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>
34 #include "mcdi_pcol.h"
35 #include "workarounds.h"
37 /**************************************************************************
41 **************************************************************************
44 /* Loopback mode names (see LOOPBACK_MODE()) */
45 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
46 const char *const efx_loopback_mode_names[] = {
47 [LOOPBACK_NONE] = "NONE",
48 [LOOPBACK_DATA] = "DATAPATH",
49 [LOOPBACK_GMAC] = "GMAC",
50 [LOOPBACK_XGMII] = "XGMII",
51 [LOOPBACK_XGXS] = "XGXS",
52 [LOOPBACK_XAUI] = "XAUI",
53 [LOOPBACK_GMII] = "GMII",
54 [LOOPBACK_SGMII] = "SGMII",
55 [LOOPBACK_XGBR] = "XGBR",
56 [LOOPBACK_XFI] = "XFI",
57 [LOOPBACK_XAUI_FAR] = "XAUI_FAR",
58 [LOOPBACK_GMII_FAR] = "GMII_FAR",
59 [LOOPBACK_SGMII_FAR] = "SGMII_FAR",
60 [LOOPBACK_XFI_FAR] = "XFI_FAR",
61 [LOOPBACK_GPHY] = "GPHY",
62 [LOOPBACK_PHYXS] = "PHYXS",
63 [LOOPBACK_PCS] = "PCS",
64 [LOOPBACK_PMAPMD] = "PMA/PMD",
65 [LOOPBACK_XPORT] = "XPORT",
66 [LOOPBACK_XGMII_WS] = "XGMII_WS",
67 [LOOPBACK_XAUI_WS] = "XAUI_WS",
68 [LOOPBACK_XAUI_WS_FAR] = "XAUI_WS_FAR",
69 [LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
70 [LOOPBACK_GMII_WS] = "GMII_WS",
71 [LOOPBACK_XFI_WS] = "XFI_WS",
72 [LOOPBACK_XFI_WS_FAR] = "XFI_WS_FAR",
73 [LOOPBACK_PHYXS_WS] = "PHYXS_WS",
76 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
77 const char *const efx_reset_type_names[] = {
78 [RESET_TYPE_INVISIBLE] = "INVISIBLE",
79 [RESET_TYPE_ALL] = "ALL",
80 [RESET_TYPE_RECOVER_OR_ALL] = "RECOVER_OR_ALL",
81 [RESET_TYPE_WORLD] = "WORLD",
82 [RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
83 [RESET_TYPE_DATAPATH] = "DATAPATH",
84 [RESET_TYPE_MC_BIST] = "MC_BIST",
85 [RESET_TYPE_DISABLE] = "DISABLE",
86 [RESET_TYPE_TX_WATCHDOG] = "TX_WATCHDOG",
87 [RESET_TYPE_INT_ERROR] = "INT_ERROR",
88 [RESET_TYPE_DMA_ERROR] = "DMA_ERROR",
89 [RESET_TYPE_TX_SKIP] = "TX_SKIP",
90 [RESET_TYPE_MC_FAILURE] = "MC_FAILURE",
91 [RESET_TYPE_MCDI_TIMEOUT] = "MCDI_TIMEOUT (FLR)",
94 /* UDP tunnel type names */
95 static const char *const efx_udp_tunnel_type_names[] = {
96 [TUNNEL_ENCAP_UDP_PORT_ENTRY_VXLAN] = "vxlan",
97 [TUNNEL_ENCAP_UDP_PORT_ENTRY_GENEVE] = "geneve",
100 void efx_get_udp_tunnel_type_name(u16 type, char *buf, size_t buflen)
102 if (type < ARRAY_SIZE(efx_udp_tunnel_type_names) &&
103 efx_udp_tunnel_type_names[type] != NULL)
104 snprintf(buf, buflen, "%s", efx_udp_tunnel_type_names[type]);
106 snprintf(buf, buflen, "type %d", type);
109 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
110 * queued onto this work queue. This is not a per-nic work queue, because
111 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
113 static struct workqueue_struct *reset_workqueue;
115 /* How often and how many times to poll for a reset while waiting for a
116 * BIST that another function started to complete.
118 #define BIST_WAIT_DELAY_MS 100
119 #define BIST_WAIT_DELAY_COUNT 100
121 /**************************************************************************
123 * Configurable values
125 *************************************************************************/
128 * Use separate channels for TX and RX events
130 * Set this to 1 to use separate channels for TX and RX. It allows us
131 * to control interrupt affinity separately for TX and RX.
133 * This is only used in MSI-X interrupt mode
135 bool efx_separate_tx_channels;
136 module_param(efx_separate_tx_channels, bool, 0444);
137 MODULE_PARM_DESC(efx_separate_tx_channels,
138 "Use separate channels for TX and RX");
140 /* This is the weight assigned to each of the (per-channel) virtual
143 static int napi_weight = 64;
145 /* This is the time (in jiffies) between invocations of the hardware
147 * On Falcon-based NICs, this will:
148 * - Check the on-board hardware monitor;
149 * - Poll the link state and reconfigure the hardware as necessary.
150 * On Siena-based NICs for power systems with EEH support, this will give EEH a
153 static unsigned int efx_monitor_interval = 1 * HZ;
155 /* Initial interrupt moderation settings. They can be modified after
156 * module load with ethtool.
158 * The default for RX should strike a balance between increasing the
159 * round-trip latency and reducing overhead.
161 static unsigned int rx_irq_mod_usec = 60;
163 /* Initial interrupt moderation settings. They can be modified after
164 * module load with ethtool.
166 * This default is chosen to ensure that a 10G link does not go idle
167 * while a TX queue is stopped after it has become full. A queue is
168 * restarted when it drops below half full. The time this takes (assuming
169 * worst case 3 descriptors per packet and 1024 descriptors) is
170 * 512 / 3 * 1.2 = 205 usec.
172 static unsigned int tx_irq_mod_usec = 150;
174 /* This is the first interrupt mode to try out of:
179 static unsigned int interrupt_mode;
181 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
182 * i.e. the number of CPUs among which we may distribute simultaneous
183 * interrupt handling.
185 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
186 * The default (0) means to assign an interrupt to each core.
188 static unsigned int rss_cpus;
189 module_param(rss_cpus, uint, 0444);
190 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
192 static bool phy_flash_cfg;
193 module_param(phy_flash_cfg, bool, 0644);
194 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
196 static unsigned irq_adapt_low_thresh = 8000;
197 module_param(irq_adapt_low_thresh, uint, 0644);
198 MODULE_PARM_DESC(irq_adapt_low_thresh,
199 "Threshold score for reducing IRQ moderation");
201 static unsigned irq_adapt_high_thresh = 16000;
202 module_param(irq_adapt_high_thresh, uint, 0644);
203 MODULE_PARM_DESC(irq_adapt_high_thresh,
204 "Threshold score for increasing IRQ moderation");
206 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
207 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
208 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
209 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
210 module_param(debug, uint, 0);
211 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
213 /**************************************************************************
215 * Utility functions and prototypes
217 *************************************************************************/
219 static int efx_soft_enable_interrupts(struct efx_nic *efx);
220 static void efx_soft_disable_interrupts(struct efx_nic *efx);
221 static void efx_remove_channel(struct efx_channel *channel);
222 static void efx_remove_channels(struct efx_nic *efx);
223 static const struct efx_channel_type efx_default_channel_type;
224 static void efx_remove_port(struct efx_nic *efx);
225 static void efx_init_napi_channel(struct efx_channel *channel);
226 static void efx_fini_napi(struct efx_nic *efx);
227 static void efx_fini_napi_channel(struct efx_channel *channel);
228 static void efx_fini_struct(struct efx_nic *efx);
229 static void efx_start_all(struct efx_nic *efx);
230 static void efx_stop_all(struct efx_nic *efx);
232 #define EFX_ASSERT_RESET_SERIALISED(efx) \
234 if ((efx->state == STATE_READY) || \
235 (efx->state == STATE_RECOVERY) || \
236 (efx->state == STATE_DISABLED)) \
240 static int efx_check_disabled(struct efx_nic *efx)
242 if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
243 netif_err(efx, drv, efx->net_dev,
244 "device is disabled due to earlier errors\n");
250 /**************************************************************************
252 * Event queue processing
254 *************************************************************************/
256 /* Process channel's event queue
258 * This function is responsible for processing the event queue of a
259 * single channel. The caller must guarantee that this function will
260 * never be concurrently called more than once on the same channel,
261 * though different channels may be being processed concurrently.
263 static int efx_process_channel(struct efx_channel *channel, int budget)
265 struct efx_tx_queue *tx_queue;
268 if (unlikely(!channel->enabled))
271 efx_for_each_channel_tx_queue(tx_queue, channel) {
272 tx_queue->pkts_compl = 0;
273 tx_queue->bytes_compl = 0;
276 spent = efx_nic_process_eventq(channel, budget);
277 if (spent && efx_channel_has_rx_queue(channel)) {
278 struct efx_rx_queue *rx_queue =
279 efx_channel_get_rx_queue(channel);
281 efx_rx_flush_packet(channel);
282 efx_fast_push_rx_descriptors(rx_queue, true);
286 efx_for_each_channel_tx_queue(tx_queue, channel) {
287 if (tx_queue->bytes_compl) {
288 netdev_tx_completed_queue(tx_queue->core_txq,
289 tx_queue->pkts_compl, tx_queue->bytes_compl);
298 * NAPI guarantees serialisation of polls of the same device, which
299 * provides the guarantee required by efx_process_channel().
301 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel)
303 int step = efx->irq_mod_step_us;
305 if (channel->irq_mod_score < irq_adapt_low_thresh) {
306 if (channel->irq_moderation_us > step) {
307 channel->irq_moderation_us -= step;
308 efx->type->push_irq_moderation(channel);
310 } else if (channel->irq_mod_score > irq_adapt_high_thresh) {
311 if (channel->irq_moderation_us <
312 efx->irq_rx_moderation_us) {
313 channel->irq_moderation_us += step;
314 efx->type->push_irq_moderation(channel);
318 channel->irq_count = 0;
319 channel->irq_mod_score = 0;
322 static int efx_poll(struct napi_struct *napi, int budget)
324 struct efx_channel *channel =
325 container_of(napi, struct efx_channel, napi_str);
326 struct efx_nic *efx = channel->efx;
329 netif_vdbg(efx, intr, efx->net_dev,
330 "channel %d NAPI poll executing on CPU %d\n",
331 channel->channel, raw_smp_processor_id());
333 spent = efx_process_channel(channel, budget);
335 if (spent < budget) {
336 if (efx_channel_has_rx_queue(channel) &&
337 efx->irq_rx_adaptive &&
338 unlikely(++channel->irq_count == 1000)) {
339 efx_update_irq_mod(efx, channel);
342 efx_filter_rfs_expire(channel);
344 /* There is no race here; although napi_disable() will
345 * only wait for napi_complete(), this isn't a problem
346 * since efx_nic_eventq_read_ack() will have no effect if
347 * interrupts have already been disabled.
349 if (napi_complete_done(napi, spent))
350 efx_nic_eventq_read_ack(channel);
356 /* Create event queue
357 * Event queue memory allocations are done only once. If the channel
358 * is reset, the memory buffer will be reused; this guards against
359 * errors during channel reset and also simplifies interrupt handling.
361 static int efx_probe_eventq(struct efx_channel *channel)
363 struct efx_nic *efx = channel->efx;
364 unsigned long entries;
366 netif_dbg(efx, probe, efx->net_dev,
367 "chan %d create event queue\n", channel->channel);
369 /* Build an event queue with room for one event per tx and rx buffer,
370 * plus some extra for link state events and MCDI completions. */
371 entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
372 EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
373 channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
375 return efx_nic_probe_eventq(channel);
378 /* Prepare channel's event queue */
379 static int efx_init_eventq(struct efx_channel *channel)
381 struct efx_nic *efx = channel->efx;
384 EFX_WARN_ON_PARANOID(channel->eventq_init);
386 netif_dbg(efx, drv, efx->net_dev,
387 "chan %d init event queue\n", channel->channel);
389 rc = efx_nic_init_eventq(channel);
391 efx->type->push_irq_moderation(channel);
392 channel->eventq_read_ptr = 0;
393 channel->eventq_init = true;
398 /* Enable event queue processing and NAPI */
399 void efx_start_eventq(struct efx_channel *channel)
401 netif_dbg(channel->efx, ifup, channel->efx->net_dev,
402 "chan %d start event queue\n", channel->channel);
404 /* Make sure the NAPI handler sees the enabled flag set */
405 channel->enabled = true;
408 napi_enable(&channel->napi_str);
409 efx_nic_eventq_read_ack(channel);
412 /* Disable event queue processing and NAPI */
413 void efx_stop_eventq(struct efx_channel *channel)
415 if (!channel->enabled)
418 napi_disable(&channel->napi_str);
419 channel->enabled = false;
422 static void efx_fini_eventq(struct efx_channel *channel)
424 if (!channel->eventq_init)
427 netif_dbg(channel->efx, drv, channel->efx->net_dev,
428 "chan %d fini event queue\n", channel->channel);
430 efx_nic_fini_eventq(channel);
431 channel->eventq_init = false;
434 static void efx_remove_eventq(struct efx_channel *channel)
436 netif_dbg(channel->efx, drv, channel->efx->net_dev,
437 "chan %d remove event queue\n", channel->channel);
439 efx_nic_remove_eventq(channel);
442 /**************************************************************************
446 *************************************************************************/
448 /* Allocate and initialise a channel structure. */
449 static struct efx_channel *
450 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
452 struct efx_channel *channel;
453 struct efx_rx_queue *rx_queue;
454 struct efx_tx_queue *tx_queue;
457 channel = kzalloc(sizeof(*channel), GFP_KERNEL);
462 channel->channel = i;
463 channel->type = &efx_default_channel_type;
465 for (j = 0; j < EFX_TXQ_TYPES; j++) {
466 tx_queue = &channel->tx_queue[j];
468 tx_queue->queue = i * EFX_TXQ_TYPES + j;
469 tx_queue->channel = channel;
472 rx_queue = &channel->rx_queue;
474 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
479 /* Allocate and initialise a channel structure, copying parameters
480 * (but not resources) from an old channel structure.
482 static struct efx_channel *
483 efx_copy_channel(const struct efx_channel *old_channel)
485 struct efx_channel *channel;
486 struct efx_rx_queue *rx_queue;
487 struct efx_tx_queue *tx_queue;
490 channel = kmalloc(sizeof(*channel), GFP_KERNEL);
494 *channel = *old_channel;
496 channel->napi_dev = NULL;
497 INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
498 channel->napi_str.napi_id = 0;
499 channel->napi_str.state = 0;
500 memset(&channel->eventq, 0, sizeof(channel->eventq));
502 for (j = 0; j < EFX_TXQ_TYPES; j++) {
503 tx_queue = &channel->tx_queue[j];
504 if (tx_queue->channel)
505 tx_queue->channel = channel;
506 tx_queue->buffer = NULL;
507 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
510 rx_queue = &channel->rx_queue;
511 rx_queue->buffer = NULL;
512 memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
513 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
518 static int efx_probe_channel(struct efx_channel *channel)
520 struct efx_tx_queue *tx_queue;
521 struct efx_rx_queue *rx_queue;
524 netif_dbg(channel->efx, probe, channel->efx->net_dev,
525 "creating channel %d\n", channel->channel);
527 rc = channel->type->pre_probe(channel);
531 rc = efx_probe_eventq(channel);
535 efx_for_each_channel_tx_queue(tx_queue, channel) {
536 rc = efx_probe_tx_queue(tx_queue);
541 efx_for_each_channel_rx_queue(rx_queue, channel) {
542 rc = efx_probe_rx_queue(rx_queue);
550 efx_remove_channel(channel);
555 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
557 struct efx_nic *efx = channel->efx;
561 number = channel->channel;
562 if (efx->tx_channel_offset == 0) {
564 } else if (channel->channel < efx->tx_channel_offset) {
568 number -= efx->tx_channel_offset;
570 snprintf(buf, len, "%s%s-%d", efx->name, type, number);
573 static void efx_set_channel_names(struct efx_nic *efx)
575 struct efx_channel *channel;
577 efx_for_each_channel(channel, efx)
578 channel->type->get_name(channel,
579 efx->msi_context[channel->channel].name,
580 sizeof(efx->msi_context[0].name));
583 static int efx_probe_channels(struct efx_nic *efx)
585 struct efx_channel *channel;
588 /* Restart special buffer allocation */
589 efx->next_buffer_table = 0;
591 /* Probe channels in reverse, so that any 'extra' channels
592 * use the start of the buffer table. This allows the traffic
593 * channels to be resized without moving them or wasting the
594 * entries before them.
596 efx_for_each_channel_rev(channel, efx) {
597 rc = efx_probe_channel(channel);
599 netif_err(efx, probe, efx->net_dev,
600 "failed to create channel %d\n",
605 efx_set_channel_names(efx);
610 efx_remove_channels(efx);
614 /* Channels are shutdown and reinitialised whilst the NIC is running
615 * to propagate configuration changes (mtu, checksum offload), or
616 * to clear hardware error conditions
618 static void efx_start_datapath(struct efx_nic *efx)
620 netdev_features_t old_features = efx->net_dev->features;
621 bool old_rx_scatter = efx->rx_scatter;
622 struct efx_tx_queue *tx_queue;
623 struct efx_rx_queue *rx_queue;
624 struct efx_channel *channel;
627 /* Calculate the rx buffer allocation parameters required to
628 * support the current MTU, including padding for header
629 * alignment and overruns.
631 efx->rx_dma_len = (efx->rx_prefix_size +
632 EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
633 efx->type->rx_buffer_padding);
634 rx_buf_len = (sizeof(struct efx_rx_page_state) +
635 efx->rx_ip_align + efx->rx_dma_len);
636 if (rx_buf_len <= PAGE_SIZE) {
637 efx->rx_scatter = efx->type->always_rx_scatter;
638 efx->rx_buffer_order = 0;
639 } else if (efx->type->can_rx_scatter) {
640 BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
641 BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
642 2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
643 EFX_RX_BUF_ALIGNMENT) >
645 efx->rx_scatter = true;
646 efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
647 efx->rx_buffer_order = 0;
649 efx->rx_scatter = false;
650 efx->rx_buffer_order = get_order(rx_buf_len);
653 efx_rx_config_page_split(efx);
654 if (efx->rx_buffer_order)
655 netif_dbg(efx, drv, efx->net_dev,
656 "RX buf len=%u; page order=%u batch=%u\n",
657 efx->rx_dma_len, efx->rx_buffer_order,
658 efx->rx_pages_per_batch);
660 netif_dbg(efx, drv, efx->net_dev,
661 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
662 efx->rx_dma_len, efx->rx_page_buf_step,
663 efx->rx_bufs_per_page, efx->rx_pages_per_batch);
665 /* Restore previously fixed features in hw_features and remove
666 * features which are fixed now
668 efx->net_dev->hw_features |= efx->net_dev->features;
669 efx->net_dev->hw_features &= ~efx->fixed_features;
670 efx->net_dev->features |= efx->fixed_features;
671 if (efx->net_dev->features != old_features)
672 netdev_features_change(efx->net_dev);
674 /* RX filters may also have scatter-enabled flags */
675 if (efx->rx_scatter != old_rx_scatter)
676 efx->type->filter_update_rx_scatter(efx);
678 /* We must keep at least one descriptor in a TX ring empty.
679 * We could avoid this when the queue size does not exactly
680 * match the hardware ring size, but it's not that important.
681 * Therefore we stop the queue when one more skb might fill
682 * the ring completely. We wake it when half way back to
685 efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
686 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
688 /* Initialise the channels */
689 efx_for_each_channel(channel, efx) {
690 efx_for_each_channel_tx_queue(tx_queue, channel) {
691 efx_init_tx_queue(tx_queue);
692 atomic_inc(&efx->active_queues);
695 efx_for_each_channel_rx_queue(rx_queue, channel) {
696 efx_init_rx_queue(rx_queue);
697 atomic_inc(&efx->active_queues);
698 efx_stop_eventq(channel);
699 efx_fast_push_rx_descriptors(rx_queue, false);
700 efx_start_eventq(channel);
703 WARN_ON(channel->rx_pkt_n_frags);
706 efx_ptp_start_datapath(efx);
708 if (netif_device_present(efx->net_dev))
709 netif_tx_wake_all_queues(efx->net_dev);
712 static void efx_stop_datapath(struct efx_nic *efx)
714 struct efx_channel *channel;
715 struct efx_tx_queue *tx_queue;
716 struct efx_rx_queue *rx_queue;
719 EFX_ASSERT_RESET_SERIALISED(efx);
720 BUG_ON(efx->port_enabled);
722 efx_ptp_stop_datapath(efx);
725 efx_for_each_channel(channel, efx) {
726 efx_for_each_channel_rx_queue(rx_queue, channel)
727 rx_queue->refill_enabled = false;
730 efx_for_each_channel(channel, efx) {
731 /* RX packet processing is pipelined, so wait for the
732 * NAPI handler to complete. At least event queue 0
733 * might be kept active by non-data events, so don't
734 * use napi_synchronize() but actually disable NAPI
737 if (efx_channel_has_rx_queue(channel)) {
738 efx_stop_eventq(channel);
739 efx_start_eventq(channel);
743 rc = efx->type->fini_dmaq(efx);
745 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
747 netif_dbg(efx, drv, efx->net_dev,
748 "successfully flushed all queues\n");
751 efx_for_each_channel(channel, efx) {
752 efx_for_each_channel_rx_queue(rx_queue, channel)
753 efx_fini_rx_queue(rx_queue);
754 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
755 efx_fini_tx_queue(tx_queue);
759 static void efx_remove_channel(struct efx_channel *channel)
761 struct efx_tx_queue *tx_queue;
762 struct efx_rx_queue *rx_queue;
764 netif_dbg(channel->efx, drv, channel->efx->net_dev,
765 "destroy chan %d\n", channel->channel);
767 efx_for_each_channel_rx_queue(rx_queue, channel)
768 efx_remove_rx_queue(rx_queue);
769 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
770 efx_remove_tx_queue(tx_queue);
771 efx_remove_eventq(channel);
772 channel->type->post_remove(channel);
775 static void efx_remove_channels(struct efx_nic *efx)
777 struct efx_channel *channel;
779 efx_for_each_channel(channel, efx)
780 efx_remove_channel(channel);
784 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
786 struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
787 u32 old_rxq_entries, old_txq_entries;
788 unsigned i, next_buffer_table = 0;
791 rc = efx_check_disabled(efx);
795 /* Not all channels should be reallocated. We must avoid
796 * reallocating their buffer table entries.
798 efx_for_each_channel(channel, efx) {
799 struct efx_rx_queue *rx_queue;
800 struct efx_tx_queue *tx_queue;
802 if (channel->type->copy)
804 next_buffer_table = max(next_buffer_table,
805 channel->eventq.index +
806 channel->eventq.entries);
807 efx_for_each_channel_rx_queue(rx_queue, channel)
808 next_buffer_table = max(next_buffer_table,
809 rx_queue->rxd.index +
810 rx_queue->rxd.entries);
811 efx_for_each_channel_tx_queue(tx_queue, channel)
812 next_buffer_table = max(next_buffer_table,
813 tx_queue->txd.index +
814 tx_queue->txd.entries);
817 efx_device_detach_sync(efx);
819 efx_soft_disable_interrupts(efx);
821 /* Clone channels (where possible) */
822 memset(other_channel, 0, sizeof(other_channel));
823 for (i = 0; i < efx->n_channels; i++) {
824 channel = efx->channel[i];
825 if (channel->type->copy)
826 channel = channel->type->copy(channel);
831 other_channel[i] = channel;
834 /* Swap entry counts and channel pointers */
835 old_rxq_entries = efx->rxq_entries;
836 old_txq_entries = efx->txq_entries;
837 efx->rxq_entries = rxq_entries;
838 efx->txq_entries = txq_entries;
839 for (i = 0; i < efx->n_channels; i++) {
840 channel = efx->channel[i];
841 efx->channel[i] = other_channel[i];
842 other_channel[i] = channel;
845 /* Restart buffer table allocation */
846 efx->next_buffer_table = next_buffer_table;
848 for (i = 0; i < efx->n_channels; i++) {
849 channel = efx->channel[i];
850 if (!channel->type->copy)
852 rc = efx_probe_channel(channel);
855 efx_init_napi_channel(efx->channel[i]);
859 /* Destroy unused channel structures */
860 for (i = 0; i < efx->n_channels; i++) {
861 channel = other_channel[i];
862 if (channel && channel->type->copy) {
863 efx_fini_napi_channel(channel);
864 efx_remove_channel(channel);
869 rc2 = efx_soft_enable_interrupts(efx);
872 netif_err(efx, drv, efx->net_dev,
873 "unable to restart interrupts on channel reallocation\n");
874 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
877 efx_device_attach_if_not_resetting(efx);
883 efx->rxq_entries = old_rxq_entries;
884 efx->txq_entries = old_txq_entries;
885 for (i = 0; i < efx->n_channels; i++) {
886 channel = efx->channel[i];
887 efx->channel[i] = other_channel[i];
888 other_channel[i] = channel;
893 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
895 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
898 static const struct efx_channel_type efx_default_channel_type = {
899 .pre_probe = efx_channel_dummy_op_int,
900 .post_remove = efx_channel_dummy_op_void,
901 .get_name = efx_get_channel_name,
902 .copy = efx_copy_channel,
903 .keep_eventq = false,
906 int efx_channel_dummy_op_int(struct efx_channel *channel)
911 void efx_channel_dummy_op_void(struct efx_channel *channel)
915 /**************************************************************************
919 **************************************************************************/
921 /* This ensures that the kernel is kept informed (via
922 * netif_carrier_on/off) of the link status, and also maintains the
923 * link status's stop on the port's TX queue.
925 void efx_link_status_changed(struct efx_nic *efx)
927 struct efx_link_state *link_state = &efx->link_state;
929 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
930 * that no events are triggered between unregister_netdev() and the
931 * driver unloading. A more general condition is that NETDEV_CHANGE
932 * can only be generated between NETDEV_UP and NETDEV_DOWN */
933 if (!netif_running(efx->net_dev))
936 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
937 efx->n_link_state_changes++;
940 netif_carrier_on(efx->net_dev);
942 netif_carrier_off(efx->net_dev);
945 /* Status message for kernel log */
947 netif_info(efx, link, efx->net_dev,
948 "link up at %uMbps %s-duplex (MTU %d)\n",
949 link_state->speed, link_state->fd ? "full" : "half",
952 netif_info(efx, link, efx->net_dev, "link down\n");
955 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
957 efx->link_advertising = advertising;
959 if (advertising & ADVERTISED_Pause)
960 efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
962 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
963 if (advertising & ADVERTISED_Asym_Pause)
964 efx->wanted_fc ^= EFX_FC_TX;
968 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
970 efx->wanted_fc = wanted_fc;
971 if (efx->link_advertising) {
972 if (wanted_fc & EFX_FC_RX)
973 efx->link_advertising |= (ADVERTISED_Pause |
974 ADVERTISED_Asym_Pause);
976 efx->link_advertising &= ~(ADVERTISED_Pause |
977 ADVERTISED_Asym_Pause);
978 if (wanted_fc & EFX_FC_TX)
979 efx->link_advertising ^= ADVERTISED_Asym_Pause;
983 static void efx_fini_port(struct efx_nic *efx);
985 /* We assume that efx->type->reconfigure_mac will always try to sync RX
986 * filters and therefore needs to read-lock the filter table against freeing
988 void efx_mac_reconfigure(struct efx_nic *efx)
990 down_read(&efx->filter_sem);
991 efx->type->reconfigure_mac(efx);
992 up_read(&efx->filter_sem);
995 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
996 * the MAC appropriately. All other PHY configuration changes are pushed
997 * through phy_op->set_settings(), and pushed asynchronously to the MAC
998 * through efx_monitor().
1000 * Callers must hold the mac_lock
1002 int __efx_reconfigure_port(struct efx_nic *efx)
1004 enum efx_phy_mode phy_mode;
1007 WARN_ON(!mutex_is_locked(&efx->mac_lock));
1009 /* Disable PHY transmit in mac level loopbacks */
1010 phy_mode = efx->phy_mode;
1011 if (LOOPBACK_INTERNAL(efx))
1012 efx->phy_mode |= PHY_MODE_TX_DISABLED;
1014 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
1016 rc = efx->type->reconfigure_port(efx);
1019 efx->phy_mode = phy_mode;
1024 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1026 int efx_reconfigure_port(struct efx_nic *efx)
1030 EFX_ASSERT_RESET_SERIALISED(efx);
1032 mutex_lock(&efx->mac_lock);
1033 rc = __efx_reconfigure_port(efx);
1034 mutex_unlock(&efx->mac_lock);
1039 /* Asynchronous work item for changing MAC promiscuity and multicast
1040 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
1042 static void efx_mac_work(struct work_struct *data)
1044 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
1046 mutex_lock(&efx->mac_lock);
1047 if (efx->port_enabled)
1048 efx_mac_reconfigure(efx);
1049 mutex_unlock(&efx->mac_lock);
1052 static int efx_probe_port(struct efx_nic *efx)
1056 netif_dbg(efx, probe, efx->net_dev, "create port\n");
1059 efx->phy_mode = PHY_MODE_SPECIAL;
1061 /* Connect up MAC/PHY operations table */
1062 rc = efx->type->probe_port(efx);
1066 /* Initialise MAC address to permanent address */
1067 ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1072 static int efx_init_port(struct efx_nic *efx)
1076 netif_dbg(efx, drv, efx->net_dev, "init port\n");
1078 mutex_lock(&efx->mac_lock);
1080 rc = efx->phy_op->init(efx);
1084 efx->port_initialized = true;
1086 /* Reconfigure the MAC before creating dma queues (required for
1087 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1088 efx_mac_reconfigure(efx);
1090 /* Ensure the PHY advertises the correct flow control settings */
1091 rc = efx->phy_op->reconfigure(efx);
1092 if (rc && rc != -EPERM)
1095 mutex_unlock(&efx->mac_lock);
1099 efx->phy_op->fini(efx);
1101 mutex_unlock(&efx->mac_lock);
1105 static void efx_start_port(struct efx_nic *efx)
1107 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1108 BUG_ON(efx->port_enabled);
1110 mutex_lock(&efx->mac_lock);
1111 efx->port_enabled = true;
1113 /* Ensure MAC ingress/egress is enabled */
1114 efx_mac_reconfigure(efx);
1116 mutex_unlock(&efx->mac_lock);
1119 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1120 * and the async self-test, wait for them to finish and prevent them
1121 * being scheduled again. This doesn't cover online resets, which
1122 * should only be cancelled when removing the device.
1124 static void efx_stop_port(struct efx_nic *efx)
1126 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1128 EFX_ASSERT_RESET_SERIALISED(efx);
1130 mutex_lock(&efx->mac_lock);
1131 efx->port_enabled = false;
1132 mutex_unlock(&efx->mac_lock);
1134 /* Serialise against efx_set_multicast_list() */
1135 netif_addr_lock_bh(efx->net_dev);
1136 netif_addr_unlock_bh(efx->net_dev);
1138 cancel_delayed_work_sync(&efx->monitor_work);
1139 efx_selftest_async_cancel(efx);
1140 cancel_work_sync(&efx->mac_work);
1143 static void efx_fini_port(struct efx_nic *efx)
1145 netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1147 if (!efx->port_initialized)
1150 efx->phy_op->fini(efx);
1151 efx->port_initialized = false;
1153 efx->link_state.up = false;
1154 efx_link_status_changed(efx);
1157 static void efx_remove_port(struct efx_nic *efx)
1159 netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1161 efx->type->remove_port(efx);
1164 /**************************************************************************
1168 **************************************************************************/
1170 static LIST_HEAD(efx_primary_list);
1171 static LIST_HEAD(efx_unassociated_list);
1173 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1175 return left->type == right->type &&
1176 left->vpd_sn && right->vpd_sn &&
1177 !strcmp(left->vpd_sn, right->vpd_sn);
1180 static void efx_associate(struct efx_nic *efx)
1182 struct efx_nic *other, *next;
1184 if (efx->primary == efx) {
1185 /* Adding primary function; look for secondaries */
1187 netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1188 list_add_tail(&efx->node, &efx_primary_list);
1190 list_for_each_entry_safe(other, next, &efx_unassociated_list,
1192 if (efx_same_controller(efx, other)) {
1193 list_del(&other->node);
1194 netif_dbg(other, probe, other->net_dev,
1195 "moving to secondary list of %s %s\n",
1196 pci_name(efx->pci_dev),
1197 efx->net_dev->name);
1198 list_add_tail(&other->node,
1199 &efx->secondary_list);
1200 other->primary = efx;
1204 /* Adding secondary function; look for primary */
1206 list_for_each_entry(other, &efx_primary_list, node) {
1207 if (efx_same_controller(efx, other)) {
1208 netif_dbg(efx, probe, efx->net_dev,
1209 "adding to secondary list of %s %s\n",
1210 pci_name(other->pci_dev),
1211 other->net_dev->name);
1212 list_add_tail(&efx->node,
1213 &other->secondary_list);
1214 efx->primary = other;
1219 netif_dbg(efx, probe, efx->net_dev,
1220 "adding to unassociated list\n");
1221 list_add_tail(&efx->node, &efx_unassociated_list);
1225 static void efx_dissociate(struct efx_nic *efx)
1227 struct efx_nic *other, *next;
1229 list_del(&efx->node);
1230 efx->primary = NULL;
1232 list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1233 list_del(&other->node);
1234 netif_dbg(other, probe, other->net_dev,
1235 "moving to unassociated list\n");
1236 list_add_tail(&other->node, &efx_unassociated_list);
1237 other->primary = NULL;
1241 /* This configures the PCI device to enable I/O and DMA. */
1242 static int efx_init_io(struct efx_nic *efx)
1244 struct pci_dev *pci_dev = efx->pci_dev;
1245 dma_addr_t dma_mask = efx->type->max_dma_mask;
1246 unsigned int mem_map_size = efx->type->mem_map_size(efx);
1249 netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1251 bar = efx->type->mem_bar;
1253 rc = pci_enable_device(pci_dev);
1255 netif_err(efx, probe, efx->net_dev,
1256 "failed to enable PCI device\n");
1260 pci_set_master(pci_dev);
1262 /* Set the PCI DMA mask. Try all possibilities from our
1263 * genuine mask down to 32 bits, because some architectures
1264 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1265 * masks event though they reject 46 bit masks.
1267 while (dma_mask > 0x7fffffffUL) {
1268 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1274 netif_err(efx, probe, efx->net_dev,
1275 "could not find a suitable DMA mask\n");
1278 netif_dbg(efx, probe, efx->net_dev,
1279 "using DMA mask %llx\n", (unsigned long long) dma_mask);
1281 efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1282 rc = pci_request_region(pci_dev, bar, "sfc");
1284 netif_err(efx, probe, efx->net_dev,
1285 "request for memory BAR failed\n");
1289 efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1290 if (!efx->membase) {
1291 netif_err(efx, probe, efx->net_dev,
1292 "could not map memory BAR at %llx+%x\n",
1293 (unsigned long long)efx->membase_phys, mem_map_size);
1297 netif_dbg(efx, probe, efx->net_dev,
1298 "memory BAR at %llx+%x (virtual %p)\n",
1299 (unsigned long long)efx->membase_phys, mem_map_size,
1305 pci_release_region(efx->pci_dev, bar);
1307 efx->membase_phys = 0;
1309 pci_disable_device(efx->pci_dev);
1314 static void efx_fini_io(struct efx_nic *efx)
1318 netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1321 iounmap(efx->membase);
1322 efx->membase = NULL;
1325 if (efx->membase_phys) {
1326 bar = efx->type->mem_bar;
1327 pci_release_region(efx->pci_dev, bar);
1328 efx->membase_phys = 0;
1331 /* Don't disable bus-mastering if VFs are assigned */
1332 if (!pci_vfs_assigned(efx->pci_dev))
1333 pci_disable_device(efx->pci_dev);
1336 void efx_set_default_rx_indir_table(struct efx_nic *efx)
1340 for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1341 efx->rx_indir_table[i] =
1342 ethtool_rxfh_indir_default(i, efx->rss_spread);
1345 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1347 cpumask_var_t thread_mask;
1354 if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1355 netif_warn(efx, probe, efx->net_dev,
1356 "RSS disabled due to allocation failure\n");
1361 for_each_online_cpu(cpu) {
1362 if (!cpumask_test_cpu(cpu, thread_mask)) {
1364 cpumask_or(thread_mask, thread_mask,
1365 topology_sibling_cpumask(cpu));
1369 free_cpumask_var(thread_mask);
1372 if (count > EFX_MAX_RX_QUEUES) {
1373 netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn,
1374 "Reducing number of rx queues from %u to %u.\n",
1375 count, EFX_MAX_RX_QUEUES);
1376 count = EFX_MAX_RX_QUEUES;
1379 /* If RSS is requested for the PF *and* VFs then we can't write RSS
1380 * table entries that are inaccessible to VFs
1382 #ifdef CONFIG_SFC_SRIOV
1383 if (efx->type->sriov_wanted) {
1384 if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1385 count > efx_vf_size(efx)) {
1386 netif_warn(efx, probe, efx->net_dev,
1387 "Reducing number of RSS channels from %u to %u for "
1388 "VF support. Increase vf-msix-limit to use more "
1389 "channels on the PF.\n",
1390 count, efx_vf_size(efx));
1391 count = efx_vf_size(efx);
1399 /* Probe the number and type of interrupts we are able to obtain, and
1400 * the resulting numbers of channels and RX queues.
1402 static int efx_probe_interrupts(struct efx_nic *efx)
1404 unsigned int extra_channels = 0;
1408 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1409 if (efx->extra_channel_type[i])
1412 if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1413 struct msix_entry xentries[EFX_MAX_CHANNELS];
1414 unsigned int n_channels;
1416 n_channels = efx_wanted_parallelism(efx);
1417 if (efx_separate_tx_channels)
1419 n_channels += extra_channels;
1420 n_channels = min(n_channels, efx->max_channels);
1422 for (i = 0; i < n_channels; i++)
1423 xentries[i].entry = i;
1424 rc = pci_enable_msix_range(efx->pci_dev,
1425 xentries, 1, n_channels);
1427 /* Fall back to single channel MSI */
1428 netif_err(efx, drv, efx->net_dev,
1429 "could not enable MSI-X\n");
1430 if (efx->type->min_interrupt_mode >= EFX_INT_MODE_MSI)
1431 efx->interrupt_mode = EFX_INT_MODE_MSI;
1434 } else if (rc < n_channels) {
1435 netif_err(efx, drv, efx->net_dev,
1436 "WARNING: Insufficient MSI-X vectors"
1437 " available (%d < %u).\n", rc, n_channels);
1438 netif_err(efx, drv, efx->net_dev,
1439 "WARNING: Performance may be reduced.\n");
1444 efx->n_channels = n_channels;
1445 if (n_channels > extra_channels)
1446 n_channels -= extra_channels;
1447 if (efx_separate_tx_channels) {
1448 efx->n_tx_channels = min(max(n_channels / 2,
1450 efx->max_tx_channels);
1451 efx->n_rx_channels = max(n_channels -
1455 efx->n_tx_channels = min(n_channels,
1456 efx->max_tx_channels);
1457 efx->n_rx_channels = n_channels;
1459 for (i = 0; i < efx->n_channels; i++)
1460 efx_get_channel(efx, i)->irq =
1465 /* Try single interrupt MSI */
1466 if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1467 efx->n_channels = 1;
1468 efx->n_rx_channels = 1;
1469 efx->n_tx_channels = 1;
1470 rc = pci_enable_msi(efx->pci_dev);
1472 efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1474 netif_err(efx, drv, efx->net_dev,
1475 "could not enable MSI\n");
1476 if (efx->type->min_interrupt_mode >= EFX_INT_MODE_LEGACY)
1477 efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1483 /* Assume legacy interrupts */
1484 if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1485 efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
1486 efx->n_rx_channels = 1;
1487 efx->n_tx_channels = 1;
1488 efx->legacy_irq = efx->pci_dev->irq;
1491 /* Assign extra channels if possible */
1492 j = efx->n_channels;
1493 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1494 if (!efx->extra_channel_type[i])
1496 if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1497 efx->n_channels <= extra_channels) {
1498 efx->extra_channel_type[i]->handle_no_channel(efx);
1501 efx_get_channel(efx, j)->type =
1502 efx->extra_channel_type[i];
1506 /* RSS might be usable on VFs even if it is disabled on the PF */
1507 #ifdef CONFIG_SFC_SRIOV
1508 if (efx->type->sriov_wanted) {
1509 efx->rss_spread = ((efx->n_rx_channels > 1 ||
1510 !efx->type->sriov_wanted(efx)) ?
1511 efx->n_rx_channels : efx_vf_size(efx));
1515 efx->rss_spread = efx->n_rx_channels;
1520 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1522 struct efx_channel *channel, *end_channel;
1525 BUG_ON(efx->state == STATE_DISABLED);
1527 efx->irq_soft_enabled = true;
1530 efx_for_each_channel(channel, efx) {
1531 if (!channel->type->keep_eventq) {
1532 rc = efx_init_eventq(channel);
1536 efx_start_eventq(channel);
1539 efx_mcdi_mode_event(efx);
1543 end_channel = channel;
1544 efx_for_each_channel(channel, efx) {
1545 if (channel == end_channel)
1547 efx_stop_eventq(channel);
1548 if (!channel->type->keep_eventq)
1549 efx_fini_eventq(channel);
1555 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1557 struct efx_channel *channel;
1559 if (efx->state == STATE_DISABLED)
1562 efx_mcdi_mode_poll(efx);
1564 efx->irq_soft_enabled = false;
1567 if (efx->legacy_irq)
1568 synchronize_irq(efx->legacy_irq);
1570 efx_for_each_channel(channel, efx) {
1572 synchronize_irq(channel->irq);
1574 efx_stop_eventq(channel);
1575 if (!channel->type->keep_eventq)
1576 efx_fini_eventq(channel);
1579 /* Flush the asynchronous MCDI request queue */
1580 efx_mcdi_flush_async(efx);
1583 static int efx_enable_interrupts(struct efx_nic *efx)
1585 struct efx_channel *channel, *end_channel;
1588 BUG_ON(efx->state == STATE_DISABLED);
1590 if (efx->eeh_disabled_legacy_irq) {
1591 enable_irq(efx->legacy_irq);
1592 efx->eeh_disabled_legacy_irq = false;
1595 efx->type->irq_enable_master(efx);
1597 efx_for_each_channel(channel, efx) {
1598 if (channel->type->keep_eventq) {
1599 rc = efx_init_eventq(channel);
1605 rc = efx_soft_enable_interrupts(efx);
1612 end_channel = channel;
1613 efx_for_each_channel(channel, efx) {
1614 if (channel == end_channel)
1616 if (channel->type->keep_eventq)
1617 efx_fini_eventq(channel);
1620 efx->type->irq_disable_non_ev(efx);
1625 static void efx_disable_interrupts(struct efx_nic *efx)
1627 struct efx_channel *channel;
1629 efx_soft_disable_interrupts(efx);
1631 efx_for_each_channel(channel, efx) {
1632 if (channel->type->keep_eventq)
1633 efx_fini_eventq(channel);
1636 efx->type->irq_disable_non_ev(efx);
1639 static void efx_remove_interrupts(struct efx_nic *efx)
1641 struct efx_channel *channel;
1643 /* Remove MSI/MSI-X interrupts */
1644 efx_for_each_channel(channel, efx)
1646 pci_disable_msi(efx->pci_dev);
1647 pci_disable_msix(efx->pci_dev);
1649 /* Remove legacy interrupt */
1650 efx->legacy_irq = 0;
1653 static void efx_set_channels(struct efx_nic *efx)
1655 struct efx_channel *channel;
1656 struct efx_tx_queue *tx_queue;
1658 efx->tx_channel_offset =
1659 efx_separate_tx_channels ?
1660 efx->n_channels - efx->n_tx_channels : 0;
1662 /* We need to mark which channels really have RX and TX
1663 * queues, and adjust the TX queue numbers if we have separate
1664 * RX-only and TX-only channels.
1666 efx_for_each_channel(channel, efx) {
1667 if (channel->channel < efx->n_rx_channels)
1668 channel->rx_queue.core_index = channel->channel;
1670 channel->rx_queue.core_index = -1;
1672 efx_for_each_channel_tx_queue(tx_queue, channel)
1673 tx_queue->queue -= (efx->tx_channel_offset *
1678 static int efx_probe_nic(struct efx_nic *efx)
1682 netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1684 /* Carry out hardware-type specific initialisation */
1685 rc = efx->type->probe(efx);
1690 if (!efx->max_channels || !efx->max_tx_channels) {
1691 netif_err(efx, drv, efx->net_dev,
1692 "Insufficient resources to allocate"
1698 /* Determine the number of channels and queues by trying
1699 * to hook in MSI-X interrupts.
1701 rc = efx_probe_interrupts(efx);
1705 efx_set_channels(efx);
1707 /* dimension_resources can fail with EAGAIN */
1708 rc = efx->type->dimension_resources(efx);
1709 if (rc != 0 && rc != -EAGAIN)
1713 /* try again with new max_channels */
1714 efx_remove_interrupts(efx);
1716 } while (rc == -EAGAIN);
1718 if (efx->n_channels > 1)
1719 netdev_rss_key_fill(&efx->rx_hash_key,
1720 sizeof(efx->rx_hash_key));
1721 efx_set_default_rx_indir_table(efx);
1723 netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1724 netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1726 /* Initialise the interrupt moderation settings */
1727 efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
1728 efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1734 efx_remove_interrupts(efx);
1736 efx->type->remove(efx);
1740 static void efx_remove_nic(struct efx_nic *efx)
1742 netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1744 efx_remove_interrupts(efx);
1745 efx->type->remove(efx);
1748 static int efx_probe_filters(struct efx_nic *efx)
1752 spin_lock_init(&efx->filter_lock);
1753 init_rwsem(&efx->filter_sem);
1754 mutex_lock(&efx->mac_lock);
1755 down_write(&efx->filter_sem);
1756 rc = efx->type->filter_table_probe(efx);
1760 #ifdef CONFIG_RFS_ACCEL
1761 if (efx->type->offload_features & NETIF_F_NTUPLE) {
1762 struct efx_channel *channel;
1765 efx_for_each_channel(channel, efx) {
1766 channel->rps_flow_id =
1767 kcalloc(efx->type->max_rx_ip_filters,
1768 sizeof(*channel->rps_flow_id),
1770 if (!channel->rps_flow_id)
1774 i < efx->type->max_rx_ip_filters;
1776 channel->rps_flow_id[i] =
1777 RPS_FLOW_ID_INVALID;
1781 efx_for_each_channel(channel, efx)
1782 kfree(channel->rps_flow_id);
1783 efx->type->filter_table_remove(efx);
1788 efx->rps_expire_index = efx->rps_expire_channel = 0;
1792 up_write(&efx->filter_sem);
1793 mutex_unlock(&efx->mac_lock);
1797 static void efx_remove_filters(struct efx_nic *efx)
1799 #ifdef CONFIG_RFS_ACCEL
1800 struct efx_channel *channel;
1802 efx_for_each_channel(channel, efx)
1803 kfree(channel->rps_flow_id);
1805 down_write(&efx->filter_sem);
1806 efx->type->filter_table_remove(efx);
1807 up_write(&efx->filter_sem);
1810 static void efx_restore_filters(struct efx_nic *efx)
1812 down_read(&efx->filter_sem);
1813 efx->type->filter_table_restore(efx);
1814 up_read(&efx->filter_sem);
1817 /**************************************************************************
1819 * NIC startup/shutdown
1821 *************************************************************************/
1823 static int efx_probe_all(struct efx_nic *efx)
1827 rc = efx_probe_nic(efx);
1829 netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1833 rc = efx_probe_port(efx);
1835 netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1839 BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1840 if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1844 efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1846 #ifdef CONFIG_SFC_SRIOV
1847 rc = efx->type->vswitching_probe(efx);
1848 if (rc) /* not fatal; the PF will still work fine */
1849 netif_warn(efx, probe, efx->net_dev,
1850 "failed to setup vswitching rc=%d;"
1851 " VFs may not function\n", rc);
1854 rc = efx_probe_filters(efx);
1856 netif_err(efx, probe, efx->net_dev,
1857 "failed to create filter tables\n");
1861 rc = efx_probe_channels(efx);
1868 efx_remove_filters(efx);
1870 #ifdef CONFIG_SFC_SRIOV
1871 efx->type->vswitching_remove(efx);
1874 efx_remove_port(efx);
1876 efx_remove_nic(efx);
1881 /* If the interface is supposed to be running but is not, start
1882 * the hardware and software data path, regular activity for the port
1883 * (MAC statistics, link polling, etc.) and schedule the port to be
1884 * reconfigured. Interrupts must already be enabled. This function
1885 * is safe to call multiple times, so long as the NIC is not disabled.
1886 * Requires the RTNL lock.
1888 static void efx_start_all(struct efx_nic *efx)
1890 EFX_ASSERT_RESET_SERIALISED(efx);
1891 BUG_ON(efx->state == STATE_DISABLED);
1893 /* Check that it is appropriate to restart the interface. All
1894 * of these flags are safe to read under just the rtnl lock */
1895 if (efx->port_enabled || !netif_running(efx->net_dev) ||
1899 efx_start_port(efx);
1900 efx_start_datapath(efx);
1902 /* Start the hardware monitor if there is one */
1903 if (efx->type->monitor != NULL)
1904 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1905 efx_monitor_interval);
1907 /* Link state detection is normally event-driven; we have
1908 * to poll now because we could have missed a change
1910 mutex_lock(&efx->mac_lock);
1911 if (efx->phy_op->poll(efx))
1912 efx_link_status_changed(efx);
1913 mutex_unlock(&efx->mac_lock);
1915 efx->type->start_stats(efx);
1916 efx->type->pull_stats(efx);
1917 spin_lock_bh(&efx->stats_lock);
1918 efx->type->update_stats(efx, NULL, NULL);
1919 spin_unlock_bh(&efx->stats_lock);
1922 /* Quiesce the hardware and software data path, and regular activity
1923 * for the port without bringing the link down. Safe to call multiple
1924 * times with the NIC in almost any state, but interrupts should be
1925 * enabled. Requires the RTNL lock.
1927 static void efx_stop_all(struct efx_nic *efx)
1929 EFX_ASSERT_RESET_SERIALISED(efx);
1931 /* port_enabled can be read safely under the rtnl lock */
1932 if (!efx->port_enabled)
1935 /* update stats before we go down so we can accurately count
1938 efx->type->pull_stats(efx);
1939 spin_lock_bh(&efx->stats_lock);
1940 efx->type->update_stats(efx, NULL, NULL);
1941 spin_unlock_bh(&efx->stats_lock);
1942 efx->type->stop_stats(efx);
1945 /* Stop the kernel transmit interface. This is only valid if
1946 * the device is stopped or detached; otherwise the watchdog
1947 * may fire immediately.
1949 WARN_ON(netif_running(efx->net_dev) &&
1950 netif_device_present(efx->net_dev));
1951 netif_tx_disable(efx->net_dev);
1953 efx_stop_datapath(efx);
1956 static void efx_remove_all(struct efx_nic *efx)
1958 efx_remove_channels(efx);
1959 efx_remove_filters(efx);
1960 #ifdef CONFIG_SFC_SRIOV
1961 efx->type->vswitching_remove(efx);
1963 efx_remove_port(efx);
1964 efx_remove_nic(efx);
1967 /**************************************************************************
1969 * Interrupt moderation
1971 **************************************************************************/
1972 unsigned int efx_usecs_to_ticks(struct efx_nic *efx, unsigned int usecs)
1976 if (usecs * 1000 < efx->timer_quantum_ns)
1977 return 1; /* never round down to 0 */
1978 return usecs * 1000 / efx->timer_quantum_ns;
1981 unsigned int efx_ticks_to_usecs(struct efx_nic *efx, unsigned int ticks)
1983 /* We must round up when converting ticks to microseconds
1984 * because we round down when converting the other way.
1986 return DIV_ROUND_UP(ticks * efx->timer_quantum_ns, 1000);
1989 /* Set interrupt moderation parameters */
1990 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1991 unsigned int rx_usecs, bool rx_adaptive,
1992 bool rx_may_override_tx)
1994 struct efx_channel *channel;
1995 unsigned int timer_max_us;
1997 EFX_ASSERT_RESET_SERIALISED(efx);
1999 timer_max_us = efx->timer_max_ns / 1000;
2001 if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
2004 if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
2005 !rx_may_override_tx) {
2006 netif_err(efx, drv, efx->net_dev, "Channels are shared. "
2007 "RX and TX IRQ moderation must be equal\n");
2011 efx->irq_rx_adaptive = rx_adaptive;
2012 efx->irq_rx_moderation_us = rx_usecs;
2013 efx_for_each_channel(channel, efx) {
2014 if (efx_channel_has_rx_queue(channel))
2015 channel->irq_moderation_us = rx_usecs;
2016 else if (efx_channel_has_tx_queues(channel))
2017 channel->irq_moderation_us = tx_usecs;
2023 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
2024 unsigned int *rx_usecs, bool *rx_adaptive)
2026 *rx_adaptive = efx->irq_rx_adaptive;
2027 *rx_usecs = efx->irq_rx_moderation_us;
2029 /* If channels are shared between RX and TX, so is IRQ
2030 * moderation. Otherwise, IRQ moderation is the same for all
2031 * TX channels and is not adaptive.
2033 if (efx->tx_channel_offset == 0) {
2034 *tx_usecs = *rx_usecs;
2036 struct efx_channel *tx_channel;
2038 tx_channel = efx->channel[efx->tx_channel_offset];
2039 *tx_usecs = tx_channel->irq_moderation_us;
2043 /**************************************************************************
2047 **************************************************************************/
2049 /* Run periodically off the general workqueue */
2050 static void efx_monitor(struct work_struct *data)
2052 struct efx_nic *efx = container_of(data, struct efx_nic,
2055 netif_vdbg(efx, timer, efx->net_dev,
2056 "hardware monitor executing on CPU %d\n",
2057 raw_smp_processor_id());
2058 BUG_ON(efx->type->monitor == NULL);
2060 /* If the mac_lock is already held then it is likely a port
2061 * reconfiguration is already in place, which will likely do
2062 * most of the work of monitor() anyway. */
2063 if (mutex_trylock(&efx->mac_lock)) {
2064 if (efx->port_enabled)
2065 efx->type->monitor(efx);
2066 mutex_unlock(&efx->mac_lock);
2069 queue_delayed_work(efx->workqueue, &efx->monitor_work,
2070 efx_monitor_interval);
2073 /**************************************************************************
2077 *************************************************************************/
2080 * Context: process, rtnl_lock() held.
2082 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
2084 struct efx_nic *efx = netdev_priv(net_dev);
2085 struct mii_ioctl_data *data = if_mii(ifr);
2087 if (cmd == SIOCSHWTSTAMP)
2088 return efx_ptp_set_ts_config(efx, ifr);
2089 if (cmd == SIOCGHWTSTAMP)
2090 return efx_ptp_get_ts_config(efx, ifr);
2092 /* Convert phy_id from older PRTAD/DEVAD format */
2093 if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2094 (data->phy_id & 0xfc00) == 0x0400)
2095 data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2097 return mdio_mii_ioctl(&efx->mdio, data, cmd);
2100 /**************************************************************************
2104 **************************************************************************/
2106 static void efx_init_napi_channel(struct efx_channel *channel)
2108 struct efx_nic *efx = channel->efx;
2110 channel->napi_dev = efx->net_dev;
2111 netif_napi_add(channel->napi_dev, &channel->napi_str,
2112 efx_poll, napi_weight);
2115 static void efx_init_napi(struct efx_nic *efx)
2117 struct efx_channel *channel;
2119 efx_for_each_channel(channel, efx)
2120 efx_init_napi_channel(channel);
2123 static void efx_fini_napi_channel(struct efx_channel *channel)
2125 if (channel->napi_dev)
2126 netif_napi_del(&channel->napi_str);
2128 channel->napi_dev = NULL;
2131 static void efx_fini_napi(struct efx_nic *efx)
2133 struct efx_channel *channel;
2135 efx_for_each_channel(channel, efx)
2136 efx_fini_napi_channel(channel);
2139 /**************************************************************************
2141 * Kernel netpoll interface
2143 *************************************************************************/
2145 #ifdef CONFIG_NET_POLL_CONTROLLER
2147 /* Although in the common case interrupts will be disabled, this is not
2148 * guaranteed. However, all our work happens inside the NAPI callback,
2149 * so no locking is required.
2151 static void efx_netpoll(struct net_device *net_dev)
2153 struct efx_nic *efx = netdev_priv(net_dev);
2154 struct efx_channel *channel;
2156 efx_for_each_channel(channel, efx)
2157 efx_schedule_channel(channel);
2162 /**************************************************************************
2164 * Kernel net device interface
2166 *************************************************************************/
2168 /* Context: process, rtnl_lock() held. */
2169 int efx_net_open(struct net_device *net_dev)
2171 struct efx_nic *efx = netdev_priv(net_dev);
2174 netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2175 raw_smp_processor_id());
2177 rc = efx_check_disabled(efx);
2180 if (efx->phy_mode & PHY_MODE_SPECIAL)
2182 if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2185 /* Notify the kernel of the link state polled during driver load,
2186 * before the monitor starts running */
2187 efx_link_status_changed(efx);
2190 if (efx->state == STATE_DISABLED || efx->reset_pending)
2191 netif_device_detach(efx->net_dev);
2192 efx_selftest_async_start(efx);
2196 /* Context: process, rtnl_lock() held.
2197 * Note that the kernel will ignore our return code; this method
2198 * should really be a void.
2200 int efx_net_stop(struct net_device *net_dev)
2202 struct efx_nic *efx = netdev_priv(net_dev);
2204 netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2205 raw_smp_processor_id());
2207 /* Stop the device and flush all the channels */
2213 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2214 static void efx_net_stats(struct net_device *net_dev,
2215 struct rtnl_link_stats64 *stats)
2217 struct efx_nic *efx = netdev_priv(net_dev);
2219 spin_lock_bh(&efx->stats_lock);
2220 efx->type->update_stats(efx, NULL, stats);
2221 spin_unlock_bh(&efx->stats_lock);
2224 /* Context: netif_tx_lock held, BHs disabled. */
2225 static void efx_watchdog(struct net_device *net_dev)
2227 struct efx_nic *efx = netdev_priv(net_dev);
2229 netif_err(efx, tx_err, efx->net_dev,
2230 "TX stuck with port_enabled=%d: resetting channels\n",
2233 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2237 /* Context: process, rtnl_lock() held. */
2238 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2240 struct efx_nic *efx = netdev_priv(net_dev);
2243 rc = efx_check_disabled(efx);
2247 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2249 efx_device_detach_sync(efx);
2252 mutex_lock(&efx->mac_lock);
2253 net_dev->mtu = new_mtu;
2254 efx_mac_reconfigure(efx);
2255 mutex_unlock(&efx->mac_lock);
2258 efx_device_attach_if_not_resetting(efx);
2262 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2264 struct efx_nic *efx = netdev_priv(net_dev);
2265 struct sockaddr *addr = data;
2266 u8 *new_addr = addr->sa_data;
2270 if (!is_valid_ether_addr(new_addr)) {
2271 netif_err(efx, drv, efx->net_dev,
2272 "invalid ethernet MAC address requested: %pM\n",
2274 return -EADDRNOTAVAIL;
2277 /* save old address */
2278 ether_addr_copy(old_addr, net_dev->dev_addr);
2279 ether_addr_copy(net_dev->dev_addr, new_addr);
2280 if (efx->type->set_mac_address) {
2281 rc = efx->type->set_mac_address(efx);
2283 ether_addr_copy(net_dev->dev_addr, old_addr);
2288 /* Reconfigure the MAC */
2289 mutex_lock(&efx->mac_lock);
2290 efx_mac_reconfigure(efx);
2291 mutex_unlock(&efx->mac_lock);
2296 /* Context: netif_addr_lock held, BHs disabled. */
2297 static void efx_set_rx_mode(struct net_device *net_dev)
2299 struct efx_nic *efx = netdev_priv(net_dev);
2301 if (efx->port_enabled)
2302 queue_work(efx->workqueue, &efx->mac_work);
2303 /* Otherwise efx_start_port() will do this */
2306 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2308 struct efx_nic *efx = netdev_priv(net_dev);
2311 /* If disabling RX n-tuple filtering, clear existing filters */
2312 if (net_dev->features & ~data & NETIF_F_NTUPLE) {
2313 rc = efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2318 /* If Rx VLAN filter is changed, update filters via mac_reconfigure.
2319 * If rx-fcs is changed, mac_reconfigure updates that too.
2321 if ((net_dev->features ^ data) & (NETIF_F_HW_VLAN_CTAG_FILTER |
2323 /* efx_set_rx_mode() will schedule MAC work to update filters
2324 * when a new features are finally set in net_dev.
2326 efx_set_rx_mode(net_dev);
2332 static int efx_get_phys_port_id(struct net_device *net_dev,
2333 struct netdev_phys_item_id *ppid)
2335 struct efx_nic *efx = netdev_priv(net_dev);
2337 if (efx->type->get_phys_port_id)
2338 return efx->type->get_phys_port_id(efx, ppid);
2343 static int efx_get_phys_port_name(struct net_device *net_dev,
2344 char *name, size_t len)
2346 struct efx_nic *efx = netdev_priv(net_dev);
2348 if (snprintf(name, len, "p%u", efx->port_num) >= len)
2353 static int efx_vlan_rx_add_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2355 struct efx_nic *efx = netdev_priv(net_dev);
2357 if (efx->type->vlan_rx_add_vid)
2358 return efx->type->vlan_rx_add_vid(efx, proto, vid);
2363 static int efx_vlan_rx_kill_vid(struct net_device *net_dev, __be16 proto, u16 vid)
2365 struct efx_nic *efx = netdev_priv(net_dev);
2367 if (efx->type->vlan_rx_kill_vid)
2368 return efx->type->vlan_rx_kill_vid(efx, proto, vid);
2373 static int efx_udp_tunnel_type_map(enum udp_parsable_tunnel_type in)
2376 case UDP_TUNNEL_TYPE_VXLAN:
2377 return TUNNEL_ENCAP_UDP_PORT_ENTRY_VXLAN;
2378 case UDP_TUNNEL_TYPE_GENEVE:
2379 return TUNNEL_ENCAP_UDP_PORT_ENTRY_GENEVE;
2385 static void efx_udp_tunnel_add(struct net_device *dev, struct udp_tunnel_info *ti)
2387 struct efx_nic *efx = netdev_priv(dev);
2388 struct efx_udp_tunnel tnl;
2389 int efx_tunnel_type;
2391 efx_tunnel_type = efx_udp_tunnel_type_map(ti->type);
2392 if (efx_tunnel_type < 0)
2395 tnl.type = (u16)efx_tunnel_type;
2396 tnl.port = ti->port;
2398 if (efx->type->udp_tnl_add_port)
2399 (void)efx->type->udp_tnl_add_port(efx, tnl);
2402 static void efx_udp_tunnel_del(struct net_device *dev, struct udp_tunnel_info *ti)
2404 struct efx_nic *efx = netdev_priv(dev);
2405 struct efx_udp_tunnel tnl;
2406 int efx_tunnel_type;
2408 efx_tunnel_type = efx_udp_tunnel_type_map(ti->type);
2409 if (efx_tunnel_type < 0)
2412 tnl.type = (u16)efx_tunnel_type;
2413 tnl.port = ti->port;
2415 if (efx->type->udp_tnl_del_port)
2416 (void)efx->type->udp_tnl_del_port(efx, tnl);
2419 static const struct net_device_ops efx_netdev_ops = {
2420 .ndo_open = efx_net_open,
2421 .ndo_stop = efx_net_stop,
2422 .ndo_get_stats64 = efx_net_stats,
2423 .ndo_tx_timeout = efx_watchdog,
2424 .ndo_start_xmit = efx_hard_start_xmit,
2425 .ndo_validate_addr = eth_validate_addr,
2426 .ndo_do_ioctl = efx_ioctl,
2427 .ndo_change_mtu = efx_change_mtu,
2428 .ndo_set_mac_address = efx_set_mac_address,
2429 .ndo_set_rx_mode = efx_set_rx_mode,
2430 .ndo_set_features = efx_set_features,
2431 .ndo_vlan_rx_add_vid = efx_vlan_rx_add_vid,
2432 .ndo_vlan_rx_kill_vid = efx_vlan_rx_kill_vid,
2433 #ifdef CONFIG_SFC_SRIOV
2434 .ndo_set_vf_mac = efx_sriov_set_vf_mac,
2435 .ndo_set_vf_vlan = efx_sriov_set_vf_vlan,
2436 .ndo_set_vf_spoofchk = efx_sriov_set_vf_spoofchk,
2437 .ndo_get_vf_config = efx_sriov_get_vf_config,
2438 .ndo_set_vf_link_state = efx_sriov_set_vf_link_state,
2440 .ndo_get_phys_port_id = efx_get_phys_port_id,
2441 .ndo_get_phys_port_name = efx_get_phys_port_name,
2442 #ifdef CONFIG_NET_POLL_CONTROLLER
2443 .ndo_poll_controller = efx_netpoll,
2445 .ndo_setup_tc = efx_setup_tc,
2446 #ifdef CONFIG_RFS_ACCEL
2447 .ndo_rx_flow_steer = efx_filter_rfs,
2449 .ndo_udp_tunnel_add = efx_udp_tunnel_add,
2450 .ndo_udp_tunnel_del = efx_udp_tunnel_del,
2453 static void efx_update_name(struct efx_nic *efx)
2455 strcpy(efx->name, efx->net_dev->name);
2456 efx_mtd_rename(efx);
2457 efx_set_channel_names(efx);
2460 static int efx_netdev_event(struct notifier_block *this,
2461 unsigned long event, void *ptr)
2463 struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2465 if ((net_dev->netdev_ops == &efx_netdev_ops) &&
2466 event == NETDEV_CHANGENAME)
2467 efx_update_name(netdev_priv(net_dev));
2472 static struct notifier_block efx_netdev_notifier = {
2473 .notifier_call = efx_netdev_event,
2477 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2479 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2480 return sprintf(buf, "%d\n", efx->phy_type);
2482 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2484 #ifdef CONFIG_SFC_MCDI_LOGGING
2485 static ssize_t show_mcdi_log(struct device *dev, struct device_attribute *attr,
2488 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2489 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2491 return scnprintf(buf, PAGE_SIZE, "%d\n", mcdi->logging_enabled);
2493 static ssize_t set_mcdi_log(struct device *dev, struct device_attribute *attr,
2494 const char *buf, size_t count)
2496 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2497 struct efx_mcdi_iface *mcdi = efx_mcdi(efx);
2498 bool enable = count > 0 && *buf != '0';
2500 mcdi->logging_enabled = enable;
2503 static DEVICE_ATTR(mcdi_logging, 0644, show_mcdi_log, set_mcdi_log);
2506 static int efx_register_netdev(struct efx_nic *efx)
2508 struct net_device *net_dev = efx->net_dev;
2509 struct efx_channel *channel;
2512 net_dev->watchdog_timeo = 5 * HZ;
2513 net_dev->irq = efx->pci_dev->irq;
2514 net_dev->netdev_ops = &efx_netdev_ops;
2515 if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
2516 net_dev->priv_flags |= IFF_UNICAST_FLT;
2517 net_dev->ethtool_ops = &efx_ethtool_ops;
2518 net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2519 net_dev->min_mtu = EFX_MIN_MTU;
2520 net_dev->max_mtu = EFX_MAX_MTU;
2524 /* Enable resets to be scheduled and check whether any were
2525 * already requested. If so, the NIC is probably hosed so we
2528 efx->state = STATE_READY;
2529 smp_mb(); /* ensure we change state before checking reset_pending */
2530 if (efx->reset_pending) {
2531 netif_err(efx, probe, efx->net_dev,
2532 "aborting probe due to scheduled reset\n");
2537 rc = dev_alloc_name(net_dev, net_dev->name);
2540 efx_update_name(efx);
2542 /* Always start with carrier off; PHY events will detect the link */
2543 netif_carrier_off(net_dev);
2545 rc = register_netdevice(net_dev);
2549 efx_for_each_channel(channel, efx) {
2550 struct efx_tx_queue *tx_queue;
2551 efx_for_each_channel_tx_queue(tx_queue, channel)
2552 efx_init_tx_queue_core_txq(tx_queue);
2559 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2561 netif_err(efx, drv, efx->net_dev,
2562 "failed to init net dev attributes\n");
2563 goto fail_registered;
2565 #ifdef CONFIG_SFC_MCDI_LOGGING
2566 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2568 netif_err(efx, drv, efx->net_dev,
2569 "failed to init net dev attributes\n");
2570 goto fail_attr_mcdi_logging;
2576 #ifdef CONFIG_SFC_MCDI_LOGGING
2577 fail_attr_mcdi_logging:
2578 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2582 efx_dissociate(efx);
2583 unregister_netdevice(net_dev);
2585 efx->state = STATE_UNINIT;
2587 netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2591 static void efx_unregister_netdev(struct efx_nic *efx)
2596 BUG_ON(netdev_priv(efx->net_dev) != efx);
2598 if (efx_dev_registered(efx)) {
2599 strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2600 #ifdef CONFIG_SFC_MCDI_LOGGING
2601 device_remove_file(&efx->pci_dev->dev, &dev_attr_mcdi_logging);
2603 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2604 unregister_netdev(efx->net_dev);
2608 /**************************************************************************
2610 * Device reset and suspend
2612 **************************************************************************/
2614 /* Tears down the entire software state and most of the hardware state
2616 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2618 EFX_ASSERT_RESET_SERIALISED(efx);
2620 if (method == RESET_TYPE_MCDI_TIMEOUT)
2621 efx->type->prepare_flr(efx);
2624 efx_disable_interrupts(efx);
2626 mutex_lock(&efx->mac_lock);
2627 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2628 method != RESET_TYPE_DATAPATH)
2629 efx->phy_op->fini(efx);
2630 efx->type->fini(efx);
2633 /* This function will always ensure that the locks acquired in
2634 * efx_reset_down() are released. A failure return code indicates
2635 * that we were unable to reinitialise the hardware, and the
2636 * driver should be disabled. If ok is false, then the rx and tx
2637 * engines are not restarted, pending a RESET_DISABLE. */
2638 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2642 EFX_ASSERT_RESET_SERIALISED(efx);
2644 if (method == RESET_TYPE_MCDI_TIMEOUT)
2645 efx->type->finish_flr(efx);
2647 /* Ensure that SRAM is initialised even if we're disabling the device */
2648 rc = efx->type->init(efx);
2650 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2657 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2658 method != RESET_TYPE_DATAPATH) {
2659 rc = efx->phy_op->init(efx);
2662 rc = efx->phy_op->reconfigure(efx);
2663 if (rc && rc != -EPERM)
2664 netif_err(efx, drv, efx->net_dev,
2665 "could not restore PHY settings\n");
2668 rc = efx_enable_interrupts(efx);
2672 #ifdef CONFIG_SFC_SRIOV
2673 rc = efx->type->vswitching_restore(efx);
2674 if (rc) /* not fatal; the PF will still work fine */
2675 netif_warn(efx, probe, efx->net_dev,
2676 "failed to restore vswitching rc=%d;"
2677 " VFs may not function\n", rc);
2680 down_read(&efx->filter_sem);
2681 efx_restore_filters(efx);
2682 up_read(&efx->filter_sem);
2683 if (efx->type->sriov_reset)
2684 efx->type->sriov_reset(efx);
2686 mutex_unlock(&efx->mac_lock);
2690 if (efx->type->udp_tnl_push_ports)
2691 efx->type->udp_tnl_push_ports(efx);
2696 efx->port_initialized = false;
2698 mutex_unlock(&efx->mac_lock);
2703 /* Reset the NIC using the specified method. Note that the reset may
2704 * fail, in which case the card will be left in an unusable state.
2706 * Caller must hold the rtnl_lock.
2708 int efx_reset(struct efx_nic *efx, enum reset_type method)
2713 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2714 RESET_TYPE(method));
2716 efx_device_detach_sync(efx);
2717 efx_reset_down(efx, method);
2719 rc = efx->type->reset(efx, method);
2721 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2725 /* Clear flags for the scopes we covered. We assume the NIC and
2726 * driver are now quiescent so that there is no race here.
2728 if (method < RESET_TYPE_MAX_METHOD)
2729 efx->reset_pending &= -(1 << (method + 1));
2730 else /* it doesn't fit into the well-ordered scope hierarchy */
2731 __clear_bit(method, &efx->reset_pending);
2733 /* Reinitialise bus-mastering, which may have been turned off before
2734 * the reset was scheduled. This is still appropriate, even in the
2735 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2736 * can respond to requests. */
2737 pci_set_master(efx->pci_dev);
2740 /* Leave device stopped if necessary */
2742 method == RESET_TYPE_DISABLE ||
2743 method == RESET_TYPE_RECOVER_OR_DISABLE;
2744 rc2 = efx_reset_up(efx, method, !disabled);
2752 dev_close(efx->net_dev);
2753 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2754 efx->state = STATE_DISABLED;
2756 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2757 efx_device_attach_if_not_resetting(efx);
2762 /* Try recovery mechanisms.
2763 * For now only EEH is supported.
2764 * Returns 0 if the recovery mechanisms are unsuccessful.
2765 * Returns a non-zero value otherwise.
2767 int efx_try_recovery(struct efx_nic *efx)
2770 /* A PCI error can occur and not be seen by EEH because nothing
2771 * happens on the PCI bus. In this case the driver may fail and
2772 * schedule a 'recover or reset', leading to this recovery handler.
2773 * Manually call the eeh failure check function.
2775 struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2776 if (eeh_dev_check_failure(eehdev)) {
2777 /* The EEH mechanisms will handle the error and reset the
2778 * device if necessary.
2786 static void efx_wait_for_bist_end(struct efx_nic *efx)
2790 for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2791 if (efx_mcdi_poll_reboot(efx))
2793 msleep(BIST_WAIT_DELAY_MS);
2796 netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2798 /* Either way unset the BIST flag. If we found no reboot we probably
2799 * won't recover, but we should try.
2801 efx->mc_bist_for_other_fn = false;
2804 /* The worker thread exists so that code that cannot sleep can
2805 * schedule a reset for later.
2807 static void efx_reset_work(struct work_struct *data)
2809 struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2810 unsigned long pending;
2811 enum reset_type method;
2813 pending = READ_ONCE(efx->reset_pending);
2814 method = fls(pending) - 1;
2816 if (method == RESET_TYPE_MC_BIST)
2817 efx_wait_for_bist_end(efx);
2819 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2820 method == RESET_TYPE_RECOVER_OR_ALL) &&
2821 efx_try_recovery(efx))
2829 /* We checked the state in efx_schedule_reset() but it may
2830 * have changed by now. Now that we have the RTNL lock,
2831 * it cannot change again.
2833 if (efx->state == STATE_READY)
2834 (void)efx_reset(efx, method);
2839 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2841 enum reset_type method;
2843 if (efx->state == STATE_RECOVERY) {
2844 netif_dbg(efx, drv, efx->net_dev,
2845 "recovering: skip scheduling %s reset\n",
2851 case RESET_TYPE_INVISIBLE:
2852 case RESET_TYPE_ALL:
2853 case RESET_TYPE_RECOVER_OR_ALL:
2854 case RESET_TYPE_WORLD:
2855 case RESET_TYPE_DISABLE:
2856 case RESET_TYPE_RECOVER_OR_DISABLE:
2857 case RESET_TYPE_DATAPATH:
2858 case RESET_TYPE_MC_BIST:
2859 case RESET_TYPE_MCDI_TIMEOUT:
2861 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2862 RESET_TYPE(method));
2865 method = efx->type->map_reset_reason(type);
2866 netif_dbg(efx, drv, efx->net_dev,
2867 "scheduling %s reset for %s\n",
2868 RESET_TYPE(method), RESET_TYPE(type));
2872 set_bit(method, &efx->reset_pending);
2873 smp_mb(); /* ensure we change reset_pending before checking state */
2875 /* If we're not READY then just leave the flags set as the cue
2876 * to abort probing or reschedule the reset later.
2878 if (READ_ONCE(efx->state) != STATE_READY)
2881 /* efx_process_channel() will no longer read events once a
2882 * reset is scheduled. So switch back to poll'd MCDI completions. */
2883 efx_mcdi_mode_poll(efx);
2885 queue_work(reset_workqueue, &efx->reset_work);
2888 /**************************************************************************
2890 * List of NICs we support
2892 **************************************************************************/
2894 /* PCI device ID table */
2895 static const struct pci_device_id efx_pci_table[] = {
2896 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803), /* SFC9020 */
2897 .driver_data = (unsigned long) &siena_a0_nic_type},
2898 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813), /* SFL9021 */
2899 .driver_data = (unsigned long) &siena_a0_nic_type},
2900 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903), /* SFC9120 PF */
2901 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2902 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903), /* SFC9120 VF */
2903 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2904 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923), /* SFC9140 PF */
2905 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2906 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1923), /* SFC9140 VF */
2907 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2908 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0a03), /* SFC9220 PF */
2909 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2910 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1a03), /* SFC9220 VF */
2911 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
2912 {0} /* end of list */
2915 /**************************************************************************
2917 * Dummy PHY/MAC operations
2919 * Can be used for some unimplemented operations
2920 * Needed so all function pointers are valid and do not have to be tested
2923 **************************************************************************/
2924 int efx_port_dummy_op_int(struct efx_nic *efx)
2928 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2930 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2935 static const struct efx_phy_operations efx_dummy_phy_operations = {
2936 .init = efx_port_dummy_op_int,
2937 .reconfigure = efx_port_dummy_op_int,
2938 .poll = efx_port_dummy_op_poll,
2939 .fini = efx_port_dummy_op_void,
2942 /**************************************************************************
2946 **************************************************************************/
2948 /* This zeroes out and then fills in the invariants in a struct
2949 * efx_nic (including all sub-structures).
2951 static int efx_init_struct(struct efx_nic *efx,
2952 struct pci_dev *pci_dev, struct net_device *net_dev)
2954 int rc = -ENOMEM, i;
2956 /* Initialise common structures */
2957 INIT_LIST_HEAD(&efx->node);
2958 INIT_LIST_HEAD(&efx->secondary_list);
2959 spin_lock_init(&efx->biu_lock);
2960 #ifdef CONFIG_SFC_MTD
2961 INIT_LIST_HEAD(&efx->mtd_list);
2963 INIT_WORK(&efx->reset_work, efx_reset_work);
2964 INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2965 INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2966 efx->pci_dev = pci_dev;
2967 efx->msg_enable = debug;
2968 efx->state = STATE_UNINIT;
2969 strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2971 efx->net_dev = net_dev;
2972 efx->rx_prefix_size = efx->type->rx_prefix_size;
2974 NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2975 efx->rx_packet_hash_offset =
2976 efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2977 efx->rx_packet_ts_offset =
2978 efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2979 spin_lock_init(&efx->stats_lock);
2980 mutex_init(&efx->mac_lock);
2981 efx->phy_op = &efx_dummy_phy_operations;
2982 efx->mdio.dev = net_dev;
2983 INIT_WORK(&efx->mac_work, efx_mac_work);
2984 init_waitqueue_head(&efx->flush_wq);
2986 for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2987 efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2988 if (!efx->channel[i])
2990 efx->msi_context[i].efx = efx;
2991 efx->msi_context[i].index = i;
2994 /* Higher numbered interrupt modes are less capable! */
2995 if (WARN_ON_ONCE(efx->type->max_interrupt_mode >
2996 efx->type->min_interrupt_mode)) {
3000 efx->interrupt_mode = max(efx->type->max_interrupt_mode,
3002 efx->interrupt_mode = min(efx->type->min_interrupt_mode,
3005 /* Would be good to use the net_dev name, but we're too early */
3006 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
3008 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
3009 if (!efx->workqueue)
3015 efx_fini_struct(efx);
3019 static void efx_fini_struct(struct efx_nic *efx)
3023 for (i = 0; i < EFX_MAX_CHANNELS; i++)
3024 kfree(efx->channel[i]);
3028 if (efx->workqueue) {
3029 destroy_workqueue(efx->workqueue);
3030 efx->workqueue = NULL;
3034 void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
3036 u64 n_rx_nodesc_trunc = 0;
3037 struct efx_channel *channel;
3039 efx_for_each_channel(channel, efx)
3040 n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
3041 stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
3042 stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
3045 /**************************************************************************
3049 **************************************************************************/
3051 /* Main body of final NIC shutdown code
3052 * This is called only at module unload (or hotplug removal).
3054 static void efx_pci_remove_main(struct efx_nic *efx)
3056 /* Flush reset_work. It can no longer be scheduled since we
3059 BUG_ON(efx->state == STATE_READY);
3060 cancel_work_sync(&efx->reset_work);
3062 efx_disable_interrupts(efx);
3063 efx_nic_fini_interrupt(efx);
3065 efx->type->fini(efx);
3067 efx_remove_all(efx);
3070 /* Final NIC shutdown
3071 * This is called only at module unload (or hotplug removal). A PF can call
3072 * this on its VFs to ensure they are unbound first.
3074 static void efx_pci_remove(struct pci_dev *pci_dev)
3076 struct efx_nic *efx;
3078 efx = pci_get_drvdata(pci_dev);
3082 /* Mark the NIC as fini, then stop the interface */
3084 efx_dissociate(efx);
3085 dev_close(efx->net_dev);
3086 efx_disable_interrupts(efx);
3087 efx->state = STATE_UNINIT;
3090 if (efx->type->sriov_fini)
3091 efx->type->sriov_fini(efx);
3093 efx_unregister_netdev(efx);
3095 efx_mtd_remove(efx);
3097 efx_pci_remove_main(efx);
3100 netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
3102 efx_fini_struct(efx);
3103 free_netdev(efx->net_dev);
3105 pci_disable_pcie_error_reporting(pci_dev);
3108 /* NIC VPD information
3109 * Called during probe to display the part number of the
3110 * installed NIC. VPD is potentially very large but this should
3111 * always appear within the first 512 bytes.
3113 #define SFC_VPD_LEN 512
3114 static void efx_probe_vpd_strings(struct efx_nic *efx)
3116 struct pci_dev *dev = efx->pci_dev;
3117 char vpd_data[SFC_VPD_LEN];
3119 int ro_start, ro_size, i, j;
3121 /* Get the vpd data from the device */
3122 vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
3123 if (vpd_size <= 0) {
3124 netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
3128 /* Get the Read only section */
3129 ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
3131 netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
3135 ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
3137 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3138 if (i + j > vpd_size)
3141 /* Get the Part number */
3142 i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
3144 netif_err(efx, drv, efx->net_dev, "Part number not found\n");
3148 j = pci_vpd_info_field_size(&vpd_data[i]);
3149 i += PCI_VPD_INFO_FLD_HDR_SIZE;
3150 if (i + j > vpd_size) {
3151 netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
3155 netif_info(efx, drv, efx->net_dev,
3156 "Part Number : %.*s\n", j, &vpd_data[i]);
3158 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
3160 i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
3162 netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
3166 j = pci_vpd_info_field_size(&vpd_data[i]);
3167 i += PCI_VPD_INFO_FLD_HDR_SIZE;
3168 if (i + j > vpd_size) {
3169 netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
3173 efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
3177 snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
3181 /* Main body of NIC initialisation
3182 * This is called at module load (or hotplug insertion, theoretically).
3184 static int efx_pci_probe_main(struct efx_nic *efx)
3188 /* Do start-of-day initialisation */
3189 rc = efx_probe_all(efx);
3195 rc = efx->type->init(efx);
3197 netif_err(efx, probe, efx->net_dev,
3198 "failed to initialise NIC\n");
3202 rc = efx_init_port(efx);
3204 netif_err(efx, probe, efx->net_dev,
3205 "failed to initialise port\n");
3209 rc = efx_nic_init_interrupt(efx);
3212 rc = efx_enable_interrupts(efx);
3219 efx_nic_fini_interrupt(efx);
3223 efx->type->fini(efx);
3226 efx_remove_all(efx);
3231 static int efx_pci_probe_post_io(struct efx_nic *efx)
3233 struct net_device *net_dev = efx->net_dev;
3234 int rc = efx_pci_probe_main(efx);
3239 if (efx->type->sriov_init) {
3240 rc = efx->type->sriov_init(efx);
3242 netif_err(efx, probe, efx->net_dev,
3243 "SR-IOV can't be enabled rc %d\n", rc);
3246 /* Determine netdevice features */
3247 net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
3248 NETIF_F_TSO | NETIF_F_RXCSUM | NETIF_F_RXALL);
3249 if (efx->type->offload_features & (NETIF_F_IPV6_CSUM | NETIF_F_HW_CSUM))
3250 net_dev->features |= NETIF_F_TSO6;
3251 /* Check whether device supports TSO */
3252 if (!efx->type->tso_versions || !efx->type->tso_versions(efx))
3253 net_dev->features &= ~NETIF_F_ALL_TSO;
3254 /* Mask for features that also apply to VLAN devices */
3255 net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
3256 NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
3259 net_dev->hw_features |= net_dev->features & ~efx->fixed_features;
3261 /* Disable receiving frames with bad FCS, by default. */
3262 net_dev->features &= ~NETIF_F_RXALL;
3264 /* Disable VLAN filtering by default. It may be enforced if
3265 * the feature is fixed (i.e. VLAN filters are required to
3266 * receive VLAN tagged packets due to vPort restrictions).
3268 net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
3269 net_dev->features |= efx->fixed_features;
3271 rc = efx_register_netdev(efx);
3275 efx_pci_remove_main(efx);
3279 /* NIC initialisation
3281 * This is called at module load (or hotplug insertion,
3282 * theoretically). It sets up PCI mappings, resets the NIC,
3283 * sets up and registers the network devices with the kernel and hooks
3284 * the interrupt service routine. It does not prepare the device for
3285 * transmission; this is left to the first time one of the network
3286 * interfaces is brought up (i.e. efx_net_open).
3288 static int efx_pci_probe(struct pci_dev *pci_dev,
3289 const struct pci_device_id *entry)
3291 struct net_device *net_dev;
3292 struct efx_nic *efx;
3295 /* Allocate and initialise a struct net_device and struct efx_nic */
3296 net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
3300 efx = netdev_priv(net_dev);
3301 efx->type = (const struct efx_nic_type *) entry->driver_data;
3302 efx->fixed_features |= NETIF_F_HIGHDMA;
3304 pci_set_drvdata(pci_dev, efx);
3305 SET_NETDEV_DEV(net_dev, &pci_dev->dev);
3306 rc = efx_init_struct(efx, pci_dev, net_dev);
3310 netif_info(efx, probe, efx->net_dev,
3311 "Solarflare NIC detected\n");
3313 if (!efx->type->is_vf)
3314 efx_probe_vpd_strings(efx);
3316 /* Set up basic I/O (BAR mappings etc) */
3317 rc = efx_init_io(efx);
3321 rc = efx_pci_probe_post_io(efx);
3323 /* On failure, retry once immediately.
3324 * If we aborted probe due to a scheduled reset, dismiss it.
3326 efx->reset_pending = 0;
3327 rc = efx_pci_probe_post_io(efx);
3329 /* On another failure, retry once more
3330 * after a 50-305ms delay.
3334 get_random_bytes(&r, 1);
3335 msleep((unsigned int)r + 50);
3336 efx->reset_pending = 0;
3337 rc = efx_pci_probe_post_io(efx);
3343 netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3345 /* Try to create MTDs, but allow this to fail */
3347 rc = efx_mtd_probe(efx);
3349 if (rc && rc != -EPERM)
3350 netif_warn(efx, probe, efx->net_dev,
3351 "failed to create MTDs (%d)\n", rc);
3353 rc = pci_enable_pcie_error_reporting(pci_dev);
3354 if (rc && rc != -EINVAL)
3355 netif_notice(efx, probe, efx->net_dev,
3356 "PCIE error reporting unavailable (%d).\n",
3359 if (efx->type->udp_tnl_push_ports)
3360 efx->type->udp_tnl_push_ports(efx);
3367 efx_fini_struct(efx);
3370 netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3371 free_netdev(net_dev);
3375 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3376 * enabled on success
3378 #ifdef CONFIG_SFC_SRIOV
3379 static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
3382 struct efx_nic *efx = pci_get_drvdata(dev);
3384 if (efx->type->sriov_configure) {
3385 rc = efx->type->sriov_configure(efx, num_vfs);
3395 static int efx_pm_freeze(struct device *dev)
3397 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3401 if (efx->state != STATE_DISABLED) {
3402 efx->state = STATE_UNINIT;
3404 efx_device_detach_sync(efx);
3407 efx_disable_interrupts(efx);
3415 static int efx_pm_thaw(struct device *dev)
3418 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3422 if (efx->state != STATE_DISABLED) {
3423 rc = efx_enable_interrupts(efx);
3427 mutex_lock(&efx->mac_lock);
3428 efx->phy_op->reconfigure(efx);
3429 mutex_unlock(&efx->mac_lock);
3433 efx_device_attach_if_not_resetting(efx);
3435 efx->state = STATE_READY;
3437 efx->type->resume_wol(efx);
3442 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3443 queue_work(reset_workqueue, &efx->reset_work);
3453 static int efx_pm_poweroff(struct device *dev)
3455 struct pci_dev *pci_dev = to_pci_dev(dev);
3456 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3458 efx->type->fini(efx);
3460 efx->reset_pending = 0;
3462 pci_save_state(pci_dev);
3463 return pci_set_power_state(pci_dev, PCI_D3hot);
3466 /* Used for both resume and restore */
3467 static int efx_pm_resume(struct device *dev)
3469 struct pci_dev *pci_dev = to_pci_dev(dev);
3470 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3473 rc = pci_set_power_state(pci_dev, PCI_D0);
3476 pci_restore_state(pci_dev);
3477 rc = pci_enable_device(pci_dev);
3480 pci_set_master(efx->pci_dev);
3481 rc = efx->type->reset(efx, RESET_TYPE_ALL);
3484 rc = efx->type->init(efx);
3487 rc = efx_pm_thaw(dev);
3491 static int efx_pm_suspend(struct device *dev)
3496 rc = efx_pm_poweroff(dev);
3502 static const struct dev_pm_ops efx_pm_ops = {
3503 .suspend = efx_pm_suspend,
3504 .resume = efx_pm_resume,
3505 .freeze = efx_pm_freeze,
3506 .thaw = efx_pm_thaw,
3507 .poweroff = efx_pm_poweroff,
3508 .restore = efx_pm_resume,
3511 /* A PCI error affecting this device was detected.
3512 * At this point MMIO and DMA may be disabled.
3513 * Stop the software path and request a slot reset.
3515 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3516 enum pci_channel_state state)
3518 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3519 struct efx_nic *efx = pci_get_drvdata(pdev);
3521 if (state == pci_channel_io_perm_failure)
3522 return PCI_ERS_RESULT_DISCONNECT;
3526 if (efx->state != STATE_DISABLED) {
3527 efx->state = STATE_RECOVERY;
3528 efx->reset_pending = 0;
3530 efx_device_detach_sync(efx);
3533 efx_disable_interrupts(efx);
3535 status = PCI_ERS_RESULT_NEED_RESET;
3537 /* If the interface is disabled we don't want to do anything
3540 status = PCI_ERS_RESULT_RECOVERED;
3545 pci_disable_device(pdev);
3550 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3551 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3553 struct efx_nic *efx = pci_get_drvdata(pdev);
3554 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3557 if (pci_enable_device(pdev)) {
3558 netif_err(efx, hw, efx->net_dev,
3559 "Cannot re-enable PCI device after reset.\n");
3560 status = PCI_ERS_RESULT_DISCONNECT;
3563 rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3565 netif_err(efx, hw, efx->net_dev,
3566 "pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3567 /* Non-fatal error. Continue. */
3573 /* Perform the actual reset and resume I/O operations. */
3574 static void efx_io_resume(struct pci_dev *pdev)
3576 struct efx_nic *efx = pci_get_drvdata(pdev);
3581 if (efx->state == STATE_DISABLED)
3584 rc = efx_reset(efx, RESET_TYPE_ALL);
3586 netif_err(efx, hw, efx->net_dev,
3587 "efx_reset failed after PCI error (%d)\n", rc);
3589 efx->state = STATE_READY;
3590 netif_dbg(efx, hw, efx->net_dev,
3591 "Done resetting and resuming IO after PCI error.\n");
3598 /* For simplicity and reliability, we always require a slot reset and try to
3599 * reset the hardware when a pci error affecting the device is detected.
3600 * We leave both the link_reset and mmio_enabled callback unimplemented:
3601 * with our request for slot reset the mmio_enabled callback will never be
3602 * called, and the link_reset callback is not used by AER or EEH mechanisms.
3604 static const struct pci_error_handlers efx_err_handlers = {
3605 .error_detected = efx_io_error_detected,
3606 .slot_reset = efx_io_slot_reset,
3607 .resume = efx_io_resume,
3610 static struct pci_driver efx_pci_driver = {
3611 .name = KBUILD_MODNAME,
3612 .id_table = efx_pci_table,
3613 .probe = efx_pci_probe,
3614 .remove = efx_pci_remove,
3615 .driver.pm = &efx_pm_ops,
3616 .err_handler = &efx_err_handlers,
3617 #ifdef CONFIG_SFC_SRIOV
3618 .sriov_configure = efx_pci_sriov_configure,
3622 /**************************************************************************
3624 * Kernel module interface
3626 *************************************************************************/
3628 module_param(interrupt_mode, uint, 0444);
3629 MODULE_PARM_DESC(interrupt_mode,
3630 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3632 static int __init efx_init_module(void)
3636 printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3638 rc = register_netdevice_notifier(&efx_netdev_notifier);
3642 #ifdef CONFIG_SFC_SRIOV
3643 rc = efx_init_sriov();
3648 reset_workqueue = create_singlethread_workqueue("sfc_reset");
3649 if (!reset_workqueue) {
3654 rc = pci_register_driver(&efx_pci_driver);
3661 destroy_workqueue(reset_workqueue);
3663 #ifdef CONFIG_SFC_SRIOV
3667 unregister_netdevice_notifier(&efx_netdev_notifier);
3672 static void __exit efx_exit_module(void)
3674 printk(KERN_INFO "Solarflare NET driver unloading\n");
3676 pci_unregister_driver(&efx_pci_driver);
3677 destroy_workqueue(reset_workqueue);
3678 #ifdef CONFIG_SFC_SRIOV
3681 unregister_netdevice_notifier(&efx_netdev_notifier);
3685 module_init(efx_init_module);
3686 module_exit(efx_exit_module);
3688 MODULE_AUTHOR("Solarflare Communications and "
3689 "Michael Brown <mbrown@fensystems.co.uk>");
3690 MODULE_DESCRIPTION("Solarflare network driver");
3691 MODULE_LICENSE("GPL");
3692 MODULE_DEVICE_TABLE(pci, efx_pci_table);
3693 MODULE_VERSION(EFX_DRIVER_VERSION);