1 /****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2013 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 <linux/module.h>
12 #include <linux/pci.h>
13 #include <linux/netdevice.h>
14 #include <linux/etherdevice.h>
15 #include <linux/delay.h>
16 #include <linux/notifier.h>
18 #include <linux/tcp.h>
20 #include <linux/ethtool.h>
21 #include <linux/topology.h>
22 #include <linux/gfp.h>
23 #include <linux/aer.h>
24 #include <linux/interrupt.h>
25 #include "net_driver.h"
27 #include <net/udp_tunnel.h>
35 #include "mcdi_pcol.h"
36 #include "workarounds.h"
38 /**************************************************************************
42 **************************************************************************
45 /* Loopback mode names (see LOOPBACK_MODE()) */
46 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
47 const char *const efx_loopback_mode_names[] = {
48 [LOOPBACK_NONE] = "NONE",
49 [LOOPBACK_DATA] = "DATAPATH",
50 [LOOPBACK_GMAC] = "GMAC",
51 [LOOPBACK_XGMII] = "XGMII",
52 [LOOPBACK_XGXS] = "XGXS",
53 [LOOPBACK_XAUI] = "XAUI",
54 [LOOPBACK_GMII] = "GMII",
55 [LOOPBACK_SGMII] = "SGMII",
56 [LOOPBACK_XGBR] = "XGBR",
57 [LOOPBACK_XFI] = "XFI",
58 [LOOPBACK_XAUI_FAR] = "XAUI_FAR",
59 [LOOPBACK_GMII_FAR] = "GMII_FAR",
60 [LOOPBACK_SGMII_FAR] = "SGMII_FAR",
61 [LOOPBACK_XFI_FAR] = "XFI_FAR",
62 [LOOPBACK_GPHY] = "GPHY",
63 [LOOPBACK_PHYXS] = "PHYXS",
64 [LOOPBACK_PCS] = "PCS",
65 [LOOPBACK_PMAPMD] = "PMA/PMD",
66 [LOOPBACK_XPORT] = "XPORT",
67 [LOOPBACK_XGMII_WS] = "XGMII_WS",
68 [LOOPBACK_XAUI_WS] = "XAUI_WS",
69 [LOOPBACK_XAUI_WS_FAR] = "XAUI_WS_FAR",
70 [LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
71 [LOOPBACK_GMII_WS] = "GMII_WS",
72 [LOOPBACK_XFI_WS] = "XFI_WS",
73 [LOOPBACK_XFI_WS_FAR] = "XFI_WS_FAR",
74 [LOOPBACK_PHYXS_WS] = "PHYXS_WS",
77 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
78 const char *const efx_reset_type_names[] = {
79 [RESET_TYPE_INVISIBLE] = "INVISIBLE",
80 [RESET_TYPE_ALL] = "ALL",
81 [RESET_TYPE_RECOVER_OR_ALL] = "RECOVER_OR_ALL",
82 [RESET_TYPE_WORLD] = "WORLD",
83 [RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
84 [RESET_TYPE_DATAPATH] = "DATAPATH",
85 [RESET_TYPE_MC_BIST] = "MC_BIST",
86 [RESET_TYPE_DISABLE] = "DISABLE",
87 [RESET_TYPE_TX_WATCHDOG] = "TX_WATCHDOG",
88 [RESET_TYPE_INT_ERROR] = "INT_ERROR",
89 [RESET_TYPE_DMA_ERROR] = "DMA_ERROR",
90 [RESET_TYPE_TX_SKIP] = "TX_SKIP",
91 [RESET_TYPE_MC_FAILURE] = "MC_FAILURE",
92 [RESET_TYPE_MCDI_TIMEOUT] = "MCDI_TIMEOUT (FLR)",
95 /* UDP tunnel type names */
96 static const char *const efx_udp_tunnel_type_names[] = {
97 [TUNNEL_ENCAP_UDP_PORT_ENTRY_VXLAN] = "vxlan",
98 [TUNNEL_ENCAP_UDP_PORT_ENTRY_GENEVE] = "geneve",
101 void efx_get_udp_tunnel_type_name(u16 type, char *buf, size_t buflen)
103 if (type < ARRAY_SIZE(efx_udp_tunnel_type_names) &&
104 efx_udp_tunnel_type_names[type] != NULL)
105 snprintf(buf, buflen, "%s", efx_udp_tunnel_type_names[type]);
107 snprintf(buf, buflen, "type %d", type);
110 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
111 * queued onto this work queue. This is not a per-nic work queue, because
112 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
114 static struct workqueue_struct *reset_workqueue;
116 /* How often and how many times to poll for a reset while waiting for a
117 * BIST that another function started to complete.
119 #define BIST_WAIT_DELAY_MS 100
120 #define BIST_WAIT_DELAY_COUNT 100
122 /**************************************************************************
124 * Configurable values
126 *************************************************************************/
129 * Use separate channels for TX and RX events
131 * Set this to 1 to use separate channels for TX and RX. It allows us
132 * to control interrupt affinity separately for TX and RX.
134 * This is only used in MSI-X interrupt mode
136 bool efx_separate_tx_channels;
137 module_param(efx_separate_tx_channels, bool, 0444);
138 MODULE_PARM_DESC(efx_separate_tx_channels,
139 "Use separate channels for TX and RX");
141 /* This is the weight assigned to each of the (per-channel) virtual
144 static int napi_weight = 64;
146 /* This is the time (in jiffies) between invocations of the hardware
148 * On Falcon-based NICs, this will:
149 * - Check the on-board hardware monitor;
150 * - Poll the link state and reconfigure the hardware as necessary.
151 * On Siena-based NICs for power systems with EEH support, this will give EEH a
154 static unsigned int efx_monitor_interval = 1 * HZ;
156 /* Initial interrupt moderation settings. They can be modified after
157 * module load with ethtool.
159 * The default for RX should strike a balance between increasing the
160 * round-trip latency and reducing overhead.
162 static unsigned int rx_irq_mod_usec = 60;
164 /* Initial interrupt moderation settings. They can be modified after
165 * module load with ethtool.
167 * This default is chosen to ensure that a 10G link does not go idle
168 * while a TX queue is stopped after it has become full. A queue is
169 * restarted when it drops below half full. The time this takes (assuming
170 * worst case 3 descriptors per packet and 1024 descriptors) is
171 * 512 / 3 * 1.2 = 205 usec.
173 static unsigned int tx_irq_mod_usec = 150;
175 /* This is the first interrupt mode to try out of:
180 static unsigned int interrupt_mode;
182 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
183 * i.e. the number of CPUs among which we may distribute simultaneous
184 * interrupt handling.
186 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
187 * The default (0) means to assign an interrupt to each core.
189 static unsigned int rss_cpus;
190 module_param(rss_cpus, uint, 0444);
191 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
193 static bool phy_flash_cfg;
194 module_param(phy_flash_cfg, bool, 0644);
195 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
197 static unsigned irq_adapt_low_thresh = 8000;
198 module_param(irq_adapt_low_thresh, uint, 0644);
199 MODULE_PARM_DESC(irq_adapt_low_thresh,
200 "Threshold score for reducing IRQ moderation");
202 static unsigned irq_adapt_high_thresh = 16000;
203 module_param(irq_adapt_high_thresh, uint, 0644);
204 MODULE_PARM_DESC(irq_adapt_high_thresh,
205 "Threshold score for increasing IRQ moderation");
207 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
208 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
209 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
210 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
211 module_param(debug, uint, 0);
212 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
214 /**************************************************************************
216 * Utility functions and prototypes
218 *************************************************************************/
220 static int efx_soft_enable_interrupts(struct efx_nic *efx);
221 static void efx_soft_disable_interrupts(struct efx_nic *efx);
222 static void efx_remove_channel(struct efx_channel *channel);
223 static void efx_remove_channels(struct efx_nic *efx);
224 static const struct efx_channel_type efx_default_channel_type;
225 static void efx_remove_port(struct efx_nic *efx);
226 static void efx_init_napi_channel(struct efx_channel *channel);
227 static void efx_fini_napi(struct efx_nic *efx);
228 static void efx_fini_napi_channel(struct efx_channel *channel);
229 static void efx_fini_struct(struct efx_nic *efx);
230 static void efx_start_all(struct efx_nic *efx);
231 static void efx_stop_all(struct efx_nic *efx);
233 #define EFX_ASSERT_RESET_SERIALISED(efx) \
235 if ((efx->state == STATE_READY) || \
236 (efx->state == STATE_RECOVERY) || \
237 (efx->state == STATE_DISABLED)) \
241 static int efx_check_disabled(struct efx_nic *efx)
243 if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
244 netif_err(efx, drv, efx->net_dev,
245 "device is disabled due to earlier errors\n");
251 /**************************************************************************
253 * Event queue processing
255 *************************************************************************/
257 /* Process channel's event queue
259 * This function is responsible for processing the event queue of a
260 * single channel. The caller must guarantee that this function will
261 * never be concurrently called more than once on the same channel,
262 * though different channels may be being processed concurrently.
264 static int efx_process_channel(struct efx_channel *channel, int budget)
266 struct efx_tx_queue *tx_queue;
269 if (unlikely(!channel->enabled))
272 efx_for_each_channel_tx_queue(tx_queue, channel) {
273 tx_queue->pkts_compl = 0;
274 tx_queue->bytes_compl = 0;
277 spent = efx_nic_process_eventq(channel, budget);
278 if (spent && efx_channel_has_rx_queue(channel)) {
279 struct efx_rx_queue *rx_queue =
280 efx_channel_get_rx_queue(channel);
282 efx_rx_flush_packet(channel);
283 efx_fast_push_rx_descriptors(rx_queue, true);
287 efx_for_each_channel_tx_queue(tx_queue, channel) {
288 if (tx_queue->bytes_compl) {
289 netdev_tx_completed_queue(tx_queue->core_txq,
290 tx_queue->pkts_compl, tx_queue->bytes_compl);
299 * NAPI guarantees serialisation of polls of the same device, which
300 * provides the guarantee required by efx_process_channel().
302 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel)
304 int step = efx->irq_mod_step_us;
306 if (channel->irq_mod_score < irq_adapt_low_thresh) {
307 if (channel->irq_moderation_us > step) {
308 channel->irq_moderation_us -= step;
309 efx->type->push_irq_moderation(channel);
311 } else if (channel->irq_mod_score > irq_adapt_high_thresh) {
312 if (channel->irq_moderation_us <
313 efx->irq_rx_moderation_us) {
314 channel->irq_moderation_us += step;
315 efx->type->push_irq_moderation(channel);
319 channel->irq_count = 0;
320 channel->irq_mod_score = 0;
323 static int efx_poll(struct napi_struct *napi, int budget)
325 struct efx_channel *channel =
326 container_of(napi, struct efx_channel, napi_str);
327 struct efx_nic *efx = channel->efx;
330 netif_vdbg(efx, intr, efx->net_dev,
331 "channel %d NAPI poll executing on CPU %d\n",
332 channel->channel, raw_smp_processor_id());
334 spent = efx_process_channel(channel, budget);
336 if (spent < budget) {
337 if (efx_channel_has_rx_queue(channel) &&
338 efx->irq_rx_adaptive &&
339 unlikely(++channel->irq_count == 1000)) {
340 efx_update_irq_mod(efx, channel);
343 #ifdef CONFIG_RFS_ACCEL
344 /* Perhaps expire some ARFS filters */
345 schedule_work(&channel->filter_work);
348 /* There is no race here; although napi_disable() will
349 * only wait for napi_complete(), this isn't a problem
350 * since efx_nic_eventq_read_ack() will have no effect if
351 * interrupts have already been disabled.
353 if (napi_complete_done(napi, spent))
354 efx_nic_eventq_read_ack(channel);
360 /* Create event queue
361 * Event queue memory allocations are done only once. If the channel
362 * is reset, the memory buffer will be reused; this guards against
363 * errors during channel reset and also simplifies interrupt handling.
365 static int efx_probe_eventq(struct efx_channel *channel)
367 struct efx_nic *efx = channel->efx;
368 unsigned long entries;
370 netif_dbg(efx, probe, efx->net_dev,
371 "chan %d create event queue\n", channel->channel);
373 /* Build an event queue with room for one event per tx and rx buffer,
374 * plus some extra for link state events and MCDI completions. */
375 entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
376 EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
377 channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
379 return efx_nic_probe_eventq(channel);
382 /* Prepare channel's event queue */
383 static int efx_init_eventq(struct efx_channel *channel)
385 struct efx_nic *efx = channel->efx;
388 EFX_WARN_ON_PARANOID(channel->eventq_init);
390 netif_dbg(efx, drv, efx->net_dev,
391 "chan %d init event queue\n", channel->channel);
393 rc = efx_nic_init_eventq(channel);
395 efx->type->push_irq_moderation(channel);
396 channel->eventq_read_ptr = 0;
397 channel->eventq_init = true;
402 /* Enable event queue processing and NAPI */
403 void efx_start_eventq(struct efx_channel *channel)
405 netif_dbg(channel->efx, ifup, channel->efx->net_dev,
406 "chan %d start event queue\n", channel->channel);
408 /* Make sure the NAPI handler sees the enabled flag set */
409 channel->enabled = true;
412 napi_enable(&channel->napi_str);
413 efx_nic_eventq_read_ack(channel);
416 /* Disable event queue processing and NAPI */
417 void efx_stop_eventq(struct efx_channel *channel)
419 if (!channel->enabled)
422 napi_disable(&channel->napi_str);
423 channel->enabled = false;
426 static void efx_fini_eventq(struct efx_channel *channel)
428 if (!channel->eventq_init)
431 netif_dbg(channel->efx, drv, channel->efx->net_dev,
432 "chan %d fini event queue\n", channel->channel);
434 efx_nic_fini_eventq(channel);
435 channel->eventq_init = false;
438 static void efx_remove_eventq(struct efx_channel *channel)
440 netif_dbg(channel->efx, drv, channel->efx->net_dev,
441 "chan %d remove event queue\n", channel->channel);
443 efx_nic_remove_eventq(channel);
446 /**************************************************************************
450 *************************************************************************/
452 /* Allocate and initialise a channel structure. */
453 static struct efx_channel *
454 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
456 struct efx_channel *channel;
457 struct efx_rx_queue *rx_queue;
458 struct efx_tx_queue *tx_queue;
461 channel = kzalloc(sizeof(*channel), GFP_KERNEL);
466 channel->channel = i;
467 channel->type = &efx_default_channel_type;
469 for (j = 0; j < EFX_TXQ_TYPES; j++) {
470 tx_queue = &channel->tx_queue[j];
472 tx_queue->queue = i * EFX_TXQ_TYPES + j;
473 tx_queue->channel = channel;
476 #ifdef CONFIG_RFS_ACCEL
477 INIT_WORK(&channel->filter_work, efx_filter_rfs_expire);
480 rx_queue = &channel->rx_queue;
482 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
487 /* Allocate and initialise a channel structure, copying parameters
488 * (but not resources) from an old channel structure.
490 static struct efx_channel *
491 efx_copy_channel(const struct efx_channel *old_channel)
493 struct efx_channel *channel;
494 struct efx_rx_queue *rx_queue;
495 struct efx_tx_queue *tx_queue;
498 channel = kmalloc(sizeof(*channel), GFP_KERNEL);
502 *channel = *old_channel;
504 channel->napi_dev = NULL;
505 INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
506 channel->napi_str.napi_id = 0;
507 channel->napi_str.state = 0;
508 memset(&channel->eventq, 0, sizeof(channel->eventq));
510 for (j = 0; j < EFX_TXQ_TYPES; j++) {
511 tx_queue = &channel->tx_queue[j];
512 if (tx_queue->channel)
513 tx_queue->channel = channel;
514 tx_queue->buffer = NULL;
515 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
518 rx_queue = &channel->rx_queue;
519 rx_queue->buffer = NULL;
520 memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
521 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
522 #ifdef CONFIG_RFS_ACCEL
523 INIT_WORK(&channel->filter_work, efx_filter_rfs_expire);
529 static int efx_probe_channel(struct efx_channel *channel)
531 struct efx_tx_queue *tx_queue;
532 struct efx_rx_queue *rx_queue;
535 netif_dbg(channel->efx, probe, channel->efx->net_dev,
536 "creating channel %d\n", channel->channel);
538 rc = channel->type->pre_probe(channel);
542 rc = efx_probe_eventq(channel);
546 efx_for_each_channel_tx_queue(tx_queue, channel) {
547 rc = efx_probe_tx_queue(tx_queue);
552 efx_for_each_channel_rx_queue(rx_queue, channel) {
553 rc = efx_probe_rx_queue(rx_queue);
561 efx_remove_channel(channel);
566 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
568 struct efx_nic *efx = channel->efx;
572 number = channel->channel;
573 if (efx->tx_channel_offset == 0) {
575 } else if (channel->channel < efx->tx_channel_offset) {
579 number -= efx->tx_channel_offset;
581 snprintf(buf, len, "%s%s-%d", efx->name, type, number);
584 static void efx_set_channel_names(struct efx_nic *efx)
586 struct efx_channel *channel;
588 efx_for_each_channel(channel, efx)
589 channel->type->get_name(channel,
590 efx->msi_context[channel->channel].name,
591 sizeof(efx->msi_context[0].name));
594 static int efx_probe_channels(struct efx_nic *efx)
596 struct efx_channel *channel;
599 /* Restart special buffer allocation */
600 efx->next_buffer_table = 0;
602 /* Probe channels in reverse, so that any 'extra' channels
603 * use the start of the buffer table. This allows the traffic
604 * channels to be resized without moving them or wasting the
605 * entries before them.
607 efx_for_each_channel_rev(channel, efx) {
608 rc = efx_probe_channel(channel);
610 netif_err(efx, probe, efx->net_dev,
611 "failed to create channel %d\n",
616 efx_set_channel_names(efx);
621 efx_remove_channels(efx);
625 /* Channels are shutdown and reinitialised whilst the NIC is running
626 * to propagate configuration changes (mtu, checksum offload), or
627 * to clear hardware error conditions
629 static void efx_start_datapath(struct efx_nic *efx)
631 netdev_features_t old_features = efx->net_dev->features;
632 bool old_rx_scatter = efx->rx_scatter;
633 struct efx_tx_queue *tx_queue;
634 struct efx_rx_queue *rx_queue;
635 struct efx_channel *channel;
638 /* Calculate the rx buffer allocation parameters required to
639 * support the current MTU, including padding for header
640 * alignment and overruns.
642 efx->rx_dma_len = (efx->rx_prefix_size +
643 EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
644 efx->type->rx_buffer_padding);
645 rx_buf_len = (sizeof(struct efx_rx_page_state) +
646 efx->rx_ip_align + efx->rx_dma_len);
647 if (rx_buf_len <= PAGE_SIZE) {
648 efx->rx_scatter = efx->type->always_rx_scatter;
649 efx->rx_buffer_order = 0;
650 } else if (efx->type->can_rx_scatter) {
651 BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
652 BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
653 2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
654 EFX_RX_BUF_ALIGNMENT) >
656 efx->rx_scatter = true;
657 efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
658 efx->rx_buffer_order = 0;
660 efx->rx_scatter = false;
661 efx->rx_buffer_order = get_order(rx_buf_len);
664 efx_rx_config_page_split(efx);
665 if (efx->rx_buffer_order)
666 netif_dbg(efx, drv, efx->net_dev,
667 "RX buf len=%u; page order=%u batch=%u\n",
668 efx->rx_dma_len, efx->rx_buffer_order,
669 efx->rx_pages_per_batch);
671 netif_dbg(efx, drv, efx->net_dev,
672 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
673 efx->rx_dma_len, efx->rx_page_buf_step,
674 efx->rx_bufs_per_page, efx->rx_pages_per_batch);
676 /* Restore previously fixed features in hw_features and remove
677 * features which are fixed now
679 efx->net_dev->hw_features |= efx->net_dev->features;
680 efx->net_dev->hw_features &= ~efx->fixed_features;
681 efx->net_dev->features |= efx->fixed_features;
682 if (efx->net_dev->features != old_features)
683 netdev_features_change(efx->net_dev);
685 /* RX filters may also have scatter-enabled flags */
686 if (efx->rx_scatter != old_rx_scatter)
687 efx->type->filter_update_rx_scatter(efx);
689 /* We must keep at least one descriptor in a TX ring empty.
690 * We could avoid this when the queue size does not exactly
691 * match the hardware ring size, but it's not that important.
692 * Therefore we stop the queue when one more skb might fill
693 * the ring completely. We wake it when half way back to
696 efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
697 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
699 /* Initialise the channels */
700 efx_for_each_channel(channel, efx) {
701 efx_for_each_channel_tx_queue(tx_queue, channel) {
702 efx_init_tx_queue(tx_queue);
703 atomic_inc(&efx->active_queues);
706 efx_for_each_channel_rx_queue(rx_queue, channel) {
707 efx_init_rx_queue(rx_queue);
708 atomic_inc(&efx->active_queues);
709 efx_stop_eventq(channel);
710 efx_fast_push_rx_descriptors(rx_queue, false);
711 efx_start_eventq(channel);
714 WARN_ON(channel->rx_pkt_n_frags);
717 efx_ptp_start_datapath(efx);
719 if (netif_device_present(efx->net_dev))
720 netif_tx_wake_all_queues(efx->net_dev);
723 static void efx_stop_datapath(struct efx_nic *efx)
725 struct efx_channel *channel;
726 struct efx_tx_queue *tx_queue;
727 struct efx_rx_queue *rx_queue;
730 EFX_ASSERT_RESET_SERIALISED(efx);
731 BUG_ON(efx->port_enabled);
733 efx_ptp_stop_datapath(efx);
736 efx_for_each_channel(channel, efx) {
737 efx_for_each_channel_rx_queue(rx_queue, channel)
738 rx_queue->refill_enabled = false;
741 efx_for_each_channel(channel, efx) {
742 /* RX packet processing is pipelined, so wait for the
743 * NAPI handler to complete. At least event queue 0
744 * might be kept active by non-data events, so don't
745 * use napi_synchronize() but actually disable NAPI
748 if (efx_channel_has_rx_queue(channel)) {
749 efx_stop_eventq(channel);
750 efx_start_eventq(channel);
754 rc = efx->type->fini_dmaq(efx);
756 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
758 netif_dbg(efx, drv, efx->net_dev,
759 "successfully flushed all queues\n");
762 efx_for_each_channel(channel, efx) {
763 efx_for_each_channel_rx_queue(rx_queue, channel)
764 efx_fini_rx_queue(rx_queue);
765 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
766 efx_fini_tx_queue(tx_queue);
770 static void efx_remove_channel(struct efx_channel *channel)
772 struct efx_tx_queue *tx_queue;
773 struct efx_rx_queue *rx_queue;
775 netif_dbg(channel->efx, drv, channel->efx->net_dev,
776 "destroy chan %d\n", channel->channel);
778 efx_for_each_channel_rx_queue(rx_queue, channel)
779 efx_remove_rx_queue(rx_queue);
780 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
781 efx_remove_tx_queue(tx_queue);
782 efx_remove_eventq(channel);
783 channel->type->post_remove(channel);
786 static void efx_remove_channels(struct efx_nic *efx)
788 struct efx_channel *channel;
790 efx_for_each_channel(channel, efx)
791 efx_remove_channel(channel);
795 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
797 struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
798 u32 old_rxq_entries, old_txq_entries;
799 unsigned i, next_buffer_table = 0;
802 rc = efx_check_disabled(efx);
806 /* Not all channels should be reallocated. We must avoid
807 * reallocating their buffer table entries.
809 efx_for_each_channel(channel, efx) {
810 struct efx_rx_queue *rx_queue;
811 struct efx_tx_queue *tx_queue;
813 if (channel->type->copy)
815 next_buffer_table = max(next_buffer_table,
816 channel->eventq.index +
817 channel->eventq.entries);
818 efx_for_each_channel_rx_queue(rx_queue, channel)
819 next_buffer_table = max(next_buffer_table,
820 rx_queue->rxd.index +
821 rx_queue->rxd.entries);
822 efx_for_each_channel_tx_queue(tx_queue, channel)
823 next_buffer_table = max(next_buffer_table,
824 tx_queue->txd.index +
825 tx_queue->txd.entries);
828 efx_device_detach_sync(efx);
830 efx_soft_disable_interrupts(efx);
832 /* Clone channels (where possible) */
833 memset(other_channel, 0, sizeof(other_channel));
834 for (i = 0; i < efx->n_channels; i++) {
835 channel = efx->channel[i];
836 if (channel->type->copy)
837 channel = channel->type->copy(channel);
842 other_channel[i] = channel;
845 /* Swap entry counts and channel pointers */
846 old_rxq_entries = efx->rxq_entries;
847 old_txq_entries = efx->txq_entries;
848 efx->rxq_entries = rxq_entries;
849 efx->txq_entries = txq_entries;
850 for (i = 0; i < efx->n_channels; i++) {
851 channel = efx->channel[i];
852 efx->channel[i] = other_channel[i];
853 other_channel[i] = channel;
856 /* Restart buffer table allocation */
857 efx->next_buffer_table = next_buffer_table;
859 for (i = 0; i < efx->n_channels; i++) {
860 channel = efx->channel[i];
861 if (!channel->type->copy)
863 rc = efx_probe_channel(channel);
866 efx_init_napi_channel(efx->channel[i]);
870 /* Destroy unused channel structures */
871 for (i = 0; i < efx->n_channels; i++) {
872 channel = other_channel[i];
873 if (channel && channel->type->copy) {
874 efx_fini_napi_channel(channel);
875 efx_remove_channel(channel);
880 rc2 = efx_soft_enable_interrupts(efx);
883 netif_err(efx, drv, efx->net_dev,
884 "unable to restart interrupts on channel reallocation\n");
885 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
888 efx_device_attach_if_not_resetting(efx);
894 efx->rxq_entries = old_rxq_entries;
895 efx->txq_entries = old_txq_entries;
896 for (i = 0; i < efx->n_channels; i++) {
897 channel = efx->channel[i];
898 efx->channel[i] = other_channel[i];
899 other_channel[i] = channel;
904 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
906 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
909 static bool efx_default_channel_want_txqs(struct efx_channel *channel)
911 return channel->channel - channel->efx->tx_channel_offset <
912 channel->efx->n_tx_channels;
915 static const struct efx_channel_type efx_default_channel_type = {
916 .pre_probe = efx_channel_dummy_op_int,
917 .post_remove = efx_channel_dummy_op_void,
918 .get_name = efx_get_channel_name,
919 .copy = efx_copy_channel,
920 .want_txqs = efx_default_channel_want_txqs,
921 .keep_eventq = false,
925 int efx_channel_dummy_op_int(struct efx_channel *channel)
930 void efx_channel_dummy_op_void(struct efx_channel *channel)
934 /**************************************************************************
938 **************************************************************************/
940 /* This ensures that the kernel is kept informed (via
941 * netif_carrier_on/off) of the link status, and also maintains the
942 * link status's stop on the port's TX queue.
944 void efx_link_status_changed(struct efx_nic *efx)
946 struct efx_link_state *link_state = &efx->link_state;
948 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
949 * that no events are triggered between unregister_netdev() and the
950 * driver unloading. A more general condition is that NETDEV_CHANGE
951 * can only be generated between NETDEV_UP and NETDEV_DOWN */
952 if (!netif_running(efx->net_dev))
955 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
956 efx->n_link_state_changes++;
959 netif_carrier_on(efx->net_dev);
961 netif_carrier_off(efx->net_dev);
964 /* Status message for kernel log */
966 netif_info(efx, link, efx->net_dev,
967 "link up at %uMbps %s-duplex (MTU %d)\n",
968 link_state->speed, link_state->fd ? "full" : "half",
971 netif_info(efx, link, efx->net_dev, "link down\n");
974 void efx_link_set_advertising(struct efx_nic *efx,
975 const unsigned long *advertising)
977 memcpy(efx->link_advertising, advertising,
978 sizeof(__ETHTOOL_DECLARE_LINK_MODE_MASK()));
980 efx->link_advertising[0] |= ADVERTISED_Autoneg;
981 if (advertising[0] & ADVERTISED_Pause)
982 efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
984 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
985 if (advertising[0] & ADVERTISED_Asym_Pause)
986 efx->wanted_fc ^= EFX_FC_TX;
989 /* Equivalent to efx_link_set_advertising with all-zeroes, except does not
990 * force the Autoneg bit on.
992 void efx_link_clear_advertising(struct efx_nic *efx)
994 bitmap_zero(efx->link_advertising, __ETHTOOL_LINK_MODE_MASK_NBITS);
995 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
998 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
1000 efx->wanted_fc = wanted_fc;
1001 if (efx->link_advertising[0]) {
1002 if (wanted_fc & EFX_FC_RX)
1003 efx->link_advertising[0] |= (ADVERTISED_Pause |
1004 ADVERTISED_Asym_Pause);
1006 efx->link_advertising[0] &= ~(ADVERTISED_Pause |
1007 ADVERTISED_Asym_Pause);
1008 if (wanted_fc & EFX_FC_TX)
1009 efx->link_advertising[0] ^= ADVERTISED_Asym_Pause;
1013 static void efx_fini_port(struct efx_nic *efx);
1015 /* We assume that efx->type->reconfigure_mac will always try to sync RX
1016 * filters and therefore needs to read-lock the filter table against freeing
1018 void efx_mac_reconfigure(struct efx_nic *efx)
1020 down_read(&efx->filter_sem);
1021 efx->type->reconfigure_mac(efx);
1022 up_read(&efx->filter_sem);
1025 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
1026 * the MAC appropriately. All other PHY configuration changes are pushed
1027 * through phy_op->set_settings(), and pushed asynchronously to the MAC
1028 * through efx_monitor().
1030 * Callers must hold the mac_lock
1032 int __efx_reconfigure_port(struct efx_nic *efx)
1034 enum efx_phy_mode phy_mode;
1037 WARN_ON(!mutex_is_locked(&efx->mac_lock));
1039 /* Disable PHY transmit in mac level loopbacks */
1040 phy_mode = efx->phy_mode;
1041 if (LOOPBACK_INTERNAL(efx))
1042 efx->phy_mode |= PHY_MODE_TX_DISABLED;
1044 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
1046 rc = efx->type->reconfigure_port(efx);
1049 efx->phy_mode = phy_mode;
1054 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1056 int efx_reconfigure_port(struct efx_nic *efx)
1060 EFX_ASSERT_RESET_SERIALISED(efx);
1062 mutex_lock(&efx->mac_lock);
1063 rc = __efx_reconfigure_port(efx);
1064 mutex_unlock(&efx->mac_lock);
1069 /* Asynchronous work item for changing MAC promiscuity and multicast
1070 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
1072 static void efx_mac_work(struct work_struct *data)
1074 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1076 mutex_lock(&efx->mac_lock);
1077 if (efx->port_enabled)
1078 efx_mac_reconfigure(efx);
1079 mutex_unlock(&efx->mac_lock);
1082 static int efx_probe_port(struct efx_nic *efx)
1086 netif_dbg(efx, probe, efx->net_dev, "create port\n");
1089 efx->phy_mode = PHY_MODE_SPECIAL;
1091 /* Connect up MAC/PHY operations table */
1092 rc = efx->type->probe_port(efx);
1096 /* Initialise MAC address to permanent address */
1097 ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1102 static int efx_init_port(struct efx_nic *efx)
1106 netif_dbg(efx, drv, efx->net_dev, "init port\n");
1108 mutex_lock(&efx->mac_lock);
1110 rc = efx->phy_op->init(efx);
1114 efx->port_initialized = true;
1116 /* Reconfigure the MAC before creating dma queues (required for
1117 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1118 efx_mac_reconfigure(efx);
1120 /* Ensure the PHY advertises the correct flow control settings */
1121 rc = efx->phy_op->reconfigure(efx);
1122 if (rc && rc != -EPERM)
1125 mutex_unlock(&efx->mac_lock);
1129 efx->phy_op->fini(efx);
1131 mutex_unlock(&efx->mac_lock);
1135 static void efx_start_port(struct efx_nic *efx)
1137 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1138 BUG_ON(efx->port_enabled);
1140 mutex_lock(&efx->mac_lock);
1141 efx->port_enabled = true;
1143 /* Ensure MAC ingress/egress is enabled */
1144 efx_mac_reconfigure(efx);
1146 mutex_unlock(&efx->mac_lock);
1149 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1150 * and the async self-test, wait for them to finish and prevent them
1151 * being scheduled again. This doesn't cover online resets, which
1152 * should only be cancelled when removing the device.
1154 static void efx_stop_port(struct efx_nic *efx)
1156 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1158 EFX_ASSERT_RESET_SERIALISED(efx);
1160 mutex_lock(&efx->mac_lock);
1161 efx->port_enabled = false;
1162 mutex_unlock(&efx->mac_lock);
1164 /* Serialise against efx_set_multicast_list() */
1165 netif_addr_lock_bh(efx->net_dev);
1166 netif_addr_unlock_bh(efx->net_dev);
1168 cancel_delayed_work_sync(&efx->monitor_work);
1169 efx_selftest_async_cancel(efx);
1170 cancel_work_sync(&efx->mac_work);
1173 static void efx_fini_port(struct efx_nic *efx)
1175 netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1177 if (!efx->port_initialized)
1180 efx->phy_op->fini(efx);
1181 efx->port_initialized = false;
1183 efx->link_state.up = false;
1184 efx_link_status_changed(efx);
1187 static void efx_remove_port(struct efx_nic *efx)
1189 netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1191 efx->type->remove_port(efx);
1194 /**************************************************************************
1198 **************************************************************************/
1200 static LIST_HEAD(efx_primary_list);
1201 static LIST_HEAD(efx_unassociated_list);
1203 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1205 return left->type == right->type &&
1206 left->vpd_sn && right->vpd_sn &&
1207 !strcmp(left->vpd_sn, right->vpd_sn);
1210 static void efx_associate(struct efx_nic *efx)
1212 struct efx_nic *other, *next;
1214 if (efx->primary == efx) {
1215 /* Adding primary function; look for secondaries */
1217 netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1218 list_add_tail(&efx->node, &efx_primary_list);
1220 list_for_each_entry_safe(other, next, &efx_unassociated_list,
1222 if (efx_same_controller(efx, other)) {
1223 list_del(&other->node);
1224 netif_dbg(other, probe, other->net_dev,
1225 "moving to secondary list of %s %s\n",
1226 pci_name(efx->pci_dev),
1227 efx->net_dev->name);
1228 list_add_tail(&other->node,
1229 &efx->secondary_list);
1230 other->primary = efx;
1234 /* Adding secondary function; look for primary */
1236 list_for_each_entry(other, &efx_primary_list, node) {
1237 if (efx_same_controller(efx, other)) {
1238 netif_dbg(efx, probe, efx->net_dev,
1239 "adding to secondary list of %s %s\n",
1240 pci_name(other->pci_dev),
1241 other->net_dev->name);
1242 list_add_tail(&efx->node,
1243 &other->secondary_list);
1244 efx->primary = other;
1249 netif_dbg(efx, probe, efx->net_dev,
1250 "adding to unassociated list\n");
1251 list_add_tail(&efx->node, &efx_unassociated_list);
1255 static void efx_dissociate(struct efx_nic *efx)
1257 struct efx_nic *other, *next;
1259 list_del(&efx->node);
1260 efx->primary = NULL;
1262 list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1263 list_del(&other->node);
1264 netif_dbg(other, probe, other->net_dev,
1265 "moving to unassociated list\n");
1266 list_add_tail(&other->node, &efx_unassociated_list);
1267 other->primary = NULL;
1271 /* This configures the PCI device to enable I/O and DMA. */
1272 static int efx_init_io(struct efx_nic *efx)
1274 struct pci_dev *pci_dev = efx->pci_dev;
1275 dma_addr_t dma_mask = efx->type->max_dma_mask;
1276 unsigned int mem_map_size = efx->type->mem_map_size(efx);
1279 netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1281 bar = efx->type->mem_bar(efx);
1283 rc = pci_enable_device(pci_dev);
1285 netif_err(efx, probe, efx->net_dev,
1286 "failed to enable PCI device\n");
1290 pci_set_master(pci_dev);
1292 /* Set the PCI DMA mask. Try all possibilities from our genuine mask
1293 * down to 32 bits, because some architectures will allow 40 bit
1294 * masks event though they reject 46 bit masks.
1296 while (dma_mask > 0x7fffffffUL) {
1297 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1303 netif_err(efx, probe, efx->net_dev,
1304 "could not find a suitable DMA mask\n");
1307 netif_dbg(efx, probe, efx->net_dev,
1308 "using DMA mask %llx\n", (unsigned long long) dma_mask);
1310 efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1311 rc = pci_request_region(pci_dev, bar, "sfc");
1313 netif_err(efx, probe, efx->net_dev,
1314 "request for memory BAR failed\n");
1318 efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1319 if (!efx->membase) {
1320 netif_err(efx, probe, efx->net_dev,
1321 "could not map memory BAR at %llx+%x\n",
1322 (unsigned long long)efx->membase_phys, mem_map_size);
1326 netif_dbg(efx, probe, efx->net_dev,
1327 "memory BAR at %llx+%x (virtual %p)\n",
1328 (unsigned long long)efx->membase_phys, mem_map_size,
1334 pci_release_region(efx->pci_dev, bar);
1336 efx->membase_phys = 0;
1338 pci_disable_device(efx->pci_dev);
1343 static void efx_fini_io(struct efx_nic *efx)
1347 netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1350 iounmap(efx->membase);
1351 efx->membase = NULL;
1354 if (efx->membase_phys) {
1355 bar = efx->type->mem_bar(efx);
1356 pci_release_region(efx->pci_dev, bar);
1357 efx->membase_phys = 0;
1360 /* Don't disable bus-mastering if VFs are assigned */
1361 if (!pci_vfs_assigned(efx->pci_dev))
1362 pci_disable_device(efx->pci_dev);
1365 void efx_set_default_rx_indir_table(struct efx_nic *efx,
1366 struct efx_rss_context *ctx)
1370 for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++)
1371 ctx->rx_indir_table[i] =
1372 ethtool_rxfh_indir_default(i, efx->rss_spread);
1375 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1377 cpumask_var_t thread_mask;
1384 if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1385 netif_warn(efx, probe, efx->net_dev,
1386 "RSS disabled due to allocation failure\n");
1391 for_each_online_cpu(cpu) {
1392 if (!cpumask_test_cpu(cpu, thread_mask)) {
1394 cpumask_or(thread_mask, thread_mask,
1395 topology_sibling_cpumask(cpu));
1399 free_cpumask_var(thread_mask);
1402 if (count > EFX_MAX_RX_QUEUES) {
1403 netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn,
1404 "Reducing number of rx queues from %u to %u.\n",
1405 count, EFX_MAX_RX_QUEUES);
1406 count = EFX_MAX_RX_QUEUES;
1409 /* If RSS is requested for the PF *and* VFs then we can't write RSS
1410 * table entries that are inaccessible to VFs
1412 #ifdef CONFIG_SFC_SRIOV
1413 if (efx->type->sriov_wanted) {
1414 if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1415 count > efx_vf_size(efx)) {
1416 netif_warn(efx, probe, efx->net_dev,
1417 "Reducing number of RSS channels from %u to %u for "
1418 "VF support. Increase vf-msix-limit to use more "
1419 "channels on the PF.\n",
1420 count, efx_vf_size(efx));
1421 count = efx_vf_size(efx);
1429 /* Probe the number and type of interrupts we are able to obtain, and
1430 * the resulting numbers of channels and RX queues.
1432 static int efx_probe_interrupts(struct efx_nic *efx)
1434 unsigned int extra_channels = 0;
1438 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1439 if (efx->extra_channel_type[i])
1442 if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1443 struct msix_entry xentries[EFX_MAX_CHANNELS];
1444 unsigned int n_channels;
1446 n_channels = efx_wanted_parallelism(efx);
1447 if (efx_separate_tx_channels)
1449 n_channels += extra_channels;
1450 n_channels = min(n_channels, efx->max_channels);
1452 for (i = 0; i < n_channels; i++)
1453 xentries[i].entry = i;
1454 rc = pci_enable_msix_range(efx->pci_dev,
1455 xentries, 1, n_channels);
1457 /* Fall back to single channel MSI */
1458 netif_err(efx, drv, efx->net_dev,
1459 "could not enable MSI-X\n");
1460 if (efx->type->min_interrupt_mode >= EFX_INT_MODE_MSI)
1461 efx->interrupt_mode = EFX_INT_MODE_MSI;
1464 } else if (rc < n_channels) {
1465 netif_err(efx, drv, efx->net_dev,
1466 "WARNING: Insufficient MSI-X vectors"
1467 " available (%d < %u).\n", rc, n_channels);
1468 netif_err(efx, drv, efx->net_dev,
1469 "WARNING: Performance may be reduced.\n");
1474 efx->n_channels = n_channels;
1475 if (n_channels > extra_channels)
1476 n_channels -= extra_channels;
1477 if (efx_separate_tx_channels) {
1478 efx->n_tx_channels = min(max(n_channels / 2,
1480 efx->max_tx_channels);
1481 efx->n_rx_channels = max(n_channels -
1485 efx->n_tx_channels = min(n_channels,
1486 efx->max_tx_channels);
1487 efx->n_rx_channels = n_channels;
1489 for (i = 0; i < efx->n_channels; i++)
1490 efx_get_channel(efx, i)->irq =
1495 /* Try single interrupt MSI */
1496 if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1497 efx->n_channels = 1;
1498 efx->n_rx_channels = 1;
1499 efx->n_tx_channels = 1;
1500 rc = pci_enable_msi(efx->pci_dev);
1502 efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1504 netif_err(efx, drv, efx->net_dev,
1505 "could not enable MSI\n");
1506 if (efx->type->min_interrupt_mode >= EFX_INT_MODE_LEGACY)
1507 efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1513 /* Assume legacy interrupts */
1514 if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1515 efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
1516 efx->n_rx_channels = 1;
1517 efx->n_tx_channels = 1;
1518 efx->legacy_irq = efx->pci_dev->irq;
1521 /* Assign extra channels if possible */
1522 efx->n_extra_tx_channels = 0;
1523 j = efx->n_channels;
1524 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1525 if (!efx->extra_channel_type[i])
1527 if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1528 efx->n_channels <= extra_channels) {
1529 efx->extra_channel_type[i]->handle_no_channel(efx);
1532 efx_get_channel(efx, j)->type =
1533 efx->extra_channel_type[i];
1534 if (efx_channel_has_tx_queues(efx_get_channel(efx, j)))
1535 efx->n_extra_tx_channels++;
1539 /* RSS might be usable on VFs even if it is disabled on the PF */
1540 #ifdef CONFIG_SFC_SRIOV
1541 if (efx->type->sriov_wanted) {
1542 efx->rss_spread = ((efx->n_rx_channels > 1 ||
1543 !efx->type->sriov_wanted(efx)) ?
1544 efx->n_rx_channels : efx_vf_size(efx));
1548 efx->rss_spread = efx->n_rx_channels;
1553 #if defined(CONFIG_SMP)
1554 static void efx_set_interrupt_affinity(struct efx_nic *efx)
1556 struct efx_channel *channel;
1559 efx_for_each_channel(channel, efx) {
1560 cpu = cpumask_local_spread(channel->channel,
1561 pcibus_to_node(efx->pci_dev->bus));
1562 irq_set_affinity_hint(channel->irq, cpumask_of(cpu));
1566 static void efx_clear_interrupt_affinity(struct efx_nic *efx)
1568 struct efx_channel *channel;
1570 efx_for_each_channel(channel, efx)
1571 irq_set_affinity_hint(channel->irq, NULL);
1575 efx_set_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
1580 efx_clear_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
1583 #endif /* CONFIG_SMP */
1585 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1587 struct efx_channel *channel, *end_channel;
1590 BUG_ON(efx->state == STATE_DISABLED);
1592 efx->irq_soft_enabled = true;
1595 efx_for_each_channel(channel, efx) {
1596 if (!channel->type->keep_eventq) {
1597 rc = efx_init_eventq(channel);
1601 efx_start_eventq(channel);
1604 efx_mcdi_mode_event(efx);
1608 end_channel = channel;
1609 efx_for_each_channel(channel, efx) {
1610 if (channel == end_channel)
1612 efx_stop_eventq(channel);
1613 if (!channel->type->keep_eventq)
1614 efx_fini_eventq(channel);
1620 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1622 struct efx_channel *channel;
1624 if (efx->state == STATE_DISABLED)
1627 efx_mcdi_mode_poll(efx);
1629 efx->irq_soft_enabled = false;
1632 if (efx->legacy_irq)
1633 synchronize_irq(efx->legacy_irq);
1635 efx_for_each_channel(channel, efx) {
1637 synchronize_irq(channel->irq);
1639 efx_stop_eventq(channel);
1640 if (!channel->type->keep_eventq)
1641 efx_fini_eventq(channel);
1644 /* Flush the asynchronous MCDI request queue */
1645 efx_mcdi_flush_async(efx);
1648 static int efx_enable_interrupts(struct efx_nic *efx)
1650 struct efx_channel *channel, *end_channel;
1653 BUG_ON(efx->state == STATE_DISABLED);
1655 if (efx->eeh_disabled_legacy_irq) {
1656 enable_irq(efx->legacy_irq);
1657 efx->eeh_disabled_legacy_irq = false;
1660 efx->type->irq_enable_master(efx);
1662 efx_for_each_channel(channel, efx) {
1663 if (channel->type->keep_eventq) {
1664 rc = efx_init_eventq(channel);
1670 rc = efx_soft_enable_interrupts(efx);
1677 end_channel = channel;
1678 efx_for_each_channel(channel, efx) {
1679 if (channel == end_channel)
1681 if (channel->type->keep_eventq)
1682 efx_fini_eventq(channel);
1685 efx->type->irq_disable_non_ev(efx);
1690 static void efx_disable_interrupts(struct efx_nic *efx)
1692 struct efx_channel *channel;
1694 efx_soft_disable_interrupts(efx);
1696 efx_for_each_channel(channel, efx) {
1697 if (channel->type->keep_eventq)
1698 efx_fini_eventq(channel);
1701 efx->type->irq_disable_non_ev(efx);
1704 static void efx_remove_interrupts(struct efx_nic *efx)
1706 struct efx_channel *channel;
1708 /* Remove MSI/MSI-X interrupts */
1709 efx_for_each_channel(channel, efx)
1711 pci_disable_msi(efx->pci_dev);
1712 pci_disable_msix(efx->pci_dev);
1714 /* Remove legacy interrupt */
1715 efx->legacy_irq = 0;
1718 static void efx_set_channels(struct efx_nic *efx)
1720 struct efx_channel *channel;
1721 struct efx_tx_queue *tx_queue;
1723 efx->tx_channel_offset =
1724 efx_separate_tx_channels ?
1725 efx->n_channels - efx->n_tx_channels : 0;
1727 /* We need to mark which channels really have RX and TX
1728 * queues, and adjust the TX queue numbers if we have separate
1729 * RX-only and TX-only channels.
1731 efx_for_each_channel(channel, efx) {
1732 if (channel->channel < efx->n_rx_channels)
1733 channel->rx_queue.core_index = channel->channel;
1735 channel->rx_queue.core_index = -1;
1737 efx_for_each_channel_tx_queue(tx_queue, channel)
1738 tx_queue->queue -= (efx->tx_channel_offset *
1743 static int efx_probe_nic(struct efx_nic *efx)
1747 netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1749 /* Carry out hardware-type specific initialisation */
1750 rc = efx->type->probe(efx);
1755 if (!efx->max_channels || !efx->max_tx_channels) {
1756 netif_err(efx, drv, efx->net_dev,
1757 "Insufficient resources to allocate"
1763 /* Determine the number of channels and queues by trying
1764 * to hook in MSI-X interrupts.
1766 rc = efx_probe_interrupts(efx);
1770 efx_set_channels(efx);
1772 /* dimension_resources can fail with EAGAIN */
1773 rc = efx->type->dimension_resources(efx);
1774 if (rc != 0 && rc != -EAGAIN)
1778 /* try again with new max_channels */
1779 efx_remove_interrupts(efx);
1781 } while (rc == -EAGAIN);
1783 if (efx->n_channels > 1)
1784 netdev_rss_key_fill(efx->rss_context.rx_hash_key,
1785 sizeof(efx->rss_context.rx_hash_key));
1786 efx_set_default_rx_indir_table(efx, &efx->rss_context);
1788 netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1789 netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1791 /* Initialise the interrupt moderation settings */
1792 efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
1793 efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1799 efx_remove_interrupts(efx);
1801 efx->type->remove(efx);
1805 static void efx_remove_nic(struct efx_nic *efx)
1807 netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1809 efx_remove_interrupts(efx);
1810 efx->type->remove(efx);
1813 static int efx_probe_filters(struct efx_nic *efx)
1817 init_rwsem(&efx->filter_sem);
1818 mutex_lock(&efx->mac_lock);
1819 down_write(&efx->filter_sem);
1820 rc = efx->type->filter_table_probe(efx);
1824 #ifdef CONFIG_RFS_ACCEL
1825 if (efx->type->offload_features & NETIF_F_NTUPLE) {
1826 struct efx_channel *channel;
1829 efx_for_each_channel(channel, efx) {
1830 channel->rps_flow_id =
1831 kcalloc(efx->type->max_rx_ip_filters,
1832 sizeof(*channel->rps_flow_id),
1834 if (!channel->rps_flow_id)
1838 i < efx->type->max_rx_ip_filters;
1840 channel->rps_flow_id[i] =
1841 RPS_FLOW_ID_INVALID;
1845 efx_for_each_channel(channel, efx)
1846 kfree(channel->rps_flow_id);
1847 efx->type->filter_table_remove(efx);
1852 efx->rps_expire_index = efx->rps_expire_channel = 0;
1856 up_write(&efx->filter_sem);
1857 mutex_unlock(&efx->mac_lock);
1861 static void efx_remove_filters(struct efx_nic *efx)
1863 #ifdef CONFIG_RFS_ACCEL
1864 struct efx_channel *channel;
1866 efx_for_each_channel(channel, efx)
1867 kfree(channel->rps_flow_id);
1869 down_write(&efx->filter_sem);
1870 efx->type->filter_table_remove(efx);
1871 up_write(&efx->filter_sem);
1874 static void efx_restore_filters(struct efx_nic *efx)
1876 down_read(&efx->filter_sem);
1877 efx->type->filter_table_restore(efx);
1878 up_read(&efx->filter_sem);
1881 /**************************************************************************
1883 * NIC startup/shutdown
1885 *************************************************************************/
1887 static int efx_probe_all(struct efx_nic *efx)
1891 rc = efx_probe_nic(efx);
1893 netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1897 rc = efx_probe_port(efx);
1899 netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1903 BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1904 if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1908 efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1910 #ifdef CONFIG_SFC_SRIOV
1911 rc = efx->type->vswitching_probe(efx);
1912 if (rc) /* not fatal; the PF will still work fine */
1913 netif_warn(efx, probe, efx->net_dev,
1914 "failed to setup vswitching rc=%d;"
1915 " VFs may not function\n", rc);
1918 rc = efx_probe_filters(efx);
1920 netif_err(efx, probe, efx->net_dev,
1921 "failed to create filter tables\n");
1925 rc = efx_probe_channels(efx);
1932 efx_remove_filters(efx);
1934 #ifdef CONFIG_SFC_SRIOV
1935 efx->type->vswitching_remove(efx);
1938 efx_remove_port(efx);
1940 efx_remove_nic(efx);
1945 /* If the interface is supposed to be running but is not, start
1946 * the hardware and software data path, regular activity for the port
1947 * (MAC statistics, link polling, etc.) and schedule the port to be
1948 * reconfigured. Interrupts must already be enabled. This function
1949 * is safe to call multiple times, so long as the NIC is not disabled.
1950 * Requires the RTNL lock.
1952 static void efx_start_all(struct efx_nic *efx)
1954 EFX_ASSERT_RESET_SERIALISED(efx);
1955 BUG_ON(efx->state == STATE_DISABLED);
1957 /* Check that it is appropriate to restart the interface. All
1958 * of these flags are safe to read under just the rtnl lock */
1959 if (efx->port_enabled || !netif_running(efx->net_dev) ||
1963 efx_start_port(efx);
1964 efx_start_datapath(efx);
1966 /* Start the hardware monitor if there is one */
1967 if (efx->type->monitor != NULL)
1968 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1969 efx_monitor_interval);
1971 /* Link state detection is normally event-driven; we have
1972 * to poll now because we could have missed a change
1974 mutex_lock(&efx->mac_lock);
1975 if (efx->phy_op->poll(efx))
1976 efx_link_status_changed(efx);
1977 mutex_unlock(&efx->mac_lock);
1979 efx->type->start_stats(efx);
1980 efx->type->pull_stats(efx);
1981 spin_lock_bh(&efx->stats_lock);
1982 efx->type->update_stats(efx, NULL, NULL);
1983 spin_unlock_bh(&efx->stats_lock);
1986 /* Quiesce the hardware and software data path, and regular activity
1987 * for the port without bringing the link down. Safe to call multiple
1988 * times with the NIC in almost any state, but interrupts should be
1989 * enabled. Requires the RTNL lock.
1991 static void efx_stop_all(struct efx_nic *efx)
1993 EFX_ASSERT_RESET_SERIALISED(efx);
1995 /* port_enabled can be read safely under the rtnl lock */
1996 if (!efx->port_enabled)
1999 /* update stats before we go down so we can accurately count
2002 efx->type->pull_stats(efx);
2003 spin_lock_bh(&efx->stats_lock);
2004 efx->type->update_stats(efx, NULL, NULL);
2005 spin_unlock_bh(&efx->stats_lock);
2006 efx->type->stop_stats(efx);
2009 /* Stop the kernel transmit interface. This is only valid if
2010 * the device is stopped or detached; otherwise the watchdog
2011 * may fire immediately.
2013 WARN_ON(netif_running(efx->net_dev) &&
2014 netif_device_present(efx->net_dev));
2015 netif_tx_disable(efx->net_dev);
2017 efx_stop_datapath(efx);
2020 static void efx_remove_all(struct efx_nic *efx)
2022 efx_remove_channels(efx);
2023 efx_remove_filters(efx);
2024 #ifdef CONFIG_SFC_SRIOV
2025 efx->type->vswitching_remove(efx);
2027 efx_remove_port(efx);
2028 efx_remove_nic(efx);
2031 /**************************************************************************
2033 * Interrupt moderation
2035 **************************************************************************/
2036 unsigned int efx_usecs_to_ticks(struct efx_nic *efx, unsigned int usecs)
2040 if (usecs * 1000 < efx->timer_quantum_ns)
2041 return 1; /* never round down to 0 */
2042 return usecs * 1000 / efx->timer_quantum_ns;
2045 unsigned int efx_ticks_to_usecs(struct efx_nic *efx, unsigned int ticks)
2047 /* We must round up when converting ticks to microseconds
2048 * because we round down when converting the other way.
2050 return DIV_ROUND_UP(ticks * efx->timer_quantum_ns, 1000);
2053 /* Set interrupt moderation parameters */
2054 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
2055 unsigned int rx_usecs, bool rx_adaptive,
2056 bool rx_may_override_tx)
2058 struct efx_channel *channel;
2059 unsigned int timer_max_us;
2061 EFX_ASSERT_RESET_SERIALISED(efx);
2063 timer_max_us = efx->timer_max_ns / 1000;
2065 if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
2068 if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
2069 !rx_may_override_tx) {
2070 netif_err(efx, drv, efx->net_dev, "Channels are shared. "
2071 "RX and TX IRQ moderation must be equal\n");
2075 efx->irq_rx_adaptive = rx_adaptive;
2076 efx->irq_rx_moderation_us = rx_usecs;
2077 efx_for_each_channel(channel, efx) {
2078 if (efx_channel_has_rx_queue(channel))
2079 channel->irq_moderation_us = rx_usecs;
2080 else if (efx_channel_has_tx_queues(channel))
2081 channel->irq_moderation_us = tx_usecs;
2087 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
2088 unsigned int *rx_usecs, bool *rx_adaptive)
2090 *rx_adaptive = efx->irq_rx_adaptive;
2091 *rx_usecs = efx->irq_rx_moderation_us;
2093 /* If channels are shared between RX and TX, so is IRQ
2094 * moderation. Otherwise, IRQ moderation is the same for all
2095 * TX channels and is not adaptive.
2097 if (efx->tx_channel_offset == 0) {
2098 *tx_usecs = *rx_usecs;
2100 struct efx_channel *tx_channel;
2102 tx_channel = efx->channel[efx->tx_channel_offset];
2103 *tx_usecs = tx_channel->irq_moderation_us;
2107 /**************************************************************************
2111 **************************************************************************/
2113 /* Run periodically off the general workqueue */
2114 static void efx_monitor(struct work_struct *data)
2116 struct efx_nic *efx = container_of(data, struct efx_nic,
2119 netif_vdbg(efx, timer, efx->net_dev,
2120 "hardware monitor executing on CPU %d\n",
2121 raw_smp_processor_id());
2122 BUG_ON(efx->type->monitor == NULL);
2124 /* If the mac_lock is already held then it is likely a port
2125 * reconfiguration is already in place, which will likely do
2126 * most of the work of monitor() anyway. */
2127 if (mutex_trylock(&efx->mac_lock)) {
2128 if (efx->port_enabled)
2129 efx->type->monitor(efx);
2130 mutex_unlock(&efx->mac_lock);
2133 queue_delayed_work(efx->workqueue, &efx->monitor_work,
2134 efx_monitor_interval);
2137 /**************************************************************************
2141 *************************************************************************/
2144 * Context: process, rtnl_lock() held.
2146 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
2148 struct efx_nic *efx = netdev_priv(net_dev);
2149 struct mii_ioctl_data *data = if_mii(ifr);
2151 if (cmd == SIOCSHWTSTAMP)
2152 return efx_ptp_set_ts_config(efx, ifr);
2153 if (cmd == SIOCGHWTSTAMP)
2154 return efx_ptp_get_ts_config(efx, ifr);
2156 /* Convert phy_id from older PRTAD/DEVAD format */
2157 if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2158 (data->phy_id & 0xfc00) == 0x0400)
2159 data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2161 return mdio_mii_ioctl(&efx->mdio, data, cmd);
2164 /**************************************************************************
2168 **************************************************************************/
2170 static void efx_init_napi_channel(struct efx_channel *channel)
2172 struct efx_nic *efx = channel->efx;
2174 channel->napi_dev = efx->net_dev;
2175 netif_napi_add(channel->napi_dev, &channel->napi_str,
2176 efx_poll, napi_weight);
2179 static void efx_init_napi(struct efx_nic *efx)
2181 struct efx_channel *channel;
2183 efx_for_each_channel(channel, efx)
2184 efx_init_napi_channel(channel);
2187 static void efx_fini_napi_channel(struct efx_channel *channel)
2189 if (channel->napi_dev)
2190 netif_napi_del(&channel->napi_str);
2192 channel->napi_dev = NULL;
2195 static void efx_fini_napi(struct efx_nic *efx)
2197 struct efx_channel *channel;
2199 efx_for_each_channel(channel, efx)
2200 efx_fini_napi_channel(channel);
2203 /**************************************************************************
2205 * Kernel netpoll interface
2207 *************************************************************************/
2209 #ifdef CONFIG_NET_POLL_CONTROLLER
2211 /* Although in the common case interrupts will be disabled, this is not
2212 * guaranteed. However, all our work happens inside the NAPI callback,
2213 * so no locking is required.
2215 static void efx_netpoll(struct net_device *net_dev)
2217 struct efx_nic *efx = netdev_priv(net_dev);
2218 struct efx_channel *channel;
2220 efx_for_each_channel(channel, efx)
2221 efx_schedule_channel(channel);
2226 /**************************************************************************
2228 * Kernel net device interface
2230 *************************************************************************/
2232 /* Context: process, rtnl_lock() held. */
2233 int efx_net_open(struct net_device *net_dev)
2235 struct efx_nic *efx = netdev_priv(net_dev);
2238 netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2239 raw_smp_processor_id());
2241 rc = efx_check_disabled(efx);
2244 if (efx->phy_mode & PHY_MODE_SPECIAL)
2246 if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2249 /* Notify the kernel of the link state polled during driver load,
2250 * before the monitor starts running */
2251 efx_link_status_changed(efx);
2254 if (efx->state == STATE_DISABLED || efx->reset_pending)
2255 netif_device_detach(efx->net_dev);
2256 efx_selftest_async_start(efx);
2260 /* Context: process, rtnl_lock() held.
2261 * Note that the kernel will ignore our return code; this method
2262 * should really be a void.
2264 int efx_net_stop(struct net_device *net_dev)
2266 struct efx_nic *efx = netdev_priv(net_dev);
2268 netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2269 raw_smp_processor_id());
2271 /* Stop the device and flush all the channels */
2277 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2278 static void efx_net_stats(struct net_device *net_dev,
2279 struct rtnl_link_stats64 *stats)
2281 struct efx_nic *efx = netdev_priv(net_dev);
2283 spin_lock_bh(&efx->stats_lock);
2284 efx->type->update_stats(efx, NULL, stats);
2285 spin_unlock_bh(&efx->stats_lock);
2288 /* Context: netif_tx_lock held, BHs disabled. */
2289 static void efx_watchdog(struct net_device *net_dev)
2291 struct efx_nic *efx = netdev_priv(net_dev);
2293 netif_err(efx, tx_err, efx->net_dev,
2294 "TX stuck with port_enabled=%d: resetting channels\n",
2297 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2301 /* Context: process, rtnl_lock() held. */
2302 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2304 struct efx_nic *efx = netdev_priv(net_dev);
2307 rc = efx_check_disabled(efx);
2311 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2313 efx_device_detach_sync(efx);
2316 mutex_lock(&efx->mac_lock);
2317 net_dev->mtu = new_mtu;
2318 efx_mac_reconfigure(efx);
2319 mutex_unlock(&efx->mac_lock);
2322 efx_device_attach_if_not_resetting(efx);
2326 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2328 struct efx_nic *efx = netdev_priv(net_dev);
2329 struct sockaddr *addr = data;
2330 u8 *new_addr = addr->sa_data;
2334 if (!is_valid_ether_addr(new_addr)) {
2335 netif_err(efx, drv, efx->net_dev,
2336 "invalid ethernet MAC address requested: %pM\n",
2338 return -EADDRNOTAVAIL;
2341 /* save old address */
2342 ether_addr_copy(old_addr, net_dev->dev_addr);
2343 ether_addr_copy(net_dev->dev_addr, new_addr);
2344 if (efx->type->set_mac_address) {
2345 rc = efx->type->set_mac_address(efx);
2347 ether_addr_copy(net_dev->dev_addr, old_addr);
2352 /* Reconfigure the MAC */
2353 mutex_lock(&efx->mac_lock);
2354 efx_mac_reconfigure(efx);
2355 mutex_unlock(&efx->mac_lock);
2360 /* Context: netif_addr_lock held, BHs disabled. */
2361 static void efx_set_rx_mode(struct net_device *net_dev)
2363 struct efx_nic *efx = netdev_priv(net_dev);
2365 if (efx->port_enabled)
2366 queue_work(efx->workqueue, &efx->mac_work);
2367 /* Otherwise efx_start_port() will do this */
2370 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2372 struct efx_nic *efx = netdev_priv(net_dev);
2375 /* If disabling RX n-tuple filtering, clear existing filters */
2376 if (net_dev->features & ~data & NETIF_F_NTUPLE) {
2377 rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2382 /* If Rx VLAN filter is changed, update filters via mac_reconfigure.
2383 * If rx-fcs is changed, mac_reconfigure updates that too.
2385 if ((net_dev->features ^ data) & (NETIF_F_HW_VLAN_CTAG_FILTER |
2387 /* efx_set_rx_mode() will schedule MAC work to update filters
2388 * when a new features are finally set in net_dev.
2390 efx_set_rx_mode(net_dev);
2396 static int efx_get_phys_port_id(struct net_device *net_dev,
2397 struct netdev_phys_item_id *ppid)
2399 struct efx_nic *efx = netdev_priv(net_dev);
2401 if (efx->type->get_phys_port_id)
2402 return efx->type->get_phys_port_id(efx, ppid);
2407 static int efx_get_phys_port_name(struct net_device *net_dev,
2408 char *name, size_t len)
2410 struct efx_nic *efx = netdev_priv(net_dev);
2412 if (snprintf(name, len, "p%u", efx->port_num) >= len)
2417 static int efx_vlan_rx_add_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2419 struct efx_nic *efx = netdev_priv(net_dev);
2421 if (efx->type->vlan_rx_add_vid)
2422 return efx->type->vlan_rx_add_vid(efx, proto, vid);
2427 static int efx_vlan_rx_kill_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2429 struct efx_nic *efx = netdev_priv(net_dev);
2431 if (efx->type->vlan_rx_kill_vid)
2432 return efx->type->vlan_rx_kill_vid(efx, proto, vid);
2437 static int efx_udp_tunnel_type_map(enum udp_parsable_tunnel_type in)
2440 case UDP_TUNNEL_TYPE_VXLAN:
2441 return TUNNEL_ENCAP_UDP_PORT_ENTRY_VXLAN;
2442 case UDP_TUNNEL_TYPE_GENEVE:
2443 return TUNNEL_ENCAP_UDP_PORT_ENTRY_GENEVE;
2449 static void efx_udp_tunnel_add(struct net_device *dev, struct udp_tunnel_info *ti)
2451 struct efx_nic *efx = netdev_priv(dev);
2452 struct efx_udp_tunnel tnl;
2453 int efx_tunnel_type;
2455 efx_tunnel_type = efx_udp_tunnel_type_map(ti->type);
2456 if (efx_tunnel_type < 0)
2459 tnl.type = (u16)efx_tunnel_type;
2460 tnl.port = ti->port;
2462 if (efx->type->udp_tnl_add_port)
2463 (void)efx->type->udp_tnl_add_port(efx, tnl);
2466 static void efx_udp_tunnel_del(struct net_device *dev, struct udp_tunnel_info *ti)
2468 struct efx_nic *efx = netdev_priv(dev);
2469 struct efx_udp_tunnel tnl;
2470 int efx_tunnel_type;
2472 efx_tunnel_type = efx_udp_tunnel_type_map(ti->type);
2473 if (efx_tunnel_type < 0)
2476 tnl.type = (u16)efx_tunnel_type;
2477 tnl.port = ti->port;
2479 if (efx->type->udp_tnl_del_port)
2480 (void)efx->type->udp_tnl_del_port(efx, tnl);
2483 static const struct net_device_ops efx_netdev_ops = {
2484 .ndo_open = efx_net_open,
2485 .ndo_stop = efx_net_stop,
2486 .ndo_get_stats64 = efx_net_stats,
2487 .ndo_tx_timeout = efx_watchdog,
2488 .ndo_start_xmit = efx_hard_start_xmit,
2489 .ndo_validate_addr = eth_validate_addr,
2490 .ndo_do_ioctl = efx_ioctl,
2491 .ndo_change_mtu = efx_change_mtu,
2492 .ndo_set_mac_address = efx_set_mac_address,
2493 .ndo_set_rx_mode = efx_set_rx_mode,
2494 .ndo_set_features = efx_set_features,
2495 .ndo_vlan_rx_add_vid = efx_vlan_rx_add_vid,
2496 .ndo_vlan_rx_kill_vid = efx_vlan_rx_kill_vid,
2497 #ifdef CONFIG_SFC_SRIOV
2498 .ndo_set_vf_mac = efx_sriov_set_vf_mac,
2499 .ndo_set_vf_vlan = efx_sriov_set_vf_vlan,
2500 .ndo_set_vf_spoofchk = efx_sriov_set_vf_spoofchk,
2501 .ndo_get_vf_config = efx_sriov_get_vf_config,
2502 .ndo_set_vf_link_state = efx_sriov_set_vf_link_state,
2504 .ndo_get_phys_port_id = efx_get_phys_port_id,
2505 .ndo_get_phys_port_name = efx_get_phys_port_name,
2506 #ifdef CONFIG_NET_POLL_CONTROLLER
2507 .ndo_poll_controller = efx_netpoll,
2509 .ndo_setup_tc = efx_setup_tc,
2510 #ifdef CONFIG_RFS_ACCEL
2511 .ndo_rx_flow_steer = efx_filter_rfs,
2513 .ndo_udp_tunnel_add = efx_udp_tunnel_add,
2514 .ndo_udp_tunnel_del = efx_udp_tunnel_del,
2517 static void efx_update_name(struct efx_nic *efx)
2519 strcpy(efx->name, efx->net_dev->name);
2520 efx_mtd_rename(efx);
2521 efx_set_channel_names(efx);
2524 static int efx_netdev_event(struct notifier_block *this,
2525 unsigned long event, void *ptr)
2527 struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2529 if ((net_dev->netdev_ops == &efx_netdev_ops) &&
2530 event == NETDEV_CHANGENAME)
2531 efx_update_name(netdev_priv(net_dev));
2536 static struct notifier_block efx_netdev_notifier = {
2537 .notifier_call = efx_netdev_event,
2541 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2543 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2544 return sprintf(buf, "%d\n", efx->phy_type);
2546 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2548 #ifdef CONFIG_SFC_MCDI_LOGGING
2549 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
2552 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2553 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2555 return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
2557 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
2558 const char *buf, size_t count)
2560 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2561 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2562 bool enable = count > 0 && *buf != '0';
2564 mcdi->logging_enabled = enable;
2567 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
2570 static int efx_register_netdev(struct efx_nic *efx)
2572 struct net_device *net_dev = efx->net_dev;
2573 struct efx_channel *channel;
2576 net_dev->watchdog_timeo = 5 * HZ;
2577 net_dev->irq = efx->pci_dev->irq;
2578 net_dev->netdev_ops = &efx_netdev_ops;
2579 if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
2580 net_dev->priv_flags |= IFF_UNICAST_FLT;
2581 net_dev->ethtool_ops = &efx_ethtool_ops;
2582 net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2583 net_dev->min_mtu = EFX_MIN_MTU;
2584 net_dev->max_mtu = EFX_MAX_MTU;
2588 /* Enable resets to be scheduled and check whether any were
2589 * already requested. If so, the NIC is probably hosed so we
2592 efx->state = STATE_READY;
2593 smp_mb(); /* ensure we change state before checking reset_pending */
2594 if (efx->reset_pending) {
2595 netif_err(efx, probe, efx->net_dev,
2596 "aborting probe due to scheduled reset\n");
2601 rc = dev_alloc_name(net_dev, net_dev->name);
2604 efx_update_name(efx);
2606 /* Always start with carrier off; PHY events will detect the link */
2607 netif_carrier_off(net_dev);
2609 rc = register_netdevice(net_dev);
2613 efx_for_each_channel(channel, efx) {
2614 struct efx_tx_queue *tx_queue;
2615 efx_for_each_channel_tx_queue(tx_queue, channel)
2616 efx_init_tx_queue_core_txq(tx_queue);
2623 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2625 netif_err(efx, drv, efx->net_dev,
2626 "failed to init net dev attributes\n");
2627 goto fail_registered;
2629 #ifdef CONFIG_SFC_MCDI_LOGGING
2630 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2632 netif_err(efx, drv, efx->net_dev,
2633 "failed to init net dev attributes\n");
2634 goto fail_attr_mcdi_logging;
2640 #ifdef CONFIG_SFC_MCDI_LOGGING
2641 fail_attr_mcdi_logging:
2642 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2646 efx_dissociate(efx);
2647 unregister_netdevice(net_dev);
2649 efx->state = STATE_UNINIT;
2651 netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2655 static void efx_unregister_netdev(struct efx_nic *efx)
2660 BUG_ON(netdev_priv(efx->net_dev) != efx);
2662 if (efx_dev_registered(efx)) {
2663 strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2664 #ifdef CONFIG_SFC_MCDI_LOGGING
2665 device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2667 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2668 unregister_netdev(efx->net_dev);
2672 /**************************************************************************
2674 * Device reset and suspend
2676 **************************************************************************/
2678 /* Tears down the entire software state and most of the hardware state
2680 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2682 EFX_ASSERT_RESET_SERIALISED(efx);
2684 if (method == RESET_TYPE_MCDI_TIMEOUT)
2685 efx->type->prepare_flr(efx);
2688 efx_disable_interrupts(efx);
2690 mutex_lock(&efx->mac_lock);
2691 mutex_lock(&efx->rss_lock);
2692 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2693 method != RESET_TYPE_DATAPATH)
2694 efx->phy_op->fini(efx);
2695 efx->type->fini(efx);
2698 /* This function will always ensure that the locks acquired in
2699 * efx_reset_down() are released. A failure return code indicates
2700 * that we were unable to reinitialise the hardware, and the
2701 * driver should be disabled. If ok is false, then the rx and tx
2702 * engines are not restarted, pending a RESET_DISABLE. */
2703 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2707 EFX_ASSERT_RESET_SERIALISED(efx);
2709 if (method == RESET_TYPE_MCDI_TIMEOUT)
2710 efx->type->finish_flr(efx);
2712 /* Ensure that SRAM is initialised even if we're disabling the device */
2713 rc = efx->type->init(efx);
2715 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2722 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2723 method != RESET_TYPE_DATAPATH) {
2724 rc = efx->phy_op->init(efx);
2727 rc = efx->phy_op->reconfigure(efx);
2728 if (rc && rc != -EPERM)
2729 netif_err(efx, drv, efx->net_dev,
2730 "could not restore PHY settings\n");
2733 rc = efx_enable_interrupts(efx);
2737 #ifdef CONFIG_SFC_SRIOV
2738 rc = efx->type->vswitching_restore(efx);
2739 if (rc) /* not fatal; the PF will still work fine */
2740 netif_warn(efx, probe, efx->net_dev,
2741 "failed to restore vswitching rc=%d;"
2742 " VFs may not function\n", rc);
2745 if (efx->type->rx_restore_rss_contexts)
2746 efx->type->rx_restore_rss_contexts(efx);
2747 mutex_unlock(&efx->rss_lock);
2748 down_read(&efx->filter_sem);
2749 efx_restore_filters(efx);
2750 up_read(&efx->filter_sem);
2751 if (efx->type->sriov_reset)
2752 efx->type->sriov_reset(efx);
2754 mutex_unlock(&efx->mac_lock);
2758 if (efx->type->udp_tnl_push_ports)
2759 efx->type->udp_tnl_push_ports(efx);
2764 efx->port_initialized = false;
2766 mutex_unlock(&efx->rss_lock);
2767 mutex_unlock(&efx->mac_lock);
2772 /* Reset the NIC using the specified method. Note that the reset may
2773 * fail, in which case the card will be left in an unusable state.
2775 * Caller must hold the rtnl_lock.
2777 int efx_reset(struct efx_nic *efx, enum reset_type method)
2782 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2783 RESET_TYPE(method));
2785 efx_device_detach_sync(efx);
2786 efx_reset_down(efx, method);
2788 rc = efx->type->reset(efx, method);
2790 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2794 /* Clear flags for the scopes we covered. We assume the NIC and
2795 * driver are now quiescent so that there is no race here.
2797 if (method < RESET_TYPE_MAX_METHOD)
2798 efx->reset_pending &= -(1 << (method + 1));
2799 else /* it doesn't fit into the well-ordered scope hierarchy */
2800 __clear_bit(method, &efx->reset_pending);
2802 /* Reinitialise bus-mastering, which may have been turned off before
2803 * the reset was scheduled. This is still appropriate, even in the
2804 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2805 * can respond to requests. */
2806 pci_set_master(efx->pci_dev);
2809 /* Leave device stopped if necessary */
2811 method == RESET_TYPE_DISABLE ||
2812 method == RESET_TYPE_RECOVER_OR_DISABLE;
2813 rc2 = efx_reset_up(efx, method, !disabled);
2821 dev_close(efx->net_dev);
2822 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2823 efx->state = STATE_DISABLED;
2825 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2826 efx_device_attach_if_not_resetting(efx);
2831 /* Try recovery mechanisms.
2832 * For now only EEH is supported.
2833 * Returns 0 if the recovery mechanisms are unsuccessful.
2834 * Returns a non-zero value otherwise.
2836 int efx_try_recovery(struct efx_nic *efx)
2839 /* A PCI error can occur and not be seen by EEH because nothing
2840 * happens on the PCI bus. In this case the driver may fail and
2841 * schedule a 'recover or reset', leading to this recovery handler.
2842 * Manually call the eeh failure check function.
2844 struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2845 if (eeh_dev_check_failure(eehdev)) {
2846 /* The EEH mechanisms will handle the error and reset the
2847 * device if necessary.
2855 static void efx_wait_for_bist_end(struct efx_nic *efx)
2859 for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2860 if (efx_mcdi_poll_reboot(efx))
2862 msleep(BIST_WAIT_DELAY_MS);
2865 netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2867 /* Either way unset the BIST flag. If we found no reboot we probably
2868 * won't recover, but we should try.
2870 efx->mc_bist_for_other_fn = false;
2873 /* The worker thread exists so that code that cannot sleep can
2874 * schedule a reset for later.
2876 static void efx_reset_work(struct work_struct *data)
2878 struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2879 unsigned long pending;
2880 enum reset_type method;
2882 pending = READ_ONCE(efx->reset_pending);
2883 method = fls(pending) - 1;
2885 if (method == RESET_TYPE_MC_BIST)
2886 efx_wait_for_bist_end(efx);
2888 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2889 method == RESET_TYPE_RECOVER_OR_ALL) &&
2890 efx_try_recovery(efx))
2898 /* We checked the state in efx_schedule_reset() but it may
2899 * have changed by now. Now that we have the RTNL lock,
2900 * it cannot change again.
2902 if (efx->state == STATE_READY)
2903 (void)efx_reset(efx, method);
2908 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2910 enum reset_type method;
2912 if (efx->state == STATE_RECOVERY) {
2913 netif_dbg(efx, drv, efx->net_dev,
2914 "recovering: skip scheduling %s reset\n",
2920 case RESET_TYPE_INVISIBLE:
2921 case RESET_TYPE_ALL:
2922 case RESET_TYPE_RECOVER_OR_ALL:
2923 case RESET_TYPE_WORLD:
2924 case RESET_TYPE_DISABLE:
2925 case RESET_TYPE_RECOVER_OR_DISABLE:
2926 case RESET_TYPE_DATAPATH:
2927 case RESET_TYPE_MC_BIST:
2928 case RESET_TYPE_MCDI_TIMEOUT:
2930 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2931 RESET_TYPE(method));
2934 method = efx->type->map_reset_reason(type);
2935 netif_dbg(efx, drv, efx->net_dev,
2936 "scheduling %s reset for %s\n",
2937 RESET_TYPE(method), RESET_TYPE(type));
2941 set_bit(method, &efx->reset_pending);
2942 smp_mb(); /* ensure we change reset_pending before checking state */
2944 /* If we're not READY then just leave the flags set as the cue
2945 * to abort probing or reschedule the reset later.
2947 if (READ_ONCE(efx->state) != STATE_READY)
2950 /* efx_process_channel() will no longer read events once a
2951 * reset is scheduled. So switch back to poll'd MCDI completions. */
2952 efx_mcdi_mode_poll(efx);
2954 queue_work(reset_workqueue, &efx->reset_work);
2957 /**************************************************************************
2959 * List of NICs we support
2961 **************************************************************************/
2963 /* PCI device ID table */
2964 static const struct pci_device_id efx_pci_table[] = {
2965 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803), /* SFC9020 */
2966 .driver_data = (unsigned long) &siena_a0_nic_type},
2967 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813), /* SFL9021 */
2968 .driver_data = (unsigned long) &siena_a0_nic_type},
2969 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903), /* SFC9120 PF */
2970 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2971 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903), /* SFC9120 VF */
2972 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2973 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923), /* SFC9140 PF */
2974 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2975 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1923), /* SFC9140 VF */
2976 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2977 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0a03), /* SFC9220 PF */
2978 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2979 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1a03), /* SFC9220 VF */
2980 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2981 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0b03), /* SFC9250 PF */
2982 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2983 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1b03), /* SFC9250 VF */
2984 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2985 {0} /* end of list */
2988 /**************************************************************************
2990 * Dummy PHY/MAC operations
2992 * Can be used for some unimplemented operations
2993 * Needed so all function pointers are valid and do not have to be tested
2996 **************************************************************************/
2997 int efx_port_dummy_op_int(struct efx_nic *efx)
3001 void efx_port_dummy_op_void(struct efx_nic *efx) {}
3003 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
3008 static const struct efx_phy_operations efx_dummy_phy_operations = {
3009 .init = efx_port_dummy_op_int,
3010 .reconfigure = efx_port_dummy_op_int,
3011 .poll = efx_port_dummy_op_poll,
3012 .fini = efx_port_dummy_op_void,
3015 /**************************************************************************
3019 **************************************************************************/
3021 /* This zeroes out and then fills in the invariants in a struct
3022 * efx_nic (including all sub-structures).
3024 static int efx_init_struct(struct efx_nic *efx,
3025 struct pci_dev *pci_dev, struct net_device *net_dev)
3027 int rc = -ENOMEM, i;
3029 /* Initialise common structures */
3030 INIT_LIST_HEAD(&efx->node);
3031 INIT_LIST_HEAD(&efx->secondary_list);
3032 spin_lock_init(&efx->biu_lock);
3033 #ifdef CONFIG_SFC_MTD
3034 INIT_LIST_HEAD(&efx->mtd_list);
3036 INIT_WORK(&efx->reset_work, efx_reset_work);
3037 INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
3038 INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
3039 efx->pci_dev = pci_dev;
3040 efx->msg_enable = debug;
3041 efx->state = STATE_UNINIT;
3042 strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
3044 efx->net_dev = net_dev;
3045 efx->rx_prefix_size = efx->type->rx_prefix_size;
3047 NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
3048 efx->rx_packet_hash_offset =
3049 efx->type->rx_hash_offset - efx->type->rx_prefix_size;
3050 efx->rx_packet_ts_offset =
3051 efx->type->rx_ts_offset - efx->type->rx_prefix_size;
3052 INIT_LIST_HEAD(&efx->rss_context.list);
3053 mutex_init(&efx->rss_lock);
3054 spin_lock_init(&efx->stats_lock);
3055 efx->vi_stride = EFX_DEFAULT_VI_STRIDE;
3056 efx->num_mac_stats = MC_CMD_MAC_NSTATS;
3057 BUILD_BUG_ON(MC_CMD_MAC_NSTATS - 1 != MC_CMD_MAC_GENERATION_END);
3058 mutex_init(&efx->mac_lock);
3059 #ifdef CONFIG_RFS_ACCEL
3060 mutex_init(&efx->rps_mutex);
3061 spin_lock_init(&efx->rps_hash_lock);
3062 /* Failure to allocate is not fatal, but may degrade ARFS performance */
3063 efx->rps_hash_table = kcalloc(EFX_ARFS_HASH_TABLE_SIZE,
3064 sizeof(*efx->rps_hash_table), GFP_KERNEL);
3066 efx->phy_op = &efx_dummy_phy_operations;
3067 efx->mdio.dev = net_dev;
3068 INIT_WORK(&efx->mac_work, efx_mac_work);
3069 init_waitqueue_head(&efx->flush_wq);
3071 for (i = 0; i < EFX_MAX_CHANNELS; i++) {
3072 efx->channel[i] = efx_alloc_channel(efx, i, NULL);
3073 if (!efx->channel[i])
3075 efx->msi_context[i].efx = efx;
3076 efx->msi_context[i].index = i;
3079 /* Higher numbered interrupt modes are less capable! */
3080 if (WARN_ON_ONCE(efx->type->max_interrupt_mode >
3081 efx->type->min_interrupt_mode)) {
3085 efx->interrupt_mode = max(efx->type->max_interrupt_mode,
3087 efx->interrupt_mode = min(efx->type->min_interrupt_mode,
3090 /* Would be good to use the net_dev name, but we're too early */
3091 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
3093 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
3094 if (!efx->workqueue)
3100 efx_fini_struct(efx);
3104 static void efx_fini_struct(struct efx_nic *efx)
3108 #ifdef CONFIG_RFS_ACCEL
3109 kfree(efx->rps_hash_table);
3112 for (i = 0; i < EFX_MAX_CHANNELS; i++)
3113 kfree(efx->channel[i]);
3117 if (efx->workqueue) {
3118 destroy_workqueue(efx->workqueue);
3119 efx->workqueue = NULL;
3123 void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
3125 u64 n_rx_nodesc_trunc = 0;
3126 struct efx_channel *channel;
3128 efx_for_each_channel(channel, efx)
3129 n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
3130 stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
3131 stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
3134 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
3135 const struct efx_filter_spec *right)
3137 if ((left->match_flags ^ right->match_flags) |
3138 ((left->flags ^ right->flags) &
3139 (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
3142 return memcmp(&left->outer_vid, &right->outer_vid,
3143 sizeof(struct efx_filter_spec) -
3144 offsetof(struct efx_filter_spec, outer_vid)) == 0;
3147 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
3149 BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
3150 return jhash2((const u32 *)&spec->outer_vid,
3151 (sizeof(struct efx_filter_spec) -
3152 offsetof(struct efx_filter_spec, outer_vid)) / 4,
3156 #ifdef CONFIG_RFS_ACCEL
3157 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
3160 if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
3161 /* ARFS is currently updating this entry, leave it */
3164 if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
3165 /* ARFS tried and failed to update this, so it's probably out
3166 * of date. Remove the filter and the ARFS rule entry.
3168 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
3171 } else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
3172 /* ARFS has moved on, so old filter is not needed. Since we did
3173 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
3174 * not be removed by efx_rps_hash_del() subsequently.
3179 /* Remove it iff ARFS wants to. */
3183 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
3184 const struct efx_filter_spec *spec)
3186 u32 hash = efx_filter_spec_hash(spec);
3188 WARN_ON(!spin_is_locked(&efx->rps_hash_lock));
3189 if (!efx->rps_hash_table)
3191 return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
3194 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
3195 const struct efx_filter_spec *spec)
3197 struct efx_arfs_rule *rule;
3198 struct hlist_head *head;
3199 struct hlist_node *node;
3201 head = efx_rps_hash_bucket(efx, spec);
3204 hlist_for_each(node, head) {
3205 rule = container_of(node, struct efx_arfs_rule, node);
3206 if (efx_filter_spec_equal(spec, &rule->spec))
3212 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
3213 const struct efx_filter_spec *spec,
3216 struct efx_arfs_rule *rule;
3217 struct hlist_head *head;
3218 struct hlist_node *node;
3220 head = efx_rps_hash_bucket(efx, spec);
3223 hlist_for_each(node, head) {
3224 rule = container_of(node, struct efx_arfs_rule, node);
3225 if (efx_filter_spec_equal(spec, &rule->spec)) {
3230 rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
3233 memcpy(&rule->spec, spec, sizeof(rule->spec));
3234 hlist_add_head(&rule->node, head);
3239 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
3241 struct efx_arfs_rule *rule;
3242 struct hlist_head *head;
3243 struct hlist_node *node;
3245 head = efx_rps_hash_bucket(efx, spec);
3248 hlist_for_each(node, head) {
3249 rule = container_of(node, struct efx_arfs_rule, node);
3250 if (efx_filter_spec_equal(spec, &rule->spec)) {
3251 /* Someone already reused the entry. We know that if
3252 * this check doesn't fire (i.e. filter_id == REMOVING)
3253 * then the REMOVING mark was put there by our caller,
3254 * because caller is holding a lock on filter table and
3255 * only holders of that lock set REMOVING.
3257 if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
3264 /* We didn't find it. */
3269 /* RSS contexts. We're using linked lists and crappy O(n) algorithms, because
3270 * (a) this is an infrequent control-plane operation and (b) n is small (max 64)
3272 struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx)
3274 struct list_head *head = &efx->rss_context.list;
3275 struct efx_rss_context *ctx, *new;
3276 u32 id = 1; /* Don't use zero, that refers to the master RSS context */
3278 WARN_ON(!mutex_is_locked(&efx->rss_lock));
3280 /* Search for first gap in the numbering */
3281 list_for_each_entry(ctx, head, list) {
3282 if (ctx->user_id != id)
3285 /* Check for wrap. If this happens, we have nearly 2^32
3286 * allocated RSS contexts, which seems unlikely.
3288 if (WARN_ON_ONCE(!id))
3292 /* Create the new entry */
3293 new = kmalloc(sizeof(struct efx_rss_context), GFP_KERNEL);
3296 new->context_id = EFX_EF10_RSS_CONTEXT_INVALID;
3297 new->rx_hash_udp_4tuple = false;
3299 /* Insert the new entry into the gap */
3301 list_add_tail(&new->list, &ctx->list);
3305 struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id)
3307 struct list_head *head = &efx->rss_context.list;
3308 struct efx_rss_context *ctx;
3310 WARN_ON(!mutex_is_locked(&efx->rss_lock));
3312 list_for_each_entry(ctx, head, list)
3313 if (ctx->user_id == id)
3318 void efx_free_rss_context_entry(struct efx_rss_context *ctx)
3320 list_del(&ctx->list);
3324 /**************************************************************************
3328 **************************************************************************/
3330 /* Main body of final NIC shutdown code
3331 * This is called only at module unload (or hotplug removal).
3333 static void efx_pci_remove_main(struct efx_nic *efx)
3335 /* Flush reset_work. It can no longer be scheduled since we
3338 BUG_ON(efx->state == STATE_READY);
3339 cancel_work_sync(&efx->reset_work);
3341 efx_disable_interrupts(efx);
3342 efx_clear_interrupt_affinity(efx);
3343 efx_nic_fini_interrupt(efx);
3345 efx->type->fini(efx);
3347 efx_remove_all(efx);
3350 /* Final NIC shutdown
3351 * This is called only at module unload (or hotplug removal). A PF can call
3352 * this on its VFs to ensure they are unbound first.
3354 static void efx_pci_remove(struct pci_dev *pci_dev)
3356 struct efx_nic *efx;
3358 efx = pci_get_drvdata(pci_dev);
3362 /* Mark the NIC as fini, then stop the interface */
3364 efx_dissociate(efx);
3365 dev_close(efx->net_dev);
3366 efx_disable_interrupts(efx);
3367 efx->state = STATE_UNINIT;
3370 if (efx->type->sriov_fini)
3371 efx->type->sriov_fini(efx);
3373 efx_unregister_netdev(efx);
3375 efx_mtd_remove(efx);
3377 efx_pci_remove_main(efx);
3380 netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
3382 efx_fini_struct(efx);
3383 free_netdev(efx->net_dev);
3385 pci_disable_pcie_error_reporting(pci_dev);
3388 /* NIC VPD information
3389 * Called during probe to display the part number of the
3390 * installed NIC. VPD is potentially very large but this should
3391 * always appear within the first 512 bytes.
3393 #define SFC_VPD_LEN 512
3394 static void efx_probe_vpd_strings(struct efx_nic *efx)
3396 struct pci_dev *dev = efx->pci_dev;
3397 char vpd_data[SFC_VPD_LEN];
3399 int ro_start, ro_size, i, j;
3401 /* Get the vpd data from the device */
3402 vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
3403 if (vpd_size <= 0) {
3404 netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
3408 /* Get the Read only section */
3409 ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
3411 netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
3415 ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
3417 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3418 if (i + j > vpd_size)
3421 /* Get the Part number */
3422 i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
3424 netif_err(efx, drv, efx->net_dev, "Part number not found\n");
3428 j = pci_vpd_info_field_size(&vpd_data[i]);
3429 i += PCI_VPD_INFO_FLD_HDR_SIZE;
3430 if (i + j > vpd_size) {
3431 netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
3435 netif_info(efx, drv, efx->net_dev,
3436 "Part Number : %.*s\n", j, &vpd_data[i]);
3438 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3440 i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
3442 netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
3446 j = pci_vpd_info_field_size(&vpd_data[i]);
3447 i += PCI_VPD_INFO_FLD_HDR_SIZE;
3448 if (i + j > vpd_size) {
3449 netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
3453 efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
3457 snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
3461 /* Main body of NIC initialisation
3462 * This is called at module load (or hotplug insertion, theoretically).
3464 static int efx_pci_probe_main(struct efx_nic *efx)
3468 /* Do start-of-day initialisation */
3469 rc = efx_probe_all(efx);
3475 rc = efx->type->init(efx);
3477 netif_err(efx, probe, efx->net_dev,
3478 "failed to initialise NIC\n");
3482 rc = efx_init_port(efx);
3484 netif_err(efx, probe, efx->net_dev,
3485 "failed to initialise port\n");
3489 rc = efx_nic_init_interrupt(efx);
3493 efx_set_interrupt_affinity(efx);
3494 rc = efx_enable_interrupts(efx);
3501 efx_clear_interrupt_affinity(efx);
3502 efx_nic_fini_interrupt(efx);
3506 efx->type->fini(efx);
3509 efx_remove_all(efx);
3514 static int efx_pci_probe_post_io(struct efx_nic *efx)
3516 struct net_device *net_dev = efx->net_dev;
3517 int rc = efx_pci_probe_main(efx);
3522 if (efx->type->sriov_init) {
3523 rc = efx->type->sriov_init(efx);
3525 netif_err(efx, probe, efx->net_dev,
3526 "SR-IOV can't be enabled rc %d\n", rc);
3529 /* Determine netdevice features */
3530 net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
3531 NETIF_F_TSO | NETIF_F_RXCSUM | NETIF_F_RXALL);
3532 if (efx->type->offload_features & (NETIF_F_IPV6_CSUM | NETIF_F_HW_CSUM))
3533 net_dev->features |= NETIF_F_TSO6;
3534 /* Check whether device supports TSO */
3535 if (!efx->type->tso_versions || !efx->type->tso_versions(efx))
3536 net_dev->features &= ~NETIF_F_ALL_TSO;
3537 /* Mask for features that also apply to VLAN devices */
3538 net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
3539 NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
3542 net_dev->hw_features |= net_dev->features & ~efx->fixed_features;
3544 /* Disable receiving frames with bad FCS, by default. */
3545 net_dev->features &= ~NETIF_F_RXALL;
3547 /* Disable VLAN filtering by default. It may be enforced if
3548 * the feature is fixed (i.e. VLAN filters are required to
3549 * receive VLAN tagged packets due to vPort restrictions).
3551 net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
3552 net_dev->features |= efx->fixed_features;
3554 rc = efx_register_netdev(efx);
3558 efx_pci_remove_main(efx);
3562 /* NIC initialisation
3564 * This is called at module load (or hotplug insertion,
3565 * theoretically). It sets up PCI mappings, resets the NIC,
3566 * sets up and registers the network devices with the kernel and hooks
3567 * the interrupt service routine. It does not prepare the device for
3568 * transmission; this is left to the first time one of the network
3569 * interfaces is brought up (i.e. efx_net_open).
3571 static int efx_pci_probe(struct pci_dev *pci_dev,
3572 const struct pci_device_id *entry)
3574 struct net_device *net_dev;
3575 struct efx_nic *efx;
3578 /* Allocate and initialise a struct net_device and struct efx_nic */
3579 net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
3583 efx = netdev_priv(net_dev);
3584 efx->type = (const struct efx_nic_type *) entry->driver_data;
3585 efx->fixed_features |= NETIF_F_HIGHDMA;
3587 pci_set_drvdata(pci_dev, efx);
3588 SET_NETDEV_DEV(net_dev, &pci_dev->dev);
3589 rc = efx_init_struct(efx, pci_dev, net_dev);
3593 netif_info(efx, probe, efx->net_dev,
3594 "Solarflare NIC detected\n");
3596 if (!efx->type->is_vf)
3597 efx_probe_vpd_strings(efx);
3599 /* Set up basic I/O (BAR mappings etc) */
3600 rc = efx_init_io(efx);
3604 rc = efx_pci_probe_post_io(efx);
3606 /* On failure, retry once immediately.
3607 * If we aborted probe due to a scheduled reset, dismiss it.
3609 efx->reset_pending = 0;
3610 rc = efx_pci_probe_post_io(efx);
3612 /* On another failure, retry once more
3613 * after a 50-305ms delay.
3617 get_random_bytes(&r, 1);
3618 msleep((unsigned int)r + 50);
3619 efx->reset_pending = 0;
3620 rc = efx_pci_probe_post_io(efx);
3626 netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3628 /* Try to create MTDs, but allow this to fail */
3630 rc = efx_mtd_probe(efx);
3632 if (rc && rc != -EPERM)
3633 netif_warn(efx, probe, efx->net_dev,
3634 "failed to create MTDs (%d)\n", rc);
3636 rc = pci_enable_pcie_error_reporting(pci_dev);
3637 if (rc && rc != -EINVAL)
3638 netif_notice(efx, probe, efx->net_dev,
3639 "PCIE error reporting unavailable (%d).\n",
3642 if (efx->type->udp_tnl_push_ports)
3643 efx->type->udp_tnl_push_ports(efx);
3650 efx_fini_struct(efx);
3653 netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3654 free_netdev(net_dev);
3658 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3659 * enabled on success
3661 #ifdef CONFIG_SFC_SRIOV
3662 static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
3665 struct efx_nic *efx = pci_get_drvdata(dev);
3667 if (efx->type->sriov_configure) {
3668 rc = efx->type->sriov_configure(efx, num_vfs);
3678 static int efx_pm_freeze(struct device *dev)
3680 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3684 if (efx->state != STATE_DISABLED) {
3685 efx->state = STATE_UNINIT;
3687 efx_device_detach_sync(efx);
3690 efx_disable_interrupts(efx);
3698 static int efx_pm_thaw(struct device *dev)
3701 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3705 if (efx->state != STATE_DISABLED) {
3706 rc = efx_enable_interrupts(efx);
3710 mutex_lock(&efx->mac_lock);
3711 efx->phy_op->reconfigure(efx);
3712 mutex_unlock(&efx->mac_lock);
3716 efx_device_attach_if_not_resetting(efx);
3718 efx->state = STATE_READY;
3720 efx->type->resume_wol(efx);
3725 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3726 queue_work(reset_workqueue, &efx->reset_work);
3736 static int efx_pm_poweroff(struct device *dev)
3738 struct pci_dev *pci_dev = to_pci_dev(dev);
3739 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3741 efx->type->fini(efx);
3743 efx->reset_pending = 0;
3745 pci_save_state(pci_dev);
3746 return pci_set_power_state(pci_dev, PCI_D3hot);
3749 /* Used for both resume and restore */
3750 static int efx_pm_resume(struct device *dev)
3752 struct pci_dev *pci_dev = to_pci_dev(dev);
3753 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3756 rc = pci_set_power_state(pci_dev, PCI_D0);
3759 pci_restore_state(pci_dev);
3760 rc = pci_enable_device(pci_dev);
3763 pci_set_master(efx->pci_dev);
3764 rc = efx->type->reset(efx, RESET_TYPE_ALL);
3767 rc = efx->type->init(efx);
3770 rc = efx_pm_thaw(dev);
3774 static int efx_pm_suspend(struct device *dev)
3779 rc = efx_pm_poweroff(dev);
3785 static const struct dev_pm_ops efx_pm_ops = {
3786 .suspend = efx_pm_suspend,
3787 .resume = efx_pm_resume,
3788 .freeze = efx_pm_freeze,
3789 .thaw = efx_pm_thaw,
3790 .poweroff = efx_pm_poweroff,
3791 .restore = efx_pm_resume,
3794 /* A PCI error affecting this device was detected.
3795 * At this point MMIO and DMA may be disabled.
3796 * Stop the software path and request a slot reset.
3798 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3799 enum pci_channel_state state)
3801 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3802 struct efx_nic *efx = pci_get_drvdata(pdev);
3804 if (state == pci_channel_io_perm_failure)
3805 return PCI_ERS_RESULT_DISCONNECT;
3809 if (efx->state != STATE_DISABLED) {
3810 efx->state = STATE_RECOVERY;
3811 efx->reset_pending = 0;
3813 efx_device_detach_sync(efx);
3816 efx_disable_interrupts(efx);
3818 status = PCI_ERS_RESULT_NEED_RESET;
3820 /* If the interface is disabled we don't want to do anything
3823 status = PCI_ERS_RESULT_RECOVERED;
3828 pci_disable_device(pdev);
3833 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3834 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3836 struct efx_nic *efx = pci_get_drvdata(pdev);
3837 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3840 if (pci_enable_device(pdev)) {
3841 netif_err(efx, hw, efx->net_dev,
3842 "Cannot re-enable PCI device after reset.\n");
3843 status = PCI_ERS_RESULT_DISCONNECT;
3846 rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3848 netif_err(efx, hw, efx->net_dev,
3849 "pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3850 /* Non-fatal error. Continue. */
3856 /* Perform the actual reset and resume I/O operations. */
3857 static void efx_io_resume(struct pci_dev *pdev)
3859 struct efx_nic *efx = pci_get_drvdata(pdev);
3864 if (efx->state == STATE_DISABLED)
3867 rc = efx_reset(efx, RESET_TYPE_ALL);
3869 netif_err(efx, hw, efx->net_dev,
3870 "efx_reset failed after PCI error (%d)\n", rc);
3872 efx->state = STATE_READY;
3873 netif_dbg(efx, hw, efx->net_dev,
3874 "Done resetting and resuming IO after PCI error.\n");
3881 /* For simplicity and reliability, we always require a slot reset and try to
3882 * reset the hardware when a pci error affecting the device is detected.
3883 * We leave both the link_reset and mmio_enabled callback unimplemented:
3884 * with our request for slot reset the mmio_enabled callback will never be
3885 * called, and the link_reset callback is not used by AER or EEH mechanisms.
3887 static const struct pci_error_handlers efx_err_handlers = {
3888 .error_detected = efx_io_error_detected,
3889 .slot_reset = efx_io_slot_reset,
3890 .resume = efx_io_resume,
3893 static struct pci_driver efx_pci_driver = {
3894 .name = KBUILD_MODNAME,
3895 .id_table = efx_pci_table,
3896 .probe = efx_pci_probe,
3897 .remove = efx_pci_remove,
3898 .driver.pm = &efx_pm_ops,
3899 .err_handler = &efx_err_handlers,
3900 #ifdef CONFIG_SFC_SRIOV
3901 .sriov_configure = efx_pci_sriov_configure,
3905 /**************************************************************************
3907 * Kernel module interface
3909 *************************************************************************/
3911 module_param(interrupt_mode, uint, 0444);
3912 MODULE_PARM_DESC(interrupt_mode,
3913 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3915 static int __init efx_init_module(void)
3919 printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3921 rc = register_netdevice_notifier(&efx_netdev_notifier);
3925 #ifdef CONFIG_SFC_SRIOV
3926 rc = efx_init_sriov();
3931 reset_workqueue = create_singlethread_workqueue("sfc_reset");
3932 if (!reset_workqueue) {
3937 rc = pci_register_driver(&efx_pci_driver);
3944 destroy_workqueue(reset_workqueue);
3946 #ifdef CONFIG_SFC_SRIOV
3950 unregister_netdevice_notifier(&efx_netdev_notifier);
3955 static void __exit efx_exit_module(void)
3957 printk(KERN_INFO "Solarflare NET driver unloading\n");
3959 pci_unregister_driver(&efx_pci_driver);
3960 destroy_workqueue(reset_workqueue);
3961 #ifdef CONFIG_SFC_SRIOV
3964 unregister_netdevice_notifier(&efx_netdev_notifier);
3968 module_init(efx_init_module);
3969 module_exit(efx_exit_module);
3971 MODULE_AUTHOR("Solarflare Communications and "
3972 "Michael Brown <mbrown@fensystems.co.uk>");
3973 MODULE_DESCRIPTION("Solarflare network driver");
3974 MODULE_LICENSE("GPL");
3975 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3976 MODULE_VERSION(EFX_DRIVER_VERSION);