xref: /openbmc/linux/drivers/net/ethernet/sfc/efx.c (revision f35e839a)
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 <linux/aer.h>
26 #include "net_driver.h"
27 #include "efx.h"
28 #include "nic.h"
29 #include "selftest.h"
30 
31 #include "mcdi.h"
32 #include "workarounds.h"
33 
34 /**************************************************************************
35  *
36  * Type name strings
37  *
38  **************************************************************************
39  */
40 
41 /* Loopback mode names (see LOOPBACK_MODE()) */
42 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
43 const char *const efx_loopback_mode_names[] = {
44 	[LOOPBACK_NONE]		= "NONE",
45 	[LOOPBACK_DATA]		= "DATAPATH",
46 	[LOOPBACK_GMAC]		= "GMAC",
47 	[LOOPBACK_XGMII]	= "XGMII",
48 	[LOOPBACK_XGXS]		= "XGXS",
49 	[LOOPBACK_XAUI]		= "XAUI",
50 	[LOOPBACK_GMII]		= "GMII",
51 	[LOOPBACK_SGMII]	= "SGMII",
52 	[LOOPBACK_XGBR]		= "XGBR",
53 	[LOOPBACK_XFI]		= "XFI",
54 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
55 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
56 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
57 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
58 	[LOOPBACK_GPHY]		= "GPHY",
59 	[LOOPBACK_PHYXS]	= "PHYXS",
60 	[LOOPBACK_PCS]		= "PCS",
61 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
62 	[LOOPBACK_XPORT]	= "XPORT",
63 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
64 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
65 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
66 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
67 	[LOOPBACK_GMII_WS]	= "GMII_WS",
68 	[LOOPBACK_XFI_WS]	= "XFI_WS",
69 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
70 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
71 };
72 
73 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
74 const char *const efx_reset_type_names[] = {
75 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
76 	[RESET_TYPE_ALL]                = "ALL",
77 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
78 	[RESET_TYPE_WORLD]              = "WORLD",
79 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
80 	[RESET_TYPE_DISABLE]            = "DISABLE",
81 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
82 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
83 	[RESET_TYPE_RX_RECOVERY]        = "RX_RECOVERY",
84 	[RESET_TYPE_RX_DESC_FETCH]      = "RX_DESC_FETCH",
85 	[RESET_TYPE_TX_DESC_FETCH]      = "TX_DESC_FETCH",
86 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
87 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
88 };
89 
90 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
91  * queued onto this work queue. This is not a per-nic work queue, because
92  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
93  */
94 static struct workqueue_struct *reset_workqueue;
95 
96 /**************************************************************************
97  *
98  * Configurable values
99  *
100  *************************************************************************/
101 
102 /*
103  * Use separate channels for TX and RX events
104  *
105  * Set this to 1 to use separate channels for TX and RX. It allows us
106  * to control interrupt affinity separately for TX and RX.
107  *
108  * This is only used in MSI-X interrupt mode
109  */
110 static bool separate_tx_channels;
111 module_param(separate_tx_channels, bool, 0444);
112 MODULE_PARM_DESC(separate_tx_channels,
113 		 "Use separate channels for TX and RX");
114 
115 /* This is the weight assigned to each of the (per-channel) virtual
116  * NAPI devices.
117  */
118 static int napi_weight = 64;
119 
120 /* This is the time (in jiffies) between invocations of the hardware
121  * monitor.
122  * On Falcon-based NICs, this will:
123  * - Check the on-board hardware monitor;
124  * - Poll the link state and reconfigure the hardware as necessary.
125  * On Siena-based NICs for power systems with EEH support, this will give EEH a
126  * chance to start.
127  */
128 static unsigned int efx_monitor_interval = 1 * HZ;
129 
130 /* Initial interrupt moderation settings.  They can be modified after
131  * module load with ethtool.
132  *
133  * The default for RX should strike a balance between increasing the
134  * round-trip latency and reducing overhead.
135  */
136 static unsigned int rx_irq_mod_usec = 60;
137 
138 /* Initial interrupt moderation settings.  They can be modified after
139  * module load with ethtool.
140  *
141  * This default is chosen to ensure that a 10G link does not go idle
142  * while a TX queue is stopped after it has become full.  A queue is
143  * restarted when it drops below half full.  The time this takes (assuming
144  * worst case 3 descriptors per packet and 1024 descriptors) is
145  *   512 / 3 * 1.2 = 205 usec.
146  */
147 static unsigned int tx_irq_mod_usec = 150;
148 
149 /* This is the first interrupt mode to try out of:
150  * 0 => MSI-X
151  * 1 => MSI
152  * 2 => legacy
153  */
154 static unsigned int interrupt_mode;
155 
156 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
157  * i.e. the number of CPUs among which we may distribute simultaneous
158  * interrupt handling.
159  *
160  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
161  * The default (0) means to assign an interrupt to each core.
162  */
163 static unsigned int rss_cpus;
164 module_param(rss_cpus, uint, 0444);
165 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
166 
167 static bool phy_flash_cfg;
168 module_param(phy_flash_cfg, bool, 0644);
169 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
170 
171 static unsigned irq_adapt_low_thresh = 8000;
172 module_param(irq_adapt_low_thresh, uint, 0644);
173 MODULE_PARM_DESC(irq_adapt_low_thresh,
174 		 "Threshold score for reducing IRQ moderation");
175 
176 static unsigned irq_adapt_high_thresh = 16000;
177 module_param(irq_adapt_high_thresh, uint, 0644);
178 MODULE_PARM_DESC(irq_adapt_high_thresh,
179 		 "Threshold score for increasing IRQ moderation");
180 
181 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
182 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
183 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
184 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
185 module_param(debug, uint, 0);
186 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
187 
188 /**************************************************************************
189  *
190  * Utility functions and prototypes
191  *
192  *************************************************************************/
193 
194 static void efx_start_interrupts(struct efx_nic *efx, bool may_keep_eventq);
195 static void efx_stop_interrupts(struct efx_nic *efx, bool may_keep_eventq);
196 static void efx_remove_channel(struct efx_channel *channel);
197 static void efx_remove_channels(struct efx_nic *efx);
198 static const struct efx_channel_type efx_default_channel_type;
199 static void efx_remove_port(struct efx_nic *efx);
200 static void efx_init_napi_channel(struct efx_channel *channel);
201 static void efx_fini_napi(struct efx_nic *efx);
202 static void efx_fini_napi_channel(struct efx_channel *channel);
203 static void efx_fini_struct(struct efx_nic *efx);
204 static void efx_start_all(struct efx_nic *efx);
205 static void efx_stop_all(struct efx_nic *efx);
206 
207 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
208 	do {						\
209 		if ((efx->state == STATE_READY) ||	\
210 		    (efx->state == STATE_RECOVERY) ||	\
211 		    (efx->state == STATE_DISABLED))	\
212 			ASSERT_RTNL();			\
213 	} while (0)
214 
215 static int efx_check_disabled(struct efx_nic *efx)
216 {
217 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
218 		netif_err(efx, drv, efx->net_dev,
219 			  "device is disabled due to earlier errors\n");
220 		return -EIO;
221 	}
222 	return 0;
223 }
224 
225 /**************************************************************************
226  *
227  * Event queue processing
228  *
229  *************************************************************************/
230 
231 /* Process channel's event queue
232  *
233  * This function is responsible for processing the event queue of a
234  * single channel.  The caller must guarantee that this function will
235  * never be concurrently called more than once on the same channel,
236  * though different channels may be being processed concurrently.
237  */
238 static int efx_process_channel(struct efx_channel *channel, int budget)
239 {
240 	int spent;
241 
242 	if (unlikely(!channel->enabled))
243 		return 0;
244 
245 	spent = efx_nic_process_eventq(channel, budget);
246 	if (spent && efx_channel_has_rx_queue(channel)) {
247 		struct efx_rx_queue *rx_queue =
248 			efx_channel_get_rx_queue(channel);
249 
250 		efx_rx_flush_packet(channel);
251 		if (rx_queue->enabled)
252 			efx_fast_push_rx_descriptors(rx_queue);
253 	}
254 
255 	return spent;
256 }
257 
258 /* Mark channel as finished processing
259  *
260  * Note that since we will not receive further interrupts for this
261  * channel before we finish processing and call the eventq_read_ack()
262  * method, there is no need to use the interrupt hold-off timers.
263  */
264 static inline void efx_channel_processed(struct efx_channel *channel)
265 {
266 	/* The interrupt handler for this channel may set work_pending
267 	 * as soon as we acknowledge the events we've seen.  Make sure
268 	 * it's cleared before then. */
269 	channel->work_pending = false;
270 	smp_wmb();
271 
272 	efx_nic_eventq_read_ack(channel);
273 }
274 
275 /* NAPI poll handler
276  *
277  * NAPI guarantees serialisation of polls of the same device, which
278  * provides the guarantee required by efx_process_channel().
279  */
280 static int efx_poll(struct napi_struct *napi, int budget)
281 {
282 	struct efx_channel *channel =
283 		container_of(napi, struct efx_channel, napi_str);
284 	struct efx_nic *efx = channel->efx;
285 	int spent;
286 
287 	netif_vdbg(efx, intr, efx->net_dev,
288 		   "channel %d NAPI poll executing on CPU %d\n",
289 		   channel->channel, raw_smp_processor_id());
290 
291 	spent = efx_process_channel(channel, budget);
292 
293 	if (spent < budget) {
294 		if (efx_channel_has_rx_queue(channel) &&
295 		    efx->irq_rx_adaptive &&
296 		    unlikely(++channel->irq_count == 1000)) {
297 			if (unlikely(channel->irq_mod_score <
298 				     irq_adapt_low_thresh)) {
299 				if (channel->irq_moderation > 1) {
300 					channel->irq_moderation -= 1;
301 					efx->type->push_irq_moderation(channel);
302 				}
303 			} else if (unlikely(channel->irq_mod_score >
304 					    irq_adapt_high_thresh)) {
305 				if (channel->irq_moderation <
306 				    efx->irq_rx_moderation) {
307 					channel->irq_moderation += 1;
308 					efx->type->push_irq_moderation(channel);
309 				}
310 			}
311 			channel->irq_count = 0;
312 			channel->irq_mod_score = 0;
313 		}
314 
315 		efx_filter_rfs_expire(channel);
316 
317 		/* There is no race here; although napi_disable() will
318 		 * only wait for napi_complete(), this isn't a problem
319 		 * since efx_channel_processed() will have no effect if
320 		 * interrupts have already been disabled.
321 		 */
322 		napi_complete(napi);
323 		efx_channel_processed(channel);
324 	}
325 
326 	return spent;
327 }
328 
329 /* Process the eventq of the specified channel immediately on this CPU
330  *
331  * Disable hardware generated interrupts, wait for any existing
332  * processing to finish, then directly poll (and ack ) the eventq.
333  * Finally reenable NAPI and interrupts.
334  *
335  * This is for use only during a loopback self-test.  It must not
336  * deliver any packets up the stack as this can result in deadlock.
337  */
338 void efx_process_channel_now(struct efx_channel *channel)
339 {
340 	struct efx_nic *efx = channel->efx;
341 
342 	BUG_ON(channel->channel >= efx->n_channels);
343 	BUG_ON(!channel->enabled);
344 	BUG_ON(!efx->loopback_selftest);
345 
346 	/* Disable interrupts and wait for ISRs to complete */
347 	efx_nic_disable_interrupts(efx);
348 	if (efx->legacy_irq) {
349 		synchronize_irq(efx->legacy_irq);
350 		efx->legacy_irq_enabled = false;
351 	}
352 	if (channel->irq)
353 		synchronize_irq(channel->irq);
354 
355 	/* Wait for any NAPI processing to complete */
356 	napi_disable(&channel->napi_str);
357 
358 	/* Poll the channel */
359 	efx_process_channel(channel, channel->eventq_mask + 1);
360 
361 	/* Ack the eventq. This may cause an interrupt to be generated
362 	 * when they are reenabled */
363 	efx_channel_processed(channel);
364 
365 	napi_enable(&channel->napi_str);
366 	if (efx->legacy_irq)
367 		efx->legacy_irq_enabled = true;
368 	efx_nic_enable_interrupts(efx);
369 }
370 
371 /* Create event queue
372  * Event queue memory allocations are done only once.  If the channel
373  * is reset, the memory buffer will be reused; this guards against
374  * errors during channel reset and also simplifies interrupt handling.
375  */
376 static int efx_probe_eventq(struct efx_channel *channel)
377 {
378 	struct efx_nic *efx = channel->efx;
379 	unsigned long entries;
380 
381 	netif_dbg(efx, probe, efx->net_dev,
382 		  "chan %d create event queue\n", channel->channel);
383 
384 	/* Build an event queue with room for one event per tx and rx buffer,
385 	 * plus some extra for link state events and MCDI completions. */
386 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
387 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
388 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
389 
390 	return efx_nic_probe_eventq(channel);
391 }
392 
393 /* Prepare channel's event queue */
394 static void efx_init_eventq(struct efx_channel *channel)
395 {
396 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
397 		  "chan %d init event queue\n", channel->channel);
398 
399 	channel->eventq_read_ptr = 0;
400 
401 	efx_nic_init_eventq(channel);
402 }
403 
404 /* Enable event queue processing and NAPI */
405 static void efx_start_eventq(struct efx_channel *channel)
406 {
407 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
408 		  "chan %d start event queue\n", channel->channel);
409 
410 	/* The interrupt handler for this channel may set work_pending
411 	 * as soon as we enable it.  Make sure it's cleared before
412 	 * then.  Similarly, make sure it sees the enabled flag set.
413 	 */
414 	channel->work_pending = false;
415 	channel->enabled = true;
416 	smp_wmb();
417 
418 	napi_enable(&channel->napi_str);
419 	efx_nic_eventq_read_ack(channel);
420 }
421 
422 /* Disable event queue processing and NAPI */
423 static void efx_stop_eventq(struct efx_channel *channel)
424 {
425 	if (!channel->enabled)
426 		return;
427 
428 	napi_disable(&channel->napi_str);
429 	channel->enabled = false;
430 }
431 
432 static void efx_fini_eventq(struct efx_channel *channel)
433 {
434 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
435 		  "chan %d fini event queue\n", channel->channel);
436 
437 	efx_nic_fini_eventq(channel);
438 }
439 
440 static void efx_remove_eventq(struct efx_channel *channel)
441 {
442 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
443 		  "chan %d remove event queue\n", channel->channel);
444 
445 	efx_nic_remove_eventq(channel);
446 }
447 
448 /**************************************************************************
449  *
450  * Channel handling
451  *
452  *************************************************************************/
453 
454 /* Allocate and initialise a channel structure. */
455 static struct efx_channel *
456 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
457 {
458 	struct efx_channel *channel;
459 	struct efx_rx_queue *rx_queue;
460 	struct efx_tx_queue *tx_queue;
461 	int j;
462 
463 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
464 	if (!channel)
465 		return NULL;
466 
467 	channel->efx = efx;
468 	channel->channel = i;
469 	channel->type = &efx_default_channel_type;
470 
471 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
472 		tx_queue = &channel->tx_queue[j];
473 		tx_queue->efx = efx;
474 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
475 		tx_queue->channel = channel;
476 	}
477 
478 	rx_queue = &channel->rx_queue;
479 	rx_queue->efx = efx;
480 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
481 		    (unsigned long)rx_queue);
482 
483 	return channel;
484 }
485 
486 /* Allocate and initialise a channel structure, copying parameters
487  * (but not resources) from an old channel structure.
488  */
489 static struct efx_channel *
490 efx_copy_channel(const struct efx_channel *old_channel)
491 {
492 	struct efx_channel *channel;
493 	struct efx_rx_queue *rx_queue;
494 	struct efx_tx_queue *tx_queue;
495 	int j;
496 
497 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
498 	if (!channel)
499 		return NULL;
500 
501 	*channel = *old_channel;
502 
503 	channel->napi_dev = NULL;
504 	memset(&channel->eventq, 0, sizeof(channel->eventq));
505 
506 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
507 		tx_queue = &channel->tx_queue[j];
508 		if (tx_queue->channel)
509 			tx_queue->channel = channel;
510 		tx_queue->buffer = NULL;
511 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
512 	}
513 
514 	rx_queue = &channel->rx_queue;
515 	rx_queue->buffer = NULL;
516 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
517 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
518 		    (unsigned long)rx_queue);
519 
520 	return channel;
521 }
522 
523 static int efx_probe_channel(struct efx_channel *channel)
524 {
525 	struct efx_tx_queue *tx_queue;
526 	struct efx_rx_queue *rx_queue;
527 	int rc;
528 
529 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
530 		  "creating channel %d\n", channel->channel);
531 
532 	rc = channel->type->pre_probe(channel);
533 	if (rc)
534 		goto fail;
535 
536 	rc = efx_probe_eventq(channel);
537 	if (rc)
538 		goto fail;
539 
540 	efx_for_each_channel_tx_queue(tx_queue, channel) {
541 		rc = efx_probe_tx_queue(tx_queue);
542 		if (rc)
543 			goto fail;
544 	}
545 
546 	efx_for_each_channel_rx_queue(rx_queue, channel) {
547 		rc = efx_probe_rx_queue(rx_queue);
548 		if (rc)
549 			goto fail;
550 	}
551 
552 	channel->n_rx_frm_trunc = 0;
553 
554 	return 0;
555 
556 fail:
557 	efx_remove_channel(channel);
558 	return rc;
559 }
560 
561 static void
562 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
563 {
564 	struct efx_nic *efx = channel->efx;
565 	const char *type;
566 	int number;
567 
568 	number = channel->channel;
569 	if (efx->tx_channel_offset == 0) {
570 		type = "";
571 	} else if (channel->channel < efx->tx_channel_offset) {
572 		type = "-rx";
573 	} else {
574 		type = "-tx";
575 		number -= efx->tx_channel_offset;
576 	}
577 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
578 }
579 
580 static void efx_set_channel_names(struct efx_nic *efx)
581 {
582 	struct efx_channel *channel;
583 
584 	efx_for_each_channel(channel, efx)
585 		channel->type->get_name(channel,
586 					efx->channel_name[channel->channel],
587 					sizeof(efx->channel_name[0]));
588 }
589 
590 static int efx_probe_channels(struct efx_nic *efx)
591 {
592 	struct efx_channel *channel;
593 	int rc;
594 
595 	/* Restart special buffer allocation */
596 	efx->next_buffer_table = 0;
597 
598 	/* Probe channels in reverse, so that any 'extra' channels
599 	 * use the start of the buffer table. This allows the traffic
600 	 * channels to be resized without moving them or wasting the
601 	 * entries before them.
602 	 */
603 	efx_for_each_channel_rev(channel, efx) {
604 		rc = efx_probe_channel(channel);
605 		if (rc) {
606 			netif_err(efx, probe, efx->net_dev,
607 				  "failed to create channel %d\n",
608 				  channel->channel);
609 			goto fail;
610 		}
611 	}
612 	efx_set_channel_names(efx);
613 
614 	return 0;
615 
616 fail:
617 	efx_remove_channels(efx);
618 	return rc;
619 }
620 
621 /* Channels are shutdown and reinitialised whilst the NIC is running
622  * to propagate configuration changes (mtu, checksum offload), or
623  * to clear hardware error conditions
624  */
625 static void efx_start_datapath(struct efx_nic *efx)
626 {
627 	bool old_rx_scatter = efx->rx_scatter;
628 	struct efx_tx_queue *tx_queue;
629 	struct efx_rx_queue *rx_queue;
630 	struct efx_channel *channel;
631 	size_t rx_buf_len;
632 
633 	/* Calculate the rx buffer allocation parameters required to
634 	 * support the current MTU, including padding for header
635 	 * alignment and overruns.
636 	 */
637 	efx->rx_dma_len = (efx->type->rx_buffer_hash_size +
638 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
639 			   efx->type->rx_buffer_padding);
640 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
641 		      EFX_PAGE_IP_ALIGN + efx->rx_dma_len);
642 	if (rx_buf_len <= PAGE_SIZE) {
643 		efx->rx_scatter = false;
644 		efx->rx_buffer_order = 0;
645 	} else if (efx->type->can_rx_scatter) {
646 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
647 			     EFX_PAGE_IP_ALIGN + EFX_RX_USR_BUF_SIZE >
648 			     PAGE_SIZE / 2);
649 		efx->rx_scatter = true;
650 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
651 		efx->rx_buffer_order = 0;
652 	} else {
653 		efx->rx_scatter = false;
654 		efx->rx_buffer_order = get_order(rx_buf_len);
655 	}
656 
657 	efx_rx_config_page_split(efx);
658 	if (efx->rx_buffer_order)
659 		netif_dbg(efx, drv, efx->net_dev,
660 			  "RX buf len=%u; page order=%u batch=%u\n",
661 			  efx->rx_dma_len, efx->rx_buffer_order,
662 			  efx->rx_pages_per_batch);
663 	else
664 		netif_dbg(efx, drv, efx->net_dev,
665 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
666 			  efx->rx_dma_len, efx->rx_page_buf_step,
667 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
668 
669 	/* RX filters also have scatter-enabled flags */
670 	if (efx->rx_scatter != old_rx_scatter)
671 		efx_filter_update_rx_scatter(efx);
672 
673 	/* We must keep at least one descriptor in a TX ring empty.
674 	 * We could avoid this when the queue size does not exactly
675 	 * match the hardware ring size, but it's not that important.
676 	 * Therefore we stop the queue when one more skb might fill
677 	 * the ring completely.  We wake it when half way back to
678 	 * empty.
679 	 */
680 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
681 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
682 
683 	/* Initialise the channels */
684 	efx_for_each_channel(channel, efx) {
685 		efx_for_each_channel_tx_queue(tx_queue, channel)
686 			efx_init_tx_queue(tx_queue);
687 
688 		efx_for_each_channel_rx_queue(rx_queue, channel) {
689 			efx_init_rx_queue(rx_queue);
690 			efx_nic_generate_fill_event(rx_queue);
691 		}
692 
693 		WARN_ON(channel->rx_pkt_n_frags);
694 	}
695 
696 	if (netif_device_present(efx->net_dev))
697 		netif_tx_wake_all_queues(efx->net_dev);
698 }
699 
700 static void efx_stop_datapath(struct efx_nic *efx)
701 {
702 	struct efx_channel *channel;
703 	struct efx_tx_queue *tx_queue;
704 	struct efx_rx_queue *rx_queue;
705 	struct pci_dev *dev = efx->pci_dev;
706 	int rc;
707 
708 	EFX_ASSERT_RESET_SERIALISED(efx);
709 	BUG_ON(efx->port_enabled);
710 
711 	/* Only perform flush if dma is enabled */
712 	if (dev->is_busmaster && efx->state != STATE_RECOVERY) {
713 		rc = efx_nic_flush_queues(efx);
714 
715 		if (rc && EFX_WORKAROUND_7803(efx)) {
716 			/* Schedule a reset to recover from the flush failure. The
717 			 * descriptor caches reference memory we're about to free,
718 			 * but falcon_reconfigure_mac_wrapper() won't reconnect
719 			 * the MACs because of the pending reset. */
720 			netif_err(efx, drv, efx->net_dev,
721 				  "Resetting to recover from flush failure\n");
722 			efx_schedule_reset(efx, RESET_TYPE_ALL);
723 		} else if (rc) {
724 			netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
725 		} else {
726 			netif_dbg(efx, drv, efx->net_dev,
727 				  "successfully flushed all queues\n");
728 		}
729 	}
730 
731 	efx_for_each_channel(channel, efx) {
732 		/* RX packet processing is pipelined, so wait for the
733 		 * NAPI handler to complete.  At least event queue 0
734 		 * might be kept active by non-data events, so don't
735 		 * use napi_synchronize() but actually disable NAPI
736 		 * temporarily.
737 		 */
738 		if (efx_channel_has_rx_queue(channel)) {
739 			efx_stop_eventq(channel);
740 			efx_start_eventq(channel);
741 		}
742 
743 		efx_for_each_channel_rx_queue(rx_queue, channel)
744 			efx_fini_rx_queue(rx_queue);
745 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
746 			efx_fini_tx_queue(tx_queue);
747 	}
748 }
749 
750 static void efx_remove_channel(struct efx_channel *channel)
751 {
752 	struct efx_tx_queue *tx_queue;
753 	struct efx_rx_queue *rx_queue;
754 
755 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
756 		  "destroy chan %d\n", channel->channel);
757 
758 	efx_for_each_channel_rx_queue(rx_queue, channel)
759 		efx_remove_rx_queue(rx_queue);
760 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
761 		efx_remove_tx_queue(tx_queue);
762 	efx_remove_eventq(channel);
763 	channel->type->post_remove(channel);
764 }
765 
766 static void efx_remove_channels(struct efx_nic *efx)
767 {
768 	struct efx_channel *channel;
769 
770 	efx_for_each_channel(channel, efx)
771 		efx_remove_channel(channel);
772 }
773 
774 int
775 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
776 {
777 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
778 	u32 old_rxq_entries, old_txq_entries;
779 	unsigned i, next_buffer_table = 0;
780 	int rc;
781 
782 	rc = efx_check_disabled(efx);
783 	if (rc)
784 		return rc;
785 
786 	/* Not all channels should be reallocated. We must avoid
787 	 * reallocating their buffer table entries.
788 	 */
789 	efx_for_each_channel(channel, efx) {
790 		struct efx_rx_queue *rx_queue;
791 		struct efx_tx_queue *tx_queue;
792 
793 		if (channel->type->copy)
794 			continue;
795 		next_buffer_table = max(next_buffer_table,
796 					channel->eventq.index +
797 					channel->eventq.entries);
798 		efx_for_each_channel_rx_queue(rx_queue, channel)
799 			next_buffer_table = max(next_buffer_table,
800 						rx_queue->rxd.index +
801 						rx_queue->rxd.entries);
802 		efx_for_each_channel_tx_queue(tx_queue, channel)
803 			next_buffer_table = max(next_buffer_table,
804 						tx_queue->txd.index +
805 						tx_queue->txd.entries);
806 	}
807 
808 	efx_device_detach_sync(efx);
809 	efx_stop_all(efx);
810 	efx_stop_interrupts(efx, true);
811 
812 	/* Clone channels (where possible) */
813 	memset(other_channel, 0, sizeof(other_channel));
814 	for (i = 0; i < efx->n_channels; i++) {
815 		channel = efx->channel[i];
816 		if (channel->type->copy)
817 			channel = channel->type->copy(channel);
818 		if (!channel) {
819 			rc = -ENOMEM;
820 			goto out;
821 		}
822 		other_channel[i] = channel;
823 	}
824 
825 	/* Swap entry counts and channel pointers */
826 	old_rxq_entries = efx->rxq_entries;
827 	old_txq_entries = efx->txq_entries;
828 	efx->rxq_entries = rxq_entries;
829 	efx->txq_entries = txq_entries;
830 	for (i = 0; i < efx->n_channels; i++) {
831 		channel = efx->channel[i];
832 		efx->channel[i] = other_channel[i];
833 		other_channel[i] = channel;
834 	}
835 
836 	/* Restart buffer table allocation */
837 	efx->next_buffer_table = next_buffer_table;
838 
839 	for (i = 0; i < efx->n_channels; i++) {
840 		channel = efx->channel[i];
841 		if (!channel->type->copy)
842 			continue;
843 		rc = efx_probe_channel(channel);
844 		if (rc)
845 			goto rollback;
846 		efx_init_napi_channel(efx->channel[i]);
847 	}
848 
849 out:
850 	/* Destroy unused channel structures */
851 	for (i = 0; i < efx->n_channels; i++) {
852 		channel = other_channel[i];
853 		if (channel && channel->type->copy) {
854 			efx_fini_napi_channel(channel);
855 			efx_remove_channel(channel);
856 			kfree(channel);
857 		}
858 	}
859 
860 	efx_start_interrupts(efx, true);
861 	efx_start_all(efx);
862 	netif_device_attach(efx->net_dev);
863 	return rc;
864 
865 rollback:
866 	/* Swap back */
867 	efx->rxq_entries = old_rxq_entries;
868 	efx->txq_entries = old_txq_entries;
869 	for (i = 0; i < efx->n_channels; i++) {
870 		channel = efx->channel[i];
871 		efx->channel[i] = other_channel[i];
872 		other_channel[i] = channel;
873 	}
874 	goto out;
875 }
876 
877 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
878 {
879 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
880 }
881 
882 static const struct efx_channel_type efx_default_channel_type = {
883 	.pre_probe		= efx_channel_dummy_op_int,
884 	.post_remove		= efx_channel_dummy_op_void,
885 	.get_name		= efx_get_channel_name,
886 	.copy			= efx_copy_channel,
887 	.keep_eventq		= false,
888 };
889 
890 int efx_channel_dummy_op_int(struct efx_channel *channel)
891 {
892 	return 0;
893 }
894 
895 void efx_channel_dummy_op_void(struct efx_channel *channel)
896 {
897 }
898 
899 /**************************************************************************
900  *
901  * Port handling
902  *
903  **************************************************************************/
904 
905 /* This ensures that the kernel is kept informed (via
906  * netif_carrier_on/off) of the link status, and also maintains the
907  * link status's stop on the port's TX queue.
908  */
909 void efx_link_status_changed(struct efx_nic *efx)
910 {
911 	struct efx_link_state *link_state = &efx->link_state;
912 
913 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
914 	 * that no events are triggered between unregister_netdev() and the
915 	 * driver unloading. A more general condition is that NETDEV_CHANGE
916 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
917 	if (!netif_running(efx->net_dev))
918 		return;
919 
920 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
921 		efx->n_link_state_changes++;
922 
923 		if (link_state->up)
924 			netif_carrier_on(efx->net_dev);
925 		else
926 			netif_carrier_off(efx->net_dev);
927 	}
928 
929 	/* Status message for kernel log */
930 	if (link_state->up)
931 		netif_info(efx, link, efx->net_dev,
932 			   "link up at %uMbps %s-duplex (MTU %d)%s\n",
933 			   link_state->speed, link_state->fd ? "full" : "half",
934 			   efx->net_dev->mtu,
935 			   (efx->promiscuous ? " [PROMISC]" : ""));
936 	else
937 		netif_info(efx, link, efx->net_dev, "link down\n");
938 }
939 
940 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
941 {
942 	efx->link_advertising = advertising;
943 	if (advertising) {
944 		if (advertising & ADVERTISED_Pause)
945 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
946 		else
947 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
948 		if (advertising & ADVERTISED_Asym_Pause)
949 			efx->wanted_fc ^= EFX_FC_TX;
950 	}
951 }
952 
953 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
954 {
955 	efx->wanted_fc = wanted_fc;
956 	if (efx->link_advertising) {
957 		if (wanted_fc & EFX_FC_RX)
958 			efx->link_advertising |= (ADVERTISED_Pause |
959 						  ADVERTISED_Asym_Pause);
960 		else
961 			efx->link_advertising &= ~(ADVERTISED_Pause |
962 						   ADVERTISED_Asym_Pause);
963 		if (wanted_fc & EFX_FC_TX)
964 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
965 	}
966 }
967 
968 static void efx_fini_port(struct efx_nic *efx);
969 
970 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
971  * the MAC appropriately. All other PHY configuration changes are pushed
972  * through phy_op->set_settings(), and pushed asynchronously to the MAC
973  * through efx_monitor().
974  *
975  * Callers must hold the mac_lock
976  */
977 int __efx_reconfigure_port(struct efx_nic *efx)
978 {
979 	enum efx_phy_mode phy_mode;
980 	int rc;
981 
982 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
983 
984 	/* Serialise the promiscuous flag with efx_set_rx_mode. */
985 	netif_addr_lock_bh(efx->net_dev);
986 	netif_addr_unlock_bh(efx->net_dev);
987 
988 	/* Disable PHY transmit in mac level loopbacks */
989 	phy_mode = efx->phy_mode;
990 	if (LOOPBACK_INTERNAL(efx))
991 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
992 	else
993 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
994 
995 	rc = efx->type->reconfigure_port(efx);
996 
997 	if (rc)
998 		efx->phy_mode = phy_mode;
999 
1000 	return rc;
1001 }
1002 
1003 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1004  * disabled. */
1005 int efx_reconfigure_port(struct efx_nic *efx)
1006 {
1007 	int rc;
1008 
1009 	EFX_ASSERT_RESET_SERIALISED(efx);
1010 
1011 	mutex_lock(&efx->mac_lock);
1012 	rc = __efx_reconfigure_port(efx);
1013 	mutex_unlock(&efx->mac_lock);
1014 
1015 	return rc;
1016 }
1017 
1018 /* Asynchronous work item for changing MAC promiscuity and multicast
1019  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1020  * MAC directly. */
1021 static void efx_mac_work(struct work_struct *data)
1022 {
1023 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1024 
1025 	mutex_lock(&efx->mac_lock);
1026 	if (efx->port_enabled)
1027 		efx->type->reconfigure_mac(efx);
1028 	mutex_unlock(&efx->mac_lock);
1029 }
1030 
1031 static int efx_probe_port(struct efx_nic *efx)
1032 {
1033 	int rc;
1034 
1035 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1036 
1037 	if (phy_flash_cfg)
1038 		efx->phy_mode = PHY_MODE_SPECIAL;
1039 
1040 	/* Connect up MAC/PHY operations table */
1041 	rc = efx->type->probe_port(efx);
1042 	if (rc)
1043 		return rc;
1044 
1045 	/* Initialise MAC address to permanent address */
1046 	memcpy(efx->net_dev->dev_addr, efx->net_dev->perm_addr, ETH_ALEN);
1047 
1048 	return 0;
1049 }
1050 
1051 static int efx_init_port(struct efx_nic *efx)
1052 {
1053 	int rc;
1054 
1055 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1056 
1057 	mutex_lock(&efx->mac_lock);
1058 
1059 	rc = efx->phy_op->init(efx);
1060 	if (rc)
1061 		goto fail1;
1062 
1063 	efx->port_initialized = true;
1064 
1065 	/* Reconfigure the MAC before creating dma queues (required for
1066 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1067 	efx->type->reconfigure_mac(efx);
1068 
1069 	/* Ensure the PHY advertises the correct flow control settings */
1070 	rc = efx->phy_op->reconfigure(efx);
1071 	if (rc)
1072 		goto fail2;
1073 
1074 	mutex_unlock(&efx->mac_lock);
1075 	return 0;
1076 
1077 fail2:
1078 	efx->phy_op->fini(efx);
1079 fail1:
1080 	mutex_unlock(&efx->mac_lock);
1081 	return rc;
1082 }
1083 
1084 static void efx_start_port(struct efx_nic *efx)
1085 {
1086 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1087 	BUG_ON(efx->port_enabled);
1088 
1089 	mutex_lock(&efx->mac_lock);
1090 	efx->port_enabled = true;
1091 
1092 	/* efx_mac_work() might have been scheduled after efx_stop_port(),
1093 	 * and then cancelled by efx_flush_all() */
1094 	efx->type->reconfigure_mac(efx);
1095 
1096 	mutex_unlock(&efx->mac_lock);
1097 }
1098 
1099 /* Prevent efx_mac_work() and efx_monitor() from working */
1100 static void efx_stop_port(struct efx_nic *efx)
1101 {
1102 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1103 
1104 	mutex_lock(&efx->mac_lock);
1105 	efx->port_enabled = false;
1106 	mutex_unlock(&efx->mac_lock);
1107 
1108 	/* Serialise against efx_set_multicast_list() */
1109 	netif_addr_lock_bh(efx->net_dev);
1110 	netif_addr_unlock_bh(efx->net_dev);
1111 }
1112 
1113 static void efx_fini_port(struct efx_nic *efx)
1114 {
1115 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1116 
1117 	if (!efx->port_initialized)
1118 		return;
1119 
1120 	efx->phy_op->fini(efx);
1121 	efx->port_initialized = false;
1122 
1123 	efx->link_state.up = false;
1124 	efx_link_status_changed(efx);
1125 }
1126 
1127 static void efx_remove_port(struct efx_nic *efx)
1128 {
1129 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1130 
1131 	efx->type->remove_port(efx);
1132 }
1133 
1134 /**************************************************************************
1135  *
1136  * NIC handling
1137  *
1138  **************************************************************************/
1139 
1140 /* This configures the PCI device to enable I/O and DMA. */
1141 static int efx_init_io(struct efx_nic *efx)
1142 {
1143 	struct pci_dev *pci_dev = efx->pci_dev;
1144 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1145 	int rc;
1146 
1147 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1148 
1149 	rc = pci_enable_device(pci_dev);
1150 	if (rc) {
1151 		netif_err(efx, probe, efx->net_dev,
1152 			  "failed to enable PCI device\n");
1153 		goto fail1;
1154 	}
1155 
1156 	pci_set_master(pci_dev);
1157 
1158 	/* Set the PCI DMA mask.  Try all possibilities from our
1159 	 * genuine mask down to 32 bits, because some architectures
1160 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1161 	 * masks event though they reject 46 bit masks.
1162 	 */
1163 	while (dma_mask > 0x7fffffffUL) {
1164 		if (dma_supported(&pci_dev->dev, dma_mask)) {
1165 			rc = dma_set_mask(&pci_dev->dev, dma_mask);
1166 			if (rc == 0)
1167 				break;
1168 		}
1169 		dma_mask >>= 1;
1170 	}
1171 	if (rc) {
1172 		netif_err(efx, probe, efx->net_dev,
1173 			  "could not find a suitable DMA mask\n");
1174 		goto fail2;
1175 	}
1176 	netif_dbg(efx, probe, efx->net_dev,
1177 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1178 	rc = dma_set_coherent_mask(&pci_dev->dev, dma_mask);
1179 	if (rc) {
1180 		/* dma_set_coherent_mask() is not *allowed* to
1181 		 * fail with a mask that dma_set_mask() accepted,
1182 		 * but just in case...
1183 		 */
1184 		netif_err(efx, probe, efx->net_dev,
1185 			  "failed to set consistent DMA mask\n");
1186 		goto fail2;
1187 	}
1188 
1189 	efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1190 	rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1191 	if (rc) {
1192 		netif_err(efx, probe, efx->net_dev,
1193 			  "request for memory BAR failed\n");
1194 		rc = -EIO;
1195 		goto fail3;
1196 	}
1197 	efx->membase = ioremap_nocache(efx->membase_phys,
1198 				       efx->type->mem_map_size);
1199 	if (!efx->membase) {
1200 		netif_err(efx, probe, efx->net_dev,
1201 			  "could not map memory BAR at %llx+%x\n",
1202 			  (unsigned long long)efx->membase_phys,
1203 			  efx->type->mem_map_size);
1204 		rc = -ENOMEM;
1205 		goto fail4;
1206 	}
1207 	netif_dbg(efx, probe, efx->net_dev,
1208 		  "memory BAR at %llx+%x (virtual %p)\n",
1209 		  (unsigned long long)efx->membase_phys,
1210 		  efx->type->mem_map_size, efx->membase);
1211 
1212 	return 0;
1213 
1214  fail4:
1215 	pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1216  fail3:
1217 	efx->membase_phys = 0;
1218  fail2:
1219 	pci_disable_device(efx->pci_dev);
1220  fail1:
1221 	return rc;
1222 }
1223 
1224 static void efx_fini_io(struct efx_nic *efx)
1225 {
1226 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1227 
1228 	if (efx->membase) {
1229 		iounmap(efx->membase);
1230 		efx->membase = NULL;
1231 	}
1232 
1233 	if (efx->membase_phys) {
1234 		pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1235 		efx->membase_phys = 0;
1236 	}
1237 
1238 	pci_disable_device(efx->pci_dev);
1239 }
1240 
1241 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1242 {
1243 	cpumask_var_t thread_mask;
1244 	unsigned int count;
1245 	int cpu;
1246 
1247 	if (rss_cpus) {
1248 		count = rss_cpus;
1249 	} else {
1250 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1251 			netif_warn(efx, probe, efx->net_dev,
1252 				   "RSS disabled due to allocation failure\n");
1253 			return 1;
1254 		}
1255 
1256 		count = 0;
1257 		for_each_online_cpu(cpu) {
1258 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1259 				++count;
1260 				cpumask_or(thread_mask, thread_mask,
1261 					   topology_thread_cpumask(cpu));
1262 			}
1263 		}
1264 
1265 		free_cpumask_var(thread_mask);
1266 	}
1267 
1268 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1269 	 * table entries that are inaccessible to VFs
1270 	 */
1271 	if (efx_sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1272 	    count > efx_vf_size(efx)) {
1273 		netif_warn(efx, probe, efx->net_dev,
1274 			   "Reducing number of RSS channels from %u to %u for "
1275 			   "VF support. Increase vf-msix-limit to use more "
1276 			   "channels on the PF.\n",
1277 			   count, efx_vf_size(efx));
1278 		count = efx_vf_size(efx);
1279 	}
1280 
1281 	return count;
1282 }
1283 
1284 static int
1285 efx_init_rx_cpu_rmap(struct efx_nic *efx, struct msix_entry *xentries)
1286 {
1287 #ifdef CONFIG_RFS_ACCEL
1288 	unsigned int i;
1289 	int rc;
1290 
1291 	efx->net_dev->rx_cpu_rmap = alloc_irq_cpu_rmap(efx->n_rx_channels);
1292 	if (!efx->net_dev->rx_cpu_rmap)
1293 		return -ENOMEM;
1294 	for (i = 0; i < efx->n_rx_channels; i++) {
1295 		rc = irq_cpu_rmap_add(efx->net_dev->rx_cpu_rmap,
1296 				      xentries[i].vector);
1297 		if (rc) {
1298 			free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1299 			efx->net_dev->rx_cpu_rmap = NULL;
1300 			return rc;
1301 		}
1302 	}
1303 #endif
1304 	return 0;
1305 }
1306 
1307 /* Probe the number and type of interrupts we are able to obtain, and
1308  * the resulting numbers of channels and RX queues.
1309  */
1310 static int efx_probe_interrupts(struct efx_nic *efx)
1311 {
1312 	unsigned int max_channels =
1313 		min(efx->type->phys_addr_channels, EFX_MAX_CHANNELS);
1314 	unsigned int extra_channels = 0;
1315 	unsigned int i, j;
1316 	int rc;
1317 
1318 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1319 		if (efx->extra_channel_type[i])
1320 			++extra_channels;
1321 
1322 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1323 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1324 		unsigned int n_channels;
1325 
1326 		n_channels = efx_wanted_parallelism(efx);
1327 		if (separate_tx_channels)
1328 			n_channels *= 2;
1329 		n_channels += extra_channels;
1330 		n_channels = min(n_channels, max_channels);
1331 
1332 		for (i = 0; i < n_channels; i++)
1333 			xentries[i].entry = i;
1334 		rc = pci_enable_msix(efx->pci_dev, xentries, n_channels);
1335 		if (rc > 0) {
1336 			netif_err(efx, drv, efx->net_dev,
1337 				  "WARNING: Insufficient MSI-X vectors"
1338 				  " available (%d < %u).\n", rc, n_channels);
1339 			netif_err(efx, drv, efx->net_dev,
1340 				  "WARNING: Performance may be reduced.\n");
1341 			EFX_BUG_ON_PARANOID(rc >= n_channels);
1342 			n_channels = rc;
1343 			rc = pci_enable_msix(efx->pci_dev, xentries,
1344 					     n_channels);
1345 		}
1346 
1347 		if (rc == 0) {
1348 			efx->n_channels = n_channels;
1349 			if (n_channels > extra_channels)
1350 				n_channels -= extra_channels;
1351 			if (separate_tx_channels) {
1352 				efx->n_tx_channels = max(n_channels / 2, 1U);
1353 				efx->n_rx_channels = max(n_channels -
1354 							 efx->n_tx_channels,
1355 							 1U);
1356 			} else {
1357 				efx->n_tx_channels = n_channels;
1358 				efx->n_rx_channels = n_channels;
1359 			}
1360 			rc = efx_init_rx_cpu_rmap(efx, xentries);
1361 			if (rc) {
1362 				pci_disable_msix(efx->pci_dev);
1363 				return rc;
1364 			}
1365 			for (i = 0; i < efx->n_channels; i++)
1366 				efx_get_channel(efx, i)->irq =
1367 					xentries[i].vector;
1368 		} else {
1369 			/* Fall back to single channel MSI */
1370 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1371 			netif_err(efx, drv, efx->net_dev,
1372 				  "could not enable MSI-X\n");
1373 		}
1374 	}
1375 
1376 	/* Try single interrupt MSI */
1377 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1378 		efx->n_channels = 1;
1379 		efx->n_rx_channels = 1;
1380 		efx->n_tx_channels = 1;
1381 		rc = pci_enable_msi(efx->pci_dev);
1382 		if (rc == 0) {
1383 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1384 		} else {
1385 			netif_err(efx, drv, efx->net_dev,
1386 				  "could not enable MSI\n");
1387 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1388 		}
1389 	}
1390 
1391 	/* Assume legacy interrupts */
1392 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1393 		efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1394 		efx->n_rx_channels = 1;
1395 		efx->n_tx_channels = 1;
1396 		efx->legacy_irq = efx->pci_dev->irq;
1397 	}
1398 
1399 	/* Assign extra channels if possible */
1400 	j = efx->n_channels;
1401 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1402 		if (!efx->extra_channel_type[i])
1403 			continue;
1404 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1405 		    efx->n_channels <= extra_channels) {
1406 			efx->extra_channel_type[i]->handle_no_channel(efx);
1407 		} else {
1408 			--j;
1409 			efx_get_channel(efx, j)->type =
1410 				efx->extra_channel_type[i];
1411 		}
1412 	}
1413 
1414 	/* RSS might be usable on VFs even if it is disabled on the PF */
1415 	efx->rss_spread = ((efx->n_rx_channels > 1 || !efx_sriov_wanted(efx)) ?
1416 			   efx->n_rx_channels : efx_vf_size(efx));
1417 
1418 	return 0;
1419 }
1420 
1421 /* Enable interrupts, then probe and start the event queues */
1422 static void efx_start_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1423 {
1424 	struct efx_channel *channel;
1425 
1426 	BUG_ON(efx->state == STATE_DISABLED);
1427 
1428 	if (efx->legacy_irq)
1429 		efx->legacy_irq_enabled = true;
1430 	efx_nic_enable_interrupts(efx);
1431 
1432 	efx_for_each_channel(channel, efx) {
1433 		if (!channel->type->keep_eventq || !may_keep_eventq)
1434 			efx_init_eventq(channel);
1435 		efx_start_eventq(channel);
1436 	}
1437 
1438 	efx_mcdi_mode_event(efx);
1439 }
1440 
1441 static void efx_stop_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1442 {
1443 	struct efx_channel *channel;
1444 
1445 	if (efx->state == STATE_DISABLED)
1446 		return;
1447 
1448 	efx_mcdi_mode_poll(efx);
1449 
1450 	efx_nic_disable_interrupts(efx);
1451 	if (efx->legacy_irq) {
1452 		synchronize_irq(efx->legacy_irq);
1453 		efx->legacy_irq_enabled = false;
1454 	}
1455 
1456 	efx_for_each_channel(channel, efx) {
1457 		if (channel->irq)
1458 			synchronize_irq(channel->irq);
1459 
1460 		efx_stop_eventq(channel);
1461 		if (!channel->type->keep_eventq || !may_keep_eventq)
1462 			efx_fini_eventq(channel);
1463 	}
1464 }
1465 
1466 static void efx_remove_interrupts(struct efx_nic *efx)
1467 {
1468 	struct efx_channel *channel;
1469 
1470 	/* Remove MSI/MSI-X interrupts */
1471 	efx_for_each_channel(channel, efx)
1472 		channel->irq = 0;
1473 	pci_disable_msi(efx->pci_dev);
1474 	pci_disable_msix(efx->pci_dev);
1475 
1476 	/* Remove legacy interrupt */
1477 	efx->legacy_irq = 0;
1478 }
1479 
1480 static void efx_set_channels(struct efx_nic *efx)
1481 {
1482 	struct efx_channel *channel;
1483 	struct efx_tx_queue *tx_queue;
1484 
1485 	efx->tx_channel_offset =
1486 		separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1487 
1488 	/* We need to mark which channels really have RX and TX
1489 	 * queues, and adjust the TX queue numbers if we have separate
1490 	 * RX-only and TX-only channels.
1491 	 */
1492 	efx_for_each_channel(channel, efx) {
1493 		if (channel->channel < efx->n_rx_channels)
1494 			channel->rx_queue.core_index = channel->channel;
1495 		else
1496 			channel->rx_queue.core_index = -1;
1497 
1498 		efx_for_each_channel_tx_queue(tx_queue, channel)
1499 			tx_queue->queue -= (efx->tx_channel_offset *
1500 					    EFX_TXQ_TYPES);
1501 	}
1502 }
1503 
1504 static int efx_probe_nic(struct efx_nic *efx)
1505 {
1506 	size_t i;
1507 	int rc;
1508 
1509 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1510 
1511 	/* Carry out hardware-type specific initialisation */
1512 	rc = efx->type->probe(efx);
1513 	if (rc)
1514 		return rc;
1515 
1516 	/* Determine the number of channels and queues by trying to hook
1517 	 * in MSI-X interrupts. */
1518 	rc = efx_probe_interrupts(efx);
1519 	if (rc)
1520 		goto fail;
1521 
1522 	efx->type->dimension_resources(efx);
1523 
1524 	if (efx->n_channels > 1)
1525 		get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1526 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1527 		efx->rx_indir_table[i] =
1528 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1529 
1530 	efx_set_channels(efx);
1531 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1532 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1533 
1534 	/* Initialise the interrupt moderation settings */
1535 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1536 				true);
1537 
1538 	return 0;
1539 
1540 fail:
1541 	efx->type->remove(efx);
1542 	return rc;
1543 }
1544 
1545 static void efx_remove_nic(struct efx_nic *efx)
1546 {
1547 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1548 
1549 	efx_remove_interrupts(efx);
1550 	efx->type->remove(efx);
1551 }
1552 
1553 /**************************************************************************
1554  *
1555  * NIC startup/shutdown
1556  *
1557  *************************************************************************/
1558 
1559 static int efx_probe_all(struct efx_nic *efx)
1560 {
1561 	int rc;
1562 
1563 	rc = efx_probe_nic(efx);
1564 	if (rc) {
1565 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1566 		goto fail1;
1567 	}
1568 
1569 	rc = efx_probe_port(efx);
1570 	if (rc) {
1571 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1572 		goto fail2;
1573 	}
1574 
1575 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1576 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1577 		rc = -EINVAL;
1578 		goto fail3;
1579 	}
1580 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1581 
1582 	rc = efx_probe_filters(efx);
1583 	if (rc) {
1584 		netif_err(efx, probe, efx->net_dev,
1585 			  "failed to create filter tables\n");
1586 		goto fail3;
1587 	}
1588 
1589 	rc = efx_probe_channels(efx);
1590 	if (rc)
1591 		goto fail4;
1592 
1593 	return 0;
1594 
1595  fail4:
1596 	efx_remove_filters(efx);
1597  fail3:
1598 	efx_remove_port(efx);
1599  fail2:
1600 	efx_remove_nic(efx);
1601  fail1:
1602 	return rc;
1603 }
1604 
1605 /* If the interface is supposed to be running but is not, start
1606  * the hardware and software data path, regular activity for the port
1607  * (MAC statistics, link polling, etc.) and schedule the port to be
1608  * reconfigured.  Interrupts must already be enabled.  This function
1609  * is safe to call multiple times, so long as the NIC is not disabled.
1610  * Requires the RTNL lock.
1611  */
1612 static void efx_start_all(struct efx_nic *efx)
1613 {
1614 	EFX_ASSERT_RESET_SERIALISED(efx);
1615 	BUG_ON(efx->state == STATE_DISABLED);
1616 
1617 	/* Check that it is appropriate to restart the interface. All
1618 	 * of these flags are safe to read under just the rtnl lock */
1619 	if (efx->port_enabled || !netif_running(efx->net_dev))
1620 		return;
1621 
1622 	efx_start_port(efx);
1623 	efx_start_datapath(efx);
1624 
1625 	/* Start the hardware monitor if there is one */
1626 	if (efx->type->monitor != NULL)
1627 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1628 				   efx_monitor_interval);
1629 
1630 	/* If link state detection is normally event-driven, we have
1631 	 * to poll now because we could have missed a change
1632 	 */
1633 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1634 		mutex_lock(&efx->mac_lock);
1635 		if (efx->phy_op->poll(efx))
1636 			efx_link_status_changed(efx);
1637 		mutex_unlock(&efx->mac_lock);
1638 	}
1639 
1640 	efx->type->start_stats(efx);
1641 }
1642 
1643 /* Flush all delayed work. Should only be called when no more delayed work
1644  * will be scheduled. This doesn't flush pending online resets (efx_reset),
1645  * since we're holding the rtnl_lock at this point. */
1646 static void efx_flush_all(struct efx_nic *efx)
1647 {
1648 	/* Make sure the hardware monitor and event self-test are stopped */
1649 	cancel_delayed_work_sync(&efx->monitor_work);
1650 	efx_selftest_async_cancel(efx);
1651 	/* Stop scheduled port reconfigurations */
1652 	cancel_work_sync(&efx->mac_work);
1653 }
1654 
1655 /* Quiesce the hardware and software data path, and regular activity
1656  * for the port without bringing the link down.  Safe to call multiple
1657  * times with the NIC in almost any state, but interrupts should be
1658  * enabled.  Requires the RTNL lock.
1659  */
1660 static void efx_stop_all(struct efx_nic *efx)
1661 {
1662 	EFX_ASSERT_RESET_SERIALISED(efx);
1663 
1664 	/* port_enabled can be read safely under the rtnl lock */
1665 	if (!efx->port_enabled)
1666 		return;
1667 
1668 	efx->type->stop_stats(efx);
1669 	efx_stop_port(efx);
1670 
1671 	/* Flush efx_mac_work(), refill_workqueue, monitor_work */
1672 	efx_flush_all(efx);
1673 
1674 	/* Stop the kernel transmit interface.  This is only valid if
1675 	 * the device is stopped or detached; otherwise the watchdog
1676 	 * may fire immediately.
1677 	 */
1678 	WARN_ON(netif_running(efx->net_dev) &&
1679 		netif_device_present(efx->net_dev));
1680 	netif_tx_disable(efx->net_dev);
1681 
1682 	efx_stop_datapath(efx);
1683 }
1684 
1685 static void efx_remove_all(struct efx_nic *efx)
1686 {
1687 	efx_remove_channels(efx);
1688 	efx_remove_filters(efx);
1689 	efx_remove_port(efx);
1690 	efx_remove_nic(efx);
1691 }
1692 
1693 /**************************************************************************
1694  *
1695  * Interrupt moderation
1696  *
1697  **************************************************************************/
1698 
1699 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1700 {
1701 	if (usecs == 0)
1702 		return 0;
1703 	if (usecs * 1000 < quantum_ns)
1704 		return 1; /* never round down to 0 */
1705 	return usecs * 1000 / quantum_ns;
1706 }
1707 
1708 /* Set interrupt moderation parameters */
1709 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1710 			    unsigned int rx_usecs, bool rx_adaptive,
1711 			    bool rx_may_override_tx)
1712 {
1713 	struct efx_channel *channel;
1714 	unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1715 						efx->timer_quantum_ns,
1716 						1000);
1717 	unsigned int tx_ticks;
1718 	unsigned int rx_ticks;
1719 
1720 	EFX_ASSERT_RESET_SERIALISED(efx);
1721 
1722 	if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1723 		return -EINVAL;
1724 
1725 	tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1726 	rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1727 
1728 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1729 	    !rx_may_override_tx) {
1730 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1731 			  "RX and TX IRQ moderation must be equal\n");
1732 		return -EINVAL;
1733 	}
1734 
1735 	efx->irq_rx_adaptive = rx_adaptive;
1736 	efx->irq_rx_moderation = rx_ticks;
1737 	efx_for_each_channel(channel, efx) {
1738 		if (efx_channel_has_rx_queue(channel))
1739 			channel->irq_moderation = rx_ticks;
1740 		else if (efx_channel_has_tx_queues(channel))
1741 			channel->irq_moderation = tx_ticks;
1742 	}
1743 
1744 	return 0;
1745 }
1746 
1747 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1748 			    unsigned int *rx_usecs, bool *rx_adaptive)
1749 {
1750 	/* We must round up when converting ticks to microseconds
1751 	 * because we round down when converting the other way.
1752 	 */
1753 
1754 	*rx_adaptive = efx->irq_rx_adaptive;
1755 	*rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1756 				 efx->timer_quantum_ns,
1757 				 1000);
1758 
1759 	/* If channels are shared between RX and TX, so is IRQ
1760 	 * moderation.  Otherwise, IRQ moderation is the same for all
1761 	 * TX channels and is not adaptive.
1762 	 */
1763 	if (efx->tx_channel_offset == 0)
1764 		*tx_usecs = *rx_usecs;
1765 	else
1766 		*tx_usecs = DIV_ROUND_UP(
1767 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1768 			efx->timer_quantum_ns,
1769 			1000);
1770 }
1771 
1772 /**************************************************************************
1773  *
1774  * Hardware monitor
1775  *
1776  **************************************************************************/
1777 
1778 /* Run periodically off the general workqueue */
1779 static void efx_monitor(struct work_struct *data)
1780 {
1781 	struct efx_nic *efx = container_of(data, struct efx_nic,
1782 					   monitor_work.work);
1783 
1784 	netif_vdbg(efx, timer, efx->net_dev,
1785 		   "hardware monitor executing on CPU %d\n",
1786 		   raw_smp_processor_id());
1787 	BUG_ON(efx->type->monitor == NULL);
1788 
1789 	/* If the mac_lock is already held then it is likely a port
1790 	 * reconfiguration is already in place, which will likely do
1791 	 * most of the work of monitor() anyway. */
1792 	if (mutex_trylock(&efx->mac_lock)) {
1793 		if (efx->port_enabled)
1794 			efx->type->monitor(efx);
1795 		mutex_unlock(&efx->mac_lock);
1796 	}
1797 
1798 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1799 			   efx_monitor_interval);
1800 }
1801 
1802 /**************************************************************************
1803  *
1804  * ioctls
1805  *
1806  *************************************************************************/
1807 
1808 /* Net device ioctl
1809  * Context: process, rtnl_lock() held.
1810  */
1811 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1812 {
1813 	struct efx_nic *efx = netdev_priv(net_dev);
1814 	struct mii_ioctl_data *data = if_mii(ifr);
1815 
1816 	if (cmd == SIOCSHWTSTAMP)
1817 		return efx_ptp_ioctl(efx, ifr, cmd);
1818 
1819 	/* Convert phy_id from older PRTAD/DEVAD format */
1820 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1821 	    (data->phy_id & 0xfc00) == 0x0400)
1822 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1823 
1824 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
1825 }
1826 
1827 /**************************************************************************
1828  *
1829  * NAPI interface
1830  *
1831  **************************************************************************/
1832 
1833 static void efx_init_napi_channel(struct efx_channel *channel)
1834 {
1835 	struct efx_nic *efx = channel->efx;
1836 
1837 	channel->napi_dev = efx->net_dev;
1838 	netif_napi_add(channel->napi_dev, &channel->napi_str,
1839 		       efx_poll, napi_weight);
1840 }
1841 
1842 static void efx_init_napi(struct efx_nic *efx)
1843 {
1844 	struct efx_channel *channel;
1845 
1846 	efx_for_each_channel(channel, efx)
1847 		efx_init_napi_channel(channel);
1848 }
1849 
1850 static void efx_fini_napi_channel(struct efx_channel *channel)
1851 {
1852 	if (channel->napi_dev)
1853 		netif_napi_del(&channel->napi_str);
1854 	channel->napi_dev = NULL;
1855 }
1856 
1857 static void efx_fini_napi(struct efx_nic *efx)
1858 {
1859 	struct efx_channel *channel;
1860 
1861 	efx_for_each_channel(channel, efx)
1862 		efx_fini_napi_channel(channel);
1863 }
1864 
1865 /**************************************************************************
1866  *
1867  * Kernel netpoll interface
1868  *
1869  *************************************************************************/
1870 
1871 #ifdef CONFIG_NET_POLL_CONTROLLER
1872 
1873 /* Although in the common case interrupts will be disabled, this is not
1874  * guaranteed. However, all our work happens inside the NAPI callback,
1875  * so no locking is required.
1876  */
1877 static void efx_netpoll(struct net_device *net_dev)
1878 {
1879 	struct efx_nic *efx = netdev_priv(net_dev);
1880 	struct efx_channel *channel;
1881 
1882 	efx_for_each_channel(channel, efx)
1883 		efx_schedule_channel(channel);
1884 }
1885 
1886 #endif
1887 
1888 /**************************************************************************
1889  *
1890  * Kernel net device interface
1891  *
1892  *************************************************************************/
1893 
1894 /* Context: process, rtnl_lock() held. */
1895 static int efx_net_open(struct net_device *net_dev)
1896 {
1897 	struct efx_nic *efx = netdev_priv(net_dev);
1898 	int rc;
1899 
1900 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
1901 		  raw_smp_processor_id());
1902 
1903 	rc = efx_check_disabled(efx);
1904 	if (rc)
1905 		return rc;
1906 	if (efx->phy_mode & PHY_MODE_SPECIAL)
1907 		return -EBUSY;
1908 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
1909 		return -EIO;
1910 
1911 	/* Notify the kernel of the link state polled during driver load,
1912 	 * before the monitor starts running */
1913 	efx_link_status_changed(efx);
1914 
1915 	efx_start_all(efx);
1916 	efx_selftest_async_start(efx);
1917 	return 0;
1918 }
1919 
1920 /* Context: process, rtnl_lock() held.
1921  * Note that the kernel will ignore our return code; this method
1922  * should really be a void.
1923  */
1924 static int efx_net_stop(struct net_device *net_dev)
1925 {
1926 	struct efx_nic *efx = netdev_priv(net_dev);
1927 
1928 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
1929 		  raw_smp_processor_id());
1930 
1931 	/* Stop the device and flush all the channels */
1932 	efx_stop_all(efx);
1933 
1934 	return 0;
1935 }
1936 
1937 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1938 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
1939 					       struct rtnl_link_stats64 *stats)
1940 {
1941 	struct efx_nic *efx = netdev_priv(net_dev);
1942 	struct efx_mac_stats *mac_stats = &efx->mac_stats;
1943 
1944 	spin_lock_bh(&efx->stats_lock);
1945 
1946 	efx->type->update_stats(efx);
1947 
1948 	stats->rx_packets = mac_stats->rx_packets;
1949 	stats->tx_packets = mac_stats->tx_packets;
1950 	stats->rx_bytes = mac_stats->rx_bytes;
1951 	stats->tx_bytes = mac_stats->tx_bytes;
1952 	stats->rx_dropped = efx->n_rx_nodesc_drop_cnt;
1953 	stats->multicast = mac_stats->rx_multicast;
1954 	stats->collisions = mac_stats->tx_collision;
1955 	stats->rx_length_errors = (mac_stats->rx_gtjumbo +
1956 				   mac_stats->rx_length_error);
1957 	stats->rx_crc_errors = mac_stats->rx_bad;
1958 	stats->rx_frame_errors = mac_stats->rx_align_error;
1959 	stats->rx_fifo_errors = mac_stats->rx_overflow;
1960 	stats->rx_missed_errors = mac_stats->rx_missed;
1961 	stats->tx_window_errors = mac_stats->tx_late_collision;
1962 
1963 	stats->rx_errors = (stats->rx_length_errors +
1964 			    stats->rx_crc_errors +
1965 			    stats->rx_frame_errors +
1966 			    mac_stats->rx_symbol_error);
1967 	stats->tx_errors = (stats->tx_window_errors +
1968 			    mac_stats->tx_bad);
1969 
1970 	spin_unlock_bh(&efx->stats_lock);
1971 
1972 	return stats;
1973 }
1974 
1975 /* Context: netif_tx_lock held, BHs disabled. */
1976 static void efx_watchdog(struct net_device *net_dev)
1977 {
1978 	struct efx_nic *efx = netdev_priv(net_dev);
1979 
1980 	netif_err(efx, tx_err, efx->net_dev,
1981 		  "TX stuck with port_enabled=%d: resetting channels\n",
1982 		  efx->port_enabled);
1983 
1984 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
1985 }
1986 
1987 
1988 /* Context: process, rtnl_lock() held. */
1989 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
1990 {
1991 	struct efx_nic *efx = netdev_priv(net_dev);
1992 	int rc;
1993 
1994 	rc = efx_check_disabled(efx);
1995 	if (rc)
1996 		return rc;
1997 	if (new_mtu > EFX_MAX_MTU)
1998 		return -EINVAL;
1999 
2000 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2001 
2002 	efx_device_detach_sync(efx);
2003 	efx_stop_all(efx);
2004 
2005 	mutex_lock(&efx->mac_lock);
2006 	net_dev->mtu = new_mtu;
2007 	efx->type->reconfigure_mac(efx);
2008 	mutex_unlock(&efx->mac_lock);
2009 
2010 	efx_start_all(efx);
2011 	netif_device_attach(efx->net_dev);
2012 	return 0;
2013 }
2014 
2015 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2016 {
2017 	struct efx_nic *efx = netdev_priv(net_dev);
2018 	struct sockaddr *addr = data;
2019 	char *new_addr = addr->sa_data;
2020 
2021 	if (!is_valid_ether_addr(new_addr)) {
2022 		netif_err(efx, drv, efx->net_dev,
2023 			  "invalid ethernet MAC address requested: %pM\n",
2024 			  new_addr);
2025 		return -EADDRNOTAVAIL;
2026 	}
2027 
2028 	memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
2029 	efx_sriov_mac_address_changed(efx);
2030 
2031 	/* Reconfigure the MAC */
2032 	mutex_lock(&efx->mac_lock);
2033 	efx->type->reconfigure_mac(efx);
2034 	mutex_unlock(&efx->mac_lock);
2035 
2036 	return 0;
2037 }
2038 
2039 /* Context: netif_addr_lock held, BHs disabled. */
2040 static void efx_set_rx_mode(struct net_device *net_dev)
2041 {
2042 	struct efx_nic *efx = netdev_priv(net_dev);
2043 	struct netdev_hw_addr *ha;
2044 	union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2045 	u32 crc;
2046 	int bit;
2047 
2048 	efx->promiscuous = !!(net_dev->flags & IFF_PROMISC);
2049 
2050 	/* Build multicast hash table */
2051 	if (efx->promiscuous || (net_dev->flags & IFF_ALLMULTI)) {
2052 		memset(mc_hash, 0xff, sizeof(*mc_hash));
2053 	} else {
2054 		memset(mc_hash, 0x00, sizeof(*mc_hash));
2055 		netdev_for_each_mc_addr(ha, net_dev) {
2056 			crc = ether_crc_le(ETH_ALEN, ha->addr);
2057 			bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2058 			__set_bit_le(bit, mc_hash);
2059 		}
2060 
2061 		/* Broadcast packets go through the multicast hash filter.
2062 		 * ether_crc_le() of the broadcast address is 0xbe2612ff
2063 		 * so we always add bit 0xff to the mask.
2064 		 */
2065 		__set_bit_le(0xff, mc_hash);
2066 	}
2067 
2068 	if (efx->port_enabled)
2069 		queue_work(efx->workqueue, &efx->mac_work);
2070 	/* Otherwise efx_start_port() will do this */
2071 }
2072 
2073 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2074 {
2075 	struct efx_nic *efx = netdev_priv(net_dev);
2076 
2077 	/* If disabling RX n-tuple filtering, clear existing filters */
2078 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
2079 		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2080 
2081 	return 0;
2082 }
2083 
2084 static const struct net_device_ops efx_netdev_ops = {
2085 	.ndo_open		= efx_net_open,
2086 	.ndo_stop		= efx_net_stop,
2087 	.ndo_get_stats64	= efx_net_stats,
2088 	.ndo_tx_timeout		= efx_watchdog,
2089 	.ndo_start_xmit		= efx_hard_start_xmit,
2090 	.ndo_validate_addr	= eth_validate_addr,
2091 	.ndo_do_ioctl		= efx_ioctl,
2092 	.ndo_change_mtu		= efx_change_mtu,
2093 	.ndo_set_mac_address	= efx_set_mac_address,
2094 	.ndo_set_rx_mode	= efx_set_rx_mode,
2095 	.ndo_set_features	= efx_set_features,
2096 #ifdef CONFIG_SFC_SRIOV
2097 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2098 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2099 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2100 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2101 #endif
2102 #ifdef CONFIG_NET_POLL_CONTROLLER
2103 	.ndo_poll_controller = efx_netpoll,
2104 #endif
2105 	.ndo_setup_tc		= efx_setup_tc,
2106 #ifdef CONFIG_RFS_ACCEL
2107 	.ndo_rx_flow_steer	= efx_filter_rfs,
2108 #endif
2109 };
2110 
2111 static void efx_update_name(struct efx_nic *efx)
2112 {
2113 	strcpy(efx->name, efx->net_dev->name);
2114 	efx_mtd_rename(efx);
2115 	efx_set_channel_names(efx);
2116 }
2117 
2118 static int efx_netdev_event(struct notifier_block *this,
2119 			    unsigned long event, void *ptr)
2120 {
2121 	struct net_device *net_dev = ptr;
2122 
2123 	if (net_dev->netdev_ops == &efx_netdev_ops &&
2124 	    event == NETDEV_CHANGENAME)
2125 		efx_update_name(netdev_priv(net_dev));
2126 
2127 	return NOTIFY_DONE;
2128 }
2129 
2130 static struct notifier_block efx_netdev_notifier = {
2131 	.notifier_call = efx_netdev_event,
2132 };
2133 
2134 static ssize_t
2135 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2136 {
2137 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2138 	return sprintf(buf, "%d\n", efx->phy_type);
2139 }
2140 static DEVICE_ATTR(phy_type, 0644, show_phy_type, NULL);
2141 
2142 static int efx_register_netdev(struct efx_nic *efx)
2143 {
2144 	struct net_device *net_dev = efx->net_dev;
2145 	struct efx_channel *channel;
2146 	int rc;
2147 
2148 	net_dev->watchdog_timeo = 5 * HZ;
2149 	net_dev->irq = efx->pci_dev->irq;
2150 	net_dev->netdev_ops = &efx_netdev_ops;
2151 	SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
2152 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2153 
2154 	rtnl_lock();
2155 
2156 	/* Enable resets to be scheduled and check whether any were
2157 	 * already requested.  If so, the NIC is probably hosed so we
2158 	 * abort.
2159 	 */
2160 	efx->state = STATE_READY;
2161 	smp_mb(); /* ensure we change state before checking reset_pending */
2162 	if (efx->reset_pending) {
2163 		netif_err(efx, probe, efx->net_dev,
2164 			  "aborting probe due to scheduled reset\n");
2165 		rc = -EIO;
2166 		goto fail_locked;
2167 	}
2168 
2169 	rc = dev_alloc_name(net_dev, net_dev->name);
2170 	if (rc < 0)
2171 		goto fail_locked;
2172 	efx_update_name(efx);
2173 
2174 	/* Always start with carrier off; PHY events will detect the link */
2175 	netif_carrier_off(net_dev);
2176 
2177 	rc = register_netdevice(net_dev);
2178 	if (rc)
2179 		goto fail_locked;
2180 
2181 	efx_for_each_channel(channel, efx) {
2182 		struct efx_tx_queue *tx_queue;
2183 		efx_for_each_channel_tx_queue(tx_queue, channel)
2184 			efx_init_tx_queue_core_txq(tx_queue);
2185 	}
2186 
2187 	rtnl_unlock();
2188 
2189 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2190 	if (rc) {
2191 		netif_err(efx, drv, efx->net_dev,
2192 			  "failed to init net dev attributes\n");
2193 		goto fail_registered;
2194 	}
2195 
2196 	return 0;
2197 
2198 fail_registered:
2199 	rtnl_lock();
2200 	unregister_netdevice(net_dev);
2201 fail_locked:
2202 	efx->state = STATE_UNINIT;
2203 	rtnl_unlock();
2204 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2205 	return rc;
2206 }
2207 
2208 static void efx_unregister_netdev(struct efx_nic *efx)
2209 {
2210 	struct efx_channel *channel;
2211 	struct efx_tx_queue *tx_queue;
2212 
2213 	if (!efx->net_dev)
2214 		return;
2215 
2216 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2217 
2218 	/* Free up any skbs still remaining. This has to happen before
2219 	 * we try to unregister the netdev as running their destructors
2220 	 * may be needed to get the device ref. count to 0. */
2221 	efx_for_each_channel(channel, efx) {
2222 		efx_for_each_channel_tx_queue(tx_queue, channel)
2223 			efx_release_tx_buffers(tx_queue);
2224 	}
2225 
2226 	strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2227 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2228 
2229 	rtnl_lock();
2230 	unregister_netdevice(efx->net_dev);
2231 	efx->state = STATE_UNINIT;
2232 	rtnl_unlock();
2233 }
2234 
2235 /**************************************************************************
2236  *
2237  * Device reset and suspend
2238  *
2239  **************************************************************************/
2240 
2241 /* Tears down the entire software state and most of the hardware state
2242  * before reset.  */
2243 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2244 {
2245 	EFX_ASSERT_RESET_SERIALISED(efx);
2246 
2247 	efx_stop_all(efx);
2248 	efx_stop_interrupts(efx, false);
2249 
2250 	mutex_lock(&efx->mac_lock);
2251 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2252 		efx->phy_op->fini(efx);
2253 	efx->type->fini(efx);
2254 }
2255 
2256 /* This function will always ensure that the locks acquired in
2257  * efx_reset_down() are released. A failure return code indicates
2258  * that we were unable to reinitialise the hardware, and the
2259  * driver should be disabled. If ok is false, then the rx and tx
2260  * engines are not restarted, pending a RESET_DISABLE. */
2261 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2262 {
2263 	int rc;
2264 
2265 	EFX_ASSERT_RESET_SERIALISED(efx);
2266 
2267 	rc = efx->type->init(efx);
2268 	if (rc) {
2269 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2270 		goto fail;
2271 	}
2272 
2273 	if (!ok)
2274 		goto fail;
2275 
2276 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2277 		rc = efx->phy_op->init(efx);
2278 		if (rc)
2279 			goto fail;
2280 		if (efx->phy_op->reconfigure(efx))
2281 			netif_err(efx, drv, efx->net_dev,
2282 				  "could not restore PHY settings\n");
2283 	}
2284 
2285 	efx->type->reconfigure_mac(efx);
2286 
2287 	efx_start_interrupts(efx, false);
2288 	efx_restore_filters(efx);
2289 	efx_sriov_reset(efx);
2290 
2291 	mutex_unlock(&efx->mac_lock);
2292 
2293 	efx_start_all(efx);
2294 
2295 	return 0;
2296 
2297 fail:
2298 	efx->port_initialized = false;
2299 
2300 	mutex_unlock(&efx->mac_lock);
2301 
2302 	return rc;
2303 }
2304 
2305 /* Reset the NIC using the specified method.  Note that the reset may
2306  * fail, in which case the card will be left in an unusable state.
2307  *
2308  * Caller must hold the rtnl_lock.
2309  */
2310 int efx_reset(struct efx_nic *efx, enum reset_type method)
2311 {
2312 	int rc, rc2;
2313 	bool disabled;
2314 
2315 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2316 		   RESET_TYPE(method));
2317 
2318 	efx_device_detach_sync(efx);
2319 	efx_reset_down(efx, method);
2320 
2321 	rc = efx->type->reset(efx, method);
2322 	if (rc) {
2323 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2324 		goto out;
2325 	}
2326 
2327 	/* Clear flags for the scopes we covered.  We assume the NIC and
2328 	 * driver are now quiescent so that there is no race here.
2329 	 */
2330 	efx->reset_pending &= -(1 << (method + 1));
2331 
2332 	/* Reinitialise bus-mastering, which may have been turned off before
2333 	 * the reset was scheduled. This is still appropriate, even in the
2334 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2335 	 * can respond to requests. */
2336 	pci_set_master(efx->pci_dev);
2337 
2338 out:
2339 	/* Leave device stopped if necessary */
2340 	disabled = rc ||
2341 		method == RESET_TYPE_DISABLE ||
2342 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2343 	rc2 = efx_reset_up(efx, method, !disabled);
2344 	if (rc2) {
2345 		disabled = true;
2346 		if (!rc)
2347 			rc = rc2;
2348 	}
2349 
2350 	if (disabled) {
2351 		dev_close(efx->net_dev);
2352 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2353 		efx->state = STATE_DISABLED;
2354 	} else {
2355 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2356 		netif_device_attach(efx->net_dev);
2357 	}
2358 	return rc;
2359 }
2360 
2361 /* Try recovery mechanisms.
2362  * For now only EEH is supported.
2363  * Returns 0 if the recovery mechanisms are unsuccessful.
2364  * Returns a non-zero value otherwise.
2365  */
2366 static int efx_try_recovery(struct efx_nic *efx)
2367 {
2368 #ifdef CONFIG_EEH
2369 	/* A PCI error can occur and not be seen by EEH because nothing
2370 	 * happens on the PCI bus. In this case the driver may fail and
2371 	 * schedule a 'recover or reset', leading to this recovery handler.
2372 	 * Manually call the eeh failure check function.
2373 	 */
2374 	struct eeh_dev *eehdev =
2375 		of_node_to_eeh_dev(pci_device_to_OF_node(efx->pci_dev));
2376 
2377 	if (eeh_dev_check_failure(eehdev)) {
2378 		/* The EEH mechanisms will handle the error and reset the
2379 		 * device if necessary.
2380 		 */
2381 		return 1;
2382 	}
2383 #endif
2384 	return 0;
2385 }
2386 
2387 /* The worker thread exists so that code that cannot sleep can
2388  * schedule a reset for later.
2389  */
2390 static void efx_reset_work(struct work_struct *data)
2391 {
2392 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2393 	unsigned long pending;
2394 	enum reset_type method;
2395 
2396 	pending = ACCESS_ONCE(efx->reset_pending);
2397 	method = fls(pending) - 1;
2398 
2399 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2400 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2401 	    efx_try_recovery(efx))
2402 		return;
2403 
2404 	if (!pending)
2405 		return;
2406 
2407 	rtnl_lock();
2408 
2409 	/* We checked the state in efx_schedule_reset() but it may
2410 	 * have changed by now.  Now that we have the RTNL lock,
2411 	 * it cannot change again.
2412 	 */
2413 	if (efx->state == STATE_READY)
2414 		(void)efx_reset(efx, method);
2415 
2416 	rtnl_unlock();
2417 }
2418 
2419 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2420 {
2421 	enum reset_type method;
2422 
2423 	if (efx->state == STATE_RECOVERY) {
2424 		netif_dbg(efx, drv, efx->net_dev,
2425 			  "recovering: skip scheduling %s reset\n",
2426 			  RESET_TYPE(type));
2427 		return;
2428 	}
2429 
2430 	switch (type) {
2431 	case RESET_TYPE_INVISIBLE:
2432 	case RESET_TYPE_ALL:
2433 	case RESET_TYPE_RECOVER_OR_ALL:
2434 	case RESET_TYPE_WORLD:
2435 	case RESET_TYPE_DISABLE:
2436 	case RESET_TYPE_RECOVER_OR_DISABLE:
2437 		method = type;
2438 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2439 			  RESET_TYPE(method));
2440 		break;
2441 	default:
2442 		method = efx->type->map_reset_reason(type);
2443 		netif_dbg(efx, drv, efx->net_dev,
2444 			  "scheduling %s reset for %s\n",
2445 			  RESET_TYPE(method), RESET_TYPE(type));
2446 		break;
2447 	}
2448 
2449 	set_bit(method, &efx->reset_pending);
2450 	smp_mb(); /* ensure we change reset_pending before checking state */
2451 
2452 	/* If we're not READY then just leave the flags set as the cue
2453 	 * to abort probing or reschedule the reset later.
2454 	 */
2455 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2456 		return;
2457 
2458 	/* efx_process_channel() will no longer read events once a
2459 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2460 	efx_mcdi_mode_poll(efx);
2461 
2462 	queue_work(reset_workqueue, &efx->reset_work);
2463 }
2464 
2465 /**************************************************************************
2466  *
2467  * List of NICs we support
2468  *
2469  **************************************************************************/
2470 
2471 /* PCI device ID table */
2472 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2473 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2474 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2475 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2476 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2477 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2478 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2479 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2480 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2481 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2482 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2483 	{0}			/* end of list */
2484 };
2485 
2486 /**************************************************************************
2487  *
2488  * Dummy PHY/MAC operations
2489  *
2490  * Can be used for some unimplemented operations
2491  * Needed so all function pointers are valid and do not have to be tested
2492  * before use
2493  *
2494  **************************************************************************/
2495 int efx_port_dummy_op_int(struct efx_nic *efx)
2496 {
2497 	return 0;
2498 }
2499 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2500 
2501 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2502 {
2503 	return false;
2504 }
2505 
2506 static const struct efx_phy_operations efx_dummy_phy_operations = {
2507 	.init		 = efx_port_dummy_op_int,
2508 	.reconfigure	 = efx_port_dummy_op_int,
2509 	.poll		 = efx_port_dummy_op_poll,
2510 	.fini		 = efx_port_dummy_op_void,
2511 };
2512 
2513 /**************************************************************************
2514  *
2515  * Data housekeeping
2516  *
2517  **************************************************************************/
2518 
2519 /* This zeroes out and then fills in the invariants in a struct
2520  * efx_nic (including all sub-structures).
2521  */
2522 static int efx_init_struct(struct efx_nic *efx,
2523 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2524 {
2525 	int i;
2526 
2527 	/* Initialise common structures */
2528 	spin_lock_init(&efx->biu_lock);
2529 #ifdef CONFIG_SFC_MTD
2530 	INIT_LIST_HEAD(&efx->mtd_list);
2531 #endif
2532 	INIT_WORK(&efx->reset_work, efx_reset_work);
2533 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2534 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2535 	efx->pci_dev = pci_dev;
2536 	efx->msg_enable = debug;
2537 	efx->state = STATE_UNINIT;
2538 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2539 
2540 	efx->net_dev = net_dev;
2541 	spin_lock_init(&efx->stats_lock);
2542 	mutex_init(&efx->mac_lock);
2543 	efx->phy_op = &efx_dummy_phy_operations;
2544 	efx->mdio.dev = net_dev;
2545 	INIT_WORK(&efx->mac_work, efx_mac_work);
2546 	init_waitqueue_head(&efx->flush_wq);
2547 
2548 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2549 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2550 		if (!efx->channel[i])
2551 			goto fail;
2552 	}
2553 
2554 	EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS);
2555 
2556 	/* Higher numbered interrupt modes are less capable! */
2557 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2558 				  interrupt_mode);
2559 
2560 	/* Would be good to use the net_dev name, but we're too early */
2561 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2562 		 pci_name(pci_dev));
2563 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2564 	if (!efx->workqueue)
2565 		goto fail;
2566 
2567 	return 0;
2568 
2569 fail:
2570 	efx_fini_struct(efx);
2571 	return -ENOMEM;
2572 }
2573 
2574 static void efx_fini_struct(struct efx_nic *efx)
2575 {
2576 	int i;
2577 
2578 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2579 		kfree(efx->channel[i]);
2580 
2581 	if (efx->workqueue) {
2582 		destroy_workqueue(efx->workqueue);
2583 		efx->workqueue = NULL;
2584 	}
2585 }
2586 
2587 /**************************************************************************
2588  *
2589  * PCI interface
2590  *
2591  **************************************************************************/
2592 
2593 /* Main body of final NIC shutdown code
2594  * This is called only at module unload (or hotplug removal).
2595  */
2596 static void efx_pci_remove_main(struct efx_nic *efx)
2597 {
2598 	/* Flush reset_work. It can no longer be scheduled since we
2599 	 * are not READY.
2600 	 */
2601 	BUG_ON(efx->state == STATE_READY);
2602 	cancel_work_sync(&efx->reset_work);
2603 
2604 #ifdef CONFIG_RFS_ACCEL
2605 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
2606 	efx->net_dev->rx_cpu_rmap = NULL;
2607 #endif
2608 	efx_stop_interrupts(efx, false);
2609 	efx_nic_fini_interrupt(efx);
2610 	efx_fini_port(efx);
2611 	efx->type->fini(efx);
2612 	efx_fini_napi(efx);
2613 	efx_remove_all(efx);
2614 }
2615 
2616 /* Final NIC shutdown
2617  * This is called only at module unload (or hotplug removal).
2618  */
2619 static void efx_pci_remove(struct pci_dev *pci_dev)
2620 {
2621 	struct efx_nic *efx;
2622 
2623 	efx = pci_get_drvdata(pci_dev);
2624 	if (!efx)
2625 		return;
2626 
2627 	/* Mark the NIC as fini, then stop the interface */
2628 	rtnl_lock();
2629 	dev_close(efx->net_dev);
2630 	efx_stop_interrupts(efx, false);
2631 	rtnl_unlock();
2632 
2633 	efx_sriov_fini(efx);
2634 	efx_unregister_netdev(efx);
2635 
2636 	efx_mtd_remove(efx);
2637 
2638 	efx_pci_remove_main(efx);
2639 
2640 	efx_fini_io(efx);
2641 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2642 
2643 	efx_fini_struct(efx);
2644 	pci_set_drvdata(pci_dev, NULL);
2645 	free_netdev(efx->net_dev);
2646 
2647 	pci_disable_pcie_error_reporting(pci_dev);
2648 };
2649 
2650 /* NIC VPD information
2651  * Called during probe to display the part number of the
2652  * installed NIC.  VPD is potentially very large but this should
2653  * always appear within the first 512 bytes.
2654  */
2655 #define SFC_VPD_LEN 512
2656 static void efx_print_product_vpd(struct efx_nic *efx)
2657 {
2658 	struct pci_dev *dev = efx->pci_dev;
2659 	char vpd_data[SFC_VPD_LEN];
2660 	ssize_t vpd_size;
2661 	int i, j;
2662 
2663 	/* Get the vpd data from the device */
2664 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2665 	if (vpd_size <= 0) {
2666 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2667 		return;
2668 	}
2669 
2670 	/* Get the Read only section */
2671 	i = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2672 	if (i < 0) {
2673 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2674 		return;
2675 	}
2676 
2677 	j = pci_vpd_lrdt_size(&vpd_data[i]);
2678 	i += PCI_VPD_LRDT_TAG_SIZE;
2679 	if (i + j > vpd_size)
2680 		j = vpd_size - i;
2681 
2682 	/* Get the Part number */
2683 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2684 	if (i < 0) {
2685 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2686 		return;
2687 	}
2688 
2689 	j = pci_vpd_info_field_size(&vpd_data[i]);
2690 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
2691 	if (i + j > vpd_size) {
2692 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2693 		return;
2694 	}
2695 
2696 	netif_info(efx, drv, efx->net_dev,
2697 		   "Part Number : %.*s\n", j, &vpd_data[i]);
2698 }
2699 
2700 
2701 /* Main body of NIC initialisation
2702  * This is called at module load (or hotplug insertion, theoretically).
2703  */
2704 static int efx_pci_probe_main(struct efx_nic *efx)
2705 {
2706 	int rc;
2707 
2708 	/* Do start-of-day initialisation */
2709 	rc = efx_probe_all(efx);
2710 	if (rc)
2711 		goto fail1;
2712 
2713 	efx_init_napi(efx);
2714 
2715 	rc = efx->type->init(efx);
2716 	if (rc) {
2717 		netif_err(efx, probe, efx->net_dev,
2718 			  "failed to initialise NIC\n");
2719 		goto fail3;
2720 	}
2721 
2722 	rc = efx_init_port(efx);
2723 	if (rc) {
2724 		netif_err(efx, probe, efx->net_dev,
2725 			  "failed to initialise port\n");
2726 		goto fail4;
2727 	}
2728 
2729 	rc = efx_nic_init_interrupt(efx);
2730 	if (rc)
2731 		goto fail5;
2732 	efx_start_interrupts(efx, false);
2733 
2734 	return 0;
2735 
2736  fail5:
2737 	efx_fini_port(efx);
2738  fail4:
2739 	efx->type->fini(efx);
2740  fail3:
2741 	efx_fini_napi(efx);
2742 	efx_remove_all(efx);
2743  fail1:
2744 	return rc;
2745 }
2746 
2747 /* NIC initialisation
2748  *
2749  * This is called at module load (or hotplug insertion,
2750  * theoretically).  It sets up PCI mappings, resets the NIC,
2751  * sets up and registers the network devices with the kernel and hooks
2752  * the interrupt service routine.  It does not prepare the device for
2753  * transmission; this is left to the first time one of the network
2754  * interfaces is brought up (i.e. efx_net_open).
2755  */
2756 static int efx_pci_probe(struct pci_dev *pci_dev,
2757 			 const struct pci_device_id *entry)
2758 {
2759 	struct net_device *net_dev;
2760 	struct efx_nic *efx;
2761 	int rc;
2762 
2763 	/* Allocate and initialise a struct net_device and struct efx_nic */
2764 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2765 				     EFX_MAX_RX_QUEUES);
2766 	if (!net_dev)
2767 		return -ENOMEM;
2768 	efx = netdev_priv(net_dev);
2769 	efx->type = (const struct efx_nic_type *) entry->driver_data;
2770 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2771 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
2772 			      NETIF_F_RXCSUM);
2773 	if (efx->type->offload_features & NETIF_F_V6_CSUM)
2774 		net_dev->features |= NETIF_F_TSO6;
2775 	/* Mask for features that also apply to VLAN devices */
2776 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2777 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2778 				   NETIF_F_RXCSUM);
2779 	/* All offloads can be toggled */
2780 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2781 	pci_set_drvdata(pci_dev, efx);
2782 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2783 	rc = efx_init_struct(efx, pci_dev, net_dev);
2784 	if (rc)
2785 		goto fail1;
2786 
2787 	netif_info(efx, probe, efx->net_dev,
2788 		   "Solarflare NIC detected\n");
2789 
2790 	efx_print_product_vpd(efx);
2791 
2792 	/* Set up basic I/O (BAR mappings etc) */
2793 	rc = efx_init_io(efx);
2794 	if (rc)
2795 		goto fail2;
2796 
2797 	rc = efx_pci_probe_main(efx);
2798 	if (rc)
2799 		goto fail3;
2800 
2801 	rc = efx_register_netdev(efx);
2802 	if (rc)
2803 		goto fail4;
2804 
2805 	rc = efx_sriov_init(efx);
2806 	if (rc)
2807 		netif_err(efx, probe, efx->net_dev,
2808 			  "SR-IOV can't be enabled rc %d\n", rc);
2809 
2810 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2811 
2812 	/* Try to create MTDs, but allow this to fail */
2813 	rtnl_lock();
2814 	rc = efx_mtd_probe(efx);
2815 	rtnl_unlock();
2816 	if (rc)
2817 		netif_warn(efx, probe, efx->net_dev,
2818 			   "failed to create MTDs (%d)\n", rc);
2819 
2820 	rc = pci_enable_pcie_error_reporting(pci_dev);
2821 	if (rc && rc != -EINVAL)
2822 		netif_warn(efx, probe, efx->net_dev,
2823 			   "pci_enable_pcie_error_reporting failed (%d)\n", rc);
2824 
2825 	return 0;
2826 
2827  fail4:
2828 	efx_pci_remove_main(efx);
2829  fail3:
2830 	efx_fini_io(efx);
2831  fail2:
2832 	efx_fini_struct(efx);
2833  fail1:
2834 	pci_set_drvdata(pci_dev, NULL);
2835 	WARN_ON(rc > 0);
2836 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2837 	free_netdev(net_dev);
2838 	return rc;
2839 }
2840 
2841 static int efx_pm_freeze(struct device *dev)
2842 {
2843 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2844 
2845 	rtnl_lock();
2846 
2847 	if (efx->state != STATE_DISABLED) {
2848 		efx->state = STATE_UNINIT;
2849 
2850 		efx_device_detach_sync(efx);
2851 
2852 		efx_stop_all(efx);
2853 		efx_stop_interrupts(efx, false);
2854 	}
2855 
2856 	rtnl_unlock();
2857 
2858 	return 0;
2859 }
2860 
2861 static int efx_pm_thaw(struct device *dev)
2862 {
2863 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2864 
2865 	rtnl_lock();
2866 
2867 	if (efx->state != STATE_DISABLED) {
2868 		efx_start_interrupts(efx, false);
2869 
2870 		mutex_lock(&efx->mac_lock);
2871 		efx->phy_op->reconfigure(efx);
2872 		mutex_unlock(&efx->mac_lock);
2873 
2874 		efx_start_all(efx);
2875 
2876 		netif_device_attach(efx->net_dev);
2877 
2878 		efx->state = STATE_READY;
2879 
2880 		efx->type->resume_wol(efx);
2881 	}
2882 
2883 	rtnl_unlock();
2884 
2885 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2886 	queue_work(reset_workqueue, &efx->reset_work);
2887 
2888 	return 0;
2889 }
2890 
2891 static int efx_pm_poweroff(struct device *dev)
2892 {
2893 	struct pci_dev *pci_dev = to_pci_dev(dev);
2894 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2895 
2896 	efx->type->fini(efx);
2897 
2898 	efx->reset_pending = 0;
2899 
2900 	pci_save_state(pci_dev);
2901 	return pci_set_power_state(pci_dev, PCI_D3hot);
2902 }
2903 
2904 /* Used for both resume and restore */
2905 static int efx_pm_resume(struct device *dev)
2906 {
2907 	struct pci_dev *pci_dev = to_pci_dev(dev);
2908 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2909 	int rc;
2910 
2911 	rc = pci_set_power_state(pci_dev, PCI_D0);
2912 	if (rc)
2913 		return rc;
2914 	pci_restore_state(pci_dev);
2915 	rc = pci_enable_device(pci_dev);
2916 	if (rc)
2917 		return rc;
2918 	pci_set_master(efx->pci_dev);
2919 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
2920 	if (rc)
2921 		return rc;
2922 	rc = efx->type->init(efx);
2923 	if (rc)
2924 		return rc;
2925 	efx_pm_thaw(dev);
2926 	return 0;
2927 }
2928 
2929 static int efx_pm_suspend(struct device *dev)
2930 {
2931 	int rc;
2932 
2933 	efx_pm_freeze(dev);
2934 	rc = efx_pm_poweroff(dev);
2935 	if (rc)
2936 		efx_pm_resume(dev);
2937 	return rc;
2938 }
2939 
2940 static const struct dev_pm_ops efx_pm_ops = {
2941 	.suspend	= efx_pm_suspend,
2942 	.resume		= efx_pm_resume,
2943 	.freeze		= efx_pm_freeze,
2944 	.thaw		= efx_pm_thaw,
2945 	.poweroff	= efx_pm_poweroff,
2946 	.restore	= efx_pm_resume,
2947 };
2948 
2949 /* A PCI error affecting this device was detected.
2950  * At this point MMIO and DMA may be disabled.
2951  * Stop the software path and request a slot reset.
2952  */
2953 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
2954 					      enum pci_channel_state state)
2955 {
2956 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
2957 	struct efx_nic *efx = pci_get_drvdata(pdev);
2958 
2959 	if (state == pci_channel_io_perm_failure)
2960 		return PCI_ERS_RESULT_DISCONNECT;
2961 
2962 	rtnl_lock();
2963 
2964 	if (efx->state != STATE_DISABLED) {
2965 		efx->state = STATE_RECOVERY;
2966 		efx->reset_pending = 0;
2967 
2968 		efx_device_detach_sync(efx);
2969 
2970 		efx_stop_all(efx);
2971 		efx_stop_interrupts(efx, false);
2972 
2973 		status = PCI_ERS_RESULT_NEED_RESET;
2974 	} else {
2975 		/* If the interface is disabled we don't want to do anything
2976 		 * with it.
2977 		 */
2978 		status = PCI_ERS_RESULT_RECOVERED;
2979 	}
2980 
2981 	rtnl_unlock();
2982 
2983 	pci_disable_device(pdev);
2984 
2985 	return status;
2986 }
2987 
2988 /* Fake a successfull reset, which will be performed later in efx_io_resume. */
2989 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
2990 {
2991 	struct efx_nic *efx = pci_get_drvdata(pdev);
2992 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
2993 	int rc;
2994 
2995 	if (pci_enable_device(pdev)) {
2996 		netif_err(efx, hw, efx->net_dev,
2997 			  "Cannot re-enable PCI device after reset.\n");
2998 		status =  PCI_ERS_RESULT_DISCONNECT;
2999 	}
3000 
3001 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3002 	if (rc) {
3003 		netif_err(efx, hw, efx->net_dev,
3004 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3005 		/* Non-fatal error. Continue. */
3006 	}
3007 
3008 	return status;
3009 }
3010 
3011 /* Perform the actual reset and resume I/O operations. */
3012 static void efx_io_resume(struct pci_dev *pdev)
3013 {
3014 	struct efx_nic *efx = pci_get_drvdata(pdev);
3015 	int rc;
3016 
3017 	rtnl_lock();
3018 
3019 	if (efx->state == STATE_DISABLED)
3020 		goto out;
3021 
3022 	rc = efx_reset(efx, RESET_TYPE_ALL);
3023 	if (rc) {
3024 		netif_err(efx, hw, efx->net_dev,
3025 			  "efx_reset failed after PCI error (%d)\n", rc);
3026 	} else {
3027 		efx->state = STATE_READY;
3028 		netif_dbg(efx, hw, efx->net_dev,
3029 			  "Done resetting and resuming IO after PCI error.\n");
3030 	}
3031 
3032 out:
3033 	rtnl_unlock();
3034 }
3035 
3036 /* For simplicity and reliability, we always require a slot reset and try to
3037  * reset the hardware when a pci error affecting the device is detected.
3038  * We leave both the link_reset and mmio_enabled callback unimplemented:
3039  * with our request for slot reset the mmio_enabled callback will never be
3040  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3041  */
3042 static struct pci_error_handlers efx_err_handlers = {
3043 	.error_detected = efx_io_error_detected,
3044 	.slot_reset	= efx_io_slot_reset,
3045 	.resume		= efx_io_resume,
3046 };
3047 
3048 static struct pci_driver efx_pci_driver = {
3049 	.name		= KBUILD_MODNAME,
3050 	.id_table	= efx_pci_table,
3051 	.probe		= efx_pci_probe,
3052 	.remove		= efx_pci_remove,
3053 	.driver.pm	= &efx_pm_ops,
3054 	.err_handler	= &efx_err_handlers,
3055 };
3056 
3057 /**************************************************************************
3058  *
3059  * Kernel module interface
3060  *
3061  *************************************************************************/
3062 
3063 module_param(interrupt_mode, uint, 0444);
3064 MODULE_PARM_DESC(interrupt_mode,
3065 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3066 
3067 static int __init efx_init_module(void)
3068 {
3069 	int rc;
3070 
3071 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3072 
3073 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3074 	if (rc)
3075 		goto err_notifier;
3076 
3077 	rc = efx_init_sriov();
3078 	if (rc)
3079 		goto err_sriov;
3080 
3081 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3082 	if (!reset_workqueue) {
3083 		rc = -ENOMEM;
3084 		goto err_reset;
3085 	}
3086 
3087 	rc = pci_register_driver(&efx_pci_driver);
3088 	if (rc < 0)
3089 		goto err_pci;
3090 
3091 	return 0;
3092 
3093  err_pci:
3094 	destroy_workqueue(reset_workqueue);
3095  err_reset:
3096 	efx_fini_sriov();
3097  err_sriov:
3098 	unregister_netdevice_notifier(&efx_netdev_notifier);
3099  err_notifier:
3100 	return rc;
3101 }
3102 
3103 static void __exit efx_exit_module(void)
3104 {
3105 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3106 
3107 	pci_unregister_driver(&efx_pci_driver);
3108 	destroy_workqueue(reset_workqueue);
3109 	efx_fini_sriov();
3110 	unregister_netdevice_notifier(&efx_netdev_notifier);
3111 
3112 }
3113 
3114 module_init(efx_init_module);
3115 module_exit(efx_exit_module);
3116 
3117 MODULE_AUTHOR("Solarflare Communications and "
3118 	      "Michael Brown <mbrown@fensystems.co.uk>");
3119 MODULE_DESCRIPTION("Solarflare Communications network driver");
3120 MODULE_LICENSE("GPL");
3121 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3122