] = {
43 	[LOOPBACK_NONE]		= "NONE",
44 	[LOOPBACK_DATA]		= "DATAPATH",
45 	[LOOPBACK_GMAC]		= "GMAC",
46 	[LOOPBACK_XGMII]	= "XGMII",
47 	[LOOPBACK_XGXS]		= "XGXS",
48 	[LOOPBACK_XAUI]		= "XAUI",
49 	[LOOPBACK_GMII]		= "GMII",
50 	[LOOPBACK_SGMII]	= "SGMII",
51 	[LOOPBACK_XGBR]		= "XGBR",
52 	[LOOPBACK_XFI]		= "XFI",
53 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
54 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
55 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
56 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
57 	[LOOPBACK_GPHY]		= "GPHY",
58 	[LOOPBACK_PHYXS]	= "PHYXS",
59 	[LOOPBACK_PCS]		= "PCS",
60 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
61 	[LOOPBACK_XPORT]	= "XPORT",
62 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
63 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
64 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
65 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
66 	[LOOPBACK_GMII_WS]	= "GMII_WS",
67 	[LOOPBACK_XFI_WS]	= "XFI_WS",
68 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
69 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
70 };
71 
72 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
73 const char *const efx_reset_type_names[] = {
74 	[RESET_TYPE_INVISIBLE]     = "INVISIBLE",
75 	[RESET_TYPE_ALL]           = "ALL",
76 	[RESET_TYPE_WORLD]         = "WORLD",
77 	[RESET_TYPE_DISABLE]       = "DISABLE",
78 	[RESET_TYPE_TX_WATCHDOG]   = "TX_WATCHDOG",
79 	[RESET_TYPE_INT_ERROR]     = "INT_ERROR",
80 	[RESET_TYPE_RX_RECOVERY]   = "RX_RECOVERY",
81 	[RESET_TYPE_RX_DESC_FETCH] = "RX_DESC_FETCH",
82 	[RESET_TYPE_TX_DESC_FETCH] = "TX_DESC_FETCH",
83 	[RESET_TYPE_TX_SKIP]       = "TX_SKIP",
84 	[RESET_TYPE_MC_FAILURE]    = "MC_FAILURE",
85 };
86 
87 #define EFX_MAX_MTU (9 * 1024)
88 
89 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
90  * queued onto this work queue. This is not a per-nic work queue, because
91  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
92  */
93 static struct workqueue_struct *reset_workqueue;
94 
95 /**************************************************************************
96  *
97  * Configurable values
98  *
99  *************************************************************************/
100 
101 /*
102  * Use separate channels for TX and RX events
103  *
104  * Set this to 1 to use separate channels for TX and RX. It allows us
105  * to control interrupt affinity separately for TX and RX.
106  *
107  * This is only used in MSI-X interrupt mode
108  */
109 static bool separate_tx_channels;
110 module_param(separate_tx_channels, bool, 0444);
111 MODULE_PARM_DESC(separate_tx_channels,
112 		 "Use separate channels for TX and RX");
113 
114 /* This is the weight assigned to each of the (per-channel) virtual
115  * NAPI devices.
116  */
117 static int napi_weight = 64;
118 
119 /* This is the time (in jiffies) between invocations of the hardware
120  * monitor.  On Falcon-based NICs, this will:
121  * - Check the on-board hardware monitor;
122  * - Poll the link state and reconfigure the hardware as necessary.
123  */
124 static unsigned int efx_monitor_interval = 1 * HZ;
125 
126 /* Initial interrupt moderation settings.  They can be modified after
127  * module load with ethtool.
128  *
129  * The default for RX should strike a balance between increasing the
130  * round-trip latency and reducing overhead.
131  */
132 static unsigned int rx_irq_mod_usec = 60;
133 
134 /* Initial interrupt moderation settings.  They can be modified after
135  * module load with ethtool.
136  *
137  * This default is chosen to ensure that a 10G link does not go idle
138  * while a TX queue is stopped after it has become full.  A queue is
139  * restarted when it drops below half full.  The time this takes (assuming
140  * worst case 3 descriptors per packet and 1024 descriptors) is
141  *   512 / 3 * 1.2 = 205 usec.
142  */
143 static unsigned int tx_irq_mod_usec = 150;
144 
145 /* This is the first interrupt mode to try out of:
146  * 0 => MSI-X
147  * 1 => MSI
148  * 2 => legacy
149  */
150 static unsigned int interrupt_mode;
151 
152 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
153  * i.e. the number of CPUs among which we may distribute simultaneous
154  * interrupt handling.
155  *
156  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
157  * The default (0) means to assign an interrupt to each core.
158  */
159 static unsigned int rss_cpus;
160 module_param(rss_cpus, uint, 0444);
161 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
162 
163 static bool phy_flash_cfg;
164 module_param(phy_flash_cfg, bool, 0644);
165 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
166 
167 static unsigned irq_adapt_low_thresh = 8000;
168 module_param(irq_adapt_low_thresh, uint, 0644);
169 MODULE_PARM_DESC(irq_adapt_low_thresh,
170 		 "Threshold score for reducing IRQ moderation");
171 
172 static unsigned irq_adapt_high_thresh = 16000;
173 module_param(irq_adapt_high_thresh, uint, 0644);
174 MODULE_PARM_DESC(irq_adapt_high_thresh,
175 		 "Threshold score for increasing IRQ moderation");
176 
177 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
178 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
179 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
180 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
181 module_param(debug, uint, 0);
182 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
183 
184 /**************************************************************************
185  *
186  * Utility functions and prototypes
187  *
188  *************************************************************************/
189 
190 static void efx_start_interrupts(struct efx_nic *efx, bool may_keep_eventq);
191 static void efx_stop_interrupts(struct efx_nic *efx, bool may_keep_eventq);
192 static void efx_remove_channel(struct efx_channel *channel);
193 static void efx_remove_channels(struct efx_nic *efx);
194 static const struct efx_channel_type efx_default_channel_type;
195 static void efx_remove_port(struct efx_nic *efx);
196 static void efx_init_napi_channel(struct efx_channel *channel);
197 static void efx_fini_napi(struct efx_nic *efx);
198 static void efx_fini_napi_channel(struct efx_channel *channel);
199 static void efx_fini_struct(struct efx_nic *efx);
200 static void efx_start_all(struct efx_nic *efx);
201 static void efx_stop_all(struct efx_nic *efx);
202 
203 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
204 	do {						\
205 		if ((efx->state == STATE_READY) ||	\
206 		    (efx->state == STATE_DISABLED))	\
207 			ASSERT_RTNL();			\
208 	} while (0)
209 
210 static int efx_check_disabled(struct efx_nic *efx)
211 {
212 	if (efx->state == STATE_DISABLED) {
213 		netif_err(efx, drv, efx->net_dev,
214 			  "device is disabled due to earlier errors\n");
215 		return -EIO;
216 	}
217 	return 0;
218 }
219 
220 /**************************************************************************
221  *
222  * Event queue processing
223  *
224  *************************************************************************/
225 
226 /* Process channel's event queue
227  *
228  * This function is responsible for processing the event queue of a
229  * single channel.  The caller must guarantee that this function will
230  * never be concurrently called more than once on the same channel,
231  * though different channels may be being processed concurrently.
232  */
233 static int efx_process_channel(struct efx_channel *channel, int budget)
234 {
235 	int spent;
236 
237 	if (unlikely(!channel->enabled))
238 		return 0;
239 
240 	spent = efx_nic_process_eventq(channel, budget);
241 	if (spent && efx_channel_has_rx_queue(channel)) {
242 		struct efx_rx_queue *rx_queue =
243 			efx_channel_get_rx_queue(channel);
244 
245 		/* Deliver last RX packet. */
246 		if (channel->rx_pkt) {
247 			__efx_rx_packet(channel, channel->rx_pkt);
248 			channel->rx_pkt = NULL;
249 		}
250 		if (rx_queue->enabled) {
251 			efx_rx_strategy(channel);
252 			efx_fast_push_rx_descriptors(rx_queue);
253 		}
254 	}
255 
256 	return spent;
257 }
258 
259 /* Mark channel as finished processing
260  *
261  * Note that since we will not receive further interrupts for this
262  * channel before we finish processing and call the eventq_read_ack()
263  * method, there is no need to use the interrupt hold-off timers.
264  */
265 static inline void efx_channel_processed(struct efx_channel *channel)
266 {
267 	/* The interrupt handler for this channel may set work_pending
268 	 * as soon as we acknowledge the events we've seen.  Make sure
269 	 * it's cleared before then. */
270 	channel->work_pending = false;
271 	smp_wmb();
272 
273 	efx_nic_eventq_read_ack(channel);
274 }
275 
276 /* NAPI poll handler
277  *
278  * NAPI guarantees serialisation of polls of the same device, which
279  * provides the guarantee required by efx_process_channel().
280  */
281 static int efx_poll(struct napi_struct *napi, int budget)
282 {
283 	struct efx_channel *channel =
284 		container_of(napi, struct efx_channel, napi_str);
285 	struct efx_nic *efx = channel->efx;
286 	int spent;
287 
288 	netif_vdbg(efx, intr, efx->net_dev,
289 		   "channel %d NAPI poll executing on CPU %d\n",
290 		   channel->channel, raw_smp_processor_id());
291 
292 	spent = efx_process_channel(channel, budget);
293 
294 	if (spent < budget) {
295 		if (efx_channel_has_rx_queue(channel) &&
296 		    efx->irq_rx_adaptive &&
297 		    unlikely(++channel->irq_count == 1000)) {
298 			if (unlikely(channel->irq_mod_score <
299 				     irq_adapt_low_thresh)) {
300 				if (channel->irq_moderation > 1) {
301 					channel->irq_moderation -= 1;
302 					efx->type->push_irq_moderation(channel);
303 				}
304 			} else if (unlikely(channel->irq_mod_score >
305 					    irq_adapt_high_thresh)) {
306 				if (channel->irq_moderation <
307 				    efx->irq_rx_moderation) {
308 					channel->irq_moderation += 1;
309 					efx->type->push_irq_moderation(channel);
310 				}
311 			}
312 			channel->irq_count = 0;
313 			channel->irq_mod_score = 0;
314 		}
315 
316 		efx_filter_rfs_expire(channel);
317 
318 		/* There is no race here; although napi_disable() will
319 		 * only wait for napi_complete(), this isn't a problem
320 		 * since efx_channel_processed() will have no effect if
321 		 * interrupts have already been disabled.
322 		 */
323 		napi_complete(napi);
324 		efx_channel_processed(channel);
325 	}
326 
327 	return spent;
328 }
329 
330 /* Process the eventq of the specified channel immediately on this CPU
331  *
332  * Disable hardware generated interrupts, wait for any existing
333  * processing to finish, then directly poll (and ack ) the eventq.
334  * Finally reenable NAPI and interrupts.
335  *
336  * This is for use only during a loopback self-test.  It must not
337  * deliver any packets up the stack as this can result in deadlock.
338  */
339 void efx_process_channel_now(struct efx_channel *channel)
340 {
341 	struct efx_nic *efx = channel->efx;
342 
343 	BUG_ON(channel->channel >= efx->n_channels);
344 	BUG_ON(!channel->enabled);
345 	BUG_ON(!efx->loopback_selftest);
346 
347 	/* Disable interrupts and wait for ISRs to complete */
348 	efx_nic_disable_interrupts(efx);
349 	if (efx->legacy_irq) {
350 		synchronize_irq(efx->legacy_irq);
351 		efx->legacy_irq_enabled = false;
352 	}
353 	if (channel->irq)
354 		synchronize_irq(channel->irq);
355 
356 	/* Wait for any NAPI processing to complete */
357 	napi_disable(&channel->napi_str);
358 
359 	/* Poll the channel */
360 	efx_process_channel(channel, channel->eventq_mask + 1);
361 
362 	/* Ack the eventq. This may cause an interrupt to be generated
363 	 * when they are reenabled */
364 	efx_channel_processed(channel);
365 
366 	napi_enable(&channel->napi_str);
367 	if (efx->legacy_irq)
368 		efx->legacy_irq_enabled = true;
369 	efx_nic_enable_interrupts(efx);
370 }
371 
372 /* Create event queue
373  * Event queue memory allocations are done only once.  If the channel
374  * is reset, the memory buffer will be reused; this guards against
375  * errors during channel reset and also simplifies interrupt handling.
376  */
377 static int efx_probe_eventq(struct efx_channel *channel)
378 {
379 	struct efx_nic *efx = channel->efx;
380 	unsigned long entries;
381 
382 	netif_dbg(efx, probe, efx->net_dev,
383 		  "chan %d create event queue\n", channel->channel);
384 
385 	/* Build an event queue with room for one event per tx and rx buffer,
386 	 * plus some extra for link state events and MCDI completions. */
387 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
388 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
389 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
390 
391 	return efx_nic_probe_eventq(channel);
392 }
393 
394 /* Prepare channel's event queue */
395 static void efx_init_eventq(struct efx_channel *channel)
396 {
397 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
398 		  "chan %d init event queue\n", channel->channel);
399 
400 	channel->eventq_read_ptr = 0;
401 
402 	efx_nic_init_eventq(channel);
403 }
404 
405 /* Enable event queue processing and NAPI */
406 static void efx_start_eventq(struct efx_channel *channel)
407 {
408 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
409 		  "chan %d start event queue\n", channel->channel);
410 
411 	/* The interrupt handler for this channel may set work_pending
412 	 * as soon as we enable it.  Make sure it's cleared before
413 	 * then.  Similarly, make sure it sees the enabled flag set.
414 	 */
415 	channel->work_pending = false;
416 	channel->enabled = true;
417 	smp_wmb();
418 
419 	napi_enable(&channel->napi_str);
420 	efx_nic_eventq_read_ack(channel);
421 }
422 
423 /* Disable event queue processing and NAPI */
424 static void efx_stop_eventq(struct efx_channel *channel)
425 {
426 	if (!channel->enabled)
427 		return;
428 
429 	napi_disable(&channel->napi_str);
430 	channel->enabled = false;
431 }
432 
433 static void efx_fini_eventq(struct efx_channel *channel)
434 {
435 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
436 		  "chan %d fini event queue\n", channel->channel);
437 
438 	efx_nic_fini_eventq(channel);
439 }
440 
441 static void efx_remove_eventq(struct efx_channel *channel)
442 {
443 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
444 		  "chan %d remove event queue\n", channel->channel);
445 
446 	efx_nic_remove_eventq(channel);
447 }
448 
449 /**************************************************************************
450  *
451  * Channel handling
452  *
453  *************************************************************************/
454 
455 /* Allocate and initialise a channel structure. */
456 static struct efx_channel *
457 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
458 {
459 	struct efx_channel *channel;
460 	struct efx_rx_queue *rx_queue;
461 	struct efx_tx_queue *tx_queue;
462 	int j;
463 
464 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
465 	if (!channel)
466 		return NULL;
467 
468 	channel->efx = efx;
469 	channel->channel = i;
470 	channel->type = &efx_default_channel_type;
471 
472 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
473 		tx_queue = &channel->tx_queue[j];
474 		tx_queue->efx = efx;
475 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
476 		tx_queue->channel = channel;
477 	}
478 
479 	rx_queue = &channel->rx_queue;
480 	rx_queue->efx = efx;
481 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
482 		    (unsigned long)rx_queue);
483 
484 	return channel;
485 }
486 
487 /* Allocate and initialise a channel structure, copying parameters
488  * (but not resources) from an old channel structure.
489  */
490 static struct efx_channel *
491 efx_copy_channel(const struct efx_channel *old_channel)
492 {
493 	struct efx_channel *channel;
494 	struct efx_rx_queue *rx_queue;
495 	struct efx_tx_queue *tx_queue;
496 	int j;
497 
498 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
499 	if (!channel)
500 		return NULL;
501 
502 	*channel = *old_channel;
503 
504 	channel->napi_dev = NULL;
505 	memset(&channel->eventq, 0, sizeof(channel->eventq));
506 
507 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
508 		tx_queue = &channel->tx_queue[j];
509 		if (tx_queue->channel)
510 			tx_queue->channel = channel;
511 		tx_queue->buffer = NULL;
512 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
513 	}
514 
515 	rx_queue = &channel->rx_queue;
516 	rx_queue->buffer = NULL;
517 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
518 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
519 		    (unsigned long)rx_queue);
520 
521 	return channel;
522 }
523 
524 static int efx_probe_channel(struct efx_channel *channel)
525 {
526 	struct efx_tx_queue *tx_queue;
527 	struct efx_rx_queue *rx_queue;
528 	int rc;
529 
530 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
531 		  "creating channel %d\n", channel->channel);
532 
533 	rc = channel->type->pre_probe(channel);
534 	if (rc)
535 		goto fail;
536 
537 	rc = efx_probe_eventq(channel);
538 	if (rc)
539 		goto fail;
540 
541 	efx_for_each_channel_tx_queue(tx_queue, channel) {
542 		rc = efx_probe_tx_queue(tx_queue);
543 		if (rc)
544 			goto fail;
545 	}
546 
547 	efx_for_each_channel_rx_queue(rx_queue, channel) {
548 		rc = efx_probe_rx_queue(rx_queue);
549 		if (rc)
550 			goto fail;
551 	}
552 
553 	channel->n_rx_frm_trunc = 0;
554 
555 	return 0;
556 
557 fail:
558 	efx_remove_channel(channel);
559 	return rc;
560 }
561 
562 static void
563 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
564 {
565 	struct efx_nic *efx = channel->efx;
566 	const char *type;
567 	int number;
568 
569 	number = channel->channel;
570 	if (efx->tx_channel_offset == 0) {
571 		type = "";
572 	} else if (channel->channel < efx->tx_channel_offset) {
573 		type = "-rx";
574 	} else {
575 		type = "-tx";
576 		number -= efx->tx_channel_offset;
577 	}
578 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
579 }
580 
581 static void efx_set_channel_names(struct efx_nic *efx)
582 {
583 	struct efx_channel *channel;
584 
585 	efx_for_each_channel(channel, efx)
586 		channel->type->get_name(channel,
587 					efx->channel_name[channel->channel],
588 					sizeof(efx->channel_name[0]));
589 }
590 
591 static int efx_probe_channels(struct efx_nic *efx)
592 {
593 	struct efx_channel *channel;
594 	int rc;
595 
596 	/* Restart special buffer allocation */
597 	efx->next_buffer_table = 0;
598 
599 	/* Probe channels in reverse, so that any 'extra' channels
600 	 * use the start of the buffer table. This allows the traffic
601 	 * channels to be resized without moving them or wasting the
602 	 * entries before them.
603 	 */
604 	efx_for_each_channel_rev(channel, efx) {
605 		rc = efx_probe_channel(channel);
606 		if (rc) {
607 			netif_err(efx, probe, efx->net_dev,
608 				  "failed to create channel %d\n",
609 				  channel->channel);
610 			goto fail;
611 		}
612 	}
613 	efx_set_channel_names(efx);
614 
615 	return 0;
616 
617 fail:
618 	efx_remove_channels(efx);
619 	return rc;
620 }
621 
622 /* Channels are shutdown and reinitialised whilst the NIC is running
623  * to propagate configuration changes (mtu, checksum offload), or
624  * to clear hardware error conditions
625  */
626 static void efx_start_datapath(struct efx_nic *efx)
627 {
628 	struct efx_tx_queue *tx_queue;
629 	struct efx_rx_queue *rx_queue;
630 	struct efx_channel *channel;
631 
632 	/* Calculate the rx buffer allocation parameters required to
633 	 * support the current MTU, including padding for header
634 	 * alignment and overruns.
635 	 */
636 	efx->rx_buffer_len = (max(EFX_PAGE_IP_ALIGN, NET_IP_ALIGN) +
637 			      EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
638 			      efx->type->rx_buffer_hash_size +
639 			      efx->type->rx_buffer_padding);
640 	efx->rx_buffer_order = get_order(efx->rx_buffer_len +
641 					 sizeof(struct efx_rx_page_state));
642 
643 	/* We must keep at least one descriptor in a TX ring empty.
644 	 * We could avoid this when the queue size does not exactly
645 	 * match the hardware ring size, but it's not that important.
646 	 * Therefore we stop the queue when one more skb might fill
647 	 * the ring completely.  We wake it when half way back to
648 	 * empty.
649 	 */
650 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
651 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
652 
653 	/* Initialise the channels */
654 	efx_for_each_channel(channel, efx) {
655 		efx_for_each_channel_tx_queue(tx_queue, channel)
656 			efx_init_tx_queue(tx_queue);
657 
658 		/* The rx buffer allocation strategy is MTU dependent */
659 		efx_rx_strategy(channel);
660 
661 		efx_for_each_channel_rx_queue(rx_queue, channel) {
662 			efx_init_rx_queue(rx_queue);
663 			efx_nic_generate_fill_event(rx_queue);
664 		}
665 
666 		WARN_ON(channel->rx_pkt != NULL);
667 		efx_rx_strategy(channel);
668 	}
669 
670 	if (netif_device_present(efx->net_dev))
671 		netif_tx_wake_all_queues(efx->net_dev);
672 }
673 
674 static void efx_stop_datapath(struct efx_nic *efx)
675 {
676 	struct efx_channel *channel;
677 	struct efx_tx_queue *tx_queue;
678 	struct efx_rx_queue *rx_queue;
679 	struct pci_dev *dev = efx->pci_dev;
680 	int rc;
681 
682 	EFX_ASSERT_RESET_SERIALISED(efx);
683 	BUG_ON(efx->port_enabled);
684 
685 	/* Only perform flush if dma is enabled */
686 	if (dev->is_busmaster) {
687 		rc = efx_nic_flush_queues(efx);
688 
689 		if (rc && EFX_WORKAROUND_7803(efx)) {
690 			/* Schedule a reset to recover from the flush failure. The
691 			 * descriptor caches reference memory we're about to free,
692 			 * but falcon_reconfigure_mac_wrapper() won't reconnect
693 			 * the MACs because of the pending reset. */
694 			netif_err(efx, drv, efx->net_dev,
695 				  "Resetting to recover from flush failure\n");
696 			efx_schedule_reset(efx, RESET_TYPE_ALL);
697 		} else if (rc) {
698 			netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
699 		} else {
700 			netif_dbg(efx, drv, efx->net_dev,
701 				  "successfully flushed all queues\n");
702 		}
703 	}
704 
705 	efx_for_each_channel(channel, efx) {
706 		/* RX packet processing is pipelined, so wait for the
707 		 * NAPI handler to complete.  At least event queue 0
708 		 * might be kept active by non-data events, so don't
709 		 * use napi_synchronize() but actually disable NAPI
710 		 * temporarily.
711 		 */
712 		if (efx_channel_has_rx_queue(channel)) {
713 			efx_stop_eventq(channel);
714 			efx_start_eventq(channel);
715 		}
716 
717 		efx_for_each_channel_rx_queue(rx_queue, channel)
718 			efx_fini_rx_queue(rx_queue);
719 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
720 			efx_fini_tx_queue(tx_queue);
721 	}
722 }
723 
724 static void efx_remove_channel(struct efx_channel *channel)
725 {
726 	struct efx_tx_queue *tx_queue;
727 	struct efx_rx_queue *rx_queue;
728 
729 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
730 		  "destroy chan %d\n", channel->channel);
731 
732 	efx_for_each_channel_rx_queue(rx_queue, channel)
733 		efx_remove_rx_queue(rx_queue);
734 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
735 		efx_remove_tx_queue(tx_queue);
736 	efx_remove_eventq(channel);
737 	channel->type->post_remove(channel);
738 }
739 
740 static void efx_remove_channels(struct efx_nic *efx)
741 {
742 	struct efx_channel *channel;
743 
744 	efx_for_each_channel(channel, efx)
745 		efx_remove_channel(channel);
746 }
747 
748 int
749 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
750 {
751 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
752 	u32 old_rxq_entries, old_txq_entries;
753 	unsigned i, next_buffer_table = 0;
754 	int rc;
755 
756 	rc = efx_check_disabled(efx);
757 	if (rc)
758 		return rc;
759 
760 	/* Not all channels should be reallocated. We must avoid
761 	 * reallocating their buffer table entries.
762 	 */
763 	efx_for_each_channel(channel, efx) {
764 		struct efx_rx_queue *rx_queue;
765 		struct efx_tx_queue *tx_queue;
766 
767 		if (channel->type->copy)
768 			continue;
769 		next_buffer_table = max(next_buffer_table,
770 					channel->eventq.index +
771 					channel->eventq.entries);
772 		efx_for_each_channel_rx_queue(rx_queue, channel)
773 			next_buffer_table = max(next_buffer_table,
774 						rx_queue->rxd.index +
775 						rx_queue->rxd.entries);
776 		efx_for_each_channel_tx_queue(tx_queue, channel)
777 			next_buffer_table = max(next_buffer_table,
778 						tx_queue->txd.index +
779 						tx_queue->txd.entries);
780 	}
781 
782 	efx_device_detach_sync(efx);
783 	efx_stop_all(efx);
784 	efx_stop_interrupts(efx, true);
785 
786 	/* Clone channels (where possible) */
787 	memset(other_channel, 0, sizeof(other_channel));
788 	for (i = 0; i < efx->n_channels; i++) {
789 		channel = efx->channel[i];
790 		if (channel->type->copy)
791 			channel = channel->type->copy(channel);
792 		if (!channel) {
793 			rc = -ENOMEM;
794 			goto out;
795 		}
796 		other_channel[i] = channel;
797 	}
798 
799 	/* Swap entry counts and channel pointers */
800 	old_rxq_entries = efx->rxq_entries;
801 	old_txq_entries = efx->txq_entries;
802 	efx->rxq_entries = rxq_entries;
803 	efx->txq_entries = txq_entries;
804 	for (i = 0; i < efx->n_channels; i++) {
805 		channel = efx->channel[i];
806 		efx->channel[i] = other_channel[i];
807 		other_channel[i] = channel;
808 	}
809 
810 	/* Restart buffer table allocation */
811 	efx->next_buffer_table = next_buffer_table;
812 
813 	for (i = 0; i < efx->n_channels; i++) {
814 		channel = efx->channel[i];
815 		if (!channel->type->copy)
816 			continue;
817 		rc = efx_probe_channel(channel);
818 		if (rc)
819 			goto rollback;
820 		efx_init_napi_channel(efx->channel[i]);
821 	}
822 
823 out:
824 	/* Destroy unused channel structures */
825 	for (i = 0; i < efx->n_channels; i++) {
826 		channel = other_channel[i];
827 		if (channel && channel->type->copy) {
828 			efx_fini_napi_channel(channel);
829 			efx_remove_channel(channel);
830 			kfree(channel);
831 		}
832 	}
833 
834 	efx_start_interrupts(efx, true);
835 	efx_start_all(efx);
836 	netif_device_attach(efx->net_dev);
837 	return rc;
838 
839 rollback:
840 	/* Swap back */
841 	efx->rxq_entries = old_rxq_entries;
842 	efx->txq_entries = old_txq_entries;
843 	for (i = 0; i < efx->n_channels; i++) {
844 		channel = efx->channel[i];
845 		efx->channel[i] = other_channel[i];
846 		other_channel[i] = channel;
847 	}
848 	goto out;
849 }
850 
851 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
852 {
853 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
854 }
855 
856 static const struct efx_channel_type efx_default_channel_type = {
857 	.pre_probe		= efx_channel_dummy_op_int,
858 	.post_remove		= efx_channel_dummy_op_void,
859 	.get_name		= efx_get_channel_name,
860 	.copy			= efx_copy_channel,
861 	.keep_eventq		= false,
862 };
863 
864 int efx_channel_dummy_op_int(struct efx_channel *channel)
865 {
866 	return 0;
867 }
868 
869 void efx_channel_dummy_op_void(struct efx_channel *channel)
870 {
871 }
872 
873 /**************************************************************************
874  *
875  * Port handling
876  *
877  **************************************************************************/
878 
879 /* This ensures that the kernel is kept informed (via
880  * netif_carrier_on/off) of the link status, and also maintains the
881  * link status's stop on the port's TX queue.
882  */
883 void efx_link_status_changed(struct efx_nic *efx)
884 {
885 	struct efx_link_state *link_state = &efx->link_state;
886 
887 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
888 	 * that no events are triggered between unregister_netdev() and the
889 	 * driver unloading. A more general condition is that NETDEV_CHANGE
890 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
891 	if (!netif_running(efx->net_dev))
892 		return;
893 
894 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
895 		efx->n_link_state_changes++;
896 
897 		if (link_state->up)
898 			netif_carrier_on(efx->net_dev);
899 		else
900 			netif_carrier_off(efx->net_dev);
901 	}
902 
903 	/* Status message for kernel log */
904 	if (link_state->up)
905 		netif_info(efx, link, efx->net_dev,
906 			   "link up at %uMbps %s-duplex (MTU %d)%s\n",
907 			   link_state->speed, link_state->fd ? "full" : "half",
908 			   efx->net_dev->mtu,
909 			   (efx->promiscuous ? " [PROMISC]" : ""));
910 	else
911 		netif_info(efx, link, efx->net_dev, "link down\n");
912 }
913 
914 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
915 {
916 	efx->link_advertising = advertising;
917 	if (advertising) {
918 		if (advertising & ADVERTISED_Pause)
919 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
920 		else
921 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
922 		if (advertising & ADVERTISED_Asym_Pause)
923 			efx->wanted_fc ^= EFX_FC_TX;
924 	}
925 }
926 
927 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
928 {
929 	efx->wanted_fc = wanted_fc;
930 	if (efx->link_advertising) {
931 		if (wanted_fc & EFX_FC_RX)
932 			efx->link_advertising |= (ADVERTISED_Pause |
933 						  ADVERTISED_Asym_Pause);
934 		else
935 			efx->link_advertising &= ~(ADVERTISED_Pause |
936 						   ADVERTISED_Asym_Pause);
937 		if (wanted_fc & EFX_FC_TX)
938 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
939 	}
940 }
941 
942 static void efx_fini_port(struct efx_nic *efx);
943 
944 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
945  * the MAC appropriately. All other PHY configuration changes are pushed
946  * through phy_op->set_settings(), and pushed asynchronously to the MAC
947  * through efx_monitor().
948  *
949  * Callers must hold the mac_lock
950  */
951 int __efx_reconfigure_port(struct efx_nic *efx)
952 {
953 	enum efx_phy_mode phy_mode;
954 	int rc;
955 
956 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
957 
958 	/* Serialise the promiscuous flag with efx_set_rx_mode. */
959 	netif_addr_lock_bh(efx->net_dev);
960 	netif_addr_unlock_bh(efx->net_dev);
961 
962 	/* Disable PHY transmit in mac level loopbacks */
963 	phy_mode = efx->phy_mode;
964 	if (LOOPBACK_INTERNAL(efx))
965 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
966 	else
967 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
968 
969 	rc = efx->type->reconfigure_port(efx);
970 
971 	if (rc)
972 		efx->phy_mode = phy_mode;
973 
974 	return rc;
975 }
976 
977 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
978  * disabled. */
979 int efx_reconfigure_port(struct efx_nic *efx)
980 {
981 	int rc;
982 
983 	EFX_ASSERT_RESET_SERIALISED(efx);
984 
985 	mutex_lock(&efx->mac_lock);
986 	rc = __efx_reconfigure_port(efx);
987 	mutex_unlock(&efx->mac_lock);
988 
989 	return rc;
990 }
991 
992 /* Asynchronous work item for changing MAC promiscuity and multicast
993  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
994  * MAC directly. */
995 static void efx_mac_work(struct work_struct *data)
996 {
997 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
998 
999 	mutex_lock(&efx->mac_lock);
1000 	if (efx->port_enabled)
1001 		efx->type->reconfigure_mac(efx);
1002 	mutex_unlock(&efx->mac_lock);
1003 }
1004 
1005 static int efx_probe_port(struct efx_nic *efx)
1006 {
1007 	int rc;
1008 
1009 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1010 
1011 	if (phy_flash_cfg)
1012 		efx->phy_mode = PHY_MODE_SPECIAL;
1013 
1014 	/* Connect up MAC/PHY operations table */
1015 	rc = efx->type->probe_port(efx);
1016 	if (rc)
1017 		return rc;
1018 
1019 	/* Initialise MAC address to permanent address */
1020 	memcpy(efx->net_dev->dev_addr, efx->net_dev->perm_addr, ETH_ALEN);
1021 
1022 	return 0;
1023 }
1024 
1025 static int efx_init_port(struct efx_nic *efx)
1026 {
1027 	int rc;
1028 
1029 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1030 
1031 	mutex_lock(&efx->mac_lock);
1032 
1033 	rc = efx->phy_op->init(efx);
1034 	if (rc)
1035 		goto fail1;
1036 
1037 	efx->port_initialized = true;
1038 
1039 	/* Reconfigure the MAC before creating dma queues (required for
1040 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1041 	efx->type->reconfigure_mac(efx);
1042 
1043 	/* Ensure the PHY advertises the correct flow control settings */
1044 	rc = efx->phy_op->reconfigure(efx);
1045 	if (rc)
1046 		goto fail2;
1047 
1048 	mutex_unlock(&efx->mac_lock);
1049 	return 0;
1050 
1051 fail2:
1052 	efx->phy_op->fini(efx);
1053 fail1:
1054 	mutex_unlock(&efx->mac_lock);
1055 	return rc;
1056 }
1057 
1058 static void efx_start_port(struct efx_nic *efx)
1059 {
1060 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1061 	BUG_ON(efx->port_enabled);
1062 
1063 	mutex_lock(&efx->mac_lock);
1064 	efx->port_enabled = true;
1065 
1066 	/* efx_mac_work() might have been scheduled after efx_stop_port(),
1067 	 * and then cancelled by efx_flush_all() */
1068 	efx->type->reconfigure_mac(efx);
1069 
1070 	mutex_unlock(&efx->mac_lock);
1071 }
1072 
1073 /* Prevent efx_mac_work() and efx_monitor() from working */
1074 static void efx_stop_port(struct efx_nic *efx)
1075 {
1076 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1077 
1078 	mutex_lock(&efx->mac_lock);
1079 	efx->port_enabled = false;
1080 	mutex_unlock(&efx->mac_lock);
1081 
1082 	/* Serialise against efx_set_multicast_list() */
1083 	netif_addr_lock_bh(efx->net_dev);
1084 	netif_addr_unlock_bh(efx->net_dev);
1085 }
1086 
1087 static void efx_fini_port(struct efx_nic *efx)
1088 {
1089 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1090 
1091 	if (!efx->port_initialized)
1092 		return;
1093 
1094 	efx->phy_op->fini(efx);
1095 	efx->port_initialized = false;
1096 
1097 	efx->link_state.up = false;
1098 	efx_link_status_changed(efx);
1099 }
1100 
1101 static void efx_remove_port(struct efx_nic *efx)
1102 {
1103 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1104 
1105 	efx->type->remove_port(efx);
1106 }
1107 
1108 /**************************************************************************
1109  *
1110  * NIC handling
1111  *
1112  **************************************************************************/
1113 
1114 /* This configures the PCI device to enable I/O and DMA. */
1115 static int efx_init_io(struct efx_nic *efx)
1116 {
1117 	struct pci_dev *pci_dev = efx->pci_dev;
1118 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1119 	int rc;
1120 
1121 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1122 
1123 	rc = pci_enable_device(pci_dev);
1124 	if (rc) {
1125 		netif_err(efx, probe, efx->net_dev,
1126 			  "failed to enable PCI device\n");
1127 		goto fail1;
1128 	}
1129 
1130 	pci_set_master(pci_dev);
1131 
1132 	/* Set the PCI DMA mask.  Try all possibilities from our
1133 	 * genuine mask down to 32 bits, because some architectures
1134 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1135 	 * masks event though they reject 46 bit masks.
1136 	 */
1137 	while (dma_mask > 0x7fffffffUL) {
1138 		if (dma_supported(&pci_dev->dev, dma_mask)) {
1139 			rc = dma_set_mask(&pci_dev->dev, dma_mask);
1140 			if (rc == 0)
1141 				break;
1142 		}
1143 		dma_mask >>= 1;
1144 	}
1145 	if (rc) {
1146 		netif_err(efx, probe, efx->net_dev,
1147 			  "could not find a suitable DMA mask\n");
1148 		goto fail2;
1149 	}
1150 	netif_dbg(efx, probe, efx->net_dev,
1151 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1152 	rc = dma_set_coherent_mask(&pci_dev->dev, dma_mask);
1153 	if (rc) {
1154 		/* dma_set_coherent_mask() is not *allowed* to
1155 		 * fail with a mask that dma_set_mask() accepted,
1156 		 * but just in case...
1157 		 */
1158 		netif_err(efx, probe, efx->net_dev,
1159 			  "failed to set consistent DMA mask\n");
1160 		goto fail2;
1161 	}
1162 
1163 	efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1164 	rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1165 	if (rc) {
1166 		netif_err(efx, probe, efx->net_dev,
1167 			  "request for memory BAR failed\n");
1168 		rc = -EIO;
1169 		goto fail3;
1170 	}
1171 	efx->membase = ioremap_nocache(efx->membase_phys,
1172 				       efx->type->mem_map_size);
1173 	if (!efx->membase) {
1174 		netif_err(efx, probe, efx->net_dev,
1175 			  "could not map memory BAR at %llx+%x\n",
1176 			  (unsigned long long)efx->membase_phys,
1177 			  efx->type->mem_map_size);
1178 		rc = -ENOMEM;
1179 		goto fail4;
1180 	}
1181 	netif_dbg(efx, probe, efx->net_dev,
1182 		  "memory BAR at %llx+%x (virtual %p)\n",
1183 		  (unsigned long long)efx->membase_phys,
1184 		  efx->type->mem_map_size, efx->membase);
1185 
1186 	return 0;
1187 
1188  fail4:
1189 	pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1190  fail3:
1191 	efx->membase_phys = 0;
1192  fail2:
1193 	pci_disable_device(efx->pci_dev);
1194  fail1:
1195 	return rc;
1196 }
1197 
1198 static void efx_fini_io(struct efx_nic *efx)
1199 {
1200 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1201 
1202 	if (efx->membase) {
1203 		iounmap(efx->membase);
1204 		efx->membase = NULL;
1205 	}
1206 
1207 	if (efx->membase_phys) {
1208 		pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1209 		efx->membase_phys = 0;
1210 	}
1211 
1212 	pci_disable_device(efx->pci_dev);
1213 }
1214 
1215 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1216 {
1217 	cpumask_var_t thread_mask;
1218 	unsigned int count;
1219 	int cpu;
1220 
1221 	if (rss_cpus) {
1222 		count = rss_cpus;
1223 	} else {
1224 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1225 			netif_warn(efx, probe, efx->net_dev,
1226 				   "RSS disabled due to allocation failure\n");
1227 			return 1;
1228 		}
1229 
1230 		count = 0;
1231 		for_each_online_cpu(cpu) {
1232 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1233 				++count;
1234 				cpumask_or(thread_mask, thread_mask,
1235 					   topology_thread_cpumask(cpu));
1236 			}
1237 		}
1238 
1239 		free_cpumask_var(thread_mask);
1240 	}
1241 
1242 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1243 	 * table entries that are inaccessible to VFs
1244 	 */
1245 	if (efx_sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1246 	    count > efx_vf_size(efx)) {
1247 		netif_warn(efx, probe, efx->net_dev,
1248 			   "Reducing number of RSS channels from %u to %u for "
1249 			   "VF support. Increase vf-msix-limit to use more "
1250 			   "channels on the PF.\n",
1251 			   count, efx_vf_size(efx));
1252 		count = efx_vf_size(efx);
1253 	}
1254 
1255 	return count;
1256 }
1257 
1258 static int
1259 efx_init_rx_cpu_rmap(struct efx_nic *efx, struct msix_entry *xentries)
1260 {
1261 #ifdef CONFIG_RFS_ACCEL
1262 	unsigned int i;
1263 	int rc;
1264 
1265 	efx->net_dev->rx_cpu_rmap = alloc_irq_cpu_rmap(efx->n_rx_channels);
1266 	if (!efx->net_dev->rx_cpu_rmap)
1267 		return -ENOMEM;
1268 	for (i = 0; i < efx->n_rx_channels; i++) {
1269 		rc = irq_cpu_rmap_add(efx->net_dev->rx_cpu_rmap,
1270 				      xentries[i].vector);
1271 		if (rc) {
1272 			free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1273 			efx->net_dev->rx_cpu_rmap = NULL;
1274 			return rc;
1275 		}
1276 	}
1277 #endif
1278 	return 0;
1279 }
1280 
1281 /* Probe the number and type of interrupts we are able to obtain, and
1282  * the resulting numbers of channels and RX queues.
1283  */
1284 static int efx_probe_interrupts(struct efx_nic *efx)
1285 {
1286 	unsigned int max_channels =
1287 		min(efx->type->phys_addr_channels, EFX_MAX_CHANNELS);
1288 	unsigned int extra_channels = 0;
1289 	unsigned int i, j;
1290 	int rc;
1291 
1292 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1293 		if (efx->extra_channel_type[i])
1294 			++extra_channels;
1295 
1296 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1297 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1298 		unsigned int n_channels;
1299 
1300 		n_channels = efx_wanted_parallelism(efx);
1301 		if (separate_tx_channels)
1302 			n_channels *= 2;
1303 		n_channels += extra_channels;
1304 		n_channels = min(n_channels, max_channels);
1305 
1306 		for (i = 0; i < n_channels; i++)
1307 			xentries[i].entry = i;
1308 		rc = pci_enable_msix(efx->pci_dev, xentries, n_channels);
1309 		if (rc > 0) {
1310 			netif_err(efx, drv, efx->net_dev,
1311 				  "WARNING: Insufficient MSI-X vectors"
1312 				  " available (%d < %u).\n", rc, n_channels);
1313 			netif_err(efx, drv, efx->net_dev,
1314 				  "WARNING: Performance may be reduced.\n");
1315 			EFX_BUG_ON_PARANOID(rc >= n_channels);
1316 			n_channels = rc;
1317 			rc = pci_enable_msix(efx->pci_dev, xentries,
1318 					     n_channels);
1319 		}
1320 
1321 		if (rc == 0) {
1322 			efx->n_channels = n_channels;
1323 			if (n_channels > extra_channels)
1324 				n_channels -= extra_channels;
1325 			if (separate_tx_channels) {
1326 				efx->n_tx_channels = max(n_channels / 2, 1U);
1327 				efx->n_rx_channels = max(n_channels -
1328 							 efx->n_tx_channels,
1329 							 1U);
1330 			} else {
1331 				efx->n_tx_channels = n_channels;
1332 				efx->n_rx_channels = n_channels;
1333 			}
1334 			rc = efx_init_rx_cpu_rmap(efx, xentries);
1335 			if (rc) {
1336 				pci_disable_msix(efx->pci_dev);
1337 				return rc;
1338 			}
1339 			for (i = 0; i < efx->n_channels; i++)
1340 				efx_get_channel(efx, i)->irq =
1341 					xentries[i].vector;
1342 		} else {
1343 			/* Fall back to single channel MSI */
1344 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1345 			netif_err(efx, drv, efx->net_dev,
1346 				  "could not enable MSI-X\n");
1347 		}
1348 	}
1349 
1350 	/* Try single interrupt MSI */
1351 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1352 		efx->n_channels = 1;
1353 		efx->n_rx_channels = 1;
1354 		efx->n_tx_channels = 1;
1355 		rc = pci_enable_msi(efx->pci_dev);
1356 		if (rc == 0) {
1357 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1358 		} else {
1359 			netif_err(efx, drv, efx->net_dev,
1360 				  "could not enable MSI\n");
1361 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1362 		}
1363 	}
1364 
1365 	/* Assume legacy interrupts */
1366 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1367 		efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1368 		efx->n_rx_channels = 1;
1369 		efx->n_tx_channels = 1;
1370 		efx->legacy_irq = efx->pci_dev->irq;
1371 	}
1372 
1373 	/* Assign extra channels if possible */
1374 	j = efx->n_channels;
1375 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1376 		if (!efx->extra_channel_type[i])
1377 			continue;
1378 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1379 		    efx->n_channels <= extra_channels) {
1380 			efx->extra_channel_type[i]->handle_no_channel(efx);
1381 		} else {
1382 			--j;
1383 			efx_get_channel(efx, j)->type =
1384 				efx->extra_channel_type[i];
1385 		}
1386 	}
1387 
1388 	/* RSS might be usable on VFs even if it is disabled on the PF */
1389 	efx->rss_spread = ((efx->n_rx_channels > 1 || !efx_sriov_wanted(efx)) ?
1390 			   efx->n_rx_channels : efx_vf_size(efx));
1391 
1392 	return 0;
1393 }
1394 
1395 /* Enable interrupts, then probe and start the event queues */
1396 static void efx_start_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1397 {
1398 	struct efx_channel *channel;
1399 
1400 	BUG_ON(efx->state == STATE_DISABLED);
1401 
1402 	if (efx->legacy_irq)
1403 		efx->legacy_irq_enabled = true;
1404 	efx_nic_enable_interrupts(efx);
1405 
1406 	efx_for_each_channel(channel, efx) {
1407 		if (!channel->type->keep_eventq || !may_keep_eventq)
1408 			efx_init_eventq(channel);
1409 		efx_start_eventq(channel);
1410 	}
1411 
1412 	efx_mcdi_mode_event(efx);
1413 }
1414 
1415 static void efx_stop_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1416 {
1417 	struct efx_channel *channel;
1418 
1419 	if (efx->state == STATE_DISABLED)
1420 		return;
1421 
1422 	efx_mcdi_mode_poll(efx);
1423 
1424 	efx_nic_disable_interrupts(efx);
1425 	if (efx->legacy_irq) {
1426 		synchronize_irq(efx->legacy_irq);
1427 		efx->legacy_irq_enabled = false;
1428 	}
1429 
1430 	efx_for_each_channel(channel, efx) {
1431 		if (channel->irq)
1432 			synchronize_irq(channel->irq);
1433 
1434 		efx_stop_eventq(channel);
1435 		if (!channel->type->keep_eventq || !may_keep_eventq)
1436 			efx_fini_eventq(channel);
1437 	}
1438 }
1439 
1440 static void efx_remove_interrupts(struct efx_nic *efx)
1441 {
1442 	struct efx_channel *channel;
1443 
1444 	/* Remove MSI/MSI-X interrupts */
1445 	efx_for_each_channel(channel, efx)
1446 		channel->irq = 0;
1447 	pci_disable_msi(efx->pci_dev);
1448 	pci_disable_msix(efx->pci_dev);
1449 
1450 	/* Remove legacy interrupt */
1451 	efx->legacy_irq = 0;
1452 }
1453 
1454 static void efx_set_channels(struct efx_nic *efx)
1455 {
1456 	struct efx_channel *channel;
1457 	struct efx_tx_queue *tx_queue;
1458 
1459 	efx->tx_channel_offset =
1460 		separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1461 
1462 	/* We need to mark which channels really have RX and TX
1463 	 * queues, and adjust the TX queue numbers if we have separate
1464 	 * RX-only and TX-only channels.
1465 	 */
1466 	efx_for_each_channel(channel, efx) {
1467 		if (channel->channel < efx->n_rx_channels)
1468 			channel->rx_queue.core_index = channel->channel;
1469 		else
1470 			channel->rx_queue.core_index = -1;
1471 
1472 		efx_for_each_channel_tx_queue(tx_queue, channel)
1473 			tx_queue->queue -= (efx->tx_channel_offset *
1474 					    EFX_TXQ_TYPES);
1475 	}
1476 }
1477 
1478 static int efx_probe_nic(struct efx_nic *efx)
1479 {
1480 	size_t i;
1481 	int rc;
1482 
1483 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1484 
1485 	/* Carry out hardware-type specific initialisation */
1486 	rc = efx->type->probe(efx);
1487 	if (rc)
1488 		return rc;
1489 
1490 	/* Determine the number of channels and queues by trying to hook
1491 	 * in MSI-X interrupts. */
1492 	rc = efx_probe_interrupts(efx);
1493 	if (rc)
1494 		goto fail;
1495 
1496 	efx->type->dimension_resources(efx);
1497 
1498 	if (efx->n_channels > 1)
1499 		get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1500 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1501 		efx->rx_indir_table[i] =
1502 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1503 
1504 	efx_set_channels(efx);
1505 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1506 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1507 
1508 	/* Initialise the interrupt moderation settings */
1509 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1510 				true);
1511 
1512 	return 0;
1513 
1514 fail:
1515 	efx->type->remove(efx);
1516 	return rc;
1517 }
1518 
1519 static void efx_remove_nic(struct efx_nic *efx)
1520 {
1521 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1522 
1523 	efx_remove_interrupts(efx);
1524 	efx->type->remove(efx);
1525 }
1526 
1527 /**************************************************************************
1528  *
1529  * NIC startup/shutdown
1530  *
1531  *************************************************************************/
1532 
1533 static int efx_probe_all(struct efx_nic *efx)
1534 {
1535 	int rc;
1536 
1537 	rc = efx_probe_nic(efx);
1538 	if (rc) {
1539 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1540 		goto fail1;
1541 	}
1542 
1543 	rc = efx_probe_port(efx);
1544 	if (rc) {
1545 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1546 		goto fail2;
1547 	}
1548 
1549 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1550 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1551 		rc = -EINVAL;
1552 		goto fail3;
1553 	}
1554 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1555 
1556 	rc = efx_probe_filters(efx);
1557 	if (rc) {
1558 		netif_err(efx, probe, efx->net_dev,
1559 			  "failed to create filter tables\n");
1560 		goto fail3;
1561 	}
1562 
1563 	rc = efx_probe_channels(efx);
1564 	if (rc)
1565 		goto fail4;
1566 
1567 	return 0;
1568 
1569  fail4:
1570 	efx_remove_filters(efx);
1571  fail3:
1572 	efx_remove_port(efx);
1573  fail2:
1574 	efx_remove_nic(efx);
1575  fail1:
1576 	return rc;
1577 }
1578 
1579 /* If the interface is supposed to be running but is not, start
1580  * the hardware and software data path, regular activity for the port
1581  * (MAC statistics, link polling, etc.) and schedule the port to be
1582  * reconfigured.  Interrupts must already be enabled.  This function
1583  * is safe to call multiple times, so long as the NIC is not disabled.
1584  * Requires the RTNL lock.
1585  */
1586 static void efx_start_all(struct efx_nic *efx)
1587 {
1588 	EFX_ASSERT_RESET_SERIALISED(efx);
1589 	BUG_ON(efx->state == STATE_DISABLED);
1590 
1591 	/* Check that it is appropriate to restart the interface. All
1592 	 * of these flags are safe to read under just the rtnl lock */
1593 	if (efx->port_enabled || !netif_running(efx->net_dev))
1594 		return;
1595 
1596 	efx_start_port(efx);
1597 	efx_start_datapath(efx);
1598 
1599 	/* Start the hardware monitor if there is one. Otherwise (we're link
1600 	 * event driven), we have to poll the PHY because after an event queue
1601 	 * flush, we could have a missed a link state change */
1602 	if (efx->type->monitor != NULL) {
1603 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1604 				   efx_monitor_interval);
1605 	} else {
1606 		mutex_lock(&efx->mac_lock);
1607 		if (efx->phy_op->poll(efx))
1608 			efx_link_status_changed(efx);
1609 		mutex_unlock(&efx->mac_lock);
1610 	}
1611 
1612 	efx->type->start_stats(efx);
1613 }
1614 
1615 /* Flush all delayed work. Should only be called when no more delayed work
1616  * will be scheduled. This doesn't flush pending online resets (efx_reset),
1617  * since we're holding the rtnl_lock at this point. */
1618 static void efx_flush_all(struct efx_nic *efx)
1619 {
1620 	/* Make sure the hardware monitor and event self-test are stopped */
1621 	cancel_delayed_work_sync(&efx->monitor_work);
1622 	efx_selftest_async_cancel(efx);
1623 	/* Stop scheduled port reconfigurations */
1624 	cancel_work_sync(&efx->mac_work);
1625 }
1626 
1627 /* Quiesce the hardware and software data path, and regular activity
1628  * for the port without bringing the link down.  Safe to call multiple
1629  * times with the NIC in almost any state, but interrupts should be
1630  * enabled.  Requires the RTNL lock.
1631  */
1632 static void efx_stop_all(struct efx_nic *efx)
1633 {
1634 	EFX_ASSERT_RESET_SERIALISED(efx);
1635 
1636 	/* port_enabled can be read safely under the rtnl lock */
1637 	if (!efx->port_enabled)
1638 		return;
1639 
1640 	efx->type->stop_stats(efx);
1641 	efx_stop_port(efx);
1642 
1643 	/* Flush efx_mac_work(), refill_workqueue, monitor_work */
1644 	efx_flush_all(efx);
1645 
1646 	/* Stop the kernel transmit interface.  This is only valid if
1647 	 * the device is stopped or detached; otherwise the watchdog
1648 	 * may fire immediately.
1649 	 */
1650 	WARN_ON(netif_running(efx->net_dev) &&
1651 		netif_device_present(efx->net_dev));
1652 	netif_tx_disable(efx->net_dev);
1653 
1654 	efx_stop_datapath(efx);
1655 }
1656 
1657 static void efx_remove_all(struct efx_nic *efx)
1658 {
1659 	efx_remove_channels(efx);
1660 	efx_remove_filters(efx);
1661 	efx_remove_port(efx);
1662 	efx_remove_nic(efx);
1663 }
1664 
1665 /**************************************************************************
1666  *
1667  * Interrupt moderation
1668  *
1669  **************************************************************************/
1670 
1671 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1672 {
1673 	if (usecs == 0)
1674 		return 0;
1675 	if (usecs * 1000 < quantum_ns)
1676 		return 1; /* never round down to 0 */
1677 	return usecs * 1000 / quantum_ns;
1678 }
1679 
1680 /* Set interrupt moderation parameters */
1681 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1682 			    unsigned int rx_usecs, bool rx_adaptive,
1683 			    bool rx_may_override_tx)
1684 {
1685 	struct efx_channel *channel;
1686 	unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1687 						efx->timer_quantum_ns,
1688 						1000);
1689 	unsigned int tx_ticks;
1690 	unsigned int rx_ticks;
1691 
1692 	EFX_ASSERT_RESET_SERIALISED(efx);
1693 
1694 	if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1695 		return -EINVAL;
1696 
1697 	tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1698 	rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1699 
1700 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1701 	    !rx_may_override_tx) {
1702 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1703 			  "RX and TX IRQ moderation must be equal\n");
1704 		return -EINVAL;
1705 	}
1706 
1707 	efx->irq_rx_adaptive = rx_adaptive;
1708 	efx->irq_rx_moderation = rx_ticks;
1709 	efx_for_each_channel(channel, efx) {
1710 		if (efx_channel_has_rx_queue(channel))
1711 			channel->irq_moderation = rx_ticks;
1712 		else if (efx_channel_has_tx_queues(channel))
1713 			channel->irq_moderation = tx_ticks;
1714 	}
1715 
1716 	return 0;
1717 }
1718 
1719 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1720 			    unsigned int *rx_usecs, bool *rx_adaptive)
1721 {
1722 	/* We must round up when converting ticks to microseconds
1723 	 * because we round down when converting the other way.
1724 	 */
1725 
1726 	*rx_adaptive = efx->irq_rx_adaptive;
1727 	*rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1728 				 efx->timer_quantum_ns,
1729 				 1000);
1730 
1731 	/* If channels are shared between RX and TX, so is IRQ
1732 	 * moderation.  Otherwise, IRQ moderation is the same for all
1733 	 * TX channels and is not adaptive.
1734 	 */
1735 	if (efx->tx_channel_offset == 0)
1736 		*tx_usecs = *rx_usecs;
1737 	else
1738 		*tx_usecs = DIV_ROUND_UP(
1739 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1740 			efx->timer_quantum_ns,
1741 			1000);
1742 }
1743 
1744 /**************************************************************************
1745  *
1746  * Hardware monitor
1747  *
1748  **************************************************************************/
1749 
1750 /* Run periodically off the general workqueue */
1751 static void efx_monitor(struct work_struct *data)
1752 {
1753 	struct efx_nic *efx = container_of(data, struct efx_nic,
1754 					   monitor_work.work);
1755 
1756 	netif_vdbg(efx, timer, efx->net_dev,
1757 		   "hardware monitor executing on CPU %d\n",
1758 		   raw_smp_processor_id());
1759 	BUG_ON(efx->type->monitor == NULL);
1760 
1761 	/* If the mac_lock is already held then it is likely a port
1762 	 * reconfiguration is already in place, which will likely do
1763 	 * most of the work of monitor() anyway. */
1764 	if (mutex_trylock(&efx->mac_lock)) {
1765 		if (efx->port_enabled)
1766 			efx->type->monitor(efx);
1767 		mutex_unlock(&efx->mac_lock);
1768 	}
1769 
1770 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1771 			   efx_monitor_interval);
1772 }
1773 
1774 /**************************************************************************
1775  *
1776  * ioctls
1777  *
1778  *************************************************************************/
1779 
1780 /* Net device ioctl
1781  * Context: process, rtnl_lock() held.
1782  */
1783 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1784 {
1785 	struct efx_nic *efx = netdev_priv(net_dev);
1786 	struct mii_ioctl_data *data = if_mii(ifr);
1787 
1788 	if (cmd == SIOCSHWTSTAMP)
1789 		return efx_ptp_ioctl(efx, ifr, cmd);
1790 
1791 	/* Convert phy_id from older PRTAD/DEVAD format */
1792 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1793 	    (data->phy_id & 0xfc00) == 0x0400)
1794 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1795 
1796 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
1797 }
1798 
1799 /**************************************************************************
1800  *
1801  * NAPI interface
1802  *
1803  **************************************************************************/
1804 
1805 static void efx_init_napi_channel(struct efx_channel *channel)
1806 {
1807 	struct efx_nic *efx = channel->efx;
1808 
1809 	channel->napi_dev = efx->net_dev;
1810 	netif_napi_add(channel->napi_dev, &channel->napi_str,
1811 		       efx_poll, napi_weight);
1812 }
1813 
1814 static void efx_init_napi(struct efx_nic *efx)
1815 {
1816 	struct efx_channel *channel;
1817 
1818 	efx_for_each_channel(channel, efx)
1819 		efx_init_napi_channel(channel);
1820 }
1821 
1822 static void efx_fini_napi_channel(struct efx_channel *channel)
1823 {
1824 	if (channel->napi_dev)
1825 		netif_napi_del(&channel->napi_str);
1826 	channel->napi_dev = NULL;
1827 }
1828 
1829 static void efx_fini_napi(struct efx_nic *efx)
1830 {
1831 	struct efx_channel *channel;
1832 
1833 	efx_for_each_channel(channel, efx)
1834 		efx_fini_napi_channel(channel);
1835 }
1836 
1837 /**************************************************************************
1838  *
1839  * Kernel netpoll interface
1840  *
1841  *************************************************************************/
1842 
1843 #ifdef CONFIG_NET_POLL_CONTROLLER
1844 
1845 /* Although in the common case interrupts will be disabled, this is not
1846  * guaranteed. However, all our work happens inside the NAPI callback,
1847  * so no locking is required.
1848  */
1849 static void efx_netpoll(struct net_device *net_dev)
1850 {
1851 	struct efx_nic *efx = netdev_priv(net_dev);
1852 	struct efx_channel *channel;
1853 
1854 	efx_for_each_channel(channel, efx)
1855 		efx_schedule_channel(channel);
1856 }
1857 
1858 #endif
1859 
1860 /**************************************************************************
1861  *
1862  * Kernel net device interface
1863  *
1864  *************************************************************************/
1865 
1866 /* Context: process, rtnl_lock() held. */
1867 static int efx_net_open(struct net_device *net_dev)
1868 {
1869 	struct efx_nic *efx = netdev_priv(net_dev);
1870 	int rc;
1871 
1872 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
1873 		  raw_smp_processor_id());
1874 
1875 	rc = efx_check_disabled(efx);
1876 	if (rc)
1877 		return rc;
1878 	if (efx->phy_mode & PHY_MODE_SPECIAL)
1879 		return -EBUSY;
1880 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
1881 		return -EIO;
1882 
1883 	/* Notify the kernel of the link state polled during driver load,
1884 	 * before the monitor starts running */
1885 	efx_link_status_changed(efx);
1886 
1887 	efx_start_all(efx);
1888 	efx_selftest_async_start(efx);
1889 	return 0;
1890 }
1891 
1892 /* Context: process, rtnl_lock() held.
1893  * Note that the kernel will ignore our return code; this method
1894  * should really be a void.
1895  */
1896 static int efx_net_stop(struct net_device *net_dev)
1897 {
1898 	struct efx_nic *efx = netdev_priv(net_dev);
1899 
1900 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
1901 		  raw_smp_processor_id());
1902 
1903 	/* Stop the device and flush all the channels */
1904 	efx_stop_all(efx);
1905 
1906 	return 0;
1907 }
1908 
1909 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1910 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
1911 					       struct rtnl_link_stats64 *stats)
1912 {
1913 	struct efx_nic *efx = netdev_priv(net_dev);
1914 	struct efx_mac_stats *mac_stats = &efx->mac_stats;
1915 
1916 	spin_lock_bh(&efx->stats_lock);
1917 
1918 	efx->type->update_stats(efx);
1919 
1920 	stats->rx_packets = mac_stats->rx_packets;
1921 	stats->tx_packets = mac_stats->tx_packets;
1922 	stats->rx_bytes = mac_stats->rx_bytes;
1923 	stats->tx_bytes = mac_stats->tx_bytes;
1924 	stats->rx_dropped = efx->n_rx_nodesc_drop_cnt;
1925 	stats->multicast = mac_stats->rx_multicast;
1926 	stats->collisions = mac_stats->tx_collision;
1927 	stats->rx_length_errors = (mac_stats->rx_gtjumbo +
1928 				   mac_stats->rx_length_error);
1929 	stats->rx_crc_errors = mac_stats->rx_bad;
1930 	stats->rx_frame_errors = mac_stats->rx_align_error;
1931 	stats->rx_fifo_errors = mac_stats->rx_overflow;
1932 	stats->rx_missed_errors = mac_stats->rx_missed;
1933 	stats->tx_window_errors = mac_stats->tx_late_collision;
1934 
1935 	stats->rx_errors = (stats->rx_length_errors +
1936 			    stats->rx_crc_errors +
1937 			    stats->rx_frame_errors +
1938 			    mac_stats->rx_symbol_error);
1939 	stats->tx_errors = (stats->tx_window_errors +
1940 			    mac_stats->tx_bad);
1941 
1942 	spin_unlock_bh(&efx->stats_lock);
1943 
1944 	return stats;
1945 }
1946 
1947 /* Context: netif_tx_lock held, BHs disabled. */
1948 static void efx_watchdog(struct net_device *net_dev)
1949 {
1950 	struct efx_nic *efx = netdev_priv(net_dev);
1951 
1952 	netif_err(efx, tx_err, efx->net_dev,
1953 		  "TX stuck with port_enabled=%d: resetting channels\n",
1954 		  efx->port_enabled);
1955 
1956 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
1957 }
1958 
1959 
1960 /* Context: process, rtnl_lock() held. */
1961 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
1962 {
1963 	struct efx_nic *efx = netdev_priv(net_dev);
1964 	int rc;
1965 
1966 	rc = efx_check_disabled(efx);
1967 	if (rc)
1968 		return rc;
1969 	if (new_mtu > EFX_MAX_MTU)
1970 		return -EINVAL;
1971 
1972 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
1973 
1974 	efx_device_detach_sync(efx);
1975 	efx_stop_all(efx);
1976 
1977 	mutex_lock(&efx->mac_lock);
1978 	net_dev->mtu = new_mtu;
1979 	efx->type->reconfigure_mac(efx);
1980 	mutex_unlock(&efx->mac_lock);
1981 
1982 	efx_start_all(efx);
1983 	netif_device_attach(efx->net_dev);
1984 	return 0;
1985 }
1986 
1987 static int efx_set_mac_address(struct net_device *net_dev, void *data)
1988 {
1989 	struct efx_nic *efx = netdev_priv(net_dev);
1990 	struct sockaddr *addr = data;
1991 	char *new_addr = addr->sa_data;
1992 
1993 	if (!is_valid_ether_addr(new_addr)) {
1994 		netif_err(efx, drv, efx->net_dev,
1995 			  "invalid ethernet MAC address requested: %pM\n",
1996 			  new_addr);
1997 		return -EADDRNOTAVAIL;
1998 	}
1999 
2000 	memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
2001 	efx_sriov_mac_address_changed(efx);
2002 
2003 	/* Reconfigure the MAC */
2004 	mutex_lock(&efx->mac_lock);
2005 	efx->type->reconfigure_mac(efx);
2006 	mutex_unlock(&efx->mac_lock);
2007 
2008 	return 0;
2009 }
2010 
2011 /* Context: netif_addr_lock held, BHs disabled. */
2012 static void efx_set_rx_mode(struct net_device *net_dev)
2013 {
2014 	struct efx_nic *efx = netdev_priv(net_dev);
2015 	struct netdev_hw_addr *ha;
2016 	union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2017 	u32 crc;
2018 	int bit;
2019 
2020 	efx->promiscuous = !!(net_dev->flags & IFF_PROMISC);
2021 
2022 	/* Build multicast hash table */
2023 	if (efx->promiscuous || (net_dev->flags & IFF_ALLMULTI)) {
2024 		memset(mc_hash, 0xff, sizeof(*mc_hash));
2025 	} else {
2026 		memset(mc_hash, 0x00, sizeof(*mc_hash));
2027 		netdev_for_each_mc_addr(ha, net_dev) {
2028 			crc = ether_crc_le(ETH_ALEN, ha->addr);
2029 			bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2030 			__set_bit_le(bit, mc_hash);
2031 		}
2032 
2033 		/* Broadcast packets go through the multicast hash filter.
2034 		 * ether_crc_le() of the broadcast address is 0xbe2612ff
2035 		 * so we always add bit 0xff to the mask.
2036 		 */
2037 		__set_bit_le(0xff, mc_hash);
2038 	}
2039 
2040 	if (efx->port_enabled)
2041 		queue_work(efx->workqueue, &efx->mac_work);
2042 	/* Otherwise efx_start_port() will do this */
2043 }
2044 
2045 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2046 {
2047 	struct efx_nic *efx = netdev_priv(net_dev);
2048 
2049 	/* If disabling RX n-tuple filtering, clear existing filters */
2050 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
2051 		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2052 
2053 	return 0;
2054 }
2055 
2056 static const struct net_device_ops efx_netdev_ops = {
2057 	.ndo_open		= efx_net_open,
2058 	.ndo_stop		= efx_net_stop,
2059 	.ndo_get_stats64	= efx_net_stats,
2060 	.ndo_tx_timeout		= efx_watchdog,
2061 	.ndo_start_xmit		= efx_hard_start_xmit,
2062 	.ndo_validate_addr	= eth_validate_addr,
2063 	.ndo_do_ioctl		= efx_ioctl,
2064 	.ndo_change_mtu		= efx_change_mtu,
2065 	.ndo_set_mac_address	= efx_set_mac_address,
2066 	.ndo_set_rx_mode	= efx_set_rx_mode,
2067 	.ndo_set_features	= efx_set_features,
2068 #ifdef CONFIG_SFC_SRIOV
2069 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2070 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2071 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2072 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2073 #endif
2074 #ifdef CONFIG_NET_POLL_CONTROLLER
2075 	.ndo_poll_controller = efx_netpoll,
2076 #endif
2077 	.ndo_setup_tc		= efx_setup_tc,
2078 #ifdef CONFIG_RFS_ACCEL
2079 	.ndo_rx_flow_steer	= efx_filter_rfs,
2080 #endif
2081 };
2082 
2083 static void efx_update_name(struct efx_nic *efx)
2084 {
2085 	strcpy(efx->name, efx->net_dev->name);
2086 	efx_mtd_rename(efx);
2087 	efx_set_channel_names(efx);
2088 }
2089 
2090 static int efx_netdev_event(struct notifier_block *this,
2091 			    unsigned long event, void *ptr)
2092 {
2093 	struct net_device *net_dev = ptr;
2094 
2095 	if (net_dev->netdev_ops == &efx_netdev_ops &&
2096 	    event == NETDEV_CHANGENAME)
2097 		efx_update_name(netdev_priv(net_dev));
2098 
2099 	return NOTIFY_DONE;
2100 }
2101 
2102 static struct notifier_block efx_netdev_notifier = {
2103 	.notifier_call = efx_netdev_event,
2104 };
2105 
2106 static ssize_t
2107 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2108 {
2109 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2110 	return sprintf(buf, "%d\n", efx->phy_type);
2111 }
2112 static DEVICE_ATTR(phy_type, 0644, show_phy_type, NULL);
2113 
2114 static int efx_register_netdev(struct efx_nic *efx)
2115 {
2116 	struct net_device *net_dev = efx->net_dev;
2117 	struct efx_channel *channel;
2118 	int rc;
2119 
2120 	net_dev->watchdog_timeo = 5 * HZ;
2121 	net_dev->irq = efx->pci_dev->irq;
2122 	net_dev->netdev_ops = &efx_netdev_ops;
2123 	SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
2124 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2125 
2126 	rtnl_lock();
2127 
2128 	/* Enable resets to be scheduled and check whether any were
2129 	 * already requested.  If so, the NIC is probably hosed so we
2130 	 * abort.
2131 	 */
2132 	efx->state = STATE_READY;
2133 	smp_mb(); /* ensure we change state before checking reset_pending */
2134 	if (efx->reset_pending) {
2135 		netif_err(efx, probe, efx->net_dev,
2136 			  "aborting probe due to scheduled reset\n");
2137 		rc = -EIO;
2138 		goto fail_locked;
2139 	}
2140 
2141 	rc = dev_alloc_name(net_dev, net_dev->name);
2142 	if (rc < 0)
2143 		goto fail_locked;
2144 	efx_update_name(efx);
2145 
2146 	/* Always start with carrier off; PHY events will detect the link */
2147 	netif_carrier_off(net_dev);
2148 
2149 	rc = register_netdevice(net_dev);
2150 	if (rc)
2151 		goto fail_locked;
2152 
2153 	efx_for_each_channel(channel, efx) {
2154 		struct efx_tx_queue *tx_queue;
2155 		efx_for_each_channel_tx_queue(tx_queue, channel)
2156 			efx_init_tx_queue_core_txq(tx_queue);
2157 	}
2158 
2159 	rtnl_unlock();
2160 
2161 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2162 	if (rc) {
2163 		netif_err(efx, drv, efx->net_dev,
2164 			  "failed to init net dev attributes\n");
2165 		goto fail_registered;
2166 	}
2167 
2168 	return 0;
2169 
2170 fail_registered:
2171 	rtnl_lock();
2172 	unregister_netdevice(net_dev);
2173 fail_locked:
2174 	efx->state = STATE_UNINIT;
2175 	rtnl_unlock();
2176 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2177 	return rc;
2178 }
2179 
2180 static void efx_unregister_netdev(struct efx_nic *efx)
2181 {
2182 	struct efx_channel *channel;
2183 	struct efx_tx_queue *tx_queue;
2184 
2185 	if (!efx->net_dev)
2186 		return;
2187 
2188 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2189 
2190 	/* Free up any skbs still remaining. This has to happen before
2191 	 * we try to unregister the netdev as running their destructors
2192 	 * may be needed to get the device ref. count to 0. */
2193 	efx_for_each_channel(channel, efx) {
2194 		efx_for_each_channel_tx_queue(tx_queue, channel)
2195 			efx_release_tx_buffers(tx_queue);
2196 	}
2197 
2198 	strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2199 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2200 
2201 	rtnl_lock();
2202 	unregister_netdevice(efx->net_dev);
2203 	efx->state = STATE_UNINIT;
2204 	rtnl_unlock();
2205 }
2206 
2207 /**************************************************************************
2208  *
2209  * Device reset and suspend
2210  *
2211  **************************************************************************/
2212 
2213 /* Tears down the entire software state and most of the hardware state
2214  * before reset.  */
2215 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2216 {
2217 	EFX_ASSERT_RESET_SERIALISED(efx);
2218 
2219 	efx_stop_all(efx);
2220 	efx_stop_interrupts(efx, false);
2221 
2222 	mutex_lock(&efx->mac_lock);
2223 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2224 		efx->phy_op->fini(efx);
2225 	efx->type->fini(efx);
2226 }
2227 
2228 /* This function will always ensure that the locks acquired in
2229  * efx_reset_down() are released. A failure return code indicates
2230  * that we were unable to reinitialise the hardware, and the
2231  * driver should be disabled. If ok is false, then the rx and tx
2232  * engines are not restarted, pending a RESET_DISABLE. */
2233 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2234 {
2235 	int rc;
2236 
2237 	EFX_ASSERT_RESET_SERIALISED(efx);
2238 
2239 	rc = efx->type->init(efx);
2240 	if (rc) {
2241 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2242 		goto fail;
2243 	}
2244 
2245 	if (!ok)
2246 		goto fail;
2247 
2248 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2249 		rc = efx->phy_op->init(efx);
2250 		if (rc)
2251 			goto fail;
2252 		if (efx->phy_op->reconfigure(efx))
2253 			netif_err(efx, drv, efx->net_dev,
2254 				  "could not restore PHY settings\n");
2255 	}
2256 
2257 	efx->type->reconfigure_mac(efx);
2258 
2259 	efx_start_interrupts(efx, false);
2260 	efx_restore_filters(efx);
2261 	efx_sriov_reset(efx);
2262 
2263 	mutex_unlock(&efx->mac_lock);
2264 
2265 	efx_start_all(efx);
2266 
2267 	return 0;
2268 
2269 fail:
2270 	efx->port_initialized = false;
2271 
2272 	mutex_unlock(&efx->mac_lock);
2273 
2274 	return rc;
2275 }
2276 
2277 /* Reset the NIC using the specified method.  Note that the reset may
2278  * fail, in which case the card will be left in an unusable state.
2279  *
2280  * Caller must hold the rtnl_lock.
2281  */
2282 int efx_reset(struct efx_nic *efx, enum reset_type method)
2283 {
2284 	int rc, rc2;
2285 	bool disabled;
2286 
2287 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2288 		   RESET_TYPE(method));
2289 
2290 	efx_device_detach_sync(efx);
2291 	efx_reset_down(efx, method);
2292 
2293 	rc = efx->type->reset(efx, method);
2294 	if (rc) {
2295 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2296 		goto out;
2297 	}
2298 
2299 	/* Clear flags for the scopes we covered.  We assume the NIC and
2300 	 * driver are now quiescent so that there is no race here.
2301 	 */
2302 	efx->reset_pending &= -(1 << (method + 1));
2303 
2304 	/* Reinitialise bus-mastering, which may have been turned off before
2305 	 * the reset was scheduled. This is still appropriate, even in the
2306 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2307 	 * can respond to requests. */
2308 	pci_set_master(efx->pci_dev);
2309 
2310 out:
2311 	/* Leave device stopped if necessary */
2312 	disabled = rc || method == RESET_TYPE_DISABLE;
2313 	rc2 = efx_reset_up(efx, method, !disabled);
2314 	if (rc2) {
2315 		disabled = true;
2316 		if (!rc)
2317 			rc = rc2;
2318 	}
2319 
2320 	if (disabled) {
2321 		dev_close(efx->net_dev);
2322 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2323 		efx->state = STATE_DISABLED;
2324 	} else {
2325 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2326 		netif_device_attach(efx->net_dev);
2327 	}
2328 	return rc;
2329 }
2330 
2331 /* The worker thread exists so that code that cannot sleep can
2332  * schedule a reset for later.
2333  */
2334 static void efx_reset_work(struct work_struct *data)
2335 {
2336 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2337 	unsigned long pending = ACCESS_ONCE(efx->reset_pending);
2338 
2339 	if (!pending)
2340 		return;
2341 
2342 	rtnl_lock();
2343 
2344 	/* We checked the state in efx_schedule_reset() but it may
2345 	 * have changed by now.  Now that we have the RTNL lock,
2346 	 * it cannot change again.
2347 	 */
2348 	if (efx->state == STATE_READY)
2349 		(void)efx_reset(efx, fls(pending) - 1);
2350 
2351 	rtnl_unlock();
2352 }
2353 
2354 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2355 {
2356 	enum reset_type method;
2357 
2358 	switch (type) {
2359 	case RESET_TYPE_INVISIBLE:
2360 	case RESET_TYPE_ALL:
2361 	case RESET_TYPE_WORLD:
2362 	case RESET_TYPE_DISABLE:
2363 		method = type;
2364 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2365 			  RESET_TYPE(method));
2366 		break;
2367 	default:
2368 		method = efx->type->map_reset_reason(type);
2369 		netif_dbg(efx, drv, efx->net_dev,
2370 			  "scheduling %s reset for %s\n",
2371 			  RESET_TYPE(method), RESET_TYPE(type));
2372 		break;
2373 	}
2374 
2375 	set_bit(method, &efx->reset_pending);
2376 	smp_mb(); /* ensure we change reset_pending before checking state */
2377 
2378 	/* If we're not READY then just leave the flags set as the cue
2379 	 * to abort probing or reschedule the reset later.
2380 	 */
2381 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2382 		return;
2383 
2384 	/* efx_process_channel() will no longer read events once a
2385 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2386 	efx_mcdi_mode_poll(efx);
2387 
2388 	queue_work(reset_workqueue, &efx->reset_work);
2389 }
2390 
2391 /**************************************************************************
2392  *
2393  * List of NICs we support
2394  *
2395  **************************************************************************/
2396 
2397 /* PCI device ID table */
2398 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2399 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2400 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2401 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2402 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2403 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2404 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2405 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2406 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2407 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2408 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2409 	{0}			/* end of list */
2410 };
2411 
2412 /**************************************************************************
2413  *
2414  * Dummy PHY/MAC operations
2415  *
2416  * Can be used for some unimplemented operations
2417  * Needed so all function pointers are valid and do not have to be tested
2418  * before use
2419  *
2420  **************************************************************************/
2421 int efx_port_dummy_op_int(struct efx_nic *efx)
2422 {
2423 	return 0;
2424 }
2425 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2426 
2427 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2428 {
2429 	return false;
2430 }
2431 
2432 static const struct efx_phy_operations efx_dummy_phy_operations = {
2433 	.init		 = efx_port_dummy_op_int,
2434 	.reconfigure	 = efx_port_dummy_op_int,
2435 	.poll		 = efx_port_dummy_op_poll,
2436 	.fini		 = efx_port_dummy_op_void,
2437 };
2438 
2439 /**************************************************************************
2440  *
2441  * Data housekeeping
2442  *
2443  **************************************************************************/
2444 
2445 /* This zeroes out and then fills in the invariants in a struct
2446  * efx_nic (including all sub-structures).
2447  */
2448 static int efx_init_struct(struct efx_nic *efx,
2449 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2450 {
2451 	int i;
2452 
2453 	/* Initialise common structures */
2454 	spin_lock_init(&efx->biu_lock);
2455 #ifdef CONFIG_SFC_MTD
2456 	INIT_LIST_HEAD(&efx->mtd_list);
2457 #endif
2458 	INIT_WORK(&efx->reset_work, efx_reset_work);
2459 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2460 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2461 	efx->pci_dev = pci_dev;
2462 	efx->msg_enable = debug;
2463 	efx->state = STATE_UNINIT;
2464 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2465 
2466 	efx->net_dev = net_dev;
2467 	spin_lock_init(&efx->stats_lock);
2468 	mutex_init(&efx->mac_lock);
2469 	efx->phy_op = &efx_dummy_phy_operations;
2470 	efx->mdio.dev = net_dev;
2471 	INIT_WORK(&efx->mac_work, efx_mac_work);
2472 	init_waitqueue_head(&efx->flush_wq);
2473 
2474 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2475 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2476 		if (!efx->channel[i])
2477 			goto fail;
2478 	}
2479 
2480 	EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS);
2481 
2482 	/* Higher numbered interrupt modes are less capable! */
2483 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2484 				  interrupt_mode);
2485 
2486 	/* Would be good to use the net_dev name, but we're too early */
2487 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2488 		 pci_name(pci_dev));
2489 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2490 	if (!efx->workqueue)
2491 		goto fail;
2492 
2493 	return 0;
2494 
2495 fail:
2496 	efx_fini_struct(efx);
2497 	return -ENOMEM;
2498 }
2499 
2500 static void efx_fini_struct(struct efx_nic *efx)
2501 {
2502 	int i;
2503 
2504 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2505 		kfree(efx->channel[i]);
2506 
2507 	if (efx->workqueue) {
2508 		destroy_workqueue(efx->workqueue);
2509 		efx->workqueue = NULL;
2510 	}
2511 }
2512 
2513 /**************************************************************************
2514  *
2515  * PCI interface
2516  *
2517  **************************************************************************/
2518 
2519 /* Main body of final NIC shutdown code
2520  * This is called only at module unload (or hotplug removal).
2521  */
2522 static void efx_pci_remove_main(struct efx_nic *efx)
2523 {
2524 	/* Flush reset_work. It can no longer be scheduled since we
2525 	 * are not READY.
2526 	 */
2527 	BUG_ON(efx->state == STATE_READY);
2528 	cancel_work_sync(&efx->reset_work);
2529 
2530 #ifdef CONFIG_RFS_ACCEL
2531 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
2532 	efx->net_dev->rx_cpu_rmap = NULL;
2533 #endif
2534 	efx_stop_interrupts(efx, false);
2535 	efx_nic_fini_interrupt(efx);
2536 	efx_fini_port(efx);
2537 	efx->type->fini(efx);
2538 	efx_fini_napi(efx);
2539 	efx_remove_all(efx);
2540 }
2541 
2542 /* Final NIC shutdown
2543  * This is called only at module unload (or hotplug removal).
2544  */
2545 static void efx_pci_remove(struct pci_dev *pci_dev)
2546 {
2547 	struct efx_nic *efx;
2548 
2549 	efx = pci_get_drvdata(pci_dev);
2550 	if (!efx)
2551 		return;
2552 
2553 	/* Mark the NIC as fini, then stop the interface */
2554 	rtnl_lock();
2555 	dev_close(efx->net_dev);
2556 	efx_stop_interrupts(efx, false);
2557 	rtnl_unlock();
2558 
2559 	efx_sriov_fini(efx);
2560 	efx_unregister_netdev(efx);
2561 
2562 	efx_mtd_remove(efx);
2563 
2564 	efx_pci_remove_main(efx);
2565 
2566 	efx_fini_io(efx);
2567 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2568 
2569 	efx_fini_struct(efx);
2570 	pci_set_drvdata(pci_dev, NULL);
2571 	free_netdev(efx->net_dev);
2572 };
2573 
2574 /* NIC VPD information
2575  * Called during probe to display the part number of the
2576  * installed NIC.  VPD is potentially very large but this should
2577  * always appear within the first 512 bytes.
2578  */
2579 #define SFC_VPD_LEN 512
2580 static void efx_print_product_vpd(struct efx_nic *efx)
2581 {
2582 	struct pci_dev *dev = efx->pci_dev;
2583 	char vpd_data[SFC_VPD_LEN];
2584 	ssize_t vpd_size;
2585 	int i, j;
2586 
2587 	/* Get the vpd data from the device */
2588 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2589 	if (vpd_size <= 0) {
2590 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2591 		return;
2592 	}
2593 
2594 	/* Get the Read only section */
2595 	i = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2596 	if (i < 0) {
2597 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2598 		return;
2599 	}
2600 
2601 	j = pci_vpd_lrdt_size(&vpd_data[i]);
2602 	i += PCI_VPD_LRDT_TAG_SIZE;
2603 	if (i + j > vpd_size)
2604 		j = vpd_size - i;
2605 
2606 	/* Get the Part number */
2607 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2608 	if (i < 0) {
2609 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2610 		return;
2611 	}
2612 
2613 	j = pci_vpd_info_field_size(&vpd_data[i]);
2614 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
2615 	if (i + j > vpd_size) {
2616 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2617 		return;
2618 	}
2619 
2620 	netif_info(efx, drv, efx->net_dev,
2621 		   "Part Number : %.*s\n", j, &vpd_data[i]);
2622 }
2623 
2624 
2625 /* Main body of NIC initialisation
2626  * This is called at module load (or hotplug insertion, theoretically).
2627  */
2628 static int efx_pci_probe_main(struct efx_nic *efx)
2629 {
2630 	int rc;
2631 
2632 	/* Do start-of-day initialisation */
2633 	rc = efx_probe_all(efx);
2634 	if (rc)
2635 		goto fail1;
2636 
2637 	efx_init_napi(efx);
2638 
2639 	rc = efx->type->init(efx);
2640 	if (rc) {
2641 		netif_err(efx, probe, efx->net_dev,
2642 			  "failed to initialise NIC\n");
2643 		goto fail3;
2644 	}
2645 
2646 	rc = efx_init_port(efx);
2647 	if (rc) {
2648 		netif_err(efx, probe, efx->net_dev,
2649 			  "failed to initialise port\n");
2650 		goto fail4;
2651 	}
2652 
2653 	rc = efx_nic_init_interrupt(efx);
2654 	if (rc)
2655 		goto fail5;
2656 	efx_start_interrupts(efx, false);
2657 
2658 	return 0;
2659 
2660  fail5:
2661 	efx_fini_port(efx);
2662  fail4:
2663 	efx->type->fini(efx);
2664  fail3:
2665 	efx_fini_napi(efx);
2666 	efx_remove_all(efx);
2667  fail1:
2668 	return rc;
2669 }
2670 
2671 /* NIC initialisation
2672  *
2673  * This is called at module load (or hotplug insertion,
2674  * theoretically).  It sets up PCI mappings, resets the NIC,
2675  * sets up and registers the network devices with the kernel and hooks
2676  * the interrupt service routine.  It does not prepare the device for
2677  * transmission; this is left to the first time one of the network
2678  * interfaces is brought up (i.e. efx_net_open).
2679  */
2680 static int efx_pci_probe(struct pci_dev *pci_dev,
2681 			 const struct pci_device_id *entry)
2682 {
2683 	struct net_device *net_dev;
2684 	struct efx_nic *efx;
2685 	int rc;
2686 
2687 	/* Allocate and initialise a struct net_device and struct efx_nic */
2688 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2689 				     EFX_MAX_RX_QUEUES);
2690 	if (!net_dev)
2691 		return -ENOMEM;
2692 	efx = netdev_priv(net_dev);
2693 	efx->type = (const struct efx_nic_type *) entry->driver_data;
2694 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2695 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
2696 			      NETIF_F_RXCSUM);
2697 	if (efx->type->offload_features & NETIF_F_V6_CSUM)
2698 		net_dev->features |= NETIF_F_TSO6;
2699 	/* Mask for features that also apply to VLAN devices */
2700 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2701 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2702 				   NETIF_F_RXCSUM);
2703 	/* All offloads can be toggled */
2704 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2705 	pci_set_drvdata(pci_dev, efx);
2706 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2707 	rc = efx_init_struct(efx, pci_dev, net_dev);
2708 	if (rc)
2709 		goto fail1;
2710 
2711 	netif_info(efx, probe, efx->net_dev,
2712 		   "Solarflare NIC detected\n");
2713 
2714 	efx_print_product_vpd(efx);
2715 
2716 	/* Set up basic I/O (BAR mappings etc) */
2717 	rc = efx_init_io(efx);
2718 	if (rc)
2719 		goto fail2;
2720 
2721 	rc = efx_pci_probe_main(efx);
2722 	if (rc)
2723 		goto fail3;
2724 
2725 	rc = efx_register_netdev(efx);
2726 	if (rc)
2727 		goto fail4;
2728 
2729 	rc = efx_sriov_init(efx);
2730 	if (rc)
2731 		netif_err(efx, probe, efx->net_dev,
2732 			  "SR-IOV can't be enabled rc %d\n", rc);
2733 
2734 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2735 
2736 	/* Try to create MTDs, but allow this to fail */
2737 	rtnl_lock();
2738 	rc = efx_mtd_probe(efx);
2739 	rtnl_unlock();
2740 	if (rc)
2741 		netif_warn(efx, probe, efx->net_dev,
2742 			   "failed to create MTDs (%d)\n", rc);
2743 
2744 	return 0;
2745 
2746  fail4:
2747 	efx_pci_remove_main(efx);
2748  fail3:
2749 	efx_fini_io(efx);
2750  fail2:
2751 	efx_fini_struct(efx);
2752  fail1:
2753 	pci_set_drvdata(pci_dev, NULL);
2754 	WARN_ON(rc > 0);
2755 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2756 	free_netdev(net_dev);
2757 	return rc;
2758 }
2759 
2760 static int efx_pm_freeze(struct device *dev)
2761 {
2762 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2763 
2764 	rtnl_lock();
2765 
2766 	if (efx->state != STATE_DISABLED) {
2767 		efx->state = STATE_UNINIT;
2768 
2769 		efx_device_detach_sync(efx);
2770 
2771 		efx_stop_all(efx);
2772 		efx_stop_interrupts(efx, false);
2773 	}
2774 
2775 	rtnl_unlock();
2776 
2777 	return 0;
2778 }
2779 
2780 static int efx_pm_thaw(struct device *dev)
2781 {
2782 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2783 
2784 	rtnl_lock();
2785 
2786 	if (efx->state != STATE_DISABLED) {
2787 		efx_start_interrupts(efx, false);
2788 
2789 		mutex_lock(&efx->mac_lock);
2790 		efx->phy_op->reconfigure(efx);
2791 		mutex_unlock(&efx->mac_lock);
2792 
2793 		efx_start_all(efx);
2794 
2795 		netif_device_attach(efx->net_dev);
2796 
2797 		efx->state = STATE_READY;
2798 
2799 		efx->type->resume_wol(efx);
2800 	}
2801 
2802 	rtnl_unlock();
2803 
2804 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2805 	queue_work(reset_workqueue, &efx->reset_work);
2806 
2807 	return 0;
2808 }
2809 
2810 static int efx_pm_poweroff(struct device *dev)
2811 {
2812 	struct pci_dev *pci_dev = to_pci_dev(dev);
2813 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2814 
2815 	efx->type->fini(efx);
2816 
2817 	efx->reset_pending = 0;
2818 
2819 	pci_save_state(pci_dev);
2820 	return pci_set_power_state(pci_dev, PCI_D3hot);
2821 }
2822 
2823 /* Used for both resume and restore */
2824 static int efx_pm_resume(struct device *dev)
2825 {
2826 	struct pci_dev *pci_dev = to_pci_dev(dev);
2827 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2828 	int rc;
2829 
2830 	rc = pci_set_power_state(pci_dev, PCI_D0);
2831 	if (rc)
2832 		return rc;
2833 	pci_restore_state(pci_dev);
2834 	rc = pci_enable_device(pci_dev);
2835 	if (rc)
2836 		return rc;
2837 	pci_set_master(efx->pci_dev);
2838 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
2839 	if (rc)
2840 		return rc;
2841 	rc = efx->type->init(efx);
2842 	if (rc)
2843 		return rc;
2844 	efx_pm_thaw(dev);
2845 	return 0;
2846 }
2847 
2848 static int efx_pm_suspend(struct device *dev)
2849 {
2850 	int rc;
2851 
2852 	efx_pm_freeze(dev);
2853 	rc = efx_pm_poweroff(dev);
2854 	if (rc)
2855 		efx_pm_resume(dev);
2856 	return rc;
2857 }
2858 
2859 static const struct dev_pm_ops efx_pm_ops = {
2860 	.suspend	= efx_pm_suspend,
2861 	.resume		= efx_pm_resume,
2862 	.freeze		= efx_pm_freeze,
2863 	.thaw		= efx_pm_thaw,
2864 	.poweroff	= efx_pm_poweroff,
2865 	.restore	= efx_pm_resume,
2866 };
2867 
2868 static struct pci_driver efx_pci_driver = {
2869 	.name		= KBUILD_MODNAME,
2870 	.id_table	= efx_pci_table,
2871 	.probe		= efx_pci_probe,
2872 	.remove		= efx_pci_remove,
2873 	.driver.pm	= &efx_pm_ops,
2874 };
2875 
2876 /**************************************************************************
2877  *
2878  * Kernel module interface
2879  *
2880  *************************************************************************/
2881 
2882 module_param(interrupt_mode, uint, 0444);
2883 MODULE_PARM_DESC(interrupt_mode,
2884 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2885 
2886 static int __init efx_init_module(void)
2887 {
2888 	int rc;
2889 
2890 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
2891 
2892 	rc = register_netdevice_notifier(&efx_netdev_notifier);
2893 	if (rc)
2894 		goto err_notifier;
2895 
2896 	rc = efx_init_sriov();
2897 	if (rc)
2898 		goto err_sriov;
2899 
2900 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
2901 	if (!reset_workqueue) {
2902 		rc = -ENOMEM;
2903 		goto err_reset;
2904 	}
2905 
2906 	rc = pci_register_driver(&efx_pci_driver);
2907 	if (rc < 0)
2908 		goto err_pci;
2909 
2910 	return 0;
2911 
2912  err_pci:
2913 	destroy_workqueue(reset_workqueue);
2914  err_reset:
2915 	efx_fini_sriov();
2916  err_sriov:
2917 	unregister_netdevice_notifier(&efx_netdev_notifier);
2918  err_notifier:
2919 	return rc;
2920 }
2921 
2922 static void __exit efx_exit_module(void)
2923 {
2924 	printk(KERN_INFO "Solarflare NET driver unloading\n");
2925 
2926 	pci_unregister_driver(&efx_pci_driver);
2927 	destroy_workqueue(reset_workqueue);
2928 	efx_fini_sriov();
2929 	unregister_netdevice_notifier(&efx_netdev_notifier);
2930 
2931 }
2932 
2933 module_init(efx_init_module);
2934 module_exit(efx_exit_module);
2935 
2936 MODULE_AUTHOR("Solarflare Communications and "
2937 	      "Michael Brown <mbrown@fensystems.co.uk>");
2938 MODULE_DESCRIPTION("Solarflare Communications network driver");
2939 MODULE_LICENSE("GPL");
2940 MODULE_DEVICE_TABLE(pci, efx_pci_table);
2941