1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2018 Solarflare Communications Inc.
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.
11 #include "net_driver.h"
12 #include <linux/filter.h>
13 #include <linux/module.h>
14 #include <linux/netdevice.h>
16 #include "efx_common.h"
17 #include "efx_channels.h"
21 #include "rx_common.h"
22 #include "tx_common.h"
24 #include "mcdi_port_common.h"
26 #include "mcdi_pcol.h"
27 #include "ef100_rep.h"
29 static 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);
33 module_param(debug, uint, 0);
34 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
36 /* This is the time (in jiffies) between invocations of the hardware
39 static unsigned int efx_monitor_interval = 1 * HZ;
41 /* How often and how many times to poll for a reset while waiting for a
42 * BIST that another function started to complete.
44 #define BIST_WAIT_DELAY_MS 100
45 #define BIST_WAIT_DELAY_COUNT 100
47 /* Default stats update time */
48 #define STATS_PERIOD_MS_DEFAULT 1000
50 static const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
51 static 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)",
68 #define RESET_TYPE(type) \
69 STRING_TABLE_LOOKUP(type, efx_reset_type)
71 /* Loopback mode names (see LOOPBACK_MODE()) */
72 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
73 const 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",
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.
107 static struct workqueue_struct *reset_workqueue;
109 int efx_create_reset_workqueue(void)
111 reset_workqueue = create_singlethread_workqueue("sfc_reset");
112 if (!reset_workqueue) {
113 printk(KERN_ERR "Failed to create reset workqueue\n");
120 void efx_queue_reset_work(struct efx_nic *efx)
122 queue_work(reset_workqueue, &efx->reset_work);
125 void efx_flush_reset_workqueue(struct efx_nic *efx)
127 cancel_work_sync(&efx->reset_work);
130 void efx_destroy_reset_workqueue(void)
132 if (reset_workqueue) {
133 destroy_workqueue(reset_workqueue);
134 reset_workqueue = NULL;
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
141 void efx_mac_reconfigure(struct efx_nic *efx, bool mtu_only)
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);
150 /* Asynchronous work item for changing MAC promiscuity and multicast
151 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
154 static void efx_mac_work(struct work_struct *data)
156 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
158 mutex_lock(&efx->mac_lock);
159 if (efx->port_enabled)
160 efx_mac_reconfigure(efx, false);
161 mutex_unlock(&efx->mac_lock);
164 int efx_set_mac_address(struct net_device *net_dev, void *data)
166 struct efx_nic *efx = efx_netdev_priv(net_dev);
167 struct sockaddr *addr = data;
168 u8 *new_addr = addr->sa_data;
172 if (!is_valid_ether_addr(new_addr)) {
173 netif_err(efx, drv, efx->net_dev,
174 "invalid ethernet MAC address requested: %pM\n",
176 return -EADDRNOTAVAIL;
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);
185 eth_hw_addr_set(net_dev, old_addr);
190 /* Reconfigure the MAC */
191 mutex_lock(&efx->mac_lock);
192 efx_mac_reconfigure(efx, false);
193 mutex_unlock(&efx->mac_lock);
198 /* Context: netif_addr_lock held, BHs disabled. */
199 void efx_set_rx_mode(struct net_device *net_dev)
201 struct efx_nic *efx = efx_netdev_priv(net_dev);
203 if (efx->port_enabled)
204 queue_work(efx->workqueue, &efx->mac_work);
205 /* Otherwise efx_start_port() will do this */
208 int efx_set_features(struct net_device *net_dev, netdev_features_t data)
210 struct efx_nic *efx = efx_netdev_priv(net_dev);
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);
220 /* If Rx VLAN filter is changed, update filters via mac_reconfigure.
221 * If rx-fcs is changed, mac_reconfigure updates that too.
223 if ((net_dev->features ^ data) & (NETIF_F_HW_VLAN_CTAG_FILTER |
225 /* efx_set_rx_mode() will schedule MAC work to update filters
226 * when a new features are finally set in net_dev.
228 efx_set_rx_mode(net_dev);
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.
238 void efx_link_status_changed(struct efx_nic *efx)
240 struct efx_link_state *link_state = &efx->link_state;
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
247 if (!netif_running(efx->net_dev))
250 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
251 efx->n_link_state_changes++;
254 netif_carrier_on(efx->net_dev);
256 netif_carrier_off(efx->net_dev);
259 /* Status message for kernel log */
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",
266 netif_info(efx, link, efx->net_dev, "link down\n");
269 unsigned int efx_xdp_max_mtu(struct efx_nic *efx)
271 /* The maximum MTU that we can fit in a single page, allowing for
272 * framing, overhead and XDP headroom + tailroom.
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;
278 return PAGE_SIZE - overhead;
281 /* Context: process, rtnl_lock() held. */
282 int efx_change_mtu(struct net_device *net_dev, int new_mtu)
284 struct efx_nic *efx = efx_netdev_priv(net_dev);
287 rc = efx_check_disabled(efx);
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));
299 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
301 efx_device_detach_sync(efx);
304 mutex_lock(&efx->mac_lock);
305 net_dev->mtu = new_mtu;
306 efx_mac_reconfigure(efx, true);
307 mutex_unlock(&efx->mac_lock);
310 efx_device_attach_if_not_resetting(efx);
314 /**************************************************************************
318 **************************************************************************/
320 /* Run periodically off the general workqueue */
321 static void efx_monitor(struct work_struct *data)
323 struct efx_nic *efx = container_of(data, struct efx_nic,
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);
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.
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);
341 efx_start_monitor(efx);
344 void efx_start_monitor(struct efx_nic *efx)
346 if (efx->type->monitor)
347 queue_delayed_work(efx->workqueue, &efx->monitor_work,
348 efx_monitor_interval);
351 /**************************************************************************
353 * Event queue processing
355 *************************************************************************/
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
361 static void efx_start_datapath(struct efx_nic *efx)
363 netdev_features_t old_features = efx->net_dev->features;
364 bool old_rx_scatter = efx->rx_scatter;
367 /* Calculate the rx buffer allocation parameters required to
368 * support the current MTU, including padding for header
369 * alignment and overruns.
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);
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) >
386 efx->rx_scatter = true;
387 efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
388 efx->rx_buffer_order = 0;
390 efx->rx_scatter = false;
391 efx->rx_buffer_order = get_order(rx_buf_len);
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);
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);
406 /* Restore previously fixed features in hw_features and remove
407 * features which are fixed now
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);
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);
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
427 efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
428 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
430 /* Initialise the channels */
431 efx_start_channels(efx);
433 efx_ptp_start_datapath(efx);
435 if (netif_device_present(efx->net_dev))
436 netif_tx_wake_all_queues(efx->net_dev);
439 static void efx_stop_datapath(struct efx_nic *efx)
441 EFX_ASSERT_RESET_SERIALISED(efx);
442 BUG_ON(efx->port_enabled);
444 efx_ptp_stop_datapath(efx);
446 efx_stop_channels(efx);
449 /**************************************************************************
453 **************************************************************************/
455 /* Equivalent to efx_link_set_advertising with all-zeroes, except does not
456 * force the Autoneg bit on.
458 void efx_link_clear_advertising(struct efx_nic *efx)
460 bitmap_zero(efx->link_advertising, __ETHTOOL_LINK_MODE_MASK_NBITS);
461 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
464 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
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);
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;
479 static void efx_start_port(struct efx_nic *efx)
481 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
482 BUG_ON(efx->port_enabled);
484 mutex_lock(&efx->mac_lock);
485 efx->port_enabled = true;
487 /* Ensure MAC ingress/egress is enabled */
488 efx_mac_reconfigure(efx, false);
490 mutex_unlock(&efx->mac_lock);
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.
498 static void efx_stop_port(struct efx_nic *efx)
500 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
502 EFX_ASSERT_RESET_SERIALISED(efx);
504 mutex_lock(&efx->mac_lock);
505 efx->port_enabled = false;
506 mutex_unlock(&efx->mac_lock);
508 /* Serialise against efx_set_multicast_list() */
509 netif_addr_lock_bh(efx->net_dev);
510 netif_addr_unlock_bh(efx->net_dev);
512 cancel_delayed_work_sync(&efx->monitor_work);
513 efx_selftest_async_cancel(efx);
514 cancel_work_sync(&efx->mac_work);
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.
524 void efx_start_all(struct efx_nic *efx)
526 EFX_ASSERT_RESET_SERIALISED(efx);
527 BUG_ON(efx->state == STATE_DISABLED);
529 /* Check that it is appropriate to restart the interface. All
530 * of these flags are safe to read under just the rtnl lock
532 if (efx->port_enabled || !netif_running(efx->net_dev) ||
537 efx_start_datapath(efx);
539 /* Start the hardware monitor if there is one */
540 efx_start_monitor(efx);
542 efx_selftest_async_start(efx);
544 /* Link state detection is normally event-driven; we have
545 * to poll now because we could have missed a change
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);
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);
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.
566 void efx_stop_all(struct efx_nic *efx)
568 EFX_ASSERT_RESET_SERIALISED(efx);
570 /* port_enabled can be read safely under the rtnl lock */
571 if (!efx->port_enabled)
574 if (efx->type->update_stats) {
575 /* update stats before we go down so we can accurately count
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);
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.
591 WARN_ON(netif_running(efx->net_dev) &&
592 netif_device_present(efx->net_dev));
593 netif_tx_disable(efx->net_dev);
595 efx_stop_datapath(efx);
598 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
599 void efx_net_stats(struct net_device *net_dev, struct rtnl_link_stats64 *stats)
601 struct efx_nic *efx = efx_netdev_priv(net_dev);
603 spin_lock_bh(&efx->stats_lock);
604 efx_nic_update_stats_atomic(efx, NULL, stats);
605 spin_unlock_bh(&efx->stats_lock);
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().
613 * Callers must hold the mac_lock
615 int __efx_reconfigure_port(struct efx_nic *efx)
617 enum efx_phy_mode phy_mode;
620 WARN_ON(!mutex_is_locked(&efx->mac_lock));
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;
627 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
629 if (efx->type->reconfigure_port)
630 rc = efx->type->reconfigure_port(efx);
633 efx->phy_mode = phy_mode;
638 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
641 int efx_reconfigure_port(struct efx_nic *efx)
645 EFX_ASSERT_RESET_SERIALISED(efx);
647 mutex_lock(&efx->mac_lock);
648 rc = __efx_reconfigure_port(efx);
649 mutex_unlock(&efx->mac_lock);
654 /**************************************************************************
656 * Device reset and suspend
658 **************************************************************************/
660 static void efx_wait_for_bist_end(struct efx_nic *efx)
664 for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
665 if (efx_mcdi_poll_reboot(efx))
667 msleep(BIST_WAIT_DELAY_MS);
670 netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
672 /* Either way unset the BIST flag. If we found no reboot we probably
673 * won't recover, but we should try.
675 efx->mc_bist_for_other_fn = false;
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.
683 int efx_try_recovery(struct efx_nic *efx)
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.
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.
702 /* Tears down the entire software state and most of the hardware state
705 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
707 EFX_ASSERT_RESET_SERIALISED(efx);
709 if (method == RESET_TYPE_MCDI_TIMEOUT)
710 efx->type->prepare_flr(efx);
713 efx_disable_interrupts(efx);
715 mutex_lock(&efx->mac_lock);
716 down_write(&efx->filter_sem);
717 mutex_lock(&efx->rss_lock);
718 efx->type->fini(efx);
721 /* Context: netif_tx_lock held, BHs disabled. */
722 void efx_watchdog(struct net_device *net_dev, unsigned int txqueue)
724 struct efx_nic *efx = efx_netdev_priv(net_dev);
726 netif_err(efx, tx_err, efx->net_dev,
727 "TX stuck with port_enabled=%d: resetting channels\n",
730 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
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.
739 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
743 EFX_ASSERT_RESET_SERIALISED(efx);
745 if (method == RESET_TYPE_MCDI_TIMEOUT)
746 efx->type->finish_flr(efx);
748 /* Ensure that SRAM is initialised even if we're disabling the device */
749 rc = efx->type->init(efx);
751 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
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");
766 rc = efx_enable_interrupts(efx);
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);
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);
784 mutex_unlock(&efx->mac_lock);
788 if (efx->type->udp_tnl_push_ports)
789 efx->type->udp_tnl_push_ports(efx);
794 efx->port_initialized = false;
796 mutex_unlock(&efx->rss_lock);
797 up_write(&efx->filter_sem);
798 mutex_unlock(&efx->mac_lock);
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.
806 * Caller must hold the rtnl_lock.
808 int efx_reset(struct efx_nic *efx, enum reset_type method)
813 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
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.
820 if (efx_nic_rev(efx) != EFX_REV_EF100)
821 efx_reset_down(efx, method);
823 rc = efx->type->reset(efx, method);
825 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
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.
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);
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.
842 pci_set_master(efx->pci_dev);
845 /* Leave device stopped if necessary */
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);
858 dev_close(efx->net_dev);
859 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
860 efx->state = STATE_DISABLED;
862 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
863 efx_device_attach_if_not_resetting(efx);
868 /* The worker thread exists so that code that cannot sleep can
869 * schedule a reset for later.
871 static void efx_reset_work(struct work_struct *data)
873 struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
874 unsigned long pending;
875 enum reset_type method;
877 pending = READ_ONCE(efx->reset_pending);
878 method = fls(pending) - 1;
880 if (method == RESET_TYPE_MC_BIST)
881 efx_wait_for_bist_end(efx);
883 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
884 method == RESET_TYPE_RECOVER_OR_ALL) &&
885 efx_try_recovery(efx))
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.
897 if (efx_net_active(efx->state))
898 (void)efx_reset(efx, method);
903 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
905 enum reset_type method;
907 if (efx_recovering(efx->state)) {
908 netif_dbg(efx, drv, efx->net_dev,
909 "recovering: skip scheduling %s reset\n",
915 case RESET_TYPE_INVISIBLE:
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:
925 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
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));
936 set_bit(method, &efx->reset_pending);
937 smp_mb(); /* ensure we change reset_pending before checking state */
939 /* If we're not READY then just leave the flags set as the cue
940 * to abort probing or reschedule the reset later.
942 if (!efx_net_active(READ_ONCE(efx->state)))
945 /* efx_process_channel() will no longer read events once a
946 * reset is scheduled. So switch back to poll'd MCDI completions.
948 efx_mcdi_mode_poll(efx);
950 efx_queue_reset_work(efx);
953 /**************************************************************************
955 * Dummy NIC operations
957 * Can be used for some unimplemented operations
958 * Needed so all function pointers are valid and do not have to be tested
961 **************************************************************************/
962 int efx_port_dummy_op_int(struct efx_nic *efx)
966 void efx_port_dummy_op_void(struct efx_nic *efx) {}
968 /**************************************************************************
972 **************************************************************************/
974 /* This zeroes out and then fills in the invariants in a struct
975 * efx_nic (including all sub-structures).
977 int efx_init_struct(struct efx_nic *efx, struct pci_dev *pci_dev)
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);
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));
996 efx->rx_prefix_size = efx->type->rx_prefix_size;
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);
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);
1025 efx->tx_queues_per_channel = 1;
1026 efx->rxq_entries = EFX_DEFAULT_DMAQ_SIZE;
1027 efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1029 efx->mem_bar = UINT_MAX;
1031 rc = efx_init_channels(efx);
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",
1038 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
1039 if (!efx->workqueue) {
1047 efx_fini_struct(efx);
1051 void efx_fini_struct(struct efx_nic *efx)
1053 #ifdef CONFIG_RFS_ACCEL
1054 kfree(efx->rps_hash_table);
1057 efx_fini_channels(efx);
1061 if (efx->workqueue) {
1062 destroy_workqueue(efx->workqueue);
1063 efx->workqueue = NULL;
1067 /* This configures the PCI device to enable I/O and DMA. */
1068 int efx_init_io(struct efx_nic *efx, int bar, dma_addr_t dma_mask,
1069 unsigned int mem_map_size)
1071 struct pci_dev *pci_dev = efx->pci_dev;
1074 efx->mem_bar = UINT_MAX;
1075 pci_dbg(pci_dev, "initialising I/O bar=%d\n", bar);
1077 rc = pci_enable_device(pci_dev);
1079 pci_err(pci_dev, "failed to enable PCI device\n");
1083 pci_set_master(pci_dev);
1085 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1087 pci_err(efx->pci_dev, "could not find a suitable DMA mask\n");
1090 pci_dbg(efx->pci_dev, "using DMA mask %llx\n", (unsigned long long)dma_mask);
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",
1101 rc = pci_request_region(pci_dev, bar, "sfc");
1103 pci_err(efx->pci_dev,
1104 "request for memory BAR[%d] failed\n", 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);
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,
1125 pci_release_region(efx->pci_dev, bar);
1127 efx->membase_phys = 0;
1129 pci_disable_device(efx->pci_dev);
1134 void efx_fini_io(struct efx_nic *efx)
1136 pci_dbg(efx->pci_dev, "shutting down I/O\n");
1139 iounmap(efx->membase);
1140 efx->membase = NULL;
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;
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);
1154 #ifdef CONFIG_SFC_MCDI_LOGGING
1155 static ssize_t mcdi_logging_show(struct device *dev,
1156 struct device_attribute *attr,
1159 struct efx_nic *efx = dev_get_drvdata(dev);
1160 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
1162 return sysfs_emit(buf, "%d\n", mcdi->logging_enabled);
1165 static ssize_t mcdi_logging_store(struct device *dev,
1166 struct device_attribute *attr,
1167 const char *buf, size_t count)
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';
1173 mcdi->logging_enabled = enable;
1177 static DEVICE_ATTR_RW(mcdi_logging);
1179 void efx_init_mcdi_logging(struct efx_nic *efx)
1181 int rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
1184 netif_warn(efx, drv, efx->net_dev,
1185 "failed to init net dev attributes\n");
1189 void efx_fini_mcdi_logging(struct efx_nic *efx)
1191 device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
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.
1199 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
1200 pci_channel_state_t state)
1202 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
1203 struct efx_nic *efx = pci_get_drvdata(pdev);
1205 if (state == pci_channel_io_perm_failure)
1206 return PCI_ERS_RESULT_DISCONNECT;
1210 if (efx->state != STATE_DISABLED) {
1211 efx->state = efx_recover(efx->state);
1212 efx->reset_pending = 0;
1214 efx_device_detach_sync(efx);
1216 if (efx_net_active(efx->state)) {
1218 efx_disable_interrupts(efx);
1221 status = PCI_ERS_RESULT_NEED_RESET;
1223 /* If the interface is disabled we don't want to do anything
1226 status = PCI_ERS_RESULT_RECOVERED;
1231 pci_disable_device(pdev);
1236 /* Fake a successful reset, which will be performed later in efx_io_resume. */
1237 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
1239 struct efx_nic *efx = pci_get_drvdata(pdev);
1240 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
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;
1251 /* Perform the actual reset and resume I/O operations. */
1252 static void efx_io_resume(struct pci_dev *pdev)
1254 struct efx_nic *efx = pci_get_drvdata(pdev);
1259 if (efx->state == STATE_DISABLED)
1262 rc = efx_reset(efx, RESET_TYPE_ALL);
1264 netif_err(efx, hw, efx->net_dev,
1265 "efx_reset failed after PCI error (%d)\n", rc);
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");
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.
1282 const 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,
1288 /* Determine whether the NIC will be able to handle TX offloads for a given
1289 * encapsulated packet.
1291 static bool efx_can_encap_offloads(struct efx_nic *efx, struct sk_buff *skb)
1293 struct gre_base_hdr *greh;
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.
1301 if (WARN_ON_ONCE(!efx->type->udp_tnl_has_port))
1304 /* Determine encapsulation protocol in use */
1305 switch (skb->protocol) {
1306 case htons(ETH_P_IP):
1307 ipproto = ip_hdr(skb)->protocol;
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.
1313 ipproto = ipv6_hdr(skb)->nexthdr;
1316 /* Not IP, so can't offload it */
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.)
1330 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER)
1332 if (ntohs(skb->inner_protocol) != ETH_P_TEB)
1334 if (skb_inner_mac_header(skb) - skb_transport_header(skb) != 8)
1336 greh = (struct gre_base_hdr *)skb_transport_header(skb);
1337 return !(greh->flags & (GRE_CSUM | GRE_SEQ));
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.
1343 dst_port = udp_hdr(skb)->dest;
1344 return efx->type->udp_tnl_has_port(efx, dst_port);
1350 netdev_features_t efx_features_check(struct sk_buff *skb, struct net_device *dev,
1351 netdev_features_t features)
1353 struct efx_nic *efx = efx_netdev_priv(dev);
1355 if (skb->encapsulation) {
1356 if (features & NETIF_F_GSO_MASK)
1357 /* Hardware can only do TSO with at most 208 bytes
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 |
1371 int efx_get_phys_port_id(struct net_device *net_dev,
1372 struct netdev_phys_item_id *ppid)
1374 struct efx_nic *efx = efx_netdev_priv(net_dev);
1376 if (efx->type->get_phys_port_id)
1377 return efx->type->get_phys_port_id(efx, ppid);
1382 int efx_get_phys_port_name(struct net_device *net_dev, char *name, size_t len)
1384 struct efx_nic *efx = efx_netdev_priv(net_dev);
1386 if (snprintf(name, len, "p%u", efx->port_num) >= len)
1391 void efx_detach_reps(struct efx_nic *efx)
1393 struct net_device *rep_dev;
1394 struct efx_rep *efv;
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;
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);
1410 void efx_attach_reps(struct efx_nic *efx)
1412 struct net_device *rep_dev;
1413 struct efx_rep *efv;
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;
1421 netif_tx_wake_all_queues(rep_dev);
1422 netif_carrier_on(rep_dev);