xref: /linux/drivers/net/ethernet/sfc/efx.c (revision ff5599816711d2e67da2d7561fd36ac48debd433)
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2011 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/crc32.h>
21 #include <linux/ethtool.h>
22 #include <linux/topology.h>
23 #include <linux/gfp.h>
24 #include <linux/aer.h>
25 #include <linux/interrupt.h>
26 #include "net_driver.h"
27 #include "efx.h"
28 #include "nic.h"
29 #include "selftest.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_DISABLE]            = "DISABLE",
81 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
82 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
83 	[RESET_TYPE_RX_RECOVERY]        = "RX_RECOVERY",
84 	[RESET_TYPE_RX_DESC_FETCH]      = "RX_DESC_FETCH",
85 	[RESET_TYPE_TX_DESC_FETCH]      = "TX_DESC_FETCH",
86 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
87 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
88 };
89 
90 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
91  * queued onto this work queue. This is not a per-nic work queue, because
92  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
93  */
94 static struct workqueue_struct *reset_workqueue;
95 
96 /**************************************************************************
97  *
98  * Configurable values
99  *
100  *************************************************************************/
101 
102 /*
103  * Use separate channels for TX and RX events
104  *
105  * Set this to 1 to use separate channels for TX and RX. It allows us
106  * to control interrupt affinity separately for TX and RX.
107  *
108  * This is only used in MSI-X interrupt mode
109  */
110 static bool separate_tx_channels;
111 module_param(separate_tx_channels, bool, 0444);
112 MODULE_PARM_DESC(separate_tx_channels,
113 		 "Use separate channels for TX and RX");
114 
115 /* This is the weight assigned to each of the (per-channel) virtual
116  * NAPI devices.
117  */
118 static int napi_weight = 64;
119 
120 /* This is the time (in jiffies) between invocations of the hardware
121  * monitor.
122  * On Falcon-based NICs, this will:
123  * - Check the on-board hardware monitor;
124  * - Poll the link state and reconfigure the hardware as necessary.
125  * On Siena-based NICs for power systems with EEH support, this will give EEH a
126  * chance to start.
127  */
128 static unsigned int efx_monitor_interval = 1 * HZ;
129 
130 /* Initial interrupt moderation settings.  They can be modified after
131  * module load with ethtool.
132  *
133  * The default for RX should strike a balance between increasing the
134  * round-trip latency and reducing overhead.
135  */
136 static unsigned int rx_irq_mod_usec = 60;
137 
138 /* Initial interrupt moderation settings.  They can be modified after
139  * module load with ethtool.
140  *
141  * This default is chosen to ensure that a 10G link does not go idle
142  * while a TX queue is stopped after it has become full.  A queue is
143  * restarted when it drops below half full.  The time this takes (assuming
144  * worst case 3 descriptors per packet and 1024 descriptors) is
145  *   512 / 3 * 1.2 = 205 usec.
146  */
147 static unsigned int tx_irq_mod_usec = 150;
148 
149 /* This is the first interrupt mode to try out of:
150  * 0 => MSI-X
151  * 1 => MSI
152  * 2 => legacy
153  */
154 static unsigned int interrupt_mode;
155 
156 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
157  * i.e. the number of CPUs among which we may distribute simultaneous
158  * interrupt handling.
159  *
160  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
161  * The default (0) means to assign an interrupt to each core.
162  */
163 static unsigned int rss_cpus;
164 module_param(rss_cpus, uint, 0444);
165 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
166 
167 static bool phy_flash_cfg;
168 module_param(phy_flash_cfg, bool, 0644);
169 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
170 
171 static unsigned irq_adapt_low_thresh = 8000;
172 module_param(irq_adapt_low_thresh, uint, 0644);
173 MODULE_PARM_DESC(irq_adapt_low_thresh,
174 		 "Threshold score for reducing IRQ moderation");
175 
176 static unsigned irq_adapt_high_thresh = 16000;
177 module_param(irq_adapt_high_thresh, uint, 0644);
178 MODULE_PARM_DESC(irq_adapt_high_thresh,
179 		 "Threshold score for increasing IRQ moderation");
180 
181 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
182 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
183 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
184 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
185 module_param(debug, uint, 0);
186 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
187 
188 /**************************************************************************
189  *
190  * Utility functions and prototypes
191  *
192  *************************************************************************/
193 
194 static void efx_start_interrupts(struct efx_nic *efx, bool may_keep_eventq);
195 static void efx_stop_interrupts(struct efx_nic *efx, bool may_keep_eventq);
196 static void efx_remove_channel(struct efx_channel *channel);
197 static void efx_remove_channels(struct efx_nic *efx);
198 static const struct efx_channel_type efx_default_channel_type;
199 static void efx_remove_port(struct efx_nic *efx);
200 static void efx_init_napi_channel(struct efx_channel *channel);
201 static void efx_fini_napi(struct efx_nic *efx);
202 static void efx_fini_napi_channel(struct efx_channel *channel);
203 static void efx_fini_struct(struct efx_nic *efx);
204 static void efx_start_all(struct efx_nic *efx);
205 static void efx_stop_all(struct efx_nic *efx);
206 
207 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
208 	do {						\
209 		if ((efx->state == STATE_READY) ||	\
210 		    (efx->state == STATE_RECOVERY) ||	\
211 		    (efx->state == STATE_DISABLED))	\
212 			ASSERT_RTNL();			\
213 	} while (0)
214 
215 static int efx_check_disabled(struct efx_nic *efx)
216 {
217 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
218 		netif_err(efx, drv, efx->net_dev,
219 			  "device is disabled due to earlier errors\n");
220 		return -EIO;
221 	}
222 	return 0;
223 }
224 
225 /**************************************************************************
226  *
227  * Event queue processing
228  *
229  *************************************************************************/
230 
231 /* Process channel's event queue
232  *
233  * This function is responsible for processing the event queue of a
234  * single channel.  The caller must guarantee that this function will
235  * never be concurrently called more than once on the same channel,
236  * though different channels may be being processed concurrently.
237  */
238 static int efx_process_channel(struct efx_channel *channel, int budget)
239 {
240 	int spent;
241 
242 	if (unlikely(!channel->enabled))
243 		return 0;
244 
245 	spent = efx_nic_process_eventq(channel, budget);
246 	if (spent && efx_channel_has_rx_queue(channel)) {
247 		struct efx_rx_queue *rx_queue =
248 			efx_channel_get_rx_queue(channel);
249 
250 		efx_rx_flush_packet(channel);
251 		if (rx_queue->enabled)
252 			efx_fast_push_rx_descriptors(rx_queue);
253 	}
254 
255 	return spent;
256 }
257 
258 /* Mark channel as finished processing
259  *
260  * Note that since we will not receive further interrupts for this
261  * channel before we finish processing and call the eventq_read_ack()
262  * method, there is no need to use the interrupt hold-off timers.
263  */
264 static inline void efx_channel_processed(struct efx_channel *channel)
265 {
266 	/* The interrupt handler for this channel may set work_pending
267 	 * as soon as we acknowledge the events we've seen.  Make sure
268 	 * it's cleared before then. */
269 	channel->work_pending = false;
270 	smp_wmb();
271 
272 	efx_nic_eventq_read_ack(channel);
273 }
274 
275 /* NAPI poll handler
276  *
277  * NAPI guarantees serialisation of polls of the same device, which
278  * provides the guarantee required by efx_process_channel().
279  */
280 static int efx_poll(struct napi_struct *napi, int budget)
281 {
282 	struct efx_channel *channel =
283 		container_of(napi, struct efx_channel, napi_str);
284 	struct efx_nic *efx = channel->efx;
285 	int spent;
286 
287 	netif_vdbg(efx, intr, efx->net_dev,
288 		   "channel %d NAPI poll executing on CPU %d\n",
289 		   channel->channel, raw_smp_processor_id());
290 
291 	spent = efx_process_channel(channel, budget);
292 
293 	if (spent < budget) {
294 		if (efx_channel_has_rx_queue(channel) &&
295 		    efx->irq_rx_adaptive &&
296 		    unlikely(++channel->irq_count == 1000)) {
297 			if (unlikely(channel->irq_mod_score <
298 				     irq_adapt_low_thresh)) {
299 				if (channel->irq_moderation > 1) {
300 					channel->irq_moderation -= 1;
301 					efx->type->push_irq_moderation(channel);
302 				}
303 			} else if (unlikely(channel->irq_mod_score >
304 					    irq_adapt_high_thresh)) {
305 				if (channel->irq_moderation <
306 				    efx->irq_rx_moderation) {
307 					channel->irq_moderation += 1;
308 					efx->type->push_irq_moderation(channel);
309 				}
310 			}
311 			channel->irq_count = 0;
312 			channel->irq_mod_score = 0;
313 		}
314 
315 		efx_filter_rfs_expire(channel);
316 
317 		/* There is no race here; although napi_disable() will
318 		 * only wait for napi_complete(), this isn't a problem
319 		 * since efx_channel_processed() will have no effect if
320 		 * interrupts have already been disabled.
321 		 */
322 		napi_complete(napi);
323 		efx_channel_processed(channel);
324 	}
325 
326 	return spent;
327 }
328 
329 /* Process the eventq of the specified channel immediately on this CPU
330  *
331  * Disable hardware generated interrupts, wait for any existing
332  * processing to finish, then directly poll (and ack ) the eventq.
333  * Finally reenable NAPI and interrupts.
334  *
335  * This is for use only during a loopback self-test.  It must not
336  * deliver any packets up the stack as this can result in deadlock.
337  */
338 void efx_process_channel_now(struct efx_channel *channel)
339 {
340 	struct efx_nic *efx = channel->efx;
341 
342 	BUG_ON(channel->channel >= efx->n_channels);
343 	BUG_ON(!channel->enabled);
344 	BUG_ON(!efx->loopback_selftest);
345 
346 	/* Disable interrupts and wait for ISRs to complete */
347 	efx_nic_disable_interrupts(efx);
348 	if (efx->legacy_irq) {
349 		synchronize_irq(efx->legacy_irq);
350 		efx->legacy_irq_enabled = false;
351 	}
352 	if (channel->irq)
353 		synchronize_irq(channel->irq);
354 
355 	/* Wait for any NAPI processing to complete */
356 	napi_disable(&channel->napi_str);
357 
358 	/* Poll the channel */
359 	efx_process_channel(channel, channel->eventq_mask + 1);
360 
361 	/* Ack the eventq. This may cause an interrupt to be generated
362 	 * when they are reenabled */
363 	efx_channel_processed(channel);
364 
365 	napi_enable(&channel->napi_str);
366 	if (efx->legacy_irq)
367 		efx->legacy_irq_enabled = true;
368 	efx_nic_enable_interrupts(efx);
369 }
370 
371 /* Create event queue
372  * Event queue memory allocations are done only once.  If the channel
373  * is reset, the memory buffer will be reused; this guards against
374  * errors during channel reset and also simplifies interrupt handling.
375  */
376 static int efx_probe_eventq(struct efx_channel *channel)
377 {
378 	struct efx_nic *efx = channel->efx;
379 	unsigned long entries;
380 
381 	netif_dbg(efx, probe, efx->net_dev,
382 		  "chan %d create event queue\n", channel->channel);
383 
384 	/* Build an event queue with room for one event per tx and rx buffer,
385 	 * plus some extra for link state events and MCDI completions. */
386 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
387 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
388 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
389 
390 	return efx_nic_probe_eventq(channel);
391 }
392 
393 /* Prepare channel's event queue */
394 static void efx_init_eventq(struct efx_channel *channel)
395 {
396 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
397 		  "chan %d init event queue\n", channel->channel);
398 
399 	channel->eventq_read_ptr = 0;
400 
401 	efx_nic_init_eventq(channel);
402 }
403 
404 /* Enable event queue processing and NAPI */
405 static void efx_start_eventq(struct efx_channel *channel)
406 {
407 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
408 		  "chan %d start event queue\n", channel->channel);
409 
410 	/* The interrupt handler for this channel may set work_pending
411 	 * as soon as we enable it.  Make sure it's cleared before
412 	 * then.  Similarly, make sure it sees the enabled flag set.
413 	 */
414 	channel->work_pending = false;
415 	channel->enabled = true;
416 	smp_wmb();
417 
418 	napi_enable(&channel->napi_str);
419 	efx_nic_eventq_read_ack(channel);
420 }
421 
422 /* Disable event queue processing and NAPI */
423 static void efx_stop_eventq(struct efx_channel *channel)
424 {
425 	if (!channel->enabled)
426 		return;
427 
428 	napi_disable(&channel->napi_str);
429 	channel->enabled = false;
430 }
431 
432 static void efx_fini_eventq(struct efx_channel *channel)
433 {
434 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
435 		  "chan %d fini event queue\n", channel->channel);
436 
437 	efx_nic_fini_eventq(channel);
438 }
439 
440 static void efx_remove_eventq(struct efx_channel *channel)
441 {
442 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
443 		  "chan %d remove event queue\n", channel->channel);
444 
445 	efx_nic_remove_eventq(channel);
446 }
447 
448 /**************************************************************************
449  *
450  * Channel handling
451  *
452  *************************************************************************/
453 
454 /* Allocate and initialise a channel structure. */
455 static struct efx_channel *
456 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
457 {
458 	struct efx_channel *channel;
459 	struct efx_rx_queue *rx_queue;
460 	struct efx_tx_queue *tx_queue;
461 	int j;
462 
463 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
464 	if (!channel)
465 		return NULL;
466 
467 	channel->efx = efx;
468 	channel->channel = i;
469 	channel->type = &efx_default_channel_type;
470 
471 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
472 		tx_queue = &channel->tx_queue[j];
473 		tx_queue->efx = efx;
474 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
475 		tx_queue->channel = channel;
476 	}
477 
478 	rx_queue = &channel->rx_queue;
479 	rx_queue->efx = efx;
480 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
481 		    (unsigned long)rx_queue);
482 
483 	return channel;
484 }
485 
486 /* Allocate and initialise a channel structure, copying parameters
487  * (but not resources) from an old channel structure.
488  */
489 static struct efx_channel *
490 efx_copy_channel(const struct efx_channel *old_channel)
491 {
492 	struct efx_channel *channel;
493 	struct efx_rx_queue *rx_queue;
494 	struct efx_tx_queue *tx_queue;
495 	int j;
496 
497 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
498 	if (!channel)
499 		return NULL;
500 
501 	*channel = *old_channel;
502 
503 	channel->napi_dev = NULL;
504 	memset(&channel->eventq, 0, sizeof(channel->eventq));
505 
506 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
507 		tx_queue = &channel->tx_queue[j];
508 		if (tx_queue->channel)
509 			tx_queue->channel = channel;
510 		tx_queue->buffer = NULL;
511 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
512 	}
513 
514 	rx_queue = &channel->rx_queue;
515 	rx_queue->buffer = NULL;
516 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
517 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
518 		    (unsigned long)rx_queue);
519 
520 	return channel;
521 }
522 
523 static int efx_probe_channel(struct efx_channel *channel)
524 {
525 	struct efx_tx_queue *tx_queue;
526 	struct efx_rx_queue *rx_queue;
527 	int rc;
528 
529 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
530 		  "creating channel %d\n", channel->channel);
531 
532 	rc = channel->type->pre_probe(channel);
533 	if (rc)
534 		goto fail;
535 
536 	rc = efx_probe_eventq(channel);
537 	if (rc)
538 		goto fail;
539 
540 	efx_for_each_channel_tx_queue(tx_queue, channel) {
541 		rc = efx_probe_tx_queue(tx_queue);
542 		if (rc)
543 			goto fail;
544 	}
545 
546 	efx_for_each_channel_rx_queue(rx_queue, channel) {
547 		rc = efx_probe_rx_queue(rx_queue);
548 		if (rc)
549 			goto fail;
550 	}
551 
552 	channel->n_rx_frm_trunc = 0;
553 
554 	return 0;
555 
556 fail:
557 	efx_remove_channel(channel);
558 	return rc;
559 }
560 
561 static void
562 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
563 {
564 	struct efx_nic *efx = channel->efx;
565 	const char *type;
566 	int number;
567 
568 	number = channel->channel;
569 	if (efx->tx_channel_offset == 0) {
570 		type = "";
571 	} else if (channel->channel < efx->tx_channel_offset) {
572 		type = "-rx";
573 	} else {
574 		type = "-tx";
575 		number -= efx->tx_channel_offset;
576 	}
577 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
578 }
579 
580 static void efx_set_channel_names(struct efx_nic *efx)
581 {
582 	struct efx_channel *channel;
583 
584 	efx_for_each_channel(channel, efx)
585 		channel->type->get_name(channel,
586 					efx->channel_name[channel->channel],
587 					sizeof(efx->channel_name[0]));
588 }
589 
590 static int efx_probe_channels(struct efx_nic *efx)
591 {
592 	struct efx_channel *channel;
593 	int rc;
594 
595 	/* Restart special buffer allocation */
596 	efx->next_buffer_table = 0;
597 
598 	/* Probe channels in reverse, so that any 'extra' channels
599 	 * use the start of the buffer table. This allows the traffic
600 	 * channels to be resized without moving them or wasting the
601 	 * entries before them.
602 	 */
603 	efx_for_each_channel_rev(channel, efx) {
604 		rc = efx_probe_channel(channel);
605 		if (rc) {
606 			netif_err(efx, probe, efx->net_dev,
607 				  "failed to create channel %d\n",
608 				  channel->channel);
609 			goto fail;
610 		}
611 	}
612 	efx_set_channel_names(efx);
613 
614 	return 0;
615 
616 fail:
617 	efx_remove_channels(efx);
618 	return rc;
619 }
620 
621 /* Channels are shutdown and reinitialised whilst the NIC is running
622  * to propagate configuration changes (mtu, checksum offload), or
623  * to clear hardware error conditions
624  */
625 static void efx_start_datapath(struct efx_nic *efx)
626 {
627 	bool old_rx_scatter = efx->rx_scatter;
628 	struct efx_tx_queue *tx_queue;
629 	struct efx_rx_queue *rx_queue;
630 	struct efx_channel *channel;
631 	size_t rx_buf_len;
632 
633 	/* Calculate the rx buffer allocation parameters required to
634 	 * support the current MTU, including padding for header
635 	 * alignment and overruns.
636 	 */
637 	efx->rx_dma_len = (efx->type->rx_buffer_hash_size +
638 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
639 			   efx->type->rx_buffer_padding);
640 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
641 		      NET_IP_ALIGN + efx->rx_dma_len);
642 	if (rx_buf_len <= PAGE_SIZE) {
643 		efx->rx_scatter = false;
644 		efx->rx_buffer_order = 0;
645 	} else if (efx->type->can_rx_scatter) {
646 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
647 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
648 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
649 				       EFX_RX_BUF_ALIGNMENT) >
650 			     PAGE_SIZE);
651 		efx->rx_scatter = true;
652 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
653 		efx->rx_buffer_order = 0;
654 	} else {
655 		efx->rx_scatter = false;
656 		efx->rx_buffer_order = get_order(rx_buf_len);
657 	}
658 
659 	efx_rx_config_page_split(efx);
660 	if (efx->rx_buffer_order)
661 		netif_dbg(efx, drv, efx->net_dev,
662 			  "RX buf len=%u; page order=%u batch=%u\n",
663 			  efx->rx_dma_len, efx->rx_buffer_order,
664 			  efx->rx_pages_per_batch);
665 	else
666 		netif_dbg(efx, drv, efx->net_dev,
667 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
668 			  efx->rx_dma_len, efx->rx_page_buf_step,
669 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
670 
671 	/* RX filters also have scatter-enabled flags */
672 	if (efx->rx_scatter != old_rx_scatter)
673 		efx_filter_update_rx_scatter(efx);
674 
675 	/* We must keep at least one descriptor in a TX ring empty.
676 	 * We could avoid this when the queue size does not exactly
677 	 * match the hardware ring size, but it's not that important.
678 	 * Therefore we stop the queue when one more skb might fill
679 	 * the ring completely.  We wake it when half way back to
680 	 * empty.
681 	 */
682 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
683 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
684 
685 	/* Initialise the channels */
686 	efx_for_each_channel(channel, efx) {
687 		efx_for_each_channel_tx_queue(tx_queue, channel)
688 			efx_init_tx_queue(tx_queue);
689 
690 		efx_for_each_channel_rx_queue(rx_queue, channel) {
691 			efx_init_rx_queue(rx_queue);
692 			efx_nic_generate_fill_event(rx_queue);
693 		}
694 
695 		WARN_ON(channel->rx_pkt_n_frags);
696 	}
697 
698 	if (netif_device_present(efx->net_dev))
699 		netif_tx_wake_all_queues(efx->net_dev);
700 }
701 
702 static void efx_stop_datapath(struct efx_nic *efx)
703 {
704 	struct efx_channel *channel;
705 	struct efx_tx_queue *tx_queue;
706 	struct efx_rx_queue *rx_queue;
707 	struct pci_dev *dev = efx->pci_dev;
708 	int rc;
709 
710 	EFX_ASSERT_RESET_SERIALISED(efx);
711 	BUG_ON(efx->port_enabled);
712 
713 	/* Only perform flush if dma is enabled */
714 	if (dev->is_busmaster && efx->state != STATE_RECOVERY) {
715 		rc = efx_nic_flush_queues(efx);
716 
717 		if (rc && EFX_WORKAROUND_7803(efx)) {
718 			/* Schedule a reset to recover from the flush failure. The
719 			 * descriptor caches reference memory we're about to free,
720 			 * but falcon_reconfigure_mac_wrapper() won't reconnect
721 			 * the MACs because of the pending reset. */
722 			netif_err(efx, drv, efx->net_dev,
723 				  "Resetting to recover from flush failure\n");
724 			efx_schedule_reset(efx, RESET_TYPE_ALL);
725 		} else if (rc) {
726 			netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
727 		} else {
728 			netif_dbg(efx, drv, efx->net_dev,
729 				  "successfully flushed all queues\n");
730 		}
731 	}
732 
733 	efx_for_each_channel(channel, efx) {
734 		/* RX packet processing is pipelined, so wait for the
735 		 * NAPI handler to complete.  At least event queue 0
736 		 * might be kept active by non-data events, so don't
737 		 * use napi_synchronize() but actually disable NAPI
738 		 * temporarily.
739 		 */
740 		if (efx_channel_has_rx_queue(channel)) {
741 			efx_stop_eventq(channel);
742 			efx_start_eventq(channel);
743 		}
744 
745 		efx_for_each_channel_rx_queue(rx_queue, channel)
746 			efx_fini_rx_queue(rx_queue);
747 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
748 			efx_fini_tx_queue(tx_queue);
749 	}
750 }
751 
752 static void efx_remove_channel(struct efx_channel *channel)
753 {
754 	struct efx_tx_queue *tx_queue;
755 	struct efx_rx_queue *rx_queue;
756 
757 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
758 		  "destroy chan %d\n", channel->channel);
759 
760 	efx_for_each_channel_rx_queue(rx_queue, channel)
761 		efx_remove_rx_queue(rx_queue);
762 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
763 		efx_remove_tx_queue(tx_queue);
764 	efx_remove_eventq(channel);
765 	channel->type->post_remove(channel);
766 }
767 
768 static void efx_remove_channels(struct efx_nic *efx)
769 {
770 	struct efx_channel *channel;
771 
772 	efx_for_each_channel(channel, efx)
773 		efx_remove_channel(channel);
774 }
775 
776 int
777 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
778 {
779 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
780 	u32 old_rxq_entries, old_txq_entries;
781 	unsigned i, next_buffer_table = 0;
782 	int rc;
783 
784 	rc = efx_check_disabled(efx);
785 	if (rc)
786 		return rc;
787 
788 	/* Not all channels should be reallocated. We must avoid
789 	 * reallocating their buffer table entries.
790 	 */
791 	efx_for_each_channel(channel, efx) {
792 		struct efx_rx_queue *rx_queue;
793 		struct efx_tx_queue *tx_queue;
794 
795 		if (channel->type->copy)
796 			continue;
797 		next_buffer_table = max(next_buffer_table,
798 					channel->eventq.index +
799 					channel->eventq.entries);
800 		efx_for_each_channel_rx_queue(rx_queue, channel)
801 			next_buffer_table = max(next_buffer_table,
802 						rx_queue->rxd.index +
803 						rx_queue->rxd.entries);
804 		efx_for_each_channel_tx_queue(tx_queue, channel)
805 			next_buffer_table = max(next_buffer_table,
806 						tx_queue->txd.index +
807 						tx_queue->txd.entries);
808 	}
809 
810 	efx_device_detach_sync(efx);
811 	efx_stop_all(efx);
812 	efx_stop_interrupts(efx, true);
813 
814 	/* Clone channels (where possible) */
815 	memset(other_channel, 0, sizeof(other_channel));
816 	for (i = 0; i < efx->n_channels; i++) {
817 		channel = efx->channel[i];
818 		if (channel->type->copy)
819 			channel = channel->type->copy(channel);
820 		if (!channel) {
821 			rc = -ENOMEM;
822 			goto out;
823 		}
824 		other_channel[i] = channel;
825 	}
826 
827 	/* Swap entry counts and channel pointers */
828 	old_rxq_entries = efx->rxq_entries;
829 	old_txq_entries = efx->txq_entries;
830 	efx->rxq_entries = rxq_entries;
831 	efx->txq_entries = txq_entries;
832 	for (i = 0; i < efx->n_channels; i++) {
833 		channel = efx->channel[i];
834 		efx->channel[i] = other_channel[i];
835 		other_channel[i] = channel;
836 	}
837 
838 	/* Restart buffer table allocation */
839 	efx->next_buffer_table = next_buffer_table;
840 
841 	for (i = 0; i < efx->n_channels; i++) {
842 		channel = efx->channel[i];
843 		if (!channel->type->copy)
844 			continue;
845 		rc = efx_probe_channel(channel);
846 		if (rc)
847 			goto rollback;
848 		efx_init_napi_channel(efx->channel[i]);
849 	}
850 
851 out:
852 	/* Destroy unused channel structures */
853 	for (i = 0; i < efx->n_channels; i++) {
854 		channel = other_channel[i];
855 		if (channel && channel->type->copy) {
856 			efx_fini_napi_channel(channel);
857 			efx_remove_channel(channel);
858 			kfree(channel);
859 		}
860 	}
861 
862 	efx_start_interrupts(efx, true);
863 	efx_start_all(efx);
864 	netif_device_attach(efx->net_dev);
865 	return rc;
866 
867 rollback:
868 	/* Swap back */
869 	efx->rxq_entries = old_rxq_entries;
870 	efx->txq_entries = old_txq_entries;
871 	for (i = 0; i < efx->n_channels; i++) {
872 		channel = efx->channel[i];
873 		efx->channel[i] = other_channel[i];
874 		other_channel[i] = channel;
875 	}
876 	goto out;
877 }
878 
879 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
880 {
881 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
882 }
883 
884 static const struct efx_channel_type efx_default_channel_type = {
885 	.pre_probe		= efx_channel_dummy_op_int,
886 	.post_remove		= efx_channel_dummy_op_void,
887 	.get_name		= efx_get_channel_name,
888 	.copy			= efx_copy_channel,
889 	.keep_eventq		= false,
890 };
891 
892 int efx_channel_dummy_op_int(struct efx_channel *channel)
893 {
894 	return 0;
895 }
896 
897 void efx_channel_dummy_op_void(struct efx_channel *channel)
898 {
899 }
900 
901 /**************************************************************************
902  *
903  * Port handling
904  *
905  **************************************************************************/
906 
907 /* This ensures that the kernel is kept informed (via
908  * netif_carrier_on/off) of the link status, and also maintains the
909  * link status's stop on the port's TX queue.
910  */
911 void efx_link_status_changed(struct efx_nic *efx)
912 {
913 	struct efx_link_state *link_state = &efx->link_state;
914 
915 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
916 	 * that no events are triggered between unregister_netdev() and the
917 	 * driver unloading. A more general condition is that NETDEV_CHANGE
918 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
919 	if (!netif_running(efx->net_dev))
920 		return;
921 
922 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
923 		efx->n_link_state_changes++;
924 
925 		if (link_state->up)
926 			netif_carrier_on(efx->net_dev);
927 		else
928 			netif_carrier_off(efx->net_dev);
929 	}
930 
931 	/* Status message for kernel log */
932 	if (link_state->up)
933 		netif_info(efx, link, efx->net_dev,
934 			   "link up at %uMbps %s-duplex (MTU %d)%s\n",
935 			   link_state->speed, link_state->fd ? "full" : "half",
936 			   efx->net_dev->mtu,
937 			   (efx->promiscuous ? " [PROMISC]" : ""));
938 	else
939 		netif_info(efx, link, efx->net_dev, "link down\n");
940 }
941 
942 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
943 {
944 	efx->link_advertising = advertising;
945 	if (advertising) {
946 		if (advertising & ADVERTISED_Pause)
947 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
948 		else
949 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
950 		if (advertising & ADVERTISED_Asym_Pause)
951 			efx->wanted_fc ^= EFX_FC_TX;
952 	}
953 }
954 
955 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
956 {
957 	efx->wanted_fc = wanted_fc;
958 	if (efx->link_advertising) {
959 		if (wanted_fc & EFX_FC_RX)
960 			efx->link_advertising |= (ADVERTISED_Pause |
961 						  ADVERTISED_Asym_Pause);
962 		else
963 			efx->link_advertising &= ~(ADVERTISED_Pause |
964 						   ADVERTISED_Asym_Pause);
965 		if (wanted_fc & EFX_FC_TX)
966 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
967 	}
968 }
969 
970 static void efx_fini_port(struct efx_nic *efx);
971 
972 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
973  * the MAC appropriately. All other PHY configuration changes are pushed
974  * through phy_op->set_settings(), and pushed asynchronously to the MAC
975  * through efx_monitor().
976  *
977  * Callers must hold the mac_lock
978  */
979 int __efx_reconfigure_port(struct efx_nic *efx)
980 {
981 	enum efx_phy_mode phy_mode;
982 	int rc;
983 
984 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
985 
986 	/* Serialise the promiscuous flag with efx_set_rx_mode. */
987 	netif_addr_lock_bh(efx->net_dev);
988 	netif_addr_unlock_bh(efx->net_dev);
989 
990 	/* Disable PHY transmit in mac level loopbacks */
991 	phy_mode = efx->phy_mode;
992 	if (LOOPBACK_INTERNAL(efx))
993 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
994 	else
995 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
996 
997 	rc = efx->type->reconfigure_port(efx);
998 
999 	if (rc)
1000 		efx->phy_mode = phy_mode;
1001 
1002 	return rc;
1003 }
1004 
1005 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1006  * disabled. */
1007 int efx_reconfigure_port(struct efx_nic *efx)
1008 {
1009 	int rc;
1010 
1011 	EFX_ASSERT_RESET_SERIALISED(efx);
1012 
1013 	mutex_lock(&efx->mac_lock);
1014 	rc = __efx_reconfigure_port(efx);
1015 	mutex_unlock(&efx->mac_lock);
1016 
1017 	return rc;
1018 }
1019 
1020 /* Asynchronous work item for changing MAC promiscuity and multicast
1021  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1022  * MAC directly. */
1023 static void efx_mac_work(struct work_struct *data)
1024 {
1025 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1026 
1027 	mutex_lock(&efx->mac_lock);
1028 	if (efx->port_enabled)
1029 		efx->type->reconfigure_mac(efx);
1030 	mutex_unlock(&efx->mac_lock);
1031 }
1032 
1033 static int efx_probe_port(struct efx_nic *efx)
1034 {
1035 	int rc;
1036 
1037 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1038 
1039 	if (phy_flash_cfg)
1040 		efx->phy_mode = PHY_MODE_SPECIAL;
1041 
1042 	/* Connect up MAC/PHY operations table */
1043 	rc = efx->type->probe_port(efx);
1044 	if (rc)
1045 		return rc;
1046 
1047 	/* Initialise MAC address to permanent address */
1048 	memcpy(efx->net_dev->dev_addr, efx->net_dev->perm_addr, ETH_ALEN);
1049 
1050 	return 0;
1051 }
1052 
1053 static int efx_init_port(struct efx_nic *efx)
1054 {
1055 	int rc;
1056 
1057 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1058 
1059 	mutex_lock(&efx->mac_lock);
1060 
1061 	rc = efx->phy_op->init(efx);
1062 	if (rc)
1063 		goto fail1;
1064 
1065 	efx->port_initialized = true;
1066 
1067 	/* Reconfigure the MAC before creating dma queues (required for
1068 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1069 	efx->type->reconfigure_mac(efx);
1070 
1071 	/* Ensure the PHY advertises the correct flow control settings */
1072 	rc = efx->phy_op->reconfigure(efx);
1073 	if (rc)
1074 		goto fail2;
1075 
1076 	mutex_unlock(&efx->mac_lock);
1077 	return 0;
1078 
1079 fail2:
1080 	efx->phy_op->fini(efx);
1081 fail1:
1082 	mutex_unlock(&efx->mac_lock);
1083 	return rc;
1084 }
1085 
1086 static void efx_start_port(struct efx_nic *efx)
1087 {
1088 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1089 	BUG_ON(efx->port_enabled);
1090 
1091 	mutex_lock(&efx->mac_lock);
1092 	efx->port_enabled = true;
1093 
1094 	/* efx_mac_work() might have been scheduled after efx_stop_port(),
1095 	 * and then cancelled by efx_flush_all() */
1096 	efx->type->reconfigure_mac(efx);
1097 
1098 	mutex_unlock(&efx->mac_lock);
1099 }
1100 
1101 /* Prevent efx_mac_work() and efx_monitor() from working */
1102 static void efx_stop_port(struct efx_nic *efx)
1103 {
1104 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1105 
1106 	mutex_lock(&efx->mac_lock);
1107 	efx->port_enabled = false;
1108 	mutex_unlock(&efx->mac_lock);
1109 
1110 	/* Serialise against efx_set_multicast_list() */
1111 	netif_addr_lock_bh(efx->net_dev);
1112 	netif_addr_unlock_bh(efx->net_dev);
1113 }
1114 
1115 static void efx_fini_port(struct efx_nic *efx)
1116 {
1117 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1118 
1119 	if (!efx->port_initialized)
1120 		return;
1121 
1122 	efx->phy_op->fini(efx);
1123 	efx->port_initialized = false;
1124 
1125 	efx->link_state.up = false;
1126 	efx_link_status_changed(efx);
1127 }
1128 
1129 static void efx_remove_port(struct efx_nic *efx)
1130 {
1131 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1132 
1133 	efx->type->remove_port(efx);
1134 }
1135 
1136 /**************************************************************************
1137  *
1138  * NIC handling
1139  *
1140  **************************************************************************/
1141 
1142 /* This configures the PCI device to enable I/O and DMA. */
1143 static int efx_init_io(struct efx_nic *efx)
1144 {
1145 	struct pci_dev *pci_dev = efx->pci_dev;
1146 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1147 	int rc;
1148 
1149 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1150 
1151 	rc = pci_enable_device(pci_dev);
1152 	if (rc) {
1153 		netif_err(efx, probe, efx->net_dev,
1154 			  "failed to enable PCI device\n");
1155 		goto fail1;
1156 	}
1157 
1158 	pci_set_master(pci_dev);
1159 
1160 	/* Set the PCI DMA mask.  Try all possibilities from our
1161 	 * genuine mask down to 32 bits, because some architectures
1162 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1163 	 * masks event though they reject 46 bit masks.
1164 	 */
1165 	while (dma_mask > 0x7fffffffUL) {
1166 		if (dma_supported(&pci_dev->dev, dma_mask)) {
1167 			rc = dma_set_mask(&pci_dev->dev, dma_mask);
1168 			if (rc == 0)
1169 				break;
1170 		}
1171 		dma_mask >>= 1;
1172 	}
1173 	if (rc) {
1174 		netif_err(efx, probe, efx->net_dev,
1175 			  "could not find a suitable DMA mask\n");
1176 		goto fail2;
1177 	}
1178 	netif_dbg(efx, probe, efx->net_dev,
1179 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1180 	rc = dma_set_coherent_mask(&pci_dev->dev, dma_mask);
1181 	if (rc) {
1182 		/* dma_set_coherent_mask() is not *allowed* to
1183 		 * fail with a mask that dma_set_mask() accepted,
1184 		 * but just in case...
1185 		 */
1186 		netif_err(efx, probe, efx->net_dev,
1187 			  "failed to set consistent DMA mask\n");
1188 		goto fail2;
1189 	}
1190 
1191 	efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1192 	rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1193 	if (rc) {
1194 		netif_err(efx, probe, efx->net_dev,
1195 			  "request for memory BAR failed\n");
1196 		rc = -EIO;
1197 		goto fail3;
1198 	}
1199 	efx->membase = ioremap_nocache(efx->membase_phys,
1200 				       efx->type->mem_map_size);
1201 	if (!efx->membase) {
1202 		netif_err(efx, probe, efx->net_dev,
1203 			  "could not map memory BAR at %llx+%x\n",
1204 			  (unsigned long long)efx->membase_phys,
1205 			  efx->type->mem_map_size);
1206 		rc = -ENOMEM;
1207 		goto fail4;
1208 	}
1209 	netif_dbg(efx, probe, efx->net_dev,
1210 		  "memory BAR at %llx+%x (virtual %p)\n",
1211 		  (unsigned long long)efx->membase_phys,
1212 		  efx->type->mem_map_size, efx->membase);
1213 
1214 	return 0;
1215 
1216  fail4:
1217 	pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1218  fail3:
1219 	efx->membase_phys = 0;
1220  fail2:
1221 	pci_disable_device(efx->pci_dev);
1222  fail1:
1223 	return rc;
1224 }
1225 
1226 static void efx_fini_io(struct efx_nic *efx)
1227 {
1228 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1229 
1230 	if (efx->membase) {
1231 		iounmap(efx->membase);
1232 		efx->membase = NULL;
1233 	}
1234 
1235 	if (efx->membase_phys) {
1236 		pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1237 		efx->membase_phys = 0;
1238 	}
1239 
1240 	pci_disable_device(efx->pci_dev);
1241 }
1242 
1243 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1244 {
1245 	cpumask_var_t thread_mask;
1246 	unsigned int count;
1247 	int cpu;
1248 
1249 	if (rss_cpus) {
1250 		count = rss_cpus;
1251 	} else {
1252 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1253 			netif_warn(efx, probe, efx->net_dev,
1254 				   "RSS disabled due to allocation failure\n");
1255 			return 1;
1256 		}
1257 
1258 		count = 0;
1259 		for_each_online_cpu(cpu) {
1260 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1261 				++count;
1262 				cpumask_or(thread_mask, thread_mask,
1263 					   topology_thread_cpumask(cpu));
1264 			}
1265 		}
1266 
1267 		free_cpumask_var(thread_mask);
1268 	}
1269 
1270 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1271 	 * table entries that are inaccessible to VFs
1272 	 */
1273 	if (efx_sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1274 	    count > efx_vf_size(efx)) {
1275 		netif_warn(efx, probe, efx->net_dev,
1276 			   "Reducing number of RSS channels from %u to %u for "
1277 			   "VF support. Increase vf-msix-limit to use more "
1278 			   "channels on the PF.\n",
1279 			   count, efx_vf_size(efx));
1280 		count = efx_vf_size(efx);
1281 	}
1282 
1283 	return count;
1284 }
1285 
1286 /* Probe the number and type of interrupts we are able to obtain, and
1287  * the resulting numbers of channels and RX queues.
1288  */
1289 static int efx_probe_interrupts(struct efx_nic *efx)
1290 {
1291 	unsigned int max_channels =
1292 		min(efx->type->phys_addr_channels, EFX_MAX_CHANNELS);
1293 	unsigned int extra_channels = 0;
1294 	unsigned int i, j;
1295 	int rc;
1296 
1297 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1298 		if (efx->extra_channel_type[i])
1299 			++extra_channels;
1300 
1301 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1302 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1303 		unsigned int n_channels;
1304 
1305 		n_channels = efx_wanted_parallelism(efx);
1306 		if (separate_tx_channels)
1307 			n_channels *= 2;
1308 		n_channels += extra_channels;
1309 		n_channels = min(n_channels, max_channels);
1310 
1311 		for (i = 0; i < n_channels; i++)
1312 			xentries[i].entry = i;
1313 		rc = pci_enable_msix(efx->pci_dev, xentries, n_channels);
1314 		if (rc > 0) {
1315 			netif_err(efx, drv, efx->net_dev,
1316 				  "WARNING: Insufficient MSI-X vectors"
1317 				  " available (%d < %u).\n", rc, n_channels);
1318 			netif_err(efx, drv, efx->net_dev,
1319 				  "WARNING: Performance may be reduced.\n");
1320 			EFX_BUG_ON_PARANOID(rc >= n_channels);
1321 			n_channels = rc;
1322 			rc = pci_enable_msix(efx->pci_dev, xentries,
1323 					     n_channels);
1324 		}
1325 
1326 		if (rc == 0) {
1327 			efx->n_channels = n_channels;
1328 			if (n_channels > extra_channels)
1329 				n_channels -= extra_channels;
1330 			if (separate_tx_channels) {
1331 				efx->n_tx_channels = max(n_channels / 2, 1U);
1332 				efx->n_rx_channels = max(n_channels -
1333 							 efx->n_tx_channels,
1334 							 1U);
1335 			} else {
1336 				efx->n_tx_channels = n_channels;
1337 				efx->n_rx_channels = n_channels;
1338 			}
1339 			for (i = 0; i < efx->n_channels; i++)
1340 				efx_get_channel(efx, i)->irq =
1341 					xentries[i].vector;
1342 		} else {
1343 			/* Fall back to single channel MSI */
1344 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1345 			netif_err(efx, drv, efx->net_dev,
1346 				  "could not enable MSI-X\n");
1347 		}
1348 	}
1349 
1350 	/* Try single interrupt MSI */
1351 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1352 		efx->n_channels = 1;
1353 		efx->n_rx_channels = 1;
1354 		efx->n_tx_channels = 1;
1355 		rc = pci_enable_msi(efx->pci_dev);
1356 		if (rc == 0) {
1357 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1358 		} else {
1359 			netif_err(efx, drv, efx->net_dev,
1360 				  "could not enable MSI\n");
1361 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1362 		}
1363 	}
1364 
1365 	/* Assume legacy interrupts */
1366 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1367 		efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1368 		efx->n_rx_channels = 1;
1369 		efx->n_tx_channels = 1;
1370 		efx->legacy_irq = efx->pci_dev->irq;
1371 	}
1372 
1373 	/* Assign extra channels if possible */
1374 	j = efx->n_channels;
1375 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1376 		if (!efx->extra_channel_type[i])
1377 			continue;
1378 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1379 		    efx->n_channels <= extra_channels) {
1380 			efx->extra_channel_type[i]->handle_no_channel(efx);
1381 		} else {
1382 			--j;
1383 			efx_get_channel(efx, j)->type =
1384 				efx->extra_channel_type[i];
1385 		}
1386 	}
1387 
1388 	/* RSS might be usable on VFs even if it is disabled on the PF */
1389 	efx->rss_spread = ((efx->n_rx_channels > 1 || !efx_sriov_wanted(efx)) ?
1390 			   efx->n_rx_channels : efx_vf_size(efx));
1391 
1392 	return 0;
1393 }
1394 
1395 /* Enable interrupts, then probe and start the event queues */
1396 static void efx_start_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1397 {
1398 	struct efx_channel *channel;
1399 
1400 	BUG_ON(efx->state == STATE_DISABLED);
1401 
1402 	if (efx->eeh_disabled_legacy_irq) {
1403 		enable_irq(efx->legacy_irq);
1404 		efx->eeh_disabled_legacy_irq = false;
1405 	}
1406 	if (efx->legacy_irq)
1407 		efx->legacy_irq_enabled = true;
1408 	efx_nic_enable_interrupts(efx);
1409 
1410 	efx_for_each_channel(channel, efx) {
1411 		if (!channel->type->keep_eventq || !may_keep_eventq)
1412 			efx_init_eventq(channel);
1413 		efx_start_eventq(channel);
1414 	}
1415 
1416 	efx_mcdi_mode_event(efx);
1417 }
1418 
1419 static void efx_stop_interrupts(struct efx_nic *efx, bool may_keep_eventq)
1420 {
1421 	struct efx_channel *channel;
1422 
1423 	if (efx->state == STATE_DISABLED)
1424 		return;
1425 
1426 	efx_mcdi_mode_poll(efx);
1427 
1428 	efx_nic_disable_interrupts(efx);
1429 	if (efx->legacy_irq) {
1430 		synchronize_irq(efx->legacy_irq);
1431 		efx->legacy_irq_enabled = false;
1432 	}
1433 
1434 	efx_for_each_channel(channel, efx) {
1435 		if (channel->irq)
1436 			synchronize_irq(channel->irq);
1437 
1438 		efx_stop_eventq(channel);
1439 		if (!channel->type->keep_eventq || !may_keep_eventq)
1440 			efx_fini_eventq(channel);
1441 	}
1442 }
1443 
1444 static void efx_remove_interrupts(struct efx_nic *efx)
1445 {
1446 	struct efx_channel *channel;
1447 
1448 	/* Remove MSI/MSI-X interrupts */
1449 	efx_for_each_channel(channel, efx)
1450 		channel->irq = 0;
1451 	pci_disable_msi(efx->pci_dev);
1452 	pci_disable_msix(efx->pci_dev);
1453 
1454 	/* Remove legacy interrupt */
1455 	efx->legacy_irq = 0;
1456 }
1457 
1458 static void efx_set_channels(struct efx_nic *efx)
1459 {
1460 	struct efx_channel *channel;
1461 	struct efx_tx_queue *tx_queue;
1462 
1463 	efx->tx_channel_offset =
1464 		separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1465 
1466 	/* We need to mark which channels really have RX and TX
1467 	 * queues, and adjust the TX queue numbers if we have separate
1468 	 * RX-only and TX-only channels.
1469 	 */
1470 	efx_for_each_channel(channel, efx) {
1471 		if (channel->channel < efx->n_rx_channels)
1472 			channel->rx_queue.core_index = channel->channel;
1473 		else
1474 			channel->rx_queue.core_index = -1;
1475 
1476 		efx_for_each_channel_tx_queue(tx_queue, channel)
1477 			tx_queue->queue -= (efx->tx_channel_offset *
1478 					    EFX_TXQ_TYPES);
1479 	}
1480 }
1481 
1482 static int efx_probe_nic(struct efx_nic *efx)
1483 {
1484 	size_t i;
1485 	int rc;
1486 
1487 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1488 
1489 	/* Carry out hardware-type specific initialisation */
1490 	rc = efx->type->probe(efx);
1491 	if (rc)
1492 		return rc;
1493 
1494 	/* Determine the number of channels and queues by trying to hook
1495 	 * in MSI-X interrupts. */
1496 	rc = efx_probe_interrupts(efx);
1497 	if (rc)
1498 		goto fail;
1499 
1500 	efx->type->dimension_resources(efx);
1501 
1502 	if (efx->n_channels > 1)
1503 		get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1504 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1505 		efx->rx_indir_table[i] =
1506 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1507 
1508 	efx_set_channels(efx);
1509 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1510 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1511 
1512 	/* Initialise the interrupt moderation settings */
1513 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1514 				true);
1515 
1516 	return 0;
1517 
1518 fail:
1519 	efx->type->remove(efx);
1520 	return rc;
1521 }
1522 
1523 static void efx_remove_nic(struct efx_nic *efx)
1524 {
1525 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1526 
1527 	efx_remove_interrupts(efx);
1528 	efx->type->remove(efx);
1529 }
1530 
1531 /**************************************************************************
1532  *
1533  * NIC startup/shutdown
1534  *
1535  *************************************************************************/
1536 
1537 static int efx_probe_all(struct efx_nic *efx)
1538 {
1539 	int rc;
1540 
1541 	rc = efx_probe_nic(efx);
1542 	if (rc) {
1543 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1544 		goto fail1;
1545 	}
1546 
1547 	rc = efx_probe_port(efx);
1548 	if (rc) {
1549 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1550 		goto fail2;
1551 	}
1552 
1553 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1554 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1555 		rc = -EINVAL;
1556 		goto fail3;
1557 	}
1558 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1559 
1560 	rc = efx_probe_filters(efx);
1561 	if (rc) {
1562 		netif_err(efx, probe, efx->net_dev,
1563 			  "failed to create filter tables\n");
1564 		goto fail3;
1565 	}
1566 
1567 	rc = efx_probe_channels(efx);
1568 	if (rc)
1569 		goto fail4;
1570 
1571 	return 0;
1572 
1573  fail4:
1574 	efx_remove_filters(efx);
1575  fail3:
1576 	efx_remove_port(efx);
1577  fail2:
1578 	efx_remove_nic(efx);
1579  fail1:
1580 	return rc;
1581 }
1582 
1583 /* If the interface is supposed to be running but is not, start
1584  * the hardware and software data path, regular activity for the port
1585  * (MAC statistics, link polling, etc.) and schedule the port to be
1586  * reconfigured.  Interrupts must already be enabled.  This function
1587  * is safe to call multiple times, so long as the NIC is not disabled.
1588  * Requires the RTNL lock.
1589  */
1590 static void efx_start_all(struct efx_nic *efx)
1591 {
1592 	EFX_ASSERT_RESET_SERIALISED(efx);
1593 	BUG_ON(efx->state == STATE_DISABLED);
1594 
1595 	/* Check that it is appropriate to restart the interface. All
1596 	 * of these flags are safe to read under just the rtnl lock */
1597 	if (efx->port_enabled || !netif_running(efx->net_dev))
1598 		return;
1599 
1600 	efx_start_port(efx);
1601 	efx_start_datapath(efx);
1602 
1603 	/* Start the hardware monitor if there is one */
1604 	if (efx->type->monitor != NULL)
1605 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1606 				   efx_monitor_interval);
1607 
1608 	/* If link state detection is normally event-driven, we have
1609 	 * to poll now because we could have missed a change
1610 	 */
1611 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1612 		mutex_lock(&efx->mac_lock);
1613 		if (efx->phy_op->poll(efx))
1614 			efx_link_status_changed(efx);
1615 		mutex_unlock(&efx->mac_lock);
1616 	}
1617 
1618 	efx->type->start_stats(efx);
1619 }
1620 
1621 /* Flush all delayed work. Should only be called when no more delayed work
1622  * will be scheduled. This doesn't flush pending online resets (efx_reset),
1623  * since we're holding the rtnl_lock at this point. */
1624 static void efx_flush_all(struct efx_nic *efx)
1625 {
1626 	/* Make sure the hardware monitor and event self-test are stopped */
1627 	cancel_delayed_work_sync(&efx->monitor_work);
1628 	efx_selftest_async_cancel(efx);
1629 	/* Stop scheduled port reconfigurations */
1630 	cancel_work_sync(&efx->mac_work);
1631 }
1632 
1633 /* Quiesce the hardware and software data path, and regular activity
1634  * for the port without bringing the link down.  Safe to call multiple
1635  * times with the NIC in almost any state, but interrupts should be
1636  * enabled.  Requires the RTNL lock.
1637  */
1638 static void efx_stop_all(struct efx_nic *efx)
1639 {
1640 	EFX_ASSERT_RESET_SERIALISED(efx);
1641 
1642 	/* port_enabled can be read safely under the rtnl lock */
1643 	if (!efx->port_enabled)
1644 		return;
1645 
1646 	efx->type->stop_stats(efx);
1647 	efx_stop_port(efx);
1648 
1649 	/* Flush efx_mac_work(), refill_workqueue, monitor_work */
1650 	efx_flush_all(efx);
1651 
1652 	/* Stop the kernel transmit interface.  This is only valid if
1653 	 * the device is stopped or detached; otherwise the watchdog
1654 	 * may fire immediately.
1655 	 */
1656 	WARN_ON(netif_running(efx->net_dev) &&
1657 		netif_device_present(efx->net_dev));
1658 	netif_tx_disable(efx->net_dev);
1659 
1660 	efx_stop_datapath(efx);
1661 }
1662 
1663 static void efx_remove_all(struct efx_nic *efx)
1664 {
1665 	efx_remove_channels(efx);
1666 	efx_remove_filters(efx);
1667 	efx_remove_port(efx);
1668 	efx_remove_nic(efx);
1669 }
1670 
1671 /**************************************************************************
1672  *
1673  * Interrupt moderation
1674  *
1675  **************************************************************************/
1676 
1677 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1678 {
1679 	if (usecs == 0)
1680 		return 0;
1681 	if (usecs * 1000 < quantum_ns)
1682 		return 1; /* never round down to 0 */
1683 	return usecs * 1000 / quantum_ns;
1684 }
1685 
1686 /* Set interrupt moderation parameters */
1687 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1688 			    unsigned int rx_usecs, bool rx_adaptive,
1689 			    bool rx_may_override_tx)
1690 {
1691 	struct efx_channel *channel;
1692 	unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1693 						efx->timer_quantum_ns,
1694 						1000);
1695 	unsigned int tx_ticks;
1696 	unsigned int rx_ticks;
1697 
1698 	EFX_ASSERT_RESET_SERIALISED(efx);
1699 
1700 	if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1701 		return -EINVAL;
1702 
1703 	tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1704 	rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1705 
1706 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1707 	    !rx_may_override_tx) {
1708 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1709 			  "RX and TX IRQ moderation must be equal\n");
1710 		return -EINVAL;
1711 	}
1712 
1713 	efx->irq_rx_adaptive = rx_adaptive;
1714 	efx->irq_rx_moderation = rx_ticks;
1715 	efx_for_each_channel(channel, efx) {
1716 		if (efx_channel_has_rx_queue(channel))
1717 			channel->irq_moderation = rx_ticks;
1718 		else if (efx_channel_has_tx_queues(channel))
1719 			channel->irq_moderation = tx_ticks;
1720 	}
1721 
1722 	return 0;
1723 }
1724 
1725 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1726 			    unsigned int *rx_usecs, bool *rx_adaptive)
1727 {
1728 	/* We must round up when converting ticks to microseconds
1729 	 * because we round down when converting the other way.
1730 	 */
1731 
1732 	*rx_adaptive = efx->irq_rx_adaptive;
1733 	*rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1734 				 efx->timer_quantum_ns,
1735 				 1000);
1736 
1737 	/* If channels are shared between RX and TX, so is IRQ
1738 	 * moderation.  Otherwise, IRQ moderation is the same for all
1739 	 * TX channels and is not adaptive.
1740 	 */
1741 	if (efx->tx_channel_offset == 0)
1742 		*tx_usecs = *rx_usecs;
1743 	else
1744 		*tx_usecs = DIV_ROUND_UP(
1745 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1746 			efx->timer_quantum_ns,
1747 			1000);
1748 }
1749 
1750 /**************************************************************************
1751  *
1752  * Hardware monitor
1753  *
1754  **************************************************************************/
1755 
1756 /* Run periodically off the general workqueue */
1757 static void efx_monitor(struct work_struct *data)
1758 {
1759 	struct efx_nic *efx = container_of(data, struct efx_nic,
1760 					   monitor_work.work);
1761 
1762 	netif_vdbg(efx, timer, efx->net_dev,
1763 		   "hardware monitor executing on CPU %d\n",
1764 		   raw_smp_processor_id());
1765 	BUG_ON(efx->type->monitor == NULL);
1766 
1767 	/* If the mac_lock is already held then it is likely a port
1768 	 * reconfiguration is already in place, which will likely do
1769 	 * most of the work of monitor() anyway. */
1770 	if (mutex_trylock(&efx->mac_lock)) {
1771 		if (efx->port_enabled)
1772 			efx->type->monitor(efx);
1773 		mutex_unlock(&efx->mac_lock);
1774 	}
1775 
1776 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1777 			   efx_monitor_interval);
1778 }
1779 
1780 /**************************************************************************
1781  *
1782  * ioctls
1783  *
1784  *************************************************************************/
1785 
1786 /* Net device ioctl
1787  * Context: process, rtnl_lock() held.
1788  */
1789 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1790 {
1791 	struct efx_nic *efx = netdev_priv(net_dev);
1792 	struct mii_ioctl_data *data = if_mii(ifr);
1793 
1794 	if (cmd == SIOCSHWTSTAMP)
1795 		return efx_ptp_ioctl(efx, ifr, cmd);
1796 
1797 	/* Convert phy_id from older PRTAD/DEVAD format */
1798 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1799 	    (data->phy_id & 0xfc00) == 0x0400)
1800 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1801 
1802 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
1803 }
1804 
1805 /**************************************************************************
1806  *
1807  * NAPI interface
1808  *
1809  **************************************************************************/
1810 
1811 static void efx_init_napi_channel(struct efx_channel *channel)
1812 {
1813 	struct efx_nic *efx = channel->efx;
1814 
1815 	channel->napi_dev = efx->net_dev;
1816 	netif_napi_add(channel->napi_dev, &channel->napi_str,
1817 		       efx_poll, napi_weight);
1818 }
1819 
1820 static void efx_init_napi(struct efx_nic *efx)
1821 {
1822 	struct efx_channel *channel;
1823 
1824 	efx_for_each_channel(channel, efx)
1825 		efx_init_napi_channel(channel);
1826 }
1827 
1828 static void efx_fini_napi_channel(struct efx_channel *channel)
1829 {
1830 	if (channel->napi_dev)
1831 		netif_napi_del(&channel->napi_str);
1832 	channel->napi_dev = NULL;
1833 }
1834 
1835 static void efx_fini_napi(struct efx_nic *efx)
1836 {
1837 	struct efx_channel *channel;
1838 
1839 	efx_for_each_channel(channel, efx)
1840 		efx_fini_napi_channel(channel);
1841 }
1842 
1843 /**************************************************************************
1844  *
1845  * Kernel netpoll interface
1846  *
1847  *************************************************************************/
1848 
1849 #ifdef CONFIG_NET_POLL_CONTROLLER
1850 
1851 /* Although in the common case interrupts will be disabled, this is not
1852  * guaranteed. However, all our work happens inside the NAPI callback,
1853  * so no locking is required.
1854  */
1855 static void efx_netpoll(struct net_device *net_dev)
1856 {
1857 	struct efx_nic *efx = netdev_priv(net_dev);
1858 	struct efx_channel *channel;
1859 
1860 	efx_for_each_channel(channel, efx)
1861 		efx_schedule_channel(channel);
1862 }
1863 
1864 #endif
1865 
1866 /**************************************************************************
1867  *
1868  * Kernel net device interface
1869  *
1870  *************************************************************************/
1871 
1872 /* Context: process, rtnl_lock() held. */
1873 static int efx_net_open(struct net_device *net_dev)
1874 {
1875 	struct efx_nic *efx = netdev_priv(net_dev);
1876 	int rc;
1877 
1878 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
1879 		  raw_smp_processor_id());
1880 
1881 	rc = efx_check_disabled(efx);
1882 	if (rc)
1883 		return rc;
1884 	if (efx->phy_mode & PHY_MODE_SPECIAL)
1885 		return -EBUSY;
1886 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
1887 		return -EIO;
1888 
1889 	/* Notify the kernel of the link state polled during driver load,
1890 	 * before the monitor starts running */
1891 	efx_link_status_changed(efx);
1892 
1893 	efx_start_all(efx);
1894 	efx_selftest_async_start(efx);
1895 	return 0;
1896 }
1897 
1898 /* Context: process, rtnl_lock() held.
1899  * Note that the kernel will ignore our return code; this method
1900  * should really be a void.
1901  */
1902 static int efx_net_stop(struct net_device *net_dev)
1903 {
1904 	struct efx_nic *efx = netdev_priv(net_dev);
1905 
1906 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
1907 		  raw_smp_processor_id());
1908 
1909 	/* Stop the device and flush all the channels */
1910 	efx_stop_all(efx);
1911 
1912 	return 0;
1913 }
1914 
1915 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1916 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
1917 					       struct rtnl_link_stats64 *stats)
1918 {
1919 	struct efx_nic *efx = netdev_priv(net_dev);
1920 	struct efx_mac_stats *mac_stats = &efx->mac_stats;
1921 
1922 	spin_lock_bh(&efx->stats_lock);
1923 
1924 	efx->type->update_stats(efx);
1925 
1926 	stats->rx_packets = mac_stats->rx_packets;
1927 	stats->tx_packets = mac_stats->tx_packets;
1928 	stats->rx_bytes = mac_stats->rx_bytes;
1929 	stats->tx_bytes = mac_stats->tx_bytes;
1930 	stats->rx_dropped = efx->n_rx_nodesc_drop_cnt;
1931 	stats->multicast = mac_stats->rx_multicast;
1932 	stats->collisions = mac_stats->tx_collision;
1933 	stats->rx_length_errors = (mac_stats->rx_gtjumbo +
1934 				   mac_stats->rx_length_error);
1935 	stats->rx_crc_errors = mac_stats->rx_bad;
1936 	stats->rx_frame_errors = mac_stats->rx_align_error;
1937 	stats->rx_fifo_errors = mac_stats->rx_overflow;
1938 	stats->rx_missed_errors = mac_stats->rx_missed;
1939 	stats->tx_window_errors = mac_stats->tx_late_collision;
1940 
1941 	stats->rx_errors = (stats->rx_length_errors +
1942 			    stats->rx_crc_errors +
1943 			    stats->rx_frame_errors +
1944 			    mac_stats->rx_symbol_error);
1945 	stats->tx_errors = (stats->tx_window_errors +
1946 			    mac_stats->tx_bad);
1947 
1948 	spin_unlock_bh(&efx->stats_lock);
1949 
1950 	return stats;
1951 }
1952 
1953 /* Context: netif_tx_lock held, BHs disabled. */
1954 static void efx_watchdog(struct net_device *net_dev)
1955 {
1956 	struct efx_nic *efx = netdev_priv(net_dev);
1957 
1958 	netif_err(efx, tx_err, efx->net_dev,
1959 		  "TX stuck with port_enabled=%d: resetting channels\n",
1960 		  efx->port_enabled);
1961 
1962 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
1963 }
1964 
1965 
1966 /* Context: process, rtnl_lock() held. */
1967 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
1968 {
1969 	struct efx_nic *efx = netdev_priv(net_dev);
1970 	int rc;
1971 
1972 	rc = efx_check_disabled(efx);
1973 	if (rc)
1974 		return rc;
1975 	if (new_mtu > EFX_MAX_MTU)
1976 		return -EINVAL;
1977 
1978 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
1979 
1980 	efx_device_detach_sync(efx);
1981 	efx_stop_all(efx);
1982 
1983 	mutex_lock(&efx->mac_lock);
1984 	net_dev->mtu = new_mtu;
1985 	efx->type->reconfigure_mac(efx);
1986 	mutex_unlock(&efx->mac_lock);
1987 
1988 	efx_start_all(efx);
1989 	netif_device_attach(efx->net_dev);
1990 	return 0;
1991 }
1992 
1993 static int efx_set_mac_address(struct net_device *net_dev, void *data)
1994 {
1995 	struct efx_nic *efx = netdev_priv(net_dev);
1996 	struct sockaddr *addr = data;
1997 	char *new_addr = addr->sa_data;
1998 
1999 	if (!is_valid_ether_addr(new_addr)) {
2000 		netif_err(efx, drv, efx->net_dev,
2001 			  "invalid ethernet MAC address requested: %pM\n",
2002 			  new_addr);
2003 		return -EADDRNOTAVAIL;
2004 	}
2005 
2006 	memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
2007 	efx_sriov_mac_address_changed(efx);
2008 
2009 	/* Reconfigure the MAC */
2010 	mutex_lock(&efx->mac_lock);
2011 	efx->type->reconfigure_mac(efx);
2012 	mutex_unlock(&efx->mac_lock);
2013 
2014 	return 0;
2015 }
2016 
2017 /* Context: netif_addr_lock held, BHs disabled. */
2018 static void efx_set_rx_mode(struct net_device *net_dev)
2019 {
2020 	struct efx_nic *efx = netdev_priv(net_dev);
2021 	struct netdev_hw_addr *ha;
2022 	union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2023 	u32 crc;
2024 	int bit;
2025 
2026 	efx->promiscuous = !!(net_dev->flags & IFF_PROMISC);
2027 
2028 	/* Build multicast hash table */
2029 	if (efx->promiscuous || (net_dev->flags & IFF_ALLMULTI)) {
2030 		memset(mc_hash, 0xff, sizeof(*mc_hash));
2031 	} else {
2032 		memset(mc_hash, 0x00, sizeof(*mc_hash));
2033 		netdev_for_each_mc_addr(ha, net_dev) {
2034 			crc = ether_crc_le(ETH_ALEN, ha->addr);
2035 			bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2036 			__set_bit_le(bit, mc_hash);
2037 		}
2038 
2039 		/* Broadcast packets go through the multicast hash filter.
2040 		 * ether_crc_le() of the broadcast address is 0xbe2612ff
2041 		 * so we always add bit 0xff to the mask.
2042 		 */
2043 		__set_bit_le(0xff, mc_hash);
2044 	}
2045 
2046 	if (efx->port_enabled)
2047 		queue_work(efx->workqueue, &efx->mac_work);
2048 	/* Otherwise efx_start_port() will do this */
2049 }
2050 
2051 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2052 {
2053 	struct efx_nic *efx = netdev_priv(net_dev);
2054 
2055 	/* If disabling RX n-tuple filtering, clear existing filters */
2056 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
2057 		efx_filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2058 
2059 	return 0;
2060 }
2061 
2062 static const struct net_device_ops efx_netdev_ops = {
2063 	.ndo_open		= efx_net_open,
2064 	.ndo_stop		= efx_net_stop,
2065 	.ndo_get_stats64	= efx_net_stats,
2066 	.ndo_tx_timeout		= efx_watchdog,
2067 	.ndo_start_xmit		= efx_hard_start_xmit,
2068 	.ndo_validate_addr	= eth_validate_addr,
2069 	.ndo_do_ioctl		= efx_ioctl,
2070 	.ndo_change_mtu		= efx_change_mtu,
2071 	.ndo_set_mac_address	= efx_set_mac_address,
2072 	.ndo_set_rx_mode	= efx_set_rx_mode,
2073 	.ndo_set_features	= efx_set_features,
2074 #ifdef CONFIG_SFC_SRIOV
2075 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2076 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2077 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2078 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2079 #endif
2080 #ifdef CONFIG_NET_POLL_CONTROLLER
2081 	.ndo_poll_controller = efx_netpoll,
2082 #endif
2083 	.ndo_setup_tc		= efx_setup_tc,
2084 #ifdef CONFIG_RFS_ACCEL
2085 	.ndo_rx_flow_steer	= efx_filter_rfs,
2086 #endif
2087 };
2088 
2089 static void efx_update_name(struct efx_nic *efx)
2090 {
2091 	strcpy(efx->name, efx->net_dev->name);
2092 	efx_mtd_rename(efx);
2093 	efx_set_channel_names(efx);
2094 }
2095 
2096 static int efx_netdev_event(struct notifier_block *this,
2097 			    unsigned long event, void *ptr)
2098 {
2099 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2100 
2101 	if (net_dev->netdev_ops == &efx_netdev_ops &&
2102 	    event == NETDEV_CHANGENAME)
2103 		efx_update_name(netdev_priv(net_dev));
2104 
2105 	return NOTIFY_DONE;
2106 }
2107 
2108 static struct notifier_block efx_netdev_notifier = {
2109 	.notifier_call = efx_netdev_event,
2110 };
2111 
2112 static ssize_t
2113 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2114 {
2115 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2116 	return sprintf(buf, "%d\n", efx->phy_type);
2117 }
2118 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2119 
2120 static int efx_register_netdev(struct efx_nic *efx)
2121 {
2122 	struct net_device *net_dev = efx->net_dev;
2123 	struct efx_channel *channel;
2124 	int rc;
2125 
2126 	net_dev->watchdog_timeo = 5 * HZ;
2127 	net_dev->irq = efx->pci_dev->irq;
2128 	net_dev->netdev_ops = &efx_netdev_ops;
2129 	SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
2130 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2131 
2132 	rtnl_lock();
2133 
2134 	/* Enable resets to be scheduled and check whether any were
2135 	 * already requested.  If so, the NIC is probably hosed so we
2136 	 * abort.
2137 	 */
2138 	efx->state = STATE_READY;
2139 	smp_mb(); /* ensure we change state before checking reset_pending */
2140 	if (efx->reset_pending) {
2141 		netif_err(efx, probe, efx->net_dev,
2142 			  "aborting probe due to scheduled reset\n");
2143 		rc = -EIO;
2144 		goto fail_locked;
2145 	}
2146 
2147 	rc = dev_alloc_name(net_dev, net_dev->name);
2148 	if (rc < 0)
2149 		goto fail_locked;
2150 	efx_update_name(efx);
2151 
2152 	/* Always start with carrier off; PHY events will detect the link */
2153 	netif_carrier_off(net_dev);
2154 
2155 	rc = register_netdevice(net_dev);
2156 	if (rc)
2157 		goto fail_locked;
2158 
2159 	efx_for_each_channel(channel, efx) {
2160 		struct efx_tx_queue *tx_queue;
2161 		efx_for_each_channel_tx_queue(tx_queue, channel)
2162 			efx_init_tx_queue_core_txq(tx_queue);
2163 	}
2164 
2165 	rtnl_unlock();
2166 
2167 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2168 	if (rc) {
2169 		netif_err(efx, drv, efx->net_dev,
2170 			  "failed to init net dev attributes\n");
2171 		goto fail_registered;
2172 	}
2173 
2174 	return 0;
2175 
2176 fail_registered:
2177 	rtnl_lock();
2178 	unregister_netdevice(net_dev);
2179 fail_locked:
2180 	efx->state = STATE_UNINIT;
2181 	rtnl_unlock();
2182 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2183 	return rc;
2184 }
2185 
2186 static void efx_unregister_netdev(struct efx_nic *efx)
2187 {
2188 	struct efx_channel *channel;
2189 	struct efx_tx_queue *tx_queue;
2190 
2191 	if (!efx->net_dev)
2192 		return;
2193 
2194 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2195 
2196 	/* Free up any skbs still remaining. This has to happen before
2197 	 * we try to unregister the netdev as running their destructors
2198 	 * may be needed to get the device ref. count to 0. */
2199 	efx_for_each_channel(channel, efx) {
2200 		efx_for_each_channel_tx_queue(tx_queue, channel)
2201 			efx_release_tx_buffers(tx_queue);
2202 	}
2203 
2204 	strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2205 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2206 
2207 	rtnl_lock();
2208 	unregister_netdevice(efx->net_dev);
2209 	efx->state = STATE_UNINIT;
2210 	rtnl_unlock();
2211 }
2212 
2213 /**************************************************************************
2214  *
2215  * Device reset and suspend
2216  *
2217  **************************************************************************/
2218 
2219 /* Tears down the entire software state and most of the hardware state
2220  * before reset.  */
2221 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2222 {
2223 	EFX_ASSERT_RESET_SERIALISED(efx);
2224 
2225 	efx_stop_all(efx);
2226 	efx_stop_interrupts(efx, false);
2227 
2228 	mutex_lock(&efx->mac_lock);
2229 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2230 		efx->phy_op->fini(efx);
2231 	efx->type->fini(efx);
2232 }
2233 
2234 /* This function will always ensure that the locks acquired in
2235  * efx_reset_down() are released. A failure return code indicates
2236  * that we were unable to reinitialise the hardware, and the
2237  * driver should be disabled. If ok is false, then the rx and tx
2238  * engines are not restarted, pending a RESET_DISABLE. */
2239 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2240 {
2241 	int rc;
2242 
2243 	EFX_ASSERT_RESET_SERIALISED(efx);
2244 
2245 	rc = efx->type->init(efx);
2246 	if (rc) {
2247 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2248 		goto fail;
2249 	}
2250 
2251 	if (!ok)
2252 		goto fail;
2253 
2254 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2255 		rc = efx->phy_op->init(efx);
2256 		if (rc)
2257 			goto fail;
2258 		if (efx->phy_op->reconfigure(efx))
2259 			netif_err(efx, drv, efx->net_dev,
2260 				  "could not restore PHY settings\n");
2261 	}
2262 
2263 	efx->type->reconfigure_mac(efx);
2264 
2265 	efx_start_interrupts(efx, false);
2266 	efx_restore_filters(efx);
2267 	efx_sriov_reset(efx);
2268 
2269 	mutex_unlock(&efx->mac_lock);
2270 
2271 	efx_start_all(efx);
2272 
2273 	return 0;
2274 
2275 fail:
2276 	efx->port_initialized = false;
2277 
2278 	mutex_unlock(&efx->mac_lock);
2279 
2280 	return rc;
2281 }
2282 
2283 /* Reset the NIC using the specified method.  Note that the reset may
2284  * fail, in which case the card will be left in an unusable state.
2285  *
2286  * Caller must hold the rtnl_lock.
2287  */
2288 int efx_reset(struct efx_nic *efx, enum reset_type method)
2289 {
2290 	int rc, rc2;
2291 	bool disabled;
2292 
2293 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2294 		   RESET_TYPE(method));
2295 
2296 	efx_device_detach_sync(efx);
2297 	efx_reset_down(efx, method);
2298 
2299 	rc = efx->type->reset(efx, method);
2300 	if (rc) {
2301 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2302 		goto out;
2303 	}
2304 
2305 	/* Clear flags for the scopes we covered.  We assume the NIC and
2306 	 * driver are now quiescent so that there is no race here.
2307 	 */
2308 	efx->reset_pending &= -(1 << (method + 1));
2309 
2310 	/* Reinitialise bus-mastering, which may have been turned off before
2311 	 * the reset was scheduled. This is still appropriate, even in the
2312 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2313 	 * can respond to requests. */
2314 	pci_set_master(efx->pci_dev);
2315 
2316 out:
2317 	/* Leave device stopped if necessary */
2318 	disabled = rc ||
2319 		method == RESET_TYPE_DISABLE ||
2320 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2321 	rc2 = efx_reset_up(efx, method, !disabled);
2322 	if (rc2) {
2323 		disabled = true;
2324 		if (!rc)
2325 			rc = rc2;
2326 	}
2327 
2328 	if (disabled) {
2329 		dev_close(efx->net_dev);
2330 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2331 		efx->state = STATE_DISABLED;
2332 	} else {
2333 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2334 		netif_device_attach(efx->net_dev);
2335 	}
2336 	return rc;
2337 }
2338 
2339 /* Try recovery mechanisms.
2340  * For now only EEH is supported.
2341  * Returns 0 if the recovery mechanisms are unsuccessful.
2342  * Returns a non-zero value otherwise.
2343  */
2344 int efx_try_recovery(struct efx_nic *efx)
2345 {
2346 #ifdef CONFIG_EEH
2347 	/* A PCI error can occur and not be seen by EEH because nothing
2348 	 * happens on the PCI bus. In this case the driver may fail and
2349 	 * schedule a 'recover or reset', leading to this recovery handler.
2350 	 * Manually call the eeh failure check function.
2351 	 */
2352 	struct eeh_dev *eehdev =
2353 		of_node_to_eeh_dev(pci_device_to_OF_node(efx->pci_dev));
2354 
2355 	if (eeh_dev_check_failure(eehdev)) {
2356 		/* The EEH mechanisms will handle the error and reset the
2357 		 * device if necessary.
2358 		 */
2359 		return 1;
2360 	}
2361 #endif
2362 	return 0;
2363 }
2364 
2365 /* The worker thread exists so that code that cannot sleep can
2366  * schedule a reset for later.
2367  */
2368 static void efx_reset_work(struct work_struct *data)
2369 {
2370 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2371 	unsigned long pending;
2372 	enum reset_type method;
2373 
2374 	pending = ACCESS_ONCE(efx->reset_pending);
2375 	method = fls(pending) - 1;
2376 
2377 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2378 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2379 	    efx_try_recovery(efx))
2380 		return;
2381 
2382 	if (!pending)
2383 		return;
2384 
2385 	rtnl_lock();
2386 
2387 	/* We checked the state in efx_schedule_reset() but it may
2388 	 * have changed by now.  Now that we have the RTNL lock,
2389 	 * it cannot change again.
2390 	 */
2391 	if (efx->state == STATE_READY)
2392 		(void)efx_reset(efx, method);
2393 
2394 	rtnl_unlock();
2395 }
2396 
2397 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2398 {
2399 	enum reset_type method;
2400 
2401 	if (efx->state == STATE_RECOVERY) {
2402 		netif_dbg(efx, drv, efx->net_dev,
2403 			  "recovering: skip scheduling %s reset\n",
2404 			  RESET_TYPE(type));
2405 		return;
2406 	}
2407 
2408 	switch (type) {
2409 	case RESET_TYPE_INVISIBLE:
2410 	case RESET_TYPE_ALL:
2411 	case RESET_TYPE_RECOVER_OR_ALL:
2412 	case RESET_TYPE_WORLD:
2413 	case RESET_TYPE_DISABLE:
2414 	case RESET_TYPE_RECOVER_OR_DISABLE:
2415 		method = type;
2416 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2417 			  RESET_TYPE(method));
2418 		break;
2419 	default:
2420 		method = efx->type->map_reset_reason(type);
2421 		netif_dbg(efx, drv, efx->net_dev,
2422 			  "scheduling %s reset for %s\n",
2423 			  RESET_TYPE(method), RESET_TYPE(type));
2424 		break;
2425 	}
2426 
2427 	set_bit(method, &efx->reset_pending);
2428 	smp_mb(); /* ensure we change reset_pending before checking state */
2429 
2430 	/* If we're not READY then just leave the flags set as the cue
2431 	 * to abort probing or reschedule the reset later.
2432 	 */
2433 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2434 		return;
2435 
2436 	/* efx_process_channel() will no longer read events once a
2437 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2438 	efx_mcdi_mode_poll(efx);
2439 
2440 	queue_work(reset_workqueue, &efx->reset_work);
2441 }
2442 
2443 /**************************************************************************
2444  *
2445  * List of NICs we support
2446  *
2447  **************************************************************************/
2448 
2449 /* PCI device ID table */
2450 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2451 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2452 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2453 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2454 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2455 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2456 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2457 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2458 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2459 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2460 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2461 	{0}			/* end of list */
2462 };
2463 
2464 /**************************************************************************
2465  *
2466  * Dummy PHY/MAC operations
2467  *
2468  * Can be used for some unimplemented operations
2469  * Needed so all function pointers are valid and do not have to be tested
2470  * before use
2471  *
2472  **************************************************************************/
2473 int efx_port_dummy_op_int(struct efx_nic *efx)
2474 {
2475 	return 0;
2476 }
2477 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2478 
2479 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2480 {
2481 	return false;
2482 }
2483 
2484 static const struct efx_phy_operations efx_dummy_phy_operations = {
2485 	.init		 = efx_port_dummy_op_int,
2486 	.reconfigure	 = efx_port_dummy_op_int,
2487 	.poll		 = efx_port_dummy_op_poll,
2488 	.fini		 = efx_port_dummy_op_void,
2489 };
2490 
2491 /**************************************************************************
2492  *
2493  * Data housekeeping
2494  *
2495  **************************************************************************/
2496 
2497 /* This zeroes out and then fills in the invariants in a struct
2498  * efx_nic (including all sub-structures).
2499  */
2500 static int efx_init_struct(struct efx_nic *efx,
2501 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2502 {
2503 	int i;
2504 
2505 	/* Initialise common structures */
2506 	spin_lock_init(&efx->biu_lock);
2507 #ifdef CONFIG_SFC_MTD
2508 	INIT_LIST_HEAD(&efx->mtd_list);
2509 #endif
2510 	INIT_WORK(&efx->reset_work, efx_reset_work);
2511 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2512 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2513 	efx->pci_dev = pci_dev;
2514 	efx->msg_enable = debug;
2515 	efx->state = STATE_UNINIT;
2516 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2517 
2518 	efx->net_dev = net_dev;
2519 	spin_lock_init(&efx->stats_lock);
2520 	mutex_init(&efx->mac_lock);
2521 	efx->phy_op = &efx_dummy_phy_operations;
2522 	efx->mdio.dev = net_dev;
2523 	INIT_WORK(&efx->mac_work, efx_mac_work);
2524 	init_waitqueue_head(&efx->flush_wq);
2525 
2526 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2527 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2528 		if (!efx->channel[i])
2529 			goto fail;
2530 	}
2531 
2532 	EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS);
2533 
2534 	/* Higher numbered interrupt modes are less capable! */
2535 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2536 				  interrupt_mode);
2537 
2538 	/* Would be good to use the net_dev name, but we're too early */
2539 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2540 		 pci_name(pci_dev));
2541 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2542 	if (!efx->workqueue)
2543 		goto fail;
2544 
2545 	return 0;
2546 
2547 fail:
2548 	efx_fini_struct(efx);
2549 	return -ENOMEM;
2550 }
2551 
2552 static void efx_fini_struct(struct efx_nic *efx)
2553 {
2554 	int i;
2555 
2556 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2557 		kfree(efx->channel[i]);
2558 
2559 	if (efx->workqueue) {
2560 		destroy_workqueue(efx->workqueue);
2561 		efx->workqueue = NULL;
2562 	}
2563 }
2564 
2565 /**************************************************************************
2566  *
2567  * PCI interface
2568  *
2569  **************************************************************************/
2570 
2571 /* Main body of final NIC shutdown code
2572  * This is called only at module unload (or hotplug removal).
2573  */
2574 static void efx_pci_remove_main(struct efx_nic *efx)
2575 {
2576 	/* Flush reset_work. It can no longer be scheduled since we
2577 	 * are not READY.
2578 	 */
2579 	BUG_ON(efx->state == STATE_READY);
2580 	cancel_work_sync(&efx->reset_work);
2581 
2582 	efx_stop_interrupts(efx, false);
2583 	efx_nic_fini_interrupt(efx);
2584 	efx_fini_port(efx);
2585 	efx->type->fini(efx);
2586 	efx_fini_napi(efx);
2587 	efx_remove_all(efx);
2588 }
2589 
2590 /* Final NIC shutdown
2591  * This is called only at module unload (or hotplug removal).
2592  */
2593 static void efx_pci_remove(struct pci_dev *pci_dev)
2594 {
2595 	struct efx_nic *efx;
2596 
2597 	efx = pci_get_drvdata(pci_dev);
2598 	if (!efx)
2599 		return;
2600 
2601 	/* Mark the NIC as fini, then stop the interface */
2602 	rtnl_lock();
2603 	dev_close(efx->net_dev);
2604 	efx_stop_interrupts(efx, false);
2605 	rtnl_unlock();
2606 
2607 	efx_sriov_fini(efx);
2608 	efx_unregister_netdev(efx);
2609 
2610 	efx_mtd_remove(efx);
2611 
2612 	efx_pci_remove_main(efx);
2613 
2614 	efx_fini_io(efx);
2615 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2616 
2617 	efx_fini_struct(efx);
2618 	pci_set_drvdata(pci_dev, NULL);
2619 	free_netdev(efx->net_dev);
2620 
2621 	pci_disable_pcie_error_reporting(pci_dev);
2622 };
2623 
2624 /* NIC VPD information
2625  * Called during probe to display the part number of the
2626  * installed NIC.  VPD is potentially very large but this should
2627  * always appear within the first 512 bytes.
2628  */
2629 #define SFC_VPD_LEN 512
2630 static void efx_print_product_vpd(struct efx_nic *efx)
2631 {
2632 	struct pci_dev *dev = efx->pci_dev;
2633 	char vpd_data[SFC_VPD_LEN];
2634 	ssize_t vpd_size;
2635 	int i, j;
2636 
2637 	/* Get the vpd data from the device */
2638 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2639 	if (vpd_size <= 0) {
2640 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2641 		return;
2642 	}
2643 
2644 	/* Get the Read only section */
2645 	i = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2646 	if (i < 0) {
2647 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2648 		return;
2649 	}
2650 
2651 	j = pci_vpd_lrdt_size(&vpd_data[i]);
2652 	i += PCI_VPD_LRDT_TAG_SIZE;
2653 	if (i + j > vpd_size)
2654 		j = vpd_size - i;
2655 
2656 	/* Get the Part number */
2657 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2658 	if (i < 0) {
2659 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2660 		return;
2661 	}
2662 
2663 	j = pci_vpd_info_field_size(&vpd_data[i]);
2664 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
2665 	if (i + j > vpd_size) {
2666 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2667 		return;
2668 	}
2669 
2670 	netif_info(efx, drv, efx->net_dev,
2671 		   "Part Number : %.*s\n", j, &vpd_data[i]);
2672 }
2673 
2674 
2675 /* Main body of NIC initialisation
2676  * This is called at module load (or hotplug insertion, theoretically).
2677  */
2678 static int efx_pci_probe_main(struct efx_nic *efx)
2679 {
2680 	int rc;
2681 
2682 	/* Do start-of-day initialisation */
2683 	rc = efx_probe_all(efx);
2684 	if (rc)
2685 		goto fail1;
2686 
2687 	efx_init_napi(efx);
2688 
2689 	rc = efx->type->init(efx);
2690 	if (rc) {
2691 		netif_err(efx, probe, efx->net_dev,
2692 			  "failed to initialise NIC\n");
2693 		goto fail3;
2694 	}
2695 
2696 	rc = efx_init_port(efx);
2697 	if (rc) {
2698 		netif_err(efx, probe, efx->net_dev,
2699 			  "failed to initialise port\n");
2700 		goto fail4;
2701 	}
2702 
2703 	rc = efx_nic_init_interrupt(efx);
2704 	if (rc)
2705 		goto fail5;
2706 	efx_start_interrupts(efx, false);
2707 
2708 	return 0;
2709 
2710  fail5:
2711 	efx_fini_port(efx);
2712  fail4:
2713 	efx->type->fini(efx);
2714  fail3:
2715 	efx_fini_napi(efx);
2716 	efx_remove_all(efx);
2717  fail1:
2718 	return rc;
2719 }
2720 
2721 /* NIC initialisation
2722  *
2723  * This is called at module load (or hotplug insertion,
2724  * theoretically).  It sets up PCI mappings, resets the NIC,
2725  * sets up and registers the network devices with the kernel and hooks
2726  * the interrupt service routine.  It does not prepare the device for
2727  * transmission; this is left to the first time one of the network
2728  * interfaces is brought up (i.e. efx_net_open).
2729  */
2730 static int efx_pci_probe(struct pci_dev *pci_dev,
2731 			 const struct pci_device_id *entry)
2732 {
2733 	struct net_device *net_dev;
2734 	struct efx_nic *efx;
2735 	int rc;
2736 
2737 	/* Allocate and initialise a struct net_device and struct efx_nic */
2738 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2739 				     EFX_MAX_RX_QUEUES);
2740 	if (!net_dev)
2741 		return -ENOMEM;
2742 	efx = netdev_priv(net_dev);
2743 	efx->type = (const struct efx_nic_type *) entry->driver_data;
2744 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2745 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
2746 			      NETIF_F_RXCSUM);
2747 	if (efx->type->offload_features & NETIF_F_V6_CSUM)
2748 		net_dev->features |= NETIF_F_TSO6;
2749 	/* Mask for features that also apply to VLAN devices */
2750 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2751 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2752 				   NETIF_F_RXCSUM);
2753 	/* All offloads can be toggled */
2754 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2755 	pci_set_drvdata(pci_dev, efx);
2756 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2757 	rc = efx_init_struct(efx, pci_dev, net_dev);
2758 	if (rc)
2759 		goto fail1;
2760 
2761 	netif_info(efx, probe, efx->net_dev,
2762 		   "Solarflare NIC detected\n");
2763 
2764 	efx_print_product_vpd(efx);
2765 
2766 	/* Set up basic I/O (BAR mappings etc) */
2767 	rc = efx_init_io(efx);
2768 	if (rc)
2769 		goto fail2;
2770 
2771 	rc = efx_pci_probe_main(efx);
2772 	if (rc)
2773 		goto fail3;
2774 
2775 	rc = efx_register_netdev(efx);
2776 	if (rc)
2777 		goto fail4;
2778 
2779 	rc = efx_sriov_init(efx);
2780 	if (rc)
2781 		netif_err(efx, probe, efx->net_dev,
2782 			  "SR-IOV can't be enabled rc %d\n", rc);
2783 
2784 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2785 
2786 	/* Try to create MTDs, but allow this to fail */
2787 	rtnl_lock();
2788 	rc = efx_mtd_probe(efx);
2789 	rtnl_unlock();
2790 	if (rc)
2791 		netif_warn(efx, probe, efx->net_dev,
2792 			   "failed to create MTDs (%d)\n", rc);
2793 
2794 	rc = pci_enable_pcie_error_reporting(pci_dev);
2795 	if (rc && rc != -EINVAL)
2796 		netif_warn(efx, probe, efx->net_dev,
2797 			   "pci_enable_pcie_error_reporting failed (%d)\n", rc);
2798 
2799 	return 0;
2800 
2801  fail4:
2802 	efx_pci_remove_main(efx);
2803  fail3:
2804 	efx_fini_io(efx);
2805  fail2:
2806 	efx_fini_struct(efx);
2807  fail1:
2808 	pci_set_drvdata(pci_dev, NULL);
2809 	WARN_ON(rc > 0);
2810 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2811 	free_netdev(net_dev);
2812 	return rc;
2813 }
2814 
2815 static int efx_pm_freeze(struct device *dev)
2816 {
2817 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2818 
2819 	rtnl_lock();
2820 
2821 	if (efx->state != STATE_DISABLED) {
2822 		efx->state = STATE_UNINIT;
2823 
2824 		efx_device_detach_sync(efx);
2825 
2826 		efx_stop_all(efx);
2827 		efx_stop_interrupts(efx, false);
2828 	}
2829 
2830 	rtnl_unlock();
2831 
2832 	return 0;
2833 }
2834 
2835 static int efx_pm_thaw(struct device *dev)
2836 {
2837 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2838 
2839 	rtnl_lock();
2840 
2841 	if (efx->state != STATE_DISABLED) {
2842 		efx_start_interrupts(efx, false);
2843 
2844 		mutex_lock(&efx->mac_lock);
2845 		efx->phy_op->reconfigure(efx);
2846 		mutex_unlock(&efx->mac_lock);
2847 
2848 		efx_start_all(efx);
2849 
2850 		netif_device_attach(efx->net_dev);
2851 
2852 		efx->state = STATE_READY;
2853 
2854 		efx->type->resume_wol(efx);
2855 	}
2856 
2857 	rtnl_unlock();
2858 
2859 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2860 	queue_work(reset_workqueue, &efx->reset_work);
2861 
2862 	return 0;
2863 }
2864 
2865 static int efx_pm_poweroff(struct device *dev)
2866 {
2867 	struct pci_dev *pci_dev = to_pci_dev(dev);
2868 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2869 
2870 	efx->type->fini(efx);
2871 
2872 	efx->reset_pending = 0;
2873 
2874 	pci_save_state(pci_dev);
2875 	return pci_set_power_state(pci_dev, PCI_D3hot);
2876 }
2877 
2878 /* Used for both resume and restore */
2879 static int efx_pm_resume(struct device *dev)
2880 {
2881 	struct pci_dev *pci_dev = to_pci_dev(dev);
2882 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
2883 	int rc;
2884 
2885 	rc = pci_set_power_state(pci_dev, PCI_D0);
2886 	if (rc)
2887 		return rc;
2888 	pci_restore_state(pci_dev);
2889 	rc = pci_enable_device(pci_dev);
2890 	if (rc)
2891 		return rc;
2892 	pci_set_master(efx->pci_dev);
2893 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
2894 	if (rc)
2895 		return rc;
2896 	rc = efx->type->init(efx);
2897 	if (rc)
2898 		return rc;
2899 	efx_pm_thaw(dev);
2900 	return 0;
2901 }
2902 
2903 static int efx_pm_suspend(struct device *dev)
2904 {
2905 	int rc;
2906 
2907 	efx_pm_freeze(dev);
2908 	rc = efx_pm_poweroff(dev);
2909 	if (rc)
2910 		efx_pm_resume(dev);
2911 	return rc;
2912 }
2913 
2914 static const struct dev_pm_ops efx_pm_ops = {
2915 	.suspend	= efx_pm_suspend,
2916 	.resume		= efx_pm_resume,
2917 	.freeze		= efx_pm_freeze,
2918 	.thaw		= efx_pm_thaw,
2919 	.poweroff	= efx_pm_poweroff,
2920 	.restore	= efx_pm_resume,
2921 };
2922 
2923 /* A PCI error affecting this device was detected.
2924  * At this point MMIO and DMA may be disabled.
2925  * Stop the software path and request a slot reset.
2926  */
2927 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
2928 					      enum pci_channel_state state)
2929 {
2930 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
2931 	struct efx_nic *efx = pci_get_drvdata(pdev);
2932 
2933 	if (state == pci_channel_io_perm_failure)
2934 		return PCI_ERS_RESULT_DISCONNECT;
2935 
2936 	rtnl_lock();
2937 
2938 	if (efx->state != STATE_DISABLED) {
2939 		efx->state = STATE_RECOVERY;
2940 		efx->reset_pending = 0;
2941 
2942 		efx_device_detach_sync(efx);
2943 
2944 		efx_stop_all(efx);
2945 		efx_stop_interrupts(efx, false);
2946 
2947 		status = PCI_ERS_RESULT_NEED_RESET;
2948 	} else {
2949 		/* If the interface is disabled we don't want to do anything
2950 		 * with it.
2951 		 */
2952 		status = PCI_ERS_RESULT_RECOVERED;
2953 	}
2954 
2955 	rtnl_unlock();
2956 
2957 	pci_disable_device(pdev);
2958 
2959 	return status;
2960 }
2961 
2962 /* Fake a successfull reset, which will be performed later in efx_io_resume. */
2963 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
2964 {
2965 	struct efx_nic *efx = pci_get_drvdata(pdev);
2966 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
2967 	int rc;
2968 
2969 	if (pci_enable_device(pdev)) {
2970 		netif_err(efx, hw, efx->net_dev,
2971 			  "Cannot re-enable PCI device after reset.\n");
2972 		status =  PCI_ERS_RESULT_DISCONNECT;
2973 	}
2974 
2975 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
2976 	if (rc) {
2977 		netif_err(efx, hw, efx->net_dev,
2978 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
2979 		/* Non-fatal error. Continue. */
2980 	}
2981 
2982 	return status;
2983 }
2984 
2985 /* Perform the actual reset and resume I/O operations. */
2986 static void efx_io_resume(struct pci_dev *pdev)
2987 {
2988 	struct efx_nic *efx = pci_get_drvdata(pdev);
2989 	int rc;
2990 
2991 	rtnl_lock();
2992 
2993 	if (efx->state == STATE_DISABLED)
2994 		goto out;
2995 
2996 	rc = efx_reset(efx, RESET_TYPE_ALL);
2997 	if (rc) {
2998 		netif_err(efx, hw, efx->net_dev,
2999 			  "efx_reset failed after PCI error (%d)\n", rc);
3000 	} else {
3001 		efx->state = STATE_READY;
3002 		netif_dbg(efx, hw, efx->net_dev,
3003 			  "Done resetting and resuming IO after PCI error.\n");
3004 	}
3005 
3006 out:
3007 	rtnl_unlock();
3008 }
3009 
3010 /* For simplicity and reliability, we always require a slot reset and try to
3011  * reset the hardware when a pci error affecting the device is detected.
3012  * We leave both the link_reset and mmio_enabled callback unimplemented:
3013  * with our request for slot reset the mmio_enabled callback will never be
3014  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3015  */
3016 static struct pci_error_handlers efx_err_handlers = {
3017 	.error_detected = efx_io_error_detected,
3018 	.slot_reset	= efx_io_slot_reset,
3019 	.resume		= efx_io_resume,
3020 };
3021 
3022 static struct pci_driver efx_pci_driver = {
3023 	.name		= KBUILD_MODNAME,
3024 	.id_table	= efx_pci_table,
3025 	.probe		= efx_pci_probe,
3026 	.remove		= efx_pci_remove,
3027 	.driver.pm	= &efx_pm_ops,
3028 	.err_handler	= &efx_err_handlers,
3029 };
3030 
3031 /**************************************************************************
3032  *
3033  * Kernel module interface
3034  *
3035  *************************************************************************/
3036 
3037 module_param(interrupt_mode, uint, 0444);
3038 MODULE_PARM_DESC(interrupt_mode,
3039 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3040 
3041 static int __init efx_init_module(void)
3042 {
3043 	int rc;
3044 
3045 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3046 
3047 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3048 	if (rc)
3049 		goto err_notifier;
3050 
3051 	rc = efx_init_sriov();
3052 	if (rc)
3053 		goto err_sriov;
3054 
3055 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3056 	if (!reset_workqueue) {
3057 		rc = -ENOMEM;
3058 		goto err_reset;
3059 	}
3060 
3061 	rc = pci_register_driver(&efx_pci_driver);
3062 	if (rc < 0)
3063 		goto err_pci;
3064 
3065 	return 0;
3066 
3067  err_pci:
3068 	destroy_workqueue(reset_workqueue);
3069  err_reset:
3070 	efx_fini_sriov();
3071  err_sriov:
3072 	unregister_netdevice_notifier(&efx_netdev_notifier);
3073  err_notifier:
3074 	return rc;
3075 }
3076 
3077 static void __exit efx_exit_module(void)
3078 {
3079 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3080 
3081 	pci_unregister_driver(&efx_pci_driver);
3082 	destroy_workqueue(reset_workqueue);
3083 	efx_fini_sriov();
3084 	unregister_netdevice_notifier(&efx_netdev_notifier);
3085 
3086 }
3087 
3088 module_init(efx_init_module);
3089 module_exit(efx_exit_module);
3090 
3091 MODULE_AUTHOR("Solarflare Communications and "
3092 	      "Michael Brown <mbrown@fensystems.co.uk>");
3093 MODULE_DESCRIPTION("Solarflare Communications network driver");
3094 MODULE_LICENSE("GPL");
3095 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3096