xref: /openbmc/linux/drivers/net/ethernet/sfc/efx.c (revision 8730046c)
1 /****************************************************************************
2  * Driver for Solarflare network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2013 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/ethtool.h>
21 #include <linux/topology.h>
22 #include <linux/gfp.h>
23 #include <linux/aer.h>
24 #include <linux/interrupt.h>
25 #include "net_driver.h"
26 #include "efx.h"
27 #include "nic.h"
28 #include "selftest.h"
29 #include "sriov.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_DATAPATH]           = "DATAPATH",
81 	[RESET_TYPE_MC_BIST]		= "MC_BIST",
82 	[RESET_TYPE_DISABLE]            = "DISABLE",
83 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
84 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
85 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
86 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
87 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
88 	[RESET_TYPE_MCDI_TIMEOUT]	= "MCDI_TIMEOUT (FLR)",
89 };
90 
91 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
92  * queued onto this work queue. This is not a per-nic work queue, because
93  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
94  */
95 static struct workqueue_struct *reset_workqueue;
96 
97 /* How often and how many times to poll for a reset while waiting for a
98  * BIST that another function started to complete.
99  */
100 #define BIST_WAIT_DELAY_MS	100
101 #define BIST_WAIT_DELAY_COUNT	100
102 
103 /**************************************************************************
104  *
105  * Configurable values
106  *
107  *************************************************************************/
108 
109 /*
110  * Use separate channels for TX and RX events
111  *
112  * Set this to 1 to use separate channels for TX and RX. It allows us
113  * to control interrupt affinity separately for TX and RX.
114  *
115  * This is only used in MSI-X interrupt mode
116  */
117 bool efx_separate_tx_channels;
118 module_param(efx_separate_tx_channels, bool, 0444);
119 MODULE_PARM_DESC(efx_separate_tx_channels,
120 		 "Use separate channels for TX and RX");
121 
122 /* This is the weight assigned to each of the (per-channel) virtual
123  * NAPI devices.
124  */
125 static int napi_weight = 64;
126 
127 /* This is the time (in jiffies) between invocations of the hardware
128  * monitor.
129  * On Falcon-based NICs, this will:
130  * - Check the on-board hardware monitor;
131  * - Poll the link state and reconfigure the hardware as necessary.
132  * On Siena-based NICs for power systems with EEH support, this will give EEH a
133  * chance to start.
134  */
135 static unsigned int efx_monitor_interval = 1 * HZ;
136 
137 /* Initial interrupt moderation settings.  They can be modified after
138  * module load with ethtool.
139  *
140  * The default for RX should strike a balance between increasing the
141  * round-trip latency and reducing overhead.
142  */
143 static unsigned int rx_irq_mod_usec = 60;
144 
145 /* Initial interrupt moderation settings.  They can be modified after
146  * module load with ethtool.
147  *
148  * This default is chosen to ensure that a 10G link does not go idle
149  * while a TX queue is stopped after it has become full.  A queue is
150  * restarted when it drops below half full.  The time this takes (assuming
151  * worst case 3 descriptors per packet and 1024 descriptors) is
152  *   512 / 3 * 1.2 = 205 usec.
153  */
154 static unsigned int tx_irq_mod_usec = 150;
155 
156 /* This is the first interrupt mode to try out of:
157  * 0 => MSI-X
158  * 1 => MSI
159  * 2 => legacy
160  */
161 static unsigned int interrupt_mode;
162 
163 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
164  * i.e. the number of CPUs among which we may distribute simultaneous
165  * interrupt handling.
166  *
167  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
168  * The default (0) means to assign an interrupt to each core.
169  */
170 static unsigned int rss_cpus;
171 module_param(rss_cpus, uint, 0444);
172 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
173 
174 static bool phy_flash_cfg;
175 module_param(phy_flash_cfg, bool, 0644);
176 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
177 
178 static unsigned irq_adapt_low_thresh = 8000;
179 module_param(irq_adapt_low_thresh, uint, 0644);
180 MODULE_PARM_DESC(irq_adapt_low_thresh,
181 		 "Threshold score for reducing IRQ moderation");
182 
183 static unsigned irq_adapt_high_thresh = 16000;
184 module_param(irq_adapt_high_thresh, uint, 0644);
185 MODULE_PARM_DESC(irq_adapt_high_thresh,
186 		 "Threshold score for increasing IRQ moderation");
187 
188 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
189 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
190 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
191 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
192 module_param(debug, uint, 0);
193 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
194 
195 /**************************************************************************
196  *
197  * Utility functions and prototypes
198  *
199  *************************************************************************/
200 
201 static int efx_soft_enable_interrupts(struct efx_nic *efx);
202 static void efx_soft_disable_interrupts(struct efx_nic *efx);
203 static void efx_remove_channel(struct efx_channel *channel);
204 static void efx_remove_channels(struct efx_nic *efx);
205 static const struct efx_channel_type efx_default_channel_type;
206 static void efx_remove_port(struct efx_nic *efx);
207 static void efx_init_napi_channel(struct efx_channel *channel);
208 static void efx_fini_napi(struct efx_nic *efx);
209 static void efx_fini_napi_channel(struct efx_channel *channel);
210 static void efx_fini_struct(struct efx_nic *efx);
211 static void efx_start_all(struct efx_nic *efx);
212 static void efx_stop_all(struct efx_nic *efx);
213 
214 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
215 	do {						\
216 		if ((efx->state == STATE_READY) ||	\
217 		    (efx->state == STATE_RECOVERY) ||	\
218 		    (efx->state == STATE_DISABLED))	\
219 			ASSERT_RTNL();			\
220 	} while (0)
221 
222 static int efx_check_disabled(struct efx_nic *efx)
223 {
224 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
225 		netif_err(efx, drv, efx->net_dev,
226 			  "device is disabled due to earlier errors\n");
227 		return -EIO;
228 	}
229 	return 0;
230 }
231 
232 /**************************************************************************
233  *
234  * Event queue processing
235  *
236  *************************************************************************/
237 
238 /* Process channel's event queue
239  *
240  * This function is responsible for processing the event queue of a
241  * single channel.  The caller must guarantee that this function will
242  * never be concurrently called more than once on the same channel,
243  * though different channels may be being processed concurrently.
244  */
245 static int efx_process_channel(struct efx_channel *channel, int budget)
246 {
247 	struct efx_tx_queue *tx_queue;
248 	int spent;
249 
250 	if (unlikely(!channel->enabled))
251 		return 0;
252 
253 	efx_for_each_channel_tx_queue(tx_queue, channel) {
254 		tx_queue->pkts_compl = 0;
255 		tx_queue->bytes_compl = 0;
256 	}
257 
258 	spent = efx_nic_process_eventq(channel, budget);
259 	if (spent && efx_channel_has_rx_queue(channel)) {
260 		struct efx_rx_queue *rx_queue =
261 			efx_channel_get_rx_queue(channel);
262 
263 		efx_rx_flush_packet(channel);
264 		efx_fast_push_rx_descriptors(rx_queue, true);
265 	}
266 
267 	/* Update BQL */
268 	efx_for_each_channel_tx_queue(tx_queue, channel) {
269 		if (tx_queue->bytes_compl) {
270 			netdev_tx_completed_queue(tx_queue->core_txq,
271 				tx_queue->pkts_compl, tx_queue->bytes_compl);
272 		}
273 	}
274 
275 	return spent;
276 }
277 
278 /* NAPI poll handler
279  *
280  * NAPI guarantees serialisation of polls of the same device, which
281  * provides the guarantee required by efx_process_channel().
282  */
283 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel)
284 {
285 	int step = efx->irq_mod_step_us;
286 
287 	if (channel->irq_mod_score < irq_adapt_low_thresh) {
288 		if (channel->irq_moderation_us > step) {
289 			channel->irq_moderation_us -= step;
290 			efx->type->push_irq_moderation(channel);
291 		}
292 	} else if (channel->irq_mod_score > irq_adapt_high_thresh) {
293 		if (channel->irq_moderation_us <
294 		    efx->irq_rx_moderation_us) {
295 			channel->irq_moderation_us += step;
296 			efx->type->push_irq_moderation(channel);
297 		}
298 	}
299 
300 	channel->irq_count = 0;
301 	channel->irq_mod_score = 0;
302 }
303 
304 static int efx_poll(struct napi_struct *napi, int budget)
305 {
306 	struct efx_channel *channel =
307 		container_of(napi, struct efx_channel, napi_str);
308 	struct efx_nic *efx = channel->efx;
309 	int spent;
310 
311 	if (!efx_channel_lock_napi(channel))
312 		return budget;
313 
314 	netif_vdbg(efx, intr, efx->net_dev,
315 		   "channel %d NAPI poll executing on CPU %d\n",
316 		   channel->channel, raw_smp_processor_id());
317 
318 	spent = efx_process_channel(channel, budget);
319 
320 	if (spent < budget) {
321 		if (efx_channel_has_rx_queue(channel) &&
322 		    efx->irq_rx_adaptive &&
323 		    unlikely(++channel->irq_count == 1000)) {
324 			efx_update_irq_mod(efx, channel);
325 		}
326 
327 		efx_filter_rfs_expire(channel);
328 
329 		/* There is no race here; although napi_disable() will
330 		 * only wait for napi_complete(), this isn't a problem
331 		 * since efx_nic_eventq_read_ack() will have no effect if
332 		 * interrupts have already been disabled.
333 		 */
334 		napi_complete(napi);
335 		efx_nic_eventq_read_ack(channel);
336 	}
337 
338 	efx_channel_unlock_napi(channel);
339 	return spent;
340 }
341 
342 /* Create event queue
343  * Event queue memory allocations are done only once.  If the channel
344  * is reset, the memory buffer will be reused; this guards against
345  * errors during channel reset and also simplifies interrupt handling.
346  */
347 static int efx_probe_eventq(struct efx_channel *channel)
348 {
349 	struct efx_nic *efx = channel->efx;
350 	unsigned long entries;
351 
352 	netif_dbg(efx, probe, efx->net_dev,
353 		  "chan %d create event queue\n", channel->channel);
354 
355 	/* Build an event queue with room for one event per tx and rx buffer,
356 	 * plus some extra for link state events and MCDI completions. */
357 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
358 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
359 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
360 
361 	return efx_nic_probe_eventq(channel);
362 }
363 
364 /* Prepare channel's event queue */
365 static int efx_init_eventq(struct efx_channel *channel)
366 {
367 	struct efx_nic *efx = channel->efx;
368 	int rc;
369 
370 	EFX_WARN_ON_PARANOID(channel->eventq_init);
371 
372 	netif_dbg(efx, drv, efx->net_dev,
373 		  "chan %d init event queue\n", channel->channel);
374 
375 	rc = efx_nic_init_eventq(channel);
376 	if (rc == 0) {
377 		efx->type->push_irq_moderation(channel);
378 		channel->eventq_read_ptr = 0;
379 		channel->eventq_init = true;
380 	}
381 	return rc;
382 }
383 
384 /* Enable event queue processing and NAPI */
385 void efx_start_eventq(struct efx_channel *channel)
386 {
387 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
388 		  "chan %d start event queue\n", channel->channel);
389 
390 	/* Make sure the NAPI handler sees the enabled flag set */
391 	channel->enabled = true;
392 	smp_wmb();
393 
394 	efx_channel_enable(channel);
395 	napi_enable(&channel->napi_str);
396 	efx_nic_eventq_read_ack(channel);
397 }
398 
399 /* Disable event queue processing and NAPI */
400 void efx_stop_eventq(struct efx_channel *channel)
401 {
402 	if (!channel->enabled)
403 		return;
404 
405 	napi_disable(&channel->napi_str);
406 	while (!efx_channel_disable(channel))
407 		usleep_range(1000, 20000);
408 	channel->enabled = false;
409 }
410 
411 static void efx_fini_eventq(struct efx_channel *channel)
412 {
413 	if (!channel->eventq_init)
414 		return;
415 
416 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
417 		  "chan %d fini event queue\n", channel->channel);
418 
419 	efx_nic_fini_eventq(channel);
420 	channel->eventq_init = false;
421 }
422 
423 static void efx_remove_eventq(struct efx_channel *channel)
424 {
425 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
426 		  "chan %d remove event queue\n", channel->channel);
427 
428 	efx_nic_remove_eventq(channel);
429 }
430 
431 /**************************************************************************
432  *
433  * Channel handling
434  *
435  *************************************************************************/
436 
437 /* Allocate and initialise a channel structure. */
438 static struct efx_channel *
439 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
440 {
441 	struct efx_channel *channel;
442 	struct efx_rx_queue *rx_queue;
443 	struct efx_tx_queue *tx_queue;
444 	int j;
445 
446 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
447 	if (!channel)
448 		return NULL;
449 
450 	channel->efx = efx;
451 	channel->channel = i;
452 	channel->type = &efx_default_channel_type;
453 
454 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
455 		tx_queue = &channel->tx_queue[j];
456 		tx_queue->efx = efx;
457 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
458 		tx_queue->channel = channel;
459 	}
460 
461 	rx_queue = &channel->rx_queue;
462 	rx_queue->efx = efx;
463 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
464 		    (unsigned long)rx_queue);
465 
466 	return channel;
467 }
468 
469 /* Allocate and initialise a channel structure, copying parameters
470  * (but not resources) from an old channel structure.
471  */
472 static struct efx_channel *
473 efx_copy_channel(const struct efx_channel *old_channel)
474 {
475 	struct efx_channel *channel;
476 	struct efx_rx_queue *rx_queue;
477 	struct efx_tx_queue *tx_queue;
478 	int j;
479 
480 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
481 	if (!channel)
482 		return NULL;
483 
484 	*channel = *old_channel;
485 
486 	channel->napi_dev = NULL;
487 	INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
488 	channel->napi_str.napi_id = 0;
489 	channel->napi_str.state = 0;
490 	memset(&channel->eventq, 0, sizeof(channel->eventq));
491 
492 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
493 		tx_queue = &channel->tx_queue[j];
494 		if (tx_queue->channel)
495 			tx_queue->channel = channel;
496 		tx_queue->buffer = NULL;
497 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
498 	}
499 
500 	rx_queue = &channel->rx_queue;
501 	rx_queue->buffer = NULL;
502 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
503 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
504 		    (unsigned long)rx_queue);
505 
506 	return channel;
507 }
508 
509 static int efx_probe_channel(struct efx_channel *channel)
510 {
511 	struct efx_tx_queue *tx_queue;
512 	struct efx_rx_queue *rx_queue;
513 	int rc;
514 
515 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
516 		  "creating channel %d\n", channel->channel);
517 
518 	rc = channel->type->pre_probe(channel);
519 	if (rc)
520 		goto fail;
521 
522 	rc = efx_probe_eventq(channel);
523 	if (rc)
524 		goto fail;
525 
526 	efx_for_each_channel_tx_queue(tx_queue, channel) {
527 		rc = efx_probe_tx_queue(tx_queue);
528 		if (rc)
529 			goto fail;
530 	}
531 
532 	efx_for_each_channel_rx_queue(rx_queue, channel) {
533 		rc = efx_probe_rx_queue(rx_queue);
534 		if (rc)
535 			goto fail;
536 	}
537 
538 	return 0;
539 
540 fail:
541 	efx_remove_channel(channel);
542 	return rc;
543 }
544 
545 static void
546 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
547 {
548 	struct efx_nic *efx = channel->efx;
549 	const char *type;
550 	int number;
551 
552 	number = channel->channel;
553 	if (efx->tx_channel_offset == 0) {
554 		type = "";
555 	} else if (channel->channel < efx->tx_channel_offset) {
556 		type = "-rx";
557 	} else {
558 		type = "-tx";
559 		number -= efx->tx_channel_offset;
560 	}
561 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
562 }
563 
564 static void efx_set_channel_names(struct efx_nic *efx)
565 {
566 	struct efx_channel *channel;
567 
568 	efx_for_each_channel(channel, efx)
569 		channel->type->get_name(channel,
570 					efx->msi_context[channel->channel].name,
571 					sizeof(efx->msi_context[0].name));
572 }
573 
574 static int efx_probe_channels(struct efx_nic *efx)
575 {
576 	struct efx_channel *channel;
577 	int rc;
578 
579 	/* Restart special buffer allocation */
580 	efx->next_buffer_table = 0;
581 
582 	/* Probe channels in reverse, so that any 'extra' channels
583 	 * use the start of the buffer table. This allows the traffic
584 	 * channels to be resized without moving them or wasting the
585 	 * entries before them.
586 	 */
587 	efx_for_each_channel_rev(channel, efx) {
588 		rc = efx_probe_channel(channel);
589 		if (rc) {
590 			netif_err(efx, probe, efx->net_dev,
591 				  "failed to create channel %d\n",
592 				  channel->channel);
593 			goto fail;
594 		}
595 	}
596 	efx_set_channel_names(efx);
597 
598 	return 0;
599 
600 fail:
601 	efx_remove_channels(efx);
602 	return rc;
603 }
604 
605 /* Channels are shutdown and reinitialised whilst the NIC is running
606  * to propagate configuration changes (mtu, checksum offload), or
607  * to clear hardware error conditions
608  */
609 static void efx_start_datapath(struct efx_nic *efx)
610 {
611 	netdev_features_t old_features = efx->net_dev->features;
612 	bool old_rx_scatter = efx->rx_scatter;
613 	struct efx_tx_queue *tx_queue;
614 	struct efx_rx_queue *rx_queue;
615 	struct efx_channel *channel;
616 	size_t rx_buf_len;
617 
618 	/* Calculate the rx buffer allocation parameters required to
619 	 * support the current MTU, including padding for header
620 	 * alignment and overruns.
621 	 */
622 	efx->rx_dma_len = (efx->rx_prefix_size +
623 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
624 			   efx->type->rx_buffer_padding);
625 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
626 		      efx->rx_ip_align + efx->rx_dma_len);
627 	if (rx_buf_len <= PAGE_SIZE) {
628 		efx->rx_scatter = efx->type->always_rx_scatter;
629 		efx->rx_buffer_order = 0;
630 	} else if (efx->type->can_rx_scatter) {
631 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
632 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
633 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
634 				       EFX_RX_BUF_ALIGNMENT) >
635 			     PAGE_SIZE);
636 		efx->rx_scatter = true;
637 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
638 		efx->rx_buffer_order = 0;
639 	} else {
640 		efx->rx_scatter = false;
641 		efx->rx_buffer_order = get_order(rx_buf_len);
642 	}
643 
644 	efx_rx_config_page_split(efx);
645 	if (efx->rx_buffer_order)
646 		netif_dbg(efx, drv, efx->net_dev,
647 			  "RX buf len=%u; page order=%u batch=%u\n",
648 			  efx->rx_dma_len, efx->rx_buffer_order,
649 			  efx->rx_pages_per_batch);
650 	else
651 		netif_dbg(efx, drv, efx->net_dev,
652 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
653 			  efx->rx_dma_len, efx->rx_page_buf_step,
654 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
655 
656 	/* Restore previously fixed features in hw_features and remove
657 	 * features which are fixed now
658 	 */
659 	efx->net_dev->hw_features |= efx->net_dev->features;
660 	efx->net_dev->hw_features &= ~efx->fixed_features;
661 	efx->net_dev->features |= efx->fixed_features;
662 	if (efx->net_dev->features != old_features)
663 		netdev_features_change(efx->net_dev);
664 
665 	/* RX filters may also have scatter-enabled flags */
666 	if (efx->rx_scatter != old_rx_scatter)
667 		efx->type->filter_update_rx_scatter(efx);
668 
669 	/* We must keep at least one descriptor in a TX ring empty.
670 	 * We could avoid this when the queue size does not exactly
671 	 * match the hardware ring size, but it's not that important.
672 	 * Therefore we stop the queue when one more skb might fill
673 	 * the ring completely.  We wake it when half way back to
674 	 * empty.
675 	 */
676 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
677 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
678 
679 	/* Initialise the channels */
680 	efx_for_each_channel(channel, efx) {
681 		efx_for_each_channel_tx_queue(tx_queue, channel) {
682 			efx_init_tx_queue(tx_queue);
683 			atomic_inc(&efx->active_queues);
684 		}
685 
686 		efx_for_each_channel_rx_queue(rx_queue, channel) {
687 			efx_init_rx_queue(rx_queue);
688 			atomic_inc(&efx->active_queues);
689 			efx_stop_eventq(channel);
690 			efx_fast_push_rx_descriptors(rx_queue, false);
691 			efx_start_eventq(channel);
692 		}
693 
694 		WARN_ON(channel->rx_pkt_n_frags);
695 	}
696 
697 	efx_ptp_start_datapath(efx);
698 
699 	if (netif_device_present(efx->net_dev))
700 		netif_tx_wake_all_queues(efx->net_dev);
701 }
702 
703 static void efx_stop_datapath(struct efx_nic *efx)
704 {
705 	struct efx_channel *channel;
706 	struct efx_tx_queue *tx_queue;
707 	struct efx_rx_queue *rx_queue;
708 	int rc;
709 
710 	EFX_ASSERT_RESET_SERIALISED(efx);
711 	BUG_ON(efx->port_enabled);
712 
713 	efx_ptp_stop_datapath(efx);
714 
715 	/* Stop RX refill */
716 	efx_for_each_channel(channel, efx) {
717 		efx_for_each_channel_rx_queue(rx_queue, channel)
718 			rx_queue->refill_enabled = false;
719 	}
720 
721 	efx_for_each_channel(channel, efx) {
722 		/* RX packet processing is pipelined, so wait for the
723 		 * NAPI handler to complete.  At least event queue 0
724 		 * might be kept active by non-data events, so don't
725 		 * use napi_synchronize() but actually disable NAPI
726 		 * temporarily.
727 		 */
728 		if (efx_channel_has_rx_queue(channel)) {
729 			efx_stop_eventq(channel);
730 			efx_start_eventq(channel);
731 		}
732 	}
733 
734 	rc = efx->type->fini_dmaq(efx);
735 	if (rc) {
736 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
737 	} else {
738 		netif_dbg(efx, drv, efx->net_dev,
739 			  "successfully flushed all queues\n");
740 	}
741 
742 	efx_for_each_channel(channel, efx) {
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, rc2;
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_soft_disable_interrupts(efx);
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 	rc2 = efx_soft_enable_interrupts(efx);
861 	if (rc2) {
862 		rc = rc ? rc : rc2;
863 		netif_err(efx, drv, efx->net_dev,
864 			  "unable to restart interrupts on channel reallocation\n");
865 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
866 	} else {
867 		efx_start_all(efx);
868 		netif_device_attach(efx->net_dev);
869 	}
870 	return rc;
871 
872 rollback:
873 	/* Swap back */
874 	efx->rxq_entries = old_rxq_entries;
875 	efx->txq_entries = old_txq_entries;
876 	for (i = 0; i < efx->n_channels; i++) {
877 		channel = efx->channel[i];
878 		efx->channel[i] = other_channel[i];
879 		other_channel[i] = channel;
880 	}
881 	goto out;
882 }
883 
884 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
885 {
886 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
887 }
888 
889 static const struct efx_channel_type efx_default_channel_type = {
890 	.pre_probe		= efx_channel_dummy_op_int,
891 	.post_remove		= efx_channel_dummy_op_void,
892 	.get_name		= efx_get_channel_name,
893 	.copy			= efx_copy_channel,
894 	.keep_eventq		= false,
895 };
896 
897 int efx_channel_dummy_op_int(struct efx_channel *channel)
898 {
899 	return 0;
900 }
901 
902 void efx_channel_dummy_op_void(struct efx_channel *channel)
903 {
904 }
905 
906 /**************************************************************************
907  *
908  * Port handling
909  *
910  **************************************************************************/
911 
912 /* This ensures that the kernel is kept informed (via
913  * netif_carrier_on/off) of the link status, and also maintains the
914  * link status's stop on the port's TX queue.
915  */
916 void efx_link_status_changed(struct efx_nic *efx)
917 {
918 	struct efx_link_state *link_state = &efx->link_state;
919 
920 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
921 	 * that no events are triggered between unregister_netdev() and the
922 	 * driver unloading. A more general condition is that NETDEV_CHANGE
923 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
924 	if (!netif_running(efx->net_dev))
925 		return;
926 
927 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
928 		efx->n_link_state_changes++;
929 
930 		if (link_state->up)
931 			netif_carrier_on(efx->net_dev);
932 		else
933 			netif_carrier_off(efx->net_dev);
934 	}
935 
936 	/* Status message for kernel log */
937 	if (link_state->up)
938 		netif_info(efx, link, efx->net_dev,
939 			   "link up at %uMbps %s-duplex (MTU %d)\n",
940 			   link_state->speed, link_state->fd ? "full" : "half",
941 			   efx->net_dev->mtu);
942 	else
943 		netif_info(efx, link, efx->net_dev, "link down\n");
944 }
945 
946 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
947 {
948 	efx->link_advertising = advertising;
949 	if (advertising) {
950 		if (advertising & ADVERTISED_Pause)
951 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
952 		else
953 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
954 		if (advertising & ADVERTISED_Asym_Pause)
955 			efx->wanted_fc ^= EFX_FC_TX;
956 	}
957 }
958 
959 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
960 {
961 	efx->wanted_fc = wanted_fc;
962 	if (efx->link_advertising) {
963 		if (wanted_fc & EFX_FC_RX)
964 			efx->link_advertising |= (ADVERTISED_Pause |
965 						  ADVERTISED_Asym_Pause);
966 		else
967 			efx->link_advertising &= ~(ADVERTISED_Pause |
968 						   ADVERTISED_Asym_Pause);
969 		if (wanted_fc & EFX_FC_TX)
970 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
971 	}
972 }
973 
974 static void efx_fini_port(struct efx_nic *efx);
975 
976 /* We assume that efx->type->reconfigure_mac will always try to sync RX
977  * filters and therefore needs to read-lock the filter table against freeing
978  */
979 void efx_mac_reconfigure(struct efx_nic *efx)
980 {
981 	down_read(&efx->filter_sem);
982 	efx->type->reconfigure_mac(efx);
983 	up_read(&efx->filter_sem);
984 }
985 
986 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
987  * the MAC appropriately. All other PHY configuration changes are pushed
988  * through phy_op->set_settings(), and pushed asynchronously to the MAC
989  * through efx_monitor().
990  *
991  * Callers must hold the mac_lock
992  */
993 int __efx_reconfigure_port(struct efx_nic *efx)
994 {
995 	enum efx_phy_mode phy_mode;
996 	int rc;
997 
998 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
999 
1000 	/* Disable PHY transmit in mac level loopbacks */
1001 	phy_mode = efx->phy_mode;
1002 	if (LOOPBACK_INTERNAL(efx))
1003 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
1004 	else
1005 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
1006 
1007 	rc = efx->type->reconfigure_port(efx);
1008 
1009 	if (rc)
1010 		efx->phy_mode = phy_mode;
1011 
1012 	return rc;
1013 }
1014 
1015 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1016  * disabled. */
1017 int efx_reconfigure_port(struct efx_nic *efx)
1018 {
1019 	int rc;
1020 
1021 	EFX_ASSERT_RESET_SERIALISED(efx);
1022 
1023 	mutex_lock(&efx->mac_lock);
1024 	rc = __efx_reconfigure_port(efx);
1025 	mutex_unlock(&efx->mac_lock);
1026 
1027 	return rc;
1028 }
1029 
1030 /* Asynchronous work item for changing MAC promiscuity and multicast
1031  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1032  * MAC directly. */
1033 static void efx_mac_work(struct work_struct *data)
1034 {
1035 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1036 
1037 	mutex_lock(&efx->mac_lock);
1038 	if (efx->port_enabled)
1039 		efx_mac_reconfigure(efx);
1040 	mutex_unlock(&efx->mac_lock);
1041 }
1042 
1043 static int efx_probe_port(struct efx_nic *efx)
1044 {
1045 	int rc;
1046 
1047 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1048 
1049 	if (phy_flash_cfg)
1050 		efx->phy_mode = PHY_MODE_SPECIAL;
1051 
1052 	/* Connect up MAC/PHY operations table */
1053 	rc = efx->type->probe_port(efx);
1054 	if (rc)
1055 		return rc;
1056 
1057 	/* Initialise MAC address to permanent address */
1058 	ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1059 
1060 	return 0;
1061 }
1062 
1063 static int efx_init_port(struct efx_nic *efx)
1064 {
1065 	int rc;
1066 
1067 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1068 
1069 	mutex_lock(&efx->mac_lock);
1070 
1071 	rc = efx->phy_op->init(efx);
1072 	if (rc)
1073 		goto fail1;
1074 
1075 	efx->port_initialized = true;
1076 
1077 	/* Reconfigure the MAC before creating dma queues (required for
1078 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1079 	efx_mac_reconfigure(efx);
1080 
1081 	/* Ensure the PHY advertises the correct flow control settings */
1082 	rc = efx->phy_op->reconfigure(efx);
1083 	if (rc && rc != -EPERM)
1084 		goto fail2;
1085 
1086 	mutex_unlock(&efx->mac_lock);
1087 	return 0;
1088 
1089 fail2:
1090 	efx->phy_op->fini(efx);
1091 fail1:
1092 	mutex_unlock(&efx->mac_lock);
1093 	return rc;
1094 }
1095 
1096 static void efx_start_port(struct efx_nic *efx)
1097 {
1098 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1099 	BUG_ON(efx->port_enabled);
1100 
1101 	mutex_lock(&efx->mac_lock);
1102 	efx->port_enabled = true;
1103 
1104 	/* Ensure MAC ingress/egress is enabled */
1105 	efx_mac_reconfigure(efx);
1106 
1107 	mutex_unlock(&efx->mac_lock);
1108 }
1109 
1110 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1111  * and the async self-test, wait for them to finish and prevent them
1112  * being scheduled again.  This doesn't cover online resets, which
1113  * should only be cancelled when removing the device.
1114  */
1115 static void efx_stop_port(struct efx_nic *efx)
1116 {
1117 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1118 
1119 	EFX_ASSERT_RESET_SERIALISED(efx);
1120 
1121 	mutex_lock(&efx->mac_lock);
1122 	efx->port_enabled = false;
1123 	mutex_unlock(&efx->mac_lock);
1124 
1125 	/* Serialise against efx_set_multicast_list() */
1126 	netif_addr_lock_bh(efx->net_dev);
1127 	netif_addr_unlock_bh(efx->net_dev);
1128 
1129 	cancel_delayed_work_sync(&efx->monitor_work);
1130 	efx_selftest_async_cancel(efx);
1131 	cancel_work_sync(&efx->mac_work);
1132 }
1133 
1134 static void efx_fini_port(struct efx_nic *efx)
1135 {
1136 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1137 
1138 	if (!efx->port_initialized)
1139 		return;
1140 
1141 	efx->phy_op->fini(efx);
1142 	efx->port_initialized = false;
1143 
1144 	efx->link_state.up = false;
1145 	efx_link_status_changed(efx);
1146 }
1147 
1148 static void efx_remove_port(struct efx_nic *efx)
1149 {
1150 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1151 
1152 	efx->type->remove_port(efx);
1153 }
1154 
1155 /**************************************************************************
1156  *
1157  * NIC handling
1158  *
1159  **************************************************************************/
1160 
1161 static LIST_HEAD(efx_primary_list);
1162 static LIST_HEAD(efx_unassociated_list);
1163 
1164 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1165 {
1166 	return left->type == right->type &&
1167 		left->vpd_sn && right->vpd_sn &&
1168 		!strcmp(left->vpd_sn, right->vpd_sn);
1169 }
1170 
1171 static void efx_associate(struct efx_nic *efx)
1172 {
1173 	struct efx_nic *other, *next;
1174 
1175 	if (efx->primary == efx) {
1176 		/* Adding primary function; look for secondaries */
1177 
1178 		netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1179 		list_add_tail(&efx->node, &efx_primary_list);
1180 
1181 		list_for_each_entry_safe(other, next, &efx_unassociated_list,
1182 					 node) {
1183 			if (efx_same_controller(efx, other)) {
1184 				list_del(&other->node);
1185 				netif_dbg(other, probe, other->net_dev,
1186 					  "moving to secondary list of %s %s\n",
1187 					  pci_name(efx->pci_dev),
1188 					  efx->net_dev->name);
1189 				list_add_tail(&other->node,
1190 					      &efx->secondary_list);
1191 				other->primary = efx;
1192 			}
1193 		}
1194 	} else {
1195 		/* Adding secondary function; look for primary */
1196 
1197 		list_for_each_entry(other, &efx_primary_list, node) {
1198 			if (efx_same_controller(efx, other)) {
1199 				netif_dbg(efx, probe, efx->net_dev,
1200 					  "adding to secondary list of %s %s\n",
1201 					  pci_name(other->pci_dev),
1202 					  other->net_dev->name);
1203 				list_add_tail(&efx->node,
1204 					      &other->secondary_list);
1205 				efx->primary = other;
1206 				return;
1207 			}
1208 		}
1209 
1210 		netif_dbg(efx, probe, efx->net_dev,
1211 			  "adding to unassociated list\n");
1212 		list_add_tail(&efx->node, &efx_unassociated_list);
1213 	}
1214 }
1215 
1216 static void efx_dissociate(struct efx_nic *efx)
1217 {
1218 	struct efx_nic *other, *next;
1219 
1220 	list_del(&efx->node);
1221 	efx->primary = NULL;
1222 
1223 	list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1224 		list_del(&other->node);
1225 		netif_dbg(other, probe, other->net_dev,
1226 			  "moving to unassociated list\n");
1227 		list_add_tail(&other->node, &efx_unassociated_list);
1228 		other->primary = NULL;
1229 	}
1230 }
1231 
1232 /* This configures the PCI device to enable I/O and DMA. */
1233 static int efx_init_io(struct efx_nic *efx)
1234 {
1235 	struct pci_dev *pci_dev = efx->pci_dev;
1236 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1237 	unsigned int mem_map_size = efx->type->mem_map_size(efx);
1238 	int rc, bar;
1239 
1240 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1241 
1242 	bar = efx->type->mem_bar;
1243 
1244 	rc = pci_enable_device(pci_dev);
1245 	if (rc) {
1246 		netif_err(efx, probe, efx->net_dev,
1247 			  "failed to enable PCI device\n");
1248 		goto fail1;
1249 	}
1250 
1251 	pci_set_master(pci_dev);
1252 
1253 	/* Set the PCI DMA mask.  Try all possibilities from our
1254 	 * genuine mask down to 32 bits, because some architectures
1255 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1256 	 * masks event though they reject 46 bit masks.
1257 	 */
1258 	while (dma_mask > 0x7fffffffUL) {
1259 		rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1260 		if (rc == 0)
1261 			break;
1262 		dma_mask >>= 1;
1263 	}
1264 	if (rc) {
1265 		netif_err(efx, probe, efx->net_dev,
1266 			  "could not find a suitable DMA mask\n");
1267 		goto fail2;
1268 	}
1269 	netif_dbg(efx, probe, efx->net_dev,
1270 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1271 
1272 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1273 	rc = pci_request_region(pci_dev, bar, "sfc");
1274 	if (rc) {
1275 		netif_err(efx, probe, efx->net_dev,
1276 			  "request for memory BAR failed\n");
1277 		rc = -EIO;
1278 		goto fail3;
1279 	}
1280 	efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1281 	if (!efx->membase) {
1282 		netif_err(efx, probe, efx->net_dev,
1283 			  "could not map memory BAR at %llx+%x\n",
1284 			  (unsigned long long)efx->membase_phys, mem_map_size);
1285 		rc = -ENOMEM;
1286 		goto fail4;
1287 	}
1288 	netif_dbg(efx, probe, efx->net_dev,
1289 		  "memory BAR at %llx+%x (virtual %p)\n",
1290 		  (unsigned long long)efx->membase_phys, mem_map_size,
1291 		  efx->membase);
1292 
1293 	return 0;
1294 
1295  fail4:
1296 	pci_release_region(efx->pci_dev, bar);
1297  fail3:
1298 	efx->membase_phys = 0;
1299  fail2:
1300 	pci_disable_device(efx->pci_dev);
1301  fail1:
1302 	return rc;
1303 }
1304 
1305 static void efx_fini_io(struct efx_nic *efx)
1306 {
1307 	int bar;
1308 
1309 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1310 
1311 	if (efx->membase) {
1312 		iounmap(efx->membase);
1313 		efx->membase = NULL;
1314 	}
1315 
1316 	if (efx->membase_phys) {
1317 		bar = efx->type->mem_bar;
1318 		pci_release_region(efx->pci_dev, bar);
1319 		efx->membase_phys = 0;
1320 	}
1321 
1322 	/* Don't disable bus-mastering if VFs are assigned */
1323 	if (!pci_vfs_assigned(efx->pci_dev))
1324 		pci_disable_device(efx->pci_dev);
1325 }
1326 
1327 void efx_set_default_rx_indir_table(struct efx_nic *efx)
1328 {
1329 	size_t i;
1330 
1331 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1332 		efx->rx_indir_table[i] =
1333 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1334 }
1335 
1336 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1337 {
1338 	cpumask_var_t thread_mask;
1339 	unsigned int count;
1340 	int cpu;
1341 
1342 	if (rss_cpus) {
1343 		count = rss_cpus;
1344 	} else {
1345 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1346 			netif_warn(efx, probe, efx->net_dev,
1347 				   "RSS disabled due to allocation failure\n");
1348 			return 1;
1349 		}
1350 
1351 		count = 0;
1352 		for_each_online_cpu(cpu) {
1353 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1354 				++count;
1355 				cpumask_or(thread_mask, thread_mask,
1356 					   topology_sibling_cpumask(cpu));
1357 			}
1358 		}
1359 
1360 		free_cpumask_var(thread_mask);
1361 	}
1362 
1363 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1364 	 * table entries that are inaccessible to VFs
1365 	 */
1366 #ifdef CONFIG_SFC_SRIOV
1367 	if (efx->type->sriov_wanted) {
1368 		if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1369 		    count > efx_vf_size(efx)) {
1370 			netif_warn(efx, probe, efx->net_dev,
1371 				   "Reducing number of RSS channels from %u to %u for "
1372 				   "VF support. Increase vf-msix-limit to use more "
1373 				   "channels on the PF.\n",
1374 				   count, efx_vf_size(efx));
1375 			count = efx_vf_size(efx);
1376 		}
1377 	}
1378 #endif
1379 
1380 	return count;
1381 }
1382 
1383 /* Probe the number and type of interrupts we are able to obtain, and
1384  * the resulting numbers of channels and RX queues.
1385  */
1386 static int efx_probe_interrupts(struct efx_nic *efx)
1387 {
1388 	unsigned int extra_channels = 0;
1389 	unsigned int i, j;
1390 	int rc;
1391 
1392 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1393 		if (efx->extra_channel_type[i])
1394 			++extra_channels;
1395 
1396 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1397 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1398 		unsigned int n_channels;
1399 
1400 		n_channels = efx_wanted_parallelism(efx);
1401 		if (efx_separate_tx_channels)
1402 			n_channels *= 2;
1403 		n_channels += extra_channels;
1404 		n_channels = min(n_channels, efx->max_channels);
1405 
1406 		for (i = 0; i < n_channels; i++)
1407 			xentries[i].entry = i;
1408 		rc = pci_enable_msix_range(efx->pci_dev,
1409 					   xentries, 1, n_channels);
1410 		if (rc < 0) {
1411 			/* Fall back to single channel MSI */
1412 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1413 			netif_err(efx, drv, efx->net_dev,
1414 				  "could not enable MSI-X\n");
1415 		} else if (rc < n_channels) {
1416 			netif_err(efx, drv, efx->net_dev,
1417 				  "WARNING: Insufficient MSI-X vectors"
1418 				  " available (%d < %u).\n", rc, n_channels);
1419 			netif_err(efx, drv, efx->net_dev,
1420 				  "WARNING: Performance may be reduced.\n");
1421 			n_channels = rc;
1422 		}
1423 
1424 		if (rc > 0) {
1425 			efx->n_channels = n_channels;
1426 			if (n_channels > extra_channels)
1427 				n_channels -= extra_channels;
1428 			if (efx_separate_tx_channels) {
1429 				efx->n_tx_channels = min(max(n_channels / 2,
1430 							     1U),
1431 							 efx->max_tx_channels);
1432 				efx->n_rx_channels = max(n_channels -
1433 							 efx->n_tx_channels,
1434 							 1U);
1435 			} else {
1436 				efx->n_tx_channels = min(n_channels,
1437 							 efx->max_tx_channels);
1438 				efx->n_rx_channels = n_channels;
1439 			}
1440 			for (i = 0; i < efx->n_channels; i++)
1441 				efx_get_channel(efx, i)->irq =
1442 					xentries[i].vector;
1443 		}
1444 	}
1445 
1446 	/* Try single interrupt MSI */
1447 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1448 		efx->n_channels = 1;
1449 		efx->n_rx_channels = 1;
1450 		efx->n_tx_channels = 1;
1451 		rc = pci_enable_msi(efx->pci_dev);
1452 		if (rc == 0) {
1453 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1454 		} else {
1455 			netif_err(efx, drv, efx->net_dev,
1456 				  "could not enable MSI\n");
1457 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1458 		}
1459 	}
1460 
1461 	/* Assume legacy interrupts */
1462 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1463 		efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
1464 		efx->n_rx_channels = 1;
1465 		efx->n_tx_channels = 1;
1466 		efx->legacy_irq = efx->pci_dev->irq;
1467 	}
1468 
1469 	/* Assign extra channels if possible */
1470 	j = efx->n_channels;
1471 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1472 		if (!efx->extra_channel_type[i])
1473 			continue;
1474 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1475 		    efx->n_channels <= extra_channels) {
1476 			efx->extra_channel_type[i]->handle_no_channel(efx);
1477 		} else {
1478 			--j;
1479 			efx_get_channel(efx, j)->type =
1480 				efx->extra_channel_type[i];
1481 		}
1482 	}
1483 
1484 	/* RSS might be usable on VFs even if it is disabled on the PF */
1485 #ifdef CONFIG_SFC_SRIOV
1486 	if (efx->type->sriov_wanted) {
1487 		efx->rss_spread = ((efx->n_rx_channels > 1 ||
1488 				    !efx->type->sriov_wanted(efx)) ?
1489 				   efx->n_rx_channels : efx_vf_size(efx));
1490 		return 0;
1491 	}
1492 #endif
1493 	efx->rss_spread = efx->n_rx_channels;
1494 
1495 	return 0;
1496 }
1497 
1498 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1499 {
1500 	struct efx_channel *channel, *end_channel;
1501 	int rc;
1502 
1503 	BUG_ON(efx->state == STATE_DISABLED);
1504 
1505 	efx->irq_soft_enabled = true;
1506 	smp_wmb();
1507 
1508 	efx_for_each_channel(channel, efx) {
1509 		if (!channel->type->keep_eventq) {
1510 			rc = efx_init_eventq(channel);
1511 			if (rc)
1512 				goto fail;
1513 		}
1514 		efx_start_eventq(channel);
1515 	}
1516 
1517 	efx_mcdi_mode_event(efx);
1518 
1519 	return 0;
1520 fail:
1521 	end_channel = channel;
1522 	efx_for_each_channel(channel, efx) {
1523 		if (channel == end_channel)
1524 			break;
1525 		efx_stop_eventq(channel);
1526 		if (!channel->type->keep_eventq)
1527 			efx_fini_eventq(channel);
1528 	}
1529 
1530 	return rc;
1531 }
1532 
1533 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1534 {
1535 	struct efx_channel *channel;
1536 
1537 	if (efx->state == STATE_DISABLED)
1538 		return;
1539 
1540 	efx_mcdi_mode_poll(efx);
1541 
1542 	efx->irq_soft_enabled = false;
1543 	smp_wmb();
1544 
1545 	if (efx->legacy_irq)
1546 		synchronize_irq(efx->legacy_irq);
1547 
1548 	efx_for_each_channel(channel, efx) {
1549 		if (channel->irq)
1550 			synchronize_irq(channel->irq);
1551 
1552 		efx_stop_eventq(channel);
1553 		if (!channel->type->keep_eventq)
1554 			efx_fini_eventq(channel);
1555 	}
1556 
1557 	/* Flush the asynchronous MCDI request queue */
1558 	efx_mcdi_flush_async(efx);
1559 }
1560 
1561 static int efx_enable_interrupts(struct efx_nic *efx)
1562 {
1563 	struct efx_channel *channel, *end_channel;
1564 	int rc;
1565 
1566 	BUG_ON(efx->state == STATE_DISABLED);
1567 
1568 	if (efx->eeh_disabled_legacy_irq) {
1569 		enable_irq(efx->legacy_irq);
1570 		efx->eeh_disabled_legacy_irq = false;
1571 	}
1572 
1573 	efx->type->irq_enable_master(efx);
1574 
1575 	efx_for_each_channel(channel, efx) {
1576 		if (channel->type->keep_eventq) {
1577 			rc = efx_init_eventq(channel);
1578 			if (rc)
1579 				goto fail;
1580 		}
1581 	}
1582 
1583 	rc = efx_soft_enable_interrupts(efx);
1584 	if (rc)
1585 		goto fail;
1586 
1587 	return 0;
1588 
1589 fail:
1590 	end_channel = channel;
1591 	efx_for_each_channel(channel, efx) {
1592 		if (channel == end_channel)
1593 			break;
1594 		if (channel->type->keep_eventq)
1595 			efx_fini_eventq(channel);
1596 	}
1597 
1598 	efx->type->irq_disable_non_ev(efx);
1599 
1600 	return rc;
1601 }
1602 
1603 static void efx_disable_interrupts(struct efx_nic *efx)
1604 {
1605 	struct efx_channel *channel;
1606 
1607 	efx_soft_disable_interrupts(efx);
1608 
1609 	efx_for_each_channel(channel, efx) {
1610 		if (channel->type->keep_eventq)
1611 			efx_fini_eventq(channel);
1612 	}
1613 
1614 	efx->type->irq_disable_non_ev(efx);
1615 }
1616 
1617 static void efx_remove_interrupts(struct efx_nic *efx)
1618 {
1619 	struct efx_channel *channel;
1620 
1621 	/* Remove MSI/MSI-X interrupts */
1622 	efx_for_each_channel(channel, efx)
1623 		channel->irq = 0;
1624 	pci_disable_msi(efx->pci_dev);
1625 	pci_disable_msix(efx->pci_dev);
1626 
1627 	/* Remove legacy interrupt */
1628 	efx->legacy_irq = 0;
1629 }
1630 
1631 static void efx_set_channels(struct efx_nic *efx)
1632 {
1633 	struct efx_channel *channel;
1634 	struct efx_tx_queue *tx_queue;
1635 
1636 	efx->tx_channel_offset =
1637 		efx_separate_tx_channels ?
1638 		efx->n_channels - efx->n_tx_channels : 0;
1639 
1640 	/* We need to mark which channels really have RX and TX
1641 	 * queues, and adjust the TX queue numbers if we have separate
1642 	 * RX-only and TX-only channels.
1643 	 */
1644 	efx_for_each_channel(channel, efx) {
1645 		if (channel->channel < efx->n_rx_channels)
1646 			channel->rx_queue.core_index = channel->channel;
1647 		else
1648 			channel->rx_queue.core_index = -1;
1649 
1650 		efx_for_each_channel_tx_queue(tx_queue, channel)
1651 			tx_queue->queue -= (efx->tx_channel_offset *
1652 					    EFX_TXQ_TYPES);
1653 	}
1654 }
1655 
1656 static int efx_probe_nic(struct efx_nic *efx)
1657 {
1658 	int rc;
1659 
1660 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1661 
1662 	/* Carry out hardware-type specific initialisation */
1663 	rc = efx->type->probe(efx);
1664 	if (rc)
1665 		return rc;
1666 
1667 	do {
1668 		if (!efx->max_channels || !efx->max_tx_channels) {
1669 			netif_err(efx, drv, efx->net_dev,
1670 				  "Insufficient resources to allocate"
1671 				  " any channels\n");
1672 			rc = -ENOSPC;
1673 			goto fail1;
1674 		}
1675 
1676 		/* Determine the number of channels and queues by trying
1677 		 * to hook in MSI-X interrupts.
1678 		 */
1679 		rc = efx_probe_interrupts(efx);
1680 		if (rc)
1681 			goto fail1;
1682 
1683 		efx_set_channels(efx);
1684 
1685 		/* dimension_resources can fail with EAGAIN */
1686 		rc = efx->type->dimension_resources(efx);
1687 		if (rc != 0 && rc != -EAGAIN)
1688 			goto fail2;
1689 
1690 		if (rc == -EAGAIN)
1691 			/* try again with new max_channels */
1692 			efx_remove_interrupts(efx);
1693 
1694 	} while (rc == -EAGAIN);
1695 
1696 	if (efx->n_channels > 1)
1697 		netdev_rss_key_fill(&efx->rx_hash_key,
1698 				    sizeof(efx->rx_hash_key));
1699 	efx_set_default_rx_indir_table(efx);
1700 
1701 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1702 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1703 
1704 	/* Initialise the interrupt moderation settings */
1705 	efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
1706 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1707 				true);
1708 
1709 	return 0;
1710 
1711 fail2:
1712 	efx_remove_interrupts(efx);
1713 fail1:
1714 	efx->type->remove(efx);
1715 	return rc;
1716 }
1717 
1718 static void efx_remove_nic(struct efx_nic *efx)
1719 {
1720 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1721 
1722 	efx_remove_interrupts(efx);
1723 	efx->type->remove(efx);
1724 }
1725 
1726 static int efx_probe_filters(struct efx_nic *efx)
1727 {
1728 	int rc;
1729 
1730 	spin_lock_init(&efx->filter_lock);
1731 	init_rwsem(&efx->filter_sem);
1732 	mutex_lock(&efx->mac_lock);
1733 	down_write(&efx->filter_sem);
1734 	rc = efx->type->filter_table_probe(efx);
1735 	if (rc)
1736 		goto out_unlock;
1737 
1738 #ifdef CONFIG_RFS_ACCEL
1739 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
1740 		struct efx_channel *channel;
1741 		int i, success = 1;
1742 
1743 		efx_for_each_channel(channel, efx) {
1744 			channel->rps_flow_id =
1745 				kcalloc(efx->type->max_rx_ip_filters,
1746 					sizeof(*channel->rps_flow_id),
1747 					GFP_KERNEL);
1748 			if (!channel->rps_flow_id)
1749 				success = 0;
1750 			else
1751 				for (i = 0;
1752 				     i < efx->type->max_rx_ip_filters;
1753 				     ++i)
1754 					channel->rps_flow_id[i] =
1755 						RPS_FLOW_ID_INVALID;
1756 		}
1757 
1758 		if (!success) {
1759 			efx_for_each_channel(channel, efx)
1760 				kfree(channel->rps_flow_id);
1761 			efx->type->filter_table_remove(efx);
1762 			rc = -ENOMEM;
1763 			goto out_unlock;
1764 		}
1765 
1766 		efx->rps_expire_index = efx->rps_expire_channel = 0;
1767 	}
1768 #endif
1769 out_unlock:
1770 	up_write(&efx->filter_sem);
1771 	mutex_unlock(&efx->mac_lock);
1772 	return rc;
1773 }
1774 
1775 static void efx_remove_filters(struct efx_nic *efx)
1776 {
1777 #ifdef CONFIG_RFS_ACCEL
1778 	struct efx_channel *channel;
1779 
1780 	efx_for_each_channel(channel, efx)
1781 		kfree(channel->rps_flow_id);
1782 #endif
1783 	down_write(&efx->filter_sem);
1784 	efx->type->filter_table_remove(efx);
1785 	up_write(&efx->filter_sem);
1786 }
1787 
1788 static void efx_restore_filters(struct efx_nic *efx)
1789 {
1790 	down_read(&efx->filter_sem);
1791 	efx->type->filter_table_restore(efx);
1792 	up_read(&efx->filter_sem);
1793 }
1794 
1795 /**************************************************************************
1796  *
1797  * NIC startup/shutdown
1798  *
1799  *************************************************************************/
1800 
1801 static int efx_probe_all(struct efx_nic *efx)
1802 {
1803 	int rc;
1804 
1805 	rc = efx_probe_nic(efx);
1806 	if (rc) {
1807 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1808 		goto fail1;
1809 	}
1810 
1811 	rc = efx_probe_port(efx);
1812 	if (rc) {
1813 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1814 		goto fail2;
1815 	}
1816 
1817 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1818 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1819 		rc = -EINVAL;
1820 		goto fail3;
1821 	}
1822 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1823 
1824 #ifdef CONFIG_SFC_SRIOV
1825 	rc = efx->type->vswitching_probe(efx);
1826 	if (rc) /* not fatal; the PF will still work fine */
1827 		netif_warn(efx, probe, efx->net_dev,
1828 			   "failed to setup vswitching rc=%d;"
1829 			   " VFs may not function\n", rc);
1830 #endif
1831 
1832 	rc = efx_probe_filters(efx);
1833 	if (rc) {
1834 		netif_err(efx, probe, efx->net_dev,
1835 			  "failed to create filter tables\n");
1836 		goto fail4;
1837 	}
1838 
1839 	rc = efx_probe_channels(efx);
1840 	if (rc)
1841 		goto fail5;
1842 
1843 	return 0;
1844 
1845  fail5:
1846 	efx_remove_filters(efx);
1847  fail4:
1848 #ifdef CONFIG_SFC_SRIOV
1849 	efx->type->vswitching_remove(efx);
1850 #endif
1851  fail3:
1852 	efx_remove_port(efx);
1853  fail2:
1854 	efx_remove_nic(efx);
1855  fail1:
1856 	return rc;
1857 }
1858 
1859 /* If the interface is supposed to be running but is not, start
1860  * the hardware and software data path, regular activity for the port
1861  * (MAC statistics, link polling, etc.) and schedule the port to be
1862  * reconfigured.  Interrupts must already be enabled.  This function
1863  * is safe to call multiple times, so long as the NIC is not disabled.
1864  * Requires the RTNL lock.
1865  */
1866 static void efx_start_all(struct efx_nic *efx)
1867 {
1868 	EFX_ASSERT_RESET_SERIALISED(efx);
1869 	BUG_ON(efx->state == STATE_DISABLED);
1870 
1871 	/* Check that it is appropriate to restart the interface. All
1872 	 * of these flags are safe to read under just the rtnl lock */
1873 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
1874 	    efx->reset_pending)
1875 		return;
1876 
1877 	efx_start_port(efx);
1878 	efx_start_datapath(efx);
1879 
1880 	/* Start the hardware monitor if there is one */
1881 	if (efx->type->monitor != NULL)
1882 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1883 				   efx_monitor_interval);
1884 
1885 	/* Link state detection is normally event-driven; we have
1886 	 * to poll now because we could have missed a change
1887 	 */
1888 	mutex_lock(&efx->mac_lock);
1889 	if (efx->phy_op->poll(efx))
1890 		efx_link_status_changed(efx);
1891 	mutex_unlock(&efx->mac_lock);
1892 
1893 	efx->type->start_stats(efx);
1894 	efx->type->pull_stats(efx);
1895 	spin_lock_bh(&efx->stats_lock);
1896 	efx->type->update_stats(efx, NULL, NULL);
1897 	spin_unlock_bh(&efx->stats_lock);
1898 }
1899 
1900 /* Quiesce the hardware and software data path, and regular activity
1901  * for the port without bringing the link down.  Safe to call multiple
1902  * times with the NIC in almost any state, but interrupts should be
1903  * enabled.  Requires the RTNL lock.
1904  */
1905 static void efx_stop_all(struct efx_nic *efx)
1906 {
1907 	EFX_ASSERT_RESET_SERIALISED(efx);
1908 
1909 	/* port_enabled can be read safely under the rtnl lock */
1910 	if (!efx->port_enabled)
1911 		return;
1912 
1913 	/* update stats before we go down so we can accurately count
1914 	 * rx_nodesc_drops
1915 	 */
1916 	efx->type->pull_stats(efx);
1917 	spin_lock_bh(&efx->stats_lock);
1918 	efx->type->update_stats(efx, NULL, NULL);
1919 	spin_unlock_bh(&efx->stats_lock);
1920 	efx->type->stop_stats(efx);
1921 	efx_stop_port(efx);
1922 
1923 	/* Stop the kernel transmit interface.  This is only valid if
1924 	 * the device is stopped or detached; otherwise the watchdog
1925 	 * may fire immediately.
1926 	 */
1927 	WARN_ON(netif_running(efx->net_dev) &&
1928 		netif_device_present(efx->net_dev));
1929 	netif_tx_disable(efx->net_dev);
1930 
1931 	efx_stop_datapath(efx);
1932 }
1933 
1934 static void efx_remove_all(struct efx_nic *efx)
1935 {
1936 	efx_remove_channels(efx);
1937 	efx_remove_filters(efx);
1938 #ifdef CONFIG_SFC_SRIOV
1939 	efx->type->vswitching_remove(efx);
1940 #endif
1941 	efx_remove_port(efx);
1942 	efx_remove_nic(efx);
1943 }
1944 
1945 /**************************************************************************
1946  *
1947  * Interrupt moderation
1948  *
1949  **************************************************************************/
1950 unsigned int efx_usecs_to_ticks(struct efx_nic *efx, unsigned int usecs)
1951 {
1952 	if (usecs == 0)
1953 		return 0;
1954 	if (usecs * 1000 < efx->timer_quantum_ns)
1955 		return 1; /* never round down to 0 */
1956 	return usecs * 1000 / efx->timer_quantum_ns;
1957 }
1958 
1959 unsigned int efx_ticks_to_usecs(struct efx_nic *efx, unsigned int ticks)
1960 {
1961 	/* We must round up when converting ticks to microseconds
1962 	 * because we round down when converting the other way.
1963 	 */
1964 	return DIV_ROUND_UP(ticks * efx->timer_quantum_ns, 1000);
1965 }
1966 
1967 /* Set interrupt moderation parameters */
1968 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1969 			    unsigned int rx_usecs, bool rx_adaptive,
1970 			    bool rx_may_override_tx)
1971 {
1972 	struct efx_channel *channel;
1973 	unsigned int timer_max_us;
1974 
1975 	EFX_ASSERT_RESET_SERIALISED(efx);
1976 
1977 	timer_max_us = efx->timer_max_ns / 1000;
1978 
1979 	if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
1980 		return -EINVAL;
1981 
1982 	if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
1983 	    !rx_may_override_tx) {
1984 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1985 			  "RX and TX IRQ moderation must be equal\n");
1986 		return -EINVAL;
1987 	}
1988 
1989 	efx->irq_rx_adaptive = rx_adaptive;
1990 	efx->irq_rx_moderation_us = rx_usecs;
1991 	efx_for_each_channel(channel, efx) {
1992 		if (efx_channel_has_rx_queue(channel))
1993 			channel->irq_moderation_us = rx_usecs;
1994 		else if (efx_channel_has_tx_queues(channel))
1995 			channel->irq_moderation_us = tx_usecs;
1996 	}
1997 
1998 	return 0;
1999 }
2000 
2001 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
2002 			    unsigned int *rx_usecs, bool *rx_adaptive)
2003 {
2004 	*rx_adaptive = efx->irq_rx_adaptive;
2005 	*rx_usecs = efx->irq_rx_moderation_us;
2006 
2007 	/* If channels are shared between RX and TX, so is IRQ
2008 	 * moderation.  Otherwise, IRQ moderation is the same for all
2009 	 * TX channels and is not adaptive.
2010 	 */
2011 	if (efx->tx_channel_offset == 0) {
2012 		*tx_usecs = *rx_usecs;
2013 	} else {
2014 		struct efx_channel *tx_channel;
2015 
2016 		tx_channel = efx->channel[efx->tx_channel_offset];
2017 		*tx_usecs = tx_channel->irq_moderation_us;
2018 	}
2019 }
2020 
2021 /**************************************************************************
2022  *
2023  * Hardware monitor
2024  *
2025  **************************************************************************/
2026 
2027 /* Run periodically off the general workqueue */
2028 static void efx_monitor(struct work_struct *data)
2029 {
2030 	struct efx_nic *efx = container_of(data, struct efx_nic,
2031 					   monitor_work.work);
2032 
2033 	netif_vdbg(efx, timer, efx->net_dev,
2034 		   "hardware monitor executing on CPU %d\n",
2035 		   raw_smp_processor_id());
2036 	BUG_ON(efx->type->monitor == NULL);
2037 
2038 	/* If the mac_lock is already held then it is likely a port
2039 	 * reconfiguration is already in place, which will likely do
2040 	 * most of the work of monitor() anyway. */
2041 	if (mutex_trylock(&efx->mac_lock)) {
2042 		if (efx->port_enabled)
2043 			efx->type->monitor(efx);
2044 		mutex_unlock(&efx->mac_lock);
2045 	}
2046 
2047 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
2048 			   efx_monitor_interval);
2049 }
2050 
2051 /**************************************************************************
2052  *
2053  * ioctls
2054  *
2055  *************************************************************************/
2056 
2057 /* Net device ioctl
2058  * Context: process, rtnl_lock() held.
2059  */
2060 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
2061 {
2062 	struct efx_nic *efx = netdev_priv(net_dev);
2063 	struct mii_ioctl_data *data = if_mii(ifr);
2064 
2065 	if (cmd == SIOCSHWTSTAMP)
2066 		return efx_ptp_set_ts_config(efx, ifr);
2067 	if (cmd == SIOCGHWTSTAMP)
2068 		return efx_ptp_get_ts_config(efx, ifr);
2069 
2070 	/* Convert phy_id from older PRTAD/DEVAD format */
2071 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2072 	    (data->phy_id & 0xfc00) == 0x0400)
2073 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2074 
2075 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
2076 }
2077 
2078 /**************************************************************************
2079  *
2080  * NAPI interface
2081  *
2082  **************************************************************************/
2083 
2084 static void efx_init_napi_channel(struct efx_channel *channel)
2085 {
2086 	struct efx_nic *efx = channel->efx;
2087 
2088 	channel->napi_dev = efx->net_dev;
2089 	netif_napi_add(channel->napi_dev, &channel->napi_str,
2090 		       efx_poll, napi_weight);
2091 	efx_channel_busy_poll_init(channel);
2092 }
2093 
2094 static void efx_init_napi(struct efx_nic *efx)
2095 {
2096 	struct efx_channel *channel;
2097 
2098 	efx_for_each_channel(channel, efx)
2099 		efx_init_napi_channel(channel);
2100 }
2101 
2102 static void efx_fini_napi_channel(struct efx_channel *channel)
2103 {
2104 	if (channel->napi_dev)
2105 		netif_napi_del(&channel->napi_str);
2106 
2107 	channel->napi_dev = NULL;
2108 }
2109 
2110 static void efx_fini_napi(struct efx_nic *efx)
2111 {
2112 	struct efx_channel *channel;
2113 
2114 	efx_for_each_channel(channel, efx)
2115 		efx_fini_napi_channel(channel);
2116 }
2117 
2118 /**************************************************************************
2119  *
2120  * Kernel netpoll interface
2121  *
2122  *************************************************************************/
2123 
2124 #ifdef CONFIG_NET_POLL_CONTROLLER
2125 
2126 /* Although in the common case interrupts will be disabled, this is not
2127  * guaranteed. However, all our work happens inside the NAPI callback,
2128  * so no locking is required.
2129  */
2130 static void efx_netpoll(struct net_device *net_dev)
2131 {
2132 	struct efx_nic *efx = netdev_priv(net_dev);
2133 	struct efx_channel *channel;
2134 
2135 	efx_for_each_channel(channel, efx)
2136 		efx_schedule_channel(channel);
2137 }
2138 
2139 #endif
2140 
2141 #ifdef CONFIG_NET_RX_BUSY_POLL
2142 static int efx_busy_poll(struct napi_struct *napi)
2143 {
2144 	struct efx_channel *channel =
2145 		container_of(napi, struct efx_channel, napi_str);
2146 	struct efx_nic *efx = channel->efx;
2147 	int budget = 4;
2148 	int old_rx_packets, rx_packets;
2149 
2150 	if (!netif_running(efx->net_dev))
2151 		return LL_FLUSH_FAILED;
2152 
2153 	if (!efx_channel_try_lock_poll(channel))
2154 		return LL_FLUSH_BUSY;
2155 
2156 	old_rx_packets = channel->rx_queue.rx_packets;
2157 	efx_process_channel(channel, budget);
2158 
2159 	rx_packets = channel->rx_queue.rx_packets - old_rx_packets;
2160 
2161 	/* There is no race condition with NAPI here.
2162 	 * NAPI will automatically be rescheduled if it yielded during busy
2163 	 * polling, because it was not able to take the lock and thus returned
2164 	 * the full budget.
2165 	 */
2166 	efx_channel_unlock_poll(channel);
2167 
2168 	return rx_packets;
2169 }
2170 #endif
2171 
2172 /**************************************************************************
2173  *
2174  * Kernel net device interface
2175  *
2176  *************************************************************************/
2177 
2178 /* Context: process, rtnl_lock() held. */
2179 int efx_net_open(struct net_device *net_dev)
2180 {
2181 	struct efx_nic *efx = netdev_priv(net_dev);
2182 	int rc;
2183 
2184 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2185 		  raw_smp_processor_id());
2186 
2187 	rc = efx_check_disabled(efx);
2188 	if (rc)
2189 		return rc;
2190 	if (efx->phy_mode & PHY_MODE_SPECIAL)
2191 		return -EBUSY;
2192 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2193 		return -EIO;
2194 
2195 	/* Notify the kernel of the link state polled during driver load,
2196 	 * before the monitor starts running */
2197 	efx_link_status_changed(efx);
2198 
2199 	efx_start_all(efx);
2200 	efx_selftest_async_start(efx);
2201 	return 0;
2202 }
2203 
2204 /* Context: process, rtnl_lock() held.
2205  * Note that the kernel will ignore our return code; this method
2206  * should really be a void.
2207  */
2208 int efx_net_stop(struct net_device *net_dev)
2209 {
2210 	struct efx_nic *efx = netdev_priv(net_dev);
2211 
2212 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2213 		  raw_smp_processor_id());
2214 
2215 	/* Stop the device and flush all the channels */
2216 	efx_stop_all(efx);
2217 
2218 	return 0;
2219 }
2220 
2221 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2222 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
2223 					       struct rtnl_link_stats64 *stats)
2224 {
2225 	struct efx_nic *efx = netdev_priv(net_dev);
2226 
2227 	spin_lock_bh(&efx->stats_lock);
2228 	efx->type->update_stats(efx, NULL, stats);
2229 	spin_unlock_bh(&efx->stats_lock);
2230 
2231 	return stats;
2232 }
2233 
2234 /* Context: netif_tx_lock held, BHs disabled. */
2235 static void efx_watchdog(struct net_device *net_dev)
2236 {
2237 	struct efx_nic *efx = netdev_priv(net_dev);
2238 
2239 	netif_err(efx, tx_err, efx->net_dev,
2240 		  "TX stuck with port_enabled=%d: resetting channels\n",
2241 		  efx->port_enabled);
2242 
2243 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2244 }
2245 
2246 
2247 /* Context: process, rtnl_lock() held. */
2248 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2249 {
2250 	struct efx_nic *efx = netdev_priv(net_dev);
2251 	int rc;
2252 
2253 	rc = efx_check_disabled(efx);
2254 	if (rc)
2255 		return rc;
2256 
2257 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2258 
2259 	efx_device_detach_sync(efx);
2260 	efx_stop_all(efx);
2261 
2262 	mutex_lock(&efx->mac_lock);
2263 	net_dev->mtu = new_mtu;
2264 	efx_mac_reconfigure(efx);
2265 	mutex_unlock(&efx->mac_lock);
2266 
2267 	efx_start_all(efx);
2268 	netif_device_attach(efx->net_dev);
2269 	return 0;
2270 }
2271 
2272 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2273 {
2274 	struct efx_nic *efx = netdev_priv(net_dev);
2275 	struct sockaddr *addr = data;
2276 	u8 *new_addr = addr->sa_data;
2277 	u8 old_addr[6];
2278 	int rc;
2279 
2280 	if (!is_valid_ether_addr(new_addr)) {
2281 		netif_err(efx, drv, efx->net_dev,
2282 			  "invalid ethernet MAC address requested: %pM\n",
2283 			  new_addr);
2284 		return -EADDRNOTAVAIL;
2285 	}
2286 
2287 	/* save old address */
2288 	ether_addr_copy(old_addr, net_dev->dev_addr);
2289 	ether_addr_copy(net_dev->dev_addr, new_addr);
2290 	if (efx->type->set_mac_address) {
2291 		rc = efx->type->set_mac_address(efx);
2292 		if (rc) {
2293 			ether_addr_copy(net_dev->dev_addr, old_addr);
2294 			return rc;
2295 		}
2296 	}
2297 
2298 	/* Reconfigure the MAC */
2299 	mutex_lock(&efx->mac_lock);
2300 	efx_mac_reconfigure(efx);
2301 	mutex_unlock(&efx->mac_lock);
2302 
2303 	return 0;
2304 }
2305 
2306 /* Context: netif_addr_lock held, BHs disabled. */
2307 static void efx_set_rx_mode(struct net_device *net_dev)
2308 {
2309 	struct efx_nic *efx = netdev_priv(net_dev);
2310 
2311 	if (efx->port_enabled)
2312 		queue_work(efx->workqueue, &efx->mac_work);
2313 	/* Otherwise efx_start_port() will do this */
2314 }
2315 
2316 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2317 {
2318 	struct efx_nic *efx = netdev_priv(net_dev);
2319 	int rc;
2320 
2321 	/* If disabling RX n-tuple filtering, clear existing filters */
2322 	if (net_dev->features & ~data & NETIF_F_NTUPLE) {
2323 		rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2324 		if (rc)
2325 			return rc;
2326 	}
2327 
2328 	/* If Rx VLAN filter is changed, update filters via mac_reconfigure */
2329 	if ((net_dev->features ^ data) & NETIF_F_HW_VLAN_CTAG_FILTER) {
2330 		/* efx_set_rx_mode() will schedule MAC work to update filters
2331 		 * when a new features are finally set in net_dev.
2332 		 */
2333 		efx_set_rx_mode(net_dev);
2334 	}
2335 
2336 	return 0;
2337 }
2338 
2339 static int efx_vlan_rx_add_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2340 {
2341 	struct efx_nic *efx = netdev_priv(net_dev);
2342 
2343 	if (efx->type->vlan_rx_add_vid)
2344 		return efx->type->vlan_rx_add_vid(efx, proto, vid);
2345 	else
2346 		return -EOPNOTSUPP;
2347 }
2348 
2349 static int efx_vlan_rx_kill_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2350 {
2351 	struct efx_nic *efx = netdev_priv(net_dev);
2352 
2353 	if (efx->type->vlan_rx_kill_vid)
2354 		return efx->type->vlan_rx_kill_vid(efx, proto, vid);
2355 	else
2356 		return -EOPNOTSUPP;
2357 }
2358 
2359 static const struct net_device_ops efx_netdev_ops = {
2360 	.ndo_open		= efx_net_open,
2361 	.ndo_stop		= efx_net_stop,
2362 	.ndo_get_stats64	= efx_net_stats,
2363 	.ndo_tx_timeout		= efx_watchdog,
2364 	.ndo_start_xmit		= efx_hard_start_xmit,
2365 	.ndo_validate_addr	= eth_validate_addr,
2366 	.ndo_do_ioctl		= efx_ioctl,
2367 	.ndo_change_mtu		= efx_change_mtu,
2368 	.ndo_set_mac_address	= efx_set_mac_address,
2369 	.ndo_set_rx_mode	= efx_set_rx_mode,
2370 	.ndo_set_features	= efx_set_features,
2371 	.ndo_vlan_rx_add_vid	= efx_vlan_rx_add_vid,
2372 	.ndo_vlan_rx_kill_vid	= efx_vlan_rx_kill_vid,
2373 #ifdef CONFIG_SFC_SRIOV
2374 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2375 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2376 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2377 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2378 	.ndo_set_vf_link_state  = efx_sriov_set_vf_link_state,
2379 	.ndo_get_phys_port_id   = efx_sriov_get_phys_port_id,
2380 #endif
2381 #ifdef CONFIG_NET_POLL_CONTROLLER
2382 	.ndo_poll_controller = efx_netpoll,
2383 #endif
2384 	.ndo_setup_tc		= efx_setup_tc,
2385 #ifdef CONFIG_NET_RX_BUSY_POLL
2386 	.ndo_busy_poll		= efx_busy_poll,
2387 #endif
2388 #ifdef CONFIG_RFS_ACCEL
2389 	.ndo_rx_flow_steer	= efx_filter_rfs,
2390 #endif
2391 };
2392 
2393 static void efx_update_name(struct efx_nic *efx)
2394 {
2395 	strcpy(efx->name, efx->net_dev->name);
2396 	efx_mtd_rename(efx);
2397 	efx_set_channel_names(efx);
2398 }
2399 
2400 static int efx_netdev_event(struct notifier_block *this,
2401 			    unsigned long event, void *ptr)
2402 {
2403 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2404 
2405 	if ((net_dev->netdev_ops == &efx_netdev_ops) &&
2406 	    event == NETDEV_CHANGENAME)
2407 		efx_update_name(netdev_priv(net_dev));
2408 
2409 	return NOTIFY_DONE;
2410 }
2411 
2412 static struct notifier_block efx_netdev_notifier = {
2413 	.notifier_call = efx_netdev_event,
2414 };
2415 
2416 static ssize_t
2417 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2418 {
2419 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2420 	return sprintf(buf, "%d\n", efx->phy_type);
2421 }
2422 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2423 
2424 #ifdef CONFIG_SFC_MCDI_LOGGING
2425 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
2426 			     char *buf)
2427 {
2428 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2429 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2430 
2431 	return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
2432 }
2433 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
2434 			    const char *buf, size_t count)
2435 {
2436 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2437 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2438 	bool enable = count > 0 && *buf != '0';
2439 
2440 	mcdi->logging_enabled = enable;
2441 	return count;
2442 }
2443 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
2444 #endif
2445 
2446 static int efx_register_netdev(struct efx_nic *efx)
2447 {
2448 	struct net_device *net_dev = efx->net_dev;
2449 	struct efx_channel *channel;
2450 	int rc;
2451 
2452 	net_dev->watchdog_timeo = 5 * HZ;
2453 	net_dev->irq = efx->pci_dev->irq;
2454 	net_dev->netdev_ops = &efx_netdev_ops;
2455 	if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
2456 		net_dev->priv_flags |= IFF_UNICAST_FLT;
2457 	net_dev->ethtool_ops = &efx_ethtool_ops;
2458 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2459 	net_dev->min_mtu = EFX_MIN_MTU;
2460 	net_dev->max_mtu = EFX_MAX_MTU;
2461 
2462 	rtnl_lock();
2463 
2464 	/* Enable resets to be scheduled and check whether any were
2465 	 * already requested.  If so, the NIC is probably hosed so we
2466 	 * abort.
2467 	 */
2468 	efx->state = STATE_READY;
2469 	smp_mb(); /* ensure we change state before checking reset_pending */
2470 	if (efx->reset_pending) {
2471 		netif_err(efx, probe, efx->net_dev,
2472 			  "aborting probe due to scheduled reset\n");
2473 		rc = -EIO;
2474 		goto fail_locked;
2475 	}
2476 
2477 	rc = dev_alloc_name(net_dev, net_dev->name);
2478 	if (rc < 0)
2479 		goto fail_locked;
2480 	efx_update_name(efx);
2481 
2482 	/* Always start with carrier off; PHY events will detect the link */
2483 	netif_carrier_off(net_dev);
2484 
2485 	rc = register_netdevice(net_dev);
2486 	if (rc)
2487 		goto fail_locked;
2488 
2489 	efx_for_each_channel(channel, efx) {
2490 		struct efx_tx_queue *tx_queue;
2491 		efx_for_each_channel_tx_queue(tx_queue, channel)
2492 			efx_init_tx_queue_core_txq(tx_queue);
2493 	}
2494 
2495 	efx_associate(efx);
2496 
2497 	rtnl_unlock();
2498 
2499 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2500 	if (rc) {
2501 		netif_err(efx, drv, efx->net_dev,
2502 			  "failed to init net dev attributes\n");
2503 		goto fail_registered;
2504 	}
2505 #ifdef CONFIG_SFC_MCDI_LOGGING
2506 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2507 	if (rc) {
2508 		netif_err(efx, drv, efx->net_dev,
2509 			  "failed to init net dev attributes\n");
2510 		goto fail_attr_mcdi_logging;
2511 	}
2512 #endif
2513 
2514 	return 0;
2515 
2516 #ifdef CONFIG_SFC_MCDI_LOGGING
2517 fail_attr_mcdi_logging:
2518 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2519 #endif
2520 fail_registered:
2521 	rtnl_lock();
2522 	efx_dissociate(efx);
2523 	unregister_netdevice(net_dev);
2524 fail_locked:
2525 	efx->state = STATE_UNINIT;
2526 	rtnl_unlock();
2527 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2528 	return rc;
2529 }
2530 
2531 static void efx_unregister_netdev(struct efx_nic *efx)
2532 {
2533 	if (!efx->net_dev)
2534 		return;
2535 
2536 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2537 
2538 	if (efx_dev_registered(efx)) {
2539 		strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2540 #ifdef CONFIG_SFC_MCDI_LOGGING
2541 		device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2542 #endif
2543 		device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2544 		unregister_netdev(efx->net_dev);
2545 	}
2546 }
2547 
2548 /**************************************************************************
2549  *
2550  * Device reset and suspend
2551  *
2552  **************************************************************************/
2553 
2554 /* Tears down the entire software state and most of the hardware state
2555  * before reset.  */
2556 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2557 {
2558 	EFX_ASSERT_RESET_SERIALISED(efx);
2559 
2560 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2561 		efx->type->prepare_flr(efx);
2562 
2563 	efx_stop_all(efx);
2564 	efx_disable_interrupts(efx);
2565 
2566 	mutex_lock(&efx->mac_lock);
2567 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2568 	    method != RESET_TYPE_DATAPATH)
2569 		efx->phy_op->fini(efx);
2570 	efx->type->fini(efx);
2571 }
2572 
2573 /* This function will always ensure that the locks acquired in
2574  * efx_reset_down() are released. A failure return code indicates
2575  * that we were unable to reinitialise the hardware, and the
2576  * driver should be disabled. If ok is false, then the rx and tx
2577  * engines are not restarted, pending a RESET_DISABLE. */
2578 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2579 {
2580 	int rc;
2581 
2582 	EFX_ASSERT_RESET_SERIALISED(efx);
2583 
2584 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2585 		efx->type->finish_flr(efx);
2586 
2587 	/* Ensure that SRAM is initialised even if we're disabling the device */
2588 	rc = efx->type->init(efx);
2589 	if (rc) {
2590 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2591 		goto fail;
2592 	}
2593 
2594 	if (!ok)
2595 		goto fail;
2596 
2597 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2598 	    method != RESET_TYPE_DATAPATH) {
2599 		rc = efx->phy_op->init(efx);
2600 		if (rc)
2601 			goto fail;
2602 		rc = efx->phy_op->reconfigure(efx);
2603 		if (rc && rc != -EPERM)
2604 			netif_err(efx, drv, efx->net_dev,
2605 				  "could not restore PHY settings\n");
2606 	}
2607 
2608 	rc = efx_enable_interrupts(efx);
2609 	if (rc)
2610 		goto fail;
2611 
2612 #ifdef CONFIG_SFC_SRIOV
2613 	rc = efx->type->vswitching_restore(efx);
2614 	if (rc) /* not fatal; the PF will still work fine */
2615 		netif_warn(efx, probe, efx->net_dev,
2616 			   "failed to restore vswitching rc=%d;"
2617 			   " VFs may not function\n", rc);
2618 #endif
2619 
2620 	down_read(&efx->filter_sem);
2621 	efx_restore_filters(efx);
2622 	up_read(&efx->filter_sem);
2623 	if (efx->type->sriov_reset)
2624 		efx->type->sriov_reset(efx);
2625 
2626 	mutex_unlock(&efx->mac_lock);
2627 
2628 	efx_start_all(efx);
2629 
2630 	return 0;
2631 
2632 fail:
2633 	efx->port_initialized = false;
2634 
2635 	mutex_unlock(&efx->mac_lock);
2636 
2637 	return rc;
2638 }
2639 
2640 /* Reset the NIC using the specified method.  Note that the reset may
2641  * fail, in which case the card will be left in an unusable state.
2642  *
2643  * Caller must hold the rtnl_lock.
2644  */
2645 int efx_reset(struct efx_nic *efx, enum reset_type method)
2646 {
2647 	int rc, rc2;
2648 	bool disabled;
2649 
2650 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2651 		   RESET_TYPE(method));
2652 
2653 	efx_device_detach_sync(efx);
2654 	efx_reset_down(efx, method);
2655 
2656 	rc = efx->type->reset(efx, method);
2657 	if (rc) {
2658 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2659 		goto out;
2660 	}
2661 
2662 	/* Clear flags for the scopes we covered.  We assume the NIC and
2663 	 * driver are now quiescent so that there is no race here.
2664 	 */
2665 	if (method < RESET_TYPE_MAX_METHOD)
2666 		efx->reset_pending &= -(1 << (method + 1));
2667 	else /* it doesn't fit into the well-ordered scope hierarchy */
2668 		__clear_bit(method, &efx->reset_pending);
2669 
2670 	/* Reinitialise bus-mastering, which may have been turned off before
2671 	 * the reset was scheduled. This is still appropriate, even in the
2672 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2673 	 * can respond to requests. */
2674 	pci_set_master(efx->pci_dev);
2675 
2676 out:
2677 	/* Leave device stopped if necessary */
2678 	disabled = rc ||
2679 		method == RESET_TYPE_DISABLE ||
2680 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2681 	rc2 = efx_reset_up(efx, method, !disabled);
2682 	if (rc2) {
2683 		disabled = true;
2684 		if (!rc)
2685 			rc = rc2;
2686 	}
2687 
2688 	if (disabled) {
2689 		dev_close(efx->net_dev);
2690 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2691 		efx->state = STATE_DISABLED;
2692 	} else {
2693 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2694 		netif_device_attach(efx->net_dev);
2695 	}
2696 	return rc;
2697 }
2698 
2699 /* Try recovery mechanisms.
2700  * For now only EEH is supported.
2701  * Returns 0 if the recovery mechanisms are unsuccessful.
2702  * Returns a non-zero value otherwise.
2703  */
2704 int efx_try_recovery(struct efx_nic *efx)
2705 {
2706 #ifdef CONFIG_EEH
2707 	/* A PCI error can occur and not be seen by EEH because nothing
2708 	 * happens on the PCI bus. In this case the driver may fail and
2709 	 * schedule a 'recover or reset', leading to this recovery handler.
2710 	 * Manually call the eeh failure check function.
2711 	 */
2712 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2713 	if (eeh_dev_check_failure(eehdev)) {
2714 		/* The EEH mechanisms will handle the error and reset the
2715 		 * device if necessary.
2716 		 */
2717 		return 1;
2718 	}
2719 #endif
2720 	return 0;
2721 }
2722 
2723 static void efx_wait_for_bist_end(struct efx_nic *efx)
2724 {
2725 	int i;
2726 
2727 	for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2728 		if (efx_mcdi_poll_reboot(efx))
2729 			goto out;
2730 		msleep(BIST_WAIT_DELAY_MS);
2731 	}
2732 
2733 	netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2734 out:
2735 	/* Either way unset the BIST flag. If we found no reboot we probably
2736 	 * won't recover, but we should try.
2737 	 */
2738 	efx->mc_bist_for_other_fn = false;
2739 }
2740 
2741 /* The worker thread exists so that code that cannot sleep can
2742  * schedule a reset for later.
2743  */
2744 static void efx_reset_work(struct work_struct *data)
2745 {
2746 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2747 	unsigned long pending;
2748 	enum reset_type method;
2749 
2750 	pending = ACCESS_ONCE(efx->reset_pending);
2751 	method = fls(pending) - 1;
2752 
2753 	if (method == RESET_TYPE_MC_BIST)
2754 		efx_wait_for_bist_end(efx);
2755 
2756 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2757 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2758 	    efx_try_recovery(efx))
2759 		return;
2760 
2761 	if (!pending)
2762 		return;
2763 
2764 	rtnl_lock();
2765 
2766 	/* We checked the state in efx_schedule_reset() but it may
2767 	 * have changed by now.  Now that we have the RTNL lock,
2768 	 * it cannot change again.
2769 	 */
2770 	if (efx->state == STATE_READY)
2771 		(void)efx_reset(efx, method);
2772 
2773 	rtnl_unlock();
2774 }
2775 
2776 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2777 {
2778 	enum reset_type method;
2779 
2780 	if (efx->state == STATE_RECOVERY) {
2781 		netif_dbg(efx, drv, efx->net_dev,
2782 			  "recovering: skip scheduling %s reset\n",
2783 			  RESET_TYPE(type));
2784 		return;
2785 	}
2786 
2787 	switch (type) {
2788 	case RESET_TYPE_INVISIBLE:
2789 	case RESET_TYPE_ALL:
2790 	case RESET_TYPE_RECOVER_OR_ALL:
2791 	case RESET_TYPE_WORLD:
2792 	case RESET_TYPE_DISABLE:
2793 	case RESET_TYPE_RECOVER_OR_DISABLE:
2794 	case RESET_TYPE_DATAPATH:
2795 	case RESET_TYPE_MC_BIST:
2796 	case RESET_TYPE_MCDI_TIMEOUT:
2797 		method = type;
2798 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2799 			  RESET_TYPE(method));
2800 		break;
2801 	default:
2802 		method = efx->type->map_reset_reason(type);
2803 		netif_dbg(efx, drv, efx->net_dev,
2804 			  "scheduling %s reset for %s\n",
2805 			  RESET_TYPE(method), RESET_TYPE(type));
2806 		break;
2807 	}
2808 
2809 	set_bit(method, &efx->reset_pending);
2810 	smp_mb(); /* ensure we change reset_pending before checking state */
2811 
2812 	/* If we're not READY then just leave the flags set as the cue
2813 	 * to abort probing or reschedule the reset later.
2814 	 */
2815 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2816 		return;
2817 
2818 	/* efx_process_channel() will no longer read events once a
2819 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2820 	efx_mcdi_mode_poll(efx);
2821 
2822 	queue_work(reset_workqueue, &efx->reset_work);
2823 }
2824 
2825 /**************************************************************************
2826  *
2827  * List of NICs we support
2828  *
2829  **************************************************************************/
2830 
2831 /* PCI device ID table */
2832 static const struct pci_device_id efx_pci_table[] = {
2833 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2834 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2835 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2836 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2837 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903),  /* SFC9120 PF */
2838 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2839 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903),  /* SFC9120 VF */
2840 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2841 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923),  /* SFC9140 PF */
2842 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2843 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1923),  /* SFC9140 VF */
2844 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2845 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0a03),  /* SFC9220 PF */
2846 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2847 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1a03),  /* SFC9220 VF */
2848 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2849 	{0}			/* end of list */
2850 };
2851 
2852 /**************************************************************************
2853  *
2854  * Dummy PHY/MAC operations
2855  *
2856  * Can be used for some unimplemented operations
2857  * Needed so all function pointers are valid and do not have to be tested
2858  * before use
2859  *
2860  **************************************************************************/
2861 int efx_port_dummy_op_int(struct efx_nic *efx)
2862 {
2863 	return 0;
2864 }
2865 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2866 
2867 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2868 {
2869 	return false;
2870 }
2871 
2872 static const struct efx_phy_operations efx_dummy_phy_operations = {
2873 	.init		 = efx_port_dummy_op_int,
2874 	.reconfigure	 = efx_port_dummy_op_int,
2875 	.poll		 = efx_port_dummy_op_poll,
2876 	.fini		 = efx_port_dummy_op_void,
2877 };
2878 
2879 /**************************************************************************
2880  *
2881  * Data housekeeping
2882  *
2883  **************************************************************************/
2884 
2885 /* This zeroes out and then fills in the invariants in a struct
2886  * efx_nic (including all sub-structures).
2887  */
2888 static int efx_init_struct(struct efx_nic *efx,
2889 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2890 {
2891 	int i;
2892 
2893 	/* Initialise common structures */
2894 	INIT_LIST_HEAD(&efx->node);
2895 	INIT_LIST_HEAD(&efx->secondary_list);
2896 	spin_lock_init(&efx->biu_lock);
2897 #ifdef CONFIG_SFC_MTD
2898 	INIT_LIST_HEAD(&efx->mtd_list);
2899 #endif
2900 	INIT_WORK(&efx->reset_work, efx_reset_work);
2901 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2902 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2903 	efx->pci_dev = pci_dev;
2904 	efx->msg_enable = debug;
2905 	efx->state = STATE_UNINIT;
2906 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2907 
2908 	efx->net_dev = net_dev;
2909 	efx->rx_prefix_size = efx->type->rx_prefix_size;
2910 	efx->rx_ip_align =
2911 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2912 	efx->rx_packet_hash_offset =
2913 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2914 	efx->rx_packet_ts_offset =
2915 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2916 	spin_lock_init(&efx->stats_lock);
2917 	mutex_init(&efx->mac_lock);
2918 	efx->phy_op = &efx_dummy_phy_operations;
2919 	efx->mdio.dev = net_dev;
2920 	INIT_WORK(&efx->mac_work, efx_mac_work);
2921 	init_waitqueue_head(&efx->flush_wq);
2922 
2923 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2924 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2925 		if (!efx->channel[i])
2926 			goto fail;
2927 		efx->msi_context[i].efx = efx;
2928 		efx->msi_context[i].index = i;
2929 	}
2930 
2931 	/* Higher numbered interrupt modes are less capable! */
2932 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2933 				  interrupt_mode);
2934 
2935 	/* Would be good to use the net_dev name, but we're too early */
2936 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2937 		 pci_name(pci_dev));
2938 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2939 	if (!efx->workqueue)
2940 		goto fail;
2941 
2942 	return 0;
2943 
2944 fail:
2945 	efx_fini_struct(efx);
2946 	return -ENOMEM;
2947 }
2948 
2949 static void efx_fini_struct(struct efx_nic *efx)
2950 {
2951 	int i;
2952 
2953 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2954 		kfree(efx->channel[i]);
2955 
2956 	kfree(efx->vpd_sn);
2957 
2958 	if (efx->workqueue) {
2959 		destroy_workqueue(efx->workqueue);
2960 		efx->workqueue = NULL;
2961 	}
2962 }
2963 
2964 void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
2965 {
2966 	u64 n_rx_nodesc_trunc = 0;
2967 	struct efx_channel *channel;
2968 
2969 	efx_for_each_channel(channel, efx)
2970 		n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
2971 	stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
2972 	stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
2973 }
2974 
2975 /**************************************************************************
2976  *
2977  * PCI interface
2978  *
2979  **************************************************************************/
2980 
2981 /* Main body of final NIC shutdown code
2982  * This is called only at module unload (or hotplug removal).
2983  */
2984 static void efx_pci_remove_main(struct efx_nic *efx)
2985 {
2986 	/* Flush reset_work. It can no longer be scheduled since we
2987 	 * are not READY.
2988 	 */
2989 	BUG_ON(efx->state == STATE_READY);
2990 	cancel_work_sync(&efx->reset_work);
2991 
2992 	efx_disable_interrupts(efx);
2993 	efx_nic_fini_interrupt(efx);
2994 	efx_fini_port(efx);
2995 	efx->type->fini(efx);
2996 	efx_fini_napi(efx);
2997 	efx_remove_all(efx);
2998 }
2999 
3000 /* Final NIC shutdown
3001  * This is called only at module unload (or hotplug removal).  A PF can call
3002  * this on its VFs to ensure they are unbound first.
3003  */
3004 static void efx_pci_remove(struct pci_dev *pci_dev)
3005 {
3006 	struct efx_nic *efx;
3007 
3008 	efx = pci_get_drvdata(pci_dev);
3009 	if (!efx)
3010 		return;
3011 
3012 	/* Mark the NIC as fini, then stop the interface */
3013 	rtnl_lock();
3014 	efx_dissociate(efx);
3015 	dev_close(efx->net_dev);
3016 	efx_disable_interrupts(efx);
3017 	efx->state = STATE_UNINIT;
3018 	rtnl_unlock();
3019 
3020 	if (efx->type->sriov_fini)
3021 		efx->type->sriov_fini(efx);
3022 
3023 	efx_unregister_netdev(efx);
3024 
3025 	efx_mtd_remove(efx);
3026 
3027 	efx_pci_remove_main(efx);
3028 
3029 	efx_fini_io(efx);
3030 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
3031 
3032 	efx_fini_struct(efx);
3033 	free_netdev(efx->net_dev);
3034 
3035 	pci_disable_pcie_error_reporting(pci_dev);
3036 };
3037 
3038 /* NIC VPD information
3039  * Called during probe to display the part number of the
3040  * installed NIC.  VPD is potentially very large but this should
3041  * always appear within the first 512 bytes.
3042  */
3043 #define SFC_VPD_LEN 512
3044 static void efx_probe_vpd_strings(struct efx_nic *efx)
3045 {
3046 	struct pci_dev *dev = efx->pci_dev;
3047 	char vpd_data[SFC_VPD_LEN];
3048 	ssize_t vpd_size;
3049 	int ro_start, ro_size, i, j;
3050 
3051 	/* Get the vpd data from the device */
3052 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
3053 	if (vpd_size <= 0) {
3054 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
3055 		return;
3056 	}
3057 
3058 	/* Get the Read only section */
3059 	ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
3060 	if (ro_start < 0) {
3061 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
3062 		return;
3063 	}
3064 
3065 	ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
3066 	j = ro_size;
3067 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3068 	if (i + j > vpd_size)
3069 		j = vpd_size - i;
3070 
3071 	/* Get the Part number */
3072 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
3073 	if (i < 0) {
3074 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
3075 		return;
3076 	}
3077 
3078 	j = pci_vpd_info_field_size(&vpd_data[i]);
3079 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3080 	if (i + j > vpd_size) {
3081 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
3082 		return;
3083 	}
3084 
3085 	netif_info(efx, drv, efx->net_dev,
3086 		   "Part Number : %.*s\n", j, &vpd_data[i]);
3087 
3088 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3089 	j = ro_size;
3090 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
3091 	if (i < 0) {
3092 		netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
3093 		return;
3094 	}
3095 
3096 	j = pci_vpd_info_field_size(&vpd_data[i]);
3097 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3098 	if (i + j > vpd_size) {
3099 		netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
3100 		return;
3101 	}
3102 
3103 	efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
3104 	if (!efx->vpd_sn)
3105 		return;
3106 
3107 	snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
3108 }
3109 
3110 
3111 /* Main body of NIC initialisation
3112  * This is called at module load (or hotplug insertion, theoretically).
3113  */
3114 static int efx_pci_probe_main(struct efx_nic *efx)
3115 {
3116 	int rc;
3117 
3118 	/* Do start-of-day initialisation */
3119 	rc = efx_probe_all(efx);
3120 	if (rc)
3121 		goto fail1;
3122 
3123 	efx_init_napi(efx);
3124 
3125 	rc = efx->type->init(efx);
3126 	if (rc) {
3127 		netif_err(efx, probe, efx->net_dev,
3128 			  "failed to initialise NIC\n");
3129 		goto fail3;
3130 	}
3131 
3132 	rc = efx_init_port(efx);
3133 	if (rc) {
3134 		netif_err(efx, probe, efx->net_dev,
3135 			  "failed to initialise port\n");
3136 		goto fail4;
3137 	}
3138 
3139 	rc = efx_nic_init_interrupt(efx);
3140 	if (rc)
3141 		goto fail5;
3142 	rc = efx_enable_interrupts(efx);
3143 	if (rc)
3144 		goto fail6;
3145 
3146 	return 0;
3147 
3148  fail6:
3149 	efx_nic_fini_interrupt(efx);
3150  fail5:
3151 	efx_fini_port(efx);
3152  fail4:
3153 	efx->type->fini(efx);
3154  fail3:
3155 	efx_fini_napi(efx);
3156 	efx_remove_all(efx);
3157  fail1:
3158 	return rc;
3159 }
3160 
3161 /* NIC initialisation
3162  *
3163  * This is called at module load (or hotplug insertion,
3164  * theoretically).  It sets up PCI mappings, resets the NIC,
3165  * sets up and registers the network devices with the kernel and hooks
3166  * the interrupt service routine.  It does not prepare the device for
3167  * transmission; this is left to the first time one of the network
3168  * interfaces is brought up (i.e. efx_net_open).
3169  */
3170 static int efx_pci_probe(struct pci_dev *pci_dev,
3171 			 const struct pci_device_id *entry)
3172 {
3173 	struct net_device *net_dev;
3174 	struct efx_nic *efx;
3175 	int rc;
3176 
3177 	/* Allocate and initialise a struct net_device and struct efx_nic */
3178 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
3179 				     EFX_MAX_RX_QUEUES);
3180 	if (!net_dev)
3181 		return -ENOMEM;
3182 	efx = netdev_priv(net_dev);
3183 	efx->type = (const struct efx_nic_type *) entry->driver_data;
3184 	efx->fixed_features |= NETIF_F_HIGHDMA;
3185 
3186 	pci_set_drvdata(pci_dev, efx);
3187 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
3188 	rc = efx_init_struct(efx, pci_dev, net_dev);
3189 	if (rc)
3190 		goto fail1;
3191 
3192 	netif_info(efx, probe, efx->net_dev,
3193 		   "Solarflare NIC detected\n");
3194 
3195 	if (!efx->type->is_vf)
3196 		efx_probe_vpd_strings(efx);
3197 
3198 	/* Set up basic I/O (BAR mappings etc) */
3199 	rc = efx_init_io(efx);
3200 	if (rc)
3201 		goto fail2;
3202 
3203 	rc = efx_pci_probe_main(efx);
3204 	if (rc)
3205 		goto fail3;
3206 
3207 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
3208 			      NETIF_F_TSO | NETIF_F_RXCSUM);
3209 	if (efx->type->offload_features & (NETIF_F_IPV6_CSUM | NETIF_F_HW_CSUM))
3210 		net_dev->features |= NETIF_F_TSO6;
3211 	/* Check whether device supports TSO */
3212 	if (!efx->type->tso_versions || !efx->type->tso_versions(efx))
3213 		net_dev->features &= ~NETIF_F_ALL_TSO;
3214 	/* Mask for features that also apply to VLAN devices */
3215 	net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
3216 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
3217 				   NETIF_F_RXCSUM);
3218 
3219 	net_dev->hw_features = net_dev->features & ~efx->fixed_features;
3220 
3221 	/* Disable VLAN filtering by default.  It may be enforced if
3222 	 * the feature is fixed (i.e. VLAN filters are required to
3223 	 * receive VLAN tagged packets due to vPort restrictions).
3224 	 */
3225 	net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
3226 	net_dev->features |= efx->fixed_features;
3227 
3228 	rc = efx_register_netdev(efx);
3229 	if (rc)
3230 		goto fail4;
3231 
3232 	if (efx->type->sriov_init) {
3233 		rc = efx->type->sriov_init(efx);
3234 		if (rc)
3235 			netif_err(efx, probe, efx->net_dev,
3236 				  "SR-IOV can't be enabled rc %d\n", rc);
3237 	}
3238 
3239 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3240 
3241 	/* Try to create MTDs, but allow this to fail */
3242 	rtnl_lock();
3243 	rc = efx_mtd_probe(efx);
3244 	rtnl_unlock();
3245 	if (rc && rc != -EPERM)
3246 		netif_warn(efx, probe, efx->net_dev,
3247 			   "failed to create MTDs (%d)\n", rc);
3248 
3249 	rc = pci_enable_pcie_error_reporting(pci_dev);
3250 	if (rc && rc != -EINVAL)
3251 		netif_notice(efx, probe, efx->net_dev,
3252 			     "PCIE error reporting unavailable (%d).\n",
3253 			     rc);
3254 
3255 	return 0;
3256 
3257  fail4:
3258 	efx_pci_remove_main(efx);
3259  fail3:
3260 	efx_fini_io(efx);
3261  fail2:
3262 	efx_fini_struct(efx);
3263  fail1:
3264 	WARN_ON(rc > 0);
3265 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3266 	free_netdev(net_dev);
3267 	return rc;
3268 }
3269 
3270 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3271  * enabled on success
3272  */
3273 #ifdef CONFIG_SFC_SRIOV
3274 static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
3275 {
3276 	int rc;
3277 	struct efx_nic *efx = pci_get_drvdata(dev);
3278 
3279 	if (efx->type->sriov_configure) {
3280 		rc = efx->type->sriov_configure(efx, num_vfs);
3281 		if (rc)
3282 			return rc;
3283 		else
3284 			return num_vfs;
3285 	} else
3286 		return -EOPNOTSUPP;
3287 }
3288 #endif
3289 
3290 static int efx_pm_freeze(struct device *dev)
3291 {
3292 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3293 
3294 	rtnl_lock();
3295 
3296 	if (efx->state != STATE_DISABLED) {
3297 		efx->state = STATE_UNINIT;
3298 
3299 		efx_device_detach_sync(efx);
3300 
3301 		efx_stop_all(efx);
3302 		efx_disable_interrupts(efx);
3303 	}
3304 
3305 	rtnl_unlock();
3306 
3307 	return 0;
3308 }
3309 
3310 static int efx_pm_thaw(struct device *dev)
3311 {
3312 	int rc;
3313 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3314 
3315 	rtnl_lock();
3316 
3317 	if (efx->state != STATE_DISABLED) {
3318 		rc = efx_enable_interrupts(efx);
3319 		if (rc)
3320 			goto fail;
3321 
3322 		mutex_lock(&efx->mac_lock);
3323 		efx->phy_op->reconfigure(efx);
3324 		mutex_unlock(&efx->mac_lock);
3325 
3326 		efx_start_all(efx);
3327 
3328 		netif_device_attach(efx->net_dev);
3329 
3330 		efx->state = STATE_READY;
3331 
3332 		efx->type->resume_wol(efx);
3333 	}
3334 
3335 	rtnl_unlock();
3336 
3337 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3338 	queue_work(reset_workqueue, &efx->reset_work);
3339 
3340 	return 0;
3341 
3342 fail:
3343 	rtnl_unlock();
3344 
3345 	return rc;
3346 }
3347 
3348 static int efx_pm_poweroff(struct device *dev)
3349 {
3350 	struct pci_dev *pci_dev = to_pci_dev(dev);
3351 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3352 
3353 	efx->type->fini(efx);
3354 
3355 	efx->reset_pending = 0;
3356 
3357 	pci_save_state(pci_dev);
3358 	return pci_set_power_state(pci_dev, PCI_D3hot);
3359 }
3360 
3361 /* Used for both resume and restore */
3362 static int efx_pm_resume(struct device *dev)
3363 {
3364 	struct pci_dev *pci_dev = to_pci_dev(dev);
3365 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3366 	int rc;
3367 
3368 	rc = pci_set_power_state(pci_dev, PCI_D0);
3369 	if (rc)
3370 		return rc;
3371 	pci_restore_state(pci_dev);
3372 	rc = pci_enable_device(pci_dev);
3373 	if (rc)
3374 		return rc;
3375 	pci_set_master(efx->pci_dev);
3376 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
3377 	if (rc)
3378 		return rc;
3379 	rc = efx->type->init(efx);
3380 	if (rc)
3381 		return rc;
3382 	rc = efx_pm_thaw(dev);
3383 	return rc;
3384 }
3385 
3386 static int efx_pm_suspend(struct device *dev)
3387 {
3388 	int rc;
3389 
3390 	efx_pm_freeze(dev);
3391 	rc = efx_pm_poweroff(dev);
3392 	if (rc)
3393 		efx_pm_resume(dev);
3394 	return rc;
3395 }
3396 
3397 static const struct dev_pm_ops efx_pm_ops = {
3398 	.suspend	= efx_pm_suspend,
3399 	.resume		= efx_pm_resume,
3400 	.freeze		= efx_pm_freeze,
3401 	.thaw		= efx_pm_thaw,
3402 	.poweroff	= efx_pm_poweroff,
3403 	.restore	= efx_pm_resume,
3404 };
3405 
3406 /* A PCI error affecting this device was detected.
3407  * At this point MMIO and DMA may be disabled.
3408  * Stop the software path and request a slot reset.
3409  */
3410 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3411 					      enum pci_channel_state state)
3412 {
3413 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3414 	struct efx_nic *efx = pci_get_drvdata(pdev);
3415 
3416 	if (state == pci_channel_io_perm_failure)
3417 		return PCI_ERS_RESULT_DISCONNECT;
3418 
3419 	rtnl_lock();
3420 
3421 	if (efx->state != STATE_DISABLED) {
3422 		efx->state = STATE_RECOVERY;
3423 		efx->reset_pending = 0;
3424 
3425 		efx_device_detach_sync(efx);
3426 
3427 		efx_stop_all(efx);
3428 		efx_disable_interrupts(efx);
3429 
3430 		status = PCI_ERS_RESULT_NEED_RESET;
3431 	} else {
3432 		/* If the interface is disabled we don't want to do anything
3433 		 * with it.
3434 		 */
3435 		status = PCI_ERS_RESULT_RECOVERED;
3436 	}
3437 
3438 	rtnl_unlock();
3439 
3440 	pci_disable_device(pdev);
3441 
3442 	return status;
3443 }
3444 
3445 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3446 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3447 {
3448 	struct efx_nic *efx = pci_get_drvdata(pdev);
3449 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3450 	int rc;
3451 
3452 	if (pci_enable_device(pdev)) {
3453 		netif_err(efx, hw, efx->net_dev,
3454 			  "Cannot re-enable PCI device after reset.\n");
3455 		status =  PCI_ERS_RESULT_DISCONNECT;
3456 	}
3457 
3458 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3459 	if (rc) {
3460 		netif_err(efx, hw, efx->net_dev,
3461 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3462 		/* Non-fatal error. Continue. */
3463 	}
3464 
3465 	return status;
3466 }
3467 
3468 /* Perform the actual reset and resume I/O operations. */
3469 static void efx_io_resume(struct pci_dev *pdev)
3470 {
3471 	struct efx_nic *efx = pci_get_drvdata(pdev);
3472 	int rc;
3473 
3474 	rtnl_lock();
3475 
3476 	if (efx->state == STATE_DISABLED)
3477 		goto out;
3478 
3479 	rc = efx_reset(efx, RESET_TYPE_ALL);
3480 	if (rc) {
3481 		netif_err(efx, hw, efx->net_dev,
3482 			  "efx_reset failed after PCI error (%d)\n", rc);
3483 	} else {
3484 		efx->state = STATE_READY;
3485 		netif_dbg(efx, hw, efx->net_dev,
3486 			  "Done resetting and resuming IO after PCI error.\n");
3487 	}
3488 
3489 out:
3490 	rtnl_unlock();
3491 }
3492 
3493 /* For simplicity and reliability, we always require a slot reset and try to
3494  * reset the hardware when a pci error affecting the device is detected.
3495  * We leave both the link_reset and mmio_enabled callback unimplemented:
3496  * with our request for slot reset the mmio_enabled callback will never be
3497  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3498  */
3499 static const struct pci_error_handlers efx_err_handlers = {
3500 	.error_detected = efx_io_error_detected,
3501 	.slot_reset	= efx_io_slot_reset,
3502 	.resume		= efx_io_resume,
3503 };
3504 
3505 static struct pci_driver efx_pci_driver = {
3506 	.name		= KBUILD_MODNAME,
3507 	.id_table	= efx_pci_table,
3508 	.probe		= efx_pci_probe,
3509 	.remove		= efx_pci_remove,
3510 	.driver.pm	= &efx_pm_ops,
3511 	.err_handler	= &efx_err_handlers,
3512 #ifdef CONFIG_SFC_SRIOV
3513 	.sriov_configure = efx_pci_sriov_configure,
3514 #endif
3515 };
3516 
3517 /**************************************************************************
3518  *
3519  * Kernel module interface
3520  *
3521  *************************************************************************/
3522 
3523 module_param(interrupt_mode, uint, 0444);
3524 MODULE_PARM_DESC(interrupt_mode,
3525 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3526 
3527 static int __init efx_init_module(void)
3528 {
3529 	int rc;
3530 
3531 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3532 
3533 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3534 	if (rc)
3535 		goto err_notifier;
3536 
3537 #ifdef CONFIG_SFC_SRIOV
3538 	rc = efx_init_sriov();
3539 	if (rc)
3540 		goto err_sriov;
3541 #endif
3542 
3543 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3544 	if (!reset_workqueue) {
3545 		rc = -ENOMEM;
3546 		goto err_reset;
3547 	}
3548 
3549 	rc = pci_register_driver(&efx_pci_driver);
3550 	if (rc < 0)
3551 		goto err_pci;
3552 
3553 	return 0;
3554 
3555  err_pci:
3556 	destroy_workqueue(reset_workqueue);
3557  err_reset:
3558 #ifdef CONFIG_SFC_SRIOV
3559 	efx_fini_sriov();
3560  err_sriov:
3561 #endif
3562 	unregister_netdevice_notifier(&efx_netdev_notifier);
3563  err_notifier:
3564 	return rc;
3565 }
3566 
3567 static void __exit efx_exit_module(void)
3568 {
3569 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3570 
3571 	pci_unregister_driver(&efx_pci_driver);
3572 	destroy_workqueue(reset_workqueue);
3573 #ifdef CONFIG_SFC_SRIOV
3574 	efx_fini_sriov();
3575 #endif
3576 	unregister_netdevice_notifier(&efx_netdev_notifier);
3577 
3578 }
3579 
3580 module_init(efx_init_module);
3581 module_exit(efx_exit_module);
3582 
3583 MODULE_AUTHOR("Solarflare Communications and "
3584 	      "Michael Brown <mbrown@fensystems.co.uk>");
3585 MODULE_DESCRIPTION("Solarflare network driver");
3586 MODULE_LICENSE("GPL");
3587 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3588