xref: /openbmc/linux/drivers/net/ethernet/sfc/efx.c (revision eb3fcf00)
1 /****************************************************************************
2  * Driver for Solarflare network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2013 Solarflare Communications Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published
8  * by the Free Software Foundation, incorporated herein by reference.
9  */
10 
11 #include <linux/module.h>
12 #include <linux/pci.h>
13 #include <linux/netdevice.h>
14 #include <linux/etherdevice.h>
15 #include <linux/delay.h>
16 #include <linux/notifier.h>
17 #include <linux/ip.h>
18 #include <linux/tcp.h>
19 #include <linux/in.h>
20 #include <linux/ethtool.h>
21 #include <linux/topology.h>
22 #include <linux/gfp.h>
23 #include <linux/aer.h>
24 #include <linux/interrupt.h>
25 #include "net_driver.h"
26 #include "efx.h"
27 #include "nic.h"
28 #include "selftest.h"
29 #include "sriov.h"
30 
31 #include "mcdi.h"
32 #include "workarounds.h"
33 
34 /**************************************************************************
35  *
36  * Type name strings
37  *
38  **************************************************************************
39  */
40 
41 /* Loopback mode names (see LOOPBACK_MODE()) */
42 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
43 const char *const efx_loopback_mode_names[] = {
44 	[LOOPBACK_NONE]		= "NONE",
45 	[LOOPBACK_DATA]		= "DATAPATH",
46 	[LOOPBACK_GMAC]		= "GMAC",
47 	[LOOPBACK_XGMII]	= "XGMII",
48 	[LOOPBACK_XGXS]		= "XGXS",
49 	[LOOPBACK_XAUI]		= "XAUI",
50 	[LOOPBACK_GMII]		= "GMII",
51 	[LOOPBACK_SGMII]	= "SGMII",
52 	[LOOPBACK_XGBR]		= "XGBR",
53 	[LOOPBACK_XFI]		= "XFI",
54 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
55 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
56 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
57 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
58 	[LOOPBACK_GPHY]		= "GPHY",
59 	[LOOPBACK_PHYXS]	= "PHYXS",
60 	[LOOPBACK_PCS]		= "PCS",
61 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
62 	[LOOPBACK_XPORT]	= "XPORT",
63 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
64 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
65 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
66 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
67 	[LOOPBACK_GMII_WS]	= "GMII_WS",
68 	[LOOPBACK_XFI_WS]	= "XFI_WS",
69 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
70 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
71 };
72 
73 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
74 const char *const efx_reset_type_names[] = {
75 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
76 	[RESET_TYPE_ALL]                = "ALL",
77 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
78 	[RESET_TYPE_WORLD]              = "WORLD",
79 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
80 	[RESET_TYPE_DATAPATH]           = "DATAPATH",
81 	[RESET_TYPE_MC_BIST]		= "MC_BIST",
82 	[RESET_TYPE_DISABLE]            = "DISABLE",
83 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
84 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
85 	[RESET_TYPE_RX_RECOVERY]        = "RX_RECOVERY",
86 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
87 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
88 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
89 	[RESET_TYPE_MCDI_TIMEOUT]	= "MCDI_TIMEOUT (FLR)",
90 };
91 
92 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
93  * queued onto this work queue. This is not a per-nic work queue, because
94  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
95  */
96 static struct workqueue_struct *reset_workqueue;
97 
98 /* How often and how many times to poll for a reset while waiting for a
99  * BIST that another function started to complete.
100  */
101 #define BIST_WAIT_DELAY_MS	100
102 #define BIST_WAIT_DELAY_COUNT	100
103 
104 /**************************************************************************
105  *
106  * Configurable values
107  *
108  *************************************************************************/
109 
110 /*
111  * Use separate channels for TX and RX events
112  *
113  * Set this to 1 to use separate channels for TX and RX. It allows us
114  * to control interrupt affinity separately for TX and RX.
115  *
116  * This is only used in MSI-X interrupt mode
117  */
118 bool efx_separate_tx_channels;
119 module_param(efx_separate_tx_channels, bool, 0444);
120 MODULE_PARM_DESC(efx_separate_tx_channels,
121 		 "Use separate channels for TX and RX");
122 
123 /* This is the weight assigned to each of the (per-channel) virtual
124  * NAPI devices.
125  */
126 static int napi_weight = 64;
127 
128 /* This is the time (in jiffies) between invocations of the hardware
129  * monitor.
130  * On Falcon-based NICs, this will:
131  * - Check the on-board hardware monitor;
132  * - Poll the link state and reconfigure the hardware as necessary.
133  * On Siena-based NICs for power systems with EEH support, this will give EEH a
134  * chance to start.
135  */
136 static unsigned int efx_monitor_interval = 1 * HZ;
137 
138 /* Initial interrupt moderation settings.  They can be modified after
139  * module load with ethtool.
140  *
141  * The default for RX should strike a balance between increasing the
142  * round-trip latency and reducing overhead.
143  */
144 static unsigned int rx_irq_mod_usec = 60;
145 
146 /* Initial interrupt moderation settings.  They can be modified after
147  * module load with ethtool.
148  *
149  * This default is chosen to ensure that a 10G link does not go idle
150  * while a TX queue is stopped after it has become full.  A queue is
151  * restarted when it drops below half full.  The time this takes (assuming
152  * worst case 3 descriptors per packet and 1024 descriptors) is
153  *   512 / 3 * 1.2 = 205 usec.
154  */
155 static unsigned int tx_irq_mod_usec = 150;
156 
157 /* This is the first interrupt mode to try out of:
158  * 0 => MSI-X
159  * 1 => MSI
160  * 2 => legacy
161  */
162 static unsigned int interrupt_mode;
163 
164 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
165  * i.e. the number of CPUs among which we may distribute simultaneous
166  * interrupt handling.
167  *
168  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
169  * The default (0) means to assign an interrupt to each core.
170  */
171 static unsigned int rss_cpus;
172 module_param(rss_cpus, uint, 0444);
173 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
174 
175 static bool phy_flash_cfg;
176 module_param(phy_flash_cfg, bool, 0644);
177 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
178 
179 static unsigned irq_adapt_low_thresh = 8000;
180 module_param(irq_adapt_low_thresh, uint, 0644);
181 MODULE_PARM_DESC(irq_adapt_low_thresh,
182 		 "Threshold score for reducing IRQ moderation");
183 
184 static unsigned irq_adapt_high_thresh = 16000;
185 module_param(irq_adapt_high_thresh, uint, 0644);
186 MODULE_PARM_DESC(irq_adapt_high_thresh,
187 		 "Threshold score for increasing IRQ moderation");
188 
189 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
190 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
191 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
192 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
193 module_param(debug, uint, 0);
194 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
195 
196 /**************************************************************************
197  *
198  * Utility functions and prototypes
199  *
200  *************************************************************************/
201 
202 static int efx_soft_enable_interrupts(struct efx_nic *efx);
203 static void efx_soft_disable_interrupts(struct efx_nic *efx);
204 static void efx_remove_channel(struct efx_channel *channel);
205 static void efx_remove_channels(struct efx_nic *efx);
206 static const struct efx_channel_type efx_default_channel_type;
207 static void efx_remove_port(struct efx_nic *efx);
208 static void efx_init_napi_channel(struct efx_channel *channel);
209 static void efx_fini_napi(struct efx_nic *efx);
210 static void efx_fini_napi_channel(struct efx_channel *channel);
211 static void efx_fini_struct(struct efx_nic *efx);
212 static void efx_start_all(struct efx_nic *efx);
213 static void efx_stop_all(struct efx_nic *efx);
214 
215 #define EFX_ASSERT_RESET_SERIALISED(efx)		\
216 	do {						\
217 		if ((efx->state == STATE_READY) ||	\
218 		    (efx->state == STATE_RECOVERY) ||	\
219 		    (efx->state == STATE_DISABLED))	\
220 			ASSERT_RTNL();			\
221 	} while (0)
222 
223 static int efx_check_disabled(struct efx_nic *efx)
224 {
225 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
226 		netif_err(efx, drv, efx->net_dev,
227 			  "device is disabled due to earlier errors\n");
228 		return -EIO;
229 	}
230 	return 0;
231 }
232 
233 /**************************************************************************
234  *
235  * Event queue processing
236  *
237  *************************************************************************/
238 
239 /* Process channel's event queue
240  *
241  * This function is responsible for processing the event queue of a
242  * single channel.  The caller must guarantee that this function will
243  * never be concurrently called more than once on the same channel,
244  * though different channels may be being processed concurrently.
245  */
246 static int efx_process_channel(struct efx_channel *channel, int budget)
247 {
248 	struct efx_tx_queue *tx_queue;
249 	int spent;
250 
251 	if (unlikely(!channel->enabled))
252 		return 0;
253 
254 	efx_for_each_channel_tx_queue(tx_queue, channel) {
255 		tx_queue->pkts_compl = 0;
256 		tx_queue->bytes_compl = 0;
257 	}
258 
259 	spent = efx_nic_process_eventq(channel, budget);
260 	if (spent && efx_channel_has_rx_queue(channel)) {
261 		struct efx_rx_queue *rx_queue =
262 			efx_channel_get_rx_queue(channel);
263 
264 		efx_rx_flush_packet(channel);
265 		efx_fast_push_rx_descriptors(rx_queue, true);
266 	}
267 
268 	/* Update BQL */
269 	efx_for_each_channel_tx_queue(tx_queue, channel) {
270 		if (tx_queue->bytes_compl) {
271 			netdev_tx_completed_queue(tx_queue->core_txq,
272 				tx_queue->pkts_compl, tx_queue->bytes_compl);
273 		}
274 	}
275 
276 	return spent;
277 }
278 
279 /* NAPI poll handler
280  *
281  * NAPI guarantees serialisation of polls of the same device, which
282  * provides the guarantee required by efx_process_channel().
283  */
284 static int efx_poll(struct napi_struct *napi, int budget)
285 {
286 	struct efx_channel *channel =
287 		container_of(napi, struct efx_channel, napi_str);
288 	struct efx_nic *efx = channel->efx;
289 	int spent;
290 
291 	if (!efx_channel_lock_napi(channel))
292 		return budget;
293 
294 	netif_vdbg(efx, intr, efx->net_dev,
295 		   "channel %d NAPI poll executing on CPU %d\n",
296 		   channel->channel, raw_smp_processor_id());
297 
298 	spent = efx_process_channel(channel, budget);
299 
300 	if (spent < budget) {
301 		if (efx_channel_has_rx_queue(channel) &&
302 		    efx->irq_rx_adaptive &&
303 		    unlikely(++channel->irq_count == 1000)) {
304 			if (unlikely(channel->irq_mod_score <
305 				     irq_adapt_low_thresh)) {
306 				if (channel->irq_moderation > 1) {
307 					channel->irq_moderation -= 1;
308 					efx->type->push_irq_moderation(channel);
309 				}
310 			} else if (unlikely(channel->irq_mod_score >
311 					    irq_adapt_high_thresh)) {
312 				if (channel->irq_moderation <
313 				    efx->irq_rx_moderation) {
314 					channel->irq_moderation += 1;
315 					efx->type->push_irq_moderation(channel);
316 				}
317 			}
318 			channel->irq_count = 0;
319 			channel->irq_mod_score = 0;
320 		}
321 
322 		efx_filter_rfs_expire(channel);
323 
324 		/* There is no race here; although napi_disable() will
325 		 * only wait for napi_complete(), this isn't a problem
326 		 * since efx_nic_eventq_read_ack() will have no effect if
327 		 * interrupts have already been disabled.
328 		 */
329 		napi_complete(napi);
330 		efx_nic_eventq_read_ack(channel);
331 	}
332 
333 	efx_channel_unlock_napi(channel);
334 	return spent;
335 }
336 
337 /* Create event queue
338  * Event queue memory allocations are done only once.  If the channel
339  * is reset, the memory buffer will be reused; this guards against
340  * errors during channel reset and also simplifies interrupt handling.
341  */
342 static int efx_probe_eventq(struct efx_channel *channel)
343 {
344 	struct efx_nic *efx = channel->efx;
345 	unsigned long entries;
346 
347 	netif_dbg(efx, probe, efx->net_dev,
348 		  "chan %d create event queue\n", channel->channel);
349 
350 	/* Build an event queue with room for one event per tx and rx buffer,
351 	 * plus some extra for link state events and MCDI completions. */
352 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
353 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
354 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
355 
356 	return efx_nic_probe_eventq(channel);
357 }
358 
359 /* Prepare channel's event queue */
360 static int efx_init_eventq(struct efx_channel *channel)
361 {
362 	struct efx_nic *efx = channel->efx;
363 	int rc;
364 
365 	EFX_WARN_ON_PARANOID(channel->eventq_init);
366 
367 	netif_dbg(efx, drv, efx->net_dev,
368 		  "chan %d init event queue\n", channel->channel);
369 
370 	rc = efx_nic_init_eventq(channel);
371 	if (rc == 0) {
372 		efx->type->push_irq_moderation(channel);
373 		channel->eventq_read_ptr = 0;
374 		channel->eventq_init = true;
375 	}
376 	return rc;
377 }
378 
379 /* Enable event queue processing and NAPI */
380 void efx_start_eventq(struct efx_channel *channel)
381 {
382 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
383 		  "chan %d start event queue\n", channel->channel);
384 
385 	/* Make sure the NAPI handler sees the enabled flag set */
386 	channel->enabled = true;
387 	smp_wmb();
388 
389 	efx_channel_enable(channel);
390 	napi_enable(&channel->napi_str);
391 	efx_nic_eventq_read_ack(channel);
392 }
393 
394 /* Disable event queue processing and NAPI */
395 void efx_stop_eventq(struct efx_channel *channel)
396 {
397 	if (!channel->enabled)
398 		return;
399 
400 	napi_disable(&channel->napi_str);
401 	while (!efx_channel_disable(channel))
402 		usleep_range(1000, 20000);
403 	channel->enabled = false;
404 }
405 
406 static void efx_fini_eventq(struct efx_channel *channel)
407 {
408 	if (!channel->eventq_init)
409 		return;
410 
411 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
412 		  "chan %d fini event queue\n", channel->channel);
413 
414 	efx_nic_fini_eventq(channel);
415 	channel->eventq_init = false;
416 }
417 
418 static void efx_remove_eventq(struct efx_channel *channel)
419 {
420 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
421 		  "chan %d remove event queue\n", channel->channel);
422 
423 	efx_nic_remove_eventq(channel);
424 }
425 
426 /**************************************************************************
427  *
428  * Channel handling
429  *
430  *************************************************************************/
431 
432 /* Allocate and initialise a channel structure. */
433 static struct efx_channel *
434 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
435 {
436 	struct efx_channel *channel;
437 	struct efx_rx_queue *rx_queue;
438 	struct efx_tx_queue *tx_queue;
439 	int j;
440 
441 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
442 	if (!channel)
443 		return NULL;
444 
445 	channel->efx = efx;
446 	channel->channel = i;
447 	channel->type = &efx_default_channel_type;
448 
449 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
450 		tx_queue = &channel->tx_queue[j];
451 		tx_queue->efx = efx;
452 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
453 		tx_queue->channel = channel;
454 	}
455 
456 	rx_queue = &channel->rx_queue;
457 	rx_queue->efx = efx;
458 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
459 		    (unsigned long)rx_queue);
460 
461 	return channel;
462 }
463 
464 /* Allocate and initialise a channel structure, copying parameters
465  * (but not resources) from an old channel structure.
466  */
467 static struct efx_channel *
468 efx_copy_channel(const struct efx_channel *old_channel)
469 {
470 	struct efx_channel *channel;
471 	struct efx_rx_queue *rx_queue;
472 	struct efx_tx_queue *tx_queue;
473 	int j;
474 
475 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
476 	if (!channel)
477 		return NULL;
478 
479 	*channel = *old_channel;
480 
481 	channel->napi_dev = NULL;
482 	memset(&channel->eventq, 0, sizeof(channel->eventq));
483 
484 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
485 		tx_queue = &channel->tx_queue[j];
486 		if (tx_queue->channel)
487 			tx_queue->channel = channel;
488 		tx_queue->buffer = NULL;
489 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
490 	}
491 
492 	rx_queue = &channel->rx_queue;
493 	rx_queue->buffer = NULL;
494 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
495 	setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
496 		    (unsigned long)rx_queue);
497 
498 	return channel;
499 }
500 
501 static int efx_probe_channel(struct efx_channel *channel)
502 {
503 	struct efx_tx_queue *tx_queue;
504 	struct efx_rx_queue *rx_queue;
505 	int rc;
506 
507 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
508 		  "creating channel %d\n", channel->channel);
509 
510 	rc = channel->type->pre_probe(channel);
511 	if (rc)
512 		goto fail;
513 
514 	rc = efx_probe_eventq(channel);
515 	if (rc)
516 		goto fail;
517 
518 	efx_for_each_channel_tx_queue(tx_queue, channel) {
519 		rc = efx_probe_tx_queue(tx_queue);
520 		if (rc)
521 			goto fail;
522 	}
523 
524 	efx_for_each_channel_rx_queue(rx_queue, channel) {
525 		rc = efx_probe_rx_queue(rx_queue);
526 		if (rc)
527 			goto fail;
528 	}
529 
530 	return 0;
531 
532 fail:
533 	efx_remove_channel(channel);
534 	return rc;
535 }
536 
537 static void
538 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
539 {
540 	struct efx_nic *efx = channel->efx;
541 	const char *type;
542 	int number;
543 
544 	number = channel->channel;
545 	if (efx->tx_channel_offset == 0) {
546 		type = "";
547 	} else if (channel->channel < efx->tx_channel_offset) {
548 		type = "-rx";
549 	} else {
550 		type = "-tx";
551 		number -= efx->tx_channel_offset;
552 	}
553 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
554 }
555 
556 static void efx_set_channel_names(struct efx_nic *efx)
557 {
558 	struct efx_channel *channel;
559 
560 	efx_for_each_channel(channel, efx)
561 		channel->type->get_name(channel,
562 					efx->msi_context[channel->channel].name,
563 					sizeof(efx->msi_context[0].name));
564 }
565 
566 static int efx_probe_channels(struct efx_nic *efx)
567 {
568 	struct efx_channel *channel;
569 	int rc;
570 
571 	/* Restart special buffer allocation */
572 	efx->next_buffer_table = 0;
573 
574 	/* Probe channels in reverse, so that any 'extra' channels
575 	 * use the start of the buffer table. This allows the traffic
576 	 * channels to be resized without moving them or wasting the
577 	 * entries before them.
578 	 */
579 	efx_for_each_channel_rev(channel, efx) {
580 		rc = efx_probe_channel(channel);
581 		if (rc) {
582 			netif_err(efx, probe, efx->net_dev,
583 				  "failed to create channel %d\n",
584 				  channel->channel);
585 			goto fail;
586 		}
587 	}
588 	efx_set_channel_names(efx);
589 
590 	return 0;
591 
592 fail:
593 	efx_remove_channels(efx);
594 	return rc;
595 }
596 
597 /* Channels are shutdown and reinitialised whilst the NIC is running
598  * to propagate configuration changes (mtu, checksum offload), or
599  * to clear hardware error conditions
600  */
601 static void efx_start_datapath(struct efx_nic *efx)
602 {
603 	bool old_rx_scatter = efx->rx_scatter;
604 	struct efx_tx_queue *tx_queue;
605 	struct efx_rx_queue *rx_queue;
606 	struct efx_channel *channel;
607 	size_t rx_buf_len;
608 
609 	/* Calculate the rx buffer allocation parameters required to
610 	 * support the current MTU, including padding for header
611 	 * alignment and overruns.
612 	 */
613 	efx->rx_dma_len = (efx->rx_prefix_size +
614 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
615 			   efx->type->rx_buffer_padding);
616 	rx_buf_len = (sizeof(struct efx_rx_page_state) +
617 		      efx->rx_ip_align + efx->rx_dma_len);
618 	if (rx_buf_len <= PAGE_SIZE) {
619 		efx->rx_scatter = efx->type->always_rx_scatter;
620 		efx->rx_buffer_order = 0;
621 	} else if (efx->type->can_rx_scatter) {
622 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
623 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
624 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
625 				       EFX_RX_BUF_ALIGNMENT) >
626 			     PAGE_SIZE);
627 		efx->rx_scatter = true;
628 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
629 		efx->rx_buffer_order = 0;
630 	} else {
631 		efx->rx_scatter = false;
632 		efx->rx_buffer_order = get_order(rx_buf_len);
633 	}
634 
635 	efx_rx_config_page_split(efx);
636 	if (efx->rx_buffer_order)
637 		netif_dbg(efx, drv, efx->net_dev,
638 			  "RX buf len=%u; page order=%u batch=%u\n",
639 			  efx->rx_dma_len, efx->rx_buffer_order,
640 			  efx->rx_pages_per_batch);
641 	else
642 		netif_dbg(efx, drv, efx->net_dev,
643 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
644 			  efx->rx_dma_len, efx->rx_page_buf_step,
645 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
646 
647 	/* RX filters may also have scatter-enabled flags */
648 	if (efx->rx_scatter != old_rx_scatter)
649 		efx->type->filter_update_rx_scatter(efx);
650 
651 	/* We must keep at least one descriptor in a TX ring empty.
652 	 * We could avoid this when the queue size does not exactly
653 	 * match the hardware ring size, but it's not that important.
654 	 * Therefore we stop the queue when one more skb might fill
655 	 * the ring completely.  We wake it when half way back to
656 	 * empty.
657 	 */
658 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
659 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
660 
661 	/* Initialise the channels */
662 	efx_for_each_channel(channel, efx) {
663 		efx_for_each_channel_tx_queue(tx_queue, channel) {
664 			efx_init_tx_queue(tx_queue);
665 			atomic_inc(&efx->active_queues);
666 		}
667 
668 		efx_for_each_channel_rx_queue(rx_queue, channel) {
669 			efx_init_rx_queue(rx_queue);
670 			atomic_inc(&efx->active_queues);
671 			efx_stop_eventq(channel);
672 			efx_fast_push_rx_descriptors(rx_queue, false);
673 			efx_start_eventq(channel);
674 		}
675 
676 		WARN_ON(channel->rx_pkt_n_frags);
677 	}
678 
679 	efx_ptp_start_datapath(efx);
680 
681 	if (netif_device_present(efx->net_dev))
682 		netif_tx_wake_all_queues(efx->net_dev);
683 }
684 
685 static void efx_stop_datapath(struct efx_nic *efx)
686 {
687 	struct efx_channel *channel;
688 	struct efx_tx_queue *tx_queue;
689 	struct efx_rx_queue *rx_queue;
690 	int rc;
691 
692 	EFX_ASSERT_RESET_SERIALISED(efx);
693 	BUG_ON(efx->port_enabled);
694 
695 	efx_ptp_stop_datapath(efx);
696 
697 	/* Stop RX refill */
698 	efx_for_each_channel(channel, efx) {
699 		efx_for_each_channel_rx_queue(rx_queue, channel)
700 			rx_queue->refill_enabled = false;
701 	}
702 
703 	efx_for_each_channel(channel, efx) {
704 		/* RX packet processing is pipelined, so wait for the
705 		 * NAPI handler to complete.  At least event queue 0
706 		 * might be kept active by non-data events, so don't
707 		 * use napi_synchronize() but actually disable NAPI
708 		 * temporarily.
709 		 */
710 		if (efx_channel_has_rx_queue(channel)) {
711 			efx_stop_eventq(channel);
712 			efx_start_eventq(channel);
713 		}
714 	}
715 
716 	rc = efx->type->fini_dmaq(efx);
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 		 */
723 		netif_err(efx, drv, efx->net_dev,
724 			  "Resetting to recover from flush failure\n");
725 		efx_schedule_reset(efx, RESET_TYPE_ALL);
726 	} else if (rc) {
727 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
728 	} else {
729 		netif_dbg(efx, drv, efx->net_dev,
730 			  "successfully flushed all queues\n");
731 	}
732 
733 	efx_for_each_channel(channel, efx) {
734 		efx_for_each_channel_rx_queue(rx_queue, channel)
735 			efx_fini_rx_queue(rx_queue);
736 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
737 			efx_fini_tx_queue(tx_queue);
738 	}
739 }
740 
741 static void efx_remove_channel(struct efx_channel *channel)
742 {
743 	struct efx_tx_queue *tx_queue;
744 	struct efx_rx_queue *rx_queue;
745 
746 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
747 		  "destroy chan %d\n", channel->channel);
748 
749 	efx_for_each_channel_rx_queue(rx_queue, channel)
750 		efx_remove_rx_queue(rx_queue);
751 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
752 		efx_remove_tx_queue(tx_queue);
753 	efx_remove_eventq(channel);
754 	channel->type->post_remove(channel);
755 }
756 
757 static void efx_remove_channels(struct efx_nic *efx)
758 {
759 	struct efx_channel *channel;
760 
761 	efx_for_each_channel(channel, efx)
762 		efx_remove_channel(channel);
763 }
764 
765 int
766 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
767 {
768 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
769 	u32 old_rxq_entries, old_txq_entries;
770 	unsigned i, next_buffer_table = 0;
771 	int rc, rc2;
772 
773 	rc = efx_check_disabled(efx);
774 	if (rc)
775 		return rc;
776 
777 	/* Not all channels should be reallocated. We must avoid
778 	 * reallocating their buffer table entries.
779 	 */
780 	efx_for_each_channel(channel, efx) {
781 		struct efx_rx_queue *rx_queue;
782 		struct efx_tx_queue *tx_queue;
783 
784 		if (channel->type->copy)
785 			continue;
786 		next_buffer_table = max(next_buffer_table,
787 					channel->eventq.index +
788 					channel->eventq.entries);
789 		efx_for_each_channel_rx_queue(rx_queue, channel)
790 			next_buffer_table = max(next_buffer_table,
791 						rx_queue->rxd.index +
792 						rx_queue->rxd.entries);
793 		efx_for_each_channel_tx_queue(tx_queue, channel)
794 			next_buffer_table = max(next_buffer_table,
795 						tx_queue->txd.index +
796 						tx_queue->txd.entries);
797 	}
798 
799 	efx_device_detach_sync(efx);
800 	efx_stop_all(efx);
801 	efx_soft_disable_interrupts(efx);
802 
803 	/* Clone channels (where possible) */
804 	memset(other_channel, 0, sizeof(other_channel));
805 	for (i = 0; i < efx->n_channels; i++) {
806 		channel = efx->channel[i];
807 		if (channel->type->copy)
808 			channel = channel->type->copy(channel);
809 		if (!channel) {
810 			rc = -ENOMEM;
811 			goto out;
812 		}
813 		other_channel[i] = channel;
814 	}
815 
816 	/* Swap entry counts and channel pointers */
817 	old_rxq_entries = efx->rxq_entries;
818 	old_txq_entries = efx->txq_entries;
819 	efx->rxq_entries = rxq_entries;
820 	efx->txq_entries = txq_entries;
821 	for (i = 0; i < efx->n_channels; i++) {
822 		channel = efx->channel[i];
823 		efx->channel[i] = other_channel[i];
824 		other_channel[i] = channel;
825 	}
826 
827 	/* Restart buffer table allocation */
828 	efx->next_buffer_table = next_buffer_table;
829 
830 	for (i = 0; i < efx->n_channels; i++) {
831 		channel = efx->channel[i];
832 		if (!channel->type->copy)
833 			continue;
834 		rc = efx_probe_channel(channel);
835 		if (rc)
836 			goto rollback;
837 		efx_init_napi_channel(efx->channel[i]);
838 	}
839 
840 out:
841 	/* Destroy unused channel structures */
842 	for (i = 0; i < efx->n_channels; i++) {
843 		channel = other_channel[i];
844 		if (channel && channel->type->copy) {
845 			efx_fini_napi_channel(channel);
846 			efx_remove_channel(channel);
847 			kfree(channel);
848 		}
849 	}
850 
851 	rc2 = efx_soft_enable_interrupts(efx);
852 	if (rc2) {
853 		rc = rc ? rc : rc2;
854 		netif_err(efx, drv, efx->net_dev,
855 			  "unable to restart interrupts on channel reallocation\n");
856 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
857 	} else {
858 		efx_start_all(efx);
859 		netif_device_attach(efx->net_dev);
860 	}
861 	return rc;
862 
863 rollback:
864 	/* Swap back */
865 	efx->rxq_entries = old_rxq_entries;
866 	efx->txq_entries = old_txq_entries;
867 	for (i = 0; i < efx->n_channels; i++) {
868 		channel = efx->channel[i];
869 		efx->channel[i] = other_channel[i];
870 		other_channel[i] = channel;
871 	}
872 	goto out;
873 }
874 
875 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
876 {
877 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
878 }
879 
880 static const struct efx_channel_type efx_default_channel_type = {
881 	.pre_probe		= efx_channel_dummy_op_int,
882 	.post_remove		= efx_channel_dummy_op_void,
883 	.get_name		= efx_get_channel_name,
884 	.copy			= efx_copy_channel,
885 	.keep_eventq		= false,
886 };
887 
888 int efx_channel_dummy_op_int(struct efx_channel *channel)
889 {
890 	return 0;
891 }
892 
893 void efx_channel_dummy_op_void(struct efx_channel *channel)
894 {
895 }
896 
897 /**************************************************************************
898  *
899  * Port handling
900  *
901  **************************************************************************/
902 
903 /* This ensures that the kernel is kept informed (via
904  * netif_carrier_on/off) of the link status, and also maintains the
905  * link status's stop on the port's TX queue.
906  */
907 void efx_link_status_changed(struct efx_nic *efx)
908 {
909 	struct efx_link_state *link_state = &efx->link_state;
910 
911 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
912 	 * that no events are triggered between unregister_netdev() and the
913 	 * driver unloading. A more general condition is that NETDEV_CHANGE
914 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
915 	if (!netif_running(efx->net_dev))
916 		return;
917 
918 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
919 		efx->n_link_state_changes++;
920 
921 		if (link_state->up)
922 			netif_carrier_on(efx->net_dev);
923 		else
924 			netif_carrier_off(efx->net_dev);
925 	}
926 
927 	/* Status message for kernel log */
928 	if (link_state->up)
929 		netif_info(efx, link, efx->net_dev,
930 			   "link up at %uMbps %s-duplex (MTU %d)\n",
931 			   link_state->speed, link_state->fd ? "full" : "half",
932 			   efx->net_dev->mtu);
933 	else
934 		netif_info(efx, link, efx->net_dev, "link down\n");
935 }
936 
937 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
938 {
939 	efx->link_advertising = advertising;
940 	if (advertising) {
941 		if (advertising & ADVERTISED_Pause)
942 			efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
943 		else
944 			efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
945 		if (advertising & ADVERTISED_Asym_Pause)
946 			efx->wanted_fc ^= EFX_FC_TX;
947 	}
948 }
949 
950 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
951 {
952 	efx->wanted_fc = wanted_fc;
953 	if (efx->link_advertising) {
954 		if (wanted_fc & EFX_FC_RX)
955 			efx->link_advertising |= (ADVERTISED_Pause |
956 						  ADVERTISED_Asym_Pause);
957 		else
958 			efx->link_advertising &= ~(ADVERTISED_Pause |
959 						   ADVERTISED_Asym_Pause);
960 		if (wanted_fc & EFX_FC_TX)
961 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
962 	}
963 }
964 
965 static void efx_fini_port(struct efx_nic *efx);
966 
967 /* We assume that efx->type->reconfigure_mac will always try to sync RX
968  * filters and therefore needs to read-lock the filter table against freeing
969  */
970 void efx_mac_reconfigure(struct efx_nic *efx)
971 {
972 	down_read(&efx->filter_sem);
973 	efx->type->reconfigure_mac(efx);
974 	up_read(&efx->filter_sem);
975 }
976 
977 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
978  * the MAC appropriately. All other PHY configuration changes are pushed
979  * through phy_op->set_settings(), and pushed asynchronously to the MAC
980  * through efx_monitor().
981  *
982  * Callers must hold the mac_lock
983  */
984 int __efx_reconfigure_port(struct efx_nic *efx)
985 {
986 	enum efx_phy_mode phy_mode;
987 	int rc;
988 
989 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
990 
991 	/* Disable PHY transmit in mac level loopbacks */
992 	phy_mode = efx->phy_mode;
993 	if (LOOPBACK_INTERNAL(efx))
994 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
995 	else
996 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
997 
998 	rc = efx->type->reconfigure_port(efx);
999 
1000 	if (rc)
1001 		efx->phy_mode = phy_mode;
1002 
1003 	return rc;
1004 }
1005 
1006 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1007  * disabled. */
1008 int efx_reconfigure_port(struct efx_nic *efx)
1009 {
1010 	int rc;
1011 
1012 	EFX_ASSERT_RESET_SERIALISED(efx);
1013 
1014 	mutex_lock(&efx->mac_lock);
1015 	rc = __efx_reconfigure_port(efx);
1016 	mutex_unlock(&efx->mac_lock);
1017 
1018 	return rc;
1019 }
1020 
1021 /* Asynchronous work item for changing MAC promiscuity and multicast
1022  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1023  * MAC directly. */
1024 static void efx_mac_work(struct work_struct *data)
1025 {
1026 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1027 
1028 	mutex_lock(&efx->mac_lock);
1029 	if (efx->port_enabled)
1030 		efx_mac_reconfigure(efx);
1031 	mutex_unlock(&efx->mac_lock);
1032 }
1033 
1034 static int efx_probe_port(struct efx_nic *efx)
1035 {
1036 	int rc;
1037 
1038 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1039 
1040 	if (phy_flash_cfg)
1041 		efx->phy_mode = PHY_MODE_SPECIAL;
1042 
1043 	/* Connect up MAC/PHY operations table */
1044 	rc = efx->type->probe_port(efx);
1045 	if (rc)
1046 		return rc;
1047 
1048 	/* Initialise MAC address to permanent address */
1049 	ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1050 
1051 	return 0;
1052 }
1053 
1054 static int efx_init_port(struct efx_nic *efx)
1055 {
1056 	int rc;
1057 
1058 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1059 
1060 	mutex_lock(&efx->mac_lock);
1061 
1062 	rc = efx->phy_op->init(efx);
1063 	if (rc)
1064 		goto fail1;
1065 
1066 	efx->port_initialized = true;
1067 
1068 	/* Reconfigure the MAC before creating dma queues (required for
1069 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1070 	efx_mac_reconfigure(efx);
1071 
1072 	/* Ensure the PHY advertises the correct flow control settings */
1073 	rc = efx->phy_op->reconfigure(efx);
1074 	if (rc && rc != -EPERM)
1075 		goto fail2;
1076 
1077 	mutex_unlock(&efx->mac_lock);
1078 	return 0;
1079 
1080 fail2:
1081 	efx->phy_op->fini(efx);
1082 fail1:
1083 	mutex_unlock(&efx->mac_lock);
1084 	return rc;
1085 }
1086 
1087 static void efx_start_port(struct efx_nic *efx)
1088 {
1089 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1090 	BUG_ON(efx->port_enabled);
1091 
1092 	mutex_lock(&efx->mac_lock);
1093 	efx->port_enabled = true;
1094 
1095 	/* Ensure MAC ingress/egress is enabled */
1096 	efx_mac_reconfigure(efx);
1097 
1098 	mutex_unlock(&efx->mac_lock);
1099 }
1100 
1101 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1102  * and the async self-test, wait for them to finish and prevent them
1103  * being scheduled again.  This doesn't cover online resets, which
1104  * should only be cancelled when removing the device.
1105  */
1106 static void efx_stop_port(struct efx_nic *efx)
1107 {
1108 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1109 
1110 	EFX_ASSERT_RESET_SERIALISED(efx);
1111 
1112 	mutex_lock(&efx->mac_lock);
1113 	efx->port_enabled = false;
1114 	mutex_unlock(&efx->mac_lock);
1115 
1116 	/* Serialise against efx_set_multicast_list() */
1117 	netif_addr_lock_bh(efx->net_dev);
1118 	netif_addr_unlock_bh(efx->net_dev);
1119 
1120 	cancel_delayed_work_sync(&efx->monitor_work);
1121 	efx_selftest_async_cancel(efx);
1122 	cancel_work_sync(&efx->mac_work);
1123 }
1124 
1125 static void efx_fini_port(struct efx_nic *efx)
1126 {
1127 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1128 
1129 	if (!efx->port_initialized)
1130 		return;
1131 
1132 	efx->phy_op->fini(efx);
1133 	efx->port_initialized = false;
1134 
1135 	efx->link_state.up = false;
1136 	efx_link_status_changed(efx);
1137 }
1138 
1139 static void efx_remove_port(struct efx_nic *efx)
1140 {
1141 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1142 
1143 	efx->type->remove_port(efx);
1144 }
1145 
1146 /**************************************************************************
1147  *
1148  * NIC handling
1149  *
1150  **************************************************************************/
1151 
1152 static LIST_HEAD(efx_primary_list);
1153 static LIST_HEAD(efx_unassociated_list);
1154 
1155 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1156 {
1157 	return left->type == right->type &&
1158 		left->vpd_sn && right->vpd_sn &&
1159 		!strcmp(left->vpd_sn, right->vpd_sn);
1160 }
1161 
1162 static void efx_associate(struct efx_nic *efx)
1163 {
1164 	struct efx_nic *other, *next;
1165 
1166 	if (efx->primary == efx) {
1167 		/* Adding primary function; look for secondaries */
1168 
1169 		netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1170 		list_add_tail(&efx->node, &efx_primary_list);
1171 
1172 		list_for_each_entry_safe(other, next, &efx_unassociated_list,
1173 					 node) {
1174 			if (efx_same_controller(efx, other)) {
1175 				list_del(&other->node);
1176 				netif_dbg(other, probe, other->net_dev,
1177 					  "moving to secondary list of %s %s\n",
1178 					  pci_name(efx->pci_dev),
1179 					  efx->net_dev->name);
1180 				list_add_tail(&other->node,
1181 					      &efx->secondary_list);
1182 				other->primary = efx;
1183 			}
1184 		}
1185 	} else {
1186 		/* Adding secondary function; look for primary */
1187 
1188 		list_for_each_entry(other, &efx_primary_list, node) {
1189 			if (efx_same_controller(efx, other)) {
1190 				netif_dbg(efx, probe, efx->net_dev,
1191 					  "adding to secondary list of %s %s\n",
1192 					  pci_name(other->pci_dev),
1193 					  other->net_dev->name);
1194 				list_add_tail(&efx->node,
1195 					      &other->secondary_list);
1196 				efx->primary = other;
1197 				return;
1198 			}
1199 		}
1200 
1201 		netif_dbg(efx, probe, efx->net_dev,
1202 			  "adding to unassociated list\n");
1203 		list_add_tail(&efx->node, &efx_unassociated_list);
1204 	}
1205 }
1206 
1207 static void efx_dissociate(struct efx_nic *efx)
1208 {
1209 	struct efx_nic *other, *next;
1210 
1211 	list_del(&efx->node);
1212 	efx->primary = NULL;
1213 
1214 	list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1215 		list_del(&other->node);
1216 		netif_dbg(other, probe, other->net_dev,
1217 			  "moving to unassociated list\n");
1218 		list_add_tail(&other->node, &efx_unassociated_list);
1219 		other->primary = NULL;
1220 	}
1221 }
1222 
1223 /* This configures the PCI device to enable I/O and DMA. */
1224 static int efx_init_io(struct efx_nic *efx)
1225 {
1226 	struct pci_dev *pci_dev = efx->pci_dev;
1227 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1228 	unsigned int mem_map_size = efx->type->mem_map_size(efx);
1229 	int rc, bar;
1230 
1231 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1232 
1233 	bar = efx->type->mem_bar;
1234 
1235 	rc = pci_enable_device(pci_dev);
1236 	if (rc) {
1237 		netif_err(efx, probe, efx->net_dev,
1238 			  "failed to enable PCI device\n");
1239 		goto fail1;
1240 	}
1241 
1242 	pci_set_master(pci_dev);
1243 
1244 	/* Set the PCI DMA mask.  Try all possibilities from our
1245 	 * genuine mask down to 32 bits, because some architectures
1246 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1247 	 * masks event though they reject 46 bit masks.
1248 	 */
1249 	while (dma_mask > 0x7fffffffUL) {
1250 		if (dma_supported(&pci_dev->dev, dma_mask)) {
1251 			rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1252 			if (rc == 0)
1253 				break;
1254 		}
1255 		dma_mask >>= 1;
1256 	}
1257 	if (rc) {
1258 		netif_err(efx, probe, efx->net_dev,
1259 			  "could not find a suitable DMA mask\n");
1260 		goto fail2;
1261 	}
1262 	netif_dbg(efx, probe, efx->net_dev,
1263 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1264 
1265 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1266 	rc = pci_request_region(pci_dev, bar, "sfc");
1267 	if (rc) {
1268 		netif_err(efx, probe, efx->net_dev,
1269 			  "request for memory BAR failed\n");
1270 		rc = -EIO;
1271 		goto fail3;
1272 	}
1273 	efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1274 	if (!efx->membase) {
1275 		netif_err(efx, probe, efx->net_dev,
1276 			  "could not map memory BAR at %llx+%x\n",
1277 			  (unsigned long long)efx->membase_phys, mem_map_size);
1278 		rc = -ENOMEM;
1279 		goto fail4;
1280 	}
1281 	netif_dbg(efx, probe, efx->net_dev,
1282 		  "memory BAR at %llx+%x (virtual %p)\n",
1283 		  (unsigned long long)efx->membase_phys, mem_map_size,
1284 		  efx->membase);
1285 
1286 	return 0;
1287 
1288  fail4:
1289 	pci_release_region(efx->pci_dev, bar);
1290  fail3:
1291 	efx->membase_phys = 0;
1292  fail2:
1293 	pci_disable_device(efx->pci_dev);
1294  fail1:
1295 	return rc;
1296 }
1297 
1298 static void efx_fini_io(struct efx_nic *efx)
1299 {
1300 	int bar;
1301 
1302 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1303 
1304 	if (efx->membase) {
1305 		iounmap(efx->membase);
1306 		efx->membase = NULL;
1307 	}
1308 
1309 	if (efx->membase_phys) {
1310 		bar = efx->type->mem_bar;
1311 		pci_release_region(efx->pci_dev, bar);
1312 		efx->membase_phys = 0;
1313 	}
1314 
1315 	/* Don't disable bus-mastering if VFs are assigned */
1316 	if (!pci_vfs_assigned(efx->pci_dev))
1317 		pci_disable_device(efx->pci_dev);
1318 }
1319 
1320 void efx_set_default_rx_indir_table(struct efx_nic *efx)
1321 {
1322 	size_t i;
1323 
1324 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1325 		efx->rx_indir_table[i] =
1326 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1327 }
1328 
1329 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1330 {
1331 	cpumask_var_t thread_mask;
1332 	unsigned int count;
1333 	int cpu;
1334 
1335 	if (rss_cpus) {
1336 		count = rss_cpus;
1337 	} else {
1338 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1339 			netif_warn(efx, probe, efx->net_dev,
1340 				   "RSS disabled due to allocation failure\n");
1341 			return 1;
1342 		}
1343 
1344 		count = 0;
1345 		for_each_online_cpu(cpu) {
1346 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1347 				++count;
1348 				cpumask_or(thread_mask, thread_mask,
1349 					   topology_sibling_cpumask(cpu));
1350 			}
1351 		}
1352 
1353 		free_cpumask_var(thread_mask);
1354 	}
1355 
1356 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
1357 	 * table entries that are inaccessible to VFs
1358 	 */
1359 #ifdef CONFIG_SFC_SRIOV
1360 	if (efx->type->sriov_wanted) {
1361 		if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1362 		    count > efx_vf_size(efx)) {
1363 			netif_warn(efx, probe, efx->net_dev,
1364 				   "Reducing number of RSS channels from %u to %u for "
1365 				   "VF support. Increase vf-msix-limit to use more "
1366 				   "channels on the PF.\n",
1367 				   count, efx_vf_size(efx));
1368 			count = efx_vf_size(efx);
1369 		}
1370 	}
1371 #endif
1372 
1373 	return count;
1374 }
1375 
1376 /* Probe the number and type of interrupts we are able to obtain, and
1377  * the resulting numbers of channels and RX queues.
1378  */
1379 static int efx_probe_interrupts(struct efx_nic *efx)
1380 {
1381 	unsigned int extra_channels = 0;
1382 	unsigned int i, j;
1383 	int rc;
1384 
1385 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1386 		if (efx->extra_channel_type[i])
1387 			++extra_channels;
1388 
1389 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1390 		struct msix_entry xentries[EFX_MAX_CHANNELS];
1391 		unsigned int n_channels;
1392 
1393 		n_channels = efx_wanted_parallelism(efx);
1394 		if (efx_separate_tx_channels)
1395 			n_channels *= 2;
1396 		n_channels += extra_channels;
1397 		n_channels = min(n_channels, efx->max_channels);
1398 
1399 		for (i = 0; i < n_channels; i++)
1400 			xentries[i].entry = i;
1401 		rc = pci_enable_msix_range(efx->pci_dev,
1402 					   xentries, 1, n_channels);
1403 		if (rc < 0) {
1404 			/* Fall back to single channel MSI */
1405 			efx->interrupt_mode = EFX_INT_MODE_MSI;
1406 			netif_err(efx, drv, efx->net_dev,
1407 				  "could not enable MSI-X\n");
1408 		} else if (rc < n_channels) {
1409 			netif_err(efx, drv, efx->net_dev,
1410 				  "WARNING: Insufficient MSI-X vectors"
1411 				  " available (%d < %u).\n", rc, n_channels);
1412 			netif_err(efx, drv, efx->net_dev,
1413 				  "WARNING: Performance may be reduced.\n");
1414 			n_channels = rc;
1415 		}
1416 
1417 		if (rc > 0) {
1418 			efx->n_channels = n_channels;
1419 			if (n_channels > extra_channels)
1420 				n_channels -= extra_channels;
1421 			if (efx_separate_tx_channels) {
1422 				efx->n_tx_channels = min(max(n_channels / 2,
1423 							     1U),
1424 							 efx->max_tx_channels);
1425 				efx->n_rx_channels = max(n_channels -
1426 							 efx->n_tx_channels,
1427 							 1U);
1428 			} else {
1429 				efx->n_tx_channels = min(n_channels,
1430 							 efx->max_tx_channels);
1431 				efx->n_rx_channels = n_channels;
1432 			}
1433 			for (i = 0; i < efx->n_channels; i++)
1434 				efx_get_channel(efx, i)->irq =
1435 					xentries[i].vector;
1436 		}
1437 	}
1438 
1439 	/* Try single interrupt MSI */
1440 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1441 		efx->n_channels = 1;
1442 		efx->n_rx_channels = 1;
1443 		efx->n_tx_channels = 1;
1444 		rc = pci_enable_msi(efx->pci_dev);
1445 		if (rc == 0) {
1446 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1447 		} else {
1448 			netif_err(efx, drv, efx->net_dev,
1449 				  "could not enable MSI\n");
1450 			efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1451 		}
1452 	}
1453 
1454 	/* Assume legacy interrupts */
1455 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1456 		efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
1457 		efx->n_rx_channels = 1;
1458 		efx->n_tx_channels = 1;
1459 		efx->legacy_irq = efx->pci_dev->irq;
1460 	}
1461 
1462 	/* Assign extra channels if possible */
1463 	j = efx->n_channels;
1464 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1465 		if (!efx->extra_channel_type[i])
1466 			continue;
1467 		if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1468 		    efx->n_channels <= extra_channels) {
1469 			efx->extra_channel_type[i]->handle_no_channel(efx);
1470 		} else {
1471 			--j;
1472 			efx_get_channel(efx, j)->type =
1473 				efx->extra_channel_type[i];
1474 		}
1475 	}
1476 
1477 	/* RSS might be usable on VFs even if it is disabled on the PF */
1478 #ifdef CONFIG_SFC_SRIOV
1479 	if (efx->type->sriov_wanted) {
1480 		efx->rss_spread = ((efx->n_rx_channels > 1 ||
1481 				    !efx->type->sriov_wanted(efx)) ?
1482 				   efx->n_rx_channels : efx_vf_size(efx));
1483 		return 0;
1484 	}
1485 #endif
1486 	efx->rss_spread = efx->n_rx_channels;
1487 
1488 	return 0;
1489 }
1490 
1491 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1492 {
1493 	struct efx_channel *channel, *end_channel;
1494 	int rc;
1495 
1496 	BUG_ON(efx->state == STATE_DISABLED);
1497 
1498 	efx->irq_soft_enabled = true;
1499 	smp_wmb();
1500 
1501 	efx_for_each_channel(channel, efx) {
1502 		if (!channel->type->keep_eventq) {
1503 			rc = efx_init_eventq(channel);
1504 			if (rc)
1505 				goto fail;
1506 		}
1507 		efx_start_eventq(channel);
1508 	}
1509 
1510 	efx_mcdi_mode_event(efx);
1511 
1512 	return 0;
1513 fail:
1514 	end_channel = channel;
1515 	efx_for_each_channel(channel, efx) {
1516 		if (channel == end_channel)
1517 			break;
1518 		efx_stop_eventq(channel);
1519 		if (!channel->type->keep_eventq)
1520 			efx_fini_eventq(channel);
1521 	}
1522 
1523 	return rc;
1524 }
1525 
1526 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1527 {
1528 	struct efx_channel *channel;
1529 
1530 	if (efx->state == STATE_DISABLED)
1531 		return;
1532 
1533 	efx_mcdi_mode_poll(efx);
1534 
1535 	efx->irq_soft_enabled = false;
1536 	smp_wmb();
1537 
1538 	if (efx->legacy_irq)
1539 		synchronize_irq(efx->legacy_irq);
1540 
1541 	efx_for_each_channel(channel, efx) {
1542 		if (channel->irq)
1543 			synchronize_irq(channel->irq);
1544 
1545 		efx_stop_eventq(channel);
1546 		if (!channel->type->keep_eventq)
1547 			efx_fini_eventq(channel);
1548 	}
1549 
1550 	/* Flush the asynchronous MCDI request queue */
1551 	efx_mcdi_flush_async(efx);
1552 }
1553 
1554 static int efx_enable_interrupts(struct efx_nic *efx)
1555 {
1556 	struct efx_channel *channel, *end_channel;
1557 	int rc;
1558 
1559 	BUG_ON(efx->state == STATE_DISABLED);
1560 
1561 	if (efx->eeh_disabled_legacy_irq) {
1562 		enable_irq(efx->legacy_irq);
1563 		efx->eeh_disabled_legacy_irq = false;
1564 	}
1565 
1566 	efx->type->irq_enable_master(efx);
1567 
1568 	efx_for_each_channel(channel, efx) {
1569 		if (channel->type->keep_eventq) {
1570 			rc = efx_init_eventq(channel);
1571 			if (rc)
1572 				goto fail;
1573 		}
1574 	}
1575 
1576 	rc = efx_soft_enable_interrupts(efx);
1577 	if (rc)
1578 		goto fail;
1579 
1580 	return 0;
1581 
1582 fail:
1583 	end_channel = channel;
1584 	efx_for_each_channel(channel, efx) {
1585 		if (channel == end_channel)
1586 			break;
1587 		if (channel->type->keep_eventq)
1588 			efx_fini_eventq(channel);
1589 	}
1590 
1591 	efx->type->irq_disable_non_ev(efx);
1592 
1593 	return rc;
1594 }
1595 
1596 static void efx_disable_interrupts(struct efx_nic *efx)
1597 {
1598 	struct efx_channel *channel;
1599 
1600 	efx_soft_disable_interrupts(efx);
1601 
1602 	efx_for_each_channel(channel, efx) {
1603 		if (channel->type->keep_eventq)
1604 			efx_fini_eventq(channel);
1605 	}
1606 
1607 	efx->type->irq_disable_non_ev(efx);
1608 }
1609 
1610 static void efx_remove_interrupts(struct efx_nic *efx)
1611 {
1612 	struct efx_channel *channel;
1613 
1614 	/* Remove MSI/MSI-X interrupts */
1615 	efx_for_each_channel(channel, efx)
1616 		channel->irq = 0;
1617 	pci_disable_msi(efx->pci_dev);
1618 	pci_disable_msix(efx->pci_dev);
1619 
1620 	/* Remove legacy interrupt */
1621 	efx->legacy_irq = 0;
1622 }
1623 
1624 static void efx_set_channels(struct efx_nic *efx)
1625 {
1626 	struct efx_channel *channel;
1627 	struct efx_tx_queue *tx_queue;
1628 
1629 	efx->tx_channel_offset =
1630 		efx_separate_tx_channels ?
1631 		efx->n_channels - efx->n_tx_channels : 0;
1632 
1633 	/* We need to mark which channels really have RX and TX
1634 	 * queues, and adjust the TX queue numbers if we have separate
1635 	 * RX-only and TX-only channels.
1636 	 */
1637 	efx_for_each_channel(channel, efx) {
1638 		if (channel->channel < efx->n_rx_channels)
1639 			channel->rx_queue.core_index = channel->channel;
1640 		else
1641 			channel->rx_queue.core_index = -1;
1642 
1643 		efx_for_each_channel_tx_queue(tx_queue, channel)
1644 			tx_queue->queue -= (efx->tx_channel_offset *
1645 					    EFX_TXQ_TYPES);
1646 	}
1647 }
1648 
1649 static int efx_probe_nic(struct efx_nic *efx)
1650 {
1651 	int rc;
1652 
1653 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1654 
1655 	/* Carry out hardware-type specific initialisation */
1656 	rc = efx->type->probe(efx);
1657 	if (rc)
1658 		return rc;
1659 
1660 	do {
1661 		if (!efx->max_channels || !efx->max_tx_channels) {
1662 			netif_err(efx, drv, efx->net_dev,
1663 				  "Insufficient resources to allocate"
1664 				  " any channels\n");
1665 			rc = -ENOSPC;
1666 			goto fail1;
1667 		}
1668 
1669 		/* Determine the number of channels and queues by trying
1670 		 * to hook in MSI-X interrupts.
1671 		 */
1672 		rc = efx_probe_interrupts(efx);
1673 		if (rc)
1674 			goto fail1;
1675 
1676 		efx_set_channels(efx);
1677 
1678 		/* dimension_resources can fail with EAGAIN */
1679 		rc = efx->type->dimension_resources(efx);
1680 		if (rc != 0 && rc != -EAGAIN)
1681 			goto fail2;
1682 
1683 		if (rc == -EAGAIN)
1684 			/* try again with new max_channels */
1685 			efx_remove_interrupts(efx);
1686 
1687 	} while (rc == -EAGAIN);
1688 
1689 	if (efx->n_channels > 1)
1690 		netdev_rss_key_fill(&efx->rx_hash_key,
1691 				    sizeof(efx->rx_hash_key));
1692 	efx_set_default_rx_indir_table(efx);
1693 
1694 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1695 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1696 
1697 	/* Initialise the interrupt moderation settings */
1698 	efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1699 				true);
1700 
1701 	return 0;
1702 
1703 fail2:
1704 	efx_remove_interrupts(efx);
1705 fail1:
1706 	efx->type->remove(efx);
1707 	return rc;
1708 }
1709 
1710 static void efx_remove_nic(struct efx_nic *efx)
1711 {
1712 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1713 
1714 	efx_remove_interrupts(efx);
1715 	efx->type->remove(efx);
1716 }
1717 
1718 static int efx_probe_filters(struct efx_nic *efx)
1719 {
1720 	int rc;
1721 
1722 	spin_lock_init(&efx->filter_lock);
1723 	init_rwsem(&efx->filter_sem);
1724 	down_write(&efx->filter_sem);
1725 	rc = efx->type->filter_table_probe(efx);
1726 	if (rc)
1727 		goto out_unlock;
1728 
1729 #ifdef CONFIG_RFS_ACCEL
1730 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
1731 		efx->rps_flow_id = kcalloc(efx->type->max_rx_ip_filters,
1732 					   sizeof(*efx->rps_flow_id),
1733 					   GFP_KERNEL);
1734 		if (!efx->rps_flow_id) {
1735 			efx->type->filter_table_remove(efx);
1736 			rc = -ENOMEM;
1737 			goto out_unlock;
1738 		}
1739 	}
1740 #endif
1741 out_unlock:
1742 	up_write(&efx->filter_sem);
1743 	return rc;
1744 }
1745 
1746 static void efx_remove_filters(struct efx_nic *efx)
1747 {
1748 #ifdef CONFIG_RFS_ACCEL
1749 	kfree(efx->rps_flow_id);
1750 #endif
1751 	down_write(&efx->filter_sem);
1752 	efx->type->filter_table_remove(efx);
1753 	up_write(&efx->filter_sem);
1754 }
1755 
1756 static void efx_restore_filters(struct efx_nic *efx)
1757 {
1758 	down_read(&efx->filter_sem);
1759 	efx->type->filter_table_restore(efx);
1760 	up_read(&efx->filter_sem);
1761 }
1762 
1763 /**************************************************************************
1764  *
1765  * NIC startup/shutdown
1766  *
1767  *************************************************************************/
1768 
1769 static int efx_probe_all(struct efx_nic *efx)
1770 {
1771 	int rc;
1772 
1773 	rc = efx_probe_nic(efx);
1774 	if (rc) {
1775 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1776 		goto fail1;
1777 	}
1778 
1779 	rc = efx_probe_port(efx);
1780 	if (rc) {
1781 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1782 		goto fail2;
1783 	}
1784 
1785 	BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1786 	if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1787 		rc = -EINVAL;
1788 		goto fail3;
1789 	}
1790 	efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1791 
1792 #ifdef CONFIG_SFC_SRIOV
1793 	rc = efx->type->vswitching_probe(efx);
1794 	if (rc) /* not fatal; the PF will still work fine */
1795 		netif_warn(efx, probe, efx->net_dev,
1796 			   "failed to setup vswitching rc=%d;"
1797 			   " VFs may not function\n", rc);
1798 #endif
1799 
1800 	rc = efx_probe_filters(efx);
1801 	if (rc) {
1802 		netif_err(efx, probe, efx->net_dev,
1803 			  "failed to create filter tables\n");
1804 		goto fail4;
1805 	}
1806 
1807 	rc = efx_probe_channels(efx);
1808 	if (rc)
1809 		goto fail5;
1810 
1811 	return 0;
1812 
1813  fail5:
1814 	efx_remove_filters(efx);
1815  fail4:
1816 #ifdef CONFIG_SFC_SRIOV
1817 	efx->type->vswitching_remove(efx);
1818 #endif
1819  fail3:
1820 	efx_remove_port(efx);
1821  fail2:
1822 	efx_remove_nic(efx);
1823  fail1:
1824 	return rc;
1825 }
1826 
1827 /* If the interface is supposed to be running but is not, start
1828  * the hardware and software data path, regular activity for the port
1829  * (MAC statistics, link polling, etc.) and schedule the port to be
1830  * reconfigured.  Interrupts must already be enabled.  This function
1831  * is safe to call multiple times, so long as the NIC is not disabled.
1832  * Requires the RTNL lock.
1833  */
1834 static void efx_start_all(struct efx_nic *efx)
1835 {
1836 	EFX_ASSERT_RESET_SERIALISED(efx);
1837 	BUG_ON(efx->state == STATE_DISABLED);
1838 
1839 	/* Check that it is appropriate to restart the interface. All
1840 	 * of these flags are safe to read under just the rtnl lock */
1841 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
1842 	    efx->reset_pending)
1843 		return;
1844 
1845 	efx_start_port(efx);
1846 	efx_start_datapath(efx);
1847 
1848 	/* Start the hardware monitor if there is one */
1849 	if (efx->type->monitor != NULL)
1850 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1851 				   efx_monitor_interval);
1852 
1853 	/* If link state detection is normally event-driven, we have
1854 	 * to poll now because we could have missed a change
1855 	 */
1856 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1857 		mutex_lock(&efx->mac_lock);
1858 		if (efx->phy_op->poll(efx))
1859 			efx_link_status_changed(efx);
1860 		mutex_unlock(&efx->mac_lock);
1861 	}
1862 
1863 	efx->type->start_stats(efx);
1864 	efx->type->pull_stats(efx);
1865 	spin_lock_bh(&efx->stats_lock);
1866 	efx->type->update_stats(efx, NULL, NULL);
1867 	spin_unlock_bh(&efx->stats_lock);
1868 }
1869 
1870 /* Quiesce the hardware and software data path, and regular activity
1871  * for the port without bringing the link down.  Safe to call multiple
1872  * times with the NIC in almost any state, but interrupts should be
1873  * enabled.  Requires the RTNL lock.
1874  */
1875 static void efx_stop_all(struct efx_nic *efx)
1876 {
1877 	EFX_ASSERT_RESET_SERIALISED(efx);
1878 
1879 	/* port_enabled can be read safely under the rtnl lock */
1880 	if (!efx->port_enabled)
1881 		return;
1882 
1883 	/* update stats before we go down so we can accurately count
1884 	 * rx_nodesc_drops
1885 	 */
1886 	efx->type->pull_stats(efx);
1887 	spin_lock_bh(&efx->stats_lock);
1888 	efx->type->update_stats(efx, NULL, NULL);
1889 	spin_unlock_bh(&efx->stats_lock);
1890 	efx->type->stop_stats(efx);
1891 	efx_stop_port(efx);
1892 
1893 	/* Stop the kernel transmit interface.  This is only valid if
1894 	 * the device is stopped or detached; otherwise the watchdog
1895 	 * may fire immediately.
1896 	 */
1897 	WARN_ON(netif_running(efx->net_dev) &&
1898 		netif_device_present(efx->net_dev));
1899 	netif_tx_disable(efx->net_dev);
1900 
1901 	efx_stop_datapath(efx);
1902 }
1903 
1904 static void efx_remove_all(struct efx_nic *efx)
1905 {
1906 	efx_remove_channels(efx);
1907 	efx_remove_filters(efx);
1908 #ifdef CONFIG_SFC_SRIOV
1909 	efx->type->vswitching_remove(efx);
1910 #endif
1911 	efx_remove_port(efx);
1912 	efx_remove_nic(efx);
1913 }
1914 
1915 /**************************************************************************
1916  *
1917  * Interrupt moderation
1918  *
1919  **************************************************************************/
1920 
1921 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1922 {
1923 	if (usecs == 0)
1924 		return 0;
1925 	if (usecs * 1000 < quantum_ns)
1926 		return 1; /* never round down to 0 */
1927 	return usecs * 1000 / quantum_ns;
1928 }
1929 
1930 /* Set interrupt moderation parameters */
1931 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1932 			    unsigned int rx_usecs, bool rx_adaptive,
1933 			    bool rx_may_override_tx)
1934 {
1935 	struct efx_channel *channel;
1936 	unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1937 						efx->timer_quantum_ns,
1938 						1000);
1939 	unsigned int tx_ticks;
1940 	unsigned int rx_ticks;
1941 
1942 	EFX_ASSERT_RESET_SERIALISED(efx);
1943 
1944 	if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1945 		return -EINVAL;
1946 
1947 	tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1948 	rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1949 
1950 	if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1951 	    !rx_may_override_tx) {
1952 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1953 			  "RX and TX IRQ moderation must be equal\n");
1954 		return -EINVAL;
1955 	}
1956 
1957 	efx->irq_rx_adaptive = rx_adaptive;
1958 	efx->irq_rx_moderation = rx_ticks;
1959 	efx_for_each_channel(channel, efx) {
1960 		if (efx_channel_has_rx_queue(channel))
1961 			channel->irq_moderation = rx_ticks;
1962 		else if (efx_channel_has_tx_queues(channel))
1963 			channel->irq_moderation = tx_ticks;
1964 	}
1965 
1966 	return 0;
1967 }
1968 
1969 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1970 			    unsigned int *rx_usecs, bool *rx_adaptive)
1971 {
1972 	/* We must round up when converting ticks to microseconds
1973 	 * because we round down when converting the other way.
1974 	 */
1975 
1976 	*rx_adaptive = efx->irq_rx_adaptive;
1977 	*rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1978 				 efx->timer_quantum_ns,
1979 				 1000);
1980 
1981 	/* If channels are shared between RX and TX, so is IRQ
1982 	 * moderation.  Otherwise, IRQ moderation is the same for all
1983 	 * TX channels and is not adaptive.
1984 	 */
1985 	if (efx->tx_channel_offset == 0)
1986 		*tx_usecs = *rx_usecs;
1987 	else
1988 		*tx_usecs = DIV_ROUND_UP(
1989 			efx->channel[efx->tx_channel_offset]->irq_moderation *
1990 			efx->timer_quantum_ns,
1991 			1000);
1992 }
1993 
1994 /**************************************************************************
1995  *
1996  * Hardware monitor
1997  *
1998  **************************************************************************/
1999 
2000 /* Run periodically off the general workqueue */
2001 static void efx_monitor(struct work_struct *data)
2002 {
2003 	struct efx_nic *efx = container_of(data, struct efx_nic,
2004 					   monitor_work.work);
2005 
2006 	netif_vdbg(efx, timer, efx->net_dev,
2007 		   "hardware monitor executing on CPU %d\n",
2008 		   raw_smp_processor_id());
2009 	BUG_ON(efx->type->monitor == NULL);
2010 
2011 	/* If the mac_lock is already held then it is likely a port
2012 	 * reconfiguration is already in place, which will likely do
2013 	 * most of the work of monitor() anyway. */
2014 	if (mutex_trylock(&efx->mac_lock)) {
2015 		if (efx->port_enabled)
2016 			efx->type->monitor(efx);
2017 		mutex_unlock(&efx->mac_lock);
2018 	}
2019 
2020 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
2021 			   efx_monitor_interval);
2022 }
2023 
2024 /**************************************************************************
2025  *
2026  * ioctls
2027  *
2028  *************************************************************************/
2029 
2030 /* Net device ioctl
2031  * Context: process, rtnl_lock() held.
2032  */
2033 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
2034 {
2035 	struct efx_nic *efx = netdev_priv(net_dev);
2036 	struct mii_ioctl_data *data = if_mii(ifr);
2037 
2038 	if (cmd == SIOCSHWTSTAMP)
2039 		return efx_ptp_set_ts_config(efx, ifr);
2040 	if (cmd == SIOCGHWTSTAMP)
2041 		return efx_ptp_get_ts_config(efx, ifr);
2042 
2043 	/* Convert phy_id from older PRTAD/DEVAD format */
2044 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2045 	    (data->phy_id & 0xfc00) == 0x0400)
2046 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2047 
2048 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
2049 }
2050 
2051 /**************************************************************************
2052  *
2053  * NAPI interface
2054  *
2055  **************************************************************************/
2056 
2057 static void efx_init_napi_channel(struct efx_channel *channel)
2058 {
2059 	struct efx_nic *efx = channel->efx;
2060 
2061 	channel->napi_dev = efx->net_dev;
2062 	netif_napi_add(channel->napi_dev, &channel->napi_str,
2063 		       efx_poll, napi_weight);
2064 	napi_hash_add(&channel->napi_str);
2065 	efx_channel_init_lock(channel);
2066 }
2067 
2068 static void efx_init_napi(struct efx_nic *efx)
2069 {
2070 	struct efx_channel *channel;
2071 
2072 	efx_for_each_channel(channel, efx)
2073 		efx_init_napi_channel(channel);
2074 }
2075 
2076 static void efx_fini_napi_channel(struct efx_channel *channel)
2077 {
2078 	if (channel->napi_dev) {
2079 		netif_napi_del(&channel->napi_str);
2080 		napi_hash_del(&channel->napi_str);
2081 	}
2082 	channel->napi_dev = NULL;
2083 }
2084 
2085 static void efx_fini_napi(struct efx_nic *efx)
2086 {
2087 	struct efx_channel *channel;
2088 
2089 	efx_for_each_channel(channel, efx)
2090 		efx_fini_napi_channel(channel);
2091 }
2092 
2093 /**************************************************************************
2094  *
2095  * Kernel netpoll interface
2096  *
2097  *************************************************************************/
2098 
2099 #ifdef CONFIG_NET_POLL_CONTROLLER
2100 
2101 /* Although in the common case interrupts will be disabled, this is not
2102  * guaranteed. However, all our work happens inside the NAPI callback,
2103  * so no locking is required.
2104  */
2105 static void efx_netpoll(struct net_device *net_dev)
2106 {
2107 	struct efx_nic *efx = netdev_priv(net_dev);
2108 	struct efx_channel *channel;
2109 
2110 	efx_for_each_channel(channel, efx)
2111 		efx_schedule_channel(channel);
2112 }
2113 
2114 #endif
2115 
2116 #ifdef CONFIG_NET_RX_BUSY_POLL
2117 static int efx_busy_poll(struct napi_struct *napi)
2118 {
2119 	struct efx_channel *channel =
2120 		container_of(napi, struct efx_channel, napi_str);
2121 	struct efx_nic *efx = channel->efx;
2122 	int budget = 4;
2123 	int old_rx_packets, rx_packets;
2124 
2125 	if (!netif_running(efx->net_dev))
2126 		return LL_FLUSH_FAILED;
2127 
2128 	if (!efx_channel_lock_poll(channel))
2129 		return LL_FLUSH_BUSY;
2130 
2131 	old_rx_packets = channel->rx_queue.rx_packets;
2132 	efx_process_channel(channel, budget);
2133 
2134 	rx_packets = channel->rx_queue.rx_packets - old_rx_packets;
2135 
2136 	/* There is no race condition with NAPI here.
2137 	 * NAPI will automatically be rescheduled if it yielded during busy
2138 	 * polling, because it was not able to take the lock and thus returned
2139 	 * the full budget.
2140 	 */
2141 	efx_channel_unlock_poll(channel);
2142 
2143 	return rx_packets;
2144 }
2145 #endif
2146 
2147 /**************************************************************************
2148  *
2149  * Kernel net device interface
2150  *
2151  *************************************************************************/
2152 
2153 /* Context: process, rtnl_lock() held. */
2154 int efx_net_open(struct net_device *net_dev)
2155 {
2156 	struct efx_nic *efx = netdev_priv(net_dev);
2157 	int rc;
2158 
2159 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2160 		  raw_smp_processor_id());
2161 
2162 	rc = efx_check_disabled(efx);
2163 	if (rc)
2164 		return rc;
2165 	if (efx->phy_mode & PHY_MODE_SPECIAL)
2166 		return -EBUSY;
2167 	if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2168 		return -EIO;
2169 
2170 	/* Notify the kernel of the link state polled during driver load,
2171 	 * before the monitor starts running */
2172 	efx_link_status_changed(efx);
2173 
2174 	efx_start_all(efx);
2175 	efx_selftest_async_start(efx);
2176 	return 0;
2177 }
2178 
2179 /* Context: process, rtnl_lock() held.
2180  * Note that the kernel will ignore our return code; this method
2181  * should really be a void.
2182  */
2183 int efx_net_stop(struct net_device *net_dev)
2184 {
2185 	struct efx_nic *efx = netdev_priv(net_dev);
2186 
2187 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2188 		  raw_smp_processor_id());
2189 
2190 	/* Stop the device and flush all the channels */
2191 	efx_stop_all(efx);
2192 
2193 	return 0;
2194 }
2195 
2196 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2197 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
2198 					       struct rtnl_link_stats64 *stats)
2199 {
2200 	struct efx_nic *efx = netdev_priv(net_dev);
2201 
2202 	spin_lock_bh(&efx->stats_lock);
2203 	efx->type->update_stats(efx, NULL, stats);
2204 	spin_unlock_bh(&efx->stats_lock);
2205 
2206 	return stats;
2207 }
2208 
2209 /* Context: netif_tx_lock held, BHs disabled. */
2210 static void efx_watchdog(struct net_device *net_dev)
2211 {
2212 	struct efx_nic *efx = netdev_priv(net_dev);
2213 
2214 	netif_err(efx, tx_err, efx->net_dev,
2215 		  "TX stuck with port_enabled=%d: resetting channels\n",
2216 		  efx->port_enabled);
2217 
2218 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2219 }
2220 
2221 
2222 /* Context: process, rtnl_lock() held. */
2223 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2224 {
2225 	struct efx_nic *efx = netdev_priv(net_dev);
2226 	int rc;
2227 
2228 	rc = efx_check_disabled(efx);
2229 	if (rc)
2230 		return rc;
2231 	if (new_mtu > EFX_MAX_MTU)
2232 		return -EINVAL;
2233 
2234 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2235 
2236 	efx_device_detach_sync(efx);
2237 	efx_stop_all(efx);
2238 
2239 	mutex_lock(&efx->mac_lock);
2240 	net_dev->mtu = new_mtu;
2241 	efx_mac_reconfigure(efx);
2242 	mutex_unlock(&efx->mac_lock);
2243 
2244 	efx_start_all(efx);
2245 	netif_device_attach(efx->net_dev);
2246 	return 0;
2247 }
2248 
2249 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2250 {
2251 	struct efx_nic *efx = netdev_priv(net_dev);
2252 	struct sockaddr *addr = data;
2253 	u8 *new_addr = addr->sa_data;
2254 	u8 old_addr[6];
2255 	int rc;
2256 
2257 	if (!is_valid_ether_addr(new_addr)) {
2258 		netif_err(efx, drv, efx->net_dev,
2259 			  "invalid ethernet MAC address requested: %pM\n",
2260 			  new_addr);
2261 		return -EADDRNOTAVAIL;
2262 	}
2263 
2264 	/* save old address */
2265 	ether_addr_copy(old_addr, net_dev->dev_addr);
2266 	ether_addr_copy(net_dev->dev_addr, new_addr);
2267 	if (efx->type->set_mac_address) {
2268 		rc = efx->type->set_mac_address(efx);
2269 		if (rc) {
2270 			ether_addr_copy(net_dev->dev_addr, old_addr);
2271 			return rc;
2272 		}
2273 	}
2274 
2275 	/* Reconfigure the MAC */
2276 	mutex_lock(&efx->mac_lock);
2277 	efx_mac_reconfigure(efx);
2278 	mutex_unlock(&efx->mac_lock);
2279 
2280 	return 0;
2281 }
2282 
2283 /* Context: netif_addr_lock held, BHs disabled. */
2284 static void efx_set_rx_mode(struct net_device *net_dev)
2285 {
2286 	struct efx_nic *efx = netdev_priv(net_dev);
2287 
2288 	if (efx->port_enabled)
2289 		queue_work(efx->workqueue, &efx->mac_work);
2290 	/* Otherwise efx_start_port() will do this */
2291 }
2292 
2293 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2294 {
2295 	struct efx_nic *efx = netdev_priv(net_dev);
2296 
2297 	/* If disabling RX n-tuple filtering, clear existing filters */
2298 	if (net_dev->features & ~data & NETIF_F_NTUPLE)
2299 		return efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2300 
2301 	return 0;
2302 }
2303 
2304 static const struct net_device_ops efx_netdev_ops = {
2305 	.ndo_open		= efx_net_open,
2306 	.ndo_stop		= efx_net_stop,
2307 	.ndo_get_stats64	= efx_net_stats,
2308 	.ndo_tx_timeout		= efx_watchdog,
2309 	.ndo_start_xmit		= efx_hard_start_xmit,
2310 	.ndo_validate_addr	= eth_validate_addr,
2311 	.ndo_do_ioctl		= efx_ioctl,
2312 	.ndo_change_mtu		= efx_change_mtu,
2313 	.ndo_set_mac_address	= efx_set_mac_address,
2314 	.ndo_set_rx_mode	= efx_set_rx_mode,
2315 	.ndo_set_features	= efx_set_features,
2316 #ifdef CONFIG_SFC_SRIOV
2317 	.ndo_set_vf_mac		= efx_sriov_set_vf_mac,
2318 	.ndo_set_vf_vlan	= efx_sriov_set_vf_vlan,
2319 	.ndo_set_vf_spoofchk	= efx_sriov_set_vf_spoofchk,
2320 	.ndo_get_vf_config	= efx_sriov_get_vf_config,
2321 	.ndo_set_vf_link_state  = efx_sriov_set_vf_link_state,
2322 	.ndo_get_phys_port_id   = efx_sriov_get_phys_port_id,
2323 #endif
2324 #ifdef CONFIG_NET_POLL_CONTROLLER
2325 	.ndo_poll_controller = efx_netpoll,
2326 #endif
2327 	.ndo_setup_tc		= efx_setup_tc,
2328 #ifdef CONFIG_NET_RX_BUSY_POLL
2329 	.ndo_busy_poll		= efx_busy_poll,
2330 #endif
2331 #ifdef CONFIG_RFS_ACCEL
2332 	.ndo_rx_flow_steer	= efx_filter_rfs,
2333 #endif
2334 };
2335 
2336 static void efx_update_name(struct efx_nic *efx)
2337 {
2338 	strcpy(efx->name, efx->net_dev->name);
2339 	efx_mtd_rename(efx);
2340 	efx_set_channel_names(efx);
2341 }
2342 
2343 static int efx_netdev_event(struct notifier_block *this,
2344 			    unsigned long event, void *ptr)
2345 {
2346 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2347 
2348 	if ((net_dev->netdev_ops == &efx_netdev_ops) &&
2349 	    event == NETDEV_CHANGENAME)
2350 		efx_update_name(netdev_priv(net_dev));
2351 
2352 	return NOTIFY_DONE;
2353 }
2354 
2355 static struct notifier_block efx_netdev_notifier = {
2356 	.notifier_call = efx_netdev_event,
2357 };
2358 
2359 static ssize_t
2360 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2361 {
2362 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2363 	return sprintf(buf, "%d\n", efx->phy_type);
2364 }
2365 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2366 
2367 #ifdef CONFIG_SFC_MCDI_LOGGING
2368 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
2369 			     char *buf)
2370 {
2371 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2372 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2373 
2374 	return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
2375 }
2376 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
2377 			    const char *buf, size_t count)
2378 {
2379 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2380 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2381 	bool enable = count > 0 && *buf != '0';
2382 
2383 	mcdi->logging_enabled = enable;
2384 	return count;
2385 }
2386 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
2387 #endif
2388 
2389 static int efx_register_netdev(struct efx_nic *efx)
2390 {
2391 	struct net_device *net_dev = efx->net_dev;
2392 	struct efx_channel *channel;
2393 	int rc;
2394 
2395 	net_dev->watchdog_timeo = 5 * HZ;
2396 	net_dev->irq = efx->pci_dev->irq;
2397 	net_dev->netdev_ops = &efx_netdev_ops;
2398 	if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
2399 		net_dev->priv_flags |= IFF_UNICAST_FLT;
2400 	net_dev->ethtool_ops = &efx_ethtool_ops;
2401 	net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2402 
2403 	rtnl_lock();
2404 
2405 	/* Enable resets to be scheduled and check whether any were
2406 	 * already requested.  If so, the NIC is probably hosed so we
2407 	 * abort.
2408 	 */
2409 	efx->state = STATE_READY;
2410 	smp_mb(); /* ensure we change state before checking reset_pending */
2411 	if (efx->reset_pending) {
2412 		netif_err(efx, probe, efx->net_dev,
2413 			  "aborting probe due to scheduled reset\n");
2414 		rc = -EIO;
2415 		goto fail_locked;
2416 	}
2417 
2418 	rc = dev_alloc_name(net_dev, net_dev->name);
2419 	if (rc < 0)
2420 		goto fail_locked;
2421 	efx_update_name(efx);
2422 
2423 	/* Always start with carrier off; PHY events will detect the link */
2424 	netif_carrier_off(net_dev);
2425 
2426 	rc = register_netdevice(net_dev);
2427 	if (rc)
2428 		goto fail_locked;
2429 
2430 	efx_for_each_channel(channel, efx) {
2431 		struct efx_tx_queue *tx_queue;
2432 		efx_for_each_channel_tx_queue(tx_queue, channel)
2433 			efx_init_tx_queue_core_txq(tx_queue);
2434 	}
2435 
2436 	efx_associate(efx);
2437 
2438 	rtnl_unlock();
2439 
2440 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2441 	if (rc) {
2442 		netif_err(efx, drv, efx->net_dev,
2443 			  "failed to init net dev attributes\n");
2444 		goto fail_registered;
2445 	}
2446 #ifdef CONFIG_SFC_MCDI_LOGGING
2447 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2448 	if (rc) {
2449 		netif_err(efx, drv, efx->net_dev,
2450 			  "failed to init net dev attributes\n");
2451 		goto fail_attr_mcdi_logging;
2452 	}
2453 #endif
2454 
2455 	return 0;
2456 
2457 #ifdef CONFIG_SFC_MCDI_LOGGING
2458 fail_attr_mcdi_logging:
2459 	device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2460 #endif
2461 fail_registered:
2462 	rtnl_lock();
2463 	efx_dissociate(efx);
2464 	unregister_netdevice(net_dev);
2465 fail_locked:
2466 	efx->state = STATE_UNINIT;
2467 	rtnl_unlock();
2468 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2469 	return rc;
2470 }
2471 
2472 static void efx_unregister_netdev(struct efx_nic *efx)
2473 {
2474 	if (!efx->net_dev)
2475 		return;
2476 
2477 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2478 
2479 	if (efx_dev_registered(efx)) {
2480 		strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2481 #ifdef CONFIG_SFC_MCDI_LOGGING
2482 		device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2483 #endif
2484 		device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2485 		unregister_netdev(efx->net_dev);
2486 	}
2487 }
2488 
2489 /**************************************************************************
2490  *
2491  * Device reset and suspend
2492  *
2493  **************************************************************************/
2494 
2495 /* Tears down the entire software state and most of the hardware state
2496  * before reset.  */
2497 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2498 {
2499 	EFX_ASSERT_RESET_SERIALISED(efx);
2500 
2501 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2502 		efx->type->prepare_flr(efx);
2503 
2504 	efx_stop_all(efx);
2505 	efx_disable_interrupts(efx);
2506 
2507 	mutex_lock(&efx->mac_lock);
2508 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2509 	    method != RESET_TYPE_DATAPATH)
2510 		efx->phy_op->fini(efx);
2511 	efx->type->fini(efx);
2512 }
2513 
2514 /* This function will always ensure that the locks acquired in
2515  * efx_reset_down() are released. A failure return code indicates
2516  * that we were unable to reinitialise the hardware, and the
2517  * driver should be disabled. If ok is false, then the rx and tx
2518  * engines are not restarted, pending a RESET_DISABLE. */
2519 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2520 {
2521 	int rc;
2522 
2523 	EFX_ASSERT_RESET_SERIALISED(efx);
2524 
2525 	if (method == RESET_TYPE_MCDI_TIMEOUT)
2526 		efx->type->finish_flr(efx);
2527 
2528 	/* Ensure that SRAM is initialised even if we're disabling the device */
2529 	rc = efx->type->init(efx);
2530 	if (rc) {
2531 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2532 		goto fail;
2533 	}
2534 
2535 	if (!ok)
2536 		goto fail;
2537 
2538 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2539 	    method != RESET_TYPE_DATAPATH) {
2540 		rc = efx->phy_op->init(efx);
2541 		if (rc)
2542 			goto fail;
2543 		rc = efx->phy_op->reconfigure(efx);
2544 		if (rc && rc != -EPERM)
2545 			netif_err(efx, drv, efx->net_dev,
2546 				  "could not restore PHY settings\n");
2547 	}
2548 
2549 	rc = efx_enable_interrupts(efx);
2550 	if (rc)
2551 		goto fail;
2552 
2553 #ifdef CONFIG_SFC_SRIOV
2554 	rc = efx->type->vswitching_restore(efx);
2555 	if (rc) /* not fatal; the PF will still work fine */
2556 		netif_warn(efx, probe, efx->net_dev,
2557 			   "failed to restore vswitching rc=%d;"
2558 			   " VFs may not function\n", rc);
2559 #endif
2560 
2561 	down_read(&efx->filter_sem);
2562 	efx_restore_filters(efx);
2563 	up_read(&efx->filter_sem);
2564 	if (efx->type->sriov_reset)
2565 		efx->type->sriov_reset(efx);
2566 
2567 	mutex_unlock(&efx->mac_lock);
2568 
2569 	efx_start_all(efx);
2570 
2571 	return 0;
2572 
2573 fail:
2574 	efx->port_initialized = false;
2575 
2576 	mutex_unlock(&efx->mac_lock);
2577 
2578 	return rc;
2579 }
2580 
2581 /* Reset the NIC using the specified method.  Note that the reset may
2582  * fail, in which case the card will be left in an unusable state.
2583  *
2584  * Caller must hold the rtnl_lock.
2585  */
2586 int efx_reset(struct efx_nic *efx, enum reset_type method)
2587 {
2588 	int rc, rc2;
2589 	bool disabled;
2590 
2591 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2592 		   RESET_TYPE(method));
2593 
2594 	efx_device_detach_sync(efx);
2595 	efx_reset_down(efx, method);
2596 
2597 	rc = efx->type->reset(efx, method);
2598 	if (rc) {
2599 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2600 		goto out;
2601 	}
2602 
2603 	/* Clear flags for the scopes we covered.  We assume the NIC and
2604 	 * driver are now quiescent so that there is no race here.
2605 	 */
2606 	if (method < RESET_TYPE_MAX_METHOD)
2607 		efx->reset_pending &= -(1 << (method + 1));
2608 	else /* it doesn't fit into the well-ordered scope hierarchy */
2609 		__clear_bit(method, &efx->reset_pending);
2610 
2611 	/* Reinitialise bus-mastering, which may have been turned off before
2612 	 * the reset was scheduled. This is still appropriate, even in the
2613 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2614 	 * can respond to requests. */
2615 	pci_set_master(efx->pci_dev);
2616 
2617 out:
2618 	/* Leave device stopped if necessary */
2619 	disabled = rc ||
2620 		method == RESET_TYPE_DISABLE ||
2621 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2622 	rc2 = efx_reset_up(efx, method, !disabled);
2623 	if (rc2) {
2624 		disabled = true;
2625 		if (!rc)
2626 			rc = rc2;
2627 	}
2628 
2629 	if (disabled) {
2630 		dev_close(efx->net_dev);
2631 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2632 		efx->state = STATE_DISABLED;
2633 	} else {
2634 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2635 		netif_device_attach(efx->net_dev);
2636 	}
2637 	return rc;
2638 }
2639 
2640 /* Try recovery mechanisms.
2641  * For now only EEH is supported.
2642  * Returns 0 if the recovery mechanisms are unsuccessful.
2643  * Returns a non-zero value otherwise.
2644  */
2645 int efx_try_recovery(struct efx_nic *efx)
2646 {
2647 #ifdef CONFIG_EEH
2648 	/* A PCI error can occur and not be seen by EEH because nothing
2649 	 * happens on the PCI bus. In this case the driver may fail and
2650 	 * schedule a 'recover or reset', leading to this recovery handler.
2651 	 * Manually call the eeh failure check function.
2652 	 */
2653 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2654 	if (eeh_dev_check_failure(eehdev)) {
2655 		/* The EEH mechanisms will handle the error and reset the
2656 		 * device if necessary.
2657 		 */
2658 		return 1;
2659 	}
2660 #endif
2661 	return 0;
2662 }
2663 
2664 static void efx_wait_for_bist_end(struct efx_nic *efx)
2665 {
2666 	int i;
2667 
2668 	for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2669 		if (efx_mcdi_poll_reboot(efx))
2670 			goto out;
2671 		msleep(BIST_WAIT_DELAY_MS);
2672 	}
2673 
2674 	netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2675 out:
2676 	/* Either way unset the BIST flag. If we found no reboot we probably
2677 	 * won't recover, but we should try.
2678 	 */
2679 	efx->mc_bist_for_other_fn = false;
2680 }
2681 
2682 /* The worker thread exists so that code that cannot sleep can
2683  * schedule a reset for later.
2684  */
2685 static void efx_reset_work(struct work_struct *data)
2686 {
2687 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2688 	unsigned long pending;
2689 	enum reset_type method;
2690 
2691 	pending = ACCESS_ONCE(efx->reset_pending);
2692 	method = fls(pending) - 1;
2693 
2694 	if (method == RESET_TYPE_MC_BIST)
2695 		efx_wait_for_bist_end(efx);
2696 
2697 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2698 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2699 	    efx_try_recovery(efx))
2700 		return;
2701 
2702 	if (!pending)
2703 		return;
2704 
2705 	rtnl_lock();
2706 
2707 	/* We checked the state in efx_schedule_reset() but it may
2708 	 * have changed by now.  Now that we have the RTNL lock,
2709 	 * it cannot change again.
2710 	 */
2711 	if (efx->state == STATE_READY)
2712 		(void)efx_reset(efx, method);
2713 
2714 	rtnl_unlock();
2715 }
2716 
2717 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2718 {
2719 	enum reset_type method;
2720 
2721 	if (efx->state == STATE_RECOVERY) {
2722 		netif_dbg(efx, drv, efx->net_dev,
2723 			  "recovering: skip scheduling %s reset\n",
2724 			  RESET_TYPE(type));
2725 		return;
2726 	}
2727 
2728 	switch (type) {
2729 	case RESET_TYPE_INVISIBLE:
2730 	case RESET_TYPE_ALL:
2731 	case RESET_TYPE_RECOVER_OR_ALL:
2732 	case RESET_TYPE_WORLD:
2733 	case RESET_TYPE_DISABLE:
2734 	case RESET_TYPE_RECOVER_OR_DISABLE:
2735 	case RESET_TYPE_DATAPATH:
2736 	case RESET_TYPE_MC_BIST:
2737 	case RESET_TYPE_MCDI_TIMEOUT:
2738 		method = type;
2739 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2740 			  RESET_TYPE(method));
2741 		break;
2742 	default:
2743 		method = efx->type->map_reset_reason(type);
2744 		netif_dbg(efx, drv, efx->net_dev,
2745 			  "scheduling %s reset for %s\n",
2746 			  RESET_TYPE(method), RESET_TYPE(type));
2747 		break;
2748 	}
2749 
2750 	set_bit(method, &efx->reset_pending);
2751 	smp_mb(); /* ensure we change reset_pending before checking state */
2752 
2753 	/* If we're not READY then just leave the flags set as the cue
2754 	 * to abort probing or reschedule the reset later.
2755 	 */
2756 	if (ACCESS_ONCE(efx->state) != STATE_READY)
2757 		return;
2758 
2759 	/* efx_process_channel() will no longer read events once a
2760 	 * reset is scheduled. So switch back to poll'd MCDI completions. */
2761 	efx_mcdi_mode_poll(efx);
2762 
2763 	queue_work(reset_workqueue, &efx->reset_work);
2764 }
2765 
2766 /**************************************************************************
2767  *
2768  * List of NICs we support
2769  *
2770  **************************************************************************/
2771 
2772 /* PCI device ID table */
2773 static const struct pci_device_id efx_pci_table[] = {
2774 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2775 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2776 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2777 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2778 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2779 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2780 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803),	/* SFC9020 */
2781 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2782 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813),	/* SFL9021 */
2783 	 .driver_data = (unsigned long) &siena_a0_nic_type},
2784 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903),  /* SFC9120 PF */
2785 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2786 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903),  /* SFC9120 VF */
2787 	 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2788 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923),  /* SFC9140 PF */
2789 	 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2790 	{0}			/* end of list */
2791 };
2792 
2793 /**************************************************************************
2794  *
2795  * Dummy PHY/MAC operations
2796  *
2797  * Can be used for some unimplemented operations
2798  * Needed so all function pointers are valid and do not have to be tested
2799  * before use
2800  *
2801  **************************************************************************/
2802 int efx_port_dummy_op_int(struct efx_nic *efx)
2803 {
2804 	return 0;
2805 }
2806 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2807 
2808 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2809 {
2810 	return false;
2811 }
2812 
2813 static const struct efx_phy_operations efx_dummy_phy_operations = {
2814 	.init		 = efx_port_dummy_op_int,
2815 	.reconfigure	 = efx_port_dummy_op_int,
2816 	.poll		 = efx_port_dummy_op_poll,
2817 	.fini		 = efx_port_dummy_op_void,
2818 };
2819 
2820 /**************************************************************************
2821  *
2822  * Data housekeeping
2823  *
2824  **************************************************************************/
2825 
2826 /* This zeroes out and then fills in the invariants in a struct
2827  * efx_nic (including all sub-structures).
2828  */
2829 static int efx_init_struct(struct efx_nic *efx,
2830 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2831 {
2832 	int i;
2833 
2834 	/* Initialise common structures */
2835 	INIT_LIST_HEAD(&efx->node);
2836 	INIT_LIST_HEAD(&efx->secondary_list);
2837 	spin_lock_init(&efx->biu_lock);
2838 #ifdef CONFIG_SFC_MTD
2839 	INIT_LIST_HEAD(&efx->mtd_list);
2840 #endif
2841 	INIT_WORK(&efx->reset_work, efx_reset_work);
2842 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2843 	INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2844 	efx->pci_dev = pci_dev;
2845 	efx->msg_enable = debug;
2846 	efx->state = STATE_UNINIT;
2847 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2848 
2849 	efx->net_dev = net_dev;
2850 	efx->rx_prefix_size = efx->type->rx_prefix_size;
2851 	efx->rx_ip_align =
2852 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2853 	efx->rx_packet_hash_offset =
2854 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2855 	efx->rx_packet_ts_offset =
2856 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2857 	spin_lock_init(&efx->stats_lock);
2858 	mutex_init(&efx->mac_lock);
2859 	efx->phy_op = &efx_dummy_phy_operations;
2860 	efx->mdio.dev = net_dev;
2861 	INIT_WORK(&efx->mac_work, efx_mac_work);
2862 	init_waitqueue_head(&efx->flush_wq);
2863 
2864 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2865 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2866 		if (!efx->channel[i])
2867 			goto fail;
2868 		efx->msi_context[i].efx = efx;
2869 		efx->msi_context[i].index = i;
2870 	}
2871 
2872 	/* Higher numbered interrupt modes are less capable! */
2873 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2874 				  interrupt_mode);
2875 
2876 	/* Would be good to use the net_dev name, but we're too early */
2877 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2878 		 pci_name(pci_dev));
2879 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2880 	if (!efx->workqueue)
2881 		goto fail;
2882 
2883 	return 0;
2884 
2885 fail:
2886 	efx_fini_struct(efx);
2887 	return -ENOMEM;
2888 }
2889 
2890 static void efx_fini_struct(struct efx_nic *efx)
2891 {
2892 	int i;
2893 
2894 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
2895 		kfree(efx->channel[i]);
2896 
2897 	kfree(efx->vpd_sn);
2898 
2899 	if (efx->workqueue) {
2900 		destroy_workqueue(efx->workqueue);
2901 		efx->workqueue = NULL;
2902 	}
2903 }
2904 
2905 void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
2906 {
2907 	u64 n_rx_nodesc_trunc = 0;
2908 	struct efx_channel *channel;
2909 
2910 	efx_for_each_channel(channel, efx)
2911 		n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
2912 	stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
2913 	stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
2914 }
2915 
2916 /**************************************************************************
2917  *
2918  * PCI interface
2919  *
2920  **************************************************************************/
2921 
2922 /* Main body of final NIC shutdown code
2923  * This is called only at module unload (or hotplug removal).
2924  */
2925 static void efx_pci_remove_main(struct efx_nic *efx)
2926 {
2927 	/* Flush reset_work. It can no longer be scheduled since we
2928 	 * are not READY.
2929 	 */
2930 	BUG_ON(efx->state == STATE_READY);
2931 	cancel_work_sync(&efx->reset_work);
2932 
2933 	efx_disable_interrupts(efx);
2934 	efx_nic_fini_interrupt(efx);
2935 	efx_fini_port(efx);
2936 	efx->type->fini(efx);
2937 	efx_fini_napi(efx);
2938 	efx_remove_all(efx);
2939 }
2940 
2941 /* Final NIC shutdown
2942  * This is called only at module unload (or hotplug removal).  A PF can call
2943  * this on its VFs to ensure they are unbound first.
2944  */
2945 static void efx_pci_remove(struct pci_dev *pci_dev)
2946 {
2947 	struct efx_nic *efx;
2948 
2949 	efx = pci_get_drvdata(pci_dev);
2950 	if (!efx)
2951 		return;
2952 
2953 	/* Mark the NIC as fini, then stop the interface */
2954 	rtnl_lock();
2955 	efx_dissociate(efx);
2956 	dev_close(efx->net_dev);
2957 	efx_disable_interrupts(efx);
2958 	efx->state = STATE_UNINIT;
2959 	rtnl_unlock();
2960 
2961 	if (efx->type->sriov_fini)
2962 		efx->type->sriov_fini(efx);
2963 
2964 	efx_unregister_netdev(efx);
2965 
2966 	efx_mtd_remove(efx);
2967 
2968 	efx_pci_remove_main(efx);
2969 
2970 	efx_fini_io(efx);
2971 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2972 
2973 	efx_fini_struct(efx);
2974 	free_netdev(efx->net_dev);
2975 
2976 	pci_disable_pcie_error_reporting(pci_dev);
2977 };
2978 
2979 /* NIC VPD information
2980  * Called during probe to display the part number of the
2981  * installed NIC.  VPD is potentially very large but this should
2982  * always appear within the first 512 bytes.
2983  */
2984 #define SFC_VPD_LEN 512
2985 static void efx_probe_vpd_strings(struct efx_nic *efx)
2986 {
2987 	struct pci_dev *dev = efx->pci_dev;
2988 	char vpd_data[SFC_VPD_LEN];
2989 	ssize_t vpd_size;
2990 	int ro_start, ro_size, i, j;
2991 
2992 	/* Get the vpd data from the device */
2993 	vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2994 	if (vpd_size <= 0) {
2995 		netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2996 		return;
2997 	}
2998 
2999 	/* Get the Read only section */
3000 	ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
3001 	if (ro_start < 0) {
3002 		netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
3003 		return;
3004 	}
3005 
3006 	ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
3007 	j = ro_size;
3008 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3009 	if (i + j > vpd_size)
3010 		j = vpd_size - i;
3011 
3012 	/* Get the Part number */
3013 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
3014 	if (i < 0) {
3015 		netif_err(efx, drv, efx->net_dev, "Part number not found\n");
3016 		return;
3017 	}
3018 
3019 	j = pci_vpd_info_field_size(&vpd_data[i]);
3020 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3021 	if (i + j > vpd_size) {
3022 		netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
3023 		return;
3024 	}
3025 
3026 	netif_info(efx, drv, efx->net_dev,
3027 		   "Part Number : %.*s\n", j, &vpd_data[i]);
3028 
3029 	i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3030 	j = ro_size;
3031 	i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
3032 	if (i < 0) {
3033 		netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
3034 		return;
3035 	}
3036 
3037 	j = pci_vpd_info_field_size(&vpd_data[i]);
3038 	i += PCI_VPD_INFO_FLD_HDR_SIZE;
3039 	if (i + j > vpd_size) {
3040 		netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
3041 		return;
3042 	}
3043 
3044 	efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
3045 	if (!efx->vpd_sn)
3046 		return;
3047 
3048 	snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
3049 }
3050 
3051 
3052 /* Main body of NIC initialisation
3053  * This is called at module load (or hotplug insertion, theoretically).
3054  */
3055 static int efx_pci_probe_main(struct efx_nic *efx)
3056 {
3057 	int rc;
3058 
3059 	/* Do start-of-day initialisation */
3060 	rc = efx_probe_all(efx);
3061 	if (rc)
3062 		goto fail1;
3063 
3064 	efx_init_napi(efx);
3065 
3066 	rc = efx->type->init(efx);
3067 	if (rc) {
3068 		netif_err(efx, probe, efx->net_dev,
3069 			  "failed to initialise NIC\n");
3070 		goto fail3;
3071 	}
3072 
3073 	rc = efx_init_port(efx);
3074 	if (rc) {
3075 		netif_err(efx, probe, efx->net_dev,
3076 			  "failed to initialise port\n");
3077 		goto fail4;
3078 	}
3079 
3080 	rc = efx_nic_init_interrupt(efx);
3081 	if (rc)
3082 		goto fail5;
3083 	rc = efx_enable_interrupts(efx);
3084 	if (rc)
3085 		goto fail6;
3086 
3087 	return 0;
3088 
3089  fail6:
3090 	efx_nic_fini_interrupt(efx);
3091  fail5:
3092 	efx_fini_port(efx);
3093  fail4:
3094 	efx->type->fini(efx);
3095  fail3:
3096 	efx_fini_napi(efx);
3097 	efx_remove_all(efx);
3098  fail1:
3099 	return rc;
3100 }
3101 
3102 /* NIC initialisation
3103  *
3104  * This is called at module load (or hotplug insertion,
3105  * theoretically).  It sets up PCI mappings, resets the NIC,
3106  * sets up and registers the network devices with the kernel and hooks
3107  * the interrupt service routine.  It does not prepare the device for
3108  * transmission; this is left to the first time one of the network
3109  * interfaces is brought up (i.e. efx_net_open).
3110  */
3111 static int efx_pci_probe(struct pci_dev *pci_dev,
3112 			 const struct pci_device_id *entry)
3113 {
3114 	struct net_device *net_dev;
3115 	struct efx_nic *efx;
3116 	int rc;
3117 
3118 	/* Allocate and initialise a struct net_device and struct efx_nic */
3119 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
3120 				     EFX_MAX_RX_QUEUES);
3121 	if (!net_dev)
3122 		return -ENOMEM;
3123 	efx = netdev_priv(net_dev);
3124 	efx->type = (const struct efx_nic_type *) entry->driver_data;
3125 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
3126 			      NETIF_F_HIGHDMA | NETIF_F_TSO |
3127 			      NETIF_F_RXCSUM);
3128 	if (efx->type->offload_features & NETIF_F_V6_CSUM)
3129 		net_dev->features |= NETIF_F_TSO6;
3130 	/* Mask for features that also apply to VLAN devices */
3131 	net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
3132 				   NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
3133 				   NETIF_F_RXCSUM);
3134 	/* All offloads can be toggled */
3135 	net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
3136 	pci_set_drvdata(pci_dev, efx);
3137 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
3138 	rc = efx_init_struct(efx, pci_dev, net_dev);
3139 	if (rc)
3140 		goto fail1;
3141 
3142 	netif_info(efx, probe, efx->net_dev,
3143 		   "Solarflare NIC detected\n");
3144 
3145 	if (!efx->type->is_vf)
3146 		efx_probe_vpd_strings(efx);
3147 
3148 	/* Set up basic I/O (BAR mappings etc) */
3149 	rc = efx_init_io(efx);
3150 	if (rc)
3151 		goto fail2;
3152 
3153 	rc = efx_pci_probe_main(efx);
3154 	if (rc)
3155 		goto fail3;
3156 
3157 	rc = efx_register_netdev(efx);
3158 	if (rc)
3159 		goto fail4;
3160 
3161 	if (efx->type->sriov_init) {
3162 		rc = efx->type->sriov_init(efx);
3163 		if (rc)
3164 			netif_err(efx, probe, efx->net_dev,
3165 				  "SR-IOV can't be enabled rc %d\n", rc);
3166 	}
3167 
3168 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3169 
3170 	/* Try to create MTDs, but allow this to fail */
3171 	rtnl_lock();
3172 	rc = efx_mtd_probe(efx);
3173 	rtnl_unlock();
3174 	if (rc)
3175 		netif_warn(efx, probe, efx->net_dev,
3176 			   "failed to create MTDs (%d)\n", rc);
3177 
3178 	rc = pci_enable_pcie_error_reporting(pci_dev);
3179 	if (rc && rc != -EINVAL)
3180 		netif_warn(efx, probe, efx->net_dev,
3181 			   "pci_enable_pcie_error_reporting failed (%d)\n", rc);
3182 
3183 	return 0;
3184 
3185  fail4:
3186 	efx_pci_remove_main(efx);
3187  fail3:
3188 	efx_fini_io(efx);
3189  fail2:
3190 	efx_fini_struct(efx);
3191  fail1:
3192 	WARN_ON(rc > 0);
3193 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3194 	free_netdev(net_dev);
3195 	return rc;
3196 }
3197 
3198 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3199  * enabled on success
3200  */
3201 #ifdef CONFIG_SFC_SRIOV
3202 static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
3203 {
3204 	int rc;
3205 	struct efx_nic *efx = pci_get_drvdata(dev);
3206 
3207 	if (efx->type->sriov_configure) {
3208 		rc = efx->type->sriov_configure(efx, num_vfs);
3209 		if (rc)
3210 			return rc;
3211 		else
3212 			return num_vfs;
3213 	} else
3214 		return -EOPNOTSUPP;
3215 }
3216 #endif
3217 
3218 static int efx_pm_freeze(struct device *dev)
3219 {
3220 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3221 
3222 	rtnl_lock();
3223 
3224 	if (efx->state != STATE_DISABLED) {
3225 		efx->state = STATE_UNINIT;
3226 
3227 		efx_device_detach_sync(efx);
3228 
3229 		efx_stop_all(efx);
3230 		efx_disable_interrupts(efx);
3231 	}
3232 
3233 	rtnl_unlock();
3234 
3235 	return 0;
3236 }
3237 
3238 static int efx_pm_thaw(struct device *dev)
3239 {
3240 	int rc;
3241 	struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3242 
3243 	rtnl_lock();
3244 
3245 	if (efx->state != STATE_DISABLED) {
3246 		rc = efx_enable_interrupts(efx);
3247 		if (rc)
3248 			goto fail;
3249 
3250 		mutex_lock(&efx->mac_lock);
3251 		efx->phy_op->reconfigure(efx);
3252 		mutex_unlock(&efx->mac_lock);
3253 
3254 		efx_start_all(efx);
3255 
3256 		netif_device_attach(efx->net_dev);
3257 
3258 		efx->state = STATE_READY;
3259 
3260 		efx->type->resume_wol(efx);
3261 	}
3262 
3263 	rtnl_unlock();
3264 
3265 	/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3266 	queue_work(reset_workqueue, &efx->reset_work);
3267 
3268 	return 0;
3269 
3270 fail:
3271 	rtnl_unlock();
3272 
3273 	return rc;
3274 }
3275 
3276 static int efx_pm_poweroff(struct device *dev)
3277 {
3278 	struct pci_dev *pci_dev = to_pci_dev(dev);
3279 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3280 
3281 	efx->type->fini(efx);
3282 
3283 	efx->reset_pending = 0;
3284 
3285 	pci_save_state(pci_dev);
3286 	return pci_set_power_state(pci_dev, PCI_D3hot);
3287 }
3288 
3289 /* Used for both resume and restore */
3290 static int efx_pm_resume(struct device *dev)
3291 {
3292 	struct pci_dev *pci_dev = to_pci_dev(dev);
3293 	struct efx_nic *efx = pci_get_drvdata(pci_dev);
3294 	int rc;
3295 
3296 	rc = pci_set_power_state(pci_dev, PCI_D0);
3297 	if (rc)
3298 		return rc;
3299 	pci_restore_state(pci_dev);
3300 	rc = pci_enable_device(pci_dev);
3301 	if (rc)
3302 		return rc;
3303 	pci_set_master(efx->pci_dev);
3304 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
3305 	if (rc)
3306 		return rc;
3307 	rc = efx->type->init(efx);
3308 	if (rc)
3309 		return rc;
3310 	rc = efx_pm_thaw(dev);
3311 	return rc;
3312 }
3313 
3314 static int efx_pm_suspend(struct device *dev)
3315 {
3316 	int rc;
3317 
3318 	efx_pm_freeze(dev);
3319 	rc = efx_pm_poweroff(dev);
3320 	if (rc)
3321 		efx_pm_resume(dev);
3322 	return rc;
3323 }
3324 
3325 static const struct dev_pm_ops efx_pm_ops = {
3326 	.suspend	= efx_pm_suspend,
3327 	.resume		= efx_pm_resume,
3328 	.freeze		= efx_pm_freeze,
3329 	.thaw		= efx_pm_thaw,
3330 	.poweroff	= efx_pm_poweroff,
3331 	.restore	= efx_pm_resume,
3332 };
3333 
3334 /* A PCI error affecting this device was detected.
3335  * At this point MMIO and DMA may be disabled.
3336  * Stop the software path and request a slot reset.
3337  */
3338 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3339 					      enum pci_channel_state state)
3340 {
3341 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3342 	struct efx_nic *efx = pci_get_drvdata(pdev);
3343 
3344 	if (state == pci_channel_io_perm_failure)
3345 		return PCI_ERS_RESULT_DISCONNECT;
3346 
3347 	rtnl_lock();
3348 
3349 	if (efx->state != STATE_DISABLED) {
3350 		efx->state = STATE_RECOVERY;
3351 		efx->reset_pending = 0;
3352 
3353 		efx_device_detach_sync(efx);
3354 
3355 		efx_stop_all(efx);
3356 		efx_disable_interrupts(efx);
3357 
3358 		status = PCI_ERS_RESULT_NEED_RESET;
3359 	} else {
3360 		/* If the interface is disabled we don't want to do anything
3361 		 * with it.
3362 		 */
3363 		status = PCI_ERS_RESULT_RECOVERED;
3364 	}
3365 
3366 	rtnl_unlock();
3367 
3368 	pci_disable_device(pdev);
3369 
3370 	return status;
3371 }
3372 
3373 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3374 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3375 {
3376 	struct efx_nic *efx = pci_get_drvdata(pdev);
3377 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3378 	int rc;
3379 
3380 	if (pci_enable_device(pdev)) {
3381 		netif_err(efx, hw, efx->net_dev,
3382 			  "Cannot re-enable PCI device after reset.\n");
3383 		status =  PCI_ERS_RESULT_DISCONNECT;
3384 	}
3385 
3386 	rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3387 	if (rc) {
3388 		netif_err(efx, hw, efx->net_dev,
3389 		"pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3390 		/* Non-fatal error. Continue. */
3391 	}
3392 
3393 	return status;
3394 }
3395 
3396 /* Perform the actual reset and resume I/O operations. */
3397 static void efx_io_resume(struct pci_dev *pdev)
3398 {
3399 	struct efx_nic *efx = pci_get_drvdata(pdev);
3400 	int rc;
3401 
3402 	rtnl_lock();
3403 
3404 	if (efx->state == STATE_DISABLED)
3405 		goto out;
3406 
3407 	rc = efx_reset(efx, RESET_TYPE_ALL);
3408 	if (rc) {
3409 		netif_err(efx, hw, efx->net_dev,
3410 			  "efx_reset failed after PCI error (%d)\n", rc);
3411 	} else {
3412 		efx->state = STATE_READY;
3413 		netif_dbg(efx, hw, efx->net_dev,
3414 			  "Done resetting and resuming IO after PCI error.\n");
3415 	}
3416 
3417 out:
3418 	rtnl_unlock();
3419 }
3420 
3421 /* For simplicity and reliability, we always require a slot reset and try to
3422  * reset the hardware when a pci error affecting the device is detected.
3423  * We leave both the link_reset and mmio_enabled callback unimplemented:
3424  * with our request for slot reset the mmio_enabled callback will never be
3425  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3426  */
3427 static struct pci_error_handlers efx_err_handlers = {
3428 	.error_detected = efx_io_error_detected,
3429 	.slot_reset	= efx_io_slot_reset,
3430 	.resume		= efx_io_resume,
3431 };
3432 
3433 static struct pci_driver efx_pci_driver = {
3434 	.name		= KBUILD_MODNAME,
3435 	.id_table	= efx_pci_table,
3436 	.probe		= efx_pci_probe,
3437 	.remove		= efx_pci_remove,
3438 	.driver.pm	= &efx_pm_ops,
3439 	.err_handler	= &efx_err_handlers,
3440 #ifdef CONFIG_SFC_SRIOV
3441 	.sriov_configure = efx_pci_sriov_configure,
3442 #endif
3443 };
3444 
3445 /**************************************************************************
3446  *
3447  * Kernel module interface
3448  *
3449  *************************************************************************/
3450 
3451 module_param(interrupt_mode, uint, 0444);
3452 MODULE_PARM_DESC(interrupt_mode,
3453 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3454 
3455 static int __init efx_init_module(void)
3456 {
3457 	int rc;
3458 
3459 	printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3460 
3461 	rc = register_netdevice_notifier(&efx_netdev_notifier);
3462 	if (rc)
3463 		goto err_notifier;
3464 
3465 #ifdef CONFIG_SFC_SRIOV
3466 	rc = efx_init_sriov();
3467 	if (rc)
3468 		goto err_sriov;
3469 #endif
3470 
3471 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3472 	if (!reset_workqueue) {
3473 		rc = -ENOMEM;
3474 		goto err_reset;
3475 	}
3476 
3477 	rc = pci_register_driver(&efx_pci_driver);
3478 	if (rc < 0)
3479 		goto err_pci;
3480 
3481 	return 0;
3482 
3483  err_pci:
3484 	destroy_workqueue(reset_workqueue);
3485  err_reset:
3486 #ifdef CONFIG_SFC_SRIOV
3487 	efx_fini_sriov();
3488  err_sriov:
3489 #endif
3490 	unregister_netdevice_notifier(&efx_netdev_notifier);
3491  err_notifier:
3492 	return rc;
3493 }
3494 
3495 static void __exit efx_exit_module(void)
3496 {
3497 	printk(KERN_INFO "Solarflare NET driver unloading\n");
3498 
3499 	pci_unregister_driver(&efx_pci_driver);
3500 	destroy_workqueue(reset_workqueue);
3501 #ifdef CONFIG_SFC_SRIOV
3502 	efx_fini_sriov();
3503 #endif
3504 	unregister_netdevice_notifier(&efx_netdev_notifier);
3505 
3506 }
3507 
3508 module_init(efx_init_module);
3509 module_exit(efx_exit_module);
3510 
3511 MODULE_AUTHOR("Solarflare Communications and "
3512 	      "Michael Brown <mbrown@fensystems.co.uk>");
3513 MODULE_DESCRIPTION("Solarflare network driver");
3514 MODULE_LICENSE("GPL");
3515 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3516