1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3  * Driver for Solarflare network controllers and boards
4  * Copyright 2018 Solarflare Communications Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published
8  * by the Free Software Foundation, incorporated herein by reference.
9  */
10 
11 #include "net_driver.h"
12 #include <linux/module.h>
13 #include <linux/netdevice.h>
14 #include <net/gre.h>
15 #include "efx_common.h"
16 #include "efx_channels.h"
17 #include "efx.h"
18 #include "mcdi.h"
19 #include "selftest.h"
20 #include "rx_common.h"
21 #include "tx_common.h"
22 #include "nic.h"
23 #include "mcdi_port_common.h"
24 #include "io.h"
25 #include "mcdi_pcol.h"
26 
27 static unsigned int debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
28 			     NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
29 			     NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
30 			     NETIF_MSG_TX_ERR | NETIF_MSG_HW);
31 module_param(debug, uint, 0);
32 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
33 
34 /* This is the time (in jiffies) between invocations of the hardware
35  * monitor.
36  * On Falcon-based NICs, this will:
37  * - Check the on-board hardware monitor;
38  * - Poll the link state and reconfigure the hardware as necessary.
39  * On Siena-based NICs for power systems with EEH support, this will give EEH a
40  * chance to start.
41  */
42 static unsigned int efx_monitor_interval = 1 * HZ;
43 
44 /* How often and how many times to poll for a reset while waiting for a
45  * BIST that another function started to complete.
46  */
47 #define BIST_WAIT_DELAY_MS	100
48 #define BIST_WAIT_DELAY_COUNT	100
49 
50 /* Default stats update time */
51 #define STATS_PERIOD_MS_DEFAULT 1000
52 
53 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
54 const char *const efx_reset_type_names[] = {
55 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
56 	[RESET_TYPE_ALL]                = "ALL",
57 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
58 	[RESET_TYPE_WORLD]              = "WORLD",
59 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
60 	[RESET_TYPE_DATAPATH]           = "DATAPATH",
61 	[RESET_TYPE_MC_BIST]		= "MC_BIST",
62 	[RESET_TYPE_DISABLE]            = "DISABLE",
63 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
64 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
65 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
66 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
67 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
68 	[RESET_TYPE_MCDI_TIMEOUT]	= "MCDI_TIMEOUT (FLR)",
69 };
70 
71 #define RESET_TYPE(type) \
72 	STRING_TABLE_LOOKUP(type, efx_reset_type)
73 
74 /* Loopback mode names (see LOOPBACK_MODE()) */
75 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
76 const char *const efx_loopback_mode_names[] = {
77 	[LOOPBACK_NONE]		= "NONE",
78 	[LOOPBACK_DATA]		= "DATAPATH",
79 	[LOOPBACK_GMAC]		= "GMAC",
80 	[LOOPBACK_XGMII]	= "XGMII",
81 	[LOOPBACK_XGXS]		= "XGXS",
82 	[LOOPBACK_XAUI]		= "XAUI",
83 	[LOOPBACK_GMII]		= "GMII",
84 	[LOOPBACK_SGMII]	= "SGMII",
85 	[LOOPBACK_XGBR]		= "XGBR",
86 	[LOOPBACK_XFI]		= "XFI",
87 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
88 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
89 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
90 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
91 	[LOOPBACK_GPHY]		= "GPHY",
92 	[LOOPBACK_PHYXS]	= "PHYXS",
93 	[LOOPBACK_PCS]		= "PCS",
94 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
95 	[LOOPBACK_XPORT]	= "XPORT",
96 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
97 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
98 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
99 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
100 	[LOOPBACK_GMII_WS]	= "GMII_WS",
101 	[LOOPBACK_XFI_WS]	= "XFI_WS",
102 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
103 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
104 };
105 
106 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
107  * queued onto this work queue. This is not a per-nic work queue, because
108  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
109  */
110 static struct workqueue_struct *reset_workqueue;
111 
112 int efx_create_reset_workqueue(void)
113 {
114 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
115 	if (!reset_workqueue) {
116 		printk(KERN_ERR "Failed to create reset workqueue\n");
117 		return -ENOMEM;
118 	}
119 
120 	return 0;
121 }
122 
123 void efx_queue_reset_work(struct efx_nic *efx)
124 {
125 	queue_work(reset_workqueue, &efx->reset_work);
126 }
127 
128 void efx_flush_reset_workqueue(struct efx_nic *efx)
129 {
130 	cancel_work_sync(&efx->reset_work);
131 }
132 
133 void efx_destroy_reset_workqueue(void)
134 {
135 	if (reset_workqueue) {
136 		destroy_workqueue(reset_workqueue);
137 		reset_workqueue = NULL;
138 	}
139 }
140 
141 /* We assume that efx->type->reconfigure_mac will always try to sync RX
142  * filters and therefore needs to read-lock the filter table against freeing
143  */
144 void efx_mac_reconfigure(struct efx_nic *efx, bool mtu_only)
145 {
146 	if (efx->type->reconfigure_mac) {
147 		down_read(&efx->filter_sem);
148 		efx->type->reconfigure_mac(efx, mtu_only);
149 		up_read(&efx->filter_sem);
150 	}
151 }
152 
153 /* Asynchronous work item for changing MAC promiscuity and multicast
154  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
155  * MAC directly.
156  */
157 static void efx_mac_work(struct work_struct *data)
158 {
159 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
160 
161 	mutex_lock(&efx->mac_lock);
162 	if (efx->port_enabled)
163 		efx_mac_reconfigure(efx, false);
164 	mutex_unlock(&efx->mac_lock);
165 }
166 
167 int efx_set_mac_address(struct net_device *net_dev, void *data)
168 {
169 	struct efx_nic *efx = netdev_priv(net_dev);
170 	struct sockaddr *addr = data;
171 	u8 *new_addr = addr->sa_data;
172 	u8 old_addr[6];
173 	int rc;
174 
175 	if (!is_valid_ether_addr(new_addr)) {
176 		netif_err(efx, drv, efx->net_dev,
177 			  "invalid ethernet MAC address requested: %pM\n",
178 			  new_addr);
179 		return -EADDRNOTAVAIL;
180 	}
181 
182 	/* save old address */
183 	ether_addr_copy(old_addr, net_dev->dev_addr);
184 	ether_addr_copy(net_dev->dev_addr, new_addr);
185 	if (efx->type->set_mac_address) {
186 		rc = efx->type->set_mac_address(efx);
187 		if (rc) {
188 			ether_addr_copy(net_dev->dev_addr, old_addr);
189 			return rc;
190 		}
191 	}
192 
193 	/* Reconfigure the MAC */
194 	mutex_lock(&efx->mac_lock);
195 	efx_mac_reconfigure(efx, false);
196 	mutex_unlock(&efx->mac_lock);
197 
198 	return 0;
199 }
200 
201 /* Context: netif_addr_lock held, BHs disabled. */
202 void efx_set_rx_mode(struct net_device *net_dev)
203 {
204 	struct efx_nic *efx = netdev_priv(net_dev);
205 
206 	if (efx->port_enabled)
207 		queue_work(efx->workqueue, &efx->mac_work);
208 	/* Otherwise efx_start_port() will do this */
209 }
210 
211 int efx_set_features(struct net_device *net_dev, netdev_features_t data)
212 {
213 	struct efx_nic *efx = netdev_priv(net_dev);
214 	int rc;
215 
216 	/* If disabling RX n-tuple filtering, clear existing filters */
217 	if (net_dev->features & ~data & NETIF_F_NTUPLE) {
218 		rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
219 		if (rc)
220 			return rc;
221 	}
222 
223 	/* If Rx VLAN filter is changed, update filters via mac_reconfigure.
224 	 * If rx-fcs is changed, mac_reconfigure updates that too.
225 	 */
226 	if ((net_dev->features ^ data) & (NETIF_F_HW_VLAN_CTAG_FILTER |
227 					  NETIF_F_RXFCS)) {
228 		/* efx_set_rx_mode() will schedule MAC work to update filters
229 		 * when a new features are finally set in net_dev.
230 		 */
231 		efx_set_rx_mode(net_dev);
232 	}
233 
234 	return 0;
235 }
236 
237 /* This ensures that the kernel is kept informed (via
238  * netif_carrier_on/off) of the link status, and also maintains the
239  * link status's stop on the port's TX queue.
240  */
241 void efx_link_status_changed(struct efx_nic *efx)
242 {
243 	struct efx_link_state *link_state = &efx->link_state;
244 
245 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
246 	 * that no events are triggered between unregister_netdev() and the
247 	 * driver unloading. A more general condition is that NETDEV_CHANGE
248 	 * can only be generated between NETDEV_UP and NETDEV_DOWN
249 	 */
250 	if (!netif_running(efx->net_dev))
251 		return;
252 
253 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
254 		efx->n_link_state_changes++;
255 
256 		if (link_state->up)
257 			netif_carrier_on(efx->net_dev);
258 		else
259 			netif_carrier_off(efx->net_dev);
260 	}
261 
262 	/* Status message for kernel log */
263 	if (link_state->up)
264 		netif_info(efx, link, efx->net_dev,
265 			   "link up at %uMbps %s-duplex (MTU %d)\n",
266 			   link_state->speed, link_state->fd ? "full" : "half",
267 			   efx->net_dev->mtu);
268 	else
269 		netif_info(efx, link, efx->net_dev, "link down\n");
270 }
271 
272 unsigned int efx_xdp_max_mtu(struct efx_nic *efx)
273 {
274 	/* The maximum MTU that we can fit in a single page, allowing for
275 	 * framing, overhead and XDP headroom + tailroom.
276 	 */
277 	int overhead = EFX_MAX_FRAME_LEN(0) + sizeof(struct efx_rx_page_state) +
278 		       efx->rx_prefix_size + efx->type->rx_buffer_padding +
279 		       efx->rx_ip_align + EFX_XDP_HEADROOM + EFX_XDP_TAILROOM;
280 
281 	return PAGE_SIZE - overhead;
282 }
283 
284 /* Context: process, rtnl_lock() held. */
285 int efx_change_mtu(struct net_device *net_dev, int new_mtu)
286 {
287 	struct efx_nic *efx = netdev_priv(net_dev);
288 	int rc;
289 
290 	rc = efx_check_disabled(efx);
291 	if (rc)
292 		return rc;
293 
294 	if (rtnl_dereference(efx->xdp_prog) &&
295 	    new_mtu > efx_xdp_max_mtu(efx)) {
296 		netif_err(efx, drv, efx->net_dev,
297 			  "Requested MTU of %d too big for XDP (max: %d)\n",
298 			  new_mtu, efx_xdp_max_mtu(efx));
299 		return -EINVAL;
300 	}
301 
302 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
303 
304 	efx_device_detach_sync(efx);
305 	efx_stop_all(efx);
306 
307 	mutex_lock(&efx->mac_lock);
308 	net_dev->mtu = new_mtu;
309 	efx_mac_reconfigure(efx, true);
310 	mutex_unlock(&efx->mac_lock);
311 
312 	efx_start_all(efx);
313 	efx_device_attach_if_not_resetting(efx);
314 	return 0;
315 }
316 
317 /**************************************************************************
318  *
319  * Hardware monitor
320  *
321  **************************************************************************/
322 
323 /* Run periodically off the general workqueue */
324 static void efx_monitor(struct work_struct *data)
325 {
326 	struct efx_nic *efx = container_of(data, struct efx_nic,
327 					   monitor_work.work);
328 
329 	netif_vdbg(efx, timer, efx->net_dev,
330 		   "hardware monitor executing on CPU %d\n",
331 		   raw_smp_processor_id());
332 	BUG_ON(efx->type->monitor == NULL);
333 
334 	/* If the mac_lock is already held then it is likely a port
335 	 * reconfiguration is already in place, which will likely do
336 	 * most of the work of monitor() anyway.
337 	 */
338 	if (mutex_trylock(&efx->mac_lock)) {
339 		if (efx->port_enabled && efx->type->monitor)
340 			efx->type->monitor(efx);
341 		mutex_unlock(&efx->mac_lock);
342 	}
343 
344 	efx_start_monitor(efx);
345 }
346 
347 void efx_start_monitor(struct efx_nic *efx)
348 {
349 	if (efx->type->monitor)
350 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
351 				   efx_monitor_interval);
352 }
353 
354 /**************************************************************************
355  *
356  * Event queue processing
357  *
358  *************************************************************************/
359 
360 /* Channels are shutdown and reinitialised whilst the NIC is running
361  * to propagate configuration changes (mtu, checksum offload), or
362  * to clear hardware error conditions
363  */
364 static void efx_start_datapath(struct efx_nic *efx)
365 {
366 	netdev_features_t old_features = efx->net_dev->features;
367 	bool old_rx_scatter = efx->rx_scatter;
368 	size_t rx_buf_len;
369 
370 	/* Calculate the rx buffer allocation parameters required to
371 	 * support the current MTU, including padding for header
372 	 * alignment and overruns.
373 	 */
374 	efx->rx_dma_len = (efx->rx_prefix_size +
375 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
376 			   efx->type->rx_buffer_padding);
377 	rx_buf_len = (sizeof(struct efx_rx_page_state)   + EFX_XDP_HEADROOM +
378 		      efx->rx_ip_align + efx->rx_dma_len + EFX_XDP_TAILROOM);
379 
380 	if (rx_buf_len <= PAGE_SIZE) {
381 		efx->rx_scatter = efx->type->always_rx_scatter;
382 		efx->rx_buffer_order = 0;
383 	} else if (efx->type->can_rx_scatter) {
384 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
385 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
386 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
387 				       EFX_RX_BUF_ALIGNMENT) >
388 			     PAGE_SIZE);
389 		efx->rx_scatter = true;
390 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
391 		efx->rx_buffer_order = 0;
392 	} else {
393 		efx->rx_scatter = false;
394 		efx->rx_buffer_order = get_order(rx_buf_len);
395 	}
396 
397 	efx_rx_config_page_split(efx);
398 	if (efx->rx_buffer_order)
399 		netif_dbg(efx, drv, efx->net_dev,
400 			  "RX buf len=%u; page order=%u batch=%u\n",
401 			  efx->rx_dma_len, efx->rx_buffer_order,
402 			  efx->rx_pages_per_batch);
403 	else
404 		netif_dbg(efx, drv, efx->net_dev,
405 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
406 			  efx->rx_dma_len, efx->rx_page_buf_step,
407 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
408 
409 	/* Restore previously fixed features in hw_features and remove
410 	 * features which are fixed now
411 	 */
412 	efx->net_dev->hw_features |= efx->net_dev->features;
413 	efx->net_dev->hw_features &= ~efx->fixed_features;
414 	efx->net_dev->features |= efx->fixed_features;
415 	if (efx->net_dev->features != old_features)
416 		netdev_features_change(efx->net_dev);
417 
418 	/* RX filters may also have scatter-enabled flags */
419 	if ((efx->rx_scatter != old_rx_scatter) &&
420 	    efx->type->filter_update_rx_scatter)
421 		efx->type->filter_update_rx_scatter(efx);
422 
423 	/* We must keep at least one descriptor in a TX ring empty.
424 	 * We could avoid this when the queue size does not exactly
425 	 * match the hardware ring size, but it's not that important.
426 	 * Therefore we stop the queue when one more skb might fill
427 	 * the ring completely.  We wake it when half way back to
428 	 * empty.
429 	 */
430 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
431 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
432 
433 	/* Initialise the channels */
434 	efx_start_channels(efx);
435 
436 	efx_ptp_start_datapath(efx);
437 
438 	if (netif_device_present(efx->net_dev))
439 		netif_tx_wake_all_queues(efx->net_dev);
440 }
441 
442 static void efx_stop_datapath(struct efx_nic *efx)
443 {
444 	EFX_ASSERT_RESET_SERIALISED(efx);
445 	BUG_ON(efx->port_enabled);
446 
447 	efx_ptp_stop_datapath(efx);
448 
449 	efx_stop_channels(efx);
450 }
451 
452 /**************************************************************************
453  *
454  * Port handling
455  *
456  **************************************************************************/
457 
458 /* Equivalent to efx_link_set_advertising with all-zeroes, except does not
459  * force the Autoneg bit on.
460  */
461 void efx_link_clear_advertising(struct efx_nic *efx)
462 {
463 	bitmap_zero(efx->link_advertising, __ETHTOOL_LINK_MODE_MASK_NBITS);
464 	efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
465 }
466 
467 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
468 {
469 	efx->wanted_fc = wanted_fc;
470 	if (efx->link_advertising[0]) {
471 		if (wanted_fc & EFX_FC_RX)
472 			efx->link_advertising[0] |= (ADVERTISED_Pause |
473 						     ADVERTISED_Asym_Pause);
474 		else
475 			efx->link_advertising[0] &= ~(ADVERTISED_Pause |
476 						      ADVERTISED_Asym_Pause);
477 		if (wanted_fc & EFX_FC_TX)
478 			efx->link_advertising[0] ^= ADVERTISED_Asym_Pause;
479 	}
480 }
481 
482 static void efx_start_port(struct efx_nic *efx)
483 {
484 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
485 	BUG_ON(efx->port_enabled);
486 
487 	mutex_lock(&efx->mac_lock);
488 	efx->port_enabled = true;
489 
490 	/* Ensure MAC ingress/egress is enabled */
491 	efx_mac_reconfigure(efx, false);
492 
493 	mutex_unlock(&efx->mac_lock);
494 }
495 
496 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
497  * and the async self-test, wait for them to finish and prevent them
498  * being scheduled again.  This doesn't cover online resets, which
499  * should only be cancelled when removing the device.
500  */
501 static void efx_stop_port(struct efx_nic *efx)
502 {
503 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
504 
505 	EFX_ASSERT_RESET_SERIALISED(efx);
506 
507 	mutex_lock(&efx->mac_lock);
508 	efx->port_enabled = false;
509 	mutex_unlock(&efx->mac_lock);
510 
511 	/* Serialise against efx_set_multicast_list() */
512 	netif_addr_lock_bh(efx->net_dev);
513 	netif_addr_unlock_bh(efx->net_dev);
514 
515 	cancel_delayed_work_sync(&efx->monitor_work);
516 	efx_selftest_async_cancel(efx);
517 	cancel_work_sync(&efx->mac_work);
518 }
519 
520 /* If the interface is supposed to be running but is not, start
521  * the hardware and software data path, regular activity for the port
522  * (MAC statistics, link polling, etc.) and schedule the port to be
523  * reconfigured.  Interrupts must already be enabled.  This function
524  * is safe to call multiple times, so long as the NIC is not disabled.
525  * Requires the RTNL lock.
526  */
527 void efx_start_all(struct efx_nic *efx)
528 {
529 	EFX_ASSERT_RESET_SERIALISED(efx);
530 	BUG_ON(efx->state == STATE_DISABLED);
531 
532 	/* Check that it is appropriate to restart the interface. All
533 	 * of these flags are safe to read under just the rtnl lock
534 	 */
535 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
536 	    efx->reset_pending)
537 		return;
538 
539 	efx_start_port(efx);
540 	efx_start_datapath(efx);
541 
542 	/* Start the hardware monitor if there is one */
543 	efx_start_monitor(efx);
544 
545 	/* Link state detection is normally event-driven; we have
546 	 * to poll now because we could have missed a change
547 	 */
548 	mutex_lock(&efx->mac_lock);
549 	if (efx_mcdi_phy_poll(efx))
550 		efx_link_status_changed(efx);
551 	mutex_unlock(&efx->mac_lock);
552 
553 	if (efx->type->start_stats) {
554 		efx->type->start_stats(efx);
555 		efx->type->pull_stats(efx);
556 		spin_lock_bh(&efx->stats_lock);
557 		efx->type->update_stats(efx, NULL, NULL);
558 		spin_unlock_bh(&efx->stats_lock);
559 	}
560 }
561 
562 /* Quiesce the hardware and software data path, and regular activity
563  * for the port without bringing the link down.  Safe to call multiple
564  * times with the NIC in almost any state, but interrupts should be
565  * enabled.  Requires the RTNL lock.
566  */
567 void efx_stop_all(struct efx_nic *efx)
568 {
569 	EFX_ASSERT_RESET_SERIALISED(efx);
570 
571 	/* port_enabled can be read safely under the rtnl lock */
572 	if (!efx->port_enabled)
573 		return;
574 
575 	if (efx->type->update_stats) {
576 		/* update stats before we go down so we can accurately count
577 		 * rx_nodesc_drops
578 		 */
579 		efx->type->pull_stats(efx);
580 		spin_lock_bh(&efx->stats_lock);
581 		efx->type->update_stats(efx, NULL, NULL);
582 		spin_unlock_bh(&efx->stats_lock);
583 		efx->type->stop_stats(efx);
584 	}
585 
586 	efx_stop_port(efx);
587 
588 	/* Stop the kernel transmit interface.  This is only valid if
589 	 * the device is stopped or detached; otherwise the watchdog
590 	 * may fire immediately.
591 	 */
592 	WARN_ON(netif_running(efx->net_dev) &&
593 		netif_device_present(efx->net_dev));
594 	netif_tx_disable(efx->net_dev);
595 
596 	efx_stop_datapath(efx);
597 }
598 
599 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
600 void efx_net_stats(struct net_device *net_dev, struct rtnl_link_stats64 *stats)
601 {
602 	struct efx_nic *efx = netdev_priv(net_dev);
603 
604 	spin_lock_bh(&efx->stats_lock);
605 	efx_nic_update_stats_atomic(efx, NULL, stats);
606 	spin_unlock_bh(&efx->stats_lock);
607 }
608 
609 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
610  * the MAC appropriately. All other PHY configuration changes are pushed
611  * through phy_op->set_settings(), and pushed asynchronously to the MAC
612  * through efx_monitor().
613  *
614  * Callers must hold the mac_lock
615  */
616 int __efx_reconfigure_port(struct efx_nic *efx)
617 {
618 	enum efx_phy_mode phy_mode;
619 	int rc = 0;
620 
621 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
622 
623 	/* Disable PHY transmit in mac level loopbacks */
624 	phy_mode = efx->phy_mode;
625 	if (LOOPBACK_INTERNAL(efx))
626 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
627 	else
628 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
629 
630 	if (efx->type->reconfigure_port)
631 		rc = efx->type->reconfigure_port(efx);
632 
633 	if (rc)
634 		efx->phy_mode = phy_mode;
635 
636 	return rc;
637 }
638 
639 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
640  * disabled.
641  */
642 int efx_reconfigure_port(struct efx_nic *efx)
643 {
644 	int rc;
645 
646 	EFX_ASSERT_RESET_SERIALISED(efx);
647 
648 	mutex_lock(&efx->mac_lock);
649 	rc = __efx_reconfigure_port(efx);
650 	mutex_unlock(&efx->mac_lock);
651 
652 	return rc;
653 }
654 
655 /**************************************************************************
656  *
657  * Device reset and suspend
658  *
659  **************************************************************************/
660 
661 static void efx_wait_for_bist_end(struct efx_nic *efx)
662 {
663 	int i;
664 
665 	for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
666 		if (efx_mcdi_poll_reboot(efx))
667 			goto out;
668 		msleep(BIST_WAIT_DELAY_MS);
669 	}
670 
671 	netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
672 out:
673 	/* Either way unset the BIST flag. If we found no reboot we probably
674 	 * won't recover, but we should try.
675 	 */
676 	efx->mc_bist_for_other_fn = false;
677 }
678 
679 /* Try recovery mechanisms.
680  * For now only EEH is supported.
681  * Returns 0 if the recovery mechanisms are unsuccessful.
682  * Returns a non-zero value otherwise.
683  */
684 int efx_try_recovery(struct efx_nic *efx)
685 {
686 #ifdef CONFIG_EEH
687 	/* A PCI error can occur and not be seen by EEH because nothing
688 	 * happens on the PCI bus. In this case the driver may fail and
689 	 * schedule a 'recover or reset', leading to this recovery handler.
690 	 * Manually call the eeh failure check function.
691 	 */
692 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
693 	if (eeh_dev_check_failure(eehdev)) {
694 		/* The EEH mechanisms will handle the error and reset the
695 		 * device if necessary.
696 		 */
697 		return 1;
698 	}
699 #endif
700 	return 0;
701 }
702 
703 /* Tears down the entire software state and most of the hardware state
704  * before reset.
705  */
706 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
707 {
708 	EFX_ASSERT_RESET_SERIALISED(efx);
709 
710 	if (method == RESET_TYPE_MCDI_TIMEOUT)
711 		efx->type->prepare_flr(efx);
712 
713 	efx_stop_all(efx);
714 	efx_disable_interrupts(efx);
715 
716 	mutex_lock(&efx->mac_lock);
717 	down_write(&efx->filter_sem);
718 	mutex_lock(&efx->rss_lock);
719 	efx->type->fini(efx);
720 }
721 
722 /* Context: netif_tx_lock held, BHs disabled. */
723 void efx_watchdog(struct net_device *net_dev, unsigned int txqueue)
724 {
725 	struct efx_nic *efx = netdev_priv(net_dev);
726 
727 	netif_err(efx, tx_err, efx->net_dev,
728 		  "TX stuck with port_enabled=%d: resetting channels\n",
729 		  efx->port_enabled);
730 
731 	efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
732 }
733 
734 /* This function will always ensure that the locks acquired in
735  * efx_reset_down() are released. A failure return code indicates
736  * that we were unable to reinitialise the hardware, and the
737  * driver should be disabled. If ok is false, then the rx and tx
738  * engines are not restarted, pending a RESET_DISABLE.
739  */
740 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
741 {
742 	int rc;
743 
744 	EFX_ASSERT_RESET_SERIALISED(efx);
745 
746 	if (method == RESET_TYPE_MCDI_TIMEOUT)
747 		efx->type->finish_flr(efx);
748 
749 	/* Ensure that SRAM is initialised even if we're disabling the device */
750 	rc = efx->type->init(efx);
751 	if (rc) {
752 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
753 		goto fail;
754 	}
755 
756 	if (!ok)
757 		goto fail;
758 
759 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
760 	    method != RESET_TYPE_DATAPATH) {
761 		rc = efx_mcdi_port_reconfigure(efx);
762 		if (rc && rc != -EPERM)
763 			netif_err(efx, drv, efx->net_dev,
764 				  "could not restore PHY settings\n");
765 	}
766 
767 	rc = efx_enable_interrupts(efx);
768 	if (rc)
769 		goto fail;
770 
771 #ifdef CONFIG_SFC_SRIOV
772 	rc = efx->type->vswitching_restore(efx);
773 	if (rc) /* not fatal; the PF will still work fine */
774 		netif_warn(efx, probe, efx->net_dev,
775 			   "failed to restore vswitching rc=%d;"
776 			   " VFs may not function\n", rc);
777 #endif
778 
779 	if (efx->type->rx_restore_rss_contexts)
780 		efx->type->rx_restore_rss_contexts(efx);
781 	mutex_unlock(&efx->rss_lock);
782 	efx->type->filter_table_restore(efx);
783 	up_write(&efx->filter_sem);
784 	if (efx->type->sriov_reset)
785 		efx->type->sriov_reset(efx);
786 
787 	mutex_unlock(&efx->mac_lock);
788 
789 	efx_start_all(efx);
790 
791 	if (efx->type->udp_tnl_push_ports)
792 		efx->type->udp_tnl_push_ports(efx);
793 
794 	return 0;
795 
796 fail:
797 	efx->port_initialized = false;
798 
799 	mutex_unlock(&efx->rss_lock);
800 	up_write(&efx->filter_sem);
801 	mutex_unlock(&efx->mac_lock);
802 
803 	return rc;
804 }
805 
806 /* Reset the NIC using the specified method.  Note that the reset may
807  * fail, in which case the card will be left in an unusable state.
808  *
809  * Caller must hold the rtnl_lock.
810  */
811 int efx_reset(struct efx_nic *efx, enum reset_type method)
812 {
813 	int rc, rc2 = 0;
814 	bool disabled;
815 
816 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
817 		   RESET_TYPE(method));
818 
819 	efx_device_detach_sync(efx);
820 	/* efx_reset_down() grabs locks that prevent recovery on EF100.
821 	 * EF100 reset is handled in the efx_nic_type callback below.
822 	 */
823 	if (efx_nic_rev(efx) != EFX_REV_EF100)
824 		efx_reset_down(efx, method);
825 
826 	rc = efx->type->reset(efx, method);
827 	if (rc) {
828 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
829 		goto out;
830 	}
831 
832 	/* Clear flags for the scopes we covered.  We assume the NIC and
833 	 * driver are now quiescent so that there is no race here.
834 	 */
835 	if (method < RESET_TYPE_MAX_METHOD)
836 		efx->reset_pending &= -(1 << (method + 1));
837 	else /* it doesn't fit into the well-ordered scope hierarchy */
838 		__clear_bit(method, &efx->reset_pending);
839 
840 	/* Reinitialise bus-mastering, which may have been turned off before
841 	 * the reset was scheduled. This is still appropriate, even in the
842 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
843 	 * can respond to requests.
844 	 */
845 	pci_set_master(efx->pci_dev);
846 
847 out:
848 	/* Leave device stopped if necessary */
849 	disabled = rc ||
850 		method == RESET_TYPE_DISABLE ||
851 		method == RESET_TYPE_RECOVER_OR_DISABLE;
852 	if (efx_nic_rev(efx) != EFX_REV_EF100)
853 		rc2 = efx_reset_up(efx, method, !disabled);
854 	if (rc2) {
855 		disabled = true;
856 		if (!rc)
857 			rc = rc2;
858 	}
859 
860 	if (disabled) {
861 		dev_close(efx->net_dev);
862 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
863 		efx->state = STATE_DISABLED;
864 	} else {
865 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
866 		efx_device_attach_if_not_resetting(efx);
867 	}
868 	return rc;
869 }
870 
871 /* The worker thread exists so that code that cannot sleep can
872  * schedule a reset for later.
873  */
874 static void efx_reset_work(struct work_struct *data)
875 {
876 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
877 	unsigned long pending;
878 	enum reset_type method;
879 
880 	pending = READ_ONCE(efx->reset_pending);
881 	method = fls(pending) - 1;
882 
883 	if (method == RESET_TYPE_MC_BIST)
884 		efx_wait_for_bist_end(efx);
885 
886 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
887 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
888 	    efx_try_recovery(efx))
889 		return;
890 
891 	if (!pending)
892 		return;
893 
894 	rtnl_lock();
895 
896 	/* We checked the state in efx_schedule_reset() but it may
897 	 * have changed by now.  Now that we have the RTNL lock,
898 	 * it cannot change again.
899 	 */
900 	if (efx->state == STATE_READY)
901 		(void)efx_reset(efx, method);
902 
903 	rtnl_unlock();
904 }
905 
906 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
907 {
908 	enum reset_type method;
909 
910 	if (efx->state == STATE_RECOVERY) {
911 		netif_dbg(efx, drv, efx->net_dev,
912 			  "recovering: skip scheduling %s reset\n",
913 			  RESET_TYPE(type));
914 		return;
915 	}
916 
917 	switch (type) {
918 	case RESET_TYPE_INVISIBLE:
919 	case RESET_TYPE_ALL:
920 	case RESET_TYPE_RECOVER_OR_ALL:
921 	case RESET_TYPE_WORLD:
922 	case RESET_TYPE_DISABLE:
923 	case RESET_TYPE_RECOVER_OR_DISABLE:
924 	case RESET_TYPE_DATAPATH:
925 	case RESET_TYPE_MC_BIST:
926 	case RESET_TYPE_MCDI_TIMEOUT:
927 		method = type;
928 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
929 			  RESET_TYPE(method));
930 		break;
931 	default:
932 		method = efx->type->map_reset_reason(type);
933 		netif_dbg(efx, drv, efx->net_dev,
934 			  "scheduling %s reset for %s\n",
935 			  RESET_TYPE(method), RESET_TYPE(type));
936 		break;
937 	}
938 
939 	set_bit(method, &efx->reset_pending);
940 	smp_mb(); /* ensure we change reset_pending before checking state */
941 
942 	/* If we're not READY then just leave the flags set as the cue
943 	 * to abort probing or reschedule the reset later.
944 	 */
945 	if (READ_ONCE(efx->state) != STATE_READY)
946 		return;
947 
948 	/* efx_process_channel() will no longer read events once a
949 	 * reset is scheduled. So switch back to poll'd MCDI completions.
950 	 */
951 	efx_mcdi_mode_poll(efx);
952 
953 	efx_queue_reset_work(efx);
954 }
955 
956 /**************************************************************************
957  *
958  * Dummy NIC operations
959  *
960  * Can be used for some unimplemented operations
961  * Needed so all function pointers are valid and do not have to be tested
962  * before use
963  *
964  **************************************************************************/
965 int efx_port_dummy_op_int(struct efx_nic *efx)
966 {
967 	return 0;
968 }
969 void efx_port_dummy_op_void(struct efx_nic *efx) {}
970 
971 /**************************************************************************
972  *
973  * Data housekeeping
974  *
975  **************************************************************************/
976 
977 /* This zeroes out and then fills in the invariants in a struct
978  * efx_nic (including all sub-structures).
979  */
980 int efx_init_struct(struct efx_nic *efx,
981 		    struct pci_dev *pci_dev, struct net_device *net_dev)
982 {
983 	int rc = -ENOMEM;
984 
985 	/* Initialise common structures */
986 	INIT_LIST_HEAD(&efx->node);
987 	INIT_LIST_HEAD(&efx->secondary_list);
988 	spin_lock_init(&efx->biu_lock);
989 #ifdef CONFIG_SFC_MTD
990 	INIT_LIST_HEAD(&efx->mtd_list);
991 #endif
992 	INIT_WORK(&efx->reset_work, efx_reset_work);
993 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
994 	efx_selftest_async_init(efx);
995 	efx->pci_dev = pci_dev;
996 	efx->msg_enable = debug;
997 	efx->state = STATE_UNINIT;
998 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
999 
1000 	efx->net_dev = net_dev;
1001 	efx->rx_prefix_size = efx->type->rx_prefix_size;
1002 	efx->rx_ip_align =
1003 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
1004 	efx->rx_packet_hash_offset =
1005 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
1006 	efx->rx_packet_ts_offset =
1007 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
1008 	INIT_LIST_HEAD(&efx->rss_context.list);
1009 	efx->rss_context.context_id = EFX_MCDI_RSS_CONTEXT_INVALID;
1010 	mutex_init(&efx->rss_lock);
1011 	efx->vport_id = EVB_PORT_ID_ASSIGNED;
1012 	spin_lock_init(&efx->stats_lock);
1013 	efx->vi_stride = EFX_DEFAULT_VI_STRIDE;
1014 	efx->num_mac_stats = MC_CMD_MAC_NSTATS;
1015 	BUILD_BUG_ON(MC_CMD_MAC_NSTATS - 1 != MC_CMD_MAC_GENERATION_END);
1016 	mutex_init(&efx->mac_lock);
1017 	init_rwsem(&efx->filter_sem);
1018 #ifdef CONFIG_RFS_ACCEL
1019 	mutex_init(&efx->rps_mutex);
1020 	spin_lock_init(&efx->rps_hash_lock);
1021 	/* Failure to allocate is not fatal, but may degrade ARFS performance */
1022 	efx->rps_hash_table = kcalloc(EFX_ARFS_HASH_TABLE_SIZE,
1023 				      sizeof(*efx->rps_hash_table), GFP_KERNEL);
1024 #endif
1025 	efx->mdio.dev = net_dev;
1026 	INIT_WORK(&efx->mac_work, efx_mac_work);
1027 	init_waitqueue_head(&efx->flush_wq);
1028 
1029 	efx->tx_queues_per_channel = 1;
1030 	efx->rxq_entries = EFX_DEFAULT_DMAQ_SIZE;
1031 	efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1032 
1033 	efx->mem_bar = UINT_MAX;
1034 
1035 	rc = efx_init_channels(efx);
1036 	if (rc)
1037 		goto fail;
1038 
1039 	/* Would be good to use the net_dev name, but we're too early */
1040 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
1041 		 pci_name(pci_dev));
1042 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
1043 	if (!efx->workqueue) {
1044 		rc = -ENOMEM;
1045 		goto fail;
1046 	}
1047 
1048 	return 0;
1049 
1050 fail:
1051 	efx_fini_struct(efx);
1052 	return rc;
1053 }
1054 
1055 void efx_fini_struct(struct efx_nic *efx)
1056 {
1057 #ifdef CONFIG_RFS_ACCEL
1058 	kfree(efx->rps_hash_table);
1059 #endif
1060 
1061 	efx_fini_channels(efx);
1062 
1063 	kfree(efx->vpd_sn);
1064 
1065 	if (efx->workqueue) {
1066 		destroy_workqueue(efx->workqueue);
1067 		efx->workqueue = NULL;
1068 	}
1069 }
1070 
1071 /* This configures the PCI device to enable I/O and DMA. */
1072 int efx_init_io(struct efx_nic *efx, int bar, dma_addr_t dma_mask,
1073 		unsigned int mem_map_size)
1074 {
1075 	struct pci_dev *pci_dev = efx->pci_dev;
1076 	int rc;
1077 
1078 	efx->mem_bar = UINT_MAX;
1079 
1080 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O bar=%d\n", bar);
1081 
1082 	rc = pci_enable_device(pci_dev);
1083 	if (rc) {
1084 		netif_err(efx, probe, efx->net_dev,
1085 			  "failed to enable PCI device\n");
1086 		goto fail1;
1087 	}
1088 
1089 	pci_set_master(pci_dev);
1090 
1091 	rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1092 	if (rc) {
1093 		netif_err(efx, probe, efx->net_dev,
1094 			  "could not find a suitable DMA mask\n");
1095 		goto fail2;
1096 	}
1097 	netif_dbg(efx, probe, efx->net_dev,
1098 		  "using DMA mask %llx\n", (unsigned long long)dma_mask);
1099 
1100 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1101 	if (!efx->membase_phys) {
1102 		netif_err(efx, probe, efx->net_dev,
1103 			  "ERROR: No BAR%d mapping from the BIOS. "
1104 			  "Try pci=realloc on the kernel command line\n", bar);
1105 		rc = -ENODEV;
1106 		goto fail3;
1107 	}
1108 
1109 	rc = pci_request_region(pci_dev, bar, "sfc");
1110 	if (rc) {
1111 		netif_err(efx, probe, efx->net_dev,
1112 			  "request for memory BAR[%d] failed\n", bar);
1113 		rc = -EIO;
1114 		goto fail3;
1115 	}
1116 	efx->mem_bar = bar;
1117 	efx->membase = ioremap(efx->membase_phys, mem_map_size);
1118 	if (!efx->membase) {
1119 		netif_err(efx, probe, efx->net_dev,
1120 			  "could not map memory BAR[%d] at %llx+%x\n", bar,
1121 			  (unsigned long long)efx->membase_phys, mem_map_size);
1122 		rc = -ENOMEM;
1123 		goto fail4;
1124 	}
1125 	netif_dbg(efx, probe, efx->net_dev,
1126 		  "memory BAR[%d] at %llx+%x (virtual %p)\n", bar,
1127 		  (unsigned long long)efx->membase_phys, mem_map_size,
1128 		  efx->membase);
1129 
1130 	return 0;
1131 
1132 fail4:
1133 	pci_release_region(efx->pci_dev, bar);
1134 fail3:
1135 	efx->membase_phys = 0;
1136 fail2:
1137 	pci_disable_device(efx->pci_dev);
1138 fail1:
1139 	return rc;
1140 }
1141 
1142 void efx_fini_io(struct efx_nic *efx)
1143 {
1144 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1145 
1146 	if (efx->membase) {
1147 		iounmap(efx->membase);
1148 		efx->membase = NULL;
1149 	}
1150 
1151 	if (efx->membase_phys) {
1152 		pci_release_region(efx->pci_dev, efx->mem_bar);
1153 		efx->membase_phys = 0;
1154 		efx->mem_bar = UINT_MAX;
1155 	}
1156 
1157 	/* Don't disable bus-mastering if VFs are assigned */
1158 	if (!pci_vfs_assigned(efx->pci_dev))
1159 		pci_disable_device(efx->pci_dev);
1160 }
1161 
1162 #ifdef CONFIG_SFC_MCDI_LOGGING
1163 static ssize_t mcdi_logging_show(struct device *dev,
1164 				 struct device_attribute *attr,
1165 				 char *buf)
1166 {
1167 	struct efx_nic *efx = dev_get_drvdata(dev);
1168 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1169 
1170 	return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
1171 }
1172 
1173 static ssize_t mcdi_logging_store(struct device *dev,
1174 				  struct device_attribute *attr,
1175 				  const char *buf, size_t count)
1176 {
1177 	struct efx_nic *efx = dev_get_drvdata(dev);
1178 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1179 	bool enable = count > 0 && *buf != '0';
1180 
1181 	mcdi->logging_enabled = enable;
1182 	return count;
1183 }
1184 
1185 static DEVICE_ATTR_RW(mcdi_logging);
1186 
1187 void efx_init_mcdi_logging(struct efx_nic *efx)
1188 {
1189 	int rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1190 
1191 	if (rc) {
1192 		netif_warn(efx, drv, efx->net_dev,
1193 			   "failed to init net dev attributes\n");
1194 	}
1195 }
1196 
1197 void efx_fini_mcdi_logging(struct efx_nic *efx)
1198 {
1199 	device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1200 }
1201 #endif
1202 
1203 /* A PCI error affecting this device was detected.
1204  * At this point MMIO and DMA may be disabled.
1205  * Stop the software path and request a slot reset.
1206  */
1207 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
1208 					      pci_channel_state_t state)
1209 {
1210 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
1211 	struct efx_nic *efx = pci_get_drvdata(pdev);
1212 
1213 	if (state == pci_channel_io_perm_failure)
1214 		return PCI_ERS_RESULT_DISCONNECT;
1215 
1216 	rtnl_lock();
1217 
1218 	if (efx->state != STATE_DISABLED) {
1219 		efx->state = STATE_RECOVERY;
1220 		efx->reset_pending = 0;
1221 
1222 		efx_device_detach_sync(efx);
1223 
1224 		efx_stop_all(efx);
1225 		efx_disable_interrupts(efx);
1226 
1227 		status = PCI_ERS_RESULT_NEED_RESET;
1228 	} else {
1229 		/* If the interface is disabled we don't want to do anything
1230 		 * with it.
1231 		 */
1232 		status = PCI_ERS_RESULT_RECOVERED;
1233 	}
1234 
1235 	rtnl_unlock();
1236 
1237 	pci_disable_device(pdev);
1238 
1239 	return status;
1240 }
1241 
1242 /* Fake a successful reset, which will be performed later in efx_io_resume. */
1243 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
1244 {
1245 	struct efx_nic *efx = pci_get_drvdata(pdev);
1246 	pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
1247 
1248 	if (pci_enable_device(pdev)) {
1249 		netif_err(efx, hw, efx->net_dev,
1250 			  "Cannot re-enable PCI device after reset.\n");
1251 		status =  PCI_ERS_RESULT_DISCONNECT;
1252 	}
1253 
1254 	return status;
1255 }
1256 
1257 /* Perform the actual reset and resume I/O operations. */
1258 static void efx_io_resume(struct pci_dev *pdev)
1259 {
1260 	struct efx_nic *efx = pci_get_drvdata(pdev);
1261 	int rc;
1262 
1263 	rtnl_lock();
1264 
1265 	if (efx->state == STATE_DISABLED)
1266 		goto out;
1267 
1268 	rc = efx_reset(efx, RESET_TYPE_ALL);
1269 	if (rc) {
1270 		netif_err(efx, hw, efx->net_dev,
1271 			  "efx_reset failed after PCI error (%d)\n", rc);
1272 	} else {
1273 		efx->state = STATE_READY;
1274 		netif_dbg(efx, hw, efx->net_dev,
1275 			  "Done resetting and resuming IO after PCI error.\n");
1276 	}
1277 
1278 out:
1279 	rtnl_unlock();
1280 }
1281 
1282 /* For simplicity and reliability, we always require a slot reset and try to
1283  * reset the hardware when a pci error affecting the device is detected.
1284  * We leave both the link_reset and mmio_enabled callback unimplemented:
1285  * with our request for slot reset the mmio_enabled callback will never be
1286  * called, and the link_reset callback is not used by AER or EEH mechanisms.
1287  */
1288 const struct pci_error_handlers efx_err_handlers = {
1289 	.error_detected = efx_io_error_detected,
1290 	.slot_reset	= efx_io_slot_reset,
1291 	.resume		= efx_io_resume,
1292 };
1293 
1294 /* Determine whether the NIC will be able to handle TX offloads for a given
1295  * encapsulated packet.
1296  */
1297 static bool efx_can_encap_offloads(struct efx_nic *efx, struct sk_buff *skb)
1298 {
1299 	struct gre_base_hdr *greh;
1300 	__be16 dst_port;
1301 	u8 ipproto;
1302 
1303 	/* Does the NIC support encap offloads?
1304 	 * If not, we should never get here, because we shouldn't have
1305 	 * advertised encap offload feature flags in the first place.
1306 	 */
1307 	if (WARN_ON_ONCE(!efx->type->udp_tnl_has_port))
1308 		return false;
1309 
1310 	/* Determine encapsulation protocol in use */
1311 	switch (skb->protocol) {
1312 	case htons(ETH_P_IP):
1313 		ipproto = ip_hdr(skb)->protocol;
1314 		break;
1315 	case htons(ETH_P_IPV6):
1316 		/* If there are extension headers, this will cause us to
1317 		 * think we can't offload something that we maybe could have.
1318 		 */
1319 		ipproto = ipv6_hdr(skb)->nexthdr;
1320 		break;
1321 	default:
1322 		/* Not IP, so can't offload it */
1323 		return false;
1324 	}
1325 	switch (ipproto) {
1326 	case IPPROTO_GRE:
1327 		/* We support NVGRE but not IP over GRE or random gretaps.
1328 		 * Specifically, the NIC will accept GRE as encapsulated if
1329 		 * the inner protocol is Ethernet, but only handle it
1330 		 * correctly if the GRE header is 8 bytes long.  Moreover,
1331 		 * it will not update the Checksum or Sequence Number fields
1332 		 * if they are present.  (The Routing Present flag,
1333 		 * GRE_ROUTING, cannot be set else the header would be more
1334 		 * than 8 bytes long; so we don't have to worry about it.)
1335 		 */
1336 		if (skb->inner_protocol_type != ENCAP_TYPE_ETHER)
1337 			return false;
1338 		if (ntohs(skb->inner_protocol) != ETH_P_TEB)
1339 			return false;
1340 		if (skb_inner_mac_header(skb) - skb_transport_header(skb) != 8)
1341 			return false;
1342 		greh = (struct gre_base_hdr *)skb_transport_header(skb);
1343 		return !(greh->flags & (GRE_CSUM | GRE_SEQ));
1344 	case IPPROTO_UDP:
1345 		/* If the port is registered for a UDP tunnel, we assume the
1346 		 * packet is for that tunnel, and the NIC will handle it as
1347 		 * such.  If not, the NIC won't know what to do with it.
1348 		 */
1349 		dst_port = udp_hdr(skb)->dest;
1350 		return efx->type->udp_tnl_has_port(efx, dst_port);
1351 	default:
1352 		return false;
1353 	}
1354 }
1355 
1356 netdev_features_t efx_features_check(struct sk_buff *skb, struct net_device *dev,
1357 				     netdev_features_t features)
1358 {
1359 	struct efx_nic *efx = netdev_priv(dev);
1360 
1361 	if (skb->encapsulation) {
1362 		if (features & NETIF_F_GSO_MASK)
1363 			/* Hardware can only do TSO with at most 208 bytes
1364 			 * of headers.
1365 			 */
1366 			if (skb_inner_transport_offset(skb) >
1367 			    EFX_TSO2_MAX_HDRLEN)
1368 				features &= ~(NETIF_F_GSO_MASK);
1369 		if (features & (NETIF_F_GSO_MASK | NETIF_F_CSUM_MASK))
1370 			if (!efx_can_encap_offloads(efx, skb))
1371 				features &= ~(NETIF_F_GSO_MASK |
1372 					      NETIF_F_CSUM_MASK);
1373 	}
1374 	return features;
1375 }
1376 
1377 int efx_get_phys_port_id(struct net_device *net_dev,
1378 			 struct netdev_phys_item_id *ppid)
1379 {
1380 	struct efx_nic *efx = netdev_priv(net_dev);
1381 
1382 	if (efx->type->get_phys_port_id)
1383 		return efx->type->get_phys_port_id(efx, ppid);
1384 	else
1385 		return -EOPNOTSUPP;
1386 }
1387 
1388 int efx_get_phys_port_name(struct net_device *net_dev, char *name, size_t len)
1389 {
1390 	struct efx_nic *efx = netdev_priv(net_dev);
1391 
1392 	if (snprintf(name, len, "p%u", efx->port_num) >= len)
1393 		return -EINVAL;
1394 	return 0;
1395 }
1396