xref: /openbmc/linux/drivers/net/ethernet/sfc/efx.c (revision a8c5cb99)
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 <net/gre.h>
27 #include <net/udp_tunnel.h>
28 #include "efx.h"
29 #include "nic.h"
30 #include "io.h"
31 #include "selftest.h"
32 #include "sriov.h"
33 
34 #include "mcdi.h"
35 #include "mcdi_pcol.h"
36 #include "workarounds.h"
37 
38 /**************************************************************************
39  *
40  * Type name strings
41  *
42  **************************************************************************
43  */
44 
45 /* Loopback mode names (see LOOPBACK_MODE()) */
46 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
47 const char *const efx_loopback_mode_names[] = {
48 	[LOOPBACK_NONE]		= "NONE",
49 	[LOOPBACK_DATA]		= "DATAPATH",
50 	[LOOPBACK_GMAC]		= "GMAC",
51 	[LOOPBACK_XGMII]	= "XGMII",
52 	[LOOPBACK_XGXS]		= "XGXS",
53 	[LOOPBACK_XAUI]		= "XAUI",
54 	[LOOPBACK_GMII]		= "GMII",
55 	[LOOPBACK_SGMII]	= "SGMII",
56 	[LOOPBACK_XGBR]		= "XGBR",
57 	[LOOPBACK_XFI]		= "XFI",
58 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
59 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
60 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
61 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
62 	[LOOPBACK_GPHY]		= "GPHY",
63 	[LOOPBACK_PHYXS]	= "PHYXS",
64 	[LOOPBACK_PCS]		= "PCS",
65 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
66 	[LOOPBACK_XPORT]	= "XPORT",
67 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
68 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
69 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
70 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
71 	[LOOPBACK_GMII_WS]	= "GMII_WS",
72 	[LOOPBACK_XFI_WS]	= "XFI_WS",
73 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
74 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
75 };
76 
77 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
78 const char *const efx_reset_type_names[] = {
79 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
80 	[RESET_TYPE_ALL]                = "ALL",
81 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
82 	[RESET_TYPE_WORLD]              = "WORLD",
83 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
84 	[RESET_TYPE_DATAPATH]           = "DATAPATH",
85 	[RESET_TYPE_MC_BIST]		= "MC_BIST",
86 	[RESET_TYPE_DISABLE]            = "DISABLE",
87 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
88 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
89 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
90 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
91 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
92 	[RESET_TYPE_MCDI_TIMEOUT]	= "MCDI_TIMEOUT (FLR)",
93 };
94 
95 /* UDP tunnel type names */
96 static const char *const efx_udp_tunnel_type_names[] = {
97 	[TUNNEL_ENCAP_UDP_PORT_ENTRY_VXLAN] = "vxlan",
98 	[TUNNEL_ENCAP_UDP_PORT_ENTRY_GENEVE] = "geneve",
99 };
100 
101 void efx_get_udp_tunnel_type_name(u16 type, char *buf, size_t buflen)
102 {
103 	if (type < ARRAY_SIZE(efx_udp_tunnel_type_names) &&
104 	    efx_udp_tunnel_type_names[type] != NULL)
105 		snprintf(buf, buflen, "%s", efx_udp_tunnel_type_names[type]);
106 	else
107 		snprintf(buf, buflen, "type %d", type);
108 }
109 
110 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
111  * queued onto this work queue. This is not a per-nic work queue, because
112  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
113  */
114 static struct workqueue_struct *reset_workqueue;
115 
116 /* How often and how many times to poll for a reset while waiting for a
117  * BIST that another function started to complete.
118  */
119 #define BIST_WAIT_DELAY_MS	100
120 #define BIST_WAIT_DELAY_COUNT	100
121 
122 /**************************************************************************
123  *
124  * Configurable values
125  *
126  *************************************************************************/
127 
128 /*
129  * Use separate channels for TX and RX events
130  *
131  * Set this to 1 to use separate channels for TX and RX. It allows us
132  * to control interrupt affinity separately for TX and RX.
133  *
134  * This is only used in MSI-X interrupt mode
135  */
136 bool efx_separate_tx_channels;
137 module_param(efx_separate_tx_channels, bool, 0444);
138 MODULE_PARM_DESC(efx_separate_tx_channels,
139 		 "Use separate channels for TX and RX");
140 
141 /* This is the weight assigned to each of the (per-channel) virtual
142  * NAPI devices.
143  */
144 static int napi_weight = 64;
145 
146 /* This is the time (in jiffies) between invocations of the hardware
147  * monitor.
148  * On Falcon-based NICs, this will:
149  * - Check the on-board hardware monitor;
150  * - Poll the link state and reconfigure the hardware as necessary.
151  * On Siena-based NICs for power systems with EEH support, this will give EEH a
152  * chance to start.
153  */
154 static unsigned int efx_monitor_interval = 1 * HZ;
155 
156 /* Initial interrupt moderation settings.  They can be modified after
157  * module load with ethtool.
158  *
159  * The default for RX should strike a balance between increasing the
160  * round-trip latency and reducing overhead.
161  */
162 static unsigned int rx_irq_mod_usec = 60;
163 
164 /* Initial interrupt moderation settings.  They can be modified after
165  * module load with ethtool.
166  *
167  * This default is chosen to ensure that a 10G link does not go idle
168  * while a TX queue is stopped after it has become full.  A queue is
169  * restarted when it drops below half full.  The time this takes (assuming
170  * worst case 3 descriptors per packet and 1024 descriptors) is
171  *   512 / 3 * 1.2 = 205 usec.
172  */
173 static unsigned int tx_irq_mod_usec = 150;
174 
175 /* This is the first interrupt mode to try out of:
176  * 0 => MSI-X
177  * 1 => MSI
178  * 2 => legacy
179  */
180 static unsigned int interrupt_mode;
181 
182 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
183  * i.e. the number of CPUs among which we may distribute simultaneous
184  * interrupt handling.
185  *
186  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
187  * The default (0) means to assign an interrupt to each core.
188  */
189 static unsigned int rss_cpus;
190 module_param(rss_cpus, uint, 0444);
191 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
192 
193 static bool phy_flash_cfg;
194 module_param(phy_flash_cfg, bool, 0644);
195 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
196 
197 static unsigned irq_adapt_low_thresh = 8000;
198 module_param(irq_adapt_low_thresh, uint, 0644);
199 MODULE_PARM_DESC(irq_adapt_low_thresh,
200 		 "Threshold score for reducing IRQ moderation");
201 
202 static unsigned irq_adapt_high_thresh = 16000;
203 module_param(irq_adapt_high_thresh, uint, 0644);
204 MODULE_PARM_DESC(irq_adapt_high_thresh,
205 		 "Threshold score for increasing IRQ moderation");
206 
207 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
208 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
209 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
210 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
211 module_param(debug, uint, 0);
212 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
213 
214 /**************************************************************************
215  *
216  * Utility functions and prototypes
217  *
218  *************************************************************************/
219 
220 static int efx_soft_enable_interrupts(struct efx_nic *efx);
221 static void efx_soft_disable_interrupts(struct efx_nic *efx);
222 static void efx_remove_channel(struct efx_channel *channel);
223 static void efx_remove_channels(struct efx_nic *efx);
224 static const struct efx_channel_type efx_default_channel_type;
225 static void efx_remove_port(struct efx_nic *efx);
226 static void efx_init_napi_channel(struct efx_channel *channel);
227 static void efx_fini_napi(struct efx_nic *efx);
228 static void efx_fini_napi_channel(struct efx_channel *channel);
229 static void efx_fini_struct(struct efx_nic *efx);
230 static void efx_start_all(struct efx_nic *efx);
231 static void efx_stop_all(struct efx_nic *efx);
232 
233 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
234 	do {						\
235 		if ((efx->state == STATE_READY) ||	\
236 		    (efx->state == STATE_RECOVERY) ||	\
237 		    (efx->state == STATE_DISABLED))	\
238 			ASSERT_RTNL();			\
239 	} while (0)
240 
241 static int efx_check_disabled(struct efx_nic *efx)
242 {
243 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
244 		netif_err(efx, drv, efx->net_dev,
245 			  "device is disabled due to earlier errors\n");
246 		return -EIO;
247 	}
248 	return 0;
249 }
250 
251 /**************************************************************************
252  *
253  * Event queue processing
254  *
255  *************************************************************************/
256 
257 /* Process channel's event queue
258  *
259  * This function is responsible for processing the event queue of a
260  * single channel.  The caller must guarantee that this function will
261  * never be concurrently called more than once on the same channel,
262  * though different channels may be being processed concurrently.
263  */
264 static int efx_process_channel(struct efx_channel *channel, int budget)
265 {
266 	struct efx_tx_queue *tx_queue;
267 	int spent;
268 
269 	if (unlikely(!channel->enabled))
270 		return 0;
271 
272 	efx_for_each_channel_tx_queue(tx_queue, channel) {
273 		tx_queue->pkts_compl = 0;
274 		tx_queue->bytes_compl = 0;
275 	}
276 
277 	spent = efx_nic_process_eventq(channel, budget);
278 	if (spent && efx_channel_has_rx_queue(channel)) {
279 		struct efx_rx_queue *rx_queue =
280 			efx_channel_get_rx_queue(channel);
281 
282 		efx_rx_flush_packet(channel);
283 		efx_fast_push_rx_descriptors(rx_queue, true);
284 	}
285 
286 	/* Update BQL */
287 	efx_for_each_channel_tx_queue(tx_queue, channel) {
288 		if (tx_queue->bytes_compl) {
289 			netdev_tx_completed_queue(tx_queue->core_txq,
290 				tx_queue->pkts_compl, tx_queue->bytes_compl);
291 		}
292 	}
293 
294 	return spent;
295 }
296 
297 /* NAPI poll handler
298  *
299  * NAPI guarantees serialisation of polls of the same device, which
300  * provides the guarantee required by efx_process_channel().
301  */
302 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel)
303 {
304 	int step = efx->irq_mod_step_us;
305 
306 	if (channel->irq_mod_score < irq_adapt_low_thresh) {
307 		if (channel->irq_moderation_us > step) {
308 			channel->irq_moderation_us -= step;
309 			efx->type->push_irq_moderation(channel);
310 		}
311 	} else if (channel->irq_mod_score > irq_adapt_high_thresh) {
312 		if (channel->irq_moderation_us <
313 		    efx->irq_rx_moderation_us) {
314 			channel->irq_moderation_us += step;
315 			efx->type->push_irq_moderation(channel);
316 		}
317 	}
318 
319 	channel->irq_count = 0;
320 	channel->irq_mod_score = 0;
321 }
322 
323 static int efx_poll(struct napi_struct *napi, int budget)
324 {
325 	struct efx_channel *channel =
326 		container_of(napi, struct efx_channel, napi_str);
327 	struct efx_nic *efx = channel->efx;
328 	int spent;
329 
330 	netif_vdbg(efx, intr, efx->net_dev,
331 		   "channel %d NAPI poll executing on CPU %d\n",
332 		   channel->channel, raw_smp_processor_id());
333 
334 	spent = efx_process_channel(channel, budget);
335 
336 	if (spent < budget) {
337 		if (efx_channel_has_rx_queue(channel) &&
338 		    efx->irq_rx_adaptive &&
339 		    unlikely(++channel->irq_count == 1000)) {
340 			efx_update_irq_mod(efx, channel);
341 		}
342 
343 #ifdef CONFIG_RFS_ACCEL
344 		/* Perhaps expire some ARFS filters */
345 		schedule_work(&channel->filter_work);
346 #endif
347 
348 		/* There is no race here; although napi_disable() will
349 		 * only wait for napi_complete(), this isn't a problem
350 		 * since efx_nic_eventq_read_ack() will have no effect if
351 		 * interrupts have already been disabled.
352 		 */
353 		if (napi_complete_done(napi, spent))
354 			efx_nic_eventq_read_ack(channel);
355 	}
356 
357 	return spent;
358 }
359 
360 /* Create event queue
361  * Event queue memory allocations are done only once.  If the channel
362  * is reset, the memory buffer will be reused; this guards against
363  * errors during channel reset and also simplifies interrupt handling.
364  */
365 static int efx_probe_eventq(struct efx_channel *channel)
366 {
367 	struct efx_nic *efx = channel->efx;
368 	unsigned long entries;
369 
370 	netif_dbg(efx, probe, efx->net_dev,
371 		  "chan %d create event queue\n", channel->channel);
372 
373 	/* Build an event queue with room for one event per tx and rx buffer,
374 	 * plus some extra for link state events and MCDI completions. */
375 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
376 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
377 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
378 
379 	return efx_nic_probe_eventq(channel);
380 }
381 
382 /* Prepare channel's event queue */
383 static int efx_init_eventq(struct efx_channel *channel)
384 {
385 	struct efx_nic *efx = channel->efx;
386 	int rc;
387 
388 	EFX_WARN_ON_PARANOID(channel->eventq_init);
389 
390 	netif_dbg(efx, drv, efx->net_dev,
391 		  "chan %d init event queue\n", channel->channel);
392 
393 	rc = efx_nic_init_eventq(channel);
394 	if (rc == 0) {
395 		efx->type->push_irq_moderation(channel);
396 		channel->eventq_read_ptr = 0;
397 		channel->eventq_init = true;
398 	}
399 	return rc;
400 }
401 
402 /* Enable event queue processing and NAPI */
403 void efx_start_eventq(struct efx_channel *channel)
404 {
405 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
406 		  "chan %d start event queue\n", channel->channel);
407 
408 	/* Make sure the NAPI handler sees the enabled flag set */
409 	channel->enabled = true;
410 	smp_wmb();
411 
412 	napi_enable(&channel->napi_str);
413 	efx_nic_eventq_read_ack(channel);
414 }
415 
416 /* Disable event queue processing and NAPI */
417 void efx_stop_eventq(struct efx_channel *channel)
418 {
419 	if (!channel->enabled)
420 		return;
421 
422 	napi_disable(&channel->napi_str);
423 	channel->enabled = false;
424 }
425 
426 static void efx_fini_eventq(struct efx_channel *channel)
427 {
428 	if (!channel->eventq_init)
429 		return;
430 
431 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
432 		  "chan %d fini event queue\n", channel->channel);
433 
434 	efx_nic_fini_eventq(channel);
435 	channel->eventq_init = false;
436 }
437 
438 static void efx_remove_eventq(struct efx_channel *channel)
439 {
440 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
441 		  "chan %d remove event queue\n", channel->channel);
442 
443 	efx_nic_remove_eventq(channel);
444 }
445 
446 /**************************************************************************
447  *
448  * Channel handling
449  *
450  *************************************************************************/
451 
452 /* Allocate and initialise a channel structure. */
453 static struct efx_channel *
454 efx_alloc_channel(struct efx_nic *efx, int i, 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 = kzalloc(sizeof(*channel), GFP_KERNEL);
462 	if (!channel)
463 		return NULL;
464 
465 	channel->efx = efx;
466 	channel->channel = i;
467 	channel->type = &efx_default_channel_type;
468 
469 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
470 		tx_queue = &channel->tx_queue[j];
471 		tx_queue->efx = efx;
472 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
473 		tx_queue->channel = channel;
474 	}
475 
476 #ifdef CONFIG_RFS_ACCEL
477 	INIT_WORK(&channel->filter_work, efx_filter_rfs_expire);
478 #endif
479 
480 	rx_queue = &channel->rx_queue;
481 	rx_queue->efx = efx;
482 	timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
483 
484 	return channel;
485 }
486 
487 /* Allocate and initialise a channel structure, copying parameters
488  * (but not resources) from an old channel structure.
489  */
490 static struct efx_channel *
491 efx_copy_channel(const struct efx_channel *old_channel)
492 {
493 	struct efx_channel *channel;
494 	struct efx_rx_queue *rx_queue;
495 	struct efx_tx_queue *tx_queue;
496 	int j;
497 
498 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
499 	if (!channel)
500 		return NULL;
501 
502 	*channel = *old_channel;
503 
504 	channel->napi_dev = NULL;
505 	INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
506 	channel->napi_str.napi_id = 0;
507 	channel->napi_str.state = 0;
508 	memset(&channel->eventq, 0, sizeof(channel->eventq));
509 
510 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
511 		tx_queue = &channel->tx_queue[j];
512 		if (tx_queue->channel)
513 			tx_queue->channel = channel;
514 		tx_queue->buffer = NULL;
515 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
516 	}
517 
518 	rx_queue = &channel->rx_queue;
519 	rx_queue->buffer = NULL;
520 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
521 	timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
522 #ifdef CONFIG_RFS_ACCEL
523 	INIT_WORK(&channel->filter_work, efx_filter_rfs_expire);
524 #endif
525 
526 	return channel;
527 }
528 
529 static int efx_probe_channel(struct efx_channel *channel)
530 {
531 	struct efx_tx_queue *tx_queue;
532 	struct efx_rx_queue *rx_queue;
533 	int rc;
534 
535 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
536 		  "creating channel %d\n", channel->channel);
537 
538 	rc = channel->type->pre_probe(channel);
539 	if (rc)
540 		goto fail;
541 
542 	rc = efx_probe_eventq(channel);
543 	if (rc)
544 		goto fail;
545 
546 	efx_for_each_channel_tx_queue(tx_queue, channel) {
547 		rc = efx_probe_tx_queue(tx_queue);
548 		if (rc)
549 			goto fail;
550 	}
551 
552 	efx_for_each_channel_rx_queue(rx_queue, channel) {
553 		rc = efx_probe_rx_queue(rx_queue);
554 		if (rc)
555 			goto fail;
556 	}
557 
558 	return 0;
559 
560 fail:
561 	efx_remove_channel(channel);
562 	return rc;
563 }
564 
565 static void
566 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
567 {
568 	struct efx_nic *efx = channel->efx;
569 	const char *type;
570 	int number;
571 
572 	number = channel->channel;
573 	if (efx->tx_channel_offset == 0) {
574 		type = "";
575 	} else if (channel->channel < efx->tx_channel_offset) {
576 		type = "-rx";
577 	} else {
578 		type = "-tx";
579 		number -= efx->tx_channel_offset;
580 	}
581 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
582 }
583 
584 static void efx_set_channel_names(struct efx_nic *efx)
585 {
586 	struct efx_channel *channel;
587 
588 	efx_for_each_channel(channel, efx)
589 		channel->type->get_name(channel,
590 					efx->msi_context[channel->channel].name,
591 					sizeof(efx->msi_context[0].name));
592 }
593 
594 static int efx_probe_channels(struct efx_nic *efx)
595 {
596 	struct efx_channel *channel;
597 	int rc;
598 
599 	/* Restart special buffer allocation */
600 	efx->next_buffer_table = 0;
601 
602 	/* Probe channels in reverse, so that any 'extra' channels
603 	 * use the start of the buffer table. This allows the traffic
604 	 * channels to be resized without moving them or wasting the
605 	 * entries before them.
606 	 */
607 	efx_for_each_channel_rev(channel, efx) {
608 		rc = efx_probe_channel(channel);
609 		if (rc) {
610 			netif_err(efx, probe, efx->net_dev,
611 				  "failed to create channel %d\n",
612 				  channel->channel);
613 			goto fail;
614 		}
615 	}
616 	efx_set_channel_names(efx);
617 
618 	return 0;
619 
620 fail:
621 	efx_remove_channels(efx);
622 	return rc;
623 }
624 
625 /* Channels are shutdown and reinitialised whilst the NIC is running
626  * to propagate configuration changes (mtu, checksum offload), or
627  * to clear hardware error conditions
628  */
629 static void efx_start_datapath(struct efx_nic *efx)
630 {
631 	netdev_features_t old_features = efx->net_dev->features;
632 	bool old_rx_scatter = efx->rx_scatter;
633 	struct efx_tx_queue *tx_queue;
634 	struct efx_rx_queue *rx_queue;
635 	struct efx_channel *channel;
636 	size_t rx_buf_len;
637 
638 	/* Calculate the rx buffer allocation parameters required to
639 	 * support the current MTU, including padding for header
640 	 * alignment and overruns.
641 	 */
642 	efx->rx_dma_len = (efx->rx_prefix_size +
643 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
644 			   efx->type->rx_buffer_padding);
645 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
646 		      efx->rx_ip_align + efx->rx_dma_len);
647 	if (rx_buf_len <= PAGE_SIZE) {
648 		efx->rx_scatter = efx->type->always_rx_scatter;
649 		efx->rx_buffer_order = 0;
650 	} else if (efx->type->can_rx_scatter) {
651 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
652 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
653 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
654 				       EFX_RX_BUF_ALIGNMENT) >
655 			     PAGE_SIZE);
656 		efx->rx_scatter = true;
657 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
658 		efx->rx_buffer_order = 0;
659 	} else {
660 		efx->rx_scatter = false;
661 		efx->rx_buffer_order = get_order(rx_buf_len);
662 	}
663 
664 	efx_rx_config_page_split(efx);
665 	if (efx->rx_buffer_order)
666 		netif_dbg(efx, drv, efx->net_dev,
667 			  "RX buf len=%u; page order=%u batch=%u\n",
668 			  efx->rx_dma_len, efx->rx_buffer_order,
669 			  efx->rx_pages_per_batch);
670 	else
671 		netif_dbg(efx, drv, efx->net_dev,
672 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
673 			  efx->rx_dma_len, efx->rx_page_buf_step,
674 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
675 
676 	/* Restore previously fixed features in hw_features and remove
677 	 * features which are fixed now
678 	 */
679 	efx->net_dev->hw_features |= efx->net_dev->features;
680 	efx->net_dev->hw_features &= ~efx->fixed_features;
681 	efx->net_dev->features |= efx->fixed_features;
682 	if (efx->net_dev->features != old_features)
683 		netdev_features_change(efx->net_dev);
684 
685 	/* RX filters may also have scatter-enabled flags */
686 	if (efx->rx_scatter != old_rx_scatter)
687 		efx->type->filter_update_rx_scatter(efx);
688 
689 	/* We must keep at least one descriptor in a TX ring empty.
690 	 * We could avoid this when the queue size does not exactly
691 	 * match the hardware ring size, but it's not that important.
692 	 * Therefore we stop the queue when one more skb might fill
693 	 * the ring completely.  We wake it when half way back to
694 	 * empty.
695 	 */
696 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
697 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
698 
699 	/* Initialise the channels */
700 	efx_for_each_channel(channel, efx) {
701 		efx_for_each_channel_tx_queue(tx_queue, channel) {
702 			efx_init_tx_queue(tx_queue);
703 			atomic_inc(&efx->active_queues);
704 		}
705 
706 		efx_for_each_channel_rx_queue(rx_queue, channel) {
707 			efx_init_rx_queue(rx_queue);
708 			atomic_inc(&efx->active_queues);
709 			efx_stop_eventq(channel);
710 			efx_fast_push_rx_descriptors(rx_queue, false);
711 			efx_start_eventq(channel);
712 		}
713 
714 		WARN_ON(channel->rx_pkt_n_frags);
715 	}
716 
717 	efx_ptp_start_datapath(efx);
718 
719 	if (netif_device_present(efx->net_dev))
720 		netif_tx_wake_all_queues(efx->net_dev);
721 }
722 
723 static void efx_stop_datapath(struct efx_nic *efx)
724 {
725 	struct efx_channel *channel;
726 	struct efx_tx_queue *tx_queue;
727 	struct efx_rx_queue *rx_queue;
728 	int rc;
729 
730 	EFX_ASSERT_RESET_SERIALISED(efx);
731 	BUG_ON(efx->port_enabled);
732 
733 	efx_ptp_stop_datapath(efx);
734 
735 	/* Stop RX refill */
736 	efx_for_each_channel(channel, efx) {
737 		efx_for_each_channel_rx_queue(rx_queue, channel)
738 			rx_queue->refill_enabled = false;
739 	}
740 
741 	efx_for_each_channel(channel, efx) {
742 		/* RX packet processing is pipelined, so wait for the
743 		 * NAPI handler to complete.  At least event queue 0
744 		 * might be kept active by non-data events, so don't
745 		 * use napi_synchronize() but actually disable NAPI
746 		 * temporarily.
747 		 */
748 		if (efx_channel_has_rx_queue(channel)) {
749 			efx_stop_eventq(channel);
750 			efx_start_eventq(channel);
751 		}
752 	}
753 
754 	rc = efx->type->fini_dmaq(efx);
755 	if (rc) {
756 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
757 	} else {
758 		netif_dbg(efx, drv, efx->net_dev,
759 			  "successfully flushed all queues\n");
760 	}
761 
762 	efx_for_each_channel(channel, efx) {
763 		efx_for_each_channel_rx_queue(rx_queue, channel)
764 			efx_fini_rx_queue(rx_queue);
765 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
766 			efx_fini_tx_queue(tx_queue);
767 	}
768 }
769 
770 static void efx_remove_channel(struct efx_channel *channel)
771 {
772 	struct efx_tx_queue *tx_queue;
773 	struct efx_rx_queue *rx_queue;
774 
775 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
776 		  "destroy chan %d\n", channel->channel);
777 
778 	efx_for_each_channel_rx_queue(rx_queue, channel)
779 		efx_remove_rx_queue(rx_queue);
780 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
781 		efx_remove_tx_queue(tx_queue);
782 	efx_remove_eventq(channel);
783 	channel->type->post_remove(channel);
784 }
785 
786 static void efx_remove_channels(struct efx_nic *efx)
787 {
788 	struct efx_channel *channel;
789 
790 	efx_for_each_channel(channel, efx)
791 		efx_remove_channel(channel);
792 }
793 
794 int
795 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
796 {
797 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
798 	u32 old_rxq_entries, old_txq_entries;
799 	unsigned i, next_buffer_table = 0;
800 	int rc, rc2;
801 
802 	rc = efx_check_disabled(efx);
803 	if (rc)
804 		return rc;
805 
806 	/* Not all channels should be reallocated. We must avoid
807 	 * reallocating their buffer table entries.
808 	 */
809 	efx_for_each_channel(channel, efx) {
810 		struct efx_rx_queue *rx_queue;
811 		struct efx_tx_queue *tx_queue;
812 
813 		if (channel->type->copy)
814 			continue;
815 		next_buffer_table = max(next_buffer_table,
816 					channel->eventq.index +
817 					channel->eventq.entries);
818 		efx_for_each_channel_rx_queue(rx_queue, channel)
819 			next_buffer_table = max(next_buffer_table,
820 						rx_queue->rxd.index +
821 						rx_queue->rxd.entries);
822 		efx_for_each_channel_tx_queue(tx_queue, channel)
823 			next_buffer_table = max(next_buffer_table,
824 						tx_queue->txd.index +
825 						tx_queue->txd.entries);
826 	}
827 
828 	efx_device_detach_sync(efx);
829 	efx_stop_all(efx);
830 	efx_soft_disable_interrupts(efx);
831 
832 	/* Clone channels (where possible) */
833 	memset(other_channel, 0, sizeof(other_channel));
834 	for (i = 0; i < efx->n_channels; i++) {
835 		channel = efx->channel[i];
836 		if (channel->type->copy)
837 			channel = channel->type->copy(channel);
838 		if (!channel) {
839 			rc = -ENOMEM;
840 			goto out;
841 		}
842 		other_channel[i] = channel;
843 	}
844 
845 	/* Swap entry counts and channel pointers */
846 	old_rxq_entries = efx->rxq_entries;
847 	old_txq_entries = efx->txq_entries;
848 	efx->rxq_entries = rxq_entries;
849 	efx->txq_entries = txq_entries;
850 	for (i = 0; i < efx->n_channels; i++) {
851 		channel = efx->channel[i];
852 		efx->channel[i] = other_channel[i];
853 		other_channel[i] = channel;
854 	}
855 
856 	/* Restart buffer table allocation */
857 	efx->next_buffer_table = next_buffer_table;
858 
859 	for (i = 0; i < efx->n_channels; i++) {
860 		channel = efx->channel[i];
861 		if (!channel->type->copy)
862 			continue;
863 		rc = efx_probe_channel(channel);
864 		if (rc)
865 			goto rollback;
866 		efx_init_napi_channel(efx->channel[i]);
867 	}
868 
869 out:
870 	/* Destroy unused channel structures */
871 	for (i = 0; i < efx->n_channels; i++) {
872 		channel = other_channel[i];
873 		if (channel && channel->type->copy) {
874 			efx_fini_napi_channel(channel);
875 			efx_remove_channel(channel);
876 			kfree(channel);
877 		}
878 	}
879 
880 	rc2 = efx_soft_enable_interrupts(efx);
881 	if (rc2) {
882 		rc = rc ? rc : rc2;
883 		netif_err(efx, drv, efx->net_dev,
884 			  "unable to restart interrupts on channel reallocation\n");
885 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
886 	} else {
887 		efx_start_all(efx);
888 		efx_device_attach_if_not_resetting(efx);
889 	}
890 	return rc;
891 
892 rollback:
893 	/* Swap back */
894 	efx->rxq_entries = old_rxq_entries;
895 	efx->txq_entries = old_txq_entries;
896 	for (i = 0; i < efx->n_channels; i++) {
897 		channel = efx->channel[i];
898 		efx->channel[i] = other_channel[i];
899 		other_channel[i] = channel;
900 	}
901 	goto out;
902 }
903 
904 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
905 {
906 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
907 }
908 
909 static bool efx_default_channel_want_txqs(struct efx_channel *channel)
910 {
911 	return channel->channel - channel->efx->tx_channel_offset <
912 		channel->efx->n_tx_channels;
913 }
914 
915 static const struct efx_channel_type efx_default_channel_type = {
916 	.pre_probe		= efx_channel_dummy_op_int,
917 	.post_remove		= efx_channel_dummy_op_void,
918 	.get_name		= efx_get_channel_name,
919 	.copy			= efx_copy_channel,
920 	.want_txqs		= efx_default_channel_want_txqs,
921 	.keep_eventq		= false,
922 	.want_pio		= true,
923 };
924 
925 int efx_channel_dummy_op_int(struct efx_channel *channel)
926 {
927 	return 0;
928 }
929 
930 void efx_channel_dummy_op_void(struct efx_channel *channel)
931 {
932 }
933 
934 /**************************************************************************
935  *
936  * Port handling
937  *
938  **************************************************************************/
939 
940 /* This ensures that the kernel is kept informed (via
941  * netif_carrier_on/off) of the link status, and also maintains the
942  * link status's stop on the port's TX queue.
943  */
944 void efx_link_status_changed(struct efx_nic *efx)
945 {
946 	struct efx_link_state *link_state = &efx->link_state;
947 
948 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
949 	 * that no events are triggered between unregister_netdev() and the
950 	 * driver unloading. A more general condition is that NETDEV_CHANGE
951 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
952 	if (!netif_running(efx->net_dev))
953 		return;
954 
955 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
956 		efx->n_link_state_changes++;
957 
958 		if (link_state->up)
959 			netif_carrier_on(efx->net_dev);
960 		else
961 			netif_carrier_off(efx->net_dev);
962 	}
963 
964 	/* Status message for kernel log */
965 	if (link_state->up)
966 		netif_info(efx, link, efx->net_dev,
967 			   "link up at %uMbps %s-duplex (MTU %d)\n",
968 			   link_state->speed, link_state->fd ? "full" : "half",
969 			   efx->net_dev->mtu);
970 	else
971 		netif_info(efx, link, efx->net_dev, "link down\n");
972 }
973 
974 void efx_link_set_advertising(struct efx_nic *efx,
975 			      const unsigned long *advertising)
976 {
977 	memcpy(efx->link_advertising, advertising,
978 	       sizeof(__ETHTOOL_DECLARE_LINK_MODE_MASK()));
979 
980 	efx->link_advertising[0] |= ADVERTISED_Autoneg;
981 	if (advertising[0] & ADVERTISED_Pause)
982 		efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
983 	else
984 		efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
985 	if (advertising[0] & ADVERTISED_Asym_Pause)
986 		efx->wanted_fc ^= EFX_FC_TX;
987 }
988 
989 /* Equivalent to efx_link_set_advertising with all-zeroes, except does not
990  * force the Autoneg bit on.
991  */
992 void efx_link_clear_advertising(struct efx_nic *efx)
993 {
994 	bitmap_zero(efx->link_advertising, __ETHTOOL_LINK_MODE_MASK_NBITS);
995 	efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
996 }
997 
998 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
999 {
1000 	efx->wanted_fc = wanted_fc;
1001 	if (efx->link_advertising[0]) {
1002 		if (wanted_fc & EFX_FC_RX)
1003 			efx->link_advertising[0] |= (ADVERTISED_Pause |
1004 						     ADVERTISED_Asym_Pause);
1005 		else
1006 			efx->link_advertising[0] &= ~(ADVERTISED_Pause |
1007 						      ADVERTISED_Asym_Pause);
1008 		if (wanted_fc & EFX_FC_TX)
1009 			efx->link_advertising[0] ^= ADVERTISED_Asym_Pause;
1010 	}
1011 }
1012 
1013 static void efx_fini_port(struct efx_nic *efx);
1014 
1015 /* We assume that efx->type->reconfigure_mac will always try to sync RX
1016  * filters and therefore needs to read-lock the filter table against freeing
1017  */
1018 void efx_mac_reconfigure(struct efx_nic *efx)
1019 {
1020 	down_read(&efx->filter_sem);
1021 	efx->type->reconfigure_mac(efx);
1022 	up_read(&efx->filter_sem);
1023 }
1024 
1025 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
1026  * the MAC appropriately. All other PHY configuration changes are pushed
1027  * through phy_op->set_settings(), and pushed asynchronously to the MAC
1028  * through efx_monitor().
1029  *
1030  * Callers must hold the mac_lock
1031  */
1032 int __efx_reconfigure_port(struct efx_nic *efx)
1033 {
1034 	enum efx_phy_mode phy_mode;
1035 	int rc;
1036 
1037 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
1038 
1039 	/* Disable PHY transmit in mac level loopbacks */
1040 	phy_mode = efx->phy_mode;
1041 	if (LOOPBACK_INTERNAL(efx))
1042 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
1043 	else
1044 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
1045 
1046 	rc = efx->type->reconfigure_port(efx);
1047 
1048 	if (rc)
1049 		efx->phy_mode = phy_mode;
1050 
1051 	return rc;
1052 }
1053 
1054 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1055  * disabled. */
1056 int efx_reconfigure_port(struct efx_nic *efx)
1057 {
1058 	int rc;
1059 
1060 	EFX_ASSERT_RESET_SERIALISED(efx);
1061 
1062 	mutex_lock(&efx->mac_lock);
1063 	rc = __efx_reconfigure_port(efx);
1064 	mutex_unlock(&efx->mac_lock);
1065 
1066 	return rc;
1067 }
1068 
1069 /* Asynchronous work item for changing MAC promiscuity and multicast
1070  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1071  * MAC directly. */
1072 static void efx_mac_work(struct work_struct *data)
1073 {
1074 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1075 
1076 	mutex_lock(&efx->mac_lock);
1077 	if (efx->port_enabled)
1078 		efx_mac_reconfigure(efx);
1079 	mutex_unlock(&efx->mac_lock);
1080 }
1081 
1082 static int efx_probe_port(struct efx_nic *efx)
1083 {
1084 	int rc;
1085 
1086 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1087 
1088 	if (phy_flash_cfg)
1089 		efx->phy_mode = PHY_MODE_SPECIAL;
1090 
1091 	/* Connect up MAC/PHY operations table */
1092 	rc = efx->type->probe_port(efx);
1093 	if (rc)
1094 		return rc;
1095 
1096 	/* Initialise MAC address to permanent address */
1097 	ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1098 
1099 	return 0;
1100 }
1101 
1102 static int efx_init_port(struct efx_nic *efx)
1103 {
1104 	int rc;
1105 
1106 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1107 
1108 	mutex_lock(&efx->mac_lock);
1109 
1110 	rc = efx->phy_op->init(efx);
1111 	if (rc)
1112 		goto fail1;
1113 
1114 	efx->port_initialized = true;
1115 
1116 	/* Reconfigure the MAC before creating dma queues (required for
1117 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1118 	efx_mac_reconfigure(efx);
1119 
1120 	/* Ensure the PHY advertises the correct flow control settings */
1121 	rc = efx->phy_op->reconfigure(efx);
1122 	if (rc && rc != -EPERM)
1123 		goto fail2;
1124 
1125 	mutex_unlock(&efx->mac_lock);
1126 	return 0;
1127 
1128 fail2:
1129 	efx->phy_op->fini(efx);
1130 fail1:
1131 	mutex_unlock(&efx->mac_lock);
1132 	return rc;
1133 }
1134 
1135 static void efx_start_port(struct efx_nic *efx)
1136 {
1137 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1138 	BUG_ON(efx->port_enabled);
1139 
1140 	mutex_lock(&efx->mac_lock);
1141 	efx->port_enabled = true;
1142 
1143 	/* Ensure MAC ingress/egress is enabled */
1144 	efx_mac_reconfigure(efx);
1145 
1146 	mutex_unlock(&efx->mac_lock);
1147 }
1148 
1149 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1150  * and the async self-test, wait for them to finish and prevent them
1151  * being scheduled again.  This doesn't cover online resets, which
1152  * should only be cancelled when removing the device.
1153  */
1154 static void efx_stop_port(struct efx_nic *efx)
1155 {
1156 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1157 
1158 	EFX_ASSERT_RESET_SERIALISED(efx);
1159 
1160 	mutex_lock(&efx->mac_lock);
1161 	efx->port_enabled = false;
1162 	mutex_unlock(&efx->mac_lock);
1163 
1164 	/* Serialise against efx_set_multicast_list() */
1165 	netif_addr_lock_bh(efx->net_dev);
1166 	netif_addr_unlock_bh(efx->net_dev);
1167 
1168 	cancel_delayed_work_sync(&efx->monitor_work);
1169 	efx_selftest_async_cancel(efx);
1170 	cancel_work_sync(&efx->mac_work);
1171 }
1172 
1173 static void efx_fini_port(struct efx_nic *efx)
1174 {
1175 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1176 
1177 	if (!efx->port_initialized)
1178 		return;
1179 
1180 	efx->phy_op->fini(efx);
1181 	efx->port_initialized = false;
1182 
1183 	efx->link_state.up = false;
1184 	efx_link_status_changed(efx);
1185 }
1186 
1187 static void efx_remove_port(struct efx_nic *efx)
1188 {
1189 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1190 
1191 	efx->type->remove_port(efx);
1192 }
1193 
1194 /**************************************************************************
1195  *
1196  * NIC handling
1197  *
1198  **************************************************************************/
1199 
1200 static LIST_HEAD(efx_primary_list);
1201 static LIST_HEAD(efx_unassociated_list);
1202 
1203 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1204 {
1205 	return left->type == right->type &&
1206 		left->vpd_sn && right->vpd_sn &&
1207 		!strcmp(left->vpd_sn, right->vpd_sn);
1208 }
1209 
1210 static void efx_associate(struct efx_nic *efx)
1211 {
1212 	struct efx_nic *other, *next;
1213 
1214 	if (efx->primary == efx) {
1215 		/* Adding primary function; look for secondaries */
1216 
1217 		netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1218 		list_add_tail(&efx->node, &efx_primary_list);
1219 
1220 		list_for_each_entry_safe(other, next, &efx_unassociated_list,
1221 					 node) {
1222 			if (efx_same_controller(efx, other)) {
1223 				list_del(&other->node);
1224 				netif_dbg(other, probe, other->net_dev,
1225 					  "moving to secondary list of %s %s\n",
1226 					  pci_name(efx->pci_dev),
1227 					  efx->net_dev->name);
1228 				list_add_tail(&other->node,
1229 					      &efx->secondary_list);
1230 				other->primary = efx;
1231 			}
1232 		}
1233 	} else {
1234 		/* Adding secondary function; look for primary */
1235 
1236 		list_for_each_entry(other, &efx_primary_list, node) {
1237 			if (efx_same_controller(efx, other)) {
1238 				netif_dbg(efx, probe, efx->net_dev,
1239 					  "adding to secondary list of %s %s\n",
1240 					  pci_name(other->pci_dev),
1241 					  other->net_dev->name);
1242 				list_add_tail(&efx->node,
1243 					      &other->secondary_list);
1244 				efx->primary = other;
1245 				return;
1246 			}
1247 		}
1248 
1249 		netif_dbg(efx, probe, efx->net_dev,
1250 			  "adding to unassociated list\n");
1251 		list_add_tail(&efx->node, &efx_unassociated_list);
1252 	}
1253 }
1254 
1255 static void efx_dissociate(struct efx_nic *efx)
1256 {
1257 	struct efx_nic *other, *next;
1258 
1259 	list_del(&efx->node);
1260 	efx->primary = NULL;
1261 
1262 	list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1263 		list_del(&other->node);
1264 		netif_dbg(other, probe, other->net_dev,
1265 			  "moving to unassociated list\n");
1266 		list_add_tail(&other->node, &efx_unassociated_list);
1267 		other->primary = NULL;
1268 	}
1269 }
1270 
1271 /* This configures the PCI device to enable I/O and DMA. */
1272 static int efx_init_io(struct efx_nic *efx)
1273 {
1274 	struct pci_dev *pci_dev = efx->pci_dev;
1275 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1276 	unsigned int mem_map_size = efx->type->mem_map_size(efx);
1277 	int rc, bar;
1278 
1279 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1280 
1281 	bar = efx->type->mem_bar(efx);
1282 
1283 	rc = pci_enable_device(pci_dev);
1284 	if (rc) {
1285 		netif_err(efx, probe, efx->net_dev,
1286 			  "failed to enable PCI device\n");
1287 		goto fail1;
1288 	}
1289 
1290 	pci_set_master(pci_dev);
1291 
1292 	/* Set the PCI DMA mask.  Try all possibilities from our genuine mask
1293 	 * down to 32 bits, because some architectures will allow 40 bit
1294 	 * masks event though they reject 46 bit masks.
1295 	 */
1296 	while (dma_mask > 0x7fffffffUL) {
1297 		rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1298 		if (rc == 0)
1299 			break;
1300 		dma_mask >>= 1;
1301 	}
1302 	if (rc) {
1303 		netif_err(efx, probe, efx->net_dev,
1304 			  "could not find a suitable DMA mask\n");
1305 		goto fail2;
1306 	}
1307 	netif_dbg(efx, probe, efx->net_dev,
1308 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1309 
1310 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1311 	rc = pci_request_region(pci_dev, bar, "sfc");
1312 	if (rc) {
1313 		netif_err(efx, probe, efx->net_dev,
1314 			  "request for memory BAR failed\n");
1315 		rc = -EIO;
1316 		goto fail3;
1317 	}
1318 	efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1319 	if (!efx->membase) {
1320 		netif_err(efx, probe, efx->net_dev,
1321 			  "could not map memory BAR at %llx+%x\n",
1322 			  (unsigned long long)efx->membase_phys, mem_map_size);
1323 		rc = -ENOMEM;
1324 		goto fail4;
1325 	}
1326 	netif_dbg(efx, probe, efx->net_dev,
1327 		  "memory BAR at %llx+%x (virtual %p)\n",
1328 		  (unsigned long long)efx->membase_phys, mem_map_size,
1329 		  efx->membase);
1330 
1331 	return 0;
1332 
1333  fail4:
1334 	pci_release_region(efx->pci_dev, bar);
1335  fail3:
1336 	efx->membase_phys = 0;
1337  fail2:
1338 	pci_disable_device(efx->pci_dev);
1339  fail1:
1340 	return rc;
1341 }
1342 
1343 static void efx_fini_io(struct efx_nic *efx)
1344 {
1345 	int bar;
1346 
1347 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1348 
1349 	if (efx->membase) {
1350 		iounmap(efx->membase);
1351 		efx->membase = NULL;
1352 	}
1353 
1354 	if (efx->membase_phys) {
1355 		bar = efx->type->mem_bar(efx);
1356 		pci_release_region(efx->pci_dev, bar);
1357 		efx->membase_phys = 0;
1358 	}
1359 
1360 	/* Don't disable bus-mastering if VFs are assigned */
1361 	if (!pci_vfs_assigned(efx->pci_dev))
1362 		pci_disable_device(efx->pci_dev);
1363 }
1364 
1365 void efx_set_default_rx_indir_table(struct efx_nic *efx,
1366 				    struct efx_rss_context *ctx)
1367 {
1368 	size_t i;
1369 
1370 	for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++)
1371 		ctx->rx_indir_table[i] =
1372 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1373 }
1374 
1375 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1376 {
1377 	cpumask_var_t thread_mask;
1378 	unsigned int count;
1379 	int cpu;
1380 
1381 	if (rss_cpus) {
1382 		count = rss_cpus;
1383 	} else {
1384 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1385 			netif_warn(efx, probe, efx->net_dev,
1386 				   "RSS disabled due to allocation failure\n");
1387 			return 1;
1388 		}
1389 
1390 		count = 0;
1391 		for_each_online_cpu(cpu) {
1392 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1393 				++count;
1394 				cpumask_or(thread_mask, thread_mask,
1395 					   topology_sibling_cpumask(cpu));
1396 			}
1397 		}
1398 
1399 		free_cpumask_var(thread_mask);
1400 	}
1401 
1402 	if (count > EFX_MAX_RX_QUEUES) {
1403 		netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn,
1404 			       "Reducing number of rx queues from %u to %u.\n",
1405 			       count, EFX_MAX_RX_QUEUES);
1406 		count = EFX_MAX_RX_QUEUES;
1407 	}
1408 
1409 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1410 	 * table entries that are inaccessible to VFs
1411 	 */
1412 #ifdef CONFIG_SFC_SRIOV
1413 	if (efx->type->sriov_wanted) {
1414 		if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1415 		    count > efx_vf_size(efx)) {
1416 			netif_warn(efx, probe, efx->net_dev,
1417 				   "Reducing number of RSS channels from %u to %u for "
1418 				   "VF support. Increase vf-msix-limit to use more "
1419 				   "channels on the PF.\n",
1420 				   count, efx_vf_size(efx));
1421 			count = efx_vf_size(efx);
1422 		}
1423 	}
1424 #endif
1425 
1426 	return count;
1427 }
1428 
1429 /* Probe the number and type of interrupts we are able to obtain, and
1430  * the resulting numbers of channels and RX queues.
1431  */
1432 static int efx_probe_interrupts(struct efx_nic *efx)
1433 {
1434 	unsigned int extra_channels = 0;
1435 	unsigned int i, j;
1436 	int rc;
1437 
1438 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1439 		if (efx->extra_channel_type[i])
1440 			++extra_channels;
1441 
1442 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1443 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1444 		unsigned int n_channels;
1445 
1446 		n_channels = efx_wanted_parallelism(efx);
1447 		if (efx_separate_tx_channels)
1448 			n_channels *= 2;
1449 		n_channels += extra_channels;
1450 		n_channels = min(n_channels, efx->max_channels);
1451 
1452 		for (i = 0; i < n_channels; i++)
1453 			xentries[i].entry = i;
1454 		rc = pci_enable_msix_range(efx->pci_dev,
1455 					   xentries, 1, n_channels);
1456 		if (rc < 0) {
1457 			/* Fall back to single channel MSI */
1458 			netif_err(efx, drv, efx->net_dev,
1459 				  "could not enable MSI-X\n");
1460 			if (efx->type->min_interrupt_mode >= EFX_INT_MODE_MSI)
1461 				efx->interrupt_mode = EFX_INT_MODE_MSI;
1462 			else
1463 				return rc;
1464 		} else if (rc < n_channels) {
1465 			netif_err(efx, drv, efx->net_dev,
1466 				  "WARNING: Insufficient MSI-X vectors"
1467 				  " available (%d < %u).\n", rc, n_channels);
1468 			netif_err(efx, drv, efx->net_dev,
1469 				  "WARNING: Performance may be reduced.\n");
1470 			n_channels = rc;
1471 		}
1472 
1473 		if (rc > 0) {
1474 			efx->n_channels = n_channels;
1475 			if (n_channels > extra_channels)
1476 				n_channels -= extra_channels;
1477 			if (efx_separate_tx_channels) {
1478 				efx->n_tx_channels = min(max(n_channels / 2,
1479 							     1U),
1480 							 efx->max_tx_channels);
1481 				efx->n_rx_channels = max(n_channels -
1482 							 efx->n_tx_channels,
1483 							 1U);
1484 			} else {
1485 				efx->n_tx_channels = min(n_channels,
1486 							 efx->max_tx_channels);
1487 				efx->n_rx_channels = n_channels;
1488 			}
1489 			for (i = 0; i < efx->n_channels; i++)
1490 				efx_get_channel(efx, i)->irq =
1491 					xentries[i].vector;
1492 		}
1493 	}
1494 
1495 	/* Try single interrupt MSI */
1496 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1497 		efx->n_channels = 1;
1498 		efx->n_rx_channels = 1;
1499 		efx->n_tx_channels = 1;
1500 		rc = pci_enable_msi(efx->pci_dev);
1501 		if (rc == 0) {
1502 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1503 		} else {
1504 			netif_err(efx, drv, efx->net_dev,
1505 				  "could not enable MSI\n");
1506 			if (efx->type->min_interrupt_mode >= EFX_INT_MODE_LEGACY)
1507 				efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1508 			else
1509 				return rc;
1510 		}
1511 	}
1512 
1513 	/* Assume legacy interrupts */
1514 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1515 		efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
1516 		efx->n_rx_channels = 1;
1517 		efx->n_tx_channels = 1;
1518 		efx->legacy_irq = efx->pci_dev->irq;
1519 	}
1520 
1521 	/* Assign extra channels if possible */
1522 	efx->n_extra_tx_channels = 0;
1523 	j = efx->n_channels;
1524 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1525 		if (!efx->extra_channel_type[i])
1526 			continue;
1527 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1528 		    efx->n_channels <= extra_channels) {
1529 			efx->extra_channel_type[i]->handle_no_channel(efx);
1530 		} else {
1531 			--j;
1532 			efx_get_channel(efx, j)->type =
1533 				efx->extra_channel_type[i];
1534 			if (efx_channel_has_tx_queues(efx_get_channel(efx, j)))
1535 				efx->n_extra_tx_channels++;
1536 		}
1537 	}
1538 
1539 	/* RSS might be usable on VFs even if it is disabled on the PF */
1540 #ifdef CONFIG_SFC_SRIOV
1541 	if (efx->type->sriov_wanted) {
1542 		efx->rss_spread = ((efx->n_rx_channels > 1 ||
1543 				    !efx->type->sriov_wanted(efx)) ?
1544 				   efx->n_rx_channels : efx_vf_size(efx));
1545 		return 0;
1546 	}
1547 #endif
1548 	efx->rss_spread = efx->n_rx_channels;
1549 
1550 	return 0;
1551 }
1552 
1553 #if defined(CONFIG_SMP)
1554 static void efx_set_interrupt_affinity(struct efx_nic *efx)
1555 {
1556 	struct efx_channel *channel;
1557 	unsigned int cpu;
1558 
1559 	efx_for_each_channel(channel, efx) {
1560 		cpu = cpumask_local_spread(channel->channel,
1561 					   pcibus_to_node(efx->pci_dev->bus));
1562 		irq_set_affinity_hint(channel->irq, cpumask_of(cpu));
1563 	}
1564 }
1565 
1566 static void efx_clear_interrupt_affinity(struct efx_nic *efx)
1567 {
1568 	struct efx_channel *channel;
1569 
1570 	efx_for_each_channel(channel, efx)
1571 		irq_set_affinity_hint(channel->irq, NULL);
1572 }
1573 #else
1574 static void
1575 efx_set_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
1576 {
1577 }
1578 
1579 static void
1580 efx_clear_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
1581 {
1582 }
1583 #endif /* CONFIG_SMP */
1584 
1585 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1586 {
1587 	struct efx_channel *channel, *end_channel;
1588 	int rc;
1589 
1590 	BUG_ON(efx->state == STATE_DISABLED);
1591 
1592 	efx->irq_soft_enabled = true;
1593 	smp_wmb();
1594 
1595 	efx_for_each_channel(channel, efx) {
1596 		if (!channel->type->keep_eventq) {
1597 			rc = efx_init_eventq(channel);
1598 			if (rc)
1599 				goto fail;
1600 		}
1601 		efx_start_eventq(channel);
1602 	}
1603 
1604 	efx_mcdi_mode_event(efx);
1605 
1606 	return 0;
1607 fail:
1608 	end_channel = channel;
1609 	efx_for_each_channel(channel, efx) {
1610 		if (channel == end_channel)
1611 			break;
1612 		efx_stop_eventq(channel);
1613 		if (!channel->type->keep_eventq)
1614 			efx_fini_eventq(channel);
1615 	}
1616 
1617 	return rc;
1618 }
1619 
1620 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1621 {
1622 	struct efx_channel *channel;
1623 
1624 	if (efx->state == STATE_DISABLED)
1625 		return;
1626 
1627 	efx_mcdi_mode_poll(efx);
1628 
1629 	efx->irq_soft_enabled = false;
1630 	smp_wmb();
1631 
1632 	if (efx->legacy_irq)
1633 		synchronize_irq(efx->legacy_irq);
1634 
1635 	efx_for_each_channel(channel, efx) {
1636 		if (channel->irq)
1637 			synchronize_irq(channel->irq);
1638 
1639 		efx_stop_eventq(channel);
1640 		if (!channel->type->keep_eventq)
1641 			efx_fini_eventq(channel);
1642 	}
1643 
1644 	/* Flush the asynchronous MCDI request queue */
1645 	efx_mcdi_flush_async(efx);
1646 }
1647 
1648 static int efx_enable_interrupts(struct efx_nic *efx)
1649 {
1650 	struct efx_channel *channel, *end_channel;
1651 	int rc;
1652 
1653 	BUG_ON(efx->state == STATE_DISABLED);
1654 
1655 	if (efx->eeh_disabled_legacy_irq) {
1656 		enable_irq(efx->legacy_irq);
1657 		efx->eeh_disabled_legacy_irq = false;
1658 	}
1659 
1660 	efx->type->irq_enable_master(efx);
1661 
1662 	efx_for_each_channel(channel, efx) {
1663 		if (channel->type->keep_eventq) {
1664 			rc = efx_init_eventq(channel);
1665 			if (rc)
1666 				goto fail;
1667 		}
1668 	}
1669 
1670 	rc = efx_soft_enable_interrupts(efx);
1671 	if (rc)
1672 		goto fail;
1673 
1674 	return 0;
1675 
1676 fail:
1677 	end_channel = channel;
1678 	efx_for_each_channel(channel, efx) {
1679 		if (channel == end_channel)
1680 			break;
1681 		if (channel->type->keep_eventq)
1682 			efx_fini_eventq(channel);
1683 	}
1684 
1685 	efx->type->irq_disable_non_ev(efx);
1686 
1687 	return rc;
1688 }
1689 
1690 static void efx_disable_interrupts(struct efx_nic *efx)
1691 {
1692 	struct efx_channel *channel;
1693 
1694 	efx_soft_disable_interrupts(efx);
1695 
1696 	efx_for_each_channel(channel, efx) {
1697 		if (channel->type->keep_eventq)
1698 			efx_fini_eventq(channel);
1699 	}
1700 
1701 	efx->type->irq_disable_non_ev(efx);
1702 }
1703 
1704 static void efx_remove_interrupts(struct efx_nic *efx)
1705 {
1706 	struct efx_channel *channel;
1707 
1708 	/* Remove MSI/MSI-X interrupts */
1709 	efx_for_each_channel(channel, efx)
1710 		channel->irq = 0;
1711 	pci_disable_msi(efx->pci_dev);
1712 	pci_disable_msix(efx->pci_dev);
1713 
1714 	/* Remove legacy interrupt */
1715 	efx->legacy_irq = 0;
1716 }
1717 
1718 static void efx_set_channels(struct efx_nic *efx)
1719 {
1720 	struct efx_channel *channel;
1721 	struct efx_tx_queue *tx_queue;
1722 
1723 	efx->tx_channel_offset =
1724 		efx_separate_tx_channels ?
1725 		efx->n_channels - efx->n_tx_channels : 0;
1726 
1727 	/* We need to mark which channels really have RX and TX
1728 	 * queues, and adjust the TX queue numbers if we have separate
1729 	 * RX-only and TX-only channels.
1730 	 */
1731 	efx_for_each_channel(channel, efx) {
1732 		if (channel->channel < efx->n_rx_channels)
1733 			channel->rx_queue.core_index = channel->channel;
1734 		else
1735 			channel->rx_queue.core_index = -1;
1736 
1737 		efx_for_each_channel_tx_queue(tx_queue, channel)
1738 			tx_queue->queue -= (efx->tx_channel_offset *
1739 					    EFX_TXQ_TYPES);
1740 	}
1741 }
1742 
1743 static int efx_probe_nic(struct efx_nic *efx)
1744 {
1745 	int rc;
1746 
1747 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1748 
1749 	/* Carry out hardware-type specific initialisation */
1750 	rc = efx->type->probe(efx);
1751 	if (rc)
1752 		return rc;
1753 
1754 	do {
1755 		if (!efx->max_channels || !efx->max_tx_channels) {
1756 			netif_err(efx, drv, efx->net_dev,
1757 				  "Insufficient resources to allocate"
1758 				  " any channels\n");
1759 			rc = -ENOSPC;
1760 			goto fail1;
1761 		}
1762 
1763 		/* Determine the number of channels and queues by trying
1764 		 * to hook in MSI-X interrupts.
1765 		 */
1766 		rc = efx_probe_interrupts(efx);
1767 		if (rc)
1768 			goto fail1;
1769 
1770 		efx_set_channels(efx);
1771 
1772 		/* dimension_resources can fail with EAGAIN */
1773 		rc = efx->type->dimension_resources(efx);
1774 		if (rc != 0 && rc != -EAGAIN)
1775 			goto fail2;
1776 
1777 		if (rc == -EAGAIN)
1778 			/* try again with new max_channels */
1779 			efx_remove_interrupts(efx);
1780 
1781 	} while (rc == -EAGAIN);
1782 
1783 	if (efx->n_channels > 1)
1784 		netdev_rss_key_fill(efx->rss_context.rx_hash_key,
1785 				    sizeof(efx->rss_context.rx_hash_key));
1786 	efx_set_default_rx_indir_table(efx, &efx->rss_context);
1787 
1788 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1789 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1790 
1791 	/* Initialise the interrupt moderation settings */
1792 	efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
1793 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1794 				true);
1795 
1796 	return 0;
1797 
1798 fail2:
1799 	efx_remove_interrupts(efx);
1800 fail1:
1801 	efx->type->remove(efx);
1802 	return rc;
1803 }
1804 
1805 static void efx_remove_nic(struct efx_nic *efx)
1806 {
1807 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1808 
1809 	efx_remove_interrupts(efx);
1810 	efx->type->remove(efx);
1811 }
1812 
1813 static int efx_probe_filters(struct efx_nic *efx)
1814 {
1815 	int rc;
1816 
1817 	init_rwsem(&efx->filter_sem);
1818 	mutex_lock(&efx->mac_lock);
1819 	down_write(&efx->filter_sem);
1820 	rc = efx->type->filter_table_probe(efx);
1821 	if (rc)
1822 		goto out_unlock;
1823 
1824 #ifdef CONFIG_RFS_ACCEL
1825 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
1826 		struct efx_channel *channel;
1827 		int i, success = 1;
1828 
1829 		efx_for_each_channel(channel, efx) {
1830 			channel->rps_flow_id =
1831 				kcalloc(efx->type->max_rx_ip_filters,
1832 					sizeof(*channel->rps_flow_id),
1833 					GFP_KERNEL);
1834 			if (!channel->rps_flow_id)
1835 				success = 0;
1836 			else
1837 				for (i = 0;
1838 				     i < efx->type->max_rx_ip_filters;
1839 				     ++i)
1840 					channel->rps_flow_id[i] =
1841 						RPS_FLOW_ID_INVALID;
1842 		}
1843 
1844 		if (!success) {
1845 			efx_for_each_channel(channel, efx)
1846 				kfree(channel->rps_flow_id);
1847 			efx->type->filter_table_remove(efx);
1848 			rc = -ENOMEM;
1849 			goto out_unlock;
1850 		}
1851 
1852 		efx->rps_expire_index = efx->rps_expire_channel = 0;
1853 	}
1854 #endif
1855 out_unlock:
1856 	up_write(&efx->filter_sem);
1857 	mutex_unlock(&efx->mac_lock);
1858 	return rc;
1859 }
1860 
1861 static void efx_remove_filters(struct efx_nic *efx)
1862 {
1863 #ifdef CONFIG_RFS_ACCEL
1864 	struct efx_channel *channel;
1865 
1866 	efx_for_each_channel(channel, efx)
1867 		kfree(channel->rps_flow_id);
1868 #endif
1869 	down_write(&efx->filter_sem);
1870 	efx->type->filter_table_remove(efx);
1871 	up_write(&efx->filter_sem);
1872 }
1873 
1874 
1875 /**************************************************************************
1876  *
1877  * NIC startup/shutdown
1878  *
1879  *************************************************************************/
1880 
1881 static int efx_probe_all(struct efx_nic *efx)
1882 {
1883 	int rc;
1884 
1885 	rc = efx_probe_nic(efx);
1886 	if (rc) {
1887 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1888 		goto fail1;
1889 	}
1890 
1891 	rc = efx_probe_port(efx);
1892 	if (rc) {
1893 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1894 		goto fail2;
1895 	}
1896 
1897 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1898 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1899 		rc = -EINVAL;
1900 		goto fail3;
1901 	}
1902 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1903 
1904 #ifdef CONFIG_SFC_SRIOV
1905 	rc = efx->type->vswitching_probe(efx);
1906 	if (rc) /* not fatal; the PF will still work fine */
1907 		netif_warn(efx, probe, efx->net_dev,
1908 			   "failed to setup vswitching rc=%d;"
1909 			   " VFs may not function\n", rc);
1910 #endif
1911 
1912 	rc = efx_probe_filters(efx);
1913 	if (rc) {
1914 		netif_err(efx, probe, efx->net_dev,
1915 			  "failed to create filter tables\n");
1916 		goto fail4;
1917 	}
1918 
1919 	rc = efx_probe_channels(efx);
1920 	if (rc)
1921 		goto fail5;
1922 
1923 	return 0;
1924 
1925  fail5:
1926 	efx_remove_filters(efx);
1927  fail4:
1928 #ifdef CONFIG_SFC_SRIOV
1929 	efx->type->vswitching_remove(efx);
1930 #endif
1931  fail3:
1932 	efx_remove_port(efx);
1933  fail2:
1934 	efx_remove_nic(efx);
1935  fail1:
1936 	return rc;
1937 }
1938 
1939 /* If the interface is supposed to be running but is not, start
1940  * the hardware and software data path, regular activity for the port
1941  * (MAC statistics, link polling, etc.) and schedule the port to be
1942  * reconfigured.  Interrupts must already be enabled.  This function
1943  * is safe to call multiple times, so long as the NIC is not disabled.
1944  * Requires the RTNL lock.
1945  */
1946 static void efx_start_all(struct efx_nic *efx)
1947 {
1948 	EFX_ASSERT_RESET_SERIALISED(efx);
1949 	BUG_ON(efx->state == STATE_DISABLED);
1950 
1951 	/* Check that it is appropriate to restart the interface. All
1952 	 * of these flags are safe to read under just the rtnl lock */
1953 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
1954 	    efx->reset_pending)
1955 		return;
1956 
1957 	efx_start_port(efx);
1958 	efx_start_datapath(efx);
1959 
1960 	/* Start the hardware monitor if there is one */
1961 	if (efx->type->monitor != NULL)
1962 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1963 				   efx_monitor_interval);
1964 
1965 	/* Link state detection is normally event-driven; we have
1966 	 * to poll now because we could have missed a change
1967 	 */
1968 	mutex_lock(&efx->mac_lock);
1969 	if (efx->phy_op->poll(efx))
1970 		efx_link_status_changed(efx);
1971 	mutex_unlock(&efx->mac_lock);
1972 
1973 	efx->type->start_stats(efx);
1974 	efx->type->pull_stats(efx);
1975 	spin_lock_bh(&efx->stats_lock);
1976 	efx->type->update_stats(efx, NULL, NULL);
1977 	spin_unlock_bh(&efx->stats_lock);
1978 }
1979 
1980 /* Quiesce the hardware and software data path, and regular activity
1981  * for the port without bringing the link down.  Safe to call multiple
1982  * times with the NIC in almost any state, but interrupts should be
1983  * enabled.  Requires the RTNL lock.
1984  */
1985 static void efx_stop_all(struct efx_nic *efx)
1986 {
1987 	EFX_ASSERT_RESET_SERIALISED(efx);
1988 
1989 	/* port_enabled can be read safely under the rtnl lock */
1990 	if (!efx->port_enabled)
1991 		return;
1992 
1993 	/* update stats before we go down so we can accurately count
1994 	 * rx_nodesc_drops
1995 	 */
1996 	efx->type->pull_stats(efx);
1997 	spin_lock_bh(&efx->stats_lock);
1998 	efx->type->update_stats(efx, NULL, NULL);
1999 	spin_unlock_bh(&efx->stats_lock);
2000 	efx->type->stop_stats(efx);
2001 	efx_stop_port(efx);
2002 
2003 	/* Stop the kernel transmit interface.  This is only valid if
2004 	 * the device is stopped or detached; otherwise the watchdog
2005 	 * may fire immediately.
2006 	 */
2007 	WARN_ON(netif_running(efx->net_dev) &&
2008 		netif_device_present(efx->net_dev));
2009 	netif_tx_disable(efx->net_dev);
2010 
2011 	efx_stop_datapath(efx);
2012 }
2013 
2014 static void efx_remove_all(struct efx_nic *efx)
2015 {
2016 	efx_remove_channels(efx);
2017 	efx_remove_filters(efx);
2018 #ifdef CONFIG_SFC_SRIOV
2019 	efx->type->vswitching_remove(efx);
2020 #endif
2021 	efx_remove_port(efx);
2022 	efx_remove_nic(efx);
2023 }
2024 
2025 /**************************************************************************
2026  *
2027  * Interrupt moderation
2028  *
2029  **************************************************************************/
2030 unsigned int efx_usecs_to_ticks(struct efx_nic *efx, unsigned int usecs)
2031 {
2032 	if (usecs == 0)
2033 		return 0;
2034 	if (usecs * 1000 < efx->timer_quantum_ns)
2035 		return 1; /* never round down to 0 */
2036 	return usecs * 1000 / efx->timer_quantum_ns;
2037 }
2038 
2039 unsigned int efx_ticks_to_usecs(struct efx_nic *efx, unsigned int ticks)
2040 {
2041 	/* We must round up when converting ticks to microseconds
2042 	 * because we round down when converting the other way.
2043 	 */
2044 	return DIV_ROUND_UP(ticks * efx->timer_quantum_ns, 1000);
2045 }
2046 
2047 /* Set interrupt moderation parameters */
2048 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
2049 			    unsigned int rx_usecs, bool rx_adaptive,
2050 			    bool rx_may_override_tx)
2051 {
2052 	struct efx_channel *channel;
2053 	unsigned int timer_max_us;
2054 
2055 	EFX_ASSERT_RESET_SERIALISED(efx);
2056 
2057 	timer_max_us = efx->timer_max_ns / 1000;
2058 
2059 	if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
2060 		return -EINVAL;
2061 
2062 	if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
2063 	    !rx_may_override_tx) {
2064 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
2065 			  "RX and TX IRQ moderation must be equal\n");
2066 		return -EINVAL;
2067 	}
2068 
2069 	efx->irq_rx_adaptive = rx_adaptive;
2070 	efx->irq_rx_moderation_us = rx_usecs;
2071 	efx_for_each_channel(channel, efx) {
2072 		if (efx_channel_has_rx_queue(channel))
2073 			channel->irq_moderation_us = rx_usecs;
2074 		else if (efx_channel_has_tx_queues(channel))
2075 			channel->irq_moderation_us = tx_usecs;
2076 	}
2077 
2078 	return 0;
2079 }
2080 
2081 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
2082 			    unsigned int *rx_usecs, bool *rx_adaptive)
2083 {
2084 	*rx_adaptive = efx->irq_rx_adaptive;
2085 	*rx_usecs = efx->irq_rx_moderation_us;
2086 
2087 	/* If channels are shared between RX and TX, so is IRQ
2088 	 * moderation.  Otherwise, IRQ moderation is the same for all
2089 	 * TX channels and is not adaptive.
2090 	 */
2091 	if (efx->tx_channel_offset == 0) {
2092 		*tx_usecs = *rx_usecs;
2093 	} else {
2094 		struct efx_channel *tx_channel;
2095 
2096 		tx_channel = efx->channel[efx->tx_channel_offset];
2097 		*tx_usecs = tx_channel->irq_moderation_us;
2098 	}
2099 }
2100 
2101 /**************************************************************************
2102  *
2103  * Hardware monitor
2104  *
2105  **************************************************************************/
2106 
2107 /* Run periodically off the general workqueue */
2108 static void efx_monitor(struct work_struct *data)
2109 {
2110 	struct efx_nic *efx = container_of(data, struct efx_nic,
2111 					   monitor_work.work);
2112 
2113 	netif_vdbg(efx, timer, efx->net_dev,
2114 		   "hardware monitor executing on CPU %d\n",
2115 		   raw_smp_processor_id());
2116 	BUG_ON(efx->type->monitor == NULL);
2117 
2118 	/* If the mac_lock is already held then it is likely a port
2119 	 * reconfiguration is already in place, which will likely do
2120 	 * most of the work of monitor() anyway. */
2121 	if (mutex_trylock(&efx->mac_lock)) {
2122 		if (efx->port_enabled)
2123 			efx->type->monitor(efx);
2124 		mutex_unlock(&efx->mac_lock);
2125 	}
2126 
2127 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
2128 			   efx_monitor_interval);
2129 }
2130 
2131 /**************************************************************************
2132  *
2133  * ioctls
2134  *
2135  *************************************************************************/
2136 
2137 /* Net device ioctl
2138  * Context: process, rtnl_lock() held.
2139  */
2140 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
2141 {
2142 	struct efx_nic *efx = netdev_priv(net_dev);
2143 	struct mii_ioctl_data *data = if_mii(ifr);
2144 
2145 	if (cmd == SIOCSHWTSTAMP)
2146 		return efx_ptp_set_ts_config(efx, ifr);
2147 	if (cmd == SIOCGHWTSTAMP)
2148 		return efx_ptp_get_ts_config(efx, ifr);
2149 
2150 	/* Convert phy_id from older PRTAD/DEVAD format */
2151 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2152 	    (data->phy_id & 0xfc00) == 0x0400)
2153 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2154 
2155 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
2156 }
2157 
2158 /**************************************************************************
2159  *
2160  * NAPI interface
2161  *
2162  **************************************************************************/
2163 
2164 static void efx_init_napi_channel(struct efx_channel *channel)
2165 {
2166 	struct efx_nic *efx = channel->efx;
2167 
2168 	channel->napi_dev = efx->net_dev;
2169 	netif_napi_add(channel->napi_dev, &channel->napi_str,
2170 		       efx_poll, napi_weight);
2171 }
2172 
2173 static void efx_init_napi(struct efx_nic *efx)
2174 {
2175 	struct efx_channel *channel;
2176 
2177 	efx_for_each_channel(channel, efx)
2178 		efx_init_napi_channel(channel);
2179 }
2180 
2181 static void efx_fini_napi_channel(struct efx_channel *channel)
2182 {
2183 	if (channel->napi_dev)
2184 		netif_napi_del(&channel->napi_str);
2185 
2186 	channel->napi_dev = NULL;
2187 }
2188 
2189 static void efx_fini_napi(struct efx_nic *efx)
2190 {
2191 	struct efx_channel *channel;
2192 
2193 	efx_for_each_channel(channel, efx)
2194 		efx_fini_napi_channel(channel);
2195 }
2196 
2197 /**************************************************************************
2198  *
2199  * Kernel netpoll interface
2200  *
2201  *************************************************************************/
2202 
2203 #ifdef CONFIG_NET_POLL_CONTROLLER
2204 
2205 /* Although in the common case interrupts will be disabled, this is not
2206  * guaranteed. However, all our work happens inside the NAPI callback,
2207  * so no locking is required.
2208  */
2209 static void efx_netpoll(struct net_device *net_dev)
2210 {
2211 	struct efx_nic *efx = netdev_priv(net_dev);
2212 	struct efx_channel *channel;
2213 
2214 	efx_for_each_channel(channel, efx)
2215 		efx_schedule_channel(channel);
2216 }
2217 
2218 #endif
2219 
2220 /**************************************************************************
2221  *
2222  * Kernel net device interface
2223  *
2224  *************************************************************************/
2225 
2226 /* Context: process, rtnl_lock() held. */
2227 int efx_net_open(struct net_device *net_dev)
2228 {
2229 	struct efx_nic *efx = netdev_priv(net_dev);
2230 	int rc;
2231 
2232 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2233 		  raw_smp_processor_id());
2234 
2235 	rc = efx_check_disabled(efx);
2236 	if (rc)
2237 		return rc;
2238 	if (efx->phy_mode & PHY_MODE_SPECIAL)
2239 		return -EBUSY;
2240 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2241 		return -EIO;
2242 
2243 	/* Notify the kernel of the link state polled during driver load,
2244 	 * before the monitor starts running */
2245 	efx_link_status_changed(efx);
2246 
2247 	efx_start_all(efx);
2248 	if (efx->state == STATE_DISABLED || efx->reset_pending)
2249 		netif_device_detach(efx->net_dev);
2250 	efx_selftest_async_start(efx);
2251 	return 0;
2252 }
2253 
2254 /* Context: process, rtnl_lock() held.
2255  * Note that the kernel will ignore our return code; this method
2256  * should really be a void.
2257  */
2258 int efx_net_stop(struct net_device *net_dev)
2259 {
2260 	struct efx_nic *efx = netdev_priv(net_dev);
2261 
2262 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2263 		  raw_smp_processor_id());
2264 
2265 	/* Stop the device and flush all the channels */
2266 	efx_stop_all(efx);
2267 
2268 	return 0;
2269 }
2270 
2271 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2272 static void efx_net_stats(struct net_device *net_dev,
2273 			  struct rtnl_link_stats64 *stats)
2274 {
2275 	struct efx_nic *efx = netdev_priv(net_dev);
2276 
2277 	spin_lock_bh(&efx->stats_lock);
2278 	efx->type->update_stats(efx, NULL, stats);
2279 	spin_unlock_bh(&efx->stats_lock);
2280 }
2281 
2282 /* Context: netif_tx_lock held, BHs disabled. */
2283 static void efx_watchdog(struct net_device *net_dev)
2284 {
2285 	struct efx_nic *efx = netdev_priv(net_dev);
2286 
2287 	netif_err(efx, tx_err, efx->net_dev,
2288 		  "TX stuck with port_enabled=%d: resetting channels\n",
2289 		  efx->port_enabled);
2290 
2291 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2292 }
2293 
2294 
2295 /* Context: process, rtnl_lock() held. */
2296 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2297 {
2298 	struct efx_nic *efx = netdev_priv(net_dev);
2299 	int rc;
2300 
2301 	rc = efx_check_disabled(efx);
2302 	if (rc)
2303 		return rc;
2304 
2305 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2306 
2307 	efx_device_detach_sync(efx);
2308 	efx_stop_all(efx);
2309 
2310 	mutex_lock(&efx->mac_lock);
2311 	net_dev->mtu = new_mtu;
2312 	efx_mac_reconfigure(efx);
2313 	mutex_unlock(&efx->mac_lock);
2314 
2315 	efx_start_all(efx);
2316 	efx_device_attach_if_not_resetting(efx);
2317 	return 0;
2318 }
2319 
2320 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2321 {
2322 	struct efx_nic *efx = netdev_priv(net_dev);
2323 	struct sockaddr *addr = data;
2324 	u8 *new_addr = addr->sa_data;
2325 	u8 old_addr[6];
2326 	int rc;
2327 
2328 	if (!is_valid_ether_addr(new_addr)) {
2329 		netif_err(efx, drv, efx->net_dev,
2330 			  "invalid ethernet MAC address requested: %pM\n",
2331 			  new_addr);
2332 		return -EADDRNOTAVAIL;
2333 	}
2334 
2335 	/* save old address */
2336 	ether_addr_copy(old_addr, net_dev->dev_addr);
2337 	ether_addr_copy(net_dev->dev_addr, new_addr);
2338 	if (efx->type->set_mac_address) {
2339 		rc = efx->type->set_mac_address(efx);
2340 		if (rc) {
2341 			ether_addr_copy(net_dev->dev_addr, old_addr);
2342 			return rc;
2343 		}
2344 	}
2345 
2346 	/* Reconfigure the MAC */
2347 	mutex_lock(&efx->mac_lock);
2348 	efx_mac_reconfigure(efx);
2349 	mutex_unlock(&efx->mac_lock);
2350 
2351 	return 0;
2352 }
2353 
2354 /* Context: netif_addr_lock held, BHs disabled. */
2355 static void efx_set_rx_mode(struct net_device *net_dev)
2356 {
2357 	struct efx_nic *efx = netdev_priv(net_dev);
2358 
2359 	if (efx->port_enabled)
2360 		queue_work(efx->workqueue, &efx->mac_work);
2361 	/* Otherwise efx_start_port() will do this */
2362 }
2363 
2364 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2365 {
2366 	struct efx_nic *efx = netdev_priv(net_dev);
2367 	int rc;
2368 
2369 	/* If disabling RX n-tuple filtering, clear existing filters */
2370 	if (net_dev->features & ~data & NETIF_F_NTUPLE) {
2371 		rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2372 		if (rc)
2373 			return rc;
2374 	}
2375 
2376 	/* If Rx VLAN filter is changed, update filters via mac_reconfigure.
2377 	 * If rx-fcs is changed, mac_reconfigure updates that too.
2378 	 */
2379 	if ((net_dev->features ^ data) & (NETIF_F_HW_VLAN_CTAG_FILTER |
2380 					  NETIF_F_RXFCS)) {
2381 		/* efx_set_rx_mode() will schedule MAC work to update filters
2382 		 * when a new features are finally set in net_dev.
2383 		 */
2384 		efx_set_rx_mode(net_dev);
2385 	}
2386 
2387 	return 0;
2388 }
2389 
2390 static int efx_get_phys_port_id(struct net_device *net_dev,
2391 				struct netdev_phys_item_id *ppid)
2392 {
2393 	struct efx_nic *efx = netdev_priv(net_dev);
2394 
2395 	if (efx->type->get_phys_port_id)
2396 		return efx->type->get_phys_port_id(efx, ppid);
2397 	else
2398 		return -EOPNOTSUPP;
2399 }
2400 
2401 static int efx_get_phys_port_name(struct net_device *net_dev,
2402 				  char *name, size_t len)
2403 {
2404 	struct efx_nic *efx = netdev_priv(net_dev);
2405 
2406 	if (snprintf(name, len, "p%u", efx->port_num) >= len)
2407 		return -EINVAL;
2408 	return 0;
2409 }
2410 
2411 static int efx_vlan_rx_add_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2412 {
2413 	struct efx_nic *efx = netdev_priv(net_dev);
2414 
2415 	if (efx->type->vlan_rx_add_vid)
2416 		return efx->type->vlan_rx_add_vid(efx, proto, vid);
2417 	else
2418 		return -EOPNOTSUPP;
2419 }
2420 
2421 static int efx_vlan_rx_kill_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2422 {
2423 	struct efx_nic *efx = netdev_priv(net_dev);
2424 
2425 	if (efx->type->vlan_rx_kill_vid)
2426 		return efx->type->vlan_rx_kill_vid(efx, proto, vid);
2427 	else
2428 		return -EOPNOTSUPP;
2429 }
2430 
2431 static int efx_udp_tunnel_type_map(enum udp_parsable_tunnel_type in)
2432 {
2433 	switch (in) {
2434 	case UDP_TUNNEL_TYPE_VXLAN:
2435 		return TUNNEL_ENCAP_UDP_PORT_ENTRY_VXLAN;
2436 	case UDP_TUNNEL_TYPE_GENEVE:
2437 		return TUNNEL_ENCAP_UDP_PORT_ENTRY_GENEVE;
2438 	default:
2439 		return -1;
2440 	}
2441 }
2442 
2443 static void efx_udp_tunnel_add(struct net_device *dev, struct udp_tunnel_info *ti)
2444 {
2445 	struct efx_nic *efx = netdev_priv(dev);
2446 	struct efx_udp_tunnel tnl;
2447 	int efx_tunnel_type;
2448 
2449 	efx_tunnel_type = efx_udp_tunnel_type_map(ti->type);
2450 	if (efx_tunnel_type < 0)
2451 		return;
2452 
2453 	tnl.type = (u16)efx_tunnel_type;
2454 	tnl.port = ti->port;
2455 
2456 	if (efx->type->udp_tnl_add_port)
2457 		(void)efx->type->udp_tnl_add_port(efx, tnl);
2458 }
2459 
2460 static void efx_udp_tunnel_del(struct net_device *dev, struct udp_tunnel_info *ti)
2461 {
2462 	struct efx_nic *efx = netdev_priv(dev);
2463 	struct efx_udp_tunnel tnl;
2464 	int efx_tunnel_type;
2465 
2466 	efx_tunnel_type = efx_udp_tunnel_type_map(ti->type);
2467 	if (efx_tunnel_type < 0)
2468 		return;
2469 
2470 	tnl.type = (u16)efx_tunnel_type;
2471 	tnl.port = ti->port;
2472 
2473 	if (efx->type->udp_tnl_del_port)
2474 		(void)efx->type->udp_tnl_del_port(efx, tnl);
2475 }
2476 
2477 static const struct net_device_ops efx_netdev_ops = {
2478 	.ndo_open		= efx_net_open,
2479 	.ndo_stop		= efx_net_stop,
2480 	.ndo_get_stats64	= efx_net_stats,
2481 	.ndo_tx_timeout		= efx_watchdog,
2482 	.ndo_start_xmit		= efx_hard_start_xmit,
2483 	.ndo_validate_addr	= eth_validate_addr,
2484 	.ndo_do_ioctl		= efx_ioctl,
2485 	.ndo_change_mtu		= efx_change_mtu,
2486 	.ndo_set_mac_address	= efx_set_mac_address,
2487 	.ndo_set_rx_mode	= efx_set_rx_mode,
2488 	.ndo_set_features	= efx_set_features,
2489 	.ndo_vlan_rx_add_vid	= efx_vlan_rx_add_vid,
2490 	.ndo_vlan_rx_kill_vid	= efx_vlan_rx_kill_vid,
2491 #ifdef CONFIG_SFC_SRIOV
2492 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2493 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2494 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2495 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2496 	.ndo_set_vf_link_state  = efx_sriov_set_vf_link_state,
2497 #endif
2498 	.ndo_get_phys_port_id   = efx_get_phys_port_id,
2499 	.ndo_get_phys_port_name	= efx_get_phys_port_name,
2500 #ifdef CONFIG_NET_POLL_CONTROLLER
2501 	.ndo_poll_controller = efx_netpoll,
2502 #endif
2503 	.ndo_setup_tc		= efx_setup_tc,
2504 #ifdef CONFIG_RFS_ACCEL
2505 	.ndo_rx_flow_steer	= efx_filter_rfs,
2506 #endif
2507 	.ndo_udp_tunnel_add	= efx_udp_tunnel_add,
2508 	.ndo_udp_tunnel_del	= efx_udp_tunnel_del,
2509 };
2510 
2511 static void efx_update_name(struct efx_nic *efx)
2512 {
2513 	strcpy(efx->name, efx->net_dev->name);
2514 	efx_mtd_rename(efx);
2515 	efx_set_channel_names(efx);
2516 }
2517 
2518 static int efx_netdev_event(struct notifier_block *this,
2519 			    unsigned long event, void *ptr)
2520 {
2521 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2522 
2523 	if ((net_dev->netdev_ops == &efx_netdev_ops) &&
2524 	    event == NETDEV_CHANGENAME)
2525 		efx_update_name(netdev_priv(net_dev));
2526 
2527 	return NOTIFY_DONE;
2528 }
2529 
2530 static struct notifier_block efx_netdev_notifier = {
2531 	.notifier_call = efx_netdev_event,
2532 };
2533 
2534 static ssize_t
2535 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2536 {
2537 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2538 	return sprintf(buf, "%d\n", efx->phy_type);
2539 }
2540 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2541 
2542 #ifdef CONFIG_SFC_MCDI_LOGGING
2543 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
2544 			     char *buf)
2545 {
2546 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2547 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2548 
2549 	return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
2550 }
2551 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
2552 			    const char *buf, size_t count)
2553 {
2554 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2555 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2556 	bool enable = count > 0 && *buf != '0';
2557 
2558 	mcdi->logging_enabled = enable;
2559 	return count;
2560 }
2561 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
2562 #endif
2563 
2564 static int efx_register_netdev(struct efx_nic *efx)
2565 {
2566 	struct net_device *net_dev = efx->net_dev;
2567 	struct efx_channel *channel;
2568 	int rc;
2569 
2570 	net_dev->watchdog_timeo = 5 * HZ;
2571 	net_dev->irq = efx->pci_dev->irq;
2572 	net_dev->netdev_ops = &efx_netdev_ops;
2573 	if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
2574 		net_dev->priv_flags |= IFF_UNICAST_FLT;
2575 	net_dev->ethtool_ops = &efx_ethtool_ops;
2576 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2577 	net_dev->min_mtu = EFX_MIN_MTU;
2578 	net_dev->max_mtu = EFX_MAX_MTU;
2579 
2580 	rtnl_lock();
2581 
2582 	/* Enable resets to be scheduled and check whether any were
2583 	 * already requested.  If so, the NIC is probably hosed so we
2584 	 * abort.
2585 	 */
2586 	efx->state = STATE_READY;
2587 	smp_mb(); /* ensure we change state before checking reset_pending */
2588 	if (efx->reset_pending) {
2589 		netif_err(efx, probe, efx->net_dev,
2590 			  "aborting probe due to scheduled reset\n");
2591 		rc = -EIO;
2592 		goto fail_locked;
2593 	}
2594 
2595 	rc = dev_alloc_name(net_dev, net_dev->name);
2596 	if (rc < 0)
2597 		goto fail_locked;
2598 	efx_update_name(efx);
2599 
2600 	/* Always start with carrier off; PHY events will detect the link */
2601 	netif_carrier_off(net_dev);
2602 
2603 	rc = register_netdevice(net_dev);
2604 	if (rc)
2605 		goto fail_locked;
2606 
2607 	efx_for_each_channel(channel, efx) {
2608 		struct efx_tx_queue *tx_queue;
2609 		efx_for_each_channel_tx_queue(tx_queue, channel)
2610 			efx_init_tx_queue_core_txq(tx_queue);
2611 	}
2612 
2613 	efx_associate(efx);
2614 
2615 	rtnl_unlock();
2616 
2617 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2618 	if (rc) {
2619 		netif_err(efx, drv, efx->net_dev,
2620 			  "failed to init net dev attributes\n");
2621 		goto fail_registered;
2622 	}
2623 #ifdef CONFIG_SFC_MCDI_LOGGING
2624 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2625 	if (rc) {
2626 		netif_err(efx, drv, efx->net_dev,
2627 			  "failed to init net dev attributes\n");
2628 		goto fail_attr_mcdi_logging;
2629 	}
2630 #endif
2631 
2632 	return 0;
2633 
2634 #ifdef CONFIG_SFC_MCDI_LOGGING
2635 fail_attr_mcdi_logging:
2636 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2637 #endif
2638 fail_registered:
2639 	rtnl_lock();
2640 	efx_dissociate(efx);
2641 	unregister_netdevice(net_dev);
2642 fail_locked:
2643 	efx->state = STATE_UNINIT;
2644 	rtnl_unlock();
2645 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2646 	return rc;
2647 }
2648 
2649 static void efx_unregister_netdev(struct efx_nic *efx)
2650 {
2651 	if (!efx->net_dev)
2652 		return;
2653 
2654 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2655 
2656 	if (efx_dev_registered(efx)) {
2657 		strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2658 #ifdef CONFIG_SFC_MCDI_LOGGING
2659 		device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2660 #endif
2661 		device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2662 		unregister_netdev(efx->net_dev);
2663 	}
2664 }
2665 
2666 /**************************************************************************
2667  *
2668  * Device reset and suspend
2669  *
2670  **************************************************************************/
2671 
2672 /* Tears down the entire software state and most of the hardware state
2673  * before reset.  */
2674 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2675 {
2676 	EFX_ASSERT_RESET_SERIALISED(efx);
2677 
2678 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2679 		efx->type->prepare_flr(efx);
2680 
2681 	efx_stop_all(efx);
2682 	efx_disable_interrupts(efx);
2683 
2684 	mutex_lock(&efx->mac_lock);
2685 	down_write(&efx->filter_sem);
2686 	mutex_lock(&efx->rss_lock);
2687 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2688 	    method != RESET_TYPE_DATAPATH)
2689 		efx->phy_op->fini(efx);
2690 	efx->type->fini(efx);
2691 }
2692 
2693 /* This function will always ensure that the locks acquired in
2694  * efx_reset_down() are released. A failure return code indicates
2695  * that we were unable to reinitialise the hardware, and the
2696  * driver should be disabled. If ok is false, then the rx and tx
2697  * engines are not restarted, pending a RESET_DISABLE. */
2698 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2699 {
2700 	int rc;
2701 
2702 	EFX_ASSERT_RESET_SERIALISED(efx);
2703 
2704 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2705 		efx->type->finish_flr(efx);
2706 
2707 	/* Ensure that SRAM is initialised even if we're disabling the device */
2708 	rc = efx->type->init(efx);
2709 	if (rc) {
2710 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2711 		goto fail;
2712 	}
2713 
2714 	if (!ok)
2715 		goto fail;
2716 
2717 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2718 	    method != RESET_TYPE_DATAPATH) {
2719 		rc = efx->phy_op->init(efx);
2720 		if (rc)
2721 			goto fail;
2722 		rc = efx->phy_op->reconfigure(efx);
2723 		if (rc && rc != -EPERM)
2724 			netif_err(efx, drv, efx->net_dev,
2725 				  "could not restore PHY settings\n");
2726 	}
2727 
2728 	rc = efx_enable_interrupts(efx);
2729 	if (rc)
2730 		goto fail;
2731 
2732 #ifdef CONFIG_SFC_SRIOV
2733 	rc = efx->type->vswitching_restore(efx);
2734 	if (rc) /* not fatal; the PF will still work fine */
2735 		netif_warn(efx, probe, efx->net_dev,
2736 			   "failed to restore vswitching rc=%d;"
2737 			   " VFs may not function\n", rc);
2738 #endif
2739 
2740 	if (efx->type->rx_restore_rss_contexts)
2741 		efx->type->rx_restore_rss_contexts(efx);
2742 	mutex_unlock(&efx->rss_lock);
2743 	efx->type->filter_table_restore(efx);
2744 	up_write(&efx->filter_sem);
2745 	if (efx->type->sriov_reset)
2746 		efx->type->sriov_reset(efx);
2747 
2748 	mutex_unlock(&efx->mac_lock);
2749 
2750 	efx_start_all(efx);
2751 
2752 	if (efx->type->udp_tnl_push_ports)
2753 		efx->type->udp_tnl_push_ports(efx);
2754 
2755 	return 0;
2756 
2757 fail:
2758 	efx->port_initialized = false;
2759 
2760 	mutex_unlock(&efx->rss_lock);
2761 	up_write(&efx->filter_sem);
2762 	mutex_unlock(&efx->mac_lock);
2763 
2764 	return rc;
2765 }
2766 
2767 /* Reset the NIC using the specified method.  Note that the reset may
2768  * fail, in which case the card will be left in an unusable state.
2769  *
2770  * Caller must hold the rtnl_lock.
2771  */
2772 int efx_reset(struct efx_nic *efx, enum reset_type method)
2773 {
2774 	int rc, rc2;
2775 	bool disabled;
2776 
2777 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2778 		   RESET_TYPE(method));
2779 
2780 	efx_device_detach_sync(efx);
2781 	efx_reset_down(efx, method);
2782 
2783 	rc = efx->type->reset(efx, method);
2784 	if (rc) {
2785 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2786 		goto out;
2787 	}
2788 
2789 	/* Clear flags for the scopes we covered.  We assume the NIC and
2790 	 * driver are now quiescent so that there is no race here.
2791 	 */
2792 	if (method < RESET_TYPE_MAX_METHOD)
2793 		efx->reset_pending &= -(1 << (method + 1));
2794 	else /* it doesn't fit into the well-ordered scope hierarchy */
2795 		__clear_bit(method, &efx->reset_pending);
2796 
2797 	/* Reinitialise bus-mastering, which may have been turned off before
2798 	 * the reset was scheduled. This is still appropriate, even in the
2799 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2800 	 * can respond to requests. */
2801 	pci_set_master(efx->pci_dev);
2802 
2803 out:
2804 	/* Leave device stopped if necessary */
2805 	disabled = rc ||
2806 		method == RESET_TYPE_DISABLE ||
2807 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2808 	rc2 = efx_reset_up(efx, method, !disabled);
2809 	if (rc2) {
2810 		disabled = true;
2811 		if (!rc)
2812 			rc = rc2;
2813 	}
2814 
2815 	if (disabled) {
2816 		dev_close(efx->net_dev);
2817 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2818 		efx->state = STATE_DISABLED;
2819 	} else {
2820 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2821 		efx_device_attach_if_not_resetting(efx);
2822 	}
2823 	return rc;
2824 }
2825 
2826 /* Try recovery mechanisms.
2827  * For now only EEH is supported.
2828  * Returns 0 if the recovery mechanisms are unsuccessful.
2829  * Returns a non-zero value otherwise.
2830  */
2831 int efx_try_recovery(struct efx_nic *efx)
2832 {
2833 #ifdef CONFIG_EEH
2834 	/* A PCI error can occur and not be seen by EEH because nothing
2835 	 * happens on the PCI bus. In this case the driver may fail and
2836 	 * schedule a 'recover or reset', leading to this recovery handler.
2837 	 * Manually call the eeh failure check function.
2838 	 */
2839 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2840 	if (eeh_dev_check_failure(eehdev)) {
2841 		/* The EEH mechanisms will handle the error and reset the
2842 		 * device if necessary.
2843 		 */
2844 		return 1;
2845 	}
2846 #endif
2847 	return 0;
2848 }
2849 
2850 static void efx_wait_for_bist_end(struct efx_nic *efx)
2851 {
2852 	int i;
2853 
2854 	for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2855 		if (efx_mcdi_poll_reboot(efx))
2856 			goto out;
2857 		msleep(BIST_WAIT_DELAY_MS);
2858 	}
2859 
2860 	netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2861 out:
2862 	/* Either way unset the BIST flag. If we found no reboot we probably
2863 	 * won't recover, but we should try.
2864 	 */
2865 	efx->mc_bist_for_other_fn = false;
2866 }
2867 
2868 /* The worker thread exists so that code that cannot sleep can
2869  * schedule a reset for later.
2870  */
2871 static void efx_reset_work(struct work_struct *data)
2872 {
2873 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2874 	unsigned long pending;
2875 	enum reset_type method;
2876 
2877 	pending = READ_ONCE(efx->reset_pending);
2878 	method = fls(pending) - 1;
2879 
2880 	if (method == RESET_TYPE_MC_BIST)
2881 		efx_wait_for_bist_end(efx);
2882 
2883 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2884 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2885 	    efx_try_recovery(efx))
2886 		return;
2887 
2888 	if (!pending)
2889 		return;
2890 
2891 	rtnl_lock();
2892 
2893 	/* We checked the state in efx_schedule_reset() but it may
2894 	 * have changed by now.  Now that we have the RTNL lock,
2895 	 * it cannot change again.
2896 	 */
2897 	if (efx->state == STATE_READY)
2898 		(void)efx_reset(efx, method);
2899 
2900 	rtnl_unlock();
2901 }
2902 
2903 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2904 {
2905 	enum reset_type method;
2906 
2907 	if (efx->state == STATE_RECOVERY) {
2908 		netif_dbg(efx, drv, efx->net_dev,
2909 			  "recovering: skip scheduling %s reset\n",
2910 			  RESET_TYPE(type));
2911 		return;
2912 	}
2913 
2914 	switch (type) {
2915 	case RESET_TYPE_INVISIBLE:
2916 	case RESET_TYPE_ALL:
2917 	case RESET_TYPE_RECOVER_OR_ALL:
2918 	case RESET_TYPE_WORLD:
2919 	case RESET_TYPE_DISABLE:
2920 	case RESET_TYPE_RECOVER_OR_DISABLE:
2921 	case RESET_TYPE_DATAPATH:
2922 	case RESET_TYPE_MC_BIST:
2923 	case RESET_TYPE_MCDI_TIMEOUT:
2924 		method = type;
2925 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2926 			  RESET_TYPE(method));
2927 		break;
2928 	default:
2929 		method = efx->type->map_reset_reason(type);
2930 		netif_dbg(efx, drv, efx->net_dev,
2931 			  "scheduling %s reset for %s\n",
2932 			  RESET_TYPE(method), RESET_TYPE(type));
2933 		break;
2934 	}
2935 
2936 	set_bit(method, &efx->reset_pending);
2937 	smp_mb(); /* ensure we change reset_pending before checking state */
2938 
2939 	/* If we're not READY then just leave the flags set as the cue
2940 	 * to abort probing or reschedule the reset later.
2941 	 */
2942 	if (READ_ONCE(efx->state) != STATE_READY)
2943 		return;
2944 
2945 	/* efx_process_channel() will no longer read events once a
2946 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2947 	efx_mcdi_mode_poll(efx);
2948 
2949 	queue_work(reset_workqueue, &efx->reset_work);
2950 }
2951 
2952 /**************************************************************************
2953  *
2954  * List of NICs we support
2955  *
2956  **************************************************************************/
2957 
2958 /* PCI device ID table */
2959 static const struct pci_device_id efx_pci_table[] = {
2960 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2961 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2962 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2963 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2964 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903),  /* SFC9120 PF */
2965 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2966 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903),  /* SFC9120 VF */
2967 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2968 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923),  /* SFC9140 PF */
2969 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2970 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1923),  /* SFC9140 VF */
2971 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2972 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0a03),  /* SFC9220 PF */
2973 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2974 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1a03),  /* SFC9220 VF */
2975 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2976 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0b03),  /* SFC9250 PF */
2977 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2978 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1b03),  /* SFC9250 VF */
2979 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2980 	{0}			/* end of list */
2981 };
2982 
2983 /**************************************************************************
2984  *
2985  * Dummy PHY/MAC operations
2986  *
2987  * Can be used for some unimplemented operations
2988  * Needed so all function pointers are valid and do not have to be tested
2989  * before use
2990  *
2991  **************************************************************************/
2992 int efx_port_dummy_op_int(struct efx_nic *efx)
2993 {
2994 	return 0;
2995 }
2996 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2997 
2998 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2999 {
3000 	return false;
3001 }
3002 
3003 static const struct efx_phy_operations efx_dummy_phy_operations = {
3004 	.init		 = efx_port_dummy_op_int,
3005 	.reconfigure	 = efx_port_dummy_op_int,
3006 	.poll		 = efx_port_dummy_op_poll,
3007 	.fini		 = efx_port_dummy_op_void,
3008 };
3009 
3010 /**************************************************************************
3011  *
3012  * Data housekeeping
3013  *
3014  **************************************************************************/
3015 
3016 /* This zeroes out and then fills in the invariants in a struct
3017  * efx_nic (including all sub-structures).
3018  */
3019 static int efx_init_struct(struct efx_nic *efx,
3020 			   struct pci_dev *pci_dev, struct net_device *net_dev)
3021 {
3022 	int rc = -ENOMEM, i;
3023 
3024 	/* Initialise common structures */
3025 	INIT_LIST_HEAD(&efx->node);
3026 	INIT_LIST_HEAD(&efx->secondary_list);
3027 	spin_lock_init(&efx->biu_lock);
3028 #ifdef CONFIG_SFC_MTD
3029 	INIT_LIST_HEAD(&efx->mtd_list);
3030 #endif
3031 	INIT_WORK(&efx->reset_work, efx_reset_work);
3032 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
3033 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
3034 	efx->pci_dev = pci_dev;
3035 	efx->msg_enable = debug;
3036 	efx->state = STATE_UNINIT;
3037 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
3038 
3039 	efx->net_dev = net_dev;
3040 	efx->rx_prefix_size = efx->type->rx_prefix_size;
3041 	efx->rx_ip_align =
3042 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
3043 	efx->rx_packet_hash_offset =
3044 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
3045 	efx->rx_packet_ts_offset =
3046 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
3047 	INIT_LIST_HEAD(&efx->rss_context.list);
3048 	mutex_init(&efx->rss_lock);
3049 	spin_lock_init(&efx->stats_lock);
3050 	efx->vi_stride = EFX_DEFAULT_VI_STRIDE;
3051 	efx->num_mac_stats = MC_CMD_MAC_NSTATS;
3052 	BUILD_BUG_ON(MC_CMD_MAC_NSTATS - 1 != MC_CMD_MAC_GENERATION_END);
3053 	mutex_init(&efx->mac_lock);
3054 #ifdef CONFIG_RFS_ACCEL
3055 	mutex_init(&efx->rps_mutex);
3056 	spin_lock_init(&efx->rps_hash_lock);
3057 	/* Failure to allocate is not fatal, but may degrade ARFS performance */
3058 	efx->rps_hash_table = kcalloc(EFX_ARFS_HASH_TABLE_SIZE,
3059 				      sizeof(*efx->rps_hash_table), GFP_KERNEL);
3060 #endif
3061 	efx->phy_op = &efx_dummy_phy_operations;
3062 	efx->mdio.dev = net_dev;
3063 	INIT_WORK(&efx->mac_work, efx_mac_work);
3064 	init_waitqueue_head(&efx->flush_wq);
3065 
3066 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
3067 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
3068 		if (!efx->channel[i])
3069 			goto fail;
3070 		efx->msi_context[i].efx = efx;
3071 		efx->msi_context[i].index = i;
3072 	}
3073 
3074 	/* Higher numbered interrupt modes are less capable! */
3075 	if (WARN_ON_ONCE(efx->type->max_interrupt_mode >
3076 			 efx->type->min_interrupt_mode)) {
3077 		rc = -EIO;
3078 		goto fail;
3079 	}
3080 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
3081 				  interrupt_mode);
3082 	efx->interrupt_mode = min(efx->type->min_interrupt_mode,
3083 				  interrupt_mode);
3084 
3085 	/* Would be good to use the net_dev name, but we're too early */
3086 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
3087 		 pci_name(pci_dev));
3088 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
3089 	if (!efx->workqueue)
3090 		goto fail;
3091 
3092 	return 0;
3093 
3094 fail:
3095 	efx_fini_struct(efx);
3096 	return rc;
3097 }
3098 
3099 static void efx_fini_struct(struct efx_nic *efx)
3100 {
3101 	int i;
3102 
3103 #ifdef CONFIG_RFS_ACCEL
3104 	kfree(efx->rps_hash_table);
3105 #endif
3106 
3107 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
3108 		kfree(efx->channel[i]);
3109 
3110 	kfree(efx->vpd_sn);
3111 
3112 	if (efx->workqueue) {
3113 		destroy_workqueue(efx->workqueue);
3114 		efx->workqueue = NULL;
3115 	}
3116 }
3117 
3118 void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
3119 {
3120 	u64 n_rx_nodesc_trunc = 0;
3121 	struct efx_channel *channel;
3122 
3123 	efx_for_each_channel(channel, efx)
3124 		n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
3125 	stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
3126 	stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
3127 }
3128 
3129 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
3130 			   const struct efx_filter_spec *right)
3131 {
3132 	if ((left->match_flags ^ right->match_flags) |
3133 	    ((left->flags ^ right->flags) &
3134 	     (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
3135 		return false;
3136 
3137 	return memcmp(&left->outer_vid, &right->outer_vid,
3138 		      sizeof(struct efx_filter_spec) -
3139 		      offsetof(struct efx_filter_spec, outer_vid)) == 0;
3140 }
3141 
3142 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
3143 {
3144 	BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
3145 	return jhash2((const u32 *)&spec->outer_vid,
3146 		      (sizeof(struct efx_filter_spec) -
3147 		       offsetof(struct efx_filter_spec, outer_vid)) / 4,
3148 		      0);
3149 }
3150 
3151 #ifdef CONFIG_RFS_ACCEL
3152 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
3153 			bool *force)
3154 {
3155 	if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
3156 		/* ARFS is currently updating this entry, leave it */
3157 		return false;
3158 	}
3159 	if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
3160 		/* ARFS tried and failed to update this, so it's probably out
3161 		 * of date.  Remove the filter and the ARFS rule entry.
3162 		 */
3163 		rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
3164 		*force = true;
3165 		return true;
3166 	} else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
3167 		/* ARFS has moved on, so old filter is not needed.  Since we did
3168 		 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
3169 		 * not be removed by efx_rps_hash_del() subsequently.
3170 		 */
3171 		*force = true;
3172 		return true;
3173 	}
3174 	/* Remove it iff ARFS wants to. */
3175 	return true;
3176 }
3177 
3178 static
3179 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
3180 				       const struct efx_filter_spec *spec)
3181 {
3182 	u32 hash = efx_filter_spec_hash(spec);
3183 
3184 	WARN_ON(!spin_is_locked(&efx->rps_hash_lock));
3185 	if (!efx->rps_hash_table)
3186 		return NULL;
3187 	return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
3188 }
3189 
3190 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
3191 					const struct efx_filter_spec *spec)
3192 {
3193 	struct efx_arfs_rule *rule;
3194 	struct hlist_head *head;
3195 	struct hlist_node *node;
3196 
3197 	head = efx_rps_hash_bucket(efx, spec);
3198 	if (!head)
3199 		return NULL;
3200 	hlist_for_each(node, head) {
3201 		rule = container_of(node, struct efx_arfs_rule, node);
3202 		if (efx_filter_spec_equal(spec, &rule->spec))
3203 			return rule;
3204 	}
3205 	return NULL;
3206 }
3207 
3208 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
3209 				       const struct efx_filter_spec *spec,
3210 				       bool *new)
3211 {
3212 	struct efx_arfs_rule *rule;
3213 	struct hlist_head *head;
3214 	struct hlist_node *node;
3215 
3216 	head = efx_rps_hash_bucket(efx, spec);
3217 	if (!head)
3218 		return NULL;
3219 	hlist_for_each(node, head) {
3220 		rule = container_of(node, struct efx_arfs_rule, node);
3221 		if (efx_filter_spec_equal(spec, &rule->spec)) {
3222 			*new = false;
3223 			return rule;
3224 		}
3225 	}
3226 	rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
3227 	*new = true;
3228 	if (rule) {
3229 		memcpy(&rule->spec, spec, sizeof(rule->spec));
3230 		hlist_add_head(&rule->node, head);
3231 	}
3232 	return rule;
3233 }
3234 
3235 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
3236 {
3237 	struct efx_arfs_rule *rule;
3238 	struct hlist_head *head;
3239 	struct hlist_node *node;
3240 
3241 	head = efx_rps_hash_bucket(efx, spec);
3242 	if (WARN_ON(!head))
3243 		return;
3244 	hlist_for_each(node, head) {
3245 		rule = container_of(node, struct efx_arfs_rule, node);
3246 		if (efx_filter_spec_equal(spec, &rule->spec)) {
3247 			/* Someone already reused the entry.  We know that if
3248 			 * this check doesn't fire (i.e. filter_id == REMOVING)
3249 			 * then the REMOVING mark was put there by our caller,
3250 			 * because caller is holding a lock on filter table and
3251 			 * only holders of that lock set REMOVING.
3252 			 */
3253 			if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
3254 				return;
3255 			hlist_del(node);
3256 			kfree(rule);
3257 			return;
3258 		}
3259 	}
3260 	/* We didn't find it. */
3261 	WARN_ON(1);
3262 }
3263 #endif
3264 
3265 /* RSS contexts.  We're using linked lists and crappy O(n) algorithms, because
3266  * (a) this is an infrequent control-plane operation and (b) n is small (max 64)
3267  */
3268 struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx)
3269 {
3270 	struct list_head *head = &efx->rss_context.list;
3271 	struct efx_rss_context *ctx, *new;
3272 	u32 id = 1; /* Don't use zero, that refers to the master RSS context */
3273 
3274 	WARN_ON(!mutex_is_locked(&efx->rss_lock));
3275 
3276 	/* Search for first gap in the numbering */
3277 	list_for_each_entry(ctx, head, list) {
3278 		if (ctx->user_id != id)
3279 			break;
3280 		id++;
3281 		/* Check for wrap.  If this happens, we have nearly 2^32
3282 		 * allocated RSS contexts, which seems unlikely.
3283 		 */
3284 		if (WARN_ON_ONCE(!id))
3285 			return NULL;
3286 	}
3287 
3288 	/* Create the new entry */
3289 	new = kmalloc(sizeof(struct efx_rss_context), GFP_KERNEL);
3290 	if (!new)
3291 		return NULL;
3292 	new->context_id = EFX_EF10_RSS_CONTEXT_INVALID;
3293 	new->rx_hash_udp_4tuple = false;
3294 
3295 	/* Insert the new entry into the gap */
3296 	new->user_id = id;
3297 	list_add_tail(&new->list, &ctx->list);
3298 	return new;
3299 }
3300 
3301 struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id)
3302 {
3303 	struct list_head *head = &efx->rss_context.list;
3304 	struct efx_rss_context *ctx;
3305 
3306 	WARN_ON(!mutex_is_locked(&efx->rss_lock));
3307 
3308 	list_for_each_entry(ctx, head, list)
3309 		if (ctx->user_id == id)
3310 			return ctx;
3311 	return NULL;
3312 }
3313 
3314 void efx_free_rss_context_entry(struct efx_rss_context *ctx)
3315 {
3316 	list_del(&ctx->list);
3317 	kfree(ctx);
3318 }
3319 
3320 /**************************************************************************
3321  *
3322  * PCI interface
3323  *
3324  **************************************************************************/
3325 
3326 /* Main body of final NIC shutdown code
3327  * This is called only at module unload (or hotplug removal).
3328  */
3329 static void efx_pci_remove_main(struct efx_nic *efx)
3330 {
3331 	/* Flush reset_work. It can no longer be scheduled since we
3332 	 * are not READY.
3333 	 */
3334 	BUG_ON(efx->state == STATE_READY);
3335 	cancel_work_sync(&efx->reset_work);
3336 
3337 	efx_disable_interrupts(efx);
3338 	efx_clear_interrupt_affinity(efx);
3339 	efx_nic_fini_interrupt(efx);
3340 	efx_fini_port(efx);
3341 	efx->type->fini(efx);
3342 	efx_fini_napi(efx);
3343 	efx_remove_all(efx);
3344 }
3345 
3346 /* Final NIC shutdown
3347  * This is called only at module unload (or hotplug removal).  A PF can call
3348  * this on its VFs to ensure they are unbound first.
3349  */
3350 static void efx_pci_remove(struct pci_dev *pci_dev)
3351 {
3352 	struct efx_nic *efx;
3353 
3354 	efx = pci_get_drvdata(pci_dev);
3355 	if (!efx)
3356 		return;
3357 
3358 	/* Mark the NIC as fini, then stop the interface */
3359 	rtnl_lock();
3360 	efx_dissociate(efx);
3361 	dev_close(efx->net_dev);
3362 	efx_disable_interrupts(efx);
3363 	efx->state = STATE_UNINIT;
3364 	rtnl_unlock();
3365 
3366 	if (efx->type->sriov_fini)
3367 		efx->type->sriov_fini(efx);
3368 
3369 	efx_unregister_netdev(efx);
3370 
3371 	efx_mtd_remove(efx);
3372 
3373 	efx_pci_remove_main(efx);
3374 
3375 	efx_fini_io(efx);
3376 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
3377 
3378 	efx_fini_struct(efx);
3379 	free_netdev(efx->net_dev);
3380 
3381 	pci_disable_pcie_error_reporting(pci_dev);
3382 };
3383 
3384 /* NIC VPD information
3385  * Called during probe to display the part number of the
3386  * installed NIC.  VPD is potentially very large but this should
3387  * always appear within the first 512 bytes.
3388  */
3389 #define SFC_VPD_LEN 512
3390 static void efx_probe_vpd_strings(struct efx_nic *efx)
3391 {
3392 	struct pci_dev *dev = efx->pci_dev;
3393 	char vpd_data[SFC_VPD_LEN];
3394 	ssize_t vpd_size;
3395 	int ro_start, ro_size, i, j;
3396 
3397 	/* Get the vpd data from the device */
3398 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
3399 	if (vpd_size <= 0) {
3400 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
3401 		return;
3402 	}
3403 
3404 	/* Get the Read only section */
3405 	ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
3406 	if (ro_start < 0) {
3407 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
3408 		return;
3409 	}
3410 
3411 	ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
3412 	j = ro_size;
3413 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3414 	if (i + j > vpd_size)
3415 		j = vpd_size - i;
3416 
3417 	/* Get the Part number */
3418 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
3419 	if (i < 0) {
3420 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
3421 		return;
3422 	}
3423 
3424 	j = pci_vpd_info_field_size(&vpd_data[i]);
3425 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3426 	if (i + j > vpd_size) {
3427 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
3428 		return;
3429 	}
3430 
3431 	netif_info(efx, drv, efx->net_dev,
3432 		   "Part Number : %.*s\n", j, &vpd_data[i]);
3433 
3434 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3435 	j = ro_size;
3436 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
3437 	if (i < 0) {
3438 		netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
3439 		return;
3440 	}
3441 
3442 	j = pci_vpd_info_field_size(&vpd_data[i]);
3443 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3444 	if (i + j > vpd_size) {
3445 		netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
3446 		return;
3447 	}
3448 
3449 	efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
3450 	if (!efx->vpd_sn)
3451 		return;
3452 
3453 	snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
3454 }
3455 
3456 
3457 /* Main body of NIC initialisation
3458  * This is called at module load (or hotplug insertion, theoretically).
3459  */
3460 static int efx_pci_probe_main(struct efx_nic *efx)
3461 {
3462 	int rc;
3463 
3464 	/* Do start-of-day initialisation */
3465 	rc = efx_probe_all(efx);
3466 	if (rc)
3467 		goto fail1;
3468 
3469 	efx_init_napi(efx);
3470 
3471 	down_write(&efx->filter_sem);
3472 	rc = efx->type->init(efx);
3473 	up_write(&efx->filter_sem);
3474 	if (rc) {
3475 		netif_err(efx, probe, efx->net_dev,
3476 			  "failed to initialise NIC\n");
3477 		goto fail3;
3478 	}
3479 
3480 	rc = efx_init_port(efx);
3481 	if (rc) {
3482 		netif_err(efx, probe, efx->net_dev,
3483 			  "failed to initialise port\n");
3484 		goto fail4;
3485 	}
3486 
3487 	rc = efx_nic_init_interrupt(efx);
3488 	if (rc)
3489 		goto fail5;
3490 
3491 	efx_set_interrupt_affinity(efx);
3492 	rc = efx_enable_interrupts(efx);
3493 	if (rc)
3494 		goto fail6;
3495 
3496 	return 0;
3497 
3498  fail6:
3499 	efx_clear_interrupt_affinity(efx);
3500 	efx_nic_fini_interrupt(efx);
3501  fail5:
3502 	efx_fini_port(efx);
3503  fail4:
3504 	efx->type->fini(efx);
3505  fail3:
3506 	efx_fini_napi(efx);
3507 	efx_remove_all(efx);
3508  fail1:
3509 	return rc;
3510 }
3511 
3512 static int efx_pci_probe_post_io(struct efx_nic *efx)
3513 {
3514 	struct net_device *net_dev = efx->net_dev;
3515 	int rc = efx_pci_probe_main(efx);
3516 
3517 	if (rc)
3518 		return rc;
3519 
3520 	if (efx->type->sriov_init) {
3521 		rc = efx->type->sriov_init(efx);
3522 		if (rc)
3523 			netif_err(efx, probe, efx->net_dev,
3524 				  "SR-IOV can't be enabled rc %d\n", rc);
3525 	}
3526 
3527 	/* Determine netdevice features */
3528 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
3529 			      NETIF_F_TSO | NETIF_F_RXCSUM | NETIF_F_RXALL);
3530 	if (efx->type->offload_features & (NETIF_F_IPV6_CSUM | NETIF_F_HW_CSUM))
3531 		net_dev->features |= NETIF_F_TSO6;
3532 	/* Check whether device supports TSO */
3533 	if (!efx->type->tso_versions || !efx->type->tso_versions(efx))
3534 		net_dev->features &= ~NETIF_F_ALL_TSO;
3535 	/* Mask for features that also apply to VLAN devices */
3536 	net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
3537 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
3538 				   NETIF_F_RXCSUM);
3539 
3540 	net_dev->hw_features |= net_dev->features & ~efx->fixed_features;
3541 
3542 	/* Disable receiving frames with bad FCS, by default. */
3543 	net_dev->features &= ~NETIF_F_RXALL;
3544 
3545 	/* Disable VLAN filtering by default.  It may be enforced if
3546 	 * the feature is fixed (i.e. VLAN filters are required to
3547 	 * receive VLAN tagged packets due to vPort restrictions).
3548 	 */
3549 	net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
3550 	net_dev->features |= efx->fixed_features;
3551 
3552 	rc = efx_register_netdev(efx);
3553 	if (!rc)
3554 		return 0;
3555 
3556 	efx_pci_remove_main(efx);
3557 	return rc;
3558 }
3559 
3560 /* NIC initialisation
3561  *
3562  * This is called at module load (or hotplug insertion,
3563  * theoretically).  It sets up PCI mappings, resets the NIC,
3564  * sets up and registers the network devices with the kernel and hooks
3565  * the interrupt service routine.  It does not prepare the device for
3566  * transmission; this is left to the first time one of the network
3567  * interfaces is brought up (i.e. efx_net_open).
3568  */
3569 static int efx_pci_probe(struct pci_dev *pci_dev,
3570 			 const struct pci_device_id *entry)
3571 {
3572 	struct net_device *net_dev;
3573 	struct efx_nic *efx;
3574 	int rc;
3575 
3576 	/* Allocate and initialise a struct net_device and struct efx_nic */
3577 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
3578 				     EFX_MAX_RX_QUEUES);
3579 	if (!net_dev)
3580 		return -ENOMEM;
3581 	efx = netdev_priv(net_dev);
3582 	efx->type = (const struct efx_nic_type *) entry->driver_data;
3583 	efx->fixed_features |= NETIF_F_HIGHDMA;
3584 
3585 	pci_set_drvdata(pci_dev, efx);
3586 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
3587 	rc = efx_init_struct(efx, pci_dev, net_dev);
3588 	if (rc)
3589 		goto fail1;
3590 
3591 	netif_info(efx, probe, efx->net_dev,
3592 		   "Solarflare NIC detected\n");
3593 
3594 	if (!efx->type->is_vf)
3595 		efx_probe_vpd_strings(efx);
3596 
3597 	/* Set up basic I/O (BAR mappings etc) */
3598 	rc = efx_init_io(efx);
3599 	if (rc)
3600 		goto fail2;
3601 
3602 	rc = efx_pci_probe_post_io(efx);
3603 	if (rc) {
3604 		/* On failure, retry once immediately.
3605 		 * If we aborted probe due to a scheduled reset, dismiss it.
3606 		 */
3607 		efx->reset_pending = 0;
3608 		rc = efx_pci_probe_post_io(efx);
3609 		if (rc) {
3610 			/* On another failure, retry once more
3611 			 * after a 50-305ms delay.
3612 			 */
3613 			unsigned char r;
3614 
3615 			get_random_bytes(&r, 1);
3616 			msleep((unsigned int)r + 50);
3617 			efx->reset_pending = 0;
3618 			rc = efx_pci_probe_post_io(efx);
3619 		}
3620 	}
3621 	if (rc)
3622 		goto fail3;
3623 
3624 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3625 
3626 	/* Try to create MTDs, but allow this to fail */
3627 	rtnl_lock();
3628 	rc = efx_mtd_probe(efx);
3629 	rtnl_unlock();
3630 	if (rc && rc != -EPERM)
3631 		netif_warn(efx, probe, efx->net_dev,
3632 			   "failed to create MTDs (%d)\n", rc);
3633 
3634 	rc = pci_enable_pcie_error_reporting(pci_dev);
3635 	if (rc && rc != -EINVAL)
3636 		netif_notice(efx, probe, efx->net_dev,
3637 			     "PCIE error reporting unavailable (%d).\n",
3638 			     rc);
3639 
3640 	if (efx->type->udp_tnl_push_ports)
3641 		efx->type->udp_tnl_push_ports(efx);
3642 
3643 	return 0;
3644 
3645  fail3:
3646 	efx_fini_io(efx);
3647  fail2:
3648 	efx_fini_struct(efx);
3649  fail1:
3650 	WARN_ON(rc > 0);
3651 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3652 	free_netdev(net_dev);
3653 	return rc;
3654 }
3655 
3656 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3657  * enabled on success
3658  */
3659 #ifdef CONFIG_SFC_SRIOV
3660 static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
3661 {
3662 	int rc;
3663 	struct efx_nic *efx = pci_get_drvdata(dev);
3664 
3665 	if (efx->type->sriov_configure) {
3666 		rc = efx->type->sriov_configure(efx, num_vfs);
3667 		if (rc)
3668 			return rc;
3669 		else
3670 			return num_vfs;
3671 	} else
3672 		return -EOPNOTSUPP;
3673 }
3674 #endif
3675 
3676 static int efx_pm_freeze(struct device *dev)
3677 {
3678 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3679 
3680 	rtnl_lock();
3681 
3682 	if (efx->state != STATE_DISABLED) {
3683 		efx->state = STATE_UNINIT;
3684 
3685 		efx_device_detach_sync(efx);
3686 
3687 		efx_stop_all(efx);
3688 		efx_disable_interrupts(efx);
3689 	}
3690 
3691 	rtnl_unlock();
3692 
3693 	return 0;
3694 }
3695 
3696 static int efx_pm_thaw(struct device *dev)
3697 {
3698 	int rc;
3699 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3700 
3701 	rtnl_lock();
3702 
3703 	if (efx->state != STATE_DISABLED) {
3704 		rc = efx_enable_interrupts(efx);
3705 		if (rc)
3706 			goto fail;
3707 
3708 		mutex_lock(&efx->mac_lock);
3709 		efx->phy_op->reconfigure(efx);
3710 		mutex_unlock(&efx->mac_lock);
3711 
3712 		efx_start_all(efx);
3713 
3714 		efx_device_attach_if_not_resetting(efx);
3715 
3716 		efx->state = STATE_READY;
3717 
3718 		efx->type->resume_wol(efx);
3719 	}
3720 
3721 	rtnl_unlock();
3722 
3723 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3724 	queue_work(reset_workqueue, &efx->reset_work);
3725 
3726 	return 0;
3727 
3728 fail:
3729 	rtnl_unlock();
3730 
3731 	return rc;
3732 }
3733 
3734 static int efx_pm_poweroff(struct device *dev)
3735 {
3736 	struct pci_dev *pci_dev = to_pci_dev(dev);
3737 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3738 
3739 	efx->type->fini(efx);
3740 
3741 	efx->reset_pending = 0;
3742 
3743 	pci_save_state(pci_dev);
3744 	return pci_set_power_state(pci_dev, PCI_D3hot);
3745 }
3746 
3747 /* Used for both resume and restore */
3748 static int efx_pm_resume(struct device *dev)
3749 {
3750 	struct pci_dev *pci_dev = to_pci_dev(dev);
3751 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3752 	int rc;
3753 
3754 	rc = pci_set_power_state(pci_dev, PCI_D0);
3755 	if (rc)
3756 		return rc;
3757 	pci_restore_state(pci_dev);
3758 	rc = pci_enable_device(pci_dev);
3759 	if (rc)
3760 		return rc;
3761 	pci_set_master(efx->pci_dev);
3762 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
3763 	if (rc)
3764 		return rc;
3765 	down_write(&efx->filter_sem);
3766 	rc = efx->type->init(efx);
3767 	up_write(&efx->filter_sem);
3768 	if (rc)
3769 		return rc;
3770 	rc = efx_pm_thaw(dev);
3771 	return rc;
3772 }
3773 
3774 static int efx_pm_suspend(struct device *dev)
3775 {
3776 	int rc;
3777 
3778 	efx_pm_freeze(dev);
3779 	rc = efx_pm_poweroff(dev);
3780 	if (rc)
3781 		efx_pm_resume(dev);
3782 	return rc;
3783 }
3784 
3785 static const struct dev_pm_ops efx_pm_ops = {
3786 	.suspend	= efx_pm_suspend,
3787 	.resume		= efx_pm_resume,
3788 	.freeze		= efx_pm_freeze,
3789 	.thaw		= efx_pm_thaw,
3790 	.poweroff	= efx_pm_poweroff,
3791 	.restore	= efx_pm_resume,
3792 };
3793 
3794 /* A PCI error affecting this device was detected.
3795  * At this point MMIO and DMA may be disabled.
3796  * Stop the software path and request a slot reset.
3797  */
3798 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3799 					      enum pci_channel_state state)
3800 {
3801 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3802 	struct efx_nic *efx = pci_get_drvdata(pdev);
3803 
3804 	if (state == pci_channel_io_perm_failure)
3805 		return PCI_ERS_RESULT_DISCONNECT;
3806 
3807 	rtnl_lock();
3808 
3809 	if (efx->state != STATE_DISABLED) {
3810 		efx->state = STATE_RECOVERY;
3811 		efx->reset_pending = 0;
3812 
3813 		efx_device_detach_sync(efx);
3814 
3815 		efx_stop_all(efx);
3816 		efx_disable_interrupts(efx);
3817 
3818 		status = PCI_ERS_RESULT_NEED_RESET;
3819 	} else {
3820 		/* If the interface is disabled we don't want to do anything
3821 		 * with it.
3822 		 */
3823 		status = PCI_ERS_RESULT_RECOVERED;
3824 	}
3825 
3826 	rtnl_unlock();
3827 
3828 	pci_disable_device(pdev);
3829 
3830 	return status;
3831 }
3832 
3833 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3834 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3835 {
3836 	struct efx_nic *efx = pci_get_drvdata(pdev);
3837 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3838 	int rc;
3839 
3840 	if (pci_enable_device(pdev)) {
3841 		netif_err(efx, hw, efx->net_dev,
3842 			  "Cannot re-enable PCI device after reset.\n");
3843 		status =  PCI_ERS_RESULT_DISCONNECT;
3844 	}
3845 
3846 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3847 	if (rc) {
3848 		netif_err(efx, hw, efx->net_dev,
3849 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3850 		/* Non-fatal error. Continue. */
3851 	}
3852 
3853 	return status;
3854 }
3855 
3856 /* Perform the actual reset and resume I/O operations. */
3857 static void efx_io_resume(struct pci_dev *pdev)
3858 {
3859 	struct efx_nic *efx = pci_get_drvdata(pdev);
3860 	int rc;
3861 
3862 	rtnl_lock();
3863 
3864 	if (efx->state == STATE_DISABLED)
3865 		goto out;
3866 
3867 	rc = efx_reset(efx, RESET_TYPE_ALL);
3868 	if (rc) {
3869 		netif_err(efx, hw, efx->net_dev,
3870 			  "efx_reset failed after PCI error (%d)\n", rc);
3871 	} else {
3872 		efx->state = STATE_READY;
3873 		netif_dbg(efx, hw, efx->net_dev,
3874 			  "Done resetting and resuming IO after PCI error.\n");
3875 	}
3876 
3877 out:
3878 	rtnl_unlock();
3879 }
3880 
3881 /* For simplicity and reliability, we always require a slot reset and try to
3882  * reset the hardware when a pci error affecting the device is detected.
3883  * We leave both the link_reset and mmio_enabled callback unimplemented:
3884  * with our request for slot reset the mmio_enabled callback will never be
3885  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3886  */
3887 static const struct pci_error_handlers efx_err_handlers = {
3888 	.error_detected = efx_io_error_detected,
3889 	.slot_reset	= efx_io_slot_reset,
3890 	.resume		= efx_io_resume,
3891 };
3892 
3893 static struct pci_driver efx_pci_driver = {
3894 	.name		= KBUILD_MODNAME,
3895 	.id_table	= efx_pci_table,
3896 	.probe		= efx_pci_probe,
3897 	.remove		= efx_pci_remove,
3898 	.driver.pm	= &efx_pm_ops,
3899 	.err_handler	= &efx_err_handlers,
3900 #ifdef CONFIG_SFC_SRIOV
3901 	.sriov_configure = efx_pci_sriov_configure,
3902 #endif
3903 };
3904 
3905 /**************************************************************************
3906  *
3907  * Kernel module interface
3908  *
3909  *************************************************************************/
3910 
3911 module_param(interrupt_mode, uint, 0444);
3912 MODULE_PARM_DESC(interrupt_mode,
3913 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3914 
3915 static int __init efx_init_module(void)
3916 {
3917 	int rc;
3918 
3919 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3920 
3921 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3922 	if (rc)
3923 		goto err_notifier;
3924 
3925 #ifdef CONFIG_SFC_SRIOV
3926 	rc = efx_init_sriov();
3927 	if (rc)
3928 		goto err_sriov;
3929 #endif
3930 
3931 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3932 	if (!reset_workqueue) {
3933 		rc = -ENOMEM;
3934 		goto err_reset;
3935 	}
3936 
3937 	rc = pci_register_driver(&efx_pci_driver);
3938 	if (rc < 0)
3939 		goto err_pci;
3940 
3941 	return 0;
3942 
3943  err_pci:
3944 	destroy_workqueue(reset_workqueue);
3945  err_reset:
3946 #ifdef CONFIG_SFC_SRIOV
3947 	efx_fini_sriov();
3948  err_sriov:
3949 #endif
3950 	unregister_netdevice_notifier(&efx_netdev_notifier);
3951  err_notifier:
3952 	return rc;
3953 }
3954 
3955 static void __exit efx_exit_module(void)
3956 {
3957 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3958 
3959 	pci_unregister_driver(&efx_pci_driver);
3960 	destroy_workqueue(reset_workqueue);
3961 #ifdef CONFIG_SFC_SRIOV
3962 	efx_fini_sriov();
3963 #endif
3964 	unregister_netdevice_notifier(&efx_netdev_notifier);
3965 
3966 }
3967 
3968 module_init(efx_init_module);
3969 module_exit(efx_exit_module);
3970 
3971 MODULE_AUTHOR("Solarflare Communications and "
3972 	      "Michael Brown <mbrown@fensystems.co.uk>");
3973 MODULE_DESCRIPTION("Solarflare network driver");
3974 MODULE_LICENSE("GPL");
3975 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3976 MODULE_VERSION(EFX_DRIVER_VERSION);
3977