Loading...
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/filter.h>
13#include <linux/module.h>
14#include <linux/netdevice.h>
15#include <net/gre.h>
16#include "efx_common.h"
17#include "efx_channels.h"
18#include "efx.h"
19#include "mcdi.h"
20#include "selftest.h"
21#include "rx_common.h"
22#include "tx_common.h"
23#include "nic.h"
24#include "mcdi_port_common.h"
25#include "io.h"
26#include "mcdi_pcol.h"
27#include "ef100_rep.h"
28
29static unsigned int debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
30 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
31 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
32 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
33module_param(debug, uint, 0);
34MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
35
36/* This is the time (in jiffies) between invocations of the hardware
37 * monitor.
38 */
39static unsigned int efx_monitor_interval = 1 * HZ;
40
41/* How often and how many times to poll for a reset while waiting for a
42 * BIST that another function started to complete.
43 */
44#define BIST_WAIT_DELAY_MS 100
45#define BIST_WAIT_DELAY_COUNT 100
46
47/* Default stats update time */
48#define STATS_PERIOD_MS_DEFAULT 1000
49
50static const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
51static const char *const efx_reset_type_names[] = {
52 [RESET_TYPE_INVISIBLE] = "INVISIBLE",
53 [RESET_TYPE_ALL] = "ALL",
54 [RESET_TYPE_RECOVER_OR_ALL] = "RECOVER_OR_ALL",
55 [RESET_TYPE_WORLD] = "WORLD",
56 [RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
57 [RESET_TYPE_DATAPATH] = "DATAPATH",
58 [RESET_TYPE_MC_BIST] = "MC_BIST",
59 [RESET_TYPE_DISABLE] = "DISABLE",
60 [RESET_TYPE_TX_WATCHDOG] = "TX_WATCHDOG",
61 [RESET_TYPE_INT_ERROR] = "INT_ERROR",
62 [RESET_TYPE_DMA_ERROR] = "DMA_ERROR",
63 [RESET_TYPE_TX_SKIP] = "TX_SKIP",
64 [RESET_TYPE_MC_FAILURE] = "MC_FAILURE",
65 [RESET_TYPE_MCDI_TIMEOUT] = "MCDI_TIMEOUT (FLR)",
66};
67
68#define RESET_TYPE(type) \
69 STRING_TABLE_LOOKUP(type, efx_reset_type)
70
71/* Loopback mode names (see LOOPBACK_MODE()) */
72const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
73const char *const efx_loopback_mode_names[] = {
74 [LOOPBACK_NONE] = "NONE",
75 [LOOPBACK_DATA] = "DATAPATH",
76 [LOOPBACK_GMAC] = "GMAC",
77 [LOOPBACK_XGMII] = "XGMII",
78 [LOOPBACK_XGXS] = "XGXS",
79 [LOOPBACK_XAUI] = "XAUI",
80 [LOOPBACK_GMII] = "GMII",
81 [LOOPBACK_SGMII] = "SGMII",
82 [LOOPBACK_XGBR] = "XGBR",
83 [LOOPBACK_XFI] = "XFI",
84 [LOOPBACK_XAUI_FAR] = "XAUI_FAR",
85 [LOOPBACK_GMII_FAR] = "GMII_FAR",
86 [LOOPBACK_SGMII_FAR] = "SGMII_FAR",
87 [LOOPBACK_XFI_FAR] = "XFI_FAR",
88 [LOOPBACK_GPHY] = "GPHY",
89 [LOOPBACK_PHYXS] = "PHYXS",
90 [LOOPBACK_PCS] = "PCS",
91 [LOOPBACK_PMAPMD] = "PMA/PMD",
92 [LOOPBACK_XPORT] = "XPORT",
93 [LOOPBACK_XGMII_WS] = "XGMII_WS",
94 [LOOPBACK_XAUI_WS] = "XAUI_WS",
95 [LOOPBACK_XAUI_WS_FAR] = "XAUI_WS_FAR",
96 [LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
97 [LOOPBACK_GMII_WS] = "GMII_WS",
98 [LOOPBACK_XFI_WS] = "XFI_WS",
99 [LOOPBACK_XFI_WS_FAR] = "XFI_WS_FAR",
100 [LOOPBACK_PHYXS_WS] = "PHYXS_WS",
101};
102
103/* Reset workqueue. If any NIC has a hardware failure then a reset will be
104 * queued onto this work queue. This is not a per-nic work queue, because
105 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
106 */
107static struct workqueue_struct *reset_workqueue;
108
109int efx_create_reset_workqueue(void)
110{
111 reset_workqueue = create_singlethread_workqueue("sfc_reset");
112 if (!reset_workqueue) {
113 printk(KERN_ERR "Failed to create reset workqueue\n");
114 return -ENOMEM;
115 }
116
117 return 0;
118}
119
120void efx_queue_reset_work(struct efx_nic *efx)
121{
122 queue_work(reset_workqueue, &efx->reset_work);
123}
124
125void efx_flush_reset_workqueue(struct efx_nic *efx)
126{
127 cancel_work_sync(&efx->reset_work);
128}
129
130void efx_destroy_reset_workqueue(void)
131{
132 if (reset_workqueue) {
133 destroy_workqueue(reset_workqueue);
134 reset_workqueue = NULL;
135 }
136}
137
138/* We assume that efx->type->reconfigure_mac will always try to sync RX
139 * filters and therefore needs to read-lock the filter table against freeing
140 */
141void efx_mac_reconfigure(struct efx_nic *efx, bool mtu_only)
142{
143 if (efx->type->reconfigure_mac) {
144 down_read(&efx->filter_sem);
145 efx->type->reconfigure_mac(efx, mtu_only);
146 up_read(&efx->filter_sem);
147 }
148}
149
150/* Asynchronous work item for changing MAC promiscuity and multicast
151 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
152 * MAC directly.
153 */
154static void efx_mac_work(struct work_struct *data)
155{
156 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
157
158 mutex_lock(&efx->mac_lock);
159 if (efx->port_enabled)
160 efx_mac_reconfigure(efx, false);
161 mutex_unlock(&efx->mac_lock);
162}
163
164int efx_set_mac_address(struct net_device *net_dev, void *data)
165{
166 struct efx_nic *efx = efx_netdev_priv(net_dev);
167 struct sockaddr *addr = data;
168 u8 *new_addr = addr->sa_data;
169 u8 old_addr[6];
170 int rc;
171
172 if (!is_valid_ether_addr(new_addr)) {
173 netif_err(efx, drv, efx->net_dev,
174 "invalid ethernet MAC address requested: %pM\n",
175 new_addr);
176 return -EADDRNOTAVAIL;
177 }
178
179 /* save old address */
180 ether_addr_copy(old_addr, net_dev->dev_addr);
181 eth_hw_addr_set(net_dev, new_addr);
182 if (efx->type->set_mac_address) {
183 rc = efx->type->set_mac_address(efx);
184 if (rc) {
185 eth_hw_addr_set(net_dev, old_addr);
186 return rc;
187 }
188 }
189
190 /* Reconfigure the MAC */
191 mutex_lock(&efx->mac_lock);
192 efx_mac_reconfigure(efx, false);
193 mutex_unlock(&efx->mac_lock);
194
195 return 0;
196}
197
198/* Context: netif_addr_lock held, BHs disabled. */
199void efx_set_rx_mode(struct net_device *net_dev)
200{
201 struct efx_nic *efx = efx_netdev_priv(net_dev);
202
203 if (efx->port_enabled)
204 queue_work(efx->workqueue, &efx->mac_work);
205 /* Otherwise efx_start_port() will do this */
206}
207
208int efx_set_features(struct net_device *net_dev, netdev_features_t data)
209{
210 struct efx_nic *efx = efx_netdev_priv(net_dev);
211 int rc;
212
213 /* If disabling RX n-tuple filtering, clear existing filters */
214 if (net_dev->features & ~data & NETIF_F_NTUPLE) {
215 rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
216 if (rc)
217 return rc;
218 }
219
220 /* If Rx VLAN filter is changed, update filters via mac_reconfigure.
221 * If rx-fcs is changed, mac_reconfigure updates that too.
222 */
223 if ((net_dev->features ^ data) & (NETIF_F_HW_VLAN_CTAG_FILTER |
224 NETIF_F_RXFCS)) {
225 /* efx_set_rx_mode() will schedule MAC work to update filters
226 * when a new features are finally set in net_dev.
227 */
228 efx_set_rx_mode(net_dev);
229 }
230
231 return 0;
232}
233
234/* This ensures that the kernel is kept informed (via
235 * netif_carrier_on/off) of the link status, and also maintains the
236 * link status's stop on the port's TX queue.
237 */
238void efx_link_status_changed(struct efx_nic *efx)
239{
240 struct efx_link_state *link_state = &efx->link_state;
241
242 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
243 * that no events are triggered between unregister_netdev() and the
244 * driver unloading. A more general condition is that NETDEV_CHANGE
245 * can only be generated between NETDEV_UP and NETDEV_DOWN
246 */
247 if (!netif_running(efx->net_dev))
248 return;
249
250 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
251 efx->n_link_state_changes++;
252
253 if (link_state->up)
254 netif_carrier_on(efx->net_dev);
255 else
256 netif_carrier_off(efx->net_dev);
257 }
258
259 /* Status message for kernel log */
260 if (link_state->up)
261 netif_info(efx, link, efx->net_dev,
262 "link up at %uMbps %s-duplex (MTU %d)\n",
263 link_state->speed, link_state->fd ? "full" : "half",
264 efx->net_dev->mtu);
265 else
266 netif_info(efx, link, efx->net_dev, "link down\n");
267}
268
269unsigned int efx_xdp_max_mtu(struct efx_nic *efx)
270{
271 /* The maximum MTU that we can fit in a single page, allowing for
272 * framing, overhead and XDP headroom + tailroom.
273 */
274 int overhead = EFX_MAX_FRAME_LEN(0) + sizeof(struct efx_rx_page_state) +
275 efx->rx_prefix_size + efx->type->rx_buffer_padding +
276 efx->rx_ip_align + EFX_XDP_HEADROOM + EFX_XDP_TAILROOM;
277
278 return PAGE_SIZE - overhead;
279}
280
281/* Context: process, rtnl_lock() held. */
282int efx_change_mtu(struct net_device *net_dev, int new_mtu)
283{
284 struct efx_nic *efx = efx_netdev_priv(net_dev);
285 int rc;
286
287 rc = efx_check_disabled(efx);
288 if (rc)
289 return rc;
290
291 if (rtnl_dereference(efx->xdp_prog) &&
292 new_mtu > efx_xdp_max_mtu(efx)) {
293 netif_err(efx, drv, efx->net_dev,
294 "Requested MTU of %d too big for XDP (max: %d)\n",
295 new_mtu, efx_xdp_max_mtu(efx));
296 return -EINVAL;
297 }
298
299 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
300
301 efx_device_detach_sync(efx);
302 efx_stop_all(efx);
303
304 mutex_lock(&efx->mac_lock);
305 net_dev->mtu = new_mtu;
306 efx_mac_reconfigure(efx, true);
307 mutex_unlock(&efx->mac_lock);
308
309 efx_start_all(efx);
310 efx_device_attach_if_not_resetting(efx);
311 return 0;
312}
313
314/**************************************************************************
315 *
316 * Hardware monitor
317 *
318 **************************************************************************/
319
320/* Run periodically off the general workqueue */
321static void efx_monitor(struct work_struct *data)
322{
323 struct efx_nic *efx = container_of(data, struct efx_nic,
324 monitor_work.work);
325
326 netif_vdbg(efx, timer, efx->net_dev,
327 "hardware monitor executing on CPU %d\n",
328 raw_smp_processor_id());
329 BUG_ON(efx->type->monitor == NULL);
330
331 /* If the mac_lock is already held then it is likely a port
332 * reconfiguration is already in place, which will likely do
333 * most of the work of monitor() anyway.
334 */
335 if (mutex_trylock(&efx->mac_lock)) {
336 if (efx->port_enabled && efx->type->monitor)
337 efx->type->monitor(efx);
338 mutex_unlock(&efx->mac_lock);
339 }
340
341 efx_start_monitor(efx);
342}
343
344void efx_start_monitor(struct efx_nic *efx)
345{
346 if (efx->type->monitor)
347 queue_delayed_work(efx->workqueue, &efx->monitor_work,
348 efx_monitor_interval);
349}
350
351/**************************************************************************
352 *
353 * Event queue processing
354 *
355 *************************************************************************/
356
357/* Channels are shutdown and reinitialised whilst the NIC is running
358 * to propagate configuration changes (mtu, checksum offload), or
359 * to clear hardware error conditions
360 */
361static void efx_start_datapath(struct efx_nic *efx)
362{
363 netdev_features_t old_features = efx->net_dev->features;
364 bool old_rx_scatter = efx->rx_scatter;
365 size_t rx_buf_len;
366
367 /* Calculate the rx buffer allocation parameters required to
368 * support the current MTU, including padding for header
369 * alignment and overruns.
370 */
371 efx->rx_dma_len = (efx->rx_prefix_size +
372 EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
373 efx->type->rx_buffer_padding);
374 rx_buf_len = (sizeof(struct efx_rx_page_state) + EFX_XDP_HEADROOM +
375 efx->rx_ip_align + efx->rx_dma_len + EFX_XDP_TAILROOM);
376
377 if (rx_buf_len <= PAGE_SIZE) {
378 efx->rx_scatter = efx->type->always_rx_scatter;
379 efx->rx_buffer_order = 0;
380 } else if (efx->type->can_rx_scatter) {
381 BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
382 BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
383 2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
384 EFX_RX_BUF_ALIGNMENT) >
385 PAGE_SIZE);
386 efx->rx_scatter = true;
387 efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
388 efx->rx_buffer_order = 0;
389 } else {
390 efx->rx_scatter = false;
391 efx->rx_buffer_order = get_order(rx_buf_len);
392 }
393
394 efx_rx_config_page_split(efx);
395 if (efx->rx_buffer_order)
396 netif_dbg(efx, drv, efx->net_dev,
397 "RX buf len=%u; page order=%u batch=%u\n",
398 efx->rx_dma_len, efx->rx_buffer_order,
399 efx->rx_pages_per_batch);
400 else
401 netif_dbg(efx, drv, efx->net_dev,
402 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
403 efx->rx_dma_len, efx->rx_page_buf_step,
404 efx->rx_bufs_per_page, efx->rx_pages_per_batch);
405
406 /* Restore previously fixed features in hw_features and remove
407 * features which are fixed now
408 */
409 efx->net_dev->hw_features |= efx->net_dev->features;
410 efx->net_dev->hw_features &= ~efx->fixed_features;
411 efx->net_dev->features |= efx->fixed_features;
412 if (efx->net_dev->features != old_features)
413 netdev_features_change(efx->net_dev);
414
415 /* RX filters may also have scatter-enabled flags */
416 if ((efx->rx_scatter != old_rx_scatter) &&
417 efx->type->filter_update_rx_scatter)
418 efx->type->filter_update_rx_scatter(efx);
419
420 /* We must keep at least one descriptor in a TX ring empty.
421 * We could avoid this when the queue size does not exactly
422 * match the hardware ring size, but it's not that important.
423 * Therefore we stop the queue when one more skb might fill
424 * the ring completely. We wake it when half way back to
425 * empty.
426 */
427 efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
428 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
429
430 /* Initialise the channels */
431 efx_start_channels(efx);
432
433 efx_ptp_start_datapath(efx);
434
435 if (netif_device_present(efx->net_dev))
436 netif_tx_wake_all_queues(efx->net_dev);
437}
438
439static void efx_stop_datapath(struct efx_nic *efx)
440{
441 EFX_ASSERT_RESET_SERIALISED(efx);
442 BUG_ON(efx->port_enabled);
443
444 efx_ptp_stop_datapath(efx);
445
446 efx_stop_channels(efx);
447}
448
449/**************************************************************************
450 *
451 * Port handling
452 *
453 **************************************************************************/
454
455/* Equivalent to efx_link_set_advertising with all-zeroes, except does not
456 * force the Autoneg bit on.
457 */
458void efx_link_clear_advertising(struct efx_nic *efx)
459{
460 bitmap_zero(efx->link_advertising, __ETHTOOL_LINK_MODE_MASK_NBITS);
461 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
462}
463
464void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
465{
466 efx->wanted_fc = wanted_fc;
467 if (efx->link_advertising[0]) {
468 if (wanted_fc & EFX_FC_RX)
469 efx->link_advertising[0] |= (ADVERTISED_Pause |
470 ADVERTISED_Asym_Pause);
471 else
472 efx->link_advertising[0] &= ~(ADVERTISED_Pause |
473 ADVERTISED_Asym_Pause);
474 if (wanted_fc & EFX_FC_TX)
475 efx->link_advertising[0] ^= ADVERTISED_Asym_Pause;
476 }
477}
478
479static void efx_start_port(struct efx_nic *efx)
480{
481 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
482 BUG_ON(efx->port_enabled);
483
484 mutex_lock(&efx->mac_lock);
485 efx->port_enabled = true;
486
487 /* Ensure MAC ingress/egress is enabled */
488 efx_mac_reconfigure(efx, false);
489
490 mutex_unlock(&efx->mac_lock);
491}
492
493/* Cancel work for MAC reconfiguration, periodic hardware monitoring
494 * and the async self-test, wait for them to finish and prevent them
495 * being scheduled again. This doesn't cover online resets, which
496 * should only be cancelled when removing the device.
497 */
498static void efx_stop_port(struct efx_nic *efx)
499{
500 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
501
502 EFX_ASSERT_RESET_SERIALISED(efx);
503
504 mutex_lock(&efx->mac_lock);
505 efx->port_enabled = false;
506 mutex_unlock(&efx->mac_lock);
507
508 /* Serialise against efx_set_multicast_list() */
509 netif_addr_lock_bh(efx->net_dev);
510 netif_addr_unlock_bh(efx->net_dev);
511
512 cancel_delayed_work_sync(&efx->monitor_work);
513 efx_selftest_async_cancel(efx);
514 cancel_work_sync(&efx->mac_work);
515}
516
517/* If the interface is supposed to be running but is not, start
518 * the hardware and software data path, regular activity for the port
519 * (MAC statistics, link polling, etc.) and schedule the port to be
520 * reconfigured. Interrupts must already be enabled. This function
521 * is safe to call multiple times, so long as the NIC is not disabled.
522 * Requires the RTNL lock.
523 */
524void efx_start_all(struct efx_nic *efx)
525{
526 EFX_ASSERT_RESET_SERIALISED(efx);
527 BUG_ON(efx->state == STATE_DISABLED);
528
529 /* Check that it is appropriate to restart the interface. All
530 * of these flags are safe to read under just the rtnl lock
531 */
532 if (efx->port_enabled || !netif_running(efx->net_dev) ||
533 efx->reset_pending)
534 return;
535
536 efx_start_port(efx);
537 efx_start_datapath(efx);
538
539 /* Start the hardware monitor if there is one */
540 efx_start_monitor(efx);
541
542 efx_selftest_async_start(efx);
543
544 /* Link state detection is normally event-driven; we have
545 * to poll now because we could have missed a change
546 */
547 mutex_lock(&efx->mac_lock);
548 if (efx_mcdi_phy_poll(efx))
549 efx_link_status_changed(efx);
550 mutex_unlock(&efx->mac_lock);
551
552 if (efx->type->start_stats) {
553 efx->type->start_stats(efx);
554 efx->type->pull_stats(efx);
555 spin_lock_bh(&efx->stats_lock);
556 efx->type->update_stats(efx, NULL, NULL);
557 spin_unlock_bh(&efx->stats_lock);
558 }
559}
560
561/* Quiesce the hardware and software data path, and regular activity
562 * for the port without bringing the link down. Safe to call multiple
563 * times with the NIC in almost any state, but interrupts should be
564 * enabled. Requires the RTNL lock.
565 */
566void efx_stop_all(struct efx_nic *efx)
567{
568 EFX_ASSERT_RESET_SERIALISED(efx);
569
570 /* port_enabled can be read safely under the rtnl lock */
571 if (!efx->port_enabled)
572 return;
573
574 if (efx->type->update_stats) {
575 /* update stats before we go down so we can accurately count
576 * rx_nodesc_drops
577 */
578 efx->type->pull_stats(efx);
579 spin_lock_bh(&efx->stats_lock);
580 efx->type->update_stats(efx, NULL, NULL);
581 spin_unlock_bh(&efx->stats_lock);
582 efx->type->stop_stats(efx);
583 }
584
585 efx_stop_port(efx);
586
587 /* Stop the kernel transmit interface. This is only valid if
588 * the device is stopped or detached; otherwise the watchdog
589 * may fire immediately.
590 */
591 WARN_ON(netif_running(efx->net_dev) &&
592 netif_device_present(efx->net_dev));
593 netif_tx_disable(efx->net_dev);
594
595 efx_stop_datapath(efx);
596}
597
598/* Context: process, dev_base_lock or RTNL held, non-blocking. */
599void efx_net_stats(struct net_device *net_dev, struct rtnl_link_stats64 *stats)
600{
601 struct efx_nic *efx = efx_netdev_priv(net_dev);
602
603 spin_lock_bh(&efx->stats_lock);
604 efx_nic_update_stats_atomic(efx, NULL, stats);
605 spin_unlock_bh(&efx->stats_lock);
606}
607
608/* Push loopback/power/transmit disable settings to the PHY, and reconfigure
609 * the MAC appropriately. All other PHY configuration changes are pushed
610 * through phy_op->set_settings(), and pushed asynchronously to the MAC
611 * through efx_monitor().
612 *
613 * Callers must hold the mac_lock
614 */
615int __efx_reconfigure_port(struct efx_nic *efx)
616{
617 enum efx_phy_mode phy_mode;
618 int rc = 0;
619
620 WARN_ON(!mutex_is_locked(&efx->mac_lock));
621
622 /* Disable PHY transmit in mac level loopbacks */
623 phy_mode = efx->phy_mode;
624 if (LOOPBACK_INTERNAL(efx))
625 efx->phy_mode |= PHY_MODE_TX_DISABLED;
626 else
627 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
628
629 if (efx->type->reconfigure_port)
630 rc = efx->type->reconfigure_port(efx);
631
632 if (rc)
633 efx->phy_mode = phy_mode;
634
635 return rc;
636}
637
638/* Reinitialise the MAC to pick up new PHY settings, even if the port is
639 * disabled.
640 */
641int efx_reconfigure_port(struct efx_nic *efx)
642{
643 int rc;
644
645 EFX_ASSERT_RESET_SERIALISED(efx);
646
647 mutex_lock(&efx->mac_lock);
648 rc = __efx_reconfigure_port(efx);
649 mutex_unlock(&efx->mac_lock);
650
651 return rc;
652}
653
654/**************************************************************************
655 *
656 * Device reset and suspend
657 *
658 **************************************************************************/
659
660static void efx_wait_for_bist_end(struct efx_nic *efx)
661{
662 int i;
663
664 for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
665 if (efx_mcdi_poll_reboot(efx))
666 goto out;
667 msleep(BIST_WAIT_DELAY_MS);
668 }
669
670 netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
671out:
672 /* Either way unset the BIST flag. If we found no reboot we probably
673 * won't recover, but we should try.
674 */
675 efx->mc_bist_for_other_fn = false;
676}
677
678/* Try recovery mechanisms.
679 * For now only EEH is supported.
680 * Returns 0 if the recovery mechanisms are unsuccessful.
681 * Returns a non-zero value otherwise.
682 */
683int efx_try_recovery(struct efx_nic *efx)
684{
685#ifdef CONFIG_EEH
686 /* A PCI error can occur and not be seen by EEH because nothing
687 * happens on the PCI bus. In this case the driver may fail and
688 * schedule a 'recover or reset', leading to this recovery handler.
689 * Manually call the eeh failure check function.
690 */
691 struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
692 if (eeh_dev_check_failure(eehdev)) {
693 /* The EEH mechanisms will handle the error and reset the
694 * device if necessary.
695 */
696 return 1;
697 }
698#endif
699 return 0;
700}
701
702/* Tears down the entire software state and most of the hardware state
703 * before reset.
704 */
705void efx_reset_down(struct efx_nic *efx, enum reset_type method)
706{
707 EFX_ASSERT_RESET_SERIALISED(efx);
708
709 if (method == RESET_TYPE_MCDI_TIMEOUT)
710 efx->type->prepare_flr(efx);
711
712 efx_stop_all(efx);
713 efx_disable_interrupts(efx);
714
715 mutex_lock(&efx->mac_lock);
716 down_write(&efx->filter_sem);
717 mutex_lock(&efx->rss_lock);
718 efx->type->fini(efx);
719}
720
721/* Context: netif_tx_lock held, BHs disabled. */
722void efx_watchdog(struct net_device *net_dev, unsigned int txqueue)
723{
724 struct efx_nic *efx = efx_netdev_priv(net_dev);
725
726 netif_err(efx, tx_err, efx->net_dev,
727 "TX stuck with port_enabled=%d: resetting channels\n",
728 efx->port_enabled);
729
730 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
731}
732
733/* This function will always ensure that the locks acquired in
734 * efx_reset_down() are released. A failure return code indicates
735 * that we were unable to reinitialise the hardware, and the
736 * driver should be disabled. If ok is false, then the rx and tx
737 * engines are not restarted, pending a RESET_DISABLE.
738 */
739int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
740{
741 int rc;
742
743 EFX_ASSERT_RESET_SERIALISED(efx);
744
745 if (method == RESET_TYPE_MCDI_TIMEOUT)
746 efx->type->finish_flr(efx);
747
748 /* Ensure that SRAM is initialised even if we're disabling the device */
749 rc = efx->type->init(efx);
750 if (rc) {
751 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
752 goto fail;
753 }
754
755 if (!ok)
756 goto fail;
757
758 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
759 method != RESET_TYPE_DATAPATH) {
760 rc = efx_mcdi_port_reconfigure(efx);
761 if (rc && rc != -EPERM)
762 netif_err(efx, drv, efx->net_dev,
763 "could not restore PHY settings\n");
764 }
765
766 rc = efx_enable_interrupts(efx);
767 if (rc)
768 goto fail;
769
770#ifdef CONFIG_SFC_SRIOV
771 rc = efx->type->vswitching_restore(efx);
772 if (rc) /* not fatal; the PF will still work fine */
773 netif_warn(efx, probe, efx->net_dev,
774 "failed to restore vswitching rc=%d;"
775 " VFs may not function\n", rc);
776#endif
777
778 if (efx->type->rx_restore_rss_contexts)
779 efx->type->rx_restore_rss_contexts(efx);
780 mutex_unlock(&efx->rss_lock);
781 efx->type->filter_table_restore(efx);
782 up_write(&efx->filter_sem);
783
784 mutex_unlock(&efx->mac_lock);
785
786 efx_start_all(efx);
787
788 if (efx->type->udp_tnl_push_ports)
789 efx->type->udp_tnl_push_ports(efx);
790
791 return 0;
792
793fail:
794 efx->port_initialized = false;
795
796 mutex_unlock(&efx->rss_lock);
797 up_write(&efx->filter_sem);
798 mutex_unlock(&efx->mac_lock);
799
800 return rc;
801}
802
803/* Reset the NIC using the specified method. Note that the reset may
804 * fail, in which case the card will be left in an unusable state.
805 *
806 * Caller must hold the rtnl_lock.
807 */
808int efx_reset(struct efx_nic *efx, enum reset_type method)
809{
810 int rc, rc2 = 0;
811 bool disabled;
812
813 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
814 RESET_TYPE(method));
815
816 efx_device_detach_sync(efx);
817 /* efx_reset_down() grabs locks that prevent recovery on EF100.
818 * EF100 reset is handled in the efx_nic_type callback below.
819 */
820 if (efx_nic_rev(efx) != EFX_REV_EF100)
821 efx_reset_down(efx, method);
822
823 rc = efx->type->reset(efx, method);
824 if (rc) {
825 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
826 goto out;
827 }
828
829 /* Clear flags for the scopes we covered. We assume the NIC and
830 * driver are now quiescent so that there is no race here.
831 */
832 if (method < RESET_TYPE_MAX_METHOD)
833 efx->reset_pending &= -(1 << (method + 1));
834 else /* it doesn't fit into the well-ordered scope hierarchy */
835 __clear_bit(method, &efx->reset_pending);
836
837 /* Reinitialise bus-mastering, which may have been turned off before
838 * the reset was scheduled. This is still appropriate, even in the
839 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
840 * can respond to requests.
841 */
842 pci_set_master(efx->pci_dev);
843
844out:
845 /* Leave device stopped if necessary */
846 disabled = rc ||
847 method == RESET_TYPE_DISABLE ||
848 method == RESET_TYPE_RECOVER_OR_DISABLE;
849 if (efx_nic_rev(efx) != EFX_REV_EF100)
850 rc2 = efx_reset_up(efx, method, !disabled);
851 if (rc2) {
852 disabled = true;
853 if (!rc)
854 rc = rc2;
855 }
856
857 if (disabled) {
858 dev_close(efx->net_dev);
859 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
860 efx->state = STATE_DISABLED;
861 } else {
862 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
863 efx_device_attach_if_not_resetting(efx);
864 }
865 return rc;
866}
867
868/* The worker thread exists so that code that cannot sleep can
869 * schedule a reset for later.
870 */
871static void efx_reset_work(struct work_struct *data)
872{
873 struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
874 unsigned long pending;
875 enum reset_type method;
876
877 pending = READ_ONCE(efx->reset_pending);
878 method = fls(pending) - 1;
879
880 if (method == RESET_TYPE_MC_BIST)
881 efx_wait_for_bist_end(efx);
882
883 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
884 method == RESET_TYPE_RECOVER_OR_ALL) &&
885 efx_try_recovery(efx))
886 return;
887
888 if (!pending)
889 return;
890
891 rtnl_lock();
892
893 /* We checked the state in efx_schedule_reset() but it may
894 * have changed by now. Now that we have the RTNL lock,
895 * it cannot change again.
896 */
897 if (efx_net_active(efx->state))
898 (void)efx_reset(efx, method);
899
900 rtnl_unlock();
901}
902
903void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
904{
905 enum reset_type method;
906
907 if (efx_recovering(efx->state)) {
908 netif_dbg(efx, drv, efx->net_dev,
909 "recovering: skip scheduling %s reset\n",
910 RESET_TYPE(type));
911 return;
912 }
913
914 switch (type) {
915 case RESET_TYPE_INVISIBLE:
916 case RESET_TYPE_ALL:
917 case RESET_TYPE_RECOVER_OR_ALL:
918 case RESET_TYPE_WORLD:
919 case RESET_TYPE_DISABLE:
920 case RESET_TYPE_RECOVER_OR_DISABLE:
921 case RESET_TYPE_DATAPATH:
922 case RESET_TYPE_MC_BIST:
923 case RESET_TYPE_MCDI_TIMEOUT:
924 method = type;
925 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
926 RESET_TYPE(method));
927 break;
928 default:
929 method = efx->type->map_reset_reason(type);
930 netif_dbg(efx, drv, efx->net_dev,
931 "scheduling %s reset for %s\n",
932 RESET_TYPE(method), RESET_TYPE(type));
933 break;
934 }
935
936 set_bit(method, &efx->reset_pending);
937 smp_mb(); /* ensure we change reset_pending before checking state */
938
939 /* If we're not READY then just leave the flags set as the cue
940 * to abort probing or reschedule the reset later.
941 */
942 if (!efx_net_active(READ_ONCE(efx->state)))
943 return;
944
945 /* efx_process_channel() will no longer read events once a
946 * reset is scheduled. So switch back to poll'd MCDI completions.
947 */
948 efx_mcdi_mode_poll(efx);
949
950 efx_queue_reset_work(efx);
951}
952
953/**************************************************************************
954 *
955 * Dummy NIC operations
956 *
957 * Can be used for some unimplemented operations
958 * Needed so all function pointers are valid and do not have to be tested
959 * before use
960 *
961 **************************************************************************/
962int efx_port_dummy_op_int(struct efx_nic *efx)
963{
964 return 0;
965}
966void efx_port_dummy_op_void(struct efx_nic *efx) {}
967
968/**************************************************************************
969 *
970 * Data housekeeping
971 *
972 **************************************************************************/
973
974/* This zeroes out and then fills in the invariants in a struct
975 * efx_nic (including all sub-structures).
976 */
977int efx_init_struct(struct efx_nic *efx, struct pci_dev *pci_dev)
978{
979 int rc = -ENOMEM;
980
981 /* Initialise common structures */
982 INIT_LIST_HEAD(&efx->node);
983 INIT_LIST_HEAD(&efx->secondary_list);
984 spin_lock_init(&efx->biu_lock);
985#ifdef CONFIG_SFC_MTD
986 INIT_LIST_HEAD(&efx->mtd_list);
987#endif
988 INIT_WORK(&efx->reset_work, efx_reset_work);
989 INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
990 efx_selftest_async_init(efx);
991 efx->pci_dev = pci_dev;
992 efx->msg_enable = debug;
993 efx->state = STATE_UNINIT;
994 strscpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
995
996 efx->rx_prefix_size = efx->type->rx_prefix_size;
997 efx->rx_ip_align =
998 NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
999 efx->rx_packet_hash_offset =
1000 efx->type->rx_hash_offset - efx->type->rx_prefix_size;
1001 efx->rx_packet_ts_offset =
1002 efx->type->rx_ts_offset - efx->type->rx_prefix_size;
1003 INIT_LIST_HEAD(&efx->rss_context.list);
1004 efx->rss_context.context_id = EFX_MCDI_RSS_CONTEXT_INVALID;
1005 mutex_init(&efx->rss_lock);
1006 efx->vport_id = EVB_PORT_ID_ASSIGNED;
1007 spin_lock_init(&efx->stats_lock);
1008 efx->vi_stride = EFX_DEFAULT_VI_STRIDE;
1009 efx->num_mac_stats = MC_CMD_MAC_NSTATS;
1010 BUILD_BUG_ON(MC_CMD_MAC_NSTATS - 1 != MC_CMD_MAC_GENERATION_END);
1011 mutex_init(&efx->mac_lock);
1012 init_rwsem(&efx->filter_sem);
1013#ifdef CONFIG_RFS_ACCEL
1014 mutex_init(&efx->rps_mutex);
1015 spin_lock_init(&efx->rps_hash_lock);
1016 /* Failure to allocate is not fatal, but may degrade ARFS performance */
1017 efx->rps_hash_table = kcalloc(EFX_ARFS_HASH_TABLE_SIZE,
1018 sizeof(*efx->rps_hash_table), GFP_KERNEL);
1019#endif
1020 spin_lock_init(&efx->vf_reps_lock);
1021 INIT_LIST_HEAD(&efx->vf_reps);
1022 INIT_WORK(&efx->mac_work, efx_mac_work);
1023 init_waitqueue_head(&efx->flush_wq);
1024
1025 efx->tx_queues_per_channel = 1;
1026 efx->rxq_entries = EFX_DEFAULT_DMAQ_SIZE;
1027 efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1028
1029 efx->mem_bar = UINT_MAX;
1030
1031 rc = efx_init_channels(efx);
1032 if (rc)
1033 goto fail;
1034
1035 /* Would be good to use the net_dev name, but we're too early */
1036 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
1037 pci_name(pci_dev));
1038 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
1039 if (!efx->workqueue) {
1040 rc = -ENOMEM;
1041 goto fail;
1042 }
1043
1044 return 0;
1045
1046fail:
1047 efx_fini_struct(efx);
1048 return rc;
1049}
1050
1051void efx_fini_struct(struct efx_nic *efx)
1052{
1053#ifdef CONFIG_RFS_ACCEL
1054 kfree(efx->rps_hash_table);
1055#endif
1056
1057 efx_fini_channels(efx);
1058
1059 kfree(efx->vpd_sn);
1060
1061 if (efx->workqueue) {
1062 destroy_workqueue(efx->workqueue);
1063 efx->workqueue = NULL;
1064 }
1065}
1066
1067/* This configures the PCI device to enable I/O and DMA. */
1068int efx_init_io(struct efx_nic *efx, int bar, dma_addr_t dma_mask,
1069 unsigned int mem_map_size)
1070{
1071 struct pci_dev *pci_dev = efx->pci_dev;
1072 int rc;
1073
1074 efx->mem_bar = UINT_MAX;
1075 pci_dbg(pci_dev, "initialising I/O bar=%d\n", bar);
1076
1077 rc = pci_enable_device(pci_dev);
1078 if (rc) {
1079 pci_err(pci_dev, "failed to enable PCI device\n");
1080 goto fail1;
1081 }
1082
1083 pci_set_master(pci_dev);
1084
1085 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1086 if (rc) {
1087 pci_err(efx->pci_dev, "could not find a suitable DMA mask\n");
1088 goto fail2;
1089 }
1090 pci_dbg(efx->pci_dev, "using DMA mask %llx\n", (unsigned long long)dma_mask);
1091
1092 efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1093 if (!efx->membase_phys) {
1094 pci_err(efx->pci_dev,
1095 "ERROR: No BAR%d mapping from the BIOS. Try pci=realloc on the kernel command line\n",
1096 bar);
1097 rc = -ENODEV;
1098 goto fail3;
1099 }
1100
1101 rc = pci_request_region(pci_dev, bar, "sfc");
1102 if (rc) {
1103 pci_err(efx->pci_dev,
1104 "request for memory BAR[%d] failed\n", bar);
1105 rc = -EIO;
1106 goto fail3;
1107 }
1108 efx->mem_bar = bar;
1109 efx->membase = ioremap(efx->membase_phys, mem_map_size);
1110 if (!efx->membase) {
1111 pci_err(efx->pci_dev,
1112 "could not map memory BAR[%d] at %llx+%x\n", bar,
1113 (unsigned long long)efx->membase_phys, mem_map_size);
1114 rc = -ENOMEM;
1115 goto fail4;
1116 }
1117 pci_dbg(efx->pci_dev,
1118 "memory BAR[%d] at %llx+%x (virtual %p)\n", bar,
1119 (unsigned long long)efx->membase_phys, mem_map_size,
1120 efx->membase);
1121
1122 return 0;
1123
1124fail4:
1125 pci_release_region(efx->pci_dev, bar);
1126fail3:
1127 efx->membase_phys = 0;
1128fail2:
1129 pci_disable_device(efx->pci_dev);
1130fail1:
1131 return rc;
1132}
1133
1134void efx_fini_io(struct efx_nic *efx)
1135{
1136 pci_dbg(efx->pci_dev, "shutting down I/O\n");
1137
1138 if (efx->membase) {
1139 iounmap(efx->membase);
1140 efx->membase = NULL;
1141 }
1142
1143 if (efx->membase_phys) {
1144 pci_release_region(efx->pci_dev, efx->mem_bar);
1145 efx->membase_phys = 0;
1146 efx->mem_bar = UINT_MAX;
1147 }
1148
1149 /* Don't disable bus-mastering if VFs are assigned */
1150 if (!pci_vfs_assigned(efx->pci_dev))
1151 pci_disable_device(efx->pci_dev);
1152}
1153
1154#ifdef CONFIG_SFC_MCDI_LOGGING
1155static ssize_t mcdi_logging_show(struct device *dev,
1156 struct device_attribute *attr,
1157 char *buf)
1158{
1159 struct efx_nic *efx = dev_get_drvdata(dev);
1160 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1161
1162 return sysfs_emit(buf, "%d\n", mcdi->logging_enabled);
1163}
1164
1165static ssize_t mcdi_logging_store(struct device *dev,
1166 struct device_attribute *attr,
1167 const char *buf, size_t count)
1168{
1169 struct efx_nic *efx = dev_get_drvdata(dev);
1170 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1171 bool enable = count > 0 && *buf != '0';
1172
1173 mcdi->logging_enabled = enable;
1174 return count;
1175}
1176
1177static DEVICE_ATTR_RW(mcdi_logging);
1178
1179void efx_init_mcdi_logging(struct efx_nic *efx)
1180{
1181 int rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1182
1183 if (rc) {
1184 netif_warn(efx, drv, efx->net_dev,
1185 "failed to init net dev attributes\n");
1186 }
1187}
1188
1189void efx_fini_mcdi_logging(struct efx_nic *efx)
1190{
1191 device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1192}
1193#endif
1194
1195/* A PCI error affecting this device was detected.
1196 * At this point MMIO and DMA may be disabled.
1197 * Stop the software path and request a slot reset.
1198 */
1199static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
1200 pci_channel_state_t state)
1201{
1202 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
1203 struct efx_nic *efx = pci_get_drvdata(pdev);
1204
1205 if (state == pci_channel_io_perm_failure)
1206 return PCI_ERS_RESULT_DISCONNECT;
1207
1208 rtnl_lock();
1209
1210 if (efx->state != STATE_DISABLED) {
1211 efx->state = efx_recover(efx->state);
1212 efx->reset_pending = 0;
1213
1214 efx_device_detach_sync(efx);
1215
1216 if (efx_net_active(efx->state)) {
1217 efx_stop_all(efx);
1218 efx_disable_interrupts(efx);
1219 }
1220
1221 status = PCI_ERS_RESULT_NEED_RESET;
1222 } else {
1223 /* If the interface is disabled we don't want to do anything
1224 * with it.
1225 */
1226 status = PCI_ERS_RESULT_RECOVERED;
1227 }
1228
1229 rtnl_unlock();
1230
1231 pci_disable_device(pdev);
1232
1233 return status;
1234}
1235
1236/* Fake a successful reset, which will be performed later in efx_io_resume. */
1237static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
1238{
1239 struct efx_nic *efx = pci_get_drvdata(pdev);
1240 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
1241
1242 if (pci_enable_device(pdev)) {
1243 netif_err(efx, hw, efx->net_dev,
1244 "Cannot re-enable PCI device after reset.\n");
1245 status = PCI_ERS_RESULT_DISCONNECT;
1246 }
1247
1248 return status;
1249}
1250
1251/* Perform the actual reset and resume I/O operations. */
1252static void efx_io_resume(struct pci_dev *pdev)
1253{
1254 struct efx_nic *efx = pci_get_drvdata(pdev);
1255 int rc;
1256
1257 rtnl_lock();
1258
1259 if (efx->state == STATE_DISABLED)
1260 goto out;
1261
1262 rc = efx_reset(efx, RESET_TYPE_ALL);
1263 if (rc) {
1264 netif_err(efx, hw, efx->net_dev,
1265 "efx_reset failed after PCI error (%d)\n", rc);
1266 } else {
1267 efx->state = efx_recovered(efx->state);
1268 netif_dbg(efx, hw, efx->net_dev,
1269 "Done resetting and resuming IO after PCI error.\n");
1270 }
1271
1272out:
1273 rtnl_unlock();
1274}
1275
1276/* For simplicity and reliability, we always require a slot reset and try to
1277 * reset the hardware when a pci error affecting the device is detected.
1278 * We leave both the link_reset and mmio_enabled callback unimplemented:
1279 * with our request for slot reset the mmio_enabled callback will never be
1280 * called, and the link_reset callback is not used by AER or EEH mechanisms.
1281 */
1282const struct pci_error_handlers efx_err_handlers = {
1283 .error_detected = efx_io_error_detected,
1284 .slot_reset = efx_io_slot_reset,
1285 .resume = efx_io_resume,
1286};
1287
1288/* Determine whether the NIC will be able to handle TX offloads for a given
1289 * encapsulated packet.
1290 */
1291static bool efx_can_encap_offloads(struct efx_nic *efx, struct sk_buff *skb)
1292{
1293 struct gre_base_hdr *greh;
1294 __be16 dst_port;
1295 u8 ipproto;
1296
1297 /* Does the NIC support encap offloads?
1298 * If not, we should never get here, because we shouldn't have
1299 * advertised encap offload feature flags in the first place.
1300 */
1301 if (WARN_ON_ONCE(!efx->type->udp_tnl_has_port))
1302 return false;
1303
1304 /* Determine encapsulation protocol in use */
1305 switch (skb->protocol) {
1306 case htons(ETH_P_IP):
1307 ipproto = ip_hdr(skb)->protocol;
1308 break;
1309 case htons(ETH_P_IPV6):
1310 /* If there are extension headers, this will cause us to
1311 * think we can't offload something that we maybe could have.
1312 */
1313 ipproto = ipv6_hdr(skb)->nexthdr;
1314 break;
1315 default:
1316 /* Not IP, so can't offload it */
1317 return false;
1318 }
1319 switch (ipproto) {
1320 case IPPROTO_GRE:
1321 /* We support NVGRE but not IP over GRE or random gretaps.
1322 * Specifically, the NIC will accept GRE as encapsulated if
1323 * the inner protocol is Ethernet, but only handle it
1324 * correctly if the GRE header is 8 bytes long. Moreover,
1325 * it will not update the Checksum or Sequence Number fields
1326 * if they are present. (The Routing Present flag,
1327 * GRE_ROUTING, cannot be set else the header would be more
1328 * than 8 bytes long; so we don't have to worry about it.)
1329 */
1330 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER)
1331 return false;
1332 if (ntohs(skb->inner_protocol) != ETH_P_TEB)
1333 return false;
1334 if (skb_inner_mac_header(skb) - skb_transport_header(skb) != 8)
1335 return false;
1336 greh = (struct gre_base_hdr *)skb_transport_header(skb);
1337 return !(greh->flags & (GRE_CSUM | GRE_SEQ));
1338 case IPPROTO_UDP:
1339 /* If the port is registered for a UDP tunnel, we assume the
1340 * packet is for that tunnel, and the NIC will handle it as
1341 * such. If not, the NIC won't know what to do with it.
1342 */
1343 dst_port = udp_hdr(skb)->dest;
1344 return efx->type->udp_tnl_has_port(efx, dst_port);
1345 default:
1346 return false;
1347 }
1348}
1349
1350netdev_features_t efx_features_check(struct sk_buff *skb, struct net_device *dev,
1351 netdev_features_t features)
1352{
1353 struct efx_nic *efx = efx_netdev_priv(dev);
1354
1355 if (skb->encapsulation) {
1356 if (features & NETIF_F_GSO_MASK)
1357 /* Hardware can only do TSO with at most 208 bytes
1358 * of headers.
1359 */
1360 if (skb_inner_transport_offset(skb) >
1361 EFX_TSO2_MAX_HDRLEN)
1362 features &= ~(NETIF_F_GSO_MASK);
1363 if (features & (NETIF_F_GSO_MASK | NETIF_F_CSUM_MASK))
1364 if (!efx_can_encap_offloads(efx, skb))
1365 features &= ~(NETIF_F_GSO_MASK |
1366 NETIF_F_CSUM_MASK);
1367 }
1368 return features;
1369}
1370
1371int efx_get_phys_port_id(struct net_device *net_dev,
1372 struct netdev_phys_item_id *ppid)
1373{
1374 struct efx_nic *efx = efx_netdev_priv(net_dev);
1375
1376 if (efx->type->get_phys_port_id)
1377 return efx->type->get_phys_port_id(efx, ppid);
1378 else
1379 return -EOPNOTSUPP;
1380}
1381
1382int efx_get_phys_port_name(struct net_device *net_dev, char *name, size_t len)
1383{
1384 struct efx_nic *efx = efx_netdev_priv(net_dev);
1385
1386 if (snprintf(name, len, "p%u", efx->port_num) >= len)
1387 return -EINVAL;
1388 return 0;
1389}
1390
1391void efx_detach_reps(struct efx_nic *efx)
1392{
1393 struct net_device *rep_dev;
1394 struct efx_rep *efv;
1395
1396 ASSERT_RTNL();
1397 netif_dbg(efx, drv, efx->net_dev, "Detaching VF representors\n");
1398 list_for_each_entry(efv, &efx->vf_reps, list) {
1399 rep_dev = efv->net_dev;
1400 if (!rep_dev)
1401 continue;
1402 netif_carrier_off(rep_dev);
1403 /* See efx_device_detach_sync() */
1404 netif_tx_lock_bh(rep_dev);
1405 netif_tx_stop_all_queues(rep_dev);
1406 netif_tx_unlock_bh(rep_dev);
1407 }
1408}
1409
1410void efx_attach_reps(struct efx_nic *efx)
1411{
1412 struct net_device *rep_dev;
1413 struct efx_rep *efv;
1414
1415 ASSERT_RTNL();
1416 netif_dbg(efx, drv, efx->net_dev, "Attaching VF representors\n");
1417 list_for_each_entry(efv, &efx->vf_reps, list) {
1418 rep_dev = efv->net_dev;
1419 if (!rep_dev)
1420 continue;
1421 netif_tx_wake_all_queues(rep_dev);
1422 netif_carrier_on(rep_dev);
1423 }
1424}
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/filter.h>
13#include <linux/module.h>
14#include <linux/netdevice.h>
15#include <net/gre.h>
16#include "efx_common.h"
17#include "efx_channels.h"
18#include "efx.h"
19#include "mcdi.h"
20#include "selftest.h"
21#include "rx_common.h"
22#include "tx_common.h"
23#include "nic.h"
24#include "mcdi_port_common.h"
25#include "io.h"
26#include "mcdi_pcol.h"
27#include "ef100_rep.h"
28
29static unsigned int debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
30 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
31 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
32 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
33module_param(debug, uint, 0);
34MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
35
36/* This is the time (in jiffies) between invocations of the hardware
37 * monitor.
38 * On Falcon-based NICs, this will:
39 * - Check the on-board hardware monitor;
40 * - Poll the link state and reconfigure the hardware as necessary.
41 * On Siena-based NICs for power systems with EEH support, this will give EEH a
42 * chance to start.
43 */
44static unsigned int efx_monitor_interval = 1 * HZ;
45
46/* How often and how many times to poll for a reset while waiting for a
47 * BIST that another function started to complete.
48 */
49#define BIST_WAIT_DELAY_MS 100
50#define BIST_WAIT_DELAY_COUNT 100
51
52/* Default stats update time */
53#define STATS_PERIOD_MS_DEFAULT 1000
54
55static const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
56static const char *const efx_reset_type_names[] = {
57 [RESET_TYPE_INVISIBLE] = "INVISIBLE",
58 [RESET_TYPE_ALL] = "ALL",
59 [RESET_TYPE_RECOVER_OR_ALL] = "RECOVER_OR_ALL",
60 [RESET_TYPE_WORLD] = "WORLD",
61 [RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
62 [RESET_TYPE_DATAPATH] = "DATAPATH",
63 [RESET_TYPE_MC_BIST] = "MC_BIST",
64 [RESET_TYPE_DISABLE] = "DISABLE",
65 [RESET_TYPE_TX_WATCHDOG] = "TX_WATCHDOG",
66 [RESET_TYPE_INT_ERROR] = "INT_ERROR",
67 [RESET_TYPE_DMA_ERROR] = "DMA_ERROR",
68 [RESET_TYPE_TX_SKIP] = "TX_SKIP",
69 [RESET_TYPE_MC_FAILURE] = "MC_FAILURE",
70 [RESET_TYPE_MCDI_TIMEOUT] = "MCDI_TIMEOUT (FLR)",
71};
72
73#define RESET_TYPE(type) \
74 STRING_TABLE_LOOKUP(type, efx_reset_type)
75
76/* Loopback mode names (see LOOPBACK_MODE()) */
77const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
78const char *const efx_loopback_mode_names[] = {
79 [LOOPBACK_NONE] = "NONE",
80 [LOOPBACK_DATA] = "DATAPATH",
81 [LOOPBACK_GMAC] = "GMAC",
82 [LOOPBACK_XGMII] = "XGMII",
83 [LOOPBACK_XGXS] = "XGXS",
84 [LOOPBACK_XAUI] = "XAUI",
85 [LOOPBACK_GMII] = "GMII",
86 [LOOPBACK_SGMII] = "SGMII",
87 [LOOPBACK_XGBR] = "XGBR",
88 [LOOPBACK_XFI] = "XFI",
89 [LOOPBACK_XAUI_FAR] = "XAUI_FAR",
90 [LOOPBACK_GMII_FAR] = "GMII_FAR",
91 [LOOPBACK_SGMII_FAR] = "SGMII_FAR",
92 [LOOPBACK_XFI_FAR] = "XFI_FAR",
93 [LOOPBACK_GPHY] = "GPHY",
94 [LOOPBACK_PHYXS] = "PHYXS",
95 [LOOPBACK_PCS] = "PCS",
96 [LOOPBACK_PMAPMD] = "PMA/PMD",
97 [LOOPBACK_XPORT] = "XPORT",
98 [LOOPBACK_XGMII_WS] = "XGMII_WS",
99 [LOOPBACK_XAUI_WS] = "XAUI_WS",
100 [LOOPBACK_XAUI_WS_FAR] = "XAUI_WS_FAR",
101 [LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
102 [LOOPBACK_GMII_WS] = "GMII_WS",
103 [LOOPBACK_XFI_WS] = "XFI_WS",
104 [LOOPBACK_XFI_WS_FAR] = "XFI_WS_FAR",
105 [LOOPBACK_PHYXS_WS] = "PHYXS_WS",
106};
107
108/* Reset workqueue. If any NIC has a hardware failure then a reset will be
109 * queued onto this work queue. This is not a per-nic work queue, because
110 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
111 */
112static struct workqueue_struct *reset_workqueue;
113
114int efx_create_reset_workqueue(void)
115{
116 reset_workqueue = create_singlethread_workqueue("sfc_reset");
117 if (!reset_workqueue) {
118 printk(KERN_ERR "Failed to create reset workqueue\n");
119 return -ENOMEM;
120 }
121
122 return 0;
123}
124
125void efx_queue_reset_work(struct efx_nic *efx)
126{
127 queue_work(reset_workqueue, &efx->reset_work);
128}
129
130void efx_flush_reset_workqueue(struct efx_nic *efx)
131{
132 cancel_work_sync(&efx->reset_work);
133}
134
135void efx_destroy_reset_workqueue(void)
136{
137 if (reset_workqueue) {
138 destroy_workqueue(reset_workqueue);
139 reset_workqueue = NULL;
140 }
141}
142
143/* We assume that efx->type->reconfigure_mac will always try to sync RX
144 * filters and therefore needs to read-lock the filter table against freeing
145 */
146void efx_mac_reconfigure(struct efx_nic *efx, bool mtu_only)
147{
148 if (efx->type->reconfigure_mac) {
149 down_read(&efx->filter_sem);
150 efx->type->reconfigure_mac(efx, mtu_only);
151 up_read(&efx->filter_sem);
152 }
153}
154
155/* Asynchronous work item for changing MAC promiscuity and multicast
156 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
157 * MAC directly.
158 */
159static void efx_mac_work(struct work_struct *data)
160{
161 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
162
163 mutex_lock(&efx->mac_lock);
164 if (efx->port_enabled)
165 efx_mac_reconfigure(efx, false);
166 mutex_unlock(&efx->mac_lock);
167}
168
169int efx_set_mac_address(struct net_device *net_dev, void *data)
170{
171 struct efx_nic *efx = efx_netdev_priv(net_dev);
172 struct sockaddr *addr = data;
173 u8 *new_addr = addr->sa_data;
174 u8 old_addr[6];
175 int rc;
176
177 if (!is_valid_ether_addr(new_addr)) {
178 netif_err(efx, drv, efx->net_dev,
179 "invalid ethernet MAC address requested: %pM\n",
180 new_addr);
181 return -EADDRNOTAVAIL;
182 }
183
184 /* save old address */
185 ether_addr_copy(old_addr, net_dev->dev_addr);
186 eth_hw_addr_set(net_dev, new_addr);
187 if (efx->type->set_mac_address) {
188 rc = efx->type->set_mac_address(efx);
189 if (rc) {
190 eth_hw_addr_set(net_dev, old_addr);
191 return rc;
192 }
193 }
194
195 /* Reconfigure the MAC */
196 mutex_lock(&efx->mac_lock);
197 efx_mac_reconfigure(efx, false);
198 mutex_unlock(&efx->mac_lock);
199
200 return 0;
201}
202
203/* Context: netif_addr_lock held, BHs disabled. */
204void efx_set_rx_mode(struct net_device *net_dev)
205{
206 struct efx_nic *efx = efx_netdev_priv(net_dev);
207
208 if (efx->port_enabled)
209 queue_work(efx->workqueue, &efx->mac_work);
210 /* Otherwise efx_start_port() will do this */
211}
212
213int efx_set_features(struct net_device *net_dev, netdev_features_t data)
214{
215 struct efx_nic *efx = efx_netdev_priv(net_dev);
216 int rc;
217
218 /* If disabling RX n-tuple filtering, clear existing filters */
219 if (net_dev->features & ~data & NETIF_F_NTUPLE) {
220 rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
221 if (rc)
222 return rc;
223 }
224
225 /* If Rx VLAN filter is changed, update filters via mac_reconfigure.
226 * If rx-fcs is changed, mac_reconfigure updates that too.
227 */
228 if ((net_dev->features ^ data) & (NETIF_F_HW_VLAN_CTAG_FILTER |
229 NETIF_F_RXFCS)) {
230 /* efx_set_rx_mode() will schedule MAC work to update filters
231 * when a new features are finally set in net_dev.
232 */
233 efx_set_rx_mode(net_dev);
234 }
235
236 return 0;
237}
238
239/* This ensures that the kernel is kept informed (via
240 * netif_carrier_on/off) of the link status, and also maintains the
241 * link status's stop on the port's TX queue.
242 */
243void efx_link_status_changed(struct efx_nic *efx)
244{
245 struct efx_link_state *link_state = &efx->link_state;
246
247 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
248 * that no events are triggered between unregister_netdev() and the
249 * driver unloading. A more general condition is that NETDEV_CHANGE
250 * can only be generated between NETDEV_UP and NETDEV_DOWN
251 */
252 if (!netif_running(efx->net_dev))
253 return;
254
255 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
256 efx->n_link_state_changes++;
257
258 if (link_state->up)
259 netif_carrier_on(efx->net_dev);
260 else
261 netif_carrier_off(efx->net_dev);
262 }
263
264 /* Status message for kernel log */
265 if (link_state->up)
266 netif_info(efx, link, efx->net_dev,
267 "link up at %uMbps %s-duplex (MTU %d)\n",
268 link_state->speed, link_state->fd ? "full" : "half",
269 efx->net_dev->mtu);
270 else
271 netif_info(efx, link, efx->net_dev, "link down\n");
272}
273
274unsigned int efx_xdp_max_mtu(struct efx_nic *efx)
275{
276 /* The maximum MTU that we can fit in a single page, allowing for
277 * framing, overhead and XDP headroom + tailroom.
278 */
279 int overhead = EFX_MAX_FRAME_LEN(0) + sizeof(struct efx_rx_page_state) +
280 efx->rx_prefix_size + efx->type->rx_buffer_padding +
281 efx->rx_ip_align + EFX_XDP_HEADROOM + EFX_XDP_TAILROOM;
282
283 return PAGE_SIZE - overhead;
284}
285
286/* Context: process, rtnl_lock() held. */
287int efx_change_mtu(struct net_device *net_dev, int new_mtu)
288{
289 struct efx_nic *efx = efx_netdev_priv(net_dev);
290 int rc;
291
292 rc = efx_check_disabled(efx);
293 if (rc)
294 return rc;
295
296 if (rtnl_dereference(efx->xdp_prog) &&
297 new_mtu > efx_xdp_max_mtu(efx)) {
298 netif_err(efx, drv, efx->net_dev,
299 "Requested MTU of %d too big for XDP (max: %d)\n",
300 new_mtu, efx_xdp_max_mtu(efx));
301 return -EINVAL;
302 }
303
304 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
305
306 efx_device_detach_sync(efx);
307 efx_stop_all(efx);
308
309 mutex_lock(&efx->mac_lock);
310 net_dev->mtu = new_mtu;
311 efx_mac_reconfigure(efx, true);
312 mutex_unlock(&efx->mac_lock);
313
314 efx_start_all(efx);
315 efx_device_attach_if_not_resetting(efx);
316 return 0;
317}
318
319/**************************************************************************
320 *
321 * Hardware monitor
322 *
323 **************************************************************************/
324
325/* Run periodically off the general workqueue */
326static void efx_monitor(struct work_struct *data)
327{
328 struct efx_nic *efx = container_of(data, struct efx_nic,
329 monitor_work.work);
330
331 netif_vdbg(efx, timer, efx->net_dev,
332 "hardware monitor executing on CPU %d\n",
333 raw_smp_processor_id());
334 BUG_ON(efx->type->monitor == NULL);
335
336 /* If the mac_lock is already held then it is likely a port
337 * reconfiguration is already in place, which will likely do
338 * most of the work of monitor() anyway.
339 */
340 if (mutex_trylock(&efx->mac_lock)) {
341 if (efx->port_enabled && efx->type->monitor)
342 efx->type->monitor(efx);
343 mutex_unlock(&efx->mac_lock);
344 }
345
346 efx_start_monitor(efx);
347}
348
349void efx_start_monitor(struct efx_nic *efx)
350{
351 if (efx->type->monitor)
352 queue_delayed_work(efx->workqueue, &efx->monitor_work,
353 efx_monitor_interval);
354}
355
356/**************************************************************************
357 *
358 * Event queue processing
359 *
360 *************************************************************************/
361
362/* Channels are shutdown and reinitialised whilst the NIC is running
363 * to propagate configuration changes (mtu, checksum offload), or
364 * to clear hardware error conditions
365 */
366static void efx_start_datapath(struct efx_nic *efx)
367{
368 netdev_features_t old_features = efx->net_dev->features;
369 bool old_rx_scatter = efx->rx_scatter;
370 size_t rx_buf_len;
371
372 /* Calculate the rx buffer allocation parameters required to
373 * support the current MTU, including padding for header
374 * alignment and overruns.
375 */
376 efx->rx_dma_len = (efx->rx_prefix_size +
377 EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
378 efx->type->rx_buffer_padding);
379 rx_buf_len = (sizeof(struct efx_rx_page_state) + EFX_XDP_HEADROOM +
380 efx->rx_ip_align + efx->rx_dma_len + EFX_XDP_TAILROOM);
381
382 if (rx_buf_len <= PAGE_SIZE) {
383 efx->rx_scatter = efx->type->always_rx_scatter;
384 efx->rx_buffer_order = 0;
385 } else if (efx->type->can_rx_scatter) {
386 BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
387 BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
388 2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
389 EFX_RX_BUF_ALIGNMENT) >
390 PAGE_SIZE);
391 efx->rx_scatter = true;
392 efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
393 efx->rx_buffer_order = 0;
394 } else {
395 efx->rx_scatter = false;
396 efx->rx_buffer_order = get_order(rx_buf_len);
397 }
398
399 efx_rx_config_page_split(efx);
400 if (efx->rx_buffer_order)
401 netif_dbg(efx, drv, efx->net_dev,
402 "RX buf len=%u; page order=%u batch=%u\n",
403 efx->rx_dma_len, efx->rx_buffer_order,
404 efx->rx_pages_per_batch);
405 else
406 netif_dbg(efx, drv, efx->net_dev,
407 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
408 efx->rx_dma_len, efx->rx_page_buf_step,
409 efx->rx_bufs_per_page, efx->rx_pages_per_batch);
410
411 /* Restore previously fixed features in hw_features and remove
412 * features which are fixed now
413 */
414 efx->net_dev->hw_features |= efx->net_dev->features;
415 efx->net_dev->hw_features &= ~efx->fixed_features;
416 efx->net_dev->features |= efx->fixed_features;
417 if (efx->net_dev->features != old_features)
418 netdev_features_change(efx->net_dev);
419
420 /* RX filters may also have scatter-enabled flags */
421 if ((efx->rx_scatter != old_rx_scatter) &&
422 efx->type->filter_update_rx_scatter)
423 efx->type->filter_update_rx_scatter(efx);
424
425 /* We must keep at least one descriptor in a TX ring empty.
426 * We could avoid this when the queue size does not exactly
427 * match the hardware ring size, but it's not that important.
428 * Therefore we stop the queue when one more skb might fill
429 * the ring completely. We wake it when half way back to
430 * empty.
431 */
432 efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
433 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
434
435 /* Initialise the channels */
436 efx_start_channels(efx);
437
438 efx_ptp_start_datapath(efx);
439
440 if (netif_device_present(efx->net_dev))
441 netif_tx_wake_all_queues(efx->net_dev);
442}
443
444static void efx_stop_datapath(struct efx_nic *efx)
445{
446 EFX_ASSERT_RESET_SERIALISED(efx);
447 BUG_ON(efx->port_enabled);
448
449 efx_ptp_stop_datapath(efx);
450
451 efx_stop_channels(efx);
452}
453
454/**************************************************************************
455 *
456 * Port handling
457 *
458 **************************************************************************/
459
460/* Equivalent to efx_link_set_advertising with all-zeroes, except does not
461 * force the Autoneg bit on.
462 */
463void efx_link_clear_advertising(struct efx_nic *efx)
464{
465 bitmap_zero(efx->link_advertising, __ETHTOOL_LINK_MODE_MASK_NBITS);
466 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
467}
468
469void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
470{
471 efx->wanted_fc = wanted_fc;
472 if (efx->link_advertising[0]) {
473 if (wanted_fc & EFX_FC_RX)
474 efx->link_advertising[0] |= (ADVERTISED_Pause |
475 ADVERTISED_Asym_Pause);
476 else
477 efx->link_advertising[0] &= ~(ADVERTISED_Pause |
478 ADVERTISED_Asym_Pause);
479 if (wanted_fc & EFX_FC_TX)
480 efx->link_advertising[0] ^= ADVERTISED_Asym_Pause;
481 }
482}
483
484static void efx_start_port(struct efx_nic *efx)
485{
486 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
487 BUG_ON(efx->port_enabled);
488
489 mutex_lock(&efx->mac_lock);
490 efx->port_enabled = true;
491
492 /* Ensure MAC ingress/egress is enabled */
493 efx_mac_reconfigure(efx, false);
494
495 mutex_unlock(&efx->mac_lock);
496}
497
498/* Cancel work for MAC reconfiguration, periodic hardware monitoring
499 * and the async self-test, wait for them to finish and prevent them
500 * being scheduled again. This doesn't cover online resets, which
501 * should only be cancelled when removing the device.
502 */
503static void efx_stop_port(struct efx_nic *efx)
504{
505 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
506
507 EFX_ASSERT_RESET_SERIALISED(efx);
508
509 mutex_lock(&efx->mac_lock);
510 efx->port_enabled = false;
511 mutex_unlock(&efx->mac_lock);
512
513 /* Serialise against efx_set_multicast_list() */
514 netif_addr_lock_bh(efx->net_dev);
515 netif_addr_unlock_bh(efx->net_dev);
516
517 cancel_delayed_work_sync(&efx->monitor_work);
518 efx_selftest_async_cancel(efx);
519 cancel_work_sync(&efx->mac_work);
520}
521
522/* If the interface is supposed to be running but is not, start
523 * the hardware and software data path, regular activity for the port
524 * (MAC statistics, link polling, etc.) and schedule the port to be
525 * reconfigured. Interrupts must already be enabled. This function
526 * is safe to call multiple times, so long as the NIC is not disabled.
527 * Requires the RTNL lock.
528 */
529void efx_start_all(struct efx_nic *efx)
530{
531 EFX_ASSERT_RESET_SERIALISED(efx);
532 BUG_ON(efx->state == STATE_DISABLED);
533
534 /* Check that it is appropriate to restart the interface. All
535 * of these flags are safe to read under just the rtnl lock
536 */
537 if (efx->port_enabled || !netif_running(efx->net_dev) ||
538 efx->reset_pending)
539 return;
540
541 efx_start_port(efx);
542 efx_start_datapath(efx);
543
544 /* Start the hardware monitor if there is one */
545 efx_start_monitor(efx);
546
547 /* Link state detection is normally event-driven; we have
548 * to poll now because we could have missed a change
549 */
550 mutex_lock(&efx->mac_lock);
551 if (efx_mcdi_phy_poll(efx))
552 efx_link_status_changed(efx);
553 mutex_unlock(&efx->mac_lock);
554
555 if (efx->type->start_stats) {
556 efx->type->start_stats(efx);
557 efx->type->pull_stats(efx);
558 spin_lock_bh(&efx->stats_lock);
559 efx->type->update_stats(efx, NULL, NULL);
560 spin_unlock_bh(&efx->stats_lock);
561 }
562}
563
564/* Quiesce the hardware and software data path, and regular activity
565 * for the port without bringing the link down. Safe to call multiple
566 * times with the NIC in almost any state, but interrupts should be
567 * enabled. Requires the RTNL lock.
568 */
569void efx_stop_all(struct efx_nic *efx)
570{
571 EFX_ASSERT_RESET_SERIALISED(efx);
572
573 /* port_enabled can be read safely under the rtnl lock */
574 if (!efx->port_enabled)
575 return;
576
577 if (efx->type->update_stats) {
578 /* update stats before we go down so we can accurately count
579 * rx_nodesc_drops
580 */
581 efx->type->pull_stats(efx);
582 spin_lock_bh(&efx->stats_lock);
583 efx->type->update_stats(efx, NULL, NULL);
584 spin_unlock_bh(&efx->stats_lock);
585 efx->type->stop_stats(efx);
586 }
587
588 efx_stop_port(efx);
589
590 /* Stop the kernel transmit interface. This is only valid if
591 * the device is stopped or detached; otherwise the watchdog
592 * may fire immediately.
593 */
594 WARN_ON(netif_running(efx->net_dev) &&
595 netif_device_present(efx->net_dev));
596 netif_tx_disable(efx->net_dev);
597
598 efx_stop_datapath(efx);
599}
600
601/* Context: process, dev_base_lock or RTNL held, non-blocking. */
602void efx_net_stats(struct net_device *net_dev, struct rtnl_link_stats64 *stats)
603{
604 struct efx_nic *efx = efx_netdev_priv(net_dev);
605
606 spin_lock_bh(&efx->stats_lock);
607 efx_nic_update_stats_atomic(efx, NULL, stats);
608 spin_unlock_bh(&efx->stats_lock);
609}
610
611/* Push loopback/power/transmit disable settings to the PHY, and reconfigure
612 * the MAC appropriately. All other PHY configuration changes are pushed
613 * through phy_op->set_settings(), and pushed asynchronously to the MAC
614 * through efx_monitor().
615 *
616 * Callers must hold the mac_lock
617 */
618int __efx_reconfigure_port(struct efx_nic *efx)
619{
620 enum efx_phy_mode phy_mode;
621 int rc = 0;
622
623 WARN_ON(!mutex_is_locked(&efx->mac_lock));
624
625 /* Disable PHY transmit in mac level loopbacks */
626 phy_mode = efx->phy_mode;
627 if (LOOPBACK_INTERNAL(efx))
628 efx->phy_mode |= PHY_MODE_TX_DISABLED;
629 else
630 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
631
632 if (efx->type->reconfigure_port)
633 rc = efx->type->reconfigure_port(efx);
634
635 if (rc)
636 efx->phy_mode = phy_mode;
637
638 return rc;
639}
640
641/* Reinitialise the MAC to pick up new PHY settings, even if the port is
642 * disabled.
643 */
644int efx_reconfigure_port(struct efx_nic *efx)
645{
646 int rc;
647
648 EFX_ASSERT_RESET_SERIALISED(efx);
649
650 mutex_lock(&efx->mac_lock);
651 rc = __efx_reconfigure_port(efx);
652 mutex_unlock(&efx->mac_lock);
653
654 return rc;
655}
656
657/**************************************************************************
658 *
659 * Device reset and suspend
660 *
661 **************************************************************************/
662
663static void efx_wait_for_bist_end(struct efx_nic *efx)
664{
665 int i;
666
667 for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
668 if (efx_mcdi_poll_reboot(efx))
669 goto out;
670 msleep(BIST_WAIT_DELAY_MS);
671 }
672
673 netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
674out:
675 /* Either way unset the BIST flag. If we found no reboot we probably
676 * won't recover, but we should try.
677 */
678 efx->mc_bist_for_other_fn = false;
679}
680
681/* Try recovery mechanisms.
682 * For now only EEH is supported.
683 * Returns 0 if the recovery mechanisms are unsuccessful.
684 * Returns a non-zero value otherwise.
685 */
686int efx_try_recovery(struct efx_nic *efx)
687{
688#ifdef CONFIG_EEH
689 /* A PCI error can occur and not be seen by EEH because nothing
690 * happens on the PCI bus. In this case the driver may fail and
691 * schedule a 'recover or reset', leading to this recovery handler.
692 * Manually call the eeh failure check function.
693 */
694 struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
695 if (eeh_dev_check_failure(eehdev)) {
696 /* The EEH mechanisms will handle the error and reset the
697 * device if necessary.
698 */
699 return 1;
700 }
701#endif
702 return 0;
703}
704
705/* Tears down the entire software state and most of the hardware state
706 * before reset.
707 */
708void efx_reset_down(struct efx_nic *efx, enum reset_type method)
709{
710 EFX_ASSERT_RESET_SERIALISED(efx);
711
712 if (method == RESET_TYPE_MCDI_TIMEOUT)
713 efx->type->prepare_flr(efx);
714
715 efx_stop_all(efx);
716 efx_disable_interrupts(efx);
717
718 mutex_lock(&efx->mac_lock);
719 down_write(&efx->filter_sem);
720 mutex_lock(&efx->rss_lock);
721 efx->type->fini(efx);
722}
723
724/* Context: netif_tx_lock held, BHs disabled. */
725void efx_watchdog(struct net_device *net_dev, unsigned int txqueue)
726{
727 struct efx_nic *efx = efx_netdev_priv(net_dev);
728
729 netif_err(efx, tx_err, efx->net_dev,
730 "TX stuck with port_enabled=%d: resetting channels\n",
731 efx->port_enabled);
732
733 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
734}
735
736/* This function will always ensure that the locks acquired in
737 * efx_reset_down() are released. A failure return code indicates
738 * that we were unable to reinitialise the hardware, and the
739 * driver should be disabled. If ok is false, then the rx and tx
740 * engines are not restarted, pending a RESET_DISABLE.
741 */
742int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
743{
744 int rc;
745
746 EFX_ASSERT_RESET_SERIALISED(efx);
747
748 if (method == RESET_TYPE_MCDI_TIMEOUT)
749 efx->type->finish_flr(efx);
750
751 /* Ensure that SRAM is initialised even if we're disabling the device */
752 rc = efx->type->init(efx);
753 if (rc) {
754 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
755 goto fail;
756 }
757
758 if (!ok)
759 goto fail;
760
761 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
762 method != RESET_TYPE_DATAPATH) {
763 rc = efx_mcdi_port_reconfigure(efx);
764 if (rc && rc != -EPERM)
765 netif_err(efx, drv, efx->net_dev,
766 "could not restore PHY settings\n");
767 }
768
769 rc = efx_enable_interrupts(efx);
770 if (rc)
771 goto fail;
772
773#ifdef CONFIG_SFC_SRIOV
774 rc = efx->type->vswitching_restore(efx);
775 if (rc) /* not fatal; the PF will still work fine */
776 netif_warn(efx, probe, efx->net_dev,
777 "failed to restore vswitching rc=%d;"
778 " VFs may not function\n", rc);
779#endif
780
781 if (efx->type->rx_restore_rss_contexts)
782 efx->type->rx_restore_rss_contexts(efx);
783 mutex_unlock(&efx->rss_lock);
784 efx->type->filter_table_restore(efx);
785 up_write(&efx->filter_sem);
786 if (efx->type->sriov_reset)
787 efx->type->sriov_reset(efx);
788
789 mutex_unlock(&efx->mac_lock);
790
791 efx_start_all(efx);
792
793 if (efx->type->udp_tnl_push_ports)
794 efx->type->udp_tnl_push_ports(efx);
795
796 return 0;
797
798fail:
799 efx->port_initialized = false;
800
801 mutex_unlock(&efx->rss_lock);
802 up_write(&efx->filter_sem);
803 mutex_unlock(&efx->mac_lock);
804
805 return rc;
806}
807
808/* Reset the NIC using the specified method. Note that the reset may
809 * fail, in which case the card will be left in an unusable state.
810 *
811 * Caller must hold the rtnl_lock.
812 */
813int efx_reset(struct efx_nic *efx, enum reset_type method)
814{
815 int rc, rc2 = 0;
816 bool disabled;
817
818 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
819 RESET_TYPE(method));
820
821 efx_device_detach_sync(efx);
822 /* efx_reset_down() grabs locks that prevent recovery on EF100.
823 * EF100 reset is handled in the efx_nic_type callback below.
824 */
825 if (efx_nic_rev(efx) != EFX_REV_EF100)
826 efx_reset_down(efx, method);
827
828 rc = efx->type->reset(efx, method);
829 if (rc) {
830 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
831 goto out;
832 }
833
834 /* Clear flags for the scopes we covered. We assume the NIC and
835 * driver are now quiescent so that there is no race here.
836 */
837 if (method < RESET_TYPE_MAX_METHOD)
838 efx->reset_pending &= -(1 << (method + 1));
839 else /* it doesn't fit into the well-ordered scope hierarchy */
840 __clear_bit(method, &efx->reset_pending);
841
842 /* Reinitialise bus-mastering, which may have been turned off before
843 * the reset was scheduled. This is still appropriate, even in the
844 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
845 * can respond to requests.
846 */
847 pci_set_master(efx->pci_dev);
848
849out:
850 /* Leave device stopped if necessary */
851 disabled = rc ||
852 method == RESET_TYPE_DISABLE ||
853 method == RESET_TYPE_RECOVER_OR_DISABLE;
854 if (efx_nic_rev(efx) != EFX_REV_EF100)
855 rc2 = efx_reset_up(efx, method, !disabled);
856 if (rc2) {
857 disabled = true;
858 if (!rc)
859 rc = rc2;
860 }
861
862 if (disabled) {
863 dev_close(efx->net_dev);
864 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
865 efx->state = STATE_DISABLED;
866 } else {
867 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
868 efx_device_attach_if_not_resetting(efx);
869 }
870 return rc;
871}
872
873/* The worker thread exists so that code that cannot sleep can
874 * schedule a reset for later.
875 */
876static void efx_reset_work(struct work_struct *data)
877{
878 struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
879 unsigned long pending;
880 enum reset_type method;
881
882 pending = READ_ONCE(efx->reset_pending);
883 method = fls(pending) - 1;
884
885 if (method == RESET_TYPE_MC_BIST)
886 efx_wait_for_bist_end(efx);
887
888 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
889 method == RESET_TYPE_RECOVER_OR_ALL) &&
890 efx_try_recovery(efx))
891 return;
892
893 if (!pending)
894 return;
895
896 rtnl_lock();
897
898 /* We checked the state in efx_schedule_reset() but it may
899 * have changed by now. Now that we have the RTNL lock,
900 * it cannot change again.
901 */
902 if (efx_net_active(efx->state))
903 (void)efx_reset(efx, method);
904
905 rtnl_unlock();
906}
907
908void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
909{
910 enum reset_type method;
911
912 if (efx_recovering(efx->state)) {
913 netif_dbg(efx, drv, efx->net_dev,
914 "recovering: skip scheduling %s reset\n",
915 RESET_TYPE(type));
916 return;
917 }
918
919 switch (type) {
920 case RESET_TYPE_INVISIBLE:
921 case RESET_TYPE_ALL:
922 case RESET_TYPE_RECOVER_OR_ALL:
923 case RESET_TYPE_WORLD:
924 case RESET_TYPE_DISABLE:
925 case RESET_TYPE_RECOVER_OR_DISABLE:
926 case RESET_TYPE_DATAPATH:
927 case RESET_TYPE_MC_BIST:
928 case RESET_TYPE_MCDI_TIMEOUT:
929 method = type;
930 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
931 RESET_TYPE(method));
932 break;
933 default:
934 method = efx->type->map_reset_reason(type);
935 netif_dbg(efx, drv, efx->net_dev,
936 "scheduling %s reset for %s\n",
937 RESET_TYPE(method), RESET_TYPE(type));
938 break;
939 }
940
941 set_bit(method, &efx->reset_pending);
942 smp_mb(); /* ensure we change reset_pending before checking state */
943
944 /* If we're not READY then just leave the flags set as the cue
945 * to abort probing or reschedule the reset later.
946 */
947 if (!efx_net_active(READ_ONCE(efx->state)))
948 return;
949
950 /* efx_process_channel() will no longer read events once a
951 * reset is scheduled. So switch back to poll'd MCDI completions.
952 */
953 efx_mcdi_mode_poll(efx);
954
955 efx_queue_reset_work(efx);
956}
957
958/**************************************************************************
959 *
960 * Dummy NIC operations
961 *
962 * Can be used for some unimplemented operations
963 * Needed so all function pointers are valid and do not have to be tested
964 * before use
965 *
966 **************************************************************************/
967int efx_port_dummy_op_int(struct efx_nic *efx)
968{
969 return 0;
970}
971void efx_port_dummy_op_void(struct efx_nic *efx) {}
972
973/**************************************************************************
974 *
975 * Data housekeeping
976 *
977 **************************************************************************/
978
979/* This zeroes out and then fills in the invariants in a struct
980 * efx_nic (including all sub-structures).
981 */
982int efx_init_struct(struct efx_nic *efx, struct pci_dev *pci_dev)
983{
984 int rc = -ENOMEM;
985
986 /* Initialise common structures */
987 INIT_LIST_HEAD(&efx->node);
988 INIT_LIST_HEAD(&efx->secondary_list);
989 spin_lock_init(&efx->biu_lock);
990#ifdef CONFIG_SFC_MTD
991 INIT_LIST_HEAD(&efx->mtd_list);
992#endif
993 INIT_WORK(&efx->reset_work, efx_reset_work);
994 INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
995 efx_selftest_async_init(efx);
996 efx->pci_dev = pci_dev;
997 efx->msg_enable = debug;
998 efx->state = STATE_UNINIT;
999 strscpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
1000
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 spin_lock_init(&efx->vf_reps_lock);
1026 INIT_LIST_HEAD(&efx->vf_reps);
1027 INIT_WORK(&efx->mac_work, efx_mac_work);
1028 init_waitqueue_head(&efx->flush_wq);
1029
1030 efx->tx_queues_per_channel = 1;
1031 efx->rxq_entries = EFX_DEFAULT_DMAQ_SIZE;
1032 efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1033
1034 efx->mem_bar = UINT_MAX;
1035
1036 rc = efx_init_channels(efx);
1037 if (rc)
1038 goto fail;
1039
1040 /* Would be good to use the net_dev name, but we're too early */
1041 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
1042 pci_name(pci_dev));
1043 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
1044 if (!efx->workqueue) {
1045 rc = -ENOMEM;
1046 goto fail;
1047 }
1048
1049 return 0;
1050
1051fail:
1052 efx_fini_struct(efx);
1053 return rc;
1054}
1055
1056void efx_fini_struct(struct efx_nic *efx)
1057{
1058#ifdef CONFIG_RFS_ACCEL
1059 kfree(efx->rps_hash_table);
1060#endif
1061
1062 efx_fini_channels(efx);
1063
1064 kfree(efx->vpd_sn);
1065
1066 if (efx->workqueue) {
1067 destroy_workqueue(efx->workqueue);
1068 efx->workqueue = NULL;
1069 }
1070}
1071
1072/* This configures the PCI device to enable I/O and DMA. */
1073int efx_init_io(struct efx_nic *efx, int bar, dma_addr_t dma_mask,
1074 unsigned int mem_map_size)
1075{
1076 struct pci_dev *pci_dev = efx->pci_dev;
1077 int rc;
1078
1079 efx->mem_bar = UINT_MAX;
1080 pci_dbg(pci_dev, "initialising I/O bar=%d\n", bar);
1081
1082 rc = pci_enable_device(pci_dev);
1083 if (rc) {
1084 pci_err(pci_dev, "failed to enable PCI device\n");
1085 goto fail1;
1086 }
1087
1088 pci_set_master(pci_dev);
1089
1090 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1091 if (rc) {
1092 pci_err(efx->pci_dev, "could not find a suitable DMA mask\n");
1093 goto fail2;
1094 }
1095 pci_dbg(efx->pci_dev, "using DMA mask %llx\n", (unsigned long long)dma_mask);
1096
1097 efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1098 if (!efx->membase_phys) {
1099 pci_err(efx->pci_dev,
1100 "ERROR: No BAR%d mapping from the BIOS. Try pci=realloc on the kernel command line\n",
1101 bar);
1102 rc = -ENODEV;
1103 goto fail3;
1104 }
1105
1106 rc = pci_request_region(pci_dev, bar, "sfc");
1107 if (rc) {
1108 pci_err(efx->pci_dev,
1109 "request for memory BAR[%d] failed\n", bar);
1110 rc = -EIO;
1111 goto fail3;
1112 }
1113 efx->mem_bar = bar;
1114 efx->membase = ioremap(efx->membase_phys, mem_map_size);
1115 if (!efx->membase) {
1116 pci_err(efx->pci_dev,
1117 "could not map memory BAR[%d] at %llx+%x\n", bar,
1118 (unsigned long long)efx->membase_phys, mem_map_size);
1119 rc = -ENOMEM;
1120 goto fail4;
1121 }
1122 pci_dbg(efx->pci_dev,
1123 "memory BAR[%d] at %llx+%x (virtual %p)\n", bar,
1124 (unsigned long long)efx->membase_phys, mem_map_size,
1125 efx->membase);
1126
1127 return 0;
1128
1129fail4:
1130 pci_release_region(efx->pci_dev, bar);
1131fail3:
1132 efx->membase_phys = 0;
1133fail2:
1134 pci_disable_device(efx->pci_dev);
1135fail1:
1136 return rc;
1137}
1138
1139void efx_fini_io(struct efx_nic *efx)
1140{
1141 pci_dbg(efx->pci_dev, "shutting down I/O\n");
1142
1143 if (efx->membase) {
1144 iounmap(efx->membase);
1145 efx->membase = NULL;
1146 }
1147
1148 if (efx->membase_phys) {
1149 pci_release_region(efx->pci_dev, efx->mem_bar);
1150 efx->membase_phys = 0;
1151 efx->mem_bar = UINT_MAX;
1152 }
1153
1154 /* Don't disable bus-mastering if VFs are assigned */
1155 if (!pci_vfs_assigned(efx->pci_dev))
1156 pci_disable_device(efx->pci_dev);
1157}
1158
1159#ifdef CONFIG_SFC_MCDI_LOGGING
1160static ssize_t mcdi_logging_show(struct device *dev,
1161 struct device_attribute *attr,
1162 char *buf)
1163{
1164 struct efx_nic *efx = dev_get_drvdata(dev);
1165 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1166
1167 return sysfs_emit(buf, "%d\n", mcdi->logging_enabled);
1168}
1169
1170static ssize_t mcdi_logging_store(struct device *dev,
1171 struct device_attribute *attr,
1172 const char *buf, size_t count)
1173{
1174 struct efx_nic *efx = dev_get_drvdata(dev);
1175 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1176 bool enable = count > 0 && *buf != '0';
1177
1178 mcdi->logging_enabled = enable;
1179 return count;
1180}
1181
1182static DEVICE_ATTR_RW(mcdi_logging);
1183
1184void efx_init_mcdi_logging(struct efx_nic *efx)
1185{
1186 int rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1187
1188 if (rc) {
1189 netif_warn(efx, drv, efx->net_dev,
1190 "failed to init net dev attributes\n");
1191 }
1192}
1193
1194void efx_fini_mcdi_logging(struct efx_nic *efx)
1195{
1196 device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1197}
1198#endif
1199
1200/* A PCI error affecting this device was detected.
1201 * At this point MMIO and DMA may be disabled.
1202 * Stop the software path and request a slot reset.
1203 */
1204static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
1205 pci_channel_state_t state)
1206{
1207 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
1208 struct efx_nic *efx = pci_get_drvdata(pdev);
1209
1210 if (state == pci_channel_io_perm_failure)
1211 return PCI_ERS_RESULT_DISCONNECT;
1212
1213 rtnl_lock();
1214
1215 if (efx->state != STATE_DISABLED) {
1216 efx->state = efx_recover(efx->state);
1217 efx->reset_pending = 0;
1218
1219 efx_device_detach_sync(efx);
1220
1221 if (efx_net_active(efx->state)) {
1222 efx_stop_all(efx);
1223 efx_disable_interrupts(efx);
1224 }
1225
1226 status = PCI_ERS_RESULT_NEED_RESET;
1227 } else {
1228 /* If the interface is disabled we don't want to do anything
1229 * with it.
1230 */
1231 status = PCI_ERS_RESULT_RECOVERED;
1232 }
1233
1234 rtnl_unlock();
1235
1236 pci_disable_device(pdev);
1237
1238 return status;
1239}
1240
1241/* Fake a successful reset, which will be performed later in efx_io_resume. */
1242static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
1243{
1244 struct efx_nic *efx = pci_get_drvdata(pdev);
1245 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
1246
1247 if (pci_enable_device(pdev)) {
1248 netif_err(efx, hw, efx->net_dev,
1249 "Cannot re-enable PCI device after reset.\n");
1250 status = PCI_ERS_RESULT_DISCONNECT;
1251 }
1252
1253 return status;
1254}
1255
1256/* Perform the actual reset and resume I/O operations. */
1257static void efx_io_resume(struct pci_dev *pdev)
1258{
1259 struct efx_nic *efx = pci_get_drvdata(pdev);
1260 int rc;
1261
1262 rtnl_lock();
1263
1264 if (efx->state == STATE_DISABLED)
1265 goto out;
1266
1267 rc = efx_reset(efx, RESET_TYPE_ALL);
1268 if (rc) {
1269 netif_err(efx, hw, efx->net_dev,
1270 "efx_reset failed after PCI error (%d)\n", rc);
1271 } else {
1272 efx->state = efx_recovered(efx->state);
1273 netif_dbg(efx, hw, efx->net_dev,
1274 "Done resetting and resuming IO after PCI error.\n");
1275 }
1276
1277out:
1278 rtnl_unlock();
1279}
1280
1281/* For simplicity and reliability, we always require a slot reset and try to
1282 * reset the hardware when a pci error affecting the device is detected.
1283 * We leave both the link_reset and mmio_enabled callback unimplemented:
1284 * with our request for slot reset the mmio_enabled callback will never be
1285 * called, and the link_reset callback is not used by AER or EEH mechanisms.
1286 */
1287const struct pci_error_handlers efx_err_handlers = {
1288 .error_detected = efx_io_error_detected,
1289 .slot_reset = efx_io_slot_reset,
1290 .resume = efx_io_resume,
1291};
1292
1293/* Determine whether the NIC will be able to handle TX offloads for a given
1294 * encapsulated packet.
1295 */
1296static bool efx_can_encap_offloads(struct efx_nic *efx, struct sk_buff *skb)
1297{
1298 struct gre_base_hdr *greh;
1299 __be16 dst_port;
1300 u8 ipproto;
1301
1302 /* Does the NIC support encap offloads?
1303 * If not, we should never get here, because we shouldn't have
1304 * advertised encap offload feature flags in the first place.
1305 */
1306 if (WARN_ON_ONCE(!efx->type->udp_tnl_has_port))
1307 return false;
1308
1309 /* Determine encapsulation protocol in use */
1310 switch (skb->protocol) {
1311 case htons(ETH_P_IP):
1312 ipproto = ip_hdr(skb)->protocol;
1313 break;
1314 case htons(ETH_P_IPV6):
1315 /* If there are extension headers, this will cause us to
1316 * think we can't offload something that we maybe could have.
1317 */
1318 ipproto = ipv6_hdr(skb)->nexthdr;
1319 break;
1320 default:
1321 /* Not IP, so can't offload it */
1322 return false;
1323 }
1324 switch (ipproto) {
1325 case IPPROTO_GRE:
1326 /* We support NVGRE but not IP over GRE or random gretaps.
1327 * Specifically, the NIC will accept GRE as encapsulated if
1328 * the inner protocol is Ethernet, but only handle it
1329 * correctly if the GRE header is 8 bytes long. Moreover,
1330 * it will not update the Checksum or Sequence Number fields
1331 * if they are present. (The Routing Present flag,
1332 * GRE_ROUTING, cannot be set else the header would be more
1333 * than 8 bytes long; so we don't have to worry about it.)
1334 */
1335 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER)
1336 return false;
1337 if (ntohs(skb->inner_protocol) != ETH_P_TEB)
1338 return false;
1339 if (skb_inner_mac_header(skb) - skb_transport_header(skb) != 8)
1340 return false;
1341 greh = (struct gre_base_hdr *)skb_transport_header(skb);
1342 return !(greh->flags & (GRE_CSUM | GRE_SEQ));
1343 case IPPROTO_UDP:
1344 /* If the port is registered for a UDP tunnel, we assume the
1345 * packet is for that tunnel, and the NIC will handle it as
1346 * such. If not, the NIC won't know what to do with it.
1347 */
1348 dst_port = udp_hdr(skb)->dest;
1349 return efx->type->udp_tnl_has_port(efx, dst_port);
1350 default:
1351 return false;
1352 }
1353}
1354
1355netdev_features_t efx_features_check(struct sk_buff *skb, struct net_device *dev,
1356 netdev_features_t features)
1357{
1358 struct efx_nic *efx = efx_netdev_priv(dev);
1359
1360 if (skb->encapsulation) {
1361 if (features & NETIF_F_GSO_MASK)
1362 /* Hardware can only do TSO with at most 208 bytes
1363 * of headers.
1364 */
1365 if (skb_inner_transport_offset(skb) >
1366 EFX_TSO2_MAX_HDRLEN)
1367 features &= ~(NETIF_F_GSO_MASK);
1368 if (features & (NETIF_F_GSO_MASK | NETIF_F_CSUM_MASK))
1369 if (!efx_can_encap_offloads(efx, skb))
1370 features &= ~(NETIF_F_GSO_MASK |
1371 NETIF_F_CSUM_MASK);
1372 }
1373 return features;
1374}
1375
1376int efx_get_phys_port_id(struct net_device *net_dev,
1377 struct netdev_phys_item_id *ppid)
1378{
1379 struct efx_nic *efx = efx_netdev_priv(net_dev);
1380
1381 if (efx->type->get_phys_port_id)
1382 return efx->type->get_phys_port_id(efx, ppid);
1383 else
1384 return -EOPNOTSUPP;
1385}
1386
1387int efx_get_phys_port_name(struct net_device *net_dev, char *name, size_t len)
1388{
1389 struct efx_nic *efx = efx_netdev_priv(net_dev);
1390
1391 if (snprintf(name, len, "p%u", efx->port_num) >= len)
1392 return -EINVAL;
1393 return 0;
1394}
1395
1396void efx_detach_reps(struct efx_nic *efx)
1397{
1398 struct net_device *rep_dev;
1399 struct efx_rep *efv;
1400
1401 ASSERT_RTNL();
1402 netif_dbg(efx, drv, efx->net_dev, "Detaching VF representors\n");
1403 list_for_each_entry(efv, &efx->vf_reps, list) {
1404 rep_dev = efv->net_dev;
1405 if (!rep_dev)
1406 continue;
1407 netif_carrier_off(rep_dev);
1408 /* See efx_device_detach_sync() */
1409 netif_tx_lock_bh(rep_dev);
1410 netif_tx_stop_all_queues(rep_dev);
1411 netif_tx_unlock_bh(rep_dev);
1412 }
1413}
1414
1415void efx_attach_reps(struct efx_nic *efx)
1416{
1417 struct net_device *rep_dev;
1418 struct efx_rep *efv;
1419
1420 ASSERT_RTNL();
1421 netif_dbg(efx, drv, efx->net_dev, "Attaching VF representors\n");
1422 list_for_each_entry(efv, &efx->vf_reps, list) {
1423 rep_dev = efv->net_dev;
1424 if (!rep_dev)
1425 continue;
1426 netif_tx_wake_all_queues(rep_dev);
1427 netif_carrier_on(rep_dev);
1428 }
1429}