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 "efx_common.h"
15 #include "efx_channels.h"
16 #include "efx.h"
17 #include "mcdi.h"
18 #include "selftest.h"
19 #include "rx_common.h"
20 #include "tx_common.h"
21 #include "nic.h"
22 #include "io.h"
23 #include "mcdi_pcol.h"
24 
25 static unsigned int debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
26 			     NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
27 			     NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
28 			     NETIF_MSG_TX_ERR | NETIF_MSG_HW);
29 module_param(debug, uint, 0);
30 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
31 
32 /* This is the time (in jiffies) between invocations of the hardware
33  * monitor.
34  * On Falcon-based NICs, this will:
35  * - Check the on-board hardware monitor;
36  * - Poll the link state and reconfigure the hardware as necessary.
37  * On Siena-based NICs for power systems with EEH support, this will give EEH a
38  * chance to start.
39  */
40 static unsigned int efx_monitor_interval = 1 * HZ;
41 
42 /* How often and how many times to poll for a reset while waiting for a
43  * BIST that another function started to complete.
44  */
45 #define BIST_WAIT_DELAY_MS	100
46 #define BIST_WAIT_DELAY_COUNT	100
47 
48 /* Default stats update time */
49 #define STATS_PERIOD_MS_DEFAULT 1000
50 
51 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
52 const char *const efx_reset_type_names[] = {
53 	[RESET_TYPE_INVISIBLE]          = "INVISIBLE",
54 	[RESET_TYPE_ALL]                = "ALL",
55 	[RESET_TYPE_RECOVER_OR_ALL]     = "RECOVER_OR_ALL",
56 	[RESET_TYPE_WORLD]              = "WORLD",
57 	[RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
58 	[RESET_TYPE_DATAPATH]           = "DATAPATH",
59 	[RESET_TYPE_MC_BIST]		= "MC_BIST",
60 	[RESET_TYPE_DISABLE]            = "DISABLE",
61 	[RESET_TYPE_TX_WATCHDOG]        = "TX_WATCHDOG",
62 	[RESET_TYPE_INT_ERROR]          = "INT_ERROR",
63 	[RESET_TYPE_DMA_ERROR]          = "DMA_ERROR",
64 	[RESET_TYPE_TX_SKIP]            = "TX_SKIP",
65 	[RESET_TYPE_MC_FAILURE]         = "MC_FAILURE",
66 	[RESET_TYPE_MCDI_TIMEOUT]	= "MCDI_TIMEOUT (FLR)",
67 };
68 
69 #define RESET_TYPE(type) \
70 	STRING_TABLE_LOOKUP(type, efx_reset_type)
71 
72 /* Loopback mode names (see LOOPBACK_MODE()) */
73 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
74 const char *const efx_loopback_mode_names[] = {
75 	[LOOPBACK_NONE]		= "NONE",
76 	[LOOPBACK_DATA]		= "DATAPATH",
77 	[LOOPBACK_GMAC]		= "GMAC",
78 	[LOOPBACK_XGMII]	= "XGMII",
79 	[LOOPBACK_XGXS]		= "XGXS",
80 	[LOOPBACK_XAUI]		= "XAUI",
81 	[LOOPBACK_GMII]		= "GMII",
82 	[LOOPBACK_SGMII]	= "SGMII",
83 	[LOOPBACK_XGBR]		= "XGBR",
84 	[LOOPBACK_XFI]		= "XFI",
85 	[LOOPBACK_XAUI_FAR]	= "XAUI_FAR",
86 	[LOOPBACK_GMII_FAR]	= "GMII_FAR",
87 	[LOOPBACK_SGMII_FAR]	= "SGMII_FAR",
88 	[LOOPBACK_XFI_FAR]	= "XFI_FAR",
89 	[LOOPBACK_GPHY]		= "GPHY",
90 	[LOOPBACK_PHYXS]	= "PHYXS",
91 	[LOOPBACK_PCS]		= "PCS",
92 	[LOOPBACK_PMAPMD]	= "PMA/PMD",
93 	[LOOPBACK_XPORT]	= "XPORT",
94 	[LOOPBACK_XGMII_WS]	= "XGMII_WS",
95 	[LOOPBACK_XAUI_WS]	= "XAUI_WS",
96 	[LOOPBACK_XAUI_WS_FAR]  = "XAUI_WS_FAR",
97 	[LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
98 	[LOOPBACK_GMII_WS]	= "GMII_WS",
99 	[LOOPBACK_XFI_WS]	= "XFI_WS",
100 	[LOOPBACK_XFI_WS_FAR]	= "XFI_WS_FAR",
101 	[LOOPBACK_PHYXS_WS]	= "PHYXS_WS",
102 };
103 
104 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
105  * queued onto this work queue. This is not a per-nic work queue, because
106  * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
107  */
108 static struct workqueue_struct *reset_workqueue;
109 
110 int efx_create_reset_workqueue(void)
111 {
112 	reset_workqueue = create_singlethread_workqueue("sfc_reset");
113 	if (!reset_workqueue) {
114 		printk(KERN_ERR "Failed to create reset workqueue\n");
115 		return -ENOMEM;
116 	}
117 
118 	return 0;
119 }
120 
121 void efx_queue_reset_work(struct efx_nic *efx)
122 {
123 	queue_work(reset_workqueue, &efx->reset_work);
124 }
125 
126 void efx_flush_reset_workqueue(struct efx_nic *efx)
127 {
128 	cancel_work_sync(&efx->reset_work);
129 }
130 
131 void efx_destroy_reset_workqueue(void)
132 {
133 	if (reset_workqueue) {
134 		destroy_workqueue(reset_workqueue);
135 		reset_workqueue = NULL;
136 	}
137 }
138 
139 /* We assume that efx->type->reconfigure_mac will always try to sync RX
140  * filters and therefore needs to read-lock the filter table against freeing
141  */
142 void efx_mac_reconfigure(struct efx_nic *efx)
143 {
144 	if (efx->type->reconfigure_mac) {
145 		down_read(&efx->filter_sem);
146 		efx->type->reconfigure_mac(efx);
147 		up_read(&efx->filter_sem);
148 	}
149 }
150 
151 /* Asynchronous work item for changing MAC promiscuity and multicast
152  * hash.  Avoid a drain/rx_ingress enable by reconfiguring the current
153  * MAC directly.
154  */
155 static void efx_mac_work(struct work_struct *data)
156 {
157 	struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
158 
159 	mutex_lock(&efx->mac_lock);
160 	if (efx->port_enabled)
161 		efx_mac_reconfigure(efx);
162 	mutex_unlock(&efx->mac_lock);
163 }
164 
165 /* This ensures that the kernel is kept informed (via
166  * netif_carrier_on/off) of the link status, and also maintains the
167  * link status's stop on the port's TX queue.
168  */
169 void efx_link_status_changed(struct efx_nic *efx)
170 {
171 	struct efx_link_state *link_state = &efx->link_state;
172 
173 	/* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
174 	 * that no events are triggered between unregister_netdev() and the
175 	 * driver unloading. A more general condition is that NETDEV_CHANGE
176 	 * can only be generated between NETDEV_UP and NETDEV_DOWN
177 	 */
178 	if (!netif_running(efx->net_dev))
179 		return;
180 
181 	if (link_state->up != netif_carrier_ok(efx->net_dev)) {
182 		efx->n_link_state_changes++;
183 
184 		if (link_state->up)
185 			netif_carrier_on(efx->net_dev);
186 		else
187 			netif_carrier_off(efx->net_dev);
188 	}
189 
190 	/* Status message for kernel log */
191 	if (link_state->up)
192 		netif_info(efx, link, efx->net_dev,
193 			   "link up at %uMbps %s-duplex (MTU %d)\n",
194 			   link_state->speed, link_state->fd ? "full" : "half",
195 			   efx->net_dev->mtu);
196 	else
197 		netif_info(efx, link, efx->net_dev, "link down\n");
198 }
199 
200 unsigned int efx_xdp_max_mtu(struct efx_nic *efx)
201 {
202 	/* The maximum MTU that we can fit in a single page, allowing for
203 	 * framing, overhead and XDP headroom.
204 	 */
205 	int overhead = EFX_MAX_FRAME_LEN(0) + sizeof(struct efx_rx_page_state) +
206 		       efx->rx_prefix_size + efx->type->rx_buffer_padding +
207 		       efx->rx_ip_align + XDP_PACKET_HEADROOM;
208 
209 	return PAGE_SIZE - overhead;
210 }
211 
212 /* Context: process, rtnl_lock() held. */
213 int efx_change_mtu(struct net_device *net_dev, int new_mtu)
214 {
215 	struct efx_nic *efx = netdev_priv(net_dev);
216 	int rc;
217 
218 	rc = efx_check_disabled(efx);
219 	if (rc)
220 		return rc;
221 
222 	if (rtnl_dereference(efx->xdp_prog) &&
223 	    new_mtu > efx_xdp_max_mtu(efx)) {
224 		netif_err(efx, drv, efx->net_dev,
225 			  "Requested MTU of %d too big for XDP (max: %d)\n",
226 			  new_mtu, efx_xdp_max_mtu(efx));
227 		return -EINVAL;
228 	}
229 
230 	netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
231 
232 	efx_device_detach_sync(efx);
233 	efx_stop_all(efx);
234 
235 	mutex_lock(&efx->mac_lock);
236 	net_dev->mtu = new_mtu;
237 	efx_mac_reconfigure(efx);
238 	mutex_unlock(&efx->mac_lock);
239 
240 	efx_start_all(efx);
241 	efx_device_attach_if_not_resetting(efx);
242 	return 0;
243 }
244 
245 /**************************************************************************
246  *
247  * Hardware monitor
248  *
249  **************************************************************************/
250 
251 /* Run periodically off the general workqueue */
252 static void efx_monitor(struct work_struct *data)
253 {
254 	struct efx_nic *efx = container_of(data, struct efx_nic,
255 					   monitor_work.work);
256 
257 	netif_vdbg(efx, timer, efx->net_dev,
258 		   "hardware monitor executing on CPU %d\n",
259 		   raw_smp_processor_id());
260 	BUG_ON(efx->type->monitor == NULL);
261 
262 	/* If the mac_lock is already held then it is likely a port
263 	 * reconfiguration is already in place, which will likely do
264 	 * most of the work of monitor() anyway.
265 	 */
266 	if (mutex_trylock(&efx->mac_lock)) {
267 		if (efx->port_enabled && efx->type->monitor)
268 			efx->type->monitor(efx);
269 		mutex_unlock(&efx->mac_lock);
270 	}
271 
272 	efx_start_monitor(efx);
273 }
274 
275 void efx_start_monitor(struct efx_nic *efx)
276 {
277 	if (efx->type->monitor)
278 		queue_delayed_work(efx->workqueue, &efx->monitor_work,
279 				   efx_monitor_interval);
280 }
281 
282 /**************************************************************************
283  *
284  * Event queue processing
285  *
286  *************************************************************************/
287 
288 /* Channels are shutdown and reinitialised whilst the NIC is running
289  * to propagate configuration changes (mtu, checksum offload), or
290  * to clear hardware error conditions
291  */
292 static void efx_start_datapath(struct efx_nic *efx)
293 {
294 	netdev_features_t old_features = efx->net_dev->features;
295 	bool old_rx_scatter = efx->rx_scatter;
296 	size_t rx_buf_len;
297 
298 	/* Calculate the rx buffer allocation parameters required to
299 	 * support the current MTU, including padding for header
300 	 * alignment and overruns.
301 	 */
302 	efx->rx_dma_len = (efx->rx_prefix_size +
303 			   EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
304 			   efx->type->rx_buffer_padding);
305 	rx_buf_len = (sizeof(struct efx_rx_page_state) + XDP_PACKET_HEADROOM +
306 		      efx->rx_ip_align + efx->rx_dma_len);
307 	if (rx_buf_len <= PAGE_SIZE) {
308 		efx->rx_scatter = efx->type->always_rx_scatter;
309 		efx->rx_buffer_order = 0;
310 	} else if (efx->type->can_rx_scatter) {
311 		BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
312 		BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
313 			     2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
314 				       EFX_RX_BUF_ALIGNMENT) >
315 			     PAGE_SIZE);
316 		efx->rx_scatter = true;
317 		efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
318 		efx->rx_buffer_order = 0;
319 	} else {
320 		efx->rx_scatter = false;
321 		efx->rx_buffer_order = get_order(rx_buf_len);
322 	}
323 
324 	efx_rx_config_page_split(efx);
325 	if (efx->rx_buffer_order)
326 		netif_dbg(efx, drv, efx->net_dev,
327 			  "RX buf len=%u; page order=%u batch=%u\n",
328 			  efx->rx_dma_len, efx->rx_buffer_order,
329 			  efx->rx_pages_per_batch);
330 	else
331 		netif_dbg(efx, drv, efx->net_dev,
332 			  "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
333 			  efx->rx_dma_len, efx->rx_page_buf_step,
334 			  efx->rx_bufs_per_page, efx->rx_pages_per_batch);
335 
336 	/* Restore previously fixed features in hw_features and remove
337 	 * features which are fixed now
338 	 */
339 	efx->net_dev->hw_features |= efx->net_dev->features;
340 	efx->net_dev->hw_features &= ~efx->fixed_features;
341 	efx->net_dev->features |= efx->fixed_features;
342 	if (efx->net_dev->features != old_features)
343 		netdev_features_change(efx->net_dev);
344 
345 	/* RX filters may also have scatter-enabled flags */
346 	if ((efx->rx_scatter != old_rx_scatter) &&
347 	    efx->type->filter_update_rx_scatter)
348 		efx->type->filter_update_rx_scatter(efx);
349 
350 	/* We must keep at least one descriptor in a TX ring empty.
351 	 * We could avoid this when the queue size does not exactly
352 	 * match the hardware ring size, but it's not that important.
353 	 * Therefore we stop the queue when one more skb might fill
354 	 * the ring completely.  We wake it when half way back to
355 	 * empty.
356 	 */
357 	efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
358 	efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
359 
360 	/* Initialise the channels */
361 	efx_start_channels(efx);
362 
363 	efx_ptp_start_datapath(efx);
364 
365 	if (netif_device_present(efx->net_dev))
366 		netif_tx_wake_all_queues(efx->net_dev);
367 }
368 
369 static void efx_stop_datapath(struct efx_nic *efx)
370 {
371 	EFX_ASSERT_RESET_SERIALISED(efx);
372 	BUG_ON(efx->port_enabled);
373 
374 	efx_ptp_stop_datapath(efx);
375 
376 	efx_stop_channels(efx);
377 }
378 
379 /**************************************************************************
380  *
381  * Port handling
382  *
383  **************************************************************************/
384 
385 static void efx_start_port(struct efx_nic *efx)
386 {
387 	netif_dbg(efx, ifup, efx->net_dev, "start port\n");
388 	BUG_ON(efx->port_enabled);
389 
390 	mutex_lock(&efx->mac_lock);
391 	efx->port_enabled = true;
392 
393 	/* Ensure MAC ingress/egress is enabled */
394 	efx_mac_reconfigure(efx);
395 
396 	mutex_unlock(&efx->mac_lock);
397 }
398 
399 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
400  * and the async self-test, wait for them to finish and prevent them
401  * being scheduled again.  This doesn't cover online resets, which
402  * should only be cancelled when removing the device.
403  */
404 static void efx_stop_port(struct efx_nic *efx)
405 {
406 	netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
407 
408 	EFX_ASSERT_RESET_SERIALISED(efx);
409 
410 	mutex_lock(&efx->mac_lock);
411 	efx->port_enabled = false;
412 	mutex_unlock(&efx->mac_lock);
413 
414 	/* Serialise against efx_set_multicast_list() */
415 	netif_addr_lock_bh(efx->net_dev);
416 	netif_addr_unlock_bh(efx->net_dev);
417 
418 	cancel_delayed_work_sync(&efx->monitor_work);
419 	efx_selftest_async_cancel(efx);
420 	cancel_work_sync(&efx->mac_work);
421 }
422 
423 /* If the interface is supposed to be running but is not, start
424  * the hardware and software data path, regular activity for the port
425  * (MAC statistics, link polling, etc.) and schedule the port to be
426  * reconfigured.  Interrupts must already be enabled.  This function
427  * is safe to call multiple times, so long as the NIC is not disabled.
428  * Requires the RTNL lock.
429  */
430 void efx_start_all(struct efx_nic *efx)
431 {
432 	EFX_ASSERT_RESET_SERIALISED(efx);
433 	BUG_ON(efx->state == STATE_DISABLED);
434 
435 	/* Check that it is appropriate to restart the interface. All
436 	 * of these flags are safe to read under just the rtnl lock
437 	 */
438 	if (efx->port_enabled || !netif_running(efx->net_dev) ||
439 	    efx->reset_pending)
440 		return;
441 
442 	efx_start_port(efx);
443 	efx_start_datapath(efx);
444 
445 	/* Start the hardware monitor if there is one */
446 	efx_start_monitor(efx);
447 
448 	/* Link state detection is normally event-driven; we have
449 	 * to poll now because we could have missed a change
450 	 */
451 	mutex_lock(&efx->mac_lock);
452 	if (efx->phy_op->poll(efx))
453 		efx_link_status_changed(efx);
454 	mutex_unlock(&efx->mac_lock);
455 
456 	if (efx->type->start_stats) {
457 		efx->type->start_stats(efx);
458 		efx->type->pull_stats(efx);
459 		spin_lock_bh(&efx->stats_lock);
460 		efx->type->update_stats(efx, NULL, NULL);
461 		spin_unlock_bh(&efx->stats_lock);
462 	}
463 }
464 
465 /* Quiesce the hardware and software data path, and regular activity
466  * for the port without bringing the link down.  Safe to call multiple
467  * times with the NIC in almost any state, but interrupts should be
468  * enabled.  Requires the RTNL lock.
469  */
470 void efx_stop_all(struct efx_nic *efx)
471 {
472 	EFX_ASSERT_RESET_SERIALISED(efx);
473 
474 	/* port_enabled can be read safely under the rtnl lock */
475 	if (!efx->port_enabled)
476 		return;
477 
478 	if (efx->type->update_stats) {
479 		/* update stats before we go down so we can accurately count
480 		 * rx_nodesc_drops
481 		 */
482 		efx->type->pull_stats(efx);
483 		spin_lock_bh(&efx->stats_lock);
484 		efx->type->update_stats(efx, NULL, NULL);
485 		spin_unlock_bh(&efx->stats_lock);
486 		efx->type->stop_stats(efx);
487 	}
488 
489 	efx_stop_port(efx);
490 
491 	/* Stop the kernel transmit interface.  This is only valid if
492 	 * the device is stopped or detached; otherwise the watchdog
493 	 * may fire immediately.
494 	 */
495 	WARN_ON(netif_running(efx->net_dev) &&
496 		netif_device_present(efx->net_dev));
497 	netif_tx_disable(efx->net_dev);
498 
499 	efx_stop_datapath(efx);
500 }
501 
502 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
503 void efx_net_stats(struct net_device *net_dev, struct rtnl_link_stats64 *stats)
504 {
505 	struct efx_nic *efx = netdev_priv(net_dev);
506 
507 	spin_lock_bh(&efx->stats_lock);
508 	efx->type->update_stats(efx, NULL, stats);
509 	spin_unlock_bh(&efx->stats_lock);
510 }
511 
512 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
513  * the MAC appropriately. All other PHY configuration changes are pushed
514  * through phy_op->set_settings(), and pushed asynchronously to the MAC
515  * through efx_monitor().
516  *
517  * Callers must hold the mac_lock
518  */
519 int __efx_reconfigure_port(struct efx_nic *efx)
520 {
521 	enum efx_phy_mode phy_mode;
522 	int rc = 0;
523 
524 	WARN_ON(!mutex_is_locked(&efx->mac_lock));
525 
526 	/* Disable PHY transmit in mac level loopbacks */
527 	phy_mode = efx->phy_mode;
528 	if (LOOPBACK_INTERNAL(efx))
529 		efx->phy_mode |= PHY_MODE_TX_DISABLED;
530 	else
531 		efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
532 
533 	if (efx->type->reconfigure_port)
534 		rc = efx->type->reconfigure_port(efx);
535 
536 	if (rc)
537 		efx->phy_mode = phy_mode;
538 
539 	return rc;
540 }
541 
542 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
543  * disabled.
544  */
545 int efx_reconfigure_port(struct efx_nic *efx)
546 {
547 	int rc;
548 
549 	EFX_ASSERT_RESET_SERIALISED(efx);
550 
551 	mutex_lock(&efx->mac_lock);
552 	rc = __efx_reconfigure_port(efx);
553 	mutex_unlock(&efx->mac_lock);
554 
555 	return rc;
556 }
557 
558 /**************************************************************************
559  *
560  * Device reset and suspend
561  *
562  **************************************************************************/
563 
564 static void efx_wait_for_bist_end(struct efx_nic *efx)
565 {
566 	int i;
567 
568 	for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
569 		if (efx_mcdi_poll_reboot(efx))
570 			goto out;
571 		msleep(BIST_WAIT_DELAY_MS);
572 	}
573 
574 	netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
575 out:
576 	/* Either way unset the BIST flag. If we found no reboot we probably
577 	 * won't recover, but we should try.
578 	 */
579 	efx->mc_bist_for_other_fn = false;
580 }
581 
582 /* Try recovery mechanisms.
583  * For now only EEH is supported.
584  * Returns 0 if the recovery mechanisms are unsuccessful.
585  * Returns a non-zero value otherwise.
586  */
587 int efx_try_recovery(struct efx_nic *efx)
588 {
589 #ifdef CONFIG_EEH
590 	/* A PCI error can occur and not be seen by EEH because nothing
591 	 * happens on the PCI bus. In this case the driver may fail and
592 	 * schedule a 'recover or reset', leading to this recovery handler.
593 	 * Manually call the eeh failure check function.
594 	 */
595 	struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
596 	if (eeh_dev_check_failure(eehdev)) {
597 		/* The EEH mechanisms will handle the error and reset the
598 		 * device if necessary.
599 		 */
600 		return 1;
601 	}
602 #endif
603 	return 0;
604 }
605 
606 /* Tears down the entire software state and most of the hardware state
607  * before reset.
608  */
609 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
610 {
611 	EFX_ASSERT_RESET_SERIALISED(efx);
612 
613 	if (method == RESET_TYPE_MCDI_TIMEOUT)
614 		efx->type->prepare_flr(efx);
615 
616 	efx_stop_all(efx);
617 	efx_disable_interrupts(efx);
618 
619 	mutex_lock(&efx->mac_lock);
620 	down_write(&efx->filter_sem);
621 	mutex_lock(&efx->rss_lock);
622 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
623 	    method != RESET_TYPE_DATAPATH)
624 		efx->phy_op->fini(efx);
625 	efx->type->fini(efx);
626 }
627 
628 /* This function will always ensure that the locks acquired in
629  * efx_reset_down() are released. A failure return code indicates
630  * that we were unable to reinitialise the hardware, and the
631  * driver should be disabled. If ok is false, then the rx and tx
632  * engines are not restarted, pending a RESET_DISABLE.
633  */
634 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
635 {
636 	int rc;
637 
638 	EFX_ASSERT_RESET_SERIALISED(efx);
639 
640 	if (method == RESET_TYPE_MCDI_TIMEOUT)
641 		efx->type->finish_flr(efx);
642 
643 	/* Ensure that SRAM is initialised even if we're disabling the device */
644 	rc = efx->type->init(efx);
645 	if (rc) {
646 		netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
647 		goto fail;
648 	}
649 
650 	if (!ok)
651 		goto fail;
652 
653 	if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
654 	    method != RESET_TYPE_DATAPATH) {
655 		rc = efx->phy_op->init(efx);
656 		if (rc)
657 			goto fail;
658 		rc = efx->phy_op->reconfigure(efx);
659 		if (rc && rc != -EPERM)
660 			netif_err(efx, drv, efx->net_dev,
661 				  "could not restore PHY settings\n");
662 	}
663 
664 	rc = efx_enable_interrupts(efx);
665 	if (rc)
666 		goto fail;
667 
668 #ifdef CONFIG_SFC_SRIOV
669 	rc = efx->type->vswitching_restore(efx);
670 	if (rc) /* not fatal; the PF will still work fine */
671 		netif_warn(efx, probe, efx->net_dev,
672 			   "failed to restore vswitching rc=%d;"
673 			   " VFs may not function\n", rc);
674 #endif
675 
676 	if (efx->type->rx_restore_rss_contexts)
677 		efx->type->rx_restore_rss_contexts(efx);
678 	mutex_unlock(&efx->rss_lock);
679 	efx->type->filter_table_restore(efx);
680 	up_write(&efx->filter_sem);
681 	if (efx->type->sriov_reset)
682 		efx->type->sriov_reset(efx);
683 
684 	mutex_unlock(&efx->mac_lock);
685 
686 	efx_start_all(efx);
687 
688 	if (efx->type->udp_tnl_push_ports)
689 		efx->type->udp_tnl_push_ports(efx);
690 
691 	return 0;
692 
693 fail:
694 	efx->port_initialized = false;
695 
696 	mutex_unlock(&efx->rss_lock);
697 	up_write(&efx->filter_sem);
698 	mutex_unlock(&efx->mac_lock);
699 
700 	return rc;
701 }
702 
703 /* Reset the NIC using the specified method.  Note that the reset may
704  * fail, in which case the card will be left in an unusable state.
705  *
706  * Caller must hold the rtnl_lock.
707  */
708 int efx_reset(struct efx_nic *efx, enum reset_type method)
709 {
710 	bool disabled;
711 	int rc, rc2;
712 
713 	netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
714 		   RESET_TYPE(method));
715 
716 	efx_device_detach_sync(efx);
717 	efx_reset_down(efx, method);
718 
719 	rc = efx->type->reset(efx, method);
720 	if (rc) {
721 		netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
722 		goto out;
723 	}
724 
725 	/* Clear flags for the scopes we covered.  We assume the NIC and
726 	 * driver are now quiescent so that there is no race here.
727 	 */
728 	if (method < RESET_TYPE_MAX_METHOD)
729 		efx->reset_pending &= -(1 << (method + 1));
730 	else /* it doesn't fit into the well-ordered scope hierarchy */
731 		__clear_bit(method, &efx->reset_pending);
732 
733 	/* Reinitialise bus-mastering, which may have been turned off before
734 	 * the reset was scheduled. This is still appropriate, even in the
735 	 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
736 	 * can respond to requests.
737 	 */
738 	pci_set_master(efx->pci_dev);
739 
740 out:
741 	/* Leave device stopped if necessary */
742 	disabled = rc ||
743 		method == RESET_TYPE_DISABLE ||
744 		method == RESET_TYPE_RECOVER_OR_DISABLE;
745 	rc2 = efx_reset_up(efx, method, !disabled);
746 	if (rc2) {
747 		disabled = true;
748 		if (!rc)
749 			rc = rc2;
750 	}
751 
752 	if (disabled) {
753 		dev_close(efx->net_dev);
754 		netif_err(efx, drv, efx->net_dev, "has been disabled\n");
755 		efx->state = STATE_DISABLED;
756 	} else {
757 		netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
758 		efx_device_attach_if_not_resetting(efx);
759 	}
760 	return rc;
761 }
762 
763 /* The worker thread exists so that code that cannot sleep can
764  * schedule a reset for later.
765  */
766 static void efx_reset_work(struct work_struct *data)
767 {
768 	struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
769 	unsigned long pending;
770 	enum reset_type method;
771 
772 	pending = READ_ONCE(efx->reset_pending);
773 	method = fls(pending) - 1;
774 
775 	if (method == RESET_TYPE_MC_BIST)
776 		efx_wait_for_bist_end(efx);
777 
778 	if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
779 	     method == RESET_TYPE_RECOVER_OR_ALL) &&
780 	    efx_try_recovery(efx))
781 		return;
782 
783 	if (!pending)
784 		return;
785 
786 	rtnl_lock();
787 
788 	/* We checked the state in efx_schedule_reset() but it may
789 	 * have changed by now.  Now that we have the RTNL lock,
790 	 * it cannot change again.
791 	 */
792 	if (efx->state == STATE_READY)
793 		(void)efx_reset(efx, method);
794 
795 	rtnl_unlock();
796 }
797 
798 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
799 {
800 	enum reset_type method;
801 
802 	if (efx->state == STATE_RECOVERY) {
803 		netif_dbg(efx, drv, efx->net_dev,
804 			  "recovering: skip scheduling %s reset\n",
805 			  RESET_TYPE(type));
806 		return;
807 	}
808 
809 	switch (type) {
810 	case RESET_TYPE_INVISIBLE:
811 	case RESET_TYPE_ALL:
812 	case RESET_TYPE_RECOVER_OR_ALL:
813 	case RESET_TYPE_WORLD:
814 	case RESET_TYPE_DISABLE:
815 	case RESET_TYPE_RECOVER_OR_DISABLE:
816 	case RESET_TYPE_DATAPATH:
817 	case RESET_TYPE_MC_BIST:
818 	case RESET_TYPE_MCDI_TIMEOUT:
819 		method = type;
820 		netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
821 			  RESET_TYPE(method));
822 		break;
823 	default:
824 		method = efx->type->map_reset_reason(type);
825 		netif_dbg(efx, drv, efx->net_dev,
826 			  "scheduling %s reset for %s\n",
827 			  RESET_TYPE(method), RESET_TYPE(type));
828 		break;
829 	}
830 
831 	set_bit(method, &efx->reset_pending);
832 	smp_mb(); /* ensure we change reset_pending before checking state */
833 
834 	/* If we're not READY then just leave the flags set as the cue
835 	 * to abort probing or reschedule the reset later.
836 	 */
837 	if (READ_ONCE(efx->state) != STATE_READY)
838 		return;
839 
840 	/* efx_process_channel() will no longer read events once a
841 	 * reset is scheduled. So switch back to poll'd MCDI completions.
842 	 */
843 	efx_mcdi_mode_poll(efx);
844 
845 	efx_queue_reset_work(efx);
846 }
847 
848 /**************************************************************************
849  *
850  * Dummy PHY/MAC operations
851  *
852  * Can be used for some unimplemented operations
853  * Needed so all function pointers are valid and do not have to be tested
854  * before use
855  *
856  **************************************************************************/
857 int efx_port_dummy_op_int(struct efx_nic *efx)
858 {
859 	return 0;
860 }
861 void efx_port_dummy_op_void(struct efx_nic *efx) {}
862 
863 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
864 {
865 	return false;
866 }
867 
868 static const struct efx_phy_operations efx_dummy_phy_operations = {
869 	.init		 = efx_port_dummy_op_int,
870 	.reconfigure	 = efx_port_dummy_op_int,
871 	.poll		 = efx_port_dummy_op_poll,
872 	.fini		 = efx_port_dummy_op_void,
873 };
874 
875 /**************************************************************************
876  *
877  * Data housekeeping
878  *
879  **************************************************************************/
880 
881 /* This zeroes out and then fills in the invariants in a struct
882  * efx_nic (including all sub-structures).
883  */
884 int efx_init_struct(struct efx_nic *efx,
885 		    struct pci_dev *pci_dev, struct net_device *net_dev)
886 {
887 	int rc = -ENOMEM;
888 
889 	/* Initialise common structures */
890 	INIT_LIST_HEAD(&efx->node);
891 	INIT_LIST_HEAD(&efx->secondary_list);
892 	spin_lock_init(&efx->biu_lock);
893 #ifdef CONFIG_SFC_MTD
894 	INIT_LIST_HEAD(&efx->mtd_list);
895 #endif
896 	INIT_WORK(&efx->reset_work, efx_reset_work);
897 	INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
898 	efx_selftest_async_init(efx);
899 	efx->pci_dev = pci_dev;
900 	efx->msg_enable = debug;
901 	efx->state = STATE_UNINIT;
902 	strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
903 
904 	efx->net_dev = net_dev;
905 	efx->rx_prefix_size = efx->type->rx_prefix_size;
906 	efx->rx_ip_align =
907 		NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
908 	efx->rx_packet_hash_offset =
909 		efx->type->rx_hash_offset - efx->type->rx_prefix_size;
910 	efx->rx_packet_ts_offset =
911 		efx->type->rx_ts_offset - efx->type->rx_prefix_size;
912 	INIT_LIST_HEAD(&efx->rss_context.list);
913 	mutex_init(&efx->rss_lock);
914 	spin_lock_init(&efx->stats_lock);
915 	efx->vi_stride = EFX_DEFAULT_VI_STRIDE;
916 	efx->num_mac_stats = MC_CMD_MAC_NSTATS;
917 	BUILD_BUG_ON(MC_CMD_MAC_NSTATS - 1 != MC_CMD_MAC_GENERATION_END);
918 	mutex_init(&efx->mac_lock);
919 #ifdef CONFIG_RFS_ACCEL
920 	mutex_init(&efx->rps_mutex);
921 	spin_lock_init(&efx->rps_hash_lock);
922 	/* Failure to allocate is not fatal, but may degrade ARFS performance */
923 	efx->rps_hash_table = kcalloc(EFX_ARFS_HASH_TABLE_SIZE,
924 				      sizeof(*efx->rps_hash_table), GFP_KERNEL);
925 #endif
926 	efx->phy_op = &efx_dummy_phy_operations;
927 	efx->mdio.dev = net_dev;
928 	INIT_WORK(&efx->mac_work, efx_mac_work);
929 	init_waitqueue_head(&efx->flush_wq);
930 
931 	rc = efx_init_channels(efx);
932 	if (rc)
933 		goto fail;
934 
935 	/* Would be good to use the net_dev name, but we're too early */
936 	snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
937 		 pci_name(pci_dev));
938 	efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
939 	if (!efx->workqueue) {
940 		rc = -ENOMEM;
941 		goto fail;
942 	}
943 
944 	return 0;
945 
946 fail:
947 	efx_fini_struct(efx);
948 	return rc;
949 }
950 
951 void efx_fini_struct(struct efx_nic *efx)
952 {
953 #ifdef CONFIG_RFS_ACCEL
954 	kfree(efx->rps_hash_table);
955 #endif
956 
957 	efx_fini_channels(efx);
958 
959 	kfree(efx->vpd_sn);
960 
961 	if (efx->workqueue) {
962 		destroy_workqueue(efx->workqueue);
963 		efx->workqueue = NULL;
964 	}
965 }
966 
967 /* This configures the PCI device to enable I/O and DMA. */
968 int efx_init_io(struct efx_nic *efx, int bar, dma_addr_t dma_mask,
969 		unsigned int mem_map_size)
970 {
971 	struct pci_dev *pci_dev = efx->pci_dev;
972 	int rc;
973 
974 	netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
975 
976 	rc = pci_enable_device(pci_dev);
977 	if (rc) {
978 		netif_err(efx, probe, efx->net_dev,
979 			  "failed to enable PCI device\n");
980 		goto fail1;
981 	}
982 
983 	pci_set_master(pci_dev);
984 
985 	/* Set the PCI DMA mask.  Try all possibilities from our
986 	 * genuine mask down to 32 bits, because some architectures
987 	 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
988 	 * masks event though they reject 46 bit masks.
989 	 */
990 	while (dma_mask > 0x7fffffffUL) {
991 		rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
992 		if (rc == 0)
993 			break;
994 		dma_mask >>= 1;
995 	}
996 	if (rc) {
997 		netif_err(efx, probe, efx->net_dev,
998 			  "could not find a suitable DMA mask\n");
999 		goto fail2;
1000 	}
1001 	netif_dbg(efx, probe, efx->net_dev,
1002 		  "using DMA mask %llx\n", (unsigned long long)dma_mask);
1003 
1004 	efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1005 	if (!efx->membase_phys) {
1006 		netif_err(efx, probe, efx->net_dev,
1007 			  "ERROR: No BAR%d mapping from the BIOS. "
1008 			  "Try pci=realloc on the kernel command line\n", bar);
1009 		rc = -ENODEV;
1010 		goto fail3;
1011 	}
1012 
1013 	rc = pci_request_region(pci_dev, bar, "sfc");
1014 	if (rc) {
1015 		netif_err(efx, probe, efx->net_dev,
1016 			  "request for memory BAR failed\n");
1017 		rc = -EIO;
1018 		goto fail3;
1019 	}
1020 
1021 	efx->membase = ioremap(efx->membase_phys, mem_map_size);
1022 	if (!efx->membase) {
1023 		netif_err(efx, probe, efx->net_dev,
1024 			  "could not map memory BAR at %llx+%x\n",
1025 			  (unsigned long long)efx->membase_phys, mem_map_size);
1026 		rc = -ENOMEM;
1027 		goto fail4;
1028 	}
1029 	netif_dbg(efx, probe, efx->net_dev,
1030 		  "memory BAR at %llx+%x (virtual %p)\n",
1031 		  (unsigned long long)efx->membase_phys, mem_map_size,
1032 		  efx->membase);
1033 
1034 	return 0;
1035 
1036 fail4:
1037 	pci_release_region(efx->pci_dev, bar);
1038 fail3:
1039 	efx->membase_phys = 0;
1040 fail2:
1041 	pci_disable_device(efx->pci_dev);
1042 fail1:
1043 	return rc;
1044 }
1045 
1046 void efx_fini_io(struct efx_nic *efx, int bar)
1047 {
1048 	netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1049 
1050 	if (efx->membase) {
1051 		iounmap(efx->membase);
1052 		efx->membase = NULL;
1053 	}
1054 
1055 	if (efx->membase_phys) {
1056 		pci_release_region(efx->pci_dev, bar);
1057 		efx->membase_phys = 0;
1058 	}
1059 
1060 	/* Don't disable bus-mastering if VFs are assigned */
1061 	if (!pci_vfs_assigned(efx->pci_dev))
1062 		pci_disable_device(efx->pci_dev);
1063 }
1064 
1065 #ifdef CONFIG_SFC_MCDI_LOGGING
1066 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
1067 			     char *buf)
1068 {
1069 	struct efx_nic *efx = dev_get_drvdata(dev);
1070 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1071 
1072 	return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
1073 }
1074 
1075 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
1076 			    const char *buf, size_t count)
1077 {
1078 	struct efx_nic *efx = dev_get_drvdata(dev);
1079 	struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1080 	bool enable = count > 0 && *buf != '0';
1081 
1082 	mcdi->logging_enabled = enable;
1083 	return count;
1084 }
1085 
1086 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
1087 
1088 void efx_init_mcdi_logging(struct efx_nic *efx)
1089 {
1090 	int rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1091 
1092 	if (rc) {
1093 		netif_warn(efx, drv, efx->net_dev,
1094 			   "failed to init net dev attributes\n");
1095 	}
1096 }
1097 
1098 void efx_fini_mcdi_logging(struct efx_nic *efx)
1099 {
1100 	device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1101 }
1102 #endif
1103