xref: /linux/drivers/net/ethernet/sfc/falcon/efx.c (revision 52990390f91c1c39ca742fc8f390b29891d95127)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3  * Driver for Solarflare network controllers and boards
4  * Copyright 2005-2006 Fen Systems Ltd.
5  * Copyright 2005-2013 Solarflare Communications Inc.
6  */
7 
8 #include <linux/module.h>
9 #include <linux/pci.h>
10 #include <linux/netdevice.h>
11 #include <linux/etherdevice.h>
12 #include <linux/delay.h>
13 #include <linux/notifier.h>
14 #include <linux/ip.h>
15 #include <linux/tcp.h>
16 #include <linux/in.h>
17 #include <linux/ethtool.h>
18 #include <linux/topology.h>
19 #include <linux/gfp.h>
20 #include <linux/interrupt.h>
21 #include "net_driver.h"
22 #include "efx.h"
23 #include "nic.h"
24 #include "selftest.h"
25 
26 #include "workarounds.h"
27 
28 /**************************************************************************
29  *
30  * Type name strings
31  *
32  **************************************************************************
33  */
34 
35 /* Loopback mode names (see LOOPBACK_MODE()) */
36 const unsigned int ef4_loopback_mode_max = LOOPBACK_MAX;
37 const char *const ef4_loopback_mode_names[] = {
38 	[LOOPBACK_NONE]		= "NONE",
39 	[LOOPBACK_DATA]		= "DATAPATH",
40 	[LOOPBACK_GMAC]		= "GMAC",
41 	[LOOPBACK_XGMII]	= "XGMII",
42 	[LOOPBACK_XGXS]		= "XGXS",
43 	[LOOPBACK_XAUI]		= "XAUI",
44 	[LOOPBACK_GMII]		= "GMII",
45 	[LOOPBACK_SGMII]	= "SGMII",
46 	[LOOPBACK_XGBR]		= "XGBR",
47 	[LOOPBACK_XFI]		= "XFI",
48 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
49 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
50 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
51 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
52 	[LOOPBACK_GPHY]		= "GPHY",
53 	[LOOPBACK_PHYXS]	= "PHYXS",
54 	[LOOPBACK_PCS]		= "PCS",
55 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
56 	[LOOPBACK_XPORT]	= "XPORT",
57 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
58 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
59 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
60 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
61 	[LOOPBACK_GMII_WS]	= "GMII_WS",
62 	[LOOPBACK_XFI_WS]	= "XFI_WS",
63 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
64 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
65 };
66 
67 const unsigned int ef4_reset_type_max = RESET_TYPE_MAX;
68 const char *const ef4_reset_type_names[] = {
69 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
70 	[RESET_TYPE_ALL]                = "ALL",
71 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
72 	[RESET_TYPE_WORLD]              = "WORLD",
73 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
74 	[RESET_TYPE_DATAPATH]           = "DATAPATH",
75 	[RESET_TYPE_DISABLE]            = "DISABLE",
76 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
77 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
78 	[RESET_TYPE_RX_RECOVERY]        = "RX_RECOVERY",
79 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
80 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
81 };
82 
83 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
84  * queued onto this work queue. This is not a per-nic work queue, because
85  * ef4_reset_work() acquires the rtnl lock, so resets are naturally serialised.
86  */
87 static struct workqueue_struct *reset_workqueue;
88 
89 /* How often and how many times to poll for a reset while waiting for a
90  * BIST that another function started to complete.
91  */
92 #define BIST_WAIT_DELAY_MS	100
93 #define BIST_WAIT_DELAY_COUNT	100
94 
95 /**************************************************************************
96  *
97  * Configurable values
98  *
99  *************************************************************************/
100 
101 /*
102  * Use separate channels for TX and RX events
103  *
104  * Set this to 1 to use separate channels for TX and RX. It allows us
105  * to control interrupt affinity separately for TX and RX.
106  *
107  * This is only used in MSI-X interrupt mode
108  */
109 bool ef4_separate_tx_channels;
110 module_param(ef4_separate_tx_channels, bool, 0444);
111 MODULE_PARM_DESC(ef4_separate_tx_channels,
112 		 "Use separate channels for TX and RX");
113 
114 /* This is the time (in jiffies) between invocations of the hardware
115  * monitor.
116  * On Falcon-based NICs, this will:
117  * - Check the on-board hardware monitor;
118  * - Poll the link state and reconfigure the hardware as necessary.
119  * On Siena-based NICs for power systems with EEH support, this will give EEH a
120  * chance to start.
121  */
122 static unsigned int ef4_monitor_interval = 1 * HZ;
123 
124 /* Initial interrupt moderation settings.  They can be modified after
125  * module load with ethtool.
126  *
127  * The default for RX should strike a balance between increasing the
128  * round-trip latency and reducing overhead.
129  */
130 static unsigned int rx_irq_mod_usec = 60;
131 
132 /* Initial interrupt moderation settings.  They can be modified after
133  * module load with ethtool.
134  *
135  * This default is chosen to ensure that a 10G link does not go idle
136  * while a TX queue is stopped after it has become full.  A queue is
137  * restarted when it drops below half full.  The time this takes (assuming
138  * worst case 3 descriptors per packet and 1024 descriptors) is
139  *   512 / 3 * 1.2 = 205 usec.
140  */
141 static unsigned int tx_irq_mod_usec = 150;
142 
143 /* This is the first interrupt mode to try out of:
144  * 0 => MSI-X
145  * 1 => MSI
146  * 2 => legacy
147  */
148 static unsigned int interrupt_mode;
149 
150 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
151  * i.e. the number of CPUs among which we may distribute simultaneous
152  * interrupt handling.
153  *
154  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
155  * The default (0) means to assign an interrupt to each core.
156  */
157 static unsigned int rss_cpus;
158 module_param(rss_cpus, uint, 0444);
159 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
160 
161 static bool phy_flash_cfg;
162 module_param(phy_flash_cfg, bool, 0644);
163 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
164 
165 static unsigned irq_adapt_low_thresh = 8000;
166 module_param(irq_adapt_low_thresh, uint, 0644);
167 MODULE_PARM_DESC(irq_adapt_low_thresh,
168 		 "Threshold score for reducing IRQ moderation");
169 
170 static unsigned irq_adapt_high_thresh = 16000;
171 module_param(irq_adapt_high_thresh, uint, 0644);
172 MODULE_PARM_DESC(irq_adapt_high_thresh,
173 		 "Threshold score for increasing IRQ moderation");
174 
175 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
176 			 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
177 			 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
178 			 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
179 module_param(debug, uint, 0);
180 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
181 
182 /**************************************************************************
183  *
184  * Utility functions and prototypes
185  *
186  *************************************************************************/
187 
188 static int ef4_soft_enable_interrupts(struct ef4_nic *efx);
189 static void ef4_soft_disable_interrupts(struct ef4_nic *efx);
190 static void ef4_remove_channel(struct ef4_channel *channel);
191 static void ef4_remove_channels(struct ef4_nic *efx);
192 static const struct ef4_channel_type ef4_default_channel_type;
193 static void ef4_remove_port(struct ef4_nic *efx);
194 static void ef4_init_napi_channel(struct ef4_channel *channel);
195 static void ef4_fini_napi(struct ef4_nic *efx);
196 static void ef4_fini_napi_channel(struct ef4_channel *channel);
197 static void ef4_fini_struct(struct ef4_nic *efx);
198 static void ef4_start_all(struct ef4_nic *efx);
199 static void ef4_stop_all(struct ef4_nic *efx);
200 
201 #define EF4_ASSERT_RESET_SERIALISED(efx)		\
202 	do {						\
203 		if ((efx->state == STATE_READY) ||	\
204 		    (efx->state == STATE_RECOVERY) ||	\
205 		    (efx->state == STATE_DISABLED))	\
206 			ASSERT_RTNL();			\
207 	} while (0)
208 
209 static int ef4_check_disabled(struct ef4_nic *efx)
210 {
211 	if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
212 		netif_err(efx, drv, efx->net_dev,
213 			  "device is disabled due to earlier errors\n");
214 		return -EIO;
215 	}
216 	return 0;
217 }
218 
219 /**************************************************************************
220  *
221  * Event queue processing
222  *
223  *************************************************************************/
224 
225 /* Process channel's event queue
226  *
227  * This function is responsible for processing the event queue of a
228  * single channel.  The caller must guarantee that this function will
229  * never be concurrently called more than once on the same channel,
230  * though different channels may be being processed concurrently.
231  */
232 static int ef4_process_channel(struct ef4_channel *channel, int budget)
233 {
234 	struct ef4_tx_queue *tx_queue;
235 	int spent;
236 
237 	if (unlikely(!channel->enabled))
238 		return 0;
239 
240 	ef4_for_each_channel_tx_queue(tx_queue, channel) {
241 		tx_queue->pkts_compl = 0;
242 		tx_queue->bytes_compl = 0;
243 	}
244 
245 	spent = ef4_nic_process_eventq(channel, budget);
246 	if (spent && ef4_channel_has_rx_queue(channel)) {
247 		struct ef4_rx_queue *rx_queue =
248 			ef4_channel_get_rx_queue(channel);
249 
250 		ef4_rx_flush_packet(channel);
251 		ef4_fast_push_rx_descriptors(rx_queue, true);
252 	}
253 
254 	/* Update BQL */
255 	ef4_for_each_channel_tx_queue(tx_queue, channel) {
256 		if (tx_queue->bytes_compl) {
257 			netdev_tx_completed_queue(tx_queue->core_txq,
258 				tx_queue->pkts_compl, tx_queue->bytes_compl);
259 		}
260 	}
261 
262 	return spent;
263 }
264 
265 /* NAPI poll handler
266  *
267  * NAPI guarantees serialisation of polls of the same device, which
268  * provides the guarantee required by ef4_process_channel().
269  */
270 static void ef4_update_irq_mod(struct ef4_nic *efx, struct ef4_channel *channel)
271 {
272 	int step = efx->irq_mod_step_us;
273 
274 	if (channel->irq_mod_score < irq_adapt_low_thresh) {
275 		if (channel->irq_moderation_us > step) {
276 			channel->irq_moderation_us -= step;
277 			efx->type->push_irq_moderation(channel);
278 		}
279 	} else if (channel->irq_mod_score > irq_adapt_high_thresh) {
280 		if (channel->irq_moderation_us <
281 		    efx->irq_rx_moderation_us) {
282 			channel->irq_moderation_us += step;
283 			efx->type->push_irq_moderation(channel);
284 		}
285 	}
286 
287 	channel->irq_count = 0;
288 	channel->irq_mod_score = 0;
289 }
290 
291 static int ef4_poll(struct napi_struct *napi, int budget)
292 {
293 	struct ef4_channel *channel =
294 		container_of(napi, struct ef4_channel, napi_str);
295 	struct ef4_nic *efx = channel->efx;
296 	int spent;
297 
298 	netif_vdbg(efx, intr, efx->net_dev,
299 		   "channel %d NAPI poll executing on CPU %d\n",
300 		   channel->channel, raw_smp_processor_id());
301 
302 	spent = ef4_process_channel(channel, budget);
303 
304 	if (spent < budget) {
305 		if (ef4_channel_has_rx_queue(channel) &&
306 		    efx->irq_rx_adaptive &&
307 		    unlikely(++channel->irq_count == 1000)) {
308 			ef4_update_irq_mod(efx, channel);
309 		}
310 
311 		ef4_filter_rfs_expire(channel);
312 
313 		/* There is no race here; although napi_disable() will
314 		 * only wait for napi_complete(), this isn't a problem
315 		 * since ef4_nic_eventq_read_ack() will have no effect if
316 		 * interrupts have already been disabled.
317 		 */
318 		napi_complete_done(napi, spent);
319 		ef4_nic_eventq_read_ack(channel);
320 	}
321 
322 	return spent;
323 }
324 
325 /* Create event queue
326  * Event queue memory allocations are done only once.  If the channel
327  * is reset, the memory buffer will be reused; this guards against
328  * errors during channel reset and also simplifies interrupt handling.
329  */
330 static int ef4_probe_eventq(struct ef4_channel *channel)
331 {
332 	struct ef4_nic *efx = channel->efx;
333 	unsigned long entries;
334 
335 	netif_dbg(efx, probe, efx->net_dev,
336 		  "chan %d create event queue\n", channel->channel);
337 
338 	/* Build an event queue with room for one event per tx and rx buffer,
339 	 * plus some extra for link state events and MCDI completions. */
340 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
341 	EF4_BUG_ON_PARANOID(entries > EF4_MAX_EVQ_SIZE);
342 	channel->eventq_mask = max(entries, EF4_MIN_EVQ_SIZE) - 1;
343 
344 	return ef4_nic_probe_eventq(channel);
345 }
346 
347 /* Prepare channel's event queue */
348 static int ef4_init_eventq(struct ef4_channel *channel)
349 {
350 	struct ef4_nic *efx = channel->efx;
351 	int rc;
352 
353 	EF4_WARN_ON_PARANOID(channel->eventq_init);
354 
355 	netif_dbg(efx, drv, efx->net_dev,
356 		  "chan %d init event queue\n", channel->channel);
357 
358 	rc = ef4_nic_init_eventq(channel);
359 	if (rc == 0) {
360 		efx->type->push_irq_moderation(channel);
361 		channel->eventq_read_ptr = 0;
362 		channel->eventq_init = true;
363 	}
364 	return rc;
365 }
366 
367 /* Enable event queue processing and NAPI */
368 void ef4_start_eventq(struct ef4_channel *channel)
369 {
370 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
371 		  "chan %d start event queue\n", channel->channel);
372 
373 	/* Make sure the NAPI handler sees the enabled flag set */
374 	channel->enabled = true;
375 	smp_wmb();
376 
377 	napi_enable(&channel->napi_str);
378 	ef4_nic_eventq_read_ack(channel);
379 }
380 
381 /* Disable event queue processing and NAPI */
382 void ef4_stop_eventq(struct ef4_channel *channel)
383 {
384 	if (!channel->enabled)
385 		return;
386 
387 	napi_disable(&channel->napi_str);
388 	channel->enabled = false;
389 }
390 
391 static void ef4_fini_eventq(struct ef4_channel *channel)
392 {
393 	if (!channel->eventq_init)
394 		return;
395 
396 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
397 		  "chan %d fini event queue\n", channel->channel);
398 
399 	ef4_nic_fini_eventq(channel);
400 	channel->eventq_init = false;
401 }
402 
403 static void ef4_remove_eventq(struct ef4_channel *channel)
404 {
405 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
406 		  "chan %d remove event queue\n", channel->channel);
407 
408 	ef4_nic_remove_eventq(channel);
409 }
410 
411 /**************************************************************************
412  *
413  * Channel handling
414  *
415  *************************************************************************/
416 
417 /* Allocate and initialise a channel structure. */
418 static struct ef4_channel *
419 ef4_alloc_channel(struct ef4_nic *efx, int i, struct ef4_channel *old_channel)
420 {
421 	struct ef4_channel *channel;
422 	struct ef4_rx_queue *rx_queue;
423 	struct ef4_tx_queue *tx_queue;
424 	int j;
425 
426 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
427 	if (!channel)
428 		return NULL;
429 
430 	channel->efx = efx;
431 	channel->channel = i;
432 	channel->type = &ef4_default_channel_type;
433 
434 	for (j = 0; j < EF4_TXQ_TYPES; j++) {
435 		tx_queue = &channel->tx_queue[j];
436 		tx_queue->efx = efx;
437 		tx_queue->queue = i * EF4_TXQ_TYPES + j;
438 		tx_queue->channel = channel;
439 	}
440 
441 	rx_queue = &channel->rx_queue;
442 	rx_queue->efx = efx;
443 	timer_setup(&rx_queue->slow_fill, ef4_rx_slow_fill, 0);
444 
445 	return channel;
446 }
447 
448 /* Allocate and initialise a channel structure, copying parameters
449  * (but not resources) from an old channel structure.
450  */
451 static struct ef4_channel *
452 ef4_copy_channel(const struct ef4_channel *old_channel)
453 {
454 	struct ef4_channel *channel;
455 	struct ef4_rx_queue *rx_queue;
456 	struct ef4_tx_queue *tx_queue;
457 	int j;
458 
459 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
460 	if (!channel)
461 		return NULL;
462 
463 	*channel = *old_channel;
464 
465 	channel->napi_dev = NULL;
466 	INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
467 	channel->napi_str.napi_id = 0;
468 	channel->napi_str.state = 0;
469 	memset(&channel->eventq, 0, sizeof(channel->eventq));
470 
471 	for (j = 0; j < EF4_TXQ_TYPES; j++) {
472 		tx_queue = &channel->tx_queue[j];
473 		if (tx_queue->channel)
474 			tx_queue->channel = channel;
475 		tx_queue->buffer = NULL;
476 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
477 	}
478 
479 	rx_queue = &channel->rx_queue;
480 	rx_queue->buffer = NULL;
481 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
482 	timer_setup(&rx_queue->slow_fill, ef4_rx_slow_fill, 0);
483 
484 	return channel;
485 }
486 
487 static int ef4_probe_channel(struct ef4_channel *channel)
488 {
489 	struct ef4_tx_queue *tx_queue;
490 	struct ef4_rx_queue *rx_queue;
491 	int rc;
492 
493 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
494 		  "creating channel %d\n", channel->channel);
495 
496 	rc = channel->type->pre_probe(channel);
497 	if (rc)
498 		goto fail;
499 
500 	rc = ef4_probe_eventq(channel);
501 	if (rc)
502 		goto fail;
503 
504 	ef4_for_each_channel_tx_queue(tx_queue, channel) {
505 		rc = ef4_probe_tx_queue(tx_queue);
506 		if (rc)
507 			goto fail;
508 	}
509 
510 	ef4_for_each_channel_rx_queue(rx_queue, channel) {
511 		rc = ef4_probe_rx_queue(rx_queue);
512 		if (rc)
513 			goto fail;
514 	}
515 
516 	return 0;
517 
518 fail:
519 	ef4_remove_channel(channel);
520 	return rc;
521 }
522 
523 static void
524 ef4_get_channel_name(struct ef4_channel *channel, char *buf, size_t len)
525 {
526 	struct ef4_nic *efx = channel->efx;
527 	const char *type;
528 	int number;
529 
530 	number = channel->channel;
531 	if (efx->tx_channel_offset == 0) {
532 		type = "";
533 	} else if (channel->channel < efx->tx_channel_offset) {
534 		type = "-rx";
535 	} else {
536 		type = "-tx";
537 		number -= efx->tx_channel_offset;
538 	}
539 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
540 }
541 
542 static void ef4_set_channel_names(struct ef4_nic *efx)
543 {
544 	struct ef4_channel *channel;
545 
546 	ef4_for_each_channel(channel, efx)
547 		channel->type->get_name(channel,
548 					efx->msi_context[channel->channel].name,
549 					sizeof(efx->msi_context[0].name));
550 }
551 
552 static int ef4_probe_channels(struct ef4_nic *efx)
553 {
554 	struct ef4_channel *channel;
555 	int rc;
556 
557 	/* Restart special buffer allocation */
558 	efx->next_buffer_table = 0;
559 
560 	/* Probe channels in reverse, so that any 'extra' channels
561 	 * use the start of the buffer table. This allows the traffic
562 	 * channels to be resized without moving them or wasting the
563 	 * entries before them.
564 	 */
565 	ef4_for_each_channel_rev(channel, efx) {
566 		rc = ef4_probe_channel(channel);
567 		if (rc) {
568 			netif_err(efx, probe, efx->net_dev,
569 				  "failed to create channel %d\n",
570 				  channel->channel);
571 			goto fail;
572 		}
573 	}
574 	ef4_set_channel_names(efx);
575 
576 	return 0;
577 
578 fail:
579 	ef4_remove_channels(efx);
580 	return rc;
581 }
582 
583 /* Channels are shutdown and reinitialised whilst the NIC is running
584  * to propagate configuration changes (mtu, checksum offload), or
585  * to clear hardware error conditions
586  */
587 static void ef4_start_datapath(struct ef4_nic *efx)
588 {
589 	netdev_features_t old_features = efx->net_dev->features;
590 	bool old_rx_scatter = efx->rx_scatter;
591 	struct ef4_tx_queue *tx_queue;
592 	struct ef4_rx_queue *rx_queue;
593 	struct ef4_channel *channel;
594 	size_t rx_buf_len;
595 
596 	/* Calculate the rx buffer allocation parameters required to
597 	 * support the current MTU, including padding for header
598 	 * alignment and overruns.
599 	 */
600 	efx->rx_dma_len = (efx->rx_prefix_size +
601 			   EF4_MAX_FRAME_LEN(efx->net_dev->mtu) +
602 			   efx->type->rx_buffer_padding);
603 	rx_buf_len = (sizeof(struct ef4_rx_page_state) +
604 		      efx->rx_ip_align + efx->rx_dma_len);
605 	if (rx_buf_len <= PAGE_SIZE) {
606 		efx->rx_scatter = efx->type->always_rx_scatter;
607 		efx->rx_buffer_order = 0;
608 	} else if (efx->type->can_rx_scatter) {
609 		BUILD_BUG_ON(EF4_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
610 		BUILD_BUG_ON(sizeof(struct ef4_rx_page_state) +
611 			     2 * ALIGN(NET_IP_ALIGN + EF4_RX_USR_BUF_SIZE,
612 				       EF4_RX_BUF_ALIGNMENT) >
613 			     PAGE_SIZE);
614 		efx->rx_scatter = true;
615 		efx->rx_dma_len = EF4_RX_USR_BUF_SIZE;
616 		efx->rx_buffer_order = 0;
617 	} else {
618 		efx->rx_scatter = false;
619 		efx->rx_buffer_order = get_order(rx_buf_len);
620 	}
621 
622 	ef4_rx_config_page_split(efx);
623 	if (efx->rx_buffer_order)
624 		netif_dbg(efx, drv, efx->net_dev,
625 			  "RX buf len=%u; page order=%u batch=%u\n",
626 			  efx->rx_dma_len, efx->rx_buffer_order,
627 			  efx->rx_pages_per_batch);
628 	else
629 		netif_dbg(efx, drv, efx->net_dev,
630 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
631 			  efx->rx_dma_len, efx->rx_page_buf_step,
632 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
633 
634 	/* Restore previously fixed features in hw_features and remove
635 	 * features which are fixed now
636 	 */
637 	efx->net_dev->hw_features |= efx->net_dev->features;
638 	efx->net_dev->hw_features &= ~efx->fixed_features;
639 	efx->net_dev->features |= efx->fixed_features;
640 	if (efx->net_dev->features != old_features)
641 		netdev_features_change(efx->net_dev);
642 
643 	/* RX filters may also have scatter-enabled flags */
644 	if (efx->rx_scatter != old_rx_scatter)
645 		efx->type->filter_update_rx_scatter(efx);
646 
647 	/* We must keep at least one descriptor in a TX ring empty.
648 	 * We could avoid this when the queue size does not exactly
649 	 * match the hardware ring size, but it's not that important.
650 	 * Therefore we stop the queue when one more skb might fill
651 	 * the ring completely.  We wake it when half way back to
652 	 * empty.
653 	 */
654 	efx->txq_stop_thresh = efx->txq_entries - ef4_tx_max_skb_descs(efx);
655 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
656 
657 	/* Initialise the channels */
658 	ef4_for_each_channel(channel, efx) {
659 		ef4_for_each_channel_tx_queue(tx_queue, channel) {
660 			ef4_init_tx_queue(tx_queue);
661 			atomic_inc(&efx->active_queues);
662 		}
663 
664 		ef4_for_each_channel_rx_queue(rx_queue, channel) {
665 			ef4_init_rx_queue(rx_queue);
666 			atomic_inc(&efx->active_queues);
667 			ef4_stop_eventq(channel);
668 			ef4_fast_push_rx_descriptors(rx_queue, false);
669 			ef4_start_eventq(channel);
670 		}
671 
672 		WARN_ON(channel->rx_pkt_n_frags);
673 	}
674 
675 	if (netif_device_present(efx->net_dev))
676 		netif_tx_wake_all_queues(efx->net_dev);
677 }
678 
679 static void ef4_stop_datapath(struct ef4_nic *efx)
680 {
681 	struct ef4_channel *channel;
682 	struct ef4_tx_queue *tx_queue;
683 	struct ef4_rx_queue *rx_queue;
684 	int rc;
685 
686 	EF4_ASSERT_RESET_SERIALISED(efx);
687 	BUG_ON(efx->port_enabled);
688 
689 	/* Stop RX refill */
690 	ef4_for_each_channel(channel, efx) {
691 		ef4_for_each_channel_rx_queue(rx_queue, channel)
692 			rx_queue->refill_enabled = false;
693 	}
694 
695 	ef4_for_each_channel(channel, efx) {
696 		/* RX packet processing is pipelined, so wait for the
697 		 * NAPI handler to complete.  At least event queue 0
698 		 * might be kept active by non-data events, so don't
699 		 * use napi_synchronize() but actually disable NAPI
700 		 * temporarily.
701 		 */
702 		if (ef4_channel_has_rx_queue(channel)) {
703 			ef4_stop_eventq(channel);
704 			ef4_start_eventq(channel);
705 		}
706 	}
707 
708 	rc = efx->type->fini_dmaq(efx);
709 	if (rc && EF4_WORKAROUND_7803(efx)) {
710 		/* Schedule a reset to recover from the flush failure. The
711 		 * descriptor caches reference memory we're about to free,
712 		 * but falcon_reconfigure_mac_wrapper() won't reconnect
713 		 * the MACs because of the pending reset.
714 		 */
715 		netif_err(efx, drv, efx->net_dev,
716 			  "Resetting to recover from flush failure\n");
717 		ef4_schedule_reset(efx, RESET_TYPE_ALL);
718 	} else if (rc) {
719 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
720 	} else {
721 		netif_dbg(efx, drv, efx->net_dev,
722 			  "successfully flushed all queues\n");
723 	}
724 
725 	ef4_for_each_channel(channel, efx) {
726 		ef4_for_each_channel_rx_queue(rx_queue, channel)
727 			ef4_fini_rx_queue(rx_queue);
728 		ef4_for_each_possible_channel_tx_queue(tx_queue, channel)
729 			ef4_fini_tx_queue(tx_queue);
730 	}
731 }
732 
733 static void ef4_remove_channel(struct ef4_channel *channel)
734 {
735 	struct ef4_tx_queue *tx_queue;
736 	struct ef4_rx_queue *rx_queue;
737 
738 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
739 		  "destroy chan %d\n", channel->channel);
740 
741 	ef4_for_each_channel_rx_queue(rx_queue, channel)
742 		ef4_remove_rx_queue(rx_queue);
743 	ef4_for_each_possible_channel_tx_queue(tx_queue, channel)
744 		ef4_remove_tx_queue(tx_queue);
745 	ef4_remove_eventq(channel);
746 	channel->type->post_remove(channel);
747 }
748 
749 static void ef4_remove_channels(struct ef4_nic *efx)
750 {
751 	struct ef4_channel *channel;
752 
753 	ef4_for_each_channel(channel, efx)
754 		ef4_remove_channel(channel);
755 }
756 
757 int
758 ef4_realloc_channels(struct ef4_nic *efx, u32 rxq_entries, u32 txq_entries)
759 {
760 	struct ef4_channel *other_channel[EF4_MAX_CHANNELS], *channel;
761 	u32 old_rxq_entries, old_txq_entries;
762 	unsigned i, next_buffer_table = 0;
763 	int rc, rc2;
764 
765 	rc = ef4_check_disabled(efx);
766 	if (rc)
767 		return rc;
768 
769 	/* Not all channels should be reallocated. We must avoid
770 	 * reallocating their buffer table entries.
771 	 */
772 	ef4_for_each_channel(channel, efx) {
773 		struct ef4_rx_queue *rx_queue;
774 		struct ef4_tx_queue *tx_queue;
775 
776 		if (channel->type->copy)
777 			continue;
778 		next_buffer_table = max(next_buffer_table,
779 					channel->eventq.index +
780 					channel->eventq.entries);
781 		ef4_for_each_channel_rx_queue(rx_queue, channel)
782 			next_buffer_table = max(next_buffer_table,
783 						rx_queue->rxd.index +
784 						rx_queue->rxd.entries);
785 		ef4_for_each_channel_tx_queue(tx_queue, channel)
786 			next_buffer_table = max(next_buffer_table,
787 						tx_queue->txd.index +
788 						tx_queue->txd.entries);
789 	}
790 
791 	ef4_device_detach_sync(efx);
792 	ef4_stop_all(efx);
793 	ef4_soft_disable_interrupts(efx);
794 
795 	/* Clone channels (where possible) */
796 	memset(other_channel, 0, sizeof(other_channel));
797 	for (i = 0; i < efx->n_channels; i++) {
798 		channel = efx->channel[i];
799 		if (channel->type->copy)
800 			channel = channel->type->copy(channel);
801 		if (!channel) {
802 			rc = -ENOMEM;
803 			goto out;
804 		}
805 		other_channel[i] = channel;
806 	}
807 
808 	/* Swap entry counts and channel pointers */
809 	old_rxq_entries = efx->rxq_entries;
810 	old_txq_entries = efx->txq_entries;
811 	efx->rxq_entries = rxq_entries;
812 	efx->txq_entries = txq_entries;
813 	for (i = 0; i < efx->n_channels; i++) {
814 		swap(efx->channel[i], other_channel[i]);
815 	}
816 
817 	/* Restart buffer table allocation */
818 	efx->next_buffer_table = next_buffer_table;
819 
820 	for (i = 0; i < efx->n_channels; i++) {
821 		channel = efx->channel[i];
822 		if (!channel->type->copy)
823 			continue;
824 		rc = ef4_probe_channel(channel);
825 		if (rc)
826 			goto rollback;
827 		ef4_init_napi_channel(efx->channel[i]);
828 	}
829 
830 out:
831 	/* Destroy unused channel structures */
832 	for (i = 0; i < efx->n_channels; i++) {
833 		channel = other_channel[i];
834 		if (channel && channel->type->copy) {
835 			ef4_fini_napi_channel(channel);
836 			ef4_remove_channel(channel);
837 			kfree(channel);
838 		}
839 	}
840 
841 	rc2 = ef4_soft_enable_interrupts(efx);
842 	if (rc2) {
843 		rc = rc ? rc : rc2;
844 		netif_err(efx, drv, efx->net_dev,
845 			  "unable to restart interrupts on channel reallocation\n");
846 		ef4_schedule_reset(efx, RESET_TYPE_DISABLE);
847 	} else {
848 		ef4_start_all(efx);
849 		netif_device_attach(efx->net_dev);
850 	}
851 	return rc;
852 
853 rollback:
854 	/* Swap back */
855 	efx->rxq_entries = old_rxq_entries;
856 	efx->txq_entries = old_txq_entries;
857 	for (i = 0; i < efx->n_channels; i++) {
858 		swap(efx->channel[i], other_channel[i]);
859 	}
860 	goto out;
861 }
862 
863 void ef4_schedule_slow_fill(struct ef4_rx_queue *rx_queue)
864 {
865 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
866 }
867 
868 static const struct ef4_channel_type ef4_default_channel_type = {
869 	.pre_probe		= ef4_channel_dummy_op_int,
870 	.post_remove		= ef4_channel_dummy_op_void,
871 	.get_name		= ef4_get_channel_name,
872 	.copy			= ef4_copy_channel,
873 	.keep_eventq		= false,
874 };
875 
876 int ef4_channel_dummy_op_int(struct ef4_channel *channel)
877 {
878 	return 0;
879 }
880 
881 void ef4_channel_dummy_op_void(struct ef4_channel *channel)
882 {
883 }
884 
885 /**************************************************************************
886  *
887  * Port handling
888  *
889  **************************************************************************/
890 
891 /* This ensures that the kernel is kept informed (via
892  * netif_carrier_on/off) of the link status, and also maintains the
893  * link status's stop on the port's TX queue.
894  */
895 void ef4_link_status_changed(struct ef4_nic *efx)
896 {
897 	struct ef4_link_state *link_state = &efx->link_state;
898 
899 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
900 	 * that no events are triggered between unregister_netdev() and the
901 	 * driver unloading. A more general condition is that NETDEV_CHANGE
902 	 * can only be generated between NETDEV_UP and NETDEV_DOWN */
903 	if (!netif_running(efx->net_dev))
904 		return;
905 
906 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
907 		efx->n_link_state_changes++;
908 
909 		if (link_state->up)
910 			netif_carrier_on(efx->net_dev);
911 		else
912 			netif_carrier_off(efx->net_dev);
913 	}
914 
915 	/* Status message for kernel log */
916 	if (link_state->up)
917 		netif_info(efx, link, efx->net_dev,
918 			   "link up at %uMbps %s-duplex (MTU %d)\n",
919 			   link_state->speed, link_state->fd ? "full" : "half",
920 			   efx->net_dev->mtu);
921 	else
922 		netif_info(efx, link, efx->net_dev, "link down\n");
923 }
924 
925 void ef4_link_set_advertising(struct ef4_nic *efx, u32 advertising)
926 {
927 	efx->link_advertising = advertising;
928 	if (advertising) {
929 		if (advertising & ADVERTISED_Pause)
930 			efx->wanted_fc |= (EF4_FC_TX | EF4_FC_RX);
931 		else
932 			efx->wanted_fc &= ~(EF4_FC_TX | EF4_FC_RX);
933 		if (advertising & ADVERTISED_Asym_Pause)
934 			efx->wanted_fc ^= EF4_FC_TX;
935 	}
936 }
937 
938 void ef4_link_set_wanted_fc(struct ef4_nic *efx, u8 wanted_fc)
939 {
940 	efx->wanted_fc = wanted_fc;
941 	if (efx->link_advertising) {
942 		if (wanted_fc & EF4_FC_RX)
943 			efx->link_advertising |= (ADVERTISED_Pause |
944 						  ADVERTISED_Asym_Pause);
945 		else
946 			efx->link_advertising &= ~(ADVERTISED_Pause |
947 						   ADVERTISED_Asym_Pause);
948 		if (wanted_fc & EF4_FC_TX)
949 			efx->link_advertising ^= ADVERTISED_Asym_Pause;
950 	}
951 }
952 
953 static void ef4_fini_port(struct ef4_nic *efx);
954 
955 /* We assume that efx->type->reconfigure_mac will always try to sync RX
956  * filters and therefore needs to read-lock the filter table against freeing
957  */
958 void ef4_mac_reconfigure(struct ef4_nic *efx)
959 {
960 	down_read(&efx->filter_sem);
961 	efx->type->reconfigure_mac(efx);
962 	up_read(&efx->filter_sem);
963 }
964 
965 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
966  * the MAC appropriately. All other PHY configuration changes are pushed
967  * through phy_op->set_link_ksettings(), and pushed asynchronously to the MAC
968  * through ef4_monitor().
969  *
970  * Callers must hold the mac_lock
971  */
972 int __ef4_reconfigure_port(struct ef4_nic *efx)
973 {
974 	enum ef4_phy_mode phy_mode;
975 	int rc;
976 
977 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
978 
979 	/* Disable PHY transmit in mac level loopbacks */
980 	phy_mode = efx->phy_mode;
981 	if (LOOPBACK_INTERNAL(efx))
982 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
983 	else
984 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
985 
986 	rc = efx->type->reconfigure_port(efx);
987 
988 	if (rc)
989 		efx->phy_mode = phy_mode;
990 
991 	return rc;
992 }
993 
994 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
995  * disabled. */
996 int ef4_reconfigure_port(struct ef4_nic *efx)
997 {
998 	int rc;
999 
1000 	EF4_ASSERT_RESET_SERIALISED(efx);
1001 
1002 	mutex_lock(&efx->mac_lock);
1003 	rc = __ef4_reconfigure_port(efx);
1004 	mutex_unlock(&efx->mac_lock);
1005 
1006 	return rc;
1007 }
1008 
1009 /* Asynchronous work item for changing MAC promiscuity and multicast
1010  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
1011  * MAC directly. */
1012 static void ef4_mac_work(struct work_struct *data)
1013 {
1014 	struct ef4_nic *efx = container_of(data, struct ef4_nic, mac_work);
1015 
1016 	mutex_lock(&efx->mac_lock);
1017 	if (efx->port_enabled)
1018 		ef4_mac_reconfigure(efx);
1019 	mutex_unlock(&efx->mac_lock);
1020 }
1021 
1022 static int ef4_probe_port(struct ef4_nic *efx)
1023 {
1024 	int rc;
1025 
1026 	netif_dbg(efx, probe, efx->net_dev, "create port\n");
1027 
1028 	if (phy_flash_cfg)
1029 		efx->phy_mode = PHY_MODE_SPECIAL;
1030 
1031 	/* Connect up MAC/PHY operations table */
1032 	rc = efx->type->probe_port(efx);
1033 	if (rc)
1034 		return rc;
1035 
1036 	/* Initialise MAC address to permanent address */
1037 	eth_hw_addr_set(efx->net_dev, efx->net_dev->perm_addr);
1038 
1039 	return 0;
1040 }
1041 
1042 static int ef4_init_port(struct ef4_nic *efx)
1043 {
1044 	int rc;
1045 
1046 	netif_dbg(efx, drv, efx->net_dev, "init port\n");
1047 
1048 	mutex_lock(&efx->mac_lock);
1049 
1050 	rc = efx->phy_op->init(efx);
1051 	if (rc)
1052 		goto fail1;
1053 
1054 	efx->port_initialized = true;
1055 
1056 	/* Reconfigure the MAC before creating dma queues (required for
1057 	 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1058 	ef4_mac_reconfigure(efx);
1059 
1060 	/* Ensure the PHY advertises the correct flow control settings */
1061 	rc = efx->phy_op->reconfigure(efx);
1062 	if (rc && rc != -EPERM)
1063 		goto fail2;
1064 
1065 	mutex_unlock(&efx->mac_lock);
1066 	return 0;
1067 
1068 fail2:
1069 	efx->phy_op->fini(efx);
1070 fail1:
1071 	mutex_unlock(&efx->mac_lock);
1072 	return rc;
1073 }
1074 
1075 static void ef4_start_port(struct ef4_nic *efx)
1076 {
1077 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1078 	BUG_ON(efx->port_enabled);
1079 
1080 	mutex_lock(&efx->mac_lock);
1081 	efx->port_enabled = true;
1082 
1083 	/* Ensure MAC ingress/egress is enabled */
1084 	ef4_mac_reconfigure(efx);
1085 
1086 	mutex_unlock(&efx->mac_lock);
1087 }
1088 
1089 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1090  * and the async self-test, wait for them to finish and prevent them
1091  * being scheduled again.  This doesn't cover online resets, which
1092  * should only be cancelled when removing the device.
1093  */
1094 static void ef4_stop_port(struct ef4_nic *efx)
1095 {
1096 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1097 
1098 	EF4_ASSERT_RESET_SERIALISED(efx);
1099 
1100 	mutex_lock(&efx->mac_lock);
1101 	efx->port_enabled = false;
1102 	mutex_unlock(&efx->mac_lock);
1103 
1104 	/* Serialise against ef4_set_multicast_list() */
1105 	netif_addr_lock_bh(efx->net_dev);
1106 	netif_addr_unlock_bh(efx->net_dev);
1107 
1108 	cancel_delayed_work_sync(&efx->monitor_work);
1109 	ef4_selftest_async_cancel(efx);
1110 	cancel_work_sync(&efx->mac_work);
1111 }
1112 
1113 static void ef4_fini_port(struct ef4_nic *efx)
1114 {
1115 	netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1116 
1117 	if (!efx->port_initialized)
1118 		return;
1119 
1120 	efx->phy_op->fini(efx);
1121 	efx->port_initialized = false;
1122 
1123 	efx->link_state.up = false;
1124 	ef4_link_status_changed(efx);
1125 }
1126 
1127 static void ef4_remove_port(struct ef4_nic *efx)
1128 {
1129 	netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1130 
1131 	efx->type->remove_port(efx);
1132 }
1133 
1134 /**************************************************************************
1135  *
1136  * NIC handling
1137  *
1138  **************************************************************************/
1139 
1140 static LIST_HEAD(ef4_primary_list);
1141 static LIST_HEAD(ef4_unassociated_list);
1142 
1143 static bool ef4_same_controller(struct ef4_nic *left, struct ef4_nic *right)
1144 {
1145 	return left->type == right->type &&
1146 		left->vpd_sn && right->vpd_sn &&
1147 		!strcmp(left->vpd_sn, right->vpd_sn);
1148 }
1149 
1150 static void ef4_associate(struct ef4_nic *efx)
1151 {
1152 	struct ef4_nic *other, *next;
1153 
1154 	if (efx->primary == efx) {
1155 		/* Adding primary function; look for secondaries */
1156 
1157 		netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1158 		list_add_tail(&efx->node, &ef4_primary_list);
1159 
1160 		list_for_each_entry_safe(other, next, &ef4_unassociated_list,
1161 					 node) {
1162 			if (ef4_same_controller(efx, other)) {
1163 				list_del(&other->node);
1164 				netif_dbg(other, probe, other->net_dev,
1165 					  "moving to secondary list of %s %s\n",
1166 					  pci_name(efx->pci_dev),
1167 					  efx->net_dev->name);
1168 				list_add_tail(&other->node,
1169 					      &efx->secondary_list);
1170 				other->primary = efx;
1171 			}
1172 		}
1173 	} else {
1174 		/* Adding secondary function; look for primary */
1175 
1176 		list_for_each_entry(other, &ef4_primary_list, node) {
1177 			if (ef4_same_controller(efx, other)) {
1178 				netif_dbg(efx, probe, efx->net_dev,
1179 					  "adding to secondary list of %s %s\n",
1180 					  pci_name(other->pci_dev),
1181 					  other->net_dev->name);
1182 				list_add_tail(&efx->node,
1183 					      &other->secondary_list);
1184 				efx->primary = other;
1185 				return;
1186 			}
1187 		}
1188 
1189 		netif_dbg(efx, probe, efx->net_dev,
1190 			  "adding to unassociated list\n");
1191 		list_add_tail(&efx->node, &ef4_unassociated_list);
1192 	}
1193 }
1194 
1195 static void ef4_dissociate(struct ef4_nic *efx)
1196 {
1197 	struct ef4_nic *other, *next;
1198 
1199 	list_del(&efx->node);
1200 	efx->primary = NULL;
1201 
1202 	list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1203 		list_del(&other->node);
1204 		netif_dbg(other, probe, other->net_dev,
1205 			  "moving to unassociated list\n");
1206 		list_add_tail(&other->node, &ef4_unassociated_list);
1207 		other->primary = NULL;
1208 	}
1209 }
1210 
1211 /* This configures the PCI device to enable I/O and DMA. */
1212 static int ef4_init_io(struct ef4_nic *efx)
1213 {
1214 	struct pci_dev *pci_dev = efx->pci_dev;
1215 	dma_addr_t dma_mask = efx->type->max_dma_mask;
1216 	unsigned int mem_map_size = efx->type->mem_map_size(efx);
1217 	int rc, bar;
1218 
1219 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1220 
1221 	bar = efx->type->mem_bar;
1222 
1223 	rc = pci_enable_device(pci_dev);
1224 	if (rc) {
1225 		netif_err(efx, probe, efx->net_dev,
1226 			  "failed to enable PCI device\n");
1227 		goto fail1;
1228 	}
1229 
1230 	pci_set_master(pci_dev);
1231 
1232 	/* Set the PCI DMA mask.  Try all possibilities from our genuine mask
1233 	 * down to 32 bits, because some architectures will allow 40 bit
1234 	 * masks event though they reject 46 bit masks.
1235 	 */
1236 	while (dma_mask > 0x7fffffffUL) {
1237 		rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1238 		if (rc == 0)
1239 			break;
1240 		dma_mask >>= 1;
1241 	}
1242 	if (rc) {
1243 		netif_err(efx, probe, efx->net_dev,
1244 			  "could not find a suitable DMA mask\n");
1245 		goto fail2;
1246 	}
1247 	netif_dbg(efx, probe, efx->net_dev,
1248 		  "using DMA mask %llx\n", (unsigned long long) dma_mask);
1249 
1250 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1251 	rc = pci_request_region(pci_dev, bar, "sfc");
1252 	if (rc) {
1253 		netif_err(efx, probe, efx->net_dev,
1254 			  "request for memory BAR failed\n");
1255 		rc = -EIO;
1256 		goto fail3;
1257 	}
1258 	efx->membase = ioremap(efx->membase_phys, mem_map_size);
1259 	if (!efx->membase) {
1260 		netif_err(efx, probe, efx->net_dev,
1261 			  "could not map memory BAR at %llx+%x\n",
1262 			  (unsigned long long)efx->membase_phys, mem_map_size);
1263 		rc = -ENOMEM;
1264 		goto fail4;
1265 	}
1266 	netif_dbg(efx, probe, efx->net_dev,
1267 		  "memory BAR at %llx+%x (virtual %p)\n",
1268 		  (unsigned long long)efx->membase_phys, mem_map_size,
1269 		  efx->membase);
1270 
1271 	return 0;
1272 
1273  fail4:
1274 	pci_release_region(efx->pci_dev, bar);
1275  fail3:
1276 	efx->membase_phys = 0;
1277  fail2:
1278 	pci_disable_device(efx->pci_dev);
1279  fail1:
1280 	return rc;
1281 }
1282 
1283 static void ef4_fini_io(struct ef4_nic *efx)
1284 {
1285 	int bar;
1286 
1287 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1288 
1289 	if (efx->membase) {
1290 		iounmap(efx->membase);
1291 		efx->membase = NULL;
1292 	}
1293 
1294 	if (efx->membase_phys) {
1295 		bar = efx->type->mem_bar;
1296 		pci_release_region(efx->pci_dev, bar);
1297 		efx->membase_phys = 0;
1298 	}
1299 
1300 	/* Don't disable bus-mastering if VFs are assigned */
1301 	if (!pci_vfs_assigned(efx->pci_dev))
1302 		pci_disable_device(efx->pci_dev);
1303 }
1304 
1305 void ef4_set_default_rx_indir_table(struct ef4_nic *efx)
1306 {
1307 	size_t i;
1308 
1309 	for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1310 		efx->rx_indir_table[i] =
1311 			ethtool_rxfh_indir_default(i, efx->rss_spread);
1312 }
1313 
1314 static unsigned int ef4_wanted_parallelism(struct ef4_nic *efx)
1315 {
1316 	cpumask_var_t thread_mask;
1317 	unsigned int count;
1318 	int cpu;
1319 
1320 	if (rss_cpus) {
1321 		count = rss_cpus;
1322 	} else {
1323 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1324 			netif_warn(efx, probe, efx->net_dev,
1325 				   "RSS disabled due to allocation failure\n");
1326 			return 1;
1327 		}
1328 
1329 		count = 0;
1330 		for_each_online_cpu(cpu) {
1331 			if (!cpumask_test_cpu(cpu, thread_mask)) {
1332 				++count;
1333 				cpumask_or(thread_mask, thread_mask,
1334 					   topology_sibling_cpumask(cpu));
1335 			}
1336 		}
1337 
1338 		free_cpumask_var(thread_mask);
1339 	}
1340 
1341 	if (count > EF4_MAX_RX_QUEUES) {
1342 		netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn,
1343 			       "Reducing number of rx queues from %u to %u.\n",
1344 			       count, EF4_MAX_RX_QUEUES);
1345 		count = EF4_MAX_RX_QUEUES;
1346 	}
1347 
1348 	return count;
1349 }
1350 
1351 /* Probe the number and type of interrupts we are able to obtain, and
1352  * the resulting numbers of channels and RX queues.
1353  */
1354 static int ef4_probe_interrupts(struct ef4_nic *efx)
1355 {
1356 	unsigned int extra_channels = 0;
1357 	unsigned int i, j;
1358 	int rc;
1359 
1360 	for (i = 0; i < EF4_MAX_EXTRA_CHANNELS; i++)
1361 		if (efx->extra_channel_type[i])
1362 			++extra_channels;
1363 
1364 	if (efx->interrupt_mode == EF4_INT_MODE_MSIX) {
1365 		struct msix_entry xentries[EF4_MAX_CHANNELS];
1366 		unsigned int n_channels;
1367 
1368 		n_channels = ef4_wanted_parallelism(efx);
1369 		if (ef4_separate_tx_channels)
1370 			n_channels *= 2;
1371 		n_channels += extra_channels;
1372 		n_channels = min(n_channels, efx->max_channels);
1373 
1374 		for (i = 0; i < n_channels; i++)
1375 			xentries[i].entry = i;
1376 		rc = pci_enable_msix_range(efx->pci_dev,
1377 					   xentries, 1, n_channels);
1378 		if (rc < 0) {
1379 			/* Fall back to single channel MSI */
1380 			efx->interrupt_mode = EF4_INT_MODE_MSI;
1381 			netif_err(efx, drv, efx->net_dev,
1382 				  "could not enable MSI-X\n");
1383 		} else if (rc < n_channels) {
1384 			netif_err(efx, drv, efx->net_dev,
1385 				  "WARNING: Insufficient MSI-X vectors"
1386 				  " available (%d < %u).\n", rc, n_channels);
1387 			netif_err(efx, drv, efx->net_dev,
1388 				  "WARNING: Performance may be reduced.\n");
1389 			n_channels = rc;
1390 		}
1391 
1392 		if (rc > 0) {
1393 			efx->n_channels = n_channels;
1394 			if (n_channels > extra_channels)
1395 				n_channels -= extra_channels;
1396 			if (ef4_separate_tx_channels) {
1397 				efx->n_tx_channels = min(max(n_channels / 2,
1398 							     1U),
1399 							 efx->max_tx_channels);
1400 				efx->n_rx_channels = max(n_channels -
1401 							 efx->n_tx_channels,
1402 							 1U);
1403 			} else {
1404 				efx->n_tx_channels = min(n_channels,
1405 							 efx->max_tx_channels);
1406 				efx->n_rx_channels = n_channels;
1407 			}
1408 			for (i = 0; i < efx->n_channels; i++)
1409 				ef4_get_channel(efx, i)->irq =
1410 					xentries[i].vector;
1411 		}
1412 	}
1413 
1414 	/* Try single interrupt MSI */
1415 	if (efx->interrupt_mode == EF4_INT_MODE_MSI) {
1416 		efx->n_channels = 1;
1417 		efx->n_rx_channels = 1;
1418 		efx->n_tx_channels = 1;
1419 		rc = pci_enable_msi(efx->pci_dev);
1420 		if (rc == 0) {
1421 			ef4_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1422 		} else {
1423 			netif_err(efx, drv, efx->net_dev,
1424 				  "could not enable MSI\n");
1425 			efx->interrupt_mode = EF4_INT_MODE_LEGACY;
1426 		}
1427 	}
1428 
1429 	/* Assume legacy interrupts */
1430 	if (efx->interrupt_mode == EF4_INT_MODE_LEGACY) {
1431 		efx->n_channels = 1 + (ef4_separate_tx_channels ? 1 : 0);
1432 		efx->n_rx_channels = 1;
1433 		efx->n_tx_channels = 1;
1434 		efx->legacy_irq = efx->pci_dev->irq;
1435 	}
1436 
1437 	/* Assign extra channels if possible */
1438 	j = efx->n_channels;
1439 	for (i = 0; i < EF4_MAX_EXTRA_CHANNELS; i++) {
1440 		if (!efx->extra_channel_type[i])
1441 			continue;
1442 		if (efx->interrupt_mode != EF4_INT_MODE_MSIX ||
1443 		    efx->n_channels <= extra_channels) {
1444 			efx->extra_channel_type[i]->handle_no_channel(efx);
1445 		} else {
1446 			--j;
1447 			ef4_get_channel(efx, j)->type =
1448 				efx->extra_channel_type[i];
1449 		}
1450 	}
1451 
1452 	efx->rss_spread = efx->n_rx_channels;
1453 
1454 	return 0;
1455 }
1456 
1457 static int ef4_soft_enable_interrupts(struct ef4_nic *efx)
1458 {
1459 	struct ef4_channel *channel, *end_channel;
1460 	int rc;
1461 
1462 	BUG_ON(efx->state == STATE_DISABLED);
1463 
1464 	efx->irq_soft_enabled = true;
1465 	smp_wmb();
1466 
1467 	ef4_for_each_channel(channel, efx) {
1468 		if (!channel->type->keep_eventq) {
1469 			rc = ef4_init_eventq(channel);
1470 			if (rc)
1471 				goto fail;
1472 		}
1473 		ef4_start_eventq(channel);
1474 	}
1475 
1476 	return 0;
1477 fail:
1478 	end_channel = channel;
1479 	ef4_for_each_channel(channel, efx) {
1480 		if (channel == end_channel)
1481 			break;
1482 		ef4_stop_eventq(channel);
1483 		if (!channel->type->keep_eventq)
1484 			ef4_fini_eventq(channel);
1485 	}
1486 
1487 	return rc;
1488 }
1489 
1490 static void ef4_soft_disable_interrupts(struct ef4_nic *efx)
1491 {
1492 	struct ef4_channel *channel;
1493 
1494 	if (efx->state == STATE_DISABLED)
1495 		return;
1496 
1497 	efx->irq_soft_enabled = false;
1498 	smp_wmb();
1499 
1500 	if (efx->legacy_irq)
1501 		synchronize_irq(efx->legacy_irq);
1502 
1503 	ef4_for_each_channel(channel, efx) {
1504 		if (channel->irq)
1505 			synchronize_irq(channel->irq);
1506 
1507 		ef4_stop_eventq(channel);
1508 		if (!channel->type->keep_eventq)
1509 			ef4_fini_eventq(channel);
1510 	}
1511 }
1512 
1513 static int ef4_enable_interrupts(struct ef4_nic *efx)
1514 {
1515 	struct ef4_channel *channel, *end_channel;
1516 	int rc;
1517 
1518 	BUG_ON(efx->state == STATE_DISABLED);
1519 
1520 	if (efx->eeh_disabled_legacy_irq) {
1521 		enable_irq(efx->legacy_irq);
1522 		efx->eeh_disabled_legacy_irq = false;
1523 	}
1524 
1525 	efx->type->irq_enable_master(efx);
1526 
1527 	ef4_for_each_channel(channel, efx) {
1528 		if (channel->type->keep_eventq) {
1529 			rc = ef4_init_eventq(channel);
1530 			if (rc)
1531 				goto fail;
1532 		}
1533 	}
1534 
1535 	rc = ef4_soft_enable_interrupts(efx);
1536 	if (rc)
1537 		goto fail;
1538 
1539 	return 0;
1540 
1541 fail:
1542 	end_channel = channel;
1543 	ef4_for_each_channel(channel, efx) {
1544 		if (channel == end_channel)
1545 			break;
1546 		if (channel->type->keep_eventq)
1547 			ef4_fini_eventq(channel);
1548 	}
1549 
1550 	efx->type->irq_disable_non_ev(efx);
1551 
1552 	return rc;
1553 }
1554 
1555 static void ef4_disable_interrupts(struct ef4_nic *efx)
1556 {
1557 	struct ef4_channel *channel;
1558 
1559 	ef4_soft_disable_interrupts(efx);
1560 
1561 	ef4_for_each_channel(channel, efx) {
1562 		if (channel->type->keep_eventq)
1563 			ef4_fini_eventq(channel);
1564 	}
1565 
1566 	efx->type->irq_disable_non_ev(efx);
1567 }
1568 
1569 static void ef4_remove_interrupts(struct ef4_nic *efx)
1570 {
1571 	struct ef4_channel *channel;
1572 
1573 	/* Remove MSI/MSI-X interrupts */
1574 	ef4_for_each_channel(channel, efx)
1575 		channel->irq = 0;
1576 	pci_disable_msi(efx->pci_dev);
1577 	pci_disable_msix(efx->pci_dev);
1578 
1579 	/* Remove legacy interrupt */
1580 	efx->legacy_irq = 0;
1581 }
1582 
1583 static void ef4_set_channels(struct ef4_nic *efx)
1584 {
1585 	struct ef4_channel *channel;
1586 	struct ef4_tx_queue *tx_queue;
1587 
1588 	efx->tx_channel_offset =
1589 		ef4_separate_tx_channels ?
1590 		efx->n_channels - efx->n_tx_channels : 0;
1591 
1592 	/* We need to mark which channels really have RX and TX
1593 	 * queues, and adjust the TX queue numbers if we have separate
1594 	 * RX-only and TX-only channels.
1595 	 */
1596 	ef4_for_each_channel(channel, efx) {
1597 		if (channel->channel < efx->n_rx_channels)
1598 			channel->rx_queue.core_index = channel->channel;
1599 		else
1600 			channel->rx_queue.core_index = -1;
1601 
1602 		ef4_for_each_channel_tx_queue(tx_queue, channel)
1603 			tx_queue->queue -= (efx->tx_channel_offset *
1604 					    EF4_TXQ_TYPES);
1605 	}
1606 }
1607 
1608 static int ef4_probe_nic(struct ef4_nic *efx)
1609 {
1610 	int rc;
1611 
1612 	netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1613 
1614 	/* Carry out hardware-type specific initialisation */
1615 	rc = efx->type->probe(efx);
1616 	if (rc)
1617 		return rc;
1618 
1619 	do {
1620 		if (!efx->max_channels || !efx->max_tx_channels) {
1621 			netif_err(efx, drv, efx->net_dev,
1622 				  "Insufficient resources to allocate"
1623 				  " any channels\n");
1624 			rc = -ENOSPC;
1625 			goto fail1;
1626 		}
1627 
1628 		/* Determine the number of channels and queues by trying
1629 		 * to hook in MSI-X interrupts.
1630 		 */
1631 		rc = ef4_probe_interrupts(efx);
1632 		if (rc)
1633 			goto fail1;
1634 
1635 		ef4_set_channels(efx);
1636 
1637 		/* dimension_resources can fail with EAGAIN */
1638 		rc = efx->type->dimension_resources(efx);
1639 		if (rc != 0 && rc != -EAGAIN)
1640 			goto fail2;
1641 
1642 		if (rc == -EAGAIN)
1643 			/* try again with new max_channels */
1644 			ef4_remove_interrupts(efx);
1645 
1646 	} while (rc == -EAGAIN);
1647 
1648 	if (efx->n_channels > 1)
1649 		netdev_rss_key_fill(&efx->rx_hash_key,
1650 				    sizeof(efx->rx_hash_key));
1651 	ef4_set_default_rx_indir_table(efx);
1652 
1653 	netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1654 	netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1655 
1656 	/* Initialise the interrupt moderation settings */
1657 	efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
1658 	ef4_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1659 				true);
1660 
1661 	return 0;
1662 
1663 fail2:
1664 	ef4_remove_interrupts(efx);
1665 fail1:
1666 	efx->type->remove(efx);
1667 	return rc;
1668 }
1669 
1670 static void ef4_remove_nic(struct ef4_nic *efx)
1671 {
1672 	netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1673 
1674 	ef4_remove_interrupts(efx);
1675 	efx->type->remove(efx);
1676 }
1677 
1678 static int ef4_probe_filters(struct ef4_nic *efx)
1679 {
1680 	int rc;
1681 
1682 	spin_lock_init(&efx->filter_lock);
1683 	init_rwsem(&efx->filter_sem);
1684 	mutex_lock(&efx->mac_lock);
1685 	down_write(&efx->filter_sem);
1686 	rc = efx->type->filter_table_probe(efx);
1687 	if (rc)
1688 		goto out_unlock;
1689 
1690 #ifdef CONFIG_RFS_ACCEL
1691 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
1692 		struct ef4_channel *channel;
1693 		int i, success = 1;
1694 
1695 		ef4_for_each_channel(channel, efx) {
1696 			channel->rps_flow_id =
1697 				kcalloc(efx->type->max_rx_ip_filters,
1698 					sizeof(*channel->rps_flow_id),
1699 					GFP_KERNEL);
1700 			if (!channel->rps_flow_id)
1701 				success = 0;
1702 			else
1703 				for (i = 0;
1704 				     i < efx->type->max_rx_ip_filters;
1705 				     ++i)
1706 					channel->rps_flow_id[i] =
1707 						RPS_FLOW_ID_INVALID;
1708 		}
1709 
1710 		if (!success) {
1711 			ef4_for_each_channel(channel, efx)
1712 				kfree(channel->rps_flow_id);
1713 			efx->type->filter_table_remove(efx);
1714 			rc = -ENOMEM;
1715 			goto out_unlock;
1716 		}
1717 
1718 		efx->rps_expire_index = efx->rps_expire_channel = 0;
1719 	}
1720 #endif
1721 out_unlock:
1722 	up_write(&efx->filter_sem);
1723 	mutex_unlock(&efx->mac_lock);
1724 	return rc;
1725 }
1726 
1727 static void ef4_remove_filters(struct ef4_nic *efx)
1728 {
1729 #ifdef CONFIG_RFS_ACCEL
1730 	struct ef4_channel *channel;
1731 
1732 	ef4_for_each_channel(channel, efx)
1733 		kfree(channel->rps_flow_id);
1734 #endif
1735 	down_write(&efx->filter_sem);
1736 	efx->type->filter_table_remove(efx);
1737 	up_write(&efx->filter_sem);
1738 }
1739 
1740 static void ef4_restore_filters(struct ef4_nic *efx)
1741 {
1742 	down_read(&efx->filter_sem);
1743 	efx->type->filter_table_restore(efx);
1744 	up_read(&efx->filter_sem);
1745 }
1746 
1747 /**************************************************************************
1748  *
1749  * NIC startup/shutdown
1750  *
1751  *************************************************************************/
1752 
1753 static int ef4_probe_all(struct ef4_nic *efx)
1754 {
1755 	int rc;
1756 
1757 	rc = ef4_probe_nic(efx);
1758 	if (rc) {
1759 		netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1760 		goto fail1;
1761 	}
1762 
1763 	rc = ef4_probe_port(efx);
1764 	if (rc) {
1765 		netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1766 		goto fail2;
1767 	}
1768 
1769 	BUILD_BUG_ON(EF4_DEFAULT_DMAQ_SIZE < EF4_RXQ_MIN_ENT);
1770 	if (WARN_ON(EF4_DEFAULT_DMAQ_SIZE < EF4_TXQ_MIN_ENT(efx))) {
1771 		rc = -EINVAL;
1772 		goto fail3;
1773 	}
1774 	efx->rxq_entries = efx->txq_entries = EF4_DEFAULT_DMAQ_SIZE;
1775 
1776 	rc = ef4_probe_filters(efx);
1777 	if (rc) {
1778 		netif_err(efx, probe, efx->net_dev,
1779 			  "failed to create filter tables\n");
1780 		goto fail4;
1781 	}
1782 
1783 	rc = ef4_probe_channels(efx);
1784 	if (rc)
1785 		goto fail5;
1786 
1787 	return 0;
1788 
1789  fail5:
1790 	ef4_remove_filters(efx);
1791  fail4:
1792  fail3:
1793 	ef4_remove_port(efx);
1794  fail2:
1795 	ef4_remove_nic(efx);
1796  fail1:
1797 	return rc;
1798 }
1799 
1800 /* If the interface is supposed to be running but is not, start
1801  * the hardware and software data path, regular activity for the port
1802  * (MAC statistics, link polling, etc.) and schedule the port to be
1803  * reconfigured.  Interrupts must already be enabled.  This function
1804  * is safe to call multiple times, so long as the NIC is not disabled.
1805  * Requires the RTNL lock.
1806  */
1807 static void ef4_start_all(struct ef4_nic *efx)
1808 {
1809 	EF4_ASSERT_RESET_SERIALISED(efx);
1810 	BUG_ON(efx->state == STATE_DISABLED);
1811 
1812 	/* Check that it is appropriate to restart the interface. All
1813 	 * of these flags are safe to read under just the rtnl lock */
1814 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
1815 	    efx->reset_pending)
1816 		return;
1817 
1818 	ef4_start_port(efx);
1819 	ef4_start_datapath(efx);
1820 
1821 	/* Start the hardware monitor if there is one */
1822 	if (efx->type->monitor != NULL)
1823 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
1824 				   ef4_monitor_interval);
1825 
1826 	efx->type->start_stats(efx);
1827 	efx->type->pull_stats(efx);
1828 	spin_lock_bh(&efx->stats_lock);
1829 	efx->type->update_stats(efx, NULL, NULL);
1830 	spin_unlock_bh(&efx->stats_lock);
1831 }
1832 
1833 /* Quiesce the hardware and software data path, and regular activity
1834  * for the port without bringing the link down.  Safe to call multiple
1835  * times with the NIC in almost any state, but interrupts should be
1836  * enabled.  Requires the RTNL lock.
1837  */
1838 static void ef4_stop_all(struct ef4_nic *efx)
1839 {
1840 	EF4_ASSERT_RESET_SERIALISED(efx);
1841 
1842 	/* port_enabled can be read safely under the rtnl lock */
1843 	if (!efx->port_enabled)
1844 		return;
1845 
1846 	/* update stats before we go down so we can accurately count
1847 	 * rx_nodesc_drops
1848 	 */
1849 	efx->type->pull_stats(efx);
1850 	spin_lock_bh(&efx->stats_lock);
1851 	efx->type->update_stats(efx, NULL, NULL);
1852 	spin_unlock_bh(&efx->stats_lock);
1853 	efx->type->stop_stats(efx);
1854 	ef4_stop_port(efx);
1855 
1856 	/* Stop the kernel transmit interface.  This is only valid if
1857 	 * the device is stopped or detached; otherwise the watchdog
1858 	 * may fire immediately.
1859 	 */
1860 	WARN_ON(netif_running(efx->net_dev) &&
1861 		netif_device_present(efx->net_dev));
1862 	netif_tx_disable(efx->net_dev);
1863 
1864 	ef4_stop_datapath(efx);
1865 }
1866 
1867 static void ef4_remove_all(struct ef4_nic *efx)
1868 {
1869 	ef4_remove_channels(efx);
1870 	ef4_remove_filters(efx);
1871 	ef4_remove_port(efx);
1872 	ef4_remove_nic(efx);
1873 }
1874 
1875 /**************************************************************************
1876  *
1877  * Interrupt moderation
1878  *
1879  **************************************************************************/
1880 unsigned int ef4_usecs_to_ticks(struct ef4_nic *efx, unsigned int usecs)
1881 {
1882 	if (usecs == 0)
1883 		return 0;
1884 	if (usecs * 1000 < efx->timer_quantum_ns)
1885 		return 1; /* never round down to 0 */
1886 	return usecs * 1000 / efx->timer_quantum_ns;
1887 }
1888 
1889 unsigned int ef4_ticks_to_usecs(struct ef4_nic *efx, unsigned int ticks)
1890 {
1891 	/* We must round up when converting ticks to microseconds
1892 	 * because we round down when converting the other way.
1893 	 */
1894 	return DIV_ROUND_UP(ticks * efx->timer_quantum_ns, 1000);
1895 }
1896 
1897 /* Set interrupt moderation parameters */
1898 int ef4_init_irq_moderation(struct ef4_nic *efx, unsigned int tx_usecs,
1899 			    unsigned int rx_usecs, bool rx_adaptive,
1900 			    bool rx_may_override_tx)
1901 {
1902 	struct ef4_channel *channel;
1903 	unsigned int timer_max_us;
1904 
1905 	EF4_ASSERT_RESET_SERIALISED(efx);
1906 
1907 	timer_max_us = efx->timer_max_ns / 1000;
1908 
1909 	if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
1910 		return -EINVAL;
1911 
1912 	if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
1913 	    !rx_may_override_tx) {
1914 		netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1915 			  "RX and TX IRQ moderation must be equal\n");
1916 		return -EINVAL;
1917 	}
1918 
1919 	efx->irq_rx_adaptive = rx_adaptive;
1920 	efx->irq_rx_moderation_us = rx_usecs;
1921 	ef4_for_each_channel(channel, efx) {
1922 		if (ef4_channel_has_rx_queue(channel))
1923 			channel->irq_moderation_us = rx_usecs;
1924 		else if (ef4_channel_has_tx_queues(channel))
1925 			channel->irq_moderation_us = tx_usecs;
1926 	}
1927 
1928 	return 0;
1929 }
1930 
1931 void ef4_get_irq_moderation(struct ef4_nic *efx, unsigned int *tx_usecs,
1932 			    unsigned int *rx_usecs, bool *rx_adaptive)
1933 {
1934 	*rx_adaptive = efx->irq_rx_adaptive;
1935 	*rx_usecs = efx->irq_rx_moderation_us;
1936 
1937 	/* If channels are shared between RX and TX, so is IRQ
1938 	 * moderation.  Otherwise, IRQ moderation is the same for all
1939 	 * TX channels and is not adaptive.
1940 	 */
1941 	if (efx->tx_channel_offset == 0) {
1942 		*tx_usecs = *rx_usecs;
1943 	} else {
1944 		struct ef4_channel *tx_channel;
1945 
1946 		tx_channel = efx->channel[efx->tx_channel_offset];
1947 		*tx_usecs = tx_channel->irq_moderation_us;
1948 	}
1949 }
1950 
1951 /**************************************************************************
1952  *
1953  * Hardware monitor
1954  *
1955  **************************************************************************/
1956 
1957 /* Run periodically off the general workqueue */
1958 static void ef4_monitor(struct work_struct *data)
1959 {
1960 	struct ef4_nic *efx = container_of(data, struct ef4_nic,
1961 					   monitor_work.work);
1962 
1963 	netif_vdbg(efx, timer, efx->net_dev,
1964 		   "hardware monitor executing on CPU %d\n",
1965 		   raw_smp_processor_id());
1966 	BUG_ON(efx->type->monitor == NULL);
1967 
1968 	/* If the mac_lock is already held then it is likely a port
1969 	 * reconfiguration is already in place, which will likely do
1970 	 * most of the work of monitor() anyway. */
1971 	if (mutex_trylock(&efx->mac_lock)) {
1972 		if (efx->port_enabled)
1973 			efx->type->monitor(efx);
1974 		mutex_unlock(&efx->mac_lock);
1975 	}
1976 
1977 	queue_delayed_work(efx->workqueue, &efx->monitor_work,
1978 			   ef4_monitor_interval);
1979 }
1980 
1981 /**************************************************************************
1982  *
1983  * ioctls
1984  *
1985  *************************************************************************/
1986 
1987 /* Net device ioctl
1988  * Context: process, rtnl_lock() held.
1989  */
1990 static int ef4_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1991 {
1992 	struct ef4_nic *efx = netdev_priv(net_dev);
1993 	struct mii_ioctl_data *data = if_mii(ifr);
1994 
1995 	/* Convert phy_id from older PRTAD/DEVAD format */
1996 	if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1997 	    (data->phy_id & 0xfc00) == 0x0400)
1998 		data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1999 
2000 	return mdio_mii_ioctl(&efx->mdio, data, cmd);
2001 }
2002 
2003 /**************************************************************************
2004  *
2005  * NAPI interface
2006  *
2007  **************************************************************************/
2008 
2009 static void ef4_init_napi_channel(struct ef4_channel *channel)
2010 {
2011 	struct ef4_nic *efx = channel->efx;
2012 
2013 	channel->napi_dev = efx->net_dev;
2014 	netif_napi_add(channel->napi_dev, &channel->napi_str, ef4_poll);
2015 }
2016 
2017 static void ef4_init_napi(struct ef4_nic *efx)
2018 {
2019 	struct ef4_channel *channel;
2020 
2021 	ef4_for_each_channel(channel, efx)
2022 		ef4_init_napi_channel(channel);
2023 }
2024 
2025 static void ef4_fini_napi_channel(struct ef4_channel *channel)
2026 {
2027 	if (channel->napi_dev)
2028 		netif_napi_del(&channel->napi_str);
2029 
2030 	channel->napi_dev = NULL;
2031 }
2032 
2033 static void ef4_fini_napi(struct ef4_nic *efx)
2034 {
2035 	struct ef4_channel *channel;
2036 
2037 	ef4_for_each_channel(channel, efx)
2038 		ef4_fini_napi_channel(channel);
2039 }
2040 
2041 /**************************************************************************
2042  *
2043  * Kernel net device interface
2044  *
2045  *************************************************************************/
2046 
2047 /* Context: process, rtnl_lock() held. */
2048 int ef4_net_open(struct net_device *net_dev)
2049 {
2050 	struct ef4_nic *efx = netdev_priv(net_dev);
2051 	int rc;
2052 
2053 	netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2054 		  raw_smp_processor_id());
2055 
2056 	rc = ef4_check_disabled(efx);
2057 	if (rc)
2058 		return rc;
2059 	if (efx->phy_mode & PHY_MODE_SPECIAL)
2060 		return -EBUSY;
2061 
2062 	/* Notify the kernel of the link state polled during driver load,
2063 	 * before the monitor starts running */
2064 	ef4_link_status_changed(efx);
2065 
2066 	ef4_start_all(efx);
2067 	ef4_selftest_async_start(efx);
2068 	return 0;
2069 }
2070 
2071 /* Context: process, rtnl_lock() held.
2072  * Note that the kernel will ignore our return code; this method
2073  * should really be a void.
2074  */
2075 int ef4_net_stop(struct net_device *net_dev)
2076 {
2077 	struct ef4_nic *efx = netdev_priv(net_dev);
2078 
2079 	netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2080 		  raw_smp_processor_id());
2081 
2082 	/* Stop the device and flush all the channels */
2083 	ef4_stop_all(efx);
2084 
2085 	return 0;
2086 }
2087 
2088 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2089 static void ef4_net_stats(struct net_device *net_dev,
2090 			  struct rtnl_link_stats64 *stats)
2091 {
2092 	struct ef4_nic *efx = netdev_priv(net_dev);
2093 
2094 	spin_lock_bh(&efx->stats_lock);
2095 	efx->type->update_stats(efx, NULL, stats);
2096 	spin_unlock_bh(&efx->stats_lock);
2097 }
2098 
2099 /* Context: netif_tx_lock held, BHs disabled. */
2100 static void ef4_watchdog(struct net_device *net_dev, unsigned int txqueue)
2101 {
2102 	struct ef4_nic *efx = netdev_priv(net_dev);
2103 
2104 	netif_err(efx, tx_err, efx->net_dev,
2105 		  "TX stuck with port_enabled=%d: resetting channels\n",
2106 		  efx->port_enabled);
2107 
2108 	ef4_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2109 }
2110 
2111 
2112 /* Context: process, rtnl_lock() held. */
2113 static int ef4_change_mtu(struct net_device *net_dev, int new_mtu)
2114 {
2115 	struct ef4_nic *efx = netdev_priv(net_dev);
2116 	int rc;
2117 
2118 	rc = ef4_check_disabled(efx);
2119 	if (rc)
2120 		return rc;
2121 
2122 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2123 
2124 	ef4_device_detach_sync(efx);
2125 	ef4_stop_all(efx);
2126 
2127 	mutex_lock(&efx->mac_lock);
2128 	net_dev->mtu = new_mtu;
2129 	ef4_mac_reconfigure(efx);
2130 	mutex_unlock(&efx->mac_lock);
2131 
2132 	ef4_start_all(efx);
2133 	netif_device_attach(efx->net_dev);
2134 	return 0;
2135 }
2136 
2137 static int ef4_set_mac_address(struct net_device *net_dev, void *data)
2138 {
2139 	struct ef4_nic *efx = netdev_priv(net_dev);
2140 	struct sockaddr *addr = data;
2141 	u8 *new_addr = addr->sa_data;
2142 	u8 old_addr[6];
2143 	int rc;
2144 
2145 	if (!is_valid_ether_addr(new_addr)) {
2146 		netif_err(efx, drv, efx->net_dev,
2147 			  "invalid ethernet MAC address requested: %pM\n",
2148 			  new_addr);
2149 		return -EADDRNOTAVAIL;
2150 	}
2151 
2152 	/* save old address */
2153 	ether_addr_copy(old_addr, net_dev->dev_addr);
2154 	eth_hw_addr_set(net_dev, new_addr);
2155 	if (efx->type->set_mac_address) {
2156 		rc = efx->type->set_mac_address(efx);
2157 		if (rc) {
2158 			eth_hw_addr_set(net_dev, old_addr);
2159 			return rc;
2160 		}
2161 	}
2162 
2163 	/* Reconfigure the MAC */
2164 	mutex_lock(&efx->mac_lock);
2165 	ef4_mac_reconfigure(efx);
2166 	mutex_unlock(&efx->mac_lock);
2167 
2168 	return 0;
2169 }
2170 
2171 /* Context: netif_addr_lock held, BHs disabled. */
2172 static void ef4_set_rx_mode(struct net_device *net_dev)
2173 {
2174 	struct ef4_nic *efx = netdev_priv(net_dev);
2175 
2176 	if (efx->port_enabled)
2177 		queue_work(efx->workqueue, &efx->mac_work);
2178 	/* Otherwise ef4_start_port() will do this */
2179 }
2180 
2181 static int ef4_set_features(struct net_device *net_dev, netdev_features_t data)
2182 {
2183 	struct ef4_nic *efx = netdev_priv(net_dev);
2184 	int rc;
2185 
2186 	/* If disabling RX n-tuple filtering, clear existing filters */
2187 	if (net_dev->features & ~data & NETIF_F_NTUPLE) {
2188 		rc = efx->type->filter_clear_rx(efx, EF4_FILTER_PRI_MANUAL);
2189 		if (rc)
2190 			return rc;
2191 	}
2192 
2193 	/* If Rx VLAN filter is changed, update filters via mac_reconfigure */
2194 	if ((net_dev->features ^ data) & NETIF_F_HW_VLAN_CTAG_FILTER) {
2195 		/* ef4_set_rx_mode() will schedule MAC work to update filters
2196 		 * when a new features are finally set in net_dev.
2197 		 */
2198 		ef4_set_rx_mode(net_dev);
2199 	}
2200 
2201 	return 0;
2202 }
2203 
2204 static const struct net_device_ops ef4_netdev_ops = {
2205 	.ndo_open		= ef4_net_open,
2206 	.ndo_stop		= ef4_net_stop,
2207 	.ndo_get_stats64	= ef4_net_stats,
2208 	.ndo_tx_timeout		= ef4_watchdog,
2209 	.ndo_start_xmit		= ef4_hard_start_xmit,
2210 	.ndo_validate_addr	= eth_validate_addr,
2211 	.ndo_eth_ioctl		= ef4_ioctl,
2212 	.ndo_change_mtu		= ef4_change_mtu,
2213 	.ndo_set_mac_address	= ef4_set_mac_address,
2214 	.ndo_set_rx_mode	= ef4_set_rx_mode,
2215 	.ndo_set_features	= ef4_set_features,
2216 	.ndo_setup_tc		= ef4_setup_tc,
2217 #ifdef CONFIG_RFS_ACCEL
2218 	.ndo_rx_flow_steer	= ef4_filter_rfs,
2219 #endif
2220 };
2221 
2222 static void ef4_update_name(struct ef4_nic *efx)
2223 {
2224 	strcpy(efx->name, efx->net_dev->name);
2225 	ef4_mtd_rename(efx);
2226 	ef4_set_channel_names(efx);
2227 }
2228 
2229 static int ef4_netdev_event(struct notifier_block *this,
2230 			    unsigned long event, void *ptr)
2231 {
2232 	struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2233 
2234 	if ((net_dev->netdev_ops == &ef4_netdev_ops) &&
2235 	    event == NETDEV_CHANGENAME)
2236 		ef4_update_name(netdev_priv(net_dev));
2237 
2238 	return NOTIFY_DONE;
2239 }
2240 
2241 static struct notifier_block ef4_netdev_notifier = {
2242 	.notifier_call = ef4_netdev_event,
2243 };
2244 
2245 static ssize_t
2246 phy_type_show(struct device *dev, struct device_attribute *attr, char *buf)
2247 {
2248 	struct ef4_nic *efx = dev_get_drvdata(dev);
2249 	return sprintf(buf, "%d\n", efx->phy_type);
2250 }
2251 static DEVICE_ATTR_RO(phy_type);
2252 
2253 static int ef4_register_netdev(struct ef4_nic *efx)
2254 {
2255 	struct net_device *net_dev = efx->net_dev;
2256 	struct ef4_channel *channel;
2257 	int rc;
2258 
2259 	net_dev->watchdog_timeo = 5 * HZ;
2260 	net_dev->irq = efx->pci_dev->irq;
2261 	net_dev->netdev_ops = &ef4_netdev_ops;
2262 	net_dev->ethtool_ops = &ef4_ethtool_ops;
2263 	netif_set_tso_max_segs(net_dev, EF4_TSO_MAX_SEGS);
2264 	net_dev->min_mtu = EF4_MIN_MTU;
2265 	net_dev->max_mtu = EF4_MAX_MTU;
2266 
2267 	rtnl_lock();
2268 
2269 	/* Enable resets to be scheduled and check whether any were
2270 	 * already requested.  If so, the NIC is probably hosed so we
2271 	 * abort.
2272 	 */
2273 	efx->state = STATE_READY;
2274 	smp_mb(); /* ensure we change state before checking reset_pending */
2275 	if (efx->reset_pending) {
2276 		netif_err(efx, probe, efx->net_dev,
2277 			  "aborting probe due to scheduled reset\n");
2278 		rc = -EIO;
2279 		goto fail_locked;
2280 	}
2281 
2282 	rc = dev_alloc_name(net_dev, net_dev->name);
2283 	if (rc < 0)
2284 		goto fail_locked;
2285 	ef4_update_name(efx);
2286 
2287 	/* Always start with carrier off; PHY events will detect the link */
2288 	netif_carrier_off(net_dev);
2289 
2290 	rc = register_netdevice(net_dev);
2291 	if (rc)
2292 		goto fail_locked;
2293 
2294 	ef4_for_each_channel(channel, efx) {
2295 		struct ef4_tx_queue *tx_queue;
2296 		ef4_for_each_channel_tx_queue(tx_queue, channel)
2297 			ef4_init_tx_queue_core_txq(tx_queue);
2298 	}
2299 
2300 	ef4_associate(efx);
2301 
2302 	rtnl_unlock();
2303 
2304 	rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2305 	if (rc) {
2306 		netif_err(efx, drv, efx->net_dev,
2307 			  "failed to init net dev attributes\n");
2308 		goto fail_registered;
2309 	}
2310 	return 0;
2311 
2312 fail_registered:
2313 	rtnl_lock();
2314 	ef4_dissociate(efx);
2315 	unregister_netdevice(net_dev);
2316 fail_locked:
2317 	efx->state = STATE_UNINIT;
2318 	rtnl_unlock();
2319 	netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2320 	return rc;
2321 }
2322 
2323 static void ef4_unregister_netdev(struct ef4_nic *efx)
2324 {
2325 	if (!efx->net_dev)
2326 		return;
2327 
2328 	BUG_ON(netdev_priv(efx->net_dev) != efx);
2329 
2330 	if (ef4_dev_registered(efx)) {
2331 		strscpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2332 		device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2333 		unregister_netdev(efx->net_dev);
2334 	}
2335 }
2336 
2337 /**************************************************************************
2338  *
2339  * Device reset and suspend
2340  *
2341  **************************************************************************/
2342 
2343 /* Tears down the entire software state and most of the hardware state
2344  * before reset.  */
2345 void ef4_reset_down(struct ef4_nic *efx, enum reset_type method)
2346 {
2347 	EF4_ASSERT_RESET_SERIALISED(efx);
2348 
2349 	ef4_stop_all(efx);
2350 	ef4_disable_interrupts(efx);
2351 
2352 	mutex_lock(&efx->mac_lock);
2353 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2354 	    method != RESET_TYPE_DATAPATH)
2355 		efx->phy_op->fini(efx);
2356 	efx->type->fini(efx);
2357 }
2358 
2359 /* This function will always ensure that the locks acquired in
2360  * ef4_reset_down() are released. A failure return code indicates
2361  * that we were unable to reinitialise the hardware, and the
2362  * driver should be disabled. If ok is false, then the rx and tx
2363  * engines are not restarted, pending a RESET_DISABLE. */
2364 int ef4_reset_up(struct ef4_nic *efx, enum reset_type method, bool ok)
2365 {
2366 	int rc;
2367 
2368 	EF4_ASSERT_RESET_SERIALISED(efx);
2369 
2370 	/* Ensure that SRAM is initialised even if we're disabling the device */
2371 	rc = efx->type->init(efx);
2372 	if (rc) {
2373 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2374 		goto fail;
2375 	}
2376 
2377 	if (!ok)
2378 		goto fail;
2379 
2380 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2381 	    method != RESET_TYPE_DATAPATH) {
2382 		rc = efx->phy_op->init(efx);
2383 		if (rc)
2384 			goto fail;
2385 		rc = efx->phy_op->reconfigure(efx);
2386 		if (rc && rc != -EPERM)
2387 			netif_err(efx, drv, efx->net_dev,
2388 				  "could not restore PHY settings\n");
2389 	}
2390 
2391 	rc = ef4_enable_interrupts(efx);
2392 	if (rc)
2393 		goto fail;
2394 
2395 	down_read(&efx->filter_sem);
2396 	ef4_restore_filters(efx);
2397 	up_read(&efx->filter_sem);
2398 
2399 	mutex_unlock(&efx->mac_lock);
2400 
2401 	ef4_start_all(efx);
2402 
2403 	return 0;
2404 
2405 fail:
2406 	efx->port_initialized = false;
2407 
2408 	mutex_unlock(&efx->mac_lock);
2409 
2410 	return rc;
2411 }
2412 
2413 /* Reset the NIC using the specified method.  Note that the reset may
2414  * fail, in which case the card will be left in an unusable state.
2415  *
2416  * Caller must hold the rtnl_lock.
2417  */
2418 int ef4_reset(struct ef4_nic *efx, enum reset_type method)
2419 {
2420 	int rc, rc2;
2421 	bool disabled;
2422 
2423 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2424 		   RESET_TYPE(method));
2425 
2426 	ef4_device_detach_sync(efx);
2427 	ef4_reset_down(efx, method);
2428 
2429 	rc = efx->type->reset(efx, method);
2430 	if (rc) {
2431 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2432 		goto out;
2433 	}
2434 
2435 	/* Clear flags for the scopes we covered.  We assume the NIC and
2436 	 * driver are now quiescent so that there is no race here.
2437 	 */
2438 	if (method < RESET_TYPE_MAX_METHOD)
2439 		efx->reset_pending &= -(1 << (method + 1));
2440 	else /* it doesn't fit into the well-ordered scope hierarchy */
2441 		__clear_bit(method, &efx->reset_pending);
2442 
2443 	/* Reinitialise bus-mastering, which may have been turned off before
2444 	 * the reset was scheduled. This is still appropriate, even in the
2445 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2446 	 * can respond to requests. */
2447 	pci_set_master(efx->pci_dev);
2448 
2449 out:
2450 	/* Leave device stopped if necessary */
2451 	disabled = rc ||
2452 		method == RESET_TYPE_DISABLE ||
2453 		method == RESET_TYPE_RECOVER_OR_DISABLE;
2454 	rc2 = ef4_reset_up(efx, method, !disabled);
2455 	if (rc2) {
2456 		disabled = true;
2457 		if (!rc)
2458 			rc = rc2;
2459 	}
2460 
2461 	if (disabled) {
2462 		dev_close(efx->net_dev);
2463 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2464 		efx->state = STATE_DISABLED;
2465 	} else {
2466 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2467 		netif_device_attach(efx->net_dev);
2468 	}
2469 	return rc;
2470 }
2471 
2472 /* Try recovery mechanisms.
2473  * For now only EEH is supported.
2474  * Returns 0 if the recovery mechanisms are unsuccessful.
2475  * Returns a non-zero value otherwise.
2476  */
2477 int ef4_try_recovery(struct ef4_nic *efx)
2478 {
2479 #ifdef CONFIG_EEH
2480 	/* A PCI error can occur and not be seen by EEH because nothing
2481 	 * happens on the PCI bus. In this case the driver may fail and
2482 	 * schedule a 'recover or reset', leading to this recovery handler.
2483 	 * Manually call the eeh failure check function.
2484 	 */
2485 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2486 	if (eeh_dev_check_failure(eehdev)) {
2487 		/* The EEH mechanisms will handle the error and reset the
2488 		 * device if necessary.
2489 		 */
2490 		return 1;
2491 	}
2492 #endif
2493 	return 0;
2494 }
2495 
2496 /* The worker thread exists so that code that cannot sleep can
2497  * schedule a reset for later.
2498  */
2499 static void ef4_reset_work(struct work_struct *data)
2500 {
2501 	struct ef4_nic *efx = container_of(data, struct ef4_nic, reset_work);
2502 	unsigned long pending;
2503 	enum reset_type method;
2504 
2505 	pending = READ_ONCE(efx->reset_pending);
2506 	method = fls(pending) - 1;
2507 
2508 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2509 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
2510 	    ef4_try_recovery(efx))
2511 		return;
2512 
2513 	if (!pending)
2514 		return;
2515 
2516 	rtnl_lock();
2517 
2518 	/* We checked the state in ef4_schedule_reset() but it may
2519 	 * have changed by now.  Now that we have the RTNL lock,
2520 	 * it cannot change again.
2521 	 */
2522 	if (efx->state == STATE_READY)
2523 		(void)ef4_reset(efx, method);
2524 
2525 	rtnl_unlock();
2526 }
2527 
2528 void ef4_schedule_reset(struct ef4_nic *efx, enum reset_type type)
2529 {
2530 	enum reset_type method;
2531 
2532 	if (efx->state == STATE_RECOVERY) {
2533 		netif_dbg(efx, drv, efx->net_dev,
2534 			  "recovering: skip scheduling %s reset\n",
2535 			  RESET_TYPE(type));
2536 		return;
2537 	}
2538 
2539 	switch (type) {
2540 	case RESET_TYPE_INVISIBLE:
2541 	case RESET_TYPE_ALL:
2542 	case RESET_TYPE_RECOVER_OR_ALL:
2543 	case RESET_TYPE_WORLD:
2544 	case RESET_TYPE_DISABLE:
2545 	case RESET_TYPE_RECOVER_OR_DISABLE:
2546 	case RESET_TYPE_DATAPATH:
2547 		method = type;
2548 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2549 			  RESET_TYPE(method));
2550 		break;
2551 	default:
2552 		method = efx->type->map_reset_reason(type);
2553 		netif_dbg(efx, drv, efx->net_dev,
2554 			  "scheduling %s reset for %s\n",
2555 			  RESET_TYPE(method), RESET_TYPE(type));
2556 		break;
2557 	}
2558 
2559 	set_bit(method, &efx->reset_pending);
2560 	smp_mb(); /* ensure we change reset_pending before checking state */
2561 
2562 	/* If we're not READY then just leave the flags set as the cue
2563 	 * to abort probing or reschedule the reset later.
2564 	 */
2565 	if (READ_ONCE(efx->state) != STATE_READY)
2566 		return;
2567 
2568 	queue_work(reset_workqueue, &efx->reset_work);
2569 }
2570 
2571 /**************************************************************************
2572  *
2573  * List of NICs we support
2574  *
2575  **************************************************************************/
2576 
2577 /* PCI device ID table */
2578 static const struct pci_device_id ef4_pci_table[] = {
2579 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2580 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2581 	 .driver_data = (unsigned long) &falcon_a1_nic_type},
2582 	{PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2583 		    PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2584 	 .driver_data = (unsigned long) &falcon_b0_nic_type},
2585 	{0}			/* end of list */
2586 };
2587 
2588 /**************************************************************************
2589  *
2590  * Dummy PHY/MAC operations
2591  *
2592  * Can be used for some unimplemented operations
2593  * Needed so all function pointers are valid and do not have to be tested
2594  * before use
2595  *
2596  **************************************************************************/
2597 int ef4_port_dummy_op_int(struct ef4_nic *efx)
2598 {
2599 	return 0;
2600 }
2601 void ef4_port_dummy_op_void(struct ef4_nic *efx) {}
2602 
2603 static bool ef4_port_dummy_op_poll(struct ef4_nic *efx)
2604 {
2605 	return false;
2606 }
2607 
2608 static const struct ef4_phy_operations ef4_dummy_phy_operations = {
2609 	.init		 = ef4_port_dummy_op_int,
2610 	.reconfigure	 = ef4_port_dummy_op_int,
2611 	.poll		 = ef4_port_dummy_op_poll,
2612 	.fini		 = ef4_port_dummy_op_void,
2613 };
2614 
2615 /**************************************************************************
2616  *
2617  * Data housekeeping
2618  *
2619  **************************************************************************/
2620 
2621 /* This zeroes out and then fills in the invariants in a struct
2622  * ef4_nic (including all sub-structures).
2623  */
2624 static int ef4_init_struct(struct ef4_nic *efx,
2625 			   struct pci_dev *pci_dev, struct net_device *net_dev)
2626 {
2627 	int i;
2628 
2629 	/* Initialise common structures */
2630 	INIT_LIST_HEAD(&efx->node);
2631 	INIT_LIST_HEAD(&efx->secondary_list);
2632 	spin_lock_init(&efx->biu_lock);
2633 #ifdef CONFIG_SFC_FALCON_MTD
2634 	INIT_LIST_HEAD(&efx->mtd_list);
2635 #endif
2636 	INIT_WORK(&efx->reset_work, ef4_reset_work);
2637 	INIT_DELAYED_WORK(&efx->monitor_work, ef4_monitor);
2638 	INIT_DELAYED_WORK(&efx->selftest_work, ef4_selftest_async_work);
2639 	efx->pci_dev = pci_dev;
2640 	efx->msg_enable = debug;
2641 	efx->state = STATE_UNINIT;
2642 	strscpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2643 
2644 	efx->net_dev = net_dev;
2645 	efx->rx_prefix_size = efx->type->rx_prefix_size;
2646 	efx->rx_ip_align =
2647 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2648 	efx->rx_packet_hash_offset =
2649 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2650 	efx->rx_packet_ts_offset =
2651 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2652 	spin_lock_init(&efx->stats_lock);
2653 	mutex_init(&efx->mac_lock);
2654 	efx->phy_op = &ef4_dummy_phy_operations;
2655 	efx->mdio.dev = net_dev;
2656 	INIT_WORK(&efx->mac_work, ef4_mac_work);
2657 	init_waitqueue_head(&efx->flush_wq);
2658 
2659 	for (i = 0; i < EF4_MAX_CHANNELS; i++) {
2660 		efx->channel[i] = ef4_alloc_channel(efx, i, NULL);
2661 		if (!efx->channel[i])
2662 			goto fail;
2663 		efx->msi_context[i].efx = efx;
2664 		efx->msi_context[i].index = i;
2665 	}
2666 
2667 	/* Higher numbered interrupt modes are less capable! */
2668 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2669 				  interrupt_mode);
2670 
2671 	/* Would be good to use the net_dev name, but we're too early */
2672 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2673 		 pci_name(pci_dev));
2674 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2675 	if (!efx->workqueue)
2676 		goto fail;
2677 
2678 	return 0;
2679 
2680 fail:
2681 	ef4_fini_struct(efx);
2682 	return -ENOMEM;
2683 }
2684 
2685 static void ef4_fini_struct(struct ef4_nic *efx)
2686 {
2687 	int i;
2688 
2689 	for (i = 0; i < EF4_MAX_CHANNELS; i++)
2690 		kfree(efx->channel[i]);
2691 
2692 	kfree(efx->vpd_sn);
2693 
2694 	if (efx->workqueue) {
2695 		destroy_workqueue(efx->workqueue);
2696 		efx->workqueue = NULL;
2697 	}
2698 }
2699 
2700 void ef4_update_sw_stats(struct ef4_nic *efx, u64 *stats)
2701 {
2702 	u64 n_rx_nodesc_trunc = 0;
2703 	struct ef4_channel *channel;
2704 
2705 	ef4_for_each_channel(channel, efx)
2706 		n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
2707 	stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
2708 	stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
2709 }
2710 
2711 /**************************************************************************
2712  *
2713  * PCI interface
2714  *
2715  **************************************************************************/
2716 
2717 /* Main body of final NIC shutdown code
2718  * This is called only at module unload (or hotplug removal).
2719  */
2720 static void ef4_pci_remove_main(struct ef4_nic *efx)
2721 {
2722 	/* Flush reset_work. It can no longer be scheduled since we
2723 	 * are not READY.
2724 	 */
2725 	BUG_ON(efx->state == STATE_READY);
2726 	cancel_work_sync(&efx->reset_work);
2727 
2728 	ef4_disable_interrupts(efx);
2729 	ef4_nic_fini_interrupt(efx);
2730 	ef4_fini_port(efx);
2731 	efx->type->fini(efx);
2732 	ef4_fini_napi(efx);
2733 	ef4_remove_all(efx);
2734 }
2735 
2736 /* Final NIC shutdown
2737  * This is called only at module unload (or hotplug removal).  A PF can call
2738  * this on its VFs to ensure they are unbound first.
2739  */
2740 static void ef4_pci_remove(struct pci_dev *pci_dev)
2741 {
2742 	struct ef4_nic *efx;
2743 
2744 	efx = pci_get_drvdata(pci_dev);
2745 	if (!efx)
2746 		return;
2747 
2748 	/* Mark the NIC as fini, then stop the interface */
2749 	rtnl_lock();
2750 	ef4_dissociate(efx);
2751 	dev_close(efx->net_dev);
2752 	ef4_disable_interrupts(efx);
2753 	efx->state = STATE_UNINIT;
2754 	rtnl_unlock();
2755 
2756 	ef4_unregister_netdev(efx);
2757 
2758 	ef4_mtd_remove(efx);
2759 
2760 	ef4_pci_remove_main(efx);
2761 
2762 	ef4_fini_io(efx);
2763 	netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2764 
2765 	ef4_fini_struct(efx);
2766 	free_netdev(efx->net_dev);
2767 };
2768 
2769 /* NIC VPD information
2770  * Called during probe to display the part number of the installed NIC.
2771  */
2772 static void ef4_probe_vpd_strings(struct ef4_nic *efx)
2773 {
2774 	struct pci_dev *dev = efx->pci_dev;
2775 	unsigned int vpd_size, kw_len;
2776 	u8 *vpd_data;
2777 	int start;
2778 
2779 	vpd_data = pci_vpd_alloc(dev, &vpd_size);
2780 	if (IS_ERR(vpd_data)) {
2781 		pci_warn(dev, "Unable to read VPD\n");
2782 		return;
2783 	}
2784 
2785 	start = pci_vpd_find_ro_info_keyword(vpd_data, vpd_size,
2786 					     PCI_VPD_RO_KEYWORD_PARTNO, &kw_len);
2787 	if (start < 0)
2788 		pci_warn(dev, "Part number not found or incomplete\n");
2789 	else
2790 		pci_info(dev, "Part Number : %.*s\n", kw_len, vpd_data + start);
2791 
2792 	start = pci_vpd_find_ro_info_keyword(vpd_data, vpd_size,
2793 					     PCI_VPD_RO_KEYWORD_SERIALNO, &kw_len);
2794 	if (start < 0)
2795 		pci_warn(dev, "Serial number not found or incomplete\n");
2796 	else
2797 		efx->vpd_sn = kmemdup_nul(vpd_data + start, kw_len, GFP_KERNEL);
2798 
2799 	kfree(vpd_data);
2800 }
2801 
2802 
2803 /* Main body of NIC initialisation
2804  * This is called at module load (or hotplug insertion, theoretically).
2805  */
2806 static int ef4_pci_probe_main(struct ef4_nic *efx)
2807 {
2808 	int rc;
2809 
2810 	/* Do start-of-day initialisation */
2811 	rc = ef4_probe_all(efx);
2812 	if (rc)
2813 		goto fail1;
2814 
2815 	ef4_init_napi(efx);
2816 
2817 	rc = efx->type->init(efx);
2818 	if (rc) {
2819 		netif_err(efx, probe, efx->net_dev,
2820 			  "failed to initialise NIC\n");
2821 		goto fail3;
2822 	}
2823 
2824 	rc = ef4_init_port(efx);
2825 	if (rc) {
2826 		netif_err(efx, probe, efx->net_dev,
2827 			  "failed to initialise port\n");
2828 		goto fail4;
2829 	}
2830 
2831 	rc = ef4_nic_init_interrupt(efx);
2832 	if (rc)
2833 		goto fail5;
2834 	rc = ef4_enable_interrupts(efx);
2835 	if (rc)
2836 		goto fail6;
2837 
2838 	return 0;
2839 
2840  fail6:
2841 	ef4_nic_fini_interrupt(efx);
2842  fail5:
2843 	ef4_fini_port(efx);
2844  fail4:
2845 	efx->type->fini(efx);
2846  fail3:
2847 	ef4_fini_napi(efx);
2848 	ef4_remove_all(efx);
2849  fail1:
2850 	return rc;
2851 }
2852 
2853 /* NIC initialisation
2854  *
2855  * This is called at module load (or hotplug insertion,
2856  * theoretically).  It sets up PCI mappings, resets the NIC,
2857  * sets up and registers the network devices with the kernel and hooks
2858  * the interrupt service routine.  It does not prepare the device for
2859  * transmission; this is left to the first time one of the network
2860  * interfaces is brought up (i.e. ef4_net_open).
2861  */
2862 static int ef4_pci_probe(struct pci_dev *pci_dev,
2863 			 const struct pci_device_id *entry)
2864 {
2865 	struct net_device *net_dev;
2866 	struct ef4_nic *efx;
2867 	int rc;
2868 
2869 	/* Allocate and initialise a struct net_device and struct ef4_nic */
2870 	net_dev = alloc_etherdev_mqs(sizeof(*efx), EF4_MAX_CORE_TX_QUEUES,
2871 				     EF4_MAX_RX_QUEUES);
2872 	if (!net_dev)
2873 		return -ENOMEM;
2874 	efx = netdev_priv(net_dev);
2875 	efx->type = (const struct ef4_nic_type *) entry->driver_data;
2876 	efx->fixed_features |= NETIF_F_HIGHDMA;
2877 
2878 	pci_set_drvdata(pci_dev, efx);
2879 	SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2880 	rc = ef4_init_struct(efx, pci_dev, net_dev);
2881 	if (rc)
2882 		goto fail1;
2883 
2884 	netif_info(efx, probe, efx->net_dev,
2885 		   "Solarflare NIC detected\n");
2886 
2887 	ef4_probe_vpd_strings(efx);
2888 
2889 	/* Set up basic I/O (BAR mappings etc) */
2890 	rc = ef4_init_io(efx);
2891 	if (rc)
2892 		goto fail2;
2893 
2894 	rc = ef4_pci_probe_main(efx);
2895 	if (rc)
2896 		goto fail3;
2897 
2898 	net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2899 			      NETIF_F_RXCSUM);
2900 	/* Mask for features that also apply to VLAN devices */
2901 	net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
2902 				   NETIF_F_HIGHDMA | NETIF_F_RXCSUM);
2903 
2904 	net_dev->hw_features = net_dev->features & ~efx->fixed_features;
2905 
2906 	/* Disable VLAN filtering by default.  It may be enforced if
2907 	 * the feature is fixed (i.e. VLAN filters are required to
2908 	 * receive VLAN tagged packets due to vPort restrictions).
2909 	 */
2910 	net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
2911 	net_dev->features |= efx->fixed_features;
2912 
2913 	rc = ef4_register_netdev(efx);
2914 	if (rc)
2915 		goto fail4;
2916 
2917 	netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2918 
2919 	/* Try to create MTDs, but allow this to fail */
2920 	rtnl_lock();
2921 	rc = ef4_mtd_probe(efx);
2922 	rtnl_unlock();
2923 	if (rc && rc != -EPERM)
2924 		netif_warn(efx, probe, efx->net_dev,
2925 			   "failed to create MTDs (%d)\n", rc);
2926 
2927 	return 0;
2928 
2929  fail4:
2930 	ef4_pci_remove_main(efx);
2931  fail3:
2932 	ef4_fini_io(efx);
2933  fail2:
2934 	ef4_fini_struct(efx);
2935  fail1:
2936 	WARN_ON(rc > 0);
2937 	netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2938 	free_netdev(net_dev);
2939 	return rc;
2940 }
2941 
2942 static int ef4_pm_freeze(struct device *dev)
2943 {
2944 	struct ef4_nic *efx = dev_get_drvdata(dev);
2945 
2946 	rtnl_lock();
2947 
2948 	if (efx->state != STATE_DISABLED) {
2949 		efx->state = STATE_UNINIT;
2950 
2951 		ef4_device_detach_sync(efx);
2952 
2953 		ef4_stop_all(efx);
2954 		ef4_disable_interrupts(efx);
2955 	}
2956 
2957 	rtnl_unlock();
2958 
2959 	return 0;
2960 }
2961 
2962 static int ef4_pm_thaw(struct device *dev)
2963 {
2964 	int rc;
2965 	struct ef4_nic *efx = dev_get_drvdata(dev);
2966 
2967 	rtnl_lock();
2968 
2969 	if (efx->state != STATE_DISABLED) {
2970 		rc = ef4_enable_interrupts(efx);
2971 		if (rc)
2972 			goto fail;
2973 
2974 		mutex_lock(&efx->mac_lock);
2975 		efx->phy_op->reconfigure(efx);
2976 		mutex_unlock(&efx->mac_lock);
2977 
2978 		ef4_start_all(efx);
2979 
2980 		netif_device_attach(efx->net_dev);
2981 
2982 		efx->state = STATE_READY;
2983 
2984 		efx->type->resume_wol(efx);
2985 	}
2986 
2987 	rtnl_unlock();
2988 
2989 	/* Reschedule any quenched resets scheduled during ef4_pm_freeze() */
2990 	queue_work(reset_workqueue, &efx->reset_work);
2991 
2992 	return 0;
2993 
2994 fail:
2995 	rtnl_unlock();
2996 
2997 	return rc;
2998 }
2999 
3000 static int ef4_pm_poweroff(struct device *dev)
3001 {
3002 	struct pci_dev *pci_dev = to_pci_dev(dev);
3003 	struct ef4_nic *efx = pci_get_drvdata(pci_dev);
3004 
3005 	efx->type->fini(efx);
3006 
3007 	efx->reset_pending = 0;
3008 
3009 	pci_save_state(pci_dev);
3010 	return pci_set_power_state(pci_dev, PCI_D3hot);
3011 }
3012 
3013 /* Used for both resume and restore */
3014 static int ef4_pm_resume(struct device *dev)
3015 {
3016 	struct pci_dev *pci_dev = to_pci_dev(dev);
3017 	struct ef4_nic *efx = pci_get_drvdata(pci_dev);
3018 	int rc;
3019 
3020 	rc = pci_set_power_state(pci_dev, PCI_D0);
3021 	if (rc)
3022 		return rc;
3023 	pci_restore_state(pci_dev);
3024 	rc = pci_enable_device(pci_dev);
3025 	if (rc)
3026 		return rc;
3027 	pci_set_master(efx->pci_dev);
3028 	rc = efx->type->reset(efx, RESET_TYPE_ALL);
3029 	if (rc)
3030 		return rc;
3031 	rc = efx->type->init(efx);
3032 	if (rc)
3033 		return rc;
3034 	rc = ef4_pm_thaw(dev);
3035 	return rc;
3036 }
3037 
3038 static int ef4_pm_suspend(struct device *dev)
3039 {
3040 	int rc;
3041 
3042 	ef4_pm_freeze(dev);
3043 	rc = ef4_pm_poweroff(dev);
3044 	if (rc)
3045 		ef4_pm_resume(dev);
3046 	return rc;
3047 }
3048 
3049 static const struct dev_pm_ops ef4_pm_ops = {
3050 	.suspend	= ef4_pm_suspend,
3051 	.resume		= ef4_pm_resume,
3052 	.freeze		= ef4_pm_freeze,
3053 	.thaw		= ef4_pm_thaw,
3054 	.poweroff	= ef4_pm_poweroff,
3055 	.restore	= ef4_pm_resume,
3056 };
3057 
3058 /* A PCI error affecting this device was detected.
3059  * At this point MMIO and DMA may be disabled.
3060  * Stop the software path and request a slot reset.
3061  */
3062 static pci_ers_result_t ef4_io_error_detected(struct pci_dev *pdev,
3063 					      pci_channel_state_t state)
3064 {
3065 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3066 	struct ef4_nic *efx = pci_get_drvdata(pdev);
3067 
3068 	if (state == pci_channel_io_perm_failure)
3069 		return PCI_ERS_RESULT_DISCONNECT;
3070 
3071 	rtnl_lock();
3072 
3073 	if (efx->state != STATE_DISABLED) {
3074 		efx->state = STATE_RECOVERY;
3075 		efx->reset_pending = 0;
3076 
3077 		ef4_device_detach_sync(efx);
3078 
3079 		ef4_stop_all(efx);
3080 		ef4_disable_interrupts(efx);
3081 
3082 		status = PCI_ERS_RESULT_NEED_RESET;
3083 	} else {
3084 		/* If the interface is disabled we don't want to do anything
3085 		 * with it.
3086 		 */
3087 		status = PCI_ERS_RESULT_RECOVERED;
3088 	}
3089 
3090 	rtnl_unlock();
3091 
3092 	pci_disable_device(pdev);
3093 
3094 	return status;
3095 }
3096 
3097 /* Fake a successful reset, which will be performed later in ef4_io_resume. */
3098 static pci_ers_result_t ef4_io_slot_reset(struct pci_dev *pdev)
3099 {
3100 	struct ef4_nic *efx = pci_get_drvdata(pdev);
3101 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3102 
3103 	if (pci_enable_device(pdev)) {
3104 		netif_err(efx, hw, efx->net_dev,
3105 			  "Cannot re-enable PCI device after reset.\n");
3106 		status =  PCI_ERS_RESULT_DISCONNECT;
3107 	}
3108 
3109 	return status;
3110 }
3111 
3112 /* Perform the actual reset and resume I/O operations. */
3113 static void ef4_io_resume(struct pci_dev *pdev)
3114 {
3115 	struct ef4_nic *efx = pci_get_drvdata(pdev);
3116 	int rc;
3117 
3118 	rtnl_lock();
3119 
3120 	if (efx->state == STATE_DISABLED)
3121 		goto out;
3122 
3123 	rc = ef4_reset(efx, RESET_TYPE_ALL);
3124 	if (rc) {
3125 		netif_err(efx, hw, efx->net_dev,
3126 			  "ef4_reset failed after PCI error (%d)\n", rc);
3127 	} else {
3128 		efx->state = STATE_READY;
3129 		netif_dbg(efx, hw, efx->net_dev,
3130 			  "Done resetting and resuming IO after PCI error.\n");
3131 	}
3132 
3133 out:
3134 	rtnl_unlock();
3135 }
3136 
3137 /* For simplicity and reliability, we always require a slot reset and try to
3138  * reset the hardware when a pci error affecting the device is detected.
3139  * We leave both the link_reset and mmio_enabled callback unimplemented:
3140  * with our request for slot reset the mmio_enabled callback will never be
3141  * called, and the link_reset callback is not used by AER or EEH mechanisms.
3142  */
3143 static const struct pci_error_handlers ef4_err_handlers = {
3144 	.error_detected = ef4_io_error_detected,
3145 	.slot_reset	= ef4_io_slot_reset,
3146 	.resume		= ef4_io_resume,
3147 };
3148 
3149 static struct pci_driver ef4_pci_driver = {
3150 	.name		= KBUILD_MODNAME,
3151 	.id_table	= ef4_pci_table,
3152 	.probe		= ef4_pci_probe,
3153 	.remove		= ef4_pci_remove,
3154 	.driver.pm	= &ef4_pm_ops,
3155 	.err_handler	= &ef4_err_handlers,
3156 };
3157 
3158 /**************************************************************************
3159  *
3160  * Kernel module interface
3161  *
3162  *************************************************************************/
3163 
3164 module_param(interrupt_mode, uint, 0444);
3165 MODULE_PARM_DESC(interrupt_mode,
3166 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3167 
3168 static int __init ef4_init_module(void)
3169 {
3170 	int rc;
3171 
3172 	printk(KERN_INFO "Solarflare Falcon driver v" EF4_DRIVER_VERSION "\n");
3173 
3174 	rc = register_netdevice_notifier(&ef4_netdev_notifier);
3175 	if (rc)
3176 		goto err_notifier;
3177 
3178 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
3179 	if (!reset_workqueue) {
3180 		rc = -ENOMEM;
3181 		goto err_reset;
3182 	}
3183 
3184 	rc = pci_register_driver(&ef4_pci_driver);
3185 	if (rc < 0)
3186 		goto err_pci;
3187 
3188 	return 0;
3189 
3190  err_pci:
3191 	destroy_workqueue(reset_workqueue);
3192  err_reset:
3193 	unregister_netdevice_notifier(&ef4_netdev_notifier);
3194  err_notifier:
3195 	return rc;
3196 }
3197 
3198 static void __exit ef4_exit_module(void)
3199 {
3200 	printk(KERN_INFO "Solarflare Falcon driver unloading\n");
3201 
3202 	pci_unregister_driver(&ef4_pci_driver);
3203 	destroy_workqueue(reset_workqueue);
3204 	unregister_netdevice_notifier(&ef4_netdev_notifier);
3205 
3206 }
3207 
3208 module_init(ef4_init_module);
3209 module_exit(ef4_exit_module);
3210 
3211 MODULE_AUTHOR("Solarflare Communications and "
3212 	      "Michael Brown <mbrown@fensystems.co.uk>");
3213 MODULE_DESCRIPTION("Solarflare Falcon network driver");
3214 MODULE_LICENSE("GPL");
3215 MODULE_DEVICE_TABLE(pci, ef4_pci_table);
3216 MODULE_VERSION(EF4_DRIVER_VERSION);
3217