xref: /linux/drivers/net/ethernet/sfc/efx.c (revision 95e9fd10f06cb5642028b6b851e32b8c8afb4571)
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 unsigned int separate_tx_channels;
110 module_param(separate_tx_channels, uint, 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 int phy_flash_cfg;
164 module_param(phy_flash_cfg, int, 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_RUNNING) ||	\
206 		    (efx->state == STATE_DISABLED))	\
207 			ASSERT_RTNL();			\
208 	} while (0)
209 
210 /**************************************************************************
211  *
212  * Event queue processing
213  *
214  *************************************************************************/
215 
216 /* Process channel's event queue
217  *
218  * This function is responsible for processing the event queue of a
219  * single channel.  The caller must guarantee that this function will
220  * never be concurrently called more than once on the same channel,
221  * though different channels may be being processed concurrently.
222  */
223 static int efx_process_channel(struct efx_channel *channel, int budget)
224 {
225 	int spent;
226 
227 	if (unlikely(!channel->enabled))
228 		return 0;
229 
230 	spent = efx_nic_process_eventq(channel, budget);
231 	if (spent && efx_channel_has_rx_queue(channel)) {
232 		struct efx_rx_queue *rx_queue =
233 			efx_channel_get_rx_queue(channel);
234 
235 		/* Deliver last RX packet. */
236 		if (channel->rx_pkt) {
237 			__efx_rx_packet(channel, channel->rx_pkt);
238 			channel->rx_pkt = NULL;
239 		}
240 		if (rx_queue->enabled) {
241 			efx_rx_strategy(channel);
242 			efx_fast_push_rx_descriptors(rx_queue);
243 		}
244 	}
245 
246 	return spent;
247 }
248 
249 /* Mark channel as finished processing
250  *
251  * Note that since we will not receive further interrupts for this
252  * channel before we finish processing and call the eventq_read_ack()
253  * method, there is no need to use the interrupt hold-off timers.
254  */
255 static inline void efx_channel_processed(struct efx_channel *channel)
256 {
257 	/* The interrupt handler for this channel may set work_pending
258 	 * as soon as we acknowledge the events we've seen.  Make sure
259 	 * it's cleared before then. */
260 	channel->work_pending = false;
261 	smp_wmb();
262 
263 	efx_nic_eventq_read_ack(channel);
264 }
265 
266 /* NAPI poll handler
267  *
268  * NAPI guarantees serialisation of polls of the same device, which
269  * provides the guarantee required by efx_process_channel().
270  */
271 static int efx_poll(struct napi_struct *napi, int budget)
272 {
273 	struct efx_channel *channel =
274 		container_of(napi, struct efx_channel, napi_str);
275 	struct efx_nic *efx = channel->efx;
276 	int spent;
277 
278 	netif_vdbg(efx, intr, efx->net_dev,
279 		   "channel %d NAPI poll executing on CPU %d\n",
280 		   channel->channel, raw_smp_processor_id());
281 
282 	spent = efx_process_channel(channel, budget);
283 
284 	if (spent < budget) {
285 		if (efx_channel_has_rx_queue(channel) &&
286 		    efx->irq_rx_adaptive &&
287 		    unlikely(++channel->irq_count == 1000)) {
288 			if (unlikely(channel->irq_mod_score <
289 				     irq_adapt_low_thresh)) {
290 				if (channel->irq_moderation > 1) {
291 					channel->irq_moderation -= 1;
292 					efx->type->push_irq_moderation(channel);
293 				}
294 			} else if (unlikely(channel->irq_mod_score >
295 					    irq_adapt_high_thresh)) {
296 				if (channel->irq_moderation <
297 				    efx->irq_rx_moderation) {
298 					channel->irq_moderation += 1;
299 					efx->type->push_irq_moderation(channel);
300 				}
301 			}
302 			channel->irq_count = 0;
303 			channel->irq_mod_score = 0;
304 		}
305 
306 		efx_filter_rfs_expire(channel);
307 
308 		/* There is no race here; although napi_disable() will
309 		 * only wait for napi_complete(), this isn't a problem
310 		 * since efx_channel_processed() will have no effect if
311 		 * interrupts have already been disabled.
312 		 */
313 		napi_complete(napi);
314 		efx_channel_processed(channel);
315 	}
316 
317 	return spent;
318 }
319 
320 /* Process the eventq of the specified channel immediately on this CPU
321  *
322  * Disable hardware generated interrupts, wait for any existing
323  * processing to finish, then directly poll (and ack ) the eventq.
324  * Finally reenable NAPI and interrupts.
325  *
326  * This is for use only during a loopback self-test.  It must not
327  * deliver any packets up the stack as this can result in deadlock.
328  */
329 void efx_process_channel_now(struct efx_channel *channel)
330 {
331 	struct efx_nic *efx = channel->efx;
332 
333 	BUG_ON(channel->channel >= efx->n_channels);
334 	BUG_ON(!channel->enabled);
335 	BUG_ON(!efx->loopback_selftest);
336 
337 	/* Disable interrupts and wait for ISRs to complete */
338 	efx_nic_disable_interrupts(efx);
339 	if (efx->legacy_irq) {
340 		synchronize_irq(efx->legacy_irq);
341 		efx->legacy_irq_enabled = false;
342 	}
343 	if (channel->irq)
344 		synchronize_irq(channel->irq);
345 
346 	/* Wait for any NAPI processing to complete */
347 	napi_disable(&channel->napi_str);
348 
349 	/* Poll the channel */
350 	efx_process_channel(channel, channel->eventq_mask + 1);
351 
352 	/* Ack the eventq. This may cause an interrupt to be generated
353 	 * when they are reenabled */
354 	efx_channel_processed(channel);
355 
356 	napi_enable(&channel->napi_str);
357 	if (efx->legacy_irq)
358 		efx->legacy_irq_enabled = true;
359 	efx_nic_enable_interrupts(efx);
360 }
361 
362 /* Create event queue
363  * Event queue memory allocations are done only once.  If the channel
364  * is reset, the memory buffer will be reused; this guards against
365  * errors during channel reset and also simplifies interrupt handling.
366  */
367 static int efx_probe_eventq(struct efx_channel *channel)
368 {
369 	struct efx_nic *efx = channel->efx;
370 	unsigned long entries;
371 
372 	netif_dbg(efx, probe, efx->net_dev,
373 		  "chan %d create event queue\n", channel->channel);
374 
375 	/* Build an event queue with room for one event per tx and rx buffer,
376 	 * plus some extra for link state events and MCDI completions. */
377 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
378 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
379 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
380 
381 	return efx_nic_probe_eventq(channel);
382 }
383 
384 /* Prepare channel's event queue */
385 static void efx_init_eventq(struct efx_channel *channel)
386 {
387 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
388 		  "chan %d init event queue\n", channel->channel);
389 
390 	channel->eventq_read_ptr = 0;
391 
392 	efx_nic_init_eventq(channel);
393 }
394 
395 /* Enable event queue processing and NAPI */
396 static void efx_start_eventq(struct efx_channel *channel)
397 {
398 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
399 		  "chan %d start event queue\n", channel->channel);
400 
401 	/* The interrupt handler for this channel may set work_pending
402 	 * as soon as we enable it.  Make sure it's cleared before
403 	 * then.  Similarly, make sure it sees the enabled flag set.
404 	 */
405 	channel->work_pending = false;
406 	channel->enabled = true;
407 	smp_wmb();
408 
409 	napi_enable(&channel->napi_str);
410 	efx_nic_eventq_read_ack(channel);
411 }
412 
413 /* Disable event queue processing and NAPI */
414 static void efx_stop_eventq(struct efx_channel *channel)
415 {
416 	if (!channel->enabled)
417 		return;
418 
419 	napi_disable(&channel->napi_str);
420 	channel->enabled = false;
421 }
422 
423 static void efx_fini_eventq(struct efx_channel *channel)
424 {
425 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
426 		  "chan %d fini event queue\n", channel->channel);
427 
428 	efx_nic_fini_eventq(channel);
429 }
430 
431 static void efx_remove_eventq(struct efx_channel *channel)
432 {
433 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
434 		  "chan %d remove event queue\n", channel->channel);
435 
436 	efx_nic_remove_eventq(channel);
437 }
438 
439 /**************************************************************************
440  *
441  * Channel handling
442  *
443  *************************************************************************/
444 
445 /* Allocate and initialise a channel structure. */
446 static struct efx_channel *
447 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
448 {
449 	struct efx_channel *channel;
450 	struct efx_rx_queue *rx_queue;
451 	struct efx_tx_queue *tx_queue;
452 	int j;
453 
454 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
455 	if (!channel)
456 		return NULL;
457 
458 	channel->efx = efx;
459 	channel->channel = i;
460 	channel->type = &efx_default_channel_type;
461 
462 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
463 		tx_queue = &channel->tx_queue[j];
464 		tx_queue->efx = efx;
465 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
466 		tx_queue->channel = channel;
467 	}
468 
469 	rx_queue = &channel->rx_queue;
470 	rx_queue->efx = efx;
471 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
472 		    (unsigned long)rx_queue);
473 
474 	return channel;
475 }
476 
477 /* Allocate and initialise a channel structure, copying parameters
478  * (but not resources) from an old channel structure.
479  */
480 static struct efx_channel *
481 efx_copy_channel(const struct efx_channel *old_channel)
482 {
483 	struct efx_channel *channel;
484 	struct efx_rx_queue *rx_queue;
485 	struct efx_tx_queue *tx_queue;
486 	int j;
487 
488 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
489 	if (!channel)
490 		return NULL;
491 
492 	*channel = *old_channel;
493 
494 	channel->napi_dev = NULL;
495 	memset(&channel->eventq, 0, sizeof(channel->eventq));
496 
497 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
498 		tx_queue = &channel->tx_queue[j];
499 		if (tx_queue->channel)
500 			tx_queue->channel = channel;
501 		tx_queue->buffer = NULL;
502 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
503 	}
504 
505 	rx_queue = &channel->rx_queue;
506 	rx_queue->buffer = NULL;
507 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
508 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
509 		    (unsigned long)rx_queue);
510 
511 	return channel;
512 }
513 
514 static int efx_probe_channel(struct efx_channel *channel)
515 {
516 	struct efx_tx_queue *tx_queue;
517 	struct efx_rx_queue *rx_queue;
518 	int rc;
519 
520 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
521 		  "creating channel %d\n", channel->channel);
522 
523 	rc = channel->type->pre_probe(channel);
524 	if (rc)
525 		goto fail;
526 
527 	rc = efx_probe_eventq(channel);
528 	if (rc)
529 		goto fail;
530 
531 	efx_for_each_channel_tx_queue(tx_queue, channel) {
532 		rc = efx_probe_tx_queue(tx_queue);
533 		if (rc)
534 			goto fail;
535 	}
536 
537 	efx_for_each_channel_rx_queue(rx_queue, channel) {
538 		rc = efx_probe_rx_queue(rx_queue);
539 		if (rc)
540 			goto fail;
541 	}
542 
543 	channel->n_rx_frm_trunc = 0;
544 
545 	return 0;
546 
547 fail:
548 	efx_remove_channel(channel);
549 	return rc;
550 }
551 
552 static void
553 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
554 {
555 	struct efx_nic *efx = channel->efx;
556 	const char *type;
557 	int number;
558 
559 	number = channel->channel;
560 	if (efx->tx_channel_offset == 0) {
561 		type = "";
562 	} else if (channel->channel < efx->tx_channel_offset) {
563 		type = "-rx";
564 	} else {
565 		type = "-tx";
566 		number -= efx->tx_channel_offset;
567 	}
568 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
569 }
570 
571 static void efx_set_channel_names(struct efx_nic *efx)
572 {
573 	struct efx_channel *channel;
574 
575 	efx_for_each_channel(channel, efx)
576 		channel->type->get_name(channel,
577 					efx->channel_name[channel->channel],
578 					sizeof(efx->channel_name[0]));
579 }
580 
581 static int efx_probe_channels(struct efx_nic *efx)
582 {
583 	struct efx_channel *channel;
584 	int rc;
585 
586 	/* Restart special buffer allocation */
587 	efx->next_buffer_table = 0;
588 
589 	/* Probe channels in reverse, so that any 'extra' channels
590 	 * use the start of the buffer table. This allows the traffic
591 	 * channels to be resized without moving them or wasting the
592 	 * entries before them.
593 	 */
594 	efx_for_each_channel_rev(channel, efx) {
595 		rc = efx_probe_channel(channel);
596 		if (rc) {
597 			netif_err(efx, probe, efx->net_dev,
598 				  "failed to create channel %d\n",
599 				  channel->channel);
600 			goto fail;
601 		}
602 	}
603 	efx_set_channel_names(efx);
604 
605 	return 0;
606 
607 fail:
608 	efx_remove_channels(efx);
609 	return rc;
610 }
611 
612 /* Channels are shutdown and reinitialised whilst the NIC is running
613  * to propagate configuration changes (mtu, checksum offload), or
614  * to clear hardware error conditions
615  */
616 static void efx_start_datapath(struct efx_nic *efx)
617 {
618 	struct efx_tx_queue *tx_queue;
619 	struct efx_rx_queue *rx_queue;
620 	struct efx_channel *channel;
621 
622 	/* Calculate the rx buffer allocation parameters required to
623 	 * support the current MTU, including padding for header
624 	 * alignment and overruns.
625 	 */
626 	efx->rx_buffer_len = (max(EFX_PAGE_IP_ALIGN, NET_IP_ALIGN) +
627 			      EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
628 			      efx->type->rx_buffer_hash_size +
629 			      efx->type->rx_buffer_padding);
630 	efx->rx_buffer_order = get_order(efx->rx_buffer_len +
631 					 sizeof(struct efx_rx_page_state));
632 
633 	/* Initialise the channels */
634 	efx_for_each_channel(channel, efx) {
635 		efx_for_each_channel_tx_queue(tx_queue, channel)
636 			efx_init_tx_queue(tx_queue);
637 
638 		/* The rx buffer allocation strategy is MTU dependent */
639 		efx_rx_strategy(channel);
640 
641 		efx_for_each_channel_rx_queue(rx_queue, channel) {
642 			efx_init_rx_queue(rx_queue);
643 			efx_nic_generate_fill_event(rx_queue);
644 		}
645 
646 		WARN_ON(channel->rx_pkt != NULL);
647 		efx_rx_strategy(channel);
648 	}
649 
650 	if (netif_device_present(efx->net_dev))
651 		netif_tx_wake_all_queues(efx->net_dev);
652 }
653 
654 static void efx_stop_datapath(struct efx_nic *efx)
655 {
656 	struct efx_channel *channel;
657 	struct efx_tx_queue *tx_queue;
658 	struct efx_rx_queue *rx_queue;
659 	struct pci_dev *dev = efx->pci_dev;
660 	int rc;
661 
662 	EFX_ASSERT_RESET_SERIALISED(efx);
663 	BUG_ON(efx->port_enabled);
664 
665 	/* Only perform flush if dma is enabled */
666 	if (dev->is_busmaster) {
667 		rc = efx_nic_flush_queues(efx);
668 
669 		if (rc && EFX_WORKAROUND_7803(efx)) {
670 			/* Schedule a reset to recover from the flush failure. The
671 			 * descriptor caches reference memory we're about to free,
672 			 * but falcon_reconfigure_mac_wrapper() won't reconnect
673 			 * the MACs because of the pending reset. */
674 			netif_err(efx, drv, efx->net_dev,
675 				  "Resetting to recover from flush failure\n");
676 			efx_schedule_reset(efx, RESET_TYPE_ALL);
677 		} else if (rc) {
678 			netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
679 		} else {
680 			netif_dbg(efx, drv, efx->net_dev,
681 				  "successfully flushed all queues\n");
682 		}
683 	}
684 
685 	efx_for_each_channel(channel, efx) {
686 		/* RX packet processing is pipelined, so wait for the
687 		 * NAPI handler to complete.  At least event queue 0
688 		 * might be kept active by non-data events, so don't
689 		 * use napi_synchronize() but actually disable NAPI
690 		 * temporarily.
691 		 */
692 		if (efx_channel_has_rx_queue(channel)) {
693 			efx_stop_eventq(channel);
694 			efx_start_eventq(channel);
695 		}
696 
697 		efx_for_each_channel_rx_queue(rx_queue, channel)
698 			efx_fini_rx_queue(rx_queue);
699 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
700 			efx_fini_tx_queue(tx_queue);
701 	}
702 }
703 
704 static void efx_remove_channel(struct efx_channel *channel)
705 {
706 	struct efx_tx_queue *tx_queue;
707 	struct efx_rx_queue *rx_queue;
708 
709 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
710 		  "destroy chan %d\n", channel->channel);
711 
712 	efx_for_each_channel_rx_queue(rx_queue, channel)
713 		efx_remove_rx_queue(rx_queue);
714 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
715 		efx_remove_tx_queue(tx_queue);
716 	efx_remove_eventq(channel);
717 }
718 
719 static void efx_remove_channels(struct efx_nic *efx)
720 {
721 	struct efx_channel *channel;
722 
723 	efx_for_each_channel(channel, efx)
724 		efx_remove_channel(channel);
725 }
726 
727 int
728 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
729 {
730 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
731 	u32 old_rxq_entries, old_txq_entries;
732 	unsigned i, next_buffer_table = 0;
733 	int rc = 0;
734 
735 	/* Not all channels should be reallocated. We must avoid
736 	 * reallocating their buffer table entries.
737 	 */
738 	efx_for_each_channel(channel, efx) {
739 		struct efx_rx_queue *rx_queue;
740 		struct efx_tx_queue *tx_queue;
741 
742 		if (channel->type->copy)
743 			continue;
744 		next_buffer_table = max(next_buffer_table,
745 					channel->eventq.index +
746 					channel->eventq.entries);
747 		efx_for_each_channel_rx_queue(rx_queue, channel)
748 			next_buffer_table = max(next_buffer_table,
749 						rx_queue->rxd.index +
750 						rx_queue->rxd.entries);
751 		efx_for_each_channel_tx_queue(tx_queue, channel)
752 			next_buffer_table = max(next_buffer_table,
753 						tx_queue->txd.index +
754 						tx_queue->txd.entries);
755 	}
756 
757 	efx_stop_all(efx);
758 	efx_stop_interrupts(efx, true);
759 
760 	/* Clone channels (where possible) */
761 	memset(other_channel, 0, sizeof(other_channel));
762 	for (i = 0; i < efx->n_channels; i++) {
763 		channel = efx->channel[i];
764 		if (channel->type->copy)
765 			channel = channel->type->copy(channel);
766 		if (!channel) {
767 			rc = -ENOMEM;
768 			goto out;
769 		}
770 		other_channel[i] = channel;
771 	}
772 
773 	/* Swap entry counts and channel pointers */
774 	old_rxq_entries = efx->rxq_entries;
775 	old_txq_entries = efx->txq_entries;
776 	efx->rxq_entries = rxq_entries;
777 	efx->txq_entries = txq_entries;
778 	for (i = 0; i < efx->n_channels; i++) {
779 		channel = efx->channel[i];
780 		efx->channel[i] = other_channel[i];
781 		other_channel[i] = channel;
782 	}
783 
784 	/* Restart buffer table allocation */
785 	efx->next_buffer_table = next_buffer_table;
786 
787 	for (i = 0; i < efx->n_channels; i++) {
788 		channel = efx->channel[i];
789 		if (!channel->type->copy)
790 			continue;
791 		rc = efx_probe_channel(channel);
792 		if (rc)
793 			goto rollback;
794 		efx_init_napi_channel(efx->channel[i]);
795 	}
796 
797 out:
798 	/* Destroy unused channel structures */
799 	for (i = 0; i < efx->n_channels; i++) {
800 		channel = other_channel[i];
801 		if (channel && channel->type->copy) {
802 			efx_fini_napi_channel(channel);
803 			efx_remove_channel(channel);
804 			kfree(channel);
805 		}
806 	}
807 
808 	efx_start_interrupts(efx, true);
809 	efx_start_all(efx);
810 	return rc;
811 
812 rollback:
813 	/* Swap back */
814 	efx->rxq_entries = old_rxq_entries;
815 	efx->txq_entries = old_txq_entries;
816 	for (i = 0; i < efx->n_channels; i++) {
817 		channel = efx->channel[i];
818 		efx->channel[i] = other_channel[i];
819 		other_channel[i] = channel;
820 	}
821 	goto out;
822 }
823 
824 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
825 {
826 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
827 }
828 
829 static const struct efx_channel_type efx_default_channel_type = {
830 	.pre_probe		= efx_channel_dummy_op_int,
831 	.get_name		= efx_get_channel_name,
832 	.copy			= efx_copy_channel,
833 	.keep_eventq		= false,
834 };
835 
836 int efx_channel_dummy_op_int(struct efx_channel *channel)
837 {
838 	return 0;
839 }
840 
841 /**************************************************************************
842  *
843  * Port handling
844  *
845  **************************************************************************/
846 
847 /* This ensures that the kernel is kept informed (via
848  * netif_carrier_on/off) of the link status, and also maintains the
849  * link status's stop on the port's TX queue.
850  */
851 void efx_link_status_changed(struct efx_nic *efx)
852 {
853 	struct efx_link_state *link_state = &efx->link_state;
854 
855 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
856 	 * that no events are triggered between unregister_netdev() and the
857 	 * driver unloading. A more general condition is that NETDEV_CHANGE
858 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
859 	if (!netif_running(efx->net_dev))
860 		return;
861 
862 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
863 		efx->n_link_state_changes++;
864 
865 		if (link_state->up)
866 			netif_carrier_on(efx->net_dev);
867 		else
868 			netif_carrier_off(efx->net_dev);
869 	}
870 
871 	/* Status message for kernel log */
872 	if (link_state->up)
873 		netif_info(efx, link, efx->net_dev,
874 			   "link up at %uMbps %s-duplex (MTU %d)%s\n",
875 			   link_state->speed, link_state->fd ? "full" : "half",
876 			   efx->net_dev->mtu,
877 			   (efx->promiscuous ? " [PROMISC]" : ""));
878 	else
879 		netif_info(efx, link, efx->net_dev, "link down\n");
880 }
881 
882 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
883 {
884 	efx->link_advertising = advertising;
885 	if (advertising) {
886 		if (advertising & ADVERTISED_Pause)
887 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
888 		else
889 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
890 		if (advertising & ADVERTISED_Asym_Pause)
891 			efx->wanted_fc ^= EFX_FC_TX;
892 	}
893 }
894 
895 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
896 {
897 	efx->wanted_fc = wanted_fc;
898 	if (efx->link_advertising) {
899 		if (wanted_fc & EFX_FC_RX)
900 			efx->link_advertising |= (ADVERTISED_Pause |
901 						  ADVERTISED_Asym_Pause);
902 		else
903 			efx->link_advertising &= ~(ADVERTISED_Pause |
904 						   ADVERTISED_Asym_Pause);
905 		if (wanted_fc & EFX_FC_TX)
906 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
907 	}
908 }
909 
910 static void efx_fini_port(struct efx_nic *efx);
911 
912 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
913  * the MAC appropriately. All other PHY configuration changes are pushed
914  * through phy_op->set_settings(), and pushed asynchronously to the MAC
915  * through efx_monitor().
916  *
917  * Callers must hold the mac_lock
918  */
919 int __efx_reconfigure_port(struct efx_nic *efx)
920 {
921 	enum efx_phy_mode phy_mode;
922 	int rc;
923 
924 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
925 
926 	/* Serialise the promiscuous flag with efx_set_rx_mode. */
927 	netif_addr_lock_bh(efx->net_dev);
928 	netif_addr_unlock_bh(efx->net_dev);
929 
930 	/* Disable PHY transmit in mac level loopbacks */
931 	phy_mode = efx->phy_mode;
932 	if (LOOPBACK_INTERNAL(efx))
933 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
934 	else
935 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
936 
937 	rc = efx->type->reconfigure_port(efx);
938 
939 	if (rc)
940 		efx->phy_mode = phy_mode;
941 
942 	return rc;
943 }
944 
945 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
946  * disabled. */
947 int efx_reconfigure_port(struct efx_nic *efx)
948 {
949 	int rc;
950 
951 	EFX_ASSERT_RESET_SERIALISED(efx);
952 
953 	mutex_lock(&efx->mac_lock);
954 	rc = __efx_reconfigure_port(efx);
955 	mutex_unlock(&efx->mac_lock);
956 
957 	return rc;
958 }
959 
960 /* Asynchronous work item for changing MAC promiscuity and multicast
961  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
962  * MAC directly. */
963 static void efx_mac_work(struct work_struct *data)
964 {
965 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
966 
967 	mutex_lock(&efx->mac_lock);
968 	if (efx->port_enabled)
969 		efx->type->reconfigure_mac(efx);
970 	mutex_unlock(&efx->mac_lock);
971 }
972 
973 static int efx_probe_port(struct efx_nic *efx)
974 {
975 	int rc;
976 
977 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
978 
979 	if (phy_flash_cfg)
980 		efx->phy_mode = PHY_MODE_SPECIAL;
981 
982 	/* Connect up MAC/PHY operations table */
983 	rc = efx->type->probe_port(efx);
984 	if (rc)
985 		return rc;
986 
987 	/* Initialise MAC address to permanent address */
988 	memcpy(efx->net_dev->dev_addr, efx->net_dev->perm_addr, ETH_ALEN);
989 
990 	return 0;
991 }
992 
993 static int efx_init_port(struct efx_nic *efx)
994 {
995 	int rc;
996 
997 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
998 
999 	mutex_lock(&efx->mac_lock);
1000 
1001 	rc = efx->phy_op->init(efx);
1002 	if (rc)
1003 		goto fail1;
1004 
1005 	efx->port_initialized = true;
1006 
1007 	/* Reconfigure the MAC before creating dma queues (required for
1008 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1009 	efx->type->reconfigure_mac(efx);
1010 
1011 	/* Ensure the PHY advertises the correct flow control settings */
1012 	rc = efx->phy_op->reconfigure(efx);
1013 	if (rc)
1014 		goto fail2;
1015 
1016 	mutex_unlock(&efx->mac_lock);
1017 	return 0;
1018 
1019 fail2:
1020 	efx->phy_op->fini(efx);
1021 fail1:
1022 	mutex_unlock(&efx->mac_lock);
1023 	return rc;
1024 }
1025 
1026 static void efx_start_port(struct efx_nic *efx)
1027 {
1028 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1029 	BUG_ON(efx->port_enabled);
1030 
1031 	mutex_lock(&efx->mac_lock);
1032 	efx->port_enabled = true;
1033 
1034 	/* efx_mac_work() might have been scheduled after efx_stop_port(),
1035 	 * and then cancelled by efx_flush_all() */
1036 	efx->type->reconfigure_mac(efx);
1037 
1038 	mutex_unlock(&efx->mac_lock);
1039 }
1040 
1041 /* Prevent efx_mac_work() and efx_monitor() from working */
1042 static void efx_stop_port(struct efx_nic *efx)
1043 {
1044 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1045 
1046 	mutex_lock(&efx->mac_lock);
1047 	efx->port_enabled = false;
1048 	mutex_unlock(&efx->mac_lock);
1049 
1050 	/* Serialise against efx_set_multicast_list() */
1051 	netif_addr_lock_bh(efx->net_dev);
1052 	netif_addr_unlock_bh(efx->net_dev);
1053 }
1054 
1055 static void efx_fini_port(struct efx_nic *efx)
1056 {
1057 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1058 
1059 	if (!efx->port_initialized)
1060 		return;
1061 
1062 	efx->phy_op->fini(efx);
1063 	efx->port_initialized = false;
1064 
1065 	efx->link_state.up = false;
1066 	efx_link_status_changed(efx);
1067 }
1068 
1069 static void efx_remove_port(struct efx_nic *efx)
1070 {
1071 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1072 
1073 	efx->type->remove_port(efx);
1074 }
1075 
1076 /**************************************************************************
1077  *
1078  * NIC handling
1079  *
1080  **************************************************************************/
1081 
1082 /* This configures the PCI device to enable I/O and DMA. */
1083 static int efx_init_io(struct efx_nic *efx)
1084 {
1085 	struct pci_dev *pci_dev = efx->pci_dev;
1086 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1087 	int rc;
1088 
1089 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1090 
1091 	rc = pci_enable_device(pci_dev);
1092 	if (rc) {
1093 		netif_err(efx, probe, efx->net_dev,
1094 			  "failed to enable PCI device\n");
1095 		goto fail1;
1096 	}
1097 
1098 	pci_set_master(pci_dev);
1099 
1100 	/* Set the PCI DMA mask.  Try all possibilities from our
1101 	 * genuine mask down to 32 bits, because some architectures
1102 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1103 	 * masks event though they reject 46 bit masks.
1104 	 */
1105 	while (dma_mask > 0x7fffffffUL) {
1106 		if (dma_supported(&pci_dev->dev, dma_mask)) {
1107 			rc = dma_set_mask(&pci_dev->dev, dma_mask);
1108 			if (rc == 0)
1109 				break;
1110 		}
1111 		dma_mask >>= 1;
1112 	}
1113 	if (rc) {
1114 		netif_err(efx, probe, efx->net_dev,
1115 			  "could not find a suitable DMA mask\n");
1116 		goto fail2;
1117 	}
1118 	netif_dbg(efx, probe, efx->net_dev,
1119 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1120 	rc = dma_set_coherent_mask(&pci_dev->dev, dma_mask);
1121 	if (rc) {
1122 		/* dma_set_coherent_mask() is not *allowed* to
1123 		 * fail with a mask that dma_set_mask() accepted,
1124 		 * but just in case...
1125 		 */
1126 		netif_err(efx, probe, efx->net_dev,
1127 			  "failed to set consistent DMA mask\n");
1128 		goto fail2;
1129 	}
1130 
1131 	efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1132 	rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1133 	if (rc) {
1134 		netif_err(efx, probe, efx->net_dev,
1135 			  "request for memory BAR failed\n");
1136 		rc = -EIO;
1137 		goto fail3;
1138 	}
1139 	efx->membase = ioremap_nocache(efx->membase_phys,
1140 				       efx->type->mem_map_size);
1141 	if (!efx->membase) {
1142 		netif_err(efx, probe, efx->net_dev,
1143 			  "could not map memory BAR at %llx+%x\n",
1144 			  (unsigned long long)efx->membase_phys,
1145 			  efx->type->mem_map_size);
1146 		rc = -ENOMEM;
1147 		goto fail4;
1148 	}
1149 	netif_dbg(efx, probe, efx->net_dev,
1150 		  "memory BAR at %llx+%x (virtual %p)\n",
1151 		  (unsigned long long)efx->membase_phys,
1152 		  efx->type->mem_map_size, efx->membase);
1153 
1154 	return 0;
1155 
1156  fail4:
1157 	pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1158  fail3:
1159 	efx->membase_phys = 0;
1160  fail2:
1161 	pci_disable_device(efx->pci_dev);
1162  fail1:
1163 	return rc;
1164 }
1165 
1166 static void efx_fini_io(struct efx_nic *efx)
1167 {
1168 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1169 
1170 	if (efx->membase) {
1171 		iounmap(efx->membase);
1172 		efx->membase = NULL;
1173 	}
1174 
1175 	if (efx->membase_phys) {
1176 		pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1177 		efx->membase_phys = 0;
1178 	}
1179 
1180 	pci_disable_device(efx->pci_dev);
1181 }
1182 
1183 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1184 {
1185 	cpumask_var_t thread_mask;
1186 	unsigned int count;
1187 	int cpu;
1188 
1189 	if (rss_cpus) {
1190 		count = rss_cpus;
1191 	} else {
1192 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1193 			netif_warn(efx, probe, efx->net_dev,
1194 				   "RSS disabled due to allocation failure\n");
1195 			return 1;
1196 		}
1197 
1198 		count = 0;
1199 		for_each_online_cpu(cpu) {
1200 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1201 				++count;
1202 				cpumask_or(thread_mask, thread_mask,
1203 					   topology_thread_cpumask(cpu));
1204 			}
1205 		}
1206 
1207 		free_cpumask_var(thread_mask);
1208 	}
1209 
1210 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1211 	 * table entries that are inaccessible to VFs
1212 	 */
1213 	if (efx_sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1214 	    count > efx_vf_size(efx)) {
1215 		netif_warn(efx, probe, efx->net_dev,
1216 			   "Reducing number of RSS channels from %u to %u for "
1217 			   "VF support. Increase vf-msix-limit to use more "
1218 			   "channels on the PF.\n",
1219 			   count, efx_vf_size(efx));
1220 		count = efx_vf_size(efx);
1221 	}
1222 
1223 	return count;
1224 }
1225 
1226 static int
1227 efx_init_rx_cpu_rmap(struct efx_nic *efx, struct msix_entry *xentries)
1228 {
1229 #ifdef CONFIG_RFS_ACCEL
1230 	unsigned int i;
1231 	int rc;
1232 
1233 	efx->net_dev->rx_cpu_rmap = alloc_irq_cpu_rmap(efx->n_rx_channels);
1234 	if (!efx->net_dev->rx_cpu_rmap)
1235 		return -ENOMEM;
1236 	for (i = 0; i < efx->n_rx_channels; i++) {
1237 		rc = irq_cpu_rmap_add(efx->net_dev->rx_cpu_rmap,
1238 				      xentries[i].vector);
1239 		if (rc) {
1240 			free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1241 			efx->net_dev->rx_cpu_rmap = NULL;
1242 			return rc;
1243 		}
1244 	}
1245 #endif
1246 	return 0;
1247 }
1248 
1249 /* Probe the number and type of interrupts we are able to obtain, and
1250  * the resulting numbers of channels and RX queues.
1251  */
1252 static int efx_probe_interrupts(struct efx_nic *efx)
1253 {
1254 	unsigned int max_channels =
1255 		min(efx->type->phys_addr_channels, EFX_MAX_CHANNELS);
1256 	unsigned int extra_channels = 0;
1257 	unsigned int i, j;
1258 	int rc;
1259 
1260 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1261 		if (efx->extra_channel_type[i])
1262 			++extra_channels;
1263 
1264 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1265 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1266 		unsigned int n_channels;
1267 
1268 		n_channels = efx_wanted_parallelism(efx);
1269 		if (separate_tx_channels)
1270 			n_channels *= 2;
1271 		n_channels += extra_channels;
1272 		n_channels = min(n_channels, max_channels);
1273 
1274 		for (i = 0; i < n_channels; i++)
1275 			xentries[i].entry = i;
1276 		rc = pci_enable_msix(efx->pci_dev, xentries, n_channels);
1277 		if (rc > 0) {
1278 			netif_err(efx, drv, efx->net_dev,
1279 				  "WARNING: Insufficient MSI-X vectors"
1280 				  " available (%d < %u).\n", rc, n_channels);
1281 			netif_err(efx, drv, efx->net_dev,
1282 				  "WARNING: Performance may be reduced.\n");
1283 			EFX_BUG_ON_PARANOID(rc >= n_channels);
1284 			n_channels = rc;
1285 			rc = pci_enable_msix(efx->pci_dev, xentries,
1286 					     n_channels);
1287 		}
1288 
1289 		if (rc == 0) {
1290 			efx->n_channels = n_channels;
1291 			if (n_channels > extra_channels)
1292 				n_channels -= extra_channels;
1293 			if (separate_tx_channels) {
1294 				efx->n_tx_channels = max(n_channels / 2, 1U);
1295 				efx->n_rx_channels = max(n_channels -
1296 							 efx->n_tx_channels,
1297 							 1U);
1298 			} else {
1299 				efx->n_tx_channels = n_channels;
1300 				efx->n_rx_channels = n_channels;
1301 			}
1302 			rc = efx_init_rx_cpu_rmap(efx, xentries);
1303 			if (rc) {
1304 				pci_disable_msix(efx->pci_dev);
1305 				return rc;
1306 			}
1307 			for (i = 0; i < efx->n_channels; i++)
1308 				efx_get_channel(efx, i)->irq =
1309 					xentries[i].vector;
1310 		} else {
1311 			/* Fall back to single channel MSI */
1312 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1313 			netif_err(efx, drv, efx->net_dev,
1314 				  "could not enable MSI-X\n");
1315 		}
1316 	}
1317 
1318 	/* Try single interrupt MSI */
1319 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1320 		efx->n_channels = 1;
1321 		efx->n_rx_channels = 1;
1322 		efx->n_tx_channels = 1;
1323 		rc = pci_enable_msi(efx->pci_dev);
1324 		if (rc == 0) {
1325 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1326 		} else {
1327 			netif_err(efx, drv, efx->net_dev,
1328 				  "could not enable MSI\n");
1329 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1330 		}
1331 	}
1332 
1333 	/* Assume legacy interrupts */
1334 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1335 		efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1336 		efx->n_rx_channels = 1;
1337 		efx->n_tx_channels = 1;
1338 		efx->legacy_irq = efx->pci_dev->irq;
1339 	}
1340 
1341 	/* Assign extra channels if possible */
1342 	j = efx->n_channels;
1343 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1344 		if (!efx->extra_channel_type[i])
1345 			continue;
1346 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1347 		    efx->n_channels <= extra_channels) {
1348 			efx->extra_channel_type[i]->handle_no_channel(efx);
1349 		} else {
1350 			--j;
1351 			efx_get_channel(efx, j)->type =
1352 				efx->extra_channel_type[i];
1353 		}
1354 	}
1355 
1356 	/* RSS might be usable on VFs even if it is disabled on the PF */
1357 	efx->rss_spread = ((efx->n_rx_channels > 1 || !efx_sriov_wanted(efx)) ?
1358 			   efx->n_rx_channels : efx_vf_size(efx));
1359 
1360 	return 0;
1361 }
1362 
1363 /* Enable interrupts, then probe and start the event queues */
1364 static void efx_start_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1365 {
1366 	struct efx_channel *channel;
1367 
1368 	if (efx->legacy_irq)
1369 		efx->legacy_irq_enabled = true;
1370 	efx_nic_enable_interrupts(efx);
1371 
1372 	efx_for_each_channel(channel, efx) {
1373 		if (!channel->type->keep_eventq || !may_keep_eventq)
1374 			efx_init_eventq(channel);
1375 		efx_start_eventq(channel);
1376 	}
1377 
1378 	efx_mcdi_mode_event(efx);
1379 }
1380 
1381 static void efx_stop_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1382 {
1383 	struct efx_channel *channel;
1384 
1385 	efx_mcdi_mode_poll(efx);
1386 
1387 	efx_nic_disable_interrupts(efx);
1388 	if (efx->legacy_irq) {
1389 		synchronize_irq(efx->legacy_irq);
1390 		efx->legacy_irq_enabled = false;
1391 	}
1392 
1393 	efx_for_each_channel(channel, efx) {
1394 		if (channel->irq)
1395 			synchronize_irq(channel->irq);
1396 
1397 		efx_stop_eventq(channel);
1398 		if (!channel->type->keep_eventq || !may_keep_eventq)
1399 			efx_fini_eventq(channel);
1400 	}
1401 }
1402 
1403 static void efx_remove_interrupts(struct efx_nic *efx)
1404 {
1405 	struct efx_channel *channel;
1406 
1407 	/* Remove MSI/MSI-X interrupts */
1408 	efx_for_each_channel(channel, efx)
1409 		channel->irq = 0;
1410 	pci_disable_msi(efx->pci_dev);
1411 	pci_disable_msix(efx->pci_dev);
1412 
1413 	/* Remove legacy interrupt */
1414 	efx->legacy_irq = 0;
1415 }
1416 
1417 static void efx_set_channels(struct efx_nic *efx)
1418 {
1419 	struct efx_channel *channel;
1420 	struct efx_tx_queue *tx_queue;
1421 
1422 	efx->tx_channel_offset =
1423 		separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1424 
1425 	/* We need to adjust the TX queue numbers if we have separate
1426 	 * RX-only and TX-only channels.
1427 	 */
1428 	efx_for_each_channel(channel, efx) {
1429 		efx_for_each_channel_tx_queue(tx_queue, channel)
1430 			tx_queue->queue -= (efx->tx_channel_offset *
1431 					    EFX_TXQ_TYPES);
1432 	}
1433 }
1434 
1435 static int efx_probe_nic(struct efx_nic *efx)
1436 {
1437 	size_t i;
1438 	int rc;
1439 
1440 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1441 
1442 	/* Carry out hardware-type specific initialisation */
1443 	rc = efx->type->probe(efx);
1444 	if (rc)
1445 		return rc;
1446 
1447 	/* Determine the number of channels and queues by trying to hook
1448 	 * in MSI-X interrupts. */
1449 	rc = efx_probe_interrupts(efx);
1450 	if (rc)
1451 		goto fail;
1452 
1453 	efx->type->dimension_resources(efx);
1454 
1455 	if (efx->n_channels > 1)
1456 		get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1457 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1458 		efx->rx_indir_table[i] =
1459 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1460 
1461 	efx_set_channels(efx);
1462 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1463 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1464 
1465 	/* Initialise the interrupt moderation settings */
1466 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1467 				true);
1468 
1469 	return 0;
1470 
1471 fail:
1472 	efx->type->remove(efx);
1473 	return rc;
1474 }
1475 
1476 static void efx_remove_nic(struct efx_nic *efx)
1477 {
1478 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1479 
1480 	efx_remove_interrupts(efx);
1481 	efx->type->remove(efx);
1482 }
1483 
1484 /**************************************************************************
1485  *
1486  * NIC startup/shutdown
1487  *
1488  *************************************************************************/
1489 
1490 static int efx_probe_all(struct efx_nic *efx)
1491 {
1492 	int rc;
1493 
1494 	rc = efx_probe_nic(efx);
1495 	if (rc) {
1496 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1497 		goto fail1;
1498 	}
1499 
1500 	rc = efx_probe_port(efx);
1501 	if (rc) {
1502 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1503 		goto fail2;
1504 	}
1505 
1506 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1507 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1508 		rc = -EINVAL;
1509 		goto fail3;
1510 	}
1511 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1512 
1513 	rc = efx_probe_filters(efx);
1514 	if (rc) {
1515 		netif_err(efx, probe, efx->net_dev,
1516 			  "failed to create filter tables\n");
1517 		goto fail3;
1518 	}
1519 
1520 	rc = efx_probe_channels(efx);
1521 	if (rc)
1522 		goto fail4;
1523 
1524 	return 0;
1525 
1526  fail4:
1527 	efx_remove_filters(efx);
1528  fail3:
1529 	efx_remove_port(efx);
1530  fail2:
1531 	efx_remove_nic(efx);
1532  fail1:
1533 	return rc;
1534 }
1535 
1536 /* Called after previous invocation(s) of efx_stop_all, restarts the port,
1537  * kernel transmit queues and NAPI processing, and ensures that the port is
1538  * scheduled to be reconfigured. This function is safe to call multiple
1539  * times when the NIC is in any state.
1540  */
1541 static void efx_start_all(struct efx_nic *efx)
1542 {
1543 	EFX_ASSERT_RESET_SERIALISED(efx);
1544 
1545 	/* Check that it is appropriate to restart the interface. All
1546 	 * of these flags are safe to read under just the rtnl lock */
1547 	if (efx->port_enabled)
1548 		return;
1549 	if ((efx->state != STATE_RUNNING) && (efx->state != STATE_INIT))
1550 		return;
1551 	if (!netif_running(efx->net_dev))
1552 		return;
1553 
1554 	efx_start_port(efx);
1555 	efx_start_datapath(efx);
1556 
1557 	/* Start the hardware monitor if there is one. Otherwise (we're link
1558 	 * event driven), we have to poll the PHY because after an event queue
1559 	 * flush, we could have a missed a link state change */
1560 	if (efx->type->monitor != NULL) {
1561 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1562 				   efx_monitor_interval);
1563 	} else {
1564 		mutex_lock(&efx->mac_lock);
1565 		if (efx->phy_op->poll(efx))
1566 			efx_link_status_changed(efx);
1567 		mutex_unlock(&efx->mac_lock);
1568 	}
1569 
1570 	efx->type->start_stats(efx);
1571 }
1572 
1573 /* Flush all delayed work. Should only be called when no more delayed work
1574  * will be scheduled. This doesn't flush pending online resets (efx_reset),
1575  * since we're holding the rtnl_lock at this point. */
1576 static void efx_flush_all(struct efx_nic *efx)
1577 {
1578 	/* Make sure the hardware monitor and event self-test are stopped */
1579 	cancel_delayed_work_sync(&efx->monitor_work);
1580 	efx_selftest_async_cancel(efx);
1581 	/* Stop scheduled port reconfigurations */
1582 	cancel_work_sync(&efx->mac_work);
1583 }
1584 
1585 /* Quiesce hardware and software without bringing the link down.
1586  * Safe to call multiple times, when the nic and interface is in any
1587  * state. The caller is guaranteed to subsequently be in a position
1588  * to modify any hardware and software state they see fit without
1589  * taking locks. */
1590 static void efx_stop_all(struct efx_nic *efx)
1591 {
1592 	EFX_ASSERT_RESET_SERIALISED(efx);
1593 
1594 	/* port_enabled can be read safely under the rtnl lock */
1595 	if (!efx->port_enabled)
1596 		return;
1597 
1598 	efx->type->stop_stats(efx);
1599 	efx_stop_port(efx);
1600 
1601 	/* Flush efx_mac_work(), refill_workqueue, monitor_work */
1602 	efx_flush_all(efx);
1603 
1604 	/* Stop the kernel transmit interface late, so the watchdog
1605 	 * timer isn't ticking over the flush */
1606 	netif_tx_disable(efx->net_dev);
1607 
1608 	efx_stop_datapath(efx);
1609 }
1610 
1611 static void efx_remove_all(struct efx_nic *efx)
1612 {
1613 	efx_remove_channels(efx);
1614 	efx_remove_filters(efx);
1615 	efx_remove_port(efx);
1616 	efx_remove_nic(efx);
1617 }
1618 
1619 /**************************************************************************
1620  *
1621  * Interrupt moderation
1622  *
1623  **************************************************************************/
1624 
1625 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1626 {
1627 	if (usecs == 0)
1628 		return 0;
1629 	if (usecs * 1000 < quantum_ns)
1630 		return 1; /* never round down to 0 */
1631 	return usecs * 1000 / quantum_ns;
1632 }
1633 
1634 /* Set interrupt moderation parameters */
1635 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1636 			    unsigned int rx_usecs, bool rx_adaptive,
1637 			    bool rx_may_override_tx)
1638 {
1639 	struct efx_channel *channel;
1640 	unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1641 						efx->timer_quantum_ns,
1642 						1000);
1643 	unsigned int tx_ticks;
1644 	unsigned int rx_ticks;
1645 
1646 	EFX_ASSERT_RESET_SERIALISED(efx);
1647 
1648 	if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1649 		return -EINVAL;
1650 
1651 	tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1652 	rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1653 
1654 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1655 	    !rx_may_override_tx) {
1656 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1657 			  "RX and TX IRQ moderation must be equal\n");
1658 		return -EINVAL;
1659 	}
1660 
1661 	efx->irq_rx_adaptive = rx_adaptive;
1662 	efx->irq_rx_moderation = rx_ticks;
1663 	efx_for_each_channel(channel, efx) {
1664 		if (efx_channel_has_rx_queue(channel))
1665 			channel->irq_moderation = rx_ticks;
1666 		else if (efx_channel_has_tx_queues(channel))
1667 			channel->irq_moderation = tx_ticks;
1668 	}
1669 
1670 	return 0;
1671 }
1672 
1673 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1674 			    unsigned int *rx_usecs, bool *rx_adaptive)
1675 {
1676 	/* We must round up when converting ticks to microseconds
1677 	 * because we round down when converting the other way.
1678 	 */
1679 
1680 	*rx_adaptive = efx->irq_rx_adaptive;
1681 	*rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1682 				 efx->timer_quantum_ns,
1683 				 1000);
1684 
1685 	/* If channels are shared between RX and TX, so is IRQ
1686 	 * moderation.  Otherwise, IRQ moderation is the same for all
1687 	 * TX channels and is not adaptive.
1688 	 */
1689 	if (efx->tx_channel_offset == 0)
1690 		*tx_usecs = *rx_usecs;
1691 	else
1692 		*tx_usecs = DIV_ROUND_UP(
1693 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1694 			efx->timer_quantum_ns,
1695 			1000);
1696 }
1697 
1698 /**************************************************************************
1699  *
1700  * Hardware monitor
1701  *
1702  **************************************************************************/
1703 
1704 /* Run periodically off the general workqueue */
1705 static void efx_monitor(struct work_struct *data)
1706 {
1707 	struct efx_nic *efx = container_of(data, struct efx_nic,
1708 					   monitor_work.work);
1709 
1710 	netif_vdbg(efx, timer, efx->net_dev,
1711 		   "hardware monitor executing on CPU %d\n",
1712 		   raw_smp_processor_id());
1713 	BUG_ON(efx->type->monitor == NULL);
1714 
1715 	/* If the mac_lock is already held then it is likely a port
1716 	 * reconfiguration is already in place, which will likely do
1717 	 * most of the work of monitor() anyway. */
1718 	if (mutex_trylock(&efx->mac_lock)) {
1719 		if (efx->port_enabled)
1720 			efx->type->monitor(efx);
1721 		mutex_unlock(&efx->mac_lock);
1722 	}
1723 
1724 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1725 			   efx_monitor_interval);
1726 }
1727 
1728 /**************************************************************************
1729  *
1730  * ioctls
1731  *
1732  *************************************************************************/
1733 
1734 /* Net device ioctl
1735  * Context: process, rtnl_lock() held.
1736  */
1737 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1738 {
1739 	struct efx_nic *efx = netdev_priv(net_dev);
1740 	struct mii_ioctl_data *data = if_mii(ifr);
1741 
1742 	EFX_ASSERT_RESET_SERIALISED(efx);
1743 
1744 	/* Convert phy_id from older PRTAD/DEVAD format */
1745 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1746 	    (data->phy_id & 0xfc00) == 0x0400)
1747 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1748 
1749 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
1750 }
1751 
1752 /**************************************************************************
1753  *
1754  * NAPI interface
1755  *
1756  **************************************************************************/
1757 
1758 static void efx_init_napi_channel(struct efx_channel *channel)
1759 {
1760 	struct efx_nic *efx = channel->efx;
1761 
1762 	channel->napi_dev = efx->net_dev;
1763 	netif_napi_add(channel->napi_dev, &channel->napi_str,
1764 		       efx_poll, napi_weight);
1765 }
1766 
1767 static void efx_init_napi(struct efx_nic *efx)
1768 {
1769 	struct efx_channel *channel;
1770 
1771 	efx_for_each_channel(channel, efx)
1772 		efx_init_napi_channel(channel);
1773 }
1774 
1775 static void efx_fini_napi_channel(struct efx_channel *channel)
1776 {
1777 	if (channel->napi_dev)
1778 		netif_napi_del(&channel->napi_str);
1779 	channel->napi_dev = NULL;
1780 }
1781 
1782 static void efx_fini_napi(struct efx_nic *efx)
1783 {
1784 	struct efx_channel *channel;
1785 
1786 	efx_for_each_channel(channel, efx)
1787 		efx_fini_napi_channel(channel);
1788 }
1789 
1790 /**************************************************************************
1791  *
1792  * Kernel netpoll interface
1793  *
1794  *************************************************************************/
1795 
1796 #ifdef CONFIG_NET_POLL_CONTROLLER
1797 
1798 /* Although in the common case interrupts will be disabled, this is not
1799  * guaranteed. However, all our work happens inside the NAPI callback,
1800  * so no locking is required.
1801  */
1802 static void efx_netpoll(struct net_device *net_dev)
1803 {
1804 	struct efx_nic *efx = netdev_priv(net_dev);
1805 	struct efx_channel *channel;
1806 
1807 	efx_for_each_channel(channel, efx)
1808 		efx_schedule_channel(channel);
1809 }
1810 
1811 #endif
1812 
1813 /**************************************************************************
1814  *
1815  * Kernel net device interface
1816  *
1817  *************************************************************************/
1818 
1819 /* Context: process, rtnl_lock() held. */
1820 static int efx_net_open(struct net_device *net_dev)
1821 {
1822 	struct efx_nic *efx = netdev_priv(net_dev);
1823 	EFX_ASSERT_RESET_SERIALISED(efx);
1824 
1825 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
1826 		  raw_smp_processor_id());
1827 
1828 	if (efx->state == STATE_DISABLED)
1829 		return -EIO;
1830 	if (efx->phy_mode & PHY_MODE_SPECIAL)
1831 		return -EBUSY;
1832 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
1833 		return -EIO;
1834 
1835 	/* Notify the kernel of the link state polled during driver load,
1836 	 * before the monitor starts running */
1837 	efx_link_status_changed(efx);
1838 
1839 	efx_start_all(efx);
1840 	efx_selftest_async_start(efx);
1841 	return 0;
1842 }
1843 
1844 /* Context: process, rtnl_lock() held.
1845  * Note that the kernel will ignore our return code; this method
1846  * should really be a void.
1847  */
1848 static int efx_net_stop(struct net_device *net_dev)
1849 {
1850 	struct efx_nic *efx = netdev_priv(net_dev);
1851 
1852 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
1853 		  raw_smp_processor_id());
1854 
1855 	if (efx->state != STATE_DISABLED) {
1856 		/* Stop the device and flush all the channels */
1857 		efx_stop_all(efx);
1858 	}
1859 
1860 	return 0;
1861 }
1862 
1863 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1864 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
1865 					       struct rtnl_link_stats64 *stats)
1866 {
1867 	struct efx_nic *efx = netdev_priv(net_dev);
1868 	struct efx_mac_stats *mac_stats = &efx->mac_stats;
1869 
1870 	spin_lock_bh(&efx->stats_lock);
1871 
1872 	efx->type->update_stats(efx);
1873 
1874 	stats->rx_packets = mac_stats->rx_packets;
1875 	stats->tx_packets = mac_stats->tx_packets;
1876 	stats->rx_bytes = mac_stats->rx_bytes;
1877 	stats->tx_bytes = mac_stats->tx_bytes;
1878 	stats->rx_dropped = efx->n_rx_nodesc_drop_cnt;
1879 	stats->multicast = mac_stats->rx_multicast;
1880 	stats->collisions = mac_stats->tx_collision;
1881 	stats->rx_length_errors = (mac_stats->rx_gtjumbo +
1882 				   mac_stats->rx_length_error);
1883 	stats->rx_crc_errors = mac_stats->rx_bad;
1884 	stats->rx_frame_errors = mac_stats->rx_align_error;
1885 	stats->rx_fifo_errors = mac_stats->rx_overflow;
1886 	stats->rx_missed_errors = mac_stats->rx_missed;
1887 	stats->tx_window_errors = mac_stats->tx_late_collision;
1888 
1889 	stats->rx_errors = (stats->rx_length_errors +
1890 			    stats->rx_crc_errors +
1891 			    stats->rx_frame_errors +
1892 			    mac_stats->rx_symbol_error);
1893 	stats->tx_errors = (stats->tx_window_errors +
1894 			    mac_stats->tx_bad);
1895 
1896 	spin_unlock_bh(&efx->stats_lock);
1897 
1898 	return stats;
1899 }
1900 
1901 /* Context: netif_tx_lock held, BHs disabled. */
1902 static void efx_watchdog(struct net_device *net_dev)
1903 {
1904 	struct efx_nic *efx = netdev_priv(net_dev);
1905 
1906 	netif_err(efx, tx_err, efx->net_dev,
1907 		  "TX stuck with port_enabled=%d: resetting channels\n",
1908 		  efx->port_enabled);
1909 
1910 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
1911 }
1912 
1913 
1914 /* Context: process, rtnl_lock() held. */
1915 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
1916 {
1917 	struct efx_nic *efx = netdev_priv(net_dev);
1918 
1919 	EFX_ASSERT_RESET_SERIALISED(efx);
1920 
1921 	if (new_mtu > EFX_MAX_MTU)
1922 		return -EINVAL;
1923 
1924 	efx_stop_all(efx);
1925 
1926 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
1927 
1928 	mutex_lock(&efx->mac_lock);
1929 	/* Reconfigure the MAC before enabling the dma queues so that
1930 	 * the RX buffers don't overflow */
1931 	net_dev->mtu = new_mtu;
1932 	efx->type->reconfigure_mac(efx);
1933 	mutex_unlock(&efx->mac_lock);
1934 
1935 	efx_start_all(efx);
1936 	return 0;
1937 }
1938 
1939 static int efx_set_mac_address(struct net_device *net_dev, void *data)
1940 {
1941 	struct efx_nic *efx = netdev_priv(net_dev);
1942 	struct sockaddr *addr = data;
1943 	char *new_addr = addr->sa_data;
1944 
1945 	EFX_ASSERT_RESET_SERIALISED(efx);
1946 
1947 	if (!is_valid_ether_addr(new_addr)) {
1948 		netif_err(efx, drv, efx->net_dev,
1949 			  "invalid ethernet MAC address requested: %pM\n",
1950 			  new_addr);
1951 		return -EADDRNOTAVAIL;
1952 	}
1953 
1954 	memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
1955 	efx_sriov_mac_address_changed(efx);
1956 
1957 	/* Reconfigure the MAC */
1958 	mutex_lock(&efx->mac_lock);
1959 	efx->type->reconfigure_mac(efx);
1960 	mutex_unlock(&efx->mac_lock);
1961 
1962 	return 0;
1963 }
1964 
1965 /* Context: netif_addr_lock held, BHs disabled. */
1966 static void efx_set_rx_mode(struct net_device *net_dev)
1967 {
1968 	struct efx_nic *efx = netdev_priv(net_dev);
1969 	struct netdev_hw_addr *ha;
1970 	union efx_multicast_hash *mc_hash = &efx->multicast_hash;
1971 	u32 crc;
1972 	int bit;
1973 
1974 	efx->promiscuous = !!(net_dev->flags & IFF_PROMISC);
1975 
1976 	/* Build multicast hash table */
1977 	if (efx->promiscuous || (net_dev->flags & IFF_ALLMULTI)) {
1978 		memset(mc_hash, 0xff, sizeof(*mc_hash));
1979 	} else {
1980 		memset(mc_hash, 0x00, sizeof(*mc_hash));
1981 		netdev_for_each_mc_addr(ha, net_dev) {
1982 			crc = ether_crc_le(ETH_ALEN, ha->addr);
1983 			bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
1984 			set_bit_le(bit, mc_hash->byte);
1985 		}
1986 
1987 		/* Broadcast packets go through the multicast hash filter.
1988 		 * ether_crc_le() of the broadcast address is 0xbe2612ff
1989 		 * so we always add bit 0xff to the mask.
1990 		 */
1991 		set_bit_le(0xff, mc_hash->byte);
1992 	}
1993 
1994 	if (efx->port_enabled)
1995 		queue_work(efx->workqueue, &efx->mac_work);
1996 	/* Otherwise efx_start_port() will do this */
1997 }
1998 
1999 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2000 {
2001 	struct efx_nic *efx = netdev_priv(net_dev);
2002 
2003 	/* If disabling RX n-tuple filtering, clear existing filters */
2004 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
2005 		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2006 
2007 	return 0;
2008 }
2009 
2010 static const struct net_device_ops efx_netdev_ops = {
2011 	.ndo_open		= efx_net_open,
2012 	.ndo_stop		= efx_net_stop,
2013 	.ndo_get_stats64	= efx_net_stats,
2014 	.ndo_tx_timeout		= efx_watchdog,
2015 	.ndo_start_xmit		= efx_hard_start_xmit,
2016 	.ndo_validate_addr	= eth_validate_addr,
2017 	.ndo_do_ioctl		= efx_ioctl,
2018 	.ndo_change_mtu		= efx_change_mtu,
2019 	.ndo_set_mac_address	= efx_set_mac_address,
2020 	.ndo_set_rx_mode	= efx_set_rx_mode,
2021 	.ndo_set_features	= efx_set_features,
2022 #ifdef CONFIG_SFC_SRIOV
2023 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2024 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2025 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2026 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2027 #endif
2028 #ifdef CONFIG_NET_POLL_CONTROLLER
2029 	.ndo_poll_controller = efx_netpoll,
2030 #endif
2031 	.ndo_setup_tc		= efx_setup_tc,
2032 #ifdef CONFIG_RFS_ACCEL
2033 	.ndo_rx_flow_steer	= efx_filter_rfs,
2034 #endif
2035 };
2036 
2037 static void efx_update_name(struct efx_nic *efx)
2038 {
2039 	strcpy(efx->name, efx->net_dev->name);
2040 	efx_mtd_rename(efx);
2041 	efx_set_channel_names(efx);
2042 }
2043 
2044 static int efx_netdev_event(struct notifier_block *this,
2045 			    unsigned long event, void *ptr)
2046 {
2047 	struct net_device *net_dev = ptr;
2048 
2049 	if (net_dev->netdev_ops == &efx_netdev_ops &&
2050 	    event == NETDEV_CHANGENAME)
2051 		efx_update_name(netdev_priv(net_dev));
2052 
2053 	return NOTIFY_DONE;
2054 }
2055 
2056 static struct notifier_block efx_netdev_notifier = {
2057 	.notifier_call = efx_netdev_event,
2058 };
2059 
2060 static ssize_t
2061 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2062 {
2063 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2064 	return sprintf(buf, "%d\n", efx->phy_type);
2065 }
2066 static DEVICE_ATTR(phy_type, 0644, show_phy_type, NULL);
2067 
2068 static int efx_register_netdev(struct efx_nic *efx)
2069 {
2070 	struct net_device *net_dev = efx->net_dev;
2071 	struct efx_channel *channel;
2072 	int rc;
2073 
2074 	net_dev->watchdog_timeo = 5 * HZ;
2075 	net_dev->irq = efx->pci_dev->irq;
2076 	net_dev->netdev_ops = &efx_netdev_ops;
2077 	SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
2078 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2079 
2080 	rtnl_lock();
2081 
2082 	rc = dev_alloc_name(net_dev, net_dev->name);
2083 	if (rc < 0)
2084 		goto fail_locked;
2085 	efx_update_name(efx);
2086 
2087 	rc = register_netdevice(net_dev);
2088 	if (rc)
2089 		goto fail_locked;
2090 
2091 	efx_for_each_channel(channel, efx) {
2092 		struct efx_tx_queue *tx_queue;
2093 		efx_for_each_channel_tx_queue(tx_queue, channel)
2094 			efx_init_tx_queue_core_txq(tx_queue);
2095 	}
2096 
2097 	/* Always start with carrier off; PHY events will detect the link */
2098 	netif_carrier_off(net_dev);
2099 
2100 	rtnl_unlock();
2101 
2102 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2103 	if (rc) {
2104 		netif_err(efx, drv, efx->net_dev,
2105 			  "failed to init net dev attributes\n");
2106 		goto fail_registered;
2107 	}
2108 
2109 	return 0;
2110 
2111 fail_locked:
2112 	rtnl_unlock();
2113 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2114 	return rc;
2115 
2116 fail_registered:
2117 	unregister_netdev(net_dev);
2118 	return rc;
2119 }
2120 
2121 static void efx_unregister_netdev(struct efx_nic *efx)
2122 {
2123 	struct efx_channel *channel;
2124 	struct efx_tx_queue *tx_queue;
2125 
2126 	if (!efx->net_dev)
2127 		return;
2128 
2129 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2130 
2131 	/* Free up any skbs still remaining. This has to happen before
2132 	 * we try to unregister the netdev as running their destructors
2133 	 * may be needed to get the device ref. count to 0. */
2134 	efx_for_each_channel(channel, efx) {
2135 		efx_for_each_channel_tx_queue(tx_queue, channel)
2136 			efx_release_tx_buffers(tx_queue);
2137 	}
2138 
2139 	strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2140 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2141 	unregister_netdev(efx->net_dev);
2142 }
2143 
2144 /**************************************************************************
2145  *
2146  * Device reset and suspend
2147  *
2148  **************************************************************************/
2149 
2150 /* Tears down the entire software state and most of the hardware state
2151  * before reset.  */
2152 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2153 {
2154 	EFX_ASSERT_RESET_SERIALISED(efx);
2155 
2156 	efx_stop_all(efx);
2157 	mutex_lock(&efx->mac_lock);
2158 
2159 	efx_stop_interrupts(efx, false);
2160 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2161 		efx->phy_op->fini(efx);
2162 	efx->type->fini(efx);
2163 }
2164 
2165 /* This function will always ensure that the locks acquired in
2166  * efx_reset_down() are released. A failure return code indicates
2167  * that we were unable to reinitialise the hardware, and the
2168  * driver should be disabled. If ok is false, then the rx and tx
2169  * engines are not restarted, pending a RESET_DISABLE. */
2170 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2171 {
2172 	int rc;
2173 
2174 	EFX_ASSERT_RESET_SERIALISED(efx);
2175 
2176 	rc = efx->type->init(efx);
2177 	if (rc) {
2178 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2179 		goto fail;
2180 	}
2181 
2182 	if (!ok)
2183 		goto fail;
2184 
2185 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2186 		rc = efx->phy_op->init(efx);
2187 		if (rc)
2188 			goto fail;
2189 		if (efx->phy_op->reconfigure(efx))
2190 			netif_err(efx, drv, efx->net_dev,
2191 				  "could not restore PHY settings\n");
2192 	}
2193 
2194 	efx->type->reconfigure_mac(efx);
2195 
2196 	efx_start_interrupts(efx, false);
2197 	efx_restore_filters(efx);
2198 	efx_sriov_reset(efx);
2199 
2200 	mutex_unlock(&efx->mac_lock);
2201 
2202 	efx_start_all(efx);
2203 
2204 	return 0;
2205 
2206 fail:
2207 	efx->port_initialized = false;
2208 
2209 	mutex_unlock(&efx->mac_lock);
2210 
2211 	return rc;
2212 }
2213 
2214 /* Reset the NIC using the specified method.  Note that the reset may
2215  * fail, in which case the card will be left in an unusable state.
2216  *
2217  * Caller must hold the rtnl_lock.
2218  */
2219 int efx_reset(struct efx_nic *efx, enum reset_type method)
2220 {
2221 	int rc, rc2;
2222 	bool disabled;
2223 
2224 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2225 		   RESET_TYPE(method));
2226 
2227 	netif_device_detach(efx->net_dev);
2228 	efx_reset_down(efx, method);
2229 
2230 	rc = efx->type->reset(efx, method);
2231 	if (rc) {
2232 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2233 		goto out;
2234 	}
2235 
2236 	/* Clear flags for the scopes we covered.  We assume the NIC and
2237 	 * driver are now quiescent so that there is no race here.
2238 	 */
2239 	efx->reset_pending &= -(1 << (method + 1));
2240 
2241 	/* Reinitialise bus-mastering, which may have been turned off before
2242 	 * the reset was scheduled. This is still appropriate, even in the
2243 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2244 	 * can respond to requests. */
2245 	pci_set_master(efx->pci_dev);
2246 
2247 out:
2248 	/* Leave device stopped if necessary */
2249 	disabled = rc || method == RESET_TYPE_DISABLE;
2250 	rc2 = efx_reset_up(efx, method, !disabled);
2251 	if (rc2) {
2252 		disabled = true;
2253 		if (!rc)
2254 			rc = rc2;
2255 	}
2256 
2257 	if (disabled) {
2258 		dev_close(efx->net_dev);
2259 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2260 		efx->state = STATE_DISABLED;
2261 	} else {
2262 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2263 		netif_device_attach(efx->net_dev);
2264 	}
2265 	return rc;
2266 }
2267 
2268 /* The worker thread exists so that code that cannot sleep can
2269  * schedule a reset for later.
2270  */
2271 static void efx_reset_work(struct work_struct *data)
2272 {
2273 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2274 	unsigned long pending = ACCESS_ONCE(efx->reset_pending);
2275 
2276 	if (!pending)
2277 		return;
2278 
2279 	/* If we're not RUNNING then don't reset. Leave the reset_pending
2280 	 * flags set so that efx_pci_probe_main will be retried */
2281 	if (efx->state != STATE_RUNNING) {
2282 		netif_info(efx, drv, efx->net_dev,
2283 			   "scheduled reset quenched. NIC not RUNNING\n");
2284 		return;
2285 	}
2286 
2287 	rtnl_lock();
2288 	(void)efx_reset(efx, fls(pending) - 1);
2289 	rtnl_unlock();
2290 }
2291 
2292 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2293 {
2294 	enum reset_type method;
2295 
2296 	switch (type) {
2297 	case RESET_TYPE_INVISIBLE:
2298 	case RESET_TYPE_ALL:
2299 	case RESET_TYPE_WORLD:
2300 	case RESET_TYPE_DISABLE:
2301 		method = type;
2302 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2303 			  RESET_TYPE(method));
2304 		break;
2305 	default:
2306 		method = efx->type->map_reset_reason(type);
2307 		netif_dbg(efx, drv, efx->net_dev,
2308 			  "scheduling %s reset for %s\n",
2309 			  RESET_TYPE(method), RESET_TYPE(type));
2310 		break;
2311 	}
2312 
2313 	set_bit(method, &efx->reset_pending);
2314 
2315 	/* efx_process_channel() will no longer read events once a
2316 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2317 	efx_mcdi_mode_poll(efx);
2318 
2319 	queue_work(reset_workqueue, &efx->reset_work);
2320 }
2321 
2322 /**************************************************************************
2323  *
2324  * List of NICs we support
2325  *
2326  **************************************************************************/
2327 
2328 /* PCI device ID table */
2329 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2330 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2331 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2332 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2333 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2334 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2335 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2336 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2337 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2338 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2339 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2340 	{0}			/* end of list */
2341 };
2342 
2343 /**************************************************************************
2344  *
2345  * Dummy PHY/MAC operations
2346  *
2347  * Can be used for some unimplemented operations
2348  * Needed so all function pointers are valid and do not have to be tested
2349  * before use
2350  *
2351  **************************************************************************/
2352 int efx_port_dummy_op_int(struct efx_nic *efx)
2353 {
2354 	return 0;
2355 }
2356 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2357 
2358 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2359 {
2360 	return false;
2361 }
2362 
2363 static const struct efx_phy_operations efx_dummy_phy_operations = {
2364 	.init		 = efx_port_dummy_op_int,
2365 	.reconfigure	 = efx_port_dummy_op_int,
2366 	.poll		 = efx_port_dummy_op_poll,
2367 	.fini		 = efx_port_dummy_op_void,
2368 };
2369 
2370 /**************************************************************************
2371  *
2372  * Data housekeeping
2373  *
2374  **************************************************************************/
2375 
2376 /* This zeroes out and then fills in the invariants in a struct
2377  * efx_nic (including all sub-structures).
2378  */
2379 static int efx_init_struct(struct efx_nic *efx, const struct efx_nic_type *type,
2380 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2381 {
2382 	int i;
2383 
2384 	/* Initialise common structures */
2385 	memset(efx, 0, sizeof(*efx));
2386 	spin_lock_init(&efx->biu_lock);
2387 #ifdef CONFIG_SFC_MTD
2388 	INIT_LIST_HEAD(&efx->mtd_list);
2389 #endif
2390 	INIT_WORK(&efx->reset_work, efx_reset_work);
2391 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2392 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2393 	efx->pci_dev = pci_dev;
2394 	efx->msg_enable = debug;
2395 	efx->state = STATE_INIT;
2396 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2397 
2398 	efx->net_dev = net_dev;
2399 	spin_lock_init(&efx->stats_lock);
2400 	mutex_init(&efx->mac_lock);
2401 	efx->phy_op = &efx_dummy_phy_operations;
2402 	efx->mdio.dev = net_dev;
2403 	INIT_WORK(&efx->mac_work, efx_mac_work);
2404 	init_waitqueue_head(&efx->flush_wq);
2405 
2406 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2407 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2408 		if (!efx->channel[i])
2409 			goto fail;
2410 	}
2411 
2412 	efx->type = type;
2413 
2414 	EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS);
2415 
2416 	/* Higher numbered interrupt modes are less capable! */
2417 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2418 				  interrupt_mode);
2419 
2420 	/* Would be good to use the net_dev name, but we're too early */
2421 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2422 		 pci_name(pci_dev));
2423 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2424 	if (!efx->workqueue)
2425 		goto fail;
2426 
2427 	return 0;
2428 
2429 fail:
2430 	efx_fini_struct(efx);
2431 	return -ENOMEM;
2432 }
2433 
2434 static void efx_fini_struct(struct efx_nic *efx)
2435 {
2436 	int i;
2437 
2438 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2439 		kfree(efx->channel[i]);
2440 
2441 	if (efx->workqueue) {
2442 		destroy_workqueue(efx->workqueue);
2443 		efx->workqueue = NULL;
2444 	}
2445 }
2446 
2447 /**************************************************************************
2448  *
2449  * PCI interface
2450  *
2451  **************************************************************************/
2452 
2453 /* Main body of final NIC shutdown code
2454  * This is called only at module unload (or hotplug removal).
2455  */
2456 static void efx_pci_remove_main(struct efx_nic *efx)
2457 {
2458 #ifdef CONFIG_RFS_ACCEL
2459 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
2460 	efx->net_dev->rx_cpu_rmap = NULL;
2461 #endif
2462 	efx_stop_interrupts(efx, false);
2463 	efx_nic_fini_interrupt(efx);
2464 	efx_fini_port(efx);
2465 	efx->type->fini(efx);
2466 	efx_fini_napi(efx);
2467 	efx_remove_all(efx);
2468 }
2469 
2470 /* Final NIC shutdown
2471  * This is called only at module unload (or hotplug removal).
2472  */
2473 static void efx_pci_remove(struct pci_dev *pci_dev)
2474 {
2475 	struct efx_nic *efx;
2476 
2477 	efx = pci_get_drvdata(pci_dev);
2478 	if (!efx)
2479 		return;
2480 
2481 	/* Mark the NIC as fini, then stop the interface */
2482 	rtnl_lock();
2483 	efx->state = STATE_FINI;
2484 	dev_close(efx->net_dev);
2485 
2486 	/* Allow any queued efx_resets() to complete */
2487 	rtnl_unlock();
2488 
2489 	efx_stop_interrupts(efx, false);
2490 	efx_sriov_fini(efx);
2491 	efx_unregister_netdev(efx);
2492 
2493 	efx_mtd_remove(efx);
2494 
2495 	/* Wait for any scheduled resets to complete. No more will be
2496 	 * scheduled from this point because efx_stop_all() has been
2497 	 * called, we are no longer registered with driverlink, and
2498 	 * the net_device's have been removed. */
2499 	cancel_work_sync(&efx->reset_work);
2500 
2501 	efx_pci_remove_main(efx);
2502 
2503 	efx_fini_io(efx);
2504 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2505 
2506 	efx_fini_struct(efx);
2507 	pci_set_drvdata(pci_dev, NULL);
2508 	free_netdev(efx->net_dev);
2509 };
2510 
2511 /* NIC VPD information
2512  * Called during probe to display the part number of the
2513  * installed NIC.  VPD is potentially very large but this should
2514  * always appear within the first 512 bytes.
2515  */
2516 #define SFC_VPD_LEN 512
2517 static void efx_print_product_vpd(struct efx_nic *efx)
2518 {
2519 	struct pci_dev *dev = efx->pci_dev;
2520 	char vpd_data[SFC_VPD_LEN];
2521 	ssize_t vpd_size;
2522 	int i, j;
2523 
2524 	/* Get the vpd data from the device */
2525 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2526 	if (vpd_size <= 0) {
2527 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2528 		return;
2529 	}
2530 
2531 	/* Get the Read only section */
2532 	i = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2533 	if (i < 0) {
2534 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2535 		return;
2536 	}
2537 
2538 	j = pci_vpd_lrdt_size(&vpd_data[i]);
2539 	i += PCI_VPD_LRDT_TAG_SIZE;
2540 	if (i + j > vpd_size)
2541 		j = vpd_size - i;
2542 
2543 	/* Get the Part number */
2544 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2545 	if (i < 0) {
2546 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2547 		return;
2548 	}
2549 
2550 	j = pci_vpd_info_field_size(&vpd_data[i]);
2551 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
2552 	if (i + j > vpd_size) {
2553 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2554 		return;
2555 	}
2556 
2557 	netif_info(efx, drv, efx->net_dev,
2558 		   "Part Number : %.*s\n", j, &vpd_data[i]);
2559 }
2560 
2561 
2562 /* Main body of NIC initialisation
2563  * This is called at module load (or hotplug insertion, theoretically).
2564  */
2565 static int efx_pci_probe_main(struct efx_nic *efx)
2566 {
2567 	int rc;
2568 
2569 	/* Do start-of-day initialisation */
2570 	rc = efx_probe_all(efx);
2571 	if (rc)
2572 		goto fail1;
2573 
2574 	efx_init_napi(efx);
2575 
2576 	rc = efx->type->init(efx);
2577 	if (rc) {
2578 		netif_err(efx, probe, efx->net_dev,
2579 			  "failed to initialise NIC\n");
2580 		goto fail3;
2581 	}
2582 
2583 	rc = efx_init_port(efx);
2584 	if (rc) {
2585 		netif_err(efx, probe, efx->net_dev,
2586 			  "failed to initialise port\n");
2587 		goto fail4;
2588 	}
2589 
2590 	rc = efx_nic_init_interrupt(efx);
2591 	if (rc)
2592 		goto fail5;
2593 	efx_start_interrupts(efx, false);
2594 
2595 	return 0;
2596 
2597  fail5:
2598 	efx_fini_port(efx);
2599  fail4:
2600 	efx->type->fini(efx);
2601  fail3:
2602 	efx_fini_napi(efx);
2603 	efx_remove_all(efx);
2604  fail1:
2605 	return rc;
2606 }
2607 
2608 /* NIC initialisation
2609  *
2610  * This is called at module load (or hotplug insertion,
2611  * theoretically).  It sets up PCI mappings, resets the NIC,
2612  * sets up and registers the network devices with the kernel and hooks
2613  * the interrupt service routine.  It does not prepare the device for
2614  * transmission; this is left to the first time one of the network
2615  * interfaces is brought up (i.e. efx_net_open).
2616  */
2617 static int __devinit efx_pci_probe(struct pci_dev *pci_dev,
2618 				   const struct pci_device_id *entry)
2619 {
2620 	const struct efx_nic_type *type = (const struct efx_nic_type *) entry->driver_data;
2621 	struct net_device *net_dev;
2622 	struct efx_nic *efx;
2623 	int rc;
2624 
2625 	/* Allocate and initialise a struct net_device and struct efx_nic */
2626 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2627 				     EFX_MAX_RX_QUEUES);
2628 	if (!net_dev)
2629 		return -ENOMEM;
2630 	net_dev->features |= (type->offload_features | NETIF_F_SG |
2631 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
2632 			      NETIF_F_RXCSUM);
2633 	if (type->offload_features & NETIF_F_V6_CSUM)
2634 		net_dev->features |= NETIF_F_TSO6;
2635 	/* Mask for features that also apply to VLAN devices */
2636 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2637 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2638 				   NETIF_F_RXCSUM);
2639 	/* All offloads can be toggled */
2640 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2641 	efx = netdev_priv(net_dev);
2642 	pci_set_drvdata(pci_dev, efx);
2643 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2644 	rc = efx_init_struct(efx, type, pci_dev, net_dev);
2645 	if (rc)
2646 		goto fail1;
2647 
2648 	netif_info(efx, probe, efx->net_dev,
2649 		   "Solarflare NIC detected\n");
2650 
2651 	efx_print_product_vpd(efx);
2652 
2653 	/* Set up basic I/O (BAR mappings etc) */
2654 	rc = efx_init_io(efx);
2655 	if (rc)
2656 		goto fail2;
2657 
2658 	rc = efx_pci_probe_main(efx);
2659 
2660 	/* Serialise against efx_reset(). No more resets will be
2661 	 * scheduled since efx_stop_all() has been called, and we have
2662 	 * not and never have been registered.
2663 	 */
2664 	cancel_work_sync(&efx->reset_work);
2665 
2666 	if (rc)
2667 		goto fail3;
2668 
2669 	/* If there was a scheduled reset during probe, the NIC is
2670 	 * probably hosed anyway.
2671 	 */
2672 	if (efx->reset_pending) {
2673 		rc = -EIO;
2674 		goto fail4;
2675 	}
2676 
2677 	/* Switch to the running state before we expose the device to the OS,
2678 	 * so that dev_open()|efx_start_all() will actually start the device */
2679 	efx->state = STATE_RUNNING;
2680 
2681 	rc = efx_register_netdev(efx);
2682 	if (rc)
2683 		goto fail4;
2684 
2685 	rc = efx_sriov_init(efx);
2686 	if (rc)
2687 		netif_err(efx, probe, efx->net_dev,
2688 			  "SR-IOV can't be enabled rc %d\n", rc);
2689 
2690 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2691 
2692 	/* Try to create MTDs, but allow this to fail */
2693 	rtnl_lock();
2694 	rc = efx_mtd_probe(efx);
2695 	rtnl_unlock();
2696 	if (rc)
2697 		netif_warn(efx, probe, efx->net_dev,
2698 			   "failed to create MTDs (%d)\n", rc);
2699 
2700 	return 0;
2701 
2702  fail4:
2703 	efx_pci_remove_main(efx);
2704  fail3:
2705 	efx_fini_io(efx);
2706  fail2:
2707 	efx_fini_struct(efx);
2708  fail1:
2709 	pci_set_drvdata(pci_dev, NULL);
2710 	WARN_ON(rc > 0);
2711 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2712 	free_netdev(net_dev);
2713 	return rc;
2714 }
2715 
2716 static int efx_pm_freeze(struct device *dev)
2717 {
2718 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2719 
2720 	efx->state = STATE_FINI;
2721 
2722 	netif_device_detach(efx->net_dev);
2723 
2724 	efx_stop_all(efx);
2725 	efx_stop_interrupts(efx, false);
2726 
2727 	return 0;
2728 }
2729 
2730 static int efx_pm_thaw(struct device *dev)
2731 {
2732 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2733 
2734 	efx->state = STATE_INIT;
2735 
2736 	efx_start_interrupts(efx, false);
2737 
2738 	mutex_lock(&efx->mac_lock);
2739 	efx->phy_op->reconfigure(efx);
2740 	mutex_unlock(&efx->mac_lock);
2741 
2742 	efx_start_all(efx);
2743 
2744 	netif_device_attach(efx->net_dev);
2745 
2746 	efx->state = STATE_RUNNING;
2747 
2748 	efx->type->resume_wol(efx);
2749 
2750 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2751 	queue_work(reset_workqueue, &efx->reset_work);
2752 
2753 	return 0;
2754 }
2755 
2756 static int efx_pm_poweroff(struct device *dev)
2757 {
2758 	struct pci_dev *pci_dev = to_pci_dev(dev);
2759 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2760 
2761 	efx->type->fini(efx);
2762 
2763 	efx->reset_pending = 0;
2764 
2765 	pci_save_state(pci_dev);
2766 	return pci_set_power_state(pci_dev, PCI_D3hot);
2767 }
2768 
2769 /* Used for both resume and restore */
2770 static int efx_pm_resume(struct device *dev)
2771 {
2772 	struct pci_dev *pci_dev = to_pci_dev(dev);
2773 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2774 	int rc;
2775 
2776 	rc = pci_set_power_state(pci_dev, PCI_D0);
2777 	if (rc)
2778 		return rc;
2779 	pci_restore_state(pci_dev);
2780 	rc = pci_enable_device(pci_dev);
2781 	if (rc)
2782 		return rc;
2783 	pci_set_master(efx->pci_dev);
2784 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
2785 	if (rc)
2786 		return rc;
2787 	rc = efx->type->init(efx);
2788 	if (rc)
2789 		return rc;
2790 	efx_pm_thaw(dev);
2791 	return 0;
2792 }
2793 
2794 static int efx_pm_suspend(struct device *dev)
2795 {
2796 	int rc;
2797 
2798 	efx_pm_freeze(dev);
2799 	rc = efx_pm_poweroff(dev);
2800 	if (rc)
2801 		efx_pm_resume(dev);
2802 	return rc;
2803 }
2804 
2805 static const struct dev_pm_ops efx_pm_ops = {
2806 	.suspend	= efx_pm_suspend,
2807 	.resume		= efx_pm_resume,
2808 	.freeze		= efx_pm_freeze,
2809 	.thaw		= efx_pm_thaw,
2810 	.poweroff	= efx_pm_poweroff,
2811 	.restore	= efx_pm_resume,
2812 };
2813 
2814 static struct pci_driver efx_pci_driver = {
2815 	.name		= KBUILD_MODNAME,
2816 	.id_table	= efx_pci_table,
2817 	.probe		= efx_pci_probe,
2818 	.remove		= efx_pci_remove,
2819 	.driver.pm	= &efx_pm_ops,
2820 };
2821 
2822 /**************************************************************************
2823  *
2824  * Kernel module interface
2825  *
2826  *************************************************************************/
2827 
2828 module_param(interrupt_mode, uint, 0444);
2829 MODULE_PARM_DESC(interrupt_mode,
2830 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2831 
2832 static int __init efx_init_module(void)
2833 {
2834 	int rc;
2835 
2836 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
2837 
2838 	rc = register_netdevice_notifier(&efx_netdev_notifier);
2839 	if (rc)
2840 		goto err_notifier;
2841 
2842 	rc = efx_init_sriov();
2843 	if (rc)
2844 		goto err_sriov;
2845 
2846 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
2847 	if (!reset_workqueue) {
2848 		rc = -ENOMEM;
2849 		goto err_reset;
2850 	}
2851 
2852 	rc = pci_register_driver(&efx_pci_driver);
2853 	if (rc < 0)
2854 		goto err_pci;
2855 
2856 	return 0;
2857 
2858  err_pci:
2859 	destroy_workqueue(reset_workqueue);
2860  err_reset:
2861 	efx_fini_sriov();
2862  err_sriov:
2863 	unregister_netdevice_notifier(&efx_netdev_notifier);
2864  err_notifier:
2865 	return rc;
2866 }
2867 
2868 static void __exit efx_exit_module(void)
2869 {
2870 	printk(KERN_INFO "Solarflare NET driver unloading\n");
2871 
2872 	pci_unregister_driver(&efx_pci_driver);
2873 	destroy_workqueue(reset_workqueue);
2874 	efx_fini_sriov();
2875 	unregister_netdevice_notifier(&efx_netdev_notifier);
2876 
2877 }
2878 
2879 module_init(efx_init_module);
2880 module_exit(efx_exit_module);
2881 
2882 MODULE_AUTHOR("Solarflare Communications and "
2883 	      "Michael Brown <mbrown@fensystems.co.uk>");
2884 MODULE_DESCRIPTION("Solarflare Communications network driver");
2885 MODULE_LICENSE("GPL");
2886 MODULE_DEVICE_TABLE(pci, efx_pci_table);
2887