xref: /linux/drivers/net/ethernet/sfc/efx.c (revision 005438a8eef063495ac059d128eea71b58de50e5)
1 /****************************************************************************
2  * Driver for Solarflare network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2013 Solarflare Communications Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published
8  * by the Free Software Foundation, incorporated herein by reference.
9  */
10 
11 #include <linux/module.h>
12 #include <linux/pci.h>
13 #include <linux/netdevice.h>
14 #include <linux/etherdevice.h>
15 #include <linux/delay.h>
16 #include <linux/notifier.h>
17 #include <linux/ip.h>
18 #include <linux/tcp.h>
19 #include <linux/in.h>
20 #include <linux/ethtool.h>
21 #include <linux/topology.h>
22 #include <linux/gfp.h>
23 #include <linux/aer.h>
24 #include <linux/interrupt.h>
25 #include "net_driver.h"
26 #include "efx.h"
27 #include "nic.h"
28 #include "selftest.h"
29 #include "sriov.h"
30 
31 #include "mcdi.h"
32 #include "workarounds.h"
33 
34 /**************************************************************************
35  *
36  * Type name strings
37  *
38  **************************************************************************
39  */
40 
41 /* Loopback mode names (see LOOPBACK_MODE()) */
42 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
43 const char *const efx_loopback_mode_names[] = {
44 	[LOOPBACK_NONE]		= "NONE",
45 	[LOOPBACK_DATA]		= "DATAPATH",
46 	[LOOPBACK_GMAC]		= "GMAC",
47 	[LOOPBACK_XGMII]	= "XGMII",
48 	[LOOPBACK_XGXS]		= "XGXS",
49 	[LOOPBACK_XAUI]		= "XAUI",
50 	[LOOPBACK_GMII]		= "GMII",
51 	[LOOPBACK_SGMII]	= "SGMII",
52 	[LOOPBACK_XGBR]		= "XGBR",
53 	[LOOPBACK_XFI]		= "XFI",
54 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
55 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
56 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
57 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
58 	[LOOPBACK_GPHY]		= "GPHY",
59 	[LOOPBACK_PHYXS]	= "PHYXS",
60 	[LOOPBACK_PCS]		= "PCS",
61 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
62 	[LOOPBACK_XPORT]	= "XPORT",
63 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
64 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
65 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
66 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
67 	[LOOPBACK_GMII_WS]	= "GMII_WS",
68 	[LOOPBACK_XFI_WS]	= "XFI_WS",
69 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
70 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
71 };
72 
73 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
74 const char *const efx_reset_type_names[] = {
75 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
76 	[RESET_TYPE_ALL]                = "ALL",
77 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
78 	[RESET_TYPE_WORLD]              = "WORLD",
79 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
80 	[RESET_TYPE_DATAPATH]           = "DATAPATH",
81 	[RESET_TYPE_MC_BIST]		= "MC_BIST",
82 	[RESET_TYPE_DISABLE]            = "DISABLE",
83 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
84 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
85 	[RESET_TYPE_RX_RECOVERY]        = "RX_RECOVERY",
86 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
87 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
88 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
89 	[RESET_TYPE_MCDI_TIMEOUT]	= "MCDI_TIMEOUT (FLR)",
90 };
91 
92 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
93  * queued onto this work queue. This is not a per-nic work queue, because
94  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
95  */
96 static struct workqueue_struct *reset_workqueue;
97 
98 /* How often and how many times to poll for a reset while waiting for a
99  * BIST that another function started to complete.
100  */
101 #define BIST_WAIT_DELAY_MS	100
102 #define BIST_WAIT_DELAY_COUNT	100
103 
104 /**************************************************************************
105  *
106  * Configurable values
107  *
108  *************************************************************************/
109 
110 /*
111  * Use separate channels for TX and RX events
112  *
113  * Set this to 1 to use separate channels for TX and RX. It allows us
114  * to control interrupt affinity separately for TX and RX.
115  *
116  * This is only used in MSI-X interrupt mode
117  */
118 static bool separate_tx_channels;
119 module_param(separate_tx_channels, bool, 0444);
120 MODULE_PARM_DESC(separate_tx_channels,
121 		 "Use separate channels for TX and RX");
122 
123 /* This is the weight assigned to each of the (per-channel) virtual
124  * NAPI devices.
125  */
126 static int napi_weight = 64;
127 
128 /* This is the time (in jiffies) between invocations of the hardware
129  * monitor.
130  * On Falcon-based NICs, this will:
131  * - Check the on-board hardware monitor;
132  * - Poll the link state and reconfigure the hardware as necessary.
133  * On Siena-based NICs for power systems with EEH support, this will give EEH a
134  * chance to start.
135  */
136 static unsigned int efx_monitor_interval = 1 * HZ;
137 
138 /* Initial interrupt moderation settings.  They can be modified after
139  * module load with ethtool.
140  *
141  * The default for RX should strike a balance between increasing the
142  * round-trip latency and reducing overhead.
143  */
144 static unsigned int rx_irq_mod_usec = 60;
145 
146 /* Initial interrupt moderation settings.  They can be modified after
147  * module load with ethtool.
148  *
149  * This default is chosen to ensure that a 10G link does not go idle
150  * while a TX queue is stopped after it has become full.  A queue is
151  * restarted when it drops below half full.  The time this takes (assuming
152  * worst case 3 descriptors per packet and 1024 descriptors) is
153  *   512 / 3 * 1.2 = 205 usec.
154  */
155 static unsigned int tx_irq_mod_usec = 150;
156 
157 /* This is the first interrupt mode to try out of:
158  * 0 => MSI-X
159  * 1 => MSI
160  * 2 => legacy
161  */
162 static unsigned int interrupt_mode;
163 
164 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
165  * i.e. the number of CPUs among which we may distribute simultaneous
166  * interrupt handling.
167  *
168  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
169  * The default (0) means to assign an interrupt to each core.
170  */
171 static unsigned int rss_cpus;
172 module_param(rss_cpus, uint, 0444);
173 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
174 
175 static bool phy_flash_cfg;
176 module_param(phy_flash_cfg, bool, 0644);
177 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
178 
179 static unsigned irq_adapt_low_thresh = 8000;
180 module_param(irq_adapt_low_thresh, uint, 0644);
181 MODULE_PARM_DESC(irq_adapt_low_thresh,
182 		 "Threshold score for reducing IRQ moderation");
183 
184 static unsigned irq_adapt_high_thresh = 16000;
185 module_param(irq_adapt_high_thresh, uint, 0644);
186 MODULE_PARM_DESC(irq_adapt_high_thresh,
187 		 "Threshold score for increasing IRQ moderation");
188 
189 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
190 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
191 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
192 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
193 module_param(debug, uint, 0);
194 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
195 
196 /**************************************************************************
197  *
198  * Utility functions and prototypes
199  *
200  *************************************************************************/
201 
202 static int efx_soft_enable_interrupts(struct efx_nic *efx);
203 static void efx_soft_disable_interrupts(struct efx_nic *efx);
204 static void efx_remove_channel(struct efx_channel *channel);
205 static void efx_remove_channels(struct efx_nic *efx);
206 static const struct efx_channel_type efx_default_channel_type;
207 static void efx_remove_port(struct efx_nic *efx);
208 static void efx_init_napi_channel(struct efx_channel *channel);
209 static void efx_fini_napi(struct efx_nic *efx);
210 static void efx_fini_napi_channel(struct efx_channel *channel);
211 static void efx_fini_struct(struct efx_nic *efx);
212 static void efx_start_all(struct efx_nic *efx);
213 static void efx_stop_all(struct efx_nic *efx);
214 
215 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
216 	do {						\
217 		if ((efx->state == STATE_READY) ||	\
218 		    (efx->state == STATE_RECOVERY) ||	\
219 		    (efx->state == STATE_DISABLED))	\
220 			ASSERT_RTNL();			\
221 	} while (0)
222 
223 static int efx_check_disabled(struct efx_nic *efx)
224 {
225 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
226 		netif_err(efx, drv, efx->net_dev,
227 			  "device is disabled due to earlier errors\n");
228 		return -EIO;
229 	}
230 	return 0;
231 }
232 
233 /**************************************************************************
234  *
235  * Event queue processing
236  *
237  *************************************************************************/
238 
239 /* Process channel's event queue
240  *
241  * This function is responsible for processing the event queue of a
242  * single channel.  The caller must guarantee that this function will
243  * never be concurrently called more than once on the same channel,
244  * though different channels may be being processed concurrently.
245  */
246 static int efx_process_channel(struct efx_channel *channel, int budget)
247 {
248 	int spent;
249 
250 	if (unlikely(!channel->enabled))
251 		return 0;
252 
253 	spent = efx_nic_process_eventq(channel, budget);
254 	if (spent && efx_channel_has_rx_queue(channel)) {
255 		struct efx_rx_queue *rx_queue =
256 			efx_channel_get_rx_queue(channel);
257 
258 		efx_rx_flush_packet(channel);
259 		efx_fast_push_rx_descriptors(rx_queue, true);
260 	}
261 
262 	return spent;
263 }
264 
265 /* NAPI poll handler
266  *
267  * NAPI guarantees serialisation of polls of the same device, which
268  * provides the guarantee required by efx_process_channel().
269  */
270 static int efx_poll(struct napi_struct *napi, int budget)
271 {
272 	struct efx_channel *channel =
273 		container_of(napi, struct efx_channel, napi_str);
274 	struct efx_nic *efx = channel->efx;
275 	int spent;
276 
277 	if (!efx_channel_lock_napi(channel))
278 		return budget;
279 
280 	netif_vdbg(efx, intr, efx->net_dev,
281 		   "channel %d NAPI poll executing on CPU %d\n",
282 		   channel->channel, raw_smp_processor_id());
283 
284 	spent = efx_process_channel(channel, budget);
285 
286 	if (spent < budget) {
287 		if (efx_channel_has_rx_queue(channel) &&
288 		    efx->irq_rx_adaptive &&
289 		    unlikely(++channel->irq_count == 1000)) {
290 			if (unlikely(channel->irq_mod_score <
291 				     irq_adapt_low_thresh)) {
292 				if (channel->irq_moderation > 1) {
293 					channel->irq_moderation -= 1;
294 					efx->type->push_irq_moderation(channel);
295 				}
296 			} else if (unlikely(channel->irq_mod_score >
297 					    irq_adapt_high_thresh)) {
298 				if (channel->irq_moderation <
299 				    efx->irq_rx_moderation) {
300 					channel->irq_moderation += 1;
301 					efx->type->push_irq_moderation(channel);
302 				}
303 			}
304 			channel->irq_count = 0;
305 			channel->irq_mod_score = 0;
306 		}
307 
308 		efx_filter_rfs_expire(channel);
309 
310 		/* There is no race here; although napi_disable() will
311 		 * only wait for napi_complete(), this isn't a problem
312 		 * since efx_nic_eventq_read_ack() will have no effect if
313 		 * interrupts have already been disabled.
314 		 */
315 		napi_complete(napi);
316 		efx_nic_eventq_read_ack(channel);
317 	}
318 
319 	efx_channel_unlock_napi(channel);
320 	return spent;
321 }
322 
323 /* Create event queue
324  * Event queue memory allocations are done only once.  If the channel
325  * is reset, the memory buffer will be reused; this guards against
326  * errors during channel reset and also simplifies interrupt handling.
327  */
328 static int efx_probe_eventq(struct efx_channel *channel)
329 {
330 	struct efx_nic *efx = channel->efx;
331 	unsigned long entries;
332 
333 	netif_dbg(efx, probe, efx->net_dev,
334 		  "chan %d create event queue\n", channel->channel);
335 
336 	/* Build an event queue with room for one event per tx and rx buffer,
337 	 * plus some extra for link state events and MCDI completions. */
338 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
339 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
340 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
341 
342 	return efx_nic_probe_eventq(channel);
343 }
344 
345 /* Prepare channel's event queue */
346 static int efx_init_eventq(struct efx_channel *channel)
347 {
348 	struct efx_nic *efx = channel->efx;
349 	int rc;
350 
351 	EFX_WARN_ON_PARANOID(channel->eventq_init);
352 
353 	netif_dbg(efx, drv, efx->net_dev,
354 		  "chan %d init event queue\n", channel->channel);
355 
356 	rc = efx_nic_init_eventq(channel);
357 	if (rc == 0) {
358 		efx->type->push_irq_moderation(channel);
359 		channel->eventq_read_ptr = 0;
360 		channel->eventq_init = true;
361 	}
362 	return rc;
363 }
364 
365 /* Enable event queue processing and NAPI */
366 void efx_start_eventq(struct efx_channel *channel)
367 {
368 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
369 		  "chan %d start event queue\n", channel->channel);
370 
371 	/* Make sure the NAPI handler sees the enabled flag set */
372 	channel->enabled = true;
373 	smp_wmb();
374 
375 	efx_channel_enable(channel);
376 	napi_enable(&channel->napi_str);
377 	efx_nic_eventq_read_ack(channel);
378 }
379 
380 /* Disable event queue processing and NAPI */
381 void efx_stop_eventq(struct efx_channel *channel)
382 {
383 	if (!channel->enabled)
384 		return;
385 
386 	napi_disable(&channel->napi_str);
387 	while (!efx_channel_disable(channel))
388 		usleep_range(1000, 20000);
389 	channel->enabled = false;
390 }
391 
392 static void efx_fini_eventq(struct efx_channel *channel)
393 {
394 	if (!channel->eventq_init)
395 		return;
396 
397 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
398 		  "chan %d fini event queue\n", channel->channel);
399 
400 	efx_nic_fini_eventq(channel);
401 	channel->eventq_init = false;
402 }
403 
404 static void efx_remove_eventq(struct efx_channel *channel)
405 {
406 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
407 		  "chan %d remove event queue\n", channel->channel);
408 
409 	efx_nic_remove_eventq(channel);
410 }
411 
412 /**************************************************************************
413  *
414  * Channel handling
415  *
416  *************************************************************************/
417 
418 /* Allocate and initialise a channel structure. */
419 static struct efx_channel *
420 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
421 {
422 	struct efx_channel *channel;
423 	struct efx_rx_queue *rx_queue;
424 	struct efx_tx_queue *tx_queue;
425 	int j;
426 
427 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
428 	if (!channel)
429 		return NULL;
430 
431 	channel->efx = efx;
432 	channel->channel = i;
433 	channel->type = &efx_default_channel_type;
434 
435 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
436 		tx_queue = &channel->tx_queue[j];
437 		tx_queue->efx = efx;
438 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
439 		tx_queue->channel = channel;
440 	}
441 
442 	rx_queue = &channel->rx_queue;
443 	rx_queue->efx = efx;
444 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
445 		    (unsigned long)rx_queue);
446 
447 	return channel;
448 }
449 
450 /* Allocate and initialise a channel structure, copying parameters
451  * (but not resources) from an old channel structure.
452  */
453 static struct efx_channel *
454 efx_copy_channel(const struct efx_channel *old_channel)
455 {
456 	struct efx_channel *channel;
457 	struct efx_rx_queue *rx_queue;
458 	struct efx_tx_queue *tx_queue;
459 	int j;
460 
461 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
462 	if (!channel)
463 		return NULL;
464 
465 	*channel = *old_channel;
466 
467 	channel->napi_dev = NULL;
468 	memset(&channel->eventq, 0, sizeof(channel->eventq));
469 
470 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
471 		tx_queue = &channel->tx_queue[j];
472 		if (tx_queue->channel)
473 			tx_queue->channel = channel;
474 		tx_queue->buffer = NULL;
475 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
476 	}
477 
478 	rx_queue = &channel->rx_queue;
479 	rx_queue->buffer = NULL;
480 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
481 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
482 		    (unsigned long)rx_queue);
483 
484 	return channel;
485 }
486 
487 static int efx_probe_channel(struct efx_channel *channel)
488 {
489 	struct efx_tx_queue *tx_queue;
490 	struct efx_rx_queue *rx_queue;
491 	int rc;
492 
493 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
494 		  "creating channel %d\n", channel->channel);
495 
496 	rc = channel->type->pre_probe(channel);
497 	if (rc)
498 		goto fail;
499 
500 	rc = efx_probe_eventq(channel);
501 	if (rc)
502 		goto fail;
503 
504 	efx_for_each_channel_tx_queue(tx_queue, channel) {
505 		rc = efx_probe_tx_queue(tx_queue);
506 		if (rc)
507 			goto fail;
508 	}
509 
510 	efx_for_each_channel_rx_queue(rx_queue, channel) {
511 		rc = efx_probe_rx_queue(rx_queue);
512 		if (rc)
513 			goto fail;
514 	}
515 
516 	return 0;
517 
518 fail:
519 	efx_remove_channel(channel);
520 	return rc;
521 }
522 
523 static void
524 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
525 {
526 	struct efx_nic *efx = channel->efx;
527 	const char *type;
528 	int number;
529 
530 	number = channel->channel;
531 	if (efx->tx_channel_offset == 0) {
532 		type = "";
533 	} else if (channel->channel < efx->tx_channel_offset) {
534 		type = "-rx";
535 	} else {
536 		type = "-tx";
537 		number -= efx->tx_channel_offset;
538 	}
539 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
540 }
541 
542 static void efx_set_channel_names(struct efx_nic *efx)
543 {
544 	struct efx_channel *channel;
545 
546 	efx_for_each_channel(channel, efx)
547 		channel->type->get_name(channel,
548 					efx->msi_context[channel->channel].name,
549 					sizeof(efx->msi_context[0].name));
550 }
551 
552 static int efx_probe_channels(struct efx_nic *efx)
553 {
554 	struct efx_channel *channel;
555 	int rc;
556 
557 	/* Restart special buffer allocation */
558 	efx->next_buffer_table = 0;
559 
560 	/* Probe channels in reverse, so that any 'extra' channels
561 	 * use the start of the buffer table. This allows the traffic
562 	 * channels to be resized without moving them or wasting the
563 	 * entries before them.
564 	 */
565 	efx_for_each_channel_rev(channel, efx) {
566 		rc = efx_probe_channel(channel);
567 		if (rc) {
568 			netif_err(efx, probe, efx->net_dev,
569 				  "failed to create channel %d\n",
570 				  channel->channel);
571 			goto fail;
572 		}
573 	}
574 	efx_set_channel_names(efx);
575 
576 	return 0;
577 
578 fail:
579 	efx_remove_channels(efx);
580 	return rc;
581 }
582 
583 /* Channels are shutdown and reinitialised whilst the NIC is running
584  * to propagate configuration changes (mtu, checksum offload), or
585  * to clear hardware error conditions
586  */
587 static void efx_start_datapath(struct efx_nic *efx)
588 {
589 	bool old_rx_scatter = efx->rx_scatter;
590 	struct efx_tx_queue *tx_queue;
591 	struct efx_rx_queue *rx_queue;
592 	struct efx_channel *channel;
593 	size_t rx_buf_len;
594 
595 	/* Calculate the rx buffer allocation parameters required to
596 	 * support the current MTU, including padding for header
597 	 * alignment and overruns.
598 	 */
599 	efx->rx_dma_len = (efx->rx_prefix_size +
600 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
601 			   efx->type->rx_buffer_padding);
602 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
603 		      efx->rx_ip_align + efx->rx_dma_len);
604 	if (rx_buf_len <= PAGE_SIZE) {
605 		efx->rx_scatter = efx->type->always_rx_scatter;
606 		efx->rx_buffer_order = 0;
607 	} else if (efx->type->can_rx_scatter) {
608 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
609 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
610 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
611 				       EFX_RX_BUF_ALIGNMENT) >
612 			     PAGE_SIZE);
613 		efx->rx_scatter = true;
614 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
615 		efx->rx_buffer_order = 0;
616 	} else {
617 		efx->rx_scatter = false;
618 		efx->rx_buffer_order = get_order(rx_buf_len);
619 	}
620 
621 	efx_rx_config_page_split(efx);
622 	if (efx->rx_buffer_order)
623 		netif_dbg(efx, drv, efx->net_dev,
624 			  "RX buf len=%u; page order=%u batch=%u\n",
625 			  efx->rx_dma_len, efx->rx_buffer_order,
626 			  efx->rx_pages_per_batch);
627 	else
628 		netif_dbg(efx, drv, efx->net_dev,
629 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
630 			  efx->rx_dma_len, efx->rx_page_buf_step,
631 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
632 
633 	/* RX filters may also have scatter-enabled flags */
634 	if (efx->rx_scatter != old_rx_scatter)
635 		efx->type->filter_update_rx_scatter(efx);
636 
637 	/* We must keep at least one descriptor in a TX ring empty.
638 	 * We could avoid this when the queue size does not exactly
639 	 * match the hardware ring size, but it's not that important.
640 	 * Therefore we stop the queue when one more skb might fill
641 	 * the ring completely.  We wake it when half way back to
642 	 * empty.
643 	 */
644 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
645 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
646 
647 	/* Initialise the channels */
648 	efx_for_each_channel(channel, efx) {
649 		efx_for_each_channel_tx_queue(tx_queue, channel) {
650 			efx_init_tx_queue(tx_queue);
651 			atomic_inc(&efx->active_queues);
652 		}
653 
654 		efx_for_each_channel_rx_queue(rx_queue, channel) {
655 			efx_init_rx_queue(rx_queue);
656 			atomic_inc(&efx->active_queues);
657 			efx_stop_eventq(channel);
658 			efx_fast_push_rx_descriptors(rx_queue, false);
659 			efx_start_eventq(channel);
660 		}
661 
662 		WARN_ON(channel->rx_pkt_n_frags);
663 	}
664 
665 	efx_ptp_start_datapath(efx);
666 
667 	if (netif_device_present(efx->net_dev))
668 		netif_tx_wake_all_queues(efx->net_dev);
669 }
670 
671 static void efx_stop_datapath(struct efx_nic *efx)
672 {
673 	struct efx_channel *channel;
674 	struct efx_tx_queue *tx_queue;
675 	struct efx_rx_queue *rx_queue;
676 	int rc;
677 
678 	EFX_ASSERT_RESET_SERIALISED(efx);
679 	BUG_ON(efx->port_enabled);
680 
681 	efx_ptp_stop_datapath(efx);
682 
683 	/* Stop RX refill */
684 	efx_for_each_channel(channel, efx) {
685 		efx_for_each_channel_rx_queue(rx_queue, channel)
686 			rx_queue->refill_enabled = false;
687 	}
688 
689 	efx_for_each_channel(channel, efx) {
690 		/* RX packet processing is pipelined, so wait for the
691 		 * NAPI handler to complete.  At least event queue 0
692 		 * might be kept active by non-data events, so don't
693 		 * use napi_synchronize() but actually disable NAPI
694 		 * temporarily.
695 		 */
696 		if (efx_channel_has_rx_queue(channel)) {
697 			efx_stop_eventq(channel);
698 			efx_start_eventq(channel);
699 		}
700 	}
701 
702 	rc = efx->type->fini_dmaq(efx);
703 	if (rc && EFX_WORKAROUND_7803(efx)) {
704 		/* Schedule a reset to recover from the flush failure. The
705 		 * descriptor caches reference memory we're about to free,
706 		 * but falcon_reconfigure_mac_wrapper() won't reconnect
707 		 * the MACs because of the pending reset.
708 		 */
709 		netif_err(efx, drv, efx->net_dev,
710 			  "Resetting to recover from flush failure\n");
711 		efx_schedule_reset(efx, RESET_TYPE_ALL);
712 	} else if (rc) {
713 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
714 	} else {
715 		netif_dbg(efx, drv, efx->net_dev,
716 			  "successfully flushed all queues\n");
717 	}
718 
719 	efx_for_each_channel(channel, efx) {
720 		efx_for_each_channel_rx_queue(rx_queue, channel)
721 			efx_fini_rx_queue(rx_queue);
722 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
723 			efx_fini_tx_queue(tx_queue);
724 	}
725 }
726 
727 static void efx_remove_channel(struct efx_channel *channel)
728 {
729 	struct efx_tx_queue *tx_queue;
730 	struct efx_rx_queue *rx_queue;
731 
732 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
733 		  "destroy chan %d\n", channel->channel);
734 
735 	efx_for_each_channel_rx_queue(rx_queue, channel)
736 		efx_remove_rx_queue(rx_queue);
737 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
738 		efx_remove_tx_queue(tx_queue);
739 	efx_remove_eventq(channel);
740 	channel->type->post_remove(channel);
741 }
742 
743 static void efx_remove_channels(struct efx_nic *efx)
744 {
745 	struct efx_channel *channel;
746 
747 	efx_for_each_channel(channel, efx)
748 		efx_remove_channel(channel);
749 }
750 
751 int
752 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
753 {
754 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
755 	u32 old_rxq_entries, old_txq_entries;
756 	unsigned i, next_buffer_table = 0;
757 	int rc, rc2;
758 
759 	rc = efx_check_disabled(efx);
760 	if (rc)
761 		return rc;
762 
763 	/* Not all channels should be reallocated. We must avoid
764 	 * reallocating their buffer table entries.
765 	 */
766 	efx_for_each_channel(channel, efx) {
767 		struct efx_rx_queue *rx_queue;
768 		struct efx_tx_queue *tx_queue;
769 
770 		if (channel->type->copy)
771 			continue;
772 		next_buffer_table = max(next_buffer_table,
773 					channel->eventq.index +
774 					channel->eventq.entries);
775 		efx_for_each_channel_rx_queue(rx_queue, channel)
776 			next_buffer_table = max(next_buffer_table,
777 						rx_queue->rxd.index +
778 						rx_queue->rxd.entries);
779 		efx_for_each_channel_tx_queue(tx_queue, channel)
780 			next_buffer_table = max(next_buffer_table,
781 						tx_queue->txd.index +
782 						tx_queue->txd.entries);
783 	}
784 
785 	efx_device_detach_sync(efx);
786 	efx_stop_all(efx);
787 	efx_soft_disable_interrupts(efx);
788 
789 	/* Clone channels (where possible) */
790 	memset(other_channel, 0, sizeof(other_channel));
791 	for (i = 0; i < efx->n_channels; i++) {
792 		channel = efx->channel[i];
793 		if (channel->type->copy)
794 			channel = channel->type->copy(channel);
795 		if (!channel) {
796 			rc = -ENOMEM;
797 			goto out;
798 		}
799 		other_channel[i] = channel;
800 	}
801 
802 	/* Swap entry counts and channel pointers */
803 	old_rxq_entries = efx->rxq_entries;
804 	old_txq_entries = efx->txq_entries;
805 	efx->rxq_entries = rxq_entries;
806 	efx->txq_entries = txq_entries;
807 	for (i = 0; i < efx->n_channels; i++) {
808 		channel = efx->channel[i];
809 		efx->channel[i] = other_channel[i];
810 		other_channel[i] = channel;
811 	}
812 
813 	/* Restart buffer table allocation */
814 	efx->next_buffer_table = next_buffer_table;
815 
816 	for (i = 0; i < efx->n_channels; i++) {
817 		channel = efx->channel[i];
818 		if (!channel->type->copy)
819 			continue;
820 		rc = efx_probe_channel(channel);
821 		if (rc)
822 			goto rollback;
823 		efx_init_napi_channel(efx->channel[i]);
824 	}
825 
826 out:
827 	/* Destroy unused channel structures */
828 	for (i = 0; i < efx->n_channels; i++) {
829 		channel = other_channel[i];
830 		if (channel && channel->type->copy) {
831 			efx_fini_napi_channel(channel);
832 			efx_remove_channel(channel);
833 			kfree(channel);
834 		}
835 	}
836 
837 	rc2 = efx_soft_enable_interrupts(efx);
838 	if (rc2) {
839 		rc = rc ? rc : rc2;
840 		netif_err(efx, drv, efx->net_dev,
841 			  "unable to restart interrupts on channel reallocation\n");
842 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
843 	} else {
844 		efx_start_all(efx);
845 		netif_device_attach(efx->net_dev);
846 	}
847 	return rc;
848 
849 rollback:
850 	/* Swap back */
851 	efx->rxq_entries = old_rxq_entries;
852 	efx->txq_entries = old_txq_entries;
853 	for (i = 0; i < efx->n_channels; i++) {
854 		channel = efx->channel[i];
855 		efx->channel[i] = other_channel[i];
856 		other_channel[i] = channel;
857 	}
858 	goto out;
859 }
860 
861 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
862 {
863 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
864 }
865 
866 static const struct efx_channel_type efx_default_channel_type = {
867 	.pre_probe		= efx_channel_dummy_op_int,
868 	.post_remove		= efx_channel_dummy_op_void,
869 	.get_name		= efx_get_channel_name,
870 	.copy			= efx_copy_channel,
871 	.keep_eventq		= false,
872 };
873 
874 int efx_channel_dummy_op_int(struct efx_channel *channel)
875 {
876 	return 0;
877 }
878 
879 void efx_channel_dummy_op_void(struct efx_channel *channel)
880 {
881 }
882 
883 /**************************************************************************
884  *
885  * Port handling
886  *
887  **************************************************************************/
888 
889 /* This ensures that the kernel is kept informed (via
890  * netif_carrier_on/off) of the link status, and also maintains the
891  * link status's stop on the port's TX queue.
892  */
893 void efx_link_status_changed(struct efx_nic *efx)
894 {
895 	struct efx_link_state *link_state = &efx->link_state;
896 
897 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
898 	 * that no events are triggered between unregister_netdev() and the
899 	 * driver unloading. A more general condition is that NETDEV_CHANGE
900 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
901 	if (!netif_running(efx->net_dev))
902 		return;
903 
904 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
905 		efx->n_link_state_changes++;
906 
907 		if (link_state->up)
908 			netif_carrier_on(efx->net_dev);
909 		else
910 			netif_carrier_off(efx->net_dev);
911 	}
912 
913 	/* Status message for kernel log */
914 	if (link_state->up)
915 		netif_info(efx, link, efx->net_dev,
916 			   "link up at %uMbps %s-duplex (MTU %d)\n",
917 			   link_state->speed, link_state->fd ? "full" : "half",
918 			   efx->net_dev->mtu);
919 	else
920 		netif_info(efx, link, efx->net_dev, "link down\n");
921 }
922 
923 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
924 {
925 	efx->link_advertising = advertising;
926 	if (advertising) {
927 		if (advertising & ADVERTISED_Pause)
928 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
929 		else
930 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
931 		if (advertising & ADVERTISED_Asym_Pause)
932 			efx->wanted_fc ^= EFX_FC_TX;
933 	}
934 }
935 
936 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
937 {
938 	efx->wanted_fc = wanted_fc;
939 	if (efx->link_advertising) {
940 		if (wanted_fc & EFX_FC_RX)
941 			efx->link_advertising |= (ADVERTISED_Pause |
942 						  ADVERTISED_Asym_Pause);
943 		else
944 			efx->link_advertising &= ~(ADVERTISED_Pause |
945 						   ADVERTISED_Asym_Pause);
946 		if (wanted_fc & EFX_FC_TX)
947 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
948 	}
949 }
950 
951 static void efx_fini_port(struct efx_nic *efx);
952 
953 /* We assume that efx->type->reconfigure_mac will always try to sync RX
954  * filters and therefore needs to read-lock the filter table against freeing
955  */
956 void efx_mac_reconfigure(struct efx_nic *efx)
957 {
958 	down_read(&efx->filter_sem);
959 	efx->type->reconfigure_mac(efx);
960 	up_read(&efx->filter_sem);
961 }
962 
963 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
964  * the MAC appropriately. All other PHY configuration changes are pushed
965  * through phy_op->set_settings(), and pushed asynchronously to the MAC
966  * through efx_monitor().
967  *
968  * Callers must hold the mac_lock
969  */
970 int __efx_reconfigure_port(struct efx_nic *efx)
971 {
972 	enum efx_phy_mode phy_mode;
973 	int rc;
974 
975 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
976 
977 	/* Disable PHY transmit in mac level loopbacks */
978 	phy_mode = efx->phy_mode;
979 	if (LOOPBACK_INTERNAL(efx))
980 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
981 	else
982 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
983 
984 	rc = efx->type->reconfigure_port(efx);
985 
986 	if (rc)
987 		efx->phy_mode = phy_mode;
988 
989 	return rc;
990 }
991 
992 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
993  * disabled. */
994 int efx_reconfigure_port(struct efx_nic *efx)
995 {
996 	int rc;
997 
998 	EFX_ASSERT_RESET_SERIALISED(efx);
999 
1000 	mutex_lock(&efx->mac_lock);
1001 	rc = __efx_reconfigure_port(efx);
1002 	mutex_unlock(&efx->mac_lock);
1003 
1004 	return rc;
1005 }
1006 
1007 /* Asynchronous work item for changing MAC promiscuity and multicast
1008  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1009  * MAC directly. */
1010 static void efx_mac_work(struct work_struct *data)
1011 {
1012 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1013 
1014 	mutex_lock(&efx->mac_lock);
1015 	if (efx->port_enabled)
1016 		efx_mac_reconfigure(efx);
1017 	mutex_unlock(&efx->mac_lock);
1018 }
1019 
1020 static int efx_probe_port(struct efx_nic *efx)
1021 {
1022 	int rc;
1023 
1024 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1025 
1026 	if (phy_flash_cfg)
1027 		efx->phy_mode = PHY_MODE_SPECIAL;
1028 
1029 	/* Connect up MAC/PHY operations table */
1030 	rc = efx->type->probe_port(efx);
1031 	if (rc)
1032 		return rc;
1033 
1034 	/* Initialise MAC address to permanent address */
1035 	ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1036 
1037 	return 0;
1038 }
1039 
1040 static int efx_init_port(struct efx_nic *efx)
1041 {
1042 	int rc;
1043 
1044 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1045 
1046 	mutex_lock(&efx->mac_lock);
1047 
1048 	rc = efx->phy_op->init(efx);
1049 	if (rc)
1050 		goto fail1;
1051 
1052 	efx->port_initialized = true;
1053 
1054 	/* Reconfigure the MAC before creating dma queues (required for
1055 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1056 	efx_mac_reconfigure(efx);
1057 
1058 	/* Ensure the PHY advertises the correct flow control settings */
1059 	rc = efx->phy_op->reconfigure(efx);
1060 	if (rc && rc != -EPERM)
1061 		goto fail2;
1062 
1063 	mutex_unlock(&efx->mac_lock);
1064 	return 0;
1065 
1066 fail2:
1067 	efx->phy_op->fini(efx);
1068 fail1:
1069 	mutex_unlock(&efx->mac_lock);
1070 	return rc;
1071 }
1072 
1073 static void efx_start_port(struct efx_nic *efx)
1074 {
1075 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1076 	BUG_ON(efx->port_enabled);
1077 
1078 	mutex_lock(&efx->mac_lock);
1079 	efx->port_enabled = true;
1080 
1081 	/* Ensure MAC ingress/egress is enabled */
1082 	efx_mac_reconfigure(efx);
1083 
1084 	mutex_unlock(&efx->mac_lock);
1085 }
1086 
1087 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1088  * and the async self-test, wait for them to finish and prevent them
1089  * being scheduled again.  This doesn't cover online resets, which
1090  * should only be cancelled when removing the device.
1091  */
1092 static void efx_stop_port(struct efx_nic *efx)
1093 {
1094 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1095 
1096 	EFX_ASSERT_RESET_SERIALISED(efx);
1097 
1098 	mutex_lock(&efx->mac_lock);
1099 	efx->port_enabled = false;
1100 	mutex_unlock(&efx->mac_lock);
1101 
1102 	/* Serialise against efx_set_multicast_list() */
1103 	netif_addr_lock_bh(efx->net_dev);
1104 	netif_addr_unlock_bh(efx->net_dev);
1105 
1106 	cancel_delayed_work_sync(&efx->monitor_work);
1107 	efx_selftest_async_cancel(efx);
1108 	cancel_work_sync(&efx->mac_work);
1109 }
1110 
1111 static void efx_fini_port(struct efx_nic *efx)
1112 {
1113 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1114 
1115 	if (!efx->port_initialized)
1116 		return;
1117 
1118 	efx->phy_op->fini(efx);
1119 	efx->port_initialized = false;
1120 
1121 	efx->link_state.up = false;
1122 	efx_link_status_changed(efx);
1123 }
1124 
1125 static void efx_remove_port(struct efx_nic *efx)
1126 {
1127 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1128 
1129 	efx->type->remove_port(efx);
1130 }
1131 
1132 /**************************************************************************
1133  *
1134  * NIC handling
1135  *
1136  **************************************************************************/
1137 
1138 static LIST_HEAD(efx_primary_list);
1139 static LIST_HEAD(efx_unassociated_list);
1140 
1141 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1142 {
1143 	return left->type == right->type &&
1144 		left->vpd_sn && right->vpd_sn &&
1145 		!strcmp(left->vpd_sn, right->vpd_sn);
1146 }
1147 
1148 static void efx_associate(struct efx_nic *efx)
1149 {
1150 	struct efx_nic *other, *next;
1151 
1152 	if (efx->primary == efx) {
1153 		/* Adding primary function; look for secondaries */
1154 
1155 		netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1156 		list_add_tail(&efx->node, &efx_primary_list);
1157 
1158 		list_for_each_entry_safe(other, next, &efx_unassociated_list,
1159 					 node) {
1160 			if (efx_same_controller(efx, other)) {
1161 				list_del(&other->node);
1162 				netif_dbg(other, probe, other->net_dev,
1163 					  "moving to secondary list of %s %s\n",
1164 					  pci_name(efx->pci_dev),
1165 					  efx->net_dev->name);
1166 				list_add_tail(&other->node,
1167 					      &efx->secondary_list);
1168 				other->primary = efx;
1169 			}
1170 		}
1171 	} else {
1172 		/* Adding secondary function; look for primary */
1173 
1174 		list_for_each_entry(other, &efx_primary_list, node) {
1175 			if (efx_same_controller(efx, other)) {
1176 				netif_dbg(efx, probe, efx->net_dev,
1177 					  "adding to secondary list of %s %s\n",
1178 					  pci_name(other->pci_dev),
1179 					  other->net_dev->name);
1180 				list_add_tail(&efx->node,
1181 					      &other->secondary_list);
1182 				efx->primary = other;
1183 				return;
1184 			}
1185 		}
1186 
1187 		netif_dbg(efx, probe, efx->net_dev,
1188 			  "adding to unassociated list\n");
1189 		list_add_tail(&efx->node, &efx_unassociated_list);
1190 	}
1191 }
1192 
1193 static void efx_dissociate(struct efx_nic *efx)
1194 {
1195 	struct efx_nic *other, *next;
1196 
1197 	list_del(&efx->node);
1198 	efx->primary = NULL;
1199 
1200 	list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1201 		list_del(&other->node);
1202 		netif_dbg(other, probe, other->net_dev,
1203 			  "moving to unassociated list\n");
1204 		list_add_tail(&other->node, &efx_unassociated_list);
1205 		other->primary = NULL;
1206 	}
1207 }
1208 
1209 /* This configures the PCI device to enable I/O and DMA. */
1210 static int efx_init_io(struct efx_nic *efx)
1211 {
1212 	struct pci_dev *pci_dev = efx->pci_dev;
1213 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1214 	unsigned int mem_map_size = efx->type->mem_map_size(efx);
1215 	int rc, bar;
1216 
1217 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1218 
1219 	bar = efx->type->mem_bar;
1220 
1221 	rc = pci_enable_device(pci_dev);
1222 	if (rc) {
1223 		netif_err(efx, probe, efx->net_dev,
1224 			  "failed to enable PCI device\n");
1225 		goto fail1;
1226 	}
1227 
1228 	pci_set_master(pci_dev);
1229 
1230 	/* Set the PCI DMA mask.  Try all possibilities from our
1231 	 * genuine mask down to 32 bits, because some architectures
1232 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1233 	 * masks event though they reject 46 bit masks.
1234 	 */
1235 	while (dma_mask > 0x7fffffffUL) {
1236 		if (dma_supported(&pci_dev->dev, dma_mask)) {
1237 			rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1238 			if (rc == 0)
1239 				break;
1240 		}
1241 		dma_mask >>= 1;
1242 	}
1243 	if (rc) {
1244 		netif_err(efx, probe, efx->net_dev,
1245 			  "could not find a suitable DMA mask\n");
1246 		goto fail2;
1247 	}
1248 	netif_dbg(efx, probe, efx->net_dev,
1249 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1250 
1251 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1252 	rc = pci_request_region(pci_dev, bar, "sfc");
1253 	if (rc) {
1254 		netif_err(efx, probe, efx->net_dev,
1255 			  "request for memory BAR failed\n");
1256 		rc = -EIO;
1257 		goto fail3;
1258 	}
1259 	efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1260 	if (!efx->membase) {
1261 		netif_err(efx, probe, efx->net_dev,
1262 			  "could not map memory BAR at %llx+%x\n",
1263 			  (unsigned long long)efx->membase_phys, mem_map_size);
1264 		rc = -ENOMEM;
1265 		goto fail4;
1266 	}
1267 	netif_dbg(efx, probe, efx->net_dev,
1268 		  "memory BAR at %llx+%x (virtual %p)\n",
1269 		  (unsigned long long)efx->membase_phys, mem_map_size,
1270 		  efx->membase);
1271 
1272 	return 0;
1273 
1274  fail4:
1275 	pci_release_region(efx->pci_dev, bar);
1276  fail3:
1277 	efx->membase_phys = 0;
1278  fail2:
1279 	pci_disable_device(efx->pci_dev);
1280  fail1:
1281 	return rc;
1282 }
1283 
1284 static void efx_fini_io(struct efx_nic *efx)
1285 {
1286 	int bar;
1287 
1288 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1289 
1290 	if (efx->membase) {
1291 		iounmap(efx->membase);
1292 		efx->membase = NULL;
1293 	}
1294 
1295 	if (efx->membase_phys) {
1296 		bar = efx->type->mem_bar;
1297 		pci_release_region(efx->pci_dev, bar);
1298 		efx->membase_phys = 0;
1299 	}
1300 
1301 	/* Don't disable bus-mastering if VFs are assigned */
1302 	if (!pci_vfs_assigned(efx->pci_dev))
1303 		pci_disable_device(efx->pci_dev);
1304 }
1305 
1306 void efx_set_default_rx_indir_table(struct efx_nic *efx)
1307 {
1308 	size_t i;
1309 
1310 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1311 		efx->rx_indir_table[i] =
1312 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1313 }
1314 
1315 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1316 {
1317 	cpumask_var_t thread_mask;
1318 	unsigned int count;
1319 	int cpu;
1320 
1321 	if (rss_cpus) {
1322 		count = rss_cpus;
1323 	} else {
1324 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1325 			netif_warn(efx, probe, efx->net_dev,
1326 				   "RSS disabled due to allocation failure\n");
1327 			return 1;
1328 		}
1329 
1330 		count = 0;
1331 		for_each_online_cpu(cpu) {
1332 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1333 				++count;
1334 				cpumask_or(thread_mask, thread_mask,
1335 					   topology_sibling_cpumask(cpu));
1336 			}
1337 		}
1338 
1339 		free_cpumask_var(thread_mask);
1340 	}
1341 
1342 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1343 	 * table entries that are inaccessible to VFs
1344 	 */
1345 #ifdef CONFIG_SFC_SRIOV
1346 	if (efx->type->sriov_wanted) {
1347 		if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1348 		    count > efx_vf_size(efx)) {
1349 			netif_warn(efx, probe, efx->net_dev,
1350 				   "Reducing number of RSS channels from %u to %u for "
1351 				   "VF support. Increase vf-msix-limit to use more "
1352 				   "channels on the PF.\n",
1353 				   count, efx_vf_size(efx));
1354 			count = efx_vf_size(efx);
1355 		}
1356 	}
1357 #endif
1358 
1359 	return count;
1360 }
1361 
1362 /* Probe the number and type of interrupts we are able to obtain, and
1363  * the resulting numbers of channels and RX queues.
1364  */
1365 static int efx_probe_interrupts(struct efx_nic *efx)
1366 {
1367 	unsigned int extra_channels = 0;
1368 	unsigned int i, j;
1369 	int rc;
1370 
1371 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1372 		if (efx->extra_channel_type[i])
1373 			++extra_channels;
1374 
1375 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1376 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1377 		unsigned int n_channels;
1378 
1379 		n_channels = efx_wanted_parallelism(efx);
1380 		if (separate_tx_channels)
1381 			n_channels *= 2;
1382 		n_channels += extra_channels;
1383 		n_channels = min(n_channels, efx->max_channels);
1384 
1385 		for (i = 0; i < n_channels; i++)
1386 			xentries[i].entry = i;
1387 		rc = pci_enable_msix_range(efx->pci_dev,
1388 					   xentries, 1, n_channels);
1389 		if (rc < 0) {
1390 			/* Fall back to single channel MSI */
1391 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1392 			netif_err(efx, drv, efx->net_dev,
1393 				  "could not enable MSI-X\n");
1394 		} else if (rc < n_channels) {
1395 			netif_err(efx, drv, efx->net_dev,
1396 				  "WARNING: Insufficient MSI-X vectors"
1397 				  " available (%d < %u).\n", rc, n_channels);
1398 			netif_err(efx, drv, efx->net_dev,
1399 				  "WARNING: Performance may be reduced.\n");
1400 			n_channels = rc;
1401 		}
1402 
1403 		if (rc > 0) {
1404 			efx->n_channels = n_channels;
1405 			if (n_channels > extra_channels)
1406 				n_channels -= extra_channels;
1407 			if (separate_tx_channels) {
1408 				efx->n_tx_channels = max(n_channels / 2, 1U);
1409 				efx->n_rx_channels = max(n_channels -
1410 							 efx->n_tx_channels,
1411 							 1U);
1412 			} else {
1413 				efx->n_tx_channels = n_channels;
1414 				efx->n_rx_channels = n_channels;
1415 			}
1416 			for (i = 0; i < efx->n_channels; i++)
1417 				efx_get_channel(efx, i)->irq =
1418 					xentries[i].vector;
1419 		}
1420 	}
1421 
1422 	/* Try single interrupt MSI */
1423 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1424 		efx->n_channels = 1;
1425 		efx->n_rx_channels = 1;
1426 		efx->n_tx_channels = 1;
1427 		rc = pci_enable_msi(efx->pci_dev);
1428 		if (rc == 0) {
1429 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1430 		} else {
1431 			netif_err(efx, drv, efx->net_dev,
1432 				  "could not enable MSI\n");
1433 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1434 		}
1435 	}
1436 
1437 	/* Assume legacy interrupts */
1438 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1439 		efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1440 		efx->n_rx_channels = 1;
1441 		efx->n_tx_channels = 1;
1442 		efx->legacy_irq = efx->pci_dev->irq;
1443 	}
1444 
1445 	/* Assign extra channels if possible */
1446 	j = efx->n_channels;
1447 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1448 		if (!efx->extra_channel_type[i])
1449 			continue;
1450 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1451 		    efx->n_channels <= extra_channels) {
1452 			efx->extra_channel_type[i]->handle_no_channel(efx);
1453 		} else {
1454 			--j;
1455 			efx_get_channel(efx, j)->type =
1456 				efx->extra_channel_type[i];
1457 		}
1458 	}
1459 
1460 	/* RSS might be usable on VFs even if it is disabled on the PF */
1461 #ifdef CONFIG_SFC_SRIOV
1462 	if (efx->type->sriov_wanted) {
1463 		efx->rss_spread = ((efx->n_rx_channels > 1 ||
1464 				    !efx->type->sriov_wanted(efx)) ?
1465 				   efx->n_rx_channels : efx_vf_size(efx));
1466 		return 0;
1467 	}
1468 #endif
1469 	efx->rss_spread = efx->n_rx_channels;
1470 
1471 	return 0;
1472 }
1473 
1474 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1475 {
1476 	struct efx_channel *channel, *end_channel;
1477 	int rc;
1478 
1479 	BUG_ON(efx->state == STATE_DISABLED);
1480 
1481 	efx->irq_soft_enabled = true;
1482 	smp_wmb();
1483 
1484 	efx_for_each_channel(channel, efx) {
1485 		if (!channel->type->keep_eventq) {
1486 			rc = efx_init_eventq(channel);
1487 			if (rc)
1488 				goto fail;
1489 		}
1490 		efx_start_eventq(channel);
1491 	}
1492 
1493 	efx_mcdi_mode_event(efx);
1494 
1495 	return 0;
1496 fail:
1497 	end_channel = channel;
1498 	efx_for_each_channel(channel, efx) {
1499 		if (channel == end_channel)
1500 			break;
1501 		efx_stop_eventq(channel);
1502 		if (!channel->type->keep_eventq)
1503 			efx_fini_eventq(channel);
1504 	}
1505 
1506 	return rc;
1507 }
1508 
1509 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1510 {
1511 	struct efx_channel *channel;
1512 
1513 	if (efx->state == STATE_DISABLED)
1514 		return;
1515 
1516 	efx_mcdi_mode_poll(efx);
1517 
1518 	efx->irq_soft_enabled = false;
1519 	smp_wmb();
1520 
1521 	if (efx->legacy_irq)
1522 		synchronize_irq(efx->legacy_irq);
1523 
1524 	efx_for_each_channel(channel, efx) {
1525 		if (channel->irq)
1526 			synchronize_irq(channel->irq);
1527 
1528 		efx_stop_eventq(channel);
1529 		if (!channel->type->keep_eventq)
1530 			efx_fini_eventq(channel);
1531 	}
1532 
1533 	/* Flush the asynchronous MCDI request queue */
1534 	efx_mcdi_flush_async(efx);
1535 }
1536 
1537 static int efx_enable_interrupts(struct efx_nic *efx)
1538 {
1539 	struct efx_channel *channel, *end_channel;
1540 	int rc;
1541 
1542 	BUG_ON(efx->state == STATE_DISABLED);
1543 
1544 	if (efx->eeh_disabled_legacy_irq) {
1545 		enable_irq(efx->legacy_irq);
1546 		efx->eeh_disabled_legacy_irq = false;
1547 	}
1548 
1549 	efx->type->irq_enable_master(efx);
1550 
1551 	efx_for_each_channel(channel, efx) {
1552 		if (channel->type->keep_eventq) {
1553 			rc = efx_init_eventq(channel);
1554 			if (rc)
1555 				goto fail;
1556 		}
1557 	}
1558 
1559 	rc = efx_soft_enable_interrupts(efx);
1560 	if (rc)
1561 		goto fail;
1562 
1563 	return 0;
1564 
1565 fail:
1566 	end_channel = channel;
1567 	efx_for_each_channel(channel, efx) {
1568 		if (channel == end_channel)
1569 			break;
1570 		if (channel->type->keep_eventq)
1571 			efx_fini_eventq(channel);
1572 	}
1573 
1574 	efx->type->irq_disable_non_ev(efx);
1575 
1576 	return rc;
1577 }
1578 
1579 static void efx_disable_interrupts(struct efx_nic *efx)
1580 {
1581 	struct efx_channel *channel;
1582 
1583 	efx_soft_disable_interrupts(efx);
1584 
1585 	efx_for_each_channel(channel, efx) {
1586 		if (channel->type->keep_eventq)
1587 			efx_fini_eventq(channel);
1588 	}
1589 
1590 	efx->type->irq_disable_non_ev(efx);
1591 }
1592 
1593 static void efx_remove_interrupts(struct efx_nic *efx)
1594 {
1595 	struct efx_channel *channel;
1596 
1597 	/* Remove MSI/MSI-X interrupts */
1598 	efx_for_each_channel(channel, efx)
1599 		channel->irq = 0;
1600 	pci_disable_msi(efx->pci_dev);
1601 	pci_disable_msix(efx->pci_dev);
1602 
1603 	/* Remove legacy interrupt */
1604 	efx->legacy_irq = 0;
1605 }
1606 
1607 static void efx_set_channels(struct efx_nic *efx)
1608 {
1609 	struct efx_channel *channel;
1610 	struct efx_tx_queue *tx_queue;
1611 
1612 	efx->tx_channel_offset =
1613 		separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1614 
1615 	/* We need to mark which channels really have RX and TX
1616 	 * queues, and adjust the TX queue numbers if we have separate
1617 	 * RX-only and TX-only channels.
1618 	 */
1619 	efx_for_each_channel(channel, efx) {
1620 		if (channel->channel < efx->n_rx_channels)
1621 			channel->rx_queue.core_index = channel->channel;
1622 		else
1623 			channel->rx_queue.core_index = -1;
1624 
1625 		efx_for_each_channel_tx_queue(tx_queue, channel)
1626 			tx_queue->queue -= (efx->tx_channel_offset *
1627 					    EFX_TXQ_TYPES);
1628 	}
1629 }
1630 
1631 static int efx_probe_nic(struct efx_nic *efx)
1632 {
1633 	int rc;
1634 
1635 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1636 
1637 	/* Carry out hardware-type specific initialisation */
1638 	rc = efx->type->probe(efx);
1639 	if (rc)
1640 		return rc;
1641 
1642 	/* Determine the number of channels and queues by trying to hook
1643 	 * in MSI-X interrupts. */
1644 	rc = efx_probe_interrupts(efx);
1645 	if (rc)
1646 		goto fail1;
1647 
1648 	efx_set_channels(efx);
1649 
1650 	rc = efx->type->dimension_resources(efx);
1651 	if (rc)
1652 		goto fail2;
1653 
1654 	if (efx->n_channels > 1)
1655 		netdev_rss_key_fill(&efx->rx_hash_key,
1656 				    sizeof(efx->rx_hash_key));
1657 	efx_set_default_rx_indir_table(efx);
1658 
1659 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1660 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1661 
1662 	/* Initialise the interrupt moderation settings */
1663 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1664 				true);
1665 
1666 	return 0;
1667 
1668 fail2:
1669 	efx_remove_interrupts(efx);
1670 fail1:
1671 	efx->type->remove(efx);
1672 	return rc;
1673 }
1674 
1675 static void efx_remove_nic(struct efx_nic *efx)
1676 {
1677 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1678 
1679 	efx_remove_interrupts(efx);
1680 	efx->type->remove(efx);
1681 }
1682 
1683 static int efx_probe_filters(struct efx_nic *efx)
1684 {
1685 	int rc;
1686 
1687 	spin_lock_init(&efx->filter_lock);
1688 	init_rwsem(&efx->filter_sem);
1689 	down_write(&efx->filter_sem);
1690 	rc = efx->type->filter_table_probe(efx);
1691 	if (rc)
1692 		goto out_unlock;
1693 
1694 #ifdef CONFIG_RFS_ACCEL
1695 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
1696 		efx->rps_flow_id = kcalloc(efx->type->max_rx_ip_filters,
1697 					   sizeof(*efx->rps_flow_id),
1698 					   GFP_KERNEL);
1699 		if (!efx->rps_flow_id) {
1700 			efx->type->filter_table_remove(efx);
1701 			rc = -ENOMEM;
1702 			goto out_unlock;
1703 		}
1704 	}
1705 #endif
1706 out_unlock:
1707 	up_write(&efx->filter_sem);
1708 	return rc;
1709 }
1710 
1711 static void efx_remove_filters(struct efx_nic *efx)
1712 {
1713 #ifdef CONFIG_RFS_ACCEL
1714 	kfree(efx->rps_flow_id);
1715 #endif
1716 	down_write(&efx->filter_sem);
1717 	efx->type->filter_table_remove(efx);
1718 	up_write(&efx->filter_sem);
1719 }
1720 
1721 static void efx_restore_filters(struct efx_nic *efx)
1722 {
1723 	down_read(&efx->filter_sem);
1724 	efx->type->filter_table_restore(efx);
1725 	up_read(&efx->filter_sem);
1726 }
1727 
1728 /**************************************************************************
1729  *
1730  * NIC startup/shutdown
1731  *
1732  *************************************************************************/
1733 
1734 static int efx_probe_all(struct efx_nic *efx)
1735 {
1736 	int rc;
1737 
1738 	rc = efx_probe_nic(efx);
1739 	if (rc) {
1740 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1741 		goto fail1;
1742 	}
1743 
1744 	rc = efx_probe_port(efx);
1745 	if (rc) {
1746 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1747 		goto fail2;
1748 	}
1749 
1750 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1751 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1752 		rc = -EINVAL;
1753 		goto fail3;
1754 	}
1755 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1756 
1757 #ifdef CONFIG_SFC_SRIOV
1758 	rc = efx->type->vswitching_probe(efx);
1759 	if (rc) /* not fatal; the PF will still work fine */
1760 		netif_warn(efx, probe, efx->net_dev,
1761 			   "failed to setup vswitching rc=%d;"
1762 			   " VFs may not function\n", rc);
1763 #endif
1764 
1765 	rc = efx_probe_filters(efx);
1766 	if (rc) {
1767 		netif_err(efx, probe, efx->net_dev,
1768 			  "failed to create filter tables\n");
1769 		goto fail4;
1770 	}
1771 
1772 	rc = efx_probe_channels(efx);
1773 	if (rc)
1774 		goto fail5;
1775 
1776 	return 0;
1777 
1778  fail5:
1779 	efx_remove_filters(efx);
1780  fail4:
1781 #ifdef CONFIG_SFC_SRIOV
1782 	efx->type->vswitching_remove(efx);
1783 #endif
1784  fail3:
1785 	efx_remove_port(efx);
1786  fail2:
1787 	efx_remove_nic(efx);
1788  fail1:
1789 	return rc;
1790 }
1791 
1792 /* If the interface is supposed to be running but is not, start
1793  * the hardware and software data path, regular activity for the port
1794  * (MAC statistics, link polling, etc.) and schedule the port to be
1795  * reconfigured.  Interrupts must already be enabled.  This function
1796  * is safe to call multiple times, so long as the NIC is not disabled.
1797  * Requires the RTNL lock.
1798  */
1799 static void efx_start_all(struct efx_nic *efx)
1800 {
1801 	EFX_ASSERT_RESET_SERIALISED(efx);
1802 	BUG_ON(efx->state == STATE_DISABLED);
1803 
1804 	/* Check that it is appropriate to restart the interface. All
1805 	 * of these flags are safe to read under just the rtnl lock */
1806 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
1807 	    efx->reset_pending)
1808 		return;
1809 
1810 	efx_start_port(efx);
1811 	efx_start_datapath(efx);
1812 
1813 	/* Start the hardware monitor if there is one */
1814 	if (efx->type->monitor != NULL)
1815 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1816 				   efx_monitor_interval);
1817 
1818 	/* If link state detection is normally event-driven, we have
1819 	 * to poll now because we could have missed a change
1820 	 */
1821 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1822 		mutex_lock(&efx->mac_lock);
1823 		if (efx->phy_op->poll(efx))
1824 			efx_link_status_changed(efx);
1825 		mutex_unlock(&efx->mac_lock);
1826 	}
1827 
1828 	efx->type->start_stats(efx);
1829 	efx->type->pull_stats(efx);
1830 	spin_lock_bh(&efx->stats_lock);
1831 	efx->type->update_stats(efx, NULL, NULL);
1832 	spin_unlock_bh(&efx->stats_lock);
1833 }
1834 
1835 /* Quiesce the hardware and software data path, and regular activity
1836  * for the port without bringing the link down.  Safe to call multiple
1837  * times with the NIC in almost any state, but interrupts should be
1838  * enabled.  Requires the RTNL lock.
1839  */
1840 static void efx_stop_all(struct efx_nic *efx)
1841 {
1842 	EFX_ASSERT_RESET_SERIALISED(efx);
1843 
1844 	/* port_enabled can be read safely under the rtnl lock */
1845 	if (!efx->port_enabled)
1846 		return;
1847 
1848 	/* update stats before we go down so we can accurately count
1849 	 * rx_nodesc_drops
1850 	 */
1851 	efx->type->pull_stats(efx);
1852 	spin_lock_bh(&efx->stats_lock);
1853 	efx->type->update_stats(efx, NULL, NULL);
1854 	spin_unlock_bh(&efx->stats_lock);
1855 	efx->type->stop_stats(efx);
1856 	efx_stop_port(efx);
1857 
1858 	/* Stop the kernel transmit interface.  This is only valid if
1859 	 * the device is stopped or detached; otherwise the watchdog
1860 	 * may fire immediately.
1861 	 */
1862 	WARN_ON(netif_running(efx->net_dev) &&
1863 		netif_device_present(efx->net_dev));
1864 	netif_tx_disable(efx->net_dev);
1865 
1866 	efx_stop_datapath(efx);
1867 }
1868 
1869 static void efx_remove_all(struct efx_nic *efx)
1870 {
1871 	efx_remove_channels(efx);
1872 	efx_remove_filters(efx);
1873 #ifdef CONFIG_SFC_SRIOV
1874 	efx->type->vswitching_remove(efx);
1875 #endif
1876 	efx_remove_port(efx);
1877 	efx_remove_nic(efx);
1878 }
1879 
1880 /**************************************************************************
1881  *
1882  * Interrupt moderation
1883  *
1884  **************************************************************************/
1885 
1886 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1887 {
1888 	if (usecs == 0)
1889 		return 0;
1890 	if (usecs * 1000 < quantum_ns)
1891 		return 1; /* never round down to 0 */
1892 	return usecs * 1000 / quantum_ns;
1893 }
1894 
1895 /* Set interrupt moderation parameters */
1896 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1897 			    unsigned int rx_usecs, bool rx_adaptive,
1898 			    bool rx_may_override_tx)
1899 {
1900 	struct efx_channel *channel;
1901 	unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1902 						efx->timer_quantum_ns,
1903 						1000);
1904 	unsigned int tx_ticks;
1905 	unsigned int rx_ticks;
1906 
1907 	EFX_ASSERT_RESET_SERIALISED(efx);
1908 
1909 	if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1910 		return -EINVAL;
1911 
1912 	tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1913 	rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1914 
1915 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1916 	    !rx_may_override_tx) {
1917 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1918 			  "RX and TX IRQ moderation must be equal\n");
1919 		return -EINVAL;
1920 	}
1921 
1922 	efx->irq_rx_adaptive = rx_adaptive;
1923 	efx->irq_rx_moderation = rx_ticks;
1924 	efx_for_each_channel(channel, efx) {
1925 		if (efx_channel_has_rx_queue(channel))
1926 			channel->irq_moderation = rx_ticks;
1927 		else if (efx_channel_has_tx_queues(channel))
1928 			channel->irq_moderation = tx_ticks;
1929 	}
1930 
1931 	return 0;
1932 }
1933 
1934 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1935 			    unsigned int *rx_usecs, bool *rx_adaptive)
1936 {
1937 	/* We must round up when converting ticks to microseconds
1938 	 * because we round down when converting the other way.
1939 	 */
1940 
1941 	*rx_adaptive = efx->irq_rx_adaptive;
1942 	*rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1943 				 efx->timer_quantum_ns,
1944 				 1000);
1945 
1946 	/* If channels are shared between RX and TX, so is IRQ
1947 	 * moderation.  Otherwise, IRQ moderation is the same for all
1948 	 * TX channels and is not adaptive.
1949 	 */
1950 	if (efx->tx_channel_offset == 0)
1951 		*tx_usecs = *rx_usecs;
1952 	else
1953 		*tx_usecs = DIV_ROUND_UP(
1954 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1955 			efx->timer_quantum_ns,
1956 			1000);
1957 }
1958 
1959 /**************************************************************************
1960  *
1961  * Hardware monitor
1962  *
1963  **************************************************************************/
1964 
1965 /* Run periodically off the general workqueue */
1966 static void efx_monitor(struct work_struct *data)
1967 {
1968 	struct efx_nic *efx = container_of(data, struct efx_nic,
1969 					   monitor_work.work);
1970 
1971 	netif_vdbg(efx, timer, efx->net_dev,
1972 		   "hardware monitor executing on CPU %d\n",
1973 		   raw_smp_processor_id());
1974 	BUG_ON(efx->type->monitor == NULL);
1975 
1976 	/* If the mac_lock is already held then it is likely a port
1977 	 * reconfiguration is already in place, which will likely do
1978 	 * most of the work of monitor() anyway. */
1979 	if (mutex_trylock(&efx->mac_lock)) {
1980 		if (efx->port_enabled)
1981 			efx->type->monitor(efx);
1982 		mutex_unlock(&efx->mac_lock);
1983 	}
1984 
1985 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1986 			   efx_monitor_interval);
1987 }
1988 
1989 /**************************************************************************
1990  *
1991  * ioctls
1992  *
1993  *************************************************************************/
1994 
1995 /* Net device ioctl
1996  * Context: process, rtnl_lock() held.
1997  */
1998 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1999 {
2000 	struct efx_nic *efx = netdev_priv(net_dev);
2001 	struct mii_ioctl_data *data = if_mii(ifr);
2002 
2003 	if (cmd == SIOCSHWTSTAMP)
2004 		return efx_ptp_set_ts_config(efx, ifr);
2005 	if (cmd == SIOCGHWTSTAMP)
2006 		return efx_ptp_get_ts_config(efx, ifr);
2007 
2008 	/* Convert phy_id from older PRTAD/DEVAD format */
2009 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2010 	    (data->phy_id & 0xfc00) == 0x0400)
2011 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2012 
2013 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
2014 }
2015 
2016 /**************************************************************************
2017  *
2018  * NAPI interface
2019  *
2020  **************************************************************************/
2021 
2022 static void efx_init_napi_channel(struct efx_channel *channel)
2023 {
2024 	struct efx_nic *efx = channel->efx;
2025 
2026 	channel->napi_dev = efx->net_dev;
2027 	netif_napi_add(channel->napi_dev, &channel->napi_str,
2028 		       efx_poll, napi_weight);
2029 	napi_hash_add(&channel->napi_str);
2030 	efx_channel_init_lock(channel);
2031 }
2032 
2033 static void efx_init_napi(struct efx_nic *efx)
2034 {
2035 	struct efx_channel *channel;
2036 
2037 	efx_for_each_channel(channel, efx)
2038 		efx_init_napi_channel(channel);
2039 }
2040 
2041 static void efx_fini_napi_channel(struct efx_channel *channel)
2042 {
2043 	if (channel->napi_dev) {
2044 		netif_napi_del(&channel->napi_str);
2045 		napi_hash_del(&channel->napi_str);
2046 	}
2047 	channel->napi_dev = NULL;
2048 }
2049 
2050 static void efx_fini_napi(struct efx_nic *efx)
2051 {
2052 	struct efx_channel *channel;
2053 
2054 	efx_for_each_channel(channel, efx)
2055 		efx_fini_napi_channel(channel);
2056 }
2057 
2058 /**************************************************************************
2059  *
2060  * Kernel netpoll interface
2061  *
2062  *************************************************************************/
2063 
2064 #ifdef CONFIG_NET_POLL_CONTROLLER
2065 
2066 /* Although in the common case interrupts will be disabled, this is not
2067  * guaranteed. However, all our work happens inside the NAPI callback,
2068  * so no locking is required.
2069  */
2070 static void efx_netpoll(struct net_device *net_dev)
2071 {
2072 	struct efx_nic *efx = netdev_priv(net_dev);
2073 	struct efx_channel *channel;
2074 
2075 	efx_for_each_channel(channel, efx)
2076 		efx_schedule_channel(channel);
2077 }
2078 
2079 #endif
2080 
2081 #ifdef CONFIG_NET_RX_BUSY_POLL
2082 static int efx_busy_poll(struct napi_struct *napi)
2083 {
2084 	struct efx_channel *channel =
2085 		container_of(napi, struct efx_channel, napi_str);
2086 	struct efx_nic *efx = channel->efx;
2087 	int budget = 4;
2088 	int old_rx_packets, rx_packets;
2089 
2090 	if (!netif_running(efx->net_dev))
2091 		return LL_FLUSH_FAILED;
2092 
2093 	if (!efx_channel_lock_poll(channel))
2094 		return LL_FLUSH_BUSY;
2095 
2096 	old_rx_packets = channel->rx_queue.rx_packets;
2097 	efx_process_channel(channel, budget);
2098 
2099 	rx_packets = channel->rx_queue.rx_packets - old_rx_packets;
2100 
2101 	/* There is no race condition with NAPI here.
2102 	 * NAPI will automatically be rescheduled if it yielded during busy
2103 	 * polling, because it was not able to take the lock and thus returned
2104 	 * the full budget.
2105 	 */
2106 	efx_channel_unlock_poll(channel);
2107 
2108 	return rx_packets;
2109 }
2110 #endif
2111 
2112 /**************************************************************************
2113  *
2114  * Kernel net device interface
2115  *
2116  *************************************************************************/
2117 
2118 /* Context: process, rtnl_lock() held. */
2119 int efx_net_open(struct net_device *net_dev)
2120 {
2121 	struct efx_nic *efx = netdev_priv(net_dev);
2122 	int rc;
2123 
2124 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2125 		  raw_smp_processor_id());
2126 
2127 	rc = efx_check_disabled(efx);
2128 	if (rc)
2129 		return rc;
2130 	if (efx->phy_mode & PHY_MODE_SPECIAL)
2131 		return -EBUSY;
2132 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2133 		return -EIO;
2134 
2135 	/* Notify the kernel of the link state polled during driver load,
2136 	 * before the monitor starts running */
2137 	efx_link_status_changed(efx);
2138 
2139 	efx_start_all(efx);
2140 	efx_selftest_async_start(efx);
2141 	return 0;
2142 }
2143 
2144 /* Context: process, rtnl_lock() held.
2145  * Note that the kernel will ignore our return code; this method
2146  * should really be a void.
2147  */
2148 int efx_net_stop(struct net_device *net_dev)
2149 {
2150 	struct efx_nic *efx = netdev_priv(net_dev);
2151 
2152 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2153 		  raw_smp_processor_id());
2154 
2155 	/* Stop the device and flush all the channels */
2156 	efx_stop_all(efx);
2157 
2158 	return 0;
2159 }
2160 
2161 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2162 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
2163 					       struct rtnl_link_stats64 *stats)
2164 {
2165 	struct efx_nic *efx = netdev_priv(net_dev);
2166 
2167 	spin_lock_bh(&efx->stats_lock);
2168 	efx->type->update_stats(efx, NULL, stats);
2169 	spin_unlock_bh(&efx->stats_lock);
2170 
2171 	return stats;
2172 }
2173 
2174 /* Context: netif_tx_lock held, BHs disabled. */
2175 static void efx_watchdog(struct net_device *net_dev)
2176 {
2177 	struct efx_nic *efx = netdev_priv(net_dev);
2178 
2179 	netif_err(efx, tx_err, efx->net_dev,
2180 		  "TX stuck with port_enabled=%d: resetting channels\n",
2181 		  efx->port_enabled);
2182 
2183 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2184 }
2185 
2186 
2187 /* Context: process, rtnl_lock() held. */
2188 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2189 {
2190 	struct efx_nic *efx = netdev_priv(net_dev);
2191 	int rc;
2192 
2193 	rc = efx_check_disabled(efx);
2194 	if (rc)
2195 		return rc;
2196 	if (new_mtu > EFX_MAX_MTU)
2197 		return -EINVAL;
2198 
2199 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2200 
2201 	efx_device_detach_sync(efx);
2202 	efx_stop_all(efx);
2203 
2204 	mutex_lock(&efx->mac_lock);
2205 	net_dev->mtu = new_mtu;
2206 	efx_mac_reconfigure(efx);
2207 	mutex_unlock(&efx->mac_lock);
2208 
2209 	efx_start_all(efx);
2210 	netif_device_attach(efx->net_dev);
2211 	return 0;
2212 }
2213 
2214 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2215 {
2216 	struct efx_nic *efx = netdev_priv(net_dev);
2217 	struct sockaddr *addr = data;
2218 	u8 *new_addr = addr->sa_data;
2219 	u8 old_addr[6];
2220 	int rc;
2221 
2222 	if (!is_valid_ether_addr(new_addr)) {
2223 		netif_err(efx, drv, efx->net_dev,
2224 			  "invalid ethernet MAC address requested: %pM\n",
2225 			  new_addr);
2226 		return -EADDRNOTAVAIL;
2227 	}
2228 
2229 	/* save old address */
2230 	ether_addr_copy(old_addr, net_dev->dev_addr);
2231 	ether_addr_copy(net_dev->dev_addr, new_addr);
2232 	if (efx->type->set_mac_address) {
2233 		rc = efx->type->set_mac_address(efx);
2234 		if (rc) {
2235 			ether_addr_copy(net_dev->dev_addr, old_addr);
2236 			return rc;
2237 		}
2238 	}
2239 
2240 	/* Reconfigure the MAC */
2241 	mutex_lock(&efx->mac_lock);
2242 	efx_mac_reconfigure(efx);
2243 	mutex_unlock(&efx->mac_lock);
2244 
2245 	return 0;
2246 }
2247 
2248 /* Context: netif_addr_lock held, BHs disabled. */
2249 static void efx_set_rx_mode(struct net_device *net_dev)
2250 {
2251 	struct efx_nic *efx = netdev_priv(net_dev);
2252 
2253 	if (efx->port_enabled)
2254 		queue_work(efx->workqueue, &efx->mac_work);
2255 	/* Otherwise efx_start_port() will do this */
2256 }
2257 
2258 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2259 {
2260 	struct efx_nic *efx = netdev_priv(net_dev);
2261 
2262 	/* If disabling RX n-tuple filtering, clear existing filters */
2263 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
2264 		return efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2265 
2266 	return 0;
2267 }
2268 
2269 static const struct net_device_ops efx_netdev_ops = {
2270 	.ndo_open		= efx_net_open,
2271 	.ndo_stop		= efx_net_stop,
2272 	.ndo_get_stats64	= efx_net_stats,
2273 	.ndo_tx_timeout		= efx_watchdog,
2274 	.ndo_start_xmit		= efx_hard_start_xmit,
2275 	.ndo_validate_addr	= eth_validate_addr,
2276 	.ndo_do_ioctl		= efx_ioctl,
2277 	.ndo_change_mtu		= efx_change_mtu,
2278 	.ndo_set_mac_address	= efx_set_mac_address,
2279 	.ndo_set_rx_mode	= efx_set_rx_mode,
2280 	.ndo_set_features	= efx_set_features,
2281 #ifdef CONFIG_SFC_SRIOV
2282 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2283 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2284 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2285 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2286 	.ndo_set_vf_link_state  = efx_sriov_set_vf_link_state,
2287 	.ndo_get_phys_port_id   = efx_sriov_get_phys_port_id,
2288 #endif
2289 #ifdef CONFIG_NET_POLL_CONTROLLER
2290 	.ndo_poll_controller = efx_netpoll,
2291 #endif
2292 	.ndo_setup_tc		= efx_setup_tc,
2293 #ifdef CONFIG_NET_RX_BUSY_POLL
2294 	.ndo_busy_poll		= efx_busy_poll,
2295 #endif
2296 #ifdef CONFIG_RFS_ACCEL
2297 	.ndo_rx_flow_steer	= efx_filter_rfs,
2298 #endif
2299 };
2300 
2301 static void efx_update_name(struct efx_nic *efx)
2302 {
2303 	strcpy(efx->name, efx->net_dev->name);
2304 	efx_mtd_rename(efx);
2305 	efx_set_channel_names(efx);
2306 }
2307 
2308 static int efx_netdev_event(struct notifier_block *this,
2309 			    unsigned long event, void *ptr)
2310 {
2311 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2312 
2313 	if ((net_dev->netdev_ops == &efx_netdev_ops) &&
2314 	    event == NETDEV_CHANGENAME)
2315 		efx_update_name(netdev_priv(net_dev));
2316 
2317 	return NOTIFY_DONE;
2318 }
2319 
2320 static struct notifier_block efx_netdev_notifier = {
2321 	.notifier_call = efx_netdev_event,
2322 };
2323 
2324 static ssize_t
2325 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2326 {
2327 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2328 	return sprintf(buf, "%d\n", efx->phy_type);
2329 }
2330 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2331 
2332 #ifdef CONFIG_SFC_MCDI_LOGGING
2333 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
2334 			     char *buf)
2335 {
2336 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2337 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2338 
2339 	return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
2340 }
2341 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
2342 			    const char *buf, size_t count)
2343 {
2344 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2345 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2346 	bool enable = count > 0 && *buf != '0';
2347 
2348 	mcdi->logging_enabled = enable;
2349 	return count;
2350 }
2351 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
2352 #endif
2353 
2354 static int efx_register_netdev(struct efx_nic *efx)
2355 {
2356 	struct net_device *net_dev = efx->net_dev;
2357 	struct efx_channel *channel;
2358 	int rc;
2359 
2360 	net_dev->watchdog_timeo = 5 * HZ;
2361 	net_dev->irq = efx->pci_dev->irq;
2362 	net_dev->netdev_ops = &efx_netdev_ops;
2363 	if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
2364 		net_dev->priv_flags |= IFF_UNICAST_FLT;
2365 	net_dev->ethtool_ops = &efx_ethtool_ops;
2366 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2367 
2368 	rtnl_lock();
2369 
2370 	/* Enable resets to be scheduled and check whether any were
2371 	 * already requested.  If so, the NIC is probably hosed so we
2372 	 * abort.
2373 	 */
2374 	efx->state = STATE_READY;
2375 	smp_mb(); /* ensure we change state before checking reset_pending */
2376 	if (efx->reset_pending) {
2377 		netif_err(efx, probe, efx->net_dev,
2378 			  "aborting probe due to scheduled reset\n");
2379 		rc = -EIO;
2380 		goto fail_locked;
2381 	}
2382 
2383 	rc = dev_alloc_name(net_dev, net_dev->name);
2384 	if (rc < 0)
2385 		goto fail_locked;
2386 	efx_update_name(efx);
2387 
2388 	/* Always start with carrier off; PHY events will detect the link */
2389 	netif_carrier_off(net_dev);
2390 
2391 	rc = register_netdevice(net_dev);
2392 	if (rc)
2393 		goto fail_locked;
2394 
2395 	efx_for_each_channel(channel, efx) {
2396 		struct efx_tx_queue *tx_queue;
2397 		efx_for_each_channel_tx_queue(tx_queue, channel)
2398 			efx_init_tx_queue_core_txq(tx_queue);
2399 	}
2400 
2401 	efx_associate(efx);
2402 
2403 	rtnl_unlock();
2404 
2405 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2406 	if (rc) {
2407 		netif_err(efx, drv, efx->net_dev,
2408 			  "failed to init net dev attributes\n");
2409 		goto fail_registered;
2410 	}
2411 #ifdef CONFIG_SFC_MCDI_LOGGING
2412 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2413 	if (rc) {
2414 		netif_err(efx, drv, efx->net_dev,
2415 			  "failed to init net dev attributes\n");
2416 		goto fail_attr_mcdi_logging;
2417 	}
2418 #endif
2419 
2420 	return 0;
2421 
2422 #ifdef CONFIG_SFC_MCDI_LOGGING
2423 fail_attr_mcdi_logging:
2424 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2425 #endif
2426 fail_registered:
2427 	rtnl_lock();
2428 	efx_dissociate(efx);
2429 	unregister_netdevice(net_dev);
2430 fail_locked:
2431 	efx->state = STATE_UNINIT;
2432 	rtnl_unlock();
2433 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2434 	return rc;
2435 }
2436 
2437 static void efx_unregister_netdev(struct efx_nic *efx)
2438 {
2439 	if (!efx->net_dev)
2440 		return;
2441 
2442 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2443 
2444 	if (efx_dev_registered(efx)) {
2445 		strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2446 #ifdef CONFIG_SFC_MCDI_LOGGING
2447 		device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2448 #endif
2449 		device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2450 		unregister_netdev(efx->net_dev);
2451 	}
2452 }
2453 
2454 /**************************************************************************
2455  *
2456  * Device reset and suspend
2457  *
2458  **************************************************************************/
2459 
2460 /* Tears down the entire software state and most of the hardware state
2461  * before reset.  */
2462 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2463 {
2464 	EFX_ASSERT_RESET_SERIALISED(efx);
2465 
2466 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2467 		efx->type->prepare_flr(efx);
2468 
2469 	efx_stop_all(efx);
2470 	efx_disable_interrupts(efx);
2471 
2472 	mutex_lock(&efx->mac_lock);
2473 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2474 	    method != RESET_TYPE_DATAPATH)
2475 		efx->phy_op->fini(efx);
2476 	efx->type->fini(efx);
2477 }
2478 
2479 /* This function will always ensure that the locks acquired in
2480  * efx_reset_down() are released. A failure return code indicates
2481  * that we were unable to reinitialise the hardware, and the
2482  * driver should be disabled. If ok is false, then the rx and tx
2483  * engines are not restarted, pending a RESET_DISABLE. */
2484 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2485 {
2486 	int rc;
2487 
2488 	EFX_ASSERT_RESET_SERIALISED(efx);
2489 
2490 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2491 		efx->type->finish_flr(efx);
2492 
2493 	/* Ensure that SRAM is initialised even if we're disabling the device */
2494 	rc = efx->type->init(efx);
2495 	if (rc) {
2496 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2497 		goto fail;
2498 	}
2499 
2500 	if (!ok)
2501 		goto fail;
2502 
2503 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2504 	    method != RESET_TYPE_DATAPATH) {
2505 		rc = efx->phy_op->init(efx);
2506 		if (rc)
2507 			goto fail;
2508 		rc = efx->phy_op->reconfigure(efx);
2509 		if (rc && rc != -EPERM)
2510 			netif_err(efx, drv, efx->net_dev,
2511 				  "could not restore PHY settings\n");
2512 	}
2513 
2514 	rc = efx_enable_interrupts(efx);
2515 	if (rc)
2516 		goto fail;
2517 
2518 #ifdef CONFIG_SFC_SRIOV
2519 	rc = efx->type->vswitching_restore(efx);
2520 	if (rc) /* not fatal; the PF will still work fine */
2521 		netif_warn(efx, probe, efx->net_dev,
2522 			   "failed to restore vswitching rc=%d;"
2523 			   " VFs may not function\n", rc);
2524 #endif
2525 
2526 	down_read(&efx->filter_sem);
2527 	efx_restore_filters(efx);
2528 	up_read(&efx->filter_sem);
2529 	if (efx->type->sriov_reset)
2530 		efx->type->sriov_reset(efx);
2531 
2532 	mutex_unlock(&efx->mac_lock);
2533 
2534 	efx_start_all(efx);
2535 
2536 	return 0;
2537 
2538 fail:
2539 	efx->port_initialized = false;
2540 
2541 	mutex_unlock(&efx->mac_lock);
2542 
2543 	return rc;
2544 }
2545 
2546 /* Reset the NIC using the specified method.  Note that the reset may
2547  * fail, in which case the card will be left in an unusable state.
2548  *
2549  * Caller must hold the rtnl_lock.
2550  */
2551 int efx_reset(struct efx_nic *efx, enum reset_type method)
2552 {
2553 	int rc, rc2;
2554 	bool disabled;
2555 
2556 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2557 		   RESET_TYPE(method));
2558 
2559 	efx_device_detach_sync(efx);
2560 	efx_reset_down(efx, method);
2561 
2562 	rc = efx->type->reset(efx, method);
2563 	if (rc) {
2564 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2565 		goto out;
2566 	}
2567 
2568 	/* Clear flags for the scopes we covered.  We assume the NIC and
2569 	 * driver are now quiescent so that there is no race here.
2570 	 */
2571 	if (method < RESET_TYPE_MAX_METHOD)
2572 		efx->reset_pending &= -(1 << (method + 1));
2573 	else /* it doesn't fit into the well-ordered scope hierarchy */
2574 		__clear_bit(method, &efx->reset_pending);
2575 
2576 	/* Reinitialise bus-mastering, which may have been turned off before
2577 	 * the reset was scheduled. This is still appropriate, even in the
2578 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2579 	 * can respond to requests. */
2580 	pci_set_master(efx->pci_dev);
2581 
2582 out:
2583 	/* Leave device stopped if necessary */
2584 	disabled = rc ||
2585 		method == RESET_TYPE_DISABLE ||
2586 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2587 	rc2 = efx_reset_up(efx, method, !disabled);
2588 	if (rc2) {
2589 		disabled = true;
2590 		if (!rc)
2591 			rc = rc2;
2592 	}
2593 
2594 	if (disabled) {
2595 		dev_close(efx->net_dev);
2596 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2597 		efx->state = STATE_DISABLED;
2598 	} else {
2599 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2600 		netif_device_attach(efx->net_dev);
2601 	}
2602 	return rc;
2603 }
2604 
2605 /* Try recovery mechanisms.
2606  * For now only EEH is supported.
2607  * Returns 0 if the recovery mechanisms are unsuccessful.
2608  * Returns a non-zero value otherwise.
2609  */
2610 int efx_try_recovery(struct efx_nic *efx)
2611 {
2612 #ifdef CONFIG_EEH
2613 	/* A PCI error can occur and not be seen by EEH because nothing
2614 	 * happens on the PCI bus. In this case the driver may fail and
2615 	 * schedule a 'recover or reset', leading to this recovery handler.
2616 	 * Manually call the eeh failure check function.
2617 	 */
2618 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2619 	if (eeh_dev_check_failure(eehdev)) {
2620 		/* The EEH mechanisms will handle the error and reset the
2621 		 * device if necessary.
2622 		 */
2623 		return 1;
2624 	}
2625 #endif
2626 	return 0;
2627 }
2628 
2629 static void efx_wait_for_bist_end(struct efx_nic *efx)
2630 {
2631 	int i;
2632 
2633 	for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2634 		if (efx_mcdi_poll_reboot(efx))
2635 			goto out;
2636 		msleep(BIST_WAIT_DELAY_MS);
2637 	}
2638 
2639 	netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2640 out:
2641 	/* Either way unset the BIST flag. If we found no reboot we probably
2642 	 * won't recover, but we should try.
2643 	 */
2644 	efx->mc_bist_for_other_fn = false;
2645 }
2646 
2647 /* The worker thread exists so that code that cannot sleep can
2648  * schedule a reset for later.
2649  */
2650 static void efx_reset_work(struct work_struct *data)
2651 {
2652 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2653 	unsigned long pending;
2654 	enum reset_type method;
2655 
2656 	pending = ACCESS_ONCE(efx->reset_pending);
2657 	method = fls(pending) - 1;
2658 
2659 	if (method == RESET_TYPE_MC_BIST)
2660 		efx_wait_for_bist_end(efx);
2661 
2662 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2663 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2664 	    efx_try_recovery(efx))
2665 		return;
2666 
2667 	if (!pending)
2668 		return;
2669 
2670 	rtnl_lock();
2671 
2672 	/* We checked the state in efx_schedule_reset() but it may
2673 	 * have changed by now.  Now that we have the RTNL lock,
2674 	 * it cannot change again.
2675 	 */
2676 	if (efx->state == STATE_READY)
2677 		(void)efx_reset(efx, method);
2678 
2679 	rtnl_unlock();
2680 }
2681 
2682 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2683 {
2684 	enum reset_type method;
2685 
2686 	if (efx->state == STATE_RECOVERY) {
2687 		netif_dbg(efx, drv, efx->net_dev,
2688 			  "recovering: skip scheduling %s reset\n",
2689 			  RESET_TYPE(type));
2690 		return;
2691 	}
2692 
2693 	switch (type) {
2694 	case RESET_TYPE_INVISIBLE:
2695 	case RESET_TYPE_ALL:
2696 	case RESET_TYPE_RECOVER_OR_ALL:
2697 	case RESET_TYPE_WORLD:
2698 	case RESET_TYPE_DISABLE:
2699 	case RESET_TYPE_RECOVER_OR_DISABLE:
2700 	case RESET_TYPE_DATAPATH:
2701 	case RESET_TYPE_MC_BIST:
2702 	case RESET_TYPE_MCDI_TIMEOUT:
2703 		method = type;
2704 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2705 			  RESET_TYPE(method));
2706 		break;
2707 	default:
2708 		method = efx->type->map_reset_reason(type);
2709 		netif_dbg(efx, drv, efx->net_dev,
2710 			  "scheduling %s reset for %s\n",
2711 			  RESET_TYPE(method), RESET_TYPE(type));
2712 		break;
2713 	}
2714 
2715 	set_bit(method, &efx->reset_pending);
2716 	smp_mb(); /* ensure we change reset_pending before checking state */
2717 
2718 	/* If we're not READY then just leave the flags set as the cue
2719 	 * to abort probing or reschedule the reset later.
2720 	 */
2721 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2722 		return;
2723 
2724 	/* efx_process_channel() will no longer read events once a
2725 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2726 	efx_mcdi_mode_poll(efx);
2727 
2728 	queue_work(reset_workqueue, &efx->reset_work);
2729 }
2730 
2731 /**************************************************************************
2732  *
2733  * List of NICs we support
2734  *
2735  **************************************************************************/
2736 
2737 /* PCI device ID table */
2738 static const struct pci_device_id efx_pci_table[] = {
2739 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2740 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2741 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2742 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2743 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2744 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2745 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2746 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2747 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2748 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2749 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903),  /* SFC9120 PF */
2750 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2751 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903),  /* SFC9120 VF */
2752 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2753 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923),  /* SFC9140 PF */
2754 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2755 	{0}			/* end of list */
2756 };
2757 
2758 /**************************************************************************
2759  *
2760  * Dummy PHY/MAC operations
2761  *
2762  * Can be used for some unimplemented operations
2763  * Needed so all function pointers are valid and do not have to be tested
2764  * before use
2765  *
2766  **************************************************************************/
2767 int efx_port_dummy_op_int(struct efx_nic *efx)
2768 {
2769 	return 0;
2770 }
2771 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2772 
2773 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2774 {
2775 	return false;
2776 }
2777 
2778 static const struct efx_phy_operations efx_dummy_phy_operations = {
2779 	.init		 = efx_port_dummy_op_int,
2780 	.reconfigure	 = efx_port_dummy_op_int,
2781 	.poll		 = efx_port_dummy_op_poll,
2782 	.fini		 = efx_port_dummy_op_void,
2783 };
2784 
2785 /**************************************************************************
2786  *
2787  * Data housekeeping
2788  *
2789  **************************************************************************/
2790 
2791 /* This zeroes out and then fills in the invariants in a struct
2792  * efx_nic (including all sub-structures).
2793  */
2794 static int efx_init_struct(struct efx_nic *efx,
2795 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2796 {
2797 	int i;
2798 
2799 	/* Initialise common structures */
2800 	INIT_LIST_HEAD(&efx->node);
2801 	INIT_LIST_HEAD(&efx->secondary_list);
2802 	spin_lock_init(&efx->biu_lock);
2803 #ifdef CONFIG_SFC_MTD
2804 	INIT_LIST_HEAD(&efx->mtd_list);
2805 #endif
2806 	INIT_WORK(&efx->reset_work, efx_reset_work);
2807 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2808 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2809 	efx->pci_dev = pci_dev;
2810 	efx->msg_enable = debug;
2811 	efx->state = STATE_UNINIT;
2812 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2813 
2814 	efx->net_dev = net_dev;
2815 	efx->rx_prefix_size = efx->type->rx_prefix_size;
2816 	efx->rx_ip_align =
2817 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2818 	efx->rx_packet_hash_offset =
2819 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2820 	efx->rx_packet_ts_offset =
2821 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2822 	spin_lock_init(&efx->stats_lock);
2823 	mutex_init(&efx->mac_lock);
2824 	efx->phy_op = &efx_dummy_phy_operations;
2825 	efx->mdio.dev = net_dev;
2826 	INIT_WORK(&efx->mac_work, efx_mac_work);
2827 	init_waitqueue_head(&efx->flush_wq);
2828 
2829 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2830 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2831 		if (!efx->channel[i])
2832 			goto fail;
2833 		efx->msi_context[i].efx = efx;
2834 		efx->msi_context[i].index = i;
2835 	}
2836 
2837 	/* Higher numbered interrupt modes are less capable! */
2838 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2839 				  interrupt_mode);
2840 
2841 	/* Would be good to use the net_dev name, but we're too early */
2842 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2843 		 pci_name(pci_dev));
2844 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2845 	if (!efx->workqueue)
2846 		goto fail;
2847 
2848 	return 0;
2849 
2850 fail:
2851 	efx_fini_struct(efx);
2852 	return -ENOMEM;
2853 }
2854 
2855 static void efx_fini_struct(struct efx_nic *efx)
2856 {
2857 	int i;
2858 
2859 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2860 		kfree(efx->channel[i]);
2861 
2862 	kfree(efx->vpd_sn);
2863 
2864 	if (efx->workqueue) {
2865 		destroy_workqueue(efx->workqueue);
2866 		efx->workqueue = NULL;
2867 	}
2868 }
2869 
2870 void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
2871 {
2872 	u64 n_rx_nodesc_trunc = 0;
2873 	struct efx_channel *channel;
2874 
2875 	efx_for_each_channel(channel, efx)
2876 		n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
2877 	stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
2878 	stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
2879 }
2880 
2881 /**************************************************************************
2882  *
2883  * PCI interface
2884  *
2885  **************************************************************************/
2886 
2887 /* Main body of final NIC shutdown code
2888  * This is called only at module unload (or hotplug removal).
2889  */
2890 static void efx_pci_remove_main(struct efx_nic *efx)
2891 {
2892 	/* Flush reset_work. It can no longer be scheduled since we
2893 	 * are not READY.
2894 	 */
2895 	BUG_ON(efx->state == STATE_READY);
2896 	cancel_work_sync(&efx->reset_work);
2897 
2898 	efx_disable_interrupts(efx);
2899 	efx_nic_fini_interrupt(efx);
2900 	efx_fini_port(efx);
2901 	efx->type->fini(efx);
2902 	efx_fini_napi(efx);
2903 	efx_remove_all(efx);
2904 }
2905 
2906 /* Final NIC shutdown
2907  * This is called only at module unload (or hotplug removal).  A PF can call
2908  * this on its VFs to ensure they are unbound first.
2909  */
2910 static void efx_pci_remove(struct pci_dev *pci_dev)
2911 {
2912 	struct efx_nic *efx;
2913 
2914 	efx = pci_get_drvdata(pci_dev);
2915 	if (!efx)
2916 		return;
2917 
2918 	/* Mark the NIC as fini, then stop the interface */
2919 	rtnl_lock();
2920 	efx_dissociate(efx);
2921 	dev_close(efx->net_dev);
2922 	efx_disable_interrupts(efx);
2923 	efx->state = STATE_UNINIT;
2924 	rtnl_unlock();
2925 
2926 	if (efx->type->sriov_fini)
2927 		efx->type->sriov_fini(efx);
2928 
2929 	efx_unregister_netdev(efx);
2930 
2931 	efx_mtd_remove(efx);
2932 
2933 	efx_pci_remove_main(efx);
2934 
2935 	efx_fini_io(efx);
2936 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2937 
2938 	efx_fini_struct(efx);
2939 	free_netdev(efx->net_dev);
2940 
2941 	pci_disable_pcie_error_reporting(pci_dev);
2942 };
2943 
2944 /* NIC VPD information
2945  * Called during probe to display the part number of the
2946  * installed NIC.  VPD is potentially very large but this should
2947  * always appear within the first 512 bytes.
2948  */
2949 #define SFC_VPD_LEN 512
2950 static void efx_probe_vpd_strings(struct efx_nic *efx)
2951 {
2952 	struct pci_dev *dev = efx->pci_dev;
2953 	char vpd_data[SFC_VPD_LEN];
2954 	ssize_t vpd_size;
2955 	int ro_start, ro_size, i, j;
2956 
2957 	/* Get the vpd data from the device */
2958 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2959 	if (vpd_size <= 0) {
2960 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2961 		return;
2962 	}
2963 
2964 	/* Get the Read only section */
2965 	ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2966 	if (ro_start < 0) {
2967 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2968 		return;
2969 	}
2970 
2971 	ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
2972 	j = ro_size;
2973 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2974 	if (i + j > vpd_size)
2975 		j = vpd_size - i;
2976 
2977 	/* Get the Part number */
2978 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2979 	if (i < 0) {
2980 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2981 		return;
2982 	}
2983 
2984 	j = pci_vpd_info_field_size(&vpd_data[i]);
2985 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
2986 	if (i + j > vpd_size) {
2987 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2988 		return;
2989 	}
2990 
2991 	netif_info(efx, drv, efx->net_dev,
2992 		   "Part Number : %.*s\n", j, &vpd_data[i]);
2993 
2994 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2995 	j = ro_size;
2996 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
2997 	if (i < 0) {
2998 		netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
2999 		return;
3000 	}
3001 
3002 	j = pci_vpd_info_field_size(&vpd_data[i]);
3003 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3004 	if (i + j > vpd_size) {
3005 		netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
3006 		return;
3007 	}
3008 
3009 	efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
3010 	if (!efx->vpd_sn)
3011 		return;
3012 
3013 	snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
3014 }
3015 
3016 
3017 /* Main body of NIC initialisation
3018  * This is called at module load (or hotplug insertion, theoretically).
3019  */
3020 static int efx_pci_probe_main(struct efx_nic *efx)
3021 {
3022 	int rc;
3023 
3024 	/* Do start-of-day initialisation */
3025 	rc = efx_probe_all(efx);
3026 	if (rc)
3027 		goto fail1;
3028 
3029 	efx_init_napi(efx);
3030 
3031 	rc = efx->type->init(efx);
3032 	if (rc) {
3033 		netif_err(efx, probe, efx->net_dev,
3034 			  "failed to initialise NIC\n");
3035 		goto fail3;
3036 	}
3037 
3038 	rc = efx_init_port(efx);
3039 	if (rc) {
3040 		netif_err(efx, probe, efx->net_dev,
3041 			  "failed to initialise port\n");
3042 		goto fail4;
3043 	}
3044 
3045 	rc = efx_nic_init_interrupt(efx);
3046 	if (rc)
3047 		goto fail5;
3048 	rc = efx_enable_interrupts(efx);
3049 	if (rc)
3050 		goto fail6;
3051 
3052 	return 0;
3053 
3054  fail6:
3055 	efx_nic_fini_interrupt(efx);
3056  fail5:
3057 	efx_fini_port(efx);
3058  fail4:
3059 	efx->type->fini(efx);
3060  fail3:
3061 	efx_fini_napi(efx);
3062 	efx_remove_all(efx);
3063  fail1:
3064 	return rc;
3065 }
3066 
3067 /* NIC initialisation
3068  *
3069  * This is called at module load (or hotplug insertion,
3070  * theoretically).  It sets up PCI mappings, resets the NIC,
3071  * sets up and registers the network devices with the kernel and hooks
3072  * the interrupt service routine.  It does not prepare the device for
3073  * transmission; this is left to the first time one of the network
3074  * interfaces is brought up (i.e. efx_net_open).
3075  */
3076 static int efx_pci_probe(struct pci_dev *pci_dev,
3077 			 const struct pci_device_id *entry)
3078 {
3079 	struct net_device *net_dev;
3080 	struct efx_nic *efx;
3081 	int rc;
3082 
3083 	/* Allocate and initialise a struct net_device and struct efx_nic */
3084 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
3085 				     EFX_MAX_RX_QUEUES);
3086 	if (!net_dev)
3087 		return -ENOMEM;
3088 	efx = netdev_priv(net_dev);
3089 	efx->type = (const struct efx_nic_type *) entry->driver_data;
3090 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
3091 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
3092 			      NETIF_F_RXCSUM);
3093 	if (efx->type->offload_features & NETIF_F_V6_CSUM)
3094 		net_dev->features |= NETIF_F_TSO6;
3095 	/* Mask for features that also apply to VLAN devices */
3096 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
3097 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
3098 				   NETIF_F_RXCSUM);
3099 	/* All offloads can be toggled */
3100 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
3101 	pci_set_drvdata(pci_dev, efx);
3102 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
3103 	rc = efx_init_struct(efx, pci_dev, net_dev);
3104 	if (rc)
3105 		goto fail1;
3106 
3107 	netif_info(efx, probe, efx->net_dev,
3108 		   "Solarflare NIC detected\n");
3109 
3110 	if (!efx->type->is_vf)
3111 		efx_probe_vpd_strings(efx);
3112 
3113 	/* Set up basic I/O (BAR mappings etc) */
3114 	rc = efx_init_io(efx);
3115 	if (rc)
3116 		goto fail2;
3117 
3118 	rc = efx_pci_probe_main(efx);
3119 	if (rc)
3120 		goto fail3;
3121 
3122 	rc = efx_register_netdev(efx);
3123 	if (rc)
3124 		goto fail4;
3125 
3126 	if (efx->type->sriov_init) {
3127 		rc = efx->type->sriov_init(efx);
3128 		if (rc)
3129 			netif_err(efx, probe, efx->net_dev,
3130 				  "SR-IOV can't be enabled rc %d\n", rc);
3131 	}
3132 
3133 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3134 
3135 	/* Try to create MTDs, but allow this to fail */
3136 	rtnl_lock();
3137 	rc = efx_mtd_probe(efx);
3138 	rtnl_unlock();
3139 	if (rc)
3140 		netif_warn(efx, probe, efx->net_dev,
3141 			   "failed to create MTDs (%d)\n", rc);
3142 
3143 	rc = pci_enable_pcie_error_reporting(pci_dev);
3144 	if (rc && rc != -EINVAL)
3145 		netif_warn(efx, probe, efx->net_dev,
3146 			   "pci_enable_pcie_error_reporting failed (%d)\n", rc);
3147 
3148 	return 0;
3149 
3150  fail4:
3151 	efx_pci_remove_main(efx);
3152  fail3:
3153 	efx_fini_io(efx);
3154  fail2:
3155 	efx_fini_struct(efx);
3156  fail1:
3157 	WARN_ON(rc > 0);
3158 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3159 	free_netdev(net_dev);
3160 	return rc;
3161 }
3162 
3163 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3164  * enabled on success
3165  */
3166 #ifdef CONFIG_SFC_SRIOV
3167 static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
3168 {
3169 	int rc;
3170 	struct efx_nic *efx = pci_get_drvdata(dev);
3171 
3172 	if (efx->type->sriov_configure) {
3173 		rc = efx->type->sriov_configure(efx, num_vfs);
3174 		if (rc)
3175 			return rc;
3176 		else
3177 			return num_vfs;
3178 	} else
3179 		return -EOPNOTSUPP;
3180 }
3181 #endif
3182 
3183 static int efx_pm_freeze(struct device *dev)
3184 {
3185 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3186 
3187 	rtnl_lock();
3188 
3189 	if (efx->state != STATE_DISABLED) {
3190 		efx->state = STATE_UNINIT;
3191 
3192 		efx_device_detach_sync(efx);
3193 
3194 		efx_stop_all(efx);
3195 		efx_disable_interrupts(efx);
3196 	}
3197 
3198 	rtnl_unlock();
3199 
3200 	return 0;
3201 }
3202 
3203 static int efx_pm_thaw(struct device *dev)
3204 {
3205 	int rc;
3206 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3207 
3208 	rtnl_lock();
3209 
3210 	if (efx->state != STATE_DISABLED) {
3211 		rc = efx_enable_interrupts(efx);
3212 		if (rc)
3213 			goto fail;
3214 
3215 		mutex_lock(&efx->mac_lock);
3216 		efx->phy_op->reconfigure(efx);
3217 		mutex_unlock(&efx->mac_lock);
3218 
3219 		efx_start_all(efx);
3220 
3221 		netif_device_attach(efx->net_dev);
3222 
3223 		efx->state = STATE_READY;
3224 
3225 		efx->type->resume_wol(efx);
3226 	}
3227 
3228 	rtnl_unlock();
3229 
3230 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3231 	queue_work(reset_workqueue, &efx->reset_work);
3232 
3233 	return 0;
3234 
3235 fail:
3236 	rtnl_unlock();
3237 
3238 	return rc;
3239 }
3240 
3241 static int efx_pm_poweroff(struct device *dev)
3242 {
3243 	struct pci_dev *pci_dev = to_pci_dev(dev);
3244 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3245 
3246 	efx->type->fini(efx);
3247 
3248 	efx->reset_pending = 0;
3249 
3250 	pci_save_state(pci_dev);
3251 	return pci_set_power_state(pci_dev, PCI_D3hot);
3252 }
3253 
3254 /* Used for both resume and restore */
3255 static int efx_pm_resume(struct device *dev)
3256 {
3257 	struct pci_dev *pci_dev = to_pci_dev(dev);
3258 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3259 	int rc;
3260 
3261 	rc = pci_set_power_state(pci_dev, PCI_D0);
3262 	if (rc)
3263 		return rc;
3264 	pci_restore_state(pci_dev);
3265 	rc = pci_enable_device(pci_dev);
3266 	if (rc)
3267 		return rc;
3268 	pci_set_master(efx->pci_dev);
3269 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
3270 	if (rc)
3271 		return rc;
3272 	rc = efx->type->init(efx);
3273 	if (rc)
3274 		return rc;
3275 	rc = efx_pm_thaw(dev);
3276 	return rc;
3277 }
3278 
3279 static int efx_pm_suspend(struct device *dev)
3280 {
3281 	int rc;
3282 
3283 	efx_pm_freeze(dev);
3284 	rc = efx_pm_poweroff(dev);
3285 	if (rc)
3286 		efx_pm_resume(dev);
3287 	return rc;
3288 }
3289 
3290 static const struct dev_pm_ops efx_pm_ops = {
3291 	.suspend	= efx_pm_suspend,
3292 	.resume		= efx_pm_resume,
3293 	.freeze		= efx_pm_freeze,
3294 	.thaw		= efx_pm_thaw,
3295 	.poweroff	= efx_pm_poweroff,
3296 	.restore	= efx_pm_resume,
3297 };
3298 
3299 /* A PCI error affecting this device was detected.
3300  * At this point MMIO and DMA may be disabled.
3301  * Stop the software path and request a slot reset.
3302  */
3303 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3304 					      enum pci_channel_state state)
3305 {
3306 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3307 	struct efx_nic *efx = pci_get_drvdata(pdev);
3308 
3309 	if (state == pci_channel_io_perm_failure)
3310 		return PCI_ERS_RESULT_DISCONNECT;
3311 
3312 	rtnl_lock();
3313 
3314 	if (efx->state != STATE_DISABLED) {
3315 		efx->state = STATE_RECOVERY;
3316 		efx->reset_pending = 0;
3317 
3318 		efx_device_detach_sync(efx);
3319 
3320 		efx_stop_all(efx);
3321 		efx_disable_interrupts(efx);
3322 
3323 		status = PCI_ERS_RESULT_NEED_RESET;
3324 	} else {
3325 		/* If the interface is disabled we don't want to do anything
3326 		 * with it.
3327 		 */
3328 		status = PCI_ERS_RESULT_RECOVERED;
3329 	}
3330 
3331 	rtnl_unlock();
3332 
3333 	pci_disable_device(pdev);
3334 
3335 	return status;
3336 }
3337 
3338 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3339 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3340 {
3341 	struct efx_nic *efx = pci_get_drvdata(pdev);
3342 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3343 	int rc;
3344 
3345 	if (pci_enable_device(pdev)) {
3346 		netif_err(efx, hw, efx->net_dev,
3347 			  "Cannot re-enable PCI device after reset.\n");
3348 		status =  PCI_ERS_RESULT_DISCONNECT;
3349 	}
3350 
3351 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3352 	if (rc) {
3353 		netif_err(efx, hw, efx->net_dev,
3354 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3355 		/* Non-fatal error. Continue. */
3356 	}
3357 
3358 	return status;
3359 }
3360 
3361 /* Perform the actual reset and resume I/O operations. */
3362 static void efx_io_resume(struct pci_dev *pdev)
3363 {
3364 	struct efx_nic *efx = pci_get_drvdata(pdev);
3365 	int rc;
3366 
3367 	rtnl_lock();
3368 
3369 	if (efx->state == STATE_DISABLED)
3370 		goto out;
3371 
3372 	rc = efx_reset(efx, RESET_TYPE_ALL);
3373 	if (rc) {
3374 		netif_err(efx, hw, efx->net_dev,
3375 			  "efx_reset failed after PCI error (%d)\n", rc);
3376 	} else {
3377 		efx->state = STATE_READY;
3378 		netif_dbg(efx, hw, efx->net_dev,
3379 			  "Done resetting and resuming IO after PCI error.\n");
3380 	}
3381 
3382 out:
3383 	rtnl_unlock();
3384 }
3385 
3386 /* For simplicity and reliability, we always require a slot reset and try to
3387  * reset the hardware when a pci error affecting the device is detected.
3388  * We leave both the link_reset and mmio_enabled callback unimplemented:
3389  * with our request for slot reset the mmio_enabled callback will never be
3390  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3391  */
3392 static struct pci_error_handlers efx_err_handlers = {
3393 	.error_detected = efx_io_error_detected,
3394 	.slot_reset	= efx_io_slot_reset,
3395 	.resume		= efx_io_resume,
3396 };
3397 
3398 static struct pci_driver efx_pci_driver = {
3399 	.name		= KBUILD_MODNAME,
3400 	.id_table	= efx_pci_table,
3401 	.probe		= efx_pci_probe,
3402 	.remove		= efx_pci_remove,
3403 	.driver.pm	= &efx_pm_ops,
3404 	.err_handler	= &efx_err_handlers,
3405 #ifdef CONFIG_SFC_SRIOV
3406 	.sriov_configure = efx_pci_sriov_configure,
3407 #endif
3408 };
3409 
3410 /**************************************************************************
3411  *
3412  * Kernel module interface
3413  *
3414  *************************************************************************/
3415 
3416 module_param(interrupt_mode, uint, 0444);
3417 MODULE_PARM_DESC(interrupt_mode,
3418 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3419 
3420 static int __init efx_init_module(void)
3421 {
3422 	int rc;
3423 
3424 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3425 
3426 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3427 	if (rc)
3428 		goto err_notifier;
3429 
3430 #ifdef CONFIG_SFC_SRIOV
3431 	rc = efx_init_sriov();
3432 	if (rc)
3433 		goto err_sriov;
3434 #endif
3435 
3436 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3437 	if (!reset_workqueue) {
3438 		rc = -ENOMEM;
3439 		goto err_reset;
3440 	}
3441 
3442 	rc = pci_register_driver(&efx_pci_driver);
3443 	if (rc < 0)
3444 		goto err_pci;
3445 
3446 	return 0;
3447 
3448  err_pci:
3449 	destroy_workqueue(reset_workqueue);
3450  err_reset:
3451 #ifdef CONFIG_SFC_SRIOV
3452 	efx_fini_sriov();
3453  err_sriov:
3454 #endif
3455 	unregister_netdevice_notifier(&efx_netdev_notifier);
3456  err_notifier:
3457 	return rc;
3458 }
3459 
3460 static void __exit efx_exit_module(void)
3461 {
3462 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3463 
3464 	pci_unregister_driver(&efx_pci_driver);
3465 	destroy_workqueue(reset_workqueue);
3466 #ifdef CONFIG_SFC_SRIOV
3467 	efx_fini_sriov();
3468 #endif
3469 	unregister_netdevice_notifier(&efx_netdev_notifier);
3470 
3471 }
3472 
3473 module_init(efx_init_module);
3474 module_exit(efx_exit_module);
3475 
3476 MODULE_AUTHOR("Solarflare Communications and "
3477 	      "Michael Brown <mbrown@fensystems.co.uk>");
3478 MODULE_DESCRIPTION("Solarflare network driver");
3479 MODULE_LICENSE("GPL");
3480 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3481