xref: /openbmc/linux/drivers/net/ethernet/sfc/efx.c (revision e23feb16)
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 
30 #include "mcdi.h"
31 #include "workarounds.h"
32 
33 /**************************************************************************
34  *
35  * Type name strings
36  *
37  **************************************************************************
38  */
39 
40 /* Loopback mode names (see LOOPBACK_MODE()) */
41 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
42 const char *const efx_loopback_mode_names[] = {
43 	[LOOPBACK_NONE]		= "NONE",
44 	[LOOPBACK_DATA]		= "DATAPATH",
45 	[LOOPBACK_GMAC]		= "GMAC",
46 	[LOOPBACK_XGMII]	= "XGMII",
47 	[LOOPBACK_XGXS]		= "XGXS",
48 	[LOOPBACK_XAUI]		= "XAUI",
49 	[LOOPBACK_GMII]		= "GMII",
50 	[LOOPBACK_SGMII]	= "SGMII",
51 	[LOOPBACK_XGBR]		= "XGBR",
52 	[LOOPBACK_XFI]		= "XFI",
53 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
54 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
55 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
56 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
57 	[LOOPBACK_GPHY]		= "GPHY",
58 	[LOOPBACK_PHYXS]	= "PHYXS",
59 	[LOOPBACK_PCS]		= "PCS",
60 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
61 	[LOOPBACK_XPORT]	= "XPORT",
62 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
63 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
64 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
65 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
66 	[LOOPBACK_GMII_WS]	= "GMII_WS",
67 	[LOOPBACK_XFI_WS]	= "XFI_WS",
68 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
69 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
70 };
71 
72 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
73 const char *const efx_reset_type_names[] = {
74 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
75 	[RESET_TYPE_ALL]                = "ALL",
76 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
77 	[RESET_TYPE_WORLD]              = "WORLD",
78 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
79 	[RESET_TYPE_DISABLE]            = "DISABLE",
80 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
81 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
82 	[RESET_TYPE_RX_RECOVERY]        = "RX_RECOVERY",
83 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
84 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
85 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
86 };
87 
88 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
89  * queued onto this work queue. This is not a per-nic work queue, because
90  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
91  */
92 static struct workqueue_struct *reset_workqueue;
93 
94 /**************************************************************************
95  *
96  * Configurable values
97  *
98  *************************************************************************/
99 
100 /*
101  * Use separate channels for TX and RX events
102  *
103  * Set this to 1 to use separate channels for TX and RX. It allows us
104  * to control interrupt affinity separately for TX and RX.
105  *
106  * This is only used in MSI-X interrupt mode
107  */
108 static bool separate_tx_channels;
109 module_param(separate_tx_channels, bool, 0444);
110 MODULE_PARM_DESC(separate_tx_channels,
111 		 "Use separate channels for TX and RX");
112 
113 /* This is the weight assigned to each of the (per-channel) virtual
114  * NAPI devices.
115  */
116 static int napi_weight = 64;
117 
118 /* This is the time (in jiffies) between invocations of the hardware
119  * monitor.
120  * On Falcon-based NICs, this will:
121  * - Check the on-board hardware monitor;
122  * - Poll the link state and reconfigure the hardware as necessary.
123  * On Siena-based NICs for power systems with EEH support, this will give EEH a
124  * chance to start.
125  */
126 static unsigned int efx_monitor_interval = 1 * HZ;
127 
128 /* Initial interrupt moderation settings.  They can be modified after
129  * module load with ethtool.
130  *
131  * The default for RX should strike a balance between increasing the
132  * round-trip latency and reducing overhead.
133  */
134 static unsigned int rx_irq_mod_usec = 60;
135 
136 /* Initial interrupt moderation settings.  They can be modified after
137  * module load with ethtool.
138  *
139  * This default is chosen to ensure that a 10G link does not go idle
140  * while a TX queue is stopped after it has become full.  A queue is
141  * restarted when it drops below half full.  The time this takes (assuming
142  * worst case 3 descriptors per packet and 1024 descriptors) is
143  *   512 / 3 * 1.2 = 205 usec.
144  */
145 static unsigned int tx_irq_mod_usec = 150;
146 
147 /* This is the first interrupt mode to try out of:
148  * 0 => MSI-X
149  * 1 => MSI
150  * 2 => legacy
151  */
152 static unsigned int interrupt_mode;
153 
154 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
155  * i.e. the number of CPUs among which we may distribute simultaneous
156  * interrupt handling.
157  *
158  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
159  * The default (0) means to assign an interrupt to each core.
160  */
161 static unsigned int rss_cpus;
162 module_param(rss_cpus, uint, 0444);
163 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
164 
165 static bool phy_flash_cfg;
166 module_param(phy_flash_cfg, bool, 0644);
167 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
168 
169 static unsigned irq_adapt_low_thresh = 8000;
170 module_param(irq_adapt_low_thresh, uint, 0644);
171 MODULE_PARM_DESC(irq_adapt_low_thresh,
172 		 "Threshold score for reducing IRQ moderation");
173 
174 static unsigned irq_adapt_high_thresh = 16000;
175 module_param(irq_adapt_high_thresh, uint, 0644);
176 MODULE_PARM_DESC(irq_adapt_high_thresh,
177 		 "Threshold score for increasing IRQ moderation");
178 
179 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
180 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
181 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
182 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
183 module_param(debug, uint, 0);
184 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
185 
186 /**************************************************************************
187  *
188  * Utility functions and prototypes
189  *
190  *************************************************************************/
191 
192 static int efx_soft_enable_interrupts(struct efx_nic *efx);
193 static void efx_soft_disable_interrupts(struct efx_nic *efx);
194 static void efx_remove_channel(struct efx_channel *channel);
195 static void efx_remove_channels(struct efx_nic *efx);
196 static const struct efx_channel_type efx_default_channel_type;
197 static void efx_remove_port(struct efx_nic *efx);
198 static void efx_init_napi_channel(struct efx_channel *channel);
199 static void efx_fini_napi(struct efx_nic *efx);
200 static void efx_fini_napi_channel(struct efx_channel *channel);
201 static void efx_fini_struct(struct efx_nic *efx);
202 static void efx_start_all(struct efx_nic *efx);
203 static void efx_stop_all(struct efx_nic *efx);
204 
205 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
206 	do {						\
207 		if ((efx->state == STATE_READY) ||	\
208 		    (efx->state == STATE_RECOVERY) ||	\
209 		    (efx->state == STATE_DISABLED))	\
210 			ASSERT_RTNL();			\
211 	} while (0)
212 
213 static int efx_check_disabled(struct efx_nic *efx)
214 {
215 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
216 		netif_err(efx, drv, efx->net_dev,
217 			  "device is disabled due to earlier errors\n");
218 		return -EIO;
219 	}
220 	return 0;
221 }
222 
223 /**************************************************************************
224  *
225  * Event queue processing
226  *
227  *************************************************************************/
228 
229 /* Process channel's event queue
230  *
231  * This function is responsible for processing the event queue of a
232  * single channel.  The caller must guarantee that this function will
233  * never be concurrently called more than once on the same channel,
234  * though different channels may be being processed concurrently.
235  */
236 static int efx_process_channel(struct efx_channel *channel, int budget)
237 {
238 	int spent;
239 
240 	if (unlikely(!channel->enabled))
241 		return 0;
242 
243 	spent = efx_nic_process_eventq(channel, budget);
244 	if (spent && efx_channel_has_rx_queue(channel)) {
245 		struct efx_rx_queue *rx_queue =
246 			efx_channel_get_rx_queue(channel);
247 
248 		efx_rx_flush_packet(channel);
249 		efx_fast_push_rx_descriptors(rx_queue);
250 	}
251 
252 	return spent;
253 }
254 
255 /* NAPI poll handler
256  *
257  * NAPI guarantees serialisation of polls of the same device, which
258  * provides the guarantee required by efx_process_channel().
259  */
260 static int efx_poll(struct napi_struct *napi, int budget)
261 {
262 	struct efx_channel *channel =
263 		container_of(napi, struct efx_channel, napi_str);
264 	struct efx_nic *efx = channel->efx;
265 	int spent;
266 
267 	netif_vdbg(efx, intr, efx->net_dev,
268 		   "channel %d NAPI poll executing on CPU %d\n",
269 		   channel->channel, raw_smp_processor_id());
270 
271 	spent = efx_process_channel(channel, budget);
272 
273 	if (spent < budget) {
274 		if (efx_channel_has_rx_queue(channel) &&
275 		    efx->irq_rx_adaptive &&
276 		    unlikely(++channel->irq_count == 1000)) {
277 			if (unlikely(channel->irq_mod_score <
278 				     irq_adapt_low_thresh)) {
279 				if (channel->irq_moderation > 1) {
280 					channel->irq_moderation -= 1;
281 					efx->type->push_irq_moderation(channel);
282 				}
283 			} else if (unlikely(channel->irq_mod_score >
284 					    irq_adapt_high_thresh)) {
285 				if (channel->irq_moderation <
286 				    efx->irq_rx_moderation) {
287 					channel->irq_moderation += 1;
288 					efx->type->push_irq_moderation(channel);
289 				}
290 			}
291 			channel->irq_count = 0;
292 			channel->irq_mod_score = 0;
293 		}
294 
295 		efx_filter_rfs_expire(channel);
296 
297 		/* There is no race here; although napi_disable() will
298 		 * only wait for napi_complete(), this isn't a problem
299 		 * since efx_nic_eventq_read_ack() will have no effect if
300 		 * interrupts have already been disabled.
301 		 */
302 		napi_complete(napi);
303 		efx_nic_eventq_read_ack(channel);
304 	}
305 
306 	return spent;
307 }
308 
309 /* Create event queue
310  * Event queue memory allocations are done only once.  If the channel
311  * is reset, the memory buffer will be reused; this guards against
312  * errors during channel reset and also simplifies interrupt handling.
313  */
314 static int efx_probe_eventq(struct efx_channel *channel)
315 {
316 	struct efx_nic *efx = channel->efx;
317 	unsigned long entries;
318 
319 	netif_dbg(efx, probe, efx->net_dev,
320 		  "chan %d create event queue\n", channel->channel);
321 
322 	/* Build an event queue with room for one event per tx and rx buffer,
323 	 * plus some extra for link state events and MCDI completions. */
324 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
325 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
326 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
327 
328 	return efx_nic_probe_eventq(channel);
329 }
330 
331 /* Prepare channel's event queue */
332 static int efx_init_eventq(struct efx_channel *channel)
333 {
334 	struct efx_nic *efx = channel->efx;
335 	int rc;
336 
337 	EFX_WARN_ON_PARANOID(channel->eventq_init);
338 
339 	netif_dbg(efx, drv, efx->net_dev,
340 		  "chan %d init event queue\n", channel->channel);
341 
342 	rc = efx_nic_init_eventq(channel);
343 	if (rc == 0) {
344 		efx->type->push_irq_moderation(channel);
345 		channel->eventq_read_ptr = 0;
346 		channel->eventq_init = true;
347 	}
348 	return rc;
349 }
350 
351 /* Enable event queue processing and NAPI */
352 static void efx_start_eventq(struct efx_channel *channel)
353 {
354 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
355 		  "chan %d start event queue\n", channel->channel);
356 
357 	/* Make sure the NAPI handler sees the enabled flag set */
358 	channel->enabled = true;
359 	smp_wmb();
360 
361 	napi_enable(&channel->napi_str);
362 	efx_nic_eventq_read_ack(channel);
363 }
364 
365 /* Disable event queue processing and NAPI */
366 static void efx_stop_eventq(struct efx_channel *channel)
367 {
368 	if (!channel->enabled)
369 		return;
370 
371 	napi_disable(&channel->napi_str);
372 	channel->enabled = false;
373 }
374 
375 static void efx_fini_eventq(struct efx_channel *channel)
376 {
377 	if (!channel->eventq_init)
378 		return;
379 
380 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
381 		  "chan %d fini event queue\n", channel->channel);
382 
383 	efx_nic_fini_eventq(channel);
384 	channel->eventq_init = false;
385 }
386 
387 static void efx_remove_eventq(struct efx_channel *channel)
388 {
389 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
390 		  "chan %d remove event queue\n", channel->channel);
391 
392 	efx_nic_remove_eventq(channel);
393 }
394 
395 /**************************************************************************
396  *
397  * Channel handling
398  *
399  *************************************************************************/
400 
401 /* Allocate and initialise a channel structure. */
402 static struct efx_channel *
403 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
404 {
405 	struct efx_channel *channel;
406 	struct efx_rx_queue *rx_queue;
407 	struct efx_tx_queue *tx_queue;
408 	int j;
409 
410 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
411 	if (!channel)
412 		return NULL;
413 
414 	channel->efx = efx;
415 	channel->channel = i;
416 	channel->type = &efx_default_channel_type;
417 
418 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
419 		tx_queue = &channel->tx_queue[j];
420 		tx_queue->efx = efx;
421 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
422 		tx_queue->channel = channel;
423 	}
424 
425 	rx_queue = &channel->rx_queue;
426 	rx_queue->efx = efx;
427 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
428 		    (unsigned long)rx_queue);
429 
430 	return channel;
431 }
432 
433 /* Allocate and initialise a channel structure, copying parameters
434  * (but not resources) from an old channel structure.
435  */
436 static struct efx_channel *
437 efx_copy_channel(const struct efx_channel *old_channel)
438 {
439 	struct efx_channel *channel;
440 	struct efx_rx_queue *rx_queue;
441 	struct efx_tx_queue *tx_queue;
442 	int j;
443 
444 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
445 	if (!channel)
446 		return NULL;
447 
448 	*channel = *old_channel;
449 
450 	channel->napi_dev = NULL;
451 	memset(&channel->eventq, 0, sizeof(channel->eventq));
452 
453 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
454 		tx_queue = &channel->tx_queue[j];
455 		if (tx_queue->channel)
456 			tx_queue->channel = channel;
457 		tx_queue->buffer = NULL;
458 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
459 	}
460 
461 	rx_queue = &channel->rx_queue;
462 	rx_queue->buffer = NULL;
463 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
464 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
465 		    (unsigned long)rx_queue);
466 
467 	return channel;
468 }
469 
470 static int efx_probe_channel(struct efx_channel *channel)
471 {
472 	struct efx_tx_queue *tx_queue;
473 	struct efx_rx_queue *rx_queue;
474 	int rc;
475 
476 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
477 		  "creating channel %d\n", channel->channel);
478 
479 	rc = channel->type->pre_probe(channel);
480 	if (rc)
481 		goto fail;
482 
483 	rc = efx_probe_eventq(channel);
484 	if (rc)
485 		goto fail;
486 
487 	efx_for_each_channel_tx_queue(tx_queue, channel) {
488 		rc = efx_probe_tx_queue(tx_queue);
489 		if (rc)
490 			goto fail;
491 	}
492 
493 	efx_for_each_channel_rx_queue(rx_queue, channel) {
494 		rc = efx_probe_rx_queue(rx_queue);
495 		if (rc)
496 			goto fail;
497 	}
498 
499 	channel->n_rx_frm_trunc = 0;
500 
501 	return 0;
502 
503 fail:
504 	efx_remove_channel(channel);
505 	return rc;
506 }
507 
508 static void
509 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
510 {
511 	struct efx_nic *efx = channel->efx;
512 	const char *type;
513 	int number;
514 
515 	number = channel->channel;
516 	if (efx->tx_channel_offset == 0) {
517 		type = "";
518 	} else if (channel->channel < efx->tx_channel_offset) {
519 		type = "-rx";
520 	} else {
521 		type = "-tx";
522 		number -= efx->tx_channel_offset;
523 	}
524 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
525 }
526 
527 static void efx_set_channel_names(struct efx_nic *efx)
528 {
529 	struct efx_channel *channel;
530 
531 	efx_for_each_channel(channel, efx)
532 		channel->type->get_name(channel,
533 					efx->msi_context[channel->channel].name,
534 					sizeof(efx->msi_context[0].name));
535 }
536 
537 static int efx_probe_channels(struct efx_nic *efx)
538 {
539 	struct efx_channel *channel;
540 	int rc;
541 
542 	/* Restart special buffer allocation */
543 	efx->next_buffer_table = 0;
544 
545 	/* Probe channels in reverse, so that any 'extra' channels
546 	 * use the start of the buffer table. This allows the traffic
547 	 * channels to be resized without moving them or wasting the
548 	 * entries before them.
549 	 */
550 	efx_for_each_channel_rev(channel, efx) {
551 		rc = efx_probe_channel(channel);
552 		if (rc) {
553 			netif_err(efx, probe, efx->net_dev,
554 				  "failed to create channel %d\n",
555 				  channel->channel);
556 			goto fail;
557 		}
558 	}
559 	efx_set_channel_names(efx);
560 
561 	return 0;
562 
563 fail:
564 	efx_remove_channels(efx);
565 	return rc;
566 }
567 
568 /* Channels are shutdown and reinitialised whilst the NIC is running
569  * to propagate configuration changes (mtu, checksum offload), or
570  * to clear hardware error conditions
571  */
572 static void efx_start_datapath(struct efx_nic *efx)
573 {
574 	bool old_rx_scatter = efx->rx_scatter;
575 	struct efx_tx_queue *tx_queue;
576 	struct efx_rx_queue *rx_queue;
577 	struct efx_channel *channel;
578 	size_t rx_buf_len;
579 
580 	/* Calculate the rx buffer allocation parameters required to
581 	 * support the current MTU, including padding for header
582 	 * alignment and overruns.
583 	 */
584 	efx->rx_dma_len = (efx->rx_prefix_size +
585 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
586 			   efx->type->rx_buffer_padding);
587 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
588 		      NET_IP_ALIGN + efx->rx_dma_len);
589 	if (rx_buf_len <= PAGE_SIZE) {
590 		efx->rx_scatter = efx->type->always_rx_scatter;
591 		efx->rx_buffer_order = 0;
592 	} else if (efx->type->can_rx_scatter) {
593 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
594 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
595 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
596 				       EFX_RX_BUF_ALIGNMENT) >
597 			     PAGE_SIZE);
598 		efx->rx_scatter = true;
599 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
600 		efx->rx_buffer_order = 0;
601 	} else {
602 		efx->rx_scatter = false;
603 		efx->rx_buffer_order = get_order(rx_buf_len);
604 	}
605 
606 	efx_rx_config_page_split(efx);
607 	if (efx->rx_buffer_order)
608 		netif_dbg(efx, drv, efx->net_dev,
609 			  "RX buf len=%u; page order=%u batch=%u\n",
610 			  efx->rx_dma_len, efx->rx_buffer_order,
611 			  efx->rx_pages_per_batch);
612 	else
613 		netif_dbg(efx, drv, efx->net_dev,
614 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
615 			  efx->rx_dma_len, efx->rx_page_buf_step,
616 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
617 
618 	/* RX filters may also have scatter-enabled flags */
619 	if (efx->rx_scatter != old_rx_scatter)
620 		efx->type->filter_update_rx_scatter(efx);
621 
622 	/* We must keep at least one descriptor in a TX ring empty.
623 	 * We could avoid this when the queue size does not exactly
624 	 * match the hardware ring size, but it's not that important.
625 	 * Therefore we stop the queue when one more skb might fill
626 	 * the ring completely.  We wake it when half way back to
627 	 * empty.
628 	 */
629 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
630 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
631 
632 	/* Initialise the channels */
633 	efx_for_each_channel(channel, efx) {
634 		efx_for_each_channel_tx_queue(tx_queue, channel) {
635 			efx_init_tx_queue(tx_queue);
636 			atomic_inc(&efx->active_queues);
637 		}
638 
639 		efx_for_each_channel_rx_queue(rx_queue, channel) {
640 			efx_init_rx_queue(rx_queue);
641 			atomic_inc(&efx->active_queues);
642 			efx_nic_generate_fill_event(rx_queue);
643 		}
644 
645 		WARN_ON(channel->rx_pkt_n_frags);
646 	}
647 
648 	if (netif_device_present(efx->net_dev))
649 		netif_tx_wake_all_queues(efx->net_dev);
650 }
651 
652 static void efx_stop_datapath(struct efx_nic *efx)
653 {
654 	struct efx_channel *channel;
655 	struct efx_tx_queue *tx_queue;
656 	struct efx_rx_queue *rx_queue;
657 	int rc;
658 
659 	EFX_ASSERT_RESET_SERIALISED(efx);
660 	BUG_ON(efx->port_enabled);
661 
662 	/* Stop RX refill */
663 	efx_for_each_channel(channel, efx) {
664 		efx_for_each_channel_rx_queue(rx_queue, channel)
665 			rx_queue->refill_enabled = false;
666 	}
667 
668 	efx_for_each_channel(channel, efx) {
669 		/* RX packet processing is pipelined, so wait for the
670 		 * NAPI handler to complete.  At least event queue 0
671 		 * might be kept active by non-data events, so don't
672 		 * use napi_synchronize() but actually disable NAPI
673 		 * temporarily.
674 		 */
675 		if (efx_channel_has_rx_queue(channel)) {
676 			efx_stop_eventq(channel);
677 			efx_start_eventq(channel);
678 		}
679 	}
680 
681 	rc = efx->type->fini_dmaq(efx);
682 	if (rc && EFX_WORKAROUND_7803(efx)) {
683 		/* Schedule a reset to recover from the flush failure. The
684 		 * descriptor caches reference memory we're about to free,
685 		 * but falcon_reconfigure_mac_wrapper() won't reconnect
686 		 * the MACs because of the pending reset.
687 		 */
688 		netif_err(efx, drv, efx->net_dev,
689 			  "Resetting to recover from flush failure\n");
690 		efx_schedule_reset(efx, RESET_TYPE_ALL);
691 	} else if (rc) {
692 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
693 	} else {
694 		netif_dbg(efx, drv, efx->net_dev,
695 			  "successfully flushed all queues\n");
696 	}
697 
698 	efx_for_each_channel(channel, efx) {
699 		efx_for_each_channel_rx_queue(rx_queue, channel)
700 			efx_fini_rx_queue(rx_queue);
701 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
702 			efx_fini_tx_queue(tx_queue);
703 	}
704 }
705 
706 static void efx_remove_channel(struct efx_channel *channel)
707 {
708 	struct efx_tx_queue *tx_queue;
709 	struct efx_rx_queue *rx_queue;
710 
711 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
712 		  "destroy chan %d\n", channel->channel);
713 
714 	efx_for_each_channel_rx_queue(rx_queue, channel)
715 		efx_remove_rx_queue(rx_queue);
716 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
717 		efx_remove_tx_queue(tx_queue);
718 	efx_remove_eventq(channel);
719 	channel->type->post_remove(channel);
720 }
721 
722 static void efx_remove_channels(struct efx_nic *efx)
723 {
724 	struct efx_channel *channel;
725 
726 	efx_for_each_channel(channel, efx)
727 		efx_remove_channel(channel);
728 }
729 
730 int
731 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
732 {
733 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
734 	u32 old_rxq_entries, old_txq_entries;
735 	unsigned i, next_buffer_table = 0;
736 	int rc, rc2;
737 
738 	rc = efx_check_disabled(efx);
739 	if (rc)
740 		return rc;
741 
742 	/* Not all channels should be reallocated. We must avoid
743 	 * reallocating their buffer table entries.
744 	 */
745 	efx_for_each_channel(channel, efx) {
746 		struct efx_rx_queue *rx_queue;
747 		struct efx_tx_queue *tx_queue;
748 
749 		if (channel->type->copy)
750 			continue;
751 		next_buffer_table = max(next_buffer_table,
752 					channel->eventq.index +
753 					channel->eventq.entries);
754 		efx_for_each_channel_rx_queue(rx_queue, channel)
755 			next_buffer_table = max(next_buffer_table,
756 						rx_queue->rxd.index +
757 						rx_queue->rxd.entries);
758 		efx_for_each_channel_tx_queue(tx_queue, channel)
759 			next_buffer_table = max(next_buffer_table,
760 						tx_queue->txd.index +
761 						tx_queue->txd.entries);
762 	}
763 
764 	efx_device_detach_sync(efx);
765 	efx_stop_all(efx);
766 	efx_soft_disable_interrupts(efx);
767 
768 	/* Clone channels (where possible) */
769 	memset(other_channel, 0, sizeof(other_channel));
770 	for (i = 0; i < efx->n_channels; i++) {
771 		channel = efx->channel[i];
772 		if (channel->type->copy)
773 			channel = channel->type->copy(channel);
774 		if (!channel) {
775 			rc = -ENOMEM;
776 			goto out;
777 		}
778 		other_channel[i] = channel;
779 	}
780 
781 	/* Swap entry counts and channel pointers */
782 	old_rxq_entries = efx->rxq_entries;
783 	old_txq_entries = efx->txq_entries;
784 	efx->rxq_entries = rxq_entries;
785 	efx->txq_entries = txq_entries;
786 	for (i = 0; i < efx->n_channels; i++) {
787 		channel = efx->channel[i];
788 		efx->channel[i] = other_channel[i];
789 		other_channel[i] = channel;
790 	}
791 
792 	/* Restart buffer table allocation */
793 	efx->next_buffer_table = next_buffer_table;
794 
795 	for (i = 0; i < efx->n_channels; i++) {
796 		channel = efx->channel[i];
797 		if (!channel->type->copy)
798 			continue;
799 		rc = efx_probe_channel(channel);
800 		if (rc)
801 			goto rollback;
802 		efx_init_napi_channel(efx->channel[i]);
803 	}
804 
805 out:
806 	/* Destroy unused channel structures */
807 	for (i = 0; i < efx->n_channels; i++) {
808 		channel = other_channel[i];
809 		if (channel && channel->type->copy) {
810 			efx_fini_napi_channel(channel);
811 			efx_remove_channel(channel);
812 			kfree(channel);
813 		}
814 	}
815 
816 	rc2 = efx_soft_enable_interrupts(efx);
817 	if (rc2) {
818 		rc = rc ? rc : rc2;
819 		netif_err(efx, drv, efx->net_dev,
820 			  "unable to restart interrupts on channel reallocation\n");
821 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
822 	} else {
823 		efx_start_all(efx);
824 		netif_device_attach(efx->net_dev);
825 	}
826 	return rc;
827 
828 rollback:
829 	/* Swap back */
830 	efx->rxq_entries = old_rxq_entries;
831 	efx->txq_entries = old_txq_entries;
832 	for (i = 0; i < efx->n_channels; i++) {
833 		channel = efx->channel[i];
834 		efx->channel[i] = other_channel[i];
835 		other_channel[i] = channel;
836 	}
837 	goto out;
838 }
839 
840 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
841 {
842 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
843 }
844 
845 static const struct efx_channel_type efx_default_channel_type = {
846 	.pre_probe		= efx_channel_dummy_op_int,
847 	.post_remove		= efx_channel_dummy_op_void,
848 	.get_name		= efx_get_channel_name,
849 	.copy			= efx_copy_channel,
850 	.keep_eventq		= false,
851 };
852 
853 int efx_channel_dummy_op_int(struct efx_channel *channel)
854 {
855 	return 0;
856 }
857 
858 void efx_channel_dummy_op_void(struct efx_channel *channel)
859 {
860 }
861 
862 /**************************************************************************
863  *
864  * Port handling
865  *
866  **************************************************************************/
867 
868 /* This ensures that the kernel is kept informed (via
869  * netif_carrier_on/off) of the link status, and also maintains the
870  * link status's stop on the port's TX queue.
871  */
872 void efx_link_status_changed(struct efx_nic *efx)
873 {
874 	struct efx_link_state *link_state = &efx->link_state;
875 
876 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
877 	 * that no events are triggered between unregister_netdev() and the
878 	 * driver unloading. A more general condition is that NETDEV_CHANGE
879 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
880 	if (!netif_running(efx->net_dev))
881 		return;
882 
883 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
884 		efx->n_link_state_changes++;
885 
886 		if (link_state->up)
887 			netif_carrier_on(efx->net_dev);
888 		else
889 			netif_carrier_off(efx->net_dev);
890 	}
891 
892 	/* Status message for kernel log */
893 	if (link_state->up)
894 		netif_info(efx, link, efx->net_dev,
895 			   "link up at %uMbps %s-duplex (MTU %d)\n",
896 			   link_state->speed, link_state->fd ? "full" : "half",
897 			   efx->net_dev->mtu);
898 	else
899 		netif_info(efx, link, efx->net_dev, "link down\n");
900 }
901 
902 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
903 {
904 	efx->link_advertising = advertising;
905 	if (advertising) {
906 		if (advertising & ADVERTISED_Pause)
907 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
908 		else
909 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
910 		if (advertising & ADVERTISED_Asym_Pause)
911 			efx->wanted_fc ^= EFX_FC_TX;
912 	}
913 }
914 
915 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
916 {
917 	efx->wanted_fc = wanted_fc;
918 	if (efx->link_advertising) {
919 		if (wanted_fc & EFX_FC_RX)
920 			efx->link_advertising |= (ADVERTISED_Pause |
921 						  ADVERTISED_Asym_Pause);
922 		else
923 			efx->link_advertising &= ~(ADVERTISED_Pause |
924 						   ADVERTISED_Asym_Pause);
925 		if (wanted_fc & EFX_FC_TX)
926 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
927 	}
928 }
929 
930 static void efx_fini_port(struct efx_nic *efx);
931 
932 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
933  * the MAC appropriately. All other PHY configuration changes are pushed
934  * through phy_op->set_settings(), and pushed asynchronously to the MAC
935  * through efx_monitor().
936  *
937  * Callers must hold the mac_lock
938  */
939 int __efx_reconfigure_port(struct efx_nic *efx)
940 {
941 	enum efx_phy_mode phy_mode;
942 	int rc;
943 
944 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
945 
946 	/* Disable PHY transmit in mac level loopbacks */
947 	phy_mode = efx->phy_mode;
948 	if (LOOPBACK_INTERNAL(efx))
949 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
950 	else
951 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
952 
953 	rc = efx->type->reconfigure_port(efx);
954 
955 	if (rc)
956 		efx->phy_mode = phy_mode;
957 
958 	return rc;
959 }
960 
961 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
962  * disabled. */
963 int efx_reconfigure_port(struct efx_nic *efx)
964 {
965 	int rc;
966 
967 	EFX_ASSERT_RESET_SERIALISED(efx);
968 
969 	mutex_lock(&efx->mac_lock);
970 	rc = __efx_reconfigure_port(efx);
971 	mutex_unlock(&efx->mac_lock);
972 
973 	return rc;
974 }
975 
976 /* Asynchronous work item for changing MAC promiscuity and multicast
977  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
978  * MAC directly. */
979 static void efx_mac_work(struct work_struct *data)
980 {
981 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
982 
983 	mutex_lock(&efx->mac_lock);
984 	if (efx->port_enabled)
985 		efx->type->reconfigure_mac(efx);
986 	mutex_unlock(&efx->mac_lock);
987 }
988 
989 static int efx_probe_port(struct efx_nic *efx)
990 {
991 	int rc;
992 
993 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
994 
995 	if (phy_flash_cfg)
996 		efx->phy_mode = PHY_MODE_SPECIAL;
997 
998 	/* Connect up MAC/PHY operations table */
999 	rc = efx->type->probe_port(efx);
1000 	if (rc)
1001 		return rc;
1002 
1003 	/* Initialise MAC address to permanent address */
1004 	memcpy(efx->net_dev->dev_addr, efx->net_dev->perm_addr, ETH_ALEN);
1005 
1006 	return 0;
1007 }
1008 
1009 static int efx_init_port(struct efx_nic *efx)
1010 {
1011 	int rc;
1012 
1013 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1014 
1015 	mutex_lock(&efx->mac_lock);
1016 
1017 	rc = efx->phy_op->init(efx);
1018 	if (rc)
1019 		goto fail1;
1020 
1021 	efx->port_initialized = true;
1022 
1023 	/* Reconfigure the MAC before creating dma queues (required for
1024 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1025 	efx->type->reconfigure_mac(efx);
1026 
1027 	/* Ensure the PHY advertises the correct flow control settings */
1028 	rc = efx->phy_op->reconfigure(efx);
1029 	if (rc)
1030 		goto fail2;
1031 
1032 	mutex_unlock(&efx->mac_lock);
1033 	return 0;
1034 
1035 fail2:
1036 	efx->phy_op->fini(efx);
1037 fail1:
1038 	mutex_unlock(&efx->mac_lock);
1039 	return rc;
1040 }
1041 
1042 static void efx_start_port(struct efx_nic *efx)
1043 {
1044 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1045 	BUG_ON(efx->port_enabled);
1046 
1047 	mutex_lock(&efx->mac_lock);
1048 	efx->port_enabled = true;
1049 
1050 	/* efx_mac_work() might have been scheduled after efx_stop_port(),
1051 	 * and then cancelled by efx_flush_all() */
1052 	efx->type->reconfigure_mac(efx);
1053 
1054 	mutex_unlock(&efx->mac_lock);
1055 }
1056 
1057 /* Prevent efx_mac_work() and efx_monitor() from working */
1058 static void efx_stop_port(struct efx_nic *efx)
1059 {
1060 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1061 
1062 	mutex_lock(&efx->mac_lock);
1063 	efx->port_enabled = false;
1064 	mutex_unlock(&efx->mac_lock);
1065 
1066 	/* Serialise against efx_set_multicast_list() */
1067 	netif_addr_lock_bh(efx->net_dev);
1068 	netif_addr_unlock_bh(efx->net_dev);
1069 }
1070 
1071 static void efx_fini_port(struct efx_nic *efx)
1072 {
1073 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1074 
1075 	if (!efx->port_initialized)
1076 		return;
1077 
1078 	efx->phy_op->fini(efx);
1079 	efx->port_initialized = false;
1080 
1081 	efx->link_state.up = false;
1082 	efx_link_status_changed(efx);
1083 }
1084 
1085 static void efx_remove_port(struct efx_nic *efx)
1086 {
1087 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1088 
1089 	efx->type->remove_port(efx);
1090 }
1091 
1092 /**************************************************************************
1093  *
1094  * NIC handling
1095  *
1096  **************************************************************************/
1097 
1098 /* This configures the PCI device to enable I/O and DMA. */
1099 static int efx_init_io(struct efx_nic *efx)
1100 {
1101 	struct pci_dev *pci_dev = efx->pci_dev;
1102 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1103 	unsigned int mem_map_size = efx->type->mem_map_size(efx);
1104 	int rc;
1105 
1106 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1107 
1108 	rc = pci_enable_device(pci_dev);
1109 	if (rc) {
1110 		netif_err(efx, probe, efx->net_dev,
1111 			  "failed to enable PCI device\n");
1112 		goto fail1;
1113 	}
1114 
1115 	pci_set_master(pci_dev);
1116 
1117 	/* Set the PCI DMA mask.  Try all possibilities from our
1118 	 * genuine mask down to 32 bits, because some architectures
1119 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1120 	 * masks event though they reject 46 bit masks.
1121 	 */
1122 	while (dma_mask > 0x7fffffffUL) {
1123 		if (dma_supported(&pci_dev->dev, dma_mask)) {
1124 			rc = dma_set_mask(&pci_dev->dev, dma_mask);
1125 			if (rc == 0)
1126 				break;
1127 		}
1128 		dma_mask >>= 1;
1129 	}
1130 	if (rc) {
1131 		netif_err(efx, probe, efx->net_dev,
1132 			  "could not find a suitable DMA mask\n");
1133 		goto fail2;
1134 	}
1135 	netif_dbg(efx, probe, efx->net_dev,
1136 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1137 	rc = dma_set_coherent_mask(&pci_dev->dev, dma_mask);
1138 	if (rc) {
1139 		/* dma_set_coherent_mask() is not *allowed* to
1140 		 * fail with a mask that dma_set_mask() accepted,
1141 		 * but just in case...
1142 		 */
1143 		netif_err(efx, probe, efx->net_dev,
1144 			  "failed to set consistent DMA mask\n");
1145 		goto fail2;
1146 	}
1147 
1148 	efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1149 	rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1150 	if (rc) {
1151 		netif_err(efx, probe, efx->net_dev,
1152 			  "request for memory BAR failed\n");
1153 		rc = -EIO;
1154 		goto fail3;
1155 	}
1156 	efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1157 	if (!efx->membase) {
1158 		netif_err(efx, probe, efx->net_dev,
1159 			  "could not map memory BAR at %llx+%x\n",
1160 			  (unsigned long long)efx->membase_phys, mem_map_size);
1161 		rc = -ENOMEM;
1162 		goto fail4;
1163 	}
1164 	netif_dbg(efx, probe, efx->net_dev,
1165 		  "memory BAR at %llx+%x (virtual %p)\n",
1166 		  (unsigned long long)efx->membase_phys, mem_map_size,
1167 		  efx->membase);
1168 
1169 	return 0;
1170 
1171  fail4:
1172 	pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1173  fail3:
1174 	efx->membase_phys = 0;
1175  fail2:
1176 	pci_disable_device(efx->pci_dev);
1177  fail1:
1178 	return rc;
1179 }
1180 
1181 static void efx_fini_io(struct efx_nic *efx)
1182 {
1183 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1184 
1185 	if (efx->membase) {
1186 		iounmap(efx->membase);
1187 		efx->membase = NULL;
1188 	}
1189 
1190 	if (efx->membase_phys) {
1191 		pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1192 		efx->membase_phys = 0;
1193 	}
1194 
1195 	pci_disable_device(efx->pci_dev);
1196 }
1197 
1198 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1199 {
1200 	cpumask_var_t thread_mask;
1201 	unsigned int count;
1202 	int cpu;
1203 
1204 	if (rss_cpus) {
1205 		count = rss_cpus;
1206 	} else {
1207 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1208 			netif_warn(efx, probe, efx->net_dev,
1209 				   "RSS disabled due to allocation failure\n");
1210 			return 1;
1211 		}
1212 
1213 		count = 0;
1214 		for_each_online_cpu(cpu) {
1215 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1216 				++count;
1217 				cpumask_or(thread_mask, thread_mask,
1218 					   topology_thread_cpumask(cpu));
1219 			}
1220 		}
1221 
1222 		free_cpumask_var(thread_mask);
1223 	}
1224 
1225 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1226 	 * table entries that are inaccessible to VFs
1227 	 */
1228 	if (efx_sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1229 	    count > efx_vf_size(efx)) {
1230 		netif_warn(efx, probe, efx->net_dev,
1231 			   "Reducing number of RSS channels from %u to %u for "
1232 			   "VF support. Increase vf-msix-limit to use more "
1233 			   "channels on the PF.\n",
1234 			   count, efx_vf_size(efx));
1235 		count = efx_vf_size(efx);
1236 	}
1237 
1238 	return count;
1239 }
1240 
1241 /* Probe the number and type of interrupts we are able to obtain, and
1242  * the resulting numbers of channels and RX queues.
1243  */
1244 static int efx_probe_interrupts(struct efx_nic *efx)
1245 {
1246 	unsigned int extra_channels = 0;
1247 	unsigned int i, j;
1248 	int rc;
1249 
1250 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1251 		if (efx->extra_channel_type[i])
1252 			++extra_channels;
1253 
1254 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1255 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1256 		unsigned int n_channels;
1257 
1258 		n_channels = efx_wanted_parallelism(efx);
1259 		if (separate_tx_channels)
1260 			n_channels *= 2;
1261 		n_channels += extra_channels;
1262 		n_channels = min(n_channels, efx->max_channels);
1263 
1264 		for (i = 0; i < n_channels; i++)
1265 			xentries[i].entry = i;
1266 		rc = pci_enable_msix(efx->pci_dev, xentries, n_channels);
1267 		if (rc > 0) {
1268 			netif_err(efx, drv, efx->net_dev,
1269 				  "WARNING: Insufficient MSI-X vectors"
1270 				  " available (%d < %u).\n", rc, n_channels);
1271 			netif_err(efx, drv, efx->net_dev,
1272 				  "WARNING: Performance may be reduced.\n");
1273 			EFX_BUG_ON_PARANOID(rc >= n_channels);
1274 			n_channels = rc;
1275 			rc = pci_enable_msix(efx->pci_dev, xentries,
1276 					     n_channels);
1277 		}
1278 
1279 		if (rc == 0) {
1280 			efx->n_channels = n_channels;
1281 			if (n_channels > extra_channels)
1282 				n_channels -= extra_channels;
1283 			if (separate_tx_channels) {
1284 				efx->n_tx_channels = max(n_channels / 2, 1U);
1285 				efx->n_rx_channels = max(n_channels -
1286 							 efx->n_tx_channels,
1287 							 1U);
1288 			} else {
1289 				efx->n_tx_channels = n_channels;
1290 				efx->n_rx_channels = n_channels;
1291 			}
1292 			for (i = 0; i < efx->n_channels; i++)
1293 				efx_get_channel(efx, i)->irq =
1294 					xentries[i].vector;
1295 		} else {
1296 			/* Fall back to single channel MSI */
1297 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1298 			netif_err(efx, drv, efx->net_dev,
1299 				  "could not enable MSI-X\n");
1300 		}
1301 	}
1302 
1303 	/* Try single interrupt MSI */
1304 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1305 		efx->n_channels = 1;
1306 		efx->n_rx_channels = 1;
1307 		efx->n_tx_channels = 1;
1308 		rc = pci_enable_msi(efx->pci_dev);
1309 		if (rc == 0) {
1310 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1311 		} else {
1312 			netif_err(efx, drv, efx->net_dev,
1313 				  "could not enable MSI\n");
1314 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1315 		}
1316 	}
1317 
1318 	/* Assume legacy interrupts */
1319 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1320 		efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1321 		efx->n_rx_channels = 1;
1322 		efx->n_tx_channels = 1;
1323 		efx->legacy_irq = efx->pci_dev->irq;
1324 	}
1325 
1326 	/* Assign extra channels if possible */
1327 	j = efx->n_channels;
1328 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1329 		if (!efx->extra_channel_type[i])
1330 			continue;
1331 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1332 		    efx->n_channels <= extra_channels) {
1333 			efx->extra_channel_type[i]->handle_no_channel(efx);
1334 		} else {
1335 			--j;
1336 			efx_get_channel(efx, j)->type =
1337 				efx->extra_channel_type[i];
1338 		}
1339 	}
1340 
1341 	/* RSS might be usable on VFs even if it is disabled on the PF */
1342 	efx->rss_spread = ((efx->n_rx_channels > 1 || !efx_sriov_wanted(efx)) ?
1343 			   efx->n_rx_channels : efx_vf_size(efx));
1344 
1345 	return 0;
1346 }
1347 
1348 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1349 {
1350 	struct efx_channel *channel, *end_channel;
1351 	int rc;
1352 
1353 	BUG_ON(efx->state == STATE_DISABLED);
1354 
1355 	efx->irq_soft_enabled = true;
1356 	smp_wmb();
1357 
1358 	efx_for_each_channel(channel, efx) {
1359 		if (!channel->type->keep_eventq) {
1360 			rc = efx_init_eventq(channel);
1361 			if (rc)
1362 				goto fail;
1363 		}
1364 		efx_start_eventq(channel);
1365 	}
1366 
1367 	efx_mcdi_mode_event(efx);
1368 
1369 	return 0;
1370 fail:
1371 	end_channel = channel;
1372 	efx_for_each_channel(channel, efx) {
1373 		if (channel == end_channel)
1374 			break;
1375 		efx_stop_eventq(channel);
1376 		if (!channel->type->keep_eventq)
1377 			efx_fini_eventq(channel);
1378 	}
1379 
1380 	return rc;
1381 }
1382 
1383 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1384 {
1385 	struct efx_channel *channel;
1386 
1387 	if (efx->state == STATE_DISABLED)
1388 		return;
1389 
1390 	efx_mcdi_mode_poll(efx);
1391 
1392 	efx->irq_soft_enabled = false;
1393 	smp_wmb();
1394 
1395 	if (efx->legacy_irq)
1396 		synchronize_irq(efx->legacy_irq);
1397 
1398 	efx_for_each_channel(channel, efx) {
1399 		if (channel->irq)
1400 			synchronize_irq(channel->irq);
1401 
1402 		efx_stop_eventq(channel);
1403 		if (!channel->type->keep_eventq)
1404 			efx_fini_eventq(channel);
1405 	}
1406 
1407 	/* Flush the asynchronous MCDI request queue */
1408 	efx_mcdi_flush_async(efx);
1409 }
1410 
1411 static int efx_enable_interrupts(struct efx_nic *efx)
1412 {
1413 	struct efx_channel *channel, *end_channel;
1414 	int rc;
1415 
1416 	BUG_ON(efx->state == STATE_DISABLED);
1417 
1418 	if (efx->eeh_disabled_legacy_irq) {
1419 		enable_irq(efx->legacy_irq);
1420 		efx->eeh_disabled_legacy_irq = false;
1421 	}
1422 
1423 	efx->type->irq_enable_master(efx);
1424 
1425 	efx_for_each_channel(channel, efx) {
1426 		if (channel->type->keep_eventq) {
1427 			rc = efx_init_eventq(channel);
1428 			if (rc)
1429 				goto fail;
1430 		}
1431 	}
1432 
1433 	rc = efx_soft_enable_interrupts(efx);
1434 	if (rc)
1435 		goto fail;
1436 
1437 	return 0;
1438 
1439 fail:
1440 	end_channel = channel;
1441 	efx_for_each_channel(channel, efx) {
1442 		if (channel == end_channel)
1443 			break;
1444 		if (channel->type->keep_eventq)
1445 			efx_fini_eventq(channel);
1446 	}
1447 
1448 	efx->type->irq_disable_non_ev(efx);
1449 
1450 	return rc;
1451 }
1452 
1453 static void efx_disable_interrupts(struct efx_nic *efx)
1454 {
1455 	struct efx_channel *channel;
1456 
1457 	efx_soft_disable_interrupts(efx);
1458 
1459 	efx_for_each_channel(channel, efx) {
1460 		if (channel->type->keep_eventq)
1461 			efx_fini_eventq(channel);
1462 	}
1463 
1464 	efx->type->irq_disable_non_ev(efx);
1465 }
1466 
1467 static void efx_remove_interrupts(struct efx_nic *efx)
1468 {
1469 	struct efx_channel *channel;
1470 
1471 	/* Remove MSI/MSI-X interrupts */
1472 	efx_for_each_channel(channel, efx)
1473 		channel->irq = 0;
1474 	pci_disable_msi(efx->pci_dev);
1475 	pci_disable_msix(efx->pci_dev);
1476 
1477 	/* Remove legacy interrupt */
1478 	efx->legacy_irq = 0;
1479 }
1480 
1481 static void efx_set_channels(struct efx_nic *efx)
1482 {
1483 	struct efx_channel *channel;
1484 	struct efx_tx_queue *tx_queue;
1485 
1486 	efx->tx_channel_offset =
1487 		separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1488 
1489 	/* We need to mark which channels really have RX and TX
1490 	 * queues, and adjust the TX queue numbers if we have separate
1491 	 * RX-only and TX-only channels.
1492 	 */
1493 	efx_for_each_channel(channel, efx) {
1494 		if (channel->channel < efx->n_rx_channels)
1495 			channel->rx_queue.core_index = channel->channel;
1496 		else
1497 			channel->rx_queue.core_index = -1;
1498 
1499 		efx_for_each_channel_tx_queue(tx_queue, channel)
1500 			tx_queue->queue -= (efx->tx_channel_offset *
1501 					    EFX_TXQ_TYPES);
1502 	}
1503 }
1504 
1505 static int efx_probe_nic(struct efx_nic *efx)
1506 {
1507 	size_t i;
1508 	int rc;
1509 
1510 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1511 
1512 	/* Carry out hardware-type specific initialisation */
1513 	rc = efx->type->probe(efx);
1514 	if (rc)
1515 		return rc;
1516 
1517 	/* Determine the number of channels and queues by trying to hook
1518 	 * in MSI-X interrupts. */
1519 	rc = efx_probe_interrupts(efx);
1520 	if (rc)
1521 		goto fail1;
1522 
1523 	rc = efx->type->dimension_resources(efx);
1524 	if (rc)
1525 		goto fail2;
1526 
1527 	if (efx->n_channels > 1)
1528 		get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1529 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1530 		efx->rx_indir_table[i] =
1531 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1532 
1533 	efx_set_channels(efx);
1534 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1535 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1536 
1537 	/* Initialise the interrupt moderation settings */
1538 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1539 				true);
1540 
1541 	return 0;
1542 
1543 fail2:
1544 	efx_remove_interrupts(efx);
1545 fail1:
1546 	efx->type->remove(efx);
1547 	return rc;
1548 }
1549 
1550 static void efx_remove_nic(struct efx_nic *efx)
1551 {
1552 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1553 
1554 	efx_remove_interrupts(efx);
1555 	efx->type->remove(efx);
1556 }
1557 
1558 static int efx_probe_filters(struct efx_nic *efx)
1559 {
1560 	int rc;
1561 
1562 	spin_lock_init(&efx->filter_lock);
1563 
1564 	rc = efx->type->filter_table_probe(efx);
1565 	if (rc)
1566 		return rc;
1567 
1568 #ifdef CONFIG_RFS_ACCEL
1569 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
1570 		efx->rps_flow_id = kcalloc(efx->type->max_rx_ip_filters,
1571 					   sizeof(*efx->rps_flow_id),
1572 					   GFP_KERNEL);
1573 		if (!efx->rps_flow_id) {
1574 			efx->type->filter_table_remove(efx);
1575 			return -ENOMEM;
1576 		}
1577 	}
1578 #endif
1579 
1580 	return 0;
1581 }
1582 
1583 static void efx_remove_filters(struct efx_nic *efx)
1584 {
1585 #ifdef CONFIG_RFS_ACCEL
1586 	kfree(efx->rps_flow_id);
1587 #endif
1588 	efx->type->filter_table_remove(efx);
1589 }
1590 
1591 static void efx_restore_filters(struct efx_nic *efx)
1592 {
1593 	efx->type->filter_table_restore(efx);
1594 }
1595 
1596 /**************************************************************************
1597  *
1598  * NIC startup/shutdown
1599  *
1600  *************************************************************************/
1601 
1602 static int efx_probe_all(struct efx_nic *efx)
1603 {
1604 	int rc;
1605 
1606 	rc = efx_probe_nic(efx);
1607 	if (rc) {
1608 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1609 		goto fail1;
1610 	}
1611 
1612 	rc = efx_probe_port(efx);
1613 	if (rc) {
1614 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1615 		goto fail2;
1616 	}
1617 
1618 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1619 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1620 		rc = -EINVAL;
1621 		goto fail3;
1622 	}
1623 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1624 
1625 	rc = efx_probe_filters(efx);
1626 	if (rc) {
1627 		netif_err(efx, probe, efx->net_dev,
1628 			  "failed to create filter tables\n");
1629 		goto fail3;
1630 	}
1631 
1632 	rc = efx_probe_channels(efx);
1633 	if (rc)
1634 		goto fail4;
1635 
1636 	return 0;
1637 
1638  fail4:
1639 	efx_remove_filters(efx);
1640  fail3:
1641 	efx_remove_port(efx);
1642  fail2:
1643 	efx_remove_nic(efx);
1644  fail1:
1645 	return rc;
1646 }
1647 
1648 /* If the interface is supposed to be running but is not, start
1649  * the hardware and software data path, regular activity for the port
1650  * (MAC statistics, link polling, etc.) and schedule the port to be
1651  * reconfigured.  Interrupts must already be enabled.  This function
1652  * is safe to call multiple times, so long as the NIC is not disabled.
1653  * Requires the RTNL lock.
1654  */
1655 static void efx_start_all(struct efx_nic *efx)
1656 {
1657 	EFX_ASSERT_RESET_SERIALISED(efx);
1658 	BUG_ON(efx->state == STATE_DISABLED);
1659 
1660 	/* Check that it is appropriate to restart the interface. All
1661 	 * of these flags are safe to read under just the rtnl lock */
1662 	if (efx->port_enabled || !netif_running(efx->net_dev))
1663 		return;
1664 
1665 	efx_start_port(efx);
1666 	efx_start_datapath(efx);
1667 
1668 	/* Start the hardware monitor if there is one */
1669 	if (efx->type->monitor != NULL)
1670 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1671 				   efx_monitor_interval);
1672 
1673 	/* If link state detection is normally event-driven, we have
1674 	 * to poll now because we could have missed a change
1675 	 */
1676 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1677 		mutex_lock(&efx->mac_lock);
1678 		if (efx->phy_op->poll(efx))
1679 			efx_link_status_changed(efx);
1680 		mutex_unlock(&efx->mac_lock);
1681 	}
1682 
1683 	efx->type->start_stats(efx);
1684 }
1685 
1686 /* Flush all delayed work. Should only be called when no more delayed work
1687  * will be scheduled. This doesn't flush pending online resets (efx_reset),
1688  * since we're holding the rtnl_lock at this point. */
1689 static void efx_flush_all(struct efx_nic *efx)
1690 {
1691 	/* Make sure the hardware monitor and event self-test are stopped */
1692 	cancel_delayed_work_sync(&efx->monitor_work);
1693 	efx_selftest_async_cancel(efx);
1694 	/* Stop scheduled port reconfigurations */
1695 	cancel_work_sync(&efx->mac_work);
1696 }
1697 
1698 /* Quiesce the hardware and software data path, and regular activity
1699  * for the port without bringing the link down.  Safe to call multiple
1700  * times with the NIC in almost any state, but interrupts should be
1701  * enabled.  Requires the RTNL lock.
1702  */
1703 static void efx_stop_all(struct efx_nic *efx)
1704 {
1705 	EFX_ASSERT_RESET_SERIALISED(efx);
1706 
1707 	/* port_enabled can be read safely under the rtnl lock */
1708 	if (!efx->port_enabled)
1709 		return;
1710 
1711 	efx->type->stop_stats(efx);
1712 	efx_stop_port(efx);
1713 
1714 	/* Flush efx_mac_work(), refill_workqueue, monitor_work */
1715 	efx_flush_all(efx);
1716 
1717 	/* Stop the kernel transmit interface.  This is only valid if
1718 	 * the device is stopped or detached; otherwise the watchdog
1719 	 * may fire immediately.
1720 	 */
1721 	WARN_ON(netif_running(efx->net_dev) &&
1722 		netif_device_present(efx->net_dev));
1723 	netif_tx_disable(efx->net_dev);
1724 
1725 	efx_stop_datapath(efx);
1726 }
1727 
1728 static void efx_remove_all(struct efx_nic *efx)
1729 {
1730 	efx_remove_channels(efx);
1731 	efx_remove_filters(efx);
1732 	efx_remove_port(efx);
1733 	efx_remove_nic(efx);
1734 }
1735 
1736 /**************************************************************************
1737  *
1738  * Interrupt moderation
1739  *
1740  **************************************************************************/
1741 
1742 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1743 {
1744 	if (usecs == 0)
1745 		return 0;
1746 	if (usecs * 1000 < quantum_ns)
1747 		return 1; /* never round down to 0 */
1748 	return usecs * 1000 / quantum_ns;
1749 }
1750 
1751 /* Set interrupt moderation parameters */
1752 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1753 			    unsigned int rx_usecs, bool rx_adaptive,
1754 			    bool rx_may_override_tx)
1755 {
1756 	struct efx_channel *channel;
1757 	unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1758 						efx->timer_quantum_ns,
1759 						1000);
1760 	unsigned int tx_ticks;
1761 	unsigned int rx_ticks;
1762 
1763 	EFX_ASSERT_RESET_SERIALISED(efx);
1764 
1765 	if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1766 		return -EINVAL;
1767 
1768 	tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1769 	rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1770 
1771 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1772 	    !rx_may_override_tx) {
1773 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1774 			  "RX and TX IRQ moderation must be equal\n");
1775 		return -EINVAL;
1776 	}
1777 
1778 	efx->irq_rx_adaptive = rx_adaptive;
1779 	efx->irq_rx_moderation = rx_ticks;
1780 	efx_for_each_channel(channel, efx) {
1781 		if (efx_channel_has_rx_queue(channel))
1782 			channel->irq_moderation = rx_ticks;
1783 		else if (efx_channel_has_tx_queues(channel))
1784 			channel->irq_moderation = tx_ticks;
1785 	}
1786 
1787 	return 0;
1788 }
1789 
1790 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1791 			    unsigned int *rx_usecs, bool *rx_adaptive)
1792 {
1793 	/* We must round up when converting ticks to microseconds
1794 	 * because we round down when converting the other way.
1795 	 */
1796 
1797 	*rx_adaptive = efx->irq_rx_adaptive;
1798 	*rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1799 				 efx->timer_quantum_ns,
1800 				 1000);
1801 
1802 	/* If channels are shared between RX and TX, so is IRQ
1803 	 * moderation.  Otherwise, IRQ moderation is the same for all
1804 	 * TX channels and is not adaptive.
1805 	 */
1806 	if (efx->tx_channel_offset == 0)
1807 		*tx_usecs = *rx_usecs;
1808 	else
1809 		*tx_usecs = DIV_ROUND_UP(
1810 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1811 			efx->timer_quantum_ns,
1812 			1000);
1813 }
1814 
1815 /**************************************************************************
1816  *
1817  * Hardware monitor
1818  *
1819  **************************************************************************/
1820 
1821 /* Run periodically off the general workqueue */
1822 static void efx_monitor(struct work_struct *data)
1823 {
1824 	struct efx_nic *efx = container_of(data, struct efx_nic,
1825 					   monitor_work.work);
1826 
1827 	netif_vdbg(efx, timer, efx->net_dev,
1828 		   "hardware monitor executing on CPU %d\n",
1829 		   raw_smp_processor_id());
1830 	BUG_ON(efx->type->monitor == NULL);
1831 
1832 	/* If the mac_lock is already held then it is likely a port
1833 	 * reconfiguration is already in place, which will likely do
1834 	 * most of the work of monitor() anyway. */
1835 	if (mutex_trylock(&efx->mac_lock)) {
1836 		if (efx->port_enabled)
1837 			efx->type->monitor(efx);
1838 		mutex_unlock(&efx->mac_lock);
1839 	}
1840 
1841 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1842 			   efx_monitor_interval);
1843 }
1844 
1845 /**************************************************************************
1846  *
1847  * ioctls
1848  *
1849  *************************************************************************/
1850 
1851 /* Net device ioctl
1852  * Context: process, rtnl_lock() held.
1853  */
1854 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1855 {
1856 	struct efx_nic *efx = netdev_priv(net_dev);
1857 	struct mii_ioctl_data *data = if_mii(ifr);
1858 
1859 	if (cmd == SIOCSHWTSTAMP)
1860 		return efx_ptp_ioctl(efx, ifr, cmd);
1861 
1862 	/* Convert phy_id from older PRTAD/DEVAD format */
1863 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1864 	    (data->phy_id & 0xfc00) == 0x0400)
1865 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1866 
1867 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
1868 }
1869 
1870 /**************************************************************************
1871  *
1872  * NAPI interface
1873  *
1874  **************************************************************************/
1875 
1876 static void efx_init_napi_channel(struct efx_channel *channel)
1877 {
1878 	struct efx_nic *efx = channel->efx;
1879 
1880 	channel->napi_dev = efx->net_dev;
1881 	netif_napi_add(channel->napi_dev, &channel->napi_str,
1882 		       efx_poll, napi_weight);
1883 }
1884 
1885 static void efx_init_napi(struct efx_nic *efx)
1886 {
1887 	struct efx_channel *channel;
1888 
1889 	efx_for_each_channel(channel, efx)
1890 		efx_init_napi_channel(channel);
1891 }
1892 
1893 static void efx_fini_napi_channel(struct efx_channel *channel)
1894 {
1895 	if (channel->napi_dev)
1896 		netif_napi_del(&channel->napi_str);
1897 	channel->napi_dev = NULL;
1898 }
1899 
1900 static void efx_fini_napi(struct efx_nic *efx)
1901 {
1902 	struct efx_channel *channel;
1903 
1904 	efx_for_each_channel(channel, efx)
1905 		efx_fini_napi_channel(channel);
1906 }
1907 
1908 /**************************************************************************
1909  *
1910  * Kernel netpoll interface
1911  *
1912  *************************************************************************/
1913 
1914 #ifdef CONFIG_NET_POLL_CONTROLLER
1915 
1916 /* Although in the common case interrupts will be disabled, this is not
1917  * guaranteed. However, all our work happens inside the NAPI callback,
1918  * so no locking is required.
1919  */
1920 static void efx_netpoll(struct net_device *net_dev)
1921 {
1922 	struct efx_nic *efx = netdev_priv(net_dev);
1923 	struct efx_channel *channel;
1924 
1925 	efx_for_each_channel(channel, efx)
1926 		efx_schedule_channel(channel);
1927 }
1928 
1929 #endif
1930 
1931 /**************************************************************************
1932  *
1933  * Kernel net device interface
1934  *
1935  *************************************************************************/
1936 
1937 /* Context: process, rtnl_lock() held. */
1938 static int efx_net_open(struct net_device *net_dev)
1939 {
1940 	struct efx_nic *efx = netdev_priv(net_dev);
1941 	int rc;
1942 
1943 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
1944 		  raw_smp_processor_id());
1945 
1946 	rc = efx_check_disabled(efx);
1947 	if (rc)
1948 		return rc;
1949 	if (efx->phy_mode & PHY_MODE_SPECIAL)
1950 		return -EBUSY;
1951 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
1952 		return -EIO;
1953 
1954 	/* Notify the kernel of the link state polled during driver load,
1955 	 * before the monitor starts running */
1956 	efx_link_status_changed(efx);
1957 
1958 	efx_start_all(efx);
1959 	efx_selftest_async_start(efx);
1960 	return 0;
1961 }
1962 
1963 /* Context: process, rtnl_lock() held.
1964  * Note that the kernel will ignore our return code; this method
1965  * should really be a void.
1966  */
1967 static int efx_net_stop(struct net_device *net_dev)
1968 {
1969 	struct efx_nic *efx = netdev_priv(net_dev);
1970 
1971 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
1972 		  raw_smp_processor_id());
1973 
1974 	/* Stop the device and flush all the channels */
1975 	efx_stop_all(efx);
1976 
1977 	return 0;
1978 }
1979 
1980 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1981 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
1982 					       struct rtnl_link_stats64 *stats)
1983 {
1984 	struct efx_nic *efx = netdev_priv(net_dev);
1985 
1986 	spin_lock_bh(&efx->stats_lock);
1987 	efx->type->update_stats(efx, NULL, stats);
1988 	spin_unlock_bh(&efx->stats_lock);
1989 
1990 	return stats;
1991 }
1992 
1993 /* Context: netif_tx_lock held, BHs disabled. */
1994 static void efx_watchdog(struct net_device *net_dev)
1995 {
1996 	struct efx_nic *efx = netdev_priv(net_dev);
1997 
1998 	netif_err(efx, tx_err, efx->net_dev,
1999 		  "TX stuck with port_enabled=%d: resetting channels\n",
2000 		  efx->port_enabled);
2001 
2002 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2003 }
2004 
2005 
2006 /* Context: process, rtnl_lock() held. */
2007 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2008 {
2009 	struct efx_nic *efx = netdev_priv(net_dev);
2010 	int rc;
2011 
2012 	rc = efx_check_disabled(efx);
2013 	if (rc)
2014 		return rc;
2015 	if (new_mtu > EFX_MAX_MTU)
2016 		return -EINVAL;
2017 
2018 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2019 
2020 	efx_device_detach_sync(efx);
2021 	efx_stop_all(efx);
2022 
2023 	mutex_lock(&efx->mac_lock);
2024 	net_dev->mtu = new_mtu;
2025 	efx->type->reconfigure_mac(efx);
2026 	mutex_unlock(&efx->mac_lock);
2027 
2028 	efx_start_all(efx);
2029 	netif_device_attach(efx->net_dev);
2030 	return 0;
2031 }
2032 
2033 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2034 {
2035 	struct efx_nic *efx = netdev_priv(net_dev);
2036 	struct sockaddr *addr = data;
2037 	char *new_addr = addr->sa_data;
2038 
2039 	if (!is_valid_ether_addr(new_addr)) {
2040 		netif_err(efx, drv, efx->net_dev,
2041 			  "invalid ethernet MAC address requested: %pM\n",
2042 			  new_addr);
2043 		return -EADDRNOTAVAIL;
2044 	}
2045 
2046 	memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
2047 	efx_sriov_mac_address_changed(efx);
2048 
2049 	/* Reconfigure the MAC */
2050 	mutex_lock(&efx->mac_lock);
2051 	efx->type->reconfigure_mac(efx);
2052 	mutex_unlock(&efx->mac_lock);
2053 
2054 	return 0;
2055 }
2056 
2057 /* Context: netif_addr_lock held, BHs disabled. */
2058 static void efx_set_rx_mode(struct net_device *net_dev)
2059 {
2060 	struct efx_nic *efx = netdev_priv(net_dev);
2061 
2062 	if (efx->port_enabled)
2063 		queue_work(efx->workqueue, &efx->mac_work);
2064 	/* Otherwise efx_start_port() will do this */
2065 }
2066 
2067 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2068 {
2069 	struct efx_nic *efx = netdev_priv(net_dev);
2070 
2071 	/* If disabling RX n-tuple filtering, clear existing filters */
2072 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
2073 		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2074 
2075 	return 0;
2076 }
2077 
2078 static const struct net_device_ops efx_farch_netdev_ops = {
2079 	.ndo_open		= efx_net_open,
2080 	.ndo_stop		= efx_net_stop,
2081 	.ndo_get_stats64	= efx_net_stats,
2082 	.ndo_tx_timeout		= efx_watchdog,
2083 	.ndo_start_xmit		= efx_hard_start_xmit,
2084 	.ndo_validate_addr	= eth_validate_addr,
2085 	.ndo_do_ioctl		= efx_ioctl,
2086 	.ndo_change_mtu		= efx_change_mtu,
2087 	.ndo_set_mac_address	= efx_set_mac_address,
2088 	.ndo_set_rx_mode	= efx_set_rx_mode,
2089 	.ndo_set_features	= efx_set_features,
2090 #ifdef CONFIG_SFC_SRIOV
2091 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2092 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2093 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2094 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2095 #endif
2096 #ifdef CONFIG_NET_POLL_CONTROLLER
2097 	.ndo_poll_controller = efx_netpoll,
2098 #endif
2099 	.ndo_setup_tc		= efx_setup_tc,
2100 #ifdef CONFIG_RFS_ACCEL
2101 	.ndo_rx_flow_steer	= efx_filter_rfs,
2102 #endif
2103 };
2104 
2105 static const struct net_device_ops efx_ef10_netdev_ops = {
2106 	.ndo_open		= efx_net_open,
2107 	.ndo_stop		= efx_net_stop,
2108 	.ndo_get_stats64	= efx_net_stats,
2109 	.ndo_tx_timeout		= efx_watchdog,
2110 	.ndo_start_xmit		= efx_hard_start_xmit,
2111 	.ndo_validate_addr	= eth_validate_addr,
2112 	.ndo_do_ioctl		= efx_ioctl,
2113 	.ndo_change_mtu		= efx_change_mtu,
2114 	.ndo_set_mac_address	= efx_set_mac_address,
2115 	.ndo_set_rx_mode	= efx_set_rx_mode,
2116 	.ndo_set_features	= efx_set_features,
2117 #ifdef CONFIG_NET_POLL_CONTROLLER
2118 	.ndo_poll_controller	= efx_netpoll,
2119 #endif
2120 #ifdef CONFIG_RFS_ACCEL
2121 	.ndo_rx_flow_steer	= efx_filter_rfs,
2122 #endif
2123 };
2124 
2125 static void efx_update_name(struct efx_nic *efx)
2126 {
2127 	strcpy(efx->name, efx->net_dev->name);
2128 	efx_mtd_rename(efx);
2129 	efx_set_channel_names(efx);
2130 }
2131 
2132 static int efx_netdev_event(struct notifier_block *this,
2133 			    unsigned long event, void *ptr)
2134 {
2135 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2136 
2137 	if ((net_dev->netdev_ops == &efx_farch_netdev_ops ||
2138 	     net_dev->netdev_ops == &efx_ef10_netdev_ops) &&
2139 	    event == NETDEV_CHANGENAME)
2140 		efx_update_name(netdev_priv(net_dev));
2141 
2142 	return NOTIFY_DONE;
2143 }
2144 
2145 static struct notifier_block efx_netdev_notifier = {
2146 	.notifier_call = efx_netdev_event,
2147 };
2148 
2149 static ssize_t
2150 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2151 {
2152 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2153 	return sprintf(buf, "%d\n", efx->phy_type);
2154 }
2155 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2156 
2157 static int efx_register_netdev(struct efx_nic *efx)
2158 {
2159 	struct net_device *net_dev = efx->net_dev;
2160 	struct efx_channel *channel;
2161 	int rc;
2162 
2163 	net_dev->watchdog_timeo = 5 * HZ;
2164 	net_dev->irq = efx->pci_dev->irq;
2165 	if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0) {
2166 		net_dev->netdev_ops = &efx_ef10_netdev_ops;
2167 		net_dev->priv_flags |= IFF_UNICAST_FLT;
2168 	} else {
2169 		net_dev->netdev_ops = &efx_farch_netdev_ops;
2170 	}
2171 	SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
2172 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2173 
2174 	rtnl_lock();
2175 
2176 	/* Enable resets to be scheduled and check whether any were
2177 	 * already requested.  If so, the NIC is probably hosed so we
2178 	 * abort.
2179 	 */
2180 	efx->state = STATE_READY;
2181 	smp_mb(); /* ensure we change state before checking reset_pending */
2182 	if (efx->reset_pending) {
2183 		netif_err(efx, probe, efx->net_dev,
2184 			  "aborting probe due to scheduled reset\n");
2185 		rc = -EIO;
2186 		goto fail_locked;
2187 	}
2188 
2189 	rc = dev_alloc_name(net_dev, net_dev->name);
2190 	if (rc < 0)
2191 		goto fail_locked;
2192 	efx_update_name(efx);
2193 
2194 	/* Always start with carrier off; PHY events will detect the link */
2195 	netif_carrier_off(net_dev);
2196 
2197 	rc = register_netdevice(net_dev);
2198 	if (rc)
2199 		goto fail_locked;
2200 
2201 	efx_for_each_channel(channel, efx) {
2202 		struct efx_tx_queue *tx_queue;
2203 		efx_for_each_channel_tx_queue(tx_queue, channel)
2204 			efx_init_tx_queue_core_txq(tx_queue);
2205 	}
2206 
2207 	rtnl_unlock();
2208 
2209 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2210 	if (rc) {
2211 		netif_err(efx, drv, efx->net_dev,
2212 			  "failed to init net dev attributes\n");
2213 		goto fail_registered;
2214 	}
2215 
2216 	return 0;
2217 
2218 fail_registered:
2219 	rtnl_lock();
2220 	unregister_netdevice(net_dev);
2221 fail_locked:
2222 	efx->state = STATE_UNINIT;
2223 	rtnl_unlock();
2224 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2225 	return rc;
2226 }
2227 
2228 static void efx_unregister_netdev(struct efx_nic *efx)
2229 {
2230 	if (!efx->net_dev)
2231 		return;
2232 
2233 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2234 
2235 	strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2236 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2237 
2238 	rtnl_lock();
2239 	unregister_netdevice(efx->net_dev);
2240 	efx->state = STATE_UNINIT;
2241 	rtnl_unlock();
2242 }
2243 
2244 /**************************************************************************
2245  *
2246  * Device reset and suspend
2247  *
2248  **************************************************************************/
2249 
2250 /* Tears down the entire software state and most of the hardware state
2251  * before reset.  */
2252 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2253 {
2254 	EFX_ASSERT_RESET_SERIALISED(efx);
2255 
2256 	efx_stop_all(efx);
2257 	efx_disable_interrupts(efx);
2258 
2259 	mutex_lock(&efx->mac_lock);
2260 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2261 		efx->phy_op->fini(efx);
2262 	efx->type->fini(efx);
2263 }
2264 
2265 /* This function will always ensure that the locks acquired in
2266  * efx_reset_down() are released. A failure return code indicates
2267  * that we were unable to reinitialise the hardware, and the
2268  * driver should be disabled. If ok is false, then the rx and tx
2269  * engines are not restarted, pending a RESET_DISABLE. */
2270 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2271 {
2272 	int rc;
2273 
2274 	EFX_ASSERT_RESET_SERIALISED(efx);
2275 
2276 	rc = efx->type->init(efx);
2277 	if (rc) {
2278 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2279 		goto fail;
2280 	}
2281 
2282 	if (!ok)
2283 		goto fail;
2284 
2285 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2286 		rc = efx->phy_op->init(efx);
2287 		if (rc)
2288 			goto fail;
2289 		if (efx->phy_op->reconfigure(efx))
2290 			netif_err(efx, drv, efx->net_dev,
2291 				  "could not restore PHY settings\n");
2292 	}
2293 
2294 	rc = efx_enable_interrupts(efx);
2295 	if (rc)
2296 		goto fail;
2297 	efx_restore_filters(efx);
2298 	efx_sriov_reset(efx);
2299 
2300 	mutex_unlock(&efx->mac_lock);
2301 
2302 	efx_start_all(efx);
2303 
2304 	return 0;
2305 
2306 fail:
2307 	efx->port_initialized = false;
2308 
2309 	mutex_unlock(&efx->mac_lock);
2310 
2311 	return rc;
2312 }
2313 
2314 /* Reset the NIC using the specified method.  Note that the reset may
2315  * fail, in which case the card will be left in an unusable state.
2316  *
2317  * Caller must hold the rtnl_lock.
2318  */
2319 int efx_reset(struct efx_nic *efx, enum reset_type method)
2320 {
2321 	int rc, rc2;
2322 	bool disabled;
2323 
2324 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2325 		   RESET_TYPE(method));
2326 
2327 	efx_device_detach_sync(efx);
2328 	efx_reset_down(efx, method);
2329 
2330 	rc = efx->type->reset(efx, method);
2331 	if (rc) {
2332 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2333 		goto out;
2334 	}
2335 
2336 	/* Clear flags for the scopes we covered.  We assume the NIC and
2337 	 * driver are now quiescent so that there is no race here.
2338 	 */
2339 	efx->reset_pending &= -(1 << (method + 1));
2340 
2341 	/* Reinitialise bus-mastering, which may have been turned off before
2342 	 * the reset was scheduled. This is still appropriate, even in the
2343 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2344 	 * can respond to requests. */
2345 	pci_set_master(efx->pci_dev);
2346 
2347 out:
2348 	/* Leave device stopped if necessary */
2349 	disabled = rc ||
2350 		method == RESET_TYPE_DISABLE ||
2351 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2352 	rc2 = efx_reset_up(efx, method, !disabled);
2353 	if (rc2) {
2354 		disabled = true;
2355 		if (!rc)
2356 			rc = rc2;
2357 	}
2358 
2359 	if (disabled) {
2360 		dev_close(efx->net_dev);
2361 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2362 		efx->state = STATE_DISABLED;
2363 	} else {
2364 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2365 		netif_device_attach(efx->net_dev);
2366 	}
2367 	return rc;
2368 }
2369 
2370 /* Try recovery mechanisms.
2371  * For now only EEH is supported.
2372  * Returns 0 if the recovery mechanisms are unsuccessful.
2373  * Returns a non-zero value otherwise.
2374  */
2375 int efx_try_recovery(struct efx_nic *efx)
2376 {
2377 #ifdef CONFIG_EEH
2378 	/* A PCI error can occur and not be seen by EEH because nothing
2379 	 * happens on the PCI bus. In this case the driver may fail and
2380 	 * schedule a 'recover or reset', leading to this recovery handler.
2381 	 * Manually call the eeh failure check function.
2382 	 */
2383 	struct eeh_dev *eehdev =
2384 		of_node_to_eeh_dev(pci_device_to_OF_node(efx->pci_dev));
2385 
2386 	if (eeh_dev_check_failure(eehdev)) {
2387 		/* The EEH mechanisms will handle the error and reset the
2388 		 * device if necessary.
2389 		 */
2390 		return 1;
2391 	}
2392 #endif
2393 	return 0;
2394 }
2395 
2396 /* The worker thread exists so that code that cannot sleep can
2397  * schedule a reset for later.
2398  */
2399 static void efx_reset_work(struct work_struct *data)
2400 {
2401 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2402 	unsigned long pending;
2403 	enum reset_type method;
2404 
2405 	pending = ACCESS_ONCE(efx->reset_pending);
2406 	method = fls(pending) - 1;
2407 
2408 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2409 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2410 	    efx_try_recovery(efx))
2411 		return;
2412 
2413 	if (!pending)
2414 		return;
2415 
2416 	rtnl_lock();
2417 
2418 	/* We checked the state in efx_schedule_reset() but it may
2419 	 * have changed by now.  Now that we have the RTNL lock,
2420 	 * it cannot change again.
2421 	 */
2422 	if (efx->state == STATE_READY)
2423 		(void)efx_reset(efx, method);
2424 
2425 	rtnl_unlock();
2426 }
2427 
2428 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2429 {
2430 	enum reset_type method;
2431 
2432 	if (efx->state == STATE_RECOVERY) {
2433 		netif_dbg(efx, drv, efx->net_dev,
2434 			  "recovering: skip scheduling %s reset\n",
2435 			  RESET_TYPE(type));
2436 		return;
2437 	}
2438 
2439 	switch (type) {
2440 	case RESET_TYPE_INVISIBLE:
2441 	case RESET_TYPE_ALL:
2442 	case RESET_TYPE_RECOVER_OR_ALL:
2443 	case RESET_TYPE_WORLD:
2444 	case RESET_TYPE_DISABLE:
2445 	case RESET_TYPE_RECOVER_OR_DISABLE:
2446 		method = type;
2447 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2448 			  RESET_TYPE(method));
2449 		break;
2450 	default:
2451 		method = efx->type->map_reset_reason(type);
2452 		netif_dbg(efx, drv, efx->net_dev,
2453 			  "scheduling %s reset for %s\n",
2454 			  RESET_TYPE(method), RESET_TYPE(type));
2455 		break;
2456 	}
2457 
2458 	set_bit(method, &efx->reset_pending);
2459 	smp_mb(); /* ensure we change reset_pending before checking state */
2460 
2461 	/* If we're not READY then just leave the flags set as the cue
2462 	 * to abort probing or reschedule the reset later.
2463 	 */
2464 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2465 		return;
2466 
2467 	/* efx_process_channel() will no longer read events once a
2468 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2469 	efx_mcdi_mode_poll(efx);
2470 
2471 	queue_work(reset_workqueue, &efx->reset_work);
2472 }
2473 
2474 /**************************************************************************
2475  *
2476  * List of NICs we support
2477  *
2478  **************************************************************************/
2479 
2480 /* PCI device ID table */
2481 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2482 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2483 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2484 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2485 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2486 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2487 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2488 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2489 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2490 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2491 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2492 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903),  /* SFC9120 PF */
2493 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2494 	{0}			/* end of list */
2495 };
2496 
2497 /**************************************************************************
2498  *
2499  * Dummy PHY/MAC operations
2500  *
2501  * Can be used for some unimplemented operations
2502  * Needed so all function pointers are valid and do not have to be tested
2503  * before use
2504  *
2505  **************************************************************************/
2506 int efx_port_dummy_op_int(struct efx_nic *efx)
2507 {
2508 	return 0;
2509 }
2510 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2511 
2512 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2513 {
2514 	return false;
2515 }
2516 
2517 static const struct efx_phy_operations efx_dummy_phy_operations = {
2518 	.init		 = efx_port_dummy_op_int,
2519 	.reconfigure	 = efx_port_dummy_op_int,
2520 	.poll		 = efx_port_dummy_op_poll,
2521 	.fini		 = efx_port_dummy_op_void,
2522 };
2523 
2524 /**************************************************************************
2525  *
2526  * Data housekeeping
2527  *
2528  **************************************************************************/
2529 
2530 /* This zeroes out and then fills in the invariants in a struct
2531  * efx_nic (including all sub-structures).
2532  */
2533 static int efx_init_struct(struct efx_nic *efx,
2534 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2535 {
2536 	int i;
2537 
2538 	/* Initialise common structures */
2539 	spin_lock_init(&efx->biu_lock);
2540 #ifdef CONFIG_SFC_MTD
2541 	INIT_LIST_HEAD(&efx->mtd_list);
2542 #endif
2543 	INIT_WORK(&efx->reset_work, efx_reset_work);
2544 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2545 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2546 	efx->pci_dev = pci_dev;
2547 	efx->msg_enable = debug;
2548 	efx->state = STATE_UNINIT;
2549 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2550 
2551 	efx->net_dev = net_dev;
2552 	efx->rx_prefix_size = efx->type->rx_prefix_size;
2553 	efx->rx_packet_hash_offset =
2554 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2555 	spin_lock_init(&efx->stats_lock);
2556 	mutex_init(&efx->mac_lock);
2557 	efx->phy_op = &efx_dummy_phy_operations;
2558 	efx->mdio.dev = net_dev;
2559 	INIT_WORK(&efx->mac_work, efx_mac_work);
2560 	init_waitqueue_head(&efx->flush_wq);
2561 
2562 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2563 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2564 		if (!efx->channel[i])
2565 			goto fail;
2566 		efx->msi_context[i].efx = efx;
2567 		efx->msi_context[i].index = i;
2568 	}
2569 
2570 	/* Higher numbered interrupt modes are less capable! */
2571 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2572 				  interrupt_mode);
2573 
2574 	/* Would be good to use the net_dev name, but we're too early */
2575 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2576 		 pci_name(pci_dev));
2577 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2578 	if (!efx->workqueue)
2579 		goto fail;
2580 
2581 	return 0;
2582 
2583 fail:
2584 	efx_fini_struct(efx);
2585 	return -ENOMEM;
2586 }
2587 
2588 static void efx_fini_struct(struct efx_nic *efx)
2589 {
2590 	int i;
2591 
2592 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2593 		kfree(efx->channel[i]);
2594 
2595 	if (efx->workqueue) {
2596 		destroy_workqueue(efx->workqueue);
2597 		efx->workqueue = NULL;
2598 	}
2599 }
2600 
2601 /**************************************************************************
2602  *
2603  * PCI interface
2604  *
2605  **************************************************************************/
2606 
2607 /* Main body of final NIC shutdown code
2608  * This is called only at module unload (or hotplug removal).
2609  */
2610 static void efx_pci_remove_main(struct efx_nic *efx)
2611 {
2612 	/* Flush reset_work. It can no longer be scheduled since we
2613 	 * are not READY.
2614 	 */
2615 	BUG_ON(efx->state == STATE_READY);
2616 	cancel_work_sync(&efx->reset_work);
2617 
2618 	efx_disable_interrupts(efx);
2619 	efx_nic_fini_interrupt(efx);
2620 	efx_fini_port(efx);
2621 	efx->type->fini(efx);
2622 	efx_fini_napi(efx);
2623 	efx_remove_all(efx);
2624 }
2625 
2626 /* Final NIC shutdown
2627  * This is called only at module unload (or hotplug removal).
2628  */
2629 static void efx_pci_remove(struct pci_dev *pci_dev)
2630 {
2631 	struct efx_nic *efx;
2632 
2633 	efx = pci_get_drvdata(pci_dev);
2634 	if (!efx)
2635 		return;
2636 
2637 	/* Mark the NIC as fini, then stop the interface */
2638 	rtnl_lock();
2639 	dev_close(efx->net_dev);
2640 	efx_disable_interrupts(efx);
2641 	rtnl_unlock();
2642 
2643 	efx_sriov_fini(efx);
2644 	efx_unregister_netdev(efx);
2645 
2646 	efx_mtd_remove(efx);
2647 
2648 	efx_pci_remove_main(efx);
2649 
2650 	efx_fini_io(efx);
2651 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2652 
2653 	efx_fini_struct(efx);
2654 	pci_set_drvdata(pci_dev, NULL);
2655 	free_netdev(efx->net_dev);
2656 
2657 	pci_disable_pcie_error_reporting(pci_dev);
2658 };
2659 
2660 /* NIC VPD information
2661  * Called during probe to display the part number of the
2662  * installed NIC.  VPD is potentially very large but this should
2663  * always appear within the first 512 bytes.
2664  */
2665 #define SFC_VPD_LEN 512
2666 static void efx_print_product_vpd(struct efx_nic *efx)
2667 {
2668 	struct pci_dev *dev = efx->pci_dev;
2669 	char vpd_data[SFC_VPD_LEN];
2670 	ssize_t vpd_size;
2671 	int i, j;
2672 
2673 	/* Get the vpd data from the device */
2674 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2675 	if (vpd_size <= 0) {
2676 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2677 		return;
2678 	}
2679 
2680 	/* Get the Read only section */
2681 	i = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2682 	if (i < 0) {
2683 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2684 		return;
2685 	}
2686 
2687 	j = pci_vpd_lrdt_size(&vpd_data[i]);
2688 	i += PCI_VPD_LRDT_TAG_SIZE;
2689 	if (i + j > vpd_size)
2690 		j = vpd_size - i;
2691 
2692 	/* Get the Part number */
2693 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2694 	if (i < 0) {
2695 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2696 		return;
2697 	}
2698 
2699 	j = pci_vpd_info_field_size(&vpd_data[i]);
2700 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
2701 	if (i + j > vpd_size) {
2702 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2703 		return;
2704 	}
2705 
2706 	netif_info(efx, drv, efx->net_dev,
2707 		   "Part Number : %.*s\n", j, &vpd_data[i]);
2708 }
2709 
2710 
2711 /* Main body of NIC initialisation
2712  * This is called at module load (or hotplug insertion, theoretically).
2713  */
2714 static int efx_pci_probe_main(struct efx_nic *efx)
2715 {
2716 	int rc;
2717 
2718 	/* Do start-of-day initialisation */
2719 	rc = efx_probe_all(efx);
2720 	if (rc)
2721 		goto fail1;
2722 
2723 	efx_init_napi(efx);
2724 
2725 	rc = efx->type->init(efx);
2726 	if (rc) {
2727 		netif_err(efx, probe, efx->net_dev,
2728 			  "failed to initialise NIC\n");
2729 		goto fail3;
2730 	}
2731 
2732 	rc = efx_init_port(efx);
2733 	if (rc) {
2734 		netif_err(efx, probe, efx->net_dev,
2735 			  "failed to initialise port\n");
2736 		goto fail4;
2737 	}
2738 
2739 	rc = efx_nic_init_interrupt(efx);
2740 	if (rc)
2741 		goto fail5;
2742 	rc = efx_enable_interrupts(efx);
2743 	if (rc)
2744 		goto fail6;
2745 
2746 	return 0;
2747 
2748  fail6:
2749 	efx_nic_fini_interrupt(efx);
2750  fail5:
2751 	efx_fini_port(efx);
2752  fail4:
2753 	efx->type->fini(efx);
2754  fail3:
2755 	efx_fini_napi(efx);
2756 	efx_remove_all(efx);
2757  fail1:
2758 	return rc;
2759 }
2760 
2761 /* NIC initialisation
2762  *
2763  * This is called at module load (or hotplug insertion,
2764  * theoretically).  It sets up PCI mappings, resets the NIC,
2765  * sets up and registers the network devices with the kernel and hooks
2766  * the interrupt service routine.  It does not prepare the device for
2767  * transmission; this is left to the first time one of the network
2768  * interfaces is brought up (i.e. efx_net_open).
2769  */
2770 static int efx_pci_probe(struct pci_dev *pci_dev,
2771 			 const struct pci_device_id *entry)
2772 {
2773 	struct net_device *net_dev;
2774 	struct efx_nic *efx;
2775 	int rc;
2776 
2777 	/* Allocate and initialise a struct net_device and struct efx_nic */
2778 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2779 				     EFX_MAX_RX_QUEUES);
2780 	if (!net_dev)
2781 		return -ENOMEM;
2782 	efx = netdev_priv(net_dev);
2783 	efx->type = (const struct efx_nic_type *) entry->driver_data;
2784 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2785 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
2786 			      NETIF_F_RXCSUM);
2787 	if (efx->type->offload_features & NETIF_F_V6_CSUM)
2788 		net_dev->features |= NETIF_F_TSO6;
2789 	/* Mask for features that also apply to VLAN devices */
2790 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2791 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2792 				   NETIF_F_RXCSUM);
2793 	/* All offloads can be toggled */
2794 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2795 	pci_set_drvdata(pci_dev, efx);
2796 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2797 	rc = efx_init_struct(efx, pci_dev, net_dev);
2798 	if (rc)
2799 		goto fail1;
2800 
2801 	netif_info(efx, probe, efx->net_dev,
2802 		   "Solarflare NIC detected\n");
2803 
2804 	efx_print_product_vpd(efx);
2805 
2806 	/* Set up basic I/O (BAR mappings etc) */
2807 	rc = efx_init_io(efx);
2808 	if (rc)
2809 		goto fail2;
2810 
2811 	rc = efx_pci_probe_main(efx);
2812 	if (rc)
2813 		goto fail3;
2814 
2815 	rc = efx_register_netdev(efx);
2816 	if (rc)
2817 		goto fail4;
2818 
2819 	rc = efx_sriov_init(efx);
2820 	if (rc)
2821 		netif_err(efx, probe, efx->net_dev,
2822 			  "SR-IOV can't be enabled rc %d\n", rc);
2823 
2824 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2825 
2826 	/* Try to create MTDs, but allow this to fail */
2827 	rtnl_lock();
2828 	rc = efx_mtd_probe(efx);
2829 	rtnl_unlock();
2830 	if (rc)
2831 		netif_warn(efx, probe, efx->net_dev,
2832 			   "failed to create MTDs (%d)\n", rc);
2833 
2834 	rc = pci_enable_pcie_error_reporting(pci_dev);
2835 	if (rc && rc != -EINVAL)
2836 		netif_warn(efx, probe, efx->net_dev,
2837 			   "pci_enable_pcie_error_reporting failed (%d)\n", rc);
2838 
2839 	return 0;
2840 
2841  fail4:
2842 	efx_pci_remove_main(efx);
2843  fail3:
2844 	efx_fini_io(efx);
2845  fail2:
2846 	efx_fini_struct(efx);
2847  fail1:
2848 	pci_set_drvdata(pci_dev, NULL);
2849 	WARN_ON(rc > 0);
2850 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2851 	free_netdev(net_dev);
2852 	return rc;
2853 }
2854 
2855 static int efx_pm_freeze(struct device *dev)
2856 {
2857 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2858 
2859 	rtnl_lock();
2860 
2861 	if (efx->state != STATE_DISABLED) {
2862 		efx->state = STATE_UNINIT;
2863 
2864 		efx_device_detach_sync(efx);
2865 
2866 		efx_stop_all(efx);
2867 		efx_disable_interrupts(efx);
2868 	}
2869 
2870 	rtnl_unlock();
2871 
2872 	return 0;
2873 }
2874 
2875 static int efx_pm_thaw(struct device *dev)
2876 {
2877 	int rc;
2878 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2879 
2880 	rtnl_lock();
2881 
2882 	if (efx->state != STATE_DISABLED) {
2883 		rc = efx_enable_interrupts(efx);
2884 		if (rc)
2885 			goto fail;
2886 
2887 		mutex_lock(&efx->mac_lock);
2888 		efx->phy_op->reconfigure(efx);
2889 		mutex_unlock(&efx->mac_lock);
2890 
2891 		efx_start_all(efx);
2892 
2893 		netif_device_attach(efx->net_dev);
2894 
2895 		efx->state = STATE_READY;
2896 
2897 		efx->type->resume_wol(efx);
2898 	}
2899 
2900 	rtnl_unlock();
2901 
2902 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2903 	queue_work(reset_workqueue, &efx->reset_work);
2904 
2905 	return 0;
2906 
2907 fail:
2908 	rtnl_unlock();
2909 
2910 	return rc;
2911 }
2912 
2913 static int efx_pm_poweroff(struct device *dev)
2914 {
2915 	struct pci_dev *pci_dev = to_pci_dev(dev);
2916 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2917 
2918 	efx->type->fini(efx);
2919 
2920 	efx->reset_pending = 0;
2921 
2922 	pci_save_state(pci_dev);
2923 	return pci_set_power_state(pci_dev, PCI_D3hot);
2924 }
2925 
2926 /* Used for both resume and restore */
2927 static int efx_pm_resume(struct device *dev)
2928 {
2929 	struct pci_dev *pci_dev = to_pci_dev(dev);
2930 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2931 	int rc;
2932 
2933 	rc = pci_set_power_state(pci_dev, PCI_D0);
2934 	if (rc)
2935 		return rc;
2936 	pci_restore_state(pci_dev);
2937 	rc = pci_enable_device(pci_dev);
2938 	if (rc)
2939 		return rc;
2940 	pci_set_master(efx->pci_dev);
2941 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
2942 	if (rc)
2943 		return rc;
2944 	rc = efx->type->init(efx);
2945 	if (rc)
2946 		return rc;
2947 	rc = efx_pm_thaw(dev);
2948 	return rc;
2949 }
2950 
2951 static int efx_pm_suspend(struct device *dev)
2952 {
2953 	int rc;
2954 
2955 	efx_pm_freeze(dev);
2956 	rc = efx_pm_poweroff(dev);
2957 	if (rc)
2958 		efx_pm_resume(dev);
2959 	return rc;
2960 }
2961 
2962 static const struct dev_pm_ops efx_pm_ops = {
2963 	.suspend	= efx_pm_suspend,
2964 	.resume		= efx_pm_resume,
2965 	.freeze		= efx_pm_freeze,
2966 	.thaw		= efx_pm_thaw,
2967 	.poweroff	= efx_pm_poweroff,
2968 	.restore	= efx_pm_resume,
2969 };
2970 
2971 /* A PCI error affecting this device was detected.
2972  * At this point MMIO and DMA may be disabled.
2973  * Stop the software path and request a slot reset.
2974  */
2975 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
2976 					      enum pci_channel_state state)
2977 {
2978 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
2979 	struct efx_nic *efx = pci_get_drvdata(pdev);
2980 
2981 	if (state == pci_channel_io_perm_failure)
2982 		return PCI_ERS_RESULT_DISCONNECT;
2983 
2984 	rtnl_lock();
2985 
2986 	if (efx->state != STATE_DISABLED) {
2987 		efx->state = STATE_RECOVERY;
2988 		efx->reset_pending = 0;
2989 
2990 		efx_device_detach_sync(efx);
2991 
2992 		efx_stop_all(efx);
2993 		efx_disable_interrupts(efx);
2994 
2995 		status = PCI_ERS_RESULT_NEED_RESET;
2996 	} else {
2997 		/* If the interface is disabled we don't want to do anything
2998 		 * with it.
2999 		 */
3000 		status = PCI_ERS_RESULT_RECOVERED;
3001 	}
3002 
3003 	rtnl_unlock();
3004 
3005 	pci_disable_device(pdev);
3006 
3007 	return status;
3008 }
3009 
3010 /* Fake a successfull reset, which will be performed later in efx_io_resume. */
3011 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3012 {
3013 	struct efx_nic *efx = pci_get_drvdata(pdev);
3014 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3015 	int rc;
3016 
3017 	if (pci_enable_device(pdev)) {
3018 		netif_err(efx, hw, efx->net_dev,
3019 			  "Cannot re-enable PCI device after reset.\n");
3020 		status =  PCI_ERS_RESULT_DISCONNECT;
3021 	}
3022 
3023 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3024 	if (rc) {
3025 		netif_err(efx, hw, efx->net_dev,
3026 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3027 		/* Non-fatal error. Continue. */
3028 	}
3029 
3030 	return status;
3031 }
3032 
3033 /* Perform the actual reset and resume I/O operations. */
3034 static void efx_io_resume(struct pci_dev *pdev)
3035 {
3036 	struct efx_nic *efx = pci_get_drvdata(pdev);
3037 	int rc;
3038 
3039 	rtnl_lock();
3040 
3041 	if (efx->state == STATE_DISABLED)
3042 		goto out;
3043 
3044 	rc = efx_reset(efx, RESET_TYPE_ALL);
3045 	if (rc) {
3046 		netif_err(efx, hw, efx->net_dev,
3047 			  "efx_reset failed after PCI error (%d)\n", rc);
3048 	} else {
3049 		efx->state = STATE_READY;
3050 		netif_dbg(efx, hw, efx->net_dev,
3051 			  "Done resetting and resuming IO after PCI error.\n");
3052 	}
3053 
3054 out:
3055 	rtnl_unlock();
3056 }
3057 
3058 /* For simplicity and reliability, we always require a slot reset and try to
3059  * reset the hardware when a pci error affecting the device is detected.
3060  * We leave both the link_reset and mmio_enabled callback unimplemented:
3061  * with our request for slot reset the mmio_enabled callback will never be
3062  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3063  */
3064 static struct pci_error_handlers efx_err_handlers = {
3065 	.error_detected = efx_io_error_detected,
3066 	.slot_reset	= efx_io_slot_reset,
3067 	.resume		= efx_io_resume,
3068 };
3069 
3070 static struct pci_driver efx_pci_driver = {
3071 	.name		= KBUILD_MODNAME,
3072 	.id_table	= efx_pci_table,
3073 	.probe		= efx_pci_probe,
3074 	.remove		= efx_pci_remove,
3075 	.driver.pm	= &efx_pm_ops,
3076 	.err_handler	= &efx_err_handlers,
3077 };
3078 
3079 /**************************************************************************
3080  *
3081  * Kernel module interface
3082  *
3083  *************************************************************************/
3084 
3085 module_param(interrupt_mode, uint, 0444);
3086 MODULE_PARM_DESC(interrupt_mode,
3087 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3088 
3089 static int __init efx_init_module(void)
3090 {
3091 	int rc;
3092 
3093 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3094 
3095 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3096 	if (rc)
3097 		goto err_notifier;
3098 
3099 	rc = efx_init_sriov();
3100 	if (rc)
3101 		goto err_sriov;
3102 
3103 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3104 	if (!reset_workqueue) {
3105 		rc = -ENOMEM;
3106 		goto err_reset;
3107 	}
3108 
3109 	rc = pci_register_driver(&efx_pci_driver);
3110 	if (rc < 0)
3111 		goto err_pci;
3112 
3113 	return 0;
3114 
3115  err_pci:
3116 	destroy_workqueue(reset_workqueue);
3117  err_reset:
3118 	efx_fini_sriov();
3119  err_sriov:
3120 	unregister_netdevice_notifier(&efx_netdev_notifier);
3121  err_notifier:
3122 	return rc;
3123 }
3124 
3125 static void __exit efx_exit_module(void)
3126 {
3127 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3128 
3129 	pci_unregister_driver(&efx_pci_driver);
3130 	destroy_workqueue(reset_workqueue);
3131 	efx_fini_sriov();
3132 	unregister_netdevice_notifier(&efx_netdev_notifier);
3133 
3134 }
3135 
3136 module_init(efx_init_module);
3137 module_exit(efx_exit_module);
3138 
3139 MODULE_AUTHOR("Solarflare Communications and "
3140 	      "Michael Brown <mbrown@fensystems.co.uk>");
3141 MODULE_DESCRIPTION("Solarflare Communications network driver");
3142 MODULE_LICENSE("GPL");
3143 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3144