xref: /linux/drivers/net/ethernet/sfc/efx.c (revision c33a605dd6f36618f9f658e71c09fcdcb44fc500)
1 /****************************************************************************
2  * Driver for Solarflare network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2013 Solarflare Communications Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published
8  * by the Free Software Foundation, incorporated herein by reference.
9  */
10 
11 #include <linux/module.h>
12 #include <linux/pci.h>
13 #include <linux/netdevice.h>
14 #include <linux/etherdevice.h>
15 #include <linux/delay.h>
16 #include <linux/notifier.h>
17 #include <linux/ip.h>
18 #include <linux/tcp.h>
19 #include <linux/in.h>
20 #include <linux/ethtool.h>
21 #include <linux/topology.h>
22 #include <linux/gfp.h>
23 #include <linux/aer.h>
24 #include <linux/interrupt.h>
25 #include "net_driver.h"
26 #include "efx.h"
27 #include "nic.h"
28 #include "selftest.h"
29 #include "sriov.h"
30 
31 #include "mcdi.h"
32 #include "workarounds.h"
33 
34 /**************************************************************************
35  *
36  * Type name strings
37  *
38  **************************************************************************
39  */
40 
41 /* Loopback mode names (see LOOPBACK_MODE()) */
42 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
43 const char *const efx_loopback_mode_names[] = {
44 	[LOOPBACK_NONE]		= "NONE",
45 	[LOOPBACK_DATA]		= "DATAPATH",
46 	[LOOPBACK_GMAC]		= "GMAC",
47 	[LOOPBACK_XGMII]	= "XGMII",
48 	[LOOPBACK_XGXS]		= "XGXS",
49 	[LOOPBACK_XAUI]		= "XAUI",
50 	[LOOPBACK_GMII]		= "GMII",
51 	[LOOPBACK_SGMII]	= "SGMII",
52 	[LOOPBACK_XGBR]		= "XGBR",
53 	[LOOPBACK_XFI]		= "XFI",
54 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
55 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
56 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
57 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
58 	[LOOPBACK_GPHY]		= "GPHY",
59 	[LOOPBACK_PHYXS]	= "PHYXS",
60 	[LOOPBACK_PCS]		= "PCS",
61 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
62 	[LOOPBACK_XPORT]	= "XPORT",
63 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
64 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
65 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
66 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
67 	[LOOPBACK_GMII_WS]	= "GMII_WS",
68 	[LOOPBACK_XFI_WS]	= "XFI_WS",
69 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
70 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
71 };
72 
73 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
74 const char *const efx_reset_type_names[] = {
75 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
76 	[RESET_TYPE_ALL]                = "ALL",
77 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
78 	[RESET_TYPE_WORLD]              = "WORLD",
79 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
80 	[RESET_TYPE_DATAPATH]           = "DATAPATH",
81 	[RESET_TYPE_MC_BIST]		= "MC_BIST",
82 	[RESET_TYPE_DISABLE]            = "DISABLE",
83 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
84 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
85 	[RESET_TYPE_RX_RECOVERY]        = "RX_RECOVERY",
86 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
87 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
88 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
89 	[RESET_TYPE_MCDI_TIMEOUT]	= "MCDI_TIMEOUT (FLR)",
90 };
91 
92 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
93  * queued onto this work queue. This is not a per-nic work queue, because
94  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
95  */
96 static struct workqueue_struct *reset_workqueue;
97 
98 /* How often and how many times to poll for a reset while waiting for a
99  * BIST that another function started to complete.
100  */
101 #define BIST_WAIT_DELAY_MS	100
102 #define BIST_WAIT_DELAY_COUNT	100
103 
104 /**************************************************************************
105  *
106  * Configurable values
107  *
108  *************************************************************************/
109 
110 /*
111  * Use separate channels for TX and RX events
112  *
113  * Set this to 1 to use separate channels for TX and RX. It allows us
114  * to control interrupt affinity separately for TX and RX.
115  *
116  * This is only used in MSI-X interrupt mode
117  */
118 bool efx_separate_tx_channels;
119 module_param(efx_separate_tx_channels, bool, 0444);
120 MODULE_PARM_DESC(efx_separate_tx_channels,
121 		 "Use separate channels for TX and RX");
122 
123 /* This is the weight assigned to each of the (per-channel) virtual
124  * NAPI devices.
125  */
126 static int napi_weight = 64;
127 
128 /* This is the time (in jiffies) between invocations of the hardware
129  * monitor.
130  * On Falcon-based NICs, this will:
131  * - Check the on-board hardware monitor;
132  * - Poll the link state and reconfigure the hardware as necessary.
133  * On Siena-based NICs for power systems with EEH support, this will give EEH a
134  * chance to start.
135  */
136 static unsigned int efx_monitor_interval = 1 * HZ;
137 
138 /* Initial interrupt moderation settings.  They can be modified after
139  * module load with ethtool.
140  *
141  * The default for RX should strike a balance between increasing the
142  * round-trip latency and reducing overhead.
143  */
144 static unsigned int rx_irq_mod_usec = 60;
145 
146 /* Initial interrupt moderation settings.  They can be modified after
147  * module load with ethtool.
148  *
149  * This default is chosen to ensure that a 10G link does not go idle
150  * while a TX queue is stopped after it has become full.  A queue is
151  * restarted when it drops below half full.  The time this takes (assuming
152  * worst case 3 descriptors per packet and 1024 descriptors) is
153  *   512 / 3 * 1.2 = 205 usec.
154  */
155 static unsigned int tx_irq_mod_usec = 150;
156 
157 /* This is the first interrupt mode to try out of:
158  * 0 => MSI-X
159  * 1 => MSI
160  * 2 => legacy
161  */
162 static unsigned int interrupt_mode;
163 
164 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
165  * i.e. the number of CPUs among which we may distribute simultaneous
166  * interrupt handling.
167  *
168  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
169  * The default (0) means to assign an interrupt to each core.
170  */
171 static unsigned int rss_cpus;
172 module_param(rss_cpus, uint, 0444);
173 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
174 
175 static bool phy_flash_cfg;
176 module_param(phy_flash_cfg, bool, 0644);
177 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
178 
179 static unsigned irq_adapt_low_thresh = 8000;
180 module_param(irq_adapt_low_thresh, uint, 0644);
181 MODULE_PARM_DESC(irq_adapt_low_thresh,
182 		 "Threshold score for reducing IRQ moderation");
183 
184 static unsigned irq_adapt_high_thresh = 16000;
185 module_param(irq_adapt_high_thresh, uint, 0644);
186 MODULE_PARM_DESC(irq_adapt_high_thresh,
187 		 "Threshold score for increasing IRQ moderation");
188 
189 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
190 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
191 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
192 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
193 module_param(debug, uint, 0);
194 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
195 
196 /**************************************************************************
197  *
198  * Utility functions and prototypes
199  *
200  *************************************************************************/
201 
202 static int efx_soft_enable_interrupts(struct efx_nic *efx);
203 static void efx_soft_disable_interrupts(struct efx_nic *efx);
204 static void efx_remove_channel(struct efx_channel *channel);
205 static void efx_remove_channels(struct efx_nic *efx);
206 static const struct efx_channel_type efx_default_channel_type;
207 static void efx_remove_port(struct efx_nic *efx);
208 static void efx_init_napi_channel(struct efx_channel *channel);
209 static void efx_fini_napi(struct efx_nic *efx);
210 static void efx_fini_napi_channel(struct efx_channel *channel);
211 static void efx_fini_struct(struct efx_nic *efx);
212 static void efx_start_all(struct efx_nic *efx);
213 static void efx_stop_all(struct efx_nic *efx);
214 
215 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
216 	do {						\
217 		if ((efx->state == STATE_READY) ||	\
218 		    (efx->state == STATE_RECOVERY) ||	\
219 		    (efx->state == STATE_DISABLED))	\
220 			ASSERT_RTNL();			\
221 	} while (0)
222 
223 static int efx_check_disabled(struct efx_nic *efx)
224 {
225 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
226 		netif_err(efx, drv, efx->net_dev,
227 			  "device is disabled due to earlier errors\n");
228 		return -EIO;
229 	}
230 	return 0;
231 }
232 
233 /**************************************************************************
234  *
235  * Event queue processing
236  *
237  *************************************************************************/
238 
239 /* Process channel's event queue
240  *
241  * This function is responsible for processing the event queue of a
242  * single channel.  The caller must guarantee that this function will
243  * never be concurrently called more than once on the same channel,
244  * though different channels may be being processed concurrently.
245  */
246 static int efx_process_channel(struct efx_channel *channel, int budget)
247 {
248 	struct efx_tx_queue *tx_queue;
249 	int spent;
250 
251 	if (unlikely(!channel->enabled))
252 		return 0;
253 
254 	efx_for_each_channel_tx_queue(tx_queue, channel) {
255 		tx_queue->pkts_compl = 0;
256 		tx_queue->bytes_compl = 0;
257 	}
258 
259 	spent = efx_nic_process_eventq(channel, budget);
260 	if (spent && efx_channel_has_rx_queue(channel)) {
261 		struct efx_rx_queue *rx_queue =
262 			efx_channel_get_rx_queue(channel);
263 
264 		efx_rx_flush_packet(channel);
265 		efx_fast_push_rx_descriptors(rx_queue, true);
266 	}
267 
268 	/* Update BQL */
269 	efx_for_each_channel_tx_queue(tx_queue, channel) {
270 		if (tx_queue->bytes_compl) {
271 			netdev_tx_completed_queue(tx_queue->core_txq,
272 				tx_queue->pkts_compl, tx_queue->bytes_compl);
273 		}
274 	}
275 
276 	return spent;
277 }
278 
279 /* NAPI poll handler
280  *
281  * NAPI guarantees serialisation of polls of the same device, which
282  * provides the guarantee required by efx_process_channel().
283  */
284 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel)
285 {
286 	int step = efx->irq_mod_step_us;
287 
288 	if (channel->irq_mod_score < irq_adapt_low_thresh) {
289 		if (channel->irq_moderation_us > step) {
290 			channel->irq_moderation_us -= step;
291 			efx->type->push_irq_moderation(channel);
292 		}
293 	} else if (channel->irq_mod_score > irq_adapt_high_thresh) {
294 		if (channel->irq_moderation_us <
295 		    efx->irq_rx_moderation_us) {
296 			channel->irq_moderation_us += step;
297 			efx->type->push_irq_moderation(channel);
298 		}
299 	}
300 
301 	channel->irq_count = 0;
302 	channel->irq_mod_score = 0;
303 }
304 
305 static int efx_poll(struct napi_struct *napi, int budget)
306 {
307 	struct efx_channel *channel =
308 		container_of(napi, struct efx_channel, napi_str);
309 	struct efx_nic *efx = channel->efx;
310 	int spent;
311 
312 	if (!efx_channel_lock_napi(channel))
313 		return budget;
314 
315 	netif_vdbg(efx, intr, efx->net_dev,
316 		   "channel %d NAPI poll executing on CPU %d\n",
317 		   channel->channel, raw_smp_processor_id());
318 
319 	spent = efx_process_channel(channel, budget);
320 
321 	if (spent < budget) {
322 		if (efx_channel_has_rx_queue(channel) &&
323 		    efx->irq_rx_adaptive &&
324 		    unlikely(++channel->irq_count == 1000)) {
325 			efx_update_irq_mod(efx, channel);
326 		}
327 
328 		efx_filter_rfs_expire(channel);
329 
330 		/* There is no race here; although napi_disable() will
331 		 * only wait for napi_complete(), this isn't a problem
332 		 * since efx_nic_eventq_read_ack() will have no effect if
333 		 * interrupts have already been disabled.
334 		 */
335 		napi_complete(napi);
336 		efx_nic_eventq_read_ack(channel);
337 	}
338 
339 	efx_channel_unlock_napi(channel);
340 	return spent;
341 }
342 
343 /* Create event queue
344  * Event queue memory allocations are done only once.  If the channel
345  * is reset, the memory buffer will be reused; this guards against
346  * errors during channel reset and also simplifies interrupt handling.
347  */
348 static int efx_probe_eventq(struct efx_channel *channel)
349 {
350 	struct efx_nic *efx = channel->efx;
351 	unsigned long entries;
352 
353 	netif_dbg(efx, probe, efx->net_dev,
354 		  "chan %d create event queue\n", channel->channel);
355 
356 	/* Build an event queue with room for one event per tx and rx buffer,
357 	 * plus some extra for link state events and MCDI completions. */
358 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
359 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
360 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
361 
362 	return efx_nic_probe_eventq(channel);
363 }
364 
365 /* Prepare channel's event queue */
366 static int efx_init_eventq(struct efx_channel *channel)
367 {
368 	struct efx_nic *efx = channel->efx;
369 	int rc;
370 
371 	EFX_WARN_ON_PARANOID(channel->eventq_init);
372 
373 	netif_dbg(efx, drv, efx->net_dev,
374 		  "chan %d init event queue\n", channel->channel);
375 
376 	rc = efx_nic_init_eventq(channel);
377 	if (rc == 0) {
378 		efx->type->push_irq_moderation(channel);
379 		channel->eventq_read_ptr = 0;
380 		channel->eventq_init = true;
381 	}
382 	return rc;
383 }
384 
385 /* Enable event queue processing and NAPI */
386 void efx_start_eventq(struct efx_channel *channel)
387 {
388 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
389 		  "chan %d start event queue\n", channel->channel);
390 
391 	/* Make sure the NAPI handler sees the enabled flag set */
392 	channel->enabled = true;
393 	smp_wmb();
394 
395 	efx_channel_enable(channel);
396 	napi_enable(&channel->napi_str);
397 	efx_nic_eventq_read_ack(channel);
398 }
399 
400 /* Disable event queue processing and NAPI */
401 void efx_stop_eventq(struct efx_channel *channel)
402 {
403 	if (!channel->enabled)
404 		return;
405 
406 	napi_disable(&channel->napi_str);
407 	while (!efx_channel_disable(channel))
408 		usleep_range(1000, 20000);
409 	channel->enabled = false;
410 }
411 
412 static void efx_fini_eventq(struct efx_channel *channel)
413 {
414 	if (!channel->eventq_init)
415 		return;
416 
417 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
418 		  "chan %d fini event queue\n", channel->channel);
419 
420 	efx_nic_fini_eventq(channel);
421 	channel->eventq_init = false;
422 }
423 
424 static void efx_remove_eventq(struct efx_channel *channel)
425 {
426 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
427 		  "chan %d remove event queue\n", channel->channel);
428 
429 	efx_nic_remove_eventq(channel);
430 }
431 
432 /**************************************************************************
433  *
434  * Channel handling
435  *
436  *************************************************************************/
437 
438 /* Allocate and initialise a channel structure. */
439 static struct efx_channel *
440 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
441 {
442 	struct efx_channel *channel;
443 	struct efx_rx_queue *rx_queue;
444 	struct efx_tx_queue *tx_queue;
445 	int j;
446 
447 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
448 	if (!channel)
449 		return NULL;
450 
451 	channel->efx = efx;
452 	channel->channel = i;
453 	channel->type = &efx_default_channel_type;
454 
455 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
456 		tx_queue = &channel->tx_queue[j];
457 		tx_queue->efx = efx;
458 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
459 		tx_queue->channel = channel;
460 	}
461 
462 	rx_queue = &channel->rx_queue;
463 	rx_queue->efx = efx;
464 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
465 		    (unsigned long)rx_queue);
466 
467 	return channel;
468 }
469 
470 /* Allocate and initialise a channel structure, copying parameters
471  * (but not resources) from an old channel structure.
472  */
473 static struct efx_channel *
474 efx_copy_channel(const struct efx_channel *old_channel)
475 {
476 	struct efx_channel *channel;
477 	struct efx_rx_queue *rx_queue;
478 	struct efx_tx_queue *tx_queue;
479 	int j;
480 
481 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
482 	if (!channel)
483 		return NULL;
484 
485 	*channel = *old_channel;
486 
487 	channel->napi_dev = NULL;
488 	INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
489 	channel->napi_str.napi_id = 0;
490 	channel->napi_str.state = 0;
491 	memset(&channel->eventq, 0, sizeof(channel->eventq));
492 
493 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
494 		tx_queue = &channel->tx_queue[j];
495 		if (tx_queue->channel)
496 			tx_queue->channel = channel;
497 		tx_queue->buffer = NULL;
498 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
499 	}
500 
501 	rx_queue = &channel->rx_queue;
502 	rx_queue->buffer = NULL;
503 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
504 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
505 		    (unsigned long)rx_queue);
506 
507 	return channel;
508 }
509 
510 static int efx_probe_channel(struct efx_channel *channel)
511 {
512 	struct efx_tx_queue *tx_queue;
513 	struct efx_rx_queue *rx_queue;
514 	int rc;
515 
516 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
517 		  "creating channel %d\n", channel->channel);
518 
519 	rc = channel->type->pre_probe(channel);
520 	if (rc)
521 		goto fail;
522 
523 	rc = efx_probe_eventq(channel);
524 	if (rc)
525 		goto fail;
526 
527 	efx_for_each_channel_tx_queue(tx_queue, channel) {
528 		rc = efx_probe_tx_queue(tx_queue);
529 		if (rc)
530 			goto fail;
531 	}
532 
533 	efx_for_each_channel_rx_queue(rx_queue, channel) {
534 		rc = efx_probe_rx_queue(rx_queue);
535 		if (rc)
536 			goto fail;
537 	}
538 
539 	return 0;
540 
541 fail:
542 	efx_remove_channel(channel);
543 	return rc;
544 }
545 
546 static void
547 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
548 {
549 	struct efx_nic *efx = channel->efx;
550 	const char *type;
551 	int number;
552 
553 	number = channel->channel;
554 	if (efx->tx_channel_offset == 0) {
555 		type = "";
556 	} else if (channel->channel < efx->tx_channel_offset) {
557 		type = "-rx";
558 	} else {
559 		type = "-tx";
560 		number -= efx->tx_channel_offset;
561 	}
562 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
563 }
564 
565 static void efx_set_channel_names(struct efx_nic *efx)
566 {
567 	struct efx_channel *channel;
568 
569 	efx_for_each_channel(channel, efx)
570 		channel->type->get_name(channel,
571 					efx->msi_context[channel->channel].name,
572 					sizeof(efx->msi_context[0].name));
573 }
574 
575 static int efx_probe_channels(struct efx_nic *efx)
576 {
577 	struct efx_channel *channel;
578 	int rc;
579 
580 	/* Restart special buffer allocation */
581 	efx->next_buffer_table = 0;
582 
583 	/* Probe channels in reverse, so that any 'extra' channels
584 	 * use the start of the buffer table. This allows the traffic
585 	 * channels to be resized without moving them or wasting the
586 	 * entries before them.
587 	 */
588 	efx_for_each_channel_rev(channel, efx) {
589 		rc = efx_probe_channel(channel);
590 		if (rc) {
591 			netif_err(efx, probe, efx->net_dev,
592 				  "failed to create channel %d\n",
593 				  channel->channel);
594 			goto fail;
595 		}
596 	}
597 	efx_set_channel_names(efx);
598 
599 	return 0;
600 
601 fail:
602 	efx_remove_channels(efx);
603 	return rc;
604 }
605 
606 /* Channels are shutdown and reinitialised whilst the NIC is running
607  * to propagate configuration changes (mtu, checksum offload), or
608  * to clear hardware error conditions
609  */
610 static void efx_start_datapath(struct efx_nic *efx)
611 {
612 	netdev_features_t old_features = efx->net_dev->features;
613 	bool old_rx_scatter = efx->rx_scatter;
614 	struct efx_tx_queue *tx_queue;
615 	struct efx_rx_queue *rx_queue;
616 	struct efx_channel *channel;
617 	size_t rx_buf_len;
618 
619 	/* Calculate the rx buffer allocation parameters required to
620 	 * support the current MTU, including padding for header
621 	 * alignment and overruns.
622 	 */
623 	efx->rx_dma_len = (efx->rx_prefix_size +
624 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
625 			   efx->type->rx_buffer_padding);
626 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
627 		      efx->rx_ip_align + efx->rx_dma_len);
628 	if (rx_buf_len <= PAGE_SIZE) {
629 		efx->rx_scatter = efx->type->always_rx_scatter;
630 		efx->rx_buffer_order = 0;
631 	} else if (efx->type->can_rx_scatter) {
632 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
633 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
634 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
635 				       EFX_RX_BUF_ALIGNMENT) >
636 			     PAGE_SIZE);
637 		efx->rx_scatter = true;
638 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
639 		efx->rx_buffer_order = 0;
640 	} else {
641 		efx->rx_scatter = false;
642 		efx->rx_buffer_order = get_order(rx_buf_len);
643 	}
644 
645 	efx_rx_config_page_split(efx);
646 	if (efx->rx_buffer_order)
647 		netif_dbg(efx, drv, efx->net_dev,
648 			  "RX buf len=%u; page order=%u batch=%u\n",
649 			  efx->rx_dma_len, efx->rx_buffer_order,
650 			  efx->rx_pages_per_batch);
651 	else
652 		netif_dbg(efx, drv, efx->net_dev,
653 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
654 			  efx->rx_dma_len, efx->rx_page_buf_step,
655 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
656 
657 	/* Restore previously fixed features in hw_features and remove
658 	 * features which are fixed now
659 	 */
660 	efx->net_dev->hw_features |= efx->net_dev->features;
661 	efx->net_dev->hw_features &= ~efx->fixed_features;
662 	efx->net_dev->features |= efx->fixed_features;
663 	if (efx->net_dev->features != old_features)
664 		netdev_features_change(efx->net_dev);
665 
666 	/* RX filters may also have scatter-enabled flags */
667 	if (efx->rx_scatter != old_rx_scatter)
668 		efx->type->filter_update_rx_scatter(efx);
669 
670 	/* We must keep at least one descriptor in a TX ring empty.
671 	 * We could avoid this when the queue size does not exactly
672 	 * match the hardware ring size, but it's not that important.
673 	 * Therefore we stop the queue when one more skb might fill
674 	 * the ring completely.  We wake it when half way back to
675 	 * empty.
676 	 */
677 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
678 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
679 
680 	/* Initialise the channels */
681 	efx_for_each_channel(channel, efx) {
682 		efx_for_each_channel_tx_queue(tx_queue, channel) {
683 			efx_init_tx_queue(tx_queue);
684 			atomic_inc(&efx->active_queues);
685 		}
686 
687 		efx_for_each_channel_rx_queue(rx_queue, channel) {
688 			efx_init_rx_queue(rx_queue);
689 			atomic_inc(&efx->active_queues);
690 			efx_stop_eventq(channel);
691 			efx_fast_push_rx_descriptors(rx_queue, false);
692 			efx_start_eventq(channel);
693 		}
694 
695 		WARN_ON(channel->rx_pkt_n_frags);
696 	}
697 
698 	efx_ptp_start_datapath(efx);
699 
700 	if (netif_device_present(efx->net_dev))
701 		netif_tx_wake_all_queues(efx->net_dev);
702 }
703 
704 static void efx_stop_datapath(struct efx_nic *efx)
705 {
706 	struct efx_channel *channel;
707 	struct efx_tx_queue *tx_queue;
708 	struct efx_rx_queue *rx_queue;
709 	int rc;
710 
711 	EFX_ASSERT_RESET_SERIALISED(efx);
712 	BUG_ON(efx->port_enabled);
713 
714 	efx_ptp_stop_datapath(efx);
715 
716 	/* Stop RX refill */
717 	efx_for_each_channel(channel, efx) {
718 		efx_for_each_channel_rx_queue(rx_queue, channel)
719 			rx_queue->refill_enabled = false;
720 	}
721 
722 	efx_for_each_channel(channel, efx) {
723 		/* RX packet processing is pipelined, so wait for the
724 		 * NAPI handler to complete.  At least event queue 0
725 		 * might be kept active by non-data events, so don't
726 		 * use napi_synchronize() but actually disable NAPI
727 		 * temporarily.
728 		 */
729 		if (efx_channel_has_rx_queue(channel)) {
730 			efx_stop_eventq(channel);
731 			efx_start_eventq(channel);
732 		}
733 	}
734 
735 	rc = efx->type->fini_dmaq(efx);
736 	if (rc && EFX_WORKAROUND_7803(efx)) {
737 		/* Schedule a reset to recover from the flush failure. The
738 		 * descriptor caches reference memory we're about to free,
739 		 * but falcon_reconfigure_mac_wrapper() won't reconnect
740 		 * the MACs because of the pending reset.
741 		 */
742 		netif_err(efx, drv, efx->net_dev,
743 			  "Resetting to recover from flush failure\n");
744 		efx_schedule_reset(efx, RESET_TYPE_ALL);
745 	} else if (rc) {
746 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
747 	} else {
748 		netif_dbg(efx, drv, efx->net_dev,
749 			  "successfully flushed all queues\n");
750 	}
751 
752 	efx_for_each_channel(channel, efx) {
753 		efx_for_each_channel_rx_queue(rx_queue, channel)
754 			efx_fini_rx_queue(rx_queue);
755 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
756 			efx_fini_tx_queue(tx_queue);
757 	}
758 }
759 
760 static void efx_remove_channel(struct efx_channel *channel)
761 {
762 	struct efx_tx_queue *tx_queue;
763 	struct efx_rx_queue *rx_queue;
764 
765 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
766 		  "destroy chan %d\n", channel->channel);
767 
768 	efx_for_each_channel_rx_queue(rx_queue, channel)
769 		efx_remove_rx_queue(rx_queue);
770 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
771 		efx_remove_tx_queue(tx_queue);
772 	efx_remove_eventq(channel);
773 	channel->type->post_remove(channel);
774 }
775 
776 static void efx_remove_channels(struct efx_nic *efx)
777 {
778 	struct efx_channel *channel;
779 
780 	efx_for_each_channel(channel, efx)
781 		efx_remove_channel(channel);
782 }
783 
784 int
785 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
786 {
787 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
788 	u32 old_rxq_entries, old_txq_entries;
789 	unsigned i, next_buffer_table = 0;
790 	int rc, rc2;
791 
792 	rc = efx_check_disabled(efx);
793 	if (rc)
794 		return rc;
795 
796 	/* Not all channels should be reallocated. We must avoid
797 	 * reallocating their buffer table entries.
798 	 */
799 	efx_for_each_channel(channel, efx) {
800 		struct efx_rx_queue *rx_queue;
801 		struct efx_tx_queue *tx_queue;
802 
803 		if (channel->type->copy)
804 			continue;
805 		next_buffer_table = max(next_buffer_table,
806 					channel->eventq.index +
807 					channel->eventq.entries);
808 		efx_for_each_channel_rx_queue(rx_queue, channel)
809 			next_buffer_table = max(next_buffer_table,
810 						rx_queue->rxd.index +
811 						rx_queue->rxd.entries);
812 		efx_for_each_channel_tx_queue(tx_queue, channel)
813 			next_buffer_table = max(next_buffer_table,
814 						tx_queue->txd.index +
815 						tx_queue->txd.entries);
816 	}
817 
818 	efx_device_detach_sync(efx);
819 	efx_stop_all(efx);
820 	efx_soft_disable_interrupts(efx);
821 
822 	/* Clone channels (where possible) */
823 	memset(other_channel, 0, sizeof(other_channel));
824 	for (i = 0; i < efx->n_channels; i++) {
825 		channel = efx->channel[i];
826 		if (channel->type->copy)
827 			channel = channel->type->copy(channel);
828 		if (!channel) {
829 			rc = -ENOMEM;
830 			goto out;
831 		}
832 		other_channel[i] = channel;
833 	}
834 
835 	/* Swap entry counts and channel pointers */
836 	old_rxq_entries = efx->rxq_entries;
837 	old_txq_entries = efx->txq_entries;
838 	efx->rxq_entries = rxq_entries;
839 	efx->txq_entries = txq_entries;
840 	for (i = 0; i < efx->n_channels; i++) {
841 		channel = efx->channel[i];
842 		efx->channel[i] = other_channel[i];
843 		other_channel[i] = channel;
844 	}
845 
846 	/* Restart buffer table allocation */
847 	efx->next_buffer_table = next_buffer_table;
848 
849 	for (i = 0; i < efx->n_channels; i++) {
850 		channel = efx->channel[i];
851 		if (!channel->type->copy)
852 			continue;
853 		rc = efx_probe_channel(channel);
854 		if (rc)
855 			goto rollback;
856 		efx_init_napi_channel(efx->channel[i]);
857 	}
858 
859 out:
860 	/* Destroy unused channel structures */
861 	for (i = 0; i < efx->n_channels; i++) {
862 		channel = other_channel[i];
863 		if (channel && channel->type->copy) {
864 			efx_fini_napi_channel(channel);
865 			efx_remove_channel(channel);
866 			kfree(channel);
867 		}
868 	}
869 
870 	rc2 = efx_soft_enable_interrupts(efx);
871 	if (rc2) {
872 		rc = rc ? rc : rc2;
873 		netif_err(efx, drv, efx->net_dev,
874 			  "unable to restart interrupts on channel reallocation\n");
875 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
876 	} else {
877 		efx_start_all(efx);
878 		netif_device_attach(efx->net_dev);
879 	}
880 	return rc;
881 
882 rollback:
883 	/* Swap back */
884 	efx->rxq_entries = old_rxq_entries;
885 	efx->txq_entries = old_txq_entries;
886 	for (i = 0; i < efx->n_channels; i++) {
887 		channel = efx->channel[i];
888 		efx->channel[i] = other_channel[i];
889 		other_channel[i] = channel;
890 	}
891 	goto out;
892 }
893 
894 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
895 {
896 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
897 }
898 
899 static const struct efx_channel_type efx_default_channel_type = {
900 	.pre_probe		= efx_channel_dummy_op_int,
901 	.post_remove		= efx_channel_dummy_op_void,
902 	.get_name		= efx_get_channel_name,
903 	.copy			= efx_copy_channel,
904 	.keep_eventq		= false,
905 };
906 
907 int efx_channel_dummy_op_int(struct efx_channel *channel)
908 {
909 	return 0;
910 }
911 
912 void efx_channel_dummy_op_void(struct efx_channel *channel)
913 {
914 }
915 
916 /**************************************************************************
917  *
918  * Port handling
919  *
920  **************************************************************************/
921 
922 /* This ensures that the kernel is kept informed (via
923  * netif_carrier_on/off) of the link status, and also maintains the
924  * link status's stop on the port's TX queue.
925  */
926 void efx_link_status_changed(struct efx_nic *efx)
927 {
928 	struct efx_link_state *link_state = &efx->link_state;
929 
930 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
931 	 * that no events are triggered between unregister_netdev() and the
932 	 * driver unloading. A more general condition is that NETDEV_CHANGE
933 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
934 	if (!netif_running(efx->net_dev))
935 		return;
936 
937 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
938 		efx->n_link_state_changes++;
939 
940 		if (link_state->up)
941 			netif_carrier_on(efx->net_dev);
942 		else
943 			netif_carrier_off(efx->net_dev);
944 	}
945 
946 	/* Status message for kernel log */
947 	if (link_state->up)
948 		netif_info(efx, link, efx->net_dev,
949 			   "link up at %uMbps %s-duplex (MTU %d)\n",
950 			   link_state->speed, link_state->fd ? "full" : "half",
951 			   efx->net_dev->mtu);
952 	else
953 		netif_info(efx, link, efx->net_dev, "link down\n");
954 }
955 
956 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
957 {
958 	efx->link_advertising = advertising;
959 	if (advertising) {
960 		if (advertising & ADVERTISED_Pause)
961 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
962 		else
963 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
964 		if (advertising & ADVERTISED_Asym_Pause)
965 			efx->wanted_fc ^= EFX_FC_TX;
966 	}
967 }
968 
969 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
970 {
971 	efx->wanted_fc = wanted_fc;
972 	if (efx->link_advertising) {
973 		if (wanted_fc & EFX_FC_RX)
974 			efx->link_advertising |= (ADVERTISED_Pause |
975 						  ADVERTISED_Asym_Pause);
976 		else
977 			efx->link_advertising &= ~(ADVERTISED_Pause |
978 						   ADVERTISED_Asym_Pause);
979 		if (wanted_fc & EFX_FC_TX)
980 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
981 	}
982 }
983 
984 static void efx_fini_port(struct efx_nic *efx);
985 
986 /* We assume that efx->type->reconfigure_mac will always try to sync RX
987  * filters and therefore needs to read-lock the filter table against freeing
988  */
989 void efx_mac_reconfigure(struct efx_nic *efx)
990 {
991 	down_read(&efx->filter_sem);
992 	efx->type->reconfigure_mac(efx);
993 	up_read(&efx->filter_sem);
994 }
995 
996 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
997  * the MAC appropriately. All other PHY configuration changes are pushed
998  * through phy_op->set_settings(), and pushed asynchronously to the MAC
999  * through efx_monitor().
1000  *
1001  * Callers must hold the mac_lock
1002  */
1003 int __efx_reconfigure_port(struct efx_nic *efx)
1004 {
1005 	enum efx_phy_mode phy_mode;
1006 	int rc;
1007 
1008 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
1009 
1010 	/* Disable PHY transmit in mac level loopbacks */
1011 	phy_mode = efx->phy_mode;
1012 	if (LOOPBACK_INTERNAL(efx))
1013 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
1014 	else
1015 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
1016 
1017 	rc = efx->type->reconfigure_port(efx);
1018 
1019 	if (rc)
1020 		efx->phy_mode = phy_mode;
1021 
1022 	return rc;
1023 }
1024 
1025 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1026  * disabled. */
1027 int efx_reconfigure_port(struct efx_nic *efx)
1028 {
1029 	int rc;
1030 
1031 	EFX_ASSERT_RESET_SERIALISED(efx);
1032 
1033 	mutex_lock(&efx->mac_lock);
1034 	rc = __efx_reconfigure_port(efx);
1035 	mutex_unlock(&efx->mac_lock);
1036 
1037 	return rc;
1038 }
1039 
1040 /* Asynchronous work item for changing MAC promiscuity and multicast
1041  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1042  * MAC directly. */
1043 static void efx_mac_work(struct work_struct *data)
1044 {
1045 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1046 
1047 	mutex_lock(&efx->mac_lock);
1048 	if (efx->port_enabled)
1049 		efx_mac_reconfigure(efx);
1050 	mutex_unlock(&efx->mac_lock);
1051 }
1052 
1053 static int efx_probe_port(struct efx_nic *efx)
1054 {
1055 	int rc;
1056 
1057 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1058 
1059 	if (phy_flash_cfg)
1060 		efx->phy_mode = PHY_MODE_SPECIAL;
1061 
1062 	/* Connect up MAC/PHY operations table */
1063 	rc = efx->type->probe_port(efx);
1064 	if (rc)
1065 		return rc;
1066 
1067 	/* Initialise MAC address to permanent address */
1068 	ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1069 
1070 	return 0;
1071 }
1072 
1073 static int efx_init_port(struct efx_nic *efx)
1074 {
1075 	int rc;
1076 
1077 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1078 
1079 	mutex_lock(&efx->mac_lock);
1080 
1081 	rc = efx->phy_op->init(efx);
1082 	if (rc)
1083 		goto fail1;
1084 
1085 	efx->port_initialized = true;
1086 
1087 	/* Reconfigure the MAC before creating dma queues (required for
1088 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1089 	efx_mac_reconfigure(efx);
1090 
1091 	/* Ensure the PHY advertises the correct flow control settings */
1092 	rc = efx->phy_op->reconfigure(efx);
1093 	if (rc && rc != -EPERM)
1094 		goto fail2;
1095 
1096 	mutex_unlock(&efx->mac_lock);
1097 	return 0;
1098 
1099 fail2:
1100 	efx->phy_op->fini(efx);
1101 fail1:
1102 	mutex_unlock(&efx->mac_lock);
1103 	return rc;
1104 }
1105 
1106 static void efx_start_port(struct efx_nic *efx)
1107 {
1108 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1109 	BUG_ON(efx->port_enabled);
1110 
1111 	mutex_lock(&efx->mac_lock);
1112 	efx->port_enabled = true;
1113 
1114 	/* Ensure MAC ingress/egress is enabled */
1115 	efx_mac_reconfigure(efx);
1116 
1117 	mutex_unlock(&efx->mac_lock);
1118 }
1119 
1120 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1121  * and the async self-test, wait for them to finish and prevent them
1122  * being scheduled again.  This doesn't cover online resets, which
1123  * should only be cancelled when removing the device.
1124  */
1125 static void efx_stop_port(struct efx_nic *efx)
1126 {
1127 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1128 
1129 	EFX_ASSERT_RESET_SERIALISED(efx);
1130 
1131 	mutex_lock(&efx->mac_lock);
1132 	efx->port_enabled = false;
1133 	mutex_unlock(&efx->mac_lock);
1134 
1135 	/* Serialise against efx_set_multicast_list() */
1136 	netif_addr_lock_bh(efx->net_dev);
1137 	netif_addr_unlock_bh(efx->net_dev);
1138 
1139 	cancel_delayed_work_sync(&efx->monitor_work);
1140 	efx_selftest_async_cancel(efx);
1141 	cancel_work_sync(&efx->mac_work);
1142 }
1143 
1144 static void efx_fini_port(struct efx_nic *efx)
1145 {
1146 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1147 
1148 	if (!efx->port_initialized)
1149 		return;
1150 
1151 	efx->phy_op->fini(efx);
1152 	efx->port_initialized = false;
1153 
1154 	efx->link_state.up = false;
1155 	efx_link_status_changed(efx);
1156 }
1157 
1158 static void efx_remove_port(struct efx_nic *efx)
1159 {
1160 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1161 
1162 	efx->type->remove_port(efx);
1163 }
1164 
1165 /**************************************************************************
1166  *
1167  * NIC handling
1168  *
1169  **************************************************************************/
1170 
1171 static LIST_HEAD(efx_primary_list);
1172 static LIST_HEAD(efx_unassociated_list);
1173 
1174 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1175 {
1176 	return left->type == right->type &&
1177 		left->vpd_sn && right->vpd_sn &&
1178 		!strcmp(left->vpd_sn, right->vpd_sn);
1179 }
1180 
1181 static void efx_associate(struct efx_nic *efx)
1182 {
1183 	struct efx_nic *other, *next;
1184 
1185 	if (efx->primary == efx) {
1186 		/* Adding primary function; look for secondaries */
1187 
1188 		netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1189 		list_add_tail(&efx->node, &efx_primary_list);
1190 
1191 		list_for_each_entry_safe(other, next, &efx_unassociated_list,
1192 					 node) {
1193 			if (efx_same_controller(efx, other)) {
1194 				list_del(&other->node);
1195 				netif_dbg(other, probe, other->net_dev,
1196 					  "moving to secondary list of %s %s\n",
1197 					  pci_name(efx->pci_dev),
1198 					  efx->net_dev->name);
1199 				list_add_tail(&other->node,
1200 					      &efx->secondary_list);
1201 				other->primary = efx;
1202 			}
1203 		}
1204 	} else {
1205 		/* Adding secondary function; look for primary */
1206 
1207 		list_for_each_entry(other, &efx_primary_list, node) {
1208 			if (efx_same_controller(efx, other)) {
1209 				netif_dbg(efx, probe, efx->net_dev,
1210 					  "adding to secondary list of %s %s\n",
1211 					  pci_name(other->pci_dev),
1212 					  other->net_dev->name);
1213 				list_add_tail(&efx->node,
1214 					      &other->secondary_list);
1215 				efx->primary = other;
1216 				return;
1217 			}
1218 		}
1219 
1220 		netif_dbg(efx, probe, efx->net_dev,
1221 			  "adding to unassociated list\n");
1222 		list_add_tail(&efx->node, &efx_unassociated_list);
1223 	}
1224 }
1225 
1226 static void efx_dissociate(struct efx_nic *efx)
1227 {
1228 	struct efx_nic *other, *next;
1229 
1230 	list_del(&efx->node);
1231 	efx->primary = NULL;
1232 
1233 	list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1234 		list_del(&other->node);
1235 		netif_dbg(other, probe, other->net_dev,
1236 			  "moving to unassociated list\n");
1237 		list_add_tail(&other->node, &efx_unassociated_list);
1238 		other->primary = NULL;
1239 	}
1240 }
1241 
1242 /* This configures the PCI device to enable I/O and DMA. */
1243 static int efx_init_io(struct efx_nic *efx)
1244 {
1245 	struct pci_dev *pci_dev = efx->pci_dev;
1246 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1247 	unsigned int mem_map_size = efx->type->mem_map_size(efx);
1248 	int rc, bar;
1249 
1250 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1251 
1252 	bar = efx->type->mem_bar;
1253 
1254 	rc = pci_enable_device(pci_dev);
1255 	if (rc) {
1256 		netif_err(efx, probe, efx->net_dev,
1257 			  "failed to enable PCI device\n");
1258 		goto fail1;
1259 	}
1260 
1261 	pci_set_master(pci_dev);
1262 
1263 	/* Set the PCI DMA mask.  Try all possibilities from our
1264 	 * genuine mask down to 32 bits, because some architectures
1265 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1266 	 * masks event though they reject 46 bit masks.
1267 	 */
1268 	while (dma_mask > 0x7fffffffUL) {
1269 		rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1270 		if (rc == 0)
1271 			break;
1272 		dma_mask >>= 1;
1273 	}
1274 	if (rc) {
1275 		netif_err(efx, probe, efx->net_dev,
1276 			  "could not find a suitable DMA mask\n");
1277 		goto fail2;
1278 	}
1279 	netif_dbg(efx, probe, efx->net_dev,
1280 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1281 
1282 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1283 	rc = pci_request_region(pci_dev, bar, "sfc");
1284 	if (rc) {
1285 		netif_err(efx, probe, efx->net_dev,
1286 			  "request for memory BAR failed\n");
1287 		rc = -EIO;
1288 		goto fail3;
1289 	}
1290 	efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1291 	if (!efx->membase) {
1292 		netif_err(efx, probe, efx->net_dev,
1293 			  "could not map memory BAR at %llx+%x\n",
1294 			  (unsigned long long)efx->membase_phys, mem_map_size);
1295 		rc = -ENOMEM;
1296 		goto fail4;
1297 	}
1298 	netif_dbg(efx, probe, efx->net_dev,
1299 		  "memory BAR at %llx+%x (virtual %p)\n",
1300 		  (unsigned long long)efx->membase_phys, mem_map_size,
1301 		  efx->membase);
1302 
1303 	return 0;
1304 
1305  fail4:
1306 	pci_release_region(efx->pci_dev, bar);
1307  fail3:
1308 	efx->membase_phys = 0;
1309  fail2:
1310 	pci_disable_device(efx->pci_dev);
1311  fail1:
1312 	return rc;
1313 }
1314 
1315 static void efx_fini_io(struct efx_nic *efx)
1316 {
1317 	int bar;
1318 
1319 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1320 
1321 	if (efx->membase) {
1322 		iounmap(efx->membase);
1323 		efx->membase = NULL;
1324 	}
1325 
1326 	if (efx->membase_phys) {
1327 		bar = efx->type->mem_bar;
1328 		pci_release_region(efx->pci_dev, bar);
1329 		efx->membase_phys = 0;
1330 	}
1331 
1332 	/* Don't disable bus-mastering if VFs are assigned */
1333 	if (!pci_vfs_assigned(efx->pci_dev))
1334 		pci_disable_device(efx->pci_dev);
1335 }
1336 
1337 void efx_set_default_rx_indir_table(struct efx_nic *efx)
1338 {
1339 	size_t i;
1340 
1341 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1342 		efx->rx_indir_table[i] =
1343 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1344 }
1345 
1346 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1347 {
1348 	cpumask_var_t thread_mask;
1349 	unsigned int count;
1350 	int cpu;
1351 
1352 	if (rss_cpus) {
1353 		count = rss_cpus;
1354 	} else {
1355 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1356 			netif_warn(efx, probe, efx->net_dev,
1357 				   "RSS disabled due to allocation failure\n");
1358 			return 1;
1359 		}
1360 
1361 		count = 0;
1362 		for_each_online_cpu(cpu) {
1363 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1364 				++count;
1365 				cpumask_or(thread_mask, thread_mask,
1366 					   topology_sibling_cpumask(cpu));
1367 			}
1368 		}
1369 
1370 		free_cpumask_var(thread_mask);
1371 	}
1372 
1373 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1374 	 * table entries that are inaccessible to VFs
1375 	 */
1376 #ifdef CONFIG_SFC_SRIOV
1377 	if (efx->type->sriov_wanted) {
1378 		if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1379 		    count > efx_vf_size(efx)) {
1380 			netif_warn(efx, probe, efx->net_dev,
1381 				   "Reducing number of RSS channels from %u to %u for "
1382 				   "VF support. Increase vf-msix-limit to use more "
1383 				   "channels on the PF.\n",
1384 				   count, efx_vf_size(efx));
1385 			count = efx_vf_size(efx);
1386 		}
1387 	}
1388 #endif
1389 
1390 	return count;
1391 }
1392 
1393 /* Probe the number and type of interrupts we are able to obtain, and
1394  * the resulting numbers of channels and RX queues.
1395  */
1396 static int efx_probe_interrupts(struct efx_nic *efx)
1397 {
1398 	unsigned int extra_channels = 0;
1399 	unsigned int i, j;
1400 	int rc;
1401 
1402 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1403 		if (efx->extra_channel_type[i])
1404 			++extra_channels;
1405 
1406 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1407 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1408 		unsigned int n_channels;
1409 
1410 		n_channels = efx_wanted_parallelism(efx);
1411 		if (efx_separate_tx_channels)
1412 			n_channels *= 2;
1413 		n_channels += extra_channels;
1414 		n_channels = min(n_channels, efx->max_channels);
1415 
1416 		for (i = 0; i < n_channels; i++)
1417 			xentries[i].entry = i;
1418 		rc = pci_enable_msix_range(efx->pci_dev,
1419 					   xentries, 1, n_channels);
1420 		if (rc < 0) {
1421 			/* Fall back to single channel MSI */
1422 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1423 			netif_err(efx, drv, efx->net_dev,
1424 				  "could not enable MSI-X\n");
1425 		} else if (rc < n_channels) {
1426 			netif_err(efx, drv, efx->net_dev,
1427 				  "WARNING: Insufficient MSI-X vectors"
1428 				  " available (%d < %u).\n", rc, n_channels);
1429 			netif_err(efx, drv, efx->net_dev,
1430 				  "WARNING: Performance may be reduced.\n");
1431 			n_channels = rc;
1432 		}
1433 
1434 		if (rc > 0) {
1435 			efx->n_channels = n_channels;
1436 			if (n_channels > extra_channels)
1437 				n_channels -= extra_channels;
1438 			if (efx_separate_tx_channels) {
1439 				efx->n_tx_channels = min(max(n_channels / 2,
1440 							     1U),
1441 							 efx->max_tx_channels);
1442 				efx->n_rx_channels = max(n_channels -
1443 							 efx->n_tx_channels,
1444 							 1U);
1445 			} else {
1446 				efx->n_tx_channels = min(n_channels,
1447 							 efx->max_tx_channels);
1448 				efx->n_rx_channels = n_channels;
1449 			}
1450 			for (i = 0; i < efx->n_channels; i++)
1451 				efx_get_channel(efx, i)->irq =
1452 					xentries[i].vector;
1453 		}
1454 	}
1455 
1456 	/* Try single interrupt MSI */
1457 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1458 		efx->n_channels = 1;
1459 		efx->n_rx_channels = 1;
1460 		efx->n_tx_channels = 1;
1461 		rc = pci_enable_msi(efx->pci_dev);
1462 		if (rc == 0) {
1463 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1464 		} else {
1465 			netif_err(efx, drv, efx->net_dev,
1466 				  "could not enable MSI\n");
1467 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1468 		}
1469 	}
1470 
1471 	/* Assume legacy interrupts */
1472 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1473 		efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
1474 		efx->n_rx_channels = 1;
1475 		efx->n_tx_channels = 1;
1476 		efx->legacy_irq = efx->pci_dev->irq;
1477 	}
1478 
1479 	/* Assign extra channels if possible */
1480 	j = efx->n_channels;
1481 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1482 		if (!efx->extra_channel_type[i])
1483 			continue;
1484 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1485 		    efx->n_channels <= extra_channels) {
1486 			efx->extra_channel_type[i]->handle_no_channel(efx);
1487 		} else {
1488 			--j;
1489 			efx_get_channel(efx, j)->type =
1490 				efx->extra_channel_type[i];
1491 		}
1492 	}
1493 
1494 	/* RSS might be usable on VFs even if it is disabled on the PF */
1495 #ifdef CONFIG_SFC_SRIOV
1496 	if (efx->type->sriov_wanted) {
1497 		efx->rss_spread = ((efx->n_rx_channels > 1 ||
1498 				    !efx->type->sriov_wanted(efx)) ?
1499 				   efx->n_rx_channels : efx_vf_size(efx));
1500 		return 0;
1501 	}
1502 #endif
1503 	efx->rss_spread = efx->n_rx_channels;
1504 
1505 	return 0;
1506 }
1507 
1508 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1509 {
1510 	struct efx_channel *channel, *end_channel;
1511 	int rc;
1512 
1513 	BUG_ON(efx->state == STATE_DISABLED);
1514 
1515 	efx->irq_soft_enabled = true;
1516 	smp_wmb();
1517 
1518 	efx_for_each_channel(channel, efx) {
1519 		if (!channel->type->keep_eventq) {
1520 			rc = efx_init_eventq(channel);
1521 			if (rc)
1522 				goto fail;
1523 		}
1524 		efx_start_eventq(channel);
1525 	}
1526 
1527 	efx_mcdi_mode_event(efx);
1528 
1529 	return 0;
1530 fail:
1531 	end_channel = channel;
1532 	efx_for_each_channel(channel, efx) {
1533 		if (channel == end_channel)
1534 			break;
1535 		efx_stop_eventq(channel);
1536 		if (!channel->type->keep_eventq)
1537 			efx_fini_eventq(channel);
1538 	}
1539 
1540 	return rc;
1541 }
1542 
1543 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1544 {
1545 	struct efx_channel *channel;
1546 
1547 	if (efx->state == STATE_DISABLED)
1548 		return;
1549 
1550 	efx_mcdi_mode_poll(efx);
1551 
1552 	efx->irq_soft_enabled = false;
1553 	smp_wmb();
1554 
1555 	if (efx->legacy_irq)
1556 		synchronize_irq(efx->legacy_irq);
1557 
1558 	efx_for_each_channel(channel, efx) {
1559 		if (channel->irq)
1560 			synchronize_irq(channel->irq);
1561 
1562 		efx_stop_eventq(channel);
1563 		if (!channel->type->keep_eventq)
1564 			efx_fini_eventq(channel);
1565 	}
1566 
1567 	/* Flush the asynchronous MCDI request queue */
1568 	efx_mcdi_flush_async(efx);
1569 }
1570 
1571 static int efx_enable_interrupts(struct efx_nic *efx)
1572 {
1573 	struct efx_channel *channel, *end_channel;
1574 	int rc;
1575 
1576 	BUG_ON(efx->state == STATE_DISABLED);
1577 
1578 	if (efx->eeh_disabled_legacy_irq) {
1579 		enable_irq(efx->legacy_irq);
1580 		efx->eeh_disabled_legacy_irq = false;
1581 	}
1582 
1583 	efx->type->irq_enable_master(efx);
1584 
1585 	efx_for_each_channel(channel, efx) {
1586 		if (channel->type->keep_eventq) {
1587 			rc = efx_init_eventq(channel);
1588 			if (rc)
1589 				goto fail;
1590 		}
1591 	}
1592 
1593 	rc = efx_soft_enable_interrupts(efx);
1594 	if (rc)
1595 		goto fail;
1596 
1597 	return 0;
1598 
1599 fail:
1600 	end_channel = channel;
1601 	efx_for_each_channel(channel, efx) {
1602 		if (channel == end_channel)
1603 			break;
1604 		if (channel->type->keep_eventq)
1605 			efx_fini_eventq(channel);
1606 	}
1607 
1608 	efx->type->irq_disable_non_ev(efx);
1609 
1610 	return rc;
1611 }
1612 
1613 static void efx_disable_interrupts(struct efx_nic *efx)
1614 {
1615 	struct efx_channel *channel;
1616 
1617 	efx_soft_disable_interrupts(efx);
1618 
1619 	efx_for_each_channel(channel, efx) {
1620 		if (channel->type->keep_eventq)
1621 			efx_fini_eventq(channel);
1622 	}
1623 
1624 	efx->type->irq_disable_non_ev(efx);
1625 }
1626 
1627 static void efx_remove_interrupts(struct efx_nic *efx)
1628 {
1629 	struct efx_channel *channel;
1630 
1631 	/* Remove MSI/MSI-X interrupts */
1632 	efx_for_each_channel(channel, efx)
1633 		channel->irq = 0;
1634 	pci_disable_msi(efx->pci_dev);
1635 	pci_disable_msix(efx->pci_dev);
1636 
1637 	/* Remove legacy interrupt */
1638 	efx->legacy_irq = 0;
1639 }
1640 
1641 static void efx_set_channels(struct efx_nic *efx)
1642 {
1643 	struct efx_channel *channel;
1644 	struct efx_tx_queue *tx_queue;
1645 
1646 	efx->tx_channel_offset =
1647 		efx_separate_tx_channels ?
1648 		efx->n_channels - efx->n_tx_channels : 0;
1649 
1650 	/* We need to mark which channels really have RX and TX
1651 	 * queues, and adjust the TX queue numbers if we have separate
1652 	 * RX-only and TX-only channels.
1653 	 */
1654 	efx_for_each_channel(channel, efx) {
1655 		if (channel->channel < efx->n_rx_channels)
1656 			channel->rx_queue.core_index = channel->channel;
1657 		else
1658 			channel->rx_queue.core_index = -1;
1659 
1660 		efx_for_each_channel_tx_queue(tx_queue, channel)
1661 			tx_queue->queue -= (efx->tx_channel_offset *
1662 					    EFX_TXQ_TYPES);
1663 	}
1664 }
1665 
1666 static int efx_probe_nic(struct efx_nic *efx)
1667 {
1668 	int rc;
1669 
1670 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1671 
1672 	/* Carry out hardware-type specific initialisation */
1673 	rc = efx->type->probe(efx);
1674 	if (rc)
1675 		return rc;
1676 
1677 	do {
1678 		if (!efx->max_channels || !efx->max_tx_channels) {
1679 			netif_err(efx, drv, efx->net_dev,
1680 				  "Insufficient resources to allocate"
1681 				  " any channels\n");
1682 			rc = -ENOSPC;
1683 			goto fail1;
1684 		}
1685 
1686 		/* Determine the number of channels and queues by trying
1687 		 * to hook in MSI-X interrupts.
1688 		 */
1689 		rc = efx_probe_interrupts(efx);
1690 		if (rc)
1691 			goto fail1;
1692 
1693 		efx_set_channels(efx);
1694 
1695 		/* dimension_resources can fail with EAGAIN */
1696 		rc = efx->type->dimension_resources(efx);
1697 		if (rc != 0 && rc != -EAGAIN)
1698 			goto fail2;
1699 
1700 		if (rc == -EAGAIN)
1701 			/* try again with new max_channels */
1702 			efx_remove_interrupts(efx);
1703 
1704 	} while (rc == -EAGAIN);
1705 
1706 	if (efx->n_channels > 1)
1707 		netdev_rss_key_fill(&efx->rx_hash_key,
1708 				    sizeof(efx->rx_hash_key));
1709 	efx_set_default_rx_indir_table(efx);
1710 
1711 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1712 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1713 
1714 	/* Initialise the interrupt moderation settings */
1715 	efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
1716 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1717 				true);
1718 
1719 	return 0;
1720 
1721 fail2:
1722 	efx_remove_interrupts(efx);
1723 fail1:
1724 	efx->type->remove(efx);
1725 	return rc;
1726 }
1727 
1728 static void efx_remove_nic(struct efx_nic *efx)
1729 {
1730 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1731 
1732 	efx_remove_interrupts(efx);
1733 	efx->type->remove(efx);
1734 }
1735 
1736 static int efx_probe_filters(struct efx_nic *efx)
1737 {
1738 	int rc;
1739 
1740 	spin_lock_init(&efx->filter_lock);
1741 	init_rwsem(&efx->filter_sem);
1742 	mutex_lock(&efx->mac_lock);
1743 	down_write(&efx->filter_sem);
1744 	rc = efx->type->filter_table_probe(efx);
1745 	if (rc)
1746 		goto out_unlock;
1747 
1748 #ifdef CONFIG_RFS_ACCEL
1749 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
1750 		struct efx_channel *channel;
1751 		int i, success = 1;
1752 
1753 		efx_for_each_channel(channel, efx) {
1754 			channel->rps_flow_id =
1755 				kcalloc(efx->type->max_rx_ip_filters,
1756 					sizeof(*channel->rps_flow_id),
1757 					GFP_KERNEL);
1758 			if (!channel->rps_flow_id)
1759 				success = 0;
1760 			else
1761 				for (i = 0;
1762 				     i < efx->type->max_rx_ip_filters;
1763 				     ++i)
1764 					channel->rps_flow_id[i] =
1765 						RPS_FLOW_ID_INVALID;
1766 		}
1767 
1768 		if (!success) {
1769 			efx_for_each_channel(channel, efx)
1770 				kfree(channel->rps_flow_id);
1771 			efx->type->filter_table_remove(efx);
1772 			rc = -ENOMEM;
1773 			goto out_unlock;
1774 		}
1775 
1776 		efx->rps_expire_index = efx->rps_expire_channel = 0;
1777 	}
1778 #endif
1779 out_unlock:
1780 	up_write(&efx->filter_sem);
1781 	mutex_unlock(&efx->mac_lock);
1782 	return rc;
1783 }
1784 
1785 static void efx_remove_filters(struct efx_nic *efx)
1786 {
1787 #ifdef CONFIG_RFS_ACCEL
1788 	struct efx_channel *channel;
1789 
1790 	efx_for_each_channel(channel, efx)
1791 		kfree(channel->rps_flow_id);
1792 #endif
1793 	down_write(&efx->filter_sem);
1794 	efx->type->filter_table_remove(efx);
1795 	up_write(&efx->filter_sem);
1796 }
1797 
1798 static void efx_restore_filters(struct efx_nic *efx)
1799 {
1800 	down_read(&efx->filter_sem);
1801 	efx->type->filter_table_restore(efx);
1802 	up_read(&efx->filter_sem);
1803 }
1804 
1805 /**************************************************************************
1806  *
1807  * NIC startup/shutdown
1808  *
1809  *************************************************************************/
1810 
1811 static int efx_probe_all(struct efx_nic *efx)
1812 {
1813 	int rc;
1814 
1815 	rc = efx_probe_nic(efx);
1816 	if (rc) {
1817 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1818 		goto fail1;
1819 	}
1820 
1821 	rc = efx_probe_port(efx);
1822 	if (rc) {
1823 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1824 		goto fail2;
1825 	}
1826 
1827 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1828 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1829 		rc = -EINVAL;
1830 		goto fail3;
1831 	}
1832 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1833 
1834 #ifdef CONFIG_SFC_SRIOV
1835 	rc = efx->type->vswitching_probe(efx);
1836 	if (rc) /* not fatal; the PF will still work fine */
1837 		netif_warn(efx, probe, efx->net_dev,
1838 			   "failed to setup vswitching rc=%d;"
1839 			   " VFs may not function\n", rc);
1840 #endif
1841 
1842 	rc = efx_probe_filters(efx);
1843 	if (rc) {
1844 		netif_err(efx, probe, efx->net_dev,
1845 			  "failed to create filter tables\n");
1846 		goto fail4;
1847 	}
1848 
1849 	rc = efx_probe_channels(efx);
1850 	if (rc)
1851 		goto fail5;
1852 
1853 	return 0;
1854 
1855  fail5:
1856 	efx_remove_filters(efx);
1857  fail4:
1858 #ifdef CONFIG_SFC_SRIOV
1859 	efx->type->vswitching_remove(efx);
1860 #endif
1861  fail3:
1862 	efx_remove_port(efx);
1863  fail2:
1864 	efx_remove_nic(efx);
1865  fail1:
1866 	return rc;
1867 }
1868 
1869 /* If the interface is supposed to be running but is not, start
1870  * the hardware and software data path, regular activity for the port
1871  * (MAC statistics, link polling, etc.) and schedule the port to be
1872  * reconfigured.  Interrupts must already be enabled.  This function
1873  * is safe to call multiple times, so long as the NIC is not disabled.
1874  * Requires the RTNL lock.
1875  */
1876 static void efx_start_all(struct efx_nic *efx)
1877 {
1878 	EFX_ASSERT_RESET_SERIALISED(efx);
1879 	BUG_ON(efx->state == STATE_DISABLED);
1880 
1881 	/* Check that it is appropriate to restart the interface. All
1882 	 * of these flags are safe to read under just the rtnl lock */
1883 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
1884 	    efx->reset_pending)
1885 		return;
1886 
1887 	efx_start_port(efx);
1888 	efx_start_datapath(efx);
1889 
1890 	/* Start the hardware monitor if there is one */
1891 	if (efx->type->monitor != NULL)
1892 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1893 				   efx_monitor_interval);
1894 
1895 	/* If link state detection is normally event-driven, we have
1896 	 * to poll now because we could have missed a change
1897 	 */
1898 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1899 		mutex_lock(&efx->mac_lock);
1900 		if (efx->phy_op->poll(efx))
1901 			efx_link_status_changed(efx);
1902 		mutex_unlock(&efx->mac_lock);
1903 	}
1904 
1905 	efx->type->start_stats(efx);
1906 	efx->type->pull_stats(efx);
1907 	spin_lock_bh(&efx->stats_lock);
1908 	efx->type->update_stats(efx, NULL, NULL);
1909 	spin_unlock_bh(&efx->stats_lock);
1910 }
1911 
1912 /* Quiesce the hardware and software data path, and regular activity
1913  * for the port without bringing the link down.  Safe to call multiple
1914  * times with the NIC in almost any state, but interrupts should be
1915  * enabled.  Requires the RTNL lock.
1916  */
1917 static void efx_stop_all(struct efx_nic *efx)
1918 {
1919 	EFX_ASSERT_RESET_SERIALISED(efx);
1920 
1921 	/* port_enabled can be read safely under the rtnl lock */
1922 	if (!efx->port_enabled)
1923 		return;
1924 
1925 	/* update stats before we go down so we can accurately count
1926 	 * rx_nodesc_drops
1927 	 */
1928 	efx->type->pull_stats(efx);
1929 	spin_lock_bh(&efx->stats_lock);
1930 	efx->type->update_stats(efx, NULL, NULL);
1931 	spin_unlock_bh(&efx->stats_lock);
1932 	efx->type->stop_stats(efx);
1933 	efx_stop_port(efx);
1934 
1935 	/* Stop the kernel transmit interface.  This is only valid if
1936 	 * the device is stopped or detached; otherwise the watchdog
1937 	 * may fire immediately.
1938 	 */
1939 	WARN_ON(netif_running(efx->net_dev) &&
1940 		netif_device_present(efx->net_dev));
1941 	netif_tx_disable(efx->net_dev);
1942 
1943 	efx_stop_datapath(efx);
1944 }
1945 
1946 static void efx_remove_all(struct efx_nic *efx)
1947 {
1948 	efx_remove_channels(efx);
1949 	efx_remove_filters(efx);
1950 #ifdef CONFIG_SFC_SRIOV
1951 	efx->type->vswitching_remove(efx);
1952 #endif
1953 	efx_remove_port(efx);
1954 	efx_remove_nic(efx);
1955 }
1956 
1957 /**************************************************************************
1958  *
1959  * Interrupt moderation
1960  *
1961  **************************************************************************/
1962 unsigned int efx_usecs_to_ticks(struct efx_nic *efx, unsigned int usecs)
1963 {
1964 	if (usecs == 0)
1965 		return 0;
1966 	if (usecs * 1000 < efx->timer_quantum_ns)
1967 		return 1; /* never round down to 0 */
1968 	return usecs * 1000 / efx->timer_quantum_ns;
1969 }
1970 
1971 unsigned int efx_ticks_to_usecs(struct efx_nic *efx, unsigned int ticks)
1972 {
1973 	/* We must round up when converting ticks to microseconds
1974 	 * because we round down when converting the other way.
1975 	 */
1976 	return DIV_ROUND_UP(ticks * efx->timer_quantum_ns, 1000);
1977 }
1978 
1979 /* Set interrupt moderation parameters */
1980 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1981 			    unsigned int rx_usecs, bool rx_adaptive,
1982 			    bool rx_may_override_tx)
1983 {
1984 	struct efx_channel *channel;
1985 	unsigned int timer_max_us;
1986 
1987 	EFX_ASSERT_RESET_SERIALISED(efx);
1988 
1989 	timer_max_us = efx->timer_max_ns / 1000;
1990 
1991 	if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
1992 		return -EINVAL;
1993 
1994 	if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
1995 	    !rx_may_override_tx) {
1996 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1997 			  "RX and TX IRQ moderation must be equal\n");
1998 		return -EINVAL;
1999 	}
2000 
2001 	efx->irq_rx_adaptive = rx_adaptive;
2002 	efx->irq_rx_moderation_us = rx_usecs;
2003 	efx_for_each_channel(channel, efx) {
2004 		if (efx_channel_has_rx_queue(channel))
2005 			channel->irq_moderation_us = rx_usecs;
2006 		else if (efx_channel_has_tx_queues(channel))
2007 			channel->irq_moderation_us = tx_usecs;
2008 	}
2009 
2010 	return 0;
2011 }
2012 
2013 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
2014 			    unsigned int *rx_usecs, bool *rx_adaptive)
2015 {
2016 	*rx_adaptive = efx->irq_rx_adaptive;
2017 	*rx_usecs = efx->irq_rx_moderation_us;
2018 
2019 	/* If channels are shared between RX and TX, so is IRQ
2020 	 * moderation.  Otherwise, IRQ moderation is the same for all
2021 	 * TX channels and is not adaptive.
2022 	 */
2023 	if (efx->tx_channel_offset == 0) {
2024 		*tx_usecs = *rx_usecs;
2025 	} else {
2026 		struct efx_channel *tx_channel;
2027 
2028 		tx_channel = efx->channel[efx->tx_channel_offset];
2029 		*tx_usecs = tx_channel->irq_moderation_us;
2030 	}
2031 }
2032 
2033 /**************************************************************************
2034  *
2035  * Hardware monitor
2036  *
2037  **************************************************************************/
2038 
2039 /* Run periodically off the general workqueue */
2040 static void efx_monitor(struct work_struct *data)
2041 {
2042 	struct efx_nic *efx = container_of(data, struct efx_nic,
2043 					   monitor_work.work);
2044 
2045 	netif_vdbg(efx, timer, efx->net_dev,
2046 		   "hardware monitor executing on CPU %d\n",
2047 		   raw_smp_processor_id());
2048 	BUG_ON(efx->type->monitor == NULL);
2049 
2050 	/* If the mac_lock is already held then it is likely a port
2051 	 * reconfiguration is already in place, which will likely do
2052 	 * most of the work of monitor() anyway. */
2053 	if (mutex_trylock(&efx->mac_lock)) {
2054 		if (efx->port_enabled)
2055 			efx->type->monitor(efx);
2056 		mutex_unlock(&efx->mac_lock);
2057 	}
2058 
2059 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
2060 			   efx_monitor_interval);
2061 }
2062 
2063 /**************************************************************************
2064  *
2065  * ioctls
2066  *
2067  *************************************************************************/
2068 
2069 /* Net device ioctl
2070  * Context: process, rtnl_lock() held.
2071  */
2072 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
2073 {
2074 	struct efx_nic *efx = netdev_priv(net_dev);
2075 	struct mii_ioctl_data *data = if_mii(ifr);
2076 
2077 	if (cmd == SIOCSHWTSTAMP)
2078 		return efx_ptp_set_ts_config(efx, ifr);
2079 	if (cmd == SIOCGHWTSTAMP)
2080 		return efx_ptp_get_ts_config(efx, ifr);
2081 
2082 	/* Convert phy_id from older PRTAD/DEVAD format */
2083 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2084 	    (data->phy_id & 0xfc00) == 0x0400)
2085 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2086 
2087 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
2088 }
2089 
2090 /**************************************************************************
2091  *
2092  * NAPI interface
2093  *
2094  **************************************************************************/
2095 
2096 static void efx_init_napi_channel(struct efx_channel *channel)
2097 {
2098 	struct efx_nic *efx = channel->efx;
2099 
2100 	channel->napi_dev = efx->net_dev;
2101 	netif_napi_add(channel->napi_dev, &channel->napi_str,
2102 		       efx_poll, napi_weight);
2103 	efx_channel_busy_poll_init(channel);
2104 }
2105 
2106 static void efx_init_napi(struct efx_nic *efx)
2107 {
2108 	struct efx_channel *channel;
2109 
2110 	efx_for_each_channel(channel, efx)
2111 		efx_init_napi_channel(channel);
2112 }
2113 
2114 static void efx_fini_napi_channel(struct efx_channel *channel)
2115 {
2116 	if (channel->napi_dev) {
2117 		netif_napi_del(&channel->napi_str);
2118 		napi_hash_del(&channel->napi_str);
2119 	}
2120 	channel->napi_dev = NULL;
2121 }
2122 
2123 static void efx_fini_napi(struct efx_nic *efx)
2124 {
2125 	struct efx_channel *channel;
2126 
2127 	efx_for_each_channel(channel, efx)
2128 		efx_fini_napi_channel(channel);
2129 }
2130 
2131 /**************************************************************************
2132  *
2133  * Kernel netpoll interface
2134  *
2135  *************************************************************************/
2136 
2137 #ifdef CONFIG_NET_POLL_CONTROLLER
2138 
2139 /* Although in the common case interrupts will be disabled, this is not
2140  * guaranteed. However, all our work happens inside the NAPI callback,
2141  * so no locking is required.
2142  */
2143 static void efx_netpoll(struct net_device *net_dev)
2144 {
2145 	struct efx_nic *efx = netdev_priv(net_dev);
2146 	struct efx_channel *channel;
2147 
2148 	efx_for_each_channel(channel, efx)
2149 		efx_schedule_channel(channel);
2150 }
2151 
2152 #endif
2153 
2154 #ifdef CONFIG_NET_RX_BUSY_POLL
2155 static int efx_busy_poll(struct napi_struct *napi)
2156 {
2157 	struct efx_channel *channel =
2158 		container_of(napi, struct efx_channel, napi_str);
2159 	struct efx_nic *efx = channel->efx;
2160 	int budget = 4;
2161 	int old_rx_packets, rx_packets;
2162 
2163 	if (!netif_running(efx->net_dev))
2164 		return LL_FLUSH_FAILED;
2165 
2166 	if (!efx_channel_try_lock_poll(channel))
2167 		return LL_FLUSH_BUSY;
2168 
2169 	old_rx_packets = channel->rx_queue.rx_packets;
2170 	efx_process_channel(channel, budget);
2171 
2172 	rx_packets = channel->rx_queue.rx_packets - old_rx_packets;
2173 
2174 	/* There is no race condition with NAPI here.
2175 	 * NAPI will automatically be rescheduled if it yielded during busy
2176 	 * polling, because it was not able to take the lock and thus returned
2177 	 * the full budget.
2178 	 */
2179 	efx_channel_unlock_poll(channel);
2180 
2181 	return rx_packets;
2182 }
2183 #endif
2184 
2185 /**************************************************************************
2186  *
2187  * Kernel net device interface
2188  *
2189  *************************************************************************/
2190 
2191 /* Context: process, rtnl_lock() held. */
2192 int efx_net_open(struct net_device *net_dev)
2193 {
2194 	struct efx_nic *efx = netdev_priv(net_dev);
2195 	int rc;
2196 
2197 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2198 		  raw_smp_processor_id());
2199 
2200 	rc = efx_check_disabled(efx);
2201 	if (rc)
2202 		return rc;
2203 	if (efx->phy_mode & PHY_MODE_SPECIAL)
2204 		return -EBUSY;
2205 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2206 		return -EIO;
2207 
2208 	/* Notify the kernel of the link state polled during driver load,
2209 	 * before the monitor starts running */
2210 	efx_link_status_changed(efx);
2211 
2212 	efx_start_all(efx);
2213 	efx_selftest_async_start(efx);
2214 	return 0;
2215 }
2216 
2217 /* Context: process, rtnl_lock() held.
2218  * Note that the kernel will ignore our return code; this method
2219  * should really be a void.
2220  */
2221 int efx_net_stop(struct net_device *net_dev)
2222 {
2223 	struct efx_nic *efx = netdev_priv(net_dev);
2224 
2225 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2226 		  raw_smp_processor_id());
2227 
2228 	/* Stop the device and flush all the channels */
2229 	efx_stop_all(efx);
2230 
2231 	return 0;
2232 }
2233 
2234 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2235 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
2236 					       struct rtnl_link_stats64 *stats)
2237 {
2238 	struct efx_nic *efx = netdev_priv(net_dev);
2239 
2240 	spin_lock_bh(&efx->stats_lock);
2241 	efx->type->update_stats(efx, NULL, stats);
2242 	spin_unlock_bh(&efx->stats_lock);
2243 
2244 	return stats;
2245 }
2246 
2247 /* Context: netif_tx_lock held, BHs disabled. */
2248 static void efx_watchdog(struct net_device *net_dev)
2249 {
2250 	struct efx_nic *efx = netdev_priv(net_dev);
2251 
2252 	netif_err(efx, tx_err, efx->net_dev,
2253 		  "TX stuck with port_enabled=%d: resetting channels\n",
2254 		  efx->port_enabled);
2255 
2256 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2257 }
2258 
2259 
2260 /* Context: process, rtnl_lock() held. */
2261 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2262 {
2263 	struct efx_nic *efx = netdev_priv(net_dev);
2264 	int rc;
2265 
2266 	rc = efx_check_disabled(efx);
2267 	if (rc)
2268 		return rc;
2269 	if (new_mtu > EFX_MAX_MTU) {
2270 		netif_err(efx, drv, efx->net_dev,
2271 			  "Requested MTU of %d too big (max: %d)\n",
2272 			  new_mtu, EFX_MAX_MTU);
2273 		return -EINVAL;
2274 	}
2275 	if (new_mtu < EFX_MIN_MTU) {
2276 		netif_err(efx, drv, efx->net_dev,
2277 			  "Requested MTU of %d too small (min: %d)\n",
2278 			  new_mtu, EFX_MIN_MTU);
2279 		return -EINVAL;
2280 	}
2281 
2282 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2283 
2284 	efx_device_detach_sync(efx);
2285 	efx_stop_all(efx);
2286 
2287 	mutex_lock(&efx->mac_lock);
2288 	net_dev->mtu = new_mtu;
2289 	efx_mac_reconfigure(efx);
2290 	mutex_unlock(&efx->mac_lock);
2291 
2292 	efx_start_all(efx);
2293 	netif_device_attach(efx->net_dev);
2294 	return 0;
2295 }
2296 
2297 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2298 {
2299 	struct efx_nic *efx = netdev_priv(net_dev);
2300 	struct sockaddr *addr = data;
2301 	u8 *new_addr = addr->sa_data;
2302 	u8 old_addr[6];
2303 	int rc;
2304 
2305 	if (!is_valid_ether_addr(new_addr)) {
2306 		netif_err(efx, drv, efx->net_dev,
2307 			  "invalid ethernet MAC address requested: %pM\n",
2308 			  new_addr);
2309 		return -EADDRNOTAVAIL;
2310 	}
2311 
2312 	/* save old address */
2313 	ether_addr_copy(old_addr, net_dev->dev_addr);
2314 	ether_addr_copy(net_dev->dev_addr, new_addr);
2315 	if (efx->type->set_mac_address) {
2316 		rc = efx->type->set_mac_address(efx);
2317 		if (rc) {
2318 			ether_addr_copy(net_dev->dev_addr, old_addr);
2319 			return rc;
2320 		}
2321 	}
2322 
2323 	/* Reconfigure the MAC */
2324 	mutex_lock(&efx->mac_lock);
2325 	efx_mac_reconfigure(efx);
2326 	mutex_unlock(&efx->mac_lock);
2327 
2328 	return 0;
2329 }
2330 
2331 /* Context: netif_addr_lock held, BHs disabled. */
2332 static void efx_set_rx_mode(struct net_device *net_dev)
2333 {
2334 	struct efx_nic *efx = netdev_priv(net_dev);
2335 
2336 	if (efx->port_enabled)
2337 		queue_work(efx->workqueue, &efx->mac_work);
2338 	/* Otherwise efx_start_port() will do this */
2339 }
2340 
2341 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2342 {
2343 	struct efx_nic *efx = netdev_priv(net_dev);
2344 	int rc;
2345 
2346 	/* If disabling RX n-tuple filtering, clear existing filters */
2347 	if (net_dev->features & ~data & NETIF_F_NTUPLE) {
2348 		rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2349 		if (rc)
2350 			return rc;
2351 	}
2352 
2353 	/* If Rx VLAN filter is changed, update filters via mac_reconfigure */
2354 	if ((net_dev->features ^ data) & NETIF_F_HW_VLAN_CTAG_FILTER) {
2355 		/* efx_set_rx_mode() will schedule MAC work to update filters
2356 		 * when a new features are finally set in net_dev.
2357 		 */
2358 		efx_set_rx_mode(net_dev);
2359 	}
2360 
2361 	return 0;
2362 }
2363 
2364 static int efx_vlan_rx_add_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2365 {
2366 	struct efx_nic *efx = netdev_priv(net_dev);
2367 
2368 	if (efx->type->vlan_rx_add_vid)
2369 		return efx->type->vlan_rx_add_vid(efx, proto, vid);
2370 	else
2371 		return -EOPNOTSUPP;
2372 }
2373 
2374 static int efx_vlan_rx_kill_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2375 {
2376 	struct efx_nic *efx = netdev_priv(net_dev);
2377 
2378 	if (efx->type->vlan_rx_kill_vid)
2379 		return efx->type->vlan_rx_kill_vid(efx, proto, vid);
2380 	else
2381 		return -EOPNOTSUPP;
2382 }
2383 
2384 static const struct net_device_ops efx_netdev_ops = {
2385 	.ndo_open		= efx_net_open,
2386 	.ndo_stop		= efx_net_stop,
2387 	.ndo_get_stats64	= efx_net_stats,
2388 	.ndo_tx_timeout		= efx_watchdog,
2389 	.ndo_start_xmit		= efx_hard_start_xmit,
2390 	.ndo_validate_addr	= eth_validate_addr,
2391 	.ndo_do_ioctl		= efx_ioctl,
2392 	.ndo_change_mtu		= efx_change_mtu,
2393 	.ndo_set_mac_address	= efx_set_mac_address,
2394 	.ndo_set_rx_mode	= efx_set_rx_mode,
2395 	.ndo_set_features	= efx_set_features,
2396 	.ndo_vlan_rx_add_vid	= efx_vlan_rx_add_vid,
2397 	.ndo_vlan_rx_kill_vid	= efx_vlan_rx_kill_vid,
2398 #ifdef CONFIG_SFC_SRIOV
2399 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2400 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2401 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2402 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2403 	.ndo_set_vf_link_state  = efx_sriov_set_vf_link_state,
2404 	.ndo_get_phys_port_id   = efx_sriov_get_phys_port_id,
2405 #endif
2406 #ifdef CONFIG_NET_POLL_CONTROLLER
2407 	.ndo_poll_controller = efx_netpoll,
2408 #endif
2409 	.ndo_setup_tc		= efx_setup_tc,
2410 #ifdef CONFIG_NET_RX_BUSY_POLL
2411 	.ndo_busy_poll		= efx_busy_poll,
2412 #endif
2413 #ifdef CONFIG_RFS_ACCEL
2414 	.ndo_rx_flow_steer	= efx_filter_rfs,
2415 #endif
2416 };
2417 
2418 static void efx_update_name(struct efx_nic *efx)
2419 {
2420 	strcpy(efx->name, efx->net_dev->name);
2421 	efx_mtd_rename(efx);
2422 	efx_set_channel_names(efx);
2423 }
2424 
2425 static int efx_netdev_event(struct notifier_block *this,
2426 			    unsigned long event, void *ptr)
2427 {
2428 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2429 
2430 	if ((net_dev->netdev_ops == &efx_netdev_ops) &&
2431 	    event == NETDEV_CHANGENAME)
2432 		efx_update_name(netdev_priv(net_dev));
2433 
2434 	return NOTIFY_DONE;
2435 }
2436 
2437 static struct notifier_block efx_netdev_notifier = {
2438 	.notifier_call = efx_netdev_event,
2439 };
2440 
2441 static ssize_t
2442 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2443 {
2444 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2445 	return sprintf(buf, "%d\n", efx->phy_type);
2446 }
2447 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2448 
2449 #ifdef CONFIG_SFC_MCDI_LOGGING
2450 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
2451 			     char *buf)
2452 {
2453 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2454 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2455 
2456 	return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
2457 }
2458 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
2459 			    const char *buf, size_t count)
2460 {
2461 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2462 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2463 	bool enable = count > 0 && *buf != '0';
2464 
2465 	mcdi->logging_enabled = enable;
2466 	return count;
2467 }
2468 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
2469 #endif
2470 
2471 static int efx_register_netdev(struct efx_nic *efx)
2472 {
2473 	struct net_device *net_dev = efx->net_dev;
2474 	struct efx_channel *channel;
2475 	int rc;
2476 
2477 	net_dev->watchdog_timeo = 5 * HZ;
2478 	net_dev->irq = efx->pci_dev->irq;
2479 	net_dev->netdev_ops = &efx_netdev_ops;
2480 	if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
2481 		net_dev->priv_flags |= IFF_UNICAST_FLT;
2482 	net_dev->ethtool_ops = &efx_ethtool_ops;
2483 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2484 
2485 	rtnl_lock();
2486 
2487 	/* Enable resets to be scheduled and check whether any were
2488 	 * already requested.  If so, the NIC is probably hosed so we
2489 	 * abort.
2490 	 */
2491 	efx->state = STATE_READY;
2492 	smp_mb(); /* ensure we change state before checking reset_pending */
2493 	if (efx->reset_pending) {
2494 		netif_err(efx, probe, efx->net_dev,
2495 			  "aborting probe due to scheduled reset\n");
2496 		rc = -EIO;
2497 		goto fail_locked;
2498 	}
2499 
2500 	rc = dev_alloc_name(net_dev, net_dev->name);
2501 	if (rc < 0)
2502 		goto fail_locked;
2503 	efx_update_name(efx);
2504 
2505 	/* Always start with carrier off; PHY events will detect the link */
2506 	netif_carrier_off(net_dev);
2507 
2508 	rc = register_netdevice(net_dev);
2509 	if (rc)
2510 		goto fail_locked;
2511 
2512 	efx_for_each_channel(channel, efx) {
2513 		struct efx_tx_queue *tx_queue;
2514 		efx_for_each_channel_tx_queue(tx_queue, channel)
2515 			efx_init_tx_queue_core_txq(tx_queue);
2516 	}
2517 
2518 	efx_associate(efx);
2519 
2520 	rtnl_unlock();
2521 
2522 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2523 	if (rc) {
2524 		netif_err(efx, drv, efx->net_dev,
2525 			  "failed to init net dev attributes\n");
2526 		goto fail_registered;
2527 	}
2528 #ifdef CONFIG_SFC_MCDI_LOGGING
2529 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2530 	if (rc) {
2531 		netif_err(efx, drv, efx->net_dev,
2532 			  "failed to init net dev attributes\n");
2533 		goto fail_attr_mcdi_logging;
2534 	}
2535 #endif
2536 
2537 	return 0;
2538 
2539 #ifdef CONFIG_SFC_MCDI_LOGGING
2540 fail_attr_mcdi_logging:
2541 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2542 #endif
2543 fail_registered:
2544 	rtnl_lock();
2545 	efx_dissociate(efx);
2546 	unregister_netdevice(net_dev);
2547 fail_locked:
2548 	efx->state = STATE_UNINIT;
2549 	rtnl_unlock();
2550 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2551 	return rc;
2552 }
2553 
2554 static void efx_unregister_netdev(struct efx_nic *efx)
2555 {
2556 	if (!efx->net_dev)
2557 		return;
2558 
2559 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2560 
2561 	if (efx_dev_registered(efx)) {
2562 		strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2563 #ifdef CONFIG_SFC_MCDI_LOGGING
2564 		device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2565 #endif
2566 		device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2567 		unregister_netdev(efx->net_dev);
2568 	}
2569 }
2570 
2571 /**************************************************************************
2572  *
2573  * Device reset and suspend
2574  *
2575  **************************************************************************/
2576 
2577 /* Tears down the entire software state and most of the hardware state
2578  * before reset.  */
2579 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2580 {
2581 	EFX_ASSERT_RESET_SERIALISED(efx);
2582 
2583 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2584 		efx->type->prepare_flr(efx);
2585 
2586 	efx_stop_all(efx);
2587 	efx_disable_interrupts(efx);
2588 
2589 	mutex_lock(&efx->mac_lock);
2590 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2591 	    method != RESET_TYPE_DATAPATH)
2592 		efx->phy_op->fini(efx);
2593 	efx->type->fini(efx);
2594 }
2595 
2596 /* This function will always ensure that the locks acquired in
2597  * efx_reset_down() are released. A failure return code indicates
2598  * that we were unable to reinitialise the hardware, and the
2599  * driver should be disabled. If ok is false, then the rx and tx
2600  * engines are not restarted, pending a RESET_DISABLE. */
2601 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2602 {
2603 	int rc;
2604 
2605 	EFX_ASSERT_RESET_SERIALISED(efx);
2606 
2607 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2608 		efx->type->finish_flr(efx);
2609 
2610 	/* Ensure that SRAM is initialised even if we're disabling the device */
2611 	rc = efx->type->init(efx);
2612 	if (rc) {
2613 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2614 		goto fail;
2615 	}
2616 
2617 	if (!ok)
2618 		goto fail;
2619 
2620 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2621 	    method != RESET_TYPE_DATAPATH) {
2622 		rc = efx->phy_op->init(efx);
2623 		if (rc)
2624 			goto fail;
2625 		rc = efx->phy_op->reconfigure(efx);
2626 		if (rc && rc != -EPERM)
2627 			netif_err(efx, drv, efx->net_dev,
2628 				  "could not restore PHY settings\n");
2629 	}
2630 
2631 	rc = efx_enable_interrupts(efx);
2632 	if (rc)
2633 		goto fail;
2634 
2635 #ifdef CONFIG_SFC_SRIOV
2636 	rc = efx->type->vswitching_restore(efx);
2637 	if (rc) /* not fatal; the PF will still work fine */
2638 		netif_warn(efx, probe, efx->net_dev,
2639 			   "failed to restore vswitching rc=%d;"
2640 			   " VFs may not function\n", rc);
2641 #endif
2642 
2643 	down_read(&efx->filter_sem);
2644 	efx_restore_filters(efx);
2645 	up_read(&efx->filter_sem);
2646 	if (efx->type->sriov_reset)
2647 		efx->type->sriov_reset(efx);
2648 
2649 	mutex_unlock(&efx->mac_lock);
2650 
2651 	efx_start_all(efx);
2652 
2653 	return 0;
2654 
2655 fail:
2656 	efx->port_initialized = false;
2657 
2658 	mutex_unlock(&efx->mac_lock);
2659 
2660 	return rc;
2661 }
2662 
2663 /* Reset the NIC using the specified method.  Note that the reset may
2664  * fail, in which case the card will be left in an unusable state.
2665  *
2666  * Caller must hold the rtnl_lock.
2667  */
2668 int efx_reset(struct efx_nic *efx, enum reset_type method)
2669 {
2670 	int rc, rc2;
2671 	bool disabled;
2672 
2673 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2674 		   RESET_TYPE(method));
2675 
2676 	efx_device_detach_sync(efx);
2677 	efx_reset_down(efx, method);
2678 
2679 	rc = efx->type->reset(efx, method);
2680 	if (rc) {
2681 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2682 		goto out;
2683 	}
2684 
2685 	/* Clear flags for the scopes we covered.  We assume the NIC and
2686 	 * driver are now quiescent so that there is no race here.
2687 	 */
2688 	if (method < RESET_TYPE_MAX_METHOD)
2689 		efx->reset_pending &= -(1 << (method + 1));
2690 	else /* it doesn't fit into the well-ordered scope hierarchy */
2691 		__clear_bit(method, &efx->reset_pending);
2692 
2693 	/* Reinitialise bus-mastering, which may have been turned off before
2694 	 * the reset was scheduled. This is still appropriate, even in the
2695 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2696 	 * can respond to requests. */
2697 	pci_set_master(efx->pci_dev);
2698 
2699 out:
2700 	/* Leave device stopped if necessary */
2701 	disabled = rc ||
2702 		method == RESET_TYPE_DISABLE ||
2703 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2704 	rc2 = efx_reset_up(efx, method, !disabled);
2705 	if (rc2) {
2706 		disabled = true;
2707 		if (!rc)
2708 			rc = rc2;
2709 	}
2710 
2711 	if (disabled) {
2712 		dev_close(efx->net_dev);
2713 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2714 		efx->state = STATE_DISABLED;
2715 	} else {
2716 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2717 		netif_device_attach(efx->net_dev);
2718 	}
2719 	return rc;
2720 }
2721 
2722 /* Try recovery mechanisms.
2723  * For now only EEH is supported.
2724  * Returns 0 if the recovery mechanisms are unsuccessful.
2725  * Returns a non-zero value otherwise.
2726  */
2727 int efx_try_recovery(struct efx_nic *efx)
2728 {
2729 #ifdef CONFIG_EEH
2730 	/* A PCI error can occur and not be seen by EEH because nothing
2731 	 * happens on the PCI bus. In this case the driver may fail and
2732 	 * schedule a 'recover or reset', leading to this recovery handler.
2733 	 * Manually call the eeh failure check function.
2734 	 */
2735 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2736 	if (eeh_dev_check_failure(eehdev)) {
2737 		/* The EEH mechanisms will handle the error and reset the
2738 		 * device if necessary.
2739 		 */
2740 		return 1;
2741 	}
2742 #endif
2743 	return 0;
2744 }
2745 
2746 static void efx_wait_for_bist_end(struct efx_nic *efx)
2747 {
2748 	int i;
2749 
2750 	for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2751 		if (efx_mcdi_poll_reboot(efx))
2752 			goto out;
2753 		msleep(BIST_WAIT_DELAY_MS);
2754 	}
2755 
2756 	netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2757 out:
2758 	/* Either way unset the BIST flag. If we found no reboot we probably
2759 	 * won't recover, but we should try.
2760 	 */
2761 	efx->mc_bist_for_other_fn = false;
2762 }
2763 
2764 /* The worker thread exists so that code that cannot sleep can
2765  * schedule a reset for later.
2766  */
2767 static void efx_reset_work(struct work_struct *data)
2768 {
2769 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2770 	unsigned long pending;
2771 	enum reset_type method;
2772 
2773 	pending = ACCESS_ONCE(efx->reset_pending);
2774 	method = fls(pending) - 1;
2775 
2776 	if (method == RESET_TYPE_MC_BIST)
2777 		efx_wait_for_bist_end(efx);
2778 
2779 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2780 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2781 	    efx_try_recovery(efx))
2782 		return;
2783 
2784 	if (!pending)
2785 		return;
2786 
2787 	rtnl_lock();
2788 
2789 	/* We checked the state in efx_schedule_reset() but it may
2790 	 * have changed by now.  Now that we have the RTNL lock,
2791 	 * it cannot change again.
2792 	 */
2793 	if (efx->state == STATE_READY)
2794 		(void)efx_reset(efx, method);
2795 
2796 	rtnl_unlock();
2797 }
2798 
2799 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2800 {
2801 	enum reset_type method;
2802 
2803 	if (efx->state == STATE_RECOVERY) {
2804 		netif_dbg(efx, drv, efx->net_dev,
2805 			  "recovering: skip scheduling %s reset\n",
2806 			  RESET_TYPE(type));
2807 		return;
2808 	}
2809 
2810 	switch (type) {
2811 	case RESET_TYPE_INVISIBLE:
2812 	case RESET_TYPE_ALL:
2813 	case RESET_TYPE_RECOVER_OR_ALL:
2814 	case RESET_TYPE_WORLD:
2815 	case RESET_TYPE_DISABLE:
2816 	case RESET_TYPE_RECOVER_OR_DISABLE:
2817 	case RESET_TYPE_DATAPATH:
2818 	case RESET_TYPE_MC_BIST:
2819 	case RESET_TYPE_MCDI_TIMEOUT:
2820 		method = type;
2821 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2822 			  RESET_TYPE(method));
2823 		break;
2824 	default:
2825 		method = efx->type->map_reset_reason(type);
2826 		netif_dbg(efx, drv, efx->net_dev,
2827 			  "scheduling %s reset for %s\n",
2828 			  RESET_TYPE(method), RESET_TYPE(type));
2829 		break;
2830 	}
2831 
2832 	set_bit(method, &efx->reset_pending);
2833 	smp_mb(); /* ensure we change reset_pending before checking state */
2834 
2835 	/* If we're not READY then just leave the flags set as the cue
2836 	 * to abort probing or reschedule the reset later.
2837 	 */
2838 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2839 		return;
2840 
2841 	/* efx_process_channel() will no longer read events once a
2842 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2843 	efx_mcdi_mode_poll(efx);
2844 
2845 	queue_work(reset_workqueue, &efx->reset_work);
2846 }
2847 
2848 /**************************************************************************
2849  *
2850  * List of NICs we support
2851  *
2852  **************************************************************************/
2853 
2854 /* PCI device ID table */
2855 static const struct pci_device_id efx_pci_table[] = {
2856 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2857 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2858 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2859 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2860 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2861 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2862 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2863 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2864 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2865 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2866 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903),  /* SFC9120 PF */
2867 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2868 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903),  /* SFC9120 VF */
2869 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2870 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923),  /* SFC9140 PF */
2871 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2872 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1923),  /* SFC9140 VF */
2873 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2874 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0a03),  /* SFC9220 PF */
2875 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2876 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1a03),  /* SFC9220 VF */
2877 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2878 	{0}			/* end of list */
2879 };
2880 
2881 /**************************************************************************
2882  *
2883  * Dummy PHY/MAC operations
2884  *
2885  * Can be used for some unimplemented operations
2886  * Needed so all function pointers are valid and do not have to be tested
2887  * before use
2888  *
2889  **************************************************************************/
2890 int efx_port_dummy_op_int(struct efx_nic *efx)
2891 {
2892 	return 0;
2893 }
2894 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2895 
2896 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2897 {
2898 	return false;
2899 }
2900 
2901 static const struct efx_phy_operations efx_dummy_phy_operations = {
2902 	.init		 = efx_port_dummy_op_int,
2903 	.reconfigure	 = efx_port_dummy_op_int,
2904 	.poll		 = efx_port_dummy_op_poll,
2905 	.fini		 = efx_port_dummy_op_void,
2906 };
2907 
2908 /**************************************************************************
2909  *
2910  * Data housekeeping
2911  *
2912  **************************************************************************/
2913 
2914 /* This zeroes out and then fills in the invariants in a struct
2915  * efx_nic (including all sub-structures).
2916  */
2917 static int efx_init_struct(struct efx_nic *efx,
2918 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2919 {
2920 	int i;
2921 
2922 	/* Initialise common structures */
2923 	INIT_LIST_HEAD(&efx->node);
2924 	INIT_LIST_HEAD(&efx->secondary_list);
2925 	spin_lock_init(&efx->biu_lock);
2926 #ifdef CONFIG_SFC_MTD
2927 	INIT_LIST_HEAD(&efx->mtd_list);
2928 #endif
2929 	INIT_WORK(&efx->reset_work, efx_reset_work);
2930 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2931 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2932 	efx->pci_dev = pci_dev;
2933 	efx->msg_enable = debug;
2934 	efx->state = STATE_UNINIT;
2935 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2936 
2937 	efx->net_dev = net_dev;
2938 	efx->rx_prefix_size = efx->type->rx_prefix_size;
2939 	efx->rx_ip_align =
2940 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2941 	efx->rx_packet_hash_offset =
2942 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2943 	efx->rx_packet_ts_offset =
2944 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2945 	spin_lock_init(&efx->stats_lock);
2946 	mutex_init(&efx->mac_lock);
2947 	efx->phy_op = &efx_dummy_phy_operations;
2948 	efx->mdio.dev = net_dev;
2949 	INIT_WORK(&efx->mac_work, efx_mac_work);
2950 	init_waitqueue_head(&efx->flush_wq);
2951 
2952 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2953 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2954 		if (!efx->channel[i])
2955 			goto fail;
2956 		efx->msi_context[i].efx = efx;
2957 		efx->msi_context[i].index = i;
2958 	}
2959 
2960 	/* Higher numbered interrupt modes are less capable! */
2961 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2962 				  interrupt_mode);
2963 
2964 	/* Would be good to use the net_dev name, but we're too early */
2965 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2966 		 pci_name(pci_dev));
2967 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2968 	if (!efx->workqueue)
2969 		goto fail;
2970 
2971 	return 0;
2972 
2973 fail:
2974 	efx_fini_struct(efx);
2975 	return -ENOMEM;
2976 }
2977 
2978 static void efx_fini_struct(struct efx_nic *efx)
2979 {
2980 	int i;
2981 
2982 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2983 		kfree(efx->channel[i]);
2984 
2985 	kfree(efx->vpd_sn);
2986 
2987 	if (efx->workqueue) {
2988 		destroy_workqueue(efx->workqueue);
2989 		efx->workqueue = NULL;
2990 	}
2991 }
2992 
2993 void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
2994 {
2995 	u64 n_rx_nodesc_trunc = 0;
2996 	struct efx_channel *channel;
2997 
2998 	efx_for_each_channel(channel, efx)
2999 		n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
3000 	stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
3001 	stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
3002 }
3003 
3004 /**************************************************************************
3005  *
3006  * PCI interface
3007  *
3008  **************************************************************************/
3009 
3010 /* Main body of final NIC shutdown code
3011  * This is called only at module unload (or hotplug removal).
3012  */
3013 static void efx_pci_remove_main(struct efx_nic *efx)
3014 {
3015 	/* Flush reset_work. It can no longer be scheduled since we
3016 	 * are not READY.
3017 	 */
3018 	BUG_ON(efx->state == STATE_READY);
3019 	cancel_work_sync(&efx->reset_work);
3020 
3021 	efx_disable_interrupts(efx);
3022 	efx_nic_fini_interrupt(efx);
3023 	efx_fini_port(efx);
3024 	efx->type->fini(efx);
3025 	efx_fini_napi(efx);
3026 	efx_remove_all(efx);
3027 }
3028 
3029 /* Final NIC shutdown
3030  * This is called only at module unload (or hotplug removal).  A PF can call
3031  * this on its VFs to ensure they are unbound first.
3032  */
3033 static void efx_pci_remove(struct pci_dev *pci_dev)
3034 {
3035 	struct efx_nic *efx;
3036 
3037 	efx = pci_get_drvdata(pci_dev);
3038 	if (!efx)
3039 		return;
3040 
3041 	/* Mark the NIC as fini, then stop the interface */
3042 	rtnl_lock();
3043 	efx_dissociate(efx);
3044 	dev_close(efx->net_dev);
3045 	efx_disable_interrupts(efx);
3046 	efx->state = STATE_UNINIT;
3047 	rtnl_unlock();
3048 
3049 	if (efx->type->sriov_fini)
3050 		efx->type->sriov_fini(efx);
3051 
3052 	efx_unregister_netdev(efx);
3053 
3054 	efx_mtd_remove(efx);
3055 
3056 	efx_pci_remove_main(efx);
3057 
3058 	efx_fini_io(efx);
3059 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
3060 
3061 	efx_fini_struct(efx);
3062 	free_netdev(efx->net_dev);
3063 
3064 	pci_disable_pcie_error_reporting(pci_dev);
3065 };
3066 
3067 /* NIC VPD information
3068  * Called during probe to display the part number of the
3069  * installed NIC.  VPD is potentially very large but this should
3070  * always appear within the first 512 bytes.
3071  */
3072 #define SFC_VPD_LEN 512
3073 static void efx_probe_vpd_strings(struct efx_nic *efx)
3074 {
3075 	struct pci_dev *dev = efx->pci_dev;
3076 	char vpd_data[SFC_VPD_LEN];
3077 	ssize_t vpd_size;
3078 	int ro_start, ro_size, i, j;
3079 
3080 	/* Get the vpd data from the device */
3081 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
3082 	if (vpd_size <= 0) {
3083 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
3084 		return;
3085 	}
3086 
3087 	/* Get the Read only section */
3088 	ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
3089 	if (ro_start < 0) {
3090 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
3091 		return;
3092 	}
3093 
3094 	ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
3095 	j = ro_size;
3096 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3097 	if (i + j > vpd_size)
3098 		j = vpd_size - i;
3099 
3100 	/* Get the Part number */
3101 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
3102 	if (i < 0) {
3103 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
3104 		return;
3105 	}
3106 
3107 	j = pci_vpd_info_field_size(&vpd_data[i]);
3108 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3109 	if (i + j > vpd_size) {
3110 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
3111 		return;
3112 	}
3113 
3114 	netif_info(efx, drv, efx->net_dev,
3115 		   "Part Number : %.*s\n", j, &vpd_data[i]);
3116 
3117 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3118 	j = ro_size;
3119 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
3120 	if (i < 0) {
3121 		netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
3122 		return;
3123 	}
3124 
3125 	j = pci_vpd_info_field_size(&vpd_data[i]);
3126 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3127 	if (i + j > vpd_size) {
3128 		netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
3129 		return;
3130 	}
3131 
3132 	efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
3133 	if (!efx->vpd_sn)
3134 		return;
3135 
3136 	snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
3137 }
3138 
3139 
3140 /* Main body of NIC initialisation
3141  * This is called at module load (or hotplug insertion, theoretically).
3142  */
3143 static int efx_pci_probe_main(struct efx_nic *efx)
3144 {
3145 	int rc;
3146 
3147 	/* Do start-of-day initialisation */
3148 	rc = efx_probe_all(efx);
3149 	if (rc)
3150 		goto fail1;
3151 
3152 	efx_init_napi(efx);
3153 
3154 	rc = efx->type->init(efx);
3155 	if (rc) {
3156 		netif_err(efx, probe, efx->net_dev,
3157 			  "failed to initialise NIC\n");
3158 		goto fail3;
3159 	}
3160 
3161 	rc = efx_init_port(efx);
3162 	if (rc) {
3163 		netif_err(efx, probe, efx->net_dev,
3164 			  "failed to initialise port\n");
3165 		goto fail4;
3166 	}
3167 
3168 	rc = efx_nic_init_interrupt(efx);
3169 	if (rc)
3170 		goto fail5;
3171 	rc = efx_enable_interrupts(efx);
3172 	if (rc)
3173 		goto fail6;
3174 
3175 	return 0;
3176 
3177  fail6:
3178 	efx_nic_fini_interrupt(efx);
3179  fail5:
3180 	efx_fini_port(efx);
3181  fail4:
3182 	efx->type->fini(efx);
3183  fail3:
3184 	efx_fini_napi(efx);
3185 	efx_remove_all(efx);
3186  fail1:
3187 	return rc;
3188 }
3189 
3190 /* NIC initialisation
3191  *
3192  * This is called at module load (or hotplug insertion,
3193  * theoretically).  It sets up PCI mappings, resets the NIC,
3194  * sets up and registers the network devices with the kernel and hooks
3195  * the interrupt service routine.  It does not prepare the device for
3196  * transmission; this is left to the first time one of the network
3197  * interfaces is brought up (i.e. efx_net_open).
3198  */
3199 static int efx_pci_probe(struct pci_dev *pci_dev,
3200 			 const struct pci_device_id *entry)
3201 {
3202 	struct net_device *net_dev;
3203 	struct efx_nic *efx;
3204 	int rc;
3205 
3206 	/* Allocate and initialise a struct net_device and struct efx_nic */
3207 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
3208 				     EFX_MAX_RX_QUEUES);
3209 	if (!net_dev)
3210 		return -ENOMEM;
3211 	efx = netdev_priv(net_dev);
3212 	efx->type = (const struct efx_nic_type *) entry->driver_data;
3213 	efx->fixed_features |= NETIF_F_HIGHDMA;
3214 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
3215 			      NETIF_F_TSO | NETIF_F_RXCSUM);
3216 	if (efx->type->offload_features & (NETIF_F_IPV6_CSUM | NETIF_F_HW_CSUM))
3217 		net_dev->features |= NETIF_F_TSO6;
3218 	/* Mask for features that also apply to VLAN devices */
3219 	net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
3220 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
3221 				   NETIF_F_RXCSUM);
3222 
3223 	net_dev->hw_features = net_dev->features & ~efx->fixed_features;
3224 
3225 	/* Disable VLAN filtering by default.  It may be enforced if
3226 	 * the feature is fixed (i.e. VLAN filters are required to
3227 	 * receive VLAN tagged packets due to vPort restrictions).
3228 	 */
3229 	net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
3230 	net_dev->features |= efx->fixed_features;
3231 
3232 	pci_set_drvdata(pci_dev, efx);
3233 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
3234 	rc = efx_init_struct(efx, pci_dev, net_dev);
3235 	if (rc)
3236 		goto fail1;
3237 
3238 	netif_info(efx, probe, efx->net_dev,
3239 		   "Solarflare NIC detected\n");
3240 
3241 	if (!efx->type->is_vf)
3242 		efx_probe_vpd_strings(efx);
3243 
3244 	/* Set up basic I/O (BAR mappings etc) */
3245 	rc = efx_init_io(efx);
3246 	if (rc)
3247 		goto fail2;
3248 
3249 	rc = efx_pci_probe_main(efx);
3250 	if (rc)
3251 		goto fail3;
3252 
3253 	rc = efx_register_netdev(efx);
3254 	if (rc)
3255 		goto fail4;
3256 
3257 	if (efx->type->sriov_init) {
3258 		rc = efx->type->sriov_init(efx);
3259 		if (rc)
3260 			netif_err(efx, probe, efx->net_dev,
3261 				  "SR-IOV can't be enabled rc %d\n", rc);
3262 	}
3263 
3264 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3265 
3266 	/* Try to create MTDs, but allow this to fail */
3267 	rtnl_lock();
3268 	rc = efx_mtd_probe(efx);
3269 	rtnl_unlock();
3270 	if (rc && rc != -EPERM)
3271 		netif_warn(efx, probe, efx->net_dev,
3272 			   "failed to create MTDs (%d)\n", rc);
3273 
3274 	rc = pci_enable_pcie_error_reporting(pci_dev);
3275 	if (rc && rc != -EINVAL)
3276 		netif_notice(efx, probe, efx->net_dev,
3277 			     "PCIE error reporting unavailable (%d).\n",
3278 			     rc);
3279 
3280 	return 0;
3281 
3282  fail4:
3283 	efx_pci_remove_main(efx);
3284  fail3:
3285 	efx_fini_io(efx);
3286  fail2:
3287 	efx_fini_struct(efx);
3288  fail1:
3289 	WARN_ON(rc > 0);
3290 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3291 	free_netdev(net_dev);
3292 	return rc;
3293 }
3294 
3295 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3296  * enabled on success
3297  */
3298 #ifdef CONFIG_SFC_SRIOV
3299 static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
3300 {
3301 	int rc;
3302 	struct efx_nic *efx = pci_get_drvdata(dev);
3303 
3304 	if (efx->type->sriov_configure) {
3305 		rc = efx->type->sriov_configure(efx, num_vfs);
3306 		if (rc)
3307 			return rc;
3308 		else
3309 			return num_vfs;
3310 	} else
3311 		return -EOPNOTSUPP;
3312 }
3313 #endif
3314 
3315 static int efx_pm_freeze(struct device *dev)
3316 {
3317 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3318 
3319 	rtnl_lock();
3320 
3321 	if (efx->state != STATE_DISABLED) {
3322 		efx->state = STATE_UNINIT;
3323 
3324 		efx_device_detach_sync(efx);
3325 
3326 		efx_stop_all(efx);
3327 		efx_disable_interrupts(efx);
3328 	}
3329 
3330 	rtnl_unlock();
3331 
3332 	return 0;
3333 }
3334 
3335 static int efx_pm_thaw(struct device *dev)
3336 {
3337 	int rc;
3338 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3339 
3340 	rtnl_lock();
3341 
3342 	if (efx->state != STATE_DISABLED) {
3343 		rc = efx_enable_interrupts(efx);
3344 		if (rc)
3345 			goto fail;
3346 
3347 		mutex_lock(&efx->mac_lock);
3348 		efx->phy_op->reconfigure(efx);
3349 		mutex_unlock(&efx->mac_lock);
3350 
3351 		efx_start_all(efx);
3352 
3353 		netif_device_attach(efx->net_dev);
3354 
3355 		efx->state = STATE_READY;
3356 
3357 		efx->type->resume_wol(efx);
3358 	}
3359 
3360 	rtnl_unlock();
3361 
3362 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3363 	queue_work(reset_workqueue, &efx->reset_work);
3364 
3365 	return 0;
3366 
3367 fail:
3368 	rtnl_unlock();
3369 
3370 	return rc;
3371 }
3372 
3373 static int efx_pm_poweroff(struct device *dev)
3374 {
3375 	struct pci_dev *pci_dev = to_pci_dev(dev);
3376 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3377 
3378 	efx->type->fini(efx);
3379 
3380 	efx->reset_pending = 0;
3381 
3382 	pci_save_state(pci_dev);
3383 	return pci_set_power_state(pci_dev, PCI_D3hot);
3384 }
3385 
3386 /* Used for both resume and restore */
3387 static int efx_pm_resume(struct device *dev)
3388 {
3389 	struct pci_dev *pci_dev = to_pci_dev(dev);
3390 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3391 	int rc;
3392 
3393 	rc = pci_set_power_state(pci_dev, PCI_D0);
3394 	if (rc)
3395 		return rc;
3396 	pci_restore_state(pci_dev);
3397 	rc = pci_enable_device(pci_dev);
3398 	if (rc)
3399 		return rc;
3400 	pci_set_master(efx->pci_dev);
3401 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
3402 	if (rc)
3403 		return rc;
3404 	rc = efx->type->init(efx);
3405 	if (rc)
3406 		return rc;
3407 	rc = efx_pm_thaw(dev);
3408 	return rc;
3409 }
3410 
3411 static int efx_pm_suspend(struct device *dev)
3412 {
3413 	int rc;
3414 
3415 	efx_pm_freeze(dev);
3416 	rc = efx_pm_poweroff(dev);
3417 	if (rc)
3418 		efx_pm_resume(dev);
3419 	return rc;
3420 }
3421 
3422 static const struct dev_pm_ops efx_pm_ops = {
3423 	.suspend	= efx_pm_suspend,
3424 	.resume		= efx_pm_resume,
3425 	.freeze		= efx_pm_freeze,
3426 	.thaw		= efx_pm_thaw,
3427 	.poweroff	= efx_pm_poweroff,
3428 	.restore	= efx_pm_resume,
3429 };
3430 
3431 /* A PCI error affecting this device was detected.
3432  * At this point MMIO and DMA may be disabled.
3433  * Stop the software path and request a slot reset.
3434  */
3435 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3436 					      enum pci_channel_state state)
3437 {
3438 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3439 	struct efx_nic *efx = pci_get_drvdata(pdev);
3440 
3441 	if (state == pci_channel_io_perm_failure)
3442 		return PCI_ERS_RESULT_DISCONNECT;
3443 
3444 	rtnl_lock();
3445 
3446 	if (efx->state != STATE_DISABLED) {
3447 		efx->state = STATE_RECOVERY;
3448 		efx->reset_pending = 0;
3449 
3450 		efx_device_detach_sync(efx);
3451 
3452 		efx_stop_all(efx);
3453 		efx_disable_interrupts(efx);
3454 
3455 		status = PCI_ERS_RESULT_NEED_RESET;
3456 	} else {
3457 		/* If the interface is disabled we don't want to do anything
3458 		 * with it.
3459 		 */
3460 		status = PCI_ERS_RESULT_RECOVERED;
3461 	}
3462 
3463 	rtnl_unlock();
3464 
3465 	pci_disable_device(pdev);
3466 
3467 	return status;
3468 }
3469 
3470 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3471 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3472 {
3473 	struct efx_nic *efx = pci_get_drvdata(pdev);
3474 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3475 	int rc;
3476 
3477 	if (pci_enable_device(pdev)) {
3478 		netif_err(efx, hw, efx->net_dev,
3479 			  "Cannot re-enable PCI device after reset.\n");
3480 		status =  PCI_ERS_RESULT_DISCONNECT;
3481 	}
3482 
3483 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3484 	if (rc) {
3485 		netif_err(efx, hw, efx->net_dev,
3486 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3487 		/* Non-fatal error. Continue. */
3488 	}
3489 
3490 	return status;
3491 }
3492 
3493 /* Perform the actual reset and resume I/O operations. */
3494 static void efx_io_resume(struct pci_dev *pdev)
3495 {
3496 	struct efx_nic *efx = pci_get_drvdata(pdev);
3497 	int rc;
3498 
3499 	rtnl_lock();
3500 
3501 	if (efx->state == STATE_DISABLED)
3502 		goto out;
3503 
3504 	rc = efx_reset(efx, RESET_TYPE_ALL);
3505 	if (rc) {
3506 		netif_err(efx, hw, efx->net_dev,
3507 			  "efx_reset failed after PCI error (%d)\n", rc);
3508 	} else {
3509 		efx->state = STATE_READY;
3510 		netif_dbg(efx, hw, efx->net_dev,
3511 			  "Done resetting and resuming IO after PCI error.\n");
3512 	}
3513 
3514 out:
3515 	rtnl_unlock();
3516 }
3517 
3518 /* For simplicity and reliability, we always require a slot reset and try to
3519  * reset the hardware when a pci error affecting the device is detected.
3520  * We leave both the link_reset and mmio_enabled callback unimplemented:
3521  * with our request for slot reset the mmio_enabled callback will never be
3522  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3523  */
3524 static const struct pci_error_handlers efx_err_handlers = {
3525 	.error_detected = efx_io_error_detected,
3526 	.slot_reset	= efx_io_slot_reset,
3527 	.resume		= efx_io_resume,
3528 };
3529 
3530 static struct pci_driver efx_pci_driver = {
3531 	.name		= KBUILD_MODNAME,
3532 	.id_table	= efx_pci_table,
3533 	.probe		= efx_pci_probe,
3534 	.remove		= efx_pci_remove,
3535 	.driver.pm	= &efx_pm_ops,
3536 	.err_handler	= &efx_err_handlers,
3537 #ifdef CONFIG_SFC_SRIOV
3538 	.sriov_configure = efx_pci_sriov_configure,
3539 #endif
3540 };
3541 
3542 /**************************************************************************
3543  *
3544  * Kernel module interface
3545  *
3546  *************************************************************************/
3547 
3548 module_param(interrupt_mode, uint, 0444);
3549 MODULE_PARM_DESC(interrupt_mode,
3550 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3551 
3552 static int __init efx_init_module(void)
3553 {
3554 	int rc;
3555 
3556 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3557 
3558 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3559 	if (rc)
3560 		goto err_notifier;
3561 
3562 #ifdef CONFIG_SFC_SRIOV
3563 	rc = efx_init_sriov();
3564 	if (rc)
3565 		goto err_sriov;
3566 #endif
3567 
3568 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3569 	if (!reset_workqueue) {
3570 		rc = -ENOMEM;
3571 		goto err_reset;
3572 	}
3573 
3574 	rc = pci_register_driver(&efx_pci_driver);
3575 	if (rc < 0)
3576 		goto err_pci;
3577 
3578 	return 0;
3579 
3580  err_pci:
3581 	destroy_workqueue(reset_workqueue);
3582  err_reset:
3583 #ifdef CONFIG_SFC_SRIOV
3584 	efx_fini_sriov();
3585  err_sriov:
3586 #endif
3587 	unregister_netdevice_notifier(&efx_netdev_notifier);
3588  err_notifier:
3589 	return rc;
3590 }
3591 
3592 static void __exit efx_exit_module(void)
3593 {
3594 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3595 
3596 	pci_unregister_driver(&efx_pci_driver);
3597 	destroy_workqueue(reset_workqueue);
3598 #ifdef CONFIG_SFC_SRIOV
3599 	efx_fini_sriov();
3600 #endif
3601 	unregister_netdevice_notifier(&efx_netdev_notifier);
3602 
3603 }
3604 
3605 module_init(efx_init_module);
3606 module_exit(efx_exit_module);
3607 
3608 MODULE_AUTHOR("Solarflare Communications and "
3609 	      "Michael Brown <mbrown@fensystems.co.uk>");
3610 MODULE_DESCRIPTION("Solarflare network driver");
3611 MODULE_LICENSE("GPL");
3612 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3613